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