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_t.h"
  45 #include "beuses.h"
  46 #include "besched_t.h"
  47 #include "beirgmod.h"
  48 #include "belive_t.h"
  49 #include "benode_t.h"
  50 #include "bechordal_t.h"
  51 #include "bespilloptions.h"
  52 #include "beloopana.h"
  53 #include "beirg_t.h"
  54 #include "bespill.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 bool                         move_spills      = true;
 100 static bool                         respectloopdepth = true;
 101 static bool                         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 void workset_print(const workset_t *w)
 122 {
 123         int i;
 124
 125         for(i = 0; i < w->len; ++i) {
 126                 ir_fprintf(stderr, "%+F %d\n", w->vals[i].node, w->vals[i].time);
 127         }
 128 }
 129
 130 /**
 131  * Alloc a new workset on obstack @p ob with maximum size @p max
 132  */
 133 static workset_t *new_workset(void)
 134 {
 135         workset_t *res;
 136         size_t     size = sizeof(*res) + n_regs * sizeof(res->vals[0]);
 137
 138         res  = obstack_alloc(&obst, size);
 139         memset(res, 0, size);
 140         return res;
 141 }
 142
 143 /**
 144  * Alloc a new instance on obstack and make it equal to @param workset
 145  */
 146 static workset_t *workset_clone(workset_t *workset)
 147 {
 148         workset_t *res;
 149         size_t size = sizeof(*res) + n_regs * sizeof(res->vals[0]);
 150         res = obstack_alloc(&obst, size);
 151         memcpy(res, workset, size);
 152         return res;
 153 }
 154
 155 /**
 156  * Copy workset @param src to @param tgt
 157  */
 158 static void workset_copy(workset_t *dest, const workset_t *src)
 159 {
 160         size_t size = sizeof(*src) + n_regs * sizeof(src->vals[0]);
 161         memcpy(dest, src, size);
 162 }
 163
 164 /**
 165  * Overwrites the current content array of @param ws with the
 166  * @param count locations given at memory @param locs.
 167  * Set the length of @param ws to count.
 168  */
 169 static void workset_bulk_fill(workset_t *workset, int count, const loc_t *locs)
 170 {
 171         workset->len = count;
 172         memcpy(&(workset->vals[0]), locs, count * sizeof(locs[0]));
 173 }
 174
 175 /**
 176  * Inserts the value @p val into the workset, iff it is not
 177  * already contained. The workset must not be full.
 178  */
 179 static void workset_insert(workset_t *workset, ir_node *val, bool spilled)
 180 {
 181         loc_t *loc;
 182         int    i;
 183         /* check for current regclass */
 184         assert(arch_irn_consider_in_reg_alloc(cls, val));
 185
 186         /* check if val is already contained */
 187         for (i = 0; i < workset->len; ++i) {
 188                 loc = &workset->vals[i];
 189                 if (loc->node == val) {
 190                         if (spilled) {
 191                                 loc->spilled = true;
 192                         }
 193                         return;
 194                 }
 195         }
 196
 197         /* insert val */
 198         assert(workset->len < n_regs && "Workset already full!");
 199         loc           = &workset->vals[workset->len];
 200         loc->node     = val;
 201         loc->spilled  = spilled;
 202         loc->time     = TIME_UNDEFINED;
 203         workset->len++;
 204 }
 205
 206 /**
 207  * Removes all entries from this workset
 208  */
 209 static void workset_clear(workset_t *workset)
 210 {
 211         workset->len = 0;
 212 }
 213
 214 /**
 215  * Removes the value @p val from the workset if present.
 216  */
 217 static INLINE void workset_remove(workset_t *workset, ir_node *val)
 218 {
 219         int i;
 220         for(i = 0; i < workset->len; ++i) {
 221                 if (workset->vals[i].node == val) {
 222                         workset->vals[i] = workset->vals[--workset->len];
 223                         return;
 224                 }
 225         }
 226 }
 227
 228 static INLINE const loc_t *workset_contains(const workset_t *ws,
 229                                             const ir_node *val)
 230 {
 231         int i;
 232
 233         for (i = 0; i < ws->len; ++i) {
 234                 if (ws->vals[i].node == val)
 235                         return &ws->vals[i];
 236         }
 237
 238         return NULL;
 239 }
 240
 241 /**
 242  * Iterates over all values in the working set.
 243  * @p ws The workset to iterate
 244  * @p v  A variable to put the current value in
 245  * @p i  An integer for internal use
 246  */
 247 #define workset_foreach(ws, v, i)       for(i=0; \
 248                                                                                 v=(i < ws->len) ? ws->vals[i].node : NULL, i < ws->len; \
 249                                                                                 ++i)
 250
 251 #define workset_set_time(ws, i, t) (ws)->vals[i].time=t
 252 #define workset_get_time(ws, i) (ws)->vals[i].time
 253 #define workset_set_length(ws, length) (ws)->len = length
 254 #define workset_get_length(ws) ((ws)->len)
 255 #define workset_get_val(ws, i) ((ws)->vals[i].node)
 256 #define workset_sort(ws) qsort((ws)->vals, (ws)->len, sizeof((ws)->vals[0]), loc_compare);
 257
 258 typedef struct _block_info_t
 259 {
 260         workset_t *start_workset;
 261         workset_t *end_workset;
 262 } block_info_t;
 263
 264
 265 static void *new_block_info(void)
 266 {
 267         block_info_t *res = obstack_alloc(&obst, sizeof(res[0]));
 268         memset(res, 0, sizeof(res[0]));
 269
 270         return res;
 271 }
 272
 273 #define get_block_info(block)        ((block_info_t *)get_irn_link(block))
 274 #define set_block_info(block, info)  set_irn_link(block, info)
 275
 276 /**
 277  * @return The distance to the next use or 0 if irn has dont_spill flag set
 278  */
 279 static INLINE unsigned get_distance(ir_node *from, unsigned from_step,
 280                                     const ir_node *def, int skip_from_uses)
 281 {
 282         be_next_use_t use;
 283         int           flags = arch_irn_get_flags(def);
 284         unsigned      costs;
 285         unsigned      time;
 286
 287         assert(! (flags & arch_irn_flags_ignore));
 288
 289         use  = be_get_next_use(uses, from, from_step, def, skip_from_uses);
 290         time = use.time;
 291         if (USES_IS_INFINITE(time))
 292                 return USES_INFINITY;
 293
 294         /* We have to keep nonspillable nodes in the workingset */
 295         if (flags & arch_irn_flags_dont_spill)
 296                 return 0;
 297
 298         /* give some bonus to rematerialisable nodes */
 299         if (remat_bonus > 0) {
 300                 costs = be_get_reload_costs_no_weight(senv, def, use.before);
 301                 assert(costs * remat_bonus < 1000);
 302                 time  += 1000 - (costs * remat_bonus);
 303         }
 304
 305         return time;
 306 }
 307
 308 /**
 309  * Performs the actions necessary to grant the request that:
 310  * - new_vals can be held in registers
 311  * - as few as possible other values are disposed
 312  * - the worst values get disposed
 313  *
 314  * @p is_usage indicates that the values in new_vals are used (not defined)
 315  * In this case reloads must be performed
 316  */
 317 static void displace(workset_t *new_vals, int is_usage)
 318 {
 319         ir_node **to_insert = alloca(n_regs * sizeof(to_insert[0]));
 320         bool     *spilled   = alloca(n_regs * sizeof(spilled[0]));
 321         ir_node  *val;
 322         int       i;
 323         int       len;
 324         int       spills_needed;
 325         int       demand;
 326         int       iter;
 327
 328         /* 1. Identify the number of needed slots and the values to reload */
 329         demand = 0;
 330         workset_foreach(new_vals, val, iter) {
 331                 bool reloaded = false;
 332
 333                 if (! workset_contains(ws, val)) {
 334                         DB((dbg, DBG_DECIDE, "    insert %+F\n", val));
 335                         if (is_usage) {
 336                                 DB((dbg, DBG_SPILL, "Reload %+F before %+F\n", val, instr));
 337                                 be_add_reload(senv, val, instr, cls, 1);
 338                                 reloaded = true;
 339                         }
 340                 } else {
 341                         DB((dbg, DBG_DECIDE, "    %+F already in workset\n", val));
 342                         assert(is_usage);
 343                         /* remove the value from the current workset so it is not accidently
 344                          * spilled */
 345                         workset_remove(ws, val);
 346                 }
 347                 spilled[demand]   = reloaded;
 348                 to_insert[demand] = val;
 349                 ++demand;
 350         }
 351
 352         /* 2. Make room for at least 'demand' slots */
 353         len           = workset_get_length(ws);
 354         spills_needed = len + demand - n_regs;
 355         assert(spills_needed <= len);
 356
 357         /* Only make more free room if we do not have enough */
 358         if (spills_needed > 0) {
 359                 ir_node   *curr_bb  = NULL;
 360                 workset_t *ws_start = NULL;
 361
 362                 if (move_spills) {
 363                         curr_bb  = get_nodes_block(instr);
 364                         ws_start = get_block_info(curr_bb)->start_workset;
 365                 }
 366
 367                 DB((dbg, DBG_DECIDE, "    disposing %d values\n", spills_needed));
 368
 369                 /* calculate current next-use distance for live values */
 370                 for (i = 0; i < len; ++i) {
 371                         ir_node  *val  = workset_get_val(ws, i);
 372                         unsigned  dist = get_distance(instr, instr_nr, val, !is_usage);
 373                         workset_set_time(ws, i, dist);
 374                 }
 375
 376                 /* sort entries by increasing nextuse-distance*/
 377                 workset_sort(ws);
 378
 379                 for (i = len - spills_needed; i < len; ++i) {
 380                         ir_node *val = ws->vals[i].node;
 381
 382                         DB((dbg, DBG_DECIDE, "    disposing node %+F (%u)\n", val,
 383                              workset_get_time(ws, i)));
 384
 385                         if (move_spills) {
 386                                 if (!USES_IS_INFINITE(ws->vals[i].time)
 387                                                 && !ws->vals[i].spilled) {
 388                                         ir_node *after_pos = sched_prev(instr);
 389                                         DB((dbg, DBG_DECIDE, "Spill %+F after node %+F\n", val,
 390                                                 after_pos));
 391                                         be_add_spill(senv, val, after_pos);
 392                                 }
 393                         }
 394                 }
 395
 396                 /* kill the last 'demand' entries in the array */
 397                 workset_set_length(ws, len - spills_needed);
 398         }
 399
 400         /* 3. Insert the new values into the workset */
 401         for (i = 0; i < demand; ++i) {
 402                 ir_node *val = to_insert[i];
 403
 404                 workset_insert(ws, val, spilled[i]);
 405         }
 406 }
 407
 408 enum {
 409         AVAILABLE_EVERYWHERE,
 410         AVAILABLE_NOWHERE,
 411         AVAILABLE_PARTLY,
 412         AVAILABLE_UNKNOWN
 413 };
 414
 415 static unsigned available_in_all_preds(workset_t* const* pred_worksets,
 416                                        size_t n_pred_worksets,
 417                                        const ir_node *value, bool is_local_phi)
 418 {
 419         size_t i;
 420         bool   avail_everywhere = true;
 421         bool   avail_nowhere    = true;
 422
 423         assert(n_pred_worksets > 0);
 424
 425         /* value available in all preds? */
 426         for (i = 0; i < n_pred_worksets; ++i) {
 427                 bool             found     = false;
 428                 const workset_t *p_workset = pred_worksets[i];
 429                 int              p_len     = workset_get_length(p_workset);
 430                 int              p_i;
 431                 const ir_node   *l_value;
 432
 433                 if (is_local_phi) {
 434                         assert(is_Phi(value));
 435                         l_value = get_irn_n(value, i);
 436                 } else {
 437                         l_value = value;
 438                 }
 439
 440                 for (p_i = 0; p_i < p_len; ++p_i) {
 441                         const loc_t *p_l = &p_workset->vals[p_i];
 442                         if (p_l->node != l_value)
 443                                 continue;
 444
 445                         found = true;
 446                         break;
 447                 }
 448
 449                 if (found) {
 450                         avail_nowhere = false;
 451                 } else {
 452                         avail_everywhere = false;
 453                 }
 454         }
 455
 456         if (avail_everywhere) {
 457                 assert(!avail_nowhere);
 458                 return AVAILABLE_EVERYWHERE;
 459         } else if (avail_nowhere) {
 460                 return AVAILABLE_NOWHERE;
 461         } else {
 462                 return AVAILABLE_PARTLY;
 463         }
 464 }
 465
 466 /** Decides whether a specific node should be in the start workset or not
 467  *
 468  * @param env      belady environment
 469  * @param first
 470  * @param node     the node to test
 471  * @param loop     the loop of the node
 472  */
 473 static loc_t to_take_or_not_to_take(ir_node* first, ir_node *node,
 474                                     ir_loop *loop, unsigned available)
 475 {
 476         be_next_use_t next_use;
 477         loc_t         loc;
 478
 479         loc.time    = USES_INFINITY;
 480         loc.node    = node;
 481         loc.spilled = false;
 482
 483         if (!arch_irn_consider_in_reg_alloc(cls, node)) {
 484                 loc.time = USES_INFINITY;
 485                 return loc;
 486         }
 487
 488         /* We have to keep nonspillable nodes in the workingset */
 489         if (arch_irn_get_flags(node) & arch_irn_flags_dont_spill) {
 490                 loc.time = 0;
 491                 DB((dbg, DBG_START, "    %+F taken (dontspill node)\n", node, loc.time));
 492                 return loc;
 493         }
 494
 495         next_use = be_get_next_use(uses, first, 0, node, 0);
 496         if (USES_IS_INFINITE(next_use.time)) {
 497                 // the nodes marked as live in shouldn't be dead, so it must be a phi
 498                 assert(is_Phi(node));
 499                 loc.time = USES_INFINITY;
 500                 DB((dbg, DBG_START, "    %+F not taken (dead)\n", node));
 501                 return loc;
 502         }
 503
 504         loc.time = next_use.time;
 505
 506         if (improve_known_preds) {
 507                 if (available == AVAILABLE_EVERYWHERE) {
 508                         DB((dbg, DBG_START, "    %+F taken (%u, live in all preds)\n",
 509                             node, loc.time));
 510                         return loc;
 511                 } else if(available == AVAILABLE_NOWHERE) {
 512                         DB((dbg, DBG_START, "    %+F not taken (%u, live in no pred)\n",
 513                             node, loc.time));
 514                         loc.time = USES_INFINITY;
 515                         return loc;
 516                 }
 517         }
 518
 519         if (!respectloopdepth || next_use.outermost_loop >= get_loop_depth(loop)) {
 520                 DB((dbg, DBG_START, "    %+F taken (%u, loop %d)\n", node, loc.time,
 521                     next_use.outermost_loop));
 522         } else {
 523                 loc.time = USES_PENDING;
 524                 DB((dbg, DBG_START, "    %+F delayed (outerdepth %d < loopdepth %d)\n",
 525                     node, next_use.outermost_loop, get_loop_depth(loop)));
 526         }
 527
 528         return loc;
 529 }
 530
 531 /**
 532  * Computes the start-workset for a block with multiple predecessors. We assume
 533  * that at least 1 of the predeccesors is a back-edge which means we're at the
 534  * beginning of a loop. We try to reload as much values as possible now so they
 535  * don't get reloaded inside the loop.
 536  */
 537 static void decide_start_workset(const ir_node *block)
 538 {
 539         ir_loop    *loop = get_irn_loop(block);
 540         ir_node    *first;
 541         ir_node    *node;
 542         loc_t       loc;
 543         loc_t      *starters;
 544         loc_t      *delayed;
 545         int         i, len, ws_count;
 546         int             free_slots, free_pressure_slots;
 547         unsigned    pressure;
 548         int         arity;
 549         workset_t **pred_worksets;
 550         bool        all_preds_known;
 551
 552         /* check predecessors */
 553         arity           = get_irn_arity(block);
 554         pred_worksets   = alloca(sizeof(pred_worksets[0]) * arity);
 555         all_preds_known = true;
 556         for(i = 0; i < arity; ++i) {
 557                 ir_node      *pred_block = get_Block_cfgpred_block(block, i);
 558                 block_info_t *pred_info  = get_block_info(pred_block);
 559
 560                 if (pred_info == NULL) {
 561                         pred_worksets[i] = NULL;
 562                         all_preds_known  = false;
 563                 } else {
 564                         pred_worksets[i] = pred_info->end_workset;
 565                 }
 566         }
 567
 568         /* Collect all values living at start of block */
 569         starters = NEW_ARR_F(loc_t, 0);
 570         delayed  = NEW_ARR_F(loc_t, 0);
 571
 572         DB((dbg, DBG_START, "Living at start of %+F:\n", block));
 573         first = sched_first(block);
 574
 575         /* check all Phis first */
 576         sched_foreach(block, node) {
 577                 unsigned available;
 578
 579                 if (! is_Phi(node))
 580                         break;
 581                 if (!arch_irn_consider_in_reg_alloc(cls, node))
 582                         continue;
 583
 584                 if (all_preds_known) {
 585                         available = available_in_all_preds(pred_worksets, arity, node, true);
 586                 } else {
 587                         available = AVAILABLE_UNKNOWN;
 588                 }
 589
 590                 loc = to_take_or_not_to_take(first, node, loop, available);
 591
 592                 if (! USES_IS_INFINITE(loc.time)) {
 593                         if (USES_IS_PENDING(loc.time))
 594                                 ARR_APP1(loc_t, delayed, loc);
 595                         else
 596                                 ARR_APP1(loc_t, starters, loc);
 597                 } else {
 598                         be_spill_phi(senv, node);
 599                 }
 600         }
 601
 602         /* check all Live-Ins */
 603         be_lv_foreach(lv, block, be_lv_state_in, i) {
 604                 ir_node *node = be_lv_get_irn(lv, block, i);
 605                 unsigned available;
 606
 607                 if (all_preds_known) {
 608                         available = available_in_all_preds(pred_worksets, arity, node, false);
 609                 } else {
 610                         available = AVAILABLE_UNKNOWN;
 611                 }
 612
 613                 loc = to_take_or_not_to_take(first, node, loop, available);
 614
 615                 if (! USES_IS_INFINITE(loc.time)) {
 616                         if (USES_IS_PENDING(loc.time))
 617                                 ARR_APP1(loc_t, delayed, loc);
 618                         else
 619                                 ARR_APP1(loc_t, starters, loc);
 620                 }
 621         }
 622
 623         pressure            = be_get_loop_pressure(loop_ana, cls, loop);
 624         assert(ARR_LEN(delayed) <= (signed)pressure);
 625         free_slots          = n_regs - ARR_LEN(starters);
 626         free_pressure_slots = n_regs - (pressure - ARR_LEN(delayed));
 627         free_slots          = MIN(free_slots, free_pressure_slots);
 628
 629         /* so far we only put nodes into the starters list that are used inside
 630          * the loop. If register pressure in the loop is low then we can take some
 631          * values and let them live through the loop */
 632         DB((dbg, DBG_START, "Loop pressure %d, taking %d delayed vals\n",
 633             pressure, free_slots));
 634         if (free_slots > 0) {
 635                 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         for (i = ARR_LEN(delayed) - 1; i >= 0; --i) {
 673                 ir_node *node = delayed[i].node;
 674                 if (node == NULL || !is_Phi(node) || get_nodes_block(node) != block)
 675                         continue;
 676
 677                 DB((dbg, DBG_START, "    spilling delayed phi %+F\n", node));
 678                 be_spill_phi(senv, node);
 679         }
 680         DEL_ARR_F(delayed);
 681
 682         /* Sort start values by first use */
 683         qsort(starters, ARR_LEN(starters), sizeof(starters[0]), loc_compare);
 684
 685         /* Copy the best ones from starters to start workset */
 686         ws_count = MIN(ARR_LEN(starters), n_regs);
 687         workset_clear(ws);
 688         workset_bulk_fill(ws, ws_count, starters);
 689
 690         /* spill phis (the actual phis not just their values) that are in this block
 691          * but not in the start workset */
 692         len = ARR_LEN(starters);
 693         for (i = ws_count; i < len; ++i) {
 694                 ir_node *node = starters[i].node;
 695                 if (! is_Phi(node) || get_nodes_block(node) != block)
 696                         continue;
 697
 698                 DB((dbg, DBG_START, "    spilling phi %+F\n", node));
 699                 be_spill_phi(senv, node);
 700         }
 701
 702         DEL_ARR_F(starters);
 703
 704         /* determine spill status of the values: If there's 1 pred block (which
 705          * is no backedge) where the value is spilled then we must set it to
 706          * spilled here. */
 707         for(i = 0; i < ws_count; ++i) {
 708                 loc_t   *loc     = &ws->vals[i];
 709                 ir_node *value   = loc->node;
 710                 bool     spilled;
 711                 int      n;
 712
 713                 /* phis from this block aren't spilled */
 714                 if (get_nodes_block(value) == block) {
 715                         assert(is_Phi(value));
 716                         loc->spilled = false;
 717                         continue;
 718                 }
 719
 720                 /* determine if value was spilled on any predecessor */
 721                 spilled = false;
 722                 for(n = 0; n < arity; ++n) {
 723                         workset_t *pred_workset = pred_worksets[n];
 724                         int        p_len;
 725                         int        p;
 726
 727                         if (pred_workset == NULL)
 728                                 continue;
 729
 730                         p_len = workset_get_length(pred_workset);
 731                         for(p = 0; p < p_len; ++p) {
 732                                 loc_t *l = &pred_workset->vals[p];
 733
 734                                 if (l->node != value)
 735                                         continue;
 736
 737                                 if (l->spilled) {
 738                                         spilled = true;
 739                                 }
 740                                 break;
 741                         }
 742                 }
 743
 744                 loc->spilled = spilled;
 745         }
 746 }
 747
 748 /**
 749  * For the given block @p block, decide for each values
 750  * whether it is used from a register or is reloaded
 751  * before the use.
 752  */
 753 static void process_block(ir_node *block)
 754 {
 755         workset_t       *new_vals;
 756         ir_node         *irn;
 757         int              iter;
 758         block_info_t    *block_info;
 759         int              arity;
 760
 761         /* no need to process a block twice */
 762         assert(get_block_info(block) == NULL);
 763
 764         /* construct start workset */
 765         arity = get_Block_n_cfgpreds(block);
 766         if (arity == 0) {
 767                 /* no predecessor -> empty set */
 768                 workset_clear(ws);
 769         } else if (arity == 1) {
 770                 /* one predecessor, copy it's end workset */
 771                 ir_node      *pred_block = get_Block_cfgpred_block(block, 0);
 772                 block_info_t *pred_info  = get_block_info(pred_block);
 773
 774                 assert(pred_info != NULL);
 775                 workset_copy(ws, pred_info->end_workset);
 776         } else {
 777                 /* multiple predecessors, do more advanced magic :) */
 778                 decide_start_workset(block);
 779         }
 780
 781         DB((dbg, DBG_DECIDE, "\n"));
 782         DB((dbg, DBG_DECIDE, "Decide for %+F\n", block));
 783
 784         block_info = new_block_info();
 785         set_block_info(block, block_info);
 786
 787         DB((dbg, DBG_WSETS, "Start workset for %+F:\n", block));
 788         workset_foreach(ws, irn, iter) {
 789                 DB((dbg, DBG_WSETS, "  %+F (%u)\n", irn,
 790                      workset_get_time(ws, iter)));
 791         }
 792
 793         block_info->start_workset = workset_clone(ws);
 794
 795         /* process the block from start to end */
 796         DB((dbg, DBG_WSETS, "Processing...\n"));
 797         instr_nr = 0;
 798         /* TODO: this leaks (into the obstack)... */
 799         new_vals = new_workset();
 800
 801         sched_foreach(block, irn) {
 802                 int i, arity;
 803                 assert(workset_get_length(ws) <= n_regs);
 804
 805                 /* Phis are no real instr (see insert_starters()) */
 806                 if (is_Phi(irn)) {
 807                         continue;
 808                 }
 809                 DB((dbg, DBG_DECIDE, "  ...%+F\n", irn));
 810
 811                 /* set instruction in the workset */
 812                 instr = irn;
 813
 814                 /* allocate all values _used_ by this instruction */
 815                 workset_clear(new_vals);
 816                 for(i = 0, arity = get_irn_arity(irn); i < arity; ++i) {
 817                         ir_node *in = get_irn_n(irn, i);
 818                         if (!arch_irn_consider_in_reg_alloc(cls, in))
 819                                 continue;
 820
 821                         /* (note that "spilled" is irrelevant here) */
 822                         workset_insert(new_vals, in, false);
 823                 }
 824                 displace(new_vals, 1);
 825
 826                 /* allocate all values _defined_ by this instruction */
 827                 workset_clear(new_vals);
 828                 if (get_irn_mode(irn) == mode_T) {
 829                         const ir_edge_t *edge;
 830
 831                         foreach_out_edge(irn, edge) {
 832                                 ir_node *proj = get_edge_src_irn(edge);
 833                                 if (!arch_irn_consider_in_reg_alloc(cls, proj))
 834                                         continue;
 835                                 workset_insert(new_vals, proj, false);
 836                         }
 837                 } else {
 838                         if (!arch_irn_consider_in_reg_alloc(cls, irn))
 839                                 continue;
 840                         workset_insert(new_vals, irn, false);
 841                 }
 842                 displace(new_vals, 0);
 843
 844                 instr_nr++;
 845         }
 846
 847         /* Remember end-workset for this block */
 848         block_info->end_workset = workset_clone(ws);
 849         DB((dbg, DBG_WSETS, "End workset for %+F:\n", block));
 850         workset_foreach(ws, irn, iter)
 851                 DB((dbg, DBG_WSETS, "  %+F (%u)\n", irn,
 852                      workset_get_time(ws, iter)));
 853 }
 854
 855 /**
 856  * 'decide' is block-local and makes assumptions
 857  * about the set of live-ins. Thus we must adapt the
 858  * live-outs to the live-ins at each block-border.
 859  */
 860 static void fix_block_borders(ir_node *block, void *data)
 861 {
 862         workset_t    *start_workset;
 863         int           arity;
 864         int           i;
 865         int           iter;
 866         (void) data;
 867
 868         DB((dbg, DBG_FIX, "\n"));
 869         DB((dbg, DBG_FIX, "Fixing %+F\n", block));
 870
 871         arity = get_irn_arity(block);
 872         /* can happen for endless loops */
 873         if (arity == 0)
 874                 return;
 875
 876         start_workset = get_block_info(block)->start_workset;
 877
 878         /* process all pred blocks */
 879         for (i = 0; i < arity; ++i) {
 880                 ir_node   *pred = get_Block_cfgpred_block(block, i);
 881                 workset_t *pred_end_workset = get_block_info(pred)->end_workset;
 882                 ir_node   *node;
 883
 884                 DB((dbg, DBG_FIX, "  Pred %+F\n", pred));
 885
 886                 /* spill all values not used anymore */
 887                 workset_foreach(pred_end_workset, node, iter) {
 888                         ir_node *n2;
 889                         int      iter2;
 890                         bool     found = false;
 891                         workset_foreach(start_workset, n2, iter2) {
 892                                 if (n2 == node) {
 893                                         found = true;
 894                                         break;
 895                                 }
 896                                 /* note that we do not look at phi inputs, becuase the values
 897                                  * will be either live-end and need no spill or
 898                                  * they have other users in which must be somewhere else in the
 899                                  * workset */
 900                         }
 901
 902                         if (found)
 903                                 continue;
 904
 905                         if (move_spills && be_is_live_in(lv, block, node)
 906                                         && !pred_end_workset->vals[iter].spilled) {
 907                                 ir_node *insert_point;
 908                                 if (arity > 1) {
 909                                         insert_point = be_get_end_of_block_insertion_point(pred);
 910                                         insert_point = sched_prev(insert_point);
 911                                 } else {
 912                                         insert_point = block;
 913                                 }
 914                                 DB((dbg, DBG_SPILL, "Spill %+F after %+F\n", node,
 915                                      insert_point));
 916                                 be_add_spill(senv, node, insert_point);
 917                         }
 918                 }
 919
 920                 /* reload missing values in predecessors, add missing spills */
 921                 workset_foreach(start_workset, node, iter) {
 922                         const loc_t *l    = &start_workset->vals[iter];
 923                         const loc_t *pred_loc;
 924
 925                         /* if node is a phi of the current block we reload
 926                          * the corresponding argument, else node itself */
 927                         if (is_Phi(node) && get_nodes_block(node) == block) {
 928                                 node = get_irn_n(node, i);
 929                                 assert(!l->spilled);
 930
 931                                 /* we might have unknowns as argument for the phi */
 932                                 if (!arch_irn_consider_in_reg_alloc(cls, node))
 933                                         continue;
 934                         }
 935
 936                         /* check if node is in a register at end of pred */
 937                         pred_loc = workset_contains(pred_end_workset, node);
 938                         if (pred_loc != NULL) {
 939                                 /* we might have to spill value on this path */
 940                                 if (move_spills && !pred_loc->spilled && l->spilled) {
 941                                         ir_node *insert_point
 942                                                 = be_get_end_of_block_insertion_point(pred);
 943                                         insert_point = sched_prev(insert_point);
 944                                         DB((dbg, DBG_SPILL, "Spill %+F after %+F\n", node,
 945                                             insert_point));
 946                                         be_add_spill(senv, node, insert_point);
 947                                 }
 948                         } else {
 949                                 /* node is not in register at the end of pred -> reload it */
 950                                 DB((dbg, DBG_FIX, "    reload %+F\n", node));
 951                                 DB((dbg, DBG_SPILL, "Reload %+F before %+F,%d\n", node, block, i));
 952                                 be_add_reload_on_edge(senv, node, block, i, cls, 1);
 953                         }
 954                 }
 955         }
 956 }
 957
 958 static void add_block(ir_node *block, void *data)
 959 {
 960         (void) data;
 961         ARR_APP1(ir_node*, blocklist, block);
 962 }
 963
 964 static void be_spill_belady(be_irg_t *birg, const arch_register_class_t *rcls)
 965 {
 966         int i;
 967         ir_graph *irg = be_get_birg_irg(birg);
 968
 969         be_liveness_assure_sets(be_assure_liveness(birg));
 970
 971         stat_ev_tim_push();
 972         /* construct control flow loop tree */
 973         if (! (get_irg_loopinfo_state(irg) & loopinfo_cf_consistent)) {
 974                 construct_cf_backedges(irg);
 975         }
 976         stat_ev_tim_pop("belady_time_backedges");
 977
 978         stat_ev_tim_push();
 979         be_clear_links(irg);
 980         stat_ev_tim_pop("belady_time_clear_links");
 981
 982         ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
 983
 984         /* init belady env */
 985         stat_ev_tim_push();
 986         obstack_init(&obst);
 987         cls       = rcls;
 988         lv        = be_get_birg_liveness(birg);
 989         n_regs    = cls->n_regs - be_put_ignore_regs(birg, cls, NULL);
 990         ws        = new_workset();
 991         uses      = be_begin_uses(irg, lv);
 992         loop_ana  = be_new_loop_pressure(birg, cls);
 993         senv      = be_new_spill_env(birg);
 994         blocklist = NEW_ARR_F(ir_node*, 0);
 995         irg_block_edges_walk(get_irg_start_block(irg), NULL, add_block, NULL);
 996         stat_ev_tim_pop("belady_time_init");
 997
 998         stat_ev_tim_push();
 999         /* walk blocks in reverse postorder */
1000         for (i = ARR_LEN(blocklist) - 1; i >= 0; --i) {
1001                 process_block(blocklist[i]);
1002         }
1003         DEL_ARR_F(blocklist);
1004         stat_ev_tim_pop("belady_time_belady");
1005
1006         stat_ev_tim_push();
1007         /* belady was block-local, fix the global flow by adding reloads on the
1008          * edges */
1009         irg_block_walk_graph(irg, fix_block_borders, NULL, NULL);
1010         stat_ev_tim_pop("belady_time_fix_borders");
1011
1012         ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
1013
1014         /* Insert spill/reload nodes into the graph and fix usages */
1015         be_insert_spills_reloads(senv);
1016
1017         /* clean up */
1018         be_delete_spill_env(senv);
1019         be_end_uses(uses);
1020         be_free_loop_pressure(loop_ana);
1021         obstack_free(&obst, NULL);
1022 }
1023
1024 void be_init_spillbelady(void)
1025 {
1026         static be_spiller_t belady_spiller = {
1027                 be_spill_belady
1028         };
1029         lc_opt_entry_t *be_grp       = lc_opt_get_grp(firm_opt_get_root(), "be");
1030         lc_opt_entry_t *belady_group = lc_opt_get_grp(be_grp, "belady");
1031         lc_opt_add_table(belady_group, options);
1032
1033         be_register_spiller("belady", &belady_spiller);
1034         FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
1035 }
1036
1037 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady);