1 /* SSA Dominator optimizations for trees 2 Copyright (C) 2001-2017 Free Software Foundation, Inc. 3 Contributed by Diego Novillo <dnovillo@redhat.com> 4 5 This file is part of GCC. 6 7 GCC is free software; you can redistribute it and/or modify 8 it under the terms of the GNU General Public License as published by 9 the Free Software Foundation; either version 3, or (at your option) 10 any later version. 11 12 GCC is distributed in the hope that it will be useful, 13 but WITHOUT ANY WARRANTY; without even the implied warranty of 14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 15 GNU General Public License for more details. 16 17 You should have received a copy of the GNU General Public License 18 along with GCC; see the file COPYING3. If not see 19 <http://www.gnu.org/licenses/>. */ 20 21 #include "config.h" 22 #include "system.h" 23 #include "coretypes.h" 24 #include "backend.h" 25 #include "tree.h" 26 #include "gimple.h" 27 #include "tree-pass.h" 28 #include "ssa.h" 29 #include "gimple-pretty-print.h" 30 #include "fold-const.h" 31 #include "cfganal.h" 32 #include "cfgloop.h" 33 #include "gimple-fold.h" 34 #include "tree-eh.h" 35 #include "gimple-iterator.h" 36 #include "tree-cfg.h" 37 #include "tree-into-ssa.h" 38 #include "domwalk.h" 39 #include "tree-ssa-propagate.h" 40 #include "tree-ssa-threadupdate.h" 41 #include "params.h" 42 #include "tree-ssa-scopedtables.h" 43 #include "tree-ssa-threadedge.h" 44 #include "tree-ssa-dom.h" 45 #include "gimplify.h" 46 #include "tree-cfgcleanup.h" 47 #include "dbgcnt.h" 48 49 /* This file implements optimizations on the dominator tree. */ 50 51 /* Structure for recording edge equivalences. 52 53 Computing and storing the edge equivalences instead of creating 54 them on-demand can save significant amounts of time, particularly 55 for pathological cases involving switch statements. 56 57 These structures live for a single iteration of the dominator 58 optimizer in the edge's AUX field. At the end of an iteration we 59 free each of these structures. */ 60 61 struct edge_info 62 { 63 /* If this edge creates a simple equivalence, the LHS and RHS of 64 the equivalence will be stored here. */ 65 tree lhs; 66 tree rhs; 67 68 /* Traversing an edge may also indicate one or more particular conditions 69 are true or false. */ 70 vec<cond_equivalence> cond_equivalences; 71 }; 72 73 /* Track whether or not we have changed the control flow graph. */ 74 static bool cfg_altered; 75 76 /* Bitmap of blocks that have had EH statements cleaned. We should 77 remove their dead edges eventually. */ 78 static bitmap need_eh_cleanup; 79 static vec<gimple *> need_noreturn_fixup; 80 81 /* Statistics for dominator optimizations. */ 82 struct opt_stats_d 83 { 84 long num_stmts; 85 long num_exprs_considered; 86 long num_re; 87 long num_const_prop; 88 long num_copy_prop; 89 }; 90 91 static struct opt_stats_d opt_stats; 92 93 /* Local functions. */ 94 static edge optimize_stmt (basic_block, gimple_stmt_iterator, 95 class const_and_copies *, 96 class avail_exprs_stack *); 97 static void record_equality (tree, tree, class const_and_copies *); 98 static void record_equivalences_from_phis (basic_block); 99 static void record_equivalences_from_incoming_edge (basic_block, 100 class const_and_copies *, 101 class avail_exprs_stack *); 102 static void eliminate_redundant_computations (gimple_stmt_iterator *, 103 class const_and_copies *, 104 class avail_exprs_stack *); 105 static void record_equivalences_from_stmt (gimple *, int, 106 class avail_exprs_stack *); 107 static edge single_incoming_edge_ignoring_loop_edges (basic_block); 108 static void dump_dominator_optimization_stats (FILE *file, 109 hash_table<expr_elt_hasher> *); 110 111 112 /* Free the edge_info data attached to E, if it exists. */ 113 114 void 115 free_dom_edge_info (edge e) 116 { 117 struct edge_info *edge_info = (struct edge_info *)e->aux; 118 119 if (edge_info) 120 { 121 edge_info->cond_equivalences.release (); 122 free (edge_info); 123 } 124 } 125 126 /* Allocate an EDGE_INFO for edge E and attach it to E. 127 Return the new EDGE_INFO structure. */ 128 129 static struct edge_info * 130 allocate_edge_info (edge e) 131 { 132 struct edge_info *edge_info; 133 134 /* Free the old one, if it exists. */ 135 free_dom_edge_info (e); 136 137 edge_info = XCNEW (struct edge_info); 138 139 e->aux = edge_info; 140 return edge_info; 141 } 142 143 /* Free all EDGE_INFO structures associated with edges in the CFG. 144 If a particular edge can be threaded, copy the redirection 145 target from the EDGE_INFO structure into the edge's AUX field 146 as required by code to update the CFG and SSA graph for 147 jump threading. */ 148 149 static void 150 free_all_edge_infos (void) 151 { 152 basic_block bb; 153 edge_iterator ei; 154 edge e; 155 156 FOR_EACH_BB_FN (bb, cfun) 157 { 158 FOR_EACH_EDGE (e, ei, bb->preds) 159 { 160 free_dom_edge_info (e); 161 e->aux = NULL; 162 } 163 } 164 } 165 166 /* We have finished optimizing BB, record any information implied by 167 taking a specific outgoing edge from BB. */ 168 169 static void 170 record_edge_info (basic_block bb) 171 { 172 gimple_stmt_iterator gsi = gsi_last_bb (bb); 173 struct edge_info *edge_info; 174 175 if (! gsi_end_p (gsi)) 176 { 177 gimple *stmt = gsi_stmt (gsi); 178 location_t loc = gimple_location (stmt); 179 180 if (gimple_code (stmt) == GIMPLE_SWITCH) 181 { 182 gswitch *switch_stmt = as_a <gswitch *> (stmt); 183 tree index = gimple_switch_index (switch_stmt); 184 185 if (TREE_CODE (index) == SSA_NAME) 186 { 187 int i; 188 int n_labels = gimple_switch_num_labels (switch_stmt); 189 tree *info = XCNEWVEC (tree, last_basic_block_for_fn (cfun)); 190 edge e; 191 edge_iterator ei; 192 193 for (i = 0; i < n_labels; i++) 194 { 195 tree label = gimple_switch_label (switch_stmt, i); 196 basic_block target_bb = label_to_block (CASE_LABEL (label)); 197 if (CASE_HIGH (label) 198 || !CASE_LOW (label) 199 || info[target_bb->index]) 200 info[target_bb->index] = error_mark_node; 201 else 202 info[target_bb->index] = label; 203 } 204 205 FOR_EACH_EDGE (e, ei, bb->succs) 206 { 207 basic_block target_bb = e->dest; 208 tree label = info[target_bb->index]; 209 210 if (label != NULL && label != error_mark_node) 211 { 212 tree x = fold_convert_loc (loc, TREE_TYPE (index), 213 CASE_LOW (label)); 214 edge_info = allocate_edge_info (e); 215 edge_info->lhs = index; 216 edge_info->rhs = x; 217 } 218 } 219 free (info); 220 } 221 } 222 223 /* A COND_EXPR may create equivalences too. */ 224 if (gimple_code (stmt) == GIMPLE_COND) 225 { 226 edge true_edge; 227 edge false_edge; 228 229 tree op0 = gimple_cond_lhs (stmt); 230 tree op1 = gimple_cond_rhs (stmt); 231 enum tree_code code = gimple_cond_code (stmt); 232 233 extract_true_false_edges_from_block (bb, &true_edge, &false_edge); 234 235 /* Special case comparing booleans against a constant as we 236 know the value of OP0 on both arms of the branch. i.e., we 237 can record an equivalence for OP0 rather than COND. 238 239 However, don't do this if the constant isn't zero or one. 240 Such conditionals will get optimized more thoroughly during 241 the domwalk. */ 242 if ((code == EQ_EXPR || code == NE_EXPR) 243 && TREE_CODE (op0) == SSA_NAME 244 && ssa_name_has_boolean_range (op0) 245 && is_gimple_min_invariant (op1) 246 && (integer_zerop (op1) || integer_onep (op1))) 247 { 248 tree true_val = constant_boolean_node (true, TREE_TYPE (op0)); 249 tree false_val = constant_boolean_node (false, TREE_TYPE (op0)); 250 251 if (code == EQ_EXPR) 252 { 253 edge_info = allocate_edge_info (true_edge); 254 edge_info->lhs = op0; 255 edge_info->rhs = (integer_zerop (op1) ? false_val : true_val); 256 257 edge_info = allocate_edge_info (false_edge); 258 edge_info->lhs = op0; 259 edge_info->rhs = (integer_zerop (op1) ? true_val : false_val); 260 } 261 else 262 { 263 edge_info = allocate_edge_info (true_edge); 264 edge_info->lhs = op0; 265 edge_info->rhs = (integer_zerop (op1) ? true_val : false_val); 266 267 edge_info = allocate_edge_info (false_edge); 268 edge_info->lhs = op0; 269 edge_info->rhs = (integer_zerop (op1) ? false_val : true_val); 270 } 271 } 272 else if (is_gimple_min_invariant (op0) 273 && (TREE_CODE (op1) == SSA_NAME 274 || is_gimple_min_invariant (op1))) 275 { 276 tree cond = build2 (code, boolean_type_node, op0, op1); 277 tree inverted = invert_truthvalue_loc (loc, cond); 278 bool can_infer_simple_equiv 279 = !(HONOR_SIGNED_ZEROS (op0) 280 && real_zerop (op0)); 281 struct edge_info *edge_info; 282 283 edge_info = allocate_edge_info (true_edge); 284 record_conditions (&edge_info->cond_equivalences, cond, inverted); 285 286 if (can_infer_simple_equiv && code == EQ_EXPR) 287 { 288 edge_info->lhs = op1; 289 edge_info->rhs = op0; 290 } 291 292 edge_info = allocate_edge_info (false_edge); 293 record_conditions (&edge_info->cond_equivalences, inverted, cond); 294 295 if (can_infer_simple_equiv && TREE_CODE (inverted) == EQ_EXPR) 296 { 297 edge_info->lhs = op1; 298 edge_info->rhs = op0; 299 } 300 } 301 302 else if (TREE_CODE (op0) == SSA_NAME 303 && (TREE_CODE (op1) == SSA_NAME 304 || is_gimple_min_invariant (op1))) 305 { 306 tree cond = build2 (code, boolean_type_node, op0, op1); 307 tree inverted = invert_truthvalue_loc (loc, cond); 308 bool can_infer_simple_equiv 309 = !(HONOR_SIGNED_ZEROS (op1) 310 && (TREE_CODE (op1) == SSA_NAME || real_zerop (op1))); 311 struct edge_info *edge_info; 312 313 edge_info = allocate_edge_info (true_edge); 314 record_conditions (&edge_info->cond_equivalences, cond, inverted); 315 316 if (can_infer_simple_equiv && code == EQ_EXPR) 317 { 318 edge_info->lhs = op0; 319 edge_info->rhs = op1; 320 } 321 322 edge_info = allocate_edge_info (false_edge); 323 record_conditions (&edge_info->cond_equivalences, inverted, cond); 324 325 if (can_infer_simple_equiv && TREE_CODE (inverted) == EQ_EXPR) 326 { 327 edge_info->lhs = op0; 328 edge_info->rhs = op1; 329 } 330 } 331 } 332 333 /* ??? TRUTH_NOT_EXPR can create an equivalence too. */ 334 } 335 } 336 337 338 class dom_opt_dom_walker : public dom_walker 339 { 340 public: 341 dom_opt_dom_walker (cdi_direction direction, 342 class const_and_copies *const_and_copies, 343 class avail_exprs_stack *avail_exprs_stack) 344 : dom_walker (direction, true), 345 m_const_and_copies (const_and_copies), 346 m_avail_exprs_stack (avail_exprs_stack), 347 m_dummy_cond (NULL) {} 348 349 virtual edge before_dom_children (basic_block); 350 virtual void after_dom_children (basic_block); 351 352 private: 353 354 /* Unwindable equivalences, both const/copy and expression varieties. */ 355 class const_and_copies *m_const_and_copies; 356 class avail_exprs_stack *m_avail_exprs_stack; 357 358 gcond *m_dummy_cond; 359 }; 360 361 /* Jump threading, redundancy elimination and const/copy propagation. 362 363 This pass may expose new symbols that need to be renamed into SSA. For 364 every new symbol exposed, its corresponding bit will be set in 365 VARS_TO_RENAME. */ 366 367 namespace { 368 369 const pass_data pass_data_dominator = 370 { 371 GIMPLE_PASS, /* type */ 372 "dom", /* name */ 373 OPTGROUP_NONE, /* optinfo_flags */ 374 TV_TREE_SSA_DOMINATOR_OPTS, /* tv_id */ 375 ( PROP_cfg | PROP_ssa ), /* properties_required */ 376 0, /* properties_provided */ 377 0, /* properties_destroyed */ 378 0, /* todo_flags_start */ 379 ( TODO_cleanup_cfg | TODO_update_ssa ), /* todo_flags_finish */ 380 }; 381 382 class pass_dominator : public gimple_opt_pass 383 { 384 public: 385 pass_dominator (gcc::context *ctxt) 386 : gimple_opt_pass (pass_data_dominator, ctxt), 387 may_peel_loop_headers_p (false) 388 {} 389 390 /* opt_pass methods: */ 391 opt_pass * clone () { return new pass_dominator (m_ctxt); } 392 void set_pass_param (unsigned int n, bool param) 393 { 394 gcc_assert (n == 0); 395 may_peel_loop_headers_p = param; 396 } 397 virtual bool gate (function *) { return flag_tree_dom != 0; } 398 virtual unsigned int execute (function *); 399 400 private: 401 /* This flag is used to prevent loops from being peeled repeatedly in jump 402 threading; it will be removed once we preserve loop structures throughout 403 the compilation -- we will be able to mark the affected loops directly in 404 jump threading, and avoid peeling them next time. */ 405 bool may_peel_loop_headers_p; 406 }; // class pass_dominator 407 408 unsigned int 409 pass_dominator::execute (function *fun) 410 { 411 memset (&opt_stats, 0, sizeof (opt_stats)); 412 413 /* Create our hash tables. */ 414 hash_table<expr_elt_hasher> *avail_exprs 415 = new hash_table<expr_elt_hasher> (1024); 416 class avail_exprs_stack *avail_exprs_stack 417 = new class avail_exprs_stack (avail_exprs); 418 class const_and_copies *const_and_copies = new class const_and_copies (); 419 need_eh_cleanup = BITMAP_ALLOC (NULL); 420 need_noreturn_fixup.create (0); 421 422 calculate_dominance_info (CDI_DOMINATORS); 423 cfg_altered = false; 424 425 /* We need to know loop structures in order to avoid destroying them 426 in jump threading. Note that we still can e.g. thread through loop 427 headers to an exit edge, or through loop header to the loop body, assuming 428 that we update the loop info. 429 430 TODO: We don't need to set LOOPS_HAVE_PREHEADERS generally, but due 431 to several overly conservative bail-outs in jump threading, case 432 gcc.dg/tree-ssa/pr21417.c can't be threaded if loop preheader is 433 missing. We should improve jump threading in future then 434 LOOPS_HAVE_PREHEADERS won't be needed here. */ 435 loop_optimizer_init (LOOPS_HAVE_PREHEADERS | LOOPS_HAVE_SIMPLE_LATCHES); 436 437 /* Initialize the value-handle array. */ 438 threadedge_initialize_values (); 439 440 /* We need accurate information regarding back edges in the CFG 441 for jump threading; this may include back edges that are not part of 442 a single loop. */ 443 mark_dfs_back_edges (); 444 445 /* We want to create the edge info structures before the dominator walk 446 so that they'll be in place for the jump threader, particularly when 447 threading through a join block. 448 449 The conditions will be lazily updated with global equivalences as 450 we reach them during the dominator walk. */ 451 basic_block bb; 452 FOR_EACH_BB_FN (bb, fun) 453 record_edge_info (bb); 454 455 /* Recursively walk the dominator tree optimizing statements. */ 456 dom_opt_dom_walker walker (CDI_DOMINATORS, 457 const_and_copies, 458 avail_exprs_stack); 459 walker.walk (fun->cfg->x_entry_block_ptr); 460 461 /* Look for blocks where we cleared EDGE_EXECUTABLE on an outgoing 462 edge. When found, remove jump threads which contain any outgoing 463 edge from the affected block. */ 464 if (cfg_altered) 465 { 466 FOR_EACH_BB_FN (bb, fun) 467 { 468 edge_iterator ei; 469 edge e; 470 471 /* First see if there are any edges without EDGE_EXECUTABLE 472 set. */ 473 bool found = false; 474 FOR_EACH_EDGE (e, ei, bb->succs) 475 { 476 if ((e->flags & EDGE_EXECUTABLE) == 0) 477 { 478 found = true; 479 break; 480 } 481 } 482 483 /* If there were any such edges found, then remove jump threads 484 containing any edge leaving BB. */ 485 if (found) 486 FOR_EACH_EDGE (e, ei, bb->succs) 487 remove_jump_threads_including (e); 488 } 489 } 490 491 { 492 gimple_stmt_iterator gsi; 493 basic_block bb; 494 FOR_EACH_BB_FN (bb, fun) 495 { 496 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) 497 update_stmt_if_modified (gsi_stmt (gsi)); 498 } 499 } 500 501 /* If we exposed any new variables, go ahead and put them into 502 SSA form now, before we handle jump threading. This simplifies 503 interactions between rewriting of _DECL nodes into SSA form 504 and rewriting SSA_NAME nodes into SSA form after block 505 duplication and CFG manipulation. */ 506 update_ssa (TODO_update_ssa); 507 508 free_all_edge_infos (); 509 510 /* Thread jumps, creating duplicate blocks as needed. */ 511 cfg_altered |= thread_through_all_blocks (may_peel_loop_headers_p); 512 513 if (cfg_altered) 514 free_dominance_info (CDI_DOMINATORS); 515 516 /* Removal of statements may make some EH edges dead. Purge 517 such edges from the CFG as needed. */ 518 if (!bitmap_empty_p (need_eh_cleanup)) 519 { 520 unsigned i; 521 bitmap_iterator bi; 522 523 /* Jump threading may have created forwarder blocks from blocks 524 needing EH cleanup; the new successor of these blocks, which 525 has inherited from the original block, needs the cleanup. 526 Don't clear bits in the bitmap, as that can break the bitmap 527 iterator. */ 528 EXECUTE_IF_SET_IN_BITMAP (need_eh_cleanup, 0, i, bi) 529 { 530 basic_block bb = BASIC_BLOCK_FOR_FN (fun, i); 531 if (bb == NULL) 532 continue; 533 while (single_succ_p (bb) 534 && (single_succ_edge (bb)->flags & EDGE_EH) == 0) 535 bb = single_succ (bb); 536 if (bb == EXIT_BLOCK_PTR_FOR_FN (fun)) 537 continue; 538 if ((unsigned) bb->index != i) 539 bitmap_set_bit (need_eh_cleanup, bb->index); 540 } 541 542 gimple_purge_all_dead_eh_edges (need_eh_cleanup); 543 bitmap_clear (need_eh_cleanup); 544 } 545 546 /* Fixup stmts that became noreturn calls. This may require splitting 547 blocks and thus isn't possible during the dominator walk or before 548 jump threading finished. Do this in reverse order so we don't 549 inadvertedly remove a stmt we want to fixup by visiting a dominating 550 now noreturn call first. */ 551 while (!need_noreturn_fixup.is_empty ()) 552 { 553 gimple *stmt = need_noreturn_fixup.pop (); 554 if (dump_file && dump_flags & TDF_DETAILS) 555 { 556 fprintf (dump_file, "Fixing up noreturn call "); 557 print_gimple_stmt (dump_file, stmt, 0, 0); 558 fprintf (dump_file, "\n"); 559 } 560 fixup_noreturn_call (stmt); 561 } 562 563 statistics_counter_event (fun, "Redundant expressions eliminated", 564 opt_stats.num_re); 565 statistics_counter_event (fun, "Constants propagated", 566 opt_stats.num_const_prop); 567 statistics_counter_event (fun, "Copies propagated", 568 opt_stats.num_copy_prop); 569 570 /* Debugging dumps. */ 571 if (dump_file && (dump_flags & TDF_STATS)) 572 dump_dominator_optimization_stats (dump_file, avail_exprs); 573 574 loop_optimizer_finalize (); 575 576 /* Delete our main hashtable. */ 577 delete avail_exprs; 578 avail_exprs = NULL; 579 580 /* Free asserted bitmaps and stacks. */ 581 BITMAP_FREE (need_eh_cleanup); 582 need_noreturn_fixup.release (); 583 delete avail_exprs_stack; 584 delete const_and_copies; 585 586 /* Free the value-handle array. */ 587 threadedge_finalize_values (); 588 589 return 0; 590 } 591 592 } // anon namespace 593 594 gimple_opt_pass * 595 make_pass_dominator (gcc::context *ctxt) 596 { 597 return new pass_dominator (ctxt); 598 } 599 600 601 /* A trivial wrapper so that we can present the generic jump 602 threading code with a simple API for simplifying statements. */ 603 static tree 604 simplify_stmt_for_jump_threading (gimple *stmt, 605 gimple *within_stmt ATTRIBUTE_UNUSED, 606 class avail_exprs_stack *avail_exprs_stack, 607 basic_block bb ATTRIBUTE_UNUSED) 608 { 609 return avail_exprs_stack->lookup_avail_expr (stmt, false, true); 610 } 611 612 /* Valueize hook for gimple_fold_stmt_to_constant_1. */ 613 614 static tree 615 dom_valueize (tree t) 616 { 617 if (TREE_CODE (t) == SSA_NAME) 618 { 619 tree tem = SSA_NAME_VALUE (t); 620 if (tem) 621 return tem; 622 } 623 return t; 624 } 625 626 /* We have just found an equivalence for LHS on an edge E. 627 Look backwards to other uses of LHS and see if we can derive 628 additional equivalences that are valid on edge E. */ 629 static void 630 back_propagate_equivalences (tree lhs, edge e, 631 class const_and_copies *const_and_copies) 632 { 633 use_operand_p use_p; 634 imm_use_iterator iter; 635 bitmap domby = NULL; 636 basic_block dest = e->dest; 637 638 /* Iterate over the uses of LHS to see if any dominate E->dest. 639 If so, they may create useful equivalences too. 640 641 ??? If the code gets re-organized to a worklist to catch more 642 indirect opportunities and it is made to handle PHIs then this 643 should only consider use_stmts in basic-blocks we have already visited. */ 644 FOR_EACH_IMM_USE_FAST (use_p, iter, lhs) 645 { 646 gimple *use_stmt = USE_STMT (use_p); 647 648 /* Often the use is in DEST, which we trivially know we can't use. 649 This is cheaper than the dominator set tests below. */ 650 if (dest == gimple_bb (use_stmt)) 651 continue; 652 653 /* Filter out statements that can never produce a useful 654 equivalence. */ 655 tree lhs2 = gimple_get_lhs (use_stmt); 656 if (!lhs2 || TREE_CODE (lhs2) != SSA_NAME) 657 continue; 658 659 /* Profiling has shown the domination tests here can be fairly 660 expensive. We get significant improvements by building the 661 set of blocks that dominate BB. We can then just test 662 for set membership below. 663 664 We also initialize the set lazily since often the only uses 665 are going to be in the same block as DEST. */ 666 if (!domby) 667 { 668 domby = BITMAP_ALLOC (NULL); 669 basic_block bb = get_immediate_dominator (CDI_DOMINATORS, dest); 670 while (bb) 671 { 672 bitmap_set_bit (domby, bb->index); 673 bb = get_immediate_dominator (CDI_DOMINATORS, bb); 674 } 675 } 676 677 /* This tests if USE_STMT does not dominate DEST. */ 678 if (!bitmap_bit_p (domby, gimple_bb (use_stmt)->index)) 679 continue; 680 681 /* At this point USE_STMT dominates DEST and may result in a 682 useful equivalence. Try to simplify its RHS to a constant 683 or SSA_NAME. */ 684 tree res = gimple_fold_stmt_to_constant_1 (use_stmt, dom_valueize, 685 no_follow_ssa_edges); 686 if (res && (TREE_CODE (res) == SSA_NAME || is_gimple_min_invariant (res))) 687 record_equality (lhs2, res, const_and_copies); 688 } 689 690 if (domby) 691 BITMAP_FREE (domby); 692 } 693 694 /* Record NAME has the value zero and if NAME was set from a BIT_IOR_EXPR 695 recurse into both operands recording their values as zero too. 696 RECURSION_DEPTH controls how far back we recurse through the operands 697 of the BIT_IOR_EXPR. */ 698 699 static void 700 derive_equivalences_from_bit_ior (tree name, 701 const_and_copies *const_and_copies, 702 int recursion_limit) 703 { 704 if (recursion_limit == 0) 705 return; 706 707 if (TREE_CODE (name) == SSA_NAME) 708 { 709 tree value = build_zero_cst (TREE_TYPE (name)); 710 711 /* This records the equivalence for the toplevel object. */ 712 record_equality (name, value, const_and_copies); 713 714 /* And we can recurse into each operand to potentially find more 715 equivalences. */ 716 gimple *def_stmt = SSA_NAME_DEF_STMT (name); 717 if (is_gimple_assign (def_stmt) 718 && gimple_assign_rhs_code (def_stmt) == BIT_IOR_EXPR) 719 { 720 derive_equivalences_from_bit_ior (gimple_assign_rhs1 (def_stmt), 721 const_and_copies, 722 recursion_limit - 1); 723 derive_equivalences_from_bit_ior (gimple_assign_rhs2 (def_stmt), 724 const_and_copies, 725 recursion_limit - 1); 726 } 727 } 728 } 729 730 /* Record into CONST_AND_COPIES and AVAIL_EXPRS_STACK any equivalences implied 731 by traversing edge E (which are cached in E->aux). 732 733 Callers are responsible for managing the unwinding markers. */ 734 void 735 record_temporary_equivalences (edge e, 736 class const_and_copies *const_and_copies, 737 class avail_exprs_stack *avail_exprs_stack) 738 { 739 int i; 740 struct edge_info *edge_info = (struct edge_info *) e->aux; 741 742 /* If we have info associated with this edge, record it into 743 our equivalence tables. */ 744 if (edge_info) 745 { 746 cond_equivalence *eq; 747 /* If we have 0 = COND or 1 = COND equivalences, record them 748 into our expression hash tables. */ 749 for (i = 0; edge_info->cond_equivalences.iterate (i, &eq); ++i) 750 { 751 avail_exprs_stack->record_cond (eq); 752 753 /* If the condition is testing that X == 0 is true or X != 0 is false 754 and X is set from a BIT_IOR_EXPR, then we can record equivalences 755 for the operands of the BIT_IOR_EXPR (and recurse on those). */ 756 tree op0 = eq->cond.ops.binary.opnd0; 757 tree op1 = eq->cond.ops.binary.opnd1; 758 if (TREE_CODE (op0) == SSA_NAME && integer_zerop (op1)) 759 { 760 enum tree_code code = eq->cond.ops.binary.op; 761 if ((code == EQ_EXPR && eq->value == boolean_true_node) 762 || (code == NE_EXPR && eq->value == boolean_false_node)) 763 derive_equivalences_from_bit_ior (op0, const_and_copies, 4); 764 765 /* TODO: We could handle BIT_AND_EXPR in a similar fashion 766 recording that the operands have a nonzero value. */ 767 768 /* TODO: We can handle more cases here, particularly when OP0 is 769 known to have a boolean range. */ 770 } 771 } 772 773 tree lhs = edge_info->lhs; 774 if (!lhs || TREE_CODE (lhs) != SSA_NAME) 775 return; 776 777 /* Record the simple NAME = VALUE equivalence. */ 778 tree rhs = edge_info->rhs; 779 record_equality (lhs, rhs, const_and_copies); 780 781 /* We already recorded that LHS = RHS, with canonicalization, 782 value chain following, etc. 783 784 We also want to record RHS = LHS, but without any canonicalization 785 or value chain following. */ 786 if (TREE_CODE (rhs) == SSA_NAME) 787 const_and_copies->record_const_or_copy_raw (rhs, lhs, 788 SSA_NAME_VALUE (rhs)); 789 790 /* If LHS is an SSA_NAME and RHS is a constant integer and LHS was 791 set via a widening type conversion, then we may be able to record 792 additional equivalences. */ 793 if (TREE_CODE (rhs) == INTEGER_CST) 794 { 795 gimple *defstmt = SSA_NAME_DEF_STMT (lhs); 796 797 if (defstmt 798 && is_gimple_assign (defstmt) 799 && CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (defstmt))) 800 { 801 tree old_rhs = gimple_assign_rhs1 (defstmt); 802 803 /* If the conversion widens the original value and 804 the constant is in the range of the type of OLD_RHS, 805 then convert the constant and record the equivalence. 806 807 Note that int_fits_type_p does not check the precision 808 if the upper and lower bounds are OK. */ 809 if (INTEGRAL_TYPE_P (TREE_TYPE (old_rhs)) 810 && (TYPE_PRECISION (TREE_TYPE (lhs)) 811 > TYPE_PRECISION (TREE_TYPE (old_rhs))) 812 && int_fits_type_p (rhs, TREE_TYPE (old_rhs))) 813 { 814 tree newval = fold_convert (TREE_TYPE (old_rhs), rhs); 815 record_equality (old_rhs, newval, const_and_copies); 816 } 817 } 818 } 819 820 /* Any equivalence found for LHS may result in additional 821 equivalences for other uses of LHS that we have already 822 processed. */ 823 back_propagate_equivalences (lhs, e, const_and_copies); 824 } 825 } 826 827 /* PHI nodes can create equivalences too. 828 829 Ignoring any alternatives which are the same as the result, if 830 all the alternatives are equal, then the PHI node creates an 831 equivalence. */ 832 833 static void 834 record_equivalences_from_phis (basic_block bb) 835 { 836 gphi_iterator gsi; 837 838 for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) 839 { 840 gphi *phi = gsi.phi (); 841 842 tree lhs = gimple_phi_result (phi); 843 tree rhs = NULL; 844 size_t i; 845 846 for (i = 0; i < gimple_phi_num_args (phi); i++) 847 { 848 tree t = gimple_phi_arg_def (phi, i); 849 850 /* Ignore alternatives which are the same as our LHS. Since 851 LHS is a PHI_RESULT, it is known to be a SSA_NAME, so we 852 can simply compare pointers. */ 853 if (lhs == t) 854 continue; 855 856 /* If the associated edge is not marked as executable, then it 857 can be ignored. */ 858 if ((gimple_phi_arg_edge (phi, i)->flags & EDGE_EXECUTABLE) == 0) 859 continue; 860 861 t = dom_valueize (t); 862 863 /* If we have not processed an alternative yet, then set 864 RHS to this alternative. */ 865 if (rhs == NULL) 866 rhs = t; 867 /* If we have processed an alternative (stored in RHS), then 868 see if it is equal to this one. If it isn't, then stop 869 the search. */ 870 else if (! operand_equal_for_phi_arg_p (rhs, t)) 871 break; 872 } 873 874 /* If we had no interesting alternatives, then all the RHS alternatives 875 must have been the same as LHS. */ 876 if (!rhs) 877 rhs = lhs; 878 879 /* If we managed to iterate through each PHI alternative without 880 breaking out of the loop, then we have a PHI which may create 881 a useful equivalence. We do not need to record unwind data for 882 this, since this is a true assignment and not an equivalence 883 inferred from a comparison. All uses of this ssa name are dominated 884 by this assignment, so unwinding just costs time and space. */ 885 if (i == gimple_phi_num_args (phi) 886 && may_propagate_copy (lhs, rhs)) 887 set_ssa_name_value (lhs, rhs); 888 } 889 } 890 891 /* Ignoring loop backedges, if BB has precisely one incoming edge then 892 return that edge. Otherwise return NULL. */ 893 static edge 894 single_incoming_edge_ignoring_loop_edges (basic_block bb) 895 { 896 edge retval = NULL; 897 edge e; 898 edge_iterator ei; 899 900 FOR_EACH_EDGE (e, ei, bb->preds) 901 { 902 /* A loop back edge can be identified by the destination of 903 the edge dominating the source of the edge. */ 904 if (dominated_by_p (CDI_DOMINATORS, e->src, e->dest)) 905 continue; 906 907 /* We can safely ignore edges that are not executable. */ 908 if ((e->flags & EDGE_EXECUTABLE) == 0) 909 continue; 910 911 /* If we have already seen a non-loop edge, then we must have 912 multiple incoming non-loop edges and thus we return NULL. */ 913 if (retval) 914 return NULL; 915 916 /* This is the first non-loop incoming edge we have found. Record 917 it. */ 918 retval = e; 919 } 920 921 return retval; 922 } 923 924 /* Record any equivalences created by the incoming edge to BB into 925 CONST_AND_COPIES and AVAIL_EXPRS_STACK. If BB has more than one 926 incoming edge, then no equivalence is created. */ 927 928 static void 929 record_equivalences_from_incoming_edge (basic_block bb, 930 class const_and_copies *const_and_copies, 931 class avail_exprs_stack *avail_exprs_stack) 932 { 933 edge e; 934 basic_block parent; 935 936 /* If our parent block ended with a control statement, then we may be 937 able to record some equivalences based on which outgoing edge from 938 the parent was followed. */ 939 parent = get_immediate_dominator (CDI_DOMINATORS, bb); 940 941 e = single_incoming_edge_ignoring_loop_edges (bb); 942 943 /* If we had a single incoming edge from our parent block, then enter 944 any data associated with the edge into our tables. */ 945 if (e && e->src == parent) 946 record_temporary_equivalences (e, const_and_copies, avail_exprs_stack); 947 } 948 949 /* Dump statistics for the hash table HTAB. */ 950 951 static void 952 htab_statistics (FILE *file, const hash_table<expr_elt_hasher> &htab) 953 { 954 fprintf (file, "size %ld, %ld elements, %f collision/search ratio\n", 955 (long) htab.size (), 956 (long) htab.elements (), 957 htab.collisions ()); 958 } 959 960 /* Dump SSA statistics on FILE. */ 961 962 static void 963 dump_dominator_optimization_stats (FILE *file, 964 hash_table<expr_elt_hasher> *avail_exprs) 965 { 966 fprintf (file, "Total number of statements: %6ld\n\n", 967 opt_stats.num_stmts); 968 fprintf (file, "Exprs considered for dominator optimizations: %6ld\n", 969 opt_stats.num_exprs_considered); 970 971 fprintf (file, "\nHash table statistics:\n"); 972 973 fprintf (file, " avail_exprs: "); 974 htab_statistics (file, *avail_exprs); 975 } 976 977 978 /* Similarly, but assume that X and Y are the two operands of an EQ_EXPR. 979 This constrains the cases in which we may treat this as assignment. */ 980 981 static void 982 record_equality (tree x, tree y, class const_and_copies *const_and_copies) 983 { 984 tree prev_x = NULL, prev_y = NULL; 985 986 if (tree_swap_operands_p (x, y)) 987 std::swap (x, y); 988 989 /* Most of the time tree_swap_operands_p does what we want. But there 990 are cases where we know one operand is better for copy propagation than 991 the other. Given no other code cares about ordering of equality 992 comparison operators for that purpose, we just handle the special cases 993 here. */ 994 if (TREE_CODE (x) == SSA_NAME && TREE_CODE (y) == SSA_NAME) 995 { 996 /* If one operand is a single use operand, then make it 997 X. This will preserve its single use properly and if this 998 conditional is eliminated, the computation of X can be 999 eliminated as well. */ 1000 if (has_single_use (y) && ! has_single_use (x)) 1001 std::swap (x, y); 1002 } 1003 if (TREE_CODE (x) == SSA_NAME) 1004 prev_x = SSA_NAME_VALUE (x); 1005 if (TREE_CODE (y) == SSA_NAME) 1006 prev_y = SSA_NAME_VALUE (y); 1007 1008 /* If one of the previous values is invariant, or invariant in more loops 1009 (by depth), then use that. 1010 Otherwise it doesn't matter which value we choose, just so 1011 long as we canonicalize on one value. */ 1012 if (is_gimple_min_invariant (y)) 1013 ; 1014 else if (is_gimple_min_invariant (x)) 1015 prev_x = x, x = y, y = prev_x, prev_x = prev_y; 1016 else if (prev_x && is_gimple_min_invariant (prev_x)) 1017 x = y, y = prev_x, prev_x = prev_y; 1018 else if (prev_y) 1019 y = prev_y; 1020 1021 /* After the swapping, we must have one SSA_NAME. */ 1022 if (TREE_CODE (x) != SSA_NAME) 1023 return; 1024 1025 /* For IEEE, -0.0 == 0.0, so we don't necessarily know the sign of a 1026 variable compared against zero. If we're honoring signed zeros, 1027 then we cannot record this value unless we know that the value is 1028 nonzero. */ 1029 if (HONOR_SIGNED_ZEROS (x) 1030 && (TREE_CODE (y) != REAL_CST 1031 || real_equal (&dconst0, &TREE_REAL_CST (y)))) 1032 return; 1033 1034 const_and_copies->record_const_or_copy (x, y, prev_x); 1035 } 1036 1037 /* Returns true when STMT is a simple iv increment. It detects the 1038 following situation: 1039 1040 i_1 = phi (..., i_2) 1041 i_2 = i_1 +/- ... */ 1042 1043 bool 1044 simple_iv_increment_p (gimple *stmt) 1045 { 1046 enum tree_code code; 1047 tree lhs, preinc; 1048 gimple *phi; 1049 size_t i; 1050 1051 if (gimple_code (stmt) != GIMPLE_ASSIGN) 1052 return false; 1053 1054 lhs = gimple_assign_lhs (stmt); 1055 if (TREE_CODE (lhs) != SSA_NAME) 1056 return false; 1057 1058 code = gimple_assign_rhs_code (stmt); 1059 if (code != PLUS_EXPR 1060 && code != MINUS_EXPR 1061 && code != POINTER_PLUS_EXPR) 1062 return false; 1063 1064 preinc = gimple_assign_rhs1 (stmt); 1065 if (TREE_CODE (preinc) != SSA_NAME) 1066 return false; 1067 1068 phi = SSA_NAME_DEF_STMT (preinc); 1069 if (gimple_code (phi) != GIMPLE_PHI) 1070 return false; 1071 1072 for (i = 0; i < gimple_phi_num_args (phi); i++) 1073 if (gimple_phi_arg_def (phi, i) == lhs) 1074 return true; 1075 1076 return false; 1077 } 1078 1079 /* Propagate know values from SSA_NAME_VALUE into the PHI nodes of the 1080 successors of BB. */ 1081 1082 static void 1083 cprop_into_successor_phis (basic_block bb, 1084 class const_and_copies *const_and_copies) 1085 { 1086 edge e; 1087 edge_iterator ei; 1088 1089 FOR_EACH_EDGE (e, ei, bb->succs) 1090 { 1091 int indx; 1092 gphi_iterator gsi; 1093 1094 /* If this is an abnormal edge, then we do not want to copy propagate 1095 into the PHI alternative associated with this edge. */ 1096 if (e->flags & EDGE_ABNORMAL) 1097 continue; 1098 1099 gsi = gsi_start_phis (e->dest); 1100 if (gsi_end_p (gsi)) 1101 continue; 1102 1103 /* We may have an equivalence associated with this edge. While 1104 we can not propagate it into non-dominated blocks, we can 1105 propagate them into PHIs in non-dominated blocks. */ 1106 1107 /* Push the unwind marker so we can reset the const and copies 1108 table back to its original state after processing this edge. */ 1109 const_and_copies->push_marker (); 1110 1111 /* Extract and record any simple NAME = VALUE equivalences. 1112 1113 Don't bother with [01] = COND equivalences, they're not useful 1114 here. */ 1115 struct edge_info *edge_info = (struct edge_info *) e->aux; 1116 if (edge_info) 1117 { 1118 tree lhs = edge_info->lhs; 1119 tree rhs = edge_info->rhs; 1120 1121 if (lhs && TREE_CODE (lhs) == SSA_NAME) 1122 const_and_copies->record_const_or_copy (lhs, rhs); 1123 } 1124 1125 indx = e->dest_idx; 1126 for ( ; !gsi_end_p (gsi); gsi_next (&gsi)) 1127 { 1128 tree new_val; 1129 use_operand_p orig_p; 1130 tree orig_val; 1131 gphi *phi = gsi.phi (); 1132 1133 /* The alternative may be associated with a constant, so verify 1134 it is an SSA_NAME before doing anything with it. */ 1135 orig_p = gimple_phi_arg_imm_use_ptr (phi, indx); 1136 orig_val = get_use_from_ptr (orig_p); 1137 if (TREE_CODE (orig_val) != SSA_NAME) 1138 continue; 1139 1140 /* If we have *ORIG_P in our constant/copy table, then replace 1141 ORIG_P with its value in our constant/copy table. */ 1142 new_val = SSA_NAME_VALUE (orig_val); 1143 if (new_val 1144 && new_val != orig_val 1145 && may_propagate_copy (orig_val, new_val)) 1146 propagate_value (orig_p, new_val); 1147 } 1148 1149 const_and_copies->pop_to_marker (); 1150 } 1151 } 1152 1153 edge 1154 dom_opt_dom_walker::before_dom_children (basic_block bb) 1155 { 1156 gimple_stmt_iterator gsi; 1157 1158 if (dump_file && (dump_flags & TDF_DETAILS)) 1159 fprintf (dump_file, "\n\nOptimizing block #%d\n\n", bb->index); 1160 1161 /* Push a marker on the stacks of local information so that we know how 1162 far to unwind when we finalize this block. */ 1163 m_avail_exprs_stack->push_marker (); 1164 m_const_and_copies->push_marker (); 1165 1166 record_equivalences_from_incoming_edge (bb, m_const_and_copies, 1167 m_avail_exprs_stack); 1168 1169 /* PHI nodes can create equivalences too. */ 1170 record_equivalences_from_phis (bb); 1171 1172 /* Create equivalences from redundant PHIs. PHIs are only truly 1173 redundant when they exist in the same block, so push another 1174 marker and unwind right afterwards. */ 1175 m_avail_exprs_stack->push_marker (); 1176 for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) 1177 eliminate_redundant_computations (&gsi, m_const_and_copies, 1178 m_avail_exprs_stack); 1179 m_avail_exprs_stack->pop_to_marker (); 1180 1181 edge taken_edge = NULL; 1182 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) 1183 taken_edge 1184 = optimize_stmt (bb, gsi, m_const_and_copies, m_avail_exprs_stack); 1185 1186 /* Now prepare to process dominated blocks. */ 1187 record_edge_info (bb); 1188 cprop_into_successor_phis (bb, m_const_and_copies); 1189 if (taken_edge && !dbg_cnt (dom_unreachable_edges)) 1190 return NULL; 1191 1192 return taken_edge; 1193 } 1194 1195 /* We have finished processing the dominator children of BB, perform 1196 any finalization actions in preparation for leaving this node in 1197 the dominator tree. */ 1198 1199 void 1200 dom_opt_dom_walker::after_dom_children (basic_block bb) 1201 { 1202 if (! m_dummy_cond) 1203 m_dummy_cond = gimple_build_cond (NE_EXPR, integer_zero_node, 1204 integer_zero_node, NULL, NULL); 1205 1206 thread_outgoing_edges (bb, m_dummy_cond, m_const_and_copies, 1207 m_avail_exprs_stack, 1208 simplify_stmt_for_jump_threading); 1209 1210 /* These remove expressions local to BB from the tables. */ 1211 m_avail_exprs_stack->pop_to_marker (); 1212 m_const_and_copies->pop_to_marker (); 1213 } 1214 1215 /* Search for redundant computations in STMT. If any are found, then 1216 replace them with the variable holding the result of the computation. 1217 1218 If safe, record this expression into AVAIL_EXPRS_STACK and 1219 CONST_AND_COPIES. */ 1220 1221 static void 1222 eliminate_redundant_computations (gimple_stmt_iterator* gsi, 1223 class const_and_copies *const_and_copies, 1224 class avail_exprs_stack *avail_exprs_stack) 1225 { 1226 tree expr_type; 1227 tree cached_lhs; 1228 tree def; 1229 bool insert = true; 1230 bool assigns_var_p = false; 1231 1232 gimple *stmt = gsi_stmt (*gsi); 1233 1234 if (gimple_code (stmt) == GIMPLE_PHI) 1235 def = gimple_phi_result (stmt); 1236 else 1237 def = gimple_get_lhs (stmt); 1238 1239 /* Certain expressions on the RHS can be optimized away, but can not 1240 themselves be entered into the hash tables. */ 1241 if (! def 1242 || TREE_CODE (def) != SSA_NAME 1243 || SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def) 1244 || gimple_vdef (stmt) 1245 /* Do not record equivalences for increments of ivs. This would create 1246 overlapping live ranges for a very questionable gain. */ 1247 || simple_iv_increment_p (stmt)) 1248 insert = false; 1249 1250 /* Check if the expression has been computed before. */ 1251 cached_lhs = avail_exprs_stack->lookup_avail_expr (stmt, insert, true); 1252 1253 opt_stats.num_exprs_considered++; 1254 1255 /* Get the type of the expression we are trying to optimize. */ 1256 if (is_gimple_assign (stmt)) 1257 { 1258 expr_type = TREE_TYPE (gimple_assign_lhs (stmt)); 1259 assigns_var_p = true; 1260 } 1261 else if (gimple_code (stmt) == GIMPLE_COND) 1262 expr_type = boolean_type_node; 1263 else if (is_gimple_call (stmt)) 1264 { 1265 gcc_assert (gimple_call_lhs (stmt)); 1266 expr_type = TREE_TYPE (gimple_call_lhs (stmt)); 1267 assigns_var_p = true; 1268 } 1269 else if (gswitch *swtch_stmt = dyn_cast <gswitch *> (stmt)) 1270 expr_type = TREE_TYPE (gimple_switch_index (swtch_stmt)); 1271 else if (gimple_code (stmt) == GIMPLE_PHI) 1272 /* We can't propagate into a phi, so the logic below doesn't apply. 1273 Instead record an equivalence between the cached LHS and the 1274 PHI result of this statement, provided they are in the same block. 1275 This should be sufficient to kill the redundant phi. */ 1276 { 1277 if (def && cached_lhs) 1278 const_and_copies->record_const_or_copy (def, cached_lhs); 1279 return; 1280 } 1281 else 1282 gcc_unreachable (); 1283 1284 if (!cached_lhs) 1285 return; 1286 1287 /* It is safe to ignore types here since we have already done 1288 type checking in the hashing and equality routines. In fact 1289 type checking here merely gets in the way of constant 1290 propagation. Also, make sure that it is safe to propagate 1291 CACHED_LHS into the expression in STMT. */ 1292 if ((TREE_CODE (cached_lhs) != SSA_NAME 1293 && (assigns_var_p 1294 || useless_type_conversion_p (expr_type, TREE_TYPE (cached_lhs)))) 1295 || may_propagate_copy_into_stmt (stmt, cached_lhs)) 1296 { 1297 gcc_checking_assert (TREE_CODE (cached_lhs) == SSA_NAME 1298 || is_gimple_min_invariant (cached_lhs)); 1299 1300 if (dump_file && (dump_flags & TDF_DETAILS)) 1301 { 1302 fprintf (dump_file, " Replaced redundant expr '"); 1303 print_gimple_expr (dump_file, stmt, 0, dump_flags); 1304 fprintf (dump_file, "' with '"); 1305 print_generic_expr (dump_file, cached_lhs, dump_flags); 1306 fprintf (dump_file, "'\n"); 1307 } 1308 1309 opt_stats.num_re++; 1310 1311 if (assigns_var_p 1312 && !useless_type_conversion_p (expr_type, TREE_TYPE (cached_lhs))) 1313 cached_lhs = fold_convert (expr_type, cached_lhs); 1314 1315 propagate_tree_value_into_stmt (gsi, cached_lhs); 1316 1317 /* Since it is always necessary to mark the result as modified, 1318 perhaps we should move this into propagate_tree_value_into_stmt 1319 itself. */ 1320 gimple_set_modified (gsi_stmt (*gsi), true); 1321 } 1322 } 1323 1324 /* STMT, a GIMPLE_ASSIGN, may create certain equivalences, in either 1325 the available expressions table or the const_and_copies table. 1326 Detect and record those equivalences into AVAIL_EXPRS_STACK. 1327 1328 We handle only very simple copy equivalences here. The heavy 1329 lifing is done by eliminate_redundant_computations. */ 1330 1331 static void 1332 record_equivalences_from_stmt (gimple *stmt, int may_optimize_p, 1333 class avail_exprs_stack *avail_exprs_stack) 1334 { 1335 tree lhs; 1336 enum tree_code lhs_code; 1337 1338 gcc_assert (is_gimple_assign (stmt)); 1339 1340 lhs = gimple_assign_lhs (stmt); 1341 lhs_code = TREE_CODE (lhs); 1342 1343 if (lhs_code == SSA_NAME 1344 && gimple_assign_single_p (stmt)) 1345 { 1346 tree rhs = gimple_assign_rhs1 (stmt); 1347 1348 /* If the RHS of the assignment is a constant or another variable that 1349 may be propagated, register it in the CONST_AND_COPIES table. We 1350 do not need to record unwind data for this, since this is a true 1351 assignment and not an equivalence inferred from a comparison. All 1352 uses of this ssa name are dominated by this assignment, so unwinding 1353 just costs time and space. */ 1354 if (may_optimize_p 1355 && (TREE_CODE (rhs) == SSA_NAME 1356 || is_gimple_min_invariant (rhs))) 1357 { 1358 rhs = dom_valueize (rhs); 1359 1360 if (dump_file && (dump_flags & TDF_DETAILS)) 1361 { 1362 fprintf (dump_file, "==== ASGN "); 1363 print_generic_expr (dump_file, lhs, 0); 1364 fprintf (dump_file, " = "); 1365 print_generic_expr (dump_file, rhs, 0); 1366 fprintf (dump_file, "\n"); 1367 } 1368 1369 set_ssa_name_value (lhs, rhs); 1370 } 1371 } 1372 1373 /* Make sure we can propagate &x + CST. */ 1374 if (lhs_code == SSA_NAME 1375 && gimple_assign_rhs_code (stmt) == POINTER_PLUS_EXPR 1376 && TREE_CODE (gimple_assign_rhs1 (stmt)) == ADDR_EXPR 1377 && TREE_CODE (gimple_assign_rhs2 (stmt)) == INTEGER_CST) 1378 { 1379 tree op0 = gimple_assign_rhs1 (stmt); 1380 tree op1 = gimple_assign_rhs2 (stmt); 1381 tree new_rhs 1382 = build_fold_addr_expr (fold_build2 (MEM_REF, 1383 TREE_TYPE (TREE_TYPE (op0)), 1384 unshare_expr (op0), 1385 fold_convert (ptr_type_node, 1386 op1))); 1387 if (dump_file && (dump_flags & TDF_DETAILS)) 1388 { 1389 fprintf (dump_file, "==== ASGN "); 1390 print_generic_expr (dump_file, lhs, 0); 1391 fprintf (dump_file, " = "); 1392 print_generic_expr (dump_file, new_rhs, 0); 1393 fprintf (dump_file, "\n"); 1394 } 1395 1396 set_ssa_name_value (lhs, new_rhs); 1397 } 1398 1399 /* A memory store, even an aliased store, creates a useful 1400 equivalence. By exchanging the LHS and RHS, creating suitable 1401 vops and recording the result in the available expression table, 1402 we may be able to expose more redundant loads. */ 1403 if (!gimple_has_volatile_ops (stmt) 1404 && gimple_references_memory_p (stmt) 1405 && gimple_assign_single_p (stmt) 1406 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME 1407 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))) 1408 && !is_gimple_reg (lhs)) 1409 { 1410 tree rhs = gimple_assign_rhs1 (stmt); 1411 gassign *new_stmt; 1412 1413 /* Build a new statement with the RHS and LHS exchanged. */ 1414 if (TREE_CODE (rhs) == SSA_NAME) 1415 { 1416 /* NOTE tuples. The call to gimple_build_assign below replaced 1417 a call to build_gimple_modify_stmt, which did not set the 1418 SSA_NAME_DEF_STMT on the LHS of the assignment. Doing so 1419 may cause an SSA validation failure, as the LHS may be a 1420 default-initialized name and should have no definition. I'm 1421 a bit dubious of this, as the artificial statement that we 1422 generate here may in fact be ill-formed, but it is simply 1423 used as an internal device in this pass, and never becomes 1424 part of the CFG. */ 1425 gimple *defstmt = SSA_NAME_DEF_STMT (rhs); 1426 new_stmt = gimple_build_assign (rhs, lhs); 1427 SSA_NAME_DEF_STMT (rhs) = defstmt; 1428 } 1429 else 1430 new_stmt = gimple_build_assign (rhs, lhs); 1431 1432 gimple_set_vuse (new_stmt, gimple_vdef (stmt)); 1433 1434 /* Finally enter the statement into the available expression 1435 table. */ 1436 avail_exprs_stack->lookup_avail_expr (new_stmt, true, true); 1437 } 1438 } 1439 1440 /* Replace *OP_P in STMT with any known equivalent value for *OP_P from 1441 CONST_AND_COPIES. */ 1442 1443 static void 1444 cprop_operand (gimple *stmt, use_operand_p op_p) 1445 { 1446 tree val; 1447 tree op = USE_FROM_PTR (op_p); 1448 1449 /* If the operand has a known constant value or it is known to be a 1450 copy of some other variable, use the value or copy stored in 1451 CONST_AND_COPIES. */ 1452 val = SSA_NAME_VALUE (op); 1453 if (val && val != op) 1454 { 1455 /* Do not replace hard register operands in asm statements. */ 1456 if (gimple_code (stmt) == GIMPLE_ASM 1457 && !may_propagate_copy_into_asm (op)) 1458 return; 1459 1460 /* Certain operands are not allowed to be copy propagated due 1461 to their interaction with exception handling and some GCC 1462 extensions. */ 1463 if (!may_propagate_copy (op, val)) 1464 return; 1465 1466 /* Do not propagate copies into BIVs. 1467 See PR23821 and PR62217 for how this can disturb IV and 1468 number of iteration analysis. */ 1469 if (TREE_CODE (val) != INTEGER_CST) 1470 { 1471 gimple *def = SSA_NAME_DEF_STMT (op); 1472 if (gimple_code (def) == GIMPLE_PHI 1473 && gimple_bb (def)->loop_father->header == gimple_bb (def)) 1474 return; 1475 } 1476 1477 /* Dump details. */ 1478 if (dump_file && (dump_flags & TDF_DETAILS)) 1479 { 1480 fprintf (dump_file, " Replaced '"); 1481 print_generic_expr (dump_file, op, dump_flags); 1482 fprintf (dump_file, "' with %s '", 1483 (TREE_CODE (val) != SSA_NAME ? "constant" : "variable")); 1484 print_generic_expr (dump_file, val, dump_flags); 1485 fprintf (dump_file, "'\n"); 1486 } 1487 1488 if (TREE_CODE (val) != SSA_NAME) 1489 opt_stats.num_const_prop++; 1490 else 1491 opt_stats.num_copy_prop++; 1492 1493 propagate_value (op_p, val); 1494 1495 /* And note that we modified this statement. This is now 1496 safe, even if we changed virtual operands since we will 1497 rescan the statement and rewrite its operands again. */ 1498 gimple_set_modified (stmt, true); 1499 } 1500 } 1501 1502 /* CONST_AND_COPIES is a table which maps an SSA_NAME to the current 1503 known value for that SSA_NAME (or NULL if no value is known). 1504 1505 Propagate values from CONST_AND_COPIES into the uses, vuses and 1506 vdef_ops of STMT. */ 1507 1508 static void 1509 cprop_into_stmt (gimple *stmt) 1510 { 1511 use_operand_p op_p; 1512 ssa_op_iter iter; 1513 tree last_copy_propagated_op = NULL; 1514 1515 FOR_EACH_SSA_USE_OPERAND (op_p, stmt, iter, SSA_OP_USE) 1516 { 1517 tree old_op = USE_FROM_PTR (op_p); 1518 1519 /* If we have A = B and B = A in the copy propagation tables 1520 (due to an equality comparison), avoid substituting B for A 1521 then A for B in the trivially discovered cases. This allows 1522 optimization of statements were A and B appear as input 1523 operands. */ 1524 if (old_op != last_copy_propagated_op) 1525 { 1526 cprop_operand (stmt, op_p); 1527 1528 tree new_op = USE_FROM_PTR (op_p); 1529 if (new_op != old_op && TREE_CODE (new_op) == SSA_NAME) 1530 last_copy_propagated_op = new_op; 1531 } 1532 } 1533 } 1534 1535 /* Optimize the statement in block BB pointed to by iterator SI 1536 using equivalences from CONST_AND_COPIES and AVAIL_EXPRS_STACK. 1537 1538 We try to perform some simplistic global redundancy elimination and 1539 constant propagation: 1540 1541 1- To detect global redundancy, we keep track of expressions that have 1542 been computed in this block and its dominators. If we find that the 1543 same expression is computed more than once, we eliminate repeated 1544 computations by using the target of the first one. 1545 1546 2- Constant values and copy assignments. This is used to do very 1547 simplistic constant and copy propagation. When a constant or copy 1548 assignment is found, we map the value on the RHS of the assignment to 1549 the variable in the LHS in the CONST_AND_COPIES table. */ 1550 1551 static edge 1552 optimize_stmt (basic_block bb, gimple_stmt_iterator si, 1553 class const_and_copies *const_and_copies, 1554 class avail_exprs_stack *avail_exprs_stack) 1555 { 1556 gimple *stmt, *old_stmt; 1557 bool may_optimize_p; 1558 bool modified_p = false; 1559 bool was_noreturn; 1560 edge retval = NULL; 1561 1562 old_stmt = stmt = gsi_stmt (si); 1563 was_noreturn = is_gimple_call (stmt) && gimple_call_noreturn_p (stmt); 1564 1565 if (dump_file && (dump_flags & TDF_DETAILS)) 1566 { 1567 fprintf (dump_file, "Optimizing statement "); 1568 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM); 1569 } 1570 1571 update_stmt_if_modified (stmt); 1572 opt_stats.num_stmts++; 1573 1574 /* Const/copy propagate into USES, VUSES and the RHS of VDEFs. */ 1575 cprop_into_stmt (stmt); 1576 1577 /* If the statement has been modified with constant replacements, 1578 fold its RHS before checking for redundant computations. */ 1579 if (gimple_modified_p (stmt)) 1580 { 1581 tree rhs = NULL; 1582 1583 /* Try to fold the statement making sure that STMT is kept 1584 up to date. */ 1585 if (fold_stmt (&si)) 1586 { 1587 stmt = gsi_stmt (si); 1588 gimple_set_modified (stmt, true); 1589 1590 if (dump_file && (dump_flags & TDF_DETAILS)) 1591 { 1592 fprintf (dump_file, " Folded to: "); 1593 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM); 1594 } 1595 } 1596 1597 /* We only need to consider cases that can yield a gimple operand. */ 1598 if (gimple_assign_single_p (stmt)) 1599 rhs = gimple_assign_rhs1 (stmt); 1600 else if (gimple_code (stmt) == GIMPLE_GOTO) 1601 rhs = gimple_goto_dest (stmt); 1602 else if (gswitch *swtch_stmt = dyn_cast <gswitch *> (stmt)) 1603 /* This should never be an ADDR_EXPR. */ 1604 rhs = gimple_switch_index (swtch_stmt); 1605 1606 if (rhs && TREE_CODE (rhs) == ADDR_EXPR) 1607 recompute_tree_invariant_for_addr_expr (rhs); 1608 1609 /* Indicate that maybe_clean_or_replace_eh_stmt needs to be called, 1610 even if fold_stmt updated the stmt already and thus cleared 1611 gimple_modified_p flag on it. */ 1612 modified_p = true; 1613 } 1614 1615 /* Check for redundant computations. Do this optimization only 1616 for assignments that have no volatile ops and conditionals. */ 1617 may_optimize_p = (!gimple_has_side_effects (stmt) 1618 && (is_gimple_assign (stmt) 1619 || (is_gimple_call (stmt) 1620 && gimple_call_lhs (stmt) != NULL_TREE) 1621 || gimple_code (stmt) == GIMPLE_COND 1622 || gimple_code (stmt) == GIMPLE_SWITCH)); 1623 1624 if (may_optimize_p) 1625 { 1626 if (gimple_code (stmt) == GIMPLE_CALL) 1627 { 1628 /* Resolve __builtin_constant_p. If it hasn't been 1629 folded to integer_one_node by now, it's fairly 1630 certain that the value simply isn't constant. */ 1631 tree callee = gimple_call_fndecl (stmt); 1632 if (callee 1633 && DECL_BUILT_IN_CLASS (callee) == BUILT_IN_NORMAL 1634 && DECL_FUNCTION_CODE (callee) == BUILT_IN_CONSTANT_P) 1635 { 1636 propagate_tree_value_into_stmt (&si, integer_zero_node); 1637 stmt = gsi_stmt (si); 1638 } 1639 } 1640 1641 if (gimple_code (stmt) == GIMPLE_COND) 1642 { 1643 tree lhs = gimple_cond_lhs (stmt); 1644 tree rhs = gimple_cond_rhs (stmt); 1645 1646 /* If the LHS has a range [0..1] and the RHS has a range ~[0..1], 1647 then this conditional is computable at compile time. We can just 1648 shove either 0 or 1 into the LHS, mark the statement as modified 1649 and all the right things will just happen below. 1650 1651 Note this would apply to any case where LHS has a range 1652 narrower than its type implies and RHS is outside that 1653 narrower range. Future work. */ 1654 if (TREE_CODE (lhs) == SSA_NAME 1655 && ssa_name_has_boolean_range (lhs) 1656 && TREE_CODE (rhs) == INTEGER_CST 1657 && ! (integer_zerop (rhs) || integer_onep (rhs))) 1658 { 1659 gimple_cond_set_lhs (as_a <gcond *> (stmt), 1660 fold_convert (TREE_TYPE (lhs), 1661 integer_zero_node)); 1662 gimple_set_modified (stmt, true); 1663 } 1664 } 1665 1666 update_stmt_if_modified (stmt); 1667 eliminate_redundant_computations (&si, const_and_copies, 1668 avail_exprs_stack); 1669 stmt = gsi_stmt (si); 1670 1671 /* Perform simple redundant store elimination. */ 1672 if (gimple_assign_single_p (stmt) 1673 && TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) 1674 { 1675 tree lhs = gimple_assign_lhs (stmt); 1676 tree rhs = gimple_assign_rhs1 (stmt); 1677 tree cached_lhs; 1678 gassign *new_stmt; 1679 rhs = dom_valueize (rhs); 1680 /* Build a new statement with the RHS and LHS exchanged. */ 1681 if (TREE_CODE (rhs) == SSA_NAME) 1682 { 1683 gimple *defstmt = SSA_NAME_DEF_STMT (rhs); 1684 new_stmt = gimple_build_assign (rhs, lhs); 1685 SSA_NAME_DEF_STMT (rhs) = defstmt; 1686 } 1687 else 1688 new_stmt = gimple_build_assign (rhs, lhs); 1689 gimple_set_vuse (new_stmt, gimple_vuse (stmt)); 1690 cached_lhs = avail_exprs_stack->lookup_avail_expr (new_stmt, false, 1691 false); 1692 if (cached_lhs 1693 && rhs == cached_lhs) 1694 { 1695 basic_block bb = gimple_bb (stmt); 1696 unlink_stmt_vdef (stmt); 1697 if (gsi_remove (&si, true)) 1698 { 1699 bitmap_set_bit (need_eh_cleanup, bb->index); 1700 if (dump_file && (dump_flags & TDF_DETAILS)) 1701 fprintf (dump_file, " Flagged to clear EH edges.\n"); 1702 } 1703 release_defs (stmt); 1704 return retval; 1705 } 1706 } 1707 } 1708 1709 /* Record any additional equivalences created by this statement. */ 1710 if (is_gimple_assign (stmt)) 1711 record_equivalences_from_stmt (stmt, may_optimize_p, avail_exprs_stack); 1712 1713 /* If STMT is a COND_EXPR or SWITCH_EXPR and it was modified, then we may 1714 know where it goes. */ 1715 if (gimple_modified_p (stmt) || modified_p) 1716 { 1717 tree val = NULL; 1718 1719 if (gimple_code (stmt) == GIMPLE_COND) 1720 val = fold_binary_loc (gimple_location (stmt), 1721 gimple_cond_code (stmt), boolean_type_node, 1722 gimple_cond_lhs (stmt), 1723 gimple_cond_rhs (stmt)); 1724 else if (gswitch *swtch_stmt = dyn_cast <gswitch *> (stmt)) 1725 val = gimple_switch_index (swtch_stmt); 1726 1727 if (val && TREE_CODE (val) == INTEGER_CST) 1728 { 1729 retval = find_taken_edge (bb, val); 1730 if (retval) 1731 { 1732 /* Fix the condition to be either true or false. */ 1733 if (gimple_code (stmt) == GIMPLE_COND) 1734 { 1735 if (integer_zerop (val)) 1736 gimple_cond_make_false (as_a <gcond *> (stmt)); 1737 else if (integer_onep (val)) 1738 gimple_cond_make_true (as_a <gcond *> (stmt)); 1739 else 1740 gcc_unreachable (); 1741 1742 gimple_set_modified (stmt, true); 1743 } 1744 1745 /* Further simplifications may be possible. */ 1746 cfg_altered = true; 1747 } 1748 } 1749 1750 update_stmt_if_modified (stmt); 1751 1752 /* If we simplified a statement in such a way as to be shown that it 1753 cannot trap, update the eh information and the cfg to match. */ 1754 if (maybe_clean_or_replace_eh_stmt (old_stmt, stmt)) 1755 { 1756 bitmap_set_bit (need_eh_cleanup, bb->index); 1757 if (dump_file && (dump_flags & TDF_DETAILS)) 1758 fprintf (dump_file, " Flagged to clear EH edges.\n"); 1759 } 1760 1761 if (!was_noreturn 1762 && is_gimple_call (stmt) && gimple_call_noreturn_p (stmt)) 1763 need_noreturn_fixup.safe_push (stmt); 1764 } 1765 return retval; 1766 } 1767