/* Loop invariant motion.
- Copyright (C) 2003, 2004, 2005 Free Software Foundation, Inc.
-
+ Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2010
+ Free Software Foundation, Inc.
+
This file is part of GCC.
-
+
GCC is free software; you can redistribute it and/or modify it
under the terms of the GNU General Public License as published by the
-Free Software Foundation; either version 2, or (at your option) any
+Free Software Foundation; either version 3, 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 COPYING. If not, write to the Free
-Software Foundation, 59 Temple Place - Suite 330, Boston, MA
-02111-1307, USA. */
+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 "tm.h"
#include "tree.h"
-#include "rtl.h"
#include "tm_p.h"
-#include "hard-reg-set.h"
#include "basic-block.h"
#include "output.h"
#include "diagnostic.h"
+#include "gimple-pretty-print.h"
+#include "tree-pretty-print.h"
#include "tree-flow.h"
#include "tree-dump.h"
#include "timevar.h"
#include "params.h"
#include "tree-pass.h"
#include "flags.h"
+#include "hashtab.h"
+#include "tree-affine.h"
+#include "pointer-set.h"
+#include "tree-ssa-propagate.h"
/* TODO: Support for predicated code motion. I.e.
struct depend
{
- tree stmt;
+ gimple stmt;
struct depend *next;
};
the statement is executed if the loop
is entered. */
- bool sm_done; /* True iff the store motion for a memory
- reference in the statement has already
- been executed. */
-
unsigned cost; /* Cost of the computation performed by the
statement. */
MAX_LOOP loop. */
};
-#define LIM_DATA(STMT) (TREE_CODE (STMT) == PHI_NODE \
- ? NULL \
- : (struct lim_aux_data *) (stmt_ann (STMT)->common.aux))
+/* Maps statements to their lim_aux_data. */
+
+static struct pointer_map_t *lim_aux_data_map;
-/* Description of a memory reference for store motion. */
+/* Description of a memory reference location. */
-struct mem_ref
+typedef struct mem_ref_loc
{
tree *ref; /* The reference itself. */
- tree stmt; /* The statement in that it occurs. */
- struct mem_ref *next; /* Next use in the chain. */
-};
+ gimple stmt; /* The statement in that it occurs. */
+} *mem_ref_loc_p;
+
+DEF_VEC_P(mem_ref_loc_p);
+DEF_VEC_ALLOC_P(mem_ref_loc_p, heap);
+
+/* The list of memory reference locations in a loop. */
+
+typedef struct mem_ref_locs
+{
+ VEC (mem_ref_loc_p, heap) *locs;
+} *mem_ref_locs_p;
+
+DEF_VEC_P(mem_ref_locs_p);
+DEF_VEC_ALLOC_P(mem_ref_locs_p, heap);
+
+/* Description of a memory reference. */
+
+typedef struct mem_ref
+{
+ tree mem; /* The memory itself. */
+ unsigned id; /* ID assigned to the memory reference
+ (its index in memory_accesses.refs_list) */
+ hashval_t hash; /* Its hash value. */
+ bitmap stored; /* The set of loops in that this memory location
+ is stored to. */
+ VEC (mem_ref_locs_p, heap) *accesses_in_loop;
+ /* The locations of the accesses. Vector
+ indexed by the loop number. */
+ bitmap vops; /* Vops corresponding to this memory
+ location. */
+
+ /* The following sets are computed on demand. We keep both set and
+ its complement, so that we know whether the information was
+ already computed or not. */
+ bitmap indep_loop; /* The set of loops in that the memory
+ reference is independent, meaning:
+ If it is stored in the loop, this store
+ is independent on all other loads and
+ stores.
+ If it is only loaded, then it is independent
+ on all stores in the loop. */
+ bitmap dep_loop; /* The complement of INDEP_LOOP. */
+
+ bitmap indep_ref; /* The set of memory references on that
+ this reference is independent. */
+ bitmap dep_ref; /* The complement of DEP_REF. */
+} *mem_ref_p;
+
+DEF_VEC_P(mem_ref_p);
+DEF_VEC_ALLOC_P(mem_ref_p, heap);
+
+DEF_VEC_P(bitmap);
+DEF_VEC_ALLOC_P(bitmap, heap);
+
+DEF_VEC_P(htab_t);
+DEF_VEC_ALLOC_P(htab_t, heap);
+
+/* Description of memory accesses in loops. */
+
+static struct
+{
+ /* The hash table of memory references accessed in loops. */
+ htab_t refs;
+
+ /* The list of memory references. */
+ VEC (mem_ref_p, heap) *refs_list;
+
+ /* The set of memory references accessed in each loop. */
+ VEC (bitmap, heap) *refs_in_loop;
+
+ /* The set of memory references accessed in each loop, including
+ subloops. */
+ VEC (bitmap, heap) *all_refs_in_loop;
+
+ /* The set of virtual operands clobbered in a given loop. */
+ VEC (bitmap, heap) *clobbered_vops;
+
+ /* Map from the pair (loop, virtual operand) to the set of refs that
+ touch the virtual operand in the loop. */
+ VEC (htab_t, heap) *vop_ref_map;
+
+ /* Cache for expanding memory addresses. */
+ struct pointer_map_t *ttae_cache;
+} memory_accesses;
+
+static bool ref_indep_loop_p (struct loop *, mem_ref_p);
/* Minimum cost of an expensive expression. */
#define LIM_EXPENSIVE ((unsigned) PARAM_VALUE (PARAM_LIM_EXPENSIVE))
block will be executed. */
#define ALWAYS_EXECUTED_IN(BB) ((struct loop *) (BB)->aux)
-static unsigned max_stmt_uid; /* Maximal uid of a statement. Uids to phi
- nodes are assigned using the versions of
- ssa names they define. */
+static struct lim_aux_data *
+init_lim_data (gimple stmt)
+{
+ void **p = pointer_map_insert (lim_aux_data_map, stmt);
+
+ *p = XCNEW (struct lim_aux_data);
+ return (struct lim_aux_data *) *p;
+}
+
+static struct lim_aux_data *
+get_lim_data (gimple stmt)
+{
+ void **p = pointer_map_contains (lim_aux_data_map, stmt);
+ if (!p)
+ return NULL;
-/* Returns uid of statement STMT. */
+ return (struct lim_aux_data *) *p;
+}
-static unsigned
-get_stmt_uid (tree stmt)
+/* Releases the memory occupied by DATA. */
+
+static void
+free_lim_aux_data (struct lim_aux_data *data)
+{
+ struct depend *dep, *next;
+
+ for (dep = data->depends; dep; dep = next)
+ {
+ next = dep->next;
+ free (dep);
+ }
+ free (data);
+}
+
+static void
+clear_lim_data (gimple stmt)
{
- if (TREE_CODE (stmt) == PHI_NODE)
- return SSA_NAME_VERSION (PHI_RESULT (stmt)) + max_stmt_uid;
+ void **p = pointer_map_contains (lim_aux_data_map, stmt);
+ if (!p)
+ return;
- return stmt_ann (stmt)->uid;
+ free_lim_aux_data ((struct lim_aux_data *) *p);
+ *p = NULL;
}
/* Calls CBCK for each index in memory reference ADDR_P. There are two
kinds situations handled; in each of these cases, the memory reference
and DATA are passed to the callback:
-
+
Access to an array: ARRAY_{RANGE_}REF (base, index). In this case we also
pass the pointer to the index to the callback.
Pointer dereference: INDIRECT_REF (addr). In this case we also pass the
pointer to addr to the callback.
-
+
If the callback returns false, the whole search stops and false is returned.
Otherwise the function returns true after traversing through the whole
reference *ADDR_P. */
case BIT_FIELD_REF:
case VIEW_CONVERT_EXPR:
- case ARRAY_RANGE_REF:
case REALPART_EXPR:
case IMAGPART_EXPR:
nxt = &TREE_OPERAND (*addr_p, 0);
break;
case ARRAY_REF:
+ case ARRAY_RANGE_REF:
nxt = &TREE_OPERAND (*addr_p, 0);
if (!cbck (*addr_p, &TREE_OPERAND (*addr_p, 1), data))
return false;
case PARM_DECL:
case STRING_CST:
case RESULT_DECL:
+ case VECTOR_CST:
+ case COMPLEX_CST:
+ case INTEGER_CST:
+ case REAL_CST:
+ case FIXED_CST:
+ case CONSTRUCTOR:
+ return true;
+
+ case ADDR_EXPR:
+ gcc_assert (is_gimple_min_invariant (*addr_p));
+ return true;
+
+ case TARGET_MEM_REF:
+ idx = &TMR_BASE (*addr_p);
+ if (*idx
+ && !cbck (*addr_p, idx, data))
+ return false;
+ idx = &TMR_INDEX (*addr_p);
+ if (*idx
+ && !cbck (*addr_p, idx, data))
+ return false;
return true;
default:
Otherwise return MOVE_IMPOSSIBLE. */
enum move_pos
-movement_possibility (tree stmt)
+movement_possibility (gimple stmt)
{
- tree lhs, rhs;
+ tree lhs;
+ enum move_pos ret = MOVE_POSSIBLE;
if (flag_unswitch_loops
- && TREE_CODE (stmt) == COND_EXPR)
+ && gimple_code (stmt) == GIMPLE_COND)
{
/* If we perform unswitching, force the operands of the invariant
condition to be moved out of the loop. */
- get_stmt_operands (stmt);
-
return MOVE_POSSIBLE;
}
- if (TREE_CODE (stmt) != MODIFY_EXPR)
- return MOVE_IMPOSSIBLE;
-
- if (stmt_ends_bb_p (stmt))
- return MOVE_IMPOSSIBLE;
-
- get_stmt_operands (stmt);
+ if (gimple_code (stmt) == GIMPLE_PHI
+ && gimple_phi_num_args (stmt) <= 2
+ && is_gimple_reg (gimple_phi_result (stmt))
+ && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (gimple_phi_result (stmt)))
+ return MOVE_POSSIBLE;
- if (stmt_ann (stmt)->has_volatile_ops)
+ if (gimple_get_lhs (stmt) == NULL_TREE)
return MOVE_IMPOSSIBLE;
- lhs = TREE_OPERAND (stmt, 0);
- if (TREE_CODE (lhs) == SSA_NAME
- && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs))
+ if (gimple_vdef (stmt))
return MOVE_IMPOSSIBLE;
- rhs = TREE_OPERAND (stmt, 1);
-
- if (TREE_SIDE_EFFECTS (rhs))
+ if (stmt_ends_bb_p (stmt)
+ || gimple_has_volatile_ops (stmt)
+ || gimple_has_side_effects (stmt)
+ || stmt_could_throw_p (stmt))
return MOVE_IMPOSSIBLE;
- if (TREE_CODE (lhs) != SSA_NAME
- || tree_could_trap_p (rhs))
- return MOVE_PRESERVE_EXECUTION;
-
- if (get_call_expr_in (stmt))
+ if (is_gimple_call (stmt))
{
/* While pure or const call is guaranteed to have no side effects, we
cannot move it arbitrarily. Consider code like
invalid arguments, moving out a function call that is not executed
may cause performance regressions in case the call is costly and
not executed at all. */
- return MOVE_PRESERVE_EXECUTION;
+ ret = MOVE_PRESERVE_EXECUTION;
+ lhs = gimple_call_lhs (stmt);
}
- return MOVE_POSSIBLE;
+ else if (is_gimple_assign (stmt))
+ lhs = gimple_assign_lhs (stmt);
+ else
+ return MOVE_IMPOSSIBLE;
+
+ if (TREE_CODE (lhs) == SSA_NAME
+ && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs))
+ return MOVE_IMPOSSIBLE;
+
+ if (TREE_CODE (lhs) != SSA_NAME
+ || gimple_could_trap_p (stmt))
+ return MOVE_PRESERVE_EXECUTION;
+
+ return ret;
}
/* Suppose that operand DEF is used inside the LOOP. Returns the outermost
static struct loop *
outermost_invariant_loop (tree def, struct loop *loop)
{
- tree def_stmt;
+ gimple def_stmt;
basic_block def_bb;
struct loop *max_loop;
+ struct lim_aux_data *lim_data;
- if (TREE_CODE (def) != SSA_NAME)
+ if (!def)
return superloop_at_depth (loop, 1);
+ if (TREE_CODE (def) != SSA_NAME)
+ {
+ gcc_assert (is_gimple_min_invariant (def));
+ return superloop_at_depth (loop, 1);
+ }
+
def_stmt = SSA_NAME_DEF_STMT (def);
- def_bb = bb_for_stmt (def_stmt);
+ def_bb = gimple_bb (def_stmt);
if (!def_bb)
return superloop_at_depth (loop, 1);
max_loop = find_common_loop (loop, def_bb->loop_father);
- if (LIM_DATA (def_stmt) && LIM_DATA (def_stmt)->max_loop)
+ lim_data = get_lim_data (def_stmt);
+ if (lim_data != NULL && lim_data->max_loop != NULL)
max_loop = find_common_loop (max_loop,
- LIM_DATA (def_stmt)->max_loop->outer);
+ loop_outer (lim_data->max_loop));
if (max_loop == loop)
return NULL;
- max_loop = superloop_at_depth (loop, max_loop->depth + 1);
-
- return max_loop;
-}
-
-/* Returns the outermost superloop of LOOP in that the expression EXPR is
- invariant. */
-
-static struct loop *
-outermost_invariant_loop_expr (tree expr, struct loop *loop)
-{
- enum tree_code_class class = TREE_CODE_CLASS (TREE_CODE (expr));
- unsigned i, nops;
- struct loop *max_loop = superloop_at_depth (loop, 1), *aloop;
-
- if (TREE_CODE (expr) == SSA_NAME
- || TREE_CODE (expr) == INTEGER_CST
- || is_gimple_min_invariant (expr))
- return outermost_invariant_loop (expr, loop);
-
- if (class != tcc_unary
- && class != tcc_binary
- && class != tcc_expression
- && class != tcc_comparison)
- return NULL;
-
- nops = TREE_CODE_LENGTH (TREE_CODE (expr));
- for (i = 0; i < nops; i++)
- {
- aloop = outermost_invariant_loop_expr (TREE_OPERAND (expr, i), loop);
- if (!aloop)
- return NULL;
-
- if (flow_loop_nested_p (max_loop, aloop))
- max_loop = aloop;
- }
+ max_loop = superloop_at_depth (loop, loop_depth (max_loop) + 1);
return max_loop;
}
/* DATA is a structure containing information associated with a statement
inside LOOP. DEF is one of the operands of this statement.
-
+
Find the outermost loop enclosing LOOP in that value of DEF is invariant
and record this in DATA->max_loop field. If DEF itself is defined inside
this loop as well (i.e. we need to hoist it out of the loop if we want
to hoist the statement represented by DATA), record the statement in that
DEF is defined to the DATA->depends list. Additionally if ADD_COST is true,
add the cost of the computation of DEF to the DATA->cost.
-
+
If DEF is not invariant in LOOP, return false. Otherwise return TRUE. */
static bool
add_dependency (tree def, struct lim_aux_data *data, struct loop *loop,
bool add_cost)
{
- tree def_stmt = SSA_NAME_DEF_STMT (def);
- basic_block def_bb = bb_for_stmt (def_stmt);
+ gimple def_stmt = SSA_NAME_DEF_STMT (def);
+ basic_block def_bb = gimple_bb (def_stmt);
struct loop *max_loop;
struct depend *dep;
+ struct lim_aux_data *def_data;
if (!def_bb)
return true;
if (flow_loop_nested_p (data->max_loop, max_loop))
data->max_loop = max_loop;
- if (!LIM_DATA (def_stmt))
+ def_data = get_lim_data (def_stmt);
+ if (!def_data)
return true;
if (add_cost
on it, we will be able to avoid creating a new register for
it (since it will be only used in these dependent invariants). */
&& def_bb->loop_father == loop)
- data->cost += LIM_DATA (def_stmt)->cost;
+ data->cost += def_data->cost;
- dep = xmalloc (sizeof (struct depend));
+ dep = XNEW (struct depend);
dep->stmt = def_stmt;
dep->next = data->depends;
data->depends = dep;
values. */
static unsigned
-stmt_cost (tree stmt)
+stmt_cost (gimple stmt)
{
- tree rhs;
+ tree fndecl;
unsigned cost = 1;
/* Always try to create possibilities for unswitching. */
- if (TREE_CODE (stmt) == COND_EXPR)
+ if (gimple_code (stmt) == GIMPLE_COND
+ || gimple_code (stmt) == GIMPLE_PHI)
return LIM_EXPENSIVE;
- rhs = TREE_OPERAND (stmt, 1);
-
/* Hoisting memory references out should almost surely be a win. */
- if (stmt_references_memory_p (stmt))
+ if (gimple_references_memory_p (stmt))
cost += 20;
- switch (TREE_CODE (rhs))
+ if (is_gimple_call (stmt))
{
- case CALL_EXPR:
/* We should be hoisting calls if possible. */
/* Unless the call is a builtin_constant_p; this always folds to a
constant, so moving it is useless. */
- rhs = get_callee_fndecl (rhs);
- if (DECL_BUILT_IN_CLASS (rhs) == BUILT_IN_NORMAL
- && DECL_FUNCTION_CODE (rhs) == BUILT_IN_CONSTANT_P)
+ fndecl = gimple_call_fndecl (stmt);
+ if (fndecl
+ && DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL
+ && DECL_FUNCTION_CODE (fndecl) == BUILT_IN_CONSTANT_P)
return 0;
- cost += 20;
- break;
+ return cost + 20;
+ }
+
+ if (gimple_code (stmt) != GIMPLE_ASSIGN)
+ return cost;
+ switch (gimple_assign_rhs_code (stmt))
+ {
case MULT_EXPR:
case TRUNC_DIV_EXPR:
case CEIL_DIV_EXPR:
case FLOOR_MOD_EXPR:
case ROUND_MOD_EXPR:
case TRUNC_MOD_EXPR:
+ case RDIV_EXPR:
/* Division and multiplication are usually expensive. */
cost += 20;
break;
+ case LSHIFT_EXPR:
+ case RSHIFT_EXPR:
+ cost += 20;
+ break;
+
default:
break;
}
return cost;
}
+/* Finds the outermost loop between OUTER and LOOP in that the memory reference
+ REF is independent. If REF is not independent in LOOP, NULL is returned
+ instead. */
+
+static struct loop *
+outermost_indep_loop (struct loop *outer, struct loop *loop, mem_ref_p ref)
+{
+ struct loop *aloop;
+
+ if (bitmap_bit_p (ref->stored, loop->num))
+ return NULL;
+
+ for (aloop = outer;
+ aloop != loop;
+ aloop = superloop_at_depth (loop, loop_depth (aloop) + 1))
+ if (!bitmap_bit_p (ref->stored, aloop->num)
+ && ref_indep_loop_p (aloop, ref))
+ return aloop;
+
+ if (ref_indep_loop_p (loop, ref))
+ return loop;
+ else
+ return NULL;
+}
+
+/* If there is a simple load or store to a memory reference in STMT, returns
+ the location of the memory reference, and sets IS_STORE according to whether
+ it is a store or load. Otherwise, returns NULL. */
+
+static tree *
+simple_mem_ref_in_stmt (gimple stmt, bool *is_store)
+{
+ tree *lhs;
+ enum tree_code code;
+
+ /* Recognize MEM = (SSA_NAME | invariant) and SSA_NAME = MEM patterns. */
+ if (gimple_code (stmt) != GIMPLE_ASSIGN)
+ return NULL;
+
+ code = gimple_assign_rhs_code (stmt);
+
+ lhs = gimple_assign_lhs_ptr (stmt);
+
+ if (TREE_CODE (*lhs) == SSA_NAME)
+ {
+ if (get_gimple_rhs_class (code) != GIMPLE_SINGLE_RHS
+ || !is_gimple_addressable (gimple_assign_rhs1 (stmt)))
+ return NULL;
+
+ *is_store = false;
+ return gimple_assign_rhs1_ptr (stmt);
+ }
+ else if (code == SSA_NAME
+ || (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS
+ && is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
+ {
+ *is_store = true;
+ return lhs;
+ }
+ else
+ return NULL;
+}
+
+/* Returns the memory reference contained in STMT. */
+
+static mem_ref_p
+mem_ref_in_stmt (gimple stmt)
+{
+ bool store;
+ tree *mem = simple_mem_ref_in_stmt (stmt, &store);
+ hashval_t hash;
+ mem_ref_p ref;
+
+ if (!mem)
+ return NULL;
+ gcc_assert (!store);
+
+ hash = iterative_hash_expr (*mem, 0);
+ ref = (mem_ref_p) htab_find_with_hash (memory_accesses.refs, *mem, hash);
+
+ gcc_assert (ref != NULL);
+ return ref;
+}
+
+/* From a controlling predicate in DOM determine the arguments from
+ the PHI node PHI that are chosen if the predicate evaluates to
+ true and false and store them to *TRUE_ARG_P and *FALSE_ARG_P if
+ they are non-NULL. Returns true if the arguments can be determined,
+ else return false. */
+
+static bool
+extract_true_false_args_from_phi (basic_block dom, gimple phi,
+ tree *true_arg_p, tree *false_arg_p)
+{
+ basic_block bb = gimple_bb (phi);
+ edge true_edge, false_edge, tem;
+ tree arg0 = NULL_TREE, arg1 = NULL_TREE;
+
+ /* We have to verify that one edge into the PHI node is dominated
+ by the true edge of the predicate block and the other edge
+ dominated by the false edge. This ensures that the PHI argument
+ we are going to take is completely determined by the path we
+ take from the predicate block. */
+ extract_true_false_edges_from_block (dom, &true_edge, &false_edge);
+ tem = EDGE_PRED (bb, 0);
+ if (tem == true_edge
+ || tem->src == true_edge->dest
+ || dominated_by_p (CDI_DOMINATORS,
+ tem->src, true_edge->dest))
+ arg0 = PHI_ARG_DEF (phi, tem->dest_idx);
+ else if (tem == false_edge
+ || tem->src == false_edge->dest
+ || dominated_by_p (CDI_DOMINATORS,
+ tem->src, false_edge->dest))
+ arg1 = PHI_ARG_DEF (phi, tem->dest_idx);
+ else
+ return false;
+ tem = EDGE_PRED (bb, 1);
+ if (tem == true_edge
+ || tem->src == true_edge->dest
+ || dominated_by_p (CDI_DOMINATORS,
+ tem->src, true_edge->dest))
+ arg0 = PHI_ARG_DEF (phi, tem->dest_idx);
+ else if (tem == false_edge
+ || tem->src == false_edge->dest
+ || dominated_by_p (CDI_DOMINATORS,
+ tem->src, false_edge->dest))
+ arg1 = PHI_ARG_DEF (phi, tem->dest_idx);
+ else
+ return false;
+ if (!arg0 || !arg1)
+ return false;
+
+ if (true_arg_p)
+ *true_arg_p = arg0;
+ if (false_arg_p)
+ *false_arg_p = arg1;
+
+ return true;
+}
+
/* Determine the outermost loop to that it is possible to hoist a statement
STMT and store it to LIM_DATA (STMT)->max_loop. To do this we determine
the outermost loop in that the value computed by STMT is invariant.
If MUST_PRESERVE_EXEC is true, additionally choose such a loop that
we preserve the fact whether STMT is executed. It also fills other related
information to LIM_DATA (STMT).
-
+
The function returns false if STMT cannot be hoisted outside of the loop it
is defined in, and true otherwise. */
static bool
-determine_max_movement (tree stmt, bool must_preserve_exec)
+determine_max_movement (gimple stmt, bool must_preserve_exec)
{
- basic_block bb = bb_for_stmt (stmt);
+ basic_block bb = gimple_bb (stmt);
struct loop *loop = bb->loop_father;
struct loop *level;
- struct lim_aux_data *lim_data = LIM_DATA (stmt);
+ struct lim_aux_data *lim_data = get_lim_data (stmt);
tree val;
ssa_op_iter iter;
-
+
if (must_preserve_exec)
level = ALWAYS_EXECUTED_IN (bb);
else
level = superloop_at_depth (loop, 1);
lim_data->max_loop = level;
- FOR_EACH_SSA_TREE_OPERAND (val, stmt, iter, SSA_OP_USE)
- if (!add_dependency (val, lim_data, loop, true))
- return false;
+ if (gimple_code (stmt) == GIMPLE_PHI)
+ {
+ use_operand_p use_p;
+ unsigned min_cost = UINT_MAX;
+ unsigned total_cost = 0;
+ struct lim_aux_data *def_data;
+
+ /* We will end up promoting dependencies to be unconditionally
+ evaluated. For this reason the PHI cost (and thus the
+ cost we remove from the loop by doing the invariant motion)
+ is that of the cheapest PHI argument dependency chain. */
+ FOR_EACH_PHI_ARG (use_p, stmt, iter, SSA_OP_USE)
+ {
+ val = USE_FROM_PTR (use_p);
+ if (TREE_CODE (val) != SSA_NAME)
+ continue;
+ if (!add_dependency (val, lim_data, loop, false))
+ return false;
+ def_data = get_lim_data (SSA_NAME_DEF_STMT (val));
+ if (def_data)
+ {
+ min_cost = MIN (min_cost, def_data->cost);
+ total_cost += def_data->cost;
+ }
+ }
- FOR_EACH_SSA_TREE_OPERAND (val, stmt, iter, SSA_OP_VIRTUAL_USES | SSA_OP_VIRTUAL_KILLS)
- if (!add_dependency (val, lim_data, loop, false))
- return false;
+ lim_data->cost += min_cost;
+
+ if (gimple_phi_num_args (stmt) > 1)
+ {
+ basic_block dom = get_immediate_dominator (CDI_DOMINATORS, bb);
+ gimple cond;
+ if (gsi_end_p (gsi_last_bb (dom)))
+ return false;
+ cond = gsi_stmt (gsi_last_bb (dom));
+ if (gimple_code (cond) != GIMPLE_COND)
+ return false;
+ /* Verify that this is an extended form of a diamond and
+ the PHI arguments are completely controlled by the
+ predicate in DOM. */
+ if (!extract_true_false_args_from_phi (dom, stmt, NULL, NULL))
+ return false;
+
+ /* Fold in dependencies and cost of the condition. */
+ FOR_EACH_SSA_TREE_OPERAND (val, cond, iter, SSA_OP_USE)
+ {
+ if (!add_dependency (val, lim_data, loop, false))
+ return false;
+ def_data = get_lim_data (SSA_NAME_DEF_STMT (val));
+ if (def_data)
+ total_cost += def_data->cost;
+ }
+
+ /* We want to avoid unconditionally executing very expensive
+ operations. As costs for our dependencies cannot be
+ negative just claim we are not invariand for this case.
+ We also are not sure whether the control-flow inside the
+ loop will vanish. */
+ if (total_cost - min_cost >= 2 * LIM_EXPENSIVE
+ && !(min_cost != 0
+ && total_cost / min_cost <= 2))
+ return false;
+
+ /* Assume that the control-flow in the loop will vanish.
+ ??? We should verify this and not artificially increase
+ the cost if that is not the case. */
+ lim_data->cost += stmt_cost (stmt);
+ }
+
+ return true;
+ }
+ else
+ FOR_EACH_SSA_TREE_OPERAND (val, stmt, iter, SSA_OP_USE)
+ if (!add_dependency (val, lim_data, loop, true))
+ return false;
+
+ if (gimple_vuse (stmt))
+ {
+ mem_ref_p ref = mem_ref_in_stmt (stmt);
+
+ if (ref)
+ {
+ lim_data->max_loop
+ = outermost_indep_loop (lim_data->max_loop, loop, ref);
+ if (!lim_data->max_loop)
+ return false;
+ }
+ else
+ {
+ if ((val = gimple_vuse (stmt)) != NULL_TREE)
+ {
+ if (!add_dependency (val, lim_data, loop, false))
+ return false;
+ }
+ }
+ }
lim_data->cost += stmt_cost (stmt);
operands) is hoisted at least out of the loop LEVEL. */
static void
-set_level (tree stmt, struct loop *orig_loop, struct loop *level)
+set_level (gimple stmt, struct loop *orig_loop, struct loop *level)
{
- struct loop *stmt_loop = bb_for_stmt (stmt)->loop_father;
+ struct loop *stmt_loop = gimple_bb (stmt)->loop_father;
struct depend *dep;
+ struct lim_aux_data *lim_data;
stmt_loop = find_common_loop (orig_loop, stmt_loop);
- if (LIM_DATA (stmt) && LIM_DATA (stmt)->tgt_loop)
+ lim_data = get_lim_data (stmt);
+ if (lim_data != NULL && lim_data->tgt_loop != NULL)
stmt_loop = find_common_loop (stmt_loop,
- LIM_DATA (stmt)->tgt_loop->outer);
+ loop_outer (lim_data->tgt_loop));
if (flow_loop_nested_p (stmt_loop, level))
return;
- gcc_assert (LIM_DATA (stmt));
- gcc_assert (level == LIM_DATA (stmt)->max_loop
- || flow_loop_nested_p (LIM_DATA (stmt)->max_loop, level));
+ gcc_assert (level == lim_data->max_loop
+ || flow_loop_nested_p (lim_data->max_loop, level));
- LIM_DATA (stmt)->tgt_loop = level;
- for (dep = LIM_DATA (stmt)->depends; dep; dep = dep->next)
+ lim_data->tgt_loop = level;
+ for (dep = lim_data->depends; dep; dep = dep->next)
set_level (dep->stmt, orig_loop, level);
}
information to set it more sanely. */
static void
-set_profitable_level (tree stmt)
+set_profitable_level (gimple stmt)
{
- set_level (stmt, bb_for_stmt (stmt)->loop_father, LIM_DATA (stmt)->max_loop);
+ set_level (stmt, gimple_bb (stmt)->loop_father, get_lim_data (stmt)->max_loop);
}
-/* Returns true if STMT is not a pure call. */
+/* Returns true if STMT is a call that has side effects. */
static bool
-nonpure_call_p (tree stmt)
+nonpure_call_p (gimple stmt)
{
- tree call = get_call_expr_in (stmt);
-
- if (!call)
+ if (gimple_code (stmt) != GIMPLE_CALL)
return false;
- return TREE_SIDE_EFFECTS (call) != 0;
+ return gimple_has_side_effects (stmt);
}
-/* Releases the memory occupied by DATA. */
+/* Rewrite a/b to a*(1/b). Return the invariant stmt to process. */
-static void
-free_lim_aux_data (struct lim_aux_data *data)
+static gimple
+rewrite_reciprocal (gimple_stmt_iterator *bsi)
{
- struct depend *dep, *next;
+ gimple stmt, stmt1, stmt2;
+ tree var, name, lhs, type;
+ tree real_one;
+ gimple_stmt_iterator gsi;
- for (dep = data->depends; dep; dep = next)
+ stmt = gsi_stmt (*bsi);
+ lhs = gimple_assign_lhs (stmt);
+ type = TREE_TYPE (lhs);
+
+ var = create_tmp_var (type, "reciptmp");
+ add_referenced_var (var);
+ DECL_GIMPLE_REG_P (var) = 1;
+
+ /* For vectors, create a VECTOR_CST full of 1's. */
+ if (TREE_CODE (type) == VECTOR_TYPE)
{
- next = dep->next;
- free (dep);
+ int i, len;
+ tree list = NULL_TREE;
+ real_one = build_real (TREE_TYPE (type), dconst1);
+ len = TYPE_VECTOR_SUBPARTS (type);
+ for (i = 0; i < len; i++)
+ list = tree_cons (NULL, real_one, list);
+ real_one = build_vector (type, list);
}
- free (data);
+ else
+ real_one = build_real (type, dconst1);
+
+ stmt1 = gimple_build_assign_with_ops (RDIV_EXPR,
+ var, real_one, gimple_assign_rhs2 (stmt));
+ name = make_ssa_name (var, stmt1);
+ gimple_assign_set_lhs (stmt1, name);
+
+ stmt2 = gimple_build_assign_with_ops (MULT_EXPR, lhs, name,
+ gimple_assign_rhs1 (stmt));
+
+ /* Replace division stmt with reciprocal and multiply stmts.
+ The multiply stmt is not invariant, so update iterator
+ and avoid rescanning. */
+ gsi = *bsi;
+ gsi_insert_before (bsi, stmt1, GSI_NEW_STMT);
+ gsi_replace (&gsi, stmt2, true);
+
+ /* Continue processing with invariant reciprocal statement. */
+ return stmt1;
+}
+
+/* Check if the pattern at *BSI is a bittest of the form
+ (A >> B) & 1 != 0 and in this case rewrite it to A & (1 << B) != 0. */
+
+static gimple
+rewrite_bittest (gimple_stmt_iterator *bsi)
+{
+ gimple stmt, use_stmt, stmt1, stmt2;
+ tree lhs, var, name, t, a, b;
+ use_operand_p use;
+
+ stmt = gsi_stmt (*bsi);
+ lhs = gimple_assign_lhs (stmt);
+
+ /* Verify that the single use of lhs is a comparison against zero. */
+ if (TREE_CODE (lhs) != SSA_NAME
+ || !single_imm_use (lhs, &use, &use_stmt)
+ || gimple_code (use_stmt) != GIMPLE_COND)
+ return stmt;
+ if (gimple_cond_lhs (use_stmt) != lhs
+ || (gimple_cond_code (use_stmt) != NE_EXPR
+ && gimple_cond_code (use_stmt) != EQ_EXPR)
+ || !integer_zerop (gimple_cond_rhs (use_stmt)))
+ return stmt;
+
+ /* Get at the operands of the shift. The rhs is TMP1 & 1. */
+ stmt1 = SSA_NAME_DEF_STMT (gimple_assign_rhs1 (stmt));
+ if (gimple_code (stmt1) != GIMPLE_ASSIGN)
+ return stmt;
+
+ /* There is a conversion in between possibly inserted by fold. */
+ if (CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt1)))
+ {
+ t = gimple_assign_rhs1 (stmt1);
+ if (TREE_CODE (t) != SSA_NAME
+ || !has_single_use (t))
+ return stmt;
+ stmt1 = SSA_NAME_DEF_STMT (t);
+ if (gimple_code (stmt1) != GIMPLE_ASSIGN)
+ return stmt;
+ }
+
+ /* Verify that B is loop invariant but A is not. Verify that with
+ all the stmt walking we are still in the same loop. */
+ if (gimple_assign_rhs_code (stmt1) != RSHIFT_EXPR
+ || loop_containing_stmt (stmt1) != loop_containing_stmt (stmt))
+ return stmt;
+
+ a = gimple_assign_rhs1 (stmt1);
+ b = gimple_assign_rhs2 (stmt1);
+
+ if (outermost_invariant_loop (b, loop_containing_stmt (stmt1)) != NULL
+ && outermost_invariant_loop (a, loop_containing_stmt (stmt1)) == NULL)
+ {
+ gimple_stmt_iterator rsi;
+
+ /* 1 << B */
+ var = create_tmp_var (TREE_TYPE (a), "shifttmp");
+ add_referenced_var (var);
+ t = fold_build2 (LSHIFT_EXPR, TREE_TYPE (a),
+ build_int_cst (TREE_TYPE (a), 1), b);
+ stmt1 = gimple_build_assign (var, t);
+ name = make_ssa_name (var, stmt1);
+ gimple_assign_set_lhs (stmt1, name);
+
+ /* A & (1 << B) */
+ t = fold_build2 (BIT_AND_EXPR, TREE_TYPE (a), a, name);
+ stmt2 = gimple_build_assign (var, t);
+ name = make_ssa_name (var, stmt2);
+ gimple_assign_set_lhs (stmt2, name);
+
+ /* Replace the SSA_NAME we compare against zero. Adjust
+ the type of zero accordingly. */
+ SET_USE (use, name);
+ gimple_cond_set_rhs (use_stmt, build_int_cst_type (TREE_TYPE (name), 0));
+
+ /* Don't use gsi_replace here, none of the new assignments sets
+ the variable originally set in stmt. Move bsi to stmt1, and
+ then remove the original stmt, so that we get a chance to
+ retain debug info for it. */
+ rsi = *bsi;
+ gsi_insert_before (bsi, stmt1, GSI_NEW_STMT);
+ gsi_insert_before (&rsi, stmt2, GSI_SAME_STMT);
+ gsi_remove (&rsi, true);
+
+ return stmt1;
+ }
+
+ return stmt;
}
+
/* Determine the outermost loops in that statements in basic block BB are
invariant, and record them to the LIM_DATA associated with the statements.
Callback for walk_dominator_tree. */
basic_block bb)
{
enum move_pos pos;
- block_stmt_iterator bsi;
- tree stmt;
+ gimple_stmt_iterator bsi;
+ gimple stmt;
bool maybe_never = ALWAYS_EXECUTED_IN (bb) == NULL;
struct loop *outermost = ALWAYS_EXECUTED_IN (bb);
+ struct lim_aux_data *lim_data;
- if (!bb->loop_father->outer)
+ if (!loop_outer (bb->loop_father))
return;
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "Basic block %d (loop %d -- depth %d):\n\n",
- bb->index, bb->loop_father->num, bb->loop_father->depth);
+ bb->index, bb->loop_father->num, loop_depth (bb->loop_father));
+
+ /* Look at PHI nodes, but only if there is at most two.
+ ??? We could relax this further by post-processing the inserted
+ code and transforming adjacent cond-exprs with the same predicate
+ to control flow again. */
+ bsi = gsi_start_phis (bb);
+ if (!gsi_end_p (bsi)
+ && ((gsi_next (&bsi), gsi_end_p (bsi))
+ || (gsi_next (&bsi), gsi_end_p (bsi))))
+ for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi))
+ {
+ stmt = gsi_stmt (bsi);
+
+ pos = movement_possibility (stmt);
+ if (pos == MOVE_IMPOSSIBLE)
+ continue;
+
+ lim_data = init_lim_data (stmt);
+ lim_data->always_executed_in = outermost;
+
+ if (!determine_max_movement (stmt, false))
+ {
+ lim_data->max_loop = NULL;
+ continue;
+ }
+
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ {
+ print_gimple_stmt (dump_file, stmt, 2, 0);
+ fprintf (dump_file, " invariant up to level %d, cost %d.\n\n",
+ loop_depth (lim_data->max_loop),
+ lim_data->cost);
+ }
- for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
+ if (lim_data->cost >= LIM_EXPENSIVE)
+ set_profitable_level (stmt);
+ }
+
+ for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
{
- stmt = bsi_stmt (bsi);
+ stmt = gsi_stmt (bsi);
pos = movement_possibility (stmt);
if (pos == MOVE_IMPOSSIBLE)
maybe_never = true;
outermost = NULL;
}
+ /* Make sure to note always_executed_in for stores to make
+ store-motion work. */
+ else if (stmt_makes_single_store (stmt))
+ {
+ struct lim_aux_data *lim_data = init_lim_data (stmt);
+ lim_data->always_executed_in = outermost;
+ }
continue;
}
- stmt_ann (stmt)->common.aux = xcalloc (1, sizeof (struct lim_aux_data));
- LIM_DATA (stmt)->always_executed_in = outermost;
+ if (is_gimple_assign (stmt)
+ && (get_gimple_rhs_class (gimple_assign_rhs_code (stmt))
+ == GIMPLE_BINARY_RHS))
+ {
+ tree op0 = gimple_assign_rhs1 (stmt);
+ tree op1 = gimple_assign_rhs2 (stmt);
+ struct loop *ol1 = outermost_invariant_loop (op1,
+ loop_containing_stmt (stmt));
+
+ /* If divisor is invariant, convert a/b to a*(1/b), allowing reciprocal
+ to be hoisted out of loop, saving expensive divide. */
+ if (pos == MOVE_POSSIBLE
+ && gimple_assign_rhs_code (stmt) == RDIV_EXPR
+ && flag_unsafe_math_optimizations
+ && !flag_trapping_math
+ && ol1 != NULL
+ && outermost_invariant_loop (op0, ol1) == NULL)
+ stmt = rewrite_reciprocal (&bsi);
+
+ /* If the shift count is invariant, convert (A >> B) & 1 to
+ A & (1 << B) allowing the bit mask to be hoisted out of the loop
+ saving an expensive shift. */
+ if (pos == MOVE_POSSIBLE
+ && gimple_assign_rhs_code (stmt) == BIT_AND_EXPR
+ && integer_onep (op1)
+ && TREE_CODE (op0) == SSA_NAME
+ && has_single_use (op0))
+ stmt = rewrite_bittest (&bsi);
+ }
+
+ lim_data = init_lim_data (stmt);
+ lim_data->always_executed_in = outermost;
if (maybe_never && pos == MOVE_PRESERVE_EXECUTION)
continue;
if (!determine_max_movement (stmt, pos == MOVE_PRESERVE_EXECUTION))
{
- LIM_DATA (stmt)->max_loop = NULL;
+ lim_data->max_loop = NULL;
continue;
}
if (dump_file && (dump_flags & TDF_DETAILS))
{
- print_generic_stmt_indented (dump_file, stmt, 0, 2);
+ print_gimple_stmt (dump_file, stmt, 2, 0);
fprintf (dump_file, " invariant up to level %d, cost %d.\n\n",
- LIM_DATA (stmt)->max_loop->depth,
- LIM_DATA (stmt)->cost);
+ loop_depth (lim_data->max_loop),
+ lim_data->cost);
}
- if (LIM_DATA (stmt)->cost >= LIM_EXPENSIVE)
+ if (lim_data->cost >= LIM_EXPENSIVE)
set_profitable_level (stmt);
}
}
struct dom_walk_data walk_data;
memset (&walk_data, 0, sizeof (struct dom_walk_data));
- walk_data.before_dom_children_before_stmts = determine_invariantness_stmt;
+ walk_data.dom_direction = CDI_DOMINATORS;
+ walk_data.before_dom_children = determine_invariantness_stmt;
init_walk_dominator_tree (&walk_data);
walk_dominator_tree (&walk_data, ENTRY_BLOCK_PTR);
fini_walk_dominator_tree (&walk_data);
}
-/* Commits edge insertions and updates loop structures. */
-
-void
-loop_commit_inserts (void)
-{
- unsigned old_last_basic_block, i;
- basic_block bb;
-
- old_last_basic_block = last_basic_block;
- bsi_commit_edge_inserts ();
- for (i = old_last_basic_block; i < (unsigned) last_basic_block; i++)
- {
- bb = BASIC_BLOCK (i);
- add_bb_to_loop (bb,
- find_common_loop (EDGE_SUCC (bb, 0)->dest->loop_father,
- EDGE_PRED (bb, 0)->src->loop_father));
- }
-}
-
/* Hoist the statements in basic block BB out of the loops prescribed by
data stored in LIM_DATA structures associated with each statement. Callback
for walk_dominator_tree. */
static void
-move_computations_stmt (struct dom_walk_data *dw_data ATTRIBUTE_UNUSED,
+move_computations_stmt (struct dom_walk_data *dw_data,
basic_block bb)
{
struct loop *level;
- block_stmt_iterator bsi;
- tree stmt;
+ gimple_stmt_iterator bsi;
+ gimple stmt;
unsigned cost = 0;
+ struct lim_aux_data *lim_data;
- if (!bb->loop_father->outer)
+ if (!loop_outer (bb->loop_father))
return;
- for (bsi = bsi_start (bb); !bsi_end_p (bsi); )
+ for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); )
{
- stmt = bsi_stmt (bsi);
+ gimple new_stmt;
+ stmt = gsi_stmt (bsi);
- if (!LIM_DATA (stmt))
+ lim_data = get_lim_data (stmt);
+ if (lim_data == NULL)
{
- bsi_next (&bsi);
+ gsi_next (&bsi);
continue;
}
- cost = LIM_DATA (stmt)->cost;
- level = LIM_DATA (stmt)->tgt_loop;
- free_lim_aux_data (LIM_DATA (stmt));
- stmt_ann (stmt)->common.aux = NULL;
+ cost = lim_data->cost;
+ level = lim_data->tgt_loop;
+ clear_lim_data (stmt);
if (!level)
{
- bsi_next (&bsi);
+ gsi_next (&bsi);
continue;
}
- /* We do not really want to move conditionals out of the loop; we just
- placed it here to force its operands to be moved if necessary. */
- if (TREE_CODE (stmt) == COND_EXPR)
- continue;
-
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ {
+ fprintf (dump_file, "Moving PHI node\n");
+ print_gimple_stmt (dump_file, stmt, 0, 0);
+ fprintf (dump_file, "(cost %u) out of loop %d.\n\n",
+ cost, level->num);
+ }
+
+ if (gimple_phi_num_args (stmt) == 1)
+ {
+ tree arg = PHI_ARG_DEF (stmt, 0);
+ new_stmt = gimple_build_assign_with_ops (TREE_CODE (arg),
+ gimple_phi_result (stmt),
+ arg, NULL_TREE);
+ SSA_NAME_DEF_STMT (gimple_phi_result (stmt)) = new_stmt;
+ }
+ else
+ {
+ basic_block dom = get_immediate_dominator (CDI_DOMINATORS, bb);
+ gimple cond = gsi_stmt (gsi_last_bb (dom));
+ tree arg0 = NULL_TREE, arg1 = NULL_TREE, t;
+ /* Get the PHI arguments corresponding to the true and false
+ edges of COND. */
+ extract_true_false_args_from_phi (dom, stmt, &arg0, &arg1);
+ gcc_assert (arg0 && arg1);
+ t = build2 (gimple_cond_code (cond), boolean_type_node,
+ gimple_cond_lhs (cond), gimple_cond_rhs (cond));
+ t = build3 (COND_EXPR, TREE_TYPE (gimple_phi_result (stmt)),
+ t, arg0, arg1);
+ new_stmt = gimple_build_assign_with_ops (COND_EXPR,
+ gimple_phi_result (stmt),
+ t, NULL_TREE);
+ SSA_NAME_DEF_STMT (gimple_phi_result (stmt)) = new_stmt;
+ *((unsigned int *)(dw_data->global_data)) |= TODO_cleanup_cfg;
+ }
+ gsi_insert_on_edge (loop_preheader_edge (level), new_stmt);
+ remove_phi_node (&bsi, false);
+ }
+
+ for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); )
+ {
+ stmt = gsi_stmt (bsi);
+
+ lim_data = get_lim_data (stmt);
+ if (lim_data == NULL)
+ {
+ gsi_next (&bsi);
+ continue;
+ }
+
+ cost = lim_data->cost;
+ level = lim_data->tgt_loop;
+ clear_lim_data (stmt);
+
+ if (!level)
+ {
+ gsi_next (&bsi);
+ continue;
+ }
+
+ /* We do not really want to move conditionals out of the loop; we just
+ placed it here to force its operands to be moved if necessary. */
+ if (gimple_code (stmt) == GIMPLE_COND)
+ continue;
+
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, "Moving statement\n");
- print_generic_stmt (dump_file, stmt, 0);
+ print_gimple_stmt (dump_file, stmt, 0, 0);
fprintf (dump_file, "(cost %u) out of loop %d.\n\n",
cost, level->num);
}
- bsi_insert_on_edge (loop_preheader_edge (level), stmt);
- bsi_remove (&bsi);
+
+ mark_virtual_ops_for_renaming (stmt);
+ gsi_insert_on_edge (loop_preheader_edge (level), stmt);
+ gsi_remove (&bsi, false);
}
}
/* Hoist the statements out of the loops prescribed by data stored in
LIM_DATA structures associated with each statement.*/
-static void
+static unsigned int
move_computations (void)
{
struct dom_walk_data walk_data;
+ unsigned int todo = 0;
memset (&walk_data, 0, sizeof (struct dom_walk_data));
- walk_data.before_dom_children_before_stmts = move_computations_stmt;
+ walk_data.global_data = &todo;
+ walk_data.dom_direction = CDI_DOMINATORS;
+ walk_data.before_dom_children = move_computations_stmt;
init_walk_dominator_tree (&walk_data);
walk_dominator_tree (&walk_data, ENTRY_BLOCK_PTR);
fini_walk_dominator_tree (&walk_data);
- loop_commit_inserts ();
- rewrite_into_ssa (false);
- if (!bitmap_empty_p (vars_to_rename))
- {
- /* The rewrite of ssa names may cause violation of loop closed ssa
- form invariants. TODO -- avoid these rewrites completely.
- Information in virtual phi nodes is sufficient for it. */
- rewrite_into_loop_closed_ssa (NULL);
- }
- bitmap_clear (vars_to_rename);
+ gsi_commit_edge_inserts ();
+ if (need_ssa_update_p (cfun))
+ rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa);
+
+ return todo;
}
/* Checks whether the statement defining variable *INDEX can be hoisted
static bool
may_move_till (tree ref, tree *index, void *data)
{
- struct loop *loop = data, *max_loop;
+ struct loop *loop = (struct loop *) data, *max_loop;
/* If REF is an array reference, check also that the step and the lower
bound is invariant in LOOP. */
if (TREE_CODE (ref) == ARRAY_REF)
{
- tree step = array_ref_element_size (ref);
- tree lbound = array_ref_low_bound (ref);
+ tree step = TREE_OPERAND (ref, 3);
+ tree lbound = TREE_OPERAND (ref, 2);
- max_loop = outermost_invariant_loop_expr (step, loop);
+ max_loop = outermost_invariant_loop (step, loop);
if (!max_loop)
return false;
- max_loop = outermost_invariant_loop_expr (lbound, loop);
+ max_loop = outermost_invariant_loop (lbound, loop);
if (!max_loop)
return false;
}
return true;
}
-/* Forces statements defining (invariant) SSA names in expression EXPR to be
+/* If OP is SSA NAME, force the statement that defines it to be
moved out of the LOOP. ORIG_LOOP is the loop in that EXPR is used. */
static void
-force_move_till_expr (tree expr, struct loop *orig_loop, struct loop *loop)
+force_move_till_op (tree op, struct loop *orig_loop, struct loop *loop)
{
- enum tree_code_class class = TREE_CODE_CLASS (TREE_CODE (expr));
- unsigned i, nops;
+ gimple stmt;
- if (TREE_CODE (expr) == SSA_NAME)
- {
- tree stmt = SSA_NAME_DEF_STMT (expr);
- if (IS_EMPTY_STMT (stmt))
- return;
+ if (!op
+ || is_gimple_min_invariant (op))
+ return;
- set_level (stmt, orig_loop, loop);
- return;
- }
+ gcc_assert (TREE_CODE (op) == SSA_NAME);
- if (class != tcc_unary
- && class != tcc_binary
- && class != tcc_expression
- && class != tcc_comparison)
+ stmt = SSA_NAME_DEF_STMT (op);
+ if (gimple_nop_p (stmt))
return;
- nops = TREE_CODE_LENGTH (TREE_CODE (expr));
- for (i = 0; i < nops; i++)
- force_move_till_expr (TREE_OPERAND (expr, i), orig_loop, loop);
+ set_level (stmt, orig_loop, loop);
}
/* Forces statement defining invariants in REF (and *INDEX) to be moved out of
static bool
force_move_till (tree ref, tree *index, void *data)
{
- tree stmt;
- struct fmt_data *fmt_data = data;
+ struct fmt_data *fmt_data = (struct fmt_data *) data;
if (TREE_CODE (ref) == ARRAY_REF)
{
- tree step = array_ref_element_size (ref);
- tree lbound = array_ref_low_bound (ref);
+ tree step = TREE_OPERAND (ref, 3);
+ tree lbound = TREE_OPERAND (ref, 2);
- force_move_till_expr (step, fmt_data->orig_loop, fmt_data->loop);
- force_move_till_expr (lbound, fmt_data->orig_loop, fmt_data->loop);
+ force_move_till_op (step, fmt_data->orig_loop, fmt_data->loop);
+ force_move_till_op (lbound, fmt_data->orig_loop, fmt_data->loop);
}
- if (TREE_CODE (*index) != SSA_NAME)
- return true;
+ force_move_till_op (*index, fmt_data->orig_loop, fmt_data->loop);
- stmt = SSA_NAME_DEF_STMT (*index);
- if (IS_EMPTY_STMT (stmt))
- return true;
+ return true;
+}
- set_level (stmt, fmt_data->orig_loop, fmt_data->loop);
+/* A hash function for struct mem_ref object OBJ. */
- return true;
+static hashval_t
+memref_hash (const void *obj)
+{
+ const struct mem_ref *const mem = (const struct mem_ref *) obj;
+
+ return mem->hash;
}
-/* Records memory reference *REF (that occurs in statement STMT)
- to the list MEM_REFS. */
+/* An equality function for struct mem_ref object OBJ1 with
+ memory reference OBJ2. */
+
+static int
+memref_eq (const void *obj1, const void *obj2)
+{
+ const struct mem_ref *const mem1 = (const struct mem_ref *) obj1;
+
+ return operand_equal_p (mem1->mem, (const_tree) obj2, 0);
+}
+
+/* Releases list of memory reference locations ACCS. */
static void
-record_mem_ref (struct mem_ref **mem_refs, tree stmt, tree *ref)
+free_mem_ref_locs (mem_ref_locs_p accs)
{
- struct mem_ref *aref = xmalloc (sizeof (struct mem_ref));
+ unsigned i;
+ mem_ref_loc_p loc;
- aref->stmt = stmt;
- aref->ref = ref;
+ if (!accs)
+ return;
- aref->next = *mem_refs;
- *mem_refs = aref;
+ for (i = 0; VEC_iterate (mem_ref_loc_p, accs->locs, i, loc); i++)
+ free (loc);
+ VEC_free (mem_ref_loc_p, heap, accs->locs);
+ free (accs);
}
-/* Releases list of memory references MEM_REFS. */
+/* A function to free the mem_ref object OBJ. */
static void
-free_mem_refs (struct mem_ref *mem_refs)
+memref_free (void *obj)
+{
+ struct mem_ref *const mem = (struct mem_ref *) obj;
+ unsigned i;
+ mem_ref_locs_p accs;
+
+ BITMAP_FREE (mem->stored);
+ BITMAP_FREE (mem->indep_loop);
+ BITMAP_FREE (mem->dep_loop);
+ BITMAP_FREE (mem->indep_ref);
+ BITMAP_FREE (mem->dep_ref);
+
+ for (i = 0; VEC_iterate (mem_ref_locs_p, mem->accesses_in_loop, i, accs); i++)
+ free_mem_ref_locs (accs);
+ VEC_free (mem_ref_locs_p, heap, mem->accesses_in_loop);
+
+ BITMAP_FREE (mem->vops);
+ free (mem);
+}
+
+/* Allocates and returns a memory reference description for MEM whose hash
+ value is HASH and id is ID. */
+
+static mem_ref_p
+mem_ref_alloc (tree mem, unsigned hash, unsigned id)
+{
+ mem_ref_p ref = XNEW (struct mem_ref);
+ ref->mem = mem;
+ ref->id = id;
+ ref->hash = hash;
+ ref->stored = BITMAP_ALLOC (NULL);
+ ref->indep_loop = BITMAP_ALLOC (NULL);
+ ref->dep_loop = BITMAP_ALLOC (NULL);
+ ref->indep_ref = BITMAP_ALLOC (NULL);
+ ref->dep_ref = BITMAP_ALLOC (NULL);
+ ref->accesses_in_loop = NULL;
+ ref->vops = BITMAP_ALLOC (NULL);
+
+ return ref;
+}
+
+/* Allocates and returns the new list of locations. */
+
+static mem_ref_locs_p
+mem_ref_locs_alloc (void)
{
- struct mem_ref *act;
+ mem_ref_locs_p accs = XNEW (struct mem_ref_locs);
+ accs->locs = NULL;
+ return accs;
+}
- while (mem_refs)
+/* Records memory reference location *LOC in LOOP to the memory reference
+ description REF. The reference occurs in statement STMT. */
+
+static void
+record_mem_ref_loc (mem_ref_p ref, struct loop *loop, gimple stmt, tree *loc)
+{
+ mem_ref_loc_p aref = XNEW (struct mem_ref_loc);
+ mem_ref_locs_p accs;
+ bitmap ril = VEC_index (bitmap, memory_accesses.refs_in_loop, loop->num);
+
+ if (VEC_length (mem_ref_locs_p, ref->accesses_in_loop)
+ <= (unsigned) loop->num)
+ VEC_safe_grow_cleared (mem_ref_locs_p, heap, ref->accesses_in_loop,
+ loop->num + 1);
+ accs = VEC_index (mem_ref_locs_p, ref->accesses_in_loop, loop->num);
+ if (!accs)
{
- act = mem_refs;
- mem_refs = mem_refs->next;
- free (act);
+ accs = mem_ref_locs_alloc ();
+ VEC_replace (mem_ref_locs_p, ref->accesses_in_loop, loop->num, accs);
}
+
+ aref->stmt = stmt;
+ aref->ref = loc;
+
+ VEC_safe_push (mem_ref_loc_p, heap, accs->locs, aref);
+ bitmap_set_bit (ril, ref->id);
+}
+
+/* Marks reference REF as stored in LOOP. */
+
+static void
+mark_ref_stored (mem_ref_p ref, struct loop *loop)
+{
+ for (;
+ loop != current_loops->tree_root
+ && !bitmap_bit_p (ref->stored, loop->num);
+ loop = loop_outer (loop))
+ bitmap_set_bit (ref->stored, loop->num);
}
-/* If VAR is defined in LOOP and the statement it is defined in does not belong
- to the set SEEN, add the statement to QUEUE of length IN_QUEUE and
- to the set SEEN. */
+/* Gathers memory references in statement STMT in LOOP, storing the
+ information about them in the memory_accesses structure. Marks
+ the vops accessed through unrecognized statements there as
+ well. */
static void
-maybe_queue_var (tree var, struct loop *loop,
- sbitmap seen, tree *queue, unsigned *in_queue)
-{
- tree stmt = SSA_NAME_DEF_STMT (var);
- basic_block def_bb = bb_for_stmt (stmt);
-
- if (!def_bb
- || !flow_bb_inside_loop_p (loop, def_bb)
- || TEST_BIT (seen, get_stmt_uid (stmt)))
+gather_mem_refs_stmt (struct loop *loop, gimple stmt)
+{
+ tree *mem = NULL;
+ hashval_t hash;
+ PTR *slot;
+ mem_ref_p ref;
+ tree vname;
+ bool is_stored;
+ bitmap clvops;
+ unsigned id;
+
+ if (!gimple_vuse (stmt))
return;
-
- SET_BIT (seen, get_stmt_uid (stmt));
- queue[(*in_queue)++] = stmt;
+
+ mem = simple_mem_ref_in_stmt (stmt, &is_stored);
+ if (!mem)
+ goto fail;
+
+ hash = iterative_hash_expr (*mem, 0);
+ slot = htab_find_slot_with_hash (memory_accesses.refs, *mem, hash, INSERT);
+
+ if (*slot)
+ {
+ ref = (mem_ref_p) *slot;
+ id = ref->id;
+ }
+ else
+ {
+ id = VEC_length (mem_ref_p, memory_accesses.refs_list);
+ ref = mem_ref_alloc (*mem, hash, id);
+ VEC_safe_push (mem_ref_p, heap, memory_accesses.refs_list, ref);
+ *slot = ref;
+
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ {
+ fprintf (dump_file, "Memory reference %u: ", id);
+ print_generic_expr (dump_file, ref->mem, TDF_SLIM);
+ fprintf (dump_file, "\n");
+ }
+ }
+ if (is_stored)
+ mark_ref_stored (ref, loop);
+
+ if ((vname = gimple_vuse (stmt)) != NULL_TREE)
+ bitmap_set_bit (ref->vops, DECL_UID (SSA_NAME_VAR (vname)));
+ record_mem_ref_loc (ref, loop, stmt, mem);
+ return;
+
+fail:
+ clvops = VEC_index (bitmap, memory_accesses.clobbered_vops, loop->num);
+ if ((vname = gimple_vuse (stmt)) != NULL_TREE)
+ bitmap_set_bit (clvops, DECL_UID (SSA_NAME_VAR (vname)));
+}
+
+/* Gathers memory references in loops. */
+
+static void
+gather_mem_refs_in_loops (void)
+{
+ gimple_stmt_iterator bsi;
+ basic_block bb;
+ struct loop *loop;
+ loop_iterator li;
+ bitmap clvo, clvi;
+ bitmap lrefs, alrefs, alrefso;
+
+ FOR_EACH_BB (bb)
+ {
+ loop = bb->loop_father;
+ if (loop == current_loops->tree_root)
+ continue;
+
+ for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
+ gather_mem_refs_stmt (loop, gsi_stmt (bsi));
+ }
+
+ /* Propagate the information about clobbered vops and accessed memory
+ references up the loop hierarchy. */
+ FOR_EACH_LOOP (li, loop, LI_FROM_INNERMOST)
+ {
+ lrefs = VEC_index (bitmap, memory_accesses.refs_in_loop, loop->num);
+ alrefs = VEC_index (bitmap, memory_accesses.all_refs_in_loop, loop->num);
+ bitmap_ior_into (alrefs, lrefs);
+
+ if (loop_outer (loop) == current_loops->tree_root)
+ continue;
+
+ clvi = VEC_index (bitmap, memory_accesses.clobbered_vops, loop->num);
+ clvo = VEC_index (bitmap, memory_accesses.clobbered_vops,
+ loop_outer (loop)->num);
+ bitmap_ior_into (clvo, clvi);
+
+ alrefso = VEC_index (bitmap, memory_accesses.all_refs_in_loop,
+ loop_outer (loop)->num);
+ bitmap_ior_into (alrefso, alrefs);
+ }
}
-/* If COMMON_REF is NULL, set COMMON_REF to *OP and return true.
- Otherwise return true if the memory reference *OP is equal to COMMON_REF.
- Record the reference OP to list MEM_REFS. STMT is the statement in that
- the reference occurs. */
+/* Element of the hash table that maps vops to memory references. */
-struct sra_data
+struct vop_to_refs_elt
{
- struct mem_ref **mem_refs;
- tree common_ref;
- tree stmt;
+ /* DECL_UID of the vop. */
+ unsigned uid;
+
+ /* List of the all references. */
+ bitmap refs_all;
+
+ /* List of stored references. */
+ bitmap refs_stored;
};
-static bool
-fem_single_reachable_address (tree *op, void *data)
+/* A hash function for struct vop_to_refs_elt object OBJ. */
+
+static hashval_t
+vtoe_hash (const void *obj)
{
- struct sra_data *sra_data = data;
+ const struct vop_to_refs_elt *const vtoe =
+ (const struct vop_to_refs_elt *) obj;
- if (sra_data->common_ref
- && !operand_equal_p (*op, sra_data->common_ref, 0))
- return false;
- sra_data->common_ref = *op;
+ return vtoe->uid;
+}
- record_mem_ref (sra_data->mem_refs, sra_data->stmt, op);
- return true;
+/* An equality function for struct vop_to_refs_elt object OBJ1 with
+ uid of a vop OBJ2. */
+
+static int
+vtoe_eq (const void *obj1, const void *obj2)
+{
+ const struct vop_to_refs_elt *const vtoe =
+ (const struct vop_to_refs_elt *) obj1;
+ const unsigned *const uid = (const unsigned *) obj2;
+
+ return vtoe->uid == *uid;
}
-/* Runs CALLBACK for each operand of STMT that is a memory reference. DATA
- is passed to the CALLBACK as well. The traversal stops if CALLBACK
- returns false, for_each_memref then returns false as well. Otherwise
- for_each_memref returns true. */
+/* A function to free the struct vop_to_refs_elt object. */
-static bool
-for_each_memref (tree stmt, bool (*callback)(tree *, void *), void *data)
+static void
+vtoe_free (void *obj)
{
- tree *op;
+ struct vop_to_refs_elt *const vtoe =
+ (struct vop_to_refs_elt *) obj;
- if (TREE_CODE (stmt) == RETURN_EXPR)
- stmt = TREE_OPERAND (stmt, 1);
+ BITMAP_FREE (vtoe->refs_all);
+ BITMAP_FREE (vtoe->refs_stored);
+ free (vtoe);
+}
- if (TREE_CODE (stmt) == MODIFY_EXPR)
- {
- op = &TREE_OPERAND (stmt, 0);
- if (TREE_CODE (*op) != SSA_NAME
- && !callback (op, data))
- return false;
+/* Records REF to hashtable VOP_TO_REFS for the index VOP. STORED is true
+ if the reference REF is stored. */
- op = &TREE_OPERAND (stmt, 1);
- if (TREE_CODE (*op) != SSA_NAME
- && is_gimple_lvalue (*op)
- && !callback (op, data))
- return false;
+static void
+record_vop_access (htab_t vop_to_refs, unsigned vop, unsigned ref, bool stored)
+{
+ void **slot = htab_find_slot_with_hash (vop_to_refs, &vop, vop, INSERT);
+ struct vop_to_refs_elt *vtoe;
- stmt = TREE_OPERAND (stmt, 1);
+ if (!*slot)
+ {
+ vtoe = XNEW (struct vop_to_refs_elt);
+ vtoe->uid = vop;
+ vtoe->refs_all = BITMAP_ALLOC (NULL);
+ vtoe->refs_stored = BITMAP_ALLOC (NULL);
+ *slot = vtoe;
}
+ else
+ vtoe = (struct vop_to_refs_elt *) *slot;
+
+ bitmap_set_bit (vtoe->refs_all, ref);
+ if (stored)
+ bitmap_set_bit (vtoe->refs_stored, ref);
+}
- if (TREE_CODE (stmt) == WITH_SIZE_EXPR)
- stmt = TREE_OPERAND (stmt, 0);
+/* Returns the set of references that access VOP according to the table
+ VOP_TO_REFS. */
- if (TREE_CODE (stmt) == CALL_EXPR)
+static bitmap
+get_vop_accesses (htab_t vop_to_refs, unsigned vop)
+{
+ struct vop_to_refs_elt *const vtoe =
+ (struct vop_to_refs_elt *) htab_find_with_hash (vop_to_refs, &vop, vop);
+ return vtoe->refs_all;
+}
+
+/* Returns the set of stores that access VOP according to the table
+ VOP_TO_REFS. */
+
+static bitmap
+get_vop_stores (htab_t vop_to_refs, unsigned vop)
+{
+ struct vop_to_refs_elt *const vtoe =
+ (struct vop_to_refs_elt *) htab_find_with_hash (vop_to_refs, &vop, vop);
+ return vtoe->refs_stored;
+}
+
+/* Adds REF to mapping from virtual operands to references in LOOP. */
+
+static void
+add_vop_ref_mapping (struct loop *loop, mem_ref_p ref)
+{
+ htab_t map = VEC_index (htab_t, memory_accesses.vop_ref_map, loop->num);
+ bool stored = bitmap_bit_p (ref->stored, loop->num);
+ bitmap clobbers = VEC_index (bitmap, memory_accesses.clobbered_vops,
+ loop->num);
+ bitmap_iterator bi;
+ unsigned vop;
+
+ EXECUTE_IF_AND_COMPL_IN_BITMAP (ref->vops, clobbers, 0, vop, bi)
{
- tree args;
+ record_vop_access (map, vop, ref->id, stored);
+ }
+}
- for (args = TREE_OPERAND (stmt, 1); args; args = TREE_CHAIN (args))
- {
- op = &TREE_VALUE (args);
+/* Create a mapping from virtual operands to references that touch them
+ in LOOP. */
- if (TREE_CODE (*op) != SSA_NAME
- && is_gimple_lvalue (*op)
- && !callback (op, data))
- return false;
- }
+static void
+create_vop_ref_mapping_loop (struct loop *loop)
+{
+ bitmap refs = VEC_index (bitmap, memory_accesses.refs_in_loop, loop->num);
+ struct loop *sloop;
+ bitmap_iterator bi;
+ unsigned i;
+ mem_ref_p ref;
+
+ EXECUTE_IF_SET_IN_BITMAP (refs, 0, i, bi)
+ {
+ ref = VEC_index (mem_ref_p, memory_accesses.refs_list, i);
+ for (sloop = loop; sloop != current_loops->tree_root; sloop = loop_outer (sloop))
+ add_vop_ref_mapping (sloop, ref);
}
+}
- return true;
+/* For each non-clobbered virtual operand and each loop, record the memory
+ references in this loop that touch the operand. */
+
+static void
+create_vop_ref_mapping (void)
+{
+ loop_iterator li;
+ struct loop *loop;
+
+ FOR_EACH_LOOP (li, loop, 0)
+ {
+ create_vop_ref_mapping_loop (loop);
+ }
}
-/* Determine whether all memory references inside the LOOP that correspond
- to virtual ssa names defined in statement STMT are equal.
- If so, store the list of the references to MEM_REFS, and return one
- of them. Otherwise store NULL to MEM_REFS and return NULL_TREE.
- *SEEN_CALL_STMT is set to true if the virtual operands suggest
- that the reference might be clobbered by a call inside the LOOP. */
-
-static tree
-single_reachable_address (struct loop *loop, tree stmt,
- struct mem_ref **mem_refs,
- bool *seen_call_stmt)
-{
- unsigned max_uid = max_stmt_uid + num_ssa_names;
- tree *queue = xmalloc (sizeof (tree) * max_uid);
- sbitmap seen = sbitmap_alloc (max_uid);
- unsigned in_queue = 1;
- dataflow_t df;
- unsigned i, n;
- struct sra_data sra_data;
- tree call;
- tree val;
- ssa_op_iter iter;
+/* Gathers information about memory accesses in the loops. */
+
+static void
+analyze_memory_references (void)
+{
+ unsigned i;
+ bitmap empty;
+ htab_t hempty;
+
+ memory_accesses.refs
+ = htab_create (100, memref_hash, memref_eq, memref_free);
+ memory_accesses.refs_list = NULL;
+ memory_accesses.refs_in_loop = VEC_alloc (bitmap, heap,
+ number_of_loops ());
+ memory_accesses.all_refs_in_loop = VEC_alloc (bitmap, heap,
+ number_of_loops ());
+ memory_accesses.clobbered_vops = VEC_alloc (bitmap, heap,
+ number_of_loops ());
+ memory_accesses.vop_ref_map = VEC_alloc (htab_t, heap,
+ number_of_loops ());
+
+ for (i = 0; i < number_of_loops (); i++)
+ {
+ empty = BITMAP_ALLOC (NULL);
+ VEC_quick_push (bitmap, memory_accesses.refs_in_loop, empty);
+ empty = BITMAP_ALLOC (NULL);
+ VEC_quick_push (bitmap, memory_accesses.all_refs_in_loop, empty);
+ empty = BITMAP_ALLOC (NULL);
+ VEC_quick_push (bitmap, memory_accesses.clobbered_vops, empty);
+ hempty = htab_create (10, vtoe_hash, vtoe_eq, vtoe_free);
+ VEC_quick_push (htab_t, memory_accesses.vop_ref_map, hempty);
+ }
+
+ memory_accesses.ttae_cache = NULL;
+
+ gather_mem_refs_in_loops ();
+ create_vop_ref_mapping ();
+}
- sbitmap_zero (seen);
+/* Returns true if a region of size SIZE1 at position 0 and a region of
+ size SIZE2 at position DIFF cannot overlap. */
- *mem_refs = NULL;
- sra_data.mem_refs = mem_refs;
- sra_data.common_ref = NULL_TREE;
+static bool
+cannot_overlap_p (aff_tree *diff, double_int size1, double_int size2)
+{
+ double_int d, bound;
- queue[0] = stmt;
- SET_BIT (seen, get_stmt_uid (stmt));
- *seen_call_stmt = false;
+ /* Unless the difference is a constant, we fail. */
+ if (diff->n != 0)
+ return false;
- while (in_queue)
+ d = diff->offset;
+ if (double_int_negative_p (d))
+ {
+ /* The second object is before the first one, we succeed if the last
+ element of the second object is before the start of the first one. */
+ bound = double_int_add (d, double_int_add (size2, double_int_minus_one));
+ return double_int_negative_p (bound);
+ }
+ else
{
- stmt = queue[--in_queue];
- sra_data.stmt = stmt;
+ /* We succeed if the second object starts after the first one ends. */
+ return double_int_scmp (size1, d) <= 0;
+ }
+}
- if (LIM_DATA (stmt)
- && LIM_DATA (stmt)->sm_done)
- goto fail;
+/* Returns true if MEM1 and MEM2 may alias. TTAE_CACHE is used as a cache in
+ tree_to_aff_combination_expand. */
- switch (TREE_CODE (stmt))
- {
- case MODIFY_EXPR:
- case CALL_EXPR:
- case RETURN_EXPR:
- if (!for_each_memref (stmt, fem_single_reachable_address,
- &sra_data))
- goto fail;
-
- /* If this is a function that may depend on the memory location,
- record the fact. We cannot directly refuse call clobbered
- operands here, since sra_data.common_ref did not have
- to be set yet. */
- call = get_call_expr_in (stmt);
- if (call
- && !(call_expr_flags (call) & ECF_CONST))
- *seen_call_stmt = true;
-
- /* Traverse also definitions of the VUSES (there may be other
- distinct from the one we used to get to this statement). */
- FOR_EACH_SSA_TREE_OPERAND (val, stmt, iter, SSA_OP_VIRTUAL_USES)
- maybe_queue_var (val, loop, seen, queue, &in_queue);
+static bool
+mem_refs_may_alias_p (tree mem1, tree mem2, struct pointer_map_t **ttae_cache)
+{
+ /* Perform BASE + OFFSET analysis -- if MEM1 and MEM2 are based on the same
+ object and their offset differ in such a way that the locations cannot
+ overlap, then they cannot alias. */
+ double_int size1, size2;
+ aff_tree off1, off2;
+
+ /* Perform basic offset and type-based disambiguation. */
+ if (!refs_may_alias_p (mem1, mem2))
+ return false;
- break;
+ /* The expansion of addresses may be a bit expensive, thus we only do
+ the check at -O2 and higher optimization levels. */
+ if (optimize < 2)
+ return true;
- case PHI_NODE:
- for (i = 0; i < (unsigned) PHI_NUM_ARGS (stmt); i++)
- if (TREE_CODE (PHI_ARG_DEF (stmt, i)) == SSA_NAME)
- maybe_queue_var (PHI_ARG_DEF (stmt, i), loop,
- seen, queue, &in_queue);
- break;
+ get_inner_reference_aff (mem1, &off1, &size1);
+ get_inner_reference_aff (mem2, &off2, &size2);
+ aff_combination_expand (&off1, ttae_cache);
+ aff_combination_expand (&off2, ttae_cache);
+ aff_combination_scale (&off1, double_int_minus_one);
+ aff_combination_add (&off2, &off1);
- default:
- goto fail;
- }
+ if (cannot_overlap_p (&off2, size1, size2))
+ return false;
+
+ return true;
+}
- /* Find uses of virtual names. */
- df = get_immediate_uses (stmt);
- n = num_immediate_uses (df);
+/* Rewrites location LOC by TMP_VAR. */
- for (i = 0; i < n; i++)
- {
- stmt = immediate_use (df, i);
+static void
+rewrite_mem_ref_loc (mem_ref_loc_p loc, tree tmp_var)
+{
+ mark_virtual_ops_for_renaming (loc->stmt);
+ *loc->ref = tmp_var;
+ update_stmt (loc->stmt);
+}
- if (!flow_bb_inside_loop_p (loop, bb_for_stmt (stmt)))
- continue;
+/* Adds all locations of REF in LOOP and its subloops to LOCS. */
- if (TEST_BIT (seen, get_stmt_uid (stmt)))
- continue;
- SET_BIT (seen, get_stmt_uid (stmt));
+static void
+get_all_locs_in_loop (struct loop *loop, mem_ref_p ref,
+ VEC (mem_ref_loc_p, heap) **locs)
+{
+ mem_ref_locs_p accs;
+ unsigned i;
+ mem_ref_loc_p loc;
+ bitmap refs = VEC_index (bitmap, memory_accesses.all_refs_in_loop,
+ loop->num);
+ struct loop *subloop;
- queue[in_queue++] = stmt;
+ if (!bitmap_bit_p (refs, ref->id))
+ return;
+
+ if (VEC_length (mem_ref_locs_p, ref->accesses_in_loop)
+ > (unsigned) loop->num)
+ {
+ accs = VEC_index (mem_ref_locs_p, ref->accesses_in_loop, loop->num);
+ if (accs)
+ {
+ for (i = 0; VEC_iterate (mem_ref_loc_p, accs->locs, i, loc); i++)
+ VEC_safe_push (mem_ref_loc_p, heap, *locs, loc);
}
}
- free (queue);
- sbitmap_free (seen);
+ for (subloop = loop->inner; subloop != NULL; subloop = subloop->next)
+ get_all_locs_in_loop (subloop, ref, locs);
+}
- return sra_data.common_ref;
+/* Rewrites all references to REF in LOOP by variable TMP_VAR. */
-fail:
- free_mem_refs (*mem_refs);
- *mem_refs = NULL;
- free (queue);
- sbitmap_free (seen);
+static void
+rewrite_mem_refs (struct loop *loop, mem_ref_p ref, tree tmp_var)
+{
+ unsigned i;
+ mem_ref_loc_p loc;
+ VEC (mem_ref_loc_p, heap) *locs = NULL;
- return NULL;
+ get_all_locs_in_loop (loop, ref, &locs);
+ for (i = 0; VEC_iterate (mem_ref_loc_p, locs, i, loc); i++)
+ rewrite_mem_ref_loc (loc, tmp_var);
+ VEC_free (mem_ref_loc_p, heap, locs);
}
-/* Rewrites memory references in list MEM_REFS by variable TMP_VAR. */
+/* The name and the length of the currently generated variable
+ for lsm. */
+#define MAX_LSM_NAME_LENGTH 40
+static char lsm_tmp_name[MAX_LSM_NAME_LENGTH + 1];
+static int lsm_tmp_name_length;
+
+/* Adds S to lsm_tmp_name. */
static void
-rewrite_mem_refs (tree tmp_var, struct mem_ref *mem_refs)
+lsm_tmp_name_add (const char *s)
{
- tree var;
- ssa_op_iter iter;
+ int l = strlen (s) + lsm_tmp_name_length;
+ if (l > MAX_LSM_NAME_LENGTH)
+ return;
+
+ strcpy (lsm_tmp_name + lsm_tmp_name_length, s);
+ lsm_tmp_name_length = l;
+}
+
+/* Stores the name for temporary variable that replaces REF to
+ lsm_tmp_name. */
- for (; mem_refs; mem_refs = mem_refs->next)
+static void
+gen_lsm_tmp_name (tree ref)
+{
+ const char *name;
+
+ switch (TREE_CODE (ref))
{
- FOR_EACH_SSA_TREE_OPERAND (var, mem_refs->stmt, iter, SSA_OP_ALL_VIRTUALS)
- {
- var = SSA_NAME_VAR (var);
- bitmap_set_bit (vars_to_rename, var_ann (var)->uid);
- }
+ case MISALIGNED_INDIRECT_REF:
+ case ALIGN_INDIRECT_REF:
+ case INDIRECT_REF:
+ gen_lsm_tmp_name (TREE_OPERAND (ref, 0));
+ lsm_tmp_name_add ("_");
+ break;
+
+ case BIT_FIELD_REF:
+ case VIEW_CONVERT_EXPR:
+ case ARRAY_RANGE_REF:
+ gen_lsm_tmp_name (TREE_OPERAND (ref, 0));
+ break;
+
+ case REALPART_EXPR:
+ gen_lsm_tmp_name (TREE_OPERAND (ref, 0));
+ lsm_tmp_name_add ("_RE");
+ break;
+
+ case IMAGPART_EXPR:
+ gen_lsm_tmp_name (TREE_OPERAND (ref, 0));
+ lsm_tmp_name_add ("_IM");
+ break;
- *mem_refs->ref = tmp_var;
- modify_stmt (mem_refs->stmt);
+ case COMPONENT_REF:
+ gen_lsm_tmp_name (TREE_OPERAND (ref, 0));
+ lsm_tmp_name_add ("_");
+ name = get_name (TREE_OPERAND (ref, 1));
+ if (!name)
+ name = "F";
+ lsm_tmp_name_add (name);
+ break;
+
+ case ARRAY_REF:
+ gen_lsm_tmp_name (TREE_OPERAND (ref, 0));
+ lsm_tmp_name_add ("_I");
+ break;
+
+ case SSA_NAME:
+ ref = SSA_NAME_VAR (ref);
+ /* Fallthru. */
+
+ case VAR_DECL:
+ case PARM_DECL:
+ name = get_name (ref);
+ if (!name)
+ name = "D";
+ lsm_tmp_name_add (name);
+ break;
+
+ case STRING_CST:
+ lsm_tmp_name_add ("S");
+ break;
+
+ case RESULT_DECL:
+ lsm_tmp_name_add ("R");
+ break;
+
+ case INTEGER_CST:
+ /* Nothing. */
+ break;
+
+ default:
+ gcc_unreachable ();
+ }
+}
+
+/* Determines name for temporary variable that replaces REF.
+ The name is accumulated into the lsm_tmp_name variable.
+ N is added to the name of the temporary. */
+
+char *
+get_lsm_tmp_name (tree ref, unsigned n)
+{
+ char ns[2];
+
+ lsm_tmp_name_length = 0;
+ gen_lsm_tmp_name (ref);
+ lsm_tmp_name_add ("_lsm");
+ if (n < 10)
+ {
+ ns[0] = '0' + n;
+ ns[1] = 0;
+ lsm_tmp_name_add (ns);
}
+ return lsm_tmp_name;
}
-/* Records request for store motion of memory reference REF from LOOP.
- MEM_REFS is the list of occurrences of the reference REF inside LOOP;
- these references are rewritten by a new temporary variable.
- Exits from the LOOP are stored in EXITS, there are N_EXITS of them.
- The initialization of the temporary variable is put to the preheader
- of the loop, and assignments to the reference from the temporary variable
- are emitted to exits. */
+/* Executes store motion of memory reference REF from LOOP.
+ Exits from the LOOP are stored in EXITS. The initialization of the
+ temporary variable is put to the preheader of the loop, and assignments
+ to the reference from the temporary variable are emitted to exits. */
static void
-schedule_sm (struct loop *loop, edge *exits, unsigned n_exits, tree ref,
- struct mem_ref *mem_refs)
+execute_sm (struct loop *loop, VEC (edge, heap) *exits, mem_ref_p ref)
{
- struct mem_ref *aref;
tree tmp_var;
unsigned i;
- tree load, store;
+ gimple load, store;
struct fmt_data fmt_data;
+ edge ex;
+ struct lim_aux_data *lim_data;
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, "Executing store motion of ");
- print_generic_expr (dump_file, ref, 0);
+ print_generic_expr (dump_file, ref->mem, 0);
fprintf (dump_file, " from loop %d\n", loop->num);
}
- tmp_var = make_rename_temp (TREE_TYPE (ref), "lsm_tmp");
+ tmp_var = make_rename_temp (TREE_TYPE (ref->mem),
+ get_lsm_tmp_name (ref->mem, ~0));
fmt_data.loop = loop;
fmt_data.orig_loop = loop;
- for_each_index (&ref, force_move_till, &fmt_data);
+ for_each_index (&ref->mem, force_move_till, &fmt_data);
- rewrite_mem_refs (tmp_var, mem_refs);
- for (aref = mem_refs; aref; aref = aref->next)
- if (LIM_DATA (aref->stmt))
- LIM_DATA (aref->stmt)->sm_done = true;
+ rewrite_mem_refs (loop, ref, tmp_var);
/* Emit the load & stores. */
- load = build (MODIFY_EXPR, void_type_node, tmp_var, ref);
- get_stmt_ann (load)->common.aux = xcalloc (1, sizeof (struct lim_aux_data));
- LIM_DATA (load)->max_loop = loop;
- LIM_DATA (load)->tgt_loop = loop;
+ load = gimple_build_assign (tmp_var, unshare_expr (ref->mem));
+ lim_data = init_lim_data (load);
+ lim_data->max_loop = loop;
+ lim_data->tgt_loop = loop;
/* Put this into the latch, so that we are sure it will be processed after
all dependencies. */
- bsi_insert_on_edge (loop_latch_edge (loop), load);
+ gsi_insert_on_edge (loop_latch_edge (loop), load);
- for (i = 0; i < n_exits; i++)
+ for (i = 0; VEC_iterate (edge, exits, i, ex); i++)
{
- store = build (MODIFY_EXPR, void_type_node,
- unshare_expr (ref), tmp_var);
- bsi_insert_on_edge (exits[i], store);
+ store = gimple_build_assign (unshare_expr (ref->mem), tmp_var);
+ gsi_insert_on_edge (ex, store);
}
}
-/* Returns true if REF may be clobbered by calls. */
+/* Hoists memory references MEM_REFS out of LOOP. EXITS is the list of exit
+ edges of the LOOP. */
-static bool
-is_call_clobbered_ref (tree ref)
+static void
+hoist_memory_references (struct loop *loop, bitmap mem_refs,
+ VEC (edge, heap) *exits)
{
- tree base;
+ mem_ref_p ref;
+ unsigned i;
+ bitmap_iterator bi;
- base = get_base_address (ref);
- if (!base)
- return true;
+ EXECUTE_IF_SET_IN_BITMAP (mem_refs, 0, i, bi)
+ {
+ ref = VEC_index (mem_ref_p, memory_accesses.refs_list, i);
+ execute_sm (loop, exits, ref);
+ }
+}
+
+/* Returns true if REF is always accessed in LOOP. If STORED_P is true
+ make sure REF is always stored to in LOOP. */
- if (DECL_P (base))
- return is_call_clobbered (base);
+static bool
+ref_always_accessed_p (struct loop *loop, mem_ref_p ref, bool stored_p)
+{
+ VEC (mem_ref_loc_p, heap) *locs = NULL;
+ unsigned i;
+ mem_ref_loc_p loc;
+ bool ret = false;
+ struct loop *must_exec;
+ tree base;
+ base = get_base_address (ref->mem);
if (INDIRECT_REF_P (base))
+ base = TREE_OPERAND (base, 0);
+
+ get_all_locs_in_loop (loop, ref, &locs);
+ for (i = 0; VEC_iterate (mem_ref_loc_p, locs, i, loc); i++)
{
- /* Check whether the alias tags associated with the pointer
- are call clobbered. */
- tree ptr = TREE_OPERAND (base, 0);
- struct ptr_info_def *pi = SSA_NAME_PTR_INFO (ptr);
- tree nmt = (pi) ? pi->name_mem_tag : NULL_TREE;
- tree tmt = var_ann (SSA_NAME_VAR (ptr))->type_mem_tag;
+ if (!get_lim_data (loc->stmt))
+ continue;
- if ((nmt && is_call_clobbered (nmt))
- || (tmt && is_call_clobbered (tmt)))
- return true;
+ /* If we require an always executed store make sure the statement
+ stores to the reference. */
+ if (stored_p)
+ {
+ tree lhs;
+ if (!gimple_get_lhs (loc->stmt))
+ continue;
+ lhs = get_base_address (gimple_get_lhs (loc->stmt));
+ if (!lhs)
+ continue;
+ if (INDIRECT_REF_P (lhs))
+ lhs = TREE_OPERAND (lhs, 0);
+ if (lhs != base)
+ continue;
+ }
- return false;
+ must_exec = get_lim_data (loc->stmt)->always_executed_in;
+ if (!must_exec)
+ continue;
+
+ if (must_exec == loop
+ || flow_loop_nested_p (must_exec, loop))
+ {
+ ret = true;
+ break;
+ }
}
+ VEC_free (mem_ref_loc_p, heap, locs);
+
+ return ret;
+}
+
+/* Returns true if REF1 and REF2 are independent. */
+
+static bool
+refs_independent_p (mem_ref_p ref1, mem_ref_p ref2)
+{
+ if (ref1 == ref2
+ || bitmap_bit_p (ref1->indep_ref, ref2->id))
+ return true;
+ if (bitmap_bit_p (ref1->dep_ref, ref2->id))
+ return false;
+
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ fprintf (dump_file, "Querying dependency of refs %u and %u: ",
+ ref1->id, ref2->id);
- gcc_unreachable ();
+ if (mem_refs_may_alias_p (ref1->mem, ref2->mem,
+ &memory_accesses.ttae_cache))
+ {
+ bitmap_set_bit (ref1->dep_ref, ref2->id);
+ bitmap_set_bit (ref2->dep_ref, ref1->id);
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ fprintf (dump_file, "dependent.\n");
+ return false;
+ }
+ else
+ {
+ bitmap_set_bit (ref1->indep_ref, ref2->id);
+ bitmap_set_bit (ref2->indep_ref, ref1->id);
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ fprintf (dump_file, "independent.\n");
+ return true;
+ }
}
-/* Determine whether all memory references inside LOOP corresponding to the
- virtual ssa name REG are equal to each other, and whether the address of
- this common reference can be hoisted outside of the loop. If this is true,
- prepare the statements that load the value of the memory reference to a
- temporary variable in the loop preheader, store it back on the loop exits,
- and replace all the references inside LOOP by this temporary variable.
- LOOP has N_EXITS stored in EXITS. */
+/* Records the information whether REF is independent in LOOP (according
+ to INDEP). */
static void
-determine_lsm_reg (struct loop *loop, edge *exits, unsigned n_exits, tree reg)
+record_indep_loop (struct loop *loop, mem_ref_p ref, bool indep)
{
- tree ref;
- struct mem_ref *mem_refs, *aref;
- struct loop *must_exec;
- bool sees_call;
-
- if (is_gimple_reg (reg))
- return;
-
- ref = single_reachable_address (loop, SSA_NAME_DEF_STMT (reg), &mem_refs,
- &sees_call);
- if (!ref)
- return;
+ if (indep)
+ bitmap_set_bit (ref->indep_loop, loop->num);
+ else
+ bitmap_set_bit (ref->dep_loop, loop->num);
+}
- /* If we cannot create a ssa name for the result, give up. */
- if (!is_gimple_reg_type (TREE_TYPE (ref))
- || TREE_THIS_VOLATILE (ref))
- goto fail;
+/* Returns true if REF is independent on all other memory references in
+ LOOP. */
- /* If there is a call that may use the location, give up as well. */
- if (sees_call
- && is_call_clobbered_ref (ref))
- goto fail;
+static bool
+ref_indep_loop_p_1 (struct loop *loop, mem_ref_p ref)
+{
+ bitmap clobbers, refs_to_check, refs;
+ unsigned i;
+ bitmap_iterator bi;
+ bool ret = true, stored = bitmap_bit_p (ref->stored, loop->num);
+ htab_t map;
+ mem_ref_p aref;
+
+ /* If the reference is clobbered, it is not independent. */
+ clobbers = VEC_index (bitmap, memory_accesses.clobbered_vops, loop->num);
+ if (bitmap_intersect_p (ref->vops, clobbers))
+ return false;
- if (!for_each_index (&ref, may_move_till, loop))
- goto fail;
+ refs_to_check = BITMAP_ALLOC (NULL);
- if (tree_could_trap_p (ref))
+ map = VEC_index (htab_t, memory_accesses.vop_ref_map, loop->num);
+ EXECUTE_IF_AND_COMPL_IN_BITMAP (ref->vops, clobbers, 0, i, bi)
{
- /* If the memory access is unsafe (i.e. it might trap), ensure that some
- of the statements in that it occurs is always executed when the loop
- is entered. This way we know that by moving the load from the
- reference out of the loop we will not cause the error that would not
- occur otherwise.
-
- TODO -- in fact we would like to check for anticipability of the
- reference, i.e. that on each path from loop entry to loop exit at
- least one of the statements containing the memory reference is
- executed. */
-
- for (aref = mem_refs; aref; aref = aref->next)
- {
- if (!LIM_DATA (aref->stmt))
- continue;
+ if (stored)
+ refs = get_vop_accesses (map, i);
+ else
+ refs = get_vop_stores (map, i);
- must_exec = LIM_DATA (aref->stmt)->always_executed_in;
- if (!must_exec)
- continue;
+ bitmap_ior_into (refs_to_check, refs);
+ }
- if (must_exec == loop
- || flow_loop_nested_p (must_exec, loop))
- break;
+ EXECUTE_IF_SET_IN_BITMAP (refs_to_check, 0, i, bi)
+ {
+ aref = VEC_index (mem_ref_p, memory_accesses.refs_list, i);
+ if (!refs_independent_p (ref, aref))
+ {
+ ret = false;
+ record_indep_loop (loop, aref, false);
+ break;
}
-
- if (!aref)
- goto fail;
}
- schedule_sm (loop, exits, n_exits, ref, mem_refs);
+ BITMAP_FREE (refs_to_check);
+ return ret;
+}
+
+/* Returns true if REF is independent on all other memory references in
+ LOOP. Wrapper over ref_indep_loop_p_1, caching its results. */
+
+static bool
+ref_indep_loop_p (struct loop *loop, mem_ref_p ref)
+{
+ bool ret;
+
+ if (bitmap_bit_p (ref->indep_loop, loop->num))
+ return true;
+ if (bitmap_bit_p (ref->dep_loop, loop->num))
+ return false;
+
+ ret = ref_indep_loop_p_1 (loop, ref);
+
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ fprintf (dump_file, "Querying dependencies of ref %u in loop %d: %s\n",
+ ref->id, loop->num, ret ? "independent" : "dependent");
+
+ record_indep_loop (loop, ref, ret);
-fail: ;
- free_mem_refs (mem_refs);
+ return ret;
}
-/* Checks whether LOOP (with N_EXITS exits stored in EXITS array) is suitable
- for a store motion optimization (i.e. whether we can insert statement
- on its exits). */
+/* Returns true if we can perform store motion of REF from LOOP. */
static bool
-loop_suitable_for_sm (struct loop *loop ATTRIBUTE_UNUSED, edge *exits,
- unsigned n_exits)
+can_sm_ref_p (struct loop *loop, mem_ref_p ref)
{
- unsigned i;
+ tree base;
- for (i = 0; i < n_exits; i++)
- if (exits[i]->flags & EDGE_ABNORMAL)
- return false;
+ /* Unless the reference is stored in the loop, there is nothing to do. */
+ if (!bitmap_bit_p (ref->stored, loop->num))
+ return false;
+
+ /* It should be movable. */
+ if (!is_gimple_reg_type (TREE_TYPE (ref->mem))
+ || TREE_THIS_VOLATILE (ref->mem)
+ || !for_each_index (&ref->mem, may_move_till, loop))
+ return false;
+
+ /* If it can trap, it must be always executed in LOOP.
+ Readonly memory locations may trap when storing to them, but
+ tree_could_trap_p is a predicate for rvalues, so check that
+ explicitly. */
+ base = get_base_address (ref->mem);
+ if ((tree_could_trap_p (ref->mem)
+ || (DECL_P (base) && TREE_READONLY (base)))
+ && !ref_always_accessed_p (loop, ref, true))
+ return false;
+
+ /* And it must be independent on all other memory references
+ in LOOP. */
+ if (!ref_indep_loop_p (loop, ref))
+ return false;
return true;
}
-/* Try to perform store motion for all memory references modified inside
- LOOP. */
+/* Marks the references in LOOP for that store motion should be performed
+ in REFS_TO_SM. SM_EXECUTED is the set of references for that store
+ motion was performed in one of the outer loops. */
static void
-determine_lsm_loop (struct loop *loop)
+find_refs_for_sm (struct loop *loop, bitmap sm_executed, bitmap refs_to_sm)
{
- tree phi;
- unsigned n_exits;
- edge *exits = get_loop_exit_edges (loop, &n_exits);
+ bitmap refs = VEC_index (bitmap, memory_accesses.all_refs_in_loop,
+ loop->num);
+ unsigned i;
+ bitmap_iterator bi;
+ mem_ref_p ref;
- if (!loop_suitable_for_sm (loop, exits, n_exits))
+ EXECUTE_IF_AND_COMPL_IN_BITMAP (refs, sm_executed, 0, i, bi)
{
- free (exits);
- return;
+ ref = VEC_index (mem_ref_p, memory_accesses.refs_list, i);
+ if (can_sm_ref_p (loop, ref))
+ bitmap_set_bit (refs_to_sm, i);
}
+}
+
+/* Checks whether LOOP (with exits stored in EXITS array) is suitable
+ for a store motion optimization (i.e. whether we can insert statement
+ on its exits). */
+
+static bool
+loop_suitable_for_sm (struct loop *loop ATTRIBUTE_UNUSED,
+ VEC (edge, heap) *exits)
+{
+ unsigned i;
+ edge ex;
- for (phi = phi_nodes (loop->header); phi; phi = PHI_CHAIN (phi))
- determine_lsm_reg (loop, exits, n_exits, PHI_RESULT (phi));
+ for (i = 0; VEC_iterate (edge, exits, i, ex); i++)
+ if (ex->flags & EDGE_ABNORMAL)
+ return false;
- free (exits);
+ return true;
}
/* Try to perform store motion for all memory references modified inside
- any of LOOPS. */
+ LOOP. SM_EXECUTED is the bitmap of the memory references for that
+ store motion was executed in one of the outer loops. */
static void
-determine_lsm (struct loops *loops)
+store_motion_loop (struct loop *loop, bitmap sm_executed)
{
- struct loop *loop;
- basic_block bb;
-
- if (!loops->tree_root->inner)
- return;
+ VEC (edge, heap) *exits = get_loop_exit_edges (loop);
+ struct loop *subloop;
+ bitmap sm_in_loop = BITMAP_ALLOC (NULL);
- /* Create a UID for each statement in the function. Ordering of the
- UIDs is not important for this pass. */
- max_stmt_uid = 0;
- FOR_EACH_BB (bb)
+ if (loop_suitable_for_sm (loop, exits))
{
- block_stmt_iterator bsi;
-
- for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
- stmt_ann (bsi_stmt (bsi))->uid = max_stmt_uid++;
+ find_refs_for_sm (loop, sm_executed, sm_in_loop);
+ hoist_memory_references (loop, sm_in_loop, exits);
}
+ VEC_free (edge, heap, exits);
- compute_immediate_uses (TDFA_USE_VOPS, NULL);
+ bitmap_ior_into (sm_executed, sm_in_loop);
+ for (subloop = loop->inner; subloop != NULL; subloop = subloop->next)
+ store_motion_loop (subloop, sm_executed);
+ bitmap_and_compl_into (sm_executed, sm_in_loop);
+ BITMAP_FREE (sm_in_loop);
+}
- /* Pass the loops from the outermost. For each virtual operand loop phi node
- check whether all the references inside the loop correspond to a single
- address, and if so, move them. */
+/* Try to perform store motion for all memory references modified inside
+ loops. */
- loop = loops->tree_root->inner;
- while (1)
- {
- determine_lsm_loop (loop);
+static void
+store_motion (void)
+{
+ struct loop *loop;
+ bitmap sm_executed = BITMAP_ALLOC (NULL);
- if (loop->inner)
- {
- loop = loop->inner;
- continue;
- }
- while (!loop->next)
- {
- loop = loop->outer;
- if (loop == loops->tree_root)
- {
- free_df ();
- loop_commit_inserts ();
- return;
- }
- }
- loop = loop->next;
- }
+ for (loop = current_loops->tree_root->inner; loop != NULL; loop = loop->next)
+ store_motion_loop (loop, sm_executed);
+
+ BITMAP_FREE (sm_executed);
+ gsi_commit_edge_inserts ();
}
/* Fills ALWAYS_EXECUTED_IN information for basic blocks of LOOP, i.e.
fill_always_executed_in (loop, contains_call);
}
-/* Compute the global information needed by the loop invariant motion pass.
- LOOPS is the loop tree. */
+/* Compute the global information needed by the loop invariant motion pass. */
static void
-tree_ssa_lim_initialize (struct loops *loops)
+tree_ssa_lim_initialize (void)
{
sbitmap contains_call = sbitmap_alloc (last_basic_block);
- block_stmt_iterator bsi;
+ gimple_stmt_iterator bsi;
struct loop *loop;
basic_block bb;
sbitmap_zero (contains_call);
FOR_EACH_BB (bb)
{
- for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
+ for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
{
- if (nonpure_call_p (bsi_stmt (bsi)))
+ if (nonpure_call_p (gsi_stmt (bsi)))
break;
}
- if (!bsi_end_p (bsi))
+ if (!gsi_end_p (bsi))
SET_BIT (contains_call, bb->index);
}
- for (loop = loops->tree_root->inner; loop; loop = loop->next)
+ for (loop = current_loops->tree_root->inner; loop; loop = loop->next)
fill_always_executed_in (loop, contains_call);
sbitmap_free (contains_call);
+
+ lim_aux_data_map = pointer_map_create ();
}
/* Cleans up after the invariant motion pass. */
tree_ssa_lim_finalize (void)
{
basic_block bb;
+ unsigned i;
+ bitmap b;
+ htab_t h;
FOR_EACH_BB (bb)
{
bb->aux = NULL;
}
+
+ pointer_map_destroy (lim_aux_data_map);
+
+ VEC_free (mem_ref_p, heap, memory_accesses.refs_list);
+ htab_delete (memory_accesses.refs);
+
+ for (i = 0; VEC_iterate (bitmap, memory_accesses.refs_in_loop, i, b); i++)
+ BITMAP_FREE (b);
+ VEC_free (bitmap, heap, memory_accesses.refs_in_loop);
+
+ for (i = 0; VEC_iterate (bitmap, memory_accesses.all_refs_in_loop, i, b); i++)
+ BITMAP_FREE (b);
+ VEC_free (bitmap, heap, memory_accesses.all_refs_in_loop);
+
+ for (i = 0; VEC_iterate (bitmap, memory_accesses.clobbered_vops, i, b); i++)
+ BITMAP_FREE (b);
+ VEC_free (bitmap, heap, memory_accesses.clobbered_vops);
+
+ for (i = 0; VEC_iterate (htab_t, memory_accesses.vop_ref_map, i, h); i++)
+ htab_delete (h);
+ VEC_free (htab_t, heap, memory_accesses.vop_ref_map);
+
+ if (memory_accesses.ttae_cache)
+ pointer_map_destroy (memory_accesses.ttae_cache);
}
-/* Moves invariants from LOOPS. Only "expensive" invariants are moved out --
+/* Moves invariants from loops. Only "expensive" invariants are moved out --
i.e. those that are likely to be win regardless of the register pressure. */
-void
-tree_ssa_lim (struct loops *loops)
+unsigned int
+tree_ssa_lim (void)
{
- tree_ssa_lim_initialize (loops);
+ unsigned int todo;
+
+ tree_ssa_lim_initialize ();
+
+ /* Gathers information about memory accesses in the loops. */
+ analyze_memory_references ();
/* For each statement determine the outermost loop in that it is
invariant and cost for computing the invariant. */
determine_invariantness ();
- /* For each memory reference determine whether it is possible to hoist it
- out of the loop. Force the necessary invariants to be moved out of the
- loops as well. */
- determine_lsm (loops);
+ /* Execute store motion. Force the necessary invariants to be moved
+ out of the loops as well. */
+ store_motion ();
/* Move the expressions that are expensive enough. */
- move_computations ();
+ todo = move_computations ();
tree_ssa_lim_finalize ();
+
+ return todo;
}
OSDN Git Service
