nsz Git - libfirm/blob - ir/be/bespillbelady.c

   1 /*
   2  * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
   3  *
   4  * This file is part of libFirm.
   5  *
   6  * This file may be distributed and/or modified under the terms of the
   7  * GNU General Public License version 2 as published by the Free Software
   8  * Foundation and appearing in the file LICENSE.GPL included in the
   9  * packaging of this file.
  10  *
  11  * Licensees holding valid libFirm Professional Edition licenses may use
  12  * this file in accordance with the libFirm Commercial License.
  13  * Agreement provided with the Software.
  14  *
  15  * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
  16  * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  17  * PURPOSE.
  18  */
  19
  20 /**
  21  * @file
  22  * @brief       Beladys spillalgorithm.
  23  * @author      Daniel Grund, Matthias Braun
  24  * @date        20.09.2005
  25  * @version     $Id$
  26  */
  27 #ifdef HAVE_CONFIG_H
  28 #include "config.h"
  29 #endif
  30
  31 #include <stdbool.h>
  32
  33 #include "obst.h"
  34 #include "irprintf_t.h"
  35 #include "irgraph.h"
  36 #include "irnode.h"
  37 #include "irmode.h"
  38 #include "irgwalk.h"
  39 #include "irloop.h"
  40 #include "iredges_t.h"
  41 #include "ircons_t.h"
  42 #include "irprintf.h"
  43 #include "irnodeset.h"
  44
  45 #include "beutil.h"
  46 #include "bearch_t.h"
  47 #include "beuses.h"
  48 #include "besched_t.h"
  49 #include "beirgmod.h"
  50 #include "belive_t.h"
  51 #include "benode_t.h"
  52 #include "bechordal_t.h"
  53 #include "bespilloptions.h"
  54 #include "beloopana.h"
  55 #include "beirg_t.h"
  56 #include "bespill.h"
  57 #include "bemodule.h"
  58
  59 #define DBG_SPILL     1
  60 #define DBG_WSETS     2
  61 #define DBG_FIX       4
  62 #define DBG_DECIDE    8
  63 #define DBG_START    16
  64 #define DBG_SLOTS    32
  65 #define DBG_TRACE    64
  66 #define DBG_WORKSET 128
  67 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  68
  69 #define TIME_UNDEFINED 6666
  70
  71 //#define LOOK_AT_LOOPDEPTH
  72
  73 /**
  74  * An association between a node and a point in time.
  75  */
  76 typedef struct loc_t {
  77         ir_node          *node;
  78         unsigned          time;     /**< A use time (see beuses.h). */
  79         bool              spilled;  /**< the value was already spilled on this path */
  80 } loc_t;
  81
  82 typedef struct _workset_t {
  83         int   len;          /**< current length */
  84         loc_t vals[0];      /**< inlined array of the values/distances in this working set */
  85 } workset_t;
  86
  87 static struct obstack               obst;
  88 static const arch_register_class_t *cls;
  89 static const be_lv_t               *lv;
  90 static be_loopana_t                *loop_ana;
  91 static int                          n_regs;
  92 static workset_t                   *ws;     /**< the main workset used while
  93                                                      processing a block. */
  94 static be_uses_t                   *uses;   /**< env for the next-use magic */
  95 static ir_node                     *instr;  /**< current instruction */
  96 static unsigned                     instr_nr; /**< current instruction number
  97                                                        (relative to block start) */
  98 static spill_env_t                 *senv;   /**< see bespill.h */
  99 static ir_node                    **blocklist;
 100
 101 static bool                         move_spills      = true;
 102 static bool                         respectloopdepth = true;
 103 static bool                         improve_known_preds = true;
 104 /* factor to weight the different costs of reloading/rematerializing a node
 105    (see bespill.h be_get_reload_costs_no_weight) */
 106 static int                          remat_bonus      = 10;
 107
 108 static const lc_opt_table_entry_t options[] = {
 109         LC_OPT_ENT_BOOL   ("movespills", "try to move spills out of loops", &move_spills),
 110         LC_OPT_ENT_BOOL   ("respectloopdepth", "exprimental (outermost loop cutting)", &respectloopdepth),
 111         LC_OPT_ENT_BOOL   ("improveknownpreds", "experimental (known preds cutting)", &improve_known_preds),
 112         LC_OPT_ENT_INT    ("rematbonus", "give bonus to rematerialisable nodes", &remat_bonus),
 113         LC_OPT_LAST
 114 };
 115
 116 static int loc_compare(const void *a, const void *b)
 117 {
 118         const loc_t *p = a;
 119         const loc_t *q = b;
 120         return p->time - q->time;
 121 }
 122
 123 void workset_print(const workset_t *w)
 124 {
 125         int i;
 126
 127         for(i = 0; i < w->len; ++i) {
 128                 ir_fprintf(stderr, "%+F %d\n", w->vals[i].node, w->vals[i].time);
 129         }
 130 }
 131
 132 /**
 133  * Alloc a new workset on obstack @p ob with maximum size @p max
 134  */
 135 static workset_t *new_workset(void)
 136 {
 137         workset_t *res;
 138         size_t     size = sizeof(*res) + n_regs * sizeof(res->vals[0]);
 139
 140         res  = obstack_alloc(&obst, size);
 141         memset(res, 0, size);
 142         return res;
 143 }
 144
 145 /**
 146  * Alloc a new instance on obstack and make it equal to @param workset
 147  */
 148 static workset_t *workset_clone(workset_t *workset)
 149 {
 150         workset_t *res;
 151         size_t size = sizeof(*res) + n_regs * sizeof(res->vals[0]);
 152         res = obstack_alloc(&obst, size);
 153         memcpy(res, workset, size);
 154         return res;
 155 }
 156
 157 /**
 158  * Copy workset @param src to @param tgt
 159  */
 160 static void workset_copy(workset_t *dest, const workset_t *src)
 161 {
 162         size_t size = sizeof(*src) + n_regs * sizeof(src->vals[0]);
 163         memcpy(dest, src, size);
 164 }
 165
 166 /**
 167  * Overwrites the current content array of @param ws with the
 168  * @param count locations given at memory @param locs.
