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

   1 /*
   2  * Copyright (C) 1995-2011 University of Karlsruhe.  All right reserved.
   3  *
   4  * This file is part of libFirm.
   5  *
   6  * This file may be distributed and/or modified under the terms of the
   7  * GNU General Public License version 2 as published by the Free Software
   8  * Foundation and appearing in the file LICENSE.GPL included in the
   9  * packaging of this file.
  10  *
  11  * Licensees holding valid libFirm Professional Edition licenses may use
  12  * this file in accordance with the libFirm Commercial License.
  13  * Agreement provided with the Software.
  14  *
  15  * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
  16  * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  17  * PURPOSE.
  18  */
  19
  20 /**
  21  * @file
  22  * @brief       Beladys spillalgorithm.
  23  * @author      Daniel Grund, Matthias Braun
  24  * @date        20.09.2005
  25  */
  26 #include "config.h"
  27
  28 #include <stdbool.h>
  29
  30 #include "obst.h"
  31 #include "irgraph.h"
  32 #include "irnode.h"
  33 #include "irmode.h"
  34 #include "irgwalk.h"
  35 #include "irloop.h"
  36 #include "iredges_t.h"
  37 #include "ircons_t.h"
  38 #include "irprintf.h"
  39 #include "irnodeset.h"
  40 #include "irtools.h"
  41 #include "statev_t.h"
  42 #include "util.h"
  43
  44 #include "beutil.h"
  45 #include "bearch.h"
  46 #include "beuses.h"
  47 #include "besched.h"
  48 #include "beirgmod.h"
  49 #include "belive_t.h"
  50 #include "benode.h"
  51 #include "bechordal_t.h"
  52 #include "bespill.h"
  53 #include "beloopana.h"
  54 #include "beirg.h"
  55 #include "bespillutil.h"
  56 #include "bemodule.h"
  57
  58 #define DBG_SPILL     1
  59 #define DBG_WSETS     2
  60 #define DBG_FIX       4
  61 #define DBG_DECIDE    8
  62 #define DBG_START    16
  63 #define DBG_SLOTS    32
  64 #define DBG_TRACE    64
  65 #define DBG_WORKSET 128
  66 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  67
  68 #define TIME_UNDEFINED 6666
  69
  70 /**
  71  * An association between a node and a point in time.
  72  */
  73 typedef struct loc_t {
  74         ir_node          *node;
  75         unsigned          time;     /**< A use time (see beuses.h). */
  76         bool              spilled;  /**< value was already spilled on this path */
  77 } loc_t;
  78
  79 typedef struct workset_t {
  80         unsigned len;     /**< current length */
  81         loc_t    vals[];  /**< array of the values/distances in this working set */
  82 } workset_t;
  83
  84 static struct obstack               obst;
  85 static const arch_register_class_t *cls;
  86 static const be_lv_t               *lv;
  87 static be_loopana_t                *loop_ana;
  88 static unsigned                     n_regs;
  89 static workset_t                   *ws;     /**< the main workset used while
  90                                                      processing a block. */
  91 static be_uses_t                   *uses;   /**< env for the next-use magic */
  92 static spill_env_t                 *senv;   /**< see bespill.h */
  93 static ir_node                    **blocklist;
  94
  95 static int                          move_spills      = true;
  96 static int                          respectloopdepth = true;
  97 static int                          improve_known_preds = true;
  98 /* factor to weight the different costs of reloading/rematerializing a node
  99    (see bespill.h be_get_reload_costs_no_weight) */
 100 static int                          remat_bonus      = 10;
 101
 102 static const lc_opt_table_entry_t options[] = {
 103         LC_OPT_ENT_BOOL   ("movespills", "try to move spills out of loops", &move_spills),
 104         LC_OPT_ENT_BOOL   ("respectloopdepth", "outermost loop cutting", &respectloopdepth),
 105         LC_OPT_ENT_BOOL   ("improveknownpreds", "known preds cutting", &improve_known_preds),
 106         LC_OPT_ENT_INT    ("rematbonus", "give bonus to rematerialisable nodes", &remat_bonus),
 107         LC_OPT_LAST
 108 };
 109
 110 /**
 111  * Alloc a new workset on obstack @p ob with maximum size @p max
 112  */
 113 static workset_t *new_workset(void)
 114 {
 115         return OALLOCFZ(&obst, workset_t, vals, n_regs);
 116 }
 117
 118 /**
 119  * Alloc a new instance on obstack and make it equal to @param workset
 120  */
 121 static workset_t *workset_clone(workset_t *workset)
 122 {
 123         workset_t *res = OALLOCF(&obst, workset_t, vals, n_regs);
 124         memcpy(res, workset, sizeof(*res) + n_regs * sizeof(res->vals[0]));
 125         return res;
 126 }
 127
 128 /**
 129  * Copy workset @param src to @param tgt
 130  */
 131 static void workset_copy(workset_t *dest, const workset_t *src)
 132 {
 133         size_t size = sizeof(*src) + n_regs * sizeof(src->vals[0]);
 134         memcpy(dest, src, size);
 135 }
 136
 137 /**
 138  * Overwrites the current content array of @param ws with the
 139  * @param count locations given at memory @param locs.
