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 #define TIME_UNDEFINED 6666
  72
  73 //#define LOOK_AT_LOOPDEPTH
  74
  75 /**
  76  * An association between a node and a point in time.
  77  */
  78 typedef struct loc_t {
  79         ir_node          *node;
  80         unsigned          time;     /**< A use time (see beuses.h). */
  81         bool              spilled;  /**< the value was already spilled on this path */
  82 } loc_t;
  83
  84 typedef struct _workset_t {
  85         int   len;          /**< current length */
  86         loc_t vals[0];      /**< inlined array of the values/distances in this working set */
  87 } workset_t;
  88
  89 static struct obstack               obst;
  90 static const arch_env_t            *arch_env;
  91 static const arch_register_class_t *cls;
  92 static const be_lv_t               *lv;
  93 static be_loopana_t                *loop_ana;
  94 static int                          n_regs;
  95 static workset_t                   *ws;     /**< the main workset used while
  96                                                      processing a block. */
  97 static be_uses_t                   *uses;   /**< env for the next-use magic */
  98 static ir_node                     *instr;  /**< current instruction */
  99 static unsigned                     instr_nr; /**< current instruction number
 100                                                        (relative to block start) */
 101 static ir_nodeset_t                 used;
 102 static spill_env_t                 *senv;   /**< see bespill.h */
 103 static pdeq                        *worklist;
 104
 105 static bool                         move_spills      = true;
 106 static bool                         respectloopdepth = true;
 107 static bool                         improve_known_preds = true;
 108 /* factor to weight the different costs of reloading/rematerializing a node
 109    (see bespill.h be_get_reload_costs_no_weight) */
 110 static int                          remat_bonus      = 10;
 111
 112 static const lc_opt_table_entry_t options[] = {
 113         LC_OPT_ENT_BOOL   ("movespills", "try to move spills out of loops", &move_spills),
 114         LC_OPT_ENT_BOOL   ("respectloopdepth", "exprimental (outermost loop cutting)", &respectloopdepth),
 115         LC_OPT_ENT_BOOL   ("improveknownpreds", "experimental (known preds cutting)", &improve_known_preds),
 116         LC_OPT_ENT_INT    ("rematbonus", "give bonus to rematerialisable nodes", &remat_bonus),
 117         LC_OPT_LAST
 118 };
 119
 120 static int loc_compare(const void *a, const void *b)
 121 {
 122         const loc_t *p = a;
 123         const loc_t *q = b;
 124         return p->time - q->time;
 125 }
 126
 127 void workset_print(const workset_t *w)
 128 {
 129         int i;
 130
 131         for(i = 0; i < w->len; ++i) {
 132                 ir_fprintf(stderr, "%+F %d\n", w->vals[i].node, w->vals[i].time);
 133         }
 134 }
 135
 136 /**
 137  * Alloc a new workset on obstack @p ob with maximum size @p max
 138  */
 139 static workset_t *new_workset(void)
 140 {
 141         workset_t *res;
 142         size_t     size = sizeof(*res) + n_regs * sizeof(res->vals[0]);
 143
 144         res  = obstack_alloc(&obst, size);
 145         memset(res, 0, size);
 146         return res;
 147 }
 148
 149 /**
 150  * Alloc a new instance on obstack and make it equal to @param workset
 151  */
 152 static workset_t *workset_clone(workset_t *workset)
 153 {
 154         workset_t *res;
 155         size_t size = sizeof(*res) + n_regs * sizeof(res->vals[0]);
 156         res = obstack_alloc(&obst, size);
 157         memcpy(res, workset, size);
 158         return res;
 159 }
 160
 161 /**
 162  * Copy workset @param src to @param tgt
 163  */
 164 static void workset_copy(workset_t *dest, const workset_t *src)
 165 {
 166         size_t size = sizeof(*src) + n_regs * sizeof(src->vals[0]);
 167         memcpy(dest, src, size);
 168 }
 169
 170 /**
 171  * Overwrites the current content array of @param ws with the
 172  * @param count locations given at memory @param locs.
 173  * Set the length of @param ws to count.
 174  */
 175 static void workset_bulk_fill(workset_t *workset, int count, const loc_t *locs)
 176 {
 177         workset->len = count;
 178         memcpy(&(workset->vals[0]), locs, count * sizeof(locs[0]));
 179 }
 180
 181 /**
 182  * Inserts the value @p val into the workset, iff it is not
 183  * already contained. The workset must not be full.
