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 #ifdef HAVE_CONFIG_H
  28 #include "config.h"
  29 #endif
  30
  31 #include <stdbool.h>
  32
  33 #include "obst.h"
  34 #include "irprintf_t.h"
  35 #include "irgraph.h"
  36 #include "irnode.h"
  37 #include "irmode.h"
  38 #include "irgwalk.h"
  39 #include "irloop.h"
  40 #include "iredges_t.h"
  41 #include "ircons_t.h"
  42 #include "irprintf.h"
  43 #include "irnodeset.h"
  44
  45 #include "beutil.h"
  46 #include "bearch_t.h"
  47 #include "beuses.h"
  48 #include "besched_t.h"
  49 #include "beirgmod.h"
  50 #include "belive_t.h"
  51 #include "benode_t.h"
  52 #include "bechordal_t.h"
  53 #include "bespilloptions.h"
  54 #include "beloopana.h"
  55 #include "beirg_t.h"
  56 #include "bespill.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 //#define LOOK_AT_LOOPDEPTH
  72
  73 /**
  74  * An association between a node and a point in time.
  75  */
  76 typedef struct loc_t {
  77         ir_node          *node;
  78         unsigned          time;     /**< A use time (see beuses.h). */
  79         bool              spilled;  /**< the value was already spilled on this path */
  80 } loc_t;
  81
  82 typedef struct _workset_t {
  83         int   len;          /**< current length */
  84         loc_t vals[0];      /**< inlined array of the values/distances in this working set */
  85 } workset_t;
  86
  87 static struct obstack               obst;
  88 static const arch_env_t            *arch_env;
  89 static const arch_register_class_t *cls;
  90 static const be_lv_t               *lv;
  91 static be_loopana_t                *loop_ana;
  92 static int                          n_regs;
  93 static workset_t                   *ws;     /**< the main workset used while
  94                                                      processing a block. */
  95 static be_uses_t                   *uses;   /**< env for the next-use magic */
  96 static ir_node                     *instr;  /**< current instruction */
  97 static unsigned                     instr_nr; /**< current instruction number
  98                                                        (relative to block start) */
  99 static spill_env_t                 *senv;   /**< see bespill.h */
 100 static ir_node                    **blocklist;
 101
 102 static bool                         move_spills      = true;
 103 static bool                         respectloopdepth = true;
 104 static bool                         improve_known_preds = true;
 105 /* factor to weight the different costs of reloading/rematerializing a node
 106    (see bespill.h be_get_reload_costs_no_weight) */
 107 static int                          remat_bonus      = 10;
 108
 109 static const lc_opt_table_entry_t options[] = {
 110         LC_OPT_ENT_BOOL   ("movespills", "try to move spills out of loops", &move_spills),
 111         LC_OPT_ENT_BOOL   ("respectloopdepth", "exprimental (outermost loop cutting)", &respectloopdepth),
 112         LC_OPT_ENT_BOOL   ("improveknownpreds", "experimental (known preds cutting)", &improve_known_preds),
 113         LC_OPT_ENT_INT    ("rematbonus", "give bonus to rematerialisable nodes", &remat_bonus),
 114         LC_OPT_LAST
 115 };
 116
 117 static int loc_compare(const void *a, const void *b)
 118 {
 119         const loc_t *p = a;
 120         const loc_t *q = b;
 121         return p->time - q->time;
 122 }
 123
 124 void workset_print(const workset_t *w)
 125 {
 126         int i;
 127
 128         for(i = 0; i < w->len; ++i) {
 129                 ir_fprintf(stderr, "%+F %d\n", w->vals[i].node, w->vals[i].time);
 130         }
 131 }
 132
 133 /**
 134  * Alloc a new workset on obstack @p ob with maximum size @p max
 135  */
 136 static workset_t *new_workset(void)
 137 {
 138         workset_t *res;
 139         size_t     size = sizeof(*res) + n_regs * sizeof(res->vals[0]);
 140
 141         res  = obstack_alloc(&obst, size);
 142         memset(res, 0, size);
 143         return res;
 144 }
 145
 146 /**
 147  * Alloc a new instance on obstack and make it equal to @param workset
 148  */
 149 static workset_t *workset_clone(workset_t *workset)
 150 {
 151         workset_t *res;
 152         size_t size = sizeof(*res) + n_regs * sizeof(res->vals[0]);
 153         res = obstack_alloc(&obst, size);
 154         memcpy(res, workset, size);
 155         return res;
 156 }
 157
 158 /**
 159  * Copy workset @param src to @param tgt
 160  */
 161 static void workset_copy(workset_t *dest, const workset_t *src)
 162 {
 163         size_t size = sizeof(*src) + n_regs * sizeof(src->vals[0]);
 164         memcpy(dest, src, size);
 165 }
 166
 167 /**
 168  * Overwrites the current content array of @param ws with the
 169  * @param count locations given at memory @param locs.
 170  * Set the length of @param ws to count.
 171  */
 172 static void workset_bulk_fill(workset_t *workset, int count, const loc_t *locs)
 173 {
 174         workset->len = count;
 175         memcpy(&(workset->vals[0]), locs, count * sizeof(locs[0]));
 176 }
 177
 178 /**
 179  * Inserts the value @p val into the workset, iff it is not
 180  * already contained. The workset must not be full.
 181  */
 182 static void workset_insert(workset_t *workset, ir_node *val, bool spilled)
 183 {
 184         loc_t *loc;
 185         int    i;
 186         /* check for current regclass */
 187         assert(arch_irn_consider_in_reg_alloc(arch_env, cls, val));
 188
 189         /* check if val is already contained */
 190         for (i = 0; i < workset->len; ++i) {
 191                 loc = &workset->vals[i];
 192                 if (loc->node == val) {
 193                         if (spilled) {
 194                                 loc->spilled = true;
 195                         }
 196                         return;
 197                 }
 198         }
 199
 200         /* insert val */
 201         assert(workset->len < n_regs && "Workset already full!");
 202         loc           = &workset->vals[workset->len];
 203         loc->node     = val;
 204         loc->spilled  = spilled;
 205         loc->time     = TIME_UNDEFINED;
 206         workset->len++;
 207 }
 208
 209 /**
 210  * Removes all entries from this workset
 211  */
 212 static void workset_clear(workset_t *workset)
 213 {
 214         workset->len = 0;
 215 }
 216
 217 /**
 218  * Removes the value @p val from the workset if present.
 219  */
 220 static INLINE void workset_remove(workset_t *workset, ir_node *val)
 221 {
 222         int i;
 223         for(i = 0; i < workset->len; ++i) {
 224                 if (workset->vals[i].node == val) {
 225                         workset->vals[i] = workset->vals[--workset->len];
 226                         return;
 227                 }
 228         }
 229 }
 230
 231 static INLINE const loc_t *workset_contains(const workset_t *ws,
 232                                             const ir_node *val)
 233 {
 234         int i;
 235
 236         for (i = 0; i < ws->len; ++i) {
 237                 if (ws->vals[i].node == val)
 238                         return &ws->vals[i];
 239         }
 240
 241         return NULL;
 242 }
 243
 244 /**
 245  * Iterates over all values in the working set.
 246  * @p ws The workset to iterate
 247  * @p v  A variable to put the current value in
 248  * @p i  An integer for internal use
 249  */
 250 #define workset_foreach(ws, v, i)       for(i=0; \
 251                                                                                 v=(i < ws->len) ? ws->vals[i].node : NULL, i < ws->len; \
 252                                                                                 ++i)
 253
 254 #define workset_set_time(ws, i, t) (ws)->vals[i].time=t
 255 #define workset_get_time(ws, i) (ws)->vals[i].time
 256 #define workset_set_length(ws, length) (ws)->len = length
 257 #define workset_get_length(ws) ((ws)->len)
 258 #define workset_get_val(ws, i) ((ws)->vals[i].node)
 259 #define workset_sort(ws) qsort((ws)->vals, (ws)->len, sizeof((ws)->vals[0]), loc_compare);
 260
 261 typedef struct _block_info_t
 262 {
 263         workset_t *start_workset;
 264         workset_t *end_workset;
 265 } block_info_t;
 266
 267
 268 static void *new_block_info(void)
 269 {
 270         block_info_t *res = obstack_alloc(&obst, sizeof(res[0]));
 271         memset(res, 0, sizeof(res[0]));
 272
 273         return res;
 274 }
 275
 276 #define get_block_info(block)        ((block_info_t *)get_irn_link(block))
 277 #define set_block_info(block, info)  set_irn_link(block, info)
 278
 279 /**
 280  * @return The distance to the next use or 0 if irn has dont_spill flag set
 281  */
 282 static INLINE unsigned get_distance(ir_node *from, unsigned from_step,
 283                                     const ir_node *def, int skip_from_uses)
 284 {
 285         be_next_use_t use;
 286         int           flags = arch_irn_get_flags(arch_env, def);
 287         unsigned      costs;
 288         unsigned      time;
 289
 290         assert(! (flags & arch_irn_flags_ignore));
 291
 292         use  = be_get_next_use(uses, from, from_step, 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 (flags & 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 = alloca(n_regs * sizeof(to_insert[0]));
 323         bool     *spilled   = alloca(n_regs * sizeof(spilled[0]));
 324         ir_node  *val;
 325         int       i;
 326         int       len;
 327         int       spills_needed;
 328         int       demand;
 329         int       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                 ir_node   *curr_bb  = NULL;
 363                 workset_t *ws_start = NULL;
 364
 365                 if (move_spills) {
 366                         curr_bb  = get_nodes_block(instr);
 367                         ws_start = get_block_info(curr_bb)->start_workset;
 368                 }
 369
 370                 DB((dbg, DBG_DECIDE, "    disposing %d values\n", spills_needed));
 371
 372                 /* calculate current next-use distance for live values */
 373                 for (i = 0; i < len; ++i) {
 374                         ir_node  *val  = workset_get_val(ws, i);
 375                         unsigned  dist = get_distance(instr, instr_nr, val, !is_usage);
 376                         workset_set_time(ws, i, dist);
 377                 }
 378
 379                 /* sort entries by increasing nextuse-distance*/
 380                 workset_sort(ws);
 381
 382                 for (i = len - spills_needed; i < len; ++i) {
 383                         ir_node *val = ws->vals[i].node;
 384
 385                         DB((dbg, DBG_DECIDE, "    disposing node %+F (%u)\n", val,
 386                              workset_get_time(ws, i)));
 387
 388                         if (move_spills) {
 389                                 if (!USES_IS_INFINITE(ws->vals[i].time)
 390                                                 && !ws->vals[i].spilled) {
 391                                         ir_node *after_pos = sched_prev(instr);
 392                                         DB((dbg, DBG_DECIDE, "Spill %+F after node %+F\n", val,
 393                                                 after_pos));
 394                                         be_add_spill(senv, val, after_pos);
 395                                 }
 396                         }
 397                 }
 398
 399                 /* kill the last 'demand' entries in the array */
 400                 workset_set_length(ws, len - spills_needed);
 401         }
 402
 403         /* 3. Insert the new values into the workset */
 404         for (i = 0; i < demand; ++i) {
 405                 ir_node *val = to_insert[i];
 406
 407                 workset_insert(ws, val, spilled[i]);
 408         }
 409 }
 410
 411 enum {
 412         AVAILABLE_EVERYWHERE,
 413         AVAILABLE_NOWHERE,
 414         AVAILABLE_PARTLY,
 415         AVAILABLE_UNKNOWN
 416 };
 417
 418 static unsigned available_in_all_preds(workset_t* const* pred_worksets,
 419                                        size_t n_pred_worksets,
 420                                        const ir_node *value, bool is_local_phi)
 421 {
 422         size_t i;
 423         bool   avail_everywhere = true;
 424         bool   avail_nowhere    = true;
 425
 426         assert(n_pred_worksets > 0);
 427
 428         /* value available in all preds? */
 429         for (i = 0; i < n_pred_worksets; ++i) {
 430                 bool             found     = false;
 431                 const workset_t *p_workset = pred_worksets[i];
 432                 int              p_len     = workset_get_length(p_workset);
 433                 int              p_i;
 434                 const ir_node   *l_value;
 435
 436                 if (is_local_phi) {
 437                         assert(is_Phi(value));
 438                         l_value = get_irn_n(value, i);
 439                 } else {
 440                         l_value = value;
 441                 }
 442
 443                 for (p_i = 0; p_i < p_len; ++p_i) {
 444                         const loc_t *p_l = &p_workset->vals[p_i];
 445                         if (p_l->node != l_value)
 446                                 continue;
 447
 448                         found = true;
 449                         break;
 450                 }
 451
 452                 if (found) {
 453                         avail_nowhere = false;
 454                 } else {
 455                         avail_everywhere = false;
 456                 }
 457         }
 458
 459         if (avail_everywhere) {
 460                 assert(!avail_nowhere);
 461                 return AVAILABLE_EVERYWHERE;
 462         } else if (avail_nowhere) {
 463                 return AVAILABLE_NOWHERE;
 464         } else {
 465                 return AVAILABLE_PARTLY;
 466         }
 467 }
 468
 469 /** Decides whether a specific node should be in the start workset or not
 470  *
 471  * @param env      belady environment
 472  * @param first
 473  * @param node     the node to test
 474  * @param loop     the loop of the node
 475  */
 476 static loc_t to_take_or_not_to_take(ir_node* first, ir_node *node,
 477                                     ir_loop *loop, unsigned available)
 478 {
 479         be_next_use_t next_use;
 480         loc_t         loc;
 481
 482         loc.time    = USES_INFINITY;
 483         loc.node    = node;
 484         loc.spilled = false;
 485
 486         if (!arch_irn_consider_in_reg_alloc(arch_env, cls, node)) {
 487                 loc.time = USES_INFINITY;
 488                 return loc;
 489         }
 490
 491         /* We have to keep nonspillable nodes in the workingset */
 492         if (arch_irn_get_flags(arch_env, node) & arch_irn_flags_dont_spill) {
 493                 loc.time = 0;
 494                 DB((dbg, DBG_START, "    %+F taken (dontspill node)\n", node, loc.time));
 495                 return loc;
 496         }
 497
 498         next_use = be_get_next_use(uses, first, 0, node, 0);
 499         if (USES_IS_INFINITE(next_use.time)) {
 500                 // the nodes marked as live in shouldn't be dead, so it must be a phi
 501                 assert(is_Phi(node));
 502                 loc.time = USES_INFINITY;
 503                 DB((dbg, DBG_START, "    %+F not taken (dead)\n", node));
 504                 return loc;
 505         }
 506
 507         loc.time = next_use.time;
 508
 509         if (improve_known_preds) {
 510                 if (available == AVAILABLE_EVERYWHERE) {
 511                         DB((dbg, DBG_START, "    %+F taken (%u, live in all preds)\n",
 512                             node, loc.time));
 513                         return loc;
 514                 } else if(available == AVAILABLE_NOWHERE) {
 515                         DB((dbg, DBG_START, "    %+F not taken (%u, live in no pred)\n",
 516                             node, loc.time));
 517                         loc.time = USES_INFINITY;
 518                         return loc;
 519                 }
 520         }
 521
 522         if (!respectloopdepth || next_use.outermost_loop >= get_loop_depth(loop)) {
 523                 DB((dbg, DBG_START, "    %+F taken (%u, loop %d)\n", node, loc.time,
 524                     next_use.outermost_loop));
 525         } else {
 526                 loc.time = USES_PENDING;
 527                 DB((dbg, DBG_START, "    %+F delayed (outerdepth %d < loopdepth %d)\n",
 528                     node, next_use.outermost_loop, get_loop_depth(loop)));
 529         }
 530
 531         return loc;
 532 }
 533
 534 /**
 535  * Computes the start-workset for a block with multiple predecessors. We assume
 536  * that at least 1 of the predeccesors is a back-edge which means we're at the
 537  * beginning of a loop. We try to reload as much values as possible now so they
 538  * don't get reloaded inside the loop.
 539  */
 540 static void decide_start_workset(const ir_node *block)
 541 {
 542         ir_loop    *loop = get_irn_loop(block);
 543         ir_node    *first;
 544         ir_node    *node;
 545         loc_t       loc;
 546         loc_t      *starters;
 547         loc_t      *delayed;
 548         int         i, len, ws_count;
 549         int             free_slots, free_pressure_slots;
 550         unsigned    pressure;
 551         int         arity;
 552         workset_t **pred_worksets;
 553         bool        all_preds_known;
 554
 555         /* check predecessors */
 556         arity           = get_irn_arity(block);
 557         pred_worksets   = alloca(sizeof(pred_worksets[0]) * arity);
 558         all_preds_known = true;
 559         for(i = 0; i < arity; ++i) {
 560                 ir_node      *pred_block = get_Block_cfgpred_block(block, i);
 561                 block_info_t *pred_info  = get_block_info(pred_block);
 562
 563                 if (pred_info == NULL) {
 564                         pred_worksets[i] = NULL;
 565                         all_preds_known  = false;
 566                 } else {
 567                         pred_worksets[i] = pred_info->end_workset;
 568                 }
 569         }
 570
 571         /* Collect all values living at start of block */
 572         starters = NEW_ARR_F(loc_t, 0);
 573         delayed  = NEW_ARR_F(loc_t, 0);
 574
 575         DB((dbg, DBG_START, "Living at start of %+F:\n", block));
 576         first = sched_first(block);
 577
 578         /* check all Phis first */
 579         sched_foreach(block, node) {
 580                 unsigned available;
 581
 582                 if (! is_Phi(node))
 583                         break;
 584                 if (!arch_irn_consider_in_reg_alloc(arch_env, cls, node))
 585                         continue;
 586
 587                 if (all_preds_known) {
 588                         available = available_in_all_preds(pred_worksets, arity, node, true);
 589                 } else {
 590                         available = AVAILABLE_UNKNOWN;
 591                 }
 592
 593                 loc = to_take_or_not_to_take(first, node, loop, available);
 594
 595                 if (! USES_IS_INFINITE(loc.time)) {
 596                         if (USES_IS_PENDING(loc.time))
 597                                 ARR_APP1(loc_t, delayed, loc);
 598                         else
 599                                 ARR_APP1(loc_t, starters, loc);
 600                 } else {
 601                         be_spill_phi(senv, node);
 602                 }
 603         }
 604
 605         /* check all Live-Ins */
 606         be_lv_foreach(lv, block, be_lv_state_in, i) {
 607                 ir_node *node = be_lv_get_irn(lv, block, i);
 608                 unsigned available;
 609
 610                 if (all_preds_known) {
 611                         available = available_in_all_preds(pred_worksets, arity, node, false);
 612                 } else {
 613                         available = AVAILABLE_UNKNOWN;
 614                 }
 615
 616                 loc = to_take_or_not_to_take(first, node, loop, available);
 617
 618                 if (! USES_IS_INFINITE(loc.time)) {
 619                         if (USES_IS_PENDING(loc.time))
 620                                 ARR_APP1(loc_t, delayed, loc);
 621                         else
 622                                 ARR_APP1(loc_t, starters, loc);
 623                 }
 624         }
 625
 626         pressure            = be_get_loop_pressure(loop_ana, cls, loop);
 627         assert(ARR_LEN(delayed) <= (signed)pressure);
 628         free_slots          = n_regs - ARR_LEN(starters);
 629         free_pressure_slots = n_regs - (pressure - ARR_LEN(delayed));
 630         free_slots          = MIN(free_slots, free_pressure_slots);
 631
 632         /* so far we only put nodes into the starters list that are used inside
 633          * the loop. If register pressure in the loop is low then we can take some
 634          * values and let them live through the loop */
 635         DB((dbg, DBG_START, "Loop pressure %d, taking %d delayed vals\n",
 636             pressure, free_slots));
 637         if (free_slots > 0) {
 638                 qsort(delayed, ARR_LEN(delayed), sizeof(delayed[0]), loc_compare);
 639
 640                 for (i = 0; i < ARR_LEN(delayed) && free_slots > 0; ++i) {
 641                         int    p, arity;
 642                         loc_t *loc = & delayed[i];
 643
 644                         /* don't use values which are dead in a known predecessors
 645                          * to not induce unnecessary reloads */
 646                         arity = get_irn_arity(block);
 647                         for (p = 0; p < arity; ++p) {
 648                                 ir_node      *pred_block = get_Block_cfgpred_block(block, p);
 649                                 block_info_t *pred_info  = get_block_info(pred_block);
 650
 651                                 if (pred_info == NULL)
 652                                         continue;
 653
 654                                 if (!workset_contains(pred_info->end_workset, loc->node)) {
 655                                         DB((dbg, DBG_START,
 656                                             "    delayed %+F not live at pred %+F\n", loc->node,
 657                                             pred_block));
 658                                         goto skip_delayed;
 659                                 }
 660                         }
 661
 662                         DB((dbg, DBG_START, "    delayed %+F taken\n", loc->node));
 663                         ARR_APP1(loc_t, starters, *loc);
 664                         loc->node = NULL;
 665                         --free_slots;
 666                 skip_delayed:
 667                         ;
 668                 }
 669         }
 670
 671         /* spill phis (the actual phis not just their values) that are in this block
 672          * but not in the start workset */
 673         for (i = ARR_LEN(delayed) - 1; i >= 0; --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(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         int              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 it's 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         instr_nr = 0;
 799         /* TODO: this leaks (into the obstack)... */
 800         new_vals = new_workset();
 801
 802         sched_foreach(block, irn) {
 803                 int i, arity;
 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(arch_env, 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                 if (get_irn_mode(irn) == mode_T) {
 830                         const ir_edge_t *edge;
 831
 832                         foreach_out_edge(irn, edge) {
 833                                 ir_node *proj = get_edge_src_irn(edge);
 834                                 if (!arch_irn_consider_in_reg_alloc(arch_env, cls, proj))
 835                                         continue;
 836                                 workset_insert(new_vals, proj, false);
 837                         }
 838                 } else {
 839                         if (!arch_irn_consider_in_reg_alloc(arch_env, cls, irn))
 840                                 continue;
 841                         workset_insert(new_vals, irn, false);
 842                 }
 843                 displace(new_vals, 0);
 844
 845                 instr_nr++;
 846         }
 847
 848         /* Remember end-workset for this block */
 849         block_info->end_workset = workset_clone(ws);
 850         DB((dbg, DBG_WSETS, "End workset for %+F:\n", block));
 851         workset_foreach(ws, irn, iter)
 852                 DB((dbg, DBG_WSETS, "  %+F (%u)\n", irn,
 853                      workset_get_time(ws, iter)));
 854 }
 855
 856 /**
 857  * 'decide' is block-local and makes assumptions
 858  * about the set of live-ins. Thus we must adapt the
 859  * live-outs to the live-ins at each block-border.
 860  */
 861 static void fix_block_borders(ir_node *block, void *data)
 862 {
 863         workset_t    *start_workset;
 864         int           arity;
 865         int           i;
 866         int           iter;
 867         (void) data;
 868
 869         DB((dbg, DBG_FIX, "\n"));
 870         DB((dbg, DBG_FIX, "Fixing %+F\n", block));
 871
 872         arity = get_irn_arity(block);
 873         /* can happen for endless loops */
 874         if (arity == 0)
 875                 return;
 876
 877         start_workset = get_block_info(block)->start_workset;
 878
 879         /* process all pred blocks */
 880         for (i = 0; i < arity; ++i) {
 881                 ir_node   *pred = get_Block_cfgpred_block(block, i);
 882                 workset_t *pred_end_workset = get_block_info(pred)->end_workset;
 883                 ir_node   *node;
 884
 885                 DB((dbg, DBG_FIX, "  Pred %+F\n", pred));
 886
 887                 /* spill all values not used anymore */
 888                 workset_foreach(pred_end_workset, node, iter) {
 889                         ir_node *n2;
 890                         int      iter2;
 891                         bool     found = false;
 892                         workset_foreach(start_workset, n2, iter2) {
 893                                 if (n2 == node) {
 894                                         found = true;
 895                                         break;
 896                                 }
 897                                 /* note that we do not look at phi inputs, becuase the values
 898                                  * will be either live-end and need no spill or
 899                                  * they have other users in which must be somewhere else in the
 900                                  * workset */
 901                         }
 902
 903                         if (found)
 904                                 continue;
 905
 906                         if (move_spills && be_is_live_in(lv, block, node)
 907                                         && !pred_end_workset->vals[iter].spilled) {
 908                                 ir_node *insert_point;
 909                                 if (arity > 1) {
 910                                         insert_point = be_get_end_of_block_insertion_point(pred);
 911                                         insert_point = sched_prev(insert_point);
 912                                 } else {
 913                                         insert_point = block;
 914                                 }
 915                                 DB((dbg, DBG_SPILL, "Spill %+F after %+F\n", node,
 916                                      insert_point));
 917                                 be_add_spill(senv, node, insert_point);
 918                         }
 919                 }
 920
 921                 /* reload missing values in predecessors, add missing spills */
 922                 workset_foreach(start_workset, node, iter) {
 923                         const loc_t *l    = &start_workset->vals[iter];
 924                         const loc_t *pred_loc;
 925
 926                         /* if node is a phi of the current block we reload
 927                          * the corresponding argument, else node itself */
 928                         if (is_Phi(node) && get_nodes_block(node) == block) {
 929                                 node = get_irn_n(node, i);
 930                                 assert(!l->spilled);
 931
 932                                 /* we might have unknowns as argument for the phi */
 933                                 if (!arch_irn_consider_in_reg_alloc(arch_env, cls, node))
 934                                         continue;
 935                         }
 936
 937                         /* check if node is in a register at end of pred */
 938                         pred_loc = workset_contains(pred_end_workset, node);
 939                         if (pred_loc != NULL) {
 940                                 /* we might have to spill value on this path */
 941                                 if (move_spills && !pred_loc->spilled && l->spilled) {
 942                                         ir_node *insert_point
 943                                                 = be_get_end_of_block_insertion_point(pred);
 944                                         insert_point = sched_prev(insert_point);
 945                                         DB((dbg, DBG_SPILL, "Spill %+F after %+F\n", node,
 946                                             insert_point));
 947                                         be_add_spill(senv, node, insert_point);
 948                                 }
 949                         } else {
 950                                 /* node is not in register at the end of pred -> reload it */
 951                                 DB((dbg, DBG_FIX, "    reload %+F\n", node));
 952                                 DB((dbg, DBG_SPILL, "Reload %+F before %+F,%d\n", node, block, i));
 953                                 be_add_reload_on_edge(senv, node, block, i, cls, 1);
 954                         }
 955                 }
 956         }
 957 }
 958
 959 static void add_block(ir_node *block, void *data)
 960 {
 961         (void) data;
 962         ARR_APP1(ir_node*, blocklist, block);
 963 }
 964
 965 static void be_spill_belady(be_irg_t *birg, const arch_register_class_t *rcls)
 966 {
 967         int i;
 968         ir_graph *irg = be_get_birg_irg(birg);
 969
 970         be_liveness_assure_sets(be_assure_liveness(birg));
 971
 972         stat_ev_tim_push();
 973         /* construct control flow loop tree */
 974         if (! (get_irg_loopinfo_state(irg) & loopinfo_cf_consistent)) {
 975                 construct_cf_backedges(irg);
 976         }
 977         stat_ev_tim_pop("belady_time_backedges");
 978
 979         stat_ev_tim_push();
 980         be_clear_links(irg);
 981         stat_ev_tim_pop("belady_time_clear_links");
 982
 983         ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
 984
 985         /* init belady env */
 986         stat_ev_tim_push();
 987         obstack_init(&obst);
 988         arch_env  = birg->main_env->arch_env;
 989         cls       = rcls;
 990         lv        = be_get_birg_liveness(birg);
 991         n_regs    = cls->n_regs - be_put_ignore_regs(birg, cls, NULL);
 992         ws        = new_workset();
 993         uses      = be_begin_uses(irg, lv);
 994         loop_ana  = be_new_loop_pressure(birg, cls);
 995         senv      = be_new_spill_env(birg);
 996         blocklist = NEW_ARR_F(ir_node*, 0);
 997         irg_block_edges_walk(get_irg_start_block(irg), NULL, add_block, NULL);
 998         stat_ev_tim_pop("belady_time_init");
 999
1000         stat_ev_tim_push();
1001         /* walk blocks in reverse postorder */
1002         for (i = ARR_LEN(blocklist) - 1; i >= 0; --i) {
1003                 process_block(blocklist[i]);
1004         }
1005         DEL_ARR_F(blocklist);
1006         stat_ev_tim_pop("belady_time_belady");
1007
1008         stat_ev_tim_push();
1009         /* belady was block-local, fix the global flow by adding reloads on the
1010          * edges */
1011         irg_block_walk_graph(irg, fix_block_borders, NULL, NULL);
1012         stat_ev_tim_pop("belady_time_fix_borders");
1013
1014         ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
1015
1016         /* Insert spill/reload nodes into the graph and fix usages */
1017         be_insert_spills_reloads(senv);
1018
1019         /* clean up */
1020         be_delete_spill_env(senv);
1021         be_end_uses(uses);
1022         be_free_loop_pressure(loop_ana);
1023         obstack_free(&obst, NULL);
1024 }
1025
1026 void be_init_spillbelady(void)
1027 {
1028         static be_spiller_t belady_spiller = {
1029                 be_spill_belady
1030         };
1031         lc_opt_entry_t *be_grp       = lc_opt_get_grp(firm_opt_get_root(), "be");
1032         lc_opt_entry_t *belady_group = lc_opt_get_grp(be_grp, "belady");
1033         lc_opt_add_table(belady_group, options);
1034
1035         be_register_spiller("belady", &belady_spiller);
1036         FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
1037 }
1038
1039 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady);