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