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