 184  */
 185 static void workset_insert(workset_t *workset, ir_node *val, bool spilled)
 186 {
 187         loc_t *loc;
 188         int    i;
 189         /* check for current regclass */
 190         assert(arch_irn_consider_in_reg_alloc(arch_env, cls, val));
 191
 192         /* check if val is already contained */
 193         for (i = 0; i < workset->len; ++i) {
 194                 loc = &workset->vals[i];
 195                 if (loc->node == val) {
 196                         if (spilled) {
 197                                 loc->spilled = true;
 198                         }
 199                         return;
 200                 }
 201         }
 202
 203         /* insert val */
 204         assert(workset->len < n_regs && "Workset already full!");
 205         loc           = &workset->vals[workset->len];
 206         loc->node     = val;
 207         loc->spilled  = spilled;
 208         loc->time     = TIME_UNDEFINED;
 209         workset->len++;
 210 }
 211
 212 /**
 213  * Removes all entries from this workset
 214  */
 215 static void workset_clear(workset_t *workset)
 216 {
 217         workset->len = 0;
 218 }
 219
 220 /**
 221  * Removes the value @p val from the workset if present.
 222  */
 223 static INLINE void workset_remove(workset_t *workset, ir_node *val)
 224 {
 225         int i;
 226         for(i = 0; i < workset->len; ++i) {
 227                 if (workset->vals[i].node == val) {
 228                         workset->vals[i] = workset->vals[--workset->len];
 229                         return;
 230                 }
 231         }
 232 }
 233
 234 static INLINE const loc_t *workset_contains(const workset_t *ws,
 235                                             const ir_node *val)
 236 {
 237         int i;
 238
 239         for (i = 0; i < ws->len; ++i) {
 240                 if (ws->vals[i].node == val)
 241                         return &ws->vals[i];
 242         }
 243
 244         return NULL;
 245 }
 246
 247 /**
 248  * Iterates over all values in the working set.
 249  * @p ws The workset to iterate
 250  * @p v  A variable to put the current value in
 251  * @p i  An integer for internal use
 252  */
 253 #define workset_foreach(ws, v, i)       for(i=0; \
 254                                                                                 v=(i < ws->len) ? ws->vals[i].node : NULL, i < ws->len; \
 255                                                                                 ++i)
 256
 257 #define workset_set_time(ws, i, t) (ws)->vals[i].time=t
 258 #define workset_get_time(ws, i) (ws)->vals[i].time
 259 #define workset_set_length(ws, length) (ws)->len = length
 260 #define workset_get_length(ws) ((ws)->len)
 261 #define workset_get_val(ws, i) ((ws)->vals[i].node)
 262 #define workset_sort(ws) qsort((ws)->vals, (ws)->len, sizeof((ws)->vals[0]), loc_compare);
 263
 264 typedef struct _block_info_t
 265 {
 266         workset_t *start_workset;
 267         workset_t *end_workset;
 268 } block_info_t;
 269
 270
 271 static void *new_block_info(void)
 272 {
 273         block_info_t *res = obstack_alloc(&obst, sizeof(res[0]));
 274         memset(res, 0, sizeof(res[0]));
 275
 276         return res;
 277 }
 278
 279 #define get_block_info(block)        ((block_info_t *)get_irn_link(block))
 280 #define set_block_info(block, info)  set_irn_link(block, info)
 281
 282 /**
 283  * @return The distance to the next use or 0 if irn has dont_spill flag set
 284  */
 285 static INLINE unsigned get_distance(ir_node *from, unsigned from_step,
 286                                     const ir_node *def, int skip_from_uses)
 287 {
 288         be_next_use_t use;
 289         int           flags = arch_irn_get_flags(arch_env, def);
 290         unsigned      costs;
 291         unsigned      time;
 292
 293         assert(! (flags & arch_irn_flags_ignore));
 294
 295         use = be_get_next_use(uses, from, from_step, def, skip_from_uses);
 296         if (USES_IS_INFINITE(use.time))
 297                 return USES_INFINITY;
 298
 299         /* We have to keep nonspillable nodes in the workingset */
 300         if (flags & arch_irn_flags_dont_spill)
 301                 return 0;
 302
 303         /* give some bonus to rematerialisable nodes */
 304         if (remat_bonus > 0) {
 305                 costs = be_get_reload_costs_no_weight(senv, def, use.before);
 306                 assert(costs * remat_bonus < 1000);
 307                 time  = use.time + 1000 - (costs * remat_bonus);
 308         }
 309
 310         return time;
 311 }
 312
 313 /**
 314  * Performs the actions necessary to grant the request that:
 315  * - new_vals can be held in registers
 316  * - as few as possible other values are disposed
 317  * - the worst values get disposed
 318  *
 319  * @p is_usage indicates that the values in new_vals are used (not defined)
 320  * In this case reloads must be performed
 321  */
 322 static void displace(workset_t *new_vals, int is_usage)