 140  * Set the length of @param ws to count.
 141  */
 142 static void workset_bulk_fill(workset_t *workset, int count, const loc_t *locs)
 143 {
 144         workset->len = count;
 145         memcpy(&(workset->vals[0]), locs, count * sizeof(locs[0]));
 146 }
 147
 148 /**
 149  * Inserts the value @p val into the workset, iff it is not
 150  * already contained. The workset must not be full.
 151  */
 152 static void workset_insert(workset_t *workset, ir_node *val, bool spilled)
 153 {
 154         loc_t    *loc;
 155         unsigned  i;
 156         /* check for current regclass */
 157         assert(arch_irn_consider_in_reg_alloc(cls, val));
 158
 159         /* check if val is already contained */
 160         for (i = 0; i < workset->len; ++i) {
 161                 loc = &workset->vals[i];
 162                 if (loc->node == val) {
 163                         if (spilled) {
 164                                 loc->spilled = true;
 165                         }
 166                         return;
 167                 }
 168         }
 169
 170         /* insert val */
 171         assert(workset->len < n_regs && "Workset already full!");
 172         loc           = &workset->vals[workset->len];
 173         loc->node     = val;
 174         loc->spilled  = spilled;
 175         loc->time     = TIME_UNDEFINED;
 176         workset->len++;
 177 }
 178
 179 /**
 180  * Removes all entries from this workset
 181  */
 182 static void workset_clear(workset_t *workset)
 183 {
 184         workset->len = 0;
 185 }
 186
 187 /**
 188  * Removes the value @p val from the workset if present.
 189  */
 190 static void workset_remove(workset_t *workset, ir_node *val)
 191 {
 192         unsigned i;
 193         for (i = 0; i < workset->len; ++i) {
 194                 if (workset->vals[i].node == val) {
 195                         workset->vals[i] = workset->vals[--workset->len];
 196                         return;
 197                 }
 198         }
 199 }
 200
 201 static const loc_t *workset_contains(const workset_t *ws, const ir_node *val)
 202 {
 203         unsigned i;
 204         for (i = 0; i < ws->len; ++i) {
 205                 if (ws->vals[i].node == val)
 206                         return &ws->vals[i];
 207         }
 208
 209         return NULL;
 210 }
 211
 212 static int loc_compare(const void *a, const void *b)
 213 {
 214         const loc_t   *p  = ((const loc_t*) a);
 215         const loc_t   *q  = ((const loc_t*) b);
 216         const unsigned pt = p->time;
 217         const unsigned qt = q->time;
 218
 219         if (pt < qt)
 220                 return -1;
 221         if (pt > qt)
 222                 return 1;
 223
 224         return get_irn_node_nr(p->node) - get_irn_node_nr(q->node);
 225 }
 226
 227 static void workset_sort(workset_t *workset)
 228 {
 229         qsort(workset->vals, workset->len, sizeof(workset->vals[0]), loc_compare);
 230 }
 231
 232 static inline unsigned workset_get_time(const workset_t *workset, unsigned idx)
 233 {
 234         return workset->vals[idx].time;
 235 }
 236
 237 static inline void workset_set_time(workset_t *workset, unsigned idx,
 238                                     unsigned time)
 239 {
 240         workset->vals[idx].time = time;
 241 }
 242
 243 static inline unsigned workset_get_length(const workset_t *workset)
 244 {
 245         return workset->len;
 246 }
 247
 248 static inline void workset_set_length(workset_t *workset, unsigned len)
 249 {
 250         workset->len = len;
 251 }
 252
 253 static inline ir_node *workset_get_val(const workset_t *workset, unsigned idx)
 254 {
 255         return workset->vals[idx].node;
 256 }
 257
 258 /**
 259  * Iterates over all values in the working set.
