1 /* Loop invariant motion.
2 Copyright (C) 2003 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
);
141 nxt
= &TREE_OPERAND (*addr_p
, 0);
142 return cbck (*addr_p
, nxt
, data
);
146 case VIEW_CONVERT_EXPR
:
147 case ARRAY_RANGE_REF
:
150 nxt
= &TREE_OPERAND (*addr_p
, 0);
154 nxt
= &TREE_OPERAND (*addr_p
, 0);
155 if (!cbck (*addr_p
, &TREE_OPERAND (*addr_p
, 1), data
))
171 /* If it is possible to hoist the statement STMT unconditionally,
172 returns MOVE_POSSIBLE.
173 If it is possible to hoist the statement STMT, but we must avoid making
174 it executed if it would not be executed in the original program (e.g.
175 because it may trap), return MOVE_PRESERVE_EXECUTION.
176 Otherwise return MOVE_IMPOSSIBLE. */
179 movement_possibility (tree stmt
)
183 if (flag_unswitch_loops
184 && TREE_CODE (stmt
) == COND_EXPR
)
186 /* If we perform unswitching, force the operands of the invariant
187 condition to be moved out of the loop. */
188 get_stmt_operands (stmt
);
190 return MOVE_POSSIBLE
;
193 if (TREE_CODE (stmt
) != MODIFY_EXPR
)
194 return MOVE_IMPOSSIBLE
;
196 if (stmt_ends_bb_p (stmt
))
197 return MOVE_IMPOSSIBLE
;
199 get_stmt_operands (stmt
);
201 if (stmt_ann (stmt
)->has_volatile_ops
)
202 return MOVE_IMPOSSIBLE
;
204 lhs
= TREE_OPERAND (stmt
, 0);
205 if (TREE_CODE (lhs
) == SSA_NAME
206 && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs
))
207 return MOVE_IMPOSSIBLE
;
209 rhs
= TREE_OPERAND (stmt
, 1);
211 if (TREE_SIDE_EFFECTS (rhs
))
212 return MOVE_IMPOSSIBLE
;
214 if (TREE_CODE (lhs
) != SSA_NAME
215 || tree_could_trap_p (rhs
))
216 return MOVE_PRESERVE_EXECUTION
;
218 return MOVE_POSSIBLE
;
221 /* Suppose that operand DEF is used inside the LOOP. Returns the outermost
222 loop to that we could move the expression using DEF if it did not have
223 other operands, i.e. the outermost loop enclosing LOOP in that the value
224 of DEF is invariant. */
227 outermost_invariant_loop (tree def
, struct loop
*loop
)
231 struct loop
*max_loop
;
233 if (TREE_CODE (def
) != SSA_NAME
)
234 return superloop_at_depth (loop
, 1);
236 def_stmt
= SSA_NAME_DEF_STMT (def
);
237 def_bb
= bb_for_stmt (def_stmt
);
239 return superloop_at_depth (loop
, 1);
241 max_loop
= find_common_loop (loop
, def_bb
->loop_father
);
243 if (LIM_DATA (def_stmt
) && LIM_DATA (def_stmt
)->max_loop
)
244 max_loop
= find_common_loop (max_loop
,
245 LIM_DATA (def_stmt
)->max_loop
->outer
);
246 if (max_loop
== loop
)
248 max_loop
= superloop_at_depth (loop
, max_loop
->depth
+ 1);
253 /* Returns the outermost superloop of LOOP in that the expression EXPR is
257 outermost_invariant_loop_expr (tree expr
, struct loop
*loop
)
259 char class = TREE_CODE_CLASS (TREE_CODE (expr
));
261 struct loop
*max_loop
= superloop_at_depth (loop
, 1), *aloop
;
263 if (TREE_CODE (expr
) == SSA_NAME
264 || TREE_CODE (expr
) == INTEGER_CST
265 || is_gimple_min_invariant (expr
))
266 return outermost_invariant_loop (expr
, loop
);
274 nops
= first_rtl_op (TREE_CODE (expr
));
275 for (i
= 0; i
< nops
; i
++)
277 aloop
= outermost_invariant_loop_expr (TREE_OPERAND (expr
, i
), loop
);
281 if (flow_loop_nested_p (max_loop
, aloop
))
288 /* DATA is a structure containing information associated with a statement
289 inside LOOP. DEF is one of the operands of this statement.
291 Find the outermost loop enclosing LOOP in that value of DEF is invariant
292 and record this in DATA->max_loop field. If DEF itself is defined inside
293 this loop as well (i.e. we need to hoist it out of the loop if we want
294 to hoist the statement represented by DATA), record the statement in that
295 DEF is defined to the DATA->depends list. Additionally if ADD_COST is true,
296 add the cost of the computation of DEF to the DATA->cost.
