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

   1 /** vim: set sw=4 ts=4:
   2  * @file   bespillremat.c
   3  * @date   2006-04-06
   4  * @author Adam M. Szalkowski & Sebastian Hack
   5  *
   6  * ILP based spilling & rematerialization
   7  *
   8  * Copyright (C) 2006 Universitaet Karlsruhe
   9  * Released under the GPL
  10  */
  11 #ifdef HAVE_CONFIG_H
  12 #include "config.h"
  13 #endif
  14
  15 #ifdef WITH_ILP
  16
  17 #include <math.h>
  18
  19 #include "hashptr.h"
  20 #include "debug.h"
  21 #include "obst.h"
  22 #include "set.h"
  23 #include "list.h"
  24 #include "pmap.h"
  25
  26 #include "irprintf.h"
  27 #include "irgwalk.h"
  28 #include "irdump_t.h"
  29 #include "irnode_t.h"
  30 #include "ircons_t.h"
  31 #include "irloop_t.h"
  32 #include "phiclass.h"
  33 #include "iredges.h"
  34
  35 #include <lpp/lpp.h>
  36 #include <lpp/lpp_net.h>
  37 #include <lpp/lpp_cplex.h>
  38 //#include <lc_pset.h>
  39 #include <libcore/lc_bitset.h>
  40
  41 #include "be_t.h"
  42 #include "belive_t.h"
  43 #include "besched_t.h"
  44 #include "beirgmod.h"
  45 #include "bearch.h"
  46 #include "benode_t.h"
  47 #include "beutil.h"
  48 #include "bespillremat.h"
  49 #include "bespill.h"
  50
  51 #include "bechordal_t.h"
  52
  53 #define BIGM 100000.0
  54
  55 #define DUMP_SOLUTION
  56 #define DUMP_ILP
  57 #define KEEPALIVE /* keep alive all inserted remats and dump graph with remats */
  58 #define COLLECT_REMATS /* enable rematerialization */
  59 #define COLLECT_INVERSE_REMATS /* enable placement of inverse remats */
  60 #define REMAT_WHILE_LIVE /* only remat values that are live */
  61 //#define NO_ENLARGE_L1V3N355 /* do not remat after the death of some operand */
  62 #define EXECFREQ_LOOPDEPH /* compute execution frequency from loop depth only */
  63 //#define MAY_DIE_AT_PRE_REMAT /* allow values to die after a pre remat */
  64 //#define CHECK_POST_REMAT /* check pressure after post remats (conservative but otherwise we can temporarily exceed the register pressure) */
  65 #define NO_SINGLE_USE_REMATS /* do not repair schedule */
  66
  67 #define  SOLVE
  68 #undef  SOLVE_LOCAL
  69 #define LPP_SERVER "i44pc52"
  70 #define LPP_SOLVER "cplex"
  71
  72 #ifndef EXECFREQ_LOOPDEPH
  73 #include "execfreq.h"
  74 #endif
  75
  76 #define COST_LOAD      10
  77 #define COST_STORE     50
  78 #define COST_REMAT     1
  79
  80 #define ILP_TIMEOUT    20
  81
  82 #define ILP_UNDEF               -1
  83
  84 typedef struct _spill_ilp_t {
  85         const arch_register_class_t  *cls;
  86         int                           n_regs;
  87         const be_chordal_env_t       *chordal_env;
  88         spill_env_t                  *senv;
  89         lpp_t                        *lpp;
  90         struct obstack               *obst;
  91         set                          *remat_info;
  92         pset                         *all_possible_remats;
  93         pset                         *inverse_ops;
  94 #ifdef KEEPALIVE
  95         ir_node                      *keep;
  96 #endif
  97         set                          *values; /**< for collecting all definitions of values before running ssa-construction */
  98 #ifndef EXECFREQ_LOOPDEPH
  99         set                          *execfreqs;
 100 #endif
 101         DEBUG_ONLY(firm_dbg_module_t * dbg);
 102 } spill_ilp_t;
 103
 104 typedef int ilp_var_t;
 105 typedef int ilp_cst_t;
 106
 107 typedef struct _spill_bb_t {
 108         set          *ilp;
 109         pset         *copys_needed;
 110         ilp_var_t    *reloads;
 111 } spill_bb_t;
 112
 113 typedef struct _remat_t {
 114         const ir_node        *op; /**< for copy_irn */
 115         const ir_node        *proj; /**< not NULL if the above op produces a tuple */
 116         const ir_node        *value; /**< the value which is being recomputed by this remat */
 117         int                   cost; /**< cost of this remat */
 118         int                   inverse; /**< nonzero if this is an inverse remat */
 119 } remat_t;
 120
 121 /**
 122  * Data to be attached to each IR node. For remats this contains the ilp_var
 123  * for this remat and for normal ops this contains the ilp_vars for
 124  * reloading each operand
 125  */
 126 typedef struct _op_t {
 127         int             is_remat;
 128         union {
 129                 struct {
 130                         ilp_var_t       ilp;
 131                         remat_t        *remat; /** the remat this op belongs to */
 132                         int             pre; /** 1, if this is a pressure-increasing remat */
 133                 } remat;
 134                 struct {
 135                         ilp_var_t       ilp;
 136                         ir_node        *op; /** the operation this live range belongs to */
 137                         ilp_var_t      *reloads;
 138                 } live_range;
 139         } attr;
 140 } op_t;
 141
 142 typedef struct _defs_t {
 143         ir_node   *value;
 144         ir_node   *spills;  /**< points to the first spill for this value (linked by link field) */
 145         ir_node   *remats;  /**< points to the first definition for this value (linked by link field) */
 146 } defs_t;
 147
 148 typedef struct _remat_info_t {
 149         const ir_node       *irn; /**< the irn to which these remats belong */
 150         pset                *remats; /**< possible remats for this value */
 151         pset                *remats_by_operand; /**< remats with this value as operand */
 152 } remat_info_t;
 153
 154 typedef struct _keyval_t {
 155         const void          *key;
 156         const void          *val;
 157 } keyval_t;
 158
 159 typedef struct _spill_t {
 160         ir_node      *irn;
 161         ilp_var_t     reg_in;
 162         ilp_var_t     mem_in;
 163         ilp_var_t     reg_out;
 164         ilp_var_t     mem_out;
 165         ilp_var_t     spill;
 166 } spill_t;
 167
 168 static INLINE int
 169 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
 170 {
 171         return chordal_has_class(si->chordal_env, irn);
 172 }
 173
 174 #if 0
 175 static int
 176 cmp_remat(const void *a, const void *b)
 177 {
 178         const keyval_t *p = a;
 179         const keyval_t *q = b;
 180         const remat_t  *r = p->val;
 181         const remat_t  *s = q->val;
 182
 183         assert(r && s);
 184
 185         return !(r == s || r->op == s->op);
 186 }
 187 #endif
 188 static int
 189 cmp_remat(const void *a, const void *b)
 190 {
 191         const remat_t  *r = a;
 192         const remat_t  *s = a;
 193
 194         return !(r == s || r->op == s->op);
 195 }
 196
 197 static int
 198 cmp_spill(const void *a, const void *b, size_t size)
 199 {
 200         const spill_t *p = a;
 201         const spill_t *q = b;
 202
 203 //      return !(p->irn == q->irn && p->bb == q->bb);
 204         return !(p->irn == q->irn);
 205 }
 206
 207 static keyval_t *
 208 set_find_keyval(set * set, void * key)
 209 {
 210         keyval_t     query;
 211
 212         query.key = key;
 213         return set_find(set, &query, sizeof(query), HASH_PTR(key));
 214 }
 215
 216 static keyval_t *
 217 set_insert_keyval(set * set, void * key, void * val)
 218 {
 219         keyval_t     query;
 220
 221         query.key = key;
 222         query.val = val;
 223         return set_insert(set, &query, sizeof(query), HASH_PTR(key));
 224 }
 225
 226 static defs_t *
 227 set_find_def(set * set, ir_node * value)
 228 {
 229         defs_t     query;
 230
 231         query.value = value;
 232         return set_find(set, &query, sizeof(query), HASH_PTR(value));
 233 }
 234
 235 static defs_t *
 236 set_insert_def(set * set, ir_node * value)
 237 {
 238         defs_t     query;
 239
 240         query.value = value;
 241         query.spills = NULL;
 242         query.remats = NULL;
 243         return set_insert(set, &query, sizeof(query), HASH_PTR(value));
 244 }
 245
 246 static spill_t *
 247 set_find_spill(set * set, ir_node * value)
 248 {
 249         spill_t     query;
 250
 251         query.irn = value;
 252         return set_find(set, &query, sizeof(query), HASH_PTR(value));
 253 }
 254
 255 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
 256 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
 257 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
 258 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
 259 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
 260
 261 static int
 262 cmp_remat_info(const void *a, const void *b, size_t size)
 263 {
 264         const remat_info_t *p = a;
 265         const remat_info_t *q = b;
 266
 267         return !(p->irn == q->irn);
 268 }
 269
 270 static int
 271 cmp_defs(const void *a, const void *b, size_t size)
 272 {
 273         const defs_t *p = a;
 274         const defs_t *q = b;
 275
 276         return !(p->value == q->value);
 277 }
 278
 279 static int
 280 cmp_keyval(const void *a, const void *b, size_t size)
 281 {
 282         const keyval_t *p = a;
 283         const keyval_t *q = b;
 284
 285         return !(p->key == q->key);
 286 }
 287
 288 static double
 289 execution_frequency(const spill_ilp_t * si, const ir_node * irn)
 290 {
 291 #ifdef EXECFREQ_LOOPDEPH
 292         if(is_Block(irn))
 293                 return exp(get_loop_depth(get_irn_loop(irn)) * log(10));
 294         else
 295                 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10));
 296 #else
 297         if(is_Block(irn)) {
 298                 return get_block_execfreq(si->execfreqs, irn);
 299         } else {
 300                 return get_block_execfreq(si->execfreqs, get_nodes_block(irn));
 301         }
 302 #endif
 303 }
 304
 305 /**
 306  * Checks, whether node and its operands have suitable reg classes
 307  */
 308 static INLINE int
 309 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
 310 {
 311         int             i,
 312                         n;
 313         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
 314         int               remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
 315
 316 #if 0
 317         if(!remat)
 318                 ir_fprintf(stderr, "  Node %+F is not rematerializable\n", irn);
 319 #endif
 320
 321         for (i = 0, n = get_irn_arity(irn); i < n && remat; ++i) {
 322                 ir_node        *op = get_irn_n(irn, i);
 323                 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
 324
 325 //              if(!remat)
 326 //                      ir_fprintf(stderr, "  Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
 327         }
 328
 329         return remat;
 330 }
 331
 332 /**
 333  * Try to create a remat from @p op with destination value @p dest_value
 334  */
 335 static INLINE remat_t *
 336 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
 337 {
 338         remat_t  *remat = NULL;
 339
 340 //      if(!mode_is_datab(get_irn_mode(dest_value)))
 341 //              return NULL;
 342
 343         if(dest_value == op) {
 344                 const ir_node *proj = NULL;
 345
 346                 if(is_Proj(dest_value)) {
 347                         op = get_irn_n(op, 0);
 348                         proj = dest_value;
 349                 }
 350
 351                 if(!is_rematerializable(si, op))
 352                         return NULL;
 353
 354                 remat = obstack_alloc(si->obst, sizeof(*remat));
 355                 remat->op = op;
 356                 remat->cost = COST_REMAT; /* TODO ask backend for real cost */
 357                 remat->value = dest_value;
 358                 remat->proj = proj;
 359                 remat->inverse = 0;
 360         } else {
 361                 arch_inverse_t     inverse;
 362                 int                i,
 363                                                    n;
 364
 365                 /* get the index of the operand we want to retrieve by the inverse op */
 366                 for (i = 0, n = get_irn_arity(op); i < n; ++i) {
 367                         ir_node        *arg = get_irn_n(op, i);
 368
 369                         if(arg == dest_value) break;
 370                 }
 371                 if(i == n) return NULL;
 372
 373                 DBG((si->dbg, LEVEL_5, "\t  requesting inverse op for argument %d of op %+F\n", i, op));
 374
 375                 /* else ask the backend to give an inverse op */
 376                 if(arch_get_inverse(si->chordal_env->birg->main_env->arch_env, op, i, &inverse, si->obst)) {
 377                         int   i;
 378
 379                         DBG((si->dbg, LEVEL_4, "\t  backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
 380
 381                         assert(inverse.n > 0 && "inverse op should have at least one node");
 382
 383                         for(i=0; i<inverse.n; ++i) {
 384                                 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
 385                         }
 386
 387                         if(inverse.n <= 2) {
 388                                 remat = obstack_alloc(si->obst, sizeof(*remat));
 389                                 remat->op = inverse.nodes[0];
 390                                 remat->cost = inverse.costs;
 391                                 remat->value = dest_value;
 392                                 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
 393                                 remat->inverse = 1;
 394
 395                                 assert(is_Proj(remat->proj));
 396                         } else {
 397                                 assert(0 && "I can not handle remats with more than 2 nodes");
 398                         }
 399                 }
 400         }
 401
 402         if(remat) {
 403                 if(remat->proj) {
 404                         DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
 405                 } else {
 406                         DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
 407                 }
 408         }
 409         return remat;
 410 }
 411
 412
 413 static INLINE void
 414 add_remat(const spill_ilp_t * si, const remat_t * remat)
 415 {
 416         remat_info_t    *remat_info,
 417                      query;
 418         int              i,
 419                                          n;
 420
 421         assert(remat->op);
 422         assert(remat->value);
 423
 424         query.irn = remat->value;
 425         query.remats = NULL;
 426         query.remats_by_operand = NULL;
 427         remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
 428
 429         if(remat_info->remats == NULL) {
 430                 remat_info->remats = new_pset(cmp_remat, 4096);
 431         }
 432         pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
 433
