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 #include "xmalloc.h"
  45 #include "pdeq.h"
  46
  47 #include "beutil.h"
  48 #include "bearch_t.h"
  49 #include "beuses.h"
  50 #include "besched_t.h"
  51 #include "beirgmod.h"
  52 #include "belive_t.h"
  53 #include "benode_t.h"
  54 #include "bechordal_t.h"
  55 #include "bespilloptions.h"
  56 #include "beloopana.h"
  57 #include "beirg_t.h"
  58 #include "bespill.h"
  59 #include "bemodule.h"
  60
  61 #define DBG_SPILL     1
  62 #define DBG_WSETS     2
  63 #define DBG_FIX       4
  64 #define DBG_DECIDE    8
  65 #define DBG_START    16
  66 #define DBG_SLOTS    32
  67 #define DBG_TRACE    64
  68 #define DBG_WORKSET 128
  69 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  70
  71 /* factor to weight the different costs of reloading/rematerializing a node
  72    (see bespill.h be_get_reload_costs_no_weight) */
  73 #define RELOAD_COST_FACTOR   10
  74
  75 #define TIME_UNDEFINED 6666
  76
  77 #define PLACE_SPILLS
  78
  79 /**
  80  * An association between a node and a point in time.
  81  */
  82 typedef struct loc_t {
  83         ir_node          *node;
  84         unsigned          time;     /**< A use time (see beuses.h). */
  85         bool              spilled;  /**< the value was already spilled on this path */
  86 } loc_t;
  87
  88 typedef struct _workset_t {
  89         int   len;          /**< current length */
  90         loc_t vals[0];      /**< inlined array of the values/distances in this working set */
  91 } workset_t;
  92
  93 static struct obstack               obst;
  94 static const arch_env_t            *arch_env;
  95 static const arch_register_class_t *cls;
  96 static const be_lv_t               *lv;
  97 static be_loopana_t                *loop_ana;
  98 static int                          n_regs;
  99 static workset_t                   *ws;     /**< the main workset used while
 100                                                      processing a block. */
 101 static be_uses_t                   *uses;   /**< env for the next-use magic */
 102 static ir_node                     *instr;  /**< current instruction */
 103 static unsigned                     instr_nr; /**< current instruction number
 104                                                        (relative to block start) */
 105 static ir_nodeset_t                 used;
 106 static spill_env_t                 *senv;   /**< see bespill.h */
 107 static pdeq                        *worklist;
 108
 109 static int loc_compare(const void *a, const void *b)
 110 {
 111         const loc_t *p = a;
 112         const loc_t *q = b;
 113         return p->time - q->time;
 114 }
 115
 116 void workset_print(const workset_t *w)
 117 {
 118         int i;
 119
 120         for(i = 0; i < w->len; ++i) {
 121                 ir_fprintf(stderr, "%+F %d\n", w->vals[i].node, w->vals[i].time);
 122         }
 123 }
 124
 125 /**
 126  * Alloc a new workset on obstack @p ob with maximum size @p max
 127  */
 128 static workset_t *new_workset(void)
 129 {
 130         workset_t *res;
 131         size_t     size = sizeof(*res) + n_regs * sizeof(res->vals[0]);
 132
 133         res  = obstack_alloc(&obst, size);
 134         memset(res, 0, size);
 135         return res;
 136 }
 137
 138 /**
 139  * Alloc a new instance on obstack and make it equal to @param workset
 140  */
 141 static workset_t *workset_clone(workset_t *workset)
 142 {
 143         workset_t *res;
 144         size_t size = sizeof(*res) + n_regs * sizeof(res->vals[0]);
 145         res = obstack_alloc(&obst, size);
 146         memcpy(res, workset, size);
 147         return res;
 148 }
 149
 150 /**
 151  * Copy workset @param src to @param tgt
 152  */
 153 static void workset_copy(workset_t *dest, const workset_t *src)
 154 {
 155         size_t size = sizeof(*src) + n_regs * sizeof(src->vals[0]);
 156         memcpy(dest, src, size);
 157 }
 158
 159 /**
 160  * Overwrites the current content array of @param ws with the
 161  * @param count locations given at memory @param locs.
 162  * Set the length of @param ws to count.
 163  */
 164 static void workset_bulk_fill(workset_t *workset, int count, const loc_t *locs)
 165 {
 166         workset->len = count;
 167         memcpy(&(workset->vals[0]), locs, count * sizeof(locs[0]));
 168 }
 169
 170 /**
 171  * Inserts the value @p val into the workset, iff it is not
 172  * already contained. The workset must not be full.