298 If DEF is not invariant in LOOP, return false. Otherwise return TRUE. */
301 add_dependency (tree def
, struct lim_aux_data
*data
, struct loop
*loop
,
304 tree def_stmt
= SSA_NAME_DEF_STMT (def
);
305 basic_block def_bb
= bb_for_stmt (def_stmt
);
306 struct loop
*max_loop
;
312 max_loop
= outermost_invariant_loop (def
, loop
);
316 if (flow_loop_nested_p (data
->max_loop
, max_loop
))
317 data
->max_loop
= max_loop
;
319 if (!LIM_DATA (def_stmt
))
323 /* Only add the cost if the statement defining DEF is inside LOOP,
324 i.e. if it is likely that by moving the invariants dependent
325 on it, we will be able to avoid creating a new register for
326 it (since it will be only used in these dependent invariants). */
327 && def_bb
->loop_father
== loop
)
328 data
->cost
+= LIM_DATA (def_stmt
)->cost
;
330 dep
= xmalloc (sizeof (struct depend
));
331 dep
->stmt
= def_stmt
;
332 dep
->next
= data
->depends
;
338 /* Returns an estimate for a cost of statement STMT. TODO -- the values here
339 are just ad-hoc constants. The estimates should be based on target-specific
343 stmt_cost (tree stmt
)
348 /* Always try to create possibilities for unswitching. */
349 if (TREE_CODE (stmt
) == COND_EXPR
)
350 return LIM_EXPENSIVE
;
352 lhs
= TREE_OPERAND (stmt
, 0);
353 rhs
= TREE_OPERAND (stmt
, 1);
355 /* Hoisting memory references out should almost surely be a win. */
356 if (!is_gimple_variable (lhs
))
358 if (is_gimple_addressable (rhs
) && !is_gimple_variable (rhs
))
361 switch (TREE_CODE (rhs
))
364 /* We should be hoisting calls if possible. */
366 /* Unless the call is a builtin_constant_p; this always folds to a
367 constant, so moving it is useless. */
368 rhs
= get_callee_fndecl (rhs
);
369 if (DECL_BUILT_IN (rhs
)
370 && DECL_FUNCTION_CODE (rhs
) == BUILT_IN_CONSTANT_P
)
386 /* Division and multiplication are usually expensive. */
397 /* Determine the outermost loop to that it is possible to hoist a statement
398 STMT and store it to LIM_DATA (STMT)->max_loop. To do this we determine
399 the outermost loop in that the value computed by STMT is invariant.
400 If MUST_PRESERVE_EXEC is true, additionally choose such a loop that
401 we preserve the fact whether STMT is executed. It also fills other related
402 information to LIM_DATA (STMT).
404 The function returns false if STMT cannot be hoisted outside of the loop it
405 is defined in, and true otherwise. */
408 determine_max_movement (tree stmt
, bool must_preserve_exec
)
410 basic_block bb
= bb_for_stmt (stmt
);
411 struct loop
*loop
= bb
->loop_father
;
413 struct lim_aux_data
*lim_data
= LIM_DATA (stmt
);
417 if (must_preserve_exec
)
418 level
= ALWAYS_EXECUTED_IN (bb
);
420 level
= superloop_at_depth (loop
, 1);
421 lim_data
->max_loop
= level
;
423 FOR_EACH_SSA_TREE_OPERAND (val
, stmt
, iter
, SSA_OP_USE
)
424 if (!add_dependency (val
, lim_data
, loop
, true))
427 FOR_EACH_SSA_TREE_OPERAND (val
, stmt
, iter
, SSA_OP_VIRTUAL_USES
)
428 if (!add_dependency (val
, lim_data
, loop
, false))
431 lim_data
->cost
+= stmt_cost (stmt
);
436 /* Suppose that some statement in ORIG_LOOP is hoisted to the loop LEVEL,
437 and that one of the operands of this statement is computed by STMT.
438 Ensure that STMT (together with all the statements that define its
439 operands) is hoisted at least out of the loop LEVEL. */
442 set_level (tree stmt
, struct loop
*orig_loop
, struct loop
*level
)
444 struct loop
*stmt_loop
= bb_for_stmt (stmt
)->loop_father
;
447 stmt_loop
= find_common_loop (orig_loop
, stmt_loop
);
448 if (LIM_DATA (stmt
) && LIM_DATA (stmt
)->tgt_loop
)
449 stmt_loop
= find_common_loop (stmt_loop
,
450 LIM_DATA (stmt
)->tgt_loop
->outer
);
451 if (flow_loop_nested_p (stmt_loop
, level
))
454 gcc_assert (LIM_DATA (stmt
));
455 gcc_assert (level
== LIM_DATA (stmt
)->max_loop
456 || flow_loop_nested_p (LIM_DATA (stmt
)->max_loop
, level
));
458 LIM_DATA (stmt
)->tgt_loop
= level
;
459 for (dep
= LIM_DATA (stmt
)->depends
; dep
; dep
= dep
->next
)
460 set_level (dep
->stmt
, orig_loop
, level
);
463 /* Determines an outermost loop from that we want to hoist the statement STMT.
464 For now we chose the outermost possible loop. TODO -- use profiling
465 information to set it more sanely. */
468 set_profitable_level (tree stmt
)
470 set_level (stmt
, bb_for_stmt (stmt
)->loop_father
, LIM_DATA (stmt
)->max_loop
);
473 /* Returns true if STMT is not a pure call. */
476 nonpure_call_p (tree stmt
)
478 tree call
= get_call_expr_in (stmt
);
483 return TREE_SIDE_EFFECTS (call
) != 0;
486 /* Releases the memory occupied by DATA. */
489 free_lim_aux_data (struct lim_aux_data
*data
)
491 struct depend
*dep
, *next
;
493 for (dep
= data
->depends
; dep
; dep
= next
)
501 /* Determine the outermost loops in that statements in basic block BB are
502 invariant, and record them to the LIM_DATA associated with the statements.
503 Callback for walk_dominator_tree. */
506 determine_invariantness_stmt (struct dom_walk_data
*dw_data ATTRIBUTE_UNUSED
,
510 block_stmt_iterator bsi
;
512 bool maybe_never
= ALWAYS_EXECUTED_IN (bb
) == NULL
;
513 struct loop
*outermost
= ALWAYS_EXECUTED_IN (bb
);
515 if (!bb
->loop_father
->outer
)
518 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
519 fprintf (dump_file
, "Basic block %d (loop %d -- depth %d):\n\n",
520 bb
->index
, bb
->loop_father
->num
, bb
->loop_father
->depth
);
522 for (bsi
= bsi_start (bb
); !bsi_end_p (bsi
); bsi_next (&bsi
))
524 stmt
= bsi_stmt (bsi
);
526 pos
= movement_possibility (stmt
);
527 if (pos
== MOVE_IMPOSSIBLE
)
529 if (nonpure_call_p (stmt
))
537 stmt_ann (stmt
)->common
.aux
= xcalloc (1, sizeof (struct lim_aux_data
));
538 LIM_DATA (stmt
)->always_executed_in
= outermost
;
540 if (maybe_never
&& pos
== MOVE_PRESERVE_EXECUTION
)
543 if (!determine_max_movement (stmt
, pos
== MOVE_PRESERVE_EXECUTION
))
545 LIM_DATA (stmt
)->max_loop
= NULL
;
549 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
551 print_generic_stmt_indented (dump_file
, stmt
, 0, 2);
552 fprintf (dump_file
, " invariant up to level %d, cost %d.\n\n",
553 LIM_DATA (stmt
)->max_loop
->depth
,
554 LIM_DATA (stmt
)->cost
);
557 if (LIM_DATA (stmt
)->cost
>= LIM_EXPENSIVE
)
558 set_profitable_level (stmt
);
562 /* For each statement determines the outermost loop in that it is invariant,
563 statements on whose motion it depends and the cost of the computation.
564 This information is stored to the LIM_DATA structure associated with
568 determine_invariantness (void)
570 struct dom_walk_data walk_data
;
572 memset (&walk_data
, 0, sizeof (struct dom_walk_data
));
573 walk_data
.before_dom_children_before_stmts
= determine_invariantness_stmt
;
575 init_walk_dominator_tree (&walk_data
);
576 walk_dominator_tree (&walk_data
, ENTRY_BLOCK_PTR
);
577 fini_walk_dominator_tree (&walk_data
);
580 /* Commits edge insertions and updates loop structures. */
583 loop_commit_inserts (void)
585 unsigned old_last_basic_block
, i
;
588 old_last_basic_block
= last_basic_block
;
589 bsi_commit_edge_inserts (NULL
);
590 for (i
= old_last_basic_block
; i
< (unsigned) last_basic_block
; i
++)
592 bb
= BASIC_BLOCK (i
);
594 find_common_loop (bb
->succ
->dest
->loop_father
,
595 bb
->pred
->src
->loop_father
));
599 /* Hoist the statements in basic block BB out of the loops prescribed by
600 data stored in LIM_DATA structures associated with each statement. Callback
601 for walk_dominator_tree. */
604 move_computations_stmt (struct dom_walk_data
*dw_data ATTRIBUTE_UNUSED
,
608 block_stmt_iterator bsi
;
612 if (!bb
->loop_father
->outer
)
615 for (bsi
= bsi_start (bb
); !bsi_end_p (bsi
); )
617 stmt
= bsi_stmt (bsi
);
619 if (!LIM_DATA (stmt
))
625 cost
= LIM_DATA (stmt
)->cost
;
626 level
= LIM_DATA (stmt
)->tgt_loop
;
627 free_lim_aux_data (LIM_DATA (stmt
));
628 stmt_ann (stmt
)->common
.aux
= NULL
;
636 /* We do not really want to move conditionals out of the loop; we just
637 placed it here to force its operands to be moved if necessary. */
638 if (TREE_CODE (stmt
) == COND_EXPR
)
641 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
643 fprintf (dump_file
, "Moving statement\n");
644 print_generic_stmt (dump_file
, stmt
, 0);
645 fprintf (dump_file
, "(cost %u) out of loop %d.\n\n",
648 bsi_insert_on_edge (loop_preheader_edge (level
), stmt
);
653 /* Hoist the statements out of the loops prescribed by data stored in
654 LIM_DATA structures associated with each statement.*/
657 move_computations (void)
659 struct dom_walk_data walk_data
;
661 memset (&walk_data
, 0, sizeof (struct dom_walk_data
));
662 walk_data
.before_dom_children_before_stmts
= move_computations_stmt
;
664 init_walk_dominator_tree (&walk_data
);
665 walk_dominator_tree (&walk_data
, ENTRY_BLOCK_PTR
);
666 fini_walk_dominator_tree (&walk_data
);
668 loop_commit_inserts ();
669 rewrite_into_ssa (false);
670 if (bitmap_first_set_bit (vars_to_rename
) >= 0)
672 /* The rewrite of ssa names may cause violation of loop closed ssa
673 form invariants. TODO -- avoid these rewrites completely.
674 Information in virtual phi nodes is sufficient for it. */
675 rewrite_into_loop_closed_ssa ();
677 bitmap_clear (vars_to_rename
);
680 /* Checks whether the statement defining variable *INDEX can be hoisted
681 out of the loop passed in DATA. Callback for for_each_index. */
684 may_move_till (tree ref
, tree
*index
, void *data
)
686 struct loop
*loop
= data
, *max_loop
;
688 /* If REF is an array reference, check also that the step and the lower
689 bound is invariant in LOOP. */
690 if (TREE_CODE (ref
) == ARRAY_REF
)
692 tree step
= array_ref_element_size (ref
);
693 tree lbound
= array_ref_low_bound (ref
);
695 max_loop
= outermost_invariant_loop_expr (step
, loop
);
699 max_loop
= outermost_invariant_loop_expr (lbound
, loop
);
704 max_loop
= outermost_invariant_loop (*index
, loop
);
711 /* Forces statements defining (invariant) SSA names in expression EXPR to be
712 moved out of the LOOP. ORIG_LOOP is the loop in that EXPR is used. */
715 force_move_till_expr (tree expr
, struct loop
*orig_loop
, struct loop
*loop
)
717 char class = TREE_CODE_CLASS (TREE_CODE (expr
));
720 if (TREE_CODE (expr
) == SSA_NAME
)
722 tree stmt
= SSA_NAME_DEF_STMT (expr
);
723 if (IS_EMPTY_STMT (stmt
))
726 set_level (stmt
, orig_loop
, loop
);
736 nops
= first_rtl_op (TREE_CODE (expr
));
737 for (i
= 0; i
< nops
; i
++)
738 force_move_till_expr (TREE_OPERAND (expr
, i
), orig_loop
, loop
);
741 /* Forces statement defining invariants in REF (and *INDEX) to be moved out of
742 the LOOP. The reference REF is used in the loop ORIG_LOOP. Callback for
748 struct loop
*orig_loop
;
752 force_move_till (tree ref
, tree
*index
, void *data
)
755 struct fmt_data
*fmt_data
= data
;
757 if (TREE_CODE (ref
) == ARRAY_REF
)
759 tree step
= array_ref_element_size (ref
);
760 tree lbound
= array_ref_low_bound (ref
);
762 force_move_till_expr (step
, fmt_data
->orig_loop
, fmt_data
->loop
);
763 force_move_till_expr (lbound
, fmt_data
->orig_loop
, fmt_data
->loop
);
766 if (TREE_CODE (*index
) != SSA_NAME
)
769 stmt
= SSA_NAME_DEF_STMT (*index
);
770 if (IS_EMPTY_STMT (stmt
))
773 set_level (stmt
, fmt_data
->orig_loop
, fmt_data
->loop
);
778 /* Records memory reference *REF (that occurs in statement STMT)
779 to the list MEM_REFS. */
782 record_mem_ref (struct mem_ref
**mem_refs
, tree stmt
, tree
*ref
)
784 struct mem_ref
*aref
= xmalloc (sizeof (struct mem_ref
));
789 aref
->next
= *mem_refs
;
793 /* Releases list of memory references MEM_REFS. */
796 free_mem_refs (struct mem_ref
*mem_refs
)
803 mem_refs
= mem_refs
->next
;
808 /* If VAR is defined in LOOP and the statement it is defined in does not belong
809 to the set SEEN, add the statement to QUEUE of length IN_QUEUE and
813 maybe_queue_var (tree var
, struct loop
*loop
,
814 sbitmap seen
, tree
*queue
, unsigned *in_queue
)
816 tree stmt
= SSA_NAME_DEF_STMT (var
);
817 basic_block def_bb
= bb_for_stmt (stmt
);
820 || !flow_bb_inside_loop_p (loop
, def_bb
)
821 || TEST_BIT (seen
, get_stmt_uid (stmt
)))
824 SET_BIT (seen
, get_stmt_uid (stmt
));
825 queue
[(*in_queue
)++] = stmt
;
828 /* If COMMON_REF is NULL, set COMMON_REF to *OP and return true.
829 Otherwise return true if the memory reference *OP is equal to COMMON_REF.
830 Record the reference OP to list MEM_REFS. STMT is the statement in that
831 the reference occurs. */
835 struct mem_ref
**mem_refs
;
841 fem_single_reachable_address (tree
*op
, void *data
)
843 struct sra_data
*sra_data
= data
;
845 if (sra_data
->common_ref
846 && !operand_equal_p (*op
, sra_data
->common_ref
, 0))
848 sra_data
->common_ref
= *op
;
850 record_mem_ref (sra_data
->mem_refs
, sra_data
->stmt
, op
);
854 /* Runs CALLBACK for each operand of STMT that is a memory reference. DATA
855 is passed to the CALLBACK as well. The traversal stops if CALLBACK
856 returns false, for_each_memref then returns false as well. Otherwise
857 for_each_memref returns true. */
860 for_each_memref (tree stmt
, bool (*callback
)(tree
*, void *), void *data
)
864 if (TREE_CODE (stmt
) == RETURN_EXPR
)
865 stmt
= TREE_OPERAND (stmt
, 1);
867 if (TREE_CODE (stmt
) == MODIFY_EXPR
)
869 op
= &TREE_OPERAND (stmt
, 0);
870 if (TREE_CODE (*op
) != SSA_NAME
871 && !callback (op
, data
))
874 op
= &TREE_OPERAND (stmt
, 1);
875 if (TREE_CODE (*op
) != SSA_NAME
876 && is_gimple_lvalue (*op
)
877 && !callback (op
, data
))
880 stmt
= TREE_OPERAND (stmt
, 1);
883 if (TREE_CODE (stmt
) == WITH_SIZE_EXPR
)
884 stmt
= TREE_OPERAND (stmt
, 0);
886 if (TREE_CODE (stmt
) == CALL_EXPR
)
890 for (args
= TREE_OPERAND (stmt
, 1); args
; args
= TREE_CHAIN (args
))
892 op
= &TREE_VALUE (args
);
894 if (TREE_CODE (*op
) != SSA_NAME
895 && is_gimple_lvalue (*op
)
896 && !callback (op
, data
))
904 /* Determine whether all memory references inside the LOOP that correspond
905 to virtual ssa names defined in statement STMT are equal.
906 If so, store the list of the references to MEM_REFS, and return one
907 of them. Otherwise store NULL to MEM_REFS and return NULL_TREE.
908 *SEEN_CALL_STMT is set to true if the virtual operands suggest
909 that the reference might be clobbered by a call inside the LOOP. */
912 single_reachable_address (struct loop
*loop
, tree stmt
,
913 struct mem_ref
**mem_refs
,
914 bool *seen_call_stmt
)
916 unsigned max_uid
= max_stmt_uid
+ num_ssa_names
;
917 tree
*queue
= xmalloc (sizeof (tree
) * max_uid
);
918 sbitmap seen
= sbitmap_alloc (max_uid
);
919 unsigned in_queue
= 1;
922 struct sra_data sra_data
;
930 sra_data
.mem_refs
= mem_refs
;
931 sra_data
.common_ref
= NULL_TREE
;
934 SET_BIT (seen
, get_stmt_uid (stmt
));
935 *seen_call_stmt
= false;
939 stmt
= queue
[--in_queue
];
940 sra_data
.stmt
= stmt
;
943 && LIM_DATA (stmt
)->sm_done
)
946 switch (TREE_CODE (stmt
))
951 if (!for_each_memref (stmt
, fem_single_reachable_address
,
955 /* If this is a function that may depend on the memory location,
956 record the fact. We cannot directly refuse call clobbered
957 operands here, since sra_data.common_ref did not have
959 call
= get_call_expr_in (stmt
);
961 && !(call_expr_flags (call
) & ECF_CONST
))
962 *seen_call_stmt
= true;
964 /* Traverse also definitions of the VUSES (there may be other
965 distinct from the one we used to get to this statement). */
966 FOR_EACH_SSA_TREE_OPERAND (val
, stmt
, iter
, SSA_OP_VIRTUAL_USES
)
967 maybe_queue_var (val
, loop
, seen
, queue
, &in_queue
);
972 for (i
= 0; i
< (unsigned) PHI_NUM_ARGS (stmt
); i
++)
973 maybe_queue_var (PHI_ARG_DEF (stmt
, i
), loop
,
974 seen
, queue
, &in_queue
);
981 /* Find uses of virtual names. */
982 df
= get_immediate_uses (stmt
);
983 n
= num_immediate_uses (df
);
985 for (i
= 0; i
< n
; i
++)
987 stmt
= immediate_use (df
, i
);
989 if (!flow_bb_inside_loop_p (loop
, bb_for_stmt (stmt
)))
992 if (TEST_BIT (seen
, get_stmt_uid (stmt
)))
994 SET_BIT (seen
, get_stmt_uid (stmt
));
996 queue
[in_queue
++] = stmt
;
1001 sbitmap_free (seen
);
1003 return sra_data
.common_ref
;
1006 free_mem_refs (*mem_refs
);
1009 sbitmap_free (seen
);
1014 /* Rewrites memory references in list MEM_REFS by variable TMP_VAR. */
1017 rewrite_mem_refs (tree tmp_var
, struct mem_ref
*mem_refs
)
1022 for (; mem_refs
; mem_refs
= mem_refs
->next
)
1024 FOR_EACH_SSA_TREE_OPERAND (var
, mem_refs
->stmt
, iter
,
1025 (SSA_OP_VIRTUAL_DEFS
| SSA_OP_VUSE
))
1027 var
= SSA_NAME_VAR (var
);
1028 bitmap_set_bit (vars_to_rename
, var_ann (var
)->uid
);
1031 *mem_refs
->ref
= tmp_var
;
1032 modify_stmt (mem_refs
->stmt
);
1036 /* Records request for store motion of memory reference REF from LOOP.
1037 MEM_REFS is the list of occurrences of the reference REF inside LOOP;
1038 these references are rewritten by a new temporary variable.
1039 Exits from the LOOP are stored in EXITS, there are N_EXITS of them.
1040 The initialization of the temporary variable is put to the preheader
1041 of the loop, and assignments to the reference from the temporary variable
1042 are emitted to exits. */
1045 schedule_sm (struct loop
*loop
, edge
*exits
, unsigned n_exits
, tree ref
,
1046 struct mem_ref
*mem_refs
)
1048 struct mem_ref
*aref
;
1052 struct fmt_data fmt_data
;
1054 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1056 fprintf (dump_file
, "Executing store motion of ");
1057 print_generic_expr (dump_file
, ref
, 0);
1058 fprintf (dump_file
, " from loop %d\n", loop
->num
);
1061 tmp_var
= make_rename_temp (TREE_TYPE (ref
), "lsm_tmp");
1063 fmt_data
.loop
= loop
;
1064 fmt_data
.orig_loop
= loop
;
1065 for_each_index (&ref
, force_move_till
, &fmt_data
);
1067 rewrite_mem_refs (tmp_var
, mem_refs
);
1068 for (aref
= mem_refs
; aref
; aref
= aref
->next
)
1069 if (LIM_DATA (aref
->stmt
))
1070 LIM_DATA (aref
->stmt
)->sm_done
= true;
1072 /* Emit the load & stores. */
1073 load
= build (MODIFY_EXPR
, void_type_node
, tmp_var
, ref
);
1074 get_stmt_ann (load
)->common
.aux
= xcalloc (1, sizeof (struct lim_aux_data
));
1075 LIM_DATA (load
)->max_loop
= loop
;
1076 LIM_DATA (load
)->tgt_loop
= loop
;
1078 /* Put this into the latch, so that we are sure it will be processed after
1079 all dependencies. */
1080 bsi_insert_on_edge (loop_latch_edge (loop
), load
);
1082 for (i
= 0; i
< n_exits
; i
++)
1084 store
= build (MODIFY_EXPR
, void_type_node
,
1085 unshare_expr (ref
), tmp_var
);
1086 bsi_insert_on_edge (exits
[i
], store
);
1090 /* Returns true if REF may be clobbered by calls. */
1093 is_call_clobbered_ref (tree ref
)
1097 base
= get_base_address (ref
);
1102 return is_call_clobbered (base
);
1104 if (TREE_CODE (base
) == INDIRECT_REF
)
1106 /* Check whether the alias tags associated with the pointer
1107 are call clobbered. */
1108 tree ptr
= TREE_OPERAND (base
, 0);
1109 struct ptr_info_def
*pi
= SSA_NAME_PTR_INFO (ptr
);
1110 tree nmt
= (pi
) ? pi
->name_mem_tag
: NULL_TREE
;
1111 tree tmt
= var_ann (SSA_NAME_VAR (ptr
))->type_mem_tag
;
1113 if ((nmt
&& is_call_clobbered (nmt
))
1114 || (tmt
&& is_call_clobbered (tmt
)))
1123 /* Determine whether all memory references inside LOOP corresponding to the
1124 virtual ssa name REG are equal to each other, and whether the address of
1125 this common reference can be hoisted outside of the loop. If this is true,
1126 prepare the statements that load the value of the memory reference to a
1127 temporary variable in the loop preheader, store it back on the loop exits,
1128 and replace all the references inside LOOP by this temporary variable.
1129 LOOP has N_EXITS stored in EXITS. */
1132 determine_lsm_reg (struct loop
*loop
, edge
*exits
, unsigned n_exits
, tree reg
)
1135 struct mem_ref
*mem_refs
, *aref
;
1136 struct loop
*must_exec
;
1139 if (is_gimple_reg (reg
))
1142 ref
= single_reachable_address (loop
, SSA_NAME_DEF_STMT (reg
), &mem_refs
,
1147 /* If we cannot create a ssa name for the result, give up. */
1148 if (!is_gimple_reg_type (TREE_TYPE (ref
))
1149 || TREE_THIS_VOLATILE (ref
))
1152 /* If there is a call that may use the location, give up as well. */
1154 && is_call_clobbered_ref (ref
))
1157 if (!for_each_index (&ref
, may_move_till
, loop
))
1160 if (tree_could_trap_p (ref
))
1162 /* If the memory access is unsafe (i.e. it might trap), ensure that some
1163 of the statements in that it occurs is always executed when the loop
1164 is entered. This way we know that by moving the load from the
1165 reference out of the loop we will not cause the error that would not
1168 TODO -- in fact we would like to check for anticipability of the
1169 reference, i.e. that on each path from loop entry to loop exit at
1170 least one of the statements containing the memory reference is
1173 for (aref
= mem_refs
; aref
; aref
= aref
->next
)
1175 if (!LIM_DATA (aref
->stmt
))
1178 must_exec
= LIM_DATA (aref
->stmt
)->always_executed_in
;
1182 if (must_exec
== loop
1183 || flow_loop_nested_p (must_exec
, loop
))
1191 schedule_sm (loop
, exits
, n_exits
, ref
, mem_refs
);
1194 free_mem_refs (mem_refs
);
1197 /* Checks whether LOOP (with N_EXITS exits stored in EXITS array) is suitable
1198 for a store motion optimization (i.e. whether we can insert statement
1202 loop_suitable_for_sm (struct loop
*loop ATTRIBUTE_UNUSED
, edge
*exits
,
1207 for (i
= 0; i
< n_exits
; i
++)
1208 if (exits
[i
]->flags
& EDGE_ABNORMAL
)
1214 /* Try to perform store motion for all memory references modified inside
1218 determine_lsm_loop (struct loop
*loop
)
1222 edge
*exits
= get_loop_exit_edges (loop
, &n_exits
);
1224 if (!loop_suitable_for_sm (loop
, exits
, n_exits
))
1230 for (phi
= phi_nodes (loop
->header
); phi
; phi
= TREE_CHAIN (phi
))
1231 determine_lsm_reg (loop
, exits
, n_exits
, PHI_RESULT (phi
));
1236 /* Try to perform store motion for all memory references modified inside
1240 determine_lsm (struct loops
*loops
)
1245 /* Create a UID for each statement in the function. Ordering of the
1246 UIDs is not important for this pass. */
1250 block_stmt_iterator bsi
;
1252 for (bsi
= bsi_start (bb
); !bsi_end_p (bsi
); bsi_next (&bsi
))
1253 stmt_ann (bsi_stmt (bsi
))->uid
= max_stmt_uid
++;
1256 compute_immediate_uses (TDFA_USE_VOPS
, NULL
);
1258 /* Pass the loops from the outermost. For each virtual operand loop phi node
1259 check whether all the references inside the loop correspond to a single
1260 address, and if so, move them. */
1262 loop
= loops
->tree_root
->inner
;
1265 determine_lsm_loop (loop
);
1275 if (loop
== loops
->tree_root
)
1278 loop_commit_inserts ();
1286 /* Fills ALWAYS_EXECUTED_IN information for basic blocks of LOOP, i.e.
1287 for each such basic block bb records the outermost loop for that execution
1288 of its header implies execution of bb. CONTAINS_CALL is the bitmap of
1289 blocks that contain a nonpure call. */
1292 fill_always_executed_in (struct loop
*loop
, sbitmap contains_call
)
1294 basic_block bb
= NULL
, *bbs
, last
= NULL
;
1297 struct loop
*inn_loop
= loop
;
1299 if (!loop
->header
->aux
)
1301 bbs
= get_loop_body_in_dom_order (loop
);
1303 for (i
= 0; i
< loop
->num_nodes
; i
++)
1307 if (dominated_by_p (CDI_DOMINATORS
, loop
->latch
, bb
))
1310 if (TEST_BIT (contains_call
, bb
->index
))
1313 for (e
= bb
->succ
; e
; e
= e
->succ_next
)
1314 if (!flow_bb_inside_loop_p (loop
, e
->dest
))
1319 /* A loop might be infinite (TODO use simple loop analysis
1320 to disprove this if possible). */
1321 if (bb
->flags
& BB_IRREDUCIBLE_LOOP
)
1324 if (!flow_bb_inside_loop_p (inn_loop
, bb
))
1327 if (bb
->loop_father
->header
== bb
)
1329 if (!dominated_by_p (CDI_DOMINATORS
, loop
->latch
, bb
))
1332 /* In a loop that is always entered we may proceed anyway.
1333 But record that we entered it and stop once we leave it. */
1334 inn_loop
= bb
->loop_father
;
1341 if (last
== loop
->header
)
1343 last
= get_immediate_dominator (CDI_DOMINATORS
, last
);
1349 for (loop
= loop
->inner
; loop
; loop
= loop
->next
)
1350 fill_always_executed_in (loop
, contains_call
);
1353 /* Compute the global information needed by the loop invariant motion pass.
1354 LOOPS is the loop tree. */
1357 tree_ssa_lim_initialize (struct loops
*loops
)
1359 sbitmap contains_call
= sbitmap_alloc (last_basic_block
);
1360 block_stmt_iterator bsi
;
1364 sbitmap_zero (contains_call
);
1367 for (bsi
= bsi_start (bb
); !bsi_end_p (bsi
); bsi_next (&bsi
))
1369 if (nonpure_call_p (bsi_stmt (bsi
)))
1373 if (!bsi_end_p (bsi
))
1374 SET_BIT (contains_call
, bb
->index
);
1377 for (loop
= loops
->tree_root
->inner
; loop
; loop
= loop
->next
)
1378 fill_always_executed_in (loop
, contains_call
);
1380 sbitmap_free (contains_call
);
1383 /* Cleans up after the invariant motion pass. */
1386 tree_ssa_lim_finalize (void)
1396 /* Moves invariants from LOOPS. Only "expensive" invariants are moved out --
1397 i.e. those that are likely to be win regardless of the register pressure. */
1400 tree_ssa_lim (struct loops
*loops
)
1402 tree_ssa_lim_initialize (loops
);
1404 /* For each statement determine the outermost loop in that it is
1405 invariant and cost for computing the invariant. */
1406 determine_invariantness ();
1408 /* For each memory reference determine whether it is possible to hoist it
1409 out of the loop. Force the necessary invariants to be moved out of the
1411 determine_lsm (loops
);
1413 /* Move the expressions that are expensive enough. */
1414 move_computations ();
1416 tree_ssa_lim_finalize ();