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

   1 /** vim: set sw=4 ts=4:
   2  * @file   bespillremat.c
   3  * @date   2006-04-06
   4  * @author Adam M. Szalkowski & Sebastian Hack
   5  *
   6  * ILP based spilling & rematerialization
   7  *
   8  * Copyright (C) 2006 Universitaet Karlsruhe
   9  * Released under the GPL
  10  */
  11 #ifdef HAVE_CONFIG_H
  12 #include "config.h"
  13 #endif
  14
  15 #ifdef WITH_ILP
  16
  17 #include <math.h>
  18
  19 #include "hashptr.h"
  20 #include "debug.h"
  21 #include "obst.h"
  22 #include "set.h"
  23 #include "list.h"
  24 #include "pmap.h"
  25
  26 #include "irprintf.h"
  27 #include "irgwalk.h"
  28 #include "irdump_t.h"
  29 #include "irnode_t.h"
  30 #include "ircons_t.h"
  31 #include "irloop_t.h"
  32 #include "phiclass.h"
  33 #include "iredges.h"
  34 #include "execfreq.h"
  35
  36 #include <lpp/lpp.h>
  37 #include <lpp/lpp_net.h>
  38 #include <lpp/lpp_cplex.h>
  39 //#include <lc_pset.h>
  40 #include <libcore/lc_bitset.h>
  41
  42 #include "be_t.h"
  43 #include "belive_t.h"
  44 #include "besched_t.h"
  45 #include "beirgmod.h"
  46 #include "bearch.h"
  47 #include "benode_t.h"
  48 #include "beutil.h"
  49 #include "bespillremat.h"
  50 #include "bespill.h"
  51
  52 #include "bechordal_t.h"
  53
  54 #define BIGM 100000.0
  55
  56 #define DUMP_SOLUTION
  57 #define DUMP_ILP
  58 //#define KEEPALIVE /* keep alive all inserted remats and dump graph with remats */
  59 #define COLLECT_REMATS /* enable rematerialization */
  60 #define COLLECT_INVERSE_REMATS /* enable placement of inverse remats */
  61 #define REMAT_WHILE_LIVE /* only remat values that are live */
  62 //#define NO_ENLARGE_L1V3N355 /* do not remat after the death of some operand */
  63 //#define EXECFREQ_LOOPDEPH /* compute execution frequency from loop depth only */
  64 //#define MAY_DIE_AT_PRE_REMAT /* allow values to die after a pre remat */
  65 #define CHECK_POST_REMAT /* check pressure after post remats (conservative but otherwise we can temporarily exceed the register pressure) */
  66 #define NO_SINGLE_USE_REMATS /* do not repair schedule */
  67 //#define KEEPALIVE_SPILLS
  68 //#define KEEPALIVE_RELOADS
  69 //#define GOODWIN_REDUCTION
  70
  71 #define  SOLVE
  72 //#define  SOLVE_LOCAL
  73 #define LPP_SERVER "i44pc52"
  74 #define LPP_SOLVER "cplex"
  75
  76 #define COST_LOAD      10
  77 #define COST_STORE     50
  78 #define COST_REMAT     1
  79
  80 #define ILP_TIMEOUT    90
  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         set                          *execfreqs;
  99         pset                         *spills;
 100         ir_node                      *m_unknown;
 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         if(si->execfreqs) {
 292                 if(is_Block(irn)) {
 293                         return get_block_execfreq(si->execfreqs, irn);
 294                 } else {
 295                         return get_block_execfreq(si->execfreqs, get_nodes_block(irn));
 296                 }
 297         } else {
 298                 if(is_Block(irn))
 299                         return exp(get_loop_depth(get_irn_loop(irn)) * log(10));
 300                 else
 301                         return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10));
 302         }
 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 = arch_get_op_estimated_cost(si->chordal_env->birg->main_env->arch_env, op);
 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 static int get_block_n_succs(ir_node *block) {
 803         const ir_edge_t *edge;
 804
 805         assert(edges_activated(current_ir_graph));
 806
 807         edge = get_block_succ_first(block);
 808         if (! edge)
 809                 return 0;
 810
 811         edge = get_block_succ_next(block, edge);
 812         return edge ? 2 : 1;
 813 }
 814
 815 /**
 816  * Insert (so far unused) remats into the irg to
 817  * recompute the potential liveness of all values
 818  */
 819 static void
 820 walker_remat_insertor(ir_node * bb, void * data)
 821 {
 822         spill_ilp_t    *si = data;
 823         spill_bb_t     *spill_bb;
 824         ir_node        *irn;
 825         int             i,
 826                                         n;
 827         irn_live_t     *li;
 828         pset           *live = pset_new_ptr_default();
 829
 830         DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
 831
 832         live_foreach(bb, li) {
 833                 ir_node        *value = (ir_node *) li->irn;
 834
 835                 /* add remats at end of block */
 836                 if (live_is_end(li) && has_reg_class(si, value)) {
 837                         pset_insert_ptr(live, value);
 838                 }
 839         }
 840
 841         spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
 842         set_irn_link(bb, spill_bb);
 843
 844         irn = sched_last(bb);
 845         while(!sched_is_end(irn)) {
 846                 ir_node   *next;
 847                 op_t      *op;
 848                 pset      *args;
 849                 ir_node   *arg;
 850
 851                 next = sched_prev(irn);
 852
 853                 DBG((si->dbg, LEVEL_5, "\t at %+F (next: %+F)\n", irn, next));
 854
 855                 if(is_Phi(irn) || is_Proj(irn)) {
 856                         op_t      *op;
 857
 858                         if(has_reg_class(si, irn)) {
 859                                 pset_remove_ptr(live, irn);
 860                         }
 861
 862                         op = obstack_alloc(si->obst, sizeof(*op));
 863                         op->is_remat = 0;
 864                         op->attr.live_range.reloads = NULL;
 865                         op->attr.live_range.ilp = ILP_UNDEF;
 866                         set_irn_link(irn, op);
 867
 868                         irn = next;
 869                         continue;
 870                 }
 871
 872                 op = obstack_alloc(si->obst, sizeof(*op));
 873                 op->is_remat = 0;
 874                 op->attr.live_range.ilp = ILP_UNDEF;
 875                 op->attr.live_range.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.reloads) * get_irn_arity(irn));
 876                 memset(op->attr.live_range.reloads, 0xFF, sizeof(*op->attr.live_range.reloads) * get_irn_arity(irn));
 877                 set_irn_link(irn, op);
 878
 879                 args = pset_new_ptr_default();
 880
 881                 /* collect arguments of op */
 882                 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
 883                         ir_node        *arg = get_irn_n(irn, i);
 884
 885                         pset_insert_ptr(args, arg);
 886                 }
 887
 888                 /* set args of op live in epilog */
 889                 pset_foreach(args, arg) {
 890                         if(has_reg_class(si, arg)) {
 891                                 pset_insert_ptr(live, arg);
 892                         }
 893                 }
 894
 895                 /* insert all possible remats after irn */
 896                 pset_foreach(args, arg) {
 897                         remat_info_t   *remat_info,
 898                                                     query;
 899                         remat_t        *remat;
 900
 901                         /* continue if the operand has the wrong reg class
 902                          */
 903                         if(!has_reg_class(si, arg))
 904                                 continue;
 905
 906                         query.irn = arg;
 907                         query.remats = NULL;
 908                         query.remats_by_operand = NULL;
 909                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
 910
 911                         if(!remat_info) {
 912                                 continue;
 913                         }
 914
 915                         /* do not place post remats after jumps */
 916                         if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) continue;
 917
 918                         if(remat_info->remats_by_operand) {
 919                                 pset_foreach(remat_info->remats_by_operand, remat) {
 920                                         /* do not insert remats producing the same value as one of the operands */
 921                                         if(!pset_find_ptr(args, remat->value)) {
 922                                                 DBG((si->dbg, LEVEL_4, "\t  considering remat %+F with arg %+F\n", remat->op, arg));
 923 #ifdef REMAT_WHILE_LIVE
 924                                                 if(pset_find_ptr(live, remat->value)) {
 925                                                         insert_remat_after(si, remat, irn, live);
 926                                                 }
 927 #else
 928                                                 insert_remat_after(si, remat, irn, live);
 929 #endif
 930                                         }
 931                                 }
 932                         }
 933                 }
 934
 935                 /* delete defined value from live set */
 936                 if(has_reg_class(si, irn)) {
 937                         pset_remove_ptr(live, irn);
 938                 }
 939
 940                 /* insert all possible remats before irn */
 941                 pset_foreach(args, arg) {
 942                         remat_info_t   *remat_info,
 943                                                     query;
 944                         remat_t        *remat;
 945
 946                         /* continue if the operand has the wrong reg class
 947                          */
 948                         if(!has_reg_class(si, arg))
 949                                 continue;
 950
 951                         query.irn = arg;
 952                         query.remats = NULL;
 953                         query.remats_by_operand = NULL;
 954                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
 955
 956                         if(!remat_info) {
 957                                 continue;
 958                         }
 959
 960                         if(remat_info->remats) {
 961                                 pset_foreach(remat_info->remats, remat) {
 962                                         DBG((si->dbg, LEVEL_4, "\t  considering remat %+F for arg %+F\n", remat->op, arg));
 963 #ifdef REMAT_WHILE_LIVE
 964                                         if(pset_find_ptr(live, remat->value)) {
 965                                                 insert_remat_before(si, remat, irn, live);
 966                                         }
 967 #else
 968                                         insert_remat_before(si, remat, irn, live);
 969 #endif
 970                                 }
 971                         }
 972                 }
 973
 974                 del_pset(args);
 975                 irn = next;
 976         }
 977
 978         live_foreach(bb, li) {
 979                 ir_node        *value = (ir_node *) li->irn;
 980
 981 #ifdef GOODWIN_REDUCTION
 982                 /* add remats at end if successor has multiple predecessors */
 983                 if(get_block_n_succs(bb) == 1 && get_Block_n_cfgpreds(get_block_succ_first(bb)->src) > 1) {
 984 #endif
 985                         /* add remats at end of block */
 986                         if (live_is_end(li) && has_reg_class(si, value)) {
 987                                 remat_info_t   *remat_info,
 988                                                            query;
 989                                 remat_t        *remat;
 990
 991                                 query.irn = value;
 992                                 query.remats = NULL;
 993                                 query.remats_by_operand = NULL;
 994                                 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
 995
 996                                 if(remat_info && remat_info->remats) {
 997                                         pset_foreach(remat_info->remats, remat) {
 998                                                 DBG((si->dbg, LEVEL_4, "\t  considering remat %+F at end of block %+F\n", remat->op, bb));
 999
1000                                                 insert_remat_before(si, remat, bb, NULL);
1001                                         }
1002                                 }
1003                         }
1004
1005 #ifdef GOODWIN_REDUCTION
1006                 }
1007                 if(get_Block_n_cfgpreds(bb) == 1 && get_block_n_succs(get_Block_cfgpred_block(bb,0)) > 1) {
1008 #endif
1009                         /* add remat2s at beginning of block */
1010                         if ((live_is_in(li) || (is_Phi(value) && get_nodes_block(value)==bb)) && has_reg_class(si, value)) {
1011                                 remat_info_t   *remat_info,
1012                                                            query;
1013                                 remat_t        *remat;
1014
1015                                 query.irn = value;
1016                                 query.remats = NULL;
1017                                 query.remats_by_operand = NULL;
1018                                 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1019
1020                                 if(remat_info && remat_info->remats) {
1021                                         pset_foreach(remat_info->remats, remat) {
1022                                                 DBG((si->dbg, LEVEL_4, "\t  considering remat %+F at beginning of block %+F\n", remat->op, bb));
1023
1024                                                 /* put the remat here if all its args are available */
1025                                                 insert_remat_after(si, remat, bb, NULL);
1026
1027                                         }
1028                                 }
1029                         }
1030
1031 #ifdef GOODWIN_REDUCTION
1032                 }
1033 #endif
1034
1035         }
1036 }
1037
1038 /**
1039  * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1040  */
1041 static void
1042 luke_endwalker(ir_node * bb, void * data)
1043 {
1044         spill_ilp_t    *si = (spill_ilp_t*)data;
1045         irn_live_t     *li;
1046         pset           *live;
1047         pset           *use_end;
1048         char            buf[256];
1049         ilp_cst_t       cst;
1050         ir_node        *irn;
1051         spill_bb_t     *spill_bb = get_irn_link(bb);
1052
1053
1054         live = pset_new_ptr_default();
1055         use_end = pset_new_ptr_default();
1056
1057         live_foreach(bb, li) {
1058                 irn = (ir_node *) li->irn;
1059                 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1060                         op_t      *op;
1061
1062                         pset_insert_ptr(live, irn);
1063                         op = get_irn_link(irn);
1064                         assert(!op->is_remat);
1065                 }
1066         }
1067
1068         /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1069         /* their reg_out is unimportant because it can always be set */
1070         sched_foreach_reverse(bb, irn) {
1071                 int   i,
1072                           n;
1073
1074                 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1075
1076                 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
1077                         ir_node        *irn_arg = get_irn_n(irn, i);
1078                         if(has_reg_class(si, irn_arg)) {
1079                                 pset_insert_ptr(use_end, irn);
1080                         }
1081                 }
1082         }
1083
1084         ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1085         cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1086
1087         spill_bb->ilp = new_set(cmp_spill, 16);
1088
1089         live_foreach(bb, li) {
1090                 irn = (ir_node *) li->irn;
1091                 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1092                         spill_t     query,
1093                                    *spill;
1094
1095                         query.irn = irn;
1096                         spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1097
1098                         ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1099                         spill->reg_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1100                         /* if irn is used at the end of the block, then it is live anyway */
1101                         if(!pset_find_ptr(use_end, irn))
1102                                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1103
1104                         ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1105                         spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1106
1107 #ifdef GOODWIN_REDUCTION
1108                 if(get_Block_n_cfgpreds(bb) == 1 && get_block_n_succs(get_Block_cfgpred_block(bb,0)) > 1) {
1109                         ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1110                         spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, COST_STORE*execution_frequency(si, bb));
1111                 } else {
1112                         spill->spill = ILP_UNDEF;
1113                 }
1114 #else
1115                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1116                 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, COST_STORE*execution_frequency(si, bb));
1117 #endif
1118
1119                         spill->reg_in = ILP_UNDEF;
1120                         spill->mem_in = ILP_UNDEF;
1121                 }
1122         }
1123
1124         del_pset(live);
1125         del_pset(use_end);
1126 }
1127
1128 static ir_node *
1129 next_post_remat(const ir_node * irn)
1130 {
1131         op_t      *op;
1132
1133         if(is_Block(irn)) {
1134                 irn = sched_block_first_nonphi(irn);
1135         } else {
1136                 irn = sched_next_op(irn);
1137         }
1138
1139         if(sched_is_end(irn))
1140                 return NULL;
1141
1142         op = (op_t*)get_irn_link(irn);
1143         if(op->is_remat && !op->attr.remat.pre) {
1144                 return irn;
1145         }
1146
1147         return NULL;
1148 }
1149
1150
1151 static ir_node *
1152 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
1153 {
1154         op_t      *op;
1155         ir_node   *ret;
1156
1157         if(is_Block(irn)) {
1158                 ret = sched_block_last_noncf(si, irn);
1159                 ret = sched_next(ret);
1160                 ret = sched_prev_op(ret);
1161         } else {
1162                 ret = sched_prev_op(irn);
1163         }
1164
1165         if(sched_is_end(ret) || is_Phi(ret))
1166                 return NULL;
1167
1168         op = (op_t*)get_irn_link(ret);
1169         if(op->is_remat && op->attr.remat.pre) {
1170                 return ret;
1171         }
1172
1173         return NULL;
1174 }
1175
1176 /**
1177  * Find a remat of value @p value in the epilog of @p pos
1178  */
1179 static ir_node *
1180 find_post_remat(const ir_node * value, const ir_node * pos)
1181 {
1182         while((pos = next_post_remat(pos)) != NULL) {
1183                 op_t   *op;
1184
1185                 op = get_irn_link(pos);
1186                 assert(op->is_remat && !op->attr.remat.pre);
1187
1188                 if(op->attr.remat.remat->value == value)
1189                         return (ir_node*)pos;
1190
1191 #if 0
1192         const ir_edge_t *edge;
1193                 foreach_out_edge(pos, edge) {
1194                         ir_node   *proj = get_edge_src_irn(edge);
1195                         assert(is_Proj(proj));
1196                 }
1197 #endif
1198
1199         }
1200
1201         return NULL;
1202 }
1203
1204 /**
1205  * Find a remat of value @p value in the prolog of @p pos
1206  */
1207 static ir_node *
1208 find_pre_remat(const spill_ilp_t * si, const ir_node * value, const ir_node * pos)
1209 {
1210         while((pos = next_pre_remat(si,pos)) != NULL) {
1211                 op_t   *op;
1212
1213                 op = get_irn_link(pos);
1214                 assert(op->is_remat && op->attr.remat.pre);
1215
1216                 if(op->attr.remat.remat->value == value)
1217                         return (ir_node*)pos;
1218         }
1219
1220         return NULL;
1221 }
1222
1223 static spill_t *
1224 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1225 {
1226         spill_bb_t  *spill_bb = get_irn_link(bb);
1227         spill_t     *spill,
1228                                  query;
1229         char         buf[256];
1230
1231         query.irn = irn;
1232         spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1233         if(!spill) {
1234                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1235
1236                 spill->reg_out = ILP_UNDEF;
1237                 spill->reg_in  = ILP_UNDEF;
1238                 spill->mem_in  = ILP_UNDEF;
1239
1240                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1241                 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1242
1243 #ifdef GOODWIN_REDUCTION
1244                 if(get_Block_n_cfgpreds(bb) == 1 && get_block_n_succs(get_Block_cfgpred_block(bb,0)) > 1) {
1245                         ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1246                         spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, COST_STORE*execution_frequency(si, bb));
1247                 } else {
1248                         spill->spill = ILP_UNDEF;
1249                 }
1250 #else
1251                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1252                 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, COST_STORE*execution_frequency(si, bb));
1253 #endif
1254         }
1255
1256         return spill;
1257 }
1258
1259 /**
1260  * Walk all irg blocks and emit this ILP
1261  */
1262 static void
1263 luke_blockwalker(ir_node * bb, void * data)
1264 {
1265         spill_ilp_t    *si = (spill_ilp_t*)data;
1266         ir_node        *irn;
1267         irn_live_t     *li;
1268         pset           *live;
1269         char            buf[256];
1270         ilp_cst_t       cst;
1271         spill_bb_t     *spill_bb = get_irn_link(bb);
1272         int             i;
1273         ir_node        *tmp;
1274         spill_t        *spill;
1275
1276
1277         live = pset_new_ptr_default();
1278
1279         /* do something at the end of the block */
1280
1281         /* init live values at end of block */
1282         live_foreach(bb, li) {
1283                 ir_node        *irn = (ir_node *) li->irn;
1284
1285                 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1286                         pset_insert_ptr(live, irn);
1287                 }
1288         }
1289
1290 #ifdef GOODWIN_REDUCTION
1291         if(get_block_n_succs(bb) == 1 && get_Block_n_cfgpreds(get_block_succ_first(bb)->src) > 1) {
1292                 spill_bb->reloads = obstack_alloc(si->obst, pset_count(live) * sizeof(*spill_bb->reloads));
1293                 memset(spill_bb->reloads, 0xFF, pset_count(live) * sizeof(*spill_bb->reloads));
1294         } else {
1295                 spill_bb->reloads = NULL;
1296         }
1297 #else
1298         spill_bb->reloads = obstack_alloc(si->obst, pset_count(live) * sizeof(*spill_bb->reloads));
1299         memset(spill_bb->reloads, 0xFF, pset_count(live) * sizeof(*spill_bb->reloads));
1300 #endif
1301
1302         i=0;
1303         live_foreach(bb, li) {
1304                 ir_node        *irn = (ir_node *) li->irn;
1305                 op_t           *op;
1306
1307                 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1308                         spill = set_find_spill(spill_bb->ilp, irn);
1309                         assert(spill);
1310
1311                         if(spill_bb->reloads) {
1312                                 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1313                                 spill_bb->reloads[i] = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
1314
1315                                 /* reload <= mem_out */
1316                                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1317                                 lpp_set_factor_fast(si->lpp, cst, spill_bb->reloads[i], 1.0);
1318                                 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
1319                         }
1320
1321                         op = get_irn_link(irn);
1322                         assert(!op->is_remat);
1323
1324                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1325                         op->attr.live_range.ilp = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1326                         op->attr.live_range.op = bb;
1327
1328                         ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1329                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1330
1331                         /* reg_out - reload - remat - live_range <= 0 */
1332                         lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1333                         if(spill_bb->reloads) lpp_set_factor_fast(si->lpp, cst, spill_bb->reloads[i], -1.0);
1334                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1335                         foreach_pre_remat(si, bb, tmp) {
1336                                 op_t     *remat_op = get_irn_link(tmp);
1337                                 if(remat_op->attr.remat.remat->value == irn) {
1338                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1339                                 }
1340                         }
1341
1342                         ++i;
1343                 }
1344         }
1345         DBG((si->dbg, LEVEL_4, "\t   %d values live at end of block %+F\n", pset_count(live), bb));
1346
1347         sched_foreach_reverse(bb, irn) {
1348                 op_t       *op;
1349                 op_t       *tmp_op;
1350                 int         n,
1351                                         k,
1352                                         d = 0;
1353                 ilp_cst_t       check_pre,
1354                                         check_post;
1355 #ifdef CHECK_POST_REMAT
1356                 ilp_cst_t       check_post_remat;
1357 #endif
1358                 set        *args = new_set(cmp_keyval, get_irn_arity(irn));
1359                 keyval_t   *keyval;
1360
1361                 if(is_Phi(irn))
1362                         break;
1363
1364                 op = get_irn_link(irn);
1365                 /* skip remats */
1366                 if(op->is_remat) continue;
1367                 DBG((si->dbg, LEVEL_4, "\t  at node %+F\n", irn));
1368
1369                 if(has_reg_class(si, irn)) {
1370                         assert(pset_find_ptr(live, irn));
1371                         pset_remove_ptr(live, irn);
1372                 }
1373
1374                 /* init set of irn's arguments */
1375                 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
1376                         ir_node        *irn_arg = get_irn_n(irn, i);
1377                         if(has_reg_class(si, irn_arg)) {
1378                                 set_insert_keyval(args, irn_arg, (void*)i);
1379                         }
1380                 }
1381
1382 #ifdef CHECK_POST_REMAT
1383                 /* check the register pressure after the epilog */
1384                 ir_snprintf(buf, sizeof(buf), "check_post_remat_%N", irn);
1385                 check_post_remat = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
1386
1387                 /* iterate over L\U */
1388                 pset_foreach(live, tmp) {
1389                         if(!set_find_keyval(args, tmp)) {
1390                                 /* if a live value is not used by irn */
1391                                 tmp_op = get_irn_link(tmp);
1392 //                              assert(tmp_op->attr.live_range.op != irn);
1393                                 lpp_set_factor_fast(si->lpp, check_post_remat, tmp_op->attr.live_range.ilp, 1.0);
1394                         }
1395                 }
1396                 /* iterate over following remats and remove possibly defined values again from check_post_remat */
1397                 foreach_post_remat(irn, tmp) {
1398                         op_t           *remat_op = get_irn_link(tmp);
1399                         const ir_node  *value = remat_op->attr.remat.remat->value;
1400                         op_t           *val_op = get_irn_link(value);
1401
1402                         assert(remat_op->is_remat && !remat_op->attr.remat.pre);
1403
1404                         /* values that are defined by remat2s are not counted */
1405 #ifdef REMAT_WHILE_LIVE
1406                         assert(val_op->attr.live_range.ilp);
1407                         lpp_set_factor_fast(si->lpp, check_post_remat, val_op->attr.live_range.ilp, 0.0);
1408 #else
1409                         if(val_op->attr.live_range.ilp != ILP_UNDEF) {
1410                                 lpp_set_factor_fast(si->lpp, check_post_remat, val_op->attr.live_range.ilp, 0.0);
1411                         }
1412 #endif /* REMAT_WHILE_LIVE */
1413                 }
1414 #endif /* CHECK_POST_REMAT */
1415
1416
1417                 /* new live ranges for values from L\U defined by remat2s or used by remats */
1418                 pset_foreach(live, tmp) {
1419                         ir_node     *value = tmp;//remat_op->attr.remat.remat->value;
1420                         op_t        *value_op = get_irn_link(value);
1421
1422                         if(!set_find_keyval(args, value)) {
1423                                 ilp_var_t    prev_lr = ILP_UNDEF;
1424                                 ir_node     *remat;
1425                                 cst = ILP_UNDEF;
1426
1427                                 foreach_post_remat(irn, remat) {
1428                                         op_t        *remat_op = get_irn_link(remat);
1429
1430                                         /* if value is being rematerialized by this remat */
1431                                         if(value == remat_op->attr.remat.remat->value) {
1432                                                 if(cst == ILP_UNDEF) {
1433                                                         /* next_live_range <= prev_live_range + sum remat2s */
1434                                                         ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
1435                                                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1436                                                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
1437                                                         prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1438                                                         lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
1439                                                         lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1440                                                 }
1441
1442                                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1443                                         }
1444                                 }
1445
1446 #ifdef MAY_DIE_AT_PRE_REMAT
1447                                 if(cst == ILP_UNDEF) {
1448                                         foreach_pre_remat(si, irn, remat) {
1449                                                 int          i,
1450                                                                          n;
1451
1452                                                 for (i = 0, n = get_irn_arity(remat); i < n; ++i) {
1453                                                         ir_node        *remat_arg = get_irn_n(remat, i);
1454
1455                                                         /* if value is being used by this remat */
1456                                                         if(value == remat_arg) {
1457                                                                 /* next_live_range <= prev_live_range */
1458                                                                 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
1459                                                                 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1460
1461                                                                 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
1462                                                                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1463                                                                 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
1464                                                                 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1465                                                                 goto fertig;
1466                                                         }
1467                                                         /* TODO check afterwards whether lr dies after a pre-remat (should not happen) */
1468                                                 }
1469                                         }
1470                                 }
1471 #endif
1472
1473 fertig:
1474                                 if(prev_lr != ILP_UNDEF) {
1475                                         value_op->attr.live_range.ilp = prev_lr;
1476                                         value_op->attr.live_range.op = irn;
1477                                 }
1478                         }
1479                 }
1480
1481                 /* get count of values in my register class defined by irn */
1482                 /* also add defined values to check_post_remat; do this before iterating over args */
1483                 if(get_irn_mode(irn) == mode_T) {
1484                         ir_node  *proj = sched_next(irn);
1485                         op_t     *proj_op = get_irn_link(proj);
1486
1487                         while(is_Proj(proj)) {
1488                                 if(has_reg_class(si, proj)) {
1489                                         ++d;
1490 #ifdef CHECK_POST_REMAT
1491                                         lpp_set_factor_fast(si->lpp, check_post_remat, proj_op->attr.live_range.ilp, 1.0);
1492 #endif
1493                                 }
1494                                 proj = sched_next(proj);
1495                                 proj_op = get_irn_link(proj);
1496                         }
1497                 } else {
1498                         if(has_reg_class(si, irn)) {
1499                                  d = 1;
1500 #ifdef CHECK_POST_REMAT
1501                                  lpp_set_factor_fast(si->lpp, check_post_remat, op->attr.live_range.ilp, 1.0);
1502 #endif
1503                         }
1504                 }
1505                 DBG((si->dbg, LEVEL_4, "\t   %+F produces %d values in my register class\n", irn, d));
1506
1507                 /* count how many regs irn needs for arguments */
1508                 k = set_count(args);
1509
1510                 /* check the register pressure in the prolog */
1511                 /* sum_{L\U} lr <= n - |U| */
1512                 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
1513                 check_pre = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - k);
1514
1515                 /* check the register pressure in the epilog */
1516                 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
1517                 check_post = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - d);
1518
1519                 set_foreach(args, keyval) {
1520                         ilp_var_t       next_lr;
1521                         op_t           *arg_op;
1522                         ilp_var_t       post_use;
1523                         int             p = 0;
1524                         spill_t        *spill;
1525                         ir_node        *arg = keyval->key;
1526
1527                         spill = add_to_spill_bb(si, bb, arg);
1528
1529                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
1530                         next_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1531
1532                         i = (int)keyval->val;
1533                         assert(i<n);
1534
1535                         ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
1536                         op->attr.live_range.reloads[i] = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, irn));
1537
1538                         /* reload <= mem_out */
1539                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1540                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.reloads[i], 1.0);
1541                         lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
1542
1543                         arg_op = get_irn_link(arg);
1544
1545                         /* requirement: arg must be in register for use */
1546                         /* reload + remat + live_range == 1 */
1547                         ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
1548                         cst = lpp_add_cst(si->lpp, buf, lpp_equal, 1.0);
1549
1550                         lpp_set_factor_fast(si->lpp, cst, next_lr, 1.0);
1551                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.reloads[i], 1.0);
1552                         foreach_pre_remat(si, irn, tmp) {
1553                                 op_t     *remat_op = get_irn_link(tmp);
1554                                 if(remat_op->attr.remat.remat->value == arg) {
1555                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1556                                 }
1557                         }
1558
1559                         /* the epilog stuff - including post_use, post, post_remat */
1560                         ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
1561                         post_use = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1562
1563                         lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
1564
1565                         /* arg is live throughout epilog if the next live_range is in a register */
1566                         if(pset_find_ptr(live, arg)) {
1567                                 DBG((si->dbg, LEVEL_3, "\t  arg %+F is possibly live in epilog of %+F\n", arg, irn));
1568
1569                                 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1570                                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1571                                 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
1572                                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1573
1574 #ifdef CHECK_POST_REMAT
1575                                 lpp_set_factor_fast(si->lpp, check_post_remat, arg_op->attr.live_range.ilp, 1.0);
1576 #endif
1577                         }
1578
1579                         /*forall remat2 which use arg add a similar cst*/
1580                         foreach_post_remat(irn, tmp) {
1581                                 int         i,
1582                                                         n;
1583
1584                                 for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1585                                         ir_node    *remat_arg = get_irn_n(tmp, i);
1586                                         op_t       *remat_op = get_irn_link(tmp);
1587
1588                                         if(remat_arg == arg) {
1589                                                 DBG((si->dbg, LEVEL_3, "\t  found remat with arg %+F in epilog of %+F\n", arg, irn));
1590
1591                                                 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1592                                                 cst = lpp_add_cst(si->lpp, buf, lpp_greater, 0.0);
1593                                                 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
1594                                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1595                                         }
1596                                 }
1597                         }
1598
1599                         /* new live range begins for each argument */
1600                         arg_op->attr.live_range.ilp = next_lr;
1601                         arg_op->attr.live_range.op = irn;
1602
1603                         pset_insert_ptr(live, arg);
1604                 }
1605
1606                 /* start new live ranges for values used by remats */
1607                 foreach_pre_remat(si, irn, tmp) {
1608                         int          i,
1609                                                  n;
1610
1611                         for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1612                                 ir_node        *remat_arg = get_irn_n(tmp, i);
1613                                 op_t           *arg_op = get_irn_link(remat_arg);
1614                                 ilp_var_t       prev_lr;
1615
1616                                 if(!has_reg_class(si, remat_arg)) continue;
1617
1618                                 /* if value is becoming live through use by remat */
1619                                 if(!pset_find_ptr(live, remat_arg)) {
1620                                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
1621                                         prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1622
1623                                         arg_op->attr.live_range.ilp = prev_lr;
1624                                         arg_op->attr.live_range.op = irn;
1625
1626                                         DBG((si->dbg, LEVEL_4, "  value %+F becoming live through use by remat %+F\n", remat_arg, tmp));
1627
1628                                         /* TODO ist das hier die richtige Stelle???? */
1629                                         pset_insert_ptr(live, remat_arg);
1630                                         add_to_spill_bb(si, bb, remat_arg);
1631                                 }
1632                                 /* TODO check afterwards whether lr dies after a pre-remat (should not happen) */
1633                         }
1634                 }
1635
1636                 /* iterate over L\U */
1637                 pset_foreach(live, tmp) {
1638                         if(!set_find_keyval(args, tmp)) {
1639                                 /* if a live value is not used by irn */
1640                                 tmp_op = get_irn_link(tmp);
1641 //                              assert(tmp_op->attr.live_range.op != irn);
1642                                 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
1643                                 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
1644                         }
1645                 }
1646
1647                 /* requirements for remats */
1648                 foreach_pre_remat(si, irn, tmp) {
1649                         op_t        *remat_op = get_irn_link(tmp);
1650                         int          i,
1651                                                  n;
1652
1653                         for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1654                                 ir_node        *remat_arg = get_irn_n(tmp, i);
1655                                 op_t           *arg_op = get_irn_link(remat_arg);
1656
1657                                 if(!has_reg_class(si, remat_arg)) continue;
1658
1659                                 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
1660                                 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1661                                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1662
1663                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1664                                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1665
1666                                 /* if remat arg is also used by current op then we can use reload placed for this argument */
1667                                 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
1668                                         int    index = (int)keyval->val;
1669
1670                                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.reloads[index], -1.0);
1671                                 }
1672                         }
1673                 }
1674
1675                 /* requirements for remats2
1676                  *
1677                  *  TODO unsure if this does the right thing.
1678                  *  should insert values into set if they do not become live through remat and
1679                  *  op
1680                  */
1681                 foreach_post_remat(irn, tmp) {
1682                         op_t        *remat_op = get_irn_link(tmp);
1683                         int          i,
1684                                                  n;
1685
1686                         for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1687                                 ir_node        *remat_arg = get_irn_n(tmp, i);
1688                                 op_t           *arg_op = get_irn_link(remat_arg);
1689
1690                                 if(!has_reg_class(si, remat_arg)) continue;
1691
1692                                 /* only for values in L\U, the others are handled with post_use */
1693                                 if(!set_find_keyval(args, remat_arg)) {
1694                                         /* remat <= live_rang(remat_arg) */
1695                                         ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
1696                                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1697
1698                                         /* if value is becoming live through use by remat2 */
1699                                         if(!pset_find_ptr(live, remat_arg)) {
1700                                                 ilp_var_t     lr;
1701
1702                                                 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
1703                                                 lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1704
1705                                                 arg_op->attr.live_range.ilp = lr;
1706                                                 arg_op->attr.live_range.op = irn;
1707
1708                                                 DBG((si->dbg, LEVEL_3, "  value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
1709
1710                                                 pset_insert_ptr(live, remat_arg);
1711                                                 add_to_spill_bb(si, bb, remat_arg);
1712                                         }
1713
1714                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1715                                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1716                                 }
1717                         }
1718                 }
1719
1720 #ifdef CHECK_POST_REMAT
1721                 /* iterate over following remats and add them to check_post_remat */
1722                 foreach_post_remat(irn, tmp) {
1723                         op_t           *remat_op = get_irn_link(tmp);
1724
1725                         assert(remat_op->is_remat && !remat_op->attr.remat.pre);
1726
1727                         lpp_set_factor_fast(si->lpp, check_post_remat, remat_op->attr.remat.ilp, 1.0);
1728                 }
1729 #endif
1730
1731
1732
1733                 DBG((si->dbg, LEVEL_4, "\t   %d values live at %+F\n", pset_count(live), irn));
1734
1735                 pset_foreach(live, tmp) {
1736                         assert(has_reg_class(si, tmp));
1737                 }
1738
1739                 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
1740                         ir_node        *arg = get_irn_n(irn, i);
1741
1742                         assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
1743                 }
1744
1745                 del_set(args);
1746         }
1747
1748
1749
1750         /* do something at the beginning of the block */
1751
1752         /* we are now at the beginning of the basic block, there are only \Phis in front of us */
1753         DBG((si->dbg, LEVEL_3, "\t   %d values live at beginning of block %+F\n", pset_count(live), bb));
1754
1755         pset_foreach(live, irn) {
1756                 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
1757         }
1758
1759         /* construct mem_outs for all values */
1760
1761         set_foreach(spill_bb->ilp, spill) {
1762                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
1763                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1764
1765                 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
1766                 if(spill->spill != ILP_UNDEF) {
1767                         lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
1768                 }
1769
1770                 if(pset_find_ptr(live, spill->irn)) {
1771                         DBG((si->dbg, LEVEL_5, "\t     %+F live at beginning of block %+F\n", spill->irn, bb));
1772
1773                         ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
1774                         spill->mem_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1775
1776                         lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
1777                 }
1778         }
1779
1780
1781         /* L\U is empty at bb start */
1782         /* arg is live throughout epilog if it is reg_in into this block */
1783
1784         /* check the register pressure at the beginning of the block
1785          * including remats
1786          */
1787         ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
1788         cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
1789
1790         pset_foreach(live, irn) {
1791                         spill = set_find_spill(spill_bb->ilp, irn);
1792                         assert(spill);
1793
1794                         ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
1795                         spill->reg_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1796
1797                         lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
1798         }
1799         foreach_post_remat(bb, irn) {
1800                 op_t     *remat_op = get_irn_link(irn);
1801
1802                 DBG((si->dbg, LEVEL_4, "\t  next post remat: %+F\n", irn));
1803                 assert(remat_op->is_remat && !remat_op->attr.remat.pre);
1804
1805                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1806         }
1807
1808         /* forall remat2 add requirements */
1809         foreach_post_remat(bb, tmp) {
1810                 int         i,
1811                                         n;
1812
1813                 for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1814                         ir_node    *remat_arg = get_irn_n(tmp, i);
1815                         op_t       *remat_op = get_irn_link(tmp);
1816
1817                         if(!has_reg_class(si, remat_arg)) continue;
1818
1819                         spill = set_find_spill(spill_bb->ilp, remat_arg);
1820                         assert(spill);
1821
1822                         /* TODO verify this is placed correctly */
1823                         ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
1824                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1825                         lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
1826                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1827                 }
1828         }
1829
1830         /* mem_in/reg_in for live_in values, especially phis and their arguments */
1831         pset_foreach(live, irn) {
1832                 int          p = 0,
1833                                          i,
1834                                          n;
1835
1836                 spill = set_find_spill(spill_bb->ilp, irn);
1837                 assert(spill && spill->irn == irn);
1838
1839                 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
1840                         for (i = 0, n = get_Phi_n_preds(irn); i < n; ++i) {
1841                                 ilp_cst_t       mem_in,
1842                                                                 reg_in;
1843                                 ir_node        *phi_arg = get_Phi_pred(irn, i);
1844                                 ir_node        *bb_p = get_Block_cfgpred_block(bb, i);
1845                                 spill_bb_t     *spill_bb_p = get_irn_link(bb_p);
1846                                 spill_t        *spill_p;
1847
1848                                 /* although the phi is in the right regclass one or more of
1849                                  * its arguments can be in a different one or at least to
1850                                  * ignore
1851                                  */
1852                                 if(has_reg_class(si, phi_arg)) {
1853                                         ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
1854                                         mem_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1855                                         ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
1856                                         reg_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1857
1858                                         lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
1859                                         lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
1860
1861                                         spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
