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 "irgwalk.h"
  18 #include "besched.h"
  19 #include "beutil.h"
  20 #include "irloop_t.h"
  21 #include "irgraph_t.h"
  22 #include "irprintf.h"
  23 #include "obstack.h"
  24
  25 #include "bespillbelady.h"
  26 #include "beverify.h"
  27
  28 #define DBG_LIVE                1
  29 #define DBG_LOOPANA             2
  30 #define DBG_PRESSURE    4
  31 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  32
  33 typedef struct morgan_env {
  34         const arch_env_t *arch;
  35         const arch_register_class_t *cls;
  36         ir_graph *irg;
  37         struct obstack obst;
  38         /** maximum safe register pressure */
  39         int registers_available;
  40
  41         spill_env_t *senv;
  42
  43         set *loop_attr_set;
  44         set *block_attr_set;
  45 } morgan_env_t;
  46
  47 typedef struct loop_edge {
  48         ir_node *block;
  49         int pos;
  50 } loop_edge_t;
  51
  52 typedef struct loop_attr {
  53         const ir_loop *loop;
  54         set *out_edges;
  55         set *in_edges;
  56         /** The set of all values that are live in the loop but not used in the loop */
  57         bitset_t *livethrough_unused;
  58 } loop_attr_t;
  59
  60 typedef struct block_attr {
  61         const ir_node *block;
  62         /** set of all values that are live in the block but not used in the block */
  63         bitset_t *livethrough_unused;
  64 } block_attr_t;
  65
  66 //---------------------------------------------------------------------------
  67
  68 static int loop_edge_cmp(const void* p1, const void* p2, size_t s) {
  69         loop_edge_t *e1 = (loop_edge_t*) p1;
  70         loop_edge_t *e2 = (loop_edge_t*) p2;
  71
  72         return e1->block != e2->block || e1->pos != e2->pos;
  73 }
  74
  75 static int loop_attr_cmp(const void *e1, const void *e2, size_t s) {
  76         loop_attr_t *la1 = (loop_attr_t*) e1;
  77         loop_attr_t *la2 = (loop_attr_t*) e2;
  78
  79         return la1->loop != la2->loop;
  80 }
  81
  82 static int block_attr_cmp(const void *e1, const void *e2, size_t s) {
  83         block_attr_t *b1 = (block_attr_t*) e1;
  84         block_attr_t *b2 = (block_attr_t*) e2;
  85
  86         return b1->block != b2->block;
  87 }
  88
  89 static INLINE int loop_attr_hash(const loop_attr_t *a) {
  90         return HASH_PTR(a->loop);
  91 }
  92
  93 static INLINE int block_attr_hash(const block_attr_t *b) {
  94         return HASH_PTR(b->block);
  95 }
  96
  97 static INLINE int loop_edge_hash(const loop_edge_t *e) {
  98         return HASH_PTR(e->block) ^ (e->pos * 31);
  99 }
 100
 101 static INLINE loop_attr_t *get_loop_attr(morgan_env_t *env, const ir_loop *loop) {
 102         loop_attr_t l_attr, *res;
 103         int hash;
 104         l_attr.loop = loop;
 105
 106         hash = loop_attr_hash(&l_attr);
 107         res = set_find(env->loop_attr_set, &l_attr, sizeof(l_attr), hash);
 108
 109         // create new loop_attr if none exists yet
 110         if (res == NULL) {
 111                 l_attr.out_edges = new_set(loop_edge_cmp, 1);
 112                 l_attr.in_edges = new_set(loop_edge_cmp, 1);
 113                 l_attr.livethrough_unused = bitset_obstack_alloc(&env->obst, get_irg_last_idx(env->irg));
 114                 res = set_insert(env->loop_attr_set, &l_attr, sizeof(l_attr), hash);
 115         }
 116
 117         return res;
 118 }
 119
 120 static INLINE block_attr_t *get_block_attr(morgan_env_t *env, const ir_node *block) {
 121         block_attr_t b_attr, *res;
 122         int hash;
 123         b_attr.block = block;
 124
 125         hash = block_attr_hash(&b_attr);
 126         res = set_find(env->block_attr_set, &b_attr, sizeof(b_attr), hash);
 127
