[Ada] Fix internal error on iterated array aggregate
[official-gcc.git] / gcc / tree-ssa-sccvn.cc
blobed68557f0b23b5532544a2755a1261786f3ea32e
1 /* SCC value numbering for trees
2 Copyright (C) 2006-2022 Free Software Foundation, Inc.
3 Contributed by Daniel Berlin <dan@dberlin.org>
5 This file is part of GCC.
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.
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.
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/>. */
21 #include "config.h"
22 #include "system.h"
23 #include "coretypes.h"
24 #include "splay-tree.h"
25 #include "backend.h"
26 #include "rtl.h"
27 #include "tree.h"
28 #include "gimple.h"
29 #include "ssa.h"
30 #include "expmed.h"
31 #include "insn-config.h"
32 #include "memmodel.h"
33 #include "emit-rtl.h"
34 #include "cgraph.h"
35 #include "gimple-pretty-print.h"
36 #include "alias.h"
37 #include "fold-const.h"
38 #include "stor-layout.h"
39 #include "cfganal.h"
40 #include "tree-inline.h"
41 #include "internal-fn.h"
42 #include "gimple-iterator.h"
43 #include "gimple-fold.h"
44 #include "tree-eh.h"
45 #include "gimplify.h"
46 #include "flags.h"
47 #include "dojump.h"
48 #include "explow.h"
49 #include "calls.h"
50 #include "varasm.h"
51 #include "stmt.h"
52 #include "expr.h"
53 #include "tree-dfa.h"
54 #include "tree-ssa.h"
55 #include "dumpfile.h"
56 #include "cfgloop.h"
57 #include "tree-ssa-propagate.h"
58 #include "tree-cfg.h"
59 #include "domwalk.h"
60 #include "gimple-match.h"
61 #include "stringpool.h"
62 #include "attribs.h"
63 #include "tree-pass.h"
64 #include "statistics.h"
65 #include "langhooks.h"
66 #include "ipa-utils.h"
67 #include "dbgcnt.h"
68 #include "tree-cfgcleanup.h"
69 #include "tree-ssa-loop.h"
70 #include "tree-scalar-evolution.h"
71 #include "tree-ssa-loop-niter.h"
72 #include "builtins.h"
73 #include "fold-const-call.h"
74 #include "ipa-modref-tree.h"
75 #include "ipa-modref.h"
76 #include "tree-ssa-sccvn.h"
78 /* This algorithm is based on the SCC algorithm presented by Keith
79 Cooper and L. Taylor Simpson in "SCC-Based Value numbering"
80 (http://citeseer.ist.psu.edu/41805.html). In
81 straight line code, it is equivalent to a regular hash based value
82 numbering that is performed in reverse postorder.
84 For code with cycles, there are two alternatives, both of which
85 require keeping the hashtables separate from the actual list of
86 value numbers for SSA names.
88 1. Iterate value numbering in an RPO walk of the blocks, removing
89 all the entries from the hashtable after each iteration (but
90 keeping the SSA name->value number mapping between iterations).
91 Iterate until it does not change.
93 2. Perform value numbering as part of an SCC walk on the SSA graph,
94 iterating only the cycles in the SSA graph until they do not change
95 (using a separate, optimistic hashtable for value numbering the SCC
96 operands).
98 The second is not just faster in practice (because most SSA graph
99 cycles do not involve all the variables in the graph), it also has
100 some nice properties.
102 One of these nice properties is that when we pop an SCC off the
103 stack, we are guaranteed to have processed all the operands coming from
104 *outside of that SCC*, so we do not need to do anything special to
105 ensure they have value numbers.
107 Another nice property is that the SCC walk is done as part of a DFS
108 of the SSA graph, which makes it easy to perform combining and
109 simplifying operations at the same time.
111 The code below is deliberately written in a way that makes it easy
112 to separate the SCC walk from the other work it does.
114 In order to propagate constants through the code, we track which
115 expressions contain constants, and use those while folding. In
116 theory, we could also track expressions whose value numbers are
117 replaced, in case we end up folding based on expression
118 identities.
120 In order to value number memory, we assign value numbers to vuses.
121 This enables us to note that, for example, stores to the same
122 address of the same value from the same starting memory states are
123 equivalent.
124 TODO:
126 1. We can iterate only the changing portions of the SCC's, but
127 I have not seen an SCC big enough for this to be a win.
128 2. If you differentiate between phi nodes for loops and phi nodes
129 for if-then-else, you can properly consider phi nodes in different
130 blocks for equivalence.
131 3. We could value number vuses in more cases, particularly, whole
132 structure copies.
135 /* There's no BB_EXECUTABLE but we can use BB_VISITED. */
136 #define BB_EXECUTABLE BB_VISITED
138 static vn_lookup_kind default_vn_walk_kind;
140 /* vn_nary_op hashtable helpers. */
142 struct vn_nary_op_hasher : nofree_ptr_hash <vn_nary_op_s>
144 typedef vn_nary_op_s *compare_type;
145 static inline hashval_t hash (const vn_nary_op_s *);
146 static inline bool equal (const vn_nary_op_s *, const vn_nary_op_s *);
149 /* Return the computed hashcode for nary operation P1. */
151 inline hashval_t
152 vn_nary_op_hasher::hash (const vn_nary_op_s *vno1)
154 return vno1->hashcode;
157 /* Compare nary operations P1 and P2 and return true if they are
158 equivalent. */
160 inline bool
161 vn_nary_op_hasher::equal (const vn_nary_op_s *vno1, const vn_nary_op_s *vno2)
163 return vno1 == vno2 || vn_nary_op_eq (vno1, vno2);
166 typedef hash_table<vn_nary_op_hasher> vn_nary_op_table_type;
167 typedef vn_nary_op_table_type::iterator vn_nary_op_iterator_type;
170 /* vn_phi hashtable helpers. */
172 static int
173 vn_phi_eq (const_vn_phi_t const vp1, const_vn_phi_t const vp2);
175 struct vn_phi_hasher : nofree_ptr_hash <vn_phi_s>
177 static inline hashval_t hash (const vn_phi_s *);
178 static inline bool equal (const vn_phi_s *, const vn_phi_s *);
181 /* Return the computed hashcode for phi operation P1. */
183 inline hashval_t
184 vn_phi_hasher::hash (const vn_phi_s *vp1)
186 return vp1->hashcode;
189 /* Compare two phi entries for equality, ignoring VN_TOP arguments. */
191 inline bool
192 vn_phi_hasher::equal (const vn_phi_s *vp1, const vn_phi_s *vp2)
194 return vp1 == vp2 || vn_phi_eq (vp1, vp2);
197 typedef hash_table<vn_phi_hasher> vn_phi_table_type;
198 typedef vn_phi_table_type::iterator vn_phi_iterator_type;
201 /* Compare two reference operands P1 and P2 for equality. Return true if
202 they are equal, and false otherwise. */
204 static int
205 vn_reference_op_eq (const void *p1, const void *p2)
207 const_vn_reference_op_t const vro1 = (const_vn_reference_op_t) p1;
208 const_vn_reference_op_t const vro2 = (const_vn_reference_op_t) p2;
210 return (vro1->opcode == vro2->opcode
211 /* We do not care for differences in type qualification. */
212 && (vro1->type == vro2->type
213 || (vro1->type && vro2->type
214 && types_compatible_p (TYPE_MAIN_VARIANT (vro1->type),
215 TYPE_MAIN_VARIANT (vro2->type))))
216 && expressions_equal_p (vro1->op0, vro2->op0)
217 && expressions_equal_p (vro1->op1, vro2->op1)
218 && expressions_equal_p (vro1->op2, vro2->op2)
219 && (vro1->opcode != CALL_EXPR || vro1->clique == vro2->clique));
222 /* Free a reference operation structure VP. */
224 static inline void
225 free_reference (vn_reference_s *vr)
227 vr->operands.release ();
231 /* vn_reference hashtable helpers. */
233 struct vn_reference_hasher : nofree_ptr_hash <vn_reference_s>
235 static inline hashval_t hash (const vn_reference_s *);
236 static inline bool equal (const vn_reference_s *, const vn_reference_s *);
239 /* Return the hashcode for a given reference operation P1. */
241 inline hashval_t
242 vn_reference_hasher::hash (const vn_reference_s *vr1)
244 return vr1->hashcode;
247 inline bool
248 vn_reference_hasher::equal (const vn_reference_s *v, const vn_reference_s *c)
250 return v == c || vn_reference_eq (v, c);
253 typedef hash_table<vn_reference_hasher> vn_reference_table_type;
254 typedef vn_reference_table_type::iterator vn_reference_iterator_type;
256 /* Pretty-print OPS to OUTFILE. */
258 void
259 print_vn_reference_ops (FILE *outfile, const vec<vn_reference_op_s> ops)
261 vn_reference_op_t vro;
262 unsigned int i;
263 fprintf (outfile, "{");
264 for (i = 0; ops.iterate (i, &vro); i++)
266 bool closebrace = false;
267 if (vro->opcode != SSA_NAME
268 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
270 fprintf (outfile, "%s", get_tree_code_name (vro->opcode));
271 if (vro->op0 || vro->opcode == CALL_EXPR)
273 fprintf (outfile, "<");
274 closebrace = true;
277 if (vro->op0 || vro->opcode == CALL_EXPR)
279 if (!vro->op0)
280 fprintf (outfile, internal_fn_name ((internal_fn)vro->clique));
281 else
282 print_generic_expr (outfile, vro->op0);
283 if (vro->op1)
285 fprintf (outfile, ",");
286 print_generic_expr (outfile, vro->op1);
288 if (vro->op2)
290 fprintf (outfile, ",");
291 print_generic_expr (outfile, vro->op2);
294 if (closebrace)
295 fprintf (outfile, ">");
296 if (i != ops.length () - 1)
297 fprintf (outfile, ",");
299 fprintf (outfile, "}");
302 DEBUG_FUNCTION void
303 debug_vn_reference_ops (const vec<vn_reference_op_s> ops)
305 print_vn_reference_ops (stderr, ops);
306 fputc ('\n', stderr);
309 /* The set of VN hashtables. */
311 typedef struct vn_tables_s
313 vn_nary_op_table_type *nary;
314 vn_phi_table_type *phis;
315 vn_reference_table_type *references;
316 } *vn_tables_t;
319 /* vn_constant hashtable helpers. */
321 struct vn_constant_hasher : free_ptr_hash <vn_constant_s>
323 static inline hashval_t hash (const vn_constant_s *);
324 static inline bool equal (const vn_constant_s *, const vn_constant_s *);
327 /* Hash table hash function for vn_constant_t. */
329 inline hashval_t
330 vn_constant_hasher::hash (const vn_constant_s *vc1)
332 return vc1->hashcode;
335 /* Hash table equality function for vn_constant_t. */
337 inline bool
338 vn_constant_hasher::equal (const vn_constant_s *vc1, const vn_constant_s *vc2)
340 if (vc1->hashcode != vc2->hashcode)
341 return false;
343 return vn_constant_eq_with_type (vc1->constant, vc2->constant);
346 static hash_table<vn_constant_hasher> *constant_to_value_id;
349 /* Obstack we allocate the vn-tables elements from. */
350 static obstack vn_tables_obstack;
351 /* Special obstack we never unwind. */
352 static obstack vn_tables_insert_obstack;
354 static vn_reference_t last_inserted_ref;
355 static vn_phi_t last_inserted_phi;
356 static vn_nary_op_t last_inserted_nary;
357 static vn_ssa_aux_t last_pushed_avail;
359 /* Valid hashtables storing information we have proven to be
360 correct. */
361 static vn_tables_t valid_info;
364 /* Valueization hook for simplify_replace_tree. Valueize NAME if it is
365 an SSA name, otherwise just return it. */
366 tree (*vn_valueize) (tree);
367 static tree
368 vn_valueize_for_srt (tree t, void* context ATTRIBUTE_UNUSED)
370 basic_block saved_vn_context_bb = vn_context_bb;
371 /* Look for sth available at the definition block of the argument.
372 This avoids inconsistencies between availability there which
373 decides if the stmt can be removed and availability at the
374 use site. The SSA property ensures that things available
375 at the definition are also available at uses. */
376 if (!SSA_NAME_IS_DEFAULT_DEF (t))
377 vn_context_bb = gimple_bb (SSA_NAME_DEF_STMT (t));
378 tree res = vn_valueize (t);
379 vn_context_bb = saved_vn_context_bb;
380 return res;
384 /* This represents the top of the VN lattice, which is the universal
385 value. */
387 tree VN_TOP;
389 /* Unique counter for our value ids. */
391 static unsigned int next_value_id;
392 static int next_constant_value_id;
395 /* Table of vn_ssa_aux_t's, one per ssa_name. The vn_ssa_aux_t objects
396 are allocated on an obstack for locality reasons, and to free them
397 without looping over the vec. */
399 struct vn_ssa_aux_hasher : typed_noop_remove <vn_ssa_aux_t>
401 typedef vn_ssa_aux_t value_type;
402 typedef tree compare_type;
403 static inline hashval_t hash (const value_type &);
404 static inline bool equal (const value_type &, const compare_type &);
405 static inline void mark_deleted (value_type &) {}
406 static const bool empty_zero_p = true;
407 static inline void mark_empty (value_type &e) { e = NULL; }
408 static inline bool is_deleted (value_type &) { return false; }
409 static inline bool is_empty (value_type &e) { return e == NULL; }
412 hashval_t
413 vn_ssa_aux_hasher::hash (const value_type &entry)
415 return SSA_NAME_VERSION (entry->name);
418 bool
419 vn_ssa_aux_hasher::equal (const value_type &entry, const compare_type &name)
421 return name == entry->name;
424 static hash_table<vn_ssa_aux_hasher> *vn_ssa_aux_hash;
425 typedef hash_table<vn_ssa_aux_hasher>::iterator vn_ssa_aux_iterator_type;
426 static struct obstack vn_ssa_aux_obstack;
428 static vn_nary_op_t vn_nary_op_insert_stmt (gimple *, tree);
429 static vn_nary_op_t vn_nary_op_insert_into (vn_nary_op_t,
430 vn_nary_op_table_type *);
431 static void init_vn_nary_op_from_pieces (vn_nary_op_t, unsigned int,
432 enum tree_code, tree, tree *);
433 static tree vn_lookup_simplify_result (gimple_match_op *);
434 static vn_reference_t vn_reference_lookup_or_insert_for_pieces
435 (tree, alias_set_type, alias_set_type, tree,
436 vec<vn_reference_op_s, va_heap>, tree);
438 /* Return whether there is value numbering information for a given SSA name. */
440 bool
441 has_VN_INFO (tree name)
443 return vn_ssa_aux_hash->find_with_hash (name, SSA_NAME_VERSION (name));
446 vn_ssa_aux_t
447 VN_INFO (tree name)
449 vn_ssa_aux_t *res
450 = vn_ssa_aux_hash->find_slot_with_hash (name, SSA_NAME_VERSION (name),
451 INSERT);
452 if (*res != NULL)
453 return *res;
455 vn_ssa_aux_t newinfo = *res = XOBNEW (&vn_ssa_aux_obstack, struct vn_ssa_aux);
456 memset (newinfo, 0, sizeof (struct vn_ssa_aux));
457 newinfo->name = name;
458 newinfo->valnum = VN_TOP;
459 /* We are using the visited flag to handle uses with defs not within the
460 region being value-numbered. */
461 newinfo->visited = false;
463 /* Given we create the VN_INFOs on-demand now we have to do initialization
464 different than VN_TOP here. */
465 if (SSA_NAME_IS_DEFAULT_DEF (name))
466 switch (TREE_CODE (SSA_NAME_VAR (name)))
468 case VAR_DECL:
469 /* All undefined vars are VARYING. */
470 newinfo->valnum = name;
471 newinfo->visited = true;
472 break;
474 case PARM_DECL:
475 /* Parameters are VARYING but we can record a condition
476 if we know it is a non-NULL pointer. */
477 newinfo->visited = true;
478 newinfo->valnum = name;
479 if (POINTER_TYPE_P (TREE_TYPE (name))
480 && nonnull_arg_p (SSA_NAME_VAR (name)))
482 tree ops[2];
483 ops[0] = name;
484 ops[1] = build_int_cst (TREE_TYPE (name), 0);
485 vn_nary_op_t nary;
486 /* Allocate from non-unwinding stack. */
487 nary = alloc_vn_nary_op_noinit (2, &vn_tables_insert_obstack);
488 init_vn_nary_op_from_pieces (nary, 2, NE_EXPR,
489 boolean_type_node, ops);
490 nary->predicated_values = 0;
491 nary->u.result = boolean_true_node;
492 vn_nary_op_insert_into (nary, valid_info->nary);
493 gcc_assert (nary->unwind_to == NULL);
494 /* Also do not link it into the undo chain. */
495 last_inserted_nary = nary->next;
496 nary->next = (vn_nary_op_t)(void *)-1;
497 nary = alloc_vn_nary_op_noinit (2, &vn_tables_insert_obstack);
498 init_vn_nary_op_from_pieces (nary, 2, EQ_EXPR,
499 boolean_type_node, ops);
500 nary->predicated_values = 0;
501 nary->u.result = boolean_false_node;
502 vn_nary_op_insert_into (nary, valid_info->nary);
503 gcc_assert (nary->unwind_to == NULL);
504 last_inserted_nary = nary->next;
505 nary->next = (vn_nary_op_t)(void *)-1;
506 if (dump_file && (dump_flags & TDF_DETAILS))
508 fprintf (dump_file, "Recording ");
509 print_generic_expr (dump_file, name, TDF_SLIM);
510 fprintf (dump_file, " != 0\n");
513 break;
515 case RESULT_DECL:
516 /* If the result is passed by invisible reference the default
517 def is initialized, otherwise it's uninitialized. Still
518 undefined is varying. */
519 newinfo->visited = true;
520 newinfo->valnum = name;
521 break;
523 default:
524 gcc_unreachable ();
526 return newinfo;
529 /* Return the SSA value of X. */
531 inline tree
532 SSA_VAL (tree x, bool *visited = NULL)
534 vn_ssa_aux_t tem = vn_ssa_aux_hash->find_with_hash (x, SSA_NAME_VERSION (x));
535 if (visited)
536 *visited = tem && tem->visited;
537 return tem && tem->visited ? tem->valnum : x;
540 /* Return the SSA value of the VUSE x, supporting released VDEFs
541 during elimination which will value-number the VDEF to the
542 associated VUSE (but not substitute in the whole lattice). */
544 static inline tree
545 vuse_ssa_val (tree x)
547 if (!x)
548 return NULL_TREE;
552 x = SSA_VAL (x);
553 gcc_assert (x != VN_TOP);
555 while (SSA_NAME_IN_FREE_LIST (x));
557 return x;
560 /* Similar to the above but used as callback for walk_non_aliased_vuses
561 and thus should stop at unvisited VUSE to not walk across region
562 boundaries. */
564 static tree
565 vuse_valueize (tree vuse)
569 bool visited;
570 vuse = SSA_VAL (vuse, &visited);
571 if (!visited)
572 return NULL_TREE;
573 gcc_assert (vuse != VN_TOP);
575 while (SSA_NAME_IN_FREE_LIST (vuse));
576 return vuse;
580 /* Return the vn_kind the expression computed by the stmt should be
581 associated with. */
583 enum vn_kind
584 vn_get_stmt_kind (gimple *stmt)
586 switch (gimple_code (stmt))
588 case GIMPLE_CALL:
589 return VN_REFERENCE;
590 case GIMPLE_PHI:
591 return VN_PHI;
592 case GIMPLE_ASSIGN:
594 enum tree_code code = gimple_assign_rhs_code (stmt);
595 tree rhs1 = gimple_assign_rhs1 (stmt);
596 switch (get_gimple_rhs_class (code))
598 case GIMPLE_UNARY_RHS:
599 case GIMPLE_BINARY_RHS:
600 case GIMPLE_TERNARY_RHS:
601 return VN_NARY;
602 case GIMPLE_SINGLE_RHS:
603 switch (TREE_CODE_CLASS (code))
605 case tcc_reference:
606 /* VOP-less references can go through unary case. */
607 if ((code == REALPART_EXPR
608 || code == IMAGPART_EXPR
609 || code == VIEW_CONVERT_EXPR
610 || code == BIT_FIELD_REF)
611 && (TREE_CODE (TREE_OPERAND (rhs1, 0)) == SSA_NAME
612 || is_gimple_min_invariant (TREE_OPERAND (rhs1, 0))))
613 return VN_NARY;
615 /* Fallthrough. */
616 case tcc_declaration:
617 return VN_REFERENCE;
619 case tcc_constant:
620 return VN_CONSTANT;
622 default:
623 if (code == ADDR_EXPR)
624 return (is_gimple_min_invariant (rhs1)
625 ? VN_CONSTANT : VN_REFERENCE);
626 else if (code == CONSTRUCTOR)
627 return VN_NARY;
628 return VN_NONE;
630 default:
631 return VN_NONE;
634 default:
635 return VN_NONE;
639 /* Lookup a value id for CONSTANT and return it. If it does not
640 exist returns 0. */
642 unsigned int
643 get_constant_value_id (tree constant)
645 vn_constant_s **slot;
646 struct vn_constant_s vc;
648 vc.hashcode = vn_hash_constant_with_type (constant);
649 vc.constant = constant;
650 slot = constant_to_value_id->find_slot (&vc, NO_INSERT);
651 if (slot)
652 return (*slot)->value_id;
653 return 0;
656 /* Lookup a value id for CONSTANT, and if it does not exist, create a
657 new one and return it. If it does exist, return it. */
659 unsigned int
660 get_or_alloc_constant_value_id (tree constant)
662 vn_constant_s **slot;
663 struct vn_constant_s vc;
664 vn_constant_t vcp;
666 /* If the hashtable isn't initialized we're not running from PRE and thus
667 do not need value-ids. */
668 if (!constant_to_value_id)
669 return 0;
671 vc.hashcode = vn_hash_constant_with_type (constant);
672 vc.constant = constant;
673 slot = constant_to_value_id->find_slot (&vc, INSERT);
674 if (*slot)
675 return (*slot)->value_id;
677 vcp = XNEW (struct vn_constant_s);
678 vcp->hashcode = vc.hashcode;
679 vcp->constant = constant;
680 vcp->value_id = get_next_constant_value_id ();
681 *slot = vcp;
682 return vcp->value_id;
685 /* Compute the hash for a reference operand VRO1. */
687 static void
688 vn_reference_op_compute_hash (const vn_reference_op_t vro1, inchash::hash &hstate)
690 hstate.add_int (vro1->opcode);
691 if (vro1->opcode == CALL_EXPR && !vro1->op0)
692 hstate.add_int (vro1->clique);
693 if (vro1->op0)
694 inchash::add_expr (vro1->op0, hstate);
695 if (vro1->op1)
696 inchash::add_expr (vro1->op1, hstate);
697 if (vro1->op2)
698 inchash::add_expr (vro1->op2, hstate);
701 /* Compute a hash for the reference operation VR1 and return it. */
703 static hashval_t
704 vn_reference_compute_hash (const vn_reference_t vr1)
706 inchash::hash hstate;
707 hashval_t result;
708 int i;
709 vn_reference_op_t vro;
710 poly_int64 off = -1;
711 bool deref = false;
713 FOR_EACH_VEC_ELT (vr1->operands, i, vro)
715 if (vro->opcode == MEM_REF)
716 deref = true;
717 else if (vro->opcode != ADDR_EXPR)
718 deref = false;
719 if (maybe_ne (vro->off, -1))
721 if (known_eq (off, -1))
722 off = 0;
723 off += vro->off;
725 else
727 if (maybe_ne (off, -1)
728 && maybe_ne (off, 0))
729 hstate.add_poly_int (off);
730 off = -1;
731 if (deref
732 && vro->opcode == ADDR_EXPR)
734 if (vro->op0)
736 tree op = TREE_OPERAND (vro->op0, 0);
737 hstate.add_int (TREE_CODE (op));
738 inchash::add_expr (op, hstate);
741 else
742 vn_reference_op_compute_hash (vro, hstate);
745 result = hstate.end ();
746 /* ??? We would ICE later if we hash instead of adding that in. */
747 if (vr1->vuse)
748 result += SSA_NAME_VERSION (vr1->vuse);
750 return result;
753 /* Return true if reference operations VR1 and VR2 are equivalent. This
754 means they have the same set of operands and vuses. */
756 bool
757 vn_reference_eq (const_vn_reference_t const vr1, const_vn_reference_t const vr2)
759 unsigned i, j;
761 /* Early out if this is not a hash collision. */
762 if (vr1->hashcode != vr2->hashcode)
763 return false;
765 /* The VOP needs to be the same. */
766 if (vr1->vuse != vr2->vuse)
767 return false;
769 /* If the operands are the same we are done. */
770 if (vr1->operands == vr2->operands)
771 return true;
773 if (!vr1->type || !vr2->type)
775 if (vr1->type != vr2->type)
776 return false;
778 else if (vr1->type == vr2->type)
780 else if (COMPLETE_TYPE_P (vr1->type) != COMPLETE_TYPE_P (vr2->type)
781 || (COMPLETE_TYPE_P (vr1->type)
782 && !expressions_equal_p (TYPE_SIZE (vr1->type),
783 TYPE_SIZE (vr2->type))))
784 return false;
785 else if (vr1->operands[0].opcode == CALL_EXPR
786 && !types_compatible_p (vr1->type, vr2->type))
787 return false;
788 else if (INTEGRAL_TYPE_P (vr1->type)
789 && INTEGRAL_TYPE_P (vr2->type))
791 if (TYPE_PRECISION (vr1->type) != TYPE_PRECISION (vr2->type))
792 return false;
794 else if (INTEGRAL_TYPE_P (vr1->type)
795 && (TYPE_PRECISION (vr1->type)
796 != TREE_INT_CST_LOW (TYPE_SIZE (vr1->type))))
797 return false;
798 else if (INTEGRAL_TYPE_P (vr2->type)
799 && (TYPE_PRECISION (vr2->type)
800 != TREE_INT_CST_LOW (TYPE_SIZE (vr2->type))))
801 return false;
803 i = 0;
804 j = 0;
807 poly_int64 off1 = 0, off2 = 0;
808 vn_reference_op_t vro1, vro2;
809 vn_reference_op_s tem1, tem2;
810 bool deref1 = false, deref2 = false;
811 bool reverse1 = false, reverse2 = false;
812 for (; vr1->operands.iterate (i, &vro1); i++)
814 if (vro1->opcode == MEM_REF)
815 deref1 = true;
816 /* Do not look through a storage order barrier. */
817 else if (vro1->opcode == VIEW_CONVERT_EXPR && vro1->reverse)
818 return false;
819 reverse1 |= vro1->reverse;
820 if (known_eq (vro1->off, -1))
821 break;
822 off1 += vro1->off;
824 for (; vr2->operands.iterate (j, &vro2); j++)
826 if (vro2->opcode == MEM_REF)
827 deref2 = true;
828 /* Do not look through a storage order barrier. */
829 else if (vro2->opcode == VIEW_CONVERT_EXPR && vro2->reverse)
830 return false;
831 reverse2 |= vro2->reverse;
832 if (known_eq (vro2->off, -1))
833 break;
834 off2 += vro2->off;
836 if (maybe_ne (off1, off2) || reverse1 != reverse2)
837 return false;
838 if (deref1 && vro1->opcode == ADDR_EXPR)
840 memset (&tem1, 0, sizeof (tem1));
841 tem1.op0 = TREE_OPERAND (vro1->op0, 0);
842 tem1.type = TREE_TYPE (tem1.op0);
843 tem1.opcode = TREE_CODE (tem1.op0);
844 vro1 = &tem1;
845 deref1 = false;
847 if (deref2 && vro2->opcode == ADDR_EXPR)
849 memset (&tem2, 0, sizeof (tem2));
850 tem2.op0 = TREE_OPERAND (vro2->op0, 0);
851 tem2.type = TREE_TYPE (tem2.op0);
852 tem2.opcode = TREE_CODE (tem2.op0);
853 vro2 = &tem2;
854 deref2 = false;
856 if (deref1 != deref2)
857 return false;
858 if (!vn_reference_op_eq (vro1, vro2))
859 return false;
860 ++j;
861 ++i;
863 while (vr1->operands.length () != i
864 || vr2->operands.length () != j);
866 return true;
869 /* Copy the operations present in load/store REF into RESULT, a vector of
870 vn_reference_op_s's. */
872 static void
873 copy_reference_ops_from_ref (tree ref, vec<vn_reference_op_s> *result)
875 /* For non-calls, store the information that makes up the address. */
876 tree orig = ref;
877 while (ref)
879 vn_reference_op_s temp;
881 memset (&temp, 0, sizeof (temp));
882 temp.type = TREE_TYPE (ref);
883 temp.opcode = TREE_CODE (ref);
884 temp.off = -1;
886 switch (temp.opcode)
888 case MODIFY_EXPR:
889 temp.op0 = TREE_OPERAND (ref, 1);
890 break;
891 case WITH_SIZE_EXPR:
892 temp.op0 = TREE_OPERAND (ref, 1);
893 temp.off = 0;
894 break;
895 case MEM_REF:
896 /* The base address gets its own vn_reference_op_s structure. */
897 temp.op0 = TREE_OPERAND (ref, 1);
898 if (!mem_ref_offset (ref).to_shwi (&temp.off))
899 temp.off = -1;
900 temp.clique = MR_DEPENDENCE_CLIQUE (ref);
901 temp.base = MR_DEPENDENCE_BASE (ref);
902 temp.reverse = REF_REVERSE_STORAGE_ORDER (ref);
903 break;
904 case TARGET_MEM_REF:
905 /* The base address gets its own vn_reference_op_s structure. */
906 temp.op0 = TMR_INDEX (ref);
907 temp.op1 = TMR_STEP (ref);
908 temp.op2 = TMR_OFFSET (ref);
909 temp.clique = MR_DEPENDENCE_CLIQUE (ref);
910 temp.base = MR_DEPENDENCE_BASE (ref);
911 result->safe_push (temp);
912 memset (&temp, 0, sizeof (temp));
913 temp.type = NULL_TREE;
914 temp.opcode = ERROR_MARK;
915 temp.op0 = TMR_INDEX2 (ref);
916 temp.off = -1;
917 break;
918 case BIT_FIELD_REF:
919 /* Record bits, position and storage order. */
920 temp.op0 = TREE_OPERAND (ref, 1);
921 temp.op1 = TREE_OPERAND (ref, 2);
922 if (!multiple_p (bit_field_offset (ref), BITS_PER_UNIT, &temp.off))
923 temp.off = -1;
924 temp.reverse = REF_REVERSE_STORAGE_ORDER (ref);
925 break;
926 case COMPONENT_REF:
927 /* The field decl is enough to unambiguously specify the field,
928 so use its type here. */
929 temp.type = TREE_TYPE (TREE_OPERAND (ref, 1));
930 temp.op0 = TREE_OPERAND (ref, 1);
931 temp.op1 = TREE_OPERAND (ref, 2);
932 temp.reverse = (AGGREGATE_TYPE_P (TREE_TYPE (TREE_OPERAND (ref, 0)))
933 && TYPE_REVERSE_STORAGE_ORDER
934 (TREE_TYPE (TREE_OPERAND (ref, 0))));
936 tree this_offset = component_ref_field_offset (ref);
937 if (this_offset
938 && poly_int_tree_p (this_offset))
940 tree bit_offset = DECL_FIELD_BIT_OFFSET (TREE_OPERAND (ref, 1));
941 if (TREE_INT_CST_LOW (bit_offset) % BITS_PER_UNIT == 0)
943 poly_offset_int off
944 = (wi::to_poly_offset (this_offset)
945 + (wi::to_offset (bit_offset) >> LOG2_BITS_PER_UNIT));
946 /* Probibit value-numbering zero offset components
947 of addresses the same before the pass folding
948 __builtin_object_size had a chance to run. */
949 if (TREE_CODE (orig) != ADDR_EXPR
950 || maybe_ne (off, 0)
951 || (cfun->curr_properties & PROP_objsz))
952 off.to_shwi (&temp.off);
956 break;
957 case ARRAY_RANGE_REF:
958 case ARRAY_REF:
960 tree eltype = TREE_TYPE (TREE_TYPE (TREE_OPERAND (ref, 0)));
961 /* Record index as operand. */
962 temp.op0 = TREE_OPERAND (ref, 1);
963 /* Always record lower bounds and element size. */
964 temp.op1 = array_ref_low_bound (ref);
965 /* But record element size in units of the type alignment. */
966 temp.op2 = TREE_OPERAND (ref, 3);
967 temp.align = eltype->type_common.align;
968 if (! temp.op2)
969 temp.op2 = size_binop (EXACT_DIV_EXPR, TYPE_SIZE_UNIT (eltype),
970 size_int (TYPE_ALIGN_UNIT (eltype)));
971 if (poly_int_tree_p (temp.op0)
972 && poly_int_tree_p (temp.op1)
973 && TREE_CODE (temp.op2) == INTEGER_CST)
975 poly_offset_int off = ((wi::to_poly_offset (temp.op0)
976 - wi::to_poly_offset (temp.op1))
977 * wi::to_offset (temp.op2)
978 * vn_ref_op_align_unit (&temp));
979 off.to_shwi (&temp.off);
981 temp.reverse = (AGGREGATE_TYPE_P (TREE_TYPE (TREE_OPERAND (ref, 0)))
982 && TYPE_REVERSE_STORAGE_ORDER
983 (TREE_TYPE (TREE_OPERAND (ref, 0))));
985 break;
986 case VAR_DECL:
987 if (DECL_HARD_REGISTER (ref))
989 temp.op0 = ref;
990 break;
992 /* Fallthru. */
993 case PARM_DECL:
994 case CONST_DECL:
995 case RESULT_DECL:
996 /* Canonicalize decls to MEM[&decl] which is what we end up with
997 when valueizing MEM[ptr] with ptr = &decl. */
998 temp.opcode = MEM_REF;
999 temp.op0 = build_int_cst (build_pointer_type (TREE_TYPE (ref)), 0);
1000 temp.off = 0;
1001 result->safe_push (temp);
1002 temp.opcode = ADDR_EXPR;
1003 temp.op0 = build1 (ADDR_EXPR, TREE_TYPE (temp.op0), ref);
1004 temp.type = TREE_TYPE (temp.op0);
1005 temp.off = -1;
1006 break;
1007 case STRING_CST:
1008 case INTEGER_CST:
1009 case POLY_INT_CST:
1010 case COMPLEX_CST:
1011 case VECTOR_CST:
1012 case REAL_CST:
1013 case FIXED_CST:
1014 case CONSTRUCTOR:
1015 case SSA_NAME:
1016 temp.op0 = ref;
1017 break;
1018 case ADDR_EXPR:
1019 if (is_gimple_min_invariant (ref))
1021 temp.op0 = ref;
1022 break;
1024 break;
1025 /* These are only interesting for their operands, their
1026 existence, and their type. They will never be the last
1027 ref in the chain of references (IE they require an
1028 operand), so we don't have to put anything
1029 for op* as it will be handled by the iteration */
1030 case REALPART_EXPR:
1031 temp.off = 0;
1032 break;
1033 case VIEW_CONVERT_EXPR:
1034 temp.off = 0;
1035 temp.reverse = storage_order_barrier_p (ref);
1036 break;
1037 case IMAGPART_EXPR:
1038 /* This is only interesting for its constant offset. */
1039 temp.off = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (TREE_TYPE (ref)));
1040 break;
1041 default:
1042 gcc_unreachable ();
1044 result->safe_push (temp);
1046 if (REFERENCE_CLASS_P (ref)
1047 || TREE_CODE (ref) == MODIFY_EXPR
1048 || TREE_CODE (ref) == WITH_SIZE_EXPR
1049 || (TREE_CODE (ref) == ADDR_EXPR
1050 && !is_gimple_min_invariant (ref)))
1051 ref = TREE_OPERAND (ref, 0);
1052 else
1053 ref = NULL_TREE;
1057 /* Build a alias-oracle reference abstraction in *REF from the vn_reference
1058 operands in *OPS, the reference alias set SET and the reference type TYPE.
1059 Return true if something useful was produced. */
1061 bool
1062 ao_ref_init_from_vn_reference (ao_ref *ref,
1063 alias_set_type set, alias_set_type base_set,
1064 tree type, const vec<vn_reference_op_s> &ops)
1066 unsigned i;
1067 tree base = NULL_TREE;
1068 tree *op0_p = &base;
1069 poly_offset_int offset = 0;
1070 poly_offset_int max_size;
1071 poly_offset_int size = -1;
1072 tree size_tree = NULL_TREE;
1074 /* We don't handle calls. */
1075 if (!type)
1076 return false;
1078 machine_mode mode = TYPE_MODE (type);
1079 if (mode == BLKmode)
1080 size_tree = TYPE_SIZE (type);
1081 else
1082 size = GET_MODE_BITSIZE (mode);
1083 if (size_tree != NULL_TREE
1084 && poly_int_tree_p (size_tree))
1085 size = wi::to_poly_offset (size_tree);
1087 /* Lower the final access size from the outermost expression. */
1088 const_vn_reference_op_t cst_op = &ops[0];
1089 /* Cast away constness for the sake of the const-unsafe
1090 FOR_EACH_VEC_ELT(). */
1091 vn_reference_op_t op = const_cast<vn_reference_op_t>(cst_op);
1092 size_tree = NULL_TREE;
1093 if (op->opcode == COMPONENT_REF)
1094 size_tree = DECL_SIZE (op->op0);
1095 else if (op->opcode == BIT_FIELD_REF)
1096 size_tree = op->op0;
1097 if (size_tree != NULL_TREE
1098 && poly_int_tree_p (size_tree)
1099 && (!known_size_p (size)
1100 || known_lt (wi::to_poly_offset (size_tree), size)))
1101 size = wi::to_poly_offset (size_tree);
1103 /* Initially, maxsize is the same as the accessed element size.
1104 In the following it will only grow (or become -1). */
1105 max_size = size;
1107 /* Compute cumulative bit-offset for nested component-refs and array-refs,
1108 and find the ultimate containing object. */
1109 FOR_EACH_VEC_ELT (ops, i, op)
1111 switch (op->opcode)
1113 /* These may be in the reference ops, but we cannot do anything
1114 sensible with them here. */
1115 case ADDR_EXPR:
1116 /* Apart from ADDR_EXPR arguments to MEM_REF. */
1117 if (base != NULL_TREE
1118 && TREE_CODE (base) == MEM_REF
1119 && op->op0
1120 && DECL_P (TREE_OPERAND (op->op0, 0)))
1122 const_vn_reference_op_t pop = &ops[i-1];
1123 base = TREE_OPERAND (op->op0, 0);
1124 if (known_eq (pop->off, -1))
1126 max_size = -1;
1127 offset = 0;
1129 else
1130 offset += pop->off * BITS_PER_UNIT;
1131 op0_p = NULL;
1132 break;
1134 /* Fallthru. */
1135 case CALL_EXPR:
1136 return false;
1138 /* Record the base objects. */
1139 case MEM_REF:
1140 *op0_p = build2 (MEM_REF, op->type,
1141 NULL_TREE, op->op0);
1142 MR_DEPENDENCE_CLIQUE (*op0_p) = op->clique;
1143 MR_DEPENDENCE_BASE (*op0_p) = op->base;
1144 op0_p = &TREE_OPERAND (*op0_p, 0);
1145 break;
1147 case VAR_DECL:
1148 case PARM_DECL:
1149 case RESULT_DECL:
1150 case SSA_NAME:
1151 *op0_p = op->op0;
1152 op0_p = NULL;
1153 break;
1155 /* And now the usual component-reference style ops. */
1156 case BIT_FIELD_REF:
1157 offset += wi::to_poly_offset (op->op1);
1158 break;
1160 case COMPONENT_REF:
1162 tree field = op->op0;
1163 /* We do not have a complete COMPONENT_REF tree here so we
1164 cannot use component_ref_field_offset. Do the interesting
1165 parts manually. */
1166 tree this_offset = DECL_FIELD_OFFSET (field);
1168 if (op->op1 || !poly_int_tree_p (this_offset))
1169 max_size = -1;
1170 else
1172 poly_offset_int woffset = (wi::to_poly_offset (this_offset)
1173 << LOG2_BITS_PER_UNIT);
1174 woffset += wi::to_offset (DECL_FIELD_BIT_OFFSET (field));
1175 offset += woffset;
1177 break;
1180 case ARRAY_RANGE_REF:
1181 case ARRAY_REF:
1182 /* We recorded the lower bound and the element size. */
1183 if (!poly_int_tree_p (op->op0)
1184 || !poly_int_tree_p (op->op1)
1185 || TREE_CODE (op->op2) != INTEGER_CST)
1186 max_size = -1;
1187 else
1189 poly_offset_int woffset
1190 = wi::sext (wi::to_poly_offset (op->op0)
1191 - wi::to_poly_offset (op->op1),
1192 TYPE_PRECISION (sizetype));
1193 woffset *= wi::to_offset (op->op2) * vn_ref_op_align_unit (op);
1194 woffset <<= LOG2_BITS_PER_UNIT;
1195 offset += woffset;
1197 break;
1199 case REALPART_EXPR:
1200 break;
1202 case IMAGPART_EXPR:
1203 offset += size;
1204 break;
1206 case VIEW_CONVERT_EXPR:
1207 break;
1209 case STRING_CST:
1210 case INTEGER_CST:
1211 case COMPLEX_CST:
1212 case VECTOR_CST:
1213 case REAL_CST:
1214 case CONSTRUCTOR:
1215 case CONST_DECL:
1216 return false;
1218 default:
1219 return false;
1223 if (base == NULL_TREE)
1224 return false;
1226 ref->ref = NULL_TREE;
1227 ref->base = base;
1228 ref->ref_alias_set = set;
1229 ref->base_alias_set = base_set;
1230 /* We discount volatiles from value-numbering elsewhere. */
1231 ref->volatile_p = false;
1233 if (!size.to_shwi (&ref->size) || maybe_lt (ref->size, 0))
1235 ref->offset = 0;
1236 ref->size = -1;
1237 ref->max_size = -1;
1238 return true;
1241 if (!offset.to_shwi (&ref->offset))
1243 ref->offset = 0;
1244 ref->max_size = -1;
1245 return true;
1248 if (!max_size.to_shwi (&ref->max_size) || maybe_lt (ref->max_size, 0))
1249 ref->max_size = -1;
1251 return true;
1254 /* Copy the operations present in load/store/call REF into RESULT, a vector of
1255 vn_reference_op_s's. */
1257 static void
1258 copy_reference_ops_from_call (gcall *call,
1259 vec<vn_reference_op_s> *result)
1261 vn_reference_op_s temp;
1262 unsigned i;
1263 tree lhs = gimple_call_lhs (call);
1264 int lr;
1266 /* If 2 calls have a different non-ssa lhs, vdef value numbers should be
1267 different. By adding the lhs here in the vector, we ensure that the
1268 hashcode is different, guaranteeing a different value number. */
1269 if (lhs && TREE_CODE (lhs) != SSA_NAME)
1271 memset (&temp, 0, sizeof (temp));
1272 temp.opcode = MODIFY_EXPR;
1273 temp.type = TREE_TYPE (lhs);
1274 temp.op0 = lhs;
1275 temp.off = -1;
1276 result->safe_push (temp);
1279 /* Copy the type, opcode, function, static chain and EH region, if any. */
1280 memset (&temp, 0, sizeof (temp));
1281 temp.type = gimple_call_fntype (call);
1282 temp.opcode = CALL_EXPR;
1283 temp.op0 = gimple_call_fn (call);
1284 if (gimple_call_internal_p (call))
1285 temp.clique = gimple_call_internal_fn (call);
1286 temp.op1 = gimple_call_chain (call);
1287 if (stmt_could_throw_p (cfun, call) && (lr = lookup_stmt_eh_lp (call)) > 0)
1288 temp.op2 = size_int (lr);
1289 temp.off = -1;
1290 result->safe_push (temp);
1292 /* Copy the call arguments. As they can be references as well,
1293 just chain them together. */
1294 for (i = 0; i < gimple_call_num_args (call); ++i)
1296 tree callarg = gimple_call_arg (call, i);
1297 copy_reference_ops_from_ref (callarg, result);
1301 /* Fold *& at position *I_P in a vn_reference_op_s vector *OPS. Updates
1302 *I_P to point to the last element of the replacement. */
1303 static bool
1304 vn_reference_fold_indirect (vec<vn_reference_op_s> *ops,
1305 unsigned int *i_p)
1307 unsigned int i = *i_p;
1308 vn_reference_op_t op = &(*ops)[i];
1309 vn_reference_op_t mem_op = &(*ops)[i - 1];
1310 tree addr_base;
1311 poly_int64 addr_offset = 0;
1313 /* The only thing we have to do is from &OBJ.foo.bar add the offset
1314 from .foo.bar to the preceding MEM_REF offset and replace the
1315 address with &OBJ. */
1316 addr_base = get_addr_base_and_unit_offset_1 (TREE_OPERAND (op->op0, 0),
1317 &addr_offset, vn_valueize);
1318 gcc_checking_assert (addr_base && TREE_CODE (addr_base) != MEM_REF);
1319 if (addr_base != TREE_OPERAND (op->op0, 0))
1321 poly_offset_int off
1322 = (poly_offset_int::from (wi::to_poly_wide (mem_op->op0),
1323 SIGNED)
1324 + addr_offset);
1325 mem_op->op0 = wide_int_to_tree (TREE_TYPE (mem_op->op0), off);
1326 op->op0 = build_fold_addr_expr (addr_base);
1327 if (tree_fits_shwi_p (mem_op->op0))
1328 mem_op->off = tree_to_shwi (mem_op->op0);
1329 else
1330 mem_op->off = -1;
1331 return true;
1333 return false;
1336 /* Fold *& at position *I_P in a vn_reference_op_s vector *OPS. Updates
1337 *I_P to point to the last element of the replacement. */
1338 static bool
1339 vn_reference_maybe_forwprop_address (vec<vn_reference_op_s> *ops,
1340 unsigned int *i_p)
1342 bool changed = false;
1343 vn_reference_op_t op;
1347 unsigned int i = *i_p;
1348 op = &(*ops)[i];
1349 vn_reference_op_t mem_op = &(*ops)[i - 1];
1350 gimple *def_stmt;
1351 enum tree_code code;
1352 poly_offset_int off;
1354 def_stmt = SSA_NAME_DEF_STMT (op->op0);
1355 if (!is_gimple_assign (def_stmt))
1356 return changed;
1358 code = gimple_assign_rhs_code (def_stmt);
1359 if (code != ADDR_EXPR
1360 && code != POINTER_PLUS_EXPR)
1361 return changed;
1363 off = poly_offset_int::from (wi::to_poly_wide (mem_op->op0), SIGNED);
1365 /* The only thing we have to do is from &OBJ.foo.bar add the offset
1366 from .foo.bar to the preceding MEM_REF offset and replace the
1367 address with &OBJ. */
1368 if (code == ADDR_EXPR)
1370 tree addr, addr_base;
1371 poly_int64 addr_offset;
1373 addr = gimple_assign_rhs1 (def_stmt);
1374 addr_base = get_addr_base_and_unit_offset_1 (TREE_OPERAND (addr, 0),
1375 &addr_offset,
1376 vn_valueize);
1377 /* If that didn't work because the address isn't invariant propagate
1378 the reference tree from the address operation in case the current
1379 dereference isn't offsetted. */
1380 if (!addr_base
1381 && *i_p == ops->length () - 1
1382 && known_eq (off, 0)
1383 /* This makes us disable this transform for PRE where the
1384 reference ops might be also used for code insertion which
1385 is invalid. */
1386 && default_vn_walk_kind == VN_WALKREWRITE)
1388 auto_vec<vn_reference_op_s, 32> tem;
1389 copy_reference_ops_from_ref (TREE_OPERAND (addr, 0), &tem);
1390 /* Make sure to preserve TBAA info. The only objects not
1391 wrapped in MEM_REFs that can have their address taken are
1392 STRING_CSTs. */
1393 if (tem.length () >= 2
1394 && tem[tem.length () - 2].opcode == MEM_REF)
1396 vn_reference_op_t new_mem_op = &tem[tem.length () - 2];
1397 new_mem_op->op0
1398 = wide_int_to_tree (TREE_TYPE (mem_op->op0),
1399 wi::to_poly_wide (new_mem_op->op0));
1401 else
1402 gcc_assert (tem.last ().opcode == STRING_CST);
1403 ops->pop ();
1404 ops->pop ();
1405 ops->safe_splice (tem);
1406 --*i_p;
1407 return true;
1409 if (!addr_base
1410 || TREE_CODE (addr_base) != MEM_REF
1411 || (TREE_CODE (TREE_OPERAND (addr_base, 0)) == SSA_NAME
1412 && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (TREE_OPERAND (addr_base,
1413 0))))
1414 return changed;
1416 off += addr_offset;
1417 off += mem_ref_offset (addr_base);
1418 op->op0 = TREE_OPERAND (addr_base, 0);
1420 else
1422 tree ptr, ptroff;
1423 ptr = gimple_assign_rhs1 (def_stmt);
1424 ptroff = gimple_assign_rhs2 (def_stmt);
1425 if (TREE_CODE (ptr) != SSA_NAME
1426 || SSA_NAME_OCCURS_IN_ABNORMAL_PHI (ptr)
1427 /* Make sure to not endlessly recurse.
1428 See gcc.dg/tree-ssa/20040408-1.c for an example. Can easily
1429 happen when we value-number a PHI to its backedge value. */
1430 || SSA_VAL (ptr) == op->op0
1431 || !poly_int_tree_p (ptroff))
1432 return changed;
1434 off += wi::to_poly_offset (ptroff);
1435 op->op0 = ptr;
1438 mem_op->op0 = wide_int_to_tree (TREE_TYPE (mem_op->op0), off);
1439 if (tree_fits_shwi_p (mem_op->op0))
1440 mem_op->off = tree_to_shwi (mem_op->op0);
1441 else
1442 mem_op->off = -1;
1443 /* ??? Can end up with endless recursion here!?
1444 gcc.c-torture/execute/strcmp-1.c */
1445 if (TREE_CODE (op->op0) == SSA_NAME)
1446 op->op0 = SSA_VAL (op->op0);
1447 if (TREE_CODE (op->op0) != SSA_NAME)
1448 op->opcode = TREE_CODE (op->op0);
1450 changed = true;
1452 /* Tail-recurse. */
1453 while (TREE_CODE (op->op0) == SSA_NAME);
1455 /* Fold a remaining *&. */
1456 if (TREE_CODE (op->op0) == ADDR_EXPR)
1457 vn_reference_fold_indirect (ops, i_p);
1459 return changed;
1462 /* Optimize the reference REF to a constant if possible or return
1463 NULL_TREE if not. */
1465 tree
1466 fully_constant_vn_reference_p (vn_reference_t ref)
1468 vec<vn_reference_op_s> operands = ref->operands;
1469 vn_reference_op_t op;
1471 /* Try to simplify the translated expression if it is
1472 a call to a builtin function with at most two arguments. */
1473 op = &operands[0];
1474 if (op->opcode == CALL_EXPR
1475 && (!op->op0
1476 || (TREE_CODE (op->op0) == ADDR_EXPR
1477 && TREE_CODE (TREE_OPERAND (op->op0, 0)) == FUNCTION_DECL
1478 && fndecl_built_in_p (TREE_OPERAND (op->op0, 0),
1479 BUILT_IN_NORMAL)))
1480 && operands.length () >= 2
1481 && operands.length () <= 3)
1483 vn_reference_op_t arg0, arg1 = NULL;
1484 bool anyconst = false;
1485 arg0 = &operands[1];
1486 if (operands.length () > 2)
1487 arg1 = &operands[2];
1488 if (TREE_CODE_CLASS (arg0->opcode) == tcc_constant
1489 || (arg0->opcode == ADDR_EXPR
1490 && is_gimple_min_invariant (arg0->op0)))
1491 anyconst = true;
1492 if (arg1
1493 && (TREE_CODE_CLASS (arg1->opcode) == tcc_constant
1494 || (arg1->opcode == ADDR_EXPR
1495 && is_gimple_min_invariant (arg1->op0))))
1496 anyconst = true;
1497 if (anyconst)
1499 combined_fn fn;
1500 if (op->op0)
1501 fn = as_combined_fn (DECL_FUNCTION_CODE
1502 (TREE_OPERAND (op->op0, 0)));
1503 else
1504 fn = as_combined_fn ((internal_fn) op->clique);
1505 tree folded;
1506 if (arg1)
1507 folded = fold_const_call (fn, ref->type, arg0->op0, arg1->op0);
1508 else
1509 folded = fold_const_call (fn, ref->type, arg0->op0);
1510 if (folded
1511 && is_gimple_min_invariant (folded))
1512 return folded;
1516 /* Simplify reads from constants or constant initializers. */
1517 else if (BITS_PER_UNIT == 8
1518 && ref->type
1519 && COMPLETE_TYPE_P (ref->type)
1520 && is_gimple_reg_type (ref->type))
1522 poly_int64 off = 0;
1523 HOST_WIDE_INT size;
1524 if (INTEGRAL_TYPE_P (ref->type))
1525 size = TYPE_PRECISION (ref->type);
1526 else if (tree_fits_shwi_p (TYPE_SIZE (ref->type)))
1527 size = tree_to_shwi (TYPE_SIZE (ref->type));
1528 else
1529 return NULL_TREE;
1530 if (size % BITS_PER_UNIT != 0
1531 || size > MAX_BITSIZE_MODE_ANY_MODE)
1532 return NULL_TREE;
1533 size /= BITS_PER_UNIT;
1534 unsigned i;
1535 for (i = 0; i < operands.length (); ++i)
1537 if (TREE_CODE_CLASS (operands[i].opcode) == tcc_constant)
1539 ++i;
1540 break;
1542 if (known_eq (operands[i].off, -1))
1543 return NULL_TREE;
1544 off += operands[i].off;
1545 if (operands[i].opcode == MEM_REF)
1547 ++i;
1548 break;
1551 vn_reference_op_t base = &operands[--i];
1552 tree ctor = error_mark_node;
1553 tree decl = NULL_TREE;
1554 if (TREE_CODE_CLASS (base->opcode) == tcc_constant)
1555 ctor = base->op0;
1556 else if (base->opcode == MEM_REF
1557 && base[1].opcode == ADDR_EXPR
1558 && (TREE_CODE (TREE_OPERAND (base[1].op0, 0)) == VAR_DECL
1559 || TREE_CODE (TREE_OPERAND (base[1].op0, 0)) == CONST_DECL
1560 || TREE_CODE (TREE_OPERAND (base[1].op0, 0)) == STRING_CST))
1562 decl = TREE_OPERAND (base[1].op0, 0);
1563 if (TREE_CODE (decl) == STRING_CST)
1564 ctor = decl;
1565 else
1566 ctor = ctor_for_folding (decl);
1568 if (ctor == NULL_TREE)
1569 return build_zero_cst (ref->type);
1570 else if (ctor != error_mark_node)
1572 HOST_WIDE_INT const_off;
1573 if (decl)
1575 tree res = fold_ctor_reference (ref->type, ctor,
1576 off * BITS_PER_UNIT,
1577 size * BITS_PER_UNIT, decl);
1578 if (res)
1580 STRIP_USELESS_TYPE_CONVERSION (res);
1581 if (is_gimple_min_invariant (res))
1582 return res;
1585 else if (off.is_constant (&const_off))
1587 unsigned char buf[MAX_BITSIZE_MODE_ANY_MODE / BITS_PER_UNIT];
1588 int len = native_encode_expr (ctor, buf, size, const_off);
1589 if (len > 0)
1590 return native_interpret_expr (ref->type, buf, len);
1595 return NULL_TREE;
1598 /* Return true if OPS contain a storage order barrier. */
1600 static bool
1601 contains_storage_order_barrier_p (vec<vn_reference_op_s> ops)
1603 vn_reference_op_t op;
1604 unsigned i;
1606 FOR_EACH_VEC_ELT (ops, i, op)
1607 if (op->opcode == VIEW_CONVERT_EXPR && op->reverse)
1608 return true;
1610 return false;
1613 /* Return true if OPS represent an access with reverse storage order. */
1615 static bool
1616 reverse_storage_order_for_component_p (vec<vn_reference_op_s> ops)
1618 unsigned i = 0;
1619 if (ops[i].opcode == REALPART_EXPR || ops[i].opcode == IMAGPART_EXPR)
1620 ++i;
1621 switch (ops[i].opcode)
1623 case ARRAY_REF:
1624 case COMPONENT_REF:
1625 case BIT_FIELD_REF:
1626 case MEM_REF:
1627 return ops[i].reverse;
1628 default:
1629 return false;
1633 /* Transform any SSA_NAME's in a vector of vn_reference_op_s
1634 structures into their value numbers. This is done in-place, and
1635 the vector passed in is returned. *VALUEIZED_ANYTHING will specify
1636 whether any operands were valueized. */
1638 static void
1639 valueize_refs_1 (vec<vn_reference_op_s> *orig, bool *valueized_anything,
1640 bool with_avail = false)
1642 *valueized_anything = false;
1644 for (unsigned i = 0; i < orig->length (); ++i)
1646 re_valueize:
1647 vn_reference_op_t vro = &(*orig)[i];
1648 if (vro->opcode == SSA_NAME
1649 || (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME))
1651 tree tem = with_avail ? vn_valueize (vro->op0) : SSA_VAL (vro->op0);
1652 if (tem != vro->op0)
1654 *valueized_anything = true;
1655 vro->op0 = tem;
1657 /* If it transforms from an SSA_NAME to a constant, update
1658 the opcode. */
1659 if (TREE_CODE (vro->op0) != SSA_NAME && vro->opcode == SSA_NAME)
1660 vro->opcode = TREE_CODE (vro->op0);
1662 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
1664 tree tem = with_avail ? vn_valueize (vro->op1) : SSA_VAL (vro->op1);
1665 if (tem != vro->op1)
1667 *valueized_anything = true;
1668 vro->op1 = tem;
1671 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
1673 tree tem = with_avail ? vn_valueize (vro->op2) : SSA_VAL (vro->op2);
1674 if (tem != vro->op2)
1676 *valueized_anything = true;
1677 vro->op2 = tem;
1680 /* If it transforms from an SSA_NAME to an address, fold with
1681 a preceding indirect reference. */
1682 if (i > 0
1683 && vro->op0
1684 && TREE_CODE (vro->op0) == ADDR_EXPR
1685 && (*orig)[i - 1].opcode == MEM_REF)
1687 if (vn_reference_fold_indirect (orig, &i))
1688 *valueized_anything = true;
1690 else if (i > 0
1691 && vro->opcode == SSA_NAME
1692 && (*orig)[i - 1].opcode == MEM_REF)
1694 if (vn_reference_maybe_forwprop_address (orig, &i))
1696 *valueized_anything = true;
1697 /* Re-valueize the current operand. */
1698 goto re_valueize;
1701 /* If it transforms a non-constant ARRAY_REF into a constant
1702 one, adjust the constant offset. */
1703 else if (vro->opcode == ARRAY_REF
1704 && known_eq (vro->off, -1)
1705 && poly_int_tree_p (vro->op0)
1706 && poly_int_tree_p (vro->op1)
1707 && TREE_CODE (vro->op2) == INTEGER_CST)
1709 poly_offset_int off = ((wi::to_poly_offset (vro->op0)
1710 - wi::to_poly_offset (vro->op1))
1711 * wi::to_offset (vro->op2)
1712 * vn_ref_op_align_unit (vro));
1713 off.to_shwi (&vro->off);
1718 static void
1719 valueize_refs (vec<vn_reference_op_s> *orig)
1721 bool tem;
1722 valueize_refs_1 (orig, &tem);
1725 static vec<vn_reference_op_s> shared_lookup_references;
1727 /* Create a vector of vn_reference_op_s structures from REF, a
1728 REFERENCE_CLASS_P tree. The vector is shared among all callers of
1729 this function. *VALUEIZED_ANYTHING will specify whether any
1730 operands were valueized. */
1732 static vec<vn_reference_op_s>
1733 valueize_shared_reference_ops_from_ref (tree ref, bool *valueized_anything)
1735 if (!ref)
1736 return vNULL;
1737 shared_lookup_references.truncate (0);
1738 copy_reference_ops_from_ref (ref, &shared_lookup_references);
1739 valueize_refs_1 (&shared_lookup_references, valueized_anything);
1740 return shared_lookup_references;
1743 /* Create a vector of vn_reference_op_s structures from CALL, a
1744 call statement. The vector is shared among all callers of
1745 this function. */
1747 static vec<vn_reference_op_s>
1748 valueize_shared_reference_ops_from_call (gcall *call)
1750 if (!call)
1751 return vNULL;
1752 shared_lookup_references.truncate (0);
1753 copy_reference_ops_from_call (call, &shared_lookup_references);
1754 valueize_refs (&shared_lookup_references);
1755 return shared_lookup_references;
1758 /* Lookup a SCCVN reference operation VR in the current hash table.
1759 Returns the resulting value number if it exists in the hash table,
1760 NULL_TREE otherwise. VNRESULT will be filled in with the actual
1761 vn_reference_t stored in the hashtable if something is found. */
1763 static tree
1764 vn_reference_lookup_1 (vn_reference_t vr, vn_reference_t *vnresult)
1766 vn_reference_s **slot;
1767 hashval_t hash;
1769 hash = vr->hashcode;
1770 slot = valid_info->references->find_slot_with_hash (vr, hash, NO_INSERT);
1771 if (slot)
1773 if (vnresult)
1774 *vnresult = (vn_reference_t)*slot;
1775 return ((vn_reference_t)*slot)->result;
1778 return NULL_TREE;
1782 /* Partial definition tracking support. */
1784 struct pd_range
1786 HOST_WIDE_INT offset;
1787 HOST_WIDE_INT size;
1790 struct pd_data
1792 tree rhs;
1793 HOST_WIDE_INT offset;
1794 HOST_WIDE_INT size;
1797 /* Context for alias walking. */
1799 struct vn_walk_cb_data
1801 vn_walk_cb_data (vn_reference_t vr_, tree orig_ref_, tree *last_vuse_ptr_,
1802 vn_lookup_kind vn_walk_kind_, bool tbaa_p_, tree mask_,
1803 bool redundant_store_removal_p_)
1804 : vr (vr_), last_vuse_ptr (last_vuse_ptr_), last_vuse (NULL_TREE),
1805 mask (mask_), masked_result (NULL_TREE), vn_walk_kind (vn_walk_kind_),
1806 tbaa_p (tbaa_p_), redundant_store_removal_p (redundant_store_removal_p_),
1807 saved_operands (vNULL), first_set (-2), first_base_set (-2),
1808 known_ranges (NULL)
1810 if (!last_vuse_ptr)
1811 last_vuse_ptr = &last_vuse;
1812 ao_ref_init (&orig_ref, orig_ref_);
1813 if (mask)
1815 wide_int w = wi::to_wide (mask);
1816 unsigned int pos = 0, prec = w.get_precision ();
1817 pd_data pd;
1818 pd.rhs = build_constructor (NULL_TREE, NULL);
1819 /* When bitwise and with a constant is done on a memory load,
1820 we don't really need all the bits to be defined or defined
1821 to constants, we don't really care what is in the position
1822 corresponding to 0 bits in the mask.
1823 So, push the ranges of those 0 bits in the mask as artificial
1824 zero stores and let the partial def handling code do the
1825 rest. */
1826 while (pos < prec)
1828 int tz = wi::ctz (w);
1829 if (pos + tz > prec)
1830 tz = prec - pos;
1831 if (tz)
1833 if (BYTES_BIG_ENDIAN)
1834 pd.offset = prec - pos - tz;
1835 else
1836 pd.offset = pos;
1837 pd.size = tz;
1838 void *r = push_partial_def (pd, 0, 0, 0, prec);
1839 gcc_assert (r == NULL_TREE);
1841 pos += tz;
1842 if (pos == prec)
1843 break;
1844 w = wi::lrshift (w, tz);
1845 tz = wi::ctz (wi::bit_not (w));
1846 if (pos + tz > prec)
1847 tz = prec - pos;
1848 pos += tz;
1849 w = wi::lrshift (w, tz);
1853 ~vn_walk_cb_data ();
1854 void *finish (alias_set_type, alias_set_type, tree);
1855 void *push_partial_def (pd_data pd,
1856 alias_set_type, alias_set_type, HOST_WIDE_INT,
1857 HOST_WIDE_INT);
1859 vn_reference_t vr;
1860 ao_ref orig_ref;
1861 tree *last_vuse_ptr;
1862 tree last_vuse;
1863 tree mask;
1864 tree masked_result;
1865 vn_lookup_kind vn_walk_kind;
1866 bool tbaa_p;
1867 bool redundant_store_removal_p;
1868 vec<vn_reference_op_s> saved_operands;
1870 /* The VDEFs of partial defs we come along. */
1871 auto_vec<pd_data, 2> partial_defs;
1872 /* The first defs range to avoid splay tree setup in most cases. */
1873 pd_range first_range;
1874 alias_set_type first_set;
1875 alias_set_type first_base_set;
1876 splay_tree known_ranges;
1877 obstack ranges_obstack;
1880 vn_walk_cb_data::~vn_walk_cb_data ()
1882 if (known_ranges)
1884 splay_tree_delete (known_ranges);
1885 obstack_free (&ranges_obstack, NULL);
1887 saved_operands.release ();
1890 void *
1891 vn_walk_cb_data::finish (alias_set_type set, alias_set_type base_set, tree val)
1893 if (first_set != -2)
1895 set = first_set;
1896 base_set = first_base_set;
1898 if (mask)
1900 masked_result = val;
1901 return (void *) -1;
1903 vec<vn_reference_op_s> &operands
1904 = saved_operands.exists () ? saved_operands : vr->operands;
1905 return vn_reference_lookup_or_insert_for_pieces (last_vuse, set, base_set,
1906 vr->type, operands, val);
1909 /* pd_range splay-tree helpers. */
1911 static int
1912 pd_range_compare (splay_tree_key offset1p, splay_tree_key offset2p)
1914 HOST_WIDE_INT offset1 = *(HOST_WIDE_INT *)offset1p;
1915 HOST_WIDE_INT offset2 = *(HOST_WIDE_INT *)offset2p;
1916 if (offset1 < offset2)
1917 return -1;
1918 else if (offset1 > offset2)
1919 return 1;
1920 return 0;
1923 static void *
1924 pd_tree_alloc (int size, void *data_)
1926 vn_walk_cb_data *data = (vn_walk_cb_data *)data_;
1927 return obstack_alloc (&data->ranges_obstack, size);
1930 static void
1931 pd_tree_dealloc (void *, void *)
1935 /* Push PD to the vector of partial definitions returning a
1936 value when we are ready to combine things with VUSE, SET and MAXSIZEI,
1937 NULL when we want to continue looking for partial defs or -1
1938 on failure. */
1940 void *
1941 vn_walk_cb_data::push_partial_def (pd_data pd,
1942 alias_set_type set, alias_set_type base_set,
1943 HOST_WIDE_INT offseti,
1944 HOST_WIDE_INT maxsizei)
1946 const HOST_WIDE_INT bufsize = 64;
1947 /* We're using a fixed buffer for encoding so fail early if the object
1948 we want to interpret is bigger. */
1949 if (maxsizei > bufsize * BITS_PER_UNIT
1950 || CHAR_BIT != 8
1951 || BITS_PER_UNIT != 8
1952 /* Not prepared to handle PDP endian. */
1953 || BYTES_BIG_ENDIAN != WORDS_BIG_ENDIAN)
1954 return (void *)-1;
1956 /* Turn too large constant stores into non-constant stores. */
1957 if (CONSTANT_CLASS_P (pd.rhs) && pd.size > bufsize * BITS_PER_UNIT)
1958 pd.rhs = error_mark_node;
1960 /* And for non-constant or CONSTRUCTOR stores shrink them to only keep at
1961 most a partial byte before and/or after the region. */
1962 if (!CONSTANT_CLASS_P (pd.rhs))
1964 if (pd.offset < offseti)
1966 HOST_WIDE_INT o = ROUND_DOWN (offseti - pd.offset, BITS_PER_UNIT);
1967 gcc_assert (pd.size > o);
1968 pd.size -= o;
1969 pd.offset += o;
1971 if (pd.size > maxsizei)
1972 pd.size = maxsizei + ((pd.size - maxsizei) % BITS_PER_UNIT);
1975 pd.offset -= offseti;
1977 bool pd_constant_p = (TREE_CODE (pd.rhs) == CONSTRUCTOR
1978 || CONSTANT_CLASS_P (pd.rhs));
1979 if (partial_defs.is_empty ())
1981 /* If we get a clobber upfront, fail. */
1982 if (TREE_CLOBBER_P (pd.rhs))
1983 return (void *)-1;
1984 if (!pd_constant_p)
1985 return (void *)-1;
1986 partial_defs.safe_push (pd);
1987 first_range.offset = pd.offset;
1988 first_range.size = pd.size;
1989 first_set = set;
1990 first_base_set = base_set;
1991 last_vuse_ptr = NULL;
1992 /* Continue looking for partial defs. */
1993 return NULL;
1996 if (!known_ranges)
1998 /* ??? Optimize the case where the 2nd partial def completes things. */
1999 gcc_obstack_init (&ranges_obstack);
2000 known_ranges = splay_tree_new_with_allocator (pd_range_compare, 0, 0,
2001 pd_tree_alloc,
2002 pd_tree_dealloc, this);
2003 splay_tree_insert (known_ranges,
2004 (splay_tree_key)&first_range.offset,
2005 (splay_tree_value)&first_range);
2008 pd_range newr = { pd.offset, pd.size };
2009 splay_tree_node n;
2010 pd_range *r;
2011 /* Lookup the predecessor of offset + 1 and see if we need to merge. */
2012 HOST_WIDE_INT loffset = newr.offset + 1;
2013 if ((n = splay_tree_predecessor (known_ranges, (splay_tree_key)&loffset))
2014 && ((r = (pd_range *)n->value), true)
2015 && ranges_known_overlap_p (r->offset, r->size + 1,
2016 newr.offset, newr.size))
2018 /* Ignore partial defs already covered. Here we also drop shadowed
2019 clobbers arriving here at the floor. */
2020 if (known_subrange_p (newr.offset, newr.size, r->offset, r->size))
2021 return NULL;
2022 r->size = MAX (r->offset + r->size, newr.offset + newr.size) - r->offset;
2024 else
2026 /* newr.offset wasn't covered yet, insert the range. */
2027 r = XOBNEW (&ranges_obstack, pd_range);
2028 *r = newr;
2029 splay_tree_insert (known_ranges, (splay_tree_key)&r->offset,
2030 (splay_tree_value)r);
2032 /* Merge r which now contains newr and is a member of the splay tree with
2033 adjacent overlapping ranges. */
2034 pd_range *rafter;
2035 while ((n = splay_tree_successor (known_ranges, (splay_tree_key)&r->offset))
2036 && ((rafter = (pd_range *)n->value), true)
2037 && ranges_known_overlap_p (r->offset, r->size + 1,
2038 rafter->offset, rafter->size))
2040 r->size = MAX (r->offset + r->size,
2041 rafter->offset + rafter->size) - r->offset;
2042 splay_tree_remove (known_ranges, (splay_tree_key)&rafter->offset);
2044 /* If we get a clobber, fail. */
2045 if (TREE_CLOBBER_P (pd.rhs))
2046 return (void *)-1;
2047 /* Non-constants are OK as long as they are shadowed by a constant. */
2048 if (!pd_constant_p)
2049 return (void *)-1;
2050 partial_defs.safe_push (pd);
2052 /* Now we have merged newr into the range tree. When we have covered
2053 [offseti, sizei] then the tree will contain exactly one node which has
2054 the desired properties and it will be 'r'. */
2055 if (!known_subrange_p (0, maxsizei, r->offset, r->size))
2056 /* Continue looking for partial defs. */
2057 return NULL;
2059 /* Now simply native encode all partial defs in reverse order. */
2060 unsigned ndefs = partial_defs.length ();
2061 /* We support up to 512-bit values (for V8DFmode). */
2062 unsigned char buffer[bufsize + 1];
2063 unsigned char this_buffer[bufsize + 1];
2064 int len;
2066 memset (buffer, 0, bufsize + 1);
2067 unsigned needed_len = ROUND_UP (maxsizei, BITS_PER_UNIT) / BITS_PER_UNIT;
2068 while (!partial_defs.is_empty ())
2070 pd_data pd = partial_defs.pop ();
2071 unsigned int amnt;
2072 if (TREE_CODE (pd.rhs) == CONSTRUCTOR)
2074 /* Empty CONSTRUCTOR. */
2075 if (pd.size >= needed_len * BITS_PER_UNIT)
2076 len = needed_len;
2077 else
2078 len = ROUND_UP (pd.size, BITS_PER_UNIT) / BITS_PER_UNIT;
2079 memset (this_buffer, 0, len);
2081 else
2083 len = native_encode_expr (pd.rhs, this_buffer, bufsize,
2084 MAX (0, -pd.offset) / BITS_PER_UNIT);
2085 if (len <= 0
2086 || len < (ROUND_UP (pd.size, BITS_PER_UNIT) / BITS_PER_UNIT
2087 - MAX (0, -pd.offset) / BITS_PER_UNIT))
2089 if (dump_file && (dump_flags & TDF_DETAILS))
2090 fprintf (dump_file, "Failed to encode %u "
2091 "partial definitions\n", ndefs);
2092 return (void *)-1;
2096 unsigned char *p = buffer;
2097 HOST_WIDE_INT size = pd.size;
2098 if (pd.offset < 0)
2099 size -= ROUND_DOWN (-pd.offset, BITS_PER_UNIT);
2100 this_buffer[len] = 0;
2101 if (BYTES_BIG_ENDIAN)
2103 /* LSB of this_buffer[len - 1] byte should be at
2104 pd.offset + pd.size - 1 bits in buffer. */
2105 amnt = ((unsigned HOST_WIDE_INT) pd.offset
2106 + pd.size) % BITS_PER_UNIT;
2107 if (amnt)
2108 shift_bytes_in_array_right (this_buffer, len + 1, amnt);
2109 unsigned char *q = this_buffer;
2110 unsigned int off = 0;
2111 if (pd.offset >= 0)
2113 unsigned int msk;
2114 off = pd.offset / BITS_PER_UNIT;
2115 gcc_assert (off < needed_len);
2116 p = buffer + off;
2117 if (size <= amnt)
2119 msk = ((1 << size) - 1) << (BITS_PER_UNIT - amnt);
2120 *p = (*p & ~msk) | (this_buffer[len] & msk);
2121 size = 0;
2123 else
2125 if (TREE_CODE (pd.rhs) != CONSTRUCTOR)
2126 q = (this_buffer + len
2127 - (ROUND_UP (size - amnt, BITS_PER_UNIT)
2128 / BITS_PER_UNIT));
2129 if (pd.offset % BITS_PER_UNIT)
2131 msk = -1U << (BITS_PER_UNIT
2132 - (pd.offset % BITS_PER_UNIT));
2133 *p = (*p & msk) | (*q & ~msk);
2134 p++;
2135 q++;
2136 off++;
2137 size -= BITS_PER_UNIT - (pd.offset % BITS_PER_UNIT);
2138 gcc_assert (size >= 0);
2142 else if (TREE_CODE (pd.rhs) != CONSTRUCTOR)
2144 q = (this_buffer + len
2145 - (ROUND_UP (size - amnt, BITS_PER_UNIT)
2146 / BITS_PER_UNIT));
2147 if (pd.offset % BITS_PER_UNIT)
2149 q++;
2150 size -= BITS_PER_UNIT - ((unsigned HOST_WIDE_INT) pd.offset
2151 % BITS_PER_UNIT);
2152 gcc_assert (size >= 0);
2155 if ((unsigned HOST_WIDE_INT) size / BITS_PER_UNIT + off
2156 > needed_len)
2157 size = (needed_len - off) * BITS_PER_UNIT;
2158 memcpy (p, q, size / BITS_PER_UNIT);
2159 if (size % BITS_PER_UNIT)
2161 unsigned int msk
2162 = -1U << (BITS_PER_UNIT - (size % BITS_PER_UNIT));
2163 p += size / BITS_PER_UNIT;
2164 q += size / BITS_PER_UNIT;
2165 *p = (*q & msk) | (*p & ~msk);
2168 else
2170 if (pd.offset >= 0)
2172 /* LSB of this_buffer[0] byte should be at pd.offset bits
2173 in buffer. */
2174 unsigned int msk;
2175 size = MIN (size, (HOST_WIDE_INT) needed_len * BITS_PER_UNIT);
2176 amnt = pd.offset % BITS_PER_UNIT;
2177 if (amnt)
2178 shift_bytes_in_array_left (this_buffer, len + 1, amnt);
2179 unsigned int off = pd.offset / BITS_PER_UNIT;
2180 gcc_assert (off < needed_len);
2181 size = MIN (size,
2182 (HOST_WIDE_INT) (needed_len - off) * BITS_PER_UNIT);
2183 p = buffer + off;
2184 if (amnt + size < BITS_PER_UNIT)
2186 /* Low amnt bits come from *p, then size bits
2187 from this_buffer[0] and the remaining again from
2188 *p. */
2189 msk = ((1 << size) - 1) << amnt;
2190 *p = (*p & ~msk) | (this_buffer[0] & msk);
2191 size = 0;
2193 else if (amnt)
2195 msk = -1U << amnt;
2196 *p = (*p & ~msk) | (this_buffer[0] & msk);
2197 p++;
2198 size -= (BITS_PER_UNIT - amnt);
2201 else
2203 amnt = (unsigned HOST_WIDE_INT) pd.offset % BITS_PER_UNIT;
2204 if (amnt)
2205 size -= BITS_PER_UNIT - amnt;
2206 size = MIN (size, (HOST_WIDE_INT) needed_len * BITS_PER_UNIT);
2207 if (amnt)
2208 shift_bytes_in_array_left (this_buffer, len + 1, amnt);
2210 memcpy (p, this_buffer + (amnt != 0), size / BITS_PER_UNIT);
2211 p += size / BITS_PER_UNIT;
2212 if (size % BITS_PER_UNIT)
2214 unsigned int msk = -1U << (size % BITS_PER_UNIT);
2215 *p = (this_buffer[(amnt != 0) + size / BITS_PER_UNIT]
2216 & ~msk) | (*p & msk);
2221 tree type = vr->type;
2222 /* Make sure to interpret in a type that has a range covering the whole
2223 access size. */
2224 if (INTEGRAL_TYPE_P (vr->type) && maxsizei != TYPE_PRECISION (vr->type))
2225 type = build_nonstandard_integer_type (maxsizei, TYPE_UNSIGNED (type));
2226 tree val;
2227 if (BYTES_BIG_ENDIAN)
2229 unsigned sz = needed_len;
2230 if (maxsizei % BITS_PER_UNIT)
2231 shift_bytes_in_array_right (buffer, needed_len,
2232 BITS_PER_UNIT
2233 - (maxsizei % BITS_PER_UNIT));
2234 if (INTEGRAL_TYPE_P (type))
2235 sz = GET_MODE_SIZE (SCALAR_INT_TYPE_MODE (type));
2236 if (sz > needed_len)
2238 memcpy (this_buffer + (sz - needed_len), buffer, needed_len);
2239 val = native_interpret_expr (type, this_buffer, sz);
2241 else
2242 val = native_interpret_expr (type, buffer, needed_len);
2244 else
2245 val = native_interpret_expr (type, buffer, bufsize);
2246 /* If we chop off bits because the types precision doesn't match the memory
2247 access size this is ok when optimizing reads but not when called from
2248 the DSE code during elimination. */
2249 if (val && type != vr->type)
2251 if (! int_fits_type_p (val, vr->type))
2252 val = NULL_TREE;
2253 else
2254 val = fold_convert (vr->type, val);
2257 if (val)
2259 if (dump_file && (dump_flags & TDF_DETAILS))
2260 fprintf (dump_file,
2261 "Successfully combined %u partial definitions\n", ndefs);
2262 /* We are using the alias-set of the first store we encounter which
2263 should be appropriate here. */
2264 return finish (first_set, first_base_set, val);
2266 else
2268 if (dump_file && (dump_flags & TDF_DETAILS))
2269 fprintf (dump_file,
2270 "Failed to interpret %u encoded partial definitions\n", ndefs);
2271 return (void *)-1;
2275 /* Callback for walk_non_aliased_vuses. Adjusts the vn_reference_t VR_
2276 with the current VUSE and performs the expression lookup. */
2278 static void *
2279 vn_reference_lookup_2 (ao_ref *op ATTRIBUTE_UNUSED, tree vuse, void *data_)
2281 vn_walk_cb_data *data = (vn_walk_cb_data *)data_;
2282 vn_reference_t vr = data->vr;
2283 vn_reference_s **slot;
2284 hashval_t hash;
2286 /* If we have partial definitions recorded we have to go through
2287 vn_reference_lookup_3. */
2288 if (!data->partial_defs.is_empty ())
2289 return NULL;
2291 if (data->last_vuse_ptr)
2293 *data->last_vuse_ptr = vuse;
2294 data->last_vuse = vuse;
2297 /* Fixup vuse and hash. */
2298 if (vr->vuse)
2299 vr->hashcode = vr->hashcode - SSA_NAME_VERSION (vr->vuse);
2300 vr->vuse = vuse_ssa_val (vuse);
2301 if (vr->vuse)
2302 vr->hashcode = vr->hashcode + SSA_NAME_VERSION (vr->vuse);
2304 hash = vr->hashcode;
2305 slot = valid_info->references->find_slot_with_hash (vr, hash, NO_INSERT);
2306 if (slot)
2308 if ((*slot)->result && data->saved_operands.exists ())
2309 return data->finish (vr->set, vr->base_set, (*slot)->result);
2310 return *slot;
2313 return NULL;
2316 /* Lookup an existing or insert a new vn_reference entry into the
2317 value table for the VUSE, SET, TYPE, OPERANDS reference which
2318 has the value VALUE which is either a constant or an SSA name. */
2320 static vn_reference_t
2321 vn_reference_lookup_or_insert_for_pieces (tree vuse,
2322 alias_set_type set,
2323 alias_set_type base_set,
2324 tree type,
2325 vec<vn_reference_op_s,
2326 va_heap> operands,
2327 tree value)
2329 vn_reference_s vr1;
2330 vn_reference_t result;
2331 unsigned value_id;
2332 vr1.vuse = vuse ? SSA_VAL (vuse) : NULL_TREE;
2333 vr1.operands = operands;
2334 vr1.type = type;
2335 vr1.set = set;
2336 vr1.base_set = base_set;
2337 vr1.hashcode = vn_reference_compute_hash (&vr1);
2338 if (vn_reference_lookup_1 (&vr1, &result))
2339 return result;
2340 if (TREE_CODE (value) == SSA_NAME)
2341 value_id = VN_INFO (value)->value_id;
2342 else
2343 value_id = get_or_alloc_constant_value_id (value);
2344 return vn_reference_insert_pieces (vuse, set, base_set, type,
2345 operands.copy (), value, value_id);
2348 /* Return a value-number for RCODE OPS... either by looking up an existing
2349 value-number for the possibly simplified result or by inserting the
2350 operation if INSERT is true. If SIMPLIFY is false, return a value
2351 number for the unsimplified expression. */
2353 static tree
2354 vn_nary_build_or_lookup_1 (gimple_match_op *res_op, bool insert,
2355 bool simplify)
2357 tree result = NULL_TREE;
2358 /* We will be creating a value number for
2359 RCODE (OPS...).
2360 So first simplify and lookup this expression to see if it
2361 is already available. */
2362 /* For simplification valueize. */
2363 unsigned i = 0;
2364 if (simplify)
2365 for (i = 0; i < res_op->num_ops; ++i)
2366 if (TREE_CODE (res_op->ops[i]) == SSA_NAME)
2368 tree tem = vn_valueize (res_op->ops[i]);
2369 if (!tem)
2370 break;
2371 res_op->ops[i] = tem;
2373 /* If valueization of an operand fails (it is not available), skip
2374 simplification. */
2375 bool res = false;
2376 if (i == res_op->num_ops)
2378 mprts_hook = vn_lookup_simplify_result;
2379 res = res_op->resimplify (NULL, vn_valueize);
2380 mprts_hook = NULL;
2382 gimple *new_stmt = NULL;
2383 if (res
2384 && gimple_simplified_result_is_gimple_val (res_op))
2386 /* The expression is already available. */
2387 result = res_op->ops[0];
2388 /* Valueize it, simplification returns sth in AVAIL only. */
2389 if (TREE_CODE (result) == SSA_NAME)
2390 result = SSA_VAL (result);
2392 else
2394 tree val = vn_lookup_simplify_result (res_op);
2395 if (!val && insert)
2397 gimple_seq stmts = NULL;
2398 result = maybe_push_res_to_seq (res_op, &stmts);
2399 if (result)
2401 gcc_assert (gimple_seq_singleton_p (stmts));
2402 new_stmt = gimple_seq_first_stmt (stmts);
2405 else
2406 /* The expression is already available. */
2407 result = val;
2409 if (new_stmt)
2411 /* The expression is not yet available, value-number lhs to
2412 the new SSA_NAME we created. */
2413 /* Initialize value-number information properly. */
2414 vn_ssa_aux_t result_info = VN_INFO (result);
2415 result_info->valnum = result;
2416 result_info->value_id = get_next_value_id ();
2417 result_info->visited = 1;
2418 gimple_seq_add_stmt_without_update (&VN_INFO (result)->expr,
2419 new_stmt);
2420 result_info->needs_insertion = true;
2421 /* ??? PRE phi-translation inserts NARYs without corresponding
2422 SSA name result. Re-use those but set their result according
2423 to the stmt we just built. */
2424 vn_nary_op_t nary = NULL;
2425 vn_nary_op_lookup_stmt (new_stmt, &nary);
2426 if (nary)
2428 gcc_assert (! nary->predicated_values && nary->u.result == NULL_TREE);
2429 nary->u.result = gimple_assign_lhs (new_stmt);
2431 /* As all "inserted" statements are singleton SCCs, insert
2432 to the valid table. This is strictly needed to
2433 avoid re-generating new value SSA_NAMEs for the same
2434 expression during SCC iteration over and over (the
2435 optimistic table gets cleared after each iteration).
2436 We do not need to insert into the optimistic table, as
2437 lookups there will fall back to the valid table. */
2438 else
2440 unsigned int length = vn_nary_length_from_stmt (new_stmt);
2441 vn_nary_op_t vno1
2442 = alloc_vn_nary_op_noinit (length, &vn_tables_insert_obstack);
2443 vno1->value_id = result_info->value_id;
2444 vno1->length = length;
2445 vno1->predicated_values = 0;
2446 vno1->u.result = result;
2447 init_vn_nary_op_from_stmt (vno1, as_a <gassign *> (new_stmt));
2448 vn_nary_op_insert_into (vno1, valid_info->nary);
2449 /* Also do not link it into the undo chain. */
2450 last_inserted_nary = vno1->next;
2451 vno1->next = (vn_nary_op_t)(void *)-1;
2453 if (dump_file && (dump_flags & TDF_DETAILS))
2455 fprintf (dump_file, "Inserting name ");
2456 print_generic_expr (dump_file, result);
2457 fprintf (dump_file, " for expression ");
2458 print_gimple_expr (dump_file, new_stmt, 0, TDF_SLIM);
2459 fprintf (dump_file, "\n");
2462 return result;
2465 /* Return a value-number for RCODE OPS... either by looking up an existing
2466 value-number for the simplified result or by inserting the operation. */
2468 static tree
2469 vn_nary_build_or_lookup (gimple_match_op *res_op)
2471 return vn_nary_build_or_lookup_1 (res_op, true, true);
2474 /* Try to simplify the expression RCODE OPS... of type TYPE and return
2475 its value if present. */
2477 tree
2478 vn_nary_simplify (vn_nary_op_t nary)
2480 if (nary->length > gimple_match_op::MAX_NUM_OPS)
2481 return NULL_TREE;
2482 gimple_match_op op (gimple_match_cond::UNCOND, nary->opcode,
2483 nary->type, nary->length);
2484 memcpy (op.ops, nary->op, sizeof (tree) * nary->length);
2485 return vn_nary_build_or_lookup_1 (&op, false, true);
2488 /* Elimination engine. */
2490 class eliminate_dom_walker : public dom_walker
2492 public:
2493 eliminate_dom_walker (cdi_direction, bitmap);
2494 ~eliminate_dom_walker ();
2496 virtual edge before_dom_children (basic_block);
2497 virtual void after_dom_children (basic_block);
2499 virtual tree eliminate_avail (basic_block, tree op);
2500 virtual void eliminate_push_avail (basic_block, tree op);
2501 tree eliminate_insert (basic_block, gimple_stmt_iterator *gsi, tree val);
2503 void eliminate_stmt (basic_block, gimple_stmt_iterator *);
2505 unsigned eliminate_cleanup (bool region_p = false);
2507 bool do_pre;
2508 unsigned int el_todo;
2509 unsigned int eliminations;
2510 unsigned int insertions;
2512 /* SSA names that had their defs inserted by PRE if do_pre. */
2513 bitmap inserted_exprs;
2515 /* Blocks with statements that have had their EH properties changed. */
2516 bitmap need_eh_cleanup;
2518 /* Blocks with statements that have had their AB properties changed. */
2519 bitmap need_ab_cleanup;
2521 /* Local state for the eliminate domwalk. */
2522 auto_vec<gimple *> to_remove;
2523 auto_vec<gimple *> to_fixup;
2524 auto_vec<tree> avail;
2525 auto_vec<tree> avail_stack;
2528 /* Adaptor to the elimination engine using RPO availability. */
2530 class rpo_elim : public eliminate_dom_walker
2532 public:
2533 rpo_elim(basic_block entry_)
2534 : eliminate_dom_walker (CDI_DOMINATORS, NULL), entry (entry_),
2535 m_avail_freelist (NULL) {}
2537 virtual tree eliminate_avail (basic_block, tree op);
2539 virtual void eliminate_push_avail (basic_block, tree);
2541 basic_block entry;
2542 /* Freelist of avail entries which are allocated from the vn_ssa_aux
2543 obstack. */
2544 vn_avail *m_avail_freelist;
2547 /* Global RPO state for access from hooks. */
2548 static eliminate_dom_walker *rpo_avail;
2549 basic_block vn_context_bb;
2551 /* Return true if BASE1 and BASE2 can be adjusted so they have the
2552 same address and adjust *OFFSET1 and *OFFSET2 accordingly.
2553 Otherwise return false. */
2555 static bool
2556 adjust_offsets_for_equal_base_address (tree base1, poly_int64 *offset1,
2557 tree base2, poly_int64 *offset2)
2559 poly_int64 soff;
2560 if (TREE_CODE (base1) == MEM_REF
2561 && TREE_CODE (base2) == MEM_REF)
2563 if (mem_ref_offset (base1).to_shwi (&soff))
2565 base1 = TREE_OPERAND (base1, 0);
2566 *offset1 += soff * BITS_PER_UNIT;
2568 if (mem_ref_offset (base2).to_shwi (&soff))
2570 base2 = TREE_OPERAND (base2, 0);
2571 *offset2 += soff * BITS_PER_UNIT;
2573 return operand_equal_p (base1, base2, 0);
2575 return operand_equal_p (base1, base2, OEP_ADDRESS_OF);
2578 /* Callback for walk_non_aliased_vuses. Tries to perform a lookup
2579 from the statement defining VUSE and if not successful tries to
2580 translate *REFP and VR_ through an aggregate copy at the definition
2581 of VUSE. If *DISAMBIGUATE_ONLY is true then do not perform translation
2582 of *REF and *VR. If only disambiguation was performed then
2583 *DISAMBIGUATE_ONLY is set to true. */
2585 static void *
2586 vn_reference_lookup_3 (ao_ref *ref, tree vuse, void *data_,
2587 translate_flags *disambiguate_only)
2589 vn_walk_cb_data *data = (vn_walk_cb_data *)data_;
2590 vn_reference_t vr = data->vr;
2591 gimple *def_stmt = SSA_NAME_DEF_STMT (vuse);
2592 tree base = ao_ref_base (ref);
2593 HOST_WIDE_INT offseti = 0, maxsizei, sizei = 0;
2594 static vec<vn_reference_op_s> lhs_ops;
2595 ao_ref lhs_ref;
2596 bool lhs_ref_ok = false;
2597 poly_int64 copy_size;
2599 /* First try to disambiguate after value-replacing in the definitions LHS. */
2600 if (is_gimple_assign (def_stmt))
2602 tree lhs = gimple_assign_lhs (def_stmt);
2603 bool valueized_anything = false;
2604 /* Avoid re-allocation overhead. */
2605 lhs_ops.truncate (0);
2606 basic_block saved_rpo_bb = vn_context_bb;
2607 vn_context_bb = gimple_bb (def_stmt);
2608 if (*disambiguate_only <= TR_VALUEIZE_AND_DISAMBIGUATE)
2610 copy_reference_ops_from_ref (lhs, &lhs_ops);
2611 valueize_refs_1 (&lhs_ops, &valueized_anything, true);
2613 vn_context_bb = saved_rpo_bb;
2614 ao_ref_init (&lhs_ref, lhs);
2615 lhs_ref_ok = true;
2616 if (valueized_anything
2617 && ao_ref_init_from_vn_reference
2618 (&lhs_ref, ao_ref_alias_set (&lhs_ref),
2619 ao_ref_base_alias_set (&lhs_ref), TREE_TYPE (lhs), lhs_ops)
2620 && !refs_may_alias_p_1 (ref, &lhs_ref, data->tbaa_p))
2622 *disambiguate_only = TR_VALUEIZE_AND_DISAMBIGUATE;
2623 return NULL;
2626 /* When the def is a CLOBBER we can optimistically disambiguate
2627 against it since any overlap it would be undefined behavior.
2628 Avoid this for obvious must aliases to save compile-time though.
2629 We also may not do this when the query is used for redundant
2630 store removal. */
2631 if (!data->redundant_store_removal_p
2632 && gimple_clobber_p (def_stmt)
2633 && !operand_equal_p (ao_ref_base (&lhs_ref), base, OEP_ADDRESS_OF))
2635 *disambiguate_only = TR_DISAMBIGUATE;
2636 return NULL;
2639 /* Besides valueizing the LHS we can also use access-path based
2640 disambiguation on the original non-valueized ref. */
2641 if (!ref->ref
2642 && lhs_ref_ok
2643 && data->orig_ref.ref)
2645 /* We want to use the non-valueized LHS for this, but avoid redundant
2646 work. */
2647 ao_ref *lref = &lhs_ref;
2648 ao_ref lref_alt;
2649 if (valueized_anything)
2651 ao_ref_init (&lref_alt, lhs);
2652 lref = &lref_alt;
2654 if (!refs_may_alias_p_1 (&data->orig_ref, lref, data->tbaa_p))
2656 *disambiguate_only = (valueized_anything
2657 ? TR_VALUEIZE_AND_DISAMBIGUATE
2658 : TR_DISAMBIGUATE);
2659 return NULL;
2663 /* If we reach a clobbering statement try to skip it and see if
2664 we find a VN result with exactly the same value as the
2665 possible clobber. In this case we can ignore the clobber
2666 and return the found value. */
2667 if (is_gimple_reg_type (TREE_TYPE (lhs))
2668 && types_compatible_p (TREE_TYPE (lhs), vr->type)
2669 && (ref->ref || data->orig_ref.ref))
2671 tree *saved_last_vuse_ptr = data->last_vuse_ptr;
2672 /* Do not update last_vuse_ptr in vn_reference_lookup_2. */
2673 data->last_vuse_ptr = NULL;
2674 tree saved_vuse = vr->vuse;
2675 hashval_t saved_hashcode = vr->hashcode;
2676 void *res = vn_reference_lookup_2 (ref, gimple_vuse (def_stmt), data);
2677 /* Need to restore vr->vuse and vr->hashcode. */
2678 vr->vuse = saved_vuse;
2679 vr->hashcode = saved_hashcode;
2680 data->last_vuse_ptr = saved_last_vuse_ptr;
2681 if (res && res != (void *)-1)
2683 vn_reference_t vnresult = (vn_reference_t) res;
2684 tree rhs = gimple_assign_rhs1 (def_stmt);
2685 if (TREE_CODE (rhs) == SSA_NAME)
2686 rhs = SSA_VAL (rhs);
2687 if (vnresult->result
2688 && operand_equal_p (vnresult->result, rhs, 0)
2689 /* We have to honor our promise about union type punning
2690 and also support arbitrary overlaps with
2691 -fno-strict-aliasing. So simply resort to alignment to
2692 rule out overlaps. Do this check last because it is
2693 quite expensive compared to the hash-lookup above. */
2694 && multiple_p (get_object_alignment
2695 (ref->ref ? ref->ref : data->orig_ref.ref),
2696 ref->size)
2697 && multiple_p (get_object_alignment (lhs), ref->size))
2698 return res;
2702 else if (*disambiguate_only <= TR_VALUEIZE_AND_DISAMBIGUATE
2703 && gimple_call_builtin_p (def_stmt, BUILT_IN_NORMAL)
2704 && gimple_call_num_args (def_stmt) <= 4)
2706 /* For builtin calls valueize its arguments and call the
2707 alias oracle again. Valueization may improve points-to
2708 info of pointers and constify size and position arguments.
2709 Originally this was motivated by PR61034 which has
2710 conditional calls to free falsely clobbering ref because
2711 of imprecise points-to info of the argument. */
2712 tree oldargs[4];
2713 bool valueized_anything = false;
2714 for (unsigned i = 0; i < gimple_call_num_args (def_stmt); ++i)
2716 oldargs[i] = gimple_call_arg (def_stmt, i);
2717 tree val = vn_valueize (oldargs[i]);
2718 if (val != oldargs[i])
2720 gimple_call_set_arg (def_stmt, i, val);
2721 valueized_anything = true;
2724 if (valueized_anything)
2726 bool res = call_may_clobber_ref_p_1 (as_a <gcall *> (def_stmt),
2727 ref, data->tbaa_p);
2728 for (unsigned i = 0; i < gimple_call_num_args (def_stmt); ++i)
2729 gimple_call_set_arg (def_stmt, i, oldargs[i]);
2730 if (!res)
2732 *disambiguate_only = TR_VALUEIZE_AND_DISAMBIGUATE;
2733 return NULL;
2738 if (*disambiguate_only > TR_TRANSLATE)
2739 return (void *)-1;
2741 /* If we cannot constrain the size of the reference we cannot
2742 test if anything kills it. */
2743 if (!ref->max_size_known_p ())
2744 return (void *)-1;
2746 poly_int64 offset = ref->offset;
2747 poly_int64 maxsize = ref->max_size;
2749 /* def_stmt may-defs *ref. See if we can derive a value for *ref
2750 from that definition.
2751 1) Memset. */
2752 if (is_gimple_reg_type (vr->type)
2753 && (gimple_call_builtin_p (def_stmt, BUILT_IN_MEMSET)
2754 || gimple_call_builtin_p (def_stmt, BUILT_IN_MEMSET_CHK))
2755 && (integer_zerop (gimple_call_arg (def_stmt, 1))
2756 || ((TREE_CODE (gimple_call_arg (def_stmt, 1)) == INTEGER_CST
2757 || (INTEGRAL_TYPE_P (vr->type) && known_eq (ref->size, 8)))
2758 && CHAR_BIT == 8
2759 && BITS_PER_UNIT == 8
2760 && BYTES_BIG_ENDIAN == WORDS_BIG_ENDIAN
2761 && offset.is_constant (&offseti)
2762 && ref->size.is_constant (&sizei)
2763 && (offseti % BITS_PER_UNIT == 0
2764 || TREE_CODE (gimple_call_arg (def_stmt, 1)) == INTEGER_CST)))
2765 && (poly_int_tree_p (gimple_call_arg (def_stmt, 2))
2766 || (TREE_CODE (gimple_call_arg (def_stmt, 2)) == SSA_NAME
2767 && poly_int_tree_p (SSA_VAL (gimple_call_arg (def_stmt, 2)))))
2768 && (TREE_CODE (gimple_call_arg (def_stmt, 0)) == ADDR_EXPR
2769 || TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME))
2771 tree base2;
2772 poly_int64 offset2, size2, maxsize2;
2773 bool reverse;
2774 tree ref2 = gimple_call_arg (def_stmt, 0);
2775 if (TREE_CODE (ref2) == SSA_NAME)
2777 ref2 = SSA_VAL (ref2);
2778 if (TREE_CODE (ref2) == SSA_NAME
2779 && (TREE_CODE (base) != MEM_REF
2780 || TREE_OPERAND (base, 0) != ref2))
2782 gimple *def_stmt = SSA_NAME_DEF_STMT (ref2);
2783 if (gimple_assign_single_p (def_stmt)
2784 && gimple_assign_rhs_code (def_stmt) == ADDR_EXPR)
2785 ref2 = gimple_assign_rhs1 (def_stmt);
2788 if (TREE_CODE (ref2) == ADDR_EXPR)
2790 ref2 = TREE_OPERAND (ref2, 0);
2791 base2 = get_ref_base_and_extent (ref2, &offset2, &size2, &maxsize2,
2792 &reverse);
2793 if (!known_size_p (maxsize2)
2794 || !known_eq (maxsize2, size2)
2795 || !operand_equal_p (base, base2, OEP_ADDRESS_OF))
2796 return (void *)-1;
2798 else if (TREE_CODE (ref2) == SSA_NAME)
2800 poly_int64 soff;
2801 if (TREE_CODE (base) != MEM_REF
2802 || !(mem_ref_offset (base)
2803 << LOG2_BITS_PER_UNIT).to_shwi (&soff))
2804 return (void *)-1;
2805 offset += soff;
2806 offset2 = 0;
2807 if (TREE_OPERAND (base, 0) != ref2)
2809 gimple *def = SSA_NAME_DEF_STMT (ref2);
2810 if (is_gimple_assign (def)
2811 && gimple_assign_rhs_code (def) == POINTER_PLUS_EXPR
2812 && gimple_assign_rhs1 (def) == TREE_OPERAND (base, 0)
2813 && poly_int_tree_p (gimple_assign_rhs2 (def)))
2815 tree rhs2 = gimple_assign_rhs2 (def);
2816 if (!(poly_offset_int::from (wi::to_poly_wide (rhs2),
2817 SIGNED)
2818 << LOG2_BITS_PER_UNIT).to_shwi (&offset2))
2819 return (void *)-1;
2820 ref2 = gimple_assign_rhs1 (def);
2821 if (TREE_CODE (ref2) == SSA_NAME)
2822 ref2 = SSA_VAL (ref2);
2824 else
2825 return (void *)-1;
2828 else
2829 return (void *)-1;
2830 tree len = gimple_call_arg (def_stmt, 2);
2831 HOST_WIDE_INT leni, offset2i;
2832 if (TREE_CODE (len) == SSA_NAME)
2833 len = SSA_VAL (len);
2834 /* Sometimes the above trickery is smarter than alias analysis. Take
2835 advantage of that. */
2836 if (!ranges_maybe_overlap_p (offset, maxsize, offset2,
2837 (wi::to_poly_offset (len)
2838 << LOG2_BITS_PER_UNIT)))
2839 return NULL;
2840 if (data->partial_defs.is_empty ()
2841 && known_subrange_p (offset, maxsize, offset2,
2842 wi::to_poly_offset (len) << LOG2_BITS_PER_UNIT))
2844 tree val;
2845 if (integer_zerop (gimple_call_arg (def_stmt, 1)))
2846 val = build_zero_cst (vr->type);
2847 else if (INTEGRAL_TYPE_P (vr->type)
2848 && known_eq (ref->size, 8)
2849 && offseti % BITS_PER_UNIT == 0)
2851 gimple_match_op res_op (gimple_match_cond::UNCOND, NOP_EXPR,
2852 vr->type, gimple_call_arg (def_stmt, 1));
2853 val = vn_nary_build_or_lookup (&res_op);
2854 if (!val
2855 || (TREE_CODE (val) == SSA_NAME
2856 && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (val)))
2857 return (void *)-1;
2859 else
2861 unsigned buflen = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (vr->type)) + 1;
2862 if (INTEGRAL_TYPE_P (vr->type))
2863 buflen = GET_MODE_SIZE (SCALAR_INT_TYPE_MODE (vr->type)) + 1;
2864 unsigned char *buf = XALLOCAVEC (unsigned char, buflen);
2865 memset (buf, TREE_INT_CST_LOW (gimple_call_arg (def_stmt, 1)),
2866 buflen);
2867 if (BYTES_BIG_ENDIAN)
2869 unsigned int amnt
2870 = (((unsigned HOST_WIDE_INT) offseti + sizei)
2871 % BITS_PER_UNIT);
2872 if (amnt)
2874 shift_bytes_in_array_right (buf, buflen,
2875 BITS_PER_UNIT - amnt);
2876 buf++;
2877 buflen--;
2880 else if (offseti % BITS_PER_UNIT != 0)
2882 unsigned int amnt
2883 = BITS_PER_UNIT - ((unsigned HOST_WIDE_INT) offseti
2884 % BITS_PER_UNIT);
2885 shift_bytes_in_array_left (buf, buflen, amnt);
2886 buf++;
2887 buflen--;
2889 val = native_interpret_expr (vr->type, buf, buflen);
2890 if (!val)
2891 return (void *)-1;
2893 return data->finish (0, 0, val);
2895 /* For now handle clearing memory with partial defs. */
2896 else if (known_eq (ref->size, maxsize)
2897 && integer_zerop (gimple_call_arg (def_stmt, 1))
2898 && tree_fits_poly_int64_p (len)
2899 && tree_to_poly_int64 (len).is_constant (&leni)
2900 && leni <= INTTYPE_MAXIMUM (HOST_WIDE_INT) / BITS_PER_UNIT
2901 && offset.is_constant (&offseti)
2902 && offset2.is_constant (&offset2i)
2903 && maxsize.is_constant (&maxsizei)
2904 && ranges_known_overlap_p (offseti, maxsizei, offset2i,
2905 leni << LOG2_BITS_PER_UNIT))
2907 pd_data pd;
2908 pd.rhs = build_constructor (NULL_TREE, NULL);
2909 pd.offset = offset2i;
2910 pd.size = leni << LOG2_BITS_PER_UNIT;
2911 return data->push_partial_def (pd, 0, 0, offseti, maxsizei);
2915 /* 2) Assignment from an empty CONSTRUCTOR. */
2916 else if (is_gimple_reg_type (vr->type)
2917 && gimple_assign_single_p (def_stmt)
2918 && gimple_assign_rhs_code (def_stmt) == CONSTRUCTOR
2919 && CONSTRUCTOR_NELTS (gimple_assign_rhs1 (def_stmt)) == 0)
2921 tree base2;
2922 poly_int64 offset2, size2, maxsize2;
2923 HOST_WIDE_INT offset2i, size2i;
2924 gcc_assert (lhs_ref_ok);
2925 base2 = ao_ref_base (&lhs_ref);
2926 offset2 = lhs_ref.offset;
2927 size2 = lhs_ref.size;
2928 maxsize2 = lhs_ref.max_size;
2929 if (known_size_p (maxsize2)
2930 && known_eq (maxsize2, size2)
2931 && adjust_offsets_for_equal_base_address (base, &offset,
2932 base2, &offset2))
2934 if (data->partial_defs.is_empty ()
2935 && known_subrange_p (offset, maxsize, offset2, size2))
2937 /* While technically undefined behavior do not optimize
2938 a full read from a clobber. */
2939 if (gimple_clobber_p (def_stmt))
2940 return (void *)-1;
2941 tree val = build_zero_cst (vr->type);
2942 return data->finish (ao_ref_alias_set (&lhs_ref),
2943 ao_ref_base_alias_set (&lhs_ref), val);
2945 else if (known_eq (ref->size, maxsize)
2946 && maxsize.is_constant (&maxsizei)
2947 && offset.is_constant (&offseti)
2948 && offset2.is_constant (&offset2i)
2949 && size2.is_constant (&size2i)
2950 && ranges_known_overlap_p (offseti, maxsizei,
2951 offset2i, size2i))
2953 /* Let clobbers be consumed by the partial-def tracker
2954 which can choose to ignore them if they are shadowed
2955 by a later def. */
2956 pd_data pd;
2957 pd.rhs = gimple_assign_rhs1 (def_stmt);
2958 pd.offset = offset2i;
2959 pd.size = size2i;
2960 return data->push_partial_def (pd, ao_ref_alias_set (&lhs_ref),
2961 ao_ref_base_alias_set (&lhs_ref),
2962 offseti, maxsizei);
2967 /* 3) Assignment from a constant. We can use folds native encode/interpret
2968 routines to extract the assigned bits. */
2969 else if (known_eq (ref->size, maxsize)
2970 && is_gimple_reg_type (vr->type)
2971 && !reverse_storage_order_for_component_p (vr->operands)
2972 && !contains_storage_order_barrier_p (vr->operands)
2973 && gimple_assign_single_p (def_stmt)
2974 && CHAR_BIT == 8
2975 && BITS_PER_UNIT == 8
2976 && BYTES_BIG_ENDIAN == WORDS_BIG_ENDIAN
2977 /* native_encode and native_decode operate on arrays of bytes
2978 and so fundamentally need a compile-time size and offset. */
2979 && maxsize.is_constant (&maxsizei)
2980 && offset.is_constant (&offseti)
2981 && (is_gimple_min_invariant (gimple_assign_rhs1 (def_stmt))
2982 || (TREE_CODE (gimple_assign_rhs1 (def_stmt)) == SSA_NAME
2983 && is_gimple_min_invariant (SSA_VAL (gimple_assign_rhs1 (def_stmt))))))
2985 tree lhs = gimple_assign_lhs (def_stmt);
2986 tree base2;
2987 poly_int64 offset2, size2, maxsize2;
2988 HOST_WIDE_INT offset2i, size2i;
2989 bool reverse;
2990 gcc_assert (lhs_ref_ok);
2991 base2 = ao_ref_base (&lhs_ref);
2992 offset2 = lhs_ref.offset;
2993 size2 = lhs_ref.size;
2994 maxsize2 = lhs_ref.max_size;
2995 reverse = reverse_storage_order_for_component_p (lhs);
2996 if (base2
2997 && !reverse
2998 && !storage_order_barrier_p (lhs)
2999 && known_eq (maxsize2, size2)
3000 && adjust_offsets_for_equal_base_address (base, &offset,
3001 base2, &offset2)
3002 && offset.is_constant (&offseti)
3003 && offset2.is_constant (&offset2i)
3004 && size2.is_constant (&size2i))
3006 if (data->partial_defs.is_empty ()
3007 && known_subrange_p (offseti, maxsizei, offset2, size2))
3009 /* We support up to 512-bit values (for V8DFmode). */
3010 unsigned char buffer[65];
3011 int len;
3013 tree rhs = gimple_assign_rhs1 (def_stmt);
3014 if (TREE_CODE (rhs) == SSA_NAME)
3015 rhs = SSA_VAL (rhs);
3016 len = native_encode_expr (rhs,
3017 buffer, sizeof (buffer) - 1,
3018 (offseti - offset2i) / BITS_PER_UNIT);
3019 if (len > 0 && len * BITS_PER_UNIT >= maxsizei)
3021 tree type = vr->type;
3022 unsigned char *buf = buffer;
3023 unsigned int amnt = 0;
3024 /* Make sure to interpret in a type that has a range
3025 covering the whole access size. */
3026 if (INTEGRAL_TYPE_P (vr->type)
3027 && maxsizei != TYPE_PRECISION (vr->type))
3028 type = build_nonstandard_integer_type (maxsizei,
3029 TYPE_UNSIGNED (type));
3030 if (BYTES_BIG_ENDIAN)
3032 /* For big-endian native_encode_expr stored the rhs
3033 such that the LSB of it is the LSB of buffer[len - 1].
3034 That bit is stored into memory at position
3035 offset2 + size2 - 1, i.e. in byte
3036 base + (offset2 + size2 - 1) / BITS_PER_UNIT.
3037 E.g. for offset2 1 and size2 14, rhs -1 and memory
3038 previously cleared that is:
3040 01111111|11111110
3041 Now, if we want to extract offset 2 and size 12 from
3042 it using native_interpret_expr (which actually works
3043 for integral bitfield types in terms of byte size of
3044 the mode), the native_encode_expr stored the value
3045 into buffer as
3046 XX111111|11111111
3047 and returned len 2 (the X bits are outside of
3048 precision).
3049 Let sz be maxsize / BITS_PER_UNIT if not extracting
3050 a bitfield, and GET_MODE_SIZE otherwise.
3051 We need to align the LSB of the value we want to
3052 extract as the LSB of buf[sz - 1].
3053 The LSB from memory we need to read is at position
3054 offset + maxsize - 1. */
3055 HOST_WIDE_INT sz = maxsizei / BITS_PER_UNIT;
3056 if (INTEGRAL_TYPE_P (type))
3057 sz = GET_MODE_SIZE (SCALAR_INT_TYPE_MODE (type));
3058 amnt = ((unsigned HOST_WIDE_INT) offset2i + size2i
3059 - offseti - maxsizei) % BITS_PER_UNIT;
3060 if (amnt)
3061 shift_bytes_in_array_right (buffer, len, amnt);
3062 amnt = ((unsigned HOST_WIDE_INT) offset2i + size2i
3063 - offseti - maxsizei - amnt) / BITS_PER_UNIT;
3064 if ((unsigned HOST_WIDE_INT) sz + amnt > (unsigned) len)
3065 len = 0;
3066 else
3068 buf = buffer + len - sz - amnt;
3069 len -= (buf - buffer);
3072 else
3074 amnt = ((unsigned HOST_WIDE_INT) offset2i
3075 - offseti) % BITS_PER_UNIT;
3076 if (amnt)
3078 buffer[len] = 0;
3079 shift_bytes_in_array_left (buffer, len + 1, amnt);
3080 buf = buffer + 1;
3083 tree val = native_interpret_expr (type, buf, len);
3084 /* If we chop off bits because the types precision doesn't
3085 match the memory access size this is ok when optimizing
3086 reads but not when called from the DSE code during
3087 elimination. */
3088 if (val
3089 && type != vr->type)
3091 if (! int_fits_type_p (val, vr->type))
3092 val = NULL_TREE;
3093 else
3094 val = fold_convert (vr->type, val);
3097 if (val)
3098 return data->finish (ao_ref_alias_set (&lhs_ref),
3099 ao_ref_base_alias_set (&lhs_ref), val);
3102 else if (ranges_known_overlap_p (offseti, maxsizei, offset2i,
3103 size2i))
3105 pd_data pd;
3106 tree rhs = gimple_assign_rhs1 (def_stmt);
3107 if (TREE_CODE (rhs) == SSA_NAME)
3108 rhs = SSA_VAL (rhs);
3109 pd.rhs = rhs;
3110 pd.offset = offset2i;
3111 pd.size = size2i;
3112 return data->push_partial_def (pd, ao_ref_alias_set (&lhs_ref),
3113 ao_ref_base_alias_set (&lhs_ref),
3114 offseti, maxsizei);
3119 /* 4) Assignment from an SSA name which definition we may be able
3120 to access pieces from or we can combine to a larger entity. */
3121 else if (known_eq (ref->size, maxsize)
3122 && is_gimple_reg_type (vr->type)
3123 && !reverse_storage_order_for_component_p (vr->operands)
3124 && !contains_storage_order_barrier_p (vr->operands)
3125 && gimple_assign_single_p (def_stmt)
3126 && TREE_CODE (gimple_assign_rhs1 (def_stmt)) == SSA_NAME)
3128 tree lhs = gimple_assign_lhs (def_stmt);
3129 tree base2;
3130 poly_int64 offset2, size2, maxsize2;
3131 HOST_WIDE_INT offset2i, size2i, offseti;
3132 bool reverse;
3133 gcc_assert (lhs_ref_ok);
3134 base2 = ao_ref_base (&lhs_ref);
3135 offset2 = lhs_ref.offset;
3136 size2 = lhs_ref.size;
3137 maxsize2 = lhs_ref.max_size;
3138 reverse = reverse_storage_order_for_component_p (lhs);
3139 tree def_rhs = gimple_assign_rhs1 (def_stmt);
3140 if (!reverse
3141 && !storage_order_barrier_p (lhs)
3142 && known_size_p (maxsize2)
3143 && known_eq (maxsize2, size2)
3144 && adjust_offsets_for_equal_base_address (base, &offset,
3145 base2, &offset2))
3147 if (data->partial_defs.is_empty ()
3148 && known_subrange_p (offset, maxsize, offset2, size2)
3149 /* ??? We can't handle bitfield precision extracts without
3150 either using an alternate type for the BIT_FIELD_REF and
3151 then doing a conversion or possibly adjusting the offset
3152 according to endianness. */
3153 && (! INTEGRAL_TYPE_P (vr->type)
3154 || known_eq (ref->size, TYPE_PRECISION (vr->type)))
3155 && multiple_p (ref->size, BITS_PER_UNIT))
3157 tree val = NULL_TREE;
3158 if (! INTEGRAL_TYPE_P (TREE_TYPE (def_rhs))
3159 || type_has_mode_precision_p (TREE_TYPE (def_rhs)))
3161 gimple_match_op op (gimple_match_cond::UNCOND,
3162 BIT_FIELD_REF, vr->type,
3163 SSA_VAL (def_rhs),
3164 bitsize_int (ref->size),
3165 bitsize_int (offset - offset2));
3166 val = vn_nary_build_or_lookup (&op);
3168 else if (known_eq (ref->size, size2))
3170 gimple_match_op op (gimple_match_cond::UNCOND,
3171 VIEW_CONVERT_EXPR, vr->type,
3172 SSA_VAL (def_rhs));
3173 val = vn_nary_build_or_lookup (&op);
3175 if (val
3176 && (TREE_CODE (val) != SSA_NAME
3177 || ! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (val)))
3178 return data->finish (ao_ref_alias_set (&lhs_ref),
3179 ao_ref_base_alias_set (&lhs_ref), val);
3181 else if (maxsize.is_constant (&maxsizei)
3182 && offset.is_constant (&offseti)
3183 && offset2.is_constant (&offset2i)
3184 && size2.is_constant (&size2i)
3185 && ranges_known_overlap_p (offset, maxsize, offset2, size2))
3187 pd_data pd;
3188 pd.rhs = SSA_VAL (def_rhs);
3189 pd.offset = offset2i;
3190 pd.size = size2i;
3191 return data->push_partial_def (pd, ao_ref_alias_set (&lhs_ref),
3192 ao_ref_base_alias_set (&lhs_ref),
3193 offseti, maxsizei);
3198 /* 5) For aggregate copies translate the reference through them if
3199 the copy kills ref. */
3200 else if (data->vn_walk_kind == VN_WALKREWRITE
3201 && gimple_assign_single_p (def_stmt)
3202 && (DECL_P (gimple_assign_rhs1 (def_stmt))
3203 || TREE_CODE (gimple_assign_rhs1 (def_stmt)) == MEM_REF
3204 || handled_component_p (gimple_assign_rhs1 (def_stmt))))
3206 tree base2;
3207 int i, j, k;
3208 auto_vec<vn_reference_op_s> rhs;
3209 vn_reference_op_t vro;
3210 ao_ref r;
3212 gcc_assert (lhs_ref_ok);
3214 /* See if the assignment kills REF. */
3215 base2 = ao_ref_base (&lhs_ref);
3216 if (!lhs_ref.max_size_known_p ()
3217 || (base != base2
3218 && (TREE_CODE (base) != MEM_REF
3219 || TREE_CODE (base2) != MEM_REF
3220 || TREE_OPERAND (base, 0) != TREE_OPERAND (base2, 0)
3221 || !tree_int_cst_equal (TREE_OPERAND (base, 1),
3222 TREE_OPERAND (base2, 1))))
3223 || !stmt_kills_ref_p (def_stmt, ref))
3224 return (void *)-1;
3226 /* Find the common base of ref and the lhs. lhs_ops already
3227 contains valueized operands for the lhs. */
3228 i = vr->operands.length () - 1;
3229 j = lhs_ops.length () - 1;
3230 while (j >= 0 && i >= 0
3231 && vn_reference_op_eq (&vr->operands[i], &lhs_ops[j]))
3233 i--;
3234 j--;
3237 /* ??? The innermost op should always be a MEM_REF and we already
3238 checked that the assignment to the lhs kills vr. Thus for
3239 aggregate copies using char[] types the vn_reference_op_eq
3240 may fail when comparing types for compatibility. But we really
3241 don't care here - further lookups with the rewritten operands
3242 will simply fail if we messed up types too badly. */
3243 poly_int64 extra_off = 0;
3244 if (j == 0 && i >= 0
3245 && lhs_ops[0].opcode == MEM_REF
3246 && maybe_ne (lhs_ops[0].off, -1))
3248 if (known_eq (lhs_ops[0].off, vr->operands[i].off))
3249 i--, j--;
3250 else if (vr->operands[i].opcode == MEM_REF
3251 && maybe_ne (vr->operands[i].off, -1))
3253 extra_off = vr->operands[i].off - lhs_ops[0].off;
3254 i--, j--;
3258 /* i now points to the first additional op.
3259 ??? LHS may not be completely contained in VR, one or more
3260 VIEW_CONVERT_EXPRs could be in its way. We could at least
3261 try handling outermost VIEW_CONVERT_EXPRs. */
3262 if (j != -1)
3263 return (void *)-1;
3265 /* Punt if the additional ops contain a storage order barrier. */
3266 for (k = i; k >= 0; k--)
3268 vro = &vr->operands[k];
3269 if (vro->opcode == VIEW_CONVERT_EXPR && vro->reverse)
3270 return (void *)-1;
3273 /* Now re-write REF to be based on the rhs of the assignment. */
3274 tree rhs1 = gimple_assign_rhs1 (def_stmt);
3275 copy_reference_ops_from_ref (rhs1, &rhs);
3277 /* Apply an extra offset to the inner MEM_REF of the RHS. */
3278 if (maybe_ne (extra_off, 0))
3280 if (rhs.length () < 2)
3281 return (void *)-1;
3282 int ix = rhs.length () - 2;
3283 if (rhs[ix].opcode != MEM_REF
3284 || known_eq (rhs[ix].off, -1))
3285 return (void *)-1;
3286 rhs[ix].off += extra_off;
3287 rhs[ix].op0 = int_const_binop (PLUS_EXPR, rhs[ix].op0,
3288 build_int_cst (TREE_TYPE (rhs[ix].op0),
3289 extra_off));
3292 /* Save the operands since we need to use the original ones for
3293 the hash entry we use. */
3294 if (!data->saved_operands.exists ())
3295 data->saved_operands = vr->operands.copy ();
3297 /* We need to pre-pend vr->operands[0..i] to rhs. */
3298 vec<vn_reference_op_s> old = vr->operands;
3299 if (i + 1 + rhs.length () > vr->operands.length ())
3300 vr->operands.safe_grow (i + 1 + rhs.length (), true);
3301 else
3302 vr->operands.truncate (i + 1 + rhs.length ());
3303 FOR_EACH_VEC_ELT (rhs, j, vro)
3304 vr->operands[i + 1 + j] = *vro;
3305 valueize_refs (&vr->operands);
3306 if (old == shared_lookup_references)
3307 shared_lookup_references = vr->operands;
3308 vr->hashcode = vn_reference_compute_hash (vr);
3310 /* Try folding the new reference to a constant. */
3311 tree val = fully_constant_vn_reference_p (vr);
3312 if (val)
3314 if (data->partial_defs.is_empty ())
3315 return data->finish (ao_ref_alias_set (&lhs_ref),
3316 ao_ref_base_alias_set (&lhs_ref), val);
3317 /* This is the only interesting case for partial-def handling
3318 coming from targets that like to gimplify init-ctors as
3319 aggregate copies from constant data like aarch64 for
3320 PR83518. */
3321 if (maxsize.is_constant (&maxsizei) && known_eq (ref->size, maxsize))
3323 pd_data pd;
3324 pd.rhs = val;
3325 pd.offset = 0;
3326 pd.size = maxsizei;
3327 return data->push_partial_def (pd, ao_ref_alias_set (&lhs_ref),
3328 ao_ref_base_alias_set (&lhs_ref),
3329 0, maxsizei);
3333 /* Continuing with partial defs isn't easily possible here, we
3334 have to find a full def from further lookups from here. Probably
3335 not worth the special-casing everywhere. */
3336 if (!data->partial_defs.is_empty ())
3337 return (void *)-1;
3339 /* Adjust *ref from the new operands. */
3340 ao_ref rhs1_ref;
3341 ao_ref_init (&rhs1_ref, rhs1);
3342 if (!ao_ref_init_from_vn_reference (&r, ao_ref_alias_set (&rhs1_ref),
3343 ao_ref_base_alias_set (&rhs1_ref),
3344 vr->type, vr->operands))
3345 return (void *)-1;
3346 /* This can happen with bitfields. */
3347 if (maybe_ne (ref->size, r.size))
3349 /* If the access lacks some subsetting simply apply that by
3350 shortening it. That in the end can only be successful
3351 if we can pun the lookup result which in turn requires
3352 exact offsets. */
3353 if (known_eq (r.size, r.max_size)
3354 && known_lt (ref->size, r.size))
3355 r.size = r.max_size = ref->size;
3356 else
3357 return (void *)-1;
3359 *ref = r;
3361 /* Do not update last seen VUSE after translating. */
3362 data->last_vuse_ptr = NULL;
3363 /* Invalidate the original access path since it now contains
3364 the wrong base. */
3365 data->orig_ref.ref = NULL_TREE;
3366 /* Use the alias-set of this LHS for recording an eventual result. */
3367 if (data->first_set == -2)
3369 data->first_set = ao_ref_alias_set (&lhs_ref);
3370 data->first_base_set = ao_ref_base_alias_set (&lhs_ref);
3373 /* Keep looking for the adjusted *REF / VR pair. */
3374 return NULL;
3377 /* 6) For memcpy copies translate the reference through them if the copy
3378 kills ref. But we cannot (easily) do this translation if the memcpy is
3379 a storage order barrier, i.e. is equivalent to a VIEW_CONVERT_EXPR that
3380 can modify the storage order of objects (see storage_order_barrier_p). */
3381 else if (data->vn_walk_kind == VN_WALKREWRITE
3382 && is_gimple_reg_type (vr->type)
3383 /* ??? Handle BCOPY as well. */
3384 && (gimple_call_builtin_p (def_stmt, BUILT_IN_MEMCPY)
3385 || gimple_call_builtin_p (def_stmt, BUILT_IN_MEMCPY_CHK)
3386 || gimple_call_builtin_p (def_stmt, BUILT_IN_MEMPCPY)
3387 || gimple_call_builtin_p (def_stmt, BUILT_IN_MEMPCPY_CHK)
3388 || gimple_call_builtin_p (def_stmt, BUILT_IN_MEMMOVE)
3389 || gimple_call_builtin_p (def_stmt, BUILT_IN_MEMMOVE_CHK))
3390 && (TREE_CODE (gimple_call_arg (def_stmt, 0)) == ADDR_EXPR
3391 || TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME)
3392 && (TREE_CODE (gimple_call_arg (def_stmt, 1)) == ADDR_EXPR
3393 || TREE_CODE (gimple_call_arg (def_stmt, 1)) == SSA_NAME)
3394 && (poly_int_tree_p (gimple_call_arg (def_stmt, 2), &copy_size)
3395 || (TREE_CODE (gimple_call_arg (def_stmt, 2)) == SSA_NAME
3396 && poly_int_tree_p (SSA_VAL (gimple_call_arg (def_stmt, 2)),
3397 &copy_size)))
3398 /* Handling this is more complicated, give up for now. */
3399 && data->partial_defs.is_empty ())
3401 tree lhs, rhs;
3402 ao_ref r;
3403 poly_int64 rhs_offset, lhs_offset;
3404 vn_reference_op_s op;
3405 poly_uint64 mem_offset;
3406 poly_int64 at, byte_maxsize;
3408 /* Only handle non-variable, addressable refs. */
3409 if (maybe_ne (ref->size, maxsize)
3410 || !multiple_p (offset, BITS_PER_UNIT, &at)
3411 || !multiple_p (maxsize, BITS_PER_UNIT, &byte_maxsize))
3412 return (void *)-1;
3414 /* Extract a pointer base and an offset for the destination. */
3415 lhs = gimple_call_arg (def_stmt, 0);
3416 lhs_offset = 0;
3417 if (TREE_CODE (lhs) == SSA_NAME)
3419 lhs = vn_valueize (lhs);
3420 if (TREE_CODE (lhs) == SSA_NAME)
3422 gimple *def_stmt = SSA_NAME_DEF_STMT (lhs);
3423 if (gimple_assign_single_p (def_stmt)
3424 && gimple_assign_rhs_code (def_stmt) == ADDR_EXPR)
3425 lhs = gimple_assign_rhs1 (def_stmt);
3428 if (TREE_CODE (lhs) == ADDR_EXPR)
3430 if (AGGREGATE_TYPE_P (TREE_TYPE (TREE_TYPE (lhs)))
3431 && TYPE_REVERSE_STORAGE_ORDER (TREE_TYPE (TREE_TYPE (lhs))))
3432 return (void *)-1;
3433 tree tem = get_addr_base_and_unit_offset (TREE_OPERAND (lhs, 0),
3434 &lhs_offset);
3435 if (!tem)
3436 return (void *)-1;
3437 if (TREE_CODE (tem) == MEM_REF
3438 && poly_int_tree_p (TREE_OPERAND (tem, 1), &mem_offset))
3440 lhs = TREE_OPERAND (tem, 0);
3441 if (TREE_CODE (lhs) == SSA_NAME)
3442 lhs = vn_valueize (lhs);
3443 lhs_offset += mem_offset;
3445 else if (DECL_P (tem))
3446 lhs = build_fold_addr_expr (tem);
3447 else
3448 return (void *)-1;
3450 if (TREE_CODE (lhs) != SSA_NAME
3451 && TREE_CODE (lhs) != ADDR_EXPR)
3452 return (void *)-1;
3454 /* Extract a pointer base and an offset for the source. */
3455 rhs = gimple_call_arg (def_stmt, 1);
3456 rhs_offset = 0;
3457 if (TREE_CODE (rhs) == SSA_NAME)
3458 rhs = vn_valueize (rhs);
3459 if (TREE_CODE (rhs) == ADDR_EXPR)
3461 if (AGGREGATE_TYPE_P (TREE_TYPE (TREE_TYPE (rhs)))
3462 && TYPE_REVERSE_STORAGE_ORDER (TREE_TYPE (TREE_TYPE (rhs))))
3463 return (void *)-1;
3464 tree tem = get_addr_base_and_unit_offset (TREE_OPERAND (rhs, 0),
3465 &rhs_offset);
3466 if (!tem)
3467 return (void *)-1;
3468 if (TREE_CODE (tem) == MEM_REF
3469 && poly_int_tree_p (TREE_OPERAND (tem, 1), &mem_offset))
3471 rhs = TREE_OPERAND (tem, 0);
3472 rhs_offset += mem_offset;
3474 else if (DECL_P (tem)
3475 || TREE_CODE (tem) == STRING_CST)
3476 rhs = build_fold_addr_expr (tem);
3477 else
3478 return (void *)-1;
3480 if (TREE_CODE (rhs) == SSA_NAME)
3481 rhs = SSA_VAL (rhs);
3482 else if (TREE_CODE (rhs) != ADDR_EXPR)
3483 return (void *)-1;
3485 /* The bases of the destination and the references have to agree. */
3486 if (TREE_CODE (base) == MEM_REF)
3488 if (TREE_OPERAND (base, 0) != lhs
3489 || !poly_int_tree_p (TREE_OPERAND (base, 1), &mem_offset))
3490 return (void *) -1;
3491 at += mem_offset;
3493 else if (!DECL_P (base)
3494 || TREE_CODE (lhs) != ADDR_EXPR
3495 || TREE_OPERAND (lhs, 0) != base)
3496 return (void *)-1;
3498 /* If the access is completely outside of the memcpy destination
3499 area there is no aliasing. */
3500 if (!ranges_maybe_overlap_p (lhs_offset, copy_size, at, byte_maxsize))
3501 return NULL;
3502 /* And the access has to be contained within the memcpy destination. */
3503 if (!known_subrange_p (at, byte_maxsize, lhs_offset, copy_size))
3504 return (void *)-1;
3506 /* Save the operands since we need to use the original ones for
3507 the hash entry we use. */
3508 if (!data->saved_operands.exists ())
3509 data->saved_operands = vr->operands.copy ();
3511 /* Make room for 2 operands in the new reference. */
3512 if (vr->operands.length () < 2)
3514 vec<vn_reference_op_s> old = vr->operands;
3515 vr->operands.safe_grow_cleared (2, true);
3516 if (old == shared_lookup_references)
3517 shared_lookup_references = vr->operands;
3519 else
3520 vr->operands.truncate (2);
3522 /* The looked-through reference is a simple MEM_REF. */
3523 memset (&op, 0, sizeof (op));
3524 op.type = vr->type;
3525 op.opcode = MEM_REF;
3526 op.op0 = build_int_cst (ptr_type_node, at - lhs_offset + rhs_offset);
3527 op.off = at - lhs_offset + rhs_offset;
3528 vr->operands[0] = op;
3529 op.type = TREE_TYPE (rhs);
3530 op.opcode = TREE_CODE (rhs);
3531 op.op0 = rhs;
3532 op.off = -1;
3533 vr->operands[1] = op;
3534 vr->hashcode = vn_reference_compute_hash (vr);
3536 /* Try folding the new reference to a constant. */
3537 tree val = fully_constant_vn_reference_p (vr);
3538 if (val)
3539 return data->finish (0, 0, val);
3541 /* Adjust *ref from the new operands. */
3542 if (!ao_ref_init_from_vn_reference (&r, 0, 0, vr->type, vr->operands))
3543 return (void *)-1;
3544 /* This can happen with bitfields. */
3545 if (maybe_ne (ref->size, r.size))
3546 return (void *)-1;
3547 *ref = r;
3549 /* Do not update last seen VUSE after translating. */
3550 data->last_vuse_ptr = NULL;
3551 /* Invalidate the original access path since it now contains
3552 the wrong base. */
3553 data->orig_ref.ref = NULL_TREE;
3554 /* Use the alias-set of this stmt for recording an eventual result. */
3555 if (data->first_set == -2)
3557 data->first_set = 0;
3558 data->first_base_set = 0;
3561 /* Keep looking for the adjusted *REF / VR pair. */
3562 return NULL;
3565 /* Bail out and stop walking. */
3566 return (void *)-1;
3569 /* Return a reference op vector from OP that can be used for
3570 vn_reference_lookup_pieces. The caller is responsible for releasing
3571 the vector. */
3573 vec<vn_reference_op_s>
3574 vn_reference_operands_for_lookup (tree op)
3576 bool valueized;
3577 return valueize_shared_reference_ops_from_ref (op, &valueized).copy ();
3580 /* Lookup a reference operation by it's parts, in the current hash table.
3581 Returns the resulting value number if it exists in the hash table,
3582 NULL_TREE otherwise. VNRESULT will be filled in with the actual
3583 vn_reference_t stored in the hashtable if something is found. */
3585 tree
3586 vn_reference_lookup_pieces (tree vuse, alias_set_type set,
3587 alias_set_type base_set, tree type,
3588 vec<vn_reference_op_s> operands,
3589 vn_reference_t *vnresult, vn_lookup_kind kind)
3591 struct vn_reference_s vr1;
3592 vn_reference_t tmp;
3593 tree cst;
3595 if (!vnresult)
3596 vnresult = &tmp;
3597 *vnresult = NULL;
3599 vr1.vuse = vuse_ssa_val (vuse);
3600 shared_lookup_references.truncate (0);
3601 shared_lookup_references.safe_grow (operands.length (), true);
3602 memcpy (shared_lookup_references.address (),
3603 operands.address (),
3604 sizeof (vn_reference_op_s)
3605 * operands.length ());
3606 bool valueized_p;
3607 valueize_refs_1 (&shared_lookup_references, &valueized_p);
3608 vr1.operands = shared_lookup_references;
3609 vr1.type = type;
3610 vr1.set = set;
3611 vr1.base_set = base_set;
3612 vr1.hashcode = vn_reference_compute_hash (&vr1);
3613 if ((cst = fully_constant_vn_reference_p (&vr1)))
3614 return cst;
3616 vn_reference_lookup_1 (&vr1, vnresult);
3617 if (!*vnresult
3618 && kind != VN_NOWALK
3619 && vr1.vuse)
3621 ao_ref r;
3622 unsigned limit = param_sccvn_max_alias_queries_per_access;
3623 vn_walk_cb_data data (&vr1, NULL_TREE, NULL, kind, true, NULL_TREE,
3624 false);
3625 vec<vn_reference_op_s> ops_for_ref;
3626 if (!valueized_p)
3627 ops_for_ref = vr1.operands;
3628 else
3630 /* For ao_ref_from_mem we have to ensure only available SSA names
3631 end up in base and the only convenient way to make this work
3632 for PRE is to re-valueize with that in mind. */
3633 ops_for_ref.create (operands.length ());
3634 ops_for_ref.quick_grow (operands.length ());
3635 memcpy (ops_for_ref.address (),
3636 operands.address (),
3637 sizeof (vn_reference_op_s)
3638 * operands.length ());
3639 valueize_refs_1 (&ops_for_ref, &valueized_p, true);
3641 if (ao_ref_init_from_vn_reference (&r, set, base_set, type,
3642 ops_for_ref))
3643 *vnresult
3644 = ((vn_reference_t)
3645 walk_non_aliased_vuses (&r, vr1.vuse, true, vn_reference_lookup_2,
3646 vn_reference_lookup_3, vuse_valueize,
3647 limit, &data));
3648 if (ops_for_ref != shared_lookup_references)
3649 ops_for_ref.release ();
3650 gcc_checking_assert (vr1.operands == shared_lookup_references);
3653 if (*vnresult)
3654 return (*vnresult)->result;
3656 return NULL_TREE;
3659 /* Lookup OP in the current hash table, and return the resulting value
3660 number if it exists in the hash table. Return NULL_TREE if it does
3661 not exist in the hash table or if the result field of the structure
3662 was NULL.. VNRESULT will be filled in with the vn_reference_t
3663 stored in the hashtable if one exists. When TBAA_P is false assume
3664 we are looking up a store and treat it as having alias-set zero.
3665 *LAST_VUSE_PTR will be updated with the VUSE the value lookup succeeded.
3666 MASK is either NULL_TREE, or can be an INTEGER_CST if the result of the
3667 load is bitwise anded with MASK and so we are only interested in a subset
3668 of the bits and can ignore if the other bits are uninitialized or
3669 not initialized with constants. When doing redundant store removal
3670 the caller has to set REDUNDANT_STORE_REMOVAL_P. */
3672 tree
3673 vn_reference_lookup (tree op, tree vuse, vn_lookup_kind kind,
3674 vn_reference_t *vnresult, bool tbaa_p,
3675 tree *last_vuse_ptr, tree mask,
3676 bool redundant_store_removal_p)
3678 vec<vn_reference_op_s> operands;
3679 struct vn_reference_s vr1;
3680 bool valueized_anything;
3682 if (vnresult)
3683 *vnresult = NULL;
3685 vr1.vuse = vuse_ssa_val (vuse);
3686 vr1.operands = operands
3687 = valueize_shared_reference_ops_from_ref (op, &valueized_anything);
3689 /* Handle &MEM[ptr + 5].b[1].c as POINTER_PLUS_EXPR. Avoid doing
3690 this before the pass folding __builtin_object_size had a chance to run. */
3691 if ((cfun->curr_properties & PROP_objsz)
3692 && operands[0].opcode == ADDR_EXPR
3693 && operands.last ().opcode == SSA_NAME)
3695 poly_int64 off = 0;
3696 vn_reference_op_t vro;
3697 unsigned i;
3698 for (i = 1; operands.iterate (i, &vro); ++i)
3700 if (vro->opcode == SSA_NAME)
3701 break;
3702 else if (known_eq (vro->off, -1))
3703 break;
3704 off += vro->off;
3706 if (i == operands.length () - 1
3707 /* Make sure we the offset we accumulated in a 64bit int
3708 fits the address computation carried out in target
3709 offset precision. */
3710 && (off.coeffs[0]
3711 == sext_hwi (off.coeffs[0], TYPE_PRECISION (sizetype))))
3713 gcc_assert (operands[i-1].opcode == MEM_REF);
3714 tree ops[2];
3715 ops[0] = operands[i].op0;
3716 ops[1] = wide_int_to_tree (sizetype, off);
3717 tree res = vn_nary_op_lookup_pieces (2, POINTER_PLUS_EXPR,
3718 TREE_TYPE (op), ops, NULL);
3719 if (res)
3720 return res;
3721 return NULL_TREE;
3725 vr1.type = TREE_TYPE (op);
3726 ao_ref op_ref;
3727 ao_ref_init (&op_ref, op);
3728 vr1.set = ao_ref_alias_set (&op_ref);
3729 vr1.base_set = ao_ref_base_alias_set (&op_ref);
3730 vr1.hashcode = vn_reference_compute_hash (&vr1);
3731 if (mask == NULL_TREE)
3732 if (tree cst = fully_constant_vn_reference_p (&vr1))
3733 return cst;
3735 if (kind != VN_NOWALK && vr1.vuse)
3737 vn_reference_t wvnresult;
3738 ao_ref r;
3739 unsigned limit = param_sccvn_max_alias_queries_per_access;
3740 auto_vec<vn_reference_op_s> ops_for_ref;
3741 if (valueized_anything)
3743 copy_reference_ops_from_ref (op, &ops_for_ref);
3744 bool tem;
3745 valueize_refs_1 (&ops_for_ref, &tem, true);
3747 /* Make sure to use a valueized reference if we valueized anything.
3748 Otherwise preserve the full reference for advanced TBAA. */
3749 if (!valueized_anything
3750 || !ao_ref_init_from_vn_reference (&r, vr1.set, vr1.base_set,
3751 vr1.type, ops_for_ref))
3752 ao_ref_init (&r, op);
3753 vn_walk_cb_data data (&vr1, r.ref ? NULL_TREE : op,
3754 last_vuse_ptr, kind, tbaa_p, mask,
3755 redundant_store_removal_p);
3757 wvnresult
3758 = ((vn_reference_t)
3759 walk_non_aliased_vuses (&r, vr1.vuse, tbaa_p, vn_reference_lookup_2,
3760 vn_reference_lookup_3, vuse_valueize, limit,
3761 &data));
3762 gcc_checking_assert (vr1.operands == shared_lookup_references);
3763 if (wvnresult)
3765 gcc_assert (mask == NULL_TREE);
3766 if (vnresult)
3767 *vnresult = wvnresult;
3768 return wvnresult->result;
3770 else if (mask)
3771 return data.masked_result;
3773 return NULL_TREE;
3776 if (last_vuse_ptr)
3777 *last_vuse_ptr = vr1.vuse;
3778 if (mask)
3779 return NULL_TREE;
3780 return vn_reference_lookup_1 (&vr1, vnresult);
3783 /* Lookup CALL in the current hash table and return the entry in
3784 *VNRESULT if found. Populates *VR for the hashtable lookup. */
3786 void
3787 vn_reference_lookup_call (gcall *call, vn_reference_t *vnresult,
3788 vn_reference_t vr)
3790 if (vnresult)
3791 *vnresult = NULL;
3793 tree vuse = gimple_vuse (call);
3795 vr->vuse = vuse ? SSA_VAL (vuse) : NULL_TREE;
3796 vr->operands = valueize_shared_reference_ops_from_call (call);
3797 tree lhs = gimple_call_lhs (call);
3798 /* For non-SSA return values the referece ops contain the LHS. */
3799 vr->type = ((lhs && TREE_CODE (lhs) == SSA_NAME)
3800 ? TREE_TYPE (lhs) : NULL_TREE);
3801 vr->punned = false;
3802 vr->set = 0;
3803 vr->base_set = 0;
3804 vr->hashcode = vn_reference_compute_hash (vr);
3805 vn_reference_lookup_1 (vr, vnresult);
3808 /* Insert OP into the current hash table with a value number of RESULT. */
3810 static void
3811 vn_reference_insert (tree op, tree result, tree vuse, tree vdef)
3813 vn_reference_s **slot;
3814 vn_reference_t vr1;
3815 bool tem;
3817 vec<vn_reference_op_s> operands
3818 = valueize_shared_reference_ops_from_ref (op, &tem);
3819 /* Handle &MEM[ptr + 5].b[1].c as POINTER_PLUS_EXPR. Avoid doing this
3820 before the pass folding __builtin_object_size had a chance to run. */
3821 if ((cfun->curr_properties & PROP_objsz)
3822 && operands[0].opcode == ADDR_EXPR
3823 && operands.last ().opcode == SSA_NAME)
3825 poly_int64 off = 0;
3826 vn_reference_op_t vro;
3827 unsigned i;
3828 for (i = 1; operands.iterate (i, &vro); ++i)
3830 if (vro->opcode == SSA_NAME)
3831 break;
3832 else if (known_eq (vro->off, -1))
3833 break;
3834 off += vro->off;
3836 if (i == operands.length () - 1
3837 /* Make sure we the offset we accumulated in a 64bit int
3838 fits the address computation carried out in target
3839 offset precision. */
3840 && (off.coeffs[0]
3841 == sext_hwi (off.coeffs[0], TYPE_PRECISION (sizetype))))
3843 gcc_assert (operands[i-1].opcode == MEM_REF);
3844 tree ops[2];
3845 ops[0] = operands[i].op0;
3846 ops[1] = wide_int_to_tree (sizetype, off);
3847 vn_nary_op_insert_pieces (2, POINTER_PLUS_EXPR,
3848 TREE_TYPE (op), ops, result,
3849 VN_INFO (result)->value_id);
3850 return;
3854 vr1 = XOBNEW (&vn_tables_obstack, vn_reference_s);
3855 if (TREE_CODE (result) == SSA_NAME)
3856 vr1->value_id = VN_INFO (result)->value_id;
3857 else
3858 vr1->value_id = get_or_alloc_constant_value_id (result);
3859 vr1->vuse = vuse_ssa_val (vuse);
3860 vr1->operands = operands.copy ();
3861 vr1->type = TREE_TYPE (op);
3862 vr1->punned = false;
3863 ao_ref op_ref;
3864 ao_ref_init (&op_ref, op);
3865 vr1->set = ao_ref_alias_set (&op_ref);
3866 vr1->base_set = ao_ref_base_alias_set (&op_ref);
3867 vr1->hashcode = vn_reference_compute_hash (vr1);
3868 vr1->result = TREE_CODE (result) == SSA_NAME ? SSA_VAL (result) : result;
3869 vr1->result_vdef = vdef;
3871 slot = valid_info->references->find_slot_with_hash (vr1, vr1->hashcode,
3872 INSERT);
3874 /* Because IL walking on reference lookup can end up visiting
3875 a def that is only to be visited later in iteration order
3876 when we are about to make an irreducible region reducible
3877 the def can be effectively processed and its ref being inserted
3878 by vn_reference_lookup_3 already. So we cannot assert (!*slot)
3879 but save a lookup if we deal with already inserted refs here. */
3880 if (*slot)
3882 /* We cannot assert that we have the same value either because
3883 when disentangling an irreducible region we may end up visiting
3884 a use before the corresponding def. That's a missed optimization
3885 only though. See gcc.dg/tree-ssa/pr87126.c for example. */
3886 if (dump_file && (dump_flags & TDF_DETAILS)
3887 && !operand_equal_p ((*slot)->result, vr1->result, 0))
3889 fprintf (dump_file, "Keeping old value ");
3890 print_generic_expr (dump_file, (*slot)->result);
3891 fprintf (dump_file, " because of collision\n");
3893 free_reference (vr1);
3894 obstack_free (&vn_tables_obstack, vr1);
3895 return;
3898 *slot = vr1;
3899 vr1->next = last_inserted_ref;
3900 last_inserted_ref = vr1;
3903 /* Insert a reference by it's pieces into the current hash table with
3904 a value number of RESULT. Return the resulting reference
3905 structure we created. */
3907 vn_reference_t
3908 vn_reference_insert_pieces (tree vuse, alias_set_type set,
3909 alias_set_type base_set, tree type,
3910 vec<vn_reference_op_s> operands,
3911 tree result, unsigned int value_id)
3914 vn_reference_s **slot;
3915 vn_reference_t vr1;
3917 vr1 = XOBNEW (&vn_tables_obstack, vn_reference_s);
3918 vr1->value_id = value_id;
3919 vr1->vuse = vuse_ssa_val (vuse);
3920 vr1->operands = operands;
3921 valueize_refs (&vr1->operands);
3922 vr1->type = type;
3923 vr1->punned = false;
3924 vr1->set = set;
3925 vr1->base_set = base_set;
3926 vr1->hashcode = vn_reference_compute_hash (vr1);
3927 if (result && TREE_CODE (result) == SSA_NAME)
3928 result = SSA_VAL (result);
3929 vr1->result = result;
3930 vr1->result_vdef = NULL_TREE;
3932 slot = valid_info->references->find_slot_with_hash (vr1, vr1->hashcode,
3933 INSERT);
3935 /* At this point we should have all the things inserted that we have
3936 seen before, and we should never try inserting something that
3937 already exists. */
3938 gcc_assert (!*slot);
3940 *slot = vr1;
3941 vr1->next = last_inserted_ref;
3942 last_inserted_ref = vr1;
3943 return vr1;
3946 /* Compute and return the hash value for nary operation VBO1. */
3948 hashval_t
3949 vn_nary_op_compute_hash (const vn_nary_op_t vno1)
3951 inchash::hash hstate;
3952 unsigned i;
3954 if (((vno1->length == 2
3955 && commutative_tree_code (vno1->opcode))
3956 || (vno1->length == 3
3957 && commutative_ternary_tree_code (vno1->opcode)))
3958 && tree_swap_operands_p (vno1->op[0], vno1->op[1]))
3959 std::swap (vno1->op[0], vno1->op[1]);
3960 else if (TREE_CODE_CLASS (vno1->opcode) == tcc_comparison
3961 && tree_swap_operands_p (vno1->op[0], vno1->op[1]))
3963 std::swap (vno1->op[0], vno1->op[1]);
3964 vno1->opcode = swap_tree_comparison (vno1->opcode);
3967 hstate.add_int (vno1->opcode);
3968 for (i = 0; i < vno1->length; ++i)
3969 inchash::add_expr (vno1->op[i], hstate);
3971 return hstate.end ();
3974 /* Compare nary operations VNO1 and VNO2 and return true if they are
3975 equivalent. */
3977 bool
3978 vn_nary_op_eq (const_vn_nary_op_t const vno1, const_vn_nary_op_t const vno2)
3980 unsigned i;
3982 if (vno1->hashcode != vno2->hashcode)
3983 return false;
3985 if (vno1->length != vno2->length)
3986 return false;
3988 if (vno1->opcode != vno2->opcode
3989 || !types_compatible_p (vno1->type, vno2->type))
3990 return false;
3992 for (i = 0; i < vno1->length; ++i)
3993 if (!expressions_equal_p (vno1->op[i], vno2->op[i]))
3994 return false;
3996 /* BIT_INSERT_EXPR has an implict operand as the type precision
3997 of op1. Need to check to make sure they are the same. */
3998 if (vno1->opcode == BIT_INSERT_EXPR
3999 && TREE_CODE (vno1->op[1]) == INTEGER_CST
4000 && TYPE_PRECISION (TREE_TYPE (vno1->op[1]))
4001 != TYPE_PRECISION (TREE_TYPE (vno2->op[1])))
4002 return false;
4004 return true;
4007 /* Initialize VNO from the pieces provided. */
4009 static void
4010 init_vn_nary_op_from_pieces (vn_nary_op_t vno, unsigned int length,
4011 enum tree_code code, tree type, tree *ops)
4013 vno->opcode = code;
4014 vno->length = length;
4015 vno->type = type;
4016 memcpy (&vno->op[0], ops, sizeof (tree) * length);
4019 /* Return the number of operands for a vn_nary ops structure from STMT. */
4021 unsigned int
4022 vn_nary_length_from_stmt (gimple *stmt)
4024 switch (gimple_assign_rhs_code (stmt))
4026 case REALPART_EXPR:
4027 case IMAGPART_EXPR:
4028 case VIEW_CONVERT_EXPR:
4029 return 1;
4031 case BIT_FIELD_REF:
4032 return 3;
4034 case CONSTRUCTOR:
4035 return CONSTRUCTOR_NELTS (gimple_assign_rhs1 (stmt));
4037 default:
4038 return gimple_num_ops (stmt) - 1;
4042 /* Initialize VNO from STMT. */
4044 void
4045 init_vn_nary_op_from_stmt (vn_nary_op_t vno, gassign *stmt)
4047 unsigned i;
4049 vno->opcode = gimple_assign_rhs_code (stmt);
4050 vno->type = TREE_TYPE (gimple_assign_lhs (stmt));
4051 switch (vno->opcode)
4053 case REALPART_EXPR:
4054 case IMAGPART_EXPR:
4055 case VIEW_CONVERT_EXPR:
4056 vno->length = 1;
4057 vno->op[0] = TREE_OPERAND (gimple_assign_rhs1 (stmt), 0);
4058 break;
4060 case BIT_FIELD_REF:
4061 vno->length = 3;
4062 vno->op[0] = TREE_OPERAND (gimple_assign_rhs1 (stmt), 0);
4063 vno->op[1] = TREE_OPERAND (gimple_assign_rhs1 (stmt), 1);
4064 vno->op[2] = TREE_OPERAND (gimple_assign_rhs1 (stmt), 2);
4065 break;
4067 case CONSTRUCTOR:
4068 vno->length = CONSTRUCTOR_NELTS (gimple_assign_rhs1 (stmt));
4069 for (i = 0; i < vno->length; ++i)
4070 vno->op[i] = CONSTRUCTOR_ELT (gimple_assign_rhs1 (stmt), i)->value;
4071 break;
4073 default:
4074 gcc_checking_assert (!gimple_assign_single_p (stmt));
4075 vno->length = gimple_num_ops (stmt) - 1;
4076 for (i = 0; i < vno->length; ++i)
4077 vno->op[i] = gimple_op (stmt, i + 1);
4081 /* Compute the hashcode for VNO and look for it in the hash table;
4082 return the resulting value number if it exists in the hash table.
4083 Return NULL_TREE if it does not exist in the hash table or if the
4084 result field of the operation is NULL. VNRESULT will contain the
4085 vn_nary_op_t from the hashtable if it exists. */
4087 static tree
4088 vn_nary_op_lookup_1 (vn_nary_op_t vno, vn_nary_op_t *vnresult)
4090 vn_nary_op_s **slot;
4092 if (vnresult)
4093 *vnresult = NULL;
4095 for (unsigned i = 0; i < vno->length; ++i)
4096 if (TREE_CODE (vno->op[i]) == SSA_NAME)
4097 vno->op[i] = SSA_VAL (vno->op[i]);
4099 vno->hashcode = vn_nary_op_compute_hash (vno);
4100 slot = valid_info->nary->find_slot_with_hash (vno, vno->hashcode, NO_INSERT);
4101 if (!slot)
4102 return NULL_TREE;
4103 if (vnresult)
4104 *vnresult = *slot;
4105 return (*slot)->predicated_values ? NULL_TREE : (*slot)->u.result;
4108 /* Lookup a n-ary operation by its pieces and return the resulting value
4109 number if it exists in the hash table. Return NULL_TREE if it does
4110 not exist in the hash table or if the result field of the operation
4111 is NULL. VNRESULT will contain the vn_nary_op_t from the hashtable
4112 if it exists. */
4114 tree
4115 vn_nary_op_lookup_pieces (unsigned int length, enum tree_code code,
4116 tree type, tree *ops, vn_nary_op_t *vnresult)
4118 vn_nary_op_t vno1 = XALLOCAVAR (struct vn_nary_op_s,
4119 sizeof_vn_nary_op (length));
4120 init_vn_nary_op_from_pieces (vno1, length, code, type, ops);
4121 return vn_nary_op_lookup_1 (vno1, vnresult);
4124 /* Lookup the rhs of STMT in the current hash table, and return the resulting
4125 value number if it exists in the hash table. Return NULL_TREE if
4126 it does not exist in the hash table. VNRESULT will contain the
4127 vn_nary_op_t from the hashtable if it exists. */
4129 tree
4130 vn_nary_op_lookup_stmt (gimple *stmt, vn_nary_op_t *vnresult)
4132 vn_nary_op_t vno1
4133 = XALLOCAVAR (struct vn_nary_op_s,
4134 sizeof_vn_nary_op (vn_nary_length_from_stmt (stmt)));
4135 init_vn_nary_op_from_stmt (vno1, as_a <gassign *> (stmt));
4136 return vn_nary_op_lookup_1 (vno1, vnresult);
4139 /* Allocate a vn_nary_op_t with LENGTH operands on STACK. */
4141 vn_nary_op_t
4142 alloc_vn_nary_op_noinit (unsigned int length, struct obstack *stack)
4144 return (vn_nary_op_t) obstack_alloc (stack, sizeof_vn_nary_op (length));
4147 /* Allocate and initialize a vn_nary_op_t on CURRENT_INFO's
4148 obstack. */
4150 static vn_nary_op_t
4151 alloc_vn_nary_op (unsigned int length, tree result, unsigned int value_id)
4153 vn_nary_op_t vno1 = alloc_vn_nary_op_noinit (length, &vn_tables_obstack);
4155 vno1->value_id = value_id;
4156 vno1->length = length;
4157 vno1->predicated_values = 0;
4158 vno1->u.result = result;
4160 return vno1;
4163 /* Insert VNO into TABLE. */
4165 static vn_nary_op_t
4166 vn_nary_op_insert_into (vn_nary_op_t vno, vn_nary_op_table_type *table)
4168 vn_nary_op_s **slot;
4170 gcc_assert (! vno->predicated_values
4171 || (! vno->u.values->next
4172 && vno->u.values->n == 1));
4174 for (unsigned i = 0; i < vno->length; ++i)
4175 if (TREE_CODE (vno->op[i]) == SSA_NAME)
4176 vno->op[i] = SSA_VAL (vno->op[i]);
4178 vno->hashcode = vn_nary_op_compute_hash (vno);
4179 slot = table->find_slot_with_hash (vno, vno->hashcode, INSERT);
4180 vno->unwind_to = *slot;
4181 if (*slot)
4183 /* Prefer non-predicated values.
4184 ??? Only if those are constant, otherwise, with constant predicated
4185 value, turn them into predicated values with entry-block validity
4186 (??? but we always find the first valid result currently). */
4187 if ((*slot)->predicated_values
4188 && ! vno->predicated_values)
4190 /* ??? We cannot remove *slot from the unwind stack list.
4191 For the moment we deal with this by skipping not found
4192 entries but this isn't ideal ... */
4193 *slot = vno;
4194 /* ??? Maintain a stack of states we can unwind in
4195 vn_nary_op_s? But how far do we unwind? In reality
4196 we need to push change records somewhere... Or not
4197 unwind vn_nary_op_s and linking them but instead
4198 unwind the results "list", linking that, which also
4199 doesn't move on hashtable resize. */
4200 /* We can also have a ->unwind_to recording *slot there.
4201 That way we can make u.values a fixed size array with
4202 recording the number of entries but of course we then
4203 have always N copies for each unwind_to-state. Or we
4204 make sure to only ever append and each unwinding will
4205 pop off one entry (but how to deal with predicated
4206 replaced with non-predicated here?) */
4207 vno->next = last_inserted_nary;
4208 last_inserted_nary = vno;
4209 return vno;
4211 else if (vno->predicated_values
4212 && ! (*slot)->predicated_values)
4213 return *slot;
4214 else if (vno->predicated_values
4215 && (*slot)->predicated_values)
4217 /* ??? Factor this all into a insert_single_predicated_value
4218 routine. */
4219 gcc_assert (!vno->u.values->next && vno->u.values->n == 1);
4220 basic_block vno_bb
4221 = BASIC_BLOCK_FOR_FN (cfun, vno->u.values->valid_dominated_by_p[0]);
4222 vn_pval *nval = vno->u.values;
4223 vn_pval **next = &vno->u.values;
4224 bool found = false;
4225 for (vn_pval *val = (*slot)->u.values; val; val = val->next)
4227 if (expressions_equal_p (val->result, nval->result))
4229 found = true;
4230 for (unsigned i = 0; i < val->n; ++i)
4232 basic_block val_bb
4233 = BASIC_BLOCK_FOR_FN (cfun,
4234 val->valid_dominated_by_p[i]);
4235 if (dominated_by_p (CDI_DOMINATORS, vno_bb, val_bb))
4236 /* Value registered with more generic predicate. */
4237 return *slot;
4238 else if (dominated_by_p (CDI_DOMINATORS, val_bb, vno_bb))
4239 /* Shouldn't happen, we insert in RPO order. */
4240 gcc_unreachable ();
4242 /* Append value. */
4243 *next = (vn_pval *) obstack_alloc (&vn_tables_obstack,
4244 sizeof (vn_pval)
4245 + val->n * sizeof (int));
4246 (*next)->next = NULL;
4247 (*next)->result = val->result;
4248 (*next)->n = val->n + 1;
4249 memcpy ((*next)->valid_dominated_by_p,
4250 val->valid_dominated_by_p,
4251 val->n * sizeof (int));
4252 (*next)->valid_dominated_by_p[val->n] = vno_bb->index;
4253 next = &(*next)->next;
4254 if (dump_file && (dump_flags & TDF_DETAILS))
4255 fprintf (dump_file, "Appending predicate to value.\n");
4256 continue;
4258 /* Copy other predicated values. */
4259 *next = (vn_pval *) obstack_alloc (&vn_tables_obstack,
4260 sizeof (vn_pval)
4261 + (val->n-1) * sizeof (int));
4262 memcpy (*next, val, sizeof (vn_pval) + (val->n-1) * sizeof (int));
4263 (*next)->next = NULL;
4264 next = &(*next)->next;
4266 if (!found)
4267 *next = nval;
4269 *slot = vno;
4270 vno->next = last_inserted_nary;
4271 last_inserted_nary = vno;
4272 return vno;
4275 /* While we do not want to insert things twice it's awkward to
4276 avoid it in the case where visit_nary_op pattern-matches stuff
4277 and ends up simplifying the replacement to itself. We then
4278 get two inserts, one from visit_nary_op and one from
4279 vn_nary_build_or_lookup.
4280 So allow inserts with the same value number. */
4281 if ((*slot)->u.result == vno->u.result)
4282 return *slot;
4285 /* ??? There's also optimistic vs. previous commited state merging
4286 that is problematic for the case of unwinding. */
4288 /* ??? We should return NULL if we do not use 'vno' and have the
4289 caller release it. */
4290 gcc_assert (!*slot);
4292 *slot = vno;
4293 vno->next = last_inserted_nary;
4294 last_inserted_nary = vno;
4295 return vno;
4298 /* Insert a n-ary operation into the current hash table using it's
4299 pieces. Return the vn_nary_op_t structure we created and put in
4300 the hashtable. */
4302 vn_nary_op_t
4303 vn_nary_op_insert_pieces (unsigned int length, enum tree_code code,
4304 tree type, tree *ops,
4305 tree result, unsigned int value_id)
4307 vn_nary_op_t vno1 = alloc_vn_nary_op (length, result, value_id);
4308 init_vn_nary_op_from_pieces (vno1, length, code, type, ops);
4309 return vn_nary_op_insert_into (vno1, valid_info->nary);
4312 static vn_nary_op_t
4313 vn_nary_op_insert_pieces_predicated (unsigned int length, enum tree_code code,
4314 tree type, tree *ops,
4315 tree result, unsigned int value_id,
4316 edge pred_e)
4318 /* ??? Currently tracking BBs. */
4319 if (! single_pred_p (pred_e->dest))
4321 /* Never record for backedges. */
4322 if (pred_e->flags & EDGE_DFS_BACK)
4323 return NULL;
4324 edge_iterator ei;
4325 edge e;
4326 int cnt = 0;
4327 /* Ignore backedges. */
4328 FOR_EACH_EDGE (e, ei, pred_e->dest->preds)
4329 if (! dominated_by_p (CDI_DOMINATORS, e->src, e->dest))
4330 cnt++;
4331 if (cnt != 1)
4332 return NULL;
4334 if (dump_file && (dump_flags & TDF_DETAILS)
4335 /* ??? Fix dumping, but currently we only get comparisons. */
4336 && TREE_CODE_CLASS (code) == tcc_comparison)
4338 fprintf (dump_file, "Recording on edge %d->%d ", pred_e->src->index,
4339 pred_e->dest->index);
4340 print_generic_expr (dump_file, ops[0], TDF_SLIM);
4341 fprintf (dump_file, " %s ", get_tree_code_name (code));
4342 print_generic_expr (dump_file, ops[1], TDF_SLIM);
4343 fprintf (dump_file, " == %s\n",
4344 integer_zerop (result) ? "false" : "true");
4346 vn_nary_op_t vno1 = alloc_vn_nary_op (length, NULL_TREE, value_id);
4347 init_vn_nary_op_from_pieces (vno1, length, code, type, ops);
4348 vno1->predicated_values = 1;
4349 vno1->u.values = (vn_pval *) obstack_alloc (&vn_tables_obstack,
4350 sizeof (vn_pval));
4351 vno1->u.values->next = NULL;
4352 vno1->u.values->result = result;
4353 vno1->u.values->n = 1;
4354 vno1->u.values->valid_dominated_by_p[0] = pred_e->dest->index;
4355 return vn_nary_op_insert_into (vno1, valid_info->nary);
4358 static bool
4359 dominated_by_p_w_unex (basic_block bb1, basic_block bb2, bool);
4361 static tree
4362 vn_nary_op_get_predicated_value (vn_nary_op_t vno, basic_block bb)
4364 if (! vno->predicated_values)
4365 return vno->u.result;
4366 for (vn_pval *val = vno->u.values; val; val = val->next)
4367 for (unsigned i = 0; i < val->n; ++i)
4368 /* Do not handle backedge executability optimistically since
4369 when figuring out whether to iterate we do not consider
4370 changed predication. */
4371 if (dominated_by_p_w_unex
4372 (bb, BASIC_BLOCK_FOR_FN (cfun, val->valid_dominated_by_p[i]),
4373 false))
4374 return val->result;
4375 return NULL_TREE;
4378 /* Insert the rhs of STMT into the current hash table with a value number of
4379 RESULT. */
4381 static vn_nary_op_t
4382 vn_nary_op_insert_stmt (gimple *stmt, tree result)
4384 vn_nary_op_t vno1
4385 = alloc_vn_nary_op (vn_nary_length_from_stmt (stmt),
4386 result, VN_INFO (result)->value_id);
4387 init_vn_nary_op_from_stmt (vno1, as_a <gassign *> (stmt));
4388 return vn_nary_op_insert_into (vno1, valid_info->nary);
4391 /* Compute a hashcode for PHI operation VP1 and return it. */
4393 static inline hashval_t
4394 vn_phi_compute_hash (vn_phi_t vp1)
4396 inchash::hash hstate;
4397 tree phi1op;
4398 tree type;
4399 edge e;
4400 edge_iterator ei;
4402 hstate.add_int (EDGE_COUNT (vp1->block->preds));
4403 switch (EDGE_COUNT (vp1->block->preds))
4405 case 1:
4406 break;
4407 case 2:
4408 if (vp1->block->loop_father->header == vp1->block)
4410 else
4411 break;
4412 /* Fallthru. */
4413 default:
4414 hstate.add_int (vp1->block->index);
4417 /* If all PHI arguments are constants we need to distinguish
4418 the PHI node via its type. */
4419 type = vp1->type;
4420 hstate.merge_hash (vn_hash_type (type));
4422 FOR_EACH_EDGE (e, ei, vp1->block->preds)
4424 /* Don't hash backedge values they need to be handled as VN_TOP
4425 for optimistic value-numbering. */
4426 if (e->flags & EDGE_DFS_BACK)
4427 continue;
4429 phi1op = vp1->phiargs[e->dest_idx];
4430 if (phi1op == VN_TOP)
4431 continue;
4432 inchash::add_expr (phi1op, hstate);
4435 return hstate.end ();
4439 /* Return true if COND1 and COND2 represent the same condition, set
4440 *INVERTED_P if one needs to be inverted to make it the same as
4441 the other. */
4443 static bool
4444 cond_stmts_equal_p (gcond *cond1, tree lhs1, tree rhs1,
4445 gcond *cond2, tree lhs2, tree rhs2, bool *inverted_p)
4447 enum tree_code code1 = gimple_cond_code (cond1);
4448 enum tree_code code2 = gimple_cond_code (cond2);
4450 *inverted_p = false;
4451 if (code1 == code2)
4453 else if (code1 == swap_tree_comparison (code2))
4454 std::swap (lhs2, rhs2);
4455 else if (code1 == invert_tree_comparison (code2, HONOR_NANS (lhs2)))
4456 *inverted_p = true;
4457 else if (code1 == invert_tree_comparison
4458 (swap_tree_comparison (code2), HONOR_NANS (lhs2)))
4460 std::swap (lhs2, rhs2);
4461 *inverted_p = true;
4463 else
4464 return false;
4466 return ((expressions_equal_p (lhs1, lhs2)
4467 && expressions_equal_p (rhs1, rhs2))
4468 || (commutative_tree_code (code1)
4469 && expressions_equal_p (lhs1, rhs2)
4470 && expressions_equal_p (rhs1, lhs2)));
4473 /* Compare two phi entries for equality, ignoring VN_TOP arguments. */
4475 static int
4476 vn_phi_eq (const_vn_phi_t const vp1, const_vn_phi_t const vp2)
4478 if (vp1->hashcode != vp2->hashcode)
4479 return false;
4481 if (vp1->block != vp2->block)
4483 if (EDGE_COUNT (vp1->block->preds) != EDGE_COUNT (vp2->block->preds))
4484 return false;
4486 switch (EDGE_COUNT (vp1->block->preds))
4488 case 1:
4489 /* Single-arg PHIs are just copies. */
4490 break;
4492 case 2:
4494 /* Rule out backedges into the PHI. */
4495 if (vp1->block->loop_father->header == vp1->block
4496 || vp2->block->loop_father->header == vp2->block)
4497 return false;
4499 /* If the PHI nodes do not have compatible types
4500 they are not the same. */
4501 if (!types_compatible_p (vp1->type, vp2->type))
4502 return false;
4504 basic_block idom1
4505 = get_immediate_dominator (CDI_DOMINATORS, vp1->block);
4506 basic_block idom2
4507 = get_immediate_dominator (CDI_DOMINATORS, vp2->block);
4508 /* If the immediate dominator end in switch stmts multiple
4509 values may end up in the same PHI arg via intermediate
4510 CFG merges. */
4511 if (EDGE_COUNT (idom1->succs) != 2
4512 || EDGE_COUNT (idom2->succs) != 2)
4513 return false;
4515 /* Verify the controlling stmt is the same. */
4516 gcond *last1 = safe_dyn_cast <gcond *> (last_stmt (idom1));
4517 gcond *last2 = safe_dyn_cast <gcond *> (last_stmt (idom2));
4518 if (! last1 || ! last2)
4519 return false;
4520 bool inverted_p;
4521 if (! cond_stmts_equal_p (last1, vp1->cclhs, vp1->ccrhs,
4522 last2, vp2->cclhs, vp2->ccrhs,
4523 &inverted_p))
4524 return false;
4526 /* Get at true/false controlled edges into the PHI. */
4527 edge te1, te2, fe1, fe2;
4528 if (! extract_true_false_controlled_edges (idom1, vp1->block,
4529 &te1, &fe1)
4530 || ! extract_true_false_controlled_edges (idom2, vp2->block,
4531 &te2, &fe2))
4532 return false;
4534 /* Swap edges if the second condition is the inverted of the
4535 first. */
4536 if (inverted_p)
4537 std::swap (te2, fe2);
4539 /* Since we do not know which edge will be executed we have
4540 to be careful when matching VN_TOP. Be conservative and
4541 only match VN_TOP == VN_TOP for now, we could allow
4542 VN_TOP on the not prevailing PHI though. See for example
4543 PR102920. */
4544 if (! expressions_equal_p (vp1->phiargs[te1->dest_idx],
4545 vp2->phiargs[te2->dest_idx], false)
4546 || ! expressions_equal_p (vp1->phiargs[fe1->dest_idx],
4547 vp2->phiargs[fe2->dest_idx], false))
4548 return false;
4550 return true;
4553 default:
4554 return false;
4558 /* If the PHI nodes do not have compatible types
4559 they are not the same. */
4560 if (!types_compatible_p (vp1->type, vp2->type))
4561 return false;
4563 /* Any phi in the same block will have it's arguments in the
4564 same edge order, because of how we store phi nodes. */
4565 unsigned nargs = EDGE_COUNT (vp1->block->preds);
4566 for (unsigned i = 0; i < nargs; ++i)
4568 tree phi1op = vp1->phiargs[i];
4569 tree phi2op = vp2->phiargs[i];
4570 if (phi1op == phi2op)
4571 continue;
4572 if (!expressions_equal_p (phi1op, phi2op, false))
4573 return false;
4576 return true;
4579 /* Lookup PHI in the current hash table, and return the resulting
4580 value number if it exists in the hash table. Return NULL_TREE if
4581 it does not exist in the hash table. */
4583 static tree
4584 vn_phi_lookup (gimple *phi, bool backedges_varying_p)
4586 vn_phi_s **slot;
4587 struct vn_phi_s *vp1;
4588 edge e;
4589 edge_iterator ei;
4591 vp1 = XALLOCAVAR (struct vn_phi_s,
4592 sizeof (struct vn_phi_s)
4593 + (gimple_phi_num_args (phi) - 1) * sizeof (tree));
4595 /* Canonicalize the SSA_NAME's to their value number. */
4596 FOR_EACH_EDGE (e, ei, gimple_bb (phi)->preds)
4598 tree def = PHI_ARG_DEF_FROM_EDGE (phi, e);
4599 if (TREE_CODE (def) == SSA_NAME
4600 && (!backedges_varying_p || !(e->flags & EDGE_DFS_BACK)))
4602 if (ssa_undefined_value_p (def, false))
4603 def = VN_TOP;
4604 else
4605 def = SSA_VAL (def);
4607 vp1->phiargs[e->dest_idx] = def;
4609 vp1->type = TREE_TYPE (gimple_phi_result (phi));
4610 vp1->block = gimple_bb (phi);
4611 /* Extract values of the controlling condition. */
4612 vp1->cclhs = NULL_TREE;
4613 vp1->ccrhs = NULL_TREE;
4614 basic_block idom1 = get_immediate_dominator (CDI_DOMINATORS, vp1->block);
4615 if (EDGE_COUNT (idom1->succs) == 2)
4616 if (gcond *last1 = safe_dyn_cast <gcond *> (last_stmt (idom1)))
4618 /* ??? We want to use SSA_VAL here. But possibly not
4619 allow VN_TOP. */
4620 vp1->cclhs = vn_valueize (gimple_cond_lhs (last1));
4621 vp1->ccrhs = vn_valueize (gimple_cond_rhs (last1));
4623 vp1->hashcode = vn_phi_compute_hash (vp1);
4624 slot = valid_info->phis->find_slot_with_hash (vp1, vp1->hashcode, NO_INSERT);
4625 if (!slot)
4626 return NULL_TREE;
4627 return (*slot)->result;
4630 /* Insert PHI into the current hash table with a value number of
4631 RESULT. */
4633 static vn_phi_t
4634 vn_phi_insert (gimple *phi, tree result, bool backedges_varying_p)
4636 vn_phi_s **slot;
4637 vn_phi_t vp1 = (vn_phi_t) obstack_alloc (&vn_tables_obstack,
4638 sizeof (vn_phi_s)
4639 + ((gimple_phi_num_args (phi) - 1)
4640 * sizeof (tree)));
4641 edge e;
4642 edge_iterator ei;
4644 /* Canonicalize the SSA_NAME's to their value number. */
4645 FOR_EACH_EDGE (e, ei, gimple_bb (phi)->preds)
4647 tree def = PHI_ARG_DEF_FROM_EDGE (phi, e);
4648 if (TREE_CODE (def) == SSA_NAME
4649 && (!backedges_varying_p || !(e->flags & EDGE_DFS_BACK)))
4651 if (ssa_undefined_value_p (def, false))
4652 def = VN_TOP;
4653 else
4654 def = SSA_VAL (def);
4656 vp1->phiargs[e->dest_idx] = def;
4658 vp1->value_id = VN_INFO (result)->value_id;
4659 vp1->type = TREE_TYPE (gimple_phi_result (phi));
4660 vp1->block = gimple_bb (phi);
4661 /* Extract values of the controlling condition. */
4662 vp1->cclhs = NULL_TREE;
4663 vp1->ccrhs = NULL_TREE;
4664 basic_block idom1 = get_immediate_dominator (CDI_DOMINATORS, vp1->block);
4665 if (EDGE_COUNT (idom1->succs) == 2)
4666 if (gcond *last1 = safe_dyn_cast <gcond *> (last_stmt (idom1)))
4668 /* ??? We want to use SSA_VAL here. But possibly not
4669 allow VN_TOP. */
4670 vp1->cclhs = vn_valueize (gimple_cond_lhs (last1));
4671 vp1->ccrhs = vn_valueize (gimple_cond_rhs (last1));
4673 vp1->result = result;
4674 vp1->hashcode = vn_phi_compute_hash (vp1);
4676 slot = valid_info->phis->find_slot_with_hash (vp1, vp1->hashcode, INSERT);
4677 gcc_assert (!*slot);
4679 *slot = vp1;
4680 vp1->next = last_inserted_phi;
4681 last_inserted_phi = vp1;
4682 return vp1;
4686 /* Return true if BB1 is dominated by BB2 taking into account edges
4687 that are not executable. When ALLOW_BACK is false consider not
4688 executable backedges as executable. */
4690 static bool
4691 dominated_by_p_w_unex (basic_block bb1, basic_block bb2, bool allow_back)
4693 edge_iterator ei;
4694 edge e;
4696 if (dominated_by_p (CDI_DOMINATORS, bb1, bb2))
4697 return true;
4699 /* Before iterating we'd like to know if there exists a
4700 (executable) path from bb2 to bb1 at all, if not we can
4701 directly return false. For now simply iterate once. */
4703 /* Iterate to the single executable bb1 predecessor. */
4704 if (EDGE_COUNT (bb1->preds) > 1)
4706 edge prede = NULL;
4707 FOR_EACH_EDGE (e, ei, bb1->preds)
4708 if ((e->flags & EDGE_EXECUTABLE)
4709 || (!allow_back && (e->flags & EDGE_DFS_BACK)))
4711 if (prede)
4713 prede = NULL;
4714 break;
4716 prede = e;
4718 if (prede)
4720 bb1 = prede->src;
4722 /* Re-do the dominance check with changed bb1. */
4723 if (dominated_by_p (CDI_DOMINATORS, bb1, bb2))
4724 return true;
4728 /* Iterate to the single executable bb2 successor. */
4729 edge succe = NULL;
4730 FOR_EACH_EDGE (e, ei, bb2->succs)
4731 if ((e->flags & EDGE_EXECUTABLE)
4732 || (!allow_back && (e->flags & EDGE_DFS_BACK)))
4734 if (succe)
4736 succe = NULL;
4737 break;
4739 succe = e;
4741 if (succe)
4743 /* Verify the reached block is only reached through succe.
4744 If there is only one edge we can spare us the dominator
4745 check and iterate directly. */
4746 if (EDGE_COUNT (succe->dest->preds) > 1)
4748 FOR_EACH_EDGE (e, ei, succe->dest->preds)
4749 if (e != succe
4750 && ((e->flags & EDGE_EXECUTABLE)
4751 || (!allow_back && (e->flags & EDGE_DFS_BACK))))
4753 succe = NULL;
4754 break;
4757 if (succe)
4759 bb2 = succe->dest;
4761 /* Re-do the dominance check with changed bb2. */
4762 if (dominated_by_p (CDI_DOMINATORS, bb1, bb2))
4763 return true;
4767 /* We could now iterate updating bb1 / bb2. */
4768 return false;
4771 /* Set the value number of FROM to TO, return true if it has changed
4772 as a result. */
4774 static inline bool
4775 set_ssa_val_to (tree from, tree to)
4777 vn_ssa_aux_t from_info = VN_INFO (from);
4778 tree currval = from_info->valnum; // SSA_VAL (from)
4779 poly_int64 toff, coff;
4780 bool curr_undefined = false;
4781 bool curr_invariant = false;
4783 /* The only thing we allow as value numbers are ssa_names
4784 and invariants. So assert that here. We don't allow VN_TOP
4785 as visiting a stmt should produce a value-number other than
4786 that.
4787 ??? Still VN_TOP can happen for unreachable code, so force
4788 it to varying in that case. Not all code is prepared to
4789 get VN_TOP on valueization. */
4790 if (to == VN_TOP)
4792 /* ??? When iterating and visiting PHI <undef, backedge-value>
4793 for the first time we rightfully get VN_TOP and we need to
4794 preserve that to optimize for example gcc.dg/tree-ssa/ssa-sccvn-2.c.
4795 With SCCVN we were simply lucky we iterated the other PHI
4796 cycles first and thus visited the backedge-value DEF. */
4797 if (currval == VN_TOP)
4798 goto set_and_exit;
4799 if (dump_file && (dump_flags & TDF_DETAILS))
4800 fprintf (dump_file, "Forcing value number to varying on "
4801 "receiving VN_TOP\n");
4802 to = from;
4805 gcc_checking_assert (to != NULL_TREE
4806 && ((TREE_CODE (to) == SSA_NAME
4807 && (to == from || SSA_VAL (to) == to))
4808 || is_gimple_min_invariant (to)));
4810 if (from != to)
4812 if (currval == from)
4814 if (dump_file && (dump_flags & TDF_DETAILS))
4816 fprintf (dump_file, "Not changing value number of ");
4817 print_generic_expr (dump_file, from);
4818 fprintf (dump_file, " from VARYING to ");
4819 print_generic_expr (dump_file, to);
4820 fprintf (dump_file, "\n");
4822 return false;
4824 curr_invariant = is_gimple_min_invariant (currval);
4825 curr_undefined = (TREE_CODE (currval) == SSA_NAME
4826 && ssa_undefined_value_p (currval, false));
4827 if (currval != VN_TOP
4828 && !curr_invariant
4829 && !curr_undefined
4830 && is_gimple_min_invariant (to))
4832 if (dump_file && (dump_flags & TDF_DETAILS))
4834 fprintf (dump_file, "Forcing VARYING instead of changing "
4835 "value number of ");
4836 print_generic_expr (dump_file, from);
4837 fprintf (dump_file, " from ");
4838 print_generic_expr (dump_file, currval);
4839 fprintf (dump_file, " (non-constant) to ");
4840 print_generic_expr (dump_file, to);
4841 fprintf (dump_file, " (constant)\n");
4843 to = from;
4845 else if (currval != VN_TOP
4846 && !curr_undefined
4847 && TREE_CODE (to) == SSA_NAME
4848 && ssa_undefined_value_p (to, false))
4850 if (dump_file && (dump_flags & TDF_DETAILS))
4852 fprintf (dump_file, "Forcing VARYING instead of changing "
4853 "value number of ");
4854 print_generic_expr (dump_file, from);
4855 fprintf (dump_file, " from ");
4856 print_generic_expr (dump_file, currval);
4857 fprintf (dump_file, " (non-undefined) to ");
4858 print_generic_expr (dump_file, to);
4859 fprintf (dump_file, " (undefined)\n");
4861 to = from;
4863 else if (TREE_CODE (to) == SSA_NAME
4864 && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (to))
4865 to = from;
4868 set_and_exit:
4869 if (dump_file && (dump_flags & TDF_DETAILS))
4871 fprintf (dump_file, "Setting value number of ");
4872 print_generic_expr (dump_file, from);
4873 fprintf (dump_file, " to ");
4874 print_generic_expr (dump_file, to);
4877 if (currval != to
4878 && !operand_equal_p (currval, to, 0)
4879 /* Different undefined SSA names are not actually different. See
4880 PR82320 for a testcase were we'd otherwise not terminate iteration. */
4881 && !(curr_undefined
4882 && TREE_CODE (to) == SSA_NAME
4883 && ssa_undefined_value_p (to, false))
4884 /* ??? For addresses involving volatile objects or types operand_equal_p
4885 does not reliably detect ADDR_EXPRs as equal. We know we are only
4886 getting invariant gimple addresses here, so can use
4887 get_addr_base_and_unit_offset to do this comparison. */
4888 && !(TREE_CODE (currval) == ADDR_EXPR
4889 && TREE_CODE (to) == ADDR_EXPR
4890 && (get_addr_base_and_unit_offset (TREE_OPERAND (currval, 0), &coff)
4891 == get_addr_base_and_unit_offset (TREE_OPERAND (to, 0), &toff))
4892 && known_eq (coff, toff)))
4894 if (to != from
4895 && currval != VN_TOP
4896 && !curr_undefined
4897 /* We do not want to allow lattice transitions from one value
4898 to another since that may lead to not terminating iteration
4899 (see PR95049). Since there's no convenient way to check
4900 for the allowed transition of VAL -> PHI (loop entry value,
4901 same on two PHIs, to same PHI result) we restrict the check
4902 to invariants. */
4903 && curr_invariant
4904 && is_gimple_min_invariant (to))
4906 if (dump_file && (dump_flags & TDF_DETAILS))
4907 fprintf (dump_file, " forced VARYING");
4908 to = from;
4910 if (dump_file && (dump_flags & TDF_DETAILS))
4911 fprintf (dump_file, " (changed)\n");
4912 from_info->valnum = to;
4913 return true;
4915 if (dump_file && (dump_flags & TDF_DETAILS))
4916 fprintf (dump_file, "\n");
4917 return false;
4920 /* Set all definitions in STMT to value number to themselves.
4921 Return true if a value number changed. */
4923 static bool
4924 defs_to_varying (gimple *stmt)
4926 bool changed = false;
4927 ssa_op_iter iter;
4928 def_operand_p defp;
4930 FOR_EACH_SSA_DEF_OPERAND (defp, stmt, iter, SSA_OP_ALL_DEFS)
4932 tree def = DEF_FROM_PTR (defp);
4933 changed |= set_ssa_val_to (def, def);
4935 return changed;
4938 /* Visit a copy between LHS and RHS, return true if the value number
4939 changed. */
4941 static bool
4942 visit_copy (tree lhs, tree rhs)
4944 /* Valueize. */
4945 rhs = SSA_VAL (rhs);
4947 return set_ssa_val_to (lhs, rhs);
4950 /* Lookup a value for OP in type WIDE_TYPE where the value in type of OP
4951 is the same. */
4953 static tree
4954 valueized_wider_op (tree wide_type, tree op, bool allow_truncate)
4956 if (TREE_CODE (op) == SSA_NAME)
4957 op = vn_valueize (op);
4959 /* Either we have the op widened available. */
4960 tree ops[3] = {};
4961 ops[0] = op;
4962 tree tem = vn_nary_op_lookup_pieces (1, NOP_EXPR,
4963 wide_type, ops, NULL);
4964 if (tem)
4965 return tem;
4967 /* Or the op is truncated from some existing value. */
4968 if (allow_truncate && TREE_CODE (op) == SSA_NAME)
4970 gimple *def = SSA_NAME_DEF_STMT (op);
4971 if (is_gimple_assign (def)
4972 && CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (def)))
4974 tem = gimple_assign_rhs1 (def);
4975 if (useless_type_conversion_p (wide_type, TREE_TYPE (tem)))
4977 if (TREE_CODE (tem) == SSA_NAME)
4978 tem = vn_valueize (tem);
4979 return tem;
4984 /* For constants simply extend it. */
4985 if (TREE_CODE (op) == INTEGER_CST)
4986 return wide_int_to_tree (wide_type, wi::to_wide (op));
4988 return NULL_TREE;
4991 /* Visit a nary operator RHS, value number it, and return true if the
4992 value number of LHS has changed as a result. */
4994 static bool
4995 visit_nary_op (tree lhs, gassign *stmt)
4997 vn_nary_op_t vnresult;
4998 tree result = vn_nary_op_lookup_stmt (stmt, &vnresult);
4999 if (! result && vnresult)
5000 result = vn_nary_op_get_predicated_value (vnresult, gimple_bb (stmt));
5001 if (result)
5002 return set_ssa_val_to (lhs, result);
5004 /* Do some special pattern matching for redundancies of operations
5005 in different types. */
5006 enum tree_code code = gimple_assign_rhs_code (stmt);
5007 tree type = TREE_TYPE (lhs);
5008 tree rhs1 = gimple_assign_rhs1 (stmt);
5009 switch (code)
5011 CASE_CONVERT:
5012 /* Match arithmetic done in a different type where we can easily
5013 substitute the result from some earlier sign-changed or widened
5014 operation. */
5015 if (INTEGRAL_TYPE_P (type)
5016 && TREE_CODE (rhs1) == SSA_NAME
5017 /* We only handle sign-changes, zero-extension -> & mask or
5018 sign-extension if we know the inner operation doesn't
5019 overflow. */
5020 && (((TYPE_UNSIGNED (TREE_TYPE (rhs1))
5021 || (INTEGRAL_TYPE_P (TREE_TYPE (rhs1))
5022 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (rhs1))))
5023 && TYPE_PRECISION (type) > TYPE_PRECISION (TREE_TYPE (rhs1)))
5024 || TYPE_PRECISION (type) == TYPE_PRECISION (TREE_TYPE (rhs1))))
5026 gassign *def = dyn_cast <gassign *> (SSA_NAME_DEF_STMT (rhs1));
5027 if (def
5028 && (gimple_assign_rhs_code (def) == PLUS_EXPR
5029 || gimple_assign_rhs_code (def) == MINUS_EXPR
5030 || gimple_assign_rhs_code (def) == MULT_EXPR))
5032 tree ops[3] = {};
5033 /* When requiring a sign-extension we cannot model a
5034 previous truncation with a single op so don't bother. */
5035 bool allow_truncate = TYPE_UNSIGNED (TREE_TYPE (rhs1));
5036 /* Either we have the op widened available. */
5037 ops[0] = valueized_wider_op (type, gimple_assign_rhs1 (def),
5038 allow_truncate);
5039 if (ops[0])
5040 ops[1] = valueized_wider_op (type, gimple_assign_rhs2 (def),
5041 allow_truncate);
5042 if (ops[0] && ops[1])
5044 ops[0] = vn_nary_op_lookup_pieces
5045 (2, gimple_assign_rhs_code (def), type, ops, NULL);
5046 /* We have wider operation available. */
5047 if (ops[0]
5048 /* If the leader is a wrapping operation we can
5049 insert it for code hoisting w/o introducing
5050 undefined overflow. If it is not it has to
5051 be available. See PR86554. */
5052 && (TYPE_OVERFLOW_WRAPS (TREE_TYPE (ops[0]))
5053 || (rpo_avail && vn_context_bb
5054 && rpo_avail->eliminate_avail (vn_context_bb,
5055 ops[0]))))
5057 unsigned lhs_prec = TYPE_PRECISION (type);
5058 unsigned rhs_prec = TYPE_PRECISION (TREE_TYPE (rhs1));
5059 if (lhs_prec == rhs_prec
5060 || (INTEGRAL_TYPE_P (TREE_TYPE (rhs1))
5061 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (rhs1))))
5063 gimple_match_op match_op (gimple_match_cond::UNCOND,
5064 NOP_EXPR, type, ops[0]);
5065 result = vn_nary_build_or_lookup (&match_op);
5066 if (result)
5068 bool changed = set_ssa_val_to (lhs, result);
5069 vn_nary_op_insert_stmt (stmt, result);
5070 return changed;
5073 else
5075 tree mask = wide_int_to_tree
5076 (type, wi::mask (rhs_prec, false, lhs_prec));
5077 gimple_match_op match_op (gimple_match_cond::UNCOND,
5078 BIT_AND_EXPR,
5079 TREE_TYPE (lhs),
5080 ops[0], mask);
5081 result = vn_nary_build_or_lookup (&match_op);
5082 if (result)
5084 bool changed = set_ssa_val_to (lhs, result);
5085 vn_nary_op_insert_stmt (stmt, result);
5086 return changed;
5093 break;
5094 case BIT_AND_EXPR:
5095 if (INTEGRAL_TYPE_P (type)
5096 && TREE_CODE (rhs1) == SSA_NAME
5097 && TREE_CODE (gimple_assign_rhs2 (stmt)) == INTEGER_CST
5098 && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (rhs1)
5099 && default_vn_walk_kind != VN_NOWALK
5100 && CHAR_BIT == 8
5101 && BITS_PER_UNIT == 8
5102 && BYTES_BIG_ENDIAN == WORDS_BIG_ENDIAN
5103 && !integer_all_onesp (gimple_assign_rhs2 (stmt))
5104 && !integer_zerop (gimple_assign_rhs2 (stmt)))
5106 gassign *ass = dyn_cast <gassign *> (SSA_NAME_DEF_STMT (rhs1));
5107 if (ass
5108 && !gimple_has_volatile_ops (ass)
5109 && vn_get_stmt_kind (ass) == VN_REFERENCE)
5111 tree last_vuse = gimple_vuse (ass);
5112 tree op = gimple_assign_rhs1 (ass);
5113 tree result = vn_reference_lookup (op, gimple_vuse (ass),
5114 default_vn_walk_kind,
5115 NULL, true, &last_vuse,
5116 gimple_assign_rhs2 (stmt));
5117 if (result
5118 && useless_type_conversion_p (TREE_TYPE (result),
5119 TREE_TYPE (op)))
5120 return set_ssa_val_to (lhs, result);
5123 break;
5124 case TRUNC_DIV_EXPR:
5125 if (TYPE_UNSIGNED (type))
5126 break;
5127 /* Fallthru. */
5128 case RDIV_EXPR:
5129 case MULT_EXPR:
5130 /* Match up ([-]a){/,*}([-])b with v=a{/,*}b, replacing it with -v. */
5131 if (! HONOR_SIGN_DEPENDENT_ROUNDING (type))
5133 tree rhs[2];
5134 rhs[0] = rhs1;
5135 rhs[1] = gimple_assign_rhs2 (stmt);
5136 for (unsigned i = 0; i <= 1; ++i)
5138 unsigned j = i == 0 ? 1 : 0;
5139 tree ops[2];
5140 gimple_match_op match_op (gimple_match_cond::UNCOND,
5141 NEGATE_EXPR, type, rhs[i]);
5142 ops[i] = vn_nary_build_or_lookup_1 (&match_op, false, true);
5143 ops[j] = rhs[j];
5144 if (ops[i]
5145 && (ops[0] = vn_nary_op_lookup_pieces (2, code,
5146 type, ops, NULL)))
5148 gimple_match_op match_op (gimple_match_cond::UNCOND,
5149 NEGATE_EXPR, type, ops[0]);
5150 result = vn_nary_build_or_lookup_1 (&match_op, true, false);
5151 if (result)
5153 bool changed = set_ssa_val_to (lhs, result);
5154 vn_nary_op_insert_stmt (stmt, result);
5155 return changed;
5160 break;
5161 default:
5162 break;
5165 bool changed = set_ssa_val_to (lhs, lhs);
5166 vn_nary_op_insert_stmt (stmt, lhs);
5167 return changed;
5170 /* Visit a call STMT storing into LHS. Return true if the value number
5171 of the LHS has changed as a result. */
5173 static bool
5174 visit_reference_op_call (tree lhs, gcall *stmt)
5176 bool changed = false;
5177 struct vn_reference_s vr1;
5178 vn_reference_t vnresult = NULL;
5179 tree vdef = gimple_vdef (stmt);
5180 modref_summary *summary;
5182 /* Non-ssa lhs is handled in copy_reference_ops_from_call. */
5183 if (lhs && TREE_CODE (lhs) != SSA_NAME)
5184 lhs = NULL_TREE;
5186 vn_reference_lookup_call (stmt, &vnresult, &vr1);
5188 /* If the lookup did not succeed for pure functions try to use
5189 modref info to find a candidate to CSE to. */
5190 const unsigned accesses_limit = 8;
5191 if (!vnresult
5192 && !vdef
5193 && lhs
5194 && gimple_vuse (stmt)
5195 && (((summary = get_modref_function_summary (stmt, NULL))
5196 && !summary->global_memory_read
5197 && summary->load_accesses < accesses_limit)
5198 || gimple_call_flags (stmt) & ECF_CONST))
5200 /* First search if we can do someting useful and build a
5201 vector of all loads we have to check. */
5202 bool unknown_memory_access = false;
5203 auto_vec<ao_ref, accesses_limit> accesses;
5204 unsigned load_accesses = summary ? summary->load_accesses : 0;
5205 if (!unknown_memory_access)
5206 /* Add loads done as part of setting up the call arguments.
5207 That's also necessary for CONST functions which will
5208 not have a modref summary. */
5209 for (unsigned i = 0; i < gimple_call_num_args (stmt); ++i)
5211 tree arg = gimple_call_arg (stmt, i);
5212 if (TREE_CODE (arg) != SSA_NAME
5213 && !is_gimple_min_invariant (arg))
5215 if (accesses.length () >= accesses_limit - load_accesses)
5217 unknown_memory_access = true;
5218 break;
5220 accesses.quick_grow (accesses.length () + 1);
5221 ao_ref_init (&accesses.last (), arg);
5224 if (summary && !unknown_memory_access)
5226 /* Add loads as analyzed by IPA modref. */
5227 for (auto base_node : summary->loads->bases)
5228 if (unknown_memory_access)
5229 break;
5230 else for (auto ref_node : base_node->refs)
5231 if (unknown_memory_access)
5232 break;
5233 else for (auto access_node : ref_node->accesses)
5235 accesses.quick_grow (accesses.length () + 1);
5236 ao_ref *r = &accesses.last ();
5237 if (!access_node.get_ao_ref (stmt, r))
5239 /* Initialize a ref based on the argument and
5240 unknown offset if possible. */
5241 tree arg = access_node.get_call_arg (stmt);
5242 if (arg && TREE_CODE (arg) == SSA_NAME)
5243 arg = SSA_VAL (arg);
5244 if (arg
5245 && TREE_CODE (arg) == ADDR_EXPR
5246 && (arg = get_base_address (arg))
5247 && DECL_P (arg))
5249 ao_ref_init (r, arg);
5250 r->ref = NULL_TREE;
5251 r->base = arg;
5253 else
5255 unknown_memory_access = true;
5256 break;
5259 r->base_alias_set = base_node->base;
5260 r->ref_alias_set = ref_node->ref;
5264 /* Walk the VUSE->VDEF chain optimistically trying to find an entry
5265 for the call in the hashtable. */
5266 unsigned limit = (unknown_memory_access
5268 : (param_sccvn_max_alias_queries_per_access
5269 / (accesses.length () + 1)));
5270 tree saved_vuse = vr1.vuse;
5271 hashval_t saved_hashcode = vr1.hashcode;
5272 while (limit > 0 && !vnresult && !SSA_NAME_IS_DEFAULT_DEF (vr1.vuse))
5274 vr1.hashcode = vr1.hashcode - SSA_NAME_VERSION (vr1.vuse);
5275 gimple *def = SSA_NAME_DEF_STMT (vr1.vuse);
5276 /* ??? We could use fancy stuff like in walk_non_aliased_vuses, but
5277 do not bother for now. */
5278 if (is_a <gphi *> (def))
5279 break;
5280 vr1.vuse = vuse_ssa_val (gimple_vuse (def));
5281 vr1.hashcode = vr1.hashcode + SSA_NAME_VERSION (vr1.vuse);
5282 vn_reference_lookup_1 (&vr1, &vnresult);
5283 limit--;
5286 /* If we found a candidate to CSE to verify it is valid. */
5287 if (vnresult && !accesses.is_empty ())
5289 tree vuse = vuse_ssa_val (gimple_vuse (stmt));
5290 while (vnresult && vuse != vr1.vuse)
5292 gimple *def = SSA_NAME_DEF_STMT (vuse);
5293 for (auto &ref : accesses)
5295 /* ??? stmt_may_clobber_ref_p_1 does per stmt constant
5296 analysis overhead that we might be able to cache. */
5297 if (stmt_may_clobber_ref_p_1 (def, &ref, true))
5299 vnresult = NULL;
5300 break;
5303 vuse = vuse_ssa_val (gimple_vuse (def));
5306 vr1.vuse = saved_vuse;
5307 vr1.hashcode = saved_hashcode;
5310 if (vnresult)
5312 if (vdef)
5314 if (vnresult->result_vdef)
5315 changed |= set_ssa_val_to (vdef, vnresult->result_vdef);
5316 else if (!lhs && gimple_call_lhs (stmt))
5317 /* If stmt has non-SSA_NAME lhs, value number the vdef to itself,
5318 as the call still acts as a lhs store. */
5319 changed |= set_ssa_val_to (vdef, vdef);
5320 else
5321 /* If the call was discovered to be pure or const reflect
5322 that as far as possible. */
5323 changed |= set_ssa_val_to (vdef,
5324 vuse_ssa_val (gimple_vuse (stmt)));
5327 if (!vnresult->result && lhs)
5328 vnresult->result = lhs;
5330 if (vnresult->result && lhs)
5331 changed |= set_ssa_val_to (lhs, vnresult->result);
5333 else
5335 vn_reference_t vr2;
5336 vn_reference_s **slot;
5337 tree vdef_val = vdef;
5338 if (vdef)
5340 /* If we value numbered an indirect functions function to
5341 one not clobbering memory value number its VDEF to its
5342 VUSE. */
5343 tree fn = gimple_call_fn (stmt);
5344 if (fn && TREE_CODE (fn) == SSA_NAME)
5346 fn = SSA_VAL (fn);
5347 if (TREE_CODE (fn) == ADDR_EXPR
5348 && TREE_CODE (TREE_OPERAND (fn, 0)) == FUNCTION_DECL
5349 && (flags_from_decl_or_type (TREE_OPERAND (fn, 0))
5350 & (ECF_CONST | ECF_PURE))
5351 /* If stmt has non-SSA_NAME lhs, value number the
5352 vdef to itself, as the call still acts as a lhs
5353 store. */
5354 && (lhs || gimple_call_lhs (stmt) == NULL_TREE))
5355 vdef_val = vuse_ssa_val (gimple_vuse (stmt));
5357 changed |= set_ssa_val_to (vdef, vdef_val);
5359 if (lhs)
5360 changed |= set_ssa_val_to (lhs, lhs);
5361 vr2 = XOBNEW (&vn_tables_obstack, vn_reference_s);
5362 vr2->vuse = vr1.vuse;
5363 /* As we are not walking the virtual operand chain we know the
5364 shared_lookup_references are still original so we can re-use
5365 them here. */
5366 vr2->operands = vr1.operands.copy ();
5367 vr2->type = vr1.type;
5368 vr2->punned = vr1.punned;
5369 vr2->set = vr1.set;
5370 vr2->base_set = vr1.base_set;
5371 vr2->hashcode = vr1.hashcode;
5372 vr2->result = lhs;
5373 vr2->result_vdef = vdef_val;
5374 vr2->value_id = 0;
5375 slot = valid_info->references->find_slot_with_hash (vr2, vr2->hashcode,
5376 INSERT);
5377 gcc_assert (!*slot);
5378 *slot = vr2;
5379 vr2->next = last_inserted_ref;
5380 last_inserted_ref = vr2;
5383 return changed;
5386 /* Visit a load from a reference operator RHS, part of STMT, value number it,
5387 and return true if the value number of the LHS has changed as a result. */
5389 static bool
5390 visit_reference_op_load (tree lhs, tree op, gimple *stmt)
5392 bool changed = false;
5393 tree result;
5394 vn_reference_t res;
5396 tree vuse = gimple_vuse (stmt);
5397 tree last_vuse = vuse;
5398 result = vn_reference_lookup (op, vuse, default_vn_walk_kind, &res, true, &last_vuse);
5400 /* We handle type-punning through unions by value-numbering based
5401 on offset and size of the access. Be prepared to handle a
5402 type-mismatch here via creating a VIEW_CONVERT_EXPR. */
5403 if (result
5404 && !useless_type_conversion_p (TREE_TYPE (result), TREE_TYPE (op)))
5406 /* Avoid the type punning in case the result mode has padding where
5407 the op we lookup has not. */
5408 if (maybe_lt (GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (result))),
5409 GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (op)))))
5410 result = NULL_TREE;
5411 else
5413 /* We will be setting the value number of lhs to the value number
5414 of VIEW_CONVERT_EXPR <TREE_TYPE (result)> (result).
5415 So first simplify and lookup this expression to see if it
5416 is already available. */
5417 gimple_match_op res_op (gimple_match_cond::UNCOND,
5418 VIEW_CONVERT_EXPR, TREE_TYPE (op), result);
5419 result = vn_nary_build_or_lookup (&res_op);
5420 if (result
5421 && TREE_CODE (result) == SSA_NAME
5422 && VN_INFO (result)->needs_insertion)
5423 /* Track whether this is the canonical expression for different
5424 typed loads. We use that as a stopgap measure for code
5425 hoisting when dealing with floating point loads. */
5426 res->punned = true;
5429 /* When building the conversion fails avoid inserting the reference
5430 again. */
5431 if (!result)
5432 return set_ssa_val_to (lhs, lhs);
5435 if (result)
5436 changed = set_ssa_val_to (lhs, result);
5437 else
5439 changed = set_ssa_val_to (lhs, lhs);
5440 vn_reference_insert (op, lhs, last_vuse, NULL_TREE);
5441 if (vuse && SSA_VAL (last_vuse) != SSA_VAL (vuse))
5443 if (dump_file && (dump_flags & TDF_DETAILS))
5445 fprintf (dump_file, "Using extra use virtual operand ");
5446 print_generic_expr (dump_file, last_vuse);
5447 fprintf (dump_file, "\n");
5449 vn_reference_insert (op, lhs, vuse, NULL_TREE);
5453 return changed;
5457 /* Visit a store to a reference operator LHS, part of STMT, value number it,
5458 and return true if the value number of the LHS has changed as a result. */
5460 static bool
5461 visit_reference_op_store (tree lhs, tree op, gimple *stmt)
5463 bool changed = false;
5464 vn_reference_t vnresult = NULL;
5465 tree assign;
5466 bool resultsame = false;
5467 tree vuse = gimple_vuse (stmt);
5468 tree vdef = gimple_vdef (stmt);
5470 if (TREE_CODE (op) == SSA_NAME)
5471 op = SSA_VAL (op);
5473 /* First we want to lookup using the *vuses* from the store and see
5474 if there the last store to this location with the same address
5475 had the same value.
5477 The vuses represent the memory state before the store. If the
5478 memory state, address, and value of the store is the same as the
5479 last store to this location, then this store will produce the
5480 same memory state as that store.
5482 In this case the vdef versions for this store are value numbered to those
5483 vuse versions, since they represent the same memory state after
5484 this store.
5486 Otherwise, the vdefs for the store are used when inserting into
5487 the table, since the store generates a new memory state. */
5489 vn_reference_lookup (lhs, vuse, VN_NOWALK, &vnresult, false);
5490 if (vnresult
5491 && vnresult->result)
5493 tree result = vnresult->result;
5494 gcc_checking_assert (TREE_CODE (result) != SSA_NAME
5495 || result == SSA_VAL (result));
5496 resultsame = expressions_equal_p (result, op);
5497 if (resultsame)
5499 /* If the TBAA state isn't compatible for downstream reads
5500 we cannot value-number the VDEFs the same. */
5501 ao_ref lhs_ref;
5502 ao_ref_init (&lhs_ref, lhs);
5503 alias_set_type set = ao_ref_alias_set (&lhs_ref);
5504 alias_set_type base_set = ao_ref_base_alias_set (&lhs_ref);
5505 if ((vnresult->set != set
5506 && ! alias_set_subset_of (set, vnresult->set))
5507 || (vnresult->base_set != base_set
5508 && ! alias_set_subset_of (base_set, vnresult->base_set)))
5509 resultsame = false;
5513 if (!resultsame)
5515 /* Only perform the following when being called from PRE
5516 which embeds tail merging. */
5517 if (default_vn_walk_kind == VN_WALK)
5519 assign = build2 (MODIFY_EXPR, TREE_TYPE (lhs), lhs, op);
5520 vn_reference_lookup (assign, vuse, VN_NOWALK, &vnresult, false);
5521 if (vnresult)
5523 VN_INFO (vdef)->visited = true;
5524 return set_ssa_val_to (vdef, vnresult->result_vdef);
5528 if (dump_file && (dump_flags & TDF_DETAILS))
5530 fprintf (dump_file, "No store match\n");
5531 fprintf (dump_file, "Value numbering store ");
5532 print_generic_expr (dump_file, lhs);
5533 fprintf (dump_file, " to ");
5534 print_generic_expr (dump_file, op);
5535 fprintf (dump_file, "\n");
5537 /* Have to set value numbers before insert, since insert is
5538 going to valueize the references in-place. */
5539 if (vdef)
5540 changed |= set_ssa_val_to (vdef, vdef);
5542 /* Do not insert structure copies into the tables. */
5543 if (is_gimple_min_invariant (op)
5544 || is_gimple_reg (op))
5545 vn_reference_insert (lhs, op, vdef, NULL);
5547 /* Only perform the following when being called from PRE
5548 which embeds tail merging. */
5549 if (default_vn_walk_kind == VN_WALK)
5551 assign = build2 (MODIFY_EXPR, TREE_TYPE (lhs), lhs, op);
5552 vn_reference_insert (assign, lhs, vuse, vdef);
5555 else
5557 /* We had a match, so value number the vdef to have the value
5558 number of the vuse it came from. */
5560 if (dump_file && (dump_flags & TDF_DETAILS))
5561 fprintf (dump_file, "Store matched earlier value, "
5562 "value numbering store vdefs to matching vuses.\n");
5564 changed |= set_ssa_val_to (vdef, SSA_VAL (vuse));
5567 return changed;
5570 /* Visit and value number PHI, return true if the value number
5571 changed. When BACKEDGES_VARYING_P is true then assume all
5572 backedge values are varying. When INSERTED is not NULL then
5573 this is just a ahead query for a possible iteration, set INSERTED
5574 to true if we'd insert into the hashtable. */
5576 static bool
5577 visit_phi (gimple *phi, bool *inserted, bool backedges_varying_p)
5579 tree result, sameval = VN_TOP, seen_undef = NULL_TREE;
5580 tree backedge_val = NULL_TREE;
5581 bool seen_non_backedge = false;
5582 tree sameval_base = NULL_TREE;
5583 poly_int64 soff, doff;
5584 unsigned n_executable = 0;
5585 edge_iterator ei;
5586 edge e;
5588 /* TODO: We could check for this in initialization, and replace this
5589 with a gcc_assert. */
5590 if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (PHI_RESULT (phi)))
5591 return set_ssa_val_to (PHI_RESULT (phi), PHI_RESULT (phi));
5593 /* We track whether a PHI was CSEd to to avoid excessive iterations
5594 that would be necessary only because the PHI changed arguments
5595 but not value. */
5596 if (!inserted)
5597 gimple_set_plf (phi, GF_PLF_1, false);
5599 /* See if all non-TOP arguments have the same value. TOP is
5600 equivalent to everything, so we can ignore it. */
5601 FOR_EACH_EDGE (e, ei, gimple_bb (phi)->preds)
5602 if (e->flags & EDGE_EXECUTABLE)
5604 tree def = PHI_ARG_DEF_FROM_EDGE (phi, e);
5606 if (def == PHI_RESULT (phi))
5607 continue;
5608 ++n_executable;
5609 if (TREE_CODE (def) == SSA_NAME)
5611 if (!backedges_varying_p || !(e->flags & EDGE_DFS_BACK))
5612 def = SSA_VAL (def);
5613 if (e->flags & EDGE_DFS_BACK)
5614 backedge_val = def;
5616 if (!(e->flags & EDGE_DFS_BACK))
5617 seen_non_backedge = true;
5618 if (def == VN_TOP)
5620 /* Ignore undefined defs for sameval but record one. */
5621 else if (TREE_CODE (def) == SSA_NAME
5622 && ! virtual_operand_p (def)
5623 && ssa_undefined_value_p (def, false))
5624 seen_undef = def;
5625 else if (sameval == VN_TOP)
5626 sameval = def;
5627 else if (!expressions_equal_p (def, sameval))
5629 /* We know we're arriving only with invariant addresses here,
5630 try harder comparing them. We can do some caching here
5631 which we cannot do in expressions_equal_p. */
5632 if (TREE_CODE (def) == ADDR_EXPR
5633 && TREE_CODE (sameval) == ADDR_EXPR
5634 && sameval_base != (void *)-1)
5636 if (!sameval_base)
5637 sameval_base = get_addr_base_and_unit_offset
5638 (TREE_OPERAND (sameval, 0), &soff);
5639 if (!sameval_base)
5640 sameval_base = (tree)(void *)-1;
5641 else if ((get_addr_base_and_unit_offset
5642 (TREE_OPERAND (def, 0), &doff) == sameval_base)
5643 && known_eq (soff, doff))
5644 continue;
5646 sameval = NULL_TREE;
5647 break;
5651 /* If the value we want to use is flowing over the backedge and we
5652 should take it as VARYING but it has a non-VARYING value drop to
5653 VARYING.
5654 If we value-number a virtual operand never value-number to the
5655 value from the backedge as that confuses the alias-walking code.
5656 See gcc.dg/torture/pr87176.c. If the value is the same on a
5657 non-backedge everything is OK though. */
5658 bool visited_p;
5659 if ((backedge_val
5660 && !seen_non_backedge
5661 && TREE_CODE (backedge_val) == SSA_NAME
5662 && sameval == backedge_val
5663 && (SSA_NAME_IS_VIRTUAL_OPERAND (backedge_val)
5664 || SSA_VAL (backedge_val) != backedge_val))
5665 /* Do not value-number a virtual operand to sth not visited though
5666 given that allows us to escape a region in alias walking. */
5667 || (sameval
5668 && TREE_CODE (sameval) == SSA_NAME
5669 && !SSA_NAME_IS_DEFAULT_DEF (sameval)
5670 && SSA_NAME_IS_VIRTUAL_OPERAND (sameval)
5671 && (SSA_VAL (sameval, &visited_p), !visited_p)))
5672 /* Note this just drops to VARYING without inserting the PHI into
5673 the hashes. */
5674 result = PHI_RESULT (phi);
5675 /* If none of the edges was executable keep the value-number at VN_TOP,
5676 if only a single edge is exectuable use its value. */
5677 else if (n_executable <= 1)
5678 result = seen_undef ? seen_undef : sameval;
5679 /* If we saw only undefined values and VN_TOP use one of the
5680 undefined values. */
5681 else if (sameval == VN_TOP)
5682 result = seen_undef ? seen_undef : sameval;
5683 /* First see if it is equivalent to a phi node in this block. We prefer
5684 this as it allows IV elimination - see PRs 66502 and 67167. */
5685 else if ((result = vn_phi_lookup (phi, backedges_varying_p)))
5687 if (!inserted
5688 && TREE_CODE (result) == SSA_NAME
5689 && gimple_code (SSA_NAME_DEF_STMT (result)) == GIMPLE_PHI)
5691 gimple_set_plf (SSA_NAME_DEF_STMT (result), GF_PLF_1, true);
5692 if (dump_file && (dump_flags & TDF_DETAILS))
5694 fprintf (dump_file, "Marking CSEd to PHI node ");
5695 print_gimple_expr (dump_file, SSA_NAME_DEF_STMT (result),
5696 0, TDF_SLIM);
5697 fprintf (dump_file, "\n");
5701 /* If all values are the same use that, unless we've seen undefined
5702 values as well and the value isn't constant.
5703 CCP/copyprop have the same restriction to not remove uninit warnings. */
5704 else if (sameval
5705 && (! seen_undef || is_gimple_min_invariant (sameval)))
5706 result = sameval;
5707 else
5709 result = PHI_RESULT (phi);
5710 /* Only insert PHIs that are varying, for constant value numbers
5711 we mess up equivalences otherwise as we are only comparing
5712 the immediate controlling predicates. */
5713 vn_phi_insert (phi, result, backedges_varying_p);
5714 if (inserted)
5715 *inserted = true;
5718 return set_ssa_val_to (PHI_RESULT (phi), result);
5721 /* Try to simplify RHS using equivalences and constant folding. */
5723 static tree
5724 try_to_simplify (gassign *stmt)
5726 enum tree_code code = gimple_assign_rhs_code (stmt);
5727 tree tem;
5729 /* For stores we can end up simplifying a SSA_NAME rhs. Just return
5730 in this case, there is no point in doing extra work. */
5731 if (code == SSA_NAME)
5732 return NULL_TREE;
5734 /* First try constant folding based on our current lattice. */
5735 mprts_hook = vn_lookup_simplify_result;
5736 tem = gimple_fold_stmt_to_constant_1 (stmt, vn_valueize, vn_valueize);
5737 mprts_hook = NULL;
5738 if (tem
5739 && (TREE_CODE (tem) == SSA_NAME
5740 || is_gimple_min_invariant (tem)))
5741 return tem;
5743 return NULL_TREE;
5746 /* Visit and value number STMT, return true if the value number
5747 changed. */
5749 static bool
5750 visit_stmt (gimple *stmt, bool backedges_varying_p = false)
5752 bool changed = false;
5754 if (dump_file && (dump_flags & TDF_DETAILS))
5756 fprintf (dump_file, "Value numbering stmt = ");
5757 print_gimple_stmt (dump_file, stmt, 0);
5760 if (gimple_code (stmt) == GIMPLE_PHI)
5761 changed = visit_phi (stmt, NULL, backedges_varying_p);
5762 else if (gimple_has_volatile_ops (stmt))
5763 changed = defs_to_varying (stmt);
5764 else if (gassign *ass = dyn_cast <gassign *> (stmt))
5766 enum tree_code code = gimple_assign_rhs_code (ass);
5767 tree lhs = gimple_assign_lhs (ass);
5768 tree rhs1 = gimple_assign_rhs1 (ass);
5769 tree simplified;
5771 /* Shortcut for copies. Simplifying copies is pointless,
5772 since we copy the expression and value they represent. */
5773 if (code == SSA_NAME
5774 && TREE_CODE (lhs) == SSA_NAME)
5776 changed = visit_copy (lhs, rhs1);
5777 goto done;
5779 simplified = try_to_simplify (ass);
5780 if (simplified)
5782 if (dump_file && (dump_flags & TDF_DETAILS))
5784 fprintf (dump_file, "RHS ");
5785 print_gimple_expr (dump_file, ass, 0);
5786 fprintf (dump_file, " simplified to ");
5787 print_generic_expr (dump_file, simplified);
5788 fprintf (dump_file, "\n");
5791 /* Setting value numbers to constants will occasionally
5792 screw up phi congruence because constants are not
5793 uniquely associated with a single ssa name that can be
5794 looked up. */
5795 if (simplified
5796 && is_gimple_min_invariant (simplified)
5797 && TREE_CODE (lhs) == SSA_NAME)
5799 changed = set_ssa_val_to (lhs, simplified);
5800 goto done;
5802 else if (simplified
5803 && TREE_CODE (simplified) == SSA_NAME
5804 && TREE_CODE (lhs) == SSA_NAME)
5806 changed = visit_copy (lhs, simplified);
5807 goto done;
5810 if ((TREE_CODE (lhs) == SSA_NAME
5811 /* We can substitute SSA_NAMEs that are live over
5812 abnormal edges with their constant value. */
5813 && !(gimple_assign_copy_p (ass)
5814 && is_gimple_min_invariant (rhs1))
5815 && !(simplified
5816 && is_gimple_min_invariant (simplified))
5817 && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs))
5818 /* Stores or copies from SSA_NAMEs that are live over
5819 abnormal edges are a problem. */
5820 || (code == SSA_NAME
5821 && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (rhs1)))
5822 changed = defs_to_varying (ass);
5823 else if (REFERENCE_CLASS_P (lhs)
5824 || DECL_P (lhs))
5825 changed = visit_reference_op_store (lhs, rhs1, ass);
5826 else if (TREE_CODE (lhs) == SSA_NAME)
5828 if ((gimple_assign_copy_p (ass)
5829 && is_gimple_min_invariant (rhs1))
5830 || (simplified
5831 && is_gimple_min_invariant (simplified)))
5833 if (simplified)
5834 changed = set_ssa_val_to (lhs, simplified);
5835 else
5836 changed = set_ssa_val_to (lhs, rhs1);
5838 else
5840 /* Visit the original statement. */
5841 switch (vn_get_stmt_kind (ass))
5843 case VN_NARY:
5844 changed = visit_nary_op (lhs, ass);
5845 break;
5846 case VN_REFERENCE:
5847 changed = visit_reference_op_load (lhs, rhs1, ass);
5848 break;
5849 default:
5850 changed = defs_to_varying (ass);
5851 break;
5855 else
5856 changed = defs_to_varying (ass);
5858 else if (gcall *call_stmt = dyn_cast <gcall *> (stmt))
5860 tree lhs = gimple_call_lhs (call_stmt);
5861 if (lhs && TREE_CODE (lhs) == SSA_NAME)
5863 /* Try constant folding based on our current lattice. */
5864 tree simplified = gimple_fold_stmt_to_constant_1 (call_stmt,
5865 vn_valueize);
5866 if (simplified)
5868 if (dump_file && (dump_flags & TDF_DETAILS))
5870 fprintf (dump_file, "call ");
5871 print_gimple_expr (dump_file, call_stmt, 0);
5872 fprintf (dump_file, " simplified to ");
5873 print_generic_expr (dump_file, simplified);
5874 fprintf (dump_file, "\n");
5877 /* Setting value numbers to constants will occasionally
5878 screw up phi congruence because constants are not
5879 uniquely associated with a single ssa name that can be
5880 looked up. */
5881 if (simplified
5882 && is_gimple_min_invariant (simplified))
5884 changed = set_ssa_val_to (lhs, simplified);
5885 if (gimple_vdef (call_stmt))
5886 changed |= set_ssa_val_to (gimple_vdef (call_stmt),
5887 SSA_VAL (gimple_vuse (call_stmt)));
5888 goto done;
5890 else if (simplified
5891 && TREE_CODE (simplified) == SSA_NAME)
5893 changed = visit_copy (lhs, simplified);
5894 if (gimple_vdef (call_stmt))
5895 changed |= set_ssa_val_to (gimple_vdef (call_stmt),
5896 SSA_VAL (gimple_vuse (call_stmt)));
5897 goto done;
5899 else if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs))
5901 changed = defs_to_varying (call_stmt);
5902 goto done;
5906 /* Pick up flags from a devirtualization target. */
5907 tree fn = gimple_call_fn (stmt);
5908 int extra_fnflags = 0;
5909 if (fn && TREE_CODE (fn) == SSA_NAME)
5911 fn = SSA_VAL (fn);
5912 if (TREE_CODE (fn) == ADDR_EXPR
5913 && TREE_CODE (TREE_OPERAND (fn, 0)) == FUNCTION_DECL)
5914 extra_fnflags = flags_from_decl_or_type (TREE_OPERAND (fn, 0));
5916 if ((/* Calls to the same function with the same vuse
5917 and the same operands do not necessarily return the same
5918 value, unless they're pure or const. */
5919 ((gimple_call_flags (call_stmt) | extra_fnflags)
5920 & (ECF_PURE | ECF_CONST))
5921 /* If calls have a vdef, subsequent calls won't have
5922 the same incoming vuse. So, if 2 calls with vdef have the
5923 same vuse, we know they're not subsequent.
5924 We can value number 2 calls to the same function with the
5925 same vuse and the same operands which are not subsequent
5926 the same, because there is no code in the program that can
5927 compare the 2 values... */
5928 || (gimple_vdef (call_stmt)
5929 /* ... unless the call returns a pointer which does
5930 not alias with anything else. In which case the
5931 information that the values are distinct are encoded
5932 in the IL. */
5933 && !(gimple_call_return_flags (call_stmt) & ERF_NOALIAS)
5934 /* Only perform the following when being called from PRE
5935 which embeds tail merging. */
5936 && default_vn_walk_kind == VN_WALK))
5937 /* Do not process .DEFERRED_INIT since that confuses uninit
5938 analysis. */
5939 && !gimple_call_internal_p (call_stmt, IFN_DEFERRED_INIT))
5940 changed = visit_reference_op_call (lhs, call_stmt);
5941 else
5942 changed = defs_to_varying (call_stmt);
5944 else
5945 changed = defs_to_varying (stmt);
5946 done:
5947 return changed;
5951 /* Allocate a value number table. */
5953 static void
5954 allocate_vn_table (vn_tables_t table, unsigned size)
5956 table->phis = new vn_phi_table_type (size);
5957 table->nary = new vn_nary_op_table_type (size);
5958 table->references = new vn_reference_table_type (size);
5961 /* Free a value number table. */
5963 static void
5964 free_vn_table (vn_tables_t table)
5966 /* Walk over elements and release vectors. */
5967 vn_reference_iterator_type hir;
5968 vn_reference_t vr;
5969 FOR_EACH_HASH_TABLE_ELEMENT (*table->references, vr, vn_reference_t, hir)
5970 vr->operands.release ();
5971 delete table->phis;
5972 table->phis = NULL;
5973 delete table->nary;
5974 table->nary = NULL;
5975 delete table->references;
5976 table->references = NULL;
5979 /* Set *ID according to RESULT. */
5981 static void
5982 set_value_id_for_result (tree result, unsigned int *id)
5984 if (result && TREE_CODE (result) == SSA_NAME)
5985 *id = VN_INFO (result)->value_id;
5986 else if (result && is_gimple_min_invariant (result))
5987 *id = get_or_alloc_constant_value_id (result);
5988 else
5989 *id = get_next_value_id ();
5992 /* Set the value ids in the valid hash tables. */
5994 static void
5995 set_hashtable_value_ids (void)
5997 vn_nary_op_iterator_type hin;
5998 vn_phi_iterator_type hip;
5999 vn_reference_iterator_type hir;
6000 vn_nary_op_t vno;
6001 vn_reference_t vr;
6002 vn_phi_t vp;
6004 /* Now set the value ids of the things we had put in the hash
6005 table. */
6007 FOR_EACH_HASH_TABLE_ELEMENT (*valid_info->nary, vno, vn_nary_op_t, hin)
6008 if (! vno->predicated_values)
6009 set_value_id_for_result (vno->u.result, &vno->value_id);
6011 FOR_EACH_HASH_TABLE_ELEMENT (*valid_info->phis, vp, vn_phi_t, hip)
6012 set_value_id_for_result (vp->result, &vp->value_id);
6014 FOR_EACH_HASH_TABLE_ELEMENT (*valid_info->references, vr, vn_reference_t,
6015 hir)
6016 set_value_id_for_result (vr->result, &vr->value_id);
6019 /* Return the maximum value id we have ever seen. */
6021 unsigned int
6022 get_max_value_id (void)
6024 return next_value_id;
6027 /* Return the maximum constant value id we have ever seen. */
6029 unsigned int
6030 get_max_constant_value_id (void)
6032 return -next_constant_value_id;
6035 /* Return the next unique value id. */
6037 unsigned int
6038 get_next_value_id (void)
6040 gcc_checking_assert ((int)next_value_id > 0);
6041 return next_value_id++;
6044 /* Return the next unique value id for constants. */
6046 unsigned int
6047 get_next_constant_value_id (void)
6049 gcc_checking_assert (next_constant_value_id < 0);
6050 return next_constant_value_id--;
6054 /* Compare two expressions E1 and E2 and return true if they are equal.
6055 If match_vn_top_optimistically is true then VN_TOP is equal to anything,
6056 otherwise VN_TOP only matches VN_TOP. */
6058 bool
6059 expressions_equal_p (tree e1, tree e2, bool match_vn_top_optimistically)
6061 /* The obvious case. */
6062 if (e1 == e2)
6063 return true;
6065 /* If either one is VN_TOP consider them equal. */
6066 if (match_vn_top_optimistically
6067 && (e1 == VN_TOP || e2 == VN_TOP))
6068 return true;
6070 /* SSA_NAME compare pointer equal. */
6071 if (TREE_CODE (e1) == SSA_NAME || TREE_CODE (e2) == SSA_NAME)
6072 return false;
6074 /* Now perform the actual comparison. */
6075 if (TREE_CODE (e1) == TREE_CODE (e2)
6076 && operand_equal_p (e1, e2, OEP_PURE_SAME))
6077 return true;
6079 return false;
6083 /* Return true if the nary operation NARY may trap. This is a copy
6084 of stmt_could_throw_1_p adjusted to the SCCVN IL. */
6086 bool
6087 vn_nary_may_trap (vn_nary_op_t nary)
6089 tree type;
6090 tree rhs2 = NULL_TREE;
6091 bool honor_nans = false;
6092 bool honor_snans = false;
6093 bool fp_operation = false;
6094 bool honor_trapv = false;
6095 bool handled, ret;
6096 unsigned i;
6098 if (TREE_CODE_CLASS (nary->opcode) == tcc_comparison
6099 || TREE_CODE_CLASS (nary->opcode) == tcc_unary
6100 || TREE_CODE_CLASS (nary->opcode) == tcc_binary)
6102 type = nary->type;
6103 fp_operation = FLOAT_TYPE_P (type);
6104 if (fp_operation)
6106 honor_nans = flag_trapping_math && !flag_finite_math_only;
6107 honor_snans = flag_signaling_nans != 0;
6109 else if (INTEGRAL_TYPE_P (type) && TYPE_OVERFLOW_TRAPS (type))
6110 honor_trapv = true;
6112 if (nary->length >= 2)
6113 rhs2 = nary->op[1];
6114 ret = operation_could_trap_helper_p (nary->opcode, fp_operation,
6115 honor_trapv, honor_nans, honor_snans,
6116 rhs2, &handled);
6117 if (handled && ret)
6118 return true;
6120 for (i = 0; i < nary->length; ++i)
6121 if (tree_could_trap_p (nary->op[i]))
6122 return true;
6124 return false;
6127 /* Return true if the reference operation REF may trap. */
6129 bool
6130 vn_reference_may_trap (vn_reference_t ref)
6132 switch (ref->operands[0].opcode)
6134 case MODIFY_EXPR:
6135 case CALL_EXPR:
6136 /* We do not handle calls. */
6137 return true;
6138 case ADDR_EXPR:
6139 /* And toplevel address computations never trap. */
6140 return false;
6141 default:;
6144 vn_reference_op_t op;
6145 unsigned i;
6146 FOR_EACH_VEC_ELT (ref->operands, i, op)
6148 switch (op->opcode)
6150 case WITH_SIZE_EXPR:
6151 case TARGET_MEM_REF:
6152 /* Always variable. */
6153 return true;
6154 case COMPONENT_REF:
6155 if (op->op1 && TREE_CODE (op->op1) == SSA_NAME)
6156 return true;
6157 break;
6158 case ARRAY_RANGE_REF:
6159 if (TREE_CODE (op->op0) == SSA_NAME)
6160 return true;
6161 break;
6162 case ARRAY_REF:
6164 if (TREE_CODE (op->op0) != INTEGER_CST)
6165 return true;
6167 /* !in_array_bounds */
6168 tree domain_type = TYPE_DOMAIN (ref->operands[i+1].type);
6169 if (!domain_type)
6170 return true;
6172 tree min = op->op1;
6173 tree max = TYPE_MAX_VALUE (domain_type);
6174 if (!min
6175 || !max
6176 || TREE_CODE (min) != INTEGER_CST
6177 || TREE_CODE (max) != INTEGER_CST)
6178 return true;
6180 if (tree_int_cst_lt (op->op0, min)
6181 || tree_int_cst_lt (max, op->op0))
6182 return true;
6184 break;
6186 case MEM_REF:
6187 /* Nothing interesting in itself, the base is separate. */
6188 break;
6189 /* The following are the address bases. */
6190 case SSA_NAME:
6191 return true;
6192 case ADDR_EXPR:
6193 if (op->op0)
6194 return tree_could_trap_p (TREE_OPERAND (op->op0, 0));
6195 return false;
6196 default:;
6199 return false;
6202 eliminate_dom_walker::eliminate_dom_walker (cdi_direction direction,
6203 bitmap inserted_exprs_)
6204 : dom_walker (direction), do_pre (inserted_exprs_ != NULL),
6205 el_todo (0), eliminations (0), insertions (0),
6206 inserted_exprs (inserted_exprs_)
6208 need_eh_cleanup = BITMAP_ALLOC (NULL);
6209 need_ab_cleanup = BITMAP_ALLOC (NULL);
6212 eliminate_dom_walker::~eliminate_dom_walker ()
6214 BITMAP_FREE (need_eh_cleanup);
6215 BITMAP_FREE (need_ab_cleanup);
6218 /* Return a leader for OP that is available at the current point of the
6219 eliminate domwalk. */
6221 tree
6222 eliminate_dom_walker::eliminate_avail (basic_block, tree op)
6224 tree valnum = VN_INFO (op)->valnum;
6225 if (TREE_CODE (valnum) == SSA_NAME)
6227 if (SSA_NAME_IS_DEFAULT_DEF (valnum))
6228 return valnum;
6229 if (avail.length () > SSA_NAME_VERSION (valnum))
6230 return avail[SSA_NAME_VERSION (valnum)];
6232 else if (is_gimple_min_invariant (valnum))
6233 return valnum;
6234 return NULL_TREE;
6237 /* At the current point of the eliminate domwalk make OP available. */
6239 void
6240 eliminate_dom_walker::eliminate_push_avail (basic_block, tree op)
6242 tree valnum = VN_INFO (op)->valnum;
6243 if (TREE_CODE (valnum) == SSA_NAME)
6245 if (avail.length () <= SSA_NAME_VERSION (valnum))
6246 avail.safe_grow_cleared (SSA_NAME_VERSION (valnum) + 1, true);
6247 tree pushop = op;
6248 if (avail[SSA_NAME_VERSION (valnum)])
6249 pushop = avail[SSA_NAME_VERSION (valnum)];
6250 avail_stack.safe_push (pushop);
6251 avail[SSA_NAME_VERSION (valnum)] = op;
6255 /* Insert the expression recorded by SCCVN for VAL at *GSI. Returns
6256 the leader for the expression if insertion was successful. */
6258 tree
6259 eliminate_dom_walker::eliminate_insert (basic_block bb,
6260 gimple_stmt_iterator *gsi, tree val)
6262 /* We can insert a sequence with a single assignment only. */
6263 gimple_seq stmts = VN_INFO (val)->expr;
6264 if (!gimple_seq_singleton_p (stmts))
6265 return NULL_TREE;
6266 gassign *stmt = dyn_cast <gassign *> (gimple_seq_first_stmt (stmts));
6267 if (!stmt
6268 || (!CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt))
6269 && gimple_assign_rhs_code (stmt) != VIEW_CONVERT_EXPR
6270 && gimple_assign_rhs_code (stmt) != NEGATE_EXPR
6271 && gimple_assign_rhs_code (stmt) != BIT_FIELD_REF
6272 && (gimple_assign_rhs_code (stmt) != BIT_AND_EXPR
6273 || TREE_CODE (gimple_assign_rhs2 (stmt)) != INTEGER_CST)))
6274 return NULL_TREE;
6276 tree op = gimple_assign_rhs1 (stmt);
6277 if (gimple_assign_rhs_code (stmt) == VIEW_CONVERT_EXPR
6278 || gimple_assign_rhs_code (stmt) == BIT_FIELD_REF)
6279 op = TREE_OPERAND (op, 0);
6280 tree leader = TREE_CODE (op) == SSA_NAME ? eliminate_avail (bb, op) : op;
6281 if (!leader)
6282 return NULL_TREE;
6284 tree res;
6285 stmts = NULL;
6286 if (gimple_assign_rhs_code (stmt) == BIT_FIELD_REF)
6287 res = gimple_build (&stmts, BIT_FIELD_REF,
6288 TREE_TYPE (val), leader,
6289 TREE_OPERAND (gimple_assign_rhs1 (stmt), 1),
6290 TREE_OPERAND (gimple_assign_rhs1 (stmt), 2));
6291 else if (gimple_assign_rhs_code (stmt) == BIT_AND_EXPR)
6292 res = gimple_build (&stmts, BIT_AND_EXPR,
6293 TREE_TYPE (val), leader, gimple_assign_rhs2 (stmt));
6294 else
6295 res = gimple_build (&stmts, gimple_assign_rhs_code (stmt),
6296 TREE_TYPE (val), leader);
6297 if (TREE_CODE (res) != SSA_NAME
6298 || SSA_NAME_IS_DEFAULT_DEF (res)
6299 || gimple_bb (SSA_NAME_DEF_STMT (res)))
6301 gimple_seq_discard (stmts);
6303 /* During propagation we have to treat SSA info conservatively
6304 and thus we can end up simplifying the inserted expression
6305 at elimination time to sth not defined in stmts. */
6306 /* But then this is a redundancy we failed to detect. Which means
6307 res now has two values. That doesn't play well with how
6308 we track availability here, so give up. */
6309 if (dump_file && (dump_flags & TDF_DETAILS))
6311 if (TREE_CODE (res) == SSA_NAME)
6312 res = eliminate_avail (bb, res);
6313 if (res)
6315 fprintf (dump_file, "Failed to insert expression for value ");
6316 print_generic_expr (dump_file, val);
6317 fprintf (dump_file, " which is really fully redundant to ");
6318 print_generic_expr (dump_file, res);
6319 fprintf (dump_file, "\n");
6323 return NULL_TREE;
6325 else
6327 gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);
6328 vn_ssa_aux_t vn_info = VN_INFO (res);
6329 vn_info->valnum = val;
6330 vn_info->visited = true;
6333 insertions++;
6334 if (dump_file && (dump_flags & TDF_DETAILS))
6336 fprintf (dump_file, "Inserted ");
6337 print_gimple_stmt (dump_file, SSA_NAME_DEF_STMT (res), 0);
6340 return res;
6343 void
6344 eliminate_dom_walker::eliminate_stmt (basic_block b, gimple_stmt_iterator *gsi)
6346 tree sprime = NULL_TREE;
6347 gimple *stmt = gsi_stmt (*gsi);
6348 tree lhs = gimple_get_lhs (stmt);
6349 if (lhs && TREE_CODE (lhs) == SSA_NAME
6350 && !gimple_has_volatile_ops (stmt)
6351 /* See PR43491. Do not replace a global register variable when
6352 it is a the RHS of an assignment. Do replace local register
6353 variables since gcc does not guarantee a local variable will
6354 be allocated in register.
6355 ??? The fix isn't effective here. This should instead
6356 be ensured by not value-numbering them the same but treating
6357 them like volatiles? */
6358 && !(gimple_assign_single_p (stmt)
6359 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == VAR_DECL
6360 && DECL_HARD_REGISTER (gimple_assign_rhs1 (stmt))
6361 && is_global_var (gimple_assign_rhs1 (stmt)))))
6363 sprime = eliminate_avail (b, lhs);
6364 if (!sprime)
6366 /* If there is no existing usable leader but SCCVN thinks
6367 it has an expression it wants to use as replacement,
6368 insert that. */
6369 tree val = VN_INFO (lhs)->valnum;
6370 vn_ssa_aux_t vn_info;
6371 if (val != VN_TOP
6372 && TREE_CODE (val) == SSA_NAME
6373 && (vn_info = VN_INFO (val), true)
6374 && vn_info->needs_insertion
6375 && vn_info->expr != NULL
6376 && (sprime = eliminate_insert (b, gsi, val)) != NULL_TREE)
6377 eliminate_push_avail (b, sprime);
6380 /* If this now constitutes a copy duplicate points-to
6381 and range info appropriately. This is especially
6382 important for inserted code. See tree-ssa-copy.cc
6383 for similar code. */
6384 if (sprime
6385 && TREE_CODE (sprime) == SSA_NAME)
6387 basic_block sprime_b = gimple_bb (SSA_NAME_DEF_STMT (sprime));
6388 if (POINTER_TYPE_P (TREE_TYPE (lhs))
6389 && SSA_NAME_PTR_INFO (lhs)
6390 && ! SSA_NAME_PTR_INFO (sprime))
6392 duplicate_ssa_name_ptr_info (sprime,
6393 SSA_NAME_PTR_INFO (lhs));
6394 if (b != sprime_b)
6395 reset_flow_sensitive_info (sprime);
6397 else if (INTEGRAL_TYPE_P (TREE_TYPE (lhs))
6398 && SSA_NAME_RANGE_INFO (lhs)
6399 && ! SSA_NAME_RANGE_INFO (sprime)
6400 && b == sprime_b)
6401 duplicate_ssa_name_range_info (sprime, lhs);
6404 /* Inhibit the use of an inserted PHI on a loop header when
6405 the address of the memory reference is a simple induction
6406 variable. In other cases the vectorizer won't do anything
6407 anyway (either it's loop invariant or a complicated
6408 expression). */
6409 if (sprime
6410 && TREE_CODE (sprime) == SSA_NAME
6411 && do_pre
6412 && (flag_tree_loop_vectorize || flag_tree_parallelize_loops > 1)
6413 && loop_outer (b->loop_father)
6414 && has_zero_uses (sprime)
6415 && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (sprime))
6416 && gimple_assign_load_p (stmt))
6418 gimple *def_stmt = SSA_NAME_DEF_STMT (sprime);
6419 basic_block def_bb = gimple_bb (def_stmt);
6420 if (gimple_code (def_stmt) == GIMPLE_PHI
6421 && def_bb->loop_father->header == def_bb)
6423 loop_p loop = def_bb->loop_father;
6424 ssa_op_iter iter;
6425 tree op;
6426 bool found = false;
6427 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
6429 affine_iv iv;
6430 def_bb = gimple_bb (SSA_NAME_DEF_STMT (op));
6431 if (def_bb
6432 && flow_bb_inside_loop_p (loop, def_bb)
6433 && simple_iv (loop, loop, op, &iv, true))
6435 found = true;
6436 break;
6439 if (found)
6441 if (dump_file && (dump_flags & TDF_DETAILS))
6443 fprintf (dump_file, "Not replacing ");
6444 print_gimple_expr (dump_file, stmt, 0);
6445 fprintf (dump_file, " with ");
6446 print_generic_expr (dump_file, sprime);
6447 fprintf (dump_file, " which would add a loop"
6448 " carried dependence to loop %d\n",
6449 loop->num);
6451 /* Don't keep sprime available. */
6452 sprime = NULL_TREE;
6457 if (sprime)
6459 /* If we can propagate the value computed for LHS into
6460 all uses don't bother doing anything with this stmt. */
6461 if (may_propagate_copy (lhs, sprime))
6463 /* Mark it for removal. */
6464 to_remove.safe_push (stmt);
6466 /* ??? Don't count copy/constant propagations. */
6467 if (gimple_assign_single_p (stmt)
6468 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
6469 || gimple_assign_rhs1 (stmt) == sprime))
6470 return;
6472 if (dump_file && (dump_flags & TDF_DETAILS))
6474 fprintf (dump_file, "Replaced ");
6475 print_gimple_expr (dump_file, stmt, 0);
6476 fprintf (dump_file, " with ");
6477 print_generic_expr (dump_file, sprime);
6478 fprintf (dump_file, " in all uses of ");
6479 print_gimple_stmt (dump_file, stmt, 0);
6482 eliminations++;
6483 return;
6486 /* If this is an assignment from our leader (which
6487 happens in the case the value-number is a constant)
6488 then there is nothing to do. Likewise if we run into
6489 inserted code that needed a conversion because of
6490 our type-agnostic value-numbering of loads. */
6491 if ((gimple_assign_single_p (stmt)
6492 || (is_gimple_assign (stmt)
6493 && (CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt))
6494 || gimple_assign_rhs_code (stmt) == VIEW_CONVERT_EXPR)))
6495 && sprime == gimple_assign_rhs1 (stmt))
6496 return;
6498 /* Else replace its RHS. */
6499 if (dump_file && (dump_flags & TDF_DETAILS))
6501 fprintf (dump_file, "Replaced ");
6502 print_gimple_expr (dump_file, stmt, 0);
6503 fprintf (dump_file, " with ");
6504 print_generic_expr (dump_file, sprime);
6505 fprintf (dump_file, " in ");
6506 print_gimple_stmt (dump_file, stmt, 0);
6508 eliminations++;
6510 bool can_make_abnormal_goto = (is_gimple_call (stmt)
6511 && stmt_can_make_abnormal_goto (stmt));
6512 gimple *orig_stmt = stmt;
6513 if (!useless_type_conversion_p (TREE_TYPE (lhs),
6514 TREE_TYPE (sprime)))
6516 /* We preserve conversions to but not from function or method
6517 types. This asymmetry makes it necessary to re-instantiate
6518 conversions here. */
6519 if (POINTER_TYPE_P (TREE_TYPE (lhs))
6520 && FUNC_OR_METHOD_TYPE_P (TREE_TYPE (TREE_TYPE (lhs))))
6521 sprime = fold_convert (TREE_TYPE (lhs), sprime);
6522 else
6523 gcc_unreachable ();
6525 tree vdef = gimple_vdef (stmt);
6526 tree vuse = gimple_vuse (stmt);
6527 propagate_tree_value_into_stmt (gsi, sprime);
6528 stmt = gsi_stmt (*gsi);
6529 update_stmt (stmt);
6530 /* In case the VDEF on the original stmt was released, value-number
6531 it to the VUSE. This is to make vuse_ssa_val able to skip
6532 released virtual operands. */
6533 if (vdef != gimple_vdef (stmt))
6535 gcc_assert (SSA_NAME_IN_FREE_LIST (vdef));
6536 VN_INFO (vdef)->valnum = vuse;
6539 /* If we removed EH side-effects from the statement, clean
6540 its EH information. */
6541 if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
6543 bitmap_set_bit (need_eh_cleanup,
6544 gimple_bb (stmt)->index);
6545 if (dump_file && (dump_flags & TDF_DETAILS))
6546 fprintf (dump_file, " Removed EH side-effects.\n");
6549 /* Likewise for AB side-effects. */
6550 if (can_make_abnormal_goto
6551 && !stmt_can_make_abnormal_goto (stmt))
6553 bitmap_set_bit (need_ab_cleanup,
6554 gimple_bb (stmt)->index);
6555 if (dump_file && (dump_flags & TDF_DETAILS))
6556 fprintf (dump_file, " Removed AB side-effects.\n");
6559 return;
6563 /* If the statement is a scalar store, see if the expression
6564 has the same value number as its rhs. If so, the store is
6565 dead. */
6566 if (gimple_assign_single_p (stmt)
6567 && !gimple_has_volatile_ops (stmt)
6568 && !is_gimple_reg (gimple_assign_lhs (stmt))
6569 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
6570 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
6572 tree rhs = gimple_assign_rhs1 (stmt);
6573 vn_reference_t vnresult;
6574 /* ??? gcc.dg/torture/pr91445.c shows that we lookup a boolean
6575 typed load of a byte known to be 0x11 as 1 so a store of
6576 a boolean 1 is detected as redundant. Because of this we
6577 have to make sure to lookup with a ref where its size
6578 matches the precision. */
6579 tree lookup_lhs = lhs;
6580 if (INTEGRAL_TYPE_P (TREE_TYPE (lhs))
6581 && (TREE_CODE (lhs) != COMPONENT_REF
6582 || !DECL_BIT_FIELD_TYPE (TREE_OPERAND (lhs, 1)))
6583 && !type_has_mode_precision_p (TREE_TYPE (lhs)))
6585 if (TREE_CODE (lhs) == COMPONENT_REF
6586 || TREE_CODE (lhs) == MEM_REF)
6588 tree ltype = build_nonstandard_integer_type
6589 (TREE_INT_CST_LOW (TYPE_SIZE (TREE_TYPE (lhs))),
6590 TYPE_UNSIGNED (TREE_TYPE (lhs)));
6591 if (TREE_CODE (lhs) == COMPONENT_REF)
6593 tree foff = component_ref_field_offset (lhs);
6594 tree f = TREE_OPERAND (lhs, 1);
6595 if (!poly_int_tree_p (foff))
6596 lookup_lhs = NULL_TREE;
6597 else
6598 lookup_lhs = build3 (BIT_FIELD_REF, ltype,
6599 TREE_OPERAND (lhs, 0),
6600 TYPE_SIZE (TREE_TYPE (lhs)),
6601 bit_from_pos
6602 (foff, DECL_FIELD_BIT_OFFSET (f)));
6604 else
6605 lookup_lhs = build2 (MEM_REF, ltype,
6606 TREE_OPERAND (lhs, 0),
6607 TREE_OPERAND (lhs, 1));
6609 else
6610 lookup_lhs = NULL_TREE;
6612 tree val = NULL_TREE;
6613 if (lookup_lhs)
6614 val = vn_reference_lookup (lookup_lhs, gimple_vuse (stmt),
6615 VN_WALKREWRITE, &vnresult, false,
6616 NULL, NULL_TREE, true);
6617 if (TREE_CODE (rhs) == SSA_NAME)
6618 rhs = VN_INFO (rhs)->valnum;
6619 if (val
6620 && (operand_equal_p (val, rhs, 0)
6621 /* Due to the bitfield lookups above we can get bit
6622 interpretations of the same RHS as values here. Those
6623 are redundant as well. */
6624 || (TREE_CODE (val) == SSA_NAME
6625 && gimple_assign_single_p (SSA_NAME_DEF_STMT (val))
6626 && (val = gimple_assign_rhs1 (SSA_NAME_DEF_STMT (val)))
6627 && TREE_CODE (val) == VIEW_CONVERT_EXPR
6628 && TREE_OPERAND (val, 0) == rhs)))
6630 /* We can only remove the later store if the former aliases
6631 at least all accesses the later one does or if the store
6632 was to readonly memory storing the same value. */
6633 ao_ref lhs_ref;
6634 ao_ref_init (&lhs_ref, lhs);
6635 alias_set_type set = ao_ref_alias_set (&lhs_ref);
6636 alias_set_type base_set = ao_ref_base_alias_set (&lhs_ref);
6637 if (! vnresult
6638 || ((vnresult->set == set
6639 || alias_set_subset_of (set, vnresult->set))
6640 && (vnresult->base_set == base_set
6641 || alias_set_subset_of (base_set, vnresult->base_set))))
6643 if (dump_file && (dump_flags & TDF_DETAILS))
6645 fprintf (dump_file, "Deleted redundant store ");
6646 print_gimple_stmt (dump_file, stmt, 0);
6649 /* Queue stmt for removal. */
6650 to_remove.safe_push (stmt);
6651 return;
6656 /* If this is a control statement value numbering left edges
6657 unexecuted on force the condition in a way consistent with
6658 that. */
6659 if (gcond *cond = dyn_cast <gcond *> (stmt))
6661 if ((EDGE_SUCC (b, 0)->flags & EDGE_EXECUTABLE)
6662 ^ (EDGE_SUCC (b, 1)->flags & EDGE_EXECUTABLE))
6664 if (dump_file && (dump_flags & TDF_DETAILS))
6666 fprintf (dump_file, "Removing unexecutable edge from ");
6667 print_gimple_stmt (dump_file, stmt, 0);
6669 if (((EDGE_SUCC (b, 0)->flags & EDGE_TRUE_VALUE) != 0)
6670 == ((EDGE_SUCC (b, 0)->flags & EDGE_EXECUTABLE) != 0))
6671 gimple_cond_make_true (cond);
6672 else
6673 gimple_cond_make_false (cond);
6674 update_stmt (cond);
6675 el_todo |= TODO_cleanup_cfg;
6676 return;
6680 bool can_make_abnormal_goto = stmt_can_make_abnormal_goto (stmt);
6681 bool was_noreturn = (is_gimple_call (stmt)
6682 && gimple_call_noreturn_p (stmt));
6683 tree vdef = gimple_vdef (stmt);
6684 tree vuse = gimple_vuse (stmt);
6686 /* If we didn't replace the whole stmt (or propagate the result
6687 into all uses), replace all uses on this stmt with their
6688 leaders. */
6689 bool modified = false;
6690 use_operand_p use_p;
6691 ssa_op_iter iter;
6692 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE)
6694 tree use = USE_FROM_PTR (use_p);
6695 /* ??? The call code above leaves stmt operands un-updated. */
6696 if (TREE_CODE (use) != SSA_NAME)
6697 continue;
6698 tree sprime;
6699 if (SSA_NAME_IS_DEFAULT_DEF (use))
6700 /* ??? For default defs BB shouldn't matter, but we have to
6701 solve the inconsistency between rpo eliminate and
6702 dom eliminate avail valueization first. */
6703 sprime = eliminate_avail (b, use);
6704 else
6705 /* Look for sth available at the definition block of the argument.
6706 This avoids inconsistencies between availability there which
6707 decides if the stmt can be removed and availability at the
6708 use site. The SSA property ensures that things available
6709 at the definition are also available at uses. */
6710 sprime = eliminate_avail (gimple_bb (SSA_NAME_DEF_STMT (use)), use);
6711 if (sprime && sprime != use
6712 && may_propagate_copy (use, sprime, true)
6713 /* We substitute into debug stmts to avoid excessive
6714 debug temporaries created by removed stmts, but we need
6715 to avoid doing so for inserted sprimes as we never want
6716 to create debug temporaries for them. */
6717 && (!inserted_exprs
6718 || TREE_CODE (sprime) != SSA_NAME
6719 || !is_gimple_debug (stmt)
6720 || !bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (sprime))))
6722 propagate_value (use_p, sprime);
6723 modified = true;
6727 /* Fold the stmt if modified, this canonicalizes MEM_REFs we propagated
6728 into which is a requirement for the IPA devirt machinery. */
6729 gimple *old_stmt = stmt;
6730 if (modified)
6732 /* If a formerly non-invariant ADDR_EXPR is turned into an
6733 invariant one it was on a separate stmt. */
6734 if (gimple_assign_single_p (stmt)
6735 && TREE_CODE (gimple_assign_rhs1 (stmt)) == ADDR_EXPR)
6736 recompute_tree_invariant_for_addr_expr (gimple_assign_rhs1 (stmt));
6737 gimple_stmt_iterator prev = *gsi;
6738 gsi_prev (&prev);
6739 if (fold_stmt (gsi, follow_all_ssa_edges))
6741 /* fold_stmt may have created new stmts inbetween
6742 the previous stmt and the folded stmt. Mark
6743 all defs created there as varying to not confuse
6744 the SCCVN machinery as we're using that even during
6745 elimination. */
6746 if (gsi_end_p (prev))
6747 prev = gsi_start_bb (b);
6748 else
6749 gsi_next (&prev);
6750 if (gsi_stmt (prev) != gsi_stmt (*gsi))
6753 tree def;
6754 ssa_op_iter dit;
6755 FOR_EACH_SSA_TREE_OPERAND (def, gsi_stmt (prev),
6756 dit, SSA_OP_ALL_DEFS)
6757 /* As existing DEFs may move between stmts
6758 only process new ones. */
6759 if (! has_VN_INFO (def))
6761 vn_ssa_aux_t vn_info = VN_INFO (def);
6762 vn_info->valnum = def;
6763 vn_info->visited = true;
6765 if (gsi_stmt (prev) == gsi_stmt (*gsi))
6766 break;
6767 gsi_next (&prev);
6769 while (1);
6771 stmt = gsi_stmt (*gsi);
6772 /* In case we folded the stmt away schedule the NOP for removal. */
6773 if (gimple_nop_p (stmt))
6774 to_remove.safe_push (stmt);
6777 /* Visit indirect calls and turn them into direct calls if
6778 possible using the devirtualization machinery. Do this before
6779 checking for required EH/abnormal/noreturn cleanup as devird
6780 may expose more of those. */
6781 if (gcall *call_stmt = dyn_cast <gcall *> (stmt))
6783 tree fn = gimple_call_fn (call_stmt);
6784 if (fn
6785 && flag_devirtualize
6786 && virtual_method_call_p (fn))
6788 tree otr_type = obj_type_ref_class (fn);
6789 unsigned HOST_WIDE_INT otr_tok
6790 = tree_to_uhwi (OBJ_TYPE_REF_TOKEN (fn));
6791 tree instance;
6792 ipa_polymorphic_call_context context (current_function_decl,
6793 fn, stmt, &instance);
6794 context.get_dynamic_type (instance, OBJ_TYPE_REF_OBJECT (fn),
6795 otr_type, stmt, NULL);
6796 bool final;
6797 vec <cgraph_node *> targets
6798 = possible_polymorphic_call_targets (obj_type_ref_class (fn),
6799 otr_tok, context, &final);
6800 if (dump_file)
6801 dump_possible_polymorphic_call_targets (dump_file,
6802 obj_type_ref_class (fn),
6803 otr_tok, context);
6804 if (final && targets.length () <= 1 && dbg_cnt (devirt))
6806 tree fn;
6807 if (targets.length () == 1)
6808 fn = targets[0]->decl;
6809 else
6810 fn = builtin_decl_implicit (BUILT_IN_UNREACHABLE);
6811 if (dump_enabled_p ())
6813 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, stmt,
6814 "converting indirect call to "
6815 "function %s\n",
6816 lang_hooks.decl_printable_name (fn, 2));
6818 gimple_call_set_fndecl (call_stmt, fn);
6819 /* If changing the call to __builtin_unreachable
6820 or similar noreturn function, adjust gimple_call_fntype
6821 too. */
6822 if (gimple_call_noreturn_p (call_stmt)
6823 && VOID_TYPE_P (TREE_TYPE (TREE_TYPE (fn)))
6824 && TYPE_ARG_TYPES (TREE_TYPE (fn))
6825 && (TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fn)))
6826 == void_type_node))
6827 gimple_call_set_fntype (call_stmt, TREE_TYPE (fn));
6828 maybe_remove_unused_call_args (cfun, call_stmt);
6829 modified = true;
6834 if (modified)
6836 /* When changing a call into a noreturn call, cfg cleanup
6837 is needed to fix up the noreturn call. */
6838 if (!was_noreturn
6839 && is_gimple_call (stmt) && gimple_call_noreturn_p (stmt))
6840 to_fixup.safe_push (stmt);
6841 /* When changing a condition or switch into one we know what
6842 edge will be executed, schedule a cfg cleanup. */
6843 if ((gimple_code (stmt) == GIMPLE_COND
6844 && (gimple_cond_true_p (as_a <gcond *> (stmt))
6845 || gimple_cond_false_p (as_a <gcond *> (stmt))))
6846 || (gimple_code (stmt) == GIMPLE_SWITCH
6847 && TREE_CODE (gimple_switch_index
6848 (as_a <gswitch *> (stmt))) == INTEGER_CST))
6849 el_todo |= TODO_cleanup_cfg;
6850 /* If we removed EH side-effects from the statement, clean
6851 its EH information. */
6852 if (maybe_clean_or_replace_eh_stmt (old_stmt, stmt))
6854 bitmap_set_bit (need_eh_cleanup,
6855 gimple_bb (stmt)->index);
6856 if (dump_file && (dump_flags & TDF_DETAILS))
6857 fprintf (dump_file, " Removed EH side-effects.\n");
6859 /* Likewise for AB side-effects. */
6860 if (can_make_abnormal_goto
6861 && !stmt_can_make_abnormal_goto (stmt))
6863 bitmap_set_bit (need_ab_cleanup,
6864 gimple_bb (stmt)->index);
6865 if (dump_file && (dump_flags & TDF_DETAILS))
6866 fprintf (dump_file, " Removed AB side-effects.\n");
6868 update_stmt (stmt);
6869 /* In case the VDEF on the original stmt was released, value-number
6870 it to the VUSE. This is to make vuse_ssa_val able to skip
6871 released virtual operands. */
6872 if (vdef && SSA_NAME_IN_FREE_LIST (vdef))
6873 VN_INFO (vdef)->valnum = vuse;
6876 /* Make new values available - for fully redundant LHS we
6877 continue with the next stmt above and skip this. */
6878 def_operand_p defp;
6879 FOR_EACH_SSA_DEF_OPERAND (defp, stmt, iter, SSA_OP_DEF)
6880 eliminate_push_avail (b, DEF_FROM_PTR (defp));
6883 /* Perform elimination for the basic-block B during the domwalk. */
6885 edge
6886 eliminate_dom_walker::before_dom_children (basic_block b)
6888 /* Mark new bb. */
6889 avail_stack.safe_push (NULL_TREE);
6891 /* Skip unreachable blocks marked unreachable during the SCCVN domwalk. */
6892 if (!(b->flags & BB_EXECUTABLE))
6893 return NULL;
6895 vn_context_bb = b;
6897 for (gphi_iterator gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
6899 gphi *phi = gsi.phi ();
6900 tree res = PHI_RESULT (phi);
6902 if (virtual_operand_p (res))
6904 gsi_next (&gsi);
6905 continue;
6908 tree sprime = eliminate_avail (b, res);
6909 if (sprime
6910 && sprime != res)
6912 if (dump_file && (dump_flags & TDF_DETAILS))
6914 fprintf (dump_file, "Replaced redundant PHI node defining ");
6915 print_generic_expr (dump_file, res);
6916 fprintf (dump_file, " with ");
6917 print_generic_expr (dump_file, sprime);
6918 fprintf (dump_file, "\n");
6921 /* If we inserted this PHI node ourself, it's not an elimination. */
6922 if (! inserted_exprs
6923 || ! bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
6924 eliminations++;
6926 /* If we will propagate into all uses don't bother to do
6927 anything. */
6928 if (may_propagate_copy (res, sprime))
6930 /* Mark the PHI for removal. */
6931 to_remove.safe_push (phi);
6932 gsi_next (&gsi);
6933 continue;
6936 remove_phi_node (&gsi, false);
6938 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
6939 sprime = fold_convert (TREE_TYPE (res), sprime);
6940 gimple *stmt = gimple_build_assign (res, sprime);
6941 gimple_stmt_iterator gsi2 = gsi_after_labels (b);
6942 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
6943 continue;
6946 eliminate_push_avail (b, res);
6947 gsi_next (&gsi);
6950 for (gimple_stmt_iterator gsi = gsi_start_bb (b);
6951 !gsi_end_p (gsi);
6952 gsi_next (&gsi))
6953 eliminate_stmt (b, &gsi);
6955 /* Replace destination PHI arguments. */
6956 edge_iterator ei;
6957 edge e;
6958 FOR_EACH_EDGE (e, ei, b->succs)
6959 if (e->flags & EDGE_EXECUTABLE)
6960 for (gphi_iterator gsi = gsi_start_phis (e->dest);
6961 !gsi_end_p (gsi);
6962 gsi_next (&gsi))
6964 gphi *phi = gsi.phi ();
6965 use_operand_p use_p = PHI_ARG_DEF_PTR_FROM_EDGE (phi, e);
6966 tree arg = USE_FROM_PTR (use_p);
6967 if (TREE_CODE (arg) != SSA_NAME
6968 || virtual_operand_p (arg))
6969 continue;
6970 tree sprime = eliminate_avail (b, arg);
6971 if (sprime && may_propagate_copy (arg, sprime))
6972 propagate_value (use_p, sprime);
6975 vn_context_bb = NULL;
6977 return NULL;
6980 /* Make no longer available leaders no longer available. */
6982 void
6983 eliminate_dom_walker::after_dom_children (basic_block)
6985 tree entry;
6986 while ((entry = avail_stack.pop ()) != NULL_TREE)
6988 tree valnum = VN_INFO (entry)->valnum;
6989 tree old = avail[SSA_NAME_VERSION (valnum)];
6990 if (old == entry)
6991 avail[SSA_NAME_VERSION (valnum)] = NULL_TREE;
6992 else
6993 avail[SSA_NAME_VERSION (valnum)] = entry;
6997 /* Remove queued stmts and perform delayed cleanups. */
6999 unsigned
7000 eliminate_dom_walker::eliminate_cleanup (bool region_p)
7002 statistics_counter_event (cfun, "Eliminated", eliminations);
7003 statistics_counter_event (cfun, "Insertions", insertions);
7005 /* We cannot remove stmts during BB walk, especially not release SSA
7006 names there as this confuses the VN machinery. The stmts ending
7007 up in to_remove are either stores or simple copies.
7008 Remove stmts in reverse order to make debug stmt creation possible. */
7009 while (!to_remove.is_empty ())
7011 bool do_release_defs = true;
7012 gimple *stmt = to_remove.pop ();
7014 /* When we are value-numbering a region we do not require exit PHIs to
7015 be present so we have to make sure to deal with uses outside of the
7016 region of stmts that we thought are eliminated.
7017 ??? Note we may be confused by uses in dead regions we didn't run
7018 elimination on. Rather than checking individual uses we accept
7019 dead copies to be generated here (gcc.c-torture/execute/20060905-1.c
7020 contains such example). */
7021 if (region_p)
7023 if (gphi *phi = dyn_cast <gphi *> (stmt))
7025 tree lhs = gimple_phi_result (phi);
7026 if (!has_zero_uses (lhs))
7028 if (dump_file && (dump_flags & TDF_DETAILS))
7029 fprintf (dump_file, "Keeping eliminated stmt live "
7030 "as copy because of out-of-region uses\n");
7031 tree sprime = eliminate_avail (gimple_bb (stmt), lhs);
7032 gimple *copy = gimple_build_assign (lhs, sprime);
7033 gimple_stmt_iterator gsi
7034 = gsi_after_labels (gimple_bb (stmt));
7035 gsi_insert_before (&gsi, copy, GSI_SAME_STMT);
7036 do_release_defs = false;
7039 else if (tree lhs = gimple_get_lhs (stmt))
7040 if (TREE_CODE (lhs) == SSA_NAME
7041 && !has_zero_uses (lhs))
7043 if (dump_file && (dump_flags & TDF_DETAILS))
7044 fprintf (dump_file, "Keeping eliminated stmt live "
7045 "as copy because of out-of-region uses\n");
7046 tree sprime = eliminate_avail (gimple_bb (stmt), lhs);
7047 gimple_stmt_iterator gsi = gsi_for_stmt (stmt);
7048 if (is_gimple_assign (stmt))
7050 gimple_assign_set_rhs_from_tree (&gsi, sprime);
7051 stmt = gsi_stmt (gsi);
7052 update_stmt (stmt);
7053 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
7054 bitmap_set_bit (need_eh_cleanup, gimple_bb (stmt)->index);
7055 continue;
7057 else
7059 gimple *copy = gimple_build_assign (lhs, sprime);
7060 gsi_insert_before (&gsi, copy, GSI_SAME_STMT);
7061 do_release_defs = false;
7066 if (dump_file && (dump_flags & TDF_DETAILS))
7068 fprintf (dump_file, "Removing dead stmt ");
7069 print_gimple_stmt (dump_file, stmt, 0, TDF_NONE);
7072 gimple_stmt_iterator gsi = gsi_for_stmt (stmt);
7073 if (gimple_code (stmt) == GIMPLE_PHI)
7074 remove_phi_node (&gsi, do_release_defs);
7075 else
7077 basic_block bb = gimple_bb (stmt);
7078 unlink_stmt_vdef (stmt);
7079 if (gsi_remove (&gsi, true))
7080 bitmap_set_bit (need_eh_cleanup, bb->index);
7081 if (is_gimple_call (stmt) && stmt_can_make_abnormal_goto (stmt))
7082 bitmap_set_bit (need_ab_cleanup, bb->index);
7083 if (do_release_defs)
7084 release_defs (stmt);
7087 /* Removing a stmt may expose a forwarder block. */
7088 el_todo |= TODO_cleanup_cfg;
7091 /* Fixup stmts that became noreturn calls. This may require splitting
7092 blocks and thus isn't possible during the dominator walk. Do this
7093 in reverse order so we don't inadvertedly remove a stmt we want to
7094 fixup by visiting a dominating now noreturn call first. */
7095 while (!to_fixup.is_empty ())
7097 gimple *stmt = to_fixup.pop ();
7099 if (dump_file && (dump_flags & TDF_DETAILS))
7101 fprintf (dump_file, "Fixing up noreturn call ");
7102 print_gimple_stmt (dump_file, stmt, 0);
7105 if (fixup_noreturn_call (stmt))
7106 el_todo |= TODO_cleanup_cfg;
7109 bool do_eh_cleanup = !bitmap_empty_p (need_eh_cleanup);
7110 bool do_ab_cleanup = !bitmap_empty_p (need_ab_cleanup);
7112 if (do_eh_cleanup)
7113 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
7115 if (do_ab_cleanup)
7116 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup);
7118 if (do_eh_cleanup || do_ab_cleanup)
7119 el_todo |= TODO_cleanup_cfg;
7121 return el_todo;
7124 /* Eliminate fully redundant computations. */
7126 unsigned
7127 eliminate_with_rpo_vn (bitmap inserted_exprs)
7129 eliminate_dom_walker walker (CDI_DOMINATORS, inserted_exprs);
7131 eliminate_dom_walker *saved_rpo_avail = rpo_avail;
7132 rpo_avail = &walker;
7133 walker.walk (cfun->cfg->x_entry_block_ptr);
7134 rpo_avail = saved_rpo_avail;
7136 return walker.eliminate_cleanup ();
7139 unsigned
7140 do_rpo_vn (function *fn, edge entry, bitmap exit_bbs,
7141 bool iterate, bool eliminate, vn_lookup_kind kind);
7143 void
7144 run_rpo_vn (vn_lookup_kind kind)
7146 do_rpo_vn (cfun, NULL, NULL, true, false, kind);
7148 /* ??? Prune requirement of these. */
7149 constant_to_value_id = new hash_table<vn_constant_hasher> (23);
7151 /* Initialize the value ids and prune out remaining VN_TOPs
7152 from dead code. */
7153 tree name;
7154 unsigned i;
7155 FOR_EACH_SSA_NAME (i, name, cfun)
7157 vn_ssa_aux_t info = VN_INFO (name);
7158 if (!info->visited
7159 || info->valnum == VN_TOP)
7160 info->valnum = name;
7161 if (info->valnum == name)
7162 info->value_id = get_next_value_id ();
7163 else if (is_gimple_min_invariant (info->valnum))
7164 info->value_id = get_or_alloc_constant_value_id (info->valnum);
7167 /* Propagate. */
7168 FOR_EACH_SSA_NAME (i, name, cfun)
7170 vn_ssa_aux_t info = VN_INFO (name);
7171 if (TREE_CODE (info->valnum) == SSA_NAME
7172 && info->valnum != name
7173 && info->value_id != VN_INFO (info->valnum)->value_id)
7174 info->value_id = VN_INFO (info->valnum)->value_id;
7177 set_hashtable_value_ids ();
7179 if (dump_file && (dump_flags & TDF_DETAILS))
7181 fprintf (dump_file, "Value numbers:\n");
7182 FOR_EACH_SSA_NAME (i, name, cfun)
7184 if (VN_INFO (name)->visited
7185 && SSA_VAL (name) != name)
7187 print_generic_expr (dump_file, name);
7188 fprintf (dump_file, " = ");
7189 print_generic_expr (dump_file, SSA_VAL (name));
7190 fprintf (dump_file, " (%04d)\n", VN_INFO (name)->value_id);
7196 /* Free VN associated data structures. */
7198 void
7199 free_rpo_vn (void)
7201 free_vn_table (valid_info);
7202 XDELETE (valid_info);
7203 obstack_free (&vn_tables_obstack, NULL);
7204 obstack_free (&vn_tables_insert_obstack, NULL);
7206 vn_ssa_aux_iterator_type it;
7207 vn_ssa_aux_t info;
7208 FOR_EACH_HASH_TABLE_ELEMENT (*vn_ssa_aux_hash, info, vn_ssa_aux_t, it)
7209 if (info->needs_insertion)
7210 release_ssa_name (info->name);
7211 obstack_free (&vn_ssa_aux_obstack, NULL);
7212 delete vn_ssa_aux_hash;
7214 delete constant_to_value_id;
7215 constant_to_value_id = NULL;
7218 /* Hook for maybe_push_res_to_seq, lookup the expression in the VN tables. */
7220 static tree
7221 vn_lookup_simplify_result (gimple_match_op *res_op)
7223 if (!res_op->code.is_tree_code ())
7224 return NULL_TREE;
7225 tree *ops = res_op->ops;
7226 unsigned int length = res_op->num_ops;
7227 if (res_op->code == CONSTRUCTOR
7228 /* ??? We're arriving here with SCCVNs view, decomposed CONSTRUCTOR
7229 and GIMPLEs / match-and-simplifies, CONSTRUCTOR as GENERIC tree. */
7230 && TREE_CODE (res_op->ops[0]) == CONSTRUCTOR)
7232 length = CONSTRUCTOR_NELTS (res_op->ops[0]);
7233 ops = XALLOCAVEC (tree, length);
7234 for (unsigned i = 0; i < length; ++i)
7235 ops[i] = CONSTRUCTOR_ELT (res_op->ops[0], i)->value;
7237 vn_nary_op_t vnresult = NULL;
7238 tree res = vn_nary_op_lookup_pieces (length, (tree_code) res_op->code,
7239 res_op->type, ops, &vnresult);
7240 /* If this is used from expression simplification make sure to
7241 return an available expression. */
7242 if (res && TREE_CODE (res) == SSA_NAME && mprts_hook && rpo_avail)
7243 res = rpo_avail->eliminate_avail (vn_context_bb, res);
7244 return res;
7247 /* Return a leader for OPs value that is valid at BB. */
7249 tree
7250 rpo_elim::eliminate_avail (basic_block bb, tree op)
7252 bool visited;
7253 tree valnum = SSA_VAL (op, &visited);
7254 /* If we didn't visit OP then it must be defined outside of the
7255 region we process and also dominate it. So it is available. */
7256 if (!visited)
7257 return op;
7258 if (TREE_CODE (valnum) == SSA_NAME)
7260 if (SSA_NAME_IS_DEFAULT_DEF (valnum))
7261 return valnum;
7262 vn_avail *av = VN_INFO (valnum)->avail;
7263 if (!av)
7264 return NULL_TREE;
7265 if (av->location == bb->index)
7266 /* On tramp3d 90% of the cases are here. */
7267 return ssa_name (av->leader);
7270 basic_block abb = BASIC_BLOCK_FOR_FN (cfun, av->location);
7271 /* ??? During elimination we have to use availability at the
7272 definition site of a use we try to replace. This
7273 is required to not run into inconsistencies because
7274 of dominated_by_p_w_unex behavior and removing a definition
7275 while not replacing all uses.
7276 ??? We could try to consistently walk dominators
7277 ignoring non-executable regions. The nearest common
7278 dominator of bb and abb is where we can stop walking. We
7279 may also be able to "pre-compute" (bits of) the next immediate
7280 (non-)dominator during the RPO walk when marking edges as
7281 executable. */
7282 if (dominated_by_p_w_unex (bb, abb, true))
7284 tree leader = ssa_name (av->leader);
7285 /* Prevent eliminations that break loop-closed SSA. */
7286 if (loops_state_satisfies_p (LOOP_CLOSED_SSA)
7287 && ! SSA_NAME_IS_DEFAULT_DEF (leader)
7288 && ! flow_bb_inside_loop_p (gimple_bb (SSA_NAME_DEF_STMT
7289 (leader))->loop_father,
7290 bb))
7291 return NULL_TREE;
7292 if (dump_file && (dump_flags & TDF_DETAILS))
7294 print_generic_expr (dump_file, leader);
7295 fprintf (dump_file, " is available for ");
7296 print_generic_expr (dump_file, valnum);
7297 fprintf (dump_file, "\n");
7299 /* On tramp3d 99% of the _remaining_ cases succeed at
7300 the first enty. */
7301 return leader;
7303 /* ??? Can we somehow skip to the immediate dominator
7304 RPO index (bb_to_rpo)? Again, maybe not worth, on
7305 tramp3d the worst number of elements in the vector is 9. */
7306 av = av->next;
7308 while (av);
7310 else if (valnum != VN_TOP)
7311 /* valnum is is_gimple_min_invariant. */
7312 return valnum;
7313 return NULL_TREE;
7316 /* Make LEADER a leader for its value at BB. */
7318 void
7319 rpo_elim::eliminate_push_avail (basic_block bb, tree leader)
7321 tree valnum = VN_INFO (leader)->valnum;
7322 if (valnum == VN_TOP
7323 || is_gimple_min_invariant (valnum))
7324 return;
7325 if (dump_file && (dump_flags & TDF_DETAILS))
7327 fprintf (dump_file, "Making available beyond BB%d ", bb->index);
7328 print_generic_expr (dump_file, leader);
7329 fprintf (dump_file, " for value ");
7330 print_generic_expr (dump_file, valnum);
7331 fprintf (dump_file, "\n");
7333 vn_ssa_aux_t value = VN_INFO (valnum);
7334 vn_avail *av;
7335 if (m_avail_freelist)
7337 av = m_avail_freelist;
7338 m_avail_freelist = m_avail_freelist->next;
7340 else
7341 av = XOBNEW (&vn_ssa_aux_obstack, vn_avail);
7342 av->location = bb->index;
7343 av->leader = SSA_NAME_VERSION (leader);
7344 av->next = value->avail;
7345 av->next_undo = last_pushed_avail;
7346 last_pushed_avail = value;
7347 value->avail = av;
7350 /* Valueization hook for RPO VN plus required state. */
7352 tree
7353 rpo_vn_valueize (tree name)
7355 if (TREE_CODE (name) == SSA_NAME)
7357 vn_ssa_aux_t val = VN_INFO (name);
7358 if (val)
7360 tree tem = val->valnum;
7361 if (tem != VN_TOP && tem != name)
7363 if (TREE_CODE (tem) != SSA_NAME)
7364 return tem;
7365 /* For all values we only valueize to an available leader
7366 which means we can use SSA name info without restriction. */
7367 tem = rpo_avail->eliminate_avail (vn_context_bb, tem);
7368 if (tem)
7369 return tem;
7373 return name;
7376 /* Insert on PRED_E predicates derived from CODE OPS being true besides the
7377 inverted condition. */
7379 static void
7380 insert_related_predicates_on_edge (enum tree_code code, tree *ops, edge pred_e)
7382 switch (code)
7384 case LT_EXPR:
7385 /* a < b -> a {!,<}= b */
7386 vn_nary_op_insert_pieces_predicated (2, NE_EXPR, boolean_type_node,
7387 ops, boolean_true_node, 0, pred_e);
7388 vn_nary_op_insert_pieces_predicated (2, LE_EXPR, boolean_type_node,
7389 ops, boolean_true_node, 0, pred_e);
7390 /* a < b -> ! a {>,=} b */
7391 vn_nary_op_insert_pieces_predicated (2, GT_EXPR, boolean_type_node,
7392 ops, boolean_false_node, 0, pred_e);
7393 vn_nary_op_insert_pieces_predicated (2, EQ_EXPR, boolean_type_node,
7394 ops, boolean_false_node, 0, pred_e);
7395 break;
7396 case GT_EXPR:
7397 /* a > b -> a {!,>}= b */
7398 vn_nary_op_insert_pieces_predicated (2, NE_EXPR, boolean_type_node,
7399 ops, boolean_true_node, 0, pred_e);
7400 vn_nary_op_insert_pieces_predicated (2, GE_EXPR, boolean_type_node,
7401 ops, boolean_true_node, 0, pred_e);
7402 /* a > b -> ! a {<,=} b */
7403 vn_nary_op_insert_pieces_predicated (2, LT_EXPR, boolean_type_node,
7404 ops, boolean_false_node, 0, pred_e);
7405 vn_nary_op_insert_pieces_predicated (2, EQ_EXPR, boolean_type_node,
7406 ops, boolean_false_node, 0, pred_e);
7407 break;
7408 case EQ_EXPR:
7409 /* a == b -> ! a {<,>} b */
7410 vn_nary_op_insert_pieces_predicated (2, LT_EXPR, boolean_type_node,
7411 ops, boolean_false_node, 0, pred_e);
7412 vn_nary_op_insert_pieces_predicated (2, GT_EXPR, boolean_type_node,
7413 ops, boolean_false_node, 0, pred_e);
7414 break;
7415 case LE_EXPR:
7416 case GE_EXPR:
7417 case NE_EXPR:
7418 /* Nothing besides inverted condition. */
7419 break;
7420 default:;
7424 /* Main stmt worker for RPO VN, process BB. */
7426 static unsigned
7427 process_bb (rpo_elim &avail, basic_block bb,
7428 bool bb_visited, bool iterate_phis, bool iterate, bool eliminate,
7429 bool do_region, bitmap exit_bbs, bool skip_phis)
7431 unsigned todo = 0;
7432 edge_iterator ei;
7433 edge e;
7435 vn_context_bb = bb;
7437 /* If we are in loop-closed SSA preserve this state. This is
7438 relevant when called on regions from outside of FRE/PRE. */
7439 bool lc_phi_nodes = false;
7440 if (!skip_phis
7441 && loops_state_satisfies_p (LOOP_CLOSED_SSA))
7442 FOR_EACH_EDGE (e, ei, bb->preds)
7443 if (e->src->loop_father != e->dest->loop_father
7444 && flow_loop_nested_p (e->dest->loop_father,
7445 e->src->loop_father))
7447 lc_phi_nodes = true;
7448 break;
7451 /* When we visit a loop header substitute into loop info. */
7452 if (!iterate && eliminate && bb->loop_father->header == bb)
7454 /* Keep fields in sync with substitute_in_loop_info. */
7455 if (bb->loop_father->nb_iterations)
7456 bb->loop_father->nb_iterations
7457 = simplify_replace_tree (bb->loop_father->nb_iterations,
7458 NULL_TREE, NULL_TREE, &vn_valueize_for_srt);
7461 /* Value-number all defs in the basic-block. */
7462 if (!skip_phis)
7463 for (gphi_iterator gsi = gsi_start_phis (bb); !gsi_end_p (gsi);
7464 gsi_next (&gsi))
7466 gphi *phi = gsi.phi ();
7467 tree res = PHI_RESULT (phi);
7468 vn_ssa_aux_t res_info = VN_INFO (res);
7469 if (!bb_visited)
7471 gcc_assert (!res_info->visited);
7472 res_info->valnum = VN_TOP;
7473 res_info->visited = true;
7476 /* When not iterating force backedge values to varying. */
7477 visit_stmt (phi, !iterate_phis);
7478 if (virtual_operand_p (res))
7479 continue;
7481 /* Eliminate */
7482 /* The interesting case is gcc.dg/tree-ssa/pr22230.c for correctness
7483 how we handle backedges and availability.
7484 And gcc.dg/tree-ssa/ssa-sccvn-2.c for optimization. */
7485 tree val = res_info->valnum;
7486 if (res != val && !iterate && eliminate)
7488 if (tree leader = avail.eliminate_avail (bb, res))
7490 if (leader != res
7491 /* Preserve loop-closed SSA form. */
7492 && (! lc_phi_nodes
7493 || is_gimple_min_invariant (leader)))
7495 if (dump_file && (dump_flags & TDF_DETAILS))
7497 fprintf (dump_file, "Replaced redundant PHI node "
7498 "defining ");
7499 print_generic_expr (dump_file, res);
7500 fprintf (dump_file, " with ");
7501 print_generic_expr (dump_file, leader);
7502 fprintf (dump_file, "\n");
7504 avail.eliminations++;
7506 if (may_propagate_copy (res, leader))
7508 /* Schedule for removal. */
7509 avail.to_remove.safe_push (phi);
7510 continue;
7512 /* ??? Else generate a copy stmt. */
7516 /* Only make defs available that not already are. But make
7517 sure loop-closed SSA PHI node defs are picked up for
7518 downstream uses. */
7519 if (lc_phi_nodes
7520 || res == val
7521 || ! avail.eliminate_avail (bb, res))
7522 avail.eliminate_push_avail (bb, res);
7525 /* For empty BBs mark outgoing edges executable. For non-empty BBs
7526 we do this when processing the last stmt as we have to do this
7527 before elimination which otherwise forces GIMPLE_CONDs to
7528 if (1 != 0) style when seeing non-executable edges. */
7529 if (gsi_end_p (gsi_start_bb (bb)))
7531 FOR_EACH_EDGE (e, ei, bb->succs)
7533 if (!(e->flags & EDGE_EXECUTABLE))
7535 if (dump_file && (dump_flags & TDF_DETAILS))
7536 fprintf (dump_file,
7537 "marking outgoing edge %d -> %d executable\n",
7538 e->src->index, e->dest->index);
7539 e->flags |= EDGE_EXECUTABLE;
7540 e->dest->flags |= BB_EXECUTABLE;
7542 else if (!(e->dest->flags & BB_EXECUTABLE))
7544 if (dump_file && (dump_flags & TDF_DETAILS))
7545 fprintf (dump_file,
7546 "marking destination block %d reachable\n",
7547 e->dest->index);
7548 e->dest->flags |= BB_EXECUTABLE;
7552 for (gimple_stmt_iterator gsi = gsi_start_bb (bb);
7553 !gsi_end_p (gsi); gsi_next (&gsi))
7555 ssa_op_iter i;
7556 tree op;
7557 if (!bb_visited)
7559 FOR_EACH_SSA_TREE_OPERAND (op, gsi_stmt (gsi), i, SSA_OP_ALL_DEFS)
7561 vn_ssa_aux_t op_info = VN_INFO (op);
7562 gcc_assert (!op_info->visited);
7563 op_info->valnum = VN_TOP;
7564 op_info->visited = true;
7567 /* We somehow have to deal with uses that are not defined
7568 in the processed region. Forcing unvisited uses to
7569 varying here doesn't play well with def-use following during
7570 expression simplification, so we deal with this by checking
7571 the visited flag in SSA_VAL. */
7574 visit_stmt (gsi_stmt (gsi));
7576 gimple *last = gsi_stmt (gsi);
7577 e = NULL;
7578 switch (gimple_code (last))
7580 case GIMPLE_SWITCH:
7581 e = find_taken_edge (bb, vn_valueize (gimple_switch_index
7582 (as_a <gswitch *> (last))));
7583 break;
7584 case GIMPLE_COND:
7586 tree lhs = vn_valueize (gimple_cond_lhs (last));
7587 tree rhs = vn_valueize (gimple_cond_rhs (last));
7588 tree val = gimple_simplify (gimple_cond_code (last),
7589 boolean_type_node, lhs, rhs,
7590 NULL, vn_valueize);
7591 /* If the condition didn't simplfy see if we have recorded
7592 an expression from sofar taken edges. */
7593 if (! val || TREE_CODE (val) != INTEGER_CST)
7595 vn_nary_op_t vnresult;
7596 tree ops[2];
7597 ops[0] = lhs;
7598 ops[1] = rhs;
7599 val = vn_nary_op_lookup_pieces (2, gimple_cond_code (last),
7600 boolean_type_node, ops,
7601 &vnresult);
7602 /* Did we get a predicated value? */
7603 if (! val && vnresult && vnresult->predicated_values)
7605 val = vn_nary_op_get_predicated_value (vnresult, bb);
7606 if (val && dump_file && (dump_flags & TDF_DETAILS))
7608 fprintf (dump_file, "Got predicated value ");
7609 print_generic_expr (dump_file, val, TDF_NONE);
7610 fprintf (dump_file, " for ");
7611 print_gimple_stmt (dump_file, last, TDF_SLIM);
7615 if (val)
7616 e = find_taken_edge (bb, val);
7617 if (! e)
7619 /* If we didn't manage to compute the taken edge then
7620 push predicated expressions for the condition itself
7621 and related conditions to the hashtables. This allows
7622 simplification of redundant conditions which is
7623 important as early cleanup. */
7624 edge true_e, false_e;
7625 extract_true_false_edges_from_block (bb, &true_e, &false_e);
7626 enum tree_code code = gimple_cond_code (last);
7627 enum tree_code icode
7628 = invert_tree_comparison (code, HONOR_NANS (lhs));
7629 tree ops[2];
7630 ops[0] = lhs;
7631 ops[1] = rhs;
7632 if (do_region
7633 && bitmap_bit_p (exit_bbs, true_e->dest->index))
7634 true_e = NULL;
7635 if (do_region
7636 && bitmap_bit_p (exit_bbs, false_e->dest->index))
7637 false_e = NULL;
7638 if (true_e)
7639 vn_nary_op_insert_pieces_predicated
7640 (2, code, boolean_type_node, ops,
7641 boolean_true_node, 0, true_e);
7642 if (false_e)
7643 vn_nary_op_insert_pieces_predicated
7644 (2, code, boolean_type_node, ops,
7645 boolean_false_node, 0, false_e);
7646 if (icode != ERROR_MARK)
7648 if (true_e)
7649 vn_nary_op_insert_pieces_predicated
7650 (2, icode, boolean_type_node, ops,
7651 boolean_false_node, 0, true_e);
7652 if (false_e)
7653 vn_nary_op_insert_pieces_predicated
7654 (2, icode, boolean_type_node, ops,
7655 boolean_true_node, 0, false_e);
7657 /* Relax for non-integers, inverted condition handled
7658 above. */
7659 if (INTEGRAL_TYPE_P (TREE_TYPE (lhs)))
7661 if (true_e)
7662 insert_related_predicates_on_edge (code, ops, true_e);
7663 if (false_e)
7664 insert_related_predicates_on_edge (icode, ops, false_e);
7667 break;
7669 case GIMPLE_GOTO:
7670 e = find_taken_edge (bb, vn_valueize (gimple_goto_dest (last)));
7671 break;
7672 default:
7673 e = NULL;
7675 if (e)
7677 todo = TODO_cleanup_cfg;
7678 if (!(e->flags & EDGE_EXECUTABLE))
7680 if (dump_file && (dump_flags & TDF_DETAILS))
7681 fprintf (dump_file,
7682 "marking known outgoing %sedge %d -> %d executable\n",
7683 e->flags & EDGE_DFS_BACK ? "back-" : "",
7684 e->src->index, e->dest->index);
7685 e->flags |= EDGE_EXECUTABLE;
7686 e->dest->flags |= BB_EXECUTABLE;
7688 else if (!(e->dest->flags & BB_EXECUTABLE))
7690 if (dump_file && (dump_flags & TDF_DETAILS))
7691 fprintf (dump_file,
7692 "marking destination block %d reachable\n",
7693 e->dest->index);
7694 e->dest->flags |= BB_EXECUTABLE;
7697 else if (gsi_one_before_end_p (gsi))
7699 FOR_EACH_EDGE (e, ei, bb->succs)
7701 if (!(e->flags & EDGE_EXECUTABLE))
7703 if (dump_file && (dump_flags & TDF_DETAILS))
7704 fprintf (dump_file,
7705 "marking outgoing edge %d -> %d executable\n",
7706 e->src->index, e->dest->index);
7707 e->flags |= EDGE_EXECUTABLE;
7708 e->dest->flags |= BB_EXECUTABLE;
7710 else if (!(e->dest->flags & BB_EXECUTABLE))
7712 if (dump_file && (dump_flags & TDF_DETAILS))
7713 fprintf (dump_file,
7714 "marking destination block %d reachable\n",
7715 e->dest->index);
7716 e->dest->flags |= BB_EXECUTABLE;
7721 /* Eliminate. That also pushes to avail. */
7722 if (eliminate && ! iterate)
7723 avail.eliminate_stmt (bb, &gsi);
7724 else
7725 /* If not eliminating, make all not already available defs
7726 available. */
7727 FOR_EACH_SSA_TREE_OPERAND (op, gsi_stmt (gsi), i, SSA_OP_DEF)
7728 if (! avail.eliminate_avail (bb, op))
7729 avail.eliminate_push_avail (bb, op);
7732 /* Eliminate in destination PHI arguments. Always substitute in dest
7733 PHIs, even for non-executable edges. This handles region
7734 exits PHIs. */
7735 if (!iterate && eliminate)
7736 FOR_EACH_EDGE (e, ei, bb->succs)
7737 for (gphi_iterator gsi = gsi_start_phis (e->dest);
7738 !gsi_end_p (gsi); gsi_next (&gsi))
7740 gphi *phi = gsi.phi ();
7741 use_operand_p use_p = PHI_ARG_DEF_PTR_FROM_EDGE (phi, e);
7742 tree arg = USE_FROM_PTR (use_p);
7743 if (TREE_CODE (arg) != SSA_NAME
7744 || virtual_operand_p (arg))
7745 continue;
7746 tree sprime;
7747 if (SSA_NAME_IS_DEFAULT_DEF (arg))
7749 sprime = SSA_VAL (arg);
7750 gcc_assert (TREE_CODE (sprime) != SSA_NAME
7751 || SSA_NAME_IS_DEFAULT_DEF (sprime));
7753 else
7754 /* Look for sth available at the definition block of the argument.
7755 This avoids inconsistencies between availability there which
7756 decides if the stmt can be removed and availability at the
7757 use site. The SSA property ensures that things available
7758 at the definition are also available at uses. */
7759 sprime = avail.eliminate_avail (gimple_bb (SSA_NAME_DEF_STMT (arg)),
7760 arg);
7761 if (sprime
7762 && sprime != arg
7763 && may_propagate_copy (arg, sprime))
7764 propagate_value (use_p, sprime);
7767 vn_context_bb = NULL;
7768 return todo;
7771 /* Unwind state per basic-block. */
7773 struct unwind_state
7775 /* Times this block has been visited. */
7776 unsigned visited;
7777 /* Whether to handle this as iteration point or whether to treat
7778 incoming backedge PHI values as varying. */
7779 bool iterate;
7780 /* Maximum RPO index this block is reachable from. */
7781 int max_rpo;
7782 /* Unwind state. */
7783 void *ob_top;
7784 vn_reference_t ref_top;
7785 vn_phi_t phi_top;
7786 vn_nary_op_t nary_top;
7787 vn_avail *avail_top;
7790 /* Unwind the RPO VN state for iteration. */
7792 static void
7793 do_unwind (unwind_state *to, rpo_elim &avail)
7795 gcc_assert (to->iterate);
7796 for (; last_inserted_nary != to->nary_top;
7797 last_inserted_nary = last_inserted_nary->next)
7799 vn_nary_op_t *slot;
7800 slot = valid_info->nary->find_slot_with_hash
7801 (last_inserted_nary, last_inserted_nary->hashcode, NO_INSERT);
7802 /* Predication causes the need to restore previous state. */
7803 if ((*slot)->unwind_to)
7804 *slot = (*slot)->unwind_to;
7805 else
7806 valid_info->nary->clear_slot (slot);
7808 for (; last_inserted_phi != to->phi_top;
7809 last_inserted_phi = last_inserted_phi->next)
7811 vn_phi_t *slot;
7812 slot = valid_info->phis->find_slot_with_hash
7813 (last_inserted_phi, last_inserted_phi->hashcode, NO_INSERT);
7814 valid_info->phis->clear_slot (slot);
7816 for (; last_inserted_ref != to->ref_top;
7817 last_inserted_ref = last_inserted_ref->next)
7819 vn_reference_t *slot;
7820 slot = valid_info->references->find_slot_with_hash
7821 (last_inserted_ref, last_inserted_ref->hashcode, NO_INSERT);
7822 (*slot)->operands.release ();
7823 valid_info->references->clear_slot (slot);
7825 obstack_free (&vn_tables_obstack, to->ob_top);
7827 /* Prune [rpo_idx, ] from avail. */
7828 for (; last_pushed_avail && last_pushed_avail->avail != to->avail_top;)
7830 vn_ssa_aux_t val = last_pushed_avail;
7831 vn_avail *av = val->avail;
7832 val->avail = av->next;
7833 last_pushed_avail = av->next_undo;
7834 av->next = avail.m_avail_freelist;
7835 avail.m_avail_freelist = av;
7839 /* Do VN on a SEME region specified by ENTRY and EXIT_BBS in FN.
7840 If ITERATE is true then treat backedges optimistically as not
7841 executed and iterate. If ELIMINATE is true then perform
7842 elimination, otherwise leave that to the caller. */
7844 unsigned
7845 do_rpo_vn (function *fn, edge entry, bitmap exit_bbs,
7846 bool iterate, bool eliminate, vn_lookup_kind kind)
7848 unsigned todo = 0;
7849 default_vn_walk_kind = kind;
7851 /* We currently do not support region-based iteration when
7852 elimination is requested. */
7853 gcc_assert (!entry || !iterate || !eliminate);
7854 /* When iterating we need loop info up-to-date. */
7855 gcc_assert (!iterate || !loops_state_satisfies_p (LOOPS_NEED_FIXUP));
7857 bool do_region = entry != NULL;
7858 if (!do_region)
7860 entry = single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (fn));
7861 exit_bbs = BITMAP_ALLOC (NULL);
7862 bitmap_set_bit (exit_bbs, EXIT_BLOCK);
7865 /* Clear EDGE_DFS_BACK on "all" entry edges, RPO order compute will
7866 re-mark those that are contained in the region. */
7867 edge_iterator ei;
7868 edge e;
7869 FOR_EACH_EDGE (e, ei, entry->dest->preds)
7870 e->flags &= ~EDGE_DFS_BACK;
7872 int *rpo = XNEWVEC (int, n_basic_blocks_for_fn (fn) - NUM_FIXED_BLOCKS);
7873 auto_vec<std::pair<int, int> > toplevel_scc_extents;
7874 int n = rev_post_order_and_mark_dfs_back_seme
7875 (fn, entry, exit_bbs, true, rpo, !iterate ? &toplevel_scc_extents : NULL);
7877 if (!do_region)
7878 BITMAP_FREE (exit_bbs);
7880 /* If there are any non-DFS_BACK edges into entry->dest skip
7881 processing PHI nodes for that block. This supports
7882 value-numbering loop bodies w/o the actual loop. */
7883 FOR_EACH_EDGE (e, ei, entry->dest->preds)
7884 if (e != entry
7885 && !(e->flags & EDGE_DFS_BACK))
7886 break;
7887 bool skip_entry_phis = e != NULL;
7888 if (skip_entry_phis && dump_file && (dump_flags & TDF_DETAILS))
7889 fprintf (dump_file, "Region does not contain all edges into "
7890 "the entry block, skipping its PHIs.\n");
7892 int *bb_to_rpo = XNEWVEC (int, last_basic_block_for_fn (fn));
7893 for (int i = 0; i < n; ++i)
7894 bb_to_rpo[rpo[i]] = i;
7896 unwind_state *rpo_state = XNEWVEC (unwind_state, n);
7898 rpo_elim avail (entry->dest);
7899 rpo_avail = &avail;
7901 /* Verify we have no extra entries into the region. */
7902 if (flag_checking && do_region)
7904 auto_bb_flag bb_in_region (fn);
7905 for (int i = 0; i < n; ++i)
7907 basic_block bb = BASIC_BLOCK_FOR_FN (fn, rpo[i]);
7908 bb->flags |= bb_in_region;
7910 /* We can't merge the first two loops because we cannot rely
7911 on EDGE_DFS_BACK for edges not within the region. But if
7912 we decide to always have the bb_in_region flag we can
7913 do the checking during the RPO walk itself (but then it's
7914 also easy to handle MEME conservatively). */
7915 for (int i = 0; i < n; ++i)
7917 basic_block bb = BASIC_BLOCK_FOR_FN (fn, rpo[i]);
7918 edge e;
7919 edge_iterator ei;
7920 FOR_EACH_EDGE (e, ei, bb->preds)
7921 gcc_assert (e == entry
7922 || (skip_entry_phis && bb == entry->dest)
7923 || (e->src->flags & bb_in_region));
7925 for (int i = 0; i < n; ++i)
7927 basic_block bb = BASIC_BLOCK_FOR_FN (fn, rpo[i]);
7928 bb->flags &= ~bb_in_region;
7932 /* Create the VN state. For the initial size of the various hashtables
7933 use a heuristic based on region size and number of SSA names. */
7934 unsigned region_size = (((unsigned HOST_WIDE_INT)n * num_ssa_names)
7935 / (n_basic_blocks_for_fn (fn) - NUM_FIXED_BLOCKS));
7936 VN_TOP = create_tmp_var_raw (void_type_node, "vn_top");
7937 next_value_id = 1;
7938 next_constant_value_id = -1;
7940 vn_ssa_aux_hash = new hash_table <vn_ssa_aux_hasher> (region_size * 2);
7941 gcc_obstack_init (&vn_ssa_aux_obstack);
7943 gcc_obstack_init (&vn_tables_obstack);
7944 gcc_obstack_init (&vn_tables_insert_obstack);
7945 valid_info = XCNEW (struct vn_tables_s);
7946 allocate_vn_table (valid_info, region_size);
7947 last_inserted_ref = NULL;
7948 last_inserted_phi = NULL;
7949 last_inserted_nary = NULL;
7950 last_pushed_avail = NULL;
7952 vn_valueize = rpo_vn_valueize;
7954 /* Initialize the unwind state and edge/BB executable state. */
7955 unsigned curr_scc = 0;
7956 for (int i = 0; i < n; ++i)
7958 basic_block bb = BASIC_BLOCK_FOR_FN (fn, rpo[i]);
7959 rpo_state[i].visited = 0;
7960 rpo_state[i].max_rpo = i;
7961 if (!iterate && curr_scc < toplevel_scc_extents.length ())
7963 if (i >= toplevel_scc_extents[curr_scc].first
7964 && i <= toplevel_scc_extents[curr_scc].second)
7965 rpo_state[i].max_rpo = toplevel_scc_extents[curr_scc].second;
7966 if (i == toplevel_scc_extents[curr_scc].second)
7967 curr_scc++;
7969 bb->flags &= ~BB_EXECUTABLE;
7970 bool has_backedges = false;
7971 edge e;
7972 edge_iterator ei;
7973 FOR_EACH_EDGE (e, ei, bb->preds)
7975 if (e->flags & EDGE_DFS_BACK)
7976 has_backedges = true;
7977 e->flags &= ~EDGE_EXECUTABLE;
7978 if (iterate || e == entry || (skip_entry_phis && bb == entry->dest))
7979 continue;
7981 rpo_state[i].iterate = iterate && has_backedges;
7983 entry->flags |= EDGE_EXECUTABLE;
7984 entry->dest->flags |= BB_EXECUTABLE;
7986 /* As heuristic to improve compile-time we handle only the N innermost
7987 loops and the outermost one optimistically. */
7988 if (iterate)
7990 unsigned max_depth = param_rpo_vn_max_loop_depth;
7991 for (auto loop : loops_list (cfun, LI_ONLY_INNERMOST))
7992 if (loop_depth (loop) > max_depth)
7993 for (unsigned i = 2;
7994 i < loop_depth (loop) - max_depth; ++i)
7996 basic_block header = superloop_at_depth (loop, i)->header;
7997 bool non_latch_backedge = false;
7998 edge e;
7999 edge_iterator ei;
8000 FOR_EACH_EDGE (e, ei, header->preds)
8001 if (e->flags & EDGE_DFS_BACK)
8003 /* There can be a non-latch backedge into the header
8004 which is part of an outer irreducible region. We
8005 cannot avoid iterating this block then. */
8006 if (!dominated_by_p (CDI_DOMINATORS,
8007 e->src, e->dest))
8009 if (dump_file && (dump_flags & TDF_DETAILS))
8010 fprintf (dump_file, "non-latch backedge %d -> %d "
8011 "forces iteration of loop %d\n",
8012 e->src->index, e->dest->index, loop->num);
8013 non_latch_backedge = true;
8015 else
8016 e->flags |= EDGE_EXECUTABLE;
8018 rpo_state[bb_to_rpo[header->index]].iterate = non_latch_backedge;
8022 uint64_t nblk = 0;
8023 int idx = 0;
8024 if (iterate)
8025 /* Go and process all blocks, iterating as necessary. */
8028 basic_block bb = BASIC_BLOCK_FOR_FN (fn, rpo[idx]);
8030 /* If the block has incoming backedges remember unwind state. This
8031 is required even for non-executable blocks since in irreducible
8032 regions we might reach them via the backedge and re-start iterating
8033 from there.
8034 Note we can individually mark blocks with incoming backedges to
8035 not iterate where we then handle PHIs conservatively. We do that
8036 heuristically to reduce compile-time for degenerate cases. */
8037 if (rpo_state[idx].iterate)
8039 rpo_state[idx].ob_top = obstack_alloc (&vn_tables_obstack, 0);
8040 rpo_state[idx].ref_top = last_inserted_ref;
8041 rpo_state[idx].phi_top = last_inserted_phi;
8042 rpo_state[idx].nary_top = last_inserted_nary;
8043 rpo_state[idx].avail_top
8044 = last_pushed_avail ? last_pushed_avail->avail : NULL;
8047 if (!(bb->flags & BB_EXECUTABLE))
8049 if (dump_file && (dump_flags & TDF_DETAILS))
8050 fprintf (dump_file, "Block %d: BB%d found not executable\n",
8051 idx, bb->index);
8052 idx++;
8053 continue;
8056 if (dump_file && (dump_flags & TDF_DETAILS))
8057 fprintf (dump_file, "Processing block %d: BB%d\n", idx, bb->index);
8058 nblk++;
8059 todo |= process_bb (avail, bb,
8060 rpo_state[idx].visited != 0,
8061 rpo_state[idx].iterate,
8062 iterate, eliminate, do_region, exit_bbs, false);
8063 rpo_state[idx].visited++;
8065 /* Verify if changed values flow over executable outgoing backedges
8066 and those change destination PHI values (that's the thing we
8067 can easily verify). Reduce over all such edges to the farthest
8068 away PHI. */
8069 int iterate_to = -1;
8070 edge_iterator ei;
8071 edge e;
8072 FOR_EACH_EDGE (e, ei, bb->succs)
8073 if ((e->flags & (EDGE_DFS_BACK|EDGE_EXECUTABLE))
8074 == (EDGE_DFS_BACK|EDGE_EXECUTABLE)
8075 && rpo_state[bb_to_rpo[e->dest->index]].iterate)
8077 int destidx = bb_to_rpo[e->dest->index];
8078 if (!rpo_state[destidx].visited)
8080 if (dump_file && (dump_flags & TDF_DETAILS))
8081 fprintf (dump_file, "Unvisited destination %d\n",
8082 e->dest->index);
8083 if (iterate_to == -1 || destidx < iterate_to)
8084 iterate_to = destidx;
8085 continue;
8087 if (dump_file && (dump_flags & TDF_DETAILS))
8088 fprintf (dump_file, "Looking for changed values of backedge"
8089 " %d->%d destination PHIs\n",
8090 e->src->index, e->dest->index);
8091 vn_context_bb = e->dest;
8092 gphi_iterator gsi;
8093 for (gsi = gsi_start_phis (e->dest);
8094 !gsi_end_p (gsi); gsi_next (&gsi))
8096 bool inserted = false;
8097 /* While we'd ideally just iterate on value changes
8098 we CSE PHIs and do that even across basic-block
8099 boundaries. So even hashtable state changes can
8100 be important (which is roughly equivalent to
8101 PHI argument value changes). To not excessively
8102 iterate because of that we track whether a PHI
8103 was CSEd to with GF_PLF_1. */
8104 bool phival_changed;
8105 if ((phival_changed = visit_phi (gsi.phi (),
8106 &inserted, false))
8107 || (inserted && gimple_plf (gsi.phi (), GF_PLF_1)))
8109 if (!phival_changed
8110 && dump_file && (dump_flags & TDF_DETAILS))
8111 fprintf (dump_file, "PHI was CSEd and hashtable "
8112 "state (changed)\n");
8113 if (iterate_to == -1 || destidx < iterate_to)
8114 iterate_to = destidx;
8115 break;
8118 vn_context_bb = NULL;
8120 if (iterate_to != -1)
8122 do_unwind (&rpo_state[iterate_to], avail);
8123 idx = iterate_to;
8124 if (dump_file && (dump_flags & TDF_DETAILS))
8125 fprintf (dump_file, "Iterating to %d BB%d\n",
8126 iterate_to, rpo[iterate_to]);
8127 continue;
8130 idx++;
8132 while (idx < n);
8134 else /* !iterate */
8136 /* Process all blocks greedily with a worklist that enforces RPO
8137 processing of reachable blocks. */
8138 auto_bitmap worklist;
8139 bitmap_set_bit (worklist, 0);
8140 while (!bitmap_empty_p (worklist))
8142 int idx = bitmap_first_set_bit (worklist);
8143 bitmap_clear_bit (worklist, idx);
8144 basic_block bb = BASIC_BLOCK_FOR_FN (fn, rpo[idx]);
8145 gcc_assert ((bb->flags & BB_EXECUTABLE)
8146 && !rpo_state[idx].visited);
8148 if (dump_file && (dump_flags & TDF_DETAILS))
8149 fprintf (dump_file, "Processing block %d: BB%d\n", idx, bb->index);
8151 /* When we run into predecessor edges where we cannot trust its
8152 executable state mark them executable so PHI processing will
8153 be conservative.
8154 ??? Do we need to force arguments flowing over that edge
8155 to be varying or will they even always be? */
8156 edge_iterator ei;
8157 edge e;
8158 FOR_EACH_EDGE (e, ei, bb->preds)
8159 if (!(e->flags & EDGE_EXECUTABLE)
8160 && (bb == entry->dest
8161 || (!rpo_state[bb_to_rpo[e->src->index]].visited
8162 && (rpo_state[bb_to_rpo[e->src->index]].max_rpo
8163 >= (int)idx))))
8165 if (dump_file && (dump_flags & TDF_DETAILS))
8166 fprintf (dump_file, "Cannot trust state of predecessor "
8167 "edge %d -> %d, marking executable\n",
8168 e->src->index, e->dest->index);
8169 e->flags |= EDGE_EXECUTABLE;
8172 nblk++;
8173 todo |= process_bb (avail, bb, false, false, false, eliminate,
8174 do_region, exit_bbs,
8175 skip_entry_phis && bb == entry->dest);
8176 rpo_state[idx].visited++;
8178 FOR_EACH_EDGE (e, ei, bb->succs)
8179 if ((e->flags & EDGE_EXECUTABLE)
8180 && e->dest->index != EXIT_BLOCK
8181 && (!do_region || !bitmap_bit_p (exit_bbs, e->dest->index))
8182 && !rpo_state[bb_to_rpo[e->dest->index]].visited)
8183 bitmap_set_bit (worklist, bb_to_rpo[e->dest->index]);
8187 /* If statistics or dump file active. */
8188 int nex = 0;
8189 unsigned max_visited = 1;
8190 for (int i = 0; i < n; ++i)
8192 basic_block bb = BASIC_BLOCK_FOR_FN (fn, rpo[i]);
8193 if (bb->flags & BB_EXECUTABLE)
8194 nex++;
8195 statistics_histogram_event (cfun, "RPO block visited times",
8196 rpo_state[i].visited);
8197 if (rpo_state[i].visited > max_visited)
8198 max_visited = rpo_state[i].visited;
8200 unsigned nvalues = 0, navail = 0;
8201 for (hash_table<vn_ssa_aux_hasher>::iterator i = vn_ssa_aux_hash->begin ();
8202 i != vn_ssa_aux_hash->end (); ++i)
8204 nvalues++;
8205 vn_avail *av = (*i)->avail;
8206 while (av)
8208 navail++;
8209 av = av->next;
8212 statistics_counter_event (cfun, "RPO blocks", n);
8213 statistics_counter_event (cfun, "RPO blocks visited", nblk);
8214 statistics_counter_event (cfun, "RPO blocks executable", nex);
8215 statistics_histogram_event (cfun, "RPO iterations", 10*nblk / nex);
8216 statistics_histogram_event (cfun, "RPO num values", nvalues);
8217 statistics_histogram_event (cfun, "RPO num avail", navail);
8218 statistics_histogram_event (cfun, "RPO num lattice",
8219 vn_ssa_aux_hash->elements ());
8220 if (dump_file && (dump_flags & (TDF_DETAILS|TDF_STATS)))
8222 fprintf (dump_file, "RPO iteration over %d blocks visited %" PRIu64
8223 " blocks in total discovering %d executable blocks iterating "
8224 "%d.%d times, a block was visited max. %u times\n",
8225 n, nblk, nex,
8226 (int)((10*nblk / nex)/10), (int)((10*nblk / nex)%10),
8227 max_visited);
8228 fprintf (dump_file, "RPO tracked %d values available at %d locations "
8229 "and %" PRIu64 " lattice elements\n",
8230 nvalues, navail, (uint64_t) vn_ssa_aux_hash->elements ());
8233 if (eliminate)
8235 /* When !iterate we already performed elimination during the RPO
8236 walk. */
8237 if (iterate)
8239 /* Elimination for region-based VN needs to be done within the
8240 RPO walk. */
8241 gcc_assert (! do_region);
8242 /* Note we can't use avail.walk here because that gets confused
8243 by the existing availability and it will be less efficient
8244 as well. */
8245 todo |= eliminate_with_rpo_vn (NULL);
8247 else
8248 todo |= avail.eliminate_cleanup (do_region);
8251 vn_valueize = NULL;
8252 rpo_avail = NULL;
8254 XDELETEVEC (bb_to_rpo);
8255 XDELETEVEC (rpo);
8256 XDELETEVEC (rpo_state);
8258 return todo;
8261 /* Region-based entry for RPO VN. Performs value-numbering and elimination
8262 on the SEME region specified by ENTRY and EXIT_BBS. If ENTRY is not
8263 the only edge into the region at ENTRY->dest PHI nodes in ENTRY->dest
8264 are not considered. */
8266 unsigned
8267 do_rpo_vn (function *fn, edge entry, bitmap exit_bbs)
8269 unsigned todo = do_rpo_vn (fn, entry, exit_bbs, false, true, VN_WALKREWRITE);
8270 free_rpo_vn ();
8271 return todo;
8275 namespace {
8277 const pass_data pass_data_fre =
8279 GIMPLE_PASS, /* type */
8280 "fre", /* name */
8281 OPTGROUP_NONE, /* optinfo_flags */
8282 TV_TREE_FRE, /* tv_id */
8283 ( PROP_cfg | PROP_ssa ), /* properties_required */
8284 0, /* properties_provided */
8285 0, /* properties_destroyed */
8286 0, /* todo_flags_start */
8287 0, /* todo_flags_finish */
8290 class pass_fre : public gimple_opt_pass
8292 public:
8293 pass_fre (gcc::context *ctxt)
8294 : gimple_opt_pass (pass_data_fre, ctxt), may_iterate (true)
8297 /* opt_pass methods: */
8298 opt_pass * clone () { return new pass_fre (m_ctxt); }
8299 void set_pass_param (unsigned int n, bool param)
8301 gcc_assert (n == 0);
8302 may_iterate = param;
8304 virtual bool gate (function *)
8306 return flag_tree_fre != 0 && (may_iterate || optimize > 1);
8308 virtual unsigned int execute (function *);
8310 private:
8311 bool may_iterate;
8312 }; // class pass_fre
8314 unsigned int
8315 pass_fre::execute (function *fun)
8317 unsigned todo = 0;
8319 /* At -O[1g] use the cheap non-iterating mode. */
8320 bool iterate_p = may_iterate && (optimize > 1);
8321 calculate_dominance_info (CDI_DOMINATORS);
8322 if (iterate_p)
8323 loop_optimizer_init (AVOID_CFG_MODIFICATIONS);
8325 todo = do_rpo_vn (fun, NULL, NULL, iterate_p, true, VN_WALKREWRITE);
8326 free_rpo_vn ();
8328 if (iterate_p)
8329 loop_optimizer_finalize ();
8331 if (scev_initialized_p ())
8332 scev_reset_htab ();
8334 /* For late FRE after IVOPTs and unrolling, see if we can
8335 remove some TREE_ADDRESSABLE and rewrite stuff into SSA. */
8336 if (!may_iterate)
8337 todo |= TODO_update_address_taken;
8339 return todo;
8342 } // anon namespace
8344 gimple_opt_pass *
8345 make_pass_fre (gcc::context *ctxt)
8347 return new pass_fre (ctxt);
8350 #undef BB_EXECUTABLE