 323 {
 324         ir_node **to_insert = alloca(n_regs * sizeof(to_insert[0]));
 325         bool     *spilled   = alloca(n_regs * sizeof(spilled[0]));
 326         ir_node  *val;
 327         int       i;
 328         int       len;
 329         int       spills_needed;
 330         int       demand;
 331         int       iter;
 332
 333         /* 1. Identify the number of needed slots and the values to reload */
 334         demand = 0;
 335         workset_foreach(new_vals, val, iter) {
 336                 bool reloaded = false;
 337
 338                 /* mark value as used */
 339                 if (is_usage)
 340                         ir_nodeset_insert(&used, val);
 341
 342                 if (! workset_contains(ws, val)) {
 343                         DB((dbg, DBG_DECIDE, "    insert %+F\n", val));
 344                         if (is_usage) {
 345                                 DB((dbg, DBG_SPILL, "Reload %+F before %+F\n", val, instr));
 346                                 be_add_reload(senv, val, instr, cls, 1);
 347                                 reloaded = true;
 348                         }
 349                 } else {
 350                         DB((dbg, DBG_DECIDE, "    %+F already in workset\n", val));
 351                         assert(is_usage);
 352                         /* remove the value from the current workset so it is not accidently
 353                          * spilled */
 354                         workset_remove(ws, val);
 355                 }
 356                 spilled[demand]   = reloaded;
 357                 to_insert[demand] = val;
 358                 ++demand;
 359         }
 360
 361         /* 2. Make room for at least 'demand' slots */
 362         len           = workset_get_length(ws);
 363         spills_needed = len + demand - n_regs;
 364         assert(spills_needed <= len);
 365
 366         /* Only make more free room if we do not have enough */
 367         if (spills_needed > 0) {
 368                 ir_node   *curr_bb  = NULL;
 369                 workset_t *ws_start = NULL;
 370
 371                 if (move_spills) {
 372                         curr_bb  = get_nodes_block(instr);
 373                         ws_start = get_block_info(curr_bb)->start_workset;
 374                 }
 375
 376                 DB((dbg, DBG_DECIDE, "    disposing %d values\n", spills_needed));
 377
 378                 /* calculate current next-use distance for live values */
 379                 for (i = 0; i < len; ++i) {
 380                         ir_node  *val  = workset_get_val(ws, i);
 381                         unsigned  dist = get_distance(instr, instr_nr, val, !is_usage);
 382                         workset_set_time(ws, i, dist);
 383                 }
 384
 385                 /* sort entries by increasing nextuse-distance*/
 386                 workset_sort(ws);
 387
 388                 for (i = len - spills_needed; i < len; ++i) {
 389                         ir_node *val = ws->vals[i].node;
 390
 391                         DB((dbg, DBG_DECIDE, "    disposing node %+F (%u)\n", val,
 392                              workset_get_time(ws, i)));
 393
 394                         if (move_spills) {
 395                                 if (!USES_IS_INFINITE(ws->vals[i].time)
 396                                                 && !ws->vals[i].spilled) {
 397                                         ir_node *after_pos = sched_prev(instr);
 398                                         DB((dbg, DBG_DECIDE, "Spill %+F after node %+F\n", val,
 399                                                 after_pos));
 400                                         be_add_spill(senv, val, after_pos);
 401                                 }
 402                         } else {
 403                                 /* Logic for not needed live-ins: If a value is disposed
 404                                  * before its first use, remove it from start workset
 405                                  * We don't do this for phis though     */
 406                                 if (!is_Phi(val) && ! ir_nodeset_contains(&used, val)) {
 407                                         workset_remove(ws_start, val);
 408                                         DB((dbg, DBG_DECIDE, "    (and removing %+F from start workset)\n", val));
 409                                 }
 410                         }
 411                 }
 412
 413                 /* kill the last 'demand' entries in the array */
 414                 workset_set_length(ws, len - spills_needed);
 415         }
 416
 417         /* 3. Insert the new values into the workset */
 418         for (i = 0; i < demand; ++i) {
 419                 ir_node *val = to_insert[i];
 420
 421                 workset_insert(ws, val, spilled[i]);
 422         }
 423 }
 424
 425 enum {
 426         AVAILABLE_EVERYWHERE,
 427         AVAILABLE_NOWHERE,
 428         AVAILABLE_PARTLY,
 429         AVAILABLE_UNKNOWN
