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

   1 /*
   2  * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
   3  *
   4  * This file is part of libFirm.
   5  *
   6  * This file may be distributed and/or modified under the terms of the
   7  * GNU General Public License version 2 as published by the Free Software
   8  * Foundation and appearing in the file LICENSE.GPL included in the
   9  * packaging of this file.
  10  *
  11  * Licensees holding valid libFirm Professional Edition licenses may use
  12  * this file in accordance with the libFirm Commercial License.
  13  * Agreement provided with the Software.
  14  *
  15  * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
  16  * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  17  * PURPOSE.
  18  */
  19
  20 /**
  21  * @file
  22  * @brief       Beladys spillalgorithm.
  23  * @author      Daniel Grund, Matthias Braun
  24  * @date        20.09.2005
  25  * @version     $Id$
  26  */
  27 #include "config.h"
  28
  29 #include <stdbool.h>
  30
  31 #include "obst.h"
  32 #include "irprintf_t.h"
  33 #include "irgraph.h"
  34 #include "irnode.h"
  35 #include "irmode.h"
  36 #include "irgwalk.h"
  37 #include "irloop.h"
  38 #include "iredges_t.h"
  39 #include "ircons_t.h"
  40 #include "irprintf.h"
  41 #include "irnodeset.h"
  42
  43 #include "beutil.h"
  44 #include "bearch.h"
  45 #include "beuses.h"
  46 #include "besched.h"
  47 #include "beirgmod.h"
  48 #include "belive_t.h"
  49 #include "benode.h"
  50 #include "bechordal_t.h"
  51 #include "bespill.h"
  52 #include "beloopana.h"
  53 #include "beirg.h"
  54 #include "bespillutil.h"
  55 #include "bemodule.h"
  56
  57 #define DBG_SPILL     1
  58 #define DBG_WSETS     2
  59 #define DBG_FIX       4
  60 #define DBG_DECIDE    8
  61 #define DBG_START    16
  62 #define DBG_SLOTS    32
  63 #define DBG_TRACE    64
  64 #define DBG_WORKSET 128
  65 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  66
  67 #define TIME_UNDEFINED 6666
  68
  69 /**
  70  * An association between a node and a point in time.
  71  */
  72 typedef struct loc_t {
  73         ir_node          *node;
  74         unsigned          time;     /**< A use time (see beuses.h). */
  75         bool              spilled;  /**< value was already spilled on this path */
  76 } loc_t;
  77
  78 typedef struct _workset_t {
  79         unsigned len;     /**< current length */
  80         loc_t    vals[0]; /**< array of the values/distances in this working set */
  81 } workset_t;
  82
  83 static struct obstack               obst;
  84 static const arch_register_class_t *cls;
  85 static const be_lv_t               *lv;
  86 static be_loopana_t                *loop_ana;
  87 static unsigned                     n_regs;
  88 static workset_t                   *ws;     /**< the main workset used while
  89                                                      processing a block. */
  90 static be_uses_t                   *uses;   /**< env for the next-use magic */
  91 static ir_node                     *instr;  /**< current instruction */
  92 static unsigned                     instr_nr; /**< current instruction number
  93                                                        (relative to block start) */
  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 = a;
 217         const loc_t *q = b;
 218         return p->time - q->time;
 219 }
 220
 221 static void workset_sort(workset_t *workset)
 222 {
 223         qsort(workset->vals, workset->len, sizeof(workset->vals[0]), loc_compare);
 224 }
 225
 226 static inline unsigned workset_get_time(const workset_t *workset, unsigned idx)
 227 {
 228         return workset->vals[idx].time;
 229 }
 230
 231 static inline void workset_set_time(workset_t *workset, unsigned idx,
 232                                     unsigned time)
 233 {
 234         workset->vals[idx].time = time;
 235 }
 236
 237 static inline unsigned workset_get_length(const workset_t *workset)
 238 {
 239         return workset->len;
 240 }
 241
 242 static inline void workset_set_length(workset_t *workset, unsigned len)
 243 {
 244         workset->len = len;
 245 }
 246
 247 static inline ir_node *workset_get_val(const workset_t *workset, unsigned idx)
 248 {
 249         return workset->vals[idx].node;
 250 }
 251
 252 /**
 253  * Iterates over all values in the working set.
