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

   1 /*
   2  * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
   3  *
   4  * This file is part of libFirm.
   5  *
   6  * This file may be distributed and/or modified under the terms of the
   7  * GNU General Public License version 2 as published by the Free Software
   8  * Foundation and appearing in the file LICENSE.GPL included in the
   9  * packaging of this file.
  10  *
  11  * Licensees holding valid libFirm Professional Edition licenses may use
  12  * this file in accordance with the libFirm Commercial License.
  13  * Agreement provided with the Software.
  14  *
  15  * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
  16  * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  17  * PURPOSE.
  18  */
  19
  20 /**
  21  * @file
  22  * @brief       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
  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 "obst.h"
  57 #include "raw_bitset.h"
  58 #include "unionfind.h"
  59 #include "pdeq.h"
  60 #include "hungarian.h"
  61
  62 #include "beabi.h"
  63 #include "bechordal_t.h"
  64 #include "be.h"
  65 #include "beirg.h"
  66 #include "belive_t.h"
  67 #include "bemodule.h"
  68 #include "benode.h"
  69 #include "bera.h"
  70 #include "besched.h"
  71 #include "bespill.h"
  72 #include "bespillutil.h"
  73 #include "beverify.h"
  74 #include "beutil.h"
  75
  76 #define USE_FACTOR                     1.0f
  77 #define DEF_FACTOR                     1.0f
  78 #define NEIGHBOR_FACTOR                0.2f
  79 #define AFF_SHOULD_BE_SAME             0.5f
  80 #define AFF_PHI                        1.0f
  81 #define SPLIT_DELTA                    1.0f
  82 #define MAX_OPTIMISTIC_SPLIT_RECURSION 0
  83
  84 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  85
  86 static struct obstack               obst;
  87 static ir_graph                    *irg;
  88 static const arch_register_class_t *cls;
  89 static const arch_register_req_t   *default_cls_req;
  90 static be_lv_t                     *lv;
  91 static const ir_exec_freq          *execfreqs;
  92 static unsigned                     n_regs;
  93 static unsigned                    *normal_regs;
  94 static int                         *congruence_classes;
  95 static ir_node                    **block_order;
  96 static int                          n_block_order;
  97 static int                          create_preferences        = true;
  98 static int                          create_congruence_classes = true;
  99 static int                          propagate_phi_registers   = true;
 100
 101 static const lc_opt_table_entry_t options[] = {
 102         LC_OPT_ENT_BOOL("prefs", "use preference based coloring", &create_preferences),
 103         LC_OPT_ENT_BOOL("congruences", "create congruence classes", &create_congruence_classes),
 104         LC_OPT_ENT_BOOL("prop_phi", "propagate phi registers", &propagate_phi_registers),
 105         LC_OPT_LAST
 106 };
 107
 108 /** currently active assignments (while processing a basic block)
 109  * maps registers to values(their current copies) */
 110 static ir_node **assignments;
 111
 112 /**
 113  * allocation information: last_uses, register preferences
 114  * the information is per firm-node.
 115  */
 116 struct allocation_info_t {
 117         unsigned  last_uses[2];   /**< bitset indicating last uses (input pos) */
 118         ir_node  *current_value;  /**< copy of the value that should be used */
 119         ir_node  *original_value; /**< for copies point to original value */
 120         float     prefs[0];       /**< register preferences */
 121 };
 122 typedef struct allocation_info_t allocation_info_t;
 123
 124 /** helper datastructure used when sorting register preferences */
 125 struct reg_pref_t {
 126         unsigned num;
 127         float    pref;
 128 };
 129 typedef struct reg_pref_t reg_pref_t;
 130
 131 /** per basic-block information */
 132 struct block_info_t {
 133         bool     processed;       /**< indicate whether block is processed */
 134         ir_node *assignments[0];  /**< register assignments at end of block */
 135 };
 136 typedef struct block_info_t block_info_t;
 137
 138 /**
 139  * Get the allocation info for a node.
 140  * The info is allocated on the first visit of a node.
 141  */
 142 static allocation_info_t *get_allocation_info(ir_node *node)
 143 {
 144         allocation_info_t *info = get_irn_link(node);
 145         if (info == NULL) {
 146                 info = OALLOCFZ(&obst, allocation_info_t, prefs, n_regs);
 147                 info->current_value  = node;
 148                 info->original_value = node;
 149                 set_irn_link(node, info);
 150         }
 151
 152         return info;
 153 }
 154
 155 static allocation_info_t *try_get_allocation_info(const ir_node *node)
 156 {
 157         return (allocation_info_t*) get_irn_link(node);
 158 }
 159
 160 /**
 161  * Get allocation information for a basic block
 162  */
 163 static block_info_t *get_block_info(ir_node *block)
 164 {
 165         block_info_t *info = get_irn_link(block);
 166
 167         assert(is_Block(block));
 168         if (info == NULL) {
 169                 info = OALLOCFZ(&obst, block_info_t, assignments, n_regs);
 170                 set_irn_link(block, info);
 171         }
 172
 173         return info;
 174 }
 175
 176 /**
 177  * Get default register requirement for the current register class
 178  */
 179 static const arch_register_req_t *get_default_req_current_cls(void)
 180 {
 181         if (default_cls_req == NULL) {
 182                 struct obstack      *obst = get_irg_obstack(irg);
 183                 arch_register_req_t *req  = OALLOCZ(obst, arch_register_req_t);
 184
 185                 req->type  = arch_register_req_type_normal;
 186                 req->cls   = cls;
 187                 req->width = 1;
 188
 189                 default_cls_req = req;
 190         }
 191         return default_cls_req;
 192 }
 193
 194 /**
 195  * Link the allocation info of a node to a copy.
 196  * Afterwards, both nodes uses the same allocation info.
 197  * Copy must not have an allocation info assigned yet.
 198  *
 199  * @param copy   the node that gets the allocation info assigned
 200  * @param value  the original node
 201  */
 202 static void mark_as_copy_of(ir_node *copy, ir_node *value)
 203 {
 204         ir_node           *original;
 205         allocation_info_t *info      = get_allocation_info(value);
 206         allocation_info_t *copy_info = get_allocation_info(copy);
 207
 208         /* find original value */
 209         original = info->original_value;
 210         if (original != value) {
 211                 info = get_allocation_info(original);
 212         }
 213
 214         assert(info->original_value == original);
 215         info->current_value = copy;
 216
 217         /* the copy should not be linked to something else yet */
 218         assert(copy_info->original_value == copy);
 219         copy_info->original_value = original;
 220
 221         /* copy over allocation preferences */
 222         memcpy(copy_info->prefs, info->prefs, n_regs * sizeof(copy_info->prefs[0]));
 223 }
 224
 225 /**
 226  * Calculate the penalties for every register on a node and its live neighbors.
 227  *
 228  * @param live_nodes  the set of live nodes at the current position, may be NULL
 229  * @param penalty     the penalty to subtract from
 230  * @param limited     a raw bitset containing the limited set for the node
 231  * @param node        the node
 232  */
 233 static void give_penalties_for_limits(const ir_nodeset_t *live_nodes,
 234                                       float penalty, const unsigned* limited,
 235                                       ir_node *node)
 236 {
 237         ir_nodeset_iterator_t iter;
 238         unsigned              r;
 239         unsigned              n_allowed;
 240         allocation_info_t     *info = get_allocation_info(node);
 241         ir_node               *neighbor;
 242
 243         /* give penalty for all forbidden regs */
 244         for (r = 0; r < n_regs; ++r) {
 245                 if (rbitset_is_set(limited, r))
 246                         continue;
 247
 248                 info->prefs[r] -= penalty;
 249         }
 250
 251         /* all other live values should get a penalty for allowed regs */
 252         if (live_nodes == NULL)
 253                 return;
 254
 255         penalty   *= NEIGHBOR_FACTOR;
 256         n_allowed  = rbitset_popcount(limited, n_regs);
 257         if (n_allowed > 1) {
 258                 /* only create a very weak penalty if multiple regs are allowed */
 259                 penalty = (penalty * 0.8f) / n_allowed;
 260         }
 261         foreach_ir_nodeset(live_nodes, neighbor, iter) {
 262                 allocation_info_t *neighbor_info;
 263
 264                 /* TODO: if op is used on multiple inputs we might not do a
 265                  * continue here */
 266                 if (neighbor == node)
 267                         continue;
 268
 269                 neighbor_info = get_allocation_info(neighbor);
 270                 for (r = 0; r < n_regs; ++r) {
 271                         if (!rbitset_is_set(limited, r))
 272                                 continue;
 273
 274                         neighbor_info->prefs[r] -= penalty;
 275                 }
 276         }
 277 }
 278
 279 /**
 280  * Calculate the preferences of a definition for the current register class.
 281  * If the definition uses a limited set of registers, reduce the preferences
 282  * for the limited register on the node and its neighbors.
 283  *
 284  * @param live_nodes  the set of live nodes at the current node
 285  * @param weight      the weight
 286  * @param node        the current node
 287  */
 288 static void check_defs(const ir_nodeset_t *live_nodes, float weight,
 289                        ir_node *node)
 290 {
 291         const arch_register_req_t *req = arch_get_register_req_out(node);
 292         if (req->type & arch_register_req_type_limited) {
 293                 const unsigned *limited = req->limited;
 294                 float           penalty = weight * DEF_FACTOR;
 295                 give_penalties_for_limits(live_nodes, penalty, limited, node);
 296         }
 297
 298         if (req->type & arch_register_req_type_should_be_same) {
 299                 ir_node           *insn  = skip_Proj(node);
 300                 allocation_info_t *info  = get_allocation_info(node);
 301                 int                arity = get_irn_arity(insn);
 302                 int                i;
 303
 304                 float factor = 1.0f / rbitset_popcount(&req->other_same, arity);
 305                 for (i = 0; i < arity; ++i) {
 306                         ir_node           *op;
 307                         unsigned           r;
 308                         allocation_info_t *op_info;
 309
 310                         if (!rbitset_is_set(&req->other_same, i))
 311                                 continue;
 312
 313                         op = get_irn_n(insn, i);
 314
 315                         /* if we the value at the should_be_same input doesn't die at the
 316                          * node, then it is no use to propagate the constraints (since a
 317                          * copy will emerge anyway) */
 318                         if (ir_nodeset_contains(live_nodes, op))
 319                                 continue;
 320
 321                         op_info = get_allocation_info(op);
 322                         for (r = 0; r < n_regs; ++r) {
 323                                 op_info->prefs[r] += info->prefs[r] * factor;
 324                         }
 325                 }
 326         }
 327 }
 328
 329 /**
 330  * Walker: Runs an a block calculates the preferences for any
 331  * node and every register from the considered register class.
 332  */
 333 static void analyze_block(ir_node *block, void *data)
 334 {
 335         float         weight = (float)get_block_execfreq(execfreqs, block);
 336         ir_nodeset_t  live_nodes;
 337         ir_node      *node;
 338         (void) data;
 339
 340         ir_nodeset_init(&live_nodes);
 341         be_liveness_end_of_block(lv, cls, block, &live_nodes);
 342
 343         sched_foreach_reverse(block, node) {
 344                 allocation_info_t *info;
 345                 int                i;
 346                 int                arity;
 347
 348                 if (is_Phi(node))
 349                         break;
 350
 351                 if (create_preferences) {
 352                         ir_node *value;
 353                         be_foreach_definition(node, cls, value,
 354                                 check_defs(&live_nodes, weight, value);
 355                         );
 356                 }
 357
 358                 /* mark last uses */
 359                 arity = get_irn_arity(node);
 360
 361                 /* the allocation info node currently only uses 1 unsigned value
 362                    to mark last used inputs. So we will fail for a node with more than
 363                    32 inputs. */
 364                 if (arity >= (int) sizeof(info->last_uses) * 8) {
 365                         panic("Node with more than %d inputs not supported yet",
 366                                         (int) sizeof(info->last_uses) * 8);
 367                 }
 368
 369                 info = get_allocation_info(node);
 370                 for (i = 0; i < arity; ++i) {
 371                         ir_node *op = get_irn_n(node, i);
 372                         if (!arch_irn_consider_in_reg_alloc(cls, op))
 373                                 continue;
 374
 375                         /* last usage of a value? */
 376                         if (!ir_nodeset_contains(&live_nodes, op)) {
 377                                 rbitset_set(info->last_uses, i);
 378                         }
 379                 }
 380
 381                 be_liveness_transfer(cls, node, &live_nodes);
 382
 383                 if (create_preferences) {
 384                         /* update weights based on usage constraints */
 385                         for (i = 0; i < arity; ++i) {
 386                                 const arch_register_req_t *req;
 387                                 const unsigned            *limited;
 388                                 ir_node                   *op = get_irn_n(node, i);
 389
 390                                 if (!arch_irn_consider_in_reg_alloc(cls, op))
 391                                         continue;
 392
 393                                 req = arch_get_register_req(node, i);
 394                                 if (!(req->type & arch_register_req_type_limited))
 395                                         continue;
 396
 397                                 limited = req->limited;
 398                                 give_penalties_for_limits(&live_nodes, weight * USE_FACTOR,
 399                                                           limited, op);
 400                         }
 401                 }
 402         }
 403
 404         ir_nodeset_destroy(&live_nodes);
 405 }
 406
 407 static void congruence_def(ir_nodeset_t *live_nodes, const ir_node *node)
 408 {
 409         const arch_register_req_t *req = arch_get_register_req_out(node);
 410
 411         /* should be same constraint? */
 412         if (req->type & arch_register_req_type_should_be_same) {
 413                 const ir_node *insn  = skip_Proj_const(node);
 414                 int      arity = get_irn_arity(insn);
 415                 int      i;
 416                 unsigned node_idx = get_irn_idx(node);
 417                 node_idx          = uf_find(congruence_classes, node_idx);
 418
 419                 for (i = 0; i < arity; ++i) {
 420                         ir_node               *live;
 421                         ir_node               *op;
 422                         int                    op_idx;
 423                         ir_nodeset_iterator_t  iter;
 424                         bool                   interferes = false;
 425
 426                         if (!rbitset_is_set(&req->other_same, i))
 427                                 continue;
 428
 429                         op     = get_irn_n(insn, i);
 430                         op_idx = get_irn_idx(op);
 431                         op_idx = uf_find(congruence_classes, op_idx);
 432
 433                         /* do we interfere with the value */
 434                         foreach_ir_nodeset(live_nodes, live, iter) {
 435                                 int lv_idx = get_irn_idx(live);
 436                                 lv_idx     = uf_find(congruence_classes, lv_idx);
 437                                 if (lv_idx == op_idx) {
 438                                         interferes = true;
 439                                         break;
 440                                 }
 441                         }
 442                         /* don't put in same affinity class if we interfere */
 443                         if (interferes)
 444                                 continue;
 445
 446                         node_idx = uf_union(congruence_classes, node_idx, op_idx);
 447                         DB((dbg, LEVEL_3, "Merge %+F and %+F congruence classes\n",
 448                             node, op));
 449                         /* one should_be_same is enough... */
 450                         break;
 451                 }
 452         }
 453 }
 454
 455 static void create_congruence_class(ir_node *block, void *data)
 456 {
 457         ir_nodeset_t  live_nodes;
 458         ir_node      *node;
 459
 460         (void) data;
 461         ir_nodeset_init(&live_nodes);
 462         be_liveness_end_of_block(lv, cls, block, &live_nodes);
 463
 464         /* check should be same constraints */
 465         sched_foreach_reverse(block, node) {
 466                 ir_node *value;
 467                 if (is_Phi(node))
 468                         break;
 469
 470                 be_foreach_definition(node, cls, value,
 471                         congruence_def(&live_nodes, value);
 472                 );
 473                 be_liveness_transfer(cls, node, &live_nodes);
 474         }
 475
 476         /* check phi congruence classes */
 477         sched_foreach_reverse_from(node, node) {
 478                 int i;
 479                 int arity;
 480                 int node_idx;
 481                 assert(is_Phi(node));
 482
 483                 if (!arch_irn_consider_in_reg_alloc(cls, node))
 484                         continue;
 485
 486                 node_idx = get_irn_idx(node);
 487                 node_idx = uf_find(congruence_classes, node_idx);
 488
 489                 arity = get_irn_arity(node);
 490                 for (i = 0; i < arity; ++i) {
 491                         bool                  interferes = false;
 492                         ir_nodeset_iterator_t iter;
 493                         unsigned              r;
 494                         int                   old_node_idx;
 495                         ir_node              *live;
 496                         ir_node              *phi;
 497                         allocation_info_t    *head_info;
 498                         allocation_info_t    *other_info;
 499                         ir_node              *op     = get_Phi_pred(node, i);
 500                         int                   op_idx = get_irn_idx(op);
 501                         op_idx = uf_find(congruence_classes, op_idx);
 502
 503                         /* do we interfere with the value */
 504                         foreach_ir_nodeset(&live_nodes, live, iter) {
 505                                 int lv_idx = get_irn_idx(live);
 506                                 lv_idx     = uf_find(congruence_classes, lv_idx);
 507                                 if (lv_idx == op_idx) {
 508                                         interferes = true;
 509                                         break;
 510                                 }
 511                         }
 512                         /* don't put in same affinity class if we interfere */
 513                         if (interferes)
 514                                 continue;
 515                         /* any other phi has the same input? */
 516                         sched_foreach(block, phi) {
 517                                 ir_node *oop;
 518                                 int      oop_idx;
 519                                 if (!is_Phi(phi))
 520                                         break;
 521                                 if (!arch_irn_consider_in_reg_alloc(cls, phi))
 522                                         continue;
 523                                 oop = get_Phi_pred(phi, i);
 524                                 if (oop == op)
 525                                         continue;
 526                                 oop_idx = get_irn_idx(oop);
 527                                 oop_idx = uf_find(congruence_classes, oop_idx);
 528                                 if (oop_idx == op_idx) {
 529                                         interferes = true;
 530                                         break;
 531                                 }
 532                         }
 533                         if (interferes)
 534                                 continue;
 535
 536                         /* merge the 2 congruence classes and sum up their preferences */
 537                         old_node_idx = node_idx;
 538                         node_idx = uf_union(congruence_classes, node_idx, op_idx);
 539                         DB((dbg, LEVEL_3, "Merge %+F and %+F congruence classes\n",
 540                             node, op));
 541
 542                         old_node_idx = node_idx == old_node_idx ? op_idx : old_node_idx;
 543                         head_info  = get_allocation_info(get_idx_irn(irg, node_idx));
 544                         other_info = get_allocation_info(get_idx_irn(irg, old_node_idx));
 545                         for (r = 0; r < n_regs; ++r) {
 546                                 head_info->prefs[r] += other_info->prefs[r];
 547                         }
 548                 }
 549         }
 550 }
 551
 552 static void set_congruence_prefs(ir_node *node, void *data)
 553 {
 554         allocation_info_t *info;
 555         allocation_info_t *head_info;
 556         unsigned node_idx = get_irn_idx(node);
 557         unsigned node_set = uf_find(congruence_classes, node_idx);
 558
 559         (void) data;
 560
 561         /* head of congruence class or not in any class */
 562         if (node_set == node_idx)
 563                 return;
 564
 565         if (!arch_irn_consider_in_reg_alloc(cls, node))
 566                 return;
 567
 568         head_info = get_allocation_info(get_idx_irn(irg, node_set));
 569         info      = get_allocation_info(node);
 570
 571         memcpy(info->prefs, head_info->prefs, n_regs * sizeof(info->prefs[0]));
 572 }
 573
 574 static void combine_congruence_classes(void)
 575 {
 576         size_t n = get_irg_last_idx(irg);
 577         congruence_classes = XMALLOCN(int, n);
 578         uf_init(congruence_classes, n);
 579
 580         /* create congruence classes */
 581         irg_block_walk_graph(irg, create_congruence_class, NULL, NULL);
 582         /* merge preferences */
 583         irg_walk_graph(irg, set_congruence_prefs, NULL, NULL);
 584         free(congruence_classes);
 585 }
 586
 587
 588
 589 /**
 590  * Assign register reg to the given node.
