nsz Git - libfirm/blob - ir/opt/ifconv.c

   1 /*
   2  * Copyright (C) 1995-2007 University of Karlsruhe.  All right reserved.
   3  *
   4  * This file is part of libFirm.
   5  *
   6  * This file may be distributed and/or modified under the terms of the
   7  * GNU General Public License version 2 as published by the Free Software
   8  * Foundation and appearing in the file LICENSE.GPL included in the
   9  * packaging of this file.
  10  *
  11  * Licensees holding valid libFirm Professional Edition licenses may use
  12  * this file in accordance with the libFirm Commercial License.
  13  * Agreement provided with the Software.
  14  *
  15  * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
  16  * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  17  * PURPOSE.
  18  */
  19
  20 /*
  21  * @file    ir/opt/ifconv.c
  22  * @brief   If conversion
  23  * @author  Christoph Mallon
  24  * @version $Id$
  25  */
  26
  27 #ifdef HAVE_CONFIG_H
  28 #include "config.h"
  29 #endif
  30
  31 #include <assert.h>
  32
  33 #include "obst.h"
  34 #include "irnode_t.h"
  35 #include "cdep.h"
  36 #include "ircons.h"
  37 #include "ifconv.h"
  38 #include "irdom.h"
  39 #include "irgmod.h"
  40 #include "irgopt.h"
  41 #include "irgwalk.h"
  42 #include "irtools.h"
  43 #include "return.h"
  44 #include "array.h"
  45 #include "xmalloc.h"
  46
  47 // debug
  48 #include "irdump.h"
  49 #include "debug.h"
  50
  51 DEBUG_ONLY(firm_dbg_module_t *dbg);
  52
  53 /** allow every Psi to be created. */
  54 static int default_allow_ifconv(ir_node *sel, ir_node* phi_list, int i, int j)
  55 {
  56         return 1;
  57 }
  58
  59 /**
  60  * Default options.
  61  */
  62 static const opt_if_conv_info_t default_info = {
  63         0,    /* doesn't matter for Psi */
  64         default_allow_ifconv
  65 };
  66
  67 /**
  68  * Additional block info.
  69  */
  70 typedef struct block_info {
  71         ir_node *phi;   /**< head of the Phi list */
  72         int has_pinned; /**< set if the block contains instructions that cannot be moved */
  73 } block_info;
  74
  75 #define get_block_blockinfo(block) ((block_info *)get_irn_link(block))
  76
  77 /**
  78  * Returns non-zero if a Block can be emptied.
  79  */
  80 static int can_empty_block(ir_node *block)
  81 {
  82         return !get_block_blockinfo(block)->has_pinned;
  83 }
  84
  85
  86 static ir_node* walk_to_projx(ir_node* start, const ir_node* dependency)
  87 {
  88         int arity;
  89         int i;
  90
  91         /* No need to find the conditional block if this block cannot be emptied and
  92          * therefore not moved
  93          */
  94         if (!can_empty_block(start)) return NULL;
  95
  96         arity = get_irn_arity(start);
  97         for (i = 0; i < arity; ++i) {
  98                 ir_node* pred = get_irn_n(start, i);
  99                 ir_node* pred_block = get_nodes_block(pred);
 100
 101                 if (pred_block == dependency) {
 102                         if (is_Proj(pred)) {
 103                                 assert(get_irn_mode(pred) == mode_X);
 104                                 return pred;
 105                         }
 106                         return NULL;
 107                 }
 108
 109                 if (is_Proj(pred)) {
 110                         assert(get_irn_mode(pred) == mode_X);
 111                         return NULL;
 112                 }
 113
 114                 if (is_cdep_on(pred_block, dependency)) {
 115                         return walk_to_projx(pred_block, dependency);
 116                 }
 117         }
 118         return NULL;
 119 }
 120
 121
 122 /**
 123  * Copies the DAG starting at node to the ith predecessor block of src_block
 124  * -if the node isn't in the src_block, this is a nop and the node is returned
 125  * -if the node is a phi in the src_block, the ith predecessor of the phi is
 126  *   returned
 127  * otherwise returns the copy of the passed node
 128  */
 129 static ir_node* copy_to(ir_node* node, ir_node* src_block, int i)
 130 {
 131         ir_node* dst_block;
 132         ir_node* copy;
 133         int arity;
 134         int j;
 135
 136         if (get_nodes_block(node) != src_block) return node;
 137         if (get_irn_op(node) == op_Phi) return get_irn_n(node, i);
 138
 139         copy = exact_copy(node);
 140         dst_block = get_nodes_block(get_irn_n(src_block, i));
 141         set_nodes_block(copy, dst_block);
 142
 143         DB((dbg, LEVEL_1, "Copying node %+F to block %+F, copy is %+F\n",
 144                 node, dst_block, copy));
 145
 146         arity = get_irn_arity(node);
 147         for (j = 0; j < arity; ++j) {
 148                 set_irn_n(copy, j, copy_to(get_irn_n(node, j), src_block, i));
 149                 DB((dbg, LEVEL_2, "-- pred %d is %+F\n", j, get_irn_n(copy, j)));
 150         }
 151         return copy;
 152 }
 153
 154
 155 /**
 156  * Remove predecessors i and j from node and add predecessor new_pred
 157  */
 158 static void rewire(ir_node* node, int i, int j, ir_node* new_pred)
 159 {
 160         int arity = get_irn_arity(node);
 161         ir_node **ins;
 162         int k;
 163         int l;
 164
 165         NEW_ARR_A(ir_node *, ins, arity - 1);
 166
 167         l = 0;
 168         for (k = 0; k < i; ++k) ins[l++] = get_irn_n(node, k);
 169         for (++k; k < j; ++k) ins[l++] = get_irn_n(node, k);
 170         for (++k; k < arity; ++k) ins[l++] = get_irn_n(node, k);
 171         ins[l++] = new_pred;
 172         assert(l == arity - 1);
 173         set_irn_in(node, l, ins);
 174 }
 175
 176
 177 /**
 178  * Remove the jth predecessors from the ith predecessor of block and add it to block
 179  */
 180 static void split_block(ir_node* block, int i, int j)
 181 {
 182         ir_node* pred_block = get_nodes_block(get_irn_n(block, i));
 183         int arity = get_irn_arity(block);
 184         int new_pred_arity;
 185         ir_node* phi;
 186         ir_node **ins;
 187         ir_node **pred_ins;
 188         int k;
 189
 190         DB((dbg, LEVEL_1, "Splitting predecessor %d of predecessor %d of %+F\n", j, i, block));
 191
 192         NEW_ARR_A(ir_node*, ins, arity + 1);
 193
 194         for (phi = get_block_blockinfo(block)->phi; phi != NULL; phi = get_irn_link(phi)) {
 195                 ir_node* copy = copy_to(get_irn_n(phi, i), pred_block, j);
 196
 197                 for (k = 0; k < i; ++k) ins[k] = get_irn_n(phi, k);
 198                 ins[k++] = copy;
 199                 for (; k < arity; ++k) ins[k] = get_irn_n(phi, k);
 200                 ins[k] = get_irn_n(phi, i);
 201                 assert(k == arity);
 202                 set_irn_in(phi, arity + 1, ins);
 203         }
 204
 205         for (k = 0; k < i; ++k) ins[k] = get_irn_n(block, k);
 206         ins[k++] = get_irn_n(pred_block, j);
 207         for (; k < arity; ++k) ins[k] = get_irn_n(block, k);
 208         ins[k] = get_irn_n(block, i);
 209         assert(k == arity);
 210         set_irn_in(block, arity + 1, ins);
 211
 212         new_pred_arity = get_irn_arity(pred_block) - 1;
 213         NEW_ARR_A(ir_node*, pred_ins, new_pred_arity);
 214
 215         for (phi = get_block_blockinfo(pred_block)->phi; phi != NULL; phi = get_irn_link(phi)) {
 216                 for (k = 0; k < j; ++k) pred_ins[k] = get_irn_n(phi, k);
 217                 for (; k < new_pred_arity; ++k) pred_ins[k] = get_irn_n(phi, k + 1);
 218                 assert(k == new_pred_arity);
 219                 if (new_pred_arity > 1) {
 220                         set_irn_in(phi, new_pred_arity, pred_ins);
 221                 } else {
 222                         exchange(phi, pred_ins[0]);
 223                 }
 224         }
 225
 226         for (k = 0; k < j; ++k) pred_ins[k] = get_irn_n(pred_block, k);
 227         for (; k < new_pred_arity; ++k) pred_ins[k] = get_irn_n(pred_block, k + 1);
 228         assert(k == new_pred_arity);
 229         if (new_pred_arity > 1) {
 230                 set_irn_in(pred_block, new_pred_arity, pred_ins);
 231         } else {
 232                 exchange(pred_block, get_nodes_block(pred_ins[0]));
 233         }
 234 }
 235
 236
 237 static void prepare_path(ir_node* block, int i, const ir_node* dependency)
 238 {
 239         ir_node* pred = get_nodes_block(get_irn_n(block, i));
 240         int pred_arity;
 241         int j;
 242
 243         DB((dbg, LEVEL_1, "Preparing predecessor %d of %+F\n", i, block));
 244
 245         pred_arity = get_irn_arity(pred);
 246         for (j = 0; j < pred_arity; ++j) {
 247                 ir_node* pred_pred = get_nodes_block(get_irn_n(pred, j));
 248
 249                 if (is_cdep_on(pred_pred, dependency)) {
 250                         prepare_path(pred, j, dependency);
 251                         split_block(block, i, j);
 252                         break;
 253                 }
 254         }
 255 }
 256
 257
 258 static void if_conv_walker(ir_node* block, void* env)
 259 {
 260         int arity;
 261         int i;
 262         opt_if_conv_info_t *opt_info = env;
 263
 264         /* Bail out, if there are no Phis at all */
 265         if (get_block_blockinfo(block)->phi == NULL) return;
 266
 267 restart:
 268         arity = get_irn_arity(block);
 269         for (i = 0; i < arity; ++i) {
 270                 ir_node* pred;
 271                 cdep* cdep;
 272
 273                 pred = get_nodes_block(get_irn_n(block, i));
 274                 for (cdep = find_cdep(pred); cdep != NULL; cdep = cdep->next) {
 275                         const ir_node* dependency = cdep->node;
 276                         ir_node* projx0 = walk_to_projx(pred, dependency);
 277                         ir_node* cond;
 278                         int j;
 279
 280                         if (projx0 == NULL) continue;
 281
 282                         cond = get_Proj_pred(projx0);
 283                         if (get_irn_op(cond) != op_Cond) continue;
 284
 285                         /* We only handle boolean decisions, no switches */
 286                         if (get_irn_mode(get_Cond_selector(cond)) != mode_b) continue;
 287
 288                         for (j = i + 1; j < arity; ++j) {
 289                                 ir_node* projx1;
 290                                 ir_node* conds[1];
 291                                 ir_node* vals[2];
 292                                 ir_node* psi = NULL;
 293                                 ir_node* psi_block;
 294                                 ir_node* phi;
 295
 296                                 pred = get_nodes_block(get_irn_n(block, j));
 297
 298                                 if (!is_cdep_on(pred, dependency)) continue;
 299
 300                                 projx1 = walk_to_projx(pred, dependency);
 301
 302                                 if (projx1 == NULL) continue;
 303
 304                                 phi = get_block_blockinfo(block)->phi;
 305                                 if (!opt_info->allow_ifconv(get_Cond_selector(cond), phi, i, j)) continue;
 306
 307                                 DB((dbg, LEVEL_1, "Found Cond %+F with proj %+F and %+F\n",
 308                                         cond, projx0, projx1
 309                                 ));
 310
 311                                 prepare_path(block, i, dependency);
 312                                 prepare_path(block, j, dependency);
 313                                 arity = get_irn_arity(block);
 314
 315                                 conds[0] = get_Cond_selector(cond);
 316
 317                                 psi_block = get_nodes_block(cond);
 318                                 do {
 319                                         ir_node* val_i = get_irn_n(phi, i);
 320                                         ir_node* val_j = get_irn_n(phi, j);
 321
 322                                         if (val_i == val_j) {
 323                                                 psi = val_i;
 324                                                 DB((dbg, LEVEL_2,  "Generating no psi, because both values are equal\n"));
 325                                         } else {
 326                                                 /* Something is very fishy if two predecessors of a PhiM point into
 327                                                  * one block, but not at the same memory node
 328                                                  */
 329                                                 assert(get_irn_mode(phi) != mode_M);
 330                                                 if (get_Proj_proj(projx0) == pn_Cond_true) {
 331                                                         vals[0] = val_i;
 332                                                         vals[1] = val_j;
 333                                                 } else {
 334                                                         vals[0] = val_j;
 335                                                         vals[1] = val_i;
 336                                                 }
 337
 338                                                 psi = new_r_Psi(
 339                                                         current_ir_graph, psi_block, 1, conds, vals, get_irn_mode(phi)
 340                                                 );
 341                                                 DB((dbg, LEVEL_2, "Generating %+F for %+F\n", psi, phi));
 342                                         }
 343
 344                                         if (arity == 2) {
 345                                                 exchange(phi, psi);
 346                                         } else {
 347                                                 rewire(phi, i, j, psi);
 348                                         }
 349
 350                                         phi = get_irn_link(phi);
 351                                 } while (phi != NULL);
 352
 353                                 exchange(get_nodes_block(get_irn_n(block, i)), psi_block);
 354                                 exchange(get_nodes_block(get_irn_n(block, j)), psi_block);
 355
 356                                 if (arity == 2) {
 357 #if 1
 358                                         DB((dbg, LEVEL_1,  "Welding block %+F and %+F\n", block, psi_block));
 359                                         /* copy the block-info from the Psi-block to the block before merging */
 360                                         get_block_blockinfo(psi_block)->has_pinned |= get_block_blockinfo(block)->has_pinned;
 361                                         set_irn_link(block, get_irn_link(psi_block));
 362
 363                                         set_irn_in(block, get_irn_arity(psi_block), get_irn_in(psi_block) + 1);
 364                                         exchange_cdep(psi_block, block);
 365                                         exchange(psi_block, block);
 366 #else
 367                                         DB((dbg, LEVEL_1,  "Welding block %+F to %+F\n", block, psi_block));
 368                                         get_block_blockinfo(psi_block)->has_pinned |=   get_block_blockinfo(block)->has_pinned;
 369                                         exchange(block, psi_block);
 370 #endif
 371                                         return;
 372                                 } else {
 373                                         rewire(block, i, j, new_r_Jmp(current_ir_graph, psi_block));
 374                                         goto restart;
 375                                 }
 376                         }
 377                 }
 378         }
 379 }
 380
 381 /**
 382  * Block walker: add additional data
 383  */
 384 static void init_block_link(ir_node *block, void *env)
 385 {
 386         struct obstack *obst = env;
 387         block_info *bi = obstack_alloc(obst, sizeof(*bi));
 388
 389         bi->phi = NULL;
 390         bi->has_pinned = 0;
 391         set_irn_link(block, bi);
 392 }
 393
 394
 395 /**
 396  * Daisy-chain all phis in a block
 397  * If a non-movable node is encountered set the has_pinned flag
 398  */
 399 static void collect_phis(ir_node *node, void *env)
 400 {
 401         if (is_Phi(node)) {
 402                 ir_node *block = get_nodes_block(node);
 403                 block_info *bi = get_block_blockinfo(block);
 404
 405                 set_irn_link(node, bi->phi);
 406                 bi->phi = node;
 407         }
 408         else {
 409                 if (is_no_Block(node) && get_irn_pinned(node) == op_pin_state_pinned) {
 410                         /*
 411                          * Ignore control flow nodes, these will be removed.
 412                          * This ignores Raise. That is surely bad. FIXME.
 413                          */
 414                         if (! is_cfop(node)) {
 415                                 ir_node *block = get_nodes_block(node);
 416                                 block_info *bi = get_block_blockinfo(block);
 417
 418                                 DB((dbg, LEVEL_2, "Node %+F in block %+F is unmovable\n", node, block));
 419                                 bi->has_pinned = 1;
 420                         }
 421                 }
 422         }
 423 }
 424
 425
 426 /*
 427  * Transform multiple cascaded Psis into one Psi
 428  */
 429 static ir_node* fold_psi(ir_node* psi)
 430 {
 431         int arity = get_Psi_n_conds(psi);
 432         int new_arity = 0;
 433         int i;
 434         ir_node* n;
 435         ir_node** conds;
 436         ir_node** vals;
 437         int j;
 438         int k;
 439         int a;
 440         ir_node* new_psi;
 441
 442         for (i = 0; i < arity; ++i) {
 443                 n = get_Psi_val(psi, i);
 444                 if (get_irn_op(n) == op_Psi) {
 445                         new_arity += get_Psi_n_conds(n) + 1;
 446                 } else {
 447                         ++new_arity;
 448                 }
 449         }
 450         n = get_Psi_default(psi);
 451         if (get_irn_op(n) == op_Psi) {
 452                 new_arity += get_Psi_n_conds(n);
 453         }
 454
 455         if (arity == new_arity) return psi; // no attached Psis found
 456         DB((dbg, LEVEL_1, "Folding %+F from %d to %d conds\n", psi, arity, new_arity));
 457
 458         NEW_ARR_A(ir_node *, conds, new_arity);
 459         NEW_ARR_A(ir_node *, vals, new_arity + 1);
 460         j = 0;
 461         for (i = 0; i < arity; ++i) {
 462                 ir_node* c = get_Psi_cond(psi, i);
 463
 464                 n = get_Psi_val(psi, i);
 465                 if (get_irn_op(n) == op_Psi) {
 466                         a = get_Psi_n_conds(n);
 467                         for (k = 0; k < a; ++k) {
 468                                 conds[j] = new_r_And(
 469                                         current_ir_graph, get_nodes_block(psi),
 470                                         c, get_Psi_cond(n, k), mode_b
 471                                 );
 472                                 vals[j] = get_Psi_val(n, k);
 473                                 ++j;
 474                         }
 475                         conds[j] = c;
 476                         vals[j] = get_Psi_default(n);
 477                 } else {
 478                         conds[j] = c;
 479                         vals[j] = n;
 480                 }
 481                 ++j;
 482         }
 483         n = get_Psi_default(psi);
 484         if (get_irn_op(n) == op_Psi) {
 485                 a = get_Psi_n_conds(n);
 486                 for (k = 0; k < a; ++k) {
 487                         conds[j] = get_Psi_cond(n, k);
 488                         vals[j] = get_Psi_val(n, k);
 489                         ++j;
 490                 }
 491                 vals[j] = get_Psi_default(n);
 492         } else {
 493                 vals[j] = n;
 494         }
 495         assert(j == new_arity);
 496         new_psi = new_r_Psi(
 497                 current_ir_graph, get_nodes_block(psi),
 498                 new_arity, conds, vals, get_irn_mode(psi)
 499         );
 500         DB((dbg, LEVEL_1, "Folded %+F into new %+F\n", psi, new_psi));
 501         exchange(psi, new_psi);
 502         return new_psi;
 503 }
 504
 505
 506 /*
 507  * Merge consecutive psi inputs if the data inputs are the same
 508  */
 509 static ir_node* meld_psi(ir_node* psi)
 510 {
 511         int arity = get_Psi_n_conds(psi);
 512         int new_arity;
 513         ir_node** conds;
 514         ir_node** vals;
 515         ir_node* cond;
 516         ir_node* val;
 517         int i;
 518         int j;
 519         ir_node* new_psi;
 520
 521         new_arity = 1;
 522         val = get_Psi_val(psi, 0);
 523         DB((dbg, LEVEL_1, "Pred  0 of %+F is %+F\n", psi, val));
 524         for (i = 1; i < arity; ++i) {
 525                 ir_node* v = get_Psi_val(psi, i);
 526                 DB((dbg, LEVEL_1, "Pred %2d of %+F is %+F\n", i, psi, v));
 527                 if (val != v) {
 528                         val = v;
 529                         ++new_arity;
 530                 }
 531         }
 532         DB((dbg, LEVEL_1, "Default of %+F is %+F\n", psi, get_Psi_default(psi)));
 533         if (val == get_Psi_default(psi)) --new_arity;
 534
 535         DB((dbg, LEVEL_1, "Melding Psi %+F from %d conds to %d\n", psi, arity, new_arity));
 536
 537         if (new_arity == arity) return psi;
 538
 539         /* If all data inputs of the Psi are equal, exchange the Psi with that value */
 540         if (new_arity == 0) {
 541                 exchange(psi, val);
 542                 return val;
 543         }
 544
 545         NEW_ARR_A(ir_node *, conds, new_arity);
 546         NEW_ARR_A(ir_node *, vals, new_arity + 1);
 547         cond = get_Psi_cond(psi, 0);
 548         val = get_Psi_val(psi, 0);
 549         j = 0;
 550         for (i = 1; i < arity; ++i) {
 551                 ir_node* v = get_Psi_val(psi, i);
 552
 553                 if (v == val) {
 554                         cond = new_r_Or(
 555                                 current_ir_graph, get_nodes_block(psi),
 556                                 cond, get_Psi_cond(psi, i), mode_b
 557                         );
 558                 } else {
 559                         conds[j] = cond;
 560                         vals[j] = val;
 561                         ++j;
 562                         val = v;
 563                 }
 564         }
 565         if (val != get_Psi_default(psi)) {
 566                 conds[j] = cond;
 567                 vals[j] = val;
 568                 ++j;
 569         }
 570         vals[j] = get_Psi_default(psi);
 571         assert(j == new_arity);
 572         new_psi = new_r_Psi(
 573                 current_ir_graph, get_nodes_block(psi),
 574                 new_arity, conds, vals, get_irn_mode(psi)
 575         );
 576         DB((dbg, LEVEL_1, "Molded %+F into %+F\n", psi, new_psi));
 577         exchange(psi, new_psi);
 578         return new_psi;
 579 }
 580
 581
 582 /**
 583  * Split a Psi with multiple conditions into multiple Psis with one condtition
 584  * each
 585  */
 586 static ir_node* split_psi(ir_node* psi)
 587 {
 588         int arity = get_Psi_n_conds(psi);
 589         ir_mode* mode;
 590         ir_node* block;
 591         ir_node* rval;
 592         int i;
 593
 594         if (arity == 1) return psi;
 595
 596         mode = get_irn_mode(psi);
 597         block = get_nodes_block(psi);
 598         rval = get_Psi_default(psi);
 599         for (i = arity - 1; i >= 0; --i) {
 600                 ir_node* conds[1];
 601                 ir_node* vals[2];
 602
 603                 conds[0] = get_Psi_cond(psi, i);
 604                 vals[0] = get_Psi_val(psi, i);
 605                 vals[1] = rval;
 606                 rval = new_r_Psi(
 607                         current_ir_graph, block, 1, conds, vals, mode
 608                 );
 609         }
 610         exchange(psi, rval);
 611         return rval;
 612 }
 613
 614
 615 static void optimise_psis(ir_node* node, void* env)
 616 {
 617         if (get_irn_op(node) != op_Psi) return;
 618 #if 1
 619         node = fold_psi(node);
 620 #endif
 621 #if 1
 622         node = meld_psi(node);
 623 #endif
 624 #if 1
 625         node = split_psi(node);
 626 #endif
 627 }
 628
 629
 630 void opt_if_conv(ir_graph *irg, const opt_if_conv_info_t *params)
 631 {
 632         struct obstack obst;
 633         opt_if_conv_info_t p;
 634
 635         if (! get_opt_if_conversion())
 636                 return;
 637
 638         /* get the parameters */
 639         p = (params != NULL ? *params : default_info);
 640
 641         FIRM_DBG_REGISTER(dbg, "firm.opt.ifconv");
 642
 643         DB((dbg, LEVEL_1, "Running if-conversion on %+F\n", irg));
 644
 645         normalize_one_return(irg);
 646         remove_critical_cf_edges(irg);
 647
 648         compute_cdep(irg);
 649         assure_doms(irg);
 650
 651         obstack_init(&obst);
 652         irg_block_walk_graph(irg, init_block_link, NULL, &obst);
 653         irg_walk_graph(irg, collect_phis, NULL, NULL);
 654         irg_block_walk_graph(irg, NULL, if_conv_walker, &p);
 655
 656         local_optimize_graph(irg);
 657
 658         irg_walk_graph(irg, NULL, optimise_psis, NULL);
 659
 660         obstack_free(&obst, NULL);
 661
 662         free_dom(irg);
 663         free_cdep(irg);
 664 }