 430 };
 431
 432 static unsigned available_in_all_preds(workset_t* const* pred_worksets,
 433                                        size_t n_pred_worksets,
 434                                        const ir_node *value, bool is_local_phi)
 435 {
 436         size_t i;
 437         bool   avail_everywhere = true;
 438         bool   avail_nowhere    = true;
 439
 440         assert(n_pred_worksets > 0);
 441
 442         /* value available in all preds? */
 443         for (i = 0; i < n_pred_worksets; ++i) {
 444                 bool             found     = false;
 445                 const workset_t *p_workset = pred_worksets[i];
 446                 int              p_len     = workset_get_length(p_workset);
 447                 int              p_i;
 448                 const ir_node   *l_value;
 449
 450                 if (is_local_phi) {
 451                         assert(is_Phi(value));
 452                         l_value = get_irn_n(value, i);
 453                 } else {
 454                         l_value = value;
 455                 }
 456
 457                 for (p_i = 0; p_i < p_len; ++p_i) {
 458                         const loc_t *p_l = &p_workset->vals[p_i];
 459                         if (p_l->node != l_value)
 460                                 continue;
 461
 462                         found = true;
 463                         break;
 464                 }
 465
 466                 if (found) {
 467                         avail_nowhere = false;
 468                 } else {
 469                         avail_everywhere = false;
 470                 }
 471         }
 472
 473         if (avail_everywhere) {
 474                 assert(!avail_nowhere);
 475                 return AVAILABLE_EVERYWHERE;
 476         } else if (avail_nowhere) {
 477                 return AVAILABLE_NOWHERE;
 478         } else {
 479                 return AVAILABLE_PARTLY;
 480         }
 481 }
 482
 483 /** Decides whether a specific node should be in the start workset or not
 484  *
 485  * @param env      belady environment
 486  * @param first
 487  * @param node     the node to test
 488  * @param loop     the loop of the node
 489  */
 490 static loc_t to_take_or_not_to_take(ir_node* first, ir_node *node,
 491                                     ir_loop *loop, unsigned available)
 492 {
 493         be_next_use_t next_use;
 494         loc_t         loc;
 495
 496         loc.time    = USES_INFINITY;
 497         loc.node    = node;
 498         loc.spilled = false;
 499
 500         if (!arch_irn_consider_in_reg_alloc(arch_env, cls, node)) {
 501                 loc.time = USES_INFINITY;
 502                 return loc;
 503         }
 504
 505         /* We have to keep nonspillable nodes in the workingset */
 506         if (arch_irn_get_flags(arch_env, node) & arch_irn_flags_dont_spill) {
 507                 loc.time = 0;
 508                 DB((dbg, DBG_START, "    %+F taken (dontspill node)\n", node, loc.time));
 509                 return loc;
 510         }
 511
 512         next_use = be_get_next_use(uses, first, 0, node, 0);
 513         if (USES_IS_INFINITE(next_use.time)) {
 514                 // the nodes marked as live in shouldn't be dead, so it must be a phi
 515                 assert(is_Phi(node));
 516                 loc.time = USES_INFINITY;
 517                 DB((dbg, DBG_START, "    %+F not taken (dead)\n", node));
 518                 return loc;
 519         }
 520
 521         loc.time = next_use.time;
 522
 523         if (improve_known_preds) {
 524                 if (available == AVAILABLE_EVERYWHERE) {
 525                         DB((dbg, DBG_START, "    %+F taken (%u, live in all preds)\n",
 526                             node, loc.time));
 527                         return loc;
 528                 } else if(available == AVAILABLE_NOWHERE) {
 529                         DB((dbg, DBG_START, "    %+F not taken (%u, live in no pred)\n",
 530                             node, loc.time));
 531                         loc.time = USES_INFINITY;
 532                         return loc;
 533                 }
 534         }
 535
 536         if (!respectloopdepth || next_use.outermost_loop >= get_loop_depth(loop)) {
 537                 DB((dbg, DBG_START, "    %+F taken (%u, loop %d)\n", node, loc.time,
 538                     next_use.outermost_loop));
 539         } else {
 540                 loc.time = USES_PENDING;
 541                 DB((dbg, DBG_START, "    %+F delayed (outerdepth %d < loopdepth %d)\n",
 542                     node, next_use.outermost_loop, get_loop_depth(loop)));
 543         }
 544
 545         return loc;
 546 }
 547
 548 /**
 549  * Computes the start-workset for a block with multiple predecessors. We assume
 550  * that at least 1 of the predeccesors is a back-edge which means we're at the
 551  * beginning of a loop. We try to reload as much values as possible now so they
 552  * don't get reloaded inside the loop.
 553  */
 554 static void decide_start_workset(const ir_node *block)
 555 {
 556         ir_loop    *loop = get_irn_loop(block);
 557         ir_node    *first;
 558         ir_node    *node;
 559         loc_t       loc;
 560         loc_t      *starters;
 561         loc_t      *delayed;
 562         int         i, len, ws_count;
 563         int             free_slots, free_pressure_slots;
 564         unsigned    pressure;
 565         int         arity;
 566         workset_t **pred_worksets;
 567         bool        all_preds_known;
 568
 569         /* check predecessors */
 570         arity           = get_irn_arity(block);
 571         pred_worksets   = alloca(sizeof(pred_worksets[0]) * arity);
 572         all_preds_known = true;
 573         for(i = 0; i < arity; ++i) {
 574                 ir_node      *pred_block = get_Block_cfgpred_block(block, i);
 575                 block_info_t *pred_info  = get_block_info(pred_block);
 576
 577                 if (pred_info == NULL) {
 578                         pred_worksets[i] = NULL;
 579                         all_preds_known  = false;
 580                 } else {
 581                         pred_worksets[i] = pred_info->end_workset;
 582                 }
 583         }
 584
 585         /* Collect all values living at start of block */
 586         starters = NEW_ARR_F(loc_t, 0);
 587         delayed  = NEW_ARR_F(loc_t, 0);
 588
 589         DB((dbg, DBG_START, "Living at start of %+F:\n", block));
 590         first = sched_first(block);
 591
 592         /* check all Phis first */
 593         sched_foreach(block, node) {
 594                 unsigned available;
 595
 596                 if (! is_Phi(node))
 597                         break;
 598                 if (!arch_irn_consider_in_reg_alloc(arch_env, cls, node))
 599                         continue;
 600
 601                 if (all_preds_known) {
 602                         available = available_in_all_preds(pred_worksets, arity, node, true);
 603                 } else {
 604                         available = AVAILABLE_UNKNOWN;
 605                 }
 606
 607                 loc = to_take_or_not_to_take(first, node, loop, available);
 608
 609                 if (! USES_IS_INFINITE(loc.time)) {
 610                         if (USES_IS_PENDING(loc.time))
 611                                 ARR_APP1(loc_t, delayed, loc);
 612                         else
 613                                 ARR_APP1(loc_t, starters, loc);
 614                 } else {
 615                         be_spill_phi(senv, node);
 616                 }
 617         }
 618
 619         /* check all Live-Ins */
 620         be_lv_foreach(lv, block, be_lv_state_in, i) {
 621                 ir_node *node = be_lv_get_irn(lv, block, i);
 622                 unsigned available;
 623
 624                 if (all_preds_known) {
 625                         available = available_in_all_preds(pred_worksets, arity, node, false);
 626                 } else {
 627                         available = AVAILABLE_UNKNOWN;
 628                 }
 629
 630                 loc = to_take_or_not_to_take(first, node, loop, available);
 631
 632                 if (! USES_IS_INFINITE(loc.time)) {
 633                         if (USES_IS_PENDING(loc.time))
 634                                 ARR_APP1(loc_t, delayed, loc);
 635                         else
 636                                 ARR_APP1(loc_t, starters, loc);
 637                 }
 638         }
 639
 640         pressure            = be_get_loop_pressure(loop_ana, cls, loop);
 641         assert(ARR_LEN(delayed) <= (signed)pressure);
 642         free_slots          = n_regs - ARR_LEN(starters);
 643         free_pressure_slots = n_regs - (pressure - ARR_LEN(delayed));
 644         free_slots          = MIN(free_slots, free_pressure_slots);
 645
 646         /* so far we only put nodes into the starters list that are used inside
 647          * the loop. If register pressure in the loop is low then we can take some
 648          * values and let them live through the loop */
 649         DB((dbg, DBG_START, "Loop pressure %d, taking %d delayed vals\n",
 650             pressure, free_slots));
 651         if (free_slots > 0) {
 652                 qsort(delayed, ARR_LEN(delayed), sizeof(delayed[0]), loc_compare);
 653
 654                 for (i = 0; i < ARR_LEN(delayed) && free_slots > 0; ++i) {
 655                         int    p, arity;
 656                         loc_t *loc = & delayed[i];
 657
 658                         /* don't use values which are dead in a known predecessors
 659                          * to not induce unnecessary reloads */
 660                         arity = get_irn_arity(block);
 661                         for (p = 0; p < arity; ++p) {
 662                                 ir_node      *pred_block = get_Block_cfgpred_block(block, p);
 663                                 block_info_t *pred_info  = get_block_info(pred_block);
 664
 665                                 if (pred_info == NULL)
 666                                         continue;
 667
 668                                 if (!workset_contains(pred_info->end_workset, loc->node)) {
 669                                         DB((dbg, DBG_START,
 670                                             "    delayed %+F not live at pred %+F\n", loc->node,
 671                                             pred_block));
 672                                         goto skip_delayed;
 673                                 }
 674                         }
 675
 676                         DB((dbg, DBG_START, "    delayed %+F taken\n", loc->node));
