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 //#define LOOK_AT_LOOPDEPTH
  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;  /**< the value was already spilled on this path */
  78 } loc_t;
  79
  80 typedef struct _workset_t {
  81         int   len;          /**< current length */
  82         loc_t vals[0];      /**< inlined 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 int                          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 unsigned                     instr_nr; /**< current instruction number
  95                                                        (relative to block start) */
  96 static spill_env_t                 *senv;   /**< see bespill.h */
  97 static ir_node                    **blocklist;
  98
  99 static int                          move_spills      = true;
 100 static int                          respectloopdepth = true;
 101 static int                          improve_known_preds = true;
 102 /* factor to weight the different costs of reloading/rematerializing a node
 103    (see bespill.h be_get_reload_costs_no_weight) */
 104 static int                          remat_bonus      = 10;
 105
 106 static const lc_opt_table_entry_t options[] = {
 107         LC_OPT_ENT_BOOL   ("movespills", "try to move spills out of loops", &move_spills),
 108         LC_OPT_ENT_BOOL   ("respectloopdepth", "exprimental (outermost loop cutting)", &respectloopdepth),
 109         LC_OPT_ENT_BOOL   ("improveknownpreds", "experimental (known preds cutting)", &improve_known_preds),
 110         LC_OPT_ENT_INT    ("rematbonus", "give bonus to rematerialisable nodes", &remat_bonus),
 111         LC_OPT_LAST
 112 };
 113
 114 static int loc_compare(const void *a, const void *b)
 115 {
 116         const loc_t *p = a;
 117         const loc_t *q = b;
 118         return p->time - q->time;
 119 }
 120
 121 /**
 122  * Alloc a new workset on obstack @p ob with maximum size @p max
 123  */
 124 static workset_t *new_workset(void)
 125 {
 126         return OALLOCFZ(&obst, workset_t, vals, n_regs);
 127 }
 128
 129 /**
 130  * Alloc a new instance on obstack and make it equal to @param workset
 131  */
 132 static workset_t *workset_clone(workset_t *workset)
 133 {
 134         workset_t *res = OALLOCF(&obst, workset_t, vals, n_regs);
 135         memcpy(res, workset, sizeof(*res) + n_regs * sizeof(res->vals[0]));
 136         return res;
 137 }
 138
 139 /**
 140  * Copy workset @param src to @param tgt
 141  */
 142 static void workset_copy(workset_t *dest, const workset_t *src)
 143 {
 144         size_t size = sizeof(*src) + n_regs * sizeof(src->vals[0]);
 145         memcpy(dest, src, size);
 146 }
 147
 148 /**
 149  * Overwrites the current content array of @param ws with the
 150  * @param count locations given at memory @param locs.
 151  * Set the length of @param ws to count.
 152  */
 153 static void workset_bulk_fill(workset_t *workset, int count, const loc_t *locs)
 154 {
 155         workset->len = count;
 156         memcpy(&(workset->vals[0]), locs, count * sizeof(locs[0]));
 157 }
 158
 159 /**
 160  * Inserts the value @p val into the workset, iff it is not
 161  * already contained. The workset must not be full.
 162  */
 163 static void workset_insert(workset_t *workset, ir_node *val, bool spilled)
 164 {
 165         loc_t *loc;
 166         int    i;
 167         /* check for current regclass */
 168         assert(arch_irn_consider_in_reg_alloc(cls, val));
 169
 170         /* check if val is already contained */
 171         for (i = 0; i < workset->len; ++i) {
 172                 loc = &workset->vals[i];
 173                 if (loc->node == val) {
 174                         if (spilled) {
 175                                 loc->spilled = true;
 176                         }
 177                         return;
 178                 }
 179         }
 180
 181         /* insert val */
 182         assert(workset->len < n_regs && "Workset already full!");
 183         loc           = &workset->vals[workset->len];
 184         loc->node     = val;
 185         loc->spilled  = spilled;
 186         loc->time     = TIME_UNDEFINED;
 187         workset->len++;
 188 }
 189
 190 /**
 191  * Removes all entries from this workset
 192  */
 193 static void workset_clear(workset_t *workset)
 194 {
 195         workset->len = 0;
 196 }
 197
 198 /**
 199  * Removes the value @p val from the workset if present.
 200  */
 201 static inline void workset_remove(workset_t *workset, ir_node *val)
 202 {
 203         int i;
 204         for (i = 0; i < workset->len; ++i) {
 205                 if (workset->vals[i].node == val) {
 206                         workset->vals[i] = workset->vals[--workset->len];
 207                         return;
 208                 }
 209         }
 210 }
 211
 212 static inline const loc_t *workset_contains(const workset_t *ws,
 213                                             const ir_node *val)
 214 {
 215         int i;
 216
 217         for (i = 0; i < ws->len; ++i) {
 218                 if (ws->vals[i].node == val)
 219                         return &ws->vals[i];
 220         }
 221
 222         return NULL;
 223 }
 224
 225 /**
 226  * Iterates over all values in the working set.
 227  * @p ws The workset to iterate
 228  * @p v  A variable to put the current value in
 229  * @p i  An integer for internal use
 230  */
 231 #define workset_foreach(ws, v, i)       for (i=0; \
 232                                                                                 v=(i < ws->len) ? ws->vals[i].node : NULL, i < ws->len; \
 233                                                                                 ++i)
 234
 235 #define workset_set_time(ws, i, t) (ws)->vals[i].time=t
 236 #define workset_get_time(ws, i) (ws)->vals[i].time
 237 #define workset_set_length(ws, length) (ws)->len = length
 238 #define workset_get_length(ws) ((ws)->len)
 239 #define workset_get_val(ws, i) ((ws)->vals[i].node)
 240 #define workset_sort(ws) do { qsort((ws)->vals, (ws)->len, sizeof((ws)->vals[0]), loc_compare); } while(0)
 241
 242 typedef struct _block_info_t
 243 {
 244         workset_t *start_workset;
 245         workset_t *end_workset;
 246 } block_info_t;
 247
 248
 249 static block_info_t *new_block_info(void)
 250 {
 251         return OALLOCZ(&obst, block_info_t);
 252 }
 253
 254 #define get_block_info(block)        ((block_info_t *)get_irn_link(block))
 255 #define set_block_info(block, info)  set_irn_link(block, info)
 256
 257 /**
 258  * @return The distance to the next use or 0 if irn has dont_spill flag set
 259  */
 260 static inline unsigned get_distance(ir_node *from, unsigned from_step,
 261                                     const ir_node *def, int skip_from_uses)
 262 {
 263         be_next_use_t use;
 264         unsigned      costs;
 265         unsigned      time;
 266
 267         assert(!arch_irn_is_ignore(def));
 268
 269         use  = be_get_next_use(uses, from, from_step, def, skip_from_uses);
 270         time = use.time;
 271         if (USES_IS_INFINITE(time))
 272                 return USES_INFINITY;
 273
 274         /* We have to keep nonspillable nodes in the workingset */
 275         if (arch_irn_get_flags(skip_Proj_const(def)) & arch_irn_flags_dont_spill)
 276                 return 0;
 277
 278         /* give some bonus to rematerialisable nodes */
 279         if (remat_bonus > 0) {
 280                 costs = be_get_reload_costs_no_weight(senv, def, use.before);
 281                 assert(costs * remat_bonus < 1000);
 282                 time  += 1000 - (costs * remat_bonus);
 283         }
 284
 285         return time;
 286 }
 287
 288 /**
 289  * Performs the actions necessary to grant the request that:
 290  * - new_vals can be held in registers
 291  * - as few as possible other values are disposed
 292  * - the worst values get disposed
 293  *
 294  * @p is_usage indicates that the values in new_vals are used (not defined)
 295  * In this case reloads must be performed
 296  */
 297 static void displace(workset_t *new_vals, int is_usage)
 298 {
 299         ir_node **to_insert = ALLOCAN(ir_node*, n_regs);
 300         bool     *spilled   = ALLOCAN(bool,     n_regs);
 301         ir_node  *val;
 302         int       i;
 303         int       len;
 304         int       spills_needed;
 305         int       demand;
 306         int       iter;
 307
 308         /* 1. Identify the number of needed slots and the values to reload */
 309         demand = 0;
 310         workset_foreach(new_vals, val, iter) {
 311                 bool reloaded = false;
 312
 313                 if (! workset_contains(ws, val)) {
 314                         DB((dbg, DBG_DECIDE, "    insert %+F\n", val));
 315                         if (is_usage) {
 316                                 DB((dbg, DBG_SPILL, "Reload %+F before %+F\n", val, instr));
 317                                 be_add_reload(senv, val, instr, cls, 1);
 318                                 reloaded = true;
 319                         }
 320                 } else {
 321                         DB((dbg, DBG_DECIDE, "    %+F already in workset\n", val));
 322                         assert(is_usage);
 323                         /* remove the value from the current workset so it is not accidently
 324                          * spilled */
