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 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  34
  35 typedef struct morgan_env {
  36         const be_chordal_env_t *cenv;
  37         const arch_env_t *arch;
  38         const arch_register_class_t *cls;
  39         ir_graph *irg;
  40         struct obstack obst;
  41         /** maximum safe register pressure */
  42         int registers_available;
  43
  44         spill_env_t *senv;
  45
  46         set *loop_attr_set;
  47         set *block_attr_set;
  48 } morgan_env_t;
  49
  50 typedef struct loop_edge {
  51         ir_node *block;
  52         int pos;
  53 } loop_edge_t;
  54
  55 typedef struct loop_attr {
  56         const ir_loop *loop;
  57         set *out_edges;
  58         set *in_edges;
  59         /** The set of all values that are live in the loop but not used in the loop */
  60         bitset_t *livethrough_unused;
  61 } loop_attr_t;
  62
  63 typedef struct block_attr {
  64         const ir_node *block;
  65         /** set of all values that are live in the block but not used in the block */
  66         bitset_t *livethrough_unused;
  67 } block_attr_t;
  68
  69 //---------------------------------------------------------------------------
  70
  71 static int loop_edge_cmp(const void* p1, const void* p2, size_t s) {
  72         loop_edge_t *e1 = (loop_edge_t*) p1;
  73         loop_edge_t *e2 = (loop_edge_t*) p2;
  74
  75         return e1->block != e2->block || e1->pos != e2->pos;
  76 }
  77
  78 static int loop_attr_cmp(const void *e1, const void *e2, size_t s) {
  79         loop_attr_t *la1 = (loop_attr_t*) e1;
  80         loop_attr_t *la2 = (loop_attr_t*) e2;
  81
  82         return la1->loop != la2->loop;
  83 }
  84
  85 static int block_attr_cmp(const void *e1, const void *e2, size_t s) {
  86         block_attr_t *b1 = (block_attr_t*) e1;
  87         block_attr_t *b2 = (block_attr_t*) e2;
  88
  89         return b1->block != b2->block;
  90 }
  91
  92 static INLINE int loop_attr_hash(const loop_attr_t *a) {
  93 #ifdef DEBUG_libfirm
  94         return a->loop->loop_nr;
  95 #else
  96         return HASH_PTR(a->loop);
  97 #endif
  98 }
  99
 100 static INLINE int block_attr_hash(const block_attr_t *b) {
 101         return nodeset_hash(b->block);
 102 }
 103
 104 static INLINE int loop_edge_hash(const loop_edge_t *e) {
 105         return nodeset_hash(e->block) ^ (e->pos * 31);
 106 }
 107
 108 static INLINE loop_attr_t *get_loop_attr(morgan_env_t *env, const ir_loop *loop) {
 109         loop_attr_t l_attr, *res;
 110         int hash;
 111         l_attr.loop = loop;
 112
 113         hash = loop_attr_hash(&l_attr);
 114         res = set_find(env->loop_attr_set, &l_attr, sizeof(l_attr), hash);
 115
 116         // create new loop_attr if none exists yet
 117         if (res == NULL) {
 118                 l_attr.out_edges = new_set(loop_edge_cmp, 1);
 119                 l_attr.in_edges = new_set(loop_edge_cmp, 1);
 120                 l_attr.livethrough_unused = bitset_obstack_alloc(&env->obst, get_irg_last_idx(env->irg));
 121                 res = set_insert(env->loop_attr_set, &l_attr, sizeof(l_attr), hash);
 122         }
 123
 124         return res;
 125 }
 126
 127 static INLINE block_attr_t *get_block_attr(morgan_env_t *env, const ir_node *block) {
 128         block_attr_t b_attr, *res;
 129         int hash;
 130         b_attr.block = block;
 131
 132         hash = block_attr_hash(&b_attr);
 133         res = set_find(env->block_attr_set, &b_attr, sizeof(b_attr), hash);
 134
 135         if(res == NULL) {
 136                 b_attr.livethrough_unused = bitset_obstack_alloc(&env->obst, get_irg_last_idx(env->irg));
 137                 res = set_insert(env->block_attr_set, &b_attr, sizeof(b_attr), hash);
 138         }
 139
 140         return res;
 141 }
 142
 143 //---------------------------------------------------------------------------
 144
 145 static INLINE int consider_for_spilling(const arch_env_t *env, const arch_register_class_t *cls, const ir_node *node) {
 146         if(!arch_irn_has_reg_class(env, node, -1, cls))
 147                 return 0;
 148
 149         return !(arch_irn_get_flags(env, node) & (arch_irn_flags_ignore | arch_irn_flags_dont_spill));
 150 }
 151
 152 /**
 153  * Determine edges going out of a loop (= edges that go to a block that is not inside
 154  * the loop or one of its subloops)
 155  */
 156 static INLINE void construct_loop_edges(ir_node* block, void* e) {
 157         morgan_env_t *env = (morgan_env_t*) e;
 158         int n_cfgpreds = get_Block_n_cfgpreds(block);
 159         int i;
 160         ir_loop* loop = get_irn_loop(block);
 161         loop_attr_t *loop_attr = get_loop_attr(env, loop);
 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                 ir_node* cfgpred = get_Block_cfgpred(block, i);
 166                 ir_node* cfgpred_block = get_nodes_block(cfgpred);
 167                 ir_loop* cfgpred_loop = get_irn_loop(cfgpred_block);
 168                 loop_attr_t *outedges = get_loop_attr(env, cfgpred_loop);
 169
 170                 if(cfgpred_loop == loop)
 171                         continue;
 172
 173                 // is it an edge into the loop?
 174                 if(get_loop_depth(loop) > get_loop_depth(cfgpred_loop)) {
 175                         loop_edge_t edge;
 176                         edge.block = block;
 177                         edge.pos = i;
 178                         DBG((dbg, DBG_LOOPANA, "Loop in edge from %+F (loop %d) to %+F (loop %d)\n", cfgpred_block, get_loop_loop_nr(cfgpred_loop), block, get_loop_loop_nr(loop)));
 179                         set_insert(loop_attr->in_edges, &edge, sizeof(edge), loop_edge_hash(&edge));
 180                 } else {
 181                         ir_loop *p_loop = cfgpred_loop;
 182                         while(get_loop_depth(p_loop) > get_loop_depth(loop)) {
 183                                 p_loop = get_loop_outer_loop(p_loop);
 184                         }
 185                         if(p_loop != loop) {
 186                                 loop_edge_t edge;
 187                                 edge.block = block;
 188                                 edge.pos = i;
 189                                 DBG((dbg, DBG_LOOPANA, "Loop in edge from %+F (loop %d) to %+F (loop %d)\n", cfgpred_block, get_loop_loop_nr(cfgpred_loop), block, get_loop_loop_nr(loop)));
 190                                 set_insert(loop_attr->in_edges, &edge, sizeof(edge), loop_edge_hash(&edge));
 191                         }
 192                 }
 193
 194                 // an edge out of the loop?
 195                 if(get_loop_depth(cfgpred_loop) >= get_loop_depth(loop)) {
 196                         loop_edge_t edge;
 197                         edge.block = block;
 198                         edge.pos = i;
 199                         DBG((dbg, DBG_LOOPANA, "Loop out edge from %+F (loop %d) to %+F\n", cfgpred_block, cfgpred_loop->loop_nr, block));
 200                         set_insert(outedges->out_edges, &edge, sizeof(edge), loop_edge_hash(&edge));
 201                 } else {
 202                         ir_loop *o_loop = loop;
 203
 204                         // we might jump in the middle of another inner loop which is not inside
 205                         // our loop (happens for irreducible graphs). This would be a
 206                         // real out edge then.
 207                         while(get_loop_depth(o_loop) > get_loop_depth(cfgpred_loop)) {
 208                                 o_loop = get_loop_outer_loop(o_loop);
 209                         }
 210
 211                         if(cfgpred_loop != o_loop) {
 212                                 loop_edge_t edge;
 213                                 edge.block = block;
 214                                 edge.pos = i;
 215                                 DBG((dbg, DBG_LOOPANA, "Loop out edge from %+F (loop %d) to %+F (into jump)\n", cfgpred_block, cfgpred_loop->loop_nr, block));
 216                                 set_insert(outedges->out_edges, &edge, sizeof(edge), loop_edge_hash(&edge));
 217                         }
 218                 }
 219         }
 220 }
 221
 222 static void free_loop_edges(morgan_env_t *env) {
 223         loop_attr_t *l_attr;
 224
 225         for(l_attr = set_first(env->loop_attr_set); l_attr != NULL; l_attr = set_next(env->loop_attr_set)) {
 226                 del_set(l_attr->out_edges);
 227                 del_set(l_attr->in_edges);
 228         }
 229 }
 230
 231 /**
 232  * Debugging help, shows all nodes in a (node-)bitset
 233  */
 234 static void show_nodebitset(ir_graph* irg, const bitset_t* bitset) {
 235         int i;
 236
 237         bitset_foreach(bitset, i) {
 238                 ir_node* node = get_idx_irn(irg, i);
 239                 DBG((dbg, DBG_LIVE, "\t%+F\n", node));
 240         }
 241 }
 242
 243 /**
 244  * Construct the livethrough unused set for a block
 245  */
 246 static bitset_t *construct_block_livethrough_unused(morgan_env_t* env, const ir_node* block) {
 247         block_attr_t *block_attr = get_block_attr(env, block);
 248         ir_node *node;
 249         int i;
 250
 251         DBG((dbg, DBG_LIVE, "Processing block %d\n", get_irn_node_nr(block)));
 252         // copy all live-outs into the livethrough_unused set
 253         be_lv_foreach(env->cenv->lv, block, be_lv_state_in | be_lv_state_out, i) {
 254                 ir_node *irn = be_lv_get_irn(env->cenv->lv, block, i);
 255                 int node_idx;
 256
 257                 if(!consider_for_spilling(env->arch, env->cls, irn))
 258                         continue;
 259
 260                 node_idx = get_irn_idx(irn);
 261                 bitset_set(block_attr->livethrough_unused, node_idx);
 262         }
 263
 264         /*
 265          * All values that are used within the block are not unused (and therefore not
 266          * livethrough_unused)
 267          */
 268         sched_foreach(block, node) {
 269                 int i, arity;
 270
 271                 for(i = 0, arity = get_irn_arity(node); i < arity; ++i) {
 272                         int idx = get_irn_idx(get_irn_n(node, i));
 273                         bitset_clear(block_attr->livethrough_unused, idx);
 274                 }
 275         }
 276
 277         show_nodebitset(env->irg, block_attr->livethrough_unused);
 278         return block_attr->livethrough_unused;
 279 }
 280
 281 /**
 282  * Construct the livethrough unused set for a loop (and all its subloops+blocks)
 283  */
 284 static bitset_t *construct_loop_livethrough_unused(morgan_env_t *env, const ir_loop *loop) {
 285         int i;
 286         loop_attr_t* loop_attr = get_loop_attr(env, loop);
 287
 288         DBG((dbg, DBG_LIVE, "Processing Loop %d\n", loop->loop_nr));
 289         assert(get_loop_n_elements(loop) > 0);
 290         for(i = 0; i < get_loop_n_elements(loop); ++i) {
 291                 loop_element elem = get_loop_element(loop, i);
 292                 switch (*elem.kind) {
 293                 case k_ir_node: {
 294                         bitset_t *livethrough_block_unused;
 295                         assert(is_Block(elem.node));
 296                         livethrough_block_unused = construct_block_livethrough_unused(env, elem.node);
 297                         if(i == 0) {
 298                                 bitset_copy(loop_attr->livethrough_unused, livethrough_block_unused);
 299                         } else {
 300                                 bitset_and(loop_attr->livethrough_unused, livethrough_block_unused);
 301                         }
 302                         break;
 303                 }
 304                 case k_ir_loop: {
 305                         bitset_t *livethrough_son_unused;
 306
 307                         livethrough_son_unused = construct_loop_livethrough_unused(env, elem.son);
 308                         if(i == 0) {
 309                                 bitset_copy(loop_attr->livethrough_unused, livethrough_son_unused);
 310                         } else {
 311                                 bitset_and(loop_attr->livethrough_unused, livethrough_son_unused);
 312                         }
 313                         break;
 314                 }
 315             default:
 316                         assert(0);
 317                         break;
 318                 }
 319     }
 320         DBG((dbg, DBG_LIVE, "Done with loop %d\n", loop->loop_nr));
 321
 322         // remove all unused livethroughs that are remembered for this loop from child loops and blocks
 323         for(i = 0; i < get_loop_n_elements(loop); ++i) {
 324                 const loop_element elem = get_loop_element(loop, i);
 325
 326                 if(*elem.kind == k_ir_loop) {
 327                         loop_attr_t *son_attr = get_loop_attr(env, elem.son);
 328                         bitset_andnot(son_attr->livethrough_unused, loop_attr->livethrough_unused);
 329
 330                         DBG((dbg, DBG_LIVE, "Livethroughs for loop %d:\n", loop->loop_nr));
 331                         show_nodebitset(env->irg, son_attr->livethrough_unused);
 332                 } else if(*elem.kind == k_ir_node) {
 333                         block_attr_t *block_attr = get_block_attr(env, elem.node);
 334                         bitset_andnot(block_attr->livethrough_unused, loop_attr->livethrough_unused);
 335
 336                         DBG((dbg, DBG_LIVE, "Livethroughs for block %+F\n", elem.node));
 337                         show_nodebitset(env->irg, block_attr->livethrough_unused);
 338                 } else {
 339                         assert(0);
 340                 }
 341         }
 342
 343         return loop_attr->livethrough_unused;
 344 }
 345
 346 /*---------------------------------------------------------------------------*/
 347
 348 static int reduce_register_pressure_in_block(morgan_env_t *env, const ir_node* block, int loop_unused_spills_possible) {
 349         ir_node *node;
 350         int max_pressure;
 351         int loop_unused_spills_needed;
 352         pset *live_nodes = pset_new_ptr_default();
 353
 354         be_liveness_end_of_block(env->cenv->lv, env->arch, env->cls, block, live_nodes);
 355         max_pressure = pset_count(live_nodes);
 356
 357         DBG((dbg, DBG_LIVE, "Reduce pressure to %d In Block %+F:\n", env->registers_available, block));
 358
 359         /**
 360          * Determine register pressure in block
 361          */
 362         sched_foreach_reverse(block, node) {
 363                 int pressure;
 364
 365                 if(is_Phi(node))
 366                         break;
 367
 368                 be_liveness_transfer(env->arch, env->cls, node, live_nodes);
 369                 pressure = pset_count(live_nodes);
 370                 if(pressure > max_pressure)
 371                         max_pressure = pressure;
 372         }
 373         del_pset(live_nodes);
 374
 375         DBG((dbg, DBG_PRESSURE, "\tMax Pressure in %+F: %d\n", block, max_pressure));
 376
 377         loop_unused_spills_needed = max_pressure - env->registers_available;
 378
 379         if(loop_unused_spills_needed < 0) {
 380                 loop_unused_spills_needed = 0;
 381         } else if(loop_unused_spills_needed > loop_unused_spills_possible) {
 382                 loop_unused_spills_needed = loop_unused_spills_possible;
 383         }
 384
 385         DBG((dbg, DBG_PRESSURE, "Unused spills for Block %+F needed: %d\n", block, loop_unused_spills_needed));
 386         return loop_unused_spills_needed;
 387 }
 388
 389 /**
 390  * Reduce register pressure in a loop
 391  *
 392  * @param unused_spills_possible        Number of spills from livethrough_unused variables possible in outer loops
 393  * @return                                                      Number of spills of livethrough_unused variables needed in outer loops
 394  */
 395 static int reduce_register_pressure_in_loop(morgan_env_t *env, const ir_loop *loop, int outer_spills_possible) {
 396         int i;
 397         loop_attr_t* loop_attr = get_loop_attr(env, loop);
 398         int spills_needed = 0;
 399         int spills_possible = outer_spills_possible + bitset_popcnt(loop_attr->livethrough_unused);
 400         int outer_spills_needed;
 401
 402         DBG((dbg, DBG_PRESSURE, "Reducing Pressure in loop %d\n", loop->loop_nr));
 403         for(i = 0; i < get_loop_n_elements(loop); ++i) {
 404                 loop_element elem = get_loop_element(loop, i);
 405                 switch (*elem.kind) {
 406                 case k_ir_node: {
 407                         int needed;
 408                         assert(is_Block(elem.node));
 409                         needed = reduce_register_pressure_in_block(env, elem.node, spills_possible);
 410                         assert(needed >= 0);
 411                         assert(needed <= spills_possible);
 412                         if(needed > spills_needed)
 413                                 spills_needed = needed;
 414                         break;
 415                 }
 416                 case k_ir_loop: {
 417                         int needed = reduce_register_pressure_in_loop(env, elem.son, spills_possible);
 418                         assert(needed >= 0);
 419                         assert(needed <= spills_possible);
 420                         if(needed > spills_needed)
 421                                 spills_needed = needed;
 422                         break;
 423                 }
 424             default:
 425                         assert(0);
 426                         break;
 427                 }
 428     }
 429
 430         /* calculate number of spills needed in outer loop and spill
 431          * unused livethrough nodes around this loop
 432          */
 433         if(spills_needed > outer_spills_possible) {
 434                 int spills_to_place;
 435                 outer_spills_needed = outer_spills_possible;
 436                 spills_needed -= outer_spills_possible;
 437
 438                 spills_to_place = spills_needed;
 439
 440                 bitset_foreach(loop_attr->livethrough_unused, i) {
 441                         loop_edge_t *edge;
 442                         ir_node *to_spill = get_idx_irn(env->irg, i);
 443
 444                         for(edge = set_first(loop_attr->out_edges); edge != NULL; edge = set_next(loop_attr->out_edges)) {
 445                                 DBG((dbg, DBG_PRESSURE, "Spilling node %+F around loop %d\n", to_spill, loop->loop_nr));
 446                                 be_add_reload_on_edge(env->senv, to_spill, edge->block, edge->pos);
 447                         }
 448
 449                         spills_to_place--;
 450                         if(spills_to_place <= 0) {
 451                                 break;
 452                         }
 453                 }
 454         } else {
 455                 outer_spills_needed = spills_needed;
 456         }
 457
 458         return outer_spills_needed;
 459 }
 460
 461 void be_spill_morgan(be_chordal_env_t *chordal_env) {
 462         morgan_env_t env;
 463
 464         FIRM_DBG_REGISTER(dbg, "ir.be.spillmorgan");
 465         //firm_dbg_set_mask(dbg, DBG_LOOPANA | DBG_PRESSURE);
 466
 467         env.cenv = chordal_env;
 468         env.arch = chordal_env->birg->main_env->arch_env;
 469         env.irg = chordal_env->irg;
 470         env.cls = chordal_env->cls;
 471         env.senv = be_new_spill_env(chordal_env);
 472         DEBUG_ONLY(be_set_spill_env_dbg_module(env.senv, dbg);)
 473
 474         obstack_init(&env.obst);
 475
 476         env.registers_available = env.cls->n_regs - be_put_ignore_regs(chordal_env->birg, env.cls, NULL);
 477
 478         env.loop_attr_set = new_set(loop_attr_cmp, 5);
 479         env.block_attr_set = new_set(block_attr_cmp, 20);
 480
 481         /*-- Part1: Analysis --*/
 482         be_liveness_recompute(chordal_env->lv);
 483
 484         /* construct control flow loop tree */
 485         construct_cf_backedges(chordal_env->irg);
 486
 487         /* construct loop out edges and livethrough_unused sets for loops and blocks */
 488         irg_block_walk_graph(chordal_env->irg, NULL, construct_loop_edges, &env);
 489         construct_loop_livethrough_unused(&env, get_irg_loop(env.irg));
 490
 491         /*-- Part2: Transformation --*/
 492
 493         /* spill unused livethrough values around loops and blocks where
 494          * the pressure is too high
 495          */
 496         reduce_register_pressure_in_loop(&env, get_irg_loop(env.irg), 0);
 497
 498         /* Insert real spill/reload nodes and fix usages */
 499         be_insert_spills_reloads(env.senv);
 500
 501         /* Verify the result */
 502         if (chordal_env->opts->vrfy_option == BE_CH_VRFY_WARN) {
 503                 be_verify_schedule(env.irg);
 504         } else if (chordal_env->opts->vrfy_option == BE_CH_VRFY_ASSERT) {
 505                 assert(be_verify_schedule(env.irg));
 506         }
 507
 508         if (chordal_env->opts->dump_flags & BE_CH_DUMP_SPILL)
 509                 be_dump(env.irg, "-spillmorgan", dump_ir_block_graph_sched);
 510
 511         /* cleanup */
 512         free_loop_edges(&env);
 513         del_set(env.loop_attr_set);
 514         del_set(env.block_attr_set);
 515
 516         /* fix the remaining places with too high register pressure with beladies algorithm */
 517         be_spill_belady_spill_env(chordal_env, env.senv);
 518
 519         be_delete_spill_env(env.senv);
 520         obstack_free(&env.obst, NULL);
 521 }