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