1 /* Loop manipulation code for GNU compiler.
2 Copyright (C) 2002-2013 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it under
7 the terms of the GNU General Public License as published by the Free
8 Software Foundation; either version 3, or (at your option) any later
11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
12 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"
25 #include "basic-block.h"
29 #include "gimple-iterator.h"
30 #include "tree-ssa-loop-manip.h"
33 static void copy_loops_to (struct loop
**, int,
35 static void loop_redirect_edge (edge
, basic_block
);
36 static void remove_bbs (basic_block
*, int);
37 static bool rpe_enum_p (const_basic_block
, const void *);
38 static int find_path (edge
, basic_block
**);
39 static void fix_loop_placements (struct loop
*, bool *);
40 static bool fix_bb_placement (basic_block
);
41 static void fix_bb_placements (basic_block
, bool *, bitmap
);
43 /* Checks whether basic block BB is dominated by DATA. */
45 rpe_enum_p (const_basic_block bb
, const void *data
)
47 return dominated_by_p (CDI_DOMINATORS
, bb
, (const_basic_block
) data
);
50 /* Remove basic blocks BBS. NBBS is the number of the basic blocks. */
53 remove_bbs (basic_block
*bbs
, int nbbs
)
57 for (i
= 0; i
< nbbs
; i
++)
58 delete_basic_block (bbs
[i
]);
61 /* Find path -- i.e. the basic blocks dominated by edge E and put them
62 into array BBS, that will be allocated large enough to contain them.
63 E->dest must have exactly one predecessor for this to work (it is
64 easy to achieve and we do not put it here because we do not want to
65 alter anything by this function). The number of basic blocks in the
68 find_path (edge e
, basic_block
**bbs
)
70 gcc_assert (EDGE_COUNT (e
->dest
->preds
) <= 1);
72 /* Find bbs in the path. */
73 *bbs
= XNEWVEC (basic_block
, n_basic_blocks
);
74 return dfs_enumerate_from (e
->dest
, 0, rpe_enum_p
, *bbs
,
75 n_basic_blocks
, e
->dest
);
78 /* Fix placement of basic block BB inside loop hierarchy --
79 Let L be a loop to that BB belongs. Then every successor of BB must either
80 1) belong to some superloop of loop L, or
81 2) be a header of loop K such that K->outer is superloop of L
82 Returns true if we had to move BB into other loop to enforce this condition,
83 false if the placement of BB was already correct (provided that placements
84 of its successors are correct). */
86 fix_bb_placement (basic_block bb
)
90 struct loop
*loop
= current_loops
->tree_root
, *act
;
92 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
94 if (e
->dest
== EXIT_BLOCK_PTR
)
97 act
= e
->dest
->loop_father
;
98 if (act
->header
== e
->dest
)
99 act
= loop_outer (act
);
101 if (flow_loop_nested_p (loop
, act
))
105 if (loop
== bb
->loop_father
)
108 remove_bb_from_loops (bb
);
109 add_bb_to_loop (bb
, loop
);
114 /* Fix placement of LOOP inside loop tree, i.e. find the innermost superloop
115 of LOOP to that leads at least one exit edge of LOOP, and set it
116 as the immediate superloop of LOOP. Return true if the immediate superloop
119 IRRED_INVALIDATED is set to true if a change in the loop structures might
120 invalidate the information about irreducible regions. */
123 fix_loop_placement (struct loop
*loop
, bool *irred_invalidated
)
127 vec
<edge
> exits
= get_loop_exit_edges (loop
);
128 struct loop
*father
= current_loops
->tree_root
, *act
;
131 FOR_EACH_VEC_ELT (exits
, i
, e
)
133 act
= find_common_loop (loop
, e
->dest
->loop_father
);
134 if (flow_loop_nested_p (father
, act
))
138 if (father
!= loop_outer (loop
))
140 for (act
= loop_outer (loop
); act
!= father
; act
= loop_outer (act
))
141 act
->num_nodes
-= loop
->num_nodes
;
142 flow_loop_tree_node_remove (loop
);
143 flow_loop_tree_node_add (father
, loop
);
145 /* The exit edges of LOOP no longer exits its original immediate
146 superloops; remove them from the appropriate exit lists. */
147 FOR_EACH_VEC_ELT (exits
, i
, e
)
149 /* We may need to recompute irreducible loops. */
150 if (e
->flags
& EDGE_IRREDUCIBLE_LOOP
)
151 *irred_invalidated
= true;
152 rescan_loop_exit (e
, false, false);
162 /* Fix placements of basic blocks inside loop hierarchy stored in loops; i.e.