 260  * @p ws The workset to iterate
 261  * @p v  A variable to put the current value in
 262  * @p i  An integer for internal use
 263  */
 264 #define workset_foreach(ws, v, i) \
 265         for (i=0; v=(i < ws->len) ? ws->vals[i].node : NULL, i < ws->len; ++i)
 266
 267 typedef struct block_info_t {
 268         workset_t *start_workset;
 269         workset_t *end_workset;
 270 } block_info_t;
 271
 272 static block_info_t *new_block_info(void)
 273 {
 274         return OALLOCZ(&obst, block_info_t);
 275 }
 276
 277 static inline block_info_t *get_block_info(const ir_node *block)
 278 {
 279         return (block_info_t*)get_irn_link(block);
 280 }
 281
 282 static inline void set_block_info(ir_node *block, block_info_t *info)
 283 {
 284         set_irn_link(block, info);
 285 }
 286
 287 /**
 288  * @return The distance to the next use or 0 if irn has dont_spill flag set
 289  */
 290 static unsigned get_distance(ir_node *from, const ir_node *def, int skip_from_uses)
 291 {
 292         be_next_use_t use;
 293         unsigned      costs;
 294         unsigned      time;
 295
 296         assert(!arch_irn_is_ignore(def));
 297
 298         use  = be_get_next_use(uses, from, def, skip_from_uses);
 299         time = use.time;
 300         if (USES_IS_INFINITE(time))
 301                 return USES_INFINITY;
 302
 303         /* We have to keep nonspillable nodes in the workingset */
 304         if (arch_get_irn_flags(skip_Proj_const(def)) & arch_irn_flags_dont_spill)
 305                 return 0;
 306
 307         /* give some bonus to rematerialisable nodes */
 308         if (remat_bonus > 0) {
 309                 costs = be_get_reload_costs_no_weight(senv, def, use.before);
 310                 assert(costs * remat_bonus < 1000);
 311                 time  += 1000 - (costs * remat_bonus);
 312         }
 313
 314         return time;
 315 }
 316
 317 /**
 318  * Performs the actions necessary to grant the request that:
 319  * - new_vals can be held in registers
 320  * - as few as possible other values are disposed
 321  * - the worst values get disposed
 322  *
 323  * @p is_usage indicates that the values in new_vals are used (not defined)
 324  * In this case reloads must be performed
 325  */
 326 static void displace(workset_t *const new_vals, int const is_usage, ir_node *const instr)
 327 {
 328         ir_node **to_insert = ALLOCAN(ir_node*, n_regs);
 329         bool     *spilled   = ALLOCAN(bool,     n_regs);
 330         ir_node  *val;
 331         int       i;
 332         int       len;
 333         int       spills_needed;
 334         int       demand;
 335         unsigned  iter;
 336
 337         /* 1. Identify the number of needed slots and the values to reload */
 338         demand = 0;
 339         workset_foreach(new_vals, val, iter) {
 340                 bool reloaded = false;
 341
 342                 if (! workset_contains(ws, val)) {
 343                         DB((dbg, DBG_DECIDE, "    insert %+F\n", val));
 344                         if (is_usage) {
 345                                 DB((dbg, DBG_SPILL, "Reload %+F before %+F\n", val, instr));
 346                                 be_add_reload(senv, val, instr, cls, 1);
 347                                 reloaded = true;
 348                         }
 349                 } else {
 350                         DB((dbg, DBG_DECIDE, "    %+F already in workset\n", val));
 351                         assert(is_usage);
 352                         /* remove the value from the current workset so it is not accidently
 353                          * spilled */
 354                         workset_remove(ws, val);
 355                 }
 356                 spilled[demand]   = reloaded;
 357                 to_insert[demand] = val;
 358                 ++demand;
 359         }
 360
 361         /* 2. Make room for at least 'demand' slots */
 362         len           = workset_get_length(ws);
 363         spills_needed = len + demand - n_regs;
 364         assert(spills_needed <= len);
 365
 366         /* Only make more free room if we do not have enough */
 367         if (spills_needed > 0) {
 368                 DB((dbg, DBG_DECIDE, "    disposing %d values\n", spills_needed));
 369
 370                 /* calculate current next-use distance for live values */
 371                 for (i = 0; i < len; ++i) {
 372                         ir_node  *val  = workset_get_val(ws, i);
 373                         unsigned  dist = get_distance(instr, val, !is_usage);
 374                         workset_set_time(ws, i, dist);
 375                 }
 376
 377                 /* sort entries by increasing nextuse-distance*/
 378                 workset_sort(ws);
 379
 380                 for (i = len - spills_needed; i < len; ++i) {
 381                         ir_node *val = ws->vals[i].node;
 382
 383                         DB((dbg, DBG_DECIDE, "    disposing node %+F (%u)\n", val,
 384                              workset_get_time(ws, i)));
 385
 386                         if (move_spills) {
 387                                 if (!USES_IS_INFINITE(ws->vals[i].time)
 388                                                 && !ws->vals[i].spilled) {
 389                                         ir_node *after_pos = sched_prev(instr);
 390                                         DB((dbg, DBG_DECIDE, "Spill %+F after node %+F\n", val,
 391                                                 after_pos));
