d: Add test for PR d/108167 to the testsuite [PR108167]
[official-gcc.git] / gcc / tree-ssa-sccvn.cc
blobd5b081a309fe1ed93a3e8b36371bd2365a6979a5
1 /* SCC value numbering for trees
2 Copyright (C) 2006-2023 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 rhs_off;
1794 HOST_WIDE_INT offset;
1795 HOST_WIDE_INT size;
1798 /* Context for alias walking. */
1800 struct vn_walk_cb_data
1802 vn_walk_cb_data (vn_reference_t vr_, tree orig_ref_, tree *last_vuse_ptr_,
1803 vn_lookup_kind vn_walk_kind_, bool tbaa_p_, tree mask_,
1804 bool redundant_store_removal_p_)
1805 : vr (vr_), last_vuse_ptr (last_vuse_ptr_), last_vuse (NULL_TREE),
1806 mask (mask_), masked_result (NULL_TREE), same_val (NULL_TREE),
1807 vn_walk_kind (vn_walk_kind_),
1808 tbaa_p (tbaa_p_), redundant_store_removal_p (redundant_store_removal_p_),
1809 saved_operands (vNULL), first_set (-2), first_base_set (-2),
1810 known_ranges (NULL)
1812 if (!last_vuse_ptr)
1813 last_vuse_ptr = &last_vuse;
1814 ao_ref_init (&orig_ref, orig_ref_);
1815 if (mask)
1817 wide_int w = wi::to_wide (mask);
1818 unsigned int pos = 0, prec = w.get_precision ();
1819 pd_data pd;
1820 pd.rhs = build_constructor (NULL_TREE, NULL);
1821 pd.rhs_off = 0;
1822 /* When bitwise and with a constant is done on a memory load,
1823 we don't really need all the bits to be defined or defined
1824 to constants, we don't really care what is in the position
1825 corresponding to 0 bits in the mask.
1826 So, push the ranges of those 0 bits in the mask as artificial
1827 zero stores and let the partial def handling code do the
1828 rest. */
1829 while (pos < prec)
1831 int tz = wi::ctz (w);
1832 if (pos + tz > prec)
1833 tz = prec - pos;
1834 if (tz)
1836 if (BYTES_BIG_ENDIAN)
1837 pd.offset = prec - pos - tz;
1838 else
1839 pd.offset = pos;
1840 pd.size = tz;
1841 void *r = push_partial_def (pd, 0, 0, 0, prec);
1842 gcc_assert (r == NULL_TREE);
1844 pos += tz;
1845 if (pos == prec)
1846 break;
1847 w = wi::lrshift (w, tz);
1848 tz = wi::ctz (wi::bit_not (w));
1849 if (pos + tz > prec)
1850 tz = prec - pos;
1851 pos += tz;
1852 w = wi::lrshift (w, tz);
1856 ~vn_walk_cb_data ();
1857 void *finish (alias_set_type, alias_set_type, tree);
1858 void *push_partial_def (pd_data pd,
1859 alias_set_type, alias_set_type, HOST_WIDE_INT,
1860 HOST_WIDE_INT);
1862 vn_reference_t vr;
1863 ao_ref orig_ref;
1864 tree *last_vuse_ptr;
1865 tree last_vuse;
1866 tree mask;
1867 tree masked_result;
1868 tree same_val;
1869 vn_lookup_kind vn_walk_kind;
1870 bool tbaa_p;
1871 bool redundant_store_removal_p;
1872 vec<vn_reference_op_s> saved_operands;
1874 /* The VDEFs of partial defs we come along. */
1875 auto_vec<pd_data, 2> partial_defs;
1876 /* The first defs range to avoid splay tree setup in most cases. */
1877 pd_range first_range;
1878 alias_set_type first_set;
1879 alias_set_type first_base_set;
1880 splay_tree known_ranges;
1881 obstack ranges_obstack;
1884 vn_walk_cb_data::~vn_walk_cb_data ()
1886 if (known_ranges)
1888 splay_tree_delete (known_ranges);
1889 obstack_free (&ranges_obstack, NULL);
1891 saved_operands.release ();
1894 void *
1895 vn_walk_cb_data::finish (alias_set_type set, alias_set_type base_set, tree val)
1897 if (first_set != -2)
1899 set = first_set;
1900 base_set = first_base_set;
1902 if (mask)
1904 masked_result = val;
1905 return (void *) -1;
1907 if (same_val && !operand_equal_p (val, same_val))
1908 return (void *) -1;
1909 vec<vn_reference_op_s> &operands
1910 = saved_operands.exists () ? saved_operands : vr->operands;
1911 return vn_reference_lookup_or_insert_for_pieces (last_vuse, set, base_set,
1912 vr->type, operands, val);
1915 /* pd_range splay-tree helpers. */
1917 static int
1918 pd_range_compare (splay_tree_key offset1p, splay_tree_key offset2p)
1920 HOST_WIDE_INT offset1 = *(HOST_WIDE_INT *)offset1p;
1921 HOST_WIDE_INT offset2 = *(HOST_WIDE_INT *)offset2p;
1922 if (offset1 < offset2)
1923 return -1;
1924 else if (offset1 > offset2)
1925 return 1;
1926 return 0;
1929 static void *
1930 pd_tree_alloc (int size, void *data_)
1932 vn_walk_cb_data *data = (vn_walk_cb_data *)data_;
1933 return obstack_alloc (&data->ranges_obstack, size);
1936 static void
1937 pd_tree_dealloc (void *, void *)
1941 /* Push PD to the vector of partial definitions returning a
1942 value when we are ready to combine things with VUSE, SET and MAXSIZEI,
1943 NULL when we want to continue looking for partial defs or -1
1944 on failure. */
1946 void *
1947 vn_walk_cb_data::push_partial_def (pd_data pd,
1948 alias_set_type set, alias_set_type base_set,
1949 HOST_WIDE_INT offseti,
1950 HOST_WIDE_INT maxsizei)
1952 const HOST_WIDE_INT bufsize = 64;
1953 /* We're using a fixed buffer for encoding so fail early if the object
1954 we want to interpret is bigger. */
1955 if (maxsizei > bufsize * BITS_PER_UNIT
1956 || CHAR_BIT != 8
1957 || BITS_PER_UNIT != 8
1958 /* Not prepared to handle PDP endian. */
1959 || BYTES_BIG_ENDIAN != WORDS_BIG_ENDIAN)
1960 return (void *)-1;
1962 /* Turn too large constant stores into non-constant stores. */
1963 if (CONSTANT_CLASS_P (pd.rhs) && pd.size > bufsize * BITS_PER_UNIT)
1964 pd.rhs = error_mark_node;
1966 /* And for non-constant or CONSTRUCTOR stores shrink them to only keep at
1967 most a partial byte before and/or after the region. */
1968 if (!CONSTANT_CLASS_P (pd.rhs))
1970 if (pd.offset < offseti)
1972 HOST_WIDE_INT o = ROUND_DOWN (offseti - pd.offset, BITS_PER_UNIT);
1973 gcc_assert (pd.size > o);
1974 pd.size -= o;
1975 pd.offset += o;
1977 if (pd.size > maxsizei)
1978 pd.size = maxsizei + ((pd.size - maxsizei) % BITS_PER_UNIT);
1981 pd.offset -= offseti;
1983 bool pd_constant_p = (TREE_CODE (pd.rhs) == CONSTRUCTOR
1984 || CONSTANT_CLASS_P (pd.rhs));
1985 pd_range *r;
1986 if (partial_defs.is_empty ())
1988 /* If we get a clobber upfront, fail. */
1989 if (TREE_CLOBBER_P (pd.rhs))
1990 return (void *)-1;
1991 if (!pd_constant_p)
1992 return (void *)-1;
1993 partial_defs.safe_push (pd);
1994 first_range.offset = pd.offset;
1995 first_range.size = pd.size;
1996 first_set = set;
1997 first_base_set = base_set;
1998 last_vuse_ptr = NULL;
1999 r = &first_range;
2000 /* Go check if the first partial definition was a full one in case
2001 the caller didn't optimize for this. */
2003 else
2005 if (!known_ranges)
2007 /* ??? Optimize the case where the 2nd partial def completes
2008 things. */
2009 gcc_obstack_init (&ranges_obstack);
2010 known_ranges = splay_tree_new_with_allocator (pd_range_compare, 0, 0,
2011 pd_tree_alloc,
2012 pd_tree_dealloc, this);
2013 splay_tree_insert (known_ranges,
2014 (splay_tree_key)&first_range.offset,
2015 (splay_tree_value)&first_range);
2018 pd_range newr = { pd.offset, pd.size };
2019 splay_tree_node n;
2020 /* Lookup the predecessor of offset + 1 and see if we need to merge. */
2021 HOST_WIDE_INT loffset = newr.offset + 1;
2022 if ((n = splay_tree_predecessor (known_ranges, (splay_tree_key)&loffset))
2023 && ((r = (pd_range *)n->value), true)
2024 && ranges_known_overlap_p (r->offset, r->size + 1,
2025 newr.offset, newr.size))
2027 /* Ignore partial defs already covered. Here we also drop shadowed
2028 clobbers arriving here at the floor. */
2029 if (known_subrange_p (newr.offset, newr.size, r->offset, r->size))
2030 return NULL;
2031 r->size
2032 = MAX (r->offset + r->size, newr.offset + newr.size) - r->offset;
2034 else
2036 /* newr.offset wasn't covered yet, insert the range. */
2037 r = XOBNEW (&ranges_obstack, pd_range);
2038 *r = newr;
2039 splay_tree_insert (known_ranges, (splay_tree_key)&r->offset,
2040 (splay_tree_value)r);
2042 /* Merge r which now contains newr and is a member of the splay tree with
2043 adjacent overlapping ranges. */
2044 pd_range *rafter;
2045 while ((n = splay_tree_successor (known_ranges,
2046 (splay_tree_key)&r->offset))
2047 && ((rafter = (pd_range *)n->value), true)
2048 && ranges_known_overlap_p (r->offset, r->size + 1,
2049 rafter->offset, rafter->size))
2051 r->size = MAX (r->offset + r->size,
2052 rafter->offset + rafter->size) - r->offset;
2053 splay_tree_remove (known_ranges, (splay_tree_key)&rafter->offset);
2055 /* If we get a clobber, fail. */
2056 if (TREE_CLOBBER_P (pd.rhs))
2057 return (void *)-1;
2058 /* Non-constants are OK as long as they are shadowed by a constant. */
2059 if (!pd_constant_p)
2060 return (void *)-1;
2061 partial_defs.safe_push (pd);
2064 /* Now we have merged newr into the range tree. When we have covered
2065 [offseti, sizei] then the tree will contain exactly one node which has
2066 the desired properties and it will be 'r'. */
2067 if (!known_subrange_p (0, maxsizei, r->offset, r->size))
2068 /* Continue looking for partial defs. */
2069 return NULL;
2071 /* Now simply native encode all partial defs in reverse order. */
2072 unsigned ndefs = partial_defs.length ();
2073 /* We support up to 512-bit values (for V8DFmode). */
2074 unsigned char buffer[bufsize + 1];
2075 unsigned char this_buffer[bufsize + 1];
2076 int len;
2078 memset (buffer, 0, bufsize + 1);
2079 unsigned needed_len = ROUND_UP (maxsizei, BITS_PER_UNIT) / BITS_PER_UNIT;
2080 while (!partial_defs.is_empty ())
2082 pd_data pd = partial_defs.pop ();
2083 unsigned int amnt;
2084 if (TREE_CODE (pd.rhs) == CONSTRUCTOR)
2086 /* Empty CONSTRUCTOR. */
2087 if (pd.size >= needed_len * BITS_PER_UNIT)
2088 len = needed_len;
2089 else
2090 len = ROUND_UP (pd.size, BITS_PER_UNIT) / BITS_PER_UNIT;
2091 memset (this_buffer, 0, len);
2093 else if (pd.rhs_off >= 0)
2095 len = native_encode_expr (pd.rhs, this_buffer, bufsize,
2096 (MAX (0, -pd.offset)
2097 + pd.rhs_off) / BITS_PER_UNIT);
2098 if (len <= 0
2099 || len < (ROUND_UP (pd.size, BITS_PER_UNIT) / BITS_PER_UNIT
2100 - MAX (0, -pd.offset) / BITS_PER_UNIT))
2102 if (dump_file && (dump_flags & TDF_DETAILS))
2103 fprintf (dump_file, "Failed to encode %u "
2104 "partial definitions\n", ndefs);
2105 return (void *)-1;
2108 else /* negative pd.rhs_off indicates we want to chop off first bits */
2110 if (-pd.rhs_off >= bufsize)
2111 return (void *)-1;
2112 len = native_encode_expr (pd.rhs,
2113 this_buffer + -pd.rhs_off / BITS_PER_UNIT,
2114 bufsize - -pd.rhs_off / BITS_PER_UNIT,
2115 MAX (0, -pd.offset) / BITS_PER_UNIT);
2116 if (len <= 0
2117 || len < (ROUND_UP (pd.size, BITS_PER_UNIT) / BITS_PER_UNIT
2118 - MAX (0, -pd.offset) / BITS_PER_UNIT))
2120 if (dump_file && (dump_flags & TDF_DETAILS))
2121 fprintf (dump_file, "Failed to encode %u "
2122 "partial definitions\n", ndefs);
2123 return (void *)-1;
2127 unsigned char *p = buffer;
2128 HOST_WIDE_INT size = pd.size;
2129 if (pd.offset < 0)
2130 size -= ROUND_DOWN (-pd.offset, BITS_PER_UNIT);
2131 this_buffer[len] = 0;
2132 if (BYTES_BIG_ENDIAN)
2134 /* LSB of this_buffer[len - 1] byte should be at
2135 pd.offset + pd.size - 1 bits in buffer. */
2136 amnt = ((unsigned HOST_WIDE_INT) pd.offset
2137 + pd.size) % BITS_PER_UNIT;
2138 if (amnt)
2139 shift_bytes_in_array_right (this_buffer, len + 1, amnt);
2140 unsigned char *q = this_buffer;
2141 unsigned int off = 0;
2142 if (pd.offset >= 0)
2144 unsigned int msk;
2145 off = pd.offset / BITS_PER_UNIT;
2146 gcc_assert (off < needed_len);
2147 p = buffer + off;
2148 if (size <= amnt)
2150 msk = ((1 << size) - 1) << (BITS_PER_UNIT - amnt);
2151 *p = (*p & ~msk) | (this_buffer[len] & msk);
2152 size = 0;
2154 else
2156 if (TREE_CODE (pd.rhs) != CONSTRUCTOR)
2157 q = (this_buffer + len
2158 - (ROUND_UP (size - amnt, BITS_PER_UNIT)
2159 / BITS_PER_UNIT));
2160 if (pd.offset % BITS_PER_UNIT)
2162 msk = -1U << (BITS_PER_UNIT
2163 - (pd.offset % BITS_PER_UNIT));
2164 *p = (*p & msk) | (*q & ~msk);
2165 p++;
2166 q++;
2167 off++;
2168 size -= BITS_PER_UNIT - (pd.offset % BITS_PER_UNIT);
2169 gcc_assert (size >= 0);
2173 else if (TREE_CODE (pd.rhs) != CONSTRUCTOR)
2175 q = (this_buffer + len
2176 - (ROUND_UP (size - amnt, BITS_PER_UNIT)
2177 / BITS_PER_UNIT));
2178 if (pd.offset % BITS_PER_UNIT)
2180 q++;
2181 size -= BITS_PER_UNIT - ((unsigned HOST_WIDE_INT) pd.offset
2182 % BITS_PER_UNIT);
2183 gcc_assert (size >= 0);
2186 if ((unsigned HOST_WIDE_INT) size / BITS_PER_UNIT + off
2187 > needed_len)
2188 size = (needed_len - off) * BITS_PER_UNIT;
2189 memcpy (p, q, size / BITS_PER_UNIT);
2190 if (size % BITS_PER_UNIT)
2192 unsigned int msk
2193 = -1U << (BITS_PER_UNIT - (size % BITS_PER_UNIT));
2194 p += size / BITS_PER_UNIT;
2195 q += size / BITS_PER_UNIT;
2196 *p = (*q & msk) | (*p & ~msk);
2199 else
2201 if (pd.offset >= 0)
2203 /* LSB of this_buffer[0] byte should be at pd.offset bits
2204 in buffer. */
2205 unsigned int msk;
2206 size = MIN (size, (HOST_WIDE_INT) needed_len * BITS_PER_UNIT);
2207 amnt = pd.offset % BITS_PER_UNIT;
2208 if (amnt)
2209 shift_bytes_in_array_left (this_buffer, len + 1, amnt);
2210 unsigned int off = pd.offset / BITS_PER_UNIT;
2211 gcc_assert (off < needed_len);
2212 size = MIN (size,
2213 (HOST_WIDE_INT) (needed_len - off) * BITS_PER_UNIT);
2214 p = buffer + off;
2215 if (amnt + size < BITS_PER_UNIT)
2217 /* Low amnt bits come from *p, then size bits
2218 from this_buffer[0] and the remaining again from
2219 *p. */
2220 msk = ((1 << size) - 1) << amnt;
2221 *p = (*p & ~msk) | (this_buffer[0] & msk);
2222 size = 0;
2224 else if (amnt)
2226 msk = -1U << amnt;
2227 *p = (*p & ~msk) | (this_buffer[0] & msk);
2228 p++;
2229 size -= (BITS_PER_UNIT - amnt);
2232 else
2234 amnt = (unsigned HOST_WIDE_INT) pd.offset % BITS_PER_UNIT;
2235 if (amnt)
2236 size -= BITS_PER_UNIT - amnt;
2237 size = MIN (size, (HOST_WIDE_INT) needed_len * BITS_PER_UNIT);
2238 if (amnt)
2239 shift_bytes_in_array_left (this_buffer, len + 1, amnt);
2241 memcpy (p, this_buffer + (amnt != 0), size / BITS_PER_UNIT);
2242 p += size / BITS_PER_UNIT;
2243 if (size % BITS_PER_UNIT)
2245 unsigned int msk = -1U << (size % BITS_PER_UNIT);
2246 *p = (this_buffer[(amnt != 0) + size / BITS_PER_UNIT]
2247 & ~msk) | (*p & msk);
2252 tree type = vr->type;
2253 /* Make sure to interpret in a type that has a range covering the whole
2254 access size. */
2255 if (INTEGRAL_TYPE_P (vr->type) && maxsizei != TYPE_PRECISION (vr->type))
2256 type = build_nonstandard_integer_type (maxsizei, TYPE_UNSIGNED (type));
2257 tree val;
2258 if (BYTES_BIG_ENDIAN)
2260 unsigned sz = needed_len;
2261 if (maxsizei % BITS_PER_UNIT)
2262 shift_bytes_in_array_right (buffer, needed_len,
2263 BITS_PER_UNIT
2264 - (maxsizei % BITS_PER_UNIT));
2265 if (INTEGRAL_TYPE_P (type))
2266 sz = GET_MODE_SIZE (SCALAR_INT_TYPE_MODE (type));
2267 if (sz > needed_len)
2269 memcpy (this_buffer + (sz - needed_len), buffer, needed_len);
2270 val = native_interpret_expr (type, this_buffer, sz);
2272 else
2273 val = native_interpret_expr (type, buffer, needed_len);
2275 else
2276 val = native_interpret_expr (type, buffer, bufsize);
2277 /* If we chop off bits because the types precision doesn't match the memory
2278 access size this is ok when optimizing reads but not when called from
2279 the DSE code during elimination. */
2280 if (val && type != vr->type)
2282 if (! int_fits_type_p (val, vr->type))
2283 val = NULL_TREE;
2284 else
2285 val = fold_convert (vr->type, val);
2288 if (val)
2290 if (dump_file && (dump_flags & TDF_DETAILS))
2291 fprintf (dump_file,
2292 "Successfully combined %u partial definitions\n", ndefs);
2293 /* We are using the alias-set of the first store we encounter which
2294 should be appropriate here. */
2295 return finish (first_set, first_base_set, val);
2297 else
2299 if (dump_file && (dump_flags & TDF_DETAILS))
2300 fprintf (dump_file,
2301 "Failed to interpret %u encoded partial definitions\n", ndefs);
2302 return (void *)-1;
2306 /* Callback for walk_non_aliased_vuses. Adjusts the vn_reference_t VR_
2307 with the current VUSE and performs the expression lookup. */
2309 static void *
2310 vn_reference_lookup_2 (ao_ref *op ATTRIBUTE_UNUSED, tree vuse, void *data_)
2312 vn_walk_cb_data *data = (vn_walk_cb_data *)data_;
2313 vn_reference_t vr = data->vr;
2314 vn_reference_s **slot;
2315 hashval_t hash;
2317 /* If we have partial definitions recorded we have to go through
2318 vn_reference_lookup_3. */
2319 if (!data->partial_defs.is_empty ())
2320 return NULL;
2322 if (data->last_vuse_ptr)
2324 *data->last_vuse_ptr = vuse;
2325 data->last_vuse = vuse;
2328 /* Fixup vuse and hash. */
2329 if (vr->vuse)
2330 vr->hashcode = vr->hashcode - SSA_NAME_VERSION (vr->vuse);
2331 vr->vuse = vuse_ssa_val (vuse);
2332 if (vr->vuse)
2333 vr->hashcode = vr->hashcode + SSA_NAME_VERSION (vr->vuse);
2335 hash = vr->hashcode;
2336 slot = valid_info->references->find_slot_with_hash (vr, hash, NO_INSERT);
2337 if (slot)
2339 if ((*slot)->result && data->saved_operands.exists ())
2340 return data->finish (vr->set, vr->base_set, (*slot)->result);
2341 return *slot;
2344 return NULL;
2347 /* Lookup an existing or insert a new vn_reference entry into the
2348 value table for the VUSE, SET, TYPE, OPERANDS reference which
2349 has the value VALUE which is either a constant or an SSA name. */
2351 static vn_reference_t
2352 vn_reference_lookup_or_insert_for_pieces (tree vuse,
2353 alias_set_type set,
2354 alias_set_type base_set,
2355 tree type,
2356 vec<vn_reference_op_s,
2357 va_heap> operands,
2358 tree value)
2360 vn_reference_s vr1;
2361 vn_reference_t result;
2362 unsigned value_id;
2363 vr1.vuse = vuse ? SSA_VAL (vuse) : NULL_TREE;
2364 vr1.operands = operands;
2365 vr1.type = type;
2366 vr1.set = set;
2367 vr1.base_set = base_set;
2368 vr1.hashcode = vn_reference_compute_hash (&vr1);
2369 if (vn_reference_lookup_1 (&vr1, &result))
2370 return result;
2371 if (TREE_CODE (value) == SSA_NAME)
2372 value_id = VN_INFO (value)->value_id;
2373 else
2374 value_id = get_or_alloc_constant_value_id (value);
2375 return vn_reference_insert_pieces (vuse, set, base_set, type,
2376 operands.copy (), value, value_id);
2379 /* Return a value-number for RCODE OPS... either by looking up an existing
2380 value-number for the possibly simplified result or by inserting the
2381 operation if INSERT is true. If SIMPLIFY is false, return a value
2382 number for the unsimplified expression. */
2384 static tree
2385 vn_nary_build_or_lookup_1 (gimple_match_op *res_op, bool insert,
2386 bool simplify)
2388 tree result = NULL_TREE;
2389 /* We will be creating a value number for
2390 RCODE (OPS...).
2391 So first simplify and lookup this expression to see if it
2392 is already available. */
2393 /* For simplification valueize. */
2394 unsigned i = 0;
2395 if (simplify)
2396 for (i = 0; i < res_op->num_ops; ++i)
2397 if (TREE_CODE (res_op->ops[i]) == SSA_NAME)
2399 tree tem = vn_valueize (res_op->ops[i]);
2400 if (!tem)
2401 break;
2402 res_op->ops[i] = tem;
2404 /* If valueization of an operand fails (it is not available), skip
2405 simplification. */
2406 bool res = false;
2407 if (i == res_op->num_ops)
2409 mprts_hook = vn_lookup_simplify_result;
2410 res = res_op->resimplify (NULL, vn_valueize);
2411 mprts_hook = NULL;
2413 gimple *new_stmt = NULL;
2414 if (res
2415 && gimple_simplified_result_is_gimple_val (res_op))
2417 /* The expression is already available. */
2418 result = res_op->ops[0];
2419 /* Valueize it, simplification returns sth in AVAIL only. */
2420 if (TREE_CODE (result) == SSA_NAME)
2421 result = SSA_VAL (result);
2423 else
2425 tree val = vn_lookup_simplify_result (res_op);
2426 if (!val && insert)
2428 gimple_seq stmts = NULL;
2429 result = maybe_push_res_to_seq (res_op, &stmts);
2430 if (result)
2432 gcc_assert (gimple_seq_singleton_p (stmts));
2433 new_stmt = gimple_seq_first_stmt (stmts);
2436 else
2437 /* The expression is already available. */
2438 result = val;
2440 if (new_stmt)
2442 /* The expression is not yet available, value-number lhs to
2443 the new SSA_NAME we created. */
2444 /* Initialize value-number information properly. */
2445 vn_ssa_aux_t result_info = VN_INFO (result);
2446 result_info->valnum = result;
2447 result_info->value_id = get_next_value_id ();
2448 result_info->visited = 1;
2449 gimple_seq_add_stmt_without_update (&VN_INFO (result)->expr,
2450 new_stmt);
2451 result_info->needs_insertion = true;
2452 /* ??? PRE phi-translation inserts NARYs without corresponding
2453 SSA name result. Re-use those but set their result according
2454 to the stmt we just built. */
2455 vn_nary_op_t nary = NULL;
2456 vn_nary_op_lookup_stmt (new_stmt, &nary);
2457 if (nary)
2459 gcc_assert (! nary->predicated_values && nary->u.result == NULL_TREE);
2460 nary->u.result = gimple_assign_lhs (new_stmt);
2462 /* As all "inserted" statements are singleton SCCs, insert
2463 to the valid table. This is strictly needed to
2464 avoid re-generating new value SSA_NAMEs for the same
2465 expression during SCC iteration over and over (the
2466 optimistic table gets cleared after each iteration).
2467 We do not need to insert into the optimistic table, as
2468 lookups there will fall back to the valid table. */
2469 else
2471 unsigned int length = vn_nary_length_from_stmt (new_stmt);
2472 vn_nary_op_t vno1
2473 = alloc_vn_nary_op_noinit (length, &vn_tables_insert_obstack);
2474 vno1->value_id = result_info->value_id;
2475 vno1->length = length;
2476 vno1->predicated_values = 0;
2477 vno1->u.result = result;
2478 init_vn_nary_op_from_stmt (vno1, as_a <gassign *> (new_stmt));
2479 vn_nary_op_insert_into (vno1, valid_info->nary);
2480 /* Also do not link it into the undo chain. */
2481 last_inserted_nary = vno1->next;
2482 vno1->next = (vn_nary_op_t)(void *)-1;
2484 if (dump_file && (dump_flags & TDF_DETAILS))
2486 fprintf (dump_file, "Inserting name ");
2487 print_generic_expr (dump_file, result);
2488 fprintf (dump_file, " for expression ");
2489 print_gimple_expr (dump_file, new_stmt, 0, TDF_SLIM);
2490 fprintf (dump_file, "\n");
2493 return result;
2496 /* Return a value-number for RCODE OPS... either by looking up an existing
2497 value-number for the simplified result or by inserting the operation. */
2499 static tree
2500 vn_nary_build_or_lookup (gimple_match_op *res_op)
2502 return vn_nary_build_or_lookup_1 (res_op, true, true);
2505 /* Try to simplify the expression RCODE OPS... of type TYPE and return
2506 its value if present. */
2508 tree
2509 vn_nary_simplify (vn_nary_op_t nary)
2511 if (nary->length > gimple_match_op::MAX_NUM_OPS)
2512 return NULL_TREE;
2513 gimple_match_op op (gimple_match_cond::UNCOND, nary->opcode,
2514 nary->type, nary->length);
2515 memcpy (op.ops, nary->op, sizeof (tree) * nary->length);
2516 return vn_nary_build_or_lookup_1 (&op, false, true);
2519 /* Elimination engine. */
2521 class eliminate_dom_walker : public dom_walker
2523 public:
2524 eliminate_dom_walker (cdi_direction, bitmap);
2525 ~eliminate_dom_walker ();
2527 edge before_dom_children (basic_block) final override;
2528 void after_dom_children (basic_block) final override;
2530 virtual tree eliminate_avail (basic_block, tree op);
2531 virtual void eliminate_push_avail (basic_block, tree op);
2532 tree eliminate_insert (basic_block, gimple_stmt_iterator *gsi, tree val);
2534 void eliminate_stmt (basic_block, gimple_stmt_iterator *);
2536 unsigned eliminate_cleanup (bool region_p = false);
2538 bool do_pre;
2539 unsigned int el_todo;
2540 unsigned int eliminations;
2541 unsigned int insertions;
2543 /* SSA names that had their defs inserted by PRE if do_pre. */
2544 bitmap inserted_exprs;
2546 /* Blocks with statements that have had their EH properties changed. */
2547 bitmap need_eh_cleanup;
2549 /* Blocks with statements that have had their AB properties changed. */
2550 bitmap need_ab_cleanup;
2552 /* Local state for the eliminate domwalk. */
2553 auto_vec<gimple *> to_remove;
2554 auto_vec<gimple *> to_fixup;
2555 auto_vec<tree> avail;
2556 auto_vec<tree> avail_stack;
2559 /* Adaptor to the elimination engine using RPO availability. */
2561 class rpo_elim : public eliminate_dom_walker
2563 public:
2564 rpo_elim(basic_block entry_)
2565 : eliminate_dom_walker (CDI_DOMINATORS, NULL), entry (entry_),
2566 m_avail_freelist (NULL) {}
2568 tree eliminate_avail (basic_block, tree op) final override;
2570 void eliminate_push_avail (basic_block, tree) final override;
2572 basic_block entry;
2573 /* Freelist of avail entries which are allocated from the vn_ssa_aux
2574 obstack. */
2575 vn_avail *m_avail_freelist;
2578 /* Global RPO state for access from hooks. */
2579 static eliminate_dom_walker *rpo_avail;
2580 basic_block vn_context_bb;
2582 /* Return true if BASE1 and BASE2 can be adjusted so they have the
2583 same address and adjust *OFFSET1 and *OFFSET2 accordingly.
2584 Otherwise return false. */
2586 static bool
2587 adjust_offsets_for_equal_base_address (tree base1, poly_int64 *offset1,
2588 tree base2, poly_int64 *offset2)
2590 poly_int64 soff;
2591 if (TREE_CODE (base1) == MEM_REF
2592 && TREE_CODE (base2) == MEM_REF)
2594 if (mem_ref_offset (base1).to_shwi (&soff))
2596 base1 = TREE_OPERAND (base1, 0);
2597 *offset1 += soff * BITS_PER_UNIT;
2599 if (mem_ref_offset (base2).to_shwi (&soff))
2601 base2 = TREE_OPERAND (base2, 0);
2602 *offset2 += soff * BITS_PER_UNIT;
2604 return operand_equal_p (base1, base2, 0);
2606 return operand_equal_p (base1, base2, OEP_ADDRESS_OF);
2609 /* Callback for walk_non_aliased_vuses. Tries to perform a lookup
2610 from the statement defining VUSE and if not successful tries to
2611 translate *REFP and VR_ through an aggregate copy at the definition
2612 of VUSE. If *DISAMBIGUATE_ONLY is true then do not perform translation
2613 of *REF and *VR. If only disambiguation was performed then
2614 *DISAMBIGUATE_ONLY is set to true. */
2616 static void *
2617 vn_reference_lookup_3 (ao_ref *ref, tree vuse, void *data_,
2618 translate_flags *disambiguate_only)
2620 vn_walk_cb_data *data = (vn_walk_cb_data *)data_;
2621 vn_reference_t vr = data->vr;
2622 gimple *def_stmt = SSA_NAME_DEF_STMT (vuse);
2623 tree base = ao_ref_base (ref);
2624 HOST_WIDE_INT offseti = 0, maxsizei, sizei = 0;
2625 static vec<vn_reference_op_s> lhs_ops;
2626 ao_ref lhs_ref;
2627 bool lhs_ref_ok = false;
2628 poly_int64 copy_size;
2630 /* First try to disambiguate after value-replacing in the definitions LHS. */
2631 if (is_gimple_assign (def_stmt))
2633 tree lhs = gimple_assign_lhs (def_stmt);
2634 bool valueized_anything = false;
2635 /* Avoid re-allocation overhead. */
2636 lhs_ops.truncate (0);
2637 basic_block saved_rpo_bb = vn_context_bb;
2638 vn_context_bb = gimple_bb (def_stmt);
2639 if (*disambiguate_only <= TR_VALUEIZE_AND_DISAMBIGUATE)
2641 copy_reference_ops_from_ref (lhs, &lhs_ops);
2642 valueize_refs_1 (&lhs_ops, &valueized_anything, true);
2644 vn_context_bb = saved_rpo_bb;
2645 ao_ref_init (&lhs_ref, lhs);
2646 lhs_ref_ok = true;
2647 if (valueized_anything
2648 && ao_ref_init_from_vn_reference
2649 (&lhs_ref, ao_ref_alias_set (&lhs_ref),
2650 ao_ref_base_alias_set (&lhs_ref), TREE_TYPE (lhs), lhs_ops)
2651 && !refs_may_alias_p_1 (ref, &lhs_ref, data->tbaa_p))
2653 *disambiguate_only = TR_VALUEIZE_AND_DISAMBIGUATE;
2654 return NULL;
2657 /* When the def is a CLOBBER we can optimistically disambiguate
2658 against it since any overlap it would be undefined behavior.
2659 Avoid this for obvious must aliases to save compile-time though.
2660 We also may not do this when the query is used for redundant
2661 store removal. */
2662 if (!data->redundant_store_removal_p
2663 && gimple_clobber_p (def_stmt)
2664 && !operand_equal_p (ao_ref_base (&lhs_ref), base, OEP_ADDRESS_OF))
2666 *disambiguate_only = TR_DISAMBIGUATE;
2667 return NULL;
2670 /* Besides valueizing the LHS we can also use access-path based
2671 disambiguation on the original non-valueized ref. */
2672 if (!ref->ref
2673 && lhs_ref_ok
2674 && data->orig_ref.ref)
2676 /* We want to use the non-valueized LHS for this, but avoid redundant
2677 work. */
2678 ao_ref *lref = &lhs_ref;
2679 ao_ref lref_alt;
2680 if (valueized_anything)
2682 ao_ref_init (&lref_alt, lhs);
2683 lref = &lref_alt;
2685 if (!refs_may_alias_p_1 (&data->orig_ref, lref, data->tbaa_p))
2687 *disambiguate_only = (valueized_anything
2688 ? TR_VALUEIZE_AND_DISAMBIGUATE
2689 : TR_DISAMBIGUATE);
2690 return NULL;
2694 /* If we reach a clobbering statement try to skip it and see if
2695 we find a VN result with exactly the same value as the
2696 possible clobber. In this case we can ignore the clobber
2697 and return the found value. */
2698 if (is_gimple_reg_type (TREE_TYPE (lhs))
2699 && types_compatible_p (TREE_TYPE (lhs), vr->type)
2700 && (ref->ref || data->orig_ref.ref)
2701 && !data->mask
2702 && data->partial_defs.is_empty ()
2703 && multiple_p (get_object_alignment
2704 (ref->ref ? ref->ref : data->orig_ref.ref),
2705 ref->size)
2706 && multiple_p (get_object_alignment (lhs), ref->size))
2708 tree rhs = gimple_assign_rhs1 (def_stmt);
2709 /* ??? We may not compare to ahead values which might be from
2710 a different loop iteration but only to loop invariants. Use
2711 CONSTANT_CLASS_P (unvalueized!) as conservative approximation.
2712 The one-hop lookup below doesn't have this issue since there's
2713 a virtual PHI before we ever reach a backedge to cross.
2714 We can skip multiple defs as long as they are from the same
2715 value though. */
2716 if (data->same_val
2717 && !operand_equal_p (data->same_val, rhs))
2719 else if (CONSTANT_CLASS_P (rhs))
2721 if (dump_file && (dump_flags & TDF_DETAILS))
2723 fprintf (dump_file,
2724 "Skipping possible redundant definition ");
2725 print_gimple_stmt (dump_file, def_stmt, 0);
2727 /* Delay the actual compare of the values to the end of the walk
2728 but do not update last_vuse from here. */
2729 data->last_vuse_ptr = NULL;
2730 data->same_val = rhs;
2731 return NULL;
2733 else
2735 tree *saved_last_vuse_ptr = data->last_vuse_ptr;
2736 /* Do not update last_vuse_ptr in vn_reference_lookup_2. */
2737 data->last_vuse_ptr = NULL;
2738 tree saved_vuse = vr->vuse;
2739 hashval_t saved_hashcode = vr->hashcode;
2740 void *res = vn_reference_lookup_2 (ref, gimple_vuse (def_stmt),
2741 data);
2742 /* Need to restore vr->vuse and vr->hashcode. */
2743 vr->vuse = saved_vuse;
2744 vr->hashcode = saved_hashcode;
2745 data->last_vuse_ptr = saved_last_vuse_ptr;
2746 if (res && res != (void *)-1)
2748 vn_reference_t vnresult = (vn_reference_t) res;
2749 if (TREE_CODE (rhs) == SSA_NAME)
2750 rhs = SSA_VAL (rhs);
2751 if (vnresult->result
2752 && operand_equal_p (vnresult->result, rhs, 0))
2753 return res;
2758 else if (*disambiguate_only <= TR_VALUEIZE_AND_DISAMBIGUATE
2759 && gimple_call_builtin_p (def_stmt, BUILT_IN_NORMAL)
2760 && gimple_call_num_args (def_stmt) <= 4)
2762 /* For builtin calls valueize its arguments and call the
2763 alias oracle again. Valueization may improve points-to
2764 info of pointers and constify size and position arguments.
2765 Originally this was motivated by PR61034 which has
2766 conditional calls to free falsely clobbering ref because
2767 of imprecise points-to info of the argument. */
2768 tree oldargs[4];
2769 bool valueized_anything = false;
2770 for (unsigned i = 0; i < gimple_call_num_args (def_stmt); ++i)
2772 oldargs[i] = gimple_call_arg (def_stmt, i);
2773 tree val = vn_valueize (oldargs[i]);
2774 if (val != oldargs[i])
2776 gimple_call_set_arg (def_stmt, i, val);
2777 valueized_anything = true;
2780 if (valueized_anything)
2782 bool res = call_may_clobber_ref_p_1 (as_a <gcall *> (def_stmt),
2783 ref, data->tbaa_p);
2784 for (unsigned i = 0; i < gimple_call_num_args (def_stmt); ++i)
2785 gimple_call_set_arg (def_stmt, i, oldargs[i]);
2786 if (!res)
2788 *disambiguate_only = TR_VALUEIZE_AND_DISAMBIGUATE;
2789 return NULL;
2794 if (*disambiguate_only > TR_TRANSLATE)
2795 return (void *)-1;
2797 /* If we cannot constrain the size of the reference we cannot
2798 test if anything kills it. */
2799 if (!ref->max_size_known_p ())
2800 return (void *)-1;
2802 poly_int64 offset = ref->offset;
2803 poly_int64 maxsize = ref->max_size;
2805 /* def_stmt may-defs *ref. See if we can derive a value for *ref
2806 from that definition.
2807 1) Memset. */
2808 if (is_gimple_reg_type (vr->type)
2809 && (gimple_call_builtin_p (def_stmt, BUILT_IN_MEMSET)
2810 || gimple_call_builtin_p (def_stmt, BUILT_IN_MEMSET_CHK))
2811 && (integer_zerop (gimple_call_arg (def_stmt, 1))
2812 || ((TREE_CODE (gimple_call_arg (def_stmt, 1)) == INTEGER_CST
2813 || (INTEGRAL_TYPE_P (vr->type) && known_eq (ref->size, 8)))
2814 && CHAR_BIT == 8
2815 && BITS_PER_UNIT == 8
2816 && BYTES_BIG_ENDIAN == WORDS_BIG_ENDIAN
2817 && offset.is_constant (&offseti)
2818 && ref->size.is_constant (&sizei)
2819 && (offseti % BITS_PER_UNIT == 0
2820 || TREE_CODE (gimple_call_arg (def_stmt, 1)) == INTEGER_CST)))
2821 && (poly_int_tree_p (gimple_call_arg (def_stmt, 2))
2822 || (TREE_CODE (gimple_call_arg (def_stmt, 2)) == SSA_NAME
2823 && poly_int_tree_p (SSA_VAL (gimple_call_arg (def_stmt, 2)))))
2824 && (TREE_CODE (gimple_call_arg (def_stmt, 0)) == ADDR_EXPR
2825 || TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME))
2827 tree base2;
2828 poly_int64 offset2, size2, maxsize2;
2829 bool reverse;
2830 tree ref2 = gimple_call_arg (def_stmt, 0);
2831 if (TREE_CODE (ref2) == SSA_NAME)
2833 ref2 = SSA_VAL (ref2);
2834 if (TREE_CODE (ref2) == SSA_NAME
2835 && (TREE_CODE (base) != MEM_REF
2836 || TREE_OPERAND (base, 0) != ref2))
2838 gimple *def_stmt = SSA_NAME_DEF_STMT (ref2);
2839 if (gimple_assign_single_p (def_stmt)
2840 && gimple_assign_rhs_code (def_stmt) == ADDR_EXPR)
2841 ref2 = gimple_assign_rhs1 (def_stmt);
2844 if (TREE_CODE (ref2) == ADDR_EXPR)
2846 ref2 = TREE_OPERAND (ref2, 0);
2847 base2 = get_ref_base_and_extent (ref2, &offset2, &size2, &maxsize2,
2848 &reverse);
2849 if (!known_size_p (maxsize2)
2850 || !known_eq (maxsize2, size2)
2851 || !operand_equal_p (base, base2, OEP_ADDRESS_OF))
2852 return (void *)-1;
2854 else if (TREE_CODE (ref2) == SSA_NAME)
2856 poly_int64 soff;
2857 if (TREE_CODE (base) != MEM_REF
2858 || !(mem_ref_offset (base)
2859 << LOG2_BITS_PER_UNIT).to_shwi (&soff))
2860 return (void *)-1;
2861 offset += soff;
2862 offset2 = 0;
2863 if (TREE_OPERAND (base, 0) != ref2)
2865 gimple *def = SSA_NAME_DEF_STMT (ref2);
2866 if (is_gimple_assign (def)
2867 && gimple_assign_rhs_code (def) == POINTER_PLUS_EXPR
2868 && gimple_assign_rhs1 (def) == TREE_OPERAND (base, 0)
2869 && poly_int_tree_p (gimple_assign_rhs2 (def)))
2871 tree rhs2 = gimple_assign_rhs2 (def);
2872 if (!(poly_offset_int::from (wi::to_poly_wide (rhs2),
2873 SIGNED)
2874 << LOG2_BITS_PER_UNIT).to_shwi (&offset2))
2875 return (void *)-1;
2876 ref2 = gimple_assign_rhs1 (def);
2877 if (TREE_CODE (ref2) == SSA_NAME)
2878 ref2 = SSA_VAL (ref2);
2880 else
2881 return (void *)-1;
2884 else
2885 return (void *)-1;
2886 tree len = gimple_call_arg (def_stmt, 2);
2887 HOST_WIDE_INT leni, offset2i;
2888 if (TREE_CODE (len) == SSA_NAME)
2889 len = SSA_VAL (len);
2890 /* Sometimes the above trickery is smarter than alias analysis. Take
2891 advantage of that. */
2892 if (!ranges_maybe_overlap_p (offset, maxsize, offset2,
2893 (wi::to_poly_offset (len)
2894 << LOG2_BITS_PER_UNIT)))
2895 return NULL;
2896 if (data->partial_defs.is_empty ()
2897 && known_subrange_p (offset, maxsize, offset2,
2898 wi::to_poly_offset (len) << LOG2_BITS_PER_UNIT))
2900 tree val;
2901 if (integer_zerop (gimple_call_arg (def_stmt, 1)))
2902 val = build_zero_cst (vr->type);
2903 else if (INTEGRAL_TYPE_P (vr->type)
2904 && known_eq (ref->size, 8)
2905 && offseti % BITS_PER_UNIT == 0)
2907 gimple_match_op res_op (gimple_match_cond::UNCOND, NOP_EXPR,
2908 vr->type, gimple_call_arg (def_stmt, 1));
2909 val = vn_nary_build_or_lookup (&res_op);
2910 if (!val
2911 || (TREE_CODE (val) == SSA_NAME
2912 && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (val)))
2913 return (void *)-1;
2915 else
2917 unsigned buflen = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (vr->type)) + 1;
2918 if (INTEGRAL_TYPE_P (vr->type))
2919 buflen = GET_MODE_SIZE (SCALAR_INT_TYPE_MODE (vr->type)) + 1;
2920 unsigned char *buf = XALLOCAVEC (unsigned char, buflen);
2921 memset (buf, TREE_INT_CST_LOW (gimple_call_arg (def_stmt, 1)),
2922 buflen);
2923 if (BYTES_BIG_ENDIAN)
2925 unsigned int amnt
2926 = (((unsigned HOST_WIDE_INT) offseti + sizei)
2927 % BITS_PER_UNIT);
2928 if (amnt)
2930 shift_bytes_in_array_right (buf, buflen,
2931 BITS_PER_UNIT - amnt);
2932 buf++;
2933 buflen--;
2936 else if (offseti % BITS_PER_UNIT != 0)
2938 unsigned int amnt
2939 = BITS_PER_UNIT - ((unsigned HOST_WIDE_INT) offseti
2940 % BITS_PER_UNIT);
2941 shift_bytes_in_array_left (buf, buflen, amnt);
2942 buf++;
2943 buflen--;
2945 val = native_interpret_expr (vr->type, buf, buflen);
2946 if (!val)
2947 return (void *)-1;
2949 return data->finish (0, 0, val);
2951 /* For now handle clearing memory with partial defs. */
2952 else if (known_eq (ref->size, maxsize)
2953 && integer_zerop (gimple_call_arg (def_stmt, 1))
2954 && tree_fits_poly_int64_p (len)
2955 && tree_to_poly_int64 (len).is_constant (&leni)
2956 && leni <= INTTYPE_MAXIMUM (HOST_WIDE_INT) / BITS_PER_UNIT
2957 && offset.is_constant (&offseti)
2958 && offset2.is_constant (&offset2i)
2959 && maxsize.is_constant (&maxsizei)
2960 && ranges_known_overlap_p (offseti, maxsizei, offset2i,
2961 leni << LOG2_BITS_PER_UNIT))
2963 pd_data pd;
2964 pd.rhs = build_constructor (NULL_TREE, NULL);
2965 pd.rhs_off = 0;
2966 pd.offset = offset2i;
2967 pd.size = leni << LOG2_BITS_PER_UNIT;
2968 return data->push_partial_def (pd, 0, 0, offseti, maxsizei);
2972 /* 2) Assignment from an empty CONSTRUCTOR. */
2973 else if (is_gimple_reg_type (vr->type)
2974 && gimple_assign_single_p (def_stmt)
2975 && gimple_assign_rhs_code (def_stmt) == CONSTRUCTOR
2976 && CONSTRUCTOR_NELTS (gimple_assign_rhs1 (def_stmt)) == 0)
2978 tree base2;
2979 poly_int64 offset2, size2, maxsize2;
2980 HOST_WIDE_INT offset2i, size2i;
2981 gcc_assert (lhs_ref_ok);
2982 base2 = ao_ref_base (&lhs_ref);
2983 offset2 = lhs_ref.offset;
2984 size2 = lhs_ref.size;
2985 maxsize2 = lhs_ref.max_size;
2986 if (known_size_p (maxsize2)
2987 && known_eq (maxsize2, size2)
2988 && adjust_offsets_for_equal_base_address (base, &offset,
2989 base2, &offset2))
2991 if (data->partial_defs.is_empty ()
2992 && known_subrange_p (offset, maxsize, offset2, size2))
2994 /* While technically undefined behavior do not optimize
2995 a full read from a clobber. */
2996 if (gimple_clobber_p (def_stmt))
2997 return (void *)-1;
2998 tree val = build_zero_cst (vr->type);
2999 return data->finish (ao_ref_alias_set (&lhs_ref),
3000 ao_ref_base_alias_set (&lhs_ref), val);
3002 else if (known_eq (ref->size, maxsize)
3003 && maxsize.is_constant (&maxsizei)
3004 && offset.is_constant (&offseti)
3005 && offset2.is_constant (&offset2i)
3006 && size2.is_constant (&size2i)
3007 && ranges_known_overlap_p (offseti, maxsizei,
3008 offset2i, size2i))
3010 /* Let clobbers be consumed by the partial-def tracker
3011 which can choose to ignore them if they are shadowed
3012 by a later def. */
3013 pd_data pd;
3014 pd.rhs = gimple_assign_rhs1 (def_stmt);
3015 pd.rhs_off = 0;
3016 pd.offset = offset2i;
3017 pd.size = size2i;
3018 return data->push_partial_def (pd, ao_ref_alias_set (&lhs_ref),
3019 ao_ref_base_alias_set (&lhs_ref),
3020 offseti, maxsizei);
3025 /* 3) Assignment from a constant. We can use folds native encode/interpret
3026 routines to extract the assigned bits. */
3027 else if (known_eq (ref->size, maxsize)
3028 && is_gimple_reg_type (vr->type)
3029 && !reverse_storage_order_for_component_p (vr->operands)
3030 && !contains_storage_order_barrier_p (vr->operands)
3031 && gimple_assign_single_p (def_stmt)
3032 && CHAR_BIT == 8
3033 && BITS_PER_UNIT == 8
3034 && BYTES_BIG_ENDIAN == WORDS_BIG_ENDIAN
3035 /* native_encode and native_decode operate on arrays of bytes
3036 and so fundamentally need a compile-time size and offset. */
3037 && maxsize.is_constant (&maxsizei)
3038 && offset.is_constant (&offseti)
3039 && (is_gimple_min_invariant (gimple_assign_rhs1 (def_stmt))
3040 || (TREE_CODE (gimple_assign_rhs1 (def_stmt)) == SSA_NAME
3041 && is_gimple_min_invariant (SSA_VAL (gimple_assign_rhs1 (def_stmt))))))
3043 tree lhs = gimple_assign_lhs (def_stmt);
3044 tree base2;
3045 poly_int64 offset2, size2, maxsize2;
3046 HOST_WIDE_INT offset2i, size2i;
3047 bool reverse;
3048 gcc_assert (lhs_ref_ok);
3049 base2 = ao_ref_base (&lhs_ref);
3050 offset2 = lhs_ref.offset;
3051 size2 = lhs_ref.size;
3052 maxsize2 = lhs_ref.max_size;
3053 reverse = reverse_storage_order_for_component_p (lhs);
3054 if (base2
3055 && !reverse
3056 && !storage_order_barrier_p (lhs)
3057 && known_eq (maxsize2, size2)
3058 && adjust_offsets_for_equal_base_address (base, &offset,
3059 base2, &offset2)
3060 && offset.is_constant (&offseti)
3061 && offset2.is_constant (&offset2i)
3062 && size2.is_constant (&size2i))
3064 if (data->partial_defs.is_empty ()
3065 && known_subrange_p (offseti, maxsizei, offset2, size2))
3067 /* We support up to 512-bit values (for V8DFmode). */
3068 unsigned char buffer[65];
3069 int len;
3071 tree rhs = gimple_assign_rhs1 (def_stmt);
3072 if (TREE_CODE (rhs) == SSA_NAME)
3073 rhs = SSA_VAL (rhs);
3074 len = native_encode_expr (rhs,
3075 buffer, sizeof (buffer) - 1,
3076 (offseti - offset2i) / BITS_PER_UNIT);
3077 if (len > 0 && len * BITS_PER_UNIT >= maxsizei)
3079 tree type = vr->type;
3080 unsigned char *buf = buffer;
3081 unsigned int amnt = 0;
3082 /* Make sure to interpret in a type that has a range
3083 covering the whole access size. */
3084 if (INTEGRAL_TYPE_P (vr->type)
3085 && maxsizei != TYPE_PRECISION (vr->type))
3086 type = build_nonstandard_integer_type (maxsizei,
3087 TYPE_UNSIGNED (type));
3088 if (BYTES_BIG_ENDIAN)
3090 /* For big-endian native_encode_expr stored the rhs
3091 such that the LSB of it is the LSB of buffer[len - 1].
3092 That bit is stored into memory at position
3093 offset2 + size2 - 1, i.e. in byte
3094 base + (offset2 + size2 - 1) / BITS_PER_UNIT.
3095 E.g. for offset2 1 and size2 14, rhs -1 and memory
3096 previously cleared that is:
3098 01111111|11111110
3099 Now, if we want to extract offset 2 and size 12 from
3100 it using native_interpret_expr (which actually works
3101 for integral bitfield types in terms of byte size of
3102 the mode), the native_encode_expr stored the value
3103 into buffer as
3104 XX111111|11111111
3105 and returned len 2 (the X bits are outside of
3106 precision).
3107 Let sz be maxsize / BITS_PER_UNIT if not extracting
3108 a bitfield, and GET_MODE_SIZE otherwise.
3109 We need to align the LSB of the value we want to
3110 extract as the LSB of buf[sz - 1].
3111 The LSB from memory we need to read is at position
3112 offset + maxsize - 1. */
3113 HOST_WIDE_INT sz = maxsizei / BITS_PER_UNIT;
3114 if (INTEGRAL_TYPE_P (type))
3115 sz = GET_MODE_SIZE (SCALAR_INT_TYPE_MODE (type));
3116 amnt = ((unsigned HOST_WIDE_INT) offset2i + size2i
3117 - offseti - maxsizei) % BITS_PER_UNIT;
3118 if (amnt)
3119 shift_bytes_in_array_right (buffer, len, amnt);
3120 amnt = ((unsigned HOST_WIDE_INT) offset2i + size2i
3121 - offseti - maxsizei - amnt) / BITS_PER_UNIT;
3122 if ((unsigned HOST_WIDE_INT) sz + amnt > (unsigned) len)
3123 len = 0;
3124 else
3126 buf = buffer + len - sz - amnt;
3127 len -= (buf - buffer);
3130 else
3132 amnt = ((unsigned HOST_WIDE_INT) offset2i
3133 - offseti) % BITS_PER_UNIT;
3134 if (amnt)
3136 buffer[len] = 0;
3137 shift_bytes_in_array_left (buffer, len + 1, amnt);
3138 buf = buffer + 1;
3141 tree val = native_interpret_expr (type, buf, len);
3142 /* If we chop off bits because the types precision doesn't
3143 match the memory access size this is ok when optimizing
3144 reads but not when called from the DSE code during
3145 elimination. */
3146 if (val
3147 && type != vr->type)
3149 if (! int_fits_type_p (val, vr->type))
3150 val = NULL_TREE;
3151 else
3152 val = fold_convert (vr->type, val);
3155 if (val)
3156 return data->finish (ao_ref_alias_set (&lhs_ref),
3157 ao_ref_base_alias_set (&lhs_ref), val);
3160 else if (ranges_known_overlap_p (offseti, maxsizei, offset2i,
3161 size2i))
3163 pd_data pd;
3164 tree rhs = gimple_assign_rhs1 (def_stmt);
3165 if (TREE_CODE (rhs) == SSA_NAME)
3166 rhs = SSA_VAL (rhs);
3167 pd.rhs = rhs;
3168 pd.rhs_off = 0;
3169 pd.offset = offset2i;
3170 pd.size = size2i;
3171 return data->push_partial_def (pd, ao_ref_alias_set (&lhs_ref),
3172 ao_ref_base_alias_set (&lhs_ref),
3173 offseti, maxsizei);
3178 /* 4) Assignment from an SSA name which definition we may be able
3179 to access pieces from or we can combine to a larger entity. */
3180 else if (known_eq (ref->size, maxsize)
3181 && is_gimple_reg_type (vr->type)
3182 && !reverse_storage_order_for_component_p (vr->operands)
3183 && !contains_storage_order_barrier_p (vr->operands)
3184 && gimple_assign_single_p (def_stmt)
3185 && TREE_CODE (gimple_assign_rhs1 (def_stmt)) == SSA_NAME)
3187 tree lhs = gimple_assign_lhs (def_stmt);
3188 tree base2;
3189 poly_int64 offset2, size2, maxsize2;
3190 HOST_WIDE_INT offset2i, size2i, offseti;
3191 bool reverse;
3192 gcc_assert (lhs_ref_ok);
3193 base2 = ao_ref_base (&lhs_ref);
3194 offset2 = lhs_ref.offset;
3195 size2 = lhs_ref.size;
3196 maxsize2 = lhs_ref.max_size;
3197 reverse = reverse_storage_order_for_component_p (lhs);
3198 tree def_rhs = gimple_assign_rhs1 (def_stmt);
3199 if (!reverse
3200 && !storage_order_barrier_p (lhs)
3201 && known_size_p (maxsize2)
3202 && known_eq (maxsize2, size2)
3203 && adjust_offsets_for_equal_base_address (base, &offset,
3204 base2, &offset2))
3206 if (data->partial_defs.is_empty ()
3207 && known_subrange_p (offset, maxsize, offset2, size2)
3208 /* ??? We can't handle bitfield precision extracts without
3209 either using an alternate type for the BIT_FIELD_REF and
3210 then doing a conversion or possibly adjusting the offset
3211 according to endianness. */
3212 && (! INTEGRAL_TYPE_P (vr->type)
3213 || known_eq (ref->size, TYPE_PRECISION (vr->type)))
3214 && multiple_p (ref->size, BITS_PER_UNIT))
3216 tree val = NULL_TREE;
3217 if (! INTEGRAL_TYPE_P (TREE_TYPE (def_rhs))
3218 || type_has_mode_precision_p (TREE_TYPE (def_rhs)))
3220 gimple_match_op op (gimple_match_cond::UNCOND,
3221 BIT_FIELD_REF, vr->type,
3222 SSA_VAL (def_rhs),
3223 bitsize_int (ref->size),
3224 bitsize_int (offset - offset2));
3225 val = vn_nary_build_or_lookup (&op);
3227 else if (known_eq (ref->size, size2))
3229 gimple_match_op op (gimple_match_cond::UNCOND,
3230 VIEW_CONVERT_EXPR, vr->type,
3231 SSA_VAL (def_rhs));
3232 val = vn_nary_build_or_lookup (&op);
3234 if (val
3235 && (TREE_CODE (val) != SSA_NAME
3236 || ! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (val)))
3237 return data->finish (ao_ref_alias_set (&lhs_ref),
3238 ao_ref_base_alias_set (&lhs_ref), val);
3240 else if (maxsize.is_constant (&maxsizei)
3241 && offset.is_constant (&offseti)
3242 && offset2.is_constant (&offset2i)
3243 && size2.is_constant (&size2i)
3244 && ranges_known_overlap_p (offset, maxsize, offset2, size2))
3246 pd_data pd;
3247 pd.rhs = SSA_VAL (def_rhs);
3248 pd.rhs_off = 0;
3249 pd.offset = offset2i;
3250 pd.size = size2i;
3251 return data->push_partial_def (pd, ao_ref_alias_set (&lhs_ref),
3252 ao_ref_base_alias_set (&lhs_ref),
3253 offseti, maxsizei);
3258 /* 4b) Assignment done via one of the vectorizer internal store
3259 functions where we may be able to access pieces from or we can
3260 combine to a larger entity. */
3261 else if (known_eq (ref->size, maxsize)
3262 && is_gimple_reg_type (vr->type)
3263 && !reverse_storage_order_for_component_p (vr->operands)
3264 && !contains_storage_order_barrier_p (vr->operands)
3265 && is_gimple_call (def_stmt)
3266 && gimple_call_internal_p (def_stmt)
3267 && internal_store_fn_p (gimple_call_internal_fn (def_stmt)))
3269 gcall *call = as_a <gcall *> (def_stmt);
3270 internal_fn fn = gimple_call_internal_fn (call);
3272 tree mask = NULL_TREE, len = NULL_TREE, bias = NULL_TREE;
3273 switch (fn)
3275 case IFN_MASK_STORE:
3276 mask = gimple_call_arg (call, internal_fn_mask_index (fn));
3277 mask = vn_valueize (mask);
3278 if (TREE_CODE (mask) != VECTOR_CST)
3279 return (void *)-1;
3280 break;
3281 case IFN_LEN_STORE:
3282 len = gimple_call_arg (call, 2);
3283 bias = gimple_call_arg (call, 4);
3284 if (!tree_fits_uhwi_p (len) || !tree_fits_shwi_p (bias))
3285 return (void *)-1;
3286 break;
3287 default:
3288 return (void *)-1;
3290 tree def_rhs = gimple_call_arg (call,
3291 internal_fn_stored_value_index (fn));
3292 def_rhs = vn_valueize (def_rhs);
3293 if (TREE_CODE (def_rhs) != VECTOR_CST)
3294 return (void *)-1;
3296 ao_ref_init_from_ptr_and_size (&lhs_ref,
3297 vn_valueize (gimple_call_arg (call, 0)),
3298 TYPE_SIZE_UNIT (TREE_TYPE (def_rhs)));
3299 tree base2;
3300 poly_int64 offset2, size2, maxsize2;
3301 HOST_WIDE_INT offset2i, size2i, offseti;
3302 base2 = ao_ref_base (&lhs_ref);
3303 offset2 = lhs_ref.offset;
3304 size2 = lhs_ref.size;
3305 maxsize2 = lhs_ref.max_size;
3306 if (known_size_p (maxsize2)
3307 && known_eq (maxsize2, size2)
3308 && adjust_offsets_for_equal_base_address (base, &offset,
3309 base2, &offset2)
3310 && maxsize.is_constant (&maxsizei)
3311 && offset.is_constant (&offseti)
3312 && offset2.is_constant (&offset2i)
3313 && size2.is_constant (&size2i))
3315 if (!ranges_maybe_overlap_p (offset, maxsize, offset2, size2))
3316 /* Poor-mans disambiguation. */
3317 return NULL;
3318 else if (ranges_known_overlap_p (offset, maxsize, offset2, size2))
3320 pd_data pd;
3321 pd.rhs = def_rhs;
3322 tree aa = gimple_call_arg (call, 1);
3323 alias_set_type set = get_deref_alias_set (TREE_TYPE (aa));
3324 tree vectype = TREE_TYPE (def_rhs);
3325 unsigned HOST_WIDE_INT elsz
3326 = tree_to_uhwi (TYPE_SIZE (TREE_TYPE (vectype)));
3327 if (mask)
3329 HOST_WIDE_INT start = 0, len = 0;
3330 unsigned mask_idx = 0;
3333 if (integer_zerop (VECTOR_CST_ELT (mask, mask_idx)))
3335 if (len != 0)
3337 pd.rhs_off = start;
3338 pd.offset = offset2i + start;
3339 pd.size = len;
3340 if (ranges_known_overlap_p
3341 (offset, maxsize, pd.offset, pd.size))
3343 void *res = data->push_partial_def
3344 (pd, set, set, offseti, maxsizei);
3345 if (res != NULL)
3346 return res;
3349 start = (mask_idx + 1) * elsz;
3350 len = 0;
3352 else
3353 len += elsz;
3354 mask_idx++;
3356 while (known_lt (mask_idx, TYPE_VECTOR_SUBPARTS (vectype)));
3357 if (len != 0)
3359 pd.rhs_off = start;
3360 pd.offset = offset2i + start;
3361 pd.size = len;
3362 if (ranges_known_overlap_p (offset, maxsize,
3363 pd.offset, pd.size))
3364 return data->push_partial_def (pd, set, set,
3365 offseti, maxsizei);
3368 else if (fn == IFN_LEN_STORE)
3370 pd.offset = offset2i;
3371 pd.size = (tree_to_uhwi (len)
3372 + -tree_to_shwi (bias)) * BITS_PER_UNIT;
3373 if (BYTES_BIG_ENDIAN)
3374 pd.rhs_off = pd.size - tree_to_uhwi (TYPE_SIZE (vectype));
3375 else
3376 pd.rhs_off = 0;
3377 if (ranges_known_overlap_p (offset, maxsize,
3378 pd.offset, pd.size))
3379 return data->push_partial_def (pd, set, set,
3380 offseti, maxsizei);
3382 else
3383 gcc_unreachable ();
3384 return NULL;
3389 /* 5) For aggregate copies translate the reference through them if
3390 the copy kills ref. */
3391 else if (data->vn_walk_kind == VN_WALKREWRITE
3392 && gimple_assign_single_p (def_stmt)
3393 && (DECL_P (gimple_assign_rhs1 (def_stmt))
3394 || TREE_CODE (gimple_assign_rhs1 (def_stmt)) == MEM_REF
3395 || handled_component_p (gimple_assign_rhs1 (def_stmt))))
3397 tree base2;
3398 int i, j, k;
3399 auto_vec<vn_reference_op_s> rhs;
3400 vn_reference_op_t vro;
3401 ao_ref r;
3403 gcc_assert (lhs_ref_ok);
3405 /* See if the assignment kills REF. */
3406 base2 = ao_ref_base (&lhs_ref);
3407 if (!lhs_ref.max_size_known_p ()
3408 || (base != base2
3409 && (TREE_CODE (base) != MEM_REF
3410 || TREE_CODE (base2) != MEM_REF
3411 || TREE_OPERAND (base, 0) != TREE_OPERAND (base2, 0)
3412 || !tree_int_cst_equal (TREE_OPERAND (base, 1),
3413 TREE_OPERAND (base2, 1))))
3414 || !stmt_kills_ref_p (def_stmt, ref))
3415 return (void *)-1;
3417 /* Find the common base of ref and the lhs. lhs_ops already
3418 contains valueized operands for the lhs. */
3419 i = vr->operands.length () - 1;
3420 j = lhs_ops.length () - 1;
3421 while (j >= 0 && i >= 0
3422 && vn_reference_op_eq (&vr->operands[i], &lhs_ops[j]))
3424 i--;
3425 j--;
3428 /* ??? The innermost op should always be a MEM_REF and we already
3429 checked that the assignment to the lhs kills vr. Thus for
3430 aggregate copies using char[] types the vn_reference_op_eq
3431 may fail when comparing types for compatibility. But we really
3432 don't care here - further lookups with the rewritten operands
3433 will simply fail if we messed up types too badly. */
3434 poly_int64 extra_off = 0;
3435 if (j == 0 && i >= 0
3436 && lhs_ops[0].opcode == MEM_REF
3437 && maybe_ne (lhs_ops[0].off, -1))
3439 if (known_eq (lhs_ops[0].off, vr->operands[i].off))
3440 i--, j--;
3441 else if (vr->operands[i].opcode == MEM_REF
3442 && maybe_ne (vr->operands[i].off, -1))
3444 extra_off = vr->operands[i].off - lhs_ops[0].off;
3445 i--, j--;
3449 /* i now points to the first additional op.
3450 ??? LHS may not be completely contained in VR, one or more
3451 VIEW_CONVERT_EXPRs could be in its way. We could at least
3452 try handling outermost VIEW_CONVERT_EXPRs. */
3453 if (j != -1)
3454 return (void *)-1;
3456 /* Punt if the additional ops contain a storage order barrier. */
3457 for (k = i; k >= 0; k--)
3459 vro = &vr->operands[k];
3460 if (vro->opcode == VIEW_CONVERT_EXPR && vro->reverse)
3461 return (void *)-1;
3464 /* Now re-write REF to be based on the rhs of the assignment. */
3465 tree rhs1 = gimple_assign_rhs1 (def_stmt);
3466 copy_reference_ops_from_ref (rhs1, &rhs);
3468 /* Apply an extra offset to the inner MEM_REF of the RHS. */
3469 bool force_no_tbaa = false;
3470 if (maybe_ne (extra_off, 0))
3472 if (rhs.length () < 2)
3473 return (void *)-1;
3474 int ix = rhs.length () - 2;
3475 if (rhs[ix].opcode != MEM_REF
3476 || known_eq (rhs[ix].off, -1))
3477 return (void *)-1;
3478 rhs[ix].off += extra_off;
3479 rhs[ix].op0 = int_const_binop (PLUS_EXPR, rhs[ix].op0,
3480 build_int_cst (TREE_TYPE (rhs[ix].op0),
3481 extra_off));
3482 /* When we have offsetted the RHS, reading only parts of it,
3483 we can no longer use the original TBAA type, force alias-set
3484 zero. */
3485 force_no_tbaa = true;
3488 /* Save the operands since we need to use the original ones for
3489 the hash entry we use. */
3490 if (!data->saved_operands.exists ())
3491 data->saved_operands = vr->operands.copy ();
3493 /* We need to pre-pend vr->operands[0..i] to rhs. */
3494 vec<vn_reference_op_s> old = vr->operands;
3495 if (i + 1 + rhs.length () > vr->operands.length ())
3496 vr->operands.safe_grow (i + 1 + rhs.length (), true);
3497 else
3498 vr->operands.truncate (i + 1 + rhs.length ());
3499 FOR_EACH_VEC_ELT (rhs, j, vro)
3500 vr->operands[i + 1 + j] = *vro;
3501 valueize_refs (&vr->operands);
3502 if (old == shared_lookup_references)
3503 shared_lookup_references = vr->operands;
3504 vr->hashcode = vn_reference_compute_hash (vr);
3506 /* Try folding the new reference to a constant. */
3507 tree val = fully_constant_vn_reference_p (vr);
3508 if (val)
3510 if (data->partial_defs.is_empty ())
3511 return data->finish (ao_ref_alias_set (&lhs_ref),
3512 ao_ref_base_alias_set (&lhs_ref), val);
3513 /* This is the only interesting case for partial-def handling
3514 coming from targets that like to gimplify init-ctors as
3515 aggregate copies from constant data like aarch64 for
3516 PR83518. */
3517 if (maxsize.is_constant (&maxsizei) && known_eq (ref->size, maxsize))
3519 pd_data pd;
3520 pd.rhs = val;
3521 pd.rhs_off = 0;
3522 pd.offset = 0;
3523 pd.size = maxsizei;
3524 return data->push_partial_def (pd, ao_ref_alias_set (&lhs_ref),
3525 ao_ref_base_alias_set (&lhs_ref),
3526 0, maxsizei);
3530 /* Continuing with partial defs isn't easily possible here, we
3531 have to find a full def from further lookups from here. Probably
3532 not worth the special-casing everywhere. */
3533 if (!data->partial_defs.is_empty ())
3534 return (void *)-1;
3536 /* Adjust *ref from the new operands. */
3537 ao_ref rhs1_ref;
3538 ao_ref_init (&rhs1_ref, rhs1);
3539 if (!ao_ref_init_from_vn_reference (&r,
3540 force_no_tbaa ? 0
3541 : ao_ref_alias_set (&rhs1_ref),
3542 force_no_tbaa ? 0
3543 : ao_ref_base_alias_set (&rhs1_ref),
3544 vr->type, vr->operands))
3545 return (void *)-1;
3546 /* This can happen with bitfields. */
3547 if (maybe_ne (ref->size, r.size))
3549 /* If the access lacks some subsetting simply apply that by
3550 shortening it. That in the end can only be successful
3551 if we can pun the lookup result which in turn requires
3552 exact offsets. */
3553 if (known_eq (r.size, r.max_size)
3554 && known_lt (ref->size, r.size))
3555 r.size = r.max_size = ref->size;
3556 else
3557 return (void *)-1;
3559 *ref = r;
3561 /* Do not update last seen VUSE after translating. */
3562 data->last_vuse_ptr = NULL;
3563 /* Invalidate the original access path since it now contains
3564 the wrong base. */
3565 data->orig_ref.ref = NULL_TREE;
3566 /* Use the alias-set of this LHS for recording an eventual result. */
3567 if (data->first_set == -2)
3569 data->first_set = ao_ref_alias_set (&lhs_ref);
3570 data->first_base_set = ao_ref_base_alias_set (&lhs_ref);
3573 /* Keep looking for the adjusted *REF / VR pair. */
3574 return NULL;
3577 /* 6) For memcpy copies translate the reference through them if the copy
3578 kills ref. But we cannot (easily) do this translation if the memcpy is
3579 a storage order barrier, i.e. is equivalent to a VIEW_CONVERT_EXPR that
3580 can modify the storage order of objects (see storage_order_barrier_p). */
3581 else if (data->vn_walk_kind == VN_WALKREWRITE
3582 && is_gimple_reg_type (vr->type)
3583 /* ??? Handle BCOPY as well. */
3584 && (gimple_call_builtin_p (def_stmt, BUILT_IN_MEMCPY)
3585 || gimple_call_builtin_p (def_stmt, BUILT_IN_MEMCPY_CHK)
3586 || gimple_call_builtin_p (def_stmt, BUILT_IN_MEMPCPY)
3587 || gimple_call_builtin_p (def_stmt, BUILT_IN_MEMPCPY_CHK)
3588 || gimple_call_builtin_p (def_stmt, BUILT_IN_MEMMOVE)
3589 || gimple_call_builtin_p (def_stmt, BUILT_IN_MEMMOVE_CHK))
3590 && (TREE_CODE (gimple_call_arg (def_stmt, 0)) == ADDR_EXPR
3591 || TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME)
3592 && (TREE_CODE (gimple_call_arg (def_stmt, 1)) == ADDR_EXPR
3593 || TREE_CODE (gimple_call_arg (def_stmt, 1)) == SSA_NAME)
3594 && (poly_int_tree_p (gimple_call_arg (def_stmt, 2), &copy_size)
3595 || (TREE_CODE (gimple_call_arg (def_stmt, 2)) == SSA_NAME
3596 && poly_int_tree_p (SSA_VAL (gimple_call_arg (def_stmt, 2)),
3597 &copy_size)))
3598 /* Handling this is more complicated, give up for now. */
3599 && data->partial_defs.is_empty ())
3601 tree lhs, rhs;
3602 ao_ref r;
3603 poly_int64 rhs_offset, lhs_offset;
3604 vn_reference_op_s op;
3605 poly_uint64 mem_offset;
3606 poly_int64 at, byte_maxsize;
3608 /* Only handle non-variable, addressable refs. */
3609 if (maybe_ne (ref->size, maxsize)
3610 || !multiple_p (offset, BITS_PER_UNIT, &at)
3611 || !multiple_p (maxsize, BITS_PER_UNIT, &byte_maxsize))
3612 return (void *)-1;
3614 /* Extract a pointer base and an offset for the destination. */
3615 lhs = gimple_call_arg (def_stmt, 0);
3616 lhs_offset = 0;
3617 if (TREE_CODE (lhs) == SSA_NAME)
3619 lhs = vn_valueize (lhs);
3620 if (TREE_CODE (lhs) == SSA_NAME)
3622 gimple *def_stmt = SSA_NAME_DEF_STMT (lhs);
3623 if (gimple_assign_single_p (def_stmt)
3624 && gimple_assign_rhs_code (def_stmt) == ADDR_EXPR)
3625 lhs = gimple_assign_rhs1 (def_stmt);
3628 if (TREE_CODE (lhs) == ADDR_EXPR)
3630 if (AGGREGATE_TYPE_P (TREE_TYPE (TREE_TYPE (lhs)))
3631 && TYPE_REVERSE_STORAGE_ORDER (TREE_TYPE (TREE_TYPE (lhs))))
3632 return (void *)-1;
3633 tree tem = get_addr_base_and_unit_offset (TREE_OPERAND (lhs, 0),
3634 &lhs_offset);
3635 if (!tem)
3636 return (void *)-1;
3637 if (TREE_CODE (tem) == MEM_REF
3638 && poly_int_tree_p (TREE_OPERAND (tem, 1), &mem_offset))
3640 lhs = TREE_OPERAND (tem, 0);
3641 if (TREE_CODE (lhs) == SSA_NAME)
3642 lhs = vn_valueize (lhs);
3643 lhs_offset += mem_offset;
3645 else if (DECL_P (tem))
3646 lhs = build_fold_addr_expr (tem);
3647 else
3648 return (void *)-1;
3650 if (TREE_CODE (lhs) != SSA_NAME
3651 && TREE_CODE (lhs) != ADDR_EXPR)
3652 return (void *)-1;
3654 /* Extract a pointer base and an offset for the source. */
3655 rhs = gimple_call_arg (def_stmt, 1);
3656 rhs_offset = 0;
3657 if (TREE_CODE (rhs) == SSA_NAME)
3658 rhs = vn_valueize (rhs);
3659 if (TREE_CODE (rhs) == ADDR_EXPR)
3661 if (AGGREGATE_TYPE_P (TREE_TYPE (TREE_TYPE (rhs)))
3662 && TYPE_REVERSE_STORAGE_ORDER (TREE_TYPE (TREE_TYPE (rhs))))
3663 return (void *)-1;
3664 tree tem = get_addr_base_and_unit_offset (TREE_OPERAND (rhs, 0),
3665 &rhs_offset);
3666 if (!tem)
3667 return (void *)-1;
3668 if (TREE_CODE (tem) == MEM_REF
3669 && poly_int_tree_p (TREE_OPERAND (tem, 1), &mem_offset))
3671 rhs = TREE_OPERAND (tem, 0);
3672 rhs_offset += mem_offset;
3674 else if (DECL_P (tem)
3675 || TREE_CODE (tem) == STRING_CST)
3676 rhs = build_fold_addr_expr (tem);
3677 else
3678 return (void *)-1;
3680 if (TREE_CODE (rhs) == SSA_NAME)
3681 rhs = SSA_VAL (rhs);
3682 else if (TREE_CODE (rhs) != ADDR_EXPR)
3683 return (void *)-1;
3685 /* The bases of the destination and the references have to agree. */
3686 if (TREE_CODE (base) == MEM_REF)
3688 if (TREE_OPERAND (base, 0) != lhs
3689 || !poly_int_tree_p (TREE_OPERAND (base, 1), &mem_offset))
3690 return (void *) -1;
3691 at += mem_offset;
3693 else if (!DECL_P (base)
3694 || TREE_CODE (lhs) != ADDR_EXPR
3695 || TREE_OPERAND (lhs, 0) != base)
3696 return (void *)-1;
3698 /* If the access is completely outside of the memcpy destination
3699 area there is no aliasing. */
3700 if (!ranges_maybe_overlap_p (lhs_offset, copy_size, at, byte_maxsize))
3701 return NULL;
3702 /* And the access has to be contained within the memcpy destination. */
3703 if (!known_subrange_p (at, byte_maxsize, lhs_offset, copy_size))
3704 return (void *)-1;
3706 /* Save the operands since we need to use the original ones for
3707 the hash entry we use. */
3708 if (!data->saved_operands.exists ())
3709 data->saved_operands = vr->operands.copy ();
3711 /* Make room for 2 operands in the new reference. */
3712 if (vr->operands.length () < 2)
3714 vec<vn_reference_op_s> old = vr->operands;
3715 vr->operands.safe_grow_cleared (2, true);
3716 if (old == shared_lookup_references)
3717 shared_lookup_references = vr->operands;
3719 else
3720 vr->operands.truncate (2);
3722 /* The looked-through reference is a simple MEM_REF. */
3723 memset (&op, 0, sizeof (op));
3724 op.type = vr->type;
3725 op.opcode = MEM_REF;
3726 op.op0 = build_int_cst (ptr_type_node, at - lhs_offset + rhs_offset);
3727 op.off = at - lhs_offset + rhs_offset;
3728 vr->operands[0] = op;
3729 op.type = TREE_TYPE (rhs);
3730 op.opcode = TREE_CODE (rhs);
3731 op.op0 = rhs;
3732 op.off = -1;
3733 vr->operands[1] = op;
3734 vr->hashcode = vn_reference_compute_hash (vr);
3736 /* Try folding the new reference to a constant. */
3737 tree val = fully_constant_vn_reference_p (vr);
3738 if (val)
3739 return data->finish (0, 0, val);
3741 /* Adjust *ref from the new operands. */
3742 if (!ao_ref_init_from_vn_reference (&r, 0, 0, vr->type, vr->operands))
3743 return (void *)-1;
3744 /* This can happen with bitfields. */
3745 if (maybe_ne (ref->size, r.size))
3746 return (void *)-1;
3747 *ref = r;
3749 /* Do not update last seen VUSE after translating. */
3750 data->last_vuse_ptr = NULL;
3751 /* Invalidate the original access path since it now contains
3752 the wrong base. */
3753 data->orig_ref.ref = NULL_TREE;
3754 /* Use the alias-set of this stmt for recording an eventual result. */
3755 if (data->first_set == -2)
3757 data->first_set = 0;
3758 data->first_base_set = 0;
3761 /* Keep looking for the adjusted *REF / VR pair. */
3762 return NULL;
3765 /* Bail out and stop walking. */
3766 return (void *)-1;
3769 /* Return a reference op vector from OP that can be used for
3770 vn_reference_lookup_pieces. The caller is responsible for releasing
3771 the vector. */
3773 vec<vn_reference_op_s>
3774 vn_reference_operands_for_lookup (tree op)
3776 bool valueized;
3777 return valueize_shared_reference_ops_from_ref (op, &valueized).copy ();
3780 /* Lookup a reference operation by it's parts, in the current hash table.
3781 Returns the resulting value number if it exists in the hash table,
3782 NULL_TREE otherwise. VNRESULT will be filled in with the actual
3783 vn_reference_t stored in the hashtable if something is found. */
3785 tree
3786 vn_reference_lookup_pieces (tree vuse, alias_set_type set,
3787 alias_set_type base_set, tree type,
3788 vec<vn_reference_op_s> operands,
3789 vn_reference_t *vnresult, vn_lookup_kind kind)
3791 struct vn_reference_s vr1;
3792 vn_reference_t tmp;
3793 tree cst;
3795 if (!vnresult)
3796 vnresult = &tmp;
3797 *vnresult = NULL;
3799 vr1.vuse = vuse_ssa_val (vuse);
3800 shared_lookup_references.truncate (0);
3801 shared_lookup_references.safe_grow (operands.length (), true);
3802 memcpy (shared_lookup_references.address (),
3803 operands.address (),
3804 sizeof (vn_reference_op_s)
3805 * operands.length ());
3806 bool valueized_p;
3807 valueize_refs_1 (&shared_lookup_references, &valueized_p);
3808 vr1.operands = shared_lookup_references;
3809 vr1.type = type;
3810 vr1.set = set;
3811 vr1.base_set = base_set;
3812 vr1.hashcode = vn_reference_compute_hash (&vr1);
3813 if ((cst = fully_constant_vn_reference_p (&vr1)))
3814 return cst;
3816 vn_reference_lookup_1 (&vr1, vnresult);
3817 if (!*vnresult
3818 && kind != VN_NOWALK
3819 && vr1.vuse)
3821 ao_ref r;
3822 unsigned limit = param_sccvn_max_alias_queries_per_access;
3823 vn_walk_cb_data data (&vr1, NULL_TREE, NULL, kind, true, NULL_TREE,
3824 false);
3825 vec<vn_reference_op_s> ops_for_ref;
3826 if (!valueized_p)
3827 ops_for_ref = vr1.operands;
3828 else
3830 /* For ao_ref_from_mem we have to ensure only available SSA names
3831 end up in base and the only convenient way to make this work
3832 for PRE is to re-valueize with that in mind. */
3833 ops_for_ref.create (operands.length ());
3834 ops_for_ref.quick_grow (operands.length ());
3835 memcpy (ops_for_ref.address (),
3836 operands.address (),
3837 sizeof (vn_reference_op_s)
3838 * operands.length ());
3839 valueize_refs_1 (&ops_for_ref, &valueized_p, true);
3841 if (ao_ref_init_from_vn_reference (&r, set, base_set, type,
3842 ops_for_ref))
3843 *vnresult
3844 = ((vn_reference_t)
3845 walk_non_aliased_vuses (&r, vr1.vuse, true, vn_reference_lookup_2,
3846 vn_reference_lookup_3, vuse_valueize,
3847 limit, &data));
3848 if (ops_for_ref != shared_lookup_references)
3849 ops_for_ref.release ();
3850 gcc_checking_assert (vr1.operands == shared_lookup_references);
3851 if (*vnresult
3852 && data.same_val
3853 && (!(*vnresult)->result
3854 || !operand_equal_p ((*vnresult)->result, data.same_val)))
3856 *vnresult = NULL;
3857 return NULL_TREE;
3861 if (*vnresult)
3862 return (*vnresult)->result;
3864 return NULL_TREE;
3867 /* Lookup OP in the current hash table, and return the resulting value
3868 number if it exists in the hash table. Return NULL_TREE if it does
3869 not exist in the hash table or if the result field of the structure
3870 was NULL.. VNRESULT will be filled in with the vn_reference_t
3871 stored in the hashtable if one exists. When TBAA_P is false assume
3872 we are looking up a store and treat it as having alias-set zero.
3873 *LAST_VUSE_PTR will be updated with the VUSE the value lookup succeeded.
3874 MASK is either NULL_TREE, or can be an INTEGER_CST if the result of the
3875 load is bitwise anded with MASK and so we are only interested in a subset
3876 of the bits and can ignore if the other bits are uninitialized or
3877 not initialized with constants. When doing redundant store removal
3878 the caller has to set REDUNDANT_STORE_REMOVAL_P. */
3880 tree
3881 vn_reference_lookup (tree op, tree vuse, vn_lookup_kind kind,
3882 vn_reference_t *vnresult, bool tbaa_p,
3883 tree *last_vuse_ptr, tree mask,
3884 bool redundant_store_removal_p)
3886 vec<vn_reference_op_s> operands;
3887 struct vn_reference_s vr1;
3888 bool valueized_anything;
3890 if (vnresult)
3891 *vnresult = NULL;
3893 vr1.vuse = vuse_ssa_val (vuse);
3894 vr1.operands = operands
3895 = valueize_shared_reference_ops_from_ref (op, &valueized_anything);
3897 /* Handle &MEM[ptr + 5].b[1].c as POINTER_PLUS_EXPR. Avoid doing
3898 this before the pass folding __builtin_object_size had a chance to run. */
3899 if ((cfun->curr_properties & PROP_objsz)
3900 && operands[0].opcode == ADDR_EXPR
3901 && operands.last ().opcode == SSA_NAME)
3903 poly_int64 off = 0;
3904 vn_reference_op_t vro;
3905 unsigned i;
3906 for (i = 1; operands.iterate (i, &vro); ++i)
3908 if (vro->opcode == SSA_NAME)
3909 break;
3910 else if (known_eq (vro->off, -1))
3911 break;
3912 off += vro->off;
3914 if (i == operands.length () - 1
3915 /* Make sure we the offset we accumulated in a 64bit int
3916 fits the address computation carried out in target
3917 offset precision. */
3918 && (off.coeffs[0]
3919 == sext_hwi (off.coeffs[0], TYPE_PRECISION (sizetype))))
3921 gcc_assert (operands[i-1].opcode == MEM_REF);
3922 tree ops[2];
3923 ops[0] = operands[i].op0;
3924 ops[1] = wide_int_to_tree (sizetype, off);
3925 tree res = vn_nary_op_lookup_pieces (2, POINTER_PLUS_EXPR,
3926 TREE_TYPE (op), ops, NULL);
3927 if (res)
3928 return res;
3929 return NULL_TREE;
3933 vr1.type = TREE_TYPE (op);
3934 ao_ref op_ref;
3935 ao_ref_init (&op_ref, op);
3936 vr1.set = ao_ref_alias_set (&op_ref);
3937 vr1.base_set = ao_ref_base_alias_set (&op_ref);
3938 vr1.hashcode = vn_reference_compute_hash (&vr1);
3939 if (mask == NULL_TREE)
3940 if (tree cst = fully_constant_vn_reference_p (&vr1))
3941 return cst;
3943 if (kind != VN_NOWALK && vr1.vuse)
3945 vn_reference_t wvnresult;
3946 ao_ref r;
3947 unsigned limit = param_sccvn_max_alias_queries_per_access;
3948 auto_vec<vn_reference_op_s> ops_for_ref;
3949 if (valueized_anything)
3951 copy_reference_ops_from_ref (op, &ops_for_ref);
3952 bool tem;
3953 valueize_refs_1 (&ops_for_ref, &tem, true);
3955 /* Make sure to use a valueized reference if we valueized anything.
3956 Otherwise preserve the full reference for advanced TBAA. */
3957 if (!valueized_anything
3958 || !ao_ref_init_from_vn_reference (&r, vr1.set, vr1.base_set,
3959 vr1.type, ops_for_ref))
3960 ao_ref_init (&r, op);
3961 vn_walk_cb_data data (&vr1, r.ref ? NULL_TREE : op,
3962 last_vuse_ptr, kind, tbaa_p, mask,
3963 redundant_store_removal_p);
3965 wvnresult
3966 = ((vn_reference_t)
3967 walk_non_aliased_vuses (&r, vr1.vuse, tbaa_p, vn_reference_lookup_2,
3968 vn_reference_lookup_3, vuse_valueize, limit,
3969 &data));
3970 gcc_checking_assert (vr1.operands == shared_lookup_references);
3971 if (wvnresult)
3973 gcc_assert (mask == NULL_TREE);
3974 if (data.same_val
3975 && (!wvnresult->result
3976 || !operand_equal_p (wvnresult->result, data.same_val)))
3977 return NULL_TREE;
3978 if (vnresult)
3979 *vnresult = wvnresult;
3980 return wvnresult->result;
3982 else if (mask)
3983 return data.masked_result;
3985 return NULL_TREE;
3988 if (last_vuse_ptr)
3989 *last_vuse_ptr = vr1.vuse;
3990 if (mask)
3991 return NULL_TREE;
3992 return vn_reference_lookup_1 (&vr1, vnresult);
3995 /* Lookup CALL in the current hash table and return the entry in
3996 *VNRESULT if found. Populates *VR for the hashtable lookup. */
3998 void
3999 vn_reference_lookup_call (gcall *call, vn_reference_t *vnresult,
4000 vn_reference_t vr)
4002 if (vnresult)
4003 *vnresult = NULL;
4005 tree vuse = gimple_vuse (call);
4007 vr->vuse = vuse ? SSA_VAL (vuse) : NULL_TREE;
4008 vr->operands = valueize_shared_reference_ops_from_call (call);
4009 tree lhs = gimple_call_lhs (call);
4010 /* For non-SSA return values the referece ops contain the LHS. */
4011 vr->type = ((lhs && TREE_CODE (lhs) == SSA_NAME)
4012 ? TREE_TYPE (lhs) : NULL_TREE);
4013 vr->punned = false;
4014 vr->set = 0;
4015 vr->base_set = 0;
4016 vr->hashcode = vn_reference_compute_hash (vr);
4017 vn_reference_lookup_1 (vr, vnresult);
4020 /* Insert OP into the current hash table with a value number of RESULT. */
4022 static void
4023 vn_reference_insert (tree op, tree result, tree vuse, tree vdef)
4025 vn_reference_s **slot;
4026 vn_reference_t vr1;
4027 bool tem;
4029 vec<vn_reference_op_s> operands
4030 = valueize_shared_reference_ops_from_ref (op, &tem);
4031 /* Handle &MEM[ptr + 5].b[1].c as POINTER_PLUS_EXPR. Avoid doing this
4032 before the pass folding __builtin_object_size had a chance to run. */
4033 if ((cfun->curr_properties & PROP_objsz)
4034 && operands[0].opcode == ADDR_EXPR
4035 && operands.last ().opcode == SSA_NAME)
4037 poly_int64 off = 0;
4038 vn_reference_op_t vro;
4039 unsigned i;
4040 for (i = 1; operands.iterate (i, &vro); ++i)
4042 if (vro->opcode == SSA_NAME)
4043 break;
4044 else if (known_eq (vro->off, -1))
4045 break;
4046 off += vro->off;
4048 if (i == operands.length () - 1
4049 /* Make sure we the offset we accumulated in a 64bit int
4050 fits the address computation carried out in target
4051 offset precision. */
4052 && (off.coeffs[0]
4053 == sext_hwi (off.coeffs[0], TYPE_PRECISION (sizetype))))
4055 gcc_assert (operands[i-1].opcode == MEM_REF);
4056 tree ops[2];
4057 ops[0] = operands[i].op0;
4058 ops[1] = wide_int_to_tree (sizetype, off);
4059 vn_nary_op_insert_pieces (2, POINTER_PLUS_EXPR,
4060 TREE_TYPE (op), ops, result,
4061 VN_INFO (result)->value_id);
4062 return;
4066 vr1 = XOBNEW (&vn_tables_obstack, vn_reference_s);
4067 if (TREE_CODE (result) == SSA_NAME)
4068 vr1->value_id = VN_INFO (result)->value_id;
4069 else
4070 vr1->value_id = get_or_alloc_constant_value_id (result);
4071 vr1->vuse = vuse_ssa_val (vuse);
4072 vr1->operands = operands.copy ();
4073 vr1->type = TREE_TYPE (op);
4074 vr1->punned = false;
4075 ao_ref op_ref;
4076 ao_ref_init (&op_ref, op);
4077 vr1->set = ao_ref_alias_set (&op_ref);
4078 vr1->base_set = ao_ref_base_alias_set (&op_ref);
4079 vr1->hashcode = vn_reference_compute_hash (vr1);
4080 vr1->result = TREE_CODE (result) == SSA_NAME ? SSA_VAL (result) : result;
4081 vr1->result_vdef = vdef;
4083 slot = valid_info->references->find_slot_with_hash (vr1, vr1->hashcode,
4084 INSERT);
4086 /* Because IL walking on reference lookup can end up visiting
4087 a def that is only to be visited later in iteration order
4088 when we are about to make an irreducible region reducible
4089 the def can be effectively processed and its ref being inserted
4090 by vn_reference_lookup_3 already. So we cannot assert (!*slot)
4091 but save a lookup if we deal with already inserted refs here. */
4092 if (*slot)
4094 /* We cannot assert that we have the same value either because
4095 when disentangling an irreducible region we may end up visiting
4096 a use before the corresponding def. That's a missed optimization
4097 only though. See gcc.dg/tree-ssa/pr87126.c for example. */
4098 if (dump_file && (dump_flags & TDF_DETAILS)
4099 && !operand_equal_p ((*slot)->result, vr1->result, 0))
4101 fprintf (dump_file, "Keeping old value ");
4102 print_generic_expr (dump_file, (*slot)->result);
4103 fprintf (dump_file, " because of collision\n");
4105 free_reference (vr1);
4106 obstack_free (&vn_tables_obstack, vr1);
4107 return;
4110 *slot = vr1;
4111 vr1->next = last_inserted_ref;
4112 last_inserted_ref = vr1;
4115 /* Insert a reference by it's pieces into the current hash table with
4116 a value number of RESULT. Return the resulting reference
4117 structure we created. */
4119 vn_reference_t
4120 vn_reference_insert_pieces (tree vuse, alias_set_type set,
4121 alias_set_type base_set, tree type,
4122 vec<vn_reference_op_s> operands,
4123 tree result, unsigned int value_id)
4126 vn_reference_s **slot;
4127 vn_reference_t vr1;
4129 vr1 = XOBNEW (&vn_tables_obstack, vn_reference_s);
4130 vr1->value_id = value_id;
4131 vr1->vuse = vuse_ssa_val (vuse);
4132 vr1->operands = operands;
4133 valueize_refs (&vr1->operands);
4134 vr1->type = type;
4135 vr1->punned = false;
4136 vr1->set = set;
4137 vr1->base_set = base_set;
4138 vr1->hashcode = vn_reference_compute_hash (vr1);
4139 if (result && TREE_CODE (result) == SSA_NAME)
4140 result = SSA_VAL (result);
4141 vr1->result = result;
4142 vr1->result_vdef = NULL_TREE;
4144 slot = valid_info->references->find_slot_with_hash (vr1, vr1->hashcode,
4145 INSERT);
4147 /* At this point we should have all the things inserted that we have
4148 seen before, and we should never try inserting something that
4149 already exists. */
4150 gcc_assert (!*slot);
4152 *slot = vr1;
4153 vr1->next = last_inserted_ref;
4154 last_inserted_ref = vr1;
4155 return vr1;
4158 /* Compute and return the hash value for nary operation VBO1. */
4160 hashval_t
4161 vn_nary_op_compute_hash (const vn_nary_op_t vno1)
4163 inchash::hash hstate;
4164 unsigned i;
4166 if (((vno1->length == 2
4167 && commutative_tree_code (vno1->opcode))
4168 || (vno1->length == 3
4169 && commutative_ternary_tree_code (vno1->opcode)))
4170 && tree_swap_operands_p (vno1->op[0], vno1->op[1]))
4171 std::swap (vno1->op[0], vno1->op[1]);
4172 else if (TREE_CODE_CLASS (vno1->opcode) == tcc_comparison
4173 && tree_swap_operands_p (vno1->op[0], vno1->op[1]))
4175 std::swap (vno1->op[0], vno1->op[1]);
4176 vno1->opcode = swap_tree_comparison (vno1->opcode);
4179 hstate.add_int (vno1->opcode);
4180 for (i = 0; i < vno1->length; ++i)
4181 inchash::add_expr (vno1->op[i], hstate);
4183 return hstate.end ();
4186 /* Compare nary operations VNO1 and VNO2 and return true if they are
4187 equivalent. */
4189 bool
4190 vn_nary_op_eq (const_vn_nary_op_t const vno1, const_vn_nary_op_t const vno2)
4192 unsigned i;
4194 if (vno1->hashcode != vno2->hashcode)
4195 return false;
4197 if (vno1->length != vno2->length)
4198 return false;
4200 if (vno1->opcode != vno2->opcode
4201 || !types_compatible_p (vno1->type, vno2->type))
4202 return false;
4204 for (i = 0; i < vno1->length; ++i)
4205 if (!expressions_equal_p (vno1->op[i], vno2->op[i]))
4206 return false;
4208 /* BIT_INSERT_EXPR has an implict operand as the type precision
4209 of op1. Need to check to make sure they are the same. */
4210 if (vno1->opcode == BIT_INSERT_EXPR
4211 && TREE_CODE (vno1->op[1]) == INTEGER_CST
4212 && TYPE_PRECISION (TREE_TYPE (vno1->op[1]))
4213 != TYPE_PRECISION (TREE_TYPE (vno2->op[1])))
4214 return false;
4216 return true;
4219 /* Initialize VNO from the pieces provided. */
4221 static void
4222 init_vn_nary_op_from_pieces (vn_nary_op_t vno, unsigned int length,
4223 enum tree_code code, tree type, tree *ops)
4225 vno->opcode = code;
4226 vno->length = length;
4227 vno->type = type;
4228 memcpy (&vno->op[0], ops, sizeof (tree) * length);
4231 /* Return the number of operands for a vn_nary ops structure from STMT. */
4233 unsigned int
4234 vn_nary_length_from_stmt (gimple *stmt)
4236 switch (gimple_assign_rhs_code (stmt))
4238 case REALPART_EXPR:
4239 case IMAGPART_EXPR:
4240 case VIEW_CONVERT_EXPR:
4241 return 1;
4243 case BIT_FIELD_REF:
4244 return 3;
4246 case CONSTRUCTOR:
4247 return CONSTRUCTOR_NELTS (gimple_assign_rhs1 (stmt));
4249 default:
4250 return gimple_num_ops (stmt) - 1;
4254 /* Initialize VNO from STMT. */
4256 void
4257 init_vn_nary_op_from_stmt (vn_nary_op_t vno, gassign *stmt)
4259 unsigned i;
4261 vno->opcode = gimple_assign_rhs_code (stmt);
4262 vno->type = TREE_TYPE (gimple_assign_lhs (stmt));
4263 switch (vno->opcode)
4265 case REALPART_EXPR:
4266 case IMAGPART_EXPR:
4267 case VIEW_CONVERT_EXPR:
4268 vno->length = 1;
4269 vno->op[0] = TREE_OPERAND (gimple_assign_rhs1 (stmt), 0);
4270 break;
4272 case BIT_FIELD_REF:
4273 vno->length = 3;
4274 vno->op[0] = TREE_OPERAND (gimple_assign_rhs1 (stmt), 0);
4275 vno->op[1] = TREE_OPERAND (gimple_assign_rhs1 (stmt), 1);
4276 vno->op[2] = TREE_OPERAND (gimple_assign_rhs1 (stmt), 2);
4277 break;
4279 case CONSTRUCTOR:
4280 vno->length = CONSTRUCTOR_NELTS (gimple_assign_rhs1 (stmt));
4281 for (i = 0; i < vno->length; ++i)
4282 vno->op[i] = CONSTRUCTOR_ELT (gimple_assign_rhs1 (stmt), i)->value;
4283 break;
4285 default:
4286 gcc_checking_assert (!gimple_assign_single_p (stmt));
4287 vno->length = gimple_num_ops (stmt) - 1;
4288 for (i = 0; i < vno->length; ++i)
4289 vno->op[i] = gimple_op (stmt, i + 1);
4293 /* Compute the hashcode for VNO and look for it in the hash table;
4294 return the resulting value number if it exists in the hash table.
4295 Return NULL_TREE if it does not exist in the hash table or if the
4296 result field of the operation is NULL. VNRESULT will contain the
4297 vn_nary_op_t from the hashtable if it exists. */
4299 static tree
4300 vn_nary_op_lookup_1 (vn_nary_op_t vno, vn_nary_op_t *vnresult)
4302 vn_nary_op_s **slot;
4304 if (vnresult)
4305 *vnresult = NULL;
4307 for (unsigned i = 0; i < vno->length; ++i)
4308 if (TREE_CODE (vno->op[i]) == SSA_NAME)
4309 vno->op[i] = SSA_VAL (vno->op[i]);
4311 vno->hashcode = vn_nary_op_compute_hash (vno);
4312 slot = valid_info->nary->find_slot_with_hash (vno, vno->hashcode, NO_INSERT);
4313 if (!slot)
4314 return NULL_TREE;
4315 if (vnresult)
4316 *vnresult = *slot;
4317 return (*slot)->predicated_values ? NULL_TREE : (*slot)->u.result;
4320 /* Lookup a n-ary operation by its pieces and return the resulting value
4321 number if it exists in the hash table. Return NULL_TREE if it does
4322 not exist in the hash table or if the result field of the operation
4323 is NULL. VNRESULT will contain the vn_nary_op_t from the hashtable
4324 if it exists. */
4326 tree
4327 vn_nary_op_lookup_pieces (unsigned int length, enum tree_code code,
4328 tree type, tree *ops, vn_nary_op_t *vnresult)
4330 vn_nary_op_t vno1 = XALLOCAVAR (struct vn_nary_op_s,
4331 sizeof_vn_nary_op (length));
4332 init_vn_nary_op_from_pieces (vno1, length, code, type, ops);
4333 return vn_nary_op_lookup_1 (vno1, vnresult);
4336 /* Lookup the rhs of STMT in the current hash table, and return the resulting
4337 value number if it exists in the hash table. Return NULL_TREE if
4338 it does not exist in the hash table. VNRESULT will contain the
4339 vn_nary_op_t from the hashtable if it exists. */
4341 tree
4342 vn_nary_op_lookup_stmt (gimple *stmt, vn_nary_op_t *vnresult)
4344 vn_nary_op_t vno1
4345 = XALLOCAVAR (struct vn_nary_op_s,
4346 sizeof_vn_nary_op (vn_nary_length_from_stmt (stmt)));
4347 init_vn_nary_op_from_stmt (vno1, as_a <gassign *> (stmt));
4348 return vn_nary_op_lookup_1 (vno1, vnresult);
4351 /* Allocate a vn_nary_op_t with LENGTH operands on STACK. */
4353 vn_nary_op_t
4354 alloc_vn_nary_op_noinit (unsigned int length, struct obstack *stack)
4356 return (vn_nary_op_t) obstack_alloc (stack, sizeof_vn_nary_op (length));
4359 /* Allocate and initialize a vn_nary_op_t on CURRENT_INFO's
4360 obstack. */
4362 static vn_nary_op_t
4363 alloc_vn_nary_op (unsigned int length, tree result, unsigned int value_id)
4365 vn_nary_op_t vno1 = alloc_vn_nary_op_noinit (length, &vn_tables_obstack);
4367 vno1->value_id = value_id;
4368 vno1->length = length;
4369 vno1->predicated_values = 0;
4370 vno1->u.result = result;
4372 return vno1;
4375 /* Insert VNO into TABLE. */
4377 static vn_nary_op_t
4378 vn_nary_op_insert_into (vn_nary_op_t vno, vn_nary_op_table_type *table)
4380 vn_nary_op_s **slot;
4382 gcc_assert (! vno->predicated_values
4383 || (! vno->u.values->next
4384 && vno->u.values->n == 1));
4386 for (unsigned i = 0; i < vno->length; ++i)
4387 if (TREE_CODE (vno->op[i]) == SSA_NAME)
4388 vno->op[i] = SSA_VAL (vno->op[i]);
4390 vno->hashcode = vn_nary_op_compute_hash (vno);
4391 slot = table->find_slot_with_hash (vno, vno->hashcode, INSERT);
4392 vno->unwind_to = *slot;
4393 if (*slot)
4395 /* Prefer non-predicated values.
4396 ??? Only if those are constant, otherwise, with constant predicated
4397 value, turn them into predicated values with entry-block validity
4398 (??? but we always find the first valid result currently). */
4399 if ((*slot)->predicated_values
4400 && ! vno->predicated_values)
4402 /* ??? We cannot remove *slot from the unwind stack list.
4403 For the moment we deal with this by skipping not found
4404 entries but this isn't ideal ... */
4405 *slot = vno;
4406 /* ??? Maintain a stack of states we can unwind in
4407 vn_nary_op_s? But how far do we unwind? In reality
4408 we need to push change records somewhere... Or not
4409 unwind vn_nary_op_s and linking them but instead
4410 unwind the results "list", linking that, which also
4411 doesn't move on hashtable resize. */
4412 /* We can also have a ->unwind_to recording *slot there.
4413 That way we can make u.values a fixed size array with
4414 recording the number of entries but of course we then
4415 have always N copies for each unwind_to-state. Or we
4416 make sure to only ever append and each unwinding will
4417 pop off one entry (but how to deal with predicated
4418 replaced with non-predicated here?) */
4419 vno->next = last_inserted_nary;
4420 last_inserted_nary = vno;
4421 return vno;
4423 else if (vno->predicated_values
4424 && ! (*slot)->predicated_values)
4425 return *slot;
4426 else if (vno->predicated_values
4427 && (*slot)->predicated_values)
4429 /* ??? Factor this all into a insert_single_predicated_value
4430 routine. */
4431 gcc_assert (!vno->u.values->next && vno->u.values->n == 1);
4432 basic_block vno_bb
4433 = BASIC_BLOCK_FOR_FN (cfun, vno->u.values->valid_dominated_by_p[0]);
4434 vn_pval *nval = vno->u.values;
4435 vn_pval **next = &vno->u.values;
4436 bool found = false;
4437 for (vn_pval *val = (*slot)->u.values; val; val = val->next)
4439 if (expressions_equal_p (val->result, nval->result))
4441 found = true;
4442 for (unsigned i = 0; i < val->n; ++i)
4444 basic_block val_bb
4445 = BASIC_BLOCK_FOR_FN (cfun,
4446 val->valid_dominated_by_p[i]);
4447 if (dominated_by_p (CDI_DOMINATORS, vno_bb, val_bb))
4448 /* Value registered with more generic predicate. */
4449 return *slot;
4450 else if (flag_checking)
4451 /* Shouldn't happen, we insert in RPO order. */
4452 gcc_assert (!dominated_by_p (CDI_DOMINATORS,
4453 val_bb, vno_bb));
4455 /* Append value. */
4456 *next = (vn_pval *) obstack_alloc (&vn_tables_obstack,
4457 sizeof (vn_pval)
4458 + val->n * sizeof (int));
4459 (*next)->next = NULL;
4460 (*next)->result = val->result;
4461 (*next)->n = val->n + 1;
4462 memcpy ((*next)->valid_dominated_by_p,
4463 val->valid_dominated_by_p,
4464 val->n * sizeof (int));
4465 (*next)->valid_dominated_by_p[val->n] = vno_bb->index;
4466 next = &(*next)->next;
4467 if (dump_file && (dump_flags & TDF_DETAILS))
4468 fprintf (dump_file, "Appending predicate to value.\n");
4469 continue;
4471 /* Copy other predicated values. */
4472 *next = (vn_pval *) obstack_alloc (&vn_tables_obstack,
4473 sizeof (vn_pval)
4474 + (val->n-1) * sizeof (int));
4475 memcpy (*next, val, sizeof (vn_pval) + (val->n-1) * sizeof (int));
4476 (*next)->next = NULL;
4477 next = &(*next)->next;
4479 if (!found)
4480 *next = nval;
4482 *slot = vno;
4483 vno->next = last_inserted_nary;
4484 last_inserted_nary = vno;
4485 return vno;
4488 /* While we do not want to insert things twice it's awkward to
4489 avoid it in the case where visit_nary_op pattern-matches stuff
4490 and ends up simplifying the replacement to itself. We then
4491 get two inserts, one from visit_nary_op and one from
4492 vn_nary_build_or_lookup.
4493 So allow inserts with the same value number. */
4494 if ((*slot)->u.result == vno->u.result)
4495 return *slot;
4498 /* ??? There's also optimistic vs. previous commited state merging
4499 that is problematic for the case of unwinding. */
4501 /* ??? We should return NULL if we do not use 'vno' and have the
4502 caller release it. */
4503 gcc_assert (!*slot);
4505 *slot = vno;
4506 vno->next = last_inserted_nary;
4507 last_inserted_nary = vno;
4508 return vno;
4511 /* Insert a n-ary operation into the current hash table using it's
4512 pieces. Return the vn_nary_op_t structure we created and put in
4513 the hashtable. */
4515 vn_nary_op_t
4516 vn_nary_op_insert_pieces (unsigned int length, enum tree_code code,
4517 tree type, tree *ops,
4518 tree result, unsigned int value_id)
4520 vn_nary_op_t vno1 = alloc_vn_nary_op (length, result, value_id);
4521 init_vn_nary_op_from_pieces (vno1, length, code, type, ops);
4522 return vn_nary_op_insert_into (vno1, valid_info->nary);
4525 /* Return whether we can track a predicate valid when PRED_E is executed. */
4527 static bool
4528 can_track_predicate_on_edge (edge pred_e)
4530 /* ??? As we are currently recording the destination basic-block index in
4531 vn_pval.valid_dominated_by_p and using dominance for the
4532 validity check we cannot track predicates on all edges. */
4533 if (single_pred_p (pred_e->dest))
4534 return true;
4535 /* Never record for backedges. */
4536 if (pred_e->flags & EDGE_DFS_BACK)
4537 return false;
4538 /* When there's more than one predecessor we cannot track
4539 predicate validity based on the destination block. The
4540 exception is when all other incoming edges sources are
4541 dominated by the destination block. */
4542 edge_iterator ei;
4543 edge e;
4544 FOR_EACH_EDGE (e, ei, pred_e->dest->preds)
4545 if (e != pred_e && ! dominated_by_p (CDI_DOMINATORS, e->src, e->dest))
4546 return false;
4547 return true;
4550 static vn_nary_op_t
4551 vn_nary_op_insert_pieces_predicated (unsigned int length, enum tree_code code,
4552 tree type, tree *ops,
4553 tree result, unsigned int value_id,
4554 edge pred_e)
4556 gcc_assert (can_track_predicate_on_edge (pred_e));
4558 if (dump_file && (dump_flags & TDF_DETAILS)
4559 /* ??? Fix dumping, but currently we only get comparisons. */
4560 && TREE_CODE_CLASS (code) == tcc_comparison)
4562 fprintf (dump_file, "Recording on edge %d->%d ", pred_e->src->index,
4563 pred_e->dest->index);
4564 print_generic_expr (dump_file, ops[0], TDF_SLIM);
4565 fprintf (dump_file, " %s ", get_tree_code_name (code));
4566 print_generic_expr (dump_file, ops[1], TDF_SLIM);
4567 fprintf (dump_file, " == %s\n",
4568 integer_zerop (result) ? "false" : "true");
4570 vn_nary_op_t vno1 = alloc_vn_nary_op (length, NULL_TREE, value_id);
4571 init_vn_nary_op_from_pieces (vno1, length, code, type, ops);
4572 vno1->predicated_values = 1;
4573 vno1->u.values = (vn_pval *) obstack_alloc (&vn_tables_obstack,
4574 sizeof (vn_pval));
4575 vno1->u.values->next = NULL;
4576 vno1->u.values->result = result;
4577 vno1->u.values->n = 1;
4578 vno1->u.values->valid_dominated_by_p[0] = pred_e->dest->index;
4579 return vn_nary_op_insert_into (vno1, valid_info->nary);
4582 static bool
4583 dominated_by_p_w_unex (basic_block bb1, basic_block bb2, bool);
4585 static tree
4586 vn_nary_op_get_predicated_value (vn_nary_op_t vno, basic_block bb)
4588 if (! vno->predicated_values)
4589 return vno->u.result;
4590 for (vn_pval *val = vno->u.values; val; val = val->next)
4591 for (unsigned i = 0; i < val->n; ++i)
4592 /* Do not handle backedge executability optimistically since
4593 when figuring out whether to iterate we do not consider
4594 changed predication. */
4595 if (dominated_by_p_w_unex
4596 (bb, BASIC_BLOCK_FOR_FN (cfun, val->valid_dominated_by_p[i]),
4597 false))
4598 return val->result;
4599 return NULL_TREE;
4602 /* Insert the rhs of STMT into the current hash table with a value number of
4603 RESULT. */
4605 static vn_nary_op_t
4606 vn_nary_op_insert_stmt (gimple *stmt, tree result)
4608 vn_nary_op_t vno1
4609 = alloc_vn_nary_op (vn_nary_length_from_stmt (stmt),
4610 result, VN_INFO (result)->value_id);
4611 init_vn_nary_op_from_stmt (vno1, as_a <gassign *> (stmt));
4612 return vn_nary_op_insert_into (vno1, valid_info->nary);
4615 /* Compute a hashcode for PHI operation VP1 and return it. */
4617 static inline hashval_t
4618 vn_phi_compute_hash (vn_phi_t vp1)
4620 inchash::hash hstate;
4621 tree phi1op;
4622 tree type;
4623 edge e;
4624 edge_iterator ei;
4626 hstate.add_int (EDGE_COUNT (vp1->block->preds));
4627 switch (EDGE_COUNT (vp1->block->preds))
4629 case 1:
4630 break;
4631 case 2:
4632 /* When this is a PHI node subject to CSE for different blocks
4633 avoid hashing the block index. */
4634 if (vp1->cclhs)
4635 break;
4636 /* Fallthru. */
4637 default:
4638 hstate.add_int (vp1->block->index);
4641 /* If all PHI arguments are constants we need to distinguish
4642 the PHI node via its type. */
4643 type = vp1->type;
4644 hstate.merge_hash (vn_hash_type (type));
4646 FOR_EACH_EDGE (e, ei, vp1->block->preds)
4648 /* Don't hash backedge values they need to be handled as VN_TOP
4649 for optimistic value-numbering. */
4650 if (e->flags & EDGE_DFS_BACK)
4651 continue;
4653 phi1op = vp1->phiargs[e->dest_idx];
4654 if (phi1op == VN_TOP)
4655 continue;
4656 inchash::add_expr (phi1op, hstate);
4659 return hstate.end ();
4663 /* Return true if COND1 and COND2 represent the same condition, set
4664 *INVERTED_P if one needs to be inverted to make it the same as
4665 the other. */
4667 static bool
4668 cond_stmts_equal_p (gcond *cond1, tree lhs1, tree rhs1,
4669 gcond *cond2, tree lhs2, tree rhs2, bool *inverted_p)
4671 enum tree_code code1 = gimple_cond_code (cond1);
4672 enum tree_code code2 = gimple_cond_code (cond2);
4674 *inverted_p = false;
4675 if (code1 == code2)
4677 else if (code1 == swap_tree_comparison (code2))
4678 std::swap (lhs2, rhs2);
4679 else if (code1 == invert_tree_comparison (code2, HONOR_NANS (lhs2)))
4680 *inverted_p = true;
4681 else if (code1 == invert_tree_comparison
4682 (swap_tree_comparison (code2), HONOR_NANS (lhs2)))
4684 std::swap (lhs2, rhs2);
4685 *inverted_p = true;
4687 else
4688 return false;
4690 return ((expressions_equal_p (lhs1, lhs2)
4691 && expressions_equal_p (rhs1, rhs2))
4692 || (commutative_tree_code (code1)
4693 && expressions_equal_p (lhs1, rhs2)
4694 && expressions_equal_p (rhs1, lhs2)));
4697 /* Compare two phi entries for equality, ignoring VN_TOP arguments. */
4699 static int
4700 vn_phi_eq (const_vn_phi_t const vp1, const_vn_phi_t const vp2)
4702 if (vp1->hashcode != vp2->hashcode)
4703 return false;
4705 if (vp1->block != vp2->block)
4707 if (EDGE_COUNT (vp1->block->preds) != EDGE_COUNT (vp2->block->preds))
4708 return false;
4710 switch (EDGE_COUNT (vp1->block->preds))
4712 case 1:
4713 /* Single-arg PHIs are just copies. */
4714 break;
4716 case 2:
4718 /* Make sure both PHIs are classified as CSEable. */
4719 if (! vp1->cclhs || ! vp2->cclhs)
4720 return false;
4722 /* Rule out backedges into the PHI. */
4723 gcc_checking_assert
4724 (vp1->block->loop_father->header != vp1->block
4725 && vp2->block->loop_father->header != vp2->block);
4727 /* If the PHI nodes do not have compatible types
4728 they are not the same. */
4729 if (!types_compatible_p (vp1->type, vp2->type))
4730 return false;
4732 /* If the immediate dominator end in switch stmts multiple
4733 values may end up in the same PHI arg via intermediate
4734 CFG merges. */
4735 basic_block idom1
4736 = get_immediate_dominator (CDI_DOMINATORS, vp1->block);
4737 basic_block idom2
4738 = get_immediate_dominator (CDI_DOMINATORS, vp2->block);
4739 gcc_checking_assert (EDGE_COUNT (idom1->succs) == 2
4740 && EDGE_COUNT (idom2->succs) == 2);
4742 /* Verify the controlling stmt is the same. */
4743 gcond *last1 = as_a <gcond *> (last_stmt (idom1));
4744 gcond *last2 = as_a <gcond *> (last_stmt (idom2));
4745 bool inverted_p;
4746 if (! cond_stmts_equal_p (last1, vp1->cclhs, vp1->ccrhs,
4747 last2, vp2->cclhs, vp2->ccrhs,
4748 &inverted_p))
4749 return false;
4751 /* Get at true/false controlled edges into the PHI. */
4752 edge te1, te2, fe1, fe2;
4753 if (! extract_true_false_controlled_edges (idom1, vp1->block,
4754 &te1, &fe1)
4755 || ! extract_true_false_controlled_edges (idom2, vp2->block,
4756 &te2, &fe2))
4757 return false;
4759 /* Swap edges if the second condition is the inverted of the
4760 first. */
4761 if (inverted_p)
4762 std::swap (te2, fe2);
4764 /* Since we do not know which edge will be executed we have
4765 to be careful when matching VN_TOP. Be conservative and
4766 only match VN_TOP == VN_TOP for now, we could allow
4767 VN_TOP on the not prevailing PHI though. See for example
4768 PR102920. */
4769 if (! expressions_equal_p (vp1->phiargs[te1->dest_idx],
4770 vp2->phiargs[te2->dest_idx], false)
4771 || ! expressions_equal_p (vp1->phiargs[fe1->dest_idx],
4772 vp2->phiargs[fe2->dest_idx], false))
4773 return false;
4775 return true;
4778 default:
4779 return false;
4783 /* If the PHI nodes do not have compatible types
4784 they are not the same. */
4785 if (!types_compatible_p (vp1->type, vp2->type))
4786 return false;
4788 /* Any phi in the same block will have it's arguments in the
4789 same edge order, because of how we store phi nodes. */
4790 unsigned nargs = EDGE_COUNT (vp1->block->preds);
4791 for (unsigned i = 0; i < nargs; ++i)
4793 tree phi1op = vp1->phiargs[i];
4794 tree phi2op = vp2->phiargs[i];
4795 if (phi1op == phi2op)
4796 continue;
4797 if (!expressions_equal_p (phi1op, phi2op, false))
4798 return false;
4801 return true;
4804 /* Lookup PHI in the current hash table, and return the resulting
4805 value number if it exists in the hash table. Return NULL_TREE if
4806 it does not exist in the hash table. */
4808 static tree
4809 vn_phi_lookup (gimple *phi, bool backedges_varying_p)
4811 vn_phi_s **slot;
4812 struct vn_phi_s *vp1;
4813 edge e;
4814 edge_iterator ei;
4816 vp1 = XALLOCAVAR (struct vn_phi_s,
4817 sizeof (struct vn_phi_s)
4818 + (gimple_phi_num_args (phi) - 1) * sizeof (tree));
4820 /* Canonicalize the SSA_NAME's to their value number. */
4821 FOR_EACH_EDGE (e, ei, gimple_bb (phi)->preds)
4823 tree def = PHI_ARG_DEF_FROM_EDGE (phi, e);
4824 if (TREE_CODE (def) == SSA_NAME
4825 && (!backedges_varying_p || !(e->flags & EDGE_DFS_BACK)))
4827 if (!virtual_operand_p (def)
4828 && ssa_undefined_value_p (def, false))
4829 def = VN_TOP;
4830 else
4831 def = SSA_VAL (def);
4833 vp1->phiargs[e->dest_idx] = def;
4835 vp1->type = TREE_TYPE (gimple_phi_result (phi));
4836 vp1->block = gimple_bb (phi);
4837 /* Extract values of the controlling condition. */
4838 vp1->cclhs = NULL_TREE;
4839 vp1->ccrhs = NULL_TREE;
4840 if (EDGE_COUNT (vp1->block->preds) == 2
4841 && vp1->block->loop_father->header != vp1->block)
4843 basic_block idom1 = get_immediate_dominator (CDI_DOMINATORS, vp1->block);
4844 if (EDGE_COUNT (idom1->succs) == 2)
4845 if (gcond *last1 = safe_dyn_cast <gcond *> (last_stmt (idom1)))
4847 /* ??? We want to use SSA_VAL here. But possibly not
4848 allow VN_TOP. */
4849 vp1->cclhs = vn_valueize (gimple_cond_lhs (last1));
4850 vp1->ccrhs = vn_valueize (gimple_cond_rhs (last1));
4853 vp1->hashcode = vn_phi_compute_hash (vp1);
4854 slot = valid_info->phis->find_slot_with_hash (vp1, vp1->hashcode, NO_INSERT);
4855 if (!slot)
4856 return NULL_TREE;
4857 return (*slot)->result;
4860 /* Insert PHI into the current hash table with a value number of
4861 RESULT. */
4863 static vn_phi_t
4864 vn_phi_insert (gimple *phi, tree result, bool backedges_varying_p)
4866 vn_phi_s **slot;
4867 vn_phi_t vp1 = (vn_phi_t) obstack_alloc (&vn_tables_obstack,
4868 sizeof (vn_phi_s)
4869 + ((gimple_phi_num_args (phi) - 1)
4870 * sizeof (tree)));
4871 edge e;
4872 edge_iterator ei;
4874 /* Canonicalize the SSA_NAME's to their value number. */
4875 FOR_EACH_EDGE (e, ei, gimple_bb (phi)->preds)
4877 tree def = PHI_ARG_DEF_FROM_EDGE (phi, e);
4878 if (TREE_CODE (def) == SSA_NAME
4879 && (!backedges_varying_p || !(e->flags & EDGE_DFS_BACK)))
4881 if (!virtual_operand_p (def)
4882 && ssa_undefined_value_p (def, false))
4883 def = VN_TOP;
4884 else
4885 def = SSA_VAL (def);
4887 vp1->phiargs[e->dest_idx] = def;
4889 vp1->value_id = VN_INFO (result)->value_id;
4890 vp1->type = TREE_TYPE (gimple_phi_result (phi));
4891 vp1->block = gimple_bb (phi);
4892 /* Extract values of the controlling condition. */
4893 vp1->cclhs = NULL_TREE;
4894 vp1->ccrhs = NULL_TREE;
4895 if (EDGE_COUNT (vp1->block->preds) == 2
4896 && vp1->block->loop_father->header != vp1->block)
4898 basic_block idom1 = get_immediate_dominator (CDI_DOMINATORS, vp1->block);
4899 if (EDGE_COUNT (idom1->succs) == 2)
4900 if (gcond *last1 = safe_dyn_cast <gcond *> (last_stmt (idom1)))
4902 /* ??? We want to use SSA_VAL here. But possibly not
4903 allow VN_TOP. */
4904 vp1->cclhs = vn_valueize (gimple_cond_lhs (last1));
4905 vp1->ccrhs = vn_valueize (gimple_cond_rhs (last1));
4908 vp1->result = result;
4909 vp1->hashcode = vn_phi_compute_hash (vp1);
4911 slot = valid_info->phis->find_slot_with_hash (vp1, vp1->hashcode, INSERT);
4912 gcc_assert (!*slot);
4914 *slot = vp1;
4915 vp1->next = last_inserted_phi;
4916 last_inserted_phi = vp1;
4917 return vp1;
4921 /* Return true if BB1 is dominated by BB2 taking into account edges
4922 that are not executable. When ALLOW_BACK is false consider not
4923 executable backedges as executable. */
4925 static bool
4926 dominated_by_p_w_unex (basic_block bb1, basic_block bb2, bool allow_back)
4928 edge_iterator ei;
4929 edge e;
4931 if (dominated_by_p (CDI_DOMINATORS, bb1, bb2))
4932 return true;
4934 /* Before iterating we'd like to know if there exists a
4935 (executable) path from bb2 to bb1 at all, if not we can
4936 directly return false. For now simply iterate once. */
4938 /* Iterate to the single executable bb1 predecessor. */
4939 if (EDGE_COUNT (bb1->preds) > 1)
4941 edge prede = NULL;
4942 FOR_EACH_EDGE (e, ei, bb1->preds)
4943 if ((e->flags & EDGE_EXECUTABLE)
4944 || (!allow_back && (e->flags & EDGE_DFS_BACK)))
4946 if (prede)
4948 prede = NULL;
4949 break;
4951 prede = e;
4953 if (prede)
4955 bb1 = prede->src;
4957 /* Re-do the dominance check with changed bb1. */
4958 if (dominated_by_p (CDI_DOMINATORS, bb1, bb2))
4959 return true;
4963 /* Iterate to the single executable bb2 successor. */
4964 if (EDGE_COUNT (bb2->succs) > 1)
4966 edge succe = NULL;
4967 FOR_EACH_EDGE (e, ei, bb2->succs)
4968 if ((e->flags & EDGE_EXECUTABLE)
4969 || (!allow_back && (e->flags & EDGE_DFS_BACK)))
4971 if (succe)
4973 succe = NULL;
4974 break;
4976 succe = e;
4978 if (succe)
4980 /* Verify the reached block is only reached through succe.
4981 If there is only one edge we can spare us the dominator
4982 check and iterate directly. */
4983 if (EDGE_COUNT (succe->dest->preds) > 1)
4985 FOR_EACH_EDGE (e, ei, succe->dest->preds)
4986 if (e != succe
4987 && ((e->flags & EDGE_EXECUTABLE)
4988 || (!allow_back && (e->flags & EDGE_DFS_BACK))))
4990 succe = NULL;
4991 break;
4994 if (succe)
4996 bb2 = succe->dest;
4998 /* Re-do the dominance check with changed bb2. */
4999 if (dominated_by_p (CDI_DOMINATORS, bb1, bb2))
5000 return true;
5005 /* We could now iterate updating bb1 / bb2. */
5006 return false;
5009 /* Set the value number of FROM to TO, return true if it has changed
5010 as a result. */
5012 static inline bool
5013 set_ssa_val_to (tree from, tree to)
5015 vn_ssa_aux_t from_info = VN_INFO (from);
5016 tree currval = from_info->valnum; // SSA_VAL (from)
5017 poly_int64 toff, coff;
5018 bool curr_undefined = false;
5019 bool curr_invariant = false;
5021 /* The only thing we allow as value numbers are ssa_names
5022 and invariants. So assert that here. We don't allow VN_TOP
5023 as visiting a stmt should produce a value-number other than
5024 that.
5025 ??? Still VN_TOP can happen for unreachable code, so force
5026 it to varying in that case. Not all code is prepared to
5027 get VN_TOP on valueization. */
5028 if (to == VN_TOP)
5030 /* ??? When iterating and visiting PHI <undef, backedge-value>
5031 for the first time we rightfully get VN_TOP and we need to
5032 preserve that to optimize for example gcc.dg/tree-ssa/ssa-sccvn-2.c.
5033 With SCCVN we were simply lucky we iterated the other PHI
5034 cycles first and thus visited the backedge-value DEF. */
5035 if (currval == VN_TOP)
5036 goto set_and_exit;
5037 if (dump_file && (dump_flags & TDF_DETAILS))
5038 fprintf (dump_file, "Forcing value number to varying on "
5039 "receiving VN_TOP\n");
5040 to = from;
5043 gcc_checking_assert (to != NULL_TREE
5044 && ((TREE_CODE (to) == SSA_NAME
5045 && (to == from || SSA_VAL (to) == to))
5046 || is_gimple_min_invariant (to)));
5048 if (from != to)
5050 if (currval == from)
5052 if (dump_file && (dump_flags & TDF_DETAILS))
5054 fprintf (dump_file, "Not changing value number of ");
5055 print_generic_expr (dump_file, from);
5056 fprintf (dump_file, " from VARYING to ");
5057 print_generic_expr (dump_file, to);
5058 fprintf (dump_file, "\n");
5060 return false;
5062 curr_invariant = is_gimple_min_invariant (currval);
5063 curr_undefined = (TREE_CODE (currval) == SSA_NAME
5064 && !virtual_operand_p (currval)
5065 && ssa_undefined_value_p (currval, false));
5066 if (currval != VN_TOP
5067 && !curr_invariant
5068 && !curr_undefined
5069 && is_gimple_min_invariant (to))
5071 if (dump_file && (dump_flags & TDF_DETAILS))
5073 fprintf (dump_file, "Forcing VARYING instead of changing "
5074 "value number of ");
5075 print_generic_expr (dump_file, from);
5076 fprintf (dump_file, " from ");
5077 print_generic_expr (dump_file, currval);
5078 fprintf (dump_file, " (non-constant) to ");
5079 print_generic_expr (dump_file, to);
5080 fprintf (dump_file, " (constant)\n");
5082 to = from;
5084 else if (currval != VN_TOP
5085 && !curr_undefined
5086 && TREE_CODE (to) == SSA_NAME
5087 && !virtual_operand_p (to)
5088 && ssa_undefined_value_p (to, false))
5090 if (dump_file && (dump_flags & TDF_DETAILS))
5092 fprintf (dump_file, "Forcing VARYING instead of changing "
5093 "value number of ");
5094 print_generic_expr (dump_file, from);
5095 fprintf (dump_file, " from ");
5096 print_generic_expr (dump_file, currval);
5097 fprintf (dump_file, " (non-undefined) to ");
5098 print_generic_expr (dump_file, to);
5099 fprintf (dump_file, " (undefined)\n");
5101 to = from;
5103 else if (TREE_CODE (to) == SSA_NAME
5104 && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (to))
5105 to = from;
5108 set_and_exit:
5109 if (dump_file && (dump_flags & TDF_DETAILS))
5111 fprintf (dump_file, "Setting value number of ");
5112 print_generic_expr (dump_file, from);
5113 fprintf (dump_file, " to ");
5114 print_generic_expr (dump_file, to);
5117 if (currval != to
5118 && !operand_equal_p (currval, to, 0)
5119 /* Different undefined SSA names are not actually different. See
5120 PR82320 for a testcase were we'd otherwise not terminate iteration. */
5121 && !(curr_undefined
5122 && TREE_CODE (to) == SSA_NAME
5123 && !virtual_operand_p (to)
5124 && ssa_undefined_value_p (to, false))
5125 /* ??? For addresses involving volatile objects or types operand_equal_p
5126 does not reliably detect ADDR_EXPRs as equal. We know we are only
5127 getting invariant gimple addresses here, so can use
5128 get_addr_base_and_unit_offset to do this comparison. */
5129 && !(TREE_CODE (currval) == ADDR_EXPR
5130 && TREE_CODE (to) == ADDR_EXPR
5131 && (get_addr_base_and_unit_offset (TREE_OPERAND (currval, 0), &coff)
5132 == get_addr_base_and_unit_offset (TREE_OPERAND (to, 0), &toff))
5133 && known_eq (coff, toff)))
5135 if (to != from
5136 && currval != VN_TOP
5137 && !curr_undefined
5138 /* We do not want to allow lattice transitions from one value
5139 to another since that may lead to not terminating iteration
5140 (see PR95049). Since there's no convenient way to check
5141 for the allowed transition of VAL -> PHI (loop entry value,
5142 same on two PHIs, to same PHI result) we restrict the check
5143 to invariants. */
5144 && curr_invariant
5145 && is_gimple_min_invariant (to))
5147 if (dump_file && (dump_flags & TDF_DETAILS))
5148 fprintf (dump_file, " forced VARYING");
5149 to = from;
5151 if (dump_file && (dump_flags & TDF_DETAILS))
5152 fprintf (dump_file, " (changed)\n");
5153 from_info->valnum = to;
5154 return true;
5156 if (dump_file && (dump_flags & TDF_DETAILS))
5157 fprintf (dump_file, "\n");
5158 return false;
5161 /* Set all definitions in STMT to value number to themselves.
5162 Return true if a value number changed. */
5164 static bool
5165 defs_to_varying (gimple *stmt)
5167 bool changed = false;
5168 ssa_op_iter iter;
5169 def_operand_p defp;
5171 FOR_EACH_SSA_DEF_OPERAND (defp, stmt, iter, SSA_OP_ALL_DEFS)
5173 tree def = DEF_FROM_PTR (defp);
5174 changed |= set_ssa_val_to (def, def);
5176 return changed;
5179 /* Visit a copy between LHS and RHS, return true if the value number
5180 changed. */
5182 static bool
5183 visit_copy (tree lhs, tree rhs)
5185 /* Valueize. */
5186 rhs = SSA_VAL (rhs);
5188 return set_ssa_val_to (lhs, rhs);
5191 /* Lookup a value for OP in type WIDE_TYPE where the value in type of OP
5192 is the same. */
5194 static tree
5195 valueized_wider_op (tree wide_type, tree op, bool allow_truncate)
5197 if (TREE_CODE (op) == SSA_NAME)
5198 op = vn_valueize (op);
5200 /* Either we have the op widened available. */
5201 tree ops[3] = {};
5202 ops[0] = op;
5203 tree tem = vn_nary_op_lookup_pieces (1, NOP_EXPR,
5204 wide_type, ops, NULL);
5205 if (tem)
5206 return tem;
5208 /* Or the op is truncated from some existing value. */
5209 if (allow_truncate && TREE_CODE (op) == SSA_NAME)
5211 gimple *def = SSA_NAME_DEF_STMT (op);
5212 if (is_gimple_assign (def)
5213 && CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (def)))
5215 tem = gimple_assign_rhs1 (def);
5216 if (useless_type_conversion_p (wide_type, TREE_TYPE (tem)))
5218 if (TREE_CODE (tem) == SSA_NAME)
5219 tem = vn_valueize (tem);
5220 return tem;
5225 /* For constants simply extend it. */
5226 if (TREE_CODE (op) == INTEGER_CST)
5227 return wide_int_to_tree (wide_type, wi::to_widest (op));
5229 return NULL_TREE;
5232 /* Visit a nary operator RHS, value number it, and return true if the
5233 value number of LHS has changed as a result. */
5235 static bool
5236 visit_nary_op (tree lhs, gassign *stmt)
5238 vn_nary_op_t vnresult;
5239 tree result = vn_nary_op_lookup_stmt (stmt, &vnresult);
5240 if (! result && vnresult)
5241 result = vn_nary_op_get_predicated_value (vnresult, gimple_bb (stmt));
5242 if (result)
5243 return set_ssa_val_to (lhs, result);
5245 /* Do some special pattern matching for redundancies of operations
5246 in different types. */
5247 enum tree_code code = gimple_assign_rhs_code (stmt);
5248 tree type = TREE_TYPE (lhs);
5249 tree rhs1 = gimple_assign_rhs1 (stmt);
5250 switch (code)
5252 CASE_CONVERT:
5253 /* Match arithmetic done in a different type where we can easily
5254 substitute the result from some earlier sign-changed or widened
5255 operation. */
5256 if (INTEGRAL_TYPE_P (type)
5257 && TREE_CODE (rhs1) == SSA_NAME
5258 /* We only handle sign-changes, zero-extension -> & mask or
5259 sign-extension if we know the inner operation doesn't
5260 overflow. */
5261 && (((TYPE_UNSIGNED (TREE_TYPE (rhs1))
5262 || (INTEGRAL_TYPE_P (TREE_TYPE (rhs1))
5263 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (rhs1))))
5264 && TYPE_PRECISION (type) > TYPE_PRECISION (TREE_TYPE (rhs1)))
5265 || TYPE_PRECISION (type) == TYPE_PRECISION (TREE_TYPE (rhs1))))
5267 gassign *def = dyn_cast <gassign *> (SSA_NAME_DEF_STMT (rhs1));
5268 if (def
5269 && (gimple_assign_rhs_code (def) == PLUS_EXPR
5270 || gimple_assign_rhs_code (def) == MINUS_EXPR
5271 || gimple_assign_rhs_code (def) == MULT_EXPR))
5273 tree ops[3] = {};
5274 /* When requiring a sign-extension we cannot model a
5275 previous truncation with a single op so don't bother. */
5276 bool allow_truncate = TYPE_UNSIGNED (TREE_TYPE (rhs1));
5277 /* Either we have the op widened available. */
5278 ops[0] = valueized_wider_op (type, gimple_assign_rhs1 (def),
5279 allow_truncate);
5280 if (ops[0])
5281 ops[1] = valueized_wider_op (type, gimple_assign_rhs2 (def),
5282 allow_truncate);
5283 if (ops[0] && ops[1])
5285 ops[0] = vn_nary_op_lookup_pieces
5286 (2, gimple_assign_rhs_code (def), type, ops, NULL);
5287 /* We have wider operation available. */
5288 if (ops[0]
5289 /* If the leader is a wrapping operation we can
5290 insert it for code hoisting w/o introducing
5291 undefined overflow. If it is not it has to
5292 be available. See PR86554. */
5293 && (TYPE_OVERFLOW_WRAPS (TREE_TYPE (ops[0]))
5294 || (rpo_avail && vn_context_bb
5295 && rpo_avail->eliminate_avail (vn_context_bb,
5296 ops[0]))))
5298 unsigned lhs_prec = TYPE_PRECISION (type);
5299 unsigned rhs_prec = TYPE_PRECISION (TREE_TYPE (rhs1));
5300 if (lhs_prec == rhs_prec
5301 || (INTEGRAL_TYPE_P (TREE_TYPE (rhs1))
5302 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (rhs1))))
5304 gimple_match_op match_op (gimple_match_cond::UNCOND,
5305 NOP_EXPR, type, ops[0]);
5306 result = vn_nary_build_or_lookup (&match_op);
5307 if (result)
5309 bool changed = set_ssa_val_to (lhs, result);
5310 vn_nary_op_insert_stmt (stmt, result);
5311 return changed;
5314 else
5316 tree mask = wide_int_to_tree
5317 (type, wi::mask (rhs_prec, false, lhs_prec));
5318 gimple_match_op match_op (gimple_match_cond::UNCOND,
5319 BIT_AND_EXPR,
5320 TREE_TYPE (lhs),
5321 ops[0], mask);
5322 result = vn_nary_build_or_lookup (&match_op);
5323 if (result)
5325 bool changed = set_ssa_val_to (lhs, result);
5326 vn_nary_op_insert_stmt (stmt, result);
5327 return changed;
5334 break;
5335 case BIT_AND_EXPR:
5336 if (INTEGRAL_TYPE_P (type)
5337 && TREE_CODE (rhs1) == SSA_NAME
5338 && TREE_CODE (gimple_assign_rhs2 (stmt)) == INTEGER_CST
5339 && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (rhs1)
5340 && default_vn_walk_kind != VN_NOWALK
5341 && CHAR_BIT == 8
5342 && BITS_PER_UNIT == 8
5343 && BYTES_BIG_ENDIAN == WORDS_BIG_ENDIAN
5344 && !integer_all_onesp (gimple_assign_rhs2 (stmt))
5345 && !integer_zerop (gimple_assign_rhs2 (stmt)))
5347 gassign *ass = dyn_cast <gassign *> (SSA_NAME_DEF_STMT (rhs1));
5348 if (ass
5349 && !gimple_has_volatile_ops (ass)
5350 && vn_get_stmt_kind (ass) == VN_REFERENCE)
5352 tree last_vuse = gimple_vuse (ass);
5353 tree op = gimple_assign_rhs1 (ass);
5354 tree result = vn_reference_lookup (op, gimple_vuse (ass),
5355 default_vn_walk_kind,
5356 NULL, true, &last_vuse,
5357 gimple_assign_rhs2 (stmt));
5358 if (result
5359 && useless_type_conversion_p (TREE_TYPE (result),
5360 TREE_TYPE (op)))
5361 return set_ssa_val_to (lhs, result);
5364 break;
5365 case TRUNC_DIV_EXPR:
5366 if (TYPE_UNSIGNED (type))
5367 break;
5368 /* Fallthru. */
5369 case RDIV_EXPR:
5370 case MULT_EXPR:
5371 /* Match up ([-]a){/,*}([-])b with v=a{/,*}b, replacing it with -v. */
5372 if (! HONOR_SIGN_DEPENDENT_ROUNDING (type))
5374 tree rhs[2];
5375 rhs[0] = rhs1;
5376 rhs[1] = gimple_assign_rhs2 (stmt);
5377 for (unsigned i = 0; i <= 1; ++i)
5379 unsigned j = i == 0 ? 1 : 0;
5380 tree ops[2];
5381 gimple_match_op match_op (gimple_match_cond::UNCOND,
5382 NEGATE_EXPR, type, rhs[i]);
5383 ops[i] = vn_nary_build_or_lookup_1 (&match_op, false, true);
5384 ops[j] = rhs[j];
5385 if (ops[i]
5386 && (ops[0] = vn_nary_op_lookup_pieces (2, code,
5387 type, ops, NULL)))
5389 gimple_match_op match_op (gimple_match_cond::UNCOND,
5390 NEGATE_EXPR, type, ops[0]);
5391 result = vn_nary_build_or_lookup_1 (&match_op, true, false);
5392 if (result)
5394 bool changed = set_ssa_val_to (lhs, result);
5395 vn_nary_op_insert_stmt (stmt, result);
5396 return changed;
5401 break;
5402 case LSHIFT_EXPR:
5403 /* For X << C, use the value number of X * (1 << C). */
5404 if (INTEGRAL_TYPE_P (type)
5405 && TYPE_OVERFLOW_WRAPS (type)
5406 && !TYPE_SATURATING (type))
5408 tree rhs2 = gimple_assign_rhs2 (stmt);
5409 if (TREE_CODE (rhs2) == INTEGER_CST
5410 && tree_fits_uhwi_p (rhs2)
5411 && tree_to_uhwi (rhs2) < TYPE_PRECISION (type))
5413 wide_int w = wi::set_bit_in_zero (tree_to_uhwi (rhs2),
5414 TYPE_PRECISION (type));
5415 gimple_match_op match_op (gimple_match_cond::UNCOND,
5416 MULT_EXPR, type, rhs1,
5417 wide_int_to_tree (type, w));
5418 result = vn_nary_build_or_lookup (&match_op);
5419 if (result)
5421 bool changed = set_ssa_val_to (lhs, result);
5422 if (TREE_CODE (result) == SSA_NAME)
5423 vn_nary_op_insert_stmt (stmt, result);
5424 return changed;
5428 break;
5429 default:
5430 break;
5433 bool changed = set_ssa_val_to (lhs, lhs);
5434 vn_nary_op_insert_stmt (stmt, lhs);
5435 return changed;
5438 /* Visit a call STMT storing into LHS. Return true if the value number
5439 of the LHS has changed as a result. */
5441 static bool
5442 visit_reference_op_call (tree lhs, gcall *stmt)
5444 bool changed = false;
5445 struct vn_reference_s vr1;
5446 vn_reference_t vnresult = NULL;
5447 tree vdef = gimple_vdef (stmt);
5448 modref_summary *summary;
5450 /* Non-ssa lhs is handled in copy_reference_ops_from_call. */
5451 if (lhs && TREE_CODE (lhs) != SSA_NAME)
5452 lhs = NULL_TREE;
5454 vn_reference_lookup_call (stmt, &vnresult, &vr1);
5456 /* If the lookup did not succeed for pure functions try to use
5457 modref info to find a candidate to CSE to. */
5458 const unsigned accesses_limit = 8;
5459 if (!vnresult
5460 && !vdef
5461 && lhs
5462 && gimple_vuse (stmt)
5463 && (((summary = get_modref_function_summary (stmt, NULL))
5464 && !summary->global_memory_read
5465 && summary->load_accesses < accesses_limit)
5466 || gimple_call_flags (stmt) & ECF_CONST))
5468 /* First search if we can do someting useful and build a
5469 vector of all loads we have to check. */
5470 bool unknown_memory_access = false;
5471 auto_vec<ao_ref, accesses_limit> accesses;
5472 unsigned load_accesses = summary ? summary->load_accesses : 0;
5473 if (!unknown_memory_access)
5474 /* Add loads done as part of setting up the call arguments.
5475 That's also necessary for CONST functions which will
5476 not have a modref summary. */
5477 for (unsigned i = 0; i < gimple_call_num_args (stmt); ++i)
5479 tree arg = gimple_call_arg (stmt, i);
5480 if (TREE_CODE (arg) != SSA_NAME
5481 && !is_gimple_min_invariant (arg))
5483 if (accesses.length () >= accesses_limit - load_accesses)
5485 unknown_memory_access = true;
5486 break;
5488 accesses.quick_grow (accesses.length () + 1);
5489 ao_ref_init (&accesses.last (), arg);
5492 if (summary && !unknown_memory_access)
5494 /* Add loads as analyzed by IPA modref. */
5495 for (auto base_node : summary->loads->bases)
5496 if (unknown_memory_access)
5497 break;
5498 else for (auto ref_node : base_node->refs)
5499 if (unknown_memory_access)
5500 break;
5501 else for (auto access_node : ref_node->accesses)
5503 accesses.quick_grow (accesses.length () + 1);
5504 ao_ref *r = &accesses.last ();
5505 if (!access_node.get_ao_ref (stmt, r))
5507 /* Initialize a ref based on the argument and
5508 unknown offset if possible. */
5509 tree arg = access_node.get_call_arg (stmt);
5510 if (arg && TREE_CODE (arg) == SSA_NAME)
5511 arg = SSA_VAL (arg);
5512 if (arg
5513 && TREE_CODE (arg) == ADDR_EXPR
5514 && (arg = get_base_address (arg))
5515 && DECL_P (arg))
5517 ao_ref_init (r, arg);
5518 r->ref = NULL_TREE;
5519 r->base = arg;
5521 else
5523 unknown_memory_access = true;
5524 break;
5527 r->base_alias_set = base_node->base;
5528 r->ref_alias_set = ref_node->ref;
5532 /* Walk the VUSE->VDEF chain optimistically trying to find an entry
5533 for the call in the hashtable. */
5534 unsigned limit = (unknown_memory_access
5536 : (param_sccvn_max_alias_queries_per_access
5537 / (accesses.length () + 1)));
5538 tree saved_vuse = vr1.vuse;
5539 hashval_t saved_hashcode = vr1.hashcode;
5540 while (limit > 0 && !vnresult && !SSA_NAME_IS_DEFAULT_DEF (vr1.vuse))
5542 vr1.hashcode = vr1.hashcode - SSA_NAME_VERSION (vr1.vuse);
5543 gimple *def = SSA_NAME_DEF_STMT (vr1.vuse);
5544 /* ??? We could use fancy stuff like in walk_non_aliased_vuses, but
5545 do not bother for now. */
5546 if (is_a <gphi *> (def))
5547 break;
5548 vr1.vuse = vuse_ssa_val (gimple_vuse (def));
5549 vr1.hashcode = vr1.hashcode + SSA_NAME_VERSION (vr1.vuse);
5550 vn_reference_lookup_1 (&vr1, &vnresult);
5551 limit--;
5554 /* If we found a candidate to CSE to verify it is valid. */
5555 if (vnresult && !accesses.is_empty ())
5557 tree vuse = vuse_ssa_val (gimple_vuse (stmt));
5558 while (vnresult && vuse != vr1.vuse)
5560 gimple *def = SSA_NAME_DEF_STMT (vuse);
5561 for (auto &ref : accesses)
5563 /* ??? stmt_may_clobber_ref_p_1 does per stmt constant
5564 analysis overhead that we might be able to cache. */
5565 if (stmt_may_clobber_ref_p_1 (def, &ref, true))
5567 vnresult = NULL;
5568 break;
5571 vuse = vuse_ssa_val (gimple_vuse (def));
5574 vr1.vuse = saved_vuse;
5575 vr1.hashcode = saved_hashcode;
5578 if (vnresult)
5580 if (vdef)
5582 if (vnresult->result_vdef)
5583 changed |= set_ssa_val_to (vdef, vnresult->result_vdef);
5584 else if (!lhs && gimple_call_lhs (stmt))
5585 /* If stmt has non-SSA_NAME lhs, value number the vdef to itself,
5586 as the call still acts as a lhs store. */
5587 changed |= set_ssa_val_to (vdef, vdef);
5588 else
5589 /* If the call was discovered to be pure or const reflect
5590 that as far as possible. */
5591 changed |= set_ssa_val_to (vdef,
5592 vuse_ssa_val (gimple_vuse (stmt)));
5595 if (!vnresult->result && lhs)
5596 vnresult->result = lhs;
5598 if (vnresult->result && lhs)
5599 changed |= set_ssa_val_to (lhs, vnresult->result);
5601 else
5603 vn_reference_t vr2;
5604 vn_reference_s **slot;
5605 tree vdef_val = vdef;
5606 if (vdef)
5608 /* If we value numbered an indirect functions function to
5609 one not clobbering memory value number its VDEF to its
5610 VUSE. */
5611 tree fn = gimple_call_fn (stmt);
5612 if (fn && TREE_CODE (fn) == SSA_NAME)
5614 fn = SSA_VAL (fn);
5615 if (TREE_CODE (fn) == ADDR_EXPR
5616 && TREE_CODE (TREE_OPERAND (fn, 0)) == FUNCTION_DECL
5617 && (flags_from_decl_or_type (TREE_OPERAND (fn, 0))
5618 & (ECF_CONST | ECF_PURE))
5619 /* If stmt has non-SSA_NAME lhs, value number the
5620 vdef to itself, as the call still acts as a lhs
5621 store. */
5622 && (lhs || gimple_call_lhs (stmt) == NULL_TREE))
5623 vdef_val = vuse_ssa_val (gimple_vuse (stmt));
5625 changed |= set_ssa_val_to (vdef, vdef_val);
5627 if (lhs)
5628 changed |= set_ssa_val_to (lhs, lhs);
5629 vr2 = XOBNEW (&vn_tables_obstack, vn_reference_s);
5630 vr2->vuse = vr1.vuse;
5631 /* As we are not walking the virtual operand chain we know the
5632 shared_lookup_references are still original so we can re-use
5633 them here. */
5634 vr2->operands = vr1.operands.copy ();
5635 vr2->type = vr1.type;
5636 vr2->punned = vr1.punned;
5637 vr2->set = vr1.set;
5638 vr2->base_set = vr1.base_set;
5639 vr2->hashcode = vr1.hashcode;
5640 vr2->result = lhs;
5641 vr2->result_vdef = vdef_val;
5642 vr2->value_id = 0;
5643 slot = valid_info->references->find_slot_with_hash (vr2, vr2->hashcode,
5644 INSERT);
5645 gcc_assert (!*slot);
5646 *slot = vr2;
5647 vr2->next = last_inserted_ref;
5648 last_inserted_ref = vr2;
5651 return changed;
5654 /* Visit a load from a reference operator RHS, part of STMT, value number it,
5655 and return true if the value number of the LHS has changed as a result. */
5657 static bool
5658 visit_reference_op_load (tree lhs, tree op, gimple *stmt)
5660 bool changed = false;
5661 tree result;
5662 vn_reference_t res;
5664 tree vuse = gimple_vuse (stmt);
5665 tree last_vuse = vuse;
5666 result = vn_reference_lookup (op, vuse, default_vn_walk_kind, &res, true, &last_vuse);
5668 /* We handle type-punning through unions by value-numbering based
5669 on offset and size of the access. Be prepared to handle a
5670 type-mismatch here via creating a VIEW_CONVERT_EXPR. */
5671 if (result
5672 && !useless_type_conversion_p (TREE_TYPE (result), TREE_TYPE (op)))
5674 /* Avoid the type punning in case the result mode has padding where
5675 the op we lookup has not. */
5676 if (maybe_lt (GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (result))),
5677 GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (op)))))
5678 result = NULL_TREE;
5679 else
5681 /* We will be setting the value number of lhs to the value number
5682 of VIEW_CONVERT_EXPR <TREE_TYPE (result)> (result).
5683 So first simplify and lookup this expression to see if it
5684 is already available. */
5685 gimple_match_op res_op (gimple_match_cond::UNCOND,
5686 VIEW_CONVERT_EXPR, TREE_TYPE (op), result);
5687 result = vn_nary_build_or_lookup (&res_op);
5688 if (result
5689 && TREE_CODE (result) == SSA_NAME
5690 && VN_INFO (result)->needs_insertion)
5691 /* Track whether this is the canonical expression for different
5692 typed loads. We use that as a stopgap measure for code
5693 hoisting when dealing with floating point loads. */
5694 res->punned = true;
5697 /* When building the conversion fails avoid inserting the reference
5698 again. */
5699 if (!result)
5700 return set_ssa_val_to (lhs, lhs);
5703 if (result)
5704 changed = set_ssa_val_to (lhs, result);
5705 else
5707 changed = set_ssa_val_to (lhs, lhs);
5708 vn_reference_insert (op, lhs, last_vuse, NULL_TREE);
5709 if (vuse && SSA_VAL (last_vuse) != SSA_VAL (vuse))
5711 if (dump_file && (dump_flags & TDF_DETAILS))
5713 fprintf (dump_file, "Using extra use virtual operand ");
5714 print_generic_expr (dump_file, last_vuse);
5715 fprintf (dump_file, "\n");
5717 vn_reference_insert (op, lhs, vuse, NULL_TREE);
5721 return changed;
5725 /* Visit a store to a reference operator LHS, part of STMT, value number it,
5726 and return true if the value number of the LHS has changed as a result. */
5728 static bool
5729 visit_reference_op_store (tree lhs, tree op, gimple *stmt)
5731 bool changed = false;
5732 vn_reference_t vnresult = NULL;
5733 tree assign;
5734 bool resultsame = false;
5735 tree vuse = gimple_vuse (stmt);
5736 tree vdef = gimple_vdef (stmt);
5738 if (TREE_CODE (op) == SSA_NAME)
5739 op = SSA_VAL (op);
5741 /* First we want to lookup using the *vuses* from the store and see
5742 if there the last store to this location with the same address
5743 had the same value.
5745 The vuses represent the memory state before the store. If the
5746 memory state, address, and value of the store is the same as the
5747 last store to this location, then this store will produce the
5748 same memory state as that store.
5750 In this case the vdef versions for this store are value numbered to those
5751 vuse versions, since they represent the same memory state after
5752 this store.
5754 Otherwise, the vdefs for the store are used when inserting into
5755 the table, since the store generates a new memory state. */
5757 vn_reference_lookup (lhs, vuse, VN_NOWALK, &vnresult, false);
5758 if (vnresult
5759 && vnresult->result)
5761 tree result = vnresult->result;
5762 gcc_checking_assert (TREE_CODE (result) != SSA_NAME
5763 || result == SSA_VAL (result));
5764 resultsame = expressions_equal_p (result, op);
5765 if (resultsame)
5767 /* If the TBAA state isn't compatible for downstream reads
5768 we cannot value-number the VDEFs the same. */
5769 ao_ref lhs_ref;
5770 ao_ref_init (&lhs_ref, lhs);
5771 alias_set_type set = ao_ref_alias_set (&lhs_ref);
5772 alias_set_type base_set = ao_ref_base_alias_set (&lhs_ref);
5773 if ((vnresult->set != set
5774 && ! alias_set_subset_of (set, vnresult->set))
5775 || (vnresult->base_set != base_set
5776 && ! alias_set_subset_of (base_set, vnresult->base_set)))
5777 resultsame = false;
5781 if (!resultsame)
5783 if (dump_file && (dump_flags & TDF_DETAILS))
5785 fprintf (dump_file, "No store match\n");
5786 fprintf (dump_file, "Value numbering store ");
5787 print_generic_expr (dump_file, lhs);
5788 fprintf (dump_file, " to ");
5789 print_generic_expr (dump_file, op);
5790 fprintf (dump_file, "\n");
5792 /* Have to set value numbers before insert, since insert is
5793 going to valueize the references in-place. */
5794 if (vdef)
5795 changed |= set_ssa_val_to (vdef, vdef);
5797 /* Do not insert structure copies into the tables. */
5798 if (is_gimple_min_invariant (op)
5799 || is_gimple_reg (op))
5800 vn_reference_insert (lhs, op, vdef, NULL);
5802 /* Only perform the following when being called from PRE
5803 which embeds tail merging. */
5804 if (default_vn_walk_kind == VN_WALK)
5806 assign = build2 (MODIFY_EXPR, TREE_TYPE (lhs), lhs, op);
5807 vn_reference_lookup (assign, vuse, VN_NOWALK, &vnresult, false);
5808 if (!vnresult)
5809 vn_reference_insert (assign, lhs, vuse, vdef);
5812 else
5814 /* We had a match, so value number the vdef to have the value
5815 number of the vuse it came from. */
5817 if (dump_file && (dump_flags & TDF_DETAILS))
5818 fprintf (dump_file, "Store matched earlier value, "
5819 "value numbering store vdefs to matching vuses.\n");
5821 changed |= set_ssa_val_to (vdef, SSA_VAL (vuse));
5824 return changed;
5827 /* Visit and value number PHI, return true if the value number
5828 changed. When BACKEDGES_VARYING_P is true then assume all
5829 backedge values are varying. When INSERTED is not NULL then
5830 this is just a ahead query for a possible iteration, set INSERTED
5831 to true if we'd insert into the hashtable. */
5833 static bool
5834 visit_phi (gimple *phi, bool *inserted, bool backedges_varying_p)
5836 tree result, sameval = VN_TOP, seen_undef = NULL_TREE;
5837 tree backedge_val = NULL_TREE;
5838 bool seen_non_backedge = false;
5839 tree sameval_base = NULL_TREE;
5840 poly_int64 soff, doff;
5841 unsigned n_executable = 0;
5842 edge_iterator ei;
5843 edge e, sameval_e = NULL;
5845 /* TODO: We could check for this in initialization, and replace this
5846 with a gcc_assert. */
5847 if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (PHI_RESULT (phi)))
5848 return set_ssa_val_to (PHI_RESULT (phi), PHI_RESULT (phi));
5850 /* We track whether a PHI was CSEd to to avoid excessive iterations
5851 that would be necessary only because the PHI changed arguments
5852 but not value. */
5853 if (!inserted)
5854 gimple_set_plf (phi, GF_PLF_1, false);
5856 /* See if all non-TOP arguments have the same value. TOP is
5857 equivalent to everything, so we can ignore it. */
5858 basic_block bb = gimple_bb (phi);
5859 FOR_EACH_EDGE (e, ei, bb->preds)
5860 if (e->flags & EDGE_EXECUTABLE)
5862 tree def = PHI_ARG_DEF_FROM_EDGE (phi, e);
5864 if (def == PHI_RESULT (phi))
5865 continue;
5866 ++n_executable;
5867 if (TREE_CODE (def) == SSA_NAME)
5869 if (!backedges_varying_p || !(e->flags & EDGE_DFS_BACK))
5870 def = SSA_VAL (def);
5871 if (e->flags & EDGE_DFS_BACK)
5872 backedge_val = def;
5874 if (!(e->flags & EDGE_DFS_BACK))
5875 seen_non_backedge = true;
5876 if (def == VN_TOP)
5878 /* Ignore undefined defs for sameval but record one. */
5879 else if (TREE_CODE (def) == SSA_NAME
5880 && ! virtual_operand_p (def)
5881 && ssa_undefined_value_p (def, false))
5882 seen_undef = def;
5883 else if (sameval == VN_TOP)
5885 sameval = def;
5886 sameval_e = e;
5888 else if (expressions_equal_p (def, sameval))
5889 sameval_e = NULL;
5890 else if (virtual_operand_p (def))
5892 sameval = NULL_TREE;
5893 break;
5895 else
5897 /* We know we're arriving only with invariant addresses here,
5898 try harder comparing them. We can do some caching here
5899 which we cannot do in expressions_equal_p. */
5900 if (TREE_CODE (def) == ADDR_EXPR
5901 && TREE_CODE (sameval) == ADDR_EXPR
5902 && sameval_base != (void *)-1)
5904 if (!sameval_base)
5905 sameval_base = get_addr_base_and_unit_offset
5906 (TREE_OPERAND (sameval, 0), &soff);
5907 if (!sameval_base)
5908 sameval_base = (tree)(void *)-1;
5909 else if ((get_addr_base_and_unit_offset
5910 (TREE_OPERAND (def, 0), &doff) == sameval_base)
5911 && known_eq (soff, doff))
5912 continue;
5914 /* There's also the possibility to use equivalences. */
5915 if (!FLOAT_TYPE_P (TREE_TYPE (def))
5916 /* But only do this if we didn't force any of sameval or
5917 val to VARYING because of backedge processing rules. */
5918 && (TREE_CODE (sameval) != SSA_NAME
5919 || SSA_VAL (sameval) == sameval)
5920 && (TREE_CODE (def) != SSA_NAME || SSA_VAL (def) == def))
5922 vn_nary_op_t vnresult;
5923 tree ops[2];
5924 ops[0] = def;
5925 ops[1] = sameval;
5926 tree val = vn_nary_op_lookup_pieces (2, EQ_EXPR,
5927 boolean_type_node,
5928 ops, &vnresult);
5929 if (! val && vnresult && vnresult->predicated_values)
5931 val = vn_nary_op_get_predicated_value (vnresult, e->src);
5932 if (val && integer_truep (val)
5933 && !(sameval_e && (sameval_e->flags & EDGE_DFS_BACK)))
5935 if (dump_file && (dump_flags & TDF_DETAILS))
5937 fprintf (dump_file, "Predication says ");
5938 print_generic_expr (dump_file, def, TDF_NONE);
5939 fprintf (dump_file, " and ");
5940 print_generic_expr (dump_file, sameval, TDF_NONE);
5941 fprintf (dump_file, " are equal on edge %d -> %d\n",
5942 e->src->index, e->dest->index);
5944 continue;
5946 /* If on all previous edges the value was equal to def
5947 we can change sameval to def. */
5948 if (EDGE_COUNT (bb->preds) == 2
5949 && (val = vn_nary_op_get_predicated_value
5950 (vnresult, EDGE_PRED (bb, 0)->src))
5951 && integer_truep (val)
5952 && !(e->flags & EDGE_DFS_BACK))
5954 if (dump_file && (dump_flags & TDF_DETAILS))
5956 fprintf (dump_file, "Predication says ");
5957 print_generic_expr (dump_file, def, TDF_NONE);
5958 fprintf (dump_file, " and ");
5959 print_generic_expr (dump_file, sameval, TDF_NONE);
5960 fprintf (dump_file, " are equal on edge %d -> %d\n",
5961 EDGE_PRED (bb, 0)->src->index,
5962 EDGE_PRED (bb, 0)->dest->index);
5964 sameval = def;
5965 continue;
5969 sameval = NULL_TREE;
5970 break;
5974 /* If the value we want to use is flowing over the backedge and we
5975 should take it as VARYING but it has a non-VARYING value drop to
5976 VARYING.
5977 If we value-number a virtual operand never value-number to the
5978 value from the backedge as that confuses the alias-walking code.
5979 See gcc.dg/torture/pr87176.c. If the value is the same on a
5980 non-backedge everything is OK though. */
5981 bool visited_p;
5982 if ((backedge_val
5983 && !seen_non_backedge
5984 && TREE_CODE (backedge_val) == SSA_NAME
5985 && sameval == backedge_val
5986 && (SSA_NAME_IS_VIRTUAL_OPERAND (backedge_val)
5987 || SSA_VAL (backedge_val) != backedge_val))
5988 /* Do not value-number a virtual operand to sth not visited though
5989 given that allows us to escape a region in alias walking. */
5990 || (sameval
5991 && TREE_CODE (sameval) == SSA_NAME
5992 && !SSA_NAME_IS_DEFAULT_DEF (sameval)
5993 && SSA_NAME_IS_VIRTUAL_OPERAND (sameval)
5994 && (SSA_VAL (sameval, &visited_p), !visited_p)))
5995 /* Note this just drops to VARYING without inserting the PHI into
5996 the hashes. */
5997 result = PHI_RESULT (phi);
5998 /* If none of the edges was executable keep the value-number at VN_TOP,
5999 if only a single edge is exectuable use its value. */
6000 else if (n_executable <= 1)
6001 result = seen_undef ? seen_undef : sameval;
6002 /* If we saw only undefined values and VN_TOP use one of the
6003 undefined values. */
6004 else if (sameval == VN_TOP)
6005 result = seen_undef ? seen_undef : sameval;
6006 /* First see if it is equivalent to a phi node in this block. We prefer
6007 this as it allows IV elimination - see PRs 66502 and 67167. */
6008 else if ((result = vn_phi_lookup (phi, backedges_varying_p)))
6010 if (!inserted
6011 && TREE_CODE (result) == SSA_NAME
6012 && gimple_code (SSA_NAME_DEF_STMT (result)) == GIMPLE_PHI)
6014 gimple_set_plf (SSA_NAME_DEF_STMT (result), GF_PLF_1, true);
6015 if (dump_file && (dump_flags & TDF_DETAILS))
6017 fprintf (dump_file, "Marking CSEd to PHI node ");
6018 print_gimple_expr (dump_file, SSA_NAME_DEF_STMT (result),
6019 0, TDF_SLIM);
6020 fprintf (dump_file, "\n");
6024 /* If all values are the same use that, unless we've seen undefined
6025 values as well and the value isn't constant.
6026 CCP/copyprop have the same restriction to not remove uninit warnings. */
6027 else if (sameval
6028 && (! seen_undef || is_gimple_min_invariant (sameval)))
6029 result = sameval;
6030 else
6032 result = PHI_RESULT (phi);
6033 /* Only insert PHIs that are varying, for constant value numbers
6034 we mess up equivalences otherwise as we are only comparing
6035 the immediate controlling predicates. */
6036 vn_phi_insert (phi, result, backedges_varying_p);
6037 if (inserted)
6038 *inserted = true;
6041 return set_ssa_val_to (PHI_RESULT (phi), result);
6044 /* Try to simplify RHS using equivalences and constant folding. */
6046 static tree
6047 try_to_simplify (gassign *stmt)
6049 enum tree_code code = gimple_assign_rhs_code (stmt);
6050 tree tem;
6052 /* For stores we can end up simplifying a SSA_NAME rhs. Just return
6053 in this case, there is no point in doing extra work. */
6054 if (code == SSA_NAME)
6055 return NULL_TREE;
6057 /* First try constant folding based on our current lattice. */
6058 mprts_hook = vn_lookup_simplify_result;
6059 tem = gimple_fold_stmt_to_constant_1 (stmt, vn_valueize, vn_valueize);
6060 mprts_hook = NULL;
6061 if (tem
6062 && (TREE_CODE (tem) == SSA_NAME
6063 || is_gimple_min_invariant (tem)))
6064 return tem;
6066 return NULL_TREE;
6069 /* Visit and value number STMT, return true if the value number
6070 changed. */
6072 static bool
6073 visit_stmt (gimple *stmt, bool backedges_varying_p = false)
6075 bool changed = false;
6077 if (dump_file && (dump_flags & TDF_DETAILS))
6079 fprintf (dump_file, "Value numbering stmt = ");
6080 print_gimple_stmt (dump_file, stmt, 0);
6083 if (gimple_code (stmt) == GIMPLE_PHI)
6084 changed = visit_phi (stmt, NULL, backedges_varying_p);
6085 else if (gimple_has_volatile_ops (stmt))
6086 changed = defs_to_varying (stmt);
6087 else if (gassign *ass = dyn_cast <gassign *> (stmt))
6089 enum tree_code code = gimple_assign_rhs_code (ass);
6090 tree lhs = gimple_assign_lhs (ass);
6091 tree rhs1 = gimple_assign_rhs1 (ass);
6092 tree simplified;
6094 /* Shortcut for copies. Simplifying copies is pointless,
6095 since we copy the expression and value they represent. */
6096 if (code == SSA_NAME
6097 && TREE_CODE (lhs) == SSA_NAME)
6099 changed = visit_copy (lhs, rhs1);
6100 goto done;
6102 simplified = try_to_simplify (ass);
6103 if (simplified)
6105 if (dump_file && (dump_flags & TDF_DETAILS))
6107 fprintf (dump_file, "RHS ");
6108 print_gimple_expr (dump_file, ass, 0);
6109 fprintf (dump_file, " simplified to ");
6110 print_generic_expr (dump_file, simplified);
6111 fprintf (dump_file, "\n");
6114 /* Setting value numbers to constants will occasionally
6115 screw up phi congruence because constants are not
6116 uniquely associated with a single ssa name that can be
6117 looked up. */
6118 if (simplified
6119 && is_gimple_min_invariant (simplified)
6120 && TREE_CODE (lhs) == SSA_NAME)
6122 changed = set_ssa_val_to (lhs, simplified);
6123 goto done;
6125 else if (simplified
6126 && TREE_CODE (simplified) == SSA_NAME
6127 && TREE_CODE (lhs) == SSA_NAME)
6129 changed = visit_copy (lhs, simplified);
6130 goto done;
6133 if ((TREE_CODE (lhs) == SSA_NAME
6134 /* We can substitute SSA_NAMEs that are live over
6135 abnormal edges with their constant value. */
6136 && !(gimple_assign_copy_p (ass)
6137 && is_gimple_min_invariant (rhs1))
6138 && !(simplified
6139 && is_gimple_min_invariant (simplified))
6140 && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs))
6141 /* Stores or copies from SSA_NAMEs that are live over
6142 abnormal edges are a problem. */
6143 || (code == SSA_NAME
6144 && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (rhs1)))
6145 changed = defs_to_varying (ass);
6146 else if (REFERENCE_CLASS_P (lhs)
6147 || DECL_P (lhs))
6148 changed = visit_reference_op_store (lhs, rhs1, ass);
6149 else if (TREE_CODE (lhs) == SSA_NAME)
6151 if ((gimple_assign_copy_p (ass)
6152 && is_gimple_min_invariant (rhs1))
6153 || (simplified
6154 && is_gimple_min_invariant (simplified)))
6156 if (simplified)
6157 changed = set_ssa_val_to (lhs, simplified);
6158 else
6159 changed = set_ssa_val_to (lhs, rhs1);
6161 else
6163 /* Visit the original statement. */
6164 switch (vn_get_stmt_kind (ass))
6166 case VN_NARY:
6167 changed = visit_nary_op (lhs, ass);
6168 break;
6169 case VN_REFERENCE:
6170 changed = visit_reference_op_load (lhs, rhs1, ass);
6171 break;
6172 default:
6173 changed = defs_to_varying (ass);
6174 break;
6178 else
6179 changed = defs_to_varying (ass);
6181 else if (gcall *call_stmt = dyn_cast <gcall *> (stmt))
6183 tree lhs = gimple_call_lhs (call_stmt);
6184 if (lhs && TREE_CODE (lhs) == SSA_NAME)
6186 /* Try constant folding based on our current lattice. */
6187 tree simplified = gimple_fold_stmt_to_constant_1 (call_stmt,
6188 vn_valueize);
6189 if (simplified)
6191 if (dump_file && (dump_flags & TDF_DETAILS))
6193 fprintf (dump_file, "call ");
6194 print_gimple_expr (dump_file, call_stmt, 0);
6195 fprintf (dump_file, " simplified to ");
6196 print_generic_expr (dump_file, simplified);
6197 fprintf (dump_file, "\n");
6200 /* Setting value numbers to constants will occasionally
6201 screw up phi congruence because constants are not
6202 uniquely associated with a single ssa name that can be
6203 looked up. */
6204 if (simplified
6205 && is_gimple_min_invariant (simplified))
6207 changed = set_ssa_val_to (lhs, simplified);
6208 if (gimple_vdef (call_stmt))
6209 changed |= set_ssa_val_to (gimple_vdef (call_stmt),
6210 SSA_VAL (gimple_vuse (call_stmt)));
6211 goto done;
6213 else if (simplified
6214 && TREE_CODE (simplified) == SSA_NAME)
6216 changed = visit_copy (lhs, simplified);
6217 if (gimple_vdef (call_stmt))
6218 changed |= set_ssa_val_to (gimple_vdef (call_stmt),
6219 SSA_VAL (gimple_vuse (call_stmt)));
6220 goto done;
6222 else if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs))
6224 changed = defs_to_varying (call_stmt);
6225 goto done;
6229 /* Pick up flags from a devirtualization target. */
6230 tree fn = gimple_call_fn (stmt);
6231 int extra_fnflags = 0;
6232 if (fn && TREE_CODE (fn) == SSA_NAME)
6234 fn = SSA_VAL (fn);
6235 if (TREE_CODE (fn) == ADDR_EXPR
6236 && TREE_CODE (TREE_OPERAND (fn, 0)) == FUNCTION_DECL)
6237 extra_fnflags = flags_from_decl_or_type (TREE_OPERAND (fn, 0));
6239 if ((/* Calls to the same function with the same vuse
6240 and the same operands do not necessarily return the same
6241 value, unless they're pure or const. */
6242 ((gimple_call_flags (call_stmt) | extra_fnflags)
6243 & (ECF_PURE | ECF_CONST))
6244 /* If calls have a vdef, subsequent calls won't have
6245 the same incoming vuse. So, if 2 calls with vdef have the
6246 same vuse, we know they're not subsequent.
6247 We can value number 2 calls to the same function with the
6248 same vuse and the same operands which are not subsequent
6249 the same, because there is no code in the program that can
6250 compare the 2 values... */
6251 || (gimple_vdef (call_stmt)
6252 /* ... unless the call returns a pointer which does
6253 not alias with anything else. In which case the
6254 information that the values are distinct are encoded
6255 in the IL. */
6256 && !(gimple_call_return_flags (call_stmt) & ERF_NOALIAS)
6257 /* Only perform the following when being called from PRE
6258 which embeds tail merging. */
6259 && default_vn_walk_kind == VN_WALK))
6260 /* Do not process .DEFERRED_INIT since that confuses uninit
6261 analysis. */
6262 && !gimple_call_internal_p (call_stmt, IFN_DEFERRED_INIT))
6263 changed = visit_reference_op_call (lhs, call_stmt);
6264 else
6265 changed = defs_to_varying (call_stmt);
6267 else
6268 changed = defs_to_varying (stmt);
6269 done:
6270 return changed;
6274 /* Allocate a value number table. */
6276 static void
6277 allocate_vn_table (vn_tables_t table, unsigned size)
6279 table->phis = new vn_phi_table_type (size);
6280 table->nary = new vn_nary_op_table_type (size);
6281 table->references = new vn_reference_table_type (size);
6284 /* Free a value number table. */
6286 static void
6287 free_vn_table (vn_tables_t table)
6289 /* Walk over elements and release vectors. */
6290 vn_reference_iterator_type hir;
6291 vn_reference_t vr;
6292 FOR_EACH_HASH_TABLE_ELEMENT (*table->references, vr, vn_reference_t, hir)
6293 vr->operands.release ();
6294 delete table->phis;
6295 table->phis = NULL;
6296 delete table->nary;
6297 table->nary = NULL;
6298 delete table->references;
6299 table->references = NULL;
6302 /* Set *ID according to RESULT. */
6304 static void
6305 set_value_id_for_result (tree result, unsigned int *id)
6307 if (result && TREE_CODE (result) == SSA_NAME)
6308 *id = VN_INFO (result)->value_id;
6309 else if (result && is_gimple_min_invariant (result))
6310 *id = get_or_alloc_constant_value_id (result);
6311 else
6312 *id = get_next_value_id ();
6315 /* Set the value ids in the valid hash tables. */
6317 static void
6318 set_hashtable_value_ids (void)
6320 vn_nary_op_iterator_type hin;
6321 vn_phi_iterator_type hip;
6322 vn_reference_iterator_type hir;
6323 vn_nary_op_t vno;
6324 vn_reference_t vr;
6325 vn_phi_t vp;
6327 /* Now set the value ids of the things we had put in the hash
6328 table. */
6330 FOR_EACH_HASH_TABLE_ELEMENT (*valid_info->nary, vno, vn_nary_op_t, hin)
6331 if (! vno->predicated_values)
6332 set_value_id_for_result (vno->u.result, &vno->value_id);
6334 FOR_EACH_HASH_TABLE_ELEMENT (*valid_info->phis, vp, vn_phi_t, hip)
6335 set_value_id_for_result (vp->result, &vp->value_id);
6337 FOR_EACH_HASH_TABLE_ELEMENT (*valid_info->references, vr, vn_reference_t,
6338 hir)
6339 set_value_id_for_result (vr->result, &vr->value_id);
6342 /* Return the maximum value id we have ever seen. */
6344 unsigned int
6345 get_max_value_id (void)
6347 return next_value_id;
6350 /* Return the maximum constant value id we have ever seen. */
6352 unsigned int
6353 get_max_constant_value_id (void)
6355 return -next_constant_value_id;
6358 /* Return the next unique value id. */
6360 unsigned int
6361 get_next_value_id (void)
6363 gcc_checking_assert ((int)next_value_id > 0);
6364 return next_value_id++;
6367 /* Return the next unique value id for constants. */
6369 unsigned int
6370 get_next_constant_value_id (void)
6372 gcc_checking_assert (next_constant_value_id < 0);
6373 return next_constant_value_id--;
6377 /* Compare two expressions E1 and E2 and return true if they are equal.
6378 If match_vn_top_optimistically is true then VN_TOP is equal to anything,
6379 otherwise VN_TOP only matches VN_TOP. */
6381 bool
6382 expressions_equal_p (tree e1, tree e2, bool match_vn_top_optimistically)
6384 /* The obvious case. */
6385 if (e1 == e2)
6386 return true;
6388 /* If either one is VN_TOP consider them equal. */
6389 if (match_vn_top_optimistically
6390 && (e1 == VN_TOP || e2 == VN_TOP))
6391 return true;
6393 /* SSA_NAME compare pointer equal. */
6394 if (TREE_CODE (e1) == SSA_NAME || TREE_CODE (e2) == SSA_NAME)
6395 return false;
6397 /* Now perform the actual comparison. */
6398 if (TREE_CODE (e1) == TREE_CODE (e2)
6399 && operand_equal_p (e1, e2, OEP_PURE_SAME))
6400 return true;
6402 return false;
6406 /* Return true if the nary operation NARY may trap. This is a copy
6407 of stmt_could_throw_1_p adjusted to the SCCVN IL. */
6409 bool
6410 vn_nary_may_trap (vn_nary_op_t nary)
6412 tree type;
6413 tree rhs2 = NULL_TREE;
6414 bool honor_nans = false;
6415 bool honor_snans = false;
6416 bool fp_operation = false;
6417 bool honor_trapv = false;
6418 bool handled, ret;
6419 unsigned i;
6421 if (TREE_CODE_CLASS (nary->opcode) == tcc_comparison
6422 || TREE_CODE_CLASS (nary->opcode) == tcc_unary
6423 || TREE_CODE_CLASS (nary->opcode) == tcc_binary)
6425 type = nary->type;
6426 fp_operation = FLOAT_TYPE_P (type);
6427 if (fp_operation)
6429 honor_nans = flag_trapping_math && !flag_finite_math_only;
6430 honor_snans = flag_signaling_nans != 0;
6432 else if (INTEGRAL_TYPE_P (type) && TYPE_OVERFLOW_TRAPS (type))
6433 honor_trapv = true;
6435 if (nary->length >= 2)
6436 rhs2 = nary->op[1];
6437 ret = operation_could_trap_helper_p (nary->opcode, fp_operation,
6438 honor_trapv, honor_nans, honor_snans,
6439 rhs2, &handled);
6440 if (handled && ret)
6441 return true;
6443 for (i = 0; i < nary->length; ++i)
6444 if (tree_could_trap_p (nary->op[i]))
6445 return true;
6447 return false;
6450 /* Return true if the reference operation REF may trap. */
6452 bool
6453 vn_reference_may_trap (vn_reference_t ref)
6455 switch (ref->operands[0].opcode)
6457 case MODIFY_EXPR:
6458 case CALL_EXPR:
6459 /* We do not handle calls. */
6460 return true;
6461 case ADDR_EXPR:
6462 /* And toplevel address computations never trap. */
6463 return false;
6464 default:;
6467 vn_reference_op_t op;
6468 unsigned i;
6469 FOR_EACH_VEC_ELT (ref->operands, i, op)
6471 switch (op->opcode)
6473 case WITH_SIZE_EXPR:
6474 case TARGET_MEM_REF:
6475 /* Always variable. */
6476 return true;
6477 case COMPONENT_REF:
6478 if (op->op1 && TREE_CODE (op->op1) == SSA_NAME)
6479 return true;
6480 break;
6481 case ARRAY_RANGE_REF:
6482 if (TREE_CODE (op->op0) == SSA_NAME)
6483 return true;
6484 break;
6485 case ARRAY_REF:
6487 if (TREE_CODE (op->op0) != INTEGER_CST)
6488 return true;
6490 /* !in_array_bounds */
6491 tree domain_type = TYPE_DOMAIN (ref->operands[i+1].type);
6492 if (!domain_type)
6493 return true;
6495 tree min = op->op1;
6496 tree max = TYPE_MAX_VALUE (domain_type);
6497 if (!min
6498 || !max
6499 || TREE_CODE (min) != INTEGER_CST
6500 || TREE_CODE (max) != INTEGER_CST)
6501 return true;
6503 if (tree_int_cst_lt (op->op0, min)
6504 || tree_int_cst_lt (max, op->op0))
6505 return true;
6507 break;
6509 case MEM_REF:
6510 /* Nothing interesting in itself, the base is separate. */
6511 break;
6512 /* The following are the address bases. */
6513 case SSA_NAME:
6514 return true;
6515 case ADDR_EXPR:
6516 if (op->op0)
6517 return tree_could_trap_p (TREE_OPERAND (op->op0, 0));
6518 return false;
6519 default:;
6522 return false;
6525 eliminate_dom_walker::eliminate_dom_walker (cdi_direction direction,
6526 bitmap inserted_exprs_)
6527 : dom_walker (direction), do_pre (inserted_exprs_ != NULL),
6528 el_todo (0), eliminations (0), insertions (0),
6529 inserted_exprs (inserted_exprs_)
6531 need_eh_cleanup = BITMAP_ALLOC (NULL);
6532 need_ab_cleanup = BITMAP_ALLOC (NULL);
6535 eliminate_dom_walker::~eliminate_dom_walker ()
6537 BITMAP_FREE (need_eh_cleanup);
6538 BITMAP_FREE (need_ab_cleanup);
6541 /* Return a leader for OP that is available at the current point of the
6542 eliminate domwalk. */
6544 tree
6545 eliminate_dom_walker::eliminate_avail (basic_block, tree op)
6547 tree valnum = VN_INFO (op)->valnum;
6548 if (TREE_CODE (valnum) == SSA_NAME)
6550 if (SSA_NAME_IS_DEFAULT_DEF (valnum))
6551 return valnum;
6552 if (avail.length () > SSA_NAME_VERSION (valnum))
6553 return avail[SSA_NAME_VERSION (valnum)];
6555 else if (is_gimple_min_invariant (valnum))
6556 return valnum;
6557 return NULL_TREE;
6560 /* At the current point of the eliminate domwalk make OP available. */
6562 void
6563 eliminate_dom_walker::eliminate_push_avail (basic_block, tree op)
6565 tree valnum = VN_INFO (op)->valnum;
6566 if (TREE_CODE (valnum) == SSA_NAME)
6568 if (avail.length () <= SSA_NAME_VERSION (valnum))
6569 avail.safe_grow_cleared (SSA_NAME_VERSION (valnum) + 1, true);
6570 tree pushop = op;
6571 if (avail[SSA_NAME_VERSION (valnum)])
6572 pushop = avail[SSA_NAME_VERSION (valnum)];
6573 avail_stack.safe_push (pushop);
6574 avail[SSA_NAME_VERSION (valnum)] = op;
6578 /* Insert the expression recorded by SCCVN for VAL at *GSI. Returns
6579 the leader for the expression if insertion was successful. */
6581 tree
6582 eliminate_dom_walker::eliminate_insert (basic_block bb,
6583 gimple_stmt_iterator *gsi, tree val)
6585 /* We can insert a sequence with a single assignment only. */
6586 gimple_seq stmts = VN_INFO (val)->expr;
6587 if (!gimple_seq_singleton_p (stmts))
6588 return NULL_TREE;
6589 gassign *stmt = dyn_cast <gassign *> (gimple_seq_first_stmt (stmts));
6590 if (!stmt
6591 || (!CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt))
6592 && gimple_assign_rhs_code (stmt) != VIEW_CONVERT_EXPR
6593 && gimple_assign_rhs_code (stmt) != NEGATE_EXPR
6594 && gimple_assign_rhs_code (stmt) != BIT_FIELD_REF
6595 && (gimple_assign_rhs_code (stmt) != BIT_AND_EXPR
6596 || TREE_CODE (gimple_assign_rhs2 (stmt)) != INTEGER_CST)))
6597 return NULL_TREE;
6599 tree op = gimple_assign_rhs1 (stmt);
6600 if (gimple_assign_rhs_code (stmt) == VIEW_CONVERT_EXPR
6601 || gimple_assign_rhs_code (stmt) == BIT_FIELD_REF)
6602 op = TREE_OPERAND (op, 0);
6603 tree leader = TREE_CODE (op) == SSA_NAME ? eliminate_avail (bb, op) : op;
6604 if (!leader)
6605 return NULL_TREE;
6607 tree res;
6608 stmts = NULL;
6609 if (gimple_assign_rhs_code (stmt) == BIT_FIELD_REF)
6610 res = gimple_build (&stmts, BIT_FIELD_REF,
6611 TREE_TYPE (val), leader,
6612 TREE_OPERAND (gimple_assign_rhs1 (stmt), 1),
6613 TREE_OPERAND (gimple_assign_rhs1 (stmt), 2));
6614 else if (gimple_assign_rhs_code (stmt) == BIT_AND_EXPR)
6615 res = gimple_build (&stmts, BIT_AND_EXPR,
6616 TREE_TYPE (val), leader, gimple_assign_rhs2 (stmt));
6617 else
6618 res = gimple_build (&stmts, gimple_assign_rhs_code (stmt),
6619 TREE_TYPE (val), leader);
6620 if (TREE_CODE (res) != SSA_NAME
6621 || SSA_NAME_IS_DEFAULT_DEF (res)
6622 || gimple_bb (SSA_NAME_DEF_STMT (res)))
6624 gimple_seq_discard (stmts);
6626 /* During propagation we have to treat SSA info conservatively
6627 and thus we can end up simplifying the inserted expression
6628 at elimination time to sth not defined in stmts. */
6629 /* But then this is a redundancy we failed to detect. Which means
6630 res now has two values. That doesn't play well with how
6631 we track availability here, so give up. */
6632 if (dump_file && (dump_flags & TDF_DETAILS))
6634 if (TREE_CODE (res) == SSA_NAME)
6635 res = eliminate_avail (bb, res);
6636 if (res)
6638 fprintf (dump_file, "Failed to insert expression for value ");
6639 print_generic_expr (dump_file, val);
6640 fprintf (dump_file, " which is really fully redundant to ");
6641 print_generic_expr (dump_file, res);
6642 fprintf (dump_file, "\n");
6646 return NULL_TREE;
6648 else
6650 gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);
6651 vn_ssa_aux_t vn_info = VN_INFO (res);
6652 vn_info->valnum = val;
6653 vn_info->visited = true;
6656 insertions++;
6657 if (dump_file && (dump_flags & TDF_DETAILS))
6659 fprintf (dump_file, "Inserted ");
6660 print_gimple_stmt (dump_file, SSA_NAME_DEF_STMT (res), 0);
6663 return res;
6666 void
6667 eliminate_dom_walker::eliminate_stmt (basic_block b, gimple_stmt_iterator *gsi)
6669 tree sprime = NULL_TREE;
6670 gimple *stmt = gsi_stmt (*gsi);
6671 tree lhs = gimple_get_lhs (stmt);
6672 if (lhs && TREE_CODE (lhs) == SSA_NAME
6673 && !gimple_has_volatile_ops (stmt)
6674 /* See PR43491. Do not replace a global register variable when
6675 it is a the RHS of an assignment. Do replace local register
6676 variables since gcc does not guarantee a local variable will
6677 be allocated in register.
6678 ??? The fix isn't effective here. This should instead
6679 be ensured by not value-numbering them the same but treating
6680 them like volatiles? */
6681 && !(gimple_assign_single_p (stmt)
6682 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == VAR_DECL
6683 && DECL_HARD_REGISTER (gimple_assign_rhs1 (stmt))
6684 && is_global_var (gimple_assign_rhs1 (stmt)))))
6686 sprime = eliminate_avail (b, lhs);
6687 if (!sprime)
6689 /* If there is no existing usable leader but SCCVN thinks
6690 it has an expression it wants to use as replacement,
6691 insert that. */
6692 tree val = VN_INFO (lhs)->valnum;
6693 vn_ssa_aux_t vn_info;
6694 if (val != VN_TOP
6695 && TREE_CODE (val) == SSA_NAME
6696 && (vn_info = VN_INFO (val), true)
6697 && vn_info->needs_insertion
6698 && vn_info->expr != NULL
6699 && (sprime = eliminate_insert (b, gsi, val)) != NULL_TREE)
6700 eliminate_push_avail (b, sprime);
6703 /* If this now constitutes a copy duplicate points-to
6704 and range info appropriately. This is especially
6705 important for inserted code. See tree-ssa-copy.cc
6706 for similar code. */
6707 if (sprime
6708 && TREE_CODE (sprime) == SSA_NAME)
6710 basic_block sprime_b = gimple_bb (SSA_NAME_DEF_STMT (sprime));
6711 if (POINTER_TYPE_P (TREE_TYPE (lhs))
6712 && SSA_NAME_PTR_INFO (lhs)
6713 && ! SSA_NAME_PTR_INFO (sprime))
6715 duplicate_ssa_name_ptr_info (sprime,
6716 SSA_NAME_PTR_INFO (lhs));
6717 if (b != sprime_b)
6718 reset_flow_sensitive_info (sprime);
6720 else if (INTEGRAL_TYPE_P (TREE_TYPE (lhs))
6721 && SSA_NAME_RANGE_INFO (lhs)
6722 && ! SSA_NAME_RANGE_INFO (sprime)
6723 && b == sprime_b)
6724 duplicate_ssa_name_range_info (sprime, lhs);
6727 /* Inhibit the use of an inserted PHI on a loop header when
6728 the address of the memory reference is a simple induction
6729 variable. In other cases the vectorizer won't do anything
6730 anyway (either it's loop invariant or a complicated
6731 expression). */
6732 if (sprime
6733 && TREE_CODE (sprime) == SSA_NAME
6734 && do_pre
6735 && (flag_tree_loop_vectorize || flag_tree_parallelize_loops > 1)
6736 && loop_outer (b->loop_father)
6737 && has_zero_uses (sprime)
6738 && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (sprime))
6739 && gimple_assign_load_p (stmt))
6741 gimple *def_stmt = SSA_NAME_DEF_STMT (sprime);
6742 basic_block def_bb = gimple_bb (def_stmt);
6743 if (gimple_code (def_stmt) == GIMPLE_PHI
6744 && def_bb->loop_father->header == def_bb)
6746 loop_p loop = def_bb->loop_father;
6747 ssa_op_iter iter;
6748 tree op;
6749 bool found = false;
6750 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
6752 affine_iv iv;
6753 def_bb = gimple_bb (SSA_NAME_DEF_STMT (op));
6754 if (def_bb
6755 && flow_bb_inside_loop_p (loop, def_bb)
6756 && simple_iv (loop, loop, op, &iv, true))
6758 found = true;
6759 break;
6762 if (found)
6764 if (dump_file && (dump_flags & TDF_DETAILS))
6766 fprintf (dump_file, "Not replacing ");
6767 print_gimple_expr (dump_file, stmt, 0);
6768 fprintf (dump_file, " with ");
6769 print_generic_expr (dump_file, sprime);
6770 fprintf (dump_file, " which would add a loop"
6771 " carried dependence to loop %d\n",
6772 loop->num);
6774 /* Don't keep sprime available. */
6775 sprime = NULL_TREE;
6780 if (sprime)
6782 /* If we can propagate the value computed for LHS into
6783 all uses don't bother doing anything with this stmt. */
6784 if (may_propagate_copy (lhs, sprime))
6786 /* Mark it for removal. */
6787 to_remove.safe_push (stmt);
6789 /* ??? Don't count copy/constant propagations. */
6790 if (gimple_assign_single_p (stmt)
6791 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
6792 || gimple_assign_rhs1 (stmt) == sprime))
6793 return;
6795 if (dump_file && (dump_flags & TDF_DETAILS))
6797 fprintf (dump_file, "Replaced ");
6798 print_gimple_expr (dump_file, stmt, 0);
6799 fprintf (dump_file, " with ");
6800 print_generic_expr (dump_file, sprime);
6801 fprintf (dump_file, " in all uses of ");
6802 print_gimple_stmt (dump_file, stmt, 0);
6805 eliminations++;
6806 return;
6809 /* If this is an assignment from our leader (which
6810 happens in the case the value-number is a constant)
6811 then there is nothing to do. Likewise if we run into
6812 inserted code that needed a conversion because of
6813 our type-agnostic value-numbering of loads. */
6814 if ((gimple_assign_single_p (stmt)
6815 || (is_gimple_assign (stmt)
6816 && (CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt))
6817 || gimple_assign_rhs_code (stmt) == VIEW_CONVERT_EXPR)))
6818 && sprime == gimple_assign_rhs1 (stmt))
6819 return;
6821 /* Else replace its RHS. */
6822 if (dump_file && (dump_flags & TDF_DETAILS))
6824 fprintf (dump_file, "Replaced ");
6825 print_gimple_expr (dump_file, stmt, 0);
6826 fprintf (dump_file, " with ");
6827 print_generic_expr (dump_file, sprime);
6828 fprintf (dump_file, " in ");
6829 print_gimple_stmt (dump_file, stmt, 0);
6831 eliminations++;
6833 bool can_make_abnormal_goto = (is_gimple_call (stmt)
6834 && stmt_can_make_abnormal_goto (stmt));
6835 gimple *orig_stmt = stmt;
6836 if (!useless_type_conversion_p (TREE_TYPE (lhs),
6837 TREE_TYPE (sprime)))
6839 /* We preserve conversions to but not from function or method
6840 types. This asymmetry makes it necessary to re-instantiate
6841 conversions here. */
6842 if (POINTER_TYPE_P (TREE_TYPE (lhs))
6843 && FUNC_OR_METHOD_TYPE_P (TREE_TYPE (TREE_TYPE (lhs))))
6844 sprime = fold_convert (TREE_TYPE (lhs), sprime);
6845 else
6846 gcc_unreachable ();
6848 tree vdef = gimple_vdef (stmt);
6849 tree vuse = gimple_vuse (stmt);
6850 propagate_tree_value_into_stmt (gsi, sprime);
6851 stmt = gsi_stmt (*gsi);
6852 update_stmt (stmt);
6853 /* In case the VDEF on the original stmt was released, value-number
6854 it to the VUSE. This is to make vuse_ssa_val able to skip
6855 released virtual operands. */
6856 if (vdef != gimple_vdef (stmt))
6858 gcc_assert (SSA_NAME_IN_FREE_LIST (vdef));
6859 VN_INFO (vdef)->valnum = vuse;
6862 /* If we removed EH side-effects from the statement, clean
6863 its EH information. */
6864 if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
6866 bitmap_set_bit (need_eh_cleanup,
6867 gimple_bb (stmt)->index);
6868 if (dump_file && (dump_flags & TDF_DETAILS))
6869 fprintf (dump_file, " Removed EH side-effects.\n");
6872 /* Likewise for AB side-effects. */
6873 if (can_make_abnormal_goto
6874 && !stmt_can_make_abnormal_goto (stmt))
6876 bitmap_set_bit (need_ab_cleanup,
6877 gimple_bb (stmt)->index);
6878 if (dump_file && (dump_flags & TDF_DETAILS))
6879 fprintf (dump_file, " Removed AB side-effects.\n");
6882 return;
6886 /* If the statement is a scalar store, see if the expression
6887 has the same value number as its rhs. If so, the store is
6888 dead. */
6889 if (gimple_assign_single_p (stmt)
6890 && !gimple_has_volatile_ops (stmt)
6891 && !is_gimple_reg (gimple_assign_lhs (stmt))
6892 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
6893 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
6895 tree rhs = gimple_assign_rhs1 (stmt);
6896 vn_reference_t vnresult;
6897 /* ??? gcc.dg/torture/pr91445.c shows that we lookup a boolean
6898 typed load of a byte known to be 0x11 as 1 so a store of
6899 a boolean 1 is detected as redundant. Because of this we
6900 have to make sure to lookup with a ref where its size
6901 matches the precision. */
6902 tree lookup_lhs = lhs;
6903 if (INTEGRAL_TYPE_P (TREE_TYPE (lhs))
6904 && (TREE_CODE (lhs) != COMPONENT_REF
6905 || !DECL_BIT_FIELD_TYPE (TREE_OPERAND (lhs, 1)))
6906 && !type_has_mode_precision_p (TREE_TYPE (lhs)))
6908 if (TREE_CODE (lhs) == COMPONENT_REF
6909 || TREE_CODE (lhs) == MEM_REF)
6911 tree ltype = build_nonstandard_integer_type
6912 (TREE_INT_CST_LOW (TYPE_SIZE (TREE_TYPE (lhs))),
6913 TYPE_UNSIGNED (TREE_TYPE (lhs)));
6914 if (TREE_CODE (lhs) == COMPONENT_REF)
6916 tree foff = component_ref_field_offset (lhs);
6917 tree f = TREE_OPERAND (lhs, 1);
6918 if (!poly_int_tree_p (foff))
6919 lookup_lhs = NULL_TREE;
6920 else
6921 lookup_lhs = build3 (BIT_FIELD_REF, ltype,
6922 TREE_OPERAND (lhs, 0),
6923 TYPE_SIZE (TREE_TYPE (lhs)),
6924 bit_from_pos
6925 (foff, DECL_FIELD_BIT_OFFSET (f)));
6927 else
6928 lookup_lhs = build2 (MEM_REF, ltype,
6929 TREE_OPERAND (lhs, 0),
6930 TREE_OPERAND (lhs, 1));
6932 else
6933 lookup_lhs = NULL_TREE;
6935 tree val = NULL_TREE;
6936 if (lookup_lhs)
6937 val = vn_reference_lookup (lookup_lhs, gimple_vuse (stmt),
6938 VN_WALKREWRITE, &vnresult, false,
6939 NULL, NULL_TREE, true);
6940 if (TREE_CODE (rhs) == SSA_NAME)
6941 rhs = VN_INFO (rhs)->valnum;
6942 if (val
6943 && (operand_equal_p (val, rhs, 0)
6944 /* Due to the bitfield lookups above we can get bit
6945 interpretations of the same RHS as values here. Those
6946 are redundant as well. */
6947 || (TREE_CODE (val) == SSA_NAME
6948 && gimple_assign_single_p (SSA_NAME_DEF_STMT (val))
6949 && (val = gimple_assign_rhs1 (SSA_NAME_DEF_STMT (val)))
6950 && TREE_CODE (val) == VIEW_CONVERT_EXPR
6951 && TREE_OPERAND (val, 0) == rhs)))
6953 /* We can only remove the later store if the former aliases
6954 at least all accesses the later one does or if the store
6955 was to readonly memory storing the same value. */
6956 ao_ref lhs_ref;
6957 ao_ref_init (&lhs_ref, lhs);
6958 alias_set_type set = ao_ref_alias_set (&lhs_ref);
6959 alias_set_type base_set = ao_ref_base_alias_set (&lhs_ref);
6960 if (! vnresult
6961 || ((vnresult->set == set
6962 || alias_set_subset_of (set, vnresult->set))
6963 && (vnresult->base_set == base_set
6964 || alias_set_subset_of (base_set, vnresult->base_set))))
6966 if (dump_file && (dump_flags & TDF_DETAILS))
6968 fprintf (dump_file, "Deleted redundant store ");
6969 print_gimple_stmt (dump_file, stmt, 0);
6972 /* Queue stmt for removal. */
6973 to_remove.safe_push (stmt);
6974 return;
6979 /* If this is a control statement value numbering left edges
6980 unexecuted on force the condition in a way consistent with
6981 that. */
6982 if (gcond *cond = dyn_cast <gcond *> (stmt))
6984 if ((EDGE_SUCC (b, 0)->flags & EDGE_EXECUTABLE)
6985 ^ (EDGE_SUCC (b, 1)->flags & EDGE_EXECUTABLE))
6987 if (dump_file && (dump_flags & TDF_DETAILS))
6989 fprintf (dump_file, "Removing unexecutable edge from ");
6990 print_gimple_stmt (dump_file, stmt, 0);
6992 if (((EDGE_SUCC (b, 0)->flags & EDGE_TRUE_VALUE) != 0)
6993 == ((EDGE_SUCC (b, 0)->flags & EDGE_EXECUTABLE) != 0))
6994 gimple_cond_make_true (cond);
6995 else
6996 gimple_cond_make_false (cond);
6997 update_stmt (cond);
6998 el_todo |= TODO_cleanup_cfg;
6999 return;
7003 bool can_make_abnormal_goto = stmt_can_make_abnormal_goto (stmt);
7004 bool was_noreturn = (is_gimple_call (stmt)
7005 && gimple_call_noreturn_p (stmt));
7006 tree vdef = gimple_vdef (stmt);
7007 tree vuse = gimple_vuse (stmt);
7009 /* If we didn't replace the whole stmt (or propagate the result
7010 into all uses), replace all uses on this stmt with their
7011 leaders. */
7012 bool modified = false;
7013 use_operand_p use_p;
7014 ssa_op_iter iter;
7015 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE)
7017 tree use = USE_FROM_PTR (use_p);
7018 /* ??? The call code above leaves stmt operands un-updated. */
7019 if (TREE_CODE (use) != SSA_NAME)
7020 continue;
7021 tree sprime;
7022 if (SSA_NAME_IS_DEFAULT_DEF (use))
7023 /* ??? For default defs BB shouldn't matter, but we have to
7024 solve the inconsistency between rpo eliminate and
7025 dom eliminate avail valueization first. */
7026 sprime = eliminate_avail (b, use);
7027 else
7028 /* Look for sth available at the definition block of the argument.
7029 This avoids inconsistencies between availability there which
7030 decides if the stmt can be removed and availability at the
7031 use site. The SSA property ensures that things available
7032 at the definition are also available at uses. */
7033 sprime = eliminate_avail (gimple_bb (SSA_NAME_DEF_STMT (use)), use);
7034 if (sprime && sprime != use
7035 && may_propagate_copy (use, sprime, true)
7036 /* We substitute into debug stmts to avoid excessive
7037 debug temporaries created by removed stmts, but we need
7038 to avoid doing so for inserted sprimes as we never want
7039 to create debug temporaries for them. */
7040 && (!inserted_exprs
7041 || TREE_CODE (sprime) != SSA_NAME
7042 || !is_gimple_debug (stmt)
7043 || !bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (sprime))))
7045 propagate_value (use_p, sprime);
7046 modified = true;
7050 /* Fold the stmt if modified, this canonicalizes MEM_REFs we propagated
7051 into which is a requirement for the IPA devirt machinery. */
7052 gimple *old_stmt = stmt;
7053 if (modified)
7055 /* If a formerly non-invariant ADDR_EXPR is turned into an
7056 invariant one it was on a separate stmt. */
7057 if (gimple_assign_single_p (stmt)
7058 && TREE_CODE (gimple_assign_rhs1 (stmt)) == ADDR_EXPR)
7059 recompute_tree_invariant_for_addr_expr (gimple_assign_rhs1 (stmt));
7060 gimple_stmt_iterator prev = *gsi;
7061 gsi_prev (&prev);
7062 if (fold_stmt (gsi, follow_all_ssa_edges))
7064 /* fold_stmt may have created new stmts inbetween
7065 the previous stmt and the folded stmt. Mark
7066 all defs created there as varying to not confuse
7067 the SCCVN machinery as we're using that even during
7068 elimination. */
7069 if (gsi_end_p (prev))
7070 prev = gsi_start_bb (b);
7071 else
7072 gsi_next (&prev);
7073 if (gsi_stmt (prev) != gsi_stmt (*gsi))
7076 tree def;
7077 ssa_op_iter dit;
7078 FOR_EACH_SSA_TREE_OPERAND (def, gsi_stmt (prev),
7079 dit, SSA_OP_ALL_DEFS)
7080 /* As existing DEFs may move between stmts
7081 only process new ones. */
7082 if (! has_VN_INFO (def))
7084 vn_ssa_aux_t vn_info = VN_INFO (def);
7085 vn_info->valnum = def;
7086 vn_info->visited = true;
7088 if (gsi_stmt (prev) == gsi_stmt (*gsi))
7089 break;
7090 gsi_next (&prev);
7092 while (1);
7094 stmt = gsi_stmt (*gsi);
7095 /* In case we folded the stmt away schedule the NOP for removal. */
7096 if (gimple_nop_p (stmt))
7097 to_remove.safe_push (stmt);
7100 /* Visit indirect calls and turn them into direct calls if
7101 possible using the devirtualization machinery. Do this before
7102 checking for required EH/abnormal/noreturn cleanup as devird
7103 may expose more of those. */
7104 if (gcall *call_stmt = dyn_cast <gcall *> (stmt))
7106 tree fn = gimple_call_fn (call_stmt);
7107 if (fn
7108 && flag_devirtualize
7109 && virtual_method_call_p (fn))
7111 tree otr_type = obj_type_ref_class (fn);
7112 unsigned HOST_WIDE_INT otr_tok
7113 = tree_to_uhwi (OBJ_TYPE_REF_TOKEN (fn));
7114 tree instance;
7115 ipa_polymorphic_call_context context (current_function_decl,
7116 fn, stmt, &instance);
7117 context.get_dynamic_type (instance, OBJ_TYPE_REF_OBJECT (fn),
7118 otr_type, stmt, NULL);
7119 bool final;
7120 vec <cgraph_node *> targets
7121 = possible_polymorphic_call_targets (obj_type_ref_class (fn),
7122 otr_tok, context, &final);
7123 if (dump_file)
7124 dump_possible_polymorphic_call_targets (dump_file,
7125 obj_type_ref_class (fn),
7126 otr_tok, context);
7127 if (final && targets.length () <= 1 && dbg_cnt (devirt))
7129 tree fn;
7130 if (targets.length () == 1)
7131 fn = targets[0]->decl;
7132 else
7133 fn = builtin_decl_unreachable ();
7134 if (dump_enabled_p ())
7136 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, stmt,
7137 "converting indirect call to "
7138 "function %s\n",
7139 lang_hooks.decl_printable_name (fn, 2));
7141 gimple_call_set_fndecl (call_stmt, fn);
7142 /* If changing the call to __builtin_unreachable
7143 or similar noreturn function, adjust gimple_call_fntype
7144 too. */
7145 if (gimple_call_noreturn_p (call_stmt)
7146 && VOID_TYPE_P (TREE_TYPE (TREE_TYPE (fn)))
7147 && TYPE_ARG_TYPES (TREE_TYPE (fn))
7148 && (TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fn)))
7149 == void_type_node))
7150 gimple_call_set_fntype (call_stmt, TREE_TYPE (fn));
7151 maybe_remove_unused_call_args (cfun, call_stmt);
7152 modified = true;
7157 if (modified)
7159 /* When changing a call into a noreturn call, cfg cleanup
7160 is needed to fix up the noreturn call. */
7161 if (!was_noreturn
7162 && is_gimple_call (stmt) && gimple_call_noreturn_p (stmt))
7163 to_fixup.safe_push (stmt);
7164 /* When changing a condition or switch into one we know what
7165 edge will be executed, schedule a cfg cleanup. */
7166 if ((gimple_code (stmt) == GIMPLE_COND
7167 && (gimple_cond_true_p (as_a <gcond *> (stmt))
7168 || gimple_cond_false_p (as_a <gcond *> (stmt))))
7169 || (gimple_code (stmt) == GIMPLE_SWITCH
7170 && TREE_CODE (gimple_switch_index
7171 (as_a <gswitch *> (stmt))) == INTEGER_CST))
7172 el_todo |= TODO_cleanup_cfg;
7173 /* If we removed EH side-effects from the statement, clean
7174 its EH information. */
7175 if (maybe_clean_or_replace_eh_stmt (old_stmt, stmt))
7177 bitmap_set_bit (need_eh_cleanup,
7178 gimple_bb (stmt)->index);
7179 if (dump_file && (dump_flags & TDF_DETAILS))
7180 fprintf (dump_file, " Removed EH side-effects.\n");
7182 /* Likewise for AB side-effects. */
7183 if (can_make_abnormal_goto
7184 && !stmt_can_make_abnormal_goto (stmt))
7186 bitmap_set_bit (need_ab_cleanup,
7187 gimple_bb (stmt)->index);
7188 if (dump_file && (dump_flags & TDF_DETAILS))
7189 fprintf (dump_file, " Removed AB side-effects.\n");
7191 update_stmt (stmt);
7192 /* In case the VDEF on the original stmt was released, value-number
7193 it to the VUSE. This is to make vuse_ssa_val able to skip
7194 released virtual operands. */
7195 if (vdef && SSA_NAME_IN_FREE_LIST (vdef))
7196 VN_INFO (vdef)->valnum = vuse;
7199 /* Make new values available - for fully redundant LHS we
7200 continue with the next stmt above and skip this. */
7201 def_operand_p defp;
7202 FOR_EACH_SSA_DEF_OPERAND (defp, stmt, iter, SSA_OP_DEF)
7203 eliminate_push_avail (b, DEF_FROM_PTR (defp));
7206 /* Perform elimination for the basic-block B during the domwalk. */
7208 edge
7209 eliminate_dom_walker::before_dom_children (basic_block b)
7211 /* Mark new bb. */
7212 avail_stack.safe_push (NULL_TREE);
7214 /* Skip unreachable blocks marked unreachable during the SCCVN domwalk. */
7215 if (!(b->flags & BB_EXECUTABLE))
7216 return NULL;
7218 vn_context_bb = b;
7220 for (gphi_iterator gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
7222 gphi *phi = gsi.phi ();
7223 tree res = PHI_RESULT (phi);
7225 if (virtual_operand_p (res))
7227 gsi_next (&gsi);
7228 continue;
7231 tree sprime = eliminate_avail (b, res);
7232 if (sprime
7233 && sprime != res)
7235 if (dump_file && (dump_flags & TDF_DETAILS))
7237 fprintf (dump_file, "Replaced redundant PHI node defining ");
7238 print_generic_expr (dump_file, res);
7239 fprintf (dump_file, " with ");
7240 print_generic_expr (dump_file, sprime);
7241 fprintf (dump_file, "\n");
7244 /* If we inserted this PHI node ourself, it's not an elimination. */
7245 if (! inserted_exprs
7246 || ! bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
7247 eliminations++;
7249 /* If we will propagate into all uses don't bother to do
7250 anything. */
7251 if (may_propagate_copy (res, sprime))
7253 /* Mark the PHI for removal. */
7254 to_remove.safe_push (phi);
7255 gsi_next (&gsi);
7256 continue;
7259 remove_phi_node (&gsi, false);
7261 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
7262 sprime = fold_convert (TREE_TYPE (res), sprime);
7263 gimple *stmt = gimple_build_assign (res, sprime);
7264 gimple_stmt_iterator gsi2 = gsi_after_labels (b);
7265 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
7266 continue;
7269 eliminate_push_avail (b, res);
7270 gsi_next (&gsi);
7273 for (gimple_stmt_iterator gsi = gsi_start_bb (b);
7274 !gsi_end_p (gsi);
7275 gsi_next (&gsi))
7276 eliminate_stmt (b, &gsi);
7278 /* Replace destination PHI arguments. */
7279 edge_iterator ei;
7280 edge e;
7281 FOR_EACH_EDGE (e, ei, b->succs)
7282 if (e->flags & EDGE_EXECUTABLE)
7283 for (gphi_iterator gsi = gsi_start_phis (e->dest);
7284 !gsi_end_p (gsi);
7285 gsi_next (&gsi))
7287 gphi *phi = gsi.phi ();
7288 use_operand_p use_p = PHI_ARG_DEF_PTR_FROM_EDGE (phi, e);
7289 tree arg = USE_FROM_PTR (use_p);
7290 if (TREE_CODE (arg) != SSA_NAME
7291 || virtual_operand_p (arg))
7292 continue;
7293 tree sprime = eliminate_avail (b, arg);
7294 if (sprime && may_propagate_copy (arg, sprime))
7295 propagate_value (use_p, sprime);
7298 vn_context_bb = NULL;
7300 return NULL;
7303 /* Make no longer available leaders no longer available. */
7305 void
7306 eliminate_dom_walker::after_dom_children (basic_block)
7308 tree entry;
7309 while ((entry = avail_stack.pop ()) != NULL_TREE)
7311 tree valnum = VN_INFO (entry)->valnum;
7312 tree old = avail[SSA_NAME_VERSION (valnum)];
7313 if (old == entry)
7314 avail[SSA_NAME_VERSION (valnum)] = NULL_TREE;
7315 else
7316 avail[SSA_NAME_VERSION (valnum)] = entry;
7320 /* Remove queued stmts and perform delayed cleanups. */
7322 unsigned
7323 eliminate_dom_walker::eliminate_cleanup (bool region_p)
7325 statistics_counter_event (cfun, "Eliminated", eliminations);
7326 statistics_counter_event (cfun, "Insertions", insertions);
7328 /* We cannot remove stmts during BB walk, especially not release SSA
7329 names there as this confuses the VN machinery. The stmts ending
7330 up in to_remove are either stores or simple copies.
7331 Remove stmts in reverse order to make debug stmt creation possible. */
7332 while (!to_remove.is_empty ())
7334 bool do_release_defs = true;
7335 gimple *stmt = to_remove.pop ();
7337 /* When we are value-numbering a region we do not require exit PHIs to
7338 be present so we have to make sure to deal with uses outside of the
7339 region of stmts that we thought are eliminated.
7340 ??? Note we may be confused by uses in dead regions we didn't run
7341 elimination on. Rather than checking individual uses we accept
7342 dead copies to be generated here (gcc.c-torture/execute/20060905-1.c
7343 contains such example). */
7344 if (region_p)
7346 if (gphi *phi = dyn_cast <gphi *> (stmt))
7348 tree lhs = gimple_phi_result (phi);
7349 if (!has_zero_uses (lhs))
7351 if (dump_file && (dump_flags & TDF_DETAILS))
7352 fprintf (dump_file, "Keeping eliminated stmt live "
7353 "as copy because of out-of-region uses\n");
7354 tree sprime = eliminate_avail (gimple_bb (stmt), lhs);
7355 gimple *copy = gimple_build_assign (lhs, sprime);
7356 gimple_stmt_iterator gsi
7357 = gsi_after_labels (gimple_bb (stmt));
7358 gsi_insert_before (&gsi, copy, GSI_SAME_STMT);
7359 do_release_defs = false;
7362 else if (tree lhs = gimple_get_lhs (stmt))
7363 if (TREE_CODE (lhs) == SSA_NAME
7364 && !has_zero_uses (lhs))
7366 if (dump_file && (dump_flags & TDF_DETAILS))
7367 fprintf (dump_file, "Keeping eliminated stmt live "
7368 "as copy because of out-of-region uses\n");
7369 tree sprime = eliminate_avail (gimple_bb (stmt), lhs);
7370 gimple_stmt_iterator gsi = gsi_for_stmt (stmt);
7371 if (is_gimple_assign (stmt))
7373 gimple_assign_set_rhs_from_tree (&gsi, sprime);
7374 stmt = gsi_stmt (gsi);
7375 update_stmt (stmt);
7376 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
7377 bitmap_set_bit (need_eh_cleanup, gimple_bb (stmt)->index);
7378 continue;
7380 else
7382 gimple *copy = gimple_build_assign (lhs, sprime);
7383 gsi_insert_before (&gsi, copy, GSI_SAME_STMT);
7384 do_release_defs = false;
7389 if (dump_file && (dump_flags & TDF_DETAILS))
7391 fprintf (dump_file, "Removing dead stmt ");
7392 print_gimple_stmt (dump_file, stmt, 0, TDF_NONE);
7395 gimple_stmt_iterator gsi = gsi_for_stmt (stmt);
7396 if (gimple_code (stmt) == GIMPLE_PHI)
7397 remove_phi_node (&gsi, do_release_defs);
7398 else
7400 basic_block bb = gimple_bb (stmt);
7401 unlink_stmt_vdef (stmt);
7402 if (gsi_remove (&gsi, true))
7403 bitmap_set_bit (need_eh_cleanup, bb->index);
7404 if (is_gimple_call (stmt) && stmt_can_make_abnormal_goto (stmt))
7405 bitmap_set_bit (need_ab_cleanup, bb->index);
7406 if (do_release_defs)
7407 release_defs (stmt);
7410 /* Removing a stmt may expose a forwarder block. */
7411 el_todo |= TODO_cleanup_cfg;
7414 /* Fixup stmts that became noreturn calls. This may require splitting
7415 blocks and thus isn't possible during the dominator walk. Do this
7416 in reverse order so we don't inadvertedly remove a stmt we want to
7417 fixup by visiting a dominating now noreturn call first. */
7418 while (!to_fixup.is_empty ())
7420 gimple *stmt = to_fixup.pop ();
7422 if (dump_file && (dump_flags & TDF_DETAILS))
7424 fprintf (dump_file, "Fixing up noreturn call ");
7425 print_gimple_stmt (dump_file, stmt, 0);
7428 if (fixup_noreturn_call (stmt))
7429 el_todo |= TODO_cleanup_cfg;
7432 bool do_eh_cleanup = !bitmap_empty_p (need_eh_cleanup);
7433 bool do_ab_cleanup = !bitmap_empty_p (need_ab_cleanup);
7435 if (do_eh_cleanup)
7436 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
7438 if (do_ab_cleanup)
7439 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup);
7441 if (do_eh_cleanup || do_ab_cleanup)
7442 el_todo |= TODO_cleanup_cfg;
7444 return el_todo;
7447 /* Eliminate fully redundant computations. */
7449 unsigned
7450 eliminate_with_rpo_vn (bitmap inserted_exprs)
7452 eliminate_dom_walker walker (CDI_DOMINATORS, inserted_exprs);
7454 eliminate_dom_walker *saved_rpo_avail = rpo_avail;
7455 rpo_avail = &walker;
7456 walker.walk (cfun->cfg->x_entry_block_ptr);
7457 rpo_avail = saved_rpo_avail;
7459 return walker.eliminate_cleanup ();
7462 static unsigned
7463 do_rpo_vn_1 (function *fn, edge entry, bitmap exit_bbs,
7464 bool iterate, bool eliminate, vn_lookup_kind kind);
7466 void
7467 run_rpo_vn (vn_lookup_kind kind)
7469 do_rpo_vn_1 (cfun, NULL, NULL, true, false, kind);
7471 /* ??? Prune requirement of these. */
7472 constant_to_value_id = new hash_table<vn_constant_hasher> (23);
7474 /* Initialize the value ids and prune out remaining VN_TOPs
7475 from dead code. */
7476 tree name;
7477 unsigned i;
7478 FOR_EACH_SSA_NAME (i, name, cfun)
7480 vn_ssa_aux_t info = VN_INFO (name);
7481 if (!info->visited
7482 || info->valnum == VN_TOP)
7483 info->valnum = name;
7484 if (info->valnum == name)
7485 info->value_id = get_next_value_id ();
7486 else if (is_gimple_min_invariant (info->valnum))
7487 info->value_id = get_or_alloc_constant_value_id (info->valnum);
7490 /* Propagate. */
7491 FOR_EACH_SSA_NAME (i, name, cfun)
7493 vn_ssa_aux_t info = VN_INFO (name);
7494 if (TREE_CODE (info->valnum) == SSA_NAME
7495 && info->valnum != name
7496 && info->value_id != VN_INFO (info->valnum)->value_id)
7497 info->value_id = VN_INFO (info->valnum)->value_id;
7500 set_hashtable_value_ids ();
7502 if (dump_file && (dump_flags & TDF_DETAILS))
7504 fprintf (dump_file, "Value numbers:\n");
7505 FOR_EACH_SSA_NAME (i, name, cfun)
7507 if (VN_INFO (name)->visited
7508 && SSA_VAL (name) != name)
7510 print_generic_expr (dump_file, name);
7511 fprintf (dump_file, " = ");
7512 print_generic_expr (dump_file, SSA_VAL (name));
7513 fprintf (dump_file, " (%04d)\n", VN_INFO (name)->value_id);
7519 /* Free VN associated data structures. */
7521 void
7522 free_rpo_vn (void)
7524 free_vn_table (valid_info);
7525 XDELETE (valid_info);
7526 obstack_free (&vn_tables_obstack, NULL);
7527 obstack_free (&vn_tables_insert_obstack, NULL);
7529 vn_ssa_aux_iterator_type it;
7530 vn_ssa_aux_t info;
7531 FOR_EACH_HASH_TABLE_ELEMENT (*vn_ssa_aux_hash, info, vn_ssa_aux_t, it)
7532 if (info->needs_insertion)
7533 release_ssa_name (info->name);
7534 obstack_free (&vn_ssa_aux_obstack, NULL);
7535 delete vn_ssa_aux_hash;
7537 delete constant_to_value_id;
7538 constant_to_value_id = NULL;
7541 /* Hook for maybe_push_res_to_seq, lookup the expression in the VN tables. */
7543 static tree
7544 vn_lookup_simplify_result (gimple_match_op *res_op)
7546 if (!res_op->code.is_tree_code ())
7547 return NULL_TREE;
7548 tree *ops = res_op->ops;
7549 unsigned int length = res_op->num_ops;
7550 if (res_op->code == CONSTRUCTOR
7551 /* ??? We're arriving here with SCCVNs view, decomposed CONSTRUCTOR
7552 and GIMPLEs / match-and-simplifies, CONSTRUCTOR as GENERIC tree. */
7553 && TREE_CODE (res_op->ops[0]) == CONSTRUCTOR)
7555 length = CONSTRUCTOR_NELTS (res_op->ops[0]);
7556 ops = XALLOCAVEC (tree, length);
7557 for (unsigned i = 0; i < length; ++i)
7558 ops[i] = CONSTRUCTOR_ELT (res_op->ops[0], i)->value;
7560 vn_nary_op_t vnresult = NULL;
7561 tree res = vn_nary_op_lookup_pieces (length, (tree_code) res_op->code,
7562 res_op->type, ops, &vnresult);
7563 /* If this is used from expression simplification make sure to
7564 return an available expression. */
7565 if (res && TREE_CODE (res) == SSA_NAME && mprts_hook && rpo_avail)
7566 res = rpo_avail->eliminate_avail (vn_context_bb, res);
7567 return res;
7570 /* Return a leader for OPs value that is valid at BB. */
7572 tree
7573 rpo_elim::eliminate_avail (basic_block bb, tree op)
7575 bool visited;
7576 tree valnum = SSA_VAL (op, &visited);
7577 /* If we didn't visit OP then it must be defined outside of the
7578 region we process and also dominate it. So it is available. */
7579 if (!visited)
7580 return op;
7581 if (TREE_CODE (valnum) == SSA_NAME)
7583 if (SSA_NAME_IS_DEFAULT_DEF (valnum))
7584 return valnum;
7585 vn_avail *av = VN_INFO (valnum)->avail;
7586 if (!av)
7587 return NULL_TREE;
7588 if (av->location == bb->index)
7589 /* On tramp3d 90% of the cases are here. */
7590 return ssa_name (av->leader);
7593 basic_block abb = BASIC_BLOCK_FOR_FN (cfun, av->location);
7594 /* ??? During elimination we have to use availability at the
7595 definition site of a use we try to replace. This
7596 is required to not run into inconsistencies because
7597 of dominated_by_p_w_unex behavior and removing a definition
7598 while not replacing all uses.
7599 ??? We could try to consistently walk dominators
7600 ignoring non-executable regions. The nearest common
7601 dominator of bb and abb is where we can stop walking. We
7602 may also be able to "pre-compute" (bits of) the next immediate
7603 (non-)dominator during the RPO walk when marking edges as
7604 executable. */
7605 if (dominated_by_p_w_unex (bb, abb, true))
7607 tree leader = ssa_name (av->leader);
7608 /* Prevent eliminations that break loop-closed SSA. */
7609 if (loops_state_satisfies_p (LOOP_CLOSED_SSA)
7610 && ! SSA_NAME_IS_DEFAULT_DEF (leader)
7611 && ! flow_bb_inside_loop_p (gimple_bb (SSA_NAME_DEF_STMT
7612 (leader))->loop_father,
7613 bb))
7614 return NULL_TREE;
7615 if (dump_file && (dump_flags & TDF_DETAILS))
7617 print_generic_expr (dump_file, leader);
7618 fprintf (dump_file, " is available for ");
7619 print_generic_expr (dump_file, valnum);
7620 fprintf (dump_file, "\n");
7622 /* On tramp3d 99% of the _remaining_ cases succeed at
7623 the first enty. */
7624 return leader;
7626 /* ??? Can we somehow skip to the immediate dominator
7627 RPO index (bb_to_rpo)? Again, maybe not worth, on
7628 tramp3d the worst number of elements in the vector is 9. */
7629 av = av->next;
7631 while (av);
7633 else if (valnum != VN_TOP)
7634 /* valnum is is_gimple_min_invariant. */
7635 return valnum;
7636 return NULL_TREE;
7639 /* Make LEADER a leader for its value at BB. */
7641 void
7642 rpo_elim::eliminate_push_avail (basic_block bb, tree leader)
7644 tree valnum = VN_INFO (leader)->valnum;
7645 if (valnum == VN_TOP
7646 || is_gimple_min_invariant (valnum))
7647 return;
7648 if (dump_file && (dump_flags & TDF_DETAILS))
7650 fprintf (dump_file, "Making available beyond BB%d ", bb->index);
7651 print_generic_expr (dump_file, leader);
7652 fprintf (dump_file, " for value ");
7653 print_generic_expr (dump_file, valnum);
7654 fprintf (dump_file, "\n");
7656 vn_ssa_aux_t value = VN_INFO (valnum);
7657 vn_avail *av;
7658 if (m_avail_freelist)
7660 av = m_avail_freelist;
7661 m_avail_freelist = m_avail_freelist->next;
7663 else
7664 av = XOBNEW (&vn_ssa_aux_obstack, vn_avail);
7665 av->location = bb->index;
7666 av->leader = SSA_NAME_VERSION (leader);
7667 av->next = value->avail;
7668 av->next_undo = last_pushed_avail;
7669 last_pushed_avail = value;
7670 value->avail = av;
7673 /* Valueization hook for RPO VN plus required state. */
7675 tree
7676 rpo_vn_valueize (tree name)
7678 if (TREE_CODE (name) == SSA_NAME)
7680 vn_ssa_aux_t val = VN_INFO (name);
7681 if (val)
7683 tree tem = val->valnum;
7684 if (tem != VN_TOP && tem != name)
7686 if (TREE_CODE (tem) != SSA_NAME)
7687 return tem;
7688 /* For all values we only valueize to an available leader
7689 which means we can use SSA name info without restriction. */
7690 tem = rpo_avail->eliminate_avail (vn_context_bb, tem);
7691 if (tem)
7692 return tem;
7696 return name;
7699 /* Insert on PRED_E predicates derived from CODE OPS being true besides the
7700 inverted condition. */
7702 static void
7703 insert_related_predicates_on_edge (enum tree_code code, tree *ops, edge pred_e)
7705 switch (code)
7707 case LT_EXPR:
7708 /* a < b -> a {!,<}= b */
7709 vn_nary_op_insert_pieces_predicated (2, NE_EXPR, boolean_type_node,
7710 ops, boolean_true_node, 0, pred_e);
7711 vn_nary_op_insert_pieces_predicated (2, LE_EXPR, boolean_type_node,
7712 ops, boolean_true_node, 0, pred_e);
7713 /* a < b -> ! a {>,=} b */
7714 vn_nary_op_insert_pieces_predicated (2, GT_EXPR, boolean_type_node,
7715 ops, boolean_false_node, 0, pred_e);
7716 vn_nary_op_insert_pieces_predicated (2, EQ_EXPR, boolean_type_node,
7717 ops, boolean_false_node, 0, pred_e);
7718 break;
7719 case GT_EXPR:
7720 /* a > b -> a {!,>}= b */
7721 vn_nary_op_insert_pieces_predicated (2, NE_EXPR, boolean_type_node,
7722 ops, boolean_true_node, 0, pred_e);
7723 vn_nary_op_insert_pieces_predicated (2, GE_EXPR, boolean_type_node,
7724 ops, boolean_true_node, 0, pred_e);
7725 /* a > b -> ! a {<,=} b */
7726 vn_nary_op_insert_pieces_predicated (2, LT_EXPR, boolean_type_node,
7727 ops, boolean_false_node, 0, pred_e);
7728 vn_nary_op_insert_pieces_predicated (2, EQ_EXPR, boolean_type_node,
7729 ops, boolean_false_node, 0, pred_e);
7730 break;
7731 case EQ_EXPR:
7732 /* a == b -> ! a {<,>} b */
7733 vn_nary_op_insert_pieces_predicated (2, LT_EXPR, boolean_type_node,
7734 ops, boolean_false_node, 0, pred_e);
7735 vn_nary_op_insert_pieces_predicated (2, GT_EXPR, boolean_type_node,
7736 ops, boolean_false_node, 0, pred_e);
7737 break;
7738 case LE_EXPR:
7739 case GE_EXPR:
7740 case NE_EXPR:
7741 /* Nothing besides inverted condition. */
7742 break;
7743 default:;
7747 /* Main stmt worker for RPO VN, process BB. */
7749 static unsigned
7750 process_bb (rpo_elim &avail, basic_block bb,
7751 bool bb_visited, bool iterate_phis, bool iterate, bool eliminate,
7752 bool do_region, bitmap exit_bbs, bool skip_phis)
7754 unsigned todo = 0;
7755 edge_iterator ei;
7756 edge e;
7758 vn_context_bb = bb;
7760 /* If we are in loop-closed SSA preserve this state. This is
7761 relevant when called on regions from outside of FRE/PRE. */
7762 bool lc_phi_nodes = false;
7763 if (!skip_phis
7764 && loops_state_satisfies_p (LOOP_CLOSED_SSA))
7765 FOR_EACH_EDGE (e, ei, bb->preds)
7766 if (e->src->loop_father != e->dest->loop_father
7767 && flow_loop_nested_p (e->dest->loop_father,
7768 e->src->loop_father))
7770 lc_phi_nodes = true;
7771 break;
7774 /* When we visit a loop header substitute into loop info. */
7775 if (!iterate && eliminate && bb->loop_father->header == bb)
7777 /* Keep fields in sync with substitute_in_loop_info. */
7778 if (bb->loop_father->nb_iterations)
7779 bb->loop_father->nb_iterations
7780 = simplify_replace_tree (bb->loop_father->nb_iterations,
7781 NULL_TREE, NULL_TREE, &vn_valueize_for_srt);
7784 /* Value-number all defs in the basic-block. */
7785 if (!skip_phis)
7786 for (gphi_iterator gsi = gsi_start_phis (bb); !gsi_end_p (gsi);
7787 gsi_next (&gsi))
7789 gphi *phi = gsi.phi ();
7790 tree res = PHI_RESULT (phi);
7791 vn_ssa_aux_t res_info = VN_INFO (res);
7792 if (!bb_visited)
7794 gcc_assert (!res_info->visited);
7795 res_info->valnum = VN_TOP;
7796 res_info->visited = true;
7799 /* When not iterating force backedge values to varying. */
7800 visit_stmt (phi, !iterate_phis);
7801 if (virtual_operand_p (res))
7802 continue;
7804 /* Eliminate */
7805 /* The interesting case is gcc.dg/tree-ssa/pr22230.c for correctness
7806 how we handle backedges and availability.
7807 And gcc.dg/tree-ssa/ssa-sccvn-2.c for optimization. */
7808 tree val = res_info->valnum;
7809 if (res != val && !iterate && eliminate)
7811 if (tree leader = avail.eliminate_avail (bb, res))
7813 if (leader != res
7814 /* Preserve loop-closed SSA form. */
7815 && (! lc_phi_nodes
7816 || is_gimple_min_invariant (leader)))
7818 if (dump_file && (dump_flags & TDF_DETAILS))
7820 fprintf (dump_file, "Replaced redundant PHI node "
7821 "defining ");
7822 print_generic_expr (dump_file, res);
7823 fprintf (dump_file, " with ");
7824 print_generic_expr (dump_file, leader);
7825 fprintf (dump_file, "\n");
7827 avail.eliminations++;
7829 if (may_propagate_copy (res, leader))
7831 /* Schedule for removal. */
7832 avail.to_remove.safe_push (phi);
7833 continue;
7835 /* ??? Else generate a copy stmt. */
7839 /* Only make defs available that not already are. But make
7840 sure loop-closed SSA PHI node defs are picked up for
7841 downstream uses. */
7842 if (lc_phi_nodes
7843 || res == val
7844 || ! avail.eliminate_avail (bb, res))
7845 avail.eliminate_push_avail (bb, res);
7848 /* For empty BBs mark outgoing edges executable. For non-empty BBs
7849 we do this when processing the last stmt as we have to do this
7850 before elimination which otherwise forces GIMPLE_CONDs to
7851 if (1 != 0) style when seeing non-executable edges. */
7852 if (gsi_end_p (gsi_start_bb (bb)))
7854 FOR_EACH_EDGE (e, ei, bb->succs)
7856 if (!(e->flags & EDGE_EXECUTABLE))
7858 if (dump_file && (dump_flags & TDF_DETAILS))
7859 fprintf (dump_file,
7860 "marking outgoing edge %d -> %d executable\n",
7861 e->src->index, e->dest->index);
7862 e->flags |= EDGE_EXECUTABLE;
7863 e->dest->flags |= BB_EXECUTABLE;
7865 else if (!(e->dest->flags & BB_EXECUTABLE))
7867 if (dump_file && (dump_flags & TDF_DETAILS))
7868 fprintf (dump_file,
7869 "marking destination block %d reachable\n",
7870 e->dest->index);
7871 e->dest->flags |= BB_EXECUTABLE;
7875 for (gimple_stmt_iterator gsi = gsi_start_bb (bb);
7876 !gsi_end_p (gsi); gsi_next (&gsi))
7878 ssa_op_iter i;
7879 tree op;
7880 if (!bb_visited)
7882 FOR_EACH_SSA_TREE_OPERAND (op, gsi_stmt (gsi), i, SSA_OP_ALL_DEFS)
7884 vn_ssa_aux_t op_info = VN_INFO (op);
7885 gcc_assert (!op_info->visited);
7886 op_info->valnum = VN_TOP;
7887 op_info->visited = true;
7890 /* We somehow have to deal with uses that are not defined
7891 in the processed region. Forcing unvisited uses to
7892 varying here doesn't play well with def-use following during
7893 expression simplification, so we deal with this by checking
7894 the visited flag in SSA_VAL. */
7897 visit_stmt (gsi_stmt (gsi));
7899 gimple *last = gsi_stmt (gsi);
7900 e = NULL;
7901 switch (gimple_code (last))
7903 case GIMPLE_SWITCH:
7904 e = find_taken_edge (bb, vn_valueize (gimple_switch_index
7905 (as_a <gswitch *> (last))));
7906 break;
7907 case GIMPLE_COND:
7909 tree lhs = vn_valueize (gimple_cond_lhs (last));
7910 tree rhs = vn_valueize (gimple_cond_rhs (last));
7911 tree val = gimple_simplify (gimple_cond_code (last),
7912 boolean_type_node, lhs, rhs,
7913 NULL, vn_valueize);
7914 /* If the condition didn't simplfy see if we have recorded
7915 an expression from sofar taken edges. */
7916 if (! val || TREE_CODE (val) != INTEGER_CST)
7918 vn_nary_op_t vnresult;
7919 tree ops[2];
7920 ops[0] = lhs;
7921 ops[1] = rhs;
7922 val = vn_nary_op_lookup_pieces (2, gimple_cond_code (last),
7923 boolean_type_node, ops,
7924 &vnresult);
7925 /* Did we get a predicated value? */
7926 if (! val && vnresult && vnresult->predicated_values)
7928 val = vn_nary_op_get_predicated_value (vnresult, bb);
7929 if (val && dump_file && (dump_flags & TDF_DETAILS))
7931 fprintf (dump_file, "Got predicated value ");
7932 print_generic_expr (dump_file, val, TDF_NONE);
7933 fprintf (dump_file, " for ");
7934 print_gimple_stmt (dump_file, last, TDF_SLIM);
7938 if (val)
7939 e = find_taken_edge (bb, val);
7940 if (! e)
7942 /* If we didn't manage to compute the taken edge then
7943 push predicated expressions for the condition itself
7944 and related conditions to the hashtables. This allows
7945 simplification of redundant conditions which is
7946 important as early cleanup. */
7947 edge true_e, false_e;
7948 extract_true_false_edges_from_block (bb, &true_e, &false_e);
7949 enum tree_code code = gimple_cond_code (last);
7950 enum tree_code icode
7951 = invert_tree_comparison (code, HONOR_NANS (lhs));
7952 tree ops[2];
7953 ops[0] = lhs;
7954 ops[1] = rhs;
7955 if ((do_region && bitmap_bit_p (exit_bbs, true_e->dest->index))
7956 || !can_track_predicate_on_edge (true_e))
7957 true_e = NULL;
7958 if ((do_region && bitmap_bit_p (exit_bbs, false_e->dest->index))
7959 || !can_track_predicate_on_edge (false_e))
7960 false_e = NULL;
7961 if (true_e)
7962 vn_nary_op_insert_pieces_predicated
7963 (2, code, boolean_type_node, ops,
7964 boolean_true_node, 0, true_e);
7965 if (false_e)
7966 vn_nary_op_insert_pieces_predicated
7967 (2, code, boolean_type_node, ops,
7968 boolean_false_node, 0, false_e);
7969 if (icode != ERROR_MARK)
7971 if (true_e)
7972 vn_nary_op_insert_pieces_predicated
7973 (2, icode, boolean_type_node, ops,
7974 boolean_false_node, 0, true_e);
7975 if (false_e)
7976 vn_nary_op_insert_pieces_predicated
7977 (2, icode, boolean_type_node, ops,
7978 boolean_true_node, 0, false_e);
7980 /* Relax for non-integers, inverted condition handled
7981 above. */
7982 if (INTEGRAL_TYPE_P (TREE_TYPE (lhs)))
7984 if (true_e)
7985 insert_related_predicates_on_edge (code, ops, true_e);
7986 if (false_e)
7987 insert_related_predicates_on_edge (icode, ops, false_e);
7990 break;
7992 case GIMPLE_GOTO:
7993 e = find_taken_edge (bb, vn_valueize (gimple_goto_dest (last)));
7994 break;
7995 default:
7996 e = NULL;
7998 if (e)
8000 todo = TODO_cleanup_cfg;
8001 if (!(e->flags & EDGE_EXECUTABLE))
8003 if (dump_file && (dump_flags & TDF_DETAILS))
8004 fprintf (dump_file,
8005 "marking known outgoing %sedge %d -> %d executable\n",
8006 e->flags & EDGE_DFS_BACK ? "back-" : "",
8007 e->src->index, e->dest->index);
8008 e->flags |= EDGE_EXECUTABLE;
8009 e->dest->flags |= BB_EXECUTABLE;
8011 else if (!(e->dest->flags & BB_EXECUTABLE))
8013 if (dump_file && (dump_flags & TDF_DETAILS))
8014 fprintf (dump_file,
8015 "marking destination block %d reachable\n",
8016 e->dest->index);
8017 e->dest->flags |= BB_EXECUTABLE;
8020 else if (gsi_one_before_end_p (gsi))
8022 FOR_EACH_EDGE (e, ei, bb->succs)
8024 if (!(e->flags & EDGE_EXECUTABLE))
8026 if (dump_file && (dump_flags & TDF_DETAILS))
8027 fprintf (dump_file,
8028 "marking outgoing edge %d -> %d executable\n",
8029 e->src->index, e->dest->index);
8030 e->flags |= EDGE_EXECUTABLE;
8031 e->dest->flags |= BB_EXECUTABLE;
8033 else if (!(e->dest->flags & BB_EXECUTABLE))
8035 if (dump_file && (dump_flags & TDF_DETAILS))
8036 fprintf (dump_file,
8037 "marking destination block %d reachable\n",
8038 e->dest->index);
8039 e->dest->flags |= BB_EXECUTABLE;
8044 /* Eliminate. That also pushes to avail. */
8045 if (eliminate && ! iterate)
8046 avail.eliminate_stmt (bb, &gsi);
8047 else
8048 /* If not eliminating, make all not already available defs
8049 available. */
8050 FOR_EACH_SSA_TREE_OPERAND (op, gsi_stmt (gsi), i, SSA_OP_DEF)
8051 if (! avail.eliminate_avail (bb, op))
8052 avail.eliminate_push_avail (bb, op);
8055 /* Eliminate in destination PHI arguments. Always substitute in dest
8056 PHIs, even for non-executable edges. This handles region
8057 exits PHIs. */
8058 if (!iterate && eliminate)
8059 FOR_EACH_EDGE (e, ei, bb->succs)
8060 for (gphi_iterator gsi = gsi_start_phis (e->dest);
8061 !gsi_end_p (gsi); gsi_next (&gsi))
8063 gphi *phi = gsi.phi ();
8064 use_operand_p use_p = PHI_ARG_DEF_PTR_FROM_EDGE (phi, e);
8065 tree arg = USE_FROM_PTR (use_p);
8066 if (TREE_CODE (arg) != SSA_NAME
8067 || virtual_operand_p (arg))
8068 continue;
8069 tree sprime;
8070 if (SSA_NAME_IS_DEFAULT_DEF (arg))
8072 sprime = SSA_VAL (arg);
8073 gcc_assert (TREE_CODE (sprime) != SSA_NAME
8074 || SSA_NAME_IS_DEFAULT_DEF (sprime));
8076 else
8077 /* Look for sth available at the definition block of the argument.
8078 This avoids inconsistencies between availability there which
8079 decides if the stmt can be removed and availability at the
8080 use site. The SSA property ensures that things available
8081 at the definition are also available at uses. */
8082 sprime = avail.eliminate_avail (gimple_bb (SSA_NAME_DEF_STMT (arg)),
8083 arg);
8084 if (sprime
8085 && sprime != arg
8086 && may_propagate_copy (arg, sprime))
8087 propagate_value (use_p, sprime);
8090 vn_context_bb = NULL;
8091 return todo;
8094 /* Unwind state per basic-block. */
8096 struct unwind_state
8098 /* Times this block has been visited. */
8099 unsigned visited;
8100 /* Whether to handle this as iteration point or whether to treat
8101 incoming backedge PHI values as varying. */
8102 bool iterate;
8103 /* Maximum RPO index this block is reachable from. */
8104 int max_rpo;
8105 /* Unwind state. */
8106 void *ob_top;
8107 vn_reference_t ref_top;
8108 vn_phi_t phi_top;
8109 vn_nary_op_t nary_top;
8110 vn_avail *avail_top;
8113 /* Unwind the RPO VN state for iteration. */
8115 static void
8116 do_unwind (unwind_state *to, rpo_elim &avail)
8118 gcc_assert (to->iterate);
8119 for (; last_inserted_nary != to->nary_top;
8120 last_inserted_nary = last_inserted_nary->next)
8122 vn_nary_op_t *slot;
8123 slot = valid_info->nary->find_slot_with_hash
8124 (last_inserted_nary, last_inserted_nary->hashcode, NO_INSERT);
8125 /* Predication causes the need to restore previous state. */
8126 if ((*slot)->unwind_to)
8127 *slot = (*slot)->unwind_to;
8128 else
8129 valid_info->nary->clear_slot (slot);
8131 for (; last_inserted_phi != to->phi_top;
8132 last_inserted_phi = last_inserted_phi->next)
8134 vn_phi_t *slot;
8135 slot = valid_info->phis->find_slot_with_hash
8136 (last_inserted_phi, last_inserted_phi->hashcode, NO_INSERT);
8137 valid_info->phis->clear_slot (slot);
8139 for (; last_inserted_ref != to->ref_top;
8140 last_inserted_ref = last_inserted_ref->next)
8142 vn_reference_t *slot;
8143 slot = valid_info->references->find_slot_with_hash
8144 (last_inserted_ref, last_inserted_ref->hashcode, NO_INSERT);
8145 (*slot)->operands.release ();
8146 valid_info->references->clear_slot (slot);
8148 obstack_free (&vn_tables_obstack, to->ob_top);
8150 /* Prune [rpo_idx, ] from avail. */
8151 for (; last_pushed_avail && last_pushed_avail->avail != to->avail_top;)
8153 vn_ssa_aux_t val = last_pushed_avail;
8154 vn_avail *av = val->avail;
8155 val->avail = av->next;
8156 last_pushed_avail = av->next_undo;
8157 av->next = avail.m_avail_freelist;
8158 avail.m_avail_freelist = av;
8162 /* Do VN on a SEME region specified by ENTRY and EXIT_BBS in FN.
8163 If ITERATE is true then treat backedges optimistically as not
8164 executed and iterate. If ELIMINATE is true then perform
8165 elimination, otherwise leave that to the caller. */
8167 static unsigned
8168 do_rpo_vn_1 (function *fn, edge entry, bitmap exit_bbs,
8169 bool iterate, bool eliminate, vn_lookup_kind kind)
8171 unsigned todo = 0;
8172 default_vn_walk_kind = kind;
8174 /* We currently do not support region-based iteration when
8175 elimination is requested. */
8176 gcc_assert (!entry || !iterate || !eliminate);
8177 /* When iterating we need loop info up-to-date. */
8178 gcc_assert (!iterate || !loops_state_satisfies_p (LOOPS_NEED_FIXUP));
8180 bool do_region = entry != NULL;
8181 if (!do_region)
8183 entry = single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (fn));
8184 exit_bbs = BITMAP_ALLOC (NULL);
8185 bitmap_set_bit (exit_bbs, EXIT_BLOCK);
8188 /* Clear EDGE_DFS_BACK on "all" entry edges, RPO order compute will
8189 re-mark those that are contained in the region. */
8190 edge_iterator ei;
8191 edge e;
8192 FOR_EACH_EDGE (e, ei, entry->dest->preds)
8193 e->flags &= ~EDGE_DFS_BACK;
8195 int *rpo = XNEWVEC (int, n_basic_blocks_for_fn (fn) - NUM_FIXED_BLOCKS);
8196 auto_vec<std::pair<int, int> > toplevel_scc_extents;
8197 int n = rev_post_order_and_mark_dfs_back_seme
8198 (fn, entry, exit_bbs, true, rpo, !iterate ? &toplevel_scc_extents : NULL);
8200 if (!do_region)
8201 BITMAP_FREE (exit_bbs);
8203 /* If there are any non-DFS_BACK edges into entry->dest skip
8204 processing PHI nodes for that block. This supports
8205 value-numbering loop bodies w/o the actual loop. */
8206 FOR_EACH_EDGE (e, ei, entry->dest->preds)
8207 if (e != entry
8208 && !(e->flags & EDGE_DFS_BACK))
8209 break;
8210 bool skip_entry_phis = e != NULL;
8211 if (skip_entry_phis && dump_file && (dump_flags & TDF_DETAILS))
8212 fprintf (dump_file, "Region does not contain all edges into "
8213 "the entry block, skipping its PHIs.\n");
8215 int *bb_to_rpo = XNEWVEC (int, last_basic_block_for_fn (fn));
8216 for (int i = 0; i < n; ++i)
8217 bb_to_rpo[rpo[i]] = i;
8219 unwind_state *rpo_state = XNEWVEC (unwind_state, n);
8221 rpo_elim avail (entry->dest);
8222 rpo_avail = &avail;
8224 /* Verify we have no extra entries into the region. */
8225 if (flag_checking && do_region)
8227 auto_bb_flag bb_in_region (fn);
8228 for (int i = 0; i < n; ++i)
8230 basic_block bb = BASIC_BLOCK_FOR_FN (fn, rpo[i]);
8231 bb->flags |= bb_in_region;
8233 /* We can't merge the first two loops because we cannot rely
8234 on EDGE_DFS_BACK for edges not within the region. But if
8235 we decide to always have the bb_in_region flag we can
8236 do the checking during the RPO walk itself (but then it's
8237 also easy to handle MEME conservatively). */
8238 for (int i = 0; i < n; ++i)
8240 basic_block bb = BASIC_BLOCK_FOR_FN (fn, rpo[i]);
8241 edge e;
8242 edge_iterator ei;
8243 FOR_EACH_EDGE (e, ei, bb->preds)
8244 gcc_assert (e == entry
8245 || (skip_entry_phis && bb == entry->dest)
8246 || (e->src->flags & bb_in_region));
8248 for (int i = 0; i < n; ++i)
8250 basic_block bb = BASIC_BLOCK_FOR_FN (fn, rpo[i]);
8251 bb->flags &= ~bb_in_region;
8255 /* Create the VN state. For the initial size of the various hashtables
8256 use a heuristic based on region size and number of SSA names. */
8257 unsigned region_size = (((unsigned HOST_WIDE_INT)n * num_ssa_names)
8258 / (n_basic_blocks_for_fn (fn) - NUM_FIXED_BLOCKS));
8259 VN_TOP = create_tmp_var_raw (void_type_node, "vn_top");
8260 next_value_id = 1;
8261 next_constant_value_id = -1;
8263 vn_ssa_aux_hash = new hash_table <vn_ssa_aux_hasher> (region_size * 2);
8264 gcc_obstack_init (&vn_ssa_aux_obstack);
8266 gcc_obstack_init (&vn_tables_obstack);
8267 gcc_obstack_init (&vn_tables_insert_obstack);
8268 valid_info = XCNEW (struct vn_tables_s);
8269 allocate_vn_table (valid_info, region_size);
8270 last_inserted_ref = NULL;
8271 last_inserted_phi = NULL;
8272 last_inserted_nary = NULL;
8273 last_pushed_avail = NULL;
8275 vn_valueize = rpo_vn_valueize;
8277 /* Initialize the unwind state and edge/BB executable state. */
8278 unsigned curr_scc = 0;
8279 for (int i = 0; i < n; ++i)
8281 basic_block bb = BASIC_BLOCK_FOR_FN (fn, rpo[i]);
8282 rpo_state[i].visited = 0;
8283 rpo_state[i].max_rpo = i;
8284 if (!iterate && curr_scc < toplevel_scc_extents.length ())
8286 if (i >= toplevel_scc_extents[curr_scc].first
8287 && i <= toplevel_scc_extents[curr_scc].second)
8288 rpo_state[i].max_rpo = toplevel_scc_extents[curr_scc].second;
8289 if (i == toplevel_scc_extents[curr_scc].second)
8290 curr_scc++;
8292 bb->flags &= ~BB_EXECUTABLE;
8293 bool has_backedges = false;
8294 edge e;
8295 edge_iterator ei;
8296 FOR_EACH_EDGE (e, ei, bb->preds)
8298 if (e->flags & EDGE_DFS_BACK)
8299 has_backedges = true;
8300 e->flags &= ~EDGE_EXECUTABLE;
8301 if (iterate || e == entry || (skip_entry_phis && bb == entry->dest))
8302 continue;
8304 rpo_state[i].iterate = iterate && has_backedges;
8306 entry->flags |= EDGE_EXECUTABLE;
8307 entry->dest->flags |= BB_EXECUTABLE;
8309 /* As heuristic to improve compile-time we handle only the N innermost
8310 loops and the outermost one optimistically. */
8311 if (iterate)
8313 unsigned max_depth = param_rpo_vn_max_loop_depth;
8314 for (auto loop : loops_list (cfun, LI_ONLY_INNERMOST))
8315 if (loop_depth (loop) > max_depth)
8316 for (unsigned i = 2;
8317 i < loop_depth (loop) - max_depth; ++i)
8319 basic_block header = superloop_at_depth (loop, i)->header;
8320 bool non_latch_backedge = false;
8321 edge e;
8322 edge_iterator ei;
8323 FOR_EACH_EDGE (e, ei, header->preds)
8324 if (e->flags & EDGE_DFS_BACK)
8326 /* There can be a non-latch backedge into the header
8327 which is part of an outer irreducible region. We
8328 cannot avoid iterating this block then. */
8329 if (!dominated_by_p (CDI_DOMINATORS,
8330 e->src, e->dest))
8332 if (dump_file && (dump_flags & TDF_DETAILS))
8333 fprintf (dump_file, "non-latch backedge %d -> %d "
8334 "forces iteration of loop %d\n",
8335 e->src->index, e->dest->index, loop->num);
8336 non_latch_backedge = true;
8338 else
8339 e->flags |= EDGE_EXECUTABLE;
8341 rpo_state[bb_to_rpo[header->index]].iterate = non_latch_backedge;
8345 uint64_t nblk = 0;
8346 int idx = 0;
8347 if (iterate)
8348 /* Go and process all blocks, iterating as necessary. */
8351 basic_block bb = BASIC_BLOCK_FOR_FN (fn, rpo[idx]);
8353 /* If the block has incoming backedges remember unwind state. This
8354 is required even for non-executable blocks since in irreducible
8355 regions we might reach them via the backedge and re-start iterating
8356 from there.
8357 Note we can individually mark blocks with incoming backedges to
8358 not iterate where we then handle PHIs conservatively. We do that
8359 heuristically to reduce compile-time for degenerate cases. */
8360 if (rpo_state[idx].iterate)
8362 rpo_state[idx].ob_top = obstack_alloc (&vn_tables_obstack, 0);
8363 rpo_state[idx].ref_top = last_inserted_ref;
8364 rpo_state[idx].phi_top = last_inserted_phi;
8365 rpo_state[idx].nary_top = last_inserted_nary;
8366 rpo_state[idx].avail_top
8367 = last_pushed_avail ? last_pushed_avail->avail : NULL;
8370 if (!(bb->flags & BB_EXECUTABLE))
8372 if (dump_file && (dump_flags & TDF_DETAILS))
8373 fprintf (dump_file, "Block %d: BB%d found not executable\n",
8374 idx, bb->index);
8375 idx++;
8376 continue;
8379 if (dump_file && (dump_flags & TDF_DETAILS))
8380 fprintf (dump_file, "Processing block %d: BB%d\n", idx, bb->index);
8381 nblk++;
8382 todo |= process_bb (avail, bb,
8383 rpo_state[idx].visited != 0,
8384 rpo_state[idx].iterate,
8385 iterate, eliminate, do_region, exit_bbs, false);
8386 rpo_state[idx].visited++;
8388 /* Verify if changed values flow over executable outgoing backedges
8389 and those change destination PHI values (that's the thing we
8390 can easily verify). Reduce over all such edges to the farthest
8391 away PHI. */
8392 int iterate_to = -1;
8393 edge_iterator ei;
8394 edge e;
8395 FOR_EACH_EDGE (e, ei, bb->succs)
8396 if ((e->flags & (EDGE_DFS_BACK|EDGE_EXECUTABLE))
8397 == (EDGE_DFS_BACK|EDGE_EXECUTABLE)
8398 && rpo_state[bb_to_rpo[e->dest->index]].iterate)
8400 int destidx = bb_to_rpo[e->dest->index];
8401 if (!rpo_state[destidx].visited)
8403 if (dump_file && (dump_flags & TDF_DETAILS))
8404 fprintf (dump_file, "Unvisited destination %d\n",
8405 e->dest->index);
8406 if (iterate_to == -1 || destidx < iterate_to)
8407 iterate_to = destidx;
8408 continue;
8410 if (dump_file && (dump_flags & TDF_DETAILS))
8411 fprintf (dump_file, "Looking for changed values of backedge"
8412 " %d->%d destination PHIs\n",
8413 e->src->index, e->dest->index);
8414 vn_context_bb = e->dest;
8415 gphi_iterator gsi;
8416 for (gsi = gsi_start_phis (e->dest);
8417 !gsi_end_p (gsi); gsi_next (&gsi))
8419 bool inserted = false;
8420 /* While we'd ideally just iterate on value changes
8421 we CSE PHIs and do that even across basic-block
8422 boundaries. So even hashtable state changes can
8423 be important (which is roughly equivalent to
8424 PHI argument value changes). To not excessively
8425 iterate because of that we track whether a PHI
8426 was CSEd to with GF_PLF_1. */
8427 bool phival_changed;
8428 if ((phival_changed = visit_phi (gsi.phi (),
8429 &inserted, false))
8430 || (inserted && gimple_plf (gsi.phi (), GF_PLF_1)))
8432 if (!phival_changed
8433 && dump_file && (dump_flags & TDF_DETAILS))
8434 fprintf (dump_file, "PHI was CSEd and hashtable "
8435 "state (changed)\n");
8436 if (iterate_to == -1 || destidx < iterate_to)
8437 iterate_to = destidx;
8438 break;
8441 vn_context_bb = NULL;
8443 if (iterate_to != -1)
8445 do_unwind (&rpo_state[iterate_to], avail);
8446 idx = iterate_to;
8447 if (dump_file && (dump_flags & TDF_DETAILS))
8448 fprintf (dump_file, "Iterating to %d BB%d\n",
8449 iterate_to, rpo[iterate_to]);
8450 continue;
8453 idx++;
8455 while (idx < n);
8457 else /* !iterate */
8459 /* Process all blocks greedily with a worklist that enforces RPO
8460 processing of reachable blocks. */
8461 auto_bitmap worklist;
8462 bitmap_set_bit (worklist, 0);
8463 while (!bitmap_empty_p (worklist))
8465 int idx = bitmap_first_set_bit (worklist);
8466 bitmap_clear_bit (worklist, idx);
8467 basic_block bb = BASIC_BLOCK_FOR_FN (fn, rpo[idx]);
8468 gcc_assert ((bb->flags & BB_EXECUTABLE)
8469 && !rpo_state[idx].visited);
8471 if (dump_file && (dump_flags & TDF_DETAILS))
8472 fprintf (dump_file, "Processing block %d: BB%d\n", idx, bb->index);
8474 /* When we run into predecessor edges where we cannot trust its
8475 executable state mark them executable so PHI processing will
8476 be conservative.
8477 ??? Do we need to force arguments flowing over that edge
8478 to be varying or will they even always be? */
8479 edge_iterator ei;
8480 edge e;
8481 FOR_EACH_EDGE (e, ei, bb->preds)
8482 if (!(e->flags & EDGE_EXECUTABLE)
8483 && (bb == entry->dest
8484 || (!rpo_state[bb_to_rpo[e->src->index]].visited
8485 && (rpo_state[bb_to_rpo[e->src->index]].max_rpo
8486 >= (int)idx))))
8488 if (dump_file && (dump_flags & TDF_DETAILS))
8489 fprintf (dump_file, "Cannot trust state of predecessor "
8490 "edge %d -> %d, marking executable\n",
8491 e->src->index, e->dest->index);
8492 e->flags |= EDGE_EXECUTABLE;
8495 nblk++;
8496 todo |= process_bb (avail, bb, false, false, false, eliminate,
8497 do_region, exit_bbs,
8498 skip_entry_phis && bb == entry->dest);
8499 rpo_state[idx].visited++;
8501 FOR_EACH_EDGE (e, ei, bb->succs)
8502 if ((e->flags & EDGE_EXECUTABLE)
8503 && e->dest->index != EXIT_BLOCK
8504 && (!do_region || !bitmap_bit_p (exit_bbs, e->dest->index))
8505 && !rpo_state[bb_to_rpo[e->dest->index]].visited)
8506 bitmap_set_bit (worklist, bb_to_rpo[e->dest->index]);
8510 /* If statistics or dump file active. */
8511 int nex = 0;
8512 unsigned max_visited = 1;
8513 for (int i = 0; i < n; ++i)
8515 basic_block bb = BASIC_BLOCK_FOR_FN (fn, rpo[i]);
8516 if (bb->flags & BB_EXECUTABLE)
8517 nex++;
8518 statistics_histogram_event (cfun, "RPO block visited times",
8519 rpo_state[i].visited);
8520 if (rpo_state[i].visited > max_visited)
8521 max_visited = rpo_state[i].visited;
8523 unsigned nvalues = 0, navail = 0;
8524 for (hash_table<vn_ssa_aux_hasher>::iterator i = vn_ssa_aux_hash->begin ();
8525 i != vn_ssa_aux_hash->end (); ++i)
8527 nvalues++;
8528 vn_avail *av = (*i)->avail;
8529 while (av)
8531 navail++;
8532 av = av->next;
8535 statistics_counter_event (cfun, "RPO blocks", n);
8536 statistics_counter_event (cfun, "RPO blocks visited", nblk);
8537 statistics_counter_event (cfun, "RPO blocks executable", nex);
8538 statistics_histogram_event (cfun, "RPO iterations", 10*nblk / nex);
8539 statistics_histogram_event (cfun, "RPO num values", nvalues);
8540 statistics_histogram_event (cfun, "RPO num avail", navail);
8541 statistics_histogram_event (cfun, "RPO num lattice",
8542 vn_ssa_aux_hash->elements ());
8543 if (dump_file && (dump_flags & (TDF_DETAILS|TDF_STATS)))
8545 fprintf (dump_file, "RPO iteration over %d blocks visited %" PRIu64
8546 " blocks in total discovering %d executable blocks iterating "
8547 "%d.%d times, a block was visited max. %u times\n",
8548 n, nblk, nex,
8549 (int)((10*nblk / nex)/10), (int)((10*nblk / nex)%10),
8550 max_visited);
8551 fprintf (dump_file, "RPO tracked %d values available at %d locations "
8552 "and %" PRIu64 " lattice elements\n",
8553 nvalues, navail, (uint64_t) vn_ssa_aux_hash->elements ());
8556 if (eliminate)
8558 /* When !iterate we already performed elimination during the RPO
8559 walk. */
8560 if (iterate)
8562 /* Elimination for region-based VN needs to be done within the
8563 RPO walk. */
8564 gcc_assert (! do_region);
8565 /* Note we can't use avail.walk here because that gets confused
8566 by the existing availability and it will be less efficient
8567 as well. */
8568 todo |= eliminate_with_rpo_vn (NULL);
8570 else
8571 todo |= avail.eliminate_cleanup (do_region);
8574 vn_valueize = NULL;
8575 rpo_avail = NULL;
8577 XDELETEVEC (bb_to_rpo);
8578 XDELETEVEC (rpo);
8579 XDELETEVEC (rpo_state);
8581 return todo;
8584 /* Region-based entry for RPO VN. Performs value-numbering and elimination
8585 on the SEME region specified by ENTRY and EXIT_BBS. If ENTRY is not
8586 the only edge into the region at ENTRY->dest PHI nodes in ENTRY->dest
8587 are not considered.
8588 If ITERATE is true then treat backedges optimistically as not
8589 executed and iterate. If ELIMINATE is true then perform
8590 elimination, otherwise leave that to the caller.
8591 KIND specifies the amount of work done for handling memory operations. */
8593 unsigned
8594 do_rpo_vn (function *fn, edge entry, bitmap exit_bbs,
8595 bool iterate, bool eliminate, vn_lookup_kind kind)
8597 auto_timevar tv (TV_TREE_RPO_VN);
8598 unsigned todo = do_rpo_vn_1 (fn, entry, exit_bbs, iterate, eliminate, kind);
8599 free_rpo_vn ();
8600 return todo;
8604 namespace {
8606 const pass_data pass_data_fre =
8608 GIMPLE_PASS, /* type */
8609 "fre", /* name */
8610 OPTGROUP_NONE, /* optinfo_flags */
8611 TV_TREE_FRE, /* tv_id */
8612 ( PROP_cfg | PROP_ssa ), /* properties_required */
8613 0, /* properties_provided */
8614 0, /* properties_destroyed */
8615 0, /* todo_flags_start */
8616 0, /* todo_flags_finish */
8619 class pass_fre : public gimple_opt_pass
8621 public:
8622 pass_fre (gcc::context *ctxt)
8623 : gimple_opt_pass (pass_data_fre, ctxt), may_iterate (true)
8626 /* opt_pass methods: */
8627 opt_pass * clone () final override { return new pass_fre (m_ctxt); }
8628 void set_pass_param (unsigned int n, bool param) final override
8630 gcc_assert (n == 0);
8631 may_iterate = param;
8633 bool gate (function *) final override
8635 return flag_tree_fre != 0 && (may_iterate || optimize > 1);
8637 unsigned int execute (function *) final override;
8639 private:
8640 bool may_iterate;
8641 }; // class pass_fre
8643 unsigned int
8644 pass_fre::execute (function *fun)
8646 unsigned todo = 0;
8648 /* At -O[1g] use the cheap non-iterating mode. */
8649 bool iterate_p = may_iterate && (optimize > 1);
8650 calculate_dominance_info (CDI_DOMINATORS);
8651 if (iterate_p)
8652 loop_optimizer_init (AVOID_CFG_MODIFICATIONS);
8654 todo = do_rpo_vn_1 (fun, NULL, NULL, iterate_p, true, VN_WALKREWRITE);
8655 free_rpo_vn ();
8657 if (iterate_p)
8658 loop_optimizer_finalize ();
8660 if (scev_initialized_p ())
8661 scev_reset_htab ();
8663 /* For late FRE after IVOPTs and unrolling, see if we can
8664 remove some TREE_ADDRESSABLE and rewrite stuff into SSA. */
8665 if (!may_iterate)
8666 todo |= TODO_update_address_taken;
8668 return todo;
8671 } // anon namespace
8673 gimple_opt_pass *
8674 make_pass_fre (gcc::context *ctxt)
8676 return new pass_fre (ctxt);
8679 #undef BB_EXECUTABLE