1 /* Loop invariant motion.
2 Copyright (C) 2003-2013 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it
7 under the terms of the GNU General Public License as published by the
8 Free Software Foundation; either version 3, or (at your option) any
11 GCC is distributed in the hope that it will be useful, but WITHOUT
12 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
22 #include "coretypes.h"
26 #include "basic-block.h"
27 #include "gimple-pretty-print.h"
28 #include "pointer-set.h"
29 #include "hash-table.h"
30 #include "tree-ssa-alias.h"
31 #include "internal-fn.h"
33 #include "gimple-expr.h"
37 #include "gimple-iterator.h"
38 #include "gimplify-me.h"
39 #include "gimple-ssa.h"
41 #include "tree-phinodes.h"
42 #include "ssa-iterators.h"
43 #include "stringpool.h"
44 #include "tree-ssanames.h"
45 #include "tree-ssa-loop-manip.h"
46 #include "tree-ssa-loop.h"
47 #include "tree-into-ssa.h"
51 #include "tree-pass.h"
53 #include "tree-affine.h"
54 #include "tree-ssa-propagate.h"
55 #include "trans-mem.h"
57 /* TODO: Support for predicated code motion. I.e.
68 Where COND and INV are invariants, but evaluating INV may trap or be
69 invalid from some other reason if !COND. This may be transformed to
79 /* The auxiliary data kept for each statement. */
83 struct loop
*max_loop
; /* The outermost loop in that the statement
86 struct loop
*tgt_loop
; /* The loop out of that we want to move the
89 struct loop
*always_executed_in
;
90 /* The outermost loop for that we are sure
91 the statement is executed if the loop
94 unsigned cost
; /* Cost of the computation performed by the
97 vec
<gimple
> depends
; /* Vector of statements that must be also
98 hoisted out of the loop when this statement
99 is hoisted; i.e. those that define the
100 operands of the statement and are inside of
101 the MAX_LOOP loop. */
104 /* Maps statements to their lim_aux_data. */
106 static struct pointer_map_t
*lim_aux_data_map
;
108 /* Description of a memory reference location. */
110 typedef struct mem_ref_loc
112 tree
*ref
; /* The reference itself. */
113 gimple stmt
; /* The statement in that it occurs. */
117 /* Description of a memory reference. */
119 typedef struct mem_ref
121 unsigned id
; /* ID assigned to the memory reference
122 (its index in memory_accesses.refs_list) */
123 hashval_t hash
; /* Its hash value. */
125 /* The memory access itself and associated caching of alias-oracle
129 bitmap_head stored
; /* The set of loops in that this memory location
131 vec
<vec
<mem_ref_loc
> > accesses_in_loop
;
132 /* The locations of the accesses. Vector
133 indexed by the loop number. */
135 /* The following sets are computed on demand. We keep both set and
136 its complement, so that we know whether the information was
137 already computed or not. */
138 bitmap_head indep_loop
; /* The set of loops in that the memory
139 reference is independent, meaning:
140 If it is stored in the loop, this store
141 is independent on all other loads and
143 If it is only loaded, then it is independent
144 on all stores in the loop. */
145 bitmap_head dep_loop
; /* The complement of INDEP_LOOP. */
148 /* We use two bits per loop in the ref->{in,}dep_loop bitmaps, the first
149 to record (in)dependence against stores in the loop and its subloops, the
150 second to record (in)dependence against all references in the loop
152 #define LOOP_DEP_BIT(loopnum, storedp) (2 * (loopnum) + (storedp ? 1 : 0))
154 /* Mem_ref hashtable helpers. */
156 struct mem_ref_hasher
: typed_noop_remove
<mem_ref
>
158 typedef mem_ref value_type
;
159 typedef tree_node compare_type
;
160 static inline hashval_t
hash (const value_type
*);
161 static inline bool equal (const value_type
*, const compare_type
*);
164 /* A hash function for struct mem_ref object OBJ. */
167 mem_ref_hasher::hash (const value_type
*mem
)
172 /* An equality function for struct mem_ref object MEM1 with
173 memory reference OBJ2. */
176 mem_ref_hasher::equal (const value_type
*mem1
, const compare_type
*obj2
)
178 return operand_equal_p (mem1
->mem
.ref
, (const_tree
) obj2
, 0);
182 /* Description of memory accesses in loops. */
186 /* The hash table of memory references accessed in loops. */
187 hash_table
<mem_ref_hasher
> refs
;
189 /* The list of memory references. */
190 vec
<mem_ref_p
> refs_list
;
192 /* The set of memory references accessed in each loop. */
193 vec
<bitmap_head
> refs_in_loop
;
195 /* The set of memory references stored in each loop. */
196 vec
<bitmap_head
> refs_stored_in_loop
;
198 /* The set of memory references stored in each loop, including subloops . */
199 vec
<bitmap_head
> all_refs_stored_in_loop
;
201 /* Cache for expanding memory addresses. */
202 struct pointer_map_t
*ttae_cache
;
205 /* Obstack for the bitmaps in the above data structures. */
206 static bitmap_obstack lim_bitmap_obstack
;
208 static bool ref_indep_loop_p (struct loop
*, mem_ref_p
);
210 /* Minimum cost of an expensive expression. */
211 #define LIM_EXPENSIVE ((unsigned) PARAM_VALUE (PARAM_LIM_EXPENSIVE))
213 /* The outermost loop for which execution of the header guarantees that the
214 block will be executed. */
215 #define ALWAYS_EXECUTED_IN(BB) ((struct loop *) (BB)->aux)
216 #define SET_ALWAYS_EXECUTED_IN(BB, VAL) ((BB)->aux = (void *) (VAL))
218 /* ID of the shared unanalyzable mem. */
219 #define UNANALYZABLE_MEM_ID 0
221 /* Whether the reference was analyzable. */
222 #define MEM_ANALYZABLE(REF) ((REF)->id != UNANALYZABLE_MEM_ID)
224 static struct lim_aux_data
*
225 init_lim_data (gimple stmt
)
227 void **p
= pointer_map_insert (lim_aux_data_map
, stmt
);
229 *p
= XCNEW (struct lim_aux_data
);
230 return (struct lim_aux_data
*) *p
;
233 static struct lim_aux_data
*
234 get_lim_data (gimple stmt
)
236 void **p
= pointer_map_contains (lim_aux_data_map
, stmt
);
240 return (struct lim_aux_data
*) *p
;
243 /* Releases the memory occupied by DATA. */
246 free_lim_aux_data (struct lim_aux_data
*data
)
248 data
->depends
.release ();
253 clear_lim_data (gimple stmt
)
255 void **p
= pointer_map_contains (lim_aux_data_map
, stmt
);
259 free_lim_aux_data ((struct lim_aux_data
*) *p
);
264 /* The possibilities of statement movement. */
267 MOVE_IMPOSSIBLE
, /* No movement -- side effect expression. */
268 MOVE_PRESERVE_EXECUTION
, /* Must not cause the non-executed statement
269 become executed -- memory accesses, ... */
270 MOVE_POSSIBLE
/* Unlimited movement. */
274 /* If it is possible to hoist the statement STMT unconditionally,
275 returns MOVE_POSSIBLE.
276 If it is possible to hoist the statement STMT, but we must avoid making
277 it executed if it would not be executed in the original program (e.g.
278 because it may trap), return MOVE_PRESERVE_EXECUTION.
279 Otherwise return MOVE_IMPOSSIBLE. */
282 movement_possibility (gimple stmt
)
285 enum move_pos ret
= MOVE_POSSIBLE
;
287 if (flag_unswitch_loops
288 && gimple_code (stmt
) == GIMPLE_COND
)
290 /* If we perform unswitching, force the operands of the invariant
291 condition to be moved out of the loop. */
292 return MOVE_POSSIBLE
;
295 if (gimple_code (stmt
) == GIMPLE_PHI
296 && gimple_phi_num_args (stmt
) <= 2
297 && !virtual_operand_p (gimple_phi_result (stmt
))
298 && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (gimple_phi_result (stmt
)))
299 return MOVE_POSSIBLE
;
301 if (gimple_get_lhs (stmt
) == NULL_TREE
)
302 return MOVE_IMPOSSIBLE
;
304 if (gimple_vdef (stmt
))
305 return MOVE_IMPOSSIBLE
;
307 if (stmt_ends_bb_p (stmt
)
308 || gimple_has_volatile_ops (stmt
)
309 || gimple_has_side_effects (stmt
)
310 || stmt_could_throw_p (stmt
))
311 return MOVE_IMPOSSIBLE
;
313 if (is_gimple_call (stmt
))
315 /* While pure or const call is guaranteed to have no side effects, we
316 cannot move it arbitrarily. Consider code like
318 char *s = something ();
328 Here the strlen call cannot be moved out of the loop, even though
329 s is invariant. In addition to possibly creating a call with
330 invalid arguments, moving out a function call that is not executed
331 may cause performance regressions in case the call is costly and
332 not executed at all. */
333 ret
= MOVE_PRESERVE_EXECUTION
;
334 lhs
= gimple_call_lhs (stmt
);
336 else if (is_gimple_assign (stmt
))
337 lhs
= gimple_assign_lhs (stmt
);
339 return MOVE_IMPOSSIBLE
;
341 if (TREE_CODE (lhs
) == SSA_NAME
342 && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs
))
343 return MOVE_IMPOSSIBLE
;
345 if (TREE_CODE (lhs
) != SSA_NAME
346 || gimple_could_trap_p (stmt
))
347 return MOVE_PRESERVE_EXECUTION
;
349 /* Non local loads in a transaction cannot be hoisted out. Well,
350 unless the load happens on every path out of the loop, but we
351 don't take this into account yet. */
353 && gimple_in_transaction (stmt
)
354 && gimple_assign_single_p (stmt
))
356 tree rhs
= gimple_assign_rhs1 (stmt
);
357 if (DECL_P (rhs
) && is_global_var (rhs
))
361 fprintf (dump_file
, "Cannot hoist conditional load of ");
362 print_generic_expr (dump_file
, rhs
, TDF_SLIM
);
363 fprintf (dump_file
, " because it is in a transaction.\n");
365 return MOVE_IMPOSSIBLE
;
372 /* Suppose that operand DEF is used inside the LOOP. Returns the outermost
373 loop to that we could move the expression using DEF if it did not have
374 other operands, i.e. the outermost loop enclosing LOOP in that the value
375 of DEF is invariant. */
378 outermost_invariant_loop (tree def
, struct loop
*loop
)
382 struct loop
*max_loop
;
383 struct lim_aux_data
*lim_data
;
386 return superloop_at_depth (loop
, 1);
388 if (TREE_CODE (def
) != SSA_NAME
)
390 gcc_assert (is_gimple_min_invariant (def
));
391 return superloop_at_depth (loop
, 1);
394 def_stmt
= SSA_NAME_DEF_STMT (def
);
395 def_bb
= gimple_bb (def_stmt
);
397 return superloop_at_depth (loop
, 1);
399 max_loop
= find_common_loop (loop
, def_bb
->loop_father
);
401 lim_data
= get_lim_data (def_stmt
);
402 if (lim_data
!= NULL
&& lim_data
->max_loop
!= NULL
)
403 max_loop
= find_common_loop (max_loop
,
404 loop_outer (lim_data
->max_loop
));
405 if (max_loop
== loop
)
407 max_loop
= superloop_at_depth (loop
, loop_depth (max_loop
) + 1);
412 /* DATA is a structure containing information associated with a statement
413 inside LOOP. DEF is one of the operands of this statement.
415 Find the outermost loop enclosing LOOP in that value of DEF is invariant
416 and record this in DATA->max_loop field. If DEF itself is defined inside
417 this loop as well (i.e. we need to hoist it out of the loop if we want
418 to hoist the statement represented by DATA), record the statement in that
419 DEF is defined to the DATA->depends list. Additionally if ADD_COST is true,
420 add the cost of the computation of DEF to the DATA->cost.
422 If DEF is not invariant in LOOP, return false. Otherwise return TRUE. */
425 add_dependency (tree def
, struct lim_aux_data
*data
, struct loop
*loop
,
428 gimple def_stmt
= SSA_NAME_DEF_STMT (def
);
429 basic_block def_bb
= gimple_bb (def_stmt
);
430 struct loop
*max_loop
;
431 struct lim_aux_data
*def_data
;
436 max_loop
= outermost_invariant_loop (def
, loop
);
440 if (flow_loop_nested_p (data
->max_loop
, max_loop
))
441 data
->max_loop
= max_loop
;
443 def_data
= get_lim_data (def_stmt
);
448 /* Only add the cost if the statement defining DEF is inside LOOP,
449 i.e. if it is likely that by moving the invariants dependent
450 on it, we will be able to avoid creating a new register for
451 it (since it will be only used in these dependent invariants). */
452 && def_bb
->loop_father
== loop
)
453 data
->cost
+= def_data
->cost
;
455 data
->depends
.safe_push (def_stmt
);
460 /* Returns an estimate for a cost of statement STMT. The values here
461 are just ad-hoc constants, similar to costs for inlining. */
464 stmt_cost (gimple stmt
)
466 /* Always try to create possibilities for unswitching. */
467 if (gimple_code (stmt
) == GIMPLE_COND
468 || gimple_code (stmt
) == GIMPLE_PHI
)
469 return LIM_EXPENSIVE
;
471 /* We should be hoisting calls if possible. */
472 if (is_gimple_call (stmt
))
476 /* Unless the call is a builtin_constant_p; this always folds to a
477 constant, so moving it is useless. */
478 fndecl
= gimple_call_fndecl (stmt
);
480 && DECL_BUILT_IN_CLASS (fndecl
) == BUILT_IN_NORMAL
481 && DECL_FUNCTION_CODE (fndecl
) == BUILT_IN_CONSTANT_P
)
484 return LIM_EXPENSIVE
;
487 /* Hoisting memory references out should almost surely be a win. */
488 if (gimple_references_memory_p (stmt
))
489 return LIM_EXPENSIVE
;
491 if (gimple_code (stmt
) != GIMPLE_ASSIGN
)
494 switch (gimple_assign_rhs_code (stmt
))
497 case WIDEN_MULT_EXPR
:
498 case WIDEN_MULT_PLUS_EXPR
:
499 case WIDEN_MULT_MINUS_EXPR
:
512 /* Division and multiplication are usually expensive. */
513 return LIM_EXPENSIVE
;
517 case WIDEN_LSHIFT_EXPR
:
520 /* Shifts and rotates are usually expensive. */
521 return LIM_EXPENSIVE
;
524 /* Make vector construction cost proportional to the number
526 return CONSTRUCTOR_NELTS (gimple_assign_rhs1 (stmt
));
530 /* Whether or not something is wrapped inside a PAREN_EXPR
531 should not change move cost. Nor should an intermediate
532 unpropagated SSA name copy. */
540 /* Finds the outermost loop between OUTER and LOOP in that the memory reference
541 REF is independent. If REF is not independent in LOOP, NULL is returned
545 outermost_indep_loop (struct loop
*outer
, struct loop
*loop
, mem_ref_p ref
)
549 if (bitmap_bit_p (&ref
->stored
, loop
->num
))
554 aloop
= superloop_at_depth (loop
, loop_depth (aloop
) + 1))
555 if (!bitmap_bit_p (&ref
->stored
, aloop
->num
)
556 && ref_indep_loop_p (aloop
, ref
))
559 if (ref_indep_loop_p (loop
, ref
))
565 /* If there is a simple load or store to a memory reference in STMT, returns
566 the location of the memory reference, and sets IS_STORE according to whether
567 it is a store or load. Otherwise, returns NULL. */
570 simple_mem_ref_in_stmt (gimple stmt
, bool *is_store
)
574 /* Recognize SSA_NAME = MEM and MEM = (SSA_NAME | invariant) patterns. */
575 if (!gimple_assign_single_p (stmt
))
578 lhs
= gimple_assign_lhs_ptr (stmt
);
579 rhs
= gimple_assign_rhs1_ptr (stmt
);
581 if (TREE_CODE (*lhs
) == SSA_NAME
&& gimple_vuse (stmt
))
586 else if (gimple_vdef (stmt
)
587 && (TREE_CODE (*rhs
) == SSA_NAME
|| is_gimple_min_invariant (*rhs
)))
596 /* Returns the memory reference contained in STMT. */
599 mem_ref_in_stmt (gimple stmt
)
602 tree
*mem
= simple_mem_ref_in_stmt (stmt
, &store
);
610 hash
= iterative_hash_expr (*mem
, 0);
611 ref
= memory_accesses
.refs
.find_with_hash (*mem
, hash
);
613 gcc_assert (ref
!= NULL
);
617 /* From a controlling predicate in DOM determine the arguments from
618 the PHI node PHI that are chosen if the predicate evaluates to
619 true and false and store them to *TRUE_ARG_P and *FALSE_ARG_P if
620 they are non-NULL. Returns true if the arguments can be determined,
621 else return false. */
624 extract_true_false_args_from_phi (basic_block dom
, gimple phi
,
625 tree
*true_arg_p
, tree
*false_arg_p
)
627 basic_block bb
= gimple_bb (phi
);
628 edge true_edge
, false_edge
, tem
;
629 tree arg0
= NULL_TREE
, arg1
= NULL_TREE
;
631 /* We have to verify that one edge into the PHI node is dominated
632 by the true edge of the predicate block and the other edge
633 dominated by the false edge. This ensures that the PHI argument
634 we are going to take is completely determined by the path we
635 take from the predicate block.
636 We can only use BB dominance checks below if the destination of
637 the true/false edges are dominated by their edge, thus only
638 have a single predecessor. */
639 extract_true_false_edges_from_block (dom
, &true_edge
, &false_edge
);
640 tem
= EDGE_PRED (bb
, 0);
642 || (single_pred_p (true_edge
->dest
)
643 && (tem
->src
== true_edge
->dest
644 || dominated_by_p (CDI_DOMINATORS
,
645 tem
->src
, true_edge
->dest
))))
646 arg0
= PHI_ARG_DEF (phi
, tem
->dest_idx
);
647 else if (tem
== false_edge
648 || (single_pred_p (false_edge
->dest
)
649 && (tem
->src
== false_edge
->dest
650 || dominated_by_p (CDI_DOMINATORS
,
651 tem
->src
, false_edge
->dest
))))
652 arg1
= PHI_ARG_DEF (phi
, tem
->dest_idx
);
655 tem
= EDGE_PRED (bb
, 1);
657 || (single_pred_p (true_edge
->dest
)
658 && (tem
->src
== true_edge
->dest
659 || dominated_by_p (CDI_DOMINATORS
,
660 tem
->src
, true_edge
->dest
))))
661 arg0
= PHI_ARG_DEF (phi
, tem
->dest_idx
);
662 else if (tem
== false_edge
663 || (single_pred_p (false_edge
->dest
)
664 && (tem
->src
== false_edge
->dest
665 || dominated_by_p (CDI_DOMINATORS
,
666 tem
->src
, false_edge
->dest
))))
667 arg1
= PHI_ARG_DEF (phi
, tem
->dest_idx
);
681 /* Determine the outermost loop to that it is possible to hoist a statement
682 STMT and store it to LIM_DATA (STMT)->max_loop. To do this we determine
683 the outermost loop in that the value computed by STMT is invariant.
684 If MUST_PRESERVE_EXEC is true, additionally choose such a loop that
685 we preserve the fact whether STMT is executed. It also fills other related
686 information to LIM_DATA (STMT).
688 The function returns false if STMT cannot be hoisted outside of the loop it
689 is defined in, and true otherwise. */
692 determine_max_movement (gimple stmt
, bool must_preserve_exec
)
694 basic_block bb
= gimple_bb (stmt
);
695 struct loop
*loop
= bb
->loop_father
;
697 struct lim_aux_data
*lim_data
= get_lim_data (stmt
);
701 if (must_preserve_exec
)
702 level
= ALWAYS_EXECUTED_IN (bb
);
704 level
= superloop_at_depth (loop
, 1);
705 lim_data
->max_loop
= level
;
707 if (gimple_code (stmt
) == GIMPLE_PHI
)
710 unsigned min_cost
= UINT_MAX
;
711 unsigned total_cost
= 0;
712 struct lim_aux_data
*def_data
;
714 /* We will end up promoting dependencies to be unconditionally
715 evaluated. For this reason the PHI cost (and thus the
716 cost we remove from the loop by doing the invariant motion)
717 is that of the cheapest PHI argument dependency chain. */
718 FOR_EACH_PHI_ARG (use_p
, stmt
, iter
, SSA_OP_USE
)
720 val
= USE_FROM_PTR (use_p
);
721 if (TREE_CODE (val
) != SSA_NAME
)
723 if (!add_dependency (val
, lim_data
, loop
, false))
725 def_data
= get_lim_data (SSA_NAME_DEF_STMT (val
));
728 min_cost
= MIN (min_cost
, def_data
->cost
);
729 total_cost
+= def_data
->cost
;
733 lim_data
->cost
+= min_cost
;
735 if (gimple_phi_num_args (stmt
) > 1)
737 basic_block dom
= get_immediate_dominator (CDI_DOMINATORS
, bb
);
739 if (gsi_end_p (gsi_last_bb (dom
)))
741 cond
= gsi_stmt (gsi_last_bb (dom
));
742 if (gimple_code (cond
) != GIMPLE_COND
)
744 /* Verify that this is an extended form of a diamond and
745 the PHI arguments are completely controlled by the
747 if (!extract_true_false_args_from_phi (dom
, stmt
, NULL
, NULL
))
750 /* Fold in dependencies and cost of the condition. */
751 FOR_EACH_SSA_TREE_OPERAND (val
, cond
, iter
, SSA_OP_USE
)
753 if (!add_dependency (val
, lim_data
, loop
, false))
755 def_data
= get_lim_data (SSA_NAME_DEF_STMT (val
));
757 total_cost
+= def_data
->cost
;
760 /* We want to avoid unconditionally executing very expensive
761 operations. As costs for our dependencies cannot be
762 negative just claim we are not invariand for this case.
763 We also are not sure whether the control-flow inside the
765 if (total_cost
- min_cost
>= 2 * LIM_EXPENSIVE
767 && total_cost
/ min_cost
<= 2))
770 /* Assume that the control-flow in the loop will vanish.
771 ??? We should verify this and not artificially increase
772 the cost if that is not the case. */
773 lim_data
->cost
+= stmt_cost (stmt
);
779 FOR_EACH_SSA_TREE_OPERAND (val
, stmt
, iter
, SSA_OP_USE
)
780 if (!add_dependency (val
, lim_data
, loop
, true))
783 if (gimple_vuse (stmt
))
785 mem_ref_p ref
= mem_ref_in_stmt (stmt
);
790 = outermost_indep_loop (lim_data
->max_loop
, loop
, ref
);
791 if (!lim_data
->max_loop
)
796 if ((val
= gimple_vuse (stmt
)) != NULL_TREE
)
798 if (!add_dependency (val
, lim_data
, loop
, false))
804 lim_data
->cost
+= stmt_cost (stmt
);
809 /* Suppose that some statement in ORIG_LOOP is hoisted to the loop LEVEL,
810 and that one of the operands of this statement is computed by STMT.
811 Ensure that STMT (together with all the statements that define its
812 operands) is hoisted at least out of the loop LEVEL. */
815 set_level (gimple stmt
, struct loop
*orig_loop
, struct loop
*level
)
817 struct loop
*stmt_loop
= gimple_bb (stmt
)->loop_father
;
818 struct lim_aux_data
*lim_data
;
822 stmt_loop
= find_common_loop (orig_loop
, stmt_loop
);
823 lim_data
= get_lim_data (stmt
);
824 if (lim_data
!= NULL
&& lim_data
->tgt_loop
!= NULL
)
825 stmt_loop
= find_common_loop (stmt_loop
,
826 loop_outer (lim_data
->tgt_loop
));
827 if (flow_loop_nested_p (stmt_loop
, level
))
830 gcc_assert (level
== lim_data
->max_loop
831 || flow_loop_nested_p (lim_data
->max_loop
, level
));
833 lim_data
->tgt_loop
= level
;
834 FOR_EACH_VEC_ELT (lim_data
->depends
, i
, dep_stmt
)
835 set_level (dep_stmt
, orig_loop
, level
);
838 /* Determines an outermost loop from that we want to hoist the statement STMT.
839 For now we chose the outermost possible loop. TODO -- use profiling
840 information to set it more sanely. */
843 set_profitable_level (gimple stmt
)
845 set_level (stmt
, gimple_bb (stmt
)->loop_father
, get_lim_data (stmt
)->max_loop
);
848 /* Returns true if STMT is a call that has side effects. */
851 nonpure_call_p (gimple stmt
)
853 if (gimple_code (stmt
) != GIMPLE_CALL
)
856 return gimple_has_side_effects (stmt
);
859 /* Rewrite a/b to a*(1/b). Return the invariant stmt to process. */
862 rewrite_reciprocal (gimple_stmt_iterator
*bsi
)
864 gimple stmt
, stmt1
, stmt2
;
865 tree name
, lhs
, type
;
867 gimple_stmt_iterator gsi
;
869 stmt
= gsi_stmt (*bsi
);
870 lhs
= gimple_assign_lhs (stmt
);
871 type
= TREE_TYPE (lhs
);
873 real_one
= build_one_cst (type
);
875 name
= make_temp_ssa_name (type
, NULL
, "reciptmp");
876 stmt1
= gimple_build_assign_with_ops (RDIV_EXPR
, name
, real_one
,
877 gimple_assign_rhs2 (stmt
));
879 stmt2
= gimple_build_assign_with_ops (MULT_EXPR
, lhs
, name
,
880 gimple_assign_rhs1 (stmt
));
882 /* Replace division stmt with reciprocal and multiply stmts.
883 The multiply stmt is not invariant, so update iterator
884 and avoid rescanning. */
886 gsi_insert_before (bsi
, stmt1
, GSI_NEW_STMT
);
887 gsi_replace (&gsi
, stmt2
, true);
889 /* Continue processing with invariant reciprocal statement. */
893 /* Check if the pattern at *BSI is a bittest of the form
894 (A >> B) & 1 != 0 and in this case rewrite it to A & (1 << B) != 0. */
897 rewrite_bittest (gimple_stmt_iterator
*bsi
)
899 gimple stmt
, use_stmt
, stmt1
, stmt2
;
900 tree lhs
, name
, t
, a
, b
;
903 stmt
= gsi_stmt (*bsi
);
904 lhs
= gimple_assign_lhs (stmt
);
906 /* Verify that the single use of lhs is a comparison against zero. */
907 if (TREE_CODE (lhs
) != SSA_NAME
908 || !single_imm_use (lhs
, &use
, &use_stmt
)
909 || gimple_code (use_stmt
) != GIMPLE_COND
)
911 if (gimple_cond_lhs (use_stmt
) != lhs
912 || (gimple_cond_code (use_stmt
) != NE_EXPR
913 && gimple_cond_code (use_stmt
) != EQ_EXPR
)
914 || !integer_zerop (gimple_cond_rhs (use_stmt
)))
917 /* Get at the operands of the shift. The rhs is TMP1 & 1. */
918 stmt1
= SSA_NAME_DEF_STMT (gimple_assign_rhs1 (stmt
));
919 if (gimple_code (stmt1
) != GIMPLE_ASSIGN
)
922 /* There is a conversion in between possibly inserted by fold. */
923 if (CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt1
)))
925 t
= gimple_assign_rhs1 (stmt1
);
926 if (TREE_CODE (t
) != SSA_NAME
927 || !has_single_use (t
))
929 stmt1
= SSA_NAME_DEF_STMT (t
);
930 if (gimple_code (stmt1
) != GIMPLE_ASSIGN
)
934 /* Verify that B is loop invariant but A is not. Verify that with
935 all the stmt walking we are still in the same loop. */
936 if (gimple_assign_rhs_code (stmt1
) != RSHIFT_EXPR
937 || loop_containing_stmt (stmt1
) != loop_containing_stmt (stmt
))
940 a
= gimple_assign_rhs1 (stmt1
);
941 b
= gimple_assign_rhs2 (stmt1
);
943 if (outermost_invariant_loop (b
, loop_containing_stmt (stmt1
)) != NULL
944 && outermost_invariant_loop (a
, loop_containing_stmt (stmt1
)) == NULL
)
946 gimple_stmt_iterator rsi
;
949 t
= fold_build2 (LSHIFT_EXPR
, TREE_TYPE (a
),
950 build_int_cst (TREE_TYPE (a
), 1), b
);
951 name
= make_temp_ssa_name (TREE_TYPE (a
), NULL
, "shifttmp");
952 stmt1
= gimple_build_assign (name
, t
);
955 t
= fold_build2 (BIT_AND_EXPR
, TREE_TYPE (a
), a
, name
);
956 name
= make_temp_ssa_name (TREE_TYPE (a
), NULL
, "shifttmp");
957 stmt2
= gimple_build_assign (name
, t
);
959 /* Replace the SSA_NAME we compare against zero. Adjust
960 the type of zero accordingly. */
962 gimple_cond_set_rhs (use_stmt
, build_int_cst_type (TREE_TYPE (name
), 0));
964 /* Don't use gsi_replace here, none of the new assignments sets
965 the variable originally set in stmt. Move bsi to stmt1, and
966 then remove the original stmt, so that we get a chance to
967 retain debug info for it. */
969 gsi_insert_before (bsi
, stmt1
, GSI_NEW_STMT
);
970 gsi_insert_before (&rsi
, stmt2
, GSI_SAME_STMT
);
971 gsi_remove (&rsi
, true);
979 /* For each statement determines the outermost loop in that it is invariant,
980 - statements on whose motion it depends and the cost of the computation.
981 - This information is stored to the LIM_DATA structure associated with
983 class invariantness_dom_walker
: public dom_walker
986 invariantness_dom_walker (cdi_direction direction
)
987 : dom_walker (direction
) {}
989 virtual void before_dom_children (basic_block
);
992 /* Determine the outermost loops in that statements in basic block BB are
993 invariant, and record them to the LIM_DATA associated with the statements.
994 Callback for dom_walker. */
997 invariantness_dom_walker::before_dom_children (basic_block bb
)
1000 gimple_stmt_iterator bsi
;
1002 bool maybe_never
= ALWAYS_EXECUTED_IN (bb
) == NULL
;
1003 struct loop
*outermost
= ALWAYS_EXECUTED_IN (bb
);
1004 struct lim_aux_data
*lim_data
;
1006 if (!loop_outer (bb
->loop_father
))
1009 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1010 fprintf (dump_file
, "Basic block %d (loop %d -- depth %d):\n\n",
1011 bb
->index
, bb
->loop_father
->num
, loop_depth (bb
->loop_father
));
1013 /* Look at PHI nodes, but only if there is at most two.
1014 ??? We could relax this further by post-processing the inserted
1015 code and transforming adjacent cond-exprs with the same predicate
1016 to control flow again. */
1017 bsi
= gsi_start_phis (bb
);
1018 if (!gsi_end_p (bsi
)
1019 && ((gsi_next (&bsi
), gsi_end_p (bsi
))
1020 || (gsi_next (&bsi
), gsi_end_p (bsi
))))
1021 for (bsi
= gsi_start_phis (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
1023 stmt
= gsi_stmt (bsi
);
1025 pos
= movement_possibility (stmt
);
1026 if (pos
== MOVE_IMPOSSIBLE
)
1029 lim_data
= init_lim_data (stmt
);
1030 lim_data
->always_executed_in
= outermost
;
1032 if (!determine_max_movement (stmt
, false))
1034 lim_data
->max_loop
= NULL
;
1038 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1040 print_gimple_stmt (dump_file
, stmt
, 2, 0);
1041 fprintf (dump_file
, " invariant up to level %d, cost %d.\n\n",
1042 loop_depth (lim_data
->max_loop
),
1046 if (lim_data
->cost
>= LIM_EXPENSIVE
)
1047 set_profitable_level (stmt
);
1050 for (bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
1052 stmt
= gsi_stmt (bsi
);
1054 pos
= movement_possibility (stmt
);
1055 if (pos
== MOVE_IMPOSSIBLE
)
1057 if (nonpure_call_p (stmt
))
1062 /* Make sure to note always_executed_in for stores to make
1063 store-motion work. */
1064 else if (stmt_makes_single_store (stmt
))
1066 struct lim_aux_data
*lim_data
= init_lim_data (stmt
);
1067 lim_data
->always_executed_in
= outermost
;
1072 if (is_gimple_assign (stmt
)
1073 && (get_gimple_rhs_class (gimple_assign_rhs_code (stmt
))
1074 == GIMPLE_BINARY_RHS
))
1076 tree op0
= gimple_assign_rhs1 (stmt
);
1077 tree op1
= gimple_assign_rhs2 (stmt
);
1078 struct loop
*ol1
= outermost_invariant_loop (op1
,
1079 loop_containing_stmt (stmt
));
1081 /* If divisor is invariant, convert a/b to a*(1/b), allowing reciprocal
1082 to be hoisted out of loop, saving expensive divide. */
1083 if (pos
== MOVE_POSSIBLE
1084 && gimple_assign_rhs_code (stmt
) == RDIV_EXPR
1085 && flag_unsafe_math_optimizations
1086 && !flag_trapping_math
1088 && outermost_invariant_loop (op0
, ol1
) == NULL
)
1089 stmt
= rewrite_reciprocal (&bsi
);
1091 /* If the shift count is invariant, convert (A >> B) & 1 to
1092 A & (1 << B) allowing the bit mask to be hoisted out of the loop
1093 saving an expensive shift. */
1094 if (pos
== MOVE_POSSIBLE
1095 && gimple_assign_rhs_code (stmt
) == BIT_AND_EXPR
1096 && integer_onep (op1
)
1097 && TREE_CODE (op0
) == SSA_NAME
1098 && has_single_use (op0
))
1099 stmt
= rewrite_bittest (&bsi
);
1102 lim_data
= init_lim_data (stmt
);
1103 lim_data
->always_executed_in
= outermost
;
1105 if (maybe_never
&& pos
== MOVE_PRESERVE_EXECUTION
)
1108 if (!determine_max_movement (stmt
, pos
== MOVE_PRESERVE_EXECUTION
))
1110 lim_data
->max_loop
= NULL
;
1114 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1116 print_gimple_stmt (dump_file
, stmt
, 2, 0);
1117 fprintf (dump_file
, " invariant up to level %d, cost %d.\n\n",
1118 loop_depth (lim_data
->max_loop
),
1122 if (lim_data
->cost
>= LIM_EXPENSIVE
)
1123 set_profitable_level (stmt
);
1127 class move_computations_dom_walker
: public dom_walker
1130 move_computations_dom_walker (cdi_direction direction
)
1131 : dom_walker (direction
), todo_ (0) {}
1133 virtual void before_dom_children (basic_block
);
1138 /* Return true if CODE is an operation that when operating on signed
1139 integer types involves undefined behavior on overflow and the
1140 operation can be expressed with unsigned arithmetic. */
1143 arith_code_with_undefined_signed_overflow (tree_code code
)
1151 case POINTER_PLUS_EXPR
:
1158 /* Rewrite STMT, an assignment with a signed integer or pointer arithmetic
1159 operation that can be transformed to unsigned arithmetic by converting
1160 its operand, carrying out the operation in the corresponding unsigned
1161 type and converting the result back to the original type.
1163 Returns a sequence of statements that replace STMT and also contain
1164 a modified form of STMT itself. */
1167 rewrite_to_defined_overflow (gimple stmt
)
1169 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1171 fprintf (dump_file
, "rewriting stmt with undefined signed "
1173 print_gimple_stmt (dump_file
, stmt
, 0, TDF_SLIM
);
1176 tree lhs
= gimple_assign_lhs (stmt
);
1177 tree type
= unsigned_type_for (TREE_TYPE (lhs
));
1178 gimple_seq stmts
= NULL
;
1179 for (unsigned i
= 1; i
< gimple_num_ops (stmt
); ++i
)
1181 gimple_seq stmts2
= NULL
;
1182 gimple_set_op (stmt
, i
,
1183 force_gimple_operand (fold_convert (type
,
1184 gimple_op (stmt
, i
)),
1185 &stmts2
, true, NULL_TREE
));
1186 gimple_seq_add_seq (&stmts
, stmts2
);
1188 gimple_assign_set_lhs (stmt
, make_ssa_name (type
, stmt
));
1189 if (gimple_assign_rhs_code (stmt
) == POINTER_PLUS_EXPR
)
1190 gimple_assign_set_rhs_code (stmt
, PLUS_EXPR
);
1191 gimple_seq_add_stmt (&stmts
, stmt
);
1192 gimple cvt
= gimple_build_assign_with_ops
1193 (NOP_EXPR
, lhs
, gimple_assign_lhs (stmt
), NULL_TREE
);
1194 gimple_seq_add_stmt (&stmts
, cvt
);
1199 /* Hoist the statements in basic block BB out of the loops prescribed by
1200 data stored in LIM_DATA structures associated with each statement. Callback
1201 for walk_dominator_tree. */
1204 move_computations_dom_walker::before_dom_children (basic_block bb
)
1207 gimple_stmt_iterator bsi
;
1210 struct lim_aux_data
*lim_data
;
1212 if (!loop_outer (bb
->loop_father
))
1215 for (bsi
= gsi_start_phis (bb
); !gsi_end_p (bsi
); )
1218 stmt
= gsi_stmt (bsi
);
1220 lim_data
= get_lim_data (stmt
);
1221 if (lim_data
== NULL
)
1227 cost
= lim_data
->cost
;
1228 level
= lim_data
->tgt_loop
;
1229 clear_lim_data (stmt
);
1237 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1239 fprintf (dump_file
, "Moving PHI node\n");
1240 print_gimple_stmt (dump_file
, stmt
, 0, 0);
1241 fprintf (dump_file
, "(cost %u) out of loop %d.\n\n",
1245 if (gimple_phi_num_args (stmt
) == 1)
1247 tree arg
= PHI_ARG_DEF (stmt
, 0);
1248 new_stmt
= gimple_build_assign_with_ops (TREE_CODE (arg
),
1249 gimple_phi_result (stmt
),
1254 basic_block dom
= get_immediate_dominator (CDI_DOMINATORS
, bb
);
1255 gimple cond
= gsi_stmt (gsi_last_bb (dom
));
1256 tree arg0
= NULL_TREE
, arg1
= NULL_TREE
, t
;
1257 /* Get the PHI arguments corresponding to the true and false
1259 extract_true_false_args_from_phi (dom
, stmt
, &arg0
, &arg1
);
1260 gcc_assert (arg0
&& arg1
);
1261 t
= build2 (gimple_cond_code (cond
), boolean_type_node
,
1262 gimple_cond_lhs (cond
), gimple_cond_rhs (cond
));
1263 new_stmt
= gimple_build_assign_with_ops (COND_EXPR
,
1264 gimple_phi_result (stmt
),
1266 todo_
|= TODO_cleanup_cfg
;
1268 gsi_insert_on_edge (loop_preheader_edge (level
), new_stmt
);
1269 remove_phi_node (&bsi
, false);
1272 for (bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
); )
1276 stmt
= gsi_stmt (bsi
);
1278 lim_data
= get_lim_data (stmt
);
1279 if (lim_data
== NULL
)
1285 cost
= lim_data
->cost
;
1286 level
= lim_data
->tgt_loop
;
1287 clear_lim_data (stmt
);
1295 /* We do not really want to move conditionals out of the loop; we just
1296 placed it here to force its operands to be moved if necessary. */
1297 if (gimple_code (stmt
) == GIMPLE_COND
)
1300 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1302 fprintf (dump_file
, "Moving statement\n");
1303 print_gimple_stmt (dump_file
, stmt
, 0, 0);
1304 fprintf (dump_file
, "(cost %u) out of loop %d.\n\n",
1308 e
= loop_preheader_edge (level
);
1309 gcc_assert (!gimple_vdef (stmt
));
1310 if (gimple_vuse (stmt
))
1312 /* The new VUSE is the one from the virtual PHI in the loop
1313 header or the one already present. */
1314 gimple_stmt_iterator gsi2
;
1315 for (gsi2
= gsi_start_phis (e
->dest
);
1316 !gsi_end_p (gsi2
); gsi_next (&gsi2
))
1318 gimple phi
= gsi_stmt (gsi2
);
1319 if (virtual_operand_p (gimple_phi_result (phi
)))
1321 gimple_set_vuse (stmt
, PHI_ARG_DEF_FROM_EDGE (phi
, e
));
1326 gsi_remove (&bsi
, false);
1327 /* In case this is a stmt that is not unconditionally executed
1328 when the target loop header is executed and the stmt may
1329 invoke undefined integer or pointer overflow rewrite it to
1330 unsigned arithmetic. */
1331 if (is_gimple_assign (stmt
)
1332 && INTEGRAL_TYPE_P (TREE_TYPE (gimple_assign_lhs (stmt
)))
1333 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (gimple_assign_lhs (stmt
)))
1334 && arith_code_with_undefined_signed_overflow
1335 (gimple_assign_rhs_code (stmt
))
1336 && (!ALWAYS_EXECUTED_IN (bb
)
1337 || !(ALWAYS_EXECUTED_IN (bb
) == level
1338 || flow_loop_nested_p (ALWAYS_EXECUTED_IN (bb
), level
))))
1339 gsi_insert_seq_on_edge (e
, rewrite_to_defined_overflow (stmt
));
1341 gsi_insert_on_edge (e
, stmt
);
1345 /* Hoist the statements out of the loops prescribed by data stored in
1346 LIM_DATA structures associated with each statement.*/
1349 move_computations (void)
1351 move_computations_dom_walker
walker (CDI_DOMINATORS
);
1352 walker
.walk (cfun
->cfg
->x_entry_block_ptr
);
1354 gsi_commit_edge_inserts ();
1355 if (need_ssa_update_p (cfun
))
1356 rewrite_into_loop_closed_ssa (NULL
, TODO_update_ssa
);
1358 return walker
.todo_
;
1361 /* Checks whether the statement defining variable *INDEX can be hoisted
1362 out of the loop passed in DATA. Callback for for_each_index. */
1365 may_move_till (tree ref
, tree
*index
, void *data
)
1367 struct loop
*loop
= (struct loop
*) data
, *max_loop
;
1369 /* If REF is an array reference, check also that the step and the lower
1370 bound is invariant in LOOP. */
1371 if (TREE_CODE (ref
) == ARRAY_REF
)
1373 tree step
= TREE_OPERAND (ref
, 3);
1374 tree lbound
= TREE_OPERAND (ref
, 2);
1376 max_loop
= outermost_invariant_loop (step
, loop
);
1380 max_loop
= outermost_invariant_loop (lbound
, loop
);
1385 max_loop
= outermost_invariant_loop (*index
, loop
);
1392 /* If OP is SSA NAME, force the statement that defines it to be
1393 moved out of the LOOP. ORIG_LOOP is the loop in that EXPR is used. */
1396 force_move_till_op (tree op
, struct loop
*orig_loop
, struct loop
*loop
)
1401 || is_gimple_min_invariant (op
))
1404 gcc_assert (TREE_CODE (op
) == SSA_NAME
);
1406 stmt
= SSA_NAME_DEF_STMT (op
);
1407 if (gimple_nop_p (stmt
))
1410 set_level (stmt
, orig_loop
, loop
);
1413 /* Forces statement defining invariants in REF (and *INDEX) to be moved out of
1414 the LOOP. The reference REF is used in the loop ORIG_LOOP. Callback for
1420 struct loop
*orig_loop
;
1424 force_move_till (tree ref
, tree
*index
, void *data
)
1426 struct fmt_data
*fmt_data
= (struct fmt_data
*) data
;
1428 if (TREE_CODE (ref
) == ARRAY_REF
)
1430 tree step
= TREE_OPERAND (ref
, 3);
1431 tree lbound
= TREE_OPERAND (ref
, 2);
1433 force_move_till_op (step
, fmt_data
->orig_loop
, fmt_data
->loop
);
1434 force_move_till_op (lbound
, fmt_data
->orig_loop
, fmt_data
->loop
);
1437 force_move_till_op (*index
, fmt_data
->orig_loop
, fmt_data
->loop
);
1442 /* A function to free the mem_ref object OBJ. */
1445 memref_free (struct mem_ref
*mem
)
1448 vec
<mem_ref_loc
> *accs
;
1450 FOR_EACH_VEC_ELT (mem
->accesses_in_loop
, i
, accs
)
1452 mem
->accesses_in_loop
.release ();
1457 /* Allocates and returns a memory reference description for MEM whose hash
1458 value is HASH and id is ID. */
1461 mem_ref_alloc (tree mem
, unsigned hash
, unsigned id
)
1463 mem_ref_p ref
= XNEW (struct mem_ref
);
1464 ao_ref_init (&ref
->mem
, mem
);
1467 bitmap_initialize (&ref
->stored
, &lim_bitmap_obstack
);
1468 bitmap_initialize (&ref
->indep_loop
, &lim_bitmap_obstack
);
1469 bitmap_initialize (&ref
->dep_loop
, &lim_bitmap_obstack
);
1470 ref
->accesses_in_loop
.create (0);
1475 /* Records memory reference location *LOC in LOOP to the memory reference
1476 description REF. The reference occurs in statement STMT. */
1479 record_mem_ref_loc (mem_ref_p ref
, struct loop
*loop
, gimple stmt
, tree
*loc
)
1483 if (ref
->accesses_in_loop
.length ()
1484 <= (unsigned) loop
->num
)
1485 ref
->accesses_in_loop
.safe_grow_cleared (loop
->num
+ 1);
1489 ref
->accesses_in_loop
[loop
->num
].safe_push (aref
);
1492 /* Marks reference REF as stored in LOOP. */
1495 mark_ref_stored (mem_ref_p ref
, struct loop
*loop
)
1497 while (loop
!= current_loops
->tree_root
1498 && bitmap_set_bit (&ref
->stored
, loop
->num
))
1499 loop
= loop_outer (loop
);
1502 /* Gathers memory references in statement STMT in LOOP, storing the
1503 information about them in the memory_accesses structure. Marks
1504 the vops accessed through unrecognized statements there as
1508 gather_mem_refs_stmt (struct loop
*loop
, gimple stmt
)
1517 if (!gimple_vuse (stmt
))
1520 mem
= simple_mem_ref_in_stmt (stmt
, &is_stored
);
1523 /* We use the shared mem_ref for all unanalyzable refs. */
1524 id
= UNANALYZABLE_MEM_ID
;
1525 ref
= memory_accesses
.refs_list
[id
];
1526 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1528 fprintf (dump_file
, "Unanalyzed memory reference %u: ", id
);
1529 print_gimple_stmt (dump_file
, stmt
, 0, TDF_SLIM
);
1531 is_stored
= gimple_vdef (stmt
);
1535 hash
= iterative_hash_expr (*mem
, 0);
1536 slot
= memory_accesses
.refs
.find_slot_with_hash (*mem
, hash
, INSERT
);
1539 ref
= (mem_ref_p
) *slot
;
1544 id
= memory_accesses
.refs_list
.length ();
1545 ref
= mem_ref_alloc (*mem
, hash
, id
);
1546 memory_accesses
.refs_list
.safe_push (ref
);
1549 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1551 fprintf (dump_file
, "Memory reference %u: ", id
);
1552 print_generic_expr (dump_file
, ref
->mem
.ref
, TDF_SLIM
);
1553 fprintf (dump_file
, "\n");
1557 record_mem_ref_loc (ref
, loop
, stmt
, mem
);
1559 bitmap_set_bit (&memory_accesses
.refs_in_loop
[loop
->num
], ref
->id
);
1562 bitmap_set_bit (&memory_accesses
.refs_stored_in_loop
[loop
->num
], ref
->id
);
1563 mark_ref_stored (ref
, loop
);
1568 static unsigned *bb_loop_postorder
;
1570 /* qsort sort function to sort blocks after their loop fathers postorder. */
1573 sort_bbs_in_loop_postorder_cmp (const void *bb1_
, const void *bb2_
)
1575 basic_block bb1
= *(basic_block
*)const_cast<void *>(bb1_
);
1576 basic_block bb2
= *(basic_block
*)const_cast<void *>(bb2_
);
1577 struct loop
*loop1
= bb1
->loop_father
;
1578 struct loop
*loop2
= bb2
->loop_father
;
1579 if (loop1
->num
== loop2
->num
)
1581 return bb_loop_postorder
[loop1
->num
] < bb_loop_postorder
[loop2
->num
] ? -1 : 1;
1584 /* Gathers memory references in loops. */
1587 analyze_memory_references (void)
1589 gimple_stmt_iterator bsi
;
1590 basic_block bb
, *bbs
;
1591 struct loop
*loop
, *outer
;
1594 /* Initialize bb_loop_postorder with a mapping from loop->num to
1595 its postorder index. */
1597 bb_loop_postorder
= XNEWVEC (unsigned, number_of_loops (cfun
));
1598 FOR_EACH_LOOP (loop
, LI_FROM_INNERMOST
)
1599 bb_loop_postorder
[loop
->num
] = i
++;
1600 /* Collect all basic-blocks in loops and sort them after their
1603 bbs
= XNEWVEC (basic_block
, n_basic_blocks_for_fn (cfun
) - NUM_FIXED_BLOCKS
);
1604 FOR_EACH_BB_FN (bb
, cfun
)
1605 if (bb
->loop_father
!= current_loops
->tree_root
)
1608 qsort (bbs
, n
, sizeof (basic_block
), sort_bbs_in_loop_postorder_cmp
);
1609 free (bb_loop_postorder
);
1611 /* Visit blocks in loop postorder and assign mem-ref IDs in that order.
1612 That results in better locality for all the bitmaps. */
1613 for (i
= 0; i
< n
; ++i
)
1615 basic_block bb
= bbs
[i
];
1616 for (bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
1617 gather_mem_refs_stmt (bb
->loop_father
, gsi_stmt (bsi
));
1622 /* Propagate the information about accessed memory references up
1623 the loop hierarchy. */
1624 FOR_EACH_LOOP (loop
, LI_FROM_INNERMOST
)
1626 /* Finalize the overall touched references (including subloops). */
1627 bitmap_ior_into (&memory_accesses
.all_refs_stored_in_loop
[loop
->num
],
1628 &memory_accesses
.refs_stored_in_loop
[loop
->num
]);
1630 /* Propagate the information about accessed memory references up
1631 the loop hierarchy. */
1632 outer
= loop_outer (loop
);
1633 if (outer
== current_loops
->tree_root
)
1636 bitmap_ior_into (&memory_accesses
.all_refs_stored_in_loop
[outer
->num
],
1637 &memory_accesses
.all_refs_stored_in_loop
[loop
->num
]);
1641 /* Returns true if MEM1 and MEM2 may alias. TTAE_CACHE is used as a cache in
1642 tree_to_aff_combination_expand. */
1645 mem_refs_may_alias_p (mem_ref_p mem1
, mem_ref_p mem2
,
1646 struct pointer_map_t
**ttae_cache
)
1648 /* Perform BASE + OFFSET analysis -- if MEM1 and MEM2 are based on the same
1649 object and their offset differ in such a way that the locations cannot
1650 overlap, then they cannot alias. */
1651 double_int size1
, size2
;
1652 aff_tree off1
, off2
;
1654 /* Perform basic offset and type-based disambiguation. */
1655 if (!refs_may_alias_p_1 (&mem1
->mem
, &mem2
->mem
, true))
1658 /* The expansion of addresses may be a bit expensive, thus we only do
1659 the check at -O2 and higher optimization levels. */
1663 get_inner_reference_aff (mem1
->mem
.ref
, &off1
, &size1
);
1664 get_inner_reference_aff (mem2
->mem
.ref
, &off2
, &size2
);
1665 aff_combination_expand (&off1
, ttae_cache
);
1666 aff_combination_expand (&off2
, ttae_cache
);
1667 aff_combination_scale (&off1
, double_int_minus_one
);
1668 aff_combination_add (&off2
, &off1
);
1670 if (aff_comb_cannot_overlap_p (&off2
, size1
, size2
))
1676 /* Iterates over all locations of REF in LOOP and its subloops calling
1677 fn.operator() with the location as argument. When that operator
1678 returns true the iteration is stopped and true is returned.
1679 Otherwise false is returned. */
1681 template <typename FN
>
1683 for_all_locs_in_loop (struct loop
*loop
, mem_ref_p ref
, FN fn
)
1687 struct loop
*subloop
;
1689 if (ref
->accesses_in_loop
.length () > (unsigned) loop
->num
)
1690 FOR_EACH_VEC_ELT (ref
->accesses_in_loop
[loop
->num
], i
, loc
)
1694 for (subloop
= loop
->inner
; subloop
!= NULL
; subloop
= subloop
->next
)
1695 if (for_all_locs_in_loop (subloop
, ref
, fn
))
1701 /* Rewrites location LOC by TMP_VAR. */
1703 struct rewrite_mem_ref_loc
1705 rewrite_mem_ref_loc (tree tmp_var_
) : tmp_var (tmp_var_
) {}
1706 bool operator () (mem_ref_loc_p loc
);
1711 rewrite_mem_ref_loc::operator () (mem_ref_loc_p loc
)
1713 *loc
->ref
= tmp_var
;
1714 update_stmt (loc
->stmt
);
1718 /* Rewrites all references to REF in LOOP by variable TMP_VAR. */
1721 rewrite_mem_refs (struct loop
*loop
, mem_ref_p ref
, tree tmp_var
)
1723 for_all_locs_in_loop (loop
, ref
, rewrite_mem_ref_loc (tmp_var
));
1726 /* Stores the first reference location in LOCP. */
1728 struct first_mem_ref_loc_1
1730 first_mem_ref_loc_1 (mem_ref_loc_p
*locp_
) : locp (locp_
) {}
1731 bool operator () (mem_ref_loc_p loc
);
1732 mem_ref_loc_p
*locp
;
1736 first_mem_ref_loc_1::operator () (mem_ref_loc_p loc
)
1742 /* Returns the first reference location to REF in LOOP. */
1744 static mem_ref_loc_p
1745 first_mem_ref_loc (struct loop
*loop
, mem_ref_p ref
)
1747 mem_ref_loc_p locp
= NULL
;
1748 for_all_locs_in_loop (loop
, ref
, first_mem_ref_loc_1 (&locp
));
1752 struct prev_flag_edges
{
1753 /* Edge to insert new flag comparison code. */
1754 edge append_cond_position
;
1756 /* Edge for fall through from previous flag comparison. */
1757 edge last_cond_fallthru
;
1760 /* Helper function for execute_sm. Emit code to store TMP_VAR into
1763 The store is only done if MEM has changed. We do this so no
1764 changes to MEM occur on code paths that did not originally store
1767 The common case for execute_sm will transform:
1787 This function will generate:
1806 execute_sm_if_changed (edge ex
, tree mem
, tree tmp_var
, tree flag
)
1808 basic_block new_bb
, then_bb
, old_dest
;
1809 bool loop_has_only_one_exit
;
1810 edge then_old_edge
, orig_ex
= ex
;
1811 gimple_stmt_iterator gsi
;
1813 struct prev_flag_edges
*prev_edges
= (struct prev_flag_edges
*) ex
->aux
;
1815 /* ?? Insert store after previous store if applicable. See note
1818 ex
= prev_edges
->append_cond_position
;
1820 loop_has_only_one_exit
= single_pred_p (ex
->dest
);
1822 if (loop_has_only_one_exit
)
1823 ex
= split_block_after_labels (ex
->dest
);
1825 old_dest
= ex
->dest
;
1826 new_bb
= split_edge (ex
);
1827 then_bb
= create_empty_bb (new_bb
);
1828 if (current_loops
&& new_bb
->loop_father
)
1829 add_bb_to_loop (then_bb
, new_bb
->loop_father
);
1831 gsi
= gsi_start_bb (new_bb
);
1832 stmt
= gimple_build_cond (NE_EXPR
, flag
, boolean_false_node
,
1833 NULL_TREE
, NULL_TREE
);
1834 gsi_insert_after (&gsi
, stmt
, GSI_CONTINUE_LINKING
);
1836 gsi
= gsi_start_bb (then_bb
);
1837 /* Insert actual store. */
1838 stmt
= gimple_build_assign (unshare_expr (mem
), tmp_var
);
1839 gsi_insert_after (&gsi
, stmt
, GSI_CONTINUE_LINKING
);
1841 make_edge (new_bb
, then_bb
, EDGE_TRUE_VALUE
);
1842 make_edge (new_bb
, old_dest
, EDGE_FALSE_VALUE
);
1843 then_old_edge
= make_edge (then_bb
, old_dest
, EDGE_FALLTHRU
);
1845 set_immediate_dominator (CDI_DOMINATORS
, then_bb
, new_bb
);
1849 basic_block prevbb
= prev_edges
->last_cond_fallthru
->src
;
1850 redirect_edge_succ (prev_edges
->last_cond_fallthru
, new_bb
);
1851 set_immediate_dominator (CDI_DOMINATORS
, new_bb
, prevbb
);
1852 set_immediate_dominator (CDI_DOMINATORS
, old_dest
,
1853 recompute_dominator (CDI_DOMINATORS
, old_dest
));
1856 /* ?? Because stores may alias, they must happen in the exact
1857 sequence they originally happened. Save the position right after
1858 the (_lsm) store we just created so we can continue appending after
1859 it and maintain the original order. */
1861 struct prev_flag_edges
*p
;
1864 orig_ex
->aux
= NULL
;
1865 alloc_aux_for_edge (orig_ex
, sizeof (struct prev_flag_edges
));
1866 p
= (struct prev_flag_edges
*) orig_ex
->aux
;
1867 p
->append_cond_position
= then_old_edge
;
1868 p
->last_cond_fallthru
= find_edge (new_bb
, old_dest
);
1869 orig_ex
->aux
= (void *) p
;
1872 if (!loop_has_only_one_exit
)
1873 for (gsi
= gsi_start_phis (old_dest
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1875 gimple phi
= gsi_stmt (gsi
);
1878 for (i
= 0; i
< gimple_phi_num_args (phi
); i
++)
1879 if (gimple_phi_arg_edge (phi
, i
)->src
== new_bb
)
1881 tree arg
= gimple_phi_arg_def (phi
, i
);
1882 add_phi_arg (phi
, arg
, then_old_edge
, UNKNOWN_LOCATION
);
1886 /* Remove the original fall through edge. This was the
1887 single_succ_edge (new_bb). */
1888 EDGE_SUCC (new_bb
, 0)->flags
&= ~EDGE_FALLTHRU
;
1891 /* When REF is set on the location, set flag indicating the store. */
1893 struct sm_set_flag_if_changed
1895 sm_set_flag_if_changed (tree flag_
) : flag (flag_
) {}
1896 bool operator () (mem_ref_loc_p loc
);
1901 sm_set_flag_if_changed::operator () (mem_ref_loc_p loc
)
1903 /* Only set the flag for writes. */
1904 if (is_gimple_assign (loc
->stmt
)
1905 && gimple_assign_lhs_ptr (loc
->stmt
) == loc
->ref
)
1907 gimple_stmt_iterator gsi
= gsi_for_stmt (loc
->stmt
);
1908 gimple stmt
= gimple_build_assign (flag
, boolean_true_node
);
1909 gsi_insert_after (&gsi
, stmt
, GSI_CONTINUE_LINKING
);
1914 /* Helper function for execute_sm. On every location where REF is
1915 set, set an appropriate flag indicating the store. */
1918 execute_sm_if_changed_flag_set (struct loop
*loop
, mem_ref_p ref
)
1921 char *str
= get_lsm_tmp_name (ref
->mem
.ref
, ~0, "_flag");
1922 flag
= create_tmp_reg (boolean_type_node
, str
);
1923 for_all_locs_in_loop (loop
, ref
, sm_set_flag_if_changed (flag
));
1927 /* Executes store motion of memory reference REF from LOOP.
1928 Exits from the LOOP are stored in EXITS. The initialization of the
1929 temporary variable is put to the preheader of the loop, and assignments
1930 to the reference from the temporary variable are emitted to exits. */
1933 execute_sm (struct loop
*loop
, vec
<edge
> exits
, mem_ref_p ref
)
1935 tree tmp_var
, store_flag
;
1938 struct fmt_data fmt_data
;
1940 struct lim_aux_data
*lim_data
;
1941 bool multi_threaded_model_p
= false;
1942 gimple_stmt_iterator gsi
;
1944 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1946 fprintf (dump_file
, "Executing store motion of ");
1947 print_generic_expr (dump_file
, ref
->mem
.ref
, 0);
1948 fprintf (dump_file
, " from loop %d\n", loop
->num
);
1951 tmp_var
= create_tmp_reg (TREE_TYPE (ref
->mem
.ref
),
1952 get_lsm_tmp_name (ref
->mem
.ref
, ~0));
1954 fmt_data
.loop
= loop
;
1955 fmt_data
.orig_loop
= loop
;
1956 for_each_index (&ref
->mem
.ref
, force_move_till
, &fmt_data
);
1958 if (bb_in_transaction (loop_preheader_edge (loop
)->src
)
1959 || !PARAM_VALUE (PARAM_ALLOW_STORE_DATA_RACES
))
1960 multi_threaded_model_p
= true;
1962 if (multi_threaded_model_p
)
1963 store_flag
= execute_sm_if_changed_flag_set (loop
, ref
);
1965 rewrite_mem_refs (loop
, ref
, tmp_var
);
1967 /* Emit the load code on a random exit edge or into the latch if
1968 the loop does not exit, so that we are sure it will be processed
1969 by move_computations after all dependencies. */
1970 gsi
= gsi_for_stmt (first_mem_ref_loc (loop
, ref
)->stmt
);
1972 /* FIXME/TODO: For the multi-threaded variant, we could avoid this
1973 load altogether, since the store is predicated by a flag. We
1974 could, do the load only if it was originally in the loop. */
1975 load
= gimple_build_assign (tmp_var
, unshare_expr (ref
->mem
.ref
));
1976 lim_data
= init_lim_data (load
);
1977 lim_data
->max_loop
= loop
;
1978 lim_data
->tgt_loop
= loop
;
1979 gsi_insert_before (&gsi
, load
, GSI_SAME_STMT
);
1981 if (multi_threaded_model_p
)
1983 load
= gimple_build_assign (store_flag
, boolean_false_node
);
1984 lim_data
= init_lim_data (load
);
1985 lim_data
->max_loop
= loop
;
1986 lim_data
->tgt_loop
= loop
;
1987 gsi_insert_before (&gsi
, load
, GSI_SAME_STMT
);
1990 /* Sink the store to every exit from the loop. */
1991 FOR_EACH_VEC_ELT (exits
, i
, ex
)
1992 if (!multi_threaded_model_p
)
1995 store
= gimple_build_assign (unshare_expr (ref
->mem
.ref
), tmp_var
);
1996 gsi_insert_on_edge (ex
, store
);
1999 execute_sm_if_changed (ex
, ref
->mem
.ref
, tmp_var
, store_flag
);
2002 /* Hoists memory references MEM_REFS out of LOOP. EXITS is the list of exit
2003 edges of the LOOP. */
2006 hoist_memory_references (struct loop
*loop
, bitmap mem_refs
,
2013 EXECUTE_IF_SET_IN_BITMAP (mem_refs
, 0, i
, bi
)
2015 ref
= memory_accesses
.refs_list
[i
];
2016 execute_sm (loop
, exits
, ref
);
2020 struct ref_always_accessed
2022 ref_always_accessed (struct loop
*loop_
, tree base_
, bool stored_p_
)
2023 : loop (loop_
), base (base_
), stored_p (stored_p_
) {}
2024 bool operator () (mem_ref_loc_p loc
);
2031 ref_always_accessed::operator () (mem_ref_loc_p loc
)
2033 struct loop
*must_exec
;
2035 if (!get_lim_data (loc
->stmt
))
2038 /* If we require an always executed store make sure the statement
2039 stores to the reference. */
2043 if (!gimple_get_lhs (loc
->stmt
))
2045 lhs
= get_base_address (gimple_get_lhs (loc
->stmt
));
2048 if (INDIRECT_REF_P (lhs
)
2049 || TREE_CODE (lhs
) == MEM_REF
)
2050 lhs
= TREE_OPERAND (lhs
, 0);
2055 must_exec
= get_lim_data (loc
->stmt
)->always_executed_in
;
2059 if (must_exec
== loop
2060 || flow_loop_nested_p (must_exec
, loop
))
2066 /* Returns true if REF is always accessed in LOOP. If STORED_P is true
2067 make sure REF is always stored to in LOOP. */
2070 ref_always_accessed_p (struct loop
*loop
, mem_ref_p ref
, bool stored_p
)
2072 tree base
= ao_ref_base (&ref
->mem
);
2073 if (TREE_CODE (base
) == MEM_REF
)
2074 base
= TREE_OPERAND (base
, 0);
2076 return for_all_locs_in_loop (loop
, ref
,
2077 ref_always_accessed (loop
, base
, stored_p
));
2080 /* Returns true if REF1 and REF2 are independent. */
2083 refs_independent_p (mem_ref_p ref1
, mem_ref_p ref2
)
2088 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2089 fprintf (dump_file
, "Querying dependency of refs %u and %u: ",
2090 ref1
->id
, ref2
->id
);
2092 if (mem_refs_may_alias_p (ref1
, ref2
, &memory_accesses
.ttae_cache
))
2094 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2095 fprintf (dump_file
, "dependent.\n");
2100 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2101 fprintf (dump_file
, "independent.\n");
2106 /* Mark REF dependent on stores or loads (according to STORED_P) in LOOP
2107 and its super-loops. */
2110 record_dep_loop (struct loop
*loop
, mem_ref_p ref
, bool stored_p
)
2112 /* We can propagate dependent-in-loop bits up the loop
2113 hierarchy to all outer loops. */
2114 while (loop
!= current_loops
->tree_root
2115 && bitmap_set_bit (&ref
->dep_loop
, LOOP_DEP_BIT (loop
->num
, stored_p
)))
2116 loop
= loop_outer (loop
);
2119 /* Returns true if REF is independent on all other memory references in
2123 ref_indep_loop_p_1 (struct loop
*loop
, mem_ref_p ref
, bool stored_p
)
2125 bitmap refs_to_check
;
2131 refs_to_check
= &memory_accesses
.refs_in_loop
[loop
->num
];
2133 refs_to_check
= &memory_accesses
.refs_stored_in_loop
[loop
->num
];
2135 if (bitmap_bit_p (refs_to_check
, UNANALYZABLE_MEM_ID
))
2138 EXECUTE_IF_SET_IN_BITMAP (refs_to_check
, 0, i
, bi
)
2140 aref
= memory_accesses
.refs_list
[i
];
2141 if (!refs_independent_p (ref
, aref
))
2148 /* Returns true if REF is independent on all other memory references in
2149 LOOP. Wrapper over ref_indep_loop_p_1, caching its results. */
2152 ref_indep_loop_p_2 (struct loop
*loop
, mem_ref_p ref
, bool stored_p
)
2154 stored_p
|= bitmap_bit_p (&ref
->stored
, loop
->num
);
2156 if (bitmap_bit_p (&ref
->indep_loop
, LOOP_DEP_BIT (loop
->num
, stored_p
)))
2158 if (bitmap_bit_p (&ref
->dep_loop
, LOOP_DEP_BIT (loop
->num
, stored_p
)))
2161 struct loop
*inner
= loop
->inner
;
2164 if (!ref_indep_loop_p_2 (inner
, ref
, stored_p
))
2166 inner
= inner
->next
;
2169 bool indep_p
= ref_indep_loop_p_1 (loop
, ref
, stored_p
);
2171 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2172 fprintf (dump_file
, "Querying dependencies of ref %u in loop %d: %s\n",
2173 ref
->id
, loop
->num
, indep_p
? "independent" : "dependent");
2175 /* Record the computed result in the cache. */
2178 if (bitmap_set_bit (&ref
->indep_loop
, LOOP_DEP_BIT (loop
->num
, stored_p
))
2181 /* If it's independend against all refs then it's independent
2182 against stores, too. */
2183 bitmap_set_bit (&ref
->indep_loop
, LOOP_DEP_BIT (loop
->num
, false));
2188 record_dep_loop (loop
, ref
, stored_p
);
2191 /* If it's dependent against stores it's dependent against
2193 record_dep_loop (loop
, ref
, true);
2200 /* Returns true if REF is independent on all other memory references in
2204 ref_indep_loop_p (struct loop
*loop
, mem_ref_p ref
)
2206 gcc_checking_assert (MEM_ANALYZABLE (ref
));
2208 return ref_indep_loop_p_2 (loop
, ref
, false);
2211 /* Returns true if we can perform store motion of REF from LOOP. */
2214 can_sm_ref_p (struct loop
*loop
, mem_ref_p ref
)
2218 /* Can't hoist unanalyzable refs. */
2219 if (!MEM_ANALYZABLE (ref
))
2222 /* It should be movable. */
2223 if (!is_gimple_reg_type (TREE_TYPE (ref
->mem
.ref
))
2224 || TREE_THIS_VOLATILE (ref
->mem
.ref
)
2225 || !for_each_index (&ref
->mem
.ref
, may_move_till
, loop
))
2228 /* If it can throw fail, we do not properly update EH info. */
2229 if (tree_could_throw_p (ref
->mem
.ref
))
2232 /* If it can trap, it must be always executed in LOOP.
2233 Readonly memory locations may trap when storing to them, but
2234 tree_could_trap_p is a predicate for rvalues, so check that
2236 base
= get_base_address (ref
->mem
.ref
);
2237 if ((tree_could_trap_p (ref
->mem
.ref
)
2238 || (DECL_P (base
) && TREE_READONLY (base
)))
2239 && !ref_always_accessed_p (loop
, ref
, true))
2242 /* And it must be independent on all other memory references
2244 if (!ref_indep_loop_p (loop
, ref
))
2250 /* Marks the references in LOOP for that store motion should be performed
2251 in REFS_TO_SM. SM_EXECUTED is the set of references for that store
2252 motion was performed in one of the outer loops. */
2255 find_refs_for_sm (struct loop
*loop
, bitmap sm_executed
, bitmap refs_to_sm
)
2257 bitmap refs
= &memory_accesses
.all_refs_stored_in_loop
[loop
->num
];
2262 EXECUTE_IF_AND_COMPL_IN_BITMAP (refs
, sm_executed
, 0, i
, bi
)
2264 ref
= memory_accesses
.refs_list
[i
];
2265 if (can_sm_ref_p (loop
, ref
))
2266 bitmap_set_bit (refs_to_sm
, i
);
2270 /* Checks whether LOOP (with exits stored in EXITS array) is suitable
2271 for a store motion optimization (i.e. whether we can insert statement
2275 loop_suitable_for_sm (struct loop
*loop ATTRIBUTE_UNUSED
,
2281 FOR_EACH_VEC_ELT (exits
, i
, ex
)
2282 if (ex
->flags
& (EDGE_ABNORMAL
| EDGE_EH
))
2288 /* Try to perform store motion for all memory references modified inside
2289 LOOP. SM_EXECUTED is the bitmap of the memory references for that
2290 store motion was executed in one of the outer loops. */
2293 store_motion_loop (struct loop
*loop
, bitmap sm_executed
)
2295 vec
<edge
> exits
= get_loop_exit_edges (loop
);
2296 struct loop
*subloop
;
2297 bitmap sm_in_loop
= BITMAP_ALLOC (&lim_bitmap_obstack
);
2299 if (loop_suitable_for_sm (loop
, exits
))
2301 find_refs_for_sm (loop
, sm_executed
, sm_in_loop
);
2302 hoist_memory_references (loop
, sm_in_loop
, exits
);
2306 bitmap_ior_into (sm_executed
, sm_in_loop
);
2307 for (subloop
= loop
->inner
; subloop
!= NULL
; subloop
= subloop
->next
)
2308 store_motion_loop (subloop
, sm_executed
);
2309 bitmap_and_compl_into (sm_executed
, sm_in_loop
);
2310 BITMAP_FREE (sm_in_loop
);
2313 /* Try to perform store motion for all memory references modified inside
2320 bitmap sm_executed
= BITMAP_ALLOC (&lim_bitmap_obstack
);
2322 for (loop
= current_loops
->tree_root
->inner
; loop
!= NULL
; loop
= loop
->next
)
2323 store_motion_loop (loop
, sm_executed
);
2325 BITMAP_FREE (sm_executed
);
2326 gsi_commit_edge_inserts ();
2329 /* Fills ALWAYS_EXECUTED_IN information for basic blocks of LOOP, i.e.
2330 for each such basic block bb records the outermost loop for that execution
2331 of its header implies execution of bb. CONTAINS_CALL is the bitmap of
2332 blocks that contain a nonpure call. */
2335 fill_always_executed_in_1 (struct loop
*loop
, sbitmap contains_call
)
2337 basic_block bb
= NULL
, *bbs
, last
= NULL
;
2340 struct loop
*inn_loop
= loop
;
2342 if (ALWAYS_EXECUTED_IN (loop
->header
) == NULL
)
2344 bbs
= get_loop_body_in_dom_order (loop
);
2346 for (i
= 0; i
< loop
->num_nodes
; i
++)
2351 if (dominated_by_p (CDI_DOMINATORS
, loop
->latch
, bb
))
2354 if (bitmap_bit_p (contains_call
, bb
->index
))
2357 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
2358 if (!flow_bb_inside_loop_p (loop
, e
->dest
))
2363 /* A loop might be infinite (TODO use simple loop analysis
2364 to disprove this if possible). */
2365 if (bb
->flags
& BB_IRREDUCIBLE_LOOP
)
2368 if (!flow_bb_inside_loop_p (inn_loop
, bb
))
2371 if (bb
->loop_father
->header
== bb
)
2373 if (!dominated_by_p (CDI_DOMINATORS
, loop
->latch
, bb
))
2376 /* In a loop that is always entered we may proceed anyway.
2377 But record that we entered it and stop once we leave it. */
2378 inn_loop
= bb
->loop_father
;
2384 SET_ALWAYS_EXECUTED_IN (last
, loop
);
2385 if (last
== loop
->header
)
2387 last
= get_immediate_dominator (CDI_DOMINATORS
, last
);
2393 for (loop
= loop
->inner
; loop
; loop
= loop
->next
)
2394 fill_always_executed_in_1 (loop
, contains_call
);
2397 /* Fills ALWAYS_EXECUTED_IN information for basic blocks, i.e.
2398 for each such basic block bb records the outermost loop for that execution
2399 of its header implies execution of bb. */
2402 fill_always_executed_in (void)
2404 sbitmap contains_call
= sbitmap_alloc (last_basic_block_for_fn (cfun
));
2408 bitmap_clear (contains_call
);
2409 FOR_EACH_BB_FN (bb
, cfun
)
2411 gimple_stmt_iterator gsi
;
2412 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
2414 if (nonpure_call_p (gsi_stmt (gsi
)))
2418 if (!gsi_end_p (gsi
))
2419 bitmap_set_bit (contains_call
, bb
->index
);
2422 for (loop
= current_loops
->tree_root
->inner
; loop
; loop
= loop
->next
)
2423 fill_always_executed_in_1 (loop
, contains_call
);
2425 sbitmap_free (contains_call
);
2429 /* Compute the global information needed by the loop invariant motion pass. */
2432 tree_ssa_lim_initialize (void)
2436 bitmap_obstack_initialize (&lim_bitmap_obstack
);
2437 lim_aux_data_map
= pointer_map_create ();
2440 compute_transaction_bits ();
2442 alloc_aux_for_edges (0);
2444 memory_accesses
.refs
.create (100);
2445 memory_accesses
.refs_list
.create (100);
2446 /* Allocate a special, unanalyzable mem-ref with ID zero. */
2447 memory_accesses
.refs_list
.quick_push
2448 (mem_ref_alloc (error_mark_node
, 0, UNANALYZABLE_MEM_ID
));
2450 memory_accesses
.refs_in_loop
.create (number_of_loops (cfun
));
2451 memory_accesses
.refs_in_loop
.quick_grow (number_of_loops (cfun
));
2452 memory_accesses
.refs_stored_in_loop
.create (number_of_loops (cfun
));
2453 memory_accesses
.refs_stored_in_loop
.quick_grow (number_of_loops (cfun
));
2454 memory_accesses
.all_refs_stored_in_loop
.create (number_of_loops (cfun
));
2455 memory_accesses
.all_refs_stored_in_loop
.quick_grow (number_of_loops (cfun
));
2457 for (i
= 0; i
< number_of_loops (cfun
); i
++)
2459 bitmap_initialize (&memory_accesses
.refs_in_loop
[i
],
2460 &lim_bitmap_obstack
);
2461 bitmap_initialize (&memory_accesses
.refs_stored_in_loop
[i
],
2462 &lim_bitmap_obstack
);
2463 bitmap_initialize (&memory_accesses
.all_refs_stored_in_loop
[i
],
2464 &lim_bitmap_obstack
);
2467 memory_accesses
.ttae_cache
= NULL
;
2470 /* Cleans up after the invariant motion pass. */
2473 tree_ssa_lim_finalize (void)
2479 free_aux_for_edges ();
2481 FOR_EACH_BB_FN (bb
, cfun
)
2482 SET_ALWAYS_EXECUTED_IN (bb
, NULL
);
2484 bitmap_obstack_release (&lim_bitmap_obstack
);
2485 pointer_map_destroy (lim_aux_data_map
);
2487 memory_accesses
.refs
.dispose ();
2489 FOR_EACH_VEC_ELT (memory_accesses
.refs_list
, i
, ref
)
2491 memory_accesses
.refs_list
.release ();
2493 memory_accesses
.refs_in_loop
.release ();
2494 memory_accesses
.refs_stored_in_loop
.release ();
2495 memory_accesses
.all_refs_stored_in_loop
.release ();
2497 if (memory_accesses
.ttae_cache
)
2498 free_affine_expand_cache (&memory_accesses
.ttae_cache
);
2501 /* Moves invariants from loops. Only "expensive" invariants are moved out --
2502 i.e. those that are likely to be win regardless of the register pressure. */
2509 tree_ssa_lim_initialize ();
2511 /* Gathers information about memory accesses in the loops. */
2512 analyze_memory_references ();
2514 /* Fills ALWAYS_EXECUTED_IN information for basic blocks. */
2515 fill_always_executed_in ();
2517 /* For each statement determine the outermost loop in that it is
2518 invariant and cost for computing the invariant. */
2519 invariantness_dom_walker (CDI_DOMINATORS
)
2520 .walk (cfun
->cfg
->x_entry_block_ptr
);
2522 /* Execute store motion. Force the necessary invariants to be moved
2523 out of the loops as well. */
2526 /* Move the expressions that are expensive enough. */
2527 todo
= move_computations ();
2529 tree_ssa_lim_finalize ();
2534 /* Loop invariant motion pass. */
2537 tree_ssa_loop_im (void)
2539 if (number_of_loops (cfun
) <= 1)
2542 return tree_ssa_lim ();
2546 gate_tree_ssa_loop_im (void)
2548 return flag_tree_loop_im
!= 0;
2553 const pass_data pass_data_lim
=
2555 GIMPLE_PASS
, /* type */
2557 OPTGROUP_LOOP
, /* optinfo_flags */
2558 true, /* has_gate */
2559 true, /* has_execute */
2561 PROP_cfg
, /* properties_required */
2562 0, /* properties_provided */
2563 0, /* properties_destroyed */
2564 0, /* todo_flags_start */
2565 0, /* todo_flags_finish */
2568 class pass_lim
: public gimple_opt_pass
2571 pass_lim (gcc::context
*ctxt
)
2572 : gimple_opt_pass (pass_data_lim
, ctxt
)
2575 /* opt_pass methods: */
2576 opt_pass
* clone () { return new pass_lim (m_ctxt
); }
2577 bool gate () { return gate_tree_ssa_loop_im (); }
2578 unsigned int execute () { return tree_ssa_loop_im (); }
2580 }; // class pass_lim
2585 make_pass_lim (gcc::context
*ctxt
)
2587 return new pass_lim (ctxt
);