1 /* Basic IPA optimizations and utilities. 2 Copyright (C) 2003-2019 Free Software Foundation, Inc. 3 4 This file is part of GCC. 5 6 GCC is free software; you can redistribute it and/or modify it under 7 the terms of the GNU General Public License as published by the Free 8 Software Foundation; either version 3, or (at your option) any later 9 version. 10 11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY 12 WARRANTY; without even the implied warranty of MERCHANTABILITY or 13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 14 for more details. 15 16 You should have received a copy of the GNU General Public License 17 along with GCC; see the file COPYING3. If not see 18 <http://www.gnu.org/licenses/>. */ 19 20 #include "config.h" 21 #include "system.h" 22 #include "coretypes.h" 23 #include "backend.h" 24 #include "target.h" 25 #include "tree.h" 26 #include "gimple.h" 27 #include "alloc-pool.h" 28 #include "tree-pass.h" 29 #include "stringpool.h" 30 #include "cgraph.h" 31 #include "gimplify.h" 32 #include "tree-iterator.h" 33 #include "ipa-utils.h" 34 #include "symbol-summary.h" 35 #include "tree-vrp.h" 36 #include "ipa-prop.h" 37 #include "ipa-fnsummary.h" 38 #include "dbgcnt.h" 39 #include "debug.h" 40 #include "stringpool.h" 41 #include "attribs.h" 42 43 /* Return true when NODE has ADDR reference. */ 44 45 static bool 46 has_addr_references_p (struct cgraph_node *node, 47 void *) 48 { 49 int i; 50 struct ipa_ref *ref = NULL; 51 52 for (i = 0; node->iterate_referring (i, ref); i++) 53 if (ref->use == IPA_REF_ADDR) 54 return true; 55 return false; 56 } 57 58 /* Return true when NODE can be target of an indirect call. */ 59 60 static bool 61 is_indirect_call_target_p (struct cgraph_node *node, void *) 62 { 63 return node->indirect_call_target; 64 } 65 66 /* Look for all functions inlined to NODE and update their inlined_to pointers 67 to INLINED_TO. */ 68 69 static void 70 update_inlined_to_pointer (struct cgraph_node *node, struct cgraph_node *inlined_to) 71 { 72 struct cgraph_edge *e; 73 for (e = node->callees; e; e = e->next_callee) 74 if (e->callee->global.inlined_to) 75 { 76 e->callee->global.inlined_to = inlined_to; 77 update_inlined_to_pointer (e->callee, inlined_to); 78 } 79 } 80 81 /* Add symtab NODE to queue starting at FIRST. 82 83 The queue is linked via AUX pointers and terminated by pointer to 1. 84 We enqueue nodes at two occasions: when we find them reachable or when we find 85 their bodies needed for further clonning. In the second case we mark them 86 by pointer to 2 after processing so they are re-queue when they become 87 reachable. */ 88 89 static void 90 enqueue_node (symtab_node *node, symtab_node **first, 91 hash_set<symtab_node *> *reachable) 92 { 93 /* Node is still in queue; do nothing. */ 94 if (node->aux && node->aux != (void *) 2) 95 return; 96 /* Node was already processed as unreachable, re-enqueue 97 only if it became reachable now. */ 98 if (node->aux == (void *)2 && !reachable->contains (node)) 99 return; 100 node->aux = *first; 101 *first = node; 102 } 103 104 /* Return true if NODE may get inlined later. 105 This is used to keep DECL_EXTERNAL function bodies around long enough 106 so inliner can proces them. */ 107 108 static bool 109 possible_inline_candidate_p (symtab_node *node) 110 { 111 if (symtab->state >= IPA_SSA_AFTER_INLINING) 112 return false; 113 cgraph_node *cnode = dyn_cast <cgraph_node *> (node); 114 if (!cnode) 115 return false; 116 if (DECL_UNINLINABLE (cnode->decl)) 117 return false; 118 if (opt_for_fn (cnode->decl, optimize)) 119 return true; 120 if (symtab->state >= IPA_SSA) 121 return false; 122 return lookup_attribute ("always_inline", DECL_ATTRIBUTES (node->decl)); 123 } 124 125 /* Process references. */ 126 127 static void 128 process_references (symtab_node *snode, 129 symtab_node **first, 130 hash_set<symtab_node *> *reachable) 131 { 132 int i; 133 struct ipa_ref *ref = NULL; 134 for (i = 0; snode->iterate_reference (i, ref); i++) 135 { 136 symtab_node *node = ref->referred; 137 symtab_node *body = node->ultimate_alias_target (); 138 139 if (node->definition && !node->in_other_partition 140 && ((!DECL_EXTERNAL (node->decl) || node->alias) 141 || (possible_inline_candidate_p (node) 142 /* We use variable constructors during late compilation for 143 constant folding. Keep references alive so partitioning 144 knows about potential references. */ 145 || (VAR_P (node->decl) 146 && (flag_wpa 147 || flag_incremental_link 148 == INCREMENTAL_LINK_LTO) 149 && dyn_cast <varpool_node *> (node) 150 ->ctor_useable_for_folding_p ())))) 151 { 152 /* Be sure that we will not optimize out alias target 153 body. */ 154 if (DECL_EXTERNAL (node->decl) 155 && node->alias 156 && symtab->state < IPA_SSA_AFTER_INLINING) 157 reachable->add (body); 158 reachable->add (node); 159 } 160 enqueue_node (node, first, reachable); 161 } 162 } 163 164 /* EDGE is an polymorphic call. If BEFORE_INLINING_P is set, mark 165 all its potential targets as reachable to permit later inlining if 166 devirtualization happens. After inlining still keep their declarations 167 around, so we can devirtualize to a direct call. 168 169 Also try to make trivial devirutalization when no or only one target is 170 possible. */ 171 172 static void 173 walk_polymorphic_call_targets (hash_set<void *> *reachable_call_targets, 174 struct cgraph_edge *edge, 175 symtab_node **first, 176 hash_set<symtab_node *> *reachable) 177 { 178 unsigned int i; 179 void *cache_token; 180 bool final; 181 vec <cgraph_node *>targets 182 = possible_polymorphic_call_targets 183 (edge, &final, &cache_token); 184 185 if (!reachable_call_targets->add (cache_token)) 186 { 187 for (i = 0; i < targets.length (); i++) 188 { 189 struct cgraph_node *n = targets[i]; 190 191 /* Do not bother to mark virtual methods in anonymous namespace; 192 either we will find use of virtual table defining it, or it is 193 unused. */ 194 if (TREE_CODE (TREE_TYPE (n->decl)) == METHOD_TYPE 195 && type_in_anonymous_namespace_p 196 (TYPE_METHOD_BASETYPE (TREE_TYPE (n->decl)))) 197 continue; 198 199 n->indirect_call_target = true; 200 symtab_node *body = n->function_symbol (); 201 202 /* Prior inlining, keep alive bodies of possible targets for 203 devirtualization. */ 204 if (n->definition 205 && (possible_inline_candidate_p (body) 206 && opt_for_fn (body->decl, flag_devirtualize))) 207 { 208 /* Be sure that we will not optimize out alias target 209 body. */ 210 if (DECL_EXTERNAL (n->decl) 211 && n->alias 212 && symtab->state < IPA_SSA_AFTER_INLINING) 213 reachable->add (body); 214 reachable->add (n); 215 } 216 /* Even after inlining we want to keep the possible targets in the 217 boundary, so late passes can still produce direct call even if 218 the chance for inlining is lost. */ 219 enqueue_node (n, first, reachable); 220 } 221 } 222 223 /* Very trivial devirtualization; when the type is 224 final or anonymous (so we know all its derivation) 225 and there is only one possible virtual call target, 226 make the edge direct. */ 227 if (final) 228 { 229 if (targets.length () <= 1 && dbg_cnt (devirt)) 230 { 231 cgraph_node *target, *node = edge->caller; 232 if (targets.length () == 1) 233 target = targets[0]; 234 else 235 target = cgraph_node::get_create 236 (builtin_decl_implicit (BUILT_IN_UNREACHABLE)); 237 238 if (dump_enabled_p ()) 239 { 240 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, edge->call_stmt, 241 "devirtualizing call in %s to %s\n", 242 edge->caller->dump_name (), 243 target->dump_name ()); 244 } 245 edge = edge->make_direct (target); 246 if (ipa_fn_summaries) 247 ipa_update_overall_fn_summary (node); 248 else if (edge->call_stmt) 249 edge->redirect_call_stmt_to_callee (); 250 } 251 } 252 } 253 254 /* Perform reachability analysis and reclaim all unreachable nodes. 255 256 The algorithm is basically mark&sweep but with some extra refinements: 257 258 - reachable extern inline functions needs special handling; the bodies needs 259 to stay in memory until inlining in hope that they will be inlined. 260 After inlining we release their bodies and turn them into unanalyzed 261 nodes even when they are reachable. 262 263 - virtual functions are kept in callgraph even if they seem unreachable in 264 hope calls to them will be devirtualized. 265 266 Again we remove them after inlining. In late optimization some 267 devirtualization may happen, but it is not important since we won't inline 268 the call. In theory early opts and IPA should work out all important cases. 269 270 - virtual clones needs bodies of their origins for later materialization; 271 this means that we want to keep the body even if the origin is unreachable 272 otherwise. To avoid origin from sitting in the callgraph and being 273 walked by IPA passes, we turn them into unanalyzed nodes with body 274 defined. 275 276 We maintain set of function declaration where body needs to stay in 277 body_needed_for_clonning 278 279 Inline clones represent special case: their declaration match the 280 declaration of origin and cgraph_remove_node already knows how to 281 reshape callgraph and preserve body when offline copy of function or 282 inline clone is being removed. 283 284 - C++ virtual tables keyed to other unit are represented as DECL_EXTERNAL 285 variables with DECL_INITIAL set. We finalize these and keep reachable 286 ones around for constant folding purposes. After inlining we however 287 stop walking their references to let everything static referneced by them 288 to be removed when it is otherwise unreachable. 289 290 We maintain queue of both reachable symbols (i.e. defined symbols that needs 291 to stay) and symbols that are in boundary (i.e. external symbols referenced 292 by reachable symbols or origins of clones). The queue is represented 293 as linked list by AUX pointer terminated by 1. 294 295 At the end we keep all reachable symbols. For symbols in boundary we always 296 turn definition into a declaration, but we may keep function body around 297 based on body_needed_for_clonning 298 299 All symbols that enter the queue have AUX pointer non-zero and are in the 300 boundary. Pointer set REACHABLE is used to track reachable symbols. 301 302 Every symbol can be visited twice - once as part of boundary and once 303 as real reachable symbol. enqueue_node needs to decide whether the 304 node needs to be re-queued for second processing. For this purpose 305 we set AUX pointer of processed symbols in the boundary to constant 2. */ 306 307 bool 308 symbol_table::remove_unreachable_nodes (FILE *file) 309 { 310 symtab_node *first = (symtab_node *) (void *) 1; 311 struct cgraph_node *node, *next; 312 varpool_node *vnode, *vnext; 313 bool changed = false; 314 hash_set<symtab_node *> reachable; 315 hash_set<tree> body_needed_for_clonning; 316 hash_set<void *> reachable_call_targets; 317 318 timevar_push (TV_IPA_UNREACHABLE); 319 build_type_inheritance_graph (); 320 if (file) 321 fprintf (file, "\nReclaiming functions:"); 322 if (flag_checking) 323 { 324 FOR_EACH_FUNCTION (node) 325 gcc_assert (!node->aux); 326 FOR_EACH_VARIABLE (vnode) 327 gcc_assert (!vnode->aux); 328 } 329 /* Mark functions whose bodies are obviously needed. 330 This is mostly when they can be referenced externally. Inline clones 331 are special since their declarations are shared with master clone and thus 332 cgraph_can_remove_if_no_direct_calls_and_refs_p should not be called on them. */ 333 FOR_EACH_FUNCTION (node) 334 { 335 node->used_as_abstract_origin = false; 336 node->indirect_call_target = false; 337 if (node->definition 338 && !node->global.inlined_to 339 && !node->in_other_partition 340 && !node->can_remove_if_no_direct_calls_and_refs_p ()) 341 { 342 gcc_assert (!node->global.inlined_to); 343 reachable.add (node); 344 enqueue_node (node, &first, &reachable); 345 } 346 else 347 gcc_assert (!node->aux); 348 } 349 350 /* Mark variables that are obviously needed. */ 351 FOR_EACH_DEFINED_VARIABLE (vnode) 352 if (!vnode->can_remove_if_no_refs_p() 353 && !vnode->in_other_partition) 354 { 355 reachable.add (vnode); 356 enqueue_node (vnode, &first, &reachable); 357 } 358 359 /* Perform reachability analysis. */ 360 while (first != (symtab_node *) (void *) 1) 361 { 362 bool in_boundary_p = !reachable.contains (first); 363 symtab_node *node = first; 364 365 first = (symtab_node *)first->aux; 366 367 /* If we are processing symbol in boundary, mark its AUX pointer for 368 possible later re-processing in enqueue_node. */ 369 if (in_boundary_p) 370 { 371 node->aux = (void *)2; 372 if (node->alias && node->analyzed) 373 enqueue_node (node->get_alias_target (), &first, &reachable); 374 } 375 else 376 { 377 if (TREE_CODE (node->decl) == FUNCTION_DECL 378 && DECL_ABSTRACT_ORIGIN (node->decl)) 379 { 380 struct cgraph_node *origin_node 381 = cgraph_node::get (DECL_ABSTRACT_ORIGIN (node->decl)); 382 if (origin_node && !origin_node->used_as_abstract_origin) 383 { 384 origin_node->used_as_abstract_origin = true; 385 gcc_assert (!origin_node->prev_sibling_clone); 386 gcc_assert (!origin_node->next_sibling_clone); 387 for (cgraph_node *n = origin_node->clones; n; 388 n = n->next_sibling_clone) 389 if (n->decl == DECL_ABSTRACT_ORIGIN (node->decl)) 390 n->used_as_abstract_origin = true; 391 } 392 } 393 /* If any symbol in a comdat group is reachable, force 394 all externally visible symbols in the same comdat 395 group to be reachable as well. Comdat-local symbols 396 can be discarded if all uses were inlined. */ 397 if (node->same_comdat_group) 398 { 399 symtab_node *next; 400 for (next = node->same_comdat_group; 401 next != node; 402 next = next->same_comdat_group) 403 if (!next->comdat_local_p () 404 && !reachable.add (next)) 405 enqueue_node (next, &first, &reachable); 406 } 407 /* Mark references as reachable. */ 408 process_references (node, &first, &reachable); 409 } 410 411 if (cgraph_node *cnode = dyn_cast <cgraph_node *> (node)) 412 { 413 /* Mark the callees reachable unless they are direct calls to extern 414 inline functions we decided to not inline. */ 415 if (!in_boundary_p) 416 { 417 struct cgraph_edge *e; 418 /* Keep alive possible targets for devirtualization. */ 419 if (opt_for_fn (cnode->decl, optimize) 420 && opt_for_fn (cnode->decl, flag_devirtualize)) 421 { 422 struct cgraph_edge *next; 423 for (e = cnode->indirect_calls; e; e = next) 424 { 425 next = e->next_callee; 426 if (e->indirect_info->polymorphic) 427 walk_polymorphic_call_targets (&reachable_call_targets, 428 e, &first, &reachable); 429 } 430 } 431 for (e = cnode->callees; e; e = e->next_callee) 432 { 433 symtab_node *body = e->callee->function_symbol (); 434 if (e->callee->definition 435 && !e->callee->in_other_partition 436 && (!e->inline_failed 437 || !DECL_EXTERNAL (e->callee->decl) 438 || e->callee->alias 439 || possible_inline_candidate_p (e->callee))) 440 { 441 /* Be sure that we will not optimize out alias target 442 body. */ 443 if (DECL_EXTERNAL (e->callee->decl) 444 && e->callee->alias 445 && symtab->state < IPA_SSA_AFTER_INLINING) 446 reachable.add (body); 447 reachable.add (e->callee); 448 } 449 enqueue_node (e->callee, &first, &reachable); 450 } 451 452 /* When inline clone exists, mark body to be preserved so when removing 453 offline copy of the function we don't kill it. */ 454 if (cnode->global.inlined_to) 455 body_needed_for_clonning.add (cnode->decl); 456 457 /* For non-inline clones, force their origins to the boundary and ensure 458 that body is not removed. */ 459 while (cnode->clone_of) 460 { 461 bool noninline = cnode->clone_of->decl != cnode->decl; 462 cnode = cnode->clone_of; 463 if (noninline) 464 { 465 body_needed_for_clonning.add (cnode->decl); 466 enqueue_node (cnode, &first, &reachable); 467 } 468 } 469 470 } 471 else if (cnode->thunk.thunk_p) 472 enqueue_node (cnode->callees->callee, &first, &reachable); 473 474 /* If any reachable function has simd clones, mark them as 475 reachable as well. */ 476 if (cnode->simd_clones) 477 { 478 cgraph_node *next; 479 for (next = cnode->simd_clones; 480 next; 481 next = next->simdclone->next_clone) 482 if (in_boundary_p 483 || !reachable.add (next)) 484 enqueue_node (next, &first, &reachable); 485 } 486 } 487 /* When we see constructor of external variable, keep referred nodes in the 488 boundary. This will also hold initializers of the external vars NODE 489 refers to. */ 490 varpool_node *vnode = dyn_cast <varpool_node *> (node); 491 if (vnode 492 && DECL_EXTERNAL (node->decl) 493 && !vnode->alias 494 && in_boundary_p) 495 { 496 struct ipa_ref *ref = NULL; 497 for (int i = 0; node->iterate_reference (i, ref); i++) 498 enqueue_node (ref->referred, &first, &reachable); 499 } 500 } 501 502 /* Remove unreachable functions. */ 503 for (node = first_function (); node; node = next) 504 { 505 next = next_function (node); 506 507 /* If node is not needed at all, remove it. */ 508 if (!node->aux) 509 { 510 if (file) 511 fprintf (file, " %s", node->dump_name ()); 512 node->remove (); 513 changed = true; 514 } 515 /* If node is unreachable, remove its body. */ 516 else if (!reachable.contains (node)) 517 { 518 /* We keep definitions of thunks and aliases in the boundary so 519 we can walk to the ultimate alias targets and function symbols 520 reliably. */ 521 if (node->alias || node->thunk.thunk_p) 522 ; 523 else if (!body_needed_for_clonning.contains (node->decl) 524 && !node->alias && !node->thunk.thunk_p) 525 node->release_body (); 526 else if (!node->clone_of) 527 gcc_assert (in_lto_p || DECL_RESULT (node->decl)); 528 if (node->definition && !node->alias && !node->thunk.thunk_p) 529 { 530 if (file) 531 fprintf (file, " %s", node->dump_name ()); 532 node->body_removed = true; 533 node->analyzed = false; 534 node->definition = false; 535 node->cpp_implicit_alias = false; 536 node->alias = false; 537 node->transparent_alias = false; 538 node->thunk.thunk_p = false; 539 node->weakref = false; 540 /* After early inlining we drop always_inline attributes on 541 bodies of functions that are still referenced (have their 542 address taken). */ 543 DECL_ATTRIBUTES (node->decl) 544 = remove_attribute ("always_inline", 545 DECL_ATTRIBUTES (node->decl)); 546 if (!node->in_other_partition) 547 node->local.local = false; 548 node->remove_callees (); 549 node->remove_all_references (); 550 changed = true; 551 } 552 } 553 else 554 gcc_assert (node->clone_of || !node->has_gimple_body_p () 555 || in_lto_p || DECL_RESULT (node->decl)); 556 } 557 558 /* Inline clones might be kept around so their materializing allows further 559 cloning. If the function the clone is inlined into is removed, we need 560 to turn it into normal cone. */ 561 FOR_EACH_FUNCTION (node) 562 { 563 if (node->global.inlined_to 564 && !node->callers) 565 { 566 gcc_assert (node->clones); 567 node->global.inlined_to = NULL; 568 update_inlined_to_pointer (node, node); 569 } 570 node->aux = NULL; 571 } 572 573 /* Remove unreachable variables. */ 574 if (file) 575 fprintf (file, "\nReclaiming variables:"); 576 for (vnode = first_variable (); vnode; vnode = vnext) 577 { 578 vnext = next_variable (vnode); 579 if (!vnode->aux 580 /* For can_refer_decl_in_current_unit_p we want to track for 581 all external variables if they are defined in other partition 582 or not. */ 583 && (!flag_ltrans || !DECL_EXTERNAL (vnode->decl))) 584 { 585 struct ipa_ref *ref = NULL; 586 587 /* First remove the aliases, so varpool::remove can possibly lookup 588 the constructor and save it for future use. */ 589 while (vnode->iterate_direct_aliases (0, ref)) 590 { 591 if (file) 592 fprintf (file, " %s", ref->referred->dump_name ()); 593 ref->referring->remove (); 594 } 595 if (file) 596 fprintf (file, " %s", vnode->dump_name ()); 597 vnext = next_variable (vnode); 598 /* Signal removal to the debug machinery. */ 599 if (! flag_wpa || flag_incremental_link == INCREMENTAL_LINK_LTO) 600 { 601 vnode->definition = false; 602 (*debug_hooks->late_global_decl) (vnode->decl); 603 } 604 vnode->remove (); 605 changed = true; 606 } 607 else if (!reachable.contains (vnode) && !vnode->alias) 608 { 609 tree init; 610 if (vnode->definition) 611 { 612 if (file) 613 fprintf (file, " %s", vnode->name ()); 614 changed = true; 615 } 616 /* Keep body if it may be useful for constant folding. */ 617 if ((flag_wpa || flag_incremental_link == INCREMENTAL_LINK_LTO) 618 || ((init = ctor_for_folding (vnode->decl)) == error_mark_node)) 619 vnode->remove_initializer (); 620 else 621 DECL_INITIAL (vnode->decl) = init; 622 vnode->body_removed = true; 623 vnode->definition = false; 624 vnode->analyzed = false; 625 vnode->aux = NULL; 626 627 vnode->remove_from_same_comdat_group (); 628 629 vnode->remove_all_references (); 630 } 631 else 632 vnode->aux = NULL; 633 } 634 635 /* Now update address_taken flags and try to promote functions to be local. */ 636 if (file) 637 fprintf (file, "\nClearing address taken flags:"); 638 FOR_EACH_DEFINED_FUNCTION (node) 639 if (node->address_taken 640 && !node->used_from_other_partition) 641 { 642 if (!node->call_for_symbol_and_aliases 643 (has_addr_references_p, NULL, true)) 644 { 645 if (file) 646 fprintf (file, " %s", node->name ()); 647 node->address_taken = false; 648 changed = true; 649 if (node->local_p () 650 /* Virtual functions may be kept in cgraph just because 651 of possible later devirtualization. Do not mark them as 652 local too early so we won't optimize them out before 653 we are done with polymorphic call analysis. */ 654 && (symtab->state >= IPA_SSA_AFTER_INLINING 655 || !node->call_for_symbol_and_aliases 656 (is_indirect_call_target_p, NULL, true))) 657 { 658 node->local.local = true; 659 if (file) 660 fprintf (file, " (local)"); 661 } 662 } 663 } 664 if (file) 665 fprintf (file, "\n"); 666 667 symtab_node::checking_verify_symtab_nodes (); 668 669 /* If we removed something, perhaps profile could be improved. */ 670 if (changed && (optimize || in_lto_p) && ipa_call_summaries) 671 FOR_EACH_DEFINED_FUNCTION (node) 672 ipa_propagate_frequency (node); 673 674 timevar_pop (TV_IPA_UNREACHABLE); 675 return changed; 676 } 677 678 /* Process references to VNODE and set flags WRITTEN, ADDRESS_TAKEN, READ 679 as needed, also clear EXPLICIT_REFS if the references to given variable 680 do not need to be explicit. */ 681 682 void 683 process_references (varpool_node *vnode, 684 bool *written, bool *address_taken, 685 bool *read, bool *explicit_refs) 686 { 687 int i; 688 struct ipa_ref *ref; 689 690 if (!vnode->all_refs_explicit_p () 691 || TREE_THIS_VOLATILE (vnode->decl)) 692 *explicit_refs = false; 693 694 for (i = 0; vnode->iterate_referring (i, ref) 695 && *explicit_refs && (!*written || !*address_taken || !*read); i++) 696 switch (ref->use) 697 { 698 case IPA_REF_ADDR: 699 *address_taken = true; 700 break; 701 case IPA_REF_LOAD: 702 *read = true; 703 break; 704 case IPA_REF_STORE: 705 *written = true; 706 break; 707 case IPA_REF_ALIAS: 708 process_references (dyn_cast<varpool_node *> (ref->referring), written, 709 address_taken, read, explicit_refs); 710 break; 711 } 712 } 713 714 /* Set TREE_READONLY bit. */ 715 716 bool 717 set_readonly_bit (varpool_node *vnode, void *data ATTRIBUTE_UNUSED) 718 { 719 TREE_READONLY (vnode->decl) = true; 720 return false; 721 } 722 723 /* Set writeonly bit and clear the initalizer, since it will not be needed. */ 724 725 bool 726 set_writeonly_bit (varpool_node *vnode, void *data) 727 { 728 vnode->writeonly = true; 729 if (optimize || in_lto_p) 730 { 731 DECL_INITIAL (vnode->decl) = NULL; 732 if (!vnode->alias) 733 { 734 if (vnode->num_references ()) 735 *(bool *)data = true; 736 vnode->remove_all_references (); 737 } 738 } 739 return false; 740 } 741 742 /* Clear addressale bit of VNODE. */ 743 744 bool 745 clear_addressable_bit (varpool_node *vnode, void *data ATTRIBUTE_UNUSED) 746 { 747 vnode->address_taken = false; 748 TREE_ADDRESSABLE (vnode->decl) = 0; 749 return false; 750 } 751 752 /* Discover variables that have no longer address taken, are read-only or 753 write-only and update their flags. 754 755 Return true when unreachable symbol removal should be done. 756 757 FIXME: This cannot be done in between gimplify and omp_expand since 758 readonly flag plays role on what is shared and what is not. Currently we do 759 this transformation as part of whole program visibility and re-do at 760 ipa-reference pass (to take into account clonning), but it would 761 make sense to do it before early optimizations. */ 762 763 bool 764 ipa_discover_variable_flags (void) 765 { 766 if (!flag_ipa_reference_addressable) 767 return false; 768 769 bool remove_p = false; 770 varpool_node *vnode; 771 if (dump_file) 772 fprintf (dump_file, "Clearing variable flags:"); 773 FOR_EACH_VARIABLE (vnode) 774 if (!vnode->alias 775 && (TREE_ADDRESSABLE (vnode->decl) 776 || !vnode->writeonly 777 || !TREE_READONLY (vnode->decl))) 778 { 779 bool written = false; 780 bool address_taken = false; 781 bool read = false; 782 bool explicit_refs = true; 783 784 process_references (vnode, &written, &address_taken, &read, 785 &explicit_refs); 786 if (!explicit_refs) 787 continue; 788 if (!address_taken) 789 { 790 if (TREE_ADDRESSABLE (vnode->decl) && dump_file) 791 fprintf (dump_file, " %s (non-addressable)", vnode->name ()); 792 vnode->call_for_symbol_and_aliases (clear_addressable_bit, NULL, 793 true); 794 } 795 if (!address_taken && !written 796 /* Making variable in explicit section readonly can cause section 797 type conflict. 798 See e.g. gcc.c-torture/compile/pr23237.c */ 799 && vnode->get_section () == NULL) 800 { 801 if (!TREE_READONLY (vnode->decl) && dump_file) 802 fprintf (dump_file, " %s (read-only)", vnode->name ()); 803 vnode->call_for_symbol_and_aliases (set_readonly_bit, NULL, true); 804 } 805 if (!vnode->writeonly && !read && !address_taken && written) 806 { 807 if (dump_file) 808 fprintf (dump_file, " %s (write-only)", vnode->name ()); 809 vnode->call_for_symbol_and_aliases (set_writeonly_bit, &remove_p, 810 true); 811 } 812 } 813 if (dump_file) 814 fprintf (dump_file, "\n"); 815 return remove_p; 816 } 817 818 /* Generate and emit a static constructor or destructor. WHICH must 819 be one of 'I' (for a constructor), 'D' (for a destructor). 820 BODY is a STATEMENT_LIST containing GENERIC 821 statements. PRIORITY is the initialization priority for this 822 constructor or destructor. 823 824 FINAL specify whether the externally visible name for collect2 should 825 be produced. */ 826 827 static void 828 cgraph_build_static_cdtor_1 (char which, tree body, int priority, bool final, 829 tree optimization, 830 tree target) 831 { 832 static int counter = 0; 833 char which_buf[16]; 834 tree decl, name, resdecl; 835 836 /* The priority is encoded in the constructor or destructor name. 837 collect2 will sort the names and arrange that they are called at 838 program startup. */ 839 if (final) 840 sprintf (which_buf, "%c_%.5d_%d", which, priority, counter++); 841 else 842 /* Proudce sane name but one not recognizable by collect2, just for the 843 case we fail to inline the function. */ 844 sprintf (which_buf, "sub_%c_%.5d_%d", which, priority, counter++); 845 name = get_file_function_name (which_buf); 846 847 decl = build_decl (input_location, FUNCTION_DECL, name, 848 build_function_type_list (void_type_node, NULL_TREE)); 849 current_function_decl = decl; 850 851 resdecl = build_decl (input_location, 852 RESULT_DECL, NULL_TREE, void_type_node); 853 DECL_ARTIFICIAL (resdecl) = 1; 854 DECL_RESULT (decl) = resdecl; 855 DECL_CONTEXT (resdecl) = decl; 856 857 allocate_struct_function (decl, false); 858 859 TREE_STATIC (decl) = 1; 860 TREE_USED (decl) = 1; 861 DECL_FUNCTION_SPECIFIC_OPTIMIZATION (decl) = optimization; 862 DECL_FUNCTION_SPECIFIC_TARGET (decl) = target; 863 DECL_ARTIFICIAL (decl) = 1; 864 DECL_IGNORED_P (decl) = 1; 865 DECL_NO_INSTRUMENT_FUNCTION_ENTRY_EXIT (decl) = 1; 866 DECL_SAVED_TREE (decl) = body; 867 if (!targetm.have_ctors_dtors && final) 868 { 869 TREE_PUBLIC (decl) = 1; 870 DECL_PRESERVE_P (decl) = 1; 871 } 872 DECL_UNINLINABLE (decl) = 1; 873 874 DECL_INITIAL (decl) = make_node (BLOCK); 875 BLOCK_SUPERCONTEXT (DECL_INITIAL (decl)) = decl; 876 TREE_USED (DECL_INITIAL (decl)) = 1; 877 878 DECL_SOURCE_LOCATION (decl) = input_location; 879 cfun->function_end_locus = input_location; 880 881 switch (which) 882 { 883 case 'I': 884 DECL_STATIC_CONSTRUCTOR (decl) = 1; 885 decl_init_priority_insert (decl, priority); 886 break; 887 case 'D': 888 DECL_STATIC_DESTRUCTOR (decl) = 1; 889 decl_fini_priority_insert (decl, priority); 890 break; 891 default: 892 gcc_unreachable (); 893 } 894 895 gimplify_function_tree (decl); 896 897 cgraph_node::add_new_function (decl, false); 898 899 set_cfun (NULL); 900 current_function_decl = NULL; 901 } 902 903 /* Generate and emit a static constructor or destructor. WHICH must 904 be one of 'I' (for a constructor) or 'D' (for a destructor). 905 BODY is a STATEMENT_LIST containing GENERIC 906 statements. PRIORITY is the initialization priority for this 907 constructor or destructor. */ 908 909 void 910 cgraph_build_static_cdtor (char which, tree body, int priority) 911 { 912 cgraph_build_static_cdtor_1 (which, body, priority, false, NULL, NULL); 913 } 914 915 /* When target does not have ctors and dtors, we call all constructor 916 and destructor by special initialization/destruction function 917 recognized by collect2. 918 919 When we are going to build this function, collect all constructors and 920 destructors and turn them into normal functions. */ 921 922 static void 923 record_cdtor_fn (struct cgraph_node *node, vec<tree> *ctors, vec<tree> *dtors) 924 { 925 if (DECL_STATIC_CONSTRUCTOR (node->decl)) 926 ctors->safe_push (node->decl); 927 if (DECL_STATIC_DESTRUCTOR (node->decl)) 928 dtors->safe_push (node->decl); 929 node = cgraph_node::get (node->decl); 930 DECL_DISREGARD_INLINE_LIMITS (node->decl) = 1; 931 } 932 933 /* Define global constructors/destructor functions for the CDTORS, of 934 which they are LEN. The CDTORS are sorted by initialization 935 priority. If CTOR_P is true, these are constructors; otherwise, 936 they are destructors. */ 937 938 static void 939 build_cdtor (bool ctor_p, const vec<tree> &cdtors) 940 { 941 size_t i,j; 942 size_t len = cdtors.length (); 943 944 i = 0; 945 while (i < len) 946 { 947 tree body; 948 tree fn; 949 priority_type priority; 950 951 priority = 0; 952 body = NULL_TREE; 953 j = i; 954 do 955 { 956 priority_type p; 957 fn = cdtors[j]; 958 p = ctor_p ? DECL_INIT_PRIORITY (fn) : DECL_FINI_PRIORITY (fn); 959 if (j == i) 960 priority = p; 961 else if (p != priority) 962 break; 963 j++; 964 } 965 while (j < len); 966 967 /* When there is only one cdtor and target supports them, do nothing. */ 968 if (j == i + 1 969 && targetm.have_ctors_dtors) 970 { 971 i++; 972 continue; 973 } 974 /* Find the next batch of constructors/destructors with the same 975 initialization priority. */ 976 for (;i < j; i++) 977 { 978 tree call; 979 fn = cdtors[i]; 980 call = build_call_expr (fn, 0); 981 if (ctor_p) 982 DECL_STATIC_CONSTRUCTOR (fn) = 0; 983 else 984 DECL_STATIC_DESTRUCTOR (fn) = 0; 985 /* We do not want to optimize away pure/const calls here. 986 When optimizing, these should be already removed, when not 987 optimizing, we want user to be able to breakpoint in them. */ 988 TREE_SIDE_EFFECTS (call) = 1; 989 append_to_statement_list (call, &body); 990 } 991 gcc_assert (body != NULL_TREE); 992 /* Generate a function to call all the function of like 993 priority. */ 994 cgraph_build_static_cdtor_1 (ctor_p ? 'I' : 'D', body, priority, true, 995 DECL_FUNCTION_SPECIFIC_OPTIMIZATION (cdtors[0]), 996 DECL_FUNCTION_SPECIFIC_TARGET (cdtors[0])); 997 } 998 } 999 1000 /* Comparison function for qsort. P1 and P2 are actually of type 1001 "tree *" and point to static constructors. DECL_INIT_PRIORITY is 1002 used to determine the sort order. */ 1003 1004 static int 1005 compare_ctor (const void *p1, const void *p2) 1006 { 1007 tree f1; 1008 tree f2; 1009 int priority1; 1010 int priority2; 1011 1012 f1 = *(const tree *)p1; 1013 f2 = *(const tree *)p2; 1014 priority1 = DECL_INIT_PRIORITY (f1); 1015 priority2 = DECL_INIT_PRIORITY (f2); 1016 1017 if (priority1 < priority2) 1018 return -1; 1019 else if (priority1 > priority2) 1020 return 1; 1021 else 1022 /* Ensure a stable sort. Constructors are executed in backwarding 1023 order to make LTO initialize braries first. */ 1024 return DECL_UID (f2) - DECL_UID (f1); 1025 } 1026 1027 /* Comparison function for qsort. P1 and P2 are actually of type 1028 "tree *" and point to static destructors. DECL_FINI_PRIORITY is 1029 used to determine the sort order. */ 1030 1031 static int 1032 compare_dtor (const void *p1, const void *p2) 1033 { 1034 tree f1; 1035 tree f2; 1036 int priority1; 1037 int priority2; 1038 1039 f1 = *(const tree *)p1; 1040 f2 = *(const tree *)p2; 1041 priority1 = DECL_FINI_PRIORITY (f1); 1042 priority2 = DECL_FINI_PRIORITY (f2); 1043 1044 if (priority1 < priority2) 1045 return -1; 1046 else if (priority1 > priority2) 1047 return 1; 1048 else 1049 /* Ensure a stable sort. */ 1050 return DECL_UID (f1) - DECL_UID (f2); 1051 } 1052 1053 /* Generate functions to call static constructors and destructors 1054 for targets that do not support .ctors/.dtors sections. These 1055 functions have magic names which are detected by collect2. */ 1056 1057 static void 1058 build_cdtor_fns (vec<tree> *ctors, vec<tree> *dtors) 1059 { 1060 if (!ctors->is_empty ()) 1061 { 1062 gcc_assert (!targetm.have_ctors_dtors || in_lto_p); 1063 ctors->qsort (compare_ctor); 1064 build_cdtor (/*ctor_p=*/true, *ctors); 1065 } 1066 1067 if (!dtors->is_empty ()) 1068 { 1069 gcc_assert (!targetm.have_ctors_dtors || in_lto_p); 1070 dtors->qsort (compare_dtor); 1071 build_cdtor (/*ctor_p=*/false, *dtors); 1072 } 1073 } 1074 1075 /* Look for constructors and destructors and produce function calling them. 1076 This is needed for targets not supporting ctors or dtors, but we perform the 1077 transformation also at linktime to merge possibly numerous 1078 constructors/destructors into single function to improve code locality and 1079 reduce size. */ 1080 1081 static unsigned int 1082 ipa_cdtor_merge (void) 1083 { 1084 /* A vector of FUNCTION_DECLs declared as static constructors. */ 1085 auto_vec<tree, 20> ctors; 1086 /* A vector of FUNCTION_DECLs declared as static destructors. */ 1087 auto_vec<tree, 20> dtors; 1088 struct cgraph_node *node; 1089 FOR_EACH_DEFINED_FUNCTION (node) 1090 if (DECL_STATIC_CONSTRUCTOR (node->decl) 1091 || DECL_STATIC_DESTRUCTOR (node->decl)) 1092 record_cdtor_fn (node, &ctors, &dtors); 1093 build_cdtor_fns (&ctors, &dtors); 1094 return 0; 1095 } 1096 1097 namespace { 1098 1099 const pass_data pass_data_ipa_cdtor_merge = 1100 { 1101 IPA_PASS, /* type */ 1102 "cdtor", /* name */ 1103 OPTGROUP_NONE, /* optinfo_flags */ 1104 TV_CGRAPHOPT, /* tv_id */ 1105 0, /* properties_required */ 1106 0, /* properties_provided */ 1107 0, /* properties_destroyed */ 1108 0, /* todo_flags_start */ 1109 0, /* todo_flags_finish */ 1110 }; 1111 1112 class pass_ipa_cdtor_merge : public ipa_opt_pass_d 1113 { 1114 public: 1115 pass_ipa_cdtor_merge (gcc::context *ctxt) 1116 : ipa_opt_pass_d (pass_data_ipa_cdtor_merge, ctxt, 1117 NULL, /* generate_summary */ 1118 NULL, /* write_summary */ 1119 NULL, /* read_summary */ 1120 NULL, /* write_optimization_summary */ 1121 NULL, /* read_optimization_summary */ 1122 NULL, /* stmt_fixup */ 1123 0, /* function_transform_todo_flags_start */ 1124 NULL, /* function_transform */ 1125 NULL) /* variable_transform */ 1126 {} 1127 1128 /* opt_pass methods: */ 1129 virtual bool gate (function *); 1130 virtual unsigned int execute (function *) { return ipa_cdtor_merge (); } 1131 1132 }; // class pass_ipa_cdtor_merge 1133 1134 bool 1135 pass_ipa_cdtor_merge::gate (function *) 1136 { 1137 /* Perform the pass when we have no ctors/dtors support 1138 or at LTO time to merge multiple constructors into single 1139 function. */ 1140 return !targetm.have_ctors_dtors || in_lto_p; 1141 } 1142 1143 } // anon namespace 1144 1145 ipa_opt_pass_d * 1146 make_pass_ipa_cdtor_merge (gcc::context *ctxt) 1147 { 1148 return new pass_ipa_cdtor_merge (ctxt); 1149 } 1150 1151 /* Invalid pointer representing BOTTOM for single user dataflow. */ 1152 #define BOTTOM ((cgraph_node *)(size_t) 2) 1153 1154 /* Meet operation for single user dataflow. 1155 Here we want to associate variables with sigle function that may access it. 1156 1157 FUNCTION is current single user of a variable, VAR is variable that uses it. 1158 Latttice is stored in SINGLE_USER_MAP. 1159 1160 We represent: 1161 - TOP by no entry in SIGNLE_USER_MAP 1162 - BOTTOM by BOTTOM in AUX pointer (to save lookups) 1163 - known single user by cgraph pointer in SINGLE_USER_MAP. */ 1164 1165 cgraph_node * 1166 meet (cgraph_node *function, varpool_node *var, 1167 hash_map<varpool_node *, cgraph_node *> &single_user_map) 1168 { 1169 struct cgraph_node *user, **f; 1170 1171 if (var->aux == BOTTOM) 1172 return BOTTOM; 1173 1174 f = single_user_map.get (var); 1175 if (!f) 1176 return function; 1177 user = *f; 1178 if (!function) 1179 return user; 1180 else if (function != user) 1181 return BOTTOM; 1182 else 1183 return function; 1184 } 1185 1186 /* Propagation step of single-use dataflow. 1187 1188 Check all uses of VNODE and see if they are used by single function FUNCTION. 1189 SINGLE_USER_MAP represents the dataflow lattice. */ 1190 1191 cgraph_node * 1192 propagate_single_user (varpool_node *vnode, cgraph_node *function, 1193 hash_map<varpool_node *, cgraph_node *> &single_user_map) 1194 { 1195 int i; 1196 struct ipa_ref *ref; 1197 1198 gcc_assert (!vnode->externally_visible); 1199 1200 /* If node is an alias, first meet with its target. */ 1201 if (vnode->alias) 1202 function = meet (function, vnode->get_alias_target (), single_user_map); 1203 1204 /* Check all users and see if they correspond to a single function. */ 1205 for (i = 0; vnode->iterate_referring (i, ref) && function != BOTTOM; i++) 1206 { 1207 struct cgraph_node *cnode = dyn_cast <cgraph_node *> (ref->referring); 1208 if (cnode) 1209 { 1210 if (cnode->global.inlined_to) 1211 cnode = cnode->global.inlined_to; 1212 if (!function) 1213 function = cnode; 1214 else if (function != cnode) 1215 function = BOTTOM; 1216 } 1217 else 1218 function = meet (function, dyn_cast <varpool_node *> (ref->referring), 1219 single_user_map); 1220 } 1221 return function; 1222 } 1223 1224 /* Pass setting used_by_single_function flag. 1225 This flag is set on variable when there is only one function that may 1226 possibly referr to it. */ 1227 1228 static unsigned int 1229 ipa_single_use (void) 1230 { 1231 varpool_node *first = (varpool_node *) (void *) 1; 1232 varpool_node *var; 1233 hash_map<varpool_node *, cgraph_node *> single_user_map; 1234 1235 FOR_EACH_DEFINED_VARIABLE (var) 1236 if (!var->all_refs_explicit_p ()) 1237 var->aux = BOTTOM; 1238 else 1239 { 1240 /* Enqueue symbol for dataflow. */ 1241 var->aux = first; 1242 first = var; 1243 } 1244 1245 /* The actual dataflow. */ 1246 1247 while (first != (void *) 1) 1248 { 1249 cgraph_node *user, *orig_user, **f; 1250 1251 var = first; 1252 first = (varpool_node *)first->aux; 1253 1254 f = single_user_map.get (var); 1255 if (f) 1256 orig_user = *f; 1257 else 1258 orig_user = NULL; 1259 user = propagate_single_user (var, orig_user, single_user_map); 1260 1261 gcc_checking_assert (var->aux != BOTTOM); 1262 1263 /* If user differs, enqueue all references. */ 1264 if (user != orig_user) 1265 { 1266 unsigned int i; 1267 ipa_ref *ref; 1268 1269 single_user_map.put (var, user); 1270 1271 /* Enqueue all aliases for re-processing. */ 1272 for (i = 0; var->iterate_direct_aliases (i, ref); i++) 1273 if (!ref->referring->aux) 1274 { 1275 ref->referring->aux = first; 1276 first = dyn_cast <varpool_node *> (ref->referring); 1277 } 1278 /* Enqueue all users for re-processing. */ 1279 for (i = 0; var->iterate_reference (i, ref); i++) 1280 if (!ref->referred->aux 1281 && ref->referred->definition 1282 && is_a <varpool_node *> (ref->referred)) 1283 { 1284 ref->referred->aux = first; 1285 first = dyn_cast <varpool_node *> (ref->referred); 1286 } 1287 1288 /* If user is BOTTOM, just punt on this var. */ 1289 if (user == BOTTOM) 1290 var->aux = BOTTOM; 1291 else 1292 var->aux = NULL; 1293 } 1294 else 1295 var->aux = NULL; 1296 } 1297 1298 FOR_EACH_DEFINED_VARIABLE (var) 1299 { 1300 if (var->aux != BOTTOM) 1301 { 1302 /* Not having the single user known means that the VAR is 1303 unreachable. Either someone forgot to remove unreachable 1304 variables or the reachability here is wrong. */ 1305 1306 gcc_checking_assert (single_user_map.get (var)); 1307 1308 if (dump_file) 1309 { 1310 fprintf (dump_file, "Variable %s is used by single function\n", 1311 var->dump_name ()); 1312 } 1313 var->used_by_single_function = true; 1314 } 1315 var->aux = NULL; 1316 } 1317 return 0; 1318 } 1319 1320 namespace { 1321 1322 const pass_data pass_data_ipa_single_use = 1323 { 1324 IPA_PASS, /* type */ 1325 "single-use", /* name */ 1326 OPTGROUP_NONE, /* optinfo_flags */ 1327 TV_CGRAPHOPT, /* tv_id */ 1328 0, /* properties_required */ 1329 0, /* properties_provided */ 1330 0, /* properties_destroyed */ 1331 0, /* todo_flags_start */ 1332 0, /* todo_flags_finish */ 1333 }; 1334 1335 class pass_ipa_single_use : public ipa_opt_pass_d 1336 { 1337 public: 1338 pass_ipa_single_use (gcc::context *ctxt) 1339 : ipa_opt_pass_d (pass_data_ipa_single_use, ctxt, 1340 NULL, /* generate_summary */ 1341 NULL, /* write_summary */ 1342 NULL, /* read_summary */ 1343 NULL, /* write_optimization_summary */ 1344 NULL, /* read_optimization_summary */ 1345 NULL, /* stmt_fixup */ 1346 0, /* function_transform_todo_flags_start */ 1347 NULL, /* function_transform */ 1348 NULL) /* variable_transform */ 1349 {} 1350 1351 /* opt_pass methods: */ 1352 virtual unsigned int execute (function *) { return ipa_single_use (); } 1353 1354 }; // class pass_ipa_single_use 1355 1356 } // anon namespace 1357 1358 ipa_opt_pass_d * 1359 make_pass_ipa_single_use (gcc::context *ctxt) 1360 { 1361 return new pass_ipa_single_use (ctxt); 1362 } 1363 1364 /* Materialize all clones. */ 1365 1366 namespace { 1367 1368 const pass_data pass_data_materialize_all_clones = 1369 { 1370 SIMPLE_IPA_PASS, /* type */ 1371 "materialize-all-clones", /* name */ 1372 OPTGROUP_NONE, /* optinfo_flags */ 1373 TV_IPA_OPT, /* tv_id */ 1374 0, /* properties_required */ 1375 0, /* properties_provided */ 1376 0, /* properties_destroyed */ 1377 0, /* todo_flags_start */ 1378 0, /* todo_flags_finish */ 1379 }; 1380 1381 class pass_materialize_all_clones : public simple_ipa_opt_pass 1382 { 1383 public: 1384 pass_materialize_all_clones (gcc::context *ctxt) 1385 : simple_ipa_opt_pass (pass_data_materialize_all_clones, ctxt) 1386 {} 1387 1388 /* opt_pass methods: */ 1389 virtual unsigned int execute (function *) 1390 { 1391 symtab->materialize_all_clones (); 1392 return 0; 1393 } 1394 1395 }; // class pass_materialize_all_clones 1396 1397 } // anon namespace 1398 1399 simple_ipa_opt_pass * 1400 make_pass_materialize_all_clones (gcc::context *ctxt) 1401 { 1402 return new pass_materialize_all_clones (ctxt); 1403 } 1404