 392                                         be_add_spill(senv, val, after_pos);
 393                                 }
 394                         }
 395                 }
 396
 397                 /* kill the last 'demand' entries in the array */
 398                 workset_set_length(ws, len - spills_needed);
 399         }
 400
 401         /* 3. Insert the new values into the workset */
 402         for (i = 0; i < demand; ++i) {
 403                 ir_node *val = to_insert[i];
 404
 405                 workset_insert(ws, val, spilled[i]);
 406         }
 407 }
 408
 409 enum {
 410         AVAILABLE_EVERYWHERE,
 411         AVAILABLE_NOWHERE,
 412         AVAILABLE_PARTLY,
 413         AVAILABLE_UNKNOWN
 414 };
 415
 416 static unsigned available_in_all_preds(workset_t* const* pred_worksets,
 417                                        size_t n_pred_worksets,
 418                                        const ir_node *value, bool is_local_phi)
 419 {
 420         size_t i;
 421         bool   avail_everywhere = true;
 422         bool   avail_nowhere    = true;
 423
 424         assert(n_pred_worksets > 0);
 425
 426         /* value available in all preds? */
 427         for (i = 0; i < n_pred_worksets; ++i) {
 428                 bool             found     = false;
 429                 const workset_t *p_workset = pred_worksets[i];
 430                 int              p_len     = workset_get_length(p_workset);
 431                 int              p_i;
 432                 const ir_node   *l_value;
 433
 434                 if (is_local_phi) {
 435                         assert(is_Phi(value));
 436                         l_value = get_irn_n(value, i);
 437                 } else {
 438                         l_value = value;
 439                 }
 440
 441                 for (p_i = 0; p_i < p_len; ++p_i) {
 442                         const loc_t *p_l = &p_workset->vals[p_i];
 443                         if (p_l->node != l_value)
 444                                 continue;
 445
 446                         found = true;
 447                         break;
 448                 }
 449
 450                 if (found) {
 451                         avail_nowhere = false;
 452                 } else {
 453                         avail_everywhere = false;
 454                 }
 455         }
 456
 457         if (avail_everywhere) {
 458                 assert(!avail_nowhere);
 459                 return AVAILABLE_EVERYWHERE;
 460         } else if (avail_nowhere) {
 461                 return AVAILABLE_NOWHERE;
 462         } else {
 463                 return AVAILABLE_PARTLY;
 464         }
 465 }
 466
 467 /** Decides whether a specific node should be in the start workset or not
 468  *
 469  * @param env      belady environment
 470  * @param first
 471  * @param node     the node to test
 472  * @param loop     the loop of the node
 473  */
 474 static loc_t to_take_or_not_to_take(ir_node* first, ir_node *node,
 475                                     ir_loop *loop, unsigned available)
 476 {
 477         be_next_use_t next_use;
 478         loc_t         loc;
 479
 480         loc.time    = USES_INFINITY;
 481         loc.node    = node;
 482         loc.spilled = false;
 483
 484         /* We have to keep nonspillable nodes in the workingset */
 485         if (arch_get_irn_flags(skip_Proj_const(node)) & arch_irn_flags_dont_spill) {
 486                 loc.time = 0;
 487                 DB((dbg, DBG_START, "    %+F taken (dontspill node)\n", node, loc.time));
 488                 return loc;
 489         }
 490
 491         next_use = be_get_next_use(uses, first, node, 0);
 492         if (USES_IS_INFINITE(next_use.time)) {
 493                 /* the nodes marked as live in shouldn't be dead, so it must be a phi */
 494                 assert(is_Phi(node));
 495                 loc.time = USES_INFINITY;
 496                 DB((dbg, DBG_START, "    %+F not taken (dead)\n", node));
 497                 return loc;
 498         }
 499
 500         loc.time = next_use.time;
 501
 502         if (improve_known_preds) {
 503                 if (available == AVAILABLE_EVERYWHERE) {
 504                         DB((dbg, DBG_START, "    %+F taken (%u, live in all preds)\n",
 505                             node, loc.time));
 506                         return loc;
 507                 } else if (available == AVAILABLE_NOWHERE) {
 508                         DB((dbg, DBG_START, "    %+F not taken (%u, live in no pred)\n",
 509                             node, loc.time));
 510                         loc.time = USES_INFINITY;
 511                         return loc;
 512                 }
 513         }
 514
 515         if (!respectloopdepth || next_use.outermost_loop >= get_loop_depth(loop)) {
 516                 DB((dbg, DBG_START, "    %+F taken (%u, loop %d)\n", node, loc.time,
 517                     next_use.outermost_loop));
 518         } else {
 519                 loc.time = USES_PENDING;
 520                 DB((dbg, DBG_START, "    %+F delayed (outerdepth %d < loopdepth %d)\n",
 521                     node, next_use.outermost_loop, get_loop_depth(loop)));
 522         }
 523
 524         return loc;
 525 }
 526
 527 /**
 528  * Computes the start-workset for a block with multiple predecessors. We assume
 529  * that at least 1 of the predeccesors is a back-edge which means we're at the
 530  * beginning of a loop. We try to reload as much values as possible now so they
 531  * don't get reloaded inside the loop.