 254  * @p ws The workset to iterate
 255  * @p v  A variable to put the current value in
 256  * @p i  An integer for internal use
 257  */
 258 #define workset_foreach(ws, v, i)       \
 259         for (i=0; v=(i < ws->len) ? ws->vals[i].node : NULL, i < ws->len; ++i)
 260
 261 typedef struct _block_info_t
 262 {
 263         workset_t *start_workset;
 264         workset_t *end_workset;
 265 } block_info_t;
 266
 267 static block_info_t *new_block_info(void)
 268 {
 269         return OALLOCZ(&obst, block_info_t);
 270 }
 271
 272 static inline block_info_t *get_block_info(const ir_node *block)
 273 {
 274         return get_irn_link(block);
 275 }
 276
 277 static inline void set_block_info(ir_node *block, block_info_t *info)
 278 {
 279         set_irn_link(block, info);
 280 }
 281
 282 /**
 283  * @return The distance to the next use or 0 if irn has dont_spill flag set
 284  */
 285 static unsigned get_distance(ir_node *from, unsigned from_step,
 286                              const ir_node *def, int skip_from_uses)
 287 {
 288         be_next_use_t use;
 289         unsigned      costs;
 290         unsigned      time;
 291
 292         assert(!arch_irn_is_ignore(def));
 293
 294         use  = be_get_next_use(uses, from, from_step, def, skip_from_uses);
 295         time = use.time;
 296         if (USES_IS_INFINITE(time))
 297                 return USES_INFINITY;
 298
 299         /* We have to keep nonspillable nodes in the workingset */
 300         if (arch_irn_get_flags(skip_Proj_const(def)) & arch_irn_flags_dont_spill)
 301                 return 0;
 302
 303         /* give some bonus to rematerialisable nodes */
 304         if (remat_bonus > 0) {
 305                 costs = be_get_reload_costs_no_weight(senv, def, use.before);
 306                 assert(costs * remat_bonus < 1000);
 307                 time  += 1000 - (costs * remat_bonus);
 308         }
 309
 310         return time;
 311 }
 312
 313 /**
 314  * Performs the actions necessary to grant the request that:
 315  * - new_vals can be held in registers
 316  * - as few as possible other values are disposed
 317  * - the worst values get disposed
 318  *
 319  * @p is_usage indicates that the values in new_vals are used (not defined)
 320  * In this case reloads must be performed
 321  */
 322 static void displace(workset_t *new_vals, int is_usage)
 323 {
 324         ir_node **to_insert = ALLOCAN(ir_node*, n_regs);
 325         bool     *spilled   = ALLOCAN(bool,     n_regs);
 326         ir_node  *val;
 327         int       i;
 328         int       len;
 329         int       spills_needed;
 330         int       demand;
 331         unsigned  iter;
 332
 333         /* 1. Identify the number of needed slots and the values to reload */
 334         demand = 0;
 335         workset_foreach(new_vals, val, iter) {
 336                 bool reloaded = false;
 337
 338                 if (! workset_contains(ws, val)) {
 339                         DB((dbg, DBG_DECIDE, "    insert %+F\n", val));
 340                         if (is_usage) {
 341                                 DB((dbg, DBG_SPILL, "Reload %+F before %+F\n", val, instr));
 342                                 be_add_reload(senv, val, instr, cls, 1);
 343                                 reloaded = true;
 344                         }
 345                 } else {
 346                         DB((dbg, DBG_DECIDE, "    %+F already in workset\n", val));
 347                         assert(is_usage);
 348                         /* remove the value from the current workset so it is not accidently
 349                          * spilled */
 350                         workset_remove(ws, val);
 351                 }
 352                 spilled[demand]   = reloaded;
 353                 to_insert[demand] = val;
 354                 ++demand;
 355         }
 356
 357         /* 2. Make room for at least 'demand' slots */
 358         len           = workset_get_length(ws);
 359         spills_needed = len + demand - n_regs;
 360         assert(spills_needed <= len);
 361
 362         /* Only make more free room if we do not have enough */
 363         if (spills_needed > 0) {
 364                 DB((dbg, DBG_DECIDE, "    disposing %d values\n", spills_needed));
 365
 366                 /* calculate current next-use distance for live values */
 367                 for (i = 0; i < len; ++i) {
 368                         ir_node  *val  = workset_get_val(ws, i);
 369                         unsigned  dist = get_distance(instr, instr_nr, val, !is_usage);
 370                         workset_set_time(ws, i, dist);
 371                 }
 372
 373                 /* sort entries by increasing nextuse-distance*/