 434         /* insert the remat into the remats_be_operand set of each argument of the remat op */
 435         for (i = 0, n = get_irn_arity(remat->op); i < n; ++i) {
 436                 ir_node        *arg = get_irn_n(remat->op, i);
 437
 438                 query.irn = arg;
 439                 query.remats = NULL;
 440                 query.remats_by_operand = NULL;
 441                 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
 442
 443                 if(remat_info->remats_by_operand == NULL) {
 444                         remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
 445                 }
 446                 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
 447         }
 448 }
 449
 450 static int
 451 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
 452 {
 453         const ir_edge_t   *edge = get_irn_out_edge_first(irn);
 454         int                i = 0;
 455
 456         while(edge) {
 457                 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
 458                         ++i;
 459                 }
 460                 edge = get_irn_out_edge_next(irn, edge);
 461         }
 462
 463         return i;
 464 }
 465
 466 static INLINE void
 467 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
 468 {
 469         int       i,
 470                       n;
 471         remat_t *remat;
 472
 473 #ifdef NO_SINGLE_USE_REMATS
 474         if(has_reg_class(si, op) && (get_irn_n_nonremat_edges(si, op) > 1)) {
 475 #else
 476         if(has_reg_class(si, op)) {
 477 #endif
 478                 remat = get_remat_from_op(si, op, op);
 479                 if(remat) {
 480                         add_remat(si, remat);
 481                 }
 482         }
 483
 484 #ifdef COLLECT_INVERSE_REMATS
 485         /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
 486            for each arg */
 487         for (i = 0, n = get_irn_arity(op); i < n; ++i) {
 488                 ir_node        *arg = get_irn_n(op, i);
 489
 490                 if(has_reg_class(si, arg)) {
 491                         /* try to get an inverse remat */
 492                         remat = get_remat_from_op(si, arg, op);
 493                         if(remat) {
 494                                 add_remat(si, remat);
 495                         }
 496                 }
 497         }
 498 #endif
 499
 500 }
 501
 502 static INLINE int
 503 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
 504 {
 505         ir_node *block;
 506         ir_node *def_block = get_nodes_block(val);
 507         int      ret;
 508
 509         if(val == pos)
 510                 return 0;
 511
 512         /* if pos is at end of a basic block */
 513         if(is_Block(pos)) {
 514                 ret = (pos == def_block || block_dominates(def_block, pos));
 515 //              ir_fprintf(stderr, "(def(bb)=%d) ", ret);
 516                 return ret;
 517         }
 518
 519         /* else if this is a normal operation */
 520         block = get_nodes_block(pos);
 521         if(block == def_block) {
 522                 if(!sched_is_scheduled(val)) return 1;
 523
 524                 ret = sched_comes_after(val, pos);
 525 //              ir_fprintf(stderr, "(def(same block)=%d) ",ret);
 526                 return ret;
 527         }
 528
 529         ret = block_dominates(def_block, block);
 530 //      ir_fprintf(stderr, "(def(other block)=%d) ", ret);
 531         return ret;
 532 }
 533
 534 static INLINE ir_node *
 535 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
 536 {
 537     return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->chordal_env->birg->main_env->arch_env);
 538 }
 539
 540 /**
 541  * Returns first non-Phi node of block @p bb
 542  */
 543 static INLINE ir_node *
 544 sched_block_first_nonphi(const ir_node * bb)
 545 {
 546         return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
 547 }
 548
 549 static int
 550 sched_skip_proj_predicator(const ir_node * irn, void * data)
 551 {
 552         return (is_Proj(irn));
 553 }
 554
 555 static INLINE ir_node *
 556 sched_next_nonproj(const ir_node * irn, int forward)
 557 {
 558         return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
 559 }
 560
 561 /**
 562  * Returns next operation node (non-Proj) after @p irn
 563  * or the basic block of this node
 564  */
 565 static INLINE ir_node *
 566 sched_next_op(const ir_node * irn)
 567 {
 568         ir_node *next = sched_next(irn);
 569
 570         if(is_Block(next))
 571                 return next;
 572
 573         return sched_next_nonproj(next, 1);
 574 }
 575
 576 /**
 577  * Returns previous operation node (non-Proj) before @p irn
 578  * or the basic block of this node
 579  */
 580 static INLINE ir_node *
 581 sched_prev_op(const ir_node * irn)
 582 {
 583         ir_node *prev = sched_prev(irn);
 584
 585         if(is_Block(prev))
 586                 return prev;
 587
 588         return sched_next_nonproj(prev, 0);
 589 }
 590
 591 static void
 592 sched_put_after(ir_node * insert, ir_node * irn)
 593 {
 594         if(is_Block(insert)) {
 595                 insert = sched_block_first_nonphi(insert);
 596         } else {
 597                 insert = sched_next_op(insert);
 598         }
 599         sched_add_before(insert, irn);
 600 }
 601
 602 static void
 603 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
 604 {
 605   if(is_Block(insert)) {
 606           insert = sched_block_last_noncf(si, insert);
 607   } else {
 608           insert = sched_next_nonproj(insert, 0);
 609           insert = sched_prev(insert);
 610   }
 611   sched_add_after(insert, irn);
 612 }
 613
 614 /**
 615  * Tells you whether a @p remat can be placed before the irn @p pos
 616  */
 617 static INLINE int
 618 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 619 {
 620         const ir_node   *op = remat->op;
 621         const ir_node   *prev;
 622         int        i,
 623                            n,
 624                            res = 1;
 625
 626         if(is_Block(pos)) {
 627                 prev = sched_block_last_noncf(si, pos);
 628                 prev = sched_next_nonproj(prev, 0);
 629         } else {
 630                 prev = sched_prev_op(pos);
 631         }
 632         /* do not remat if the rematted value is defined immediately before this op */
 633         if(prev == remat->op) {
 634                 return 0;
 635         }
 636
 637 #if 0
 638         /* this should be just fine, the following OP will be using this value, right? */
 639
 640         /* only remat AFTER the real definition of a value (?) */
 641         if(!value_is_defined_before(si, pos, remat->value)) {
 642 //              ir_fprintf(stderr, "error(not defined)");
 643                 return 0;
 644         }
 645 #endif
 646
 647         for(i=0, n=get_irn_arity(op); i<n && res; ++i) {
 648                 const ir_node   *arg = get_irn_n(op, i);
 649
 650 #ifdef NO_ENLARGE_L1V3N355
 651                 if(has_reg_class(si, arg) && live) {
 652                         res &= pset_find_ptr(live, arg)?1:0;
 653                 } else {
 654                         res &= value_is_defined_before(si, pos, arg);
 655                 }
 656 #else
 657                 res &= value_is_defined_before(si, pos, arg);
 658 #endif
 659         }
 660
 661         return res;
 662 }
 663
 664 /**
 665  * Tells you whether a @p remat can be placed after the irn @p pos
 666  */
 667 static INLINE int
 668 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 669 {
 670         if(is_Block(pos)) {
 671                 pos = sched_block_first_nonphi(pos);
 672         } else {
 673                 pos = sched_next_op(pos);
 674         }
 675
 676         /* only remat AFTER the real definition of a value (?) */
 677         if(!value_is_defined_before(si, pos, remat->value)) {
 678                 return 0;
 679         }
 680
 681         return can_remat_before(si, remat, pos, live);
 682 }
 683
 684 /**
 685  * Collect potetially rematerializable OPs
 686  */
 687 static void
 688 walker_remat_collector(ir_node * irn, void * data)
 689 {
 690         spill_ilp_t    *si = data;
 691
 692         if(!is_Block(irn) && !is_Phi(irn)) {
 693                 DBG((si->dbg, LEVEL_4, "\t  Processing %+F\n", irn));
 694                 get_remats_from_op(si, irn);
 695         }
 696 }
 697
 698 /**
 699  * Inserts a copy of @p irn before @p pos
 700  */
 701 static ir_node *
 702 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
 703 {
 704         ir_node     *bb;
 705         ir_node     *copy;
 706
 707         bb = is_Block(pos)?pos:get_nodes_block(pos);
 708         copy = exact_copy(irn);
 709         set_nodes_block(copy, bb);
 710         sched_put_before(si, pos, copy);
 711
 712         return copy;
 713 }
 714
 715 /**
 716  * Inserts a copy of @p irn after @p pos
 717  */
 718 static ir_node *
 719 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
 720 {
 721         ir_node     *bb;
 722         ir_node     *copy;
 723
 724         bb = is_Block(pos)?pos:get_nodes_block(pos);
 725         copy = exact_copy(irn);
 726         set_nodes_block(copy, bb);
 727         sched_put_after(pos, copy);
 728
 729         return copy;
 730 }
 731
 732 static void
 733 insert_remat_after(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 734 {
 735         char     buf[256];
 736
 737         if(can_remat_after(si, remat, pos, live)) {
 738                 ir_node         *copy,
 739                                                 *proj_copy;
 740                 op_t            *op;
 741
 742                 DBG((si->dbg, LEVEL_3, "\t  >inserting remat %+F\n", remat->op));
 743
 744                 copy = insert_copy_after(si, remat->op, pos);
 745
 746 //              ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", remat->value, pos);
 747                 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
 748                 op = obstack_alloc(si->obst, sizeof(*op));
 749                 op->is_remat = 1;
 750                 op->attr.remat.remat = remat;
 751                 op->attr.remat.pre = 0;
 752                 op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos));
 753
 754                 set_irn_link(copy, op);
 755                 pset_insert_ptr(si->all_possible_remats, copy);
 756                 if(remat->proj) {
 757                         proj_copy = insert_copy_after(si, remat->proj, copy);
 758                         set_irn_n(proj_copy, 0, copy);
 759                         set_irn_link(proj_copy, op);
 760                         pset_insert_ptr(si->all_possible_remats, proj_copy);
 761                 } else {
 762                         proj_copy = NULL;
 763                 }
 764         }
 765 }
 766
 767 static void
 768 insert_remat_before(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 769 {
 770         char     buf[256];
 771
 772         if(can_remat_before(si, remat, pos, live)) {
 773                 ir_node         *copy,
 774                                                 *proj_copy;
 775                 op_t            *op;
 776
 777                 DBG((si->dbg, LEVEL_3, "\t  >inserting remat %+F\n", remat->op));
 778
 779                 copy = insert_copy_before(si, remat->op, pos);
 780
 781 //              ir_snprintf(buf, sizeof(buf), "remat_%N_%N", remat->value, pos);
 782                 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
 783                 op = obstack_alloc(si->obst, sizeof(*op));
 784                 op->is_remat = 1;
 785                 op->attr.remat.remat = remat;
 786                 op->attr.remat.pre = 1;
 787                 op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos));
 788
 789                 set_irn_link(copy, op);
 790                 pset_insert_ptr(si->all_possible_remats, copy);
 791                 if(remat->proj) {
 792                         proj_copy = insert_copy_after(si, remat->proj, copy);
 793                         set_irn_n(proj_copy, 0, copy);
 794                         set_irn_link(proj_copy, op);
 795                         pset_insert_ptr(si->all_possible_remats, proj_copy);
 796                 } else {
 797                         proj_copy = NULL;
 798                 }
 799         }
 800 }
 801
 802
 803 /**
 804  * Insert (so far unused) remats into the irg to
 805  * recompute the potential liveness of all values
 806  */
 807 static void
 808 walker_remat_insertor(ir_node * bb, void * data)
 809 {
 810         spill_ilp_t    *si = data;
 811         spill_bb_t     *spill_bb;
 812         ir_node        *irn;
 813         int             i,
 814                                         n;
 815         irn_live_t     *li;
 816         pset           *live = pset_new_ptr_default();
 817
 818         DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
 819
 820         live_foreach(bb, li) {
 821                 ir_node        *value = (ir_node *) li->irn;
 822
 823                 /* add remats at end of block */
 824                 if (live_is_end(li) && has_reg_class(si, value)) {
 825                         pset_insert_ptr(live, value);
 826                 }
 827         }
 828
 829         spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
 830         set_irn_link(bb, spill_bb);
 831
 832         irn = sched_last(bb);
 833         while(!sched_is_end(irn)) {
 834                 ir_node   *next;
 835                 op_t      *op;
 836                 pset      *args;
 837                 ir_node   *arg;
 838
 839                 next = sched_prev(irn);
 840
 841                 DBG((si->dbg, LEVEL_5, "\t at %+F (next: %+F)\n", irn, next));
 842
 843                 if(is_Phi(irn) || is_Proj(irn)) {
 844                         op_t      *op;
 845
 846                         if(has_reg_class(si, irn)) {
 847                                 pset_remove_ptr(live, irn);
 848                         }
 849
 850                         op = obstack_alloc(si->obst, sizeof(*op));
 851                         op->is_remat = 0;
 852                         op->attr.live_range.reloads = NULL;
 853                         op->attr.live_range.ilp = ILP_UNDEF;
 854                         set_irn_link(irn, op);
 855
 856                         irn = next;
 857                         continue;
 858                 }
 859
 860                 op = obstack_alloc(si->obst, sizeof(*op));
 861                 op->is_remat = 0;
 862                 op->attr.live_range.ilp = ILP_UNDEF;
 863                 op->attr.live_range.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.reloads) * get_irn_arity(irn));
 864                 memset(op->attr.live_range.reloads, 0xFF, sizeof(*op->attr.live_range.reloads) * get_irn_arity(irn));
 865                 set_irn_link(irn, op);
 866
 867                 args = pset_new_ptr_default();
 868
 869                 /* collect arguments of op */
 870                 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
 871                         ir_node        *arg = get_irn_n(irn, i);
 872
 873                         pset_insert_ptr(args, arg);
 874                 }
 875
 876                 /* set args of op live in epilog */
 877                 pset_foreach(args, arg) {
 878                         if(has_reg_class(si, arg)) {
 879                                 pset_insert_ptr(live, arg);
 880                         }
 881                 }
 882
 883                 /* insert all possible remats after irn */
 884                 pset_foreach(args, arg) {
 885                         remat_info_t   *remat_info,
 886                                                     query;
 887                         remat_t        *remat;
 888
 889                         /* continue if the operand has the wrong reg class
 890                          */
 891                         if(!has_reg_class(si, arg))
 892                                 continue;
 893
 894                         query.irn = arg;
 895                         query.remats = NULL;
 896                         query.remats_by_operand = NULL;
 897                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
 898
 899                         if(!remat_info) {
 900                                 continue;
 901                         }
 902
 903                         /* do not place post remats after jumps */
 904                         if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) continue;
 905
 906                         if(remat_info->remats_by_operand) {
 907                                 pset_foreach(remat_info->remats_by_operand, remat) {
 908                                         /* do not insert remats producing the same value as one of the operands */
 909                                         if(!pset_find_ptr(args, remat->value)) {
 910                                                 DBG((si->dbg, LEVEL_4, "\t  considering remat %+F with arg %+F\n", remat->op, arg));
 911 #ifdef REMAT_WHILE_LIVE
 912                                                 if(pset_find_ptr(live, remat->value)) {
 913                                                         insert_remat_after(si, remat, irn, live);
 914                                                 }
 915 #else
 916                                                 insert_remat_after(si, remat, irn, live);
 917 #endif
 918                                         }
 919                                 }
 920                         }
 921                 }
 922
 923                 /* delete defined value from live set */
 924                 if(has_reg_class(si, irn)) {
 925                         pset_remove_ptr(live, irn);
 926                 }
 927
 928                 /* insert all possible remats before irn */
 929                 pset_foreach(args, arg) {
 930                         remat_info_t   *remat_info,
 931                                                     query;
 932                         remat_t        *remat;
 933
 934                         /* continue if the operand has the wrong reg class
 935                          */
 936                         if(!has_reg_class(si, arg))
 937                                 continue;
 938
 939                         query.irn = arg;
 940                         query.remats = NULL;
 941                         query.remats_by_operand = NULL;
 942                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
 943
 944                         if(!remat_info) {
 945                                 continue;
 946                         }
 947
 948                         if(remat_info->remats) {
 949                                 pset_foreach(remat_info->remats, remat) {
 950                                         DBG((si->dbg, LEVEL_4, "\t  considering remat %+F for arg %+F\n", remat->op, arg));
 951 #ifdef REMAT_WHILE_LIVE
 952                                         if(pset_find_ptr(live, remat->value)) {
 953                                                 insert_remat_before(si, remat, irn, live);
 954                                         }
 955 #else
 956                                         insert_remat_before(si, remat, irn, live);
 957 #endif
 958                                 }
 959                         }
 960                 }
 961
 962                 del_pset(args);
 963                 irn = next;
 964         }
 965
 966         live_foreach(bb, li) {
 967                 ir_node        *value = (ir_node *) li->irn;
 968
 969                 /* add remats at end of block */
 970                 if (live_is_end(li) && has_reg_class(si, value)) {
 971                         remat_info_t   *remat_info,
 972                                                     query;
 973                         remat_t        *remat;
 974
 975                         query.irn = value;
 976                         query.remats = NULL;
 977                         query.remats_by_operand = NULL;
 978                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
 979
 980                         if(remat_info && remat_info->remats) {
 981                                 pset_foreach(remat_info->remats, remat) {
 982                                         DBG((si->dbg, LEVEL_4, "\t  considering remat %+F at end of block %+F\n", remat->op, bb));
 983
 984                                         insert_remat_before(si, remat, bb, NULL);
 985                                 }
 986                         }
 987                 }
 988
 989                 /* add remat2s at beginning of block */
 990                 if ((live_is_in(li) || (is_Phi(value) && get_nodes_block(value)==bb)) && has_reg_class(si, value)) {
 991                         remat_info_t   *remat_info,
 992                                                     query;
 993                         remat_t        *remat;
 994
 995                         query.irn = value;
 996                         query.remats = NULL;
 997                         query.remats_by_operand = NULL;
 998                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
 999
1000                         if(remat_info && remat_info->remats) {
1001                                 pset_foreach(remat_info->remats, remat) {
1002                                         DBG((si->dbg, LEVEL_4, "\t  considering remat %+F at beginning of block %+F\n", remat->op, bb));
1003
1004                                         /* put the remat here if all its args are available */
1005                                         insert_remat_after(si, remat, bb, NULL);
1006
1007 #if 0
1008                                         for(i=0, n=get_irn_arity(remat->op); i<n; ++i) {
1009                                                 ir_node   *remat_arg = get_irn_n(remat->op, i);
1010
1011                                                 if(has_reg_class(si, remat_arg) && is_live_in(bb, remat_arg)) {
1012                                                         insert_remat_after(si, remat, bb, NULL);
1013                                                         break;
1014                                                 }
1015                                         }
1016 #endif
1017
1018 #if 0
1019                         if(remat_info && remat_info->remats_by_operand) {
1020                                 pset_foreach(remat_info->remats_by_operand, remat) {
1021                                         DBG((si->dbg, LEVEL_4, "\t  considering remat %+F at beginning of block %+F\n", remat->op, bb));
1022 #ifdef REMAT_WHILE_LIVE
1023                                         if(is_live_in(bb, remat->value)) {
1024                                                 insert_remat_after(si, remat, bb, NULL);
1025                                         }
1026 #else
1027                                         insert_remat_after(si, remat, bb, NULL);
1028 #endif
1029 #endif
1030                                 }
1031                         }
1032                 }
1033         }
1034 }
1035
1036 /**
1037  * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1038  */
1039 static void
1040 luke_endwalker(ir_node * bb, void * data)
1041 {
1042         spill_ilp_t    *si = (spill_ilp_t*)data;
1043         irn_live_t     *li;
1044         pset           *live;
1045         pset           *use_end;
1046         char            buf[256];
1047         ilp_cst_t       cst;
1048         ir_node        *irn;
1049         spill_bb_t     *spill_bb = get_irn_link(bb);
1050
1051
1052         live = pset_new_ptr_default();
1053         use_end = pset_new_ptr_default();
1054
1055         live_foreach(bb, li) {
1056                 irn = (ir_node *) li->irn;
1057                 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1058                         op_t      *op;
1059
1060                         pset_insert_ptr(live, irn);
1061                         op = get_irn_link(irn);
1062                         assert(!op->is_remat);
1063                 }
1064         }
1065
1066         /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1067         /* their reg_out is unimportant because it can always be set */
1068         sched_foreach_reverse(bb, irn) {
1069                 int   i,
1070                           n;
1071
1072                 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1073
1074                 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
1075                         ir_node        *irn_arg = get_irn_n(irn, i);
1076                         if(has_reg_class(si, irn_arg)) {
1077                                 pset_insert_ptr(use_end, irn);
1078                         }
1079                 }
1080         }
1081
1082         ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1083         cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1084
1085         spill_bb->ilp = new_set(cmp_spill, 16);
1086
1087         live_foreach(bb, li) {
1088                 irn = (ir_node *) li->irn;
1089                 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1090                         spill_t     query,
1091                                    *spill;
1092
1093                         query.irn = irn;
1094                         spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1095
1096                         ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1097                         spill->reg_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1098                         /* if irn is used at the end of the block, then it is live anyway */
1099                         if(!pset_find_ptr(use_end, irn))
1100                                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1101
1102                         ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", bb, irn);
1103                         spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1104
1105                         ir_snprintf(buf, sizeof(buf), "spill_%N_%N", bb, irn);
1106                         spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, COST_STORE*execution_frequency(si, irn));
1107
1108                         spill->reg_in = ILP_UNDEF;
1109                         spill->mem_in = ILP_UNDEF;
1110                 }
1111         }
1112
1113         del_pset(live);
1114         del_pset(use_end);
1115 }
1116
1117 static ir_node *
1118 next_post_remat(const ir_node * irn)
1119 {
1120         op_t      *op;
1121
1122         if(is_Block(irn)) {
1123                 irn = sched_block_first_nonphi(irn);
1124         } else {
1125                 irn = sched_next_op(irn);
1126         }
1127
1128         if(sched_is_end(irn))
1129                 return NULL;
1130
1131         op = (op_t*)get_irn_link(irn);
1132         if(op->is_remat && !op->attr.remat.pre) {
1133                 return irn;
1134         }
1135
1136         return NULL;
1137 }
1138
1139
1140 static ir_node *
1141 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
1142 {
1143         op_t      *op;
1144         ir_node   *ret;
1145
1146         if(is_Block(irn)) {
1147                 ret = sched_block_last_noncf(si, irn);
1148                 ret = sched_next(ret);
1149                 ret = sched_prev_op(ret);
1150         } else {
1151                 ret = sched_prev_op(irn);
1152         }
1153
1154         if(sched_is_end(ret) || is_Phi(ret))
1155                 return NULL;
1156
1157         op = (op_t*)get_irn_link(ret);
1158         if(op->is_remat && op->attr.remat.pre) {
1159                 return ret;
1160         }
1161
1162         return NULL;
1163 }
1164
1165 /**
1166  * Find a remat of value @p value in the epilog of @p pos
1167  */
1168 static ir_node *
1169 find_post_remat(const ir_node * value, const ir_node * pos)
1170 {
1171         while((pos = next_post_remat(pos)) != NULL) {
1172                 op_t   *op;
1173
1174                 op = get_irn_link(pos);
1175                 assert(op->is_remat && !op->attr.remat.pre);
1176
1177                 if(op->attr.remat.remat->value == value)
1178                         return (ir_node*)pos;
1179
1180 #if 0
1181         const ir_edge_t *edge;
1182                 foreach_out_edge(pos, edge) {
1183                         ir_node   *proj = get_edge_src_irn(edge);
1184                         assert(is_Proj(proj));
1185                 }
1186 #endif
1187
1188         }
1189
1190         return NULL;
1191 }
1192
1193 /**
1194  * Find a remat of value @p value in the prolog of @p pos
1195  */
1196 static ir_node *
1197 find_pre_remat(const spill_ilp_t * si, const ir_node * value, const ir_node * pos)
1198 {
1199         while((pos = next_pre_remat(si,pos)) != NULL) {
1200                 op_t   *op;
1201
1202                 op = get_irn_link(pos);
1203                 assert(op->is_remat && op->attr.remat.pre);
1204
1205                 if(op->attr.remat.remat->value == value)
1206                         return (ir_node*)pos;
1207         }
1208
1209         return NULL;
1210 }
1211
1212 static spill_t *
1213 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1214 {
1215         spill_bb_t  *spill_bb = get_irn_link(bb);
1216         spill_t     *spill,
1217                                  query;
1218         char         buf[256];
1219
1220         query.irn = irn;
1221         spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1222         if(!spill) {
1223                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1224
1225                 spill->reg_out = ILP_UNDEF;
1226                 spill->reg_in  = ILP_UNDEF;
1227                 spill->mem_in  = ILP_UNDEF;
1228
