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 "irprintf_t.h"
  32 #include "irgraph.h"
  33 #include "irnode.h"
  34 #include "irmode.h"
  35 #include "irgwalk.h"
  36 #include "irloop.h"
  37 #include "iredges_t.h"
  38 #include "ircons_t.h"
  39 #include "irprintf.h"
  40 #include "irnodeset.h"
  41 #include "irtools.h"
  42 #include "statev_t.h"
  43 #include "util.h"
  44
  45 #include "beutil.h"
  46 #include "bearch.h"
  47 #include "beuses.h"
  48 #include "besched.h"
  49 #include "beirgmod.h"
  50 #include "belive_t.h"
  51 #include "benode.h"
  52 #include "bechordal_t.h"
  53 #include "bespill.h"
  54 #include "beloopana.h"
  55 #include "beirg.h"
  56 #include "bespillutil.h"
  57 #include "bemodule.h"
  58
  59 #define DBG_SPILL     1
  60 #define DBG_WSETS     2
  61 #define DBG_FIX       4
  62 #define DBG_DECIDE    8
  63 #define DBG_START    16
  64 #define DBG_SLOTS    32
  65 #define DBG_TRACE    64
  66 #define DBG_WORKSET 128
  67 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  68
  69 #define TIME_UNDEFINED 6666
  70
  71 /**
  72  * An association between a node and a point in time.
  73  */
  74 typedef struct loc_t {
  75         ir_node          *node;
  76         unsigned          time;     /**< A use time (see beuses.h). */
  77         bool              spilled;  /**< value was already spilled on this path */
  78 } loc_t;
  79
  80 typedef struct workset_t {
  81         unsigned len;     /**< current length */
  82         loc_t    vals[];  /**< array of the values/distances in this working set */
  83 } workset_t;
  84
  85 static struct obstack               obst;
  86 static const arch_register_class_t *cls;
  87 static const be_lv_t               *lv;
  88 static be_loopana_t                *loop_ana;
  89 static unsigned                     n_regs;
  90 static workset_t                   *ws;     /**< the main workset used while
  91                                                      processing a block. */
  92 static be_uses_t                   *uses;   /**< env for the next-use magic */
  93 static ir_node                     *instr;  /**< current instruction */
  94 static spill_env_t                 *senv;   /**< see bespill.h */
  95 static ir_node                    **blocklist;
  96
  97 static int                          move_spills      = true;
  98 static int                          respectloopdepth = true;
  99 static int                          improve_known_preds = true;
 100 /* factor to weight the different costs of reloading/rematerializing a node
 101    (see bespill.h be_get_reload_costs_no_weight) */
 102 static int                          remat_bonus      = 10;
 103
 104 static const lc_opt_table_entry_t options[] = {
 105         LC_OPT_ENT_BOOL   ("movespills", "try to move spills out of loops", &move_spills),
 106         LC_OPT_ENT_BOOL   ("respectloopdepth", "outermost loop cutting", &respectloopdepth),
 107         LC_OPT_ENT_BOOL   ("improveknownpreds", "known preds cutting", &improve_known_preds),
 108         LC_OPT_ENT_INT    ("rematbonus", "give bonus to rematerialisable nodes", &remat_bonus),
 109         LC_OPT_LAST
 110 };
 111
 112 /**
 113  * Alloc a new workset on obstack @p ob with maximum size @p max
 114  */
 115 static workset_t *new_workset(void)
 116 {
 117         return OALLOCFZ(&obst, workset_t, vals, n_regs);
 118 }
 119
 120 /**
 121  * Alloc a new instance on obstack and make it equal to @param workset
 122  */
 123 static workset_t *workset_clone(workset_t *workset)
 124 {
 125         workset_t *res = OALLOCF(&obst, workset_t, vals, n_regs);
 126         memcpy(res, workset, sizeof(*res) + n_regs * sizeof(res->vals[0]));
 127         return res;
 128 }
 129
 130 /**
 131  * Copy workset @param src to @param tgt
 132  */
 133 static void workset_copy(workset_t *dest, const workset_t *src)
 134 {
 135         size_t size = sizeof(*src) + n_regs * sizeof(src->vals[0]);
 136         memcpy(dest, src, size);
 137 }
 138
 139 /**
 140  * Overwrites the current content array of @param ws with the
 141  * @param count locations given at memory @param locs.
 142  * Set the length of @param ws to count.
 143  */
 144 static void workset_bulk_fill(workset_t *workset, int count, const loc_t *locs)
 145 {
 146         workset->len = count;
 147         memcpy(&(workset->vals[0]), locs, count * sizeof(locs[0]));
 148 }
 149
 150 /**
 151  * Inserts the value @p val into the workset, iff it is not
 152  * already contained. The workset must not be full.