 374                 workset_sort(ws);
 375
 376                 for (i = len - spills_needed; i < len; ++i) {
 377                         ir_node *val = ws->vals[i].node;
 378
 379                         DB((dbg, DBG_DECIDE, "    disposing node %+F (%u)\n", val,
 380                              workset_get_time(ws, i)));
 381
 382                         if (move_spills) {
 383                                 if (!USES_IS_INFINITE(ws->vals[i].time)
 384                                                 && !ws->vals[i].spilled) {
 385                                         ir_node *after_pos = sched_prev(instr);
 386                                         DB((dbg, DBG_DECIDE, "Spill %+F after node %+F\n", val,
 387                                                 after_pos));
 388                                         be_add_spill(senv, val, after_pos);
 389                                 }
 390                         }
 391                 }
 392
 393                 /* kill the last 'demand' entries in the array */
 394                 workset_set_length(ws, len - spills_needed);
 395         }
 396
 397         /* 3. Insert the new values into the workset */
 398         for (i = 0; i < demand; ++i) {
 399                 ir_node *val = to_insert[i];
 400
 401                 workset_insert(ws, val, spilled[i]);
 402         }
 403 }
 404
 405 enum {
 406         AVAILABLE_EVERYWHERE,
 407         AVAILABLE_NOWHERE,
 408         AVAILABLE_PARTLY,
 409         AVAILABLE_UNKNOWN
 410 };
 411
 412 static unsigned available_in_all_preds(workset_t* const* pred_worksets,
 413                                        size_t n_pred_worksets,
 414                                        const ir_node *value, bool is_local_phi)
 415 {
 416         size_t i;
 417         bool   avail_everywhere = true;
 418         bool   avail_nowhere    = true;
 419
 420         assert(n_pred_worksets > 0);
 421
 422         /* value available in all preds? */
 423         for (i = 0; i < n_pred_worksets; ++i) {
 424                 bool             found     = false;
 425                 const workset_t *p_workset = pred_worksets[i];
 426                 int              p_len     = workset_get_length(p_workset);
 427                 int              p_i;
 428                 const ir_node   *l_value;
 429
 430                 if (is_local_phi) {
 431                         assert(is_Phi(value));
 432                         l_value = get_irn_n(value, i);
 433                 } else {
 434                         l_value = value;
 435                 }
 436
 437                 for (p_i = 0; p_i < p_len; ++p_i) {
 438                         const loc_t *p_l = &p_workset->vals[p_i];
 439                         if (p_l->node != l_value)
 440                                 continue;
 441
 442                         found = true;
 443                         break;
 444                 }
 445
 446                 if (found) {
 447                         avail_nowhere = false;
 448                 } else {
 449                         avail_everywhere = false;
 450                 }
 451         }
 452
 453         if (avail_everywhere) {
 454                 assert(!avail_nowhere);
 455                 return AVAILABLE_EVERYWHERE;
 456         } else if (avail_nowhere) {
 457                 return AVAILABLE_NOWHERE;
 458         } else {
 459                 return AVAILABLE_PARTLY;
 460         }
 461 }
 462
 463 /** Decides whether a specific node should be in the start workset or not
 464  *
 465  * @param env      belady environment
 466  * @param first
 467  * @param node     the node to test
 468  * @param loop     the loop of the node
 469  */
 470 static loc_t to_take_or_not_to_take(ir_node* first, ir_node *node,
 471                                     ir_loop *loop, unsigned available)
 472 {
 473         be_next_use_t next_use;
 474         loc_t         loc;
 475
 476         loc.time    = USES_INFINITY;
 477         loc.node    = node;
 478         loc.spilled = false;
 479
 480         if (!arch_irn_consider_in_reg_alloc(cls, node)) {
 481                 loc.time = USES_INFINITY;
 482                 return loc;
 483         }
 484
 485         /* We have to keep nonspillable nodes in the workingset */
 486         if (arch_irn_get_flags(skip_Proj_const(node)) & arch_irn_flags_dont_spill) {
 487                 loc.time = 0;
 488                 DB((dbg, DBG_START, "    %+F taken (dontspill node)\n", node, loc.time));
 489                 return loc;
 490         }
 491
 492         next_use = be_get_next_use(uses, first, 0, node, 0);
 493         if (USES_IS_INFINITE(next_use.time)) {
