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

   1 /*
   2  * Copyright (C) 1995-2008 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 #include "config.h"
  28
  29 #include <assert.h>
  30 #include "iroptimize.h"
  31 #include "obst.h"
  32 #include "irnode_t.h"
  33 #include "cdep.h"
  34 #include "ircons.h"
  35 #include "irdom.h"
  36 #include "irgmod.h"
  37 #include "irgopt.h"
  38 #include "irgwalk.h"
  39 #include "irtools.h"
  40 #include "array_t.h"
  41
  42 // debug
  43 #include "irdump.h"
  44 #include "debug.h"
  45
  46 DEBUG_ONLY(static firm_dbg_module_t *dbg);
  47
  48 /** allow every Mux to be created. */
  49 static int default_allow_ifconv(ir_node *sel, ir_node* phi_list, int i, int j)
  50 {
  51         (void) sel;
  52         (void) phi_list;
  53         (void) i;
  54         (void) j;
  55         return 1;
  56 }
  57
  58 /**
  59  * Default options.
  60  */
  61 static const ir_settings_if_conv_t default_info = {
  62         0,    /* doesn't matter for Mux */
  63         default_allow_ifconv
  64 };
  65
  66 /**
  67  * Returns non-zero if a Block can be emptied.
  68  */
  69 static int can_empty_block(ir_node *block) {
  70         return get_Block_mark(block) == 0;
  71 }
  72
  73
  74 static ir_node* walk_to_projx(ir_node* start, const ir_node* dependency)
  75 {
  76         int arity;
  77         int i;
  78
  79         /* No need to find the conditional block if this block cannot be emptied and
  80          * therefore not moved */
  81         if (!can_empty_block(start)) return NULL;
  82
  83         arity = get_irn_arity(start);
  84         for (i = 0; i < arity; ++i) {
  85                 ir_node* pred = get_irn_n(start, i);
  86                 ir_node* pred_block = get_nodes_block(pred);
  87
  88                 if (pred_block == dependency) {
  89                         if (is_Proj(pred)) {
  90                                 assert(get_irn_mode(pred) == mode_X);
  91                                 return pred;
  92                         }
  93                         return NULL;
  94                 }
  95
  96                 if (is_Proj(pred)) {
  97                         assert(get_irn_mode(pred) == mode_X);
  98                         return NULL;
  99                 }
 100
 101                 if (is_cdep_on(pred_block, dependency)) {
 102                         return walk_to_projx(pred_block, dependency);
 103                 }
 104         }
 105         return NULL;
 106 }
 107
 108
 109 /**
 110  * Copies the DAG starting at node to the ith predecessor block of src_block
 111  * -if the node isn't in the src_block, this is a nop and the node is returned
 112  * -if the node is a phi in the src_block, the ith predecessor of the phi is
 113  *   returned
 114  * otherwise returns the copy of the passed node
 115  */
 116 static ir_node* copy_to(ir_node* node, ir_node* src_block, int i)
 117 {
 118         ir_node* dst_block;
 119         ir_node* copy;
 120         int arity;
 121         int j;
 122
 123         if (get_nodes_block(node) != src_block) return node;
 124         if (get_irn_op(node) == op_Phi) return get_irn_n(node, i);
 125
 126         copy = exact_copy(node);
 127         dst_block = get_nodes_block(get_irn_n(src_block, i));
 128         set_nodes_block(copy, dst_block);
 129
 130         DB((dbg, LEVEL_1, "Copying node %+F to block %+F, copy is %+F\n",
 131                 node, dst_block, copy));
 132
 133         arity = get_irn_arity(node);
 134         for (j = 0; j < arity; ++j) {
 135                 set_irn_n(copy, j, copy_to(get_irn_n(node, j), src_block, i));
 136                 DB((dbg, LEVEL_2, "-- pred %d is %+F\n", j, get_irn_n(copy, j)));
 137         }
 138         return copy;
 139 }
 140
 141
 142 /**
 143  * Remove predecessors i and j from node and add predecessor new_pred
 144  */
 145 static void rewire(ir_node* node, int i, int j, ir_node* new_pred)
 146 {
 147         int arity = get_irn_arity(node);
 148         ir_node **ins;
 149         int k;
 150         int l;
 151
 152         NEW_ARR_A(ir_node *, ins, arity - 1);
 153
 154         l = 0;
 155         for (k = 0; k < i;     ++k) ins[l++] = get_irn_n(node, k);
 156         for (++k;   k < j;     ++k) ins[l++] = get_irn_n(node, k);
 157         for (++k;   k < arity; ++k) ins[l++] = get_irn_n(node, k);
 158         ins[l++] = new_pred;
 159         assert(l == arity - 1);
 160         set_irn_in(node, l, ins);
 161 }
 162
 163
 164 /**
 165  * Remove the jth predecessors from the ith predecessor of block and add it to block
 166  */
 167 static void split_block(ir_node* block, int i, int j)
 168 {
 169         ir_node* pred_block = get_nodes_block(get_irn_n(block, i));
 170         int arity = get_irn_arity(block);
 171         int new_pred_arity;
 172         ir_node *phi, *next;
 173         ir_node **ins;
 174         ir_node **pred_ins;
 175         int k;
 176
 177         DB((dbg, LEVEL_1, "Splitting predecessor %d of predecessor %d of %+F\n", j, i, block));
 178
 179         NEW_ARR_A(ir_node*, ins, arity + 1);
 180
 181         for (phi = get_Block_phis(block); phi != NULL; phi = get_Phi_next(phi)) {
 182                 ir_node* copy = copy_to(get_irn_n(phi, i), pred_block, j);
 183
 184                 for (k = 0; k < i; ++k) ins[k] = get_irn_n(phi, k);
 185                 ins[k++] = copy;
 186                 for (; k < arity; ++k) ins[k] = get_irn_n(phi, k);
 187                 ins[k] = get_irn_n(phi, i);
 188                 assert(k == arity);
 189                 set_irn_in(phi, arity + 1, ins);
 190         }
 191
 192         for (k = 0; k < i; ++k) ins[k] = get_irn_n(block, k);
 193         ins[k++] = get_irn_n(pred_block, j);
 194         for (; k < arity; ++k) ins[k] = get_irn_n(block, k);
 195         ins[k] = get_irn_n(block, i);
 196         assert(k == arity);
 197         set_irn_in(block, arity + 1, ins);
 198
 199         new_pred_arity = get_irn_arity(pred_block) - 1;
 200         NEW_ARR_A(ir_node*, pred_ins, new_pred_arity);
 201
 202         for (phi = get_Block_phis(pred_block); phi != NULL; phi = next) {
 203                 for (k = 0; k < j; ++k) pred_ins[k] = get_irn_n(phi, k);
 204                 for (; k < new_pred_arity; ++k) pred_ins[k] = get_irn_n(phi, k + 1);
 205                 assert(k == new_pred_arity);
 206                 next = get_Phi_next(phi);
 207                 if (new_pred_arity > 1) {
 208                         set_irn_in(phi, new_pred_arity, pred_ins);
 209                 } else {
 210                         exchange(phi, pred_ins[0]);
 211                 }
 212         }
 213
 214         for (k = 0; k < j; ++k) pred_ins[k] = get_irn_n(pred_block, k);
 215         for (; k < new_pred_arity; ++k) pred_ins[k] = get_irn_n(pred_block, k + 1);
 216         assert(k == new_pred_arity);
 217         if (new_pred_arity > 1) {
 218                 set_irn_in(pred_block, new_pred_arity, pred_ins);
 219         } else {
 220                 exchange(pred_block, get_nodes_block(pred_ins[0]));
 221         }
 222 }
 223
 224
 225 static void prepare_path(ir_node* block, int i, const ir_node* dependency)
 226 {
 227         ir_node* pred = get_nodes_block(get_irn_n(block, i));
 228         int pred_arity;
 229         int j;
 230
 231         DB((dbg, LEVEL_1, "Preparing predecessor %d of %+F\n", i, block));
 232
 233         pred_arity = get_irn_arity(pred);
 234         for (j = 0; j < pred_arity; ++j) {
 235                 ir_node* pred_pred = get_nodes_block(get_irn_n(pred, j));
 236
 237                 if (is_cdep_on(pred_pred, dependency)) {
 238                         prepare_path(pred, j, dependency);
 239                         split_block(block, i, j);
 240                         break;
 241                 }
 242         }
 243 }
 244
 245
 246 static void if_conv_walker(ir_node* block, void* env)
 247 {
 248         ir_settings_if_conv_t* opt_info = env;
 249         int arity;
 250         int i;
 251
 252         /* Bail out, if there are no Phis at all */
 253         if (get_Block_phis(block) == NULL) return;
 254
 255 restart:
 256         arity = get_irn_arity(block);
 257         for (i = 0; i < arity; ++i) {
 258                 ir_node* pred0;
 259                 ir_cdep* cdep;
 260
 261                 pred0 = get_Block_cfgpred_block(block, i);
 262                 for (cdep = find_cdep(pred0); cdep != NULL; cdep = cdep->next) {
 263                         const ir_node* dependency = cdep->node;
 264                         ir_node* projx0 = walk_to_projx(pred0, dependency);
 265                         ir_node* cond;
 266                         int j;
 267
 268                         if (projx0 == NULL) continue;
 269
 270                         cond = get_Proj_pred(projx0);
 271                         if (! is_Cond(cond))
 272                                 continue;
 273
 274                         /* We only handle boolean decisions, no switches */
 275                         if (get_irn_mode(get_Cond_selector(cond)) != mode_b) continue;
 276
 277                         for (j = i + 1; j < arity; ++j) {
 278                                 ir_node* projx1;
 279                                 ir_node* sel;
 280                                 ir_node* mux_block;
 281                                 ir_node* phi;
 282                                 ir_node* pred1;
 283                                 dbg_info* cond_dbg;
 284
 285                                 pred1 = get_Block_cfgpred_block(block, j);
 286
 287                                 if (!is_cdep_on(pred1, dependency)) continue;
 288
 289                                 projx1 = walk_to_projx(pred1, dependency);
 290
 291                                 if (projx1 == NULL) continue;
 292
 293                                 phi = get_Block_phis(block);
 294                                 if (!opt_info->allow_ifconv(get_Cond_selector(cond), phi, i, j)) continue;
 295
 296                                 DB((dbg, LEVEL_1, "Found Cond %+F with proj %+F and %+F\n",
 297                                         cond, projx0, projx1
 298                                 ));
 299
 300                                 prepare_path(block, i, dependency);
 301                                 prepare_path(block, j, dependency);
 302                                 arity = get_irn_arity(block);
 303
 304                                 sel = get_Cond_selector(cond);
 305
 306                                 mux_block = get_nodes_block(cond);
 307                                 cond_dbg = get_irn_dbg_info(cond);
 308                                 do {
 309                                         ir_node* val_i = get_irn_n(phi, i);
 310                                         ir_node* val_j = get_irn_n(phi, j);
 311                                         ir_node* mux;
 312                                         ir_node* next_phi;
 313
 314                                         if (val_i == val_j) {
 315                                                 mux = val_i;
 316                                                 DB((dbg, LEVEL_2,  "Generating no Mux, because both values are equal\n"));
 317                                         } else {
 318                                                 ir_node *t, *f;
 319
 320                                                 /* Something is very fishy if two predecessors of a PhiM point into
 321                                                  * one block, but not at the same memory node
 322                                                  */
 323                                                 assert(get_irn_mode(phi) != mode_M);
 324                                                 if (get_Proj_proj(projx0) == pn_Cond_true) {
 325                                                         t = val_i;
 326                                                         f = val_j;
 327                                                 } else {
 328                                                         t = val_j;
 329                                                         f = val_i;
 330                                                 }
 331
 332                                                 mux = new_rd_Mux(cond_dbg, current_ir_graph, mux_block, sel, f, t, get_irn_mode(phi));
 333                                                 DB((dbg, LEVEL_2, "Generating %+F for %+F\n", mux, phi));
 334                                         }
 335
 336                                         next_phi = get_Phi_next(phi);
 337
 338                                         if (arity == 2) {
 339                                                 exchange(phi, mux);
 340                                         } else {
 341                                                 rewire(phi, i, j, mux);
 342                                         }
 343
 344                                         phi = next_phi;
 345                                 } while (phi != NULL);
 346
 347                                 exchange(get_nodes_block(get_irn_n(block, i)), mux_block);
 348                                 exchange(get_nodes_block(get_irn_n(block, j)), mux_block);
 349
 350                                 if (arity == 2) {
 351                                         unsigned mark;
 352 #if 1
 353                                         DB((dbg, LEVEL_1,  "Welding block %+F and %+F\n", block, mux_block));
 354                                         /* copy the block-info from the Mux-block to the block before merging */
 355
 356                                         mark =  get_Block_mark(mux_block) | get_Block_mark(block);
 357                                         set_Block_mark(block, mark);
 358                                         set_Block_phis(block, get_Block_phis(mux_block));
 359
 360                                         set_irn_in(block, get_irn_arity(mux_block), get_irn_in(mux_block) + 1);
 361                                         exchange_cdep(mux_block, block);
 362                                         exchange(mux_block, block);
 363 #else
 364                                         DB((dbg, LEVEL_1,  "Welding block %+F to %+F\n", block, mux_block));
 365                                         mark =  get_Block_mark(mux_block) | get_Block_mark(block);
 366                                         /* mark both block just to be sure, should be enough to mark mux_block */
 367                                         set_Block_mark(mux_block, mark);
 368                                         exchange(block, mux_block);
 369 #endif
 370                                         return;
 371                                 } else {
 372                                         rewire(block, i, j, new_r_Jmp(current_ir_graph, mux_block));
 373                                         goto restart;
 374                                 }
 375                         }
 376                 }
 377         }
 378 }
 379
 380 /**
 381  * Block walker: clear block mark and Phi list
 382  */
 383 static void init_block_link(ir_node *block, void *env)
 384 {
 385         (void)env;
 386         set_Block_mark(block, 0);
 387         set_Block_phis(block, NULL);
 388 }
 389
 390
 391 /**
 392  * Daisy-chain all phis in a block
 393  * If a non-movable node is encountered set the has_pinned flag in its block.
 394  */
 395 static void collect_phis(ir_node *node, void *env) {
 396         (void) env;
 397
 398         if (is_Phi(node)) {
 399                 ir_node *block = get_nodes_block(node);
 400
 401                 add_Block_phi(block, node);
 402         } else {
 403                 if (is_no_Block(node) && get_irn_pinned(node) == op_pin_state_pinned) {
 404                         /*
 405                          * Ignore control flow nodes, these will be removed.
 406                          * This ignores Raise. That is surely bad. FIXME.
 407                          */
 408                         if (!is_cfop(node)) {
 409                                 ir_node *block = get_nodes_block(node);
 410
 411                                 DB((dbg, LEVEL_2, "Node %+F in block %+F is unmovable\n", node, block));
 412                                 set_Block_mark(block, 1);
 413                         }
 414                 }
 415         }
 416 }
 417
 418 static void optimise_muxs_0(ir_node* mux, void* env)
 419 {
 420         ir_node* t;
 421         ir_node* f;
 422
 423         (void) env;
 424
 425         if (!is_Mux(mux)) return;
 426
 427         t = get_Mux_true(mux);
 428         f = get_Mux_false(mux);
 429
 430         DB((dbg, LEVEL_3, "Simplify %+F T=%+F F=%+F\n", mux, t, f));
 431
 432         if (is_Unknown(t)) {
 433                 DB((dbg, LEVEL_3, "Replace Mux with unknown operand by %+F\n", f));
 434                 exchange(mux, f);
 435                 return;
 436         }
 437         if (is_Unknown(f)) {
 438                 DB((dbg, LEVEL_3, "Replace Mux with unknown operand by %+F\n", t));
 439                 exchange(mux, t);
 440                 return;
 441         }
 442
 443         if (is_Mux(t)) {
 444                 ir_graph* irg   = current_ir_graph;
 445                 ir_node*  block = get_nodes_block(mux);
 446                 ir_mode*  mode  = get_irn_mode(mux);
 447                 ir_node*  c0    = get_Mux_sel(mux);
 448                 ir_node*  c1    = get_Mux_sel(t);
 449                 ir_node*  t1    = get_Mux_true(t);
 450                 ir_node*  f1    = get_Mux_false(t);
 451                 if (f == f1) {
 452                         /* Mux(c0, Mux(c1, x, y), y) -> typical if (c0 && c1) x else y */
 453                         ir_node* and_    = new_r_And(irg, block, c0, c1, mode_b);
 454                         ir_node* new_mux = new_r_Mux(irg, block, and_, f1, t1, mode);
 455                         exchange(mux, new_mux);
 456                 } else if (f == t1) {
 457                         /* Mux(c0, Mux(c1, x, y), x) */
 458                         ir_node* not_c1 = new_r_Not(irg, block, c1, mode_b);
 459                         ir_node* and_   = new_r_And(irg, block, c0, not_c1, mode_b);
 460                         ir_node* new_mux = new_r_Mux(irg, block, and_, t1, f1, mode);
 461                         exchange(mux, new_mux);
 462                 }
 463         } else if (is_Mux(f)) {
 464                 ir_graph* irg   = current_ir_graph;
 465                 ir_node*  block = get_nodes_block(mux);
 466                 ir_mode*  mode  = get_irn_mode(mux);
 467                 ir_node*  c0    = get_Mux_sel(mux);
 468                 ir_node*  c1    = get_Mux_sel(f);
 469                 ir_node*  t1    = get_Mux_true(f);
 470                 ir_node*  f1    = get_Mux_false(f);
 471                 if (t == t1) {
 472                         /* Mux(c0, x, Mux(c1, x, y)) -> typical if (c0 || c1) x else y */
 473                         ir_node* or_     = new_r_Or(irg, block, c0, c1, mode_b);
 474                         ir_node* new_mux = new_r_Mux(irg, block, or_, f1, t1, mode);
 475                         exchange(mux, new_mux);
 476                 } else if (t == f1) {
 477                         /* Mux(c0, x, Mux(c1, y, x)) */
 478                         ir_node* not_c1  = new_r_Not(irg, block, c1, mode_b);
 479                         ir_node* or_     = new_r_Or(irg, block, c0, not_c1, mode_b);
 480                         ir_node* new_mux = new_r_Mux(irg, block, or_, t1, f1, mode);
 481                         exchange(mux, new_mux);
 482                 }
 483         }
 484 }
 485
 486
 487 static void optimise_muxs_1(ir_node* mux, void* env)
 488 {
 489         ir_node* t;
 490         ir_node* f;
 491         ir_mode* mode;
 492
 493         (void) env;
 494
 495         if (!is_Mux(mux)) return;
 496
 497         t = get_Mux_true(mux);
 498         f = get_Mux_false(mux);
 499
 500         DB((dbg, LEVEL_3, "Simplify %+F T=%+F F=%+F\n", mux, t, f));
 501
 502         mode = get_irn_mode(mux);
 503
 504         if (is_Const(t) && is_Const(f) && (mode_is_int(mode))) {
 505                 ir_node* block = get_nodes_block(mux);
 506                 ir_node* c     = get_Mux_sel(mux);
 507                 tarval* tv_t = get_Const_tarval(t);
 508                 tarval* tv_f = get_Const_tarval(f);
 509                 if (tarval_is_one(tv_t) && tarval_is_null(tv_f)) {
 510                         ir_node* conv  = new_r_Conv(current_ir_graph, block, c, mode);
 511                         exchange(mux, conv);
 512                 } else if (tarval_is_null(tv_t) && tarval_is_one(tv_f)) {
 513                         ir_node* not_  = new_r_Not(current_ir_graph, block, c, mode_b);
 514                         ir_node* conv  = new_r_Conv(current_ir_graph, block, not_, mode);
 515                         exchange(mux, conv);
 516                 }
 517         }
 518 }
 519
 520
 521 void opt_if_conv(ir_graph *irg, const ir_settings_if_conv_t *params)
 522 {
 523         ir_settings_if_conv_t p;
 524
 525         /* get the parameters */
 526         p = (params != NULL ? *params : default_info);
 527
 528         FIRM_DBG_REGISTER(dbg, "firm.opt.ifconv");
 529
 530         DB((dbg, LEVEL_1, "Running if-conversion on %+F\n", irg));
 531
 532         normalize_one_return(irg);
 533         remove_critical_cf_edges(irg);
 534
 535         compute_cdep(irg);
 536         assure_doms(irg);
 537
 538         ir_reserve_resources(irg, IR_RESOURCE_BLOCK_MARK);
 539
 540         irg_block_walk_graph(irg, init_block_link, NULL, NULL);
 541         irg_walk_graph(irg, collect_phis, NULL, NULL);
 542         irg_block_walk_graph(irg, NULL, if_conv_walker, &p);
 543
 544         ir_free_resources(irg, IR_RESOURCE_BLOCK_MARK);
 545
 546         local_optimize_graph(irg);
 547
 548         irg_walk_graph(irg, NULL, optimise_muxs_0, NULL);
 549 #if 1
 550         irg_walk_graph(irg, NULL, optimise_muxs_1, NULL);
 551 #endif
 552
 553         /* TODO: graph might be changed, handle more graceful */
 554         set_irg_outs_inconsistent(irg);
 555         set_irg_extblk_inconsistent(irg);
 556         set_irg_loopinfo_inconsistent(irg);
 557         free_dom(irg);
 558
 559         free_cdep(irg);
 560 }