1 /* Loop invariant motion.
2 Copyright (C) 2003, 2004 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 2, 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 COPYING. If not, write to the Free
18 Software Foundation, 59 Temple Place - Suite 330, Boston, MA
23 #include "coretypes.h"
28 #include "hard-reg-set.h"
29 #include "basic-block.h"
31 #include "diagnostic.h"
32 #include "tree-flow.h"
33 #include "tree-dump.h"
38 #include "tree-pass.h"
41 /* A type for the list of statements that have to be moved in order to be able
42 to hoist an invariant computation. */
50 /* The auxiliary data kept for each statement. */
54 struct loop
*max_loop
; /* The outermost loop in that the statement
57 struct loop
*tgt_loop
; /* The loop out of that we want to move the
60 struct loop
*always_executed_in
;
61 /* The outermost loop for that we are sure
62 the statement is executed if the loop
65 bool sm_done
; /* True iff the store motion for a memory
66 reference in the statement has already
69 unsigned cost
; /* Cost of the computation performed by the
72 struct depend
*depends
; /* List of statements that must be also hoisted
73 out of the loop when this statement is
74 hoisted; i.e. those that define the operands
75 of the statement and are inside of the
79 #define LIM_DATA(STMT) (TREE_CODE (STMT) == PHI_NODE \
81 : (struct lim_aux_data *) (stmt_ann (STMT)->common.aux))
83 /* Description of a memory reference for store motion. */
87 tree
*ref
; /* The reference itself. */
88 tree stmt
; /* The statement in that it occurs. */
89 struct mem_ref
*next
; /* Next use in the chain. */
92 /* Minimum cost of an expensive expression. */
93 #define LIM_EXPENSIVE ((unsigned) PARAM_VALUE (PARAM_LIM_EXPENSIVE))
95 /* The outermost loop for that execution of the header guarantees that the
96 block will be executed. */
97 #define ALWAYS_EXECUTED_IN(BB) ((struct loop *) (BB)->aux)
99 static unsigned max_stmt_uid
; /* Maximal uid of a statement. Uids to phi
100 nodes are assigned using the versions of
101 ssa names they define. */
103 /* Returns uid of statement STMT. */
106 get_stmt_uid (tree stmt
)
108 if (TREE_CODE (stmt
) == PHI_NODE
)
109 return SSA_NAME_VERSION (PHI_RESULT (stmt
)) + max_stmt_uid
;
111 return stmt_ann (stmt
)->uid
;
114 /* Calls CBCK for each index in memory reference ADDR_P. There are two
115 kinds situations handled; in each of these cases, the memory reference
116 and DATA are passed to the callback:
118 Access to an array: ARRAY_{RANGE_}REF (base, index). In this case we also
119 pass the pointer to the index to the callback.
121 Pointer dereference: INDIRECT_REF (addr). In this case we also pass the
122 pointer to addr to the callback.
124 If the callback returns false, the whole search stops and false is returned.
125 Otherwise the function returns true after traversing through the whole
126 reference *ADDR_P. */
129 for_each_index (tree
*addr_p
, bool (*cbck
) (tree
, tree
*, void *), void *data
)
133 for (; ; addr_p
= nxt
)
135 switch (TREE_CODE (*addr_p
))
138 return cbck (*addr_p
, addr_p
, data
);
140 case MISALIGNED_INDIRECT_REF
:
141 case ALIGN_INDIRECT_REF
:
143 nxt
= &TREE_OPERAND (*addr_p
, 0);
144 return cbck (*addr_p
, nxt
, data
);
147 case VIEW_CONVERT_EXPR
:
148 case ARRAY_RANGE_REF
:
151 nxt
= &TREE_OPERAND (*addr_p
, 0);
155 /* If the component has varying offset, it behaves like index
157 idx
= &TREE_OPERAND (*addr_p
, 2);
159 && !cbck (*addr_p
, idx
, data
))
162 nxt
= &TREE_OPERAND (*addr_p
, 0);
166 nxt
= &TREE_OPERAND (*addr_p
, 0);
167 if (!cbck (*addr_p
, &TREE_OPERAND (*addr_p
, 1), data
))
183 /* If it is possible to hoist the statement STMT unconditionally,
184 returns MOVE_POSSIBLE.
185 If it is possible to hoist the statement STMT, but we must avoid making
186 it executed if it would not be executed in the original program (e.g.
187 because it may trap), return MOVE_PRESERVE_EXECUTION.
188 Otherwise return MOVE_IMPOSSIBLE. */
191 movement_possibility (tree stmt
)
195 if (flag_unswitch_loops
196 && TREE_CODE (stmt
) == COND_EXPR
)
198 /* If we perform unswitching, force the operands of the invariant
199 condition to be moved out of the loop. */
200 get_stmt_operands (stmt
);
202 return MOVE_POSSIBLE
;
205 if (TREE_CODE (stmt
) != MODIFY_EXPR
)
206 return MOVE_IMPOSSIBLE
;
208 if (stmt_ends_bb_p (stmt
))
209 return MOVE_IMPOSSIBLE
;
211 get_stmt_operands (stmt
);
213 if (stmt_ann (stmt
)->has_volatile_ops
)
214 return MOVE_IMPOSSIBLE
;
216 lhs
= TREE_OPERAND (stmt
, 0);
217 if (TREE_CODE (lhs
) == SSA_NAME
218 && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs
))
219 return MOVE_IMPOSSIBLE
;
221 rhs
= TREE_OPERAND (stmt
, 1);
223 if (TREE_SIDE_EFFECTS (rhs
))
224 return MOVE_IMPOSSIBLE
;
226 if (TREE_CODE (lhs
) != SSA_NAME
227 || tree_could_trap_p (rhs
))
228 return MOVE_PRESERVE_EXECUTION
;
230 return MOVE_POSSIBLE
;
233 /* Suppose that operand DEF is used inside the LOOP. Returns the outermost
234 loop to that we could move the expression using DEF if it did not have
235 other operands, i.e. the outermost loop enclosing LOOP in that the value
236 of DEF is invariant. */
239 outermost_invariant_loop (tree def
, struct loop
*loop
)
243 struct loop
*max_loop
;
245 if (TREE_CODE (def
) != SSA_NAME
)
246 return superloop_at_depth (loop
, 1);
248 def_stmt
= SSA_NAME_DEF_STMT (def
);
249 def_bb
= bb_for_stmt (def_stmt
);
251 return superloop_at_depth (loop
, 1);
253 max_loop
= find_common_loop (loop
, def_bb
->loop_father
);
255 if (LIM_DATA (def_stmt
) && LIM_DATA (def_stmt
)->max_loop
)
256 max_loop
= find_common_loop (max_loop
,
257 LIM_DATA (def_stmt
)->max_loop
->outer
);
258 if (max_loop
== loop
)
260 max_loop
= superloop_at_depth (loop
, max_loop
->depth
+ 1);
265 /* Returns the outermost superloop of LOOP in that the expression EXPR is
269 outermost_invariant_loop_expr (tree expr
, struct loop
*loop
)
271 enum tree_code_class
class = TREE_CODE_CLASS (TREE_CODE (expr
));
273 struct loop
*max_loop
= superloop_at_depth (loop
, 1), *aloop
;
275 if (TREE_CODE (expr
) == SSA_NAME
276 || TREE_CODE (expr
) == INTEGER_CST
277 || is_gimple_min_invariant (expr
))
278 return outermost_invariant_loop (expr
, loop
);
280 if (class != tcc_unary
281 && class != tcc_binary
282 && class != tcc_expression
283 && class != tcc_comparison
)
286 nops
= first_rtl_op (TREE_CODE (expr
));
287 for (i
= 0; i
< nops
; i
++)
289 aloop
= outermost_invariant_loop_expr (TREE_OPERAND (expr
, i
), loop
);
293 if (flow_loop_nested_p (max_loop
, aloop
))
300 /* DATA is a structure containing information associated with a statement
301 inside LOOP. DEF is one of the operands of this statement.
303 Find the outermost loop enclosing LOOP in that value of DEF is invariant
304 and record this in DATA->max_loop field. If DEF itself is defined inside
305 this loop as well (i.e. we need to hoist it out of the loop if we want
306 to hoist the statement represented by DATA), record the statement in that
307 DEF is defined to the DATA->depends list. Additionally if ADD_COST is true,
308 add the cost of the computation of DEF to the DATA->cost.
310 If DEF is not invariant in LOOP, return false. Otherwise return TRUE. */
313 add_dependency (tree def
, struct lim_aux_data
*data
, struct loop
*loop
,
316 tree def_stmt
= SSA_NAME_DEF_STMT (def
);
317 basic_block def_bb
= bb_for_stmt (def_stmt
);
318 struct loop
*max_loop
;
324 max_loop
= outermost_invariant_loop (def
, loop
);
328 if (flow_loop_nested_p (data
->max_loop
, max_loop
))
329 data
->max_loop
= max_loop
;
331 if (!LIM_DATA (def_stmt
))
335 /* Only add the cost if the statement defining DEF is inside LOOP,
336 i.e. if it is likely that by moving the invariants dependent
337 on it, we will be able to avoid creating a new register for
338 it (since it will be only used in these dependent invariants). */
339 && def_bb
->loop_father
== loop
)
340 data
->cost
+= LIM_DATA (def_stmt
)->cost
;
342 dep
= xmalloc (sizeof (struct depend
));
343 dep
->stmt
= def_stmt
;
344 dep
->next
= data
->depends
;
350 /* Returns an estimate for a cost of statement STMT. TODO -- the values here
351 are just ad-hoc constants. The estimates should be based on target-specific
355 stmt_cost (tree stmt
)
360 /* Always try to create possibilities for unswitching. */
361 if (TREE_CODE (stmt
) == COND_EXPR
)
362 return LIM_EXPENSIVE
;
364 lhs
= TREE_OPERAND (stmt
, 0);
365 rhs
= TREE_OPERAND (stmt
, 1);
367 /* Hoisting memory references out should almost surely be a win. */
368 if (!is_gimple_variable (lhs
))
370 if (is_gimple_addressable (rhs
) && !is_gimple_variable (rhs
))
373 switch (TREE_CODE (rhs
))
376 /* We should be hoisting calls if possible. */
378 /* Unless the call is a builtin_constant_p; this always folds to a
379 constant, so moving it is useless. */
380 rhs
= get_callee_fndecl (rhs
);
381 if (DECL_BUILT_IN (rhs
)
382 && DECL_FUNCTION_CODE (rhs
) == BUILT_IN_CONSTANT_P
)
398 /* Division and multiplication are usually expensive. */
409 /* Determine the outermost loop to that it is possible to hoist a statement
410 STMT and store it to LIM_DATA (STMT)->max_loop. To do this we determine
411 the outermost loop in that the value computed by STMT is invariant.
412 If MUST_PRESERVE_EXEC is true, additionally choose such a loop that
413 we preserve the fact whether STMT is executed. It also fills other related
414 information to LIM_DATA (STMT).
416 The function returns false if STMT cannot be hoisted outside of the loop it
417 is defined in, and true otherwise. */
420 determine_max_movement (tree stmt
, bool must_preserve_exec
)
422 basic_block bb
= bb_for_stmt (stmt
);
423 struct loop
*loop
= bb
->loop_father
;
425 struct lim_aux_data
*lim_data
= LIM_DATA (stmt
);
429 if (must_preserve_exec
)
430 level
= ALWAYS_EXECUTED_IN (bb
);
432 level
= superloop_at_depth (loop
, 1);
433 lim_data
->max_loop
= level
;
435 FOR_EACH_SSA_TREE_OPERAND (val
, stmt
, iter
, SSA_OP_USE
)
436 if (!add_dependency (val
, lim_data
, loop
, true))
439 FOR_EACH_SSA_TREE_OPERAND (val
, stmt
, iter
, SSA_OP_VIRTUAL_USES
)
440 if (!add_dependency (val
, lim_data
, loop
, false))
443 lim_data
->cost
+= stmt_cost (stmt
);
448 /* Suppose that some statement in ORIG_LOOP is hoisted to the loop LEVEL,
449 and that one of the operands of this statement is computed by STMT.
450 Ensure that STMT (together with all the statements that define its
451 operands) is hoisted at least out of the loop LEVEL. */
454 set_level (tree stmt
, struct loop
*orig_loop
, struct loop
*level
)
456 struct loop
*stmt_loop
= bb_for_stmt (stmt
)->loop_father
;
459 stmt_loop
= find_common_loop (orig_loop
, stmt_loop
);
460 if (LIM_DATA (stmt
) && LIM_DATA (stmt
)->tgt_loop
)
461 stmt_loop
= find_common_loop (stmt_loop
,
462 LIM_DATA (stmt
)->tgt_loop
->outer
);
463 if (flow_loop_nested_p (stmt_loop
, level
))
466 gcc_assert (LIM_DATA (stmt
));
467 gcc_assert (level
== LIM_DATA (stmt
)->max_loop
468 || flow_loop_nested_p (LIM_DATA (stmt
)->max_loop
, level
));
470 LIM_DATA (stmt
)->tgt_loop
= level
;
471 for (dep
= LIM_DATA (stmt
)->depends
; dep
; dep
= dep
->next
)
472 set_level (dep
->stmt
, orig_loop
, level
);
475 /* Determines an outermost loop from that we want to hoist the statement STMT.
476 For now we chose the outermost possible loop. TODO -- use profiling
477 information to set it more sanely. */
480 set_profitable_level (tree stmt
)
482 set_level (stmt
, bb_for_stmt (stmt
)->loop_father
, LIM_DATA (stmt
)->max_loop
);
485 /* Returns true if STMT is not a pure call. */
488 nonpure_call_p (tree stmt
)
490 tree call
= get_call_expr_in (stmt
);
495 return TREE_SIDE_EFFECTS (call
) != 0;
498 /* Releases the memory occupied by DATA. */
501 free_lim_aux_data (struct lim_aux_data
*data
)
503 struct depend
*dep
, *next
;
505 for (dep
= data
->depends
; dep
; dep
= next
)
513 /* Determine the outermost loops in that statements in basic block BB are
514 invariant, and record them to the LIM_DATA associated with the statements.
515 Callback for walk_dominator_tree. */
518 determine_invariantness_stmt (struct dom_walk_data
*dw_data ATTRIBUTE_UNUSED
,
522 block_stmt_iterator bsi
;
524 bool maybe_never
= ALWAYS_EXECUTED_IN (bb
) == NULL
;
525 struct loop
*outermost
= ALWAYS_EXECUTED_IN (bb
);
527 if (!bb
->loop_father
->outer
)
530 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
531 fprintf (dump_file
, "Basic block %d (loop %d -- depth %d):\n\n",
532 bb
->index
, bb
->loop_father
->num
, bb
->loop_father
->depth
);
534 for (bsi
= bsi_start (bb
); !bsi_end_p (bsi
); bsi_next (&bsi
))
536 stmt
= bsi_stmt (bsi
);
538 pos
= movement_possibility (stmt
);
539 if (pos
== MOVE_IMPOSSIBLE
)
541 if (nonpure_call_p (stmt
))
549 stmt_ann (stmt
)->common
.aux
= xcalloc (1, sizeof (struct lim_aux_data
));
550 LIM_DATA (stmt
)->always_executed_in
= outermost
;
552 if (maybe_never
&& pos
== MOVE_PRESERVE_EXECUTION
)
555 if (!determine_max_movement (stmt
, pos
== MOVE_PRESERVE_EXECUTION
))
557 LIM_DATA (stmt
)->max_loop
= NULL
;
561 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
563 print_generic_stmt_indented (dump_file
, stmt
, 0, 2);
564 fprintf (dump_file
, " invariant up to level %d, cost %d.\n\n",
565 LIM_DATA (stmt
)->max_loop
->depth
,
566 LIM_DATA (stmt
)->cost
);
569 if (LIM_DATA (stmt
)->cost
>= LIM_EXPENSIVE
)
570 set_profitable_level (stmt
);
574 /* For each statement determines the outermost loop in that it is invariant,
575 statements on whose motion it depends and the cost of the computation.
576 This information is stored to the LIM_DATA structure associated with
580 determine_invariantness (void)
582 struct dom_walk_data walk_data
;
584 memset (&walk_data
, 0, sizeof (struct dom_walk_data
));
585 walk_data
.before_dom_children_before_stmts
= determine_invariantness_stmt
;
587 init_walk_dominator_tree (&walk_data
);
588 walk_dominator_tree (&walk_data
, ENTRY_BLOCK_PTR
);
589 fini_walk_dominator_tree (&walk_data
);
592 /* Commits edge insertions and updates loop structures. */
595 loop_commit_inserts (void)
597 unsigned old_last_basic_block
, i
;
600 old_last_basic_block
= last_basic_block
;
601 bsi_commit_edge_inserts (NULL
);
602 for (i
= old_last_basic_block
; i
< (unsigned) last_basic_block
; i
++)
604 bb
= BASIC_BLOCK (i
);
606 find_common_loop (EDGE_SUCC (bb
, 0)->dest
->loop_father
,
607 EDGE_PRED (bb
, 0)->src
->loop_father
));
611 /* Hoist the statements in basic block BB out of the loops prescribed by
612 data stored in LIM_DATA structures associated with each statement. Callback
613 for walk_dominator_tree. */
616 move_computations_stmt (struct dom_walk_data
*dw_data ATTRIBUTE_UNUSED
,
620 block_stmt_iterator bsi
;
624 if (!bb
->loop_father
->outer
)
627 for (bsi
= bsi_start (bb
); !bsi_end_p (bsi
); )
629 stmt
= bsi_stmt (bsi
);
631 if (!LIM_DATA (stmt
))
637 cost
= LIM_DATA (stmt
)->cost
;
638 level
= LIM_DATA (stmt
)->tgt_loop
;
639 free_lim_aux_data (LIM_DATA (stmt
));
640 stmt_ann (stmt
)->common
.aux
= NULL
;
648 /* We do not really want to move conditionals out of the loop; we just
649 placed it here to force its operands to be moved if necessary. */
650 if (TREE_CODE (stmt
) == COND_EXPR
)
653 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
655 fprintf (dump_file
, "Moving statement\n");
656 print_generic_stmt (dump_file
, stmt
, 0);
657 fprintf (dump_file
, "(cost %u) out of loop %d.\n\n",
660 bsi_insert_on_edge (loop_preheader_edge (level
), stmt
);
665 /* Hoist the statements out of the loops prescribed by data stored in
666 LIM_DATA structures associated with each statement.*/
669 move_computations (void)
671 struct dom_walk_data walk_data
;
673 memset (&walk_data
, 0, sizeof (struct dom_walk_data
));
674 walk_data
.before_dom_children_before_stmts
= move_computations_stmt
;
676 init_walk_dominator_tree (&walk_data
);
677 walk_dominator_tree (&walk_data
, ENTRY_BLOCK_PTR
);
678 fini_walk_dominator_tree (&walk_data
);
680 loop_commit_inserts ();
681 rewrite_into_ssa (false);
682 if (bitmap_first_set_bit (vars_to_rename
) >= 0)
684 /* The rewrite of ssa names may cause violation of loop closed ssa
685 form invariants. TODO -- avoid these rewrites completely.
686 Information in virtual phi nodes is sufficient for it. */
687 rewrite_into_loop_closed_ssa ();
689 bitmap_clear (vars_to_rename
);
692 /* Checks whether the statement defining variable *INDEX can be hoisted
693 out of the loop passed in DATA. Callback for for_each_index. */
696 may_move_till (tree ref
, tree
*index
, void *data
)
698 struct loop
*loop
= data
, *max_loop
;
700 /* If REF is an array reference, check also that the step and the lower
701 bound is invariant in LOOP. */
702 if (TREE_CODE (ref
) == ARRAY_REF
)
704 tree step
= array_ref_element_size (ref
);
705 tree lbound
= array_ref_low_bound (ref
);
707 max_loop
= outermost_invariant_loop_expr (step
, loop
);
711 max_loop
= outermost_invariant_loop_expr (lbound
, loop
);
716 max_loop
= outermost_invariant_loop (*index
, loop
);
723 /* Forces statements defining (invariant) SSA names in expression EXPR to be
724 moved out of the LOOP. ORIG_LOOP is the loop in that EXPR is used. */
727 force_move_till_expr (tree expr
, struct loop
*orig_loop
, struct loop
*loop
)
729 enum tree_code_class
class = TREE_CODE_CLASS (TREE_CODE (expr
));
732 if (TREE_CODE (expr
) == SSA_NAME
)
734 tree stmt
= SSA_NAME_DEF_STMT (expr
);
735 if (IS_EMPTY_STMT (stmt
))
738 set_level (stmt
, orig_loop
, loop
);
742 if (class != tcc_unary
743 && class != tcc_binary
744 && class != tcc_expression
745 && class != tcc_comparison
)
748 nops
= first_rtl_op (TREE_CODE (expr
));
749 for (i
= 0; i
< nops
; i
++)
750 force_move_till_expr (TREE_OPERAND (expr
, i
), orig_loop
, loop
);
753 /* Forces statement defining invariants in REF (and *INDEX) to be moved out of
754 the LOOP. The reference REF is used in the loop ORIG_LOOP. Callback for
760 struct loop
*orig_loop
;
764 force_move_till (tree ref
, tree
*index
, void *data
)
767 struct fmt_data
*fmt_data
= data
;
769 if (TREE_CODE (ref
) == ARRAY_REF
)
771 tree step
= array_ref_element_size (ref
);
772 tree lbound
= array_ref_low_bound (ref
);
774 force_move_till_expr (step
, fmt_data
->orig_loop
, fmt_data
->loop
);
775 force_move_till_expr (lbound
, fmt_data
->orig_loop
, fmt_data
->loop
);
778 if (TREE_CODE (*index
) != SSA_NAME
)
781 stmt
= SSA_NAME_DEF_STMT (*index
);
782 if (IS_EMPTY_STMT (stmt
))
785 set_level (stmt
, fmt_data
->orig_loop
, fmt_data
->loop
);
790 /* Records memory reference *REF (that occurs in statement STMT)
791 to the list MEM_REFS. */
794 record_mem_ref (struct mem_ref
**mem_refs
, tree stmt
, tree
*ref
)
796 struct mem_ref
*aref
= xmalloc (sizeof (struct mem_ref
));
801 aref
->next
= *mem_refs
;
805 /* Releases list of memory references MEM_REFS. */
808 free_mem_refs (struct mem_ref
*mem_refs
)
815 mem_refs
= mem_refs
->next
;
820 /* If VAR is defined in LOOP and the statement it is defined in does not belong
821 to the set SEEN, add the statement to QUEUE of length IN_QUEUE and
825 maybe_queue_var (tree var
, struct loop
*loop
,
826 sbitmap seen
, tree
*queue
, unsigned *in_queue
)
828 tree stmt
= SSA_NAME_DEF_STMT (var
);
829 basic_block def_bb
= bb_for_stmt (stmt
);
832 || !flow_bb_inside_loop_p (loop
, def_bb
)
833 || TEST_BIT (seen
, get_stmt_uid (stmt
)))
836 SET_BIT (seen
, get_stmt_uid (stmt
));
837 queue
[(*in_queue
)++] = stmt
;
840 /* If COMMON_REF is NULL, set COMMON_REF to *OP and return true.
841 Otherwise return true if the memory reference *OP is equal to COMMON_REF.
842 Record the reference OP to list MEM_REFS. STMT is the statement in that
843 the reference occurs. */
847 struct mem_ref
**mem_refs
;
853 fem_single_reachable_address (tree
*op
, void *data
)
855 struct sra_data
*sra_data
= data
;
857 if (sra_data
->common_ref
858 && !operand_equal_p (*op
, sra_data
->common_ref
, 0))
860 sra_data
->common_ref
= *op
;
862 record_mem_ref (sra_data
->mem_refs
, sra_data
->stmt
, op
);
866 /* Runs CALLBACK for each operand of STMT that is a memory reference. DATA
867 is passed to the CALLBACK as well. The traversal stops if CALLBACK
868 returns false, for_each_memref then returns false as well. Otherwise
869 for_each_memref returns true. */
872 for_each_memref (tree stmt
, bool (*callback
)(tree
*, void *), void *data
)
876 if (TREE_CODE (stmt
) == RETURN_EXPR
)
877 stmt
= TREE_OPERAND (stmt
, 1);
879 if (TREE_CODE (stmt
) == MODIFY_EXPR
)
881 op
= &TREE_OPERAND (stmt
, 0);
882 if (TREE_CODE (*op
) != SSA_NAME
883 && !callback (op
, data
))
886 op
= &TREE_OPERAND (stmt
, 1);
887 if (TREE_CODE (*op
) != SSA_NAME
888 && is_gimple_lvalue (*op
)
889 && !callback (op
, data
))
892 stmt
= TREE_OPERAND (stmt
, 1);
895 if (TREE_CODE (stmt
) == WITH_SIZE_EXPR
)
896 stmt
= TREE_OPERAND (stmt
, 0);
898 if (TREE_CODE (stmt
) == CALL_EXPR
)
902 for (args
= TREE_OPERAND (stmt
, 1); args
; args
= TREE_CHAIN (args
))
904 op
= &TREE_VALUE (args
);
906 if (TREE_CODE (*op
) != SSA_NAME
907 && is_gimple_lvalue (*op
)
908 && !callback (op
, data
))
916 /* Determine whether all memory references inside the LOOP that correspond
917 to virtual ssa names defined in statement STMT are equal.
918 If so, store the list of the references to MEM_REFS, and return one
919 of them. Otherwise store NULL to MEM_REFS and return NULL_TREE.
920 *SEEN_CALL_STMT is set to true if the virtual operands suggest
921 that the reference might be clobbered by a call inside the LOOP. */
924 single_reachable_address (struct loop
*loop
, tree stmt
,
925 struct mem_ref
**mem_refs
,
926 bool *seen_call_stmt
)
928 unsigned max_uid
= max_stmt_uid
+ num_ssa_names
;
929 tree
*queue
= xmalloc (sizeof (tree
) * max_uid
);
930 sbitmap seen
= sbitmap_alloc (max_uid
);
931 unsigned in_queue
= 1;
934 struct sra_data sra_data
;
942 sra_data
.mem_refs
= mem_refs
;
943 sra_data
.common_ref
= NULL_TREE
;
946 SET_BIT (seen
, get_stmt_uid (stmt
));
947 *seen_call_stmt
= false;
951 stmt
= queue
[--in_queue
];
952 sra_data
.stmt
= stmt
;
955 && LIM_DATA (stmt
)->sm_done
)
958 switch (TREE_CODE (stmt
))
963 if (!for_each_memref (stmt
, fem_single_reachable_address
,
967 /* If this is a function that may depend on the memory location,
968 record the fact. We cannot directly refuse call clobbered
969 operands here, since sra_data.common_ref did not have
971 call
= get_call_expr_in (stmt
);
973 && !(call_expr_flags (call
) & ECF_CONST
))
974 *seen_call_stmt
= true;
976 /* Traverse also definitions of the VUSES (there may be other
977 distinct from the one we used to get to this statement). */
978 FOR_EACH_SSA_TREE_OPERAND (val
, stmt
, iter
, SSA_OP_VIRTUAL_USES
)
979 maybe_queue_var (val
, loop
, seen
, queue
, &in_queue
);
984 for (i
= 0; i
< (unsigned) PHI_NUM_ARGS (stmt
); i
++)
985 if (TREE_CODE (PHI_ARG_DEF (stmt
, i
)) == SSA_NAME
)
986 maybe_queue_var (PHI_ARG_DEF (stmt
, i
), loop
,
987 seen
, queue
, &in_queue
);
994 /* Find uses of virtual names. */
995 df
= get_immediate_uses (stmt
);
996 n
= num_immediate_uses (df
);
998 for (i
= 0; i
< n
; i
++)
1000 stmt
= immediate_use (df
, i
);
1002 if (!flow_bb_inside_loop_p (loop
, bb_for_stmt (stmt
)))
1005 if (TEST_BIT (seen
, get_stmt_uid (stmt
)))
1007 SET_BIT (seen
, get_stmt_uid (stmt
));
1009 queue
[in_queue
++] = stmt
;
1014 sbitmap_free (seen
);
1016 return sra_data
.common_ref
;
1019 free_mem_refs (*mem_refs
);
1022 sbitmap_free (seen
);
1027 /* Rewrites memory references in list MEM_REFS by variable TMP_VAR. */
1030 rewrite_mem_refs (tree tmp_var
, struct mem_ref
*mem_refs
)
1035 for (; mem_refs
; mem_refs
= mem_refs
->next
)
1037 FOR_EACH_SSA_TREE_OPERAND (var
, mem_refs
->stmt
, iter
,
1038 (SSA_OP_VIRTUAL_DEFS
| SSA_OP_VUSE
))
1040 var
= SSA_NAME_VAR (var
);
1041 bitmap_set_bit (vars_to_rename
, var_ann (var
)->uid
);
1044 *mem_refs
->ref
= tmp_var
;
1045 modify_stmt (mem_refs
->stmt
);
1049 /* Records request for store motion of memory reference REF from LOOP.
1050 MEM_REFS is the list of occurrences of the reference REF inside LOOP;
1051 these references are rewritten by a new temporary variable.
1052 Exits from the LOOP are stored in EXITS, there are N_EXITS of them.
1053 The initialization of the temporary variable is put to the preheader
1054 of the loop, and assignments to the reference from the temporary variable
1055 are emitted to exits. */
1058 schedule_sm (struct loop
*loop
, edge
*exits
, unsigned n_exits
, tree ref
,
1059 struct mem_ref
*mem_refs
)
1061 struct mem_ref
*aref
;
1065 struct fmt_data fmt_data
;
1067 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1069 fprintf (dump_file
, "Executing store motion of ");
1070 print_generic_expr (dump_file
, ref
, 0);
1071 fprintf (dump_file
, " from loop %d\n", loop
->num
);
1074 tmp_var
= make_rename_temp (TREE_TYPE (ref
), "lsm_tmp");
1076 fmt_data
.loop
= loop
;
1077 fmt_data
.orig_loop
= loop
;
1078 for_each_index (&ref
, force_move_till
, &fmt_data
);
1080 rewrite_mem_refs (tmp_var
, mem_refs
);
1081 for (aref
= mem_refs
; aref
; aref
= aref
->next
)
1082 if (LIM_DATA (aref
->stmt
))
1083 LIM_DATA (aref
->stmt
)->sm_done
= true;
1085 /* Emit the load & stores. */
1086 load
= build (MODIFY_EXPR
, void_type_node
, tmp_var
, ref
);
1087 get_stmt_ann (load
)->common
.aux
= xcalloc (1, sizeof (struct lim_aux_data
));
1088 LIM_DATA (load
)->max_loop
= loop
;
1089 LIM_DATA (load
)->tgt_loop
= loop
;
1091 /* Put this into the latch, so that we are sure it will be processed after
1092 all dependencies. */
1093 bsi_insert_on_edge (loop_latch_edge (loop
), load
);
1095 for (i
= 0; i
< n_exits
; i
++)
1097 store
= build (MODIFY_EXPR
, void_type_node
,
1098 unshare_expr (ref
), tmp_var
);
1099 bsi_insert_on_edge (exits
[i
], store
);
1103 /* Returns true if REF may be clobbered by calls. */
1106 is_call_clobbered_ref (tree ref
)
1110 base
= get_base_address (ref
);
1115 return is_call_clobbered (base
);
1117 if (INDIRECT_REF_P (base
))
1119 /* Check whether the alias tags associated with the pointer
1120 are call clobbered. */
1121 tree ptr
= TREE_OPERAND (base
, 0);
1122 struct ptr_info_def
*pi
= SSA_NAME_PTR_INFO (ptr
);
1123 tree nmt
= (pi
) ? pi
->name_mem_tag
: NULL_TREE
;
1124 tree tmt
= var_ann (SSA_NAME_VAR (ptr
))->type_mem_tag
;
1126 if ((nmt
&& is_call_clobbered (nmt
))
1127 || (tmt
&& is_call_clobbered (tmt
)))
1136 /* Determine whether all memory references inside LOOP corresponding to the
1137 virtual ssa name REG are equal to each other, and whether the address of
1138 this common reference can be hoisted outside of the loop. If this is true,
1139 prepare the statements that load the value of the memory reference to a
1140 temporary variable in the loop preheader, store it back on the loop exits,
1141 and replace all the references inside LOOP by this temporary variable.
1142 LOOP has N_EXITS stored in EXITS. */
1145 determine_lsm_reg (struct loop
*loop
, edge
*exits
, unsigned n_exits
, tree reg
)
1148 struct mem_ref
*mem_refs
, *aref
;
1149 struct loop
*must_exec
;
1152 if (is_gimple_reg (reg
))
1155 ref
= single_reachable_address (loop
, SSA_NAME_DEF_STMT (reg
), &mem_refs
,
1160 /* If we cannot create a ssa name for the result, give up. */
1161 if (!is_gimple_reg_type (TREE_TYPE (ref
))
1162 || TREE_THIS_VOLATILE (ref
))
1165 /* If there is a call that may use the location, give up as well. */
1167 && is_call_clobbered_ref (ref
))
1170 if (!for_each_index (&ref
, may_move_till
, loop
))
1173 if (tree_could_trap_p (ref
))
1175 /* If the memory access is unsafe (i.e. it might trap), ensure that some
1176 of the statements in that it occurs is always executed when the loop
1177 is entered. This way we know that by moving the load from the
1178 reference out of the loop we will not cause the error that would not
1181 TODO -- in fact we would like to check for anticipability of the
1182 reference, i.e. that on each path from loop entry to loop exit at
1183 least one of the statements containing the memory reference is
1186 for (aref
= mem_refs
; aref
; aref
= aref
->next
)
1188 if (!LIM_DATA (aref
->stmt
))
1191 must_exec
= LIM_DATA (aref
->stmt
)->always_executed_in
;
1195 if (must_exec
== loop
1196 || flow_loop_nested_p (must_exec
, loop
))
1204 schedule_sm (loop
, exits
, n_exits
, ref
, mem_refs
);
1207 free_mem_refs (mem_refs
);
1210 /* Checks whether LOOP (with N_EXITS exits stored in EXITS array) is suitable
1211 for a store motion optimization (i.e. whether we can insert statement
1215 loop_suitable_for_sm (struct loop
*loop ATTRIBUTE_UNUSED
, edge
*exits
,
1220 for (i
= 0; i
< n_exits
; i
++)
1221 if (exits
[i
]->flags
& EDGE_ABNORMAL
)
1227 /* Try to perform store motion for all memory references modified inside
1231 determine_lsm_loop (struct loop
*loop
)
1235 edge
*exits
= get_loop_exit_edges (loop
, &n_exits
);
1237 if (!loop_suitable_for_sm (loop
, exits
, n_exits
))
1243 for (phi
= phi_nodes (loop
->header
); phi
; phi
= TREE_CHAIN (phi
))
1244 determine_lsm_reg (loop
, exits
, n_exits
, PHI_RESULT (phi
));
1249 /* Try to perform store motion for all memory references modified inside
1253 determine_lsm (struct loops
*loops
)
1258 /* Create a UID for each statement in the function. Ordering of the
1259 UIDs is not important for this pass. */
1263 block_stmt_iterator bsi
;
1265 for (bsi
= bsi_start (bb
); !bsi_end_p (bsi
); bsi_next (&bsi
))
1266 stmt_ann (bsi_stmt (bsi
))->uid
= max_stmt_uid
++;
1269 compute_immediate_uses (TDFA_USE_VOPS
, NULL
);
1271 /* Pass the loops from the outermost. For each virtual operand loop phi node
1272 check whether all the references inside the loop correspond to a single
1273 address, and if so, move them. */
1275 loop
= loops
->tree_root
->inner
;
1278 determine_lsm_loop (loop
);
1288 if (loop
== loops
->tree_root
)
1291 loop_commit_inserts ();
1299 /* Fills ALWAYS_EXECUTED_IN information for basic blocks of LOOP, i.e.
1300 for each such basic block bb records the outermost loop for that execution
1301 of its header implies execution of bb. CONTAINS_CALL is the bitmap of
1302 blocks that contain a nonpure call. */
1305 fill_always_executed_in (struct loop
*loop
, sbitmap contains_call
)
1307 basic_block bb
= NULL
, *bbs
, last
= NULL
;
1310 struct loop
*inn_loop
= loop
;
1312 if (!loop
->header
->aux
)
1314 bbs
= get_loop_body_in_dom_order (loop
);
1316 for (i
= 0; i
< loop
->num_nodes
; i
++)
1321 if (dominated_by_p (CDI_DOMINATORS
, loop
->latch
, bb
))
1324 if (TEST_BIT (contains_call
, bb
->index
))
1327 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1328 if (!flow_bb_inside_loop_p (loop
, e
->dest
))
1333 /* A loop might be infinite (TODO use simple loop analysis
1334 to disprove this if possible). */
1335 if (bb
->flags
& BB_IRREDUCIBLE_LOOP
)
1338 if (!flow_bb_inside_loop_p (inn_loop
, bb
))
1341 if (bb
->loop_father
->header
== bb
)
1343 if (!dominated_by_p (CDI_DOMINATORS
, loop
->latch
, bb
))
1346 /* In a loop that is always entered we may proceed anyway.
1347 But record that we entered it and stop once we leave it. */
1348 inn_loop
= bb
->loop_father
;
1355 if (last
== loop
->header
)
1357 last
= get_immediate_dominator (CDI_DOMINATORS
, last
);
1363 for (loop
= loop
->inner
; loop
; loop
= loop
->next
)
1364 fill_always_executed_in (loop
, contains_call
);
1367 /* Compute the global information needed by the loop invariant motion pass.
1368 LOOPS is the loop tree. */
1371 tree_ssa_lim_initialize (struct loops
*loops
)
1373 sbitmap contains_call
= sbitmap_alloc (last_basic_block
);
1374 block_stmt_iterator bsi
;
1378 sbitmap_zero (contains_call
);
1381 for (bsi
= bsi_start (bb
); !bsi_end_p (bsi
); bsi_next (&bsi
))
1383 if (nonpure_call_p (bsi_stmt (bsi
)))
1387 if (!bsi_end_p (bsi
))
1388 SET_BIT (contains_call
, bb
->index
);
1391 for (loop
= loops
->tree_root
->inner
; loop
; loop
= loop
->next
)
1392 fill_always_executed_in (loop
, contains_call
);
1394 sbitmap_free (contains_call
);
1397 /* Cleans up after the invariant motion pass. */
1400 tree_ssa_lim_finalize (void)
1410 /* Moves invariants from LOOPS. Only "expensive" invariants are moved out --
1411 i.e. those that are likely to be win regardless of the register pressure. */
1414 tree_ssa_lim (struct loops
*loops
)
1416 tree_ssa_lim_initialize (loops
);
1418 /* For each statement determine the outermost loop in that it is
1419 invariant and cost for computing the invariant. */
1420 determine_invariantness ();
1422 /* For each memory reference determine whether it is possible to hoist it
1423 out of the loop. Force the necessary invariants to be moved out of the
1425 determine_lsm (loops
);
1427 /* Move the expressions that are expensive enough. */
1428 move_computations ();
1430 tree_ssa_lim_finalize ();