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 "prefered" registers for live-ranges. This
  33  *    calculates for each register and each live-range a value indicating
  34  *    the usefullness. (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. Prefering registers
  38  *    with high preferences. When register constraints are not met, add copies
  39  *    and split live-ranges.
  40  *
  41  * TODO:
  42  *  - output constraints are not ensured. The algorithm fails to copy values
  43  *    away, so the registers for constrained outputs are free.
  44  *  - must_be_different constraint is not respected
  45  *  - No parallel copies at basic block borders are created, no additional phis
  46  *    created after copies have been inserted.
  47  *  - Phi color assignment should give bonus points towards registers already
  48  *    assigned at predecessors.
  49  *  - think about a smarter sequence of visiting the blocks. Sorted by
  50  *    execfreq might be good, or looptree from inner to outermost loops going
  51  *    over blocks in a reverse postorder
  52  */
  53 #include "config.h"
  54
  55 #include <float.h>
  56
  57 #include "obst.h"
  58 #include "irnode_t.h"
  59 #include "irgraph_t.h"
  60 #include "iredges_t.h"
  61 #include "ircons.h"
  62 #include "irgwalk.h"
  63 #include "execfreq.h"
  64
  65 #include "be.h"
  66 #include "bera.h"
  67 #include "belive_t.h"
  68 #include "bemodule.h"
  69 #include "bechordal_t.h"
  70 #include "besched_t.h"
  71 #include "beirg_t.h"
  72 #include "benode_t.h"
  73 #include "bespilloptions.h"
  74 #include "beverify.h"
  75
  76 #include "bipartite.h"
  77 #include "hungarian.h"
  78
  79 #define USE_FACTOR       1.0f
  80 #define DEF_FACTOR       1.0f
  81 #define NEIGHBOR_FACTOR  0.2f
  82 #define SHOULD_BE_SAME   1.0f
  83
  84 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  85
  86 static struct obstack               obst;
  87 static be_irg_t                    *birg;
  88 static ir_graph                    *irg;
  89 static const arch_register_class_t *cls;
  90 static be_lv_t                     *lv;
  91 static const ir_exec_freq          *execfreqs;
  92 static unsigned                     n_regs;
  93 static bitset_t                    *ignore_regs;
  94
  95 typedef struct assignment_t assignment_t;
  96 struct assignment_t {
  97         ir_node *value;            /**< currently assigned value */
  98 };
  99
 100 static assignment_t *assignments;
 101
 102 typedef struct allocation_info_t allocation_info_t;
 103 struct allocation_info_t {
 104         unsigned      last_uses;   /**< bitset indicating last uses (input pos) */
 105         assignment_t *current_assignment;
 106         float         prefs[0];    /**< register preferences */
 107 };
 108
 109 typedef struct reg_pref_t reg_pref_t;
 110 struct reg_pref_t {
 111         unsigned num;
 112         float    pref;
 113 };
 114
 115 static allocation_info_t *get_allocation_info(ir_node *node)
 116 {
 117         allocation_info_t *info;
 118         if (!irn_visited(node)) {
 119                 size_t size = sizeof(info[0]) + n_regs * sizeof(float);
 120                 info = obstack_alloc(&obst, size);
 121                 memset(info, 0, size);
 122                 set_irn_link(node, info);
 123                 mark_irn_visited(node);
 124         } else {
 125                 info = get_irn_link(node);
 126         }
 127
 128         return info;
 129 }
 130
 131 static void link_to(ir_node *copy, ir_node *value)
 132 {
 133         assert(!irn_visited(copy));
 134         allocation_info_t *info = get_allocation_info(value);
 135         set_irn_link(copy, info);
 136         mark_irn_visited(copy);
 137 }
 138
 139 static void give_penalties_for_limits(const ir_nodeset_t *live_nodes,
 140                                       float penalty, const unsigned* limited,
 141                                                                           ir_node *node)
 142 {
 143         ir_nodeset_iterator_t  iter;
 144         unsigned               r;
 145         allocation_info_t     *info = get_allocation_info(node);
 146
 147         /* give penalty for all forbidden regs */
 148         for (r = 0; r < n_regs; ++r) {
 149                 if (rbitset_is_set(limited, r))
 150                         continue;
 151
 152                 info->prefs[r] -= penalty;
 153         }
 154
 155         /* all other live values should get a penalty for allowed regs */
 156         if (live_nodes == NULL)
 157                 return;
 158
 159         /* TODO: reduce penalty if there are multiple allowed registers... */
 160         penalty *= NEIGHBOR_FACTOR;
 161         ir_node *neighbor;
 162         foreach_ir_nodeset(live_nodes, neighbor, iter) {
 163                 allocation_info_t *neighbor_info;
 164
 165                 /* TODO: if op is used on multiple inputs we might not do a
 166                  * continue here */
 167                 if (neighbor == node)
 168                         continue;
 169
 170                 neighbor_info = get_allocation_info(neighbor);
 171                 for (r = 0; r < n_regs; ++r) {
 172                         if (!rbitset_is_set(limited, r))
 173                                 continue;
 174
 175                         neighbor_info->prefs[r] -= penalty;
 176                 }
 177         }
 178 }
 179
 180 static void check_defs(const ir_nodeset_t *live_nodes, float weight,
 181                        ir_node *node)
 182 {
 183         if (get_irn_mode(node) == mode_T) {
 184                 const ir_edge_t *edge;
 185                 foreach_out_edge(node, edge) {
 186                         ir_node *proj = get_edge_src_irn(edge);
 187                         check_defs(live_nodes, weight, proj);
 188                 }
 189                 return;
 190         }
 191
 192         if (!arch_irn_consider_in_reg_alloc(cls, node))
 193                 return;
 194
 195         const arch_register_req_t *req = arch_get_register_req_out(node);
 196         if (req->type & arch_register_req_type_limited) {
 197                 const unsigned *limited = req->limited;
 198                 float           penalty = weight * DEF_FACTOR;
 199                 give_penalties_for_limits(live_nodes, penalty, limited, node);
 200         }
 201
 202         if (req->type & arch_register_req_type_should_be_same) {
 203                 ir_node           *insn  = skip_Proj(node);
 204                 allocation_info_t *info  = get_allocation_info(node);
 205                 int                arity = get_irn_arity(insn);
 206                 int                i;
 207
 208                 float factor = 1.0f / rbitset_popcnt(&req->other_same, arity);
 209                 for (i = 0; i < arity; ++i) {
 210                         ir_node  *op;
 211                         unsigned  r;
 212                         if (!rbitset_is_set(&req->other_same, i))
 213                                 continue;
 214
 215                         op = get_irn_n(insn, i);
 216                         allocation_info_t *op_info = get_allocation_info(op);
 217                         for (r = 0; r < n_regs; ++r) {
 218                                 if (bitset_is_set(ignore_regs, r))
 219                                         continue;
 220                                 op_info->prefs[r] += info->prefs[r] * factor;
 221                         }
 222                 }
 223         }
 224 }
 225
 226 static void analyze_block(ir_node *block, void *data)
 227 {
 228         float         weight = get_block_execfreq(execfreqs, block);
 229         ir_nodeset_t  live_nodes;
 230         ir_node      *node;
 231         (void) data;
 232
 233         ir_nodeset_init(&live_nodes);
 234         be_liveness_end_of_block(lv, cls, block, &live_nodes);
 235
 236         sched_foreach_reverse(block, node) {
 237                 int arity;
 238                 int i;
 239
 240                 if(is_Phi(node)) {
 241                         /* TODO: handle constrained phi-nodes */
 242                         break;
 243                 }
 244
 245                 /* TODO give/take penalties for should_be_same/different) */
 246                 check_defs(&live_nodes, weight, node);
 247
 248                 /* mark last uses */
 249                 arity = get_irn_arity(node);
 250                 /* I was lazy, and only allocated 1 unsigned
 251                    => maximum of 32 uses per node (rewrite if necessary) */
 252                 assert(arity <= (int) sizeof(unsigned) * 8);
 253
 254                 allocation_info_t *info = get_allocation_info(node);
 255
 256                 for (i = 0; i < arity; ++i) {
 257                         ir_node *op = get_irn_n(node, i);
 258                         if (!arch_irn_consider_in_reg_alloc(cls, op))
 259                                 continue;
 260
 261                         /* last usage of a value? */
 262                         if (!ir_nodeset_contains(&live_nodes, op)) {
 263                                 rbitset_set(&info->last_uses, i);
 264                         }
 265                 }
 266
 267                 be_liveness_transfer(cls, node, &live_nodes);
 268
 269                 /* update weights based on usage constraints */
 270                 for (i = 0; i < arity; ++i) {
 271                         ir_node  *op = get_irn_n(node, i);
 272                         if (!arch_irn_consider_in_reg_alloc(cls, op))
 273                                 continue;
 274
 275                         const arch_register_req_t *req = arch_get_register_req(node, i);
 276                         if ((req->type & arch_register_req_type_limited) == 0)
 277                                 continue;
 278
 279                         /* TODO: give penalties to neighbors for precolored nodes! */
 280
 281                         const unsigned *limited = req->limited;
 282                         give_penalties_for_limits(&live_nodes, weight * USE_FACTOR, limited,
 283                                                   op);
 284                 }
 285         }
 286
 287         ir_nodeset_destroy(&live_nodes);
 288 }
 289
 290 static void use_reg(ir_node *node, const arch_register_t *reg)
 291 {
 292         unsigned      r          = arch_register_get_index(reg);
 293         assignment_t *assignment = &assignments[r];
 294
 295         assert(assignment->value == NULL);
 296         assignment->value = node;
 297
 298         allocation_info_t *info = get_allocation_info(node);
 299         info->current_assignment = assignment;
 300
 301         arch_set_irn_register(node, reg);
 302 }
 303
 304 static int compare_reg_pref(const void *e1, const void *e2)
 305 {
 306         const reg_pref_t *rp1 = (const reg_pref_t*) e1;
 307         const reg_pref_t *rp2 = (const reg_pref_t*) e2;
 308         if (rp1->pref < rp2->pref)
 309                 return 1;
 310         if (rp1->pref > rp2->pref)
 311                 return -1;
 312         return 0;
 313 }
 314
 315 static void fill_sort_candidates(reg_pref_t *regprefs,
 316                                  const allocation_info_t *info)
 317 {
 318         unsigned r;
 319
 320         for (r = 0; r < n_regs; ++r) {
 321                 float pref = info->prefs[r];
 322                 if (bitset_is_set(ignore_regs, r)) {
 323                         pref = -10000;
 324                 }
 325                 regprefs[r].num  = r;
 326                 regprefs[r].pref = pref;
 327         }
 328         /* TODO: use a stable sort here to avoid unnecessary register jumping */
 329         qsort(regprefs, n_regs, sizeof(regprefs[0]), compare_reg_pref);
 330 }
 331
 332 static void assign_reg(const ir_node *block, ir_node *node)
 333 {
 334         assert(arch_irn_consider_in_reg_alloc(cls, node));
 335
 336         /* preassigned register? */
 337         const arch_register_t *reg = arch_get_irn_register(node);
 338         if (reg != NULL) {
 339                 DB((dbg, LEVEL_2, "Preassignment %+F -> %s\n", node, reg->name));
 340                 use_reg(node, reg);
 341                 return;
 342         }
 343
 344         /* give should_be_same boni */
 345         allocation_info_t         *info = get_allocation_info(node);
 346         const arch_register_req_t *req  = arch_get_register_req_out(node);
 347         if (req->type & arch_register_req_type_should_be_same) {
 348                 float weight = get_block_execfreq(execfreqs, block);
 349                 int   arity  = get_irn_arity(node);
 350                 int   i;
 351                 assert(arity <= (int) sizeof(req->other_same) * 8);
 352                 for (i = 0; i < arity; ++i) {
 353                         ir_node               *in;
 354                         const arch_register_t *reg;
 355                         unsigned               r;
 356                         if (!rbitset_is_set(&req->other_same, i))
 357                                 continue;
 358
 359                         in  = get_irn_n(node, i);
 360                         reg = arch_get_irn_register(in);
 361                         assert(reg != NULL);
 362                         r = arch_register_get_index(reg);
 363                         if (bitset_is_set(ignore_regs, r))
 364                                 continue;
 365                         info->prefs[r] += weight * SHOULD_BE_SAME;
 366                 }
 367         }
 368
 369         /* TODO: handle must_be_different */
 370
 371         /*  */
 372         DB((dbg, LEVEL_2, "Candidates for %+F:", node));
 373         reg_pref_t *reg_prefs = alloca(n_regs * sizeof(reg_prefs[0]));
 374         fill_sort_candidates(reg_prefs, info);
 375         unsigned i;
 376         for (i = 0; i < n_regs; ++i) {
 377                 unsigned               num = reg_prefs[i].num;
 378                 const arch_register_t *reg = arch_register_for_index(cls, num);
 379                 DB((dbg, LEVEL_2, " %s(%f)", reg->name, reg_prefs[i].pref));
 380         }
 381         DB((dbg, LEVEL_2, "\n"));
 382
 383         for (i = 0; i < n_regs; ++i) {
 384                 unsigned r = reg_prefs[i].num;
 385                 /* ignores should be last and we should have a non-ignore left */
 386                 assert(!bitset_is_set(ignore_regs, r));
 387                 /* already used? TODO: It might be better to copy the value occupying the register around here, find out when... */
 388                 if (assignments[r].value != NULL)
 389                         continue;
 390                 use_reg(node, arch_register_for_index(cls, r));
 391                 break;
 392         }
 393 }
 394
 395 static void free_reg_of_value(ir_node *node)
 396 {
 397         if (!arch_irn_consider_in_reg_alloc(cls, node))
 398                 return;
 399
 400         allocation_info_t *info       = get_allocation_info(node);
 401         assignment_t      *assignment = info->current_assignment;
 402
 403         assert(assignment != NULL);
 404
 405         unsigned r = assignment - assignments;
 406         DB((dbg, LEVEL_2, "Value %+F ended, freeing %s\n",
 407                 node, arch_register_for_index(cls, r)->name));
 408         assignment->value        = NULL;
 409         info->current_assignment = NULL;
 410 }
 411
 412 static unsigned get_current_reg(ir_node *node)
 413 {
 414         allocation_info_t *info       = get_allocation_info(node);
 415         assignment_t      *assignment = info->current_assignment;
 416         return assignment - assignments;
 417 }
 418
 419 static assignment_t *get_current_assignment(ir_node *node)
 420 {
 421         allocation_info_t *info = get_allocation_info(node);
 422         return info->current_assignment;
 423 }
 424
 425 static void permutate_values(ir_nodeset_t *live_nodes, ir_node *before,
 426                              unsigned *permutation)
 427 {
 428         ir_node **in = ALLOCAN(ir_node*, n_regs);
 429         size_t    r;
 430
 431         int i = 0;
 432         for (r = 0; r < n_regs; ++r) {
 433                 unsigned new_reg = permutation[r];
 434                 if (new_reg == r)
 435                         continue;
 436
 437                 assignment_t *assignment = &assignments[r];
 438                 ir_node      *value      = assignment->value;
 439                 if (value == NULL) {
 440                         /* nothing to do here, reg is not live */
 441                         permutation[r] = r;
 442                         continue;
 443                 }
 444
 445                 in[i++] = value;
 446
 447                 free_reg_of_value(value);
 448                 ir_nodeset_remove(live_nodes, value);
 449         }
 450
 451         ir_node *block = get_nodes_block(before);
 452         ir_node *perm  = be_new_Perm(cls, irg, block, i, in);
 453
 454         sched_add_before(before, perm);
 455
 456         i = 0;
 457         for (r = 0; r < n_regs; ++r) {
 458                 unsigned new_reg = permutation[r];
 459                 if (new_reg == r)
 460                         continue;
 461
 462                 ir_node *value = in[i];
 463                 ir_mode *mode  = get_irn_mode(value);
 464                 ir_node *proj  = new_r_Proj(irg, block, perm, mode, i);
 465
 466                 const arch_register_t *reg = arch_register_for_index(cls, new_reg);
 467
 468                 link_to(proj, value);
 469                 use_reg(proj, reg);
 470                 ir_nodeset_insert(live_nodes, proj);
 471
 472                 ++i;
 473         }
 474 }
 475
 476 /* free regs for values last used */
 477 static void free_last_uses(ir_nodeset_t *live_nodes, ir_node *node)
 478 {
 479         allocation_info_t *info  = get_allocation_info(node);
 480         int                arity = get_irn_arity(node);
 481         int                i;
 482         for (i = 0; i < arity; ++i) {
 483                 if (!rbitset_is_set(&info->last_uses, i))
 484                         continue;
 485
 486                 ir_node *op = get_irn_n(node, i);
 487                 free_reg_of_value(op);
 488                 ir_nodeset_remove(live_nodes, op);
 489         }
 490 }
 491
 492 static void enforce_constraints(ir_nodeset_t *live_nodes, ir_node *node)
 493 {
 494         int arity = get_irn_arity(node);
 495         int i;
 496
 497         /* see if any use constraints are not met */
 498         bool good = true;
 499         for (i = 0; i < arity; ++i) {
 500                 ir_node *op = get_irn_n(node, i);
 501                 if (!arch_irn_consider_in_reg_alloc(cls, op))
 502                         continue;
 503
 504                 const arch_register_req_t *req = arch_get_register_req(node, i);
 505                 if ((req->type & arch_register_req_type_limited) == 0)
 506                         continue;
 507
 508                 const unsigned *limited = req->limited;
 509                 unsigned        r       = get_current_reg(op);
 510                 if (!rbitset_is_set(limited, r)) {
 511                         good = false;
 512                         break;
 513                 }
 514         }
 515
 516         if (good)
 517                 return;
 518
 519         /* swap values around */
 520         hungarian_problem_t *bp = hungarian_new(n_regs, n_regs, HUNGARIAN_MATCH_PERFECT);
 521
 522         /* add all combinations, then remove not allowed ones */
 523         unsigned l, r;
 524         for (l = 0; l < n_regs; ++l) {
 525                 if (bitset_is_set(ignore_regs, l)) {
 526                         hungarian_add(bp, l, l, 90);
 527                         continue;
 528                 }
 529
 530                 for (r = 0; r < n_regs; ++r) {
 531                         if (bitset_is_set(ignore_regs, r))
 532                                 continue;
 533
 534                         hungarian_add(bp, l, r, l == r ? 90 : 89);
 535                 }
 536         }
 537
 538         for (i = 0; i < arity; ++i) {
 539                 ir_node *op = get_irn_n(node, i);
 540                 if (!arch_irn_consider_in_reg_alloc(cls, op))
 541                         continue;
 542
 543                 const arch_register_req_t *req = arch_get_register_req(node, i);
 544                 if ((req->type & arch_register_req_type_limited) == 0)
 545                         continue;
 546
 547                 const unsigned *limited     = req->limited;
 548                 unsigned        current_reg = get_current_reg(op);
 549                 unsigned        r;
 550                 for (r = 0; r < n_regs; ++r) {
 551                         if (rbitset_is_set(limited, r))
 552                                 continue;
 553                         hungarian_remv(bp, current_reg, r);
 554                 }
 555         }
 556
 557         hungarian_print_costmatrix(bp, 1);
 558         hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
 559
 560         unsigned *assignment = ALLOCAN(unsigned, n_regs);
 561         int dummy;
 562         int res = hungarian_solve(bp, (int*) assignment, &dummy, 0);
 563         assert(res == 0);
 564
 565         printf("Swap result:");
 566         unsigned p;
 567         for (p = 0; p < n_regs; ++p) {
 568                 printf(" %d", assignment[p]);
 569         }
 570         printf("\n");
 571
 572         hungarian_free(bp);
 573
 574         permutate_values(live_nodes, node, assignment);
 575 }
 576
 577 static void allocate_coalesce_block(ir_node *block, void *data)
 578 {
 579         int                    i;
 580         ir_nodeset_t           live_nodes;
 581         ir_nodeset_iterator_t  iter;
 582         ir_node               *node;
 583
 584         /* clear assignments */
 585         memset(assignments, 0, n_regs * sizeof(assignments[0]));
 586
 587         (void) data;
 588
 589         /* collect live-in nodes and preassigned values */
 590         ir_nodeset_init(&live_nodes);
 591         be_lv_foreach(lv, block, be_lv_state_in, i) {
 592                 node = be_lv_get_irn(lv, block, i);
 593                 if (!arch_irn_consider_in_reg_alloc(cls, node))
 594                         continue;
 595
 596                 ir_nodeset_insert(&live_nodes, node);
 597
 598                 /* fill in regs already assigned */
 599                 const arch_register_t *reg = arch_get_irn_register(node);
 600                 if (reg != NULL) {
 601                         use_reg(node, reg);
 602                 }
 603         }
 604
 605         /* handle phis... */
 606         node = sched_first(block);
 607         for ( ; is_Phi(node); node = sched_next(node)) {
 608                 if (!arch_irn_consider_in_reg_alloc(cls, node))
 609                         continue;
 610
 611                 /* fill in regs already assigned */
 612                 const arch_register_t *reg = arch_get_irn_register(node);
 613                 if (reg != NULL) {
 614                         use_reg(node, reg);
 615                 } else {
 616                         /* TODO: give boni for registers already assigned at the
 617                            predecessors */
 618                 }
 619         }
 620         ir_node *start = node;
 621
 622         /* assign regs for live-in values */
 623         foreach_ir_nodeset(&live_nodes, node, iter) {
 624                 assign_reg(block, node);
 625         }
 626
 627         /* assign instructions in the block */
 628         for (node = start; !sched_is_end(node); node = sched_next(node)) {
 629                 int arity = get_irn_arity(node);
 630                 int i;
 631
 632                 /* enforce use constraints */
 633                 enforce_constraints(&live_nodes, node);
 634
 635                 /* exchange values to copied values where needed */
 636                 for (i = 0; i < arity; ++i) {
 637                         ir_node *op = get_irn_n(node, i);
 638                         if (!arch_irn_consider_in_reg_alloc(cls, op))
 639                                 continue;
 640                         assignment_t *assignment = get_current_assignment(op);
 641                         assert(assignment != NULL);
 642                         if (op != assignment->value) {
 643                                 set_irn_n(node, i, assignment->value);
 644                         }
 645                 }
 646
 647                 free_last_uses(&live_nodes, node);
 648
 649                 /* assign output registers */
 650                 /* TODO: 2 phases: first: pre-assigned ones, 2nd real regs */
 651                 if (get_irn_mode(node) == mode_T) {
 652                         const ir_edge_t *edge;
 653                         foreach_out_edge(node, edge) {
 654                                 ir_node *proj = get_edge_src_irn(edge);
 655                                 if (!arch_irn_consider_in_reg_alloc(cls, proj))
 656                                         continue;
 657                                 assign_reg(block, proj);
 658                         }
 659                 } else if (arch_irn_consider_in_reg_alloc(cls, node)) {
 660                         assign_reg(block, node);
 661                 }
 662         }
 663
 664         foreach_ir_nodeset(&live_nodes, node, iter) {
 665                 free_reg_of_value(node);
 666         }
 667
 668         ir_nodeset_destroy(&live_nodes);
 669 }
 670
 671 void be_straight_alloc_cls(void)
 672 {
 673         n_regs         = arch_register_class_n_regs(cls);
 674         lv             = be_assure_liveness(birg);
 675         be_liveness_assure_sets(lv);
 676         be_liveness_assure_chk(lv);
 677
 678         size_t size = n_regs * sizeof(assignments[0]);
 679         assignments = obstack_alloc(&obst, size);
 680
 681         ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_IRN_VISITED);
 682         inc_irg_visited(irg);
 683
 684         irg_block_walk_graph(irg, analyze_block, NULL, NULL);
 685         irg_block_walk_graph(irg, allocate_coalesce_block, NULL, NULL);
 686
 687         ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_IRN_VISITED);
 688 }
 689
 690 static void spill(void)
 691 {
 692         /* TODO: rewrite pre_spill_prepare to work without chordal_env... */
 693         be_chordal_env_t  cenv;
 694         memset(&cenv, 0, sizeof(cenv));
 695         cenv.obst          = &obst;
 696         cenv.irg           = irg;
 697         cenv.birg          = birg;
 698         cenv.cls           = cls;
 699         cenv.ignore_colors = ignore_regs;
 700
 701         /* make sure all nodes show their real register pressure */
 702         be_pre_spill_prepare_constr(&cenv);
 703
 704         /* spill */
 705         be_do_spill(birg, cls);
 706
 707         check_for_memory_operands(irg);
 708 }
 709
 710 static void be_straight_alloc(be_irg_t *new_birg)
 711 {
 712         const arch_env_t *arch_env = new_birg->main_env->arch_env;
 713         int   n_cls                = arch_env_get_n_reg_class(arch_env);
 714         int   c;
 715
 716         obstack_init(&obst);
 717
 718         birg      = new_birg;
 719         irg       = be_get_birg_irg(birg);
 720         execfreqs = birg->exec_freq;
 721
 722         /* TODO: extract some of the stuff from bechordal allocator, like
 723          * statistics, time measurements, etc. and use them here too */
 724
 725         for (c = 0; c < n_cls; ++c) {
 726                 cls = arch_env_get_reg_class(arch_env, c);
 727                 if (arch_register_class_flags(cls) & arch_register_class_flag_manual_ra)
 728                         continue;
 729
 730                 stat_ev_ctx_push_str("bestraight_cls", cls->name);
 731
 732                 n_regs      = cls->n_regs;
 733                 ignore_regs = bitset_malloc(n_regs);
 734                 be_put_ignore_regs(birg, cls, ignore_regs);
 735
 736                 spill();
 737
 738                 /* verify schedule and register pressure */
 739                 BE_TIMER_PUSH(t_verify);
 740                 if (birg->main_env->options->vrfy_option == BE_CH_VRFY_WARN) {
 741                         be_verify_schedule(birg);
 742                         be_verify_register_pressure(birg, cls, irg);
 743                 } else if (birg->main_env->options->vrfy_option == BE_CH_VRFY_ASSERT) {
 744                         assert(be_verify_schedule(birg) && "Schedule verification failed");
 745                         assert(be_verify_register_pressure(birg, cls, irg)
 746                                 && "Register pressure verification failed");
 747                 }
 748                 BE_TIMER_POP(t_verify);
 749
 750                 BE_TIMER_PUSH(t_ra_color);
 751                 be_straight_alloc_cls();
 752                 BE_TIMER_POP(t_ra_color);
 753
 754                 bitset_free(ignore_regs);
 755
 756                 /* TODO: dump intermediate results */
 757
 758                 stat_ev_ctx_pop("bestraight_cls");
 759         }
 760
 761         obstack_free(&obst, NULL);
 762 }
 763
 764 static be_ra_t be_ra_straight = {
 765         be_straight_alloc,
 766 };
 767
 768 void be_init_straight_alloc(void)
 769 {
 770         FIRM_DBG_REGISTER(dbg, "firm.be.straightalloc");
 771
 772         be_register_allocator("straight", &be_ra_straight);
 773 }
 774
 775 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_straight_alloc);