 173  */
 174 static void workset_insert(workset_t *workset, ir_node *val, bool spilled)
 175 {
 176         loc_t *loc;
 177         int    i;
 178         /* check for current regclass */
 179         assert(arch_irn_consider_in_reg_alloc(arch_env, cls, val));
 180
 181         /* check if val is already contained */
 182         for (i = 0; i < workset->len; ++i) {
 183                 loc = &workset->vals[i];
 184                 if (loc->node == val) {
 185                         if (spilled) {
 186                                 loc->spilled = true;
 187                         }
 188                         return;
 189                 }
 190         }
 191
 192         /* insert val */
 193         assert(workset->len < n_regs && "Workset already full!");
 194         loc           = &workset->vals[workset->len];
 195         loc->node     = val;
 196         loc->spilled  = spilled;
 197         loc->time     = TIME_UNDEFINED;
 198         workset->len++;
 199 }
 200
 201 /**
 202  * Removes all entries from this workset
 203  */
 204 static void workset_clear(workset_t *workset)
 205 {
 206         workset->len = 0;
 207 }
 208
 209 /**
 210  * Removes the value @p val from the workset if present.
 211  */
 212 static INLINE void workset_remove(workset_t *workset, ir_node *val)
 213 {
 214         int i;
 215         for(i = 0; i < workset->len; ++i) {
 216                 if (workset->vals[i].node == val) {
 217                         workset->vals[i] = workset->vals[--workset->len];
 218                         return;
 219                 }
 220         }
 221 }
 222
 223 static INLINE const loc_t *workset_contains(const workset_t *ws,
 224                                             const ir_node *val)
 225 {
 226         int i;
 227
 228         for (i = 0; i < ws->len; ++i) {
 229                 if (ws->vals[i].node == val)
 230                         return &ws->vals[i];
 231         }
 232
 233         return NULL;
 234 }
 235
 236 /**
 237  * Iterates over all values in the working set.
 238  * @p ws The workset to iterate
 239  * @p v  A variable to put the current value in
 240  * @p i  An integer for internal use
 241  */
 242 #define workset_foreach(ws, v, i)       for(i=0; \
 243                                                                                 v=(i < ws->len) ? ws->vals[i].node : NULL, i < ws->len; \
 244                                                                                 ++i)
 245
 246 #define workset_set_time(ws, i, t) (ws)->vals[i].time=t
 247 #define workset_get_time(ws, i) (ws)->vals[i].time
 248 #define workset_set_length(ws, length) (ws)->len = length
 249 #define workset_get_length(ws) ((ws)->len)
 250 #define workset_get_val(ws, i) ((ws)->vals[i].node)
 251 #define workset_sort(ws) qsort((ws)->vals, (ws)->len, sizeof((ws)->vals[0]), loc_compare);
 252
 253 typedef struct _block_info_t
 254 {
 255         workset_t *start_workset;
 256         workset_t *end_workset;
 257 } block_info_t;
 258
 259
 260 static void *new_block_info(void)
 261 {
 262         block_info_t *res = obstack_alloc(&obst, sizeof(res[0]));
 263         memset(res, 0, sizeof(res[0]));
 264
 265         return res;
 266 }
 267
 268 #define get_block_info(block)        ((block_info_t *)get_irn_link(block))
 269 #define set_block_info(block, info)  set_irn_link(block, info)
 270
 271 /**
 272  * @return The distance to the next use or 0 if irn has dont_spill flag set
 273  */
 274 static INLINE unsigned get_distance(ir_node *from, unsigned from_step,
 275                                     const ir_node *def, int skip_from_uses)
 276 {
 277         be_next_use_t use;
 278         int           flags = arch_irn_get_flags(arch_env, def);
 279         unsigned      costs;
 280         unsigned      time;
 281
 282         assert(! (flags & arch_irn_flags_ignore));
 283
 284         use = be_get_next_use(uses, from, from_step, def, skip_from_uses);
 285         if (USES_IS_INFINITE(use.time))
 286                 return USES_INFINITY;
 287
 288         /* We have to keep nonspillable nodes in the workingset */
 289         if (flags & arch_irn_flags_dont_spill)
 290                 return 0;
 291
 292         costs = be_get_reload_costs_no_weight(senv, def, use.before);
 293         assert(costs * RELOAD_COST_FACTOR < 1000);
 294         time  = use.time + 1000 - (costs * RELOAD_COST_FACTOR);
 295
 296         return time;
 297 }
 298
 299 /**
 300  * Performs the actions necessary to grant the request that:
 301  * - new_vals can be held in registers
 302  * - as few as possible other values are disposed
 303  * - the worst values get disposed
 304  *
 305  * @p is_usage indicates that the values in new_vals are used (not defined)
 306  * In this case reloads must be performed
 307  */
 308 static void displace(workset_t *new_vals, int is_usage)
 309 {
 310         ir_node **to_insert = alloca(n_regs * sizeof(to_insert[0]));
 311         bool     *spilled   = alloca(n_regs * sizeof(spilled[0]));
 312         ir_node  *val;
 313         int       i;
 314         int       len;
 315         int       spills_needed;
 316         int       demand;
 317         int       iter;
 318
 319         /* 1. Identify the number of needed slots and the values to reload */
 320         demand = 0;
