1 /* Tree based points-to analysis
2 Copyright (C) 2005, 2006, 2007, 2008, 2009, 2010
3 Free Software Foundation, Inc.
4 Contributed by Daniel Berlin <dberlin@dberlin.org>
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify
9 under the terms of the GNU General Public License as published by
10 the Free Software Foundation; either version 3 of the License, or
11 (at your option) any later version.
13 GCC is distributed in the hope that it will be useful,
14 but WITHOUT ANY WARRANTY; without even the implied warranty of
15 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 GNU General Public License for more details.
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
24 #include "coretypes.h"
32 #include "hard-reg-set.h"
33 #include "basic-block.h"
36 #include "tree-flow.h"
37 #include "tree-inline.h"
38 #include "diagnostic.h"
44 #include "tree-pass.h"
46 #include "alloc-pool.h"
47 #include "splay-tree.h"
51 #include "pointer-set.h"
53 /* The idea behind this analyzer is to generate set constraints from the
54 program, then solve the resulting constraints in order to generate the
57 Set constraints are a way of modeling program analysis problems that
58 involve sets. They consist of an inclusion constraint language,
59 describing the variables (each variable is a set) and operations that
60 are involved on the variables, and a set of rules that derive facts
61 from these operations. To solve a system of set constraints, you derive
62 all possible facts under the rules, which gives you the correct sets
65 See "Efficient Field-sensitive pointer analysis for C" by "David
66 J. Pearce and Paul H. J. Kelly and Chris Hankin, at
67 http://citeseer.ist.psu.edu/pearce04efficient.html
69 Also see "Ultra-fast Aliasing Analysis using CLA: A Million Lines
70 of C Code in a Second" by ""Nevin Heintze and Olivier Tardieu" at
71 http://citeseer.ist.psu.edu/heintze01ultrafast.html
73 There are three types of real constraint expressions, DEREF,
74 ADDRESSOF, and SCALAR. Each constraint expression consists
75 of a constraint type, a variable, and an offset.
77 SCALAR is a constraint expression type used to represent x, whether
78 it appears on the LHS or the RHS of a statement.
79 DEREF is a constraint expression type used to represent *x, whether
80 it appears on the LHS or the RHS of a statement.
81 ADDRESSOF is a constraint expression used to represent &x, whether
82 it appears on the LHS or the RHS of a statement.
84 Each pointer variable in the program is assigned an integer id, and
85 each field of a structure variable is assigned an integer id as well.
87 Structure variables are linked to their list of fields through a "next
88 field" in each variable that points to the next field in offset
90 Each variable for a structure field has
92 1. "size", that tells the size in bits of that field.
93 2. "fullsize, that tells the size in bits of the entire structure.
94 3. "offset", that tells the offset in bits from the beginning of the
95 structure to this field.
107 foo.a -> id 1, size 32, offset 0, fullsize 64, next foo.b
108 foo.b -> id 2, size 32, offset 32, fullsize 64, next NULL
109 bar -> id 3, size 32, offset 0, fullsize 32, next NULL
112 In order to solve the system of set constraints, the following is
115 1. Each constraint variable x has a solution set associated with it,
118 2. Constraints are separated into direct, copy, and complex.
119 Direct constraints are ADDRESSOF constraints that require no extra
120 processing, such as P = &Q
121 Copy constraints are those of the form P = Q.
122 Complex constraints are all the constraints involving dereferences
123 and offsets (including offsetted copies).
125 3. All direct constraints of the form P = &Q are processed, such
126 that Q is added to Sol(P)
128 4. All complex constraints for a given constraint variable are stored in a
129 linked list attached to that variable's node.
131 5. A directed graph is built out of the copy constraints. Each
132 constraint variable is a node in the graph, and an edge from
133 Q to P is added for each copy constraint of the form P = Q
135 6. The graph is then walked, and solution sets are
136 propagated along the copy edges, such that an edge from Q to P
137 causes Sol(P) <- Sol(P) union Sol(Q).
139 7. As we visit each node, all complex constraints associated with
140 that node are processed by adding appropriate copy edges to the graph, or the
141 appropriate variables to the solution set.
143 8. The process of walking the graph is iterated until no solution
146 Prior to walking the graph in steps 6 and 7, We perform static
147 cycle elimination on the constraint graph, as well
148 as off-line variable substitution.
150 TODO: Adding offsets to pointer-to-structures can be handled (IE not punted
151 on and turned into anything), but isn't. You can just see what offset
152 inside the pointed-to struct it's going to access.
154 TODO: Constant bounded arrays can be handled as if they were structs of the
155 same number of elements.
157 TODO: Modeling heap and incoming pointers becomes much better if we
158 add fields to them as we discover them, which we could do.
160 TODO: We could handle unions, but to be honest, it's probably not
161 worth the pain or slowdown. */
163 /* IPA-PTA optimizations possible.
165 When the indirect function called is ANYTHING we can add disambiguation
166 based on the function signatures (or simply the parameter count which
167 is the varinfo size). We also do not need to consider functions that
168 do not have their address taken.
170 The is_global_var bit which marks escape points is overly conservative
171 in IPA mode. Split it to is_escape_point and is_global_var - only
172 externally visible globals are escape points in IPA mode. This is
173 also needed to fix the pt_solution_includes_global predicate
174 (and thus ptr_deref_may_alias_global_p).
176 The way we introduce DECL_PT_UID to avoid fixing up all points-to
177 sets in the translation unit when we copy a DECL during inlining
178 pessimizes precision. The advantage is that the DECL_PT_UID keeps
179 compile-time and memory usage overhead low - the points-to sets
180 do not grow or get unshared as they would during a fixup phase.
181 An alternative solution is to delay IPA PTA until after all
182 inlining transformations have been applied.
184 The way we propagate clobber/use information isn't optimized.
185 It should use a new complex constraint that properly filters
186 out local variables of the callee (though that would make
187 the sets invalid after inlining). OTOH we might as well
188 admit defeat to WHOPR and simply do all the clobber/use analysis
189 and propagation after PTA finished but before we threw away
190 points-to information for memory variables. WHOPR and PTA
191 do not play along well anyway - the whole constraint solving
192 would need to be done in WPA phase and it will be very interesting
193 to apply the results to local SSA names during LTRANS phase.
195 We probably should compute a per-function unit-ESCAPE solution
196 propagating it simply like the clobber / uses solutions. The
197 solution can go alongside the non-IPA espaced solution and be
198 used to query which vars escape the unit through a function.
200 We never put function decls in points-to sets so we do not
201 keep the set of called functions for indirect calls.
203 And probably more. */
205 static GTY ((if_marked ("tree_map_marked_p"), param_is (struct tree_map)))
206 htab_t heapvar_for_stmt;
208 static bool use_field_sensitive = true;
209 static int in_ipa_mode = 0;
211 /* Used for predecessor bitmaps. */
212 static bitmap_obstack predbitmap_obstack;
214 /* Used for points-to sets. */
215 static bitmap_obstack pta_obstack;
217 /* Used for oldsolution members of variables. */
218 static bitmap_obstack oldpta_obstack;
220 /* Used for per-solver-iteration bitmaps. */
221 static bitmap_obstack iteration_obstack;
223 static unsigned int create_variable_info_for (tree, const char *);
224 typedef struct constraint_graph *constraint_graph_t;
225 static void unify_nodes (constraint_graph_t, unsigned int, unsigned int, bool);
228 typedef struct constraint *constraint_t;
230 DEF_VEC_P(constraint_t);
231 DEF_VEC_ALLOC_P(constraint_t,heap);
233 #define EXECUTE_IF_IN_NONNULL_BITMAP(a, b, c, d) \
235 EXECUTE_IF_SET_IN_BITMAP (a, b, c, d)
237 static struct constraint_stats
239 unsigned int total_vars;
240 unsigned int nonpointer_vars;
241 unsigned int unified_vars_static;
242 unsigned int unified_vars_dynamic;
243 unsigned int iterations;
244 unsigned int num_edges;
245 unsigned int num_implicit_edges;
246 unsigned int points_to_sets_created;
251 /* ID of this variable */
254 /* True if this is a variable created by the constraint analysis, such as
255 heap variables and constraints we had to break up. */
256 unsigned int is_artificial_var : 1;
258 /* True if this is a special variable whose solution set should not be
260 unsigned int is_special_var : 1;
262 /* True for variables whose size is not known or variable. */
263 unsigned int is_unknown_size_var : 1;
265 /* True for (sub-)fields that represent a whole variable. */
266 unsigned int is_full_var : 1;
268 /* True if this is a heap variable. */
269 unsigned int is_heap_var : 1;
271 /* True if this is a variable tracking a restrict pointer source. */
272 unsigned int is_restrict_var : 1;
274 /* True if this field may contain pointers. */
275 unsigned int may_have_pointers : 1;
277 /* True if this field has only restrict qualified pointers. */
278 unsigned int only_restrict_pointers : 1;
280 /* True if this represents a global variable. */
281 unsigned int is_global_var : 1;
283 /* True if this represents a IPA function info. */
284 unsigned int is_fn_info : 1;
286 /* A link to the variable for the next field in this structure. */
287 struct variable_info *next;
289 /* Offset of this variable, in bits, from the base variable */
290 unsigned HOST_WIDE_INT offset;
292 /* Size of the variable, in bits. */
293 unsigned HOST_WIDE_INT size;
295 /* Full size of the base variable, in bits. */
296 unsigned HOST_WIDE_INT fullsize;
298 /* Name of this variable */
301 /* Tree that this variable is associated with. */
304 /* Points-to set for this variable. */
307 /* Old points-to set for this variable. */
310 typedef struct variable_info *varinfo_t;
312 static varinfo_t first_vi_for_offset (varinfo_t, unsigned HOST_WIDE_INT);
313 static varinfo_t first_or_preceding_vi_for_offset (varinfo_t,
314 unsigned HOST_WIDE_INT);
315 static varinfo_t lookup_vi_for_tree (tree);
317 /* Pool of variable info structures. */
318 static alloc_pool variable_info_pool;
320 DEF_VEC_P(varinfo_t);
322 DEF_VEC_ALLOC_P(varinfo_t, heap);
324 /* Table of variable info structures for constraint variables.
325 Indexed directly by variable info id. */
326 static VEC(varinfo_t,heap) *varmap;
328 /* Return the varmap element N */
330 static inline varinfo_t
331 get_varinfo (unsigned int n)
333 return VEC_index (varinfo_t, varmap, n);
336 /* Static IDs for the special variables. */
337 enum { nothing_id = 0, anything_id = 1, readonly_id = 2,
338 escaped_id = 3, nonlocal_id = 4,
339 storedanything_id = 5, integer_id = 6 };
341 struct GTY(()) heapvar_map {
343 unsigned HOST_WIDE_INT offset;
347 heapvar_map_eq (const void *p1, const void *p2)
349 const struct heapvar_map *h1 = (const struct heapvar_map *)p1;
350 const struct heapvar_map *h2 = (const struct heapvar_map *)p2;
351 return (h1->map.base.from == h2->map.base.from
352 && h1->offset == h2->offset);
356 heapvar_map_hash (struct heapvar_map *h)
358 return iterative_hash_host_wide_int (h->offset,
359 htab_hash_pointer (h->map.base.from));
362 /* Lookup a heap var for FROM, and return it if we find one. */
365 heapvar_lookup (tree from, unsigned HOST_WIDE_INT offset)
367 struct heapvar_map *h, in;
368 in.map.base.from = from;
370 h = (struct heapvar_map *) htab_find_with_hash (heapvar_for_stmt, &in,
371 heapvar_map_hash (&in));
377 /* Insert a mapping FROM->TO in the heap var for statement
381 heapvar_insert (tree from, unsigned HOST_WIDE_INT offset, tree to)
383 struct heapvar_map *h;
386 h = GGC_NEW (struct heapvar_map);
387 h->map.base.from = from;
389 h->map.hash = heapvar_map_hash (h);
391 loc = htab_find_slot_with_hash (heapvar_for_stmt, h, h->map.hash, INSERT);
392 gcc_assert (*loc == NULL);
393 *(struct heapvar_map **) loc = h;
396 /* Return a new variable info structure consisting for a variable
397 named NAME, and using constraint graph node NODE. Append it
398 to the vector of variable info structures. */
401 new_var_info (tree t, const char *name)
403 unsigned index = VEC_length (varinfo_t, varmap);
404 varinfo_t ret = (varinfo_t) pool_alloc (variable_info_pool);
409 /* Vars without decl are artificial and do not have sub-variables. */
410 ret->is_artificial_var = (t == NULL_TREE);
411 ret->is_special_var = false;
412 ret->is_unknown_size_var = false;
413 ret->is_full_var = (t == NULL_TREE);
414 ret->is_heap_var = false;
415 ret->is_restrict_var = false;
416 ret->may_have_pointers = true;
417 ret->only_restrict_pointers = false;
418 ret->is_global_var = (t == NULL_TREE);
419 ret->is_fn_info = false;
421 ret->is_global_var = is_global_var (t);
422 ret->solution = BITMAP_ALLOC (&pta_obstack);
423 ret->oldsolution = BITMAP_ALLOC (&oldpta_obstack);
428 VEC_safe_push (varinfo_t, heap, varmap, ret);
434 /* A map mapping call statements to per-stmt variables for uses
435 and clobbers specific to the call. */
436 struct pointer_map_t *call_stmt_vars;
438 /* Lookup or create the variable for the call statement CALL. */
441 get_call_vi (gimple call)
446 slot_p = pointer_map_insert (call_stmt_vars, call);
448 return (varinfo_t) *slot_p;
450 vi = new_var_info (NULL_TREE, "CALLUSED");
454 vi->is_full_var = true;
456 vi->next = vi2 = new_var_info (NULL_TREE, "CALLCLOBBERED");
460 vi2->is_full_var = true;
462 *slot_p = (void *) vi;
466 /* Lookup the variable for the call statement CALL representing
467 the uses. Returns NULL if there is nothing special about this call. */
470 lookup_call_use_vi (gimple call)
474 slot_p = pointer_map_contains (call_stmt_vars, call);
476 return (varinfo_t) *slot_p;
481 /* Lookup the variable for the call statement CALL representing
482 the clobbers. Returns NULL if there is nothing special about this call. */
485 lookup_call_clobber_vi (gimple call)
487 varinfo_t uses = lookup_call_use_vi (call);
494 /* Lookup or create the variable for the call statement CALL representing
498 get_call_use_vi (gimple call)
500 return get_call_vi (call);
503 /* Lookup or create the variable for the call statement CALL representing
506 static varinfo_t ATTRIBUTE_UNUSED
507 get_call_clobber_vi (gimple call)
509 return get_call_vi (call)->next;
513 typedef enum {SCALAR, DEREF, ADDRESSOF} constraint_expr_type;
515 /* An expression that appears in a constraint. */
517 struct constraint_expr
519 /* Constraint type. */
520 constraint_expr_type type;
522 /* Variable we are referring to in the constraint. */
525 /* Offset, in bits, of this constraint from the beginning of
526 variables it ends up referring to.
528 IOW, in a deref constraint, we would deref, get the result set,
529 then add OFFSET to each member. */
530 HOST_WIDE_INT offset;
533 /* Use 0x8000... as special unknown offset. */
534 #define UNKNOWN_OFFSET ((HOST_WIDE_INT)-1 << (HOST_BITS_PER_WIDE_INT-1))
536 typedef struct constraint_expr ce_s;
538 DEF_VEC_ALLOC_O(ce_s, heap);
539 static void get_constraint_for_1 (tree, VEC(ce_s, heap) **, bool);
540 static void get_constraint_for (tree, VEC(ce_s, heap) **);
541 static void do_deref (VEC (ce_s, heap) **);
543 /* Our set constraints are made up of two constraint expressions, one
546 As described in the introduction, our set constraints each represent an
547 operation between set valued variables.
551 struct constraint_expr lhs;
552 struct constraint_expr rhs;
555 /* List of constraints that we use to build the constraint graph from. */
557 static VEC(constraint_t,heap) *constraints;
558 static alloc_pool constraint_pool;
560 /* The constraint graph is represented as an array of bitmaps
561 containing successor nodes. */
563 struct constraint_graph
565 /* Size of this graph, which may be different than the number of
566 nodes in the variable map. */
569 /* Explicit successors of each node. */
572 /* Implicit predecessors of each node (Used for variable
574 bitmap *implicit_preds;
576 /* Explicit predecessors of each node (Used for variable substitution). */
579 /* Indirect cycle representatives, or -1 if the node has no indirect
581 int *indirect_cycles;
583 /* Representative node for a node. rep[a] == a unless the node has
587 /* Equivalence class representative for a label. This is used for
588 variable substitution. */
591 /* Pointer equivalence label for a node. All nodes with the same
592 pointer equivalence label can be unified together at some point
593 (either during constraint optimization or after the constraint
597 /* Pointer equivalence representative for a label. This is used to
598 handle nodes that are pointer equivalent but not location
599 equivalent. We can unite these once the addressof constraints
600 are transformed into initial points-to sets. */
603 /* Pointer equivalence label for each node, used during variable
605 unsigned int *pointer_label;
607 /* Location equivalence label for each node, used during location
608 equivalence finding. */
609 unsigned int *loc_label;
611 /* Pointed-by set for each node, used during location equivalence
612 finding. This is pointed-by rather than pointed-to, because it
613 is constructed using the predecessor graph. */
616 /* Points to sets for pointer equivalence. This is *not* the actual
617 points-to sets for nodes. */
620 /* Bitmap of nodes where the bit is set if the node is a direct
621 node. Used for variable substitution. */
622 sbitmap direct_nodes;
624 /* Bitmap of nodes where the bit is set if the node is address
625 taken. Used for variable substitution. */
626 bitmap address_taken;
628 /* Vector of complex constraints for each graph node. Complex
629 constraints are those involving dereferences or offsets that are
631 VEC(constraint_t,heap) **complex;
634 static constraint_graph_t graph;
636 /* During variable substitution and the offline version of indirect
637 cycle finding, we create nodes to represent dereferences and
638 address taken constraints. These represent where these start and
640 #define FIRST_REF_NODE (VEC_length (varinfo_t, varmap))
641 #define LAST_REF_NODE (FIRST_REF_NODE + (FIRST_REF_NODE - 1))
643 /* Return the representative node for NODE, if NODE has been unioned
645 This function performs path compression along the way to finding
646 the representative. */
649 find (unsigned int node)
651 gcc_assert (node < graph->size);
652 if (graph->rep[node] != node)
653 return graph->rep[node] = find (graph->rep[node]);
657 /* Union the TO and FROM nodes to the TO nodes.
658 Note that at some point in the future, we may want to do
659 union-by-rank, in which case we are going to have to return the
660 node we unified to. */
663 unite (unsigned int to, unsigned int from)
665 gcc_assert (to < graph->size && from < graph->size);
666 if (to != from && graph->rep[from] != to)
668 graph->rep[from] = to;
674 /* Create a new constraint consisting of LHS and RHS expressions. */
677 new_constraint (const struct constraint_expr lhs,
678 const struct constraint_expr rhs)
680 constraint_t ret = (constraint_t) pool_alloc (constraint_pool);
686 /* Print out constraint C to FILE. */
689 dump_constraint (FILE *file, constraint_t c)
691 if (c->lhs.type == ADDRESSOF)
693 else if (c->lhs.type == DEREF)
695 fprintf (file, "%s", get_varinfo (c->lhs.var)->name);
696 if (c->lhs.offset == UNKNOWN_OFFSET)
697 fprintf (file, " + UNKNOWN");
698 else if (c->lhs.offset != 0)
699 fprintf (file, " + " HOST_WIDE_INT_PRINT_DEC, c->lhs.offset);
700 fprintf (file, " = ");
701 if (c->rhs.type == ADDRESSOF)
703 else if (c->rhs.type == DEREF)
705 fprintf (file, "%s", get_varinfo (c->rhs.var)->name);
706 if (c->rhs.offset == UNKNOWN_OFFSET)
707 fprintf (file, " + UNKNOWN");
708 else if (c->rhs.offset != 0)
709 fprintf (file, " + " HOST_WIDE_INT_PRINT_DEC, c->rhs.offset);
710 fprintf (file, "\n");
714 void debug_constraint (constraint_t);
715 void debug_constraints (void);
716 void debug_constraint_graph (void);
717 void debug_solution_for_var (unsigned int);
718 void debug_sa_points_to_info (void);
720 /* Print out constraint C to stderr. */
723 debug_constraint (constraint_t c)
725 dump_constraint (stderr, c);
728 /* Print out all constraints to FILE */
731 dump_constraints (FILE *file, int from)
735 for (i = from; VEC_iterate (constraint_t, constraints, i, c); i++)
736 dump_constraint (file, c);
739 /* Print out all constraints to stderr. */
742 debug_constraints (void)
744 dump_constraints (stderr, 0);
747 /* Print out to FILE the edge in the constraint graph that is created by
748 constraint c. The edge may have a label, depending on the type of
749 constraint that it represents. If complex1, e.g: a = *b, then the label
750 is "=*", if complex2, e.g: *a = b, then the label is "*=", if
751 complex with an offset, e.g: a = b + 8, then the label is "+".
752 Otherwise the edge has no label. */
755 dump_constraint_edge (FILE *file, constraint_t c)
757 if (c->rhs.type != ADDRESSOF)
759 const char *src = get_varinfo (c->rhs.var)->name;
760 const char *dst = get_varinfo (c->lhs.var)->name;
761 fprintf (file, " \"%s\" -> \"%s\" ", src, dst);
762 /* Due to preprocessing of constraints, instructions like *a = *b are
763 illegal; thus, we do not have to handle such cases. */
764 if (c->lhs.type == DEREF)
765 fprintf (file, " [ label=\"*=\" ] ;\n");
766 else if (c->rhs.type == DEREF)
767 fprintf (file, " [ label=\"=*\" ] ;\n");
770 /* We must check the case where the constraint is an offset.
771 In this case, it is treated as a complex constraint. */
772 if (c->rhs.offset != c->lhs.offset)
773 fprintf (file, " [ label=\"+\" ] ;\n");
775 fprintf (file, " ;\n");
780 /* Print the constraint graph in dot format. */
783 dump_constraint_graph (FILE *file)
785 unsigned int i=0, size;
788 /* Only print the graph if it has already been initialized: */
792 /* Print the constraints used to produce the constraint graph. The
793 constraints will be printed as comments in the dot file: */
794 fprintf (file, "\n\n/* Constraints used in the constraint graph:\n");
795 dump_constraints (file, 0);
796 fprintf (file, "*/\n");
798 /* Prints the header of the dot file: */
799 fprintf (file, "\n\n// The constraint graph in dot format:\n");
800 fprintf (file, "strict digraph {\n");
801 fprintf (file, " node [\n shape = box\n ]\n");
802 fprintf (file, " edge [\n fontsize = \"12\"\n ]\n");
803 fprintf (file, "\n // List of nodes in the constraint graph:\n");
805 /* The next lines print the nodes in the graph. In order to get the
806 number of nodes in the graph, we must choose the minimum between the
807 vector VEC (varinfo_t, varmap) and graph->size. If the graph has not
808 yet been initialized, then graph->size == 0, otherwise we must only
809 read nodes that have an entry in VEC (varinfo_t, varmap). */
810 size = VEC_length (varinfo_t, varmap);
811 size = size < graph->size ? size : graph->size;
812 for (i = 0; i < size; i++)
814 const char *name = get_varinfo (graph->rep[i])->name;
815 fprintf (file, " \"%s\" ;\n", name);
818 /* Go over the list of constraints printing the edges in the constraint
820 fprintf (file, "\n // The constraint edges:\n");
821 for (i = 0; VEC_iterate (constraint_t, constraints, i, c); i++)
823 dump_constraint_edge (file, c);
825 /* Prints the tail of the dot file. By now, only the closing bracket. */
826 fprintf (file, "}\n\n\n");
829 /* Print out the constraint graph to stderr. */
832 debug_constraint_graph (void)
834 dump_constraint_graph (stderr);
839 The solver is a simple worklist solver, that works on the following
842 sbitmap changed_nodes = all zeroes;
844 For each node that is not already collapsed:
846 set bit in changed nodes
848 while (changed_count > 0)
850 compute topological ordering for constraint graph
852 find and collapse cycles in the constraint graph (updating
853 changed if necessary)
855 for each node (n) in the graph in topological order:
858 Process each complex constraint associated with the node,
859 updating changed if necessary.
861 For each outgoing edge from n, propagate the solution from n to
862 the destination of the edge, updating changed as necessary.
866 /* Return true if two constraint expressions A and B are equal. */
869 constraint_expr_equal (struct constraint_expr a, struct constraint_expr b)
871 return a.type == b.type && a.var == b.var && a.offset == b.offset;
874 /* Return true if constraint expression A is less than constraint expression
875 B. This is just arbitrary, but consistent, in order to give them an
879 constraint_expr_less (struct constraint_expr a, struct constraint_expr b)
881 if (a.type == b.type)
884 return a.offset < b.offset;
886 return a.var < b.var;
889 return a.type < b.type;
892 /* Return true if constraint A is less than constraint B. This is just
893 arbitrary, but consistent, in order to give them an ordering. */
896 constraint_less (const constraint_t a, const constraint_t b)
898 if (constraint_expr_less (a->lhs, b->lhs))
900 else if (constraint_expr_less (b->lhs, a->lhs))
903 return constraint_expr_less (a->rhs, b->rhs);
906 /* Return true if two constraints A and B are equal. */
909 constraint_equal (struct constraint a, struct constraint b)
911 return constraint_expr_equal (a.lhs, b.lhs)
912 && constraint_expr_equal (a.rhs, b.rhs);
916 /* Find a constraint LOOKFOR in the sorted constraint vector VEC */
919 constraint_vec_find (VEC(constraint_t,heap) *vec,
920 struct constraint lookfor)
928 place = VEC_lower_bound (constraint_t, vec, &lookfor, constraint_less);
929 if (place >= VEC_length (constraint_t, vec))
931 found = VEC_index (constraint_t, vec, place);
932 if (!constraint_equal (*found, lookfor))
937 /* Union two constraint vectors, TO and FROM. Put the result in TO. */
940 constraint_set_union (VEC(constraint_t,heap) **to,
941 VEC(constraint_t,heap) **from)
946 for (i = 0; VEC_iterate (constraint_t, *from, i, c); i++)
948 if (constraint_vec_find (*to, *c) == NULL)
950 unsigned int place = VEC_lower_bound (constraint_t, *to, c,
952 VEC_safe_insert (constraint_t, heap, *to, place, c);
957 /* Expands the solution in SET to all sub-fields of variables included.
958 Union the expanded result into RESULT. */
961 solution_set_expand (bitmap result, bitmap set)
967 /* In a first pass record all variables we need to add all
968 sub-fields off. This avoids quadratic behavior. */
969 EXECUTE_IF_SET_IN_BITMAP (set, 0, j, bi)
971 varinfo_t v = get_varinfo (j);
972 if (v->is_artificial_var
975 v = lookup_vi_for_tree (v->decl);
977 vars = BITMAP_ALLOC (NULL);
978 bitmap_set_bit (vars, v->id);
981 /* In the second pass now do the addition to the solution and
982 to speed up solving add it to the delta as well. */
985 EXECUTE_IF_SET_IN_BITMAP (vars, 0, j, bi)
987 varinfo_t v = get_varinfo (j);
988 for (; v != NULL; v = v->next)
989 bitmap_set_bit (result, v->id);
995 /* Take a solution set SET, add OFFSET to each member of the set, and
996 overwrite SET with the result when done. */
999 solution_set_add (bitmap set, HOST_WIDE_INT offset)
1001 bitmap result = BITMAP_ALLOC (&iteration_obstack);
1005 /* If the offset is unknown we have to expand the solution to
1007 if (offset == UNKNOWN_OFFSET)
1009 solution_set_expand (set, set);
1013 EXECUTE_IF_SET_IN_BITMAP (set, 0, i, bi)
1015 varinfo_t vi = get_varinfo (i);
1017 /* If this is a variable with just one field just set its bit
1019 if (vi->is_artificial_var
1020 || vi->is_unknown_size_var
1022 bitmap_set_bit (result, i);
1025 unsigned HOST_WIDE_INT fieldoffset = vi->offset + offset;
1027 /* If the offset makes the pointer point to before the
1028 variable use offset zero for the field lookup. */
1030 && fieldoffset > vi->offset)
1034 vi = first_or_preceding_vi_for_offset (vi, fieldoffset);
1036 bitmap_set_bit (result, vi->id);
1037 /* If the result is not exactly at fieldoffset include the next
1038 field as well. See get_constraint_for_ptr_offset for more
1040 if (vi->offset != fieldoffset
1041 && vi->next != NULL)
1042 bitmap_set_bit (result, vi->next->id);
1046 bitmap_copy (set, result);
1047 BITMAP_FREE (result);
1050 /* Union solution sets TO and FROM, and add INC to each member of FROM in the
1054 set_union_with_increment (bitmap to, bitmap from, HOST_WIDE_INT inc)
1057 return bitmap_ior_into (to, from);
1063 tmp = BITMAP_ALLOC (&iteration_obstack);
1064 bitmap_copy (tmp, from);
1065 solution_set_add (tmp, inc);
1066 res = bitmap_ior_into (to, tmp);
1072 /* Insert constraint C into the list of complex constraints for graph
1076 insert_into_complex (constraint_graph_t graph,
1077 unsigned int var, constraint_t c)
1079 VEC (constraint_t, heap) *complex = graph->complex[var];
1080 unsigned int place = VEC_lower_bound (constraint_t, complex, c,
1083 /* Only insert constraints that do not already exist. */
1084 if (place >= VEC_length (constraint_t, complex)
1085 || !constraint_equal (*c, *VEC_index (constraint_t, complex, place)))
1086 VEC_safe_insert (constraint_t, heap, graph->complex[var], place, c);
1090 /* Condense two variable nodes into a single variable node, by moving
1091 all associated info from SRC to TO. */
1094 merge_node_constraints (constraint_graph_t graph, unsigned int to,
1100 gcc_assert (find (from) == to);
1102 /* Move all complex constraints from src node into to node */
1103 for (i = 0; VEC_iterate (constraint_t, graph->complex[from], i, c); i++)
1105 /* In complex constraints for node src, we may have either
1106 a = *src, and *src = a, or an offseted constraint which are
1107 always added to the rhs node's constraints. */
1109 if (c->rhs.type == DEREF)
1111 else if (c->lhs.type == DEREF)
1116 constraint_set_union (&graph->complex[to], &graph->complex[from]);
1117 VEC_free (constraint_t, heap, graph->complex[from]);
1118 graph->complex[from] = NULL;
1122 /* Remove edges involving NODE from GRAPH. */
1125 clear_edges_for_node (constraint_graph_t graph, unsigned int node)
1127 if (graph->succs[node])
1128 BITMAP_FREE (graph->succs[node]);
1131 /* Merge GRAPH nodes FROM and TO into node TO. */
1134 merge_graph_nodes (constraint_graph_t graph, unsigned int to,
1137 if (graph->indirect_cycles[from] != -1)
1139 /* If we have indirect cycles with the from node, and we have
1140 none on the to node, the to node has indirect cycles from the
1141 from node now that they are unified.
1142 If indirect cycles exist on both, unify the nodes that they
1143 are in a cycle with, since we know they are in a cycle with
1145 if (graph->indirect_cycles[to] == -1)
1146 graph->indirect_cycles[to] = graph->indirect_cycles[from];
1149 /* Merge all the successor edges. */
1150 if (graph->succs[from])
1152 if (!graph->succs[to])
1153 graph->succs[to] = BITMAP_ALLOC (&pta_obstack);
1154 bitmap_ior_into (graph->succs[to],
1155 graph->succs[from]);
1158 clear_edges_for_node (graph, from);
1162 /* Add an indirect graph edge to GRAPH, going from TO to FROM if
1163 it doesn't exist in the graph already. */
1166 add_implicit_graph_edge (constraint_graph_t graph, unsigned int to,
1172 if (!graph->implicit_preds[to])
1173 graph->implicit_preds[to] = BITMAP_ALLOC (&predbitmap_obstack);
1175 if (bitmap_set_bit (graph->implicit_preds[to], from))
1176 stats.num_implicit_edges++;
1179 /* Add a predecessor graph edge to GRAPH, going from TO to FROM if
1180 it doesn't exist in the graph already.
1181 Return false if the edge already existed, true otherwise. */
1184 add_pred_graph_edge (constraint_graph_t graph, unsigned int to,
1187 if (!graph->preds[to])
1188 graph->preds[to] = BITMAP_ALLOC (&predbitmap_obstack);
1189 bitmap_set_bit (graph->preds[to], from);
1192 /* Add a graph edge to GRAPH, going from FROM to TO if
1193 it doesn't exist in the graph already.
1194 Return false if the edge already existed, true otherwise. */
1197 add_graph_edge (constraint_graph_t graph, unsigned int to,
1208 if (!graph->succs[from])
1209 graph->succs[from] = BITMAP_ALLOC (&pta_obstack);
1210 if (bitmap_set_bit (graph->succs[from], to))
1213 if (to < FIRST_REF_NODE && from < FIRST_REF_NODE)
1221 /* Return true if {DEST.SRC} is an existing graph edge in GRAPH. */
1224 valid_graph_edge (constraint_graph_t graph, unsigned int src,
1227 return (graph->succs[dest]
1228 && bitmap_bit_p (graph->succs[dest], src));
1231 /* Initialize the constraint graph structure to contain SIZE nodes. */
1234 init_graph (unsigned int size)
1238 graph = XCNEW (struct constraint_graph);
1240 graph->succs = XCNEWVEC (bitmap, graph->size);
1241 graph->indirect_cycles = XNEWVEC (int, graph->size);
1242 graph->rep = XNEWVEC (unsigned int, graph->size);
1243 graph->complex = XCNEWVEC (VEC(constraint_t, heap) *, size);
1244 graph->pe = XCNEWVEC (unsigned int, graph->size);
1245 graph->pe_rep = XNEWVEC (int, graph->size);
1247 for (j = 0; j < graph->size; j++)
1250 graph->pe_rep[j] = -1;
1251 graph->indirect_cycles[j] = -1;
1255 /* Build the constraint graph, adding only predecessor edges right now. */
1258 build_pred_graph (void)
1264 graph->implicit_preds = XCNEWVEC (bitmap, graph->size);
1265 graph->preds = XCNEWVEC (bitmap, graph->size);
1266 graph->pointer_label = XCNEWVEC (unsigned int, graph->size);
1267 graph->loc_label = XCNEWVEC (unsigned int, graph->size);
1268 graph->pointed_by = XCNEWVEC (bitmap, graph->size);
1269 graph->points_to = XCNEWVEC (bitmap, graph->size);
1270 graph->eq_rep = XNEWVEC (int, graph->size);
1271 graph->direct_nodes = sbitmap_alloc (graph->size);
1272 graph->address_taken = BITMAP_ALLOC (&predbitmap_obstack);
1273 sbitmap_zero (graph->direct_nodes);
1275 for (j = 0; j < FIRST_REF_NODE; j++)
1277 if (!get_varinfo (j)->is_special_var)
1278 SET_BIT (graph->direct_nodes, j);
1281 for (j = 0; j < graph->size; j++)
1282 graph->eq_rep[j] = -1;
1284 for (j = 0; j < VEC_length (varinfo_t, varmap); j++)
1285 graph->indirect_cycles[j] = -1;
1287 for (i = 0; VEC_iterate (constraint_t, constraints, i, c); i++)
1289 struct constraint_expr lhs = c->lhs;
1290 struct constraint_expr rhs = c->rhs;
1291 unsigned int lhsvar = lhs.var;
1292 unsigned int rhsvar = rhs.var;
1294 if (lhs.type == DEREF)
1297 if (rhs.offset == 0 && lhs.offset == 0 && rhs.type == SCALAR)
1298 add_pred_graph_edge (graph, FIRST_REF_NODE + lhsvar, rhsvar);
1300 else if (rhs.type == DEREF)
1303 if (rhs.offset == 0 && lhs.offset == 0 && lhs.type == SCALAR)
1304 add_pred_graph_edge (graph, lhsvar, FIRST_REF_NODE + rhsvar);
1306 RESET_BIT (graph->direct_nodes, lhsvar);
1308 else if (rhs.type == ADDRESSOF)
1313 if (graph->points_to[lhsvar] == NULL)
1314 graph->points_to[lhsvar] = BITMAP_ALLOC (&predbitmap_obstack);
1315 bitmap_set_bit (graph->points_to[lhsvar], rhsvar);
1317 if (graph->pointed_by[rhsvar] == NULL)
1318 graph->pointed_by[rhsvar] = BITMAP_ALLOC (&predbitmap_obstack);
1319 bitmap_set_bit (graph->pointed_by[rhsvar], lhsvar);
1321 /* Implicitly, *x = y */
1322 add_implicit_graph_edge (graph, FIRST_REF_NODE + lhsvar, rhsvar);
1324 /* All related variables are no longer direct nodes. */
1325 RESET_BIT (graph->direct_nodes, rhsvar);
1326 v = get_varinfo (rhsvar);
1327 if (!v->is_full_var)
1329 v = lookup_vi_for_tree (v->decl);
1332 RESET_BIT (graph->direct_nodes, v->id);
1337 bitmap_set_bit (graph->address_taken, rhsvar);
1339 else if (lhsvar > anything_id
1340 && lhsvar != rhsvar && lhs.offset == 0 && rhs.offset == 0)
1343 add_pred_graph_edge (graph, lhsvar, rhsvar);
1344 /* Implicitly, *x = *y */
1345 add_implicit_graph_edge (graph, FIRST_REF_NODE + lhsvar,
1346 FIRST_REF_NODE + rhsvar);
1348 else if (lhs.offset != 0 || rhs.offset != 0)
1350 if (rhs.offset != 0)
1351 RESET_BIT (graph->direct_nodes, lhs.var);
1352 else if (lhs.offset != 0)
1353 RESET_BIT (graph->direct_nodes, rhs.var);
1358 /* Build the constraint graph, adding successor edges. */
1361 build_succ_graph (void)
1366 for (i = 0; VEC_iterate (constraint_t, constraints, i, c); i++)
1368 struct constraint_expr lhs;
1369 struct constraint_expr rhs;
1370 unsigned int lhsvar;
1371 unsigned int rhsvar;
1378 lhsvar = find (lhs.var);
1379 rhsvar = find (rhs.var);
1381 if (lhs.type == DEREF)
1383 if (rhs.offset == 0 && lhs.offset == 0 && rhs.type == SCALAR)
1384 add_graph_edge (graph, FIRST_REF_NODE + lhsvar, rhsvar);
1386 else if (rhs.type == DEREF)
1388 if (rhs.offset == 0 && lhs.offset == 0 && lhs.type == SCALAR)
1389 add_graph_edge (graph, lhsvar, FIRST_REF_NODE + rhsvar);
1391 else if (rhs.type == ADDRESSOF)
1394 gcc_assert (find (rhs.var) == rhs.var);
1395 bitmap_set_bit (get_varinfo (lhsvar)->solution, rhsvar);
1397 else if (lhsvar > anything_id
1398 && lhsvar != rhsvar && lhs.offset == 0 && rhs.offset == 0)
1400 add_graph_edge (graph, lhsvar, rhsvar);
1404 /* Add edges from STOREDANYTHING to all non-direct nodes that can
1405 receive pointers. */
1406 t = find (storedanything_id);
1407 for (i = integer_id + 1; i < FIRST_REF_NODE; ++i)
1409 if (!TEST_BIT (graph->direct_nodes, i)
1410 && get_varinfo (i)->may_have_pointers)
1411 add_graph_edge (graph, find (i), t);
1414 /* Everything stored to ANYTHING also potentially escapes. */
1415 add_graph_edge (graph, find (escaped_id), t);
1419 /* Changed variables on the last iteration. */
1420 static unsigned int changed_count;
1421 static sbitmap changed;
1423 /* Strongly Connected Component visitation info. */
1430 unsigned int *node_mapping;
1432 VEC(unsigned,heap) *scc_stack;
1436 /* Recursive routine to find strongly connected components in GRAPH.
1437 SI is the SCC info to store the information in, and N is the id of current
1438 graph node we are processing.
1440 This is Tarjan's strongly connected component finding algorithm, as
1441 modified by Nuutila to keep only non-root nodes on the stack.
1442 The algorithm can be found in "On finding the strongly connected
1443 connected components in a directed graph" by Esko Nuutila and Eljas
1444 Soisalon-Soininen, in Information Processing Letters volume 49,
1445 number 1, pages 9-14. */
1448 scc_visit (constraint_graph_t graph, struct scc_info *si, unsigned int n)
1452 unsigned int my_dfs;
1454 SET_BIT (si->visited, n);
1455 si->dfs[n] = si->current_index ++;
1456 my_dfs = si->dfs[n];
1458 /* Visit all the successors. */
1459 EXECUTE_IF_IN_NONNULL_BITMAP (graph->succs[n], 0, i, bi)
1463 if (i > LAST_REF_NODE)
1467 if (TEST_BIT (si->deleted, w))
1470 if (!TEST_BIT (si->visited, w))
1471 scc_visit (graph, si, w);
1473 unsigned int t = find (w);
1474 unsigned int nnode = find (n);
1475 gcc_assert (nnode == n);
1477 if (si->dfs[t] < si->dfs[nnode])
1478 si->dfs[n] = si->dfs[t];
1482 /* See if any components have been identified. */
1483 if (si->dfs[n] == my_dfs)
1485 if (VEC_length (unsigned, si->scc_stack) > 0
1486 && si->dfs[VEC_last (unsigned, si->scc_stack)] >= my_dfs)
1488 bitmap scc = BITMAP_ALLOC (NULL);
1489 unsigned int lowest_node;
1492 bitmap_set_bit (scc, n);
1494 while (VEC_length (unsigned, si->scc_stack) != 0
1495 && si->dfs[VEC_last (unsigned, si->scc_stack)] >= my_dfs)
1497 unsigned int w = VEC_pop (unsigned, si->scc_stack);
1499 bitmap_set_bit (scc, w);
1502 lowest_node = bitmap_first_set_bit (scc);
1503 gcc_assert (lowest_node < FIRST_REF_NODE);
1505 /* Collapse the SCC nodes into a single node, and mark the
1507 EXECUTE_IF_SET_IN_BITMAP (scc, 0, i, bi)
1509 if (i < FIRST_REF_NODE)
1511 if (unite (lowest_node, i))
1512 unify_nodes (graph, lowest_node, i, false);
1516 unite (lowest_node, i);
1517 graph->indirect_cycles[i - FIRST_REF_NODE] = lowest_node;
1521 SET_BIT (si->deleted, n);
1524 VEC_safe_push (unsigned, heap, si->scc_stack, n);
1527 /* Unify node FROM into node TO, updating the changed count if
1528 necessary when UPDATE_CHANGED is true. */
1531 unify_nodes (constraint_graph_t graph, unsigned int to, unsigned int from,
1532 bool update_changed)
1535 gcc_assert (to != from && find (to) == to);
1536 if (dump_file && (dump_flags & TDF_DETAILS))
1537 fprintf (dump_file, "Unifying %s to %s\n",
1538 get_varinfo (from)->name,
1539 get_varinfo (to)->name);
1542 stats.unified_vars_dynamic++;
1544 stats.unified_vars_static++;
1546 merge_graph_nodes (graph, to, from);
1547 merge_node_constraints (graph, to, from);
1549 /* Mark TO as changed if FROM was changed. If TO was already marked
1550 as changed, decrease the changed count. */
1552 if (update_changed && TEST_BIT (changed, from))
1554 RESET_BIT (changed, from);
1555 if (!TEST_BIT (changed, to))
1556 SET_BIT (changed, to);
1559 gcc_assert (changed_count > 0);
1563 if (get_varinfo (from)->solution)
1565 /* If the solution changes because of the merging, we need to mark
1566 the variable as changed. */
1567 if (bitmap_ior_into (get_varinfo (to)->solution,
1568 get_varinfo (from)->solution))
1570 if (update_changed && !TEST_BIT (changed, to))
1572 SET_BIT (changed, to);
1577 BITMAP_FREE (get_varinfo (from)->solution);
1578 BITMAP_FREE (get_varinfo (from)->oldsolution);
1580 if (stats.iterations > 0)
1582 BITMAP_FREE (get_varinfo (to)->oldsolution);
1583 get_varinfo (to)->oldsolution = BITMAP_ALLOC (&oldpta_obstack);
1586 if (valid_graph_edge (graph, to, to))
1588 if (graph->succs[to])
1589 bitmap_clear_bit (graph->succs[to], to);
1593 /* Information needed to compute the topological ordering of a graph. */
1597 /* sbitmap of visited nodes. */
1599 /* Array that stores the topological order of the graph, *in
1601 VEC(unsigned,heap) *topo_order;
1605 /* Initialize and return a topological info structure. */
1607 static struct topo_info *
1608 init_topo_info (void)
1610 size_t size = graph->size;
1611 struct topo_info *ti = XNEW (struct topo_info);
1612 ti->visited = sbitmap_alloc (size);
1613 sbitmap_zero (ti->visited);
1614 ti->topo_order = VEC_alloc (unsigned, heap, 1);
1619 /* Free the topological sort info pointed to by TI. */
1622 free_topo_info (struct topo_info *ti)
1624 sbitmap_free (ti->visited);
1625 VEC_free (unsigned, heap, ti->topo_order);
1629 /* Visit the graph in topological order, and store the order in the
1630 topo_info structure. */
1633 topo_visit (constraint_graph_t graph, struct topo_info *ti,
1639 SET_BIT (ti->visited, n);
1641 if (graph->succs[n])
1642 EXECUTE_IF_SET_IN_BITMAP (graph->succs[n], 0, j, bi)
1644 if (!TEST_BIT (ti->visited, j))
1645 topo_visit (graph, ti, j);
1648 VEC_safe_push (unsigned, heap, ti->topo_order, n);
1651 /* Process a constraint C that represents x = *(y + off), using DELTA as the
1652 starting solution for y. */
1655 do_sd_constraint (constraint_graph_t graph, constraint_t c,
1658 unsigned int lhs = c->lhs.var;
1660 bitmap sol = get_varinfo (lhs)->solution;
1663 HOST_WIDE_INT roffset = c->rhs.offset;
1665 /* Our IL does not allow this. */
1666 gcc_assert (c->lhs.offset == 0);
1668 /* If the solution of Y contains anything it is good enough to transfer
1670 if (bitmap_bit_p (delta, anything_id))
1672 flag |= bitmap_set_bit (sol, anything_id);
1676 /* If we do not know at with offset the rhs is dereferenced compute
1677 the reachability set of DELTA, conservatively assuming it is
1678 dereferenced at all valid offsets. */
1679 if (roffset == UNKNOWN_OFFSET)
1681 solution_set_expand (delta, delta);
1682 /* No further offset processing is necessary. */
1686 /* For each variable j in delta (Sol(y)), add
1687 an edge in the graph from j to x, and union Sol(j) into Sol(x). */
1688 EXECUTE_IF_SET_IN_BITMAP (delta, 0, j, bi)
1690 varinfo_t v = get_varinfo (j);
1691 HOST_WIDE_INT fieldoffset = v->offset + roffset;
1695 fieldoffset = v->offset;
1696 else if (roffset != 0)
1697 v = first_vi_for_offset (v, fieldoffset);
1698 /* If the access is outside of the variable we can ignore it. */
1706 /* Adding edges from the special vars is pointless.
1707 They don't have sets that can change. */
1708 if (get_varinfo (t)->is_special_var)
1709 flag |= bitmap_ior_into (sol, get_varinfo (t)->solution);
1710 /* Merging the solution from ESCAPED needlessly increases
1711 the set. Use ESCAPED as representative instead. */
1712 else if (v->id == escaped_id)
1713 flag |= bitmap_set_bit (sol, escaped_id);
1714 else if (v->may_have_pointers
1715 && add_graph_edge (graph, lhs, t))
1716 flag |= bitmap_ior_into (sol, get_varinfo (t)->solution);
1718 /* If the variable is not exactly at the requested offset
1719 we have to include the next one. */
1720 if (v->offset == (unsigned HOST_WIDE_INT)fieldoffset
1725 fieldoffset = v->offset;
1731 /* If the LHS solution changed, mark the var as changed. */
1734 get_varinfo (lhs)->solution = sol;
1735 if (!TEST_BIT (changed, lhs))
1737 SET_BIT (changed, lhs);
1743 /* Process a constraint C that represents *(x + off) = y using DELTA
1744 as the starting solution for x. */
1747 do_ds_constraint (constraint_t c, bitmap delta)
1749 unsigned int rhs = c->rhs.var;
1750 bitmap sol = get_varinfo (rhs)->solution;
1753 HOST_WIDE_INT loff = c->lhs.offset;
1754 bool escaped_p = false;
1756 /* Our IL does not allow this. */
1757 gcc_assert (c->rhs.offset == 0);
1759 /* If the solution of y contains ANYTHING simply use the ANYTHING
1760 solution. This avoids needlessly increasing the points-to sets. */
1761 if (bitmap_bit_p (sol, anything_id))
1762 sol = get_varinfo (find (anything_id))->solution;
1764 /* If the solution for x contains ANYTHING we have to merge the
1765 solution of y into all pointer variables which we do via
1767 if (bitmap_bit_p (delta, anything_id))
1769 unsigned t = find (storedanything_id);
1770 if (add_graph_edge (graph, t, rhs))
1772 if (bitmap_ior_into (get_varinfo (t)->solution, sol))
1774 if (!TEST_BIT (changed, t))
1776 SET_BIT (changed, t);
1784 /* If we do not know at with offset the rhs is dereferenced compute
1785 the reachability set of DELTA, conservatively assuming it is
1786 dereferenced at all valid offsets. */
1787 if (loff == UNKNOWN_OFFSET)
1789 solution_set_expand (delta, delta);
1793 /* For each member j of delta (Sol(x)), add an edge from y to j and
1794 union Sol(y) into Sol(j) */
1795 EXECUTE_IF_SET_IN_BITMAP (delta, 0, j, bi)
1797 varinfo_t v = get_varinfo (j);
1799 HOST_WIDE_INT fieldoffset = v->offset + loff;
1802 fieldoffset = v->offset;
1804 v = first_vi_for_offset (v, fieldoffset);
1805 /* If the access is outside of the variable we can ignore it. */
1811 if (v->may_have_pointers)
1813 /* If v is a global variable then this is an escape point. */
1814 if (v->is_global_var
1817 t = find (escaped_id);
1818 if (add_graph_edge (graph, t, rhs)
1819 && bitmap_ior_into (get_varinfo (t)->solution, sol)
1820 && !TEST_BIT (changed, t))
1822 SET_BIT (changed, t);
1825 /* Enough to let rhs escape once. */
1829 if (v->is_special_var)
1833 if (add_graph_edge (graph, t, rhs)
1834 && bitmap_ior_into (get_varinfo (t)->solution, sol)
1835 && !TEST_BIT (changed, t))
1837 SET_BIT (changed, t);
1842 /* If the variable is not exactly at the requested offset
1843 we have to include the next one. */
1844 if (v->offset == (unsigned HOST_WIDE_INT)fieldoffset
1849 fieldoffset = v->offset;
1855 /* Handle a non-simple (simple meaning requires no iteration),
1856 constraint (IE *x = &y, x = *y, *x = y, and x = y with offsets involved). */
1859 do_complex_constraint (constraint_graph_t graph, constraint_t c, bitmap delta)
1861 if (c->lhs.type == DEREF)
1863 if (c->rhs.type == ADDRESSOF)
1870 do_ds_constraint (c, delta);
1873 else if (c->rhs.type == DEREF)
1876 if (!(get_varinfo (c->lhs.var)->is_special_var))
1877 do_sd_constraint (graph, c, delta);
1885 gcc_assert (c->rhs.type == SCALAR && c->lhs.type == SCALAR);
1886 solution = get_varinfo (c->rhs.var)->solution;
1887 tmp = get_varinfo (c->lhs.var)->solution;
1889 flag = set_union_with_increment (tmp, solution, c->rhs.offset);
1893 get_varinfo (c->lhs.var)->solution = tmp;
1894 if (!TEST_BIT (changed, c->lhs.var))
1896 SET_BIT (changed, c->lhs.var);
1903 /* Initialize and return a new SCC info structure. */
1905 static struct scc_info *
1906 init_scc_info (size_t size)
1908 struct scc_info *si = XNEW (struct scc_info);
1911 si->current_index = 0;
1912 si->visited = sbitmap_alloc (size);
1913 sbitmap_zero (si->visited);
1914 si->deleted = sbitmap_alloc (size);
1915 sbitmap_zero (si->deleted);
1916 si->node_mapping = XNEWVEC (unsigned int, size);
1917 si->dfs = XCNEWVEC (unsigned int, size);
1919 for (i = 0; i < size; i++)
1920 si->node_mapping[i] = i;
1922 si->scc_stack = VEC_alloc (unsigned, heap, 1);
1926 /* Free an SCC info structure pointed to by SI */
1929 free_scc_info (struct scc_info *si)
1931 sbitmap_free (si->visited);
1932 sbitmap_free (si->deleted);
1933 free (si->node_mapping);
1935 VEC_free (unsigned, heap, si->scc_stack);
1940 /* Find indirect cycles in GRAPH that occur, using strongly connected
1941 components, and note them in the indirect cycles map.
1943 This technique comes from Ben Hardekopf and Calvin Lin,
1944 "It Pays to be Lazy: Fast and Accurate Pointer Analysis for Millions of
1945 Lines of Code", submitted to PLDI 2007. */
1948 find_indirect_cycles (constraint_graph_t graph)
1951 unsigned int size = graph->size;
1952 struct scc_info *si = init_scc_info (size);
1954 for (i = 0; i < MIN (LAST_REF_NODE, size); i ++ )
1955 if (!TEST_BIT (si->visited, i) && find (i) == i)
1956 scc_visit (graph, si, i);
1961 /* Compute a topological ordering for GRAPH, and store the result in the
1962 topo_info structure TI. */
1965 compute_topo_order (constraint_graph_t graph,
1966 struct topo_info *ti)
1969 unsigned int size = graph->size;
1971 for (i = 0; i != size; ++i)
1972 if (!TEST_BIT (ti->visited, i) && find (i) == i)
1973 topo_visit (graph, ti, i);
1976 /* Structure used to for hash value numbering of pointer equivalence
1979 typedef struct equiv_class_label
1982 unsigned int equivalence_class;
1984 } *equiv_class_label_t;
1985 typedef const struct equiv_class_label *const_equiv_class_label_t;
1987 /* A hashtable for mapping a bitmap of labels->pointer equivalence
1989 static htab_t pointer_equiv_class_table;
1991 /* A hashtable for mapping a bitmap of labels->location equivalence
1993 static htab_t location_equiv_class_table;
1995 /* Hash function for a equiv_class_label_t */
1998 equiv_class_label_hash (const void *p)
2000 const_equiv_class_label_t const ecl = (const_equiv_class_label_t) p;
2001 return ecl->hashcode;
2004 /* Equality function for two equiv_class_label_t's. */
2007 equiv_class_label_eq (const void *p1, const void *p2)
2009 const_equiv_class_label_t const eql1 = (const_equiv_class_label_t) p1;
2010 const_equiv_class_label_t const eql2 = (const_equiv_class_label_t) p2;
2011 return (eql1->hashcode == eql2->hashcode
2012 && bitmap_equal_p (eql1->labels, eql2->labels));
2015 /* Lookup a equivalence class in TABLE by the bitmap of LABELS it
2019 equiv_class_lookup (htab_t table, bitmap labels)
2022 struct equiv_class_label ecl;
2024 ecl.labels = labels;
2025 ecl.hashcode = bitmap_hash (labels);
2027 slot = htab_find_slot_with_hash (table, &ecl,
2028 ecl.hashcode, NO_INSERT);
2032 return ((equiv_class_label_t) *slot)->equivalence_class;
2036 /* Add an equivalence class named EQUIVALENCE_CLASS with labels LABELS
2040 equiv_class_add (htab_t table, unsigned int equivalence_class,
2044 equiv_class_label_t ecl = XNEW (struct equiv_class_label);
2046 ecl->labels = labels;
2047 ecl->equivalence_class = equivalence_class;
2048 ecl->hashcode = bitmap_hash (labels);
2050 slot = htab_find_slot_with_hash (table, ecl,
2051 ecl->hashcode, INSERT);
2052 gcc_assert (!*slot);
2053 *slot = (void *) ecl;
2056 /* Perform offline variable substitution.
2058 This is a worst case quadratic time way of identifying variables
2059 that must have equivalent points-to sets, including those caused by
2060 static cycles, and single entry subgraphs, in the constraint graph.
2062 The technique is described in "Exploiting Pointer and Location
2063 Equivalence to Optimize Pointer Analysis. In the 14th International
2064 Static Analysis Symposium (SAS), August 2007." It is known as the
2065 "HU" algorithm, and is equivalent to value numbering the collapsed
2066 constraint graph including evaluating unions.
2068 The general method of finding equivalence classes is as follows:
2069 Add fake nodes (REF nodes) and edges for *a = b and a = *b constraints.
2070 Initialize all non-REF nodes to be direct nodes.
2071 For each constraint a = a U {b}, we set pts(a) = pts(a) u {fresh
2073 For each constraint containing the dereference, we also do the same
2076 We then compute SCC's in the graph and unify nodes in the same SCC,
2079 For each non-collapsed node x:
2080 Visit all unvisited explicit incoming edges.
2081 Ignoring all non-pointers, set pts(x) = Union of pts(a) for y
2083 Lookup the equivalence class for pts(x).
2084 If we found one, equivalence_class(x) = found class.
2085 Otherwise, equivalence_class(x) = new class, and new_class is
2086 added to the lookup table.
2088 All direct nodes with the same equivalence class can be replaced
2089 with a single representative node.
2090 All unlabeled nodes (label == 0) are not pointers and all edges
2091 involving them can be eliminated.
2092 We perform these optimizations during rewrite_constraints
2094 In addition to pointer equivalence class finding, we also perform
2095 location equivalence class finding. This is the set of variables
2096 that always appear together in points-to sets. We use this to
2097 compress the size of the points-to sets. */
2099 /* Current maximum pointer equivalence class id. */
2100 static int pointer_equiv_class;
2102 /* Current maximum location equivalence class id. */
2103 static int location_equiv_class;
2105 /* Recursive routine to find strongly connected components in GRAPH,
2106 and label it's nodes with DFS numbers. */
2109 condense_visit (constraint_graph_t graph, struct scc_info *si, unsigned int n)
2113 unsigned int my_dfs;
2115 gcc_assert (si->node_mapping[n] == n);
2116 SET_BIT (si->visited, n);
2117 si->dfs[n] = si->current_index ++;
2118 my_dfs = si->dfs[n];
2120 /* Visit all the successors. */
2121 EXECUTE_IF_IN_NONNULL_BITMAP (graph->preds[n], 0, i, bi)
2123 unsigned int w = si->node_mapping[i];
2125 if (TEST_BIT (si->deleted, w))
2128 if (!TEST_BIT (si->visited, w))
2129 condense_visit (graph, si, w);
2131 unsigned int t = si->node_mapping[w];
2132 unsigned int nnode = si->node_mapping[n];
2133 gcc_assert (nnode == n);
2135 if (si->dfs[t] < si->dfs[nnode])
2136 si->dfs[n] = si->dfs[t];
2140 /* Visit all the implicit predecessors. */
2141 EXECUTE_IF_IN_NONNULL_BITMAP (graph->implicit_preds[n], 0, i, bi)
2143 unsigned int w = si->node_mapping[i];
2145 if (TEST_BIT (si->deleted, w))
2148 if (!TEST_BIT (si->visited, w))
2149 condense_visit (graph, si, w);
2151 unsigned int t = si->node_mapping[w];
2152 unsigned int nnode = si->node_mapping[n];
2153 gcc_assert (nnode == n);
2155 if (si->dfs[t] < si->dfs[nnode])
2156 si->dfs[n] = si->dfs[t];
2160 /* See if any components have been identified. */
2161 if (si->dfs[n] == my_dfs)
2163 while (VEC_length (unsigned, si->scc_stack) != 0
2164 && si->dfs[VEC_last (unsigned, si->scc_stack)] >= my_dfs)
2166 unsigned int w = VEC_pop (unsigned, si->scc_stack);
2167 si->node_mapping[w] = n;
2169 if (!TEST_BIT (graph->direct_nodes, w))
2170 RESET_BIT (graph->direct_nodes, n);
2172 /* Unify our nodes. */
2173 if (graph->preds[w])
2175 if (!graph->preds[n])
2176 graph->preds[n] = BITMAP_ALLOC (&predbitmap_obstack);
2177 bitmap_ior_into (graph->preds[n], graph->preds[w]);
2179 if (graph->implicit_preds[w])
2181 if (!graph->implicit_preds[n])
2182 graph->implicit_preds[n] = BITMAP_ALLOC (&predbitmap_obstack);
2183 bitmap_ior_into (graph->implicit_preds[n],
2184 graph->implicit_preds[w]);
2186 if (graph->points_to[w])
2188 if (!graph->points_to[n])
2189 graph->points_to[n] = BITMAP_ALLOC (&predbitmap_obstack);
2190 bitmap_ior_into (graph->points_to[n],
2191 graph->points_to[w]);
2194 SET_BIT (si->deleted, n);
2197 VEC_safe_push (unsigned, heap, si->scc_stack, n);
2200 /* Label pointer equivalences. */
2203 label_visit (constraint_graph_t graph, struct scc_info *si, unsigned int n)
2207 SET_BIT (si->visited, n);
2209 if (!graph->points_to[n])
2210 graph->points_to[n] = BITMAP_ALLOC (&predbitmap_obstack);
2212 /* Label and union our incoming edges's points to sets. */
2213 EXECUTE_IF_IN_NONNULL_BITMAP (graph->preds[n], 0, i, bi)
2215 unsigned int w = si->node_mapping[i];
2216 if (!TEST_BIT (si->visited, w))
2217 label_visit (graph, si, w);
2219 /* Skip unused edges */
2220 if (w == n || graph->pointer_label[w] == 0)
2223 if (graph->points_to[w])
2224 bitmap_ior_into(graph->points_to[n], graph->points_to[w]);
2226 /* Indirect nodes get fresh variables. */
2227 if (!TEST_BIT (graph->direct_nodes, n))
2228 bitmap_set_bit (graph->points_to[n], FIRST_REF_NODE + n);
2230 if (!bitmap_empty_p (graph->points_to[n]))
2232 unsigned int label = equiv_class_lookup (pointer_equiv_class_table,
2233 graph->points_to[n]);
2236 label = pointer_equiv_class++;
2237 equiv_class_add (pointer_equiv_class_table,
2238 label, graph->points_to[n]);
2240 graph->pointer_label[n] = label;
2244 /* Perform offline variable substitution, discovering equivalence
2245 classes, and eliminating non-pointer variables. */
2247 static struct scc_info *
2248 perform_var_substitution (constraint_graph_t graph)
2251 unsigned int size = graph->size;
2252 struct scc_info *si = init_scc_info (size);
2254 bitmap_obstack_initialize (&iteration_obstack);
2255 pointer_equiv_class_table = htab_create (511, equiv_class_label_hash,
2256 equiv_class_label_eq, free);
2257 location_equiv_class_table = htab_create (511, equiv_class_label_hash,
2258 equiv_class_label_eq, free);
2259 pointer_equiv_class = 1;
2260 location_equiv_class = 1;
2262 /* Condense the nodes, which means to find SCC's, count incoming
2263 predecessors, and unite nodes in SCC's. */
2264 for (i = 0; i < FIRST_REF_NODE; i++)
2265 if (!TEST_BIT (si->visited, si->node_mapping[i]))
2266 condense_visit (graph, si, si->node_mapping[i]);
2268 sbitmap_zero (si->visited);
2269 /* Actually the label the nodes for pointer equivalences */
2270 for (i = 0; i < FIRST_REF_NODE; i++)
2271 if (!TEST_BIT (si->visited, si->node_mapping[i]))
2272 label_visit (graph, si, si->node_mapping[i]);
2274 /* Calculate location equivalence labels. */
2275 for (i = 0; i < FIRST_REF_NODE; i++)
2282 if (!graph->pointed_by[i])
2284 pointed_by = BITMAP_ALLOC (&iteration_obstack);
2286 /* Translate the pointed-by mapping for pointer equivalence
2288 EXECUTE_IF_SET_IN_BITMAP (graph->pointed_by[i], 0, j, bi)
2290 bitmap_set_bit (pointed_by,
2291 graph->pointer_label[si->node_mapping[j]]);
2293 /* The original pointed_by is now dead. */
2294 BITMAP_FREE (graph->pointed_by[i]);
2296 /* Look up the location equivalence label if one exists, or make
2298 label = equiv_class_lookup (location_equiv_class_table,
2302 label = location_equiv_class++;
2303 equiv_class_add (location_equiv_class_table,
2308 if (dump_file && (dump_flags & TDF_DETAILS))
2309 fprintf (dump_file, "Found location equivalence for node %s\n",
2310 get_varinfo (i)->name);
2311 BITMAP_FREE (pointed_by);
2313 graph->loc_label[i] = label;
2317 if (dump_file && (dump_flags & TDF_DETAILS))
2318 for (i = 0; i < FIRST_REF_NODE; i++)
2320 bool direct_node = TEST_BIT (graph->direct_nodes, i);
2322 "Equivalence classes for %s node id %d:%s are pointer: %d"
2324 direct_node ? "Direct node" : "Indirect node", i,
2325 get_varinfo (i)->name,
2326 graph->pointer_label[si->node_mapping[i]],
2327 graph->loc_label[si->node_mapping[i]]);
2330 /* Quickly eliminate our non-pointer variables. */
2332 for (i = 0; i < FIRST_REF_NODE; i++)
2334 unsigned int node = si->node_mapping[i];
2336 if (graph->pointer_label[node] == 0)
2338 if (dump_file && (dump_flags & TDF_DETAILS))
2340 "%s is a non-pointer variable, eliminating edges.\n",
2341 get_varinfo (node)->name);
2342 stats.nonpointer_vars++;
2343 clear_edges_for_node (graph, node);
2350 /* Free information that was only necessary for variable
2354 free_var_substitution_info (struct scc_info *si)
2357 free (graph->pointer_label);
2358 free (graph->loc_label);
2359 free (graph->pointed_by);
2360 free (graph->points_to);
2361 free (graph->eq_rep);
2362 sbitmap_free (graph->direct_nodes);
2363 htab_delete (pointer_equiv_class_table);
2364 htab_delete (location_equiv_class_table);
2365 bitmap_obstack_release (&iteration_obstack);
2368 /* Return an existing node that is equivalent to NODE, which has
2369 equivalence class LABEL, if one exists. Return NODE otherwise. */
2372 find_equivalent_node (constraint_graph_t graph,
2373 unsigned int node, unsigned int label)
2375 /* If the address version of this variable is unused, we can
2376 substitute it for anything else with the same label.
2377 Otherwise, we know the pointers are equivalent, but not the
2378 locations, and we can unite them later. */
2380 if (!bitmap_bit_p (graph->address_taken, node))
2382 gcc_assert (label < graph->size);
2384 if (graph->eq_rep[label] != -1)
2386 /* Unify the two variables since we know they are equivalent. */
2387 if (unite (graph->eq_rep[label], node))
2388 unify_nodes (graph, graph->eq_rep[label], node, false);
2389 return graph->eq_rep[label];
2393 graph->eq_rep[label] = node;
2394 graph->pe_rep[label] = node;
2399 gcc_assert (label < graph->size);
2400 graph->pe[node] = label;
2401 if (graph->pe_rep[label] == -1)
2402 graph->pe_rep[label] = node;
2408 /* Unite pointer equivalent but not location equivalent nodes in
2409 GRAPH. This may only be performed once variable substitution is
2413 unite_pointer_equivalences (constraint_graph_t graph)
2417 /* Go through the pointer equivalences and unite them to their
2418 representative, if they aren't already. */
2419 for (i = 0; i < FIRST_REF_NODE; i++)
2421 unsigned int label = graph->pe[i];
2424 int label_rep = graph->pe_rep[label];
2426 if (label_rep == -1)
2429 label_rep = find (label_rep);
2430 if (label_rep >= 0 && unite (label_rep, find (i)))
2431 unify_nodes (graph, label_rep, i, false);
2436 /* Move complex constraints to the GRAPH nodes they belong to. */
2439 move_complex_constraints (constraint_graph_t graph)
2444 for (i = 0; VEC_iterate (constraint_t, constraints, i, c); i++)
2448 struct constraint_expr lhs = c->lhs;
2449 struct constraint_expr rhs = c->rhs;
2451 if (lhs.type == DEREF)
2453 insert_into_complex (graph, lhs.var, c);
2455 else if (rhs.type == DEREF)
2457 if (!(get_varinfo (lhs.var)->is_special_var))
2458 insert_into_complex (graph, rhs.var, c);
2460 else if (rhs.type != ADDRESSOF && lhs.var > anything_id
2461 && (lhs.offset != 0 || rhs.offset != 0))
2463 insert_into_complex (graph, rhs.var, c);
2470 /* Optimize and rewrite complex constraints while performing
2471 collapsing of equivalent nodes. SI is the SCC_INFO that is the
2472 result of perform_variable_substitution. */
2475 rewrite_constraints (constraint_graph_t graph,
2476 struct scc_info *si)
2482 for (j = 0; j < graph->size; j++)
2483 gcc_assert (find (j) == j);
2485 for (i = 0; VEC_iterate (constraint_t, constraints, i, c); i++)
2487 struct constraint_expr lhs = c->lhs;
2488 struct constraint_expr rhs = c->rhs;
2489 unsigned int lhsvar = find (lhs.var);
2490 unsigned int rhsvar = find (rhs.var);
2491 unsigned int lhsnode, rhsnode;
2492 unsigned int lhslabel, rhslabel;
2494 lhsnode = si->node_mapping[lhsvar];
2495 rhsnode = si->node_mapping[rhsvar];
2496 lhslabel = graph->pointer_label[lhsnode];
2497 rhslabel = graph->pointer_label[rhsnode];
2499 /* See if it is really a non-pointer variable, and if so, ignore
2503 if (dump_file && (dump_flags & TDF_DETAILS))
2506 fprintf (dump_file, "%s is a non-pointer variable,"
2507 "ignoring constraint:",
2508 get_varinfo (lhs.var)->name);
2509 dump_constraint (dump_file, c);
2511 VEC_replace (constraint_t, constraints, i, NULL);
2517 if (dump_file && (dump_flags & TDF_DETAILS))
2520 fprintf (dump_file, "%s is a non-pointer variable,"
2521 "ignoring constraint:",
2522 get_varinfo (rhs.var)->name);
2523 dump_constraint (dump_file, c);
2525 VEC_replace (constraint_t, constraints, i, NULL);
2529 lhsvar = find_equivalent_node (graph, lhsvar, lhslabel);
2530 rhsvar = find_equivalent_node (graph, rhsvar, rhslabel);
2531 c->lhs.var = lhsvar;
2532 c->rhs.var = rhsvar;
2537 /* Eliminate indirect cycles involving NODE. Return true if NODE was
2538 part of an SCC, false otherwise. */
2541 eliminate_indirect_cycles (unsigned int node)
2543 if (graph->indirect_cycles[node] != -1
2544 && !bitmap_empty_p (get_varinfo (node)->solution))
2547 VEC(unsigned,heap) *queue = NULL;
2549 unsigned int to = find (graph->indirect_cycles[node]);
2552 /* We can't touch the solution set and call unify_nodes
2553 at the same time, because unify_nodes is going to do
2554 bitmap unions into it. */
2556 EXECUTE_IF_SET_IN_BITMAP (get_varinfo (node)->solution, 0, i, bi)
2558 if (find (i) == i && i != to)
2561 VEC_safe_push (unsigned, heap, queue, i);
2566 VEC_iterate (unsigned, queue, queuepos, i);
2569 unify_nodes (graph, to, i, true);
2571 VEC_free (unsigned, heap, queue);
2577 /* Solve the constraint graph GRAPH using our worklist solver.
2578 This is based on the PW* family of solvers from the "Efficient Field
2579 Sensitive Pointer Analysis for C" paper.
2580 It works by iterating over all the graph nodes, processing the complex
2581 constraints and propagating the copy constraints, until everything stops
2582 changed. This corresponds to steps 6-8 in the solving list given above. */
2585 solve_graph (constraint_graph_t graph)
2587 unsigned int size = graph->size;
2592 changed = sbitmap_alloc (size);
2593 sbitmap_zero (changed);
2595 /* Mark all initial non-collapsed nodes as changed. */
2596 for (i = 0; i < size; i++)
2598 varinfo_t ivi = get_varinfo (i);
2599 if (find (i) == i && !bitmap_empty_p (ivi->solution)
2600 && ((graph->succs[i] && !bitmap_empty_p (graph->succs[i]))
2601 || VEC_length (constraint_t, graph->complex[i]) > 0))
2603 SET_BIT (changed, i);
2608 /* Allocate a bitmap to be used to store the changed bits. */
2609 pts = BITMAP_ALLOC (&pta_obstack);
2611 while (changed_count > 0)
2614 struct topo_info *ti = init_topo_info ();
2617 bitmap_obstack_initialize (&iteration_obstack);
2619 compute_topo_order (graph, ti);
2621 while (VEC_length (unsigned, ti->topo_order) != 0)
2624 i = VEC_pop (unsigned, ti->topo_order);
2626 /* If this variable is not a representative, skip it. */
2630 /* In certain indirect cycle cases, we may merge this
2631 variable to another. */
2632 if (eliminate_indirect_cycles (i) && find (i) != i)
2635 /* If the node has changed, we need to process the
2636 complex constraints and outgoing edges again. */
2637 if (TEST_BIT (changed, i))
2642 VEC(constraint_t,heap) *complex = graph->complex[i];
2643 bool solution_empty;
2645 RESET_BIT (changed, i);
2648 /* Compute the changed set of solution bits. */
2649 bitmap_and_compl (pts, get_varinfo (i)->solution,
2650 get_varinfo (i)->oldsolution);
2652 if (bitmap_empty_p (pts))
2655 bitmap_ior_into (get_varinfo (i)->oldsolution, pts);
2657 solution = get_varinfo (i)->solution;
2658 solution_empty = bitmap_empty_p (solution);
2660 /* Process the complex constraints */
2661 for (j = 0; VEC_iterate (constraint_t, complex, j, c); j++)
2663 /* XXX: This is going to unsort the constraints in
2664 some cases, which will occasionally add duplicate
2665 constraints during unification. This does not
2666 affect correctness. */
2667 c->lhs.var = find (c->lhs.var);
2668 c->rhs.var = find (c->rhs.var);
2670 /* The only complex constraint that can change our
2671 solution to non-empty, given an empty solution,
2672 is a constraint where the lhs side is receiving
2673 some set from elsewhere. */
2674 if (!solution_empty || c->lhs.type != DEREF)
2675 do_complex_constraint (graph, c, pts);
2678 solution_empty = bitmap_empty_p (solution);
2680 if (!solution_empty)
2683 unsigned eff_escaped_id = find (escaped_id);
2685 /* Propagate solution to all successors. */
2686 EXECUTE_IF_IN_NONNULL_BITMAP (graph->succs[i],
2692 unsigned int to = find (j);
2693 tmp = get_varinfo (to)->solution;
2696 /* Don't try to propagate to ourselves. */
2700 /* If we propagate from ESCAPED use ESCAPED as
2702 if (i == eff_escaped_id)
2703 flag = bitmap_set_bit (tmp, escaped_id);
2705 flag = set_union_with_increment (tmp, pts, 0);
2709 get_varinfo (to)->solution = tmp;
2710 if (!TEST_BIT (changed, to))
2712 SET_BIT (changed, to);
2720 free_topo_info (ti);
2721 bitmap_obstack_release (&iteration_obstack);
2725 sbitmap_free (changed);
2726 bitmap_obstack_release (&oldpta_obstack);
2729 /* Map from trees to variable infos. */
2730 static struct pointer_map_t *vi_for_tree;
2733 /* Insert ID as the variable id for tree T in the vi_for_tree map. */
2736 insert_vi_for_tree (tree t, varinfo_t vi)
2738 void **slot = pointer_map_insert (vi_for_tree, t);
2740 gcc_assert (*slot == NULL);
2744 /* Find the variable info for tree T in VI_FOR_TREE. If T does not
2745 exist in the map, return NULL, otherwise, return the varinfo we found. */
2748 lookup_vi_for_tree (tree t)
2750 void **slot = pointer_map_contains (vi_for_tree, t);
2754 return (varinfo_t) *slot;
2757 /* Return a printable name for DECL */
2760 alias_get_name (tree decl)
2764 int num_printed = 0;
2766 if (DECL_ASSEMBLER_NAME_SET_P (decl))
2767 res = IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (decl));
2769 res= get_name (decl);
2777 if (TREE_CODE (decl) == SSA_NAME)
2779 num_printed = asprintf (&temp, "%s_%u",
2780 alias_get_name (SSA_NAME_VAR (decl)),
2781 SSA_NAME_VERSION (decl));
2783 else if (DECL_P (decl))
2785 num_printed = asprintf (&temp, "D.%u", DECL_UID (decl));
2787 if (num_printed > 0)
2789 res = ggc_strdup (temp);
2795 /* Find the variable id for tree T in the map.
2796 If T doesn't exist in the map, create an entry for it and return it. */
2799 get_vi_for_tree (tree t)
2801 void **slot = pointer_map_contains (vi_for_tree, t);
2803 return get_varinfo (create_variable_info_for (t, alias_get_name (t)));
2805 return (varinfo_t) *slot;
2808 /* Get a scalar constraint expression for a new temporary variable. */
2810 static struct constraint_expr
2811 new_scalar_tmp_constraint_exp (const char *name)
2813 struct constraint_expr tmp;
2816 vi = new_var_info (NULL_TREE, name);
2820 vi->is_full_var = 1;
2829 /* Get a constraint expression vector from an SSA_VAR_P node.
2830 If address_p is true, the result will be taken its address of. */
2833 get_constraint_for_ssa_var (tree t, VEC(ce_s, heap) **results, bool address_p)
2835 struct constraint_expr cexpr;
2838 /* We allow FUNCTION_DECLs here even though it doesn't make much sense. */
2839 gcc_assert (SSA_VAR_P (t) || DECL_P (t));
2841 /* For parameters, get at the points-to set for the actual parm
2843 if (TREE_CODE (t) == SSA_NAME
2844 && TREE_CODE (SSA_NAME_VAR (t)) == PARM_DECL
2845 && SSA_NAME_IS_DEFAULT_DEF (t))
2847 get_constraint_for_ssa_var (SSA_NAME_VAR (t), results, address_p);
2851 vi = get_vi_for_tree (t);
2853 cexpr.type = SCALAR;
2855 /* If we determine the result is "anything", and we know this is readonly,
2856 say it points to readonly memory instead. */
2857 if (cexpr.var == anything_id && TREE_READONLY (t))
2860 cexpr.type = ADDRESSOF;
2861 cexpr.var = readonly_id;
2864 /* If we are not taking the address of the constraint expr, add all
2865 sub-fiels of the variable as well. */
2867 && !vi->is_full_var)
2869 for (; vi; vi = vi->next)
2872 VEC_safe_push (ce_s, heap, *results, &cexpr);
2877 VEC_safe_push (ce_s, heap, *results, &cexpr);
2880 /* Process constraint T, performing various simplifications and then
2881 adding it to our list of overall constraints. */
2884 process_constraint (constraint_t t)
2886 struct constraint_expr rhs = t->rhs;
2887 struct constraint_expr lhs = t->lhs;
2889 gcc_assert (rhs.var < VEC_length (varinfo_t, varmap));
2890 gcc_assert (lhs.var < VEC_length (varinfo_t, varmap));
2892 /* If we didn't get any useful constraint from the lhs we get
2893 &ANYTHING as fallback from get_constraint_for. Deal with
2894 it here by turning it into *ANYTHING. */
2895 if (lhs.type == ADDRESSOF
2896 && lhs.var == anything_id)
2899 /* ADDRESSOF on the lhs is invalid. */
2900 gcc_assert (lhs.type != ADDRESSOF);
2902 /* We shouldn't add constraints from things that cannot have pointers.
2903 It's not completely trivial to avoid in the callers, so do it here. */
2904 if (rhs.type != ADDRESSOF
2905 && !get_varinfo (rhs.var)->may_have_pointers)
2908 /* Likewise adding to the solution of a non-pointer var isn't useful. */
2909 if (!get_varinfo (lhs.var)->may_have_pointers)
2912 /* This can happen in our IR with things like n->a = *p */
2913 if (rhs.type == DEREF && lhs.type == DEREF && rhs.var != anything_id)
2915 /* Split into tmp = *rhs, *lhs = tmp */
2916 struct constraint_expr tmplhs;
2917 tmplhs = new_scalar_tmp_constraint_exp ("doubledereftmp");
2918 process_constraint (new_constraint (tmplhs, rhs));
2919 process_constraint (new_constraint (lhs, tmplhs));
2921 else if (rhs.type == ADDRESSOF && lhs.type == DEREF)
2923 /* Split into tmp = &rhs, *lhs = tmp */
2924 struct constraint_expr tmplhs;
2925 tmplhs = new_scalar_tmp_constraint_exp ("derefaddrtmp");
2926 process_constraint (new_constraint (tmplhs, rhs));
2927 process_constraint (new_constraint (lhs, tmplhs));
2931 gcc_assert (rhs.type != ADDRESSOF || rhs.offset == 0);
2932 VEC_safe_push (constraint_t, heap, constraints, t);
2936 /* Return true if T is a type that could contain pointers. */
2939 type_could_have_pointers (tree type)
2941 if (POINTER_TYPE_P (type))
2944 if (TREE_CODE (type) == ARRAY_TYPE)
2945 return type_could_have_pointers (TREE_TYPE (type));
2947 /* A function or method can consume pointers.
2948 ??? We could be more precise here. */
2949 if (TREE_CODE (type) == FUNCTION_TYPE
2950 || TREE_CODE (type) == METHOD_TYPE)
2953 return AGGREGATE_TYPE_P (type);
2956 /* Return true if T is a variable of a type that could contain
2960 could_have_pointers (tree t)
2962 return (((TREE_CODE (t) == VAR_DECL
2963 || TREE_CODE (t) == PARM_DECL
2964 || TREE_CODE (t) == RESULT_DECL)
2965 && (TREE_PUBLIC (t) || DECL_EXTERNAL (t) || TREE_ADDRESSABLE (t)))
2966 || type_could_have_pointers (TREE_TYPE (t)));
2969 /* Return the position, in bits, of FIELD_DECL from the beginning of its
2972 static HOST_WIDE_INT
2973 bitpos_of_field (const tree fdecl)
2976 if (!host_integerp (DECL_FIELD_OFFSET (fdecl), 0)
2977 || !host_integerp (DECL_FIELD_BIT_OFFSET (fdecl), 0))
2980 return (TREE_INT_CST_LOW (DECL_FIELD_OFFSET (fdecl)) * 8
2981 + TREE_INT_CST_LOW (DECL_FIELD_BIT_OFFSET (fdecl)));
2985 /* Get constraint expressions for offsetting PTR by OFFSET. Stores the
2986 resulting constraint expressions in *RESULTS. */
2989 get_constraint_for_ptr_offset (tree ptr, tree offset,
2990 VEC (ce_s, heap) **results)
2992 struct constraint_expr c;
2994 HOST_WIDE_INT rhsunitoffset, rhsoffset;
2996 /* If we do not do field-sensitive PTA adding offsets to pointers
2997 does not change the points-to solution. */
2998 if (!use_field_sensitive)
3000 get_constraint_for (ptr, results);
3004 /* If the offset is not a non-negative integer constant that fits
3005 in a HOST_WIDE_INT, we have to fall back to a conservative
3006 solution which includes all sub-fields of all pointed-to
3007 variables of ptr. */
3008 if (offset == NULL_TREE
3009 || !host_integerp (offset, 0))
3010 rhsoffset = UNKNOWN_OFFSET;
3013 /* Make sure the bit-offset also fits. */
3014 rhsunitoffset = TREE_INT_CST_LOW (offset);
3015 rhsoffset = rhsunitoffset * BITS_PER_UNIT;
3016 if (rhsunitoffset != rhsoffset / BITS_PER_UNIT)
3017 rhsoffset = UNKNOWN_OFFSET;
3020 get_constraint_for (ptr, results);
3024 /* As we are eventually appending to the solution do not use
3025 VEC_iterate here. */
3026 n = VEC_length (ce_s, *results);
3027 for (j = 0; j < n; j++)
3030 c = *VEC_index (ce_s, *results, j);
3031 curr = get_varinfo (c.var);
3033 if (c.type == ADDRESSOF
3034 /* If this varinfo represents a full variable just use it. */
3035 && curr->is_full_var)
3037 else if (c.type == ADDRESSOF
3038 /* If we do not know the offset add all subfields. */
3039 && rhsoffset == UNKNOWN_OFFSET)
3041 varinfo_t temp = lookup_vi_for_tree (curr->decl);
3044 struct constraint_expr c2;
3046 c2.type = ADDRESSOF;
3048 if (c2.var != c.var)
3049 VEC_safe_push (ce_s, heap, *results, &c2);
3054 else if (c.type == ADDRESSOF)
3057 unsigned HOST_WIDE_INT offset = curr->offset + rhsoffset;
3059 /* Search the sub-field which overlaps with the
3060 pointed-to offset. If the result is outside of the variable
3061 we have to provide a conservative result, as the variable is
3062 still reachable from the resulting pointer (even though it
3063 technically cannot point to anything). The last and first
3064 sub-fields are such conservative results.
3065 ??? If we always had a sub-field for &object + 1 then
3066 we could represent this in a more precise way. */
3068 && curr->offset < offset)
3070 temp = first_or_preceding_vi_for_offset (curr, offset);
3072 /* If the found variable is not exactly at the pointed to
3073 result, we have to include the next variable in the
3074 solution as well. Otherwise two increments by offset / 2
3075 do not result in the same or a conservative superset
3077 if (temp->offset != offset
3078 && temp->next != NULL)
3080 struct constraint_expr c2;
3081 c2.var = temp->next->id;
3082 c2.type = ADDRESSOF;
3084 VEC_safe_push (ce_s, heap, *results, &c2);
3090 c.offset = rhsoffset;
3092 VEC_replace (ce_s, *results, j, &c);
3097 /* Given a COMPONENT_REF T, return the constraint_expr vector for it.
3098 If address_p is true the result will be taken its address of. */
3101 get_constraint_for_component_ref (tree t, VEC(ce_s, heap) **results,
3105 HOST_WIDE_INT bitsize = -1;
3106 HOST_WIDE_INT bitmaxsize = -1;
3107 HOST_WIDE_INT bitpos;
3109 struct constraint_expr *result;
3111 /* Some people like to do cute things like take the address of
3114 while (handled_component_p (forzero)
3115 || INDIRECT_REF_P (forzero))
3116 forzero = TREE_OPERAND (forzero, 0);
3118 if (CONSTANT_CLASS_P (forzero) && integer_zerop (forzero))
3120 struct constraint_expr temp;
3123 temp.var = integer_id;
3125 VEC_safe_push (ce_s, heap, *results, &temp);
3129 t = get_ref_base_and_extent (t, &bitpos, &bitsize, &bitmaxsize);
3131 /* Pretend to take the address of the base, we'll take care of
3132 adding the required subset of sub-fields below. */
3133 get_constraint_for_1 (t, results, true);
3134 gcc_assert (VEC_length (ce_s, *results) == 1);
3135 result = VEC_last (ce_s, *results);
3137 if (result->type == SCALAR
3138 && get_varinfo (result->var)->is_full_var)
3139 /* For single-field vars do not bother about the offset. */
3141 else if (result->type == SCALAR)
3143 /* In languages like C, you can access one past the end of an
3144 array. You aren't allowed to dereference it, so we can
3145 ignore this constraint. When we handle pointer subtraction,
3146 we may have to do something cute here. */
3148 if ((unsigned HOST_WIDE_INT)bitpos < get_varinfo (result->var)->fullsize
3151 /* It's also not true that the constraint will actually start at the
3152 right offset, it may start in some padding. We only care about
3153 setting the constraint to the first actual field it touches, so
3155 struct constraint_expr cexpr = *result;
3157 VEC_pop (ce_s, *results);
3159 for (curr = get_varinfo (cexpr.var); curr; curr = curr->next)
3161 if (ranges_overlap_p (curr->offset, curr->size,
3162 bitpos, bitmaxsize))
3164 cexpr.var = curr->id;
3165 VEC_safe_push (ce_s, heap, *results, &cexpr);
3170 /* If we are going to take the address of this field then
3171 to be able to compute reachability correctly add at least
3172 the last field of the variable. */
3174 && VEC_length (ce_s, *results) == 0)
3176 curr = get_varinfo (cexpr.var);
3177 while (curr->next != NULL)
3179 cexpr.var = curr->id;
3180 VEC_safe_push (ce_s, heap, *results, &cexpr);
3183 /* Assert that we found *some* field there. The user couldn't be
3184 accessing *only* padding. */
3185 /* Still the user could access one past the end of an array
3186 embedded in a struct resulting in accessing *only* padding. */
3187 gcc_assert (VEC_length (ce_s, *results) >= 1
3188 || ref_contains_array_ref (orig_t));
3190 else if (bitmaxsize == 0)
3192 if (dump_file && (dump_flags & TDF_DETAILS))
3193 fprintf (dump_file, "Access to zero-sized part of variable,"
3197 if (dump_file && (dump_flags & TDF_DETAILS))
3198 fprintf (dump_file, "Access to past the end of variable, ignoring\n");
3200 else if (result->type == DEREF)
3202 /* If we do not know exactly where the access goes say so. Note
3203 that only for non-structure accesses we know that we access
3204 at most one subfiled of any variable. */
3206 || bitsize != bitmaxsize
3207 || AGGREGATE_TYPE_P (TREE_TYPE (orig_t)))
3208 result->offset = UNKNOWN_OFFSET;
3210 result->offset = bitpos;
3212 else if (result->type == ADDRESSOF)
3214 /* We can end up here for component references on a
3215 VIEW_CONVERT_EXPR <>(&foobar). */
3216 result->type = SCALAR;
3217 result->var = anything_id;
3225 /* Dereference the constraint expression CONS, and return the result.
3226 DEREF (ADDRESSOF) = SCALAR
3227 DEREF (SCALAR) = DEREF
3228 DEREF (DEREF) = (temp = DEREF1; result = DEREF(temp))
3229 This is needed so that we can handle dereferencing DEREF constraints. */
3232 do_deref (VEC (ce_s, heap) **constraints)
3234 struct constraint_expr *c;
3237 for (i = 0; VEC_iterate (ce_s, *constraints, i, c); i++)
3239 if (c->type == SCALAR)
3241 else if (c->type == ADDRESSOF)
3243 else if (c->type == DEREF)
3245 struct constraint_expr tmplhs;
3246 tmplhs = new_scalar_tmp_constraint_exp ("dereftmp");
3247 process_constraint (new_constraint (tmplhs, *c));
3248 c->var = tmplhs.var;
3255 static void get_constraint_for_1 (tree, VEC (ce_s, heap) **, bool);
3257 /* Given a tree T, return the constraint expression for taking the
3261 get_constraint_for_address_of (tree t, VEC (ce_s, heap) **results)
3263 struct constraint_expr *c;
3266 get_constraint_for_1 (t, results, true);
3268 for (i = 0; VEC_iterate (ce_s, *results, i, c); i++)
3270 if (c->type == DEREF)
3273 c->type = ADDRESSOF;
3277 /* Given a tree T, return the constraint expression for it. */
3280 get_constraint_for_1 (tree t, VEC (ce_s, heap) **results, bool address_p)
3282 struct constraint_expr temp;
3284 /* x = integer is all glommed to a single variable, which doesn't
3285 point to anything by itself. That is, of course, unless it is an
3286 integer constant being treated as a pointer, in which case, we
3287 will return that this is really the addressof anything. This
3288 happens below, since it will fall into the default case. The only
3289 case we know something about an integer treated like a pointer is
3290 when it is the NULL pointer, and then we just say it points to
3293 Do not do that if -fno-delete-null-pointer-checks though, because
3294 in that case *NULL does not fail, so it _should_ alias *anything.
3295 It is not worth adding a new option or renaming the existing one,
3296 since this case is relatively obscure. */
3297 if ((TREE_CODE (t) == INTEGER_CST
3298 && integer_zerop (t))
3299 /* The only valid CONSTRUCTORs in gimple with pointer typed
3300 elements are zero-initializer. But in IPA mode we also
3301 process global initializers, so verify at least. */
3302 || (TREE_CODE (t) == CONSTRUCTOR
3303 && CONSTRUCTOR_NELTS (t) == 0))
3305 if (flag_delete_null_pointer_checks)
3306 temp.var = nothing_id;
3308 temp.var = anything_id;
3309 temp.type = ADDRESSOF;
3311 VEC_safe_push (ce_s, heap, *results, &temp);
3315 /* String constants are read-only. */
3316 if (TREE_CODE (t) == STRING_CST)
3318 temp.var = readonly_id;
3321 VEC_safe_push (ce_s, heap, *results, &temp);
3325 switch (TREE_CODE_CLASS (TREE_CODE (t)))
3327 case tcc_expression:
3329 switch (TREE_CODE (t))
3332 get_constraint_for_address_of (TREE_OPERAND (t, 0), results);
3340 switch (TREE_CODE (t))
3344 get_constraint_for_1 (TREE_OPERAND (t, 0), results, address_p);
3349 case ARRAY_RANGE_REF:
3351 get_constraint_for_component_ref (t, results, address_p);
3353 case VIEW_CONVERT_EXPR:
3354 get_constraint_for_1 (TREE_OPERAND (t, 0), results, address_p);
3356 /* We are missing handling for TARGET_MEM_REF here. */
3361 case tcc_exceptional:
3363 switch (TREE_CODE (t))
3367 get_constraint_for_ssa_var (t, results, address_p);
3374 VEC (ce_s, heap) *tmp = NULL;
3375 FOR_EACH_CONSTRUCTOR_VALUE (CONSTRUCTOR_ELTS (t), i, val)
3377 struct constraint_expr *rhsp;
3379 get_constraint_for_1 (val, &tmp, address_p);
3380 for (j = 0; VEC_iterate (ce_s, tmp, j, rhsp); ++j)
3381 VEC_safe_push (ce_s, heap, *results, rhsp);
3382 VEC_truncate (ce_s, tmp, 0);
3384 VEC_free (ce_s, heap, tmp);
3385 /* We do not know whether the constructor was complete,
3386 so technically we have to add &NOTHING or &ANYTHING
3387 like we do for an empty constructor as well. */
3394 case tcc_declaration:
3396 get_constraint_for_ssa_var (t, results, address_p);
3402 /* The default fallback is a constraint from anything. */
3403 temp.type = ADDRESSOF;
3404 temp.var = anything_id;
3406 VEC_safe_push (ce_s, heap, *results, &temp);
3409 /* Given a gimple tree T, return the constraint expression vector for it. */
3412 get_constraint_for (tree t, VEC (ce_s, heap) **results)
3414 gcc_assert (VEC_length (ce_s, *results) == 0);
3416 get_constraint_for_1 (t, results, false);
3420 /* Efficiently generates constraints from all entries in *RHSC to all
3421 entries in *LHSC. */
3424 process_all_all_constraints (VEC (ce_s, heap) *lhsc, VEC (ce_s, heap) *rhsc)
3426 struct constraint_expr *lhsp, *rhsp;
3429 if (VEC_length (ce_s, lhsc) <= 1
3430 || VEC_length (ce_s, rhsc) <= 1)
3432 for (i = 0; VEC_iterate (ce_s, lhsc, i, lhsp); ++i)
3433 for (j = 0; VEC_iterate (ce_s, rhsc, j, rhsp); ++j)
3434 process_constraint (new_constraint (*lhsp, *rhsp));
3438 struct constraint_expr tmp;
3439 tmp = new_scalar_tmp_constraint_exp ("allalltmp");
3440 for (i = 0; VEC_iterate (ce_s, rhsc, i, rhsp); ++i)
3441 process_constraint (new_constraint (tmp, *rhsp));
3442 for (i = 0; VEC_iterate (ce_s, lhsc, i, lhsp); ++i)
3443 process_constraint (new_constraint (*lhsp, tmp));
3447 /* Handle aggregate copies by expanding into copies of the respective
3448 fields of the structures. */
3451 do_structure_copy (tree lhsop, tree rhsop)
3453 struct constraint_expr *lhsp, *rhsp;
3454 VEC (ce_s, heap) *lhsc = NULL, *rhsc = NULL;
3457 get_constraint_for (lhsop, &lhsc);
3458 get_constraint_for (rhsop, &rhsc);
3459 lhsp = VEC_index (ce_s, lhsc, 0);
3460 rhsp = VEC_index (ce_s, rhsc, 0);
3461 if (lhsp->type == DEREF
3462 || (lhsp->type == ADDRESSOF && lhsp->var == anything_id)
3463 || rhsp->type == DEREF)
3465 if (lhsp->type == DEREF)
3467 gcc_assert (VEC_length (ce_s, lhsc) == 1);
3468 lhsp->offset = UNKNOWN_OFFSET;
3470 if (rhsp->type == DEREF)
3472 gcc_assert (VEC_length (ce_s, rhsc) == 1);
3473 rhsp->offset = UNKNOWN_OFFSET;
3475 process_all_all_constraints (lhsc, rhsc);
3477 else if (lhsp->type == SCALAR
3478 && (rhsp->type == SCALAR
3479 || rhsp->type == ADDRESSOF))
3481 HOST_WIDE_INT lhssize, lhsmaxsize, lhsoffset;
3482 HOST_WIDE_INT rhssize, rhsmaxsize, rhsoffset;
3484 get_ref_base_and_extent (lhsop, &lhsoffset, &lhssize, &lhsmaxsize);
3485 get_ref_base_and_extent (rhsop, &rhsoffset, &rhssize, &rhsmaxsize);
3486 for (j = 0; VEC_iterate (ce_s, lhsc, j, lhsp);)
3488 varinfo_t lhsv, rhsv;
3489 rhsp = VEC_index (ce_s, rhsc, k);