 494                 /* the nodes marked as live in shouldn't be dead, so it must be a phi */
 495                 assert(is_Phi(node));
 496                 loc.time = USES_INFINITY;
 497                 DB((dbg, DBG_START, "    %+F not taken (dead)\n", node));
 498                 return loc;
 499         }
 500
 501         loc.time = next_use.time;
 502
 503         if (improve_known_preds) {
 504                 if (available == AVAILABLE_EVERYWHERE) {
 505                         DB((dbg, DBG_START, "    %+F taken (%u, live in all preds)\n",
 506                             node, loc.time));
 507                         return loc;
 508                 } else if (available == AVAILABLE_NOWHERE) {
 509                         DB((dbg, DBG_START, "    %+F not taken (%u, live in no pred)\n",
 510                             node, loc.time));
 511                         loc.time = USES_INFINITY;
 512                         return loc;
 513                 }
 514         }
 515
 516         if (!respectloopdepth || next_use.outermost_loop >= get_loop_depth(loop)) {
 517                 DB((dbg, DBG_START, "    %+F taken (%u, loop %d)\n", node, loc.time,
 518                     next_use.outermost_loop));
 519         } else {
 520                 loc.time = USES_PENDING;
 521                 DB((dbg, DBG_START, "    %+F delayed (outerdepth %d < loopdepth %d)\n",
 522                     node, next_use.outermost_loop, get_loop_depth(loop)));
 523         }
 524
 525         return loc;
 526 }
 527
 528 /**
 529  * Computes the start-workset for a block with multiple predecessors. We assume
 530  * that at least 1 of the predeccesors is a back-edge which means we're at the
 531  * beginning of a loop. We try to reload as much values as possible now so they
 532  * don't get reloaded inside the loop.
 533  */
 534 static void decide_start_workset(const ir_node *block)
 535 {
 536         ir_loop    *loop = get_irn_loop(block);
 537         ir_node    *first;
 538         ir_node    *node;
 539         loc_t       loc;
 540         loc_t      *starters;
 541         loc_t      *delayed;
 542         unsigned    len;
 543         unsigned    i;
 544         int         in;
 545         unsigned    ws_count;
 546         int             free_slots, free_pressure_slots;
 547         unsigned    pressure;
 548         int         arity;
 549         workset_t **pred_worksets;
 550         bool        all_preds_known;
 551
 552         /* check predecessors */
 553         arity           = get_irn_arity(block);
 554         pred_worksets   = ALLOCAN(workset_t*, arity);
 555         all_preds_known = true;
 556         for (in = 0; in < arity; ++in) {
 557                 ir_node      *pred_block = get_Block_cfgpred_block(block, in);
 558                 block_info_t *pred_info  = get_block_info(pred_block);
 559
 560                 if (pred_info == NULL) {
 561                         pred_worksets[in] = NULL;
 562                         all_preds_known   = false;
 563                 } else {
 564                         pred_worksets[in] = pred_info->end_workset;
 565                 }
 566         }
 567
 568         /* Collect all values living at start of block */
 569         starters = NEW_ARR_F(loc_t, 0);
 570         delayed  = NEW_ARR_F(loc_t, 0);
 571
 572         DB((dbg, DBG_START, "Living at start of %+F:\n", block));
 573         first = sched_first(block);
 574
 575         /* check all Phis first */
 576         sched_foreach(block, node) {
 577                 unsigned available;
 578
 579                 if (! is_Phi(node))
 580                         break;
 581                 if (!arch_irn_consider_in_reg_alloc(cls, node))
 582                         continue;
 583
 584                 if (all_preds_known) {
 585                         available = available_in_all_preds(pred_worksets, arity, node, true);
 586                 } else {
 587                         available = AVAILABLE_UNKNOWN;
 588                 }
 589
 590                 loc = to_take_or_not_to_take(first, node, loop, available);
 591
 592                 if (! USES_IS_INFINITE(loc.time)) {
 593                         if (USES_IS_PENDING(loc.time))
 594                                 ARR_APP1(loc_t, delayed, loc);
 595                         else
 596                                 ARR_APP1(loc_t, starters, loc);
 597                 } else {
 598                         be_spill_phi(senv, node);
 599                 }
 600         }
 601
 602         /* check all Live-Ins */
 603         be_lv_foreach(lv, block, be_lv_state_in, in) {
 604                 ir_node *node = be_lv_get_irn(lv, block, in);
 605                 unsigned available;
 606
 607                 if (all_preds_known) {
 608                         available = available_in_all_preds(pred_worksets, arity, node, false);
 609                 } else {
 610                         available = AVAILABLE_UNKNOWN;
 611                 }
 612
 613                 loc = to_take_or_not_to_take(first, node, loop, available);
 614
 615                 if (! USES_IS_INFINITE(loc.time)) {
 616                         if (USES_IS_PENDING(loc.time))
 617                                 ARR_APP1(loc_t, delayed, loc);
 618                         else
 619                                 ARR_APP1(loc_t, starters, loc);
 620                 }
 621         }
 622
 623         pressure            = be_get_loop_pressure(loop_ana, cls, loop);
 624         assert(ARR_LEN(delayed) <= (signed)pressure);
 625         free_slots          = n_regs - ARR_LEN(starters);
 626         free_pressure_slots = n_regs - (pressure - ARR_LEN(delayed));
 627         free_slots          = MIN(free_slots, free_pressure_slots);
 628
 629         /* so far we only put nodes into the starters list that are used inside
 630          * the loop. If register pressure in the loop is low then we can take some
 631          * values and let them live through the loop */
 632         DB((dbg, DBG_START, "Loop pressure %d, taking %d delayed vals\n",
 633             pressure, free_slots));
 634         if (free_slots > 0) {
 635                 int i;
 636                 qsort(delayed, ARR_LEN(delayed), sizeof(delayed[0]), loc_compare);
 637
 638                 for (i = 0; i < ARR_LEN(delayed) && free_slots > 0; ++i) {
 639                         int    p, arity;
 640                         loc_t *loc = & delayed[i];
 641
 642                         if (!is_Phi(loc->node)) {
 643                                 /* don't use values which are dead in a known predecessors
 644                                  * to not induce unnecessary reloads */
 645                                 arity = get_irn_arity(block);
 646                                 for (p = 0; p < arity; ++p) {
 647                                         ir_node      *pred_block = get_Block_cfgpred_block(block, p);
 648                                         block_info_t *pred_info  = get_block_info(pred_block);
 649
 650                                         if (pred_info == NULL)
 651                                                 continue;
 652
 653                                         if (!workset_contains(pred_info->end_workset, loc->node)) {
 654                                                 DB((dbg, DBG_START,
 655                                                         "    delayed %+F not live at pred %+F\n", loc->node,
 656                                                         pred_block));
 657                                                 goto skip_delayed;
 658                                         }
 659                                 }
 660                         }
 661
 662                         DB((dbg, DBG_START, "    delayed %+F taken\n", loc->node));
 663                         ARR_APP1(loc_t, starters, *loc);
 664                         loc->node = NULL;
 665                         --free_slots;
 666                 skip_delayed:
 667                         ;
 668                 }
 669         }
 670
 671         /* spill phis (the actual phis not just their values) that are in this block
 672          * but not in the start workset */
 673         len = ARR_LEN(delayed);
 674         for (i = 0; i < len; ++i) {
 675                 ir_node *node = delayed[i].node;
 676                 if (node == NULL || !is_Phi(node) || get_nodes_block(node) != block)
 677                         continue;
 678
 679                 DB((dbg, DBG_START, "    spilling delayed phi %+F\n", node));
 680                 be_spill_phi(senv, node);
 681         }
 682         DEL_ARR_F(delayed);
 683
 684         /* Sort start values by first use */
 685         qsort(starters, ARR_LEN(starters), sizeof(starters[0]), loc_compare);
 686
 687         /* Copy the best ones from starters to start workset */
 688         ws_count = MIN((unsigned) ARR_LEN(starters), n_regs);
 689         workset_clear(ws);
 690         workset_bulk_fill(ws, ws_count, starters);
 691
 692         /* spill phis (the actual phis not just their values) that are in this block
 693          * but not in the start workset */
 694         len = ARR_LEN(starters);
 695         for (i = ws_count; i < len; ++i) {
 696                 ir_node *node = starters[i].node;
 697                 if (! is_Phi(node) || get_nodes_block(node) != block)