 321         workset_foreach(new_vals, val, iter) {
 322                 bool reloaded = false;
 323
 324                 /* mark value as used */
 325                 if (is_usage)
 326                         ir_nodeset_insert(&used, val);
 327
 328                 if (! workset_contains(ws, val)) {
 329                         DB((dbg, DBG_DECIDE, "    insert %+F\n", val));
 330                         if (is_usage) {
 331                                 DB((dbg, DBG_SPILL, "Reload %+F before %+F\n", val, instr));
 332                                 be_add_reload(senv, val, instr, cls, 1);
 333                                 reloaded = true;
 334                         }
 335                 } else {
 336                         DB((dbg, DBG_DECIDE, "    %+F already in workset\n", val));
 337                         assert(is_usage);
 338                         /* remove the value from the current workset so it is not accidently
 339                          * spilled */
 340                         workset_remove(ws, val);
 341                 }
 342                 spilled[demand]   = reloaded;
 343                 to_insert[demand] = val;
 344                 ++demand;
 345         }
 346
 347         /* 2. Make room for at least 'demand' slots */
 348         len           = workset_get_length(ws);
 349         spills_needed = len + demand - n_regs;
 350         assert(spills_needed <= len);
 351
 352         /* Only make more free room if we do not have enough */
 353         if (spills_needed > 0) {
 354 #ifndef PLACE_SPILLS
 355                 ir_node   *curr_bb  = get_nodes_block(instr);
 356                 workset_t *ws_start = get_block_info(curr_bb)->start_workset;
 357 #endif
 358
 359                 DB((dbg, DBG_DECIDE, "    disposing %d values\n", spills_needed));
 360
 361                 /* calculate current next-use distance for live values */
 362                 for (i = 0; i < len; ++i) {
 363                         ir_node  *val  = workset_get_val(ws, i);
 364                         unsigned  dist = get_distance(instr, instr_nr, val, !is_usage);
 365                         workset_set_time(ws, i, dist);
 366                 }
 367
 368                 /* sort entries by increasing nextuse-distance*/
 369                 workset_sort(ws);
 370
 371                 for (i = len - spills_needed; i < len; ++i) {
 372                         ir_node *val = ws->vals[i].node;
 373
 374                         DB((dbg, DBG_DECIDE, "    disposing node %+F (%u)\n", val,
 375                              workset_get_time(ws, i)));
 376
 377 #ifdef PLACE_SPILLS
 378                         if (!USES_IS_INFINITE(ws->vals[i].time) && !ws->vals[i].spilled) {
 379                                 ir_node *after_pos = sched_prev(instr);
 380                                 DB((dbg, DBG_DECIDE, "Spill %+F after node %+F\n", val,
 381                                     after_pos));
 382                                 be_add_spill(senv, val, after_pos);
 383                         }
 384 #endif
 385
 386 #ifndef PLACE_SPILLS
 387                         /* Logic for not needed live-ins: If a value is disposed
 388                          * before its first use, remove it from start workset
 389                          * We don't do this for phis though     */
 390                         if (!is_Phi(val) && ! ir_nodeset_contains(&used, val)) {
 391                                 workset_remove(ws_start, val);
 392                                 DB((dbg, DBG_DECIDE, "    (and removing %+F from start workset)\n", val));
 393                         }
 394 #endif
 395                 }
 396
 397                 /* kill the last 'demand' entries in the array */
 398                 workset_set_length(ws, len - spills_needed);
 399         }
 400
 401         /* 3. Insert the new values into the workset */
 402         for (i = 0; i < demand; ++i) {
 403                 ir_node *val = to_insert[i];
 404
 405                 workset_insert(ws, val, spilled[i]);
 406         }
 407 }
 408
 409 enum {
 410         AVAILABLE_EVERYWHERE,
 411         AVAILABLE_NOWHERE,
 412         AVAILABLE_PARTLY,
 413         AVAILABLE_UNKNOWN
 414 };
 415
 416 static unsigned available_in_all_preds(workset_t* const* pred_worksets,
 417                                        size_t n_pred_worksets,
 418                                        const ir_node *value, bool is_local_phi)
 419 {
 420         size_t i;
 421         bool   avail_everywhere = true;
 422         bool   avail_nowhere    = true;
 423
 424         assert(n_pred_worksets > 0);
 425
 426         /* value available in all preds? */
 427         for (i = 0; i < n_pred_worksets; ++i) {
 428                 bool             found     = false;
 429                 const workset_t *p_workset = pred_worksets[i];
 430                 int              p_len     = workset_get_length(p_workset);
 431                 int              p_i;
 432                 const ir_node   *l_value;
 433
 434                 if (is_local_phi) {
 435                         assert(is_Phi(value));
 436                         l_value = get_irn_n(value, i);
 437                 } else {
 438                         l_value = value;
 439                 }
 440
