nsz Git - libfirm/blob - ir/be/bespillmorgan.c

   1 /*
   2  * Author:      Matthias Braun
   3  * Date:                05.05.2006
   4  * Copyright:   (c) Universitaet Karlsruhe
   5  * License:     This file is protected by GPL -  GNU GENERAL PUBLIC LICENSE.
   6  *
   7  */
   8 #ifdef HAVE_CONFIG_H
   9 #include "config.h"
  10 #endif
  11
  12 #include "bespillmorgan.h"
  13
  14 #include "bechordal_t.h"
  15 #include "bespill.h"
  16 #include "belive_t.h"
  17 #include "beabi.h"
  18 #include "irgwalk.h"
  19 #include "besched.h"
  20 #include "beutil.h"
  21 #include "irloop_t.h"
  22 #include "irgraph_t.h"
  23 #include "irprintf.h"
  24 #include "obstack.h"
  25
  26 #include "bespillbelady.h"
  27 #include "beverify.h"
  28
  29 #define DBG_LIVE                1
  30 #define DBG_LOOPANA             2
  31 #define DBG_PRESSURE    4
  32 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  33
  34 typedef struct morgan_env {
  35         const be_chordal_env_t *cenv;
  36         const arch_env_t *arch;
  37         const arch_register_class_t *cls;
  38         ir_graph *irg;
  39         struct obstack obst;
  40         /** maximum safe register pressure */
  41         int registers_available;
  42
  43         spill_env_t *senv;
  44
  45         set *loop_attr_set;
  46         set *block_attr_set;
  47 } morgan_env_t;
  48
  49 typedef struct loop_edge {
  50         ir_node *block;
  51         int pos;
  52 } loop_edge_t;
  53
  54 typedef struct loop_attr {
  55         const ir_loop *loop;
  56         set *out_edges;
  57         set *in_edges;
  58         /** The set of all values that are live in the loop but not used in the loop */
  59         bitset_t *livethrough_unused;
  60 } loop_attr_t;
  61
  62 typedef struct block_attr {
  63         const ir_node *block;
  64         /** set of all values that are live in the block but not used in the block */
  65         bitset_t *livethrough_unused;
  66 } block_attr_t;
  67
  68 //---------------------------------------------------------------------------
  69
  70 static int loop_edge_cmp(const void* p1, const void* p2, size_t s) {
  71         loop_edge_t *e1 = (loop_edge_t*) p1;
  72         loop_edge_t *e2 = (loop_edge_t*) p2;
  73
  74         return e1->block != e2->block || e1->pos != e2->pos;
  75 }
  76
  77 static 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 static 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_edge_hash(const loop_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, const 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 == NULL) {
 113                 l_attr.out_edges = new_set(loop_edge_cmp, 1);
 114                 l_attr.in_edges = new_set(loop_edge_cmp, 1);
 115                 l_attr.livethrough_unused = bitset_obstack_alloc(&env->obst, get_irg_last_idx(env->irg));
 116                 res = set_insert(env->loop_attr_set, &l_attr, sizeof(l_attr), hash);
 117         }
 118
 119         return res;
 120 }
 121
 122 static INLINE block_attr_t *get_block_attr(morgan_env_t *env, const ir_node *block) {
 123         block_attr_t b_attr, *res;
 124         int hash;
 125         b_attr.block = block;
 126
 127         hash = block_attr_hash(&b_attr);
 128         res = set_find(env->block_attr_set, &b_attr, sizeof(b_attr), hash);
 129
 130         if(res == NULL) {
 131                 b_attr.livethrough_unused = bitset_obstack_alloc(&env->obst, get_irg_last_idx(env->irg));
 132                 res = set_insert(env->block_attr_set, &b_attr, sizeof(b_attr), hash);
 133         }
 134
 135         return res;
 136 }
 137
 138 //---------------------------------------------------------------------------
 139
 140 static INLINE int consider_for_spilling(const arch_env_t *env, const arch_register_class_t *cls, const ir_node *node) {
 141         if(!arch_irn_has_reg_class(env, node, -1, cls))
 142                 return 0;
 143
 144         return !(arch_irn_get_flags(env, node) & (arch_irn_flags_ignore | arch_irn_flags_dont_spill));
 145 }
 146
 147 /**
 148  * Determine edges going out of a loop (= edges that go to a block that is not inside
 149  * the loop or one of its subloops)
 150  */
 151 static INLINE void construct_loop_edges(ir_node* block, void* e) {
 152         morgan_env_t *env = (morgan_env_t*) e;
 153         int n_cfgpreds = get_Block_n_cfgpreds(block);
 154         int i;
 155         ir_loop* loop = get_irn_loop(block);
 156         loop_attr_t *loop_attr = get_loop_attr(env, loop);
 157         DBG((dbg, DBG_LOOPANA, "Loop for %+F: %d (depth %d)\n", block, loop->loop_nr, loop->depth));
 158
 159         for(i = 0; i < n_cfgpreds; ++i) {
 160                 ir_node* cfgpred = get_Block_cfgpred(block, i);
 161                 ir_node* cfgpred_block = get_nodes_block(cfgpred);
 162                 ir_loop* cfgpred_loop = get_irn_loop(cfgpred_block);
 163                 loop_attr_t *outedges = get_loop_attr(env, cfgpred_loop);
 164
 165                 if(cfgpred_loop == loop)
 166                         continue;
 167
 168                 // is it an edge into the loop?
