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 #include "config.h"
  27
  28 #include <assert.h>
  29 #include <stdbool.h>
  30
  31 #include "iroptimize.h"
  32 #include "obst.h"
  33 #include "irnode_t.h"
  34 #include "cdep.h"
  35 #include "ircons.h"
  36 #include "irgmod.h"
  37 #include "irgopt.h"
  38 #include "irgwalk.h"
  39 #include "irtools.h"
  40 #include "array_t.h"
  41 #include "irpass_t.h"
  42 #include "be.h"
  43
  44 #include "irdump.h"
  45 #include "debug.h"
  46
  47 /**
  48  * Environment for if-conversion.
  49  */
  50 typedef struct walker_env {
  51         arch_allow_ifconv_func allow_ifconv;
  52         bool                   changed; /**< Set if the graph was changed. */
  53 } walker_env;
  54
  55 DEBUG_ONLY(static firm_dbg_module_t *dbg);
  56
  57 /**
  58  * Default callback for Mux creation: don't allow any Mux nodes
  59  */
  60 static int default_allow_ifconv(ir_node *sel, ir_node *mux_false,
  61                                 ir_node *mux_true)
  62 {
  63         (void) sel;
  64         (void) mux_false;
  65         (void) mux_true;
  66         return 0;
  67 }
  68
  69 /**
  70  * Returns non-zero if a Block can be emptied.
  71  *
  72  * @param block  the block
  73  */
  74 static bool can_empty_block(ir_node *block)
  75 {
  76         return get_Block_mark(block) == 0;
  77 }
  78
  79 /**
  80  * Find the ProjX node leading from block dependency to block start.
  81  *
  82  * @param start       a block that is control depended on dependency
  83  * @param dependency  the block that decides whether start is executed
  84  *
  85  * @return a ProjX node that represent the decision control flow or
  86  *         NULL is start is not dependent at all or a block on the way
  87  *         cannot be emptied
  88  */
  89 static ir_node* walk_to_projx(ir_node* start, const ir_node* dependency)
  90 {
  91         int arity;
  92         int i;
  93
  94         /* No need to find the conditional block if this block cannot be emptied and
  95          * therefore not moved */
  96         if (!can_empty_block(start)) return NULL;
  97
  98         arity = get_irn_arity(start);
  99         for (i = 0; i < arity; ++i) {
 100                 ir_node* pred = get_irn_n(start, i);
 101                 ir_node* pred_block = get_nodes_block(skip_Proj(pred));
 102
 103                 if (pred_block == dependency) {
 104                         if (is_Proj(pred)) {
 105                                 assert(get_irn_mode(pred) == mode_X);
 106                                 /* we found it */
 107                                 return pred;
 108                         }
 109                         /* Not a Proj? Should not happen. */
 110                         return NULL;
 111                 }
 112
 113                 if (is_Proj(pred)) {
 114                         assert(get_irn_mode(pred) == mode_X);
 115                         /* another Proj but not from the control block */
 116                         return NULL;
 117                 }
 118
 119                 if (is_cdep_on(pred_block, dependency)) {
 120                         return walk_to_projx(pred_block, dependency);
 121                 }
 122         }
 123         return NULL;
 124 }
 125
 126
 127 /**
 128  * Recursively copies the DAG starting at node to the i-th predecessor
 129  * block of src_block
 130  * - if node isn't in the src_block, recursion ends and node is returned
 131  * - if node is a Phi in the src_block, the i-th predecessor of this Phi is
 132  *   returned and recursion ends
 133  * otherwise returns a copy of the passed node created in the i-th predecessor of
 134  * src_block.
 135  *
 136  * @param node       a root of a DAG
 137  * @param src_block  the block of the DAG
 138  * @param i          the position of the predecessor the DAG
 139  *                   is moved to
 140  *
 141  * @return  the root of the copied DAG
 142  */
 143 static ir_node* copy_to(ir_node* node, ir_node* src_block, int i)
 144 {
 145         ir_node* dst_block;
 146         ir_node* copy;
 147         int j;
 148
 149         if (get_nodes_block(node) != src_block) {
 150                 /* already outside src_block, do not copy */
 151                 return node;
 152         }
 153         if (is_Phi(node)) {
 154                 /* move through the Phi to the i-th predecessor */
 155                 return get_irn_n(node, i);
 156         }
 157
 158         /* else really need a copy */
 159         copy = exact_copy(node);
 160         dst_block = get_nodes_block(get_irn_n(src_block, i));
 161         set_nodes_block(copy, dst_block);
 162
 163         DB((dbg, LEVEL_1, "Copying node %+F to block %+F, copy is %+F\n",
 164                 node, dst_block, copy));
 165
 166         /* move recursively all predecessors */
 167         for (j = get_irn_arity(node) - 1; j >= 0; --j) {
 168                 set_irn_n(copy, j, copy_to(get_irn_n(node, j), src_block, i));
 169                 DB((dbg, LEVEL_2, "-- pred %d is %+F\n", j, get_irn_n(copy, j)));
 170         }
 171         return copy;
 172 }
 173
 174
 175 /**
 176  * Remove predecessors i and j (i < j) from a node and
 177  * add an additional predecessor new_pred.
 178  *
 179  * @param node      the node whose inputs are changed
 180  * @param i         the first index to remove
 181  * @param j         the second index to remove
 182  * @param new_pred  a node that is added as a new input to node
 183  */
 184 static void rewire(ir_node* node, int i, int j, ir_node* new_pred)
 185 {
 186         int arity = get_irn_arity(node);
 187         ir_node **ins;
 188         int k;
 189         int l;
 190
 191         NEW_ARR_A(ir_node *, ins, arity - 1);
 192
 193         l = 0;
 194         for (k = 0; k < i;     ++k) ins[l++] = get_irn_n(node, k);
 195         for (++k;   k < j;     ++k) ins[l++] = get_irn_n(node, k);
 196         for (++k;   k < arity; ++k) ins[l++] = get_irn_n(node, k);
 197         ins[l++] = new_pred;
 198         assert(l == arity - 1);
 199         set_irn_in(node, l, ins);
 200 }
 201
 202
 203 /**
 204  * Remove the j-th predecessors from the i-th predecessor of block and add it to block
 205  */
 206 static void split_block(ir_node* block, int i, int j)
 207 {
 208         ir_node* pred_block = get_nodes_block(get_irn_n(block, i));
 209         int arity = get_irn_arity(block);
 210         int new_pred_arity;
 211         ir_node *phi, *next;
 212         ir_node **ins;
 213         ir_node **pred_ins;
 214         int k;
 215
 216         DB((dbg, LEVEL_1, "Splitting predecessor %d of predecessor %d of %+F\n", j, i, block));
 217
 218         NEW_ARR_A(ir_node*, ins, arity + 1);
 219
 220         for (phi = get_Block_phis(block); phi != NULL; phi = get_Phi_next(phi)) {
 221                 ir_node* copy = copy_to(get_irn_n(phi, i), pred_block, j);
 222
 223                 for (k = 0; k < i; ++k) ins[k] = get_irn_n(phi, k);
 224                 ins[k++] = copy;
 225                 for (; k < arity; ++k) ins[k] = get_irn_n(phi, k);
 226                 ins[k] = get_irn_n(phi, i);
 227                 assert(k == arity);
 228                 set_irn_in(phi, arity + 1, ins);
 229         }
 230
 231         for (k = 0; k < i; ++k) ins[k] = get_irn_n(block, k);
 232         ins[k++] = get_irn_n(pred_block, j);
 233         for (; k < arity; ++k) ins[k] = get_irn_n(block, k);
 234         ins[k] = get_irn_n(block, i);
 235         assert(k == arity);
 236         set_irn_in(block, arity + 1, ins);
 237
 238         new_pred_arity = get_irn_arity(pred_block) - 1;
 239         NEW_ARR_A(ir_node*, pred_ins, new_pred_arity);
 240
 241         for (phi = get_Block_phis(pred_block); phi != NULL; phi = next) {
 242                 for (k = 0; k < j; ++k) pred_ins[k] = get_irn_n(phi, k);
 243                 for (; k < new_pred_arity; ++k) pred_ins[k] = get_irn_n(phi, k + 1);
 244                 assert(k == new_pred_arity);
 245                 next = get_Phi_next(phi);
 246                 if (new_pred_arity > 1) {
 247                         set_irn_in(phi, new_pred_arity, pred_ins);
 248                 } else {
 249                         exchange(phi, pred_ins[0]);
 250                 }
 251         }
 252
 253         for (k = 0; k < j; ++k) pred_ins[k] = get_irn_n(pred_block, k);
 254         for (; k < new_pred_arity; ++k) pred_ins[k] = get_irn_n(pred_block, k + 1);
 255         assert(k == new_pred_arity);
 256         if (new_pred_arity > 1) {
 257                 set_irn_in(pred_block, new_pred_arity, pred_ins);
 258         } else {
 259                 exchange(pred_block, get_nodes_block(pred_ins[0]));
 260         }
 261 }
 262
 263
 264 static void prepare_path(ir_node* block, int i, const ir_node* dependency)
 265 {
 266         ir_node* pred = get_nodes_block(get_irn_n(block, i));
 267         int pred_arity;
 268         int j;
 269
 270         DB((dbg, LEVEL_1, "Preparing predecessor %d of %+F\n", i, block));
 271
 272         pred_arity = get_irn_arity(pred);
 273         for (j = 0; j < pred_arity; ++j) {
 274                 ir_node* pred_pred = get_nodes_block(get_irn_n(pred, j));
 275
 276                 if (is_cdep_on(pred_pred, dependency)) {
 277                         prepare_path(pred, j, dependency);
 278                         split_block(block, i, j);
 279                         break;
 280                 }
 281         }
 282 }
 283
 284 /**
 285  * Block walker: Search for diamonds and do the if conversion.
 286  */
 287 static void if_conv_walker(ir_node *block, void *ctx)
 288 {
 289         walker_env *env = ctx;
 290         int arity;
 291         int i;
 292
 293         /* Bail out, if there are no Phis at all */
 294         if (get_Block_phis(block) == NULL) return;
 295
 296 restart:
 297         arity = get_irn_arity(block);
 298         for (i = 0; i < arity; ++i) {
 299                 ir_node* pred0;
 300                 ir_cdep* cdep;
 301
 302                 pred0 = get_Block_cfgpred_block(block, i);
 303                 for (cdep = find_cdep(pred0); cdep != NULL; cdep = cdep->next) {
 304                         const ir_node* dependency = cdep->node;
 305                         ir_node* projx0 = walk_to_projx(pred0, dependency);
 306                         ir_node* cond;
 307                         int j;
 308
 309                         if (projx0 == NULL) continue;
 310
 311                         cond = get_Proj_pred(projx0);
 312                         if (! is_Cond(cond))
 313                                 continue;
 314
 315                         /* We only handle boolean decisions, no switches */
 316                         if (get_irn_mode(get_Cond_selector(cond)) != mode_b) continue;
 317
 318                         for (j = i + 1; j < arity; ++j) {
 319                                 ir_node* projx1;
 320                                 ir_node* sel;
 321                                 ir_node* mux_block;
 322                                 ir_node* phi;
 323                                 ir_node* p;
 324                                 ir_node* pred1;
 325                                 bool     supported;
 326                                 bool     negated;
 327                                 dbg_info* cond_dbg;
 328
 329                                 pred1 = get_Block_cfgpred_block(block, j);
 330
 331                                 if (!is_cdep_on(pred1, dependency)) continue;
 332
 333                                 projx1 = walk_to_projx(pred1, dependency);
 334
 335                                 if (projx1 == NULL) continue;
 336
 337                                 sel = get_Cond_selector(cond);
 338                                 phi = get_Block_phis(block);
 339                                 supported = true;
 340                                 negated   = get_Proj_proj(projx0) == pn_Cond_false;
 341                                 for (p = phi; p != NULL; p = get_Phi_next(p)) {
 342                                         ir_node *mux_false;
 343                                         ir_node *mux_true;
 344                                         if (negated) {
 345                                                 mux_true  = get_Phi_pred(p, j);
 346                                                 mux_false = get_Phi_pred(p, i);
 347                                         } else {
 348                                                 mux_true  = get_Phi_pred(p, i);
 349                                                 mux_false = get_Phi_pred(p, j);
 350                                         }
 351                                         if (!env->allow_ifconv(sel, mux_false, mux_true)) {
 352                                                 supported = false;
 353                                                 break;
 354                                         }
 355                                 }
 356                                 if (!supported)
 357                                         continue;
 358
 359                                 DB((dbg, LEVEL_1, "Found Cond %+F with proj %+F and %+F\n",
 360                                         cond, projx0, projx1
 361                                 ));
 362
 363                                 env->changed = true;
 364                                 prepare_path(block, i, dependency);
 365                                 prepare_path(block, j, dependency);
 366                                 arity = get_irn_arity(block);
 367
 368                                 mux_block = get_nodes_block(cond);
 369                                 cond_dbg = get_irn_dbg_info(cond);
 370                                 do {
 371                                         ir_node* val_i = get_irn_n(phi, i);
 372                                         ir_node* val_j = get_irn_n(phi, j);
 373                                         ir_node* mux;
 374                                         ir_node* next_phi;
 375
 376                                         if (val_i == val_j) {
 377                                                 mux = val_i;
 378                                                 DB((dbg, LEVEL_2,  "Generating no Mux, because both values are equal\n"));
 379                                         } else {
 380                                                 ir_node *t, *f;
 381
 382                                                 /* Something is very fishy if two predecessors of a PhiM point into
 383                                                  * one block, but not at the same memory node
 384                                                  */
 385                                                 assert(get_irn_mode(phi) != mode_M);
 386                                                 if (negated) {
 387                                                         t = val_j;
 388                                                         f = val_i;
 389                                                 } else {
 390                                                         t = val_i;
 391                                                         f = val_j;
 392                                                 }
 393
 394                                                 mux = new_rd_Mux(cond_dbg, mux_block, sel, f, t, get_irn_mode(phi));
 395                                                 DB((dbg, LEVEL_2, "Generating %+F for %+F\n", mux, phi));
 396                                         }
 397
 398                                         next_phi = get_Phi_next(phi);
 399
 400                                         if (arity == 2) {
 401                                                 exchange(phi, mux);
 402                                         } else {
 403                                                 rewire(phi, i, j, mux);
 404                                         }
 405                                         phi = next_phi;
 406                                 } while (phi != NULL);
 407
 408                                 exchange(get_nodes_block(get_irn_n(block, i)), mux_block);
 409                                 exchange(get_nodes_block(get_irn_n(block, j)), mux_block);
 410
 411                                 if (arity == 2) {
 412                                         unsigned mark;
 413 #if 1
 414                                         DB((dbg, LEVEL_1,  "Welding block %+F and %+F\n", block, mux_block));
 415                                         /* copy the block-info from the Mux-block to the block before merging */
 416
 417                                         mark =  get_Block_mark(mux_block) | get_Block_mark(block);
 418                                         set_Block_mark(block, mark);
 419                                         set_Block_phis(block, get_Block_phis(mux_block));
 420
 421                                         set_irn_in(block, get_irn_arity(mux_block), get_irn_in(mux_block) + 1);
 422                                         exchange_cdep(mux_block, block);
 423                                         exchange(mux_block, block);
 424 #else
 425                                         DB((dbg, LEVEL_1,  "Welding block %+F to %+F\n", block, mux_block));
 426                                         mark =  get_Block_mark(mux_block) | get_Block_mark(block);
 427                                         /* mark both block just to be sure, should be enough to mark mux_block */
 428                                         set_Block_mark(mux_block, mark);
 429                                         exchange(block, mux_block);
 430 #endif
 431                                         return;
 432                                 } else {
 433                                         rewire(block, i, j, new_r_Jmp(mux_block));
 434                                         goto restart;
 435                                 }
 436                         }
 437                 }
 438         }
 439 }
 440
 441 /**
 442  * Block walker: clear block marks and Phi lists.
 443  */
 444 static void init_block_link(ir_node *block, void *env)
 445 {
 446         (void)env;
 447         set_Block_mark(block, 0);
 448         set_Block_phis(block, NULL);
 449 }
 450
 451
 452 /**
 453  * Daisy-chain all Phis in a block.
 454  * If a non-movable node is encountered set the has_pinned flag in its block.
 455  */
 456 static void collect_phis(ir_node *node, void *env)
 457 {
 458         (void) env;
 459
 460         if (is_Phi(node)) {
 461                 ir_node *block = get_nodes_block(node);
 462
 463                 add_Block_phi(block, node);
 464         } else {
 465                 if (!is_Block(node) && get_irn_pinned(node) == op_pin_state_pinned) {
 466                         /*
 467                          * Ignore control flow nodes (except Raise), these will be removed.
 468                          */
 469                         if (!is_cfop(node) && !is_Raise(node)) {
 470                                 ir_node *block = get_nodes_block(node);
 471
 472                                 DB((dbg, LEVEL_2, "Node %+F in block %+F is unmovable\n", node, block));
 473                                 set_Block_mark(block, 1);
 474                         }
 475                 }
 476         }
 477 }
 478
 479 void opt_if_conv(ir_graph *irg)
 480 {
 481         walker_env            env;
 482         const backend_params *be_params = be_get_backend_param();
 483
 484         /* get the parameters */
 485         if (be_params->allow_ifconv != NULL)
 486                 env.allow_ifconv = be_params->allow_ifconv;
 487         else
 488                 env.allow_ifconv = default_allow_ifconv;
 489         env.changed = false;
 490
 491         FIRM_DBG_REGISTER(dbg, "firm.opt.ifconv");
 492
 493         DB((dbg, LEVEL_1, "Running if-conversion on %+F\n", irg));
 494
 495         normalize_one_return(irg);
 496         remove_critical_cf_edges(irg);
 497
 498         compute_cdep(irg);
 499
 500         ir_reserve_resources(irg, IR_RESOURCE_BLOCK_MARK | IR_RESOURCE_PHI_LIST);
 501
 502         irg_block_walk_graph(irg, init_block_link, NULL, NULL);
 503         irg_walk_graph(irg, collect_phis, NULL, NULL);
 504         irg_block_walk_graph(irg, NULL, if_conv_walker, &env);
 505
 506         ir_free_resources(irg, IR_RESOURCE_BLOCK_MARK | IR_RESOURCE_PHI_LIST);
 507
 508         if (env.changed) {
 509                 local_optimize_graph(irg);
 510
 511                 /* graph has changed, invalidate analysis info */
 512                 set_irg_outs_inconsistent(irg);
 513                 set_irg_extblk_inconsistent(irg);
 514                 set_irg_loopinfo_inconsistent(irg);
 515                 set_irg_doms_inconsistent(irg);
 516         }
 517
 518         free_cdep(irg);
 519 }
 520
 521 ir_graph_pass_t *opt_if_conv_pass(const char *name)
 522 {
 523         return def_graph_pass(name ? name : "ifconv", opt_if_conv);
 524 }