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