1229                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1230                 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1231
1232                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1233                 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, COST_STORE*execution_frequency(si, bb));
1234         }
1235
1236         return spill;
1237 }
1238
1239 /**
1240  * Walk all irg blocks and emit this ILP
1241  */
1242 static void
1243 luke_blockwalker(ir_node * bb, void * data)
1244 {
1245         spill_ilp_t    *si = (spill_ilp_t*)data;
1246         ir_node        *irn;
1247         irn_live_t     *li;
1248         pset           *live;
1249         char            buf[256];
1250         ilp_cst_t       cst;
1251         spill_bb_t     *spill_bb = get_irn_link(bb);
1252         int             i;
1253         ir_node        *tmp;
1254         spill_t        *spill;
1255
1256
1257         live = pset_new_ptr_default();
1258
1259         /* do something at the end of the block */
1260
1261         /* init live values at end of block */
1262         live_foreach(bb, li) {
1263                 ir_node        *irn = (ir_node *) li->irn;
1264
1265                 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1266                         pset_insert_ptr(live, irn);
1267                 }
1268         }
1269
1270         spill_bb->reloads = obstack_alloc(si->obst, pset_count(live) * sizeof(*spill_bb->reloads));
1271         memset(spill_bb->reloads, 0xFF, pset_count(live) * sizeof(*spill_bb->reloads));
1272
1273         i=0;
1274         live_foreach(bb, li) {
1275                 ir_node        *irn = (ir_node *) li->irn;
1276                 op_t           *op;
1277
1278                 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1279                         spill = set_find_spill(spill_bb->ilp, irn);
1280                         assert(spill);
1281
1282                         ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1283                         spill_bb->reloads[i] = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
1284
1285                         /* reload <= mem_out */
1286                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1287                         lpp_set_factor_fast(si->lpp, cst, spill_bb->reloads[i], 1.0);
1288                         lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
1289
1290                         op = get_irn_link(irn);
1291                         assert(!op->is_remat);
1292
1293                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1294                         op->attr.live_range.ilp = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1295                         op->attr.live_range.op = bb;
1296
1297                         ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1298                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1299
1300                         /* reg_out - reload - remat - live_range <= 0 */
1301                         lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1302                         lpp_set_factor_fast(si->lpp, cst, spill_bb->reloads[i], -1.0);
1303                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1304                         foreach_pre_remat(si, bb, tmp) {
1305                                 op_t     *remat_op = get_irn_link(tmp);
1306                                 if(remat_op->attr.remat.remat->value == irn) {
1307                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1308                                 }
1309                         }
1310
1311                         ++i;
1312                 }
1313         }
1314         DBG((si->dbg, LEVEL_4, "\t   %d values live at end of block %+F\n", pset_count(live), bb));
1315
1316         sched_foreach_reverse(bb, irn) {
1317                 op_t       *op;
1318                 op_t       *tmp_op;
1319                 int         n,
1320                                         k,
1321                                         d = 0;
1322                 ilp_cst_t       check_pre,
1323                                         check_post;
1324 #ifdef CHECK_POST_REMAT
1325                 ilp_cst_t       check_post_remat;
1326 #endif
1327                 set        *args = new_set(cmp_keyval, get_irn_arity(irn));
1328                 keyval_t   *keyval;
1329
1330                 if(is_Phi(irn))
1331                         break;
1332
1333                 op = get_irn_link(irn);
1334                 /* skip remats */
1335                 if(op->is_remat) continue;
1336                 DBG((si->dbg, LEVEL_4, "\t  at node %+F\n", irn));
1337
1338                 if(has_reg_class(si, irn)) {
1339                         assert(pset_find_ptr(live, irn));
1340                         pset_remove_ptr(live, irn);
1341                 }
1342
1343                 /* init set of irn's arguments */
1344                 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
1345                         ir_node        *irn_arg = get_irn_n(irn, i);
1346                         if(has_reg_class(si, irn_arg)) {
1347                                 set_insert_keyval(args, irn_arg, (void*)i);
1348                         }
1349                 }
1350
1351 #ifdef CHECK_POST_REMAT
1352                 /* check the register pressure after the epilog */
1353                 ir_snprintf(buf, sizeof(buf), "check_post_remat_%N", irn);
1354                 check_post_remat = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
1355
1356                 /* iterate over L\U */
1357                 pset_foreach(live, tmp) {
1358                         if(!set_find_keyval(args, tmp)) {
1359                                 /* if a live value is not used by irn */
1360                                 tmp_op = get_irn_link(tmp);
1361 //                              assert(tmp_op->attr.live_range.op != irn);
1362                                 lpp_set_factor_fast(si->lpp, check_post_remat, tmp_op->attr.live_range.ilp, 1.0);
1363                         }
1364                 }
1365                 /* iterate over following remats and remove possibly defined values again from check_post_remat */
1366                 foreach_post_remat(irn, tmp) {
1367                         op_t           *remat_op = get_irn_link(tmp);
1368                         const ir_node  *value = remat_op->attr.remat.remat->value;
1369                         op_t           *val_op = get_irn_link(value);
1370
1371                         assert(remat_op->is_remat && !remat_op->attr.remat.pre);
1372
1373                         /* values that are defined by remat2s are not counted */
1374 #ifdef REMAT_WHILE_LIVE
1375                         assert(val_op->attr.live_range.ilp);
1376                         lpp_set_factor_fast(si->lpp, check_post_remat, val_op->attr.live_range.ilp, 0.0);
1377 #else
1378                         if(val_op->attr.live_range.ilp != ILP_UNDEF) {
1379                                 lpp_set_factor_fast(si->lpp, check_post_remat, val_op->attr.live_range.ilp, 0.0);
1380                         }
1381 #endif /* REMAT_WHILE_LIVE */
1382                 }
1383 #endif /* CHECK_POST_REMAT */
1384
1385
1386                 /* new live ranges for values from L\U defined by remat2s or used by remats */
1387                 pset_foreach(live, tmp) {
1388                         ir_node     *value = tmp;//remat_op->attr.remat.remat->value;
1389                         op_t        *value_op = get_irn_link(value);
1390
1391                         if(!set_find_keyval(args, value)) {
1392                                 ilp_var_t    prev_lr = ILP_UNDEF;
1393                                 ir_node     *remat;
1394                                 cst = ILP_UNDEF;
1395
1396                                 foreach_post_remat(irn, remat) {
1397                                         op_t        *remat_op = get_irn_link(remat);
1398
1399                                         /* if value is being rematerialized by this remat */
1400                                         if(value == remat_op->attr.remat.remat->value) {
1401                                                 if(cst == ILP_UNDEF) {
1402                                                         /* next_live_range <= prev_live_range + sum remat2s */
1403                                                         ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
1404                                                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1405                                                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
1406                                                         prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1407                                                         lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
1408                                                         lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1409                                                 }
1410
1411                                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1412                                         }
1413                                 }
1414
1415 #ifdef MAY_DIE_AT_PRE_REMAT
1416                                 if(cst == ILP_UNDEF) {
1417                                         foreach_pre_remat(si, irn, remat) {
1418                                                 int          i,
1419                                                                          n;
1420
1421                                                 for (i = 0, n = get_irn_arity(remat); i < n; ++i) {
1422                                                         ir_node        *remat_arg = get_irn_n(remat, i);
1423
1424                                                         /* if value is being used by this remat */
1425                                                         if(value == remat_arg) {
1426                                                                 /* next_live_range <= prev_live_range */
1427                                                                 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
1428                                                                 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1429
1430                                                                 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
1431                                                                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1432                                                                 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
1433                                                                 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1434                                                                 goto fertig;
1435                                                         }
1436                                                         /* TODO check afterwards whether lr dies after a pre-remat (should not happen) */
1437                                                 }
1438                                         }
1439                                 }
1440 #endif
1441
1442 fertig:
1443                                 if(prev_lr != ILP_UNDEF) {
1444                                         value_op->attr.live_range.ilp = prev_lr;
1445                                         value_op->attr.live_range.op = irn;
1446                                 }
1447                         }
1448                 }
1449
1450                 /* get count of values in my register class defined by irn */
1451                 /* also add defined values to check_post_remat; do this before iterating over args */
1452                 if(get_irn_mode(irn) == mode_T) {
1453                         ir_node  *proj = sched_next(irn);
1454                         op_t     *proj_op = get_irn_link(proj);
1455
1456                         while(is_Proj(proj)) {
1457                                 if(has_reg_class(si, proj)) {
1458                                         ++d;
1459 #ifdef CHECK_POST_REMAT
1460                                         lpp_set_factor_fast(si->lpp, check_post_remat, proj_op->attr.live_range.ilp, 1.0);
1461 #endif
1462                                 }
1463                                 proj = sched_next(proj);
1464                                 proj_op = get_irn_link(proj);
1465                         }
1466                 } else {
1467                         if(has_reg_class(si, irn)) {
1468                                  d = 1;
1469 #ifdef CHECK_POST_REMAT
1470                                  lpp_set_factor_fast(si->lpp, check_post_remat, op->attr.live_range.ilp, 1.0);
1471 #endif
1472                         }
1473                 }
1474                 DBG((si->dbg, LEVEL_4, "\t   %+F produces %d values in my register class\n", irn, d));
1475
1476                 /* count how many regs irn needs for arguments */
1477                 k = set_count(args);
1478
1479                 /* check the register pressure in the prolog */
1480                 /* sum_{L\U} lr <= n - |U| */
1481                 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
1482                 check_pre = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - k);
1483
1484                 /* check the register pressure in the epilog */
1485                 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
1486                 check_post = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - d);
1487
1488                 set_foreach(args, keyval) {
1489                         ilp_var_t       next_lr;
1490                         op_t           *arg_op;
1491                         ilp_var_t       post_use;
1492                         int             p = 0;
1493                         spill_t        *spill;
1494                         ir_node        *arg = keyval->key;
1495
1496                         spill = add_to_spill_bb(si, bb, arg);
1497
1498                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
1499                         next_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1500
1501                         i = (int)keyval->val;
1502                         assert(i<n);
1503
1504                         ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
1505                         op->attr.live_range.reloads[i] = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, irn));
1506
1507                         /* reload <= mem_out */
1508                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1509                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.reloads[i], 1.0);
1510                         lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
1511
1512                         arg_op = get_irn_link(arg);
1513
1514                         /* requirement: arg must be in register for use */
1515                         /* reload + remat + live_range == 1 */
1516                         ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
1517                         cst = lpp_add_cst(si->lpp, buf, lpp_equal, 1.0);
1518
1519                         lpp_set_factor_fast(si->lpp, cst, next_lr, 1.0);
1520                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.reloads[i], 1.0);
1521                         foreach_pre_remat(si, irn, tmp) {
1522                                 op_t     *remat_op = get_irn_link(tmp);
1523                                 if(remat_op->attr.remat.remat->value == arg) {
1524                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1525                                 }
1526                         }
1527
1528                         /* the epilog stuff - including post_use, post, post_remat */
1529                         ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
1530                         post_use = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1531
1532                         lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
1533
1534                         /* arg is live throughout epilog if the next live_range is in a register */
1535                         if(pset_find_ptr(live, arg)) {
1536                                 DBG((si->dbg, LEVEL_3, "\t  arg %+F is possibly live in epilog of %+F\n", arg, irn));
1537
1538                                 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1539                                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1540                                 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
1541                                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1542
1543 #ifdef CHECK_POST_REMAT
1544                                 lpp_set_factor_fast(si->lpp, check_post_remat, arg_op->attr.live_range.ilp, 1.0);
1545 #endif
1546                         }
1547
1548                         /*forall remat2 which use arg add a similar cst*/
1549                         foreach_post_remat(irn, tmp) {
1550                                 int         i,
1551                                                         n;
1552
1553                                 for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1554                                         ir_node    *remat_arg = get_irn_n(tmp, i);
1555                                         op_t       *remat_op = get_irn_link(tmp);
1556
1557                                         if(remat_arg == arg) {
1558                                                 DBG((si->dbg, LEVEL_3, "\t  found remat with arg %+F in epilog of %+F\n", arg, irn));
1559
1560                                                 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1561                                                 cst = lpp_add_cst(si->lpp, buf, lpp_greater, 0.0);
1562                                                 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
1563                                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1564                                         }
1565                                 }
1566                         }
1567
1568                         /* new live range begins for each argument */
1569                         arg_op->attr.live_range.ilp = next_lr;
1570                         arg_op->attr.live_range.op = irn;
1571
1572                         pset_insert_ptr(live, arg);
1573                 }
1574
1575                 /* start new live ranges for values used by remats */
1576                 foreach_pre_remat(si, irn, tmp) {
1577                         int          i,
1578                                                  n;
1579
1580                         for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1581                                 ir_node        *remat_arg = get_irn_n(tmp, i);
1582                                 op_t           *arg_op = get_irn_link(remat_arg);
1583                                 ilp_var_t       prev_lr;
1584
1585                                 if(!has_reg_class(si, remat_arg)) continue;
1586
1587                                 /* if value is becoming live through use by remat */
1588                                 if(!pset_find_ptr(live, remat_arg)) {
1589                                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
1590                                         prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1591
1592                                         arg_op->attr.live_range.ilp = prev_lr;
1593                                         arg_op->attr.live_range.op = irn;
1594
1595                                         DBG((si->dbg, LEVEL_4, "  value %+F becoming live through use by remat %+F\n", remat_arg, tmp));
1596
1597                                         /* TODO ist das hier die richtige Stelle???? */
1598                                         pset_insert_ptr(live, remat_arg);
1599                                         add_to_spill_bb(si, bb, remat_arg);
1600                                 }
1601                                 /* TODO check afterwards whether lr dies after a pre-remat (should not happen) */
1602                         }
1603                 }
1604
1605                 /* iterate over L\U */
1606                 pset_foreach(live, tmp) {
1607                         if(!set_find_keyval(args, tmp)) {
1608                                 /* if a live value is not used by irn */
1609                                 tmp_op = get_irn_link(tmp);
1610 //                              assert(tmp_op->attr.live_range.op != irn);
1611                                 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
1612                                 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
1613                         }
1614                 }
1615
1616                 /* requirements for remats */
1617                 foreach_pre_remat(si, irn, tmp) {
1618                         op_t        *remat_op = get_irn_link(tmp);
1619                         int          i,
1620                                                  n;
1621
1622                         for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1623                                 ir_node        *remat_arg = get_irn_n(tmp, i);
1624                                 op_t           *arg_op = get_irn_link(remat_arg);
1625
1626                                 if(!has_reg_class(si, remat_arg)) continue;
1627
1628                                 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
1629                                 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1630                                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1631
1632                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1633                                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1634
1635                                 /* if remat arg is also used by current op then we can use reload placed for this argument */
1636                                 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
1637                                         int    index = (int)keyval->val;
1638
1639                                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.reloads[index], -1.0);
1640                                 }
1641                         }
1642                 }
1643
1644                 /* requirements for remats2
1645                  *
1646                  *  TODO unsure if this does the right thing.
1647                  *  should insert values into set if they do not become live through remat and
1648                  *  op
1649                  */
1650                 foreach_post_remat(irn, tmp) {
1651                         op_t        *remat_op = get_irn_link(tmp);
1652                         int          i,
1653                                                  n;
1654
1655                         for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1656                                 ir_node        *remat_arg = get_irn_n(tmp, i);
1657                                 op_t           *arg_op = get_irn_link(remat_arg);
1658
1659                                 if(!has_reg_class(si, remat_arg)) continue;
1660
1661                                 /* only for values in L\U, the others are handled with post_use */
1662                                 if(!set_find_keyval(args, remat_arg)) {
1663                                         /* remat <= live_rang(remat_arg) */
1664                                         ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
1665                                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1666
1667                                         /* if value is becoming live through use by remat2 */
1668                                         if(!pset_find_ptr(live, remat_arg)) {
1669                                                 ilp_var_t     lr;
1670
1671                                                 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
1672                                                 lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1673
1674                                                 arg_op->attr.live_range.ilp = lr;
1675                                                 arg_op->attr.live_range.op = irn;
1676
1677                                                 DBG((si->dbg, LEVEL_3, "  value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
1678
1679                                                 pset_insert_ptr(live, remat_arg);
1680                                                 add_to_spill_bb(si, bb, remat_arg);
1681                                         }
1682
1683                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1684                                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1685                                 }
1686                         }
1687                 }
1688
1689 #ifdef CHECK_POST_REMAT
1690                 /* iterate over following remats and add them to check_post_remat */
1691                 foreach_post_remat(irn, tmp) {
1692                         op_t           *remat_op = get_irn_link(tmp);
1693
1694                         assert(remat_op->is_remat && !remat_op->attr.remat.pre);
1695
1696                         lpp_set_factor_fast(si->lpp, check_post_remat, remat_op->attr.remat.ilp, 1.0);
1697                 }
1698 #endif
1699
1700
1701
1702                 DBG((si->dbg, LEVEL_4, "\t   %d values live at %+F\n", pset_count(live), irn));
1703
1704                 pset_foreach(live, tmp) {
1705                         assert(has_reg_class(si, tmp));
1706                 }
1707
1708                 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
1709                         ir_node        *arg = get_irn_n(irn, i);
1710
1711                         assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
1712                 }
1713
1714                 del_set(args);
1715         }
1716
1717
1718
1719         /* do something at the beginning of the block */
1720
1721         /* we are now at the beginning of the basic block, there are only \Phis in front of us */
1722         DBG((si->dbg, LEVEL_3, "\t   %d values live at beginning of block %+F\n", pset_count(live), bb));
1723
1724         pset_foreach(live, irn) {
1725                 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
1726         }
1727
1728         /* construct mem_outs for all values */
1729
1730         set_foreach(spill_bb->ilp, spill) {
1731                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
1732                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1733
1734                 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
1735                 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
1736
1737                 if(pset_find_ptr(live, spill->irn)) {
1738                         DBG((si->dbg, LEVEL_5, "\t     %+F live at beginning of block %+F\n", spill->irn, bb));
1739
1740                         ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
1741                         spill->mem_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1742
1743                         lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
1744                 }
1745         }
1746
1747
1748         /* L\U is empty at bb start */
1749         /* arg is live throughout epilog if it is reg_in into this block */
1750
1751         /* check the register pressure at the beginning of the block
1752          * including remats
1753          */
1754         ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
1755         cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
1756
1757         pset_foreach(live, irn) {
1758                         spill = set_find_spill(spill_bb->ilp, irn);
1759                         assert(spill);
1760
1761                         ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
1762                         spill->reg_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1763
1764                         lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
1765         }
1766         foreach_post_remat(bb, irn) {
1767                 op_t     *remat_op = get_irn_link(irn);
1768
1769                 DBG((si->dbg, LEVEL_4, "\t  next post remat: %+F\n", irn));
1770                 assert(remat_op->is_remat && !remat_op->attr.remat.pre);
1771
1772                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1773         }
1774
1775         /* forall remat2 add requirements */
1776         foreach_post_remat(bb, tmp) {
1777                 int         i,
1778                                         n;
1779
1780                 for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1781                         ir_node    *remat_arg = get_irn_n(tmp, i);
1782                         op_t       *remat_op = get_irn_link(tmp);
1783
1784                         if(!has_reg_class(si, remat_arg)) continue;
1785
1786                         spill = set_find_spill(spill_bb->ilp, remat_arg);
1787                         assert(spill);
1788
1789                         /* TODO verify this is placed correctly */
1790                         ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
1791                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1792                         lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