 677                         ARR_APP1(loc_t, starters, *loc);
 678                         loc->node = NULL;
 679                         --free_slots;
 680                 skip_delayed:
 681                         ;
 682                 }
 683         }
 684
 685         /* spill phis (the actual phis not just their values) that are in this block
 686          * but not in the start workset */
 687         for (i = ARR_LEN(delayed) - 1; i >= 0; --i) {
 688                 ir_node *node = delayed[i].node;
 689                 if (node == NULL || !is_Phi(node) || get_nodes_block(node) != block)
 690                         continue;
 691
 692                 DB((dbg, DBG_START, "    spilling delayed phi %+F\n", node));
 693                 be_spill_phi(senv, node);
 694         }
 695         DEL_ARR_F(delayed);
 696
 697         /* Sort start values by first use */
 698         qsort(starters, ARR_LEN(starters), sizeof(starters[0]), loc_compare);
 699
 700         /* Copy the best ones from starters to start workset */
 701         ws_count = MIN(ARR_LEN(starters), n_regs);
 702         workset_clear(ws);
 703         workset_bulk_fill(ws, ws_count, starters);
 704
 705         /* spill phis (the actual phis not just their values) that are in this block
 706          * but not in the start workset */
 707         len = ARR_LEN(starters);
 708         for (i = ws_count; i < len; ++i) {
 709                 ir_node *node = starters[i].node;
 710                 if (! is_Phi(node) || get_nodes_block(node) != block)
 711                         continue;
 712
 713                 DB((dbg, DBG_START, "    spilling phi %+F\n", node));
 714                 be_spill_phi(senv, node);
 715         }
 716
 717         DEL_ARR_F(starters);
 718
 719         /* determine spill status of the values: If there's 1 pred block (which
 720          * is no backedge) where the value is spilled then we must set it to
 721          * spilled here. */
 722         for(i = 0; i < ws_count; ++i) {
 723                 loc_t   *loc     = &ws->vals[i];
 724                 ir_node *value   = loc->node;
 725                 bool     spilled;
 726                 int      n;
 727
 728                 /* phis from this block aren't spilled */
 729                 if (get_nodes_block(value) == block) {
 730                         assert(is_Phi(value));
 731                         loc->spilled = false;
 732                         continue;
 733                 }
 734
 735                 /* determine if value was spilled on any predecessor */
 736                 spilled = false;
 737                 for(n = 0; n < arity; ++n) {
 738                         workset_t *pred_workset = pred_worksets[n];
 739                         int        p_len;
 740                         int        p;
 741
 742                         if (pred_workset == NULL)
 743                                 continue;
 744
 745                         p_len = workset_get_length(pred_workset);
 746                         for(p = 0; p < p_len; ++p) {
 747                                 loc_t *l = &pred_workset->vals[p];
 748
 749                                 if (l->node != value)
 750                                         continue;
 751
 752                                 if (l->spilled) {
 753                                         spilled = true;
 754                                 }
 755                                 break;
 756                         }
 757                 }
 758
 759                 loc->spilled = spilled;
 760         }
 761 }
 762
 763 /**
 764  * For the given block @p block, decide for each values
 765  * whether it is used from a register or is reloaded
 766  * before the use.
 767  */
 768 static void belady(ir_node *block)
 769 {
 770         workset_t       *new_vals;
 771         ir_node         *irn;
 772         int              iter;
 773         block_info_t    *block_info;
 774         int              i, arity;
 775         int              has_backedges = 0;
 776         //int              first         = 0;
 777         const ir_edge_t *edge;
 778
 779         /* no need to process a block twice */
 780         if (get_block_info(block) != NULL) {
 781                 return;
 782         }
 783
 784         /* check if all predecessor blocks are processed yet (though for backedges
 785          * we have to make an exception as we can't process them first) */
 786         arity = get_Block_n_cfgpreds(block);
 787         for(i = 0; i < arity; ++i) {
 788                 ir_node      *pred_block = get_Block_cfgpred_block(block, i);
 789                 block_info_t *pred_info  = get_block_info(pred_block);
 790
 791                 if (pred_info == NULL) {
 792                         /* process predecessor first (it will be in the queue already) */
 793                         if (!is_backedge(block, i)) {
 794                                 return;
 795                         }
 796                         has_backedges = 1;
 797                 }
 798         }
 799         (void) has_backedges;
 800         if (arity == 0) {
 801                 workset_clear(ws);
 802         } else if (arity == 1) {
 803                 ir_node      *pred_block = get_Block_cfgpred_block(block, 0);
 804                 block_info_t *pred_info  = get_block_info(pred_block);
 805
 806                 assert(pred_info != NULL);
 807                 workset_copy(ws, pred_info->end_workset);
 808         } else {
 809                 /* we need 2 heuristics here, for the case when all predecessor blocks
 810                  * are known and when some are backedges (and therefore can't be known
 811                  * yet) */
 812                 decide_start_workset(block);
 813         }
 814
 815         DB((dbg, DBG_DECIDE, "\n"));
 816         DB((dbg, DBG_DECIDE, "Decide for %+F\n", block));
 817
 818         block_info = new_block_info();
 819         set_block_info(block, block_info);
 820
 821         DB((dbg, DBG_WSETS, "Start workset for %+F:\n", block));
 822         workset_foreach(ws, irn, iter) {
 823                 DB((dbg, DBG_WSETS, "  %+F (%u)\n", irn,
 824                      workset_get_time(ws, iter)));
 825         }
 826
 827         block_info->start_workset = workset_clone(ws);
 828
 829         /* process the block from start to end */
 830         DB((dbg, DBG_WSETS, "Processing...\n"));
 831         ir_nodeset_init(&used);
 832         instr_nr = 0;
 833         /* TODO: this leaks (into the obstack)... */
 834         new_vals = new_workset();
 835
 836         sched_foreach(block, irn) {
 837                 int i, arity;
 838                 assert(workset_get_length(ws) <= n_regs);
 839
 840                 /* Phis are no real instr (see insert_starters()) */
 841                 if (is_Phi(irn)) {
 842                         continue;
 843                 }
 844                 DB((dbg, DBG_DECIDE, "  ...%+F\n", irn));
 845
 846                 /* set instruction in the workset */
 847                 instr = irn;
 848
 849                 /* allocate all values _used_ by this instruction */
 850                 workset_clear(new_vals);
 851                 for(i = 0, arity = get_irn_arity(irn); i < arity; ++i) {
 852                         ir_node *in = get_irn_n(irn, i);
 853                         if (!arch_irn_consider_in_reg_alloc(arch_env, cls, in))
 854                                 continue;
 855
 856                         /* (note that "spilled" is irrelevant here) */
 857                         workset_insert(new_vals, in, false);
 858                 }
 859                 displace(new_vals, 1);
 860
 861                 /* allocate all values _defined_ by this instruction */
 862                 workset_clear(new_vals);
 863                 if (get_irn_mode(irn) == mode_T) {
 864                         const ir_edge_t *edge;
 865
 866                         foreach_out_edge(irn, edge) {
 867                                 ir_node *proj = get_edge_src_irn(edge);
 868                                 if (!arch_irn_consider_in_reg_alloc(arch_env, cls, proj))
 869                                         continue;
 870                                 workset_insert(new_vals, proj, false);
 871                         }
 872                 } else {
 873                         if (!arch_irn_consider_in_reg_alloc(arch_env, cls, irn))
 874                                 continue;
 875                         workset_insert(new_vals, irn, false);
 876                 }
 877                 displace(new_vals, 0);
 878
 879                 instr_nr++;
 880         }
 881         ir_nodeset_destroy(&used);
 882
 883         /* Remember end-workset for this block */
 884         block_info->end_workset = workset_clone(ws);
 885         DB((dbg, DBG_WSETS, "End workset for %+F:\n", block));
 886         workset_foreach(ws, irn, iter)
 887                 DB((dbg, DBG_WSETS, "  %+F (%u)\n", irn,
 888                      workset_get_time(ws, iter)));
 889
 890         /* add successor blocks into worklist */
 891         foreach_block_succ(block, edge) {
 892                 ir_node *succ = get_edge_src_irn(edge);
 893                 pdeq_putr(worklist, succ);
 894         }
 895 }
 896
 897 /**
 898  * 'decide' is block-local and makes assumptions
 899  * about the set of live-ins. Thus we must adapt the
 900  * live-outs to the live-ins at each block-border.
 901  */
 902 static void fix_block_borders(ir_node *block, void *data)
 903 {
 904         workset_t    *start_workset;
 905         int           arity;
 906         int           i;
 907         int           iter;
 908         (void) data;
 909
 910         DB((dbg, DBG_FIX, "\n"));
 911         DB((dbg, DBG_FIX, "Fixing %+F\n", block));
 912
 913         start_workset = get_block_info(block)->start_workset;
 914
 915         /* process all pred blocks */
 916         arity = get_irn_arity(block);
 917         for (i = 0; i < arity; ++i) {
 918                 ir_node   *pred = get_Block_cfgpred_block(block, i);
 919                 workset_t *pred_end_workset = get_block_info(pred)->end_workset;
 920                 ir_node   *node;
 921
 922                 DB((dbg, DBG_FIX, "  Pred %+F\n", pred));
 923
 924                 /* spill all values not used anymore */
 925                 workset_foreach(pred_end_workset, node, iter) {
 926                         ir_node *n2;
 927                         int      iter2;
 928                         bool     found = false;
 929                         workset_foreach(start_workset, n2, iter2) {
 930                                 if (n2 == node) {
 931                                         found = true;
 932                                         break;
 933                                 }
 934                                 /* note that we do not look at phi inputs, becuase the values
 935                                  * will be either live-end and need no spill or
 936                                  * they have other users in which must be somewhere else in the
 937                                  * workset */
 938                         }
 939
 940                         if (found)
 941                                 continue;
 942
 943                         if (move_spills && be_is_live_in(lv, block, node)
 944                                         && !pred_end_workset->vals[iter].spilled) {
 945                                 ir_node *insert_point;
 946                                 if (arity > 1) {
 947                                         insert_point = be_get_end_of_block_insertion_point(pred);
 948                                         insert_point = sched_prev(insert_point);
 949                                 } else {
 950                                         insert_point = block;
 951                                 }
 952                                 DB((dbg, DBG_SPILL, "Spill %+F after %+F\n", node,
 953                                      insert_point));
 954                                 be_add_spill(senv, node, insert_point);
 955                         }
 956                 }
 957
 958                 /* reload missing values in predecessors, add missing spills */
 959                 workset_foreach(start_workset, node, iter) {
 960                         const loc_t *l    = &start_workset->vals[iter];
 961                         const loc_t *pred_loc;
 962
 963                         /* if node is a phi of the current block we reload
 964                          * the corresponding argument, else node itself */
 965                         if (is_Phi(node) && get_nodes_block(node) == block) {
 966                                 node = get_irn_n(node, i);
 967                                 assert(!l->spilled);
 968
 969                                 /* we might have unknowns as argument for the phi */
 970                                 if (!arch_irn_consider_in_reg_alloc(arch_env, cls, node))
 971                                         continue;
 972                         }
 973
 974                         /* check if node is in a register at end of pred */
 975                         pred_loc = workset_contains(pred_end_workset, node);
 976                         if (pred_loc != NULL) {
 977                                 /* we might have to spill value on this path */
 978                                 if (move_spills && !pred_loc->spilled && l->spilled) {
 979                                         ir_node *insert_point
 980                                                 = be_get_end_of_block_insertion_point(pred);
 981                                         insert_point = sched_prev(insert_point);
 982                                         DB((dbg, DBG_SPILL, "Spill %+F after %+F\n", node,
 983                                             insert_point));
 984                                         be_add_spill(senv, node, insert_point);
 985                                 }
 986                         } else {
 987                                 /* node is not in register at the end of pred -> reload it */
 988                                 DB((dbg, DBG_FIX, "    reload %+F\n", node));
 989                                 DB((dbg, DBG_SPILL, "Reload %+F before %+F,%d\n", node, block, i));
 990                                 be_add_reload_on_edge(senv, node, block, i, cls, 1);
 991                         }
 992                 }
 993         }
 994 }
 995
 996 static void be_spill_belady(be_irg_t *birg, const arch_register_class_t *rcls)
 997 {
 998         ir_graph *irg = be_get_birg_irg(birg);
 999
1000         be_liveness_assure_sets(be_assure_liveness(birg));
1001
1002         /* construct control flow loop tree */
1003         if (! (get_irg_loopinfo_state(irg) & loopinfo_cf_consistent)) {
1004                 construct_cf_backedges(irg);
1005         }
1006
1007         be_clear_links(irg);
1008
1009         /* init belady env */
1010         obstack_init(&obst);
1011         arch_env = birg->main_env->arch_env;
1012         cls      = rcls;
1013         lv       = be_get_birg_liveness(birg);
1014         n_regs   = cls->n_regs - be_put_ignore_regs(birg, cls, NULL);
1015         ws       = new_workset();
1016         uses     = be_begin_uses(irg, lv);
1017         loop_ana = be_new_loop_pressure(birg);
1018         senv     = be_new_spill_env(birg);
1019         worklist = new_pdeq();
1020
1021         pdeq_putr(worklist, get_irg_start_block(irg));
1022
1023         while(!pdeq_empty(worklist)) {
1024                 ir_node *block = pdeq_getl(worklist);
1025                 belady(block);
1026         }
1027         /* end block might not be reachable in endless loops */
1028         belady(get_irg_end_block(irg));
1029
1030         del_pdeq(worklist);
1031
1032         /* belady was block-local, fix the global flow by adding reloads on the
1033          * edges */
1034         irg_block_walk_graph(irg, fix_block_borders, NULL, NULL);
1035
1036         /* Insert spill/reload nodes into the graph and fix usages */
1037         be_insert_spills_reloads(senv);
1038
1039         /* clean up */
1040         be_delete_spill_env(senv);
1041         be_end_uses(uses);
1042         be_free_loop_pressure(loop_ana);
1043         obstack_free(&obst, NULL);
1044 }
1045
1046 void be_init_spillbelady(void)
1047 {
1048         lc_opt_entry_t *be_grp       = lc_opt_get_grp(firm_opt_get_root(), "be");
1049         lc_opt_entry_t *belady_group = lc_opt_get_grp(be_grp, "belady");
1050         lc_opt_add_table(belady_group, options);
1051
1052         static be_spiller_t belady_spiller = {
1053                 be_spill_belady
1054         };
1055
1056         be_register_spiller("belady", &belady_spiller);
1057         FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
1058 }
1059
1060 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady);