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

   1 /*
   2  * This file is part of libFirm.
   3  * Copyright (C) 2012 University of Karlsruhe.
   4  */
   5
   6 /**
   7  * @file
   8  * @brief       Beladys spillalgorithm.
   9  * @author      Daniel Grund, Matthias Braun
  10  * @date        20.09.2005
  11  */
  12 #include "config.h"
  13
  14 #include <stdbool.h>
  15
  16 #include "obst.h"
  17 #include "irgraph.h"
  18 #include "irnode.h"
  19 #include "irmode.h"
  20 #include "irgwalk.h"
  21 #include "irloop.h"
  22 #include "iredges_t.h"
  23 #include "ircons_t.h"
  24 #include "irprintf.h"
  25 #include "irnodeset.h"
  26 #include "irtools.h"
  27 #include "statev_t.h"
  28 #include "util.h"
  29
  30 #include "beutil.h"
  31 #include "bearch.h"
  32 #include "beuses.h"
  33 #include "besched.h"
  34 #include "beirgmod.h"
  35 #include "belive_t.h"
  36 #include "benode.h"
  37 #include "bechordal_t.h"
  38 #include "bespill.h"
  39 #include "beloopana.h"
  40 #include "beirg.h"
  41 #include "bespillutil.h"
  42 #include "bemodule.h"
  43
  44 #define DBG_SPILL     1
  45 #define DBG_WSETS     2
  46 #define DBG_FIX       4
  47 #define DBG_DECIDE    8
  48 #define DBG_START    16
  49 #define DBG_SLOTS    32
  50 #define DBG_TRACE    64
  51 #define DBG_WORKSET 128
  52 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  53
  54 #define TIME_UNDEFINED 6666
  55
  56 /**
  57  * An association between a node and a point in time.
  58  */
  59 typedef struct loc_t {
  60         ir_node          *node;
  61         unsigned          time;     /**< A use time (see beuses.h). */
  62         bool              spilled;  /**< value was already spilled on this path */
  63 } loc_t;
  64
  65 typedef struct workset_t {
  66         unsigned len;     /**< current length */
  67         loc_t    vals[];  /**< array of the values/distances in this working set */
  68 } workset_t;
  69
  70 static struct obstack               obst;
  71 static const arch_register_class_t *cls;
  72 static const be_lv_t               *lv;
  73 static be_loopana_t                *loop_ana;
  74 static unsigned                     n_regs;
  75 static workset_t                   *ws;     /**< the main workset used while
  76                                                      processing a block. */
  77 static be_uses_t                   *uses;   /**< env for the next-use magic */
  78 static spill_env_t                 *senv;   /**< see bespill.h */
  79 static ir_node                    **blocklist;
  80
  81 static int                          move_spills      = true;
  82 static int                          respectloopdepth = true;
  83 static int                          improve_known_preds = true;
  84 /* factor to weight the different costs of reloading/rematerializing a node
  85    (see bespill.h be_get_reload_costs_no_weight) */
  86 static int                          remat_bonus      = 10;
  87
  88 static const lc_opt_table_entry_t options[] = {
  89         LC_OPT_ENT_BOOL   ("movespills", "try to move spills out of loops", &move_spills),
  90         LC_OPT_ENT_BOOL   ("respectloopdepth", "outermost loop cutting", &respectloopdepth),
  91         LC_OPT_ENT_BOOL   ("improveknownpreds", "known preds cutting", &improve_known_preds),
  92         LC_OPT_ENT_INT    ("rematbonus", "give bonus to rematerialisable nodes", &remat_bonus),
  93         LC_OPT_LAST
  94 };
  95
  96 /**
  97  * Alloc a new workset on obstack @p ob with maximum size @p max
  98  */
  99 static workset_t *new_workset(void)
 100 {
 101         return OALLOCFZ(&obst, workset_t, vals, n_regs);
 102 }
 103
 104 /**
 105  * Alloc a new instance on obstack and make it equal to @param workset
 106  */
 107 static workset_t *workset_clone(workset_t *workset)
 108 {
 109         workset_t *res = OALLOCF(&obst, workset_t, vals, n_regs);
 110         memcpy(res, workset, sizeof(*res) + n_regs * sizeof(res->vals[0]));
 111         return res;
 112 }
 113
 114 /**
 115  * Copy workset @param src to @param tgt
 116  */
 117 static void workset_copy(workset_t *dest, const workset_t *src)
 118 {
 119         size_t size = sizeof(*src) + n_regs * sizeof(src->vals[0]);
 120         memcpy(dest, src, size);
 121 }
 122
 123 /**
 124  * Overwrites the current content array of @param ws with the
 125  * @param count locations given at memory @param locs.