 169  * Set the length of @param ws to count.
 170  */
 171 static void workset_bulk_fill(workset_t *workset, int count, const loc_t *locs)
 172 {
 173         workset->len = count;
 174         memcpy(&(workset->vals[0]), locs, count * sizeof(locs[0]));
 175 }
 176
 177 /**
 178  * Inserts the value @p val into the workset, iff it is not
 179  * already contained. The workset must not be full.
 180  */
 181 static void workset_insert(workset_t *workset, ir_node *val, bool spilled)
 182 {
 183         loc_t *loc;
 184         int    i;
 185         /* check for current regclass */
 186         assert(arch_irn_consider_in_reg_alloc(cls, val));
 187
 188         /* check if val is already contained */
 189         for (i = 0; i < workset->len; ++i) {
 190                 loc = &workset->vals[i];
 191                 if (loc->node == val) {
 192                         if (spilled) {
 193                                 loc->spilled = true;
 194                         }
 195                         return;
 196                 }
 197         }
 198
 199         /* insert val */
 200         assert(workset->len < n_regs && "Workset already full!");
 201         loc           = &workset->vals[workset->len];
 202         loc->node     = val;
 203         loc->spilled  = spilled;
 204         loc->time     = TIME_UNDEFINED;
 205         workset->len++;
 206 }
 207
 208 /**
 209  * Removes all entries from this workset
 210  */
 211 static void workset_clear(workset_t *workset)
 212 {
 213         workset->len = 0;
 214 }
 215
 216 /**
 217  * Removes the value @p val from the workset if present.
 218  */
 219 static INLINE void workset_remove(workset_t *workset, ir_node *val)
 220 {
 221         int i;
 222         for(i = 0; i < workset->len; ++i) {
 223                 if (workset->vals[i].node == val) {
 224                         workset->vals[i] = workset->vals[--workset->len];
 225                         return;
 226                 }
 227         }
 228 }
 229
 230 static INLINE const loc_t *workset_contains(const workset_t *ws,
 231                                             const ir_node *val)
 232 {
 233         int i;
 234
 235         for (i = 0; i < ws->len; ++i) {
 236                 if (ws->vals[i].node == val)
 237                         return &ws->vals[i];
 238         }
 239
 240         return NULL;
 241 }
 242
 243 /**
 244  * Iterates over all values in the working set.
