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

   1 /*
   2  * Author:      Matthias Braun
   3  * Date:                05.05.2006
   4  * Copyright:   (c) Universitaet Karlsruhe
   5  * License:     This file is protected by GPL -  GNU GENERAL PUBLIC LICENSE.
   6  *
   7  */
   8 #ifdef HAVE_CONFIG_H
   9 #include "config.h"
  10 #endif
  11
  12 #include "bespillmorgan.h"
  13
  14 #include "bechordal_t.h"
  15 #include "bespill.h"
  16 #include "belive_t.h"
  17 #include "beabi.h"
  18 #include "irgwalk.h"
  19 #include "besched.h"
  20 #include "beutil.h"
  21 #include "irloop_t.h"
  22 #include "irgraph_t.h"
  23 #include "irprintf.h"
  24 #include "obstack.h"
  25
  26 #include "bespillbelady.h"
  27 #include "beverify.h"
  28 #include "benodesets.h"
  29
  30 #define DBG_LIVE                1
  31 #define DBG_LOOPANA             2
  32 #define DBG_PRESSURE    4
  33 #define DBG_SPILLS      8
  34 #define DBG_CHOOSE              16
  35 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  36
  37 typedef struct morgan_env {
  38         const be_chordal_env_t *cenv;
  39         const arch_env_t *arch;
  40         const arch_register_class_t *cls;
  41         ir_graph *irg;
  42         struct obstack obst;
  43         /** maximum safe register pressure */
  44         int registers_available;
  45
  46         spill_env_t *senv;
  47
  48         set *loop_attr_set;
  49         set *block_attr_set;
  50 } morgan_env_t;
  51
  52 typedef struct loop_edge {
  53         ir_node *block;
  54         int pos;
  55 } loop_edge_t;
  56
  57 typedef struct loop_attr {
  58         const ir_loop *loop;
  59         set *out_edges;
  60         set *in_edges;
  61         /** The set of all values that are live in the loop but not used in the loop */
  62         bitset_t *livethrough_unused;
  63 } loop_attr_t;
  64
  65 typedef struct block_attr {
  66         const ir_node *block;
  67         /** set of all values that are live in the block but not used in the block */
  68         bitset_t *livethrough_unused;
  69 } block_attr_t;
  70
  71 //---------------------------------------------------------------------------
  72
  73 static int loop_edge_cmp(const void* p1, const void* p2, size_t s) {
  74         loop_edge_t *e1 = (loop_edge_t*) p1;
  75         loop_edge_t *e2 = (loop_edge_t*) p2;
  76
  77         return e1->block != e2->block || e1->pos != e2->pos;
  78 }
  79
  80 static int loop_attr_cmp(const void *e1, const void *e2, size_t s) {
  81         loop_attr_t *la1 = (loop_attr_t*) e1;
  82         loop_attr_t *la2 = (loop_attr_t*) e2;
  83
  84         return la1->loop != la2->loop;
  85 }
  86
  87 static int block_attr_cmp(const void *e1, const void *e2, size_t s) {
  88         block_attr_t *b1 = (block_attr_t*) e1;
  89         block_attr_t *b2 = (block_attr_t*) e2;
  90
  91         return b1->block != b2->block;
  92 }
  93
  94 static INLINE int loop_attr_hash(const loop_attr_t *a) {
  95 #ifdef DEBUG_libfirm
  96         return a->loop->loop_nr;
  97 #else
  98         return HASH_PTR(a->loop);
  99 #endif
 100 }
 101
 102 static INLINE int block_attr_hash(const block_attr_t *b) {
 103         return nodeset_hash(b->block);
 104 }
 105
 106 static INLINE int loop_edge_hash(const loop_edge_t *e) {
 107         return nodeset_hash(e->block) ^ (e->pos * 31);
 108 }
 109
 110 static INLINE loop_attr_t *get_loop_attr(morgan_env_t *env, const ir_loop *loop) {
 111         loop_attr_t l_attr, *res;
 112         int hash;
 113         l_attr.loop = loop;
 114
 115         hash = loop_attr_hash(&l_attr);
 116         res = set_find(env->loop_attr_set, &l_attr, sizeof(l_attr), hash);
 117
 118         // create new loop_attr if none exists yet
 119         if (res == NULL) {
 120                 l_attr.out_edges = new_set(loop_edge_cmp, 1);
 121                 l_attr.in_edges = new_set(loop_edge_cmp, 1);
 122                 l_attr.livethrough_unused = bitset_obstack_alloc(&env->obst, get_irg_last_idx(env->irg));
 123                 res = set_insert(env->loop_attr_set, &l_attr, sizeof(l_attr), hash);
 124         }
 125
 126         return res;
 127 }
 128
 129 static INLINE block_attr_t *get_block_attr(morgan_env_t *env, const ir_node *block) {
 130         block_attr_t b_attr, *res;
 131         int hash;
 132         b_attr.block = block;
 133
 134         hash = block_attr_hash(&b_attr);
 135         res = set_find(env->block_attr_set, &b_attr, sizeof(b_attr), hash);
 136
 137         if(res == NULL) {
 138                 b_attr.livethrough_unused = NULL;
 139                 res = set_insert(env->block_attr_set, &b_attr, sizeof(b_attr), hash);
 140         }
 141
 142         return res;
 143 }
 144
 145 //---------------------------------------------------------------------------
 146
 147 static INLINE int consider_for_spilling(const arch_env_t *env, const arch_register_class_t *cls, const ir_node *node) {
 148         if(!arch_irn_has_reg_class(env, node, -1, cls))
 149                 return 0;
 150
 151         return !(arch_irn_get_flags(env, node) & (arch_irn_flags_ignore | arch_irn_flags_dont_spill));
 152 }
 153
 154 /**
 155  * Determine edges going out of a loop (= edges that go to a block that is not inside
 156  * the loop or one of its subloops)
 157  */
 158 static INLINE void construct_loop_edges(ir_node* block, void* data) {
 159         morgan_env_t *env = data;
 160         int n_cfgpreds = get_Block_n_cfgpreds(block);
 161         int i;
 162         ir_loop* loop = get_irn_loop(block);
 163         DBG((dbg, DBG_LOOPANA, "Loop for %+F: %d (depth %d)\n", block, loop->loop_nr, loop->depth));
 164
 165         for(i = 0; i < n_cfgpreds; ++i) {
 166                 loop_edge_t edge;
 167                 int hash;
 168                 ir_node* cfgpred = get_Block_cfgpred(block, i);
 169                 ir_node* cfgpred_block = get_nodes_block(cfgpred);
 170                 ir_loop* cfgpred_loop = get_irn_loop(cfgpred_block);
 171
 172                 if(cfgpred_loop == loop)
 173                         continue;
 174
 175                 assert(get_loop_depth(cfgpred_loop) != get_loop_depth(loop));
 176
 177                 edge.block = block;
 178                 edge.pos = i;
 179                 hash = loop_edge_hash(&edge);
 180
 181                 // edge out of a loop?
 182                 if(get_loop_depth(cfgpred_loop) > get_loop_depth(loop)) {
 183                         ir_loop *l;
 184
 185                         DBG((dbg, DBG_LOOPANA, "Loop out edge from %+F (loop %d) to %+F (loop %d)\n", block, get_loop_loop_nr(loop),
 186                              cfgpred_block, get_loop_loop_nr(cfgpred_loop)));
 187
 188                         /* this might be a jump out of multiple loops, so add this to all
 189                      * needed outedge sets */
 190                         l = cfgpred_loop;
 191                         do {
 192                                 loop_attr_t *l_attr = get_loop_attr(env, l);
 193                                 set_insert(l_attr->out_edges, &edge, sizeof(edge), hash);
 194
 195                                 l = get_loop_outer_loop(l);
 196                                 assert(l != NULL);
 197                         } while(l != loop);
 198                 } else {
 199                         ir_loop *l;
 200
 201                         // edge into a loop
 202                         DBG((dbg, DBG_LOOPANA, "Loop in edge from %+F (loop %d) to %+F (loop %d)\n", block, get_loop_loop_nr(loop),
 203                              cfgpred_block, get_loop_loop_nr(cfgpred_loop)));
 204
 205                         l = loop;
 206                         do {
 207                                 loop_attr_t *l_attr = get_loop_attr(env, l);
 208                                 set_insert(l_attr->in_edges, &edge, sizeof(edge), hash);
 209
 210                                 l = get_loop_outer_loop(l);
 211                         } while(l != cfgpred_loop);
 212                 }
 213         }
 214 }
 215
 216 static void free_loop_edges(morgan_env_t *env) {
 217         loop_attr_t *l_attr;
 218
 219         for(l_attr = set_first(env->loop_attr_set); l_attr != NULL; l_attr = set_next(env->loop_attr_set)) {
 220                 del_set(l_attr->out_edges);
 221                 del_set(l_attr->in_edges);
 222         }
 223 }
 224
 225 #if 0
 226 /**
 227  * Debugging help, shows all nodes in a (node-)bitset
 228  */
 229 static void show_nodebitset(ir_graph* irg, const bitset_t* bitset) {
 230         int i;
 231
 232         bitset_foreach(bitset, i) {
 233                 ir_node* node = get_idx_irn(irg, i);
 234                 ir_fprintf(stderr, " %+F", node);
 235         }
 236         fprintf(stderr, "\n");
 237 }
 238 #endif
 239
 240 static INLINE void init_livethrough_unuseds(block_attr_t *attr, morgan_env_t *env) {
 241         const ir_node *block;
 242         int i;
 243
 244         if(attr->livethrough_unused != NULL)
 245                 return;
 246
 247         block = attr->block;
 248
 249         attr->livethrough_unused = bitset_obstack_alloc(&env->obst, get_irg_last_idx(env->irg));
 250
 251         // copy all live-outs into the livethrough_unused set
 252         be_lv_foreach(env->cenv->lv, block, be_lv_state_in | be_lv_state_out, i) {
 253                 ir_node *irn = be_lv_get_irn(env->cenv->lv, block, i);
 254                 int node_idx;
 255
 256                 if(!consider_for_spilling(env->arch, env->cls, irn))
 257                         continue;
 258
 259                 node_idx = get_irn_idx(irn);
 260                 bitset_set(attr->livethrough_unused, node_idx);
 261         }
 262 }
 263
 264 /**
 265  * Construct the livethrough unused set for a block
 266  */
 267 static void construct_block_livethrough_unused(ir_node *block, void *data) {
 268         morgan_env_t* env = data;
 269         block_attr_t *block_attr = get_block_attr(env, block);
 270         ir_node *node;
 271         int n_cfgpreds;
 272         block_attr_t **pred_attrs = NULL;
 273         int i;
 274
 275         init_livethrough_unuseds(block_attr, env);
 276
 277         DBG((dbg, DBG_LIVE, "Processing block %d\n", get_irn_node_nr(block)));
 278
 279         n_cfgpreds = get_Block_n_cfgpreds(block);
 280         if(n_cfgpreds > 1) {
 281                 pred_attrs = alloca(sizeof(pred_attrs[0]) * n_cfgpreds);
 282                 for(i = 0; i < n_cfgpreds; ++i) {
 283                         ir_node *pred_block = get_Block_cfgpred_block(block, i);
 284                         pred_attrs[i] = get_block_attr(env, pred_block);
 285                         init_livethrough_unuseds(pred_attrs[i], env);
 286                 }
 287         }
 288
 289         /*
 290          * All values that are used within the block are not unused (and therefore not
 291          * livethrough_unused)
 292          */
 293         sched_foreach(block, node) {
 294                 int i, arity;
 295
 296                 // phis are really uses in the pred block
 297                 if(is_Phi(node)) {
 298                         int j;
 299                         for(j = 0; j < n_cfgpreds; ++j) {
 300                                 ir_node *used_value = get_Phi_pred(node, j);
 301                                 int idx = get_irn_idx(used_value);
 302                                 block_attr_t *pred_attr = pred_attrs[j];
 303
 304                                 bitset_clear(pred_attr->livethrough_unused, idx);
 305                         }
 306                 } else {
 307                         // mark all used values as used
 308                         for(i = 0, arity = get_irn_arity(node); i < arity; ++i) {
 309                                 int idx = get_irn_idx(get_irn_n(node, i));
 310                                 bitset_clear(block_attr->livethrough_unused, idx);
 311                         }
 312                 }
 313         }
 314 }
 315
 316 /**
 317  * Construct the livethrough unused set for a loop (and all its subloops+blocks)
 318  */
 319 static bitset_t *construct_loop_livethrough_unused(morgan_env_t *env, const ir_loop *loop) {
 320         int i;
 321         loop_attr_t* loop_attr = get_loop_attr(env, loop);
 322
 323         DBG((dbg, DBG_LIVE, "Processing Loop %d\n", loop->loop_nr));
 324         assert(get_loop_n_elements(loop) > 0);
 325         for(i = 0; i < get_loop_n_elements(loop); ++i) {
 326                 loop_element elem = get_loop_element(loop, i);
 327                 switch (*elem.kind) {
 328                 case k_ir_node: {
 329                         ir_node *block = elem.node;
 330                         block_attr_t *block_attr = get_block_attr(env, block);
 331                         bitset_t *livethrough_block_unused = block_attr->livethrough_unused;
 332
 333                         assert(is_Block(elem.node));
 334                         assert(livethrough_block_unused != NULL);
 335
 336                         if(i == 0) {
 337                                 bitset_copy(loop_attr->livethrough_unused, livethrough_block_unused);
 338                         } else {
 339                                 bitset_and(loop_attr->livethrough_unused, livethrough_block_unused);
 340                         }
 341                         break;
 342                 }
 343                 case k_ir_loop: {
 344                         bitset_t *livethrough_son_unused;
 345
 346                         livethrough_son_unused = construct_loop_livethrough_unused(env, elem.son);
 347                         if(i == 0) {
 348                                 bitset_copy(loop_attr->livethrough_unused, livethrough_son_unused);
 349                         } else {
 350                                 bitset_and(loop_attr->livethrough_unused, livethrough_son_unused);
 351                         }
 352                         break;
 353                 }
 354             default:
 355                         assert(0);
 356                         break;
 357                 }
 358     }
 359         DBG((dbg, DBG_LIVE, "Done with loop %d\n", loop->loop_nr));
 360
 361         // remove all unused livethroughs that are remembered for this loop from child loops and blocks
 362         for(i = 0; i < get_loop_n_elements(loop); ++i) {
 363                 const loop_element elem = get_loop_element(loop, i);
 364
 365                 if(*elem.kind == k_ir_loop) {
 366                         loop_attr_t *son_attr = get_loop_attr(env, elem.son);
 367                         bitset_andnot(son_attr->livethrough_unused, loop_attr->livethrough_unused);
 368
 369                         DBG((dbg, DBG_LIVE, "Livethroughs for loop %d:\n", loop->loop_nr));
 370                 } else if(*elem.kind == k_ir_node) {
 371                         block_attr_t *block_attr = get_block_attr(env, elem.node);
 372                         bitset_andnot(block_attr->livethrough_unused, loop_attr->livethrough_unused);
 373
 374                         DBG((dbg, DBG_LIVE, "Livethroughs for block %+F\n", elem.node));
 375                 } else {
 376                         assert(0);
 377                 }
 378         }
 379
 380         return loop_attr->livethrough_unused;
 381 }
 382
 383 /*---------------------------------------------------------------------------*/
 384
 385 typedef struct _spillcandidate_t {
 386         ir_node *node;
 387         int cost;
 388 } spillcandidate_t;
 389
 390 static int compare_spillcandidates(const void *d1, const void *d2) {
 391         const spillcandidate_t *cand1 = d1;
 392         const spillcandidate_t *cand2 = d2;
 393
 394         return cand1->cost - cand2->cost;
 395 }
 396
 397 static void spill_values(morgan_env_t *env, const loop_attr_t *loop_attr, int spills) {
 398         const bitset_t *cand_bitset = loop_attr->livethrough_unused;
 399         int candidatecount = bitset_popcnt(cand_bitset);
 400         spillcandidate_t *candidates;
 401         int i, c;
 402         loop_edge_t *edge;
 403
 404         assert(spills <= candidatecount);
 405
 406         candidates = alloca(sizeof(candidates[0]) * candidatecount);
 407
 408         DBG((dbg, DBG_CHOOSE, "Candidates for loop %d\n", get_loop_loop_nr(loop_attr->loop)));
 409         // build candidiatelist
 410         c = 0;
 411         bitset_foreach(cand_bitset, i) {
 412                 ir_node *node = get_idx_irn(env->irg, i);
 413                 candidates[c].node = node;
 414                 candidates[c].cost = 0;
 415
 416                 for(edge = set_first(loop_attr->out_edges); edge != NULL; edge = set_next(loop_attr->out_edges)) {
 417                         candidates[c].cost += be_get_reload_costs_on_edge(env->senv, node, edge->block, edge->pos);
 418                 }
 419                 DBG((dbg, DBG_CHOOSE, "%+F has costs %d\n", node, candidates[c].cost));
 420
 421                 c++;
 422         }
 423         assert(c == candidatecount);
 424
 425         // sort list
 426         qsort(candidates, candidatecount, sizeof(candidates[0]), compare_spillcandidates);
 427
 428         // spill values
 429         for(i = 0; i < spills; ++i) {
 430                 ir_node *to_spill = candidates[i].node;
 431                 DBG((dbg, DBG_CHOOSE, "Spilling %+F ", to_spill));
 432
 433                 for(edge = set_first(loop_attr->out_edges); edge != NULL; edge = set_next(loop_attr->out_edges)) {
 434                         be_add_reload_on_edge(env->senv, to_spill, edge->block, edge->pos);
 435                 }
 436         }
 437 }
 438
 439 static int reduce_register_pressure_in_block(morgan_env_t *env, const ir_node* block, int loop_unused_spills_possible) {
 440         ir_node *node;
 441         int max_pressure;
 442         int loop_unused_spills_needed;
 443         pset *live_nodes = pset_new_ptr_default();
 444
 445         be_liveness_end_of_block(env->cenv->lv, env->arch, env->cls, block, live_nodes);
 446         max_pressure = pset_count(live_nodes);
 447
 448         DBG((dbg, DBG_LIVE, "Reduce pressure to %d In Block %+F:\n", env->registers_available, block));
 449
 450         /**
 451          * Determine register pressure in block
 452          */
 453         sched_foreach_reverse(block, node) {
 454                 int pressure;
 455
 456                 if(is_Phi(node))
 457                         break;
 458
 459                 be_liveness_transfer(env->arch, env->cls, node, live_nodes);
 460                 pressure = pset_count(live_nodes);
 461                 if(pressure > max_pressure)
 462                         max_pressure = pressure;
 463         }
 464         del_pset(live_nodes);
 465
 466         loop_unused_spills_needed = max_pressure - env->registers_available;
 467
 468         if(loop_unused_spills_needed < 0) {
 469                 loop_unused_spills_needed = 0;
 470         } else if(loop_unused_spills_needed > loop_unused_spills_possible) {
 471                 loop_unused_spills_needed = loop_unused_spills_possible;
 472         }
 473
 474         DBG((dbg, DBG_PRESSURE, "Block %+F: max-pressure %d spills possible: %d spills used: %d\n",
 475                  block, max_pressure, loop_unused_spills_possible, loop_unused_spills_needed));
 476         return loop_unused_spills_needed;
 477 }
 478
 479 /**
 480  * Reduce register pressure in a loop
 481  *
 482  * @param unused_spills_possible        Number of spills from livethrough_unused variables possible in outer loops
 483  * @return                                                      Number of spills of livethrough_unused variables needed in outer loops
 484  */
 485 static int reduce_register_pressure_in_loop(morgan_env_t *env, const ir_loop *loop, int outer_spills_possible) {
 486         int i;
 487         loop_attr_t* loop_attr = get_loop_attr(env, loop);
 488         int spills_needed = 0;
 489         int spills_possible = outer_spills_possible + bitset_popcnt(loop_attr->livethrough_unused);
 490         int outer_spills_needed;
 491
 492         DBG((dbg, DBG_PRESSURE, "Reducing Pressure in loop %d\n", loop->loop_nr));
 493         for(i = 0; i < get_loop_n_elements(loop); ++i) {
 494                 loop_element elem = get_loop_element(loop, i);
 495                 switch (*elem.kind) {
 496                 case k_ir_node: {
 497                         int needed;
 498                         assert(is_Block(elem.node));
 499                         needed = reduce_register_pressure_in_block(env, elem.node, spills_possible);
 500                         assert(needed >= 0);
 501                         assert(needed <= spills_possible);
 502                         if(needed > spills_needed)
 503                                 spills_needed = needed;
 504                         break;
 505                 }
 506                 case k_ir_loop: {
 507                         int needed = reduce_register_pressure_in_loop(env, elem.son, spills_possible);
 508                         assert(needed >= 0);
 509                         assert(needed <= spills_possible);
 510                         if(needed > spills_needed)
 511                                 spills_needed = needed;
 512                         break;
 513                 }
 514             default:
 515                         assert(0);
 516                         break;
 517                 }
 518     }
 519
 520         /* calculate number of spills needed in outer loop and spill
 521          * unused livethrough nodes around this loop
 522          */
 523         if(spills_needed > outer_spills_possible) {
 524                 int spills_to_place;
 525                 outer_spills_needed = outer_spills_possible;
 526                 spills_needed -= outer_spills_possible;
 527
 528                 spills_to_place = spills_needed;
 529
 530                 DBG((dbg, DBG_SPILLS, "%d values unused in loop %d, spilling %d\n",
 531                  spills_possible - outer_spills_possible, loop->loop_nr, spills_to_place));
 532
 533                 spill_values(env, loop_attr, spills_to_place);
 534         } else {
 535                 outer_spills_needed = spills_needed;
 536         }
 537
 538         return outer_spills_needed;
 539 }
 540
 541 void be_spill_morgan(be_chordal_env_t *chordal_env) {
 542         ir_graph *irg = chordal_env->irg;
 543         morgan_env_t env;
 544
 545         FIRM_DBG_REGISTER(dbg, "ir.be.spillmorgan");
 546         //firm_dbg_set_mask(dbg, DBG_SPILLS | DBG_LOOPANA);
 547
 548         env.cenv = chordal_env;
 549         env.arch = chordal_env->birg->main_env->arch_env;
 550         env.irg = chordal_env->irg;
 551         env.cls = chordal_env->cls;
 552         env.senv = be_new_spill_env(chordal_env);
 553         DEBUG_ONLY(be_set_spill_env_dbg_module(env.senv, dbg);)
 554
 555         obstack_init(&env.obst);
 556
 557         env.registers_available = env.cls->n_regs - be_put_ignore_regs(chordal_env->birg, env.cls, NULL);
 558
 559         env.loop_attr_set = new_set(loop_attr_cmp, 5);
 560         env.block_attr_set = new_set(block_attr_cmp, 20);
 561
 562         /*-- Part1: Analysis --*/
 563         //Matze: I hope liveness information is up to date at this point...
 564         //be_liveness_recompute(chordal_env->lv);
 565
 566         /* construct control flow loop tree */
 567         if(! (get_irg_loopinfo_state(irg) & loopinfo_cf_consistent)) {
 568                 construct_cf_backedges(irg);
 569         }
 570
 571         /* construct loop out edges and livethrough_unused sets for loops and blocks */
 572         irg_block_walk_graph(irg, construct_block_livethrough_unused, construct_loop_edges, &env);
 573         construct_loop_livethrough_unused(&env, get_irg_loop(irg));
 574
 575         /*-- Part2: Transformation --*/
 576
 577         /* spill unused livethrough values around loops and blocks where
 578          * the pressure is too high
 579          */
 580         reduce_register_pressure_in_loop(&env, get_irg_loop(irg), 0);
 581
 582         /* Insert real spill/reload nodes and fix usages */
 583         be_insert_spills_reloads(env.senv);
 584
 585         /* Verify the result */
 586         if (chordal_env->opts->vrfy_option == BE_CH_VRFY_WARN) {
 587                 be_verify_schedule(irg);
 588         } else if (chordal_env->opts->vrfy_option == BE_CH_VRFY_ASSERT) {
 589                 assert(be_verify_schedule(irg));
 590         }
 591
 592         if (chordal_env->opts->dump_flags & BE_CH_DUMP_SPILL)
 593                 be_dump(irg, "-spillmorgan", dump_ir_block_graph_sched);
 594
 595         /* cleanup */
 596         free_loop_edges(&env);
 597         del_set(env.loop_attr_set);
 598         del_set(env.block_attr_set);
 599
 600         /* fix the remaining places with too high register pressure with beladies algorithm */
 601         be_spill_belady_spill_env(chordal_env, env.senv);
 602
 603         be_delete_spill_env(env.senv);
 604         obstack_free(&env.obst, NULL);
 605 }