 153  */
 154 static void workset_insert(workset_t *workset, ir_node *val, bool spilled)
 155 {
 156         loc_t    *loc;
 157         unsigned  i;
 158         /* check for current regclass */
 159         assert(arch_irn_consider_in_reg_alloc(cls, val));
 160
 161         /* check if val is already contained */
 162         for (i = 0; i < workset->len; ++i) {
 163                 loc = &workset->vals[i];
 164                 if (loc->node == val) {
 165                         if (spilled) {
 166                                 loc->spilled = true;
 167                         }
 168                         return;
 169                 }
 170         }
 171
 172         /* insert val */
 173         assert(workset->len < n_regs && "Workset already full!");
 174         loc           = &workset->vals[workset->len];
 175         loc->node     = val;
 176         loc->spilled  = spilled;
 177         loc->time     = TIME_UNDEFINED;
 178         workset->len++;
 179 }
 180
 181 /**
 182  * Removes all entries from this workset
 183  */
 184 static void workset_clear(workset_t *workset)
 185 {
 186         workset->len = 0;
 187 }
 188
 189 /**
 190  * Removes the value @p val from the workset if present.
 191  */
 192 static void workset_remove(workset_t *workset, ir_node *val)
 193 {
 194         unsigned i;
 195         for (i = 0; i < workset->len; ++i) {
 196                 if (workset->vals[i].node == val) {
 197                         workset->vals[i] = workset->vals[--workset->len];
 198                         return;
 199                 }
 200         }
 201 }
 202
 203 static const loc_t *workset_contains(const workset_t *ws, const ir_node *val)
 204 {
 205         unsigned i;
 206         for (i = 0; i < ws->len; ++i) {
 207                 if (ws->vals[i].node == val)
 208                         return &ws->vals[i];
 209         }
 210
 211         return NULL;
 212 }
 213
 214 static int loc_compare(const void *a, const void *b)
 215 {
 216         const loc_t   *p  = ((const loc_t*) a);
 217         const loc_t   *q  = ((const loc_t*) b);
 218         const unsigned pt = p->time;
 219         const unsigned qt = q->time;
 220
 221         if (pt < qt)
 222                 return -1;
 223         if (pt > qt)
 224                 return 1;
 225
 226         return get_irn_node_nr(p->node) - get_irn_node_nr(q->node);
 227 }
 228
 229 static void workset_sort(workset_t *workset)
 230 {
 231         qsort(workset->vals, workset->len, sizeof(workset->vals[0]), loc_compare);
 232 }
 233
 234 static inline unsigned workset_get_time(const workset_t *workset, unsigned idx)
 235 {
 236         return workset->vals[idx].time;
 237 }
 238
 239 static inline void workset_set_time(workset_t *workset, unsigned idx,
 240                                     unsigned time)
 241 {
 242         workset->vals[idx].time = time;
 243 }
 244
 245 static inline unsigned workset_get_length(const workset_t *workset)
 246 {
 247         return workset->len;
 248 }
 249
 250 static inline void workset_set_length(workset_t *workset, unsigned len)
 251 {
 252         workset->len = len;
 253 }
 254
 255 static inline ir_node *workset_get_val(const workset_t *workset, unsigned idx)
 256 {
 257         return workset->vals[idx].node;
 258 }
 259
 260 /**
 261  * Iterates over all values in the working set.
 262  * @p ws The workset to iterate
 263  * @p v  A variable to put the current value in
 264  * @p i  An integer for internal use
 265  */
 266 #define workset_foreach(ws, v, i) \
 267         for (i=0; v=(i < ws->len) ? ws->vals[i].node : NULL, i < ws->len; ++i)
 268
 269 typedef struct block_info_t {
 270         workset_t *start_workset;
 271         workset_t *end_workset;
 272 } block_info_t;
 273
 274 static block_info_t *new_block_info(void)
 275 {
 276         return OALLOCZ(&obst, block_info_t);
 277 }
 278
 279 static inline block_info_t *get_block_info(const ir_node *block)
 280 {
 281         return (block_info_t*)get_irn_link(block);
 282 }
 283
 284 static inline void set_block_info(ir_node *block, block_info_t *info)
 285 {
 286         set_irn_link(block, info);
 287 }
 288
 289 /**
 290  * @return The distance to the next use or 0 if irn has dont_spill flag set
 291  */
 292 static unsigned get_distance(ir_node *from, const ir_node *def, int skip_from_uses)
 293 {
 294         be_next_use_t use;
 295         unsigned      costs;
 296         unsigned      time;
 297
 298         assert(!arch_irn_is_ignore(def));
 299
 300         use  = be_get_next_use(uses, from, def, skip_from_uses);
 301         time = use.time;
 302         if (USES_IS_INFINITE(time))
 303                 return USES_INFINITY;
 304
 305         /* We have to keep nonspillable nodes in the workingset */
 306         if (arch_get_irn_flags(skip_Proj_const(def)) & arch_irn_flags_dont_spill)
 307                 return 0;
 308
 309         /* give some bonus to rematerialisable nodes */
 310         if (remat_bonus > 0) {
 311                 costs = be_get_reload_costs_no_weight(senv, def, use.before);
 312                 assert(costs * remat_bonus < 1000);
 313                 time  += 1000 - (costs * remat_bonus);
 314         }
 315
 316         return time;
 317 }
 318
 319 /**
 320  * Performs the actions necessary to grant the request that:
 321  * - new_vals can be held in registers
 322  * - as few as possible other values are disposed
 323  * - the worst values get disposed
 324  *
 325  * @p is_usage indicates that the values in new_vals are used (not defined)
 326  * In this case reloads must be performed
 327  */
 328 static void displace(workset_t *new_vals, int is_usage)
 329 {
 330         ir_node **to_insert = ALLOCAN(ir_node*, n_regs);
 331         bool     *spilled   = ALLOCAN(bool,     n_regs);
 332         ir_node  *val;
 333         int       i;
 334         int       len;
 335         int       spills_needed;
 336         int       demand;
 337         unsigned  iter;
 338
 339         /* 1. Identify the number of needed slots and the values to reload */
 340         demand = 0;
 341         workset_foreach(new_vals, val, iter) {
 342                 bool reloaded = false;
 343
 344                 if (! workset_contains(ws, val)) {
 345                         DB((dbg, DBG_DECIDE, "    insert %+F\n", val));
 346                         if (is_usage) {
 347                                 DB((dbg, DBG_SPILL, "Reload %+F before %+F\n", val, instr));
 348                                 be_add_reload(senv, val, instr, cls, 1);
 349                                 reloaded = true;
 350                         }
 351                 } else {
 352                         DB((dbg, DBG_DECIDE, "    %+F already in workset\n", val));
 353                         assert(is_usage);
 354                         /* remove the value from the current workset so it is not accidently
 355                          * spilled */
 356                         workset_remove(ws, val);
 357                 }
 358                 spilled[demand]   = reloaded;
 359                 to_insert[demand] = val;
 360                 ++demand;
 361         }
 362
 363         /* 2. Make room for at least 'demand' slots */
 364         len           = workset_get_length(ws);
 365         spills_needed = len + demand - n_regs;
 366         assert(spills_needed <= len);
 367
 368         /* Only make more free room if we do not have enough */
 369         if (spills_needed > 0) {
 370                 DB((dbg, DBG_DECIDE, "    disposing %d values\n", spills_needed));
 371
 372                 /* calculate current next-use distance for live values */
 373                 for (i = 0; i < len; ++i) {
 374                         ir_node  *val  = workset_get_val(ws, i);
 375                         unsigned  dist = get_distance(instr, val, !is_usage);
 376                         workset_set_time(ws, i, dist);
 377                 }
 378
 379                 /* sort entries by increasing nextuse-distance*/
 380                 workset_sort(ws);
 381
 382                 for (i = len - spills_needed; i < len; ++i) {
 383                         ir_node *val = ws->vals[i].node;
 384
 385                         DB((dbg, DBG_DECIDE, "    disposing node %+F (%u)\n", val,
 386                              workset_get_time(ws, i)));
 387
 388                         if (move_spills) {
 389                                 if (!USES_IS_INFINITE(ws->vals[i].time)
 390                                                 && !ws->vals[i].spilled) {
 391                                         ir_node *after_pos = sched_prev(instr);
 392                                         DB((dbg, DBG_DECIDE, "Spill %+F after node %+F\n", val,
 393                                                 after_pos));
 394                                         be_add_spill(senv, val, after_pos);
 395                                 }
