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_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 "irphase_t.h"
  23 #include "irprintf.h"
  24
  25 #include "bespillbelady.h"
  26 #include "beverify.h"
  27
  28 #define DBG_LIVE                1
  29 #define DBG_PRESSURE    2
  30 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  31
  32 typedef struct _morgan_env_t {
  33         const arch_env_t *arch;
  34         const arch_register_class_t *cls;
  35         ir_graph *irg;
  36         phase_t phase;
  37         // maximum safe register pressure
  38         int registers_available;
  39
  40         spill_env_t *senv;
  41
  42         set *loop_attr_set;
  43         set *block_attr_set;
  44 } morgan_env_t;
  45
  46 typedef struct _loop_out_edge_t {
  47         ir_node *block;
  48         int pos;
  49 } loop_out_edge_t;
  50
  51 typedef struct _loop_attr_t {
  52         ir_loop *loop;
  53         set *out_edges;
  54         /// The set of all values that live through the loop and are not used
  55         bitset_t *livethrough_unused;
  56 } loop_attr_t;
  57
  58 typedef struct _block_attr_t {
  59         ir_node *block;
  60         bitset_t *livethrough_unused;
  61 } block_attr_t;
  62
  63 //---------------------------------------------------------------------------
  64
  65 int loop_out_edge_cmp(const void* p1, const void* p2, size_t s) {
  66         loop_out_edge_t *e1 = (loop_out_edge_t*) p1;
  67         loop_out_edge_t *e2 = (loop_out_edge_t*) p2;
  68
  69         return e1->block != e2->block || e1->pos != e2->pos;
  70 }
  71
  72 int loop_attr_cmp(const void *e1, const void *e2, size_t s) {
  73         loop_attr_t *la1 = (loop_attr_t*) e1;
  74         loop_attr_t *la2 = (loop_attr_t*) e2;
  75
  76         return la1->loop != la2->loop;
  77 }
  78
  79 int block_attr_cmp(const void *e1, const void *e2, size_t s) {
  80         block_attr_t *b1 = (block_attr_t*) e1;
  81         block_attr_t *b2 = (block_attr_t*) e2;
  82
  83         return b1->block != b2->block;
  84 }
  85
  86 static INLINE int loop_attr_hash(const loop_attr_t *a) {
  87         return HASH_PTR(a->loop);
  88 }
  89
  90 static INLINE int block_attr_hash(const block_attr_t *b) {
  91         return HASH_PTR(b->block);
  92 }
  93
  94 static INLINE int loop_out_edge_hash(const loop_out_edge_t *e) {
  95         return HASH_PTR(e->block) ^ (e->pos * 31);
  96 }
  97
  98 static INLINE loop_attr_t *get_loop_attr(morgan_env_t *env, ir_loop *loop) {
  99         loop_attr_t l_attr, *res;
 100         int hash;
 101         l_attr.loop = loop;
 102
 103         hash = loop_attr_hash(&l_attr);
 104         res = set_find(env->loop_attr_set, &l_attr, sizeof(l_attr), hash);
 105
 106         // create new loop_attr if none exists yet
 107         if (!res) {
 108                 l_attr.out_edges = new_set(loop_out_edge_cmp, 1);
 109                 l_attr.livethrough_unused = bitset_obstack_alloc(&env->phase.obst, get_irg_last_idx(env->irg));
 110                 res = set_insert(env->loop_attr_set, &l_attr, sizeof(l_attr), hash);
 111         }
 112
 113         return res;
 114 }
 115
 116 static INLINE block_attr_t *get_block_attr(morgan_env_t *env, ir_node *block) {
 117         block_attr_t b_attr, *res;
 118         int hash;
 119         b_attr.block = block;
 120
 121         hash = block_attr_hash(&b_attr);
 122         res = set_find(env->block_attr_set, &b_attr, sizeof(b_attr), hash);
 123
 124         if(!res) {
 125                 b_attr.livethrough_unused = bitset_obstack_alloc(&env->phase.obst, get_irg_last_idx(env->irg));
 126                 res = set_insert(env->block_attr_set, &b_attr, sizeof(b_attr), hash);
 127         }
 128
 129         return res;
 130 }
 131
 132 //---------------------------------------------------------------------------
 133
 134 static INLINE int consider_for_spilling(const arch_env_t *env, const arch_register_class_t *cls, const ir_node *node) {
 135         if(!arch_irn_has_reg_class(env, node, -1, cls))
 136                 return 0;
 137
 138         return !(arch_irn_get_flags(env, node) & (arch_irn_flags_ignore | arch_irn_flags_dont_spill));
 139 }
 140
 141 /**
 142  * Determine edges going out of a loop (= edges that go to a block that is not inside
 143  * the loop or one of its subloops)
 144  */
 145 static INLINE void construct_loop_out_edges(ir_node* block, void* e) {
 146         morgan_env_t *env = (morgan_env_t*) e;
 147         int n_cfgpreds = get_Block_n_cfgpreds(block);
 148         int i;
 149         ir_loop* loop = get_irn_loop(block);
 150
 151         for(i = 0; i < n_cfgpreds; ++i) {
 152                 ir_node* cfgpred = get_Block_cfgpred(block, i);
 153                 ir_node* cfgpred_block = get_nodes_block(cfgpred);
 154                 ir_loop* cfgpred_loop = get_irn_loop(cfgpred_block);
 155                 loop_attr_t *outedges = get_loop_attr(env, cfgpred_loop);
 156
 157                 if(cfgpred_loop != loop && get_loop_depth(cfgpred_loop) >= get_loop_depth(loop)) {
 158                         loop_out_edge_t edge;
 159                         edge.block = block;
 160                         edge.pos = i;
 161                         set_insert(outedges->out_edges, &edge, sizeof(edge), loop_out_edge_hash(&edge));
 162                 }
 163         }
 164 }
 165
 166 static void free_loop_out_edges(morgan_env_t *env) {
 167         loop_attr_t *l_attr;
 168
 169         for(l_attr = set_first(env->loop_attr_set); l_attr != NULL; l_attr = set_next(env->loop_attr_set)) {
 170                 del_set(l_attr->out_edges);
 171         }
 172 }
 173
 174 /**
 175  * Debugging help, shows all nodes in a (node-)bitset
 176  */
 177 static void show_nodebitset(ir_graph* irg, bitset_t* bitset) {
 178         int i;
 179
 180         bitset_foreach(bitset, i) {
 181                 ir_node* node = get_idx_irn(irg, i);
 182                 DBG((dbg, DBG_LIVE, "\t%+F\n", node));
 183         }
 184 }
 185
 186 /**
 187  * Construct the livethrough unused set for a block
 188  */
 189 static bitset_t *construct_block_livethrough_unused(morgan_env_t* env, ir_node* block) {
 190         block_attr_t *block_attr = get_block_attr(env, block);
 191         irn_live_t *li;
 192         ir_node *node;
 193
 194         DBG((dbg, DBG_LIVE, "Processing block %d\n", get_irn_node_nr(block)));
 195         // copy all live-outs into the livethrough_unused set
 196         live_foreach(block, li) {
 197                 int node_idx;
 198
 199                 if(!live_is_in(li) || !live_is_out(li))
 200                         continue;
 201                 if(!consider_for_spilling(env->arch, env->cls, li->irn))
 202                         continue;
 203
 204                 node_idx = get_irn_idx(li->irn);
 205                 bitset_set(block_attr->livethrough_unused, node_idx);
 206         }
 207
 208         /*
 209          * All values that are used within the block are not unused (and therefore not
 210          * livethrough_unused)
 211          */
 212         sched_foreach(block, node) {
 213                 int i, arity;
 214
 215                 for(i = 0, arity = get_irn_arity(node); i < arity; ++i) {
 216                         int idx = get_irn_idx(get_irn_n(node, i));
 217                         bitset_clear(block_attr->livethrough_unused, idx);
 218                 }
 219         }
 220
 221         show_nodebitset(env->irg, block_attr->livethrough_unused);
 222         return block_attr->livethrough_unused;
 223 }
 224
 225 /**
 226  * Construct the livethrough unused set for a loop (and all its subloops+blocks)
 227  */
 228 static bitset_t *construct_loop_livethrough_unused(morgan_env_t *env, ir_loop *loop) {
 229         int i;
 230         loop_attr_t* loop_attr = get_loop_attr(env, loop);
 231
 232         DBG((dbg, DBG_LIVE, "Processing Loop %d\n", loop->loop_nr));
 233         assert(get_loop_n_elements(loop) > 0);
 234         for(i = 0; i < get_loop_n_elements(loop); ++i) {
 235                 loop_element elem = get_loop_element(loop, i);
 236                 switch (*elem.kind) {
 237                 case k_ir_node: {
 238                         bitset_t *livethrough_block_unused;
 239                         assert(is_Block(elem.node));
 240                         livethrough_block_unused = construct_block_livethrough_unused(env, elem.node);
 241                         if(i == 0) {
 242                                 bitset_copy(loop_attr->livethrough_unused, livethrough_block_unused);
 243                         } else {
 244                                 bitset_and(loop_attr->livethrough_unused, livethrough_block_unused);
 245                         }
 246                         break;
 247                 }
 248                 case k_ir_loop: {
 249                         bitset_t *livethrough_son_unused;
 250
 251                         livethrough_son_unused = construct_loop_livethrough_unused(env, elem.son);
 252                         if(i == 0) {
 253                                 bitset_copy(loop_attr->livethrough_unused, livethrough_son_unused);
 254                         } else {
 255                                 bitset_and(loop_attr->livethrough_unused, livethrough_son_unused);
 256                         }
 257                         break;
 258                 }
 259             default:
 260                         assert(0);
 261                         break;
 262                 }
 263     }
 264         DBG((dbg, DBG_LIVE, "Done with loop %d\n", loop->loop_nr));
 265
 266         // remove all unused livethroughs that are remembered for this loop from child loops and blocks
 267         for(i = 0; i < get_loop_n_elements(loop); ++i) {
 268                 const loop_element elem = get_loop_element(loop, i);
 269
 270                 if(*elem.kind == k_ir_loop) {
 271                         loop_attr_t *son_attr = get_loop_attr(env, elem.son);
 272                         bitset_andnot(son_attr->livethrough_unused, loop_attr->livethrough_unused);
 273
 274                         DBG((dbg, DBG_LIVE, "Livethroughs for loop %d:\n", loop->loop_nr));
 275                         show_nodebitset(env->irg, son_attr->livethrough_unused);
 276                 } else if(*elem.kind == k_ir_node) {
 277                         block_attr_t *block_attr = get_block_attr(env, elem.node);
 278                         bitset_andnot(block_attr->livethrough_unused, loop_attr->livethrough_unused);
 279
 280                         DBG((dbg, DBG_LIVE, "Livethroughs for block %+F\n", elem.node));
 281                         show_nodebitset(env->irg, block_attr->livethrough_unused);
 282                 } else {
 283                         assert(0);
 284                 }
 285         }
 286
 287         return loop_attr->livethrough_unused;
 288 }
 289
 290 //---------------------------------------------------------------------------
 291
 292 static int reduce_register_pressure_in_block(morgan_env_t *env, ir_node* block, int loop_unused_spills_possible) {
 293         int pressure;
 294         ir_node *irn;
 295         int max_pressure = 0;
 296         int spills_needed;
 297         int loop_unused_spills_needed;
 298         block_attr_t *block_attr = get_block_attr(env, block);
 299         int block_unused_spills_possible = bitset_popcnt(block_attr->livethrough_unused);
 300         int unused_spills_possible = loop_unused_spills_possible + block_unused_spills_possible;
 301         pset *live_nodes = pset_new_ptr_default();
 302
 303         be_liveness_end_of_block(env->arch, env->cls, block, live_nodes);
 304         pressure = pset_count(live_nodes);
 305
 306         DBG((dbg, DBG_LIVE, "Reduce pressure to %d In Block %+F:\n", env->registers_available, block));
 307
 308         /**
 309          * Walk over all irns in the schedule and check register pressure for each of them
 310          */
 311         sched_foreach_reverse(block, irn) {
 312                 // do we need more spills than possible with unused libethroughs?
 313                 int spills_needed = pressure - env->registers_available - unused_spills_possible;
 314                 if(spills_needed > 0) {
 315                         DBG((dbg, DBG_PRESSURE, "\tWARNING %d more spills needed at %+F\n", spills_needed, irn));
 316                         // TODO further spills needed
 317                         //assert(0);
 318                 }
 319                 if(pressure > max_pressure) {
 320                         max_pressure = pressure;
 321                 }
 322
 323                 /* Register pressure is only important until we reach the first phi (the rest of the block
 324                  * will only be phis.)
 325                  */
 326                 if(is_Phi(irn))
 327                         break;
 328
 329                 // update pressure
 330                 be_liveness_transfer(env->arch, env->cls, irn, live_nodes);
 331                 pressure = pset_count(live_nodes);
 332         }
 333
 334         DBG((dbg, DBG_PRESSURE, "\tMax Pressure in %+F: %d\n", block, max_pressure));
 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                         spills++;
 370                 }
 371         } else {
 372                 loop_unused_spills_needed = spills_needed;
 373         }
 374
 375         del_pset(live_nodes);
 376
 377         DBG((dbg, DBG_PRESSURE, "Unused spills for Block %+F needed: %d\n", block, loop_unused_spills_needed));
 378         return loop_unused_spills_needed;
 379 }
 380
 381 /**
 382  * Reduce register pressure in a loop
 383  *
 384  * @param unused_spills_possible        Number of spills from livethrough_unused variables possible in outer loops
 385  * @return                                                      Number of spills of livethrough_unused variables needed in outer loops
 386  */
 387 static int reduce_register_pressure_in_loop(morgan_env_t *env, ir_loop *loop, int outer_spills_possible) {
 388         int i;
 389         loop_attr_t* loop_attr = get_loop_attr(env, loop);
 390         int spills_needed = 0;
 391         int spills_possible = outer_spills_possible + bitset_popcnt(loop_attr->livethrough_unused);
 392         int outer_spills_needed;
 393
 394         DBG((dbg, DBG_PRESSURE, "Reducing Pressure in loop %d\n", loop->loop_nr));
 395         for(i = 0; i < get_loop_n_elements(loop); ++i) {
 396                 loop_element elem = get_loop_element(loop, i);
 397                 switch (*elem.kind) {
 398                 case k_ir_node: {
 399                         int needed;
 400                         assert(is_Block(elem.node));
 401                         needed = reduce_register_pressure_in_block(env, elem.node, spills_possible);
 402                         assert(needed <= spills_possible);
 403                         if(needed > spills_needed)
 404                                 spills_needed = needed;
 405                         break;
 406                 }
 407                 case k_ir_loop: {
 408                         int needed = reduce_register_pressure_in_loop(env, elem.son, spills_possible);
 409                         assert(needed <= spills_possible);
 410                         if(needed > spills_needed)
 411                                 spills_needed = needed;
 412                         break;
 413                 }
 414             default:
 415                         assert(0);
 416                         break;
 417                 }
 418     }
 419
 420         // calculate number of spills needed in outer loop and spill
 421         // unused livethrough nodes around this loop
 422         if(spills_needed > outer_spills_possible) {
 423                 outer_spills_needed = outer_spills_possible;
 424                 spills_needed -= outer_spills_possible;
 425
 426                 bitset_foreach(loop_attr->livethrough_unused, i) {
 427                         loop_out_edge_t *edge;
 428                         ir_node *to_spill = get_idx_irn(env->irg, i);
 429
 430                         for(edge = set_first(loop_attr->out_edges); edge != NULL; edge = set_next(loop_attr->out_edges)) {
 431                                 be_add_reload_on_edge(env->senv, to_spill, edge->block, edge->pos);
 432                         }
 433                 }
 434         } else {
 435                 outer_spills_needed = spills_needed;
 436         }
 437
 438         return outer_spills_needed;
 439 }
 440
 441 static void *init_phase_data(phase_t *phase, ir_node *irn, void *old) {
 442         return old;
 443 }
 444
 445 void be_spill_morgan(const be_chordal_env_t *chordal_env) {
 446         morgan_env_t env;
 447
 448         FIRM_DBG_REGISTER(dbg, "ir.be.spillmorgan");
 449
 450         env.arch = chordal_env->birg->main_env->arch_env;
 451         env.irg = chordal_env->irg;
 452         env.cls = chordal_env->cls;
 453         env.senv = be_new_spill_env(chordal_env);
 454         DEBUG_ONLY(be_set_spill_env_dbg_module(env.senv, dbg);)
 455
 456         phase_init(&env.phase, "spillmorgan", env.irg, PHASE_DEFAULT_GROWTH, init_phase_data);
 457
 458         env.registers_available = arch_count_non_ignore_regs(env.arch, env.cls);
 459
 460         env.loop_attr_set = new_set(loop_attr_cmp, 5);
 461         env.block_attr_set = new_set(block_attr_cmp, 20);
 462
 463         /*-- Part1: Analysis --*/
 464         be_liveness(env.irg);
 465
 466         // construct control flow loop tree
 467         construct_cf_backedges(chordal_env->irg);
 468
 469         // construct loop out edges and livethrough_unused sets for loops and blocks
 470         irg_block_walk_graph(chordal_env->irg, construct_loop_out_edges, NULL, &env);
 471         construct_loop_livethrough_unused(&env, get_irg_loop(env.irg));
 472
 473         /*-- Part2: Transformation --*/
 474
 475         // reduce register pressure to number of available registers
 476         reduce_register_pressure_in_loop(&env, get_irg_loop(env.irg), 0);
 477
 478         be_insert_spills_reloads(env.senv);
 479         if (chordal_env->opts->vrfy_option == BE_CH_VRFY_WARN) {
 480                 be_verify_schedule(env.irg);
 481         } else if (chordal_env->opts->vrfy_option == BE_CH_VRFY_ASSERT) {
 482                 assert(be_verify_schedule(env.irg));
 483         }
 484
 485         // we have to remove dead nodes from schedule to not confuse liveness calculation
 486         be_remove_dead_nodes_from_schedule(env.irg);
 487
 488         // cleanup
 489         if (chordal_env->opts->dump_flags & BE_CH_DUMP_SPILL)
 490                 be_dump(env.irg, "-spillmorgan", dump_ir_block_graph_sched);
 491
 492         free_loop_out_edges(&env);
 493         del_set(env.loop_attr_set);
 494         del_set(env.block_attr_set);
 495
 496         // fix the remaining places with too high register pressure with beladies algorithm
 497         be_liveness(env.irg);
 498         be_spill_belady_spill_env(chordal_env, env.senv);
 499
 500         be_delete_spill_env(env.senv);
 501         phase_free(&env.phase);
 502 }