 126  * Set the length of @param ws to count.
 127  */
 128 static void workset_bulk_fill(workset_t *workset, int count, const loc_t *locs)
 129 {
 130         workset->len = count;
 131         memcpy(&(workset->vals[0]), locs, count * sizeof(locs[0]));
 132 }
 133
 134 /**
 135  * Inserts the value @p val into the workset, iff it is not
 136  * already contained. The workset must not be full.
 137  */
 138 static void workset_insert(workset_t *workset, ir_node *val, bool spilled)
 139 {
 140         loc_t    *loc;
 141         unsigned  i;
 142         /* check for current regclass */
 143         assert(arch_irn_consider_in_reg_alloc(cls, val));
 144
 145         /* check if val is already contained */
 146         for (i = 0; i < workset->len; ++i) {
 147                 loc = &workset->vals[i];
 148                 if (loc->node == val) {
 149                         if (spilled) {
 150                                 loc->spilled = true;
 151                         }
 152                         return;
 153                 }
 154         }
 155
 156         /* insert val */
 157         assert(workset->len < n_regs && "Workset already full!");
 158         loc           = &workset->vals[workset->len];
 159         loc->node     = val;
 160         loc->spilled  = spilled;
 161         loc->time     = TIME_UNDEFINED;
 162         workset->len++;
 163 }
 164
 165 /**
 166  * Removes all entries from this workset
 167  */
 168 static void workset_clear(workset_t *workset)
 169 {
 170         workset->len = 0;
 171 }
 172
 173 /**
 174  * Removes the value @p val from the workset if present.
 175  */
 176 static void workset_remove(workset_t *workset, ir_node *val)
 177 {
 178         unsigned i;
 179         for (i = 0; i < workset->len; ++i) {
 180                 if (workset->vals[i].node == val) {
 181                         workset->vals[i] = workset->vals[--workset->len];
 182                         return;
 183                 }
 184         }
 185 }
 186
 187 static const loc_t *workset_contains(const workset_t *ws, const ir_node *val)
 188 {
 189         unsigned i;
 190         for (i = 0; i < ws->len; ++i) {
 191                 if (ws->vals[i].node == val)
 192                         return &ws->vals[i];
 193         }
 194
 195         return NULL;
 196 }
 197
 198 static int loc_compare(const void *a, const void *b)
 199 {
 200         const loc_t   *p  = ((const loc_t*) a);
 201         const loc_t   *q  = ((const loc_t*) b);
 202         const unsigned pt = p->time;
 203         const unsigned qt = q->time;
 204
 205         if (pt < qt)
 206                 return -1;
 207         if (pt > qt)
 208                 return 1;
 209
 210         return get_irn_node_nr(p->node) - get_irn_node_nr(q->node);
 211 }
 212
 213 static void workset_sort(workset_t *workset)
 214 {
 215         qsort(workset->vals, workset->len, sizeof(workset->vals[0]), loc_compare);
 216 }
 217
 218 static inline unsigned workset_get_time(const workset_t *workset, unsigned idx)
 219 {
 220         return workset->vals[idx].time;
 221 }
 222
 223 static inline void workset_set_time(workset_t *workset, unsigned idx,
 224                                     unsigned time)
 225 {
 226         workset->vals[idx].time = time;
 227 }
 228
 229 static inline unsigned workset_get_length(const workset_t *workset)
 230 {
 231         return workset->len;
 232 }
 233
 234 static inline void workset_set_length(workset_t *workset, unsigned len)
 235 {
 236         workset->len = len;
 237 }
 238
 239 static inline ir_node *workset_get_val(const workset_t *workset, unsigned idx)
 240 {
 241         return workset->vals[idx].node;
 242 }
 243
 244 /**
 245  * Iterates over all values in the working set.
