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