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 #ifdef HAVE_CONFIG_H
  28 #include "config.h"
  29 #endif
  30
  31 #include <assert.h>
  32 #include "iroptimize.h"
  33 #include "obst.h"
  34 #include "irnode_t.h"
  35 #include "cdep.h"
  36 #include "ircons.h"
  37 #include "irdom.h"
  38 #include "irgmod.h"
  39 #include "irgopt.h"
  40 #include "irgwalk.h"
  41 #include "irtools.h"
  42 #include "array.h"
  43 #include "xmalloc.h"
  44
  45 // debug
  46 #include "irdump.h"
  47 #include "debug.h"
  48
  49 DEBUG_ONLY(static firm_dbg_module_t *dbg);
  50
  51 /** allow every Mux to be created. */
  52 static int default_allow_ifconv(ir_node *sel, ir_node* phi_list, int i, int j)
  53 {
  54         (void) sel;
  55         (void) phi_list;
  56         (void) i;
  57         (void) j;
  58         return 1;
  59 }
  60
  61 /**
  62  * Default options.
  63  */
  64 static const ir_settings_if_conv_t default_info = {
  65         0,    /* doesn't matter for Mux */
  66         default_allow_ifconv
  67 };
  68
  69 /**
  70  * Returns non-zero if a Block can be emptied.
  71  */
  72 static int can_empty_block(ir_node *block) {
  73         return get_Block_mark(block) == 0;
  74 }
  75
  76
  77 static ir_node* walk_to_projx(ir_node* start, const ir_node* dependency)
  78 {
  79         int arity;
  80         int i;
  81
  82         /* No need to find the conditional block if this block cannot be emptied and
  83          * therefore not moved */
  84         if (!can_empty_block(start)) return NULL;
  85
  86         arity = get_irn_arity(start);
  87         for (i = 0; i < arity; ++i) {
  88                 ir_node* pred = get_irn_n(start, i);
  89                 ir_node* pred_block = get_nodes_block(pred);
  90
  91                 if (pred_block == dependency) {
  92                         if (is_Proj(pred)) {
  93                                 assert(get_irn_mode(pred) == mode_X);
  94                                 return pred;
  95                         }
  96                         return NULL;
  97                 }
  98
  99                 if (is_Proj(pred)) {
 100                         assert(get_irn_mode(pred) == mode_X);
 101                         return NULL;
 102                 }
 103
 104                 if (is_cdep_on(pred_block, dependency)) {
 105                         return walk_to_projx(pred_block, dependency);
 106                 }
 107         }
 108         return NULL;
 109 }
 110
 111
 112 /**
 113  * Copies the DAG starting at node to the ith predecessor block of src_block
 114  * -if the node isn't in the src_block, this is a nop and the node is returned
 115  * -if the node is a phi in the src_block, the ith predecessor of the phi is
 116  *   returned
 117  * otherwise returns the copy of the passed node
 118  */
 119 static ir_node* copy_to(ir_node* node, ir_node* src_block, int i)
 120 {
 121         ir_node* dst_block;
 122         ir_node* copy;
 123         int arity;
 124         int j;
 125
 126         if (get_nodes_block(node) != src_block) return node;
 127         if (get_irn_op(node) == op_Phi) return get_irn_n(node, i);
 128
 129         copy = exact_copy(node);
 130         dst_block = get_nodes_block(get_irn_n(src_block, i));
 131         set_nodes_block(copy, dst_block);
 132
 133         DB((dbg, LEVEL_1, "Copying node %+F to block %+F, copy is %+F\n",
 134                 node, dst_block, copy));
 135
 136         arity = get_irn_arity(node);
 137         for (j = 0; j < arity; ++j) {
 138                 set_irn_n(copy, j, copy_to(get_irn_n(node, j), src_block, i));
 139                 DB((dbg, LEVEL_2, "-- pred %d is %+F\n", j, get_irn_n(copy, j)));
 140         }
 141         return copy;
 142 }
 143
 144
 145 /**
 146  * Remove predecessors i and j from node and add predecessor new_pred
 147  */
 148 static void rewire(ir_node* node, int i, int j, ir_node* new_pred)
 149 {
 150         int arity = get_irn_arity(node);
 151         ir_node **ins;
 152         int k;
 153         int l;
 154
 155         NEW_ARR_A(ir_node *, ins, arity - 1);
 156
 157         l = 0;
 158         for (k = 0; k < i;     ++k) ins[l++] = get_irn_n(node, k);
 159         for (++k;   k < j;     ++k) ins[l++] = get_irn_n(node, k);
 160         for (++k;   k < arity; ++k) ins[l++] = get_irn_n(node, k);
 161         ins[l++] = new_pred;
 162         assert(l == arity - 1);
 163         set_irn_in(node, l, ins);
 164 }
 165
 166
 167 /**
 168  * Remove the jth predecessors from the ith predecessor of block and add it to block
 169  */
 170 static void split_block(ir_node* block, int i, int j)
 171 {
 172         ir_node* pred_block = get_nodes_block(get_irn_n(block, i));
 173         int arity = get_irn_arity(block);
 174         int new_pred_arity;
 175         ir_node *phi, *next;
 176         ir_node **ins;
 177         ir_node **pred_ins;
 178         int k;
 179
 180         DB((dbg, LEVEL_1, "Splitting predecessor %d of predecessor %d of %+F\n", j, i, block));
 181
 182         NEW_ARR_A(ir_node*, ins, arity + 1);
 183