 441                 for (p_i = 0; p_i < p_len; ++p_i) {
 442                         const loc_t *p_l = &p_workset->vals[p_i];
 443                         if (p_l->node != l_value)
 444                                 continue;
 445
 446                         found = true;
 447                         break;
 448                 }
 449
 450                 if (found) {
 451                         avail_nowhere = false;
 452                 } else {
 453                         avail_everywhere = false;
 454                 }
 455         }
 456
 457         if (avail_everywhere) {
 458                 assert(!avail_nowhere);
 459                 return AVAILABLE_EVERYWHERE;
 460         } else if (avail_nowhere) {
 461                 return AVAILABLE_NOWHERE;
 462         } else {
 463                 return AVAILABLE_PARTLY;
 464         }
 465 }
 466
 467 /** Decides whether a specific node should be in the start workset or not
 468  *
 469  * @param env      belady environment
 470  * @param first
 471  * @param node     the node to test
 472  * @param loop     the loop of the node
 473  */
 474 static loc_t to_take_or_not_to_take(ir_node* first, ir_node *node,
 475                                     ir_loop *loop, unsigned available)
 476 {
 477         be_next_use_t next_use;
 478         loc_t         loc;
 479
 480         loc.time    = USES_INFINITY;
 481         loc.node    = node;
 482         loc.spilled = false;
 483
 484         if (!arch_irn_consider_in_reg_alloc(arch_env, cls, node)) {
 485                 loc.time = USES_INFINITY;
 486                 return loc;
 487         }
 488
 489         /* We have to keep nonspillable nodes in the workingset */
 490         if (arch_irn_get_flags(arch_env, node) & arch_irn_flags_dont_spill) {
 491                 loc.time = 0;
 492                 DB((dbg, DBG_START, "    %+F taken (dontspill node)\n", node, loc.time));
 493                 return loc;
 494         }
 495
 496         next_use = be_get_next_use(uses, first, 0, node, 0);
 497         if (USES_IS_INFINITE(next_use.time)) {
 498                 // the nodes marked as live in shouldn't be dead, so it must be a phi
 499                 assert(is_Phi(node));
 500                 loc.time = USES_INFINITY;
 501                 DB((dbg, DBG_START, "    %+F not taken (dead)\n", node));
 502                 return loc;
 503         }
 504
 505         loc.time = next_use.time;
 506
 507         if (available == AVAILABLE_EVERYWHERE) {
 508                 DB((dbg, DBG_START, "    %+F taken (%u, live in all preds)\n", node,
 509                     loc.time));
 510                 return loc;
 511         } else if(available == AVAILABLE_NOWHERE) {
 512                 DB((dbg, DBG_START, "    %+F not taken (%u, live in no pred)\n", node,
 513                     loc.time));
 514                 loc.time = USES_INFINITY;
 515                 return loc;
 516         }
 517
 518         if (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         return loc;
 527 }
 528
 529 /**
 530  * Computes the start-workset for a block with multiple predecessors. We assume
 531  * that at least 1 of the predeccesors is a back-edge which means we're at the
 532  * beginning of a loop. We try to reload as much values as possible now so they
 533  * don't get reloaded inside the loop.
 534  */
 535 static void decide_start_workset(const ir_node *block)
 536 {
 537         ir_loop    *loop = get_irn_loop(block);
 538         ir_node    *first;
 539         ir_node    *node;
 540         loc_t       loc;
 541         loc_t      *starters;
 542         loc_t      *delayed;
 543         int         i, len, ws_count;
 544         int             free_slots, free_pressure_slots;
 545         unsigned    pressure;
 546         int         arity;
 547         workset_t **pred_worksets;
 548         bool        all_preds_known;
 549
 550         /* check predecessors */
 551         arity           = get_irn_arity(block);
 552         pred_worksets   = alloca(sizeof(pred_worksets[0]) * arity);
 553         all_preds_known = true;
 554         for(i = 0; i < arity; ++i) {
 555                 ir_node      *pred_block = get_Block_cfgpred_block(block, i);
 556                 block_info_t *pred_info  = get_block_info(pred_block);
 557
 558                 if (pred_info == NULL) {
 559                         pred_worksets[i] = NULL;
 560                         all_preds_known  = false;
 561                 } else {
 562                         pred_worksets[i] = pred_info->end_workset;
 563                 }
 564         }
 565
 566         /* Collect all values living at start of block */
 567         starters = NEW_ARR_F(loc_t, 0);
 568         delayed  = NEW_ARR_F(loc_t, 0);
 569
 570         DB((dbg, DBG_START, "Living at start of %+F:\n", block));
 571         first = sched_first(block);
 572
 573         /* check all Phis first */
 574         sched_foreach(block, node) {
 575                 unsigned available;
 576
 577                 if (! is_Phi(node))
 578                         break;
 579                 if (!arch_irn_consider_in_reg_alloc(arch_env, cls, node))
 580                         continue;
 581
 582                 if (all_preds_known) {
 583                         available = available_in_all_preds(pred_worksets, arity, node, true);
 584                 } else {
 585                         available = AVAILABLE_UNKNOWN;
 586                 }
 587
 588                 loc = to_take_or_not_to_take(first, node, loop, available);
 589
 590                 if (! USES_IS_INFINITE(loc.time)) {
 591                         if (USES_IS_PENDING(loc.time))
 592                                 ARR_APP1(loc_t, delayed, loc);
 593                         else
 594                                 ARR_APP1(loc_t, starters, loc);
 595                 } else {
 596                         be_spill_phi(senv, node);
 597                 }
 598         }
 599
 600         /* check all Live-Ins */
 601         be_lv_foreach(lv, block, be_lv_state_in, i) {
 602                 ir_node *node = be_lv_get_irn(lv, block, i);
 603                 unsigned available;
 604
 605                 if (all_preds_known) {
 606                         available = available_in_all_preds(pred_worksets, arity, node, false);
 607                 } else {
 608                         available = AVAILABLE_UNKNOWN;
 609                 }
 610
 611                 loc = to_take_or_not_to_take(first, node, loop, available);
 612
 613                 if (! USES_IS_INFINITE(loc.time)) {
 614                         if (USES_IS_PENDING(loc.time))
 615                                 ARR_APP1(loc_t, delayed, loc);
 616                         else
 617                                 ARR_APP1(loc_t, starters, loc);
 618                 }
 619         }
 620
 621         pressure            = be_get_loop_pressure(loop_ana, cls, loop);
 622         assert(ARR_LEN(delayed) <= (signed)pressure);
 623         free_slots          = n_regs - ARR_LEN(starters);
 624         free_pressure_slots = n_regs - (pressure - ARR_LEN(delayed));
 625         free_slots          = MIN(free_slots, free_pressure_slots);
 626
 627         /* so far we only put nodes into the starters list that are used inside
 628          * the loop. If register pressure in the loop is low then we can take some
 629          * values and let them live through the loop */
 630         if (free_slots > 0) {
 631                 qsort(delayed, ARR_LEN(delayed), sizeof(delayed[0]), loc_compare);
 632
 633                 for (i = 0; i < ARR_LEN(delayed) && i < free_slots; ++i) {
 634                         int    p, arity;
 635                         loc_t *loc = & delayed[i];
 636
 637                         /* don't use values which are dead in a known predecessors
 638                          * to not induce unnecessary reloads */
 639                         arity = get_irn_arity(block);
 640                         for (p = 0; p < arity; ++p) {
 641                                 ir_node      *pred_block = get_Block_cfgpred_block(block, p);
 642                                 block_info_t *pred_info  = get_block_info(pred_block);
 643
 644                                 if (pred_info == NULL)
 645                                         continue;
 646
 647                                 if (!workset_contains(pred_info->end_workset, loc->node)) {
 648                                         DB((dbg, DBG_START,
 649                                             "    delayed %+F not live at pred %+F\n", loc->node,
 650                                             pred_block));
 651                                         goto skip_delayed;
 652                                 }
 653                         }
 654
 655                         DB((dbg, DBG_START, "    delayed %+F taken\n", loc->node));
 656                         ARR_APP1(loc_t, starters, *loc);
 657                         loc->node = NULL;
 658                 skip_delayed:
 659                         ;
 660                 }
 661         }
 662
 663         /* spill phis (the actual phis not just their values) that are in this block
 664          * but not in the start workset */
 665         for (i = ARR_LEN(delayed) - 1; i >= 0; --i) {
 666                 ir_node *node = delayed[i].node;
 667                 if (node == NULL || !is_Phi(node) || get_nodes_block(node) != block)
 668                         continue;
 669
 670                 DB((dbg, DBG_START, "    spilling delayed phi %+F\n", node));
 671                 be_spill_phi(senv, node);
 672         }
 673         DEL_ARR_F(delayed);
 674
 675         /* Sort start values by first use */
 676         qsort(starters, ARR_LEN(starters), sizeof(starters[0]), loc_compare);
 677
 678         /* Copy the best ones from starters to start workset */
 679         ws_count = MIN(ARR_LEN(starters), n_regs);
 680         workset_clear(ws);
 681         workset_bulk_fill(ws, ws_count, starters);
 682
 683         /* spill phis (the actual phis not just their values) that are in this block
 684          * but not in the start workset */
 685         len = ARR_LEN(starters);
 686         for (i = ws_count; i < len; ++i) {
 687                 ir_node *node = starters[i].node;
 688                 if (! is_Phi(node) || get_nodes_block(node) != block)
 689                         continue;
 690
 691                 DB((dbg, DBG_START, "    spilling phi %+F\n", node));
 692                 be_spill_phi(senv, node);
 693         }
 694
 695         DEL_ARR_F(starters);
 696
 697         /* determine spill status of the values: If there's 1 pred block (which
 698          * is no backedge) where the value is spilled then we must set it to
 699          * spilled here. */
 700         for(i = 0; i < ws_count; ++i) {
 701                 loc_t   *loc     = &ws->vals[i];
 702                 ir_node *value   = loc->node;
 703                 bool     spilled;
 704                 int      n;
 705
 706                 /* phis from this block aren't spilled */
 707                 if (get_nodes_block(value) == block) {
 708                         assert(is_Phi(value));
 709                         loc->spilled = false;
 710                         continue;
 711                 }
 712
 713                 /* determine if value was spilled on any predecessor */
 714                 spilled = false;
 715                 for(n = 0; n < arity; ++n) {
 716                         workset_t *pred_workset = pred_worksets[n];
 717                         int        p_len;
 718                         int        p;
 719
 720                         if (pred_workset == NULL)
 721                                 continue;
 722
 723                         p_len = workset_get_length(pred_workset);
 724                         for(p = 0; p < p_len; ++p) {
 725                                 loc_t *l = &pred_workset->vals[p];
 726
 727                                 if (l->node != value)
 728                                         continue;
 729
 730                                 if (l->spilled) {
 731                                         spilled = true;
 732                                 }
 733                                 break;
 734                         }
 735                 }
 736
 737                 loc->spilled = spilled;
 738         }
 739 }
 740
 741 #if 0
 742 static void decide_start_workset2(const ir_node *block)
 743 {
 744         int         arity;
 745         workset_t **pred_worksets;
 746         int         p;
 747         int         i;
 748         int         len;
 749
 750         /* check if all predecessors are known */
 751         arity           = get_irn_arity(block);
 752         pred_worksets   = alloca(sizeof(pred_worksets[0]) * arity);
 753         for (i = 0; i < arity; ++i) {
 754                 ir_node      *pred_block = get_Block_cfgpred_block(block, i);
 755                 block_info_t *pred_info  = get_block_info(pred_block);
 756
 757                 if (pred_info == NULL) {
 758                         /* not all predecessors known, use decide_start_workset */
 759                         decide_start_workset(block);
 760                         return;
 761                 }
 762
 763                 pred_worksets[i] = pred_info->end_workset;
 764         }
 765
 766         /* we construct a new workset */
 767         workset_clear(ws);
 768
 769         /* take values live in all pred blocks */
 770         len = workset_get_length(pred_worksets[0]);
 771         for (p = 0; p < len; ++p) {
 772                 const loc_t *l = &pred_worksets[0]->vals[p];
 773                 ir_node     *value;
 774                 bool         spilled = false;
 775
 776                 if (USES_IS_INFINITE(l->time))
 777                         continue;
 778
 779                 /* value available in all preds? */
 780                 value = l->node;
 781                 for (i = 1; i < arity; ++i) {
 782                         bool       found     = false;
 783                         workset_t *p_workset = pred_worksets[i];
 784                         int        p_len     = workset_get_length(p_workset);
 785                         int        p_i;
 786
 787                         for (p_i = 0; p_i < p_len; ++p_i) {
 788                                 const loc_t *p_l = &p_workset->vals[p_i];
 789                                 if (p_l->node != value)
 790                                         continue;
 791
 792                                 found = true;
 793                                 if (p_l->spilled)
 794                                         spilled = true;
 795                                 break;
 796                         }
 797
 798                         if (!found)
 799                                 break;
 800                 }
 801
 802                 /* it was available in all preds */
 803                 if (i >= arity) {
 804                         workset_insert(ws, value, spilled);
 805                 }
 806         }
 807
 808         /* Copy the best ones from starters to start workset */
 809         ws_count = MIN(ARR_LEN(starters), n_regs);
 810         workset_bulk_fill(ws, ws_count, starters);
 811 }
 812 #endif
 813
 814 /**
 815  * For the given block @p block, decide for each values
 816  * whether it is used from a register or is reloaded
 817  * before the use.
 818  */
 819 static void belady(ir_node *block)
 820 {
 821         workset_t       *new_vals;
 822         ir_node         *irn;
 823         int              iter;
 824         block_info_t    *block_info;
 825         int              i, arity;
 826         int              has_backedges = 0;
 827         //int              first         = 0;
 828         const ir_edge_t *edge;
 829
 830         /* no need to process a block twice */
