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 "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 #include "xmalloc.h"
  43 #include "pdeq.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 /* factor to weight the different costs of reloading/rematerializing a node
  70    (see bespill.h be_get_reload_costs_no_weight) */
  71 #define RELOAD_COST_FACTOR   10
  72
  73 typedef enum {
  74         value_not_reloaded,       /* the value has not been reloaded */
  75         value_partially_reloaded, /* the value has been reloaded on some paths */
  76         value_reloaded            /* the value has been reloaded on all paths */
  77 } reloaded_state_t;
  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         reloaded_state_t  reloaded; /**< the value is a reloaded value */
  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, int reloaded)
 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(!loc->reloaded) {
 186                                 loc->reloaded = reloaded;
 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->reloaded = reloaded;
 197         loc->time     = 6666; /* undefined yet */
 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 int workset_contains(const workset_t *ws, const ir_node *val)
 224 {
 225         int i;
 226
 227         for(i=0; i<ws->len; ++i) {
 228                 if (ws->vals[i].node == val)
 229                         return 1;
 230         }
 231
 232         return 0;
 233 }
 234
 235 /**
 236  * Iterates over all values in the working set.
 237  * @p ws The workset to iterate
 238  * @p v  A variable to put the current value in
 239  * @p i  An integer for internal use
 240  */
 241 #define workset_foreach(ws, v, i)       for(i=0; \
 242                                                                                 v=(i < ws->len) ? ws->vals[i].node : NULL, i < ws->len; \
 243                                                                                 ++i)
 244
 245 #define workset_set_time(ws, i, t) (ws)->vals[i].time=t
 246 #define workset_get_time(ws, i) (ws)->vals[i].time
 247 #define workset_set_length(ws, length) (ws)->len = length
 248 #define workset_get_length(ws) ((ws)->len)
 249 #define workset_get_val(ws, i) ((ws)->vals[i].node)
 250 #define workset_sort(ws) qsort((ws)->vals, (ws)->len, sizeof((ws)->vals[0]), loc_compare);
 251
 252 typedef struct _block_info_t
 253 {
 254         workset_t *start_workset;
 255         workset_t *end_workset;
 256 } block_info_t;
 257
 258
 259 static void *new_block_info(void)
 260 {
 261         block_info_t *res = obstack_alloc(&obst, sizeof(res[0]));
 262         memset(res, 0, sizeof(res[0]));
 263
 264         return res;
 265 }
 266
 267 #define get_block_info(block)        ((block_info_t *)get_irn_link(block))
 268 #define set_block_info(block, info)  set_irn_link(block, info)
 269
 270 /**
 271  * @return The distance to the next use or 0 if irn has dont_spill flag set
 272  */
 273 static INLINE unsigned get_distance(ir_node *from, unsigned from_step,
 274                                     const ir_node *def, int skip_from_uses)
 275 {
 276         be_next_use_t use;
 277         int           flags = arch_irn_get_flags(arch_env, def);
 278         unsigned      costs;
 279         unsigned      time;
 280
 281         assert(! (flags & arch_irn_flags_ignore));
 282
 283         use = be_get_next_use(uses, from, from_step, def, skip_from_uses);
 284         if(USES_IS_INFINITE(use.time))
 285                 return USES_INFINITY;
 286
 287         /* We have to keep nonspillable nodes in the workingset */
 288         if(flags & arch_irn_flags_dont_spill)
 289                 return 0;
 290
 291         costs = be_get_reload_costs_no_weight(senv, def, use.before);
 292         assert(costs * RELOAD_COST_FACTOR < 1000);
 293         time  = use.time + 1000 - (costs * RELOAD_COST_FACTOR);
 294
 295         return time;
 296 }
 297
 298 /**
 299  * Performs the actions necessary to grant the request that:
 300  * - new_vals can be held in registers
 301  * - as few as possible other values are disposed
 302  * - the worst values get disposed
 303  *
 304  * @p is_usage indicates that the values in new_vals are used (not defined)
 305  * In this case reloads must be performed
 306  */
 307 static void displace(workset_t *new_vals, int is_usage)
 308 {
 309         ir_node **to_insert = alloca(n_regs * sizeof(to_insert[0]));
 310         ir_node  *val;
 311         int       i;
 312         int       len;
 313         int       spills_needed;
 314         int       demand;
 315         int       iter;
 316
 317         /* 1. Identify the number of needed slots and the values to reload */
 318         demand = 0;
 319         workset_foreach(new_vals, val, iter) {
 320                 /* mark value as used */
 321                 if (is_usage)
 322                         ir_nodeset_insert(&used, val);
 323
 324                 if (! workset_contains(ws, val)) {
 325                         DB((dbg, DBG_DECIDE, "    insert %+F\n", val));
 326                         if (is_usage) {
 327                                 DB((dbg, DBG_SPILL, "Reload %+F before %+F\n", val, instr));
 328                                 be_add_reload(senv, val, instr, cls, 1);
 329                         }
 330                 } else {
 331                         DB((dbg, DBG_DECIDE, "    %+F already in workset\n", val));
 332                         assert(is_usage);
 333                         /* remove the value from the current workset so it is not accidently
 334                          * spilled */
 335                         workset_remove(ws, val);
 336                 }
 337                 to_insert[demand++] = val;
 338         }
 339
 340         /* 2. Make room for at least 'demand' slots */
 341         len           = workset_get_length(ws);
 342         spills_needed = len + demand - n_regs;
 343         assert(spills_needed <= len);
 344
 345         /* Only make more free room if we do not have enough */
 346         if (spills_needed > 0) {
 347                 ir_node   *curr_bb  = get_nodes_block(instr);
 348                 workset_t *ws_start = get_block_info(curr_bb)->start_workset;
 349
 350                 DB((dbg, DBG_DECIDE, "    disposing %d values\n", spills_needed));
 351
 352                 /* calculate current next-use distance for live values */
 353                 for (i = 0; i < len; ++i) {
 354                         ir_node  *val  = workset_get_val(ws, i);
 355                         unsigned  dist = get_distance(instr, instr_nr, val, !is_usage);
 356                         workset_set_time(ws, i, dist);
 357                 }
 358
 359                 /* sort entries by increasing nextuse-distance*/
 360                 workset_sort(ws);
 361
 362                 /* Logic for not needed live-ins: If a value is disposed
 363                  * before its first usage, remove it from start workset
 364                  * We don't do this for phis though     */
 365                 for (i = len - spills_needed; i < len; ++i) {
 366                         ir_node *val = ws->vals[i].node;
 367
 368                         DB((dbg, DBG_DECIDE, "    disposing node %+F (%u)\n", val,
 369                              workset_get_time(ws, i)));
 370
 371                         if(!USES_IS_INFINITE(ws->vals[i].time)
 372                                         && !ws->vals[i].reloaded) {
 373                                 //be_add_spill(senv, val, instr);
 374                         }
 375
 376                         if (!is_Phi(val) && ! ir_nodeset_contains(&used, val)) {
 377                                 workset_remove(ws_start, val);
 378                                 DB((dbg, DBG_DECIDE, "    (and removing %+F from start workset)\n", val));
 379                         }
 380                 }
 381
 382                 /* kill the last 'demand' entries in the array */
 383                 workset_set_length(ws, len - spills_needed);
 384         }
 385
 386         /* 3. Insert the new values into the workset */
 387         for (i = 0; i < demand; ++i) {
 388                 ir_node *val = to_insert[i];
 389
 390                 workset_insert(ws, val, 1);
 391         }
 392 }
 393
 394 /** Decides whether a specific node should be in the start workset or not
 395  *
 396  * @param env      belady environment
 397  * @param first
 398  * @param node     the node to test
 399  * @param loop     the loop of the node
 400  */
 401 static loc_t to_take_or_not_to_take(ir_node* first, ir_node *node,
 402                                     ir_loop *loop)
 403 {
 404         be_next_use_t next_use;
 405         loc_t         loc;
 406
 407         loc.time     = USES_INFINITY;
 408         loc.node     = node;
 409         //loc.reloaded = rand() % 2; /* provoke a bug... */
 410         loc.reloaded = 0;
 411
 412         if (!arch_irn_consider_in_reg_alloc(arch_env, cls, node)) {
 413                 loc.time = USES_INFINITY;
 414                 return loc;
 415         }
 416
 417         /* We have to keep nonspillable nodes in the workingset */
 418         if(arch_irn_get_flags(arch_env, node) & arch_irn_flags_dont_spill) {
 419                 loc.time = 0;
 420                 DB((dbg, DBG_START, "    %+F taken (dontspill node)\n", node, loc.time));
 421                 return loc;
 422         }
 423
 424         next_use = be_get_next_use(uses, first, 0, node, 0);
 425         if(USES_IS_INFINITE(next_use.time)) {
 426                 // the nodes marked as live in shouldn't be dead, so it must be a phi
 427                 assert(is_Phi(node));
 428                 loc.time = USES_INFINITY;
 429                 DB((dbg, DBG_START, "    %+F not taken (dead)\n", node));
 430                 if(is_Phi(node)) {
 431                         be_spill_phi(senv, node);
 432                 }
 433                 return loc;
 434         }
 435
 436         loc.time = next_use.time;
 437
 438         if(next_use.outermost_loop >= get_loop_depth(loop)) {
 439                 DB((dbg, DBG_START, "    %+F taken (%u, loop %d)\n", node, loc.time,
 440                     next_use.outermost_loop));
 441         } else {
 442                 loc.time = USES_PENDING;
 443                 DB((dbg, DBG_START, "    %+F delayed (outerdepth %d < loopdepth %d)\n",
 444                     node, next_use.outermost_loop, get_loop_depth(loop)));
 445         }
 446         return loc;
 447 }
 448
 449 /**
 450  * Computes the start-workset for a block with multiple predecessors. We assume
 451  * that at least 1 of the predeccesors is a back-edge which means we're at the
 452  * beginning of a loop. We try to reload as much values as possible now so they
 453  * don't get reloaded inside the loop.
 454  */
 455 static void compute_live_ins(const ir_node *block)
 456 {
 457         ir_loop    *loop = get_irn_loop(block);
 458         ir_node    *first;
 459         ir_node    *node;
 460         loc_t       loc;
 461         loc_t      *starters;
 462         loc_t      *delayed;
 463         int         i, len, ws_count;
 464         int             free_slots, free_pressure_slots;
 465         unsigned    pressure;
 466         //int arity;
 467         //int         n_pred_worksets;
 468         //workset_t **pred_worksets;
 469
 470         /* Collect all values living at start of block */
 471         starters = NEW_ARR_F(loc_t, 0);
 472         delayed  = NEW_ARR_F(loc_t, 0);
 473
 474         DB((dbg, DBG_START, "Living at start of %+F:\n", block));
 475         first = sched_first(block);
 476
 477         /* check all Phis first */
 478         sched_foreach(block, node) {
 479                 if (! is_Phi(node))
 480                         break;
 481
 482                 loc = to_take_or_not_to_take(first, node, loop);
 483
 484                 if (! USES_IS_INFINITE(loc.time)) {
 485                         if (USES_IS_PENDING(loc.time))
 486                                 ARR_APP1(loc_t, delayed, loc);
 487                         else
 488                                 ARR_APP1(loc_t, starters, loc);
 489                 }
 490         }
 491
 492         /* check all Live-Ins */
 493         be_lv_foreach(lv, block, be_lv_state_in, i) {
 494                 ir_node *node = be_lv_get_irn(lv, block, i);
 495
 496                 loc = to_take_or_not_to_take(first, node, loop);
 497
 498                 if (! USES_IS_INFINITE(loc.time)) {
 499                         if (USES_IS_PENDING(loc.time))
 500                                 ARR_APP1(loc_t, delayed, loc);
 501                         else
 502                                 ARR_APP1(loc_t, starters, loc);
 503                 }
 504         }
 505
 506         pressure            = be_get_loop_pressure(loop_ana, cls, loop);
 507         assert(ARR_LEN(delayed) <= (signed)pressure);
 508         free_slots          = n_regs - ARR_LEN(starters);
 509         free_pressure_slots = n_regs - (pressure - ARR_LEN(delayed));
 510         free_slots          = MIN(free_slots, free_pressure_slots);
 511
 512         /* so far we only put nodes into the starters list that are used inside
 513          * the loop. If register pressure in the loop is low then we can take some
 514          * values and let them live through the loop */
 515         if (free_slots > 0) {
 516                 qsort(delayed, ARR_LEN(delayed), sizeof(delayed[0]), loc_compare);
 517
 518                 for (i = 0; i < ARR_LEN(delayed) && i < free_slots; ++i) {
 519                         int    p, arity;
 520                         loc_t *loc = & delayed[i];
 521
 522                         /* don't use values which are dead in a known predecessors
 523                          * to not induce unnecessary reloads */
 524                         arity = get_irn_arity(block);
 525                         for (p = 0; p < arity; ++p) {
 526                                 ir_node      *pred_block = get_Block_cfgpred_block(block, p);
 527                                 block_info_t *pred_info  = get_block_info(pred_block);
 528
 529                                 if (pred_info == NULL)
 530                                         continue;
 531
 532                                 if (!workset_contains(pred_info->end_workset, loc->node)) {
 533                                         DB((dbg, DBG_START,
 534                                             "    delayed %+F not live at pred %+F\n", loc->node,
 535                                             pred_block));
 536                                         goto skip_delayed;
 537                                 }
 538                         }
 539
 540                         DB((dbg, DBG_START, "    delayed %+F taken\n", loc->node));
 541                         ARR_APP1(loc_t, starters, *loc);
 542                         loc->node = NULL;
 543                 skip_delayed:
 544                         ;
 545                 }
 546         }
 547
 548         /* spill phis (the actual phis not just their values) that are in this block
 549          * but not in the start workset */
 550         for (i = ARR_LEN(delayed) - 1; i >= 0; --i) {
 551                 ir_node *node = delayed[i].node;
 552                 if(node == NULL || !is_Phi(node) || get_nodes_block(node) != block)
 553                         continue;
 554
 555                 DB((dbg, DBG_START, "    spilling delayed phi %+F\n", node));
 556                 be_spill_phi(senv, node);
 557         }
 558         DEL_ARR_F(delayed);
 559
 560         /* Sort start values by first use */
 561         qsort(starters, ARR_LEN(starters), sizeof(starters[0]), loc_compare);
 562
 563         /* Copy the best ones from starters to start workset */
 564         ws_count = MIN(ARR_LEN(starters), n_regs);
 565         workset_clear(ws);
 566         workset_bulk_fill(ws, ws_count, starters);
 567
 568         /* spill phis (the actual phis not just their values) that are in this block
 569          * but not in the start workset */
 570         len = ARR_LEN(starters);
 571         for (i = ws_count; i < len; ++i) {
 572                 ir_node *node = starters[i].node;
 573                 if (! is_Phi(node) || get_nodes_block(node) != block)
 574                         continue;
 575
 576                 DB((dbg, DBG_START, "    spilling phi %+F\n", node));
 577                 be_spill_phi(senv, node);
 578         }
 579
 580         DEL_ARR_F(starters);
 581
 582 #if 0
 583         /* determine reloaded status of the values: If there's 1 pred block (which
