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_t.h"
  33 #include "iredges.h"
  34 #include "execfreq.h"
  35 #include "irvrfy.h"
  36
  37 #include <lpp/lpp.h>
  38 #include <lpp/lpp_net.h>
  39 #include <lpp/lpp_cplex.h>
  40 //#include <lc_pset.h>
  41 #include <libcore/lc_bitset.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 "benode_t.h"
  49 #include "beutil.h"
  50 #include "bespillremat.h"
  51 #include "bespill.h"
  52 #include "bepressurestat.h"
  53
  54 #include "bechordal_t.h"
  55
  56 #define BIGM 100000.0
  57
  58 #define DUMP_SOLUTION
  59 #define DUMP_ILP
  60 //#define KEEPALIVE /* keep alive all inserted remats and dump graph with remats */
  61 #define COLLECT_REMATS /* enable rematerialization */
  62 #define COLLECT_INVERSE_REMATS /* enable placement of inverse remats */
  63 #define REMAT_WHILE_LIVE /* only remat values that are live */
  64 //#define NO_ENLARGE_L1V3N355 /* do not remat after the death of some operand */
  65 //#define EXECFREQ_LOOPDEPH /* compute execution frequency from loop depth only */
  66 #define MAY_DIE_AT_REMAT /* allow values to die after a pre remat */
  67 #define NO_SINGLE_USE_REMATS /* do not repair schedule */
  68 //#define KEEPALIVE_SPILLS
  69 //#define KEEPALIVE_RELOADS
  70 #define GOODWIN_REDUCTION
  71 //#define NO_MEMCOPIES
  72 //#define VERIFY_DOMINANCE
  73 #define WITH_MEMOPERANDS
  74
  75 #define  SOLVE
  76 //#define  SOLVE_LOCAL
  77 #define LPP_SERVER "i44pc52"
  78 #define LPP_SOLVER "cplex"
  79
  80 #define COST_LOAD        8
  81 #define COST_MEMOPERAND  7
  82 #define COST_STORE       50
  83 #define COST_REMAT       1
  84
  85 #define ILP_TIMEOUT    300
  86 #define MAX_PATHS      16
  87 #define ILP_UNDEF               -1
  88
  89 typedef struct _spill_ilp_t {
  90         const arch_register_class_t  *cls;
  91         int                           n_regs;
  92         const be_chordal_env_t       *chordal_env;
  93         be_lv_t                      *lv;
  94         lpp_t                        *lpp;
  95         struct obstack               *obst;
  96         set                          *remat_info;
  97         pset                         *all_possible_remats;
  98         pset                         *inverse_ops;
  99 #ifdef KEEPALIVE
 100         ir_node                      *keep;
 101 #endif
 102         set                          *values; /**< for collecting all definitions of values before running ssa-construction */
 103         pset                         *spills;
 104         set                          *interferences;
 105         ir_node                      *m_unknown;
 106 #ifdef WITH_MEMOPERANDS
 107         set                          *memoperands;
 108 #endif
 109         DEBUG_ONLY(firm_dbg_module_t * dbg);
 110 } spill_ilp_t;
 111
 112 typedef int ilp_var_t;
 113 typedef int ilp_cst_t;
 114
 115 typedef struct _spill_bb_t {
 116         set      *ilp;
 117         set      *reloads;
 118 } spill_bb_t;
 119
 120 typedef struct _remat_t {
 121         const ir_node        *op;      /**< for copy_irn */
 122         const ir_node        *value;   /**< the value which is being recomputed by this remat */
 123         ir_node              *proj;    /**< not NULL if the above op produces a tuple */
 124         int                   cost;    /**< cost of this remat */
 125         int                   inverse; /**< nonzero if this is an inverse remat */
 126 } remat_t;
 127
 128 /**
 129  * Data to be attached to each IR node. For remats this contains the ilp_var
 130  * for this remat and for normal ops this contains the ilp_vars for
 131  * reloading each operand
 132  */
 133 typedef struct _op_t {
 134         int             is_remat;
 135         union {
 136                 struct {
 137                         ilp_var_t       ilp;
 138                         remat_t        *remat; /** the remat this op belongs to */
 139                         int             pre; /** 1, if this is a pressure-increasing remat */
 140                 } remat;
 141                 struct {
 142                         ilp_var_t       ilp;
 143                         ir_node        *op; /** the operation this live range belongs to */
 144                         union {
 145                                 ilp_var_t      *reloads;
 146                                 ilp_var_t      *copies;
 147                         } args;
 148                 } live_range;
 149         } attr;
 150 } op_t;
 151
 152 typedef struct _defs_t {
 153         ir_node   *value;
 154         ir_node   *spills;  /**< points to the first spill for this value (linked by link field) */
 155         ir_node   *remats;  /**< points to the first definition for this value (linked by link field) */
 156 } defs_t;
 157
 158 typedef struct _remat_info_t {
 159         const ir_node       *irn; /**< the irn to which these remats belong */
 160         pset                *remats; /**< possible remats for this value */
 161         pset                *remats_by_operand; /**< remats with this value as operand */
 162 } remat_info_t;
 163
 164 typedef struct _keyval_t {
 165         const void          *key;
 166         const void          *val;
 167 } keyval_t;
 168
 169 typedef struct _spill_t {
 170         ir_node      *irn;
 171         ilp_var_t     reg_in;
 172         ilp_var_t     mem_in;
 173         ilp_var_t     reg_out;
 174         ilp_var_t     mem_out;
 175         ilp_var_t     spill;
 176 } spill_t;
 177
 178 #ifdef WITH_MEMOPERANDS
 179 typedef struct _memoperand_t {
 180         const ir_node       *irn; /**< the irn */
 181         unsigned int         pos; /**< the position of the argument */
 182         ilp_var_t            ilp; /**< the ilp var for this memory operand */
 183 } memoperand_t;
 184 #endif
 185
 186 static INLINE int
 187 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
 188 {
 189         return chordal_has_class(si->chordal_env, irn);
 190 }
 191
 192 #if 0
 193 static int
 194 cmp_remat(const void *a, const void *b)
 195 {
 196         const keyval_t *p = a;
 197         const keyval_t *q = b;
 198         const remat_t  *r = p->val;
 199         const remat_t  *s = q->val;
 200
 201         assert(r && s);
 202
 203         return !(r == s || r->op == s->op);
 204 }
 205 #endif
 206 static int
 207 cmp_remat(const void *a, const void *b)
 208 {
 209         const remat_t  *r = a;
 210         const remat_t  *s = a;
 211
 212         return !(r == s || r->op == s->op);
 213 }
 214
 215 static int
 216 cmp_spill(const void *a, const void *b, size_t size)
 217 {
 218         const spill_t *p = a;
 219         const spill_t *q = b;
 220
 221 //      return !(p->irn == q->irn && p->bb == q->bb);
 222         return !(p->irn == q->irn);
 223 }
 224
 225 #ifdef WITH_MEMOPERANDS
 226 static int
 227 cmp_memoperands(const void *a, const void *b, size_t size)
 228 {
 229         const memoperand_t *p = a;
 230         const memoperand_t *q = b;
 231
 232         return !(p->irn == q->irn && p->pos == q->pos);
 233 }
 234 #endif
 235
 236 static keyval_t *
 237 set_find_keyval(set * set, void * key)
 238 {
 239         keyval_t     query;
 240
 241         query.key = key;
 242         return set_find(set, &query, sizeof(query), HASH_PTR(key));
 243 }
 244
 245 static keyval_t *
 246 set_insert_keyval(set * set, void * key, void * val)
 247 {
 248         keyval_t     query;
 249
 250         query.key = key;
 251         query.val = val;
 252         return set_insert(set, &query, sizeof(query), HASH_PTR(key));
 253 }
 254
 255 static defs_t *
 256 set_find_def(set * set, ir_node * value)
 257 {
 258         defs_t     query;
 259
 260         query.value = value;
 261         return set_find(set, &query, sizeof(query), HASH_PTR(value));
 262 }
 263
 264 static defs_t *
 265 set_insert_def(set * set, ir_node * value)
 266 {
 267         defs_t     query;
 268
 269         query.value = value;
 270         query.spills = NULL;
 271         query.remats = NULL;
 272         return set_insert(set, &query, sizeof(query), HASH_PTR(value));
 273 }
 274
 275 #ifdef WITH_MEMOPERANDS
 276 static memoperand_t *
 277 set_insert_memoperand(set * set, ir_node * irn, unsigned int pos, ilp_var_t ilp)
 278 {
 279         memoperand_t     query;
 280
 281         query.irn = irn;
 282         query.pos = pos;
 283         query.ilp = ilp;
 284         return set_insert(set, &query, sizeof(query), HASH_PTR(irn)+pos);
 285 }
 286
 287 static memoperand_t *
 288 set_find_memoperand(set * set, ir_node * irn, unsigned int pos)
 289 {
 290         memoperand_t     query;
 291
 292         query.irn = irn;
 293         query.pos = pos;
 294         return set_find(set, &query, sizeof(query), HASH_PTR(irn)+pos);
 295 }
 296 #endif
 297
 298
 299 static spill_t *
 300 set_find_spill(set * set, ir_node * value)
 301 {
 302         spill_t     query;
 303
 304         query.irn = value;
 305         return set_find(set, &query, sizeof(query), HASH_PTR(value));
 306 }
 307
 308 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
 309 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
 310 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
 311 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
 312 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
 313
 314 static int
 315 cmp_remat_info(const void *a, const void *b, size_t size)
 316 {
 317         const remat_info_t *p = a;
 318         const remat_info_t *q = b;
 319
 320         return !(p->irn == q->irn);
 321 }
 322
 323 static int
 324 cmp_defs(const void *a, const void *b, size_t size)
 325 {
 326         const defs_t *p = a;
 327         const defs_t *q = b;
 328
 329         return !(p->value == q->value);
 330 }
 331
 332 static int
 333 cmp_keyval(const void *a, const void *b, size_t size)
 334 {
 335         const keyval_t *p = a;
 336         const keyval_t *q = b;
 337
 338         return !(p->key == q->key);
 339 }
 340
 341 static double
 342 execution_frequency(const spill_ilp_t *si, const ir_node * irn)
 343 {
 344 #define FUDGE 0.001
 345 #ifndef EXECFREQ_LOOPDEPH
 346         return get_block_execfreq(si->chordal_env->exec_freq, get_block(irn)) + FUDGE;
 347 #else
 348         if(is_Block(irn))
 349                 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
 350         else
 351                 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
 352 #endif
 353 }
 354
 355 static double
 356 get_cost(const spill_ilp_t * si, const ir_node * irn)
 357 {
 358         if(be_is_Spill(irn)) {
 359                 return COST_STORE;
 360         } else if(be_is_Reload(irn)){
 361                 return COST_LOAD;
 362         } else {
 363                 return arch_get_op_estimated_cost(si->chordal_env->birg->main_env->arch_env, irn);
 364         }
 365 }
 366
 367 /**
 368  * Checks, whether node and its operands have suitable reg classes
 369  */
 370 static INLINE int
 371 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
 372 {
 373         int               n;
 374         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
 375         int               remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
 376
 377 #if 0
 378         if(!remat)
 379                 ir_fprintf(stderr, "  Node %+F is not rematerializable\n", irn);
 380 #endif
 381
 382         for (n = get_irn_arity(irn)-1; n>=0 && remat; --n) {
 383                 ir_node        *op = get_irn_n(irn, n);
 384                 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
 385
 386 //              if(!remat)
 387 //                      ir_fprintf(stderr, "  Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
 388         }
 389
 390         return remat;
 391 }
 392
 393 /**
 394  * Try to create a remat from @p op with destination value @p dest_value
 395  */
 396 static INLINE remat_t *
 397 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
 398 {
 399         remat_t  *remat = NULL;
 400
 401 //      if(!mode_is_datab(get_irn_mode(dest_value)))
 402 //              return NULL;
 403
 404         if(dest_value == op) {
 405                 const ir_node *proj = NULL;
 406
 407                 if(is_Proj(dest_value)) {
 408                         op = get_irn_n(op, 0);
 409                         proj = dest_value;
 410                 }
 411
 412                 if(!is_rematerializable(si, op))
 413                         return NULL;
 414
 415                 remat = obstack_alloc(si->obst, sizeof(*remat));
 416                 remat->op = op;
 417                 remat->cost = get_cost(si, op);
 418                 remat->value = dest_value;
 419                 remat->proj = proj;
 420                 remat->inverse = 0;
 421         } else {
 422                 arch_inverse_t     inverse;
 423                 int                n;
 424
 425                 /* get the index of the operand we want to retrieve by the inverse op */
 426                 for (n = get_irn_arity(op)-1; n>=0; --n) {
 427                         ir_node        *arg = get_irn_n(op, n);
 428
 429                         if(arg == dest_value) break;
 430                 }
 431                 if(n<0) return NULL;
 432
 433                 DBG((si->dbg, LEVEL_5, "\t  requesting inverse op for argument %d of op %+F\n", n, op));
 434
 435                 /* else ask the backend to give an inverse op */
 436                 if(arch_get_inverse(si->chordal_env->birg->main_env->arch_env, op, n, &inverse, si->obst)) {
 437                         int   i;
 438
 439                         DBG((si->dbg, LEVEL_4, "\t  backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
 440
 441                         assert(inverse.n > 0 && "inverse op should have at least one node");
 442
 443                         for(i=inverse.n-1; i>=0; --i) {
 444                                 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
 445                         }
 446
 447                         if(inverse.n <= 2) {
 448                                 remat = obstack_alloc(si->obst, sizeof(*remat));
 449                                 remat->op = inverse.nodes[0];
 450                                 remat->cost = inverse.costs;
 451                                 remat->value = dest_value;
 452                                 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
 453                                 remat->inverse = 1;
 454
 455                                 assert(is_Proj(remat->proj));
 456                         } else {
 457                                 assert(0 && "I can not handle remats with more than 2 nodes");
 458                         }
 459                 }
 460         }
 461
 462         if(remat) {
 463                 if(remat->proj) {
 464                         DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
 465                 } else {
 466                         DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
 467                 }
 468         }
 469         return remat;
 470 }
 471
 472
 473 static INLINE void
 474 add_remat(const spill_ilp_t * si, const remat_t * remat)
 475 {
 476         remat_info_t    *remat_info,
 477                      query;
 478         int              n;
 479
 480         assert(remat->op);
 481         assert(remat->value);
 482
 483         query.irn = remat->value;
 484         query.remats = NULL;
 485         query.remats_by_operand = NULL;
 486         remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
 487
 488         if(remat_info->remats == NULL) {
 489                 remat_info->remats = new_pset(cmp_remat, 4096);
 490         }
 491         pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
 492
 493         /* insert the remat into the remats_be_operand set of each argument of the remat op */
 494         for (n = get_irn_arity(remat->op)-1; n>=0; --n) {
 495                 ir_node        *arg = get_irn_n(remat->op, n);
 496
 497                 query.irn = arg;
 498                 query.remats = NULL;
 499                 query.remats_by_operand = NULL;
 500                 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
 501
 502                 if(remat_info->remats_by_operand == NULL) {
 503                         remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
 504                 }
 505                 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
 506         }
 507 }
 508
 509 static int
 510 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
 511 {
 512         const ir_edge_t   *edge = get_irn_out_edge_first(irn);
 513         int                i = 0;
 514
 515         while(edge) {
 516                 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
 517                         ++i;
 518                 }
 519                 edge = get_irn_out_edge_next(irn, edge);
 520         }
 521
 522         return i;
 523 }
 524
 525 static INLINE void
 526 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
 527 {
 528         int      n;
 529         remat_t *remat;
 530
 531 #ifdef NO_SINGLE_USE_REMATS
 532         if(has_reg_class(si, op) && (get_irn_n_nonremat_edges(si, op) > 1)) {
 533 #else
 534         if(has_reg_class(si, op)) {
 535 #endif
 536                 remat = get_remat_from_op(si, op, op);
 537                 if(remat) {
 538                         add_remat(si, remat);
 539                 }
 540         }
 541
 542 #ifdef COLLECT_INVERSE_REMATS
 543         /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
 544            for each arg */
 545         for (n = get_irn_arity(op)-1; n>=0; --n) {
 546                 ir_node        *arg = get_irn_n(op, n);
 547
 548                 if(has_reg_class(si, arg)) {
 549                         /* try to get an inverse remat */
 550                         remat = get_remat_from_op(si, arg, op);
 551                         if(remat) {
 552                                 add_remat(si, remat);
 553                         }
 554                 }
 555         }
 556 #endif
 557
 558 }
 559
 560 static INLINE int
 561 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
 562 {
 563         ir_node *block;
 564         ir_node *def_block = get_nodes_block(val);
 565         int      ret;
 566
 567         if(val == pos)
 568                 return 0;
 569
 570         /* if pos is at end of a basic block */
 571         if(is_Block(pos)) {
 572                 ret = (pos == def_block || block_dominates(def_block, pos));
 573 //              ir_fprintf(stderr, "(def(bb)=%d) ", ret);
 574                 return ret;
 575         }
 576
 577         /* else if this is a normal operation */
 578         block = get_nodes_block(pos);
 579         if(block == def_block) {
 580                 if(!sched_is_scheduled(val)) return 1;
 581
 582                 ret = sched_comes_after(val, pos);
 583 //              ir_fprintf(stderr, "(def(same block)=%d) ",ret);
 584                 return ret;
 585         }
 586
 587         ret = block_dominates(def_block, block);
 588 //      ir_fprintf(stderr, "(def(other block)=%d) ", ret);
 589         return ret;
 590 }
 591
 592 static INLINE ir_node *
 593 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
 594 {
 595     return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->chordal_env->birg->main_env->arch_env);
 596 }
 597
 598 /**
 599  * Returns first non-Phi node of block @p bb
 600  */
 601 static INLINE ir_node *
 602 sched_block_first_nonphi(const ir_node * bb)
 603 {
 604         return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
 605 }
 606
 607 static int
 608 sched_skip_proj_predicator(const ir_node * irn, void * data)
 609 {
 610         return (is_Proj(irn));
 611 }
 612
 613 static INLINE ir_node *
 614 sched_next_nonproj(const ir_node * irn, int forward)
 615 {
 616         return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
 617 }
 618
 619 /**
 620  * Returns next operation node (non-Proj) after @p irn
 621  * or the basic block of this node
 622  */
 623 static INLINE ir_node *
 624 sched_next_op(const ir_node * irn)
 625 {
 626         ir_node *next = sched_next(irn);
 627
 628         if(is_Block(next))
 629                 return next;
 630
 631         return sched_next_nonproj(next, 1);
 632 }
 633
 634 /**
 635  * Returns previous operation node (non-Proj) before @p irn
 636  * or the basic block of this node
 637  */
 638 static INLINE ir_node *
 639 sched_prev_op(const ir_node * irn)
 640 {
 641         ir_node *prev = sched_prev(irn);
 642
 643         if(is_Block(prev))
 644                 return prev;
 645
 646         return sched_next_nonproj(prev, 0);
 647 }
 648
 649 static void
 650 sched_put_after(ir_node * insert, ir_node * irn)
 651 {
 652         if(is_Block(insert)) {
 653                 insert = sched_block_first_nonphi(insert);
 654         } else {
 655                 insert = sched_next_op(insert);
 656         }
 657         sched_add_before(insert, irn);
 658 }
 659
 660 static void
 661 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
 662 {
 663   if(is_Block(insert)) {
 664           insert = sched_block_last_noncf(si, insert);
 665   } else {
 666           insert = sched_next_nonproj(insert, 0);
 667           insert = sched_prev(insert);
 668   }
 669   sched_add_after(insert, irn);
 670 }
 671
 672 /**
 673  * Tells you whether a @p remat can be placed before the irn @p pos
 674  */
 675 static INLINE int
 676 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 677 {
 678         const ir_node   *op = remat->op;
 679         const ir_node   *prev;
 680         int        n,
 681                            res = 1;
 682
 683         if(is_Block(pos)) {
 684                 prev = sched_block_last_noncf(si, pos);
 685                 prev = sched_next_nonproj(prev, 0);
 686         } else {
 687                 prev = sched_prev_op(pos);
 688         }
 689         /* do not remat if the rematted value is defined immediately before this op */
 690         if(prev == remat->op) {
 691                 return 0;
 692         }
 693
 694 #if 0
 695         /* this should be just fine, the following OP will be using this value, right? */
 696
 697         /* only remat AFTER the real definition of a value (?) */
 698         if(!value_is_defined_before(si, pos, remat->value)) {
 699 //              ir_fprintf(stderr, "error(not defined)");
 700                 return 0;
 701         }
 702 #endif
 703
 704         for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
 705                 const ir_node   *arg = get_irn_n(op, n);
 706
 707 #ifdef NO_ENLARGE_L1V3N355
 708                 if(has_reg_class(si, arg) && live) {
 709                         res &= pset_find_ptr(live, arg)?1:0;
 710                 } else {
 711                         res &= value_is_defined_before(si, pos, arg);
 712                 }
 713 #else
 714                 res &= value_is_defined_before(si, pos, arg);
 715 #endif
 716         }
 717
 718         return res;
 719 }
 720
 721 /**
 722  * Tells you whether a @p remat can be placed after the irn @p pos
 723  */
 724 static INLINE int
 725 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 726 {
 727         if(is_Block(pos)) {
 728                 pos = sched_block_first_nonphi(pos);
 729         } else {
 730                 pos = sched_next_op(pos);
 731         }
 732
 733         /* only remat AFTER the real definition of a value (?) */
 734         if(!value_is_defined_before(si, pos, remat->value)) {
 735                 return 0;
 736         }
 737
 738         return can_remat_before(si, remat, pos, live);
 739 }
 740
 741 /**
 742  * Collect potetially rematerializable OPs
 743  */
 744 static void
 745 walker_remat_collector(ir_node * irn, void * data)
 746 {
 747         spill_ilp_t    *si = data;
 748
 749         if(!is_Block(irn) && !is_Phi(irn)) {
 750                 DBG((si->dbg, LEVEL_4, "\t  Processing %+F\n", irn));
 751                 get_remats_from_op(si, irn);
 752         }
 753 }
 754
 755 /**
 756  * Inserts a copy of @p irn before @p pos
 757  */
 758 static ir_node *
 759 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
 760 {
 761         ir_node     *bb;
 762         ir_node     *copy;
 763
 764         bb = is_Block(pos)?pos:get_nodes_block(pos);
 765         copy = exact_copy(irn);
 766
 767         _set_phi_class(copy, NULL);
 768         set_nodes_block(copy, bb);
 769         sched_put_before(si, pos, copy);
 770
 771         return copy;
 772 }
 773
 774 /**
 775  * Inserts a copy of @p irn after @p pos
 776  */
 777 static ir_node *
 778 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
 779 {
 780         ir_node     *bb;
 781         ir_node     *copy;
 782
 783         bb = is_Block(pos)?pos:get_nodes_block(pos);
 784         copy = exact_copy(irn);
 785
 786         _set_phi_class(copy, NULL);
 787         set_nodes_block(copy, bb);
 788         sched_put_after(pos, copy);
 789
 790         return copy;
 791 }
 792
 793 static ir_node *
 794 insert_remat_after(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 795 {
 796         char     buf[256];
 797
 798         if(can_remat_after(si, remat, pos, live)) {
 799                 ir_node         *copy,
 800                                                 *proj_copy;
 801                 op_t            *op;
 802
 803                 DBG((si->dbg, LEVEL_3, "\t  >inserting remat %+F\n", remat->op));
 804
 805                 copy = insert_copy_after(si, remat->op, pos);
 806
 807                 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
 808                 op = obstack_alloc(si->obst, sizeof(*op));
 809                 op->is_remat = 1;
 810                 op->attr.remat.remat = remat;
 811                 op->attr.remat.pre = 0;
 812                 op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos));
 813
 814                 set_irn_link(copy, op);
 815                 pset_insert_ptr(si->all_possible_remats, copy);
 816                 if(remat->proj) {
 817                         proj_copy = insert_copy_after(si, remat->proj, copy);
 818                         set_irn_n(proj_copy, 0, copy);
 819                         set_irn_link(proj_copy, op);
 820                         pset_insert_ptr(si->all_possible_remats, proj_copy);
 821                 } else {
 822                         proj_copy = NULL;
 823                 }
 824
 825                 return copy;
 826         }
 827
 828         return NULL;
 829 }
 830
 831 static ir_node *
 832 insert_remat_before(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 833 {
 834         char     buf[256];
 835
 836         if(can_remat_before(si, remat, pos, live)) {
 837                 ir_node         *copy,
 838                                                 *proj_copy;
 839                 op_t            *op;
 840
 841                 DBG((si->dbg, LEVEL_3, "\t  >inserting remat %+F\n", remat->op));
 842
 843                 copy = insert_copy_before(si, remat->op, pos);
 844
 845                 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
 846                 op = obstack_alloc(si->obst, sizeof(*op));
 847                 op->is_remat = 1;
 848                 op->attr.remat.remat = remat;
 849                 op->attr.remat.pre = 1;
 850                 op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos));
 851
 852                 set_irn_link(copy, op);
 853                 pset_insert_ptr(si->all_possible_remats, copy);
 854                 if(remat->proj) {
 855                         proj_copy = insert_copy_after(si, remat->proj, copy);
 856                         set_irn_n(proj_copy, 0, copy);
 857                         set_irn_link(proj_copy, op);
 858                         pset_insert_ptr(si->all_possible_remats, proj_copy);
 859                 } else {
 860                         proj_copy = NULL;
 861                 }
 862
 863                 return copy;
 864         }
 865
 866         return NULL;
 867 }
 868
 869 static int
 870 get_block_n_succs(const ir_node *block) {
 871         const ir_edge_t *edge;
 872
 873         assert(edges_activated(current_ir_graph));
 874
 875         edge = get_block_succ_first(block);
 876         if (! edge)
 877                 return 0;
 878
 879         edge = get_block_succ_next(block, edge);
 880         return edge ? 2 : 1;
 881 }
 882
 883 static int
 884 is_merge_edge(const ir_node * bb)
 885 {
 886 #ifdef GOODWIN_REDUCTION
 887         return get_block_n_succs(bb) == 1;
 888 #else
 889         return 1;
 890 #endif
 891 }
 892
 893 static int
 894 is_diverge_edge(const ir_node * bb)
 895 {
 896 #ifdef GOODWIN_REDUCTION
 897         return get_Block_n_cfgpreds(bb) == 1;
 898 #else
 899         return 1;
 900 #endif
 901 }
 902
 903 static void
 904 walker_regclass_copy_insertor(ir_node * irn, void * data)
 905 {
 906         spill_ilp_t    *si = data;
 907
 908         if(is_Phi(irn) && has_reg_class(si, irn)) {
 909                 int n;
 910
 911                 for(n=get_irn_arity(irn)-1; n>=0; --n) {
 912                         ir_node  *phi_arg = get_irn_n(irn, n);
 913                         ir_node  *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
 914
 915                         if(!has_reg_class(si, phi_arg)) {
 916                                 ir_node   *copy = be_new_Copy(si->cls, si->chordal_env->irg, bb, phi_arg);
 917                                 ir_node   *pos = sched_block_last_noncf(si, bb);
 918                                 op_t      *op = obstack_alloc(si->obst, sizeof(*op));
 919
 920                                 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
 921                                 sched_add_after(pos, copy);
 922                                 set_irn_n(irn, n, copy);
 923
 924                                 op->is_remat = 0;
 925                                 op->attr.live_range.args.reloads = NULL;
 926                                 op->attr.live_range.ilp = ILP_UNDEF;
 927                                 set_irn_link(copy, op);
 928                         }
 929                 }
 930         }
 931 }
 932
 933
 934 /**
 935  * Insert (so far unused) remats into the irg to
 936  * recompute the potential liveness of all values
 937  */
 938 static void
 939 walker_remat_insertor(ir_node * bb, void * data)
 940 {
 941         spill_ilp_t    *si = data;
 942         spill_bb_t     *spill_bb;
 943         ir_node        *irn;
 944         int             n, i;
 945         pset           *live = pset_new_ptr_default();
 946
 947         DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
 948
 949         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
 950                 ir_node        *value = be_lv_get_irn(si->lv, bb, i);
 951
 952                 /* add remats at end of block */
 953                 if (has_reg_class(si, value)) {
 954                         pset_insert_ptr(live, value);
 955                 }
 956         }
 957
 958         spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
 959         set_irn_link(bb, spill_bb);
 960
 961         irn = sched_last(bb);
 962         while(!sched_is_end(irn)) {
 963                 ir_node   *next;
 964                 op_t      *op;
 965                 pset      *args;
 966                 ir_node   *arg;
 967                 pset      *remat_args;
 968
 969                 next = sched_prev(irn);
 970
 971                 DBG((si->dbg, LEVEL_5, "\t at %+F (next: %+F)\n", irn, next));
 972
 973                 if(is_Phi(irn) || is_Proj(irn)) {
 974                         op_t      *op;
 975
 976                         if(has_reg_class(si, irn)) {
 977                                 pset_remove_ptr(live, irn);
 978                         }
 979
 980                         op = obstack_alloc(si->obst, sizeof(*op));
 981                         op->is_remat = 0;
 982                         op->attr.live_range.args.reloads = NULL;
 983                         op->attr.live_range.ilp = ILP_UNDEF;
 984                         set_irn_link(irn, op);
 985
 986                         irn = next;
 987                         continue;
 988                 }
 989
 990                 op = obstack_alloc(si->obst, sizeof(*op));
 991                 op->is_remat = 0;
 992                 op->attr.live_range.ilp = ILP_UNDEF;
 993                 op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
 994                 memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
 995                 set_irn_link(irn, op);
 996
 997                 args = pset_new_ptr_default();
 998
 999                 /* collect arguments of op */
1000                 for (n = get_irn_arity(irn)-1; n>=0; --n) {
1001                         ir_node        *arg = get_irn_n(irn, n);
1002
1003                         pset_insert_ptr(args, arg);
1004                 }
1005
1006                 /* set args of op already live in epilog */
1007                 pset_foreach(args, arg) {
1008                         if(has_reg_class(si, arg)) {
1009                                 pset_insert_ptr(live, arg);
1010                         }
1011                 }
1012                 /* delete defined value from live set */
1013                 if(has_reg_class(si, irn)) {
1014                         pset_remove_ptr(live, irn);
1015                 }
1016
1017
1018                 remat_args = pset_new_ptr_default();
1019
1020                 /* insert all possible remats before irn */
1021                 pset_foreach(args, arg) {
1022                         remat_info_t   *remat_info,
1023                                                     query;
1024                         remat_t        *remat;
1025
1026                         /* continue if the operand has the wrong reg class
1027                          */
1028                         if(!has_reg_class(si, arg))
1029                                 continue;
1030
1031                         query.irn = arg;
1032                         query.remats = NULL;
1033                         query.remats_by_operand = NULL;
1034                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1035
1036                         if(!remat_info) {
1037                                 continue;
1038                         }
1039
1040                         if(remat_info->remats) {
1041                                 pset_foreach(remat_info->remats, remat) {
1042                                         ir_node  *remat_irn = NULL;
1043
1044                                         DBG((si->dbg, LEVEL_4, "\t  considering remat %+F for arg %+F\n", remat->op, arg));
1045 #ifdef REMAT_WHILE_LIVE
1046                                         if(pset_find_ptr(live, remat->value)) {
1047                                                 remat_irn = insert_remat_before(si, remat, irn, live);
1048                                         }
1049 #else
1050                                         remat_irn = insert_remat_before(si, remat, irn, live);
1051 #endif
1052                                         if(remat_irn) {
1053                                                 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1054                                                         ir_node  *remat_arg = get_irn_n(remat_irn, n);
1055
1056                                                         if(!has_reg_class(si, remat_arg)) continue;
1057
1058                                                         pset_insert_ptr(remat_args, remat_arg);
1059                                                 }
1060                                         }
1061                                 }
1062                         }
1063                 }
1064
1065                 /* now we add remat args to op's args because they could also die at this op */
1066                 pset_foreach(args,arg) {
1067                         if(pset_find_ptr(remat_args, arg)) {
1068                                 pset_remove_ptr(remat_args, arg);
1069                         }
1070                 }
1071                 pset_foreach(remat_args,arg) {
1072                         pset_insert_ptr(args, arg);
1073                 }
1074
1075                 /* insert all possible remats after irn */
1076                 pset_foreach(args, arg) {
1077                         remat_info_t   *remat_info,
1078                                                     query;
1079                         remat_t        *remat;
1080
1081                         /* continue if the operand has the wrong reg class */
1082                         if(!has_reg_class(si, arg))
1083                                 continue;
1084
1085                         query.irn = arg;
1086                         query.remats = NULL;
1087                         query.remats_by_operand = NULL;
1088                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1089
1090                         if(!remat_info) {
1091                                 continue;
1092                         }
1093
1094                         /* do not place post remats after jumps */
1095                         if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) continue;
1096
1097                         if(remat_info->remats_by_operand) {
1098                                 pset_foreach(remat_info->remats_by_operand, remat) {
1099                                         /* do not insert remats producing the same value as one of the operands */
1100                                         if(!pset_find_ptr(args, remat->value)) {
1101                                                 DBG((si->dbg, LEVEL_4, "\t  considering remat %+F with arg %+F\n", remat->op, arg));
1102 #ifdef REMAT_WHILE_LIVE
1103                                                 if(pset_find_ptr(live, remat->value)) {
1104                                                         insert_remat_after(si, remat, irn, live);
1105                                                 }
1106 #else
1107                                                 insert_remat_after(si, remat, irn, live);
1108 #endif
1109                                         }
1110                                 }
1111                         }
1112                 }
1113
1114                 del_pset(remat_args);
1115                 del_pset(args);
1116                 irn = next;
1117         }
1118
1119         be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_in, i) {
1120                 ir_node        *value = be_lv_get_irn(si->lv, bb, i);
1121
1122                 /* add remats at end if successor has multiple predecessors */
1123                 if(is_merge_edge(bb)) {
1124                         /* add remats at end of block */
1125                         if (be_is_live_end(si->lv, bb, value) && has_reg_class(si, value)) {
1126                                 remat_info_t   *remat_info,
1127                                                            query;
1128                                 remat_t        *remat;
1129
1130                                 query.irn = value;
1131                                 query.remats = NULL;
1132                                 query.remats_by_operand = NULL;
1133                                 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1134
1135                                 if(remat_info && remat_info->remats) {
1136                                         pset_foreach(remat_info->remats, remat) {
1137                                                 DBG((si->dbg, LEVEL_4, "\t  considering remat %+F at end of block %+F\n", remat->op, bb));
1138
1139                                                 insert_remat_before(si, remat, bb, NULL);
1140                                         }
1141                                 }
1142                         }
1143                 }
1144                 if(is_diverge_edge(bb)) {
1145                         /* add remat2s at beginning of block */
1146                         if ((be_is_live_in(si->lv, bb, value) || (is_Phi(value) && get_nodes_block(value)==bb)) && has_reg_class(si, value)) {
1147                                 remat_info_t   *remat_info,
1148                                                            query;
1149                                 remat_t        *remat;
1150
1151                                 query.irn = value;
1152                                 query.remats = NULL;
1153                                 query.remats_by_operand = NULL;
1154                                 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1155
1156                                 if(remat_info && remat_info->remats) {
1157                                         pset_foreach(remat_info->remats, remat) {
1158                                                 DBG((si->dbg, LEVEL_4, "\t  considering remat %+F at beginning of block %+F\n", remat->op, bb));
1159
1160                                                 /* put the remat here if all its args are available */
1161                                                 insert_remat_after(si, remat, bb, NULL);
1162
1163                                         }
1164                                 }
1165                         }
1166                 }
1167         }
1168 }
1169
1170 /**
1171  * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1172  */
1173 static void
1174 luke_endwalker(ir_node * bb, void * data)
1175 {
1176         spill_ilp_t    *si = (spill_ilp_t*)data;
1177         pset           *live;
1178         pset           *use_end;
1179         char            buf[256];
1180         ilp_cst_t       cst;
1181         ir_node        *irn;
1182         spill_bb_t     *spill_bb = get_irn_link(bb);
1183         int             i;
1184
1185
1186         live = pset_new_ptr_default();
1187         use_end = pset_new_ptr_default();
1188
1189         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1190                 irn = be_lv_get_irn(si->lv, bb, i);
1191                 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1192                         op_t      *op;
1193
1194                         pset_insert_ptr(live, irn);
1195                         op = get_irn_link(irn);
1196                         assert(!op->is_remat);
1197                 }
1198         }
1199
1200         /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1201         /* their reg_out must always be set */
1202         sched_foreach_reverse(bb, irn) {
1203                 int   n;
1204
1205                 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1206
1207                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1208                         ir_node        *irn_arg = get_irn_n(irn, n);
1209
1210                         if(has_reg_class(si, irn_arg)) {
1211                                 pset_insert_ptr(use_end, irn_arg);
1212                         }
1213                 }
1214         }
1215
1216         ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1217         //cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1218         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1219
1220         spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1221
1222         /* if this is a merge edge we can reload at the end of this block */
1223         if(is_merge_edge(bb)) {
1224                 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1225         } else if(pset_count(use_end)){
1226                 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1227         } else {
1228                 spill_bb->reloads = NULL;
1229         }
1230
1231         pset_foreach(live,irn) {
1232                 spill_t     query,
1233                                         *spill;
1234                 double      spill_cost;
1235
1236
1237                 /* handle values used by control flow nodes later separately */
1238                 if(pset_find_ptr(use_end, irn)) continue;
1239
1240                 query.irn = irn;
1241                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1242
1243                 spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1244
1245                 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1246                 spill->reg_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1247                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1248
1249                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1250                 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1251
1252                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1253                 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost);
1254
1255                 if(is_merge_edge(bb)) {
1256                         ilp_var_t   reload;
1257                         ilp_cst_t   rel_cst;
1258
1259                         ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1260                         reload = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
1261                         set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1262
1263                         /* reload <= mem_out */
1264                         rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1265                         lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1266                         lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1267                 }
1268
1269                 spill->reg_in = ILP_UNDEF;
1270                 spill->mem_in = ILP_UNDEF;
1271         }
1272
1273         pset_foreach(use_end,irn) {
1274                 spill_t     query,
1275                                         *spill;
1276                 double      spill_cost;
1277                 ilp_cst_t   end_use_req,
1278                                         rel_cst;
1279                 ilp_var_t   reload;
1280
1281                 query.irn = irn;
1282                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1283
1284                 spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1285
1286                 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1287                 spill->reg_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1288                 /* if irn is used at the end of the block, then it is live anyway */
1289                 //lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1290
1291                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1292                 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1293
1294                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1295                 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost);
1296
1297                 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1298                 reload = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
1299                 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1300
1301                 /* reload <= mem_out */
1302                 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1303                 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1304                 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1305
1306                 spill->reg_in = ILP_UNDEF;
1307                 spill->mem_in = ILP_UNDEF;
1308
1309                 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1310                 end_use_req = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1);
1311                 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1312         }
1313
1314         del_pset(live);
1315         del_pset(use_end);
1316 }
1317
1318 static ir_node *
1319 next_post_remat(const ir_node * irn)
1320 {
1321         op_t      *op;
1322
1323         if(is_Block(irn)) {
1324                 irn = sched_block_first_nonphi(irn);
1325         } else {
1326                 irn = sched_next_op(irn);
1327         }
1328
1329         if(sched_is_end(irn))
1330                 return NULL;
1331
1332         op = (op_t*)get_irn_link(irn);
1333         if(op->is_remat && !op->attr.remat.pre) {
1334                 return irn;
1335         }
1336
1337         return NULL;
1338 }
1339
1340
1341 static ir_node *
1342 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
1343 {
1344         op_t      *op;
1345         ir_node   *ret;
1346
1347         if(is_Block(irn)) {
1348                 ret = sched_block_last_noncf(si, irn);
1349                 ret = sched_next(ret);
1350                 ret = sched_prev_op(ret);
1351         } else {
1352                 ret = sched_prev_op(irn);
1353         }
1354
1355         if(sched_is_end(ret) || is_Phi(ret))
1356                 return NULL;
1357
1358         op = (op_t*)get_irn_link(ret);
1359         if(op->is_remat && op->attr.remat.pre) {
1360                 return ret;
1361         }
1362
1363         return NULL;
1364 }
1365
1366 /**
1367  * Find a remat of value @p value in the epilog of @p pos
1368  */
1369 static ir_node *
1370 find_post_remat(const ir_node * value, const ir_node * pos)
1371 {
1372         while((pos = next_post_remat(pos)) != NULL) {
1373                 op_t   *op;
1374
1375                 op = get_irn_link(pos);
1376                 assert(op->is_remat && !op->attr.remat.pre);
1377
1378                 if(op->attr.remat.remat->value == value)
1379                         return (ir_node*)pos;
1380
1381 #if 0
1382         const ir_edge_t *edge;
1383                 foreach_out_edge(pos, edge) {
1384                         ir_node   *proj = get_edge_src_irn(edge);
1385                         assert(is_Proj(proj));
1386                 }
1387 #endif
1388
1389         }
1390
1391         return NULL;
1392 }
1393
1394 static spill_t *
1395 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1396 {
1397         spill_bb_t  *spill_bb = get_irn_link(bb);
1398         spill_t     *spill,
1399                                  query;
1400         char         buf[256];
1401
1402         query.irn = irn;
1403         spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1404         if(!spill) {
1405                 double   spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1406
1407                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1408
1409                 spill->reg_out = ILP_UNDEF;
1410                 spill->reg_in  = ILP_UNDEF;
1411                 spill->mem_in  = ILP_UNDEF;
1412
1413                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1414                 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1415
1416                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1417                 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost);
1418         }
1419
1420         return spill;
1421 }
1422
1423 static void
1424 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1425 {
1426         ir_node        *irn;
1427         int i;
1428
1429         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1430                 irn = be_lv_get_irn(si->lv, bb, i);
1431
1432                 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1433                         pset_insert_ptr(live, irn);
1434                 }
1435         }
1436
1437         irn = sched_last(bb);
1438
1439         /* all values eaten by control flow operations are also live until the end of the block */
1440         sched_foreach_reverse(bb, irn) {
1441                 int  i;
1442
1443                 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1444
1445                 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1446                         ir_node *arg = get_irn_n(irn,i);
1447
1448                         if(has_reg_class(si, arg)) {
1449                                 pset_insert_ptr(live, arg);
1450                         }
1451                 }
1452         }
1453 }
1454
1455 /**
1456  *  Inserts ILP-constraints and variables for memory copying before the given position
1457  */
1458 static void
1459 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1460 {
1461         const ir_node    *succ;
1462         const ir_edge_t  *edge;
1463         spill_bb_t       *spill_bb = get_irn_link(block);
1464         ir_node          *phi;
1465         int               pos;
1466         ilp_cst_t         cst;
1467         ilp_var_t         copyreg;
1468         char              buf[256];
1469         ir_node          *tmp;
1470
1471
1472         assert(edges_activated(current_ir_graph));
1473
1474         edge = get_block_succ_first(block);
1475         if(!edge) return;
1476
1477         succ = edge->src;
1478         pos = edge->pos;
1479
1480         edge = get_block_succ_next(block, edge);
1481         /* next block can only contain phis, if this is a merge edge */
1482         if(edge) return;
1483
1484         ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1485         copyreg = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1486
1487         ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1488         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1489
1490         pset_foreach(live, tmp) {
1491                 spill_t  *spill;
1492 #if 0
1493                 op_t  *op = get_irn_link(irn);
1494                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1495 #endif
1496                 spill = set_find_spill(spill_bb->ilp, tmp);
1497                 assert(spill);
1498
1499                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1500         }
1501         lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1502
1503         sched_foreach(succ, phi) {
1504                 const ir_node  *to_copy;
1505                 op_t           *to_copy_op;
1506                 spill_t        *to_copy_spill;
1507                 op_t           *phi_op = get_irn_link(phi);
1508                 ilp_var_t       reload = ILP_UNDEF;
1509
1510
1511                 if(!is_Phi(phi)) break;
1512                 if(!has_reg_class(si, phi)) continue;
1513
1514                 to_copy = get_irn_n(phi, pos);
1515
1516                 to_copy_op = get_irn_link(to_copy);
1517
1518                 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1519                 assert(to_copy_spill);
1520
1521                 if(spill_bb->reloads) {
1522                         keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1523
1524                         if(keyval) {
1525                                 reload = PTR_TO_INT(keyval->val);
1526                         }
1527                 }
1528
1529                 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N_%N", block, phi, to_copy);
1530                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1531
1532                 /* copy - reg_out - reload - remat - live_range <= 0 */
1533                 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1534                 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1535                 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1536                 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1537                 foreach_pre_remat(si, block, tmp) {
1538                         op_t     *remat_op = get_irn_link(tmp);
1539                         if(remat_op->attr.remat.remat->value == to_copy) {
1540                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1541                         }
1542                 }
1543
1544                 ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N_%N", block, phi, to_copy);
1545                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1546
1547                 /* copy - reg_out - copyreg <= 0 */
1548                 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1549                 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1550                 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1551         }
1552 }
1553
1554
1555 /**
1556  * Walk all irg blocks and emit this ILP
1557  */
1558 static void
1559 luke_blockwalker(ir_node * bb, void * data)
1560 {
1561         spill_ilp_t    *si = (spill_ilp_t*)data;
1562         ir_node        *irn;
1563         pset           *live;
1564         char            buf[256];
1565         ilp_cst_t       cst;
1566         spill_bb_t     *spill_bb = get_irn_link(bb);
1567         ir_node        *tmp;
1568         spill_t        *spill;
1569         pset           *defs = pset_new_ptr_default();
1570 #ifdef WITH_MEMOPERANDS
1571         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
1572 #endif
1573
1574
1575         live = pset_new_ptr_default();
1576
1577         /****************************************
1578          *      B A S I C  B L O C K  E N D
1579          ***************************************/
1580
1581
1582         /* init live values at end of block */
1583         get_live_end(si, bb, live);
1584
1585         pset_foreach(live, irn) {
1586                 op_t           *op;
1587                 ilp_var_t       reload = ILP_UNDEF;
1588
1589                 spill = set_find_spill(spill_bb->ilp, irn);
1590                 assert(spill);
1591
1592                 if(spill_bb->reloads) {
1593                         keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1594
1595                         if(keyval) {
1596                                 reload = PTR_TO_INT(keyval->val);
1597                         }
1598                 }
1599
1600                 op = get_irn_link(irn);
1601                 assert(!op->is_remat);
1602
1603                 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1604                 op->attr.live_range.ilp = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1605                 op->attr.live_range.op = bb;
1606
1607                 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1608                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1609
1610                 /* reg_out - reload - remat - live_range <= 0 */
1611                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1612                 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1613                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1614                 foreach_pre_remat(si, bb, tmp) {
1615                         op_t     *remat_op = get_irn_link(tmp);
1616                         if(remat_op->attr.remat.remat->value == irn) {
1617                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1618                         }
1619                 }
1620                 /* maybe we should also assure that reg_out >= live_range etc. */
1621         }
1622
1623 #ifndef NO_MEMCOPIES
1624         insert_mem_copy_position(si, live, bb);
1625 #endif
1626
1627         /*
1628          * start new live ranges for values used by remats at end of block
1629          * and assure the remat args are available
1630          */
1631         foreach_pre_remat(si, bb, tmp) {
1632                 op_t     *remat_op = get_irn_link(tmp);
1633                 int       n;
1634
1635                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1636                         ir_node        *remat_arg = get_irn_n(tmp, n);
1637                         op_t           *arg_op = get_irn_link(remat_arg);
1638                         ilp_var_t       prev_lr;
1639
1640                         if(!has_reg_class(si, remat_arg)) continue;
1641
1642                         /* if value is becoming live through use by remat */
1643                         if(!pset_find_ptr(live, remat_arg)) {
1644                                 ir_snprintf(buf, sizeof(buf), "lr_%N_end%N", remat_arg, bb);
1645                                 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1646
1647                                 arg_op->attr.live_range.ilp = prev_lr;
1648                                 arg_op->attr.live_range.op = bb;
1649
1650                                 DBG((si->dbg, LEVEL_4, "  value %+F becoming live through use by remat at end of block %+F\n", remat_arg, tmp));
1651
1652                                 pset_insert_ptr(live, remat_arg);
1653                                 add_to_spill_bb(si, bb, remat_arg);
1654                         }
1655
1656                         /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
1657                         ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1658                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1659
1660                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1661                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1662
1663                         /* use reload placed for this argument */
1664                         if(spill_bb->reloads) {
1665                                 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1666
1667                                 if(keyval) {
1668                                         ilp_var_t       reload = PTR_TO_INT(keyval->val);
1669
1670                                         lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1671                                 }
1672                         }
1673                 }
1674         }
1675         DBG((si->dbg, LEVEL_4, "\t   %d values live at end of block %+F\n", pset_count(live), bb));
1676
1677
1678
1679
1680         /**************************************
1681          *    B A S I C  B L O C K  B O D Y
1682          **************************************/
1683
1684         sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1685                 op_t       *op;
1686                 op_t       *tmp_op;
1687                 int         n,
1688                                         u = 0,
1689                                         d = 0;
1690                 ilp_cst_t       check_pre,
1691                                         check_post;
1692                 set        *args;
1693                 pset       *used;
1694                 pset       *remat_defs;
1695                 keyval_t   *keyval;
1696 #ifdef WITH_MEMOPERANDS
1697                 ilp_cst_t   one_memoperand;
1698 #endif
1699
1700                 /* iterate only until first phi */
1701                 if(is_Phi(irn))
1702                         break;
1703
1704                 op = get_irn_link(irn);
1705                 /* skip remats */
1706                 if(op->is_remat) continue;
1707                 DBG((si->dbg, LEVEL_4, "\t  at node %+F\n", irn));
1708
1709                 /* collect defined values */
1710                 if(has_reg_class(si, irn)) {
1711                         pset_insert_ptr(defs, irn);
1712                 }
1713
1714                 /* skip projs */
1715                 if(is_Proj(irn)) continue;
1716
1717                 /*
1718                  * init set of irn's arguments
1719                  * and all possibly used values around this op
1720                  * and values defined by post remats
1721                  */
1722                 args =       new_set(cmp_keyval, get_irn_arity(irn));
1723                 used =       pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1724                 remat_defs = pset_new_ptr(pset_count(live));
1725
1726                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1727                         ir_node        *irn_arg = get_irn_n(irn, n);
1728                         if(has_reg_class(si, irn_arg)) {
1729                                 set_insert_keyval(args, irn_arg, (void*)n);
1730                                 pset_insert_ptr(used, irn_arg);
1731                         }
1732                 }
1733                 foreach_post_remat(irn, tmp) {
1734                         op_t    *remat_op = get_irn_link(tmp);
1735
1736                         pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1737
1738                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1739                                 ir_node        *remat_arg = get_irn_n(tmp, n);
1740                                 if(has_reg_class(si, remat_arg)) {
1741                                         pset_insert_ptr(used, remat_arg);
1742                                 }
1743                         }
1744                 }
1745                 foreach_pre_remat(si, irn, tmp) {
1746                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1747                                 ir_node        *remat_arg = get_irn_n(tmp, n);
1748                                 if(has_reg_class(si, remat_arg)) {
1749                                         pset_insert_ptr(used, remat_arg);
1750                                 }
1751                         }
1752                 }
1753
1754                 /**********************************
1755                  *   I N  E P I L O G  O F  irn
1756                  **********************************/
1757
1758                 /* ensure each dying value is used by only one post remat */
1759                 pset_foreach(used, tmp) {
1760                         ir_node     *value = tmp;
1761                         op_t        *value_op = get_irn_link(value);
1762                         ir_node     *remat;
1763                         int          n_remats = 0;
1764
1765                         cst = ILP_UNDEF;
1766                         foreach_post_remat(irn, remat) {
1767                                 op_t  *remat_op = get_irn_link(remat);
1768
1769                                 for(n=get_irn_arity(remat)-1; n>=0; --n) {
1770                                         ir_node   *remat_arg = get_irn_n(remat, n);
1771
1772                                         /* if value is used by this remat add it to constraint */
1773                                         if(remat_arg == value) {
1774                                                 if(n_remats == 0) {
1775                                                         /* sum remat2s <= 1 + n_remats*live_range */
1776                                                         ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
1777                                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
1778                                                 }
1779
1780                                                 n_remats++;
1781                                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1782                                                 break;
1783                                         }
1784                                 }
1785                         }
1786
1787             // value_op->attr.live_range.ilp != ILP_UNDEF
1788                         if(pset_find_ptr(live, value) && cst != ILP_UNDEF) {
1789                                 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
1790                         }
1791                 }
1792
1793         /* ensure at least one value dies at post remat */
1794         foreach_post_remat(irn, tmp) {
1795             op_t     *remat_op = get_irn_link(tmp);
1796             pset     *remat_args = pset_new_ptr(get_irn_arity(tmp));
1797             ir_node  *remat_arg;
1798
1799             for(n=get_irn_arity(tmp)-1; n>=0; --n) {
1800                 remat_arg = get_irn_n(tmp, n);
1801
1802                 if(has_reg_class(si, remat_arg)) {
1803
1804                     /* does arg always die at this op? */
1805                     if(!pset_find_ptr(live, remat_arg))
1806                         goto skip_one_must_die;
1807
1808                     pset_insert_ptr(remat_args, remat_arg);
1809                 }
1810             }
1811
1812             /* remat + \sum live_range(remat_arg) <= |args| */
1813             ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp);
1814             cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
1815             lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1816
1817             pset_foreach(remat_args, remat_arg) {
1818                 op_t  *arg_op = get_irn_link(remat_arg);
1819
1820                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1821             }
1822
1823 skip_one_must_die:
1824             del_pset(remat_args);
1825         }
1826
1827                 /* new live ranges for values from L\U defined by post remats */
1828                 pset_foreach(live, tmp) {
1829                         ir_node     *value = tmp;
1830                         op_t        *value_op = get_irn_link(value);
1831
1832                         if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
1833                                 ilp_var_t    prev_lr = ILP_UNDEF;
1834                                 ir_node     *remat;
1835
1836                                 if(pset_find_ptr(remat_defs, value)) {
1837
1838                                         /* next_live_range <= prev_live_range + sum remat2s */
1839                                         ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
1840                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1841
1842                                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
1843                                         prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1844
1845                                         lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
1846                                         lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1847
1848                                         foreach_post_remat(irn, remat) {
1849                                                 op_t        *remat_op = get_irn_link(remat);
1850
1851                                                 /* if value is being rematerialized by this remat */
1852                                                 if(value == remat_op->attr.remat.remat->value) {
1853                                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1854                                                 }
1855                                         }
1856
1857                                         value_op->attr.live_range.ilp = prev_lr;
1858                                         value_op->attr.live_range.op = irn;
1859                                 }
1860                         }
1861                 }
1862
1863                 /* requirements for post remats and start live ranges from L/U' for values dying here */
1864                 foreach_post_remat(irn, tmp) {
1865                         op_t        *remat_op = get_irn_link(tmp);
1866                         int          n;
1867
1868                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1869                                 ir_node        *remat_arg = get_irn_n(tmp, n);
1870                                 op_t           *arg_op = get_irn_link(remat_arg);
1871
1872                                 if(!has_reg_class(si, remat_arg)) continue;
1873
1874                                 /* only for values in L\U (TODO and D?), the others are handled with post_use */
1875                                 if(!pset_find_ptr(used, remat_arg)) {
1876                                         /* remat <= live_range(remat_arg) */
1877                                         ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
1878                                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1879
1880                                         /* if value is becoming live through use by remat2 */
1881                                         if(!pset_find_ptr(live, remat_arg)) {
1882                                                 ilp_var_t     lr;
1883
1884                                                 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
1885                                                 lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1886
1887                                                 arg_op->attr.live_range.ilp = lr;
1888                                                 arg_op->attr.live_range.op = irn;
1889
1890                                                 DBG((si->dbg, LEVEL_3, "  value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
1891
1892                                                 pset_insert_ptr(live, remat_arg);
1893                                                 add_to_spill_bb(si, bb, remat_arg);
1894                                         }
1895
1896                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1897                                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1898                                 }
1899                         }
1900                 }
1901
1902                 d = pset_count(defs);
1903                 DBG((si->dbg, LEVEL_4, "\t   %+F produces %d values in my register class\n", irn, d));
1904
1905                 /* count how many regs irn needs for arguments */
1906                 u = set_count(args);
1907
1908
1909                 /* check the register pressure in the epilog */
1910                 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
1911                 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
1912                 check_post = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - d);
1913
1914                 /* add L\U' to check_post */
1915                 pset_foreach(live, tmp) {
1916                         if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
1917                                 /* if a live value is not used by irn */
1918                                 tmp_op = get_irn_link(tmp);
1919                                 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
1920                         }
1921                 }
1922
1923                 /***********************************************************
1924                  *  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
1925                  **********************************************************/
1926
1927
1928                 pset_foreach(used, tmp) {
1929                         ilp_var_t       prev_lr;
1930                         ilp_var_t       post_use;
1931                         int             p = 0;
1932                         spill_t        *spill;
1933                         ir_node        *arg = tmp;
1934                         op_t           *arg_op = get_irn_link(arg);
1935                         ir_node        *remat;
1936
1937                         spill = add_to_spill_bb(si, bb, arg);
1938
1939                         /* new live range for each used value */
1940                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
1941                         prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1942
1943                         /* the epilog stuff - including post_use, check_post, check_post_remat */
1944                         ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
1945                         post_use = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1946
1947                         lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
1948
1949                         /* arg is live throughout epilog if the next live_range is in a register */
1950                         if(pset_find_ptr(live, arg)) {
1951                                 DBG((si->dbg, LEVEL_3, "\t  arg %+F is possibly live in epilog of %+F\n", arg, irn));
1952
1953                                 /* post_use >= next_lr + remat */
1954                                 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1955                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1956                                 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
1957                                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1958
1959                         }
1960
1961                         /* if value is not an arg of op and not possibly defined by post remat
1962                          * then it may only die and not become live
1963                          */
1964                         if(!set_find_keyval(args, arg)) {
1965                                 /* post_use <= prev_lr */
1966                                 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
1967                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1968                                 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
1969                                 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1970
1971                                 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
1972                                         /* next_lr <= prev_lr */
1973                                         ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
1974                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1975                                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1976                                         lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1977                                 }
1978                         }
1979
1980
1981                         /* forall post remat which use arg add a similar cst */
1982                         foreach_post_remat(irn, remat) {
1983                                 int      n;
1984
1985                                 for (n=get_irn_arity(remat)-1; n>=0; --n) {
1986                                         ir_node    *remat_arg = get_irn_n(remat, n);
1987                                         op_t       *remat_op = get_irn_link(remat);
1988
1989                                         if(remat_arg == arg) {
1990                                                 DBG((si->dbg, LEVEL_3, "\t  found remat with arg %+F in epilog of %+F\n", arg, irn));
1991
1992                                                 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1993                                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1994                                                 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
1995                                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1996                                         }
1997                                 }
1998                         }
1999
2000 #ifdef WITH_MEMOPERANDS
2001                         for(n = get_irn_arity(irn)-1; n>=0; --n) {
2002                                 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2003                                         ilp_var_t       memoperand;
2004
2005                                         ir_snprintf(buf, sizeof(buf), "memoperand_%N_%d", irn, n);
2006                                         memoperand = lpp_add_var(si->lpp, buf, lpp_binary, COST_MEMOPERAND*execution_frequency(si, bb));
2007                                         set_insert_memoperand(si->memoperands, irn, n, memoperand);
2008
2009                                         ir_snprintf(buf, sizeof(buf), "nolivepost_%N_%d", irn, n);
2010                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2011
2012                                         lpp_set_factor_fast(si->lpp, cst, memoperand, 1.0);
2013                                         lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2014 //                                      if(arg_op->attr.live_range.ilp != ILP_UNDEF)
2015 //                                              lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2016                                 }
2017                         }
2018 #endif
2019
2020                         /* new live range begins for each used value */
2021                         arg_op->attr.live_range.ilp = prev_lr;
2022                         arg_op->attr.live_range.op = irn;
2023
2024                         /*if(!pset_find_ptr(live, arg)) {
2025                                 pset_insert_ptr(live, arg);
2026                                 add_to_spill_bb(si, bb, arg);
2027                         }*/
2028                         pset_insert_ptr(live, arg);
2029
2030                 }
2031
2032                 /* just to be sure */
2033                 check_post = ILP_UNDEF;
2034
2035
2036
2037
2038                 /******************
2039                  *   P R O L O G
2040                  ******************/
2041
2042                 /* check the register pressure in the prolog */
2043                 /* sum_{L\U} lr <= k - |U| */
2044                 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
2045                 check_pre = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - u);
2046
2047                 /* for the prolog remove defined values from the live set */
2048                 pset_foreach(defs, tmp) {
2049                         pset_remove_ptr(live, tmp);
2050                 }
2051
2052 #ifdef WITH_MEMOPERANDS
2053                 ir_snprintf(buf, sizeof(buf), "one_memoperand_%N", irn);
2054                 one_memoperand = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2055 #endif
2056
2057                 /***********************************************************
2058                  *  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
2059                  **********************************************************/
2060
2061
2062                 set_foreach(args, keyval) {
2063                         spill_t          *spill;
2064                         ir_node          *arg = keyval->key;
2065                         int               i = PTR_TO_INT(keyval->val);
2066                         op_t             *arg_op = get_irn_link(arg);
2067                         ilp_cst_t         requirements;
2068 #ifdef WITH_MEMOPERANDS
2069                         int               n_memoperands;
2070 #endif
2071
2072                         spill = set_find_spill(spill_bb->ilp, arg);
2073                         assert(spill);
2074
2075                         ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
2076                         op->attr.live_range.args.reloads[i] = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
2077
2078                         /* reload <= mem_out */
2079                         ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
2080                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2081                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2082                         lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2083
2084                         /* requirement: arg must be in register for use */
2085                         /* reload + remat + live_range == 1 */
2086                         ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
2087                         requirements = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1.0);
2088
2089                         lpp_set_factor_fast(si->lpp, requirements, arg_op->attr.live_range.ilp, 1.0);
2090                         lpp_set_factor_fast(si->lpp, requirements, op->attr.live_range.args.reloads[i], 1.0);
2091                         foreach_pre_remat(si, irn, tmp) {
2092                                 op_t     *remat_op = get_irn_link(tmp);
2093                                 if(remat_op->attr.remat.remat->value == arg) {
2094                                         lpp_set_factor_fast(si->lpp, requirements, remat_op->attr.remat.ilp, 1.0);
2095                                 }
2096                         }
2097
2098 #ifdef WITH_MEMOPERANDS
2099                         n_memoperands = 0;
2100                         for(n = get_irn_arity(irn)-1; n>=0; --n) {
2101                                 if(get_irn_n(irn, n) == arg) {
2102                                         n_memoperands++;
2103                                 }
2104                         }
2105                         for(n = get_irn_arity(irn)-1; n>=0; --n) {
2106                                 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2107                                         memoperand_t  *memoperand;
2108                                         memoperand = set_find_memoperand(si->memoperands, irn, n);
2109
2110                                         /* memoperand <= mem_out */
2111                                         ir_snprintf(buf, sizeof(buf), "req_memoperand_%N_%d", irn, n);
2112                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2113                                         lpp_set_factor_fast(si->lpp, cst, memoperand->ilp, 1.0);
2114                                         lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2115
2116                                         /* the memoperand is only sufficient if it is used once by the op */
2117                                         if(n_memoperands == 1)
2118                                                 lpp_set_factor_fast(si->lpp, requirements, memoperand->ilp, 1.0);
2119
2120                                         lpp_set_factor_fast(si->lpp, one_memoperand, memoperand->ilp, 1.0);
2121
2122                                         /* we have one more free register if we use a memory operand */
2123                                         lpp_set_factor_fast(si->lpp, check_pre, memoperand->ilp, -1.0);
2124                                 }
2125                         }
2126 #endif
2127                 }
2128
2129                 /* iterate over L\U */
2130                 pset_foreach(live, tmp) {
2131                         if(!set_find_keyval(args, tmp)) {
2132                                 /* if a live value is not used by irn */
2133                                 tmp_op = get_irn_link(tmp);
2134                                 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
2135                         }
2136                 }
2137
2138
2139                 /* requirements for remats */
2140                 /* start new live ranges for values used by remats */
2141                 foreach_pre_remat(si, irn, tmp) {
2142                         op_t        *remat_op = get_irn_link(tmp);
2143                         int          n;
2144
2145                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2146                                 ir_node        *remat_arg = get_irn_n(tmp, n);
2147                                 op_t           *arg_op = get_irn_link(remat_arg);
2148                                 ilp_var_t       prev_lr;
2149
2150                                 if(!has_reg_class(si, remat_arg)) continue;
2151
2152                                 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2153                                 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2154                                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2155
2156                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2157                                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2158
2159                                 /* if remat arg is also used by current op then we can use reload placed for this argument */
2160                                 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2161                                         int    index = (int)keyval->val;
2162
2163                                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2164                                 }
2165                         }
2166                 }
2167
2168
2169
2170
2171                 /*************************
2172                  *  D O N E  W I T H  O P
2173                  *************************/
2174
2175                 DBG((si->dbg, LEVEL_4, "\t   %d values live at %+F\n", pset_count(live), irn));
2176
2177                 pset_foreach(live, tmp) {
2178                         assert(has_reg_class(si, tmp));
2179                 }
2180
2181                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2182                         ir_node        *arg = get_irn_n(irn, n);
2183
2184                         assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2185                 }
2186
2187                 del_pset(remat_defs);
2188                 del_pset(used);
2189                 del_set(args);
2190                 del_pset(defs);
2191                 defs = pset_new_ptr_default();
2192         }
2193
2194
2195
2196         /***************************************
2197          *   B E G I N N I N G  O F  B L O C K
2198          ***************************************/
2199
2200
2201         /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2202         DBG((si->dbg, LEVEL_3, "\t   %d values live at beginning of block %+F\n", pset_count(live), bb));
2203
2204         pset_foreach(live, irn) {
2205                 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2206         }
2207
2208         /* construct mem_outs for all values */
2209
2210         set_foreach(spill_bb->ilp, spill) {
2211                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2212                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2213
2214                 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2215                 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2216
2217                 if(pset_find_ptr(live, spill->irn)) {
2218                         DBG((si->dbg, LEVEL_5, "\t     %+F live at beginning of block %+F\n", spill->irn, bb));
2219
2220                         ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2221                         spill->mem_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2222                         lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2223
2224                         if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2225                                 int   n;
2226                                 op_t *op = get_irn_link(spill->irn);
2227
2228                                 /* do we have to copy a phi argument? */
2229                                 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2230                                 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2231
2232                                 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2233                                         const ir_node  *arg = get_irn_n(spill->irn, n);
2234                                         double          freq=0.0;
2235                                         int             m;
2236                                         ilp_var_t       var;
2237
2238
2239                                         /* argument already done? */
2240                                         if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2241
2242                                         /* get sum of execution frequencies of blocks with the same phi argument */
2243                                         for(m=n; m>=0; --m) {
2244                                                 const ir_node  *arg2 = get_irn_n(spill->irn, m);
2245
2246                                                 if(arg==arg2) {
2247                                                         freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2248                                                 }
2249                                         }
2250
2251                                         /* copies are not for free */
2252                                         ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2253                                         var = lpp_add_var(si->lpp, buf, lpp_binary, COST_STORE * freq);
2254
2255                                         for(m=n; m>=0; --m) {
2256                                                 const ir_node  *arg2 = get_irn_n(spill->irn, m);
2257
2258                                                 if(arg==arg2) {
2259                                                         op->attr.live_range.args.copies[m] = var;
2260                                                 }
2261                                         }
2262
2263                                         /* copy <= mem_in */
2264                                         ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2265                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2266                                         lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2267                                         lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2268                                 }
2269                         }
2270                 }
2271         }
2272
2273
2274         /* L\U is empty at bb start */
2275         /* arg is live throughout epilog if it is reg_in into this block */
2276
2277         /* check the register pressure at the beginning of the block
2278          * including remats
2279          */
2280         ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2281         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
2282
2283         pset_foreach(live, irn) {
2284         ilp_cst_t  nospill;
2285
2286                 spill = set_find_spill(spill_bb->ilp, irn);
2287                 assert(spill);
2288
2289                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2290                 spill->reg_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2291
2292                 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2293
2294                 /* spill + mem_in <= 1 */
2295                 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2296                 nospill = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2297
2298                 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2299                 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2300
2301         }
2302         foreach_post_remat(bb, irn) {
2303                 op_t     *remat_op = get_irn_link(irn);
2304
2305                 DBG((si->dbg, LEVEL_4, "\t  next post remat: %+F\n", irn));
2306                 assert(remat_op->is_remat && !remat_op->attr.remat.pre);
2307
2308                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2309         }
2310
2311         /* forall post remats add requirements */
2312         foreach_post_remat(bb, tmp) {
2313                 int         n;
2314
2315                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2316                         ir_node    *remat_arg = get_irn_n(tmp, n);
2317                         op_t       *remat_op = get_irn_link(tmp);
2318
2319                         if(!has_reg_class(si, remat_arg)) continue;
2320
2321                         spill = set_find_spill(spill_bb->ilp, remat_arg);
2322                         assert(spill);
2323
2324                         /* remat <= reg_in_argument */
2325                         ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2326                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2327                         lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2328                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2329                 }
2330         }
2331
2332         /* mem_in/reg_in for live_in values, especially phis and their arguments */
2333         pset_foreach(live, irn) {
2334                 int          p = 0,
2335                                          n;
2336
2337                 spill = set_find_spill(spill_bb->ilp, irn);
2338                 assert(spill && spill->irn == irn);
2339
2340                 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2341                         for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2342                                 ilp_cst_t       mem_in,
2343                                                                 reg_in;
2344                                 ir_node        *phi_arg = get_Phi_pred(irn, n);
2345                                 ir_node        *bb_p = get_Block_cfgpred_block(bb, n);
2346                                 spill_bb_t     *spill_bb_p = get_irn_link(bb_p);
2347                                 spill_t        *spill_p;
2348                                 op_t           *op = get_irn_link(irn);
2349
2350                                 /* although the phi is in the right regclass one or more of
2351                                  * its arguments can be in a different one or at least to
2352                                  * ignore
2353                                  */
2354                                 if(has_reg_class(si, phi_arg)) {
2355                                         /* mem_in < mem_out_arg + copy */
2356                                         ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2357                                         mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2358
2359                                         /* reg_in < reg_out_arg */
2360                                         ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2361                                         reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2362
2363                                         lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2364                                         lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2365
2366                                         spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2367                                         assert(spill_p);
2368
2369                                         lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2370                                         lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0);
2371                                         lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2372                                 }
2373                         }
2374                 } else {
2375                         /* else assure the value arrives on all paths in the same resource */
2376
2377                         for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2378                                 ilp_cst_t       mem_in,
2379                                                                 reg_in;
2380                                 ir_node        *bb_p = get_Block_cfgpred_block(bb, n);
2381                                 spill_bb_t     *spill_bb_p = get_irn_link(bb_p);
2382                                 spill_t        *spill_p;
2383
2384                                 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2385                                 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2386                                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2387                                 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2388
2389                                 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2390                                 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2391
2392                                 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2393                                 assert(spill_p);
2394
2395                                 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2396                                 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2397                         }
2398                 }
2399         }
2400
2401         /* first live ranges from reg_ins */
2402         pset_foreach(live, irn) {
2403                 op_t      *op = get_irn_link(irn);
2404
2405                 spill = set_find_spill(spill_bb->ilp, irn);
2406                 assert(spill && spill->irn == irn);
2407
2408                 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2409                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2410                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2411                 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2412
2413                 foreach_post_remat(bb, tmp) {
2414                         op_t     *remat_op = get_irn_link(tmp);
2415
2416                         if(remat_op->attr.remat.remat->value == irn) {
2417                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2418                         }
2419                 }
2420         }
2421
2422         /* walk forward now and compute constraints for placing spills */
2423         /* this must only be done for values that are not defined in this block */
2424         /* TODO are these values at start of block? if yes, just check whether this is a diverge edge and skip the loop */
2425         pset_foreach(live, irn) {
2426                 /*
2427                  * if value is defined in this block we can anways place the spill directly after the def
2428                  *    -> no constraint necessary
2429                  */
2430                 if(!is_Phi(irn) && get_nodes_block(irn) == bb) continue;
2431
2432
2433                 spill = set_find_spill(spill_bb->ilp, irn);
2434                 assert(spill);
2435
2436                 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2437                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2438
2439                 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2440                 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2441
2442                 if(!is_Phi(irn)) {
2443                         sched_foreach_op(bb, tmp) {
2444                                 op_t   *op = get_irn_link(tmp);
2445
2446                                 if(is_Phi(tmp)) continue;
2447                                 assert(!is_Proj(tmp));
2448
2449                                 if(op->is_remat) {
2450                                         ir_node   *value = op->attr.remat.remat->value;
2451
2452                                         if(value == irn) {
2453                                                 /* only collect remats up to the first real use of a value */
2454                                                 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2455                                         }
2456                                 } else {
2457                                         int   n;
2458
2459                                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2460                                                 ir_node    *arg = get_irn_n(tmp, n);
2461
2462                                                 if(arg == irn) {
2463                                                         /* if a value is used stop collecting remats */
2464                             goto next_live;
2465                                                 }
2466                                         }
2467                                 }
2468                         }
2469                 }
2470 next_live:
2471         }
2472
2473         del_pset(live);
2474 }
2475
2476 typedef struct _irnlist_t {
2477         struct list_head   list;
2478         ir_node           *irn;
2479 } irnlist_t;
2480
2481 typedef struct _interference_t {
2482         struct list_head    blocklist;
2483         ir_node            *a;
2484         ir_node            *b;
2485 } interference_t;
2486
2487 static int
2488 cmp_interference(const void *a, const void *b, size_t size)
2489 {
2490         const interference_t *p = a;
2491         const interference_t *q = b;
2492
2493         return !(p->a == q->a && p->b == q->b);
2494 }
2495
2496 static interference_t *
2497 set_find_interference(set * set, ir_node * a, ir_node * b)
2498 {
2499         interference_t     query;
2500
2501         query.a = (a>b)?a:b;
2502         query.b = (a>b)?b:a;
2503
2504         return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2505 }
2506
2507 static interference_t *
2508 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2509 {
2510         interference_t     query,
2511                                           *result;
2512         irnlist_t         *list = obstack_alloc(si->obst, sizeof(*list));
2513
2514         list->irn = bb;
2515
2516         result = set_find_interference(set, a, b);
2517         if(result) {
2518
2519                 list_add(&list->list, &result->blocklist);
2520                 return result;
2521         }
2522
2523         query.a = (a>b)?a:b;
2524         query.b = (a>b)?b:a;
2525
2526         result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2527
2528         INIT_LIST_HEAD(&result->blocklist);
2529         list_add(&list->list, &result->blocklist);
2530
2531         return result;
2532 }
2533
2534 static int
2535 values_interfere_in_block(const spill_ilp_t * si, ir_node * bb, ir_node * a, ir_node * b)
2536 {
2537         const ir_edge_t *edge;
2538
2539         if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2540                 /* both values are live in, so they interfere */
2541                 return 1;
2542         }
2543
2544         /* ensure a dominates b */
2545         if(value_dominates(b,a)) {
2546                 const ir_node * t;
2547                 t = b;
2548                 b = a;
2549                 a = t;
2550         }
2551         assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2552
2553
2554         /* the following code is stolen from bera.c */
2555         if(be_is_live_end(si->lv, bb, a))
2556                 return 1;
2557
2558         foreach_out_edge(a, edge) {
2559                 const ir_node *user = edge->src;
2560                 if(get_nodes_block(user) == bb
2561                                 && !is_Phi(user)
2562                                 && b != user
2563                                 && value_dominates(b, user))
2564                         return 1;
2565         }
2566
2567         return 0;
2568 }
2569
2570 /**
2571  * Walk all irg blocks and collect interfering values inside of phi classes
2572  */
2573 static void
2574 luke_interferencewalker(ir_node * bb, void * data)
2575 {
2576         spill_ilp_t    *si = (spill_ilp_t*)data;
2577         int             l1, l2;
2578
2579         be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
2580                 ir_node        *a = be_lv_get_irn(si->lv, bb, l1);
2581                 op_t           *a_op = get_irn_link(a);
2582
2583                 if(a_op->is_remat) continue;
2584
2585                 /* a is only interesting if it is in my register class and if it is inside a phi class */
2586                 if (has_reg_class(si, a) && get_phi_class(a)) {
2587                         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)) {
2588                                 ir_node        *b = be_lv_get_irn(si->lv, bb, l2);
2589                                 op_t           *b_op = get_irn_link(b);
2590
2591                                 if(b_op->is_remat) continue;
2592
2593                                 /* a and b are only interesting if they are in the same phi class */
2594                                 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
2595                                         if(values_interfere_in_block(si, bb, a, b)) {
2596                                                 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2597                                                 set_insert_interference(si, si->interferences, a, b, bb);
2598                                         }
2599                                 }
2600                         }
2601                 }
2602         }
2603 }
2604
2605 static unsigned int copy_path_id = 0;
2606
2607 static void
2608 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2609 {
2610         ilp_cst_t  cst;
2611         ilp_var_t  copy;
2612         char       buf[256];
2613         void      *ptr;
2614
2615         ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
2616         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2617
2618         lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
2619
2620         pset_foreach(copies, ptr) {
2621                 copy = PTR_TO_INT(ptr);
2622                 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
2623         }
2624 }
2625
2626 /**
2627  * @parameter copies   contains a path of copies which lead us to irn
2628  * @parameter visited  contains a set of nodes already visited on this path
2629  */
2630 static int
2631 find_copy_path(spill_ilp_t * si, ir_node * irn, ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
2632 {
2633         ir_edge_t *edge;
2634         op_t      *op = get_irn_link(irn);
2635     pset      *visited_users = pset_new_ptr_default();
2636         int        paths = 0;
2637
2638         if(op->is_remat) return;
2639
2640         pset_insert_ptr(visited, irn);
2641
2642         if(is_Phi(irn)) {
2643                 int    n;
2644         pset  *visited_operands = pset_new_ptr(get_irn_arity(irn));
2645
2646                 /* visit all operands */
2647                 for(n=get_irn_arity(irn)-1; n>=0; --n) {
2648                         ir_node  *arg = get_irn_n(irn, n);
2649                         ilp_var_t  copy = op->attr.live_range.args.copies[n];
2650
2651                         if(!has_reg_class(si, arg)) continue;
2652             if(pset_find_ptr(visited_operands, arg)) continue;
2653             pset_insert_ptr(visited_operands, arg);
2654
2655                         if(arg == target) {
2656                                 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2657                                         del_pset(visited_operands);
2658                                         del_pset(visited_users);
2659                                         pset_remove_ptr(visited, irn);
2660                                         return paths;
2661                                 }
2662                                 pset_insert(copies, INT_TO_PTR(copy), copy);
2663                                 write_copy_path_cst(si, copies, any_interfere);
2664                                 pset_remove(copies, INT_TO_PTR(copy), copy);
2665                         } else if(!pset_find_ptr(visited, arg)) {
2666                                 pset_insert(copies, INT_TO_PTR(copy), copy);
2667                                 paths += find_copy_path(si, arg, target, any_interfere, copies, visited);
2668                                 pset_remove(copies, INT_TO_PTR(copy), copy);
2669
2670                 /*if(paths > MAX_PATHS) {
2671                     if(pset_count(copies) == 0) {
2672                         ilp_cst_t  cst;
2673                         char       buf[256];
2674
2675                         ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
2676                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
2677                         lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2678                         lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
2679                         DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
2680
2681                         paths = 0;
2682                     } else {
2683                         del_pset(visited_operands);
2684                         del_pset(visited_users);
2685                         pset_remove_ptr(visited, irn);
2686                         return paths;
2687                     }
2688                 } else if(pset_count(copies) == 0) {
2689                                         paths = 0;
2690                                 }*/
2691                         }
2692                 }
2693
2694         del_pset(visited_operands);
2695         }
2696
2697         /* visit all uses which are phis */
2698         foreach_out_edge(irn, edge) {
2699                 ir_node  *user = edge->src;
2700                 int       pos  = edge->pos;
2701                 op_t     *op = get_irn_link(user);
2702                 ilp_var_t copy;
2703
2704                 if(!is_Phi(user)) continue;
2705                 if(!has_reg_class(si, user)) continue;
2706         if(pset_find_ptr(visited_users, user)) continue;
2707         pset_insert_ptr(visited_users, user);
2708
2709                 copy = op->attr.live_range.args.copies[pos];
2710
2711                 if(user == target) {
2712                         if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2713                                 del_pset(visited_users);
2714                                 pset_remove_ptr(visited, irn);
2715                                 return paths;
2716                         }
2717                         pset_insert(copies, INT_TO_PTR(copy), copy);
2718                         write_copy_path_cst(si, copies, any_interfere);
2719                         pset_remove(copies, INT_TO_PTR(copy), copy);
2720                 } else if(!pset_find_ptr(visited, user)) {
2721                         pset_insert(copies, INT_TO_PTR(copy), copy);
2722                         paths += find_copy_path(si, user, target, any_interfere, copies, visited);
2723                         pset_remove(copies, INT_TO_PTR(copy), copy);
2724
2725             /*if(paths > MAX_PATHS) {
2726                 if(pset_count(copies) == 0) {
2727                     ilp_cst_t  cst;
2728                     char       buf[256];
2729
2730                     ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
2731                     cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
2732                     lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2733                     lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
2734                     DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
2735
2736                     paths = 0;
2737                 } else {
2738                     del_pset(visited_users);
2739                     pset_remove_ptr(visited, irn);
2740                     return paths;
2741                 }
2742             } else if(pset_count(copies) == 0) {
2743                                 paths = 0;
2744                         }*/
2745                 }
2746         }
2747
2748     del_pset(visited_users);
2749         pset_remove_ptr(visited, irn);
2750         return paths;
2751 }
2752
2753 static void
2754 gen_copy_constraints(spill_ilp_t * si, ir_node * a, ir_node * b, ilp_var_t any_interfere)
2755 {
2756         pset * copies = pset_new_ptr_default();
2757         pset * visited = pset_new_ptr_default();
2758
2759         find_copy_path(si, a, b, any_interfere, copies, visited);
2760
2761         del_pset(visited);
2762         del_pset(copies);
2763 }
2764
2765
2766 static void
2767 memcopyhandler(spill_ilp_t * si)
2768 {
2769         interference_t   *interference;
2770         char              buf[256];
2771         /* teste Speicherwerte auf Interferenz */
2772
2773         /* analyze phi classes */
2774         phi_class_compute(si->chordal_env->irg);
2775
2776         DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
2777         irg_block_walk_graph(si->chordal_env->irg, luke_interferencewalker, NULL, si);
2778
2779         /* now lets emit the ILP unequations for the crap */
2780         set_foreach(si->interferences, interference) {
2781                 irnlist_t      *irnlist;
2782                 ilp_var_t       interfere,
2783                                                 any_interfere;
2784                 ilp_cst_t       any_interfere_cst,
2785                                                 cst;
2786                 const ir_node  *a  = interference->a;
2787                 const ir_node  *b  = interference->b;
2788
2789                 /* any_interf <= \sum interf */
2790                 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
2791                 any_interfere_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2792                 any_interfere = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2793
2794                 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
2795
2796                 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
2797                         const ir_node  *bb = irnlist->irn;
2798                         spill_bb_t     *spill_bb = get_irn_link(bb);
2799                         spill_t        *spilla,
2800                                                    *spillb,
2801                                                    query;
2802                         char           buf[256];
2803
2804                         query.irn = a;
2805                         spilla = set_find_spill(spill_bb->ilp, a);
2806                         assert(spilla);
2807
2808                         query.irn = b;
2809                         spillb = set_find_spill(spill_bb->ilp, b);
2810                         assert(spillb);
2811
2812                         /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
2813                         /* 1:   mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
2814                         /* 2: - mem_in_a - spill_a + interfere <= 0 */
2815                         /* 3: - mem_in_b - spill_b + interfere <= 0 */
2816                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
2817                         interfere = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2818
2819                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
2820                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2821
2822                         lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
2823                         if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
2824                         lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
2825                         if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
2826                         lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
2827
2828                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
2829                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2830
2831                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2832                         if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
2833                         lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
2834
2835                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
2836                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2837
2838                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2839                         if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
2840                         lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
2841
2842
2843                         lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
2844
2845                         /* any_interfere >= interf */
2846                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
2847                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2848
2849                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2850                         lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2851                 }
2852
2853                 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
2854                 gen_copy_constraints(si,a,b,any_interfere);
2855         }
2856 }
2857
2858
2859 static INLINE int
2860 is_zero(double x)
2861 {
2862         return fabs(x) < 0.00001;
2863 }
2864
2865 #ifdef KEEPALIVE
2866 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
2867 {
2868         spill_ilp_t *si = get_irg_link(current_ir_graph);
2869
2870         if(pset_find_ptr(si->all_possible_remats, n)) {
2871                 op_t   *op = (op_t*)get_irn_link(n);
2872                 assert(op && op->is_remat);
2873
2874                 if(!op->attr.remat.remat->inverse) {
2875                         if(op->attr.remat.pre) {
2876                                 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
2877                         } else {
2878                                 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
2879                         }
2880
2881                         return 1;
2882                 } else {
2883                         op_t   *op = (op_t*)get_irn_link(n);
2884                         assert(op && op->is_remat);
2885
2886                         if(op->attr.remat.pre) {
2887                                 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
2888                         } else {
2889                                 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
2890                         }
2891
2892                         return 1;
2893                 }
2894         }
2895
2896         return 0;
2897 }
2898
2899 static void
2900 dump_graph_with_remats(ir_graph * irg, const char * suffix)
2901 {
2902         set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
2903         be_dump(irg, suffix, dump_ir_block_graph_sched);
2904         set_dump_node_vcgattr_hook(NULL);
2905 }
2906 #endif
2907
2908 /**
2909  * Edge hook to dump the schedule edges with annotated register pressure.
2910  */
2911 static int
2912 sched_pressure_edge_hook(FILE *F, ir_node *irn)
2913 {
2914         if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
2915                 ir_node *prev = sched_prev(irn);
2916                 fprintf(F, "edge:{sourcename:\"");
2917                 PRINT_NODEID(irn);
2918                 fprintf(F, "\" targetname:\"");
2919                 PRINT_NODEID(prev);
2920                 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
2921                 fprintf(F, "\" color:magenta}\n");
2922         }
2923         return 1;
2924 }
2925
2926 static void
2927 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
2928 {
2929         DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
2930
2931         dump_consts_local(0);
2932         set_dump_node_edge_hook(sched_pressure_edge_hook);
2933         dump_ir_block_graph(irg, suffix);
2934         set_dump_node_edge_hook(old_edge_hook);
2935 }
2936
2937 static void
2938 walker_pressure_annotator(ir_node * bb, void * data)
2939 {
2940         spill_ilp_t  *si = data;
2941         ir_node      *irn;
2942         int           n, i;
2943         pset         *live = pset_new_ptr_default();
2944         int           projs = 0;
2945
2946         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
2947                 irn = be_lv_get_irn(si->lv, bb, i);
2948
2949                 if (has_reg_class(si, irn)) {
2950                         pset_insert_ptr(live, irn);
2951                 }
2952         }
2953
2954         set_irn_link(bb, INT_TO_PTR(pset_count(live)));
2955
2956         sched_foreach_reverse(bb, irn) {
2957                 if(is_Phi(irn)) {
2958                         set_irn_link(irn, INT_TO_PTR(pset_count(live)));
2959                         continue;
2960                 }
2961
2962                 if(has_reg_class(si, irn)) {
2963                         pset_remove_ptr(live, irn);
2964                         if(is_Proj(irn)) ++projs;
2965                 }
2966
2967                 if(!is_Proj(irn)) projs = 0;
2968
2969                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2970                         ir_node    *arg = get_irn_n(irn, n);
2971
2972                         if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
2973                 }
2974                 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
2975         }
2976
2977         del_pset(live);
2978 }
2979
2980 static void
2981 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
2982 {
2983         be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
2984 }
2985
2986 #ifdef KEEPALIVE
2987 static void
2988 connect_all_remats_with_keep(spill_ilp_t * si)
2989 {
2990         ir_node   *irn;
2991         ir_node  **ins,
2992                          **pos;
2993         int        n_remats;
2994
2995
2996         n_remats = pset_count(si->all_possible_remats);
2997         if(n_remats) {
2998                 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
2999
3000                 pos = ins;
3001                 pset_foreach(si->all_possible_remats, irn) {
3002                         *pos = irn;
3003                         ++pos;
3004                 }
3005
3006                 si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins);
3007
3008                 obstack_free(si->obst, ins);
3009         }
3010 }
3011 #endif
3012
3013 static void
3014 connect_all_spills_with_keep(spill_ilp_t * si)
3015 {
3016         ir_node   *irn;
3017         ir_node  **ins,
3018                          **pos;
3019         int        n_spills;
3020         ir_node   *keep;
3021
3022
3023         n_spills = pset_count(si->spills);
3024         if(n_spills) {
3025                 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
3026
3027                 pos = ins;
3028                 pset_foreach(si->spills, irn) {
3029                         *pos = irn;
3030                         ++pos;
3031                 }
3032
3033                 keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_spills, ins);
3034
3035                 obstack_free(si->obst, ins);
3036         }
3037 }
3038
3039 /** insert a spill at an arbitrary position */
3040 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert, ir_node *ctx)
3041 {
3042         ir_node *bl     = is_Block(insert)?insert:get_nodes_block(insert);
3043         ir_graph *irg   = get_irn_irg(bl);
3044         ir_node *frame  = get_irg_frame(irg);
3045         ir_node *spill;
3046         ir_node *next;
3047
3048         const arch_register_class_t *cls       = arch_get_irn_reg_class(arch_env, irn, -1);
3049         const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
3050
3051         spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn, ctx);
3052
3053         /*
3054          * search the right insertion point. a spill of a phi cannot be put
3055          * directly after the phi, if there are some phis behind the one which
3056          * is spilled. Also, a spill of a Proj must be after all Projs of the
3057          * same tuple node.
3058          *
3059          * Here's one special case:
3060          * If the spill is in the start block, the spill must be after the frame
3061          * pointer is set up. This is done by setting insert to the end of the block
3062          * which is its default initialization (see above).
3063          */
3064
3065         if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
3066                 insert = frame;
3067
3068         for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
3069                 insert = next;
3070
3071         sched_add_after(insert, spill);
3072         return spill;
3073 }
3074
3075 static void
3076 delete_remat(spill_ilp_t * si, ir_node * remat) {
3077         int       n;
3078         ir_node  *bad = get_irg_bad(si->chordal_env->irg);
3079
3080         sched_remove(remat);
3081
3082         /* kill links to operands */
3083         for (n=get_irn_arity(remat)-1; n>=-1; --n) {
3084                 set_irn_n(remat, n, bad);
3085         }
3086 }
3087
3088 static void
3089 clean_remat_info(spill_ilp_t * si)
3090 {
3091         int            n;
3092         remat_t       *remat;
3093         remat_info_t  *remat_info;
3094         ir_node       *bad = get_irg_bad(si->chordal_env->irg);
3095
3096         set_foreach(si->remat_info, remat_info) {
3097                 if(!remat_info->remats) continue;
3098
3099                 pset_foreach(remat_info->remats, remat)
3100                 {
3101                         if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
3102                                 set_irn_n(remat->proj, -1, bad);
3103                                 set_irn_n(remat->proj, 0, bad);
3104                         }
3105
3106                         if(get_irn_n_edges(remat->op) == 0) {
3107                                 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
3108                                         set_irn_n(remat->op, n, bad);
3109                                 }
3110                         }
3111                 }
3112
3113                 if(remat_info->remats) del_pset(remat_info->remats);
3114                 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
3115         }
3116 }
3117
3118 static void
3119 delete_unnecessary_remats(spill_ilp_t * si)
3120 {
3121 #ifdef KEEPALIVE
3122         int       n;
3123         ir_node  *bad = get_irg_bad(si->chordal_env->irg);
3124
3125         if(si->keep) {
3126                 ir_node   *end = get_irg_end(si->chordal_env->irg);
3127                 ir_node  **keeps;
3128
3129                 for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
3130                         ir_node        *keep_arg = get_irn_n(si->keep, n);
3131                         op_t           *arg_op = get_irn_link(keep_arg);
3132                         lpp_name_t     *name;
3133
3134                         assert(arg_op->is_remat);
3135
3136                         name = si->lpp->vars[arg_op->attr.remat.ilp];
3137
3138                         if(is_zero(name->value)) {
3139                                 DBG((si->dbg, LEVEL_3, "\t  deleting remat %+F\n", keep_arg));
3140                                 /* TODO check whether reload is preferred over remat (could be bug) */
3141                                 delete_remat(si, keep_arg);
3142                         } else {
3143                                 if(!arg_op->attr.remat.remat->inverse) {
3144                                         if(arg_op->attr.remat.pre) {
3145                                                 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
3146                                         } else {
3147                                                 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
3148                                         }
3149                                 } else {
3150                                         if(arg_op->attr.remat.pre) {
3151                                                 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
3152                                         } else {
3153                                                 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
3154                                         }
3155                                 }
3156                         }
3157
3158                         set_irn_n(si->keep, n, bad);
3159                 }
3160 #if 0
3161                 for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
3162                         ir_node        *end_arg = get_End_keepalive(end, i);
3163
3164                         if(end_arg != si->keep) {
3165                                 obstack_grow(si->obst, &end_arg, sizeof(end_arg));
3166                         }
3167                 }
3168                 keeps = obstack_finish(si->obst);
3169                 set_End_keepalives(end, n-1, keeps);
3170                 obstack_free(si->obst, keeps);
3171 #endif
3172         } else {
3173                 DBG((si->dbg, LEVEL_2, "\t  no remats to delete (none have been inserted)\n"));
3174         }
3175 #else
3176         ir_node  *remat;
3177
3178         pset_foreach(si->all_possible_remats, remat) {
3179                 op_t           *remat_op = get_irn_link(remat);
3180                 lpp_name_t     *name = si->lpp->vars[remat_op->attr.remat.ilp];
3181
3182                 if(is_zero(name->value)) {
3183                         DBG((si->dbg, LEVEL_3, "\t  deleting remat %+F\n", remat));
3184                         /* TODO check whether reload is preferred over remat (could be bug) */
3185                         delete_remat(si, remat);
3186                 } else {
3187                         if(!remat_op->attr.remat.remat->inverse) {
3188                                 if(remat_op->attr.remat.pre) {
3189                                         DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3190                                 } else {
3191                                         DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3192                                 }
3193                         } else {
3194                                 if(remat_op->attr.remat.pre) {
3195                                         DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3196                                 } else {
3197                                         DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3198                                 }
3199                         }
3200                 }
3201         }
3202 #endif
3203 }
3204
3205 static pset *
3206 get_spills_for_value(spill_ilp_t * si, ir_node * value)
3207 {
3208         pset     *spills = pset_new_ptr_default();
3209
3210         ir_node  *next;
3211         defs_t   *defs;
3212
3213         defs = set_find_def(si->values, value);
3214
3215         if(defs && defs->spills) {
3216                 for(next = defs->spills; next; next = get_irn_link(next)) {
3217                         pset_insert_ptr(spills, next);
3218                 }
3219         }
3220
3221         return spills;
3222 }
3223
3224 static pset *
3225 get_remats_for_value(spill_ilp_t * si, ir_node * value)
3226 {
3227         pset     *remats = pset_new_ptr_default();
3228
3229         ir_node  *next;
3230         defs_t   *defs;
3231
3232         pset_insert_ptr(remats, value);
3233         defs = set_find_def(si->values, value);
3234
3235         if(defs && defs->remats) {
3236                 for(next = defs->remats; next; next = get_irn_link(next)) {
3237                         pset_insert_ptr(remats, next);
3238                 }
3239         }
3240
3241         return remats;
3242 }
3243
3244
3245 /**
3246  * @param before   The node after which the spill will be placed in the schedule
3247  */
3248 /* TODO set context properly */
3249 static ir_node *
3250 insert_spill(spill_ilp_t * si, ir_node * irn, ir_node * value, ir_node * before)
3251 {
3252         defs_t   *defs;
3253         ir_node  *spill;
3254         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3255
3256         DBG((si->dbg, LEVEL_3, "\t  inserting spill for value %+F after %+F\n", irn, before));
3257
3258         spill = be_spill2(arch_env, irn, before, irn);
3259
3260         defs = set_insert_def(si->values, value);
3261         assert(defs);
3262
3263         /* enter into the linked list */
3264         set_irn_link(spill, defs->spills);
3265         defs->spills = spill;
3266
3267 #ifdef KEEPALIVE_SPILLS
3268         pset_insert_ptr(si->spills, spill);
3269 #endif
3270
3271         return spill;
3272 }
3273
3274 /**
3275  * @param before   The Phi node which has to be spilled
3276  */
3277 static ir_node *
3278 insert_mem_phi(spill_ilp_t * si, const ir_node * phi)
3279 {
3280         ir_node   *mem_phi;
3281         ir_node  **ins;
3282         defs_t    *defs;
3283         int        n;
3284         op_t      *op = get_irn_link(phi);
3285
3286         NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3287
3288         for(n=get_irn_arity(phi)-1; n>=0; --n) {
3289                 ins[n] = si->m_unknown;
3290         }
3291
3292         mem_phi =  new_r_Phi(si->chordal_env->irg, get_nodes_block(phi), get_irn_arity(phi), ins, mode_M);
3293
3294         defs = set_insert_def(si->values, phi);
3295         assert(defs);
3296
3297         /* enter into the linked list */
3298         set_irn_link(mem_phi, defs->spills);
3299         defs->spills = mem_phi;
3300
3301         sched_add_after(phi, mem_phi);
3302
3303 #ifdef KEEPALIVE_SPILLS
3304         pset_insert_ptr(si->spills, mem_phi);
3305 #endif
3306
3307
3308         return mem_phi;
3309 }
3310
3311 /**
3312  * Add remat to list of defs, destroys link field!
3313  */
3314 static void
3315 insert_remat(spill_ilp_t * si, ir_node * remat)
3316 {
3317         defs_t   *defs;
3318         op_t     *remat_op = get_irn_link(remat);
3319
3320         assert(remat_op->is_remat);
3321
3322         defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3323         assert(defs);
3324
3325         /* enter into the linked list */
3326         set_irn_link(remat, defs->remats);
3327         defs->remats = remat;
3328 }
3329
3330
3331 /**
3332  * Add reload before operation and add to list of defs
3333  */
3334 static ir_node *
3335 insert_reload(spill_ilp_t * si, const ir_node * value, const ir_node * after)
3336 {
3337         defs_t   *defs;
3338         ir_node  *reload,
3339                          *spill;
3340         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3341
3342         DBG((si->dbg, LEVEL_3, "\t  inserting reload for value %+F before %+F\n", value, after));
3343
3344         defs = set_find_def(si->values, value);
3345
3346         spill = defs->spills;
3347         assert(spill && "no spill placed before reload");
3348
3349         reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3350
3351         /* enter into the linked list */
3352         set_irn_link(reload, defs->remats);
3353         defs->remats = reload;
3354
3355         return reload;
3356 }
3357
3358 #ifdef WITH_MEMOPERANDS
3359 void perform_memory_operand(spill_ilp_t * si, memoperand_t * memoperand)
3360 {
3361         defs_t           *defs;
3362         ir_node          *reload;
3363         ir_node          *value = get_irn_n(memoperand->irn, memoperand->pos);
3364         ir_node          *spill;
3365         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3366
3367         DBG((si->dbg, LEVEL_2, "\t  inserting memory operand for value %+F at %+F\n", value, memoperand->irn));
3368
3369         defs = set_find_def(si->values, value);
3370
3371         spill = defs->spills;
3372         assert(spill && "no spill placed before reload");
3373
3374         reload = be_reload(arch_env, si->cls, memoperand->irn, get_irn_mode(value), spill);
3375
3376         arch_perform_memory_operand(arch_env, memoperand->irn, reload, memoperand->pos);
3377         sched_remove(reload);
3378 }
3379
3380 void insert_memoperands(spill_ilp_t * si)
3381 {
3382         memoperand_t   *memoperand;
3383         lpp_name_t     *name;
3384
3385         set_foreach(si->memoperands, memoperand) {
3386                 name = si->lpp->vars[memoperand->ilp];
3387                 if(!is_zero(name->value)) {
3388                         perform_memory_operand(si, memoperand);
3389                 }
3390         }
3391 }
3392 #endif
3393
3394 static void
3395 walker_spill_placer(ir_node * bb, void * data) {
3396         spill_ilp_t   *si = (spill_ilp_t*)data;
3397         ir_node       *irn;
3398         spill_bb_t    *spill_bb = get_irn_link(bb);
3399         pset          *spills_to_do = pset_new_ptr_default();
3400         spill_t       *spill;
3401
3402         set_foreach(spill_bb->ilp, spill) {
3403                 lpp_name_t    *name;
3404
3405                 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3406                         name = si->lpp->vars[spill->mem_in];
3407                         if(!is_zero(name->value)) {
3408                                 ir_node   *mem_phi;
3409
3410                                 mem_phi = insert_mem_phi(si, spill->irn);
3411
3412                                 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3413                         }
3414                 }
3415
3416                 name = si->lpp->vars[spill->spill];
3417                 if(!is_zero(name->value)) {
3418                         /* place spill directly after definition */
3419                         if(get_nodes_block(spill->irn) == bb) {
3420                                 insert_spill(si, spill->irn, spill->irn, spill->irn);
3421                                 continue;
3422                         }
3423
3424                         /* place spill at bb start */
3425                         if(spill->reg_in > 0) {
3426                                 name = si->lpp->vars[spill->reg_in];
3427                                 if(!is_zero(name->value)) {
3428                                         insert_spill(si, spill->irn, spill->irn, bb);
3429                                         continue;
3430                                 }
3431                         }
3432                         /* place spill after a remat */
3433                         pset_insert_ptr(spills_to_do, spill->irn);
3434                 }
3435         }
3436         DBG((si->dbg, LEVEL_3, "\t  %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3437
3438
3439         for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3440                 op_t     *op = get_irn_link(irn);
3441
3442                 if(be_is_Spill(irn)) continue;
3443
3444                 if(op->is_remat) {
3445                         /* TODO fix this if we want to support remats with more than two nodes */
3446                         if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3447                                 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3448
3449                                 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3450                         }
3451                 } else {
3452                         if(pset_find_ptr(spills_to_do, irn)) {
3453                                 pset_remove_ptr(spills_to_do, irn);
3454
3455                                 insert_spill(si, irn, irn, irn);
3456                         }
3457                 }
3458
3459         }
3460
3461         assert(pset_count(spills_to_do) == 0);
3462
3463         /* afterwards free data in block */
3464         del_pset(spills_to_do);
3465 }
3466
3467 static ir_node *
3468 insert_mem_copy(spill_ilp_t * si, const ir_node * bb, const ir_node * value)
3469 {
3470         ir_node          *insert_pos = bb;
3471         ir_node          *spill;
3472         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3473
3474         /* find last definition of arg value in block */
3475         ir_node  *next;
3476         defs_t   *defs;
3477         int       last = 0;
3478
3479         defs = set_find_def(si->values, value);
3480
3481         if(defs && defs->remats) {
3482                 for(next = defs->remats; next; next = get_irn_link(next)) {
3483                         if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3484                                 last = sched_get_time_step(next);
3485                                 insert_pos = next;
3486                         }
3487                 }
3488         }
3489
3490         if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3491                 last = sched_get_time_step(value);
3492                 insert_pos = value;
3493         }
3494
3495         DBG((si->dbg, LEVEL_2, "\t  inserting mem copy for value %+F after %+F\n", value, insert_pos));
3496
3497         spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos, value);
3498
3499         return spill;
3500 }
3501
3502 static void
3503 phim_fixer(spill_ilp_t *si) {
3504         defs_t  *defs;
3505
3506         set_foreach(si->values, defs) {
3507                 const ir_node  *phi = defs->value;
3508                 op_t           *op = get_irn_link(phi);
3509                 ir_node        *phi_m = NULL;
3510                 ir_node        *next = defs->spills;
3511                 int             n;
3512
3513                 if(!is_Phi(phi)) continue;
3514
3515                 while(next) {
3516                         if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3517                                 phi_m = next;
3518                                 break;
3519                         } else {
3520                                 next = get_irn_link(next);
3521                         }
3522                 }
3523                 if(!phi_m) continue;
3524
3525                 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3526                         const ir_node  *value = get_irn_n(phi, n);
3527                         defs_t         *val_defs = set_find_def(si->values, value);
3528                         ir_node        *arg = get_irn_n(phi_m, n);
3529
3530                         /* a spill of this value */
3531                         ir_node      *spill;
3532
3533
3534 #ifndef NO_MEMCOPIES
3535                         ir_node    *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3536                         lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3537
3538                         if(!is_zero(name->value)) {
3539                                 spill = insert_mem_copy(si, pred, value);
3540                         } else {
3541                                 spill = val_defs->spills;
3542                         }
3543 #else
3544                         spill = val_defs->spills;
3545 #endif
3546                         assert(spill && "no spill placed before PhiM");
3547                         set_irn_n(phi_m, n, spill);
3548                 }
3549         }
3550 }
3551
3552 static void
3553 walker_reload_placer(ir_node * bb, void * data) {
3554         spill_ilp_t   *si = (spill_ilp_t*)data;
3555         ir_node       *irn;
3556         spill_bb_t    *spill_bb = get_irn_link(bb);
3557         int            i;
3558
3559         /* reloads at end of block */
3560         if(spill_bb->reloads) {
3561                 keyval_t    *keyval;
3562
3563                 set_foreach(spill_bb->reloads, keyval) {
3564                         ir_node        *irn = (ir_node*)keyval->key;
3565                         ilp_var_t       reload = PTR_TO_INT(keyval->val);
3566                         lpp_name_t     *name;
3567
3568                         name = si->lpp->vars[reload];
3569                         if(!is_zero(name->value)) {
3570                                 ir_node    *reload;
3571                                 ir_node    *insert_pos = bb;
3572                                 ir_node    *prev = sched_block_last_noncf(si, bb);
3573                                 op_t       *prev_op = get_irn_link(prev);
3574
3575                                 while(be_is_Spill(prev)) {
3576                                         prev = sched_prev(prev);
3577                                 }
3578
3579                                 prev_op = get_irn_link(prev);
3580
3581                                 /* insert reload before pre-remats */
3582                                 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3583                                                 && prev_op->is_remat && prev_op->attr.remat.pre) {
3584                                         insert_pos = prev;
3585
3586                                         do {
3587                                                 prev = sched_prev(prev);
3588                                         } while(be_is_Spill(prev));
3589
3590                                         prev_op = get_irn_link(prev);
3591
3592                                 }
3593
3594                                 reload = insert_reload(si, irn, insert_pos);
3595
3596 #ifdef KEEPALIVE_RELOADS
3597                                 pset_insert_ptr(si->spills, reload);
3598 #endif
3599                         }
3600                 }
3601         }
3602
3603         /* walk and insert more reloads and collect remats */
3604         sched_foreach_reverse(bb, irn) {
3605                 op_t     *op = get_irn_link(irn);
3606
3607                 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3608                 if(is_Phi(irn)) break;
3609
3610                 if(op->is_remat) {
3611                         if(get_irn_mode(irn) != mode_T) {
3612                                 insert_remat(si, irn);
3613                         }
3614                 } else {
3615                         int    n;
3616
3617                         for (n=get_irn_arity(irn)-1; n>=0; --n) {
3618                                 ir_node    *arg = get_irn_n(irn, n);
3619
3620                                 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3621                                         lpp_name_t    *name;
3622
3623                                         name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3624                                         if(!is_zero(name->value)) {
3625                                                 ir_node    *reload;
3626                                                 ir_node    *insert_pos = irn;
3627                                                 ir_node    *prev = sched_prev(insert_pos);
3628                                                 op_t       *prev_op;
3629
3630                                                 while(be_is_Spill(prev)) {
3631                                                         prev = sched_prev(prev);
3632                                                 }
3633
3634                                                 prev_op = get_irn_link(prev);
3635
3636                                                 /* insert reload before pre-remats */
3637                                                 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3638                                                                 && prev_op->is_remat && prev_op->attr.remat.pre) {
3639                                                         insert_pos = prev;
3640
3641                                                         do {
3642                                                                 prev = sched_prev(prev);
3643                                                         } while(be_is_Spill(prev));
3644
3645                                                         prev_op = get_irn_link(prev);
3646
3647                                                 }
3648
3649                                                 reload = insert_reload(si, arg, insert_pos);
3650
3651                                                 set_irn_n(irn, n, reload);
3652
3653 #ifdef KEEPALIVE_RELOADS
3654                                                 pset_insert_ptr(si->spills, reload);
3655 #endif
3656                                         }
3657                                 }
3658                         }
3659                 }
3660         }
3661
3662         del_set(spill_bb->ilp);
3663         if(spill_bb->reloads) del_set(spill_bb->reloads);
3664 }
3665
3666 static void
3667 walker_collect_used(ir_node * irn, void * data)
3668 {
3669         lc_bitset_t   *used = data;
3670
3671         lc_bitset_set(used, get_irn_idx(irn));
3672 }
3673
3674 struct kill_helper {
3675         lc_bitset_t  *used;
3676         spill_ilp_t  *si;
3677 };
3678
3679 static void
3680 walker_kill_unused(ir_node * bb, void * data)
3681 {
3682         struct kill_helper *kh = data;
3683         const ir_node      *bad = get_irg_bad(get_irn_irg(bb));
3684         ir_node            *irn;
3685
3686
3687         for(irn=sched_first(bb); !sched_is_end(irn);) {
3688                 ir_node     *next = sched_next(irn);
3689                 int          n;
3690
3691                 if(!lc_bitset_is_set(kh->used, get_irn_idx(irn))) {
3692                         if(be_is_Spill(irn) || be_is_Reload(irn)) {
3693                                 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)));
3694 #if 0
3695                                 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
3696 #endif
3697                         }
3698
3699                         sched_remove(irn);
3700
3701                         set_nodes_block(irn, bad);
3702                         for (n=get_irn_arity(irn)-1; n>=0; --n) {
3703                                 set_irn_n(irn, n, bad);
3704                         }
3705                 }
3706                 irn = next;
3707         }
3708 }
3709
3710 static void
3711 kill_all_unused_values_in_schedule(spill_ilp_t * si)
3712 {
3713         struct kill_helper kh;
3714
3715         kh.used = lc_bitset_malloc(get_irg_last_idx(si->chordal_env->irg));
3716         kh.si = si;
3717
3718         irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, kh.used);
3719         irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, &kh);
3720
3721         lc_bitset_free(kh.used);
3722 }
3723
3724 static void
3725 print_irn_pset(pset * p)
3726 {
3727         ir_node   *irn;
3728
3729         pset_foreach(p, irn) {
3730                 ir_printf("%+F\n", irn);
3731         }
3732 }
3733
3734 static void
3735 dump_phi_class(spill_ilp_t * si, pset * phiclass, const char * file)
3736 {
3737     FILE           *f = fopen(file, "w");
3738     ir_node        *irn;
3739     interference_t *interference;
3740
3741     pset_break(phiclass);
3742     set_break(si->interferences);
3743
3744     ir_fprintf(f, "digraph phiclass {\n");
3745
3746     pset_foreach(phiclass, irn) {
3747         if(is_Phi(irn))
3748             ir_fprintf(f, "  %F%N [shape=box]\n",irn,irn);
3749     }
3750
3751     pset_foreach(phiclass, irn) {
3752         int n;
3753
3754         if(!is_Phi(irn)) continue;
3755
3756         for(n=get_irn_arity(irn)-1; n>=0; --n) {
3757             ir_node  *arg = get_irn_n(irn, n);
3758
3759             ir_fprintf(f, "  %F%N -> %F%N\n",irn,irn,arg,arg);
3760         }
3761     }
3762
3763     set_foreach(si->interferences, interference) {
3764         const ir_node  *a  = interference->a;
3765         const ir_node  *b  = interference->b;
3766         if(get_phi_class(a) == phiclass) {
3767             ir_fprintf(f, "  %F%N -> %F%N [color=red,dir=none,style=bold]\n",a,a,b,b);
3768         }
3769     }
3770
3771     ir_fprintf(f, "}");
3772     fclose(f);
3773 }
3774
3775 static void
3776 rewire_uses(spill_ilp_t * si)
3777 {
3778         dom_front_info_t     *dfi = be_compute_dominance_frontiers(si->chordal_env->irg);
3779         defs_t               *defs;
3780         pset                 *ignore = pset_new_ptr(1);
3781
3782         pset_insert_ptr(ignore, get_irg_end(si->chordal_env->irg));
3783
3784         /* then fix uses of spills */
3785         set_foreach(si->values, defs) {
3786                 pset     *reloads;
3787                 pset     *spills;
3788                 ir_node  *next = defs->remats;
3789                 int remats = 0;
3790
3791                 reloads = pset_new_ptr_default();
3792
3793                 while(next) {
3794                         if(be_is_Reload(next)) {
3795                                 pset_insert_ptr(reloads, next);
3796                         } else {
3797                                 ++remats;
3798                         }
3799                         next = get_irn_link(next);
3800                 }
3801
3802                 spills = get_spills_for_value(si, defs->value);
3803                 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));
3804                 if(pset_count(spills) > 1) {
3805                         //assert(pset_count(reloads) > 0);
3806                         //                              print_irn_pset(spills);
3807                         //                              print_irn_pset(reloads);
3808
3809                         be_ssa_constr_set_ignore(dfi, si->lv, spills, ignore);
3810                 }
3811
3812                 del_pset(reloads);
3813                 del_pset(spills);
3814         }
3815
3816         /* first fix uses of remats and reloads */
3817         set_foreach(si->values, defs) {
3818                 pset     *nodes;
3819                 ir_node  *next = defs->remats;
3820
3821                 if(next) {
3822                         nodes = pset_new_ptr_default();
3823                         pset_insert_ptr(nodes, defs->value);
3824
3825                         while(next) {
3826                                 pset_insert_ptr(nodes, next);
3827                                 next = get_irn_link(next);
3828                         }
3829
3830                         if(pset_count(nodes) > 1) {
3831                                 DBG((si->dbg, LEVEL_4, "\t    %d new definitions for value %+F\n", pset_count(nodes)-1, defs->value));
3832                                 be_ssa_constr_set(dfi, si->lv, nodes);
3833                         }
3834
3835                         del_pset(nodes);
3836                 }
3837         }
3838
3839 //      remove_unused_defs(si);
3840
3841         be_free_dominance_frontiers(dfi);
3842 }
3843
3844
3845 static void
3846 writeback_results(spill_ilp_t * si)
3847 {
3848         /* walk through the graph and collect all spills, reloads and remats for a value */
3849
3850         si->values = new_set(cmp_defs, 4096);
3851
3852         DBG((si->dbg, LEVEL_1, "Applying results\n"));
3853         delete_unnecessary_remats(si);
3854         si->m_unknown = new_r_Unknown(si->chordal_env->irg, mode_M);
3855         irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
3856         irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
3857 #ifdef WITH_MEMOPERANDS
3858         insert_memoperands(si);
3859 #endif
3860         phim_fixer(si);
3861
3862         /* clean the remat info! there are still back-edges leading there! */
3863         clean_remat_info(si);
3864
3865         rewire_uses(si);
3866
3867         connect_all_spills_with_keep(si);
3868
3869         del_set(si->values);
3870 }
3871
3872 static int
3873 get_n_regs(spill_ilp_t * si)
3874 {
3875         int     arch_n_regs = arch_register_class_n_regs(si->cls);
3876         int     free = 0;
3877         int     i;
3878
3879         for(i=0; i<arch_n_regs; i++) {
3880                 if(!arch_register_type_is(&si->cls->regs[i], ignore)) {
3881                         free++;
3882                 }
3883         }
3884
3885         DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", free, si->cls->name));
3886         return free;
3887 }
3888
3889 static void
3890 walker_reload_mover(ir_node * bb, void * data)
3891 {
3892         spill_ilp_t   *si = data;
3893         ir_node           *tmp;
3894
3895         sched_foreach(bb, tmp) {
3896                 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
3897                         ir_node       *reload = tmp;
3898                         ir_node       *irn = tmp;
3899
3900                         /* move reload upwards */
3901
3902                         int pressure = (int)get_irn_link(reload);
3903                         if(pressure < si->n_regs) {
3904                                 irn = sched_prev(reload);
3905                                 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
3906                                 sched_remove(reload);
3907                                 pressure = (int)get_irn_link(irn);
3908
3909                                 while(pressure < si->n_regs) {
3910                                         if( sched_is_end(irn) ||
3911                                            (be_is_Reload(irn) && has_reg_class(si, irn)) ||
3912                                            /* do not move reload before its spill */
3913                                            (irn == be_get_Reload_mem(reload)) ) break;
3914
3915                                         set_irn_link(irn, INT_TO_PTR(pressure+1));
3916                                         DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
3917                                         irn = sched_prev(irn);
3918
3919                                         pressure = (int)get_irn_link(irn);
3920                                 }
3921
3922                                 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
3923                                 sched_put_after(irn, reload);
3924                         }
3925                 }
3926         }
3927 }
3928
3929 static void
3930 move_reloads_upward(spill_ilp_t * si)
3931 {
3932         irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si);
3933 }
3934
3935
3936 /**
3937  * Walk all irg blocks and check for interfering spills inside of phi classes
3938  */
3939 static void
3940 luke_meminterferencechecker(ir_node * bb, void * data)
3941 {
3942         spill_ilp_t    *si = (spill_ilp_t*)data;
3943         int             l1, l2;
3944
3945         be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
3946                 ir_node        *a = be_lv_get_irn(si->lv, bb, l1);
3947
3948                 if(!be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T)) continue;
3949
3950                 /* a is only interesting if it is in my register class and if it is inside a phi class */
3951                 if (has_reg_class(si, a) && get_phi_class(a)) {
3952                         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)) {
3953                                 ir_node        *b = be_lv_get_irn(si->lv, bb, l2);
3954
3955                                 if(!be_is_Spill(b) && (!is_Phi(b) || get_irn_mode(b) != mode_T)) continue;
3956
3957                                 /* a and b are only interesting if they are in the same phi class */
3958                                 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
3959                                         if(values_interfere_in_block(si, bb, a, b)) {
3960                                                 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
3961                                         }
3962                                 }
3963                         }
3964                 }
3965         }
3966 }
3967
3968 static void
3969 verify_phiclasses(spill_ilp_t * si)
3970 {
3971         /* analyze phi classes */
3972         phi_class_compute(si->chordal_env->irg);
3973
3974         DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
3975         irg_block_walk_graph(si->chordal_env->irg, luke_meminterferencechecker, NULL, si);
3976 }
3977
3978 static void
3979 walker_spillslotassigner(ir_node * irn, void * data)
3980 {
3981         spill_ilp_t            *si = (spill_ilp_t*)data;
3982         void                   *cls;
3983
3984         if(!be_is_Spill(irn)) return;
3985
3986         /* set spill context to phi class if it has one ;) */
3987
3988         cls = get_phi_class(irn);
3989         if(cls)
3990                 be_set_Spill_context(irn, cls);
3991         else
3992                 be_set_Spill_context(irn, irn);
3993 }
3994
3995
3996 static void
3997 assign_spillslots(spill_ilp_t * si)
3998 {
3999         DBG((si->dbg, LEVEL_2, "\t calling spill slot assigner\n"));
4000         irg_walk_graph(si->chordal_env->irg, walker_spillslotassigner, NULL, si);
4001 }
4002
4003 void
4004 be_spill_remat(const be_chordal_env_t * chordal_env)
4005 {
4006         char            problem_name[256];
4007         char            dump_suffix[256];
4008         char            dump_suffix2[256];
4009         struct obstack  obst;
4010         spill_ilp_t     si;
4011
4012         ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name);
4013         ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name);
4014         ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name);
4015
4016         FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
4017         DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
4018
4019 #ifdef VERIFY_DOMINANCE
4020     be_check_dominance(chordal_env->irg);
4021 #endif
4022
4023         obstack_init(&obst);
4024         si.chordal_env = chordal_env;
4025         si.obst = &obst;
4026         si.cls = chordal_env->cls;
4027         si.lpp = new_lpp(problem_name, lpp_minimize);
4028         si.remat_info = new_set(cmp_remat_info, 4096);
4029         si.interferences = new_set(cmp_interference, 32);
4030 #ifdef WITH_MEMOPERANDS
4031         si.memoperands = new_set(cmp_memoperands, 128);
4032 #endif
4033         si.all_possible_remats = pset_new_ptr_default();
4034         si.spills = pset_new_ptr_default();
4035         si.inverse_ops = pset_new_ptr_default();
4036         si.lv = chordal_env->lv;
4037 #ifdef KEEPALIVE
4038         si.keep = NULL;
4039 #endif
4040         si.n_regs = get_n_regs(&si);
4041
4042         set_irg_link(chordal_env->irg, &si);
4043         compute_doms(chordal_env->irg);
4044
4045         /* compute phi classes */
4046 //      phi_class_compute(chordal_env->irg);
4047
4048         be_analyze_regpressure(chordal_env, "-pre");
4049
4050 #ifdef COLLECT_REMATS
4051         /* collect remats */
4052         DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
4053         irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
4054 #endif
4055
4056         /* insert possible remats */
4057         DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
4058         irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
4059         DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
4060
4061 #ifdef KEEPALIVE
4062         DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
4063         connect_all_remats_with_keep(&si);
4064         /* dump graph with inserted remats */
4065         dump_graph_with_remats(chordal_env->irg, dump_suffix);
4066 #endif
4067
4068         /* insert copies for phi arguments not in my regclass */
4069         irg_walk_graph(chordal_env->irg, walker_regclass_copy_insertor, NULL, &si);
4070
4071         /* recompute liveness */
4072         DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
4073         be_liveness_recompute(si.lv);
4074
4075         /* build the ILP */
4076
4077         DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
4078         DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
4079         irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si);
4080
4081         DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
4082         irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
4083
4084 #ifndef NO_MEMCOPIES
4085         DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
4086         memcopyhandler(&si);
4087 #endif
4088
4089 #ifdef DUMP_ILP
4090         {
4091                 FILE           *f;
4092                 char            buf[256];
4093
4094                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
4095                 if ((f = fopen(buf, "wt")) != NULL) {
4096                         lpp_dump_plain(si.lpp, f);
4097                         fclose(f);
4098                 }
4099         }
4100 #endif
4101
4102 #ifdef SOLVE
4103         DBG((si.dbg, LEVEL_1, "\tSolving %F\n", chordal_env->irg));
4104 #ifdef ILP_TIMEOUT
4105         lpp_set_time_limit(si.lpp, ILP_TIMEOUT);
4106 #endif
4107
4108 #ifdef SOLVE_LOCAL
4109         lpp_solve_cplex(si.lpp);
4110 #else
4111         lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
4112 #endif
4113         assert(lpp_is_sol_valid(si.lpp)
4114                && "solution of ILP must be valid");
4115
4116         DBG((si.dbg, LEVEL_1, "\t%s: iterations: %d, solution time: %g, objective function: %g\n", problem_name, si.lpp->iterations, si.lpp->sol_time, is_zero(si.lpp->objval)?0.0:si.lpp->objval));
4117
4118 #ifdef DUMP_SOLUTION
4119         {
4120                 FILE           *f;
4121                 char            buf[256];
4122
4123                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
4124                 if ((f = fopen(buf, "wt")) != NULL) {
4125                         int             i;
4126                         for (i = 0; i < si.lpp->var_next; ++i) {
4127                                 lpp_name_t     *name = si.lpp->vars[i];
4128                                 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
4129                         }
4130                         fclose(f);
4131                 }
4132         }
4133 #endif
4134
4135         writeback_results(&si);
4136
4137 #endif                          /* SOLVE */
4138
4139         kill_all_unused_values_in_schedule(&si);
4140
4141 #if defined(KEEPALIVE_SPILLS) || defined(KEEPALIVE_RELOADS)
4142         be_dump(chordal_env->irg, "-spills-placed", dump_ir_block_graph);
4143 #endif
4144
4145         // move reloads upwards
4146         be_liveness_recompute(si.lv);
4147         irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4148         move_reloads_upward(&si);
4149
4150 #ifndef NO_MEMCOPIES
4151         verify_phiclasses(&si);
4152         assign_spillslots(&si);
4153 #endif
4154
4155         irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4156
4157         dump_pressure_graph(&si, dump_suffix2);
4158
4159         be_analyze_regpressure(chordal_env, "-post");
4160
4161 #ifdef VERIFY_DOMINANCE
4162         be_check_dominance(chordal_env->irg);
4163 #endif
4164
4165         free_dom(chordal_env->irg);
4166         del_set(si.interferences);
4167         del_pset(si.inverse_ops);
4168         del_pset(si.all_possible_remats);
4169 #ifdef WITH_MEMOPERANDS
4170         del_set(si.memoperands);
4171 #endif
4172         del_pset(si.spills);
4173         free_lpp(si.lpp);
4174         obstack_free(&obst, NULL);
4175         DBG((si.dbg, LEVEL_1, "\tdone.\n"));
4176 }
4177
4178 #else                           /* WITH_ILP */
4179
4180 static void
4181 only_that_you_can_compile_without_WITH_ILP_defined(void)
4182 {
4183 }
4184
4185 #endif                          /* WITH_ILP */