 245  * @p ws The workset to iterate
 246  * @p v  A variable to put the current value in
 247  * @p i  An integer for internal use
 248  */
 249 #define workset_foreach(ws, v, i)       for(i=0; \
 250                                                                                 v=(i < ws->len) ? ws->vals[i].node : NULL, i < ws->len; \
 251                                                                                 ++i)
 252
 253 #define workset_set_time(ws, i, t) (ws)->vals[i].time=t
 254 #define workset_get_time(ws, i) (ws)->vals[i].time
 255 #define workset_set_length(ws, length) (ws)->len = length
 256 #define workset_get_length(ws) ((ws)->len)
 257 #define workset_get_val(ws, i) ((ws)->vals[i].node)
 258 #define workset_sort(ws) qsort((ws)->vals, (ws)->len, sizeof((ws)->vals[0]), loc_compare);
 259
 260 typedef struct _block_info_t
 261 {
 262         workset_t *start_workset;
 263         workset_t *end_workset;
 264 } block_info_t;
 265
 266
 267 static void *new_block_info(void)
 268 {
 269         block_info_t *res = obstack_alloc(&obst, sizeof(res[0]));
 270         memset(res, 0, sizeof(res[0]));
 271
 272         return res;
 273 }
 274
 275 #define get_block_info(block)        ((block_info_t *)get_irn_link(block))
 276 #define set_block_info(block, info)  set_irn_link(block, info)
 277
 278 /**
 279  * @return The distance to the next use or 0 if irn has dont_spill flag set
 280  */
 281 static INLINE unsigned get_distance(ir_node *from, unsigned from_step,
 282                                     const ir_node *def, int skip_from_uses)
 283 {
 284         be_next_use_t use;
 285         int           flags = arch_irn_get_flags(def);
 286         unsigned      costs;
 287         unsigned      time;
 288
 289         assert(! (flags & arch_irn_flags_ignore));
 290
 291         use  = be_get_next_use(uses, from, from_step, def, skip_from_uses);
 292         time = use.time;
 293         if (USES_IS_INFINITE(time))
 294                 return USES_INFINITY;
 295
 296         /* We have to keep nonspillable nodes in the workingset */
 297         if (flags & arch_irn_flags_dont_spill)
 298                 return 0;
 299
 300         /* give some bonus to rematerialisable nodes */
 301         if (remat_bonus > 0) {
 302                 costs = be_get_reload_costs_no_weight(senv, def, use.before);
 303                 assert(costs * remat_bonus < 1000);
 304                 time  += 1000 - (costs * remat_bonus);
 305         }
 306
 307         return time;
 308 }
 309
 310 /**
 311  * Performs the actions necessary to grant the request that:
 312  * - new_vals can be held in registers
 313  * - as few as possible other values are disposed
 314  * - the worst values get disposed
 315  *
 316  * @p is_usage indicates that the values in new_vals are used (not defined)
 317  * In this case reloads must be performed
 318  */
 319 static void displace(workset_t *new_vals, int is_usage)
 320 {
 321         ir_node **to_insert = alloca(n_regs * sizeof(to_insert[0]));
 322         bool     *spilled   = alloca(n_regs * sizeof(spilled[0]));
 323         ir_node  *val;
 324         int       i;
 325         int       len;
 326         int       spills_needed;
 327         int       demand;
 328         int       iter;
 329
 330         /* 1. Identify the number of needed slots and the values to reload */
 331         demand = 0;
 332         workset_foreach(new_vals, val, iter) {
 333                 bool reloaded = false;
 334
 335                 if (! workset_contains(ws, val)) {
 336                         DB((dbg, DBG_DECIDE, "    insert %+F\n", val));
 337                         if (is_usage) {
 338                                 DB((dbg, DBG_SPILL, "Reload %+F before %+F\n", val, instr));
 339                                 be_add_reload(senv, val, instr, cls, 1);
 340                                 reloaded = true;
 341                         }
 342                 } else {
 343                         DB((dbg, DBG_DECIDE, "    %+F already in workset\n", val));
 344                         assert(is_usage);
 345                         /* remove the value from the current workset so it is not accidently
 346                          * spilled */
 347                         workset_remove(ws, val);
 348                 }
 349                 spilled[demand]   = reloaded;
 350                 to_insert[demand] = val;
 351                 ++demand;
 352         }
 353
 354         /* 2. Make room for at least 'demand' slots */
 355         len           = workset_get_length(ws);
 356         spills_needed = len + demand - n_regs;
 357         assert(spills_needed <= len);
 358
 359         /* Only make more free room if we do not have enough */
 360         if (spills_needed > 0) {
 361                 ir_node   *curr_bb  = NULL;
 362                 workset_t *ws_start = NULL;
 363
 364                 if (move_spills) {
 365                         curr_bb  = get_nodes_block(instr);
 366                         ws_start = get_block_info(curr_bb)->start_workset;
 367                 }
 368
 369                 DB((dbg, DBG_DECIDE, "    disposing %d values\n", spills_needed));
 370
 371                 /* calculate current next-use distance for live values */
 372                 for (i = 0; i < len; ++i) {
 373                         ir_node  *val  = workset_get_val(ws, i);
 374                         unsigned  dist = get_distance(instr, instr_nr, val, !is_usage);
 375                         workset_set_time(ws, i, dist);
 376                 }
 377
 378                 /* sort entries by increasing nextuse-distance*/
 379                 workset_sort(ws);
 380
 381                 for (i = len - spills_needed; i < len; ++i) {
 382                         ir_node *val = ws->vals[i].node;
 383
 384                         DB((dbg, DBG_DECIDE, "    disposing node %+F (%u)\n", val,
 385                              workset_get_time(ws, i)));
 386
 387                         if (move_spills) {
 388                                 if (!USES_IS_INFINITE(ws->vals[i].time)
 389                                                 && !ws->vals[i].spilled) {
 390                                         ir_node *after_pos = sched_prev(instr);
 391                                         DB((dbg, DBG_DECIDE, "Spill %+F after node %+F\n", val,
 392                                                 after_pos));
 393                                         be_add_spill(senv, val, after_pos);
 394                                 }
 395                         }
 396                 }
 397
 398                 /* kill the last 'demand' entries in the array */
 399                 workset_set_length(ws, len - spills_needed);
 400         }
 401
 402         /* 3. Insert the new values into the workset */
 403         for (i = 0; i < demand; ++i) {
 404                 ir_node *val = to_insert[i];
 405
 406                 workset_insert(ws, val, spilled[i]);
 407         }
 408 }
 409
 410 enum {
 411         AVAILABLE_EVERYWHERE,
 412         AVAILABLE_NOWHERE,
 413         AVAILABLE_PARTLY,
 414         AVAILABLE_UNKNOWN
 415 };
 416
 417 static unsigned available_in_all_preds(workset_t* const* pred_worksets,
 418                                        size_t n_pred_worksets,
 419                                        const ir_node *value, bool is_local_phi)
 420 {
 421         size_t i;
 422         bool   avail_everywhere = true;
 423         bool   avail_nowhere    = true;
 424
 425         assert(n_pred_worksets > 0);
 426
 427         /* value available in all preds? */
 428         for (i = 0; i < n_pred_worksets; ++i) {
 429                 bool             found     = false;
 430                 const workset_t *p_workset = pred_worksets[i];
 431                 int              p_len     = workset_get_length(p_workset);
 432                 int              p_i;
 433                 const ir_node   *l_value;
 434
 435                 if (is_local_phi) {
 436                         assert(is_Phi(value));
 437                         l_value = get_irn_n(value, i);
 438                 } else {
 439                         l_value = value;
 440                 }
 441
 442                 for (p_i = 0; p_i < p_len; ++p_i) {
 443                         const loc_t *p_l = &p_workset->vals[p_i];
 444                         if (p_l->node != l_value)
 445                                 continue;
 446
 447                         found = true;
 448                         break;
 449                 }
 450
 451                 if (found) {
 452                         avail_nowhere = false;
 453                 } else {
 454                         avail_everywhere = false;
 455                 }
 456         }
 457
 458         if (avail_everywhere) {
 459                 assert(!avail_nowhere);
 460                 return AVAILABLE_EVERYWHERE;
 461         } else if (avail_nowhere) {
 462                 return AVAILABLE_NOWHERE;
 463         } else {
 464                 return AVAILABLE_PARTLY;
 465         }
 466 }
 467
 468 /** Decides whether a specific node should be in the start workset or not
 469  *
 470  * @param env      belady environment
 471  * @param first
 472  * @param node     the node to test
 473  * @param loop     the loop of the node
 474  */
 475 static loc_t to_take_or_not_to_take(ir_node* first, ir_node *node,
 476                                     ir_loop *loop, unsigned available)
 477 {
 478         be_next_use_t next_use;
 479         loc_t         loc;
 480
 481         loc.time    = USES_INFINITY;
 482         loc.node    = node;
 483         loc.spilled = false;
 484
 485         if (!arch_irn_consider_in_reg_alloc(cls, node)) {
 486                 loc.time = USES_INFINITY;
 487                 return loc;
 488         }
 489
 490         /* We have to keep nonspillable nodes in the workingset */
 491         if (arch_irn_get_flags(node) & arch_irn_flags_dont_spill) {
 492                 loc.time = 0;
 493                 DB((dbg, DBG_START, "    %+F taken (dontspill node)\n", node, loc.time));
 494                 return loc;
 495         }
 496
 497         next_use = be_get_next_use(uses, first, 0, node, 0);
 498         if (USES_IS_INFINITE(next_use.time)) {
 499                 // the nodes marked as live in shouldn't be dead, so it must be a phi
 500                 assert(is_Phi(node));
 501                 loc.time = USES_INFINITY;
 502                 DB((dbg, DBG_START, "    %+F not taken (dead)\n", node));
 503                 return loc;
 504         }
 505
 506         loc.time = next_use.time;
 507
 508         if (improve_known_preds) {
 509                 if (available == AVAILABLE_EVERYWHERE) {
 510                         DB((dbg, DBG_START, "    %+F taken (%u, live in all preds)\n",
 511                             node, loc.time));
 512                         return loc;
 513                 } else if(available == AVAILABLE_NOWHERE) {
 514                         DB((dbg, DBG_START, "    %+F not taken (%u, live in no pred)\n",
 515                             node, loc.time));
 516                         loc.time = USES_INFINITY;
 517                         return loc;
 518                 }
 519         }
 520
 521         if (!respectloopdepth || next_use.outermost_loop >= get_loop_depth(loop)) {
 522                 DB((dbg, DBG_START, "    %+F taken (%u, loop %d)\n", node, loc.time,
 523                     next_use.outermost_loop));
 524         } else {
 525                 loc.time = USES_PENDING;
 526                 DB((dbg, DBG_START, "    %+F delayed (outerdepth %d < loopdepth %d)\n",
 527                     node, next_use.outermost_loop, get_loop_depth(loop)));
 528         }
 529
 530         return loc;
 531 }
 532
 533 /**
 534  * Computes the start-workset for a block with multiple predecessors. We assume
 535  * that at least 1 of the predeccesors is a back-edge which means we're at the
 536  * beginning of a loop. We try to reload as much values as possible now so they
 537  * don't get reloaded inside the loop.
 538  */
 539 static void decide_start_workset(const ir_node *block)
 540 {
 541         ir_loop    *loop = get_irn_loop(block);
 542         ir_node    *first;
 543         ir_node    *node;
 544         loc_t       loc;
 545         loc_t      *starters;
 546         loc_t      *delayed;
 547         int         i, len, ws_count;
 548         int             free_slots, free_pressure_slots;
 549         unsigned    pressure;
 550         int         arity;
 551         workset_t **pred_worksets;
 552         bool        all_preds_known;
 553
 554         /* check predecessors */
 555         arity           = get_irn_arity(block);
 556         pred_worksets   = alloca(sizeof(pred_worksets[0]) * arity);
 557         all_preds_known = true;
 558         for(i = 0; i < arity; ++i) {
 559                 ir_node      *pred_block = get_Block_cfgpred_block(block, i);
 560                 block_info_t *pred_info  = get_block_info(pred_block);
 561
 562                 if (pred_info == NULL) {
 563                         pred_worksets[i] = NULL;
 564                         all_preds_known  = false;
 565                 } else {
 566                         pred_worksets[i] = pred_info->end_workset;
 567                 }
 568         }
 569
 570         /* Collect all values living at start of block */
 571         starters = NEW_ARR_F(loc_t, 0);
 572         delayed  = NEW_ARR_F(loc_t, 0);
 573
 574         DB((dbg, DBG_START, "Living at start of %+F:\n", block));
 575         first = sched_first(block);
 576
 577         /* check all Phis first */
 578         sched_foreach(block, node) {
 579                 unsigned available;
 580
 581                 if (! is_Phi(node))
 582                         break;
 583                 if (!arch_irn_consider_in_reg_alloc(cls, node))
 584                         continue;
 585
 586                 if (all_preds_known) {
 587                         available = available_in_all_preds(pred_worksets, arity, node, true);
 588                 } else {
 589                         available = AVAILABLE_UNKNOWN;
 590                 }
 591
 592                 loc = to_take_or_not_to_take(first, node, loop, available);
 593
 594                 if (! USES_IS_INFINITE(loc.time)) {
 595                         if (USES_IS_PENDING(loc.time))
 596                                 ARR_APP1(loc_t, delayed, loc);
 597                         else
 598                                 ARR_APP1(loc_t, starters, loc);
 599                 } else {
 600                         be_spill_phi(senv, node);
 601                 }
 602         }
 603
 604         /* check all Live-Ins */
 605         be_lv_foreach(lv, block, be_lv_state_in, i) {
 606                 ir_node *node = be_lv_get_irn(lv, block, i);
 607                 unsigned available;
 608
 609                 if (all_preds_known) {
 610                         available = available_in_all_preds(pred_worksets, arity, node, false);
 611                 } else {
 612                         available = AVAILABLE_UNKNOWN;
 613                 }
 614
 615                 loc = to_take_or_not_to_take(first, node, loop, available);
 616
 617                 if (! USES_IS_INFINITE(loc.time)) {
 618                         if (USES_IS_PENDING(loc.time))
 619                                 ARR_APP1(loc_t, delayed, loc);
 620                         else
 621                                 ARR_APP1(loc_t, starters, loc);
 622                 }
 623         }
 624
 625         pressure            = be_get_loop_pressure(loop_ana, cls, loop);
 626         assert(ARR_LEN(delayed) <= (signed)pressure);
 627         free_slots          = n_regs - ARR_LEN(starters);
 628         free_pressure_slots = n_regs - (pressure - ARR_LEN(delayed));
 629         free_slots          = MIN(free_slots, free_pressure_slots);
 630
 631         /* so far we only put nodes into the starters list that are used inside
 632          * the loop. If register pressure in the loop is low then we can take some
 633          * values and let them live through the loop */
 634         DB((dbg, DBG_START, "Loop pressure %d, taking %d delayed vals\n",
 635             pressure, free_slots));
 636         if (free_slots > 0) {
 637                 qsort(delayed, ARR_LEN(delayed), sizeof(delayed[0]), loc_compare);
 638
 639                 for (i = 0; i < ARR_LEN(delayed) && free_slots > 0; ++i) {
 640                         int    p, arity;
 641                         loc_t *loc = & delayed[i];
 642
 643                         if (!is_Phi(loc->node)) {
 644                                 /* don't use values which are dead in a known predecessors
 645                                  * to not induce unnecessary reloads */
 646                                 arity = get_irn_arity(block);
 647                                 for (p = 0; p < arity; ++p) {
 648                                         ir_node      *pred_block = get_Block_cfgpred_block(block, p);
 649                                         block_info_t *pred_info  = get_block_info(pred_block);
 650
 651                                         if (pred_info == NULL)
 652                                                 continue;
 653
 654                                         if (!workset_contains(pred_info->end_workset, loc->node)) {
 655                                                 DB((dbg, DBG_START,
 656                                                         "    delayed %+F not live at pred %+F\n", loc->node,
 657                                                         pred_block));
 658                                                 goto skip_delayed;
 659                                         }
 660                                 }
 661                         }
 662
 663                         DB((dbg, DBG_START, "    delayed %+F taken\n", loc->node));
 664                         ARR_APP1(loc_t, starters, *loc);
 665                         loc->node = NULL;
 666                         --free_slots;
 667                 skip_delayed:
 668                         ;
 669                 }
 670         }
 671
 672         /* spill phis (the actual phis not just their values) that are in this block
 673          * but not in the start workset */
 674         for (i = ARR_LEN(delayed) - 1; i >= 0; --i) {
 675                 ir_node *node = delayed[i].node;
 676                 if (node == NULL || !is_Phi(node) || get_nodes_block(node) != block)
 677                         continue;
 678
 679                 DB((dbg, DBG_START, "    spilling delayed phi %+F\n", node));
 680                 be_spill_phi(senv, node);
 681         }
 682         DEL_ARR_F(delayed);
 683
 684         /* Sort start values by first use */
 685         qsort(starters, ARR_LEN(starters), sizeof(starters[0]), loc_compare);
 686
 687         /* Copy the best ones from starters to start workset */
 688         ws_count = MIN(ARR_LEN(starters), n_regs);
 689         workset_clear(ws);
 690         workset_bulk_fill(ws, ws_count, starters);
 691
 692         /* spill phis (the actual phis not just their values) that are in this block
 693          * but not in the start workset */
 694         len = ARR_LEN(starters);
 695         for (i = ws_count; i < len; ++i) {
 696                 ir_node *node = starters[i].node;
 697                 if (! is_Phi(node) || get_nodes_block(node) != block)
 698                         continue;
 699
 700                 DB((dbg, DBG_START, "    spilling phi %+F\n", node));
 701                 be_spill_phi(senv, node);
 702         }
 703
 704         DEL_ARR_F(starters);
 705
 706         /* determine spill status of the values: If there's 1 pred block (which
 707          * is no backedge) where the value is spilled then we must set it to
 708          * spilled here. */
 709         for(i = 0; i < ws_count; ++i) {
 710                 loc_t   *loc     = &ws->vals[i];
 711                 ir_node *value   = loc->node;
 712                 bool     spilled;
 713                 int      n;
 714
 715                 /* phis from this block aren't spilled */
 716                 if (get_nodes_block(value) == block) {
 717                         assert(is_Phi(value));
 718                         loc->spilled = false;
 719                         continue;
 720                 }
 721
 722                 /* determine if value was spilled on any predecessor */
 723                 spilled = false;
 724                 for(n = 0; n < arity; ++n) {
 725                         workset_t *pred_workset = pred_worksets[n];
 726                         int        p_len;
 727                         int        p;
 728
 729                         if (pred_workset == NULL)
 730                                 continue;
 731
 732                         p_len = workset_get_length(pred_workset);
 733                         for(p = 0; p < p_len; ++p) {
 734                                 loc_t *l = &pred_workset->vals[p];
 735
 736                                 if (l->node != value)
 737                                         continue;
 738
 739                                 if (l->spilled) {
 740                                         spilled = true;
 741                                 }
 742                                 break;
 743                         }
 744                 }
 745
 746                 loc->spilled = spilled;
 747         }
 748 }
 749
 750 /**
 751  * For the given block @p block, decide for each values
 752  * whether it is used from a register or is reloaded
 753  * before the use.
 754  */
 755 static void process_block(ir_node *block)
 756 {
 757         workset_t       *new_vals;
 758         ir_node         *irn;
 759         int              iter;
 760         block_info_t    *block_info;
 761         int              arity;
 762
 763         /* no need to process a block twice */
 764         assert(get_block_info(block) == NULL);
 765
 766         /* construct start workset */
 767         arity = get_Block_n_cfgpreds(block);
 768         if (arity == 0) {
 769                 /* no predecessor -> empty set */
 770                 workset_clear(ws);
 771         } else if (arity == 1) {
 772                 /* one predecessor, copy it's end workset */
 773                 ir_node      *pred_block = get_Block_cfgpred_block(block, 0);
 774                 block_info_t *pred_info  = get_block_info(pred_block);
 775
 776                 assert(pred_info != NULL);
 777                 workset_copy(ws, pred_info->end_workset);
 778         } else {
 779                 /* multiple predecessors, do more advanced magic :) */
 780                 decide_start_workset(block);
 781         }
 782
 783         DB((dbg, DBG_DECIDE, "\n"));
 784         DB((dbg, DBG_DECIDE, "Decide for %+F\n", block));
 785
 786         block_info = new_block_info();
 787         set_block_info(block, block_info);
 788
 789         DB((dbg, DBG_WSETS, "Start workset for %+F:\n", block));
 790         workset_foreach(ws, irn, iter) {
 791                 DB((dbg, DBG_WSETS, "  %+F (%u)\n", irn,
 792                      workset_get_time(ws, iter)));
 793         }
 794
 795         block_info->start_workset = workset_clone(ws);
 796
 797         /* process the block from start to end */
 798         DB((dbg, DBG_WSETS, "Processing...\n"));
 799         instr_nr = 0;
 800         /* TODO: this leaks (into the obstack)... */
 801         new_vals = new_workset();
 802
 803         sched_foreach(block, irn) {
 804                 int i, arity;
 805                 assert(workset_get_length(ws) <= n_regs);
 806
 807                 /* Phis are no real instr (see insert_starters()) */
 808                 if (is_Phi(irn)) {
 809                         continue;
 810                 }
 811                 DB((dbg, DBG_DECIDE, "  ...%+F\n", irn));
 812
 813                 /* set instruction in the workset */
 814                 instr = irn;
 815
 816                 /* allocate all values _used_ by this instruction */
 817                 workset_clear(new_vals);
 818                 for(i = 0, arity = get_irn_arity(irn); i < arity; ++i) {
 819                         ir_node *in = get_irn_n(irn, i);
 820                         if (!arch_irn_consider_in_reg_alloc(cls, in))
 821                                 continue;
 822
 823                         /* (note that "spilled" is irrelevant here) */
 824                         workset_insert(new_vals, in, false);
 825                 }
 826                 displace(new_vals, 1);
 827
 828                 /* allocate all values _defined_ by this instruction */
 829                 workset_clear(new_vals);
 830                 if (get_irn_mode(irn) == mode_T) {
 831                         const ir_edge_t *edge;
 832
 833                         foreach_out_edge(irn, edge) {
 834                                 ir_node *proj = get_edge_src_irn(edge);
 835                                 if (!arch_irn_consider_in_reg_alloc(cls, proj))
 836                                         continue;
 837                                 workset_insert(new_vals, proj, false);
 838                         }
 839                 } else {
 840                         if (!arch_irn_consider_in_reg_alloc(cls, irn))
 841                                 continue;
 842                         workset_insert(new_vals, irn, false);
 843                 }
 844                 displace(new_vals, 0);
 845
 846                 instr_nr++;
 847         }
 848
 849         /* Remember end-workset for this block */
 850         block_info->end_workset = workset_clone(ws);
 851         DB((dbg, DBG_WSETS, "End workset for %+F:\n", block));
 852         workset_foreach(ws, irn, iter)
 853                 DB((dbg, DBG_WSETS, "  %+F (%u)\n", irn,
 854                      workset_get_time(ws, iter)));
 855 }
 856
 857 /**
 858  * 'decide' is block-local and makes assumptions
 859  * about the set of live-ins. Thus we must adapt the
 860  * live-outs to the live-ins at each block-border.
 861  */
 862 static void fix_block_borders(ir_node *block, void *data)
 863 {
 864         workset_t    *start_workset;
 865         int           arity;
 866         int           i;
 867         int           iter;
 868         (void) data;
 869
 870         DB((dbg, DBG_FIX, "\n"));
 871         DB((dbg, DBG_FIX, "Fixing %+F\n", block));
 872
 873         arity = get_irn_arity(block);
 874         /* can happen for endless loops */
 875         if (arity == 0)
 876                 return;
 877
 878         start_workset = get_block_info(block)->start_workset;
 879
 880         /* process all pred blocks */
 881         for (i = 0; i < arity; ++i) {
 882                 ir_node   *pred = get_Block_cfgpred_block(block, i);
 883                 workset_t *pred_end_workset = get_block_info(pred)->end_workset;
 884                 ir_node   *node;
 885
 886                 DB((dbg, DBG_FIX, "  Pred %+F\n", pred));
 887
 888                 /* spill all values not used anymore */
 889                 workset_foreach(pred_end_workset, node, iter) {
 890                         ir_node *n2;
 891                         int      iter2;
 892                         bool     found = false;
 893                         workset_foreach(start_workset, n2, iter2) {
 894                                 if (n2 == node) {
 895                                         found = true;
 896                                         break;
 897                                 }
 898                                 /* note that we do not look at phi inputs, becuase the values
 899                                  * will be either live-end and need no spill or
 900                                  * they have other users in which must be somewhere else in the
 901                                  * workset */
 902                         }
 903
 904                         if (found)
 905                                 continue;
 906
 907                         if (move_spills && be_is_live_in(lv, block, node)
 908                                         && !pred_end_workset->vals[iter].spilled) {
 909                                 ir_node *insert_point;
 910                                 if (arity > 1) {
 911                                         insert_point = be_get_end_of_block_insertion_point(pred);
 912                                         insert_point = sched_prev(insert_point);
 913                                 } else {
 914                                         insert_point = block;
 915                                 }
 916                                 DB((dbg, DBG_SPILL, "Spill %+F after %+F\n", node,
 917                                      insert_point));
 918                                 be_add_spill(senv, node, insert_point);
 919                         }
 920                 }
 921
 922                 /* reload missing values in predecessors, add missing spills */
 923                 workset_foreach(start_workset, node, iter) {
 924                         const loc_t *l    = &start_workset->vals[iter];
 925                         const loc_t *pred_loc;
 926
 927                         /* if node is a phi of the current block we reload
 928                          * the corresponding argument, else node itself */
 929                         if (is_Phi(node) && get_nodes_block(node) == block) {
 930                                 node = get_irn_n(node, i);
 931                                 assert(!l->spilled);
 932
 933                                 /* we might have unknowns as argument for the phi */
 934                                 if (!arch_irn_consider_in_reg_alloc(cls, node))
 935                                         continue;
 936                         }
 937
 938                         /* check if node is in a register at end of pred */
 939                         pred_loc = workset_contains(pred_end_workset, node);
 940                         if (pred_loc != NULL) {
 941                                 /* we might have to spill value on this path */
 942                                 if (move_spills && !pred_loc->spilled && l->spilled) {
 943                                         ir_node *insert_point
 944                                                 = be_get_end_of_block_insertion_point(pred);
 945                                         insert_point = sched_prev(insert_point);
 946                                         DB((dbg, DBG_SPILL, "Spill %+F after %+F\n", node,
 947                                             insert_point));
 948                                         be_add_spill(senv, node, insert_point);
 949                                 }
 950                         } else {
 951                                 /* node is not in register at the end of pred -> reload it */
 952                                 DB((dbg, DBG_FIX, "    reload %+F\n", node));
 953                                 DB((dbg, DBG_SPILL, "Reload %+F before %+F,%d\n", node, block, i));
 954                                 be_add_reload_on_edge(senv, node, block, i, cls, 1);
 955                         }
 956                 }
 957         }
 958 }
 959
 960 static void add_block(ir_node *block, void *data)
 961 {
 962         (void) data;
 963         ARR_APP1(ir_node*, blocklist, block);
 964 }
 965
 966 static void be_spill_belady(be_irg_t *birg, const arch_register_class_t *rcls)
 967 {
 968         int i;
 969         ir_graph *irg = be_get_birg_irg(birg);
 970
 971         be_liveness_assure_sets(be_assure_liveness(birg));
 972
 973         stat_ev_tim_push();
 974         /* construct control flow loop tree */
 975         if (! (get_irg_loopinfo_state(irg) & loopinfo_cf_consistent)) {
 976                 construct_cf_backedges(irg);
 977         }
 978         stat_ev_tim_pop("belady_time_backedges");
 979
 980         stat_ev_tim_push();
 981         be_clear_links(irg);
 982         stat_ev_tim_pop("belady_time_clear_links");
 983
 984         ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
 985
 986         /* init belady env */
 987         stat_ev_tim_push();
 988         obstack_init(&obst);
 989         cls       = rcls;
 990         lv        = be_get_birg_liveness(birg);
 991         n_regs    = cls->n_regs - be_put_ignore_regs(birg, cls, NULL);
 992         ws        = new_workset();
 993         uses      = be_begin_uses(irg, lv);
 994         loop_ana  = be_new_loop_pressure(birg, cls);
 995         senv      = be_new_spill_env(birg);
 996         blocklist = NEW_ARR_F(ir_node*, 0);
 997         irg_block_edges_walk(get_irg_start_block(irg), NULL, add_block, NULL);
 998         stat_ev_tim_pop("belady_time_init");
 999
1000         stat_ev_tim_push();
1001         /* walk blocks in reverse postorder */
1002         for (i = ARR_LEN(blocklist) - 1; i >= 0; --i) {
1003                 process_block(blocklist[i]);
1004         }
1005         DEL_ARR_F(blocklist);
1006         stat_ev_tim_pop("belady_time_belady");
1007
1008         stat_ev_tim_push();
1009         /* belady was block-local, fix the global flow by adding reloads on the
1010          * edges */
1011         irg_block_walk_graph(irg, fix_block_borders, NULL, NULL);
1012         stat_ev_tim_pop("belady_time_fix_borders");
1013
1014         ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
1015
1016         /* Insert spill/reload nodes into the graph and fix usages */
1017         be_insert_spills_reloads(senv);
1018
1019         /* clean up */
1020         be_delete_spill_env(senv);
1021         be_end_uses(uses);
1022         be_free_loop_pressure(loop_ana);
1023         obstack_free(&obst, NULL);
1024 }
1025
1026 void be_init_spillbelady(void)
1027 {
1028         static be_spiller_t belady_spiller = {
1029                 be_spill_belady
1030         };
1031         lc_opt_entry_t *be_grp       = lc_opt_get_grp(firm_opt_get_root(), "be");
1032         lc_opt_entry_t *belady_group = lc_opt_get_grp(be_grp, "belady");
1033         lc_opt_add_table(belady_group, options);
1034
1035         be_register_spiller("belady", &belady_spiller);
1036         FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
1037 }
1038
1039 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady);