/* Loop invariant motion.
- Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008 Free Software
- Foundation, Inc.
-
+ Copyright (C) 2003-2013 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 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 COPYING3. If not see
<http://www.gnu.org/licenses/>. */
#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 "tree-flow.h"
-#include "tree-dump.h"
-#include "timevar.h"
+#include "gimple-pretty-print.h"
+#include "gimple.h"
+#include "gimplify.h"
+#include "gimple-iterator.h"
+#include "gimplify-me.h"
+#include "gimple-ssa.h"
+#include "tree-cfg.h"
+#include "tree-phinodes.h"
+#include "ssa-iterators.h"
+#include "stringpool.h"
+#include "tree-ssanames.h"
+#include "tree-ssa-loop-manip.h"
+#include "tree-ssa-loop.h"
+#include "tree-into-ssa.h"
#include "cfgloop.h"
#include "domwalk.h"
#include "params.h"
#include "tree-pass.h"
#include "flags.h"
-#include "real.h"
-#include "hashtab.h"
+#include "hash-table.h"
#include "tree-affine.h"
#include "pointer-set.h"
#include "tree-ssa-propagate.h"
+#include "trans-mem.h"
/* TODO: Support for predicated code motion. I.e.
}
}
- Where COND and INV are is invariants, but evaluating INV may trap or be
+ Where COND and INV are invariants, but evaluating INV may trap or be
invalid from some other reason if !COND. This may be transformed to
if (cond)
something;
} */
-/* A type for the list of statements that have to be moved in order to be able
- to hoist an invariant computation. */
-
-struct depend
-{
- gimple stmt;
- struct depend *next;
-};
-
/* The auxiliary data kept for each statement. */
struct lim_aux_data
unsigned cost; /* Cost of the computation performed by the
statement. */
- struct depend *depends; /* List of statements that must be also hoisted
- out of the loop when this statement is
- hoisted; i.e. those that define the operands
- of the statement and are inside of the
- MAX_LOOP loop. */
+ vec<gimple> depends; /* Vector of statements that must be also
+ hoisted out of the loop when this statement
+ is hoisted; i.e. those that define the
+ operands of the statement and are inside of
+ the MAX_LOOP loop. */
};
/* Maps statements to their lim_aux_data. */
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
+
+ /* The memory access itself and associated caching of alias-oracle
+ query meta-data. */
+ ao_ref mem;
+
+ bitmap_head stored; /* The set of loops in that this memory location
is stored to. */
- VEC (mem_ref_locs_p, heap) *accesses_in_loop;
+ vec<vec<mem_ref_loc> > 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
+ bitmap_head 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. */
+ bitmap_head dep_loop; /* The complement of INDEP_LOOP. */
} *mem_ref_p;
-DEF_VEC_P(mem_ref_p);
-DEF_VEC_ALLOC_P(mem_ref_p, heap);
+/* We use two bits per loop in the ref->{in,}dep_loop bitmaps, the first
+ to record (in)dependence against stores in the loop and its subloops, the
+ second to record (in)dependence against all references in the loop
+ and its subloops. */
+#define LOOP_DEP_BIT(loopnum, storedp) (2 * (loopnum) + (storedp ? 1 : 0))
-DEF_VEC_P(bitmap);
-DEF_VEC_ALLOC_P(bitmap, heap);
+/* Mem_ref hashtable helpers. */
+
+struct mem_ref_hasher : typed_noop_remove <mem_ref>
+{
+ typedef mem_ref value_type;
+ typedef tree_node compare_type;
+ static inline hashval_t hash (const value_type *);
+ static inline bool equal (const value_type *, const compare_type *);
+};
+
+/* A hash function for struct mem_ref object OBJ. */
+
+inline hashval_t
+mem_ref_hasher::hash (const value_type *mem)
+{
+ return mem->hash;
+}
+
+/* An equality function for struct mem_ref object MEM1 with
+ memory reference OBJ2. */
+
+inline bool
+mem_ref_hasher::equal (const value_type *mem1, const compare_type *obj2)
+{
+ return operand_equal_p (mem1->mem.ref, (const_tree) obj2, 0);
+}
-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;
+ hash_table <mem_ref_hasher> refs;
/* The list of memory references. */
- VEC (mem_ref_p, heap) *refs_list;
+ vec<mem_ref_p> refs_list;
/* The set of memory references accessed in each loop. */
- VEC (bitmap, heap) *refs_in_loop;
+ vec<bitmap_head> refs_in_loop;
- /* The set of memory references accessed in each loop, including
- subloops. */
- VEC (bitmap, heap) *all_refs_in_loop;
+ /* The set of memory references stored in each loop. */
+ vec<bitmap_head> refs_stored_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;
+ /* The set of memory references stored in each loop, including subloops . */
+ vec<bitmap_head> all_refs_stored_in_loop;
/* Cache for expanding memory addresses. */
struct pointer_map_t *ttae_cache;
} memory_accesses;
+/* Obstack for the bitmaps in the above data structures. */
+static bitmap_obstack lim_bitmap_obstack;
+
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))
-/* The outermost loop for that execution of the header guarantees that the
+/* The outermost loop for which execution of the header guarantees that the
block will be executed. */
#define ALWAYS_EXECUTED_IN(BB) ((struct loop *) (BB)->aux)
+#define SET_ALWAYS_EXECUTED_IN(BB, VAL) ((BB)->aux = (void *) (VAL))
+
+/* ID of the shared unanalyzable mem. */
+#define UNANALYZABLE_MEM_ID 0
+
+/* Whether the reference was analyzable. */
+#define MEM_ANALYZABLE(REF) ((REF)->id != UNANALYZABLE_MEM_ID)
static struct lim_aux_data *
init_lim_data (gimple stmt)
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);
- }
+ data->depends.release ();
free (data);
}
*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. */
-
-bool
-for_each_index (tree *addr_p, bool (*cbck) (tree, tree *, void *), void *data)
-{
- tree *nxt, *idx;
-
- for (; ; addr_p = nxt)
- {
- switch (TREE_CODE (*addr_p))
- {
- case SSA_NAME:
- return cbck (*addr_p, addr_p, data);
-
- case MISALIGNED_INDIRECT_REF:
- case ALIGN_INDIRECT_REF:
- case INDIRECT_REF:
- nxt = &TREE_OPERAND (*addr_p, 0);
- return cbck (*addr_p, nxt, data);
-
- case BIT_FIELD_REF:
- case VIEW_CONVERT_EXPR:
- case REALPART_EXPR:
- case IMAGPART_EXPR:
- nxt = &TREE_OPERAND (*addr_p, 0);
- break;
-
- case COMPONENT_REF:
- /* If the component has varying offset, it behaves like index
- as well. */
- idx = &TREE_OPERAND (*addr_p, 2);
- if (*idx
- && !cbck (*addr_p, idx, data))
- return false;
-
- 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;
- break;
-
- case VAR_DECL:
- 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;
+/* The possibilities of statement movement. */
+enum move_pos
+ {
+ MOVE_IMPOSSIBLE, /* No movement -- side effect expression. */
+ MOVE_PRESERVE_EXECUTION, /* Must not cause the non-executed statement
+ become executed -- memory accesses, ... */
+ MOVE_POSSIBLE /* Unlimited movement. */
+ };
- default:
- gcc_unreachable ();
- }
- }
-}
/* If it is possible to hoist the statement STMT unconditionally,
returns MOVE_POSSIBLE.
return MOVE_POSSIBLE;
}
+ if (gimple_code (stmt) == GIMPLE_PHI
+ && gimple_phi_num_args (stmt) <= 2
+ && !virtual_operand_p (gimple_phi_result (stmt))
+ && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (gimple_phi_result (stmt)))
+ return MOVE_POSSIBLE;
+
if (gimple_get_lhs (stmt) == NULL_TREE)
return MOVE_IMPOSSIBLE;
- if (!ZERO_SSA_OPERANDS (stmt, SSA_OP_VIRTUAL_DEFS))
+ if (gimple_vdef (stmt))
return MOVE_IMPOSSIBLE;
if (stmt_ends_bb_p (stmt)
|| gimple_could_trap_p (stmt))
return MOVE_PRESERVE_EXECUTION;
+ /* Non local loads in a transaction cannot be hoisted out. Well,
+ unless the load happens on every path out of the loop, but we
+ don't take this into account yet. */
+ if (flag_tm
+ && gimple_in_transaction (stmt)
+ && gimple_assign_single_p (stmt))
+ {
+ tree rhs = gimple_assign_rhs1 (stmt);
+ if (DECL_P (rhs) && is_global_var (rhs))
+ {
+ if (dump_file)
+ {
+ fprintf (dump_file, "Cannot hoist conditional load of ");
+ print_generic_expr (dump_file, rhs, TDF_SLIM);
+ fprintf (dump_file, " because it is in a transaction.\n");
+ }
+ return MOVE_IMPOSSIBLE;
+ }
+ }
+
return ret;
}
/* 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
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)
&& def_bb->loop_father == loop)
data->cost += def_data->cost;
- dep = XNEW (struct depend);
- dep->stmt = def_stmt;
- dep->next = data->depends;
- data->depends = dep;
+ data->depends.safe_push (def_stmt);
return true;
}
-/* Returns an estimate for a cost of statement STMT. TODO -- the values here
- are just ad-hoc constants. The estimates should be based on target-specific
- values. */
+/* Returns an estimate for a cost of statement STMT. The values here
+ are just ad-hoc constants, similar to costs for inlining. */
static unsigned
stmt_cost (gimple stmt)
{
- tree fndecl;
- unsigned cost = 1;
-
/* Always try to create possibilities for unswitching. */
- if (gimple_code (stmt) == GIMPLE_COND)
+ if (gimple_code (stmt) == GIMPLE_COND
+ || gimple_code (stmt) == GIMPLE_PHI)
return LIM_EXPENSIVE;
- /* Hoisting memory references out should almost surely be a win. */
- if (gimple_references_memory_p (stmt))
- cost += 20;
-
+ /* We should be hoisting calls if possible. */
if (is_gimple_call (stmt))
{
- /* We should be hoisting calls if possible. */
+ tree fndecl;
/* Unless the call is a builtin_constant_p; this always folds to a
constant, so moving it is useless. */
&& DECL_FUNCTION_CODE (fndecl) == BUILT_IN_CONSTANT_P)
return 0;
- return cost + 20;
+ return LIM_EXPENSIVE;
}
+ /* Hoisting memory references out should almost surely be a win. */
+ if (gimple_references_memory_p (stmt))
+ return LIM_EXPENSIVE;
+
if (gimple_code (stmt) != GIMPLE_ASSIGN)
- return cost;
+ return 1;
switch (gimple_assign_rhs_code (stmt))
{
case MULT_EXPR:
+ case WIDEN_MULT_EXPR:
+ case WIDEN_MULT_PLUS_EXPR:
+ case WIDEN_MULT_MINUS_EXPR:
+ case DOT_PROD_EXPR:
+ case FMA_EXPR:
case TRUNC_DIV_EXPR:
case CEIL_DIV_EXPR:
case FLOOR_DIV_EXPR:
case TRUNC_MOD_EXPR:
case RDIV_EXPR:
/* Division and multiplication are usually expensive. */
- cost += 20;
- break;
+ return LIM_EXPENSIVE;
case LSHIFT_EXPR:
case RSHIFT_EXPR:
- cost += 20;
- break;
+ case WIDEN_LSHIFT_EXPR:
+ case LROTATE_EXPR:
+ case RROTATE_EXPR:
+ /* Shifts and rotates are usually expensive. */
+ return LIM_EXPENSIVE;
+
+ case CONSTRUCTOR:
+ /* Make vector construction cost proportional to the number
+ of elements. */
+ return CONSTRUCTOR_NELTS (gimple_assign_rhs1 (stmt));
+
+ case SSA_NAME:
+ case PAREN_EXPR:
+ /* Whether or not something is wrapped inside a PAREN_EXPR
+ should not change move cost. Nor should an intermediate
+ unpropagated SSA name copy. */
+ return 0;
default:
- break;
+ return 1;
}
-
- return cost;
}
/* Finds the outermost loop between OUTER and LOOP in that the memory reference
{
struct loop *aloop;
- if (bitmap_bit_p (ref->stored, loop->num))
+ 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)
+ if (!bitmap_bit_p (&ref->stored, aloop->num)
&& ref_indep_loop_p (aloop, ref))
return aloop;
static tree *
simple_mem_ref_in_stmt (gimple stmt, bool *is_store)
{
- tree *lhs;
- enum tree_code code;
+ tree *lhs, *rhs;
- /* Recognize MEM = (SSA_NAME | invariant) and SSA_NAME = MEM patterns. */
- if (gimple_code (stmt) != GIMPLE_ASSIGN)
+ /* Recognize SSA_NAME = MEM and MEM = (SSA_NAME | invariant) patterns. */
+ if (!gimple_assign_single_p (stmt))
return NULL;
- code = gimple_assign_rhs_code (stmt);
-
lhs = gimple_assign_lhs_ptr (stmt);
+ rhs = gimple_assign_rhs1_ptr (stmt);
- if (TREE_CODE (*lhs) == SSA_NAME)
+ if (TREE_CODE (*lhs) == SSA_NAME && gimple_vuse (stmt))
{
- 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);
+ return rhs;
}
- else if (code == SSA_NAME
- || (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS
- && is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
+ else if (gimple_vdef (stmt)
+ && (TREE_CODE (*rhs) == SSA_NAME || is_gimple_min_invariant (*rhs)))
{
*is_store = true;
return lhs;
gcc_assert (!store);
hash = iterative_hash_expr (*mem, 0);
- ref = (mem_ref_p) htab_find_with_hash (memory_accesses.refs, *mem, hash);
+ ref = memory_accesses.refs.find_with_hash (*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.
+ We can only use BB dominance checks below if the destination of
+ the true/false edges are dominated by their edge, thus only
+ have a single predecessor. */
+ extract_true_false_edges_from_block (dom, &true_edge, &false_edge);
+ tem = EDGE_PRED (bb, 0);
+ if (tem == true_edge
+ || (single_pred_p (true_edge->dest)
+ && (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
+ || (single_pred_p (false_edge->dest)
+ && (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
+ || (single_pred_p (true_edge->dest)
+ && (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
+ || (single_pred_p (false_edge->dest)
+ && (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. */
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;
+ }
+ }
+
+ 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;
- if (!ZERO_SSA_OPERANDS (stmt, SSA_OP_VIRTUAL_USES))
+ /* 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);
}
else
{
- FOR_EACH_SSA_TREE_OPERAND (val, stmt, iter, SSA_OP_VIRTUAL_USES)
+ if ((val = gimple_vuse (stmt)) != NULL_TREE)
{
if (!add_dependency (val, lim_data, loop, false))
return false;
set_level (gimple stmt, struct loop *orig_loop, struct loop *level)
{
struct loop *stmt_loop = gimple_bb (stmt)->loop_father;
- struct depend *dep;
struct lim_aux_data *lim_data;
+ gimple dep_stmt;
+ unsigned i;
stmt_loop = find_common_loop (orig_loop, stmt_loop);
lim_data = get_lim_data (stmt);
|| flow_loop_nested_p (lim_data->max_loop, level));
lim_data->tgt_loop = level;
- for (dep = lim_data->depends; dep; dep = dep->next)
- set_level (dep->stmt, orig_loop, level);
+ FOR_EACH_VEC_ELT (lim_data->depends, i, dep_stmt)
+ set_level (dep_stmt, orig_loop, level);
}
/* Determines an outermost loop from that we want to hoist the statement STMT.
rewrite_reciprocal (gimple_stmt_iterator *bsi)
{
gimple stmt, stmt1, stmt2;
- tree var, name, lhs, type;
+ tree name, lhs, type;
+ tree real_one;
+ gimple_stmt_iterator gsi;
stmt = gsi_stmt (*bsi);
lhs = gimple_assign_lhs (stmt);
type = TREE_TYPE (lhs);
- var = create_tmp_var (type, "reciptmp");
- add_referenced_var (var);
+ real_one = build_one_cst (type);
- stmt1 = gimple_build_assign_with_ops (RDIV_EXPR,
- var, build_real (type, dconst1), gimple_assign_rhs2 (stmt));
- name = make_ssa_name (var, stmt1);
- gimple_assign_set_lhs (stmt1, name);
+ name = make_temp_ssa_name (type, NULL, "reciptmp");
+ stmt1 = gimple_build_assign_with_ops (RDIV_EXPR, name, real_one,
+ gimple_assign_rhs2 (stmt));
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_replace (bsi, stmt1, true);
- gsi_insert_after (bsi, stmt2, GSI_NEW_STMT);
+ gsi = *bsi;
+ gsi_insert_before (bsi, stmt1, GSI_NEW_STMT);
+ gsi_replace (&gsi, stmt2, true);
/* Continue processing with invariant reciprocal statement. */
return stmt1;
rewrite_bittest (gimple_stmt_iterator *bsi)
{
gimple stmt, use_stmt, stmt1, stmt2;
- tree lhs, var, name, t, a, b;
+ tree lhs, name, t, a, b;
use_operand_p use;
stmt = gsi_stmt (*bsi);
return stmt;
/* There is a conversion in between possibly inserted by fold. */
- if (gimple_assign_rhs_code (stmt1) == NOP_EXPR
- || gimple_assign_rhs_code (stmt1) == CONVERT_EXPR)
+ if (CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt1)))
{
t = gimple_assign_rhs1 (stmt1);
if (TREE_CODE (t) != SSA_NAME
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);
+ name = make_temp_ssa_name (TREE_TYPE (a), NULL, "shifttmp");
+ stmt1 = gimple_build_assign (name, t);
/* 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);
+ name = make_temp_ssa_name (TREE_TYPE (a), NULL, "shifttmp");
+ stmt2 = gimple_build_assign (name, t);
/* 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));
- gsi_insert_before (bsi, stmt1, GSI_SAME_STMT);
- gsi_replace (bsi, stmt2, true);
+ /* 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;
}
+/* For each statement determines the outermost loop in that it is invariant,
+ - statements on whose motion it depends and the cost of the computation.
+ - This information is stored to the LIM_DATA structure associated with
+ - each statement. */
+class invariantness_dom_walker : public dom_walker
+{
+public:
+ invariantness_dom_walker (cdi_direction direction)
+ : dom_walker (direction) {}
+
+ virtual void before_dom_children (basic_block);
+};
/* 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. */
+ Callback for dom_walker. */
-static void
-determine_invariantness_stmt (struct dom_walk_data *dw_data ATTRIBUTE_UNUSED,
- basic_block bb)
+void
+invariantness_dom_walker::before_dom_children (basic_block bb)
{
enum move_pos pos;
gimple_stmt_iterator bsi;
fprintf (dump_file, "Basic block %d (loop %d -- depth %d):\n\n",
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);
+ }
+
+ if (lim_data->cost >= LIM_EXPENSIVE)
+ set_profitable_level (stmt);
+ }
+
for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
{
stmt = gsi_stmt (bsi);
}
}
-/* For each statement determines the outermost loop in that it is invariant,
- statements on whose motion it depends and the cost of the computation.
- This information is stored to the LIM_DATA structure associated with
- each statement. */
+class move_computations_dom_walker : public dom_walker
+{
+public:
+ move_computations_dom_walker (cdi_direction direction)
+ : dom_walker (direction), todo_ (0) {}
-static void
-determine_invariantness (void)
+ virtual void before_dom_children (basic_block);
+
+ unsigned int todo_;
+};
+
+/* Return true if CODE is an operation that when operating on signed
+ integer types involves undefined behavior on overflow and the
+ operation can be expressed with unsigned arithmetic. */
+
+static bool
+arith_code_with_undefined_signed_overflow (tree_code code)
+{
+ switch (code)
+ {
+ case PLUS_EXPR:
+ case MINUS_EXPR:
+ case MULT_EXPR:
+ case NEGATE_EXPR:
+ case POINTER_PLUS_EXPR:
+ return true;
+ default:
+ return false;
+ }
+}
+
+/* Rewrite STMT, an assignment with a signed integer or pointer arithmetic
+ operation that can be transformed to unsigned arithmetic by converting
+ its operand, carrying out the operation in the corresponding unsigned
+ type and converting the result back to the original type.
+
+ Returns a sequence of statements that replace STMT and also contain
+ a modified form of STMT itself. */
+
+static gimple_seq
+rewrite_to_defined_overflow (gimple stmt)
{
- struct dom_walk_data walk_data;
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ {
+ fprintf (dump_file, "rewriting stmt with undefined signed "
+ "overflow ");
+ print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
+ }
- memset (&walk_data, 0, sizeof (struct dom_walk_data));
- walk_data.dom_direction = CDI_DOMINATORS;
- walk_data.before_dom_children_before_stmts = determine_invariantness_stmt;
+ tree lhs = gimple_assign_lhs (stmt);
+ tree type = unsigned_type_for (TREE_TYPE (lhs));
+ gimple_seq stmts = NULL;
+ for (unsigned i = 1; i < gimple_num_ops (stmt); ++i)
+ {
+ gimple_seq stmts2 = NULL;
+ gimple_set_op (stmt, i,
+ force_gimple_operand (fold_convert (type,
+ gimple_op (stmt, i)),
+ &stmts2, true, NULL_TREE));
+ gimple_seq_add_seq (&stmts, stmts2);
+ }
+ gimple_assign_set_lhs (stmt, make_ssa_name (type, stmt));
+ if (gimple_assign_rhs_code (stmt) == POINTER_PLUS_EXPR)
+ gimple_assign_set_rhs_code (stmt, PLUS_EXPR);
+ gimple_seq_add_stmt (&stmts, stmt);
+ gimple cvt = gimple_build_assign_with_ops
+ (NOP_EXPR, lhs, gimple_assign_lhs (stmt), NULL_TREE);
+ gimple_seq_add_stmt (&stmts, cvt);
- init_walk_dominator_tree (&walk_data);
- walk_dominator_tree (&walk_data, ENTRY_BLOCK_PTR);
- fini_walk_dominator_tree (&walk_data);
+ return stmts;
}
/* 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,
- basic_block bb)
+void
+move_computations_dom_walker::before_dom_children (basic_block bb)
{
struct loop *level;
gimple_stmt_iterator bsi;
if (!loop_outer (bb->loop_father))
return;
+ for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); )
+ {
+ gimple new_stmt;
+ 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;
+ }
+
+ 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);
+ }
+ 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));
+ new_stmt = gimple_build_assign_with_ops (COND_EXPR,
+ gimple_phi_result (stmt),
+ t, arg0, arg1);
+ todo_ |= 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); )
{
+ edge e;
+
stmt = gsi_stmt (bsi);
lim_data = get_lim_data (stmt);
cost, level->num);
}
- mark_virtual_ops_for_renaming (stmt);
- gsi_insert_on_edge (loop_preheader_edge (level), stmt);
+ e = loop_preheader_edge (level);
+ gcc_assert (!gimple_vdef (stmt));
+ if (gimple_vuse (stmt))
+ {
+ /* The new VUSE is the one from the virtual PHI in the loop
+ header or the one already present. */
+ gimple_stmt_iterator gsi2;
+ for (gsi2 = gsi_start_phis (e->dest);
+ !gsi_end_p (gsi2); gsi_next (&gsi2))
+ {
+ gimple phi = gsi_stmt (gsi2);
+ if (virtual_operand_p (gimple_phi_result (phi)))
+ {
+ gimple_set_vuse (stmt, PHI_ARG_DEF_FROM_EDGE (phi, e));
+ break;
+ }
+ }
+ }
gsi_remove (&bsi, false);
+ /* In case this is a stmt that is not unconditionally executed
+ when the target loop header is executed and the stmt may
+ invoke undefined integer or pointer overflow rewrite it to
+ unsigned arithmetic. */
+ if (is_gimple_assign (stmt)
+ && INTEGRAL_TYPE_P (TREE_TYPE (gimple_assign_lhs (stmt)))
+ && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (gimple_assign_lhs (stmt)))
+ && arith_code_with_undefined_signed_overflow
+ (gimple_assign_rhs_code (stmt))
+ && (!ALWAYS_EXECUTED_IN (bb)
+ || !(ALWAYS_EXECUTED_IN (bb) == level
+ || flow_loop_nested_p (ALWAYS_EXECUTED_IN (bb), level))))
+ gsi_insert_seq_on_edge (e, rewrite_to_defined_overflow (stmt));
+ else
+ gsi_insert_on_edge (e, stmt);
}
}
/* 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;
-
- memset (&walk_data, 0, sizeof (struct dom_walk_data));
- walk_data.dom_direction = CDI_DOMINATORS;
- walk_data.before_dom_children_before_stmts = move_computations_stmt;
-
- init_walk_dominator_tree (&walk_data);
- walk_dominator_tree (&walk_data, ENTRY_BLOCK_PTR);
- fini_walk_dominator_tree (&walk_data);
+ move_computations_dom_walker walker (CDI_DOMINATORS);
+ walker.walk (cfun->cfg->x_entry_block_ptr);
gsi_commit_edge_inserts ();
- if (need_ssa_update_p ())
+ if (need_ssa_update_p (cfun))
rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa);
+
+ return walker.todo_;
}
/* Checks whether the statement defining variable *INDEX can be hoisted
return;
gcc_assert (TREE_CODE (op) == SSA_NAME);
-
+
stmt = SSA_NAME_DEF_STMT (op);
if (gimple_nop_p (stmt))
return;
return true;
}
-/* A hash function for struct mem_ref object OBJ. */
-
-static hashval_t
-memref_hash (const void *obj)
-{
- const struct mem_ref *const mem = (const struct mem_ref *) obj;
-
- return mem->hash;
-}
-
-/* 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
-free_mem_ref_locs (mem_ref_locs_p accs)
-{
- unsigned i;
- mem_ref_loc_p loc;
-
- if (!accs)
- return;
-
- 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);
-}
-
/* A function to free the mem_ref object OBJ. */
static void
-memref_free (void *obj)
+memref_free (struct mem_ref *mem)
{
- 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);
+ vec<mem_ref_loc> *accs;
- 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);
+ FOR_EACH_VEC_ELT (mem->accesses_in_loop, i, accs)
+ accs->release ();
+ mem->accesses_in_loop.release ();
- BITMAP_FREE (mem->vops);
free (mem);
}
mem_ref_alloc (tree mem, unsigned hash, unsigned id)
{
mem_ref_p ref = XNEW (struct mem_ref);
- ref->mem = mem;
+ ao_ref_init (&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);
+ bitmap_initialize (&ref->stored, &lim_bitmap_obstack);
+ bitmap_initialize (&ref->indep_loop, &lim_bitmap_obstack);
+ bitmap_initialize (&ref->dep_loop, &lim_bitmap_obstack);
+ ref->accesses_in_loop.create (0);
return ref;
}
-/* Allocates and returns the new list of locations. */
-
-static mem_ref_locs_p
-mem_ref_locs_alloc (void)
-{
- mem_ref_locs_p accs = XNEW (struct mem_ref_locs);
- accs->locs = NULL;
- return accs;
-}
-
/* 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);
+ mem_ref_loc aref;
- if (VEC_length (mem_ref_locs_p, ref->accesses_in_loop)
+ if (ref->accesses_in_loop.length ()
<= (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)
- {
- accs = mem_ref_locs_alloc ();
- VEC_replace (mem_ref_locs_p, ref->accesses_in_loop, loop->num, accs);
- }
+ ref->accesses_in_loop.safe_grow_cleared (loop->num + 1);
- aref->stmt = stmt;
- aref->ref = loc;
-
- VEC_safe_push (mem_ref_loc_p, heap, accs->locs, aref);
- bitmap_set_bit (ril, ref->id);
+ aref.stmt = stmt;
+ aref.ref = loc;
+ ref->accesses_in_loop[loop->num].safe_push (aref);
}
/* 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);
+ while (loop != current_loops->tree_root
+ && bitmap_set_bit (&ref->stored, loop->num))
+ loop = loop_outer (loop);
}
/* Gathers memory references in statement STMT in LOOP, storing the
{
tree *mem = NULL;
hashval_t hash;
- PTR *slot;
+ mem_ref **slot;
mem_ref_p ref;
- ssa_op_iter oi;
- tree vname;
bool is_stored;
- bitmap clvops;
unsigned id;
- if (ZERO_SSA_OPERANDS (stmt, SSA_OP_ALL_VIRTUALS))
+ if (!gimple_vuse (stmt))
return;
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;
+ /* We use the shared mem_ref for all unanalyzable refs. */
+ id = UNANALYZABLE_MEM_ID;
+ ref = memory_accesses.refs_list[id];
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ {
+ fprintf (dump_file, "Unanalyzed memory reference %u: ", id);
+ print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
+ }
+ is_stored = gimple_vdef (stmt);
}
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))
+ hash = iterative_hash_expr (*mem, 0);
+ slot = memory_accesses.refs.find_slot_with_hash (*mem, hash, INSERT);
+ if (*slot)
+ {
+ ref = (mem_ref_p) *slot;
+ id = ref->id;
+ }
+ else
{
- fprintf (dump_file, "Memory reference %u: ", id);
- print_generic_expr (dump_file, ref->mem, TDF_SLIM);
- fprintf (dump_file, "\n");
+ id = memory_accesses.refs_list.length ();
+ ref = mem_ref_alloc (*mem, hash, id);
+ memory_accesses.refs_list.safe_push (ref);
+ *slot = ref;
+
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ {
+ fprintf (dump_file, "Memory reference %u: ", id);
+ print_generic_expr (dump_file, ref->mem.ref, TDF_SLIM);
+ fprintf (dump_file, "\n");
+ }
}
+
+ record_mem_ref_loc (ref, loop, stmt, mem);
}
+ bitmap_set_bit (&memory_accesses.refs_in_loop[loop->num], ref->id);
if (is_stored)
- mark_ref_stored (ref, loop);
-
- FOR_EACH_SSA_TREE_OPERAND (vname, stmt, oi, SSA_OP_VIRTUAL_USES)
- bitmap_set_bit (ref->vops, DECL_UID (SSA_NAME_VAR (vname)));
- record_mem_ref_loc (ref, loop, stmt, mem);
+ {
+ bitmap_set_bit (&memory_accesses.refs_stored_in_loop[loop->num], ref->id);
+ mark_ref_stored (ref, loop);
+ }
return;
-
-fail:
- clvops = VEC_index (bitmap, memory_accesses.clobbered_vops, loop->num);
- FOR_EACH_SSA_TREE_OPERAND (vname, stmt, oi, SSA_OP_VIRTUAL_USES)
- bitmap_set_bit (clvops, DECL_UID (SSA_NAME_VAR (vname)));
+}
+
+static unsigned *bb_loop_postorder;
+
+/* qsort sort function to sort blocks after their loop fathers postorder. */
+
+static int
+sort_bbs_in_loop_postorder_cmp (const void *bb1_, const void *bb2_)
+{
+ basic_block bb1 = *(basic_block *)const_cast<void *>(bb1_);
+ basic_block bb2 = *(basic_block *)const_cast<void *>(bb2_);
+ struct loop *loop1 = bb1->loop_father;
+ struct loop *loop2 = bb2->loop_father;
+ if (loop1->num == loop2->num)
+ return 0;
+ return bb_loop_postorder[loop1->num] < bb_loop_postorder[loop2->num] ? -1 : 1;
}
/* Gathers memory references in loops. */
static void
-gather_mem_refs_in_loops (void)
+analyze_memory_references (void)
{
gimple_stmt_iterator bsi;
- basic_block bb;
- struct loop *loop;
- loop_iterator li;
- bitmap clvo, clvi;
- bitmap lrefs, alrefs, alrefso;
-
+ basic_block bb, *bbs;
+ struct loop *loop, *outer;
+ unsigned i, n;
+
+ /* Initialize bb_loop_postorder with a mapping from loop->num to
+ its postorder index. */
+ i = 0;
+ bb_loop_postorder = XNEWVEC (unsigned, number_of_loops (cfun));
+ FOR_EACH_LOOP (loop, LI_FROM_INNERMOST)
+ bb_loop_postorder[loop->num] = i++;
+ /* Collect all basic-blocks in loops and sort them after their
+ loops postorder. */
+ i = 0;
+ bbs = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun) - NUM_FIXED_BLOCKS);
FOR_EACH_BB (bb)
+ if (bb->loop_father != current_loops->tree_root)
+ bbs[i++] = bb;
+ n = i;
+ qsort (bbs, n, sizeof (basic_block), sort_bbs_in_loop_postorder_cmp);
+ free (bb_loop_postorder);
+
+ /* Visit blocks in loop postorder and assign mem-ref IDs in that order.
+ That results in better locality for all the bitmaps. */
+ for (i = 0; i < n; ++i)
{
- loop = bb->loop_father;
- if (loop == current_loops->tree_root)
- continue;
-
+ basic_block bb = bbs[i];
for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
- gather_mem_refs_stmt (loop, gsi_stmt (bsi));
+ gather_mem_refs_stmt (bb->loop_father, 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);
+ free (bbs);
- if (loop_outer (loop) == current_loops->tree_root)
+ /* Propagate the information about accessed memory references up
+ the loop hierarchy. */
+ FOR_EACH_LOOP (loop, LI_FROM_INNERMOST)
+ {
+ /* Finalize the overall touched references (including subloops). */
+ bitmap_ior_into (&memory_accesses.all_refs_stored_in_loop[loop->num],
+ &memory_accesses.refs_stored_in_loop[loop->num]);
+
+ /* Propagate the information about accessed memory references up
+ the loop hierarchy. */
+ outer = loop_outer (loop);
+ if (outer == 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);
+ bitmap_ior_into (&memory_accesses.all_refs_stored_in_loop[outer->num],
+ &memory_accesses.all_refs_stored_in_loop[loop->num]);
}
}
-/* Element of the hash table that maps vops to memory references. */
+/* Returns true if MEM1 and MEM2 may alias. TTAE_CACHE is used as a cache in
+ tree_to_aff_combination_expand. */
-struct vop_to_refs_elt
+static bool
+mem_refs_may_alias_p (mem_ref_p mem1, mem_ref_p mem2,
+ struct pointer_map_t **ttae_cache)
{
- /* DECL_UID of the vop. */
- unsigned uid;
-
- /* List of the all references. */
- bitmap refs_all;
-
- /* List of stored references. */
- bitmap refs_stored;
-};
-
-/* A hash function for struct vop_to_refs_elt object OBJ. */
+ /* 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. */
+ widest_int size1, size2;
+ aff_tree off1, off2;
-static hashval_t
-vtoe_hash (const void *obj)
-{
- const struct vop_to_refs_elt *const vtoe =
- (const struct vop_to_refs_elt *) obj;
+ /* Perform basic offset and type-based disambiguation. */
+ if (!refs_may_alias_p_1 (&mem1->mem, &mem2->mem, true))
+ return false;
- return vtoe->uid;
-}
+ /* 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;
-/* An equality function for struct vop_to_refs_elt object OBJ1 with
- uid of a vop OBJ2. */
+ get_inner_reference_aff (mem1->mem.ref, &off1, &size1);
+ get_inner_reference_aff (mem2->mem.ref, &off2, &size2);
+ aff_combination_expand (&off1, ttae_cache);
+ aff_combination_expand (&off2, ttae_cache);
+ aff_combination_scale (&off1, -1);
+ aff_combination_add (&off2, &off1);
-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;
+ if (aff_comb_cannot_overlap_p (&off2, size1, size2))
+ return false;
- return vtoe->uid == *uid;
+ return true;
}
-/* A function to free the struct vop_to_refs_elt object. */
+/* Iterates over all locations of REF in LOOP and its subloops calling
+ fn.operator() with the location as argument. When that operator
+ returns true the iteration is stopped and true is returned.
+ Otherwise false is returned. */
-static void
-vtoe_free (void *obj)
+template <typename FN>
+static bool
+for_all_locs_in_loop (struct loop *loop, mem_ref_p ref, FN fn)
{
- struct vop_to_refs_elt *const vtoe =
- (struct vop_to_refs_elt *) obj;
-
- BITMAP_FREE (vtoe->refs_all);
- BITMAP_FREE (vtoe->refs_stored);
- free (vtoe);
-}
-
-/* Records REF to hashtable VOP_TO_REFS for the index VOP. STORED is true
- if the reference REF is stored. */
+ unsigned i;
+ mem_ref_loc_p loc;
+ struct loop *subloop;
-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;
+ if (ref->accesses_in_loop.length () > (unsigned) loop->num)
+ FOR_EACH_VEC_ELT (ref->accesses_in_loop[loop->num], i, loc)
+ if (fn (loc))
+ return true;
- 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;
+ for (subloop = loop->inner; subloop != NULL; subloop = subloop->next)
+ if (for_all_locs_in_loop (subloop, ref, fn))
+ return true;
- bitmap_set_bit (vtoe->refs_all, ref);
- if (stored)
- bitmap_set_bit (vtoe->refs_stored, ref);
+ return false;
}
-/* Returns the set of references that access VOP according to the table
- VOP_TO_REFS. */
+/* Rewrites location LOC by TMP_VAR. */
-static bitmap
-get_vop_accesses (htab_t vop_to_refs, unsigned vop)
+struct rewrite_mem_ref_loc
{
- 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. */
+ rewrite_mem_ref_loc (tree tmp_var_) : tmp_var (tmp_var_) {}
+ bool operator () (mem_ref_loc_p loc);
+ tree tmp_var;
+};
-static bitmap
-get_vop_stores (htab_t vop_to_refs, unsigned vop)
+bool
+rewrite_mem_ref_loc::operator () (mem_ref_loc_p loc)
{
- 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;
+ *loc->ref = tmp_var;
+ update_stmt (loc->stmt);
+ return false;
}
-/* Adds REF to mapping from virtual operands to references in LOOP. */
+/* Rewrites all references to REF in LOOP by variable TMP_VAR. */
static void
-add_vop_ref_mapping (struct loop *loop, mem_ref_p ref)
+rewrite_mem_refs (struct loop *loop, mem_ref_p ref, tree tmp_var)
{
- 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)
- {
- record_vop_access (map, vop, ref->id, stored);
- }
+ for_all_locs_in_loop (loop, ref, rewrite_mem_ref_loc (tmp_var));
}
-/* Create a mapping from virtual operands to references that touch them
- in LOOP. */
+/* Stores the first reference location in LOCP. */
-static void
-create_vop_ref_mapping_loop (struct loop *loop)
+struct first_mem_ref_loc_1
{
- 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);
- }
-}
-
-/* For each non-clobbered virtual operand and each loop, record the memory
- references in this loop that touch the operand. */
+ first_mem_ref_loc_1 (mem_ref_loc_p *locp_) : locp (locp_) {}
+ bool operator () (mem_ref_loc_p loc);
+ mem_ref_loc_p *locp;
+};
-static void
-create_vop_ref_mapping (void)
+bool
+first_mem_ref_loc_1::operator () (mem_ref_loc_p loc)
{
- loop_iterator li;
- struct loop *loop;
-
- FOR_EACH_LOOP (li, loop, 0)
- {
- create_vop_ref_mapping_loop (loop);
- }
+ *locp = loc;
+ return true;
}
-/* Gathers information about memory accesses in the loops. */
+/* Returns the first reference location to REF in LOOP. */
-static void
-analyze_memory_references (void)
+static mem_ref_loc_p
+first_mem_ref_loc (struct loop *loop, mem_ref_p ref)
{
- 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 ();
+ mem_ref_loc_p locp = NULL;
+ for_all_locs_in_loop (loop, ref, first_mem_ref_loc_1 (&locp));
+ return locp;
}
-/* Returns true if a region of size SIZE1 at position 0 and a region of
- size SIZE2 at position DIFF cannot overlap. */
-
-static bool
-cannot_overlap_p (aff_tree *diff, double_int size1, double_int size2)
-{
- double_int d, bound;
+struct prev_flag_edges {
+ /* Edge to insert new flag comparison code. */
+ edge append_cond_position;
- /* Unless the difference is a constant, we fail. */
- if (diff->n != 0)
- return false;
+ /* Edge for fall through from previous flag comparison. */
+ edge last_cond_fallthru;
+};
- 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
- {
- /* We succeed if the second object starts after the first one ends. */
- return double_int_scmp (size1, d) <= 0;
- }
-}
+/* Helper function for execute_sm. Emit code to store TMP_VAR into
+ MEM along edge EX.
-/* Returns true if MEM1 and MEM2 may alias. TTAE_CACHE is used as a cache in
- tree_to_aff_combination_expand. */
+ The store is only done if MEM has changed. We do this so no
+ changes to MEM occur on code paths that did not originally store
+ into it.
-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;
+ The common case for execute_sm will transform:
- /* 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;
-
- 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);
+ for (...) {
+ if (foo)
+ stuff;
+ else
+ MEM = TMP_VAR;
+ }
- if (cannot_overlap_p (&off2, size1, size2))
- return false;
+ into:
- return true;
-}
+ lsm = MEM;
+ for (...) {
+ if (foo)
+ stuff;
+ else
+ lsm = TMP_VAR;
+ }
+ MEM = lsm;
-/* Rewrites location LOC by TMP_VAR. */
+ This function will generate:
-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);
-}
+ lsm = MEM;
-/* Adds all locations of REF in LOOP and its subloops to LOCS. */
+ lsm_flag = false;
+ ...
+ for (...) {
+ if (foo)
+ stuff;
+ else {
+ lsm = TMP_VAR;
+ lsm_flag = true;
+ }
+ }
+ if (lsm_flag) <--
+ MEM = lsm; <--
+*/
static void
-get_all_locs_in_loop (struct loop *loop, mem_ref_p ref,
- VEC (mem_ref_loc_p, heap) **locs)
+execute_sm_if_changed (edge ex, tree mem, tree tmp_var, tree flag)
{
- 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;
+ basic_block new_bb, then_bb, old_dest;
+ bool loop_has_only_one_exit;
+ edge then_old_edge, orig_ex = ex;
+ gimple_stmt_iterator gsi;
+ gimple stmt;
+ struct prev_flag_edges *prev_edges = (struct prev_flag_edges *) ex->aux;
- if (!bitmap_bit_p (refs, ref->id))
- return;
+ /* ?? Insert store after previous store if applicable. See note
+ below. */
+ if (prev_edges)
+ ex = prev_edges->append_cond_position;
- 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);
- }
- }
+ loop_has_only_one_exit = single_pred_p (ex->dest);
- for (subloop = loop->inner; subloop != NULL; subloop = subloop->next)
- get_all_locs_in_loop (subloop, ref, locs);
-}
+ if (loop_has_only_one_exit)
+ ex = split_block_after_labels (ex->dest);
-/* Rewrites all references to REF in LOOP by variable TMP_VAR. */
+ old_dest = ex->dest;
+ new_bb = split_edge (ex);
+ then_bb = create_empty_bb (new_bb);
+ if (current_loops && new_bb->loop_father)
+ add_bb_to_loop (then_bb, new_bb->loop_father);
-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;
+ gsi = gsi_start_bb (new_bb);
+ stmt = gimple_build_cond (NE_EXPR, flag, boolean_false_node,
+ NULL_TREE, NULL_TREE);
+ gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING);
- 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);
-}
+ gsi = gsi_start_bb (then_bb);
+ /* Insert actual store. */
+ stmt = gimple_build_assign (unshare_expr (mem), tmp_var);
+ gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING);
-/* 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;
+ make_edge (new_bb, then_bb, EDGE_TRUE_VALUE);
+ make_edge (new_bb, old_dest, EDGE_FALSE_VALUE);
+ then_old_edge = make_edge (then_bb, old_dest, EDGE_FALLTHRU);
-/* Adds S to lsm_tmp_name. */
+ set_immediate_dominator (CDI_DOMINATORS, then_bb, new_bb);
-static void
-lsm_tmp_name_add (const char *s)
-{
- int l = strlen (s) + lsm_tmp_name_length;
- if (l > MAX_LSM_NAME_LENGTH)
- return;
+ if (prev_edges)
+ {
+ basic_block prevbb = prev_edges->last_cond_fallthru->src;
+ redirect_edge_succ (prev_edges->last_cond_fallthru, new_bb);
+ set_immediate_dominator (CDI_DOMINATORS, new_bb, prevbb);
+ set_immediate_dominator (CDI_DOMINATORS, old_dest,
+ recompute_dominator (CDI_DOMINATORS, old_dest));
+ }
- strcpy (lsm_tmp_name + lsm_tmp_name_length, s);
- lsm_tmp_name_length = l;
+ /* ?? Because stores may alias, they must happen in the exact
+ sequence they originally happened. Save the position right after
+ the (_lsm) store we just created so we can continue appending after
+ it and maintain the original order. */
+ {
+ struct prev_flag_edges *p;
+
+ if (orig_ex->aux)
+ orig_ex->aux = NULL;
+ alloc_aux_for_edge (orig_ex, sizeof (struct prev_flag_edges));
+ p = (struct prev_flag_edges *) orig_ex->aux;
+ p->append_cond_position = then_old_edge;
+ p->last_cond_fallthru = find_edge (new_bb, old_dest);
+ orig_ex->aux = (void *) p;
+ }
+
+ if (!loop_has_only_one_exit)
+ for (gsi = gsi_start_phis (old_dest); !gsi_end_p (gsi); gsi_next (&gsi))
+ {
+ gimple phi = gsi_stmt (gsi);
+ unsigned i;
+
+ for (i = 0; i < gimple_phi_num_args (phi); i++)
+ if (gimple_phi_arg_edge (phi, i)->src == new_bb)
+ {
+ tree arg = gimple_phi_arg_def (phi, i);
+ add_phi_arg (phi, arg, then_old_edge, UNKNOWN_LOCATION);
+ update_stmt (phi);
+ }
+ }
+ /* Remove the original fall through edge. This was the
+ single_succ_edge (new_bb). */
+ EDGE_SUCC (new_bb, 0)->flags &= ~EDGE_FALLTHRU;
}
-/* Stores the name for temporary variable that replaces REF to
- lsm_tmp_name. */
+/* When REF is set on the location, set flag indicating the store. */
-static void
-gen_lsm_tmp_name (tree ref)
+struct sm_set_flag_if_changed
{
- const char *name;
+ sm_set_flag_if_changed (tree flag_) : flag (flag_) {}
+ bool operator () (mem_ref_loc_p loc);
+ tree flag;
+};
- switch (TREE_CODE (ref))
+bool
+sm_set_flag_if_changed::operator () (mem_ref_loc_p loc)
+{
+ /* Only set the flag for writes. */
+ if (is_gimple_assign (loc->stmt)
+ && gimple_assign_lhs_ptr (loc->stmt) == loc->ref)
{
- 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;
-
- 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 ("_");
- lsm_tmp_name_add (name);
-
- 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;
-
- default:
- gcc_unreachable ();
+ gimple_stmt_iterator gsi = gsi_for_stmt (loc->stmt);
+ gimple stmt = gimple_build_assign (flag, boolean_true_node);
+ gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING);
}
+ return false;
}
-/* 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. */
+/* Helper function for execute_sm. On every location where REF is
+ set, set an appropriate flag indicating the store. */
-char *
-get_lsm_tmp_name (tree ref, unsigned n)
+static tree
+execute_sm_if_changed_flag_set (struct loop *loop, mem_ref_p ref)
{
- 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;
+ tree flag;
+ char *str = get_lsm_tmp_name (ref->mem.ref, ~0, "_flag");
+ flag = create_tmp_reg (boolean_type_node, str);
+ for_all_locs_in_loop (loop, ref, sm_set_flag_if_changed (flag));
+ return flag;
}
/* Executes store motion of memory reference REF from LOOP.
to the reference from the temporary variable are emitted to exits. */
static void
-execute_sm (struct loop *loop, VEC (edge, heap) *exits, mem_ref_p ref)
+execute_sm (struct loop *loop, vec<edge> exits, mem_ref_p ref)
{
- tree tmp_var;
+ tree tmp_var, store_flag;
unsigned i;
- gimple load, store;
+ gimple load;
struct fmt_data fmt_data;
edge ex;
struct lim_aux_data *lim_data;
+ bool multi_threaded_model_p = false;
+ gimple_stmt_iterator gsi;
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, "Executing store motion of ");
- print_generic_expr (dump_file, ref->mem, 0);
+ print_generic_expr (dump_file, ref->mem.ref, 0);
fprintf (dump_file, " from loop %d\n", loop->num);
}
- tmp_var = make_rename_temp (TREE_TYPE (ref->mem),
- get_lsm_tmp_name (ref->mem, ~0));
+ tmp_var = create_tmp_reg (TREE_TYPE (ref->mem.ref),
+ get_lsm_tmp_name (ref->mem.ref, ~0));
fmt_data.loop = loop;
fmt_data.orig_loop = loop;
- for_each_index (&ref->mem, force_move_till, &fmt_data);
+ for_each_index (&ref->mem.ref, force_move_till, &fmt_data);
+
+ if (bb_in_transaction (loop_preheader_edge (loop)->src)
+ || !PARAM_VALUE (PARAM_ALLOW_STORE_DATA_RACES))
+ multi_threaded_model_p = true;
+
+ if (multi_threaded_model_p)
+ store_flag = execute_sm_if_changed_flag_set (loop, ref);
rewrite_mem_refs (loop, ref, tmp_var);
- /* Emit the load & stores. */
- load = gimple_build_assign (tmp_var, unshare_expr (ref->mem));
+ /* Emit the load code on a random exit edge or into the latch if
+ the loop does not exit, so that we are sure it will be processed
+ by move_computations after all dependencies. */
+ gsi = gsi_for_stmt (first_mem_ref_loc (loop, ref)->stmt);
+
+ /* FIXME/TODO: For the multi-threaded variant, we could avoid this
+ load altogether, since the store is predicated by a flag. We
+ could, do the load only if it was originally in the loop. */
+ load = gimple_build_assign (tmp_var, unshare_expr (ref->mem.ref));
lim_data = init_lim_data (load);
lim_data->max_loop = loop;
lim_data->tgt_loop = loop;
+ gsi_insert_before (&gsi, load, GSI_SAME_STMT);
- /* Put this into the latch, so that we are sure it will be processed after
- all dependencies. */
- gsi_insert_on_edge (loop_latch_edge (loop), load);
-
- for (i = 0; VEC_iterate (edge, exits, i, ex); i++)
+ if (multi_threaded_model_p)
{
- store = gimple_build_assign (unshare_expr (ref->mem), tmp_var);
- gsi_insert_on_edge (ex, store);
+ load = gimple_build_assign (store_flag, boolean_false_node);
+ lim_data = init_lim_data (load);
+ lim_data->max_loop = loop;
+ lim_data->tgt_loop = loop;
+ gsi_insert_before (&gsi, load, GSI_SAME_STMT);
}
+
+ /* Sink the store to every exit from the loop. */
+ FOR_EACH_VEC_ELT (exits, i, ex)
+ if (!multi_threaded_model_p)
+ {
+ gimple store;
+ store = gimple_build_assign (unshare_expr (ref->mem.ref), tmp_var);
+ gsi_insert_on_edge (ex, store);
+ }
+ else
+ execute_sm_if_changed (ex, ref->mem.ref, tmp_var, store_flag);
}
/* Hoists memory references MEM_REFS out of LOOP. EXITS is the list of exit
static void
hoist_memory_references (struct loop *loop, bitmap mem_refs,
- VEC (edge, heap) *exits)
+ vec<edge> exits)
{
mem_ref_p ref;
unsigned i;
EXECUTE_IF_SET_IN_BITMAP (mem_refs, 0, i, bi)
{
- ref = VEC_index (mem_ref_p, memory_accesses.refs_list, i);
+ ref = memory_accesses.refs_list[i];
execute_sm (loop, exits, ref);
}
}
-/* Returns true if REF is always accessed in LOOP. */
+struct ref_always_accessed
+{
+ ref_always_accessed (struct loop *loop_, tree base_, bool stored_p_)
+ : loop (loop_), base (base_), stored_p (stored_p_) {}
+ bool operator () (mem_ref_loc_p loc);
+ struct loop *loop;
+ tree base;
+ bool stored_p;
+};
-static bool
-ref_always_accessed_p (struct loop *loop, mem_ref_p ref)
+bool
+ref_always_accessed::operator () (mem_ref_loc_p loc)
{
- VEC (mem_ref_loc_p, heap) *locs = NULL;
- unsigned i;
- mem_ref_loc_p loc;
- bool ret = false;
struct loop *must_exec;
- get_all_locs_in_loop (loop, ref, &locs);
- for (i = 0; VEC_iterate (mem_ref_loc_p, locs, i, loc); i++)
+ if (!get_lim_data (loc->stmt))
+ return false;
+
+ /* If we require an always executed store make sure the statement
+ stores to the reference. */
+ if (stored_p)
{
- if (!get_lim_data (loc->stmt))
- continue;
+ tree lhs;
+ if (!gimple_get_lhs (loc->stmt))
+ return false;
+ lhs = get_base_address (gimple_get_lhs (loc->stmt));
+ if (!lhs)
+ return false;
+ if (INDIRECT_REF_P (lhs)
+ || TREE_CODE (lhs) == MEM_REF)
+ lhs = TREE_OPERAND (lhs, 0);
+ if (lhs != base)
+ return false;
+ }
- must_exec = get_lim_data (loc->stmt)->always_executed_in;
- if (!must_exec)
- continue;
+ must_exec = get_lim_data (loc->stmt)->always_executed_in;
+ if (!must_exec)
+ return false;
- if (must_exec == loop
- || flow_loop_nested_p (must_exec, loop))
- {
- ret = true;
- break;
- }
- }
- VEC_free (mem_ref_loc_p, heap, locs);
+ if (must_exec == loop
+ || flow_loop_nested_p (must_exec, loop))
+ return true;
- return ret;
+ return false;
+}
+
+/* Returns true if REF is always accessed in LOOP. If STORED_P is true
+ make sure REF is always stored to in LOOP. */
+
+static bool
+ref_always_accessed_p (struct loop *loop, mem_ref_p ref, bool stored_p)
+{
+ tree base = ao_ref_base (&ref->mem);
+ if (TREE_CODE (base) == MEM_REF)
+ base = TREE_OPERAND (base, 0);
+
+ return for_all_locs_in_loop (loop, ref,
+ ref_always_accessed (loop, base, stored_p));
}
/* 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))
+ if (ref1 == ref2)
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);
- if (mem_refs_may_alias_p (ref1->mem, ref2->mem,
- &memory_accesses.ttae_cache))
+ if (mem_refs_may_alias_p (ref1, ref2, &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;
}
}
-/* Records the information whether REF is independent in LOOP (according
- to INDEP). */
+/* Mark REF dependent on stores or loads (according to STORED_P) in LOOP
+ and its super-loops. */
static void
-record_indep_loop (struct loop *loop, mem_ref_p ref, bool indep)
+record_dep_loop (struct loop *loop, mem_ref_p ref, bool stored_p)
{
- if (indep)
- bitmap_set_bit (ref->indep_loop, loop->num);
- else
- bitmap_set_bit (ref->dep_loop, loop->num);
+ /* We can propagate dependent-in-loop bits up the loop
+ hierarchy to all outer loops. */
+ while (loop != current_loops->tree_root
+ && bitmap_set_bit (&ref->dep_loop, LOOP_DEP_BIT (loop->num, stored_p)))
+ loop = loop_outer (loop);
}
/* Returns true if REF is independent on all other memory references in
LOOP. */
static bool
-ref_indep_loop_p_1 (struct loop *loop, mem_ref_p ref)
+ref_indep_loop_p_1 (struct loop *loop, mem_ref_p ref, bool stored_p)
{
- bitmap clobbers, refs_to_check, refs;
+ bitmap refs_to_check;
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;
-
- refs_to_check = BITMAP_ALLOC (NULL);
-
- 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 (stored)
- refs = get_vop_accesses (map, i);
- else
- refs = get_vop_stores (map, i);
+ if (stored_p)
+ refs_to_check = &memory_accesses.refs_in_loop[loop->num];
+ else
+ refs_to_check = &memory_accesses.refs_stored_in_loop[loop->num];
- bitmap_ior_into (refs_to_check, refs);
- }
+ if (bitmap_bit_p (refs_to_check, UNANALYZABLE_MEM_ID))
+ return false;
EXECUTE_IF_SET_IN_BITMAP (refs_to_check, 0, i, bi)
{
- aref = VEC_index (mem_ref_p, memory_accesses.refs_list, i);
+ aref = memory_accesses.refs_list[i];
if (!refs_independent_p (ref, aref))
- {
- ret = false;
- record_indep_loop (loop, aref, false);
- break;
- }
+ return false;
}
- BITMAP_FREE (refs_to_check);
- return ret;
+ return true;
}
/* 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)
+ref_indep_loop_p_2 (struct loop *loop, mem_ref_p ref, bool stored_p)
{
- bool ret;
+ stored_p |= bitmap_bit_p (&ref->stored, loop->num);
- if (bitmap_bit_p (ref->indep_loop, loop->num))
+ if (bitmap_bit_p (&ref->indep_loop, LOOP_DEP_BIT (loop->num, stored_p)))
return true;
- if (bitmap_bit_p (ref->dep_loop, loop->num))
+ if (bitmap_bit_p (&ref->dep_loop, LOOP_DEP_BIT (loop->num, stored_p)))
return false;
- ret = ref_indep_loop_p_1 (loop, ref);
+ struct loop *inner = loop->inner;
+ while (inner)
+ {
+ if (!ref_indep_loop_p_2 (inner, ref, stored_p))
+ return false;
+ inner = inner->next;
+ }
+
+ bool indep_p = ref_indep_loop_p_1 (loop, ref, stored_p);
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");
+ ref->id, loop->num, indep_p ? "independent" : "dependent");
+
+ /* Record the computed result in the cache. */
+ if (indep_p)
+ {
+ if (bitmap_set_bit (&ref->indep_loop, LOOP_DEP_BIT (loop->num, stored_p))
+ && stored_p)
+ {
+ /* If it's independend against all refs then it's independent
+ against stores, too. */
+ bitmap_set_bit (&ref->indep_loop, LOOP_DEP_BIT (loop->num, false));
+ }
+ }
+ else
+ {
+ record_dep_loop (loop, ref, stored_p);
+ if (!stored_p)
+ {
+ /* If it's dependent against stores it's dependent against
+ all refs, too. */
+ record_dep_loop (loop, ref, true);
+ }
+ }
- record_indep_loop (loop, ref, ret);
+ return indep_p;
+}
- return ret;
+/* Returns true if REF is independent on all other memory references in
+ LOOP. */
+
+static bool
+ref_indep_loop_p (struct loop *loop, mem_ref_p ref)
+{
+ gcc_checking_assert (MEM_ANALYZABLE (ref));
+
+ return ref_indep_loop_p_2 (loop, ref, false);
}
/* Returns true if we can perform store motion of REF from LOOP. */
static bool
can_sm_ref_p (struct loop *loop, mem_ref_p ref)
{
- /* Unless the reference is stored in the loop, there is nothing to do. */
- if (!bitmap_bit_p (ref->stored, loop->num))
+ tree base;
+
+ /* Can't hoist unanalyzable refs. */
+ if (!MEM_ANALYZABLE (ref))
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))
+ if (!is_gimple_reg_type (TREE_TYPE (ref->mem.ref))
+ || TREE_THIS_VOLATILE (ref->mem.ref)
+ || !for_each_index (&ref->mem.ref, may_move_till, loop))
return false;
- /* If it can trap, it must be always executed in LOOP. */
- if (tree_could_trap_p (ref->mem)
- && !ref_always_accessed_p (loop, ref))
+ /* If it can throw fail, we do not properly update EH info. */
+ if (tree_could_throw_p (ref->mem.ref))
+ 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.ref);
+ if ((tree_could_trap_p (ref->mem.ref)
+ || (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
static void
find_refs_for_sm (struct loop *loop, bitmap sm_executed, bitmap refs_to_sm)
{
- bitmap refs = VEC_index (bitmap, memory_accesses.all_refs_in_loop,
- loop->num);
+ bitmap refs = &memory_accesses.all_refs_stored_in_loop[loop->num];
unsigned i;
bitmap_iterator bi;
mem_ref_p ref;
EXECUTE_IF_AND_COMPL_IN_BITMAP (refs, sm_executed, 0, i, bi)
{
- ref = VEC_index (mem_ref_p, memory_accesses.refs_list, i);
+ ref = memory_accesses.refs_list[i];
if (can_sm_ref_p (loop, ref))
bitmap_set_bit (refs_to_sm, i);
}
static bool
loop_suitable_for_sm (struct loop *loop ATTRIBUTE_UNUSED,
- VEC (edge, heap) *exits)
+ vec<edge> exits)
{
unsigned i;
edge ex;
- for (i = 0; VEC_iterate (edge, exits, i, ex); i++)
- if (ex->flags & EDGE_ABNORMAL)
+ FOR_EACH_VEC_ELT (exits, i, ex)
+ if (ex->flags & (EDGE_ABNORMAL | EDGE_EH))
return false;
return true;
static void
store_motion_loop (struct loop *loop, bitmap sm_executed)
{
- VEC (edge, heap) *exits = get_loop_exit_edges (loop);
+ vec<edge> exits = get_loop_exit_edges (loop);
struct loop *subloop;
- bitmap sm_in_loop = BITMAP_ALLOC (NULL);
+ bitmap sm_in_loop = BITMAP_ALLOC (&lim_bitmap_obstack);
if (loop_suitable_for_sm (loop, exits))
{
find_refs_for_sm (loop, sm_executed, sm_in_loop);
hoist_memory_references (loop, sm_in_loop, exits);
}
- VEC_free (edge, heap, exits);
+ exits.release ();
bitmap_ior_into (sm_executed, sm_in_loop);
for (subloop = loop->inner; subloop != NULL; subloop = subloop->next)
store_motion (void)
{
struct loop *loop;
- bitmap sm_executed = BITMAP_ALLOC (NULL);
+ bitmap sm_executed = BITMAP_ALLOC (&lim_bitmap_obstack);
for (loop = current_loops->tree_root->inner; loop != NULL; loop = loop->next)
store_motion_loop (loop, sm_executed);
blocks that contain a nonpure call. */
static void
-fill_always_executed_in (struct loop *loop, sbitmap contains_call)
+fill_always_executed_in_1 (struct loop *loop, sbitmap contains_call)
{
basic_block bb = NULL, *bbs, last = NULL;
unsigned i;
edge e;
struct loop *inn_loop = loop;
- if (!loop->header->aux)
+ if (ALWAYS_EXECUTED_IN (loop->header) == NULL)
{
bbs = get_loop_body_in_dom_order (loop);
if (dominated_by_p (CDI_DOMINATORS, loop->latch, bb))
last = bb;
- if (TEST_BIT (contains_call, bb->index))
+ if (bitmap_bit_p (contains_call, bb->index))
break;
FOR_EACH_EDGE (e, ei, bb->succs)
while (1)
{
- last->aux = loop;
+ SET_ALWAYS_EXECUTED_IN (last, loop);
if (last == loop->header)
break;
last = get_immediate_dominator (CDI_DOMINATORS, last);
}
for (loop = loop->inner; loop; loop = loop->next)
- fill_always_executed_in (loop, contains_call);
+ fill_always_executed_in_1 (loop, contains_call);
}
-/* Compute the global information needed by the loop invariant motion pass. */
+/* Fills ALWAYS_EXECUTED_IN information for basic blocks, i.e.
+ for each such basic block bb records the outermost loop for that execution
+ of its header implies execution of bb. */
static void
-tree_ssa_lim_initialize (void)
+fill_always_executed_in (void)
{
sbitmap contains_call = sbitmap_alloc (last_basic_block);
- gimple_stmt_iterator bsi;
- struct loop *loop;
basic_block bb;
+ struct loop *loop;
- sbitmap_zero (contains_call);
+ bitmap_clear (contains_call);
FOR_EACH_BB (bb)
{
- for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
+ gimple_stmt_iterator gsi;
+ for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
{
- if (nonpure_call_p (gsi_stmt (bsi)))
+ if (nonpure_call_p (gsi_stmt (gsi)))
break;
}
- if (!gsi_end_p (bsi))
- SET_BIT (contains_call, bb->index);
+ if (!gsi_end_p (gsi))
+ bitmap_set_bit (contains_call, bb->index);
}
for (loop = current_loops->tree_root->inner; loop; loop = loop->next)
- fill_always_executed_in (loop, contains_call);
+ fill_always_executed_in_1 (loop, contains_call);
sbitmap_free (contains_call);
+}
+
+
+/* Compute the global information needed by the loop invariant motion pass. */
+static void
+tree_ssa_lim_initialize (void)
+{
+ unsigned i;
+
+ bitmap_obstack_initialize (&lim_bitmap_obstack);
lim_aux_data_map = pointer_map_create ();
+
+ if (flag_tm)
+ compute_transaction_bits ();
+
+ alloc_aux_for_edges (0);
+
+ memory_accesses.refs.create (100);
+ memory_accesses.refs_list.create (100);
+ /* Allocate a special, unanalyzable mem-ref with ID zero. */
+ memory_accesses.refs_list.quick_push
+ (mem_ref_alloc (error_mark_node, 0, UNANALYZABLE_MEM_ID));
+
+ memory_accesses.refs_in_loop.create (number_of_loops (cfun));
+ memory_accesses.refs_in_loop.quick_grow (number_of_loops (cfun));
+ memory_accesses.refs_stored_in_loop.create (number_of_loops (cfun));
+ memory_accesses.refs_stored_in_loop.quick_grow (number_of_loops (cfun));
+ memory_accesses.all_refs_stored_in_loop.create (number_of_loops (cfun));
+ memory_accesses.all_refs_stored_in_loop.quick_grow (number_of_loops (cfun));
+
+ for (i = 0; i < number_of_loops (cfun); i++)
+ {
+ bitmap_initialize (&memory_accesses.refs_in_loop[i],
+ &lim_bitmap_obstack);
+ bitmap_initialize (&memory_accesses.refs_stored_in_loop[i],
+ &lim_bitmap_obstack);
+ bitmap_initialize (&memory_accesses.all_refs_stored_in_loop[i],
+ &lim_bitmap_obstack);
+ }
+
+ memory_accesses.ttae_cache = NULL;
}
/* Cleans up after the invariant motion pass. */
{
basic_block bb;
unsigned i;
- bitmap b;
- htab_t h;
+ mem_ref_p ref;
+
+ free_aux_for_edges ();
FOR_EACH_BB (bb)
- {
- bb->aux = NULL;
- }
+ SET_ALWAYS_EXECUTED_IN (bb, NULL);
+ bitmap_obstack_release (&lim_bitmap_obstack);
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);
+ memory_accesses.refs.dispose ();
- 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_EACH_VEC_ELT (memory_accesses.refs_list, i, ref)
+ memref_free (ref);
+ memory_accesses.refs_list.release ();
- 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);
+ memory_accesses.refs_in_loop.release ();
+ memory_accesses.refs_stored_in_loop.release ();
+ memory_accesses.all_refs_stored_in_loop.release ();
if (memory_accesses.ttae_cache)
- pointer_map_destroy (memory_accesses.ttae_cache);
+ free_affine_expand_cache (&memory_accesses.ttae_cache);
}
/* 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
+unsigned int
tree_ssa_lim (void)
{
+ unsigned int todo;
+
tree_ssa_lim_initialize ();
/* Gathers information about memory accesses in the loops. */
analyze_memory_references ();
+ /* Fills ALWAYS_EXECUTED_IN information for basic blocks. */
+ fill_always_executed_in ();
+
/* For each statement determine the outermost loop in that it is
invariant and cost for computing the invariant. */
- determine_invariantness ();
+ invariantness_dom_walker (CDI_DOMINATORS)
+ .walk (cfun->cfg->x_entry_block_ptr);
/* 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;
+}
+
+/* Loop invariant motion pass. */
+
+static unsigned int
+tree_ssa_loop_im (void)
+{
+ if (number_of_loops (cfun) <= 1)
+ return 0;
+
+ return tree_ssa_lim ();
+}
+
+static bool
+gate_tree_ssa_loop_im (void)
+{
+ return flag_tree_loop_im != 0;
}
+
+namespace {
+
+const pass_data pass_data_lim =
+{
+ GIMPLE_PASS, /* type */
+ "lim", /* name */
+ OPTGROUP_LOOP, /* optinfo_flags */
+ true, /* has_gate */
+ true, /* has_execute */
+ TV_LIM, /* tv_id */
+ PROP_cfg, /* properties_required */
+ 0, /* properties_provided */
+ 0, /* properties_destroyed */
+ 0, /* todo_flags_start */
+ 0, /* todo_flags_finish */
+};
+
+class pass_lim : public gimple_opt_pass
+{
+public:
+ pass_lim (gcc::context *ctxt)
+ : gimple_opt_pass (pass_data_lim, ctxt)
+ {}
+
+ /* opt_pass methods: */
+ opt_pass * clone () { return new pass_lim (m_ctxt); }
+ bool gate () { return gate_tree_ssa_loop_im (); }
+ unsigned int execute () { return tree_ssa_loop_im (); }
+
+}; // class pass_lim
+
+} // anon namespace
+
+gimple_opt_pass *
+make_pass_lim (gcc::context *ctxt)
+{
+ return new pass_lim (ctxt);
+}
+
+
projects / thirdparty / gcc.git / blobdiff