1793                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1794                 }
1795         }
1796
1797         /* mem_in/reg_in for live_in values, especially phis and their arguments */
1798         pset_foreach(live, irn) {
1799                 int          p = 0,
1800                                          i,
1801                                          n;
1802
1803                 spill = set_find_spill(spill_bb->ilp, irn);
1804                 assert(spill && spill->irn == irn);
1805
1806                 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
1807                         for (i = 0, n = get_Phi_n_preds(irn); i < n; ++i) {
1808                                 ilp_cst_t       mem_in,
1809                                                                 reg_in;
1810                                 ir_node        *phi_arg = get_Phi_pred(irn, i);
1811                                 ir_node        *bb_p = get_Block_cfgpred_block(bb, i);
1812                                 spill_bb_t     *spill_bb_p = get_irn_link(bb_p);
1813                                 spill_t        *spill_p;
1814
1815                                 /* although the phi is in the right regclass one or more of
1816                                  * its arguments can be in a different one or at least to
1817                                  * ignore
1818                                  */
1819                                 if(has_reg_class(si, phi_arg)) {
1820                                         ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
1821                                         mem_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1822                                         ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
1823                                         reg_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1824
1825                                         lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
1826                                         lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
1827
1828                                         spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
1829                                         assert(spill_p);
1830
1831                                         lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
1832                                         lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
1833                                 }
1834                         }
1835                 } else {
1836                         /* else assure the value arrives on all paths in the same resource */
1837
1838                         for (i = 0, n = get_Block_n_cfgpreds(bb); i < n; ++i) {
1839                                 ilp_cst_t       mem_in,
1840                                                                 reg_in;
1841                                 ir_node        *bb_p = get_Block_cfgpred_block(bb, i);
1842                                 spill_bb_t     *spill_bb_p = get_irn_link(bb_p);
1843                                 spill_t        *spill_p;
1844
1845                                 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
1846                                 mem_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1847                                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
1848                                 reg_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1849
1850                                 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
1851                                 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
1852
1853                                 spill_p = set_find_spill(spill_bb_p->ilp, irn);
1854                                 assert(spill_p);
1855
1856                                 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
1857                                 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
1858                         }
1859                 }
1860         }
1861
1862         /* first live ranges from reg_ins */
1863         pset_foreach(live, irn) {
1864                 op_t      *op = get_irn_link(irn);
1865
1866                 spill = set_find_spill(spill_bb->ilp, irn);
1867                 assert(spill && spill->irn == irn);
1868
1869                 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
1870                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1871                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1872                 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
1873
1874                 foreach_post_remat(bb, tmp) {
1875                         op_t     *remat_op = get_irn_link(tmp);
1876
1877                         if(remat_op->attr.remat.remat->value == irn) {
1878                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1879                         }
1880                 }
1881         }
1882
1883         /* walk forward now and compute constraints for placing spills */
1884         /* this must only be done for values that are not defined in this block */
1885         pset_foreach(live, irn) {
1886                 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", spill->irn, bb);
1887                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1888
1889                 spill = set_find_spill(spill_bb->ilp, irn);
1890                 assert(spill);
1891
1892                 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
1893                 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
1894
1895                 sched_foreach_op(bb, tmp) {
1896                         op_t   *op = get_irn_link(tmp);
1897
1898                         if(is_Phi(tmp)) continue;
1899                         assert(!is_Proj(tmp));
1900
1901                         if(op->is_remat) {
1902                                 ir_node   *value = op->attr.remat.remat->value;
1903
1904                                 if(value == irn) {
1905                                         /* only collect remats up to the first use of a value */
1906                                         lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
1907                                 }
1908                         } else {
1909                                 int i,
1910                                         n;
1911
1912                                 for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1913                                         ir_node    *arg = get_irn_n(tmp, i);
1914
1915                                         if(arg == irn) {
1916                                                 /* if a value is used stop collecting remats */
1917                                                 cst = ILP_UNDEF;
1918                                         }
1919                                         break;
1920                                 }
1921                         }
1922                         if(cst == ILP_UNDEF) break;
1923                 }
1924         }
1925
1926
1927         /* if a value is used by a mem-phi, then mem_in of this value is 0 (has to be spilled again into a different slot)
1928            mem_in(phi) -> not mem_in(orig_value) TODO: how does this depend on a certain predecessor?
1929          */
1930
1931         /* mem_in of mem-phi has associated costs (but first one is free) */
1932         /* define n_mem_copies as positive integer in each predecessor block,
1933            #mem_in into this block from predecessor block - 1 weighted with SPILL_COST*execfreq(predecessor)
1934            TODO
1935          */
1936
1937
1938         del_pset(live);
1939 }
1940
1941
1942 #if 0
1943          * Speicherkopienminimierung: teste Speicherwerte auf Interferenz
1944          * und weise Spillkontexte zu. Sorge bei Phis dafuer, dass gleiche
1945          * Kontexte zusammenfliessen (Operanden und Ergebnis hat gleichen
1946          * Kontext)
1947 #endif
1948
1949 static INLINE int
1950 is_zero(double x)
1951 {
1952         return fabs(x) < 0.00001;
1953 }
1954
1955 #if 0
1956 static int
1957 is_spilled(const spill_ilp_t * si, const live_range_t * lr)
1958 {
1959         return !is_zero(lpp_get_var_sol(si->lpp, lr->in_mem_var));
1960 }
1961 #endif
1962
1963 static int
1964 is_mem_phi(const ir_node * phi, void *data)
1965 {
1966         spill_ilp_t    *si = data;
1967 //      return is_spilled(si, get_use_head(si, phi)->closest_use);
1968         return 0;
1969 }
1970
1971 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
1972 {
1973         spill_ilp_t *si = get_irg_link(current_ir_graph);
1974
1975         if(pset_find_ptr(si->all_possible_remats, n)) {
1976                 op_t   *op = (op_t*)get_irn_link(n);
1977                 assert(op && op->is_remat);
1978
1979                 if(!op->attr.remat.remat->inverse) {
1980                         if(op->attr.remat.pre) {
1981                                 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
1982                         } else {
1983                                 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
1984                         }
1985
1986                         return 1;
1987                 } else {
1988                         op_t   *op = (op_t*)get_irn_link(n);
1989                         assert(op && op->is_remat);
1990
1991                         if(op->attr.remat.pre) {
1992                                 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
1993                         } else {
1994                                 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
1995                         }
1996
1997                         return 1;
1998                 }
1999         }
2000
2001         return 0;
2002 }
2003
2004 static void
2005 dump_graph_with_remats(ir_graph * irg, const char * suffix)
2006 {
2007         set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
2008         be_dump(irg, suffix, dump_ir_block_graph_sched);
2009         set_dump_node_vcgattr_hook(NULL);
2010 }
2011
2012 /**
2013  * Edge hook to dump the schedule edges with annotated register pressure.
2014  */
2015 static int
2016 sched_pressure_edge_hook(FILE *F, ir_node *irn)
2017 {
2018         if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
2019                 ir_node *prev = sched_prev(irn);
2020                 fprintf(F, "edge:{sourcename:\"");
2021                 PRINT_NODEID(irn);
2022                 fprintf(F, "\" targetname:\"");
2023                 PRINT_NODEID(prev);
2024                 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
2025                 fprintf(F, "\" color:magenta}\n");
2026         }
2027         return 1;
2028 }
2029
2030 static void
2031 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
2032 {
2033         DUMP_NODE_EDGE_FUNC old = get_dump_node_edge_hook();
2034
2035         dump_consts_local(0);
2036         set_dump_node_edge_hook(sched_pressure_edge_hook);
2037         dump_ir_block_graph(irg, suffix);
2038         set_dump_node_edge_hook(old);
2039 }
2040
2041 static void
2042 walker_pressure_annotator(ir_node * bb, void * data)
2043 {
2044         spill_ilp_t  *si = data;
2045         ir_node      *irn;
2046         irn_live_t   *li;
2047         int           i,
2048                                   n;
2049         pset         *live = pset_new_ptr_default();
2050         int           projs = 0;
2051
2052         live_foreach(bb, li) {
2053                 irn = (ir_node *) li->irn;
2054
2055                 if (live_is_end(li) && has_reg_class(si, irn)) {
2056                         pset_insert_ptr(live, irn);
2057                 }
2058         }
2059
2060         set_irn_link(bb, INT_TO_PTR(pset_count(live)));
2061
2062         sched_foreach_reverse(bb, irn) {
2063                 if(is_Phi(irn)) {
2064                         set_irn_link(irn, INT_TO_PTR(pset_count(live)));
2065                         continue;
2066                 }
2067
2068                 if(has_reg_class(si, irn)) {
2069                         pset_remove_ptr(live, irn);
2070                         if(is_Proj(irn)) ++projs;
2071                 }
2072
2073                 if(!is_Proj(irn)) projs = 0;
2074
2075                 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
2076                         ir_node    *arg = get_irn_n(irn, i);
2077
2078                         if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
2079                 }
2080                 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
2081         }
2082
2083         del_pset(live);
2084 }
2085
2086 static void
2087 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
2088 {
2089         be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
2090 }
2091
2092 #ifdef KEEPALIVE
2093 static void
2094 connect_all_remats_with_keep(spill_ilp_t * si)
2095 {
2096         ir_node   *irn;
2097         ir_node  **ins,
2098                          **pos;
2099         int        n_remats;
2100
2101
2102         n_remats = pset_count(si->all_possible_remats);
2103         if(n_remats) {
2104                 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
2105
2106                 pos = ins;
2107                 pset_foreach(si->all_possible_remats, irn) {
2108                         *pos = irn;
2109                         ++pos;
2110                 }
2111
2112                 si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins);
2113
2114                 obstack_free(si->obst, ins);
2115         }
2116 }
2117 #endif
2118
2119 /** insert a spill at an arbitrary position */
2120 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert, ir_node *ctx)
2121 {
2122         ir_node *bl     = is_Block(insert)?insert:get_nodes_block(insert);
2123         ir_graph *irg   = get_irn_irg(bl);
2124         ir_node *frame  = get_irg_frame(irg);
2125         ir_node *spill;
2126         ir_node *next;
2127
2128         const arch_register_class_t *cls       = arch_get_irn_reg_class(arch_env, irn, -1);
2129         const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
2130
2131         spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn, ctx);
2132
2133         /*
2134          * search the right insertion point. a spill of a phi cannot be put
2135          * directly after the phi, if there are some phis behind the one which
2136          * is spilled. Also, a spill of a Proj must be after all Projs of the
2137          * same tuple node.
2138          *
2139          * Here's one special case:
2140          * If the spill is in the start block, the spill must be after the frame
2141          * pointer is set up. This is done by setting insert to the end of the block
2142          * which is its default initialization (see above).
2143          */
2144
2145         if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
2146                 insert = frame;
2147
2148         for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
2149                 insert = next;
2150
2151         sched_add_after(insert, spill);
2152         return spill;
2153 }
2154
2155 static void
2156 delete_remat(spill_ilp_t * si, ir_node * remat) {
2157         int       i,
2158                       n;
2159         ir_node  *bad = get_irg_bad(si->chordal_env->irg);
2160
2161         sched_remove(remat);
2162
2163         /* kill links to operands */
2164         for (i = -1, n = get_irn_arity(remat); i < n; ++i) {
2165                 set_irn_n(remat, i, bad);
2166         }
2167 }
2168
2169 static void
2170 clean_remat_info(spill_ilp_t * si)
2171 {
2172         int            i,
2173                                n;
2174         remat_t       *remat;
2175         remat_info_t  *remat_info;
2176         ir_node       *bad = get_irg_bad(si->chordal_env->irg);
2177
2178         set_foreach(si->remat_info, remat_info) {
2179                 if(!remat_info->remats) continue;
2180
2181                 pset_foreach(remat_info->remats, remat)
2182                 {
2183                         if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
2184                                 set_irn_n(remat->proj, -1, bad);
2185                                 set_irn_n(remat->proj, 0, bad);
2186                         }
2187
2188                         if(get_irn_n_edges(remat->op) == 0) {
2189                                 for (i = -1, n = get_irn_arity(remat->op); i < n; ++i) {
2190                                         set_irn_n(remat->op, i, bad);
2191                                 }
2192                         }
2193                 }
2194
2195                 if(remat_info->remats) del_pset(remat_info->remats);
2196                 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
2197         }
2198 }
2199
2200 static void
2201 delete_unnecessary_remats(spill_ilp_t * si)
2202 {
2203         int       i,
2204                       n;
2205         ir_node  *bad = get_irg_bad(si->chordal_env->irg);
2206
2207 #ifdef KEEPALIVE
2208         if(si->keep) {
2209                 ir_node   *end = get_irg_end(si->chordal_env->irg);
2210                 ir_node  **keeps;
2211
2212                 for (i = 0, n = get_irn_arity(si->keep); i < n; ++i) {
2213                         ir_node        *keep_arg = get_irn_n(si->keep, i);
2214                         op_t           *arg_op = get_irn_link(keep_arg);
2215                         lpp_name_t     *name;
2216
2217                         assert(arg_op->is_remat);
2218
2219                         name = si->lpp->vars[arg_op->attr.remat.ilp];
2220
2221                         if(is_zero(name->value)) {
2222                                 DBG((si->dbg, LEVEL_3, "\t  deleting remat %+F\n", keep_arg));
2223                                 /* TODO check whether reload is preferred over remat (could be bug) */
2224                                 delete_remat(si, keep_arg);
2225                         } else {
2226                                 if(!arg_op->attr.remat.remat->inverse) {
2227                                         if(arg_op->attr.remat.pre) {
2228                                                 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
2229                                         } else {
2230                                                 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
2231                                         }
2232                                 } else {
2233                                         if(arg_op->attr.remat.pre) {
2234                                                 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
2235                                         } else {
2236                                                 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
2237                                         }
2238                                 }
2239                         }
2240
2241                         set_irn_n(si->keep, i, bad);
2242                 }
2243 #if 0
2244                 for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
2245                         ir_node        *end_arg = get_End_keepalive(end, i);
2246
2247                         if(end_arg != si->keep) {
2248                                 obstack_grow(si->obst, &end_arg, sizeof(end_arg));
2249                         }
2250                 }
2251                 keeps = obstack_finish(si->obst);
2252                 set_End_keepalives(end, n-1, keeps);
2253                 obstack_free(si->obst, keeps);
2254 #endif
2255         } else {
2256                 DBG((si->dbg, LEVEL_2, "\t  no remats to delete (none have been inserted)\n"));
2257         }
2258 #else
2259         ir_node  *remat;
2260
2261         pset_foreach(si->all_possible_remats, remat) {
2262                 op_t           *remat_op = get_irn_link(remat);
2263                 lpp_name_t     *name = si->lpp->vars[remat_op->attr.remat.ilp];
2264
2265                 if(is_zero(name->value)) {
2266                         DBG((si->dbg, LEVEL_3, "\t  deleting remat %+F\n", remat));
2267                         /* TODO check whether reload is preferred over remat (could be bug) */
2268                         delete_remat(si, remat);
2269                 } else {
2270                         if(remat_op->attr.remat.pre) {
2271                                 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
2272                         } else {
2273                                 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
2274                         }
2275                 }
2276         }
2277 #endif
2278 }
2279
2280 /**
2281  * @param before   The node after which the spill will be placed in the schedule
2282  */
2283 /* TODO set context properly */
2284 static ir_node *
2285 insert_spill(spill_ilp_t * si, const ir_node * irn, const ir_node * value, const ir_node * before)
2286 {
2287         defs_t   *defs;
2288         ir_node  *spill;
2289         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
2290
2291         DBG((si->dbg, LEVEL_3, "\t  inserting spill for value %+F after %+F\n", irn, before));
2292
2293         spill = be_spill2(arch_env, irn, before, irn);
2294
2295         defs = set_insert_def(si->values, value);
2296         assert(defs);
2297
2298         /* enter into the linked list */
2299         set_irn_link(spill, defs->spills);
2300         defs->spills = spill;
2301
2302 #ifdef KEEPALIVE_SPILLS
2303         keep_alive(spill);
2304 #endif
2305
2306         return spill;
2307 }
2308
2309
2310 /**
2311  * Add remat to list of defs, destroys link field!
2312  */
2313 static void
2314 insert_remat(spill_ilp_t * si, ir_node * remat)
2315 {
2316         defs_t   *defs;
2317         op_t     *remat_op = get_irn_link(remat);
2318
2319         assert(remat_op->is_remat);
2320
2321         defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
2322         assert(defs);
2323
2324         /* enter into the linked list */
2325         set_irn_link(remat, defs->remats);
2326         defs->remats = remat;
2327 }
2328
2329 static void
2330 collect_spills(spill_ilp_t * si, ir_node * value, pset * spills, pset * visited)
2331 {
2332         ir_node  *next;
2333         defs_t   *defs;
2334
2335         defs = set_find_def(si->values, value);
2336
2337         if(defs && defs->spills) {
2338                 for(next = defs->spills; next; next = get_irn_link(next)) {
2339                         pset_insert_ptr(spills, next);
2340                 }
2341         } else if (is_Phi(value)) {
2342                 /* recursion */
2343                 if(!pset_find_ptr(visited, value)) {
2344                         int    i,
2345                                    n;
2346
2347                         pset_insert_ptr(visited, value);
2348                         for(i=0, n=get_irn_arity(value); i<n; ++i) {
2349                                 ir_node    *arg = get_irn_n(value, i);
2350
2351                                 collect_spills(si, arg, spills, visited);
2352                         }
2353                 }
2354         } else {
2355 //              assert(0 && "Phi operand not spilled");
2356         }
2357 }
2358
2359 static pset *
2360 get_spills_for_value(spill_ilp_t * si, ir_node * value)
2361 {
2362         pset     *spills = pset_new_ptr_default();
2363         pset     *visited = pset_new_ptr_default();
2364
2365         collect_spills(si, value, spills, visited);
2366         del_pset(visited);
2367
2368         return spills;
2369 }
2370
2371 /**
2372  * Add reload before operation and add to list of defs
2373  */
2374 static ir_node *
2375 insert_reload(spill_ilp_t * si, const ir_node * value, const ir_node * after)
2376 {
2377         defs_t   *defs;
2378         ir_node  *reload,
2379                          *spill;
2380         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
2381
2382         DBG((si->dbg, LEVEL_3, "\t  inserting reload for value %+F before %+F\n", value, after));
2383
2384         defs = set_find_def(si->values, value);
2385         /* get a spill of this value */
2386         if((!defs || !defs->spills) && is_Phi(value)) {
2387                 pset  *spills;
2388
2389                 spills = get_spills_for_value(si, value);
2390
2391                 spill = pset_first(spills);
2392                 del_pset(spills);
2393
2394                 if(!defs) {
2395                         defs = set_insert_def(si->values, value);
2396                 }
2397         } else {
2398                 spill = defs->spills;
2399         }
2400         assert(spill && "no spill placed before reload");
2401
2402         reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
2403
2404         /* enter into the linked list */
2405         set_irn_link(reload, defs->remats);
2406         defs->remats = reload;
2407
2408         return reload;
2409
2410 }
2411
2412 static void
2413 walker_spill_placer(ir_node * bb, void * data) {
2414         spill_ilp_t   *si = (spill_ilp_t*)data;
2415         ir_node       *irn;
2416         spill_bb_t    *spill_bb = get_irn_link(bb);
2417         pset          *spills_to_do = pset_new_ptr_default();
2418         spill_t       *spill;
2419
2420         set_foreach(spill_bb->ilp, spill) {
2421                 lpp_name_t    *name;
2422
2423                 assert(spill->spill > 0);
2424
2425                 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2426                         name = si->lpp->vars[spill->mem_in];
2427                         if(!is_zero(name->value)) {
2428                                 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F\n", spill->irn));
2429                         }
2430                 }
2431
2432                 name = si->lpp->vars[spill->spill];
2433                 if(!is_zero(name->value)) {
2434                         if(spill->reg_in > 0) {
2435                                 name = si->lpp->vars[spill->reg_in];
2436                                 if(!is_zero(name->value)) {
2437                                         insert_spill(si, spill->irn, spill->irn, bb);
2438                                         continue;
2439                                 }
2440                         }
2441                         pset_insert_ptr(spills_to_do, spill->irn);
2442                 }
2443         }
2444         DBG((si->dbg, LEVEL_3, "\t  %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
2445
2446
2447         for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
2448                 op_t     *op = get_irn_link(irn);
2449
2450                 if(be_is_Spill(irn)) continue;
2451
2452                 if(op->is_remat) {
2453                         /* TODO fix this if we want to support remats with more than two nodes */
2454                         if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
2455                                 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
2456
2457                                 insert_spill(si, irn, op->attr.remat.remat->value, irn);
2458                         }
2459                 } else {
2460                         if(pset_find_ptr(spills_to_do, irn)) {
2461                                 pset_remove_ptr(spills_to_do, irn);
2462
2463                                 insert_spill(si, irn, irn, irn);
2464                         }
2465                 }
2466
2467         }
2468
2469         assert(pset_count(spills_to_do) == 0);
2470
2471         /* afterwards free data in block */
2472         del_pset(spills_to_do);
2473 }
2474
2475 static void
2476 walker_reload_placer(ir_node * bb, void * data) {
2477         spill_ilp_t   *si = (spill_ilp_t*)data;
2478         ir_node       *irn;
2479         spill_bb_t    *spill_bb = get_irn_link(bb);
2480         int            i;
2481         irn_live_t    *li;
2482
2483         sched_foreach_reverse(bb, irn) {
2484                 op_t     *op = get_irn_link(irn);
2485
2486                 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
2487                 if(is_Phi(irn)) break;
2488
2489                 if(op->is_remat) {
2490                         if(get_irn_mode(irn) != mode_T) {
2491                                 insert_remat(si, irn);
2492                         }
2493                 } else {
2494                         int    n;
2495
2496                         for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
2497                                 ir_node    *arg = get_irn_n(irn, i);
2498
2499                                 if(op->attr.live_range.reloads && op->attr.live_range.reloads[i] != ILP_UNDEF) {
2500                                         lpp_name_t    *name;
2501
2502                                         name = si->lpp->vars[op->attr.live_range.reloads[i]];
2503                                         if(!is_zero(name->value)) {
2504                                                 ir_node    *reload;
2505                                                 ir_node    *insert_pos = irn;
2506                                                 ir_node    *prev = sched_prev(insert_pos);
2507                                                 op_t       *prev_op = get_irn_link(prev);
2508
2509                                                 /* insert reload before pre-remats */
2510                                                 while(!sched_is_end(prev) && !be_is_Reload(prev) && !be_is_Spill(prev)
2511                                                                 && prev_op->is_remat && prev_op->attr.remat.pre) {
2512                                                         insert_pos = prev;
2513
2514                                                         prev = sched_prev(insert_pos);
2515                                                         prev_op = get_irn_link(prev);
2516                                                 }
2517
2518                                                 reload = insert_reload(si, arg, insert_pos);
2519
2520                                                 set_irn_n(irn, i, reload);
2521
2522
2523 #ifdef KEEPALIVE_SPILLS
2524                                                 keep_alive(reload);
2525 #endif
2526                                         }
2527                                 }
2528                         }
2529                 }
2530         }
2531
2532         i=0;
2533         live_foreach(bb, li) {
2534                 ir_node        *irn = (ir_node *) li->irn;
2535
2536                 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
2537                         lpp_name_t    *name;
2538
2539                         name = si->lpp->vars[spill_bb->reloads[i]];
2540                         if(!is_zero(name->value)) {
2541                                 ir_node    *reload;
2542                                 ir_node    *insert_pos = bb;
2543                                 ir_node    *prev = sched_prev(insert_pos);
2544                                 op_t       *prev_op = get_irn_link(prev);
2545
2546                                 /* insert reload before pre-remats */
2547                                 while(!sched_is_end(prev) && !be_is_Reload(prev) && !be_is_Spill(prev)
2548                                                 && prev_op->is_remat && prev_op->attr.remat.pre) {
2549                                         insert_pos = prev;
2550
2551                                         prev = sched_prev(insert_pos);
2552                                         prev_op = get_irn_link(prev);
2553                                 }
2554
2555                                 reload = insert_reload(si, irn, insert_pos);
2556
2557 #ifdef KEEPALIVE_SPILLS
2558                                 keep_alive(reload);
2559 #endif
2560                         }
2561                         ++i;
2562                 }
2563         }
2564
2565         del_set(spill_bb->ilp);
2566 }
2567
2568 static void
2569 walker_collect_used(ir_node * irn, void * data)
2570 {
2571         lc_bitset_t   *used = data;
2572
2573         lc_bitset_set(used, get_irn_idx(irn));
2574 }
2575
2576 static void
2577 walker_kill_unused(ir_node * bb, void * data)
2578 {
2579         lc_bitset_t     *used = data;
2580         const ir_node   *bad = get_irg_bad(get_irn_irg(bb));
2581         ir_node         *irn;
2582
2583
2584         for(irn=sched_first(bb); !sched_is_end(irn);) {
2585                 ir_node     *next = sched_next(irn);
2586                 int          i,
2587                                          n;
2588
2589                 if(!lc_bitset_is_set(used, get_irn_idx(irn))) {
2590                         assert(!be_is_Spill(irn) && !be_is_Reload(irn) && "something is fishy, spill or remat is unused");
2591
2592                         sched_remove(irn);
2593
2594                         set_nodes_block(irn, bad);
2595                         for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
2596                                 ir_node             *arg = get_irn_n(irn, i);
2597
2598                                 set_irn_n(irn, i, bad);
2599                         }
2600                 }
2601                 irn = next;
2602         }
2603 }
2604
2605 static void
2606 kill_all_unused_values_in_schedule(spill_ilp_t * si)
2607 {
2608         lc_bitset_t   *used = lc_bitset_malloc(get_irg_last_idx(si->chordal_env->irg));
2609
2610         irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, used);
2611         irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, used);
2612
2613         lc_bitset_free(used);
2614 }
2615
2616 static void
2617 print_irn_pset(pset * p)
2618 {
2619         ir_node   *irn;
2620
2621         pset_foreach(p, irn) {
2622                 ir_printf("%+F\n", irn);
2623         }
2624 }
2625
2626 static void
2627 rewire_uses(spill_ilp_t * si)
2628 {
2629         dom_front_info_t     *dfi = be_compute_dominance_frontiers(si->chordal_env->irg);
2630         defs_t               *defs;
2631
2632         /* then fix uses of spills */
2633         set_foreach(si->values, defs) {
2634                 pset     *reloads;
2635                 pset     *spills;
2636                 ir_node  *next = defs->remats;
2637                 int remats = 0;
2638
2639                 if(next) {
2640                         reloads = pset_new_ptr_default();
2641
2642                         while(next) {
2643                                 if(be_is_Reload(next)) {
2644                                         pset_insert_ptr(reloads, next);
2645                                 } else {
2646                                         ++remats;
2647                                 }
2648                                 next = get_irn_link(next);
2649                         }
2650
2651                         spills = get_spills_for_value(si, defs->value);
2652                         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));
2653                         if(pset_count(spills) > 1) {
2654                                 assert(pset_count(reloads) > 0);
2655 //                              print_irn_pset(spills);
2656 //                              print_irn_pset(reloads);
2657                                 be_ssa_constr_set_uses(dfi, spills, reloads);
2658                         }
2659
2660                         del_pset(reloads);
2661                         del_pset(spills);
2662                 }
2663         }
2664
2665         /* first fix uses of remats and reloads */
2666         set_foreach(si->values, defs) {
2667                 pset     *nodes;
2668                 ir_node  *next = defs->remats;
2669
2670                 if(next) {
2671                         nodes = pset_new_ptr_default();
2672                         pset_insert_ptr(nodes, defs->value);
2673
2674                         while(next) {
2675                                 pset_insert_ptr(nodes, next);
2676                                 next = get_irn_link(next);
2677                         }
2678
2679                         if(pset_count(nodes) > 1) {
2680                                 DBG((si->dbg, LEVEL_4, "\t    %d new definitions for value %+F\n", pset_count(nodes)-1, defs->value));
2681                                 be_ssa_constr_set(dfi, nodes);
2682                         }
2683
2684                         del_pset(nodes);
2685                 }
2686         }
2687
2688 //      remove_unused_defs(si);
2689
2690         be_free_dominance_frontiers(dfi);
2691 }
2692
2693 static void
2694 writeback_results(spill_ilp_t * si)
2695 {
2696         /* walk through the graph and collect all spills, reloads and remats for a value */
2697
2698         si->values = new_set(cmp_defs, 4096);
2699
2700         DBG((si->dbg, LEVEL_1, "Applying results\n"));
2701         delete_unnecessary_remats(si);
2702         irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
2703         irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
2704
2705         /* clean the remat info! there are still back-edges leading there! */
2706         clean_remat_info(si);
2707
2708         rewire_uses(si);
2709
2710         del_set(si->values);
2711 }
2712
2713 static int
2714 get_n_regs(spill_ilp_t * si)
2715 {
2716         int     arch_n_regs = arch_register_class_n_regs(si->cls);
2717         int     free = 0;
2718         int     i;
2719
2720         for(i=0; i<arch_n_regs; i++) {
2721                 if(!arch_register_type_is(&si->cls->regs[i], ignore)) {
2722                         free++;
2723                 }
2724         }
2725
2726         DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", free, si->cls->name));
2727         return free;
2728 }
2729
2730 static void
2731 walker_reload_mover(ir_node * bb, void * data)
2732 {
2733         spill_ilp_t   *si = data;
2734         ir_node           *tmp;
2735
2736         sched_foreach(bb, tmp) {
2737                 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
2738                         ir_node       *reload = tmp;
2739                         ir_node       *irn = tmp;
2740
2741                         /* move reload upwards */
2742
2743                         int pressure = (int)get_irn_link(reload);
2744                         if(pressure < si->n_regs) {
2745                                 irn = sched_prev(reload);
2746                                 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
2747                                 sched_remove(reload);
2748                                 pressure = (int)get_irn_link(irn);
2749
2750                                 while(pressure < si->n_regs) {
2751                                         if(sched_is_end(irn) || (be_is_Reload(irn) && has_reg_class(si, irn))) break;
2752
2753                                         set_irn_link(irn, INT_TO_PTR(pressure+1));
2754                                         DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
2755                                         irn = sched_prev(irn);
2756
2757                                         pressure = (int)get_irn_link(irn);
2758                                 }
2759
2760                                 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
2761                                 sched_put_after(irn, reload);
2762                         }
2763                 }
2764         }
2765 }
2766
2767 static void
2768 move_reloads_upward(spill_ilp_t * si)
2769 {
2770         irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si);
2771 }
2772
2773 void
2774 be_spill_remat(const be_chordal_env_t * chordal_env)
2775 {
2776         char            problem_name[256];
2777         char            dump_suffix[256];
2778         char            dump_suffix2[256];
2779         char            dump_suffix3[256];
2780         struct obstack  obst;
2781         spill_ilp_t     si;
2782
2783         ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name);
2784         ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name);
2785         ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name);
2786         ir_snprintf(dump_suffix3, sizeof(dump_suffix3), "-%s-reloads_moved", chordal_env->cls->name);
2787
2788         FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
2789         DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
2790
2791         obstack_init(&obst);
2792         si.chordal_env = chordal_env;
2793         si.obst = &obst;
2794         si.senv = be_new_spill_env(chordal_env, is_mem_phi, &si);
2795         si.cls = chordal_env->cls;
2796         si.lpp = new_lpp(problem_name, lpp_minimize);
2797         si.remat_info = new_set(cmp_remat_info, 4096);
2798         si.all_possible_remats = pset_new_ptr_default();
2799         si.inverse_ops = pset_new_ptr_default();
2800 #ifndef EXECFREQ_LOOPDEPH
2801         si.execfreqs = compute_execfreq(chordal_env->irg);
2802 #endif
2803 #ifdef KEEPALIVE
2804         si.keep = NULL;
2805 #endif
2806         si.n_regs = get_n_regs(&si);
2807
2808         set_irg_link(chordal_env->irg, &si);
2809         compute_doms(chordal_env->irg);
2810
2811 #ifdef COLLECT_REMATS
2812         /* collect remats */
2813         DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
2814         irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
2815 #endif
2816
2817         /* insert possible remats */
2818         DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
2819         irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
2820         DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
2821
2822 #ifdef KEEPALIVE
2823         DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
2824         connect_all_remats_with_keep(&si);
2825         /* dump graph with inserted remats */
2826         dump_graph_with_remats(chordal_env->irg, dump_suffix);
2827 #endif
2828
2829
2830         /* recompute liveness */
2831         DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
2832         be_liveness(chordal_env->irg);
2833
2834         /* build the ILP */
2835
2836         DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
2837         DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
2838         irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si);
2839
2840         DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
2841         irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
2842
2843 #ifdef DUMP_ILP
2844         {
2845                 FILE           *f;
2846                 char            buf[256];
2847
2848                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
2849                 if ((f = fopen(buf, "wt")) != NULL) {
2850                         lpp_dump_plain(si.lpp, f);
2851                         fclose(f);
2852                 }
2853         }
2854 #endif
2855
2856 #ifdef SOLVE
2857         DBG((si.dbg, LEVEL_1, "\tSolving %F\n", chordal_env->irg));
2858         lpp_set_time_limit(si.lpp, ILP_TIMEOUT);
2859
2860 #ifdef SOLVE_LOCAL
2861         lpp_solve_cplex(si.lpp);
2862 #else
2863         lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
2864 #endif
2865         assert(lpp_is_sol_valid(si.lpp)
2866                && "solution of ILP must be valid");
2867
2868         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, si.lpp->objval));
2869
2870 #ifdef DUMP_SOLUTION
2871         {
2872                 FILE           *f;
2873                 char            buf[256];
2874
2875                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
2876                 if ((f = fopen(buf, "wt")) != NULL) {
2877                         int             i;
2878                         for (i = 0; i < si.lpp->var_next; ++i) {
2879                                 lpp_name_t     *name = si.lpp->vars[i];
2880                                 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
2881                         }
2882                         fclose(f);
2883                 }
2884         }
2885 #endif
2886
2887         writeback_results(&si);
2888
2889 #endif                          /* SOLVE */
2890
2891         kill_all_unused_values_in_schedule(&si);
2892
2893         be_liveness(chordal_env->irg);
2894         irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
2895
2896         dump_pressure_graph(&si, dump_suffix2);
2897
2898         // TODO fix temporarily exceeded regpressure due to remat2s
2899
2900         // TODO insert copys to fix interferences in memory
2901
2902         // move reloads upwards
2903         move_reloads_upward(&si);
2904         irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
2905         dump_pressure_graph(&si, dump_suffix3);
2906
2907         free_dom(chordal_env->irg);
2908         del_pset(si.inverse_ops);
2909         del_pset(si.all_possible_remats);
2910 #ifndef EXECFREQ_LOOPDEPH
2911         del_set(si.execfreqs);
2912 #endif
2913         free_lpp(si.lpp);
2914         obstack_free(&obst, NULL);
2915 //      exit(0);
2916 }
2917
2918 #else                           /* WITH_ILP */
2919
2920 static void
2921 only_that_you_can_compile_without_WITH_ILP_defined(void)
2922 {
2923 }
2924
2925 #endif                          /* WITH_ILP */