 396                         }
 397                 }
 398
 399                 /* kill the last 'demand' entries in the array */
 400                 workset_set_length(ws, len - spills_needed);
 401         }
 402
 403         /* 3. Insert the new values into the workset */
 404         for (i = 0; i < demand; ++i) {
 405                 ir_node *val = to_insert[i];
 406
 407                 workset_insert(ws, val, spilled[i]);
 408         }
 409 }
 410
 411 enum {
 412         AVAILABLE_EVERYWHERE,
 413         AVAILABLE_NOWHERE,
 414         AVAILABLE_PARTLY,
 415         AVAILABLE_UNKNOWN
 416 };
 417
 418 static unsigned available_in_all_preds(workset_t* const* pred_worksets,
 419                                        size_t n_pred_worksets,
 420                                        const ir_node *value, bool is_local_phi)
 421 {
 422         size_t i;
 423         bool   avail_everywhere = true;
 424         bool   avail_nowhere    = true;
 425
 426         assert(n_pred_worksets > 0);
 427
 428         /* value available in all preds? */
 429         for (i = 0; i < n_pred_worksets; ++i) {
 430                 bool             found     = false;
 431                 const workset_t *p_workset = pred_worksets[i];
 432                 int              p_len     = workset_get_length(p_workset);
 433                 int              p_i;
 434                 const ir_node   *l_value;
 435
 436                 if (is_local_phi) {
 437                         assert(is_Phi(value));
 438                         l_value = get_irn_n(value, i);
 439                 } else {
 440                         l_value = value;
 441                 }
 442
 443                 for (p_i = 0; p_i < p_len; ++p_i) {
 444                         const loc_t *p_l = &p_workset->vals[p_i];
 445                         if (p_l->node != l_value)
 446                                 continue;
 447
 448                         found = true;
 449                         break;
 450                 }
 451
 452                 if (found) {
 453                         avail_nowhere = false;
 454                 } else {
 455                         avail_everywhere = false;
 456                 }
 457         }
 458
 459         if (avail_everywhere) {
 460                 assert(!avail_nowhere);
 461                 return AVAILABLE_EVERYWHERE;
 462         } else if (avail_nowhere) {
 463                 return AVAILABLE_NOWHERE;
 464         } else {
 465                 return AVAILABLE_PARTLY;
 466         }
 467 }
 468
 469 /** Decides whether a specific node should be in the start workset or not
 470  *
 471  * @param env      belady environment
 472  * @param first
 473  * @param node     the node to test
 474  * @param loop     the loop of the node
 475  */
 476 static loc_t to_take_or_not_to_take(ir_node* first, ir_node *node,
 477                                     ir_loop *loop, unsigned available)
 478 {
 479         be_next_use_t next_use;
 480         loc_t         loc;
 481
 482         loc.time    = USES_INFINITY;
 483         loc.node    = node;
 484         loc.spilled = false;
 485
 486         if (!arch_irn_consider_in_reg_alloc(cls, node)) {
 487                 loc.time = USES_INFINITY;
 488                 return loc;
 489         }
 490
 491         /* We have to keep nonspillable nodes in the workingset */
 492         if (arch_get_irn_flags(skip_Proj_const(node)) & arch_irn_flags_dont_spill) {
 493                 loc.time = 0;
 494                 DB((dbg, DBG_START, "    %+F taken (dontspill node)\n", node, loc.time));
 495                 return loc;
 496         }
 497
 498         next_use = be_get_next_use(uses, first, node, 0);
 499         if (USES_IS_INFINITE(next_use.time)) {
 500                 /* the nodes marked as live in shouldn't be dead, so it must be a phi */
 501                 assert(is_Phi(node));
 502                 loc.time = USES_INFINITY;
 503                 DB((dbg, DBG_START, "    %+F not taken (dead)\n", node));
 504                 return loc;
 505         }
 506
 507         loc.time = next_use.time;
 508
 509         if (improve_known_preds) {
 510                 if (available == AVAILABLE_EVERYWHERE) {
 511                         DB((dbg, DBG_START, "    %+F taken (%u, live in all preds)\n",
 512                             node, loc.time));
 513                         return loc;
 514                 } else if (available == AVAILABLE_NOWHERE) {
 515                         DB((dbg, DBG_START, "    %+F not taken (%u, live in no pred)\n",
 516                             node, loc.time));
 517                         loc.time = USES_INFINITY;
 518                         return loc;
 519                 }
 520         }
 521
 522         if (!respectloopdepth || next_use.outermost_loop >= get_loop_depth(loop)) {
 523                 DB((dbg, DBG_START, "    %+F taken (%u, loop %d)\n", node, loc.time,
 524                     next_use.outermost_loop));
 525         } else {
 526                 loc.time = USES_PENDING;
 527                 DB((dbg, DBG_START, "    %+F delayed (outerdepth %d < loopdepth %d)\n",
 528                     node, next_use.outermost_loop, get_loop_depth(loop)));
 529         }
 530
 531         return loc;
 532 }
 533
 534 /**
 535  * Computes the start-workset for a block with multiple predecessors. We assume
 536  * that at least 1 of the predeccesors is a back-edge which means we're at the
 537  * beginning of a loop. We try to reload as much values as possible now so they
 538  * don't get reloaded inside the loop.
 539  */
 540 static void decide_start_workset(const ir_node *block)
 541 {
 542         ir_loop    *loop = get_irn_loop(block);
 543         ir_node    *first;
 544         loc_t       loc;
 545         loc_t      *starters;
 546         loc_t      *delayed;
 547         unsigned    len;
 548         unsigned    i;
 549         unsigned    ws_count;
 550         int         free_slots, free_pressure_slots;
 551         unsigned    pressure;
 552         int         arity;
 553         workset_t **pred_worksets;
 554         bool        all_preds_known;
 555
 556         /* check predecessors */
 557         arity           = get_irn_arity(block);
 558         pred_worksets   = ALLOCAN(workset_t*, arity);
 559         all_preds_known = true;
 560         for (int in = 0; in < arity; ++in) {
 561                 ir_node      *pred_block = get_Block_cfgpred_block(block, in);
 562                 block_info_t *pred_info  = get_block_info(pred_block);
 563
 564                 if (pred_info == NULL) {
 565                         pred_worksets[in] = NULL;
 566                         all_preds_known   = false;
 567                 } else {
 568                         pred_worksets[in] = pred_info->end_workset;
 569                 }
 570         }
 571
 572         /* Collect all values living at start of block */
 573         starters = NEW_ARR_F(loc_t, 0);
 574         delayed  = NEW_ARR_F(loc_t, 0);
 575
 576         DB((dbg, DBG_START, "Living at start of %+F:\n", block));
 577         first = sched_first(block);
 578
 579         /* check all Phis first */
 580         sched_foreach(block, node) {
 581                 unsigned available;
 582
 583                 if (! is_Phi(node))
 584                         break;
 585                 if (!arch_irn_consider_in_reg_alloc(cls, node))
 586                         continue;
 587
 588                 if (all_preds_known) {
 589                         available = available_in_all_preds(pred_worksets, arity, node, true);
 590                 } else {
 591                         available = AVAILABLE_UNKNOWN;
 592                 }
 593
 594                 loc = to_take_or_not_to_take(first, node, loop, available);
 595
 596                 if (! USES_IS_INFINITE(loc.time)) {
 597                         if (USES_IS_PENDING(loc.time))
 598                                 ARR_APP1(loc_t, delayed, loc);
 599                         else
 600                                 ARR_APP1(loc_t, starters, loc);
 601                 } else {
 602                         be_spill_phi(senv, node);
 603                 }
 604         }
 605
 606         /* check all Live-Ins */
 607         be_lv_foreach(lv, block, be_lv_state_in, node) {
 608                 unsigned available;
 609
 610                 if (all_preds_known) {
 611                         available = available_in_all_preds(pred_worksets, arity, node, false);
 612                 } else {
 613                         available = AVAILABLE_UNKNOWN;
 614                 }
 615
 616                 loc = to_take_or_not_to_take(first, node, loop, available);
 617
 618                 if (! USES_IS_INFINITE(loc.time)) {
 619                         if (USES_IS_PENDING(loc.time))
 620                                 ARR_APP1(loc_t, delayed, loc);
 621                         else
 622                                 ARR_APP1(loc_t, starters, loc);
 623                 }
 624         }
 625
 626         pressure            = be_get_loop_pressure(loop_ana, cls, loop);
 627         assert(ARR_LEN(delayed) <= pressure);
 628         free_slots          = n_regs - ARR_LEN(starters);
 629         free_pressure_slots = n_regs - (pressure - ARR_LEN(delayed));
 630         free_slots          = MIN(free_slots, free_pressure_slots);
 631
 632         /* so far we only put nodes into the starters list that are used inside
 633          * the loop. If register pressure in the loop is low then we can take some
 634          * values and let them live through the loop */
 635         DB((dbg, DBG_START, "Loop pressure %d, taking %d delayed vals\n",
 636             pressure, free_slots));
 637         if (free_slots > 0) {
 638                 size_t i;
 639
 640                 qsort(delayed, ARR_LEN(delayed), sizeof(delayed[0]), loc_compare);
 641
 642                 for (i = 0; i < ARR_LEN(delayed) && free_slots > 0; ++i) {
 643                         int    p, arity;
 644                         loc_t *loc = & delayed[i];
 645
 646                         if (!is_Phi(loc->node)) {
 647                                 /* don't use values which are dead in a known predecessors
 648                                  * to not induce unnecessary reloads */
 649                                 arity = get_irn_arity(block);
 650                                 for (p = 0; p < arity; ++p) {
 651                                         ir_node      *pred_block = get_Block_cfgpred_block(block, p);
 652                                         block_info_t *pred_info  = get_block_info(pred_block);
 653
 654                                         if (pred_info == NULL)
 655                                                 continue;
 656
 657                                         if (!workset_contains(pred_info->end_workset, loc->node)) {
 658                                                 DB((dbg, DBG_START,
 659                                                         "    delayed %+F not live at pred %+F\n", loc->node,
 660                                                         pred_block));
 661                                                 goto skip_delayed;
 662                                         }
 663                                 }
 664                         }
 665
 666                         DB((dbg, DBG_START, "    delayed %+F taken\n", loc->node));
 667                         ARR_APP1(loc_t, starters, *loc);
 668                         loc->node = NULL;
 669                         --free_slots;
 670                 skip_delayed:
 671                         ;
 672                 }
 673         }
 674
 675         /* spill phis (the actual phis not just their values) that are in this block
 676          * but not in the start workset */
 677         len = ARR_LEN(delayed);
 678         for (i = 0; i < len; ++i) {
 679                 ir_node *node = delayed[i].node;
 680                 if (node == NULL || !is_Phi(node) || get_nodes_block(node) != block)
 681                         continue;
 682
 683                 DB((dbg, DBG_START, "    spilling delayed phi %+F\n", node));
 684                 be_spill_phi(senv, node);
 685         }
 686         DEL_ARR_F(delayed);
 687
 688         /* Sort start values by first use */
 689         qsort(starters, ARR_LEN(starters), sizeof(starters[0]), loc_compare);
 690
 691         /* Copy the best ones from starters to start workset */
 692         ws_count = MIN((unsigned) ARR_LEN(starters), n_regs);
 693         workset_clear(ws);
 694         workset_bulk_fill(ws, ws_count, starters);
 695
 696         /* spill phis (the actual phis not just their values) that are in this block
 697          * but not in the start workset */
 698         len = ARR_LEN(starters);
 699         for (i = ws_count; i < len; ++i) {
 700                 ir_node *node = starters[i].node;
 701                 if (! is_Phi(node) || get_nodes_block(node) != block)
 702                         continue;
 703
 704                 DB((dbg, DBG_START, "    spilling phi %+F\n", node));
 705                 be_spill_phi(senv, node);
 706         }
 707
 708         DEL_ARR_F(starters);
 709
 710         /* determine spill status of the values: If there's 1 pred block (which
 711          * is no backedge) where the value is spilled then we must set it to
 712          * spilled here. */
 713         for (i = 0; i < ws_count; ++i) {
 714                 loc_t   *loc     = &ws->vals[i];
 715                 ir_node *value   = loc->node;
 716                 bool     spilled;
 717                 int      n;
 718
 719                 /* phis from this block aren't spilled */
 720                 if (get_nodes_block(value) == block) {
 721                         assert(is_Phi(value));
 722                         loc->spilled = false;
 723                         continue;
 724                 }
 725
 726                 /* determine if value was spilled on any predecessor */
 727                 spilled = false;
 728                 for (n = 0; n < arity; ++n) {
 729                         workset_t *pred_workset = pred_worksets[n];
 730                         int        p_len;
 731                         int        p;
 732
 733                         if (pred_workset == NULL)
 734                                 continue;
 735
 736                         p_len = workset_get_length(pred_workset);
 737                         for (p = 0; p < p_len; ++p) {
 738                                 loc_t *l = &pred_workset->vals[p];
 739
 740                                 if (l->node != value)
 741                                         continue;
 742
 743                                 if (l->spilled) {
 744                                         spilled = true;
 745                                 }
 746                                 break;
 747                         }
 748                 }
 749
 750                 loc->spilled = spilled;
 751         }
 752 }
 753
 754 /**
 755  * For the given block @p block, decide for each values
 756  * whether it is used from a register or is reloaded
 757  * before the use.
 758  */
 759 static void process_block(ir_node *block)
 760 {
 761         workset_t    *new_vals;
 762         unsigned      iter;
 763         block_info_t *block_info;
 764         int           arity;
 765
 766         /* no need to process a block twice */
 767         assert(get_block_info(block) == NULL);
 768
 769         /* construct start workset */
 770         arity = get_Block_n_cfgpreds(block);
 771         if (arity == 0) {
 772                 /* no predecessor -> empty set */
 773                 workset_clear(ws);
 774         } else if (arity == 1) {
 775                 /* one predecessor, copy its end workset */
 776                 ir_node      *pred_block = get_Block_cfgpred_block(block, 0);
 777                 block_info_t *pred_info  = get_block_info(pred_block);
 778
 779                 assert(pred_info != NULL);
 780                 workset_copy(ws, pred_info->end_workset);
 781         } else {
 782                 /* multiple predecessors, do more advanced magic :) */
 783                 decide_start_workset(block);
 784         }
 785
 786         DB((dbg, DBG_DECIDE, "\n"));
 787         DB((dbg, DBG_DECIDE, "Decide for %+F\n", block));
 788
 789         block_info = new_block_info();
 790         set_block_info(block, block_info);
 791
 792         DB((dbg, DBG_WSETS, "Start workset for %+F:\n", block));
 793         {
 794                 ir_node *irn;
 795                 workset_foreach(ws, irn, iter) {
 796                         DB((dbg, DBG_WSETS, "  %+F (%u)\n", irn, workset_get_time(ws, iter)));
 797                 }
 798         }
 799
 800         block_info->start_workset = workset_clone(ws);
 801
 802         /* process the block from start to end */
 803         DB((dbg, DBG_WSETS, "Processing...\n"));
 804         /* TODO: this leaks (into the obstack)... */
 805         new_vals = new_workset();
 806
 807         sched_foreach(block, irn) {
 808                 int i, arity;
 809                 assert(workset_get_length(ws) <= n_regs);
 810
 811                 /* Phis are no real instr (see insert_starters()) */
 812                 if (is_Phi(irn)) {
 813                         continue;
 814                 }
 815                 DB((dbg, DBG_DECIDE, "  ...%+F\n", irn));
 816
 817                 /* set instruction in the workset */
 818                 instr = irn;
 819
 820                 /* allocate all values _used_ by this instruction */
 821                 workset_clear(new_vals);
 822                 for (i = 0, arity = get_irn_arity(irn); i < arity; ++i) {
 823                         ir_node *in = get_irn_n(irn, i);
 824                         if (!arch_irn_consider_in_reg_alloc(cls, in))
 825                                 continue;
 826
 827                         /* (note that "spilled" is irrelevant here) */
 828                         workset_insert(new_vals, in, false);
 829                 }
 830                 displace(new_vals, 1);
 831
 832                 /* allocate all values _defined_ by this instruction */
 833                 workset_clear(new_vals);
 834                 be_foreach_definition(irn, cls, value,
 835                         assert(req_->width == 1);
 836                         workset_insert(new_vals, value, false);
 837                 );
 838                 displace(new_vals, 0);
 839         }
 840
 841         /* Remember end-workset for this block */
 842         block_info->end_workset = workset_clone(ws);
 843         DB((dbg, DBG_WSETS, "End workset for %+F:\n", block));
 844         {
 845                 ir_node *irn;
 846                 workset_foreach(ws, irn, iter)
 847                         DB((dbg, DBG_WSETS, "  %+F (%u)\n", irn, workset_get_time(ws, iter)));
 848         }
 849 }
 850
 851 /**
 852  * 'decide' is block-local and makes assumptions
 853  * about the set of live-ins. Thus we must adapt the
 854  * live-outs to the live-ins at each block-border.
 855  */
 856 static void fix_block_borders(ir_node *block, void *data)
 857 {
 858         workset_t *start_workset;
 859         int        arity;
 860         int        i;
 861         unsigned   iter;
 862         (void) data;
 863
 864         DB((dbg, DBG_FIX, "\n"));
 865         DB((dbg, DBG_FIX, "Fixing %+F\n", block));
 866
 867         arity = get_irn_arity(block);
 868         /* can happen for endless loops */
 869         if (arity == 0)
 870                 return;
 871
 872         start_workset = get_block_info(block)->start_workset;
 873
 874         /* process all pred blocks */
 875         for (i = 0; i < arity; ++i) {
 876                 ir_node   *pred = get_Block_cfgpred_block(block, i);
 877                 workset_t *pred_end_workset = get_block_info(pred)->end_workset;
 878                 ir_node   *node;
 879
 880                 DB((dbg, DBG_FIX, "  Pred %+F\n", pred));
 881
 882                 /* spill all values not used anymore */
 883                 workset_foreach(pred_end_workset, node, iter) {
 884                         ir_node *n2;
 885                         unsigned iter2;
 886                         bool     found = false;
 887                         workset_foreach(start_workset, n2, iter2) {
 888                                 if (n2 == node) {
 889                                         found = true;
 890                                         break;
 891                                 }
 892                                 /* note that we do not look at phi inputs, becuase the values
 893                                  * will be either live-end and need no spill or
 894                                  * they have other users in which must be somewhere else in the
 895                                  * workset */
 896                         }
 897
 898                         if (found)
