testsuite: Correct requirements for vadsdu*, vslv and vsrv testcases.
[official-gcc.git] / gcc / tree-ssa-sccvn.c
blobc139adb6130b4f0cd43e265c6bc88e9f40f61b28
1 /* SCC value numbering for trees
2 Copyright (C) 2006-2020 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-fold.h"
43 #include "tree-eh.h"
44 #include "gimplify.h"
45 #include "flags.h"
46 #include "dojump.h"
47 #include "explow.h"
48 #include "calls.h"
49 #include "varasm.h"
50 #include "stmt.h"
51 #include "expr.h"
52 #include "tree-dfa.h"
53 #include "tree-ssa.h"
54 #include "dumpfile.h"
55 #include "cfgloop.h"
56 #include "tree-ssa-propagate.h"
57 #include "tree-cfg.h"
58 #include "domwalk.h"
59 #include "gimple-iterator.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 "tree-ssa-sccvn.h"
75 /* This algorithm is based on the SCC algorithm presented by Keith
76 Cooper and L. Taylor Simpson in "SCC-Based Value numbering"
77 (http://citeseer.ist.psu.edu/41805.html). In
78 straight line code, it is equivalent to a regular hash based value
79 numbering that is performed in reverse postorder.
81 For code with cycles, there are two alternatives, both of which
82 require keeping the hashtables separate from the actual list of
83 value numbers for SSA names.
85 1. Iterate value numbering in an RPO walk of the blocks, removing
86 all the entries from the hashtable after each iteration (but
87 keeping the SSA name->value number mapping between iterations).
88 Iterate until it does not change.
90 2. Perform value numbering as part of an SCC walk on the SSA graph,
91 iterating only the cycles in the SSA graph until they do not change
92 (using a separate, optimistic hashtable for value numbering the SCC
93 operands).
95 The second is not just faster in practice (because most SSA graph
96 cycles do not involve all the variables in the graph), it also has
97 some nice properties.
99 One of these nice properties is that when we pop an SCC off the
100 stack, we are guaranteed to have processed all the operands coming from
101 *outside of that SCC*, so we do not need to do anything special to
102 ensure they have value numbers.
104 Another nice property is that the SCC walk is done as part of a DFS
105 of the SSA graph, which makes it easy to perform combining and
106 simplifying operations at the same time.
108 The code below is deliberately written in a way that makes it easy
109 to separate the SCC walk from the other work it does.
111 In order to propagate constants through the code, we track which
112 expressions contain constants, and use those while folding. In
113 theory, we could also track expressions whose value numbers are
114 replaced, in case we end up folding based on expression
115 identities.
117 In order to value number memory, we assign value numbers to vuses.
118 This enables us to note that, for example, stores to the same
119 address of the same value from the same starting memory states are
120 equivalent.
121 TODO:
123 1. We can iterate only the changing portions of the SCC's, but
124 I have not seen an SCC big enough for this to be a win.
125 2. If you differentiate between phi nodes for loops and phi nodes
126 for if-then-else, you can properly consider phi nodes in different
127 blocks for equivalence.
128 3. We could value number vuses in more cases, particularly, whole
129 structure copies.
132 /* There's no BB_EXECUTABLE but we can use BB_VISITED. */
133 #define BB_EXECUTABLE BB_VISITED
135 static vn_lookup_kind default_vn_walk_kind;
137 /* vn_nary_op hashtable helpers. */
139 struct vn_nary_op_hasher : nofree_ptr_hash <vn_nary_op_s>
141 typedef vn_nary_op_s *compare_type;
142 static inline hashval_t hash (const vn_nary_op_s *);
143 static inline bool equal (const vn_nary_op_s *, const vn_nary_op_s *);
146 /* Return the computed hashcode for nary operation P1. */
148 inline hashval_t
149 vn_nary_op_hasher::hash (const vn_nary_op_s *vno1)
151 return vno1->hashcode;
154 /* Compare nary operations P1 and P2 and return true if they are
155 equivalent. */
157 inline bool
158 vn_nary_op_hasher::equal (const vn_nary_op_s *vno1, const vn_nary_op_s *vno2)
160 return vno1 == vno2 || vn_nary_op_eq (vno1, vno2);
163 typedef hash_table<vn_nary_op_hasher> vn_nary_op_table_type;
164 typedef vn_nary_op_table_type::iterator vn_nary_op_iterator_type;
167 /* vn_phi hashtable helpers. */
169 static int
170 vn_phi_eq (const_vn_phi_t const vp1, const_vn_phi_t const vp2);
172 struct vn_phi_hasher : nofree_ptr_hash <vn_phi_s>
174 static inline hashval_t hash (const vn_phi_s *);
175 static inline bool equal (const vn_phi_s *, const vn_phi_s *);
178 /* Return the computed hashcode for phi operation P1. */
180 inline hashval_t
181 vn_phi_hasher::hash (const vn_phi_s *vp1)
183 return vp1->hashcode;
186 /* Compare two phi entries for equality, ignoring VN_TOP arguments. */
188 inline bool
189 vn_phi_hasher::equal (const vn_phi_s *vp1, const vn_phi_s *vp2)
191 return vp1 == vp2 || vn_phi_eq (vp1, vp2);
194 typedef hash_table<vn_phi_hasher> vn_phi_table_type;
195 typedef vn_phi_table_type::iterator vn_phi_iterator_type;
198 /* Compare two reference operands P1 and P2 for equality. Return true if
199 they are equal, and false otherwise. */
201 static int
202 vn_reference_op_eq (const void *p1, const void *p2)
204 const_vn_reference_op_t const vro1 = (const_vn_reference_op_t) p1;
205 const_vn_reference_op_t const vro2 = (const_vn_reference_op_t) p2;
207 return (vro1->opcode == vro2->opcode
208 /* We do not care for differences in type qualification. */
209 && (vro1->type == vro2->type
210 || (vro1->type && vro2->type
211 && types_compatible_p (TYPE_MAIN_VARIANT (vro1->type),
212 TYPE_MAIN_VARIANT (vro2->type))))
213 && expressions_equal_p (vro1->op0, vro2->op0)
214 && expressions_equal_p (vro1->op1, vro2->op1)
215 && expressions_equal_p (vro1->op2, vro2->op2));
218 /* Free a reference operation structure VP. */
220 static inline void
221 free_reference (vn_reference_s *vr)
223 vr->operands.release ();
227 /* vn_reference hashtable helpers. */
229 struct vn_reference_hasher : nofree_ptr_hash <vn_reference_s>
231 static inline hashval_t hash (const vn_reference_s *);
232 static inline bool equal (const vn_reference_s *, const vn_reference_s *);
235 /* Return the hashcode for a given reference operation P1. */
237 inline hashval_t
238 vn_reference_hasher::hash (const vn_reference_s *vr1)
240 return vr1->hashcode;
243 inline bool
244 vn_reference_hasher::equal (const vn_reference_s *v, const vn_reference_s *c)
246 return v == c || vn_reference_eq (v, c);
249 typedef hash_table<vn_reference_hasher> vn_reference_table_type;
250 typedef vn_reference_table_type::iterator vn_reference_iterator_type;
253 /* The set of VN hashtables. */
255 typedef struct vn_tables_s
257 vn_nary_op_table_type *nary;
258 vn_phi_table_type *phis;
259 vn_reference_table_type *references;
260 } *vn_tables_t;
263 /* vn_constant hashtable helpers. */
265 struct vn_constant_hasher : free_ptr_hash <vn_constant_s>
267 static inline hashval_t hash (const vn_constant_s *);
268 static inline bool equal (const vn_constant_s *, const vn_constant_s *);
271 /* Hash table hash function for vn_constant_t. */
273 inline hashval_t
274 vn_constant_hasher::hash (const vn_constant_s *vc1)
276 return vc1->hashcode;
279 /* Hash table equality function for vn_constant_t. */
281 inline bool
282 vn_constant_hasher::equal (const vn_constant_s *vc1, const vn_constant_s *vc2)
284 if (vc1->hashcode != vc2->hashcode)
285 return false;
287 return vn_constant_eq_with_type (vc1->constant, vc2->constant);
290 static hash_table<vn_constant_hasher> *constant_to_value_id;
291 static bitmap constant_value_ids;
294 /* Obstack we allocate the vn-tables elements from. */
295 static obstack vn_tables_obstack;
296 /* Special obstack we never unwind. */
297 static obstack vn_tables_insert_obstack;
299 static vn_reference_t last_inserted_ref;
300 static vn_phi_t last_inserted_phi;
301 static vn_nary_op_t last_inserted_nary;
303 /* Valid hashtables storing information we have proven to be
304 correct. */
305 static vn_tables_t valid_info;
308 /* Valueization hook. Valueize NAME if it is an SSA name, otherwise
309 just return it. */
310 tree (*vn_valueize) (tree);
311 tree vn_valueize_wrapper (tree t, void* context ATTRIBUTE_UNUSED)
313 return vn_valueize (t);
317 /* This represents the top of the VN lattice, which is the universal
318 value. */
320 tree VN_TOP;
322 /* Unique counter for our value ids. */
324 static unsigned int next_value_id;
327 /* Table of vn_ssa_aux_t's, one per ssa_name. The vn_ssa_aux_t objects
328 are allocated on an obstack for locality reasons, and to free them
329 without looping over the vec. */
331 struct vn_ssa_aux_hasher : typed_noop_remove <vn_ssa_aux_t>
333 typedef vn_ssa_aux_t value_type;
334 typedef tree compare_type;
335 static inline hashval_t hash (const value_type &);
336 static inline bool equal (const value_type &, const compare_type &);
337 static inline void mark_deleted (value_type &) {}
338 static const bool empty_zero_p = true;
339 static inline void mark_empty (value_type &e) { e = NULL; }
340 static inline bool is_deleted (value_type &) { return false; }
341 static inline bool is_empty (value_type &e) { return e == NULL; }
344 hashval_t
345 vn_ssa_aux_hasher::hash (const value_type &entry)
347 return SSA_NAME_VERSION (entry->name);
350 bool
351 vn_ssa_aux_hasher::equal (const value_type &entry, const compare_type &name)
353 return name == entry->name;
356 static hash_table<vn_ssa_aux_hasher> *vn_ssa_aux_hash;
357 typedef hash_table<vn_ssa_aux_hasher>::iterator vn_ssa_aux_iterator_type;
358 static struct obstack vn_ssa_aux_obstack;
360 static vn_nary_op_t vn_nary_op_insert_stmt (gimple *, tree);
361 static unsigned int vn_nary_length_from_stmt (gimple *);
362 static vn_nary_op_t alloc_vn_nary_op_noinit (unsigned int, obstack *);
363 static vn_nary_op_t vn_nary_op_insert_into (vn_nary_op_t,
364 vn_nary_op_table_type *, bool);
365 static void init_vn_nary_op_from_stmt (vn_nary_op_t, gimple *);
366 static void init_vn_nary_op_from_pieces (vn_nary_op_t, unsigned int,
367 enum tree_code, tree, tree *);
368 static tree vn_lookup_simplify_result (gimple_match_op *);
369 static vn_reference_t vn_reference_lookup_or_insert_for_pieces
370 (tree, alias_set_type, alias_set_type, tree,
371 vec<vn_reference_op_s, va_heap>, tree);
373 /* Return whether there is value numbering information for a given SSA name. */
375 bool
376 has_VN_INFO (tree name)
378 return vn_ssa_aux_hash->find_with_hash (name, SSA_NAME_VERSION (name));
381 vn_ssa_aux_t
382 VN_INFO (tree name)
384 vn_ssa_aux_t *res
385 = vn_ssa_aux_hash->find_slot_with_hash (name, SSA_NAME_VERSION (name),
386 INSERT);
387 if (*res != NULL)
388 return *res;
390 vn_ssa_aux_t newinfo = *res = XOBNEW (&vn_ssa_aux_obstack, struct vn_ssa_aux);
391 memset (newinfo, 0, sizeof (struct vn_ssa_aux));
392 newinfo->name = name;
393 newinfo->valnum = VN_TOP;
394 /* We are using the visited flag to handle uses with defs not within the
395 region being value-numbered. */
396 newinfo->visited = false;
398 /* Given we create the VN_INFOs on-demand now we have to do initialization
399 different than VN_TOP here. */
400 if (SSA_NAME_IS_DEFAULT_DEF (name))
401 switch (TREE_CODE (SSA_NAME_VAR (name)))
403 case VAR_DECL:
404 /* All undefined vars are VARYING. */
405 newinfo->valnum = name;
406 newinfo->visited = true;
407 break;
409 case PARM_DECL:
410 /* Parameters are VARYING but we can record a condition
411 if we know it is a non-NULL pointer. */
412 newinfo->visited = true;
413 newinfo->valnum = name;
414 if (POINTER_TYPE_P (TREE_TYPE (name))
415 && nonnull_arg_p (SSA_NAME_VAR (name)))
417 tree ops[2];
418 ops[0] = name;
419 ops[1] = build_int_cst (TREE_TYPE (name), 0);
420 vn_nary_op_t nary;
421 /* Allocate from non-unwinding stack. */
422 nary = alloc_vn_nary_op_noinit (2, &vn_tables_insert_obstack);
423 init_vn_nary_op_from_pieces (nary, 2, NE_EXPR,
424 boolean_type_node, ops);
425 nary->predicated_values = 0;
426 nary->u.result = boolean_true_node;
427 vn_nary_op_insert_into (nary, valid_info->nary, true);
428 gcc_assert (nary->unwind_to == NULL);
429 /* Also do not link it into the undo chain. */
430 last_inserted_nary = nary->next;
431 nary->next = (vn_nary_op_t)(void *)-1;
432 nary = alloc_vn_nary_op_noinit (2, &vn_tables_insert_obstack);
433 init_vn_nary_op_from_pieces (nary, 2, EQ_EXPR,
434 boolean_type_node, ops);
435 nary->predicated_values = 0;
436 nary->u.result = boolean_false_node;
437 vn_nary_op_insert_into (nary, valid_info->nary, true);
438 gcc_assert (nary->unwind_to == NULL);
439 last_inserted_nary = nary->next;
440 nary->next = (vn_nary_op_t)(void *)-1;
441 if (dump_file && (dump_flags & TDF_DETAILS))
443 fprintf (dump_file, "Recording ");
444 print_generic_expr (dump_file, name, TDF_SLIM);
445 fprintf (dump_file, " != 0\n");
448 break;
450 case RESULT_DECL:
451 /* If the result is passed by invisible reference the default
452 def is initialized, otherwise it's uninitialized. Still
453 undefined is varying. */
454 newinfo->visited = true;
455 newinfo->valnum = name;
456 break;
458 default:
459 gcc_unreachable ();
461 return newinfo;
464 /* Return the SSA value of X. */
466 inline tree
467 SSA_VAL (tree x, bool *visited = NULL)
469 vn_ssa_aux_t tem = vn_ssa_aux_hash->find_with_hash (x, SSA_NAME_VERSION (x));
470 if (visited)
471 *visited = tem && tem->visited;
472 return tem && tem->visited ? tem->valnum : x;
475 /* Return the SSA value of the VUSE x, supporting released VDEFs
476 during elimination which will value-number the VDEF to the
477 associated VUSE (but not substitute in the whole lattice). */
479 static inline tree
480 vuse_ssa_val (tree x)
482 if (!x)
483 return NULL_TREE;
487 x = SSA_VAL (x);
488 gcc_assert (x != VN_TOP);
490 while (SSA_NAME_IN_FREE_LIST (x));
492 return x;
495 /* Similar to the above but used as callback for walk_non_aliases_vuses
496 and thus should stop at unvisited VUSE to not walk across region
497 boundaries. */
499 static tree
500 vuse_valueize (tree vuse)
504 bool visited;
505 vuse = SSA_VAL (vuse, &visited);
506 if (!visited)
507 return NULL_TREE;
508 gcc_assert (vuse != VN_TOP);
510 while (SSA_NAME_IN_FREE_LIST (vuse));
511 return vuse;
515 /* Return the vn_kind the expression computed by the stmt should be
516 associated with. */
518 enum vn_kind
519 vn_get_stmt_kind (gimple *stmt)
521 switch (gimple_code (stmt))
523 case GIMPLE_CALL:
524 return VN_REFERENCE;
525 case GIMPLE_PHI:
526 return VN_PHI;
527 case GIMPLE_ASSIGN:
529 enum tree_code code = gimple_assign_rhs_code (stmt);
530 tree rhs1 = gimple_assign_rhs1 (stmt);
531 switch (get_gimple_rhs_class (code))
533 case GIMPLE_UNARY_RHS:
534 case GIMPLE_BINARY_RHS:
535 case GIMPLE_TERNARY_RHS:
536 return VN_NARY;
537 case GIMPLE_SINGLE_RHS:
538 switch (TREE_CODE_CLASS (code))
540 case tcc_reference:
541 /* VOP-less references can go through unary case. */
542 if ((code == REALPART_EXPR
543 || code == IMAGPART_EXPR
544 || code == VIEW_CONVERT_EXPR
545 || code == BIT_FIELD_REF)
546 && TREE_CODE (TREE_OPERAND (rhs1, 0)) == SSA_NAME)
547 return VN_NARY;
549 /* Fallthrough. */
550 case tcc_declaration:
551 return VN_REFERENCE;
553 case tcc_constant:
554 return VN_CONSTANT;
556 default:
557 if (code == ADDR_EXPR)
558 return (is_gimple_min_invariant (rhs1)
559 ? VN_CONSTANT : VN_REFERENCE);
560 else if (code == CONSTRUCTOR)
561 return VN_NARY;
562 return VN_NONE;
564 default:
565 return VN_NONE;
568 default:
569 return VN_NONE;
573 /* Lookup a value id for CONSTANT and return it. If it does not
574 exist returns 0. */
576 unsigned int
577 get_constant_value_id (tree constant)
579 vn_constant_s **slot;
580 struct vn_constant_s vc;
582 vc.hashcode = vn_hash_constant_with_type (constant);
583 vc.constant = constant;
584 slot = constant_to_value_id->find_slot (&vc, NO_INSERT);
585 if (slot)
586 return (*slot)->value_id;
587 return 0;
590 /* Lookup a value id for CONSTANT, and if it does not exist, create a
591 new one and return it. If it does exist, return it. */
593 unsigned int
594 get_or_alloc_constant_value_id (tree constant)
596 vn_constant_s **slot;
597 struct vn_constant_s vc;
598 vn_constant_t vcp;
600 /* If the hashtable isn't initialized we're not running from PRE and thus
601 do not need value-ids. */
602 if (!constant_to_value_id)
603 return 0;
605 vc.hashcode = vn_hash_constant_with_type (constant);
606 vc.constant = constant;
607 slot = constant_to_value_id->find_slot (&vc, INSERT);
608 if (*slot)
609 return (*slot)->value_id;
611 vcp = XNEW (struct vn_constant_s);
612 vcp->hashcode = vc.hashcode;
613 vcp->constant = constant;
614 vcp->value_id = get_next_value_id ();
615 *slot = vcp;
616 bitmap_set_bit (constant_value_ids, vcp->value_id);
617 return vcp->value_id;
620 /* Return true if V is a value id for a constant. */
622 bool
623 value_id_constant_p (unsigned int v)
625 return bitmap_bit_p (constant_value_ids, v);
628 /* Compute the hash for a reference operand VRO1. */
630 static void
631 vn_reference_op_compute_hash (const vn_reference_op_t vro1, inchash::hash &hstate)
633 hstate.add_int (vro1->opcode);
634 if (vro1->op0)
635 inchash::add_expr (vro1->op0, hstate);
636 if (vro1->op1)
637 inchash::add_expr (vro1->op1, hstate);
638 if (vro1->op2)
639 inchash::add_expr (vro1->op2, hstate);
642 /* Compute a hash for the reference operation VR1 and return it. */
644 static hashval_t
645 vn_reference_compute_hash (const vn_reference_t vr1)
647 inchash::hash hstate;
648 hashval_t result;
649 int i;
650 vn_reference_op_t vro;
651 poly_int64 off = -1;
652 bool deref = false;
654 FOR_EACH_VEC_ELT (vr1->operands, i, vro)
656 if (vro->opcode == MEM_REF)
657 deref = true;
658 else if (vro->opcode != ADDR_EXPR)
659 deref = false;
660 if (maybe_ne (vro->off, -1))
662 if (known_eq (off, -1))
663 off = 0;
664 off += vro->off;
666 else
668 if (maybe_ne (off, -1)
669 && maybe_ne (off, 0))
670 hstate.add_poly_int (off);
671 off = -1;
672 if (deref
673 && vro->opcode == ADDR_EXPR)
675 if (vro->op0)
677 tree op = TREE_OPERAND (vro->op0, 0);
678 hstate.add_int (TREE_CODE (op));
679 inchash::add_expr (op, hstate);
682 else
683 vn_reference_op_compute_hash (vro, hstate);
686 result = hstate.end ();
687 /* ??? We would ICE later if we hash instead of adding that in. */
688 if (vr1->vuse)
689 result += SSA_NAME_VERSION (vr1->vuse);
691 return result;
694 /* Return true if reference operations VR1 and VR2 are equivalent. This
695 means they have the same set of operands and vuses. */
697 bool
698 vn_reference_eq (const_vn_reference_t const vr1, const_vn_reference_t const vr2)
700 unsigned i, j;
702 /* Early out if this is not a hash collision. */
703 if (vr1->hashcode != vr2->hashcode)
704 return false;
706 /* The VOP needs to be the same. */
707 if (vr1->vuse != vr2->vuse)
708 return false;
710 /* If the operands are the same we are done. */
711 if (vr1->operands == vr2->operands)
712 return true;
714 if (!expressions_equal_p (TYPE_SIZE (vr1->type), TYPE_SIZE (vr2->type)))
715 return false;
717 if (INTEGRAL_TYPE_P (vr1->type)
718 && INTEGRAL_TYPE_P (vr2->type))
720 if (TYPE_PRECISION (vr1->type) != TYPE_PRECISION (vr2->type))
721 return false;
723 else if (INTEGRAL_TYPE_P (vr1->type)
724 && (TYPE_PRECISION (vr1->type)
725 != TREE_INT_CST_LOW (TYPE_SIZE (vr1->type))))
726 return false;
727 else if (INTEGRAL_TYPE_P (vr2->type)
728 && (TYPE_PRECISION (vr2->type)
729 != TREE_INT_CST_LOW (TYPE_SIZE (vr2->type))))
730 return false;
732 i = 0;
733 j = 0;
736 poly_int64 off1 = 0, off2 = 0;
737 vn_reference_op_t vro1, vro2;
738 vn_reference_op_s tem1, tem2;
739 bool deref1 = false, deref2 = false;
740 for (; vr1->operands.iterate (i, &vro1); i++)
742 if (vro1->opcode == MEM_REF)
743 deref1 = true;
744 /* Do not look through a storage order barrier. */
745 else if (vro1->opcode == VIEW_CONVERT_EXPR && vro1->reverse)
746 return false;
747 if (known_eq (vro1->off, -1))
748 break;
749 off1 += vro1->off;
751 for (; vr2->operands.iterate (j, &vro2); j++)
753 if (vro2->opcode == MEM_REF)
754 deref2 = true;
755 /* Do not look through a storage order barrier. */
756 else if (vro2->opcode == VIEW_CONVERT_EXPR && vro2->reverse)
757 return false;
758 if (known_eq (vro2->off, -1))
759 break;
760 off2 += vro2->off;
762 if (maybe_ne (off1, off2))
763 return false;
764 if (deref1 && vro1->opcode == ADDR_EXPR)
766 memset (&tem1, 0, sizeof (tem1));
767 tem1.op0 = TREE_OPERAND (vro1->op0, 0);
768 tem1.type = TREE_TYPE (tem1.op0);
769 tem1.opcode = TREE_CODE (tem1.op0);
770 vro1 = &tem1;
771 deref1 = false;
773 if (deref2 && vro2->opcode == ADDR_EXPR)
775 memset (&tem2, 0, sizeof (tem2));
776 tem2.op0 = TREE_OPERAND (vro2->op0, 0);
777 tem2.type = TREE_TYPE (tem2.op0);
778 tem2.opcode = TREE_CODE (tem2.op0);
779 vro2 = &tem2;
780 deref2 = false;
782 if (deref1 != deref2)
783 return false;
784 if (!vn_reference_op_eq (vro1, vro2))
785 return false;
786 ++j;
787 ++i;
789 while (vr1->operands.length () != i
790 || vr2->operands.length () != j);
792 return true;
795 /* Copy the operations present in load/store REF into RESULT, a vector of
796 vn_reference_op_s's. */
798 static void
799 copy_reference_ops_from_ref (tree ref, vec<vn_reference_op_s> *result)
801 /* For non-calls, store the information that makes up the address. */
802 tree orig = ref;
803 while (ref)
805 vn_reference_op_s temp;
807 memset (&temp, 0, sizeof (temp));
808 temp.type = TREE_TYPE (ref);
809 temp.opcode = TREE_CODE (ref);
810 temp.off = -1;
812 switch (temp.opcode)
814 case MODIFY_EXPR:
815 temp.op0 = TREE_OPERAND (ref, 1);
816 break;
817 case WITH_SIZE_EXPR:
818 temp.op0 = TREE_OPERAND (ref, 1);
819 temp.off = 0;
820 break;
821 case MEM_REF:
822 /* The base address gets its own vn_reference_op_s structure. */
823 temp.op0 = TREE_OPERAND (ref, 1);
824 if (!mem_ref_offset (ref).to_shwi (&temp.off))
825 temp.off = -1;
826 temp.clique = MR_DEPENDENCE_CLIQUE (ref);
827 temp.base = MR_DEPENDENCE_BASE (ref);
828 temp.reverse = REF_REVERSE_STORAGE_ORDER (ref);
829 break;
830 case TARGET_MEM_REF:
831 /* The base address gets its own vn_reference_op_s structure. */
832 temp.op0 = TMR_INDEX (ref);
833 temp.op1 = TMR_STEP (ref);
834 temp.op2 = TMR_OFFSET (ref);
835 temp.clique = MR_DEPENDENCE_CLIQUE (ref);
836 temp.base = MR_DEPENDENCE_BASE (ref);
837 result->safe_push (temp);
838 memset (&temp, 0, sizeof (temp));
839 temp.type = NULL_TREE;
840 temp.opcode = ERROR_MARK;
841 temp.op0 = TMR_INDEX2 (ref);
842 temp.off = -1;
843 break;
844 case BIT_FIELD_REF:
845 /* Record bits, position and storage order. */
846 temp.op0 = TREE_OPERAND (ref, 1);
847 temp.op1 = TREE_OPERAND (ref, 2);
848 if (!multiple_p (bit_field_offset (ref), BITS_PER_UNIT, &temp.off))
849 temp.off = -1;
850 temp.reverse = REF_REVERSE_STORAGE_ORDER (ref);
851 break;
852 case COMPONENT_REF:
853 /* The field decl is enough to unambiguously specify the field,
854 a matching type is not necessary and a mismatching type
855 is always a spurious difference. */
856 temp.type = NULL_TREE;
857 temp.op0 = TREE_OPERAND (ref, 1);
858 temp.op1 = TREE_OPERAND (ref, 2);
860 tree this_offset = component_ref_field_offset (ref);
861 if (this_offset
862 && poly_int_tree_p (this_offset))
864 tree bit_offset = DECL_FIELD_BIT_OFFSET (TREE_OPERAND (ref, 1));
865 if (TREE_INT_CST_LOW (bit_offset) % BITS_PER_UNIT == 0)
867 poly_offset_int off
868 = (wi::to_poly_offset (this_offset)
869 + (wi::to_offset (bit_offset) >> LOG2_BITS_PER_UNIT));
870 /* Probibit value-numbering zero offset components
871 of addresses the same before the pass folding
872 __builtin_object_size had a chance to run
873 (checking cfun->after_inlining does the
874 trick here). */
875 if (TREE_CODE (orig) != ADDR_EXPR
876 || maybe_ne (off, 0)
877 || cfun->after_inlining)
878 off.to_shwi (&temp.off);
882 break;
883 case ARRAY_RANGE_REF:
884 case ARRAY_REF:
886 tree eltype = TREE_TYPE (TREE_TYPE (TREE_OPERAND (ref, 0)));
887 /* Record index as operand. */
888 temp.op0 = TREE_OPERAND (ref, 1);
889 /* Always record lower bounds and element size. */
890 temp.op1 = array_ref_low_bound (ref);
891 /* But record element size in units of the type alignment. */
892 temp.op2 = TREE_OPERAND (ref, 3);
893 temp.align = eltype->type_common.align;
894 if (! temp.op2)
895 temp.op2 = size_binop (EXACT_DIV_EXPR, TYPE_SIZE_UNIT (eltype),
896 size_int (TYPE_ALIGN_UNIT (eltype)));
897 if (poly_int_tree_p (temp.op0)
898 && poly_int_tree_p (temp.op1)
899 && TREE_CODE (temp.op2) == INTEGER_CST)
901 poly_offset_int off = ((wi::to_poly_offset (temp.op0)
902 - wi::to_poly_offset (temp.op1))
903 * wi::to_offset (temp.op2)
904 * vn_ref_op_align_unit (&temp));
905 off.to_shwi (&temp.off);
908 break;
909 case VAR_DECL:
910 if (DECL_HARD_REGISTER (ref))
912 temp.op0 = ref;
913 break;
915 /* Fallthru. */
916 case PARM_DECL:
917 case CONST_DECL:
918 case RESULT_DECL:
919 /* Canonicalize decls to MEM[&decl] which is what we end up with
920 when valueizing MEM[ptr] with ptr = &decl. */
921 temp.opcode = MEM_REF;
922 temp.op0 = build_int_cst (build_pointer_type (TREE_TYPE (ref)), 0);
923 temp.off = 0;
924 result->safe_push (temp);
925 temp.opcode = ADDR_EXPR;
926 temp.op0 = build1 (ADDR_EXPR, TREE_TYPE (temp.op0), ref);
927 temp.type = TREE_TYPE (temp.op0);
928 temp.off = -1;
929 break;
930 case STRING_CST:
931 case INTEGER_CST:
932 case POLY_INT_CST:
933 case COMPLEX_CST:
934 case VECTOR_CST:
935 case REAL_CST:
936 case FIXED_CST:
937 case CONSTRUCTOR:
938 case SSA_NAME:
939 temp.op0 = ref;
940 break;
941 case ADDR_EXPR:
942 if (is_gimple_min_invariant (ref))
944 temp.op0 = ref;
945 break;
947 break;
948 /* These are only interesting for their operands, their
949 existence, and their type. They will never be the last
950 ref in the chain of references (IE they require an
951 operand), so we don't have to put anything
952 for op* as it will be handled by the iteration */
953 case REALPART_EXPR:
954 temp.off = 0;
955 break;
956 case VIEW_CONVERT_EXPR:
957 temp.off = 0;
958 temp.reverse = storage_order_barrier_p (ref);
959 break;
960 case IMAGPART_EXPR:
961 /* This is only interesting for its constant offset. */
962 temp.off = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (TREE_TYPE (ref)));
963 break;
964 default:
965 gcc_unreachable ();
967 result->safe_push (temp);
969 if (REFERENCE_CLASS_P (ref)
970 || TREE_CODE (ref) == MODIFY_EXPR
971 || TREE_CODE (ref) == WITH_SIZE_EXPR
972 || (TREE_CODE (ref) == ADDR_EXPR
973 && !is_gimple_min_invariant (ref)))
974 ref = TREE_OPERAND (ref, 0);
975 else
976 ref = NULL_TREE;
980 /* Build a alias-oracle reference abstraction in *REF from the vn_reference
981 operands in *OPS, the reference alias set SET and the reference type TYPE.
982 Return true if something useful was produced. */
984 bool
985 ao_ref_init_from_vn_reference (ao_ref *ref,
986 alias_set_type set, alias_set_type base_set,
987 tree type, vec<vn_reference_op_s> ops)
989 vn_reference_op_t op;
990 unsigned i;
991 tree base = NULL_TREE;
992 tree *op0_p = &base;
993 poly_offset_int offset = 0;
994 poly_offset_int max_size;
995 poly_offset_int size = -1;
996 tree size_tree = NULL_TREE;
998 /* First get the final access size from just the outermost expression. */
999 op = &ops[0];
1000 if (op->opcode == COMPONENT_REF)
1001 size_tree = DECL_SIZE (op->op0);
1002 else if (op->opcode == BIT_FIELD_REF)
1003 size_tree = op->op0;
1004 else
1006 machine_mode mode = TYPE_MODE (type);
1007 if (mode == BLKmode)
1008 size_tree = TYPE_SIZE (type);
1009 else
1010 size = GET_MODE_BITSIZE (mode);
1012 if (size_tree != NULL_TREE
1013 && poly_int_tree_p (size_tree))
1014 size = wi::to_poly_offset (size_tree);
1016 /* Initially, maxsize is the same as the accessed element size.
1017 In the following it will only grow (or become -1). */
1018 max_size = size;
1020 /* Compute cumulative bit-offset for nested component-refs and array-refs,
1021 and find the ultimate containing object. */
1022 FOR_EACH_VEC_ELT (ops, i, op)
1024 switch (op->opcode)
1026 /* These may be in the reference ops, but we cannot do anything
1027 sensible with them here. */
1028 case ADDR_EXPR:
1029 /* Apart from ADDR_EXPR arguments to MEM_REF. */
1030 if (base != NULL_TREE
1031 && TREE_CODE (base) == MEM_REF
1032 && op->op0
1033 && DECL_P (TREE_OPERAND (op->op0, 0)))
1035 vn_reference_op_t pop = &ops[i-1];
1036 base = TREE_OPERAND (op->op0, 0);
1037 if (known_eq (pop->off, -1))
1039 max_size = -1;
1040 offset = 0;
1042 else
1043 offset += pop->off * BITS_PER_UNIT;
1044 op0_p = NULL;
1045 break;
1047 /* Fallthru. */
1048 case CALL_EXPR:
1049 return false;
1051 /* Record the base objects. */
1052 case MEM_REF:
1053 *op0_p = build2 (MEM_REF, op->type,
1054 NULL_TREE, op->op0);
1055 MR_DEPENDENCE_CLIQUE (*op0_p) = op->clique;
1056 MR_DEPENDENCE_BASE (*op0_p) = op->base;
1057 op0_p = &TREE_OPERAND (*op0_p, 0);
1058 break;
1060 case VAR_DECL:
1061 case PARM_DECL:
1062 case RESULT_DECL:
1063 case SSA_NAME:
1064 *op0_p = op->op0;
1065 op0_p = NULL;
1066 break;
1068 /* And now the usual component-reference style ops. */
1069 case BIT_FIELD_REF:
1070 offset += wi::to_poly_offset (op->op1);
1071 break;
1073 case COMPONENT_REF:
1075 tree field = op->op0;
1076 /* We do not have a complete COMPONENT_REF tree here so we
1077 cannot use component_ref_field_offset. Do the interesting
1078 parts manually. */
1079 tree this_offset = DECL_FIELD_OFFSET (field);
1081 if (op->op1 || !poly_int_tree_p (this_offset))
1082 max_size = -1;
1083 else
1085 poly_offset_int woffset = (wi::to_poly_offset (this_offset)
1086 << LOG2_BITS_PER_UNIT);
1087 woffset += wi::to_offset (DECL_FIELD_BIT_OFFSET (field));
1088 offset += woffset;
1090 break;
1093 case ARRAY_RANGE_REF:
1094 case ARRAY_REF:
1095 /* We recorded the lower bound and the element size. */
1096 if (!poly_int_tree_p (op->op0)
1097 || !poly_int_tree_p (op->op1)
1098 || TREE_CODE (op->op2) != INTEGER_CST)
1099 max_size = -1;
1100 else
1102 poly_offset_int woffset
1103 = wi::sext (wi::to_poly_offset (op->op0)
1104 - wi::to_poly_offset (op->op1),
1105 TYPE_PRECISION (TREE_TYPE (op->op0)));
1106 woffset *= wi::to_offset (op->op2) * vn_ref_op_align_unit (op);
1107 woffset <<= LOG2_BITS_PER_UNIT;
1108 offset += woffset;
1110 break;
1112 case REALPART_EXPR:
1113 break;
1115 case IMAGPART_EXPR:
1116 offset += size;
1117 break;
1119 case VIEW_CONVERT_EXPR:
1120 break;
1122 case STRING_CST:
1123 case INTEGER_CST:
1124 case COMPLEX_CST:
1125 case VECTOR_CST:
1126 case REAL_CST:
1127 case CONSTRUCTOR:
1128 case CONST_DECL:
1129 return false;
1131 default:
1132 return false;
1136 if (base == NULL_TREE)
1137 return false;
1139 ref->ref = NULL_TREE;
1140 ref->base = base;
1141 ref->ref_alias_set = set;
1142 ref->base_alias_set = base_set;
1143 /* We discount volatiles from value-numbering elsewhere. */
1144 ref->volatile_p = false;
1146 if (!size.to_shwi (&ref->size) || maybe_lt (ref->size, 0))
1148 ref->offset = 0;
1149 ref->size = -1;
1150 ref->max_size = -1;
1151 return true;
1154 if (!offset.to_shwi (&ref->offset))
1156 ref->offset = 0;
1157 ref->max_size = -1;
1158 return true;
1161 if (!max_size.to_shwi (&ref->max_size) || maybe_lt (ref->max_size, 0))
1162 ref->max_size = -1;
1164 return true;
1167 /* Copy the operations present in load/store/call REF into RESULT, a vector of
1168 vn_reference_op_s's. */
1170 static void
1171 copy_reference_ops_from_call (gcall *call,
1172 vec<vn_reference_op_s> *result)
1174 vn_reference_op_s temp;
1175 unsigned i;
1176 tree lhs = gimple_call_lhs (call);
1177 int lr;
1179 /* If 2 calls have a different non-ssa lhs, vdef value numbers should be
1180 different. By adding the lhs here in the vector, we ensure that the
1181 hashcode is different, guaranteeing a different value number. */
1182 if (lhs && TREE_CODE (lhs) != SSA_NAME)
1184 memset (&temp, 0, sizeof (temp));
1185 temp.opcode = MODIFY_EXPR;
1186 temp.type = TREE_TYPE (lhs);
1187 temp.op0 = lhs;
1188 temp.off = -1;
1189 result->safe_push (temp);
1192 /* Copy the type, opcode, function, static chain and EH region, if any. */
1193 memset (&temp, 0, sizeof (temp));
1194 temp.type = gimple_call_fntype (call);
1195 temp.opcode = CALL_EXPR;
1196 temp.op0 = gimple_call_fn (call);
1197 temp.op1 = gimple_call_chain (call);
1198 if (stmt_could_throw_p (cfun, call) && (lr = lookup_stmt_eh_lp (call)) > 0)
1199 temp.op2 = size_int (lr);
1200 temp.off = -1;
1201 result->safe_push (temp);
1203 /* Copy the call arguments. As they can be references as well,
1204 just chain them together. */
1205 for (i = 0; i < gimple_call_num_args (call); ++i)
1207 tree callarg = gimple_call_arg (call, i);
1208 copy_reference_ops_from_ref (callarg, result);
1212 /* Fold *& at position *I_P in a vn_reference_op_s vector *OPS. Updates
1213 *I_P to point to the last element of the replacement. */
1214 static bool
1215 vn_reference_fold_indirect (vec<vn_reference_op_s> *ops,
1216 unsigned int *i_p)
1218 unsigned int i = *i_p;
1219 vn_reference_op_t op = &(*ops)[i];
1220 vn_reference_op_t mem_op = &(*ops)[i - 1];
1221 tree addr_base;
1222 poly_int64 addr_offset = 0;
1224 /* The only thing we have to do is from &OBJ.foo.bar add the offset
1225 from .foo.bar to the preceding MEM_REF offset and replace the
1226 address with &OBJ. */
1227 addr_base = get_addr_base_and_unit_offset_1 (TREE_OPERAND (op->op0, 0),
1228 &addr_offset, vn_valueize);
1229 gcc_checking_assert (addr_base && TREE_CODE (addr_base) != MEM_REF);
1230 if (addr_base != TREE_OPERAND (op->op0, 0))
1232 poly_offset_int off
1233 = (poly_offset_int::from (wi::to_poly_wide (mem_op->op0),
1234 SIGNED)
1235 + addr_offset);
1236 mem_op->op0 = wide_int_to_tree (TREE_TYPE (mem_op->op0), off);
1237 op->op0 = build_fold_addr_expr (addr_base);
1238 if (tree_fits_shwi_p (mem_op->op0))
1239 mem_op->off = tree_to_shwi (mem_op->op0);
1240 else
1241 mem_op->off = -1;
1242 return true;
1244 return false;
1247 /* Fold *& at position *I_P in a vn_reference_op_s vector *OPS. Updates
1248 *I_P to point to the last element of the replacement. */
1249 static bool
1250 vn_reference_maybe_forwprop_address (vec<vn_reference_op_s> *ops,
1251 unsigned int *i_p)
1253 bool changed = false;
1254 vn_reference_op_t op;
1258 unsigned int i = *i_p;
1259 op = &(*ops)[i];
1260 vn_reference_op_t mem_op = &(*ops)[i - 1];
1261 gimple *def_stmt;
1262 enum tree_code code;
1263 poly_offset_int off;
1265 def_stmt = SSA_NAME_DEF_STMT (op->op0);
1266 if (!is_gimple_assign (def_stmt))
1267 return changed;
1269 code = gimple_assign_rhs_code (def_stmt);
1270 if (code != ADDR_EXPR
1271 && code != POINTER_PLUS_EXPR)
1272 return changed;
1274 off = poly_offset_int::from (wi::to_poly_wide (mem_op->op0), SIGNED);
1276 /* The only thing we have to do is from &OBJ.foo.bar add the offset
1277 from .foo.bar to the preceding MEM_REF offset and replace the
1278 address with &OBJ. */
1279 if (code == ADDR_EXPR)
1281 tree addr, addr_base;
1282 poly_int64 addr_offset;
1284 addr = gimple_assign_rhs1 (def_stmt);
1285 addr_base = get_addr_base_and_unit_offset_1 (TREE_OPERAND (addr, 0),
1286 &addr_offset,
1287 vn_valueize);
1288 /* If that didn't work because the address isn't invariant propagate
1289 the reference tree from the address operation in case the current
1290 dereference isn't offsetted. */
1291 if (!addr_base
1292 && *i_p == ops->length () - 1
1293 && known_eq (off, 0)
1294 /* This makes us disable this transform for PRE where the
1295 reference ops might be also used for code insertion which
1296 is invalid. */
1297 && default_vn_walk_kind == VN_WALKREWRITE)
1299 auto_vec<vn_reference_op_s, 32> tem;
1300 copy_reference_ops_from_ref (TREE_OPERAND (addr, 0), &tem);
1301 /* Make sure to preserve TBAA info. The only objects not
1302 wrapped in MEM_REFs that can have their address taken are
1303 STRING_CSTs. */
1304 if (tem.length () >= 2
1305 && tem[tem.length () - 2].opcode == MEM_REF)
1307 vn_reference_op_t new_mem_op = &tem[tem.length () - 2];
1308 new_mem_op->op0
1309 = wide_int_to_tree (TREE_TYPE (mem_op->op0),
1310 wi::to_poly_wide (new_mem_op->op0));
1312 else
1313 gcc_assert (tem.last ().opcode == STRING_CST);
1314 ops->pop ();
1315 ops->pop ();
1316 ops->safe_splice (tem);
1317 --*i_p;
1318 return true;
1320 if (!addr_base
1321 || TREE_CODE (addr_base) != MEM_REF
1322 || (TREE_CODE (TREE_OPERAND (addr_base, 0)) == SSA_NAME
1323 && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (TREE_OPERAND (addr_base,
1324 0))))
1325 return changed;
1327 off += addr_offset;
1328 off += mem_ref_offset (addr_base);
1329 op->op0 = TREE_OPERAND (addr_base, 0);
1331 else
1333 tree ptr, ptroff;
1334 ptr = gimple_assign_rhs1 (def_stmt);
1335 ptroff = gimple_assign_rhs2 (def_stmt);
1336 if (TREE_CODE (ptr) != SSA_NAME
1337 || SSA_NAME_OCCURS_IN_ABNORMAL_PHI (ptr)
1338 /* Make sure to not endlessly recurse.
1339 See gcc.dg/tree-ssa/20040408-1.c for an example. Can easily
1340 happen when we value-number a PHI to its backedge value. */
1341 || SSA_VAL (ptr) == op->op0
1342 || !poly_int_tree_p (ptroff))
1343 return changed;
1345 off += wi::to_poly_offset (ptroff);
1346 op->op0 = ptr;
1349 mem_op->op0 = wide_int_to_tree (TREE_TYPE (mem_op->op0), off);
1350 if (tree_fits_shwi_p (mem_op->op0))
1351 mem_op->off = tree_to_shwi (mem_op->op0);
1352 else
1353 mem_op->off = -1;
1354 /* ??? Can end up with endless recursion here!?
1355 gcc.c-torture/execute/strcmp-1.c */
1356 if (TREE_CODE (op->op0) == SSA_NAME)
1357 op->op0 = SSA_VAL (op->op0);
1358 if (TREE_CODE (op->op0) != SSA_NAME)
1359 op->opcode = TREE_CODE (op->op0);
1361 changed = true;
1363 /* Tail-recurse. */
1364 while (TREE_CODE (op->op0) == SSA_NAME);
1366 /* Fold a remaining *&. */
1367 if (TREE_CODE (op->op0) == ADDR_EXPR)
1368 vn_reference_fold_indirect (ops, i_p);
1370 return changed;
1373 /* Optimize the reference REF to a constant if possible or return
1374 NULL_TREE if not. */
1376 tree
1377 fully_constant_vn_reference_p (vn_reference_t ref)
1379 vec<vn_reference_op_s> operands = ref->operands;
1380 vn_reference_op_t op;
1382 /* Try to simplify the translated expression if it is
1383 a call to a builtin function with at most two arguments. */
1384 op = &operands[0];
1385 if (op->opcode == CALL_EXPR
1386 && TREE_CODE (op->op0) == ADDR_EXPR
1387 && TREE_CODE (TREE_OPERAND (op->op0, 0)) == FUNCTION_DECL
1388 && fndecl_built_in_p (TREE_OPERAND (op->op0, 0))
1389 && operands.length () >= 2
1390 && operands.length () <= 3)
1392 vn_reference_op_t arg0, arg1 = NULL;
1393 bool anyconst = false;
1394 arg0 = &operands[1];
1395 if (operands.length () > 2)
1396 arg1 = &operands[2];
1397 if (TREE_CODE_CLASS (arg0->opcode) == tcc_constant
1398 || (arg0->opcode == ADDR_EXPR
1399 && is_gimple_min_invariant (arg0->op0)))
1400 anyconst = true;
1401 if (arg1
1402 && (TREE_CODE_CLASS (arg1->opcode) == tcc_constant
1403 || (arg1->opcode == ADDR_EXPR
1404 && is_gimple_min_invariant (arg1->op0))))
1405 anyconst = true;
1406 if (anyconst)
1408 tree folded = build_call_expr (TREE_OPERAND (op->op0, 0),
1409 arg1 ? 2 : 1,
1410 arg0->op0,
1411 arg1 ? arg1->op0 : NULL);
1412 if (folded
1413 && TREE_CODE (folded) == NOP_EXPR)
1414 folded = TREE_OPERAND (folded, 0);
1415 if (folded
1416 && is_gimple_min_invariant (folded))
1417 return folded;
1421 /* Simplify reads from constants or constant initializers. */
1422 else if (BITS_PER_UNIT == 8
1423 && COMPLETE_TYPE_P (ref->type)
1424 && is_gimple_reg_type (ref->type))
1426 poly_int64 off = 0;
1427 HOST_WIDE_INT size;
1428 if (INTEGRAL_TYPE_P (ref->type))
1429 size = TYPE_PRECISION (ref->type);
1430 else if (tree_fits_shwi_p (TYPE_SIZE (ref->type)))
1431 size = tree_to_shwi (TYPE_SIZE (ref->type));
1432 else
1433 return NULL_TREE;
1434 if (size % BITS_PER_UNIT != 0
1435 || size > MAX_BITSIZE_MODE_ANY_MODE)
1436 return NULL_TREE;
1437 size /= BITS_PER_UNIT;
1438 unsigned i;
1439 for (i = 0; i < operands.length (); ++i)
1441 if (TREE_CODE_CLASS (operands[i].opcode) == tcc_constant)
1443 ++i;
1444 break;
1446 if (known_eq (operands[i].off, -1))
1447 return NULL_TREE;
1448 off += operands[i].off;
1449 if (operands[i].opcode == MEM_REF)
1451 ++i;
1452 break;
1455 vn_reference_op_t base = &operands[--i];
1456 tree ctor = error_mark_node;
1457 tree decl = NULL_TREE;
1458 if (TREE_CODE_CLASS (base->opcode) == tcc_constant)
1459 ctor = base->op0;
1460 else if (base->opcode == MEM_REF
1461 && base[1].opcode == ADDR_EXPR
1462 && (TREE_CODE (TREE_OPERAND (base[1].op0, 0)) == VAR_DECL
1463 || TREE_CODE (TREE_OPERAND (base[1].op0, 0)) == CONST_DECL
1464 || TREE_CODE (TREE_OPERAND (base[1].op0, 0)) == STRING_CST))
1466 decl = TREE_OPERAND (base[1].op0, 0);
1467 if (TREE_CODE (decl) == STRING_CST)
1468 ctor = decl;
1469 else
1470 ctor = ctor_for_folding (decl);
1472 if (ctor == NULL_TREE)
1473 return build_zero_cst (ref->type);
1474 else if (ctor != error_mark_node)
1476 HOST_WIDE_INT const_off;
1477 if (decl)
1479 tree res = fold_ctor_reference (ref->type, ctor,
1480 off * BITS_PER_UNIT,
1481 size * BITS_PER_UNIT, decl);
1482 if (res)
1484 STRIP_USELESS_TYPE_CONVERSION (res);
1485 if (is_gimple_min_invariant (res))
1486 return res;
1489 else if (off.is_constant (&const_off))
1491 unsigned char buf[MAX_BITSIZE_MODE_ANY_MODE / BITS_PER_UNIT];
1492 int len = native_encode_expr (ctor, buf, size, const_off);
1493 if (len > 0)
1494 return native_interpret_expr (ref->type, buf, len);
1499 return NULL_TREE;
1502 /* Return true if OPS contain a storage order barrier. */
1504 static bool
1505 contains_storage_order_barrier_p (vec<vn_reference_op_s> ops)
1507 vn_reference_op_t op;
1508 unsigned i;
1510 FOR_EACH_VEC_ELT (ops, i, op)
1511 if (op->opcode == VIEW_CONVERT_EXPR && op->reverse)
1512 return true;
1514 return false;
1517 /* Transform any SSA_NAME's in a vector of vn_reference_op_s
1518 structures into their value numbers. This is done in-place, and
1519 the vector passed in is returned. *VALUEIZED_ANYTHING will specify
1520 whether any operands were valueized. */
1522 static vec<vn_reference_op_s>
1523 valueize_refs_1 (vec<vn_reference_op_s> orig, bool *valueized_anything,
1524 bool with_avail = false)
1526 vn_reference_op_t vro;
1527 unsigned int i;
1529 *valueized_anything = false;
1531 FOR_EACH_VEC_ELT (orig, i, vro)
1533 if (vro->opcode == SSA_NAME
1534 || (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME))
1536 tree tem = with_avail ? vn_valueize (vro->op0) : SSA_VAL (vro->op0);
1537 if (tem != vro->op0)
1539 *valueized_anything = true;
1540 vro->op0 = tem;
1542 /* If it transforms from an SSA_NAME to a constant, update
1543 the opcode. */
1544 if (TREE_CODE (vro->op0) != SSA_NAME && vro->opcode == SSA_NAME)
1545 vro->opcode = TREE_CODE (vro->op0);
1547 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
1549 tree tem = with_avail ? vn_valueize (vro->op1) : SSA_VAL (vro->op1);
1550 if (tem != vro->op1)
1552 *valueized_anything = true;
1553 vro->op1 = tem;
1556 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
1558 tree tem = with_avail ? vn_valueize (vro->op2) : SSA_VAL (vro->op2);
1559 if (tem != vro->op2)
1561 *valueized_anything = true;
1562 vro->op2 = tem;
1565 /* If it transforms from an SSA_NAME to an address, fold with
1566 a preceding indirect reference. */
1567 if (i > 0
1568 && vro->op0
1569 && TREE_CODE (vro->op0) == ADDR_EXPR
1570 && orig[i - 1].opcode == MEM_REF)
1572 if (vn_reference_fold_indirect (&orig, &i))
1573 *valueized_anything = true;
1575 else if (i > 0
1576 && vro->opcode == SSA_NAME
1577 && orig[i - 1].opcode == MEM_REF)
1579 if (vn_reference_maybe_forwprop_address (&orig, &i))
1580 *valueized_anything = true;
1582 /* If it transforms a non-constant ARRAY_REF into a constant
1583 one, adjust the constant offset. */
1584 else if (vro->opcode == ARRAY_REF
1585 && known_eq (vro->off, -1)
1586 && poly_int_tree_p (vro->op0)
1587 && poly_int_tree_p (vro->op1)
1588 && TREE_CODE (vro->op2) == INTEGER_CST)
1590 poly_offset_int off = ((wi::to_poly_offset (vro->op0)
1591 - wi::to_poly_offset (vro->op1))
1592 * wi::to_offset (vro->op2)
1593 * vn_ref_op_align_unit (vro));
1594 off.to_shwi (&vro->off);
1598 return orig;
1601 static vec<vn_reference_op_s>
1602 valueize_refs (vec<vn_reference_op_s> orig)
1604 bool tem;
1605 return valueize_refs_1 (orig, &tem);
1608 static vec<vn_reference_op_s> shared_lookup_references;
1610 /* Create a vector of vn_reference_op_s structures from REF, a
1611 REFERENCE_CLASS_P tree. The vector is shared among all callers of
1612 this function. *VALUEIZED_ANYTHING will specify whether any
1613 operands were valueized. */
1615 static vec<vn_reference_op_s>
1616 valueize_shared_reference_ops_from_ref (tree ref, bool *valueized_anything)
1618 if (!ref)
1619 return vNULL;
1620 shared_lookup_references.truncate (0);
1621 copy_reference_ops_from_ref (ref, &shared_lookup_references);
1622 shared_lookup_references = valueize_refs_1 (shared_lookup_references,
1623 valueized_anything);
1624 return shared_lookup_references;
1627 /* Create a vector of vn_reference_op_s structures from CALL, a
1628 call statement. The vector is shared among all callers of
1629 this function. */
1631 static vec<vn_reference_op_s>
1632 valueize_shared_reference_ops_from_call (gcall *call)
1634 if (!call)
1635 return vNULL;
1636 shared_lookup_references.truncate (0);
1637 copy_reference_ops_from_call (call, &shared_lookup_references);
1638 shared_lookup_references = valueize_refs (shared_lookup_references);
1639 return shared_lookup_references;
1642 /* Lookup a SCCVN reference operation VR in the current hash table.
1643 Returns the resulting value number if it exists in the hash table,
1644 NULL_TREE otherwise. VNRESULT will be filled in with the actual
1645 vn_reference_t stored in the hashtable if something is found. */
1647 static tree
1648 vn_reference_lookup_1 (vn_reference_t vr, vn_reference_t *vnresult)
1650 vn_reference_s **slot;
1651 hashval_t hash;
1653 hash = vr->hashcode;
1654 slot = valid_info->references->find_slot_with_hash (vr, hash, NO_INSERT);
1655 if (slot)
1657 if (vnresult)
1658 *vnresult = (vn_reference_t)*slot;
1659 return ((vn_reference_t)*slot)->result;
1662 return NULL_TREE;
1666 /* Partial definition tracking support. */
1668 struct pd_range
1670 HOST_WIDE_INT offset;
1671 HOST_WIDE_INT size;
1674 struct pd_data
1676 tree rhs;
1677 HOST_WIDE_INT offset;
1678 HOST_WIDE_INT size;
1681 /* Context for alias walking. */
1683 struct vn_walk_cb_data
1685 vn_walk_cb_data (vn_reference_t vr_, tree orig_ref_, tree *last_vuse_ptr_,
1686 vn_lookup_kind vn_walk_kind_, bool tbaa_p_, tree mask_)
1687 : vr (vr_), last_vuse_ptr (last_vuse_ptr_), last_vuse (NULL_TREE),
1688 mask (mask_), masked_result (NULL_TREE), vn_walk_kind (vn_walk_kind_),
1689 tbaa_p (tbaa_p_), saved_operands (vNULL), first_set (-2),
1690 first_base_set (-2), known_ranges (NULL)
1692 if (!last_vuse_ptr)
1693 last_vuse_ptr = &last_vuse;
1694 ao_ref_init (&orig_ref, orig_ref_);
1695 if (mask)
1697 wide_int w = wi::to_wide (mask);
1698 unsigned int pos = 0, prec = w.get_precision ();
1699 pd_data pd;
1700 pd.rhs = build_constructor (NULL_TREE, NULL);
1701 /* When bitwise and with a constant is done on a memory load,
1702 we don't really need all the bits to be defined or defined
1703 to constants, we don't really care what is in the position
1704 corresponding to 0 bits in the mask.
1705 So, push the ranges of those 0 bits in the mask as artificial
1706 zero stores and let the partial def handling code do the
1707 rest. */
1708 while (pos < prec)
1710 int tz = wi::ctz (w);
1711 if (pos + tz > prec)
1712 tz = prec - pos;
1713 if (tz)
1715 if (BYTES_BIG_ENDIAN)
1716 pd.offset = prec - pos - tz;
1717 else
1718 pd.offset = pos;
1719 pd.size = tz;
1720 void *r = push_partial_def (pd, 0, 0, 0, prec);
1721 gcc_assert (r == NULL_TREE);
1723 pos += tz;
1724 if (pos == prec)
1725 break;
1726 w = wi::lrshift (w, tz);
1727 tz = wi::ctz (wi::bit_not (w));
1728 if (pos + tz > prec)
1729 tz = prec - pos;
1730 pos += tz;
1731 w = wi::lrshift (w, tz);
1735 ~vn_walk_cb_data ();
1736 void *finish (alias_set_type, alias_set_type, tree);
1737 void *push_partial_def (pd_data pd,
1738 alias_set_type, alias_set_type, HOST_WIDE_INT,
1739 HOST_WIDE_INT);
1741 vn_reference_t vr;
1742 ao_ref orig_ref;
1743 tree *last_vuse_ptr;
1744 tree last_vuse;
1745 tree mask;
1746 tree masked_result;
1747 vn_lookup_kind vn_walk_kind;
1748 bool tbaa_p;
1749 vec<vn_reference_op_s> saved_operands;
1751 /* The VDEFs of partial defs we come along. */
1752 auto_vec<pd_data, 2> partial_defs;
1753 /* The first defs range to avoid splay tree setup in most cases. */
1754 pd_range first_range;
1755 alias_set_type first_set;
1756 alias_set_type first_base_set;
1757 splay_tree known_ranges;
1758 obstack ranges_obstack;
1761 vn_walk_cb_data::~vn_walk_cb_data ()
1763 if (known_ranges)
1765 splay_tree_delete (known_ranges);
1766 obstack_free (&ranges_obstack, NULL);
1768 saved_operands.release ();
1771 void *
1772 vn_walk_cb_data::finish (alias_set_type set, alias_set_type base_set, tree val)
1774 if (first_set != -2)
1776 set = first_set;
1777 base_set = first_base_set;
1779 if (mask)
1781 masked_result = val;
1782 return (void *) -1;
1784 vec<vn_reference_op_s> &operands
1785 = saved_operands.exists () ? saved_operands : vr->operands;
1786 return vn_reference_lookup_or_insert_for_pieces (last_vuse, set, base_set,
1787 vr->type, operands, val);
1790 /* pd_range splay-tree helpers. */
1792 static int
1793 pd_range_compare (splay_tree_key offset1p, splay_tree_key offset2p)
1795 HOST_WIDE_INT offset1 = *(HOST_WIDE_INT *)offset1p;
1796 HOST_WIDE_INT offset2 = *(HOST_WIDE_INT *)offset2p;
1797 if (offset1 < offset2)
1798 return -1;
1799 else if (offset1 > offset2)
1800 return 1;
1801 return 0;
1804 static void *
1805 pd_tree_alloc (int size, void *data_)
1807 vn_walk_cb_data *data = (vn_walk_cb_data *)data_;
1808 return obstack_alloc (&data->ranges_obstack, size);
1811 static void
1812 pd_tree_dealloc (void *, void *)
1816 /* Push PD to the vector of partial definitions returning a
1817 value when we are ready to combine things with VUSE, SET and MAXSIZEI,
1818 NULL when we want to continue looking for partial defs or -1
1819 on failure. */
1821 void *
1822 vn_walk_cb_data::push_partial_def (pd_data pd,
1823 alias_set_type set, alias_set_type base_set,
1824 HOST_WIDE_INT offseti,
1825 HOST_WIDE_INT maxsizei)
1827 const HOST_WIDE_INT bufsize = 64;
1828 /* We're using a fixed buffer for encoding so fail early if the object
1829 we want to interpret is bigger. */
1830 if (maxsizei > bufsize * BITS_PER_UNIT
1831 || CHAR_BIT != 8
1832 || BITS_PER_UNIT != 8
1833 /* Not prepared to handle PDP endian. */
1834 || BYTES_BIG_ENDIAN != WORDS_BIG_ENDIAN)
1835 return (void *)-1;
1837 /* Turn too large constant stores into non-constant stores. */
1838 if (CONSTANT_CLASS_P (pd.rhs) && pd.size > bufsize * BITS_PER_UNIT)
1839 pd.rhs = error_mark_node;
1841 /* And for non-constant or CONSTRUCTOR stores shrink them to only keep at
1842 most a partial byte before and/or after the region. */
1843 if (!CONSTANT_CLASS_P (pd.rhs))
1845 if (pd.offset < offseti)
1847 HOST_WIDE_INT o = ROUND_DOWN (offseti - pd.offset, BITS_PER_UNIT);
1848 gcc_assert (pd.size > o);
1849 pd.size -= o;
1850 pd.offset += o;
1852 if (pd.size > maxsizei)
1853 pd.size = maxsizei + ((pd.size - maxsizei) % BITS_PER_UNIT);
1856 pd.offset -= offseti;
1858 bool pd_constant_p = (TREE_CODE (pd.rhs) == CONSTRUCTOR
1859 || CONSTANT_CLASS_P (pd.rhs));
1860 if (partial_defs.is_empty ())
1862 /* If we get a clobber upfront, fail. */
1863 if (TREE_CLOBBER_P (pd.rhs))
1864 return (void *)-1;
1865 if (!pd_constant_p)
1866 return (void *)-1;
1867 partial_defs.safe_push (pd);
1868 first_range.offset = pd.offset;
1869 first_range.size = pd.size;
1870 first_set = set;
1871 first_base_set = base_set;
1872 last_vuse_ptr = NULL;
1873 /* Continue looking for partial defs. */
1874 return NULL;
1877 if (!known_ranges)
1879 /* ??? Optimize the case where the 2nd partial def completes things. */
1880 gcc_obstack_init (&ranges_obstack);
1881 known_ranges = splay_tree_new_with_allocator (pd_range_compare, 0, 0,
1882 pd_tree_alloc,
1883 pd_tree_dealloc, this);
1884 splay_tree_insert (known_ranges,
1885 (splay_tree_key)&first_range.offset,
1886 (splay_tree_value)&first_range);
1889 pd_range newr = { pd.offset, pd.size };
1890 splay_tree_node n;
1891 pd_range *r;
1892 /* Lookup the predecessor of offset + 1 and see if we need to merge. */
1893 HOST_WIDE_INT loffset = newr.offset + 1;
1894 if ((n = splay_tree_predecessor (known_ranges, (splay_tree_key)&loffset))
1895 && ((r = (pd_range *)n->value), true)
1896 && ranges_known_overlap_p (r->offset, r->size + 1,
1897 newr.offset, newr.size))
1899 /* Ignore partial defs already covered. Here we also drop shadowed
1900 clobbers arriving here at the floor. */
1901 if (known_subrange_p (newr.offset, newr.size, r->offset, r->size))
1902 return NULL;
1903 r->size = MAX (r->offset + r->size, newr.offset + newr.size) - r->offset;
1905 else
1907 /* newr.offset wasn't covered yet, insert the range. */
1908 r = XOBNEW (&ranges_obstack, pd_range);
1909 *r = newr;
1910 splay_tree_insert (known_ranges, (splay_tree_key)&r->offset,
1911 (splay_tree_value)r);
1913 /* Merge r which now contains newr and is a member of the splay tree with
1914 adjacent overlapping ranges. */
1915 pd_range *rafter;
1916 while ((n = splay_tree_successor (known_ranges, (splay_tree_key)&r->offset))
1917 && ((rafter = (pd_range *)n->value), true)
1918 && ranges_known_overlap_p (r->offset, r->size + 1,
1919 rafter->offset, rafter->size))
1921 r->size = MAX (r->offset + r->size,
1922 rafter->offset + rafter->size) - r->offset;
1923 splay_tree_remove (known_ranges, (splay_tree_key)&rafter->offset);
1925 /* If we get a clobber, fail. */
1926 if (TREE_CLOBBER_P (pd.rhs))
1927 return (void *)-1;
1928 /* Non-constants are OK as long as they are shadowed by a constant. */
1929 if (!pd_constant_p)
1930 return (void *)-1;
1931 partial_defs.safe_push (pd);
1933 /* Now we have merged newr into the range tree. When we have covered
1934 [offseti, sizei] then the tree will contain exactly one node which has
1935 the desired properties and it will be 'r'. */
1936 if (!known_subrange_p (0, maxsizei, r->offset, r->size))
1937 /* Continue looking for partial defs. */
1938 return NULL;
1940 /* Now simply native encode all partial defs in reverse order. */
1941 unsigned ndefs = partial_defs.length ();
1942 /* We support up to 512-bit values (for V8DFmode). */
1943 unsigned char buffer[bufsize + 1];
1944 unsigned char this_buffer[bufsize + 1];
1945 int len;
1947 memset (buffer, 0, bufsize + 1);
1948 unsigned needed_len = ROUND_UP (maxsizei, BITS_PER_UNIT) / BITS_PER_UNIT;
1949 while (!partial_defs.is_empty ())
1951 pd_data pd = partial_defs.pop ();
1952 unsigned int amnt;
1953 if (TREE_CODE (pd.rhs) == CONSTRUCTOR)
1955 /* Empty CONSTRUCTOR. */
1956 if (pd.size >= needed_len * BITS_PER_UNIT)
1957 len = needed_len;
1958 else
1959 len = ROUND_UP (pd.size, BITS_PER_UNIT) / BITS_PER_UNIT;
1960 memset (this_buffer, 0, len);
1962 else
1964 len = native_encode_expr (pd.rhs, this_buffer, bufsize,
1965 MAX (0, -pd.offset) / BITS_PER_UNIT);
1966 if (len <= 0
1967 || len < (ROUND_UP (pd.size, BITS_PER_UNIT) / BITS_PER_UNIT
1968 - MAX (0, -pd.offset) / BITS_PER_UNIT))
1970 if (dump_file && (dump_flags & TDF_DETAILS))
1971 fprintf (dump_file, "Failed to encode %u "
1972 "partial definitions\n", ndefs);
1973 return (void *)-1;
1977 unsigned char *p = buffer;
1978 HOST_WIDE_INT size = pd.size;
1979 if (pd.offset < 0)
1980 size -= ROUND_DOWN (-pd.offset, BITS_PER_UNIT);
1981 this_buffer[len] = 0;
1982 if (BYTES_BIG_ENDIAN)
1984 /* LSB of this_buffer[len - 1] byte should be at
1985 pd.offset + pd.size - 1 bits in buffer. */
1986 amnt = ((unsigned HOST_WIDE_INT) pd.offset
1987 + pd.size) % BITS_PER_UNIT;
1988 if (amnt)
1989 shift_bytes_in_array_right (this_buffer, len + 1, amnt);
1990 unsigned char *q = this_buffer;
1991 unsigned int off = 0;
1992 if (pd.offset >= 0)
1994 unsigned int msk;
1995 off = pd.offset / BITS_PER_UNIT;
1996 gcc_assert (off < needed_len);
1997 p = buffer + off;
1998 if (size <= amnt)
2000 msk = ((1 << size) - 1) << (BITS_PER_UNIT - amnt);
2001 *p = (*p & ~msk) | (this_buffer[len] & msk);
2002 size = 0;
2004 else
2006 if (TREE_CODE (pd.rhs) != CONSTRUCTOR)
2007 q = (this_buffer + len
2008 - (ROUND_UP (size - amnt, BITS_PER_UNIT)
2009 / BITS_PER_UNIT));
2010 if (pd.offset % BITS_PER_UNIT)
2012 msk = -1U << (BITS_PER_UNIT
2013 - (pd.offset % BITS_PER_UNIT));
2014 *p = (*p & msk) | (*q & ~msk);
2015 p++;
2016 q++;
2017 off++;
2018 size -= BITS_PER_UNIT - (pd.offset % BITS_PER_UNIT);
2019 gcc_assert (size >= 0);
2023 else if (TREE_CODE (pd.rhs) != CONSTRUCTOR)
2025 q = (this_buffer + len
2026 - (ROUND_UP (size - amnt, BITS_PER_UNIT)
2027 / BITS_PER_UNIT));
2028 if (pd.offset % BITS_PER_UNIT)
2030 q++;
2031 size -= BITS_PER_UNIT - ((unsigned HOST_WIDE_INT) pd.offset
2032 % BITS_PER_UNIT);
2033 gcc_assert (size >= 0);
2036 if ((unsigned HOST_WIDE_INT) size / BITS_PER_UNIT + off
2037 > needed_len)
2038 size = (needed_len - off) * BITS_PER_UNIT;
2039 memcpy (p, q, size / BITS_PER_UNIT);
2040 if (size % BITS_PER_UNIT)
2042 unsigned int msk
2043 = -1U << (BITS_PER_UNIT - (size % BITS_PER_UNIT));
2044 p += size / BITS_PER_UNIT;
2045 q += size / BITS_PER_UNIT;
2046 *p = (*q & msk) | (*p & ~msk);
2049 else
2051 size = MIN (size, (HOST_WIDE_INT) needed_len * BITS_PER_UNIT);
2052 if (pd.offset >= 0)
2054 /* LSB of this_buffer[0] byte should be at pd.offset bits
2055 in buffer. */
2056 unsigned int msk;
2057 amnt = pd.offset % BITS_PER_UNIT;
2058 if (amnt)
2059 shift_bytes_in_array_left (this_buffer, len + 1, amnt);
2060 unsigned int off = pd.offset / BITS_PER_UNIT;
2061 gcc_assert (off < needed_len);
2062 size = MIN (size,
2063 (HOST_WIDE_INT) (needed_len - off) * BITS_PER_UNIT);
2064 p = buffer + off;
2065 if (amnt + size < BITS_PER_UNIT)
2067 /* Low amnt bits come from *p, then size bits
2068 from this_buffer[0] and the remaining again from
2069 *p. */
2070 msk = ((1 << size) - 1) << amnt;
2071 *p = (*p & ~msk) | (this_buffer[0] & msk);
2072 size = 0;
2074 else if (amnt)
2076 msk = -1U << amnt;
2077 *p = (*p & ~msk) | (this_buffer[0] & msk);
2078 p++;
2079 size -= (BITS_PER_UNIT - amnt);
2082 else
2084 amnt = (unsigned HOST_WIDE_INT) pd.offset % BITS_PER_UNIT;
2085 if (amnt)
2086 shift_bytes_in_array_left (this_buffer, len + 1, amnt);
2088 memcpy (p, this_buffer + (amnt != 0), size / BITS_PER_UNIT);
2089 p += size / BITS_PER_UNIT;
2090 if (size % BITS_PER_UNIT)
2092 unsigned int msk = -1U << (size % BITS_PER_UNIT);
2093 *p = (this_buffer[(amnt != 0) + size / BITS_PER_UNIT]
2094 & ~msk) | (*p & msk);
2099 tree type = vr->type;
2100 /* Make sure to interpret in a type that has a range covering the whole
2101 access size. */
2102 if (INTEGRAL_TYPE_P (vr->type) && maxsizei != TYPE_PRECISION (vr->type))
2103 type = build_nonstandard_integer_type (maxsizei, TYPE_UNSIGNED (type));
2104 tree val;
2105 if (BYTES_BIG_ENDIAN)
2107 unsigned sz = needed_len;
2108 if (maxsizei % BITS_PER_UNIT)
2109 shift_bytes_in_array_right (buffer, needed_len,
2110 BITS_PER_UNIT
2111 - (maxsizei % BITS_PER_UNIT));
2112 if (INTEGRAL_TYPE_P (type))
2113 sz = GET_MODE_SIZE (SCALAR_INT_TYPE_MODE (type));
2114 if (sz > needed_len)
2116 memcpy (this_buffer + (sz - needed_len), buffer, needed_len);
2117 val = native_interpret_expr (type, this_buffer, sz);
2119 else
2120 val = native_interpret_expr (type, buffer, needed_len);
2122 else
2123 val = native_interpret_expr (type, buffer, bufsize);
2124 /* If we chop off bits because the types precision doesn't match the memory
2125 access size this is ok when optimizing reads but not when called from
2126 the DSE code during elimination. */
2127 if (val && type != vr->type)
2129 if (! int_fits_type_p (val, vr->type))
2130 val = NULL_TREE;
2131 else
2132 val = fold_convert (vr->type, val);
2135 if (val)
2137 if (dump_file && (dump_flags & TDF_DETAILS))
2138 fprintf (dump_file,
2139 "Successfully combined %u partial definitions\n", ndefs);
2140 /* We are using the alias-set of the first store we encounter which
2141 should be appropriate here. */
2142 return finish (first_set, first_base_set, val);
2144 else
2146 if (dump_file && (dump_flags & TDF_DETAILS))
2147 fprintf (dump_file,
2148 "Failed to interpret %u encoded partial definitions\n", ndefs);
2149 return (void *)-1;
2153 /* Callback for walk_non_aliased_vuses. Adjusts the vn_reference_t VR_
2154 with the current VUSE and performs the expression lookup. */
2156 static void *
2157 vn_reference_lookup_2 (ao_ref *op ATTRIBUTE_UNUSED, tree vuse, void *data_)
2159 vn_walk_cb_data *data = (vn_walk_cb_data *)data_;
2160 vn_reference_t vr = data->vr;
2161 vn_reference_s **slot;
2162 hashval_t hash;
2164 /* If we have partial definitions recorded we have to go through
2165 vn_reference_lookup_3. */
2166 if (!data->partial_defs.is_empty ())
2167 return NULL;
2169 if (data->last_vuse_ptr)
2171 *data->last_vuse_ptr = vuse;
2172 data->last_vuse = vuse;
2175 /* Fixup vuse and hash. */
2176 if (vr->vuse)
2177 vr->hashcode = vr->hashcode - SSA_NAME_VERSION (vr->vuse);
2178 vr->vuse = vuse_ssa_val (vuse);
2179 if (vr->vuse)
2180 vr->hashcode = vr->hashcode + SSA_NAME_VERSION (vr->vuse);
2182 hash = vr->hashcode;
2183 slot = valid_info->references->find_slot_with_hash (vr, hash, NO_INSERT);
2184 if (slot)
2186 if ((*slot)->result && data->saved_operands.exists ())
2187 return data->finish (vr->set, vr->base_set, (*slot)->result);
2188 return *slot;
2191 return NULL;
2194 /* Lookup an existing or insert a new vn_reference entry into the
2195 value table for the VUSE, SET, TYPE, OPERANDS reference which
2196 has the value VALUE which is either a constant or an SSA name. */
2198 static vn_reference_t
2199 vn_reference_lookup_or_insert_for_pieces (tree vuse,
2200 alias_set_type set,
2201 alias_set_type base_set,
2202 tree type,
2203 vec<vn_reference_op_s,
2204 va_heap> operands,
2205 tree value)
2207 vn_reference_s vr1;
2208 vn_reference_t result;
2209 unsigned value_id;
2210 vr1.vuse = vuse ? SSA_VAL (vuse) : NULL_TREE;
2211 vr1.operands = operands;
2212 vr1.type = type;
2213 vr1.set = set;
2214 vr1.base_set = base_set;
2215 vr1.hashcode = vn_reference_compute_hash (&vr1);
2216 if (vn_reference_lookup_1 (&vr1, &result))
2217 return result;
2218 if (TREE_CODE (value) == SSA_NAME)
2219 value_id = VN_INFO (value)->value_id;
2220 else
2221 value_id = get_or_alloc_constant_value_id (value);
2222 return vn_reference_insert_pieces (vuse, set, base_set, type,
2223 operands.copy (), value, value_id);
2226 /* Return a value-number for RCODE OPS... either by looking up an existing
2227 value-number for the simplified result or by inserting the operation if
2228 INSERT is true. */
2230 static tree
2231 vn_nary_build_or_lookup_1 (gimple_match_op *res_op, bool insert)
2233 tree result = NULL_TREE;
2234 /* We will be creating a value number for
2235 RCODE (OPS...).
2236 So first simplify and lookup this expression to see if it
2237 is already available. */
2238 /* For simplification valueize. */
2239 unsigned i;
2240 for (i = 0; i < res_op->num_ops; ++i)
2241 if (TREE_CODE (res_op->ops[i]) == SSA_NAME)
2243 tree tem = vn_valueize (res_op->ops[i]);
2244 if (!tem)
2245 break;
2246 res_op->ops[i] = tem;
2248 /* If valueization of an operand fails (it is not available), skip
2249 simplification. */
2250 bool res = false;
2251 if (i == res_op->num_ops)
2253 mprts_hook = vn_lookup_simplify_result;
2254 res = res_op->resimplify (NULL, vn_valueize);
2255 mprts_hook = NULL;
2257 gimple *new_stmt = NULL;
2258 if (res
2259 && gimple_simplified_result_is_gimple_val (res_op))
2261 /* The expression is already available. */
2262 result = res_op->ops[0];
2263 /* Valueize it, simplification returns sth in AVAIL only. */
2264 if (TREE_CODE (result) == SSA_NAME)
2265 result = SSA_VAL (result);
2267 else
2269 tree val = vn_lookup_simplify_result (res_op);
2270 if (!val && insert)
2272 gimple_seq stmts = NULL;
2273 result = maybe_push_res_to_seq (res_op, &stmts);
2274 if (result)
2276 gcc_assert (gimple_seq_singleton_p (stmts));
2277 new_stmt = gimple_seq_first_stmt (stmts);
2280 else
2281 /* The expression is already available. */
2282 result = val;
2284 if (new_stmt)
2286 /* The expression is not yet available, value-number lhs to
2287 the new SSA_NAME we created. */
2288 /* Initialize value-number information properly. */
2289 vn_ssa_aux_t result_info = VN_INFO (result);
2290 result_info->valnum = result;
2291 result_info->value_id = get_next_value_id ();
2292 result_info->visited = 1;
2293 gimple_seq_add_stmt_without_update (&VN_INFO (result)->expr,
2294 new_stmt);
2295 result_info->needs_insertion = true;
2296 /* ??? PRE phi-translation inserts NARYs without corresponding
2297 SSA name result. Re-use those but set their result according
2298 to the stmt we just built. */
2299 vn_nary_op_t nary = NULL;
2300 vn_nary_op_lookup_stmt (new_stmt, &nary);
2301 if (nary)
2303 gcc_assert (! nary->predicated_values && nary->u.result == NULL_TREE);
2304 nary->u.result = gimple_assign_lhs (new_stmt);
2306 /* As all "inserted" statements are singleton SCCs, insert
2307 to the valid table. This is strictly needed to
2308 avoid re-generating new value SSA_NAMEs for the same
2309 expression during SCC iteration over and over (the
2310 optimistic table gets cleared after each iteration).
2311 We do not need to insert into the optimistic table, as
2312 lookups there will fall back to the valid table. */
2313 else
2315 unsigned int length = vn_nary_length_from_stmt (new_stmt);
2316 vn_nary_op_t vno1
2317 = alloc_vn_nary_op_noinit (length, &vn_tables_insert_obstack);
2318 vno1->value_id = result_info->value_id;
2319 vno1->length = length;
2320 vno1->predicated_values = 0;
2321 vno1->u.result = result;
2322 init_vn_nary_op_from_stmt (vno1, new_stmt);
2323 vn_nary_op_insert_into (vno1, valid_info->nary, true);
2324 /* Also do not link it into the undo chain. */
2325 last_inserted_nary = vno1->next;
2326 vno1->next = (vn_nary_op_t)(void *)-1;
2328 if (dump_file && (dump_flags & TDF_DETAILS))
2330 fprintf (dump_file, "Inserting name ");
2331 print_generic_expr (dump_file, result);
2332 fprintf (dump_file, " for expression ");
2333 print_gimple_expr (dump_file, new_stmt, 0, TDF_SLIM);
2334 fprintf (dump_file, "\n");
2337 return result;
2340 /* Return a value-number for RCODE OPS... either by looking up an existing
2341 value-number for the simplified result or by inserting the operation. */
2343 static tree
2344 vn_nary_build_or_lookup (gimple_match_op *res_op)
2346 return vn_nary_build_or_lookup_1 (res_op, true);
2349 /* Try to simplify the expression RCODE OPS... of type TYPE and return
2350 its value if present. */
2352 tree
2353 vn_nary_simplify (vn_nary_op_t nary)
2355 if (nary->length > gimple_match_op::MAX_NUM_OPS)
2356 return NULL_TREE;
2357 gimple_match_op op (gimple_match_cond::UNCOND, nary->opcode,
2358 nary->type, nary->length);
2359 memcpy (op.ops, nary->op, sizeof (tree) * nary->length);
2360 return vn_nary_build_or_lookup_1 (&op, false);
2363 /* Elimination engine. */
2365 class eliminate_dom_walker : public dom_walker
2367 public:
2368 eliminate_dom_walker (cdi_direction, bitmap);
2369 ~eliminate_dom_walker ();
2371 virtual edge before_dom_children (basic_block);
2372 virtual void after_dom_children (basic_block);
2374 virtual tree eliminate_avail (basic_block, tree op);
2375 virtual void eliminate_push_avail (basic_block, tree op);
2376 tree eliminate_insert (basic_block, gimple_stmt_iterator *gsi, tree val);
2378 void eliminate_stmt (basic_block, gimple_stmt_iterator *);
2380 unsigned eliminate_cleanup (bool region_p = false);
2382 bool do_pre;
2383 unsigned int el_todo;
2384 unsigned int eliminations;
2385 unsigned int insertions;
2387 /* SSA names that had their defs inserted by PRE if do_pre. */
2388 bitmap inserted_exprs;
2390 /* Blocks with statements that have had their EH properties changed. */
2391 bitmap need_eh_cleanup;
2393 /* Blocks with statements that have had their AB properties changed. */
2394 bitmap need_ab_cleanup;
2396 /* Local state for the eliminate domwalk. */
2397 auto_vec<gimple *> to_remove;
2398 auto_vec<gimple *> to_fixup;
2399 auto_vec<tree> avail;
2400 auto_vec<tree> avail_stack;
2403 /* Adaptor to the elimination engine using RPO availability. */
2405 class rpo_elim : public eliminate_dom_walker
2407 public:
2408 rpo_elim(basic_block entry_)
2409 : eliminate_dom_walker (CDI_DOMINATORS, NULL), entry (entry_),
2410 m_avail_freelist (NULL) {}
2412 virtual tree eliminate_avail (basic_block, tree op);
2414 virtual void eliminate_push_avail (basic_block, tree);
2416 basic_block entry;
2417 /* Freelist of avail entries which are allocated from the vn_ssa_aux
2418 obstack. */
2419 vn_avail *m_avail_freelist;
2422 /* Global RPO state for access from hooks. */
2423 static eliminate_dom_walker *rpo_avail;
2424 basic_block vn_context_bb;
2426 /* Return true if BASE1 and BASE2 can be adjusted so they have the
2427 same address and adjust *OFFSET1 and *OFFSET2 accordingly.
2428 Otherwise return false. */
2430 static bool
2431 adjust_offsets_for_equal_base_address (tree base1, poly_int64 *offset1,
2432 tree base2, poly_int64 *offset2)
2434 poly_int64 soff;
2435 if (TREE_CODE (base1) == MEM_REF
2436 && TREE_CODE (base2) == MEM_REF)
2438 if (mem_ref_offset (base1).to_shwi (&soff))
2440 base1 = TREE_OPERAND (base1, 0);
2441 *offset1 += soff * BITS_PER_UNIT;
2443 if (mem_ref_offset (base2).to_shwi (&soff))
2445 base2 = TREE_OPERAND (base2, 0);
2446 *offset2 += soff * BITS_PER_UNIT;
2448 return operand_equal_p (base1, base2, 0);
2450 return operand_equal_p (base1, base2, OEP_ADDRESS_OF);
2453 /* Callback for walk_non_aliased_vuses. Tries to perform a lookup
2454 from the statement defining VUSE and if not successful tries to
2455 translate *REFP and VR_ through an aggregate copy at the definition
2456 of VUSE. If *DISAMBIGUATE_ONLY is true then do not perform translation
2457 of *REF and *VR. If only disambiguation was performed then
2458 *DISAMBIGUATE_ONLY is set to true. */
2460 static void *
2461 vn_reference_lookup_3 (ao_ref *ref, tree vuse, void *data_,
2462 translate_flags *disambiguate_only)
2464 vn_walk_cb_data *data = (vn_walk_cb_data *)data_;
2465 vn_reference_t vr = data->vr;
2466 gimple *def_stmt = SSA_NAME_DEF_STMT (vuse);
2467 tree base = ao_ref_base (ref);
2468 HOST_WIDE_INT offseti = 0, maxsizei, sizei = 0;
2469 static vec<vn_reference_op_s> lhs_ops;
2470 ao_ref lhs_ref;
2471 bool lhs_ref_ok = false;
2472 poly_int64 copy_size;
2474 /* First try to disambiguate after value-replacing in the definitions LHS. */
2475 if (is_gimple_assign (def_stmt))
2477 tree lhs = gimple_assign_lhs (def_stmt);
2478 bool valueized_anything = false;
2479 /* Avoid re-allocation overhead. */
2480 lhs_ops.truncate (0);
2481 basic_block saved_rpo_bb = vn_context_bb;
2482 vn_context_bb = gimple_bb (def_stmt);
2483 if (*disambiguate_only <= TR_VALUEIZE_AND_DISAMBIGUATE)
2485 copy_reference_ops_from_ref (lhs, &lhs_ops);
2486 lhs_ops = valueize_refs_1 (lhs_ops, &valueized_anything, true);
2488 vn_context_bb = saved_rpo_bb;
2489 ao_ref_init (&lhs_ref, lhs);
2490 lhs_ref_ok = true;
2491 if (valueized_anything
2492 && ao_ref_init_from_vn_reference
2493 (&lhs_ref, ao_ref_alias_set (&lhs_ref),
2494 ao_ref_base_alias_set (&lhs_ref), TREE_TYPE (lhs), lhs_ops)
2495 && !refs_may_alias_p_1 (ref, &lhs_ref, data->tbaa_p))
2497 *disambiguate_only = TR_VALUEIZE_AND_DISAMBIGUATE;
2498 return NULL;
2501 /* Besides valueizing the LHS we can also use access-path based
2502 disambiguation on the original non-valueized ref. */
2503 if (!ref->ref
2504 && lhs_ref_ok
2505 && data->orig_ref.ref)
2507 /* We want to use the non-valueized LHS for this, but avoid redundant
2508 work. */
2509 ao_ref *lref = &lhs_ref;
2510 ao_ref lref_alt;
2511 if (valueized_anything)
2513 ao_ref_init (&lref_alt, lhs);
2514 lref = &lref_alt;
2516 if (!refs_may_alias_p_1 (&data->orig_ref, lref, data->tbaa_p))
2518 *disambiguate_only = (valueized_anything
2519 ? TR_VALUEIZE_AND_DISAMBIGUATE
2520 : TR_DISAMBIGUATE);
2521 return NULL;
2525 /* If we reach a clobbering statement try to skip it and see if
2526 we find a VN result with exactly the same value as the
2527 possible clobber. In this case we can ignore the clobber
2528 and return the found value. */
2529 if (is_gimple_reg_type (TREE_TYPE (lhs))
2530 && types_compatible_p (TREE_TYPE (lhs), vr->type)
2531 && (ref->ref || data->orig_ref.ref))
2533 tree *saved_last_vuse_ptr = data->last_vuse_ptr;
2534 /* Do not update last_vuse_ptr in vn_reference_lookup_2. */
2535 data->last_vuse_ptr = NULL;
2536 tree saved_vuse = vr->vuse;
2537 hashval_t saved_hashcode = vr->hashcode;
2538 void *res = vn_reference_lookup_2 (ref, gimple_vuse (def_stmt), data);
2539 /* Need to restore vr->vuse and vr->hashcode. */
2540 vr->vuse = saved_vuse;
2541 vr->hashcode = saved_hashcode;
2542 data->last_vuse_ptr = saved_last_vuse_ptr;
2543 if (res && res != (void *)-1)
2545 vn_reference_t vnresult = (vn_reference_t) res;
2546 tree rhs = gimple_assign_rhs1 (def_stmt);
2547 if (TREE_CODE (rhs) == SSA_NAME)
2548 rhs = SSA_VAL (rhs);
2549 if (vnresult->result
2550 && operand_equal_p (vnresult->result, rhs, 0)
2551 /* We have to honor our promise about union type punning
2552 and also support arbitrary overlaps with
2553 -fno-strict-aliasing. So simply resort to alignment to
2554 rule out overlaps. Do this check last because it is
2555 quite expensive compared to the hash-lookup above. */
2556 && multiple_p (get_object_alignment
2557 (ref->ref ? ref->ref : data->orig_ref.ref),
2558 ref->size)
2559 && multiple_p (get_object_alignment (lhs), ref->size))
2560 return res;
2564 else if (*disambiguate_only <= TR_VALUEIZE_AND_DISAMBIGUATE
2565 && gimple_call_builtin_p (def_stmt, BUILT_IN_NORMAL)
2566 && gimple_call_num_args (def_stmt) <= 4)
2568 /* For builtin calls valueize its arguments and call the
2569 alias oracle again. Valueization may improve points-to
2570 info of pointers and constify size and position arguments.
2571 Originally this was motivated by PR61034 which has
2572 conditional calls to free falsely clobbering ref because
2573 of imprecise points-to info of the argument. */
2574 tree oldargs[4];
2575 bool valueized_anything = false;
2576 for (unsigned i = 0; i < gimple_call_num_args (def_stmt); ++i)
2578 oldargs[i] = gimple_call_arg (def_stmt, i);
2579 tree val = vn_valueize (oldargs[i]);
2580 if (val != oldargs[i])
2582 gimple_call_set_arg (def_stmt, i, val);
2583 valueized_anything = true;
2586 if (valueized_anything)
2588 bool res = call_may_clobber_ref_p_1 (as_a <gcall *> (def_stmt),
2589 ref, data->tbaa_p);
2590 for (unsigned i = 0; i < gimple_call_num_args (def_stmt); ++i)
2591 gimple_call_set_arg (def_stmt, i, oldargs[i]);
2592 if (!res)
2594 *disambiguate_only = TR_VALUEIZE_AND_DISAMBIGUATE;
2595 return NULL;
2600 if (*disambiguate_only > TR_TRANSLATE)
2601 return (void *)-1;
2603 /* If we cannot constrain the size of the reference we cannot
2604 test if anything kills it. */
2605 if (!ref->max_size_known_p ())
2606 return (void *)-1;
2608 poly_int64 offset = ref->offset;
2609 poly_int64 maxsize = ref->max_size;
2611 /* def_stmt may-defs *ref. See if we can derive a value for *ref
2612 from that definition.
2613 1) Memset. */
2614 if (is_gimple_reg_type (vr->type)
2615 && (gimple_call_builtin_p (def_stmt, BUILT_IN_MEMSET)
2616 || gimple_call_builtin_p (def_stmt, BUILT_IN_MEMSET_CHK))
2617 && (integer_zerop (gimple_call_arg (def_stmt, 1))
2618 || ((TREE_CODE (gimple_call_arg (def_stmt, 1)) == INTEGER_CST
2619 || (INTEGRAL_TYPE_P (vr->type) && known_eq (ref->size, 8)))
2620 && CHAR_BIT == 8
2621 && BITS_PER_UNIT == 8
2622 && BYTES_BIG_ENDIAN == WORDS_BIG_ENDIAN
2623 && offset.is_constant (&offseti)
2624 && ref->size.is_constant (&sizei)
2625 && (offseti % BITS_PER_UNIT == 0
2626 || TREE_CODE (gimple_call_arg (def_stmt, 1)) == INTEGER_CST)))
2627 && (poly_int_tree_p (gimple_call_arg (def_stmt, 2))
2628 || (TREE_CODE (gimple_call_arg (def_stmt, 2)) == SSA_NAME
2629 && poly_int_tree_p (SSA_VAL (gimple_call_arg (def_stmt, 2)))))
2630 && (TREE_CODE (gimple_call_arg (def_stmt, 0)) == ADDR_EXPR
2631 || TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME))
2633 tree base2;
2634 poly_int64 offset2, size2, maxsize2;
2635 bool reverse;
2636 tree ref2 = gimple_call_arg (def_stmt, 0);
2637 if (TREE_CODE (ref2) == SSA_NAME)
2639 ref2 = SSA_VAL (ref2);
2640 if (TREE_CODE (ref2) == SSA_NAME
2641 && (TREE_CODE (base) != MEM_REF
2642 || TREE_OPERAND (base, 0) != ref2))
2644 gimple *def_stmt = SSA_NAME_DEF_STMT (ref2);
2645 if (gimple_assign_single_p (def_stmt)
2646 && gimple_assign_rhs_code (def_stmt) == ADDR_EXPR)
2647 ref2 = gimple_assign_rhs1 (def_stmt);
2650 if (TREE_CODE (ref2) == ADDR_EXPR)
2652 ref2 = TREE_OPERAND (ref2, 0);
2653 base2 = get_ref_base_and_extent (ref2, &offset2, &size2, &maxsize2,
2654 &reverse);
2655 if (!known_size_p (maxsize2)
2656 || !known_eq (maxsize2, size2)
2657 || !operand_equal_p (base, base2, OEP_ADDRESS_OF))
2658 return (void *)-1;
2660 else if (TREE_CODE (ref2) == SSA_NAME)
2662 poly_int64 soff;
2663 if (TREE_CODE (base) != MEM_REF
2664 || !(mem_ref_offset (base)
2665 << LOG2_BITS_PER_UNIT).to_shwi (&soff))
2666 return (void *)-1;
2667 offset += soff;
2668 offset2 = 0;
2669 if (TREE_OPERAND (base, 0) != ref2)
2671 gimple *def = SSA_NAME_DEF_STMT (ref2);
2672 if (is_gimple_assign (def)
2673 && gimple_assign_rhs_code (def) == POINTER_PLUS_EXPR
2674 && gimple_assign_rhs1 (def) == TREE_OPERAND (base, 0)
2675 && poly_int_tree_p (gimple_assign_rhs2 (def)))
2677 tree rhs2 = gimple_assign_rhs2 (def);
2678 if (!(poly_offset_int::from (wi::to_poly_wide (rhs2),
2679 SIGNED)
2680 << LOG2_BITS_PER_UNIT).to_shwi (&offset2))
2681 return (void *)-1;
2682 ref2 = gimple_assign_rhs1 (def);
2683 if (TREE_CODE (ref2) == SSA_NAME)
2684 ref2 = SSA_VAL (ref2);
2686 else
2687 return (void *)-1;
2690 else
2691 return (void *)-1;
2692 tree len = gimple_call_arg (def_stmt, 2);
2693 HOST_WIDE_INT leni, offset2i;
2694 if (TREE_CODE (len) == SSA_NAME)
2695 len = SSA_VAL (len);
2696 /* Sometimes the above trickery is smarter than alias analysis. Take
2697 advantage of that. */
2698 if (!ranges_maybe_overlap_p (offset, maxsize, offset2,
2699 (wi::to_poly_offset (len)
2700 << LOG2_BITS_PER_UNIT)))
2701 return NULL;
2702 if (data->partial_defs.is_empty ()
2703 && known_subrange_p (offset, maxsize, offset2,
2704 wi::to_poly_offset (len) << LOG2_BITS_PER_UNIT))
2706 tree val;
2707 if (integer_zerop (gimple_call_arg (def_stmt, 1)))
2708 val = build_zero_cst (vr->type);
2709 else if (INTEGRAL_TYPE_P (vr->type)
2710 && known_eq (ref->size, 8)
2711 && offseti % BITS_PER_UNIT == 0)
2713 gimple_match_op res_op (gimple_match_cond::UNCOND, NOP_EXPR,
2714 vr->type, gimple_call_arg (def_stmt, 1));
2715 val = vn_nary_build_or_lookup (&res_op);
2716 if (!val
2717 || (TREE_CODE (val) == SSA_NAME
2718 && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (val)))
2719 return (void *)-1;
2721 else
2723 unsigned buflen = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (vr->type)) + 1;
2724 if (INTEGRAL_TYPE_P (vr->type))
2725 buflen = GET_MODE_SIZE (SCALAR_INT_TYPE_MODE (vr->type)) + 1;
2726 unsigned char *buf = XALLOCAVEC (unsigned char, buflen);
2727 memset (buf, TREE_INT_CST_LOW (gimple_call_arg (def_stmt, 1)),
2728 buflen);
2729 if (BYTES_BIG_ENDIAN)
2731 unsigned int amnt
2732 = (((unsigned HOST_WIDE_INT) offseti + sizei)
2733 % BITS_PER_UNIT);
2734 if (amnt)
2736 shift_bytes_in_array_right (buf, buflen,
2737 BITS_PER_UNIT - amnt);
2738 buf++;
2739 buflen--;
2742 else if (offseti % BITS_PER_UNIT != 0)
2744 unsigned int amnt
2745 = BITS_PER_UNIT - ((unsigned HOST_WIDE_INT) offseti
2746 % BITS_PER_UNIT);
2747 shift_bytes_in_array_left (buf, buflen, amnt);
2748 buf++;
2749 buflen--;
2751 val = native_interpret_expr (vr->type, buf, buflen);
2752 if (!val)
2753 return (void *)-1;
2755 return data->finish (0, 0, val);
2757 /* For now handle clearing memory with partial defs. */
2758 else if (known_eq (ref->size, maxsize)
2759 && integer_zerop (gimple_call_arg (def_stmt, 1))
2760 && tree_fits_poly_int64_p (len)
2761 && tree_to_poly_int64 (len).is_constant (&leni)
2762 && leni <= INTTYPE_MAXIMUM (HOST_WIDE_INT) / BITS_PER_UNIT
2763 && offset.is_constant (&offseti)
2764 && offset2.is_constant (&offset2i)
2765 && maxsize.is_constant (&maxsizei)
2766 && ranges_known_overlap_p (offseti, maxsizei, offset2i,
2767 leni << LOG2_BITS_PER_UNIT))
2769 pd_data pd;
2770 pd.rhs = build_constructor (NULL_TREE, NULL);
2771 pd.offset = offset2i;
2772 pd.size = leni << LOG2_BITS_PER_UNIT;
2773 return data->push_partial_def (pd, 0, 0, offseti, maxsizei);
2777 /* 2) Assignment from an empty CONSTRUCTOR. */
2778 else if (is_gimple_reg_type (vr->type)
2779 && gimple_assign_single_p (def_stmt)
2780 && gimple_assign_rhs_code (def_stmt) == CONSTRUCTOR
2781 && CONSTRUCTOR_NELTS (gimple_assign_rhs1 (def_stmt)) == 0)
2783 tree base2;
2784 poly_int64 offset2, size2, maxsize2;
2785 HOST_WIDE_INT offset2i, size2i;
2786 gcc_assert (lhs_ref_ok);
2787 base2 = ao_ref_base (&lhs_ref);
2788 offset2 = lhs_ref.offset;
2789 size2 = lhs_ref.size;
2790 maxsize2 = lhs_ref.max_size;
2791 if (known_size_p (maxsize2)
2792 && known_eq (maxsize2, size2)
2793 && adjust_offsets_for_equal_base_address (base, &offset,
2794 base2, &offset2))
2796 if (data->partial_defs.is_empty ()
2797 && known_subrange_p (offset, maxsize, offset2, size2))
2799 /* While technically undefined behavior do not optimize
2800 a full read from a clobber. */
2801 if (gimple_clobber_p (def_stmt))
2802 return (void *)-1;
2803 tree val = build_zero_cst (vr->type);
2804 return data->finish (ao_ref_alias_set (&lhs_ref),
2805 ao_ref_base_alias_set (&lhs_ref), val);
2807 else if (known_eq (ref->size, maxsize)
2808 && maxsize.is_constant (&maxsizei)
2809 && offset.is_constant (&offseti)
2810 && offset2.is_constant (&offset2i)
2811 && size2.is_constant (&size2i)
2812 && ranges_known_overlap_p (offseti, maxsizei,
2813 offset2i, size2i))
2815 /* Let clobbers be consumed by the partial-def tracker
2816 which can choose to ignore them if they are shadowed
2817 by a later def. */
2818 pd_data pd;
2819 pd.rhs = gimple_assign_rhs1 (def_stmt);
2820 pd.offset = offset2i;
2821 pd.size = size2i;
2822 return data->push_partial_def (pd, ao_ref_alias_set (&lhs_ref),
2823 ao_ref_base_alias_set (&lhs_ref),
2824 offseti, maxsizei);
2829 /* 3) Assignment from a constant. We can use folds native encode/interpret
2830 routines to extract the assigned bits. */
2831 else if (known_eq (ref->size, maxsize)
2832 && is_gimple_reg_type (vr->type)
2833 && !contains_storage_order_barrier_p (vr->operands)
2834 && gimple_assign_single_p (def_stmt)
2835 && CHAR_BIT == 8
2836 && BITS_PER_UNIT == 8
2837 && BYTES_BIG_ENDIAN == WORDS_BIG_ENDIAN
2838 /* native_encode and native_decode operate on arrays of bytes
2839 and so fundamentally need a compile-time size and offset. */
2840 && maxsize.is_constant (&maxsizei)
2841 && offset.is_constant (&offseti)
2842 && (is_gimple_min_invariant (gimple_assign_rhs1 (def_stmt))
2843 || (TREE_CODE (gimple_assign_rhs1 (def_stmt)) == SSA_NAME
2844 && is_gimple_min_invariant (SSA_VAL (gimple_assign_rhs1 (def_stmt))))))
2846 tree lhs = gimple_assign_lhs (def_stmt);
2847 tree base2;
2848 poly_int64 offset2, size2, maxsize2;
2849 HOST_WIDE_INT offset2i, size2i;
2850 bool reverse;
2851 gcc_assert (lhs_ref_ok);
2852 base2 = ao_ref_base (&lhs_ref);
2853 offset2 = lhs_ref.offset;
2854 size2 = lhs_ref.size;
2855 maxsize2 = lhs_ref.max_size;
2856 reverse = reverse_storage_order_for_component_p (lhs);
2857 if (base2
2858 && !reverse
2859 && !storage_order_barrier_p (lhs)
2860 && known_eq (maxsize2, size2)
2861 && adjust_offsets_for_equal_base_address (base, &offset,
2862 base2, &offset2)
2863 && offset.is_constant (&offseti)
2864 && offset2.is_constant (&offset2i)
2865 && size2.is_constant (&size2i))
2867 if (data->partial_defs.is_empty ()
2868 && known_subrange_p (offseti, maxsizei, offset2, size2))
2870 /* We support up to 512-bit values (for V8DFmode). */
2871 unsigned char buffer[65];
2872 int len;
2874 tree rhs = gimple_assign_rhs1 (def_stmt);
2875 if (TREE_CODE (rhs) == SSA_NAME)
2876 rhs = SSA_VAL (rhs);
2877 len = native_encode_expr (rhs,
2878 buffer, sizeof (buffer) - 1,
2879 (offseti - offset2i) / BITS_PER_UNIT);
2880 if (len > 0 && len * BITS_PER_UNIT >= maxsizei)
2882 tree type = vr->type;
2883 unsigned char *buf = buffer;
2884 unsigned int amnt = 0;
2885 /* Make sure to interpret in a type that has a range
2886 covering the whole access size. */
2887 if (INTEGRAL_TYPE_P (vr->type)
2888 && maxsizei != TYPE_PRECISION (vr->type))
2889 type = build_nonstandard_integer_type (maxsizei,
2890 TYPE_UNSIGNED (type));
2891 if (BYTES_BIG_ENDIAN)
2893 /* For big-endian native_encode_expr stored the rhs
2894 such that the LSB of it is the LSB of buffer[len - 1].
2895 That bit is stored into memory at position
2896 offset2 + size2 - 1, i.e. in byte
2897 base + (offset2 + size2 - 1) / BITS_PER_UNIT.
2898 E.g. for offset2 1 and size2 14, rhs -1 and memory
2899 previously cleared that is:
2901 01111111|11111110
2902 Now, if we want to extract offset 2 and size 12 from
2903 it using native_interpret_expr (which actually works
2904 for integral bitfield types in terms of byte size of
2905 the mode), the native_encode_expr stored the value
2906 into buffer as
2907 XX111111|11111111
2908 and returned len 2 (the X bits are outside of
2909 precision).
2910 Let sz be maxsize / BITS_PER_UNIT if not extracting
2911 a bitfield, and GET_MODE_SIZE otherwise.
2912 We need to align the LSB of the value we want to
2913 extract as the LSB of buf[sz - 1].
2914 The LSB from memory we need to read is at position
2915 offset + maxsize - 1. */
2916 HOST_WIDE_INT sz = maxsizei / BITS_PER_UNIT;
2917 if (INTEGRAL_TYPE_P (type))
2918 sz = GET_MODE_SIZE (SCALAR_INT_TYPE_MODE (type));
2919 amnt = ((unsigned HOST_WIDE_INT) offset2i + size2i
2920 - offseti - maxsizei) % BITS_PER_UNIT;
2921 if (amnt)
2922 shift_bytes_in_array_right (buffer, len, amnt);
2923 amnt = ((unsigned HOST_WIDE_INT) offset2i + size2i
2924 - offseti - maxsizei - amnt) / BITS_PER_UNIT;
2925 if ((unsigned HOST_WIDE_INT) sz + amnt > (unsigned) len)
2926 len = 0;
2927 else
2929 buf = buffer + len - sz - amnt;
2930 len -= (buf - buffer);
2933 else
2935 amnt = ((unsigned HOST_WIDE_INT) offset2i
2936 - offseti) % BITS_PER_UNIT;
2937 if (amnt)
2939 buffer[len] = 0;
2940 shift_bytes_in_array_left (buffer, len + 1, amnt);
2941 buf = buffer + 1;
2944 tree val = native_interpret_expr (type, buf, len);
2945 /* If we chop off bits because the types precision doesn't
2946 match the memory access size this is ok when optimizing
2947 reads but not when called from the DSE code during
2948 elimination. */
2949 if (val
2950 && type != vr->type)
2952 if (! int_fits_type_p (val, vr->type))
2953 val = NULL_TREE;
2954 else
2955 val = fold_convert (vr->type, val);
2958 if (val)
2959 return data->finish (ao_ref_alias_set (&lhs_ref),
2960 ao_ref_base_alias_set (&lhs_ref), val);
2963 else if (ranges_known_overlap_p (offseti, maxsizei, offset2i,
2964 size2i))
2966 pd_data pd;
2967 tree rhs = gimple_assign_rhs1 (def_stmt);
2968 if (TREE_CODE (rhs) == SSA_NAME)
2969 rhs = SSA_VAL (rhs);
2970 pd.rhs = rhs;
2971 pd.offset = offset2i;
2972 pd.size = size2i;
2973 return data->push_partial_def (pd, ao_ref_alias_set (&lhs_ref),
2974 ao_ref_base_alias_set (&lhs_ref),
2975 offseti, maxsizei);
2980 /* 4) Assignment from an SSA name which definition we may be able
2981 to access pieces from or we can combine to a larger entity. */
2982 else if (known_eq (ref->size, maxsize)
2983 && is_gimple_reg_type (vr->type)
2984 && !contains_storage_order_barrier_p (vr->operands)
2985 && gimple_assign_single_p (def_stmt)
2986 && TREE_CODE (gimple_assign_rhs1 (def_stmt)) == SSA_NAME)
2988 tree lhs = gimple_assign_lhs (def_stmt);
2989 tree base2;
2990 poly_int64 offset2, size2, maxsize2;
2991 HOST_WIDE_INT offset2i, size2i, offseti;
2992 bool reverse;
2993 gcc_assert (lhs_ref_ok);
2994 base2 = ao_ref_base (&lhs_ref);
2995 offset2 = lhs_ref.offset;
2996 size2 = lhs_ref.size;
2997 maxsize2 = lhs_ref.max_size;
2998 reverse = reverse_storage_order_for_component_p (lhs);
2999 tree def_rhs = gimple_assign_rhs1 (def_stmt);
3000 if (!reverse
3001 && !storage_order_barrier_p (lhs)
3002 && known_size_p (maxsize2)
3003 && known_eq (maxsize2, size2)
3004 && adjust_offsets_for_equal_base_address (base, &offset,
3005 base2, &offset2))
3007 if (data->partial_defs.is_empty ()
3008 && known_subrange_p (offset, maxsize, offset2, size2)
3009 /* ??? We can't handle bitfield precision extracts without
3010 either using an alternate type for the BIT_FIELD_REF and
3011 then doing a conversion or possibly adjusting the offset
3012 according to endianness. */
3013 && (! INTEGRAL_TYPE_P (vr->type)
3014 || known_eq (ref->size, TYPE_PRECISION (vr->type)))
3015 && multiple_p (ref->size, BITS_PER_UNIT))
3017 tree val = NULL_TREE;
3018 if (! INTEGRAL_TYPE_P (TREE_TYPE (def_rhs))
3019 || type_has_mode_precision_p (TREE_TYPE (def_rhs)))
3021 gimple_match_op op (gimple_match_cond::UNCOND,
3022 BIT_FIELD_REF, vr->type,
3023 SSA_VAL (def_rhs),
3024 bitsize_int (ref->size),
3025 bitsize_int (offset - offset2));
3026 val = vn_nary_build_or_lookup (&op);
3028 else if (known_eq (ref->size, size2))
3030 gimple_match_op op (gimple_match_cond::UNCOND,
3031 VIEW_CONVERT_EXPR, vr->type,
3032 SSA_VAL (def_rhs));
3033 val = vn_nary_build_or_lookup (&op);
3035 if (val
3036 && (TREE_CODE (val) != SSA_NAME
3037 || ! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (val)))
3038 return data->finish (ao_ref_alias_set (&lhs_ref),
3039 ao_ref_base_alias_set (&lhs_ref), val);
3041 else if (maxsize.is_constant (&maxsizei)
3042 && offset.is_constant (&offseti)
3043 && offset2.is_constant (&offset2i)
3044 && size2.is_constant (&size2i)
3045 && ranges_known_overlap_p (offset, maxsize, offset2, size2))
3047 pd_data pd;
3048 pd.rhs = SSA_VAL (def_rhs);
3049 pd.offset = offset2i;
3050 pd.size = size2i;
3051 return data->push_partial_def (pd, ao_ref_alias_set (&lhs_ref),
3052 ao_ref_base_alias_set (&lhs_ref),
3053 offseti, maxsizei);
3058 /* 5) For aggregate copies translate the reference through them if
3059 the copy kills ref. */
3060 else if (data->vn_walk_kind == VN_WALKREWRITE
3061 && gimple_assign_single_p (def_stmt)
3062 && (DECL_P (gimple_assign_rhs1 (def_stmt))
3063 || TREE_CODE (gimple_assign_rhs1 (def_stmt)) == MEM_REF
3064 || handled_component_p (gimple_assign_rhs1 (def_stmt))))
3066 tree base2;
3067 int i, j, k;
3068 auto_vec<vn_reference_op_s> rhs;
3069 vn_reference_op_t vro;
3070 ao_ref r;
3072 gcc_assert (lhs_ref_ok);
3074 /* See if the assignment kills REF. */
3075 base2 = ao_ref_base (&lhs_ref);
3076 if (!lhs_ref.max_size_known_p ()
3077 || (base != base2
3078 && (TREE_CODE (base) != MEM_REF
3079 || TREE_CODE (base2) != MEM_REF
3080 || TREE_OPERAND (base, 0) != TREE_OPERAND (base2, 0)
3081 || !tree_int_cst_equal (TREE_OPERAND (base, 1),
3082 TREE_OPERAND (base2, 1))))
3083 || !stmt_kills_ref_p (def_stmt, ref))
3084 return (void *)-1;
3086 /* Find the common base of ref and the lhs. lhs_ops already
3087 contains valueized operands for the lhs. */
3088 i = vr->operands.length () - 1;
3089 j = lhs_ops.length () - 1;
3090 while (j >= 0 && i >= 0
3091 && vn_reference_op_eq (&vr->operands[i], &lhs_ops[j]))
3093 i--;
3094 j--;
3097 /* ??? The innermost op should always be a MEM_REF and we already
3098 checked that the assignment to the lhs kills vr. Thus for
3099 aggregate copies using char[] types the vn_reference_op_eq
3100 may fail when comparing types for compatibility. But we really
3101 don't care here - further lookups with the rewritten operands
3102 will simply fail if we messed up types too badly. */
3103 poly_int64 extra_off = 0;
3104 if (j == 0 && i >= 0
3105 && lhs_ops[0].opcode == MEM_REF
3106 && maybe_ne (lhs_ops[0].off, -1))
3108 if (known_eq (lhs_ops[0].off, vr->operands[i].off))
3109 i--, j--;
3110 else if (vr->operands[i].opcode == MEM_REF
3111 && maybe_ne (vr->operands[i].off, -1))
3113 extra_off = vr->operands[i].off - lhs_ops[0].off;
3114 i--, j--;
3118 /* i now points to the first additional op.
3119 ??? LHS may not be completely contained in VR, one or more
3120 VIEW_CONVERT_EXPRs could be in its way. We could at least
3121 try handling outermost VIEW_CONVERT_EXPRs. */
3122 if (j != -1)
3123 return (void *)-1;
3125 /* Punt if the additional ops contain a storage order barrier. */
3126 for (k = i; k >= 0; k--)
3128 vro = &vr->operands[k];
3129 if (vro->opcode == VIEW_CONVERT_EXPR && vro->reverse)
3130 return (void *)-1;
3133 /* Now re-write REF to be based on the rhs of the assignment. */
3134 tree rhs1 = gimple_assign_rhs1 (def_stmt);
3135 copy_reference_ops_from_ref (rhs1, &rhs);
3137 /* Apply an extra offset to the inner MEM_REF of the RHS. */
3138 if (maybe_ne (extra_off, 0))
3140 if (rhs.length () < 2)
3141 return (void *)-1;
3142 int ix = rhs.length () - 2;
3143 if (rhs[ix].opcode != MEM_REF
3144 || known_eq (rhs[ix].off, -1))
3145 return (void *)-1;
3146 rhs[ix].off += extra_off;
3147 rhs[ix].op0 = int_const_binop (PLUS_EXPR, rhs[ix].op0,
3148 build_int_cst (TREE_TYPE (rhs[ix].op0),
3149 extra_off));
3152 /* Save the operands since we need to use the original ones for
3153 the hash entry we use. */
3154 if (!data->saved_operands.exists ())
3155 data->saved_operands = vr->operands.copy ();
3157 /* We need to pre-pend vr->operands[0..i] to rhs. */
3158 vec<vn_reference_op_s> old = vr->operands;
3159 if (i + 1 + rhs.length () > vr->operands.length ())
3160 vr->operands.safe_grow (i + 1 + rhs.length (), true);
3161 else
3162 vr->operands.truncate (i + 1 + rhs.length ());
3163 FOR_EACH_VEC_ELT (rhs, j, vro)
3164 vr->operands[i + 1 + j] = *vro;
3165 vr->operands = valueize_refs (vr->operands);
3166 if (old == shared_lookup_references)
3167 shared_lookup_references = vr->operands;
3168 vr->hashcode = vn_reference_compute_hash (vr);
3170 /* Try folding the new reference to a constant. */
3171 tree val = fully_constant_vn_reference_p (vr);
3172 if (val)
3174 if (data->partial_defs.is_empty ())
3175 return data->finish (ao_ref_alias_set (&lhs_ref),
3176 ao_ref_base_alias_set (&lhs_ref), val);
3177 /* This is the only interesting case for partial-def handling
3178 coming from targets that like to gimplify init-ctors as
3179 aggregate copies from constant data like aarch64 for
3180 PR83518. */
3181 if (maxsize.is_constant (&maxsizei) && known_eq (ref->size, maxsize))
3183 pd_data pd;
3184 pd.rhs = val;
3185 pd.offset = 0;
3186 pd.size = maxsizei;
3187 return data->push_partial_def (pd, ao_ref_alias_set (&lhs_ref),
3188 ao_ref_base_alias_set (&lhs_ref),
3189 0, maxsizei);
3193 /* Continuing with partial defs isn't easily possible here, we
3194 have to find a full def from further lookups from here. Probably
3195 not worth the special-casing everywhere. */
3196 if (!data->partial_defs.is_empty ())
3197 return (void *)-1;
3199 /* Adjust *ref from the new operands. */
3200 ao_ref rhs1_ref;
3201 ao_ref_init (&rhs1_ref, rhs1);
3202 if (!ao_ref_init_from_vn_reference (&r, ao_ref_alias_set (&rhs1_ref),
3203 ao_ref_base_alias_set (&rhs1_ref),
3204 vr->type, vr->operands))
3205 return (void *)-1;
3206 /* This can happen with bitfields. */
3207 if (maybe_ne (ref->size, r.size))
3208 return (void *)-1;
3209 *ref = r;
3211 /* Do not update last seen VUSE after translating. */
3212 data->last_vuse_ptr = NULL;
3213 /* Invalidate the original access path since it now contains
3214 the wrong base. */
3215 data->orig_ref.ref = NULL_TREE;
3216 /* Use the alias-set of this LHS for recording an eventual result. */
3217 if (data->first_set == -2)
3219 data->first_set = ao_ref_alias_set (&lhs_ref);
3220 data->first_base_set = ao_ref_base_alias_set (&lhs_ref);
3223 /* Keep looking for the adjusted *REF / VR pair. */
3224 return NULL;
3227 /* 6) For memcpy copies translate the reference through them if the copy
3228 kills ref. But we cannot (easily) do this translation if the memcpy is
3229 a storage order barrier, i.e. is equivalent to a VIEW_CONVERT_EXPR that
3230 can modify the storage order of objects (see storage_order_barrier_p). */
3231 else if (data->vn_walk_kind == VN_WALKREWRITE
3232 && is_gimple_reg_type (vr->type)
3233 /* ??? Handle BCOPY as well. */
3234 && (gimple_call_builtin_p (def_stmt, BUILT_IN_MEMCPY)
3235 || gimple_call_builtin_p (def_stmt, BUILT_IN_MEMCPY_CHK)
3236 || gimple_call_builtin_p (def_stmt, BUILT_IN_MEMPCPY)
3237 || gimple_call_builtin_p (def_stmt, BUILT_IN_MEMPCPY_CHK)
3238 || gimple_call_builtin_p (def_stmt, BUILT_IN_MEMMOVE)
3239 || gimple_call_builtin_p (def_stmt, BUILT_IN_MEMMOVE_CHK))
3240 && (TREE_CODE (gimple_call_arg (def_stmt, 0)) == ADDR_EXPR
3241 || TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME)
3242 && (TREE_CODE (gimple_call_arg (def_stmt, 1)) == ADDR_EXPR
3243 || TREE_CODE (gimple_call_arg (def_stmt, 1)) == SSA_NAME)
3244 && (poly_int_tree_p (gimple_call_arg (def_stmt, 2), &copy_size)
3245 || (TREE_CODE (gimple_call_arg (def_stmt, 2)) == SSA_NAME
3246 && poly_int_tree_p (SSA_VAL (gimple_call_arg (def_stmt, 2)),
3247 &copy_size)))
3248 /* Handling this is more complicated, give up for now. */
3249 && data->partial_defs.is_empty ())
3251 tree lhs, rhs;
3252 ao_ref r;
3253 poly_int64 rhs_offset, lhs_offset;
3254 vn_reference_op_s op;
3255 poly_uint64 mem_offset;
3256 poly_int64 at, byte_maxsize;
3258 /* Only handle non-variable, addressable refs. */
3259 if (maybe_ne (ref->size, maxsize)
3260 || !multiple_p (offset, BITS_PER_UNIT, &at)
3261 || !multiple_p (maxsize, BITS_PER_UNIT, &byte_maxsize))
3262 return (void *)-1;
3264 /* Extract a pointer base and an offset for the destination. */
3265 lhs = gimple_call_arg (def_stmt, 0);
3266 lhs_offset = 0;
3267 if (TREE_CODE (lhs) == SSA_NAME)
3269 lhs = vn_valueize (lhs);
3270 if (TREE_CODE (lhs) == SSA_NAME)
3272 gimple *def_stmt = SSA_NAME_DEF_STMT (lhs);
3273 if (gimple_assign_single_p (def_stmt)
3274 && gimple_assign_rhs_code (def_stmt) == ADDR_EXPR)
3275 lhs = gimple_assign_rhs1 (def_stmt);
3278 if (TREE_CODE (lhs) == ADDR_EXPR)
3280 if (AGGREGATE_TYPE_P (TREE_TYPE (TREE_TYPE (lhs)))
3281 && TYPE_REVERSE_STORAGE_ORDER (TREE_TYPE (TREE_TYPE (lhs))))
3282 return (void *)-1;
3283 tree tem = get_addr_base_and_unit_offset (TREE_OPERAND (lhs, 0),
3284 &lhs_offset);
3285 if (!tem)
3286 return (void *)-1;
3287 if (TREE_CODE (tem) == MEM_REF
3288 && poly_int_tree_p (TREE_OPERAND (tem, 1), &mem_offset))
3290 lhs = TREE_OPERAND (tem, 0);
3291 if (TREE_CODE (lhs) == SSA_NAME)
3292 lhs = vn_valueize (lhs);
3293 lhs_offset += mem_offset;
3295 else if (DECL_P (tem))
3296 lhs = build_fold_addr_expr (tem);
3297 else
3298 return (void *)-1;
3300 if (TREE_CODE (lhs) != SSA_NAME
3301 && TREE_CODE (lhs) != ADDR_EXPR)
3302 return (void *)-1;
3304 /* Extract a pointer base and an offset for the source. */
3305 rhs = gimple_call_arg (def_stmt, 1);
3306 rhs_offset = 0;
3307 if (TREE_CODE (rhs) == SSA_NAME)
3308 rhs = vn_valueize (rhs);
3309 if (TREE_CODE (rhs) == ADDR_EXPR)
3311 if (AGGREGATE_TYPE_P (TREE_TYPE (TREE_TYPE (rhs)))
3312 && TYPE_REVERSE_STORAGE_ORDER (TREE_TYPE (TREE_TYPE (rhs))))
3313 return (void *)-1;
3314 tree tem = get_addr_base_and_unit_offset (TREE_OPERAND (rhs, 0),
3315 &rhs_offset);
3316 if (!tem)
3317 return (void *)-1;
3318 if (TREE_CODE (tem) == MEM_REF
3319 && poly_int_tree_p (TREE_OPERAND (tem, 1), &mem_offset))
3321 rhs = TREE_OPERAND (tem, 0);
3322 rhs_offset += mem_offset;
3324 else if (DECL_P (tem)
3325 || TREE_CODE (tem) == STRING_CST)
3326 rhs = build_fold_addr_expr (tem);
3327 else
3328 return (void *)-1;
3330 if (TREE_CODE (rhs) == SSA_NAME)
3331 rhs = SSA_VAL (rhs);
3332 else if (TREE_CODE (rhs) != ADDR_EXPR)
3333 return (void *)-1;
3335 /* The bases of the destination and the references have to agree. */
3336 if (TREE_CODE (base) == MEM_REF)
3338 if (TREE_OPERAND (base, 0) != lhs
3339 || !poly_int_tree_p (TREE_OPERAND (base, 1), &mem_offset))
3340 return (void *) -1;
3341 at += mem_offset;
3343 else if (!DECL_P (base)
3344 || TREE_CODE (lhs) != ADDR_EXPR
3345 || TREE_OPERAND (lhs, 0) != base)
3346 return (void *)-1;
3348 /* If the access is completely outside of the memcpy destination
3349 area there is no aliasing. */
3350 if (!ranges_maybe_overlap_p (lhs_offset, copy_size, at, byte_maxsize))
3351 return NULL;
3352 /* And the access has to be contained within the memcpy destination. */
3353 if (!known_subrange_p (at, byte_maxsize, lhs_offset, copy_size))
3354 return (void *)-1;
3356 /* Save the operands since we need to use the original ones for
3357 the hash entry we use. */
3358 if (!data->saved_operands.exists ())
3359 data->saved_operands = vr->operands.copy ();
3361 /* Make room for 2 operands in the new reference. */
3362 if (vr->operands.length () < 2)
3364 vec<vn_reference_op_s> old = vr->operands;
3365 vr->operands.safe_grow_cleared (2, true);
3366 if (old == shared_lookup_references)
3367 shared_lookup_references = vr->operands;
3369 else
3370 vr->operands.truncate (2);
3372 /* The looked-through reference is a simple MEM_REF. */
3373 memset (&op, 0, sizeof (op));
3374 op.type = vr->type;
3375 op.opcode = MEM_REF;
3376 op.op0 = build_int_cst (ptr_type_node, at - lhs_offset + rhs_offset);
3377 op.off = at - lhs_offset + rhs_offset;
3378 vr->operands[0] = op;
3379 op.type = TREE_TYPE (rhs);
3380 op.opcode = TREE_CODE (rhs);
3381 op.op0 = rhs;
3382 op.off = -1;
3383 vr->operands[1] = op;
3384 vr->hashcode = vn_reference_compute_hash (vr);
3386 /* Try folding the new reference to a constant. */
3387 tree val = fully_constant_vn_reference_p (vr);
3388 if (val)
3389 return data->finish (0, 0, val);
3391 /* Adjust *ref from the new operands. */
3392 if (!ao_ref_init_from_vn_reference (&r, 0, 0, vr->type, vr->operands))
3393 return (void *)-1;
3394 /* This can happen with bitfields. */
3395 if (maybe_ne (ref->size, r.size))
3396 return (void *)-1;
3397 *ref = r;
3399 /* Do not update last seen VUSE after translating. */
3400 data->last_vuse_ptr = NULL;
3401 /* Invalidate the original access path since it now contains
3402 the wrong base. */
3403 data->orig_ref.ref = NULL_TREE;
3404 /* Use the alias-set of this stmt for recording an eventual result. */
3405 if (data->first_set == -2)
3407 data->first_set = 0;
3408 data->first_base_set = 0;
3411 /* Keep looking for the adjusted *REF / VR pair. */
3412 return NULL;
3415 /* Bail out and stop walking. */
3416 return (void *)-1;
3419 /* Return a reference op vector from OP that can be used for
3420 vn_reference_lookup_pieces. The caller is responsible for releasing
3421 the vector. */
3423 vec<vn_reference_op_s>
3424 vn_reference_operands_for_lookup (tree op)
3426 bool valueized;
3427 return valueize_shared_reference_ops_from_ref (op, &valueized).copy ();
3430 /* Lookup a reference operation by it's parts, in the current hash table.
3431 Returns the resulting value number if it exists in the hash table,
3432 NULL_TREE otherwise. VNRESULT will be filled in with the actual
3433 vn_reference_t stored in the hashtable if something is found. */
3435 tree
3436 vn_reference_lookup_pieces (tree vuse, alias_set_type set,
3437 alias_set_type base_set, tree type,
3438 vec<vn_reference_op_s> operands,
3439 vn_reference_t *vnresult, vn_lookup_kind kind)
3441 struct vn_reference_s vr1;
3442 vn_reference_t tmp;
3443 tree cst;
3445 if (!vnresult)
3446 vnresult = &tmp;
3447 *vnresult = NULL;
3449 vr1.vuse = vuse_ssa_val (vuse);
3450 shared_lookup_references.truncate (0);
3451 shared_lookup_references.safe_grow (operands.length (), true);
3452 memcpy (shared_lookup_references.address (),
3453 operands.address (),
3454 sizeof (vn_reference_op_s)
3455 * operands.length ());
3456 vr1.operands = operands = shared_lookup_references
3457 = valueize_refs (shared_lookup_references);
3458 vr1.type = type;
3459 vr1.set = set;
3460 vr1.base_set = base_set;
3461 vr1.hashcode = vn_reference_compute_hash (&vr1);
3462 if ((cst = fully_constant_vn_reference_p (&vr1)))
3463 return cst;
3465 vn_reference_lookup_1 (&vr1, vnresult);
3466 if (!*vnresult
3467 && kind != VN_NOWALK
3468 && vr1.vuse)
3470 ao_ref r;
3471 unsigned limit = param_sccvn_max_alias_queries_per_access;
3472 vn_walk_cb_data data (&vr1, NULL_TREE, NULL, kind, true, NULL_TREE);
3473 if (ao_ref_init_from_vn_reference (&r, set, base_set, type,
3474 vr1.operands))
3475 *vnresult
3476 = ((vn_reference_t)
3477 walk_non_aliased_vuses (&r, vr1.vuse, true, vn_reference_lookup_2,
3478 vn_reference_lookup_3, vuse_valueize,
3479 limit, &data));
3480 gcc_checking_assert (vr1.operands == shared_lookup_references);
3483 if (*vnresult)
3484 return (*vnresult)->result;
3486 return NULL_TREE;
3489 /* Lookup OP in the current hash table, and return the resulting value
3490 number if it exists in the hash table. Return NULL_TREE if it does
3491 not exist in the hash table or if the result field of the structure
3492 was NULL.. VNRESULT will be filled in with the vn_reference_t
3493 stored in the hashtable if one exists. When TBAA_P is false assume
3494 we are looking up a store and treat it as having alias-set zero.
3495 *LAST_VUSE_PTR will be updated with the VUSE the value lookup succeeded.
3496 MASK is either NULL_TREE, or can be an INTEGER_CST if the result of the
3497 load is bitwise anded with MASK and so we are only interested in a subset
3498 of the bits and can ignore if the other bits are uninitialized or
3499 not initialized with constants. */
3501 tree
3502 vn_reference_lookup (tree op, tree vuse, vn_lookup_kind kind,
3503 vn_reference_t *vnresult, bool tbaa_p,
3504 tree *last_vuse_ptr, tree mask)
3506 vec<vn_reference_op_s> operands;
3507 struct vn_reference_s vr1;
3508 bool valuezied_anything;
3510 if (vnresult)
3511 *vnresult = NULL;
3513 vr1.vuse = vuse_ssa_val (vuse);
3514 vr1.operands = operands
3515 = valueize_shared_reference_ops_from_ref (op, &valuezied_anything);
3516 vr1.type = TREE_TYPE (op);
3517 ao_ref op_ref;
3518 ao_ref_init (&op_ref, op);
3519 vr1.set = ao_ref_alias_set (&op_ref);
3520 vr1.base_set = ao_ref_base_alias_set (&op_ref);
3521 vr1.hashcode = vn_reference_compute_hash (&vr1);
3522 if (mask == NULL_TREE)
3523 if (tree cst = fully_constant_vn_reference_p (&vr1))
3524 return cst;
3526 if (kind != VN_NOWALK && vr1.vuse)
3528 vn_reference_t wvnresult;
3529 ao_ref r;
3530 unsigned limit = param_sccvn_max_alias_queries_per_access;
3531 /* Make sure to use a valueized reference if we valueized anything.
3532 Otherwise preserve the full reference for advanced TBAA. */
3533 if (!valuezied_anything
3534 || !ao_ref_init_from_vn_reference (&r, vr1.set, vr1.base_set,
3535 vr1.type, vr1.operands))
3536 ao_ref_init (&r, op);
3537 vn_walk_cb_data data (&vr1, r.ref ? NULL_TREE : op,
3538 last_vuse_ptr, kind, tbaa_p, mask);
3540 wvnresult
3541 = ((vn_reference_t)
3542 walk_non_aliased_vuses (&r, vr1.vuse, tbaa_p, vn_reference_lookup_2,
3543 vn_reference_lookup_3, vuse_valueize, limit,
3544 &data));
3545 gcc_checking_assert (vr1.operands == shared_lookup_references);
3546 if (wvnresult)
3548 gcc_assert (mask == NULL_TREE);
3549 if (vnresult)
3550 *vnresult = wvnresult;
3551 return wvnresult->result;
3553 else if (mask)
3554 return data.masked_result;
3556 return NULL_TREE;
3559 if (last_vuse_ptr)
3560 *last_vuse_ptr = vr1.vuse;
3561 if (mask)
3562 return NULL_TREE;
3563 return vn_reference_lookup_1 (&vr1, vnresult);
3566 /* Lookup CALL in the current hash table and return the entry in
3567 *VNRESULT if found. Populates *VR for the hashtable lookup. */
3569 void
3570 vn_reference_lookup_call (gcall *call, vn_reference_t *vnresult,
3571 vn_reference_t vr)
3573 if (vnresult)
3574 *vnresult = NULL;
3576 tree vuse = gimple_vuse (call);
3578 vr->vuse = vuse ? SSA_VAL (vuse) : NULL_TREE;
3579 vr->operands = valueize_shared_reference_ops_from_call (call);
3580 vr->type = gimple_expr_type (call);
3581 vr->punned = false;
3582 vr->set = 0;
3583 vr->base_set = 0;
3584 vr->hashcode = vn_reference_compute_hash (vr);
3585 vn_reference_lookup_1 (vr, vnresult);
3588 /* Insert OP into the current hash table with a value number of RESULT. */
3590 static void
3591 vn_reference_insert (tree op, tree result, tree vuse, tree vdef)
3593 vn_reference_s **slot;
3594 vn_reference_t vr1;
3595 bool tem;
3597 vr1 = XOBNEW (&vn_tables_obstack, vn_reference_s);
3598 if (TREE_CODE (result) == SSA_NAME)
3599 vr1->value_id = VN_INFO (result)->value_id;
3600 else
3601 vr1->value_id = get_or_alloc_constant_value_id (result);
3602 vr1->vuse = vuse_ssa_val (vuse);
3603 vr1->operands = valueize_shared_reference_ops_from_ref (op, &tem).copy ();
3604 vr1->type = TREE_TYPE (op);
3605 vr1->punned = false;
3606 ao_ref op_ref;
3607 ao_ref_init (&op_ref, op);
3608 vr1->set = ao_ref_alias_set (&op_ref);
3609 vr1->base_set = ao_ref_base_alias_set (&op_ref);
3610 vr1->hashcode = vn_reference_compute_hash (vr1);
3611 vr1->result = TREE_CODE (result) == SSA_NAME ? SSA_VAL (result) : result;
3612 vr1->result_vdef = vdef;
3614 slot = valid_info->references->find_slot_with_hash (vr1, vr1->hashcode,
3615 INSERT);
3617 /* Because IL walking on reference lookup can end up visiting
3618 a def that is only to be visited later in iteration order
3619 when we are about to make an irreducible region reducible
3620 the def can be effectively processed and its ref being inserted
3621 by vn_reference_lookup_3 already. So we cannot assert (!*slot)
3622 but save a lookup if we deal with already inserted refs here. */
3623 if (*slot)
3625 /* We cannot assert that we have the same value either because
3626 when disentangling an irreducible region we may end up visiting
3627 a use before the corresponding def. That's a missed optimization
3628 only though. See gcc.dg/tree-ssa/pr87126.c for example. */
3629 if (dump_file && (dump_flags & TDF_DETAILS)
3630 && !operand_equal_p ((*slot)->result, vr1->result, 0))
3632 fprintf (dump_file, "Keeping old value ");
3633 print_generic_expr (dump_file, (*slot)->result);
3634 fprintf (dump_file, " because of collision\n");
3636 free_reference (vr1);
3637 obstack_free (&vn_tables_obstack, vr1);
3638 return;
3641 *slot = vr1;
3642 vr1->next = last_inserted_ref;
3643 last_inserted_ref = vr1;
3646 /* Insert a reference by it's pieces into the current hash table with
3647 a value number of RESULT. Return the resulting reference
3648 structure we created. */
3650 vn_reference_t
3651 vn_reference_insert_pieces (tree vuse, alias_set_type set,
3652 alias_set_type base_set, tree type,
3653 vec<vn_reference_op_s> operands,
3654 tree result, unsigned int value_id)
3657 vn_reference_s **slot;
3658 vn_reference_t vr1;
3660 vr1 = XOBNEW (&vn_tables_obstack, vn_reference_s);
3661 vr1->value_id = value_id;
3662 vr1->vuse = vuse_ssa_val (vuse);
3663 vr1->operands = valueize_refs (operands);
3664 vr1->type = type;
3665 vr1->punned = false;
3666 vr1->set = set;
3667 vr1->base_set = base_set;
3668 vr1->hashcode = vn_reference_compute_hash (vr1);
3669 if (result && TREE_CODE (result) == SSA_NAME)
3670 result = SSA_VAL (result);
3671 vr1->result = result;
3673 slot = valid_info->references->find_slot_with_hash (vr1, vr1->hashcode,
3674 INSERT);
3676 /* At this point we should have all the things inserted that we have
3677 seen before, and we should never try inserting something that
3678 already exists. */
3679 gcc_assert (!*slot);
3681 *slot = vr1;
3682 vr1->next = last_inserted_ref;
3683 last_inserted_ref = vr1;
3684 return vr1;
3687 /* Compute and return the hash value for nary operation VBO1. */
3689 static hashval_t
3690 vn_nary_op_compute_hash (const vn_nary_op_t vno1)
3692 inchash::hash hstate;
3693 unsigned i;
3695 for (i = 0; i < vno1->length; ++i)
3696 if (TREE_CODE (vno1->op[i]) == SSA_NAME)
3697 vno1->op[i] = SSA_VAL (vno1->op[i]);
3699 if (((vno1->length == 2
3700 && commutative_tree_code (vno1->opcode))
3701 || (vno1->length == 3
3702 && commutative_ternary_tree_code (vno1->opcode)))
3703 && tree_swap_operands_p (vno1->op[0], vno1->op[1]))
3704 std::swap (vno1->op[0], vno1->op[1]);
3705 else if (TREE_CODE_CLASS (vno1->opcode) == tcc_comparison
3706 && tree_swap_operands_p (vno1->op[0], vno1->op[1]))
3708 std::swap (vno1->op[0], vno1->op[1]);
3709 vno1->opcode = swap_tree_comparison (vno1->opcode);
3712 hstate.add_int (vno1->opcode);
3713 for (i = 0; i < vno1->length; ++i)
3714 inchash::add_expr (vno1->op[i], hstate);
3716 return hstate.end ();
3719 /* Compare nary operations VNO1 and VNO2 and return true if they are
3720 equivalent. */
3722 bool
3723 vn_nary_op_eq (const_vn_nary_op_t const vno1, const_vn_nary_op_t const vno2)
3725 unsigned i;
3727 if (vno1->hashcode != vno2->hashcode)
3728 return false;
3730 if (vno1->length != vno2->length)
3731 return false;
3733 if (vno1->opcode != vno2->opcode
3734 || !types_compatible_p (vno1->type, vno2->type))
3735 return false;
3737 for (i = 0; i < vno1->length; ++i)
3738 if (!expressions_equal_p (vno1->op[i], vno2->op[i]))
3739 return false;
3741 /* BIT_INSERT_EXPR has an implict operand as the type precision
3742 of op1. Need to check to make sure they are the same. */
3743 if (vno1->opcode == BIT_INSERT_EXPR
3744 && TREE_CODE (vno1->op[1]) == INTEGER_CST
3745 && TYPE_PRECISION (TREE_TYPE (vno1->op[1]))
3746 != TYPE_PRECISION (TREE_TYPE (vno2->op[1])))
3747 return false;
3749 return true;
3752 /* Initialize VNO from the pieces provided. */
3754 static void
3755 init_vn_nary_op_from_pieces (vn_nary_op_t vno, unsigned int length,
3756 enum tree_code code, tree type, tree *ops)
3758 vno->opcode = code;
3759 vno->length = length;
3760 vno->type = type;
3761 memcpy (&vno->op[0], ops, sizeof (tree) * length);
3764 /* Return the number of operands for a vn_nary ops structure from STMT. */
3766 static unsigned int
3767 vn_nary_length_from_stmt (gimple *stmt)
3769 switch (gimple_assign_rhs_code (stmt))
3771 case REALPART_EXPR:
3772 case IMAGPART_EXPR:
3773 case VIEW_CONVERT_EXPR:
3774 return 1;
3776 case BIT_FIELD_REF:
3777 return 3;
3779 case CONSTRUCTOR:
3780 return CONSTRUCTOR_NELTS (gimple_assign_rhs1 (stmt));
3782 default:
3783 return gimple_num_ops (stmt) - 1;
3787 /* Initialize VNO from STMT. */
3789 static void
3790 init_vn_nary_op_from_stmt (vn_nary_op_t vno, gimple *stmt)
3792 unsigned i;
3794 vno->opcode = gimple_assign_rhs_code (stmt);
3795 vno->type = gimple_expr_type (stmt);
3796 switch (vno->opcode)
3798 case REALPART_EXPR:
3799 case IMAGPART_EXPR:
3800 case VIEW_CONVERT_EXPR:
3801 vno->length = 1;
3802 vno->op[0] = TREE_OPERAND (gimple_assign_rhs1 (stmt), 0);
3803 break;
3805 case BIT_FIELD_REF:
3806 vno->length = 3;
3807 vno->op[0] = TREE_OPERAND (gimple_assign_rhs1 (stmt), 0);
3808 vno->op[1] = TREE_OPERAND (gimple_assign_rhs1 (stmt), 1);
3809 vno->op[2] = TREE_OPERAND (gimple_assign_rhs1 (stmt), 2);
3810 break;
3812 case CONSTRUCTOR:
3813 vno->length = CONSTRUCTOR_NELTS (gimple_assign_rhs1 (stmt));
3814 for (i = 0; i < vno->length; ++i)
3815 vno->op[i] = CONSTRUCTOR_ELT (gimple_assign_rhs1 (stmt), i)->value;
3816 break;
3818 default:
3819 gcc_checking_assert (!gimple_assign_single_p (stmt));
3820 vno->length = gimple_num_ops (stmt) - 1;
3821 for (i = 0; i < vno->length; ++i)
3822 vno->op[i] = gimple_op (stmt, i + 1);
3826 /* Compute the hashcode for VNO and look for it in the hash table;
3827 return the resulting value number if it exists in the hash table.
3828 Return NULL_TREE if it does not exist in the hash table or if the
3829 result field of the operation is NULL. VNRESULT will contain the
3830 vn_nary_op_t from the hashtable if it exists. */
3832 static tree
3833 vn_nary_op_lookup_1 (vn_nary_op_t vno, vn_nary_op_t *vnresult)
3835 vn_nary_op_s **slot;
3837 if (vnresult)
3838 *vnresult = NULL;
3840 vno->hashcode = vn_nary_op_compute_hash (vno);
3841 slot = valid_info->nary->find_slot_with_hash (vno, vno->hashcode, NO_INSERT);
3842 if (!slot)
3843 return NULL_TREE;
3844 if (vnresult)
3845 *vnresult = *slot;
3846 return (*slot)->predicated_values ? NULL_TREE : (*slot)->u.result;
3849 /* Lookup a n-ary operation by its pieces and return the resulting value
3850 number if it exists in the hash table. Return NULL_TREE if it does
3851 not exist in the hash table or if the result field of the operation
3852 is NULL. VNRESULT will contain the vn_nary_op_t from the hashtable
3853 if it exists. */
3855 tree
3856 vn_nary_op_lookup_pieces (unsigned int length, enum tree_code code,
3857 tree type, tree *ops, vn_nary_op_t *vnresult)
3859 vn_nary_op_t vno1 = XALLOCAVAR (struct vn_nary_op_s,
3860 sizeof_vn_nary_op (length));
3861 init_vn_nary_op_from_pieces (vno1, length, code, type, ops);
3862 return vn_nary_op_lookup_1 (vno1, vnresult);
3865 /* Lookup the rhs of STMT in the current hash table, and return the resulting
3866 value number if it exists in the hash table. Return NULL_TREE if
3867 it does not exist in the hash table. VNRESULT will contain the
3868 vn_nary_op_t from the hashtable if it exists. */
3870 tree
3871 vn_nary_op_lookup_stmt (gimple *stmt, vn_nary_op_t *vnresult)
3873 vn_nary_op_t vno1
3874 = XALLOCAVAR (struct vn_nary_op_s,
3875 sizeof_vn_nary_op (vn_nary_length_from_stmt (stmt)));
3876 init_vn_nary_op_from_stmt (vno1, stmt);
3877 return vn_nary_op_lookup_1 (vno1, vnresult);
3880 /* Allocate a vn_nary_op_t with LENGTH operands on STACK. */
3882 static vn_nary_op_t
3883 alloc_vn_nary_op_noinit (unsigned int length, struct obstack *stack)
3885 return (vn_nary_op_t) obstack_alloc (stack, sizeof_vn_nary_op (length));
3888 /* Allocate and initialize a vn_nary_op_t on CURRENT_INFO's
3889 obstack. */
3891 static vn_nary_op_t
3892 alloc_vn_nary_op (unsigned int length, tree result, unsigned int value_id)
3894 vn_nary_op_t vno1 = alloc_vn_nary_op_noinit (length, &vn_tables_obstack);
3896 vno1->value_id = value_id;
3897 vno1->length = length;
3898 vno1->predicated_values = 0;
3899 vno1->u.result = result;
3901 return vno1;
3904 /* Insert VNO into TABLE. If COMPUTE_HASH is true, then compute
3905 VNO->HASHCODE first. */
3907 static vn_nary_op_t
3908 vn_nary_op_insert_into (vn_nary_op_t vno, vn_nary_op_table_type *table,
3909 bool compute_hash)
3911 vn_nary_op_s **slot;
3913 if (compute_hash)
3915 vno->hashcode = vn_nary_op_compute_hash (vno);
3916 gcc_assert (! vno->predicated_values
3917 || (! vno->u.values->next
3918 && vno->u.values->n == 1));
3921 slot = table->find_slot_with_hash (vno, vno->hashcode, INSERT);
3922 vno->unwind_to = *slot;
3923 if (*slot)
3925 /* Prefer non-predicated values.
3926 ??? Only if those are constant, otherwise, with constant predicated
3927 value, turn them into predicated values with entry-block validity
3928 (??? but we always find the first valid result currently). */
3929 if ((*slot)->predicated_values
3930 && ! vno->predicated_values)
3932 /* ??? We cannot remove *slot from the unwind stack list.
3933 For the moment we deal with this by skipping not found
3934 entries but this isn't ideal ... */
3935 *slot = vno;
3936 /* ??? Maintain a stack of states we can unwind in
3937 vn_nary_op_s? But how far do we unwind? In reality
3938 we need to push change records somewhere... Or not
3939 unwind vn_nary_op_s and linking them but instead
3940 unwind the results "list", linking that, which also
3941 doesn't move on hashtable resize. */
3942 /* We can also have a ->unwind_to recording *slot there.
3943 That way we can make u.values a fixed size array with
3944 recording the number of entries but of course we then
3945 have always N copies for each unwind_to-state. Or we
3946 make sure to only ever append and each unwinding will
3947 pop off one entry (but how to deal with predicated
3948 replaced with non-predicated here?) */
3949 vno->next = last_inserted_nary;
3950 last_inserted_nary = vno;
3951 return vno;
3953 else if (vno->predicated_values
3954 && ! (*slot)->predicated_values)
3955 return *slot;
3956 else if (vno->predicated_values
3957 && (*slot)->predicated_values)
3959 /* ??? Factor this all into a insert_single_predicated_value
3960 routine. */
3961 gcc_assert (!vno->u.values->next && vno->u.values->n == 1);
3962 basic_block vno_bb
3963 = BASIC_BLOCK_FOR_FN (cfun, vno->u.values->valid_dominated_by_p[0]);
3964 vn_pval *nval = vno->u.values;
3965 vn_pval **next = &vno->u.values;
3966 bool found = false;
3967 for (vn_pval *val = (*slot)->u.values; val; val = val->next)
3969 if (expressions_equal_p (val->result, vno->u.values->result))
3971 found = true;
3972 for (unsigned i = 0; i < val->n; ++i)
3974 basic_block val_bb
3975 = BASIC_BLOCK_FOR_FN (cfun,
3976 val->valid_dominated_by_p[i]);
3977 if (dominated_by_p (CDI_DOMINATORS, vno_bb, val_bb))
3978 /* Value registered with more generic predicate. */
3979 return *slot;
3980 else if (dominated_by_p (CDI_DOMINATORS, val_bb, vno_bb))
3981 /* Shouldn't happen, we insert in RPO order. */
3982 gcc_unreachable ();
3984 /* Append value. */
3985 *next = (vn_pval *) obstack_alloc (&vn_tables_obstack,
3986 sizeof (vn_pval)
3987 + val->n * sizeof (int));
3988 (*next)->next = NULL;
3989 (*next)->result = val->result;
3990 (*next)->n = val->n + 1;
3991 memcpy ((*next)->valid_dominated_by_p,
3992 val->valid_dominated_by_p,
3993 val->n * sizeof (int));
3994 (*next)->valid_dominated_by_p[val->n] = vno_bb->index;
3995 next = &(*next)->next;
3996 if (dump_file && (dump_flags & TDF_DETAILS))
3997 fprintf (dump_file, "Appending predicate to value.\n");
3998 continue;
4000 /* Copy other predicated values. */
4001 *next = (vn_pval *) obstack_alloc (&vn_tables_obstack,
4002 sizeof (vn_pval)
4003 + (val->n-1) * sizeof (int));
4004 memcpy (*next, val, sizeof (vn_pval) + (val->n-1) * sizeof (int));
4005 (*next)->next = NULL;
4006 next = &(*next)->next;
4008 if (!found)
4009 *next = nval;
4011 *slot = vno;
4012 vno->next = last_inserted_nary;
4013 last_inserted_nary = vno;
4014 return vno;
4017 /* While we do not want to insert things twice it's awkward to
4018 avoid it in the case where visit_nary_op pattern-matches stuff
4019 and ends up simplifying the replacement to itself. We then
4020 get two inserts, one from visit_nary_op and one from
4021 vn_nary_build_or_lookup.
4022 So allow inserts with the same value number. */
4023 if ((*slot)->u.result == vno->u.result)
4024 return *slot;
4027 /* ??? There's also optimistic vs. previous commited state merging
4028 that is problematic for the case of unwinding. */
4030 /* ??? We should return NULL if we do not use 'vno' and have the
4031 caller release it. */
4032 gcc_assert (!*slot);
4034 *slot = vno;
4035 vno->next = last_inserted_nary;
4036 last_inserted_nary = vno;
4037 return vno;
4040 /* Insert a n-ary operation into the current hash table using it's
4041 pieces. Return the vn_nary_op_t structure we created and put in
4042 the hashtable. */
4044 vn_nary_op_t
4045 vn_nary_op_insert_pieces (unsigned int length, enum tree_code code,
4046 tree type, tree *ops,
4047 tree result, unsigned int value_id)
4049 vn_nary_op_t vno1 = alloc_vn_nary_op (length, result, value_id);
4050 init_vn_nary_op_from_pieces (vno1, length, code, type, ops);
4051 return vn_nary_op_insert_into (vno1, valid_info->nary, true);
4054 static vn_nary_op_t
4055 vn_nary_op_insert_pieces_predicated (unsigned int length, enum tree_code code,
4056 tree type, tree *ops,
4057 tree result, unsigned int value_id,
4058 edge pred_e)
4060 /* ??? Currently tracking BBs. */
4061 if (! single_pred_p (pred_e->dest))
4063 /* Never record for backedges. */
4064 if (pred_e->flags & EDGE_DFS_BACK)
4065 return NULL;
4066 edge_iterator ei;
4067 edge e;
4068 int cnt = 0;
4069 /* Ignore backedges. */
4070 FOR_EACH_EDGE (e, ei, pred_e->dest->preds)
4071 if (! dominated_by_p (CDI_DOMINATORS, e->src, e->dest))
4072 cnt++;
4073 if (cnt != 1)
4074 return NULL;
4076 if (dump_file && (dump_flags & TDF_DETAILS)
4077 /* ??? Fix dumping, but currently we only get comparisons. */
4078 && TREE_CODE_CLASS (code) == tcc_comparison)
4080 fprintf (dump_file, "Recording on edge %d->%d ", pred_e->src->index,
4081 pred_e->dest->index);
4082 print_generic_expr (dump_file, ops[0], TDF_SLIM);
4083 fprintf (dump_file, " %s ", get_tree_code_name (code));
4084 print_generic_expr (dump_file, ops[1], TDF_SLIM);
4085 fprintf (dump_file, " == %s\n",
4086 integer_zerop (result) ? "false" : "true");
4088 vn_nary_op_t vno1 = alloc_vn_nary_op (length, NULL_TREE, value_id);
4089 init_vn_nary_op_from_pieces (vno1, length, code, type, ops);
4090 vno1->predicated_values = 1;
4091 vno1->u.values = (vn_pval *) obstack_alloc (&vn_tables_obstack,
4092 sizeof (vn_pval));
4093 vno1->u.values->next = NULL;
4094 vno1->u.values->result = result;
4095 vno1->u.values->n = 1;
4096 vno1->u.values->valid_dominated_by_p[0] = pred_e->dest->index;
4097 return vn_nary_op_insert_into (vno1, valid_info->nary, true);
4100 static bool
4101 dominated_by_p_w_unex (basic_block bb1, basic_block bb2);
4103 static tree
4104 vn_nary_op_get_predicated_value (vn_nary_op_t vno, basic_block bb)
4106 if (! vno->predicated_values)
4107 return vno->u.result;
4108 for (vn_pval *val = vno->u.values; val; val = val->next)
4109 for (unsigned i = 0; i < val->n; ++i)
4110 if (dominated_by_p_w_unex (bb,
4111 BASIC_BLOCK_FOR_FN
4112 (cfun, val->valid_dominated_by_p[i])))
4113 return val->result;
4114 return NULL_TREE;
4117 /* Insert the rhs of STMT into the current hash table with a value number of
4118 RESULT. */
4120 static vn_nary_op_t
4121 vn_nary_op_insert_stmt (gimple *stmt, tree result)
4123 vn_nary_op_t vno1
4124 = alloc_vn_nary_op (vn_nary_length_from_stmt (stmt),
4125 result, VN_INFO (result)->value_id);
4126 init_vn_nary_op_from_stmt (vno1, stmt);
4127 return vn_nary_op_insert_into (vno1, valid_info->nary, true);
4130 /* Compute a hashcode for PHI operation VP1 and return it. */
4132 static inline hashval_t
4133 vn_phi_compute_hash (vn_phi_t vp1)
4135 inchash::hash hstate (EDGE_COUNT (vp1->block->preds) > 2
4136 ? vp1->block->index : EDGE_COUNT (vp1->block->preds));
4137 tree phi1op;
4138 tree type;
4139 edge e;
4140 edge_iterator ei;
4142 /* If all PHI arguments are constants we need to distinguish
4143 the PHI node via its type. */
4144 type = vp1->type;
4145 hstate.merge_hash (vn_hash_type (type));
4147 FOR_EACH_EDGE (e, ei, vp1->block->preds)
4149 /* Don't hash backedge values they need to be handled as VN_TOP
4150 for optimistic value-numbering. */
4151 if (e->flags & EDGE_DFS_BACK)
4152 continue;
4154 phi1op = vp1->phiargs[e->dest_idx];
4155 if (phi1op == VN_TOP)
4156 continue;
4157 inchash::add_expr (phi1op, hstate);
4160 return hstate.end ();
4164 /* Return true if COND1 and COND2 represent the same condition, set
4165 *INVERTED_P if one needs to be inverted to make it the same as
4166 the other. */
4168 static bool
4169 cond_stmts_equal_p (gcond *cond1, tree lhs1, tree rhs1,
4170 gcond *cond2, tree lhs2, tree rhs2, bool *inverted_p)
4172 enum tree_code code1 = gimple_cond_code (cond1);
4173 enum tree_code code2 = gimple_cond_code (cond2);
4175 *inverted_p = false;
4176 if (code1 == code2)
4178 else if (code1 == swap_tree_comparison (code2))
4179 std::swap (lhs2, rhs2);
4180 else if (code1 == invert_tree_comparison (code2, HONOR_NANS (lhs2)))
4181 *inverted_p = true;
4182 else if (code1 == invert_tree_comparison
4183 (swap_tree_comparison (code2), HONOR_NANS (lhs2)))
4185 std::swap (lhs2, rhs2);
4186 *inverted_p = true;
4188 else
4189 return false;
4191 return ((expressions_equal_p (lhs1, lhs2)
4192 && expressions_equal_p (rhs1, rhs2))
4193 || (commutative_tree_code (code1)
4194 && expressions_equal_p (lhs1, rhs2)
4195 && expressions_equal_p (rhs1, lhs2)));
4198 /* Compare two phi entries for equality, ignoring VN_TOP arguments. */
4200 static int
4201 vn_phi_eq (const_vn_phi_t const vp1, const_vn_phi_t const vp2)
4203 if (vp1->hashcode != vp2->hashcode)
4204 return false;
4206 if (vp1->block != vp2->block)
4208 if (EDGE_COUNT (vp1->block->preds) != EDGE_COUNT (vp2->block->preds))
4209 return false;
4211 switch (EDGE_COUNT (vp1->block->preds))
4213 case 1:
4214 /* Single-arg PHIs are just copies. */
4215 break;
4217 case 2:
4219 /* Rule out backedges into the PHI. */
4220 if (vp1->block->loop_father->header == vp1->block
4221 || vp2->block->loop_father->header == vp2->block)
4222 return false;
4224 /* If the PHI nodes do not have compatible types
4225 they are not the same. */
4226 if (!types_compatible_p (vp1->type, vp2->type))
4227 return false;
4229 basic_block idom1
4230 = get_immediate_dominator (CDI_DOMINATORS, vp1->block);
4231 basic_block idom2
4232 = get_immediate_dominator (CDI_DOMINATORS, vp2->block);
4233 /* If the immediate dominator end in switch stmts multiple
4234 values may end up in the same PHI arg via intermediate
4235 CFG merges. */
4236 if (EDGE_COUNT (idom1->succs) != 2
4237 || EDGE_COUNT (idom2->succs) != 2)
4238 return false;
4240 /* Verify the controlling stmt is the same. */
4241 gcond *last1 = safe_dyn_cast <gcond *> (last_stmt (idom1));
4242 gcond *last2 = safe_dyn_cast <gcond *> (last_stmt (idom2));
4243 if (! last1 || ! last2)
4244 return false;
4245 bool inverted_p;
4246 if (! cond_stmts_equal_p (last1, vp1->cclhs, vp1->ccrhs,
4247 last2, vp2->cclhs, vp2->ccrhs,
4248 &inverted_p))
4249 return false;
4251 /* Get at true/false controlled edges into the PHI. */
4252 edge te1, te2, fe1, fe2;
4253 if (! extract_true_false_controlled_edges (idom1, vp1->block,
4254 &te1, &fe1)
4255 || ! extract_true_false_controlled_edges (idom2, vp2->block,
4256 &te2, &fe2))
4257 return false;
4259 /* Swap edges if the second condition is the inverted of the
4260 first. */
4261 if (inverted_p)
4262 std::swap (te2, fe2);
4264 /* ??? Handle VN_TOP specially. */
4265 if (! expressions_equal_p (vp1->phiargs[te1->dest_idx],
4266 vp2->phiargs[te2->dest_idx])
4267 || ! expressions_equal_p (vp1->phiargs[fe1->dest_idx],
4268 vp2->phiargs[fe2->dest_idx]))
4269 return false;
4271 return true;
4274 default:
4275 return false;
4279 /* If the PHI nodes do not have compatible types
4280 they are not the same. */
4281 if (!types_compatible_p (vp1->type, vp2->type))
4282 return false;
4284 /* Any phi in the same block will have it's arguments in the
4285 same edge order, because of how we store phi nodes. */
4286 for (unsigned i = 0; i < EDGE_COUNT (vp1->block->preds); ++i)
4288 tree phi1op = vp1->phiargs[i];
4289 tree phi2op = vp2->phiargs[i];
4290 if (phi1op == VN_TOP || phi2op == VN_TOP)
4291 continue;
4292 if (!expressions_equal_p (phi1op, phi2op))
4293 return false;
4296 return true;
4299 /* Lookup PHI in the current hash table, and return the resulting
4300 value number if it exists in the hash table. Return NULL_TREE if
4301 it does not exist in the hash table. */
4303 static tree
4304 vn_phi_lookup (gimple *phi, bool backedges_varying_p)
4306 vn_phi_s **slot;
4307 struct vn_phi_s *vp1;
4308 edge e;
4309 edge_iterator ei;
4311 vp1 = XALLOCAVAR (struct vn_phi_s,
4312 sizeof (struct vn_phi_s)
4313 + (gimple_phi_num_args (phi) - 1) * sizeof (tree));
4315 /* Canonicalize the SSA_NAME's to their value number. */
4316 FOR_EACH_EDGE (e, ei, gimple_bb (phi)->preds)
4318 tree def = PHI_ARG_DEF_FROM_EDGE (phi, e);
4319 if (TREE_CODE (def) == SSA_NAME
4320 && (!backedges_varying_p || !(e->flags & EDGE_DFS_BACK)))
4321 def = SSA_VAL (def);
4322 vp1->phiargs[e->dest_idx] = def;
4324 vp1->type = TREE_TYPE (gimple_phi_result (phi));
4325 vp1->block = gimple_bb (phi);
4326 /* Extract values of the controlling condition. */
4327 vp1->cclhs = NULL_TREE;
4328 vp1->ccrhs = NULL_TREE;
4329 basic_block idom1 = get_immediate_dominator (CDI_DOMINATORS, vp1->block);
4330 if (EDGE_COUNT (idom1->succs) == 2)
4331 if (gcond *last1 = safe_dyn_cast <gcond *> (last_stmt (idom1)))
4333 /* ??? We want to use SSA_VAL here. But possibly not
4334 allow VN_TOP. */
4335 vp1->cclhs = vn_valueize (gimple_cond_lhs (last1));
4336 vp1->ccrhs = vn_valueize (gimple_cond_rhs (last1));
4338 vp1->hashcode = vn_phi_compute_hash (vp1);
4339 slot = valid_info->phis->find_slot_with_hash (vp1, vp1->hashcode, NO_INSERT);
4340 if (!slot)
4341 return NULL_TREE;
4342 return (*slot)->result;
4345 /* Insert PHI into the current hash table with a value number of
4346 RESULT. */
4348 static vn_phi_t
4349 vn_phi_insert (gimple *phi, tree result, bool backedges_varying_p)
4351 vn_phi_s **slot;
4352 vn_phi_t vp1 = (vn_phi_t) obstack_alloc (&vn_tables_obstack,
4353 sizeof (vn_phi_s)
4354 + ((gimple_phi_num_args (phi) - 1)
4355 * sizeof (tree)));
4356 edge e;
4357 edge_iterator ei;
4359 /* Canonicalize the SSA_NAME's to their value number. */
4360 FOR_EACH_EDGE (e, ei, gimple_bb (phi)->preds)
4362 tree def = PHI_ARG_DEF_FROM_EDGE (phi, e);
4363 if (TREE_CODE (def) == SSA_NAME
4364 && (!backedges_varying_p || !(e->flags & EDGE_DFS_BACK)))
4365 def = SSA_VAL (def);
4366 vp1->phiargs[e->dest_idx] = def;
4368 vp1->value_id = VN_INFO (result)->value_id;
4369 vp1->type = TREE_TYPE (gimple_phi_result (phi));
4370 vp1->block = gimple_bb (phi);
4371 /* Extract values of the controlling condition. */
4372 vp1->cclhs = NULL_TREE;
4373 vp1->ccrhs = NULL_TREE;
4374 basic_block idom1 = get_immediate_dominator (CDI_DOMINATORS, vp1->block);
4375 if (EDGE_COUNT (idom1->succs) == 2)
4376 if (gcond *last1 = safe_dyn_cast <gcond *> (last_stmt (idom1)))
4378 /* ??? We want to use SSA_VAL here. But possibly not
4379 allow VN_TOP. */
4380 vp1->cclhs = vn_valueize (gimple_cond_lhs (last1));
4381 vp1->ccrhs = vn_valueize (gimple_cond_rhs (last1));
4383 vp1->result = result;
4384 vp1->hashcode = vn_phi_compute_hash (vp1);
4386 slot = valid_info->phis->find_slot_with_hash (vp1, vp1->hashcode, INSERT);
4387 gcc_assert (!*slot);
4389 *slot = vp1;
4390 vp1->next = last_inserted_phi;
4391 last_inserted_phi = vp1;
4392 return vp1;
4396 /* Return true if BB1 is dominated by BB2 taking into account edges
4397 that are not executable. */
4399 static bool
4400 dominated_by_p_w_unex (basic_block bb1, basic_block bb2)
4402 edge_iterator ei;
4403 edge e;
4405 if (dominated_by_p (CDI_DOMINATORS, bb1, bb2))
4406 return true;
4408 /* Before iterating we'd like to know if there exists a
4409 (executable) path from bb2 to bb1 at all, if not we can
4410 directly return false. For now simply iterate once. */
4412 /* Iterate to the single executable bb1 predecessor. */
4413 if (EDGE_COUNT (bb1->preds) > 1)
4415 edge prede = NULL;
4416 FOR_EACH_EDGE (e, ei, bb1->preds)
4417 if (e->flags & EDGE_EXECUTABLE)
4419 if (prede)
4421 prede = NULL;
4422 break;
4424 prede = e;
4426 if (prede)
4428 bb1 = prede->src;
4430 /* Re-do the dominance check with changed bb1. */
4431 if (dominated_by_p (CDI_DOMINATORS, bb1, bb2))
4432 return true;
4436 /* Iterate to the single executable bb2 successor. */
4437 edge succe = NULL;
4438 FOR_EACH_EDGE (e, ei, bb2->succs)
4439 if (e->flags & EDGE_EXECUTABLE)
4441 if (succe)
4443 succe = NULL;
4444 break;
4446 succe = e;
4448 if (succe)
4450 /* Verify the reached block is only reached through succe.
4451 If there is only one edge we can spare us the dominator
4452 check and iterate directly. */
4453 if (EDGE_COUNT (succe->dest->preds) > 1)
4455 FOR_EACH_EDGE (e, ei, succe->dest->preds)
4456 if (e != succe
4457 && (e->flags & EDGE_EXECUTABLE))
4459 succe = NULL;
4460 break;
4463 if (succe)
4465 bb2 = succe->dest;
4467 /* Re-do the dominance check with changed bb2. */
4468 if (dominated_by_p (CDI_DOMINATORS, bb1, bb2))
4469 return true;
4473 /* We could now iterate updating bb1 / bb2. */
4474 return false;
4477 /* Set the value number of FROM to TO, return true if it has changed
4478 as a result. */
4480 static inline bool
4481 set_ssa_val_to (tree from, tree to)
4483 vn_ssa_aux_t from_info = VN_INFO (from);
4484 tree currval = from_info->valnum; // SSA_VAL (from)
4485 poly_int64 toff, coff;
4486 bool curr_undefined = false;
4487 bool curr_invariant = false;
4489 /* The only thing we allow as value numbers are ssa_names
4490 and invariants. So assert that here. We don't allow VN_TOP
4491 as visiting a stmt should produce a value-number other than
4492 that.
4493 ??? Still VN_TOP can happen for unreachable code, so force
4494 it to varying in that case. Not all code is prepared to
4495 get VN_TOP on valueization. */
4496 if (to == VN_TOP)
4498 /* ??? When iterating and visiting PHI <undef, backedge-value>
4499 for the first time we rightfully get VN_TOP and we need to
4500 preserve that to optimize for example gcc.dg/tree-ssa/ssa-sccvn-2.c.
4501 With SCCVN we were simply lucky we iterated the other PHI
4502 cycles first and thus visited the backedge-value DEF. */
4503 if (currval == VN_TOP)
4504 goto set_and_exit;
4505 if (dump_file && (dump_flags & TDF_DETAILS))
4506 fprintf (dump_file, "Forcing value number to varying on "
4507 "receiving VN_TOP\n");
4508 to = from;
4511 gcc_checking_assert (to != NULL_TREE
4512 && ((TREE_CODE (to) == SSA_NAME
4513 && (to == from || SSA_VAL (to) == to))
4514 || is_gimple_min_invariant (to)));
4516 if (from != to)
4518 if (currval == from)
4520 if (dump_file && (dump_flags & TDF_DETAILS))
4522 fprintf (dump_file, "Not changing value number of ");
4523 print_generic_expr (dump_file, from);
4524 fprintf (dump_file, " from VARYING to ");
4525 print_generic_expr (dump_file, to);
4526 fprintf (dump_file, "\n");
4528 return false;
4530 curr_invariant = is_gimple_min_invariant (currval);
4531 curr_undefined = (TREE_CODE (currval) == SSA_NAME
4532 && ssa_undefined_value_p (currval, false));
4533 if (currval != VN_TOP
4534 && !curr_invariant
4535 && !curr_undefined
4536 && is_gimple_min_invariant (to))
4538 if (dump_file && (dump_flags & TDF_DETAILS))
4540 fprintf (dump_file, "Forcing VARYING instead of changing "
4541 "value number of ");
4542 print_generic_expr (dump_file, from);
4543 fprintf (dump_file, " from ");
4544 print_generic_expr (dump_file, currval);
4545 fprintf (dump_file, " (non-constant) to ");
4546 print_generic_expr (dump_file, to);
4547 fprintf (dump_file, " (constant)\n");
4549 to = from;
4551 else if (currval != VN_TOP
4552 && !curr_undefined
4553 && TREE_CODE (to) == SSA_NAME
4554 && ssa_undefined_value_p (to, false))
4556 if (dump_file && (dump_flags & TDF_DETAILS))
4558 fprintf (dump_file, "Forcing VARYING instead of changing "
4559 "value number of ");
4560 print_generic_expr (dump_file, from);
4561 fprintf (dump_file, " from ");
4562 print_generic_expr (dump_file, currval);
4563 fprintf (dump_file, " (non-undefined) to ");
4564 print_generic_expr (dump_file, to);
4565 fprintf (dump_file, " (undefined)\n");
4567 to = from;
4569 else if (TREE_CODE (to) == SSA_NAME
4570 && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (to))
4571 to = from;
4574 set_and_exit:
4575 if (dump_file && (dump_flags & TDF_DETAILS))
4577 fprintf (dump_file, "Setting value number of ");
4578 print_generic_expr (dump_file, from);
4579 fprintf (dump_file, " to ");
4580 print_generic_expr (dump_file, to);
4583 if (currval != to
4584 && !operand_equal_p (currval, to, 0)
4585 /* Different undefined SSA names are not actually different. See
4586 PR82320 for a testcase were we'd otherwise not terminate iteration. */
4587 && !(curr_undefined
4588 && TREE_CODE (to) == SSA_NAME
4589 && ssa_undefined_value_p (to, false))
4590 /* ??? For addresses involving volatile objects or types operand_equal_p
4591 does not reliably detect ADDR_EXPRs as equal. We know we are only
4592 getting invariant gimple addresses here, so can use
4593 get_addr_base_and_unit_offset to do this comparison. */
4594 && !(TREE_CODE (currval) == ADDR_EXPR
4595 && TREE_CODE (to) == ADDR_EXPR
4596 && (get_addr_base_and_unit_offset (TREE_OPERAND (currval, 0), &coff)
4597 == get_addr_base_and_unit_offset (TREE_OPERAND (to, 0), &toff))
4598 && known_eq (coff, toff)))
4600 if (to != from
4601 && currval != VN_TOP
4602 && !curr_undefined
4603 /* We do not want to allow lattice transitions from one value
4604 to another since that may lead to not terminating iteration
4605 (see PR95049). Since there's no convenient way to check
4606 for the allowed transition of VAL -> PHI (loop entry value,
4607 same on two PHIs, to same PHI result) we restrict the check
4608 to invariants. */
4609 && curr_invariant
4610 && is_gimple_min_invariant (to))
4612 if (dump_file && (dump_flags & TDF_DETAILS))
4613 fprintf (dump_file, " forced VARYING");
4614 to = from;
4616 if (dump_file && (dump_flags & TDF_DETAILS))
4617 fprintf (dump_file, " (changed)\n");
4618 from_info->valnum = to;
4619 return true;
4621 if (dump_file && (dump_flags & TDF_DETAILS))
4622 fprintf (dump_file, "\n");
4623 return false;
4626 /* Set all definitions in STMT to value number to themselves.
4627 Return true if a value number changed. */
4629 static bool
4630 defs_to_varying (gimple *stmt)
4632 bool changed = false;
4633 ssa_op_iter iter;
4634 def_operand_p defp;
4636 FOR_EACH_SSA_DEF_OPERAND (defp, stmt, iter, SSA_OP_ALL_DEFS)
4638 tree def = DEF_FROM_PTR (defp);
4639 changed |= set_ssa_val_to (def, def);
4641 return changed;
4644 /* Visit a copy between LHS and RHS, return true if the value number
4645 changed. */
4647 static bool
4648 visit_copy (tree lhs, tree rhs)
4650 /* Valueize. */
4651 rhs = SSA_VAL (rhs);
4653 return set_ssa_val_to (lhs, rhs);
4656 /* Lookup a value for OP in type WIDE_TYPE where the value in type of OP
4657 is the same. */
4659 static tree
4660 valueized_wider_op (tree wide_type, tree op)
4662 if (TREE_CODE (op) == SSA_NAME)
4663 op = vn_valueize (op);
4665 /* Either we have the op widened available. */
4666 tree ops[3] = {};
4667 ops[0] = op;
4668 tree tem = vn_nary_op_lookup_pieces (1, NOP_EXPR,
4669 wide_type, ops, NULL);
4670 if (tem)
4671 return tem;
4673 /* Or the op is truncated from some existing value. */
4674 if (TREE_CODE (op) == SSA_NAME)
4676 gimple *def = SSA_NAME_DEF_STMT (op);
4677 if (is_gimple_assign (def)
4678 && CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (def)))
4680 tem = gimple_assign_rhs1 (def);
4681 if (useless_type_conversion_p (wide_type, TREE_TYPE (tem)))
4683 if (TREE_CODE (tem) == SSA_NAME)
4684 tem = vn_valueize (tem);
4685 return tem;
4690 /* For constants simply extend it. */
4691 if (TREE_CODE (op) == INTEGER_CST)
4692 return wide_int_to_tree (wide_type, wi::to_wide (op));
4694 return NULL_TREE;
4697 /* Visit a nary operator RHS, value number it, and return true if the
4698 value number of LHS has changed as a result. */
4700 static bool
4701 visit_nary_op (tree lhs, gassign *stmt)
4703 vn_nary_op_t vnresult;
4704 tree result = vn_nary_op_lookup_stmt (stmt, &vnresult);
4705 if (! result && vnresult)
4706 result = vn_nary_op_get_predicated_value (vnresult, gimple_bb (stmt));
4707 if (result)
4708 return set_ssa_val_to (lhs, result);
4710 /* Do some special pattern matching for redundancies of operations
4711 in different types. */
4712 enum tree_code code = gimple_assign_rhs_code (stmt);
4713 tree type = TREE_TYPE (lhs);
4714 tree rhs1 = gimple_assign_rhs1 (stmt);
4715 switch (code)
4717 CASE_CONVERT:
4718 /* Match arithmetic done in a different type where we can easily
4719 substitute the result from some earlier sign-changed or widened
4720 operation. */
4721 if (INTEGRAL_TYPE_P (type)
4722 && TREE_CODE (rhs1) == SSA_NAME
4723 /* We only handle sign-changes, zero-extension -> & mask or
4724 sign-extension if we know the inner operation doesn't
4725 overflow. */
4726 && (((TYPE_UNSIGNED (TREE_TYPE (rhs1))
4727 || (INTEGRAL_TYPE_P (TREE_TYPE (rhs1))
4728 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (rhs1))))
4729 && TYPE_PRECISION (type) > TYPE_PRECISION (TREE_TYPE (rhs1)))
4730 || TYPE_PRECISION (type) == TYPE_PRECISION (TREE_TYPE (rhs1))))
4732 gassign *def = dyn_cast <gassign *> (SSA_NAME_DEF_STMT (rhs1));
4733 if (def
4734 && (gimple_assign_rhs_code (def) == PLUS_EXPR
4735 || gimple_assign_rhs_code (def) == MINUS_EXPR
4736 || gimple_assign_rhs_code (def) == MULT_EXPR))
4738 tree ops[3] = {};
4739 /* Either we have the op widened available. */
4740 ops[0] = valueized_wider_op (type,
4741 gimple_assign_rhs1 (def));
4742 if (ops[0])
4743 ops[1] = valueized_wider_op (type,
4744 gimple_assign_rhs2 (def));
4745 if (ops[0] && ops[1])
4747 ops[0] = vn_nary_op_lookup_pieces
4748 (2, gimple_assign_rhs_code (def), type, ops, NULL);
4749 /* We have wider operation available. */
4750 if (ops[0]
4751 /* If the leader is a wrapping operation we can
4752 insert it for code hoisting w/o introducing
4753 undefined overflow. If it is not it has to
4754 be available. See PR86554. */
4755 && (TYPE_OVERFLOW_WRAPS (TREE_TYPE (ops[0]))
4756 || (rpo_avail && vn_context_bb
4757 && rpo_avail->eliminate_avail (vn_context_bb,
4758 ops[0]))))
4760 unsigned lhs_prec = TYPE_PRECISION (type);
4761 unsigned rhs_prec = TYPE_PRECISION (TREE_TYPE (rhs1));
4762 if (lhs_prec == rhs_prec
4763 || (INTEGRAL_TYPE_P (TREE_TYPE (rhs1))
4764 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (rhs1))))
4766 gimple_match_op match_op (gimple_match_cond::UNCOND,
4767 NOP_EXPR, type, ops[0]);
4768 result = vn_nary_build_or_lookup (&match_op);
4769 if (result)
4771 bool changed = set_ssa_val_to (lhs, result);
4772 vn_nary_op_insert_stmt (stmt, result);
4773 return changed;
4776 else
4778 tree mask = wide_int_to_tree
4779 (type, wi::mask (rhs_prec, false, lhs_prec));
4780 gimple_match_op match_op (gimple_match_cond::UNCOND,
4781 BIT_AND_EXPR,
4782 TREE_TYPE (lhs),
4783 ops[0], mask);
4784 result = vn_nary_build_or_lookup (&match_op);
4785 if (result)
4787 bool changed = set_ssa_val_to (lhs, result);
4788 vn_nary_op_insert_stmt (stmt, result);
4789 return changed;
4796 break;
4797 case BIT_AND_EXPR:
4798 if (INTEGRAL_TYPE_P (type)
4799 && TREE_CODE (rhs1) == SSA_NAME
4800 && TREE_CODE (gimple_assign_rhs2 (stmt)) == INTEGER_CST
4801 && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (rhs1)
4802 && default_vn_walk_kind != VN_NOWALK
4803 && CHAR_BIT == 8
4804 && BITS_PER_UNIT == 8
4805 && BYTES_BIG_ENDIAN == WORDS_BIG_ENDIAN
4806 && !integer_all_onesp (gimple_assign_rhs2 (stmt))
4807 && !integer_zerop (gimple_assign_rhs2 (stmt)))
4809 gassign *ass = dyn_cast <gassign *> (SSA_NAME_DEF_STMT (rhs1));
4810 if (ass
4811 && !gimple_has_volatile_ops (ass)
4812 && vn_get_stmt_kind (ass) == VN_REFERENCE)
4814 tree last_vuse = gimple_vuse (ass);
4815 tree op = gimple_assign_rhs1 (ass);
4816 tree result = vn_reference_lookup (op, gimple_vuse (ass),
4817 default_vn_walk_kind,
4818 NULL, true, &last_vuse,
4819 gimple_assign_rhs2 (stmt));
4820 if (result
4821 && useless_type_conversion_p (TREE_TYPE (result),
4822 TREE_TYPE (op)))
4823 return set_ssa_val_to (lhs, result);
4826 break;
4827 case TRUNC_DIV_EXPR:
4828 if (TYPE_UNSIGNED (type))
4829 break;
4830 /* Fallthru. */
4831 case RDIV_EXPR:
4832 case MULT_EXPR:
4833 /* Match up ([-]a){/,*}([-])b with v=a{/,*}b, replacing it with -v. */
4834 if (! HONOR_SIGN_DEPENDENT_ROUNDING (type))
4836 tree rhs[2];
4837 rhs[0] = rhs1;
4838 rhs[1] = gimple_assign_rhs2 (stmt);
4839 for (unsigned i = 0; i <= 1; ++i)
4841 unsigned j = i == 0 ? 1 : 0;
4842 tree ops[2];
4843 gimple_match_op match_op (gimple_match_cond::UNCOND,
4844 NEGATE_EXPR, type, rhs[i]);
4845 ops[i] = vn_nary_build_or_lookup_1 (&match_op, false);
4846 ops[j] = rhs[j];
4847 if (ops[i]
4848 && (ops[0] = vn_nary_op_lookup_pieces (2, code,
4849 type, ops, NULL)))
4851 gimple_match_op match_op (gimple_match_cond::UNCOND,
4852 NEGATE_EXPR, type, ops[0]);
4853 result = vn_nary_build_or_lookup (&match_op);
4854 if (result)
4856 bool changed = set_ssa_val_to (lhs, result);
4857 vn_nary_op_insert_stmt (stmt, result);
4858 return changed;
4863 break;
4864 default:
4865 break;
4868 bool changed = set_ssa_val_to (lhs, lhs);
4869 vn_nary_op_insert_stmt (stmt, lhs);
4870 return changed;
4873 /* Visit a call STMT storing into LHS. Return true if the value number
4874 of the LHS has changed as a result. */
4876 static bool
4877 visit_reference_op_call (tree lhs, gcall *stmt)
4879 bool changed = false;
4880 struct vn_reference_s vr1;
4881 vn_reference_t vnresult = NULL;
4882 tree vdef = gimple_vdef (stmt);
4884 /* Non-ssa lhs is handled in copy_reference_ops_from_call. */
4885 if (lhs && TREE_CODE (lhs) != SSA_NAME)
4886 lhs = NULL_TREE;
4888 vn_reference_lookup_call (stmt, &vnresult, &vr1);
4889 if (vnresult)
4891 if (vnresult->result_vdef && vdef)
4892 changed |= set_ssa_val_to (vdef, vnresult->result_vdef);
4893 else if (vdef)
4894 /* If the call was discovered to be pure or const reflect
4895 that as far as possible. */
4896 changed |= set_ssa_val_to (vdef, vuse_ssa_val (gimple_vuse (stmt)));
4898 if (!vnresult->result && lhs)
4899 vnresult->result = lhs;
4901 if (vnresult->result && lhs)
4902 changed |= set_ssa_val_to (lhs, vnresult->result);
4904 else
4906 vn_reference_t vr2;
4907 vn_reference_s **slot;
4908 tree vdef_val = vdef;
4909 if (vdef)
4911 /* If we value numbered an indirect functions function to
4912 one not clobbering memory value number its VDEF to its
4913 VUSE. */
4914 tree fn = gimple_call_fn (stmt);
4915 if (fn && TREE_CODE (fn) == SSA_NAME)
4917 fn = SSA_VAL (fn);
4918 if (TREE_CODE (fn) == ADDR_EXPR
4919 && TREE_CODE (TREE_OPERAND (fn, 0)) == FUNCTION_DECL
4920 && (flags_from_decl_or_type (TREE_OPERAND (fn, 0))
4921 & (ECF_CONST | ECF_PURE)))
4922 vdef_val = vuse_ssa_val (gimple_vuse (stmt));
4924 changed |= set_ssa_val_to (vdef, vdef_val);
4926 if (lhs)
4927 changed |= set_ssa_val_to (lhs, lhs);
4928 vr2 = XOBNEW (&vn_tables_obstack, vn_reference_s);
4929 vr2->vuse = vr1.vuse;
4930 /* As we are not walking the virtual operand chain we know the
4931 shared_lookup_references are still original so we can re-use
4932 them here. */
4933 vr2->operands = vr1.operands.copy ();
4934 vr2->type = vr1.type;
4935 vr2->punned = vr1.punned;
4936 vr2->set = vr1.set;
4937 vr2->base_set = vr1.base_set;
4938 vr2->hashcode = vr1.hashcode;
4939 vr2->result = lhs;
4940 vr2->result_vdef = vdef_val;
4941 vr2->value_id = 0;
4942 slot = valid_info->references->find_slot_with_hash (vr2, vr2->hashcode,
4943 INSERT);
4944 gcc_assert (!*slot);
4945 *slot = vr2;
4946 vr2->next = last_inserted_ref;
4947 last_inserted_ref = vr2;
4950 return changed;
4953 /* Visit a load from a reference operator RHS, part of STMT, value number it,
4954 and return true if the value number of the LHS has changed as a result. */
4956 static bool
4957 visit_reference_op_load (tree lhs, tree op, gimple *stmt)
4959 bool changed = false;
4960 tree last_vuse;
4961 tree result;
4962 vn_reference_t res;
4964 last_vuse = gimple_vuse (stmt);
4965 result = vn_reference_lookup (op, gimple_vuse (stmt),
4966 default_vn_walk_kind, &res, true, &last_vuse);
4968 /* We handle type-punning through unions by value-numbering based
4969 on offset and size of the access. Be prepared to handle a
4970 type-mismatch here via creating a VIEW_CONVERT_EXPR. */
4971 if (result
4972 && !useless_type_conversion_p (TREE_TYPE (result), TREE_TYPE (op)))
4974 /* Avoid the type punning in case the result mode has padding where
4975 the op we lookup has not. */
4976 if (maybe_lt (GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (result))),
4977 GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (op)))))
4978 result = NULL_TREE;
4979 else
4981 /* We will be setting the value number of lhs to the value number
4982 of VIEW_CONVERT_EXPR <TREE_TYPE (result)> (result).
4983 So first simplify and lookup this expression to see if it
4984 is already available. */
4985 gimple_match_op res_op (gimple_match_cond::UNCOND,
4986 VIEW_CONVERT_EXPR, TREE_TYPE (op), result);
4987 result = vn_nary_build_or_lookup (&res_op);
4988 if (result
4989 && TREE_CODE (result) == SSA_NAME
4990 && VN_INFO (result)->needs_insertion)
4991 /* Track whether this is the canonical expression for different
4992 typed loads. We use that as a stopgap measure for code
4993 hoisting when dealing with floating point loads. */
4994 res->punned = true;
4997 /* When building the conversion fails avoid inserting the reference
4998 again. */
4999 if (!result)
5000 return set_ssa_val_to (lhs, lhs);
5003 if (result)
5004 changed = set_ssa_val_to (lhs, result);
5005 else
5007 changed = set_ssa_val_to (lhs, lhs);
5008 vn_reference_insert (op, lhs, last_vuse, NULL_TREE);
5011 return changed;
5015 /* Visit a store to a reference operator LHS, part of STMT, value number it,
5016 and return true if the value number of the LHS has changed as a result. */
5018 static bool
5019 visit_reference_op_store (tree lhs, tree op, gimple *stmt)
5021 bool changed = false;
5022 vn_reference_t vnresult = NULL;
5023 tree assign;
5024 bool resultsame = false;
5025 tree vuse = gimple_vuse (stmt);
5026 tree vdef = gimple_vdef (stmt);
5028 if (TREE_CODE (op) == SSA_NAME)
5029 op = SSA_VAL (op);
5031 /* First we want to lookup using the *vuses* from the store and see
5032 if there the last store to this location with the same address
5033 had the same value.
5035 The vuses represent the memory state before the store. If the
5036 memory state, address, and value of the store is the same as the
5037 last store to this location, then this store will produce the
5038 same memory state as that store.
5040 In this case the vdef versions for this store are value numbered to those
5041 vuse versions, since they represent the same memory state after
5042 this store.
5044 Otherwise, the vdefs for the store are used when inserting into
5045 the table, since the store generates a new memory state. */
5047 vn_reference_lookup (lhs, vuse, VN_NOWALK, &vnresult, false);
5048 if (vnresult
5049 && vnresult->result)
5051 tree result = vnresult->result;
5052 gcc_checking_assert (TREE_CODE (result) != SSA_NAME
5053 || result == SSA_VAL (result));
5054 resultsame = expressions_equal_p (result, op);
5055 if (resultsame)
5057 /* If the TBAA state isn't compatible for downstream reads
5058 we cannot value-number the VDEFs the same. */
5059 ao_ref lhs_ref;
5060 ao_ref_init (&lhs_ref, lhs);
5061 alias_set_type set = ao_ref_alias_set (&lhs_ref);
5062 alias_set_type base_set = ao_ref_base_alias_set (&lhs_ref);
5063 if ((vnresult->set != set
5064 && ! alias_set_subset_of (set, vnresult->set))
5065 || (vnresult->base_set != base_set
5066 && ! alias_set_subset_of (base_set, vnresult->base_set)))
5067 resultsame = false;
5071 if (!resultsame)
5073 /* Only perform the following when being called from PRE
5074 which embeds tail merging. */
5075 if (default_vn_walk_kind == VN_WALK)
5077 assign = build2 (MODIFY_EXPR, TREE_TYPE (lhs), lhs, op);
5078 vn_reference_lookup (assign, vuse, VN_NOWALK, &vnresult, false);
5079 if (vnresult)
5081 VN_INFO (vdef)->visited = true;
5082 return set_ssa_val_to (vdef, vnresult->result_vdef);
5086 if (dump_file && (dump_flags & TDF_DETAILS))
5088 fprintf (dump_file, "No store match\n");
5089 fprintf (dump_file, "Value numbering store ");
5090 print_generic_expr (dump_file, lhs);
5091 fprintf (dump_file, " to ");
5092 print_generic_expr (dump_file, op);
5093 fprintf (dump_file, "\n");
5095 /* Have to set value numbers before insert, since insert is
5096 going to valueize the references in-place. */
5097 if (vdef)
5098 changed |= set_ssa_val_to (vdef, vdef);
5100 /* Do not insert structure copies into the tables. */
5101 if (is_gimple_min_invariant (op)
5102 || is_gimple_reg (op))
5103 vn_reference_insert (lhs, op, vdef, NULL);
5105 /* Only perform the following when being called from PRE
5106 which embeds tail merging. */
5107 if (default_vn_walk_kind == VN_WALK)
5109 assign = build2 (MODIFY_EXPR, TREE_TYPE (lhs), lhs, op);
5110 vn_reference_insert (assign, lhs, vuse, vdef);
5113 else
5115 /* We had a match, so value number the vdef to have the value
5116 number of the vuse it came from. */
5118 if (dump_file && (dump_flags & TDF_DETAILS))
5119 fprintf (dump_file, "Store matched earlier value, "
5120 "value numbering store vdefs to matching vuses.\n");
5122 changed |= set_ssa_val_to (vdef, SSA_VAL (vuse));
5125 return changed;
5128 /* Visit and value number PHI, return true if the value number
5129 changed. When BACKEDGES_VARYING_P is true then assume all
5130 backedge values are varying. When INSERTED is not NULL then
5131 this is just a ahead query for a possible iteration, set INSERTED
5132 to true if we'd insert into the hashtable. */
5134 static bool
5135 visit_phi (gimple *phi, bool *inserted, bool backedges_varying_p)
5137 tree result, sameval = VN_TOP, seen_undef = NULL_TREE;
5138 tree backedge_val = NULL_TREE;
5139 bool seen_non_backedge = false;
5140 tree sameval_base = NULL_TREE;
5141 poly_int64 soff, doff;
5142 unsigned n_executable = 0;
5143 edge_iterator ei;
5144 edge e;
5146 /* TODO: We could check for this in initialization, and replace this
5147 with a gcc_assert. */
5148 if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (PHI_RESULT (phi)))
5149 return set_ssa_val_to (PHI_RESULT (phi), PHI_RESULT (phi));
5151 /* We track whether a PHI was CSEd to to avoid excessive iterations
5152 that would be necessary only because the PHI changed arguments
5153 but not value. */
5154 if (!inserted)
5155 gimple_set_plf (phi, GF_PLF_1, false);
5157 /* See if all non-TOP arguments have the same value. TOP is
5158 equivalent to everything, so we can ignore it. */
5159 FOR_EACH_EDGE (e, ei, gimple_bb (phi)->preds)
5160 if (e->flags & EDGE_EXECUTABLE)
5162 tree def = PHI_ARG_DEF_FROM_EDGE (phi, e);
5164 ++n_executable;
5165 if (TREE_CODE (def) == SSA_NAME)
5167 if (!backedges_varying_p || !(e->flags & EDGE_DFS_BACK))
5168 def = SSA_VAL (def);
5169 if (e->flags & EDGE_DFS_BACK)
5170 backedge_val = def;
5172 if (!(e->flags & EDGE_DFS_BACK))
5173 seen_non_backedge = true;
5174 if (def == VN_TOP)
5176 /* Ignore undefined defs for sameval but record one. */
5177 else if (TREE_CODE (def) == SSA_NAME
5178 && ! virtual_operand_p (def)
5179 && ssa_undefined_value_p (def, false))
5180 seen_undef = def;
5181 else if (sameval == VN_TOP)
5182 sameval = def;
5183 else if (!expressions_equal_p (def, sameval))
5185 /* We know we're arriving only with invariant addresses here,
5186 try harder comparing them. We can do some caching here
5187 which we cannot do in expressions_equal_p. */
5188 if (TREE_CODE (def) == ADDR_EXPR
5189 && TREE_CODE (sameval) == ADDR_EXPR
5190 && sameval_base != (void *)-1)
5192 if (!sameval_base)
5193 sameval_base = get_addr_base_and_unit_offset
5194 (TREE_OPERAND (sameval, 0), &soff);
5195 if (!sameval_base)
5196 sameval_base = (tree)(void *)-1;
5197 else if ((get_addr_base_and_unit_offset
5198 (TREE_OPERAND (def, 0), &doff) == sameval_base)
5199 && known_eq (soff, doff))
5200 continue;
5202 sameval = NULL_TREE;
5203 break;
5207 /* If the value we want to use is flowing over the backedge and we
5208 should take it as VARYING but it has a non-VARYING value drop to
5209 VARYING.
5210 If we value-number a virtual operand never value-number to the
5211 value from the backedge as that confuses the alias-walking code.
5212 See gcc.dg/torture/pr87176.c. If the value is the same on a
5213 non-backedge everything is OK though. */
5214 bool visited_p;
5215 if ((backedge_val
5216 && !seen_non_backedge
5217 && TREE_CODE (backedge_val) == SSA_NAME
5218 && sameval == backedge_val
5219 && (SSA_NAME_IS_VIRTUAL_OPERAND (backedge_val)
5220 || SSA_VAL (backedge_val) != backedge_val))
5221 /* Do not value-number a virtual operand to sth not visited though
5222 given that allows us to escape a region in alias walking. */
5223 || (sameval
5224 && TREE_CODE (sameval) == SSA_NAME
5225 && !SSA_NAME_IS_DEFAULT_DEF (sameval)
5226 && SSA_NAME_IS_VIRTUAL_OPERAND (sameval)
5227 && (SSA_VAL (sameval, &visited_p), !visited_p)))
5228 /* Note this just drops to VARYING without inserting the PHI into
5229 the hashes. */
5230 result = PHI_RESULT (phi);
5231 /* If none of the edges was executable keep the value-number at VN_TOP,
5232 if only a single edge is exectuable use its value. */
5233 else if (n_executable <= 1)
5234 result = seen_undef ? seen_undef : sameval;
5235 /* If we saw only undefined values and VN_TOP use one of the
5236 undefined values. */
5237 else if (sameval == VN_TOP)
5238 result = seen_undef ? seen_undef : sameval;
5239 /* First see if it is equivalent to a phi node in this block. We prefer
5240 this as it allows IV elimination - see PRs 66502 and 67167. */
5241 else if ((result = vn_phi_lookup (phi, backedges_varying_p)))
5243 if (!inserted
5244 && TREE_CODE (result) == SSA_NAME
5245 && gimple_code (SSA_NAME_DEF_STMT (result)) == GIMPLE_PHI)
5247 gimple_set_plf (SSA_NAME_DEF_STMT (result), GF_PLF_1, true);
5248 if (dump_file && (dump_flags & TDF_DETAILS))
5250 fprintf (dump_file, "Marking CSEd to PHI node ");
5251 print_gimple_expr (dump_file, SSA_NAME_DEF_STMT (result),
5252 0, TDF_SLIM);
5253 fprintf (dump_file, "\n");
5257 /* If all values are the same use that, unless we've seen undefined
5258 values as well and the value isn't constant.
5259 CCP/copyprop have the same restriction to not remove uninit warnings. */
5260 else if (sameval
5261 && (! seen_undef || is_gimple_min_invariant (sameval)))
5262 result = sameval;
5263 else
5265 result = PHI_RESULT (phi);
5266 /* Only insert PHIs that are varying, for constant value numbers
5267 we mess up equivalences otherwise as we are only comparing
5268 the immediate controlling predicates. */
5269 vn_phi_insert (phi, result, backedges_varying_p);
5270 if (inserted)
5271 *inserted = true;
5274 return set_ssa_val_to (PHI_RESULT (phi), result);
5277 /* Try to simplify RHS using equivalences and constant folding. */
5279 static tree
5280 try_to_simplify (gassign *stmt)
5282 enum tree_code code = gimple_assign_rhs_code (stmt);
5283 tree tem;
5285 /* For stores we can end up simplifying a SSA_NAME rhs. Just return
5286 in this case, there is no point in doing extra work. */
5287 if (code == SSA_NAME)
5288 return NULL_TREE;
5290 /* First try constant folding based on our current lattice. */
5291 mprts_hook = vn_lookup_simplify_result;
5292 tem = gimple_fold_stmt_to_constant_1 (stmt, vn_valueize, vn_valueize);
5293 mprts_hook = NULL;
5294 if (tem
5295 && (TREE_CODE (tem) == SSA_NAME
5296 || is_gimple_min_invariant (tem)))
5297 return tem;
5299 return NULL_TREE;
5302 /* Visit and value number STMT, return true if the value number
5303 changed. */
5305 static bool
5306 visit_stmt (gimple *stmt, bool backedges_varying_p = false)
5308 bool changed = false;
5310 if (dump_file && (dump_flags & TDF_DETAILS))
5312 fprintf (dump_file, "Value numbering stmt = ");
5313 print_gimple_stmt (dump_file, stmt, 0);
5316 if (gimple_code (stmt) == GIMPLE_PHI)
5317 changed = visit_phi (stmt, NULL, backedges_varying_p);
5318 else if (gimple_has_volatile_ops (stmt))
5319 changed = defs_to_varying (stmt);
5320 else if (gassign *ass = dyn_cast <gassign *> (stmt))
5322 enum tree_code code = gimple_assign_rhs_code (ass);
5323 tree lhs = gimple_assign_lhs (ass);
5324 tree rhs1 = gimple_assign_rhs1 (ass);
5325 tree simplified;
5327 /* Shortcut for copies. Simplifying copies is pointless,
5328 since we copy the expression and value they represent. */
5329 if (code == SSA_NAME
5330 && TREE_CODE (lhs) == SSA_NAME)
5332 changed = visit_copy (lhs, rhs1);
5333 goto done;
5335 simplified = try_to_simplify (ass);
5336 if (simplified)
5338 if (dump_file && (dump_flags & TDF_DETAILS))
5340 fprintf (dump_file, "RHS ");
5341 print_gimple_expr (dump_file, ass, 0);
5342 fprintf (dump_file, " simplified to ");
5343 print_generic_expr (dump_file, simplified);
5344 fprintf (dump_file, "\n");
5347 /* Setting value numbers to constants will occasionally
5348 screw up phi congruence because constants are not
5349 uniquely associated with a single ssa name that can be
5350 looked up. */
5351 if (simplified
5352 && is_gimple_min_invariant (simplified)
5353 && TREE_CODE (lhs) == SSA_NAME)
5355 changed = set_ssa_val_to (lhs, simplified);
5356 goto done;
5358 else if (simplified
5359 && TREE_CODE (simplified) == SSA_NAME
5360 && TREE_CODE (lhs) == SSA_NAME)
5362 changed = visit_copy (lhs, simplified);
5363 goto done;
5366 if ((TREE_CODE (lhs) == SSA_NAME
5367 /* We can substitute SSA_NAMEs that are live over
5368 abnormal edges with their constant value. */
5369 && !(gimple_assign_copy_p (ass)
5370 && is_gimple_min_invariant (rhs1))
5371 && !(simplified
5372 && is_gimple_min_invariant (simplified))
5373 && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs))
5374 /* Stores or copies from SSA_NAMEs that are live over
5375 abnormal edges are a problem. */
5376 || (code == SSA_NAME
5377 && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (rhs1)))
5378 changed = defs_to_varying (ass);
5379 else if (REFERENCE_CLASS_P (lhs)
5380 || DECL_P (lhs))
5381 changed = visit_reference_op_store (lhs, rhs1, ass);
5382 else if (TREE_CODE (lhs) == SSA_NAME)
5384 if ((gimple_assign_copy_p (ass)
5385 && is_gimple_min_invariant (rhs1))
5386 || (simplified
5387 && is_gimple_min_invariant (simplified)))
5389 if (simplified)
5390 changed = set_ssa_val_to (lhs, simplified);
5391 else
5392 changed = set_ssa_val_to (lhs, rhs1);
5394 else
5396 /* Visit the original statement. */
5397 switch (vn_get_stmt_kind (ass))
5399 case VN_NARY:
5400 changed = visit_nary_op (lhs, ass);
5401 break;
5402 case VN_REFERENCE:
5403 changed = visit_reference_op_load (lhs, rhs1, ass);
5404 break;
5405 default:
5406 changed = defs_to_varying (ass);
5407 break;
5411 else
5412 changed = defs_to_varying (ass);
5414 else if (gcall *call_stmt = dyn_cast <gcall *> (stmt))
5416 tree lhs = gimple_call_lhs (call_stmt);
5417 if (lhs && TREE_CODE (lhs) == SSA_NAME)
5419 /* Try constant folding based on our current lattice. */
5420 tree simplified = gimple_fold_stmt_to_constant_1 (call_stmt,
5421 vn_valueize);
5422 if (simplified)
5424 if (dump_file && (dump_flags & TDF_DETAILS))
5426 fprintf (dump_file, "call ");
5427 print_gimple_expr (dump_file, call_stmt, 0);
5428 fprintf (dump_file, " simplified to ");
5429 print_generic_expr (dump_file, simplified);
5430 fprintf (dump_file, "\n");
5433 /* Setting value numbers to constants will occasionally
5434 screw up phi congruence because constants are not
5435 uniquely associated with a single ssa name that can be
5436 looked up. */
5437 if (simplified
5438 && is_gimple_min_invariant (simplified))
5440 changed = set_ssa_val_to (lhs, simplified);
5441 if (gimple_vdef (call_stmt))
5442 changed |= set_ssa_val_to (gimple_vdef (call_stmt),
5443 SSA_VAL (gimple_vuse (call_stmt)));
5444 goto done;
5446 else if (simplified
5447 && TREE_CODE (simplified) == SSA_NAME)
5449 changed = visit_copy (lhs, simplified);
5450 if (gimple_vdef (call_stmt))
5451 changed |= set_ssa_val_to (gimple_vdef (call_stmt),
5452 SSA_VAL (gimple_vuse (call_stmt)));
5453 goto done;
5455 else if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs))
5457 changed = defs_to_varying (call_stmt);
5458 goto done;
5462 /* Pick up flags from a devirtualization target. */
5463 tree fn = gimple_call_fn (stmt);
5464 int extra_fnflags = 0;
5465 if (fn && TREE_CODE (fn) == SSA_NAME)
5467 fn = SSA_VAL (fn);
5468 if (TREE_CODE (fn) == ADDR_EXPR
5469 && TREE_CODE (TREE_OPERAND (fn, 0)) == FUNCTION_DECL)
5470 extra_fnflags = flags_from_decl_or_type (TREE_OPERAND (fn, 0));
5472 if (!gimple_call_internal_p (call_stmt)
5473 && (/* Calls to the same function with the same vuse
5474 and the same operands do not necessarily return the same
5475 value, unless they're pure or const. */
5476 ((gimple_call_flags (call_stmt) | extra_fnflags)
5477 & (ECF_PURE | ECF_CONST))
5478 /* If calls have a vdef, subsequent calls won't have
5479 the same incoming vuse. So, if 2 calls with vdef have the
5480 same vuse, we know they're not subsequent.
5481 We can value number 2 calls to the same function with the
5482 same vuse and the same operands which are not subsequent
5483 the same, because there is no code in the program that can
5484 compare the 2 values... */
5485 || (gimple_vdef (call_stmt)
5486 /* ... unless the call returns a pointer which does
5487 not alias with anything else. In which case the
5488 information that the values are distinct are encoded
5489 in the IL. */
5490 && !(gimple_call_return_flags (call_stmt) & ERF_NOALIAS)
5491 /* Only perform the following when being called from PRE
5492 which embeds tail merging. */
5493 && default_vn_walk_kind == VN_WALK)))
5494 changed = visit_reference_op_call (lhs, call_stmt);
5495 else
5496 changed = defs_to_varying (call_stmt);
5498 else
5499 changed = defs_to_varying (stmt);
5500 done:
5501 return changed;
5505 /* Allocate a value number table. */
5507 static void
5508 allocate_vn_table (vn_tables_t table, unsigned size)
5510 table->phis = new vn_phi_table_type (size);
5511 table->nary = new vn_nary_op_table_type (size);
5512 table->references = new vn_reference_table_type (size);
5515 /* Free a value number table. */
5517 static void
5518 free_vn_table (vn_tables_t table)
5520 /* Walk over elements and release vectors. */
5521 vn_reference_iterator_type hir;
5522 vn_reference_t vr;
5523 FOR_EACH_HASH_TABLE_ELEMENT (*table->references, vr, vn_reference_t, hir)
5524 vr->operands.release ();
5525 delete table->phis;
5526 table->phis = NULL;
5527 delete table->nary;
5528 table->nary = NULL;
5529 delete table->references;
5530 table->references = NULL;
5533 /* Set *ID according to RESULT. */
5535 static void
5536 set_value_id_for_result (tree result, unsigned int *id)
5538 if (result && TREE_CODE (result) == SSA_NAME)
5539 *id = VN_INFO (result)->value_id;
5540 else if (result && is_gimple_min_invariant (result))
5541 *id = get_or_alloc_constant_value_id (result);
5542 else
5543 *id = get_next_value_id ();
5546 /* Set the value ids in the valid hash tables. */
5548 static void
5549 set_hashtable_value_ids (void)
5551 vn_nary_op_iterator_type hin;
5552 vn_phi_iterator_type hip;
5553 vn_reference_iterator_type hir;
5554 vn_nary_op_t vno;
5555 vn_reference_t vr;
5556 vn_phi_t vp;
5558 /* Now set the value ids of the things we had put in the hash
5559 table. */
5561 FOR_EACH_HASH_TABLE_ELEMENT (*valid_info->nary, vno, vn_nary_op_t, hin)
5562 if (! vno->predicated_values)
5563 set_value_id_for_result (vno->u.result, &vno->value_id);
5565 FOR_EACH_HASH_TABLE_ELEMENT (*valid_info->phis, vp, vn_phi_t, hip)
5566 set_value_id_for_result (vp->result, &vp->value_id);
5568 FOR_EACH_HASH_TABLE_ELEMENT (*valid_info->references, vr, vn_reference_t,
5569 hir)
5570 set_value_id_for_result (vr->result, &vr->value_id);
5573 /* Return the maximum value id we have ever seen. */
5575 unsigned int
5576 get_max_value_id (void)
5578 return next_value_id;
5581 /* Return the next unique value id. */
5583 unsigned int
5584 get_next_value_id (void)
5586 return next_value_id++;
5590 /* Compare two expressions E1 and E2 and return true if they are equal. */
5592 bool
5593 expressions_equal_p (tree e1, tree e2)
5595 /* The obvious case. */
5596 if (e1 == e2)
5597 return true;
5599 /* If either one is VN_TOP consider them equal. */
5600 if (e1 == VN_TOP || e2 == VN_TOP)
5601 return true;
5603 /* If only one of them is null, they cannot be equal. */
5604 if (!e1 || !e2)
5605 return false;
5607 /* Now perform the actual comparison. */
5608 if (TREE_CODE (e1) == TREE_CODE (e2)
5609 && operand_equal_p (e1, e2, OEP_PURE_SAME))
5610 return true;
5612 return false;
5616 /* Return true if the nary operation NARY may trap. This is a copy
5617 of stmt_could_throw_1_p adjusted to the SCCVN IL. */
5619 bool
5620 vn_nary_may_trap (vn_nary_op_t nary)
5622 tree type;
5623 tree rhs2 = NULL_TREE;
5624 bool honor_nans = false;
5625 bool honor_snans = false;
5626 bool fp_operation = false;
5627 bool honor_trapv = false;
5628 bool handled, ret;
5629 unsigned i;
5631 if (TREE_CODE_CLASS (nary->opcode) == tcc_comparison
5632 || TREE_CODE_CLASS (nary->opcode) == tcc_unary
5633 || TREE_CODE_CLASS (nary->opcode) == tcc_binary)
5635 type = nary->type;
5636 fp_operation = FLOAT_TYPE_P (type);
5637 if (fp_operation)
5639 honor_nans = flag_trapping_math && !flag_finite_math_only;
5640 honor_snans = flag_signaling_nans != 0;
5642 else if (INTEGRAL_TYPE_P (type) && TYPE_OVERFLOW_TRAPS (type))
5643 honor_trapv = true;
5645 if (nary->length >= 2)
5646 rhs2 = nary->op[1];
5647 ret = operation_could_trap_helper_p (nary->opcode, fp_operation,
5648 honor_trapv, honor_nans, honor_snans,
5649 rhs2, &handled);
5650 if (handled && ret)
5651 return true;
5653 for (i = 0; i < nary->length; ++i)
5654 if (tree_could_trap_p (nary->op[i]))
5655 return true;
5657 return false;
5660 /* Return true if the reference operation REF may trap. */
5662 bool
5663 vn_reference_may_trap (vn_reference_t ref)
5665 switch (ref->operands[0].opcode)
5667 case MODIFY_EXPR:
5668 case CALL_EXPR:
5669 /* We do not handle calls. */
5670 case ADDR_EXPR:
5671 /* And toplevel address computations never trap. */
5672 return false;
5673 default:;
5676 vn_reference_op_t op;
5677 unsigned i;
5678 FOR_EACH_VEC_ELT (ref->operands, i, op)
5680 switch (op->opcode)
5682 case WITH_SIZE_EXPR:
5683 case TARGET_MEM_REF:
5684 /* Always variable. */
5685 return true;
5686 case COMPONENT_REF:
5687 if (op->op1 && TREE_CODE (op->op1) == SSA_NAME)
5688 return true;
5689 break;
5690 case ARRAY_RANGE_REF:
5691 case ARRAY_REF:
5692 if (TREE_CODE (op->op0) == SSA_NAME)
5693 return true;
5694 break;
5695 case MEM_REF:
5696 /* Nothing interesting in itself, the base is separate. */
5697 break;
5698 /* The following are the address bases. */
5699 case SSA_NAME:
5700 return true;
5701 case ADDR_EXPR:
5702 if (op->op0)
5703 return tree_could_trap_p (TREE_OPERAND (op->op0, 0));
5704 return false;
5705 default:;
5708 return false;
5711 eliminate_dom_walker::eliminate_dom_walker (cdi_direction direction,
5712 bitmap inserted_exprs_)
5713 : dom_walker (direction), do_pre (inserted_exprs_ != NULL),
5714 el_todo (0), eliminations (0), insertions (0),
5715 inserted_exprs (inserted_exprs_)
5717 need_eh_cleanup = BITMAP_ALLOC (NULL);
5718 need_ab_cleanup = BITMAP_ALLOC (NULL);
5721 eliminate_dom_walker::~eliminate_dom_walker ()
5723 BITMAP_FREE (need_eh_cleanup);
5724 BITMAP_FREE (need_ab_cleanup);
5727 /* Return a leader for OP that is available at the current point of the
5728 eliminate domwalk. */
5730 tree
5731 eliminate_dom_walker::eliminate_avail (basic_block, tree op)
5733 tree valnum = VN_INFO (op)->valnum;
5734 if (TREE_CODE (valnum) == SSA_NAME)
5736 if (SSA_NAME_IS_DEFAULT_DEF (valnum))
5737 return valnum;
5738 if (avail.length () > SSA_NAME_VERSION (valnum))
5739 return avail[SSA_NAME_VERSION (valnum)];
5741 else if (is_gimple_min_invariant (valnum))
5742 return valnum;
5743 return NULL_TREE;
5746 /* At the current point of the eliminate domwalk make OP available. */
5748 void
5749 eliminate_dom_walker::eliminate_push_avail (basic_block, tree op)
5751 tree valnum = VN_INFO (op)->valnum;
5752 if (TREE_CODE (valnum) == SSA_NAME)
5754 if (avail.length () <= SSA_NAME_VERSION (valnum))
5755 avail.safe_grow_cleared (SSA_NAME_VERSION (valnum) + 1, true);
5756 tree pushop = op;
5757 if (avail[SSA_NAME_VERSION (valnum)])
5758 pushop = avail[SSA_NAME_VERSION (valnum)];
5759 avail_stack.safe_push (pushop);
5760 avail[SSA_NAME_VERSION (valnum)] = op;
5764 /* Insert the expression recorded by SCCVN for VAL at *GSI. Returns
5765 the leader for the expression if insertion was successful. */
5767 tree
5768 eliminate_dom_walker::eliminate_insert (basic_block bb,
5769 gimple_stmt_iterator *gsi, tree val)
5771 /* We can insert a sequence with a single assignment only. */
5772 gimple_seq stmts = VN_INFO (val)->expr;
5773 if (!gimple_seq_singleton_p (stmts))
5774 return NULL_TREE;
5775 gassign *stmt = dyn_cast <gassign *> (gimple_seq_first_stmt (stmts));
5776 if (!stmt
5777 || (!CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt))
5778 && gimple_assign_rhs_code (stmt) != VIEW_CONVERT_EXPR
5779 && gimple_assign_rhs_code (stmt) != NEGATE_EXPR
5780 && gimple_assign_rhs_code (stmt) != BIT_FIELD_REF
5781 && (gimple_assign_rhs_code (stmt) != BIT_AND_EXPR
5782 || TREE_CODE (gimple_assign_rhs2 (stmt)) != INTEGER_CST)))
5783 return NULL_TREE;
5785 tree op = gimple_assign_rhs1 (stmt);
5786 if (gimple_assign_rhs_code (stmt) == VIEW_CONVERT_EXPR
5787 || gimple_assign_rhs_code (stmt) == BIT_FIELD_REF)
5788 op = TREE_OPERAND (op, 0);
5789 tree leader = TREE_CODE (op) == SSA_NAME ? eliminate_avail (bb, op) : op;
5790 if (!leader)
5791 return NULL_TREE;
5793 tree res;
5794 stmts = NULL;
5795 if (gimple_assign_rhs_code (stmt) == BIT_FIELD_REF)
5796 res = gimple_build (&stmts, BIT_FIELD_REF,
5797 TREE_TYPE (val), leader,
5798 TREE_OPERAND (gimple_assign_rhs1 (stmt), 1),
5799 TREE_OPERAND (gimple_assign_rhs1 (stmt), 2));
5800 else if (gimple_assign_rhs_code (stmt) == BIT_AND_EXPR)
5801 res = gimple_build (&stmts, BIT_AND_EXPR,
5802 TREE_TYPE (val), leader, gimple_assign_rhs2 (stmt));
5803 else
5804 res = gimple_build (&stmts, gimple_assign_rhs_code (stmt),
5805 TREE_TYPE (val), leader);
5806 if (TREE_CODE (res) != SSA_NAME
5807 || SSA_NAME_IS_DEFAULT_DEF (res)
5808 || gimple_bb (SSA_NAME_DEF_STMT (res)))
5810 gimple_seq_discard (stmts);
5812 /* During propagation we have to treat SSA info conservatively
5813 and thus we can end up simplifying the inserted expression
5814 at elimination time to sth not defined in stmts. */
5815 /* But then this is a redundancy we failed to detect. Which means
5816 res now has two values. That doesn't play well with how
5817 we track availability here, so give up. */
5818 if (dump_file && (dump_flags & TDF_DETAILS))
5820 if (TREE_CODE (res) == SSA_NAME)
5821 res = eliminate_avail (bb, res);
5822 if (res)
5824 fprintf (dump_file, "Failed to insert expression for value ");
5825 print_generic_expr (dump_file, val);
5826 fprintf (dump_file, " which is really fully redundant to ");
5827 print_generic_expr (dump_file, res);
5828 fprintf (dump_file, "\n");
5832 return NULL_TREE;
5834 else
5836 gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);
5837 VN_INFO (res)->valnum = val;
5838 VN_INFO (res)->visited = true;
5841 insertions++;
5842 if (dump_file && (dump_flags & TDF_DETAILS))
5844 fprintf (dump_file, "Inserted ");
5845 print_gimple_stmt (dump_file, SSA_NAME_DEF_STMT (res), 0);
5848 return res;
5851 void
5852 eliminate_dom_walker::eliminate_stmt (basic_block b, gimple_stmt_iterator *gsi)
5854 tree sprime = NULL_TREE;
5855 gimple *stmt = gsi_stmt (*gsi);
5856 tree lhs = gimple_get_lhs (stmt);
5857 if (lhs && TREE_CODE (lhs) == SSA_NAME
5858 && !gimple_has_volatile_ops (stmt)
5859 /* See PR43491. Do not replace a global register variable when
5860 it is a the RHS of an assignment. Do replace local register
5861 variables since gcc does not guarantee a local variable will
5862 be allocated in register.
5863 ??? The fix isn't effective here. This should instead
5864 be ensured by not value-numbering them the same but treating
5865 them like volatiles? */
5866 && !(gimple_assign_single_p (stmt)
5867 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == VAR_DECL
5868 && DECL_HARD_REGISTER (gimple_assign_rhs1 (stmt))
5869 && is_global_var (gimple_assign_rhs1 (stmt)))))
5871 sprime = eliminate_avail (b, lhs);
5872 if (!sprime)
5874 /* If there is no existing usable leader but SCCVN thinks
5875 it has an expression it wants to use as replacement,
5876 insert that. */
5877 tree val = VN_INFO (lhs)->valnum;
5878 if (val != VN_TOP
5879 && TREE_CODE (val) == SSA_NAME
5880 && VN_INFO (val)->needs_insertion
5881 && VN_INFO (val)->expr != NULL
5882 && (sprime = eliminate_insert (b, gsi, val)) != NULL_TREE)
5883 eliminate_push_avail (b, sprime);
5886 /* If this now constitutes a copy duplicate points-to
5887 and range info appropriately. This is especially
5888 important for inserted code. See tree-ssa-copy.c
5889 for similar code. */
5890 if (sprime
5891 && TREE_CODE (sprime) == SSA_NAME)
5893 basic_block sprime_b = gimple_bb (SSA_NAME_DEF_STMT (sprime));
5894 if (POINTER_TYPE_P (TREE_TYPE (lhs))
5895 && SSA_NAME_PTR_INFO (lhs)
5896 && ! SSA_NAME_PTR_INFO (sprime))
5898 duplicate_ssa_name_ptr_info (sprime,
5899 SSA_NAME_PTR_INFO (lhs));
5900 if (b != sprime_b)
5901 reset_flow_sensitive_info (sprime);
5903 else if (INTEGRAL_TYPE_P (TREE_TYPE (lhs))
5904 && SSA_NAME_RANGE_INFO (lhs)
5905 && ! SSA_NAME_RANGE_INFO (sprime)
5906 && b == sprime_b)
5907 duplicate_ssa_name_range_info (sprime,
5908 SSA_NAME_RANGE_TYPE (lhs),
5909 SSA_NAME_RANGE_INFO (lhs));
5912 /* Inhibit the use of an inserted PHI on a loop header when
5913 the address of the memory reference is a simple induction
5914 variable. In other cases the vectorizer won't do anything
5915 anyway (either it's loop invariant or a complicated
5916 expression). */
5917 if (sprime
5918 && TREE_CODE (sprime) == SSA_NAME
5919 && do_pre
5920 && (flag_tree_loop_vectorize || flag_tree_parallelize_loops > 1)
5921 && loop_outer (b->loop_father)
5922 && has_zero_uses (sprime)
5923 && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (sprime))
5924 && gimple_assign_load_p (stmt))
5926 gimple *def_stmt = SSA_NAME_DEF_STMT (sprime);
5927 basic_block def_bb = gimple_bb (def_stmt);
5928 if (gimple_code (def_stmt) == GIMPLE_PHI
5929 && def_bb->loop_father->header == def_bb)
5931 loop_p loop = def_bb->loop_father;
5932 ssa_op_iter iter;
5933 tree op;
5934 bool found = false;
5935 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
5937 affine_iv iv;
5938 def_bb = gimple_bb (SSA_NAME_DEF_STMT (op));
5939 if (def_bb
5940 && flow_bb_inside_loop_p (loop, def_bb)
5941 && simple_iv (loop, loop, op, &iv, true))
5943 found = true;
5944 break;
5947 if (found)
5949 if (dump_file && (dump_flags & TDF_DETAILS))
5951 fprintf (dump_file, "Not replacing ");
5952 print_gimple_expr (dump_file, stmt, 0);
5953 fprintf (dump_file, " with ");
5954 print_generic_expr (dump_file, sprime);
5955 fprintf (dump_file, " which would add a loop"
5956 " carried dependence to loop %d\n",
5957 loop->num);
5959 /* Don't keep sprime available. */
5960 sprime = NULL_TREE;
5965 if (sprime)
5967 /* If we can propagate the value computed for LHS into
5968 all uses don't bother doing anything with this stmt. */
5969 if (may_propagate_copy (lhs, sprime))
5971 /* Mark it for removal. */
5972 to_remove.safe_push (stmt);
5974 /* ??? Don't count copy/constant propagations. */
5975 if (gimple_assign_single_p (stmt)
5976 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
5977 || gimple_assign_rhs1 (stmt) == sprime))
5978 return;
5980 if (dump_file && (dump_flags & TDF_DETAILS))
5982 fprintf (dump_file, "Replaced ");
5983 print_gimple_expr (dump_file, stmt, 0);
5984 fprintf (dump_file, " with ");
5985 print_generic_expr (dump_file, sprime);
5986 fprintf (dump_file, " in all uses of ");
5987 print_gimple_stmt (dump_file, stmt, 0);
5990 eliminations++;
5991 return;
5994 /* If this is an assignment from our leader (which
5995 happens in the case the value-number is a constant)
5996 then there is nothing to do. Likewise if we run into
5997 inserted code that needed a conversion because of
5998 our type-agnostic value-numbering of loads. */
5999 if ((gimple_assign_single_p (stmt)
6000 || (is_gimple_assign (stmt)
6001 && (CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt))
6002 || gimple_assign_rhs_code (stmt) == VIEW_CONVERT_EXPR)))
6003 && sprime == gimple_assign_rhs1 (stmt))
6004 return;
6006 /* Else replace its RHS. */
6007 if (dump_file && (dump_flags & TDF_DETAILS))
6009 fprintf (dump_file, "Replaced ");
6010 print_gimple_expr (dump_file, stmt, 0);
6011 fprintf (dump_file, " with ");
6012 print_generic_expr (dump_file, sprime);
6013 fprintf (dump_file, " in ");
6014 print_gimple_stmt (dump_file, stmt, 0);
6016 eliminations++;
6018 bool can_make_abnormal_goto = (is_gimple_call (stmt)
6019 && stmt_can_make_abnormal_goto (stmt));
6020 gimple *orig_stmt = stmt;
6021 if (!useless_type_conversion_p (TREE_TYPE (lhs),
6022 TREE_TYPE (sprime)))
6024 /* We preserve conversions to but not from function or method
6025 types. This asymmetry makes it necessary to re-instantiate
6026 conversions here. */
6027 if (POINTER_TYPE_P (TREE_TYPE (lhs))
6028 && FUNC_OR_METHOD_TYPE_P (TREE_TYPE (TREE_TYPE (lhs))))
6029 sprime = fold_convert (TREE_TYPE (lhs), sprime);
6030 else
6031 gcc_unreachable ();
6033 tree vdef = gimple_vdef (stmt);
6034 tree vuse = gimple_vuse (stmt);
6035 propagate_tree_value_into_stmt (gsi, sprime);
6036 stmt = gsi_stmt (*gsi);
6037 update_stmt (stmt);
6038 /* In case the VDEF on the original stmt was released, value-number
6039 it to the VUSE. This is to make vuse_ssa_val able to skip
6040 released virtual operands. */
6041 if (vdef != gimple_vdef (stmt))
6043 gcc_assert (SSA_NAME_IN_FREE_LIST (vdef));
6044 VN_INFO (vdef)->valnum = vuse;
6047 /* If we removed EH side-effects from the statement, clean
6048 its EH information. */
6049 if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
6051 bitmap_set_bit (need_eh_cleanup,
6052 gimple_bb (stmt)->index);
6053 if (dump_file && (dump_flags & TDF_DETAILS))
6054 fprintf (dump_file, " Removed EH side-effects.\n");
6057 /* Likewise for AB side-effects. */
6058 if (can_make_abnormal_goto
6059 && !stmt_can_make_abnormal_goto (stmt))
6061 bitmap_set_bit (need_ab_cleanup,
6062 gimple_bb (stmt)->index);
6063 if (dump_file && (dump_flags & TDF_DETAILS))
6064 fprintf (dump_file, " Removed AB side-effects.\n");
6067 return;
6071 /* If the statement is a scalar store, see if the expression
6072 has the same value number as its rhs. If so, the store is
6073 dead. */
6074 if (gimple_assign_single_p (stmt)
6075 && !gimple_has_volatile_ops (stmt)
6076 && !is_gimple_reg (gimple_assign_lhs (stmt))
6077 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
6078 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
6080 tree rhs = gimple_assign_rhs1 (stmt);
6081 vn_reference_t vnresult;
6082 /* ??? gcc.dg/torture/pr91445.c shows that we lookup a boolean
6083 typed load of a byte known to be 0x11 as 1 so a store of
6084 a boolean 1 is detected as redundant. Because of this we
6085 have to make sure to lookup with a ref where its size
6086 matches the precision. */
6087 tree lookup_lhs = lhs;
6088 if (INTEGRAL_TYPE_P (TREE_TYPE (lhs))
6089 && (TREE_CODE (lhs) != COMPONENT_REF
6090 || !DECL_BIT_FIELD_TYPE (TREE_OPERAND (lhs, 1)))
6091 && !type_has_mode_precision_p (TREE_TYPE (lhs)))
6093 if (TREE_CODE (lhs) == COMPONENT_REF
6094 || TREE_CODE (lhs) == MEM_REF)
6096 tree ltype = build_nonstandard_integer_type
6097 (TREE_INT_CST_LOW (TYPE_SIZE (TREE_TYPE (lhs))),
6098 TYPE_UNSIGNED (TREE_TYPE (lhs)));
6099 if (TREE_CODE (lhs) == COMPONENT_REF)
6101 tree foff = component_ref_field_offset (lhs);
6102 tree f = TREE_OPERAND (lhs, 1);
6103 if (!poly_int_tree_p (foff))
6104 lookup_lhs = NULL_TREE;
6105 else
6106 lookup_lhs = build3 (BIT_FIELD_REF, ltype,
6107 TREE_OPERAND (lhs, 0),
6108 TYPE_SIZE (TREE_TYPE (lhs)),
6109 bit_from_pos
6110 (foff, DECL_FIELD_BIT_OFFSET (f)));
6112 else
6113 lookup_lhs = build2 (MEM_REF, ltype,
6114 TREE_OPERAND (lhs, 0),
6115 TREE_OPERAND (lhs, 1));
6117 else
6118 lookup_lhs = NULL_TREE;
6120 tree val = NULL_TREE;
6121 if (lookup_lhs)
6122 val = vn_reference_lookup (lookup_lhs, gimple_vuse (stmt),
6123 VN_WALKREWRITE, &vnresult, false);
6124 if (TREE_CODE (rhs) == SSA_NAME)
6125 rhs = VN_INFO (rhs)->valnum;
6126 if (val
6127 && (operand_equal_p (val, rhs, 0)
6128 /* Due to the bitfield lookups above we can get bit
6129 interpretations of the same RHS as values here. Those
6130 are redundant as well. */
6131 || (TREE_CODE (val) == SSA_NAME
6132 && gimple_assign_single_p (SSA_NAME_DEF_STMT (val))
6133 && (val = gimple_assign_rhs1 (SSA_NAME_DEF_STMT (val)))
6134 && TREE_CODE (val) == VIEW_CONVERT_EXPR
6135 && TREE_OPERAND (val, 0) == rhs)))
6137 /* We can only remove the later store if the former aliases
6138 at least all accesses the later one does or if the store
6139 was to readonly memory storing the same value. */
6140 ao_ref lhs_ref;
6141 ao_ref_init (&lhs_ref, lhs);
6142 alias_set_type set = ao_ref_alias_set (&lhs_ref);
6143 alias_set_type base_set = ao_ref_base_alias_set (&lhs_ref);
6144 if (! vnresult
6145 || ((vnresult->set == set
6146 || alias_set_subset_of (set, vnresult->set))
6147 && (vnresult->base_set == base_set
6148 || alias_set_subset_of (base_set, vnresult->base_set))))
6150 if (dump_file && (dump_flags & TDF_DETAILS))
6152 fprintf (dump_file, "Deleted redundant store ");
6153 print_gimple_stmt (dump_file, stmt, 0);
6156 /* Queue stmt for removal. */
6157 to_remove.safe_push (stmt);
6158 return;
6163 /* If this is a control statement value numbering left edges
6164 unexecuted on force the condition in a way consistent with
6165 that. */
6166 if (gcond *cond = dyn_cast <gcond *> (stmt))
6168 if ((EDGE_SUCC (b, 0)->flags & EDGE_EXECUTABLE)
6169 ^ (EDGE_SUCC (b, 1)->flags & EDGE_EXECUTABLE))
6171 if (dump_file && (dump_flags & TDF_DETAILS))
6173 fprintf (dump_file, "Removing unexecutable edge from ");
6174 print_gimple_stmt (dump_file, stmt, 0);
6176 if (((EDGE_SUCC (b, 0)->flags & EDGE_TRUE_VALUE) != 0)
6177 == ((EDGE_SUCC (b, 0)->flags & EDGE_EXECUTABLE) != 0))
6178 gimple_cond_make_true (cond);
6179 else
6180 gimple_cond_make_false (cond);
6181 update_stmt (cond);
6182 el_todo |= TODO_cleanup_cfg;
6183 return;
6187 bool can_make_abnormal_goto = stmt_can_make_abnormal_goto (stmt);
6188 bool was_noreturn = (is_gimple_call (stmt)
6189 && gimple_call_noreturn_p (stmt));
6190 tree vdef = gimple_vdef (stmt);
6191 tree vuse = gimple_vuse (stmt);
6193 /* If we didn't replace the whole stmt (or propagate the result
6194 into all uses), replace all uses on this stmt with their
6195 leaders. */
6196 bool modified = false;
6197 use_operand_p use_p;
6198 ssa_op_iter iter;
6199 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE)
6201 tree use = USE_FROM_PTR (use_p);
6202 /* ??? The call code above leaves stmt operands un-updated. */
6203 if (TREE_CODE (use) != SSA_NAME)
6204 continue;
6205 tree sprime;
6206 if (SSA_NAME_IS_DEFAULT_DEF (use))
6207 /* ??? For default defs BB shouldn't matter, but we have to
6208 solve the inconsistency between rpo eliminate and
6209 dom eliminate avail valueization first. */
6210 sprime = eliminate_avail (b, use);
6211 else
6212 /* Look for sth available at the definition block of the argument.
6213 This avoids inconsistencies between availability there which
6214 decides if the stmt can be removed and availability at the
6215 use site. The SSA property ensures that things available
6216 at the definition are also available at uses. */
6217 sprime = eliminate_avail (gimple_bb (SSA_NAME_DEF_STMT (use)), use);
6218 if (sprime && sprime != use
6219 && may_propagate_copy (use, sprime)
6220 /* We substitute into debug stmts to avoid excessive
6221 debug temporaries created by removed stmts, but we need
6222 to avoid doing so for inserted sprimes as we never want
6223 to create debug temporaries for them. */
6224 && (!inserted_exprs
6225 || TREE_CODE (sprime) != SSA_NAME
6226 || !is_gimple_debug (stmt)
6227 || !bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (sprime))))
6229 propagate_value (use_p, sprime);
6230 modified = true;
6234 /* Fold the stmt if modified, this canonicalizes MEM_REFs we propagated
6235 into which is a requirement for the IPA devirt machinery. */
6236 gimple *old_stmt = stmt;
6237 if (modified)
6239 /* If a formerly non-invariant ADDR_EXPR is turned into an
6240 invariant one it was on a separate stmt. */
6241 if (gimple_assign_single_p (stmt)
6242 && TREE_CODE (gimple_assign_rhs1 (stmt)) == ADDR_EXPR)
6243 recompute_tree_invariant_for_addr_expr (gimple_assign_rhs1 (stmt));
6244 gimple_stmt_iterator prev = *gsi;
6245 gsi_prev (&prev);
6246 if (fold_stmt (gsi))
6248 /* fold_stmt may have created new stmts inbetween
6249 the previous stmt and the folded stmt. Mark
6250 all defs created there as varying to not confuse
6251 the SCCVN machinery as we're using that even during
6252 elimination. */
6253 if (gsi_end_p (prev))
6254 prev = gsi_start_bb (b);
6255 else
6256 gsi_next (&prev);
6257 if (gsi_stmt (prev) != gsi_stmt (*gsi))
6260 tree def;
6261 ssa_op_iter dit;
6262 FOR_EACH_SSA_TREE_OPERAND (def, gsi_stmt (prev),
6263 dit, SSA_OP_ALL_DEFS)
6264 /* As existing DEFs may move between stmts
6265 only process new ones. */
6266 if (! has_VN_INFO (def))
6268 VN_INFO (def)->valnum = def;
6269 VN_INFO (def)->visited = true;
6271 if (gsi_stmt (prev) == gsi_stmt (*gsi))
6272 break;
6273 gsi_next (&prev);
6275 while (1);
6277 stmt = gsi_stmt (*gsi);
6278 /* In case we folded the stmt away schedule the NOP for removal. */
6279 if (gimple_nop_p (stmt))
6280 to_remove.safe_push (stmt);
6283 /* Visit indirect calls and turn them into direct calls if
6284 possible using the devirtualization machinery. Do this before
6285 checking for required EH/abnormal/noreturn cleanup as devird
6286 may expose more of those. */
6287 if (gcall *call_stmt = dyn_cast <gcall *> (stmt))
6289 tree fn = gimple_call_fn (call_stmt);
6290 if (fn
6291 && flag_devirtualize
6292 && virtual_method_call_p (fn))
6294 tree otr_type = obj_type_ref_class (fn);
6295 unsigned HOST_WIDE_INT otr_tok
6296 = tree_to_uhwi (OBJ_TYPE_REF_TOKEN (fn));
6297 tree instance;
6298 ipa_polymorphic_call_context context (current_function_decl,
6299 fn, stmt, &instance);
6300 context.get_dynamic_type (instance, OBJ_TYPE_REF_OBJECT (fn),
6301 otr_type, stmt, NULL);
6302 bool final;
6303 vec <cgraph_node *> targets
6304 = possible_polymorphic_call_targets (obj_type_ref_class (fn),
6305 otr_tok, context, &final);
6306 if (dump_file)
6307 dump_possible_polymorphic_call_targets (dump_file,
6308 obj_type_ref_class (fn),
6309 otr_tok, context);
6310 if (final && targets.length () <= 1 && dbg_cnt (devirt))
6312 tree fn;
6313 if (targets.length () == 1)
6314 fn = targets[0]->decl;
6315 else
6316 fn = builtin_decl_implicit (BUILT_IN_UNREACHABLE);
6317 if (dump_enabled_p ())
6319 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, stmt,
6320 "converting indirect call to "
6321 "function %s\n",
6322 lang_hooks.decl_printable_name (fn, 2));
6324 gimple_call_set_fndecl (call_stmt, fn);
6325 /* If changing the call to __builtin_unreachable
6326 or similar noreturn function, adjust gimple_call_fntype
6327 too. */
6328 if (gimple_call_noreturn_p (call_stmt)
6329 && VOID_TYPE_P (TREE_TYPE (TREE_TYPE (fn)))
6330 && TYPE_ARG_TYPES (TREE_TYPE (fn))
6331 && (TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fn)))
6332 == void_type_node))
6333 gimple_call_set_fntype (call_stmt, TREE_TYPE (fn));
6334 maybe_remove_unused_call_args (cfun, call_stmt);
6335 modified = true;
6340 if (modified)
6342 /* When changing a call into a noreturn call, cfg cleanup
6343 is needed to fix up the noreturn call. */
6344 if (!was_noreturn
6345 && is_gimple_call (stmt) && gimple_call_noreturn_p (stmt))
6346 to_fixup.safe_push (stmt);
6347 /* When changing a condition or switch into one we know what
6348 edge will be executed, schedule a cfg cleanup. */
6349 if ((gimple_code (stmt) == GIMPLE_COND
6350 && (gimple_cond_true_p (as_a <gcond *> (stmt))
6351 || gimple_cond_false_p (as_a <gcond *> (stmt))))
6352 || (gimple_code (stmt) == GIMPLE_SWITCH
6353 && TREE_CODE (gimple_switch_index
6354 (as_a <gswitch *> (stmt))) == INTEGER_CST))
6355 el_todo |= TODO_cleanup_cfg;
6356 /* If we removed EH side-effects from the statement, clean
6357 its EH information. */
6358 if (maybe_clean_or_replace_eh_stmt (old_stmt, stmt))
6360 bitmap_set_bit (need_eh_cleanup,
6361 gimple_bb (stmt)->index);
6362 if (dump_file && (dump_flags & TDF_DETAILS))
6363 fprintf (dump_file, " Removed EH side-effects.\n");
6365 /* Likewise for AB side-effects. */
6366 if (can_make_abnormal_goto
6367 && !stmt_can_make_abnormal_goto (stmt))
6369 bitmap_set_bit (need_ab_cleanup,
6370 gimple_bb (stmt)->index);
6371 if (dump_file && (dump_flags & TDF_DETAILS))
6372 fprintf (dump_file, " Removed AB side-effects.\n");
6374 update_stmt (stmt);
6375 /* In case the VDEF on the original stmt was released, value-number
6376 it to the VUSE. This is to make vuse_ssa_val able to skip
6377 released virtual operands. */
6378 if (vdef && SSA_NAME_IN_FREE_LIST (vdef))
6379 VN_INFO (vdef)->valnum = vuse;
6382 /* Make new values available - for fully redundant LHS we
6383 continue with the next stmt above and skip this. */
6384 def_operand_p defp;
6385 FOR_EACH_SSA_DEF_OPERAND (defp, stmt, iter, SSA_OP_DEF)
6386 eliminate_push_avail (b, DEF_FROM_PTR (defp));
6389 /* Perform elimination for the basic-block B during the domwalk. */
6391 edge
6392 eliminate_dom_walker::before_dom_children (basic_block b)
6394 /* Mark new bb. */
6395 avail_stack.safe_push (NULL_TREE);
6397 /* Skip unreachable blocks marked unreachable during the SCCVN domwalk. */
6398 if (!(b->flags & BB_EXECUTABLE))
6399 return NULL;
6401 vn_context_bb = b;
6403 for (gphi_iterator gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
6405 gphi *phi = gsi.phi ();
6406 tree res = PHI_RESULT (phi);
6408 if (virtual_operand_p (res))
6410 gsi_next (&gsi);
6411 continue;
6414 tree sprime = eliminate_avail (b, res);
6415 if (sprime
6416 && sprime != res)
6418 if (dump_file && (dump_flags & TDF_DETAILS))
6420 fprintf (dump_file, "Replaced redundant PHI node defining ");
6421 print_generic_expr (dump_file, res);
6422 fprintf (dump_file, " with ");
6423 print_generic_expr (dump_file, sprime);
6424 fprintf (dump_file, "\n");
6427 /* If we inserted this PHI node ourself, it's not an elimination. */
6428 if (! inserted_exprs
6429 || ! bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
6430 eliminations++;
6432 /* If we will propagate into all uses don't bother to do
6433 anything. */
6434 if (may_propagate_copy (res, sprime))
6436 /* Mark the PHI for removal. */
6437 to_remove.safe_push (phi);
6438 gsi_next (&gsi);
6439 continue;
6442 remove_phi_node (&gsi, false);
6444 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
6445 sprime = fold_convert (TREE_TYPE (res), sprime);
6446 gimple *stmt = gimple_build_assign (res, sprime);
6447 gimple_stmt_iterator gsi2 = gsi_after_labels (b);
6448 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
6449 continue;
6452 eliminate_push_avail (b, res);
6453 gsi_next (&gsi);
6456 for (gimple_stmt_iterator gsi = gsi_start_bb (b);
6457 !gsi_end_p (gsi);
6458 gsi_next (&gsi))
6459 eliminate_stmt (b, &gsi);
6461 /* Replace destination PHI arguments. */
6462 edge_iterator ei;
6463 edge e;
6464 FOR_EACH_EDGE (e, ei, b->succs)
6465 if (e->flags & EDGE_EXECUTABLE)
6466 for (gphi_iterator gsi = gsi_start_phis (e->dest);
6467 !gsi_end_p (gsi);
6468 gsi_next (&gsi))
6470 gphi *phi = gsi.phi ();
6471 use_operand_p use_p = PHI_ARG_DEF_PTR_FROM_EDGE (phi, e);
6472 tree arg = USE_FROM_PTR (use_p);
6473 if (TREE_CODE (arg) != SSA_NAME
6474 || virtual_operand_p (arg))
6475 continue;
6476 tree sprime = eliminate_avail (b, arg);
6477 if (sprime && may_propagate_copy (arg, sprime))
6478 propagate_value (use_p, sprime);
6481 vn_context_bb = NULL;
6483 return NULL;
6486 /* Make no longer available leaders no longer available. */
6488 void
6489 eliminate_dom_walker::after_dom_children (basic_block)
6491 tree entry;
6492 while ((entry = avail_stack.pop ()) != NULL_TREE)
6494 tree valnum = VN_INFO (entry)->valnum;
6495 tree old = avail[SSA_NAME_VERSION (valnum)];
6496 if (old == entry)
6497 avail[SSA_NAME_VERSION (valnum)] = NULL_TREE;
6498 else
6499 avail[SSA_NAME_VERSION (valnum)] = entry;
6503 /* Remove queued stmts and perform delayed cleanups. */
6505 unsigned
6506 eliminate_dom_walker::eliminate_cleanup (bool region_p)
6508 statistics_counter_event (cfun, "Eliminated", eliminations);
6509 statistics_counter_event (cfun, "Insertions", insertions);
6511 /* We cannot remove stmts during BB walk, especially not release SSA
6512 names there as this confuses the VN machinery. The stmts ending
6513 up in to_remove are either stores or simple copies.
6514 Remove stmts in reverse order to make debug stmt creation possible. */
6515 while (!to_remove.is_empty ())
6517 bool do_release_defs = true;
6518 gimple *stmt = to_remove.pop ();
6520 /* When we are value-numbering a region we do not require exit PHIs to
6521 be present so we have to make sure to deal with uses outside of the
6522 region of stmts that we thought are eliminated.
6523 ??? Note we may be confused by uses in dead regions we didn't run
6524 elimination on. Rather than checking individual uses we accept
6525 dead copies to be generated here (gcc.c-torture/execute/20060905-1.c
6526 contains such example). */
6527 if (region_p)
6529 if (gphi *phi = dyn_cast <gphi *> (stmt))
6531 tree lhs = gimple_phi_result (phi);
6532 if (!has_zero_uses (lhs))
6534 if (dump_file && (dump_flags & TDF_DETAILS))
6535 fprintf (dump_file, "Keeping eliminated stmt live "
6536 "as copy because of out-of-region uses\n");
6537 tree sprime = eliminate_avail (gimple_bb (stmt), lhs);
6538 gimple *copy = gimple_build_assign (lhs, sprime);
6539 gimple_stmt_iterator gsi
6540 = gsi_after_labels (gimple_bb (stmt));
6541 gsi_insert_before (&gsi, copy, GSI_SAME_STMT);
6542 do_release_defs = false;
6545 else if (tree lhs = gimple_get_lhs (stmt))
6546 if (TREE_CODE (lhs) == SSA_NAME
6547 && !has_zero_uses (lhs))
6549 if (dump_file && (dump_flags & TDF_DETAILS))
6550 fprintf (dump_file, "Keeping eliminated stmt live "
6551 "as copy because of out-of-region uses\n");
6552 tree sprime = eliminate_avail (gimple_bb (stmt), lhs);
6553 gimple_stmt_iterator gsi = gsi_for_stmt (stmt);
6554 if (is_gimple_assign (stmt))
6556 gimple_assign_set_rhs_from_tree (&gsi, sprime);
6557 stmt = gsi_stmt (gsi);
6558 update_stmt (stmt);
6559 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
6560 bitmap_set_bit (need_eh_cleanup, gimple_bb (stmt)->index);
6561 continue;
6563 else
6565 gimple *copy = gimple_build_assign (lhs, sprime);
6566 gsi_insert_before (&gsi, copy, GSI_SAME_STMT);
6567 do_release_defs = false;
6572 if (dump_file && (dump_flags & TDF_DETAILS))
6574 fprintf (dump_file, "Removing dead stmt ");
6575 print_gimple_stmt (dump_file, stmt, 0, TDF_NONE);
6578 gimple_stmt_iterator gsi = gsi_for_stmt (stmt);
6579 if (gimple_code (stmt) == GIMPLE_PHI)
6580 remove_phi_node (&gsi, do_release_defs);
6581 else
6583 basic_block bb = gimple_bb (stmt);
6584 unlink_stmt_vdef (stmt);
6585 if (gsi_remove (&gsi, true))
6586 bitmap_set_bit (need_eh_cleanup, bb->index);
6587 if (is_gimple_call (stmt) && stmt_can_make_abnormal_goto (stmt))
6588 bitmap_set_bit (need_ab_cleanup, bb->index);
6589 if (do_release_defs)
6590 release_defs (stmt);
6593 /* Removing a stmt may expose a forwarder block. */
6594 el_todo |= TODO_cleanup_cfg;
6597 /* Fixup stmts that became noreturn calls. This may require splitting
6598 blocks and thus isn't possible during the dominator walk. Do this
6599 in reverse order so we don't inadvertedly remove a stmt we want to
6600 fixup by visiting a dominating now noreturn call first. */
6601 while (!to_fixup.is_empty ())
6603 gimple *stmt = to_fixup.pop ();
6605 if (dump_file && (dump_flags & TDF_DETAILS))
6607 fprintf (dump_file, "Fixing up noreturn call ");
6608 print_gimple_stmt (dump_file, stmt, 0);
6611 if (fixup_noreturn_call (stmt))
6612 el_todo |= TODO_cleanup_cfg;
6615 bool do_eh_cleanup = !bitmap_empty_p (need_eh_cleanup);
6616 bool do_ab_cleanup = !bitmap_empty_p (need_ab_cleanup);
6618 if (do_eh_cleanup)
6619 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
6621 if (do_ab_cleanup)
6622 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup);
6624 if (do_eh_cleanup || do_ab_cleanup)
6625 el_todo |= TODO_cleanup_cfg;
6627 return el_todo;
6630 /* Eliminate fully redundant computations. */
6632 unsigned
6633 eliminate_with_rpo_vn (bitmap inserted_exprs)
6635 eliminate_dom_walker walker (CDI_DOMINATORS, inserted_exprs);
6637 eliminate_dom_walker *saved_rpo_avail = rpo_avail;
6638 rpo_avail = &walker;
6639 walker.walk (cfun->cfg->x_entry_block_ptr);
6640 rpo_avail = saved_rpo_avail;
6642 return walker.eliminate_cleanup ();
6645 static unsigned
6646 do_rpo_vn (function *fn, edge entry, bitmap exit_bbs,
6647 bool iterate, bool eliminate);
6649 void
6650 run_rpo_vn (vn_lookup_kind kind)
6652 default_vn_walk_kind = kind;
6653 do_rpo_vn (cfun, NULL, NULL, true, false);
6655 /* ??? Prune requirement of these. */
6656 constant_to_value_id = new hash_table<vn_constant_hasher> (23);
6657 constant_value_ids = BITMAP_ALLOC (NULL);
6659 /* Initialize the value ids and prune out remaining VN_TOPs
6660 from dead code. */
6661 tree name;
6662 unsigned i;
6663 FOR_EACH_SSA_NAME (i, name, cfun)
6665 vn_ssa_aux_t info = VN_INFO (name);
6666 if (!info->visited
6667 || info->valnum == VN_TOP)
6668 info->valnum = name;
6669 if (info->valnum == name)
6670 info->value_id = get_next_value_id ();
6671 else if (is_gimple_min_invariant (info->valnum))
6672 info->value_id = get_or_alloc_constant_value_id (info->valnum);
6675 /* Propagate. */
6676 FOR_EACH_SSA_NAME (i, name, cfun)
6678 vn_ssa_aux_t info = VN_INFO (name);
6679 if (TREE_CODE (info->valnum) == SSA_NAME
6680 && info->valnum != name
6681 && info->value_id != VN_INFO (info->valnum)->value_id)
6682 info->value_id = VN_INFO (info->valnum)->value_id;
6685 set_hashtable_value_ids ();
6687 if (dump_file && (dump_flags & TDF_DETAILS))
6689 fprintf (dump_file, "Value numbers:\n");
6690 FOR_EACH_SSA_NAME (i, name, cfun)
6692 if (VN_INFO (name)->visited
6693 && SSA_VAL (name) != name)
6695 print_generic_expr (dump_file, name);
6696 fprintf (dump_file, " = ");
6697 print_generic_expr (dump_file, SSA_VAL (name));
6698 fprintf (dump_file, " (%04d)\n", VN_INFO (name)->value_id);
6704 /* Free VN associated data structures. */
6706 void
6707 free_rpo_vn (void)
6709 free_vn_table (valid_info);
6710 XDELETE (valid_info);
6711 obstack_free (&vn_tables_obstack, NULL);
6712 obstack_free (&vn_tables_insert_obstack, NULL);
6714 vn_ssa_aux_iterator_type it;
6715 vn_ssa_aux_t info;
6716 FOR_EACH_HASH_TABLE_ELEMENT (*vn_ssa_aux_hash, info, vn_ssa_aux_t, it)
6717 if (info->needs_insertion)
6718 release_ssa_name (info->name);
6719 obstack_free (&vn_ssa_aux_obstack, NULL);
6720 delete vn_ssa_aux_hash;
6722 delete constant_to_value_id;
6723 constant_to_value_id = NULL;
6724 BITMAP_FREE (constant_value_ids);
6727 /* Hook for maybe_push_res_to_seq, lookup the expression in the VN tables. */
6729 static tree
6730 vn_lookup_simplify_result (gimple_match_op *res_op)
6732 if (!res_op->code.is_tree_code ())
6733 return NULL_TREE;
6734 tree *ops = res_op->ops;
6735 unsigned int length = res_op->num_ops;
6736 if (res_op->code == CONSTRUCTOR
6737 /* ??? We're arriving here with SCCVNs view, decomposed CONSTRUCTOR
6738 and GIMPLEs / match-and-simplifies, CONSTRUCTOR as GENERIC tree. */
6739 && TREE_CODE (res_op->ops[0]) == CONSTRUCTOR)
6741 length = CONSTRUCTOR_NELTS (res_op->ops[0]);
6742 ops = XALLOCAVEC (tree, length);
6743 for (unsigned i = 0; i < length; ++i)
6744 ops[i] = CONSTRUCTOR_ELT (res_op->ops[0], i)->value;
6746 vn_nary_op_t vnresult = NULL;
6747 tree res = vn_nary_op_lookup_pieces (length, (tree_code) res_op->code,
6748 res_op->type, ops, &vnresult);
6749 /* If this is used from expression simplification make sure to
6750 return an available expression. */
6751 if (res && TREE_CODE (res) == SSA_NAME && mprts_hook && rpo_avail)
6752 res = rpo_avail->eliminate_avail (vn_context_bb, res);
6753 return res;
6756 /* Return a leader for OPs value that is valid at BB. */
6758 tree
6759 rpo_elim::eliminate_avail (basic_block bb, tree op)
6761 bool visited;
6762 tree valnum = SSA_VAL (op, &visited);
6763 /* If we didn't visit OP then it must be defined outside of the
6764 region we process and also dominate it. So it is available. */
6765 if (!visited)
6766 return op;
6767 if (TREE_CODE (valnum) == SSA_NAME)
6769 if (SSA_NAME_IS_DEFAULT_DEF (valnum))
6770 return valnum;
6771 vn_avail *av = VN_INFO (valnum)->avail;
6772 if (!av)
6773 return NULL_TREE;
6774 if (av->location == bb->index)
6775 /* On tramp3d 90% of the cases are here. */
6776 return ssa_name (av->leader);
6779 basic_block abb = BASIC_BLOCK_FOR_FN (cfun, av->location);
6780 /* ??? During elimination we have to use availability at the
6781 definition site of a use we try to replace. This
6782 is required to not run into inconsistencies because
6783 of dominated_by_p_w_unex behavior and removing a definition
6784 while not replacing all uses.
6785 ??? We could try to consistently walk dominators
6786 ignoring non-executable regions. The nearest common
6787 dominator of bb and abb is where we can stop walking. We
6788 may also be able to "pre-compute" (bits of) the next immediate
6789 (non-)dominator during the RPO walk when marking edges as
6790 executable. */
6791 if (dominated_by_p_w_unex (bb, abb))
6793 tree leader = ssa_name (av->leader);
6794 /* Prevent eliminations that break loop-closed SSA. */
6795 if (loops_state_satisfies_p (LOOP_CLOSED_SSA)
6796 && ! SSA_NAME_IS_DEFAULT_DEF (leader)
6797 && ! flow_bb_inside_loop_p (gimple_bb (SSA_NAME_DEF_STMT
6798 (leader))->loop_father,
6799 bb))
6800 return NULL_TREE;
6801 if (dump_file && (dump_flags & TDF_DETAILS))
6803 print_generic_expr (dump_file, leader);
6804 fprintf (dump_file, " is available for ");
6805 print_generic_expr (dump_file, valnum);
6806 fprintf (dump_file, "\n");
6808 /* On tramp3d 99% of the _remaining_ cases succeed at
6809 the first enty. */
6810 return leader;
6812 /* ??? Can we somehow skip to the immediate dominator
6813 RPO index (bb_to_rpo)? Again, maybe not worth, on
6814 tramp3d the worst number of elements in the vector is 9. */
6815 av = av->next;
6817 while (av);
6819 else if (valnum != VN_TOP)
6820 /* valnum is is_gimple_min_invariant. */
6821 return valnum;
6822 return NULL_TREE;
6825 /* Make LEADER a leader for its value at BB. */
6827 void
6828 rpo_elim::eliminate_push_avail (basic_block bb, tree leader)
6830 tree valnum = VN_INFO (leader)->valnum;
6831 if (valnum == VN_TOP
6832 || is_gimple_min_invariant (valnum))
6833 return;
6834 if (dump_file && (dump_flags & TDF_DETAILS))
6836 fprintf (dump_file, "Making available beyond BB%d ", bb->index);
6837 print_generic_expr (dump_file, leader);
6838 fprintf (dump_file, " for value ");
6839 print_generic_expr (dump_file, valnum);
6840 fprintf (dump_file, "\n");
6842 vn_ssa_aux_t value = VN_INFO (valnum);
6843 vn_avail *av;
6844 if (m_avail_freelist)
6846 av = m_avail_freelist;
6847 m_avail_freelist = m_avail_freelist->next;
6849 else
6850 av = XOBNEW (&vn_ssa_aux_obstack, vn_avail);
6851 av->location = bb->index;
6852 av->leader = SSA_NAME_VERSION (leader);
6853 av->next = value->avail;
6854 value->avail = av;
6857 /* Valueization hook for RPO VN plus required state. */
6859 tree
6860 rpo_vn_valueize (tree name)
6862 if (TREE_CODE (name) == SSA_NAME)
6864 vn_ssa_aux_t val = VN_INFO (name);
6865 if (val)
6867 tree tem = val->valnum;
6868 if (tem != VN_TOP && tem != name)
6870 if (TREE_CODE (tem) != SSA_NAME)
6871 return tem;
6872 /* For all values we only valueize to an available leader
6873 which means we can use SSA name info without restriction. */
6874 tem = rpo_avail->eliminate_avail (vn_context_bb, tem);
6875 if (tem)
6876 return tem;
6880 return name;
6883 /* Insert on PRED_E predicates derived from CODE OPS being true besides the
6884 inverted condition. */
6886 static void
6887 insert_related_predicates_on_edge (enum tree_code code, tree *ops, edge pred_e)
6889 switch (code)
6891 case LT_EXPR:
6892 /* a < b -> a {!,<}= b */
6893 vn_nary_op_insert_pieces_predicated (2, NE_EXPR, boolean_type_node,
6894 ops, boolean_true_node, 0, pred_e);
6895 vn_nary_op_insert_pieces_predicated (2, LE_EXPR, boolean_type_node,
6896 ops, boolean_true_node, 0, pred_e);
6897 /* a < b -> ! a {>,=} b */
6898 vn_nary_op_insert_pieces_predicated (2, GT_EXPR, boolean_type_node,
6899 ops, boolean_false_node, 0, pred_e);
6900 vn_nary_op_insert_pieces_predicated (2, EQ_EXPR, boolean_type_node,
6901 ops, boolean_false_node, 0, pred_e);
6902 break;
6903 case GT_EXPR:
6904 /* a > b -> a {!,>}= b */
6905 vn_nary_op_insert_pieces_predicated (2, NE_EXPR, boolean_type_node,
6906 ops, boolean_true_node, 0, pred_e);
6907 vn_nary_op_insert_pieces_predicated (2, GE_EXPR, boolean_type_node,
6908 ops, boolean_true_node, 0, pred_e);
6909 /* a > b -> ! a {<,=} b */
6910 vn_nary_op_insert_pieces_predicated (2, LT_EXPR, boolean_type_node,
6911 ops, boolean_false_node, 0, pred_e);
6912 vn_nary_op_insert_pieces_predicated (2, EQ_EXPR, boolean_type_node,
6913 ops, boolean_false_node, 0, pred_e);
6914 break;
6915 case EQ_EXPR:
6916 /* a == b -> ! a {<,>} b */
6917 vn_nary_op_insert_pieces_predicated (2, LT_EXPR, boolean_type_node,
6918 ops, boolean_false_node, 0, pred_e);
6919 vn_nary_op_insert_pieces_predicated (2, GT_EXPR, boolean_type_node,
6920 ops, boolean_false_node, 0, pred_e);
6921 break;
6922 case LE_EXPR:
6923 case GE_EXPR:
6924 case NE_EXPR:
6925 /* Nothing besides inverted condition. */
6926 break;
6927 default:;
6931 /* Main stmt worker for RPO VN, process BB. */
6933 static unsigned
6934 process_bb (rpo_elim &avail, basic_block bb,
6935 bool bb_visited, bool iterate_phis, bool iterate, bool eliminate,
6936 bool do_region, bitmap exit_bbs, bool skip_phis)
6938 unsigned todo = 0;
6939 edge_iterator ei;
6940 edge e;
6942 vn_context_bb = bb;
6944 /* If we are in loop-closed SSA preserve this state. This is
6945 relevant when called on regions from outside of FRE/PRE. */
6946 bool lc_phi_nodes = false;
6947 if (!skip_phis
6948 && loops_state_satisfies_p (LOOP_CLOSED_SSA))
6949 FOR_EACH_EDGE (e, ei, bb->preds)
6950 if (e->src->loop_father != e->dest->loop_father
6951 && flow_loop_nested_p (e->dest->loop_father,
6952 e->src->loop_father))
6954 lc_phi_nodes = true;
6955 break;
6958 /* When we visit a loop header substitute into loop info. */
6959 if (!iterate && eliminate && bb->loop_father->header == bb)
6961 /* Keep fields in sync with substitute_in_loop_info. */
6962 if (bb->loop_father->nb_iterations)
6963 bb->loop_father->nb_iterations
6964 = simplify_replace_tree (bb->loop_father->nb_iterations,
6965 NULL_TREE, NULL_TREE, &vn_valueize_wrapper);
6968 /* Value-number all defs in the basic-block. */
6969 if (!skip_phis)
6970 for (gphi_iterator gsi = gsi_start_phis (bb); !gsi_end_p (gsi);
6971 gsi_next (&gsi))
6973 gphi *phi = gsi.phi ();
6974 tree res = PHI_RESULT (phi);
6975 vn_ssa_aux_t res_info = VN_INFO (res);
6976 if (!bb_visited)
6978 gcc_assert (!res_info->visited);
6979 res_info->valnum = VN_TOP;
6980 res_info->visited = true;
6983 /* When not iterating force backedge values to varying. */
6984 visit_stmt (phi, !iterate_phis);
6985 if (virtual_operand_p (res))
6986 continue;
6988 /* Eliminate */
6989 /* The interesting case is gcc.dg/tree-ssa/pr22230.c for correctness
6990 how we handle backedges and availability.
6991 And gcc.dg/tree-ssa/ssa-sccvn-2.c for optimization. */
6992 tree val = res_info->valnum;
6993 if (res != val && !iterate && eliminate)
6995 if (tree leader = avail.eliminate_avail (bb, res))
6997 if (leader != res
6998 /* Preserve loop-closed SSA form. */
6999 && (! lc_phi_nodes
7000 || is_gimple_min_invariant (leader)))
7002 if (dump_file && (dump_flags & TDF_DETAILS))
7004 fprintf (dump_file, "Replaced redundant PHI node "
7005 "defining ");
7006 print_generic_expr (dump_file, res);
7007 fprintf (dump_file, " with ");
7008 print_generic_expr (dump_file, leader);
7009 fprintf (dump_file, "\n");
7011 avail.eliminations++;
7013 if (may_propagate_copy (res, leader))
7015 /* Schedule for removal. */
7016 avail.to_remove.safe_push (phi);
7017 continue;
7019 /* ??? Else generate a copy stmt. */
7023 /* Only make defs available that not already are. But make
7024 sure loop-closed SSA PHI node defs are picked up for
7025 downstream uses. */
7026 if (lc_phi_nodes
7027 || res == val
7028 || ! avail.eliminate_avail (bb, res))
7029 avail.eliminate_push_avail (bb, res);
7032 /* For empty BBs mark outgoing edges executable. For non-empty BBs
7033 we do this when processing the last stmt as we have to do this
7034 before elimination which otherwise forces GIMPLE_CONDs to
7035 if (1 != 0) style when seeing non-executable edges. */
7036 if (gsi_end_p (gsi_start_bb (bb)))
7038 FOR_EACH_EDGE (e, ei, bb->succs)
7040 if (!(e->flags & EDGE_EXECUTABLE))
7042 if (dump_file && (dump_flags & TDF_DETAILS))
7043 fprintf (dump_file,
7044 "marking outgoing edge %d -> %d executable\n",
7045 e->src->index, e->dest->index);
7046 e->flags |= EDGE_EXECUTABLE;
7047 e->dest->flags |= BB_EXECUTABLE;
7049 else if (!(e->dest->flags & BB_EXECUTABLE))
7051 if (dump_file && (dump_flags & TDF_DETAILS))
7052 fprintf (dump_file,
7053 "marking destination block %d reachable\n",
7054 e->dest->index);
7055 e->dest->flags |= BB_EXECUTABLE;
7059 for (gimple_stmt_iterator gsi = gsi_start_bb (bb);
7060 !gsi_end_p (gsi); gsi_next (&gsi))
7062 ssa_op_iter i;
7063 tree op;
7064 if (!bb_visited)
7066 FOR_EACH_SSA_TREE_OPERAND (op, gsi_stmt (gsi), i, SSA_OP_ALL_DEFS)
7068 vn_ssa_aux_t op_info = VN_INFO (op);
7069 gcc_assert (!op_info->visited);
7070 op_info->valnum = VN_TOP;
7071 op_info->visited = true;
7074 /* We somehow have to deal with uses that are not defined
7075 in the processed region. Forcing unvisited uses to
7076 varying here doesn't play well with def-use following during
7077 expression simplification, so we deal with this by checking
7078 the visited flag in SSA_VAL. */
7081 visit_stmt (gsi_stmt (gsi));
7083 gimple *last = gsi_stmt (gsi);
7084 e = NULL;
7085 switch (gimple_code (last))
7087 case GIMPLE_SWITCH:
7088 e = find_taken_edge (bb, vn_valueize (gimple_switch_index
7089 (as_a <gswitch *> (last))));
7090 break;
7091 case GIMPLE_COND:
7093 tree lhs = vn_valueize (gimple_cond_lhs (last));
7094 tree rhs = vn_valueize (gimple_cond_rhs (last));
7095 tree val = gimple_simplify (gimple_cond_code (last),
7096 boolean_type_node, lhs, rhs,
7097 NULL, vn_valueize);
7098 /* If the condition didn't simplfy see if we have recorded
7099 an expression from sofar taken edges. */
7100 if (! val || TREE_CODE (val) != INTEGER_CST)
7102 vn_nary_op_t vnresult;
7103 tree ops[2];
7104 ops[0] = lhs;
7105 ops[1] = rhs;
7106 val = vn_nary_op_lookup_pieces (2, gimple_cond_code (last),
7107 boolean_type_node, ops,
7108 &vnresult);
7109 /* Did we get a predicated value? */
7110 if (! val && vnresult && vnresult->predicated_values)
7112 val = vn_nary_op_get_predicated_value (vnresult, bb);
7113 if (val && dump_file && (dump_flags & TDF_DETAILS))
7115 fprintf (dump_file, "Got predicated value ");
7116 print_generic_expr (dump_file, val, TDF_NONE);
7117 fprintf (dump_file, " for ");
7118 print_gimple_stmt (dump_file, last, TDF_SLIM);
7122 if (val)
7123 e = find_taken_edge (bb, val);
7124 if (! e)
7126 /* If we didn't manage to compute the taken edge then
7127 push predicated expressions for the condition itself
7128 and related conditions to the hashtables. This allows
7129 simplification of redundant conditions which is
7130 important as early cleanup. */
7131 edge true_e, false_e;
7132 extract_true_false_edges_from_block (bb, &true_e, &false_e);
7133 enum tree_code code = gimple_cond_code (last);
7134 enum tree_code icode
7135 = invert_tree_comparison (code, HONOR_NANS (lhs));
7136 tree ops[2];
7137 ops[0] = lhs;
7138 ops[1] = rhs;
7139 if (do_region
7140 && bitmap_bit_p (exit_bbs, true_e->dest->index))
7141 true_e = NULL;
7142 if (do_region
7143 && bitmap_bit_p (exit_bbs, false_e->dest->index))
7144 false_e = NULL;
7145 if (true_e)
7146 vn_nary_op_insert_pieces_predicated
7147 (2, code, boolean_type_node, ops,
7148 boolean_true_node, 0, true_e);
7149 if (false_e)
7150 vn_nary_op_insert_pieces_predicated
7151 (2, code, boolean_type_node, ops,
7152 boolean_false_node, 0, false_e);
7153 if (icode != ERROR_MARK)
7155 if (true_e)
7156 vn_nary_op_insert_pieces_predicated
7157 (2, icode, boolean_type_node, ops,
7158 boolean_false_node, 0, true_e);
7159 if (false_e)
7160 vn_nary_op_insert_pieces_predicated
7161 (2, icode, boolean_type_node, ops,
7162 boolean_true_node, 0, false_e);
7164 /* Relax for non-integers, inverted condition handled
7165 above. */
7166 if (INTEGRAL_TYPE_P (TREE_TYPE (lhs)))
7168 if (true_e)
7169 insert_related_predicates_on_edge (code, ops, true_e);
7170 if (false_e)
7171 insert_related_predicates_on_edge (icode, ops, false_e);
7174 break;
7176 case GIMPLE_GOTO:
7177 e = find_taken_edge (bb, vn_valueize (gimple_goto_dest (last)));
7178 break;
7179 default:
7180 e = NULL;
7182 if (e)
7184 todo = TODO_cleanup_cfg;
7185 if (!(e->flags & EDGE_EXECUTABLE))
7187 if (dump_file && (dump_flags & TDF_DETAILS))
7188 fprintf (dump_file,
7189 "marking known outgoing %sedge %d -> %d executable\n",
7190 e->flags & EDGE_DFS_BACK ? "back-" : "",
7191 e->src->index, e->dest->index);
7192 e->flags |= EDGE_EXECUTABLE;
7193 e->dest->flags |= BB_EXECUTABLE;
7195 else if (!(e->dest->flags & BB_EXECUTABLE))
7197 if (dump_file && (dump_flags & TDF_DETAILS))
7198 fprintf (dump_file,
7199 "marking destination block %d reachable\n",
7200 e->dest->index);
7201 e->dest->flags |= BB_EXECUTABLE;
7204 else if (gsi_one_before_end_p (gsi))
7206 FOR_EACH_EDGE (e, ei, bb->succs)
7208 if (!(e->flags & EDGE_EXECUTABLE))
7210 if (dump_file && (dump_flags & TDF_DETAILS))
7211 fprintf (dump_file,
7212 "marking outgoing edge %d -> %d executable\n",
7213 e->src->index, e->dest->index);
7214 e->flags |= EDGE_EXECUTABLE;
7215 e->dest->flags |= BB_EXECUTABLE;
7217 else if (!(e->dest->flags & BB_EXECUTABLE))
7219 if (dump_file && (dump_flags & TDF_DETAILS))
7220 fprintf (dump_file,
7221 "marking destination block %d reachable\n",
7222 e->dest->index);
7223 e->dest->flags |= BB_EXECUTABLE;
7228 /* Eliminate. That also pushes to avail. */
7229 if (eliminate && ! iterate)
7230 avail.eliminate_stmt (bb, &gsi);
7231 else
7232 /* If not eliminating, make all not already available defs
7233 available. */
7234 FOR_EACH_SSA_TREE_OPERAND (op, gsi_stmt (gsi), i, SSA_OP_DEF)
7235 if (! avail.eliminate_avail (bb, op))
7236 avail.eliminate_push_avail (bb, op);
7239 /* Eliminate in destination PHI arguments. Always substitute in dest
7240 PHIs, even for non-executable edges. This handles region
7241 exits PHIs. */
7242 if (!iterate && eliminate)
7243 FOR_EACH_EDGE (e, ei, bb->succs)
7244 for (gphi_iterator gsi = gsi_start_phis (e->dest);
7245 !gsi_end_p (gsi); gsi_next (&gsi))
7247 gphi *phi = gsi.phi ();
7248 use_operand_p use_p = PHI_ARG_DEF_PTR_FROM_EDGE (phi, e);
7249 tree arg = USE_FROM_PTR (use_p);
7250 if (TREE_CODE (arg) != SSA_NAME
7251 || virtual_operand_p (arg))
7252 continue;
7253 tree sprime;
7254 if (SSA_NAME_IS_DEFAULT_DEF (arg))
7256 sprime = SSA_VAL (arg);
7257 gcc_assert (TREE_CODE (sprime) != SSA_NAME
7258 || SSA_NAME_IS_DEFAULT_DEF (sprime));
7260 else
7261 /* Look for sth available at the definition block of the argument.
7262 This avoids inconsistencies between availability there which
7263 decides if the stmt can be removed and availability at the
7264 use site. The SSA property ensures that things available
7265 at the definition are also available at uses. */
7266 sprime = avail.eliminate_avail (gimple_bb (SSA_NAME_DEF_STMT (arg)),
7267 arg);
7268 if (sprime
7269 && sprime != arg
7270 && may_propagate_copy (arg, sprime))
7271 propagate_value (use_p, sprime);
7274 vn_context_bb = NULL;
7275 return todo;
7278 /* Unwind state per basic-block. */
7280 struct unwind_state
7282 /* Times this block has been visited. */
7283 unsigned visited;
7284 /* Whether to handle this as iteration point or whether to treat
7285 incoming backedge PHI values as varying. */
7286 bool iterate;
7287 /* Maximum RPO index this block is reachable from. */
7288 int max_rpo;
7289 /* Unwind state. */
7290 void *ob_top;
7291 vn_reference_t ref_top;
7292 vn_phi_t phi_top;
7293 vn_nary_op_t nary_top;
7296 /* Unwind the RPO VN state for iteration. */
7298 static void
7299 do_unwind (unwind_state *to, int rpo_idx, rpo_elim &avail, int *bb_to_rpo)
7301 gcc_assert (to->iterate);
7302 for (; last_inserted_nary != to->nary_top;
7303 last_inserted_nary = last_inserted_nary->next)
7305 vn_nary_op_t *slot;
7306 slot = valid_info->nary->find_slot_with_hash
7307 (last_inserted_nary, last_inserted_nary->hashcode, NO_INSERT);
7308 /* Predication causes the need to restore previous state. */
7309 if ((*slot)->unwind_to)
7310 *slot = (*slot)->unwind_to;
7311 else
7312 valid_info->nary->clear_slot (slot);
7314 for (; last_inserted_phi != to->phi_top;
7315 last_inserted_phi = last_inserted_phi->next)
7317 vn_phi_t *slot;
7318 slot = valid_info->phis->find_slot_with_hash
7319 (last_inserted_phi, last_inserted_phi->hashcode, NO_INSERT);
7320 valid_info->phis->clear_slot (slot);
7322 for (; last_inserted_ref != to->ref_top;
7323 last_inserted_ref = last_inserted_ref->next)
7325 vn_reference_t *slot;
7326 slot = valid_info->references->find_slot_with_hash
7327 (last_inserted_ref, last_inserted_ref->hashcode, NO_INSERT);
7328 (*slot)->operands.release ();
7329 valid_info->references->clear_slot (slot);
7331 obstack_free (&vn_tables_obstack, to->ob_top);
7333 /* Prune [rpo_idx, ] from avail. */
7334 /* ??? This is O(number-of-values-in-region) which is
7335 O(region-size) rather than O(iteration-piece). */
7336 for (hash_table<vn_ssa_aux_hasher>::iterator i = vn_ssa_aux_hash->begin ();
7337 i != vn_ssa_aux_hash->end (); ++i)
7339 while ((*i)->avail)
7341 if (bb_to_rpo[(*i)->avail->location] < rpo_idx)
7342 break;
7343 vn_avail *av = (*i)->avail;
7344 (*i)->avail = (*i)->avail->next;
7345 av->next = avail.m_avail_freelist;
7346 avail.m_avail_freelist = av;
7351 /* Do VN on a SEME region specified by ENTRY and EXIT_BBS in FN.
7352 If ITERATE is true then treat backedges optimistically as not
7353 executed and iterate. If ELIMINATE is true then perform
7354 elimination, otherwise leave that to the caller. */
7356 static unsigned
7357 do_rpo_vn (function *fn, edge entry, bitmap exit_bbs,
7358 bool iterate, bool eliminate)
7360 unsigned todo = 0;
7362 /* We currently do not support region-based iteration when
7363 elimination is requested. */
7364 gcc_assert (!entry || !iterate || !eliminate);
7365 /* When iterating we need loop info up-to-date. */
7366 gcc_assert (!iterate || !loops_state_satisfies_p (LOOPS_NEED_FIXUP));
7368 bool do_region = entry != NULL;
7369 if (!do_region)
7371 entry = single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (fn));
7372 exit_bbs = BITMAP_ALLOC (NULL);
7373 bitmap_set_bit (exit_bbs, EXIT_BLOCK);
7376 /* Clear EDGE_DFS_BACK on "all" entry edges, RPO order compute will
7377 re-mark those that are contained in the region. */
7378 edge_iterator ei;
7379 edge e;
7380 FOR_EACH_EDGE (e, ei, entry->dest->preds)
7381 e->flags &= ~EDGE_DFS_BACK;
7383 int *rpo = XNEWVEC (int, n_basic_blocks_for_fn (fn) - NUM_FIXED_BLOCKS);
7384 auto_vec<std::pair<int, int> > toplevel_scc_extents;
7385 int n = rev_post_order_and_mark_dfs_back_seme
7386 (fn, entry, exit_bbs, true, rpo, !iterate ? &toplevel_scc_extents : NULL);
7388 if (!do_region)
7389 BITMAP_FREE (exit_bbs);
7391 /* If there are any non-DFS_BACK edges into entry->dest skip
7392 processing PHI nodes for that block. This supports
7393 value-numbering loop bodies w/o the actual loop. */
7394 FOR_EACH_EDGE (e, ei, entry->dest->preds)
7395 if (e != entry
7396 && !(e->flags & EDGE_DFS_BACK))
7397 break;
7398 bool skip_entry_phis = e != NULL;
7399 if (skip_entry_phis && dump_file && (dump_flags & TDF_DETAILS))
7400 fprintf (dump_file, "Region does not contain all edges into "
7401 "the entry block, skipping its PHIs.\n");
7403 int *bb_to_rpo = XNEWVEC (int, last_basic_block_for_fn (fn));
7404 for (int i = 0; i < n; ++i)
7405 bb_to_rpo[rpo[i]] = i;
7407 unwind_state *rpo_state = XNEWVEC (unwind_state, n);
7409 rpo_elim avail (entry->dest);
7410 rpo_avail = &avail;
7412 /* Verify we have no extra entries into the region. */
7413 if (flag_checking && do_region)
7415 auto_bb_flag bb_in_region (fn);
7416 for (int i = 0; i < n; ++i)
7418 basic_block bb = BASIC_BLOCK_FOR_FN (fn, rpo[i]);
7419 bb->flags |= bb_in_region;
7421 /* We can't merge the first two loops because we cannot rely
7422 on EDGE_DFS_BACK for edges not within the region. But if
7423 we decide to always have the bb_in_region flag we can
7424 do the checking during the RPO walk itself (but then it's
7425 also easy to handle MEME conservatively). */
7426 for (int i = 0; i < n; ++i)
7428 basic_block bb = BASIC_BLOCK_FOR_FN (fn, rpo[i]);
7429 edge e;
7430 edge_iterator ei;
7431 FOR_EACH_EDGE (e, ei, bb->preds)
7432 gcc_assert (e == entry
7433 || (skip_entry_phis && bb == entry->dest)
7434 || (e->src->flags & bb_in_region));
7436 for (int i = 0; i < n; ++i)
7438 basic_block bb = BASIC_BLOCK_FOR_FN (fn, rpo[i]);
7439 bb->flags &= ~bb_in_region;
7443 /* Create the VN state. For the initial size of the various hashtables
7444 use a heuristic based on region size and number of SSA names. */
7445 unsigned region_size = (((unsigned HOST_WIDE_INT)n * num_ssa_names)
7446 / (n_basic_blocks_for_fn (fn) - NUM_FIXED_BLOCKS));
7447 VN_TOP = create_tmp_var_raw (void_type_node, "vn_top");
7448 next_value_id = 1;
7450 vn_ssa_aux_hash = new hash_table <vn_ssa_aux_hasher> (region_size * 2);
7451 gcc_obstack_init (&vn_ssa_aux_obstack);
7453 gcc_obstack_init (&vn_tables_obstack);
7454 gcc_obstack_init (&vn_tables_insert_obstack);
7455 valid_info = XCNEW (struct vn_tables_s);
7456 allocate_vn_table (valid_info, region_size);
7457 last_inserted_ref = NULL;
7458 last_inserted_phi = NULL;
7459 last_inserted_nary = NULL;
7461 vn_valueize = rpo_vn_valueize;
7463 /* Initialize the unwind state and edge/BB executable state. */
7464 unsigned curr_scc = 0;
7465 for (int i = 0; i < n; ++i)
7467 basic_block bb = BASIC_BLOCK_FOR_FN (fn, rpo[i]);
7468 rpo_state[i].visited = 0;
7469 rpo_state[i].max_rpo = i;
7470 if (!iterate && curr_scc < toplevel_scc_extents.length ())
7472 if (i >= toplevel_scc_extents[curr_scc].first
7473 && i <= toplevel_scc_extents[curr_scc].second)
7474 rpo_state[i].max_rpo = toplevel_scc_extents[curr_scc].second;
7475 if (i == toplevel_scc_extents[curr_scc].second)
7476 curr_scc++;
7478 bb->flags &= ~BB_EXECUTABLE;
7479 bool has_backedges = false;
7480 edge e;
7481 edge_iterator ei;
7482 FOR_EACH_EDGE (e, ei, bb->preds)
7484 if (e->flags & EDGE_DFS_BACK)
7485 has_backedges = true;
7486 e->flags &= ~EDGE_EXECUTABLE;
7487 if (iterate || e == entry || (skip_entry_phis && bb == entry->dest))
7488 continue;
7490 rpo_state[i].iterate = iterate && has_backedges;
7492 entry->flags |= EDGE_EXECUTABLE;
7493 entry->dest->flags |= BB_EXECUTABLE;
7495 /* As heuristic to improve compile-time we handle only the N innermost
7496 loops and the outermost one optimistically. */
7497 if (iterate)
7499 loop_p loop;
7500 unsigned max_depth = param_rpo_vn_max_loop_depth;
7501 FOR_EACH_LOOP (loop, LI_ONLY_INNERMOST)
7502 if (loop_depth (loop) > max_depth)
7503 for (unsigned i = 2;
7504 i < loop_depth (loop) - max_depth; ++i)
7506 basic_block header = superloop_at_depth (loop, i)->header;
7507 bool non_latch_backedge = false;
7508 edge e;
7509 edge_iterator ei;
7510 FOR_EACH_EDGE (e, ei, header->preds)
7511 if (e->flags & EDGE_DFS_BACK)
7513 /* There can be a non-latch backedge into the header
7514 which is part of an outer irreducible region. We
7515 cannot avoid iterating this block then. */
7516 if (!dominated_by_p (CDI_DOMINATORS,
7517 e->src, e->dest))
7519 if (dump_file && (dump_flags & TDF_DETAILS))
7520 fprintf (dump_file, "non-latch backedge %d -> %d "
7521 "forces iteration of loop %d\n",
7522 e->src->index, e->dest->index, loop->num);
7523 non_latch_backedge = true;
7525 else
7526 e->flags |= EDGE_EXECUTABLE;
7528 rpo_state[bb_to_rpo[header->index]].iterate = non_latch_backedge;
7532 uint64_t nblk = 0;
7533 int idx = 0;
7534 if (iterate)
7535 /* Go and process all blocks, iterating as necessary. */
7538 basic_block bb = BASIC_BLOCK_FOR_FN (fn, rpo[idx]);
7540 /* If the block has incoming backedges remember unwind state. This
7541 is required even for non-executable blocks since in irreducible
7542 regions we might reach them via the backedge and re-start iterating
7543 from there.
7544 Note we can individually mark blocks with incoming backedges to
7545 not iterate where we then handle PHIs conservatively. We do that
7546 heuristically to reduce compile-time for degenerate cases. */
7547 if (rpo_state[idx].iterate)
7549 rpo_state[idx].ob_top = obstack_alloc (&vn_tables_obstack, 0);
7550 rpo_state[idx].ref_top = last_inserted_ref;
7551 rpo_state[idx].phi_top = last_inserted_phi;
7552 rpo_state[idx].nary_top = last_inserted_nary;
7555 if (!(bb->flags & BB_EXECUTABLE))
7557 if (dump_file && (dump_flags & TDF_DETAILS))
7558 fprintf (dump_file, "Block %d: BB%d found not executable\n",
7559 idx, bb->index);
7560 idx++;
7561 continue;
7564 if (dump_file && (dump_flags & TDF_DETAILS))
7565 fprintf (dump_file, "Processing block %d: BB%d\n", idx, bb->index);
7566 nblk++;
7567 todo |= process_bb (avail, bb,
7568 rpo_state[idx].visited != 0,
7569 rpo_state[idx].iterate,
7570 iterate, eliminate, do_region, exit_bbs, false);
7571 rpo_state[idx].visited++;
7573 /* Verify if changed values flow over executable outgoing backedges
7574 and those change destination PHI values (that's the thing we
7575 can easily verify). Reduce over all such edges to the farthest
7576 away PHI. */
7577 int iterate_to = -1;
7578 edge_iterator ei;
7579 edge e;
7580 FOR_EACH_EDGE (e, ei, bb->succs)
7581 if ((e->flags & (EDGE_DFS_BACK|EDGE_EXECUTABLE))
7582 == (EDGE_DFS_BACK|EDGE_EXECUTABLE)
7583 && rpo_state[bb_to_rpo[e->dest->index]].iterate)
7585 int destidx = bb_to_rpo[e->dest->index];
7586 if (!rpo_state[destidx].visited)
7588 if (dump_file && (dump_flags & TDF_DETAILS))
7589 fprintf (dump_file, "Unvisited destination %d\n",
7590 e->dest->index);
7591 if (iterate_to == -1 || destidx < iterate_to)
7592 iterate_to = destidx;
7593 continue;
7595 if (dump_file && (dump_flags & TDF_DETAILS))
7596 fprintf (dump_file, "Looking for changed values of backedge"
7597 " %d->%d destination PHIs\n",
7598 e->src->index, e->dest->index);
7599 vn_context_bb = e->dest;
7600 gphi_iterator gsi;
7601 for (gsi = gsi_start_phis (e->dest);
7602 !gsi_end_p (gsi); gsi_next (&gsi))
7604 bool inserted = false;
7605 /* While we'd ideally just iterate on value changes
7606 we CSE PHIs and do that even across basic-block
7607 boundaries. So even hashtable state changes can
7608 be important (which is roughly equivalent to
7609 PHI argument value changes). To not excessively
7610 iterate because of that we track whether a PHI
7611 was CSEd to with GF_PLF_1. */
7612 bool phival_changed;
7613 if ((phival_changed = visit_phi (gsi.phi (),
7614 &inserted, false))
7615 || (inserted && gimple_plf (gsi.phi (), GF_PLF_1)))
7617 if (!phival_changed
7618 && dump_file && (dump_flags & TDF_DETAILS))
7619 fprintf (dump_file, "PHI was CSEd and hashtable "
7620 "state (changed)\n");
7621 if (iterate_to == -1 || destidx < iterate_to)
7622 iterate_to = destidx;
7623 break;
7626 vn_context_bb = NULL;
7628 if (iterate_to != -1)
7630 do_unwind (&rpo_state[iterate_to], iterate_to, avail, bb_to_rpo);
7631 idx = iterate_to;
7632 if (dump_file && (dump_flags & TDF_DETAILS))
7633 fprintf (dump_file, "Iterating to %d BB%d\n",
7634 iterate_to, rpo[iterate_to]);
7635 continue;
7638 idx++;
7640 while (idx < n);
7642 else /* !iterate */
7644 /* Process all blocks greedily with a worklist that enforces RPO
7645 processing of reachable blocks. */
7646 auto_bitmap worklist;
7647 bitmap_set_bit (worklist, 0);
7648 while (!bitmap_empty_p (worklist))
7650 int idx = bitmap_first_set_bit (worklist);
7651 bitmap_clear_bit (worklist, idx);
7652 basic_block bb = BASIC_BLOCK_FOR_FN (fn, rpo[idx]);
7653 gcc_assert ((bb->flags & BB_EXECUTABLE)
7654 && !rpo_state[idx].visited);
7656 if (dump_file && (dump_flags & TDF_DETAILS))
7657 fprintf (dump_file, "Processing block %d: BB%d\n", idx, bb->index);
7659 /* When we run into predecessor edges where we cannot trust its
7660 executable state mark them executable so PHI processing will
7661 be conservative.
7662 ??? Do we need to force arguments flowing over that edge
7663 to be varying or will they even always be? */
7664 edge_iterator ei;
7665 edge e;
7666 FOR_EACH_EDGE (e, ei, bb->preds)
7667 if (!(e->flags & EDGE_EXECUTABLE)
7668 && (bb == entry->dest
7669 || (!rpo_state[bb_to_rpo[e->src->index]].visited
7670 && (rpo_state[bb_to_rpo[e->src->index]].max_rpo
7671 >= (int)idx))))
7673 if (dump_file && (dump_flags & TDF_DETAILS))
7674 fprintf (dump_file, "Cannot trust state of predecessor "
7675 "edge %d -> %d, marking executable\n",
7676 e->src->index, e->dest->index);
7677 e->flags |= EDGE_EXECUTABLE;
7680 nblk++;
7681 todo |= process_bb (avail, bb, false, false, false, eliminate,
7682 do_region, exit_bbs,
7683 skip_entry_phis && bb == entry->dest);
7684 rpo_state[idx].visited++;
7686 FOR_EACH_EDGE (e, ei, bb->succs)
7687 if ((e->flags & EDGE_EXECUTABLE)
7688 && e->dest->index != EXIT_BLOCK
7689 && (!do_region || !bitmap_bit_p (exit_bbs, e->dest->index))
7690 && !rpo_state[bb_to_rpo[e->dest->index]].visited)
7691 bitmap_set_bit (worklist, bb_to_rpo[e->dest->index]);
7695 /* If statistics or dump file active. */
7696 int nex = 0;
7697 unsigned max_visited = 1;
7698 for (int i = 0; i < n; ++i)
7700 basic_block bb = BASIC_BLOCK_FOR_FN (fn, rpo[i]);
7701 if (bb->flags & BB_EXECUTABLE)
7702 nex++;
7703 statistics_histogram_event (cfun, "RPO block visited times",
7704 rpo_state[i].visited);
7705 if (rpo_state[i].visited > max_visited)
7706 max_visited = rpo_state[i].visited;
7708 unsigned nvalues = 0, navail = 0;
7709 for (hash_table<vn_ssa_aux_hasher>::iterator i = vn_ssa_aux_hash->begin ();
7710 i != vn_ssa_aux_hash->end (); ++i)
7712 nvalues++;
7713 vn_avail *av = (*i)->avail;
7714 while (av)
7716 navail++;
7717 av = av->next;
7720 statistics_counter_event (cfun, "RPO blocks", n);
7721 statistics_counter_event (cfun, "RPO blocks visited", nblk);
7722 statistics_counter_event (cfun, "RPO blocks executable", nex);
7723 statistics_histogram_event (cfun, "RPO iterations", 10*nblk / nex);
7724 statistics_histogram_event (cfun, "RPO num values", nvalues);
7725 statistics_histogram_event (cfun, "RPO num avail", navail);
7726 statistics_histogram_event (cfun, "RPO num lattice",
7727 vn_ssa_aux_hash->elements ());
7728 if (dump_file && (dump_flags & (TDF_DETAILS|TDF_STATS)))
7730 fprintf (dump_file, "RPO iteration over %d blocks visited %" PRIu64
7731 " blocks in total discovering %d executable blocks iterating "
7732 "%d.%d times, a block was visited max. %u times\n",
7733 n, nblk, nex,
7734 (int)((10*nblk / nex)/10), (int)((10*nblk / nex)%10),
7735 max_visited);
7736 fprintf (dump_file, "RPO tracked %d values available at %d locations "
7737 "and %" PRIu64 " lattice elements\n",
7738 nvalues, navail, (uint64_t) vn_ssa_aux_hash->elements ());
7741 if (eliminate)
7743 /* When !iterate we already performed elimination during the RPO
7744 walk. */
7745 if (iterate)
7747 /* Elimination for region-based VN needs to be done within the
7748 RPO walk. */
7749 gcc_assert (! do_region);
7750 /* Note we can't use avail.walk here because that gets confused
7751 by the existing availability and it will be less efficient
7752 as well. */
7753 todo |= eliminate_with_rpo_vn (NULL);
7755 else
7756 todo |= avail.eliminate_cleanup (do_region);
7759 vn_valueize = NULL;
7760 rpo_avail = NULL;
7762 XDELETEVEC (bb_to_rpo);
7763 XDELETEVEC (rpo);
7764 XDELETEVEC (rpo_state);
7766 return todo;
7769 /* Region-based entry for RPO VN. Performs value-numbering and elimination
7770 on the SEME region specified by ENTRY and EXIT_BBS. If ENTRY is not
7771 the only edge into the region at ENTRY->dest PHI nodes in ENTRY->dest
7772 are not considered. */
7774 unsigned
7775 do_rpo_vn (function *fn, edge entry, bitmap exit_bbs)
7777 default_vn_walk_kind = VN_WALKREWRITE;
7778 unsigned todo = do_rpo_vn (fn, entry, exit_bbs, false, true);
7779 free_rpo_vn ();
7780 return todo;
7784 namespace {
7786 const pass_data pass_data_fre =
7788 GIMPLE_PASS, /* type */
7789 "fre", /* name */
7790 OPTGROUP_NONE, /* optinfo_flags */
7791 TV_TREE_FRE, /* tv_id */
7792 ( PROP_cfg | PROP_ssa ), /* properties_required */
7793 0, /* properties_provided */
7794 0, /* properties_destroyed */
7795 0, /* todo_flags_start */
7796 0, /* todo_flags_finish */
7799 class pass_fre : public gimple_opt_pass
7801 public:
7802 pass_fre (gcc::context *ctxt)
7803 : gimple_opt_pass (pass_data_fre, ctxt), may_iterate (true)
7806 /* opt_pass methods: */
7807 opt_pass * clone () { return new pass_fre (m_ctxt); }
7808 void set_pass_param (unsigned int n, bool param)
7810 gcc_assert (n == 0);
7811 may_iterate = param;
7813 virtual bool gate (function *)
7815 return flag_tree_fre != 0 && (may_iterate || optimize > 1);
7817 virtual unsigned int execute (function *);
7819 private:
7820 bool may_iterate;
7821 }; // class pass_fre
7823 unsigned int
7824 pass_fre::execute (function *fun)
7826 unsigned todo = 0;
7828 /* At -O[1g] use the cheap non-iterating mode. */
7829 bool iterate_p = may_iterate && (optimize > 1);
7830 calculate_dominance_info (CDI_DOMINATORS);
7831 if (iterate_p)
7832 loop_optimizer_init (AVOID_CFG_MODIFICATIONS);
7834 default_vn_walk_kind = VN_WALKREWRITE;
7835 todo = do_rpo_vn (fun, NULL, NULL, iterate_p, true);
7836 free_rpo_vn ();
7838 if (iterate_p)
7839 loop_optimizer_finalize ();
7841 /* For late FRE after IVOPTs and unrolling, see if we can
7842 remove some TREE_ADDRESSABLE and rewrite stuff into SSA. */
7843 if (!may_iterate)
7844 todo |= TODO_update_address_taken;
7846 return todo;
7849 } // anon namespace
7851 gimple_opt_pass *
7852 make_pass_fre (gcc::context *ctxt)
7854 return new pass_fre (ctxt);
7857 #undef BB_EXECUTABLE