 591  *
 592  * @param node  the node
 593  * @param reg   the register
 594  */
 595 static void use_reg(ir_node *node, const arch_register_t *reg)
 596 {
 597         unsigned r = arch_register_get_index(reg);
 598         assignments[r] = node;
 599         arch_set_irn_register(node, reg);
 600 }
 601
 602 static void free_reg_of_value(ir_node *node)
 603 {
 604         const arch_register_t *reg;
 605         unsigned               r;
 606
 607         if (!arch_irn_consider_in_reg_alloc(cls, node))
 608                 return;
 609
 610         reg        = arch_get_irn_register(node);
 611         r          = arch_register_get_index(reg);
 612         /* assignment->value may be NULL if a value is used at 2 inputs
 613            so it gets freed twice. */
 614         assert(assignments[r] == node || assignments[r] == NULL);
 615         assignments[r] = NULL;
 616 }
 617
 618 /**
 619  * Compare two register preferences in decreasing order.
 620  */
 621 static int compare_reg_pref(const void *e1, const void *e2)
 622 {
 623         const reg_pref_t *rp1 = (const reg_pref_t*) e1;
 624         const reg_pref_t *rp2 = (const reg_pref_t*) e2;
 625         if (rp1->pref < rp2->pref)
 626                 return 1;
 627         if (rp1->pref > rp2->pref)
 628                 return -1;
 629         return 0;
 630 }
 631
 632 static void fill_sort_candidates(reg_pref_t *regprefs,
 633                                  const allocation_info_t *info)
 634 {
 635         unsigned r;
 636
 637         for (r = 0; r < n_regs; ++r) {
 638                 float pref = info->prefs[r];
 639                 regprefs[r].num  = r;
 640                 regprefs[r].pref = pref;
 641         }
 642         /* TODO: use a stable sort here to avoid unnecessary register jumping */
 643         qsort(regprefs, n_regs, sizeof(regprefs[0]), compare_reg_pref);
 644 }
 645
 646 static bool try_optimistic_split(ir_node *to_split, ir_node *before,
 647                                  float pref, float pref_delta,
 648                                  unsigned *forbidden_regs, int recursion)
 649 {
 650         const arch_register_t *from_reg;
 651         const arch_register_t *reg;
 652         ir_node               *original_insn;
 653         ir_node               *block;
 654         ir_node               *copy;
 655         unsigned               r;
 656         unsigned               from_r;
 657         unsigned               i;
 658         allocation_info_t     *info = get_allocation_info(to_split);
 659         reg_pref_t            *prefs;
 660         float                  delta;
 661         float                  split_threshold;
 662
 663         (void) pref;
 664
 665         /* stupid hack: don't optimisticallt split don't spill nodes...
 666          * (so we don't split away the values produced because of
 667          *  must_be_different constraints) */
 668         original_insn = skip_Proj(info->original_value);
 669         if (arch_irn_get_flags(original_insn) & arch_irn_flags_dont_spill)
 670                 return false;
 671
 672         from_reg        = arch_get_irn_register(to_split);
 673         from_r          = arch_register_get_index(from_reg);
 674         block           = get_nodes_block(before);
 675         split_threshold = (float)get_block_execfreq(execfreqs, block) * SPLIT_DELTA;
 676
 677         if (pref_delta < split_threshold*0.5)
 678                 return false;
 679
 680         /* find the best free position where we could move to */
 681         prefs = ALLOCAN(reg_pref_t, n_regs);
 682         fill_sort_candidates(prefs, info);
 683         for (i = 0; i < n_regs; ++i) {
 684                 float apref;
 685                 float apref_delta;
 686                 bool  res;
 687                 bool  old_source_state;
 688
 689                 /* we need a normal register which is not an output register
 690                    an different from the current register of to_split */
 691                 r = prefs[i].num;
 692                 if (!rbitset_is_set(normal_regs, r))
 693                         continue;
 694                 if (rbitset_is_set(forbidden_regs, r))
 695                         continue;
 696                 if (r == from_r)
 697                         continue;
 698
 699                 /* is the split worth it? */
 700                 delta = pref_delta + prefs[i].pref;
 701                 if (delta < split_threshold) {
 702                         DB((dbg, LEVEL_3, "Not doing optimistical split of %+F (depth %d), win %f too low\n",
 703                                 to_split, recursion, delta));
 704                         return false;
 705                 }
 706
 707                 /* if the register is free then we can do the split */
 708                 if (assignments[r] == NULL)
 709                         break;
 710
 711                 /* otherwise we might try recursively calling optimistic_split */
 712                 if (recursion+1 > MAX_OPTIMISTIC_SPLIT_RECURSION)
 713                         continue;
 714
 715                 apref        = prefs[i].pref;
 716                 apref_delta  = i+1 < n_regs ? apref - prefs[i+1].pref : 0;
 717                 apref_delta += pref_delta - split_threshold;
 718
 719                 /* our source register isn't a useful destination for recursive
 720                    splits */
 721                 old_source_state = rbitset_is_set(forbidden_regs, from_r);
 722                 rbitset_set(forbidden_regs, from_r);
 723                 /* try recursive split */
 724                 res = try_optimistic_split(assignments[r], before, apref,
 725                                            apref_delta, forbidden_regs, recursion+1);
 726                 /* restore our destination */
 727                 if (old_source_state) {
 728                         rbitset_set(forbidden_regs, from_r);
 729                 } else {
 730                         rbitset_clear(forbidden_regs, from_r);
 731                 }
 732
 733                 if (res)
 734                         break;
 735         }
 736         if (i >= n_regs)
 737                 return false;
 738
 739         reg  = arch_register_for_index(cls, r);
 740         copy = be_new_Copy(cls, block, to_split);
 741         mark_as_copy_of(copy, to_split);
 742         /* hacky, but correct here */
 743         if (assignments[arch_register_get_index(from_reg)] == to_split)
 744                 free_reg_of_value(to_split);
 745         use_reg(copy, reg);
 746         sched_add_before(before, copy);
 747
 748         DB((dbg, LEVEL_3,
 749             "Optimistic live-range split %+F move %+F(%s) -> %s before %+F (win %f, depth %d)\n",
 750             copy, to_split, from_reg->name, reg->name, before, delta, recursion));
 751         return true;
 752 }
 753
 754 /**
 755  * Determine and assign a register for node @p node
 756  */
 757 static void assign_reg(const ir_node *block, ir_node *node,
 758                        unsigned *forbidden_regs)
 759 {
 760         const arch_register_t     *reg;
 761         allocation_info_t         *info;
 762         const arch_register_req_t *req;
 763         reg_pref_t                *reg_prefs;
 764         ir_node                   *in_node;
 765         unsigned                   i;
 766         const unsigned            *allowed_regs;
 767         unsigned                   r;
 768
 769         assert(!is_Phi(node));
 770         assert(arch_irn_consider_in_reg_alloc(cls, node));
 771
 772         /* preassigned register? */
 773         reg = arch_get_irn_register(node);
 774         if (reg != NULL) {
 775                 DB((dbg, LEVEL_2, "Preassignment %+F -> %s\n", node, reg->name));
 776                 use_reg(node, reg);
 777                 return;
 778         }
 779
 780         /* give should_be_same boni */
 781         info = get_allocation_info(node);
 782         req  = arch_get_register_req_out(node);
 783
 784         in_node = skip_Proj(node);
 785         if (req->type & arch_register_req_type_should_be_same) {
 786                 float weight = (float)get_block_execfreq(execfreqs, block);
 787                 int   arity  = get_irn_arity(in_node);
 788                 int   i;
 789
 790                 assert(arity <= (int) sizeof(req->other_same) * 8);
 791                 for (i = 0; i < arity; ++i) {
 792                         ir_node               *in;
 793                         const arch_register_t *reg;
 794                         unsigned               r;
 795                         if (!rbitset_is_set(&req->other_same, i))
 796                                 continue;
 797
 798                         in  = get_irn_n(in_node, i);
 799                         reg = arch_get_irn_register(in);
 800                         assert(reg != NULL);
 801                         r = arch_register_get_index(reg);
 802
 803                         /* if the value didn't die here then we should not propagate the
 804                          * should_be_same info */
 805                         if (assignments[r] == in)
 806                                 continue;
 807
 808                         info->prefs[r] += weight * AFF_SHOULD_BE_SAME;
 809                 }
 810         }
 811
 812         /* create list of register candidates and sort by their preference */
 813         DB((dbg, LEVEL_2, "Candidates for %+F:", node));
 814         reg_prefs = alloca(n_regs * sizeof(reg_prefs[0]));
 815         fill_sort_candidates(reg_prefs, info);
 816         for (i = 0; i < n_regs; ++i) {
 817                 unsigned num = reg_prefs[i].num;
 818                 const arch_register_t *reg;
 819
 820                 if (!rbitset_is_set(normal_regs, num))
 821                         continue;
 822
 823                 reg = arch_register_for_index(cls, num);
 824                 DB((dbg, LEVEL_2, " %s(%f)", reg->name, reg_prefs[i].pref));
 825         }
 826         DB((dbg, LEVEL_2, "\n"));
 827
 828         allowed_regs = normal_regs;
 829         if (req->type & arch_register_req_type_limited) {
 830                 allowed_regs = req->limited;
 831         }
 832
 833         for (i = 0; i < n_regs; ++i) {
 834                 float   pref, delta;
 835                 ir_node *before;
 836                 bool    res;
 837
 838                 r = reg_prefs[i].num;
 839                 if (!rbitset_is_set(allowed_regs, r))
 840                         continue;
 841                 if (assignments[r] == NULL)
 842                         break;
 843                 pref   = reg_prefs[i].pref;
 844                 delta  = i+1 < n_regs ? pref - reg_prefs[i+1].pref : 0;
 845                 before = skip_Proj(node);
 846                 res    = try_optimistic_split(assignments[r], before,
 847                                               pref, delta, forbidden_regs, 0);
 848                 if (res)
 849                         break;
 850         }
 851         if (i >= n_regs) {
 852                 /* the common reason to hit this panic is when 1 of your nodes is not
 853                  * register pressure faithful */
 854                 panic("No register left for %+F\n", node);
 855         }
 856
 857         reg = arch_register_for_index(cls, r);
 858         DB((dbg, LEVEL_2, "Assign %+F -> %s\n", node, reg->name));
 859         use_reg(node, reg);
 860 }
 861
 862 /**
 863  * Add an permutation in front of a node and change the assignments
 864  * due to this permutation.
 865  *
 866  * To understand this imagine a permutation like this:
 867  *
 868  * 1 -> 2
 869  * 2 -> 3
 870  * 3 -> 1, 5
 871  * 4 -> 6
 872  * 5
 873  * 6
 874  * 7 -> 7
 875  *
 876  * First we count how many destinations a single value has. At the same time
 877  * we can be sure that each destination register has at most 1 source register
 878  * (it can have 0 which means we don't care what value is in it).
 879  * We ignore all fullfilled permuations (like 7->7)
 880  * In a first pass we create as much copy instructions as possible as they
 881  * are generally cheaper than exchanges. We do this by counting into how many
 882  * destinations a register has to be copied (in the example it's 2 for register
 883  * 3, or 1 for the registers 1,2,4 and 7).
 884  * We can then create a copy into every destination register when the usecount
 885  * of that register is 0 (= noone else needs the value in the register).
 886  *
 887  * After this step we should have cycles left. We implement a cyclic permutation
 888  * of n registers with n-1 transpositions.
 889  *
 890  * @param live_nodes   the set of live nodes, updated due to live range split
 891  * @param before       the node before we add the permutation
 892  * @param permutation  the permutation array indices are the destination
 893  *                     registers, the values in the array are the source
 894  *                     registers.
 895  */
 896 static void permute_values(ir_nodeset_t *live_nodes, ir_node *before,
 897                              unsigned *permutation)
 898 {
 899         unsigned  *n_used = ALLOCANZ(unsigned, n_regs);
 900         ir_node   *block;
 901         unsigned   r;
 902
 903         /* determine how often each source register needs to be read */
 904         for (r = 0; r < n_regs; ++r) {
 905                 unsigned  old_reg = permutation[r];
 906                 ir_node  *value;
 907
 908                 value = assignments[old_reg];
 909                 if (value == NULL) {
 910                         /* nothing to do here, reg is not live. Mark it as fixpoint
 911                          * so we ignore it in the next steps */
 912                         permutation[r] = r;
 913                         continue;
 914                 }
 915
 916                 ++n_used[old_reg];
 917         }
 918
 919         block = get_nodes_block(before);
 920
 921         /* step1: create copies where immediately possible */
 922         for (r = 0; r < n_regs; /* empty */) {
 923                 ir_node *copy;
 924                 ir_node *src;
 925                 const arch_register_t *reg;
 926                 unsigned               old_r = permutation[r];
 927
 928                 /* - no need to do anything for fixed points.
 929                    - we can't copy if the value in the dest reg is still needed */
 930                 if (old_r == r || n_used[r] > 0) {
 931                         ++r;
 932                         continue;
 933                 }
 934
 935                 /* create a copy */
 936                 src  = assignments[old_r];
 937                 copy = be_new_Copy(cls, block, src);
 938                 sched_add_before(before, copy);
 939                 reg = arch_register_for_index(cls, r);
 940                 DB((dbg, LEVEL_2, "Copy %+F (from %+F, before %+F) -> %s\n",
 941                     copy, src, before, reg->name));
 942                 mark_as_copy_of(copy, src);
 943                 use_reg(copy, reg);
 944
 945                 if (live_nodes != NULL) {
 946                         ir_nodeset_insert(live_nodes, copy);
 947                 }
 948
 949                 /* old register has 1 user less, permutation is resolved */
 950                 assert(arch_register_get_index(arch_get_irn_register(src)) == old_r);
 951                 permutation[r] = r;
 952
 953                 assert(n_used[old_r] > 0);
 954                 --n_used[old_r];
 955                 if (n_used[old_r] == 0) {
 956                         if (live_nodes != NULL) {
 957                                 ir_nodeset_remove(live_nodes, src);
 958                         }
 959                         free_reg_of_value(src);
 960                 }
 961
 962                 /* advance or jump back (if this copy enabled another copy) */
 963                 if (old_r < r && n_used[old_r] == 0) {
 964                         r = old_r;
 965                 } else {
 966                         ++r;
 967                 }
 968         }
 969
 970         /* at this point we only have "cycles" left which we have to resolve with
 971          * perm instructions
 972          * TODO: if we have free registers left, then we should really use copy
 973          * instructions for any cycle longer than 2 registers...
 974          * (this is probably architecture dependent, there might be archs where
 975          *  copies are preferable even for 2-cycles) */
 976
 977         /* create perms with the rest */
 978         for (r = 0; r < n_regs; /* empty */) {
 979                 const arch_register_t *reg;
 980                 unsigned  old_r = permutation[r];
 981                 unsigned  r2;
 982                 ir_node  *in[2];
 983                 ir_node  *perm;
 984                 ir_node  *proj0;
 985                 ir_node  *proj1;
 986
 987                 if (old_r == r) {
 988                         ++r;
 989                         continue;
 990                 }
 991
 992                 /* we shouldn't have copies from 1 value to multiple destinations left*/
 993                 assert(n_used[old_r] == 1);
 994
 995                 /* exchange old_r and r2; after that old_r is a fixed point */
 996                 r2 = permutation[old_r];
 997
 998                 in[0] = assignments[r2];
 999                 in[1] = assignments[old_r];
1000                 perm = be_new_Perm(cls, block, 2, in);
1001                 sched_add_before(before, perm);
1002                 DB((dbg, LEVEL_2, "Perm %+F (perm %+F,%+F, before %+F)\n",
1003                     perm, in[0], in[1], before));
1004
1005                 proj0 = new_r_Proj(perm, get_irn_mode(in[0]), 0);
1006                 mark_as_copy_of(proj0, in[0]);
1007                 reg = arch_register_for_index(cls, old_r);
1008                 use_reg(proj0, reg);
1009
1010                 proj1 = new_r_Proj(perm, get_irn_mode(in[1]), 1);
1011                 mark_as_copy_of(proj1, in[1]);
1012                 reg = arch_register_for_index(cls, r2);
1013                 use_reg(proj1, reg);
1014
1015                 /* 1 value is now in the correct register */
1016                 permutation[old_r] = old_r;
1017                 /* the source of r changed to r2 */
1018                 permutation[r] = r2;
1019
1020                 /* if we have reached a fixpoint update data structures */
1021                 if (live_nodes != NULL) {
1022                         ir_nodeset_remove(live_nodes, in[0]);
1023                         ir_nodeset_remove(live_nodes, in[1]);
1024                         ir_nodeset_remove(live_nodes, proj0);
1025                         ir_nodeset_insert(live_nodes, proj1);
1026                 }
1027         }
1028
1029 #ifdef DEBUG_libfirm
1030         /* now we should only have fixpoints left */
1031         for (r = 0; r < n_regs; ++r) {
1032                 assert(permutation[r] == r);
1033         }
1034 #endif
1035 }
1036
1037 /**
1038  * Free regs for values last used.
1039  *
1040  * @param live_nodes   set of live nodes, will be updated
1041  * @param node         the node to consider
1042  */
1043 static void free_last_uses(ir_nodeset_t *live_nodes, ir_node *node)
1044 {
1045         allocation_info_t     *info      = get_allocation_info(node);
1046         const unsigned        *last_uses = info->last_uses;
1047         int                    arity     = get_irn_arity(node);
1048         int                    i;
1049
1050         for (i = 0; i < arity; ++i) {
1051                 ir_node *op;
1052
1053                 /* check if one operand is the last use */
1054                 if (!rbitset_is_set(last_uses, i))
1055                         continue;
1056
1057                 op = get_irn_n(node, i);
1058                 free_reg_of_value(op);
1059                 ir_nodeset_remove(live_nodes, op);
1060         }
1061 }
1062
1063 /**
1064  * change inputs of a node to the current value (copies/perms)
1065  */
1066 static void rewire_inputs(ir_node *node)
1067 {
1068         int i;
1069         int arity = get_irn_arity(node);
1070
1071         for (i = 0; i < arity; ++i) {
1072                 ir_node           *op = get_irn_n(node, i);
1073                 allocation_info_t *info = try_get_allocation_info(op);
1074
1075                 if (info == NULL)
1076                         continue;
1077
1078                 info = get_allocation_info(info->original_value);
1079                 if (info->current_value != op) {
1080                         set_irn_n(node, i, info->current_value);
1081                 }
1082         }
1083 }
1084
1085 /**
1086  * Create a bitset of registers occupied with value living through an
1087  * instruction
1088  */
1089 static void determine_live_through_regs(unsigned *bitset, ir_node *node)
1090 {
1091         const allocation_info_t *info = get_allocation_info(node);
1092         unsigned r;
1093         int i;
1094         int arity;
1095
1096         /* mark all used registers as potentially live-through */
1097         for (r = 0; r < n_regs; ++r) {
1098                 if (assignments[r] == NULL)
1099                         continue;
1100                 if (!rbitset_is_set(normal_regs, r))
1101                         continue;
1102
1103                 rbitset_set(bitset, r);
1104         }
1105
1106         /* remove registers of value dying at the instruction */
1107         arity = get_irn_arity(node);
1108         for (i = 0; i < arity; ++i) {
1109                 ir_node               *op;
1110                 const arch_register_t *reg;
1111
1112                 if (!rbitset_is_set(info->last_uses, i))
1113                         continue;
1114
1115                 op  = get_irn_n(node, i);
1116                 reg = arch_get_irn_register(op);
1117                 rbitset_clear(bitset, arch_register_get_index(reg));
1118         }
1119 }
1120
1121 /**
1122  * Enforce constraints at a node by live range splits.
1123  *
1124  * @param  live_nodes  the set of live nodes, might be changed
1125  * @param  node        the current node
1126  */
1127 static void enforce_constraints(ir_nodeset_t *live_nodes, ir_node *node,
1128                                 unsigned *forbidden_regs)
1129 {
1130         int arity = get_irn_arity(node);
1131         int i, res;
1132         hungarian_problem_t *bp;
1133         unsigned l, r;
1134         unsigned *assignment;
1135         ir_node  *value;
1136
1137         /* construct a list of register occupied by live-through values */
1138         unsigned *live_through_regs = NULL;
1139
1140         /* see if any use constraints are not met */
1141         bool good = true;
1142         for (i = 0; i < arity; ++i) {
1143                 ir_node                   *op = get_irn_n(node, i);
1144                 const arch_register_t     *reg;
1145                 const arch_register_req_t *req;
1146                 const unsigned            *limited;
1147                 unsigned                  r;
1148
1149                 if (!arch_irn_consider_in_reg_alloc(cls, op))
1150                         continue;
1151
1152                 /* are there any limitations for the i'th operand? */
1153                 req = arch_get_register_req(node, i);
1154                 if (!(req->type & arch_register_req_type_limited))
1155                         continue;
1156
1157                 limited = req->limited;
1158                 reg     = arch_get_irn_register(op);
1159                 r       = arch_register_get_index(reg);
1160                 if (!rbitset_is_set(limited, r)) {
1161                         /* found an assignment outside the limited set */
1162                         good = false;
1163                         break;
1164                 }
1165         }
1166
1167         /* is any of the live-throughs using a constrained output register? */
1168         be_foreach_definition(node, cls, value,
1169                 if (! (req_->type & arch_register_req_type_limited))
1170                         continue;
1171                 if (live_through_regs == NULL) {
1172                         rbitset_alloca(live_through_regs, n_regs);
1173                         determine_live_through_regs(live_through_regs, node);
1174                 }
1175                 rbitset_or(forbidden_regs, req_->limited, n_regs);
1176                 if (rbitsets_have_common(req_->limited, live_through_regs, n_regs))
1177                         good = false;
1178         );
1179
1180         if (good)
1181                 return;
1182
1183         /* create these arrays if we haven't yet */
1184         if (live_through_regs == NULL) {
1185                 rbitset_alloca(live_through_regs, n_regs);
1186         }
1187
1188         /* at this point we have to construct a bipartite matching problem to see
1189          * which values should go to which registers
1190          * Note: We're building the matrix in "reverse" - source registers are
1191          *       right, destinations left because this will produce the solution
1192          *       in the format required for permute_values.
1193          */
1194         bp = hungarian_new(n_regs, n_regs, HUNGARIAN_MATCH_PERFECT);
1195
1196         /* add all combinations, then remove not allowed ones */
1197         for (l = 0; l < n_regs; ++l) {
1198                 if (!rbitset_is_set(normal_regs, l)) {
1199                         hungarian_add(bp, l, l, 1);
1200                         continue;
1201                 }
1202
1203                 for (r = 0; r < n_regs; ++r) {
1204                         if (!rbitset_is_set(normal_regs, r))
1205                                 continue;
1206                         /* livethrough values may not use constrainted output registers */
1207                         if (rbitset_is_set(live_through_regs, l)
1208                                         && rbitset_is_set(forbidden_regs, r))
1209                                 continue;
1210
1211                         hungarian_add(bp, r, l, l == r ? 9 : 8);
1212                 }
1213         }
1214
1215         for (i = 0; i < arity; ++i) {
1216                 ir_node                   *op = get_irn_n(node, i);
1217                 const arch_register_t     *reg;
1218                 const arch_register_req_t *req;
1219                 const unsigned            *limited;
1220                 unsigned                   current_reg;
1221
1222                 if (!arch_irn_consider_in_reg_alloc(cls, op))
1223                         continue;
1224
1225                 req = arch_get_register_req(node, i);
1226                 if (!(req->type & arch_register_req_type_limited))
1227                         continue;
1228
1229                 limited     = req->limited;
1230                 reg         = arch_get_irn_register(op);
1231                 current_reg = arch_register_get_index(reg);
1232                 for (r = 0; r < n_regs; ++r) {
1233                         if (rbitset_is_set(limited, r))
1234                                 continue;
1235                         hungarian_remove(bp, r, current_reg);
1236                 }
1237         }
1238
1239         //hungarian_print_cost_matrix(bp, 1);
1240         hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
1241
1242         assignment = ALLOCAN(unsigned, n_regs);
1243         res = hungarian_solve(bp, assignment, NULL, 0);
1244         assert(res == 0);
1245
1246 #if 0
1247         fprintf(stderr, "Swap result:");
1248         for (i = 0; i < (int) n_regs; ++i) {
1249                 fprintf(stderr, " %d", assignment[i]);
1250         }
1251         fprintf(stderr, "\n");
1252 #endif
1253
1254         hungarian_free(bp);
1255
1256         permute_values(live_nodes, node, assignment);
1257 }
1258
1259 /** test wether a node @p n is a copy of the value of node @p of */
1260 static bool is_copy_of(ir_node *value, ir_node *test_value)
1261 {
1262         allocation_info_t *test_info;
1263         allocation_info_t *info;
1264
1265         if (value == test_value)
1266                 return true;
1267
1268         info      = get_allocation_info(value);
1269         test_info = get_allocation_info(test_value);
1270         return test_info->original_value == info->original_value;
1271 }
1272
1273 /**
1274  * find a value in the end-assignment of a basic block
1275  * @returns the index into the assignment array if found
1276  *          -1 if not found
1277  */
1278 static int find_value_in_block_info(block_info_t *info, ir_node *value)
1279 {
1280         unsigned   r;
1281         ir_node  **assignments = info->assignments;
1282         for (r = 0; r < n_regs; ++r) {
1283                 ir_node *a_value = assignments[r];
1284
1285                 if (a_value == NULL)
1286                         continue;
1287                 if (is_copy_of(a_value, value))
1288                         return (int) r;
1289         }
1290
1291         return -1;
1292 }
1293
1294 /**
1295  * Create the necessary permutations at the end of a basic block to fullfill
1296  * the register assignment for phi-nodes in the next block
1297  */
1298 static void add_phi_permutations(ir_node *block, int p)
1299 {
1300         unsigned   r;
1301         unsigned  *permutation;
1302         ir_node  **old_assignments;
1303         bool       need_permutation;
1304         ir_node   *node;
1305         ir_node   *pred = get_Block_cfgpred_block(block, p);
1306
1307         block_info_t *pred_info = get_block_info(pred);
1308
1309         /* predecessor not processed yet? nothing to do */
1310         if (!pred_info->processed)
1311                 return;
1312
1313         permutation = ALLOCAN(unsigned, n_regs);
1314         for (r = 0; r < n_regs; ++r) {
1315                 permutation[r] = r;
1316         }
1317
1318         /* check phi nodes */
1319         need_permutation = false;
1320         node = sched_first(block);
1321         for ( ; is_Phi(node); node = sched_next(node)) {
1322                 const arch_register_t *reg;
1323                 int                    regn;
1324                 int                    a;
1325                 ir_node               *op;
1326
1327                 if (!arch_irn_consider_in_reg_alloc(cls, node))
1328                         continue;
1329
1330                 op = get_Phi_pred(node, p);
1331                 if (!arch_irn_consider_in_reg_alloc(cls, op))
1332                         continue;
1333
1334                 a = find_value_in_block_info(pred_info, op);
1335                 assert(a >= 0);
1336
1337                 reg  = arch_get_irn_register(node);
1338                 regn = arch_register_get_index(reg);
1339                 if (regn != a) {
1340                         permutation[regn] = a;
1341                         need_permutation  = true;
1342                 }
1343         }
1344
1345         if (need_permutation) {
1346                 /* permute values at end of predecessor */
1347                 old_assignments = assignments;
1348                 assignments     = pred_info->assignments;
1349                 permute_values(NULL, be_get_end_of_block_insertion_point(pred),
1350                                                  permutation);
1351                 assignments = old_assignments;
1352         }
1353
1354         /* change phi nodes to use the copied values */
1355         node = sched_first(block);
1356         for ( ; is_Phi(node); node = sched_next(node)) {
1357                 int      a;
1358                 ir_node *op;
1359
1360                 if (!arch_irn_consider_in_reg_alloc(cls, node))
1361                         continue;
1362
1363                 op = get_Phi_pred(node, p);
1364                 /* no need to do anything for Unknown inputs */
1365                 if (!arch_irn_consider_in_reg_alloc(cls, op))
1366                         continue;
1367
1368                 /* we have permuted all values into the correct registers so we can
1369                    simply query which value occupies the phis register in the
1370                    predecessor */
1371                 a  = arch_register_get_index(arch_get_irn_register(node));
1372                 op = pred_info->assignments[a];
1373                 set_Phi_pred(node, p, op);
1374         }
1375 }
1376
1377 /**
1378  * Set preferences for a phis register based on the registers used on the
1379  * phi inputs.
1380  */
1381 static void adapt_phi_prefs(ir_node *phi)
1382 {
1383         int i;
1384         int arity = get_irn_arity(phi);
1385         ir_node           *block = get_nodes_block(phi);
1386         allocation_info_t *info  = get_allocation_info(phi);
1387
1388         for (i = 0; i < arity; ++i) {
1389                 ir_node               *op  = get_irn_n(phi, i);
1390                 const arch_register_t *reg = arch_get_irn_register(op);
1391                 ir_node               *pred_block;
1392                 block_info_t          *pred_block_info;
1393                 float                  weight;
1394                 unsigned               r;
1395
1396                 if (reg == NULL)
1397                         continue;
1398                 /* we only give the bonus if the predecessor already has registers
1399                  * assigned, otherwise we only see a dummy value
1400                  * and any conclusions about its register are useless */
1401                 pred_block = get_Block_cfgpred_block(block, i);
1402                 pred_block_info = get_block_info(pred_block);
1403                 if (!pred_block_info->processed)
1404                         continue;
1405
1406                 /* give bonus for already assigned register */
1407                 weight = (float)get_block_execfreq(execfreqs, pred_block);
1408                 r      = arch_register_get_index(reg);
1409                 info->prefs[r] += weight * AFF_PHI;
1410         }
1411 }
1412
1413 /**
1414  * After a phi has been assigned a register propagate preference inputs
1415  * to the phi inputs.
1416  */
1417 static void propagate_phi_register(ir_node *phi, unsigned assigned_r)
1418 {
1419         int      i;
1420         ir_node *block = get_nodes_block(phi);
1421         int      arity = get_irn_arity(phi);
1422
1423         for (i = 0; i < arity; ++i) {
1424                 ir_node           *op         = get_Phi_pred(phi, i);
1425                 allocation_info_t *info       = get_allocation_info(op);
1426                 ir_node           *pred_block = get_Block_cfgpred_block(block, i);
1427                 unsigned           r;
1428                 float              weight
1429                         = (float)get_block_execfreq(execfreqs, pred_block) * AFF_PHI;
1430
1431                 if (info->prefs[assigned_r] >= weight)
1432                         continue;
1433
1434                 /* promote the prefered register */
1435                 for (r = 0; r < n_regs; ++r) {
1436                         if (info->prefs[r] > -weight) {
1437                                 info->prefs[r] = -weight;
1438                         }
1439                 }
1440                 info->prefs[assigned_r] = weight;
1441
1442                 if (is_Phi(op))
1443                         propagate_phi_register(op, assigned_r);
1444         }
1445 }
1446
1447 static void assign_phi_registers(ir_node *block)
1448 {
1449         int                  n_phis = 0;
1450         int                  n;
1451         int                  res;
1452         unsigned            *assignment;
1453         ir_node             *node;
1454         hungarian_problem_t *bp;
1455
1456         /* count phi nodes */
1457         sched_foreach(block, node) {
1458                 if (!is_Phi(node))
1459                         break;
1460                 if (!arch_irn_consider_in_reg_alloc(cls, node))
1461                         continue;
1462                 ++n_phis;
1463         }
1464
1465         if (n_phis == 0)
1466                 return;
1467
1468         /* build a bipartite matching problem for all phi nodes */
1469         bp = hungarian_new(n_phis, n_regs, HUNGARIAN_MATCH_PERFECT);
1470         n  = 0;
1471         sched_foreach(block, node) {
1472                 unsigned r;
1473
1474                 allocation_info_t *info;
1475                 if (!is_Phi(node))
1476                         break;
1477                 if (!arch_irn_consider_in_reg_alloc(cls, node))
1478                         continue;
1479
1480                 /* give boni for predecessor colorings */
1481                 adapt_phi_prefs(node);
1482                 /* add stuff to bipartite problem */
1483                 info = get_allocation_info(node);
1484                 DB((dbg, LEVEL_3, "Prefs for %+F: ", node));
1485                 for (r = 0; r < n_regs; ++r) {
1486                         float costs;
1487
1488                         if (!rbitset_is_set(normal_regs, r))
1489                                 continue;
1490
1491                         costs = info->prefs[r];
1492                         costs = costs < 0 ? -logf(-costs+1) : logf(costs+1);
1493                         costs *= 100;
1494                         costs += 10000;
1495                         hungarian_add(bp, n, r, (int)costs);
1496                         DB((dbg, LEVEL_3, " %s(%f)", arch_register_for_index(cls, r)->name,
1497                                                 info->prefs[r]));
1498                 }
1499                 DB((dbg, LEVEL_3, "\n"));
1500                 ++n;
1501         }
1502
1503         //hungarian_print_cost_matrix(bp, 7);
1504         hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
1505
1506         assignment = ALLOCAN(unsigned, n_regs);
1507         res        = hungarian_solve(bp, assignment, NULL, 0);
1508         assert(res == 0);
1509
1510         /* apply results */
1511         n = 0;
1512         sched_foreach(block, node) {
1513                 unsigned               r;
1514                 const arch_register_t *reg;
1515
1516                 if (!is_Phi(node))
1517                         break;
1518                 if (!arch_irn_consider_in_reg_alloc(cls, node))
1519                         continue;
1520
1521                 r = assignment[n++];
1522                 assert(rbitset_is_set(normal_regs, r));
1523                 reg = arch_register_for_index(cls, r);
1524                 DB((dbg, LEVEL_2, "Assign %+F -> %s\n", node, reg->name));
1525                 use_reg(node, reg);
1526
1527                 /* adapt preferences for phi inputs */
1528                 if (propagate_phi_registers)
1529                         propagate_phi_register(node, r);
1530         }
1531 }
1532
1533 /**
1534  * Walker: assign registers to all nodes of a block that
1535  * need registers from the currently considered register class.