1862                                         assert(spill_p);
1863
1864                                         lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
1865                                         lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
1866                                 }
1867                         }
1868                 } else {
1869                         /* else assure the value arrives on all paths in the same resource */
1870
1871                         for (i = 0, n = get_Block_n_cfgpreds(bb); i < n; ++i) {
1872                                 ilp_cst_t       mem_in,
1873                                                                 reg_in;
1874                                 ir_node        *bb_p = get_Block_cfgpred_block(bb, i);
1875                                 spill_bb_t     *spill_bb_p = get_irn_link(bb_p);
1876                                 spill_t        *spill_p;
1877
1878                                 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
1879                                 mem_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1880                                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
1881                                 reg_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1882
1883                                 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
1884                                 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
1885
1886                                 spill_p = set_find_spill(spill_bb_p->ilp, irn);
1887                                 assert(spill_p);
1888
1889                                 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
1890                                 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
1891                         }
1892                 }
1893         }
1894
1895         /* first live ranges from reg_ins */
1896         pset_foreach(live, irn) {
1897                 op_t      *op = get_irn_link(irn);
1898
1899                 spill = set_find_spill(spill_bb->ilp, irn);
1900                 assert(spill && spill->irn == irn);
1901
1902                 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
1903                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1904                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1905                 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
1906
1907                 foreach_post_remat(bb, tmp) {
1908                         op_t     *remat_op = get_irn_link(tmp);
1909
1910                         if(remat_op->attr.remat.remat->value == irn) {
1911                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1912                         }
1913                 }
1914         }
1915
1916         /* walk forward now and compute constraints for placing spills */
1917         /* this must only be done for values that are not defined in this block */
1918 #ifdef GOODWIN_REDUCTION
1919         if(get_Block_n_cfgpreds(bb) == 1 && get_block_n_succs(get_Block_cfgpred_block(bb,0)) > 1) {
1920 #endif
1921                 pset_foreach(live, irn) {
1922                         ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", spill->irn, bb);
1923                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1924
1925                         spill = set_find_spill(spill_bb->ilp, irn);
1926                         assert(spill);
1927
1928                         lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
1929                         lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
1930
1931                         sched_foreach_op(bb, tmp) {
1932                                 op_t   *op = get_irn_link(tmp);
1933
1934                                 if(is_Phi(tmp)) continue;
1935                                 assert(!is_Proj(tmp));
1936
1937                                 if(op->is_remat) {
1938                                         ir_node   *value = op->attr.remat.remat->value;
1939
1940                                         if(value == irn) {
1941                                                 /* only collect remats up to the first use of a value */
1942                                                 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
1943                                         }
1944                                 } else {
1945                                         int i,
1946                                                 n;
1947
1948                                         for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1949                                                 ir_node    *arg = get_irn_n(tmp, i);
1950
1951                                                 if(arg == irn) {
1952                                                         /* if a value is used stop collecting remats */
1953                                                         cst = ILP_UNDEF;
1954                                                 }
1955                                                 break;
1956                                         }
1957                                 }
1958                                 if(cst == ILP_UNDEF) break;
1959                         }
1960                 }
1961 #ifdef GOODWIN_REDUCTION
1962         }
1963 #endif
1964
1965
1966         /* 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)
1967            mem_in(phi) -> not mem_in(orig_value) TODO: how does this depend on a certain predecessor?
1968          */
1969
1970         /* mem_in of mem-phi has associated costs (but first one is free) */
1971         /* define n_mem_copies as positive integer in each predecessor block,
1972            #mem_in into this block from predecessor block - 1 weighted with SPILL_COST*execfreq(predecessor)
1973            TODO
1974          */
1975
1976
1977         del_pset(live);
1978 }
1979
1980
1981 #if 0
1982          * Speicherkopienminimierung: teste Speicherwerte auf Interferenz
1983          * und weise Spillkontexte zu. Sorge bei Phis dafuer, dass gleiche
1984          * Kontexte zusammenfliessen (Operanden und Ergebnis hat gleichen
1985          * Kontext)
1986 #endif
1987
1988 static INLINE int
1989 is_zero(double x)
1990 {
1991         return fabs(x) < 0.00001;
1992 }
1993
1994 #if 0
1995 static int
1996 is_spilled(const spill_ilp_t * si, const live_range_t * lr)
1997 {
1998         return !is_zero(lpp_get_var_sol(si->lpp, lr->in_mem_var));
1999 }
2000 #endif
2001
2002 static int
2003 is_mem_phi(const ir_node * phi, void *data)
2004 {
2005         spill_ilp_t    *si = data;
2006 //      return is_spilled(si, get_use_head(si, phi)->closest_use);
2007         return 0;
2008 }
2009
2010 #ifdef KEEPALIVE
2011 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
2012 {
2013         spill_ilp_t *si = get_irg_link(current_ir_graph);
2014
2015         if(pset_find_ptr(si->all_possible_remats, n)) {
2016                 op_t   *op = (op_t*)get_irn_link(n);
2017                 assert(op && op->is_remat);
2018
2019                 if(!op->attr.remat.remat->inverse) {
2020                         if(op->attr.remat.pre) {
2021                                 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
2022                         } else {
2023                                 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
2024                         }
2025
2026                         return 1;
2027                 } else {
2028                         op_t   *op = (op_t*)get_irn_link(n);
2029                         assert(op && op->is_remat);
2030
2031                         if(op->attr.remat.pre) {
2032                                 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
2033                         } else {
2034                                 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
2035                         }
2036
2037                         return 1;
2038                 }
2039         }
2040
2041         return 0;
2042 }
2043
2044 static void
2045 dump_graph_with_remats(ir_graph * irg, const char * suffix)
2046 {
2047         set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
2048         be_dump(irg, suffix, dump_ir_block_graph_sched);
2049         set_dump_node_vcgattr_hook(NULL);
2050 }
2051 #endif
2052
2053 /**
2054  * Edge hook to dump the schedule edges with annotated register pressure.
2055  */
2056 static int
2057 sched_pressure_edge_hook(FILE *F, ir_node *irn)
2058 {
2059         if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
2060                 ir_node *prev = sched_prev(irn);
2061                 fprintf(F, "edge:{sourcename:\"");
2062                 PRINT_NODEID(irn);
2063                 fprintf(F, "\" targetname:\"");
2064                 PRINT_NODEID(prev);
2065                 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
2066                 fprintf(F, "\" color:magenta}\n");
2067         }
2068         return 1;
2069 }
2070
2071 static void
2072 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
2073 {
2074         DUMP_NODE_EDGE_FUNC old = get_dump_node_edge_hook();
2075
2076         dump_consts_local(0);
2077         set_dump_node_edge_hook(sched_pressure_edge_hook);
2078         dump_ir_block_graph(irg, suffix);
2079         set_dump_node_edge_hook(old);
2080 }
2081
2082 static void
2083 walker_pressure_annotator(ir_node * bb, void * data)
2084 {
2085         spill_ilp_t  *si = data;
2086         ir_node      *irn;
2087         irn_live_t   *li;
2088         int           i,
2089                                   n;
2090         pset         *live = pset_new_ptr_default();
2091         int           projs = 0;
2092
2093         live_foreach(bb, li) {
2094                 irn = (ir_node *) li->irn;
2095
2096                 if (live_is_end(li) && has_reg_class(si, irn)) {
2097                         pset_insert_ptr(live, irn);
2098                 }
2099         }
2100
2101         set_irn_link(bb, INT_TO_PTR(pset_count(live)));
2102
2103         sched_foreach_reverse(bb, irn) {
2104                 if(is_Phi(irn)) {
2105                         set_irn_link(irn, INT_TO_PTR(pset_count(live)));
2106                         continue;
2107                 }
2108
2109                 if(has_reg_class(si, irn)) {
2110                         pset_remove_ptr(live, irn);
2111                         if(is_Proj(irn)) ++projs;
2112                 }
2113
2114                 if(!is_Proj(irn)) projs = 0;
2115
2116                 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
2117                         ir_node    *arg = get_irn_n(irn, i);
2118
2119                         if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
2120                 }
2121                 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
2122         }
2123
2124         del_pset(live);
2125 }
2126
2127 static void
2128 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
2129 {
2130         be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
2131 }
2132
2133 #ifdef KEEPALIVE
2134 static void
2135 connect_all_remats_with_keep(spill_ilp_t * si)
2136 {
2137         ir_node   *irn;
2138         ir_node  **ins,
2139                          **pos;
2140         int        n_remats;
2141
2142
2143         n_remats = pset_count(si->all_possible_remats);
2144         if(n_remats) {
2145                 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
2146
2147                 pos = ins;
2148                 pset_foreach(si->all_possible_remats, irn) {
2149                         *pos = irn;
2150                         ++pos;
2151                 }
2152
2153                 si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins);
2154
2155                 obstack_free(si->obst, ins);
2156         }
2157 }
2158 #endif
2159
2160 static void
2161 connect_all_spills_with_keep(spill_ilp_t * si)
2162 {
2163         ir_node   *irn;
2164         ir_node  **ins,
2165                          **pos;
2166         int        n_spills;
2167         ir_node   *keep;
2168
2169
2170         n_spills = pset_count(si->spills);
2171         if(n_spills) {
2172                 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
2173
2174                 pos = ins;
2175                 pset_foreach(si->spills, irn) {
2176                         *pos = irn;
2177                         ++pos;
2178                 }
2179
2180                 keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_spills, ins);
2181
2182                 obstack_free(si->obst, ins);
2183         }
2184 }
2185
2186 /** insert a spill at an arbitrary position */
2187 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert, ir_node *ctx)
2188 {
2189         ir_node *bl     = is_Block(insert)?insert:get_nodes_block(insert);
2190         ir_graph *irg   = get_irn_irg(bl);
2191         ir_node *frame  = get_irg_frame(irg);
2192         ir_node *spill;
2193         ir_node *next;
2194
2195         const arch_register_class_t *cls       = arch_get_irn_reg_class(arch_env, irn, -1);
2196         const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
2197
2198         spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn, ctx);
2199
2200         /*
2201          * search the right insertion point. a spill of a phi cannot be put
2202          * directly after the phi, if there are some phis behind the one which
2203          * is spilled. Also, a spill of a Proj must be after all Projs of the
2204          * same tuple node.
2205          *
2206          * Here's one special case:
2207          * If the spill is in the start block, the spill must be after the frame
2208          * pointer is set up. This is done by setting insert to the end of the block
2209          * which is its default initialization (see above).
2210          */
2211
2212         if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
2213                 insert = frame;
2214
2215         for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
2216                 insert = next;
2217
2218         sched_add_after(insert, spill);
2219         return spill;
2220 }
2221
2222 static void
2223 delete_remat(spill_ilp_t * si, ir_node * remat) {
2224         int       i,
2225                       n;
2226         ir_node  *bad = get_irg_bad(si->chordal_env->irg);
2227
2228         sched_remove(remat);
2229
2230         /* kill links to operands */
2231         for (i = -1, n = get_irn_arity(remat); i < n; ++i) {
2232                 set_irn_n(remat, i, bad);
2233         }
2234 }
2235
2236 static void
2237 clean_remat_info(spill_ilp_t * si)
2238 {
2239         int            i,
2240                                n;
2241         remat_t       *remat;
2242         remat_info_t  *remat_info;
2243         ir_node       *bad = get_irg_bad(si->chordal_env->irg);
2244
2245         set_foreach(si->remat_info, remat_info) {
2246                 if(!remat_info->remats) continue;
2247
2248                 pset_foreach(remat_info->remats, remat)
2249                 {
2250                         if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
2251                                 set_irn_n(remat->proj, -1, bad);
2252                                 set_irn_n(remat->proj, 0, bad);
2253                         }
2254
2255                         if(get_irn_n_edges(remat->op) == 0) {
2256                                 for (i = -1, n = get_irn_arity(remat->op); i < n; ++i) {
2257                                         set_irn_n(remat->op, i, bad);
2258                                 }
2259                         }
2260                 }
2261
2262                 if(remat_info->remats) del_pset(remat_info->remats);
2263                 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
2264         }
2265 }
2266
2267 static void
2268 delete_unnecessary_remats(spill_ilp_t * si)
2269 {
2270 #ifdef KEEPALIVE
2271         int       i,
2272                       n;
2273         ir_node  *bad = get_irg_bad(si->chordal_env->irg);
2274
2275         if(si->keep) {
2276                 ir_node   *end = get_irg_end(si->chordal_env->irg);
2277                 ir_node  **keeps;
2278
2279                 for (i = 0, n = get_irn_arity(si->keep); i < n; ++i) {
2280                         ir_node        *keep_arg = get_irn_n(si->keep, i);
2281                         op_t           *arg_op = get_irn_link(keep_arg);
2282                         lpp_name_t     *name;
2283
2284                         assert(arg_op->is_remat);
2285
2286                         name = si->lpp->vars[arg_op->attr.remat.ilp];
2287
2288                         if(is_zero(name->value)) {
2289                                 DBG((si->dbg, LEVEL_3, "\t  deleting remat %+F\n", keep_arg));
2290                                 /* TODO check whether reload is preferred over remat (could be bug) */
2291                                 delete_remat(si, keep_arg);
2292                         } else {
2293                                 if(!arg_op->attr.remat.remat->inverse) {
2294                                         if(arg_op->attr.remat.pre) {
2295                                                 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
2296                                         } else {
2297                                                 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
2298                                         }
2299                                 } else {
2300                                         if(arg_op->attr.remat.pre) {
2301                                                 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
2302                                         } else {
2303                                                 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
2304                                         }
2305                                 }
2306                         }
2307
2308                         set_irn_n(si->keep, i, bad);
2309                 }
2310 #if 0
2311                 for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
2312                         ir_node        *end_arg = get_End_keepalive(end, i);
2313
2314                         if(end_arg != si->keep) {
2315                                 obstack_grow(si->obst, &end_arg, sizeof(end_arg));
2316                         }
2317                 }
2318                 keeps = obstack_finish(si->obst);
2319                 set_End_keepalives(end, n-1, keeps);
2320                 obstack_free(si->obst, keeps);
2321 #endif
2322         } else {
2323                 DBG((si->dbg, LEVEL_2, "\t  no remats to delete (none have been inserted)\n"));
2324         }
2325 #else
2326         ir_node  *remat;
2327
2328         pset_foreach(si->all_possible_remats, remat) {
2329                 op_t           *remat_op = get_irn_link(remat);
2330                 lpp_name_t     *name = si->lpp->vars[remat_op->attr.remat.ilp];
2331
2332                 if(is_zero(name->value)) {
2333                         DBG((si->dbg, LEVEL_3, "\t  deleting remat %+F\n", remat));
2334                         /* TODO check whether reload is preferred over remat (could be bug) */
2335                         delete_remat(si, remat);
2336                 } else {
2337                         if(!remat_op->attr.remat.remat->inverse) {
2338                                 if(remat_op->attr.remat.pre) {
2339                                         DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
2340                                 } else {
2341                                         DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
2342                                 }
2343                         } else {
2344                                 if(remat_op->attr.remat.pre) {
2345                                         DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
2346                                 } else {
2347                                         DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
2348                                 }
2349                         }
2350                 }
2351         }
2352 #endif
2353 }
2354
2355 /**
2356  * @param before   The node after which the spill will be placed in the schedule
2357  */
2358 /* TODO set context properly */
2359 static ir_node *
2360 insert_spill(spill_ilp_t * si, const ir_node * irn, const ir_node * value, const ir_node * before)
2361 {
2362         defs_t   *defs;
2363         ir_node  *spill;
2364         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
2365
2366         DBG((si->dbg, LEVEL_3, "\t  inserting spill for value %+F after %+F\n", irn, before));
2367
2368         spill = be_spill2(arch_env, irn, before, irn);
2369
2370         defs = set_insert_def(si->values, value);
2371         assert(defs);
2372
2373         /* enter into the linked list */
2374         set_irn_link(spill, defs->spills);
2375         defs->spills = spill;
2376
2377 #ifdef KEEPALIVE_SPILLS
2378         pset_insert_ptr(si->spills, spill);
2379 #endif
2380
2381         return spill;
2382 }
2383
2384 /**
2385  * @param before   The Phi node which has to be spilled
2386  */
2387 static ir_node *
2388 insert_mem_phi(spill_ilp_t * si, const ir_node * phi)
2389 {
2390         ir_node   *mem_phi;
2391         ir_node  **ins;
2392         defs_t    *defs;
2393         int        i,
2394                            n;
2395
2396         NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
2397
2398         for(i=0,n=get_irn_arity(phi); i<n; ++i) {
2399                 ins[i] = si->m_unknown;
2400         }
2401
2402         mem_phi =  new_r_Phi(si->chordal_env->irg, get_nodes_block(phi), get_irn_arity(phi), ins, mode_M);
2403
2404         defs = set_insert_def(si->values, phi);
2405         assert(defs);
2406
2407         /* enter into the linked list */
2408         set_irn_link(mem_phi, defs->spills);
2409         defs->spills = mem_phi;
2410
2411 #ifdef KEEPALIVE_SPILLS
2412         pset_insert_ptr(si->spills, mem_phi);
2413 #endif
2414
2415         return mem_phi;
2416 }
2417
2418 /**
2419  * Add remat to list of defs, destroys link field!
2420  */
2421 static void
2422 insert_remat(spill_ilp_t * si, ir_node * remat)
2423 {
2424         defs_t   *defs;
2425         op_t     *remat_op = get_irn_link(remat);
2426
2427         assert(remat_op->is_remat);
2428
2429         defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
2430         assert(defs);
2431
2432         /* enter into the linked list */
2433         set_irn_link(remat, defs->remats);
2434         defs->remats = remat;
2435 }
2436
2437 #if 0
2438 static void
2439 collect_spills(spill_ilp_t * si, ir_node * value, pset * spills, pset * visited)
2440 {
2441         ir_node  *next;
2442         defs_t   *defs;
2443
2444         defs = set_find_def(si->values, value);
2445
2446         if(defs && defs->spills) {
2447                 for(next = defs->spills; next; next = get_irn_link(next)) {
2448                         pset_insert_ptr(spills, next);
2449                 }
2450         } else if (is_Phi(value)) {
2451                 /* recursion */
2452                 if(!pset_find_ptr(visited, value)) {
2453                         int    i,
2454                                    n;
2455
2456                         pset_insert_ptr(visited, value);
2457                         for(i=0, n=get_irn_arity(value); i<n; ++i) {
2458                                 ir_node    *arg = get_irn_n(value, i);
2459
2460                                 collect_spills(si, arg, spills, visited);
2461                         }
2462                 }
2463         } else {
2464 //              assert(0 && "Phi operand not spilled");
2465         }
2466 }
2467 #endif
2468
2469 static pset *
2470 get_spills_for_value(spill_ilp_t * si, ir_node * value)
2471 {
2472         pset     *spills = pset_new_ptr_default();
2473 //      pset     *visited = pset_new_ptr_default();
2474
2475 //      collect_spills(si, value, spills, visited);
2476 //      del_pset(visited);
2477         ir_node  *next;
2478         defs_t   *defs;
2479
2480         defs = set_find_def(si->values, value);
2481
2482         if(defs && defs->spills) {
2483                 for(next = defs->spills; next; next = get_irn_link(next)) {
2484                         pset_insert_ptr(spills, next);
2485                 }
2486         }
2487
2488         return spills;
2489 }
2490
2491 /**
2492  * Add reload before operation and add to list of defs
2493  */
2494 static ir_node *
2495 insert_reload(spill_ilp_t * si, const ir_node * value, const ir_node * after)
2496 {
2497         defs_t   *defs;
2498         ir_node  *reload,
2499                          *spill;
2500         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
2501
2502         DBG((si->dbg, LEVEL_3, "\t  inserting reload for value %+F before %+F\n", value, after));
2503
2504         defs = set_find_def(si->values, value);
2505         /* get a spill of this value */
2506 #if 0
2507         if((!defs || !defs->spills) && is_Phi(value)) {
2508                 pset  *spills;
2509
2510                 spills = get_spills_for_value(si, value);
2511
2512                 spill = pset_first(spills);
2513                 del_pset(spills);
2514
2515                 if(!defs) {
2516                         defs = set_insert_def(si->values, value);
2517                 }
2518                 defs->spills = spill;
2519                 set_irn_link(spill, NULL);
2520         } else {
2521                 spill = defs->spills;
2522         }
2523 #endif
2524         spill = defs->spills;
2525         assert(spill && "no spill placed before reload");
2526
2527         reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
2528
2529         /* enter into the linked list */
2530         set_irn_link(reload, defs->remats);
2531         defs->remats = reload;
2532
2533         return reload;
2534 }
2535
2536 static void
2537 walker_spill_placer(ir_node * bb, void * data) {
2538         spill_ilp_t   *si = (spill_ilp_t*)data;
2539         ir_node       *irn;
2540         spill_bb_t    *spill_bb = get_irn_link(bb);
2541         pset          *spills_to_do = pset_new_ptr_default();
2542         spill_t       *spill;
2543
2544         set_foreach(spill_bb->ilp, spill) {
2545                 lpp_name_t    *name;
2546
2547                 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2548                         name = si->lpp->vars[spill->mem_in];
2549                         if(!is_zero(name->value)) {
2550                                 ir_node   *mem_phi;
2551
2552                                 mem_phi = insert_mem_phi(si, spill->irn);
2553
2554                                 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
2555                         }
2556                 }
2557
2558                 if(spill->spill != ILP_UNDEF) {
2559                         name = si->lpp->vars[spill->spill];
2560                         if(!is_zero(name->value)) {
2561                                 if(spill->reg_in > 0) {
2562                                         name = si->lpp->vars[spill->reg_in];
2563                                         if(!is_zero(name->value)) {
2564                                                 insert_spill(si, spill->irn, spill->irn, bb);
2565                                                 continue;
2566                                         }
2567                                 }
2568                                 pset_insert_ptr(spills_to_do, spill->irn);
2569                         }
2570                 }
2571         }
2572         DBG((si->dbg, LEVEL_3, "\t  %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
2573
2574
2575         for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
2576                 op_t     *op = get_irn_link(irn);
2577
2578                 if(be_is_Spill(irn)) continue;
2579
2580                 if(op->is_remat) {
2581                         /* TODO fix this if we want to support remats with more than two nodes */
2582                         if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
2583                                 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
2584
2585                                 insert_spill(si, irn, op->attr.remat.remat->value, irn);
2586                         }
2587                 } else {
2588                         if(pset_find_ptr(spills_to_do, irn)) {
2589                                 pset_remove_ptr(spills_to_do, irn);
2590
2591                                 insert_spill(si, irn, irn, irn);
2592                         }
2593                 }
2594
2595         }
2596
2597         assert(pset_count(spills_to_do) == 0);
2598
2599         /* afterwards free data in block */
2600         del_pset(spills_to_do);
2601 }
2602
2603 static void
2604 phim_fixer(spill_ilp_t *si) {
2605         defs_t  *defs;
2606
2607         set_foreach(si->values, defs) {
2608                 const ir_node  *phi = defs->value;
2609                 const ir_node  *phi_m = defs->spills;
2610                 int       i,
2611                                   n;
2612
2613                 if(!is_Phi(phi)) continue;
2614                 if(!phi_m || !is_Phi(phi_m) || get_irn_mode(phi_m) != mode_M) continue;
2615
2616                 for(i=0,n=get_irn_arity(phi); i<n; ++i) {
2617                         const ir_node  *value = get_irn_n(phi, i);
2618                         defs_t         *val_defs = set_find_def(si->values, value);
2619
2620                         /* get a spill of this value */
2621                         ir_node      *spill = val_defs->spills;
2622
2623                         assert(spill && "no spill placed before PhiM");
2624
2625                         set_irn_n(phi_m, i, spill);
2626                 }
2627         }
2628 }
2629
2630 static void
2631 walker_reload_placer(ir_node * bb, void * data) {
2632         spill_ilp_t   *si = (spill_ilp_t*)data;
2633         ir_node       *irn;
2634         spill_bb_t    *spill_bb = get_irn_link(bb);
2635         int            i;
2636         irn_live_t    *li;
2637
2638         sched_foreach_reverse(bb, irn) {
2639                 op_t     *op = get_irn_link(irn);
2640
2641                 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
2642                 if(is_Phi(irn)) break;
2643
2644                 if(op->is_remat) {
2645                         if(get_irn_mode(irn) != mode_T) {
2646                                 insert_remat(si, irn);
2647                         }
2648                 } else {
2649                         int    n;
2650
2651                         for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
2652                                 ir_node    *arg = get_irn_n(irn, i);
2653
2654                                 if(op->attr.live_range.reloads && op->attr.live_range.reloads[i] != ILP_UNDEF) {
2655                                         lpp_name_t    *name;
2656
2657                                         name = si->lpp->vars[op->attr.live_range.reloads[i]];
2658                                         if(!is_zero(name->value)) {
2659                                                 ir_node    *reload;
2660                                                 ir_node    *insert_pos = irn;
2661                                                 ir_node    *prev = sched_prev(insert_pos);
2662                                                 op_t       *prev_op = get_irn_link(prev);
2663
2664                                                 /* insert reload before pre-remats */
2665                                                 while(!sched_is_end(prev) && !be_is_Reload(prev) && !be_is_Spill(prev)
2666                                                                 && prev_op->is_remat && prev_op->attr.remat.pre) {
2667                                                         insert_pos = prev;
2668
2669                                                         prev = sched_prev(insert_pos);
2670                                                         prev_op = get_irn_link(prev);
2671                                                 }
2672
2673                                                 reload = insert_reload(si, arg, insert_pos);
2674
2675                                                 set_irn_n(irn, i, reload);
2676
2677 #ifdef KEEPALIVE_RELOADS
2678                                                 pset_insert_ptr(si->spills, reload);
2679 #endif
2680                                         }
2681                                 }
2682                         }
2683                 }
2684         }
2685
2686         /* reloads at end of block */
2687         if(spill_bb->reloads) {
2688                 i=0;
2689                 live_foreach(bb, li) {
2690                         ir_node        *irn = (ir_node *) li->irn;
2691
2692                         if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
2693                                 lpp_name_t    *name;
2694
2695                                 name = si->lpp->vars[spill_bb->reloads[i]];
2696                                 if(!is_zero(name->value)) {
2697                                         ir_node    *reload;
2698                                         ir_node    *insert_pos = bb;
2699                                         ir_node    *prev = sched_prev(insert_pos);
2700                                         op_t       *prev_op = get_irn_link(prev);
2701
2702                                         /* insert reload before pre-remats */
2703                                         while(!sched_is_end(prev) && !be_is_Reload(prev) && !be_is_Spill(prev)
2704                                                         && prev_op->is_remat && prev_op->attr.remat.pre) {
2705                                                 insert_pos = prev;
2706
2707                                                 prev = sched_prev(insert_pos);
2708                                                 prev_op = get_irn_link(prev);
2709                                         }
2710
2711                                         reload = insert_reload(si, irn, insert_pos);
2712
2713 #ifdef KEEPALIVE_RELOADS
2714                                         pset_insert_ptr(si->spills, reload);
2715 #endif
2716                                 }
2717                                 ++i;
2718                         }
2719                 }
2720         }
2721
2722         del_set(spill_bb->ilp);
2723 }
2724
2725 static void
2726 walker_collect_used(ir_node * irn, void * data)
2727 {
2728         lc_bitset_t   *used = data;
2729
2730         lc_bitset_set(used, get_irn_idx(irn));
2731 }
2732
2733 static void
2734 walker_kill_unused(ir_node * bb, void * data)
2735 {
2736         lc_bitset_t     *used = data;
2737         const ir_node   *bad = get_irg_bad(get_irn_irg(bb));
2738         ir_node         *irn;
2739
2740
2741         for(irn=sched_first(bb); !sched_is_end(irn);) {
2742                 ir_node     *next = sched_next(irn);
2743                 int          i,
2744                                          n;
2745
2746                 if(!lc_bitset_is_set(used, get_irn_idx(irn))) {
2747                         assert(!be_is_Spill(irn) && !be_is_Reload(irn) && "something is fishy, spill or remat is unused");
2748
2749                         sched_remove(irn);
2750
2751                         set_nodes_block(irn, bad);
2752                         for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
2753                                 set_irn_n(irn, i, bad);
2754                         }
2755                 }
2756                 irn = next;
2757         }
2758 }
2759
2760 static void
2761 kill_all_unused_values_in_schedule(spill_ilp_t * si)
2762 {
2763         lc_bitset_t   *used = lc_bitset_malloc(get_irg_last_idx(si->chordal_env->irg));
2764
2765         irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, used);
2766         irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, used);
2767
2768         lc_bitset_free(used);
2769 }
2770
2771 static void
2772 print_irn_pset(pset * p)
2773 {
2774         ir_node   *irn;
2775
2776         pset_foreach(p, irn) {
2777                 ir_printf("%+F\n", irn);
2778         }
2779 }
2780
2781 static void
2782 rewire_uses(spill_ilp_t * si)
2783 {
2784         dom_front_info_t     *dfi = be_compute_dominance_frontiers(si->chordal_env->irg);
2785         defs_t               *defs;
2786
2787         /* then fix uses of spills */
2788         set_foreach(si->values, defs) {
2789                 pset     *reloads;
2790                 pset     *spills;
2791                 ir_node  *next = defs->remats;
2792                 int remats = 0;
2793
2794                 if(next) {
2795                         reloads = pset_new_ptr_default();
2796
2797                         while(next) {
2798                                 if(be_is_Reload(next)) {
2799                                         pset_insert_ptr(reloads, next);
2800                                 } else {
2801                                         ++remats;
2802                                 }
2803                                 next = get_irn_link(next);
2804                         }
2805
2806                         spills = get_spills_for_value(si, defs->value);
2807                         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));
2808                         if(pset_count(spills) > 1) {
2809                                 assert(pset_count(reloads) > 0);
2810 //                              print_irn_pset(spills);
2811 //                              print_irn_pset(reloads);
2812
2813 //                              be_ssa_constr_set_uses(dfi, spills, reloads);
2814                                 be_ssa_constr_set(dfi, spills);
2815                         }
2816
2817                         del_pset(reloads);
2818                         del_pset(spills);
2819                 }
2820         }
2821
2822         /* first fix uses of remats and reloads */
2823         set_foreach(si->values, defs) {
2824                 pset     *nodes;
2825                 ir_node  *next = defs->remats;
2826
2827                 if(next) {
2828                         nodes = pset_new_ptr_default();
2829                         pset_insert_ptr(nodes, defs->value);
2830
2831                         while(next) {
2832                                 pset_insert_ptr(nodes, next);
2833                                 next = get_irn_link(next);
2834                         }
2835
2836                         if(pset_count(nodes) > 1) {
2837                                 DBG((si->dbg, LEVEL_4, "\t    %d new definitions for value %+F\n", pset_count(nodes)-1, defs->value));
2838                                 be_ssa_constr_set(dfi, nodes);
2839                         }
2840
2841                         del_pset(nodes);
2842                 }
2843         }
2844
2845 //      remove_unused_defs(si);
2846
2847         be_free_dominance_frontiers(dfi);
2848 }
2849
2850 static void
2851 writeback_results(spill_ilp_t * si)
2852 {
2853         /* walk through the graph and collect all spills, reloads and remats for a value */
2854
2855         si->values = new_set(cmp_defs, 4096);
2856
2857         DBG((si->dbg, LEVEL_1, "Applying results\n"));
2858         delete_unnecessary_remats(si);
2859         si->m_unknown = new_r_Unknown(si->chordal_env->irg, mode_M);
2860         irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
2861         phim_fixer(si);
2862         irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
2863
2864         /* clean the remat info! there are still back-edges leading there! */
2865         clean_remat_info(si);
2866
2867         rewire_uses(si);
2868
2869         connect_all_spills_with_keep(si);
2870
2871         del_set(si->values);
2872 }
2873
2874 static int
2875 get_n_regs(spill_ilp_t * si)
2876 {
2877         int     arch_n_regs = arch_register_class_n_regs(si->cls);
2878         int     free = 0;
2879         int     i;
2880
2881         for(i=0; i<arch_n_regs; i++) {
2882                 if(!arch_register_type_is(&si->cls->regs[i], ignore)) {
2883                         free++;
2884                 }
2885         }
2886
2887         DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", free, si->cls->name));
2888         return free;
2889 }
2890
2891 static void
2892 walker_reload_mover(ir_node * bb, void * data)
2893 {
2894         spill_ilp_t   *si = data;
2895         ir_node           *tmp;
2896
2897         sched_foreach(bb, tmp) {
2898                 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
2899                         ir_node       *reload = tmp;
2900                         ir_node       *irn = tmp;
2901
2902                         /* move reload upwards */
2903
2904                         int pressure = (int)get_irn_link(reload);
2905                         if(pressure < si->n_regs) {
2906                                 irn = sched_prev(reload);
2907                                 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
2908                                 sched_remove(reload);
2909                                 pressure = (int)get_irn_link(irn);
2910
2911                                 while(pressure < si->n_regs) {
2912                                         if(sched_is_end(irn) || (be_is_Reload(irn) && has_reg_class(si, irn))) break;
2913
2914                                         set_irn_link(irn, INT_TO_PTR(pressure+1));
2915                                         DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
2916                                         irn = sched_prev(irn);
2917
2918                                         pressure = (int)get_irn_link(irn);
2919                                 }
2920
2921                                 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
2922                                 sched_put_after(irn, reload);
2923                         }
2924                 }
2925         }
2926 }
2927
2928 static void
2929 move_reloads_upward(spill_ilp_t * si)
2930 {
2931         irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si);
2932 }
2933
2934 void
2935 be_spill_remat(const be_chordal_env_t * chordal_env)
2936 {
2937         char            problem_name[256];
2938         char            dump_suffix[256];
2939         char            dump_suffix2[256];
2940         char            dump_suffix3[256];
2941         struct obstack  obst;
2942         spill_ilp_t     si;
2943
2944         ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name);
2945         ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name);
2946         ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name);
2947         ir_snprintf(dump_suffix3, sizeof(dump_suffix3), "-%s-reloads_moved", chordal_env->cls->name);
2948
2949         FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
2950         DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
2951
2952         obstack_init(&obst);
2953         si.chordal_env = chordal_env;
2954         si.obst = &obst;
2955         si.senv = be_new_spill_env(chordal_env, is_mem_phi, &si);
2956         si.cls = chordal_env->cls;
2957         si.lpp = new_lpp(problem_name, lpp_minimize);
2958         si.remat_info = new_set(cmp_remat_info, 4096);
2959         si.all_possible_remats = pset_new_ptr_default();
2960         si.spills = pset_new_ptr_default();
2961         si.inverse_ops = pset_new_ptr_default();
2962 #ifndef EXECFREQ_LOOPDEPH
2963         si.execfreqs = compute_execfreq(chordal_env->irg);
2964 #else
2965         si.execfreqs = NULL;
2966 #endif
2967 #ifdef KEEPALIVE
2968         si.keep = NULL;
2969 #endif
2970         si.n_regs = get_n_regs(&si);
2971
2972         set_irg_link(chordal_env->irg, &si);
2973         compute_doms(chordal_env->irg);
2974
2975 #ifdef COLLECT_REMATS
2976         /* collect remats */
2977         DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
2978         irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
2979 #endif
2980
2981         /* insert possible remats */
2982         DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
2983         irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
2984         DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
2985
2986 #ifdef KEEPALIVE
2987         DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
2988         connect_all_remats_with_keep(&si);
2989         /* dump graph with inserted remats */
2990         dump_graph_with_remats(chordal_env->irg, dump_suffix);
2991 #endif
2992
2993
2994         /* recompute liveness */
2995         DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
2996         be_liveness(chordal_env->irg);
2997
2998         /* build the ILP */
2999
3000         DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
3001         DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
3002         irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si);
3003
3004         DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
3005         irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
3006
3007 #ifdef DUMP_ILP
3008         {
3009                 FILE           *f;
3010                 char            buf[256];
3011
3012                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
3013                 if ((f = fopen(buf, "wt")) != NULL) {
3014                         lpp_dump_plain(si.lpp, f);
3015                         fclose(f);
3016                 }
3017         }
3018 #endif
3019
3020 #ifdef SOLVE
3021         DBG((si.dbg, LEVEL_1, "\tSolving %F\n", chordal_env->irg));
3022         lpp_set_time_limit(si.lpp, ILP_TIMEOUT);
3023
3024 #ifdef SOLVE_LOCAL
3025         lpp_solve_cplex(si.lpp);
3026 #else
3027         lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
3028 #endif
3029         assert(lpp_is_sol_valid(si.lpp)
3030                && "solution of ILP must be valid");
3031
3032         DBG((si.dbg, LEVEL_1, "\t%s: iterations: %d, solution time: %g, objective function: %g\n", problem_name, si.lpp->iterations, si.lpp->sol_time, is_zero(si.lpp->objval)?0.0:si.lpp->objval));
3033
3034 #ifdef DUMP_SOLUTION
3035         {
3036                 FILE           *f;
3037                 char            buf[256];
3038
3039                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
3040                 if ((f = fopen(buf, "wt")) != NULL) {
3041                         int             i;
3042                         for (i = 0; i < si.lpp->var_next; ++i) {
3043                                 lpp_name_t     *name = si.lpp->vars[i];
3044                                 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
3045                         }
3046                         fclose(f);
3047                 }
3048         }
3049 #endif
3050
3051         writeback_results(&si);
3052
3053 #endif                          /* SOLVE */
3054
3055         kill_all_unused_values_in_schedule(&si);
3056
3057 //      be_dump(chordal_env->irg, "-bla", dump_ir_block_graph);
3058
3059         be_liveness(chordal_env->irg);
3060         irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
3061
3062         dump_pressure_graph(&si, dump_suffix2);
3063
3064         // TODO fix temporarily exceeded regpressure due to remat2s
3065
3066         // TODO insert copys to fix interferences in memory
3067
3068         // move reloads upwards
3069         move_reloads_upward(&si);
3070         irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
3071         dump_pressure_graph(&si, dump_suffix3);
3072
3073         free_dom(chordal_env->irg);
3074         del_pset(si.inverse_ops);
3075         del_pset(si.all_possible_remats);
3076         del_pset(si.spills);
3077 #ifndef EXECFREQ_LOOPDEPH
3078         free_execfreq(si.execfreqs);
3079 #endif
3080         free_lpp(si.lpp);
3081         obstack_free(&obst, NULL);
3082 //      exit(0);
3083 }
3084
3085 #else                           /* WITH_ILP */
3086
3087 static void
3088 only_that_you_can_compile_without_WITH_ILP_defined(void)
3089 {
3090 }
3091
3092 #endif                          /* WITH_ILP */