 899                                 continue;
 900
 901                         if (move_spills && be_is_live_in(lv, block, node)
 902                                         && !pred_end_workset->vals[iter].spilled) {
 903                                 ir_node *insert_point;
 904                                 if (arity > 1) {
 905                                         insert_point = be_get_end_of_block_insertion_point(pred);
 906                                         insert_point = sched_prev(insert_point);
 907                                 } else {
 908                                         insert_point = block;
 909                                 }
 910                                 DB((dbg, DBG_SPILL, "Spill %+F after %+F\n", node,
 911                                      insert_point));
 912                                 be_add_spill(senv, node, insert_point);
 913                         }
 914                 }
 915
 916                 /* reload missing values in predecessors, add missing spills */
 917                 workset_foreach(start_workset, node, iter) {
 918                         const loc_t *l    = &start_workset->vals[iter];
 919                         const loc_t *pred_loc;
 920
 921                         /* if node is a phi of the current block we reload
 922                          * the corresponding argument, else node itself */
 923                         if (is_Phi(node) && get_nodes_block(node) == block) {
 924                                 node = get_irn_n(node, i);
 925                                 assert(!l->spilled);
 926
 927                                 /* we might have unknowns as argument for the phi */
 928                                 if (!arch_irn_consider_in_reg_alloc(cls, node))
 929                                         continue;
 930                         }
 931
 932                         /* check if node is in a register at end of pred */
 933                         pred_loc = workset_contains(pred_end_workset, node);
 934                         if (pred_loc != NULL) {
 935                                 /* we might have to spill value on this path */
 936                                 if (move_spills && !pred_loc->spilled && l->spilled) {
 937                                         ir_node *insert_point
 938                                                 = be_get_end_of_block_insertion_point(pred);
 939                                         insert_point = sched_prev(insert_point);
 940                                         DB((dbg, DBG_SPILL, "Spill %+F after %+F\n", node,
 941                                             insert_point));
 942                                         be_add_spill(senv, node, insert_point);
 943                                 }
 944                         } else {
 945                                 /* node is not in register at the end of pred -> reload it */
 946                                 DB((dbg, DBG_FIX, "    reload %+F\n", node));
 947                                 DB((dbg, DBG_SPILL, "Reload %+F before %+F,%d\n", node, block, i));
 948                                 be_add_reload_on_edge(senv, node, block, i, cls, 1);
 949                         }
 950                 }
 951         }
 952 }
 953
 954 static void be_spill_belady(ir_graph *irg, const arch_register_class_t *rcls)
 955 {
 956         int i;
 957
 958         be_assure_live_sets(irg);
 959
 960         stat_ev_tim_push();
 961         assure_loopinfo(irg);
 962         stat_ev_tim_pop("belady_time_backedges");
 963
 964         stat_ev_tim_push();
 965         be_clear_links(irg);
 966         stat_ev_tim_pop("belady_time_clear_links");
 967
 968         ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
 969
 970         /* init belady env */
 971         stat_ev_tim_push();
 972         obstack_init(&obst);
 973         cls       = rcls;
 974         lv        = be_get_irg_liveness(irg);
 975         n_regs    = be_get_n_allocatable_regs(irg, cls);
 976         ws        = new_workset();
 977         uses      = be_begin_uses(irg, lv);
 978         loop_ana  = be_new_loop_pressure(irg, cls);
 979         senv      = be_new_spill_env(irg);
 980         blocklist = be_get_cfgpostorder(irg);
 981         stat_ev_tim_pop("belady_time_init");
 982
 983         stat_ev_tim_push();
 984         /* walk blocks in reverse postorder */
 985         for (i = ARR_LEN(blocklist) - 1; i >= 0; --i) {
 986                 process_block(blocklist[i]);
 987         }
 988         DEL_ARR_F(blocklist);
 989         stat_ev_tim_pop("belady_time_belady");
 990
 991         stat_ev_tim_push();
 992         /* belady was block-local, fix the global flow by adding reloads on the
 993          * edges */
 994         irg_block_walk_graph(irg, fix_block_borders, NULL, NULL);
 995         stat_ev_tim_pop("belady_time_fix_borders");
 996
 997         ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
 998
 999         /* Insert spill/reload nodes into the graph and fix usages */
1000         be_insert_spills_reloads(senv);
1001
1002         /* clean up */
1003         be_delete_spill_env(senv);
1004         be_end_uses(uses);
1005         be_free_loop_pressure(loop_ana);
1006         obstack_free(&obst, NULL);
1007 }
1008
1009 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady)
1010 void be_init_spillbelady(void)
1011 {
1012         static be_spiller_t belady_spiller = {
1013                 be_spill_belady
1014         };
1015         lc_opt_entry_t *be_grp       = lc_opt_get_grp(firm_opt_get_root(), "be");
1016         lc_opt_entry_t *belady_group = lc_opt_get_grp(be_grp, "belady");
1017         lc_opt_add_table(belady_group, options);
1018
1019         be_register_spiller("belady", &belady_spiller);
1020         FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
1021 }