 169                 if(get_loop_depth(loop) > get_loop_depth(cfgpred_loop)) {
 170                         loop_edge_t edge;
 171                         edge.block = block;
 172                         edge.pos = i;
 173                         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)));
 174                         set_insert(loop_attr->in_edges, &edge, sizeof(edge), loop_edge_hash(&edge));
 175                 } else {
 176                         ir_loop *p_loop = cfgpred_loop;
 177                         while(get_loop_depth(p_loop) > get_loop_depth(loop)) {
 178                                 p_loop = get_loop_outer_loop(p_loop);
 179                         }
 180                         if(p_loop != loop) {
 181                                 loop_edge_t edge;
 182                                 edge.block = block;
 183                                 edge.pos = i;
 184                                 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)));
 185                                 set_insert(loop_attr->in_edges, &edge, sizeof(edge), loop_edge_hash(&edge));
 186                         }
 187                 }
 188
 189                 // an edge out of the loop?
 190                 if(get_loop_depth(cfgpred_loop) >= get_loop_depth(loop)) {
 191                         loop_edge_t edge;
 192                         edge.block = block;
 193                         edge.pos = i;
 194                         DBG((dbg, DBG_LOOPANA, "Loop out edge from %+F (loop %d) to %+F\n", cfgpred_block, cfgpred_loop->loop_nr, block));
 195                         set_insert(outedges->out_edges, &edge, sizeof(edge), loop_edge_hash(&edge));
 196                 } else {
 197                         ir_loop *o_loop = loop;
 198
 199                         // we might jump in the middle of another inner loop which is not inside
 200                         // our loop (happens for irreducible graphs). This would be a
 201                         // real out edge then.
 202                         while(get_loop_depth(o_loop) > get_loop_depth(cfgpred_loop)) {
 203                                 o_loop = get_loop_outer_loop(o_loop);
 204                         }
 205
 206                         if(cfgpred_loop != o_loop) {
 207                                 loop_edge_t edge;
 208                                 edge.block = block;
 209                                 edge.pos = i;
 210                                 DBG((dbg, DBG_LOOPANA, "Loop out edge from %+F (loop %d) to %+F (into jump)\n", cfgpred_block, cfgpred_loop->loop_nr, block));
 211                                 set_insert(outedges->out_edges, &edge, sizeof(edge), loop_edge_hash(&edge));
 212                         }
 213                 }
 214         }
 215 }
 216
 217 static void free_loop_edges(morgan_env_t *env) {
 218         loop_attr_t *l_attr;
 219
 220         for(l_attr = set_first(env->loop_attr_set); l_attr != NULL; l_attr = set_next(env->loop_attr_set)) {
 221                 del_set(l_attr->out_edges);
 222                 del_set(l_attr->in_edges);
 223         }
 224 }
 225
 226 /**
 227  * Debugging help, shows all nodes in a (node-)bitset
 228  */
 229 static void show_nodebitset(ir_graph* irg, const bitset_t* bitset) {
 230         int i;
 231
 232         bitset_foreach(bitset, i) {
 233                 ir_node* node = get_idx_irn(irg, i);
 234                 DBG((dbg, DBG_LIVE, "\t%+F\n", node));
 235         }
 236 }
 237
 238 /**
 239  * Construct the livethrough unused set for a block
 240  */
 241 static bitset_t *construct_block_livethrough_unused(morgan_env_t* env, const ir_node* block) {
 242         block_attr_t *block_attr = get_block_attr(env, block);
 243         ir_node *node;
 244         int i;
 245
 246         DBG((dbg, DBG_LIVE, "Processing block %d\n", get_irn_node_nr(block)));
 247         // copy all live-outs into the livethrough_unused set
 248         be_lv_foreach(env->cenv->lv, block, be_lv_state_in | be_lv_state_out, i) {
 249                 ir_node *irn = be_lv_get_irn(env->cenv->lv, block, i);
 250                 int node_idx;
 251
 252                 if(!consider_for_spilling(env->arch, env->cls, irn))
 253                         continue;
 254
 255                 node_idx = get_irn_idx(irn);
 256                 bitset_set(block_attr->livethrough_unused, node_idx);
 257         }
 258
 259         /*
 260          * All values that are used within the block are not unused (and therefore not
 261          * livethrough_unused)
 262          */
 263         sched_foreach(block, node) {
 264                 int i, arity;
 265
 266                 for(i = 0, arity = get_irn_arity(node); i < arity; ++i) {
 267                         int idx = get_irn_idx(get_irn_n(node, i));
 268                         bitset_clear(block_attr->livethrough_unused, idx);
 269                 }
 270         }
 271
 272         show_nodebitset(env->irg, block_attr->livethrough_unused);
 273         return block_attr->livethrough_unused;
 274 }
 275
 276 /**
 277  * Construct the livethrough unused set for a loop (and all its subloops+blocks)
 278  */
 279 static bitset_t *construct_loop_livethrough_unused(morgan_env_t *env, const ir_loop *loop) {
 280         int i;
 281         loop_attr_t* loop_attr = get_loop_attr(env, loop);
 282
 283         DBG((dbg, DBG_LIVE, "Processing Loop %d\n", loop->loop_nr));
 284         assert(get_loop_n_elements(loop) > 0);
 285         for(i = 0; i < get_loop_n_elements(loop); ++i) {
 286                 loop_element elem = get_loop_element(loop, i);
 287                 switch (*elem.kind) {
 288                 case k_ir_node: {
 289                         bitset_t *livethrough_block_unused;
 290                         assert(is_Block(elem.node));
 291                         livethrough_block_unused = construct_block_livethrough_unused(env, elem.node);
 292                         if(i == 0) {
 293                                 bitset_copy(loop_attr->livethrough_unused, livethrough_block_unused);
 294                         } else {
 295                                 bitset_and(loop_attr->livethrough_unused, livethrough_block_unused);
 296                         }
 297                         break;
 298                 }
 299                 case k_ir_loop: {
 300                         bitset_t *livethrough_son_unused;
 301
 302                         livethrough_son_unused = construct_loop_livethrough_unused(env, elem.son);
 303                         if(i == 0) {
 304                                 bitset_copy(loop_attr->livethrough_unused, livethrough_son_unused);
 305                         } else {
 306                                 bitset_and(loop_attr->livethrough_unused, livethrough_son_unused);
 307                         }
 308                         break;
 309                 }
 310             default:
 311                         assert(0);
 312                         break;
 313                 }
 314     }
 315         DBG((dbg, DBG_LIVE, "Done with loop %d\n", loop->loop_nr));
 316
 317         // remove all unused livethroughs that are remembered for this loop from child loops and blocks
 318         for(i = 0; i < get_loop_n_elements(loop); ++i) {
 319                 const loop_element elem = get_loop_element(loop, i);
 320
 321                 if(*elem.kind == k_ir_loop) {
 322                         loop_attr_t *son_attr = get_loop_attr(env, elem.son);
 323                         bitset_andnot(son_attr->livethrough_unused, loop_attr->livethrough_unused);
 324
 325                         DBG((dbg, DBG_LIVE, "Livethroughs for loop %d:\n", loop->loop_nr));
 326                         show_nodebitset(env->irg, son_attr->livethrough_unused);
 327                 } else if(*elem.kind == k_ir_node) {
 328                         block_attr_t *block_attr = get_block_attr(env, elem.node);
 329                         bitset_andnot(block_attr->livethrough_unused, loop_attr->livethrough_unused);
 330
 331                         DBG((dbg, DBG_LIVE, "Livethroughs for block %+F\n", elem.node));
 332                         show_nodebitset(env->irg, block_attr->livethrough_unused);
 333                 } else {
 334                         assert(0);
 335                 }
 336         }
 337
 338         return loop_attr->livethrough_unused;
 339 }
 340
 341 /*---------------------------------------------------------------------------*/
 342
 343 static int reduce_register_pressure_in_block(morgan_env_t *env, const ir_node* block, int loop_unused_spills_possible) {
 344         ir_node *node;
 345         int max_pressure;
 346         int loop_unused_spills_needed;
 347         pset *live_nodes = pset_new_ptr_default();
 348
 349         be_liveness_end_of_block(env->cenv->lv, env->arch, env->cls, block, live_nodes);
 350         max_pressure = pset_count(live_nodes);
 351
 352         DBG((dbg, DBG_LIVE, "Reduce pressure to %d In Block %+F:\n", env->registers_available, block));
 353
 354         /**
 355          * Determine register pressure in block
 356          */
 357         sched_foreach_reverse(block, node) {
 358                 int pressure;
 359
 360                 if(is_Phi(node))
 361                         break;
 362
 363                 be_liveness_transfer(env->arch, env->cls, node, live_nodes);
 364                 pressure = pset_count(live_nodes);
 365                 if(pressure > max_pressure)
 366                         max_pressure = pressure;
 367         }
 368         del_pset(live_nodes);
 369
 370         DBG((dbg, DBG_PRESSURE, "\tMax Pressure in %+F: %d\n", block, max_pressure));
 371
 372         loop_unused_spills_needed = max_pressure - env->registers_available;
 373
 374         if(loop_unused_spills_needed < 0) {
 375                 loop_unused_spills_needed = 0;
 376         } else if(loop_unused_spills_needed > loop_unused_spills_possible) {
 377                 loop_unused_spills_needed = loop_unused_spills_possible;
 378         }
 379
 380         DBG((dbg, DBG_PRESSURE, "Unused spills for Block %+F needed: %d\n", block, loop_unused_spills_needed));
 381         return loop_unused_spills_needed;
 382 }
 383
 384 /**
 385  * Reduce register pressure in a loop
 386  *
 387  * @param unused_spills_possible        Number of spills from livethrough_unused variables possible in outer loops
 388  * @return                                                      Number of spills of livethrough_unused variables needed in outer loops
 389  */
 390 static int reduce_register_pressure_in_loop(morgan_env_t *env, const ir_loop *loop, int outer_spills_possible) {
 391         int i;
 392         loop_attr_t* loop_attr = get_loop_attr(env, loop);
 393         int spills_needed = 0;
 394         int spills_possible = outer_spills_possible + bitset_popcnt(loop_attr->livethrough_unused);
 395         int outer_spills_needed;
 396
 397         DBG((dbg, DBG_PRESSURE, "Reducing Pressure in loop %d\n", loop->loop_nr));
 398         for(i = 0; i < get_loop_n_elements(loop); ++i) {
 399                 loop_element elem = get_loop_element(loop, i);
 400                 switch (*elem.kind) {
 401                 case k_ir_node: {
 402                         int needed;
 403                         assert(is_Block(elem.node));
 404                         needed = reduce_register_pressure_in_block(env, elem.node, spills_possible);
 405                         assert(needed >= 0);
 406                         assert(needed <= spills_possible);
 407                         if(needed > spills_needed)
 408                                 spills_needed = needed;
 409                         break;
 410                 }
 411                 case k_ir_loop: {
 412                         int needed = reduce_register_pressure_in_loop(env, elem.son, spills_possible);
 413                         assert(needed >= 0);
 414                         assert(needed <= spills_possible);
 415                         if(needed > spills_needed)
 416                                 spills_needed = needed;
 417                         break;
 418                 }
 419             default:
 420                         assert(0);
 421                         break;
 422                 }
 423     }
 424
 425         /* calculate number of spills needed in outer loop and spill
 426          * unused livethrough nodes around this loop
 427          */
 428         if(spills_needed > outer_spills_possible) {
 429                 int spills_to_place;
 430                 outer_spills_needed = outer_spills_possible;
 431                 spills_needed -= outer_spills_possible;
 432
 433                 spills_to_place = spills_needed;
 434
 435                 bitset_foreach(loop_attr->livethrough_unused, i) {
 436                         loop_edge_t *edge;
 437                         ir_node *to_spill = get_idx_irn(env->irg, i);
 438
 439                         for(edge = set_first(loop_attr->out_edges); edge != NULL; edge = set_next(loop_attr->out_edges)) {
 440                                 DBG((dbg, DBG_PRESSURE, "Spilling node %+F around loop %d\n", to_spill, loop->loop_nr));
 441                                 be_add_reload_on_edge(env->senv, to_spill, edge->block, edge->pos);
 442                         }
 443
 444                         spills_to_place--;
 445                         if(spills_to_place <= 0) {
 446                                 break;
 447                         }
 448                 }
 449         } else {
 450                 outer_spills_needed = spills_needed;
 451         }
 452
 453         return outer_spills_needed;
 454 }
 455
 456 void be_spill_morgan(be_chordal_env_t *chordal_env) {
 457         morgan_env_t env;
 458
 459         FIRM_DBG_REGISTER(dbg, "ir.be.spillmorgan");
 460         //firm_dbg_set_mask(dbg, DBG_LOOPANA | DBG_PRESSURE);
 461
 462         env.cenv = chordal_env;
 463         env.arch = chordal_env->birg->main_env->arch_env;
 464         env.irg = chordal_env->irg;
 465         env.cls = chordal_env->cls;
 466         env.senv = be_new_spill_env(chordal_env);
 467         DEBUG_ONLY(be_set_spill_env_dbg_module(env.senv, dbg);)
 468
 469         obstack_init(&env.obst);
 470
 471         env.registers_available = env.cls->n_regs - be_put_ignore_regs(chordal_env->birg, env.cls, NULL);
 472
 473         env.loop_attr_set = new_set(loop_attr_cmp, 5);
 474         env.block_attr_set = new_set(block_attr_cmp, 20);
 475
 476         /*-- Part1: Analysis --*/
 477         be_liveness_recompute(chordal_env->lv);
 478
 479         /* construct control flow loop tree */
 480         construct_cf_backedges(chordal_env->irg);
 481
 482         /* construct loop out edges and livethrough_unused sets for loops and blocks */
 483         irg_block_walk_graph(chordal_env->irg, NULL, construct_loop_edges, &env);
 484         construct_loop_livethrough_unused(&env, get_irg_loop(env.irg));
 485
 486         /*-- Part2: Transformation --*/
 487
 488         /* spill unused livethrough values around loops and blocks where
 489          * the pressure is too high
 490          */
 491         reduce_register_pressure_in_loop(&env, get_irg_loop(env.irg), 0);
 492
 493         /* Insert real spill/reload nodes and fix usages */
 494         be_insert_spills_reloads(env.senv);
 495
 496         /* Verify the result */
 497         if (chordal_env->opts->vrfy_option == BE_CH_VRFY_WARN) {
 498                 be_verify_schedule(env.irg);
 499         } else if (chordal_env->opts->vrfy_option == BE_CH_VRFY_ASSERT) {
 500                 assert(be_verify_schedule(env.irg));
 501         }
 502
 503         if (chordal_env->opts->dump_flags & BE_CH_DUMP_SPILL)
 504                 be_dump(env.irg, "-spillmorgan", dump_ir_block_graph_sched);
 505
 506         /* cleanup */
 507         free_loop_edges(&env);
 508         del_set(env.loop_attr_set);
 509         del_set(env.block_attr_set);
 510
 511         /* fix the remaining places with too high register pressure with beladies algorithm */
 512         be_spill_belady_spill_env(chordal_env, env.senv);
 513
 514         be_delete_spill_env(env.senv);
 515         obstack_free(&env.obst, NULL);
 516 }