 532  */
 533 static void decide_start_workset(const ir_node *block)
 534 {
 535         ir_loop    *loop = get_irn_loop(block);
 536         ir_node    *first;
 537         loc_t       loc;
 538         loc_t      *starters;
 539         loc_t      *delayed;
 540         unsigned    len;
 541         unsigned    i;
 542         unsigned    ws_count;
 543         int         free_slots, free_pressure_slots;
 544         unsigned    pressure;
 545         int         arity;
 546         workset_t **pred_worksets;
 547         bool        all_preds_known;
 548
 549         /* check predecessors */
 550         arity           = get_irn_arity(block);
 551         pred_worksets   = ALLOCAN(workset_t*, arity);
 552         all_preds_known = true;
 553         for (int in = 0; in < arity; ++in) {
 554                 ir_node      *pred_block = get_Block_cfgpred_block(block, in);
 555                 block_info_t *pred_info  = get_block_info(pred_block);
 556
 557                 if (pred_info == NULL) {
 558                         pred_worksets[in] = NULL;
 559                         all_preds_known   = false;
 560                 } else {
 561                         pred_worksets[in] = pred_info->end_workset;
 562                 }
 563         }
 564
 565         /* Collect all values living at start of block */
 566         starters = NEW_ARR_F(loc_t, 0);
 567         delayed  = NEW_ARR_F(loc_t, 0);
 568
 569         DB((dbg, DBG_START, "Living at start of %+F:\n", block));
 570         first = sched_first(block);
 571
 572         /* check all Phis first */
 573         sched_foreach(block, node) {
 574                 unsigned available;
 575
 576                 if (! is_Phi(node))
 577                         break;
 578                 if (!arch_irn_consider_in_reg_alloc(cls, node))
 579                         continue;
 580
 581                 if (all_preds_known) {
 582                         available = available_in_all_preds(pred_worksets, arity, node, true);
 583                 } else {
 584                         available = AVAILABLE_UNKNOWN;
 585                 }
 586
 587                 loc = to_take_or_not_to_take(first, node, loop, available);
 588
 589                 if (! USES_IS_INFINITE(loc.time)) {
 590                         if (USES_IS_PENDING(loc.time))
 591                                 ARR_APP1(loc_t, delayed, loc);
 592                         else
 593                                 ARR_APP1(loc_t, starters, loc);
 594                 } else {
 595                         be_spill_phi(senv, node);
 596                 }
 597         }
 598
 599         /* check all Live-Ins */
 600         be_lv_foreach_cls(lv, block, be_lv_state_in, cls, node) {
 601                 unsigned available;
 602                 if (all_preds_known) {
 603                         available = available_in_all_preds(pred_worksets, arity, node, false);
 604                 } else {
 605                         available = AVAILABLE_UNKNOWN;
 606                 }
 607
 608                 loc = to_take_or_not_to_take(first, node, loop, available);
 609
 610                 if (! USES_IS_INFINITE(loc.time)) {
 611                         if (USES_IS_PENDING(loc.time))
 612                                 ARR_APP1(loc_t, delayed, loc);
 613                         else
 614                                 ARR_APP1(loc_t, starters, loc);
 615                 }
 616         }
 617
 618         pressure            = be_get_loop_pressure(loop_ana, cls, loop);
 619         assert(ARR_LEN(delayed) <= pressure);
 620         free_slots          = n_regs - ARR_LEN(starters);
 621         free_pressure_slots = n_regs - (pressure - ARR_LEN(delayed));
 622         free_slots          = MIN(free_slots, free_pressure_slots);
 623
 624         /* so far we only put nodes into the starters list that are used inside
 625          * the loop. If register pressure in the loop is low then we can take some
 626          * values and let them live through the loop */
 627         DB((dbg, DBG_START, "Loop pressure %d, taking %d delayed vals\n",
 628             pressure, free_slots));
 629         if (free_slots > 0) {
 630                 size_t i;
 631
 632                 qsort(delayed, ARR_LEN(delayed), sizeof(delayed[0]), loc_compare);
 633
 634                 for (i = 0; i < ARR_LEN(delayed) && free_slots > 0; ++i) {
 635                         int    p, arity;
 636                         loc_t *loc = & delayed[i];
 637
 638                         if (!is_Phi(loc->node)) {
 639                                 /* don't use values which are dead in a known predecessors
 640                                  * to not induce unnecessary reloads */
 641                                 arity = get_irn_arity(block);
 642                                 for (p = 0; p < arity; ++p) {
 643                                         ir_node      *pred_block = get_Block_cfgpred_block(block, p);
 644                                         block_info_t *pred_info  = get_block_info(pred_block);
 645
 646                                         if (pred_info == NULL)
 647                                                 continue;
 648
 649                                         if (!workset_contains(pred_info->end_workset, loc->node)) {
 650                                                 DB((dbg, DBG_START,
 651                                                         "    delayed %+F not live at pred %+F\n", loc->node,
 652                                                         pred_block));
 653                                                 goto skip_delayed;
 654                                         }
 655                                 }
 656                         }
 657
 658                         DB((dbg, DBG_START, "    delayed %+F taken\n", loc->node));
 659                         ARR_APP1(loc_t, starters, *loc);
 660                         loc->node = NULL;
 661                         --free_slots;
 662                 skip_delayed:
 663                         ;
 664                 }
 665         }
 666
 667         /* spill phis (the actual phis not just their values) that are in this block
 668          * but not in the start workset */
 669         len = ARR_LEN(delayed);
 670         for (i = 0; i < len; ++i) {
 671                 ir_node *node = delayed[i].node;
 672                 if (node == NULL || !is_Phi(node) || get_nodes_block(node) != block)
 673                         continue;
 674
 675                 DB((dbg, DBG_START, "    spilling delayed phi %+F\n", node));
 676                 be_spill_phi(senv, node);
 677         }
 678         DEL_ARR_F(delayed);
 679
 680         /* Sort start values by first use */
 681         qsort(starters, ARR_LEN(starters), sizeof(starters[0]), loc_compare);
 682
 683         /* Copy the best ones from starters to start workset */
 684         ws_count = MIN((unsigned) ARR_LEN(starters), n_regs);
 685         workset_clear(ws);
 686         workset_bulk_fill(ws, ws_count, starters);
 687
 688         /* spill phis (the actual phis not just their values) that are in this block
 689          * but not in the start workset */
 690         len = ARR_LEN(starters);
 691         for (i = ws_count; i < len; ++i) {
 692                 ir_node *node = starters[i].node;
 693                 if (! is_Phi(node) || get_nodes_block(node) != block)
 694                         continue;
 695
 696                 DB((dbg, DBG_START, "    spilling phi %+F\n", node));
 697                 be_spill_phi(senv, node);
 698         }
 699
 700         DEL_ARR_F(starters);
 701
 702         /* determine spill status of the values: If there's 1 pred block (which
 703          * is no backedge) where the value is spilled then we must set it to
 704          * spilled here. */
 705         for (i = 0; i < ws_count; ++i) {
 706                 loc_t   *loc     = &ws->vals[i];
 707                 ir_node *value   = loc->node;
 708                 bool     spilled;
 709                 int      n;
 710
 711                 /* phis from this block aren't spilled */
 712                 if (get_nodes_block(value) == block) {
 713                         assert(is_Phi(value));
 714                         loc->spilled = false;
 715                         continue;
 716                 }
 717
 718                 /* determine if value was spilled on any predecessor */
 719                 spilled = false;
 720                 for (n = 0; n < arity; ++n) {
 721                         workset_t *pred_workset = pred_worksets[n];
 722                         int        p_len;
 723                         int        p;
 724
 725                         if (pred_workset == NULL)
 726                                 continue;
 727
 728                         p_len = workset_get_length(pred_workset);
 729                         for (p = 0; p < p_len; ++p) {
 730                                 loc_t *l = &pred_workset->vals[p];
 731
 732                                 if (l->node != value)
 733                                         continue;
 734
 735                                 if (l->spilled) {
 736                                         spilled = true;
 737                                 }
 738                                 break;
 739                         }
 740                 }
 741
 742                 loc->spilled = spilled;
 743         }
 744 }
 745
 746 /**
 747  * For the given block @p block, decide for each values
 748  * whether it is used from a register or is reloaded
 749  * before the use.
 750  */
 751 static void process_block(ir_node *block)
 752 {
 753         workset_t    *new_vals;
 754         unsigned      iter;
 755         block_info_t *block_info;
 756         int           arity;
 757
 758         /* no need to process a block twice */
 759         assert(get_block_info(block) == NULL);
 760
 761         /* construct start workset */
 762         arity = get_Block_n_cfgpreds(block);
 763         if (arity == 0) {
 764                 /* no predecessor -> empty set */
 765                 workset_clear(ws);
 766         } else if (arity == 1) {
 767                 /* one predecessor, copy its end workset */
 768                 ir_node      *pred_block = get_Block_cfgpred_block(block, 0);
 769                 block_info_t *pred_info  = get_block_info(pred_block);
 770
 771                 assert(pred_info != NULL);
 772                 workset_copy(ws, pred_info->end_workset);
 773         } else {
 774                 /* multiple predecessors, do more advanced magic :) */
 775                 decide_start_workset(block);
 776         }
 777
 778         DB((dbg, DBG_DECIDE, "\n"));
 779         DB((dbg, DBG_DECIDE, "Decide for %+F\n", block));
 780
 781         block_info = new_block_info();
 782         set_block_info(block, block_info);
 783
 784         DB((dbg, DBG_WSETS, "Start workset for %+F:\n", block));
 785         {
 786                 ir_node *irn;
 787                 workset_foreach(ws, irn, iter) {
 788                         DB((dbg, DBG_WSETS, "  %+F (%u)\n", irn, workset_get_time(ws, iter)));
 789                 }
 790         }
 791
 792         block_info->start_workset = workset_clone(ws);
 793
 794         /* process the block from start to end */
 795         DB((dbg, DBG_WSETS, "Processing...\n"));
 796         /* TODO: this leaks (into the obstack)... */
 797         new_vals = new_workset();
 798
 799         sched_foreach(block, irn) {
 800                 assert(workset_get_length(ws) <= n_regs);
 801
 802                 /* Phis are no real instr (see insert_starters()) */
 803                 if (is_Phi(irn)) {
 804                         continue;
 805                 }
 806                 DB((dbg, DBG_DECIDE, "  ...%+F\n", irn));
 807
 808                 /* allocate all values _used_ by this instruction */
 809                 workset_clear(new_vals);
 810                 be_foreach_use(irn, cls, in_req_, in, in_req,
 811                         /* (note that "spilled" is irrelevant here) */
 812                         workset_insert(new_vals, in, false);
 813                 );
 814                 displace(new_vals, 1, irn);
 815
 816                 /* allocate all values _defined_ by this instruction */
 817                 workset_clear(new_vals);
 818                 be_foreach_definition(irn, cls, value, req,
 819                         assert(req->width == 1);
 820                         workset_insert(new_vals, value, false);
 821                 );
 822                 displace(new_vals, 0, irn);
 823         }
 824
 825         /* Remember end-workset for this block */
 826         block_info->end_workset = workset_clone(ws);
 827         DB((dbg, DBG_WSETS, "End workset for %+F:\n", block));
 828         {
 829                 ir_node *irn;
 830                 workset_foreach(ws, irn, iter)
 831                         DB((dbg, DBG_WSETS, "  %+F (%u)\n", irn, workset_get_time(ws, iter)));
 832         }
 833 }
 834
 835 /**
 836  * 'decide' is block-local and makes assumptions
 837  * about the set of live-ins. Thus we must adapt the
 838  * live-outs to the live-ins at each block-border.
 839  */
 840 static void fix_block_borders(ir_node *block, void *data)
 841 {
 842         workset_t *start_workset;
 843         int        arity;
 844         int        i;
 845         unsigned   iter;
 846         (void) data;
 847
 848         DB((dbg, DBG_FIX, "\n"));
 849         DB((dbg, DBG_FIX, "Fixing %+F\n", block));
 850
 851         arity = get_irn_arity(block);
 852         /* can happen for endless loops */
 853         if (arity == 0)
 854                 return;
 855
 856         start_workset = get_block_info(block)->start_workset;
 857
 858         /* process all pred blocks */
 859         for (i = 0; i < arity; ++i) {
 860                 ir_node   *pred = get_Block_cfgpred_block(block, i);
 861                 workset_t *pred_end_workset = get_block_info(pred)->end_workset;
 862                 ir_node   *node;
 863
 864                 DB((dbg, DBG_FIX, "  Pred %+F\n", pred));
 865
 866                 /* spill all values not used anymore */
 867                 workset_foreach(pred_end_workset, node, iter) {
 868                         ir_node *n2;
 869                         unsigned iter2;
 870                         bool     found = false;
 871                         workset_foreach(start_workset, n2, iter2) {
 872                                 if (n2 == node) {
 873                                         found = true;
 874                                         break;
 875                                 }
 876                                 /* note that we do not look at phi inputs, becuase the values
 877                                  * will be either live-end and need no spill or
 878                                  * they have other users in which must be somewhere else in the
 879                                  * workset */
 880                         }
 881
 882                         if (found)
 883                                 continue;
 884
 885                         if (move_spills && be_is_live_in(lv, block, node)
 886                                         && !pred_end_workset->vals[iter].spilled) {
