blob: 581bdcf565243a010ae12b750c196b72ab0cc4e3 [file] [log] [blame]
/* Tree based points-to analysis
Copyright (C) 2005-2022 Free Software Foundation, Inc.
Contributed by Daniel Berlin <dberlin@dberlin.org>
This file is part of GCC.
GCC is free software; you can redistribute it and/or modify
under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 3 of the License, or
(at your option) any later version.
GCC is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with GCC; see the file COPYING3. If not see
<http://www.gnu.org/licenses/>. */
#include "config.h"
#include "system.h"
#include "coretypes.h"
#include "backend.h"
#include "rtl.h"
#include "tree.h"
#include "gimple.h"
#include "alloc-pool.h"
#include "tree-pass.h"
#include "ssa.h"
#include "cgraph.h"
#include "tree-pretty-print.h"
#include "diagnostic-core.h"
#include "fold-const.h"
#include "stor-layout.h"
#include "stmt.h"
#include "gimple-iterator.h"
#include "tree-into-ssa.h"
#include "tree-dfa.h"
#include "gimple-walk.h"
#include "varasm.h"
#include "stringpool.h"
#include "attribs.h"
#include "tree-ssa.h"
#include "tree-cfg.h"
#include "gimple-range.h"
#include "ipa-modref-tree.h"
#include "ipa-modref.h"
#include "attr-fnspec.h"
/* The idea behind this analyzer is to generate set constraints from the
program, then solve the resulting constraints in order to generate the
points-to sets.
Set constraints are a way of modeling program analysis problems that
involve sets. They consist of an inclusion constraint language,
describing the variables (each variable is a set) and operations that
are involved on the variables, and a set of rules that derive facts
from these operations. To solve a system of set constraints, you derive
all possible facts under the rules, which gives you the correct sets
as a consequence.
See "Efficient Field-sensitive pointer analysis for C" by "David
J. Pearce and Paul H. J. Kelly and Chris Hankin", at
http://citeseer.ist.psu.edu/pearce04efficient.html
Also see "Ultra-fast Aliasing Analysis using CLA: A Million Lines
of C Code in a Second" by "Nevin Heintze and Olivier Tardieu" at
http://citeseer.ist.psu.edu/heintze01ultrafast.html
There are three types of real constraint expressions, DEREF,
ADDRESSOF, and SCALAR. Each constraint expression consists
of a constraint type, a variable, and an offset.
SCALAR is a constraint expression type used to represent x, whether
it appears on the LHS or the RHS of a statement.
DEREF is a constraint expression type used to represent *x, whether
it appears on the LHS or the RHS of a statement.
ADDRESSOF is a constraint expression used to represent &x, whether
it appears on the LHS or the RHS of a statement.
Each pointer variable in the program is assigned an integer id, and
each field of a structure variable is assigned an integer id as well.
Structure variables are linked to their list of fields through a "next
field" in each variable that points to the next field in offset
order.
Each variable for a structure field has
1. "size", that tells the size in bits of that field.
2. "fullsize", that tells the size in bits of the entire structure.
3. "offset", that tells the offset in bits from the beginning of the
structure to this field.
Thus,
struct f
{
int a;
int b;
} foo;
int *bar;
looks like
foo.a -> id 1, size 32, offset 0, fullsize 64, next foo.b
foo.b -> id 2, size 32, offset 32, fullsize 64, next NULL
bar -> id 3, size 32, offset 0, fullsize 32, next NULL
In order to solve the system of set constraints, the following is
done:
1. Each constraint variable x has a solution set associated with it,
Sol(x).
2. Constraints are separated into direct, copy, and complex.
Direct constraints are ADDRESSOF constraints that require no extra
processing, such as P = &Q
Copy constraints are those of the form P = Q.
Complex constraints are all the constraints involving dereferences
and offsets (including offsetted copies).
3. All direct constraints of the form P = &Q are processed, such
that Q is added to Sol(P)
4. All complex constraints for a given constraint variable are stored in a
linked list attached to that variable's node.
5. A directed graph is built out of the copy constraints. Each
constraint variable is a node in the graph, and an edge from
Q to P is added for each copy constraint of the form P = Q
6. The graph is then walked, and solution sets are
propagated along the copy edges, such that an edge from Q to P
causes Sol(P) <- Sol(P) union Sol(Q).
7. As we visit each node, all complex constraints associated with
that node are processed by adding appropriate copy edges to the graph, or the
appropriate variables to the solution set.
8. The process of walking the graph is iterated until no solution
sets change.
Prior to walking the graph in steps 6 and 7, We perform static
cycle elimination on the constraint graph, as well
as off-line variable substitution.
TODO: Adding offsets to pointer-to-structures can be handled (IE not punted
on and turned into anything), but isn't. You can just see what offset
inside the pointed-to struct it's going to access.
TODO: Constant bounded arrays can be handled as if they were structs of the
same number of elements.
TODO: Modeling heap and incoming pointers becomes much better if we
add fields to them as we discover them, which we could do.
TODO: We could handle unions, but to be honest, it's probably not
worth the pain or slowdown. */
/* IPA-PTA optimizations possible.
When the indirect function called is ANYTHING we can add disambiguation
based on the function signatures (or simply the parameter count which
is the varinfo size). We also do not need to consider functions that
do not have their address taken.
The is_global_var bit which marks escape points is overly conservative
in IPA mode. Split it to is_escape_point and is_global_var - only
externally visible globals are escape points in IPA mode.
There is now is_ipa_escape_point but this is only used in a few
selected places.
The way we introduce DECL_PT_UID to avoid fixing up all points-to
sets in the translation unit when we copy a DECL during inlining
pessimizes precision. The advantage is that the DECL_PT_UID keeps
compile-time and memory usage overhead low - the points-to sets
do not grow or get unshared as they would during a fixup phase.
An alternative solution is to delay IPA PTA until after all
inlining transformations have been applied.
The way we propagate clobber/use information isn't optimized.
It should use a new complex constraint that properly filters
out local variables of the callee (though that would make
the sets invalid after inlining). OTOH we might as well
admit defeat to WHOPR and simply do all the clobber/use analysis
and propagation after PTA finished but before we threw away
points-to information for memory variables. WHOPR and PTA
do not play along well anyway - the whole constraint solving
would need to be done in WPA phase and it will be very interesting
to apply the results to local SSA names during LTRANS phase.
We probably should compute a per-function unit-ESCAPE solution
propagating it simply like the clobber / uses solutions. The
solution can go alongside the non-IPA escaped solution and be
used to query which vars escape the unit through a function.
This is also required to make the escaped-HEAP trick work in IPA mode.
We never put function decls in points-to sets so we do not
keep the set of called functions for indirect calls.
And probably more. */
static bool use_field_sensitive = true;
static int in_ipa_mode = 0;
/* Used for predecessor bitmaps. */
static bitmap_obstack predbitmap_obstack;
/* Used for points-to sets. */
static bitmap_obstack pta_obstack;
/* Used for oldsolution members of variables. */
static bitmap_obstack oldpta_obstack;
/* Used for per-solver-iteration bitmaps. */
static bitmap_obstack iteration_obstack;
static unsigned int create_variable_info_for (tree, const char *, bool);
typedef struct constraint_graph *constraint_graph_t;
static void unify_nodes (constraint_graph_t, unsigned int, unsigned int, bool);
struct constraint;
typedef struct constraint *constraint_t;
#define EXECUTE_IF_IN_NONNULL_BITMAP(a, b, c, d) \
if (a) \
EXECUTE_IF_SET_IN_BITMAP (a, b, c, d)
static struct constraint_stats
{
unsigned int total_vars;
unsigned int nonpointer_vars;
unsigned int unified_vars_static;
unsigned int unified_vars_dynamic;
unsigned int iterations;
unsigned int num_edges;
unsigned int num_implicit_edges;
unsigned int points_to_sets_created;
} stats;
struct variable_info
{
/* ID of this variable */
unsigned int id;
/* True if this is a variable created by the constraint analysis, such as
heap variables and constraints we had to break up. */
unsigned int is_artificial_var : 1;
/* True if this is a special variable whose solution set should not be
changed. */
unsigned int is_special_var : 1;
/* True for variables whose size is not known or variable. */
unsigned int is_unknown_size_var : 1;
/* True for (sub-)fields that represent a whole variable. */
unsigned int is_full_var : 1;
/* True if this is a heap variable. */
unsigned int is_heap_var : 1;
/* True if this is a register variable. */
unsigned int is_reg_var : 1;
/* True if this field may contain pointers. */
unsigned int may_have_pointers : 1;
/* True if this field has only restrict qualified pointers. */
unsigned int only_restrict_pointers : 1;
/* True if this represents a heap var created for a restrict qualified
pointer. */
unsigned int is_restrict_var : 1;
/* True if this represents a global variable. */
unsigned int is_global_var : 1;
/* True if this represents a module escape point for IPA analysis. */
unsigned int is_ipa_escape_point : 1;
/* True if this represents a IPA function info. */
unsigned int is_fn_info : 1;
/* True if this appears as RHS in a ADDRESSOF constraint. */
unsigned int address_taken : 1;
/* ??? Store somewhere better. */
unsigned short ruid;
/* The ID of the variable for the next field in this structure
or zero for the last field in this structure. */
unsigned next;
/* The ID of the variable for the first field in this structure. */
unsigned head;
/* Offset of this variable, in bits, from the base variable */
unsigned HOST_WIDE_INT offset;
/* Size of the variable, in bits. */
unsigned HOST_WIDE_INT size;
/* Full size of the base variable, in bits. */
unsigned HOST_WIDE_INT fullsize;
/* In IPA mode the shadow UID in case the variable needs to be duplicated in
the final points-to solution because it reaches its containing
function recursively. Zero if none is needed. */
unsigned int shadow_var_uid;
/* Name of this variable */
const char *name;
/* Tree that this variable is associated with. */
tree decl;
/* Points-to set for this variable. */
bitmap solution;
/* Old points-to set for this variable. */
bitmap oldsolution;
};
typedef struct variable_info *varinfo_t;
static varinfo_t first_vi_for_offset (varinfo_t, unsigned HOST_WIDE_INT);
static varinfo_t first_or_preceding_vi_for_offset (varinfo_t,
unsigned HOST_WIDE_INT);
static varinfo_t lookup_vi_for_tree (tree);
static inline bool type_can_have_subvars (const_tree);
static void make_param_constraints (varinfo_t);
/* Pool of variable info structures. */
static object_allocator<variable_info> variable_info_pool
("Variable info pool");
/* Map varinfo to final pt_solution. */
static hash_map<varinfo_t, pt_solution *> *final_solutions;
struct obstack final_solutions_obstack;
/* Table of variable info structures for constraint variables.
Indexed directly by variable info id. */
static vec<varinfo_t> varmap;
/* Return the varmap element N */
static inline varinfo_t
get_varinfo (unsigned int n)
{
return varmap[n];
}
/* Return the next variable in the list of sub-variables of VI
or NULL if VI is the last sub-variable. */
static inline varinfo_t
vi_next (varinfo_t vi)
{
return get_varinfo (vi->next);
}
/* Static IDs for the special variables. Variable ID zero is unused
and used as terminator for the sub-variable chain. */
enum { nothing_id = 1, anything_id = 2, string_id = 3,
escaped_id = 4, nonlocal_id = 5,
storedanything_id = 6, integer_id = 7 };
/* Return a new variable info structure consisting for a variable
named NAME, and using constraint graph node NODE. Append it
to the vector of variable info structures. */
static varinfo_t
new_var_info (tree t, const char *name, bool add_id)
{
unsigned index = varmap.length ();
varinfo_t ret = variable_info_pool.allocate ();
if (dump_file && add_id)
{
char *tempname = xasprintf ("%s(%d)", name, index);
name = ggc_strdup (tempname);
free (tempname);
}
ret->id = index;
ret->name = name;
ret->decl = t;
/* Vars without decl are artificial and do not have sub-variables. */
ret->is_artificial_var = (t == NULL_TREE);
ret->is_special_var = false;
ret->is_unknown_size_var = false;
ret->is_full_var = (t == NULL_TREE);
ret->is_heap_var = false;
ret->may_have_pointers = true;
ret->only_restrict_pointers = false;
ret->is_restrict_var = false;
ret->ruid = 0;
ret->is_global_var = (t == NULL_TREE);
ret->is_ipa_escape_point = false;
ret->is_fn_info = false;
ret->address_taken = false;
if (t && DECL_P (t))
ret->is_global_var = (is_global_var (t)
/* We have to treat even local register variables
as escape points. */
|| (VAR_P (t) && DECL_HARD_REGISTER (t)));
ret->is_reg_var = (t && TREE_CODE (t) == SSA_NAME);
ret->solution = BITMAP_ALLOC (&pta_obstack);
ret->oldsolution = NULL;
ret->next = 0;
ret->shadow_var_uid = 0;
ret->head = ret->id;
stats.total_vars++;
varmap.safe_push (ret);
return ret;
}
/* A map mapping call statements to per-stmt variables for uses
and clobbers specific to the call. */
static hash_map<gimple *, varinfo_t> *call_stmt_vars;
/* Lookup or create the variable for the call statement CALL. */
static varinfo_t
get_call_vi (gcall *call)
{
varinfo_t vi, vi2;
bool existed;
varinfo_t *slot_p = &call_stmt_vars->get_or_insert (call, &existed);
if (existed)
return *slot_p;
vi = new_var_info (NULL_TREE, "CALLUSED", true);
vi->offset = 0;
vi->size = 1;
vi->fullsize = 2;
vi->is_full_var = true;
vi->is_reg_var = true;
vi2 = new_var_info (NULL_TREE, "CALLCLOBBERED", true);
vi2->offset = 1;
vi2->size = 1;
vi2->fullsize = 2;
vi2->is_full_var = true;
vi2->is_reg_var = true;
vi->next = vi2->id;
*slot_p = vi;
return vi;
}
/* Lookup the variable for the call statement CALL representing
the uses. Returns NULL if there is nothing special about this call. */
static varinfo_t
lookup_call_use_vi (gcall *call)
{
varinfo_t *slot_p = call_stmt_vars->get (call);
if (slot_p)
return *slot_p;
return NULL;
}
/* Lookup the variable for the call statement CALL representing
the clobbers. Returns NULL if there is nothing special about this call. */
static varinfo_t
lookup_call_clobber_vi (gcall *call)
{
varinfo_t uses = lookup_call_use_vi (call);
if (!uses)
return NULL;
return vi_next (uses);
}
/* Lookup or create the variable for the call statement CALL representing
the uses. */
static varinfo_t
get_call_use_vi (gcall *call)
{
return get_call_vi (call);
}
/* Lookup or create the variable for the call statement CALL representing
the clobbers. */
static varinfo_t ATTRIBUTE_UNUSED
get_call_clobber_vi (gcall *call)
{
return vi_next (get_call_vi (call));
}
enum constraint_expr_type {SCALAR, DEREF, ADDRESSOF};
/* An expression that appears in a constraint. */
struct constraint_expr
{
/* Constraint type. */
constraint_expr_type type;
/* Variable we are referring to in the constraint. */
unsigned int var;
/* Offset, in bits, of this constraint from the beginning of
variables it ends up referring to.
IOW, in a deref constraint, we would deref, get the result set,
then add OFFSET to each member. */
HOST_WIDE_INT offset;
};
/* Use 0x8000... as special unknown offset. */
#define UNKNOWN_OFFSET HOST_WIDE_INT_MIN
typedef struct constraint_expr ce_s;
static void get_constraint_for_1 (tree, vec<ce_s> *, bool, bool);
static void get_constraint_for (tree, vec<ce_s> *);
static void get_constraint_for_rhs (tree, vec<ce_s> *);
static void do_deref (vec<ce_s> *);
/* Our set constraints are made up of two constraint expressions, one
LHS, and one RHS.
As described in the introduction, our set constraints each represent an
operation between set valued variables.
*/
struct constraint
{
struct constraint_expr lhs;
struct constraint_expr rhs;
};
/* List of constraints that we use to build the constraint graph from. */
static vec<constraint_t> constraints;
static object_allocator<constraint> constraint_pool ("Constraint pool");
/* The constraint graph is represented as an array of bitmaps
containing successor nodes. */
struct constraint_graph
{
/* Size of this graph, which may be different than the number of
nodes in the variable map. */
unsigned int size;
/* Explicit successors of each node. */
bitmap *succs;
/* Implicit predecessors of each node (Used for variable
substitution). */
bitmap *implicit_preds;
/* Explicit predecessors of each node (Used for variable substitution). */
bitmap *preds;
/* Indirect cycle representatives, or -1 if the node has no indirect
cycles. */
int *indirect_cycles;
/* Representative node for a node. rep[a] == a unless the node has
been unified. */
unsigned int *rep;
/* Equivalence class representative for a label. This is used for
variable substitution. */
int *eq_rep;
/* Pointer equivalence label for a node. All nodes with the same
pointer equivalence label can be unified together at some point
(either during constraint optimization or after the constraint
graph is built). */
unsigned int *pe;
/* Pointer equivalence representative for a label. This is used to
handle nodes that are pointer equivalent but not location
equivalent. We can unite these once the addressof constraints
are transformed into initial points-to sets. */
int *pe_rep;
/* Pointer equivalence label for each node, used during variable
substitution. */
unsigned int *pointer_label;
/* Location equivalence label for each node, used during location
equivalence finding. */
unsigned int *loc_label;
/* Pointed-by set for each node, used during location equivalence
finding. This is pointed-by rather than pointed-to, because it
is constructed using the predecessor graph. */
bitmap *pointed_by;
/* Points to sets for pointer equivalence. This is *not* the actual
points-to sets for nodes. */
bitmap *points_to;
/* Bitmap of nodes where the bit is set if the node is a direct
node. Used for variable substitution. */
sbitmap direct_nodes;
/* Bitmap of nodes where the bit is set if the node is address
taken. Used for variable substitution. */
bitmap address_taken;
/* Vector of complex constraints for each graph node. Complex
constraints are those involving dereferences or offsets that are
not 0. */
vec<constraint_t> *complex;
};
static constraint_graph_t graph;
/* During variable substitution and the offline version of indirect
cycle finding, we create nodes to represent dereferences and
address taken constraints. These represent where these start and
end. */
#define FIRST_REF_NODE (varmap).length ()
#define LAST_REF_NODE (FIRST_REF_NODE + (FIRST_REF_NODE - 1))
/* Return the representative node for NODE, if NODE has been unioned
with another NODE.
This function performs path compression along the way to finding
the representative. */
static unsigned int
find (unsigned int node)
{
gcc_checking_assert (node < graph->size);
if (graph->rep[node] != node)
return graph->rep[node] = find (graph->rep[node]);
return node;
}
/* Union the TO and FROM nodes to the TO nodes.
Note that at some point in the future, we may want to do
union-by-rank, in which case we are going to have to return the
node we unified to. */
static bool
unite (unsigned int to, unsigned int from)
{
gcc_checking_assert (to < graph->size && from < graph->size);
if (to != from && graph->rep[from] != to)
{
graph->rep[from] = to;
return true;
}
return false;
}
/* Create a new constraint consisting of LHS and RHS expressions. */
static constraint_t
new_constraint (const struct constraint_expr lhs,
const struct constraint_expr rhs)
{
constraint_t ret = constraint_pool.allocate ();
ret->lhs = lhs;
ret->rhs = rhs;
return ret;
}
/* Print out constraint C to FILE. */
static void
dump_constraint (FILE *file, constraint_t c)
{
if (c->lhs.type == ADDRESSOF)
fprintf (file, "&");
else if (c->lhs.type == DEREF)
fprintf (file, "*");
if (dump_file)
fprintf (file, "%s", get_varinfo (c->lhs.var)->name);
else
fprintf (file, "V%d", c->lhs.var);
if (c->lhs.offset == UNKNOWN_OFFSET)
fprintf (file, " + UNKNOWN");
else if (c->lhs.offset != 0)
fprintf (file, " + " HOST_WIDE_INT_PRINT_DEC, c->lhs.offset);
fprintf (file, " = ");
if (c->rhs.type == ADDRESSOF)
fprintf (file, "&");
else if (c->rhs.type == DEREF)
fprintf (file, "*");
if (dump_file)
fprintf (file, "%s", get_varinfo (c->rhs.var)->name);
else
fprintf (file, "V%d", c->rhs.var);
if (c->rhs.offset == UNKNOWN_OFFSET)
fprintf (file, " + UNKNOWN");
else if (c->rhs.offset != 0)
fprintf (file, " + " HOST_WIDE_INT_PRINT_DEC, c->rhs.offset);
}
void debug_constraint (constraint_t);
void debug_constraints (void);
void debug_constraint_graph (void);
void debug_solution_for_var (unsigned int);
void debug_sa_points_to_info (void);
void debug_varinfo (varinfo_t);
void debug_varmap (void);
/* Print out constraint C to stderr. */
DEBUG_FUNCTION void
debug_constraint (constraint_t c)
{
dump_constraint (stderr, c);
fprintf (stderr, "\n");
}
/* Print out all constraints to FILE */
static void
dump_constraints (FILE *file, int from)
{
int i;
constraint_t c;
for (i = from; constraints.iterate (i, &c); i++)
if (c)
{
dump_constraint (file, c);
fprintf (file, "\n");
}
}
/* Print out all constraints to stderr. */
DEBUG_FUNCTION void
debug_constraints (void)
{
dump_constraints (stderr, 0);
}
/* Print the constraint graph in dot format. */
static void
dump_constraint_graph (FILE *file)
{
unsigned int i;
/* Only print the graph if it has already been initialized: */
if (!graph)
return;
/* Prints the header of the dot file: */
fprintf (file, "strict digraph {\n");
fprintf (file, " node [\n shape = box\n ]\n");
fprintf (file, " edge [\n fontsize = \"12\"\n ]\n");
fprintf (file, "\n // List of nodes and complex constraints in "
"the constraint graph:\n");
/* The next lines print the nodes in the graph together with the
complex constraints attached to them. */
for (i = 1; i < graph->size; i++)
{
if (i == FIRST_REF_NODE)
continue;
if (find (i) != i)
continue;
if (i < FIRST_REF_NODE)
fprintf (file, "\"%s\"", get_varinfo (i)->name);
else
fprintf (file, "\"*%s\"", get_varinfo (i - FIRST_REF_NODE)->name);
if (graph->complex[i].exists ())
{
unsigned j;
constraint_t c;
fprintf (file, " [label=\"\\N\\n");
for (j = 0; graph->complex[i].iterate (j, &c); ++j)
{
dump_constraint (file, c);
fprintf (file, "\\l");
}
fprintf (file, "\"]");
}
fprintf (file, ";\n");
}
/* Go over the edges. */
fprintf (file, "\n // Edges in the constraint graph:\n");
for (i = 1; i < graph->size; i++)
{
unsigned j;
bitmap_iterator bi;
if (find (i) != i)
continue;
EXECUTE_IF_IN_NONNULL_BITMAP (graph->succs[i], 0, j, bi)
{
unsigned to = find (j);
if (i == to)
continue;
if (i < FIRST_REF_NODE)
fprintf (file, "\"%s\"", get_varinfo (i)->name);
else
fprintf (file, "\"*%s\"", get_varinfo (i - FIRST_REF_NODE)->name);
fprintf (file, " -> ");
if (to < FIRST_REF_NODE)
fprintf (file, "\"%s\"", get_varinfo (to)->name);
else
fprintf (file, "\"*%s\"", get_varinfo (to - FIRST_REF_NODE)->name);
fprintf (file, ";\n");
}
}
/* Prints the tail of the dot file. */
fprintf (file, "}\n");
}
/* Print out the constraint graph to stderr. */
DEBUG_FUNCTION void
debug_constraint_graph (void)
{
dump_constraint_graph (stderr);
}
/* SOLVER FUNCTIONS
The solver is a simple worklist solver, that works on the following
algorithm:
sbitmap changed_nodes = all zeroes;
changed_count = 0;
For each node that is not already collapsed:
changed_count++;
set bit in changed nodes
while (changed_count > 0)
{
compute topological ordering for constraint graph
find and collapse cycles in the constraint graph (updating
changed if necessary)
for each node (n) in the graph in topological order:
changed_count--;
Process each complex constraint associated with the node,
updating changed if necessary.
For each outgoing edge from n, propagate the solution from n to
the destination of the edge, updating changed as necessary.
} */
/* Return true if two constraint expressions A and B are equal. */
static bool
constraint_expr_equal (struct constraint_expr a, struct constraint_expr b)
{
return a.type == b.type && a.var == b.var && a.offset == b.offset;
}
/* Return true if constraint expression A is less than constraint expression
B. This is just arbitrary, but consistent, in order to give them an
ordering. */
static bool
constraint_expr_less (struct constraint_expr a, struct constraint_expr b)
{
if (a.type == b.type)
{
if (a.var == b.var)
return a.offset < b.offset;
else
return a.var < b.var;
}
else
return a.type < b.type;
}
/* Return true if constraint A is less than constraint B. This is just
arbitrary, but consistent, in order to give them an ordering. */
static bool
constraint_less (const constraint_t &a, const constraint_t &b)
{
if (constraint_expr_less (a->lhs, b->lhs))
return true;
else if (constraint_expr_less (b->lhs, a->lhs))
return false;
else
return constraint_expr_less (a->rhs, b->rhs);
}
/* Return true if two constraints A and B are equal. */
static bool
constraint_equal (struct constraint a, struct constraint b)
{
return constraint_expr_equal (a.lhs, b.lhs)
&& constraint_expr_equal (a.rhs, b.rhs);
}
/* Find a constraint LOOKFOR in the sorted constraint vector VEC */
static constraint_t
constraint_vec_find (vec<constraint_t> vec,
struct constraint lookfor)
{
unsigned int place;
constraint_t found;
if (!vec.exists ())
return NULL;
place = vec.lower_bound (&lookfor, constraint_less);
if (place >= vec.length ())
return NULL;
found = vec[place];
if (!constraint_equal (*found, lookfor))
return NULL;
return found;
}
/* Union two constraint vectors, TO and FROM. Put the result in TO.
Returns true of TO set is changed. */
static bool
constraint_set_union (vec<constraint_t> *to,
vec<constraint_t> *from)
{
int i;
constraint_t c;
bool any_change = false;
FOR_EACH_VEC_ELT (*from, i, c)
{
if (constraint_vec_find (*to, *c) == NULL)
{
unsigned int place = to->lower_bound (c, constraint_less);
to->safe_insert (place, c);
any_change = true;
}
}
return any_change;
}
/* Expands the solution in SET to all sub-fields of variables included. */
static bitmap
solution_set_expand (bitmap set, bitmap *expanded)
{
bitmap_iterator bi;
unsigned j;
if (*expanded)
return *expanded;
*expanded = BITMAP_ALLOC (&iteration_obstack);
/* In a first pass expand to the head of the variables we need to
add all sub-fields off. This avoids quadratic behavior. */
EXECUTE_IF_SET_IN_BITMAP (set, 0, j, bi)
{
varinfo_t v = get_varinfo (j);
if (v->is_artificial_var
|| v->is_full_var)
continue;
bitmap_set_bit (*expanded, v->head);
}
/* In the second pass now expand all head variables with subfields. */
EXECUTE_IF_SET_IN_BITMAP (*expanded, 0, j, bi)
{
varinfo_t v = get_varinfo (j);
if (v->head != j)
continue;
for (v = vi_next (v); v != NULL; v = vi_next (v))
bitmap_set_bit (*expanded, v->id);
}
/* And finally set the rest of the bits from SET. */
bitmap_ior_into (*expanded, set);
return *expanded;
}
/* Union solution sets TO and DELTA, and add INC to each member of DELTA in the
process. */
static bool
set_union_with_increment (bitmap to, bitmap delta, HOST_WIDE_INT inc,
bitmap *expanded_delta)
{
bool changed = false;
bitmap_iterator bi;
unsigned int i;
/* If the solution of DELTA contains anything it is good enough to transfer
this to TO. */
if (bitmap_bit_p (delta, anything_id))
return bitmap_set_bit (to, anything_id);
/* If the offset is unknown we have to expand the solution to
all subfields. */
if (inc == UNKNOWN_OFFSET)
{
delta = solution_set_expand (delta, expanded_delta);
changed |= bitmap_ior_into (to, delta);
return changed;
}
/* For non-zero offset union the offsetted solution into the destination. */
EXECUTE_IF_SET_IN_BITMAP (delta, 0, i, bi)
{
varinfo_t vi = get_varinfo (i);
/* If this is a variable with just one field just set its bit
in the result. */
if (vi->is_artificial_var
|| vi->is_unknown_size_var
|| vi->is_full_var)
changed |= bitmap_set_bit (to, i);
else
{
HOST_WIDE_INT fieldoffset = vi->offset + inc;
unsigned HOST_WIDE_INT size = vi->size;
/* If the offset makes the pointer point to before the
variable use offset zero for the field lookup. */
if (fieldoffset < 0)
vi = get_varinfo (vi->head);
else
vi = first_or_preceding_vi_for_offset (vi, fieldoffset);
do
{
changed |= bitmap_set_bit (to, vi->id);
if (vi->is_full_var
|| vi->next == 0)
break;
/* We have to include all fields that overlap the current field
shifted by inc. */
vi = vi_next (vi);
}
while (vi->offset < fieldoffset + size);
}
}
return changed;
}
/* Insert constraint C into the list of complex constraints for graph
node VAR. */
static void
insert_into_complex (constraint_graph_t graph,
unsigned int var, constraint_t c)
{
vec<constraint_t> complex = graph->complex[var];
unsigned int place = complex.lower_bound (c, constraint_less);
/* Only insert constraints that do not already exist. */
if (place >= complex.length ()
|| !constraint_equal (*c, *complex[place]))
graph->complex[var].safe_insert (place, c);
}
/* Condense two variable nodes into a single variable node, by moving
all associated info from FROM to TO. Returns true if TO node's
constraint set changes after the merge. */
static bool
merge_node_constraints (constraint_graph_t graph, unsigned int to,
unsigned int from)
{
unsigned int i;
constraint_t c;
bool any_change = false;
gcc_checking_assert (find (from) == to);
/* Move all complex constraints from src node into to node */
FOR_EACH_VEC_ELT (graph->complex[from], i, c)
{
/* In complex constraints for node FROM, we may have either
a = *FROM, and *FROM = a, or an offseted constraint which are
always added to the rhs node's constraints. */
if (c->rhs.type == DEREF)
c->rhs.var = to;
else if (c->lhs.type == DEREF)
c->lhs.var = to;
else
c->rhs.var = to;
}
any_change = constraint_set_union (&graph->complex[to],
&graph->complex[from]);
graph->complex[from].release ();
return any_change;
}
/* Remove edges involving NODE from GRAPH. */
static void
clear_edges_for_node (constraint_graph_t graph, unsigned int node)
{
if (graph->succs[node])
BITMAP_FREE (graph->succs[node]);
}
/* Merge GRAPH nodes FROM and TO into node TO. */
static void
merge_graph_nodes (constraint_graph_t graph, unsigned int to,
unsigned int from)
{
if (graph->indirect_cycles[from] != -1)
{
/* If we have indirect cycles with the from node, and we have
none on the to node, the to node has indirect cycles from the
from node now that they are unified.
If indirect cycles exist on both, unify the nodes that they
are in a cycle with, since we know they are in a cycle with
each other. */
if (graph->indirect_cycles[to] == -1)
graph->indirect_cycles[to] = graph->indirect_cycles[from];
}
/* Merge all the successor edges. */
if (graph->succs[from])
{
if (!graph->succs[to])
graph->succs[to] = BITMAP_ALLOC (&pta_obstack);
bitmap_ior_into (graph->succs[to],
graph->succs[from]);
}
clear_edges_for_node (graph, from);
}
/* Add an indirect graph edge to GRAPH, going from TO to FROM if
it doesn't exist in the graph already. */
static void
add_implicit_graph_edge (constraint_graph_t graph, unsigned int to,
unsigned int from)
{
if (to == from)
return;
if (!graph->implicit_preds[to])
graph->implicit_preds[to] = BITMAP_ALLOC (&predbitmap_obstack);
if (bitmap_set_bit (graph->implicit_preds[to], from))
stats.num_implicit_edges++;
}
/* Add a predecessor graph edge to GRAPH, going from TO to FROM if
it doesn't exist in the graph already.
Return false if the edge already existed, true otherwise. */
static void
add_pred_graph_edge (constraint_graph_t graph, unsigned int to,
unsigned int from)
{
if (!graph->preds[to])
graph->preds[to] = BITMAP_ALLOC (&predbitmap_obstack);
bitmap_set_bit (graph->preds[to], from);
}
/* Add a graph edge to GRAPH, going from FROM to TO if
it doesn't exist in the graph already.
Return false if the edge already existed, true otherwise. */
static bool
add_graph_edge (constraint_graph_t graph, unsigned int to,
unsigned int from)
{
if (to == from)
{
return false;
}
else
{
bool r = false;
if (!graph->succs[from])
graph->succs[from] = BITMAP_ALLOC (&pta_obstack);
/* The graph solving process does not avoid "triangles", thus
there can be multiple paths from a node to another involving
intermediate other nodes. That causes extra copying which is
most difficult to avoid when the intermediate node is ESCAPED
because there are no edges added from ESCAPED. Avoid
adding the direct edge FROM -> TO when we have FROM -> ESCAPED
and TO contains ESCAPED.
??? Note this is only a heuristic, it does not prevent the
situation from occuring. The heuristic helps PR38474 and
PR99912 significantly. */
if (to < FIRST_REF_NODE
&& bitmap_bit_p (graph->succs[from], find (escaped_id))
&& bitmap_bit_p (get_varinfo (find (to))->solution, escaped_id))
return false;
if (bitmap_set_bit (graph->succs[from], to))
{
r = true;
if (to < FIRST_REF_NODE && from < FIRST_REF_NODE)
stats.num_edges++;
}
return r;
}
}
/* Initialize the constraint graph structure to contain SIZE nodes. */
static void
init_graph (unsigned int size)
{
unsigned int j;
graph = XCNEW (struct constraint_graph);
graph->size = size;
graph->succs = XCNEWVEC (bitmap, graph->size);
graph->indirect_cycles = XNEWVEC (int, graph->size);
graph->rep = XNEWVEC (unsigned int, graph->size);
/* ??? Macros do not support template types with multiple arguments,
so we use a typedef to work around it. */
typedef vec<constraint_t> vec_constraint_t_heap;
graph->complex = XCNEWVEC (vec_constraint_t_heap, size);
graph->pe = XCNEWVEC (unsigned int, graph->size);
graph->pe_rep = XNEWVEC (int, graph->size);
for (j = 0; j < graph->size; j++)
{
graph->rep[j] = j;
graph->pe_rep[j] = -1;
graph->indirect_cycles[j] = -1;
}
}
/* Build the constraint graph, adding only predecessor edges right now. */
static void
build_pred_graph (void)
{
int i;
constraint_t c;
unsigned int j;
graph->implicit_preds = XCNEWVEC (bitmap, graph->size);
graph->preds = XCNEWVEC (bitmap, graph->size);
graph->pointer_label = XCNEWVEC (unsigned int, graph->size);
graph->loc_label = XCNEWVEC (unsigned int, graph->size);
graph->pointed_by = XCNEWVEC (bitmap, graph->size);
graph->points_to = XCNEWVEC (bitmap, graph->size);
graph->eq_rep = XNEWVEC (int, graph->size);
graph->direct_nodes = sbitmap_alloc (graph->size);
graph->address_taken = BITMAP_ALLOC (&predbitmap_obstack);
bitmap_clear (graph->direct_nodes);
for (j = 1; j < FIRST_REF_NODE; j++)
{
if (!get_varinfo (j)->is_special_var)
bitmap_set_bit (graph->direct_nodes, j);
}
for (j = 0; j < graph->size; j++)
graph->eq_rep[j] = -1;
for (j = 0; j < varmap.length (); j++)
graph->indirect_cycles[j] = -1;
FOR_EACH_VEC_ELT (constraints, i, c)
{
struct constraint_expr lhs = c->lhs;
struct constraint_expr rhs = c->rhs;
unsigned int lhsvar = lhs.var;
unsigned int rhsvar = rhs.var;
if (lhs.type == DEREF)
{
/* *x = y. */
if (rhs.offset == 0 && lhs.offset == 0 && rhs.type == SCALAR)
add_pred_graph_edge (graph, FIRST_REF_NODE + lhsvar, rhsvar);
}
else if (rhs.type == DEREF)
{
/* x = *y */
if (rhs.offset == 0 && lhs.offset == 0 && lhs.type == SCALAR)
add_pred_graph_edge (graph, lhsvar, FIRST_REF_NODE + rhsvar);
else
bitmap_clear_bit (graph->direct_nodes, lhsvar);
}
else if (rhs.type == ADDRESSOF)
{
varinfo_t v;
/* x = &y */
if (graph->points_to[lhsvar] == NULL)
graph->points_to[lhsvar] = BITMAP_ALLOC (&predbitmap_obstack);
bitmap_set_bit (graph->points_to[lhsvar], rhsvar);
if (graph->pointed_by[rhsvar] == NULL)
graph->pointed_by[rhsvar] = BITMAP_ALLOC (&predbitmap_obstack);
bitmap_set_bit (graph->pointed_by[rhsvar], lhsvar);
/* Implicitly, *x = y */
add_implicit_graph_edge (graph, FIRST_REF_NODE + lhsvar, rhsvar);
/* All related variables are no longer direct nodes. */
bitmap_clear_bit (graph->direct_nodes, rhsvar);
v = get_varinfo (rhsvar);
if (!v->is_full_var)
{
v = get_varinfo (v->head);
do
{
bitmap_clear_bit (graph->direct_nodes, v->id);
v = vi_next (v);
}
while (v != NULL);
}
bitmap_set_bit (graph->address_taken, rhsvar);
}
else if (lhsvar > anything_id
&& lhsvar != rhsvar && lhs.offset == 0 && rhs.offset == 0)
{
/* x = y */
add_pred_graph_edge (graph, lhsvar, rhsvar);
/* Implicitly, *x = *y */
add_implicit_graph_edge (graph, FIRST_REF_NODE + lhsvar,
FIRST_REF_NODE + rhsvar);
}
else if (lhs.offset != 0 || rhs.offset != 0)
{
if (rhs.offset != 0)
bitmap_clear_bit (graph->direct_nodes, lhs.var);
else if (lhs.offset != 0)
bitmap_clear_bit (graph->direct_nodes, rhs.var);
}
}
}
/* Build the constraint graph, adding successor edges. */
static void
build_succ_graph (void)
{
unsigned i, t;
constraint_t c;
FOR_EACH_VEC_ELT (constraints, i, c)
{
struct constraint_expr lhs;
struct constraint_expr rhs;
unsigned int lhsvar;
unsigned int rhsvar;
if (!c)
continue;
lhs = c->lhs;
rhs = c->rhs;
lhsvar = find (lhs.var);
rhsvar = find (rhs.var);
if (lhs.type == DEREF)
{
if (rhs.offset == 0 && lhs.offset == 0 && rhs.type == SCALAR)
add_graph_edge (graph, FIRST_REF_NODE + lhsvar, rhsvar);
}
else if (rhs.type == DEREF)
{
if (rhs.offset == 0 && lhs.offset == 0 && lhs.type == SCALAR)
add_graph_edge (graph, lhsvar, FIRST_REF_NODE + rhsvar);
}
else if (rhs.type == ADDRESSOF)
{
/* x = &y */
gcc_checking_assert (find (rhs.var) == rhs.var);
bitmap_set_bit (get_varinfo (lhsvar)->solution, rhsvar);
}
else if (lhsvar > anything_id
&& lhsvar != rhsvar && lhs.offset == 0 && rhs.offset == 0)
{
add_graph_edge (graph, lhsvar, rhsvar);
}
}
/* Add edges from STOREDANYTHING to all non-direct nodes that can
receive pointers. */
t = find (storedanything_id);
for (i = integer_id + 1; i < FIRST_REF_NODE; ++i)
{
if (!bitmap_bit_p (graph->direct_nodes, i)
&& get_varinfo (i)->may_have_pointers)
add_graph_edge (graph, find (i), t);
}
/* Everything stored to ANYTHING also potentially escapes. */
add_graph_edge (graph, find (escaped_id), t);
}
/* Changed variables on the last iteration. */
static bitmap changed;
/* Strongly Connected Component visitation info. */
class scc_info
{
public:
scc_info (size_t size);
~scc_info ();
auto_sbitmap visited;
auto_sbitmap deleted;
unsigned int *dfs;
unsigned int *node_mapping;
int current_index;
auto_vec<unsigned> scc_stack;
};
/* Recursive routine to find strongly connected components in GRAPH.
SI is the SCC info to store the information in, and N is the id of current
graph node we are processing.
This is Tarjan's strongly connected component finding algorithm, as
modified by Nuutila to keep only non-root nodes on the stack.
The algorithm can be found in "On finding the strongly connected
connected components in a directed graph" by Esko Nuutila and Eljas
Soisalon-Soininen, in Information Processing Letters volume 49,
number 1, pages 9-14. */
static void
scc_visit (constraint_graph_t graph, class scc_info *si, unsigned int n)
{
unsigned int i;
bitmap_iterator bi;
unsigned int my_dfs;
bitmap_set_bit (si->visited, n);
si->dfs[n] = si->current_index ++;
my_dfs = si->dfs[n];
/* Visit all the successors. */
EXECUTE_IF_IN_NONNULL_BITMAP (graph->succs[n], 0, i, bi)
{
unsigned int w;
if (i > LAST_REF_NODE)
break;
w = find (i);
if (bitmap_bit_p (si->deleted, w))
continue;
if (!bitmap_bit_p (si->visited, w))
scc_visit (graph, si, w);
unsigned int t = find (w);
gcc_checking_assert (find (n) == n);
if (si->dfs[t] < si->dfs[n])
si->dfs[n] = si->dfs[t];
}
/* See if any components have been identified. */
if (si->dfs[n] == my_dfs)
{
if (si->scc_stack.length () > 0
&& si->dfs[si->scc_stack.last ()] >= my_dfs)
{
bitmap scc = BITMAP_ALLOC (NULL);
unsigned int lowest_node;
bitmap_iterator bi;
bitmap_set_bit (scc, n);
while (si->scc_stack.length () != 0
&& si->dfs[si->scc_stack.last ()] >= my_dfs)
{
unsigned int w = si->scc_stack.pop ();
bitmap_set_bit (scc, w);
}
lowest_node = bitmap_first_set_bit (scc);
gcc_assert (lowest_node < FIRST_REF_NODE);
/* Collapse the SCC nodes into a single node, and mark the
indirect cycles. */
EXECUTE_IF_SET_IN_BITMAP (scc, 0, i, bi)
{
if (i < FIRST_REF_NODE)
{
if (unite (lowest_node, i))
unify_nodes (graph, lowest_node, i, false);
}
else
{
unite (lowest_node, i);
graph->indirect_cycles[i - FIRST_REF_NODE] = lowest_node;
}
}
}
bitmap_set_bit (si->deleted, n);
}
else
si->scc_stack.safe_push (n);
}
/* Unify node FROM into node TO, updating the changed count if
necessary when UPDATE_CHANGED is true. */
static void
unify_nodes (constraint_graph_t graph, unsigned int to, unsigned int from,
bool update_changed)
{
gcc_checking_assert (to != from && find (to) == to);
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "Unifying %s to %s\n",
get_varinfo (from)->name,
get_varinfo (to)->name);
if (update_changed)
stats.unified_vars_dynamic++;
else
stats.unified_vars_static++;
merge_graph_nodes (graph, to, from);
if (merge_node_constraints (graph, to, from))
{
if (update_changed)
bitmap_set_bit (changed, to);
}
/* Mark TO as changed if FROM was changed. If TO was already marked
as changed, decrease the changed count. */
if (update_changed
&& bitmap_clear_bit (changed, from))
bitmap_set_bit (changed, to);
varinfo_t fromvi = get_varinfo (from);
if (fromvi->solution)
{
/* If the solution changes because of the merging, we need to mark
the variable as changed. */
varinfo_t tovi = get_varinfo (to);
if (bitmap_ior_into (tovi->solution, fromvi->solution))
{
if (update_changed)
bitmap_set_bit (changed, to);
}
BITMAP_FREE (fromvi->solution);
if (fromvi->oldsolution)
BITMAP_FREE (fromvi->oldsolution);
if (stats.iterations > 0
&& tovi->oldsolution)
BITMAP_FREE (tovi->oldsolution);
}
if (graph->succs[to])
bitmap_clear_bit (graph->succs[to], to);
}
/* Information needed to compute the topological ordering of a graph. */
struct topo_info
{
/* sbitmap of visited nodes. */
sbitmap visited;
/* Array that stores the topological order of the graph, *in
reverse*. */
vec<unsigned> topo_order;
};
/* Initialize and return a topological info structure. */
static struct topo_info *
init_topo_info (void)
{
size_t size = graph->size;
struct topo_info *ti = XNEW (struct topo_info);
ti->visited = sbitmap_alloc (size);
bitmap_clear (ti->visited);
ti->topo_order.create (1);
return ti;
}
/* Free the topological sort info pointed to by TI. */
static void
free_topo_info (struct topo_info *ti)
{
sbitmap_free (ti->visited);
ti->topo_order.release ();
free (ti);
}
/* Visit the graph in topological order, and store the order in the
topo_info structure. */
static void
topo_visit (constraint_graph_t graph, struct topo_info *ti,
unsigned int n)
{
bitmap_iterator bi;
unsigned int j;
bitmap_set_bit (ti->visited, n);
if (graph->succs[n])
EXECUTE_IF_SET_IN_BITMAP (graph->succs[n], 0, j, bi)
{
if (!bitmap_bit_p (ti->visited, j))
topo_visit (graph, ti, j);
}
ti->topo_order.safe_push (n);
}
/* Process a constraint C that represents x = *(y + off), using DELTA as the
starting solution for y. */
static void
do_sd_constraint (constraint_graph_t graph, constraint_t c,
bitmap delta, bitmap *expanded_delta)
{
unsigned int lhs = c->lhs.var;
bool flag = false;
bitmap sol = get_varinfo (lhs)->solution;
unsigned int j;
bitmap_iterator bi;
HOST_WIDE_INT roffset = c->rhs.offset;
/* Our IL does not allow this. */
gcc_checking_assert (c->lhs.offset == 0);
/* If the solution of Y contains anything it is good enough to transfer
this to the LHS. */
if (bitmap_bit_p (delta, anything_id))
{
flag |= bitmap_set_bit (sol, anything_id);
goto done;
}
/* If we do not know at with offset the rhs is dereferenced compute
the reachability set of DELTA, conservatively assuming it is
dereferenced at all valid offsets. */
if (roffset == UNKNOWN_OFFSET)
{
delta = solution_set_expand (delta, expanded_delta);
/* No further offset processing is necessary. */
roffset = 0;
}
/* For each variable j in delta (Sol(y)), add
an edge in the graph from j to x, and union Sol(j) into Sol(x). */
EXECUTE_IF_SET_IN_BITMAP (delta, 0, j, bi)
{
varinfo_t v = get_varinfo (j);
HOST_WIDE_INT fieldoffset = v->offset + roffset;
unsigned HOST_WIDE_INT size = v->size;
unsigned int t;
if (v->is_full_var)
;
else if (roffset != 0)
{
if (fieldoffset < 0)
v = get_varinfo (v->head);
else
v = first_or_preceding_vi_for_offset (v, fieldoffset);
}
/* We have to include all fields that overlap the current field
shifted by roffset. */
do
{
t = find (v->id);
/* Adding edges from the special vars is pointless.
They don't have sets that can change. */
if (get_varinfo (t)->is_special_var)
flag |= bitmap_ior_into (sol, get_varinfo (t)->solution);
/* Merging the solution from ESCAPED needlessly increases
the set. Use ESCAPED as representative instead. */
else if (v->id == escaped_id)
flag |= bitmap_set_bit (sol, escaped_id);
else if (v->may_have_pointers
&& add_graph_edge (graph, lhs, t))
flag |= bitmap_ior_into (sol, get_varinfo (t)->solution);
if (v->is_full_var
|| v->next == 0)
break;
v = vi_next (v);
}
while (v->offset < fieldoffset + size);
}
done:
/* If the LHS solution changed, mark the var as changed. */
if (flag)
{
get_varinfo (lhs)->solution = sol;
bitmap_set_bit (changed, lhs);
}
}
/* Process a constraint C that represents *(x + off) = y using DELTA
as the starting solution for x. */
static void
do_ds_constraint (constraint_t c, bitmap delta, bitmap *expanded_delta)
{
unsigned int rhs = c->rhs.var;
bitmap sol = get_varinfo (rhs)->solution;
unsigned int j;
bitmap_iterator bi;
HOST_WIDE_INT loff = c->lhs.offset;
bool escaped_p = false;
/* Our IL does not allow this. */
gcc_checking_assert (c->rhs.offset == 0);
/* If the solution of y contains ANYTHING simply use the ANYTHING
solution. This avoids needlessly increasing the points-to sets. */
if (bitmap_bit_p (sol, anything_id))
sol = get_varinfo (find (anything_id))->solution;
/* If the solution for x contains ANYTHING we have to merge the
solution of y into all pointer variables which we do via
STOREDANYTHING. */
if (bitmap_bit_p (delta, anything_id))
{
unsigned t = find (storedanything_id);
if (add_graph_edge (graph, t, rhs))
{
if (bitmap_ior_into (get_varinfo (t)->solution, sol))
bitmap_set_bit (changed, t);
}
return;
}
/* If we do not know at with offset the rhs is dereferenced compute
the reachability set of DELTA, conservatively assuming it is
dereferenced at all valid offsets. */
if (loff == UNKNOWN_OFFSET)
{
delta = solution_set_expand (delta, expanded_delta);
loff = 0;
}
/* For each member j of delta (Sol(x)), add an edge from y to j and
union Sol(y) into Sol(j) */
EXECUTE_IF_SET_IN_BITMAP (delta, 0, j, bi)
{
varinfo_t v = get_varinfo (j);
unsigned int t;
HOST_WIDE_INT fieldoffset = v->offset + loff;
unsigned HOST_WIDE_INT size = v->size;
if (v->is_full_var)
;
else if (loff != 0)
{
if (fieldoffset < 0)
v = get_varinfo (v->head);
else
v = first_or_preceding_vi_for_offset (v, fieldoffset);
}
/* We have to include all fields that overlap the current field
shifted by loff. */
do
{
if (v->may_have_pointers)
{
/* If v is a global variable then this is an escape point. */
if (v->is_global_var
&& !escaped_p)
{
t = find (escaped_id);
if (add_graph_edge (graph, t, rhs)
&& bitmap_ior_into (get_varinfo (t)->solution, sol))
bitmap_set_bit (changed, t);
/* Enough to let rhs escape once. */
escaped_p = true;
}
if (v->is_special_var)
break;
t = find (v->id);
if (add_graph_edge (graph, t, rhs)
&& bitmap_ior_into (get_varinfo (t)->solution, sol))
bitmap_set_bit (changed, t);
}
if (v->is_full_var
|| v->next == 0)
break;
v = vi_next (v);
}
while (v->offset < fieldoffset + size);
}
}
/* Handle a non-simple (simple meaning requires no iteration),
constraint (IE *x = &y, x = *y, *x = y, and x = y with offsets involved). */
static void
do_complex_constraint (constraint_graph_t graph, constraint_t c, bitmap delta,
bitmap *expanded_delta)
{
if (c->lhs.type == DEREF)
{
if (c->rhs.type == ADDRESSOF)
{
gcc_unreachable ();
}
else
{
/* *x = y */
do_ds_constraint (c, delta, expanded_delta);
}
}
else if (c->rhs.type == DEREF)
{
/* x = *y */
if (!(get_varinfo (c->lhs.var)->is_special_var))
do_sd_constraint (graph, c, delta, expanded_delta);
}
else
{
bitmap tmp;
bool flag = false;
gcc_checking_assert (c->rhs.type == SCALAR && c->lhs.type == SCALAR
&& c->rhs.offset != 0 && c->lhs.offset == 0);
tmp = get_varinfo (c->lhs.var)->solution;
flag = set_union_with_increment (tmp, delta, c->rhs.offset,
expanded_delta);
if (flag)
bitmap_set_bit (changed, c->lhs.var);
}
}
/* Initialize and return a new SCC info structure. */
scc_info::scc_info (size_t size) :
visited (size), deleted (size), current_index (0), scc_stack (1)
{
bitmap_clear (visited);
bitmap_clear (deleted);
node_mapping = XNEWVEC (unsigned int, size);
dfs = XCNEWVEC (unsigned int, size);
for (size_t i = 0; i < size; i++)
node_mapping[i] = i;
}
/* Free an SCC info structure pointed to by SI */
scc_info::~scc_info ()
{
free (node_mapping);
free (dfs);
}
/* Find indirect cycles in GRAPH that occur, using strongly connected
components, and note them in the indirect cycles map.
This technique comes from Ben Hardekopf and Calvin Lin,
"It Pays to be Lazy: Fast and Accurate Pointer Analysis for Millions of
Lines of Code", submitted to PLDI 2007. */
static void
find_indirect_cycles (constraint_graph_t graph)
{
unsigned int i;
unsigned int size = graph->size;
scc_info si (size);
for (i = 0; i < MIN (LAST_REF_NODE, size); i ++ )
if (!bitmap_bit_p (si.visited, i) && find (i) == i)
scc_visit (graph, &si, i);
}
/* Compute a topological ordering for GRAPH, and store the result in the
topo_info structure TI. */
static void
compute_topo_order (constraint_graph_t graph,
struct topo_info *ti)
{
unsigned int i;
unsigned int size = graph->size;
for (i = 0; i != size; ++i)
if (!bitmap_bit_p (ti->visited, i) && find (i) == i)
topo_visit (graph, ti, i);
}
/* Structure used to for hash value numbering of pointer equivalence
classes. */
typedef struct equiv_class_label
{
hashval_t hashcode;
unsigned int equivalence_class;
bitmap labels;
} *equiv_class_label_t;
typedef const struct equiv_class_label *const_equiv_class_label_t;
/* Equiv_class_label hashtable helpers. */
struct equiv_class_hasher : nofree_ptr_hash <equiv_class_label>
{
static inline hashval_t hash (const equiv_class_label *);
static inline bool equal (const equiv_class_label *,
const equiv_class_label *);
};
/* Hash function for a equiv_class_label_t */
inline hashval_t
equiv_class_hasher::hash (const equiv_class_label *ecl)
{
return ecl->hashcode;
}
/* Equality function for two equiv_class_label_t's. */
inline bool
equiv_class_hasher::equal (const equiv_class_label *eql1,
const equiv_class_label *eql2)
{
return (eql1->hashcode == eql2->hashcode
&& bitmap_equal_p (eql1->labels, eql2->labels));
}
/* A hashtable for mapping a bitmap of labels->pointer equivalence
classes. */
static hash_table<equiv_class_hasher> *pointer_equiv_class_table;
/* A hashtable for mapping a bitmap of labels->location equivalence
classes. */
static hash_table<equiv_class_hasher> *location_equiv_class_table;
struct obstack equiv_class_obstack;
/* Lookup a equivalence class in TABLE by the bitmap of LABELS with
hash HAS it contains. Sets *REF_LABELS to the bitmap LABELS
is equivalent to. */
static equiv_class_label *
equiv_class_lookup_or_add (hash_table<equiv_class_hasher> *table,
bitmap labels)
{
equiv_class_label **slot;
equiv_class_label ecl;
ecl.labels = labels;
ecl.hashcode = bitmap_hash (labels);
slot = table->find_slot (&ecl, INSERT);
if (!*slot)
{
*slot = XOBNEW (&equiv_class_obstack, struct equiv_class_label);
(*slot)->labels = labels;
(*slot)->hashcode = ecl.hashcode;
(*slot)->equivalence_class = 0;
}
return *slot;
}
/* Perform offline variable substitution.
This is a worst case quadratic time way of identifying variables
that must have equivalent points-to sets, including those caused by
static cycles, and single entry subgraphs, in the constraint graph.
The technique is described in "Exploiting Pointer and Location
Equivalence to Optimize Pointer Analysis. In the 14th International
Static Analysis Symposium (SAS), August 2007." It is known as the
"HU" algorithm, and is equivalent to value numbering the collapsed
constraint graph including evaluating unions.
The general method of finding equivalence classes is as follows:
Add fake nodes (REF nodes) and edges for *a = b and a = *b constraints.
Initialize all non-REF nodes to be direct nodes.
For each constraint a = a U {b}, we set pts(a) = pts(a) u {fresh
variable}
For each constraint containing the dereference, we also do the same
thing.
We then compute SCC's in the graph and unify nodes in the same SCC,
including pts sets.
For each non-collapsed node x:
Visit all unvisited explicit incoming edges.
Ignoring all non-pointers, set pts(x) = Union of pts(a) for y
where y->x.
Lookup the equivalence class for pts(x).
If we found one, equivalence_class(x) = found class.
Otherwise, equivalence_class(x) = new class, and new_class is
added to the lookup table.
All direct nodes with the same equivalence class can be replaced
with a single representative node.
All unlabeled nodes (label == 0) are not pointers and all edges
involving them can be eliminated.
We perform these optimizations during rewrite_constraints
In addition to pointer equivalence class finding, we also perform
location equivalence class finding. This is the set of variables
that always appear together in points-to sets. We use this to
compress the size of the points-to sets. */
/* Current maximum pointer equivalence class id. */
static int pointer_equiv_class;
/* Current maximum location equivalence class id. */
static int location_equiv_class;
/* Recursive routine to find strongly connected components in GRAPH,
and label it's nodes with DFS numbers. */
static void
condense_visit (constraint_graph_t graph, class scc_info *si, unsigned int n)
{
unsigned int i;
bitmap_iterator bi;
unsigned int my_dfs;
gcc_checking_assert (si->node_mapping[n] == n);
bitmap_set_bit (si->visited, n);
si->dfs[n] = si->current_index ++;
my_dfs = si->dfs[n];
/* Visit all the successors. */
EXECUTE_IF_IN_NONNULL_BITMAP (graph->preds[n], 0, i, bi)
{
unsigned int w = si->node_mapping[i];
if (bitmap_bit_p (si->deleted, w))
continue;
if (!bitmap_bit_p (si->visited, w))
condense_visit (graph, si, w);
unsigned int t = si->node_mapping[w];
gcc_checking_assert (si->node_mapping[n] == n);
if (si->dfs[t] < si->dfs[n])
si->dfs[n] = si->dfs[t];
}
/* Visit all the implicit predecessors. */
EXECUTE_IF_IN_NONNULL_BITMAP (graph->implicit_preds[n], 0, i, bi)
{
unsigned int w = si->node_mapping[i];
if (bitmap_bit_p (si->deleted, w))
continue;
if (!bitmap_bit_p (si->visited, w))
condense_visit (graph, si, w);
unsigned int t = si->node_mapping[w];
gcc_assert (si->node_mapping[n] == n);
if (si->dfs[t] < si->dfs[n])
si->dfs[n] = si->dfs[t];
}
/* See if any components have been identified. */
if (si->dfs[n] == my_dfs)
{
if (si->scc_stack.length () != 0
&& si->dfs[si->scc_stack.last ()] >= my_dfs)
{
/* Find the first node of the SCC and do non-bitmap work. */
bool direct_p = true;
unsigned first = si->scc_stack.length ();
do
{
--first;
unsigned int w = si->scc_stack[first];
si->node_mapping[w] = n;
if (!bitmap_bit_p (graph->direct_nodes, w))
direct_p = false;
}
while (first > 0
&& si->dfs[si->scc_stack[first - 1]] >= my_dfs);
if (!direct_p)
bitmap_clear_bit (graph->direct_nodes, n);
/* Want to reduce to node n, push that first. */
si->scc_stack.reserve (1);
si->scc_stack.quick_push (si->scc_stack[first]);
si->scc_stack[first] = n;
unsigned scc_size = si->scc_stack.length () - first;
unsigned split = scc_size / 2;
unsigned carry = scc_size - split * 2;
while (split > 0)
{
for (unsigned i = 0; i < split; ++i)
{
unsigned a = si->scc_stack[first + i];
unsigned b = si->scc_stack[first + split + carry + i];
/* Unify our nodes. */
if (graph->preds[b])
{
if (!graph->preds[a])
std::swap (graph->preds[a], graph->preds[b]);
else
bitmap_ior_into_and_free (graph->preds[a],
&graph->preds[b]);
}
if (graph->implicit_preds[b])
{
if (!graph->implicit_preds[a])
std::swap (graph->implicit_preds[a],
graph->implicit_preds[b]);
else
bitmap_ior_into_and_free (graph->implicit_preds[a],
&graph->implicit_preds[b]);
}
if (graph->points_to[b])
{
if (!graph->points_to[a])
std::swap (graph->points_to[a], graph->points_to[b]);
else
bitmap_ior_into_and_free (graph->points_to[a],
&graph->points_to[b]);
}
}
unsigned remain = split + carry;
split = remain / 2;
carry = remain - split * 2;
}
/* Actually pop the SCC. */
si->scc_stack.truncate (first);
}
bitmap_set_bit (si->deleted, n);
}
else
si->scc_stack.safe_push (n);
}
/* Label pointer equivalences.
This performs a value numbering of the constraint graph to
discover which variables will always have the same points-to sets
under the current set of constraints.
The way it value numbers is to store the set of points-to bits
generated by the constraints and graph edges. This is just used as a
hash and equality comparison. The *actual set of points-to bits* is
completely irrelevant, in that we don't care about being able to
extract them later.
The equality values (currently bitmaps) just have to satisfy a few
constraints, the main ones being:
1. The combining operation must be order independent.
2. The end result of a given set of operations must be unique iff the
combination of input values is unique
3. Hashable. */
static void
label_visit (constraint_graph_t graph, class scc_info *si, unsigned int n)
{
unsigned int i, first_pred;
bitmap_iterator bi;
bitmap_set_bit (si->visited, n);
/* Label and union our incoming edges's points to sets. */
first_pred = -1U;
EXECUTE_IF_IN_NONNULL_BITMAP (graph->preds[n], 0, i, bi)
{
unsigned int w = si->node_mapping[i];
if (!bitmap_bit_p (si->visited, w))
label_visit (graph, si, w);
/* Skip unused edges */
if (w == n || graph->pointer_label[w] == 0)
continue;
if (graph->points_to[w])
{
if (!graph->points_to[n])
{
if (first_pred == -1U)
first_pred = w;
else
{
graph->points_to[n] = BITMAP_ALLOC (&predbitmap_obstack);
bitmap_ior (graph->points_to[n],
graph->points_to[first_pred],
graph->points_to[w]);
}
}
else
bitmap_ior_into (graph->points_to[n], graph->points_to[w]);
}
}
/* Indirect nodes get fresh variables and a new pointer equiv class. */
if (!bitmap_bit_p (graph->direct_nodes, n))
{
if (!graph->points_to[n])
{
graph->points_to[n] = BITMAP_ALLOC (&predbitmap_obstack);
if (first_pred != -1U)
bitmap_copy (graph->points_to[n], graph->points_to[first_pred]);
}
bitmap_set_bit (graph->points_to[n], FIRST_REF_NODE + n);
graph->pointer_label[n] = pointer_equiv_class++;
equiv_class_label_t ecl;
ecl = equiv_class_lookup_or_add (pointer_equiv_class_table,
graph->points_to[n]);
ecl->equivalence_class = graph->pointer_label[n];
return;
}
/* If there was only a single non-empty predecessor the pointer equiv
class is the same. */
if (!graph->points_to[n])
{
if (first_pred != -1U)
{
graph->pointer_label[n] = graph->pointer_label[first_pred];
graph->points_to[n] = graph->points_to[first_pred];
}
return;
}
if (!bitmap_empty_p (graph->points_to[n]))
{
equiv_class_label_t ecl;
ecl = equiv_class_lookup_or_add (pointer_equiv_class_table,
graph->points_to[n]);
if (ecl->equivalence_class == 0)
ecl->equivalence_class = pointer_equiv_class++;
else
{
BITMAP_FREE (graph->points_to[n]);
graph->points_to[n] = ecl->labels;
}
graph->pointer_label[n] = ecl->equivalence_class;
}
}
/* Print the pred graph in dot format. */
static void
dump_pred_graph (class scc_info *si, FILE *file)
{
unsigned int i;
/* Only print the graph if it has already been initialized: */
if (!graph)
return;
/* Prints the header of the dot file: */
fprintf (file, "strict digraph {\n");
fprintf (file, " node [\n shape = box\n ]\n");
fprintf (file, " edge [\n fontsize = \"12\"\n ]\n");
fprintf (file, "\n // List of nodes and complex constraints in "
"the constraint graph:\n");
/* The next lines print the nodes in the graph together with the
complex constraints attached to them. */
for (i = 1; i < graph->size; i++)
{
if (i == FIRST_REF_NODE)
continue;
if (si->node_mapping[i] != i)
continue;
if (i < FIRST_REF_NODE)
fprintf (file, "\"%s\"", get_varinfo (i)->name);
else
fprintf (file, "\"*%s\"", get_varinfo (i - FIRST_REF_NODE)->name);
if (graph->points_to[i]
&& !bitmap_empty_p (graph->points_to[i]))
{
if (i < FIRST_REF_NODE)
fprintf (file, "[label=\"%s = {", get_varinfo (i)->name);
else
fprintf (file, "[label=\"*%s = {",
get_varinfo (i - FIRST_REF_NODE)->name);
unsigned j;
bitmap_iterator bi;
EXECUTE_IF_SET_IN_BITMAP (graph->points_to[i], 0, j, bi)
fprintf (file, " %d", j);
fprintf (file, " }\"]");
}
fprintf (file, ";\n");
}
/* Go over the edges. */
fprintf (file, "\n // Edges in the constraint graph:\n");
for (i = 1; i < graph->size; i++)
{
unsigned j;
bitmap_iterator bi;
if (si->node_mapping[i] != i)
continue;
EXECUTE_IF_IN_NONNULL_BITMAP (graph->preds[i], 0, j, bi)
{
unsigned from = si->node_mapping[j];
if (from < FIRST_REF_NODE)
fprintf (file, "\"%s\"", get_varinfo (from)->name);
else
fprintf (file, "\"*%s\"", get_varinfo (from - FIRST_REF_NODE)->name);
fprintf (file, " -> ");
if (i < FIRST_REF_NODE)
fprintf (file, "\"%s\"", get_varinfo (i)->name);
else
fprintf (file, "\"*%s\"", get_varinfo (i - FIRST_REF_NODE)->name);
fprintf (file, ";\n");
}
}
/* Prints the tail of the dot file. */
fprintf (file, "}\n");
}
/* Perform offline variable substitution, discovering equivalence
classes, and eliminating non-pointer variables. */
static class scc_info *
perform_var_substitution (constraint_graph_t graph)
{
unsigned int i;
unsigned int size = graph->size;
scc_info *si = new scc_info (size);
bitmap_obstack_initialize (&iteration_obstack);
gcc_obstack_init (&equiv_class_obstack);
pointer_equiv_class_table = new hash_table<equiv_class_hasher> (511);
location_equiv_class_table
= new hash_table<equiv_class_hasher> (511);
pointer_equiv_class = 1;
location_equiv_class = 1;
/* Condense the nodes, which means to find SCC's, count incoming
predecessors, and unite nodes in SCC's. */
for (i = 1; i < FIRST_REF_NODE; i++)
if (!bitmap_bit_p (si->visited, si->node_mapping[i]))
condense_visit (graph, si, si->node_mapping[i]);
if (dump_file && (dump_flags & TDF_GRAPH))
{
fprintf (dump_file, "\n\n// The constraint graph before var-substitution "
"in dot format:\n");
dump_pred_graph (si, dump_file);
fprintf (dump_file, "\n\n");
}
bitmap_clear (si->visited);
/* Actually the label the nodes for pointer equivalences */
for (i = 1; i < FIRST_REF_NODE; i++)
if (!bitmap_bit_p (si->visited, si->node_mapping[i]))
label_visit (graph, si, si->node_mapping[i]);
/* Calculate location equivalence labels. */
for (i = 1; i < FIRST_REF_NODE; i++)
{
bitmap pointed_by;
bitmap_iterator bi;
unsigned int j;
if (!graph->pointed_by[i])
continue;
pointed_by = BITMAP_ALLOC (&iteration_obstack);
/* Translate the pointed-by mapping for pointer equivalence
labels. */
EXECUTE_IF_SET_IN_BITMAP (graph->pointed_by[i], 0, j, bi)
{
bitmap_set_bit (pointed_by,
graph->pointer_label[si->node_mapping[j]]);
}
/* The original pointed_by is now dead. */
BITMAP_FREE (graph->pointed_by[i]);
/* Look up the location equivalence label if one exists, or make
one otherwise. */
equiv_class_label_t ecl;
ecl = equiv_class_lookup_or_add (location_equiv_class_table, pointed_by);
if (ecl->equivalence_class == 0)
ecl->equivalence_class = location_equiv_class++;
else
{
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "Found location equivalence for node %s\n",
get_varinfo (i)->name);
BITMAP_FREE (pointed_by);
}
graph->loc_label[i] = ecl->equivalence_class;
}
if (dump_file && (dump_flags & TDF_DETAILS))
for (i = 1; i < FIRST_REF_NODE; i++)
{
unsigned j = si->node_mapping[i];
if (j != i)
{
fprintf (dump_file, "%s node id %d ",
bitmap_bit_p (graph->direct_nodes, i)
? "Direct" : "Indirect", i);
if (i < FIRST_REF_NODE)
fprintf (dump_file, "\"%s\"", get_varinfo (i)->name);
else
fprintf (dump_file, "\"*%s\"",
get_varinfo (i - FIRST_REF_NODE)->name);
fprintf (dump_file, " mapped to SCC leader node id %d ", j);
if (j < FIRST_REF_NODE)
fprintf (dump_file, "\"%s\"\n", get_varinfo (j)->name);
else
fprintf (dump_file, "\"*%s\"\n",
get_varinfo (j - FIRST_REF_NODE)->name);
}
else
{
fprintf (dump_file,
"Equivalence classes for %s node id %d ",
bitmap_bit_p (graph->direct_nodes, i)
? "direct" : "indirect", i);
if (i < FIRST_REF_NODE)
fprintf (dump_file, "\"%s\"", get_varinfo (i)->name);
else
fprintf (dump_file, "\"*%s\"",
get_varinfo (i - FIRST_REF_NODE)->name);
fprintf (dump_file,
": pointer %d, location %d\n",
graph->pointer_label[i], graph->loc_label[i]);
}
}
/* Quickly eliminate our non-pointer variables. */
for (i = 1; i < FIRST_REF_NODE; i++)
{
unsigned int node = si->node_mapping[i];
if (graph->pointer_label[node] == 0)
{
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file,
"%s is a non-pointer variable, eliminating edges.\n",
get_varinfo (node)->name);
stats.nonpointer_vars++;
clear_edges_for_node (graph, node);
}
}
return si;
}
/* Free information that was only necessary for variable
substitution. */
static void
free_var_substitution_info (class scc_info *si)
{
delete si;
free (graph->pointer_label);
free (graph->loc_label);
free (graph->pointed_by);
free (graph->points_to);
free (graph->eq_rep);
sbitmap_free (graph->direct_nodes);
delete pointer_equiv_class_table;
pointer_equiv_class_table = NULL;
delete location_equiv_class_table;
location_equiv_class_table = NULL;
obstack_free (&equiv_class_obstack, NULL);
bitmap_obstack_release (&iteration_obstack);
}
/* Return an existing node that is equivalent to NODE, which has
equivalence class LABEL, if one exists. Return NODE otherwise. */
static unsigned int
find_equivalent_node (constraint_graph_t graph,
unsigned int node, unsigned int label)
{
/* If the address version of this variable is unused, we can
substitute it for anything else with the same label.
Otherwise, we know the pointers are equivalent, but not the
locations, and we can unite them later. */
if (!bitmap_bit_p (graph->address_taken, node))
{
gcc_checking_assert (label < graph->size);
if (graph->eq_rep[label] != -1)
{
/* Unify the two variables since we know they are equivalent. */
if (unite (graph->eq_rep[label], node))
unify_nodes (graph, graph->eq_rep[label], node, false);
return graph->eq_rep[label];
}
else
{
graph->eq_rep[label] = node;
graph->pe_rep[label] = node;
}
}
else
{
gcc_checking_assert (label < graph->size);
graph->pe[node] = label;
if (graph->pe_rep[label] == -1)
graph->pe_rep[label] = node;
}
return node;
}
/* Unite pointer equivalent but not location equivalent nodes in
GRAPH. This may only be performed once variable substitution is
finished. */
static void
unite_pointer_equivalences (constraint_graph_t graph)
{
unsigned int i;
/* Go through the pointer equivalences and unite them to their
representative, if they aren't already. */
for (i = 1; i < FIRST_REF_NODE; i++)
{
unsigned int label = graph->pe[i];
if (label)
{
int label_rep = graph->pe_rep[label];
if (label_rep == -1)
continue;
label_rep = find (label_rep);
if (label_rep >= 0 && unite (label_rep, find (i)))
unify_nodes (graph, label_rep, i, false);
}
}
}
/* Move complex constraints to the GRAPH nodes they belong to. */
static void
move_complex_constraints (constraint_graph_t graph)
{
int i;
constraint_t c;
FOR_EACH_VEC_ELT (constraints, i, c)
{
if (c)
{
struct constraint_expr lhs = c->lhs;
struct constraint_expr rhs = c->rhs;
if (lhs.type == DEREF)
{
insert_into_complex (graph, lhs.var, c);
}
else if (rhs.type == DEREF)
{
if (!(get_varinfo (lhs.var)->is_special_var))
insert_into_complex (graph, rhs.var, c);
}
else if (rhs.type != ADDRESSOF && lhs.var > anything_id
&& (lhs.offset != 0 || rhs.offset != 0))
{
insert_into_complex (graph, rhs.var, c);
}
}
}
}
/* Optimize and rewrite complex constraints while performing
collapsing of equivalent nodes. SI is the SCC_INFO that is the
result of perform_variable_substitution. */
static void
rewrite_constraints (constraint_graph_t graph,
class scc_info *si)
{
int i;
constraint_t c;
if (flag_checking)
{
for (unsigned int j = 0; j < graph->size; j++)
gcc_assert (find (j) == j);
}
FOR_EACH_VEC_ELT (constraints, i, c)
{
struct constraint_expr lhs = c->lhs;
struct constraint_expr rhs = c->rhs;
unsigned int lhsvar = find (lhs.var);
unsigned int rhsvar = find (rhs.var);
unsigned int lhsnode, rhsnode;
unsigned int lhslabel, rhslabel;
lhsnode = si->node_mapping[lhsvar];
rhsnode = si->node_mapping[rhsvar];
lhslabel = graph->pointer_label[lhsnode];
rhslabel = graph->pointer_label[rhsnode];
/* See if it is really a non-pointer variable, and if so, ignore
the constraint. */
if (lhslabel == 0)
{
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, "%s is a non-pointer variable, "
"ignoring constraint:",
get_varinfo (lhs.var)->name);
dump_constraint (dump_file, c);
fprintf (dump_file, "\n");
}
constraints[i] = NULL;
continue;
}
if (rhslabel == 0)
{
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, "%s is a non-pointer variable, "
"ignoring constraint:",
get_varinfo (rhs.var)->name);
dump_constraint (dump_file, c);
fprintf (dump_file, "\n");
}
constraints[i] = NULL;
continue;
}
lhsvar = find_equivalent_node (graph, lhsvar, lhslabel);
rhsvar = find_equivalent_node (graph, rhsvar, rhslabel);
c->lhs.var = lhsvar;
c->rhs.var = rhsvar;
}
}
/* Eliminate indirect cycles involving NODE. Return true if NODE was
part of an SCC, false otherwise. */
static bool
eliminate_indirect_cycles (unsigned int node)
{
if (graph->indirect_cycles[node] != -1
&& !bitmap_empty_p (get_varinfo (node)->solution))
{
unsigned int i;
auto_vec<unsigned> queue;
int queuepos;
unsigned int to = find (graph->indirect_cycles[node]);
bitmap_iterator bi;
/* We can't touch the solution set and call unify_nodes
at the same time, because unify_nodes is going to do
bitmap unions into it. */
EXECUTE_IF_SET_IN_BITMAP (get_varinfo (node)->solution, 0, i, bi)
{
if (find (i) == i && i != to)
{
if (unite (to, i))
queue.safe_push (i);
}
}
for (queuepos = 0;
queue.iterate (queuepos, &i);
queuepos++)
{
unify_nodes (graph, to, i, true);
}
return true;
}
return false;
}
/* Solve the constraint graph GRAPH using our worklist solver.
This is based on the PW* family of solvers from the "Efficient Field
Sensitive Pointer Analysis for C" paper.
It works by iterating over all the graph nodes, processing the complex
constraints and propagating the copy constraints, until everything stops
changed. This corresponds to steps 6-8 in the solving list given above. */
static void
solve_graph (constraint_graph_t graph)
{
unsigned int size = graph->size;
unsigned int i;
bitmap pts;
changed = BITMAP_ALLOC (NULL);
/* Mark all initial non-collapsed nodes as changed. */
for (i = 1; i < size; i++)
{
varinfo_t ivi = get_varinfo (i);
if (find (i) == i && !bitmap_empty_p (ivi->solution)
&& ((graph->succs[i] && !bitmap_empty_p (graph->succs[i]))
|| graph->complex[i].length () > 0))
bitmap_set_bit (changed, i);
}
/* Allocate a bitmap to be used to store the changed bits. */
pts = BITMAP_ALLOC (&pta_obstack);
while (!bitmap_empty_p (changed))
{
unsigned int i;
struct topo_info *ti = init_topo_info ();
stats.iterations++;
bitmap_obstack_initialize (&iteration_obstack);
compute_topo_order (graph, ti);
while (ti->topo_order.length () != 0)
{
i = ti->topo_order.pop ();
/* If this variable is not a representative, skip it. */
if (find (i) != i)
continue;
/* In certain indirect cycle cases, we may merge this
variable to another. */
if (eliminate_indirect_cycles (i) && find (i) != i)
continue;
/* If the node has changed, we need to process the
complex constraints and outgoing edges again. */
if (bitmap_clear_bit (changed, i))
{
unsigned int j;
constraint_t c;
bitmap solution;
vec<constraint_t> complex = graph->complex[i];
varinfo_t vi = get_varinfo (i);
bool solution_empty;
/* Compute the changed set of solution bits. If anything
is in the solution just propagate that. */
if (bitmap_bit_p (vi->solution, anything_id))
{
/* If anything is also in the old solution there is
nothing to do.
??? But we shouldn't ended up with "changed" set ... */
if (vi->oldsolution
&& bitmap_bit_p (vi->oldsolution, anything_id))
continue;
bitmap_copy (pts, get_varinfo (find (anything_id))->solution);
}
else if (vi->oldsolution)
bitmap_and_compl (pts, vi->solution, vi->oldsolution);
else
bitmap_copy (pts, vi->solution);
if (bitmap_empty_p (pts))
continue;
if (vi->oldsolution)
bitmap_ior_into (vi->oldsolution, pts);
else
{
vi->oldsolution = BITMAP_ALLOC (&oldpta_obstack);
bitmap_copy (vi->oldsolution, pts);
}
solution = vi->solution;
solution_empty = bitmap_empty_p (solution);
/* Process the complex constraints */
bitmap expanded_pts = NULL;
FOR_EACH_VEC_ELT (complex, j, c)
{
/* XXX: This is going to unsort the constraints in
some cases, which will occasionally add duplicate
constraints during unification. This does not
affect correctness. */
c->lhs.var = find (c->lhs.var);
c->rhs.var = find (c->rhs.var);
/* The only complex constraint that can change our
solution to non-empty, given an empty solution,
is a constraint where the lhs side is receiving
some set from elsewhere. */
if (!solution_empty || c->lhs.type != DEREF)
do_complex_constraint (graph, c, pts, &expanded_pts);
}
BITMAP_FREE (expanded_pts);
solution_empty = bitmap_empty_p (solution);
if (!solution_empty)
{
bitmap_iterator bi;
unsigned eff_escaped_id = find (escaped_id);
/* Propagate solution to all successors. */
unsigned to_remove = ~0U;
EXECUTE_IF_IN_NONNULL_BITMAP (graph->succs[i],
0, j, bi)
{
if (to_remove != ~0U)
{
bitmap_clear_bit (graph->succs[i], to_remove);
to_remove = ~0U;
}
unsigned int to = find (j);
if (to != j)
{
/* Update the succ graph, avoiding duplicate
work. */
to_remove = j;
if (! bitmap_set_bit (graph->succs[i], to))
continue;
/* We eventually end up processing 'to' twice
as it is undefined whether bitmap iteration
iterates over bits set during iteration.
Play safe instead of doing tricks. */
}
/* Don't try to propagate to ourselves. */
if (to == i)
continue;
bitmap tmp = get_varinfo (to)->solution;
bool flag = false;
/* If we propagate from ESCAPED use ESCAPED as
placeholder. */
if (i == eff_escaped_id)
flag = bitmap_set_bit (tmp, escaped_id);
else
flag = bitmap_ior_into (tmp, pts);
if (flag)
bitmap_set_bit (changed, to);
}
if (to_remove != ~0U)
bitmap_clear_bit (graph->succs[i], to_remove);
}
}
}
free_topo_info (ti);
bitmap_obstack_release (&iteration_obstack);
}
BITMAP_FREE (pts);
BITMAP_FREE (changed);
bitmap_obstack_release (&oldpta_obstack);
}
/* Map from trees to variable infos. */
static hash_map<tree, varinfo_t> *vi_for_tree;
/* Insert ID as the variable id for tree T in the vi_for_tree map. */
static void
insert_vi_for_tree (tree t, varinfo_t vi)
{
gcc_assert (vi);
gcc_assert (!vi_for_tree->put (t, vi));
}
/* Find the variable info for tree T in VI_FOR_TREE. If T does not
exist in the map, return NULL, otherwise, return the varinfo we found. */
static varinfo_t
lookup_vi_for_tree (tree t)
{
varinfo_t *slot = vi_for_tree->get (t);
if (slot == NULL)
return NULL;
return *slot;
}
/* Return a printable name for DECL */
static const char *
alias_get_name (tree decl)
{
const char *res = "NULL";
if (dump_file)
{
char *temp = NULL;
if (TREE_CODE (decl) == SSA_NAME)
{
res = get_name (decl);
temp = xasprintf ("%s_%u", res ? res : "", SSA_NAME_VERSION (decl));
}
else if (HAS_DECL_ASSEMBLER_NAME_P (decl)
&& DECL_ASSEMBLER_NAME_SET_P (decl))
res = IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME_RAW (decl));
else if (DECL_P (decl))
{
res = get_name (decl);
if (!res)
temp = xasprintf ("D.%u", DECL_UID (decl));
}
if (temp)
{
res = ggc_strdup (temp);
free (temp);
}
}
return res;
}
/* Find the variable id for tree T in the map.
If T doesn't exist in the map, create an entry for it and return it. */
static varinfo_t
get_vi_for_tree (tree t)
{
varinfo_t *slot = vi_for_tree->get (t);
if (slot == NULL)
{
unsigned int id = create_variable_info_for (t, alias_get_name (t), false);
return get_varinfo (id);
}
return *slot;
}
/* Get a scalar constraint expression for a new temporary variable. */
static struct constraint_expr
new_scalar_tmp_constraint_exp (const char *name, bool add_id)
{
struct constraint_expr tmp;
varinfo_t vi;
vi = new_var_info (NULL_TREE, name, add_id);
vi->offset = 0;
vi->size = -1;
vi->fullsize = -1;
vi->is_full_var = 1;
vi->is_reg_var = 1;
tmp.var = vi->id;
tmp.type = SCALAR;
tmp.offset = 0;
return tmp;
}
/* Get a constraint expression vector from an SSA_VAR_P node.
If address_p is true, the result will be taken its address of. */
static void
get_constraint_for_ssa_var (tree t, vec<ce_s> *results, bool address_p)
{
struct constraint_expr cexpr;
varinfo_t vi;
/* We allow FUNCTION_DECLs here even though it doesn't make much sense. */
gcc_assert (TREE_CODE (t) == SSA_NAME || DECL_P (t));
if (TREE_CODE (t) == SSA_NAME
&& SSA_NAME_IS_DEFAULT_DEF (t))
{
/* For parameters, get at the points-to set for the actual parm
decl. */
if (TREE_CODE (SSA_NAME_VAR (t)) == PARM_DECL
|| TREE_CODE (SSA_NAME_VAR (t)) == RESULT_DECL)
{
get_constraint_for_ssa_var (SSA_NAME_VAR (t), results, address_p);
return;
}
/* For undefined SSA names return nothing. */
else if (!ssa_defined_default_def_p (t))
{
cexpr.var = nothing_id;
cexpr.type = SCALAR;
cexpr.offset = 0;
results->safe_push (cexpr);
return;
}
}
/* For global variables resort to the alias target. */
if (VAR_P (t) && (TREE_STATIC (t) || DECL_EXTERNAL (t)))
{
varpool_node *node = varpool_node::get (t);
if (node && node->alias && node->analyzed)
{
node = node->ultimate_alias_target ();
/* Canonicalize the PT uid of all aliases to the ultimate target.
??? Hopefully the set of aliases can't change in a way that
changes the ultimate alias target. */
gcc_assert ((! DECL_PT_UID_SET_P (node->decl)
|| DECL_PT_UID (node->decl) == DECL_UID (node->decl))
&& (! DECL_PT_UID_SET_P (t)
|| DECL_PT_UID (t) == DECL_UID (node->decl)));
DECL_PT_UID (t) = DECL_UID (node->decl);
t = node->decl;
}
/* If this is decl may bind to NULL note that. */
if (address_p
&& (! node || ! node->nonzero_address ()))
{
cexpr.var = nothing_id;
cexpr.type = SCALAR;
cexpr.offset = 0;
results->safe_push (cexpr);
}
}
vi = get_vi_for_tree (t);
cexpr.var = vi->id;
cexpr.type = SCALAR;
cexpr.offset = 0;
/* If we are not taking the address of the constraint expr, add all
sub-fiels of the variable as well. */
if (!address_p
&& !vi->is_full_var)
{
for (; vi; vi = vi_next (vi))
{
cexpr.var = vi->id;
results->safe_push (cexpr);
}
return;
}
results->safe_push (cexpr);
}
/* Process constraint T, performing various simplifications and then
adding it to our list of overall constraints. */
static void
process_constraint (constraint_t t)
{
struct constraint_expr rhs = t->rhs;
struct constraint_expr lhs = t->lhs;
gcc_assert (rhs.var < varmap.length ());
gcc_assert (lhs.var < varmap.length ());
/* If we didn't get any useful constraint from the lhs we get
&ANYTHING as fallback from get_constraint_for. Deal with
it here by turning it into *ANYTHING. */
if (lhs.type == ADDRESSOF
&& lhs.var == anything_id)
lhs.type = DEREF;
/* ADDRESSOF on the lhs is invalid. */
gcc_assert (lhs.type != ADDRESSOF);
/* We shouldn't add constraints from things that cannot have pointers.
It's not completely trivial to avoid in the callers, so do it here. */
if (rhs.type != ADDRESSOF
&& !get_varinfo (rhs.var)->may_have_pointers)
return;
/* Likewise adding to the solution of a non-pointer var isn't useful. */
if (!get_varinfo (lhs.var)->may_have_pointers)
return;
/* This can happen in our IR with things like n->a = *p */
if (rhs.type == DEREF && lhs.type == DEREF && rhs.var != anything_id)
{
/* Split into tmp = *rhs, *lhs = tmp */
struct constraint_expr tmplhs;
tmplhs = new_scalar_tmp_constraint_exp ("doubledereftmp", true);
process_constraint (new_constraint (tmplhs, rhs));
process_constraint (new_constraint (lhs, tmplhs));
}
else if ((rhs.type != SCALAR || rhs.offset != 0) && lhs.type == DEREF)
{
/* Split into tmp = &rhs, *lhs = tmp */
struct constraint_expr tmplhs;
tmplhs = new_scalar_tmp_constraint_exp ("derefaddrtmp", true);
process_constraint (new_constraint (tmplhs, rhs));
process_constraint (new_constraint (lhs, tmplhs));
}
else
{
gcc_assert (rhs.type != ADDRESSOF || rhs.offset == 0);
if (rhs.type == ADDRESSOF)
get_varinfo (get_varinfo (rhs.var)->head)->address_taken = true;
constraints.safe_push (t);
}
}
/* Return the position, in bits, of FIELD_DECL from the beginning of its
structure. */
static HOST_WIDE_INT
bitpos_of_field (const tree fdecl)
{
if (!tree_fits_shwi_p (DECL_FIELD_OFFSET (fdecl))
|| !tree_fits_shwi_p (DECL_FIELD_BIT_OFFSET (fdecl)))
return -1;
return (tree_to_shwi (DECL_FIELD_OFFSET (fdecl)) * BITS_PER_UNIT
+ tree_to_shwi (DECL_FIELD_BIT_OFFSET (fdecl)));
}
/* Get constraint expressions for offsetting PTR by OFFSET. Stores the
resulting constraint expressions in *RESULTS. */
static void
get_constraint_for_ptr_offset (tree ptr, tree offset,
vec<ce_s> *results)
{
struct constraint_expr c;
unsigned int j, n;
HOST_WIDE_INT rhsoffset;
/* If we do not do field-sensitive PTA adding offsets to pointers
does not change the points-to solution. */
if (!use_field_sensitive)
{
get_constraint_for_rhs (ptr, results);
return;
}
/* If the offset is not a non-negative integer constant that fits
in a HOST_WIDE_INT, we have to fall back to a conservative
solution which includes all sub-fields of all pointed-to
variables of ptr. */
if (offset == NULL_TREE
|| TREE_CODE (offset) != INTEGER_CST)
rhsoffset = UNKNOWN_OFFSET;
else
{
/* Sign-extend the offset. */
offset_int soffset = offset_int::from (wi::to_wide (offset), SIGNED);
if (!wi::fits_shwi_p (soffset))
rhsoffset = UNKNOWN_OFFSET;
else
{
/* Make sure the bit-offset also fits. */
HOST_WIDE_INT rhsunitoffset = soffset.to_shwi ();
rhsoffset = rhsunitoffset * (unsigned HOST_WIDE_INT) BITS_PER_UNIT;
if (rhsunitoffset != rhsoffset / BITS_PER_UNIT)
rhsoffset = UNKNOWN_OFFSET;
}
}
get_constraint_for_rhs (ptr, results);
if (rhsoffset == 0)
return;
/* As we are eventually appending to the solution do not use
vec::iterate here. */
n = results->length ();
for (j = 0; j < n; j++)
{
varinfo_t curr;
c = (*results)[j];
curr = get_varinfo (c.var);
if (c.type == ADDRESSOF
/* If this varinfo represents a full variable just use it. */
&& curr->is_full_var)
;
else if (<