 831         if (get_block_info(block) != NULL) {
 832                 return;
 833         }
 834
 835         /* check if all predecessor blocks are processed yet (though for backedges
 836          * we have to make an exception as we can't process them first) */
 837         arity = get_Block_n_cfgpreds(block);
 838         for(i = 0; i < arity; ++i) {
 839                 ir_node      *pred_block = get_Block_cfgpred_block(block, i);
 840                 block_info_t *pred_info  = get_block_info(pred_block);
 841
 842                 if (pred_info == NULL) {
 843                         /* process predecessor first (it will be in the queue already) */
 844                         if (!is_backedge(block, i)) {
 845                                 return;
 846                         }
 847                         has_backedges = 1;
 848                 }
 849         }
 850         (void) has_backedges;
 851         if (arity == 0) {
 852                 workset_clear(ws);
 853         } else if (arity == 1) {
 854                 ir_node      *pred_block = get_Block_cfgpred_block(block, 0);
 855                 block_info_t *pred_info  = get_block_info(pred_block);
 856
 857                 assert(pred_info != NULL);
 858                 workset_copy(ws, pred_info->end_workset);
 859         } else {
 860                 /* we need 2 heuristics here, for the case when all predecessor blocks
 861                  * are known and when some are backedges (and therefore can't be known
 862                  * yet) */
 863                 decide_start_workset(block);
 864         }
 865
 866         DB((dbg, DBG_DECIDE, "\n"));
 867         DB((dbg, DBG_DECIDE, "Decide for %+F\n", block));
 868
 869         block_info = new_block_info();
 870         set_block_info(block, block_info);
 871
 872         DB((dbg, DBG_WSETS, "Start workset for %+F:\n", block));
 873         workset_foreach(ws, irn, iter) {
 874                 DB((dbg, DBG_WSETS, "  %+F (%u)\n", irn,
 875                      workset_get_time(ws, iter)));
 876         }
 877
 878         block_info->start_workset = workset_clone(ws);
 879
 880         /* process the block from start to end */
 881         DB((dbg, DBG_WSETS, "Processing...\n"));
 882         ir_nodeset_init(&used);
 883         instr_nr = 0;
 884         /* TODO: this leaks (into the obstack)... */
 885         new_vals = new_workset();
 886
 887         sched_foreach(block, irn) {
 888                 int i, arity;
 889                 assert(workset_get_length(ws) <= n_regs);
 890
 891                 /* Phis are no real instr (see insert_starters()) */
 892                 if (is_Phi(irn)) {
 893                         continue;
 894                 }
 895                 DB((dbg, DBG_DECIDE, "  ...%+F\n", irn));
 896
 897                 /* set instruction in the workset */
 898                 instr = irn;
 899
 900                 /* allocate all values _used_ by this instruction */
 901                 workset_clear(new_vals);
 902                 for(i = 0, arity = get_irn_arity(irn); i < arity; ++i) {
 903                         ir_node *in = get_irn_n(irn, i);
 904                         if (!arch_irn_consider_in_reg_alloc(arch_env, cls, in))
 905                                 continue;
 906
 907                         /* (note that "spilled" is irrelevant here) */
 908                         workset_insert(new_vals, in, false);
 909                 }
 910                 displace(new_vals, 1);
 911
 912                 /* allocate all values _defined_ by this instruction */
 913                 workset_clear(new_vals);
 914                 if (get_irn_mode(irn) == mode_T) {
 915                         const ir_edge_t *edge;
 916
 917                         foreach_out_edge(irn, edge) {
 918                                 ir_node *proj = get_edge_src_irn(edge);
 919                                 if (!arch_irn_consider_in_reg_alloc(arch_env, cls, proj))
 920                                         continue;
 921                                 workset_insert(new_vals, proj, false);
 922                         }
 923                 } else {
 924                         if (!arch_irn_consider_in_reg_alloc(arch_env, cls, irn))
 925                                 continue;
 926                         workset_insert(new_vals, irn, false);
 927                 }
 928                 displace(new_vals, 0);
 929
 930                 instr_nr++;
 931         }
 932         ir_nodeset_destroy(&used);
 933
 934         /* Remember end-workset for this block */
 935         block_info->end_workset = workset_clone(ws);
 936         DB((dbg, DBG_WSETS, "End workset for %+F:\n", block));
 937         workset_foreach(ws, irn, iter)
 938                 DB((dbg, DBG_WSETS, "  %+F (%u)\n", irn,
 939                      workset_get_time(ws, iter)));
 940
 941         /* add successor blocks into worklist */
 942         foreach_block_succ(block, edge) {
 943                 ir_node *succ = get_edge_src_irn(edge);
 944                 pdeq_putr(worklist, succ);
 945         }
 946 }
 947
 948 /**
 949  * 'decide' is block-local and makes assumptions
 950  * about the set of live-ins. Thus we must adapt the
 951  * live-outs to the live-ins at each block-border.
 952  */
 953 static void fix_block_borders(ir_node *block, void *data)
 954 {
 955         workset_t    *start_workset;
 956         int           arity;
 957         int           i;
 958         int           iter;
 959         (void) data;
 960
 961         DB((dbg, DBG_FIX, "\n"));
 962         DB((dbg, DBG_FIX, "Fixing %+F\n", block));
 963
 964         start_workset = get_block_info(block)->start_workset;
 965
 966         /* process all pred blocks */
 967         arity = get_irn_arity(block);
 968         for (i = 0; i < arity; ++i) {
 969                 ir_node   *pred = get_Block_cfgpred_block(block, i);
 970                 workset_t *pred_end_workset = get_block_info(pred)->end_workset;
 971                 ir_node   *node;
 972
 973                 DB((dbg, DBG_FIX, "  Pred %+F\n", pred));
 974
 975                 /* spill all values not used anymore */
 976                 workset_foreach(pred_end_workset, node, iter) {
 977                         ir_node *n2;
 978                         int      iter2;
 979                         bool     found = false;
 980                         workset_foreach(start_workset, n2, iter2) {
 981                                 if (n2 == node) {
 982                                         found = true;
 983                                         break;
 984                                 }
 985                                 /* note that we do not look at phi inputs, becuase the values
 986                                  * will be either live-end and need no spill or
 987                                  * they have other users in which must be somewhere else in the
 988                                  * workset */
 989                         }
 990
 991                         if (found)
 992                                 continue;
 993
 994 #ifdef PLACE_SPILLS
 995                         if (be_is_live_in(lv, block, node)
 996                                         && !pred_end_workset->vals[iter].spilled) {
 997                                 ir_node *insert_point;
 998                                 if (arity > 1) {
 999                                         insert_point = be_get_end_of_block_insertion_point(pred);
1000                                         insert_point = sched_prev(insert_point);
1001                                 } else {
1002                                         insert_point = block;
1003                                 }
1004                                 DB((dbg, DBG_SPILL, "Spill %+F after %+F\n", node,
1005                                      insert_point));
1006                                 be_add_spill(senv, node, insert_point);
1007                         }
1008 #endif
1009                 }
1010
1011                 /* reload missing values in predecessors, add missing spills */
1012                 workset_foreach(start_workset, node, iter) {
1013                         const loc_t *l    = &start_workset->vals[iter];
1014                         const loc_t *pred_loc;
1015
1016                         /* if node is a phi of the current block we reload
1017                          * the corresponding argument, else node itself */
1018                         if (is_Phi(node) && get_nodes_block(node) == block) {
1019                                 node = get_irn_n(node, i);
1020                                 assert(!l->spilled);
1021
1022                                 /* we might have unknowns as argument for the phi */
1023                                 if (!arch_irn_consider_in_reg_alloc(arch_env, cls, node))
1024                                         continue;
1025                         }
1026
1027                         /* check if node is in a register at end of pred */
1028                         pred_loc = workset_contains(pred_end_workset, node);
1029                         if (pred_loc != NULL) {
1030 #ifdef PLACE_SPILLS
1031                                 /* we might have to spill value on this path */
1032                                 if (!pred_loc->spilled && l->spilled) {
1033                                         ir_node *insert_point
1034                                                 = be_get_end_of_block_insertion_point(pred);
1035                                         insert_point = sched_prev(insert_point);
1036                                         DB((dbg, DBG_SPILL, "Spill %+F after %+F\n", node,
1037                                             insert_point));
1038                                         be_add_spill(senv, node, insert_point);
1039                                 }
1040 #endif
1041                         } else {
1042                                 /* node is not in register at the end of pred -> reload it */
1043                                 DB((dbg, DBG_FIX, "    reload %+F\n", node));
1044                                 DB((dbg, DBG_SPILL, "Reload %+F before %+F,%d\n", node, block, i));
1045                                 be_add_reload_on_edge(senv, node, block, i, cls, 1);
1046                         }
1047                 }
1048         }
1049 }
1050
1051 static void be_spill_belady(be_irg_t *birg, const arch_register_class_t *rcls)
1052 {
1053         ir_graph *irg = be_get_birg_irg(birg);
1054
1055         be_liveness_assure_sets(be_assure_liveness(birg));
1056
1057         /* construct control flow loop tree */
1058         if (! (get_irg_loopinfo_state(irg) & loopinfo_cf_consistent)) {
1059                 construct_cf_backedges(irg);
1060         }
1061
1062         be_clear_links(irg);
1063
1064         /* init belady env */
1065         obstack_init(&obst);
1066         arch_env = birg->main_env->arch_env;
1067         cls      = rcls;
1068         lv       = be_get_birg_liveness(birg);
1069         n_regs   = cls->n_regs - be_put_ignore_regs(birg, cls, NULL);
1070         ws       = new_workset();
1071         uses     = be_begin_uses(irg, lv);
1072         loop_ana = be_new_loop_pressure(birg);
1073         senv     = be_new_spill_env(birg);
1074         worklist = new_pdeq();
1075
1076         pdeq_putr(worklist, get_irg_start_block(irg));
1077
1078         while(!pdeq_empty(worklist)) {
1079                 ir_node *block = pdeq_getl(worklist);
1080                 belady(block);
1081         }
1082         /* end block might not be reachable in endless loops */
1083         belady(get_irg_end_block(irg));
1084
1085         del_pdeq(worklist);
1086
1087         /* belady was block-local, fix the global flow by adding reloads on the
1088          * edges */
1089         irg_block_walk_graph(irg, fix_block_borders, NULL, NULL);
1090
1091         /* Insert spill/reload nodes into the graph and fix usages */
1092         be_insert_spills_reloads(senv);
1093
1094         /* clean up */
1095         be_delete_spill_env(senv);
1096         be_end_uses(uses);
1097         be_free_loop_pressure(loop_ana);
1098         obstack_free(&obst, NULL);
1099 }
1100
1101 void be_init_spillbelady(void)
1102 {
1103         static be_spiller_t belady_spiller = {
1104                 be_spill_belady
1105         };
1106
1107         be_register_spiller("belady", &belady_spiller);
1108         FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
1109 }
1110
1111 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady);