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 /**
  70  * An association between a node and a point in time.
  71  */
  72 typedef struct loc_t {
  73         ir_node          *node;
  74         unsigned          time;     /**< A use time (see beuses.h). */
  75         bool              spilled;  /**< the value was already spilled on this path */
  76 } loc_t;
  77
  78 typedef struct _workset_t {
  79         int   len;          /**< current length */
  80         loc_t vals[0];      /**< inlined array of the values/distances in this working set */
  81 } workset_t;
  82
  83 static struct obstack               obst;
  84 static const arch_register_class_t *cls;
  85 static const be_lv_t               *lv;
  86 static be_loopana_t                *loop_ana;
  87 static int                          n_regs;
  88 static workset_t                   *ws;     /**< the main workset used while
  89                                                      processing a block. */
  90 static be_uses_t                   *uses;   /**< env for the next-use magic */
  91 static ir_node                     *instr;  /**< current instruction */
  92 static unsigned                     instr_nr; /**< current instruction number
  93                                                        (relative to block start) */
  94 static spill_env_t                 *senv;   /**< see bespill.h */
  95 static ir_node                    **blocklist;
  96
  97 static int                          move_spills      = true;
  98 static int                          respectloopdepth = true;
  99 static int                          improve_known_preds = true;
 100 /* factor to weight the different costs of reloading/rematerializing a node
 101    (see bespill.h be_get_reload_costs_no_weight) */
 102 static int                          remat_bonus      = 10;
 103
 104 static const lc_opt_table_entry_t options[] = {
 105         LC_OPT_ENT_BOOL   ("movespills", "try to move spills out of loops", &move_spills),
 106         LC_OPT_ENT_BOOL   ("respectloopdepth", "exprimental (outermost loop cutting)", &respectloopdepth),
 107         LC_OPT_ENT_BOOL   ("improveknownpreds", "experimental (known preds cutting)", &improve_known_preds),
 108         LC_OPT_ENT_INT    ("rematbonus", "give bonus to rematerialisable nodes", &remat_bonus),
 109         LC_OPT_LAST
 110 };
 111
 112 static int loc_compare(const void *a, const void *b)
 113 {
 114         const loc_t *p = a;
 115         const loc_t *q = b;
 116         return p->time - q->time;
 117 }
 118
 119 /**
 120  * Alloc a new workset on obstack @p ob with maximum size @p max
 121  */
 122 static workset_t *new_workset(void)
 123 {
 124         return OALLOCFZ(&obst, workset_t, vals, n_regs);
 125 }
 126
 127 /**
 128  * Alloc a new instance on obstack and make it equal to @param workset
 129  */
 130 static workset_t *workset_clone(workset_t *workset)
 131 {
 132         workset_t *res = OALLOCF(&obst, workset_t, vals, n_regs);
 133         memcpy(res, workset, sizeof(*res) + n_regs * sizeof(res->vals[0]));
 134         return res;
 135 }
 136
 137 /**
 138  * Copy workset @param src to @param tgt
 139  */
 140 static void workset_copy(workset_t *dest, const workset_t *src)
 141 {
 142         size_t size = sizeof(*src) + n_regs * sizeof(src->vals[0]);
 143         memcpy(dest, src, size);
 144 }
 145
 146 /**
 147  * Overwrites the current content array of @param ws with the
 148  * @param count locations given at memory @param locs.
 149  * Set the length of @param ws to count.
 150  */
 151 static void workset_bulk_fill(workset_t *workset, int count, const loc_t *locs)
 152 {
 153         workset->len = count;
 154         memcpy(&(workset->vals[0]), locs, count * sizeof(locs[0]));
 155 }
 156
 157 /**
 158  * Inserts the value @p val into the workset, iff it is not
 159  * already contained. The workset must not be full.
 160  */
 161 static void workset_insert(workset_t *workset, ir_node *val, bool spilled)
 162 {
 163         loc_t *loc;
 164         int    i;
 165         /* check for current regclass */
 166         assert(arch_irn_consider_in_reg_alloc(cls, val));
 167
 168         /* check if val is already contained */
 169         for (i = 0; i < workset->len; ++i) {
 170                 loc = &workset->vals[i];
 171                 if (loc->node == val) {
 172                         if (spilled) {
 173                                 loc->spilled = true;
 174                         }
 175                         return;
 176                 }
 177         }
 178
 179         /* insert val */
 180         assert(workset->len < n_regs && "Workset already full!");
 181         loc           = &workset->vals[workset->len];
 182         loc->node     = val;
 183         loc->spilled  = spilled;
 184         loc->time     = TIME_UNDEFINED;
 185         workset->len++;
 186 }
 187
 188 /**
 189  * Removes all entries from this workset
 190  */
 191 static void workset_clear(workset_t *workset)
 192 {
 193         workset->len = 0;
 194 }
 195
 196 /**
 197  * Removes the value @p val from the workset if present.
 198  */
 199 static inline void workset_remove(workset_t *workset, ir_node *val)
 200 {
 201         int i;
 202         for (i = 0; i < workset->len; ++i) {
 203                 if (workset->vals[i].node == val) {
 204                         workset->vals[i] = workset->vals[--workset->len];
 205                         return;
 206                 }
 207         }
 208 }
 209
 210 static inline const loc_t *workset_contains(const workset_t *ws,
 211                                             const ir_node *val)
 212 {
 213         int i;
 214
 215         for (i = 0; i < ws->len; ++i) {
 216                 if (ws->vals[i].node == val)
 217                         return &ws->vals[i];
 218         }
 219
 220         return NULL;
 221 }
 222
 223 /**
 224  * Iterates over all values in the working set.
 225  * @p ws The workset to iterate
 226  * @p v  A variable to put the current value in
 227  * @p i  An integer for internal use
 228  */
 229 #define workset_foreach(ws, v, i)       for (i=0; \
 230                                                                                 v=(i < ws->len) ? ws->vals[i].node : NULL, i < ws->len; \
 231                                                                                 ++i)
 232
 233 #define workset_set_time(ws, i, t) (ws)->vals[i].time=t
 234 #define workset_get_time(ws, i) (ws)->vals[i].time
 235 #define workset_set_length(ws, length) (ws)->len = length
 236 #define workset_get_length(ws) ((ws)->len)
 237 #define workset_get_val(ws, i) ((ws)->vals[i].node)
 238 #define workset_sort(ws) do { qsort((ws)->vals, (ws)->len, sizeof((ws)->vals[0]), loc_compare); } while(0)
 239
 240 typedef struct _block_info_t
 241 {
 242         workset_t *start_workset;
 243         workset_t *end_workset;
 244 } block_info_t;
 245
 246
 247 static block_info_t *new_block_info(void)
 248 {
 249         return OALLOCZ(&obst, block_info_t);
 250 }
 251
 252 #define get_block_info(block)        ((block_info_t *)get_irn_link(block))
 253 #define set_block_info(block, info)  set_irn_link(block, info)
 254
 255 /**
 256  * @return The distance to the next use or 0 if irn has dont_spill flag set
 257  */
 258 static inline unsigned get_distance(ir_node *from, unsigned from_step,
 259                                     const ir_node *def, int skip_from_uses)
 260 {
 261         be_next_use_t use;
 262         unsigned      costs;
 263         unsigned      time;
 264
 265         assert(!arch_irn_is_ignore(def));
 266
 267         use  = be_get_next_use(uses, from, from_step, def, skip_from_uses);
 268         time = use.time;
 269         if (USES_IS_INFINITE(time))
 270                 return USES_INFINITY;
 271
 272         /* We have to keep nonspillable nodes in the workingset */
 273         if (arch_irn_get_flags(skip_Proj_const(def)) & arch_irn_flags_dont_spill)
 274                 return 0;
 275
 276         /* give some bonus to rematerialisable nodes */
 277         if (remat_bonus > 0) {
 278                 costs = be_get_reload_costs_no_weight(senv, def, use.before);
 279                 assert(costs * remat_bonus < 1000);
 280                 time  += 1000 - (costs * remat_bonus);
 281         }
 282
 283         return time;
 284 }
 285
 286 /**
 287  * Performs the actions necessary to grant the request that:
 288  * - new_vals can be held in registers
 289  * - as few as possible other values are disposed
 290  * - the worst values get disposed
 291  *
 292  * @p is_usage indicates that the values in new_vals are used (not defined)
 293  * In this case reloads must be performed
 294  */
 295 static void displace(workset_t *new_vals, int is_usage)
 296 {
 297         ir_node **to_insert = ALLOCAN(ir_node*, n_regs);
 298         bool     *spilled   = ALLOCAN(bool,     n_regs);
 299         ir_node  *val;
 300         int       i;
 301         int       len;
 302         int       spills_needed;
 303         int       demand;
 304         int       iter;
 305
 306         /* 1. Identify the number of needed slots and the values to reload */
 307         demand = 0;
 308         workset_foreach(new_vals, val, iter) {
 309                 bool reloaded = false;
 310
 311                 if (! workset_contains(ws, val)) {
 312                         DB((dbg, DBG_DECIDE, "    insert %+F\n", val));
 313                         if (is_usage) {
 314                                 DB((dbg, DBG_SPILL, "Reload %+F before %+F\n", val, instr));
 315                                 be_add_reload(senv, val, instr, cls, 1);
 316                                 reloaded = true;
 317                         }
 318                 } else {
 319                         DB((dbg, DBG_DECIDE, "    %+F already in workset\n", val));
 320                         assert(is_usage);
 321                         /* remove the value from the current workset so it is not accidently
 322                          * spilled */
 323                         workset_remove(ws, val);
 324                 }
 325                 spilled[demand]   = reloaded;
 326                 to_insert[demand] = val;
 327                 ++demand;
 328         }
 329
 330         /* 2. Make room for at least 'demand' slots */
 331         len           = workset_get_length(ws);
 332         spills_needed = len + demand - n_regs;
 333         assert(spills_needed <= len);
 334
 335         /* Only make more free room if we do not have enough */
 336         if (spills_needed > 0) {
 337                 DB((dbg, DBG_DECIDE, "    disposing %d values\n", spills_needed));
 338
 339                 /* calculate current next-use distance for live values */
 340                 for (i = 0; i < len; ++i) {
 341                         ir_node  *val  = workset_get_val(ws, i);
 342                         unsigned  dist = get_distance(instr, instr_nr, val, !is_usage);
 343                         workset_set_time(ws, i, dist);
 344                 }
 345
 346                 /* sort entries by increasing nextuse-distance*/
 347                 workset_sort(ws);
 348
 349                 for (i = len - spills_needed; i < len; ++i) {
 350                         ir_node *val = ws->vals[i].node;
 351
 352                         DB((dbg, DBG_DECIDE, "    disposing node %+F (%u)\n", val,
 353                              workset_get_time(ws, i)));
 354
 355                         if (move_spills) {
 356                                 if (!USES_IS_INFINITE(ws->vals[i].time)
 357                                                 && !ws->vals[i].spilled) {
 358                                         ir_node *after_pos = sched_prev(instr);
 359                                         DB((dbg, DBG_DECIDE, "Spill %+F after node %+F\n", val,
 360                                                 after_pos));
 361                                         be_add_spill(senv, val, after_pos);
 362                                 }
 363                         }
 364                 }
 365
 366                 /* kill the last 'demand' entries in the array */
 367                 workset_set_length(ws, len - spills_needed);
 368         }
 369
 370         /* 3. Insert the new values into the workset */
 371         for (i = 0; i < demand; ++i) {
 372                 ir_node *val = to_insert[i];
 373
 374                 workset_insert(ws, val, spilled[i]);
 375         }
 376 }
 377
 378 enum {
 379         AVAILABLE_EVERYWHERE,
 380         AVAILABLE_NOWHERE,
 381         AVAILABLE_PARTLY,
 382         AVAILABLE_UNKNOWN
 383 };
 384
 385 static unsigned available_in_all_preds(workset_t* const* pred_worksets,
 386                                        size_t n_pred_worksets,
 387                                        const ir_node *value, bool is_local_phi)
 388 {
 389         size_t i;
 390         bool   avail_everywhere = true;
 391         bool   avail_nowhere    = true;
 392
 393         assert(n_pred_worksets > 0);
 394
 395         /* value available in all preds? */
 396         for (i = 0; i < n_pred_worksets; ++i) {
 397                 bool             found     = false;
 398                 const workset_t *p_workset = pred_worksets[i];
 399                 int              p_len     = workset_get_length(p_workset);
 400                 int              p_i;
 401                 const ir_node   *l_value;
 402
 403                 if (is_local_phi) {
 404                         assert(is_Phi(value));
 405                         l_value = get_irn_n(value, i);
 406                 } else {
 407                         l_value = value;
 408                 }
 409
 410                 for (p_i = 0; p_i < p_len; ++p_i) {
 411                         const loc_t *p_l = &p_workset->vals[p_i];
 412                         if (p_l->node != l_value)
 413                                 continue;
 414
 415                         found = true;
 416                         break;
 417                 }
 418
 419                 if (found) {
 420                         avail_nowhere = false;
 421                 } else {
 422                         avail_everywhere = false;
 423                 }
 424         }
 425
 426         if (avail_everywhere) {
 427                 assert(!avail_nowhere);
 428                 return AVAILABLE_EVERYWHERE;
 429         } else if (avail_nowhere) {
 430                 return AVAILABLE_NOWHERE;
 431         } else {
 432                 return AVAILABLE_PARTLY;
 433         }
 434 }
 435
 436 /** Decides whether a specific node should be in the start workset or not
 437  *
 438  * @param env      belady environment
 439  * @param first
 440  * @param node     the node to test
 441  * @param loop     the loop of the node
 442  */
 443 static loc_t to_take_or_not_to_take(ir_node* first, ir_node *node,
 444                                     ir_loop *loop, unsigned available)
 445 {
 446         be_next_use_t next_use;
 447         loc_t         loc;
 448
 449         loc.time    = USES_INFINITY;
 450         loc.node    = node;
 451         loc.spilled = false;
 452
 453         if (!arch_irn_consider_in_reg_alloc(cls, node)) {
 454                 loc.time = USES_INFINITY;
 455                 return loc;
 456         }
 457
 458         /* We have to keep nonspillable nodes in the workingset */
 459         if (arch_irn_get_flags(skip_Proj_const(node)) & arch_irn_flags_dont_spill) {
 460                 loc.time = 0;
 461                 DB((dbg, DBG_START, "    %+F taken (dontspill node)\n", node, loc.time));
 462                 return loc;
 463         }
 464
 465         next_use = be_get_next_use(uses, first, 0, node, 0);
 466         if (USES_IS_INFINITE(next_use.time)) {
 467                 // the nodes marked as live in shouldn't be dead, so it must be a phi
 468                 assert(is_Phi(node));
 469                 loc.time = USES_INFINITY;
 470                 DB((dbg, DBG_START, "    %+F not taken (dead)\n", node));
 471                 return loc;
 472         }
 473
 474         loc.time = next_use.time;
 475
 476         if (improve_known_preds) {
 477                 if (available == AVAILABLE_EVERYWHERE) {
 478                         DB((dbg, DBG_START, "    %+F taken (%u, live in all preds)\n",
 479                             node, loc.time));
 480                         return loc;
 481                 } else if (available == AVAILABLE_NOWHERE) {
 482                         DB((dbg, DBG_START, "    %+F not taken (%u, live in no pred)\n",
 483                             node, loc.time));
 484                         loc.time = USES_INFINITY;
 485                         return loc;
 486                 }
 487         }
 488
 489         if (!respectloopdepth || next_use.outermost_loop >= get_loop_depth(loop)) {
 490                 DB((dbg, DBG_START, "    %+F taken (%u, loop %d)\n", node, loc.time,
 491                     next_use.outermost_loop));
 492         } else {
 493                 loc.time = USES_PENDING;
 494                 DB((dbg, DBG_START, "    %+F delayed (outerdepth %d < loopdepth %d)\n",
 495                     node, next_use.outermost_loop, get_loop_depth(loop)));
 496         }
 497
 498         return loc;
 499 }
 500
 501 /**
 502  * Computes the start-workset for a block with multiple predecessors. We assume
 503  * that at least 1 of the predeccesors is a back-edge which means we're at the
 504  * beginning of a loop. We try to reload as much values as possible now so they
 505  * don't get reloaded inside the loop.
 506  */
 507 static void decide_start_workset(const ir_node *block)
 508 {
 509         ir_loop    *loop = get_irn_loop(block);
 510         ir_node    *first;
 511         ir_node    *node;
 512         loc_t       loc;
 513         loc_t      *starters;
 514         loc_t      *delayed;
 515         int         i, len, ws_count;
 516         int             free_slots, free_pressure_slots;
 517         unsigned    pressure;
 518         int         arity;
 519         workset_t **pred_worksets;
 520         bool        all_preds_known;
 521
 522         /* check predecessors */
 523         arity           = get_irn_arity(block);
 524         pred_worksets   = ALLOCAN(workset_t*, arity);
 525         all_preds_known = true;
 526         for (i = 0; i < arity; ++i) {
 527                 ir_node      *pred_block = get_Block_cfgpred_block(block, i);
 528                 block_info_t *pred_info  = get_block_info(pred_block);
 529
 530                 if (pred_info == NULL) {
 531                         pred_worksets[i] = NULL;
 532                         all_preds_known  = false;
 533                 } else {
 534                         pred_worksets[i] = pred_info->end_workset;
 535                 }
 536         }
 537
 538         /* Collect all values living at start of block */
 539         starters = NEW_ARR_F(loc_t, 0);
 540         delayed  = NEW_ARR_F(loc_t, 0);
 541
 542         DB((dbg, DBG_START, "Living at start of %+F:\n", block));
 543         first = sched_first(block);
 544
 545         /* check all Phis first */
 546         sched_foreach(block, node) {
 547                 unsigned available;
 548
 549                 if (! is_Phi(node))
 550                         break;
 551                 if (!arch_irn_consider_in_reg_alloc(cls, node))
 552                         continue;
 553
 554                 if (all_preds_known) {
 555                         available = available_in_all_preds(pred_worksets, arity, node, true);
 556                 } else {
 557                         available = AVAILABLE_UNKNOWN;
 558                 }
 559
 560                 loc = to_take_or_not_to_take(first, node, loop, available);
 561
 562                 if (! USES_IS_INFINITE(loc.time)) {
 563                         if (USES_IS_PENDING(loc.time))
 564                                 ARR_APP1(loc_t, delayed, loc);
 565                         else
 566                                 ARR_APP1(loc_t, starters, loc);
 567                 } else {
 568                         be_spill_phi(senv, node);
 569                 }
 570         }
 571
 572         /* check all Live-Ins */
 573         be_lv_foreach(lv, block, be_lv_state_in, i) {
 574                 ir_node *node = be_lv_get_irn(lv, block, i);
 575                 unsigned available;
 576
 577                 if (all_preds_known) {
 578                         available = available_in_all_preds(pred_worksets, arity, node, false);
 579                 } else {
 580                         available = AVAILABLE_UNKNOWN;
 581                 }
 582
 583                 loc = to_take_or_not_to_take(first, node, loop, available);
 584
 585                 if (! USES_IS_INFINITE(loc.time)) {
 586                         if (USES_IS_PENDING(loc.time))
 587                                 ARR_APP1(loc_t, delayed, loc);
 588                         else
 589                                 ARR_APP1(loc_t, starters, loc);
 590                 }
 591         }
 592
 593         pressure            = be_get_loop_pressure(loop_ana, cls, loop);
 594         assert(ARR_LEN(delayed) <= (signed)pressure);
 595         free_slots          = n_regs - ARR_LEN(starters);
 596         free_pressure_slots = n_regs - (pressure - ARR_LEN(delayed));
 597         free_slots          = MIN(free_slots, free_pressure_slots);
 598
 599         /* so far we only put nodes into the starters list that are used inside
 600          * the loop. If register pressure in the loop is low then we can take some
 601          * values and let them live through the loop */
 602         DB((dbg, DBG_START, "Loop pressure %d, taking %d delayed vals\n",
 603             pressure, free_slots));
 604         if (free_slots > 0) {
 605                 qsort(delayed, ARR_LEN(delayed), sizeof(delayed[0]), loc_compare);
 606
 607                 for (i = 0; i < ARR_LEN(delayed) && free_slots > 0; ++i) {
 608                         int    p, arity;
 609                         loc_t *loc = & delayed[i];
 610
 611                         if (!is_Phi(loc->node)) {
 612                                 /* don't use values which are dead in a known predecessors
 613                                  * to not induce unnecessary reloads */
 614                                 arity = get_irn_arity(block);
 615                                 for (p = 0; p < arity; ++p) {
 616                                         ir_node      *pred_block = get_Block_cfgpred_block(block, p);
 617                                         block_info_t *pred_info  = get_block_info(pred_block);
 618
 619                                         if (pred_info == NULL)
 620                                                 continue;
 621
 622                                         if (!workset_contains(pred_info->end_workset, loc->node)) {
 623                                                 DB((dbg, DBG_START,
 624                                                         "    delayed %+F not live at pred %+F\n", loc->node,
 625                                                         pred_block));
 626                                                 goto skip_delayed;
 627                                         }
 628                                 }
 629                         }
 630
 631                         DB((dbg, DBG_START, "    delayed %+F taken\n", loc->node));
 632                         ARR_APP1(loc_t, starters, *loc);
 633                         loc->node = NULL;
 634                         --free_slots;
 635                 skip_delayed:
 636                         ;
 637                 }
 638         }
 639
 640         /* spill phis (the actual phis not just their values) that are in this block
 641          * but not in the start workset */
 642         for (i = ARR_LEN(delayed) - 1; i >= 0; --i) {
 643                 ir_node *node = delayed[i].node;
 644                 if (node == NULL || !is_Phi(node) || get_nodes_block(node) != block)
 645                         continue;
 646
 647                 DB((dbg, DBG_START, "    spilling delayed phi %+F\n", node));
 648                 be_spill_phi(senv, node);
 649         }
 650         DEL_ARR_F(delayed);
 651
 652         /* Sort start values by first use */
 653         qsort(starters, ARR_LEN(starters), sizeof(starters[0]), loc_compare);
 654
 655         /* Copy the best ones from starters to start workset */
 656         ws_count = MIN(ARR_LEN(starters), n_regs);
 657         workset_clear(ws);
 658         workset_bulk_fill(ws, ws_count, starters);
 659
 660         /* spill phis (the actual phis not just their values) that are in this block
 661          * but not in the start workset */
 662         len = ARR_LEN(starters);
 663         for (i = ws_count; i < len; ++i) {
 664                 ir_node *node = starters[i].node;
 665                 if (! is_Phi(node) || get_nodes_block(node) != block)
 666                         continue;
 667
 668                 DB((dbg, DBG_START, "    spilling phi %+F\n", node));
 669                 be_spill_phi(senv, node);
 670         }
 671
 672         DEL_ARR_F(starters);
 673
 674         /* determine spill status of the values: If there's 1 pred block (which
 675          * is no backedge) where the value is spilled then we must set it to
 676          * spilled here. */
 677         for (i = 0; i < ws_count; ++i) {
 678                 loc_t   *loc     = &ws->vals[i];
 679                 ir_node *value   = loc->node;
 680                 bool     spilled;
 681                 int      n;
 682
 683                 /* phis from this block aren't spilled */
 684                 if (get_nodes_block(value) == block) {
 685                         assert(is_Phi(value));
 686                         loc->spilled = false;
 687                         continue;
 688                 }
 689
 690                 /* determine if value was spilled on any predecessor */
 691                 spilled = false;
 692                 for (n = 0; n < arity; ++n) {
 693                         workset_t *pred_workset = pred_worksets[n];
 694                         int        p_len;
 695                         int        p;
 696
 697                         if (pred_workset == NULL)
 698                                 continue;
 699
 700                         p_len = workset_get_length(pred_workset);
 701                         for (p = 0; p < p_len; ++p) {
 702                                 loc_t *l = &pred_workset->vals[p];
 703
 704                                 if (l->node != value)
 705                                         continue;
 706
 707                                 if (l->spilled) {
 708                                         spilled = true;
 709                                 }
 710                                 break;
 711                         }
 712                 }
 713
 714                 loc->spilled = spilled;
 715         }
 716 }
 717
 718 /**
 719  * For the given block @p block, decide for each values
 720  * whether it is used from a register or is reloaded
 721  * before the use.
 722  */
 723 static void process_block(ir_node *block)
 724 {
 725         workset_t       *new_vals;
 726         ir_node         *irn;
 727         int              iter;
 728         block_info_t    *block_info;
 729         int              arity;
 730
 731         /* no need to process a block twice */
 732         assert(get_block_info(block) == NULL);
 733
 734         /* construct start workset */
 735         arity = get_Block_n_cfgpreds(block);
 736         if (arity == 0) {
 737                 /* no predecessor -> empty set */
 738                 workset_clear(ws);
 739         } else if (arity == 1) {
 740                 /* one predecessor, copy it's end workset */
 741                 ir_node      *pred_block = get_Block_cfgpred_block(block, 0);
 742                 block_info_t *pred_info  = get_block_info(pred_block);
 743
 744                 assert(pred_info != NULL);
 745                 workset_copy(ws, pred_info->end_workset);
 746         } else {
 747                 /* multiple predecessors, do more advanced magic :) */
 748                 decide_start_workset(block);
 749         }
 750
 751         DB((dbg, DBG_DECIDE, "\n"));
 752         DB((dbg, DBG_DECIDE, "Decide for %+F\n", block));
 753
 754         block_info = new_block_info();
 755         set_block_info(block, block_info);
 756
 757         DB((dbg, DBG_WSETS, "Start workset for %+F:\n", block));
 758         workset_foreach(ws, irn, iter) {
 759                 DB((dbg, DBG_WSETS, "  %+F (%u)\n", irn,
 760                      workset_get_time(ws, iter)));
 761         }
 762
 763         block_info->start_workset = workset_clone(ws);
 764
 765         /* process the block from start to end */
 766         DB((dbg, DBG_WSETS, "Processing...\n"));
 767         instr_nr = 0;
 768         /* TODO: this leaks (into the obstack)... */
 769         new_vals = new_workset();
 770
 771         sched_foreach(block, irn) {
 772                 int i, arity;
 773                 ir_node *value;
 774                 assert(workset_get_length(ws) <= n_regs);
 775
 776                 /* Phis are no real instr (see insert_starters()) */
 777                 if (is_Phi(irn)) {
 778                         continue;
 779                 }
 780                 DB((dbg, DBG_DECIDE, "  ...%+F\n", irn));
 781
 782                 /* set instruction in the workset */
 783                 instr = irn;
 784
 785                 /* allocate all values _used_ by this instruction */
 786                 workset_clear(new_vals);
 787                 for (i = 0, arity = get_irn_arity(irn); i < arity; ++i) {
 788                         ir_node *in = get_irn_n(irn, i);
 789                         if (!arch_irn_consider_in_reg_alloc(cls, in))
 790                                 continue;
 791
 792                         /* (note that "spilled" is irrelevant here) */
 793                         workset_insert(new_vals, in, false);
 794                 }
 795                 displace(new_vals, 1);
 796
 797                 /* allocate all values _defined_ by this instruction */
 798                 workset_clear(new_vals);
 799                 be_foreach_definition(irn, cls, value,
 800                         workset_insert(new_vals, value, false);
 801                 );
 802                 displace(new_vals, 0);
 803
 804                 instr_nr++;
 805         }
 806
 807         /* Remember end-workset for this block */
 808         block_info->end_workset = workset_clone(ws);
 809         DB((dbg, DBG_WSETS, "End workset for %+F:\n", block));
 810         workset_foreach(ws, irn, iter)
 811                 DB((dbg, DBG_WSETS, "  %+F (%u)\n", irn,
 812                      workset_get_time(ws, iter)));
 813 }
 814
 815 /**
 816  * 'decide' is block-local and makes assumptions
 817  * about the set of live-ins. Thus we must adapt the
 818  * live-outs to the live-ins at each block-border.
 819  */
 820 static void fix_block_borders(ir_node *block, void *data)
 821 {
 822         workset_t    *start_workset;
 823         int           arity;
 824         int           i;
 825         int           iter;
 826         (void) data;
 827
 828         DB((dbg, DBG_FIX, "\n"));
 829         DB((dbg, DBG_FIX, "Fixing %+F\n", block));
 830
 831         arity = get_irn_arity(block);
 832         /* can happen for endless loops */
 833         if (arity == 0)
 834                 return;
 835
 836         start_workset = get_block_info(block)->start_workset;
 837
 838         /* process all pred blocks */
 839         for (i = 0; i < arity; ++i) {
 840                 ir_node   *pred = get_Block_cfgpred_block(block, i);
 841                 workset_t *pred_end_workset = get_block_info(pred)->end_workset;
 842                 ir_node   *node;
 843
 844                 DB((dbg, DBG_FIX, "  Pred %+F\n", pred));
 845
 846                 /* spill all values not used anymore */
 847                 workset_foreach(pred_end_workset, node, iter) {
 848                         ir_node *n2;
 849                         int      iter2;
 850                         bool     found = false;
 851                         workset_foreach(start_workset, n2, iter2) {
 852                                 if (n2 == node) {
 853                                         found = true;
 854                                         break;
 855                                 }
 856                                 /* note that we do not look at phi inputs, becuase the values
 857                                  * will be either live-end and need no spill or
 858                                  * they have other users in which must be somewhere else in the
 859                                  * workset */
 860                         }
 861
 862                         if (found)
 863                                 continue;
 864
 865                         if (move_spills && be_is_live_in(lv, block, node)
 866                                         && !pred_end_workset->vals[iter].spilled) {
 867                                 ir_node *insert_point;
 868                                 if (arity > 1) {
 869                                         insert_point = be_get_end_of_block_insertion_point(pred);
 870                                         insert_point = sched_prev(insert_point);
 871                                 } else {
 872                                         insert_point = block;
 873                                 }
 874                                 DB((dbg, DBG_SPILL, "Spill %+F after %+F\n", node,
 875                                      insert_point));
 876                                 be_add_spill(senv, node, insert_point);
 877                         }
 878                 }
 879
 880                 /* reload missing values in predecessors, add missing spills */
 881                 workset_foreach(start_workset, node, iter) {
 882                         const loc_t *l    = &start_workset->vals[iter];
 883                         const loc_t *pred_loc;
 884
 885                         /* if node is a phi of the current block we reload
 886                          * the corresponding argument, else node itself */
 887                         if (is_Phi(node) && get_nodes_block(node) == block) {
 888                                 node = get_irn_n(node, i);
 889                                 assert(!l->spilled);
 890
 891                                 /* we might have unknowns as argument for the phi */
 892                                 if (!arch_irn_consider_in_reg_alloc(cls, node))
 893                                         continue;
 894                         }
 895
 896                         /* check if node is in a register at end of pred */
 897                         pred_loc = workset_contains(pred_end_workset, node);
 898                         if (pred_loc != NULL) {
 899                                 /* we might have to spill value on this path */
 900                                 if (move_spills && !pred_loc->spilled && l->spilled) {
 901                                         ir_node *insert_point
 902                                                 = be_get_end_of_block_insertion_point(pred);
 903                                         insert_point = sched_prev(insert_point);
 904                                         DB((dbg, DBG_SPILL, "Spill %+F after %+F\n", node,
 905                                             insert_point));
 906                                         be_add_spill(senv, node, insert_point);
 907                                 }
 908                         } else {
 909                                 /* node is not in register at the end of pred -> reload it */
 910                                 DB((dbg, DBG_FIX, "    reload %+F\n", node));
 911                                 DB((dbg, DBG_SPILL, "Reload %+F before %+F,%d\n", node, block, i));
 912                                 be_add_reload_on_edge(senv, node, block, i, cls, 1);
 913                         }
 914                 }
 915         }
 916 }
 917
 918 static void be_spill_belady(ir_graph *irg, const arch_register_class_t *rcls)
 919 {
 920         int i;
 921
 922         be_liveness_assure_sets(be_assure_liveness(irg));
 923
 924         stat_ev_tim_push();
 925         /* construct control flow loop tree */
 926         if (! (get_irg_loopinfo_state(irg) & loopinfo_cf_consistent)) {
 927                 construct_cf_backedges(irg);
 928         }
 929         stat_ev_tim_pop("belady_time_backedges");
 930
 931         stat_ev_tim_push();
 932         be_clear_links(irg);
 933         stat_ev_tim_pop("belady_time_clear_links");
 934
 935         ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
 936
 937         /* init belady env */
 938         stat_ev_tim_push();
 939         obstack_init(&obst);
 940         cls       = rcls;
 941         lv        = be_get_irg_liveness(irg);
 942         n_regs    = cls->n_regs - be_put_ignore_regs(irg, cls, NULL);
 943         ws        = new_workset();
 944         uses      = be_begin_uses(irg, lv);
 945         loop_ana  = be_new_loop_pressure(irg, cls);
 946         senv      = be_new_spill_env(irg);
 947         blocklist = be_get_cfgpostorder(irg);
 948         stat_ev_tim_pop("belady_time_init");
 949
 950         stat_ev_tim_push();
 951         /* walk blocks in reverse postorder */
 952         for (i = ARR_LEN(blocklist) - 1; i >= 0; --i) {
 953                 process_block(blocklist[i]);
 954         }
 955         DEL_ARR_F(blocklist);
 956         stat_ev_tim_pop("belady_time_belady");
 957
 958         stat_ev_tim_push();
 959         /* belady was block-local, fix the global flow by adding reloads on the
 960          * edges */
 961         irg_block_walk_graph(irg, fix_block_borders, NULL, NULL);
 962         stat_ev_tim_pop("belady_time_fix_borders");
 963
 964         ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
 965
 966         /* Insert spill/reload nodes into the graph and fix usages */
 967         be_insert_spills_reloads(senv);
 968
 969         /* clean up */
 970         be_delete_spill_env(senv);
 971         be_end_uses(uses);
 972         be_free_loop_pressure(loop_ana);
 973         obstack_free(&obst, NULL);
 974 }
 975
 976 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady);
 977 void be_init_spillbelady(void)
 978 {
 979         static be_spiller_t belady_spiller = {
 980                 be_spill_belady
 981         };
 982         lc_opt_entry_t *be_grp       = lc_opt_get_grp(firm_opt_get_root(), "be");
 983         lc_opt_entry_t *belady_group = lc_opt_get_grp(be_grp, "belady");
 984         lc_opt_add_table(belady_group, options);
 985
 986         be_register_spiller("belady", &belady_spiller);
 987         FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
 988 }