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