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 #include "bespillbelady.h"
  34 #include "beverify.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 *node, void *data) {
 143         // TODO what is this for?
 144
 145         return 0;
 146 }
 147
 148 //---------------------------------------------------------------------------
 149
 150 /**
 151  * Determine edges going out of a loop (= edges that go to a block that is not inside
 152  * the loop or one of its subloops)
 153  */
 154 static INLINE void construct_loop_out_edges(ir_node* block, void* e) {
 155         morgan_env_t *env = (morgan_env_t*) e;
 156         int n_cfgpreds = get_Block_n_cfgpreds(block);
 157         int i;
 158         ir_loop* loop = get_irn_loop(block);
 159
 160         for(i = 0; i < n_cfgpreds; ++i) {
 161                 ir_node* cfgpred = get_Block_cfgpred(block, i);
 162                 ir_node* cfgpred_block = get_nodes_block(cfgpred);
 163                 ir_loop* cfgpred_loop = get_irn_loop(cfgpred_block);
 164                 loop_attr_t *outedges = get_loop_attr(env, cfgpred_loop);
 165
 166                 if(cfgpred_loop != loop && get_loop_depth(cfgpred_loop) >= get_loop_depth(loop)) {
 167                         loop_out_edge_t edge;
 168                         edge.block = block;
 169                         edge.pos = i;
 170                         set_insert(outedges->out_edges, &edge, sizeof(edge), loop_out_edge_hash(&edge));
 171                 }
 172         }
 173 }
 174
 175 static void free_loop_out_edges(morgan_env_t *env) {
 176         loop_attr_t *l_attr;
 177
 178         for(l_attr = set_first(env->loop_attr_set); l_attr != NULL; l_attr = set_next(env->loop_attr_set)) {
 179                 del_set(l_attr->out_edges);
 180         }
 181 }
 182
 183 /**
 184  * Construct the livethrough unused information for a block
 185  */
 186 static bitset_t *construct_block_livethrough_unused(morgan_env_t* env, ir_node* block) {
 187         int i;
 188         int node_idx;
 189         ir_node *irn;
 190         block_attr_t *block_attr = get_block_attr(env, block);
 191
 192         /*
 193          * This is the first block in a sequence, all variables that are livethrough this block are potential
 194          * candidates for livethrough_unused
 195          */
 196         irn_live_t *li;
 197
 198         // copy all live-outs into the livethrough_unused set
 199         live_foreach(block, li) {
 200                 if(!live_is_in(li) || !live_is_out(li))
 201                         continue;
 202                 if(!arch_irn_consider_in_reg_alloc(env->arch, env->cls, li->irn))
 203                         continue;
 204
 205                 node_idx = get_irn_idx(li->irn);
 206                 bitset_set(block_attr->livethrough_unused, node_idx);
 207         }
 208
 209         /*
 210          * All values that are used within the block are not unused (and therefore not
 211          * livethrough_unused)
 212          */
 213         sched_foreach(block, irn) {
 214                 be_insn_t *insn = be_scan_insn(&env->insn_env, irn);
 215
 216                 for(i = insn->use_start; i < insn->n_ops; ++i) {
 217                         const be_operand_t *op = &insn->ops[i];
 218                         int idx = get_irn_idx(op->irn);
 219                         bitset_clear(block_attr->livethrough_unused, idx);
 220                 }
 221         }
 222
 223         return block_attr->livethrough_unused;
 224 }
 225
 226 /**
 227  * Debugging help, shows all nodes in a (node-)bitset
 228  */
 229 static void show_nodebitset(ir_graph* irg, 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 static bitset_t *construct_loop_livethrough_unused(morgan_env_t *env, ir_loop *loop) {
 239         int i;
 240         loop_attr_t* loop_attr = get_loop_attr(env, loop);
 241
 242         DBG((dbg, DBG_LIVE, "Processing Loop %d\n", loop->loop_nr));
 243         assert(get_loop_n_elements(loop) > 0);
 244         for(i = 0; i < get_loop_n_elements(loop); ++i) {
 245                 loop_element elem = get_loop_element(loop, i);
 246                 switch (*elem.kind) {
 247                 case k_ir_node: {
 248                         bitset_t *livethrough_block_unused;
 249                         assert(is_Block(elem.node));
 250                         livethrough_block_unused = construct_block_livethrough_unused(env, elem.node);
 251                         if(i == 0) {
 252                                 bitset_copy(loop_attr->livethrough_unused, livethrough_block_unused);
 253                         } else {
 254                                 bitset_and(loop_attr->livethrough_unused, livethrough_block_unused);
 255                         }
 256                         break;
 257                 }
 258                 case k_ir_loop: {
 259                         bitset_t *livethrough_son_unused;
 260
 261                         livethrough_son_unused = construct_loop_livethrough_unused(env, elem.son);
 262                         if(i == 0) {
 263                                 bitset_copy(loop_attr->livethrough_unused, livethrough_son_unused);
 264                         } else {
 265                                 bitset_and(loop_attr->livethrough_unused, livethrough_son_unused);
 266                         }
 267                         break;
 268                 }
 269             default:
 270                         assert(0);
 271                         break;
 272                 }
 273     }
 274
 275         // remove all unused livethroughs that are remembered for this loop from child loops and blocks
 276         for(i = 0; i < get_loop_n_elements(loop); ++i) {
 277                 const loop_element elem = get_loop_element(loop, i);
 278
 279                 if(*elem.kind == k_ir_loop) {
 280                         loop_attr_t *son_attr = get_loop_attr(env, elem.son);
 281                         bitset_andnot(son_attr->livethrough_unused, loop_attr->livethrough_unused);
 282
 283                         DBG((dbg, DBG_LIVE, "Livethroughs for loop %d:\n", loop->loop_nr));
 284                         show_nodebitset(env->irg, son_attr->livethrough_unused);
 285                 } else if(*elem.kind == k_ir_node) {
 286                         block_attr_t *block_attr = get_block_attr(env, elem.node);
 287                         bitset_andnot(block_attr->livethrough_unused, loop_attr->livethrough_unused);
 288
 289                         DBG((dbg, DBG_LIVE, "Livethroughs for block %+F\n", elem.node));
 290                         show_nodebitset(env->irg, block_attr->livethrough_unused);
 291                 } else {
 292                         assert(0);
 293                 }
 294         }
 295
 296         return loop_attr->livethrough_unused;
 297 }
 298
 299 //---------------------------------------------------------------------------
 300
 301 static int reduce_register_pressure_in_block(morgan_env_t *env, ir_node* block, int loop_unused_spills_possible) {
 302         int pressure;
 303         ir_node *irn;
 304         int max_pressure = 0;
 305         int spills_needed;
 306         int loop_unused_spills_needed;
 307         block_attr_t *block_attr = get_block_attr(env, block);
 308         int block_unused_spills_possible = bitset_popcnt(block_attr->livethrough_unused);
 309         int unused_spills_possible = loop_unused_spills_possible + block_unused_spills_possible;
 310         pset *live_nodes = pset_new_ptr_default();
 311
 312         be_liveness_end_of_block(env->arch, env->cls, block, live_nodes);
 313         pressure = pset_count(live_nodes);
 314
 315         DBG((dbg, DBG_LIVE, "Reduce pressure to %d In Block %+F:\n", env->registers_available, block));
 316
 317         /**
 318          * Walk over all irns in the schedule and check register pressure for each of them
 319          */
 320         sched_foreach_reverse(block, irn) {
 321                 // do we need more spills than possible with unused libethroughs?
 322                 int spills_needed = pressure - unused_spills_possible - env->registers_available;
 323                 if(spills_needed > 0) {
 324                         DBG((dbg, DBG_PRESSURE, "\tWARNING %d more spills needed at %+F\n", spills_needed, irn));
 325                         // TODO further spills needed
 326                         //assert(0);
 327                 }
 328                 if(pressure > max_pressure) {
 329                         max_pressure = pressure;
 330                 }
 331
 332                 /* Register pressure is only important until we reach the first phi (the rest of the block
 333                  * will only be phis.)
 334                  */
 335                 if(is_Phi(irn))
 336                         break;
 337
 338                 // update pressure
 339                 {
 340                         int pressure_old = pressure;
 341                         be_liveness_transfer(env->arch, env->cls, irn, live_nodes);
 342                         pressure = pset_count(live_nodes);
 343                         DBG((dbg, DBG_PRESSURE, "\tPressure at %+F - before: %d after: %d\n", irn, pressure_old, pressure));
 344                 }
 345         }
 346
 347         /*
 348          * Calculate number of spills from loop_unused_spills_possible that we want to use,
 349          * and spill unused livethroughs from the block if we still don't have enough registers
 350          */
 351         spills_needed = max_pressure - env->registers_available;
 352         if(spills_needed < 0) {
 353                 loop_unused_spills_needed = 0;
 354         } else if(spills_needed > loop_unused_spills_possible) {
 355                 int i, spills;
 356                 int block_unused_spills_needed;
 357
 358                 loop_unused_spills_needed = loop_unused_spills_possible;
 359                 block_unused_spills_needed = spills_needed - loop_unused_spills_possible;
 360                 if(block_unused_spills_needed > block_unused_spills_possible) {
 361                         block_unused_spills_needed = block_unused_spills_possible;
 362                 }
 363
 364                 spills = 0;
 365                 /*
 366                  * Spill/Reload unused livethroughs from the block
 367                  */
 368                 bitset_foreach(block_attr->livethrough_unused, i) {
 369                         ir_node *to_spill;
 370                         const ir_edge_t *edge;
 371
 372                         if(spills >= block_unused_spills_needed)
 373                                 break;
 374
 375                         to_spill = get_idx_irn(env->irg, i);
 376                         foreach_block_succ(block, edge) {
 377                                 DBG((dbg, DBG_PRESSURE, "Spilling node %+F around block %+F\n", to_spill, block));
 378                                 be_add_reload_on_edge(env->senv, to_spill, edge->src, edge->pos);
 379                         }
 380                 }
 381         } else {
 382                 loop_unused_spills_needed = spills_needed;
 383         }
 384
 385         del_pset(live_nodes);
 386
 387         DBG((dbg, DBG_PRESSURE, "Unused spills for Block %+F needed: %d\n", block, loop_unused_spills_needed));
 388         return loop_unused_spills_needed;
 389 }
 390
 391 /**
 392  * Reduce register pressure in a loop
 393  *
 394  * @param unused_spills_possible        Number of spills from livethrough_unused variables possible in outer loops
 395  * @return                                                      Number of spills of livethrough_unused variables needed in outer loops
 396  */
 397 static int reduce_register_pressure_in_loop(morgan_env_t *env, ir_loop *loop, int outer_spills_possible) {
 398         int i;
 399         loop_attr_t* loop_attr = get_loop_attr(env, loop);
 400         int spills_needed = 0;
 401         int spills_possible = outer_spills_possible + bitset_popcnt(loop_attr->livethrough_unused);
 402         int outer_spills_needed;
 403
 404         DBG((dbg, DBG_PRESSURE, "Reducing Pressure in loop %d\n", loop->loop_nr));
 405         for(i = 0; i < get_loop_n_elements(loop); ++i) {
 406                 loop_element elem = get_loop_element(loop, i);
 407                 switch (*elem.kind) {
 408                 case k_ir_node: {
 409                         int needed;
 410                         assert(is_Block(elem.node));
 411                         needed = reduce_register_pressure_in_block(env, elem.node, spills_possible);
 412                         assert(needed <= spills_possible);
 413                         if(needed > spills_needed)
 414                                 spills_needed = needed;
 415                         break;
 416                 }
 417                 case k_ir_loop: {
 418                         int needed = reduce_register_pressure_in_loop(env, elem.son, spills_possible);
 419                         assert(needed <= spills_possible);
 420                         if(needed > spills_needed)
 421                                 spills_needed = needed;
 422                         break;
 423                 }
 424             default:
 425                         assert(0);
 426                         break;
 427                 }
 428     }
 429
 430         // calculate number of spills needed in outer loop and spill
 431         // unused livethrough nodes around this loop
 432         if(spills_needed > outer_spills_possible) {
 433                 outer_spills_needed = outer_spills_possible;
 434                 spills_needed -= outer_spills_possible;
 435
 436                 bitset_foreach(loop_attr->livethrough_unused, i) {
 437                         loop_out_edge_t *edge;
 438                         ir_node *to_spill = get_idx_irn(env->irg, i);
 439
 440                         for(edge = set_first(loop_attr->out_edges); edge != NULL; edge = set_next(loop_attr->out_edges)) {
 441                                 be_add_reload_on_edge(env->senv, to_spill, edge->block, edge->pos);
 442                         }
 443                 }
 444         } else {
 445                 outer_spills_needed = spills_needed;
 446         }
 447
 448         return outer_spills_needed;
 449 }
 450
 451 static void *init_phase_data(phase_t *phase, ir_node *irn, void *old) {
 452         return old;
 453 }
 454
 455 void be_spill_morgan(const be_chordal_env_t *chordal_env) {
 456         morgan_env_t env;
 457
 458         FIRM_DBG_REGISTER(dbg, "ir.be.spillmorgan");
 459
 460         env.arch = chordal_env->birg->main_env->arch_env;
 461         env.irg = chordal_env->irg;
 462         env.cls = chordal_env->cls;
 463         env.senv = be_new_spill_env(chordal_env, is_mem_phi, NULL);
 464         DEBUG_ONLY(be_set_spill_env_dbg_module(env.senv, dbg);)
 465         env.uses = be_begin_uses(env.irg, env.arch, env.cls);
 466
 467         phase_init(&env.phase, "spillmorgan", env.irg, PHASE_DEFAULT_GROWTH, init_phase_data);
 468
 469         env.registers_available = arch_count_non_ignore_regs(env.arch, env.cls);
 470
 471         be_insn_env_init(&env.insn_env, chordal_env->birg, chordal_env->cls, &env.phase.obst);
 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(env.irg);
 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, construct_loop_out_edges, NULL, &env);
 484         construct_loop_livethrough_unused(&env, get_irg_loop(env.irg));
 485
 486         /*-- Part2: Transformation --*/
 487
 488         // reduce register pressure to number of available registers
 489         reduce_register_pressure_in_loop(&env, get_irg_loop(env.irg), 0);
 490
 491         be_insert_spills_reloads(env.senv);
 492         if (chordal_env->opts->vrfy_option == BE_CH_VRFY_WARN)
 493                 be_verify_schedule(env.irg);
 494         else if (chordal_env->opts->vrfy_option == BE_CH_VRFY_ASSERT)
 495                 assert(be_verify_schedule(env.irg));
 496
 497         // cleanup
 498         be_end_uses(env.uses);
 499         be_dump(env.irg, "-spillmorgan", dump_ir_block_graph_sched);
 500         free_loop_out_edges(&env);
 501         del_set(env.loop_attr_set);
 502         del_set(env.block_attr_set);
 503
 504         // fix the remaining places with too high register pressure with beladies algorithm
 505         be_remove_dead_nodes_from_schedule(env.irg);
 506         be_liveness(env.irg);
 507         be_spill_belady_spill_env(chordal_env, env.senv);
 508
 509         be_delete_spill_env(env.senv);
 510         phase_free(&env.phase);
 511 }