1536  */
1537 static void allocate_coalesce_block(ir_node *block, void *data)
1538 {
1539         int                    i;
1540         ir_nodeset_t           live_nodes;
1541         ir_node               *node;
1542         int                    n_preds;
1543         block_info_t          *block_info;
1544         block_info_t         **pred_block_infos;
1545         ir_node              **phi_ins;
1546         unsigned              *forbidden_regs; /**< collects registers which must
1547                                                 not be used for optimistic splits */
1548
1549         (void) data;
1550         DB((dbg, LEVEL_2, "* Block %+F\n", block));
1551
1552         /* clear assignments */
1553         block_info  = get_block_info(block);
1554         assignments = block_info->assignments;
1555
1556         ir_nodeset_init(&live_nodes);
1557
1558         /* gather regalloc infos of predecessor blocks */
1559         n_preds             = get_Block_n_cfgpreds(block);
1560         pred_block_infos    = ALLOCAN(block_info_t*, n_preds);
1561         for (i = 0; i < n_preds; ++i) {
1562                 ir_node      *pred      = get_Block_cfgpred_block(block, i);
1563                 block_info_t *pred_info = get_block_info(pred);
1564                 pred_block_infos[i]     = pred_info;
1565         }
1566
1567         phi_ins = ALLOCAN(ir_node*, n_preds);
1568
1569         /* collect live-in nodes and preassigned values */
1570         be_lv_foreach(lv, block, be_lv_state_in, i) {
1571                 const arch_register_t *reg;
1572                 int                    p;
1573                 bool                   need_phi = false;
1574
1575                 node = be_lv_get_irn(lv, block, i);
1576                 if (!arch_irn_consider_in_reg_alloc(cls, node))
1577                         continue;
1578
1579                 /* check all predecessors for this value, if it is not everywhere the
1580                    same or unknown then we have to construct a phi
1581                    (we collect the potential phi inputs here) */
1582                 for (p = 0; p < n_preds; ++p) {
1583                         block_info_t *pred_info = pred_block_infos[p];
1584
1585                         if (!pred_info->processed) {
1586                                 /* use node for now, it will get fixed later */
1587                                 phi_ins[p] = node;
1588                                 need_phi   = true;
1589                         } else {
1590                                 int a = find_value_in_block_info(pred_info, node);
1591
1592                                 /* must live out of predecessor */
1593                                 assert(a >= 0);
1594                                 phi_ins[p] = pred_info->assignments[a];
1595                                 /* different value from last time? then we need a phi */
1596                                 if (p > 0 && phi_ins[p-1] != phi_ins[p]) {
1597                                         need_phi = true;
1598                                 }
1599                         }
1600                 }
1601
1602                 if (need_phi) {
1603                         ir_mode                   *mode = get_irn_mode(node);
1604                         const arch_register_req_t *req  = get_default_req_current_cls();
1605                         ir_node                   *phi;
1606
1607                         phi = new_r_Phi(block, n_preds, phi_ins, mode);
1608                         be_set_phi_reg_req(phi, req);
1609
1610                         DB((dbg, LEVEL_3, "Create Phi %+F (for %+F) -", phi, node));
1611 #ifdef DEBUG_libfirm
1612                         {
1613                                 int i;
1614                                 for (i = 0; i < n_preds; ++i) {
1615                                         DB((dbg, LEVEL_3, " %+F", phi_ins[i]));
1616                                 }
1617                                 DB((dbg, LEVEL_3, "\n"));
1618                         }
1619 #endif
1620                         mark_as_copy_of(phi, node);
1621                         sched_add_after(block, phi);
1622
1623                         node = phi;
1624                 } else {
1625                         allocation_info_t *info = get_allocation_info(node);
1626                         info->current_value = phi_ins[0];
1627
1628                         /* Grab 1 of the inputs we constructed (might not be the same as
1629                          * "node" as we could see the same copy of the value in all
1630                          * predecessors */
1631                         node = phi_ins[0];
1632                 }
1633
1634                 /* if the node already has a register assigned use it */
1635                 reg = arch_get_irn_register(node);
1636                 if (reg != NULL) {
1637                         use_reg(node, reg);
1638                 }
1639
1640                 /* remember that this node is live at the beginning of the block */
1641                 ir_nodeset_insert(&live_nodes, node);
1642         }
1643
1644         rbitset_alloca(forbidden_regs, n_regs);
1645
1646         /* handle phis... */
1647         assign_phi_registers(block);
1648
1649         /* all live-ins must have a register */
1650 #ifdef DEBUG_libfirm
1651         {
1652                 ir_nodeset_iterator_t  iter;
1653                 foreach_ir_nodeset(&live_nodes, node, iter) {
1654                         const arch_register_t *reg = arch_get_irn_register(node);
1655                         assert(reg != NULL);
1656                 }
1657         }
1658 #endif
1659
1660         /* assign instructions in the block */
1661         sched_foreach(block, node) {
1662                 int i;
1663                 int arity;
1664                 ir_node *value;
1665
1666                 /* phis are already assigned */
1667                 if (is_Phi(node))
1668                         continue;
1669
1670                 rewire_inputs(node);
1671
1672                 /* enforce use constraints */
1673                 rbitset_clear_all(forbidden_regs, n_regs);
1674                 enforce_constraints(&live_nodes, node, forbidden_regs);
1675
1676                 rewire_inputs(node);
1677
1678                 /* we may not use registers used for inputs for optimistic splits */
1679                 arity = get_irn_arity(node);
1680                 for (i = 0; i < arity; ++i) {
1681                         ir_node *op = get_irn_n(node, i);
1682                         const arch_register_t *reg;
1683                         if (!arch_irn_consider_in_reg_alloc(cls, op))
1684                                 continue;
1685
1686                         reg = arch_get_irn_register(op);
1687                         rbitset_set(forbidden_regs, arch_register_get_index(reg));
1688                 }
1689
1690                 /* free registers of values last used at this instruction */
1691                 free_last_uses(&live_nodes, node);
1692
1693                 /* assign output registers */
1694                 /* TODO: 2 phases: first: pre-assigned ones, 2nd real regs */
1695                 be_foreach_definition(node, cls, value,
1696                         assign_reg(block, value, forbidden_regs);
1697                 );
1698         }
1699
1700         ir_nodeset_destroy(&live_nodes);
1701         assignments = NULL;
1702
1703         block_info->processed = true;
1704
1705         /* permute values at end of predecessor blocks in case of phi-nodes */
1706         if (n_preds > 1) {
1707                 int p;
1708                 for (p = 0; p < n_preds; ++p) {
1709                         add_phi_permutations(block, p);
1710                 }
1711         }
1712
1713         /* if we have exactly 1 successor then we might be able to produce phi
1714            copies now */
1715         if (get_irn_n_edges_kind(block, EDGE_KIND_BLOCK) == 1) {
1716                 const ir_edge_t *edge
1717                         = get_irn_out_edge_first_kind(block, EDGE_KIND_BLOCK);
1718                 ir_node      *succ      = get_edge_src_irn(edge);
1719                 int           p         = get_edge_src_pos(edge);
1720                 block_info_t *succ_info = get_block_info(succ);
1721
1722                 if (succ_info->processed) {
1723                         add_phi_permutations(succ, p);
1724                 }
1725         }
1726 }
1727
1728 typedef struct block_costs_t block_costs_t;
1729 struct block_costs_t {
1730         float costs;   /**< costs of the block */
1731         int   dfs_num; /**< depth first search number (to detect backedges) */
1732 };
1733
1734 static int cmp_block_costs(const void *d1, const void *d2)
1735 {
1736         const ir_node       * const *block1 = d1;
1737         const ir_node       * const *block2 = d2;
1738         const block_costs_t *info1  = get_irn_link(*block1);
1739         const block_costs_t *info2  = get_irn_link(*block2);
1740         return QSORT_CMP(info2->costs, info1->costs);
1741 }
1742
1743 static void determine_block_order(void)
1744 {
1745         int i;
1746         ir_node **blocklist = be_get_cfgpostorder(irg);
1747         int       n_blocks  = ARR_LEN(blocklist);
1748         int       dfs_num   = 0;
1749         pdeq     *worklist  = new_pdeq();
1750         ir_node **order     = XMALLOCN(ir_node*, n_blocks);
1751         int       order_p   = 0;
1752
1753         /* clear block links... */
1754         for (i = 0; i < n_blocks; ++i) {
1755                 ir_node *block = blocklist[i];
1756                 set_irn_link(block, NULL);
1757         }
1758
1759         /* walk blocks in reverse postorder, the costs for each block are the
1760          * sum of the costs of its predecessors (excluding the costs on backedges
1761          * which we can't determine) */
1762         for (i = n_blocks-1; i >= 0; --i) {
1763                 block_costs_t *cost_info;
1764                 ir_node *block = blocklist[i];
1765
1766                 float execfreq   = (float)get_block_execfreq(execfreqs, block);
1767                 float costs      = execfreq;
1768                 int   n_cfgpreds = get_Block_n_cfgpreds(block);
1769                 int   p;
1770                 for (p = 0; p < n_cfgpreds; ++p) {
1771                         ir_node       *pred_block = get_Block_cfgpred_block(block, p);
1772                         block_costs_t *pred_costs = get_irn_link(pred_block);
1773                         /* we don't have any info for backedges */
1774                         if (pred_costs == NULL)
1775                                 continue;
1776                         costs += pred_costs->costs;
1777                 }
1778
1779                 cost_info          = OALLOCZ(&obst, block_costs_t);
1780                 cost_info->costs   = costs;
1781                 cost_info->dfs_num = dfs_num++;
1782                 set_irn_link(block, cost_info);
1783         }
1784
1785         /* sort array by block costs */
1786         qsort(blocklist, n_blocks, sizeof(blocklist[0]), cmp_block_costs);
1787
1788         ir_reserve_resources(irg, IR_RESOURCE_BLOCK_VISITED);
1789         inc_irg_block_visited(irg);
1790
1791         for (i = 0; i < n_blocks; ++i) {
1792                 ir_node       *block = blocklist[i];
1793                 if (Block_block_visited(block))
1794                         continue;
1795
1796                 /* continually add predecessors with highest costs to worklist
1797                  * (without using backedges) */
1798                 do {
1799                         block_costs_t *info       = get_irn_link(block);
1800                         ir_node       *best_pred  = NULL;
1801                         float          best_costs = -1;
1802                         int            n_cfgpred  = get_Block_n_cfgpreds(block);
1803                         int            i;
1804
1805                         pdeq_putr(worklist, block);
1806                         mark_Block_block_visited(block);
1807                         for (i = 0; i < n_cfgpred; ++i) {
1808                                 ir_node       *pred_block = get_Block_cfgpred_block(block, i);
1809                                 block_costs_t *pred_info  = get_irn_link(pred_block);
1810
1811                                 /* ignore backedges */
1812                                 if (pred_info->dfs_num > info->dfs_num)
1813                                         continue;
1814
1815                                 if (info->costs > best_costs) {
1816                                         best_costs = info->costs;
1817                                         best_pred  = pred_block;
1818                                 }
1819                         }
1820                         block = best_pred;
1821                 } while (block != NULL && !Block_block_visited(block));
1822
1823                 /* now put all nodes in the worklist in our final order */
1824                 while (!pdeq_empty(worklist)) {
1825                         ir_node *pblock = pdeq_getr(worklist);
1826                         assert(order_p < n_blocks);
1827                         order[order_p++] = pblock;
1828                 }
1829         }
1830         assert(order_p == n_blocks);
1831         del_pdeq(worklist);
1832
1833         ir_free_resources(irg, IR_RESOURCE_BLOCK_VISITED);
1834
1835         DEL_ARR_F(blocklist);
1836
1837         obstack_free(&obst, NULL);
1838         obstack_init(&obst);
1839
1840         block_order   = order;
1841         n_block_order = n_blocks;
1842 }
1843
1844 /**
1845  * Run the register allocator for the current register class.
1846  */
1847 static void be_pref_alloc_cls(void)
1848 {
1849         int i;
1850
1851         lv = be_assure_liveness(irg);
1852         be_liveness_assure_sets(lv);
1853
1854         ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
1855
1856         DB((dbg, LEVEL_2, "=== Allocating registers of %s ===\n", cls->name));
1857
1858         be_clear_links(irg);
1859
1860         irg_block_walk_graph(irg, NULL, analyze_block, NULL);
1861         if (create_congruence_classes)
1862                 combine_congruence_classes();
1863
1864         for (i = 0; i < n_block_order; ++i) {
1865                 ir_node *block = block_order[i];
1866                 allocate_coalesce_block(block, NULL);
1867         }
1868
1869         ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
1870 }
1871
1872 static void dump(int mask, ir_graph *irg, const char *suffix)
1873 {
1874         if (be_get_irg_options(irg)->dump_flags & mask)
1875                 dump_ir_graph(irg, suffix);
1876 }
1877
1878 /**
1879  * Run the spiller on the current graph.
1880  */
1881 static void spill(void)
1882 {
1883         /* make sure all nodes show their real register pressure */
1884         be_timer_push(T_RA_CONSTR);
1885         be_pre_spill_prepare_constr(irg, cls);
1886         be_timer_pop(T_RA_CONSTR);
1887
1888         dump(DUMP_RA, irg, "-spillprepare");
1889
1890         /* spill */
1891         be_timer_push(T_RA_SPILL);
1892         be_do_spill(irg, cls);
1893         be_timer_pop(T_RA_SPILL);
1894
1895         be_timer_push(T_RA_SPILL_APPLY);
1896         check_for_memory_operands(irg);
1897         be_timer_pop(T_RA_SPILL_APPLY);
1898
1899         dump(DUMP_RA, irg, "-spill");
1900 }
1901
1902 /**
1903  * The pref register allocator for a whole procedure.
1904  */
1905 static void be_pref_alloc(ir_graph *new_irg)
1906 {
1907         const arch_env_t *arch_env = be_get_irg_arch_env(new_irg);
1908         int   n_cls                = arch_env_get_n_reg_class(arch_env);
1909         int   c;
1910
1911         obstack_init(&obst);
1912
1913         irg       = new_irg;
1914         execfreqs = be_get_irg_exec_freq(irg);
1915
1916         /* determine a good coloring order */
1917         determine_block_order();
1918
1919         for (c = 0; c < n_cls; ++c) {
1920                 cls             = arch_env_get_reg_class(arch_env, c);
1921                 default_cls_req = NULL;
1922                 if (arch_register_class_flags(cls) & arch_register_class_flag_manual_ra)
1923                         continue;
1924
1925                 stat_ev_ctx_push_str("regcls", cls->name);
1926
1927                 n_regs      = arch_register_class_n_regs(cls);
1928                 normal_regs = rbitset_malloc(n_regs);
1929                 be_abi_set_non_ignore_regs(be_get_irg_abi(irg), cls, normal_regs);
1930
1931                 spill();
1932
1933                 /* verify schedule and register pressure */
1934                 be_timer_push(T_VERIFY);
1935                 if (be_get_irg_options(irg)->verify_option == BE_VERIFY_WARN) {
1936                         be_verify_schedule(irg);
1937                         be_verify_register_pressure(irg, cls);
1938                 } else if (be_get_irg_options(irg)->verify_option == BE_VERIFY_ASSERT) {
1939                         assert(be_verify_schedule(irg) && "Schedule verification failed");
1940                         assert(be_verify_register_pressure(irg, cls)
1941                                 && "Register pressure verification failed");
1942                 }
1943                 be_timer_pop(T_VERIFY);
1944
1945                 be_timer_push(T_RA_COLOR);
1946                 be_pref_alloc_cls();
1947                 be_timer_pop(T_RA_COLOR);
1948
1949                 /* we most probably constructed new Phis so liveness info is invalid
1950                  * now */
1951                 /* TODO: test liveness_introduce */
1952                 be_liveness_invalidate(lv);
1953                 free(normal_regs);
1954
1955                 stat_ev_ctx_pop("regcls");
1956         }
1957
1958         be_timer_push(T_RA_SPILL_APPLY);
1959         be_abi_fix_stack_nodes(irg);
1960         be_timer_pop(T_RA_SPILL_APPLY);
1961
1962         be_timer_push(T_VERIFY);
1963         if (be_get_irg_options(irg)->verify_option == BE_VERIFY_WARN) {
1964                 be_verify_register_allocation(irg);
1965         } else if (be_get_irg_options(irg)->verify_option == BE_VERIFY_ASSERT) {
1966                 assert(be_verify_register_allocation(irg)
1967                        && "Register allocation invalid");
1968         }
1969         be_timer_pop(T_VERIFY);
1970
1971         obstack_free(&obst, NULL);
1972 }
1973
1974 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_pref_alloc);
1975 void be_init_pref_alloc(void)
1976 {
1977         static be_ra_t be_ra_pref = {
1978                 be_pref_alloc,
1979         };
1980         lc_opt_entry_t *be_grp              = lc_opt_get_grp(firm_opt_get_root(), "be");
1981         lc_opt_entry_t *prefalloc_group = lc_opt_get_grp(be_grp, "prefalloc");
1982         lc_opt_add_table(prefalloc_group, options);
1983
1984         be_register_allocator("pref", &be_ra_pref);
1985         FIRM_DBG_REGISTER(dbg, "firm.be.prefalloc");
1986 }