 584          * is no backedge) where the value is reloaded then we must set it to
 585          * reloaded here. We place spills in all pred where the value was not yet
 586          * reloaded to be sure we have a spill on each path */
 587         n_pred_worksets = 0;
 588         arity           = get_irn_arity(block);
 589         pred_worksets   = alloca(sizeof(pred_worksets[0]) * arity);
 590         for(i = 0; i < arity; ++i) {
 591                 ir_node      *pred_block = get_Block_cfgpred_block(block, i);
 592                 block_info_t *pred_info  = get_block_info(pred_block);
 593                 if(pred_info == NULL)
 594                         continue;
 595
 596                 pred_worksets[n_pred_worksets] = pred_info->end_workset;
 597                 ++n_pred_worksets;
 598         }
 599
 600         for(i = 0; i < ws_count; ++i) {
 601                 loc_t   *loc   = &ws->vals[i];
 602                 ir_node *value = loc->node;
 603                 int      reloaded;
 604                 int      n;
 605
 606                 /* phis from this block aren't reloaded */
 607                 if(get_nodes_block(value) == block) {
 608                         assert(is_Phi(value));
 609                         loc->reloaded = value_not_reloaded;
 610                         continue;
 611                 }
 612
 613                 /* was the value reloaded on any of the other inputs */
 614                 reloaded = 0;
 615                 arity    = get_Block_n_cfgpreds(block);
 616                 for(n = 0; n < n_pred_worksets; ++n) {
 617                         workset_t *pred_workset = pred_worksets[n];
 618                         int        p_len        = workset_get_length(pred_workset);
 619                         int        p;
 620
 621                         for(p = 0; p < p_len; ++p) {
 622                                 loc_t *l = &pred_workset->vals[p];
 623                                 if(l->node == value) {
 624                                         if(l->reloaded) {
 625                                                 reloaded = 1;
 626                                         }
 627                                         break;
 628                                 }
 629                         }
 630                         if(p >= p_len) {
 631                                 reloaded = 1;
 632                                 break;
 633                         }
 634                 }
 635         }
 636 #endif
 637 }
 638
 639 /**
 640  * For the given block @p block, decide for each values
 641  * whether it is used from a register or is reloaded
 642  * before the use.
 643  */
 644 static void belady(ir_node *block)
 645 {
 646         workset_t       *new_vals;
 647         ir_node         *irn;
 648         int              iter;
 649         block_info_t    *block_info;
 650         int              i, arity;
 651         int              has_backedges = 0;
 652         //int              first         = 0;
 653         const ir_edge_t *edge;
 654
 655         /* no need to process a block twice */
 656         if(get_block_info(block) != NULL) {
 657                 return;
 658         }
 659
 660         /* check if all predecessor blocks are processed yet (though for backedges
 661          * we have to make an exception as we can't process them first) */
 662         arity = get_Block_n_cfgpreds(block);
 663         for(i = 0; i < arity; ++i) {
 664                 ir_node      *pred_block = get_Block_cfgpred_block(block, i);
 665                 block_info_t *pred_info  = get_block_info(pred_block);
 666
 667                 if(pred_info == NULL) {
 668                         /* process predecessor first (it will be in the queue already) */
 669                         if(!is_backedge(block, i)) {
 670                                 return;
 671                         }
 672                         has_backedges = 1;
 673                 }
 674         }
 675         (void) has_backedges;
 676         if(arity == 0) {
 677                 workset_clear(ws);
 678         } else if(arity == 1) {
 679                 ir_node      *pred_block = get_Block_cfgpred_block(block, 0);
 680                 block_info_t *pred_info  = get_block_info(pred_block);
 681
 682                 assert(pred_info != NULL);
 683                 workset_copy(ws, pred_info->end_workset);
 684         } else {
 685                 /* we need 2 heuristics here, for the case when all predecessor blocks
 686                  * are known and when some are backedges (and therefore can't be known
 687                  * yet) */
 688                 compute_live_ins(block);
 689         }
 690
 691         DB((dbg, DBG_DECIDE, "\n"));
 692         DB((dbg, DBG_DECIDE, "Decide for %+F\n", block));
 693
 694         block_info = new_block_info();
 695         set_block_info(block, block_info);
 696
 697         DB((dbg, DBG_WSETS, "Start workset for %+F:\n", block));
 698         workset_foreach(ws, irn, iter) {
 699                 DB((dbg, DBG_WSETS, "  %+F (%u)\n", irn,
 700                      workset_get_time(ws, iter)));
 701         }
 702
 703         block_info->start_workset = workset_clone(ws);
 704
 705         /* process the block from start to end */
 706         DB((dbg, DBG_WSETS, "Processing...\n"));
 707         ir_nodeset_init(&used);
 708         instr_nr = 0;
 709         /* TODO: this leaks (into the obstack)... */
 710         new_vals = new_workset();
 711
 712         sched_foreach(block, irn) {
 713                 int i, arity;
 714                 assert(workset_get_length(ws) <= n_regs);
 715
 716                 /* Phis are no real instr (see insert_starters()) */
 717                 if (is_Phi(irn)) {
 718                         continue;
 719                 }
 720                 DB((dbg, DBG_DECIDE, "  ...%+F\n", irn));
 721
 722                 /* set instruction in the workset */
 723                 instr = irn;
 724
 725                 /* allocate all values _used_ by this instruction */
 726                 workset_clear(new_vals);
 727                 for(i = 0, arity = get_irn_arity(irn); i < arity; ++i) {
 728                         ir_node *in = get_irn_n(irn, i);
 729                         if (!arch_irn_consider_in_reg_alloc(arch_env, cls, in))
 730                                 continue;
 731
 732                         /* (note that reloaded_value is irrelevant here) */
 733                         workset_insert(new_vals, in, 0);
 734                 }
 735                 displace(new_vals, 1);
 736
 737                 /* allocate all values _defined_ by this instruction */
 738                 workset_clear(new_vals);
 739                 if (get_irn_mode(irn) == mode_T) {
 740                         const ir_edge_t *edge;
 741
 742                         foreach_out_edge(irn, edge) {
 743                                 ir_node *proj = get_edge_src_irn(edge);
 744                                 if (!arch_irn_consider_in_reg_alloc(arch_env, cls, proj))
 745                                         continue;
 746                                 workset_insert(new_vals, proj, 0);
 747                         }
 748                 } else {
 749                         if (!arch_irn_consider_in_reg_alloc(arch_env, cls, irn))
 750                                 continue;
 751                         workset_insert(new_vals, irn, 0);
 752                 }
 753                 displace(new_vals, 0);
 754
 755                 instr_nr++;
 756         }
 757         ir_nodeset_destroy(&used);
 758
 759         /* Remember end-workset for this block */
 760         block_info->end_workset = workset_clone(ws);
 761         DB((dbg, DBG_WSETS, "End workset for %+F:\n", block));
 762         workset_foreach(ws, irn, iter)
 763                 DB((dbg, DBG_WSETS, "  %+F (%u)\n", irn,
 764                      workset_get_time(ws, iter)));
 765
 766         /* add successor blocks into worklist */
 767         foreach_block_succ(block, edge) {
 768                 ir_node *succ = get_edge_src_irn(edge);
 769                 pdeq_putr(worklist, succ);
 770         }
 771 }
 772
 773 /**
 774  * 'decide' is block-local and makes assumptions
 775  * about the set of live-ins. Thus we must adapt the
 776  * live-outs to the live-ins at each block-border.
 777  */
 778 static void fix_block_borders(ir_node *block, void *data)
 779 {
 780         workset_t    *start_workset;
 781         int           arity;
 782         int           i;
 783         int           iter;
 784         (void) data;
 785
 786         DB((dbg, DBG_FIX, "\n"));
 787         DB((dbg, DBG_FIX, "Fixing %+F\n", block));
 788
 789         start_workset = get_block_info(block)->start_workset;
 790
 791         /* process all pred blocks */
 792         arity = get_irn_arity(block);
 793         for (i = 0; i < arity; ++i) {
 794                 ir_node   *pred = get_Block_cfgpred_block(block, i);
 795                 workset_t *pred_end_workset = get_block_info(pred)->end_workset;
 796                 ir_node   *node;
 797
 798                 DB((dbg, DBG_FIX, "  Pred %+F\n", pred));
 799
 800                 /* spill all values not used anymore */
 801                 workset_foreach(pred_end_workset, node, iter) {
 802                         ir_node *n2;
 803                         int      iter2;
 804                         int      found = 0;
 805                         workset_foreach(start_workset, n2, iter2) {
 806                                 if(n2 == node) {
 807                                         found = 1;
 808                                         break;
 809                                 }
 810                                 /* note that we do not look at phi inputs, becuase the values
 811                                  * will be either live-end and need no spill or
 812                                  * they have other users in which must be somewhere else in the
 813                                  * workset */
 814                         }
 815
 816 #if 0
 817                         if(!found && be_is_live_out(lv, pred, node)
 818                                         && !pred_end_workset->vals[iter].reloaded) {
 819                                 ir_node *insert_point
 820                                         = be_get_end_of_block_insertion_point(pred);
 821                                 DB((dbg, DBG_SPILL, "Spill %+F before %+F\n", node,
 822                                      insert_point));
 823                                 be_add_spill(senv, node, insert_point);
 824                         }
 825 #endif
 826                 }
 827
 828                 /* reload missing values in predecessors */
 829                 workset_foreach(start_workset, node, iter) {
 830                         /* if node is a phi of the current block we reload
 831                          * the corresponding argument, else node itself */
 832                         if(is_Phi(node) && block == get_nodes_block(node)) {
 833                                 node = get_irn_n(node, i);
 834
 835                                 /* we might have unknowns as argument for the phi */
 836                                 if(!arch_irn_consider_in_reg_alloc(arch_env, cls, node))
 837                                         continue;
 838                         }
 839
 840                         /* check if node is in a register at end of pred */
 841                         if(workset_contains(pred_end_workset, node))
 842                                 continue;
 843
 844                         /* node is not in memory at the end of pred -> reload it */
 845                         DB((dbg, DBG_FIX, "    reload %+F\n", node));
 846                         DB((dbg, DBG_SPILL, "Reload %+F before %+F,%d\n", node, block, i));
 847                         be_add_reload_on_edge(senv, node, block, i, cls, 1);
 848                 }
 849         }
 850 }
 851
 852 static void be_spill_belady(be_irg_t *birg, const arch_register_class_t *rcls)
 853 {
 854         ir_graph *irg = be_get_birg_irg(birg);
 855
 856         be_liveness_assure_sets(be_assure_liveness(birg));
 857
 858         /* construct control flow loop tree */
 859         if(! (get_irg_loopinfo_state(irg) & loopinfo_cf_consistent)) {
 860                 construct_cf_backedges(irg);
 861         }
 862
 863         be_clear_links(irg);
 864
 865         /* init belady env */
 866         obstack_init(&obst);
 867         arch_env = birg->main_env->arch_env;
 868         cls      = rcls;
 869         lv       = be_get_birg_liveness(birg);
 870         n_regs   = cls->n_regs - be_put_ignore_regs(birg, cls, NULL);
 871         ws       = new_workset();
 872         uses     = be_begin_uses(irg, lv);
 873         loop_ana = be_new_loop_pressure(birg);
 874         senv     = be_new_spill_env(birg);
 875         worklist = new_pdeq();
 876
 877         pdeq_putr(worklist, get_irg_start_block(irg));
 878
 879         while(!pdeq_empty(worklist)) {
 880                 ir_node *block = pdeq_getl(worklist);
 881                 belady(block);
 882         }
 883         /* end block might not be reachable in endless loops */
 884         belady(get_irg_end_block(irg));
 885
 886         del_pdeq(worklist);
 887
 888         /* belady was block-local, fix the global flow by adding reloads on the
 889          * edges */
 890         irg_block_walk_graph(irg, fix_block_borders, NULL, NULL);
 891
 892         /* Insert spill/reload nodes into the graph and fix usages */
 893         be_insert_spills_reloads(senv);
 894
 895         /* clean up */
 896         be_delete_spill_env(senv);
 897         be_end_uses(uses);
 898         be_free_loop_pressure(loop_ana);
 899         obstack_free(&obst, NULL);
 900 }
 901
 902 void be_init_spillbelady(void)
 903 {
 904         static be_spiller_t belady_spiller = {
 905                 be_spill_belady
 906         };
 907
 908         be_register_spiller("belady", &belady_spiller);
 909         FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
 910 }
 911
 912 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady);