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_t.h"
  50 #include "bechordal_t.h"
  51 #include "bespill.h"
  52 #include "beloopana.h"
  53 #include "beirg.h"
  54 #include "bespillutil.h"
  55 #include "bemodule.h"
  56
  57 #define DBG_SPILL     1
  58 #define DBG_WSETS     2
  59 #define DBG_FIX       4
  60 #define DBG_DECIDE    8
  61 #define DBG_START    16
  62 #define DBG_SLOTS    32
  63 #define DBG_TRACE    64
  64 #define DBG_WORKSET 128
  65 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  66
  67 #define TIME_UNDEFINED 6666
  68
  69 //#define LOOK_AT_LOOPDEPTH
  70
  71 /**
  72  * An association between a node and a point in time.
  73  */
  74 typedef struct loc_t {
  75         ir_node          *node;
  76         unsigned          time;     /**< A use time (see beuses.h). */
  77         bool              spilled;  /**< the value was already spilled on this path */
  78 } loc_t;
  79
  80 typedef struct _workset_t {
  81         int   len;          /**< current length */
  82         loc_t vals[0];      /**< inlined array of the values/distances in this working set */
  83 } workset_t;
  84
  85 static struct obstack               obst;
  86 static const arch_register_class_t *cls;
  87 static const be_lv_t               *lv;
  88 static be_loopana_t                *loop_ana;
  89 static int                          n_regs;
  90 static workset_t                   *ws;     /**< the main workset used while
  91                                                      processing a block. */
  92 static be_uses_t                   *uses;   /**< env for the next-use magic */
  93 static ir_node                     *instr;  /**< current instruction */
  94 static unsigned                     instr_nr; /**< current instruction number
  95                                                        (relative to block start) */
  96 static spill_env_t                 *senv;   /**< see bespill.h */
  97 static ir_node                    **blocklist;
  98
  99 static 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         unsigned      costs;
 284         unsigned      time;
 285
 286         assert(!arch_irn_is_ignore(def));
 287
 288         use  = be_get_next_use(uses, from, from_step, def, skip_from_uses);
 289         time = use.time;
 290         if (USES_IS_INFINITE(time))
 291                 return USES_INFINITY;
 292
 293         /* We have to keep nonspillable nodes in the workingset */
 294         if (arch_irn_get_flags(def) & arch_irn_flags_dont_spill)
 295                 return 0;
 296
 297         /* give some bonus to rematerialisable nodes */
 298         if (remat_bonus > 0) {
 299                 costs = be_get_reload_costs_no_weight(senv, def, use.before);
 300                 assert(costs * remat_bonus < 1000);
 301                 time  += 1000 - (costs * remat_bonus);
 302         }
 303
 304         return time;
 305 }
 306
 307 /**
 308  * Performs the actions necessary to grant the request that:
 309  * - new_vals can be held in registers
 310  * - as few as possible other values are disposed
 311  * - the worst values get disposed
 312  *
 313  * @p is_usage indicates that the values in new_vals are used (not defined)
 314  * In this case reloads must be performed
 315  */
 316 static void displace(workset_t *new_vals, int is_usage)
 317 {
 318         ir_node **to_insert = ALLOCAN(ir_node*, n_regs);
 319         bool     *spilled   = ALLOCAN(bool,     n_regs);
 320         ir_node  *val;
 321         int       i;
 322         int       len;
 323         int       spills_needed;
 324         int       demand;
 325         int       iter;
 326
 327         /* 1. Identify the number of needed slots and the values to reload */
 328         demand = 0;
 329         workset_foreach(new_vals, val, iter) {
 330                 bool reloaded = false;
 331
 332                 if (! workset_contains(ws, val)) {
 333                         DB((dbg, DBG_DECIDE, "    insert %+F\n", val));
 334                         if (is_usage) {
 335                                 DB((dbg, DBG_SPILL, "Reload %+F before %+F\n", val, instr));
