1 /* SSA-PRE for trees. 2 Copyright (C) 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010 3 Free Software Foundation, Inc. 4 Contributed by Daniel Berlin <dan@dberlin.org> and Steven Bosscher 5 <stevenb@suse.de> 6 7 This file is part of GCC. 8 9 GCC is free software; you can redistribute it and/or modify 10 it under the terms of the GNU General Public License as published by 11 the Free Software Foundation; either version 3, or (at your option) 12 any later version. 13 14 GCC is distributed in the hope that it will be useful, 15 but WITHOUT ANY WARRANTY; without even the implied warranty of 16 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 17 GNU General Public License for more details. 18 19 You should have received a copy of the GNU General Public License 20 along with GCC; see the file COPYING3. If not see 21 <http://www.gnu.org/licenses/>. */ 22 23 #include "config.h" 24 #include "system.h" 25 #include "coretypes.h" 26 #include "tm.h" 27 #include "tree.h" 28 #include "basic-block.h" 29 #include "tree-pretty-print.h" 30 #include "gimple-pretty-print.h" 31 #include "tree-inline.h" 32 #include "tree-flow.h" 33 #include "gimple.h" 34 #include "tree-dump.h" 35 #include "timevar.h" 36 #include "fibheap.h" 37 #include "hashtab.h" 38 #include "tree-iterator.h" 39 #include "alloc-pool.h" 40 #include "obstack.h" 41 #include "tree-pass.h" 42 #include "flags.h" 43 #include "bitmap.h" 44 #include "langhooks.h" 45 #include "cfgloop.h" 46 #include "tree-ssa-sccvn.h" 47 #include "tree-scalar-evolution.h" 48 #include "params.h" 49 #include "dbgcnt.h" 50 51 /* TODO: 52 53 1. Avail sets can be shared by making an avail_find_leader that 54 walks up the dominator tree and looks in those avail sets. 55 This might affect code optimality, it's unclear right now. 56 2. Strength reduction can be performed by anticipating expressions 57 we can repair later on. 58 3. We can do back-substitution or smarter value numbering to catch 59 commutative expressions split up over multiple statements. 60 */ 61 62 /* For ease of terminology, "expression node" in the below refers to 63 every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs 64 represent the actual statement containing the expressions we care about, 65 and we cache the value number by putting it in the expression. */ 66 67 /* Basic algorithm 68 69 First we walk the statements to generate the AVAIL sets, the 70 EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the 71 generation of values/expressions by a given block. We use them 72 when computing the ANTIC sets. The AVAIL sets consist of 73 SSA_NAME's that represent values, so we know what values are 74 available in what blocks. AVAIL is a forward dataflow problem. In 75 SSA, values are never killed, so we don't need a kill set, or a 76 fixpoint iteration, in order to calculate the AVAIL sets. In 77 traditional parlance, AVAIL sets tell us the downsafety of the 78 expressions/values. 79 80 Next, we generate the ANTIC sets. These sets represent the 81 anticipatable expressions. ANTIC is a backwards dataflow 82 problem. An expression is anticipatable in a given block if it could 83 be generated in that block. This means that if we had to perform 84 an insertion in that block, of the value of that expression, we 85 could. Calculating the ANTIC sets requires phi translation of 86 expressions, because the flow goes backwards through phis. We must 87 iterate to a fixpoint of the ANTIC sets, because we have a kill 88 set. Even in SSA form, values are not live over the entire 89 function, only from their definition point onwards. So we have to 90 remove values from the ANTIC set once we go past the definition 91 point of the leaders that make them up. 92 compute_antic/compute_antic_aux performs this computation. 93 94 Third, we perform insertions to make partially redundant 95 expressions fully redundant. 96 97 An expression is partially redundant (excluding partial 98 anticipation) if: 99 100 1. It is AVAIL in some, but not all, of the predecessors of a 101 given block. 102 2. It is ANTIC in all the predecessors. 103 104 In order to make it fully redundant, we insert the expression into 105 the predecessors where it is not available, but is ANTIC. 106 107 For the partial anticipation case, we only perform insertion if it 108 is partially anticipated in some block, and fully available in all 109 of the predecessors. 110 111 insert/insert_aux/do_regular_insertion/do_partial_partial_insertion 112 performs these steps. 113 114 Fourth, we eliminate fully redundant expressions. 115 This is a simple statement walk that replaces redundant 116 calculations with the now available values. */ 117 118 /* Representations of value numbers: 119 120 Value numbers are represented by a representative SSA_NAME. We 121 will create fake SSA_NAME's in situations where we need a 122 representative but do not have one (because it is a complex 123 expression). In order to facilitate storing the value numbers in 124 bitmaps, and keep the number of wasted SSA_NAME's down, we also 125 associate a value_id with each value number, and create full blown 126 ssa_name's only where we actually need them (IE in operands of 127 existing expressions). 128 129 Theoretically you could replace all the value_id's with 130 SSA_NAME_VERSION, but this would allocate a large number of 131 SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number. 132 It would also require an additional indirection at each point we 133 use the value id. */ 134 135 /* Representation of expressions on value numbers: 136 137 Expressions consisting of value numbers are represented the same 138 way as our VN internally represents them, with an additional 139 "pre_expr" wrapping around them in order to facilitate storing all 140 of the expressions in the same sets. */ 141 142 /* Representation of sets: 143 144 The dataflow sets do not need to be sorted in any particular order 145 for the majority of their lifetime, are simply represented as two 146 bitmaps, one that keeps track of values present in the set, and one 147 that keeps track of expressions present in the set. 148 149 When we need them in topological order, we produce it on demand by 150 transforming the bitmap into an array and sorting it into topo 151 order. */ 152 153 /* Type of expression, used to know which member of the PRE_EXPR union 154 is valid. */ 155 156 enum pre_expr_kind 157 { 158 NAME, 159 NARY, 160 REFERENCE, 161 CONSTANT 162 }; 163 164 typedef union pre_expr_union_d 165 { 166 tree name; 167 tree constant; 168 vn_nary_op_t nary; 169 vn_reference_t reference; 170 } pre_expr_union; 171 172 typedef struct pre_expr_d 173 { 174 enum pre_expr_kind kind; 175 unsigned int id; 176 pre_expr_union u; 177 } *pre_expr; 178 179 #define PRE_EXPR_NAME(e) (e)->u.name 180 #define PRE_EXPR_NARY(e) (e)->u.nary 181 #define PRE_EXPR_REFERENCE(e) (e)->u.reference 182 #define PRE_EXPR_CONSTANT(e) (e)->u.constant 183 184 static int 185 pre_expr_eq (const void *p1, const void *p2) 186 { 187 const struct pre_expr_d *e1 = (const struct pre_expr_d *) p1; 188 const struct pre_expr_d *e2 = (const struct pre_expr_d *) p2; 189 190 if (e1->kind != e2->kind) 191 return false; 192 193 switch (e1->kind) 194 { 195 case CONSTANT: 196 return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1), 197 PRE_EXPR_CONSTANT (e2)); 198 case NAME: 199 return PRE_EXPR_NAME (e1) == PRE_EXPR_NAME (e2); 200 case NARY: 201 return vn_nary_op_eq (PRE_EXPR_NARY (e1), PRE_EXPR_NARY (e2)); 202 case REFERENCE: 203 return vn_reference_eq (PRE_EXPR_REFERENCE (e1), 204 PRE_EXPR_REFERENCE (e2)); 205 default: 206 gcc_unreachable (); 207 } 208 } 209 210 static hashval_t 211 pre_expr_hash (const void *p1) 212 { 213 const struct pre_expr_d *e = (const struct pre_expr_d *) p1; 214 switch (e->kind) 215 { 216 case CONSTANT: 217 return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e)); 218 case NAME: 219 return SSA_NAME_VERSION (PRE_EXPR_NAME (e)); 220 case NARY: 221 return PRE_EXPR_NARY (e)->hashcode; 222 case REFERENCE: 223 return PRE_EXPR_REFERENCE (e)->hashcode; 224 default: 225 gcc_unreachable (); 226 } 227 } 228 229 230 /* Next global expression id number. */ 231 static unsigned int next_expression_id; 232 233 /* Mapping from expression to id number we can use in bitmap sets. */ 234 DEF_VEC_P (pre_expr); 235 DEF_VEC_ALLOC_P (pre_expr, heap); 236 static VEC(pre_expr, heap) *expressions; 237 static htab_t expression_to_id; 238 static VEC(unsigned, heap) *name_to_id; 239 240 /* Allocate an expression id for EXPR. */ 241 242 static inline unsigned int 243 alloc_expression_id (pre_expr expr) 244 { 245 void **slot; 246 /* Make sure we won't overflow. */ 247 gcc_assert (next_expression_id + 1 > next_expression_id); 248 expr->id = next_expression_id++; 249 VEC_safe_push (pre_expr, heap, expressions, expr); 250 if (expr->kind == NAME) 251 { 252 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr)); 253 /* VEC_safe_grow_cleared allocates no headroom. Avoid frequent 254 re-allocations by using VEC_reserve upfront. There is no 255 VEC_quick_grow_cleared unfortunately. */ 256 VEC_reserve (unsigned, heap, name_to_id, num_ssa_names); 257 VEC_safe_grow_cleared (unsigned, heap, name_to_id, num_ssa_names); 258 gcc_assert (VEC_index (unsigned, name_to_id, version) == 0); 259 VEC_replace (unsigned, name_to_id, version, expr->id); 260 } 261 else 262 { 263 slot = htab_find_slot (expression_to_id, expr, INSERT); 264 gcc_assert (!*slot); 265 *slot = expr; 266 } 267 return next_expression_id - 1; 268 } 269 270 /* Return the expression id for tree EXPR. */ 271 272 static inline unsigned int 273 get_expression_id (const pre_expr expr) 274 { 275 return expr->id; 276 } 277 278 static inline unsigned int 279 lookup_expression_id (const pre_expr expr) 280 { 281 void **slot; 282 283 if (expr->kind == NAME) 284 { 285 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr)); 286 if (VEC_length (unsigned, name_to_id) <= version) 287 return 0; 288 return VEC_index (unsigned, name_to_id, version); 289 } 290 else 291 { 292 slot = htab_find_slot (expression_to_id, expr, NO_INSERT); 293 if (!slot) 294 return 0; 295 return ((pre_expr)*slot)->id; 296 } 297 } 298 299 /* Return the existing expression id for EXPR, or create one if one 300 does not exist yet. */ 301 302 static inline unsigned int 303 get_or_alloc_expression_id (pre_expr expr) 304 { 305 unsigned int id = lookup_expression_id (expr); 306 if (id == 0) 307 return alloc_expression_id (expr); 308 return expr->id = id; 309 } 310 311 /* Return the expression that has expression id ID */ 312 313 static inline pre_expr 314 expression_for_id (unsigned int id) 315 { 316 return VEC_index (pre_expr, expressions, id); 317 } 318 319 /* Free the expression id field in all of our expressions, 320 and then destroy the expressions array. */ 321 322 static void 323 clear_expression_ids (void) 324 { 325 VEC_free (pre_expr, heap, expressions); 326 } 327 328 static alloc_pool pre_expr_pool; 329 330 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */ 331 332 static pre_expr 333 get_or_alloc_expr_for_name (tree name) 334 { 335 struct pre_expr_d expr; 336 pre_expr result; 337 unsigned int result_id; 338 339 expr.kind = NAME; 340 expr.id = 0; 341 PRE_EXPR_NAME (&expr) = name; 342 result_id = lookup_expression_id (&expr); 343 if (result_id != 0) 344 return expression_for_id (result_id); 345 346 result = (pre_expr) pool_alloc (pre_expr_pool); 347 result->kind = NAME; 348 PRE_EXPR_NAME (result) = name; 349 alloc_expression_id (result); 350 return result; 351 } 352 353 static bool in_fre = false; 354 355 /* An unordered bitmap set. One bitmap tracks values, the other, 356 expressions. */ 357 typedef struct bitmap_set 358 { 359 bitmap_head expressions; 360 bitmap_head values; 361 } *bitmap_set_t; 362 363 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \ 364 EXECUTE_IF_SET_IN_BITMAP(&(set)->expressions, 0, (id), (bi)) 365 366 #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \ 367 EXECUTE_IF_SET_IN_BITMAP(&(set)->values, 0, (id), (bi)) 368 369 /* Mapping from value id to expressions with that value_id. */ 370 DEF_VEC_P (bitmap_set_t); 371 DEF_VEC_ALLOC_P (bitmap_set_t, heap); 372 static VEC(bitmap_set_t, heap) *value_expressions; 373 374 /* Sets that we need to keep track of. */ 375 typedef struct bb_bitmap_sets 376 { 377 /* The EXP_GEN set, which represents expressions/values generated in 378 a basic block. */ 379 bitmap_set_t exp_gen; 380 381 /* The PHI_GEN set, which represents PHI results generated in a 382 basic block. */ 383 bitmap_set_t phi_gen; 384 385 /* The TMP_GEN set, which represents results/temporaries generated 386 in a basic block. IE the LHS of an expression. */ 387 bitmap_set_t tmp_gen; 388 389 /* The AVAIL_OUT set, which represents which values are available in 390 a given basic block. */ 391 bitmap_set_t avail_out; 392 393 /* The ANTIC_IN set, which represents which values are anticipatable 394 in a given basic block. */ 395 bitmap_set_t antic_in; 396 397 /* The PA_IN set, which represents which values are 398 partially anticipatable in a given basic block. */ 399 bitmap_set_t pa_in; 400 401 /* The NEW_SETS set, which is used during insertion to augment the 402 AVAIL_OUT set of blocks with the new insertions performed during 403 the current iteration. */ 404 bitmap_set_t new_sets; 405 406 /* A cache for value_dies_in_block_x. */ 407 bitmap expr_dies; 408 409 /* True if we have visited this block during ANTIC calculation. */ 410 unsigned int visited : 1; 411 412 /* True we have deferred processing this block during ANTIC 413 calculation until its successor is processed. */ 414 unsigned int deferred : 1; 415 416 /* True when the block contains a call that might not return. */ 417 unsigned int contains_may_not_return_call : 1; 418 } *bb_value_sets_t; 419 420 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen 421 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen 422 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen 423 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out 424 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in 425 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in 426 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets 427 #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies 428 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited 429 #define BB_DEFERRED(BB) ((bb_value_sets_t) ((BB)->aux))->deferred 430 #define BB_MAY_NOTRETURN(BB) ((bb_value_sets_t) ((BB)->aux))->contains_may_not_return_call 431 432 433 /* Basic block list in postorder. */ 434 static int *postorder; 435 436 /* This structure is used to keep track of statistics on what 437 optimization PRE was able to perform. */ 438 static struct 439 { 440 /* The number of RHS computations eliminated by PRE. */ 441 int eliminations; 442 443 /* The number of new expressions/temporaries generated by PRE. */ 444 int insertions; 445 446 /* The number of inserts found due to partial anticipation */ 447 int pa_insert; 448 449 /* The number of new PHI nodes added by PRE. */ 450 int phis; 451 } pre_stats; 452 453 static bool do_partial_partial; 454 static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int, gimple); 455 static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr); 456 static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr); 457 static void bitmap_set_copy (bitmap_set_t, bitmap_set_t); 458 static bool bitmap_set_contains_value (bitmap_set_t, unsigned int); 459 static void bitmap_insert_into_set (bitmap_set_t, pre_expr); 460 static void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr, 461 unsigned int, bool); 462 static bitmap_set_t bitmap_set_new (void); 463 static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *, 464 gimple, tree); 465 static tree find_or_generate_expression (basic_block, pre_expr, gimple_seq *, 466 gimple); 467 static unsigned int get_expr_value_id (pre_expr); 468 469 /* We can add and remove elements and entries to and from sets 470 and hash tables, so we use alloc pools for them. */ 471 472 static alloc_pool bitmap_set_pool; 473 static bitmap_obstack grand_bitmap_obstack; 474 475 /* To avoid adding 300 temporary variables when we only need one, we 476 only create one temporary variable, on demand, and build ssa names 477 off that. We do have to change the variable if the types don't 478 match the current variable's type. */ 479 static tree pretemp; 480 static tree storetemp; 481 static tree prephitemp; 482 483 /* Set of blocks with statements that have had their EH properties changed. */ 484 static bitmap need_eh_cleanup; 485 486 /* Set of blocks with statements that have had their AB properties changed. */ 487 static bitmap need_ab_cleanup; 488 489 /* The phi_translate_table caches phi translations for a given 490 expression and predecessor. */ 491 492 static htab_t phi_translate_table; 493 494 /* A three tuple {e, pred, v} used to cache phi translations in the 495 phi_translate_table. */ 496 497 typedef struct expr_pred_trans_d 498 { 499 /* The expression. */ 500 pre_expr e; 501 502 /* The predecessor block along which we translated the expression. */ 503 basic_block pred; 504 505 /* The value that resulted from the translation. */ 506 pre_expr v; 507 508 /* The hashcode for the expression, pred pair. This is cached for 509 speed reasons. */ 510 hashval_t hashcode; 511 } *expr_pred_trans_t; 512 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t; 513 514 /* Return the hash value for a phi translation table entry. */ 515 516 static hashval_t 517 expr_pred_trans_hash (const void *p) 518 { 519 const_expr_pred_trans_t const ve = (const_expr_pred_trans_t) p; 520 return ve->hashcode; 521 } 522 523 /* Return true if two phi translation table entries are the same. 524 P1 and P2 should point to the expr_pred_trans_t's to be compared.*/ 525 526 static int 527 expr_pred_trans_eq (const void *p1, const void *p2) 528 { 529 const_expr_pred_trans_t const ve1 = (const_expr_pred_trans_t) p1; 530 const_expr_pred_trans_t const ve2 = (const_expr_pred_trans_t) p2; 531 basic_block b1 = ve1->pred; 532 basic_block b2 = ve2->pred; 533 534 /* If they are not translations for the same basic block, they can't 535 be equal. */ 536 if (b1 != b2) 537 return false; 538 return pre_expr_eq (ve1->e, ve2->e); 539 } 540 541 /* Search in the phi translation table for the translation of 542 expression E in basic block PRED. 543 Return the translated value, if found, NULL otherwise. */ 544 545 static inline pre_expr 546 phi_trans_lookup (pre_expr e, basic_block pred) 547 { 548 void **slot; 549 struct expr_pred_trans_d ept; 550 551 ept.e = e; 552 ept.pred = pred; 553 ept.hashcode = iterative_hash_hashval_t (pre_expr_hash (e), pred->index); 554 slot = htab_find_slot_with_hash (phi_translate_table, &ept, ept.hashcode, 555 NO_INSERT); 556 if (!slot) 557 return NULL; 558 else 559 return ((expr_pred_trans_t) *slot)->v; 560 } 561 562 563 /* Add the tuple mapping from {expression E, basic block PRED} to 564 value V, to the phi translation table. */ 565 566 static inline void 567 phi_trans_add (pre_expr e, pre_expr v, basic_block pred) 568 { 569 void **slot; 570 expr_pred_trans_t new_pair = XNEW (struct expr_pred_trans_d); 571 new_pair->e = e; 572 new_pair->pred = pred; 573 new_pair->v = v; 574 new_pair->hashcode = iterative_hash_hashval_t (pre_expr_hash (e), 575 pred->index); 576 577 slot = htab_find_slot_with_hash (phi_translate_table, new_pair, 578 new_pair->hashcode, INSERT); 579 free (*slot); 580 *slot = (void *) new_pair; 581 } 582 583 584 /* Add expression E to the expression set of value id V. */ 585 586 void 587 add_to_value (unsigned int v, pre_expr e) 588 { 589 bitmap_set_t set; 590 591 gcc_assert (get_expr_value_id (e) == v); 592 593 if (v >= VEC_length (bitmap_set_t, value_expressions)) 594 { 595 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions, 596 v + 1); 597 } 598 599 set = VEC_index (bitmap_set_t, value_expressions, v); 600 if (!set) 601 { 602 set = bitmap_set_new (); 603 VEC_replace (bitmap_set_t, value_expressions, v, set); 604 } 605 606 bitmap_insert_into_set_1 (set, e, v, true); 607 } 608 609 /* Create a new bitmap set and return it. */ 610 611 static bitmap_set_t 612 bitmap_set_new (void) 613 { 614 bitmap_set_t ret = (bitmap_set_t) pool_alloc (bitmap_set_pool); 615 bitmap_initialize (&ret->expressions, &grand_bitmap_obstack); 616 bitmap_initialize (&ret->values, &grand_bitmap_obstack); 617 return ret; 618 } 619 620 /* Return the value id for a PRE expression EXPR. */ 621 622 static unsigned int 623 get_expr_value_id (pre_expr expr) 624 { 625 switch (expr->kind) 626 { 627 case CONSTANT: 628 { 629 unsigned int id; 630 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr)); 631 if (id == 0) 632 { 633 id = get_or_alloc_constant_value_id (PRE_EXPR_CONSTANT (expr)); 634 add_to_value (id, expr); 635 } 636 return id; 637 } 638 case NAME: 639 return VN_INFO (PRE_EXPR_NAME (expr))->value_id; 640 case NARY: 641 return PRE_EXPR_NARY (expr)->value_id; 642 case REFERENCE: 643 return PRE_EXPR_REFERENCE (expr)->value_id; 644 default: 645 gcc_unreachable (); 646 } 647 } 648 649 /* Remove an expression EXPR from a bitmapped set. */ 650 651 static void 652 bitmap_remove_from_set (bitmap_set_t set, pre_expr expr) 653 { 654 unsigned int val = get_expr_value_id (expr); 655 if (!value_id_constant_p (val)) 656 { 657 bitmap_clear_bit (&set->values, val); 658 bitmap_clear_bit (&set->expressions, get_expression_id (expr)); 659 } 660 } 661 662 static void 663 bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr, 664 unsigned int val, bool allow_constants) 665 { 666 if (allow_constants || !value_id_constant_p (val)) 667 { 668 /* We specifically expect this and only this function to be able to 669 insert constants into a set. */ 670 bitmap_set_bit (&set->values, val); 671 bitmap_set_bit (&set->expressions, get_or_alloc_expression_id (expr)); 672 } 673 } 674 675 /* Insert an expression EXPR into a bitmapped set. */ 676 677 static void 678 bitmap_insert_into_set (bitmap_set_t set, pre_expr expr) 679 { 680 bitmap_insert_into_set_1 (set, expr, get_expr_value_id (expr), false); 681 } 682 683 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */ 684 685 static void 686 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig) 687 { 688 bitmap_copy (&dest->expressions, &orig->expressions); 689 bitmap_copy (&dest->values, &orig->values); 690 } 691 692 693 /* Free memory used up by SET. */ 694 static void 695 bitmap_set_free (bitmap_set_t set) 696 { 697 bitmap_clear (&set->expressions); 698 bitmap_clear (&set->values); 699 } 700 701 702 /* Generate an topological-ordered array of bitmap set SET. */ 703 704 static VEC(pre_expr, heap) * 705 sorted_array_from_bitmap_set (bitmap_set_t set) 706 { 707 unsigned int i, j; 708 bitmap_iterator bi, bj; 709 VEC(pre_expr, heap) *result; 710 711 /* Pre-allocate roughly enough space for the array. */ 712 result = VEC_alloc (pre_expr, heap, bitmap_count_bits (&set->values)); 713 714 FOR_EACH_VALUE_ID_IN_SET (set, i, bi) 715 { 716 /* The number of expressions having a given value is usually 717 relatively small. Thus, rather than making a vector of all 718 the expressions and sorting it by value-id, we walk the values 719 and check in the reverse mapping that tells us what expressions 720 have a given value, to filter those in our set. As a result, 721 the expressions are inserted in value-id order, which means 722 topological order. 723 724 If this is somehow a significant lose for some cases, we can 725 choose which set to walk based on the set size. */ 726 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, i); 727 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bj) 728 { 729 if (bitmap_bit_p (&set->expressions, j)) 730 VEC_safe_push (pre_expr, heap, result, expression_for_id (j)); 731 } 732 } 733 734 return result; 735 } 736 737 /* Perform bitmapped set operation DEST &= ORIG. */ 738 739 static void 740 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig) 741 { 742 bitmap_iterator bi; 743 unsigned int i; 744 745 if (dest != orig) 746 { 747 bitmap_head temp; 748 bitmap_initialize (&temp, &grand_bitmap_obstack); 749 750 bitmap_and_into (&dest->values, &orig->values); 751 bitmap_copy (&temp, &dest->expressions); 752 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi) 753 { 754 pre_expr expr = expression_for_id (i); 755 unsigned int value_id = get_expr_value_id (expr); 756 if (!bitmap_bit_p (&dest->values, value_id)) 757 bitmap_clear_bit (&dest->expressions, i); 758 } 759 bitmap_clear (&temp); 760 } 761 } 762 763 /* Subtract all values and expressions contained in ORIG from DEST. */ 764 765 static bitmap_set_t 766 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig) 767 { 768 bitmap_set_t result = bitmap_set_new (); 769 bitmap_iterator bi; 770 unsigned int i; 771 772 bitmap_and_compl (&result->expressions, &dest->expressions, 773 &orig->expressions); 774 775 FOR_EACH_EXPR_ID_IN_SET (result, i, bi) 776 { 777 pre_expr expr = expression_for_id (i); 778 unsigned int value_id = get_expr_value_id (expr); 779 bitmap_set_bit (&result->values, value_id); 780 } 781 782 return result; 783 } 784 785 /* Subtract all the values in bitmap set B from bitmap set A. */ 786 787 static void 788 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b) 789 { 790 unsigned int i; 791 bitmap_iterator bi; 792 bitmap_head temp; 793 794 bitmap_initialize (&temp, &grand_bitmap_obstack); 795 796 bitmap_copy (&temp, &a->expressions); 797 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi) 798 { 799 pre_expr expr = expression_for_id (i); 800 if (bitmap_set_contains_value (b, get_expr_value_id (expr))) 801 bitmap_remove_from_set (a, expr); 802 } 803 bitmap_clear (&temp); 804 } 805 806 807 /* Return true if bitmapped set SET contains the value VALUE_ID. */ 808 809 static bool 810 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id) 811 { 812 if (value_id_constant_p (value_id)) 813 return true; 814 815 if (!set || bitmap_empty_p (&set->expressions)) 816 return false; 817 818 return bitmap_bit_p (&set->values, value_id); 819 } 820 821 static inline bool 822 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr) 823 { 824 return bitmap_bit_p (&set->expressions, get_expression_id (expr)); 825 } 826 827 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */ 828 829 static void 830 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor, 831 const pre_expr expr) 832 { 833 bitmap_set_t exprset; 834 unsigned int i; 835 bitmap_iterator bi; 836 837 if (value_id_constant_p (lookfor)) 838 return; 839 840 if (!bitmap_set_contains_value (set, lookfor)) 841 return; 842 843 /* The number of expressions having a given value is usually 844 significantly less than the total number of expressions in SET. 845 Thus, rather than check, for each expression in SET, whether it 846 has the value LOOKFOR, we walk the reverse mapping that tells us 847 what expressions have a given value, and see if any of those 848 expressions are in our set. For large testcases, this is about 849 5-10x faster than walking the bitmap. If this is somehow a 850 significant lose for some cases, we can choose which set to walk 851 based on the set size. */ 852 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor); 853 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi) 854 { 855 if (bitmap_clear_bit (&set->expressions, i)) 856 { 857 bitmap_set_bit (&set->expressions, get_expression_id (expr)); 858 return; 859 } 860 } 861 862 gcc_unreachable (); 863 } 864 865 /* Return true if two bitmap sets are equal. */ 866 867 static bool 868 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b) 869 { 870 return bitmap_equal_p (&a->values, &b->values); 871 } 872 873 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists, 874 and add it otherwise. */ 875 876 static void 877 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr) 878 { 879 unsigned int val = get_expr_value_id (expr); 880 881 if (bitmap_set_contains_value (set, val)) 882 bitmap_set_replace_value (set, val, expr); 883 else 884 bitmap_insert_into_set (set, expr); 885 } 886 887 /* Insert EXPR into SET if EXPR's value is not already present in 888 SET. */ 889 890 static void 891 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr) 892 { 893 unsigned int val = get_expr_value_id (expr); 894 895 gcc_checking_assert (expr->id == get_or_alloc_expression_id (expr)); 896 897 /* Constant values are always considered to be part of the set. */ 898 if (value_id_constant_p (val)) 899 return; 900 901 /* If the value membership changed, add the expression. */ 902 if (bitmap_set_bit (&set->values, val)) 903 bitmap_set_bit (&set->expressions, expr->id); 904 } 905 906 /* Print out EXPR to outfile. */ 907 908 static void 909 print_pre_expr (FILE *outfile, const pre_expr expr) 910 { 911 switch (expr->kind) 912 { 913 case CONSTANT: 914 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0); 915 break; 916 case NAME: 917 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0); 918 break; 919 case NARY: 920 { 921 unsigned int i; 922 vn_nary_op_t nary = PRE_EXPR_NARY (expr); 923 fprintf (outfile, "{%s,", tree_code_name [nary->opcode]); 924 for (i = 0; i < nary->length; i++) 925 { 926 print_generic_expr (outfile, nary->op[i], 0); 927 if (i != (unsigned) nary->length - 1) 928 fprintf (outfile, ","); 929 } 930 fprintf (outfile, "}"); 931 } 932 break; 933 934 case REFERENCE: 935 { 936 vn_reference_op_t vro; 937 unsigned int i; 938 vn_reference_t ref = PRE_EXPR_REFERENCE (expr); 939 fprintf (outfile, "{"); 940 for (i = 0; 941 VEC_iterate (vn_reference_op_s, ref->operands, i, vro); 942 i++) 943 { 944 bool closebrace = false; 945 if (vro->opcode != SSA_NAME 946 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration) 947 { 948 fprintf (outfile, "%s", tree_code_name [vro->opcode]); 949 if (vro->op0) 950 { 951 fprintf (outfile, "<"); 952 closebrace = true; 953 } 954 } 955 if (vro->op0) 956 { 957 print_generic_expr (outfile, vro->op0, 0); 958 if (vro->op1) 959 { 960 fprintf (outfile, ","); 961 print_generic_expr (outfile, vro->op1, 0); 962 } 963 if (vro->op2) 964 { 965 fprintf (outfile, ","); 966 print_generic_expr (outfile, vro->op2, 0); 967 } 968 } 969 if (closebrace) 970 fprintf (outfile, ">"); 971 if (i != VEC_length (vn_reference_op_s, ref->operands) - 1) 972 fprintf (outfile, ","); 973 } 974 fprintf (outfile, "}"); 975 if (ref->vuse) 976 { 977 fprintf (outfile, "@"); 978 print_generic_expr (outfile, ref->vuse, 0); 979 } 980 } 981 break; 982 } 983 } 984 void debug_pre_expr (pre_expr); 985 986 /* Like print_pre_expr but always prints to stderr. */ 987 DEBUG_FUNCTION void 988 debug_pre_expr (pre_expr e) 989 { 990 print_pre_expr (stderr, e); 991 fprintf (stderr, "\n"); 992 } 993 994 /* Print out SET to OUTFILE. */ 995 996 static void 997 print_bitmap_set (FILE *outfile, bitmap_set_t set, 998 const char *setname, int blockindex) 999 { 1000 fprintf (outfile, "%s[%d] := { ", setname, blockindex); 1001 if (set) 1002 { 1003 bool first = true; 1004 unsigned i; 1005 bitmap_iterator bi; 1006 1007 FOR_EACH_EXPR_ID_IN_SET (set, i, bi) 1008 { 1009 const pre_expr expr = expression_for_id (i); 1010 1011 if (!first) 1012 fprintf (outfile, ", "); 1013 first = false; 1014 print_pre_expr (outfile, expr); 1015 1016 fprintf (outfile, " (%04d)", get_expr_value_id (expr)); 1017 } 1018 } 1019 fprintf (outfile, " }\n"); 1020 } 1021 1022 void debug_bitmap_set (bitmap_set_t); 1023 1024 DEBUG_FUNCTION void 1025 debug_bitmap_set (bitmap_set_t set) 1026 { 1027 print_bitmap_set (stderr, set, "debug", 0); 1028 } 1029 1030 /* Print out the expressions that have VAL to OUTFILE. */ 1031 1032 void 1033 print_value_expressions (FILE *outfile, unsigned int val) 1034 { 1035 bitmap_set_t set = VEC_index (bitmap_set_t, value_expressions, val); 1036 if (set) 1037 { 1038 char s[10]; 1039 sprintf (s, "%04d", val); 1040 print_bitmap_set (outfile, set, s, 0); 1041 } 1042 } 1043 1044 1045 DEBUG_FUNCTION void 1046 debug_value_expressions (unsigned int val) 1047 { 1048 print_value_expressions (stderr, val); 1049 } 1050 1051 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to 1052 represent it. */ 1053 1054 static pre_expr 1055 get_or_alloc_expr_for_constant (tree constant) 1056 { 1057 unsigned int result_id; 1058 unsigned int value_id; 1059 struct pre_expr_d expr; 1060 pre_expr newexpr; 1061 1062 expr.kind = CONSTANT; 1063 PRE_EXPR_CONSTANT (&expr) = constant; 1064 result_id = lookup_expression_id (&expr); 1065 if (result_id != 0) 1066 return expression_for_id (result_id); 1067 1068 newexpr = (pre_expr) pool_alloc (pre_expr_pool); 1069 newexpr->kind = CONSTANT; 1070 PRE_EXPR_CONSTANT (newexpr) = constant; 1071 alloc_expression_id (newexpr); 1072 value_id = get_or_alloc_constant_value_id (constant); 1073 add_to_value (value_id, newexpr); 1074 return newexpr; 1075 } 1076 1077 /* Given a value id V, find the actual tree representing the constant 1078 value if there is one, and return it. Return NULL if we can't find 1079 a constant. */ 1080 1081 static tree 1082 get_constant_for_value_id (unsigned int v) 1083 { 1084 if (value_id_constant_p (v)) 1085 { 1086 unsigned int i; 1087 bitmap_iterator bi; 1088 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, v); 1089 1090 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi) 1091 { 1092 pre_expr expr = expression_for_id (i); 1093 if (expr->kind == CONSTANT) 1094 return PRE_EXPR_CONSTANT (expr); 1095 } 1096 } 1097 return NULL; 1098 } 1099 1100 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it. 1101 Currently only supports constants and SSA_NAMES. */ 1102 static pre_expr 1103 get_or_alloc_expr_for (tree t) 1104 { 1105 if (TREE_CODE (t) == SSA_NAME) 1106 return get_or_alloc_expr_for_name (t); 1107 else if (is_gimple_min_invariant (t)) 1108 return get_or_alloc_expr_for_constant (t); 1109 else 1110 { 1111 /* More complex expressions can result from SCCVN expression 1112 simplification that inserts values for them. As they all 1113 do not have VOPs the get handled by the nary ops struct. */ 1114 vn_nary_op_t result; 1115 unsigned int result_id; 1116 vn_nary_op_lookup (t, &result); 1117 if (result != NULL) 1118 { 1119 pre_expr e = (pre_expr) pool_alloc (pre_expr_pool); 1120 e->kind = NARY; 1121 PRE_EXPR_NARY (e) = result; 1122 result_id = lookup_expression_id (e); 1123 if (result_id != 0) 1124 { 1125 pool_free (pre_expr_pool, e); 1126 e = expression_for_id (result_id); 1127 return e; 1128 } 1129 alloc_expression_id (e); 1130 return e; 1131 } 1132 } 1133 return NULL; 1134 } 1135 1136 /* Return the folded version of T if T, when folded, is a gimple 1137 min_invariant. Otherwise, return T. */ 1138 1139 static pre_expr 1140 fully_constant_expression (pre_expr e) 1141 { 1142 switch (e->kind) 1143 { 1144 case CONSTANT: 1145 return e; 1146 case NARY: 1147 { 1148 vn_nary_op_t nary = PRE_EXPR_NARY (e); 1149 switch (TREE_CODE_CLASS (nary->opcode)) 1150 { 1151 case tcc_binary: 1152 case tcc_comparison: 1153 { 1154 /* We have to go from trees to pre exprs to value ids to 1155 constants. */ 1156 tree naryop0 = nary->op[0]; 1157 tree naryop1 = nary->op[1]; 1158 tree result; 1159 if (!is_gimple_min_invariant (naryop0)) 1160 { 1161 pre_expr rep0 = get_or_alloc_expr_for (naryop0); 1162 unsigned int vrep0 = get_expr_value_id (rep0); 1163 tree const0 = get_constant_for_value_id (vrep0); 1164 if (const0) 1165 naryop0 = fold_convert (TREE_TYPE (naryop0), const0); 1166 } 1167 if (!is_gimple_min_invariant (naryop1)) 1168 { 1169 pre_expr rep1 = get_or_alloc_expr_for (naryop1); 1170 unsigned int vrep1 = get_expr_value_id (rep1); 1171 tree const1 = get_constant_for_value_id (vrep1); 1172 if (const1) 1173 naryop1 = fold_convert (TREE_TYPE (naryop1), const1); 1174 } 1175 result = fold_binary (nary->opcode, nary->type, 1176 naryop0, naryop1); 1177 if (result && is_gimple_min_invariant (result)) 1178 return get_or_alloc_expr_for_constant (result); 1179 /* We might have simplified the expression to a 1180 SSA_NAME for example from x_1 * 1. But we cannot 1181 insert a PHI for x_1 unconditionally as x_1 might 1182 not be available readily. */ 1183 return e; 1184 } 1185 case tcc_reference: 1186 if (nary->opcode != REALPART_EXPR 1187 && nary->opcode != IMAGPART_EXPR 1188 && nary->opcode != VIEW_CONVERT_EXPR) 1189 return e; 1190 /* Fallthrough. */ 1191 case tcc_unary: 1192 { 1193 /* We have to go from trees to pre exprs to value ids to 1194 constants. */ 1195 tree naryop0 = nary->op[0]; 1196 tree const0, result; 1197 if (is_gimple_min_invariant (naryop0)) 1198 const0 = naryop0; 1199 else 1200 { 1201 pre_expr rep0 = get_or_alloc_expr_for (naryop0); 1202 unsigned int vrep0 = get_expr_value_id (rep0); 1203 const0 = get_constant_for_value_id (vrep0); 1204 } 1205 result = NULL; 1206 if (const0) 1207 { 1208 tree type1 = TREE_TYPE (nary->op[0]); 1209 const0 = fold_convert (type1, const0); 1210 result = fold_unary (nary->opcode, nary->type, const0); 1211 } 1212 if (result && is_gimple_min_invariant (result)) 1213 return get_or_alloc_expr_for_constant (result); 1214 return e; 1215 } 1216 default: 1217 return e; 1218 } 1219 } 1220 case REFERENCE: 1221 { 1222 vn_reference_t ref = PRE_EXPR_REFERENCE (e); 1223 tree folded; 1224 if ((folded = fully_constant_vn_reference_p (ref))) 1225 return get_or_alloc_expr_for_constant (folded); 1226 return e; 1227 } 1228 default: 1229 return e; 1230 } 1231 return e; 1232 } 1233 1234 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that 1235 it has the value it would have in BLOCK. Set *SAME_VALID to true 1236 in case the new vuse doesn't change the value id of the OPERANDS. */ 1237 1238 static tree 1239 translate_vuse_through_block (VEC (vn_reference_op_s, heap) *operands, 1240 alias_set_type set, tree type, tree vuse, 1241 basic_block phiblock, 1242 basic_block block, bool *same_valid) 1243 { 1244 gimple phi = SSA_NAME_DEF_STMT (vuse); 1245 ao_ref ref; 1246 edge e = NULL; 1247 bool use_oracle; 1248 1249 *same_valid = true; 1250 1251 if (gimple_bb (phi) != phiblock) 1252 return vuse; 1253 1254 use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands); 1255 1256 /* Use the alias-oracle to find either the PHI node in this block, 1257 the first VUSE used in this block that is equivalent to vuse or 1258 the first VUSE which definition in this block kills the value. */ 1259 if (gimple_code (phi) == GIMPLE_PHI) 1260 e = find_edge (block, phiblock); 1261 else if (use_oracle) 1262 while (!stmt_may_clobber_ref_p_1 (phi, &ref)) 1263 { 1264 vuse = gimple_vuse (phi); 1265 phi = SSA_NAME_DEF_STMT (vuse); 1266 if (gimple_bb (phi) != phiblock) 1267 return vuse; 1268 if (gimple_code (phi) == GIMPLE_PHI) 1269 { 1270 e = find_edge (block, phiblock); 1271 break; 1272 } 1273 } 1274 else 1275 return NULL_TREE; 1276 1277 if (e) 1278 { 1279 if (use_oracle) 1280 { 1281 bitmap visited = NULL; 1282 /* Try to find a vuse that dominates this phi node by skipping 1283 non-clobbering statements. */ 1284 vuse = get_continuation_for_phi (phi, &ref, &visited, false); 1285 if (visited) 1286 BITMAP_FREE (visited); 1287 } 1288 else 1289 vuse = NULL_TREE; 1290 if (!vuse) 1291 { 1292 /* If we didn't find any, the value ID can't stay the same, 1293 but return the translated vuse. */ 1294 *same_valid = false; 1295 vuse = PHI_ARG_DEF (phi, e->dest_idx); 1296 } 1297 /* ??? We would like to return vuse here as this is the canonical 1298 upmost vdef that this reference is associated with. But during 1299 insertion of the references into the hash tables we only ever 1300 directly insert with their direct gimple_vuse, hence returning 1301 something else would make us not find the other expression. */ 1302 return PHI_ARG_DEF (phi, e->dest_idx); 1303 } 1304 1305 return NULL_TREE; 1306 } 1307 1308 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or* 1309 SET2. This is used to avoid making a set consisting of the union 1310 of PA_IN and ANTIC_IN during insert. */ 1311 1312 static inline pre_expr 1313 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2) 1314 { 1315 pre_expr result; 1316 1317 result = bitmap_find_leader (set1, val, NULL); 1318 if (!result && set2) 1319 result = bitmap_find_leader (set2, val, NULL); 1320 return result; 1321 } 1322 1323 /* Get the tree type for our PRE expression e. */ 1324 1325 static tree 1326 get_expr_type (const pre_expr e) 1327 { 1328 switch (e->kind) 1329 { 1330 case NAME: 1331 return TREE_TYPE (PRE_EXPR_NAME (e)); 1332 case CONSTANT: 1333 return TREE_TYPE (PRE_EXPR_CONSTANT (e)); 1334 case REFERENCE: 1335 return PRE_EXPR_REFERENCE (e)->type; 1336 case NARY: 1337 return PRE_EXPR_NARY (e)->type; 1338 } 1339 gcc_unreachable(); 1340 } 1341 1342 /* Get a representative SSA_NAME for a given expression. 1343 Since all of our sub-expressions are treated as values, we require 1344 them to be SSA_NAME's for simplicity. 1345 Prior versions of GVNPRE used to use "value handles" here, so that 1346 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In 1347 either case, the operands are really values (IE we do not expect 1348 them to be usable without finding leaders). */ 1349 1350 static tree 1351 get_representative_for (const pre_expr e) 1352 { 1353 tree exprtype; 1354 tree name; 1355 unsigned int value_id = get_expr_value_id (e); 1356 1357 switch (e->kind) 1358 { 1359 case NAME: 1360 return PRE_EXPR_NAME (e); 1361 case CONSTANT: 1362 return PRE_EXPR_CONSTANT (e); 1363 case NARY: 1364 case REFERENCE: 1365 { 1366 /* Go through all of the expressions representing this value 1367 and pick out an SSA_NAME. */ 1368 unsigned int i; 1369 bitmap_iterator bi; 1370 bitmap_set_t exprs = VEC_index (bitmap_set_t, value_expressions, 1371 value_id); 1372 FOR_EACH_EXPR_ID_IN_SET (exprs, i, bi) 1373 { 1374 pre_expr rep = expression_for_id (i); 1375 if (rep->kind == NAME) 1376 return PRE_EXPR_NAME (rep); 1377 } 1378 } 1379 break; 1380 } 1381 /* If we reached here we couldn't find an SSA_NAME. This can 1382 happen when we've discovered a value that has never appeared in 1383 the program as set to an SSA_NAME, most likely as the result of 1384 phi translation. */ 1385 if (dump_file) 1386 { 1387 fprintf (dump_file, 1388 "Could not find SSA_NAME representative for expression:"); 1389 print_pre_expr (dump_file, e); 1390 fprintf (dump_file, "\n"); 1391 } 1392 1393 exprtype = get_expr_type (e); 1394 1395 /* Build and insert the assignment of the end result to the temporary 1396 that we will return. */ 1397 if (!pretemp || exprtype != TREE_TYPE (pretemp)) 1398 { 1399 pretemp = create_tmp_reg (exprtype, "pretmp"); 1400 add_referenced_var (pretemp); 1401 } 1402 1403 name = make_ssa_name (pretemp, gimple_build_nop ()); 1404 VN_INFO_GET (name)->value_id = value_id; 1405 if (e->kind == CONSTANT) 1406 VN_INFO (name)->valnum = PRE_EXPR_CONSTANT (e); 1407 else 1408 VN_INFO (name)->valnum = name; 1409 1410 add_to_value (value_id, get_or_alloc_expr_for_name (name)); 1411 if (dump_file) 1412 { 1413 fprintf (dump_file, "Created SSA_NAME representative "); 1414 print_generic_expr (dump_file, name, 0); 1415 fprintf (dump_file, " for expression:"); 1416 print_pre_expr (dump_file, e); 1417 fprintf (dump_file, "\n"); 1418 } 1419 1420 return name; 1421 } 1422 1423 1424 1425 static pre_expr 1426 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2, 1427 basic_block pred, basic_block phiblock); 1428 1429 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of 1430 the phis in PRED. Return NULL if we can't find a leader for each part 1431 of the translated expression. */ 1432 1433 static pre_expr 1434 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2, 1435 basic_block pred, basic_block phiblock) 1436 { 1437 switch (expr->kind) 1438 { 1439 case NARY: 1440 { 1441 unsigned int i; 1442 bool changed = false; 1443 vn_nary_op_t nary = PRE_EXPR_NARY (expr); 1444 vn_nary_op_t newnary = XALLOCAVAR (struct vn_nary_op_s, 1445 sizeof_vn_nary_op (nary->length)); 1446 memcpy (newnary, nary, sizeof_vn_nary_op (nary->length)); 1447 1448 for (i = 0; i < newnary->length; i++) 1449 { 1450 if (TREE_CODE (newnary->op[i]) != SSA_NAME) 1451 continue; 1452 else 1453 { 1454 pre_expr leader, result; 1455 unsigned int op_val_id = VN_INFO (newnary->op[i])->value_id; 1456 leader = find_leader_in_sets (op_val_id, set1, set2); 1457 result = phi_translate (leader, set1, set2, pred, phiblock); 1458 if (result && result != leader) 1459 { 1460 tree name = get_representative_for (result); 1461 if (!name) 1462 return NULL; 1463 newnary->op[i] = name; 1464 } 1465 else if (!result) 1466 return NULL; 1467 1468 changed |= newnary->op[i] != nary->op[i]; 1469 } 1470 } 1471 if (changed) 1472 { 1473 pre_expr constant; 1474 unsigned int new_val_id; 1475 1476 tree result = vn_nary_op_lookup_pieces (newnary->length, 1477 newnary->opcode, 1478 newnary->type, 1479 &newnary->op[0], 1480 &nary); 1481 if (result && is_gimple_min_invariant (result)) 1482 return get_or_alloc_expr_for_constant (result); 1483 1484 expr = (pre_expr) pool_alloc (pre_expr_pool); 1485 expr->kind = NARY; 1486 expr->id = 0; 1487 if (nary) 1488 { 1489 PRE_EXPR_NARY (expr) = nary; 1490 constant = fully_constant_expression (expr); 1491 if (constant != expr) 1492 return constant; 1493 1494 new_val_id = nary->value_id; 1495 get_or_alloc_expression_id (expr); 1496 } 1497 else 1498 { 1499 new_val_id = get_next_value_id (); 1500 VEC_safe_grow_cleared (bitmap_set_t, heap, 1501 value_expressions, 1502 get_max_value_id() + 1); 1503 nary = vn_nary_op_insert_pieces (newnary->length, 1504 newnary->opcode, 1505 newnary->type, 1506 &newnary->op[0], 1507 result, new_val_id); 1508 PRE_EXPR_NARY (expr) = nary; 1509 constant = fully_constant_expression (expr); 1510 if (constant != expr) 1511 return constant; 1512 get_or_alloc_expression_id (expr); 1513 } 1514 add_to_value (new_val_id, expr); 1515 } 1516 return expr; 1517 } 1518 break; 1519 1520 case REFERENCE: 1521 { 1522 vn_reference_t ref = PRE_EXPR_REFERENCE (expr); 1523 VEC (vn_reference_op_s, heap) *operands = ref->operands; 1524 tree vuse = ref->vuse; 1525 tree newvuse = vuse; 1526 VEC (vn_reference_op_s, heap) *newoperands = NULL; 1527 bool changed = false, same_valid = true; 1528 unsigned int i, j, n; 1529 vn_reference_op_t operand; 1530 vn_reference_t newref; 1531 1532 for (i = 0, j = 0; 1533 VEC_iterate (vn_reference_op_s, operands, i, operand); i++, j++) 1534 { 1535 pre_expr opresult; 1536 pre_expr leader; 1537 tree op[3]; 1538 tree type = operand->type; 1539 vn_reference_op_s newop = *operand; 1540 op[0] = operand->op0; 1541 op[1] = operand->op1; 1542 op[2] = operand->op2; 1543 for (n = 0; n < 3; ++n) 1544 { 1545 unsigned int op_val_id; 1546 if (!op[n]) 1547 continue; 1548 if (TREE_CODE (op[n]) != SSA_NAME) 1549 { 1550 /* We can't possibly insert these. */ 1551 if (n != 0 1552 && !is_gimple_min_invariant (op[n])) 1553 break; 1554 continue; 1555 } 1556 op_val_id = VN_INFO (op[n])->value_id; 1557 leader = find_leader_in_sets (op_val_id, set1, set2); 1558 if (!leader) 1559 break; 1560 /* Make sure we do not recursively translate ourselves 1561 like for translating a[n_1] with the leader for 1562 n_1 being a[n_1]. */ 1563 if (get_expression_id (leader) != get_expression_id (expr)) 1564 { 1565 opresult = phi_translate (leader, set1, set2, 1566 pred, phiblock); 1567 if (!opresult) 1568 break; 1569 if (opresult != leader) 1570 { 1571 tree name = get_representative_for (opresult); 1572 if (!name) 1573 break; 1574 changed |= name != op[n]; 1575 op[n] = name; 1576 } 1577 } 1578 } 1579 if (n != 3) 1580 { 1581 if (newoperands) 1582 VEC_free (vn_reference_op_s, heap, newoperands); 1583 return NULL; 1584 } 1585 if (!newoperands) 1586 newoperands = VEC_copy (vn_reference_op_s, heap, operands); 1587 /* We may have changed from an SSA_NAME to a constant */ 1588 if (newop.opcode == SSA_NAME && TREE_CODE (op[0]) != SSA_NAME) 1589 newop.opcode = TREE_CODE (op[0]); 1590 newop.type = type; 1591 newop.op0 = op[0]; 1592 newop.op1 = op[1]; 1593 newop.op2 = op[2]; 1594 /* If it transforms a non-constant ARRAY_REF into a constant 1595 one, adjust the constant offset. */ 1596 if (newop.opcode == ARRAY_REF 1597 && newop.off == -1 1598 && TREE_CODE (op[0]) == INTEGER_CST 1599 && TREE_CODE (op[1]) == INTEGER_CST 1600 && TREE_CODE (op[2]) == INTEGER_CST) 1601 { 1602 double_int off = tree_to_double_int (op[0]); 1603 off = double_int_add (off, 1604 double_int_neg 1605 (tree_to_double_int (op[1]))); 1606 off = double_int_mul (off, tree_to_double_int (op[2])); 1607 if (double_int_fits_in_shwi_p (off)) 1608 newop.off = off.low; 1609 } 1610 VEC_replace (vn_reference_op_s, newoperands, j, &newop); 1611 /* If it transforms from an SSA_NAME to an address, fold with 1612 a preceding indirect reference. */ 1613 if (j > 0 && op[0] && TREE_CODE (op[0]) == ADDR_EXPR 1614 && VEC_index (vn_reference_op_s, 1615 newoperands, j - 1)->opcode == MEM_REF) 1616 vn_reference_fold_indirect (&newoperands, &j); 1617 } 1618 if (i != VEC_length (vn_reference_op_s, operands)) 1619 { 1620 if (newoperands) 1621 VEC_free (vn_reference_op_s, heap, newoperands); 1622 return NULL; 1623 } 1624 1625 if (vuse) 1626 { 1627 newvuse = translate_vuse_through_block (newoperands, 1628 ref->set, ref->type, 1629 vuse, phiblock, pred, 1630 &same_valid); 1631 if (newvuse == NULL_TREE) 1632 { 1633 VEC_free (vn_reference_op_s, heap, newoperands); 1634 return NULL; 1635 } 1636 } 1637 1638 if (changed || newvuse != vuse) 1639 { 1640 unsigned int new_val_id; 1641 pre_expr constant; 1642 bool converted = false; 1643 1644 tree result = vn_reference_lookup_pieces (newvuse, ref->set, 1645 ref->type, 1646 newoperands, 1647 &newref, VN_WALK); 1648 if (result) 1649 VEC_free (vn_reference_op_s, heap, newoperands); 1650 1651 if (result 1652 && !useless_type_conversion_p (ref->type, TREE_TYPE (result))) 1653 { 1654 result = fold_build1 (VIEW_CONVERT_EXPR, ref->type, result); 1655 converted = true; 1656 } 1657 else if (!result && newref 1658 && !useless_type_conversion_p (ref->type, newref->type)) 1659 { 1660 VEC_free (vn_reference_op_s, heap, newoperands); 1661 return NULL; 1662 } 1663 1664 if (result && is_gimple_min_invariant (result)) 1665 { 1666 gcc_assert (!newoperands); 1667 return get_or_alloc_expr_for_constant (result); 1668 } 1669 1670 expr = (pre_expr) pool_alloc (pre_expr_pool); 1671 expr->kind = REFERENCE; 1672 expr->id = 0; 1673 1674 if (converted) 1675 { 1676 vn_nary_op_t nary; 1677 tree nresult; 1678 1679 gcc_assert (CONVERT_EXPR_P (result) 1680 || TREE_CODE (result) == VIEW_CONVERT_EXPR); 1681 1682 nresult = vn_nary_op_lookup_pieces (1, TREE_CODE (result), 1683 TREE_TYPE (result), 1684 &TREE_OPERAND (result, 0), 1685 &nary); 1686 if (nresult && is_gimple_min_invariant (nresult)) 1687 return get_or_alloc_expr_for_constant (nresult); 1688 1689 expr->kind = NARY; 1690 if (nary) 1691 { 1692 PRE_EXPR_NARY (expr) = nary; 1693 constant = fully_constant_expression (expr); 1694 if (constant != expr) 1695 return constant; 1696 1697 new_val_id = nary->value_id; 1698 get_or_alloc_expression_id (expr); 1699 } 1700 else 1701 { 1702 new_val_id = get_next_value_id (); 1703 VEC_safe_grow_cleared (bitmap_set_t, heap, 1704 value_expressions, 1705 get_max_value_id() + 1); 1706 nary = vn_nary_op_insert_pieces (1, TREE_CODE (result), 1707 TREE_TYPE (result), 1708 &TREE_OPERAND (result, 0), 1709 NULL_TREE, 1710 new_val_id); 1711 PRE_EXPR_NARY (expr) = nary; 1712 constant = fully_constant_expression (expr); 1713 if (constant != expr) 1714 return constant; 1715 get_or_alloc_expression_id (expr); 1716 } 1717 } 1718 else if (newref) 1719 { 1720 PRE_EXPR_REFERENCE (expr) = newref; 1721 constant = fully_constant_expression (expr); 1722 if (constant != expr) 1723 return constant; 1724 1725 new_val_id = newref->value_id; 1726 get_or_alloc_expression_id (expr); 1727 } 1728 else 1729 { 1730 if (changed || !same_valid) 1731 { 1732 new_val_id = get_next_value_id (); 1733 VEC_safe_grow_cleared (bitmap_set_t, heap, 1734 value_expressions, 1735 get_max_value_id() + 1); 1736 } 1737 else 1738 new_val_id = ref->value_id; 1739 newref = vn_reference_insert_pieces (newvuse, ref->set, 1740 ref->type, 1741 newoperands, 1742 result, new_val_id); 1743 newoperands = NULL; 1744 PRE_EXPR_REFERENCE (expr) = newref; 1745 constant = fully_constant_expression (expr); 1746 if (constant != expr) 1747 return constant; 1748 get_or_alloc_expression_id (expr); 1749 } 1750 add_to_value (new_val_id, expr); 1751 } 1752 VEC_free (vn_reference_op_s, heap, newoperands); 1753 return expr; 1754 } 1755 break; 1756 1757 case NAME: 1758 { 1759 gimple phi = NULL; 1760 edge e; 1761 gimple def_stmt; 1762 tree name = PRE_EXPR_NAME (expr); 1763 1764 def_stmt = SSA_NAME_DEF_STMT (name); 1765 if (gimple_code (def_stmt) == GIMPLE_PHI 1766 && gimple_bb (def_stmt) == phiblock) 1767 phi = def_stmt; 1768 else 1769 return expr; 1770 1771 e = find_edge (pred, gimple_bb (phi)); 1772 if (e) 1773 { 1774 tree def = PHI_ARG_DEF (phi, e->dest_idx); 1775 pre_expr newexpr; 1776 1777 if (TREE_CODE (def) == SSA_NAME) 1778 def = VN_INFO (def)->valnum; 1779 1780 /* Handle constant. */ 1781 if (is_gimple_min_invariant (def)) 1782 return get_or_alloc_expr_for_constant (def); 1783 1784 if (TREE_CODE (def) == SSA_NAME && ssa_undefined_value_p (def)) 1785 return NULL; 1786 1787 newexpr = get_or_alloc_expr_for_name (def); 1788 return newexpr; 1789 } 1790 } 1791 return expr; 1792 1793 default: 1794 gcc_unreachable (); 1795 } 1796 } 1797 1798 /* Wrapper around phi_translate_1 providing caching functionality. */ 1799 1800 static pre_expr 1801 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2, 1802 basic_block pred, basic_block phiblock) 1803 { 1804 pre_expr phitrans; 1805 1806 if (!expr) 1807 return NULL; 1808 1809 /* Constants contain no values that need translation. */ 1810 if (expr->kind == CONSTANT) 1811 return expr; 1812 1813 if (value_id_constant_p (get_expr_value_id (expr))) 1814 return expr; 1815 1816 if (expr->kind != NAME) 1817 { 1818 phitrans = phi_trans_lookup (expr, pred); 1819 if (phitrans) 1820 return phitrans; 1821 } 1822 1823 /* Translate. */ 1824 phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock); 1825 1826 /* Don't add empty translations to the cache. Neither add 1827 translations of NAMEs as those are cheap to translate. */ 1828 if (phitrans 1829 && expr->kind != NAME) 1830 phi_trans_add (expr, phitrans, pred); 1831 1832 return phitrans; 1833 } 1834 1835 1836 /* For each expression in SET, translate the values through phi nodes 1837 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting 1838 expressions in DEST. */ 1839 1840 static void 1841 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred, 1842 basic_block phiblock) 1843 { 1844 VEC (pre_expr, heap) *exprs; 1845 pre_expr expr; 1846 int i; 1847 1848 if (gimple_seq_empty_p (phi_nodes (phiblock))) 1849 { 1850 bitmap_set_copy (dest, set); 1851 return; 1852 } 1853 1854 exprs = sorted_array_from_bitmap_set (set); 1855 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr) 1856 { 1857 pre_expr translated; 1858 translated = phi_translate (expr, set, NULL, pred, phiblock); 1859 if (!translated) 1860 continue; 1861 1862 /* We might end up with multiple expressions from SET being 1863 translated to the same value. In this case we do not want 1864 to retain the NARY or REFERENCE expression but prefer a NAME 1865 which would be the leader. */ 1866 if (translated->kind == NAME) 1867 bitmap_value_replace_in_set (dest, translated); 1868 else 1869 bitmap_value_insert_into_set (dest, translated); 1870 } 1871 VEC_free (pre_expr, heap, exprs); 1872 } 1873 1874 /* Find the leader for a value (i.e., the name representing that 1875 value) in a given set, and return it. If STMT is non-NULL it 1876 makes sure the defining statement for the leader dominates it. 1877 Return NULL if no leader is found. */ 1878 1879 static pre_expr 1880 bitmap_find_leader (bitmap_set_t set, unsigned int val, gimple stmt) 1881 { 1882 if (value_id_constant_p (val)) 1883 { 1884 unsigned int i; 1885 bitmap_iterator bi; 1886 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val); 1887 1888 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi) 1889 { 1890 pre_expr expr = expression_for_id (i); 1891 if (expr->kind == CONSTANT) 1892 return expr; 1893 } 1894 } 1895 if (bitmap_set_contains_value (set, val)) 1896 { 1897 /* Rather than walk the entire bitmap of expressions, and see 1898 whether any of them has the value we are looking for, we look 1899 at the reverse mapping, which tells us the set of expressions 1900 that have a given value (IE value->expressions with that 1901 value) and see if any of those expressions are in our set. 1902 The number of expressions per value is usually significantly 1903 less than the number of expressions in the set. In fact, for 1904 large testcases, doing it this way is roughly 5-10x faster 1905 than walking the bitmap. 1906 If this is somehow a significant lose for some cases, we can 1907 choose which set to walk based on which set is smaller. */ 1908 unsigned int i; 1909 bitmap_iterator bi; 1910 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val); 1911 1912 EXECUTE_IF_AND_IN_BITMAP (&exprset->expressions, 1913 &set->expressions, 0, i, bi) 1914 { 1915 pre_expr val = expression_for_id (i); 1916 /* At the point where stmt is not null, there should always 1917 be an SSA_NAME first in the list of expressions. */ 1918 if (stmt) 1919 { 1920 gimple def_stmt = SSA_NAME_DEF_STMT (PRE_EXPR_NAME (val)); 1921 if (gimple_code (def_stmt) != GIMPLE_PHI 1922 && gimple_bb (def_stmt) == gimple_bb (stmt) 1923 /* PRE insertions are at the end of the basic-block 1924 and have UID 0. */ 1925 && (gimple_uid (def_stmt) == 0 1926 || gimple_uid (def_stmt) >= gimple_uid (stmt))) 1927 continue; 1928 } 1929 return val; 1930 } 1931 } 1932 return NULL; 1933 } 1934 1935 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of 1936 BLOCK by seeing if it is not killed in the block. Note that we are 1937 only determining whether there is a store that kills it. Because 1938 of the order in which clean iterates over values, we are guaranteed 1939 that altered operands will have caused us to be eliminated from the 1940 ANTIC_IN set already. */ 1941 1942 static bool 1943 value_dies_in_block_x (pre_expr expr, basic_block block) 1944 { 1945 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse; 1946 vn_reference_t refx = PRE_EXPR_REFERENCE (expr); 1947 gimple def; 1948 gimple_stmt_iterator gsi; 1949 unsigned id = get_expression_id (expr); 1950 bool res = false; 1951 ao_ref ref; 1952 1953 if (!vuse) 1954 return false; 1955 1956 /* Lookup a previously calculated result. */ 1957 if (EXPR_DIES (block) 1958 && bitmap_bit_p (EXPR_DIES (block), id * 2)) 1959 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1); 1960 1961 /* A memory expression {e, VUSE} dies in the block if there is a 1962 statement that may clobber e. If, starting statement walk from the 1963 top of the basic block, a statement uses VUSE there can be no kill 1964 inbetween that use and the original statement that loaded {e, VUSE}, 1965 so we can stop walking. */ 1966 ref.base = NULL_TREE; 1967 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi)) 1968 { 1969 tree def_vuse, def_vdef; 1970 def = gsi_stmt (gsi); 1971 def_vuse = gimple_vuse (def); 1972 def_vdef = gimple_vdef (def); 1973 1974 /* Not a memory statement. */ 1975 if (!def_vuse) 1976 continue; 1977 1978 /* Not a may-def. */ 1979 if (!def_vdef) 1980 { 1981 /* A load with the same VUSE, we're done. */ 1982 if (def_vuse == vuse) 1983 break; 1984 1985 continue; 1986 } 1987 1988 /* Init ref only if we really need it. */ 1989 if (ref.base == NULL_TREE 1990 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type, 1991 refx->operands)) 1992 { 1993 res = true; 1994 break; 1995 } 1996 /* If the statement may clobber expr, it dies. */ 1997 if (stmt_may_clobber_ref_p_1 (def, &ref)) 1998 { 1999 res = true; 2000 break; 2001 } 2002 } 2003 2004 /* Remember the result. */ 2005 if (!EXPR_DIES (block)) 2006 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack); 2007 bitmap_set_bit (EXPR_DIES (block), id * 2); 2008 if (res) 2009 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1); 2010 2011 return res; 2012 } 2013 2014 2015 #define union_contains_value(SET1, SET2, VAL) \ 2016 (bitmap_set_contains_value ((SET1), (VAL)) \ 2017 || ((SET2) && bitmap_set_contains_value ((SET2), (VAL)))) 2018 2019 /* Determine if vn_reference_op_t VRO is legal in SET1 U SET2. 2020 */ 2021 static bool 2022 vro_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, 2023 vn_reference_op_t vro) 2024 { 2025 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME) 2026 { 2027 struct pre_expr_d temp; 2028 temp.kind = NAME; 2029 temp.id = 0; 2030 PRE_EXPR_NAME (&temp) = vro->op0; 2031 temp.id = lookup_expression_id (&temp); 2032 if (temp.id == 0) 2033 return false; 2034 if (!union_contains_value (set1, set2, 2035 get_expr_value_id (&temp))) 2036 return false; 2037 } 2038 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME) 2039 { 2040 struct pre_expr_d temp; 2041 temp.kind = NAME; 2042 temp.id = 0; 2043 PRE_EXPR_NAME (&temp) = vro->op1; 2044 temp.id = lookup_expression_id (&temp); 2045 if (temp.id == 0) 2046 return false; 2047 if (!union_contains_value (set1, set2, 2048 get_expr_value_id (&temp))) 2049 return false; 2050 } 2051 2052 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME) 2053 { 2054 struct pre_expr_d temp; 2055 temp.kind = NAME; 2056 temp.id = 0; 2057 PRE_EXPR_NAME (&temp) = vro->op2; 2058 temp.id = lookup_expression_id (&temp); 2059 if (temp.id == 0) 2060 return false; 2061 if (!union_contains_value (set1, set2, 2062 get_expr_value_id (&temp))) 2063 return false; 2064 } 2065 2066 return true; 2067 } 2068 2069 /* Determine if the expression EXPR is valid in SET1 U SET2. 2070 ONLY SET2 CAN BE NULL. 2071 This means that we have a leader for each part of the expression 2072 (if it consists of values), or the expression is an SSA_NAME. 2073 For loads/calls, we also see if the vuse is killed in this block. */ 2074 2075 static bool 2076 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr, 2077 basic_block block) 2078 { 2079 switch (expr->kind) 2080 { 2081 case NAME: 2082 return bitmap_set_contains_expr (AVAIL_OUT (block), expr); 2083 case NARY: 2084 { 2085 unsigned int i; 2086 vn_nary_op_t nary = PRE_EXPR_NARY (expr); 2087 for (i = 0; i < nary->length; i++) 2088 { 2089 if (TREE_CODE (nary->op[i]) == SSA_NAME) 2090 { 2091 struct pre_expr_d temp; 2092 temp.kind = NAME; 2093 temp.id = 0; 2094 PRE_EXPR_NAME (&temp) = nary->op[i]; 2095 temp.id = lookup_expression_id (&temp); 2096 if (temp.id == 0) 2097 return false; 2098 if (!union_contains_value (set1, set2, 2099 get_expr_value_id (&temp))) 2100 return false; 2101 } 2102 } 2103 /* If the NARY may trap make sure the block does not contain 2104 a possible exit point. 2105 ??? This is overly conservative if we translate AVAIL_OUT 2106 as the available expression might be after the exit point. */ 2107 if (BB_MAY_NOTRETURN (block) 2108 && vn_nary_may_trap (nary)) 2109 return false; 2110 return true; 2111 } 2112 break; 2113 case REFERENCE: 2114 { 2115 vn_reference_t ref = PRE_EXPR_REFERENCE (expr); 2116 vn_reference_op_t vro; 2117 unsigned int i; 2118 2119 FOR_EACH_VEC_ELT (vn_reference_op_s, ref->operands, i, vro) 2120 { 2121 if (!vro_valid_in_sets (set1, set2, vro)) 2122 return false; 2123 } 2124 if (ref->vuse) 2125 { 2126 gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse); 2127 if (!gimple_nop_p (def_stmt) 2128 && gimple_bb (def_stmt) != block 2129 && !dominated_by_p (CDI_DOMINATORS, 2130 block, gimple_bb (def_stmt))) 2131 return false; 2132 } 2133 return !value_dies_in_block_x (expr, block); 2134 } 2135 default: 2136 gcc_unreachable (); 2137 } 2138 } 2139 2140 /* Clean the set of expressions that are no longer valid in SET1 or 2141 SET2. This means expressions that are made up of values we have no 2142 leaders for in SET1 or SET2. This version is used for partial 2143 anticipation, which means it is not valid in either ANTIC_IN or 2144 PA_IN. */ 2145 2146 static void 2147 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block) 2148 { 2149 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set1); 2150 pre_expr expr; 2151 int i; 2152 2153 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr) 2154 { 2155 if (!valid_in_sets (set1, set2, expr, block)) 2156 bitmap_remove_from_set (set1, expr); 2157 } 2158 VEC_free (pre_expr, heap, exprs); 2159 } 2160 2161 /* Clean the set of expressions that are no longer valid in SET. This 2162 means expressions that are made up of values we have no leaders for 2163 in SET. */ 2164 2165 static void 2166 clean (bitmap_set_t set, basic_block block) 2167 { 2168 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set); 2169 pre_expr expr; 2170 int i; 2171 2172 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr) 2173 { 2174 if (!valid_in_sets (set, NULL, expr, block)) 2175 bitmap_remove_from_set (set, expr); 2176 } 2177 VEC_free (pre_expr, heap, exprs); 2178 } 2179 2180 static sbitmap has_abnormal_preds; 2181 2182 /* List of blocks that may have changed during ANTIC computation and 2183 thus need to be iterated over. */ 2184 2185 static sbitmap changed_blocks; 2186 2187 /* Decide whether to defer a block for a later iteration, or PHI 2188 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we 2189 should defer the block, and true if we processed it. */ 2190 2191 static bool 2192 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source, 2193 basic_block block, basic_block phiblock) 2194 { 2195 if (!BB_VISITED (phiblock)) 2196 { 2197 SET_BIT (changed_blocks, block->index); 2198 BB_VISITED (block) = 0; 2199 BB_DEFERRED (block) = 1; 2200 return false; 2201 } 2202 else 2203 phi_translate_set (dest, source, block, phiblock); 2204 return true; 2205 } 2206 2207 /* Compute the ANTIC set for BLOCK. 2208 2209 If succs(BLOCK) > 1 then 2210 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK) 2211 else if succs(BLOCK) == 1 then 2212 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)]) 2213 2214 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK]) 2215 */ 2216 2217 static bool 2218 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge) 2219 { 2220 bool changed = false; 2221 bitmap_set_t S, old, ANTIC_OUT; 2222 bitmap_iterator bi; 2223 unsigned int bii; 2224 edge e; 2225 edge_iterator ei; 2226 2227 old = ANTIC_OUT = S = NULL; 2228 BB_VISITED (block) = 1; 2229 2230 /* If any edges from predecessors are abnormal, antic_in is empty, 2231 so do nothing. */ 2232 if (block_has_abnormal_pred_edge) 2233 goto maybe_dump_sets; 2234 2235 old = ANTIC_IN (block); 2236 ANTIC_OUT = bitmap_set_new (); 2237 2238 /* If the block has no successors, ANTIC_OUT is empty. */ 2239 if (EDGE_COUNT (block->succs) == 0) 2240 ; 2241 /* If we have one successor, we could have some phi nodes to 2242 translate through. */ 2243 else if (single_succ_p (block)) 2244 { 2245 basic_block succ_bb = single_succ (block); 2246 2247 /* We trade iterations of the dataflow equations for having to 2248 phi translate the maximal set, which is incredibly slow 2249 (since the maximal set often has 300+ members, even when you 2250 have a small number of blocks). 2251 Basically, we defer the computation of ANTIC for this block 2252 until we have processed it's successor, which will inevitably 2253 have a *much* smaller set of values to phi translate once 2254 clean has been run on it. 2255 The cost of doing this is that we technically perform more 2256 iterations, however, they are lower cost iterations. 2257 2258 Timings for PRE on tramp3d-v4: 2259 without maximal set fix: 11 seconds 2260 with maximal set fix/without deferring: 26 seconds 2261 with maximal set fix/with deferring: 11 seconds 2262 */ 2263 2264 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb), 2265 block, succ_bb)) 2266 { 2267 changed = true; 2268 goto maybe_dump_sets; 2269 } 2270 } 2271 /* If we have multiple successors, we take the intersection of all of 2272 them. Note that in the case of loop exit phi nodes, we may have 2273 phis to translate through. */ 2274 else 2275 { 2276 VEC(basic_block, heap) * worklist; 2277 size_t i; 2278 basic_block bprime, first = NULL; 2279 2280 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs)); 2281 FOR_EACH_EDGE (e, ei, block->succs) 2282 { 2283 if (!first 2284 && BB_VISITED (e->dest)) 2285 first = e->dest; 2286 else if (BB_VISITED (e->dest)) 2287 VEC_quick_push (basic_block, worklist, e->dest); 2288 } 2289 2290 /* Of multiple successors we have to have visited one already. */ 2291 if (!first) 2292 { 2293 SET_BIT (changed_blocks, block->index); 2294 BB_VISITED (block) = 0; 2295 BB_DEFERRED (block) = 1; 2296 changed = true; 2297 VEC_free (basic_block, heap, worklist); 2298 goto maybe_dump_sets; 2299 } 2300 2301 if (!gimple_seq_empty_p (phi_nodes (first))) 2302 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first); 2303 else 2304 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first)); 2305 2306 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime) 2307 { 2308 if (!gimple_seq_empty_p (phi_nodes (bprime))) 2309 { 2310 bitmap_set_t tmp = bitmap_set_new (); 2311 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime); 2312 bitmap_set_and (ANTIC_OUT, tmp); 2313 bitmap_set_free (tmp); 2314 } 2315 else 2316 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime)); 2317 } 2318 VEC_free (basic_block, heap, worklist); 2319 } 2320 2321 /* Generate ANTIC_OUT - TMP_GEN. */ 2322 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block)); 2323 2324 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */ 2325 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block), 2326 TMP_GEN (block)); 2327 2328 /* Then union in the ANTIC_OUT - TMP_GEN values, 2329 to get ANTIC_OUT U EXP_GEN - TMP_GEN */ 2330 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi) 2331 bitmap_value_insert_into_set (ANTIC_IN (block), 2332 expression_for_id (bii)); 2333 2334 clean (ANTIC_IN (block), block); 2335 2336 if (!bitmap_set_equal (old, ANTIC_IN (block))) 2337 { 2338 changed = true; 2339 SET_BIT (changed_blocks, block->index); 2340 FOR_EACH_EDGE (e, ei, block->preds) 2341 SET_BIT (changed_blocks, e->src->index); 2342 } 2343 else 2344 RESET_BIT (changed_blocks, block->index); 2345 2346 maybe_dump_sets: 2347 if (dump_file && (dump_flags & TDF_DETAILS)) 2348 { 2349 if (!BB_DEFERRED (block) || BB_VISITED (block)) 2350 { 2351 if (ANTIC_OUT) 2352 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index); 2353 2354 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN", 2355 block->index); 2356 2357 if (S) 2358 print_bitmap_set (dump_file, S, "S", block->index); 2359 } 2360 else 2361 { 2362 fprintf (dump_file, 2363 "Block %d was deferred for a future iteration.\n", 2364 block->index); 2365 } 2366 } 2367 if (old) 2368 bitmap_set_free (old); 2369 if (S) 2370 bitmap_set_free (S); 2371 if (ANTIC_OUT) 2372 bitmap_set_free (ANTIC_OUT); 2373 return changed; 2374 } 2375 2376 /* Compute PARTIAL_ANTIC for BLOCK. 2377 2378 If succs(BLOCK) > 1 then 2379 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not 2380 in ANTIC_OUT for all succ(BLOCK) 2381 else if succs(BLOCK) == 1 then 2382 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)]) 2383 2384 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK] 2385 - ANTIC_IN[BLOCK]) 2386 2387 */ 2388 static bool 2389 compute_partial_antic_aux (basic_block block, 2390 bool block_has_abnormal_pred_edge) 2391 { 2392 bool changed = false; 2393 bitmap_set_t old_PA_IN; 2394 bitmap_set_t PA_OUT; 2395 edge e; 2396 edge_iterator ei; 2397 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH); 2398 2399 old_PA_IN = PA_OUT = NULL; 2400 2401 /* If any edges from predecessors are abnormal, antic_in is empty, 2402 so do nothing. */ 2403 if (block_has_abnormal_pred_edge) 2404 goto maybe_dump_sets; 2405 2406 /* If there are too many partially anticipatable values in the 2407 block, phi_translate_set can take an exponential time: stop 2408 before the translation starts. */ 2409 if (max_pa 2410 && single_succ_p (block) 2411 && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa) 2412 goto maybe_dump_sets; 2413 2414 old_PA_IN = PA_IN (block); 2415 PA_OUT = bitmap_set_new (); 2416 2417 /* If the block has no successors, ANTIC_OUT is empty. */ 2418 if (EDGE_COUNT (block->succs) == 0) 2419 ; 2420 /* If we have one successor, we could have some phi nodes to 2421 translate through. Note that we can't phi translate across DFS 2422 back edges in partial antic, because it uses a union operation on 2423 the successors. For recurrences like IV's, we will end up 2424 generating a new value in the set on each go around (i + 3 (VH.1) 2425 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */ 2426 else if (single_succ_p (block)) 2427 { 2428 basic_block succ = single_succ (block); 2429 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK)) 2430 phi_translate_set (PA_OUT, PA_IN (succ), block, succ); 2431 } 2432 /* If we have multiple successors, we take the union of all of 2433 them. */ 2434 else 2435 { 2436 VEC(basic_block, heap) * worklist; 2437 size_t i; 2438 basic_block bprime; 2439 2440 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs)); 2441 FOR_EACH_EDGE (e, ei, block->succs) 2442 { 2443 if (e->flags & EDGE_DFS_BACK) 2444 continue; 2445 VEC_quick_push (basic_block, worklist, e->dest); 2446 } 2447 if (VEC_length (basic_block, worklist) > 0) 2448 { 2449 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime) 2450 { 2451 unsigned int i; 2452 bitmap_iterator bi; 2453 2454 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi) 2455 bitmap_value_insert_into_set (PA_OUT, 2456 expression_for_id (i)); 2457 if (!gimple_seq_empty_p (phi_nodes (bprime))) 2458 { 2459 bitmap_set_t pa_in = bitmap_set_new (); 2460 phi_translate_set (pa_in, PA_IN (bprime), block, bprime); 2461 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi) 2462 bitmap_value_insert_into_set (PA_OUT, 2463 expression_for_id (i)); 2464 bitmap_set_free (pa_in); 2465 } 2466 else 2467 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi) 2468 bitmap_value_insert_into_set (PA_OUT, 2469 expression_for_id (i)); 2470 } 2471 } 2472 VEC_free (basic_block, heap, worklist); 2473 } 2474 2475 /* PA_IN starts with PA_OUT - TMP_GEN. 2476 Then we subtract things from ANTIC_IN. */ 2477 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block)); 2478 2479 /* For partial antic, we want to put back in the phi results, since 2480 we will properly avoid making them partially antic over backedges. */ 2481 bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values); 2482 bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions); 2483 2484 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */ 2485 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block)); 2486 2487 dependent_clean (PA_IN (block), ANTIC_IN (block), block); 2488 2489 if (!bitmap_set_equal (old_PA_IN, PA_IN (block))) 2490 { 2491 changed = true; 2492 SET_BIT (changed_blocks, block->index); 2493 FOR_EACH_EDGE (e, ei, block->preds) 2494 SET_BIT (changed_blocks, e->src->index); 2495 } 2496 else 2497 RESET_BIT (changed_blocks, block->index); 2498 2499 maybe_dump_sets: 2500 if (dump_file && (dump_flags & TDF_DETAILS)) 2501 { 2502 if (PA_OUT) 2503 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index); 2504 2505 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index); 2506 } 2507 if (old_PA_IN) 2508 bitmap_set_free (old_PA_IN); 2509 if (PA_OUT) 2510 bitmap_set_free (PA_OUT); 2511 return changed; 2512 } 2513 2514 /* Compute ANTIC and partial ANTIC sets. */ 2515 2516 static void 2517 compute_antic (void) 2518 { 2519 bool changed = true; 2520 int num_iterations = 0; 2521 basic_block block; 2522 int i; 2523 2524 /* If any predecessor edges are abnormal, we punt, so antic_in is empty. 2525 We pre-build the map of blocks with incoming abnormal edges here. */ 2526 has_abnormal_preds = sbitmap_alloc (last_basic_block); 2527 sbitmap_zero (has_abnormal_preds); 2528 2529 FOR_EACH_BB (block) 2530 { 2531 edge_iterator ei; 2532 edge e; 2533 2534 FOR_EACH_EDGE (e, ei, block->preds) 2535 { 2536 e->flags &= ~EDGE_DFS_BACK; 2537 if (e->flags & EDGE_ABNORMAL) 2538 { 2539 SET_BIT (has_abnormal_preds, block->index); 2540 break; 2541 } 2542 } 2543 2544 BB_VISITED (block) = 0; 2545 BB_DEFERRED (block) = 0; 2546 2547 /* While we are here, give empty ANTIC_IN sets to each block. */ 2548 ANTIC_IN (block) = bitmap_set_new (); 2549 PA_IN (block) = bitmap_set_new (); 2550 } 2551 2552 /* At the exit block we anticipate nothing. */ 2553 ANTIC_IN (EXIT_BLOCK_PTR) = bitmap_set_new (); 2554 BB_VISITED (EXIT_BLOCK_PTR) = 1; 2555 PA_IN (EXIT_BLOCK_PTR) = bitmap_set_new (); 2556 2557 changed_blocks = sbitmap_alloc (last_basic_block + 1); 2558 sbitmap_ones (changed_blocks); 2559 while (changed) 2560 { 2561 if (dump_file && (dump_flags & TDF_DETAILS)) 2562 fprintf (dump_file, "Starting iteration %d\n", num_iterations); 2563 /* ??? We need to clear our PHI translation cache here as the 2564 ANTIC sets shrink and we restrict valid translations to 2565 those having operands with leaders in ANTIC. Same below 2566 for PA ANTIC computation. */ 2567 num_iterations++; 2568 changed = false; 2569 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1; i >= 0; i--) 2570 { 2571 if (TEST_BIT (changed_blocks, postorder[i])) 2572 { 2573 basic_block block = BASIC_BLOCK (postorder[i]); 2574 changed |= compute_antic_aux (block, 2575 TEST_BIT (has_abnormal_preds, 2576 block->index)); 2577 } 2578 } 2579 /* Theoretically possible, but *highly* unlikely. */ 2580 gcc_checking_assert (num_iterations < 500); 2581 } 2582 2583 statistics_histogram_event (cfun, "compute_antic iterations", 2584 num_iterations); 2585 2586 if (do_partial_partial) 2587 { 2588 sbitmap_ones (changed_blocks); 2589 mark_dfs_back_edges (); 2590 num_iterations = 0; 2591 changed = true; 2592 while (changed) 2593 { 2594 if (dump_file && (dump_flags & TDF_DETAILS)) 2595 fprintf (dump_file, "Starting iteration %d\n", num_iterations); 2596 num_iterations++; 2597 changed = false; 2598 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1 ; i >= 0; i--) 2599 { 2600 if (TEST_BIT (changed_blocks, postorder[i])) 2601 { 2602 basic_block block = BASIC_BLOCK (postorder[i]); 2603 changed 2604 |= compute_partial_antic_aux (block, 2605 TEST_BIT (has_abnormal_preds, 2606 block->index)); 2607 } 2608 } 2609 /* Theoretically possible, but *highly* unlikely. */ 2610 gcc_checking_assert (num_iterations < 500); 2611 } 2612 statistics_histogram_event (cfun, "compute_partial_antic iterations", 2613 num_iterations); 2614 } 2615 sbitmap_free (has_abnormal_preds); 2616 sbitmap_free (changed_blocks); 2617 } 2618 2619 /* Return true if OP is a tree which we can perform PRE on. 2620 This may not match the operations we can value number, but in 2621 a perfect world would. */ 2622 2623 static bool 2624 can_PRE_operation (tree op) 2625 { 2626 return UNARY_CLASS_P (op) 2627 || BINARY_CLASS_P (op) 2628 || COMPARISON_CLASS_P (op) 2629 || TREE_CODE (op) == MEM_REF 2630 || TREE_CODE (op) == COMPONENT_REF 2631 || TREE_CODE (op) == VIEW_CONVERT_EXPR 2632 || TREE_CODE (op) == CALL_EXPR 2633 || TREE_CODE (op) == ARRAY_REF; 2634 } 2635 2636 2637 /* Inserted expressions are placed onto this worklist, which is used 2638 for performing quick dead code elimination of insertions we made 2639 that didn't turn out to be necessary. */ 2640 static bitmap inserted_exprs; 2641 2642 /* Pool allocated fake store expressions are placed onto this 2643 worklist, which, after performing dead code elimination, is walked 2644 to see which expressions need to be put into GC'able memory */ 2645 static VEC(gimple, heap) *need_creation; 2646 2647 /* The actual worker for create_component_ref_by_pieces. */ 2648 2649 static tree 2650 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref, 2651 unsigned int *operand, gimple_seq *stmts, 2652 gimple domstmt) 2653 { 2654 vn_reference_op_t currop = VEC_index (vn_reference_op_s, ref->operands, 2655 *operand); 2656 tree genop; 2657 ++*operand; 2658 switch (currop->opcode) 2659 { 2660 case CALL_EXPR: 2661 { 2662 tree folded, sc = NULL_TREE; 2663 unsigned int nargs = 0; 2664 tree fn, *args; 2665 if (TREE_CODE (currop->op0) == FUNCTION_DECL) 2666 fn = currop->op0; 2667 else 2668 { 2669 pre_expr op0 = get_or_alloc_expr_for (currop->op0); 2670 fn = find_or_generate_expression (block, op0, stmts, domstmt); 2671 if (!fn) 2672 return NULL_TREE; 2673 } 2674 if (currop->op1) 2675 { 2676 pre_expr scexpr = get_or_alloc_expr_for (currop->op1); 2677 sc = find_or_generate_expression (block, scexpr, stmts, domstmt); 2678 if (!sc) 2679 return NULL_TREE; 2680 } 2681 args = XNEWVEC (tree, VEC_length (vn_reference_op_s, 2682 ref->operands) - 1); 2683 while (*operand < VEC_length (vn_reference_op_s, ref->operands)) 2684 { 2685 args[nargs] = create_component_ref_by_pieces_1 (block, ref, 2686 operand, stmts, 2687 domstmt); 2688 if (!args[nargs]) 2689 { 2690 free (args); 2691 return NULL_TREE; 2692 } 2693 nargs++; 2694 } 2695 folded = build_call_array (currop->type, 2696 (TREE_CODE (fn) == FUNCTION_DECL 2697 ? build_fold_addr_expr (fn) : fn), 2698 nargs, args); 2699 free (args); 2700 if (sc) 2701 CALL_EXPR_STATIC_CHAIN (folded) = sc; 2702 return folded; 2703 } 2704 break; 2705 case MEM_REF: 2706 { 2707 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand, 2708 stmts, domstmt); 2709 tree offset = currop->op0; 2710 if (!baseop) 2711 return NULL_TREE; 2712 if (TREE_CODE (baseop) == ADDR_EXPR 2713 && handled_component_p (TREE_OPERAND (baseop, 0))) 2714 { 2715 HOST_WIDE_INT off; 2716 tree base; 2717 base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0), 2718 &off); 2719 gcc_assert (base); 2720 offset = int_const_binop (PLUS_EXPR, offset, 2721 build_int_cst (TREE_TYPE (offset), 2722 off)); 2723 baseop = build_fold_addr_expr (base); 2724 } 2725 return fold_build2 (MEM_REF, currop->type, baseop, offset); 2726 } 2727 break; 2728 case TARGET_MEM_REF: 2729 { 2730 pre_expr op0expr, op1expr; 2731 tree genop0 = NULL_TREE, genop1 = NULL_TREE; 2732 vn_reference_op_t nextop = VEC_index (vn_reference_op_s, ref->operands, 2733 ++*operand); 2734 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand, 2735 stmts, domstmt); 2736 if (!baseop) 2737 return NULL_TREE; 2738 if (currop->op0) 2739 { 2740 op0expr = get_or_alloc_expr_for (currop->op0); 2741 genop0 = find_or_generate_expression (block, op0expr, 2742 stmts, domstmt); 2743 if (!genop0) 2744 return NULL_TREE; 2745 } 2746 if (nextop->op0) 2747 { 2748 op1expr = get_or_alloc_expr_for (nextop->op0); 2749 genop1 = find_or_generate_expression (block, op1expr, 2750 stmts, domstmt); 2751 if (!genop1) 2752 return NULL_TREE; 2753 } 2754 return build5 (TARGET_MEM_REF, currop->type, 2755 baseop, currop->op2, genop0, currop->op1, genop1); 2756 } 2757 break; 2758 case ADDR_EXPR: 2759 if (currop->op0) 2760 { 2761 gcc_assert (is_gimple_min_invariant (currop->op0)); 2762 return currop->op0; 2763 } 2764 /* Fallthrough. */ 2765 case REALPART_EXPR: 2766 case IMAGPART_EXPR: 2767 case VIEW_CONVERT_EXPR: 2768 { 2769 tree folded; 2770 tree genop0 = create_component_ref_by_pieces_1 (block, ref, 2771 operand, 2772 stmts, domstmt); 2773 if (!genop0) 2774 return NULL_TREE; 2775 folded = fold_build1 (currop->opcode, currop->type, 2776 genop0); 2777 return folded; 2778 } 2779 break; 2780 case WITH_SIZE_EXPR: 2781 { 2782 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand, 2783 stmts, domstmt); 2784 pre_expr op1expr = get_or_alloc_expr_for (currop->op0); 2785 tree genop1; 2786 2787 if (!genop0) 2788 return NULL_TREE; 2789 2790 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt); 2791 if (!genop1) 2792 return NULL_TREE; 2793 2794 return fold_build2 (currop->opcode, currop->type, genop0, genop1); 2795 } 2796 break; 2797 case BIT_FIELD_REF: 2798 { 2799 tree folded; 2800 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand, 2801 stmts, domstmt); 2802 pre_expr op1expr = get_or_alloc_expr_for (currop->op0); 2803 pre_expr op2expr = get_or_alloc_expr_for (currop->op1); 2804 tree genop1; 2805 tree genop2; 2806 2807 if (!genop0) 2808 return NULL_TREE; 2809 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt); 2810 if (!genop1) 2811 return NULL_TREE; 2812 genop2 = find_or_generate_expression (block, op2expr, stmts, domstmt); 2813 if (!genop2) 2814 return NULL_TREE; 2815 folded = fold_build3 (BIT_FIELD_REF, currop->type, genop0, genop1, 2816 genop2); 2817 return folded; 2818 } 2819 2820 /* For array ref vn_reference_op's, operand 1 of the array ref 2821 is op0 of the reference op and operand 3 of the array ref is 2822 op1. */ 2823 case ARRAY_RANGE_REF: 2824 case ARRAY_REF: 2825 { 2826 tree genop0; 2827 tree genop1 = currop->op0; 2828 pre_expr op1expr; 2829 tree genop2 = currop->op1; 2830 pre_expr op2expr; 2831 tree genop3 = currop->op2; 2832 pre_expr op3expr; 2833 genop0 = create_component_ref_by_pieces_1 (block, ref, operand, 2834 stmts, domstmt); 2835 if (!genop0) 2836 return NULL_TREE; 2837 op1expr = get_or_alloc_expr_for (genop1); 2838 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt); 2839 if (!genop1) 2840 return NULL_TREE; 2841 if (genop2) 2842 { 2843 tree domain_type = TYPE_DOMAIN (TREE_TYPE (genop0)); 2844 /* Drop zero minimum index if redundant. */ 2845 if (integer_zerop (genop2) 2846 && (!domain_type 2847 || integer_zerop (TYPE_MIN_VALUE (domain_type)))) 2848 genop2 = NULL_TREE; 2849 else 2850 { 2851 op2expr = get_or_alloc_expr_for (genop2); 2852 genop2 = find_or_generate_expression (block, op2expr, stmts, 2853 domstmt); 2854 if (!genop2) 2855 return NULL_TREE; 2856 } 2857 } 2858 if (genop3) 2859 { 2860 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0)); 2861 /* We can't always put a size in units of the element alignment 2862 here as the element alignment may be not visible. See 2863 PR43783. Simply drop the element size for constant 2864 sizes. */ 2865 if (tree_int_cst_equal (genop3, TYPE_SIZE_UNIT (elmt_type))) 2866 genop3 = NULL_TREE; 2867 else 2868 { 2869 genop3 = size_binop (EXACT_DIV_EXPR, genop3, 2870 size_int (TYPE_ALIGN_UNIT (elmt_type))); 2871 op3expr = get_or_alloc_expr_for (genop3); 2872 genop3 = find_or_generate_expression (block, op3expr, stmts, 2873 domstmt); 2874 if (!genop3) 2875 return NULL_TREE; 2876 } 2877 } 2878 return build4 (currop->opcode, currop->type, genop0, genop1, 2879 genop2, genop3); 2880 } 2881 case COMPONENT_REF: 2882 { 2883 tree op0; 2884 tree op1; 2885 tree genop2 = currop->op1; 2886 pre_expr op2expr; 2887 op0 = create_component_ref_by_pieces_1 (block, ref, operand, 2888 stmts, domstmt); 2889 if (!op0) 2890 return NULL_TREE; 2891 /* op1 should be a FIELD_DECL, which are represented by 2892 themselves. */ 2893 op1 = currop->op0; 2894 if (genop2) 2895 { 2896 op2expr = get_or_alloc_expr_for (genop2); 2897 genop2 = find_or_generate_expression (block, op2expr, stmts, 2898 domstmt); 2899 if (!genop2) 2900 return NULL_TREE; 2901 } 2902 2903 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1, 2904 genop2); 2905 } 2906 break; 2907 case SSA_NAME: 2908 { 2909 pre_expr op0expr = get_or_alloc_expr_for (currop->op0); 2910 genop = find_or_generate_expression (block, op0expr, stmts, domstmt); 2911 return genop; 2912 } 2913 case STRING_CST: 2914 case INTEGER_CST: 2915 case COMPLEX_CST: 2916 case VECTOR_CST: 2917 case REAL_CST: 2918 case CONSTRUCTOR: 2919 case VAR_DECL: 2920 case PARM_DECL: 2921 case CONST_DECL: 2922 case RESULT_DECL: 2923 case FUNCTION_DECL: 2924 return currop->op0; 2925 2926 default: 2927 gcc_unreachable (); 2928 } 2929 } 2930 2931 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the 2932 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with 2933 trying to rename aggregates into ssa form directly, which is a no no. 2934 2935 Thus, this routine doesn't create temporaries, it just builds a 2936 single access expression for the array, calling 2937 find_or_generate_expression to build the innermost pieces. 2938 2939 This function is a subroutine of create_expression_by_pieces, and 2940 should not be called on it's own unless you really know what you 2941 are doing. */ 2942 2943 static tree 2944 create_component_ref_by_pieces (basic_block block, vn_reference_t ref, 2945 gimple_seq *stmts, gimple domstmt) 2946 { 2947 unsigned int op = 0; 2948 return create_component_ref_by_pieces_1 (block, ref, &op, stmts, domstmt); 2949 } 2950 2951 /* Find a leader for an expression, or generate one using 2952 create_expression_by_pieces if it's ANTIC but 2953 complex. 2954 BLOCK is the basic_block we are looking for leaders in. 2955 EXPR is the expression to find a leader or generate for. 2956 STMTS is the statement list to put the inserted expressions on. 2957 Returns the SSA_NAME of the LHS of the generated expression or the 2958 leader. 2959 DOMSTMT if non-NULL is a statement that should be dominated by 2960 all uses in the generated expression. If DOMSTMT is non-NULL this 2961 routine can fail and return NULL_TREE. Otherwise it will assert 2962 on failure. */ 2963 2964 static tree 2965 find_or_generate_expression (basic_block block, pre_expr expr, 2966 gimple_seq *stmts, gimple domstmt) 2967 { 2968 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block), 2969 get_expr_value_id (expr), domstmt); 2970 tree genop = NULL; 2971 if (leader) 2972 { 2973 if (leader->kind == NAME) 2974 genop = PRE_EXPR_NAME (leader); 2975 else if (leader->kind == CONSTANT) 2976 genop = PRE_EXPR_CONSTANT (leader); 2977 } 2978 2979 /* If it's still NULL, it must be a complex expression, so generate 2980 it recursively. Not so if inserting expressions for values generated 2981 by SCCVN. */ 2982 if (genop == NULL 2983 && !domstmt) 2984 { 2985 bitmap_set_t exprset; 2986 unsigned int lookfor = get_expr_value_id (expr); 2987 bool handled = false; 2988 bitmap_iterator bi; 2989 unsigned int i; 2990 2991 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor); 2992 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi) 2993 { 2994 pre_expr temp = expression_for_id (i); 2995 if (temp->kind != NAME) 2996 { 2997 handled = true; 2998 genop = create_expression_by_pieces (block, temp, stmts, 2999 domstmt, 3000 get_expr_type (expr)); 3001 break; 3002 } 3003 } 3004 if (!handled && domstmt) 3005 return NULL_TREE; 3006 3007 gcc_assert (handled); 3008 } 3009 return genop; 3010 } 3011 3012 #define NECESSARY GF_PLF_1 3013 3014 /* Create an expression in pieces, so that we can handle very complex 3015 expressions that may be ANTIC, but not necessary GIMPLE. 3016 BLOCK is the basic block the expression will be inserted into, 3017 EXPR is the expression to insert (in value form) 3018 STMTS is a statement list to append the necessary insertions into. 3019 3020 This function will die if we hit some value that shouldn't be 3021 ANTIC but is (IE there is no leader for it, or its components). 3022 This function may also generate expressions that are themselves 3023 partially or fully redundant. Those that are will be either made 3024 fully redundant during the next iteration of insert (for partially 3025 redundant ones), or eliminated by eliminate (for fully redundant 3026 ones). 3027 3028 If DOMSTMT is non-NULL then we make sure that all uses in the 3029 expressions dominate that statement. In this case the function 3030 can return NULL_TREE to signal failure. */ 3031 3032 static tree 3033 create_expression_by_pieces (basic_block block, pre_expr expr, 3034 gimple_seq *stmts, gimple domstmt, tree type) 3035 { 3036 tree temp, name; 3037 tree folded; 3038 gimple_seq forced_stmts = NULL; 3039 unsigned int value_id; 3040 gimple_stmt_iterator gsi; 3041 tree exprtype = type ? type : get_expr_type (expr); 3042 pre_expr nameexpr; 3043 gimple newstmt; 3044 3045 switch (expr->kind) 3046 { 3047 /* We may hit the NAME/CONSTANT case if we have to convert types 3048 that value numbering saw through. */ 3049 case NAME: 3050 folded = PRE_EXPR_NAME (expr); 3051 break; 3052 case CONSTANT: 3053 folded = PRE_EXPR_CONSTANT (expr); 3054 break; 3055 case REFERENCE: 3056 { 3057 vn_reference_t ref = PRE_EXPR_REFERENCE (expr); 3058 folded = create_component_ref_by_pieces (block, ref, stmts, domstmt); 3059 } 3060 break; 3061 case NARY: 3062 { 3063 vn_nary_op_t nary = PRE_EXPR_NARY (expr); 3064 tree *genop = XALLOCAVEC (tree, nary->length); 3065 unsigned i; 3066 for (i = 0; i < nary->length; ++i) 3067 { 3068 pre_expr op = get_or_alloc_expr_for (nary->op[i]); 3069 genop[i] = find_or_generate_expression (block, op, 3070 stmts, domstmt); 3071 if (!genop[i]) 3072 return NULL_TREE; 3073 /* Ensure genop[] is properly typed for POINTER_PLUS_EXPR. It 3074 may have conversions stripped. */ 3075 if (nary->opcode == POINTER_PLUS_EXPR) 3076 { 3077 if (i == 0) 3078 genop[i] = fold_convert (nary->type, genop[i]); 3079 else if (i == 1) 3080 genop[i] = convert_to_ptrofftype (genop[i]); 3081 } 3082 else 3083 genop[i] = fold_convert (TREE_TYPE (nary->op[i]), genop[i]); 3084 } 3085 if (nary->opcode == CONSTRUCTOR) 3086 { 3087 VEC(constructor_elt,gc) *elts = NULL; 3088 for (i = 0; i < nary->length; ++i) 3089 CONSTRUCTOR_APPEND_ELT (elts, NULL_TREE, genop[i]); 3090 folded = build_constructor (nary->type, elts); 3091 } 3092 else 3093 { 3094 switch (nary->length) 3095 { 3096 case 1: 3097 folded = fold_build1 (nary->opcode, nary->type, 3098 genop[0]); 3099 break; 3100 case 2: 3101 folded = fold_build2 (nary->opcode, nary->type, 3102 genop[0], genop[1]); 3103 break; 3104 case 3: 3105 folded = fold_build3 (nary->opcode, nary->type, 3106 genop[0], genop[1], genop[2]); 3107 break; 3108 default: 3109 gcc_unreachable (); 3110 } 3111 } 3112 } 3113 break; 3114 default: 3115 return NULL_TREE; 3116 } 3117 3118 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded))) 3119 folded = fold_convert (exprtype, folded); 3120 3121 /* Force the generated expression to be a sequence of GIMPLE 3122 statements. 3123 We have to call unshare_expr because force_gimple_operand may 3124 modify the tree we pass to it. */ 3125 folded = force_gimple_operand (unshare_expr (folded), &forced_stmts, 3126 false, NULL); 3127 3128 /* If we have any intermediate expressions to the value sets, add them 3129 to the value sets and chain them in the instruction stream. */ 3130 if (forced_stmts) 3131 { 3132 gsi = gsi_start (forced_stmts); 3133 for (; !gsi_end_p (gsi); gsi_next (&gsi)) 3134 { 3135 gimple stmt = gsi_stmt (gsi); 3136 tree forcedname = gimple_get_lhs (stmt); 3137 pre_expr nameexpr; 3138 3139 if (TREE_CODE (forcedname) == SSA_NAME) 3140 { 3141 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname)); 3142 VN_INFO_GET (forcedname)->valnum = forcedname; 3143 VN_INFO (forcedname)->value_id = get_next_value_id (); 3144 nameexpr = get_or_alloc_expr_for_name (forcedname); 3145 add_to_value (VN_INFO (forcedname)->value_id, nameexpr); 3146 if (!in_fre) 3147 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr); 3148 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr); 3149 } 3150 mark_symbols_for_renaming (stmt); 3151 } 3152 gimple_seq_add_seq (stmts, forced_stmts); 3153 } 3154 3155 /* Build and insert the assignment of the end result to the temporary 3156 that we will return. */ 3157 if (!pretemp || exprtype != TREE_TYPE (pretemp)) 3158 pretemp = create_tmp_reg (exprtype, "pretmp"); 3159 3160 temp = pretemp; 3161 add_referenced_var (temp); 3162 3163 newstmt = gimple_build_assign (temp, folded); 3164 name = make_ssa_name (temp, newstmt); 3165 gimple_assign_set_lhs (newstmt, name); 3166 gimple_set_plf (newstmt, NECESSARY, false); 3167 3168 gimple_seq_add_stmt (stmts, newstmt); 3169 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (name)); 3170 3171 /* All the symbols in NEWEXPR should be put into SSA form. */ 3172 mark_symbols_for_renaming (newstmt); 3173 3174 /* Fold the last statement. */ 3175 gsi = gsi_last (*stmts); 3176 if (fold_stmt_inplace (&gsi)) 3177 update_stmt (gsi_stmt (gsi)); 3178 3179 /* Add a value number to the temporary. 3180 The value may already exist in either NEW_SETS, or AVAIL_OUT, because 3181 we are creating the expression by pieces, and this particular piece of 3182 the expression may have been represented. There is no harm in replacing 3183 here. */ 3184 VN_INFO_GET (name)->valnum = name; 3185 value_id = get_expr_value_id (expr); 3186 VN_INFO (name)->value_id = value_id; 3187 nameexpr = get_or_alloc_expr_for_name (name); 3188 add_to_value (value_id, nameexpr); 3189 if (NEW_SETS (block)) 3190 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr); 3191 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr); 3192 3193 pre_stats.insertions++; 3194 if (dump_file && (dump_flags & TDF_DETAILS)) 3195 { 3196 fprintf (dump_file, "Inserted "); 3197 print_gimple_stmt (dump_file, newstmt, 0, 0); 3198 fprintf (dump_file, " in predecessor %d\n", block->index); 3199 } 3200 3201 return name; 3202 } 3203 3204 3205 /* Returns true if we want to inhibit the insertions of PHI nodes 3206 for the given EXPR for basic block BB (a member of a loop). 3207 We want to do this, when we fear that the induction variable we 3208 create might inhibit vectorization. */ 3209 3210 static bool 3211 inhibit_phi_insertion (basic_block bb, pre_expr expr) 3212 { 3213 vn_reference_t vr = PRE_EXPR_REFERENCE (expr); 3214 VEC (vn_reference_op_s, heap) *ops = vr->operands; 3215 vn_reference_op_t op; 3216 unsigned i; 3217 3218 /* If we aren't going to vectorize we don't inhibit anything. */ 3219 if (!flag_tree_vectorize) 3220 return false; 3221 3222 /* Otherwise we inhibit the insertion when the address of the 3223 memory reference is a simple induction variable. In other 3224 cases the vectorizer won't do anything anyway (either it's 3225 loop invariant or a complicated expression). */ 3226 FOR_EACH_VEC_ELT (vn_reference_op_s, ops, i, op) 3227 { 3228 switch (op->opcode) 3229 { 3230 case ARRAY_REF: 3231 case ARRAY_RANGE_REF: 3232 if (TREE_CODE (op->op0) != SSA_NAME) 3233 break; 3234 /* Fallthru. */ 3235 case SSA_NAME: 3236 { 3237 basic_block defbb = gimple_bb (SSA_NAME_DEF_STMT (op->op0)); 3238 affine_iv iv; 3239 /* Default defs are loop invariant. */ 3240 if (!defbb) 3241 break; 3242 /* Defined outside this loop, also loop invariant. */ 3243 if (!flow_bb_inside_loop_p (bb->loop_father, defbb)) 3244 break; 3245 /* If it's a simple induction variable inhibit insertion, 3246 the vectorizer might be interested in this one. */ 3247 if (simple_iv (bb->loop_father, bb->loop_father, 3248 op->op0, &iv, true)) 3249 return true; 3250 /* No simple IV, vectorizer can't do anything, hence no 3251 reason to inhibit the transformation for this operand. */ 3252 break; 3253 } 3254 default: 3255 break; 3256 } 3257 } 3258 return false; 3259 } 3260 3261 /* Insert the to-be-made-available values of expression EXPRNUM for each 3262 predecessor, stored in AVAIL, into the predecessors of BLOCK, and 3263 merge the result with a phi node, given the same value number as 3264 NODE. Return true if we have inserted new stuff. */ 3265 3266 static bool 3267 insert_into_preds_of_block (basic_block block, unsigned int exprnum, 3268 pre_expr *avail) 3269 { 3270 pre_expr expr = expression_for_id (exprnum); 3271 pre_expr newphi; 3272 unsigned int val = get_expr_value_id (expr); 3273 edge pred; 3274 bool insertions = false; 3275 bool nophi = false; 3276 basic_block bprime; 3277 pre_expr eprime; 3278 edge_iterator ei; 3279 tree type = get_expr_type (expr); 3280 tree temp; 3281 gimple phi; 3282 3283 if (dump_file && (dump_flags & TDF_DETAILS)) 3284 { 3285 fprintf (dump_file, "Found partial redundancy for expression "); 3286 print_pre_expr (dump_file, expr); 3287 fprintf (dump_file, " (%04d)\n", val); 3288 } 3289 3290 /* Make sure we aren't creating an induction variable. */ 3291 if (block->loop_depth > 0 && EDGE_COUNT (block->preds) == 2) 3292 { 3293 bool firstinsideloop = false; 3294 bool secondinsideloop = false; 3295 firstinsideloop = flow_bb_inside_loop_p (block->loop_father, 3296 EDGE_PRED (block, 0)->src); 3297 secondinsideloop = flow_bb_inside_loop_p (block->loop_father, 3298 EDGE_PRED (block, 1)->src); 3299 /* Induction variables only have one edge inside the loop. */ 3300 if ((firstinsideloop ^ secondinsideloop) 3301 && (expr->kind != REFERENCE 3302 || inhibit_phi_insertion (block, expr))) 3303 { 3304 if (dump_file && (dump_flags & TDF_DETAILS)) 3305 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n"); 3306 nophi = true; 3307 } 3308 } 3309 3310 /* Make the necessary insertions. */ 3311 FOR_EACH_EDGE (pred, ei, block->preds) 3312 { 3313 gimple_seq stmts = NULL; 3314 tree builtexpr; 3315 bprime = pred->src; 3316 eprime = avail[bprime->index]; 3317 3318 if (eprime->kind != NAME && eprime->kind != CONSTANT) 3319 { 3320 builtexpr = create_expression_by_pieces (bprime, 3321 eprime, 3322 &stmts, NULL, 3323 type); 3324 gcc_assert (!(pred->flags & EDGE_ABNORMAL)); 3325 gsi_insert_seq_on_edge (pred, stmts); 3326 avail[bprime->index] = get_or_alloc_expr_for_name (builtexpr); 3327 insertions = true; 3328 } 3329 else if (eprime->kind == CONSTANT) 3330 { 3331 /* Constants may not have the right type, fold_convert 3332 should give us back a constant with the right type. 3333 */ 3334 tree constant = PRE_EXPR_CONSTANT (eprime); 3335 if (!useless_type_conversion_p (type, TREE_TYPE (constant))) 3336 { 3337 tree builtexpr = fold_convert (type, constant); 3338 if (!is_gimple_min_invariant (builtexpr)) 3339 { 3340 tree forcedexpr = force_gimple_operand (builtexpr, 3341 &stmts, true, 3342 NULL); 3343 if (!is_gimple_min_invariant (forcedexpr)) 3344 { 3345 if (forcedexpr != builtexpr) 3346 { 3347 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime); 3348 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime); 3349 } 3350 if (stmts) 3351 { 3352 gimple_stmt_iterator gsi; 3353 gsi = gsi_start (stmts); 3354 for (; !gsi_end_p (gsi); gsi_next (&gsi)) 3355 { 3356 gimple stmt = gsi_stmt (gsi); 3357 tree lhs = gimple_get_lhs (stmt); 3358 if (TREE_CODE (lhs) == SSA_NAME) 3359 bitmap_set_bit (inserted_exprs, 3360 SSA_NAME_VERSION (lhs)); 3361 gimple_set_plf (stmt, NECESSARY, false); 3362 } 3363 gsi_insert_seq_on_edge (pred, stmts); 3364 } 3365 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr); 3366 } 3367 } 3368 else 3369 avail[bprime->index] = get_or_alloc_expr_for_constant (builtexpr); 3370 } 3371 } 3372 else if (eprime->kind == NAME) 3373 { 3374 /* We may have to do a conversion because our value 3375 numbering can look through types in certain cases, but 3376 our IL requires all operands of a phi node have the same 3377 type. */ 3378 tree name = PRE_EXPR_NAME (eprime); 3379 if (!useless_type_conversion_p (type, TREE_TYPE (name))) 3380 { 3381 tree builtexpr; 3382 tree forcedexpr; 3383 builtexpr = fold_convert (type, name); 3384 forcedexpr = force_gimple_operand (builtexpr, 3385 &stmts, true, 3386 NULL); 3387 3388 if (forcedexpr != name) 3389 { 3390 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum; 3391 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id; 3392 } 3393 3394 if (stmts) 3395 { 3396 gimple_stmt_iterator gsi; 3397 gsi = gsi_start (stmts); 3398 for (; !gsi_end_p (gsi); gsi_next (&gsi)) 3399 { 3400 gimple stmt = gsi_stmt (gsi); 3401 tree lhs = gimple_get_lhs (stmt); 3402 if (TREE_CODE (lhs) == SSA_NAME) 3403 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (lhs)); 3404 gimple_set_plf (stmt, NECESSARY, false); 3405 } 3406 gsi_insert_seq_on_edge (pred, stmts); 3407 } 3408 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr); 3409 } 3410 } 3411 } 3412 /* If we didn't want a phi node, and we made insertions, we still have 3413 inserted new stuff, and thus return true. If we didn't want a phi node, 3414 and didn't make insertions, we haven't added anything new, so return 3415 false. */ 3416 if (nophi && insertions) 3417 return true; 3418 else if (nophi && !insertions) 3419 return false; 3420 3421 /* Now build a phi for the new variable. */ 3422 if (!prephitemp || TREE_TYPE (prephitemp) != type) 3423 prephitemp = create_tmp_var (type, "prephitmp"); 3424 3425 temp = prephitemp; 3426 add_referenced_var (temp); 3427 3428 if (TREE_CODE (type) == COMPLEX_TYPE 3429 || TREE_CODE (type) == VECTOR_TYPE) 3430 DECL_GIMPLE_REG_P (temp) = 1; 3431 phi = create_phi_node (temp, block); 3432 3433 gimple_set_plf (phi, NECESSARY, false); 3434 VN_INFO_GET (gimple_phi_result (phi))->valnum = gimple_phi_result (phi); 3435 VN_INFO (gimple_phi_result (phi))->value_id = val; 3436 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (gimple_phi_result (phi))); 3437 FOR_EACH_EDGE (pred, ei, block->preds) 3438 { 3439 pre_expr ae = avail[pred->src->index]; 3440 gcc_assert (get_expr_type (ae) == type 3441 || useless_type_conversion_p (type, get_expr_type (ae))); 3442 if (ae->kind == CONSTANT) 3443 add_phi_arg (phi, PRE_EXPR_CONSTANT (ae), pred, UNKNOWN_LOCATION); 3444 else 3445 add_phi_arg (phi, PRE_EXPR_NAME (avail[pred->src->index]), pred, 3446 UNKNOWN_LOCATION); 3447 } 3448 3449 newphi = get_or_alloc_expr_for_name (gimple_phi_result (phi)); 3450 add_to_value (val, newphi); 3451 3452 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing 3453 this insertion, since we test for the existence of this value in PHI_GEN 3454 before proceeding with the partial redundancy checks in insert_aux. 3455 3456 The value may exist in AVAIL_OUT, in particular, it could be represented 3457 by the expression we are trying to eliminate, in which case we want the 3458 replacement to occur. If it's not existing in AVAIL_OUT, we want it 3459 inserted there. 3460 3461 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of 3462 this block, because if it did, it would have existed in our dominator's 3463 AVAIL_OUT, and would have been skipped due to the full redundancy check. 3464 */ 3465 3466 bitmap_insert_into_set (PHI_GEN (block), newphi); 3467 bitmap_value_replace_in_set (AVAIL_OUT (block), 3468 newphi); 3469 bitmap_insert_into_set (NEW_SETS (block), 3470 newphi); 3471 3472 if (dump_file && (dump_flags & TDF_DETAILS)) 3473 { 3474 fprintf (dump_file, "Created phi "); 3475 print_gimple_stmt (dump_file, phi, 0, 0); 3476 fprintf (dump_file, " in block %d\n", block->index); 3477 } 3478 pre_stats.phis++; 3479 return true; 3480 } 3481 3482 3483 3484 /* Perform insertion of partially redundant values. 3485 For BLOCK, do the following: 3486 1. Propagate the NEW_SETS of the dominator into the current block. 3487 If the block has multiple predecessors, 3488 2a. Iterate over the ANTIC expressions for the block to see if 3489 any of them are partially redundant. 3490 2b. If so, insert them into the necessary predecessors to make 3491 the expression fully redundant. 3492 2c. Insert a new PHI merging the values of the predecessors. 3493 2d. Insert the new PHI, and the new expressions, into the 3494 NEW_SETS set. 3495 3. Recursively call ourselves on the dominator children of BLOCK. 3496 3497 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by 3498 do_regular_insertion and do_partial_insertion. 3499 3500 */ 3501 3502 static bool 3503 do_regular_insertion (basic_block block, basic_block dom) 3504 { 3505 bool new_stuff = false; 3506 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (ANTIC_IN (block)); 3507 pre_expr expr; 3508 int i; 3509 3510 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr) 3511 { 3512 if (expr->kind == NARY 3513 || expr->kind == REFERENCE) 3514 { 3515 pre_expr *avail; 3516 unsigned int val; 3517 bool by_some = false; 3518 bool cant_insert = false; 3519 bool all_same = true; 3520 pre_expr first_s = NULL; 3521 edge pred; 3522 basic_block bprime; 3523 pre_expr eprime = NULL; 3524 edge_iterator ei; 3525 pre_expr edoubleprime = NULL; 3526 bool do_insertion = false; 3527 3528 val = get_expr_value_id (expr); 3529 if (bitmap_set_contains_value (PHI_GEN (block), val)) 3530 continue; 3531 if (bitmap_set_contains_value (AVAIL_OUT (dom), val)) 3532 { 3533 if (dump_file && (dump_flags & TDF_DETAILS)) 3534 fprintf (dump_file, "Found fully redundant value\n"); 3535 continue; 3536 } 3537 3538 avail = XCNEWVEC (pre_expr, last_basic_block); 3539 FOR_EACH_EDGE (pred, ei, block->preds) 3540 { 3541 unsigned int vprime; 3542 3543 /* We should never run insertion for the exit block 3544 and so not come across fake pred edges. */ 3545 gcc_assert (!(pred->flags & EDGE_FAKE)); 3546 bprime = pred->src; 3547 eprime = phi_translate (expr, ANTIC_IN (block), NULL, 3548 bprime, block); 3549 3550 /* eprime will generally only be NULL if the 3551 value of the expression, translated 3552 through the PHI for this predecessor, is 3553 undefined. If that is the case, we can't 3554 make the expression fully redundant, 3555 because its value is undefined along a 3556 predecessor path. We can thus break out 3557 early because it doesn't matter what the 3558 rest of the results are. */ 3559 if (eprime == NULL) 3560 { 3561 cant_insert = true; 3562 break; 3563 } 3564 3565 eprime = fully_constant_expression (eprime); 3566 vprime = get_expr_value_id (eprime); 3567 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime), 3568 vprime, NULL); 3569 if (edoubleprime == NULL) 3570 { 3571 avail[bprime->index] = eprime; 3572 all_same = false; 3573 } 3574 else 3575 { 3576 avail[bprime->index] = edoubleprime; 3577 by_some = true; 3578 /* We want to perform insertions to remove a redundancy on 3579 a path in the CFG we want to optimize for speed. */ 3580 if (optimize_edge_for_speed_p (pred)) 3581 do_insertion = true; 3582 if (first_s == NULL) 3583 first_s = edoubleprime; 3584 else if (!pre_expr_eq (first_s, edoubleprime)) 3585 all_same = false; 3586 } 3587 } 3588 /* If we can insert it, it's not the same value 3589 already existing along every predecessor, and 3590 it's defined by some predecessor, it is 3591 partially redundant. */ 3592 if (!cant_insert && !all_same && by_some) 3593 { 3594 if (!do_insertion) 3595 { 3596 if (dump_file && (dump_flags & TDF_DETAILS)) 3597 { 3598 fprintf (dump_file, "Skipping partial redundancy for " 3599 "expression "); 3600 print_pre_expr (dump_file, expr); 3601 fprintf (dump_file, " (%04d), no redundancy on to be " 3602 "optimized for speed edge\n", val); 3603 } 3604 } 3605 else if (dbg_cnt (treepre_insert) 3606 && insert_into_preds_of_block (block, 3607 get_expression_id (expr), 3608 avail)) 3609 new_stuff = true; 3610 } 3611 /* If all edges produce the same value and that value is 3612 an invariant, then the PHI has the same value on all 3613 edges. Note this. */ 3614 else if (!cant_insert && all_same) 3615 { 3616 tree exprtype = get_expr_type (expr); 3617 tree temp; 3618 gimple assign; 3619 pre_expr newe; 3620 gimple_stmt_iterator gsi; 3621 3622 gcc_assert (edoubleprime->kind == CONSTANT 3623 || edoubleprime->kind == NAME); 3624 3625 if (!pretemp || TREE_TYPE (pretemp) != exprtype) 3626 { 3627 pretemp = create_tmp_reg (exprtype, "pretmp"); 3628 add_referenced_var (pretemp); 3629 } 3630 temp = make_ssa_name (pretemp, NULL); 3631 assign = gimple_build_assign (temp, 3632 edoubleprime->kind == CONSTANT ? PRE_EXPR_CONSTANT (edoubleprime) : PRE_EXPR_NAME (edoubleprime)); 3633 gsi = gsi_after_labels (block); 3634 gsi_insert_before (&gsi, assign, GSI_NEW_STMT); 3635 3636 gimple_set_plf (assign, NECESSARY, false); 3637 VN_INFO_GET (temp)->value_id = val; 3638 VN_INFO (temp)->valnum = temp; 3639 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp)); 3640 newe = get_or_alloc_expr_for_name (temp); 3641 add_to_value (val, newe); 3642 bitmap_value_replace_in_set (AVAIL_OUT (block), newe); 3643 bitmap_insert_into_set (NEW_SETS (block), newe); 3644 } 3645 free (avail); 3646 } 3647 } 3648 3649 VEC_free (pre_expr, heap, exprs); 3650 return new_stuff; 3651 } 3652 3653 3654 /* Perform insertion for partially anticipatable expressions. There 3655 is only one case we will perform insertion for these. This case is 3656 if the expression is partially anticipatable, and fully available. 3657 In this case, we know that putting it earlier will enable us to 3658 remove the later computation. */ 3659 3660 3661 static bool 3662 do_partial_partial_insertion (basic_block block, basic_block dom) 3663 { 3664 bool new_stuff = false; 3665 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (PA_IN (block)); 3666 pre_expr expr; 3667 int i; 3668 3669 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr) 3670 { 3671 if (expr->kind == NARY 3672 || expr->kind == REFERENCE) 3673 { 3674 pre_expr *avail; 3675 unsigned int val; 3676 bool by_all = true; 3677 bool cant_insert = false; 3678 edge pred; 3679 basic_block bprime; 3680 pre_expr eprime = NULL; 3681 edge_iterator ei; 3682 3683 val = get_expr_value_id (expr); 3684 if (bitmap_set_contains_value (PHI_GEN (block), val)) 3685 continue; 3686 if (bitmap_set_contains_value (AVAIL_OUT (dom), val)) 3687 continue; 3688 3689 avail = XCNEWVEC (pre_expr, last_basic_block); 3690 FOR_EACH_EDGE (pred, ei, block->preds) 3691 { 3692 unsigned int vprime; 3693 pre_expr edoubleprime; 3694 3695 /* We should never run insertion for the exit block 3696 and so not come across fake pred edges. */ 3697 gcc_assert (!(pred->flags & EDGE_FAKE)); 3698 bprime = pred->src; 3699 eprime = phi_translate (expr, ANTIC_IN (block), 3700 PA_IN (block), 3701 bprime, block); 3702 3703 /* eprime will generally only be NULL if the 3704 value of the expression, translated 3705 through the PHI for this predecessor, is 3706 undefined. If that is the case, we can't 3707 make the expression fully redundant, 3708 because its value is undefined along a 3709 predecessor path. We can thus break out 3710 early because it doesn't matter what the 3711 rest of the results are. */ 3712 if (eprime == NULL) 3713 { 3714 cant_insert = true; 3715 break; 3716 } 3717 3718 eprime = fully_constant_expression (eprime); 3719 vprime = get_expr_value_id (eprime); 3720 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime), 3721 vprime, NULL); 3722 if (edoubleprime == NULL) 3723 { 3724 by_all = false; 3725 break; 3726 } 3727 else 3728 avail[bprime->index] = edoubleprime; 3729 3730 } 3731 3732 /* If we can insert it, it's not the same value 3733 already existing along every predecessor, and 3734 it's defined by some predecessor, it is 3735 partially redundant. */ 3736 if (!cant_insert && by_all && dbg_cnt (treepre_insert)) 3737 { 3738 pre_stats.pa_insert++; 3739 if (insert_into_preds_of_block (block, get_expression_id (expr), 3740 avail)) 3741 new_stuff = true; 3742 } 3743 free (avail); 3744 } 3745 } 3746 3747 VEC_free (pre_expr, heap, exprs); 3748 return new_stuff; 3749 } 3750 3751 static bool 3752 insert_aux (basic_block block) 3753 { 3754 basic_block son; 3755 bool new_stuff = false; 3756 3757 if (block) 3758 { 3759 basic_block dom; 3760 dom = get_immediate_dominator (CDI_DOMINATORS, block); 3761 if (dom) 3762 { 3763 unsigned i; 3764 bitmap_iterator bi; 3765 bitmap_set_t newset = NEW_SETS (dom); 3766 if (newset) 3767 { 3768 /* Note that we need to value_replace both NEW_SETS, and 3769 AVAIL_OUT. For both the case of NEW_SETS, the value may be 3770 represented by some non-simple expression here that we want 3771 to replace it with. */ 3772 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi) 3773 { 3774 pre_expr expr = expression_for_id (i); 3775 bitmap_value_replace_in_set (NEW_SETS (block), expr); 3776 bitmap_value_replace_in_set (AVAIL_OUT (block), expr); 3777 } 3778 } 3779 if (!single_pred_p (block)) 3780 { 3781 new_stuff |= do_regular_insertion (block, dom); 3782 if (do_partial_partial) 3783 new_stuff |= do_partial_partial_insertion (block, dom); 3784 } 3785 } 3786 } 3787 for (son = first_dom_son (CDI_DOMINATORS, block); 3788 son; 3789 son = next_dom_son (CDI_DOMINATORS, son)) 3790 { 3791 new_stuff |= insert_aux (son); 3792 } 3793 3794 return new_stuff; 3795 } 3796 3797 /* Perform insertion of partially redundant values. */ 3798 3799 static void 3800 insert (void) 3801 { 3802 bool new_stuff = true; 3803 basic_block bb; 3804 int num_iterations = 0; 3805 3806 FOR_ALL_BB (bb) 3807 NEW_SETS (bb) = bitmap_set_new (); 3808 3809 while (new_stuff) 3810 { 3811 num_iterations++; 3812 new_stuff = insert_aux (ENTRY_BLOCK_PTR); 3813 } 3814 statistics_histogram_event (cfun, "insert iterations", num_iterations); 3815 } 3816 3817 3818 /* Add OP to EXP_GEN (block), and possibly to the maximal set. */ 3819 3820 static void 3821 add_to_exp_gen (basic_block block, tree op) 3822 { 3823 if (!in_fre) 3824 { 3825 pre_expr result; 3826 if (TREE_CODE (op) == SSA_NAME && ssa_undefined_value_p (op)) 3827 return; 3828 result = get_or_alloc_expr_for_name (op); 3829 bitmap_value_insert_into_set (EXP_GEN (block), result); 3830 } 3831 } 3832 3833 /* Create value ids for PHI in BLOCK. */ 3834 3835 static void 3836 make_values_for_phi (gimple phi, basic_block block) 3837 { 3838 tree result = gimple_phi_result (phi); 3839 3840 /* We have no need for virtual phis, as they don't represent 3841 actual computations. */ 3842 if (is_gimple_reg (result)) 3843 { 3844 pre_expr e = get_or_alloc_expr_for_name (result); 3845 add_to_value (get_expr_value_id (e), e); 3846 bitmap_insert_into_set (PHI_GEN (block), e); 3847 bitmap_value_insert_into_set (AVAIL_OUT (block), e); 3848 if (!in_fre) 3849 { 3850 unsigned i; 3851 for (i = 0; i < gimple_phi_num_args (phi); ++i) 3852 { 3853 tree arg = gimple_phi_arg_def (phi, i); 3854 if (TREE_CODE (arg) == SSA_NAME) 3855 { 3856 e = get_or_alloc_expr_for_name (arg); 3857 add_to_value (get_expr_value_id (e), e); 3858 } 3859 } 3860 } 3861 } 3862 } 3863 3864 /* Compute the AVAIL set for all basic blocks. 3865 3866 This function performs value numbering of the statements in each basic 3867 block. The AVAIL sets are built from information we glean while doing 3868 this value numbering, since the AVAIL sets contain only one entry per 3869 value. 3870 3871 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)]. 3872 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */ 3873 3874 static void 3875 compute_avail (void) 3876 { 3877 3878 basic_block block, son; 3879 basic_block *worklist; 3880 size_t sp = 0; 3881 unsigned i; 3882 3883 /* We pretend that default definitions are defined in the entry block. 3884 This includes function arguments and the static chain decl. */ 3885 for (i = 1; i < num_ssa_names; ++i) 3886 { 3887 tree name = ssa_name (i); 3888 pre_expr e; 3889 if (!name 3890 || !SSA_NAME_IS_DEFAULT_DEF (name) 3891 || has_zero_uses (name) 3892 || !is_gimple_reg (name)) 3893 continue; 3894 3895 e = get_or_alloc_expr_for_name (name); 3896 add_to_value (get_expr_value_id (e), e); 3897 if (!in_fre) 3898 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e); 3899 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e); 3900 } 3901 3902 /* Allocate the worklist. */ 3903 worklist = XNEWVEC (basic_block, n_basic_blocks); 3904 3905 /* Seed the algorithm by putting the dominator children of the entry 3906 block on the worklist. */ 3907 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR); 3908 son; 3909 son = next_dom_son (CDI_DOMINATORS, son)) 3910 worklist[sp++] = son; 3911 3912 /* Loop until the worklist is empty. */ 3913 while (sp) 3914 { 3915 gimple_stmt_iterator gsi; 3916 gimple stmt; 3917 basic_block dom; 3918 unsigned int stmt_uid = 1; 3919 3920 /* Pick a block from the worklist. */ 3921 block = worklist[--sp]; 3922 3923 /* Initially, the set of available values in BLOCK is that of 3924 its immediate dominator. */ 3925 dom = get_immediate_dominator (CDI_DOMINATORS, block); 3926 if (dom) 3927 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom)); 3928 3929 /* Generate values for PHI nodes. */ 3930 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi)) 3931 make_values_for_phi (gsi_stmt (gsi), block); 3932 3933 BB_MAY_NOTRETURN (block) = 0; 3934 3935 /* Now compute value numbers and populate value sets with all 3936 the expressions computed in BLOCK. */ 3937 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi)) 3938 { 3939 ssa_op_iter iter; 3940 tree op; 3941 3942 stmt = gsi_stmt (gsi); 3943 gimple_set_uid (stmt, stmt_uid++); 3944 3945 /* Cache whether the basic-block has any non-visible side-effect 3946 or control flow. 3947 If this isn't a call or it is the last stmt in the 3948 basic-block then the CFG represents things correctly. */ 3949 if (is_gimple_call (stmt) && !stmt_ends_bb_p (stmt)) 3950 { 3951 /* Non-looping const functions always return normally. 3952 Otherwise the call might not return or have side-effects 3953 that forbids hoisting possibly trapping expressions 3954 before it. */ 3955 int flags = gimple_call_flags (stmt); 3956 if (!(flags & ECF_CONST) 3957 || (flags & ECF_LOOPING_CONST_OR_PURE)) 3958 BB_MAY_NOTRETURN (block) = 1; 3959 } 3960 3961 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF) 3962 { 3963 pre_expr e = get_or_alloc_expr_for_name (op); 3964 3965 add_to_value (get_expr_value_id (e), e); 3966 if (!in_fre) 3967 bitmap_insert_into_set (TMP_GEN (block), e); 3968 bitmap_value_insert_into_set (AVAIL_OUT (block), e); 3969 } 3970 3971 if (gimple_has_side_effects (stmt) || stmt_could_throw_p (stmt)) 3972 continue; 3973 3974 switch (gimple_code (stmt)) 3975 { 3976 case GIMPLE_RETURN: 3977 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE) 3978 add_to_exp_gen (block, op); 3979 continue; 3980 3981 case GIMPLE_CALL: 3982 { 3983 vn_reference_t ref; 3984 unsigned int i; 3985 vn_reference_op_t vro; 3986 pre_expr result = NULL; 3987 VEC(vn_reference_op_s, heap) *ops = NULL; 3988 3989 /* We can value number only calls to real functions. */ 3990 if (gimple_call_internal_p (stmt)) 3991 continue; 3992 3993 copy_reference_ops_from_call (stmt, &ops); 3994 vn_reference_lookup_pieces (gimple_vuse (stmt), 0, 3995 gimple_expr_type (stmt), 3996 ops, &ref, VN_NOWALK); 3997 VEC_free (vn_reference_op_s, heap, ops); 3998 if (!ref) 3999 continue; 4000 4001 for (i = 0; VEC_iterate (vn_reference_op_s, 4002 ref->operands, i, 4003 vro); i++) 4004 { 4005 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME) 4006 add_to_exp_gen (block, vro->op0); 4007 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME) 4008 add_to_exp_gen (block, vro->op1); 4009 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME) 4010 add_to_exp_gen (block, vro->op2); 4011 } 4012 result = (pre_expr) pool_alloc (pre_expr_pool); 4013 result->kind = REFERENCE; 4014 result->id = 0; 4015 PRE_EXPR_REFERENCE (result) = ref; 4016 4017 get_or_alloc_expression_id (result); 4018 add_to_value (get_expr_value_id (result), result); 4019 if (!in_fre) 4020 bitmap_value_insert_into_set (EXP_GEN (block), result); 4021 continue; 4022 } 4023 4024 case GIMPLE_ASSIGN: 4025 { 4026 pre_expr result = NULL; 4027 switch (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt))) 4028 { 4029 case tcc_unary: 4030 case tcc_binary: 4031 case tcc_comparison: 4032 { 4033 vn_nary_op_t nary; 4034 unsigned int i; 4035 4036 vn_nary_op_lookup_pieces (gimple_num_ops (stmt) - 1, 4037 gimple_assign_rhs_code (stmt), 4038 gimple_expr_type (stmt), 4039 gimple_assign_rhs1_ptr (stmt), 4040 &nary); 4041 4042 if (!nary) 4043 continue; 4044 4045 for (i = 0; i < nary->length; i++) 4046 if (TREE_CODE (nary->op[i]) == SSA_NAME) 4047 add_to_exp_gen (block, nary->op[i]); 4048 4049 result = (pre_expr) pool_alloc (pre_expr_pool); 4050 result->kind = NARY; 4051 result->id = 0; 4052 PRE_EXPR_NARY (result) = nary; 4053 break; 4054 } 4055 4056 case tcc_declaration: 4057 case tcc_reference: 4058 { 4059 vn_reference_t ref; 4060 unsigned int i; 4061 vn_reference_op_t vro; 4062 4063 vn_reference_lookup (gimple_assign_rhs1 (stmt), 4064 gimple_vuse (stmt), 4065 VN_WALK, &ref); 4066 if (!ref) 4067 continue; 4068 4069 for (i = 0; VEC_iterate (vn_reference_op_s, 4070 ref->operands, i, 4071 vro); i++) 4072 { 4073 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME) 4074 add_to_exp_gen (block, vro->op0); 4075 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME) 4076 add_to_exp_gen (block, vro->op1); 4077 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME) 4078 add_to_exp_gen (block, vro->op2); 4079 } 4080 result = (pre_expr) pool_alloc (pre_expr_pool); 4081 result->kind = REFERENCE; 4082 result->id = 0; 4083 PRE_EXPR_REFERENCE (result) = ref; 4084 break; 4085 } 4086 4087 default: 4088 /* For any other statement that we don't 4089 recognize, simply add all referenced 4090 SSA_NAMEs to EXP_GEN. */ 4091 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE) 4092 add_to_exp_gen (block, op); 4093 continue; 4094 } 4095 4096 get_or_alloc_expression_id (result); 4097 add_to_value (get_expr_value_id (result), result); 4098 if (!in_fre) 4099 bitmap_value_insert_into_set (EXP_GEN (block), result); 4100 4101 continue; 4102 } 4103 default: 4104 break; 4105 } 4106 } 4107 4108 /* Put the dominator children of BLOCK on the worklist of blocks 4109 to compute available sets for. */ 4110 for (son = first_dom_son (CDI_DOMINATORS, block); 4111 son; 4112 son = next_dom_son (CDI_DOMINATORS, son)) 4113 worklist[sp++] = son; 4114 } 4115 4116 free (worklist); 4117 } 4118 4119 /* Insert the expression for SSA_VN that SCCVN thought would be simpler 4120 than the available expressions for it. The insertion point is 4121 right before the first use in STMT. Returns the SSA_NAME that should 4122 be used for replacement. */ 4123 4124 static tree 4125 do_SCCVN_insertion (gimple stmt, tree ssa_vn) 4126 { 4127 basic_block bb = gimple_bb (stmt); 4128 gimple_stmt_iterator gsi; 4129 gimple_seq stmts = NULL; 4130 tree expr; 4131 pre_expr e; 4132 4133 /* First create a value expression from the expression we want 4134 to insert and associate it with the value handle for SSA_VN. */ 4135 e = get_or_alloc_expr_for (vn_get_expr_for (ssa_vn)); 4136 if (e == NULL) 4137 return NULL_TREE; 4138 4139 /* Then use create_expression_by_pieces to generate a valid 4140 expression to insert at this point of the IL stream. */ 4141 expr = create_expression_by_pieces (bb, e, &stmts, stmt, NULL); 4142 if (expr == NULL_TREE) 4143 return NULL_TREE; 4144 gsi = gsi_for_stmt (stmt); 4145 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT); 4146 4147 return expr; 4148 } 4149 4150 /* Eliminate fully redundant computations. */ 4151 4152 static unsigned int 4153 eliminate (void) 4154 { 4155 VEC (gimple, heap) *to_remove = NULL; 4156 VEC (gimple, heap) *to_update = NULL; 4157 basic_block b; 4158 unsigned int todo = 0; 4159 gimple_stmt_iterator gsi; 4160 gimple stmt; 4161 unsigned i; 4162 4163 FOR_EACH_BB (b) 4164 { 4165 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi)) 4166 { 4167 tree lhs = NULL_TREE; 4168 tree rhs = NULL_TREE; 4169 4170 stmt = gsi_stmt (gsi); 4171 4172 if (gimple_has_lhs (stmt)) 4173 lhs = gimple_get_lhs (stmt); 4174 4175 if (gimple_assign_single_p (stmt)) 4176 rhs = gimple_assign_rhs1 (stmt); 4177 4178 /* Lookup the RHS of the expression, see if we have an 4179 available computation for it. If so, replace the RHS with 4180 the available computation. 4181 4182 See PR43491. 4183 We don't replace global register variable when it is a the RHS of 4184 a single assign. We do replace local register variable since gcc 4185 does not guarantee local variable will be allocated in register. */ 4186 if (gimple_has_lhs (stmt) 4187 && TREE_CODE (lhs) == SSA_NAME 4188 && !gimple_assign_ssa_name_copy_p (stmt) 4189 && (!gimple_assign_single_p (stmt) 4190 || (!is_gimple_min_invariant (rhs) 4191 && (gimple_assign_rhs_code (stmt) != VAR_DECL 4192 || !is_global_var (rhs) 4193 || !DECL_HARD_REGISTER (rhs)))) 4194 && !gimple_has_volatile_ops (stmt) 4195 && !has_zero_uses (lhs)) 4196 { 4197 tree sprime = NULL; 4198 pre_expr lhsexpr = get_or_alloc_expr_for_name (lhs); 4199 pre_expr sprimeexpr; 4200 gimple orig_stmt = stmt; 4201 4202 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b), 4203 get_expr_value_id (lhsexpr), 4204 NULL); 4205 4206 if (sprimeexpr) 4207 { 4208 if (sprimeexpr->kind == CONSTANT) 4209 sprime = PRE_EXPR_CONSTANT (sprimeexpr); 4210 else if (sprimeexpr->kind == NAME) 4211 sprime = PRE_EXPR_NAME (sprimeexpr); 4212 else 4213 gcc_unreachable (); 4214 } 4215 4216 /* If there is no existing leader but SCCVN knows this 4217 value is constant, use that constant. */ 4218 if (!sprime && is_gimple_min_invariant (VN_INFO (lhs)->valnum)) 4219 { 4220 sprime = VN_INFO (lhs)->valnum; 4221 if (!useless_type_conversion_p (TREE_TYPE (lhs), 4222 TREE_TYPE (sprime))) 4223 sprime = fold_convert (TREE_TYPE (lhs), sprime); 4224 4225 if (dump_file && (dump_flags & TDF_DETAILS)) 4226 { 4227 fprintf (dump_file, "Replaced "); 4228 print_gimple_expr (dump_file, stmt, 0, 0); 4229 fprintf (dump_file, " with "); 4230 print_generic_expr (dump_file, sprime, 0); 4231 fprintf (dump_file, " in "); 4232 print_gimple_stmt (dump_file, stmt, 0, 0); 4233 } 4234 pre_stats.eliminations++; 4235 propagate_tree_value_into_stmt (&gsi, sprime); 4236 stmt = gsi_stmt (gsi); 4237 update_stmt (stmt); 4238 4239 /* If we removed EH side-effects from the statement, clean 4240 its EH information. */ 4241 if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt)) 4242 { 4243 bitmap_set_bit (need_eh_cleanup, 4244 gimple_bb (stmt)->index); 4245 if (dump_file && (dump_flags & TDF_DETAILS)) 4246 fprintf (dump_file, " Removed EH side-effects.\n"); 4247 } 4248 continue; 4249 } 4250 4251 /* If there is no existing usable leader but SCCVN thinks 4252 it has an expression it wants to use as replacement, 4253 insert that. */ 4254 if (!sprime || sprime == lhs) 4255 { 4256 tree val = VN_INFO (lhs)->valnum; 4257 if (val != VN_TOP 4258 && TREE_CODE (val) == SSA_NAME 4259 && VN_INFO (val)->needs_insertion 4260 && can_PRE_operation (vn_get_expr_for (val))) 4261 sprime = do_SCCVN_insertion (stmt, val); 4262 } 4263 if (sprime 4264 && sprime != lhs 4265 && (rhs == NULL_TREE 4266 || TREE_CODE (rhs) != SSA_NAME 4267 || may_propagate_copy (rhs, sprime))) 4268 { 4269 bool can_make_abnormal_goto 4270 = is_gimple_call (stmt) 4271 && stmt_can_make_abnormal_goto (stmt); 4272 4273 gcc_assert (sprime != rhs); 4274 4275 if (dump_file && (dump_flags & TDF_DETAILS)) 4276 { 4277 fprintf (dump_file, "Replaced "); 4278 print_gimple_expr (dump_file, stmt, 0, 0); 4279 fprintf (dump_file, " with "); 4280 print_generic_expr (dump_file, sprime, 0); 4281 fprintf (dump_file, " in "); 4282 print_gimple_stmt (dump_file, stmt, 0, 0); 4283 } 4284 4285 if (TREE_CODE (sprime) == SSA_NAME) 4286 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), 4287 NECESSARY, true); 4288 /* We need to make sure the new and old types actually match, 4289 which may require adding a simple cast, which fold_convert 4290 will do for us. */ 4291 if ((!rhs || TREE_CODE (rhs) != SSA_NAME) 4292 && !useless_type_conversion_p (gimple_expr_type (stmt), 4293 TREE_TYPE (sprime))) 4294 sprime = fold_convert (gimple_expr_type (stmt), sprime); 4295 4296 pre_stats.eliminations++; 4297 propagate_tree_value_into_stmt (&gsi, sprime); 4298 stmt = gsi_stmt (gsi); 4299 update_stmt (stmt); 4300 4301 /* If we removed EH side-effects from the statement, clean 4302 its EH information. */ 4303 if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt)) 4304 { 4305 bitmap_set_bit (need_eh_cleanup, 4306 gimple_bb (stmt)->index); 4307 if (dump_file && (dump_flags & TDF_DETAILS)) 4308 fprintf (dump_file, " Removed EH side-effects.\n"); 4309 } 4310 4311 /* Likewise for AB side-effects. */ 4312 if (can_make_abnormal_goto 4313 && !stmt_can_make_abnormal_goto (stmt)) 4314 { 4315 bitmap_set_bit (need_ab_cleanup, 4316 gimple_bb (stmt)->index); 4317 if (dump_file && (dump_flags & TDF_DETAILS)) 4318 fprintf (dump_file, " Removed AB side-effects.\n"); 4319 } 4320 } 4321 } 4322 /* If the statement is a scalar store, see if the expression 4323 has the same value number as its rhs. If so, the store is 4324 dead. */ 4325 else if (gimple_assign_single_p (stmt) 4326 && !gimple_has_volatile_ops (stmt) 4327 && !is_gimple_reg (gimple_assign_lhs (stmt)) 4328 && (TREE_CODE (rhs) == SSA_NAME 4329 || is_gimple_min_invariant (rhs))) 4330 { 4331 tree val; 4332 val = vn_reference_lookup (gimple_assign_lhs (stmt), 4333 gimple_vuse (stmt), VN_WALK, NULL); 4334 if (TREE_CODE (rhs) == SSA_NAME) 4335 rhs = VN_INFO (rhs)->valnum; 4336 if (val 4337 && operand_equal_p (val, rhs, 0)) 4338 { 4339 if (dump_file && (dump_flags & TDF_DETAILS)) 4340 { 4341 fprintf (dump_file, "Deleted redundant store "); 4342 print_gimple_stmt (dump_file, stmt, 0, 0); 4343 } 4344 4345 /* Queue stmt for removal. */ 4346 VEC_safe_push (gimple, heap, to_remove, stmt); 4347 } 4348 } 4349 /* Visit COND_EXPRs and fold the comparison with the 4350 available value-numbers. */ 4351 else if (gimple_code (stmt) == GIMPLE_COND) 4352 { 4353 tree op0 = gimple_cond_lhs (stmt); 4354 tree op1 = gimple_cond_rhs (stmt); 4355 tree result; 4356 4357 if (TREE_CODE (op0) == SSA_NAME) 4358 op0 = VN_INFO (op0)->valnum; 4359 if (TREE_CODE (op1) == SSA_NAME) 4360 op1 = VN_INFO (op1)->valnum; 4361 result = fold_binary (gimple_cond_code (stmt), boolean_type_node, 4362 op0, op1); 4363 if (result && TREE_CODE (result) == INTEGER_CST) 4364 { 4365 if (integer_zerop (result)) 4366 gimple_cond_make_false (stmt); 4367 else 4368 gimple_cond_make_true (stmt); 4369 update_stmt (stmt); 4370 todo = TODO_cleanup_cfg; 4371 } 4372 } 4373 /* Visit indirect calls and turn them into direct calls if 4374 possible. */ 4375 if (is_gimple_call (stmt)) 4376 { 4377 tree orig_fn = gimple_call_fn (stmt); 4378 tree fn; 4379 if (!orig_fn) 4380 continue; 4381 if (TREE_CODE (orig_fn) == SSA_NAME) 4382 fn = VN_INFO (orig_fn)->valnum; 4383 else if (TREE_CODE (orig_fn) == OBJ_TYPE_REF 4384 && TREE_CODE (OBJ_TYPE_REF_EXPR (orig_fn)) == SSA_NAME) 4385 fn = VN_INFO (OBJ_TYPE_REF_EXPR (orig_fn))->valnum; 4386 else 4387 continue; 4388 if (gimple_call_addr_fndecl (fn) != NULL_TREE 4389 && useless_type_conversion_p (TREE_TYPE (orig_fn), 4390 TREE_TYPE (fn))) 4391 { 4392 bool can_make_abnormal_goto 4393 = stmt_can_make_abnormal_goto (stmt); 4394 bool was_noreturn = gimple_call_noreturn_p (stmt); 4395 4396 if (dump_file && (dump_flags & TDF_DETAILS)) 4397 { 4398 fprintf (dump_file, "Replacing call target with "); 4399 print_generic_expr (dump_file, fn, 0); 4400 fprintf (dump_file, " in "); 4401 print_gimple_stmt (dump_file, stmt, 0, 0); 4402 } 4403 4404 gimple_call_set_fn (stmt, fn); 4405 VEC_safe_push (gimple, heap, to_update, stmt); 4406 4407 /* When changing a call into a noreturn call, cfg cleanup 4408 is needed to fix up the noreturn call. */ 4409 if (!was_noreturn && gimple_call_noreturn_p (stmt)) 4410 todo |= TODO_cleanup_cfg; 4411 4412 /* If we removed EH side-effects from the statement, clean 4413 its EH information. */ 4414 if (maybe_clean_or_replace_eh_stmt (stmt, stmt)) 4415 { 4416 bitmap_set_bit (need_eh_cleanup, 4417 gimple_bb (stmt)->index); 4418 if (dump_file && (dump_flags & TDF_DETAILS)) 4419 fprintf (dump_file, " Removed EH side-effects.\n"); 4420 } 4421 4422 /* Likewise for AB side-effects. */ 4423 if (can_make_abnormal_goto 4424 && !stmt_can_make_abnormal_goto (stmt)) 4425 { 4426 bitmap_set_bit (need_ab_cleanup, 4427 gimple_bb (stmt)->index); 4428 if (dump_file && (dump_flags & TDF_DETAILS)) 4429 fprintf (dump_file, " Removed AB side-effects.\n"); 4430 } 4431 4432 /* Changing an indirect call to a direct call may 4433 have exposed different semantics. This may 4434 require an SSA update. */ 4435 todo |= TODO_update_ssa_only_virtuals; 4436 } 4437 } 4438 } 4439 4440 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);) 4441 { 4442 gimple stmt, phi = gsi_stmt (gsi); 4443 tree sprime = NULL_TREE, res = PHI_RESULT (phi); 4444 pre_expr sprimeexpr, resexpr; 4445 gimple_stmt_iterator gsi2; 4446 4447 /* We want to perform redundant PHI elimination. Do so by 4448 replacing the PHI with a single copy if possible. 4449 Do not touch inserted, single-argument or virtual PHIs. */ 4450 if (gimple_phi_num_args (phi) == 1 4451 || !is_gimple_reg (res)) 4452 { 4453 gsi_next (&gsi); 4454 continue; 4455 } 4456 4457 resexpr = get_or_alloc_expr_for_name (res); 4458 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b), 4459 get_expr_value_id (resexpr), NULL); 4460 if (sprimeexpr) 4461 { 4462 if (sprimeexpr->kind == CONSTANT) 4463 sprime = PRE_EXPR_CONSTANT (sprimeexpr); 4464 else if (sprimeexpr->kind == NAME) 4465 sprime = PRE_EXPR_NAME (sprimeexpr); 4466 else 4467 gcc_unreachable (); 4468 } 4469 if (!sprime && is_gimple_min_invariant (VN_INFO (res)->valnum)) 4470 { 4471 sprime = VN_INFO (res)->valnum; 4472 if (!useless_type_conversion_p (TREE_TYPE (res), 4473 TREE_TYPE (sprime))) 4474 sprime = fold_convert (TREE_TYPE (res), sprime); 4475 } 4476 if (!sprime 4477 || sprime == res) 4478 { 4479 gsi_next (&gsi); 4480 continue; 4481 } 4482 4483 if (dump_file && (dump_flags & TDF_DETAILS)) 4484 { 4485 fprintf (dump_file, "Replaced redundant PHI node defining "); 4486 print_generic_expr (dump_file, res, 0); 4487 fprintf (dump_file, " with "); 4488 print_generic_expr (dump_file, sprime, 0); 4489 fprintf (dump_file, "\n"); 4490 } 4491 4492 remove_phi_node (&gsi, false); 4493 4494 if (!bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)) 4495 && TREE_CODE (sprime) == SSA_NAME) 4496 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true); 4497 4498 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime))) 4499 sprime = fold_convert (TREE_TYPE (res), sprime); 4500 stmt = gimple_build_assign (res, sprime); 4501 SSA_NAME_DEF_STMT (res) = stmt; 4502 gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY)); 4503 4504 gsi2 = gsi_after_labels (b); 4505 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT); 4506 /* Queue the copy for eventual removal. */ 4507 VEC_safe_push (gimple, heap, to_remove, stmt); 4508 /* If we inserted this PHI node ourself, it's not an elimination. */ 4509 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))) 4510 pre_stats.phis--; 4511 else 4512 pre_stats.eliminations++; 4513 } 4514 } 4515 4516 /* We cannot remove stmts during BB walk, especially not release SSA 4517 names there as this confuses the VN machinery. The stmts ending 4518 up in to_remove are either stores or simple copies. */ 4519 FOR_EACH_VEC_ELT (gimple, to_remove, i, stmt) 4520 { 4521 tree lhs = gimple_assign_lhs (stmt); 4522 tree rhs = gimple_assign_rhs1 (stmt); 4523 use_operand_p use_p; 4524 gimple use_stmt; 4525 4526 /* If there is a single use only, propagate the equivalency 4527 instead of keeping the copy. */ 4528 if (TREE_CODE (lhs) == SSA_NAME 4529 && TREE_CODE (rhs) == SSA_NAME 4530 && single_imm_use (lhs, &use_p, &use_stmt) 4531 && may_propagate_copy (USE_FROM_PTR (use_p), rhs)) 4532 { 4533 SET_USE (use_p, rhs); 4534 update_stmt (use_stmt); 4535 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (lhs)) 4536 && TREE_CODE (rhs) == SSA_NAME) 4537 gimple_set_plf (SSA_NAME_DEF_STMT (rhs), NECESSARY, true); 4538 } 4539 4540 /* If this is a store or a now unused copy, remove it. */ 4541 if (TREE_CODE (lhs) != SSA_NAME 4542 || has_zero_uses (lhs)) 4543 { 4544 basic_block bb = gimple_bb (stmt); 4545 gsi = gsi_for_stmt (stmt); 4546 unlink_stmt_vdef (stmt); 4547 gsi_remove (&gsi, true); 4548 /* ??? gsi_remove doesn't tell us whether the stmt was 4549 in EH tables and thus whether we need to purge EH edges. 4550 Simply schedule the block for a cleanup. */ 4551 bitmap_set_bit (need_eh_cleanup, bb->index); 4552 if (TREE_CODE (lhs) == SSA_NAME) 4553 bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs)); 4554 release_defs (stmt); 4555 } 4556 } 4557 VEC_free (gimple, heap, to_remove); 4558 4559 /* We cannot update call statements with virtual operands during 4560 SSA walk. This might remove them which in turn makes our 4561 VN lattice invalid. */ 4562 FOR_EACH_VEC_ELT (gimple, to_update, i, stmt) 4563 update_stmt (stmt); 4564 VEC_free (gimple, heap, to_update); 4565 4566 return todo; 4567 } 4568 4569 /* Borrow a bit of tree-ssa-dce.c for the moment. 4570 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though 4571 this may be a bit faster, and we may want critical edges kept split. */ 4572 4573 /* If OP's defining statement has not already been determined to be necessary, 4574 mark that statement necessary. Return the stmt, if it is newly 4575 necessary. */ 4576 4577 static inline gimple 4578 mark_operand_necessary (tree op) 4579 { 4580 gimple stmt; 4581 4582 gcc_assert (op); 4583 4584 if (TREE_CODE (op) != SSA_NAME) 4585 return NULL; 4586 4587 stmt = SSA_NAME_DEF_STMT (op); 4588 gcc_assert (stmt); 4589 4590 if (gimple_plf (stmt, NECESSARY) 4591 || gimple_nop_p (stmt)) 4592 return NULL; 4593 4594 gimple_set_plf (stmt, NECESSARY, true); 4595 return stmt; 4596 } 4597 4598 /* Because we don't follow exactly the standard PRE algorithm, and decide not 4599 to insert PHI nodes sometimes, and because value numbering of casts isn't 4600 perfect, we sometimes end up inserting dead code. This simple DCE-like 4601 pass removes any insertions we made that weren't actually used. */ 4602 4603 static void 4604 remove_dead_inserted_code (void) 4605 { 4606 bitmap worklist; 4607 unsigned i; 4608 bitmap_iterator bi; 4609 gimple t; 4610 4611 worklist = BITMAP_ALLOC (NULL); 4612 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi) 4613 { 4614 t = SSA_NAME_DEF_STMT (ssa_name (i)); 4615 if (gimple_plf (t, NECESSARY)) 4616 bitmap_set_bit (worklist, i); 4617 } 4618 while (!bitmap_empty_p (worklist)) 4619 { 4620 i = bitmap_first_set_bit (worklist); 4621 bitmap_clear_bit (worklist, i); 4622 t = SSA_NAME_DEF_STMT (ssa_name (i)); 4623 4624 /* PHI nodes are somewhat special in that each PHI alternative has 4625 data and control dependencies. All the statements feeding the 4626 PHI node's arguments are always necessary. */ 4627 if (gimple_code (t) == GIMPLE_PHI) 4628 { 4629 unsigned k; 4630 4631 for (k = 0; k < gimple_phi_num_args (t); k++) 4632 { 4633 tree arg = PHI_ARG_DEF (t, k); 4634 if (TREE_CODE (arg) == SSA_NAME) 4635 { 4636 gimple n = mark_operand_necessary (arg); 4637 if (n) 4638 bitmap_set_bit (worklist, SSA_NAME_VERSION (arg)); 4639 } 4640 } 4641 } 4642 else 4643 { 4644 /* Propagate through the operands. Examine all the USE, VUSE and 4645 VDEF operands in this statement. Mark all the statements 4646 which feed this statement's uses as necessary. */ 4647 ssa_op_iter iter; 4648 tree use; 4649 4650 /* The operands of VDEF expressions are also needed as they 4651 represent potential definitions that may reach this 4652 statement (VDEF operands allow us to follow def-def 4653 links). */ 4654 4655 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES) 4656 { 4657 gimple n = mark_operand_necessary (use); 4658 if (n) 4659 bitmap_set_bit (worklist, SSA_NAME_VERSION (use)); 4660 } 4661 } 4662 } 4663 4664 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi) 4665 { 4666 t = SSA_NAME_DEF_STMT (ssa_name (i)); 4667 if (!gimple_plf (t, NECESSARY)) 4668 { 4669 gimple_stmt_iterator gsi; 4670 4671 if (dump_file && (dump_flags & TDF_DETAILS)) 4672 { 4673 fprintf (dump_file, "Removing unnecessary insertion:"); 4674 print_gimple_stmt (dump_file, t, 0, 0); 4675 } 4676 4677 gsi = gsi_for_stmt (t); 4678 if (gimple_code (t) == GIMPLE_PHI) 4679 remove_phi_node (&gsi, true); 4680 else 4681 { 4682 gsi_remove (&gsi, true); 4683 release_defs (t); 4684 } 4685 } 4686 } 4687 BITMAP_FREE (worklist); 4688 } 4689 4690 /* Compute a reverse post-order in *POST_ORDER. If INCLUDE_ENTRY_EXIT is 4691 true, then then ENTRY_BLOCK and EXIT_BLOCK are included. Returns 4692 the number of visited blocks. */ 4693 4694 static int 4695 my_rev_post_order_compute (int *post_order, bool include_entry_exit) 4696 { 4697 edge_iterator *stack; 4698 int sp; 4699 int post_order_num = 0; 4700 sbitmap visited; 4701 4702 if (include_entry_exit) 4703 post_order[post_order_num++] = EXIT_BLOCK; 4704 4705 /* Allocate stack for back-tracking up CFG. */ 4706 stack = XNEWVEC (edge_iterator, n_basic_blocks + 1); 4707 sp = 0; 4708 4709 /* Allocate bitmap to track nodes that have been visited. */ 4710 visited = sbitmap_alloc (last_basic_block); 4711 4712 /* None of the nodes in the CFG have been visited yet. */ 4713 sbitmap_zero (visited); 4714 4715 /* Push the last edge on to the stack. */ 4716 stack[sp++] = ei_start (EXIT_BLOCK_PTR->preds); 4717 4718 while (sp) 4719 { 4720 edge_iterator ei; 4721 basic_block src; 4722 basic_block dest; 4723 4724 /* Look at the edge on the top of the stack. */ 4725 ei = stack[sp - 1]; 4726 src = ei_edge (ei)->src; 4727 dest = ei_edge (ei)->dest; 4728 4729 /* Check if the edge destination has been visited yet. */ 4730 if (src != ENTRY_BLOCK_PTR && ! TEST_BIT (visited, src->index)) 4731 { 4732 /* Mark that we have visited the destination. */ 4733 SET_BIT (visited, src->index); 4734 4735 if (EDGE_COUNT (src->preds) > 0) 4736 /* Since the DEST node has been visited for the first 4737 time, check its successors. */ 4738 stack[sp++] = ei_start (src->preds); 4739 else 4740 post_order[post_order_num++] = src->index; 4741 } 4742 else 4743 { 4744 if (ei_one_before_end_p (ei) && dest != EXIT_BLOCK_PTR) 4745 post_order[post_order_num++] = dest->index; 4746 4747 if (!ei_one_before_end_p (ei)) 4748 ei_next (&stack[sp - 1]); 4749 else 4750 sp--; 4751 } 4752 } 4753 4754 if (include_entry_exit) 4755 post_order[post_order_num++] = ENTRY_BLOCK; 4756 4757 free (stack); 4758 sbitmap_free (visited); 4759 return post_order_num; 4760 } 4761 4762 4763 /* Initialize data structures used by PRE. */ 4764 4765 static void 4766 init_pre (bool do_fre) 4767 { 4768 basic_block bb; 4769 4770 next_expression_id = 1; 4771 expressions = NULL; 4772 VEC_safe_push (pre_expr, heap, expressions, NULL); 4773 value_expressions = VEC_alloc (bitmap_set_t, heap, get_max_value_id () + 1); 4774 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions, 4775 get_max_value_id() + 1); 4776 name_to_id = NULL; 4777 4778 in_fre = do_fre; 4779 4780 inserted_exprs = BITMAP_ALLOC (NULL); 4781 need_creation = NULL; 4782 pretemp = NULL_TREE; 4783 storetemp = NULL_TREE; 4784 prephitemp = NULL_TREE; 4785 4786 connect_infinite_loops_to_exit (); 4787 memset (&pre_stats, 0, sizeof (pre_stats)); 4788 4789 4790 postorder = XNEWVEC (int, n_basic_blocks - NUM_FIXED_BLOCKS); 4791 my_rev_post_order_compute (postorder, false); 4792 4793 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets)); 4794 4795 calculate_dominance_info (CDI_POST_DOMINATORS); 4796 calculate_dominance_info (CDI_DOMINATORS); 4797 4798 bitmap_obstack_initialize (&grand_bitmap_obstack); 4799 phi_translate_table = htab_create (5110, expr_pred_trans_hash, 4800 expr_pred_trans_eq, free); 4801 expression_to_id = htab_create (num_ssa_names * 3, 4802 pre_expr_hash, 4803 pre_expr_eq, NULL); 4804 bitmap_set_pool = create_alloc_pool ("Bitmap sets", 4805 sizeof (struct bitmap_set), 30); 4806 pre_expr_pool = create_alloc_pool ("pre_expr nodes", 4807 sizeof (struct pre_expr_d), 30); 4808 FOR_ALL_BB (bb) 4809 { 4810 EXP_GEN (bb) = bitmap_set_new (); 4811 PHI_GEN (bb) = bitmap_set_new (); 4812 TMP_GEN (bb) = bitmap_set_new (); 4813 AVAIL_OUT (bb) = bitmap_set_new (); 4814 } 4815 4816 need_eh_cleanup = BITMAP_ALLOC (NULL); 4817 need_ab_cleanup = BITMAP_ALLOC (NULL); 4818 } 4819 4820 4821 /* Deallocate data structures used by PRE. */ 4822 4823 static unsigned 4824 fini_pre (bool do_fre) 4825 { 4826 bool do_eh_cleanup = !bitmap_empty_p (need_eh_cleanup); 4827 bool do_ab_cleanup = !bitmap_empty_p (need_ab_cleanup); 4828 unsigned todo = 0; 4829 4830 free (postorder); 4831 VEC_free (bitmap_set_t, heap, value_expressions); 4832 BITMAP_FREE (inserted_exprs); 4833 VEC_free (gimple, heap, need_creation); 4834 bitmap_obstack_release (&grand_bitmap_obstack); 4835 free_alloc_pool (bitmap_set_pool); 4836 free_alloc_pool (pre_expr_pool); 4837 htab_delete (phi_translate_table); 4838 htab_delete (expression_to_id); 4839 VEC_free (unsigned, heap, name_to_id); 4840 4841 free_aux_for_blocks (); 4842 4843 free_dominance_info (CDI_POST_DOMINATORS); 4844 4845 if (do_eh_cleanup) 4846 gimple_purge_all_dead_eh_edges (need_eh_cleanup); 4847 4848 if (do_ab_cleanup) 4849 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup); 4850 4851 BITMAP_FREE (need_eh_cleanup); 4852 BITMAP_FREE (need_ab_cleanup); 4853 4854 if (do_eh_cleanup || do_ab_cleanup) 4855 todo = TODO_cleanup_cfg; 4856 4857 if (!do_fre) 4858 loop_optimizer_finalize (); 4859 4860 return todo; 4861 } 4862 4863 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller 4864 only wants to do full redundancy elimination. */ 4865 4866 static unsigned int 4867 execute_pre (bool do_fre) 4868 { 4869 unsigned int todo = 0; 4870 4871 do_partial_partial = optimize > 2 && optimize_function_for_speed_p (cfun); 4872 4873 /* This has to happen before SCCVN runs because 4874 loop_optimizer_init may create new phis, etc. */ 4875 if (!do_fre) 4876 loop_optimizer_init (LOOPS_NORMAL); 4877 4878 if (!run_scc_vn (do_fre ? VN_WALKREWRITE : VN_WALK)) 4879 { 4880 if (!do_fre) 4881 loop_optimizer_finalize (); 4882 4883 return 0; 4884 } 4885 4886 init_pre (do_fre); 4887 scev_initialize (); 4888 4889 /* Collect and value number expressions computed in each basic block. */ 4890 compute_avail (); 4891 4892 if (dump_file && (dump_flags & TDF_DETAILS)) 4893 { 4894 basic_block bb; 4895 4896 FOR_ALL_BB (bb) 4897 { 4898 print_bitmap_set (dump_file, EXP_GEN (bb), "exp_gen", bb->index); 4899 print_bitmap_set (dump_file, PHI_GEN (bb), "phi_gen", bb->index); 4900 print_bitmap_set (dump_file, TMP_GEN (bb), "tmp_gen", bb->index); 4901 print_bitmap_set (dump_file, AVAIL_OUT (bb), "avail_out", bb->index); 4902 } 4903 } 4904 4905 /* Insert can get quite slow on an incredibly large number of basic 4906 blocks due to some quadratic behavior. Until this behavior is 4907 fixed, don't run it when he have an incredibly large number of 4908 bb's. If we aren't going to run insert, there is no point in 4909 computing ANTIC, either, even though it's plenty fast. */ 4910 if (!do_fre && n_basic_blocks < 4000) 4911 { 4912 compute_antic (); 4913 insert (); 4914 } 4915 4916 /* Make sure to remove fake edges before committing our inserts. 4917 This makes sure we don't end up with extra critical edges that 4918 we would need to split. */ 4919 remove_fake_exit_edges (); 4920 gsi_commit_edge_inserts (); 4921 4922 /* Remove all the redundant expressions. */ 4923 todo |= eliminate (); 4924 4925 statistics_counter_event (cfun, "Insertions", pre_stats.insertions); 4926 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert); 4927 statistics_counter_event (cfun, "New PHIs", pre_stats.phis); 4928 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations); 4929 4930 clear_expression_ids (); 4931 if (!do_fre) 4932 { 4933 remove_dead_inserted_code (); 4934 todo |= TODO_verify_flow; 4935 } 4936 4937 scev_finalize (); 4938 todo |= fini_pre (do_fre); 4939 4940 if (!do_fre) 4941 /* TODO: tail_merge_optimize may merge all predecessors of a block, in which 4942 case we can merge the block with the remaining predecessor of the block. 4943 It should either: 4944 - call merge_blocks after each tail merge iteration 4945 - call merge_blocks after all tail merge iterations 4946 - mark TODO_cleanup_cfg when necessary 4947 - share the cfg cleanup with fini_pre. */ 4948 todo |= tail_merge_optimize (todo); 4949 free_scc_vn (); 4950 4951 /* Tail merging invalidates the virtual SSA web, together with 4952 cfg-cleanup opportunities exposed by PRE this will wreck the 4953 SSA updating machinery. So make sure to run update-ssa 4954 manually, before eventually scheduling cfg-cleanup as part of 4955 the todo. */ 4956 update_ssa (TODO_update_ssa_only_virtuals); 4957 4958 return todo; 4959 } 4960 4961 /* Gate and execute functions for PRE. */ 4962 4963 static unsigned int 4964 do_pre (void) 4965 { 4966 return execute_pre (false); 4967 } 4968 4969 static bool 4970 gate_pre (void) 4971 { 4972 return flag_tree_pre != 0; 4973 } 4974 4975 struct gimple_opt_pass pass_pre = 4976 { 4977 { 4978 GIMPLE_PASS, 4979 "pre", /* name */ 4980 gate_pre, /* gate */ 4981 do_pre, /* execute */ 4982 NULL, /* sub */ 4983 NULL, /* next */ 4984 0, /* static_pass_number */ 4985 TV_TREE_PRE, /* tv_id */ 4986 PROP_no_crit_edges | PROP_cfg 4987 | PROP_ssa, /* properties_required */ 4988 0, /* properties_provided */ 4989 0, /* properties_destroyed */ 4990 TODO_rebuild_alias, /* todo_flags_start */ 4991 TODO_ggc_collect | TODO_verify_ssa /* todo_flags_finish */ 4992 } 4993 }; 4994 4995 4996 /* Gate and execute functions for FRE. */ 4997 4998 static unsigned int 4999 execute_fre (void) 5000 { 5001 return execute_pre (true); 5002 } 5003 5004 static bool 5005 gate_fre (void) 5006 { 5007 return flag_tree_fre != 0; 5008 } 5009 5010 struct gimple_opt_pass pass_fre = 5011 { 5012 { 5013 GIMPLE_PASS, 5014 "fre", /* name */ 5015 gate_fre, /* gate */ 5016 execute_fre, /* execute */ 5017 NULL, /* sub */ 5018 NULL, /* next */ 5019 0, /* static_pass_number */ 5020 TV_TREE_FRE, /* tv_id */ 5021 PROP_cfg | PROP_ssa, /* properties_required */ 5022 0, /* properties_provided */ 5023 0, /* properties_destroyed */ 5024 0, /* todo_flags_start */ 5025 TODO_ggc_collect | TODO_verify_ssa /* todo_flags_finish */ 5026 } 5027 }; 5028