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