 246  * @p ws The workset to iterate
 247  * @p v  A variable to put the current value in
 248  * @p i  An integer for internal use
 249  */
 250 #define workset_foreach(ws, v, i) \
 251         for (i=0; v=(i < ws->len) ? ws->vals[i].node : NULL, i < ws->len; ++i)
 252
 253 typedef struct block_info_t {
 254         workset_t *start_workset;
 255         workset_t *end_workset;
 256 } block_info_t;
 257
 258 static block_info_t *new_block_info(void)
 259 {
 260         return OALLOCZ(&obst, block_info_t);
 261 }
 262
 263 static inline block_info_t *get_block_info(const ir_node *block)
 264 {
 265         return (block_info_t*)get_irn_link(block);
 266 }
 267
 268 static inline void set_block_info(ir_node *block, block_info_t *info)
 269 {
 270         set_irn_link(block, info);
 271 }
 272
 273 /**
 274  * @return The distance to the next use or 0 if irn has dont_spill flag set
 275  */
 276 static unsigned get_distance(ir_node *from, const ir_node *def, int skip_from_uses)
 277 {
 278         be_next_use_t use;
 279         unsigned      costs;
 280         unsigned      time;
 281
 282         assert(!arch_irn_is_ignore(def));
 283
 284         use  = be_get_next_use(uses, from, def, skip_from_uses);
 285         time = use.time;
 286         if (USES_IS_INFINITE(time))
 287                 return USES_INFINITY;
 288
 289         /* We have to keep nonspillable nodes in the workingset */
 290         if (arch_get_irn_flags(skip_Proj_const(def)) & arch_irn_flags_dont_spill)
 291                 return 0;
 292
 293         /* give some bonus to rematerialisable nodes */
 294         if (remat_bonus > 0) {
 295                 costs = be_get_reload_costs_no_weight(senv, def, use.before);
 296                 assert(costs * remat_bonus < 1000);
 297                 time  += 1000 - (costs * remat_bonus);
 298         }
 299
 300         return time;
 301 }
 302
 303 /**
 304  * Performs the actions necessary to grant the request that:
 305  * - new_vals can be held in registers
 306  * - as few as possible other values are disposed
 307  * - the worst values get disposed
 308  *
 309  * @p is_usage indicates that the values in new_vals are used (not defined)
 310  * In this case reloads must be performed
 311  */
 312 static void displace(workset_t *const new_vals, int const is_usage, ir_node *const instr)
 313 {
 314         ir_node **to_insert = ALLOCAN(ir_node*, n_regs);
 315         bool     *spilled   = ALLOCAN(bool,     n_regs);
 316         ir_node  *val;
 317         int       i;
 318         int       len;
 319         int       spills_needed;
 320         int       demand;
 321         unsigned  iter;
 322
 323         /* 1. Identify the number of needed slots and the values to reload */
 324         demand = 0;
 325         workset_foreach(new_vals, val, iter) {
 326                 bool reloaded = false;
 327
 328                 if (! workset_contains(ws, val)) {
 329                         DB((dbg, DBG_DECIDE, "    insert %+F\n", val));
 330                         if (is_usage) {
 331                                 DB((dbg, DBG_SPILL, "Reload %+F before %+F\n", val, instr));
 332                                 be_add_reload(senv, val, instr, cls, 1);
 333                                 reloaded = true;
 334                         }
 335                 } else {
 336                         DB((dbg, DBG_DECIDE, "    %+F already in workset\n", val));
 337                         assert(is_usage);
 338                         /* remove the value from the current workset so it is not accidently
 339                          * spilled */
 340                         workset_remove(ws, val);
 341                 }
 342                 spilled[demand]   = reloaded;
 343                 to_insert[demand] = val;
 344                 ++demand;
 345         }
 346
 347         /* 2. Make room for at least 'demand' slots */
 348         len           = workset_get_length(ws);
 349         spills_needed = len + demand - n_regs;
 350         assert(spills_needed <= len);
 351
 352         /* Only make more free room if we do not have enough */
 353         if (spills_needed > 0) {
 354                 DB((dbg, DBG_DECIDE, "    disposing %d values\n", spills_needed));
 355
 356                 /* calculate current next-use distance for live values */
 357                 for (i = 0; i < len; ++i) {
 358                         ir_node  *val  = workset_get_val(ws, i);
 359                         unsigned  dist = get_distance(instr, val, !is_usage);
 360                         workset_set_time(ws, i, dist);
 361                 }
 362
 363                 /* sort entries by increasing nextuse-distance*/
 364                 workset_sort(ws);
 365
 366                 for (i = len - spills_needed; i < len; ++i) {
 367                         ir_node *val = ws->vals[i].node;
 368
 369                         DB((dbg, DBG_DECIDE, "    disposing node %+F (%u)\n", val,
 370                              workset_get_time(ws, i)));
 371
 372                         if (move_spills) {
 373                                 if (!USES_IS_INFINITE(ws->vals[i].time)
 374                                                 && !ws->vals[i].spilled) {
 375                                         ir_node *after_pos = sched_prev(instr);
 376                                         DB((dbg, DBG_DECIDE, "Spill %+F after node %+F\n", val,
 377                                                 after_pos));
 378                                         be_add_spill(senv, val, after_pos);
 379                                 }
 380                         }
 381                 }
 382
 383                 /* kill the last 'demand' entries in the array */
 384                 workset_set_length(ws, len - spills_needed);
 385         }
 386
 387         /* 3. Insert the new values into the workset */
 388         for (i = 0; i < demand; ++i) {
 389                 ir_node *val = to_insert[i];
 390
 391                 workset_insert(ws, val, spilled[i]);
 392         }
 393 }
 394
 395 enum {
 396         AVAILABLE_EVERYWHERE,
 397         AVAILABLE_NOWHERE,
 398         AVAILABLE_PARTLY,
 399         AVAILABLE_UNKNOWN
 400 };
 401
 402 static unsigned available_in_all_preds(workset_t* const* pred_worksets,
 403                                        size_t n_pred_worksets,
 404                                        const ir_node *value, bool is_local_phi)
 405 {
 406         size_t i;
 407         bool   avail_everywhere = true;
 408         bool   avail_nowhere    = true;
 409
 410         assert(n_pred_worksets > 0);
 411
 412         /* value available in all preds? */
 413         for (i = 0; i < n_pred_worksets; ++i) {
 414                 bool             found     = false;
 415                 const workset_t *p_workset = pred_worksets[i];
 416                 int              p_len     = workset_get_length(p_workset);
 417                 int              p_i;
 418                 const ir_node   *l_value;
 419
 420                 if (is_local_phi) {
 421                         assert(is_Phi(value));
 422                         l_value = get_irn_n(value, i);
 423                 } else {
 424                         l_value = value;
 425                 }
 426
 427                 for (p_i = 0; p_i < p_len; ++p_i) {
 428                         const loc_t *p_l = &p_workset->vals[p_i];
 429                         if (p_l->node != l_value)
 430                                 continue;
 431
 432                         found = true;
 433                         break;
 434                 }
 435
 436                 if (found) {
 437                         avail_nowhere = false;
 438                 } else {
 439                         avail_everywhere = false;
 440                 }
 441         }
 442
 443         if (avail_everywhere) {
 444                 assert(!avail_nowhere);
 445                 return AVAILABLE_EVERYWHERE;
 446         } else if (avail_nowhere) {
 447                 return AVAILABLE_NOWHERE;
 448         } else {
 449                 return AVAILABLE_PARTLY;
 450         }
 451 }
 452
 453 /** Decides whether a specific node should be in the start workset or not
 454  *
 455  * @param env      belady environment
 456  * @param first
 457  * @param node     the node to test
 458  * @param loop     the loop of the node
 459  */
 460 static loc_t to_take_or_not_to_take(ir_node* first, ir_node *node,
 461                                     ir_loop *loop, unsigned available)
 462 {
 463         be_next_use_t next_use;
 464         loc_t         loc;
 465
 466         loc.time    = USES_INFINITY;
 467         loc.node    = node;
 468         loc.spilled = false;
 469
 470         /* We have to keep nonspillable nodes in the workingset */
 471         if (arch_get_irn_flags(skip_Proj_const(node)) & arch_irn_flags_dont_spill) {
 472                 loc.time = 0;
 473                 DB((dbg, DBG_START, "    %+F taken (dontspill node)\n", node, loc.time));
 474                 return loc;
 475         }
 476
 477         next_use = be_get_next_use(uses, first, node, 0);
 478         if (USES_IS_INFINITE(next_use.time)) {
 479                 /* the nodes marked as live in shouldn't be dead, so it must be a phi */
 480                 assert(is_Phi(node));
 481                 loc.time = USES_INFINITY;
 482                 DB((dbg, DBG_START, "    %+F not taken (dead)\n", node));
 483                 return loc;
 484         }
 485
 486         loc.time = next_use.time;
 487
 488         if (improve_known_preds) {
 489                 if (available == AVAILABLE_EVERYWHERE) {
 490                         DB((dbg, DBG_START, "    %+F taken (%u, live in all preds)\n",
 491                             node, loc.time));
 492                         return loc;
 493                 } else if (available == AVAILABLE_NOWHERE) {
 494                         DB((dbg, DBG_START, "    %+F not taken (%u, live in no pred)\n",
 495                             node, loc.time));
 496                         loc.time = USES_INFINITY;
 497                         return loc;
 498                 }
 499         }
 500
 501         if (!respectloopdepth || next_use.outermost_loop >= get_loop_depth(loop)) {
 502                 DB((dbg, DBG_START, "    %+F taken (%u, loop %d)\n", node, loc.time,
 503                     next_use.outermost_loop));
 504         } else {
 505                 loc.time = USES_PENDING;
 506                 DB((dbg, DBG_START, "    %+F delayed (outerdepth %d < loopdepth %d)\n",
 507                     node, next_use.outermost_loop, get_loop_depth(loop)));
 508         }
 509
 510         return loc;
 511 }
 512
 513 /**
 514  * Computes the start-workset for a block with multiple predecessors. We assume
 515  * that at least 1 of the predeccesors is a back-edge which means we're at the
 516  * beginning of a loop. We try to reload as much values as possible now so they
 517  * don't get reloaded inside the loop.