 698                         continue;
 699
 700                 DB((dbg, DBG_START, "    spilling phi %+F\n", node));
 701                 be_spill_phi(senv, node);
 702         }
 703
 704         DEL_ARR_F(starters);
 705
 706         /* determine spill status of the values: If there's 1 pred block (which
 707          * is no backedge) where the value is spilled then we must set it to
 708          * spilled here. */
 709         for (i = 0; i < ws_count; ++i) {
 710                 loc_t   *loc     = &ws->vals[i];
 711                 ir_node *value   = loc->node;
 712                 bool     spilled;
 713                 int      n;
 714
 715                 /* phis from this block aren't spilled */
 716                 if (get_nodes_block(value) == block) {
 717                         assert(is_Phi(value));
 718                         loc->spilled = false;
 719                         continue;
 720                 }
 721
 722                 /* determine if value was spilled on any predecessor */
 723                 spilled = false;
 724                 for (n = 0; n < arity; ++n) {
 725                         workset_t *pred_workset = pred_worksets[n];
 726                         int        p_len;
 727                         int        p;
 728
 729                         if (pred_workset == NULL)
 730                                 continue;
 731
 732                         p_len = workset_get_length(pred_workset);
 733                         for (p = 0; p < p_len; ++p) {
 734                                 loc_t *l = &pred_workset->vals[p];
 735
 736                                 if (l->node != value)
 737                                         continue;
 738
 739                                 if (l->spilled) {
 740                                         spilled = true;
 741                                 }
 742                                 break;
 743                         }
 744                 }
 745
 746                 loc->spilled = spilled;
 747         }
 748 }
 749
 750 /**
 751  * For the given block @p block, decide for each values
 752  * whether it is used from a register or is reloaded
 753  * before the use.
 754  */
 755 static void process_block(ir_node *block)
 756 {
 757         workset_t    *new_vals;
 758         ir_node      *irn;
 759         unsigned      iter;
 760         block_info_t *block_info;
 761         int           arity;
 762
 763         /* no need to process a block twice */
 764         assert(get_block_info(block) == NULL);
 765
 766         /* construct start workset */
 767         arity = get_Block_n_cfgpreds(block);
 768         if (arity == 0) {
 769                 /* no predecessor -> empty set */
 770                 workset_clear(ws);
 771         } else if (arity == 1) {
 772                 /* one predecessor, copy it's end workset */
 773                 ir_node      *pred_block = get_Block_cfgpred_block(block, 0);
 774                 block_info_t *pred_info  = get_block_info(pred_block);
 775
 776                 assert(pred_info != NULL);
 777                 workset_copy(ws, pred_info->end_workset);
 778         } else {
 779                 /* multiple predecessors, do more advanced magic :) */
 780                 decide_start_workset(block);
 781         }
 782
 783         DB((dbg, DBG_DECIDE, "\n"));
 784         DB((dbg, DBG_DECIDE, "Decide for %+F\n", block));
 785
 786         block_info = new_block_info();
 787         set_block_info(block, block_info);
 788
 789         DB((dbg, DBG_WSETS, "Start workset for %+F:\n", block));
 790         workset_foreach(ws, irn, iter) {
 791                 DB((dbg, DBG_WSETS, "  %+F (%u)\n", irn,
 792                      workset_get_time(ws, iter)));
 793         }
 794
 795         block_info->start_workset = workset_clone(ws);
 796
 797         /* process the block from start to end */
 798         DB((dbg, DBG_WSETS, "Processing...\n"));
 799         instr_nr = 0;
 800         /* TODO: this leaks (into the obstack)... */
 801         new_vals = new_workset();
 802
 803         sched_foreach(block, irn) {
 804                 int i, arity;
 805                 ir_node *value;
 806                 assert(workset_get_length(ws) <= n_regs);
 807
 808                 /* Phis are no real instr (see insert_starters()) */
 809                 if (is_Phi(irn)) {
 810                         continue;
 811                 }
 812                 DB((dbg, DBG_DECIDE, "  ...%+F\n", irn));
 813
 814                 /* set instruction in the workset */
 815                 instr = irn;
 816
 817                 /* allocate all values _used_ by this instruction */
 818                 workset_clear(new_vals);
 819                 for (i = 0, arity = get_irn_arity(irn); i < arity; ++i) {
 820                         ir_node *in = get_irn_n(irn, i);
 821                         if (!arch_irn_consider_in_reg_alloc(cls, in))
 822                                 continue;
 823
 824                         /* (note that "spilled" is irrelevant here) */
 825                         workset_insert(new_vals, in, false);
 826                 }
 827                 displace(new_vals, 1);
 828
 829                 /* allocate all values _defined_ by this instruction */
 830                 workset_clear(new_vals);
 831                 be_foreach_definition(irn, cls, value,
 832                         assert(req_->width == 1);
 833                         workset_insert(new_vals, value, false);
 834                 );
 835                 displace(new_vals, 0);
 836
 837                 instr_nr++;
 838         }
 839
 840         /* Remember end-workset for this block */
 841         block_info->end_workset = workset_clone(ws);
 842         DB((dbg, DBG_WSETS, "End workset for %+F:\n", block));
 843         workset_foreach(ws, irn, iter)
 844                 DB((dbg, DBG_WSETS, "  %+F (%u)\n", irn, workset_get_time(ws, iter)));
 845 }
 846
 847 /**
 848  * 'decide' is block-local and makes assumptions
 849  * about the set of live-ins. Thus we must adapt the
 850  * live-outs to the live-ins at each block-border.
 851  */
 852 static void fix_block_borders(ir_node *block, void *data)
 853 {
 854         workset_t *start_workset;
 855         int        arity;
 856         int        i;
 857         unsigned   iter;
 858         (void) data;
 859
 860         DB((dbg, DBG_FIX, "\n"));
 861         DB((dbg, DBG_FIX, "Fixing %+F\n", block));
 862
 863         arity = get_irn_arity(block);
 864         /* can happen for endless loops */
 865         if (arity == 0)
 866                 return;
 867
 868         start_workset = get_block_info(block)->start_workset;
 869
 870         /* process all pred blocks */
 871         for (i = 0; i < arity; ++i) {
 872                 ir_node   *pred = get_Block_cfgpred_block(block, i);
 873                 workset_t *pred_end_workset = get_block_info(pred)->end_workset;
 874                 ir_node   *node;
 875
 876                 DB((dbg, DBG_FIX, "  Pred %+F\n", pred));
 877
 878                 /* spill all values not used anymore */
 879                 workset_foreach(pred_end_workset, node, iter) {
 880                         ir_node *n2;
 881                         unsigned iter2;
 882                         bool     found = false;
 883                         workset_foreach(start_workset, n2, iter2) {
 884                                 if (n2 == node) {
 885                                         found = true;
 886                                         break;
 887                                 }
 888                                 /* note that we do not look at phi inputs, becuase the values
 889                                  * will be either live-end and need no spill or
 890                                  * they have other users in which must be somewhere else in the
 891                                  * workset */
 892                         }
 893
 894                         if (found)
 895                                 continue;
 896
 897                         if (move_spills && be_is_live_in(lv, block, node)
 898                                         && !pred_end_workset->vals[iter].spilled) {
 899                                 ir_node *insert_point;
 900                                 if (arity > 1) {
 901                                         insert_point = be_get_end_of_block_insertion_point(pred);
 902                                         insert_point = sched_prev(insert_point);
 903                                 } else {
 904                                         insert_point = block;
 905                                 }
 906                                 DB((dbg, DBG_SPILL, "Spill %+F after %+F\n", node,
 907                                      insert_point));
 908                                 be_add_spill(senv, node, insert_point);
 909                         }
 910                 }
 911
 912                 /* reload missing values in predecessors, add missing spills */
 913                 workset_foreach(start_workset, node, iter) {
 914                         const loc_t *l    = &start_workset->vals[iter];
 915                         const loc_t *pred_loc;
 916
 917                         /* if node is a phi of the current block we reload
 918                          * the corresponding argument, else node itself */
 919                         if (is_Phi(node) && get_nodes_block(node) == block) {
 920                                 node = get_irn_n(node, i);
 921                                 assert(!l->spilled);
 922
 923                                 /* we might have unknowns as argument for the phi */
 924                                 if (!arch_irn_consider_in_reg_alloc(cls, node))
 925                                         continue;
 926                         }
 927
 928                         /* check if node is in a register at end of pred */
 929                         pred_loc = workset_contains(pred_end_workset, node);
 930                         if (pred_loc != NULL) {
 931                                 /* we might have to spill value on this path */
 932                                 if (move_spills && !pred_loc->spilled && l->spilled) {
 933                                         ir_node *insert_point
 934                                                 = be_get_end_of_block_insertion_point(pred);
 935                                         insert_point = sched_prev(insert_point);
 936                                         DB((dbg, DBG_SPILL, "Spill %+F after %+F\n", node,
 937                                             insert_point));
 938                                         be_add_spill(senv, node, insert_point);
 939                                 }
 940                         } else {
 941                                 /* node is not in register at the end of pred -> reload it */
 942                                 DB((dbg, DBG_FIX, "    reload %+F\n", node));
 943                                 DB((dbg, DBG_SPILL, "Reload %+F before %+F,%d\n", node, block, i));
 944                                 be_add_reload_on_edge(senv, node, block, i, cls, 1);
 945                         }
 946                 }
 947         }
 948 }
 949
 950 static void be_spill_belady(ir_graph *irg, const arch_register_class_t *rcls)
 951 {
 952         int i;
 953
 954         be_liveness_assure_sets(be_assure_liveness(irg));
 955
 956         stat_ev_tim_push();
 957         /* construct control flow loop tree */
 958         if (! (get_irg_loopinfo_state(irg) & loopinfo_cf_consistent)) {
 959                 construct_cf_backedges(irg);
 960         }
 961         stat_ev_tim_pop("belady_time_backedges");
 962
 963         stat_ev_tim_push();
 964         be_clear_links(irg);
 965         stat_ev_tim_pop("belady_time_clear_links");
 966
 967         ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
 968
 969         /* init belady env */
 970         stat_ev_tim_push();
 971         obstack_init(&obst);
 972         cls       = rcls;
 973         lv        = be_get_irg_liveness(irg);
 974         n_regs    = cls->n_regs - be_put_ignore_regs(irg, cls, NULL);
 975         ws        = new_workset();
 976         uses      = be_begin_uses(irg, lv);
 977         loop_ana  = be_new_loop_pressure(irg, cls);
 978         senv      = be_new_spill_env(irg);
 979         blocklist = be_get_cfgpostorder(irg);
 980         stat_ev_tim_pop("belady_time_init");
 981
 982         stat_ev_tim_push();
 983         /* walk blocks in reverse postorder */
 984         for (i = ARR_LEN(blocklist) - 1; i >= 0; --i) {
 985                 process_block(blocklist[i]);
 986         }
 987         DEL_ARR_F(blocklist);
 988         stat_ev_tim_pop("belady_time_belady");
 989
 990         stat_ev_tim_push();
 991         /* belady was block-local, fix the global flow by adding reloads on the
 992          * edges */
 993         irg_block_walk_graph(irg, fix_block_borders, NULL, NULL);
 994         stat_ev_tim_pop("belady_time_fix_borders");
 995
 996         ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
 997
 998         /* Insert spill/reload nodes into the graph and fix usages */
 999         be_insert_spills_reloads(senv);
1000
1001         /* clean up */
1002         be_delete_spill_env(senv);
1003         be_end_uses(uses);
1004         be_free_loop_pressure(loop_ana);
1005         obstack_free(&obst, NULL);
1006 }
1007
1008 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady);
1009 void be_init_spillbelady(void)
1010 {
1011         static be_spiller_t belady_spiller = {
1012                 be_spill_belady
1013         };
1014         lc_opt_entry_t *be_grp       = lc_opt_get_grp(firm_opt_get_root(), "be");
1015         lc_opt_entry_t *belady_group = lc_opt_get_grp(be_grp, "belady");
1016         lc_opt_add_table(belady_group, options);
1017
1018         be_register_spiller("belady", &belady_spiller);
1019         FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
1020 }