nsz Git - libfirm/blob - ir/be/benewalloc.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       New approach to allocation and copy coalescing
  23  * @author      Matthias Braun
  24  * @date        14.2.2009
  25  * @version     $Id$
  26  *
  27  * ... WE NEED A NAME FOR THIS ...
  28  *
  29  * Only a proof of concept at this moment...
  30  *
  31  * The idea is to allocate registers in 2 passes:
  32  * 1. A first pass to determine "preferred" registers for live-ranges. This
  33  *    calculates for each register and each live-range a value indicating
  34  *    the usefulness. (You can roughly think of the value as the negative
  35  *    costs needed for copies when the value is in the specific registers...)
  36  *
  37  * 2. Walk blocks and assigns registers in a greedy fashion. Preferring
  38  *    registers with high preferences. When register constraints are not met,
  39  *    add copies and split live-ranges.
  40  *
  41  * TODO:
  42  *  - make use of free registers in the permute_values code
  43  *  - think about a smarter sequence of visiting the blocks. Sorted by
  44  *    execfreq might be good, or looptree from inner to outermost loops going
  45  *    over blocks in a reverse postorder
  46  *  - propagate preferences through Phis
  47  */
  48 #include "config.h"
  49
  50 #include <float.h>
  51 #include <stdbool.h>
  52 #include <math.h>
  53
  54 #include "error.h"
  55 #include "execfreq.h"
  56 #include "ircons.h"
  57 #include "irdom.h"
  58 #include "iredges_t.h"
  59 #include "irgraph_t.h"
  60 #include "irgwalk.h"
  61 #include "irnode_t.h"
  62 #include "irprintf.h"
  63 #include "obst.h"
  64 #include "raw_bitset.h"
  65 #include "unionfind.h"
  66 #include "pdeq.h"
  67 #include "hungarian.h"
  68
  69 #include "beabi.h"
  70 #include "bechordal_t.h"
  71 #include "be.h"
  72 #include "beirg.h"
  73 #include "belive_t.h"
  74 #include "bemodule.h"
  75 #include "benode_t.h"
  76 #include "bera.h"
  77 #include "besched.h"
  78 #include "bespill.h"
  79 #include "bespillutil.h"
  80 #include "beverify.h"
  81 #include "beutil.h"
  82
  83 #define USE_FACTOR         1.0f
  84 #define DEF_FACTOR         1.0f
  85 #define NEIGHBOR_FACTOR    0.2f
  86 #define AFF_SHOULD_BE_SAME 0.5f
  87 #define AFF_PHI            1.0f
  88 #define SPLIT_DELTA        1.0f
  89
  90 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  91
  92 static struct obstack               obst;
  93 static be_irg_t                    *birg;
  94 static ir_graph                    *irg;
  95 static const arch_register_class_t *cls;
  96 static const arch_register_req_t   *default_cls_req;
  97 static be_lv_t                     *lv;
  98 static const ir_exec_freq          *execfreqs;
  99 static unsigned                     n_regs;
 100 static unsigned                    *normal_regs;
 101 static int                         *congruence_classes;
 102 static ir_node                    **block_order;
 103 static int                          n_block_order;
 104
 105 /** currently active assignments (while processing a basic block)
 106  * maps registers to values(their current copies) */
 107 static ir_node **assignments;
 108
 109 /**
 110  * allocation information: last_uses, register preferences
 111  * the information is per firm-node.
 112  */
 113 struct allocation_info_t {
 114         unsigned  last_uses;      /**< bitset indicating last uses (input pos) */
 115         ir_node  *current_value;  /**< copy of the value that should be used */
 116         ir_node  *original_value; /**< for copies point to original value */
 117         float     prefs[0];       /**< register preferences */
 118 };
 119 typedef struct allocation_info_t allocation_info_t;
 120
 121 /** helper datastructure used when sorting register preferences */
 122 struct reg_pref_t {
 123         unsigned num;
 124         float    pref;
 125 };
 126 typedef struct reg_pref_t reg_pref_t;
 127
 128 /** per basic-block information */
 129 struct block_info_t {
 130         bool     processed;       /**< indicate wether block is processed */
 131         ir_node *assignments[0];  /**< register assignments at end of block */
 132 };
 133 typedef struct block_info_t block_info_t;
 134
 135 /**
 136  * Get the allocation info for a node.
 137  * The info is allocated on the first visit of a node.
 138  */
 139 static allocation_info_t *get_allocation_info(ir_node *node)
 140 {
 141         allocation_info_t *info = get_irn_link(node);
 142         if (info == NULL) {
 143                 info = OALLOCFZ(&obst, allocation_info_t, prefs, n_regs);
 144                 info->current_value  = node;
 145                 info->original_value = node;
 146                 set_irn_link(node, info);
 147         }
 148
 149         return info;
 150 }
 151
 152 /**
 153  * Get allocation information for a basic block
 154  */
 155 static block_info_t *get_block_info(ir_node *block)
 156 {
 157         block_info_t *info = get_irn_link(block);
 158
 159         assert(is_Block(block));
 160         if (info == NULL) {
 161                 info = OALLOCFZ(&obst, block_info_t, assignments, n_regs);
 162                 set_irn_link(block, info);
 163         }
 164
 165         return info;
 166 }
 167
 168 /**
 169  * Get default register requirement for the current register class
 170  */
 171 static const arch_register_req_t *get_default_req_current_cls(void)
 172 {
 173         if (default_cls_req == NULL) {
 174                 struct obstack      *obst = get_irg_obstack(irg);
 175                 arch_register_req_t *req  = OALLOCZ(obst, arch_register_req_t);
 176
 177                 req->type = arch_register_req_type_normal;
 178                 req->cls  = cls;
 179
 180                 default_cls_req = req;
 181         }
 182         return default_cls_req;
 183 }
 184
 185 /**
 186  * Link the allocation info of a node to a copy.
 187  * Afterwards, both nodes uses the same allocation info.
 188  * Copy must not have an allocation info assigned yet.
 189  *
 190  * @param copy   the node that gets the allocation info assigned
 191  * @param value  the original node
 192  */
 193 static void mark_as_copy_of(ir_node *copy, ir_node *value)
 194 {
 195         ir_node           *original;
 196         allocation_info_t *info      = get_allocation_info(value);
 197         allocation_info_t *copy_info = get_allocation_info(copy);
 198
 199         /* find original value */
 200         original = info->original_value;
 201         if (original != value) {
 202                 info = get_allocation_info(original);
 203         }
 204
 205         assert(info->original_value == original);
 206         info->current_value = copy;
 207
 208         /* the copy should not be linked to something else yet */
 209         assert(copy_info->original_value == copy);
 210         copy_info->original_value = original;
 211
 212         /* copy over allocation preferences */
 213         memcpy(copy_info->prefs, info->prefs, n_regs * sizeof(copy_info->prefs[0]));
 214 }
 215
 216 /**
 217  * Calculate the penalties for every register on a node and its live neighbors.
 218  *
 219  * @param live_nodes  the set of live nodes at the current position, may be NULL
 220  * @param penalty     the penalty to subtract from
 221  * @param limited     a raw bitset containing the limited set for the node
 222  * @param node        the node
 223  */
 224 static void give_penalties_for_limits(const ir_nodeset_t *live_nodes,
 225                                       float penalty, const unsigned* limited,
 226                                       ir_node *node)
 227 {
 228         ir_nodeset_iterator_t iter;
 229         unsigned              r;
 230         allocation_info_t     *info = get_allocation_info(node);
 231         ir_node               *neighbor;
 232
 233         /* give penalty for all forbidden regs */
 234         for (r = 0; r < n_regs; ++r) {
 235                 if (rbitset_is_set(limited, r))
 236                         continue;
 237
 238                 info->prefs[r] -= penalty;
 239         }
 240
 241         /* all other live values should get a penalty for allowed regs */
 242         if (live_nodes == NULL)
 243                 return;
 244
 245         /* TODO: reduce penalty if there are multiple allowed registers... */
 246         penalty *= NEIGHBOR_FACTOR;
 247         foreach_ir_nodeset(live_nodes, neighbor, iter) {
 248                 allocation_info_t *neighbor_info;
 249
 250                 /* TODO: if op is used on multiple inputs we might not do a
 251                  * continue here */
 252                 if (neighbor == node)
 253                         continue;
 254
 255                 neighbor_info = get_allocation_info(neighbor);
 256                 for (r = 0; r < n_regs; ++r) {
 257                         if (!rbitset_is_set(limited, r))
 258                                 continue;
 259
 260                         neighbor_info->prefs[r] -= penalty;
 261                 }
 262         }
 263 }
 264
 265 /**
 266  * Calculate the preferences of a definition for the current register class.
 267  * If the definition uses a limited set of registers, reduce the preferences
 268  * for the limited register on the node and its neighbors.
 269  *
 270  * @param live_nodes  the set of live nodes at the current node
 271  * @param weight      the weight
 272  * @param node        the current node
 273  */
 274 static void check_defs(const ir_nodeset_t *live_nodes, float weight,
 275                        ir_node *node)
 276 {
 277         const arch_register_req_t *req;
 278
 279         if (get_irn_mode(node) == mode_T) {
 280                 const ir_edge_t *edge;
 281                 foreach_out_edge(node, edge) {
 282                         ir_node *proj = get_edge_src_irn(edge);
 283                         check_defs(live_nodes, weight, proj);
 284                 }
 285                 return;
 286         }
 287
 288         if (!arch_irn_consider_in_reg_alloc(cls, node))
 289                 return;
 290
 291         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_popcnt(&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 = 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                 check_defs(&live_nodes, weight, node);
 352
 353                 /* mark last uses */
 354                 arity = get_irn_arity(node);
 355
 356                 /* the allocation info node currently only uses 1 unsigned value
 357                    to mark last used inputs. So we will fail for a node with more than
 358                    32 inputs. */
 359                 if (arity >= (int) sizeof(unsigned) * 8) {
 360                         panic("Node with more than %d inputs not supported yet",
 361                                         (int) sizeof(unsigned) * 8);
 362                 }
 363
 364                 info = get_allocation_info(node);
 365                 for (i = 0; i < arity; ++i) {
 366                         ir_node *op = get_irn_n(node, i);
 367                         if (!arch_irn_consider_in_reg_alloc(cls, op))
 368                                 continue;
 369
 370                         /* last usage of a value? */
 371                         if (!ir_nodeset_contains(&live_nodes, op)) {
 372                                 rbitset_set(&info->last_uses, i);
 373                         }
 374                 }
 375
 376                 be_liveness_transfer(cls, node, &live_nodes);
 377
 378                 /* update weights based on usage constraints */
 379                 for (i = 0; i < arity; ++i) {
 380                         const arch_register_req_t *req;
 381                         const unsigned            *limited;
 382                         ir_node                   *op = get_irn_n(node, i);
 383
 384                         if (!arch_irn_consider_in_reg_alloc(cls, op))
 385                                 continue;
 386
 387                         req = arch_get_register_req(node, i);
 388                         if (!(req->type & arch_register_req_type_limited))
 389                                 continue;
 390
 391                         limited = req->limited;
 392                         give_penalties_for_limits(&live_nodes, weight * USE_FACTOR, limited,
 393                                                   op);
 394                 }
 395         }
 396
 397         ir_nodeset_destroy(&live_nodes);
 398 }
 399
 400 static void congruence_def(ir_nodeset_t *live_nodes, ir_node *node)
 401 {
 402         const arch_register_req_t *req;
 403
 404         if (get_irn_mode(node) == mode_T) {
 405                 const ir_edge_t *edge;
 406                 foreach_out_edge(node, edge) {
 407                         ir_node *def = get_edge_src_irn(edge);
 408                         congruence_def(live_nodes, def);
 409                 }
 410                 return;
 411         }
 412
 413         if (!arch_irn_consider_in_reg_alloc(cls, node))
 414                 return;
 415
 416         /* should be same constraint? */
 417         req = arch_get_register_req_out(node);
 418         if (req->type & arch_register_req_type_should_be_same) {
 419                 ir_node *insn  = skip_Proj(node);
 420                 int      arity = get_irn_arity(insn);
 421                 int      i;
 422                 unsigned node_idx = get_irn_idx(node);
 423                 node_idx          = uf_find(congruence_classes, node_idx);
 424
 425                 for (i = 0; i < arity; ++i) {
 426                         ir_node               *live;
 427                         ir_node               *op;
 428                         int                    op_idx;
 429                         ir_nodeset_iterator_t  iter;
 430                         bool                   interferes = false;
 431
 432                         if (!rbitset_is_set(&req->other_same, i))
 433                                 continue;
 434
 435                         op     = get_irn_n(insn, i);
 436                         op_idx = get_irn_idx(op);
 437                         op_idx = uf_find(congruence_classes, op_idx);
 438
 439                         /* do we interfere with the value */
 440                         foreach_ir_nodeset(live_nodes, live, iter) {
 441                                 int lv_idx = get_irn_idx(live);
 442                                 lv_idx     = uf_find(congruence_classes, lv_idx);
 443                                 if (lv_idx == op_idx) {
 444                                         interferes = true;
 445                                         break;
 446                                 }
 447                         }
 448                         /* don't put in same affinity class if we interfere */
 449                         if (interferes)
 450                                 continue;
 451
 452                         node_idx = uf_union(congruence_classes, node_idx, op_idx);
 453                         DB((dbg, LEVEL_3, "Merge %+F and %+F congruence classes\n",
 454                             node, op));
 455                         /* one should_be_same is enough... */
 456                         break;
 457                 }
 458         }
 459 }
 460
 461 static void create_congurence_class(ir_node *block, void *data)
 462 {
 463         ir_nodeset_t  live_nodes;
 464         ir_node      *node;
 465
 466         (void) data;
 467         ir_nodeset_init(&live_nodes);
 468         be_liveness_end_of_block(lv, cls, block, &live_nodes);
 469
 470         /* check should be same constraints */
 471         sched_foreach_reverse(block, node) {
 472                 if (is_Phi(node))
 473                         break;
 474
 475                 congruence_def(&live_nodes, node);
 476                 be_liveness_transfer(cls, node, &live_nodes);
 477         }
 478
 479         /* check phi congruence classes */
 480         sched_foreach_reverse_from(node, node) {
 481                 int i;
 482                 int arity;
 483                 int node_idx;
 484                 assert(is_Phi(node));
 485
 486                 if (!arch_irn_consider_in_reg_alloc(cls, node))
 487                         continue;
 488
 489                 node_idx = get_irn_idx(node);
 490                 node_idx = uf_find(congruence_classes, node_idx);
 491
 492                 arity = get_irn_arity(node);
 493                 for (i = 0; i < arity; ++i) {
 494                         bool                  interferes = false;
 495                         ir_nodeset_iterator_t iter;
 496                         ir_node *live;
 497                         ir_node *phi;
 498                         ir_node *op     = get_Phi_pred(node, i);
 499                         int      op_idx = get_irn_idx(op);
 500                         op_idx = uf_find(congruence_classes, op_idx);
 501
 502                         /* do we interfere with the value */
 503                         foreach_ir_nodeset(&live_nodes, live, iter) {
 504                                 int lv_idx = get_irn_idx(live);
 505                                 lv_idx     = uf_find(congruence_classes, lv_idx);
 506                                 if (lv_idx == op_idx) {
 507                                         interferes = true;
 508                                         break;
 509                                 }
 510                         }
 511                         /* don't put in same affinity class if we interfere */
 512                         if (interferes)
 513                                 continue;
 514                         /* any other phi has the same input? */
 515                         sched_foreach(block, phi) {
 516                                 ir_node *oop;
 517                                 int      oop_idx;
 518                                 if (!is_Phi(phi))
 519                                         break;
 520                                 if (!arch_irn_consider_in_reg_alloc(cls, phi))
 521                                         continue;
 522                                 oop = get_Phi_pred(phi, i);
 523                                 if (oop == op)
 524                                         continue;
 525                                 oop_idx = get_irn_idx(oop);
 526                                 oop_idx = uf_find(congruence_classes, oop_idx);
 527                                 if (oop_idx == op_idx) {
 528                                         interferes = true;
 529                                         break;
 530                                 }
 531                         }
 532                         if (interferes)
 533                                 continue;
 534
 535                         node_idx = uf_union(congruence_classes, node_idx, op_idx);
 536                         DB((dbg, LEVEL_3, "Merge %+F and %+F congruence classes\n",
 537                             node, op));
 538                 }
 539         }
 540 }
 541
 542 static void merge_congruence_prefs(ir_node *node, void *data)
 543 {
 544         allocation_info_t *info;
 545         allocation_info_t *head_info;
 546         unsigned node_idx = get_irn_idx(node);
 547         unsigned node_set = uf_find(congruence_classes, node_idx);
 548         unsigned r;
 549
 550         (void) data;
 551
 552         /* head of congruence class or not in any class */
 553         if (node_set == node_idx)
 554                 return;
 555
 556         if (!arch_irn_consider_in_reg_alloc(cls, node))
 557                 return;
 558
 559         head_info = get_allocation_info(get_idx_irn(irg, node_set));
 560         info      = get_allocation_info(node);
 561
 562         for (r = 0; r < n_regs; ++r) {
 563                 head_info->prefs[r] += info->prefs[r];
 564         }
 565 }
 566
 567 static void set_congruence_prefs(ir_node *node, void *data)
 568 {
 569         allocation_info_t *info;
 570         allocation_info_t *head_info;
 571         unsigned node_idx = get_irn_idx(node);
 572         unsigned node_set = uf_find(congruence_classes, node_idx);
 573
 574         (void) data;
 575
 576         /* head of congruence class or not in any class */
 577         if (node_set == node_idx)
 578                 return;
 579
 580         if (!arch_irn_consider_in_reg_alloc(cls, node))
 581                 return;
 582
 583         head_info = get_allocation_info(get_idx_irn(irg, node_set));
 584         info      = get_allocation_info(node);
 585
 586         memcpy(info->prefs, head_info->prefs, n_regs * sizeof(info->prefs[0]));
 587 }
 588
 589 static void combine_congruence_classes(void)
 590 {
 591         size_t n = get_irg_last_idx(irg);
 592         congruence_classes = XMALLOCN(int, n);
 593         uf_init(congruence_classes, n);
 594
 595         /* create congruence classes */
 596         irg_block_walk_graph(irg, create_congurence_class, NULL, NULL);
 597         /* merge preferences */
 598         irg_walk_graph(irg, merge_congruence_prefs, NULL, NULL);
 599         irg_walk_graph(irg, set_congruence_prefs, NULL, NULL);
 600 }
 601
 602
 603
 604
 605
 606 /**
 607  * Assign register reg to the given node.
 608  *
 609  * @param node  the node
 610  * @param reg   the register
 611  */
 612 static void use_reg(ir_node *node, const arch_register_t *reg)
 613 {
 614         unsigned r = arch_register_get_index(reg);
 615         assignments[r] = node;
 616         arch_set_irn_register(node, reg);
 617 }
 618
 619 static void free_reg_of_value(ir_node *node)
 620 {
 621         const arch_register_t *reg;
 622         unsigned               r;
 623
 624         if (!arch_irn_consider_in_reg_alloc(cls, node))
 625                 return;
 626
 627         reg        = arch_get_irn_register(node);
 628         r          = arch_register_get_index(reg);
 629         /* assignment->value may be NULL if a value is used at 2 inputs
 630            so it gets freed twice. */
 631         assert(assignments[r] == node || assignments[r] == NULL);
 632         assignments[r] = NULL;
 633 }
 634
 635 /**
 636  * Compare two register preferences in decreasing order.
 637  */
 638 static int compare_reg_pref(const void *e1, const void *e2)
 639 {
 640         const reg_pref_t *rp1 = (const reg_pref_t*) e1;
 641         const reg_pref_t *rp2 = (const reg_pref_t*) e2;
 642         if (rp1->pref < rp2->pref)
 643                 return 1;
 644         if (rp1->pref > rp2->pref)
 645                 return -1;
 646         return 0;
 647 }
 648
 649 static void fill_sort_candidates(reg_pref_t *regprefs,
 650                                  const allocation_info_t *info)
 651 {
 652         unsigned r;
 653
 654         for (r = 0; r < n_regs; ++r) {
 655                 float pref = info->prefs[r];
 656                 regprefs[r].num  = r;
 657                 regprefs[r].pref = pref;
 658         }
 659         /* TODO: use a stable sort here to avoid unnecessary register jumping */
 660         qsort(regprefs, n_regs, sizeof(regprefs[0]), compare_reg_pref);
 661 }
 662
 663 static bool try_optimistic_split(ir_node *to_split, ir_node *before,
 664                                  float pref, float pref_delta,
 665                                  unsigned *output_regs)
 666 {
 667         const arch_register_t *reg;
 668         ir_node               *original_insn;
 669         ir_node               *block;
 670         ir_node               *copy;
 671         unsigned               r;
 672         unsigned               i;
 673         allocation_info_t     *info = get_allocation_info(to_split);
 674         reg_pref_t            *prefs;
 675         float                  delta;
 676         float                  split_threshold;
 677
 678         (void) pref;
 679
 680         /* stupid hack: don't optimisticallt split don't spill nodes...
 681          * (so we don't split away the values produced because of
 682          *  must_be_different constraints) */
 683         original_insn = skip_Proj(info->original_value);
 684         if (arch_irn_get_flags(original_insn) & arch_irn_flags_dont_spill)
 685                 return false;
 686
 687         /* find the best free position where we could move to */
 688         prefs = ALLOCAN(reg_pref_t, n_regs);
 689         fill_sort_candidates(prefs, info);
 690         for (i = 0; i < n_regs; ++i) {
 691                 r = prefs[i].num;
 692                 if (!rbitset_is_set(normal_regs, r))
 693                         continue;
 694                 if (rbitset_is_set(output_regs, r))
 695                         continue;
 696                 if (assignments[r] == NULL)
 697                         break;
 698         }
 699         if (i >= n_regs) {
 700                 return false;
 701         }
 702         /* TODO: use execfreq somehow... */
 703         block = get_nodes_block(before);
 704         delta = pref_delta + prefs[i].pref;
 705         split_threshold = get_block_execfreq(execfreqs, block) * SPLIT_DELTA;
 706         if (delta < split_threshold) {
 707                 DB((dbg, LEVEL_3, "Not doing optimistical split, win %f too low\n",
 708                     delta));
 709                 return false;
 710         }
 711
 712         reg   = arch_register_for_index(cls, r);
 713         copy = be_new_Copy(cls, block, to_split);
 714         mark_as_copy_of(copy, to_split);
 715         free_reg_of_value(to_split);
 716         use_reg(copy, reg);
 717         sched_add_before(before, copy);
 718
 719         DB((dbg, LEVEL_3,
 720             "Optimistic live-range split %+F move %+F -> %s before %+F (win %f)\n",
 721             copy, to_split, reg->name, before, delta));
 722         return true;
 723 }
 724
 725 /**
 726  * Determine and assign a register for node @p node
 727  */
 728 static void assign_reg(const ir_node *block, ir_node *node,
 729                        unsigned *output_regs)
 730 {
 731         const arch_register_t     *reg;
 732         allocation_info_t         *info;
 733         const arch_register_req_t *req;
 734         reg_pref_t                *reg_prefs;
 735         ir_node                   *in_node;
 736         unsigned                   i;
 737         const unsigned            *allowed_regs;
 738         unsigned                   r;
 739
 740         assert(arch_irn_consider_in_reg_alloc(cls, node));
 741
 742         /* preassigned register? */
 743         reg = arch_get_irn_register(node);
 744         if (reg != NULL) {
 745                 DB((dbg, LEVEL_2, "Preassignment %+F -> %s\n", node, reg->name));
 746                 use_reg(node, reg);
 747                 return;
 748         }
 749
 750         /* give should_be_same boni */
 751         info = get_allocation_info(node);
 752         req  = arch_get_register_req_out(node);
 753
 754         in_node = skip_Proj(node);
 755         if (req->type & arch_register_req_type_should_be_same) {
 756                 float weight = get_block_execfreq(execfreqs, block);
 757                 int   arity  = get_irn_arity(in_node);
 758                 int   i;
 759
 760                 assert(arity <= (int) sizeof(req->other_same) * 8);
 761                 for (i = 0; i < arity; ++i) {
 762                         ir_node               *in;
 763                         const arch_register_t *reg;
 764                         unsigned               r;
 765                         if (!rbitset_is_set(&req->other_same, i))
 766                                 continue;
 767
 768                         in  = get_irn_n(in_node, i);
 769                         reg = arch_get_irn_register(in);
 770                         assert(reg != NULL);
 771                         r = arch_register_get_index(reg);
 772
 773                         /* if the value didn't die here then we should not propagate the
 774                          * should_be_same info */
 775                         if (assignments[r] == in)
 776                                 continue;
 777
 778                         info->prefs[r] += weight * AFF_SHOULD_BE_SAME;
 779                 }
 780         }
 781
 782         /* create list of register candidates and sort by their preference */
 783         DB((dbg, LEVEL_2, "Candidates for %+F:", node));
 784         reg_prefs = alloca(n_regs * sizeof(reg_prefs[0]));
 785         fill_sort_candidates(reg_prefs, info);
 786         for (i = 0; i < n_regs; ++i) {
 787                 unsigned num = reg_prefs[i].num;
 788                 const arch_register_t *reg;
 789
 790                 if (!rbitset_is_set(normal_regs, num))
 791                         continue;
 792
 793                 reg = arch_register_for_index(cls, num);
 794                 DB((dbg, LEVEL_2, " %s(%f)", reg->name, reg_prefs[i].pref));
 795         }
 796         DB((dbg, LEVEL_2, "\n"));
 797
 798         allowed_regs = normal_regs;
 799         if (req->type & arch_register_req_type_limited) {
 800                 allowed_regs = req->limited;
 801         }
 802
 803         for (i = 0; i < n_regs; ++i) {
 804                 r = reg_prefs[i].num;
 805                 if (!rbitset_is_set(allowed_regs, r))
 806                         continue;
 807                 if (assignments[r] == NULL)
 808                         break;
 809                 if (!is_Phi(node)) {
 810                         float    pref   = reg_prefs[i].pref;
 811                         float    delta  = i+1 < n_regs ? pref - reg_prefs[i+1].pref : 0;
 812                         ir_node *before = skip_Proj(node);
 813                         bool     res    = try_optimistic_split(assignments[r], before,
 814                                                                pref, delta,
 815                                                                output_regs);
 816                         if (res)
 817                                 break;
 818                 }
 819         }
 820         if (i >= n_regs) {
 821                 panic("No register left for %+F\n", node);
 822         }
 823
 824         reg = arch_register_for_index(cls, r);
 825         DB((dbg, LEVEL_2, "Assign %+F -> %s\n", node, reg->name));
 826         use_reg(node, reg);
 827 }
 828
 829 /**
 830  * Add an permutation in front of a node and change the assignments
 831  * due to this permutation.
 832  *
 833  * To understand this imagine a permutation like this:
 834  *
 835  * 1 -> 2
 836  * 2 -> 3
 837  * 3 -> 1, 5
 838  * 4 -> 6
 839  * 5
 840  * 6
 841  * 7 -> 7
 842  *
 843  * First we count how many destinations a single value has. At the same time
 844  * we can be sure that each destination register has at most 1 source register
 845  * (it can have 0 which means we don't care what value is in it).
 846  * We ignore all fullfilled permuations (like 7->7)
 847  * In a first pass we create as much copy instructions as possible as they
 848  * are generally cheaper than exchanges. We do this by counting into how many
 849  * destinations a register has to be copied (in the example it's 2 for register
 850  * 3, or 1 for the registers 1,2,4 and 7).
 851  * We can then create a copy into every destination register when the usecount
 852  * of that register is 0 (= noone else needs the value in the register).
 853  *
 854  * After this step we should have cycles left. We implement a cyclic permutation
 855  * of n registers with n-1 transpositions.
 856  *
 857  * @param live_nodes   the set of live nodes, updated due to live range split
 858  * @param before       the node before we add the permutation
 859  * @param permutation  the permutation array indices are the destination
 860  *                     registers, the values in the array are the source
 861  *                     registers.
 862  */
 863 static void permute_values(ir_nodeset_t *live_nodes, ir_node *before,
 864                              unsigned *permutation)
 865 {
 866         unsigned  *n_used = ALLOCANZ(unsigned, n_regs);
 867         ir_node   *block;
 868         unsigned   r;
 869
 870         /* determine how often each source register needs to be read */
 871         for (r = 0; r < n_regs; ++r) {
 872                 unsigned  old_reg = permutation[r];
 873                 ir_node  *value;
 874
 875                 value = assignments[old_reg];
 876                 if (value == NULL) {
 877                         /* nothing to do here, reg is not live. Mark it as fixpoint
 878                          * so we ignore it in the next steps */
 879                         permutation[r] = r;
 880                         continue;
 881                 }
 882
 883                 ++n_used[old_reg];
 884         }
 885
 886         block = get_nodes_block(before);
 887
 888         /* step1: create copies where immediately possible */
 889         for (r = 0; r < n_regs; /* empty */) {
 890                 ir_node *copy;
 891                 ir_node *src;
 892                 const arch_register_t *reg;
 893                 unsigned               old_r = permutation[r];
 894
 895                 /* - no need to do anything for fixed points.
 896                    - we can't copy if the value in the dest reg is still needed */
 897                 if (old_r == r || n_used[r] > 0) {
 898                         ++r;
 899                         continue;
 900                 }
 901
 902                 /* create a copy */
 903                 src  = assignments[old_r];
 904                 copy = be_new_Copy(cls, block, src);
 905                 sched_add_before(before, copy);
 906                 reg = arch_register_for_index(cls, r);
 907                 DB((dbg, LEVEL_2, "Copy %+F (from %+F, before %+F) -> %s\n",
 908                     copy, src, before, reg->name));
 909                 mark_as_copy_of(copy, src);
 910                 use_reg(copy, reg);
 911
 912                 if (live_nodes != NULL) {
 913                         ir_nodeset_insert(live_nodes, copy);
 914                 }
 915
 916                 /* old register has 1 user less, permutation is resolved */
 917                 assert(arch_register_get_index(arch_get_irn_register(src)) == old_r);
 918                 permutation[r] = r;
 919
 920                 assert(n_used[old_r] > 0);
 921                 --n_used[old_r];
 922                 if (n_used[old_r] == 0) {
 923                         if (live_nodes != NULL) {
 924                                 ir_nodeset_remove(live_nodes, src);
 925                         }
 926                         free_reg_of_value(src);
 927                 }
 928
 929                 /* advance or jump back (if this copy enabled another copy) */
 930                 if (old_r < r && n_used[old_r] == 0) {
 931                         r = old_r;
 932                 } else {
 933                         ++r;
 934                 }
 935         }
 936
 937         /* at this point we only have "cycles" left which we have to resolve with
 938          * perm instructions
 939          * TODO: if we have free registers left, then we should really use copy
 940          * instructions for any cycle longer than 2 registers...
 941          * (this is probably architecture dependent, there might be archs where
 942          *  copies are preferable even for 2-cycles) */
 943
 944         /* create perms with the rest */
 945         for (r = 0; r < n_regs; /* empty */) {
 946                 const arch_register_t *reg;
 947                 unsigned  old_r = permutation[r];
 948                 unsigned  r2;
 949                 ir_node  *in[2];
 950                 ir_node  *perm;
 951                 ir_node  *proj0;
 952                 ir_node  *proj1;
 953
 954                 if (old_r == r) {
 955                         ++r;
 956                         continue;
 957                 }
 958
 959                 /* we shouldn't have copies from 1 value to multiple destinations left*/
 960                 assert(n_used[old_r] == 1);
 961
 962                 /* exchange old_r and r2; after that old_r is a fixed point */
 963                 r2 = permutation[old_r];
 964
 965                 in[0] = assignments[r2];
 966                 in[1] = assignments[old_r];
 967                 perm = be_new_Perm(cls, block, 2, in);
 968                 sched_add_before(before, perm);
 969                 DB((dbg, LEVEL_2, "Perm %+F (perm %+F,%+F, before %+F)\n",
 970                     perm, in[0], in[1], before));
 971
 972                 proj0 = new_r_Proj(block, perm, get_irn_mode(in[0]), 0);
 973                 mark_as_copy_of(proj0, in[0]);
 974                 reg = arch_register_for_index(cls, old_r);
 975                 use_reg(proj0, reg);
 976
 977                 proj1 = new_r_Proj(block, perm, get_irn_mode(in[1]), 1);
 978                 mark_as_copy_of(proj1, in[1]);
 979                 reg = arch_register_for_index(cls, r2);
 980                 use_reg(proj1, reg);
 981
 982                 /* 1 value is now in the correct register */
 983                 permutation[old_r] = old_r;
 984                 /* the source of r changed to r2 */
 985                 permutation[r] = r2;
 986
 987                 /* if we have reached a fixpoint update data structures */
 988                 if (live_nodes != NULL) {
 989                         ir_nodeset_remove(live_nodes, in[0]);
 990                         ir_nodeset_remove(live_nodes, in[1]);
 991                         ir_nodeset_remove(live_nodes, proj0);
 992                         ir_nodeset_insert(live_nodes, proj1);
 993                 }
 994         }
 995
 996 #ifdef DEBUG_libfirm
 997         /* now we should only have fixpoints left */
 998         for (r = 0; r < n_regs; ++r) {
 999                 assert(permutation[r] == r);
1000         }
1001 #endif
1002 }
1003
1004 /**
1005  * Free regs for values last used.
1006  *
1007  * @param live_nodes   set of live nodes, will be updated
1008  * @param node         the node to consider
1009  */
1010 static void free_last_uses(ir_nodeset_t *live_nodes, ir_node *node)
1011 {
1012         allocation_info_t     *info      = get_allocation_info(node);
1013         const unsigned        *last_uses = &info->last_uses;
1014         int                    arity     = get_irn_arity(node);
1015         int                    i;
1016
1017         for (i = 0; i < arity; ++i) {
1018                 ir_node *op;
1019
1020                 /* check if one operand is the last use */
1021                 if (!rbitset_is_set(last_uses, i))
1022                         continue;
1023
1024                 op = get_irn_n(node, i);
1025                 free_reg_of_value(op);
1026                 ir_nodeset_remove(live_nodes, op);
1027         }
1028 }
1029
1030 /**
1031  * change inputs of a node to the current value (copies/perms)
1032  */
1033 static void rewire_inputs(ir_node *node)
1034 {
1035         int i;
1036         int arity = get_irn_arity(node);
1037
1038         for (i = 0; i < arity; ++i) {
1039                 ir_node           *op = get_irn_n(node, i);
1040                 allocation_info_t *info;
1041
1042                 if (!arch_irn_consider_in_reg_alloc(cls, op))
1043                         continue;
1044
1045                 info = get_allocation_info(op);
1046                 info = get_allocation_info(info->original_value);
1047                 if (info->current_value != op) {
1048                         set_irn_n(node, i, info->current_value);
1049                 }
1050         }
1051 }
1052
1053 /**
1054  * Create a bitset of registers occupied with value living through an
1055  * instruction
1056  */
1057 static void determine_live_through_regs(unsigned *bitset, ir_node *node)
1058 {
1059         const allocation_info_t *info = get_allocation_info(node);
1060         unsigned r;
1061         int i;
1062         int arity;
1063
1064         /* mark all used registers as potentially live-through */
1065         for (r = 0; r < n_regs; ++r) {
1066                 if (assignments[r] == NULL)
1067                         continue;
1068                 if (!rbitset_is_set(normal_regs, r))
1069                         continue;
1070
1071                 rbitset_set(bitset, r);
1072         }
1073
1074         /* remove registers of value dying at the instruction */
1075         arity = get_irn_arity(node);
1076         for (i = 0; i < arity; ++i) {
1077                 ir_node               *op;
1078                 const arch_register_t *reg;
1079
1080                 if (!rbitset_is_set(&info->last_uses, i))
1081                         continue;
1082
1083                 op  = get_irn_n(node, i);
1084                 reg = arch_get_irn_register(op);
1085                 rbitset_clear(bitset, arch_register_get_index(reg));
1086         }
1087 }
1088
1089 /**
1090  * Enforce constraints at a node by live range splits.
1091  *
1092  * @param  live_nodes  the set of live nodes, might be changed
1093  * @param  node        the current node
1094  */
1095 static void enforce_constraints(ir_nodeset_t *live_nodes, ir_node *node,
1096                                 unsigned *output_regs)
1097 {
1098         int arity = get_irn_arity(node);
1099         int i, res;
1100         hungarian_problem_t *bp;
1101         unsigned l, r;
1102         unsigned *assignment;
1103
1104         /* construct a list of register occupied by live-through values */
1105         unsigned *live_through_regs = NULL;
1106
1107         /* see if any use constraints are not met */
1108         bool good = true;
1109         for (i = 0; i < arity; ++i) {
1110                 ir_node                   *op = get_irn_n(node, i);
1111                 const arch_register_t     *reg;
1112                 const arch_register_req_t *req;
1113                 const unsigned            *limited;
1114                 unsigned                  r;
1115
1116                 if (!arch_irn_consider_in_reg_alloc(cls, op))
1117                         continue;
1118
1119                 /* are there any limitations for the i'th operand? */
1120                 req = arch_get_register_req(node, i);
1121                 if (!(req->type & arch_register_req_type_limited))
1122                         continue;
1123
1124                 limited = req->limited;
1125                 reg     = arch_get_irn_register(op);
1126                 r       = arch_register_get_index(reg);
1127                 if (!rbitset_is_set(limited, r)) {
1128                         /* found an assignment outside the limited set */
1129                         good = false;
1130                         break;
1131                 }
1132         }
1133
1134         /* is any of the live-throughs using a constrained output register? */
1135         if (get_irn_mode(node) == mode_T) {
1136                 const ir_edge_t *edge;
1137
1138                 foreach_out_edge(node, edge) {
1139                         ir_node *proj = get_edge_src_irn(edge);
1140                         const arch_register_req_t *req;
1141
1142                         if (!arch_irn_consider_in_reg_alloc(cls, proj))
1143                                 continue;
1144
1145                         req = arch_get_register_req_out(proj);
1146                         if (!(req->type & arch_register_req_type_limited))
1147                                 continue;
1148
1149                         if (live_through_regs == NULL) {
1150                                 rbitset_alloca(live_through_regs, n_regs);
1151                                 determine_live_through_regs(live_through_regs, node);
1152                         }
1153
1154                         rbitset_or(output_regs, req->limited, n_regs);
1155                         if (rbitsets_have_common(req->limited, live_through_regs, n_regs)) {
1156                                 good = false;
1157                         }
1158                 }
1159         } else {
1160                 if (arch_irn_consider_in_reg_alloc(cls, node)) {
1161                         const arch_register_req_t *req = arch_get_register_req_out(node);
1162                         if (req->type & arch_register_req_type_limited) {
1163                                 rbitset_alloca(live_through_regs, n_regs);
1164                                 determine_live_through_regs(live_through_regs, node);
1165                                 if (rbitsets_have_common(req->limited, live_through_regs, n_regs)) {
1166                                         good = false;
1167                                         rbitset_or(output_regs, req->limited, n_regs);
1168                                 }
1169                         }
1170                 }
1171         }
1172
1173         if (good)
1174                 return;
1175
1176         /* create these arrays if we haven't yet */
1177         if (live_through_regs == NULL) {
1178                 rbitset_alloca(live_through_regs, n_regs);
1179         }
1180
1181         /* at this point we have to construct a bipartite matching problem to see
1182          * which values should go to which registers
1183          * Note: We're building the matrix in "reverse" - source registers are
1184          *       right, destinations left because this will produce the solution
1185          *       in the format required for permute_values.
1186          */
1187         bp = hungarian_new(n_regs, n_regs, HUNGARIAN_MATCH_PERFECT);
1188
1189         /* add all combinations, then remove not allowed ones */
1190         for (l = 0; l < n_regs; ++l) {
1191                 if (!rbitset_is_set(normal_regs, l)) {
1192                         hungarian_add(bp, l, l, 1);
1193                         continue;
1194                 }
1195
1196                 for (r = 0; r < n_regs; ++r) {
1197                         if (!rbitset_is_set(normal_regs, r))
1198                                 continue;
1199                         /* livethrough values may not use constrainted output registers */
1200                         if (rbitset_is_set(live_through_regs, l)
1201                                         && rbitset_is_set(output_regs, r))
1202                                 continue;
1203
1204                         hungarian_add(bp, r, l, l == r ? 9 : 8);
1205                 }
1206         }
1207
1208         for (i = 0; i < arity; ++i) {
1209                 ir_node                   *op = get_irn_n(node, i);
1210                 const arch_register_t     *reg;
1211                 const arch_register_req_t *req;
1212                 const unsigned            *limited;
1213                 unsigned                   current_reg;
1214
1215                 if (!arch_irn_consider_in_reg_alloc(cls, op))
1216                         continue;
1217
1218                 req = arch_get_register_req(node, i);
1219                 if (!(req->type & arch_register_req_type_limited))
1220                         continue;
1221
1222                 limited     = req->limited;
1223                 reg         = arch_get_irn_register(op);
1224                 current_reg = arch_register_get_index(reg);
1225                 for (r = 0; r < n_regs; ++r) {
1226                         if (rbitset_is_set(limited, r))
1227                                 continue;
1228                         hungarian_remv(bp, r, current_reg);
1229                 }
1230         }
1231
1232         //hungarian_print_cost_matrix(bp, 1);
1233         hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
1234
1235         assignment = ALLOCAN(unsigned, n_regs);
1236         res = hungarian_solve(bp, (int*) assignment, NULL, 0);
1237         assert(res == 0);
1238
1239 #if 0
1240         fprintf(stderr, "Swap result:");
1241         for (i = 0; i < (int) n_regs; ++i) {
1242                 fprintf(stderr, " %d", assignment[i]);
1243         }
1244         fprintf(stderr, "\n");
1245 #endif
1246
1247         hungarian_free(bp);
1248
1249         permute_values(live_nodes, node, assignment);
1250 }
1251
1252 /** test wether a node @p n is a copy of the value of node @p of */
1253 static bool is_copy_of(ir_node *value, ir_node *test_value)
1254 {
1255         allocation_info_t *test_info;
1256         allocation_info_t *info;
1257
1258         if (value == test_value)
1259                 return true;
1260
1261         info      = get_allocation_info(value);
1262         test_info = get_allocation_info(test_value);
1263         return test_info->original_value == info->original_value;
1264 }
1265
1266 /**
1267  * find a value in the end-assignment of a basic block
1268  * @returns the index into the assignment array if found
1269  *          -1 if not found
1270  */
1271 static int find_value_in_block_info(block_info_t *info, ir_node *value)
1272 {
1273         unsigned   r;
1274         ir_node  **assignments = info->assignments;
1275         for (r = 0; r < n_regs; ++r) {
1276                 ir_node *a_value = assignments[r];
1277
1278                 if (a_value == NULL)
1279                         continue;
1280                 if (is_copy_of(a_value, value))
1281                         return (int) r;
1282         }
1283
1284         return -1;
1285 }
1286
1287 /**
1288  * Create the necessary permutations at the end of a basic block to fullfill
1289  * the register assignment for phi-nodes in the next block
1290  */
1291 static void add_phi_permutations(ir_node *block, int p)
1292 {
1293         unsigned   r;
1294         unsigned  *permutation;
1295         ir_node  **old_assignments;
1296         bool       need_permutation;
1297         ir_node   *node;
1298         ir_node   *pred = get_Block_cfgpred_block(block, p);
1299
1300         block_info_t *pred_info = get_block_info(pred);
1301
1302         /* predecessor not processed yet? nothing to do */
1303         if (!pred_info->processed)
1304                 return;
1305
1306         permutation = ALLOCAN(unsigned, n_regs);
1307         for (r = 0; r < n_regs; ++r) {
1308                 permutation[r] = r;
1309         }
1310
1311         /* check phi nodes */
1312         need_permutation = false;
1313         node = sched_first(block);
1314         for ( ; is_Phi(node); node = sched_next(node)) {
1315                 const arch_register_t *reg;
1316                 int                    regn;
1317                 int                    a;
1318                 ir_node               *op;
1319
1320                 if (!arch_irn_consider_in_reg_alloc(cls, node))
1321                         continue;
1322
1323                 op = get_Phi_pred(node, p);
1324                 if (!arch_irn_consider_in_reg_alloc(cls, op))
1325                         continue;
1326
1327                 a  = find_value_in_block_info(pred_info, op);
1328                 assert(a >= 0);
1329
1330                 reg  = arch_get_irn_register(node);
1331                 regn = arch_register_get_index(reg);
1332                 if (regn != a) {
1333                         permutation[regn] = a;
1334                         need_permutation  = true;
1335                 }
1336         }
1337
1338         if (need_permutation) {
1339                 /* permute values at end of predecessor */
1340                 old_assignments = assignments;
1341                 assignments     = pred_info->assignments;
1342                 permute_values(NULL, be_get_end_of_block_insertion_point(pred),
1343                                                  permutation);
1344                 assignments = old_assignments;
1345         }
1346
1347         /* change phi nodes to use the copied values */
1348         node = sched_first(block);
1349         for ( ; is_Phi(node); node = sched_next(node)) {
1350                 int      a;
1351                 ir_node *op;
1352
1353                 if (!arch_irn_consider_in_reg_alloc(cls, node))
1354                         continue;
1355
1356                 op = get_Phi_pred(node, p);
1357                 /* no need to do anything for Unknown inputs */
1358                 if (!arch_irn_consider_in_reg_alloc(cls, op))
1359                         continue;
1360
1361                 /* we have permuted all values into the correct registers so we can
1362                    simply query which value occupies the phis register in the
1363                    predecessor */
1364                 a  = arch_register_get_index(arch_get_irn_register(node));
1365                 op = pred_info->assignments[a];
1366                 set_Phi_pred(node, p, op);
1367         }
1368 }
1369
1370 /**
1371  * Set preferences for a phis register based on the registers used on the
1372  * phi inputs.
1373  */
1374 static void adapt_phi_prefs(ir_node *phi)
1375 {
1376         int i;
1377         int arity = get_irn_arity(phi);
1378         ir_node           *block = get_nodes_block(phi);
1379         allocation_info_t *info  = get_allocation_info(phi);
1380
1381         for (i = 0; i < arity; ++i) {
1382                 ir_node               *op  = get_irn_n(phi, i);
1383                 const arch_register_t *reg = arch_get_irn_register(op);
1384                 ir_node               *pred_block;
1385                 block_info_t          *pred_block_info;
1386                 float                  weight;
1387                 unsigned               r;
1388
1389                 if (reg == NULL)
1390                         continue;
1391                 /* we only give the bonus if the predecessor already has registers
1392                  * assigned, otherwise we only see a dummy value
1393                  * and any conclusions about its register are useless */
1394                 pred_block = get_Block_cfgpred_block(block, i);
1395                 pred_block_info = get_block_info(pred_block);
1396                 if (!pred_block_info->processed)
1397                         continue;
1398
1399                 /* give bonus for already assigned register */
1400                 weight = get_block_execfreq(execfreqs, pred_block);
1401                 r      = arch_register_get_index(reg);
1402                 info->prefs[r] += weight * AFF_PHI;
1403         }
1404 }
1405
1406 /**
1407  * After a phi has been assigned a register propagate preference inputs
1408  * to the phi inputs.
1409  */
1410 static void propagate_phi_register(ir_node *phi, unsigned r)
1411 {
1412         int                    i;
1413         ir_node               *block = get_nodes_block(phi);
1414         int                    arity = get_irn_arity(phi);
1415
1416         for (i = 0; i < arity; ++i) {
1417                 ir_node           *op         = get_Phi_pred(phi, i);
1418                 allocation_info_t *info       = get_allocation_info(op);
1419                 ir_node           *pred_block = get_Block_cfgpred_block(block, i);
1420                 float              weight
1421                         = get_block_execfreq(execfreqs, pred_block) * AFF_PHI;
1422
1423                 if (info->prefs[r] >= weight)
1424                         continue;
1425
1426                 /* promote the prefered register */
1427                 info->prefs[r] = AFF_PHI * weight;
1428                 if (is_Phi(op))
1429                         propagate_phi_register(op, r);
1430         }
1431 }
1432
1433 static void assign_phi_registers(ir_node *block)
1434 {
1435         int                  n_phis = 0;
1436         int                  n;
1437         int                  res;
1438         int                 *assignment;
1439         ir_node             *node;
1440         hungarian_problem_t *bp;
1441
1442         /* count phi nodes */
1443         sched_foreach(block, node) {
1444                 if (!is_Phi(node))
1445                         break;
1446                 if (!arch_irn_consider_in_reg_alloc(cls, node))
1447                         continue;
1448                 ++n_phis;
1449         }
1450
1451         if (n_phis == 0)
1452                 return;
1453
1454         /* build a bipartite matching problem for all phi nodes */
1455         bp = hungarian_new(n_phis, n_regs, HUNGARIAN_MATCH_PERFECT);
1456         n  = 0;
1457         sched_foreach(block, node) {
1458                 unsigned r;
1459
1460                 allocation_info_t *info;
1461                 if (!is_Phi(node))
1462                         break;
1463                 if (!arch_irn_consider_in_reg_alloc(cls, node))
1464                         continue;
1465
1466                 /* give boni for predecessor colorings */
1467                 adapt_phi_prefs(node);
1468                 /* add stuff to bipartite problem */
1469                 info = get_allocation_info(node);
1470                 DB((dbg, LEVEL_3, "Prefs for %+F: ", node));
1471                 for (r = 0; r < n_regs; ++r) {
1472                         float costs;
1473
1474                         if (!rbitset_is_set(normal_regs, r))
1475                                 continue;
1476
1477                         costs = info->prefs[r];
1478                         costs = costs < 0 ? -logf(-costs+1) : logf(costs+1);
1479                         costs *= 100;
1480                         costs += 10000;
1481                         hungarian_add(bp, n, r, costs);
1482                         DB((dbg, LEVEL_3, " %s(%f)", arch_register_for_index(cls, r)->name,
1483                                                 info->prefs[r]));
1484                 }
1485                 DB((dbg, LEVEL_3, "\n"));
1486                 ++n;
1487         }
1488
1489         //hungarian_print_cost_matrix(bp, 7);
1490         hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
1491
1492         assignment = ALLOCAN(int, n_regs);
1493         res        = hungarian_solve(bp, assignment, NULL, 0);
1494         assert(res == 0);
1495
1496         /* apply results */
1497         n = 0;
1498         sched_foreach(block, node) {
1499                 unsigned               r;
1500                 const arch_register_t *reg;
1501
1502                 if (!is_Phi(node))
1503                         break;
1504                 if (!arch_irn_consider_in_reg_alloc(cls, node))
1505                         continue;
1506
1507                 r   = assignment[n++];
1508                 assert(rbitset_is_set(normal_regs, r));
1509                 reg = arch_register_for_index(cls, r);
1510                 DB((dbg, LEVEL_2, "Assign %+F -> %s\n", node, reg->name));
1511                 use_reg(node, reg);
1512
1513                 /* adapt preferences for phi inputs */
1514                 propagate_phi_register(node, r);
1515         }
1516 }
1517
1518 /**
1519  * Walker: assign registers to all nodes of a block that
1520  * need registers from the currently considered register class.
1521  */
1522 static void allocate_coalesce_block(ir_node *block, void *data)
1523 {
1524         int                    i;
1525         ir_nodeset_t           live_nodes;
1526         ir_nodeset_iterator_t  iter;
1527         ir_node               *node;
1528         int                    n_preds;
1529         block_info_t          *block_info;
1530         block_info_t         **pred_block_infos;
1531         ir_node              **phi_ins;
1532         unsigned              *output_regs; /**< collects registers which must not
1533                                                                                      be used for optimistic splits */
1534
1535         (void) data;
1536         DB((dbg, LEVEL_2, "* Block %+F\n", block));
1537
1538         /* clear assignments */
1539         block_info  = get_block_info(block);
1540         assignments = block_info->assignments;
1541
1542         ir_nodeset_init(&live_nodes);
1543
1544         /* gather regalloc infos of predecessor blocks */
1545         n_preds             = get_Block_n_cfgpreds(block);
1546         pred_block_infos    = ALLOCAN(block_info_t*, n_preds);
1547         for (i = 0; i < n_preds; ++i) {
1548                 ir_node      *pred      = get_Block_cfgpred_block(block, i);
1549                 block_info_t *pred_info = get_block_info(pred);
1550                 pred_block_infos[i]     = pred_info;
1551         }
1552
1553         phi_ins = ALLOCAN(ir_node*, n_preds);
1554
1555         /* collect live-in nodes and preassigned values */
1556         be_lv_foreach(lv, block, be_lv_state_in, i) {
1557                 const arch_register_t *reg;
1558                 int                    p;
1559                 bool                   need_phi = false;
1560
1561                 node = be_lv_get_irn(lv, block, i);
1562                 if (!arch_irn_consider_in_reg_alloc(cls, node))
1563                         continue;
1564
1565                 /* check all predecessors for this value, if it is not everywhere the
1566                    same or unknown then we have to construct a phi
1567                    (we collect the potential phi inputs here) */
1568                 for (p = 0; p < n_preds; ++p) {
1569                         block_info_t *pred_info = pred_block_infos[p];
1570
1571                         if (!pred_info->processed) {
1572                                 /* use node for now, it will get fixed later */
1573                                 phi_ins[p] = node;
1574                                 need_phi   = true;
1575                         } else {
1576                                 int a = find_value_in_block_info(pred_info, node);
1577
1578                                 /* must live out of predecessor */
1579                                 assert(a >= 0);
1580                                 phi_ins[p] = pred_info->assignments[a];
1581                                 /* different value from last time? then we need a phi */
1582                                 if (p > 0 && phi_ins[p-1] != phi_ins[p]) {
1583                                         need_phi = true;
1584                                 }
1585                         }
1586                 }
1587
1588                 if (need_phi) {
1589                         ir_mode                   *mode = get_irn_mode(node);
1590                         const arch_register_req_t *req  = get_default_req_current_cls();
1591                         ir_node                   *phi;
1592                         int                        i;
1593
1594                         phi = new_r_Phi(block, n_preds, phi_ins, mode);
1595                         be_set_phi_reg_req(phi, req);
1596
1597                         DB((dbg, LEVEL_3, "Create Phi %+F (for %+F) -", phi, node));
1598 #ifdef DEBUG_libfirm
1599                         for (i = 0; i < n_preds; ++i) {
1600                                 DB((dbg, LEVEL_3, " %+F", phi_ins[i]));
1601                         }
1602                         DB((dbg, LEVEL_3, "\n"));
1603 #endif
1604                         mark_as_copy_of(phi, node);
1605                         sched_add_after(block, phi);
1606
1607                         node = phi;
1608                 } else {
1609                         allocation_info_t *info = get_allocation_info(node);
1610                         info->current_value = phi_ins[0];
1611
1612                         /* Grab 1 of the inputs we constructed (might not be the same as
1613                          * "node" as we could see the same copy of the value in all
1614                          * predecessors */
1615                         node = phi_ins[0];
1616                 }
1617
1618                 /* if the node already has a register assigned use it */
1619                 reg = arch_get_irn_register(node);
1620                 if (reg != NULL) {
1621                         /* TODO: consult pred-block infos here. The value could be copied
1622                            away in some/all predecessor blocks. We need to construct
1623                            phi-nodes in this case.
1624                            We even need to construct some Phi_0 like constructs in cases
1625                            where the predecessor allocation is not determined yet. */
1626                         use_reg(node, reg);
1627                 }
1628
1629                 /* remember that this node is live at the beginning of the block */
1630                 ir_nodeset_insert(&live_nodes, node);
1631         }
1632
1633         rbitset_alloca(output_regs, n_regs);
1634
1635         /* handle phis... */
1636         assign_phi_registers(block);
1637
1638         /* assign regs for live-in values */
1639         foreach_ir_nodeset(&live_nodes, node, iter) {
1640                 const arch_register_t *reg = arch_get_irn_register(node);
1641                 if (reg != NULL)
1642                         continue;
1643
1644                 assign_reg(block, node, output_regs);
1645                 /* shouldn't happen if we color in dominance order */
1646                 assert (!is_Phi(node));
1647         }
1648
1649         /* assign instructions in the block */
1650         sched_foreach(block, node) {
1651                 unsigned r;
1652
1653                 /* phis are already assigned */
1654                 if (is_Phi(node))
1655                         continue;
1656
1657                 rewire_inputs(node);
1658
1659                 /* enforce use constraints */
1660                 rbitset_clear_all(output_regs, n_regs);
1661                 enforce_constraints(&live_nodes, node, output_regs);
1662                 /* we may not use registers occupied here for optimistic splits */
1663                 for (r = 0; r < n_regs; ++r) {
1664                         if (assignments[r] != NULL)
1665                                 rbitset_set(output_regs, r);
1666                 }
1667
1668                 rewire_inputs(node);
1669
1670                 /* free registers of values last used at this instruction */
1671                 free_last_uses(&live_nodes, node);
1672
1673                 /* assign output registers */
1674                 /* TODO: 2 phases: first: pre-assigned ones, 2nd real regs */
1675                 if (get_irn_mode(node) == mode_T) {
1676                         const ir_edge_t *edge;
1677                         foreach_out_edge(node, edge) {
1678                                 ir_node *proj = get_edge_src_irn(edge);
1679                                 if (!arch_irn_consider_in_reg_alloc(cls, proj))
1680                                         continue;
1681                                 assign_reg(block, proj, output_regs);
1682                         }
1683                 } else if (arch_irn_consider_in_reg_alloc(cls, node)) {
1684                         assign_reg(block, node, output_regs);
1685                 }
1686         }
1687
1688         ir_nodeset_destroy(&live_nodes);
1689         assignments = NULL;
1690
1691         block_info->processed = true;
1692
1693         /* permute values at end of predecessor blocks in case of phi-nodes */
1694         if (n_preds > 1) {
1695                 int p;
1696                 for (p = 0; p < n_preds; ++p) {
1697                         add_phi_permutations(block, p);
1698                 }
1699         }
1700
1701         /* if we have exactly 1 successor then we might be able to produce phi
1702            copies now */
1703         if (get_irn_n_edges_kind(block, EDGE_KIND_BLOCK) == 1) {
1704                 const ir_edge_t *edge
1705                         = get_irn_out_edge_first_kind(block, EDGE_KIND_BLOCK);
1706                 ir_node      *succ      = get_edge_src_irn(edge);
1707                 int           p         = get_edge_src_pos(edge);
1708                 block_info_t *succ_info = get_block_info(succ);
1709
1710                 if (succ_info->processed) {
1711                         add_phi_permutations(succ, p);
1712                 }
1713         }
1714 }
1715
1716 typedef struct block_costs_t block_costs_t;
1717 struct block_costs_t {
1718         float costs;   /**< costs of the block */
1719         int   dfs_num; /**< depth first search number (to detect backedges) */
1720 };
1721
1722 static int cmp_block_costs(const void *d1, const void *d2)
1723 {
1724         const ir_node       * const *block1 = d1;
1725         const ir_node       * const *block2 = d2;
1726         const block_costs_t *info1  = get_irn_link(*block1);
1727         const block_costs_t *info2  = get_irn_link(*block2);
1728         return QSORT_CMP(info2->costs, info1->costs);
1729 }
1730
1731 static void determine_block_order(void)
1732 {
1733         int i;
1734         ir_node **blocklist = be_get_cfgpostorder(irg);
1735         int       n_blocks  = ARR_LEN(blocklist);
1736         int       dfs_num   = 0;
1737         pdeq     *worklist  = new_pdeq();
1738         ir_node **order     = XMALLOCN(ir_node*, n_blocks);
1739         int       order_p   = 0;
1740
1741         /* clear block links... */
1742         for (i = 0; i < n_blocks; ++i) {
1743                 ir_node *block = blocklist[i];
1744                 set_irn_link(block, NULL);
1745         }
1746
1747         /* walk blocks in reverse postorder, the costs for each block are the
1748          * sum of the costs of its predecessors (excluding the costs on backedges
1749          * which we can't determine) */
1750         for (i = n_blocks-1; i >= 0; --i) {
1751                 block_costs_t *cost_info;
1752                 ir_node *block = blocklist[i];
1753
1754                 float execfreq   = get_block_execfreq(execfreqs, block);
1755                 float costs      = execfreq;
1756                 int   n_cfgpreds = get_Block_n_cfgpreds(block);
1757                 int   p;
1758                 for (p = 0; p < n_cfgpreds; ++p) {
1759                         ir_node       *pred_block = get_Block_cfgpred_block(block, p);
1760                         block_costs_t *pred_costs = get_irn_link(pred_block);
1761                         /* we don't have any info for backedges */
1762                         if (pred_costs == NULL)
1763                                 continue;
1764                         costs += pred_costs->costs;
1765                 }
1766
1767                 cost_info          = OALLOCZ(&obst, block_costs_t);
1768                 cost_info->costs   = costs;
1769                 cost_info->dfs_num = dfs_num++;
1770                 set_irn_link(block, cost_info);
1771         }
1772
1773         /* sort array by block costs */
1774         qsort(blocklist, n_blocks, sizeof(blocklist[0]), cmp_block_costs);
1775
1776         ir_reserve_resources(irg, IR_RESOURCE_BLOCK_VISITED);
1777         inc_irg_block_visited(irg);
1778
1779         for (i = 0; i < n_blocks; ++i) {
1780                 ir_node       *block = blocklist[i];
1781                 if (Block_block_visited(block))
1782                         continue;
1783
1784                 /* continually add predecessors with highest costs to worklist
1785                  * (without using backedges) */
1786                 do {
1787                         block_costs_t *info       = get_irn_link(block);
1788                         ir_node       *best_pred  = NULL;
1789                         float          best_costs = -1;
1790                         int            n_cfgpred  = get_Block_n_cfgpreds(block);
1791                         int            i;
1792
1793                         pdeq_putr(worklist, block);
1794                         mark_Block_block_visited(block);
1795                         for (i = 0; i < n_cfgpred; ++i) {
1796                                 ir_node       *pred_block = get_Block_cfgpred_block(block, i);
1797                                 block_costs_t *pred_info  = get_irn_link(pred_block);
1798
1799                                 /* ignore backedges */
1800                                 if (pred_info->dfs_num > info->dfs_num)
1801                                         continue;
1802
1803                                 if (info->costs > best_costs) {
1804                                         best_costs = info->costs;
1805                                         best_pred  = pred_block;
1806                                 }
1807                         }
1808                         block = best_pred;
1809                 } while(block != NULL && !Block_block_visited(block));
1810
1811                 /* now put all nodes in the worklist in our final order */
1812                 while (!pdeq_empty(worklist)) {
1813                         ir_node *pblock = pdeq_getr(worklist);
1814                         assert(order_p < n_blocks);
1815                         order[order_p++] = pblock;
1816                 }
1817         }
1818         assert(order_p == n_blocks);
1819         del_pdeq(worklist);
1820
1821         ir_free_resources(irg, IR_RESOURCE_BLOCK_VISITED);
1822
1823         DEL_ARR_F(blocklist);
1824
1825         obstack_free(&obst, NULL);
1826         obstack_init(&obst);
1827
1828         block_order   = order;
1829         n_block_order = n_blocks;
1830 }
1831
1832 /**
1833  * Run the register allocator for the current register class.
1834  */
1835 static void be_straight_alloc_cls(void)
1836 {
1837         int i;
1838
1839         lv = be_assure_liveness(birg);
1840         be_liveness_assure_sets(lv);
1841         be_liveness_assure_chk(lv);
1842
1843         ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
1844
1845         DB((dbg, LEVEL_2, "=== Allocating registers of %s ===\n", cls->name));
1846
1847         be_clear_links(irg);
1848         irg_block_walk_graph(irg, NULL, analyze_block, NULL);
1849         combine_congruence_classes();
1850
1851         for (i = 0; i < n_block_order; ++i) {
1852                 ir_node *block = block_order[i];
1853                 allocate_coalesce_block(block, NULL);
1854         }
1855
1856         ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
1857 }
1858
1859 static void dump(int mask, ir_graph *irg, const char *suffix,
1860                  void (*dumper)(ir_graph *, const char *))
1861 {
1862         if(birg->main_env->options->dump_flags & mask)
1863                 be_dump(irg, suffix, dumper);
1864 }
1865
1866 /**
1867  * Run the spiller on the current graph.
1868  */
1869 static void spill(void)
1870 {
1871         /* make sure all nodes show their real register pressure */
1872         BE_TIMER_PUSH(t_ra_constr);
1873         be_pre_spill_prepare_constr(birg, cls);
1874         BE_TIMER_POP(t_ra_constr);
1875
1876         dump(DUMP_RA, irg, "-spillprepare", dump_ir_block_graph_sched);
1877
1878         /* spill */
1879         BE_TIMER_PUSH(t_ra_spill);
1880         be_do_spill(birg, cls);
1881         BE_TIMER_POP(t_ra_spill);
1882
1883         BE_TIMER_PUSH(t_ra_spill_apply);
1884         check_for_memory_operands(irg);
1885         BE_TIMER_POP(t_ra_spill_apply);
1886
1887         dump(DUMP_RA, irg, "-spill", dump_ir_block_graph_sched);
1888 }
1889
1890 /**
1891  * The straight register allocator for a whole procedure.
1892  */
1893 static void be_straight_alloc(be_irg_t *new_birg)
1894 {
1895         const arch_env_t *arch_env = new_birg->main_env->arch_env;
1896         int   n_cls                = arch_env_get_n_reg_class(arch_env);
1897         int   c;
1898
1899         obstack_init(&obst);
1900
1901         birg      = new_birg;
1902         irg       = be_get_birg_irg(birg);
1903         execfreqs = birg->exec_freq;
1904
1905         /* determine a good coloring order */
1906         determine_block_order();
1907
1908         for (c = 0; c < n_cls; ++c) {
1909                 cls             = arch_env_get_reg_class(arch_env, c);
1910                 default_cls_req = NULL;
1911                 if (arch_register_class_flags(cls) & arch_register_class_flag_manual_ra)
1912                         continue;
1913
1914                 stat_ev_ctx_push_str("regcls", cls->name);
1915
1916                 n_regs      = arch_register_class_n_regs(cls);
1917                 normal_regs = rbitset_malloc(n_regs);
1918                 be_abi_set_non_ignore_regs(birg->abi, cls, normal_regs);
1919
1920                 spill();
1921
1922                 /* verify schedule and register pressure */
1923                 BE_TIMER_PUSH(t_verify);
1924                 if (birg->main_env->options->vrfy_option == BE_VRFY_WARN) {
1925                         be_verify_schedule(birg);
1926                         be_verify_register_pressure(birg, cls, irg);
1927                 } else if (birg->main_env->options->vrfy_option == BE_VRFY_ASSERT) {
1928                         assert(be_verify_schedule(birg) && "Schedule verification failed");
1929                         assert(be_verify_register_pressure(birg, cls, irg)
1930                                 && "Register pressure verification failed");
1931                 }
1932                 BE_TIMER_POP(t_verify);
1933
1934                 BE_TIMER_PUSH(t_ra_color);
1935                 be_straight_alloc_cls();
1936                 BE_TIMER_POP(t_ra_color);
1937
1938                 /* we most probably constructed new Phis so liveness info is invalid
1939                  * now */
1940                 /* TODO: test liveness_introduce */
1941                 be_liveness_invalidate(lv);
1942                 free(normal_regs);
1943
1944                 stat_ev_ctx_pop("regcls");
1945         }
1946
1947         BE_TIMER_PUSH(t_ra_spill_apply);
1948         be_abi_fix_stack_nodes(birg->abi);
1949         BE_TIMER_POP(t_ra_spill_apply);
1950
1951         BE_TIMER_PUSH(t_verify);
1952         if (birg->main_env->options->vrfy_option == BE_VRFY_WARN) {
1953                 be_verify_register_allocation(birg);
1954         } else if (birg->main_env->options->vrfy_option == BE_VRFY_ASSERT) {
1955                 assert(be_verify_register_allocation(birg)
1956                                 && "Register allocation invalid");
1957         }
1958         BE_TIMER_POP(t_verify);
1959
1960         obstack_free(&obst, NULL);
1961 }
1962
1963 /**
1964  * Initializes this module.
1965  */
1966 void be_init_straight_alloc(void)
1967 {
1968         static be_ra_t be_ra_straight = {
1969                 be_straight_alloc,
1970         };
1971
1972         FIRM_DBG_REGISTER(dbg, "firm.be.straightalloc");
1973
1974         be_register_allocator("straight", &be_ra_straight);
1975 }
1976
1977 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_straight_alloc);