 518  */
 519 static void decide_start_workset(ir_node *const block)
 520 {
 521         ir_loop    *loop = get_irn_loop(block);
 522         ir_node    *first;
 523         loc_t       loc;
 524         loc_t      *starters;
 525         loc_t      *delayed;
 526         unsigned    len;
 527         unsigned    i;
 528         unsigned    ws_count;
 529         int         free_slots, free_pressure_slots;
 530         unsigned    pressure;
 531         int         arity;
 532         workset_t **pred_worksets;
 533         bool        all_preds_known;
 534
 535         /* check predecessors */
 536         arity           = get_irn_arity(block);
 537         pred_worksets   = ALLOCAN(workset_t*, arity);
 538         all_preds_known = true;
 539         for (int in = 0; in < arity; ++in) {
 540                 ir_node      *pred_block = get_Block_cfgpred_block(block, in);
 541                 block_info_t *pred_info  = get_block_info(pred_block);
 542
 543                 if (pred_info == NULL) {
 544                         pred_worksets[in] = NULL;
 545                         all_preds_known   = false;
 546                 } else {
 547                         pred_worksets[in] = pred_info->end_workset;
 548                 }
 549         }
 550
 551         /* Collect all values living at start of block */
 552         starters = NEW_ARR_F(loc_t, 0);
 553         delayed  = NEW_ARR_F(loc_t, 0);
 554
 555         DB((dbg, DBG_START, "Living at start of %+F:\n", block));
 556         first = sched_first(block);
 557
 558         /* check all Phis first */
 559         sched_foreach(block, node) {
 560                 unsigned available;
 561
 562                 if (! is_Phi(node))
 563                         break;
 564                 if (!arch_irn_consider_in_reg_alloc(cls, node))
 565                         continue;
 566
 567                 if (all_preds_known) {
 568                         available = available_in_all_preds(pred_worksets, arity, node, true);
 569                 } else {
 570                         available = AVAILABLE_UNKNOWN;
 571                 }
 572
 573                 loc = to_take_or_not_to_take(first, node, loop, available);
 574
 575                 if (! USES_IS_INFINITE(loc.time)) {
 576                         if (USES_IS_PENDING(loc.time))
 577                                 ARR_APP1(loc_t, delayed, loc);
 578                         else
 579                                 ARR_APP1(loc_t, starters, loc);
 580                 } else {
 581                         be_spill_phi(senv, node);
 582                 }
 583         }
 584
 585         /* check all Live-Ins */
 586         be_lv_foreach_cls(lv, block, be_lv_state_in, cls, node) {
 587                 unsigned available;
 588                 if (all_preds_known) {
 589                         available = available_in_all_preds(pred_worksets, arity, node, false);
 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                 }
 602         }
 603
 604         pressure            = be_get_loop_pressure(loop_ana, cls, loop);
 605         assert(ARR_LEN(delayed) <= pressure);
 606         free_slots          = n_regs - ARR_LEN(starters);
 607         free_pressure_slots = n_regs - (pressure - ARR_LEN(delayed));
 608         free_slots          = MIN(free_slots, free_pressure_slots);
 609
 610         /* so far we only put nodes into the starters list that are used inside
 611          * the loop. If register pressure in the loop is low then we can take some
 612          * values and let them live through the loop */
 613         DB((dbg, DBG_START, "Loop pressure %d, taking %d delayed vals\n",
 614             pressure, free_slots));
 615         if (free_slots > 0) {
 616                 size_t i;
 617
 618                 qsort(delayed, ARR_LEN(delayed), sizeof(delayed[0]), loc_compare);
 619
 620                 for (i = 0; i < ARR_LEN(delayed) && free_slots > 0; ++i) {
 621                         int    p, arity;
 622                         loc_t *loc = & delayed[i];
 623
 624                         if (!is_Phi(loc->node)) {
 625                                 /* don't use values which are dead in a known predecessors
 626                                  * to not induce unnecessary reloads */
 627                                 arity = get_irn_arity(block);
 628                                 for (p = 0; p < arity; ++p) {
 629                                         ir_node      *pred_block = get_Block_cfgpred_block(block, p);
 630                                         block_info_t *pred_info  = get_block_info(pred_block);