 336                                 be_add_reload(senv, val, instr, cls, 1);
 337                                 reloaded = true;
 338                         }
 339                 } else {
 340                         DB((dbg, DBG_DECIDE, "    %+F already in workset\n", val));
 341                         assert(is_usage);
 342                         /* remove the value from the current workset so it is not accidently
 343                          * spilled */
 344                         workset_remove(ws, val);
 345                 }
 346                 spilled[demand]   = reloaded;
 347                 to_insert[demand] = val;
 348                 ++demand;
 349         }
 350
 351         /* 2. Make room for at least 'demand' slots */
 352         len           = workset_get_length(ws);
 353         spills_needed = len + demand - n_regs;
 354         assert(spills_needed <= len);
 355
 356         /* Only make more free room if we do not have enough */
 357         if (spills_needed > 0) {
 358                 ir_node   *curr_bb  = NULL;
 359                 workset_t *ws_start = NULL;
 360
 361                 if (move_spills) {
 362                         curr_bb  = get_nodes_block(instr);
 363                         ws_start = get_block_info(curr_bb)->start_workset;
 364                 }
 365
 366                 DB((dbg, DBG_DECIDE, "    disposing %d values\n", spills_needed));
 367
 368                 /* calculate current next-use distance for live values */
 369                 for (i = 0; i < len; ++i) {
 370                         ir_node  *val  = workset_get_val(ws, i);
 371                         unsigned  dist = get_distance(instr, instr_nr, val, !is_usage);
 372                         workset_set_time(ws, i, dist);
 373                 }
 374
 375                 /* sort entries by increasing nextuse-distance*/
 376                 workset_sort(ws);
 377
 378                 for (i = len - spills_needed; i < len; ++i) {
 379                         ir_node *val = ws->vals[i].node;
 380
 381                         DB((dbg, DBG_DECIDE, "    disposing node %+F (%u)\n", val,
 382                              workset_get_time(ws, i)));
 383
 384                         if (move_spills) {
 385                                 if (!USES_IS_INFINITE(ws->vals[i].time)
 386                                                 && !ws->vals[i].spilled) {
 387                                         ir_node *after_pos = sched_prev(instr);
 388                                         DB((dbg, DBG_DECIDE, "Spill %+F after node %+F\n", val,
 389                                                 after_pos));
 390                                         be_add_spill(senv, val, after_pos);
 391                                 }
 392                         }
 393                 }
 394
 395                 /* kill the last 'demand' entries in the array */
 396                 workset_set_length(ws, len - spills_needed);
 397         }
 398
 399         /* 3. Insert the new values into the workset */
 400         for (i = 0; i < demand; ++i) {
 401                 ir_node *val = to_insert[i];
 402
 403                 workset_insert(ws, val, spilled[i]);
 404         }
 405 }
 406
 407 enum {
 408         AVAILABLE_EVERYWHERE,
 409         AVAILABLE_NOWHERE,
 410         AVAILABLE_PARTLY,
 411         AVAILABLE_UNKNOWN
 412 };
 413
 414 static unsigned available_in_all_preds(workset_t* const* pred_worksets,
 415                                        size_t n_pred_worksets,
 416                                        const ir_node *value, bool is_local_phi)
 417 {
 418         size_t i;
 419         bool   avail_everywhere = true;
 420         bool   avail_nowhere    = true;
 421
 422         assert(n_pred_worksets > 0);
 423
 424         /* value available in all preds? */
 425         for (i = 0; i < n_pred_worksets; ++i) {
 426                 bool             found     = false;
 427                 const workset_t *p_workset = pred_worksets[i];
 428                 int              p_len     = workset_get_length(p_workset);
 429                 int              p_i;
 430                 const ir_node   *l_value;
 431
 432                 if (is_local_phi) {
 433                         assert(is_Phi(value));
 434                         l_value = get_irn_n(value, i);
 435                 } else {
 436                         l_value = value;
 437                 }
 438
 439                 for (p_i = 0; p_i < p_len; ++p_i) {
 440                         const loc_t *p_l = &p_workset->vals[p_i];
 441                         if (p_l->node != l_value)
 442                                 continue;
 443
 444                         found = true;
 445                         break;
 446                 }
 447
 448                 if (found) {
 449                         avail_nowhere = false;
 450                 } else {
 451                         avail_everywhere = false;
 452                 }
 453         }
 454
 455         if (avail_everywhere) {
 456                 assert(!avail_nowhere);
 457                 return AVAILABLE_EVERYWHERE;
 458         } else if (avail_nowhere) {
 459                 return AVAILABLE_NOWHERE;
 460         } else {
 461                 return AVAILABLE_PARTLY;
 462         }
 463 }
 464
 465 /** Decides whether a specific node should be in the start workset or not
 466  *
 467  * @param env      belady environment
 468  * @param first
 469  * @param node     the node to test
 470  * @param loop     the loop of the node
 471  */
 472 static loc_t to_take_or_not_to_take(ir_node* first, ir_node *node,
 473                                     ir_loop *loop, unsigned available)
 474 {
 475         be_next_use_t next_use;
 476         loc_t         loc;
 477
 478         loc.time    = USES_INFINITY;
 479         loc.node    = node;
 480         loc.spilled = false;
 481
 482         if (!arch_irn_consider_in_reg_alloc(cls, node)) {
 483                 loc.time = USES_INFINITY;
 484                 return loc;
 485         }
 486
 487         /* We have to keep nonspillable nodes in the workingset */
 488         if (arch_irn_get_flags(node) & arch_irn_flags_dont_spill) {
 489                 loc.time = 0;
 490                 DB((dbg, DBG_START, "    %+F taken (dontspill node)\n", node, loc.time));
 491                 return loc;
 492         }
 493
 494         next_use = be_get_next_use(uses, first, 0, node, 0);
 495         if (USES_IS_INFINITE(next_use.time)) {
 496                 // the nodes marked as live in shouldn't be dead, so it must be a phi
 497                 assert(is_Phi(node));
 498                 loc.time = USES_INFINITY;
 499                 DB((dbg, DBG_START, "    %+F not taken (dead)\n", node));
 500                 return loc;
 501         }
 502
 503         loc.time = next_use.time;
 504
 505         if (improve_known_preds) {
 506                 if (available == AVAILABLE_EVERYWHERE) {
 507                         DB((dbg, DBG_START, "    %+F taken (%u, live in all preds)\n",
 508                             node, loc.time));
 509                         return loc;
 510                 } else if(available == AVAILABLE_NOWHERE) {
 511                         DB((dbg, DBG_START, "    %+F not taken (%u, live in no pred)\n",
 512                             node, loc.time));
 513                         loc.time = USES_INFINITY;
 514                         return loc;
 515                 }
 516         }
 517
 518         if (!respectloopdepth || next_use.outermost_loop >= get_loop_depth(loop)) {
 519                 DB((dbg, DBG_START, "    %+F taken (%u, loop %d)\n", node, loc.time,
 520                     next_use.outermost_loop));
 521         } else {
 522                 loc.time = USES_PENDING;
 523                 DB((dbg, DBG_START, "    %+F delayed (outerdepth %d < loopdepth %d)\n",
 524                     node, next_use.outermost_loop, get_loop_depth(loop)));
 525         }
 526
 527         return loc;
 528 }
 529
 530 /**
 531  * Computes the start-workset for a block with multiple predecessors. We assume
 532  * that at least 1 of the predeccesors is a back-edge which means we're at the
 533  * beginning of a loop. We try to reload as much values as possible now so they
 534  * don't get reloaded inside the loop.
 535  */
 536 static void decide_start_workset(const ir_node *block)
 537 {
 538         ir_loop    *loop = get_irn_loop(block);
 539         ir_node    *first;
 540         ir_node    *node;
 541         loc_t       loc;
 542         loc_t      *starters;
 543         loc_t      *delayed;
 544         int         i, len, ws_count;
 545         int             free_slots, free_pressure_slots;
 546         unsigned    pressure;
 547         int         arity;
 548         workset_t **pred_worksets;
 549         bool        all_preds_known;
 550
 551         /* check predecessors */
 552         arity           = get_irn_arity(block);
 553         pred_worksets   = ALLOCAN(workset_t*, arity);
 554         all_preds_known = true;
 555         for(i = 0; i < arity; ++i) {
 556                 ir_node      *pred_block = get_Block_cfgpred_block(block, i);
 557                 block_info_t *pred_info  = get_block_info(pred_block);
 558
 559                 if (pred_info == NULL) {
 560                         pred_worksets[i] = NULL;
 561                         all_preds_known  = false;
 562                 } else {
 563                         pred_worksets[i] = pred_info->end_workset;
 564                 }
 565         }
 566
 567         /* Collect all values living at start of block */
 568         starters = NEW_ARR_F(loc_t, 0);
 569         delayed  = NEW_ARR_F(loc_t, 0);
 570
 571         DB((dbg, DBG_START, "Living at start of %+F:\n", block));
 572         first = sched_first(block);
 573
 574         /* check all Phis first */
 575         sched_foreach(block, node) {
 576                 unsigned available;
 577
 578                 if (! is_Phi(node))
 579                         break;
 580                 if (!arch_irn_consider_in_reg_alloc(cls, node))
 581                         continue;
 582
 583                 if (all_preds_known) {
 584                         available = available_in_all_preds(pred_worksets, arity, node, true);
 585                 } else {
 586                         available = AVAILABLE_UNKNOWN;
 587                 }
 588
 589                 loc = to_take_or_not_to_take(first, node, loop, available);
 590
 591                 if (! USES_IS_INFINITE(loc.time)) {
 592                         if (USES_IS_PENDING(loc.time))
 593                                 ARR_APP1(loc_t, delayed, loc);
 594                         else
 595                                 ARR_APP1(loc_t, starters, loc);
 596                 } else {
 597                         be_spill_phi(senv, node);
 598                 }
 599         }
 600
 601         /* check all Live-Ins */
 602         be_lv_foreach(lv, block, be_lv_state_in, i) {
 603                 ir_node *node = be_lv_get_irn(lv, block, i);
 604                 unsigned available;
 605
 606                 if (all_preds_known) {
 607                         available = available_in_all_preds(pred_worksets, arity, node, false);
 608                 } else {
 609                         available = AVAILABLE_UNKNOWN;
 610                 }
 611
 612                 loc = to_take_or_not_to_take(first, node, loop, available);
 613
 614                 if (! USES_IS_INFINITE(loc.time)) {
 615                         if (USES_IS_PENDING(loc.time))
 616                                 ARR_APP1(loc_t, delayed, loc);
 617                         else
 618                                 ARR_APP1(loc_t, starters, loc);
 619                 }
 620         }
 621
 622         pressure            = be_get_loop_pressure(loop_ana, cls, loop);
 623         assert(ARR_LEN(delayed) <= (signed)pressure);
 624         free_slots          = n_regs - ARR_LEN(starters);
 625         free_pressure_slots = n_regs - (pressure - ARR_LEN(delayed));
 626         free_slots          = MIN(free_slots, free_pressure_slots);
 627
 628         /* so far we only put nodes into the starters list that are used inside
 629          * the loop. If register pressure in the loop is low then we can take some
 630          * values and let them live through the loop */
 631         DB((dbg, DBG_START, "Loop pressure %d, taking %d delayed vals\n",
 632             pressure, free_slots));
 633         if (free_slots > 0) {
 634                 qsort(delayed, ARR_LEN(delayed), sizeof(delayed[0]), loc_compare);
 635
 636                 for (i = 0; i < ARR_LEN(delayed) && free_slots > 0; ++i) {
 637                         int    p, arity;
 638                         loc_t *loc = & delayed[i];
 639
 640                         if (!is_Phi(loc->node)) {
 641                                 /* don't use values which are dead in a known predecessors
 642                                  * to not induce unnecessary reloads */
 643                                 arity = get_irn_arity(block);
 644                                 for (p = 0; p < arity; ++p) {
 645                                         ir_node      *pred_block = get_Block_cfgpred_block(block, p);
 646                                         block_info_t *pred_info  = get_block_info(pred_block);
 647
 648                                         if (pred_info == NULL)
 649                                                 continue;
 650
 651                                         if (!workset_contains(pred_info->end_workset, loc->node)) {
 652                                                 DB((dbg, DBG_START,
 653                                                         "    delayed %+F not live at pred %+F\n", loc->node,
 654                                                         pred_block));
 655                                                 goto skip_delayed;
 656                                         }
 657                                 }
 658                         }
 659
 660                         DB((dbg, DBG_START, "    delayed %+F taken\n", loc->node));
 661                         ARR_APP1(loc_t, starters, *loc);
 662                         loc->node = NULL;
 663                         --free_slots;
 664                 skip_delayed:
 665                         ;
 666                 }
 667         }
 668
 669         /* spill phis (the actual phis not just their values) that are in this block
 670          * but not in the start workset */
 671         for (i = ARR_LEN(delayed) - 1; i >= 0; --i) {
 672                 ir_node *node = delayed[i].node;
 673                 if (node == NULL || !is_Phi(node) || get_nodes_block(node) != block)
 674                         continue;
 675
 676                 DB((dbg, DBG_START, "    spilling delayed phi %+F\n", node));
 677                 be_spill_phi(senv, node);
 678         }
 679         DEL_ARR_F(delayed);
 680
 681         /* Sort start values by first use */
 682         qsort(starters, ARR_LEN(starters), sizeof(starters[0]), loc_compare);
 683
 684         /* Copy the best ones from starters to start workset */
 685         ws_count = MIN(ARR_LEN(starters), n_regs);
 686         workset_clear(ws);
 687         workset_bulk_fill(ws, ws_count, starters);
 688
 689         /* spill phis (the actual phis not just their values) that are in this block
 690          * but not in the start workset */
 691         len = ARR_LEN(starters);
 692         for (i = ws_count; i < len; ++i) {
 693                 ir_node *node = starters[i].node;
 694                 if (! is_Phi(node) || get_nodes_block(node) != block)
 695                         continue;
 696
 697                 DB((dbg, DBG_START, "    spilling phi %+F\n", node));
 698                 be_spill_phi(senv, node);
 699         }
 700
 701         DEL_ARR_F(starters);
 702
 703         /* determine spill status of the values: If there's 1 pred block (which
 704          * is no backedge) where the value is spilled then we must set it to
 705          * spilled here. */
 706         for(i = 0; i < ws_count; ++i) {
 707                 loc_t   *loc     = &ws->vals[i];
 708                 ir_node *value   = loc->node;
 709                 bool     spilled;
 710                 int      n;
 711
 712                 /* phis from this block aren't spilled */
 713                 if (get_nodes_block(value) == block) {
 714                         assert(is_Phi(value));
 715                         loc->spilled = false;
 716                         continue;
 717                 }
 718
 719                 /* determine if value was spilled on any predecessor */
 720                 spilled = false;
 721                 for(n = 0; n < arity; ++n) {
 722                         workset_t *pred_workset = pred_worksets[n];
 723                         int        p_len;
 724                         int        p;
 725
 726                         if (pred_workset == NULL)
 727                                 continue;
 728
 729                         p_len = workset_get_length(pred_workset);
 730                         for(p = 0; p < p_len; ++p) {
 731                                 loc_t *l = &pred_workset->vals[p];
 732
 733                                 if (l->node != value)
 734                                         continue;
 735
 736                                 if (l->spilled) {
 737                                         spilled = true;
 738                                 }
 739                                 break;
 740                         }
 741                 }
 742
 743                 loc->spilled = spilled;
 744         }
 745 }
 746
 747 /**
 748  * For the given block @p block, decide for each values
 749  * whether it is used from a register or is reloaded
 750  * before the use.
 751  */
 752 static void process_block(ir_node *block)
 753 {
 754         workset_t       *new_vals;
 755         ir_node         *irn;
 756         int              iter;
 757         block_info_t    *block_info;
 758         int              arity;
 759
 760         /* no need to process a block twice */
 761         assert(get_block_info(block) == NULL);
 762
 763         /* construct start workset */
 764         arity = get_Block_n_cfgpreds(block);
 765         if (arity == 0) {
 766                 /* no predecessor -> empty set */
 767                 workset_clear(ws);
 768         } else if (arity == 1) {
 769                 /* one predecessor, copy it's end workset */
 770                 ir_node      *pred_block = get_Block_cfgpred_block(block, 0);
 771                 block_info_t *pred_info  = get_block_info(pred_block);
 772
 773                 assert(pred_info != NULL);
 774                 workset_copy(ws, pred_info->end_workset);
 775         } else {
 776                 /* multiple predecessors, do more advanced magic :) */
 777                 decide_start_workset(block);
 778         }
 779
 780         DB((dbg, DBG_DECIDE, "\n"));
 781         DB((dbg, DBG_DECIDE, "Decide for %+F\n", block));
 782
 783         block_info = new_block_info();
 784         set_block_info(block, block_info);
 785
 786         DB((dbg, DBG_WSETS, "Start workset for %+F:\n", block));
 787         workset_foreach(ws, irn, iter) {
 788                 DB((dbg, DBG_WSETS, "  %+F (%u)\n", irn,
 789                      workset_get_time(ws, iter)));
 790         }
 791
 792         block_info->start_workset = workset_clone(ws);
 793
 794         /* process the block from start to end */
 795         DB((dbg, DBG_WSETS, "Processing...\n"));
 796         instr_nr = 0;
 797         /* TODO: this leaks (into the obstack)... */
 798         new_vals = new_workset();
 799
 800         sched_foreach(block, irn) {
 801                 int i, arity;
 802                 assert(workset_get_length(ws) <= n_regs);
 803
 804                 /* Phis are no real instr (see insert_starters()) */
 805                 if (is_Phi(irn)) {
 806                         continue;
 807                 }
 808                 DB((dbg, DBG_DECIDE, "  ...%+F\n", irn));
 809
 810                 /* set instruction in the workset */
 811                 instr = irn;
 812
 813                 /* allocate all values _used_ by this instruction */
 814                 workset_clear(new_vals);
 815                 for(i = 0, arity = get_irn_arity(irn); i < arity; ++i) {
 816                         ir_node *in = get_irn_n(irn, i);
 817                         if (!arch_irn_consider_in_reg_alloc(cls, in))
 818                                 continue;
 819
 820                         /* (note that "spilled" is irrelevant here) */
 821                         workset_insert(new_vals, in, false);
 822                 }
 823                 displace(new_vals, 1);
 824
 825                 /* allocate all values _defined_ by this instruction */
 826                 workset_clear(new_vals);
 827                 if (get_irn_mode(irn) == mode_T) {
 828                         const ir_edge_t *edge;
 829
 830                         foreach_out_edge(irn, edge) {
 831                                 ir_node *proj = get_edge_src_irn(edge);
 832                                 if (!arch_irn_consider_in_reg_alloc(cls, proj))
 833                                         continue;
 834                                 workset_insert(new_vals, proj, false);
 835                         }
 836                 } else {
 837                         if (!arch_irn_consider_in_reg_alloc(cls, irn))
 838                                 continue;
 839                         workset_insert(new_vals, irn, false);
 840                 }
 841                 displace(new_vals, 0);
 842
 843                 instr_nr++;
 844         }
 845
 846         /* Remember end-workset for this block */
 847         block_info->end_workset = workset_clone(ws);
 848         DB((dbg, DBG_WSETS, "End workset for %+F:\n", block));
 849         workset_foreach(ws, irn, iter)
 850                 DB((dbg, DBG_WSETS, "  %+F (%u)\n", irn,
 851                      workset_get_time(ws, iter)));
 852 }
 853
 854 /**
 855  * 'decide' is block-local and makes assumptions
 856  * about the set of live-ins. Thus we must adapt the
 857  * live-outs to the live-ins at each block-border.
 858  */
 859 static void fix_block_borders(ir_node *block, void *data)
 860 {
 861         workset_t    *start_workset;
 862         int           arity;
 863         int           i;
 864         int           iter;
 865         (void) data;
 866
 867         DB((dbg, DBG_FIX, "\n"));
 868         DB((dbg, DBG_FIX, "Fixing %+F\n", block));
 869
 870         arity = get_irn_arity(block);
 871         /* can happen for endless loops */
 872         if (arity == 0)
 873                 return;
 874
 875         start_workset = get_block_info(block)->start_workset;
 876
 877         /* process all pred blocks */
 878         for (i = 0; i < arity; ++i) {
 879                 ir_node   *pred = get_Block_cfgpred_block(block, i);
 880                 workset_t *pred_end_workset = get_block_info(pred)->end_workset;
 881                 ir_node   *node;
 882
 883                 DB((dbg, DBG_FIX, "  Pred %+F\n", pred));
 884
 885                 /* spill all values not used anymore */
 886                 workset_foreach(pred_end_workset, node, iter) {
 887                         ir_node *n2;
 888                         int      iter2;
 889                         bool     found = false;
 890                         workset_foreach(start_workset, n2, iter2) {
 891                                 if (n2 == node) {
 892                                         found = true;
 893                                         break;
 894                                 }
 895                                 /* note that we do not look at phi inputs, becuase the values
 896                                  * will be either live-end and need no spill or
 897                                  * they have other users in which must be somewhere else in the
 898                                  * workset */
 899                         }
 900
 901                         if (found)
 902                                 continue;
 903
 904                         if (move_spills && be_is_live_in(lv, block, node)
 905                                         && !pred_end_workset->vals[iter].spilled) {
 906                                 ir_node *insert_point;
 907                                 if (arity > 1) {
 908                                         insert_point = be_get_end_of_block_insertion_point(pred);
 909                                         insert_point = sched_prev(insert_point);
 910                                 } else {
 911                                         insert_point = block;
 912                                 }
 913                                 DB((dbg, DBG_SPILL, "Spill %+F after %+F\n", node,
 914                                      insert_point));
 915                                 be_add_spill(senv, node, insert_point);
 916                         }
 917                 }
 918
 919                 /* reload missing values in predecessors, add missing spills */
 920                 workset_foreach(start_workset, node, iter) {
 921                         const loc_t *l    = &start_workset->vals[iter];
 922                         const loc_t *pred_loc;
 923
 924                         /* if node is a phi of the current block we reload
 925                          * the corresponding argument, else node itself */
 926                         if (is_Phi(node) && get_nodes_block(node) == block) {
 927                                 node = get_irn_n(node, i);
 928                                 assert(!l->spilled);
 929
 930                                 /* we might have unknowns as argument for the phi */
 931                                 if (!arch_irn_consider_in_reg_alloc(cls, node))
 932                                         continue;
 933                         }
 934
 935                         /* check if node is in a register at end of pred */
 936                         pred_loc = workset_contains(pred_end_workset, node);
 937                         if (pred_loc != NULL) {
 938                                 /* we might have to spill value on this path */
 939                                 if (move_spills && !pred_loc->spilled && l->spilled) {
 940                                         ir_node *insert_point
 941                                                 = be_get_end_of_block_insertion_point(pred);
 942                                         insert_point = sched_prev(insert_point);
 943                                         DB((dbg, DBG_SPILL, "Spill %+F after %+F\n", node,
 944                                             insert_point));
 945                                         be_add_spill(senv, node, insert_point);
 946                                 }
 947                         } else {
 948                                 /* node is not in register at the end of pred -> reload it */
 949                                 DB((dbg, DBG_FIX, "    reload %+F\n", node));
 950                                 DB((dbg, DBG_SPILL, "Reload %+F before %+F,%d\n", node, block, i));
 951                                 be_add_reload_on_edge(senv, node, block, i, cls, 1);
 952                         }
 953                 }
 954         }
 955 }
 956
 957 static void add_block(ir_node *block, void *data)
 958 {
 959         (void) data;
 960         ARR_APP1(ir_node*, blocklist, block);
 961 }
 962
 963 static void be_spill_belady(be_irg_t *birg, const arch_register_class_t *rcls)
 964 {
 965         int i;
 966         ir_graph *irg = be_get_birg_irg(birg);
 967
 968         be_liveness_assure_sets(be_assure_liveness(birg));
 969
 970         stat_ev_tim_push();
 971         /* construct control flow loop tree */
 972         if (! (get_irg_loopinfo_state(irg) & loopinfo_cf_consistent)) {
 973                 construct_cf_backedges(irg);
 974         }
 975         stat_ev_tim_pop("belady_time_backedges");
 976
 977         stat_ev_tim_push();
 978         be_clear_links(irg);
 979         stat_ev_tim_pop("belady_time_clear_links");
 980
 981         ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
 982
 983         /* init belady env */
 984         stat_ev_tim_push();
 985         obstack_init(&obst);
 986         cls       = rcls;
 987         lv        = be_get_birg_liveness(birg);
 988         n_regs    = cls->n_regs - be_put_ignore_regs(birg, cls, NULL);
 989         ws        = new_workset();
 990         uses      = be_begin_uses(irg, lv);
 991         loop_ana  = be_new_loop_pressure(birg, cls);
 992         senv      = be_new_spill_env(birg);
 993         blocklist = NEW_ARR_F(ir_node*, 0);
 994         irg_block_edges_walk(get_irg_start_block(irg), NULL, add_block, NULL);
 995         stat_ev_tim_pop("belady_time_init");
 996
 997         stat_ev_tim_push();
 998         /* walk blocks in reverse postorder */
 999         for (i = ARR_LEN(blocklist) - 1; i >= 0; --i) {
1000                 process_block(blocklist[i]);
1001         }
1002         DEL_ARR_F(blocklist);
1003         stat_ev_tim_pop("belady_time_belady");
1004
1005         stat_ev_tim_push();
1006         /* belady was block-local, fix the global flow by adding reloads on the
1007          * edges */
1008         irg_block_walk_graph(irg, fix_block_borders, NULL, NULL);
1009         stat_ev_tim_pop("belady_time_fix_borders");
1010
1011         ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
1012
1013         /* Insert spill/reload nodes into the graph and fix usages */
1014         be_insert_spills_reloads(senv);
1015
1016         /* clean up */
1017         be_delete_spill_env(senv);
1018         be_end_uses(uses);
1019         be_free_loop_pressure(loop_ana);
1020         obstack_free(&obst, NULL);
1021 }
1022
1023 void be_init_spillbelady(void)
1024 {
1025         static be_spiller_t belady_spiller = {
1026                 be_spill_belady
1027         };
1028         lc_opt_entry_t *be_grp       = lc_opt_get_grp(firm_opt_get_root(), "be");
1029         lc_opt_entry_t *belady_group = lc_opt_get_grp(be_grp, "belady");
1030         lc_opt_add_table(belady_group, options);
1031
1032         be_register_spiller("belady", &belady_spiller);
1033         FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
1034 }
1035
1036 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady);