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  */
  25 #include "config.h"
  26
  27 #include <assert.h>
  28 #include <stdbool.h>
  29
  30 #include "iroptimize.h"
  31 #include "obst.h"
  32 #include "irnode_t.h"
  33 #include "cdep_t.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 #include "irpass_t.h"
  41 #include "be.h"
  42 #include "opt_manage.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  * Returns non-zero if a Block can be emptied.
  59  *
  60  * @param block  the block
  61  */
  62 static bool can_empty_block(ir_node *block)
  63 {
  64         return get_Block_mark(block) == 0;
  65 }
  66
  67 /**
  68  * Find the ProjX node leading from block dependency to block start.
  69  *
  70  * @param start       a block that is control depended on dependency
  71  * @param dependency  the block that decides whether start is executed
  72  *
  73  * @return a ProjX node that represent the decision control flow or
  74  *         NULL is start is not dependent at all or a block on the way
  75  *         cannot be emptied
  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(skip_Proj(pred));
  90
  91                 if (pred_block == dependency) {
  92                         if (is_Proj(pred)) {
  93                                 assert(get_irn_mode(pred) == mode_X);
  94                                 /* we found it */
  95                                 return pred;
  96                         }
  97                         /* Not a Proj? Should not happen. */
  98                         return NULL;
  99                 }
 100
 101                 if (is_Proj(pred)) {
 102                         assert(get_irn_mode(pred) == mode_X);
 103                         /* another Proj but not from the control block */
 104                         return NULL;
 105                 }
 106
 107                 if (is_cdep_on(pred_block, dependency)) {
 108                         return walk_to_projx(pred_block, dependency);
 109                 }
 110         }
 111         return NULL;
 112 }
 113
 114
 115 /**
 116  * Recursively copies the DAG starting at node to the i-th predecessor
 117  * block of src_block
 118  * - if node isn't in the src_block, recursion ends and node is returned
 119  * - if node is a Phi in the src_block, the i-th predecessor of this Phi is
 120  *   returned and recursion ends
 121  * otherwise returns a copy of the passed node created in the i-th predecessor of
 122  * src_block.
 123  *
 124  * @param node       a root of a DAG
 125  * @param src_block  the block of the DAG
 126  * @param i          the position of the predecessor the DAG
 127  *                   is moved to
 128  *
 129  * @return  the root of the copied DAG
 130  */
 131 static ir_node* copy_to(ir_node* node, ir_node* src_block, int i)
 132 {
 133         ir_node* dst_block;
 134         ir_node* copy;
 135         int j;
 136
 137         if (get_nodes_block(node) != src_block) {
 138                 /* already outside src_block, do not copy */
 139                 return node;
 140         }
 141         if (is_Phi(node)) {
 142                 /* move through the Phi to the i-th predecessor */
 143                 return get_irn_n(node, i);
 144         }
 145
 146         /* else really need a copy */
 147         copy = exact_copy(node);
 148         dst_block = get_nodes_block(get_irn_n(src_block, i));
 149         set_nodes_block(copy, dst_block);
 150
 151         DB((dbg, LEVEL_1, "Copying node %+F to block %+F, copy is %+F\n",
 152                 node, dst_block, copy));
 153
 154         /* move recursively all predecessors */
 155         for (j = get_irn_arity(node) - 1; j >= 0; --j) {
 156                 set_irn_n(copy, j, copy_to(get_irn_n(node, j), src_block, i));
 157                 DB((dbg, LEVEL_2, "-- pred %d is %+F\n", j, get_irn_n(copy, j)));
 158         }
 159         return copy;
 160 }
 161
 162
 163 /**
 164  * Remove predecessors i and j (i < j) from a node and
 165  * add an additional predecessor new_pred.
 166  *
 167  * @param node      the node whose inputs are changed
 168  * @param i         the first index to remove
 169  * @param j         the second index to remove
 170  * @param new_pred  a node that is added as a new input to node
 171  */
 172 static void rewire(ir_node* node, int i, int j, ir_node* new_pred)
 173 {
 174         int arity = get_irn_arity(node);
 175         ir_node **ins;
 176         int k;
 177         int l;
 178
 179         NEW_ARR_A(ir_node *, ins, arity - 1);
 180
 181         l = 0;
 182         for (k = 0; k < i;     ++k) ins[l++] = get_irn_n(node, k);
 183         for (++k;   k < j;     ++k) ins[l++] = get_irn_n(node, k);
 184         for (++k;   k < arity; ++k) ins[l++] = get_irn_n(node, k);
 185         ins[l++] = new_pred;
 186         assert(l == arity - 1);
 187         set_irn_in(node, l, ins);
 188 }
 189
 190
 191 /**
 192  * Remove the j-th predecessor from the i-th predecessor of block and add it to block
 193  */
 194 static void split_block(ir_node* block, int i, int j)
 195 {
 196         ir_node  *pred_block = get_nodes_block(get_irn_n(block, i));
 197         int       arity      = get_irn_arity(block);
 198         ir_node **ins        = ALLOCAN(ir_node*, arity + 1);
 199         int       new_pred_arity;
 200         ir_node  *phi;
 201         ir_node  *next;
 202         ir_node **pred_ins;
 203         int       k;
 204
 205         DB((dbg, LEVEL_1, "Splitting predecessor %d of predecessor %d of %+F\n", j, i, block));
 206
 207         for (phi = get_Block_phis(block); phi != NULL; phi = get_Phi_next(phi)) {
 208                 ir_node* copy = copy_to(get_irn_n(phi, i), pred_block, j);
 209
 210                 for (k = 0; k < i; ++k) ins[k] = get_irn_n(phi, k);
 211                 ins[k++] = copy;
 212                 for (; k < arity; ++k) ins[k] = get_irn_n(phi, k);
 213                 ins[k++] = get_irn_n(phi, i);
 214                 set_irn_in(phi, k, ins);
 215         }
 216
 217         for (k = 0; k < i; ++k) ins[k] = get_irn_n(block, k);
 218         ins[k++] = get_irn_n(pred_block, j);
 219         for (; k < arity; ++k) ins[k] = get_irn_n(block, k);
 220         ins[k++] = get_irn_n(block, i);
 221         set_irn_in(block, k, ins);
 222
 223         new_pred_arity = get_irn_arity(pred_block) - 1;
 224         pred_ins       = ALLOCAN(ir_node*, new_pred_arity);
 225
 226         for (phi = get_Block_phis(pred_block); phi != NULL; phi = next) {
 227                 next = get_Phi_next(phi);
 228                 for (k = 0; k != j;              ++k) pred_ins[k] = get_irn_n(phi, k);
 229                 for (;      k != new_pred_arity; ++k) pred_ins[k] = get_irn_n(phi, k + 1);
 230                 if (k == 1) {
 231                         exchange(phi, pred_ins[0]);
 232                 } else {
 233                         set_irn_in(phi, k, pred_ins);
 234                 }
 235         }
 236
 237         for (k = 0; k != j;              ++k) pred_ins[k] = get_irn_n(pred_block, k);
 238         for (;      k != new_pred_arity; ++k) pred_ins[k] = get_irn_n(pred_block, k + 1);
 239         if (k == 1) {
 240                 exchange(pred_block, get_nodes_block(pred_ins[0]));
 241         } else {
 242                 set_irn_in(pred_block, k, pred_ins);
 243         }
 244 }
 245
 246
 247 static void prepare_path(ir_node* block, int i, const ir_node* dependency)
 248 {
 249         ir_node* pred = get_nodes_block(get_irn_n(block, i));
 250         int pred_arity;
 251         int j;
 252
 253         DB((dbg, LEVEL_1, "Preparing predecessor %d of %+F\n", i, block));
 254
 255         pred_arity = get_irn_arity(pred);
 256         for (j = 0; j < pred_arity; ++j) {
 257                 ir_node* pred_pred = get_nodes_block(get_irn_n(pred, j));
 258
 259                 if (pred_pred != dependency && is_cdep_on(pred_pred, dependency)) {
 260                         prepare_path(pred, j, dependency);
 261                         split_block(block, i, j);
 262                         break;
 263                 }
 264         }
 265 }
 266
 267 /**
 268  * Block walker: Search for diamonds and do the if conversion.
 269  */
 270 static void if_conv_walker(ir_node *block, void *ctx)
 271 {
 272         walker_env *env = (walker_env*)ctx;
 273         int arity;
 274         int i;
 275
 276         /* Bail out, if there are no Phis at all */
 277         if (get_Block_phis(block) == NULL) return;
 278
 279 restart:
 280         arity = get_irn_arity(block);
 281         for (i = 0; i < arity; ++i) {
 282                 ir_node* pred0;
 283                 ir_cdep* cdep;
 284
 285                 pred0 = get_Block_cfgpred_block(block, i);
 286                 if (pred0 == block) continue;
 287
 288                 for (cdep = find_cdep(pred0); cdep != NULL; cdep = get_cdep_next(cdep)) {
 289                         const ir_node* dependency = get_cdep_node(cdep);
 290                         ir_node* projx0 = walk_to_projx(pred0, dependency);
 291                         ir_node* cond;
 292                         int j;
 293
 294                         if (projx0 == NULL) continue;
 295
 296                         cond = get_Proj_pred(projx0);
 297                         if (! is_Cond(cond))
 298                                 continue;
 299
 300                         /* We only handle boolean decisions, no switches */
 301                         if (get_irn_mode(get_Cond_selector(cond)) != mode_b) continue;
 302
 303                         for (j = i + 1; j < arity; ++j) {
 304                                 ir_node* projx1;
 305                                 ir_node* sel;
 306                                 ir_node* mux_block;
 307                                 ir_node* phi;
 308                                 ir_node* p;
 309                                 ir_node* pred1;
 310                                 bool     supported;
 311                                 bool     negated;
 312                                 dbg_info* cond_dbg;
 313
 314                                 pred1 = get_Block_cfgpred_block(block, j);
 315                                 if (pred1 == block) continue;
 316
 317                                 if (!is_cdep_on(pred1, dependency)) continue;
 318
 319                                 projx1 = walk_to_projx(pred1, dependency);
 320
 321                                 if (projx1 == NULL) continue;
 322
 323                                 sel = get_Cond_selector(cond);
 324                                 phi = get_Block_phis(block);
 325                                 supported = true;
 326                                 negated   = get_Proj_proj(projx0) == pn_Cond_false;
 327                                 for (p = phi; p != NULL; p = get_Phi_next(p)) {
 328                                         ir_node *mux_false;
 329                                         ir_node *mux_true;
 330                                         if (negated) {
 331                                                 mux_true  = get_Phi_pred(p, j);
 332                                                 mux_false = get_Phi_pred(p, i);
 333                                         } else {
 334                                                 mux_true  = get_Phi_pred(p, i);
 335                                                 mux_false = get_Phi_pred(p, j);
 336                                         }
 337                                         if (!env->allow_ifconv(sel, mux_false, mux_true)) {
 338                                                 supported = false;
 339                                                 break;
 340                                         }
 341                                 }
 342                                 if (!supported)
 343                                         continue;
 344
 345                                 DB((dbg, LEVEL_1, "Found Cond %+F with proj %+F and %+F\n",
 346                                         cond, projx0, projx1
 347                                 ));
 348
 349                                 /* remove critical edges */
 350                                 env->changed = true;
 351                                 prepare_path(block, i, dependency);
 352                                 prepare_path(block, j, dependency);
 353                                 arity = get_irn_arity(block);
 354
 355                                 mux_block = get_nodes_block(cond);
 356                                 cond_dbg = get_irn_dbg_info(cond);
 357                                 do { /* generate Mux nodes in mux_block for Phis in block */
 358                                         ir_node* val_i = get_irn_n(phi, i);
 359                                         ir_node* val_j = get_irn_n(phi, j);
 360                                         ir_node* mux;
 361                                         ir_node* next_phi;
 362
 363                                         if (val_i == val_j) {
 364                                                 mux = val_i;
 365                                                 DB((dbg, LEVEL_2,  "Generating no Mux, because both values are equal\n"));
 366                                         } else {
 367                                                 ir_node *t, *f;
 368
 369                                                 /* Something is very fishy if two predecessors of a PhiM point into
 370                                                  * one block, but not at the same memory node
 371                                                  */
 372                                                 assert(get_irn_mode(phi) != mode_M);
 373                                                 if (negated) {
 374                                                         t = val_j;
 375                                                         f = val_i;
 376                                                 } else {
 377                                                         t = val_i;
 378                                                         f = val_j;
 379                                                 }
 380
 381                                                 mux = new_rd_Mux(cond_dbg, mux_block, sel, f, t, get_irn_mode(phi));
 382                                                 DB((dbg, LEVEL_2, "Generating %+F for %+F\n", mux, phi));
 383                                         }
 384
 385                                         next_phi = get_Phi_next(phi);
 386
 387                                         if (arity == 2) {
 388                                                 exchange(phi, mux);
 389                                         } else {
 390                                                 rewire(phi, i, j, mux);
 391                                         }
 392                                         phi = next_phi;
 393                                 } while (phi != NULL);
 394
 395                                 /* move mux operands into mux_block */
 396                                 exchange(get_nodes_block(get_irn_n(block, i)), mux_block);
 397                                 exchange(get_nodes_block(get_irn_n(block, j)), mux_block);
 398
 399                                 if (arity == 2) {
 400                                         unsigned mark;
 401 #if 0
 402                                         DB((dbg, LEVEL_1,  "Welding block %+F and %+F\n", block, mux_block));
 403                                         /* copy the block-info from the Mux-block to the block before merging */
 404
 405                                         mark =  get_Block_mark(mux_block) | get_Block_mark(block);
 406                                         set_Block_mark(block, mark);
 407                                         set_Block_phis(block, get_Block_phis(mux_block));
 408
 409                                         set_irn_in(block, get_irn_arity(mux_block), get_irn_in(mux_block) + 1);
 410                                         exchange_cdep(mux_block, block);
 411                                         exchange(mux_block, block);
 412 #else
 413                                         DB((dbg, LEVEL_1,  "Welding block %+F to %+F\n", block, mux_block));
 414                                         mark =  get_Block_mark(mux_block) | get_Block_mark(block);
 415                                         /* mark both block just to be sure, should be enough to mark mux_block */
 416                                         set_Block_mark(mux_block, mark);
 417                                         exchange(block, mux_block);
 418 #endif
 419                                         return;
 420                                 } else {
 421                                         rewire(block, i, j, new_r_Jmp(mux_block));
 422                                         goto restart;
 423                                 }
 424                         }
 425                 }
 426         }
 427 }
 428
 429 /**
 430  * Block walker: clear block marks and Phi lists.
 431  */
 432 static void init_block_link(ir_node *block, void *env)
 433 {
 434         (void)env;
 435         set_Block_mark(block, 0);
 436         set_Block_phis(block, NULL);
 437 }
 438
 439
 440 /**
 441  * Daisy-chain all Phis in a block.
 442  * If a non-movable node is encountered set the has_pinned flag in its block.
 443  */
 444 static void collect_phis(ir_node *node, void *env)
 445 {
 446         (void) env;
 447
 448         if (is_Phi(node)) {
 449                 ir_node *block = get_nodes_block(node);
 450
 451                 add_Block_phi(block, node);
 452         } else {
 453                 if (!is_Block(node) && get_irn_pinned(node) == op_pin_state_pinned) {
 454                         /*
 455                          * Ignore control flow nodes (except Raise), these will be removed.
 456                          */
 457                         if (!is_cfop(node) && !is_Raise(node)) {
 458                                 ir_node *block = get_nodes_block(node);
 459
 460                                 DB((dbg, LEVEL_2, "Node %+F in block %+F is unmovable\n", node, block));
 461                                 set_Block_mark(block, 1);
 462                         }
 463                 }
 464         }
 465 }
 466
 467 static ir_graph_state_t do_ifconv(ir_graph *irg)
 468 {
 469         walker_env            env;
 470         const backend_params *be_params = be_get_backend_param();
 471
 472         /* get the parameters */
 473         env.allow_ifconv = be_params->allow_ifconv;
 474         env.changed      = false;
 475
 476         FIRM_DBG_REGISTER(dbg, "firm.opt.ifconv");
 477
 478         DB((dbg, LEVEL_1, "Running if-conversion on %+F\n", irg));
 479
 480         compute_cdep(irg);
 481
 482         ir_reserve_resources(irg, IR_RESOURCE_BLOCK_MARK | IR_RESOURCE_PHI_LIST);
 483
 484         irg_block_walk_graph(irg, init_block_link, NULL, NULL);
 485         irg_walk_graph(irg, collect_phis, NULL, NULL);
 486         irg_block_walk_graph(irg, NULL, if_conv_walker, &env);
 487
 488         ir_free_resources(irg, IR_RESOURCE_BLOCK_MARK | IR_RESOURCE_PHI_LIST);
 489
 490         if (env.changed) {
 491                 local_optimize_graph(irg);
 492         }
 493
 494         free_cdep(irg);
 495
 496         return IR_GRAPH_STATE_NO_CRITICAL_EDGES | IR_GRAPH_STATE_ONE_RETURN;
 497 }
 498
 499 static optdesc_t opt_ifconv = {
 500         "if-conversion",
 501         IR_GRAPH_STATE_NO_CRITICAL_EDGES | IR_GRAPH_STATE_NO_UNREACHABLE_CODE | IR_GRAPH_STATE_NO_BADS | IR_GRAPH_STATE_ONE_RETURN,
 502         do_ifconv,
 503 };
 504
 505 void opt_if_conv(ir_graph *irg)
 506 {
 507         perform_irg_optimization(irg, &opt_ifconv);
 508 }
 509
 510 ir_graph_pass_t *opt_if_conv_pass(const char *name)
 511 {
 512         return def_graph_pass(name ? name : "ifconv", opt_if_conv);
 513 }