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 "beinsn_t.h"
  20 #include "irgwalk.h"
  21 #include "besched.h"
  22 #include "beutil.h"
  23 #include "beuses.h"
  24 #include "interval_analysis.h"
  25 #include "irloop.h"
  26 #include "irloop_t.h"
  27 #include "irgraph.h"
  28 #include "irgraph_t.h"
  29 #include "irphase.h"
  30 #include "irphase_t.h"
  31 #include "irprintf.h"
  32
  33 // remove me later
  34 #include "bespillbelady.h"
  35
  36 #define DBG_LIVE                1
  37 #define DBG_PRESSURE    2
  38 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  39
  40 typedef struct _morgan_env_t {
  41         const arch_env_t *arch;
  42         const arch_register_class_t *cls;
  43         ir_graph *irg;
  44         phase_t phase;
  45         // maximum safe register pressure
  46         int registers_available;
  47
  48         be_insn_env_t insn_env;
  49         spill_env_t *senv;
  50         be_uses_t *uses;
  51
  52         set *loop_attr_set;
  53         set *block_attr_set;
  54 } morgan_env_t;
  55
  56 typedef struct _loop_out_edge_t {
  57         ir_node *block;
  58         int pos;
  59 } loop_out_edge_t;
  60
  61 typedef struct _loop_attr_t {
  62         ir_loop *loop;
  63         set *out_edges;
  64         /// The set of all values that live through the loop and are not used
  65         bitset_t *livethrough_unused;
  66 } loop_attr_t;
  67
  68 typedef struct _block_attr_t {
  69         ir_node *block;
  70         bitset_t *livethrough_unused;
  71 } block_attr_t;
  72
  73 //---------------------------------------------------------------------------
  74
  75 int loop_out_edge_cmp(const void* p1, const void* p2, size_t s) {
  76         loop_out_edge_t *e1 = (loop_out_edge_t*) p1;
  77         loop_out_edge_t *e2 = (loop_out_edge_t*) p2;
  78
  79         return e1->block != e2->block || e1->pos != e2->pos;
  80 }
  81
  82 int loop_attr_cmp(const void *e1, const void *e2, size_t s) {
  83         loop_attr_t *la1 = (loop_attr_t*) e1;
  84         loop_attr_t *la2 = (loop_attr_t*) e2;
  85
  86         return la1->loop != la2->loop;
  87 }
  88
  89 int block_attr_cmp(const void *e1, const void *e2, size_t s) {
  90         block_attr_t *b1 = (block_attr_t*) e1;
  91         block_attr_t *b2 = (block_attr_t*) e2;
  92
  93         return b1->block != b2->block;
  94 }
  95
  96 static INLINE int loop_attr_hash(const loop_attr_t *a) {
  97         return HASH_PTR(a->loop);
  98 }
  99
 100 static INLINE int block_attr_hash(const block_attr_t *b) {
 101         return HASH_PTR(b->block);
 102 }
 103
 104 static INLINE int loop_out_edge_hash(const loop_out_edge_t *e) {
 105         return HASH_PTR(e->block) ^ (e->pos * 31);
 106 }
 107
 108 static INLINE loop_attr_t *get_loop_attr(morgan_env_t *env, ir_loop *loop) {
 109         loop_attr_t l_attr, *res;
 110         int hash;
 111         l_attr.loop = loop;
 112
 113         hash = loop_attr_hash(&l_attr);
 114         res = set_find(env->loop_attr_set, &l_attr, sizeof(l_attr), hash);
 115
 116         // create new loop_attr if none exists yet
 117         if (!res) {
 118                 l_attr.out_edges = new_set(loop_out_edge_cmp, 1);
 119                 l_attr.livethrough_unused = bitset_obstack_alloc(&env->phase.obst, get_irg_last_idx(env->irg));
 120                 res = set_insert(env->loop_attr_set, &l_attr, sizeof(l_attr), hash);
 121         }
 122
 123         return res;
 124 }
 125
 126 static INLINE block_attr_t *get_block_attr(morgan_env_t *env, ir_node *block) {
 127         block_attr_t b_attr, *res;
 128         int hash;
 129         b_attr.block = block;
 130
 131         hash = block_attr_hash(&b_attr);
 132         res = set_find(env->block_attr_set, &b_attr, sizeof(b_attr), hash);
 133
 134         if(!res) {
 135                 b_attr.livethrough_unused = bitset_obstack_alloc(&env->phase.obst, get_irg_last_idx(env->irg));
 136                 res = set_insert(env->block_attr_set, &b_attr, sizeof(b_attr), hash);
 137         }
 138
 139         return res;
 140 }
 141
 142 static int is_mem_phi(const ir_node *irn, void *data) {
 143         // TODO what is this for?
 144         return 0;
 145 }
 146
 147 //---------------------------------------------------------------------------
 148
 149 /**
 150  * Determine edges going out of a loop (= edges that go to a block that is not inside
 151  * the loop or one of its subloops)
 152  */
 153 static INLINE void construct_loop_out_edges(ir_node* block, void* e) {
 154         morgan_env_t *env = (morgan_env_t*) e;
 155         int n_cfgpreds = get_Block_n_cfgpreds(block);
 156         int i;
 157         ir_loop* loop = get_irn_loop(block);
 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 && get_loop_depth(cfgpred_loop) >= get_loop_depth(loop)) {
 166                         loop_out_edge_t edge;
 167                         edge.block = block;
 168                         edge.pos = i;
 169                         set_insert(outedges->out_edges, &edge, sizeof(edge), loop_out_edge_hash(&edge));
 170                 }
 171         }
 172 }
 173
 174 /**
 175  * Construct the livethrough unused information for a block
 176  */
 177 static bitset_t *construct_block_livethrough_unused(morgan_env_t* env, ir_node* block) {
 178         int i;
 179         int node_idx;
 180         ir_node *irn;
 181         block_attr_t *block_attr = get_block_attr(env, block);
 182
 183         /*
 184          * This is the first block in a sequence, all variables that are livethrough this block are potential
 185          * candidates for livethrough_unused
 186          */
 187         irn_live_t *li;
 188
 189         // copy all live-outs into the livethrough_unused set
 190         live_foreach(block, li) {
 191                 if(!live_is_in(li) || !live_is_out(li))
 192                         continue;
 193                 if(!arch_irn_consider_in_reg_alloc(env->arch, env->cls, li->irn))
 194                         continue;
 195
 196                 node_idx = get_irn_idx(li->irn);
 197                 bitset_set(block_attr->livethrough_unused, node_idx);
 198         }
 199
 200         /*
 201          * All values that are used within the block are not unused (and therefore not
 202          * livethrough_unused)
 203          */
 204         sched_foreach(block, irn) {
 205                 be_insn_t *insn = be_scan_insn(&env->insn_env, irn);
 206
 207                 for(i = insn->use_start; i < insn->n_ops; ++i) {
 208                         const be_operand_t *op = &insn->ops[i];
 209                         int idx = get_irn_idx(op->irn);
 210                         bitset_clear(block_attr->livethrough_unused, idx);
 211                 }
 212         }
 213
 214         return block_attr->livethrough_unused;
 215 }
 216
 217 /**
 218  * Debugging help, shows all nodes in a (node-)bitset
 219  */
 220 static void show_nodebitset(ir_graph* irg, bitset_t* bitset) {
 221         int i;
 222
 223         bitset_foreach(bitset, i) {
 224                 ir_node* node = get_idx_irn(irg, i);
 225                 DBG((dbg, DBG_LIVE, "\t%+F\n", node));
 226         }
 227 }
 228
 229 static bitset_t *construct_loop_livethrough_unused(morgan_env_t *env, ir_loop *loop) {
 230         int i;
 231         loop_attr_t* loop_attr = get_loop_attr(env, loop);
 232
 233         DBG((dbg, DBG_LIVE, "Processing Loop %d\n", loop->loop_nr));
 234         assert(get_loop_n_elements(loop) > 0);
 235         for(i = 0; i < get_loop_n_elements(loop); ++i) {
 236                 loop_element elem = get_loop_element(loop, i);
 237                 switch (*elem.kind) {
 238                 case k_ir_node: {
 239                         bitset_t *livethrough_block_unused;
 240                         assert(is_Block(elem.node));
 241                         livethrough_block_unused = construct_block_livethrough_unused(env, elem.node);
 242                         if(i == 0) {
 243                                 bitset_copy(loop_attr->livethrough_unused, livethrough_block_unused);
 244                         } else {
 245                                 bitset_and(loop_attr->livethrough_unused, livethrough_block_unused);
 246                         }
 247                         break;
 248                 }
 249                 case k_ir_loop: {
 250                         bitset_t *livethrough_son_unused;
 251
 252                         livethrough_son_unused = construct_loop_livethrough_unused(env, elem.son);
 253                         if(i == 0) {
 254                                 bitset_copy(loop_attr->livethrough_unused, livethrough_son_unused);
 255                         } else {
 256                                 bitset_and(loop_attr->livethrough_unused, livethrough_son_unused);
 257                         }
 258                         break;
 259                 }
 260             default:
 261                         assert(0);
 262                         break;
 263                 }
 264     }
 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 - unused_spills_possible - env->registers_available;
 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                 {
 331                         int pressure_old = pressure;
 332                         be_liveness_transfer(env->arch, env->cls, irn, live_nodes);
 333                         pressure = pset_count(live_nodes);
 334                         DBG((dbg, DBG_PRESSURE, "\tPressure at %+F - before: %d after: %d\n", irn, pressure_old, pressure));
 335                 }
 336         }
 337
 338         /*
 339          * Calculate number of spills from loop_unused_spills_possible that we want to use,
 340          * and spill unused livethroughs from the block if we still don't have enough registers
 341          */
 342         spills_needed = max_pressure - env->registers_available;
 343         if(spills_needed < 0) {
 344                 loop_unused_spills_needed = 0;
 345         } else if(spills_needed > loop_unused_spills_possible) {
 346                 int i, spills;
 347                 int block_unused_spills_needed;
 348
 349                 loop_unused_spills_needed = loop_unused_spills_possible;
 350                 block_unused_spills_needed = spills_needed - loop_unused_spills_possible;
 351                 if(block_unused_spills_needed > block_unused_spills_possible) {
 352                         block_unused_spills_needed = block_unused_spills_possible;
 353                 }
 354
 355                 spills = 0;
 356                 /*
 357                  * Spill/Reload unused livethroughs from the block
 358                  */
 359                 bitset_foreach(block_attr->livethrough_unused, i) {
 360                         ir_node *to_spill;
 361                         const ir_edge_t *edge;
 362
 363                         if(spills >= block_unused_spills_needed)
 364                                 break;
 365
 366                         to_spill = get_idx_irn(env->irg, i);
 367                         foreach_block_succ(block, edge) {
 368                                 DBG((dbg, DBG_PRESSURE, "Spilling node %+F around block %+F\n", to_spill, block));
 369                                 be_add_reload_on_edge(env->senv, to_spill, edge->src, edge->pos);
 370                         }
 371                 }
 372         } else {
 373                 loop_unused_spills_needed = spills_needed;
 374         }
 375
 376         del_pset(live_nodes);
 377
 378         DBG((dbg, DBG_PRESSURE, "Unused spills for Block %+F needed: %d\n", block, loop_unused_spills_needed));
 379         return loop_unused_spills_needed;
 380 }
 381
 382 /**
 383  * Reduce register pressure in a loop
 384  *
 385  * @param unused_spills_possible        Number of spills from livethrough_unused variables possible in outer loops
 386  * @return                                                      Number of spills of livethrough_unused variables needed in outer loops
 387  */
 388 static int reduce_register_pressure_in_loop(morgan_env_t *env, ir_loop *loop, int outer_spills_possible) {
 389         int i;
 390         loop_attr_t* loop_attr = get_loop_attr(env, loop);
 391         int spills_needed = 0;
 392         int spills_possible = outer_spills_possible + bitset_popcnt(loop_attr->livethrough_unused);
 393         int outer_spills_needed;
 394
 395         DBG((dbg, DBG_PRESSURE, "Reducing Pressure in loop %d\n", loop->loop_nr));
 396         for(i = 0; i < get_loop_n_elements(loop); ++i) {
 397                 loop_element elem = get_loop_element(loop, i);
 398                 switch (*elem.kind) {
 399                 case k_ir_node: {
 400                         int needed;
 401                         assert(is_Block(elem.node));
 402                         needed = reduce_register_pressure_in_block(env, elem.node, spills_possible);
 403                         assert(needed <= spills_possible);
 404                         if(needed > spills_needed)
 405                                 spills_needed = needed;
 406                         break;
 407                 }
 408                 case k_ir_loop: {
 409                         int needed = reduce_register_pressure_in_loop(env, elem.son, spills_possible);
 410                         assert(needed <= spills_possible);
 411                         if(needed > spills_needed)
 412                                 spills_needed = needed;
 413                         break;
 414                 }
 415             default:
 416                         assert(0);
 417                         break;
 418                 }
 419     }
 420
 421         // calculate number of spills needed in outer loop and spill
 422         // unused livethrough nodes around this loop
 423         if(spills_needed > outer_spills_possible) {
 424                 outer_spills_needed = outer_spills_possible;
 425                 spills_needed -= outer_spills_possible;
 426
 427                 bitset_foreach(loop_attr->livethrough_unused, i) {
 428                         loop_out_edge_t *edge;
 429                         ir_node *to_spill = get_idx_irn(env->irg, i);
 430
 431                         for(edge = set_first(loop_attr->out_edges); edge != NULL; edge = set_next(loop_attr->out_edges)) {
 432                                 be_add_reload_on_edge(env->senv, to_spill, edge->block, edge->pos);
 433                         }
 434                 }
 435         } else {
 436                 outer_spills_needed = spills_needed;
 437         }
 438
 439         return outer_spills_needed;
 440 }
 441
 442 static void *init_phase_data(phase_t *phase, ir_node *irn, void *old) {
 443         return old;
 444 }
 445
 446 typedef struct _liveness_dump_env_t {
 447         const be_chordal_env_t *chordal_env;
 448         FILE *f;
 449 } liveness_dump_env_t;
 450
 451 /**
 452  * Pre-walker: dump liveness data to a file
 453  */
 454 static void dump_liveness_walker(ir_node *bl, void *data)
 455 {
 456         liveness_dump_env_t *env = (liveness_dump_env_t*) data;
 457         FILE *f = env->f;
 458         const irn_live_t *li;
 459         ir_node* irn;
 460         int in = 0, end = 0, out = 0;
 461         int max_pressure = 0;
 462         pset *live_nodes;
 463
 464         // collect some statistics
 465         live_foreach(bl, li) {
 466                 const ir_node* irn = li->irn;
 467                 if(!arch_irn_consider_in_reg_alloc(env->chordal_env->birg->main_env->arch_env, env->chordal_env->cls, irn))
 468                         continue;
 469
 470                 if(live_is_in(li))
 471                         in++;
 472                 if(live_is_end(li))
 473                         end++;
 474                 if(live_is_out(li))
 475                         out++;
 476         }
 477
 478         // collect register pressure info
 479         live_nodes = pset_new_ptr_default();
 480         be_liveness_end_of_block(env->chordal_env->birg->main_env->arch_env, env->chordal_env->cls, bl, live_nodes);
 481         max_pressure = pset_count(live_nodes);
 482         sched_foreach_reverse(bl, irn) {
 483                 int pressure;
 484
 485                 if(is_Phi(irn))
 486                         break;
 487
 488                 be_liveness_transfer(env->chordal_env->birg->main_env->arch_env, env->chordal_env->cls, irn, live_nodes);
 489                 pressure = pset_count(live_nodes);
 490                 if(pressure > max_pressure)
 491                         max_pressure = pressure;
 492         }
 493         del_pset(live_nodes);
 494
 495         ir_fprintf(f, "%+20F (%d in) (%d end) (%d out) (max_pressure %d)\n", bl, in, end, out, max_pressure);
 496         live_foreach(bl, li) {
 497                 const ir_node* irn = li->irn;
 498                 if(!arch_irn_consider_in_reg_alloc(env->chordal_env->birg->main_env->arch_env, env->chordal_env->cls, irn))
 499                         continue;
 500
 501                 ir_fprintf(f, "\t%+30F %4s %4s %4s\n",
 502                         irn,
 503                         live_is_in(li) ? "in" : "",
 504                         live_is_end(li) ? "end" : "",
 505                         live_is_out(li) ? "out" : "");
 506         }
 507 }
 508
 509 static void dump_liveness_info(const be_chordal_env_t *chordal_env, const char* name) {
 510         char buf[128];
 511         liveness_dump_env_t env;
 512
 513         env.chordal_env = chordal_env;
 514         ir_snprintf(buf, sizeof(buf), "%F_%s_%s-live.txt", chordal_env->irg, chordal_env->cls->name, name);
 515         env.f = fopen(buf, "wt");
 516         if(env.f == NULL)
 517                 return;
 518
 519         irg_block_walk_graph(chordal_env->irg, dump_liveness_walker, NULL, &env);
 520         fclose(env.f);
 521 }
 522
 523
 524 void be_spill_morgan(const be_chordal_env_t *chordal_env) {
 525         morgan_env_t env;
 526
 527         FIRM_DBG_REGISTER(dbg, "ir.be.spillmorgan");
 528         //firm_dbg_set_mask(dbg, DBG_LIVE | DBG_PRESSURE);
 529
 530         env.arch = chordal_env->birg->main_env->arch_env;
 531         env.irg = chordal_env->irg;
 532         env.cls = chordal_env->cls;
 533         env.senv = be_new_spill_env(chordal_env, is_mem_phi, NULL);
 534         DEBUG_ONLY(be_set_spill_env_dbg_module(env.senv, dbg);)
 535         env.uses = be_begin_uses(env.irg, env.arch, env.cls);
 536
 537         phase_init(&env.phase, "spillmorgan", env.irg, PHASE_DEFAULT_GROWTH, init_phase_data);
 538
 539         env.registers_available = arch_count_non_ignore_regs(env.arch, env.cls);
 540
 541         be_insn_env_init(&env.insn_env, chordal_env->birg, chordal_env->cls, &env.phase.obst);
 542
 543         env.loop_attr_set = new_set(loop_attr_cmp, 5);
 544         env.block_attr_set = new_set(block_attr_cmp, 20);
 545
 546
 547         /*-- Part1: Analysis --*/
 548         be_liveness(env.irg);
 549
 550         // construct control flow loop tree
 551         construct_cf_backedges(chordal_env->irg);
 552
 553         // construct loop out edges and livethrough_unused sets for loops and blocks
 554         irg_block_walk_graph(chordal_env->irg, construct_loop_out_edges, NULL, &env);
 555         construct_loop_livethrough_unused(&env, get_irg_loop(env.irg));
 556
 557         /*-- Part2: Transformation --*/
 558
 559         // reduce register pressure to number of available registers
 560         reduce_register_pressure_in_loop(&env, get_irg_loop(env.irg), 0);
 561
 562         be_insert_spills_reloads(env.senv, NULL);
 563
 564         // cleanup
 565         be_end_uses(env.uses);
 566         be_dump(env.irg, "-spillmorgan", dump_ir_block_graph_sched);
 567         del_set(env.loop_attr_set);
 568         del_set(env.block_attr_set);
 569
 570         be_liveness(env.irg);
 571         dump_liveness_info(chordal_env, "spillmorgan");
 572
 573         // fix the remaining places with too high register pressure with beladies algorithm
 574         be_spill_belady_spill_env(chordal_env, env.senv);
 575
 576         be_liveness(env.irg);
 577         dump_liveness_info(chordal_env, "spillcomplete");
 578
 579         be_delete_spill_env(env.senv);
 580         phase_free(&env.phase);
 581 }