 325                         workset_remove(ws, val);
 326                 }
 327                 spilled[demand]   = reloaded;
 328                 to_insert[demand] = val;
 329                 ++demand;
 330         }
 331
 332         /* 2. Make room for at least 'demand' slots */
 333         len           = workset_get_length(ws);
 334         spills_needed = len + demand - n_regs;
 335         assert(spills_needed <= len);
 336
 337         /* Only make more free room if we do not have enough */
 338         if (spills_needed > 0) {
 339                 DB((dbg, DBG_DECIDE, "    disposing %d values\n", spills_needed));
 340
 341                 /* calculate current next-use distance for live values */
 342                 for (i = 0; i < len; ++i) {
 343                         ir_node  *val  = workset_get_val(ws, i);
 344                         unsigned  dist = get_distance(instr, instr_nr, val, !is_usage);
 345                         workset_set_time(ws, i, dist);
 346                 }
 347
 348                 /* sort entries by increasing nextuse-distance*/
 349                 workset_sort(ws);
 350
 351                 for (i = len - spills_needed; i < len; ++i) {
 352                         ir_node *val = ws->vals[i].node;
 353
 354                         DB((dbg, DBG_DECIDE, "    disposing node %+F (%u)\n", val,
 355                              workset_get_time(ws, i)));
 356
 357                         if (move_spills) {
 358                                 if (!USES_IS_INFINITE(ws->vals[i].time)
 359                                                 && !ws->vals[i].spilled) {
 360                                         ir_node *after_pos = sched_prev(instr);
 361                                         DB((dbg, DBG_DECIDE, "Spill %+F after node %+F\n", val,
 362                                                 after_pos));
 363                                         be_add_spill(senv, val, after_pos);
 364                                 }
 365                         }
 366                 }
 367
 368                 /* kill the last 'demand' entries in the array */
 369                 workset_set_length(ws, len - spills_needed);
 370         }
 371
 372         /* 3. Insert the new values into the workset */
 373         for (i = 0; i < demand; ++i) {
 374                 ir_node *val = to_insert[i];
 375
 376                 workset_insert(ws, val, spilled[i]);
 377         }
 378 }
 379
 380 enum {
 381         AVAILABLE_EVERYWHERE,
 382         AVAILABLE_NOWHERE,
 383         AVAILABLE_PARTLY,
 384         AVAILABLE_UNKNOWN
 385 };
 386
 387 static unsigned available_in_all_preds(workset_t* const* pred_worksets,
 388                                        size_t n_pred_worksets,
 389                                        const ir_node *value, bool is_local_phi)
 390 {
 391         size_t i;
 392         bool   avail_everywhere = true;
 393         bool   avail_nowhere    = true;
 394
 395         assert(n_pred_worksets > 0);
 396
 397         /* value available in all preds? */
 398         for (i = 0; i < n_pred_worksets; ++i) {
 399                 bool             found     = false;
 400                 const workset_t *p_workset = pred_worksets[i];
 401                 int              p_len     = workset_get_length(p_workset);
 402                 int              p_i;
 403                 const ir_node   *l_value;
 404
 405                 if (is_local_phi) {
 406                         assert(is_Phi(value));
 407                         l_value = get_irn_n(value, i);
 408                 } else {
 409                         l_value = value;
 410                 }
 411
 412                 for (p_i = 0; p_i < p_len; ++p_i) {
 413                         const loc_t *p_l = &p_workset->vals[p_i];
 414                         if (p_l->node != l_value)
 415                                 continue;
 416
 417                         found = true;
 418                         break;
 419                 }
 420
 421                 if (found) {
 422                         avail_nowhere = false;
 423                 } else {
 424                         avail_everywhere = false;
 425                 }
 426         }
 427
 428         if (avail_everywhere) {
 429                 assert(!avail_nowhere);
 430                 return AVAILABLE_EVERYWHERE;
 431         } else if (avail_nowhere) {
 432                 return AVAILABLE_NOWHERE;
 433         } else {
 434                 return AVAILABLE_PARTLY;
 435         }
 436 }
 437
 438 /** Decides whether a specific node should be in the start workset or not
 439  *
 440  * @param env      belady environment
 441  * @param first
 442  * @param node     the node to test
 443  * @param loop     the loop of the node
 444  */
 445 static loc_t to_take_or_not_to_take(ir_node* first, ir_node *node,
 446                                     ir_loop *loop, unsigned available)
 447 {
 448         be_next_use_t next_use;
 449         loc_t         loc;
 450
 451         loc.time    = USES_INFINITY;
 452         loc.node    = node;
 453         loc.spilled = false;
 454
 455         if (!arch_irn_consider_in_reg_alloc(cls, node)) {
 456                 loc.time = USES_INFINITY;
 457                 return loc;
 458         }
 459
 460         /* We have to keep nonspillable nodes in the workingset */
 461         if (arch_irn_get_flags(skip_Proj_const(node)) & arch_irn_flags_dont_spill) {
 462                 loc.time = 0;
 463                 DB((dbg, DBG_START, "    %+F taken (dontspill node)\n", node, loc.time));
 464                 return loc;
 465         }
 466
 467         next_use = be_get_next_use(uses, first, 0, node, 0);
 468         if (USES_IS_INFINITE(next_use.time)) {
 469                 // the nodes marked as live in shouldn't be dead, so it must be a phi
 470                 assert(is_Phi(node));
 471                 loc.time = USES_INFINITY;
 472                 DB((dbg, DBG_START, "    %+F not taken (dead)\n", node));
 473                 return loc;
 474         }
 475
 476         loc.time = next_use.time;
 477
 478         if (improve_known_preds) {
 479                 if (available == AVAILABLE_EVERYWHERE) {
 480                         DB((dbg, DBG_START, "    %+F taken (%u, live in all preds)\n",
 481                             node, loc.time));
 482                         return loc;
 483                 } else if (available == AVAILABLE_NOWHERE) {
 484                         DB((dbg, DBG_START, "    %+F not taken (%u, live in no pred)\n",
 485                             node, loc.time));
 486                         loc.time = USES_INFINITY;
 487                         return loc;
 488                 }
 489         }
 490
 491         if (!respectloopdepth || next_use.outermost_loop >= get_loop_depth(loop)) {
 492                 DB((dbg, DBG_START, "    %+F taken (%u, loop %d)\n", node, loc.time,
 493                     next_use.outermost_loop));
 494         } else {
 495                 loc.time = USES_PENDING;
 496                 DB((dbg, DBG_START, "    %+F delayed (outerdepth %d < loopdepth %d)\n",
 497                     node, next_use.outermost_loop, get_loop_depth(loop)));
 498         }
 499
 500         return loc;
 501 }
 502
 503 /**
 504  * Computes the start-workset for a block with multiple predecessors. We assume
 505  * that at least 1 of the predeccesors is a back-edge which means we're at the
 506  * beginning of a loop. We try to reload as much values as possible now so they
 507  * don't get reloaded inside the loop.
 508  */
 509 static void decide_start_workset(const ir_node *block)
 510 {
 511         ir_loop    *loop = get_irn_loop(block);
 512         ir_node    *first;
 513         ir_node    *node;
 514         loc_t       loc;
 515         loc_t      *starters;
 516         loc_t      *delayed;
 517         int         i, len, ws_count;
 518         int             free_slots, free_pressure_slots;
 519         unsigned    pressure;
 520         int         arity;
 521         workset_t **pred_worksets;
 522         bool        all_preds_known;
 523
 524         /* check predecessors */
 525         arity           = get_irn_arity(block);
 526         pred_worksets   = ALLOCAN(workset_t*, arity);
 527         all_preds_known = true;
 528         for (i = 0; i < arity; ++i) {
 529                 ir_node      *pred_block = get_Block_cfgpred_block(block, i);
 530                 block_info_t *pred_info  = get_block_info(pred_block);
 531
 532                 if (pred_info == NULL) {
 533                         pred_worksets[i] = NULL;
 534                         all_preds_known  = false;
 535                 } else {
 536                         pred_worksets[i] = pred_info->end_workset;
 537                 }
 538         }
 539
 540         /* Collect all values living at start of block */
 541         starters = NEW_ARR_F(loc_t, 0);
 542         delayed  = NEW_ARR_F(loc_t, 0);
 543
 544         DB((dbg, DBG_START, "Living at start of %+F:\n", block));
 545         first = sched_first(block);
 546
 547         /* check all Phis first */
 548         sched_foreach(block, node) {
 549                 unsigned available;
 550
 551                 if (! is_Phi(node))
 552                         break;
 553                 if (!arch_irn_consider_in_reg_alloc(cls, node))
 554                         continue;
 555
 556                 if (all_preds_known) {
 557                         available = available_in_all_preds(pred_worksets, arity, node, true);
 558                 } else {
 559                         available = AVAILABLE_UNKNOWN;
 560                 }
 561
 562                 loc = to_take_or_not_to_take(first, node, loop, available);
 563
 564                 if (! USES_IS_INFINITE(loc.time)) {
 565                         if (USES_IS_PENDING(loc.time))
 566                                 ARR_APP1(loc_t, delayed, loc);
 567                         else
 568                                 ARR_APP1(loc_t, starters, loc);
 569                 } else {
 570                         be_spill_phi(senv, node);
 571                 }
 572         }
 573
 574         /* check all Live-Ins */
 575         be_lv_foreach(lv, block, be_lv_state_in, i) {
 576                 ir_node *node = be_lv_get_irn(lv, block, i);
 577                 unsigned available;
 578
 579                 if (all_preds_known) {
 580                         available = available_in_all_preds(pred_worksets, arity, node, false);
 581                 } else {
 582                         available = AVAILABLE_UNKNOWN;
 583                 }
 584
 585                 loc = to_take_or_not_to_take(first, node, loop, available);
 586
 587                 if (! USES_IS_INFINITE(loc.time)) {
 588                         if (USES_IS_PENDING(loc.time))
 589                                 ARR_APP1(loc_t, delayed, loc);
 590                         else
 591                                 ARR_APP1(loc_t, starters, loc);
 592                 }
 593         }
 594
 595         pressure            = be_get_loop_pressure(loop_ana, cls, loop);
 596         assert(ARR_LEN(delayed) <= (signed)pressure);
 597         free_slots          = n_regs - ARR_LEN(starters);
 598         free_pressure_slots = n_regs - (pressure - ARR_LEN(delayed));
 599         free_slots          = MIN(free_slots, free_pressure_slots);
 600
 601         /* so far we only put nodes into the starters list that are used inside
 602          * the loop. If register pressure in the loop is low then we can take some
 603          * values and let them live through the loop */
 604         DB((dbg, DBG_START, "Loop pressure %d, taking %d delayed vals\n",
 605             pressure, free_slots));
 606         if (free_slots > 0) {
 607                 qsort(delayed, ARR_LEN(delayed), sizeof(delayed[0]), loc_compare);
 608
 609                 for (i = 0; i < ARR_LEN(delayed) && free_slots > 0; ++i) {
 610                         int    p, arity;
 611                         loc_t *loc = & delayed[i];
 612
 613                         if (!is_Phi(loc->node)) {
 614                                 /* don't use values which are dead in a known predecessors
 615                                  * to not induce unnecessary reloads */
 616                                 arity = get_irn_arity(block);
 617                                 for (p = 0; p < arity; ++p) {
 618                                         ir_node      *pred_block = get_Block_cfgpred_block(block, p);
 619                                         block_info_t *pred_info  = get_block_info(pred_block);
 620
 621                                         if (pred_info == NULL)
 622                                                 continue;
 623
 624                                         if (!workset_contains(pred_info->end_workset, loc->node)) {
 625                                                 DB((dbg, DBG_START,
 626                                                         "    delayed %+F not live at pred %+F\n", loc->node,
 627                                                         pred_block));
 628                                                 goto skip_delayed;
 629                                         }
 630                                 }
 631                         }
 632
 633                         DB((dbg, DBG_START, "    delayed %+F taken\n", loc->node));
 634                         ARR_APP1(loc_t, starters, *loc);
 635                         loc->node = NULL;
 636                         --free_slots;
 637                 skip_delayed:
 638                         ;
 639                 }
 640         }
 641
 642         /* spill phis (the actual phis not just their values) that are in this block
 643          * but not in the start workset */
 644         for (i = ARR_LEN(delayed) - 1; i >= 0; --i) {
 645                 ir_node *node = delayed[i].node;
 646                 if (node == NULL || !is_Phi(node) || get_nodes_block(node) != block)
 647                         continue;
 648
 649                 DB((dbg, DBG_START, "    spilling delayed phi %+F\n", node));
 650                 be_spill_phi(senv, node);
 651         }
 652         DEL_ARR_F(delayed);
 653
 654         /* Sort start values by first use */
 655         qsort(starters, ARR_LEN(starters), sizeof(starters[0]), loc_compare);
 656
 657         /* Copy the best ones from starters to start workset */
 658         ws_count = MIN(ARR_LEN(starters), n_regs);
 659         workset_clear(ws);
 660         workset_bulk_fill(ws, ws_count, starters);
 661
 662         /* spill phis (the actual phis not just their values) that are in this block
 663          * but not in the start workset */
 664         len = ARR_LEN(starters);
 665         for (i = ws_count; i < len; ++i) {
 666                 ir_node *node = starters[i].node;
 667                 if (! is_Phi(node) || get_nodes_block(node) != block)
 668                         continue;
 669
 670                 DB((dbg, DBG_START, "    spilling phi %+F\n", node));
 671                 be_spill_phi(senv, node);
 672         }
 673
 674         DEL_ARR_F(starters);
 675
 676         /* determine spill status of the values: If there's 1 pred block (which
 677          * is no backedge) where the value is spilled then we must set it to
 678          * spilled here. */
 679         for (i = 0; i < ws_count; ++i) {
 680                 loc_t   *loc     = &ws->vals[i];
 681                 ir_node *value   = loc->node;
 682                 bool     spilled;
 683                 int      n;
 684
 685                 /* phis from this block aren't spilled */
 686                 if (get_nodes_block(value) == block) {
 687                         assert(is_Phi(value));
 688                         loc->spilled = false;
 689                         continue;
 690                 }
 691
 692                 /* determine if value was spilled on any predecessor */
 693                 spilled = false;
 694                 for (n = 0; n < arity; ++n) {
 695                         workset_t *pred_workset = pred_worksets[n];
 696                         int        p_len;
 697                         int        p;
 698
 699                         if (pred_workset == NULL)
 700                                 continue;
 701
 702                         p_len = workset_get_length(pred_workset);
 703                         for (p = 0; p < p_len; ++p) {
 704                                 loc_t *l = &pred_workset->vals[p];
 705
 706                                 if (l->node != value)
 707                                         continue;
 708
 709                                 if (l->spilled) {
 710                                         spilled = true;
 711                                 }
 712                                 break;
 713                         }
 714                 }
 715
 716                 loc->spilled = spilled;
 717         }
 718 }
 719
 720 /**
 721  * For the given block @p block, decide for each values
 722  * whether it is used from a register or is reloaded
 723  * before the use.
 724  */
 725 static void process_block(ir_node *block)
 726 {
 727         workset_t       *new_vals;
 728         ir_node         *irn;
 729         int              iter;
 730         block_info_t    *block_info;
 731         int              arity;
 732
 733         /* no need to process a block twice */
 734         assert(get_block_info(block) == NULL);
 735
 736         /* construct start workset */
 737         arity = get_Block_n_cfgpreds(block);
 738         if (arity == 0) {
 739                 /* no predecessor -> empty set */
 740                 workset_clear(ws);
 741         } else if (arity == 1) {
 742                 /* one predecessor, copy it's end workset */
 743                 ir_node      *pred_block = get_Block_cfgpred_block(block, 0);
 744                 block_info_t *pred_info  = get_block_info(pred_block);
 745
 746                 assert(pred_info != NULL);
 747                 workset_copy(ws, pred_info->end_workset);
 748         } else {
 749                 /* multiple predecessors, do more advanced magic :) */
 750                 decide_start_workset(block);
 751         }
 752
 753         DB((dbg, DBG_DECIDE, "\n"));
 754         DB((dbg, DBG_DECIDE, "Decide for %+F\n", block));
 755
 756         block_info = new_block_info();
 757         set_block_info(block, block_info);
 758
 759         DB((dbg, DBG_WSETS, "Start workset for %+F:\n", block));
 760         workset_foreach(ws, irn, iter) {
 761                 DB((dbg, DBG_WSETS, "  %+F (%u)\n", irn,
 762                      workset_get_time(ws, iter)));
 763         }
 764
 765         block_info->start_workset = workset_clone(ws);
 766
 767         /* process the block from start to end */
 768         DB((dbg, DBG_WSETS, "Processing...\n"));
 769         instr_nr = 0;
 770         /* TODO: this leaks (into the obstack)... */
 771         new_vals = new_workset();
 772
 773         sched_foreach(block, irn) {
 774                 int i, arity;
 775                 assert(workset_get_length(ws) <= n_regs);
 776
 777                 /* Phis are no real instr (see insert_starters()) */
 778                 if (is_Phi(irn)) {
 779                         continue;
 780                 }
 781                 DB((dbg, DBG_DECIDE, "  ...%+F\n", irn));
 782
 783                 /* set instruction in the workset */
 784                 instr = irn;
 785
 786                 /* allocate all values _used_ by this instruction */
 787                 workset_clear(new_vals);
 788                 for (i = 0, arity = get_irn_arity(irn); i < arity; ++i) {
 789                         ir_node *in = get_irn_n(irn, i);
 790                         if (!arch_irn_consider_in_reg_alloc(cls, in))
 791                                 continue;
 792
 793                         /* (note that "spilled" is irrelevant here) */
 794                         workset_insert(new_vals, in, false);
 795                 }
 796                 displace(new_vals, 1);
 797
 798                 /* allocate all values _defined_ by this instruction */
 799                 workset_clear(new_vals);
 800                 if (get_irn_mode(irn) == mode_T) {
 801                         const ir_edge_t *edge;
 802
 803                         foreach_out_edge(irn, edge) {
 804                                 ir_node *proj = get_edge_src_irn(edge);
 805                                 if (!arch_irn_consider_in_reg_alloc(cls, proj))
 806                                         continue;
 807                                 workset_insert(new_vals, proj, false);
 808                         }
 809                 } else {
 810                         if (!arch_irn_consider_in_reg_alloc(cls, irn))
 811                                 continue;
 812                         workset_insert(new_vals, irn, false);
 813                 }
 814                 displace(new_vals, 0);
 815
 816                 instr_nr++;
 817         }
 818
 819         /* Remember end-workset for this block */
 820         block_info->end_workset = workset_clone(ws);
 821         DB((dbg, DBG_WSETS, "End workset for %+F:\n", block));
 822         workset_foreach(ws, irn, iter)
 823                 DB((dbg, DBG_WSETS, "  %+F (%u)\n", irn,
 824                      workset_get_time(ws, iter)));
 825 }
 826
 827 /**
 828  * 'decide' is block-local and makes assumptions
 829  * about the set of live-ins. Thus we must adapt the
 830  * live-outs to the live-ins at each block-border.
 831  */
 832 static void fix_block_borders(ir_node *block, void *data)
 833 {
 834         workset_t    *start_workset;
 835         int           arity;
 836         int           i;
 837         int           iter;
 838         (void) data;
 839
 840         DB((dbg, DBG_FIX, "\n"));
 841         DB((dbg, DBG_FIX, "Fixing %+F\n", block));
 842
 843         arity = get_irn_arity(block);
 844         /* can happen for endless loops */
 845         if (arity == 0)
 846                 return;
 847
 848         start_workset = get_block_info(block)->start_workset;
 849
 850         /* process all pred blocks */
 851         for (i = 0; i < arity; ++i) {
 852                 ir_node   *pred = get_Block_cfgpred_block(block, i);
 853                 workset_t *pred_end_workset = get_block_info(pred)->end_workset;
 854                 ir_node   *node;
 855
 856                 DB((dbg, DBG_FIX, "  Pred %+F\n", pred));
 857
 858                 /* spill all values not used anymore */
 859                 workset_foreach(pred_end_workset, node, iter) {
 860                         ir_node *n2;
 861                         int      iter2;
 862                         bool     found = false;
 863                         workset_foreach(start_workset, n2, iter2) {
 864                                 if (n2 == node) {
 865                                         found = true;
 866                                         break;
 867                                 }
 868                                 /* note that we do not look at phi inputs, becuase the values
 869                                  * will be either live-end and need no spill or
 870                                  * they have other users in which must be somewhere else in the
 871                                  * workset */
 872                         }
 873
 874                         if (found)
 875                                 continue;
 876
 877                         if (move_spills && be_is_live_in(lv, block, node)
 878                                         && !pred_end_workset->vals[iter].spilled) {
 879                                 ir_node *insert_point;
 880                                 if (arity > 1) {
 881                                         insert_point = be_get_end_of_block_insertion_point(pred);
 882                                         insert_point = sched_prev(insert_point);
 883                                 } else {
 884                                         insert_point = block;
 885                                 }
 886                                 DB((dbg, DBG_SPILL, "Spill %+F after %+F\n", node,
 887                                      insert_point));
 888                                 be_add_spill(senv, node, insert_point);
 889                         }
 890                 }
 891
 892                 /* reload missing values in predecessors, add missing spills */
 893                 workset_foreach(start_workset, node, iter) {
 894                         const loc_t *l    = &start_workset->vals[iter];
 895                         const loc_t *pred_loc;
 896
 897                         /* if node is a phi of the current block we reload
 898                          * the corresponding argument, else node itself */
 899                         if (is_Phi(node) && get_nodes_block(node) == block) {
 900                                 node = get_irn_n(node, i);
 901                                 assert(!l->spilled);
 902
 903                                 /* we might have unknowns as argument for the phi */
 904                                 if (!arch_irn_consider_in_reg_alloc(cls, node))
 905                                         continue;
 906                         }
 907
 908                         /* check if node is in a register at end of pred */
 909                         pred_loc = workset_contains(pred_end_workset, node);
 910                         if (pred_loc != NULL) {
 911                                 /* we might have to spill value on this path */
 912                                 if (move_spills && !pred_loc->spilled && l->spilled) {
 913                                         ir_node *insert_point
 914                                                 = be_get_end_of_block_insertion_point(pred);
 915                                         insert_point = sched_prev(insert_point);
 916                                         DB((dbg, DBG_SPILL, "Spill %+F after %+F\n", node,
 917                                             insert_point));
 918                                         be_add_spill(senv, node, insert_point);
 919                                 }
 920                         } else {
 921                                 /* node is not in register at the end of pred -> reload it */
 922                                 DB((dbg, DBG_FIX, "    reload %+F\n", node));
 923                                 DB((dbg, DBG_SPILL, "Reload %+F before %+F,%d\n", node, block, i));
 924                                 be_add_reload_on_edge(senv, node, block, i, cls, 1);
 925                         }
 926                 }
 927         }
 928 }
 929
 930 static void be_spill_belady(ir_graph *irg, const arch_register_class_t *rcls)
 931 {
 932         int i;
 933
 934         be_liveness_assure_sets(be_assure_liveness(irg));
 935
 936         stat_ev_tim_push();
 937         /* construct control flow loop tree */
 938         if (! (get_irg_loopinfo_state(irg) & loopinfo_cf_consistent)) {
 939                 construct_cf_backedges(irg);
 940         }
 941         stat_ev_tim_pop("belady_time_backedges");
 942
 943         stat_ev_tim_push();
 944         be_clear_links(irg);
 945         stat_ev_tim_pop("belady_time_clear_links");
 946
 947         ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
 948
 949         /* init belady env */
 950         stat_ev_tim_push();
 951         obstack_init(&obst);
 952         cls       = rcls;
 953         lv        = be_get_irg_liveness(irg);
 954         n_regs    = cls->n_regs - be_put_ignore_regs(irg, cls, NULL);
 955         ws        = new_workset();
 956         uses      = be_begin_uses(irg, lv);
 957         loop_ana  = be_new_loop_pressure(irg, cls);
 958         senv      = be_new_spill_env(irg);
 959         blocklist = be_get_cfgpostorder(irg);
 960         stat_ev_tim_pop("belady_time_init");
 961
 962         stat_ev_tim_push();
 963         /* walk blocks in reverse postorder */
 964         for (i = ARR_LEN(blocklist) - 1; i >= 0; --i) {
 965                 process_block(blocklist[i]);
 966         }
 967         DEL_ARR_F(blocklist);
 968         stat_ev_tim_pop("belady_time_belady");
 969
 970         stat_ev_tim_push();
 971         /* belady was block-local, fix the global flow by adding reloads on the
 972          * edges */
 973         irg_block_walk_graph(irg, fix_block_borders, NULL, NULL);
 974         stat_ev_tim_pop("belady_time_fix_borders");
 975
 976         ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
 977
 978         /* Insert spill/reload nodes into the graph and fix usages */
 979         be_insert_spills_reloads(senv);
 980
 981         /* clean up */
 982         be_delete_spill_env(senv);
 983         be_end_uses(uses);
 984         be_free_loop_pressure(loop_ana);
 985         obstack_free(&obst, NULL);
 986 }
 987
 988 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady);
 989 void be_init_spillbelady(void)
 990 {
 991         static be_spiller_t belady_spiller = {
 992                 be_spill_belady
 993         };
 994         lc_opt_entry_t *be_grp       = lc_opt_get_grp(firm_opt_get_root(), "be");
 995         lc_opt_entry_t *belady_group = lc_opt_get_grp(be_grp, "belady");
 996         lc_opt_add_table(belady_group, options);
 997
 998         be_register_spiller("belady", &belady_spiller);
 999         FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
1000 }