 184         for (phi = get_Block_phis(block); phi != NULL; phi = get_Phi_next(phi)) {
 185                 ir_node* copy = copy_to(get_irn_n(phi, i), pred_block, j);
 186
 187                 for (k = 0; k < i; ++k) ins[k] = get_irn_n(phi, k);
 188                 ins[k++] = copy;
 189                 for (; k < arity; ++k) ins[k] = get_irn_n(phi, k);
 190                 ins[k] = get_irn_n(phi, i);
 191                 assert(k == arity);
 192                 set_irn_in(phi, arity + 1, ins);
 193         }
 194
 195         for (k = 0; k < i; ++k) ins[k] = get_irn_n(block, k);
 196         ins[k++] = get_irn_n(pred_block, j);
 197         for (; k < arity; ++k) ins[k] = get_irn_n(block, k);
 198         ins[k] = get_irn_n(block, i);
 199         assert(k == arity);
 200         set_irn_in(block, arity + 1, ins);
 201
 202         new_pred_arity = get_irn_arity(pred_block) - 1;
 203         NEW_ARR_A(ir_node*, pred_ins, new_pred_arity);
 204
 205         for (phi = get_Block_phis(pred_block); phi != NULL; phi = next) {
 206                 for (k = 0; k < j; ++k) pred_ins[k] = get_irn_n(phi, k);
 207                 for (; k < new_pred_arity; ++k) pred_ins[k] = get_irn_n(phi, k + 1);
 208                 assert(k == new_pred_arity);
 209                 next = get_Phi_next(phi);
 210                 if (new_pred_arity > 1) {
 211                         set_irn_in(phi, new_pred_arity, pred_ins);
 212                 } else {
 213                         exchange(phi, pred_ins[0]);
 214                 }
 215         }
 216
 217         for (k = 0; k < j; ++k) pred_ins[k] = get_irn_n(pred_block, k);
 218         for (; k < new_pred_arity; ++k) pred_ins[k] = get_irn_n(pred_block, k + 1);
 219         assert(k == new_pred_arity);
 220         if (new_pred_arity > 1) {
 221                 set_irn_in(pred_block, new_pred_arity, pred_ins);
 222         } else {
 223                 exchange(pred_block, get_nodes_block(pred_ins[0]));
 224         }
 225 }
 226
 227
 228 static void prepare_path(ir_node* block, int i, const ir_node* dependency)
 229 {
 230         ir_node* pred = get_nodes_block(get_irn_n(block, i));
 231         int pred_arity;
 232         int j;
 233
 234         DB((dbg, LEVEL_1, "Preparing predecessor %d of %+F\n", i, block));
 235
 236         pred_arity = get_irn_arity(pred);
 237         for (j = 0; j < pred_arity; ++j) {
 238                 ir_node* pred_pred = get_nodes_block(get_irn_n(pred, j));
 239
 240                 if (is_cdep_on(pred_pred, dependency)) {
 241                         prepare_path(pred, j, dependency);
 242                         split_block(block, i, j);
 243                         break;
 244                 }
 245         }
 246 }
 247
 248
 249 static void if_conv_walker(ir_node* block, void* env)
 250 {
 251         ir_settings_if_conv_t* opt_info = env;
 252         int arity;
 253         int i;
 254
 255         /* Bail out, if there are no Phis at all */
 256         if (get_Block_phis(block) == NULL) return;
 257
 258 restart:
 259         arity = get_irn_arity(block);
 260         for (i = 0; i < arity; ++i) {
 261                 ir_node* pred0;
 262                 ir_cdep* cdep;
 263
 264                 pred0 = get_Block_cfgpred_block(block, i);
 265                 for (cdep = find_cdep(pred0); cdep != NULL; cdep = cdep->next) {
 266                         const ir_node* dependency = cdep->node;
 267                         ir_node* projx0 = walk_to_projx(pred0, dependency);
 268                         ir_node* cond;
 269                         int j;
 270
 271                         if (projx0 == NULL) continue;
 272
 273                         cond = get_Proj_pred(projx0);
 274                         if (! is_Cond(cond))
 275                                 continue;
 276
 277                         /* We only handle boolean decisions, no switches */
 278                         if (get_irn_mode(get_Cond_selector(cond)) != mode_b) continue;
 279
 280                         for (j = i + 1; j < arity; ++j) {
 281                                 ir_node* projx1;
 282                                 ir_node* sel;
 283                                 ir_node* mux_block;
 284                                 ir_node* phi;
 285                                 ir_node* pred1;
 286                                 dbg_info* cond_dbg;
 287
 288                                 pred1 = get_Block_cfgpred_block(block, j);
 289
 290                                 if (!is_cdep_on(pred1, dependency)) continue;
 291
 292                                 projx1 = walk_to_projx(pred1, dependency);
 293
 294                                 if (projx1 == NULL) continue;
 295
 296                                 phi = get_Block_phis(block);
 297                                 if (!opt_info->allow_ifconv(get_Cond_selector(cond), phi, i, j)) continue;
 298
 299                                 DB((dbg, LEVEL_1, "Found Cond %+F with proj %+F and %+F\n",
 300                                         cond, projx0, projx1
 301                                 ));
 302
 303                                 prepare_path(block, i, dependency);
 304                                 prepare_path(block, j, dependency);
 305                                 arity = get_irn_arity(block);
 306
 307                                 sel = get_Cond_selector(cond);
 308
 309                                 mux_block = get_nodes_block(cond);
 310                                 cond_dbg = get_irn_dbg_info(cond);
 311                                 do {
 312                                         ir_node* val_i = get_irn_n(phi, i);
 313                                         ir_node* val_j = get_irn_n(phi, j);
 314                                         ir_node* mux;
 315                                         ir_node* next_phi;
 316
 317                                         if (val_i == val_j) {
 318                                                 mux = val_i;
 319                                                 DB((dbg, LEVEL_2,  "Generating no Mux, because both values are equal\n"));
 320                                         } else {
 321                                                 ir_node *t, *f;
 322
 323                                                 /* Something is very fishy if two predecessors of a PhiM point into
 324                                                  * one block, but not at the same memory node
 325                                                  */
 326                                                 assert(get_irn_mode(phi) != mode_M);
 327                                                 if (get_Proj_proj(projx0) == pn_Cond_true) {
 328                                                         t = val_i;
 329                                                         f = val_j;
 330                                                 } else {
 331                                                         t = val_j;
 332                                                         f = val_i;
 333                                                 }
 334
 335                                                 mux = new_rd_Mux(cond_dbg, current_ir_graph, mux_block, sel, f, t, get_irn_mode(phi));
 336                                                 DB((dbg, LEVEL_2, "Generating %+F for %+F\n", mux, phi));
 337                                         }
 338
 339                                         next_phi = get_Phi_next(phi);
 340
 341                                         if (arity == 2) {
 342                                                 exchange(phi, mux);
 343                                         } else {
 344                                                 rewire(phi, i, j, mux);
 345                                         }
 346
 347                                         phi = next_phi;
 348                                 } while (phi != NULL);
 349
 350                                 exchange(get_nodes_block(get_irn_n(block, i)), mux_block);
 351                                 exchange(get_nodes_block(get_irn_n(block, j)), mux_block);
 352
 353                                 if (arity == 2) {
 354                                         unsigned mark;
 355 #if 1
 356                                         DB((dbg, LEVEL_1,  "Welding block %+F and %+F\n", block, mux_block));
 357                                         /* copy the block-info from the Mux-block to the block before merging */
 358
 359                                         mark =  get_Block_mark(mux_block) | get_Block_mark(block);
 360                                         set_Block_mark(block, mark);
 361                                         set_Block_phis(block, get_Block_phis(mux_block));
 362
 363                                         set_irn_in(block, get_irn_arity(mux_block), get_irn_in(mux_block) + 1);
 364                                         exchange_cdep(mux_block, block);
 365                                         exchange(mux_block, block);
 366 #else
 367                                         DB((dbg, LEVEL_1,  "Welding block %+F to %+F\n", block, mux_block));
 368                                         mark =  get_Block_mark(mux_block) | get_Block_mark(block);
 369                                         /* mark both block just to be sure, should be enough to mark mux_block */
 370                                         set_Block_mark(mux_block, mark);
 371                                         exchange(block, mux_block);
 372 #endif
 373                                         return;
 374                                 } else {
 375                                         rewire(block, i, j, new_r_Jmp(current_ir_graph, mux_block));
 376                                         goto restart;
 377                                 }
 378                         }
 379                 }
 380         }
 381 }
 382
 383 /**
 384  * Block walker: clear block mark and Phi list
 385  */
 386 static void init_block_link(ir_node *block, void *env)
 387 {
 388         (void)env;
 389         set_Block_mark(block, 0);
 390         set_Block_phis(block, NULL);
 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 in its block.
 397  */
 398 static void collect_phis(ir_node *node, void *env) {
 399         (void) env;
 400
 401         if (is_Phi(node)) {
 402                 ir_node *block = get_nodes_block(node);
 403
 404                 add_Block_phi(block, node);
 405         } else {
 406                 if (is_no_Block(node) && get_irn_pinned(node) == op_pin_state_pinned) {
 407                         /*
 408                          * Ignore control flow nodes, these will be removed.
 409                          * This ignores Raise. That is surely bad. FIXME.
 410                          */
 411                         if (!is_cfop(node)) {
 412                                 ir_node *block = get_nodes_block(node);
 413
 414                                 DB((dbg, LEVEL_2, "Node %+F in block %+F is unmovable\n", node, block));
 415                                 set_Block_mark(block, 1);
 416                         }
 417                 }
 418         }
 419 }
 420
 421 static void optimise_muxs_0(ir_node* mux, void* env)
 422 {
 423         ir_node* t;
 424         ir_node* f;
 425
 426         (void) env;
 427
 428         if (!is_Mux(mux)) return;
 429
 430         t = get_Mux_true(mux);
 431         f = get_Mux_false(mux);
 432
 433         DB((dbg, LEVEL_3, "Simplify %+F T=%+F F=%+F\n", mux, t, f));
 434
 435         if (is_Unknown(t)) {
 436                 DB((dbg, LEVEL_3, "Replace Mux with unknown operand by %+F\n", f));
 437                 exchange(mux, f);
 438                 return;
 439         }
 440         if (is_Unknown(f)) {
 441                 DB((dbg, LEVEL_3, "Replace Mux with unknown operand by %+F\n", t));
 442                 exchange(mux, t);
 443                 return;
 444         }
 445
 446         if (is_Mux(t)) {
 447                 ir_graph* irg   = current_ir_graph;
 448                 ir_node*  block = get_nodes_block(mux);
 449                 ir_mode*  mode  = get_irn_mode(mux);
 450                 ir_node*  c0    = get_Mux_sel(mux);
 451                 ir_node*  c1    = get_Mux_sel(t);
 452                 ir_node*  t1    = get_Mux_true(t);
 453                 ir_node*  f1    = get_Mux_false(t);
 454                 if (f == f1) {
 455                         /* Mux(c0, Mux(c1, x, y), y) -> typical if (c0 && c1) x else y */
 456                         ir_node* and_    = new_r_And(irg, block, c0, c1, mode_b);
 457                         ir_node* new_mux = new_r_Mux(irg, block, and_, f1, t1, mode);
 458                         exchange(mux, new_mux);
 459                 } else if (f == t1) {
 460                         /* Mux(c0, Mux(c1, x, y), x) */
 461                         ir_node* not_c1 = new_r_Not(irg, block, c1, mode_b);
 462                         ir_node* and_   = new_r_And(irg, block, c0, not_c1, mode_b);
 463                         ir_node* new_mux = new_r_Mux(irg, block, and_, t1, f1, mode);
 464                         exchange(mux, new_mux);
 465                 }
 466         } else if (is_Mux(f)) {
 467                 ir_graph* irg   = current_ir_graph;
 468                 ir_node*  block = get_nodes_block(mux);
 469                 ir_mode*  mode  = get_irn_mode(mux);
 470                 ir_node*  c0    = get_Mux_sel(mux);
 471                 ir_node*  c1    = get_Mux_sel(f);
 472                 ir_node*  t1    = get_Mux_true(f);
 473                 ir_node*  f1    = get_Mux_false(f);
 474                 if (t == t1) {
 475                         /* Mux(c0, x, Mux(c1, x, y)) -> typical if (c0 || c1) x else y */
 476                         ir_node* or_     = new_r_Or(irg, block, c0, c1, mode_b);
 477                         ir_node* new_mux = new_r_Mux(irg, block, or_, f1, t1, mode);
 478                         exchange(mux, new_mux);
 479                 } else if (t == f1) {
 480                         /* Mux(c0, x, Mux(c1, y, x)) */
 481                         ir_node* not_c1  = new_r_Not(irg, block, c1, mode_b);
 482                         ir_node* or_     = new_r_Or(irg, block, c0, not_c1, mode_b);
 483                         ir_node* new_mux = new_r_Mux(irg, block, or_, t1, f1, mode);
 484                         exchange(mux, new_mux);
 485                 }
 486         }
 487 }
 488
 489
 490 static void optimise_muxs_1(ir_node* mux, void* env)
 491 {
 492         ir_node* t;
 493         ir_node* f;
 494         ir_mode* mode;
 495
 496         (void) env;
 497
 498         if (!is_Mux(mux)) return;
 499
 500         t = get_Mux_true(mux);
 501         f = get_Mux_false(mux);
 502
 503         DB((dbg, LEVEL_3, "Simplify %+F T=%+F F=%+F\n", mux, t, f));
 504
 505         mode = get_irn_mode(mux);
 506
 507         if (is_Const(t) && is_Const(f) && (mode_is_int(mode))) {
 508                 ir_node* block = get_nodes_block(mux);
 509                 ir_node* c     = get_Mux_sel(mux);
 510                 tarval* tv_t = get_Const_tarval(t);
 511                 tarval* tv_f = get_Const_tarval(f);
 512                 if (tarval_is_one(tv_t) && tarval_is_null(tv_f)) {
 513                         ir_node* conv  = new_r_Conv(current_ir_graph, block, c, mode);
 514                         exchange(mux, conv);
 515                 } else if (tarval_is_null(tv_t) && tarval_is_one(tv_f)) {
 516                         ir_node* not_  = new_r_Not(current_ir_graph, block, c, mode_b);
 517                         ir_node* conv  = new_r_Conv(current_ir_graph, block, not_, mode);
 518                         exchange(mux, conv);
 519                 }
 520         }
 521 }
 522
 523
 524 void opt_if_conv(ir_graph *irg, const ir_settings_if_conv_t *params)
 525 {
 526         ir_settings_if_conv_t p;
 527
 528         /* get the parameters */
 529         p = (params != NULL ? *params : default_info);
 530
 531         FIRM_DBG_REGISTER(dbg, "firm.opt.ifconv");
 532
 533         DB((dbg, LEVEL_1, "Running if-conversion on %+F\n", irg));
 534
 535         normalize_one_return(irg);
 536         remove_critical_cf_edges(irg);
 537
 538         compute_cdep(irg);
 539         assure_doms(irg);
 540
 541         ir_reserve_resources(irg, IR_RESOURCE_BLOCK_MARK);
 542
 543         irg_block_walk_graph(irg, init_block_link, NULL, NULL);
 544         irg_walk_graph(irg, collect_phis, NULL, NULL);
 545         irg_block_walk_graph(irg, NULL, if_conv_walker, &p);
 546
 547         ir_free_resources(irg, IR_RESOURCE_BLOCK_MARK);
 548
 549         local_optimize_graph(irg);
 550
 551         irg_walk_graph(irg, NULL, optimise_muxs_0, NULL);
 552 #if 1
 553         irg_walk_graph(irg, NULL, optimise_muxs_1, NULL);
 554 #endif
 555
 556         /* TODO: graph might be changed, handle more graceful */
 557         set_irg_outs_inconsistent(irg);
 558         set_irg_extblk_inconsistent(irg);
 559         set_irg_loopinfo_inconsistent(irg);
 560         free_dom(irg);
 561
 562         free_cdep(irg);
 563 }