 631
 632                                         if (pred_info == NULL)
 633                                                 continue;
 634
 635                                         if (!workset_contains(pred_info->end_workset, loc->node)) {
 636                                                 DB((dbg, DBG_START,
 637                                                         "    delayed %+F not live at pred %+F\n", loc->node,
 638                                                         pred_block));
 639                                                 goto skip_delayed;
 640                                         }
 641                                 }
 642                         }
 643
 644                         DB((dbg, DBG_START, "    delayed %+F taken\n", loc->node));
 645                         ARR_APP1(loc_t, starters, *loc);
 646                         loc->node = NULL;
 647                         --free_slots;
 648                 skip_delayed:
 649                         ;
 650                 }
 651         }
 652
 653         /* spill phis (the actual phis not just their values) that are in this block
 654          * but not in the start workset */
 655         len = ARR_LEN(delayed);
 656         for (i = 0; i < len; ++i) {
 657                 ir_node *node = delayed[i].node;
 658                 if (node == NULL || !is_Phi(node) || get_nodes_block(node) != block)
 659                         continue;
 660
 661                 DB((dbg, DBG_START, "    spilling delayed phi %+F\n", node));
 662                 be_spill_phi(senv, node);
 663         }
 664         DEL_ARR_F(delayed);
 665
 666         /* Sort start values by first use */
 667         qsort(starters, ARR_LEN(starters), sizeof(starters[0]), loc_compare);
 668
 669         /* Copy the best ones from starters to start workset */
 670         ws_count = MIN((unsigned) ARR_LEN(starters), n_regs);
 671         workset_clear(ws);
 672         workset_bulk_fill(ws, ws_count, starters);
 673
 674         /* spill phis (the actual phis not just their values) that are in this block
 675          * but not in the start workset */
 676         len = ARR_LEN(starters);
 677         for (i = ws_count; i < len; ++i) {
 678                 ir_node *node = starters[i].node;
 679                 if (! is_Phi(node) || get_nodes_block(node) != block)
 680                         continue;
 681
 682                 DB((dbg, DBG_START, "    spilling phi %+F\n", node));
 683                 be_spill_phi(senv, node);
 684         }
 685
 686         DEL_ARR_F(starters);
 687
 688         /* determine spill status of the values: If there's 1 pred block (which
 689          * is no backedge) where the value is spilled then we must set it to
 690          * spilled here. */
 691         for (i = 0; i < ws_count; ++i) {
 692                 loc_t   *loc     = &ws->vals[i];
 693                 ir_node *value   = loc->node;
 694                 bool     spilled;
 695                 int      n;
 696
 697                 /* phis from this block aren't spilled */
 698                 if (get_nodes_block(value) == block) {
 699                         assert(is_Phi(value));
 700                         loc->spilled = false;
 701                         continue;
 702                 }
 703
 704                 /* determine if value was spilled on any predecessor */
 705                 spilled = false;
 706                 for (n = 0; n < arity; ++n) {
 707                         workset_t *pred_workset = pred_worksets[n];
 708                         int        p_len;
 709                         int        p;
 710
 711                         if (pred_workset == NULL)
 712                                 continue;
 713
 714                         p_len = workset_get_length(pred_workset);
 715                         for (p = 0; p < p_len; ++p) {
 716                                 loc_t *l = &pred_workset->vals[p];
 717
 718                                 if (l->node != value)
 719                                         continue;
 720
 721                                 if (l->spilled) {
 722                                         spilled = true;
 723                                 }
 724                                 break;
 725                         }
 726                 }
 727
 728                 loc->spilled = spilled;
 729         }
 730 }
 731
 732 /**
 733  * For the given block @p block, decide for each values
 734  * whether it is used from a register or is reloaded
 735  * before the use.
 736  */
 737 static void process_block(ir_node *block)
 738 {
 739         workset_t    *new_vals;
 740         unsigned      iter;
 741         block_info_t *block_info;
 742         int           arity;
 743
 744         /* no need to process a block twice */
 745         assert(get_block_info(block) == NULL);
 746
 747         /* construct start workset */
 748         arity = get_Block_n_cfgpreds(block);
 749         if (arity == 0) {
 750                 /* no predecessor -> empty set */
 751                 workset_clear(ws);
 752         } else if (arity == 1) {
 753                 /* one predecessor, copy its end workset */
 754                 ir_node      *pred_block = get_Block_cfgpred_block(block, 0);
 755                 block_info_t *pred_info  = get_block_info(pred_block);
 756
 757                 assert(pred_info != NULL);
 758                 workset_copy(ws, pred_info->end_workset);
 759         } else {
 760                 /* multiple predecessors, do more advanced magic :) */
 761                 decide_start_workset(block);
 762         }
 763
 764         DB((dbg, DBG_DECIDE, "\n"));
 765         DB((dbg, DBG_DECIDE, "Decide for %+F\n", block));
 766
 767         block_info = new_block_info();
 768         set_block_info(block, block_info);
 769
 770         DB((dbg, DBG_WSETS, "Start workset for %+F:\n", block));
 771         {
 772                 ir_node *irn;
 773                 workset_foreach(ws, irn, iter) {
 774                         DB((dbg, DBG_WSETS, "  %+F (%u)\n", irn, workset_get_time(ws, iter)));
 775                 }
 776         }
 777
 778         block_info->start_workset = workset_clone(ws);
 779
 780         /* process the block from start to end */
 781         DB((dbg, DBG_WSETS, "Processing...\n"));
 782         /* TODO: this leaks (into the obstack)... */
 783         new_vals = new_workset();
 784
 785         sched_foreach(block, irn) {
 786                 assert(workset_get_length(ws) <= n_regs);
 787
 788                 /* Phis are no real instr (see insert_starters()) */
 789                 if (is_Phi(irn)) {
 790                         continue;
 791                 }
 792                 DB((dbg, DBG_DECIDE, "  ...%+F\n", irn));
 793
 794                 /* allocate all values _used_ by this instruction */
 795                 workset_clear(new_vals);
 796                 be_foreach_use(irn, cls, in_req_, in, in_req,
 797                         /* (note that "spilled" is irrelevant here) */
 798                         workset_insert(new_vals, in, false);
 799                 );
 800                 displace(new_vals, 1, irn);
 801
 802                 /* allocate all values _defined_ by this instruction */
 803                 workset_clear(new_vals);
 804                 be_foreach_definition(irn, cls, value, req,
 805                         assert(req->width == 1);
 806                         workset_insert(new_vals, value, false);
 807                 );
 808                 displace(new_vals, 0, irn);
 809         }
 810
 811         /* Remember end-workset for this block */
 812         block_info->end_workset = workset_clone(ws);
 813         DB((dbg, DBG_WSETS, "End workset for %+F:\n", block));
 814         {
 815                 ir_node *irn;
 816                 workset_foreach(ws, irn, iter)
 817                         DB((dbg, DBG_WSETS, "  %+F (%u)\n", irn, workset_get_time(ws, iter)));
 818         }
 819 }
 820
 821 /**
 822  * 'decide' is block-local and makes assumptions
 823  * about the set of live-ins. Thus we must adapt the
 824  * live-outs to the live-ins at each block-border.
 825  */
 826 static void fix_block_borders(ir_node *block, void *data)
 827 {
 828         workset_t *start_workset;
 829         int        arity;
 830         int        i;
 831         unsigned   iter;
 832         (void) data;
 833
 834         DB((dbg, DBG_FIX, "\n"));
 835         DB((dbg, DBG_FIX, "Fixing %+F\n", block));
 836
 837         arity = get_irn_arity(block);
 838         /* can happen for endless loops */
 839         if (arity == 0)
 840                 return;
 841
 842         start_workset = get_block_info(block)->start_workset;
 843
 844         /* process all pred blocks */
 845         for (i = 0; i < arity; ++i) {
 846                 ir_node   *pred = get_Block_cfgpred_block(block, i);
 847                 workset_t *pred_end_workset = get_block_info(pred)->end_workset;
 848                 ir_node   *node;
 849
 850                 DB((dbg, DBG_FIX, "  Pred %+F\n", pred));
 851
 852                 /* spill all values not used anymore */
 853                 workset_foreach(pred_end_workset, node, iter) {
 854                         ir_node *n2;
 855                         unsigned iter2;
 856                         bool     found = false;
 857                         workset_foreach(start_workset, n2, iter2) {
 858                                 if (n2 == node) {
 859                                         found = true;
 860                                         break;
 861                                 }
 862                                 /* note that we do not look at phi inputs, becuase the values
 863                                  * will be either live-end and need no spill or
 864                                  * they have other users in which must be somewhere else in the
 865                                  * workset */
 866                         }
 867
 868                         if (found)
 869                                 continue;
 870
 871                         if (move_spills && be_is_live_in(lv, block, node)
 872                                         && !pred_end_workset->vals[iter].spilled) {
 873                                 ir_node *insert_point;
 874                                 if (arity > 1) {
 875                                         insert_point = be_get_end_of_block_insertion_point(pred);
 876                                         insert_point = sched_prev(insert_point);
 877                                 } else {
 878                                         insert_point = block;
 879                                 }
 880                                 DB((dbg, DBG_SPILL, "Spill %+F after %+F\n", node,
 881                                      insert_point));
 882                                 be_add_spill(senv, node, insert_point);
 883                         }
 884                 }
 885
 886                 /* reload missing values in predecessors, add missing spills */
 887                 workset_foreach(start_workset, node, iter) {
 888                         const loc_t *l    = &start_workset->vals[iter];
 889                         const loc_t *pred_loc;
 890
 891                         /* if node is a phi of the current block we reload
 892                          * the corresponding argument, else node itself */
 893                         if (is_Phi(node) && get_nodes_block(node) == block) {
 894                                 node = get_irn_n(node, i);
 895                                 assert(!l->spilled);
 896
 897                                 /* we might have unknowns as argument for the phi */
 898                                 if (!arch_irn_consider_in_reg_alloc(cls, node))
 899                                         continue;
 900                         }
 901
 902                         /* check if node is in a register at end of pred */
 903                         pred_loc = workset_contains(pred_end_workset, node);
 904                         if (pred_loc != NULL) {
 905                                 /* we might have to spill value on this path */
 906                                 if (move_spills && !pred_loc->spilled && l->spilled) {
 907                                         ir_node *insert_point
 908                                                 = be_get_end_of_block_insertion_point(pred);
 909                                         insert_point = sched_prev(insert_point);
 910                                         DB((dbg, DBG_SPILL, "Spill %+F after %+F\n", node,
 911                                             insert_point));
 912                                         be_add_spill(senv, node, insert_point);
 913                                 }
 914                         } else {
 915                                 /* node is not in register at the end of pred -> reload it */
 916                                 DB((dbg, DBG_FIX, "    reload %+F\n", node));
 917                                 DB((dbg, DBG_SPILL, "Reload %+F before %+F,%d\n", node, block, i));
 918                                 be_add_reload_on_edge(senv, node, block, i, cls, 1);
 919                         }
 920                 }
 921         }
 922 }
 923
 924 static void be_spill_belady(ir_graph *irg, const arch_register_class_t *rcls)
 925 {
 926         int i;
 927
 928         be_assure_live_sets(irg);
 929
 930         stat_ev_tim_push();
 931         assure_loopinfo(irg);
 932         stat_ev_tim_pop("belady_time_backedges");
 933
 934         stat_ev_tim_push();
 935         be_clear_links(irg);
 936         stat_ev_tim_pop("belady_time_clear_links");
 937
 938         ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
 939
 940         /* init belady env */
 941         stat_ev_tim_push();
 942         obstack_init(&obst);
 943         cls       = rcls;
 944         lv        = be_get_irg_liveness(irg);
 945         n_regs    = be_get_n_allocatable_regs(irg, cls);
 946         ws        = new_workset();
 947         uses      = be_begin_uses(irg, lv);
 948         loop_ana  = be_new_loop_pressure(irg, cls);
 949         senv      = be_new_spill_env(irg);
 950         blocklist = be_get_cfgpostorder(irg);
 951         stat_ev_tim_pop("belady_time_init");
 952
 953         stat_ev_tim_push();
 954         /* walk blocks in reverse postorder */
 955         for (i = ARR_LEN(blocklist) - 1; i >= 0; --i) {
 956                 process_block(blocklist[i]);
 957         }
 958         DEL_ARR_F(blocklist);
 959         stat_ev_tim_pop("belady_time_belady");
 960
 961         stat_ev_tim_push();
 962         /* belady was block-local, fix the global flow by adding reloads on the
 963          * edges */
 964         irg_block_walk_graph(irg, fix_block_borders, NULL, NULL);
 965         stat_ev_tim_pop("belady_time_fix_borders");
 966
 967         ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
 968
 969         /* Insert spill/reload nodes into the graph and fix usages */
 970         be_insert_spills_reloads(senv);
 971
 972         /* clean up */
 973         be_delete_spill_env(senv);
 974         be_end_uses(uses);
 975         be_free_loop_pressure(loop_ana);
 976         obstack_free(&obst, NULL);
 977 }
 978
 979 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady)
 980 void be_init_spillbelady(void)
 981 {
 982         static be_spiller_t belady_spiller = {
 983                 be_spill_belady
 984         };
 985         lc_opt_entry_t *be_grp       = lc_opt_get_grp(firm_opt_get_root(), "be");
 986         lc_opt_entry_t *belady_group = lc_opt_get_grp(be_grp, "belady");
 987         lc_opt_add_table(belady_group, options);
 988
 989         be_register_spiller("belady", &belady_spiller);
 990         FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
 991 }