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

   1 /** vim: set sw=4 ts=4:
   2  * @file   bespillremat.c
   3  * @date   2006-04-06
   4  * @author Adam M. Szalkowski & Sebastian Hack
   5  *
   6  * ILP based spilling & rematerialization
   7  *
   8  * Copyright (C) 2006 Universitaet Karlsruhe
   9  * Released under the GPL
  10  */
  11 #ifdef HAVE_CONFIG_H
  12 #include "config.h"
  13 #endif
  14
  15 #ifdef WITH_ILP
  16
  17 #include <math.h>
  18
  19 #include "hashptr.h"
  20 #include "debug.h"
  21 #include "obst.h"
  22 #include "set.h"
  23 #include "list.h"
  24 #include "pmap.h"
  25
  26 #include "irprintf.h"
  27 #include "irgwalk.h"
  28 #include "irdump_t.h"
  29 #include "irnode_t.h"
  30 #include "ircons_t.h"
  31 #include "irloop_t.h"
  32 #include "phiclass.h"
  33 #include "iredges.h"
  34 #include "execfreq.h"
  35 #include "irvrfy.h"
  36 #include "irbackedge_t.h"
  37
  38 #include <lpp/lpp.h>
  39 #include <lpp/mps.h>
  40 #include <lpp/lpp_net.h>
  41 #include <lpp/lpp_cplex.h>
  42
  43 #include "be_t.h"
  44 #include "belive_t.h"
  45 #include "besched_t.h"
  46 #include "beirgmod.h"
  47 #include "bearch.h"
  48 #include "beabi.h"
  49 #include "benode_t.h"
  50 #include "beutil.h"
  51 #include "bespillremat.h"
  52 #include "bespill.h"
  53 #include "bepressurestat.h"
  54 #include "beprofile.h"
  55 #include "bespilloptions.h"
  56
  57 #include "bechordal_t.h"
  58
  59 #include <libcore/lc_opts.h>
  60 #include <libcore/lc_opts_enum.h>
  61
  62 #define DUMP_PROBLEM       1
  63 #define DUMP_MPS           2
  64 #define DUMP_SOLUTION      4
  65 #define DUMP_STATS         8
  66 #define DUMP_PRESSURE      16
  67
  68 #define KEEPALIVE_REMATS   1
  69 #define KEEPALIVE_SPILLS   2
  70 #define KEEPALIVE_RELOADS  4
  71
  72 #define VERIFY_MEMINTERF   1
  73 #define VERIFY_DOMINANCE   2
  74
  75 #define REMATS_NONE        0
  76 #define REMATS_BRIGGS      1
  77 #define REMATS_NOINVERSE   2
  78 #define REMATS_ALL         3
  79
  80 static unsigned opt_dump_flags = 0;
  81 static int opt_log = 0;
  82 static unsigned opt_keep_alive = 0;
  83 static int opt_goodwin = 1;
  84 static int opt_memcopies = 1;
  85 static int opt_memoperands = 1;
  86 static int opt_verify = VERIFY_MEMINTERF;
  87 static unsigned opt_remats = REMATS_ALL;
  88 static int opt_repair_schedule = 0;
  89 static int opt_no_enlarge_liveness = 0;
  90 static int opt_remat_while_live = 1;
  91 static int opt_timeout = 300;
  92 static double opt_cost_reload = 8.0;
  93 static double opt_cost_memoperand =  7.0;
  94 static double opt_cost_spill =  15.0;
  95 static double opt_cost_remat =  1.0;
  96
  97
  98 static const lc_opt_enum_mask_items_t dump_items[] = {
  99         { "problem",  DUMP_PROBLEM  },
 100         { "mps",      DUMP_MPS      },
 101         { "solution", DUMP_SOLUTION },
 102         { "stats",    DUMP_STATS },
 103         { "pressure", DUMP_PRESSURE },
 104         { NULL,       0 }
 105 };
 106
 107 static lc_opt_enum_mask_var_t dump_var = {
 108         &opt_dump_flags, dump_items
 109 };
 110
 111 static const lc_opt_enum_mask_items_t keepalive_items[] = {
 112         { "remats",  KEEPALIVE_REMATS  },
 113         { "spills",  KEEPALIVE_SPILLS  },
 114         { "reloads", KEEPALIVE_RELOADS },
 115         { NULL,      0 }
 116 };
 117
 118 static lc_opt_enum_mask_var_t keep_alive_var = {
 119         &opt_keep_alive, keepalive_items
 120 };
 121
 122 static const lc_opt_enum_mask_items_t remats_items[] = {
 123         { "none",      REMATS_NONE      },
 124         { "briggs",    REMATS_BRIGGS    },
 125         { "noinverse", REMATS_NOINVERSE },
 126         { "all",       REMATS_ALL       },
 127         { NULL,        0 }
 128 };
 129
 130 static lc_opt_enum_mask_var_t remats_var = {
 131         &opt_remats, remats_items
 132 };
 133
 134 static const lc_opt_table_entry_t options[] = {
 135         LC_OPT_ENT_ENUM_MASK("keepalive", "keep alive inserted nodes",                              &keep_alive_var),
 136
 137         LC_OPT_ENT_BOOL     ("goodwin",  "activate goodwin reduction",                              &opt_goodwin),
 138         LC_OPT_ENT_BOOL     ("memcopies",  "activate memcopy handling",                             &opt_memcopies),
 139         LC_OPT_ENT_BOOL     ("memoperands",  "activate memoperands",                                &opt_memoperands),
 140         LC_OPT_ENT_ENUM_INT ("remats",  "type of remats to insert",                                 &remats_var),
 141         LC_OPT_ENT_BOOL     ("repair_schedule",  "repair the schedule by rematting once used nodes",&opt_repair_schedule),
 142         LC_OPT_ENT_BOOL     ("no_enlage_liveness",  "do not enlarge liveness of operands of remats",&opt_no_enlarge_liveness),
 143         LC_OPT_ENT_BOOL     ("remat_while_live",  "only remat where rematted value was live",       &opt_remat_while_live),
 144
 145         LC_OPT_ENT_ENUM_MASK("dump", "dump problem, solution or statistical data",                  &dump_var),
 146         LC_OPT_ENT_BOOL     ("log",  "activate the lpp log",                                        &opt_log),
 147         LC_OPT_ENT_INT      ("timeout",  "ILP solver timeout",                                      &opt_timeout),
 148
 149         LC_OPT_ENT_DBL      ("cost_reload",  "cost of a reload",                                    &opt_cost_reload),
 150         LC_OPT_ENT_DBL      ("cost_memoperand",  "cost of a memory operand",                        &opt_cost_memoperand),
 151         LC_OPT_ENT_DBL      ("cost_spill",  "cost of a spill instruction",                          &opt_cost_spill),
 152         LC_OPT_ENT_DBL      ("cost_remat",  "cost of a rematerialization",                          &opt_cost_remat),
 153         { NULL }
 154 };
 155
 156 //#define EXECFREQ_LOOPDEPH   /* compute execution frequency from loop depth only */
 157 //#define SCHEDULE_PHIM   /* insert phim nodes into schedule */
 158
 159 #define  SOLVE
 160 //#define  SOLVE_LOCAL
 161 #define LPP_SERVER "i44pc52"
 162 #define LPP_SOLVER "cplex"
 163
 164
 165 #define MAX_PATHS      INT_MAX
 166 #define ILP_UNDEF               -1
 167
 168 typedef struct _spill_ilp_t {
 169         const arch_register_class_t  *cls;
 170         int                           n_regs;
 171         be_irg_t                     *birg;
 172         be_lv_t                      *lv;
 173         lpp_t                        *lpp;
 174         struct obstack               *obst;
 175         set                          *remat_info;
 176         pset                         *all_possible_remats;
 177         pset                         *inverse_ops;
 178         ir_node                      *keep;
 179         set                          *values; /**< for collecting all definitions of values before running ssa-construction */
 180         pset                         *spills;
 181         set                          *interferences;
 182         ir_node                      *m_unknown;
 183         set                          *memoperands;
 184         phi_classes_t                *pc;
 185 #ifndef SCHEDULE_PHIM
 186         pset                         *phims;
 187 #endif
 188         DEBUG_ONLY(firm_dbg_module_t * dbg);
 189 } spill_ilp_t;
 190
 191 typedef int ilp_var_t;
 192 typedef int ilp_cst_t;
 193
 194 typedef struct _spill_bb_t {
 195         set      *ilp;
 196         set      *reloads;
 197 } spill_bb_t;
 198
 199 typedef struct _remat_t {
 200         const ir_node        *op;      /**< for copy_irn */
 201         const ir_node        *value;   /**< the value which is being recomputed by this remat */
 202         const ir_node        *proj;    /**< not NULL if the above op produces a tuple */
 203         int                   cost;    /**< cost of this remat */
 204         int                   inverse; /**< nonzero if this is an inverse remat */
 205 } remat_t;
 206
 207 /**
 208  * Data to be attached to each IR node. For remats this contains the ilp_var
 209  * for this remat and for normal ops this contains the ilp_vars for
 210  * reloading each operand
 211  */
 212 typedef struct _op_t {
 213         int             is_remat;
 214         union {
 215                 struct {
 216                         ilp_var_t       ilp;
 217                         const remat_t  *remat; /** the remat this op belongs to */
 218                         int             pre; /** 1, if this is a pressure-increasing remat */
 219                 } remat;
 220                 struct {
 221                         ilp_var_t       ilp;
 222                         ir_node        *op; /** the operation this live range belongs to */
 223                         union {
 224                                 ilp_var_t      *reloads;
 225                                 ilp_var_t      *copies;
 226                         } args;
 227                 } live_range;
 228         } attr;
 229 } op_t;
 230
 231 typedef struct _defs_t {
 232         const ir_node   *value;
 233         ir_node         *spills;  /**< points to the first spill for this value (linked by link field) */
 234         ir_node         *remats;  /**< points to the first definition for this value (linked by link field) */
 235 } defs_t;
 236
 237 typedef struct _remat_info_t {
 238         const ir_node       *irn; /**< the irn to which these remats belong */
 239         pset                *remats; /**< possible remats for this value */
 240         pset                *remats_by_operand; /**< remats with this value as operand */
 241 } remat_info_t;
 242
 243 typedef struct _keyval_t {
 244         const void          *key;
 245         const void          *val;
 246 } keyval_t;
 247
 248 typedef struct _spill_t {
 249         ir_node            *irn;
 250         ilp_var_t           reg_in;
 251         ilp_var_t           mem_in;
 252         ilp_var_t           reg_out;
 253         ilp_var_t           mem_out;
 254         ilp_var_t           spill;
 255 } spill_t;
 256
 257 typedef struct _memoperand_t {
 258         ir_node             *irn; /**< the irn */
 259         unsigned int         pos; /**< the position of the argument */
 260         ilp_var_t            ilp; /**< the ilp var for this memory operand */
 261 } memoperand_t;
 262
 263 static INLINE int
 264 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
 265 {
 266         return arch_irn_consider_in_reg_alloc(si->birg->main_env->arch_env,
 267                                               si->cls, irn);
 268 }
 269
 270 #if 0
 271 static int
 272 cmp_remat(const void *a, const void *b)
 273 {
 274         const keyval_t *p = a;
 275         const keyval_t *q = b;
 276         const remat_t  *r = p->val;
 277         const remat_t  *s = q->val;
 278
 279         assert(r && s);
 280
 281         return !(r == s || r->op == s->op);
 282 }
 283 #endif
 284 static int
 285 cmp_remat(const void *a, const void *b)
 286 {
 287         const remat_t  *r = a;
 288         const remat_t  *s = a;
 289
 290         return !(r == s || r->op == s->op);
 291 }
 292
 293 static int
 294 cmp_spill(const void *a, const void *b, size_t size)
 295 {
 296         const spill_t *p = a;
 297         const spill_t *q = b;
 298
 299 //      return !(p->irn == q->irn && p->bb == q->bb);
 300         return !(p->irn == q->irn);
 301 }
 302
 303 static int
 304 cmp_memoperands(const void *a, const void *b, size_t size)
 305 {
 306         const memoperand_t *p = a;
 307         const memoperand_t *q = b;
 308
 309         return !(p->irn == q->irn && p->pos == q->pos);
 310 }
 311
 312 static keyval_t *
 313 set_find_keyval(set * set, const void * key)
 314 {
 315         keyval_t     query;
 316
 317         query.key = key;
 318         return set_find(set, &query, sizeof(query), HASH_PTR(key));
 319 }
 320
 321 static keyval_t *
 322 set_insert_keyval(set * set, void * key, void * val)
 323 {
 324         keyval_t     query;
 325
 326         query.key = key;
 327         query.val = val;
 328         return set_insert(set, &query, sizeof(query), HASH_PTR(key));
 329 }
 330
 331 static defs_t *
 332 set_find_def(set * set, const ir_node * value)
 333 {
 334         defs_t     query;
 335
 336         query.value = value;
 337         return set_find(set, &query, sizeof(query), HASH_PTR(value));
 338 }
 339
 340 static defs_t *
 341 set_insert_def(set * set, const ir_node * value)
 342 {
 343         defs_t     query;
 344
 345         query.value = value;
 346         query.spills = NULL;
 347         query.remats = NULL;
 348         return set_insert(set, &query, sizeof(query), HASH_PTR(value));
 349 }
 350
 351 static memoperand_t *
 352 set_insert_memoperand(set * set, ir_node * irn, unsigned int pos, ilp_var_t ilp)
 353 {
 354         memoperand_t     query;
 355
 356         query.irn = irn;
 357         query.pos = pos;
 358         query.ilp = ilp;
 359         return set_insert(set, &query, sizeof(query), HASH_PTR(irn)+pos);
 360 }
 361
 362 static memoperand_t *
 363 set_find_memoperand(set * set, const ir_node * irn, unsigned int pos)
 364 {
 365         memoperand_t     query;
 366
 367         query.irn = (ir_node*)irn;
 368         query.pos = pos;
 369         return set_find(set, &query, sizeof(query), HASH_PTR(irn)+pos);
 370 }
 371
 372
 373 static spill_t *
 374 set_find_spill(set * set, const ir_node * value)
 375 {
 376         spill_t     query;
 377
 378         query.irn = (ir_node*)value;
 379         return set_find(set, &query, sizeof(query), HASH_PTR(value));
 380 }
 381
 382 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
 383 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
 384 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
 385 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
 386 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
 387
 388 static int
 389 cmp_remat_info(const void *a, const void *b, size_t size)
 390 {
 391         const remat_info_t *p = a;
 392         const remat_info_t *q = b;
 393
 394         return !(p->irn == q->irn);
 395 }
 396
 397 static int
 398 cmp_defs(const void *a, const void *b, size_t size)
 399 {
 400         const defs_t *p = a;
 401         const defs_t *q = b;
 402
 403         return !(p->value == q->value);
 404 }
 405
 406 static int
 407 cmp_keyval(const void *a, const void *b, size_t size)
 408 {
 409         const keyval_t *p = a;
 410         const keyval_t *q = b;
 411
 412         return !(p->key == q->key);
 413 }
 414
 415 static double
 416 execution_frequency(const spill_ilp_t *si, const ir_node * irn)
 417 {
 418 #define FUDGE 0.001
 419         if(be_profile_has_data())
 420                 return ((double)be_profile_get_block_execcount(get_block(irn))) + FUDGE;
 421
 422 #ifndef EXECFREQ_LOOPDEPH
 423         return get_block_execfreq(si->birg->exec_freq, get_block(irn)) + FUDGE;
 424 #else
 425         if(is_Block(irn))
 426                 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
 427         else
 428                 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
 429 #endif
 430 }
 431
 432 static double
 433 get_cost(const spill_ilp_t * si, const ir_node * irn)
 434 {
 435         if(be_is_Spill(irn)) {
 436                 return opt_cost_spill;
 437         } else if(be_is_Reload(irn)){
 438                 return opt_cost_reload;
 439         } else {
 440                 return arch_get_op_estimated_cost(si->birg->main_env->arch_env, irn);
 441         }
 442 }
 443
 444 /**
 445  * Checks, whether node and its operands have suitable reg classes
 446  */
 447 static INLINE int
 448 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
 449 {
 450         int               n;
 451         const arch_env_t *arch_env = si->birg->main_env->arch_env;
 452         int               remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
 453
 454 #if 0
 455         if(!remat)
 456                 ir_fprintf(stderr, "  Node %+F is not rematerializable\n", irn);
 457 #endif
 458
 459         for (n = get_irn_arity(irn)-1; n>=0 && remat; --n) {
 460                 ir_node        *op = get_irn_n(irn, n);
 461                 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
 462
 463 //              if(!remat)
 464 //                      ir_fprintf(stderr, "  Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
 465         }
 466
 467         return remat;
 468 }
 469
 470 /**
 471  * Try to create a remat from @p op with destination value @p dest_value
 472  */
 473 static INLINE remat_t *
 474 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
 475 {
 476         remat_t  *remat = NULL;
 477
 478 //      if(!mode_is_datab(get_irn_mode(dest_value)))
 479 //              return NULL;
 480
 481         if(dest_value == op) {
 482                 const ir_node *proj = NULL;
 483
 484                 if(is_Proj(dest_value)) {
 485                         op = get_Proj_pred(op);
 486                         proj = dest_value;
 487                 }
 488
 489                 if(!is_rematerializable(si, op))
 490                         return NULL;
 491
 492                 remat          = obstack_alloc(si->obst, sizeof(*remat));
 493                 remat->op      = op;
 494                 remat->cost    = (int)get_cost(si, op);
 495                 remat->value   = dest_value;
 496                 remat->proj    = proj;
 497                 remat->inverse = 0;
 498         } else {
 499                 arch_inverse_t     inverse;
 500                 int                n;
 501
 502                 /* get the index of the operand we want to retrieve by the inverse op */
 503                 for (n = get_irn_arity(op)-1; n>=0; --n) {
 504                         ir_node        *arg = get_irn_n(op, n);
 505
 506                         if(arg == dest_value) break;
 507                 }
 508                 if(n<0) return NULL;
 509
 510                 DBG((si->dbg, LEVEL_5, "\t  requesting inverse op for argument %d of op %+F\n", n, op));
 511
 512                 /* else ask the backend to give an inverse op */
 513                 if(arch_get_inverse(si->birg->main_env->arch_env, op, n, &inverse, si->obst)) {
 514                         int   i;
 515
 516                         DBG((si->dbg, LEVEL_4, "\t  backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
 517
 518                         assert(inverse.n > 0 && "inverse op should have at least one node");
 519
 520                         for(i=inverse.n-1; i>=0; --i) {
 521                                 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
 522                         }
 523
 524                         if(inverse.n <= 2) {
 525                                 remat = obstack_alloc(si->obst, sizeof(*remat));
 526                                 remat->op = inverse.nodes[0];
 527                                 remat->cost = inverse.costs;
 528                                 remat->value = dest_value;
 529                                 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
 530                                 remat->inverse = 1;
 531
 532                                 // Matze: commented this out, this doesn't seem to be true if
 533                                 // the inverse is a simple operation with only 1 result...
 534                                 //assert(is_Proj(remat->proj));
 535                         } else {
 536                                 assert(0 && "I can not handle remats with more than 2 nodes");
 537                         }
 538                 }
 539         }
 540
 541         if(remat) {
 542                 if(remat->proj) {
 543                         DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
 544                 } else {
 545                         DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
 546                 }
 547         }
 548         return remat;
 549 }
 550
 551
 552 static INLINE void
 553 add_remat(const spill_ilp_t * si, const remat_t * remat)
 554 {
 555         remat_info_t    *remat_info,
 556                      query;
 557         int              n;
 558
 559         assert(remat->op);
 560         assert(remat->value);
 561
 562         query.irn = remat->value;
 563         query.remats = NULL;
 564         query.remats_by_operand = NULL;
 565         remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
 566
 567         if(remat_info->remats == NULL) {
 568                 remat_info->remats = new_pset(cmp_remat, 4096);
 569         }
 570         pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
 571
 572         /* insert the remat into the remats_be_operand set of each argument of the remat op */
 573         for (n = get_irn_arity(remat->op)-1; n>=0; --n) {
 574                 ir_node        *arg = get_irn_n(remat->op, n);
 575
 576                 query.irn = arg;
 577                 query.remats = NULL;
 578                 query.remats_by_operand = NULL;
 579                 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
 580
 581                 if(remat_info->remats_by_operand == NULL) {
 582                         remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
 583                 }
 584                 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
 585         }
 586 }
 587
 588 static int
 589 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
 590 {
 591         const ir_edge_t   *edge = get_irn_out_edge_first(irn);
 592         int                i = 0;
 593
 594         while(edge) {
 595                 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
 596                         ++i;
 597                 }
 598                 edge = get_irn_out_edge_next(irn, edge);
 599         }
 600
 601         return i;
 602 }
 603
 604 static int
 605 get_irn_n_nonignore_args(const spill_ilp_t * si, const ir_node * irn)
 606 {
 607         int n;
 608         int ret = 0;
 609
 610         if(is_Proj(irn))
 611                 irn = get_Proj_pred(irn);
 612
 613         for(n=get_irn_arity(irn)-1; n>=0; --n) {
 614                 const ir_node  *arg = get_irn_n(irn, n);
 615
 616                 if(has_reg_class(si, arg)) ++ret;
 617         }
 618
 619         return ret;
 620 }
 621
 622 static INLINE void
 623 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
 624 {
 625         int      n;
 626         remat_t *remat;
 627
 628         if( has_reg_class(si, op)
 629         && (opt_repair_schedule || get_irn_n_nonremat_edges(si, op) > 1)
 630         && (opt_remats !=  REMATS_BRIGGS || get_irn_n_nonignore_args(si, op) == 0)
 631         ) {
 632                 remat = get_remat_from_op(si, op, op);
 633                 if(remat) {
 634                         add_remat(si, remat);
 635                 }
 636         }
 637
 638         if(opt_remats == REMATS_ALL) {
 639                 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
 640                    for each arg */
 641                 for (n = get_irn_arity(op)-1; n>=0; --n) {
 642                         ir_node        *arg = get_irn_n(op, n);
 643
 644                         if(has_reg_class(si, arg)) {
 645                                 /* try to get an inverse remat */
 646                                 remat = get_remat_from_op(si, arg, op);
 647                                 if(remat) {
 648                                         add_remat(si, remat);
 649                                 }
 650                         }
 651                 }
 652         }
 653 }
 654
 655 static INLINE int
 656 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
 657 {
 658         ir_node *block;
 659         ir_node *def_block = get_nodes_block(val);
 660         int      ret;
 661
 662         if(val == pos)
 663                 return 0;
 664
 665         /* if pos is at end of a basic block */
 666         if(is_Block(pos)) {
 667                 ret = (pos == def_block || block_dominates(def_block, pos));
 668 //              ir_fprintf(stderr, "(def(bb)=%d) ", ret);
 669                 return ret;
 670         }
 671
 672         /* else if this is a normal operation */
 673         block = get_nodes_block(pos);
 674         if(block == def_block) {
 675                 if(!sched_is_scheduled(val)) return 1;
 676
 677                 ret = sched_comes_after(val, pos);
 678 //              ir_fprintf(stderr, "(def(same block)=%d) ",ret);
 679                 return ret;
 680         }
 681
 682         ret = block_dominates(def_block, block);
 683 //      ir_fprintf(stderr, "(def(other block)=%d) ", ret);
 684         return ret;
 685 }
 686
 687 static INLINE ir_node *
 688 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
 689 {
 690     return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->birg->main_env->arch_env);
 691 }
 692
 693 /**
 694  * Returns first non-Phi node of block @p bb
 695  */
 696 static INLINE ir_node *
 697 sched_block_first_nonphi(const ir_node * bb)
 698 {
 699         return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
 700 }
 701
 702 static int
 703 sched_skip_proj_predicator(const ir_node * irn, void * data)
 704 {
 705         return (is_Proj(irn));
 706 }
 707
 708 static INLINE ir_node *
 709 sched_next_nonproj(const ir_node * irn, int forward)
 710 {
 711         return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
 712 }
 713
 714 /**
 715  * Returns next operation node (non-Proj) after @p irn
 716  * or the basic block of this node
 717  */
 718 static INLINE ir_node *
 719 sched_next_op(const ir_node * irn)
 720 {
 721         ir_node *next = sched_next(irn);
 722
 723         if(is_Block(next))
 724                 return next;
 725
 726         return sched_next_nonproj(next, 1);
 727 }
 728
 729 /**
 730  * Returns previous operation node (non-Proj) before @p irn
 731  * or the basic block of this node
 732  */
 733 static INLINE ir_node *
 734 sched_prev_op(const ir_node * irn)
 735 {
 736         ir_node *prev = sched_prev(irn);
 737
 738         if(is_Block(prev))
 739                 return prev;
 740
 741         return sched_next_nonproj(prev, 0);
 742 }
 743
 744 static void
 745 sched_put_after(ir_node * insert, ir_node * irn)
 746 {
 747         if(is_Block(insert)) {
 748                 insert = sched_block_first_nonphi(insert);
 749         } else {
 750                 insert = sched_next_op(insert);
 751         }
 752         sched_reset(irn);
 753         sched_add_before(insert, irn);
 754 }
 755
 756 static void
 757 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
 758 {
 759   if(is_Block(insert)) {
 760           insert = sched_block_last_noncf(si, insert);
 761   } else {
 762           insert = sched_next_nonproj(insert, 0);
 763           insert = sched_prev(insert);
 764   }
 765   sched_reset(irn);
 766   sched_add_after(insert, irn);
 767 }
 768
 769 static ir_node *
 770 next_post_remat(const ir_node * irn)
 771 {
 772         op_t      *op;
 773     ir_node   *next;
 774
 775         if(is_Block(irn)) {
 776                 next = sched_block_first_nonphi(irn);
 777         } else {
 778                 next = sched_next_op(irn);
 779         }
 780
 781         if(sched_is_end(next))
 782                 return NULL;
 783
 784         op = get_irn_link(next);
 785         if(op->is_remat && !op->attr.remat.pre) {
 786                 return next;
 787         }
 788
 789         return NULL;
 790 }
 791
 792
 793 static ir_node *
 794 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
 795 {
 796         op_t      *op;
 797         ir_node   *ret;
 798
 799         if(is_Block(irn)) {
 800                 ret = sched_block_last_noncf(si, irn);
 801                 ret = sched_next(ret);
 802                 ret = sched_prev_op(ret);
 803         } else {
 804                 ret = sched_prev_op(irn);
 805         }
 806
 807         if(sched_is_end(ret) || is_Phi(ret))
 808                 return NULL;
 809
 810         op = (op_t*)get_irn_link(ret);
 811         if(op->is_remat && op->attr.remat.pre) {
 812                 return ret;
 813         }
 814
 815         return NULL;
 816 }
 817
 818 /**
 819  * Tells you whether a @p remat can be placed before the irn @p pos
 820  */
 821 static INLINE int
 822 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 823 {
 824         const ir_node   *op = remat->op;
 825         const ir_node   *prev;
 826         int        n,
 827                            res = 1;
 828
 829         if(is_Block(pos)) {
 830                 prev = sched_block_last_noncf(si, pos);
 831                 prev = sched_next_nonproj(prev, 0);
 832         } else {
 833                 prev = sched_prev_op(pos);
 834         }
 835         /* do not remat if the rematted value is defined immediately before this op */
 836         if(prev == remat->op) {
 837                 return 0;
 838         }
 839
 840 #if 0
 841         /* this should be just fine, the following OP will be using this value, right? */
 842
 843         /* only remat AFTER the real definition of a value (?) */
 844         if(!value_is_defined_before(si, pos, remat->value)) {
 845 //              ir_fprintf(stderr, "error(not defined)");
 846                 return 0;
 847         }
 848 #endif
 849
 850         for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
 851                 const ir_node   *arg = get_irn_n(op, n);
 852
 853                 if(opt_no_enlarge_liveness) {
 854                         if(has_reg_class(si, arg) && live) {
 855                                 res &= pset_find_ptr((pset*)live, arg)?1:0;
 856                         } else {
 857                                 res &= value_is_defined_before(si, pos, arg);
 858                         }
 859                 } else {
 860                         res &= value_is_defined_before(si, pos, arg);
 861                 }
 862         }
 863
 864         return res;
 865 }
 866
 867 /**
 868  * Tells you whether a @p remat can be placed after the irn @p pos
 869  */
 870 static INLINE int
 871 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 872 {
 873         if(is_Block(pos)) {
 874                 pos = sched_block_first_nonphi(pos);
 875         } else {
 876                 pos = sched_next_op(pos);
 877         }
 878
 879         /* only remat AFTER the real definition of a value (?) */
 880         if(!value_is_defined_before(si, pos, remat->value)) {
 881                 return 0;
 882         }
 883
 884         return can_remat_before(si, remat, pos, live);
 885 }
 886
 887 /**
 888  * Collect potetially rematerializable OPs
 889  */
 890 static void
 891 walker_remat_collector(ir_node * irn, void * data)
 892 {
 893         spill_ilp_t    *si = data;
 894
 895         if(!is_Block(irn) && !is_Phi(irn)) {
 896                 DBG((si->dbg, LEVEL_4, "\t  Processing %+F\n", irn));
 897                 get_remats_from_op(si, irn);
 898         }
 899 }
 900
 901 /**
 902  * Inserts a copy of @p irn before @p pos
 903  */
 904 static ir_node *
 905 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
 906 {
 907         ir_node     *bb;
 908         ir_node     *copy;
 909
 910         bb = is_Block(pos)?pos:get_nodes_block(pos);
 911         copy = exact_copy(irn);
 912
 913         set_phi_class(si->pc, copy, NULL);
 914         set_nodes_block(copy, bb);
 915         sched_put_before(si, pos, copy);
 916
 917         return copy;
 918 }
 919
 920 /**
 921  * Inserts a copy of @p irn after @p pos
 922  */
 923 static ir_node *
 924 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
 925 {
 926         ir_node     *bb;
 927         ir_node     *copy;
 928
 929         bb = is_Block(pos)?pos:get_nodes_block(pos);
 930         copy = exact_copy(irn);
 931
 932         set_phi_class(si->pc, copy, NULL);
 933         set_nodes_block(copy, bb);
 934         sched_put_after(pos, copy);
 935
 936         return copy;
 937 }
 938
 939 static ir_node *
 940 insert_remat_after(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
 941 {
 942         char     buf[256];
 943
 944         if(can_remat_after(si, remat, pos, live)) {
 945                 ir_node         *copy,
 946                                                 *proj_copy;
 947                 op_t            *op;
 948
 949                 DBG((si->dbg, LEVEL_3, "\t  >inserting remat2 %+F\n", remat->op));
 950
 951                 copy = insert_copy_after(si, remat->op, pos);
 952
 953                 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
 954                 op = obstack_alloc(si->obst, sizeof(*op));
 955                 op->is_remat = 1;
 956                 op->attr.remat.remat = remat;
 957                 op->attr.remat.pre = 0;
 958                 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
 959
 960                 set_irn_link(copy, op);
 961                 pset_insert_ptr(si->all_possible_remats, copy);
 962                 if(remat->proj) {
 963                         proj_copy = insert_copy_after(si, remat->proj, copy);
 964                         set_irn_n(proj_copy, 0, copy);
 965                         set_irn_link(proj_copy, op);
 966                         pset_insert_ptr(si->all_possible_remats, proj_copy);
 967                 } else {
 968                         proj_copy = NULL;
 969                 }
 970
 971                 return copy;
 972         }
 973
 974         return NULL;
 975 }
 976
 977 static ir_node *
 978 insert_remat_before(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
 979 {
 980         char     buf[256];
 981
 982         if(can_remat_before(si, remat, pos, live)) {
 983                 ir_node         *copy,
 984                                                 *proj_copy;
 985                 op_t            *op;
 986
 987                 DBG((si->dbg, LEVEL_3, "\t  >inserting remat %+F\n", remat->op));
 988
 989                 copy = insert_copy_before(si, remat->op, pos);
 990
 991                 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
 992                 op = obstack_alloc(si->obst, sizeof(*op));
 993                 op->is_remat = 1;
 994                 op->attr.remat.remat = remat;
 995                 op->attr.remat.pre = 1;
 996                 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
 997
 998                 set_irn_link(copy, op);
 999                 pset_insert_ptr(si->all_possible_remats, copy);
1000                 if(remat->proj) {
1001                         proj_copy = insert_copy_after(si, remat->proj, copy);
1002                         set_irn_n(proj_copy, 0, copy);
1003                         set_irn_link(proj_copy, op);
1004                         pset_insert_ptr(si->all_possible_remats, proj_copy);
1005                 } else {
1006                         proj_copy = NULL;
1007                 }
1008
1009                 return copy;
1010         }
1011
1012         return NULL;
1013 }
1014
1015 static int
1016 get_block_n_succs(const ir_node *block) {
1017         const ir_edge_t *edge;
1018
1019         assert(edges_activated(current_ir_graph));
1020
1021         edge = get_block_succ_first(block);
1022         if (! edge)
1023                 return 0;
1024
1025         edge = get_block_succ_next(block, edge);
1026         return edge ? 2 : 1;
1027 }
1028
1029 static int
1030 is_start_block(const ir_node * bb)
1031 {
1032         return get_irg_start_block(get_irn_irg(bb)) == bb;
1033 }
1034
1035 static int
1036 is_merge_edge(const ir_node * bb)
1037 {
1038         if(is_start_block(bb))
1039                 return 0;
1040
1041         if(opt_goodwin)
1042                 return get_block_n_succs(bb) == 1;
1043         else
1044                 return 1;
1045 }
1046
1047 static int
1048 is_diverge_edge(const ir_node * bb)
1049 {
1050         if(is_start_block(bb))
1051                 return 0;
1052
1053         if(opt_goodwin)
1054                 return get_Block_n_cfgpreds(bb) == 1;
1055         else
1056                 return 1;
1057 }
1058
1059 static void
1060 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1061 {
1062         ir_node        *irn;
1063         int i;
1064
1065         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1066                 irn = be_lv_get_irn(si->lv, bb, i);
1067
1068                 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1069                         pset_insert_ptr(live, irn);
1070                 }
1071         }
1072
1073         irn = sched_last(bb);
1074
1075         /* all values eaten by control flow operations are also live until the end of the block */
1076         sched_foreach_reverse(bb, irn) {
1077                 int  i;
1078
1079                 if(!sched_skip_cf_predicator(irn, si->birg->main_env->arch_env)) break;
1080
1081                 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1082                         ir_node *arg = get_irn_n(irn,i);
1083
1084                         if(has_reg_class(si, arg)) {
1085                                 pset_insert_ptr(live, arg);
1086                         }
1087                 }
1088         }
1089         /*
1090          * find values that are used by remats at end of block
1091          * and insert them into live set
1092          */
1093         foreach_pre_remat(si, bb, irn) {
1094                 int       n;
1095
1096                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1097                         ir_node        *remat_arg = get_irn_n(irn, n);
1098
1099                         if(!has_reg_class(si, remat_arg)) continue;
1100
1101                         /* if value is becoming live through use by remat */
1102                         if(!pset_find_ptr(live, remat_arg)) {
1103                                 DBG((si->dbg, LEVEL_4, "  value %+F becoming live through use by remat at end of block %+F\n", remat_arg, irn));
1104
1105                                 pset_insert_ptr(live, remat_arg);
1106                         }
1107                 }
1108         }
1109 }
1110
1111 static void
1112 walker_regclass_copy_insertor(ir_node * irn, void * data)
1113 {
1114         spill_ilp_t    *si = data;
1115
1116         if(is_Phi(irn) && has_reg_class(si, irn)) {
1117                 int n;
1118
1119                 for(n=get_irn_arity(irn)-1; n>=0; --n) {
1120                         ir_node  *phi_arg = get_irn_n(irn, n);
1121                         ir_node  *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
1122
1123                         if(!has_reg_class(si, phi_arg)) {
1124                                 ir_node   *copy = be_new_Copy(si->cls, si->birg->irg, bb, phi_arg);
1125                                 ir_node   *pos = sched_block_last_noncf(si, bb);
1126                                 op_t      *op = obstack_alloc(si->obst, sizeof(*op));
1127
1128                                 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
1129                                 sched_add_after(pos, copy);
1130                                 set_irn_n(irn, n, copy);
1131
1132                                 op->is_remat = 0;
1133                                 op->attr.live_range.args.reloads = NULL;
1134                                 op->attr.live_range.ilp = ILP_UNDEF;
1135                                 set_irn_link(copy, op);
1136                         }
1137                 }
1138         }
1139 }
1140
1141 /**
1142  * Insert (so far unused) remats into the irg to
1143  * recompute the potential liveness of all values
1144  */
1145 static void
1146 walker_remat_insertor(ir_node * bb, void * data)
1147 {
1148         spill_ilp_t    *si = data;
1149         ir_node        *irn;
1150         int             n, i;
1151         pset           *live;
1152         pset           *post_remats;
1153         remat_t        *remat;
1154
1155         /* skip start block, no remats to do there */
1156         if(is_start_block(bb)) return;
1157
1158         DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
1159
1160         live = pset_new_ptr_default();
1161         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1162                 ir_node        *value = be_lv_get_irn(si->lv, bb, i);
1163
1164                 /* add remats at end of block */
1165                 if (has_reg_class(si, value)) {
1166                         pset_insert_ptr(live, value);
1167                 }
1168         }
1169
1170         irn = sched_last(bb);
1171         while(!sched_is_end(irn)) {
1172                 ir_node   *next;
1173                 pset      *args;
1174                 ir_node   *arg;
1175                 pset      *used;
1176
1177                 next = sched_prev(irn);
1178
1179                 /* delete defined value from live set */
1180                 if(has_reg_class(si, irn)) {
1181                         pset_remove_ptr(live, irn);
1182                 }
1183
1184                 if(is_Phi(irn) || is_Proj(irn)) {
1185                         irn = next;
1186                         continue;
1187                 }
1188
1189                 args = pset_new_ptr_default();
1190                 used = pset_new_ptr_default();
1191
1192                 /* collect arguments of op and set args of op already live in epilog */
1193                 for (n = get_irn_arity(irn)-1; n>=0; --n) {
1194                         ir_node        *arg = get_irn_n(irn, n);
1195
1196                         pset_insert_ptr(args, arg);
1197                         if(has_reg_class(si, arg)) {
1198                                 pset_insert_ptr(live, arg);
1199                                 pset_insert_ptr(used, arg);
1200                         }
1201                 }
1202
1203                 /* insert all possible remats before irn */
1204                 pset_foreach(args, arg) {
1205                         remat_info_t   *remat_info,
1206                                                     query;
1207
1208                         /* continue if the operand has the wrong reg class */
1209                         if(!has_reg_class(si, arg))
1210                                 continue;
1211
1212                         query.irn = arg;
1213                         query.remats = NULL;
1214                         query.remats_by_operand = NULL;
1215                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1216
1217                         if(!remat_info) {
1218                                 continue;
1219                         }
1220
1221                         if(remat_info->remats) {
1222                                 pset_foreach(remat_info->remats, remat) {
1223                                         ir_node  *remat_irn = NULL;
1224
1225                                         DBG((si->dbg, LEVEL_4, "\t  considering remat %+F for arg %+F\n", remat->op, arg));
1226                                         remat_irn = insert_remat_before(si, remat, irn, live);
1227
1228                                         if(remat_irn) {
1229                                                 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1230                                                         ir_node  *remat_arg = get_irn_n(remat_irn, n);
1231
1232                                                         /* collect args of remats which are not args of op */
1233                                                         if(has_reg_class(si, remat_arg) && !pset_find_ptr(args, remat_arg)) {
1234                                                                 pset_insert_ptr(used, remat_arg);
1235                                                         }
1236                                                 }
1237                                         }
1238                                 }
1239                         }
1240                 }
1241
1242                 /* do not place post remats after jumps */
1243                 if(sched_skip_cf_predicator(irn, si->birg->main_env->arch_env)) {
1244                         del_pset(used);
1245                         del_pset(args);
1246                         break;
1247                 }
1248
1249                 /* insert all possible remats after irn */
1250                 post_remats = pset_new_ptr_default();
1251                 pset_foreach(used, arg) {
1252                         remat_info_t   *remat_info,
1253                                                     query;
1254
1255                         /* continue if the operand has the wrong reg class */
1256                         if(!has_reg_class(si, arg))
1257                                 continue;
1258
1259                         query.irn = arg;
1260                         query.remats = NULL;
1261                         query.remats_by_operand = NULL;
1262                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1263
1264                         if(!remat_info) {
1265                                 continue;
1266                         }
1267
1268                         if(remat_info->remats_by_operand) {
1269                                 pset_foreach(remat_info->remats_by_operand, remat) {
1270                                         /* do not insert remats producing the same value as one of the operands */
1271                                         if(!pset_find_ptr(args, remat->value)) {
1272                                                 DBG((si->dbg, LEVEL_4, "\t  considering remat %+F with arg %+F\n", remat->op, arg));
1273
1274                                                 /* only remat values that can be used by real ops */
1275                                                 if(!opt_remat_while_live || pset_find_ptr(live, remat->value)) {
1276                                                         pset_insert_ptr(post_remats, remat);
1277                                                 }
1278                                         }
1279                                 }
1280                         }
1281                 }
1282                 pset_foreach(post_remats, remat) {
1283                         insert_remat_after(si, remat, irn, live);
1284                 }
1285                 del_pset(post_remats);
1286
1287                 del_pset(used);
1288                 del_pset(args);
1289                 irn = next;
1290         }
1291
1292         /* add remats at end if successor has multiple predecessors */
1293         if(is_merge_edge(bb)) {
1294                 pset     *live_out = pset_new_ptr_default();
1295                 ir_node  *value;
1296
1297                 get_live_end(si, bb, live_out);
1298
1299                 /* add remats at end of block */
1300                 pset_foreach(live_out, value) {
1301                         remat_info_t   *remat_info,
1302                                                    query;
1303
1304                         query.irn = value;
1305                         query.remats = NULL;
1306                         query.remats_by_operand = NULL;
1307                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1308
1309                         if(remat_info && remat_info->remats) {
1310                                 pset_foreach(remat_info->remats, remat) {
1311                                         DBG((si->dbg, LEVEL_4, "\t  considering remat %+F at end of block %+F\n", remat->op, bb));
1312
1313                                         insert_remat_before(si, remat, bb, live_out);
1314                                 }
1315                         }
1316                 }
1317                 del_pset(live_out);
1318         }
1319
1320         if(is_diverge_edge(bb)) {
1321                 pset     *live_in = pset_new_ptr_default();
1322                 ir_node  *value;
1323
1324                 be_lv_foreach(si->lv, bb, be_lv_state_in, i) {
1325                         value = be_lv_get_irn(si->lv, bb, i);
1326
1327                         if(has_reg_class(si, value)) {
1328                                 pset_insert_ptr(live_in, value);
1329                         }
1330                 }
1331                 /* add phis to live_in */
1332                 sched_foreach(bb, value) {
1333                         if(!is_Phi(value)) break;
1334
1335                         if(has_reg_class(si, value)) {
1336                                 pset_insert_ptr(live_in, value);
1337                         }
1338                 }
1339
1340                 /* add remat2s at beginning of block */
1341                 post_remats = pset_new_ptr_default();
1342                 pset_foreach(live_in, value) {
1343                         remat_info_t   *remat_info,
1344                                                    query;
1345
1346                         query.irn = value;
1347                         query.remats = NULL;
1348                         query.remats_by_operand = NULL;
1349                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1350
1351                         if(remat_info && remat_info->remats_by_operand) {
1352                                 pset_foreach(remat_info->remats_by_operand, remat) {
1353                                         DBG((si->dbg, LEVEL_4, "\t  considering remat2 %+F at beginning of block %+F\n", remat->op, bb));
1354
1355                                         /* put the remat here if all its args are available and result is still live */
1356                                         if(!opt_remat_while_live || pset_find_ptr(live_in, remat->value)) {
1357                                                 pset_insert_ptr(post_remats, remat);
1358                                         }
1359                                 }
1360                         }
1361                 }
1362                 pset_foreach(post_remats, remat) {
1363                         insert_remat_after(si, remat, bb, live_in);
1364                 }
1365                 del_pset(post_remats);
1366                 del_pset(live_in);
1367         }
1368 }
1369
1370 static int
1371 can_be_copied(const ir_node * bb, const ir_node * irn)
1372 {
1373         const ir_edge_t *edge    = get_block_succ_first(bb);
1374         const ir_node   *next_bb = edge->src;
1375         int             pos      = edge->pos;
1376         const ir_node   *phi;
1377
1378         assert(is_merge_edge(bb));
1379
1380         sched_foreach(next_bb, phi) {
1381                 const ir_node  *phi_arg;
1382
1383                 if(!is_Phi(phi)) break;
1384
1385                 phi_arg = get_irn_n(phi, pos);
1386
1387                 if(phi_arg == irn) {
1388                         return 1;
1389                 }
1390         }
1391         return 0;
1392 }
1393
1394 /**
1395  * Initialize additional node info
1396  */
1397 static void
1398 luke_initializer(ir_node * bb, void * data)
1399 {
1400         spill_ilp_t    *si = (spill_ilp_t*)data;
1401         spill_bb_t     *spill_bb;
1402         ir_node        *irn;
1403
1404         spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
1405         set_irn_link(bb, spill_bb);
1406
1407         sched_foreach(bb, irn) {
1408                 op_t      *op;
1409
1410                 op = obstack_alloc(si->obst, sizeof(*op));
1411                 op->is_remat = 0;
1412                 op->attr.live_range.ilp = ILP_UNDEF;
1413                 if(is_Phi(irn)) {
1414                         if(opt_memcopies) {
1415                                 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(irn));
1416                                 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(irn));
1417                         }
1418                 } else if(!is_Proj(irn)) {
1419                         op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1420                         memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1421                 } else {
1422                         op->attr.live_range.args.reloads = NULL;
1423                 }
1424                 set_irn_link(irn, op);
1425         }
1426 }
1427
1428
1429 /**
1430  * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1431  */
1432 static void
1433 luke_endwalker(ir_node * bb, void * data)
1434 {
1435         spill_ilp_t    *si = (spill_ilp_t*)data;
1436         pset           *live;
1437         pset           *use_end;
1438         char            buf[256];
1439         ilp_cst_t       cst;
1440         ir_node        *irn;
1441         spill_bb_t     *spill_bb = get_irn_link(bb);
1442         int             i;
1443
1444         live = pset_new_ptr_default();
1445         use_end = pset_new_ptr_default();
1446
1447         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1448                 irn = be_lv_get_irn(si->lv, bb, i);
1449                 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1450                         pset_insert_ptr(live, irn);
1451                 }
1452         }
1453         /*
1454          * find values that are used by remats at end of block
1455          * and insert them into live set
1456          */
1457         foreach_pre_remat(si, bb, irn) {
1458                 int       n;
1459
1460                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1461                         ir_node        *remat_arg = get_irn_n(irn, n);
1462
1463                         if(has_reg_class(si, remat_arg)) {
1464                                 pset_insert_ptr(live, remat_arg);
1465                         }
1466                 }
1467         }
1468
1469         /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1470         /* their reg_out must always be set */
1471         sched_foreach_reverse(bb, irn) {
1472                 int   n;
1473
1474                 if(!sched_skip_cf_predicator(irn, si->birg->main_env->arch_env)) break;
1475
1476                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1477                         ir_node        *irn_arg = get_irn_n(irn, n);
1478
1479                         if(has_reg_class(si, irn_arg)) {
1480                                 pset_insert_ptr(use_end, irn_arg);
1481                         }
1482                 }
1483         }
1484
1485         ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1486         //cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1487         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1488
1489         spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1490
1491         /* if this is a merge edge we can reload at the end of this block */
1492         if(is_merge_edge(bb)) {
1493                 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1494         } else if(pset_count(use_end)){
1495                 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1496         } else {
1497                 spill_bb->reloads = NULL;
1498         }
1499
1500         pset_foreach(live,irn) {
1501                 spill_t     query,
1502                                         *spill;
1503                 double      spill_cost;
1504                 int         default_spilled;
1505
1506
1507                 /* handle values used by control flow nodes later separately */
1508                 if(pset_find_ptr(use_end, irn)) continue;
1509
1510                 query.irn = irn;
1511                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1512
1513                 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1514
1515                 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1516                 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1517                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1518
1519                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1520                 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1521
1522                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1523                 /* by default spill value right after definition */
1524                 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1525                 spill->spill    = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1526
1527                 if(is_merge_edge(bb)) {
1528                         ilp_var_t   reload;
1529                         ilp_cst_t   rel_cst;
1530
1531                         ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1532                         reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), can_be_copied(bb, irn));
1533                         set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1534
1535                         /* reload <= mem_out */
1536                         rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1537                         lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1538                         lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1539                 }
1540
1541                 spill->reg_in = ILP_UNDEF;
1542                 spill->mem_in = ILP_UNDEF;
1543         }
1544
1545         pset_foreach(use_end,irn) {
1546                 spill_t     query,
1547                                         *spill;
1548                 double      spill_cost;
1549                 ilp_cst_t   end_use_req,
1550                                         rel_cst;
1551                 ilp_var_t   reload;
1552                 int         default_spilled;
1553
1554                 query.irn = irn;
1555                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1556
1557                 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1558
1559                 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1560                 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1561
1562                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1563                 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1564
1565                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1566                 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1567                 spill->spill    = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1568
1569                 /* reload for use be control flow op */
1570                 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1571                 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
1572                 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1573
1574                 /* reload <= mem_out */
1575                 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1576                 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1577                 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1578
1579                 spill->reg_in = ILP_UNDEF;
1580                 spill->mem_in = ILP_UNDEF;
1581
1582                 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1583                 end_use_req = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1);
1584                 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1585         }
1586
1587         del_pset(live);
1588         del_pset(use_end);
1589 }
1590
1591 /**
1592  * Find a remat of value @p value in the epilog of @p pos
1593  */
1594 static ir_node *
1595 find_post_remat(const ir_node * value, const ir_node * pos)
1596 {
1597         while((pos = next_post_remat(pos)) != NULL) {
1598                 op_t   *op;
1599
1600                 op = get_irn_link(pos);
1601                 assert(op->is_remat && !op->attr.remat.pre);
1602
1603                 if(op->attr.remat.remat->value == value)
1604                         return (ir_node*)pos;
1605
1606 #if 0
1607         const ir_edge_t *edge;
1608                 foreach_out_edge(pos, edge) {
1609                         ir_node   *proj = get_edge_src_irn(edge);
1610                         assert(is_Proj(proj));
1611                 }
1612 #endif
1613
1614         }
1615
1616         return NULL;
1617 }
1618
1619 static spill_t *
1620 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1621 {
1622         spill_bb_t  *spill_bb = get_irn_link(bb);
1623         spill_t     *spill,
1624                                  query;
1625         char         buf[256];
1626         int          default_spilled;
1627
1628         query.irn = irn;
1629         spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1630         if(!spill) {
1631                 double   spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1632
1633                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1634
1635                 spill->reg_out = ILP_UNDEF;
1636                 spill->reg_in  = ILP_UNDEF;
1637                 spill->mem_in  = ILP_UNDEF;
1638
1639                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1640                 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1641
1642                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1643                 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1644                 spill->spill    = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1645         }
1646
1647         return spill;
1648 }
1649
1650 /**
1651  *  Inserts ILP-constraints and variables for memory copying before the given position
1652  */
1653 static void
1654 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1655 {
1656         const ir_node    *succ;
1657         const ir_edge_t  *edge;
1658         spill_bb_t       *spill_bb = get_irn_link(block);
1659         ir_node          *phi;
1660         int               pos;
1661         ilp_cst_t         cst;
1662         ilp_var_t         copyreg;
1663         char              buf[256];
1664         ir_node          *tmp;
1665
1666
1667         assert(edges_activated(current_ir_graph));
1668
1669         edge = get_block_succ_first(block);
1670         if(!edge) return;
1671
1672         succ = edge->src;
1673         pos = edge->pos;
1674
1675         edge = get_block_succ_next(block, edge);
1676         /* next block can only contain phis, if this is a merge edge */
1677         if(edge) return;
1678
1679         ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1680         copyreg = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1681
1682         ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1683         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1684
1685         pset_foreach(live, tmp) {
1686                 spill_t  *spill;
1687 #if 0
1688                 op_t  *op = get_irn_link(irn);
1689                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1690 #endif
1691                 spill = set_find_spill(spill_bb->ilp, tmp);
1692                 assert(spill);
1693
1694                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1695         }
1696         lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1697
1698         sched_foreach(succ, phi) {
1699                 const ir_node  *to_copy;
1700                 op_t           *to_copy_op;
1701                 spill_t        *to_copy_spill;
1702                 op_t           *phi_op = get_irn_link(phi);
1703                 ilp_var_t       reload = ILP_UNDEF;
1704
1705
1706                 if(!is_Phi(phi)) break;
1707                 if(!has_reg_class(si, phi)) continue;
1708
1709                 to_copy = get_irn_n(phi, pos);
1710                 to_copy_op = get_irn_link(to_copy);
1711
1712                 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1713                 assert(to_copy_spill);
1714
1715                 if(spill_bb->reloads) {
1716                         keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1717
1718                         if(keyval) {
1719                                 reload = PTR_TO_INT(keyval->val);
1720                         }
1721                 }
1722
1723                 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N_%N", block, phi, to_copy);
1724                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1725
1726                 /* copy - reg_out - reload - remat - live_range <= 0 */
1727                 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1728                 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1729                 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1730                 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1731                 foreach_pre_remat(si, block, tmp) {
1732                         op_t     *remat_op = get_irn_link(tmp);
1733                         if(remat_op->attr.remat.remat->value == to_copy) {
1734                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1735                         }
1736                 }
1737
1738                 ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N_%N", block, phi, to_copy);
1739                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1740
1741                 /* copy - reg_out - copyreg <= 0 */
1742                 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1743                 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1744                 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1745         }
1746 }
1747
1748
1749 /**
1750  * Walk all irg blocks and emit this ILP
1751  */
1752 static void
1753 luke_blockwalker(ir_node * bb, void * data)
1754 {
1755         spill_ilp_t    *si = (spill_ilp_t*)data;
1756         ir_node        *irn;
1757         pset           *live;
1758         char            buf[256];
1759         ilp_cst_t       cst;
1760         spill_bb_t     *spill_bb = get_irn_link(bb);
1761         ir_node        *tmp;
1762         spill_t        *spill;
1763         pset           *defs = pset_new_ptr_default();
1764         const arch_env_t *arch_env = si->birg->main_env->arch_env;
1765
1766         live = pset_new_ptr_default();
1767
1768         /****************************************
1769          *      B A S I C  B L O C K  E N D
1770          ***************************************/
1771
1772
1773         /* init live values at end of block */
1774         get_live_end(si, bb, live);
1775
1776         pset_foreach(live, irn) {
1777                 op_t           *op;
1778                 ilp_var_t       reload = ILP_UNDEF;
1779
1780                 spill = set_find_spill(spill_bb->ilp, irn);
1781                 assert(spill);
1782
1783                 if(spill_bb->reloads) {
1784                         keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1785
1786                         if(keyval) {
1787                                 reload = PTR_TO_INT(keyval->val);
1788                         }
1789                 }
1790
1791                 op = get_irn_link(irn);
1792                 assert(!op->is_remat);
1793
1794                 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1795                 op->attr.live_range.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1796                 op->attr.live_range.op = bb;
1797
1798                 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1799                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1800
1801                 /* reg_out - reload - remat - live_range <= 0 */
1802                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1803                 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1804                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1805                 foreach_pre_remat(si, bb, tmp) {
1806                         op_t     *remat_op = get_irn_link(tmp);
1807                         if(remat_op->attr.remat.remat->value == irn) {
1808                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1809                         }
1810                 }
1811                 ir_snprintf(buf, sizeof(buf), "reg_out2_%N_%N", bb, irn);
1812                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
1813
1814                 /* value may only die at bb end if it is used for a mem copy */
1815                 /* reg_out + \sum copy - reload - remat - live_range >= 0 */
1816                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1817                 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1818                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1819                 foreach_pre_remat(si, bb, tmp) {
1820                         op_t     *remat_op = get_irn_link(tmp);
1821                         if(remat_op->attr.remat.remat->value == irn) {
1822                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1823                         }
1824                 }
1825                 if(is_merge_edge(bb)) {
1826                         const ir_edge_t *edge = get_block_succ_first(bb);
1827                         const ir_node   *next_bb = edge->src;
1828                         int              pos = edge->pos;
1829                         const ir_node   *phi;
1830
1831                         sched_foreach(next_bb, phi) {
1832                                 const ir_node  *phi_arg;
1833
1834                                 if(!is_Phi(phi)) break;
1835
1836                                 phi_arg = get_irn_n(phi, pos);
1837
1838                                 if(phi_arg == irn) {
1839                                         op_t      *phi_op = get_irn_link(phi);
1840                                         ilp_var_t  copy = phi_op->attr.live_range.args.copies[pos];
1841
1842                                         lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
1843                                 }
1844                         }
1845                 }
1846         }
1847
1848         if(opt_memcopies)
1849                 insert_mem_copy_position(si, live, bb);
1850
1851         /*
1852          * assure the remat args are available
1853          */
1854         foreach_pre_remat(si, bb, tmp) {
1855                 op_t     *remat_op = get_irn_link(tmp);
1856                 int       n;
1857
1858                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1859                         ir_node        *remat_arg = get_irn_n(tmp, n);
1860                         op_t           *arg_op = get_irn_link(remat_arg);
1861
1862                         if(!has_reg_class(si, remat_arg)) continue;
1863
1864                         spill = set_find_spill(spill_bb->ilp, remat_arg);
1865                         assert(spill);
1866
1867                         /* arguments of remats have to be live until the very end of the block
1868                          * remat = reg_out(remat_arg) and (reload(remat_arg) or live_range(remat_arg)),
1869                          * no remats, they could be in wrong order
1870                          */
1871
1872                         ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1873                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1874
1875                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 3.0);
1876                         lpp_set_factor_fast(si->lpp, cst, spill->reg_out, -2.0);
1877                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1878
1879                         /* use reload placed for this argument */
1880                         if(spill_bb->reloads) {
1881                                 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1882
1883                                 if(keyval) {
1884                                         ilp_var_t       reload = PTR_TO_INT(keyval->val);
1885
1886                                         lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1887                                 }
1888                         }
1889                 }
1890         }
1891         DBG((si->dbg, LEVEL_4, "\t   %d values live at end of block %+F\n", pset_count(live), bb));
1892
1893
1894
1895
1896         /**************************************
1897          *    B A S I C  B L O C K  B O D Y
1898          **************************************/
1899
1900         sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1901                 op_t       *op;
1902                 op_t       *tmp_op;
1903                 int         n,
1904                                         u = 0,
1905                                         d = 0;
1906                 ilp_cst_t       check_pre,
1907                                         check_post;
1908                 set        *args;
1909                 pset       *used;
1910                 pset       *remat_defs;
1911                 keyval_t   *keyval;
1912                 ilp_cst_t   one_memoperand = -1;
1913
1914                 /* iterate only until first phi */
1915                 if(is_Phi(irn))
1916                         break;
1917
1918                 op = get_irn_link(irn);
1919                 /* skip remats */
1920                 if(op->is_remat) continue;
1921
1922                 DBG((si->dbg, LEVEL_4, "\t  at node %+F\n", irn));
1923
1924                 /* collect defined values */
1925                 if(has_reg_class(si, irn)) {
1926                         pset_insert_ptr(defs, irn);
1927                 }
1928
1929                 /* skip projs */
1930                 if(is_Proj(irn)) continue;
1931
1932                 /*
1933                  * init set of irn's arguments
1934                  * and all possibly used values around this op
1935                  * and values defined by post remats
1936                  */
1937                 args =       new_set(cmp_keyval, get_irn_arity(irn));
1938                 used =       pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1939                 remat_defs = pset_new_ptr(pset_count(live));
1940
1941                 if(!is_start_block(bb) || !be_is_Barrier(irn)) {
1942                         for (n=get_irn_arity(irn)-1; n>=0; --n) {
1943                                 ir_node        *irn_arg = get_irn_n(irn, n);
1944                                 if(has_reg_class(si, irn_arg)) {
1945                                         set_insert_keyval(args, irn_arg, (void*)n);
1946                                         pset_insert_ptr(used, irn_arg);
1947                                 }
1948                         }
1949                         foreach_post_remat(irn, tmp) {
1950                                 op_t    *remat_op = get_irn_link(tmp);
1951
1952                                 pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1953
1954                                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1955                                         ir_node        *remat_arg = get_irn_n(tmp, n);
1956                                         if(has_reg_class(si, remat_arg)) {
1957                                                 pset_insert_ptr(used, remat_arg);
1958                                         }
1959                                 }
1960                         }
1961                         foreach_pre_remat(si, irn, tmp) {
1962                                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1963                                         ir_node        *remat_arg = get_irn_n(tmp, n);
1964                                         if(has_reg_class(si, remat_arg)) {
1965                                                 pset_insert_ptr(used, remat_arg);
1966                                         }
1967                                 }
1968                         }
1969                 }
1970
1971                 /**********************************
1972                  *   I N  E P I L O G  O F  irn
1973                  **********************************/
1974
1975                 /* ensure each dying value is used by only one post remat */
1976                 pset_foreach(used, tmp) {
1977                         ir_node     *value = tmp;
1978                         op_t        *value_op = get_irn_link(value);
1979                         ir_node     *remat;
1980                         int          n_remats = 0;
1981
1982                         cst = ILP_UNDEF;
1983                         foreach_post_remat(irn, remat) {
1984                                 op_t  *remat_op = get_irn_link(remat);
1985
1986                                 for(n=get_irn_arity(remat)-1; n>=0; --n) {
1987                                         ir_node   *remat_arg = get_irn_n(remat, n);
1988
1989                                         /* if value is used by this remat add it to constraint */
1990                                         if(remat_arg == value) {
1991                                                 if(n_remats == 0) {
1992                                                         /* sum remat2s <= 1 + n_remats*live_range */
1993                                                         ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
1994                                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
1995                                                 }
1996
1997                                                 n_remats++;
1998                                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1999                                                 break;
2000                                         }
2001                                 }
2002                         }
2003
2004                         if(pset_find_ptr(live, value) && cst != ILP_UNDEF) {
2005                                 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
2006                         }
2007                 }
2008
2009         /* ensure at least one value dies at post remat */
2010         foreach_post_remat(irn, tmp) {
2011             op_t     *remat_op = get_irn_link(tmp);
2012             pset     *remat_args = pset_new_ptr(get_irn_arity(tmp));
2013             ir_node  *remat_arg;
2014
2015             for(n=get_irn_arity(tmp)-1; n>=0; --n) {
2016                 remat_arg = get_irn_n(tmp, n);
2017
2018                 if(has_reg_class(si, remat_arg)) {
2019
2020                     /* does arg always die at this op? */
2021                     if(!pset_find_ptr(live, remat_arg))
2022                         goto skip_one_must_die;
2023
2024                     pset_insert_ptr(remat_args, remat_arg);
2025                 }
2026             }
2027
2028             /* remat + \sum live_range(remat_arg) <= |args| */
2029             ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp);
2030             cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2031             lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2032
2033             pset_foreach(remat_args, remat_arg) {
2034                 op_t  *arg_op = get_irn_link(remat_arg);
2035
2036                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2037             }
2038
2039 skip_one_must_die:
2040             del_pset(remat_args);
2041         }
2042
2043                 /* new live ranges for values from L\U defined by post remats */
2044                 pset_foreach(live, tmp) {
2045                         ir_node     *value = tmp;
2046                         op_t        *value_op = get_irn_link(value);
2047
2048                         if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
2049                                 ilp_var_t    prev_lr = ILP_UNDEF;
2050                                 ir_node     *remat;
2051
2052                                 if(pset_find_ptr(remat_defs, value)) {
2053
2054                                         /* next_live_range <= prev_live_range + sum remat2s */
2055                                         ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
2056                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2057
2058                                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
2059                                         prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2060
2061                                         lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
2062                                         lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2063
2064                                         foreach_post_remat(irn, remat) {
2065                                                 op_t        *remat_op = get_irn_link(remat);
2066
2067                                                 /* if value is being rematerialized by this remat */
2068                                                 if(value == remat_op->attr.remat.remat->value) {
2069                                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2070                                                 }
2071                                         }
2072
2073                                         value_op->attr.live_range.ilp = prev_lr;
2074                                         value_op->attr.live_range.op = irn;
2075                                 }
2076                         }
2077                 }
2078
2079                 /* requirements for post remats and start live ranges from L/U' for values dying here */
2080                 foreach_post_remat(irn, tmp) {
2081                         op_t        *remat_op = get_irn_link(tmp);
2082                         int          n;
2083
2084                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2085                                 ir_node        *remat_arg = get_irn_n(tmp, n);
2086                                 op_t           *arg_op = get_irn_link(remat_arg);
2087
2088                                 if(!has_reg_class(si, remat_arg)) continue;
2089
2090                                 /* only for values in L\U (TODO and D?), the others are handled with post_use */
2091                                 if(!pset_find_ptr(used, remat_arg)) {
2092                                         /* remat <= live_range(remat_arg) */
2093                                         ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
2094                                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2095
2096                                         /* if value is becoming live through use by remat2 */
2097                                         if(!pset_find_ptr(live, remat_arg)) {
2098                                                 ilp_var_t     lr;
2099
2100                                                 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
2101                                                 lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2102
2103                                                 arg_op->attr.live_range.ilp = lr;
2104                                                 arg_op->attr.live_range.op = irn;
2105
2106                                                 DBG((si->dbg, LEVEL_3, "  value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
2107
2108                                                 pset_insert_ptr(live, remat_arg);
2109                                                 add_to_spill_bb(si, bb, remat_arg);
2110                                         }
2111
2112                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2113                                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2114                                 }
2115                         }
2116                 }
2117
2118                 d = pset_count(defs);
2119                 DBG((si->dbg, LEVEL_4, "\t   %+F produces %d values in my register class\n", irn, d));
2120
2121                 /* count how many regs irn needs for arguments */
2122                 u = set_count(args);
2123
2124
2125                 /* check the register pressure in the epilog */
2126                 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
2127                 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
2128                 check_post = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - d);
2129
2130                 /* add L\U' to check_post */
2131                 pset_foreach(live, tmp) {
2132                         if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
2133                                 /* if a live value is not used by irn */
2134                                 tmp_op = get_irn_link(tmp);
2135                                 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
2136                         }
2137                 }
2138
2139                 /***********************************************************
2140                  *  I T E R A T I O N  O V E R  U S E S  F O R  E P I L O G
2141                  **********************************************************/
2142
2143
2144                 pset_foreach(used, tmp) {
2145                         ilp_var_t       prev_lr;
2146                         ilp_var_t       post_use;
2147                         int             p = 0;
2148                         spill_t        *spill;
2149                         ir_node        *arg = tmp;
2150                         op_t           *arg_op = get_irn_link(arg);
2151                         ir_node        *remat;
2152
2153                         spill = add_to_spill_bb(si, bb, arg);
2154
2155                         /* new live range for each used value */
2156                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
2157                         prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2158
2159                         /* the epilog stuff - including post_use, check_post, check_post_remat */
2160                         ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
2161                         post_use = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2162
2163                         lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
2164
2165                         /* arg is live throughout epilog if the next live_range is in a register */
2166                         if(pset_find_ptr(live, arg)) {
2167                                 DBG((si->dbg, LEVEL_3, "\t  arg %+F is possibly live in epilog of %+F\n", arg, irn));
2168
2169                                 /* post_use >= next_lr + remat */
2170                                 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2171                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2172                                 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2173                                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2174                         }
2175
2176                         /* forall post remat which use arg add a similar cst */
2177                         foreach_post_remat(irn, remat) {
2178                                 int      n;
2179
2180                                 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2181                                         ir_node    *remat_arg = get_irn_n(remat, n);
2182                                         op_t       *remat_op = get_irn_link(remat);
2183
2184                                         if(remat_arg == arg) {
2185                                                 DBG((si->dbg, LEVEL_3, "\t  found remat with arg %+F in epilog of %+F\n", arg, irn));
2186
2187                                                 /* post_use >= remat */
2188                                                 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2189                                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2190                                                 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2191                                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2192                                         }
2193                                 }
2194                         }
2195
2196                         /* if value is not an arg of op and not possibly defined by post remat
2197                          * then it may only die and not become live
2198                          */
2199                         if(!set_find_keyval(args, arg)) {
2200                                 /* post_use <= prev_lr */
2201                                 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
2202                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2203                                 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2204                                 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2205
2206                                 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
2207                                         /* next_lr <= prev_lr */
2208                                         ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
2209                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2210                                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2211                                         lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2212                                 }
2213                         }
2214
2215                         if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2216                                 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2217                                         if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2218                                                 ilp_var_t       memoperand;
2219
2220                                                 ir_snprintf(buf, sizeof(buf), "memoperand_%N_%d", irn, n);
2221                                                 memoperand = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_memoperand*execution_frequency(si, bb), 0.0);
2222                                                 set_insert_memoperand(si->memoperands, irn, n, memoperand);
2223
2224                                                 ir_snprintf(buf, sizeof(buf), "nolivepost_%N_%d", irn, n);
2225                                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2226
2227                                                 lpp_set_factor_fast(si->lpp, cst, memoperand, 1.0);
2228                                                 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2229                                         }
2230                                 }
2231                         }
2232
2233                         /* new live range begins for each used value */
2234                         arg_op->attr.live_range.ilp = prev_lr;
2235                         arg_op->attr.live_range.op = irn;
2236
2237                         pset_insert_ptr(live, arg);
2238                 }
2239
2240                 /* just to be sure */
2241                 check_post = ILP_UNDEF;
2242
2243                 /* allow original defintions to be removed */
2244                 if(opt_repair_schedule) {
2245                         pset_foreach(defs, tmp) {
2246                                 op_t      *tmp_op = get_irn_link(tmp);
2247                                 spill_t   *spill = set_find_spill(spill_bb->ilp, tmp);
2248 #if 1
2249                                 ilp_var_t  delete;
2250                                 assert(spill);
2251
2252                                 ir_snprintf(buf, sizeof(buf), "delete_%N", tmp);
2253                                 delete = lpp_add_var_default(si->lpp, buf, lpp_binary, -1.0*get_cost(si, irn)*execution_frequency(si, bb), 0.0);
2254
2255                                 /* op may not be killed if its first live_range is 1 */
2256                                 ir_snprintf(buf, sizeof(buf), "killorig-lr_%N", tmp);
2257                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2258                                 lpp_set_factor_fast(si->lpp, cst, delete, 1.0);
2259                                 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, 1.0);
2260
2261                                 /* op may not be killed if it is spilled after the definition */
2262                                 ir_snprintf(buf, sizeof(buf), "killorig-spill_%N", tmp);
2263                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2264                                 lpp_set_factor_fast(si->lpp, cst, delete, 1.0);
2265                                 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2266 #else
2267                                 ilp_var_t  keep;
2268                                 assert(spill);
2269
2270                                 ir_snprintf(buf, sizeof(buf), "keep_%N", tmp);
2271                                 keep = lpp_add_var_default(si->lpp, buf, lpp_binary, get_cost(si, irn)*execution_frequency(si, bb), 1.0);
2272
2273                                 /* op may not be killed if its first live_range is 1 */
2274                                 ir_snprintf(buf, sizeof(buf), "killorig-lr_%N", tmp);
2275                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
2276                                 lpp_set_factor_fast(si->lpp, cst, keep, 1.0);
2277                                 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, -1.0);
2278
2279                                 /* op may not be killed if it is spilled after the definition */
2280                                 ir_snprintf(buf, sizeof(buf), "killorig-spill_%N", tmp);
2281                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
2282                                 lpp_set_factor_fast(si->lpp, cst, keep, 1.0);
2283                                 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2284 #endif
2285                         }
2286                 } else {
2287 #if 0
2288                         pset_foreach(defs, tmp) {
2289                                 op_t      *tmp_op = get_irn_link(tmp);
2290                                 spill_t   *spill = set_find_spill(spill_bb->ilp, tmp);
2291                                 assert(spill);
2292
2293                                 /* live_range or spill should be 1
2294                                    TODO: lr should be live until first use */
2295                                 ir_snprintf(buf, sizeof(buf), "nokillorig_%N", tmp);
2296                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 1.0);
2297                                 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, 1.0);
2298                                 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2299                         }
2300 #endif
2301                 }
2302
2303
2304                 /******************
2305                  *   P R O L O G
2306                  ******************/
2307
2308                 /* check the register pressure in the prolog */
2309                 /* sum_{L\U} lr <= k - |U| */
2310                 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
2311                 check_pre = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - u);
2312
2313                 /* for the prolog remove defined values from the live set */
2314                 pset_foreach(defs, tmp) {
2315                         pset_remove_ptr(live, tmp);
2316                 }
2317
2318                 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2319                         ir_snprintf(buf, sizeof(buf), "one_memoperand_%N", irn);
2320                         one_memoperand = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2321                 }
2322
2323                 /***********************************************************
2324                  *  I T E R A T I O N  O V E R  A R G S  F O R  P R O L O G
2325                  **********************************************************/
2326
2327
2328                 set_foreach(args, keyval) {
2329                         spill_t          *spill;
2330                         const ir_node    *arg = keyval->key;
2331                         int               i = PTR_TO_INT(keyval->val);
2332                         op_t             *arg_op = get_irn_link(arg);
2333                         ilp_cst_t         requirements;
2334                         int               n_memoperands;
2335
2336                         spill = set_find_spill(spill_bb->ilp, arg);
2337                         assert(spill);
2338
2339                         ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
2340                         op->attr.live_range.args.reloads[i] = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
2341
2342                         /* reload <= mem_out */
2343                         ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
2344                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2345                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2346                         lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2347
2348                         /* requirement: arg must be in register for use */
2349                         /* reload + remat + live_range == 1 */
2350                         ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
2351                         requirements = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1.0);
2352
2353                         lpp_set_factor_fast(si->lpp, requirements, arg_op->attr.live_range.ilp, 1.0);
2354                         lpp_set_factor_fast(si->lpp, requirements, op->attr.live_range.args.reloads[i], 1.0);
2355                         foreach_pre_remat(si, irn, tmp) {
2356                                 op_t     *remat_op = get_irn_link(tmp);
2357                                 if(remat_op->attr.remat.remat->value == arg) {
2358                                         lpp_set_factor_fast(si->lpp, requirements, remat_op->attr.remat.ilp, 1.0);
2359                                 }
2360                         }
2361
2362                         if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2363                                 n_memoperands = 0;
2364                                 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2365                                         if(get_irn_n(irn, n) == arg) {
2366                                                 n_memoperands++;
2367                                         }
2368                                 }
2369                                 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2370                                         if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2371                                                 memoperand_t  *memoperand;
2372                                                 memoperand = set_find_memoperand(si->memoperands, irn, n);
2373
2374                                                 /* memoperand <= mem_out */
2375                                                 ir_snprintf(buf, sizeof(buf), "req_memoperand_%N_%d", irn, n);
2376                                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2377                                                 lpp_set_factor_fast(si->lpp, cst, memoperand->ilp, 1.0);
2378                                                 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2379
2380                                                 /* the memoperand is only sufficient if it is used once by the op */
2381                                                 if(n_memoperands == 1)
2382                                                         lpp_set_factor_fast(si->lpp, requirements, memoperand->ilp, 1.0);
2383
2384                                                 lpp_set_factor_fast(si->lpp, one_memoperand, memoperand->ilp, 1.0);
2385
2386                                                 /* we have one more free register if we use a memory operand */
2387                                                 lpp_set_factor_fast(si->lpp, check_pre, memoperand->ilp, -1.0);
2388                                         }
2389                                 }
2390                         }
2391                 }
2392
2393                 /* iterate over L\U */
2394                 pset_foreach(live, tmp) {
2395                         if(!set_find_keyval(args, tmp)) {
2396                                 /* if a live value is not used by irn */
2397                                 tmp_op = get_irn_link(tmp);
2398                                 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
2399                         }
2400                 }
2401
2402                 /* requirements for remats */
2403                 foreach_pre_remat(si, irn, tmp) {
2404                         op_t        *remat_op = get_irn_link(tmp);
2405                         int          n;
2406
2407                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2408                                 ir_node        *remat_arg = get_irn_n(tmp, n);
2409                                 op_t           *arg_op = get_irn_link(remat_arg);
2410
2411                                 if(!has_reg_class(si, remat_arg)) continue;
2412
2413                                 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2414                                 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2415                                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2416
2417                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2418                                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2419
2420                                 /* if remat arg is also used by current op then we can use reload placed for this argument */
2421                                 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2422                                         int    index = (int)keyval->val;
2423
2424                                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2425                                 }
2426                         }
2427                 }
2428
2429
2430
2431
2432                 /*************************
2433                  *  D O N E  W I T H  O P
2434                  *************************/
2435
2436                 DBG((si->dbg, LEVEL_4, "\t   %d values live at %+F\n", pset_count(live), irn));
2437
2438                 pset_foreach(live, tmp) {
2439                         assert(has_reg_class(si, tmp));
2440                 }
2441
2442                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2443                         ir_node        *arg = get_irn_n(irn, n);
2444
2445                         assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2446                 }
2447
2448                 del_pset(remat_defs);
2449                 del_pset(used);
2450                 del_set(args);
2451                 del_pset(defs);
2452                 defs = pset_new_ptr_default();
2453
2454                 /* skip everything above barrier in start block */
2455                 if(is_start_block(bb) && be_is_Barrier(irn)) {
2456                         assert(pset_count(live) == 0);
2457                         break;
2458                 }
2459
2460         }
2461         del_pset(defs);
2462
2463
2464
2465         /***************************************
2466          *   B E G I N N I N G  O F  B L O C K
2467          ***************************************/
2468
2469
2470         /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2471         DBG((si->dbg, LEVEL_3, "\t   %d values live at beginning of block %+F\n", pset_count(live), bb));
2472
2473         pset_foreach(live, irn) {
2474                 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2475         }
2476
2477         /* construct mem_outs for all values */
2478         set_foreach(spill_bb->ilp, spill) {
2479                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2480                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2481
2482                 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2483                 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2484
2485                 if(pset_find_ptr(live, spill->irn)) {
2486                         int default_spilled;
2487                         DBG((si->dbg, LEVEL_5, "\t     %+F live at beginning of block %+F\n", spill->irn, bb));
2488
2489                         ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2490                         default_spilled = be_is_live_in(si->lv, bb, spill->irn) || is_Phi(spill->irn);
2491                         spill->mem_in   = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, default_spilled);
2492                         lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2493
2494                         if(opt_memcopies && is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2495                                 int   n;
2496                                 op_t *op = get_irn_link(spill->irn);
2497
2498                                 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2499                                         const ir_node  *arg = get_irn_n(spill->irn, n);
2500                                         double          freq=0.0;
2501                                         int             m;
2502                                         ilp_var_t       var;
2503
2504
2505                                         /* argument already done? */
2506                                         if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2507
2508                                         /* get sum of execution frequencies of blocks with the same phi argument */
2509                                         for(m=n; m>=0; --m) {
2510                                                 const ir_node  *arg2 = get_irn_n(spill->irn, m);
2511
2512                                                 if(arg==arg2) {
2513                                                         freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2514                                                 }
2515                                         }
2516
2517                                         /* copies are not for free */
2518                                         ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2519                                         var = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_spill * freq, 1.0);
2520
2521                                         for(m=n; m>=0; --m) {
2522                                                 const ir_node  *arg2 = get_irn_n(spill->irn, m);
2523
2524                                                 if(arg==arg2) {
2525                                                         op->attr.live_range.args.copies[m] = var;
2526                                                 }
2527                                         }
2528
2529 #if 0
2530                                         /* copy <= mem_in */
2531                                         ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2532                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2533                                         lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2534                                         lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2535 #endif
2536                                 }
2537                         }
2538                 }
2539         }
2540
2541         foreach_post_remat(bb, tmp) {
2542                 int         n;
2543                 op_t       *remat_op = get_irn_link(tmp);
2544                 pset       *remat_args = pset_new_ptr(get_irn_arity(tmp));
2545                 ir_node    *remat_arg;
2546
2547                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2548                         remat_arg = get_irn_n(tmp, n);
2549
2550                         if(has_reg_class(si, remat_arg)) {
2551                                 pset_insert_ptr(remat_args, remat_arg);
2552                         }
2553                 }
2554
2555                 /* remat + \sum live_range(remat_arg) <= |args| */
2556                 ir_snprintf(buf, sizeof(buf), "one_must_die_%N", tmp);
2557                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2558                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2559
2560                 pset_foreach(remat_args, remat_arg) {
2561                         if(pset_find_ptr(live, remat_arg)) {
2562                                 op_t       *remat_arg_op = get_irn_link(remat_arg);
2563                                 lpp_set_factor_fast(si->lpp, cst, remat_arg_op->attr.live_range.ilp, 1.0);
2564                         }
2565                 }
2566                 del_pset(remat_args);
2567         }
2568
2569         foreach_post_remat(bb, tmp) {
2570                 int  n;
2571
2572                 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
2573                         ir_node  *remat_arg = get_irn_n(tmp, n);
2574
2575                         /* if value is becoming live through use by remat2 */
2576                         if(has_reg_class(si, remat_arg) && !pset_find_ptr(live, remat_arg)) {
2577                                 op_t       *remat_arg_op = get_irn_link(remat_arg);
2578                                 ilp_cst_t   nomem;
2579
2580                                 DBG((si->dbg, LEVEL_3, "  value %+F becoming live through use by remat2 at bb start %+F\n", remat_arg, tmp));
2581
2582                                 pset_insert_ptr(live, remat_arg);
2583                                 spill = add_to_spill_bb(si, bb, remat_arg);
2584                                 remat_arg_op->attr.live_range.ilp = ILP_UNDEF;
2585
2586                                 /* we need reg_in and mem_in for this value; they will be referenced later */
2587                                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", remat_arg, bb);
2588                                 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2589                                 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", remat_arg, bb);
2590                                 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2591
2592
2593                                 /* optimization: all memory stuff should be 0, for we do not want to insert reloads for remats */
2594                                 ir_snprintf(buf, sizeof(buf), "nomem_%N_%N", remat_arg, bb);
2595                                 nomem = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0.0);
2596                                 lpp_set_factor_fast(si->lpp, nomem, spill->spill, 1.0);
2597                         }
2598                 }
2599         }
2600
2601         /* L\U is empty at bb start */
2602         /* arg is live throughout epilog if it is reg_in into this block */
2603
2604         /* check the register pressure at the beginning of the block
2605          * including remats
2606          */
2607         /* reg_in entspricht post_use */
2608
2609         ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2610         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
2611
2612         pset_foreach(live, irn) {
2613         ilp_cst_t  nospill;
2614
2615                 spill = set_find_spill(spill_bb->ilp, irn);
2616                 assert(spill);
2617
2618                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2619                 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2620
2621                 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2622
2623                 /* spill + mem_in <= 1 */
2624                 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2625                 nospill = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2626
2627                 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2628                 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2629
2630         } /* post_remats are NOT included in register pressure check because
2631            they do not increase regpressure */
2632
2633         /* mem_in/reg_in for live_in values, especially phis and their arguments */
2634         pset_foreach(live, irn) {
2635                 int          p = 0,
2636                                          n;
2637
2638                 spill = set_find_spill(spill_bb->ilp, irn);
2639                 assert(spill && spill->irn == irn);
2640
2641                 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2642                         for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2643                                 ilp_cst_t       mem_in,
2644                                                                 reg_in;
2645                                 ir_node        *phi_arg = get_Phi_pred(irn, n);
2646                                 ir_node        *bb_p = get_Block_cfgpred_block(bb, n);
2647                                 spill_bb_t     *spill_bb_p = get_irn_link(bb_p);
2648                                 spill_t        *spill_p;
2649                                 op_t           *op = get_irn_link(irn);
2650
2651                                 /* although the phi is in the right regclass one or more of
2652                                  * its arguments can be in a different one or at least to
2653                                  * ignore
2654                                  */
2655                                 if(has_reg_class(si, phi_arg)) {
2656                                         /* mem_in < mem_out_arg + copy */
2657                                         ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2658                                         mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2659
2660                                         /* reg_in < reg_out_arg */
2661                                         ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2662                                         reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2663
2664                                         lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2665                                         lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2666
2667                                         spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2668                                         assert(spill_p);
2669
2670                                         lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2671                                         if(opt_memcopies)
2672                                                 lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0);
2673
2674                                         lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2675                                 }
2676                         }
2677                 } else {
2678                         /* else assure the value arrives on all paths in the same resource */
2679
2680                         for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2681                                 ilp_cst_t       mem_in,
2682                                                                 reg_in;
2683                                 ir_node        *bb_p = get_Block_cfgpred_block(bb, n);
2684                                 spill_bb_t     *spill_bb_p = get_irn_link(bb_p);
2685                                 spill_t        *spill_p;
2686
2687                                 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2688                                 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2689                                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2690                                 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2691
2692                                 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2693                                 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2694
2695                                 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2696                                 assert(spill_p);
2697
2698                                 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2699                                 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2700                         }
2701                 }
2702         }
2703
2704         foreach_post_remat(bb, tmp) {
2705                 int         n;
2706
2707                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2708                         ir_node    *remat_arg = get_irn_n(tmp, n);
2709                         op_t       *remat_op = get_irn_link(tmp);
2710
2711                         if(!has_reg_class(si, remat_arg)) continue;
2712
2713                         spill = set_find_spill(spill_bb->ilp, remat_arg);
2714                         assert(spill);
2715
2716                         ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2717                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2718                         lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2719                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2720                 }
2721         }
2722
2723         pset_foreach(live, irn) {
2724                 const op_t      *op = get_irn_link(irn);
2725                 const ir_node   *remat;
2726                 int              n_remats = 0;
2727
2728                 cst = ILP_UNDEF;
2729
2730                 foreach_post_remat(bb, remat) {
2731                         int   n;
2732
2733                         for (n=get_irn_arity(remat)-1; n>=0; --n) {
2734                                 const ir_node  *arg = get_irn_n(remat, n);
2735
2736                                 if(arg == irn) {
2737                                         const op_t   *remat_op = get_irn_link(remat);
2738
2739                                         if(cst == ILP_UNDEF) {
2740                                                 /* sum remat2s <= 1 + n_remats*live_range */
2741                                                 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", irn, bb);
2742                                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2743                                         }
2744                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2745                                         ++n_remats;
2746                                         break;
2747                                 }
2748                         }
2749                 }
2750                 if(cst != ILP_UNDEF && op->attr.live_range.ilp != ILP_UNDEF) {
2751                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -n_remats);
2752                 }
2753         }
2754
2755         /* first live ranges from reg_ins */
2756         pset_foreach(live, irn) {
2757                 op_t      *op = get_irn_link(irn);
2758
2759                 if(op->attr.live_range.ilp != ILP_UNDEF) {
2760
2761                         spill = set_find_spill(spill_bb->ilp, irn);
2762                         assert(spill && spill->irn == irn);
2763
2764                         ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2765                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2766                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2767                         lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2768
2769                         foreach_post_remat(bb, tmp) {
2770                                 op_t     *remat_op = get_irn_link(tmp);
2771
2772                                 if(remat_op->attr.remat.remat->value == irn) {
2773                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2774                                 }
2775                         }
2776                 }
2777         }
2778
2779         /* walk forward now and compute constraints for placing spills */
2780         /* this must only be done for values that are not defined in this block */
2781         pset_foreach(live, irn) {
2782                 /*
2783                  * if value is defined in this block we can anways place the spill directly after the def
2784                  *    -> no constraint necessary
2785                  */
2786                 if(!is_Phi(irn) && get_nodes_block(irn) == bb) {
2787                         assert(0);
2788                 }
2789
2790
2791                 spill = set_find_spill(spill_bb->ilp, irn);
2792                 assert(spill);
2793
2794                 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2795                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2796
2797                 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2798                 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2799
2800                 if(!is_Phi(irn)) {
2801                         sched_foreach_op(bb, tmp) {
2802                                 op_t   *op = get_irn_link(tmp);
2803
2804                                 if(is_Phi(tmp)) continue;
2805                                 assert(!is_Proj(tmp));
2806
2807                                 if(op->is_remat) {
2808                                         const ir_node   *value = op->attr.remat.remat->value;
2809
2810                                         if(value == irn) {
2811                                                 /* only collect remats up to the first real use of a value */
2812                                                 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2813                                         }
2814                                 } else {
2815                                         int   n;
2816
2817                                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2818                                                 ir_node    *arg = get_irn_n(tmp, n);
2819
2820                                                 if(arg == irn) {
2821                                                         /* if a value is used stop collecting remats */
2822                             goto next_live;
2823                                                 }
2824                                         }
2825                                 }
2826                         }
2827                 }
2828 next_live: ;
2829         }
2830
2831         del_pset(live);
2832 }
2833
2834 typedef struct _irnlist_t {
2835         struct list_head   list;
2836         ir_node           *irn;
2837 } irnlist_t;
2838
2839 typedef struct _interference_t {
2840         struct list_head    blocklist;
2841         ir_node            *a;
2842         ir_node            *b;
2843 } interference_t;
2844
2845 static int
2846 cmp_interference(const void *a, const void *b, size_t size)
2847 {
2848         const interference_t *p = a;
2849         const interference_t *q = b;
2850
2851         return !(p->a == q->a && p->b == q->b);
2852 }
2853
2854 static interference_t *
2855 set_find_interference(set * set, ir_node * a, ir_node * b)
2856 {
2857         interference_t     query;
2858
2859         query.a = (a>b)?a:b;
2860         query.b = (a>b)?b:a;
2861
2862         return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2863 }
2864
2865 static interference_t *
2866 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2867 {
2868         interference_t     query,
2869                                           *result;
2870         irnlist_t         *list = obstack_alloc(si->obst, sizeof(*list));
2871
2872         list->irn = bb;
2873
2874         result = set_find_interference(set, a, b);
2875         if(result) {
2876
2877                 list_add(&list->list, &result->blocklist);
2878                 return result;
2879         }
2880
2881         query.a = (a>b)?a:b;
2882         query.b = (a>b)?b:a;
2883
2884         result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2885
2886         INIT_LIST_HEAD(&result->blocklist);
2887         list_add(&list->list, &result->blocklist);
2888
2889         return result;
2890 }
2891
2892 static int
2893 values_interfere_in_block(const spill_ilp_t * si, const ir_node * bb, const ir_node * a, const ir_node * b)
2894 {
2895         const ir_edge_t *edge;
2896
2897         if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2898                 /* both values are live in, so they interfere */
2899                 return 1;
2900         }
2901
2902         /* ensure a dominates b */
2903         if(value_dominates(b,a)) {
2904                 const ir_node * t;
2905                 t = b;
2906                 b = a;
2907                 a = t;
2908         }
2909         assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2910
2911
2912         /* the following code is stolen from bera.c */
2913         if(be_is_live_end(si->lv, bb, a))
2914                 return 1;
2915
2916         foreach_out_edge(a, edge) {
2917                 const ir_node *user = edge->src;
2918                 if(get_nodes_block(user) == bb
2919                                 && !is_Phi(user)
2920                                 && b != user
2921                                 && !pset_find_ptr(si->inverse_ops, user)
2922                                 && value_dominates(b, user))
2923                         return 1;
2924         }
2925
2926         return 0;
2927 }
2928
2929 /**
2930  * Walk all irg blocks and collect interfering values inside of phi classes
2931  */
2932 static void
2933 luke_interferencewalker(ir_node * bb, void * data)
2934 {
2935         spill_ilp_t    *si = (spill_ilp_t*)data;
2936         int             l1, l2;
2937
2938         be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
2939                 ir_node        *a = be_lv_get_irn(si->lv, bb, l1);
2940                 op_t           *a_op = get_irn_link(a);
2941
2942
2943                 /* a is only interesting if it is in my register class and if it is inside a phi class */
2944                 if (has_reg_class(si, a) && get_phi_class(si->pc, a)) {
2945                         if (a_op->is_remat || pset_find_ptr(si->inverse_ops, a))
2946                                 continue;
2947
2948                         for (l2 = _be_lv_next_irn(si->lv, bb, 0xff, l1 + 1); l2 >= 0; l2 = _be_lv_next_irn(si->lv, bb, 0xff, l2 + 1)) {
2949                                 ir_node *b    = be_lv_get_irn(si->lv, bb, l2);
2950                                 op_t    *b_op = get_irn_link(b);
2951
2952                                 /* a and b are only interesting if they are in the same phi class */
2953                                 if (has_reg_class(si, b) && get_phi_class(si->pc, a) == get_phi_class(si->pc, b)) {
2954                                         if (b_op->is_remat || pset_find_ptr(si->inverse_ops, b))
2955                                                 continue;
2956
2957                                         if (values_interfere_in_block(si, bb, a, b)) {
2958                                                 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2959                                                 set_insert_interference(si, si->interferences, a, b, bb);
2960                                         }
2961                                 }
2962                         }
2963                 }
2964         }
2965 }
2966
2967 static unsigned int copy_path_id = 0;
2968
2969 static void
2970 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2971 {
2972         ilp_cst_t  cst;
2973         ilp_var_t  copy;
2974         char       buf[256];
2975         void      *ptr;
2976
2977         ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
2978         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2979
2980         lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
2981
2982         pset_foreach(copies, ptr) {
2983                 copy = PTR_TO_INT(ptr);
2984                 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
2985         }
2986 }
2987
2988 /**
2989  * @parameter copies   contains a path of copies which lead us to irn
2990  * @parameter visited  contains a set of nodes already visited on this path
2991  */
2992 static int
2993 find_copy_path(spill_ilp_t * si, const ir_node * irn, const ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
2994 {
2995         const ir_edge_t *edge;
2996         op_t            *op = get_irn_link(irn);
2997     pset            *visited_users = pset_new_ptr_default();
2998         int              paths = 0;
2999
3000         if(op->is_remat) return 0;
3001
3002         pset_insert_ptr(visited, irn);
3003
3004         if(is_Phi(irn)) {
3005                 int    n;
3006         pset  *visited_operands = pset_new_ptr(get_irn_arity(irn));
3007
3008                 /* visit all operands */
3009                 for(n=get_irn_arity(irn)-1; n>=0; --n) {
3010                         ir_node  *arg = get_irn_n(irn, n);
3011                         ilp_var_t  copy = op->attr.live_range.args.copies[n];
3012
3013                         if(!has_reg_class(si, arg)) continue;
3014             if(pset_find_ptr(visited_operands, arg)) continue;
3015             pset_insert_ptr(visited_operands, arg);
3016
3017                         if(arg == target) {
3018                                 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
3019                                         del_pset(visited_operands);
3020                                         del_pset(visited_users);
3021                                         pset_remove_ptr(visited, irn);
3022                                         return paths;
3023                                 }
3024                                 pset_insert(copies, INT_TO_PTR(copy), copy);
3025                                 write_copy_path_cst(si, copies, any_interfere);
3026                                 pset_remove(copies, INT_TO_PTR(copy), copy);
3027                         } else if(!pset_find_ptr(visited, arg)) {
3028                                 pset_insert(copies, INT_TO_PTR(copy), copy);
3029                                 paths += find_copy_path(si, arg, target, any_interfere, copies, visited);
3030                                 pset_remove(copies, INT_TO_PTR(copy), copy);
3031
3032                 if(paths > MAX_PATHS) {
3033                     if(pset_count(copies) == 0) {
3034                         ilp_cst_t  cst;
3035                         char       buf[256];
3036
3037                         ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
3038                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
3039                         lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3040                         lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
3041                         DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
3042
3043                         paths = 0;
3044                     } else {
3045                         del_pset(visited_operands);
3046                         del_pset(visited_users);
3047                         pset_remove_ptr(visited, irn);
3048                         return paths;
3049                     }
3050                 } else if(pset_count(copies) == 0) {
3051                                         paths = 0;
3052                                 }
3053                         }
3054                 }
3055
3056         del_pset(visited_operands);
3057         }
3058
3059         /* visit all uses which are phis */
3060         foreach_out_edge(irn, edge) {
3061                 ir_node  *user = edge->src;
3062                 int       pos  = edge->pos;
3063                 op_t     *op = get_irn_link(user);
3064                 ilp_var_t copy;
3065
3066                 if(!is_Phi(user)) continue;
3067                 if(!has_reg_class(si, user)) continue;
3068         if(pset_find_ptr(visited_users, user)) continue;
3069         pset_insert_ptr(visited_users, user);
3070
3071                 copy = op->attr.live_range.args.copies[pos];
3072
3073                 if(user == target) {
3074                         if(++paths > MAX_PATHS && pset_count(copies) != 0) {
3075                                 del_pset(visited_users);
3076                                 pset_remove_ptr(visited, irn);
3077                                 return paths;
3078                         }
3079                         pset_insert(copies, INT_TO_PTR(copy), copy);
3080                         write_copy_path_cst(si, copies, any_interfere);
3081                         pset_remove(copies, INT_TO_PTR(copy), copy);
3082                 } else if(!pset_find_ptr(visited, user)) {
3083                         pset_insert(copies, INT_TO_PTR(copy), copy);
3084                         paths += find_copy_path(si, user, target, any_interfere, copies, visited);
3085                         pset_remove(copies, INT_TO_PTR(copy), copy);
3086
3087             if(paths > MAX_PATHS) {
3088                 if(pset_count(copies) == 0) {
3089                     ilp_cst_t  cst;
3090                     char       buf[256];
3091
3092                     ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
3093                     cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
3094                     lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3095                     lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
3096                     DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
3097
3098                     paths = 0;
3099                 } else {
3100                     del_pset(visited_users);
3101                     pset_remove_ptr(visited, irn);
3102                     return paths;
3103                 }
3104             } else if(pset_count(copies) == 0) {
3105                                 paths = 0;
3106                         }
3107                 }
3108         }
3109
3110     del_pset(visited_users);
3111         pset_remove_ptr(visited, irn);
3112         return paths;
3113 }
3114
3115 static void
3116 gen_copy_constraints(spill_ilp_t * si, const ir_node * a, const ir_node * b, ilp_var_t any_interfere)
3117 {
3118         pset * copies = pset_new_ptr_default();
3119         pset * visited = pset_new_ptr_default();
3120
3121         find_copy_path(si, a, b, any_interfere, copies, visited);
3122
3123         del_pset(visited);
3124         del_pset(copies);
3125 }
3126
3127
3128 static void
3129 memcopyhandler(spill_ilp_t * si)
3130 {
3131         interference_t   *interference;
3132         char              buf[256];
3133         /* teste Speicherwerte auf Interferenz */
3134
3135         DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
3136         irg_block_walk_graph(si->birg->irg, luke_interferencewalker, NULL, si);
3137
3138         /* now lets emit the ILP unequations for the crap */
3139         set_foreach(si->interferences, interference) {
3140                 irnlist_t      *irnlist;
3141                 ilp_var_t      interfere, any_interfere;
3142                 ilp_cst_t      any_interfere_cst, cst;
3143                 const ir_node  *a  = interference->a;
3144                 const ir_node  *b  = interference->b;
3145
3146                 /* any_interf <= \sum interf */
3147                 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
3148                 any_interfere_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3149                 any_interfere     = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3150
3151                 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
3152
3153                 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
3154                         const ir_node  *bb = irnlist->irn;
3155                         spill_bb_t     *spill_bb = get_irn_link(bb);
3156                         spill_t        *spilla,
3157                                                    *spillb;
3158                         char           buf[256];
3159
3160                         spilla = set_find_spill(spill_bb->ilp, a);
3161                         assert(spilla);
3162
3163                         spillb = set_find_spill(spill_bb->ilp, b);
3164                         assert(spillb);
3165
3166                         /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
3167                         /* 1:   mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
3168                         /* 2: - mem_in_a - spill_a + interfere <= 0 */
3169                         /* 3: - mem_in_b - spill_b + interfere <= 0 */
3170                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
3171                         interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3172
3173                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
3174                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
3175
3176                         lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
3177                         if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
3178                         lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
3179                         if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
3180                         lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
3181
3182                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
3183                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3184
3185                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3186                         if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
3187                         lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
3188
3189                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
3190                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3191
3192                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3193                         if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
3194                         lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
3195
3196
3197                         lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
3198
3199                         /* any_interfere >= interf */
3200                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
3201                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3202
3203                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3204                         lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3205                 }
3206
3207                 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
3208                 gen_copy_constraints(si,a,b,any_interfere);
3209         }
3210 }
3211
3212
3213 static INLINE int
3214 is_zero(double x)
3215 {
3216         return fabs(x) < 0.00001;
3217 }
3218
3219 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
3220 {
3221         spill_ilp_t *si = get_irg_link(current_ir_graph);
3222
3223         if(pset_find_ptr(si->all_possible_remats, n)) {
3224                 op_t   *op = (op_t*)get_irn_link(n);
3225                 assert(op && op->is_remat);
3226
3227                 if(!op->attr.remat.remat->inverse) {
3228                         if(op->attr.remat.pre) {
3229                                 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
3230                         } else {
3231                                 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
3232                         }
3233
3234                         return 1;
3235                 } else {
3236                         op_t   *op = (op_t*)get_irn_link(n);
3237                         assert(op && op->is_remat);
3238
3239                         if(op->attr.remat.pre) {
3240                                 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
3241                         } else {
3242                                 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
3243                         }
3244
3245                         return 1;
3246                 }
3247         }
3248
3249         return 0;
3250 }
3251
3252 static void
3253 dump_graph_with_remats(ir_graph * irg, const char * suffix)
3254 {
3255         set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
3256         be_dump(irg, suffix, dump_ir_block_graph_sched);
3257         set_dump_node_vcgattr_hook(NULL);
3258 }
3259
3260 /**
3261  * Edge hook to dump the schedule edges with annotated register pressure.
3262  */
3263 static int
3264 sched_pressure_edge_hook(FILE *F, ir_node *irn)
3265 {
3266         if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
3267                 ir_node *prev = sched_prev(irn);
3268                 fprintf(F, "edge:{sourcename:\"");
3269                 PRINT_NODEID(irn);
3270                 fprintf(F, "\" targetname:\"");
3271                 PRINT_NODEID(prev);
3272                 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
3273                 fprintf(F, "\" color:magenta}\n");
3274         }
3275         return 1;
3276 }
3277
3278 static void
3279 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
3280 {
3281         DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
3282
3283         dump_consts_local(0);
3284         set_dump_node_edge_hook(sched_pressure_edge_hook);
3285         dump_ir_block_graph(irg, suffix);
3286         set_dump_node_edge_hook(old_edge_hook);
3287 }
3288
3289 static void
3290 walker_pressure_annotator(ir_node * bb, void * data)
3291 {
3292         spill_ilp_t  *si = data;
3293         ir_node      *irn;
3294         int           n, i;
3295         pset         *live = pset_new_ptr_default();
3296         int           projs = 0;
3297
3298         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
3299                 irn = be_lv_get_irn(si->lv, bb, i);
3300
3301                 if (has_reg_class(si, irn)) {
3302                         pset_insert_ptr(live, irn);
3303                 }
3304         }
3305
3306         set_irn_link(bb, INT_TO_PTR(pset_count(live)));
3307
3308         sched_foreach_reverse(bb, irn) {
3309                 if(is_Phi(irn)) {
3310                         set_irn_link(irn, INT_TO_PTR(pset_count(live)));
3311                         continue;
3312                 }
3313
3314                 if(has_reg_class(si, irn)) {
3315                         pset_remove_ptr(live, irn);
3316                         if(is_Proj(irn)) ++projs;
3317                 }
3318
3319                 if(!is_Proj(irn)) projs = 0;
3320
3321                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3322                         ir_node    *arg = get_irn_n(irn, n);
3323
3324                         if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
3325                 }
3326                 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
3327         }
3328
3329         del_pset(live);
3330 }
3331
3332 static void
3333 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
3334 {
3335         be_dump(si->birg->irg, suffix, dump_ir_block_graph_sched_pressure);
3336 }
3337
3338 static void
3339 connect_all_remats_with_keep(spill_ilp_t * si)
3340 {
3341         ir_node   *irn;
3342         ir_node  **ins,
3343                          **pos;
3344         int        n_remats;
3345
3346
3347         n_remats = pset_count(si->all_possible_remats);
3348         if(n_remats) {
3349                 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
3350
3351                 pos = ins;
3352                 pset_foreach(si->all_possible_remats, irn) {
3353                         *pos = irn;
3354                         ++pos;
3355                 }
3356
3357                 si->keep = be_new_Keep(si->cls, si->birg->irg, get_irg_end_block(si->birg->irg), n_remats, ins);
3358
3359                 obstack_free(si->obst, ins);
3360         }
3361 }
3362
3363 static void
3364 connect_all_spills_with_keep(spill_ilp_t * si)
3365 {
3366         ir_node   *irn;
3367         ir_node  **ins,
3368                          **pos;
3369         int        n_spills;
3370         ir_node   *keep;
3371
3372
3373         n_spills = pset_count(si->spills);
3374         if(n_spills) {
3375                 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
3376
3377                 pos = ins;
3378                 pset_foreach(si->spills, irn) {
3379                         *pos = irn;
3380                         ++pos;
3381                 }
3382
3383                 keep = be_new_Keep(si->cls, si->birg->irg, get_irg_end_block(si->birg->irg), n_spills, ins);
3384
3385                 obstack_free(si->obst, ins);
3386         }
3387 }
3388
3389 /** insert a spill at an arbitrary position */
3390 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert)
3391 {
3392         ir_node  *bl    = is_Block(insert) ? insert : get_nodes_block(insert);
3393         ir_graph *irg   = get_irn_irg(bl);
3394         ir_node  *frame = get_irg_frame(irg);
3395         ir_node  *spill;
3396         ir_node  *next;
3397         const arch_register_class_t *cls       = arch_get_irn_reg_class(arch_env, irn, -1);
3398         const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
3399
3400         spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn);
3401
3402         /*
3403          * search the right insertion point. a spill of a phi cannot be put
3404          * directly after the phi, if there are some phis behind the one which
3405          * is spilled. Also, a spill of a Proj must be after all Projs of the
3406          * same tuple node.
3407          *
3408          * Here's one special case:
3409          * If the spill is in the start block, the spill must be after the frame
3410          * pointer is set up. This is done by setting insert to the end of the block
3411          * which is its default initialization (see above).
3412          */
3413
3414         if (bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
3415                 insert = frame;
3416
3417         for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
3418                 insert = next;
3419
3420         sched_add_after(insert, spill);
3421         return spill;
3422 }
3423
3424 static void
3425 delete_remat(spill_ilp_t * si, ir_node * remat) {
3426         int       n;
3427         ir_node  *bad = get_irg_bad(si->birg->irg);
3428
3429         sched_remove(remat);
3430
3431         /* kill links to operands */
3432         for (n=get_irn_arity(remat)-1; n>=-1; --n) {
3433                 set_irn_n(remat, n, bad);
3434         }
3435 }
3436
3437 static void
3438 clean_remat_info(spill_ilp_t * si)
3439 {
3440         int            n;
3441         remat_t       *remat;
3442         remat_info_t  *remat_info;
3443         ir_node       *bad = get_irg_bad(si->birg->irg);
3444
3445         set_foreach(si->remat_info, remat_info) {
3446                 if(!remat_info->remats) continue;
3447
3448                 pset_foreach(remat_info->remats, remat)
3449                 {
3450                         if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
3451                                 if(sched_is_scheduled(remat->proj)) {
3452                                         sched_remove((ir_node*)remat->proj);
3453                                 }
3454                                 set_irn_n((ir_node*)remat->proj, -1, bad);
3455                                 set_irn_n((ir_node*)remat->proj, 0, bad);
3456                         }
3457
3458                         if(get_irn_n_edges(remat->op) == 0) {
3459                                 if(sched_is_scheduled(remat->op)) {
3460                                         sched_remove((ir_node*)remat->op);
3461                                 }
3462                                 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
3463                                         set_irn_n((ir_node*)remat->op, n, bad);
3464                                 }
3465                         }
3466                 }
3467
3468                 if(remat_info->remats) del_pset(remat_info->remats);
3469                 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
3470         }
3471 }
3472
3473 static void
3474 delete_unnecessary_remats(spill_ilp_t * si)
3475 {
3476         if(opt_keep_alive & KEEPALIVE_REMATS) {
3477                 int       n;
3478                 ir_node  *bad = get_irg_bad(si->birg->irg);
3479
3480                 if(si->keep) {
3481                         for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
3482                                 ir_node        *keep_arg = get_irn_n(si->keep, n);
3483                                 op_t           *arg_op = get_irn_link(keep_arg);
3484                                 lpp_name_t     *name;
3485
3486                                 assert(arg_op->is_remat);
3487
3488                                 name = si->lpp->vars[arg_op->attr.remat.ilp];
3489
3490                                 if(is_zero(name->value)) {
3491                                         DBG((si->dbg, LEVEL_3, "\t  deleting remat %+F\n", keep_arg));
3492                                         /* TODO check whether reload is preferred over remat (could be bug) */
3493                                         delete_remat(si, keep_arg);
3494                                 } else {
3495                                         if(!arg_op->attr.remat.remat->inverse) {
3496                                                 if(arg_op->attr.remat.pre) {
3497                                                         DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
3498                                                 } else {
3499                                                         DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
3500                                                 }
3501                                         } else {
3502                                                 if(arg_op->attr.remat.pre) {
3503                                                         DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
3504                                                 } else {
3505                                                         DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
3506                                                 }
3507                                         }
3508                                 }
3509
3510                                 set_irn_n(si->keep, n, bad);
3511                         }
3512                 } else {
3513                         DBG((si->dbg, LEVEL_2, "\t  no remats to delete (none have been inserted)\n"));
3514                 }
3515         } else {
3516                 ir_node  *remat;
3517
3518                 pset_foreach(si->all_possible_remats, remat) {
3519                         op_t           *remat_op = get_irn_link(remat);
3520                         lpp_name_t     *name = si->lpp->vars[remat_op->attr.remat.ilp];
3521
3522                         if(is_zero(name->value)) {
3523                                 DBG((si->dbg, LEVEL_3, "\t  deleting remat %+F\n", remat));
3524                                 /* TODO check whether reload is preferred over remat (could be bug) */
3525                                 delete_remat(si, remat);
3526                         } else {
3527                                 if(!remat_op->attr.remat.remat->inverse) {
3528                                         if(remat_op->attr.remat.pre) {
3529                                                 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3530                                         } else {
3531                                                 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3532                                         }
3533                                 } else {
3534                                         if(remat_op->attr.remat.pre) {
3535                                                 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3536                                         } else {
3537                                                 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3538                                         }
3539                                 }
3540                         }
3541                 }
3542         }
3543 }
3544
3545 static pset *
3546 get_spills_for_value(spill_ilp_t * si, const ir_node * value)
3547 {
3548         pset     *spills = pset_new_ptr_default();
3549
3550         const ir_node  *next;
3551         defs_t         *defs;
3552
3553         defs = set_find_def(si->values, value);
3554
3555         if(defs && defs->spills) {
3556                 for(next = defs->spills; next; next = get_irn_link(next)) {
3557                         pset_insert_ptr(spills, next);
3558                 }
3559         }
3560
3561         return spills;
3562 }
3563
3564 static ir_node *
3565 new_r_PhiM_nokeep(ir_graph * irg, ir_node *block, int arity, ir_node **in)
3566 {
3567         ir_node  *res;
3568
3569         assert( get_irn_arity(block) == arity );
3570
3571         res = new_ir_node(NULL, irg, block, op_Phi, mode_M, arity, in);
3572         res->attr.phi_backedge = new_backedge_arr(irg->obst, arity);
3573
3574         return res;
3575 }
3576
3577 /**
3578  * @param before   The node after which the spill will be placed in the schedule
3579  */
3580 static ir_node *
3581 insert_spill(spill_ilp_t * si, ir_node * irn, const ir_node * value, ir_node * before)
3582 {
3583         defs_t   *defs;
3584         ir_node  *spill;
3585         const arch_env_t *arch_env = si->birg->main_env->arch_env;
3586
3587         DBG((si->dbg, LEVEL_3, "\t  inserting spill for value %+F after %+F\n", irn, before));
3588
3589         spill = be_spill2(arch_env, irn, before);
3590
3591         defs = set_insert_def(si->values, value);
3592         assert(defs);
3593
3594         /* enter into the linked list */
3595         set_irn_link(spill, defs->spills);
3596         defs->spills = spill;
3597
3598         if(opt_keep_alive & KEEPALIVE_SPILLS)
3599                 pset_insert_ptr(si->spills, spill);
3600
3601         return spill;
3602 }
3603
3604 /**
3605  * @param before   The Phi node which has to be spilled
3606  */
3607 static ir_node *
3608 insert_mem_phi(spill_ilp_t * si, ir_node * phi)
3609 {
3610         ir_node   *mem_phi;
3611         ir_node  **ins;
3612         defs_t    *defs;
3613         int        n;
3614
3615         NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3616
3617         for(n=get_irn_arity(phi)-1; n>=0; --n) {
3618                 ins[n] = si->m_unknown;
3619         }
3620
3621         mem_phi =  new_r_PhiM_nokeep(si->birg->irg, get_nodes_block(phi), get_irn_arity(phi), ins);
3622
3623         defs = set_insert_def(si->values, phi);
3624         assert(defs);
3625
3626         /* enter into the linked list */
3627         set_irn_link(mem_phi, defs->spills);
3628         defs->spills = mem_phi;
3629
3630 #ifdef SCHEDULE_PHIM
3631         sched_add_after(phi, mem_phi);
3632 #else
3633         pset_insert_ptr(si->phims, mem_phi);
3634 #endif
3635
3636         if(opt_keep_alive & KEEPALIVE_SPILLS)
3637                 pset_insert_ptr(si->spills, mem_phi);
3638
3639
3640         return mem_phi;
3641 }
3642
3643 /**
3644  * Add remat to list of defs, destroys link field!
3645  */
3646 static void
3647 insert_remat(spill_ilp_t * si, ir_node * remat)
3648 {
3649         defs_t   *defs;
3650         op_t     *remat_op = get_irn_link(remat);
3651
3652         assert(remat_op->is_remat);
3653
3654         defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3655         assert(defs);
3656
3657         /* enter into the linked list */
3658         set_irn_link(remat, defs->remats);
3659         defs->remats = remat;
3660 }
3661
3662
3663 /**
3664  * Add reload before operation and add to list of defs
3665  */
3666 static ir_node *
3667 insert_reload(spill_ilp_t * si, const ir_node * value, ir_node * after)
3668 {
3669         defs_t   *defs;
3670         ir_node  *reload,
3671                          *spill;
3672         const arch_env_t *arch_env = si->birg->main_env->arch_env;
3673
3674         DBG((si->dbg, LEVEL_3, "\t  inserting reload for value %+F before %+F\n", value, after));
3675
3676         defs = set_find_def(si->values, value);
3677
3678         spill = defs->spills;
3679         assert(spill && "no spill placed before reload");
3680
3681         reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3682
3683         /* enter into the linked list */
3684         set_irn_link(reload, defs->remats);
3685         defs->remats = reload;
3686
3687         return reload;
3688 }
3689
3690 void perform_memory_operand(spill_ilp_t * si, memoperand_t * memoperand)
3691 {
3692         defs_t           *defs;
3693         ir_node          *value = get_irn_n(memoperand->irn, memoperand->pos);
3694         ir_node          *spill;
3695         const arch_env_t *arch_env = si->birg->main_env->arch_env;
3696
3697         DBG((si->dbg, LEVEL_2, "\t  inserting memory operand for value %+F at %+F\n", value, memoperand->irn));
3698
3699         defs = set_find_def(si->values, value);
3700
3701         spill = defs->spills;
3702         assert(spill && "no spill placed before reload");
3703
3704         arch_perform_memory_operand(arch_env, memoperand->irn, spill, memoperand->pos);
3705 }
3706
3707 void insert_memoperands(spill_ilp_t * si)
3708 {
3709         memoperand_t   *memoperand;
3710         lpp_name_t     *name;
3711
3712         set_foreach(si->memoperands, memoperand) {
3713                 name = si->lpp->vars[memoperand->ilp];
3714                 if(!is_zero(name->value)) {
3715                         perform_memory_operand(si, memoperand);
3716                 }
3717         }
3718 }
3719
3720 static void
3721 walker_spill_placer(ir_node * bb, void * data) {
3722         spill_ilp_t   *si = (spill_ilp_t*)data;
3723         ir_node       *irn;
3724         spill_bb_t    *spill_bb = get_irn_link(bb);
3725         pset          *spills_to_do = pset_new_ptr_default();
3726         spill_t       *spill;
3727
3728         set_foreach(spill_bb->ilp, spill) {
3729                 lpp_name_t    *name;
3730
3731                 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3732                         name = si->lpp->vars[spill->mem_in];
3733                         if(!is_zero(name->value)) {
3734                                 ir_node   *mem_phi;
3735
3736                                 mem_phi = insert_mem_phi(si, spill->irn);
3737
3738                                 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3739                         }
3740                 }
3741
3742                 name = si->lpp->vars[spill->spill];
3743                 if(!is_zero(name->value)) {
3744                         /* place spill directly after definition */
3745                         if(get_nodes_block(spill->irn) == bb) {
3746                                 insert_spill(si, spill->irn, spill->irn, spill->irn);
3747                                 continue;
3748                         }
3749
3750                         /* place spill at bb start */
3751                         if(spill->reg_in > 0) {
3752                                 name = si->lpp->vars[spill->reg_in];
3753                                 if(!is_zero(name->value)) {
3754                                         insert_spill(si, spill->irn, spill->irn, bb);
3755                                         continue;
3756                                 }
3757                         }
3758                         /* place spill after a remat */
3759                         pset_insert_ptr(spills_to_do, spill->irn);
3760                 }
3761         }
3762         DBG((si->dbg, LEVEL_3, "\t  %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3763
3764
3765         for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3766                 op_t     *op = get_irn_link(irn);
3767
3768                 if(be_is_Spill(irn)) continue;
3769
3770                 if(op->is_remat) {
3771                         /* TODO fix this if we want to support remats with more than two nodes */
3772                         if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3773                                 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3774
3775                                 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3776                         }
3777                 } else {
3778                         if(pset_find_ptr(spills_to_do, irn)) {
3779                                 pset_remove_ptr(spills_to_do, irn);
3780
3781                                 insert_spill(si, irn, irn, irn);
3782                         }
3783                 }
3784
3785         }
3786
3787         assert(pset_count(spills_to_do) == 0);
3788
3789         /* afterwards free data in block */
3790         del_pset(spills_to_do);
3791 }
3792
3793 static ir_node *
3794 insert_mem_copy(spill_ilp_t * si, ir_node * bb, ir_node * value)
3795 {
3796         ir_node          *insert_pos = bb;
3797         ir_node          *spill;
3798         const arch_env_t *arch_env = si->birg->main_env->arch_env;
3799
3800         /* find last definition of arg value in block */
3801         ir_node  *next;
3802         defs_t   *defs;
3803         int       last = 0;
3804
3805         defs = set_find_def(si->values, value);
3806
3807         if(defs && defs->remats) {
3808                 for(next = defs->remats; next; next = get_irn_link(next)) {
3809                         if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3810                                 last = sched_get_time_step(next);
3811                                 insert_pos = next;
3812                         }
3813                 }
3814         }
3815
3816         if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3817                 last = sched_get_time_step(value);
3818                 insert_pos = value;
3819         }
3820
3821         DBG((si->dbg, LEVEL_2, "\t  inserting mem copy for value %+F after %+F\n", value, insert_pos));
3822
3823         spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos);
3824
3825         return spill;
3826 }
3827
3828 static void
3829 phim_fixer(spill_ilp_t *si) {
3830         defs_t  *defs;
3831
3832         set_foreach(si->values, defs) {
3833                 const ir_node  *phi = defs->value;
3834                 op_t           *op = get_irn_link(phi);
3835                 ir_node        *phi_m = NULL;
3836                 ir_node        *next = defs->spills;
3837                 int             n;
3838
3839                 if(!is_Phi(phi)) continue;
3840
3841                 while(next) {
3842                         if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3843                                 phi_m = next;
3844                                 break;
3845                         } else {
3846                                 next = get_irn_link(next);
3847                         }
3848                 }
3849                 if(!phi_m) continue;
3850
3851                 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3852                         ir_node        *value = get_irn_n(phi, n);
3853                         defs_t         *val_defs = set_find_def(si->values, value);
3854
3855                         /* a spill of this value */
3856                         ir_node      *spill;
3857
3858
3859                         if(opt_memcopies) {
3860                                 ir_node    *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3861                                 lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3862
3863                                 if(!is_zero(name->value)) {
3864                                         spill = insert_mem_copy(si, pred, value);
3865                                 } else {
3866                                         spill = val_defs->spills;
3867                                 }
3868                         } else {
3869                                 spill = val_defs->spills;
3870                         }
3871
3872                         assert(spill && "no spill placed before PhiM");
3873                         set_irn_n(phi_m, n, spill);
3874                 }
3875         }
3876 }
3877
3878 static void
3879 walker_reload_placer(ir_node * bb, void * data) {
3880         spill_ilp_t   *si = (spill_ilp_t*)data;
3881         ir_node       *irn;
3882         spill_bb_t    *spill_bb = get_irn_link(bb);
3883
3884         /* reloads at end of block */
3885         if(spill_bb->reloads) {
3886                 keyval_t    *keyval;
3887
3888                 set_foreach(spill_bb->reloads, keyval) {
3889                         ir_node        *irn = (ir_node*)keyval->key;
3890                         ilp_var_t       reload = PTR_TO_INT(keyval->val);
3891                         lpp_name_t     *name;
3892
3893                         name = si->lpp->vars[reload];
3894                         if(!is_zero(name->value)) {
3895                                 ir_node    *reload;
3896                                 ir_node    *insert_pos = bb;
3897                                 ir_node    *prev = sched_block_last_noncf(si, bb);
3898                                 op_t       *prev_op = get_irn_link(prev);
3899
3900                                 while(be_is_Spill(prev)) {
3901                                         prev = sched_prev(prev);
3902                                 }
3903
3904                                 prev_op = get_irn_link(prev);
3905
3906                                 /* insert reload before pre-remats */
3907                                 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3908                                                 && prev_op->is_remat && prev_op->attr.remat.pre) {
3909                                         insert_pos = prev;
3910
3911                                         do {
3912                                                 prev = sched_prev(prev);
3913                                         } while(be_is_Spill(prev));
3914
3915                                         prev_op = get_irn_link(prev);
3916
3917                                 }
3918
3919                                 reload = insert_reload(si, irn, insert_pos);
3920
3921                                 if(opt_keep_alive & KEEPALIVE_RELOADS)
3922                                         pset_insert_ptr(si->spills, reload);
3923                         }
3924                 }
3925         }
3926
3927         /* walk and insert more reloads and collect remats */
3928         sched_foreach_reverse(bb, irn) {
3929                 op_t     *op = get_irn_link(irn);
3930
3931                 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3932                 if(is_Phi(irn)) break;
3933
3934                 if(op->is_remat) {
3935                         if(get_irn_mode(irn) != mode_T) {
3936                                 insert_remat(si, irn);
3937                         }
3938                 } else {
3939                         int    n;
3940
3941                         for (n=get_irn_arity(irn)-1; n>=0; --n) {
3942                                 ir_node    *arg = get_irn_n(irn, n);
3943
3944                                 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3945                                         lpp_name_t    *name;
3946
3947                                         name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3948                                         if(!is_zero(name->value)) {
3949                                                 ir_node    *reload;
3950                                                 ir_node    *insert_pos = irn;
3951                                                 ir_node    *prev = sched_prev(insert_pos);
3952                                                 op_t       *prev_op;
3953
3954                                                 while(be_is_Spill(prev)) {
3955                                                         prev = sched_prev(prev);
3956                                                 }
3957
3958                                                 prev_op = get_irn_link(prev);
3959
3960                                                 /* insert reload before pre-remats */
3961                                                 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3962                                                                 && prev_op->is_remat && prev_op->attr.remat.pre) {
3963                                                         insert_pos = prev;
3964
3965                                                         do {
3966                                                                 prev = sched_prev(prev);
3967                                                         } while(be_is_Spill(prev));
3968
3969                                                         prev_op = get_irn_link(prev);
3970
3971                                                 }
3972
3973                                                 reload = insert_reload(si, arg, insert_pos);
3974
3975                                                 assert(reload && "no reload returned");
3976                                                 set_irn_n(irn, n, reload);
3977
3978                                                 if(opt_keep_alive & KEEPALIVE_RELOADS)
3979                                                         pset_insert_ptr(si->spills, reload);
3980                                         }
3981                                 }
3982                         }
3983                 }
3984         }
3985
3986         del_set(spill_bb->ilp);
3987         if(spill_bb->reloads) del_set(spill_bb->reloads);
3988 }
3989
3990 static void
3991 walker_collect_used(ir_node * irn, void * data)
3992 {
3993         bitset_t   *used = data;
3994
3995         bitset_set(used, get_irn_idx(irn));
3996 }
3997
3998 struct kill_helper {
3999         bitset_t  *used;
4000         spill_ilp_t  *si;
4001 };
4002
4003 static void
4004 walker_kill_unused(ir_node * bb, void * data)
4005 {
4006         struct kill_helper *kh = data;
4007         ir_node            *bad = get_irg_bad(get_irn_irg(bb));
4008         ir_node            *irn;
4009
4010
4011         for(irn=sched_first(bb); !sched_is_end(irn);) {
4012                 ir_node     *next = sched_next(irn);
4013                 int          n;
4014
4015                 if(!bitset_is_set(kh->used, get_irn_idx(irn))) {
4016                         if(be_is_Spill(irn) || be_is_Reload(irn)) {
4017                                 DBG((kh->si->dbg, LEVEL_1, "\t SUBOPTIMAL! %+F IS UNUSED (cost: %g)\n", irn, get_cost(kh->si, irn)*execution_frequency(kh->si, bb)));
4018 #if 0
4019                                 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
4020 #endif
4021                         }
4022
4023                         sched_remove(irn);
4024
4025                         set_nodes_block(irn, bad);
4026                         for (n=get_irn_arity(irn)-1; n>=0; --n) {
4027                                 set_irn_n(irn, n, bad);
4028                         }
4029                 }
4030                 irn = next;
4031         }
4032 }
4033
4034 #ifndef SCHEDULE_PHIM
4035 static void
4036 kill_unused_phims(spill_ilp_t * si, struct kill_helper * kh)
4037 {
4038         ir_node  *phi;
4039         ir_node  *bad = get_irg_bad(si->birg->irg);
4040         int       n;
4041
4042         pset_foreach(si->phims, phi) {
4043                 if(!bitset_is_set(kh->used, get_irn_idx(phi))) {
4044
4045                         set_nodes_block(phi, bad);
4046                         for (n=get_irn_arity(phi)-1; n>=0; --n) {
4047                                 set_irn_n(phi, n, bad);
4048                         }
4049                 }
4050         }
4051 }
4052 #endif
4053
4054 static void
4055 kill_all_unused_values_in_schedule(spill_ilp_t * si)
4056 {
4057         struct kill_helper  kh;
4058
4059         kh.used = bitset_malloc(get_irg_last_idx(si->birg->irg));
4060         kh.si = si;
4061
4062         irg_walk_graph(si->birg->irg, walker_collect_used, NULL, kh.used);
4063 #ifndef SCHEDULE_PHIM
4064         kill_unused_phims(si, &kh);
4065 #endif
4066         irg_block_walk_graph(si->birg->irg, walker_kill_unused, NULL, &kh);
4067
4068         bitset_free(kh.used);
4069 }
4070
4071 void
4072 print_irn_pset(pset * p)
4073 {
4074         ir_node   *irn;
4075
4076         pset_foreach(p, irn) {
4077                 ir_printf("%+F\n", irn);
4078         }
4079 }
4080
4081 void
4082 dump_phi_class(spill_ilp_t *si, ir_node **phiclass, const char * file)
4083 {
4084     FILE           *f = fopen(file, "w");
4085     ir_node        *irn;
4086     interference_t *interference;
4087     int            i;
4088
4089     set_break(si->interferences);
4090
4091     ir_fprintf(f, "digraph phiclass {\n");
4092
4093     for (i = ARR_LEN(phiclass) - 1; i >= 0; --i) {
4094         irn = phiclass[i];
4095         if (is_Phi(irn))
4096             ir_fprintf(f, "  %F%N [shape=box]\n", irn, irn);
4097     }
4098
4099     for (i = ARR_LEN(phiclass) - 1; i >= 0; --i) {
4100         int n;
4101
4102         irn = phiclass[i];
4103         if (! is_Phi(irn))
4104             continue;
4105
4106         for (n = get_irn_arity(irn) - 1; n >= 0; --n) {
4107             ir_node  *arg = get_irn_n(irn, n);
4108
4109             ir_fprintf(f, "  %F%N -> %F%N\n", irn, irn, arg, arg);
4110         }
4111     }
4112
4113     set_foreach(si->interferences, interference) {
4114         const ir_node *a = interference->a;
4115         const ir_node *b = interference->b;
4116         if (get_phi_class(si->pc, (ir_node *)a) == phiclass) {
4117             ir_fprintf(f, "  %F%N -> %F%N [color=red,dir=none,style=bold]\n", a, a, b, b);
4118         }
4119     }
4120
4121     ir_fprintf(f, "}");
4122     fclose(f);
4123 }
4124
4125 static void
4126 rewire_uses(spill_ilp_t * si)
4127 {
4128         defs_t               *defs;
4129         pset                 *ignore = pset_new_ptr(1);
4130         be_dom_front_info_t *dom_front = si->birg->dom_front;
4131
4132         pset_insert_ptr(ignore, get_irg_end(si->birg->irg));
4133
4134         /* then fix uses of spills */
4135         set_foreach(si->values, defs) {
4136                 pset           *reloads;
4137                 pset           *spills;
4138                 const ir_node  *next = defs->remats;
4139                 int remats = 0;
4140
4141                 reloads = pset_new_ptr_default();
4142
4143                 while(next) {
4144                         if(be_is_Reload(next)) {
4145                                 pset_insert_ptr(reloads, next);
4146                         } else {
4147                                 ++remats;
4148                         }
4149                         next = get_irn_link(next);
4150                 }
4151
4152                 spills = get_spills_for_value(si, defs->value);
4153                 DBG((si->dbg, LEVEL_2, "\t  %d remats, %d reloads, and %d spills for value %+F\n", remats, pset_count(reloads), pset_count(spills), defs->value));
4154                 if(pset_count(spills) > 1) {
4155                         //assert(pset_count(reloads) > 0);
4156                         //                              print_irn_pset(spills);
4157                         //                              print_irn_pset(reloads);
4158
4159                         be_ssa_constr_set_ignore(dom_front, si->lv, spills, ignore);
4160                 }
4161
4162                 del_pset(reloads);
4163                 del_pset(spills);
4164         }
4165
4166         /* first fix uses of remats and reloads */
4167         set_foreach(si->values, defs) {
4168                 pset           *nodes;
4169                 const ir_node  *next = defs->remats;
4170                 int             orig_kept = 0;
4171
4172                 if(next) {
4173                         nodes = pset_new_ptr_default();
4174                         if(sched_is_scheduled(defs->value)) {
4175                                 pset_insert_ptr(nodes, defs->value);
4176                                 orig_kept = 1;
4177                         }
4178
4179                         while(next) {
4180                                 pset_insert_ptr(nodes, next);
4181                                 next = get_irn_link(next);
4182                         }
4183
4184                         DBG((si->dbg, LEVEL_4, "\t    %d new definitions for value %+F\n", pset_count(nodes)-orig_kept, defs->value));
4185                         be_ssa_constr_set(dom_front, si->lv, nodes);
4186
4187                         del_pset(nodes);
4188                 }
4189         }
4190
4191 //      remove_unused_defs(si);
4192 }
4193
4194
4195 static void
4196 writeback_results(spill_ilp_t * si)
4197 {
4198         /* walk through the graph and collect all spills, reloads and remats for a value */
4199
4200         si->values = new_set(cmp_defs, 4096);
4201
4202         DBG((si->dbg, LEVEL_1, "Applying results\n"));
4203         delete_unnecessary_remats(si);
4204         si->m_unknown = new_r_Unknown(si->birg->irg, mode_M);
4205         irg_block_walk_graph(si->birg->irg, walker_spill_placer, NULL, si);
4206         irg_block_walk_graph(si->birg->irg, walker_reload_placer, NULL, si);
4207         if(opt_memoperands)
4208                 insert_memoperands(si);
4209         phim_fixer(si);
4210
4211         /* clean the remat info! there are still back-edges leading there! */
4212         clean_remat_info(si);
4213
4214         rewire_uses(si);
4215
4216         connect_all_spills_with_keep(si);
4217
4218         del_set(si->values);
4219 }
4220
4221 static int
4222 get_n_regs(spill_ilp_t * si)
4223 {
4224         int       arch_n_regs = arch_register_class_n_regs(si->cls);
4225
4226         bitset_t *arch_regs = bitset_malloc(arch_n_regs);
4227         bitset_t *abi_regs = bitset_malloc(arch_n_regs);
4228
4229         arch_put_non_ignore_regs(si->birg->main_env->arch_env, si->cls, arch_regs);
4230     be_abi_put_ignore_regs(si->birg->abi, si->cls, abi_regs);
4231
4232         bitset_andnot(arch_regs, abi_regs);
4233         arch_n_regs = bitset_popcnt(arch_regs);
4234
4235         bitset_free(arch_regs);
4236         bitset_free(abi_regs);
4237
4238         DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", arch_n_regs, si->cls->name));
4239         return arch_n_regs;
4240 }
4241
4242 static void
4243 walker_reload_mover(ir_node * bb, void * data)
4244 {
4245         spill_ilp_t   *si = data;
4246         ir_node           *tmp;
4247
4248         sched_foreach(bb, tmp) {
4249                 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
4250                         ir_node       *reload = tmp;
4251                         ir_node       *irn = tmp;
4252
4253                         /* move reload upwards */
4254
4255                         int pressure = (int)get_irn_link(reload);
4256                         if(pressure < si->n_regs) {
4257                                 irn = sched_prev(reload);
4258                                 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
4259                                 sched_remove(reload);
4260                                 pressure = (int)get_irn_link(irn);
4261
4262                                 while(pressure < si->n_regs) {
4263                                         if( sched_is_end(irn) ||
4264                                            (be_is_Reload(irn) && has_reg_class(si, irn)) ||
4265                                            /* do not move reload before its spill */
4266                                            (irn == be_get_Reload_mem(reload)) ||
4267                                            /* do not move before phi */
4268                                            is_Phi(irn)) break;
4269
4270                                         set_irn_link(irn, INT_TO_PTR(pressure+1));
4271                                         DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
4272                                         irn = sched_prev(irn);
4273
4274                                         pressure = (int)get_irn_link(irn);
4275                                 }
4276
4277                                 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
4278                                 sched_put_after(irn, reload);
4279                         }
4280                 }
4281         }
4282 }
4283
4284 static void
4285 move_reloads_upward(spill_ilp_t * si)
4286 {
4287         irg_block_walk_graph(si->birg->irg, walker_reload_mover, NULL, si);
4288 }
4289
4290
4291 /**
4292  * Walk all irg blocks and check for interfering spills inside of phi classes
4293  */
4294 static void
4295 luke_meminterferencechecker(ir_node * bb, void * data)
4296 {
4297         spill_ilp_t *si = (spill_ilp_t*)data;
4298         int         l1, l2;
4299
4300         be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
4301                 ir_node        *a = be_lv_get_irn(si->lv, bb, l1);
4302
4303                 if (! be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T))
4304                         continue;
4305
4306                 /* a is only interesting if it is in my register class and if it is inside a phi class */
4307                 if (has_reg_class(si, a) && get_phi_class(si->pc, a)) {
4308                         for (l2 = _be_lv_next_irn(si->lv, bb, 0xff, l1 + 1); l2 >= 0; l2 = _be_lv_next_irn(si->lv, bb, 0xff, l2 + 1)) {
4309                                 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
4310
4311                                 if (! be_is_Spill(b) && (! is_Phi(b) || get_irn_mode(b) != mode_T))
4312                                         continue;
4313
4314                                 /* a and b are only interesting if they are in the same phi class */
4315                                 if (has_reg_class(si, b) && get_phi_class(si->pc, a) == get_phi_class(si->pc, b)) {
4316                                         if (values_interfere_in_block(si, bb, a, b)) {
4317                                                 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
4318                                         }
4319                                 }
4320                         }
4321                 }
4322         }
4323 }
4324
4325 static void
4326 verify_phiclasses(spill_ilp_t * si)
4327 {
4328         /* analyze phi classes */
4329         phi_class_free(si->pc);
4330         si->pc = phi_class_new_from_irg(si->birg->irg, 0);
4331
4332         DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
4333         irg_block_walk_graph(si->birg->irg, luke_meminterferencechecker, NULL, si);
4334 }
4335
4336 void
4337 be_spill_remat(be_irg_t *birg, const arch_register_class_t *cls)
4338 {
4339         char            buf[256];
4340         char            problem_name[256];
4341         char            dump_suffix[256];
4342         char            dump_suffix2[256];
4343         struct obstack  obst;
4344         spill_ilp_t     si;
4345         ir_graph       *irg = be_get_birg_irg(birg);
4346
4347         ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", irg, cls->name);
4348         ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", cls->name);
4349         ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", cls->name);
4350
4351         FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
4352         DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
4353
4354         if(opt_verify & VERIFY_DOMINANCE)
4355                 be_check_dominance(irg);
4356
4357         be_assure_dom_front(birg);
4358         be_assure_liveness(birg);
4359
4360         obstack_init(&obst);
4361         si.obst                = &obst;
4362         si.birg                = birg;
4363         si.cls                 = cls;
4364         si.lpp                 = new_lpp(problem_name, lpp_minimize);
4365         si.remat_info          = new_set(cmp_remat_info, 4096);
4366         si.interferences       = new_set(cmp_interference, 32);
4367         si.memoperands         = new_set(cmp_memoperands, 128);
4368         si.all_possible_remats = pset_new_ptr_default();
4369         si.spills              = pset_new_ptr_default();
4370         si.inverse_ops         = pset_new_ptr_default();
4371         si.lv                  = birg->lv;
4372         si.keep                = NULL;
4373         si.n_regs              = get_n_regs(&si);
4374
4375         set_irg_link(irg, &si);
4376         compute_doms(irg);
4377
4378         /* compute phi classes */
4379         // phi_class_compute(irg);
4380
4381         if(opt_dump_flags & DUMP_STATS)
4382                 be_analyze_regpressure(birg, cls, "-pre");
4383
4384         DBG((si.dbg, LEVEL_2, "\t initializing\n"));
4385         irg_block_walk_graph(irg, luke_initializer, NULL, &si);
4386
4387         if(opt_remats) {
4388                 /* collect remats */
4389                 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
4390                 irg_walk_graph(irg, walker_remat_collector, NULL, &si);
4391         }
4392
4393         /* insert possible remats */
4394         DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
4395         irg_block_walk_graph(irg, walker_remat_insertor, NULL, &si);
4396         DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
4397
4398         if(opt_keep_alive & KEEPALIVE_REMATS) {
4399                 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
4400                 connect_all_remats_with_keep(&si);
4401                 /* dump graph with inserted remats */
4402                 dump_graph_with_remats(irg, dump_suffix);
4403         }
4404
4405         /* insert copies for phi arguments not in my regclass */
4406         irg_walk_graph(irg, walker_regclass_copy_insertor, NULL, &si);
4407
4408         /* recompute liveness */
4409         DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
4410         be_liveness_recompute(si.lv);
4411
4412         /* build the ILP */
4413         DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
4414         DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
4415         irg_block_walk_graph(irg, luke_endwalker, NULL, &si);
4416
4417         DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
4418         irg_block_walk_graph(irg, luke_blockwalker, NULL, &si);
4419
4420         si.pc = phi_class_new_from_irg(birg->irg, 0);
4421         if (opt_memcopies) {
4422                 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
4423                 memcopyhandler(&si);
4424         }
4425
4426         if (opt_dump_flags & DUMP_PROBLEM) {
4427                 FILE           *f;
4428                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
4429                 if ((f = fopen(buf, "wt")) != NULL) {
4430                         lpp_dump_plain(si.lpp, f);
4431                         fclose(f);
4432                 }
4433         }
4434
4435         if (opt_dump_flags & DUMP_MPS) {
4436                 FILE *f;
4437
4438                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mps", problem_name);
4439                 if ((f = fopen(buf, "wt")) != NULL) {
4440                         mps_write_mps(si.lpp, s_mps_fixed, f);
4441                         fclose(f);
4442                 }
4443
4444                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mst", problem_name);
4445                 if ((f = fopen(buf, "wt")) != NULL) {
4446                         mps_write_mst(si.lpp, s_mps_fixed, f);
4447                         fclose(f);
4448                 }
4449         }
4450
4451         lpp_check_startvals(si.lpp);
4452
4453 #ifdef SOLVE
4454         DBG((si.dbg, LEVEL_1, "\tSolving %s (%d variables, %d constraints)\n", problem_name, si.lpp->var_next, si.lpp->cst_next));
4455         lpp_set_time_limit(si.lpp, opt_timeout);
4456
4457         if(opt_log)
4458                 lpp_set_log(si.lpp, stdout);
4459
4460 #ifdef SOLVE_LOCAL
4461         lpp_solve_cplex(si.lpp);
4462 #else
4463         lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
4464 #endif
4465         assert(lpp_is_sol_valid(si.lpp)
4466                && "solution of ILP must be valid");
4467
4468         DBG((si.dbg, LEVEL_1, "\t%s: iterations: %d, solution time: %g, objective function: %g, best bound: %g\n", problem_name, si.lpp->iterations, si.lpp->sol_time, is_zero(si.lpp->objval)?0.0:si.lpp->objval, is_zero(si.lpp->best_bound)?0.0:si.lpp->best_bound));
4469
4470         if(opt_dump_flags & DUMP_SOLUTION) {
4471                 FILE           *f;
4472                 char            buf[256];
4473
4474                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
4475                 if ((f = fopen(buf, "wt")) != NULL) {
4476                         int             i;
4477                         for (i = 0; i < si.lpp->var_next; ++i) {
4478                                 lpp_name_t     *name = si.lpp->vars[i];
4479                                 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
4480                         }
4481                         fclose(f);
4482                 }
4483         }
4484
4485 #ifndef SCHEDULE_PHIM
4486         si.phims = pset_new_ptr_default();
4487 #endif
4488         writeback_results(&si);
4489
4490
4491 #endif                          /* SOLVE */
4492
4493         kill_all_unused_values_in_schedule(&si);
4494
4495 #if !defined(SCHEDULE_PHIM) && defined(SOLVE)
4496         del_pset(si.phims);
4497 #endif
4498
4499         if(opt_keep_alive & (KEEPALIVE_SPILLS | KEEPALIVE_RELOADS))
4500                 be_dump(irg, "-spills-placed", dump_ir_block_graph);
4501
4502         // move reloads upwards
4503         be_liveness_recompute(si.lv);
4504         irg_block_walk_graph(irg, walker_pressure_annotator, NULL, &si);
4505         move_reloads_upward(&si);
4506
4507         if(opt_memcopies) {
4508                 verify_phiclasses(&si);
4509         }
4510
4511         irg_block_walk_graph(irg, walker_pressure_annotator, NULL, &si);
4512
4513         if(opt_dump_flags & DUMP_PRESSURE)
4514                 dump_pressure_graph(&si, dump_suffix2);
4515
4516         if(opt_dump_flags & DUMP_STATS)
4517                 be_analyze_regpressure(birg, cls, "-post");
4518
4519         if(opt_verify & VERIFY_DOMINANCE)
4520                 be_check_dominance(irg);
4521
4522         free_dom(irg);
4523         del_set(si.interferences);
4524         del_pset(si.inverse_ops);
4525         del_pset(si.all_possible_remats);
4526         del_set(si.memoperands);
4527         del_pset(si.spills);
4528         free_lpp(si.lpp);
4529         phi_class_free(si.pc);
4530         obstack_free(&obst, NULL);
4531         DBG((si.dbg, LEVEL_1, "\tdone.\n"));
4532 }
4533
4534 void be_init_spillremat(void)
4535 {
4536         static be_spiller_t remat_spiller = {
4537                 be_spill_remat
4538         };
4539         lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
4540         lc_opt_entry_t *ra_grp = lc_opt_get_grp(be_grp, "ra");
4541         lc_opt_entry_t *chordal_grp = lc_opt_get_grp(ra_grp, "chordal");
4542         lc_opt_entry_t *remat_grp = lc_opt_get_grp(chordal_grp, "remat");
4543
4544         be_register_spiller("remat", &remat_spiller);
4545         lc_opt_add_table(remat_grp, options);
4546 }
4547
4548 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillremat);
4549
4550 #else                           /* WITH_ILP */
4551
4552 static void INLINE
4553 only_that_you_can_compile_without_WITH_ILP_defined(void)
4554 {
4555 }
4556
4557 #endif                          /* WITH_ILP */