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 be_chordal_env_t *cenv;
  35         const arch_env_t *arch;
  36         const arch_register_class_t *cls;
  37         ir_graph *irg;
  38         struct obstack obst;
  39         /** maximum safe register pressure */
  40         int registers_available;
  41
  42         spill_env_t *senv;
  43
  44         set *loop_attr_set;
  45         set *block_attr_set;
  46 } morgan_env_t;
  47
  48 typedef struct loop_edge {
  49         ir_node *block;
  50         int pos;
  51 } loop_edge_t;
  52
  53 typedef struct loop_attr {
  54         const ir_loop *loop;
  55         set *out_edges;
  56         set *in_edges;
  57         /** The set of all values that are live in the loop but not used in the loop */
  58         bitset_t *livethrough_unused;
  59 } loop_attr_t;
  60
  61 typedef struct block_attr {
  62         const ir_node *block;
  63         /** set of all values that are live in the block but not used in the block */
  64         bitset_t *livethrough_unused;
  65 } block_attr_t;
  66
  67 //---------------------------------------------------------------------------
  68
  69 static int loop_edge_cmp(const void* p1, const void* p2, size_t s) {
  70         loop_edge_t *e1 = (loop_edge_t*) p1;
  71         loop_edge_t *e2 = (loop_edge_t*) p2;
  72
  73         return e1->block != e2->block || e1->pos != e2->pos;
  74 }
  75
  76 static int loop_attr_cmp(const void *e1, const void *e2, size_t s) {
  77         loop_attr_t *la1 = (loop_attr_t*) e1;
  78         loop_attr_t *la2 = (loop_attr_t*) e2;
  79
  80         return la1->loop != la2->loop;
  81 }
  82
  83 static int block_attr_cmp(const void *e1, const void *e2, size_t s) {
  84         block_attr_t *b1 = (block_attr_t*) e1;
  85         block_attr_t *b2 = (block_attr_t*) e2;
  86
  87         return b1->block != b2->block;
  88 }
  89
  90 static INLINE int loop_attr_hash(const loop_attr_t *a) {
  91         return HASH_PTR(a->loop);
  92 }
  93
  94 static INLINE int block_attr_hash(const block_attr_t *b) {
  95         return HASH_PTR(b->block);
  96 }
  97
  98 static INLINE int loop_edge_hash(const loop_edge_t *e) {
  99         return HASH_PTR(e->block) ^ (e->pos * 31);
 100 }
 101
 102 static INLINE loop_attr_t *get_loop_attr(morgan_env_t *env, const ir_loop *loop) {
 103         loop_attr_t l_attr, *res;
 104         int hash;
 105         l_attr.loop = loop;
 106
 107         hash = loop_attr_hash(&l_attr);
 108         res = set_find(env->loop_attr_set, &l_attr, sizeof(l_attr), hash);
 109
 110         // create new loop_attr if none exists yet
 111         if (res == NULL) {
 112                 l_attr.out_edges = new_set(loop_edge_cmp, 1);
 113                 l_attr.in_edges = new_set(loop_edge_cmp, 1);
 114                 l_attr.livethrough_unused = bitset_obstack_alloc(&env->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, const 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 == NULL) {
 130                 b_attr.livethrough_unused = bitset_obstack_alloc(&env->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 static INLINE int consider_for_spilling(const arch_env_t *env, const arch_register_class_t *cls, const ir_node *node) {
 140         if(!arch_irn_has_reg_class(env, node, -1, cls))
 141                 return 0;
 142
 143         return !(arch_irn_get_flags(env, node) & (arch_irn_flags_ignore | arch_irn_flags_dont_spill));
 144 }
 145
 146 /**
 147  * Determine edges going out of a loop (= edges that go to a block that is not inside
 148  * the loop or one of its subloops)
 149  */
 150 static INLINE void construct_loop_edges(ir_node* block, void* e) {
 151         morgan_env_t *env = (morgan_env_t*) e;
 152         int n_cfgpreds = get_Block_n_cfgpreds(block);
 153         int i;
 154         ir_loop* loop = get_irn_loop(block);
 155         loop_attr_t *loop_attr = get_loop_attr(env, loop);
 156         DBG((dbg, DBG_LOOPANA, "Loop for %+F: %d (depth %d)\n", block, loop->loop_nr, loop->depth));
 157
 158         for(i = 0; i < n_cfgpreds; ++i) {
 159                 ir_node* cfgpred = get_Block_cfgpred(block, i);
 160                 ir_node* cfgpred_block = get_nodes_block(cfgpred);
 161                 ir_loop* cfgpred_loop = get_irn_loop(cfgpred_block);
 162                 loop_attr_t *outedges = get_loop_attr(env, cfgpred_loop);
 163
 164                 if(cfgpred_loop == loop)
 165                         continue;
 166
 167                 // is it an edge into the loop?
 168                 if(get_loop_depth(loop) > get_loop_depth(cfgpred_loop)) {
 169                         loop_edge_t edge;
 170                         edge.block = block;
 171                         edge.pos = i;
 172                         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)));
 173                         set_insert(loop_attr->in_edges, &edge, sizeof(edge), loop_edge_hash(&edge));
 174                 } else {
 175                         ir_loop *p_loop = cfgpred_loop;
 176                         while(get_loop_depth(p_loop) > get_loop_depth(loop)) {
 177                                 p_loop = get_loop_outer_loop(p_loop);
 178                         }
 179                         if(p_loop != loop) {
 180                                 loop_edge_t edge;
 181                                 edge.block = block;
 182                                 edge.pos = i;
 183                                 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)));
 184                                 set_insert(loop_attr->in_edges, &edge, sizeof(edge), loop_edge_hash(&edge));
 185                         }
 186                 }
 187
 188                 // an edge out of the loop?
 189                 if(get_loop_depth(cfgpred_loop) >= get_loop_depth(loop)) {
 190                         loop_edge_t edge;
 191                         edge.block = block;
 192                         edge.pos = i;
 193                         DBG((dbg, DBG_LOOPANA, "Loop out edge from %+F (loop %d) to %+F\n", cfgpred_block, cfgpred_loop->loop_nr, block));
 194                         set_insert(outedges->out_edges, &edge, sizeof(edge), loop_edge_hash(&edge));
 195                 } else {
 196                         ir_loop *o_loop = loop;
 197
 198                         // we might jump in the middle of another inner loop which is not inside
 199                         // our loop (happens for irreducible graphs). This would be a
 200                         // real out edge then.
 201                         while(get_loop_depth(o_loop) > get_loop_depth(cfgpred_loop)) {
 202                                 o_loop = get_loop_outer_loop(o_loop);
 203                         }
 204
 205                         if(cfgpred_loop != o_loop) {
 206                                 loop_edge_t edge;
 207                                 edge.block = block;
 208                                 edge.pos = i;
 209                                 DBG((dbg, DBG_LOOPANA, "Loop out edge from %+F (loop %d) to %+F (into jump)\n", cfgpred_block, cfgpred_loop->loop_nr, block));
 210                                 set_insert(outedges->out_edges, &edge, sizeof(edge), loop_edge_hash(&edge));
 211                         }
 212                 }
 213         }
 214 }
 215
 216 static void free_loop_edges(morgan_env_t *env) {
 217         loop_attr_t *l_attr;
 218
 219         for(l_attr = set_first(env->loop_attr_set); l_attr != NULL; l_attr = set_next(env->loop_attr_set)) {
 220                 del_set(l_attr->out_edges);
 221                 del_set(l_attr->in_edges);
 222         }
 223 }
 224
 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                 DBG((dbg, DBG_LIVE, "\t%+F\n", node));
 234         }
 235 }
 236
 237 /**
 238  * Construct the livethrough unused set for a block
 239  */
 240 static bitset_t *construct_block_livethrough_unused(morgan_env_t* env, const ir_node* block) {
 241         block_attr_t *block_attr = get_block_attr(env, block);
 242         ir_node *node;
 243         int i;
 244
 245         DBG((dbg, DBG_LIVE, "Processing block %d\n", get_irn_node_nr(block)));
 246         // copy all live-outs into the livethrough_unused set
 247         be_lv_foreach(env->cenv->lv, block, be_lv_state_in | be_lv_state_out, i) {
 248                 ir_node *irn = be_lv_get_irn(env->cenv->lv, block, i);
 249                 int node_idx;
 250
 251                 /*
 252                 if(!live_is_in(li) || !live_is_out(li))
 253                         continue;
 254                 */
 255                 if(!consider_for_spilling(env->arch, env->cls, irn))
 256                         continue;
 257
 258                 node_idx = get_irn_idx(irn);
 259                 bitset_set(block_attr->livethrough_unused, node_idx);
 260         }
 261
 262         /*
 263          * All values that are used within the block are not unused (and therefore not
 264          * livethrough_unused)
 265          */
 266         sched_foreach(block, node) {
 267                 int i, arity;
 268
 269                 for(i = 0, arity = get_irn_arity(node); i < arity; ++i) {
 270                         int idx = get_irn_idx(get_irn_n(node, i));
 271                         bitset_clear(block_attr->livethrough_unused, idx);
 272                 }
 273         }
 274
 275         show_nodebitset(env->irg, block_attr->livethrough_unused);
 276         return block_attr->livethrough_unused;
 277 }
 278
 279 /**
 280  * Construct the livethrough unused set for a loop (and all its subloops+blocks)
 281  */
 282 static bitset_t *construct_loop_livethrough_unused(morgan_env_t *env, const ir_loop *loop) {
 283         int i;
 284         loop_attr_t* loop_attr = get_loop_attr(env, loop);
 285
 286         DBG((dbg, DBG_LIVE, "Processing Loop %d\n", loop->loop_nr));
 287         assert(get_loop_n_elements(loop) > 0);
 288         for(i = 0; i < get_loop_n_elements(loop); ++i) {
 289                 loop_element elem = get_loop_element(loop, i);
 290                 switch (*elem.kind) {
 291                 case k_ir_node: {
 292                         bitset_t *livethrough_block_unused;
 293                         assert(is_Block(elem.node));
 294                         livethrough_block_unused = construct_block_livethrough_unused(env, elem.node);
 295                         if(i == 0) {
 296                                 bitset_copy(loop_attr->livethrough_unused, livethrough_block_unused);
 297                         } else {
 298                                 bitset_and(loop_attr->livethrough_unused, livethrough_block_unused);
 299                         }
 300                         break;
 301                 }
 302                 case k_ir_loop: {
 303                         bitset_t *livethrough_son_unused;
 304
 305                         livethrough_son_unused = construct_loop_livethrough_unused(env, elem.son);
 306                         if(i == 0) {
 307                                 bitset_copy(loop_attr->livethrough_unused, livethrough_son_unused);
 308                         } else {
 309                                 bitset_and(loop_attr->livethrough_unused, livethrough_son_unused);
 310                         }
 311                         break;
 312                 }
 313             default:
 314                         assert(0);
 315                         break;
 316                 }
 317     }
 318         DBG((dbg, DBG_LIVE, "Done with loop %d\n", loop->loop_nr));
 319
 320         // remove all unused livethroughs that are remembered for this loop from child loops and blocks
 321         for(i = 0; i < get_loop_n_elements(loop); ++i) {
 322                 const loop_element elem = get_loop_element(loop, i);
 323
 324                 if(*elem.kind == k_ir_loop) {
 325                         loop_attr_t *son_attr = get_loop_attr(env, elem.son);
 326                         bitset_andnot(son_attr->livethrough_unused, loop_attr->livethrough_unused);
 327
 328                         DBG((dbg, DBG_LIVE, "Livethroughs for loop %d:\n", loop->loop_nr));
 329                         show_nodebitset(env->irg, son_attr->livethrough_unused);
 330                 } else if(*elem.kind == k_ir_node) {
 331                         block_attr_t *block_attr = get_block_attr(env, elem.node);
 332                         bitset_andnot(block_attr->livethrough_unused, loop_attr->livethrough_unused);
 333
 334                         DBG((dbg, DBG_LIVE, "Livethroughs for block %+F\n", elem.node));
 335                         show_nodebitset(env->irg, block_attr->livethrough_unused);
 336                 } else {
 337                         assert(0);
 338                 }
 339         }
 340
 341         return loop_attr->livethrough_unused;
 342 }
 343
 344 /*---------------------------------------------------------------------------*/
 345
 346 static int reduce_register_pressure_in_block(morgan_env_t *env, const ir_node* block, int loop_unused_spills_possible) {
 347         int pressure;
 348         ir_node *irn;
 349         int max_pressure = 0;
 350         int spills_needed;
 351         int loop_unused_spills_needed;
 352         block_attr_t *block_attr = get_block_attr(env, block);
 353         int block_unused_spills_possible = bitset_popcnt(block_attr->livethrough_unused);
 354         int unused_spills_possible = loop_unused_spills_possible + block_unused_spills_possible;
 355         pset *live_nodes = pset_new_ptr_default();
 356
 357         be_liveness_end_of_block(env->cenv->lv, env->arch, env->cls, block, live_nodes);
 358         pressure = pset_count(live_nodes);
 359
 360         DBG((dbg, DBG_LIVE, "Reduce pressure to %d In Block %+F:\n", env->registers_available, block));
 361
 362         /**
 363          * Walk over all irns in the schedule and check register pressure for each of them
 364          */
 365         sched_foreach_reverse(block, irn) {
 366                 // do we need more spills than possible with unused libethroughs?
 367                 int spills_needed = pressure - env->registers_available - unused_spills_possible;
 368                 if(spills_needed > 0) {
 369                         DBG((dbg, DBG_PRESSURE, "\tWARNING %d more spills needed at %+F\n", spills_needed, irn));
 370                         // TODO further spills needed
 371                         //assert(0);
 372                 }
 373                 if(pressure > max_pressure) {
 374                         max_pressure = pressure;
 375                 }
 376
 377                 /* Register pressure is only important until we reach the first phi (the rest of the block
 378                  * will only be phis.)
 379                  */
 380                 if(is_Phi(irn))
 381                         break;
 382
 383                 // update pressure
 384                 be_liveness_transfer(env->arch, env->cls, irn, live_nodes);
 385                 pressure = pset_count(live_nodes);
 386         }
 387
 388         DBG((dbg, DBG_PRESSURE, "\tMax Pressure in %+F: %d\n", block, max_pressure));
 389
 390         /*
 391          * Calculate number of spills from loop_unused_spills_possible that we want to use,
 392          * and spill unused livethroughs from the block if we still don't have enough registers
 393          */
 394         spills_needed = max_pressure - env->registers_available;
 395         if(spills_needed < 0) {
 396                 loop_unused_spills_needed = 0;
 397         } else if(spills_needed > loop_unused_spills_possible) {
 398                 int i, spills;
 399                 int block_unused_spills_needed;
 400
 401                 loop_unused_spills_needed = loop_unused_spills_possible;
 402                 block_unused_spills_needed = spills_needed - loop_unused_spills_possible;
 403                 if(block_unused_spills_needed > block_unused_spills_possible) {
 404                         block_unused_spills_needed = block_unused_spills_possible;
 405                 }
 406
 407                 spills = 0;
 408                 /*
 409                  * Spill/Reload unused livethroughs from the block
 410                  */
 411                 bitset_foreach(block_attr->livethrough_unused, i) {
 412                         ir_node *to_spill;
 413                         const ir_edge_t *edge;
 414
 415                         if(spills >= block_unused_spills_needed)
 416                                 break;
 417
 418                         to_spill = get_idx_irn(env->irg, i);
 419                         foreach_block_succ(block, edge) {
 420                                 DBG((dbg, DBG_PRESSURE, "Spilling node %+F around block %+F\n", to_spill, block));
 421                                 be_add_reload_on_edge(env->senv, to_spill, edge->src, edge->pos);
 422                         }
 423                         spills++;
 424                 }
 425         } else {
 426                 loop_unused_spills_needed = spills_needed;
 427         }
 428
 429         del_pset(live_nodes);
 430
 431         DBG((dbg, DBG_PRESSURE, "Unused spills for Block %+F needed: %d\n", block, loop_unused_spills_needed));
 432         return loop_unused_spills_needed;
 433 }
 434
 435 /**
 436  * Reduce register pressure in a loop
 437  *
 438  * @param unused_spills_possible        Number of spills from livethrough_unused variables possible in outer loops
 439  * @return                                                      Number of spills of livethrough_unused variables needed in outer loops
 440  */
 441 static int reduce_register_pressure_in_loop(morgan_env_t *env, const ir_loop *loop, int outer_spills_possible) {
 442         int i;
 443         loop_attr_t* loop_attr = get_loop_attr(env, loop);
 444         int spills_needed = 0;
 445         int spills_possible = outer_spills_possible + bitset_popcnt(loop_attr->livethrough_unused);
 446         int outer_spills_needed;
 447
 448         DBG((dbg, DBG_PRESSURE, "Reducing Pressure in loop %d\n", loop->loop_nr));
 449         for(i = 0; i < get_loop_n_elements(loop); ++i) {
 450                 loop_element elem = get_loop_element(loop, i);
 451                 switch (*elem.kind) {
 452                 case k_ir_node: {
 453                         int needed;
 454                         assert(is_Block(elem.node));
 455                         needed = reduce_register_pressure_in_block(env, elem.node, spills_possible);
 456                         assert(needed <= spills_possible);
 457                         if(needed > spills_needed)
 458                                 spills_needed = needed;
 459                         break;
 460                 }
 461                 case k_ir_loop: {
 462                         int needed = reduce_register_pressure_in_loop(env, elem.son, spills_possible);
 463                         assert(needed <= spills_possible);
 464                         if(needed > spills_needed)
 465                                 spills_needed = needed;
 466                         break;
 467                 }
 468             default:
 469                         assert(0);
 470                         break;
 471                 }
 472     }
 473
 474         /* calculate number of spills needed in outer loop and spill
 475          * unused livethrough nodes around this loop
 476          */
 477         if(spills_needed > outer_spills_possible) {
 478                 int spills_to_place;
 479                 outer_spills_needed = outer_spills_possible;
 480                 spills_needed -= outer_spills_possible;
 481
 482                 spills_to_place = spills_needed;
 483
 484                 bitset_foreach(loop_attr->livethrough_unused, i) {
 485                         loop_edge_t *edge;
 486                         ir_node *to_spill = get_idx_irn(env->irg, i);
 487
 488                         for(edge = set_first(loop_attr->out_edges); edge != NULL; edge = set_next(loop_attr->out_edges)) {
 489                                 DBG((dbg, DBG_PRESSURE, "Spilling node %+F around loop %d\n", to_spill, loop->loop_nr));
 490                                 be_add_reload_on_edge(env->senv, to_spill, edge->block, edge->pos);
 491                         }
 492
 493                         spills_to_place--;
 494                         if(spills_to_place <= 0) {
 495                                 break;
 496                         }
 497                 }
 498         } else {
 499                 outer_spills_needed = spills_needed;
 500         }
 501
 502         return outer_spills_needed;
 503 }
 504
 505 void be_spill_morgan(be_chordal_env_t *chordal_env) {
 506         morgan_env_t env;
 507
 508         FIRM_DBG_REGISTER(dbg, "ir.be.spillmorgan");
 509         //firm_dbg_set_mask(dbg, DBG_LOOPANA | DBG_PRESSURE);
 510
 511         env.cenv = chordal_env;
 512         env.arch = chordal_env->birg->main_env->arch_env;
 513         env.irg = chordal_env->irg;
 514         env.cls = chordal_env->cls;
 515         env.senv = be_new_spill_env(chordal_env);
 516         DEBUG_ONLY(be_set_spill_env_dbg_module(env.senv, dbg);)
 517
 518         obstack_init(&env.obst);
 519
 520         env.registers_available = arch_count_non_ignore_regs(env.arch, env.cls);
 521
 522         env.loop_attr_set = new_set(loop_attr_cmp, 5);
 523         env.block_attr_set = new_set(block_attr_cmp, 20);
 524
 525         /*-- Part1: Analysis --*/
 526         be_liveness_recompute(chordal_env->lv);
 527
 528         /* construct control flow loop tree */
 529         construct_cf_backedges(chordal_env->irg);
 530
 531         /* construct loop out edges and livethrough_unused sets for loops and blocks */
 532         irg_block_walk_graph(chordal_env->irg, NULL, construct_loop_edges, &env);
 533         construct_loop_livethrough_unused(&env, get_irg_loop(env.irg));
 534
 535         /*-- Part2: Transformation --*/
 536
 537         /* spill unused livethrough values around loops and blocks where
 538          * the pressure is too high
 539          */
 540         reduce_register_pressure_in_loop(&env, get_irg_loop(env.irg), 0);
 541
 542         /* Insert real spill/reload nodes and fix usages */
 543         be_insert_spills_reloads(env.senv);
 544
 545         /* Verify the result */
 546         if (chordal_env->opts->vrfy_option == BE_CH_VRFY_WARN) {
 547                 be_verify_schedule(env.irg);
 548         } else if (chordal_env->opts->vrfy_option == BE_CH_VRFY_ASSERT) {
 549                 assert(be_verify_schedule(env.irg));
 550         }
 551
 552         if (chordal_env->opts->dump_flags & BE_CH_DUMP_SPILL)
 553                 be_dump(env.irg, "-spillmorgan", dump_ir_block_graph_sched);
 554
 555         /* cleanup */
 556         free_loop_edges(&env);
 557         del_set(env.loop_attr_set);
 558         del_set(env.block_attr_set);
 559
 560         /* fix the remaining places with too high register pressure with beladies algorithm */
 561         be_spill_belady_spill_env(chordal_env, env.senv);
 562
 563         be_delete_spill_env(env.senv);
 564         obstack_free(&env.obst, NULL);
 565 }