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 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.h"
  15 #include "bechordal_t.h"
  16 #include "bespill.h"
  17 #include "belive.h"
  18 #include "belive_t.h"
  19 #include "irgwalk.h"
  20 #include "besched.h"
  21 #include "beutil.h"
  22 #include "irloop.h"
  23 #include "irloop_t.h"
  24 #include "irgraph.h"
  25 #include "irgraph_t.h"
  26 #include "irphase.h"
  27 #include "irphase_t.h"
  28 #include "irprintf.h"
  29
  30 #include "bespillbelady.h"
  31 #include "beverify.h"
  32
  33 #define DBG_LIVE                1
  34 #define DBG_PRESSURE    2
  35 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  36
  37 typedef struct _morgan_env_t {
  38         const arch_env_t *arch;
  39         const arch_register_class_t *cls;
  40         ir_graph *irg;
  41         phase_t phase;
  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_out_edge_t {
  52         ir_node *block;
  53         int pos;
  54 } loop_out_edge_t;
  55
  56 typedef struct _loop_attr_t {
  57         ir_loop *loop;
  58         set *out_edges;
  59         /// The set of all values that live through the loop and are not used
  60         bitset_t *livethrough_unused;
  61 } loop_attr_t;
  62
  63 typedef struct _block_attr_t {
  64         ir_node *block;
  65         bitset_t *livethrough_unused;
  66 } block_attr_t;
  67
  68 //---------------------------------------------------------------------------
  69
  70 int loop_out_edge_cmp(const void* p1, const void* p2, size_t s) {
  71         loop_out_edge_t *e1 = (loop_out_edge_t*) p1;
  72         loop_out_edge_t *e2 = (loop_out_edge_t*) p2;
  73
  74         return e1->block != e2->block || e1->pos != e2->pos;
  75 }
  76
  77 int loop_attr_cmp(const void *e1, const void *e2, size_t s) {
  78         loop_attr_t *la1 = (loop_attr_t*) e1;
  79         loop_attr_t *la2 = (loop_attr_t*) e2;
  80
  81         return la1->loop != la2->loop;
  82 }
  83
  84 int block_attr_cmp(const void *e1, const void *e2, size_t s) {
  85         block_attr_t *b1 = (block_attr_t*) e1;
  86         block_attr_t *b2 = (block_attr_t*) e2;
  87
  88         return b1->block != b2->block;
  89 }
  90
  91 static INLINE int loop_attr_hash(const loop_attr_t *a) {
  92         return HASH_PTR(a->loop);
  93 }
  94
  95 static INLINE int block_attr_hash(const block_attr_t *b) {
  96         return HASH_PTR(b->block);
  97 }
  98
  99 static INLINE int loop_out_edge_hash(const loop_out_edge_t *e) {
 100         return HASH_PTR(e->block) ^ (e->pos * 31);
 101 }
 102
 103 static INLINE loop_attr_t *get_loop_attr(morgan_env_t *env, ir_loop *loop) {
 104         loop_attr_t l_attr, *res;
 105         int hash;
 106         l_attr.loop = loop;
 107
 108         hash = loop_attr_hash(&l_attr);
 109         res = set_find(env->loop_attr_set, &l_attr, sizeof(l_attr), hash);
 110
 111         // create new loop_attr if none exists yet
 112         if (!res) {
 113                 l_attr.out_edges = new_set(loop_out_edge_cmp, 1);
 114                 l_attr.livethrough_unused = bitset_obstack_alloc(&env->phase.obst, get_irg_last_idx(env->irg));
 115                 res = set_insert(env->loop_attr_set, &l_attr, sizeof(l_attr), hash);
 116         }
 117
 118         return res;
 119 }
 120
 121 static INLINE block_attr_t *get_block_attr(morgan_env_t *env, ir_node *block) {
 122         block_attr_t b_attr, *res;
 123         int hash;
 124         b_attr.block = block;
 125
 126         hash = block_attr_hash(&b_attr);
 127         res = set_find(env->block_attr_set, &b_attr, sizeof(b_attr), hash);
 128
 129         if(!res) {
 130                 b_attr.livethrough_unused = bitset_obstack_alloc(&env->phase.obst, get_irg_last_idx(env->irg));
 131                 res = set_insert(env->block_attr_set, &b_attr, sizeof(b_attr), hash);
 132         }
 133
 134         return res;
 135 }
 136
 137 //---------------------------------------------------------------------------
 138
 139 /**
 140  * Determine edges going out of a loop (= edges that go to a block that is not inside
 141  * the loop or one of its subloops)
 142  */
 143 static INLINE void construct_loop_out_edges(ir_node* block, void* e) {
 144         morgan_env_t *env = (morgan_env_t*) e;
 145         int n_cfgpreds = get_Block_n_cfgpreds(block);
 146         int i;
 147         ir_loop* loop = get_irn_loop(block);
 148
 149         for(i = 0; i < n_cfgpreds; ++i) {
 150                 ir_node* cfgpred = get_Block_cfgpred(block, i);
 151                 ir_node* cfgpred_block = get_nodes_block(cfgpred);
 152                 ir_loop* cfgpred_loop = get_irn_loop(cfgpred_block);
 153                 loop_attr_t *outedges = get_loop_attr(env, cfgpred_loop);
 154
 155                 if(cfgpred_loop != loop && get_loop_depth(cfgpred_loop) >= get_loop_depth(loop)) {
 156                         loop_out_edge_t edge;
 157                         edge.block = block;
 158                         edge.pos = i;
 159                         set_insert(outedges->out_edges, &edge, sizeof(edge), loop_out_edge_hash(&edge));
 160                 }
 161         }
 162 }
 163
 164 static void free_loop_out_edges(morgan_env_t *env) {
 165         loop_attr_t *l_attr;
 166
 167         for(l_attr = set_first(env->loop_attr_set); l_attr != NULL; l_attr = set_next(env->loop_attr_set)) {
 168                 del_set(l_attr->out_edges);
 169         }
 170 }
 171
 172 /**
 173  * Debugging help, shows all nodes in a (node-)bitset
 174  */
 175 static void show_nodebitset(ir_graph* irg, bitset_t* bitset) {
 176         int i;
 177
 178         bitset_foreach(bitset, i) {
 179                 ir_node* node = get_idx_irn(irg, i);
 180                 DBG((dbg, DBG_LIVE, "\t%+F\n", node));
 181         }
 182 }
 183
 184 /**
 185  * Construct the livethrough unused set for a block
 186  */
 187 static bitset_t *construct_block_livethrough_unused(morgan_env_t* env, ir_node* block) {
 188         block_attr_t *block_attr = get_block_attr(env, block);
 189         irn_live_t *li;
 190         ir_node *node;
 191
 192         DBG((dbg, DBG_LIVE, "Processing block %d\n", get_irn_node_nr(block)));
 193         // copy all live-outs into the livethrough_unused set
 194         live_foreach(block, li) {
 195                 int node_idx;
 196
 197                 if(!live_is_in(li) || !live_is_out(li))
 198                         continue;
 199                 if(!arch_irn_consider_in_reg_alloc(env->arch, env->cls, li->irn))
 200                         continue;
 201
 202                 node_idx = get_irn_idx(li->irn);
 203                 bitset_set(block_attr->livethrough_unused, node_idx);
 204         }
 205
 206         /*
 207          * All values that are used within the block are not unused (and therefore not
 208          * livethrough_unused)
 209          */
 210         sched_foreach(block, node) {
 211                 int i, arity;
 212
 213                 for(i = 0, arity = get_irn_arity(node); i < arity; ++i) {
 214                         int idx = get_irn_idx(get_irn_n(node, i));
 215                         bitset_clear(block_attr->livethrough_unused, idx);
 216                 }
 217         }
 218
 219         show_nodebitset(env->irg, block_attr->livethrough_unused);
 220         return block_attr->livethrough_unused;
 221 }
 222
 223 /**
 224  * Construct the livethrough unused set for a loop (and all its subloops+blocks)
 225  */
 226 static bitset_t *construct_loop_livethrough_unused(morgan_env_t *env, ir_loop *loop) {
 227         int i;
 228         loop_attr_t* loop_attr = get_loop_attr(env, loop);
 229
 230         DBG((dbg, DBG_LIVE, "Processing Loop %d\n", loop->loop_nr));
 231         assert(get_loop_n_elements(loop) > 0);
 232         for(i = 0; i < get_loop_n_elements(loop); ++i) {
 233                 loop_element elem = get_loop_element(loop, i);
 234                 switch (*elem.kind) {
 235                 case k_ir_node: {
 236                         bitset_t *livethrough_block_unused;
 237                         assert(is_Block(elem.node));
 238                         livethrough_block_unused = construct_block_livethrough_unused(env, elem.node);
 239                         if(i == 0) {
 240                                 bitset_copy(loop_attr->livethrough_unused, livethrough_block_unused);
 241                         } else {
 242                                 bitset_and(loop_attr->livethrough_unused, livethrough_block_unused);
 243                         }
 244                         break;
 245                 }
 246                 case k_ir_loop: {
 247                         bitset_t *livethrough_son_unused;
 248
 249                         livethrough_son_unused = construct_loop_livethrough_unused(env, elem.son);
 250                         if(i == 0) {
 251                                 bitset_copy(loop_attr->livethrough_unused, livethrough_son_unused);
 252                         } else {
 253                                 bitset_and(loop_attr->livethrough_unused, livethrough_son_unused);
 254                         }
 255                         break;
 256                 }
 257             default:
 258                         assert(0);
 259                         break;
 260                 }
 261     }
 262         DBG((dbg, DBG_LIVE, "Done with loop %d\n", loop->loop_nr));
 263
 264         // remove all unused livethroughs that are remembered for this loop from child loops and blocks
 265         for(i = 0; i < get_loop_n_elements(loop); ++i) {
 266                 const loop_element elem = get_loop_element(loop, i);
 267
 268                 if(*elem.kind == k_ir_loop) {
 269                         loop_attr_t *son_attr = get_loop_attr(env, elem.son);
 270                         bitset_andnot(son_attr->livethrough_unused, loop_attr->livethrough_unused);
 271
 272                         DBG((dbg, DBG_LIVE, "Livethroughs for loop %d:\n", loop->loop_nr));
 273                         show_nodebitset(env->irg, son_attr->livethrough_unused);
 274                 } else if(*elem.kind == k_ir_node) {
 275                         block_attr_t *block_attr = get_block_attr(env, elem.node);
 276                         bitset_andnot(block_attr->livethrough_unused, loop_attr->livethrough_unused);
 277
 278                         DBG((dbg, DBG_LIVE, "Livethroughs for block %+F\n", elem.node));
 279                         show_nodebitset(env->irg, block_attr->livethrough_unused);
 280                 } else {
 281                         assert(0);
 282                 }
 283         }
 284
 285         return loop_attr->livethrough_unused;
 286 }
 287
 288 //---------------------------------------------------------------------------
 289
 290 static int reduce_register_pressure_in_block(morgan_env_t *env, ir_node* block, int loop_unused_spills_possible) {
 291         int pressure;
 292         ir_node *irn;
 293         int max_pressure = 0;
 294         int spills_needed;
 295         int loop_unused_spills_needed;
 296         block_attr_t *block_attr = get_block_attr(env, block);
 297         int block_unused_spills_possible = bitset_popcnt(block_attr->livethrough_unused);
 298         int unused_spills_possible = loop_unused_spills_possible + block_unused_spills_possible;
 299         pset *live_nodes = pset_new_ptr_default();
 300
 301         be_liveness_end_of_block(env->arch, env->cls, block, live_nodes);
 302         pressure = pset_count(live_nodes);
 303
 304         DBG((dbg, DBG_LIVE, "Reduce pressure to %d In Block %+F:\n", env->registers_available, block));
 305
 306         /**
 307          * Walk over all irns in the schedule and check register pressure for each of them
 308          */
 309         sched_foreach_reverse(block, irn) {
 310                 // do we need more spills than possible with unused libethroughs?
 311                 int spills_needed = pressure - unused_spills_possible - env->registers_available;
 312                 if(spills_needed > 0) {
 313                         DBG((dbg, DBG_PRESSURE, "\tWARNING %d more spills needed at %+F\n", spills_needed, irn));
 314                         // TODO further spills needed
 315                         //assert(0);
 316                 }
 317                 if(pressure > max_pressure) {
 318                         max_pressure = pressure;
 319                 }
 320
 321                 /* Register pressure is only important until we reach the first phi (the rest of the block
 322                  * will only be phis.)
 323                  */
 324                 if(is_Phi(irn))
 325                         break;
 326
 327                 // update pressure
 328                 {
 329                         int pressure_old = pressure;
 330                         be_liveness_transfer(env->arch, env->cls, irn, live_nodes);
 331                         pressure = pset_count(live_nodes);
 332                         DBG((dbg, DBG_PRESSURE, "\tPressure at %+F - before: %d after: %d\n", irn, pressure_old, pressure));
 333                 }
 334         }
 335
 336         /*
 337          * Calculate number of spills from loop_unused_spills_possible that we want to use,
 338          * and spill unused livethroughs from the block if we still don't have enough registers
 339          */
 340         spills_needed = max_pressure - env->registers_available;
 341         if(spills_needed < 0) {
 342                 loop_unused_spills_needed = 0;
 343         } else if(spills_needed > loop_unused_spills_possible) {
 344                 int i, spills;
 345                 int block_unused_spills_needed;
 346
 347                 loop_unused_spills_needed = loop_unused_spills_possible;
 348                 block_unused_spills_needed = spills_needed - loop_unused_spills_possible;
 349                 if(block_unused_spills_needed > block_unused_spills_possible) {
 350                         block_unused_spills_needed = block_unused_spills_possible;
 351                 }
 352
 353                 spills = 0;
 354                 /*
 355                  * Spill/Reload unused livethroughs from the block
 356                  */
 357                 bitset_foreach(block_attr->livethrough_unused, i) {
 358                         ir_node *to_spill;
 359                         const ir_edge_t *edge;
 360
 361                         if(spills >= block_unused_spills_needed)
 362                                 break;
 363
 364                         to_spill = get_idx_irn(env->irg, i);
 365                         foreach_block_succ(block, edge) {
 366                                 DBG((dbg, DBG_PRESSURE, "Spilling node %+F around block %+F\n", to_spill, block));
 367                                 be_add_reload_on_edge(env->senv, to_spill, edge->src, edge->pos);
 368                         }
 369                 }
 370         } else {
 371                 loop_unused_spills_needed = spills_needed;
 372         }
 373
 374         del_pset(live_nodes);
 375
 376         DBG((dbg, DBG_PRESSURE, "Unused spills for Block %+F needed: %d\n", block, loop_unused_spills_needed));
 377         return loop_unused_spills_needed;
 378 }
 379
 380 /**
 381  * Reduce register pressure in a loop
 382  *
 383  * @param unused_spills_possible        Number of spills from livethrough_unused variables possible in outer loops
 384  * @return                                                      Number of spills of livethrough_unused variables needed in outer loops
 385  */
 386 static int reduce_register_pressure_in_loop(morgan_env_t *env, ir_loop *loop, int outer_spills_possible) {
 387         int i;
 388         loop_attr_t* loop_attr = get_loop_attr(env, loop);
 389         int spills_needed = 0;
 390         int spills_possible = outer_spills_possible + bitset_popcnt(loop_attr->livethrough_unused);
 391         int outer_spills_needed;
 392
 393         DBG((dbg, DBG_PRESSURE, "Reducing Pressure in loop %d\n", loop->loop_nr));
 394         for(i = 0; i < get_loop_n_elements(loop); ++i) {
 395                 loop_element elem = get_loop_element(loop, i);
 396                 switch (*elem.kind) {
 397                 case k_ir_node: {
 398                         int needed;
 399                         assert(is_Block(elem.node));
 400                         needed = reduce_register_pressure_in_block(env, elem.node, spills_possible);
 401                         assert(needed <= spills_possible);
 402                         if(needed > spills_needed)
 403                                 spills_needed = needed;
 404                         break;
 405                 }
 406                 case k_ir_loop: {
 407                         int needed = reduce_register_pressure_in_loop(env, elem.son, spills_possible);
 408                         assert(needed <= spills_possible);
 409                         if(needed > spills_needed)
 410                                 spills_needed = needed;
 411                         break;
 412                 }
 413             default:
 414                         assert(0);
 415                         break;
 416                 }
 417     }
 418
 419         // calculate number of spills needed in outer loop and spill
 420         // unused livethrough nodes around this loop
 421         if(spills_needed > outer_spills_possible) {
 422                 outer_spills_needed = outer_spills_possible;
 423                 spills_needed -= outer_spills_possible;
 424
 425                 bitset_foreach(loop_attr->livethrough_unused, i) {
 426                         loop_out_edge_t *edge;
 427                         ir_node *to_spill = get_idx_irn(env->irg, i);
 428
 429                         for(edge = set_first(loop_attr->out_edges); edge != NULL; edge = set_next(loop_attr->out_edges)) {
 430                                 if(is_Phi(to_spill)) {
 431                                         be_spill_phi(env->senv, to_spill);
 432                                 }
 433
 434                                 be_add_reload_on_edge(env->senv, to_spill, edge->block, edge->pos);
 435                         }
 436                 }
 437         } else {
 438                 outer_spills_needed = spills_needed;
 439         }
 440
 441         return outer_spills_needed;
 442 }
 443
 444 static void *init_phase_data(phase_t *phase, ir_node *irn, void *old) {
 445         return old;
 446 }
 447
 448 void be_spill_morgan(const be_chordal_env_t *chordal_env) {
 449         morgan_env_t env;
 450
 451         FIRM_DBG_REGISTER(dbg, "ir.be.spillmorgan");
 452
 453         env.arch = chordal_env->birg->main_env->arch_env;
 454         env.irg = chordal_env->irg;
 455         env.cls = chordal_env->cls;
 456         env.senv = be_new_spill_env(chordal_env);
 457         DEBUG_ONLY(be_set_spill_env_dbg_module(env.senv, dbg);)
 458
 459         phase_init(&env.phase, "spillmorgan", env.irg, PHASE_DEFAULT_GROWTH, init_phase_data);
 460
 461         env.registers_available = arch_count_non_ignore_regs(env.arch, env.cls);
 462
 463         env.loop_attr_set = new_set(loop_attr_cmp, 5);
 464         env.block_attr_set = new_set(block_attr_cmp, 20);
 465
 466         /*-- Part1: Analysis --*/
 467         be_liveness(env.irg);
 468
 469         // construct control flow loop tree
 470         construct_cf_backedges(chordal_env->irg);
 471
 472         // construct loop out edges and livethrough_unused sets for loops and blocks
 473         irg_block_walk_graph(chordal_env->irg, construct_loop_out_edges, NULL, &env);
 474         construct_loop_livethrough_unused(&env, get_irg_loop(env.irg));
 475
 476         /*-- Part2: Transformation --*/
 477
 478         // reduce register pressure to number of available registers
 479         reduce_register_pressure_in_loop(&env, get_irg_loop(env.irg), 0);
 480
 481         be_insert_spills_reloads(env.senv);
 482         if (chordal_env->opts->vrfy_option == BE_CH_VRFY_WARN) {
 483                 be_verify_schedule(env.irg);
 484         } else if (chordal_env->opts->vrfy_option == BE_CH_VRFY_ASSERT) {
 485                 assert(be_verify_schedule(env.irg));
 486         }
 487
 488         // cleanup
 489         be_dump(env.irg, "-spillmorgan", dump_ir_block_graph_sched);
 490         free_loop_out_edges(&env);
 491         del_set(env.loop_attr_set);
 492         del_set(env.block_attr_set);
 493
 494         // fix the remaining places with too high register pressure with beladies algorithm
 495
 496         // we have to remove dead nodes from schedule to not confuse liveness calculation
 497         be_remove_dead_nodes_from_schedule(env.irg);
 498         be_liveness(env.irg);
 499         be_spill_belady_spill_env(chordal_env, env.senv);
 500
 501         be_delete_spill_env(env.senv);
 502         phase_free(&env.phase);
 503 }