 887                                 ir_node *insert_point;
 888                                 if (arity > 1) {
 889                                         insert_point = be_get_end_of_block_insertion_point(pred);
 890                                         insert_point = sched_prev(insert_point);
 891                                 } else {
 892                                         insert_point = block;
 893                                 }
 894                                 DB((dbg, DBG_SPILL, "Spill %+F after %+F\n", node,
 895                                      insert_point));
 896                                 be_add_spill(senv, node, insert_point);
 897                         }
 898                 }
 899
 900                 /* reload missing values in predecessors, add missing spills */
 901                 workset_foreach(start_workset, node, iter) {
 902                         const loc_t *l    = &start_workset->vals[iter];
 903                         const loc_t *pred_loc;
 904
 905                         /* if node is a phi of the current block we reload
 906                          * the corresponding argument, else node itself */
 907                         if (is_Phi(node) && get_nodes_block(node) == block) {
 908                                 node = get_irn_n(node, i);
 909                                 assert(!l->spilled);
 910
 911                                 /* we might have unknowns as argument for the phi */
 912                                 if (!arch_irn_consider_in_reg_alloc(cls, node))
 913                                         continue;
 914                         }
 915
 916                         /* check if node is in a register at end of pred */
 917                         pred_loc = workset_contains(pred_end_workset, node);
 918                         if (pred_loc != NULL) {
 919                                 /* we might have to spill value on this path */
 920                                 if (move_spills && !pred_loc->spilled && l->spilled) {
 921                                         ir_node *insert_point
 922                                                 = be_get_end_of_block_insertion_point(pred);
 923                                         insert_point = sched_prev(insert_point);
 924                                         DB((dbg, DBG_SPILL, "Spill %+F after %+F\n", node,
 925                                             insert_point));
 926                                         be_add_spill(senv, node, insert_point);
 927                                 }
 928                         } else {
 929                                 /* node is not in register at the end of pred -> reload it */
 930                                 DB((dbg, DBG_FIX, "    reload %+F\n", node));
 931                                 DB((dbg, DBG_SPILL, "Reload %+F before %+F,%d\n", node, block, i));
 932                                 be_add_reload_on_edge(senv, node, block, i, cls, 1);
 933                         }
 934                 }
 935         }
 936 }
 937
 938 static void be_spill_belady(ir_graph *irg, const arch_register_class_t *rcls)
 939 {
 940         int i;
 941
 942         be_assure_live_sets(irg);
 943
 944         stat_ev_tim_push();
 945         assure_loopinfo(irg);
 946         stat_ev_tim_pop("belady_time_backedges");
 947
 948         stat_ev_tim_push();
 949         be_clear_links(irg);
 950         stat_ev_tim_pop("belady_time_clear_links");
 951
 952         ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
 953
 954         /* init belady env */
 955         stat_ev_tim_push();
 956         obstack_init(&obst);
 957         cls       = rcls;
 958         lv        = be_get_irg_liveness(irg);
 959         n_regs    = be_get_n_allocatable_regs(irg, cls);
 960         ws        = new_workset();
 961         uses      = be_begin_uses(irg, lv);
 962         loop_ana  = be_new_loop_pressure(irg, cls);
 963         senv      = be_new_spill_env(irg);
 964         blocklist = be_get_cfgpostorder(irg);
 965         stat_ev_tim_pop("belady_time_init");
 966
 967         stat_ev_tim_push();
 968         /* walk blocks in reverse postorder */
 969         for (i = ARR_LEN(blocklist) - 1; i >= 0; --i) {
 970                 process_block(blocklist[i]);
 971         }
 972         DEL_ARR_F(blocklist);
 973         stat_ev_tim_pop("belady_time_belady");
 974
 975         stat_ev_tim_push();
 976         /* belady was block-local, fix the global flow by adding reloads on the
 977          * edges */
 978         irg_block_walk_graph(irg, fix_block_borders, NULL, NULL);
 979         stat_ev_tim_pop("belady_time_fix_borders");
 980
 981         ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
 982
 983         /* Insert spill/reload nodes into the graph and fix usages */
 984         be_insert_spills_reloads(senv);
 985
 986         /* clean up */
 987         be_delete_spill_env(senv);
 988         be_end_uses(uses);
 989         be_free_loop_pressure(loop_ana);
 990         obstack_free(&obst, NULL);
 991 }
 992
 993 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady)
 994 void be_init_spillbelady(void)
 995 {
 996         static be_spiller_t belady_spiller = {
 997                 be_spill_belady
 998         };
 999         lc_opt_entry_t *be_grp       = lc_opt_get_grp(firm_opt_get_root(), "be");
1000         lc_opt_entry_t *belady_group = lc_opt_get_grp(be_grp, "belady");
1001         lc_opt_add_table(belady_group, options);
1002
1003         be_register_spiller("belady", &belady_spiller);
1004         FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
1005 }