 128         if(res == NULL) {
 129                 b_attr.livethrough_unused = bitset_obstack_alloc(&env->obst, get_irg_last_idx(env->irg));
 130                 res = set_insert(env->block_attr_set, &b_attr, sizeof(b_attr), hash);
 131         }
 132
 133         return res;
 134 }
 135
 136 //---------------------------------------------------------------------------
 137
 138 static INLINE int consider_for_spilling(const arch_env_t *env, const arch_register_class_t *cls, const ir_node *node) {
 139         if(!arch_irn_has_reg_class(env, node, -1, cls))
 140                 return 0;
 141
 142         return !(arch_irn_get_flags(env, node) & (arch_irn_flags_ignore | arch_irn_flags_dont_spill));
 143 }
 144
 145 /**
 146  * Determine edges going out of a loop (= edges that go to a block that is not inside
 147  * the loop or one of its subloops)
 148  */
 149 static INLINE void construct_loop_edges(ir_node* block, void* e) {
 150         morgan_env_t *env = (morgan_env_t*) e;
 151         int n_cfgpreds = get_Block_n_cfgpreds(block);
 152         int i;
 153         ir_loop* loop = get_irn_loop(block);
 154         loop_attr_t *loop_attr = get_loop_attr(env, loop);
 155         DBG((dbg, DBG_LOOPANA, "Loop for %+F: %d (depth %d)\n", block, loop->loop_nr, loop->depth));
 156
 157         for(i = 0; i < n_cfgpreds; ++i) {
 158                 ir_node* cfgpred = get_Block_cfgpred(block, i);
 159                 ir_node* cfgpred_block = get_nodes_block(cfgpred);
 160                 ir_loop* cfgpred_loop = get_irn_loop(cfgpred_block);
 161                 loop_attr_t *outedges = get_loop_attr(env, cfgpred_loop);
 162
 163                 if(cfgpred_loop == loop)
 164                         continue;
 165
 166                 // is it an edge into the loop?
 167                 if(get_loop_depth(loop) > get_loop_depth(cfgpred_loop)) {
 168                         loop_edge_t edge;
 169                         edge.block = block;
 170                         edge.pos = i;
 171                         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)));
 172                         set_insert(loop_attr->in_edges, &edge, sizeof(edge), loop_edge_hash(&edge));
 173                 } else {
 174                         ir_loop *p_loop = cfgpred_loop;
 175                         while(get_loop_depth(p_loop) > get_loop_depth(loop)) {
 176                                 p_loop = get_loop_outer_loop(p_loop);
 177                         }
 178                         if(p_loop != loop) {
 179                                 loop_edge_t edge;
 180                                 edge.block = block;
 181                                 edge.pos = i;
 182                                 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)));
 183                                 set_insert(loop_attr->in_edges, &edge, sizeof(edge), loop_edge_hash(&edge));
 184                         }
 185                 }
 186
 187                 // an edge out of the loop?
 188                 if(get_loop_depth(cfgpred_loop) >= get_loop_depth(loop)) {
 189                         loop_edge_t edge;
 190                         edge.block = block;
 191                         edge.pos = i;
 192                         DBG((dbg, DBG_LOOPANA, "Loop out edge from %+F (loop %d) to %+F\n", cfgpred_block, cfgpred_loop->loop_nr, block));
 193                         set_insert(outedges->out_edges, &edge, sizeof(edge), loop_edge_hash(&edge));
 194                 } else {
 195                         ir_loop *o_loop = loop;
 196
 197                         // we might jump in the middle of another inner loop which is not inside
 198                         // our loop (happens for irreducible graphs). This would be a
 199                         // real out edge then.
 200                         while(get_loop_depth(o_loop) > get_loop_depth(cfgpred_loop)) {
 201                                 o_loop = get_loop_outer_loop(o_loop);
 202                         }
 203
 204                         if(cfgpred_loop != o_loop) {
 205                                 loop_edge_t edge;
 206                                 edge.block = block;
 207                                 edge.pos = i;
 208                                 DBG((dbg, DBG_LOOPANA, "Loop out edge from %+F (loop %d) to %+F (into jump)\n", cfgpred_block, cfgpred_loop->loop_nr, block));
 209                                 set_insert(outedges->out_edges, &edge, sizeof(edge), loop_edge_hash(&edge));
 210                         }
 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 /**
 225  * Debugging help, shows all nodes in a (node-)bitset
 226  */
 227 static void show_nodebitset(ir_graph* irg, const bitset_t* bitset) {
 228         int i;
 229
 230         bitset_foreach(bitset, i) {
 231                 ir_node* node = get_idx_irn(irg, i);
 232                 DBG((dbg, DBG_LIVE, "\t%+F\n", node));
 233         }
 234 }
 235
 236 /**
 237  * Construct the livethrough unused set for a block
 238  */
 239 static bitset_t *construct_block_livethrough_unused(morgan_env_t* env, const ir_node* block) {
 240         block_attr_t *block_attr = get_block_attr(env, block);
 241         irn_live_t *li;
 242         ir_node *node;
 243
 244         DBG((dbg, DBG_LIVE, "Processing block %d\n", get_irn_node_nr(block)));
 245         // copy all live-outs into the livethrough_unused set
 246         live_foreach(block, li) {
 247                 int node_idx;
 248
 249                 if(!live_is_in(li) || !live_is_out(li))
 250                         continue;
 251                 if(!consider_for_spilling(env->arch, env->cls, li->irn))
 252                         continue;
 253
 254                 node_idx = get_irn_idx(li->irn);
 255                 bitset_set(block_attr->livethrough_unused, node_idx);
 256         }
 257
 258         /*
 259          * All values that are used within the block are not unused (and therefore not
 260          * livethrough_unused)
 261          */
 262         sched_foreach(block, node) {
 263                 int i, arity;
 264
 265                 for(i = 0, arity = get_irn_arity(node); i < arity; ++i) {
 266                         int idx = get_irn_idx(get_irn_n(node, i));
 267                         bitset_clear(block_attr->livethrough_unused, idx);
 268                 }
 269         }
 270
 271         show_nodebitset(env->irg, block_attr->livethrough_unused);
 272         return block_attr->livethrough_unused;
 273 }
 274
 275 /**
 276  * Construct the livethrough unused set for a loop (and all its subloops+blocks)
 277  */
 278 static bitset_t *construct_loop_livethrough_unused(morgan_env_t *env, const ir_loop *loop) {
 279         int i;
 280         loop_attr_t* loop_attr = get_loop_attr(env, loop);
 281
 282         DBG((dbg, DBG_LIVE, "Processing Loop %d\n", loop->loop_nr));
 283         assert(get_loop_n_elements(loop) > 0);
 284         for(i = 0; i < get_loop_n_elements(loop); ++i) {
 285                 loop_element elem = get_loop_element(loop, i);
 286                 switch (*elem.kind) {
 287                 case k_ir_node: {
 288                         bitset_t *livethrough_block_unused;
 289                         assert(is_Block(elem.node));
 290                         livethrough_block_unused = construct_block_livethrough_unused(env, elem.node);
 291                         if(i == 0) {
 292                                 bitset_copy(loop_attr->livethrough_unused, livethrough_block_unused);
 293                         } else {
 294                                 bitset_and(loop_attr->livethrough_unused, livethrough_block_unused);
 295                         }
 296                         break;
 297                 }
 298                 case k_ir_loop: {
 299                         bitset_t *livethrough_son_unused;
 300
 301                         livethrough_son_unused = construct_loop_livethrough_unused(env, elem.son);
 302                         if(i == 0) {
 303                                 bitset_copy(loop_attr->livethrough_unused, livethrough_son_unused);
 304                         } else {
 305                                 bitset_and(loop_attr->livethrough_unused, livethrough_son_unused);
 306                         }
 307                         break;
 308                 }
 309             default:
 310                         assert(0);
 311                         break;
 312                 }
 313     }
 314         DBG((dbg, DBG_LIVE, "Done with loop %d\n", loop->loop_nr));
 315
 316         // remove all unused livethroughs that are remembered for this loop from child loops and blocks
 317         for(i = 0; i < get_loop_n_elements(loop); ++i) {
 318                 const loop_element elem = get_loop_element(loop, i);
 319
 320                 if(*elem.kind == k_ir_loop) {
 321                         loop_attr_t *son_attr = get_loop_attr(env, elem.son);
 322                         bitset_andnot(son_attr->livethrough_unused, loop_attr->livethrough_unused);
 323
 324                         DBG((dbg, DBG_LIVE, "Livethroughs for loop %d:\n", loop->loop_nr));
 325                         show_nodebitset(env->irg, son_attr->livethrough_unused);
 326                 } else if(*elem.kind == k_ir_node) {
 327                         block_attr_t *block_attr = get_block_attr(env, elem.node);
 328                         bitset_andnot(block_attr->livethrough_unused, loop_attr->livethrough_unused);
 329
 330                         DBG((dbg, DBG_LIVE, "Livethroughs for block %+F\n", elem.node));
 331                         show_nodebitset(env->irg, block_attr->livethrough_unused);
 332                 } else {
 333                         assert(0);
 334                 }
 335         }
 336
 337         return loop_attr->livethrough_unused;
 338 }
 339
 340 /*---------------------------------------------------------------------------*/
 341
 342 static int reduce_register_pressure_in_block(morgan_env_t *env, const ir_node* block, int loop_unused_spills_possible) {
 343         int pressure;
 344         ir_node *irn;
 345         int max_pressure = 0;
 346         int spills_needed;
 347         int loop_unused_spills_needed;
 348         block_attr_t *block_attr = get_block_attr(env, block);
 349         int block_unused_spills_possible = bitset_popcnt(block_attr->livethrough_unused);
 350         int unused_spills_possible = loop_unused_spills_possible + block_unused_spills_possible;
 351         pset *live_nodes = pset_new_ptr_default();
 352
 353         be_liveness_end_of_block(env->arch, env->cls, block, live_nodes);
 354         pressure = pset_count(live_nodes);
 355
 356         DBG((dbg, DBG_LIVE, "Reduce pressure to %d In Block %+F:\n", env->registers_available, block));
 357
 358         /**
 359          * Walk over all irns in the schedule and check register pressure for each of them
 360          */
 361         sched_foreach_reverse(block, irn) {
 362                 // do we need more spills than possible with unused libethroughs?
 363                 int spills_needed = pressure - env->registers_available - unused_spills_possible;
 364                 if(spills_needed > 0) {
 365                         DBG((dbg, DBG_PRESSURE, "\tWARNING %d more spills needed at %+F\n", spills_needed, irn));
 366                         // TODO further spills needed
 367                         //assert(0);
 368                 }
 369                 if(pressure > max_pressure) {
 370                         max_pressure = pressure;
 371                 }
 372
 373                 /* Register pressure is only important until we reach the first phi (the rest of the block
 374                  * will only be phis.)
 375                  */
 376                 if(is_Phi(irn))
 377                         break;
 378
 379                 // update pressure
 380                 be_liveness_transfer(env->arch, env->cls, irn, live_nodes);
 381                 pressure = pset_count(live_nodes);
 382         }
 383
 384         DBG((dbg, DBG_PRESSURE, "\tMax Pressure in %+F: %d\n", block, max_pressure));
 385
 386         /*
 387          * Calculate number of spills from loop_unused_spills_possible that we want to use,
 388          * and spill unused livethroughs from the block if we still don't have enough registers
 389          */
 390         spills_needed = max_pressure - env->registers_available;
 391         if(spills_needed < 0) {
 392                 loop_unused_spills_needed = 0;
 393         } else if(spills_needed > loop_unused_spills_possible) {
 394                 int i, spills;
 395                 int block_unused_spills_needed;
 396
 397                 loop_unused_spills_needed = loop_unused_spills_possible;
 398                 block_unused_spills_needed = spills_needed - loop_unused_spills_possible;
 399                 if(block_unused_spills_needed > block_unused_spills_possible) {
 400                         block_unused_spills_needed = block_unused_spills_possible;
 401                 }
 402
 403                 spills = 0;
 404                 /*
 405                  * Spill/Reload unused livethroughs from the block
 406                  */
 407                 bitset_foreach(block_attr->livethrough_unused, i) {
 408                         ir_node *to_spill;
 409                         const ir_edge_t *edge;
 410
 411                         if(spills >= block_unused_spills_needed)
 412                                 break;
 413
 414                         to_spill = get_idx_irn(env->irg, i);
 415                         foreach_block_succ(block, edge) {
 416                                 DBG((dbg, DBG_PRESSURE, "Spilling node %+F around block %+F\n", to_spill, block));
 417                                 be_add_reload_on_edge(env->senv, to_spill, edge->src, edge->pos);
 418                         }
 419                         spills++;
 420                 }
 421         } else {
 422                 loop_unused_spills_needed = spills_needed;
 423         }
 424
 425         del_pset(live_nodes);
 426
 427         DBG((dbg, DBG_PRESSURE, "Unused spills for Block %+F needed: %d\n", block, loop_unused_spills_needed));
 428         return loop_unused_spills_needed;
 429 }
 430
 431 /**
 432  * Reduce register pressure in a loop
 433  *
 434  * @param unused_spills_possible        Number of spills from livethrough_unused variables possible in outer loops
 435  * @return                                                      Number of spills of livethrough_unused variables needed in outer loops
 436  */
 437 static int reduce_register_pressure_in_loop(morgan_env_t *env, const ir_loop *loop, int outer_spills_possible) {
 438         int i;
 439         loop_attr_t* loop_attr = get_loop_attr(env, loop);
 440         int spills_needed = 0;
 441         int spills_possible = outer_spills_possible + bitset_popcnt(loop_attr->livethrough_unused);
 442         int outer_spills_needed;
 443
 444         DBG((dbg, DBG_PRESSURE, "Reducing Pressure in loop %d\n", loop->loop_nr));
 445         for(i = 0; i < get_loop_n_elements(loop); ++i) {
 446                 loop_element elem = get_loop_element(loop, i);
 447                 switch (*elem.kind) {
 448                 case k_ir_node: {
 449                         int needed;
 450                         assert(is_Block(elem.node));
 451                         needed = reduce_register_pressure_in_block(env, elem.node, spills_possible);
 452                         assert(needed <= spills_possible);
 453                         if(needed > spills_needed)
 454                                 spills_needed = needed;
 455                         break;
 456                 }
 457                 case k_ir_loop: {
 458                         int needed = reduce_register_pressure_in_loop(env, elem.son, spills_possible);
 459                         assert(needed <= spills_possible);
 460                         if(needed > spills_needed)
 461                                 spills_needed = needed;
 462                         break;
 463                 }
 464             default:
 465                         assert(0);
 466                         break;
 467                 }
 468     }
 469
 470         /* calculate number of spills needed in outer loop and spill
 471          * unused livethrough nodes around this loop
 472          */
 473         if(spills_needed > outer_spills_possible) {
 474                 int spills_to_place;
 475                 outer_spills_needed = outer_spills_possible;
 476                 spills_needed -= outer_spills_possible;
 477
 478                 spills_to_place = spills_needed;
 479
 480                 bitset_foreach(loop_attr->livethrough_unused, i) {
 481                         loop_edge_t *edge;
 482                         ir_node *to_spill = get_idx_irn(env->irg, i);
 483
 484                         for(edge = set_first(loop_attr->out_edges); edge != NULL; edge = set_next(loop_attr->out_edges)) {
 485                                 DBG((dbg, DBG_PRESSURE, "Spilling node %+F around loop %d\n", to_spill, loop->loop_nr));
 486                                 be_add_reload_on_edge(env->senv, to_spill, edge->block, edge->pos);
 487                         }
 488
 489                         spills_to_place--;
 490                         if(spills_to_place <= 0) {
 491                                 break;
 492                         }
 493                 }
 494         } else {
 495                 outer_spills_needed = spills_needed;
 496         }
 497
 498         return outer_spills_needed;
 499 }
 500
 501 void be_spill_morgan(const be_chordal_env_t *chordal_env) {
 502         morgan_env_t env;
 503
 504         FIRM_DBG_REGISTER(dbg, "ir.be.spillmorgan");
 505         //firm_dbg_set_mask(dbg, DBG_LOOPANA | DBG_PRESSURE);
 506
 507         env.arch = chordal_env->birg->main_env->arch_env;
 508         env.irg = chordal_env->irg;
 509         env.cls = chordal_env->cls;
 510         env.senv = be_new_spill_env(chordal_env);
 511         DEBUG_ONLY(be_set_spill_env_dbg_module(env.senv, dbg);)
 512
 513         obstack_init(&env.obst);
 514
 515         env.registers_available = arch_count_non_ignore_regs(env.arch, env.cls);
 516
 517         env.loop_attr_set = new_set(loop_attr_cmp, 5);
 518         env.block_attr_set = new_set(block_attr_cmp, 20);
 519
 520         /*-- Part1: Analysis --*/
 521         be_liveness(env.irg);
 522
 523         /* construct control flow loop tree */
 524         construct_cf_backedges(chordal_env->irg);
 525
 526         /* construct loop out edges and livethrough_unused sets for loops and blocks */
 527         irg_block_walk_graph(chordal_env->irg, NULL, construct_loop_edges, &env);
 528         construct_loop_livethrough_unused(&env, get_irg_loop(env.irg));
 529
 530         /*-- Part2: Transformation --*/
 531
 532         /* spill unused livethrough values around loops and blocks where
 533          * the pressure is too high
 534          */
 535         reduce_register_pressure_in_loop(&env, get_irg_loop(env.irg), 0);
 536
 537         /* Place copies for spilled phis */
 538         be_place_copies(env.senv);
 539         /* Insert real spill/reload nodes and fix usages */
 540         be_insert_spills_reloads(env.senv);
 541
 542         /* Verify the result */
 543         if (chordal_env->opts->vrfy_option == BE_CH_VRFY_WARN) {
 544                 be_verify_schedule(env.irg);
 545         } else if (chordal_env->opts->vrfy_option == BE_CH_VRFY_ASSERT) {
 546                 assert(be_verify_schedule(env.irg));
 547         }
 548
 549         if (chordal_env->opts->dump_flags & BE_CH_DUMP_SPILL)
 550                 be_dump(env.irg, "-spillmorgan", dump_ir_block_graph_sched);
 551
 552         /* cleanup */
 553         free_loop_edges(&env);
 554         del_set(env.loop_attr_set);
 555         del_set(env.block_attr_set);
 556
 557         /* fix the remaining places with too high register pressure with beladies algorithm */
 558
 559         /* we have to remove dead nodes from schedule to not confuse liveness calculation */
 560         be_remove_dead_nodes_from_schedule(env.irg);
 561         be_liveness(env.irg);
 562
 563         be_spill_belady_spill_env(chordal_env, env.senv);
 564
 565         be_delete_spill_env(env.senv);
 566         obstack_free(&env.obst, NULL);
 567 }