163 enforce condition condition stated in description of fix_bb_placement. We
164 start from basic block FROM that had some of its successors removed, so that
165 his placement no longer has to be correct, and iteratively fix placement of
166 its predecessors that may change if placement of FROM changed. Also fix
167 placement of subloops of FROM->loop_father, that might also be altered due
168 to this change; the condition for them is similar, except that instead of
169 successors we consider edges coming out of the loops.
171 If the changes may invalidate the information about irreducible regions,
172 IRRED_INVALIDATED is set to true.
174 If LOOP_CLOSED_SSA_INVLIDATED is non-zero then all basic blocks with
175 changed loop_father are collected there. */
178 fix_bb_placements (basic_block from
,
179 bool *irred_invalidated
,
180 bitmap loop_closed_ssa_invalidated
)
183 basic_block
*queue
, *qtop
, *qbeg
, *qend
;
184 struct loop
*base_loop
, *target_loop
;
187 /* We pass through blocks back-reachable from FROM, testing whether some
188 of their successors moved to outer loop. It may be necessary to
189 iterate several times, but it is finite, as we stop unless we move
190 the basic block up the loop structure. The whole story is a bit
191 more complicated due to presence of subloops, those are moved using
192 fix_loop_placement. */
194 base_loop
= from
->loop_father
;
195 /* If we are already in the outermost loop, the basic blocks cannot be moved
196 outside of it. If FROM is the header of the base loop, it cannot be moved
197 outside of it, either. In both cases, we can end now. */
198 if (base_loop
== current_loops
->tree_root
199 || from
== base_loop
->header
)
202 in_queue
= sbitmap_alloc (last_basic_block
);
203 bitmap_clear (in_queue
);
204 bitmap_set_bit (in_queue
, from
->index
);
205 /* Prevent us from going out of the base_loop. */
206 bitmap_set_bit (in_queue
, base_loop
->header
->index
);
208 queue
= XNEWVEC (basic_block
, base_loop
->num_nodes
+ 1);
209 qtop
= queue
+ base_loop
->num_nodes
+ 1;
221 bitmap_clear_bit (in_queue
, from
->index
);
223 if (from
->loop_father
->header
== from
)
225 /* Subloop header, maybe move the loop upward. */
226 if (!fix_loop_placement (from
->loop_father
, irred_invalidated
))
228 target_loop
= loop_outer (from
->loop_father
);
229 if (loop_closed_ssa_invalidated
)
231 basic_block
*bbs
= get_loop_body (from
->loop_father
);
232 for (unsigned i
= 0; i
< from
->loop_father
->num_nodes
; ++i
)
233 bitmap_set_bit (loop_closed_ssa_invalidated
, bbs
[i
]->index
);
239 /* Ordinary basic block. */
240 if (!fix_bb_placement (from
))
242 target_loop
= from
->loop_father
;
243 if (loop_closed_ssa_invalidated
)
244 bitmap_set_bit (loop_closed_ssa_invalidated
, from
->index
);
247 FOR_EACH_EDGE (e
, ei
, from
->succs
)
249 if (e
->flags
& EDGE_IRREDUCIBLE_LOOP
)
250 *irred_invalidated
= true;
253 /* Something has changed, insert predecessors into queue. */
254 FOR_EACH_EDGE (e
, ei
, from
->preds
)
256 basic_block pred
= e
->src
;
259 if (e
->flags
& EDGE_IRREDUCIBLE_LOOP
)
260 *irred_invalidated
= true;
262 if (bitmap_bit_p (in_queue
, pred
->index
))
265 /* If it is subloop, then it either was not moved, or
266 the path up the loop tree from base_loop do not contain
268 nca
= find_common_loop (pred
->loop_father
, base_loop
);
269 if (pred
->loop_father
!= base_loop
271 || nca
!= pred
->loop_father
))
272 pred
= pred
->loop_father
->header
;
273 else if (!flow_loop_nested_p (target_loop
, pred
->loop_father
))
275 /* If PRED is already higher in the loop hierarchy than the
276 TARGET_LOOP to that we moved FROM, the change of the position
277 of FROM does not affect the position of PRED, so there is no
278 point in processing it. */
282 if (bitmap_bit_p (in_queue
, pred
->index
))
285 /* Schedule the basic block. */
290 bitmap_set_bit (in_queue
, pred
->index
);
297 /* Removes path beginning at edge E, i.e. remove basic blocks dominated by E
298 and update loop structures and dominators. Return true if we were able
299 to remove the path, false otherwise (and nothing is affected then). */
304 basic_block
*rem_bbs
, *bord_bbs
, from
, bb
;
305 vec
<basic_block
> dom_bbs
;
306 int i
, nrem
, n_bord_bbs
;
308 bool irred_invalidated
= false;
312 if (!can_remove_branch_p (e
))
315 /* Keep track of whether we need to update information about irreducible
316 regions. This is the case if the removed area is a part of the
317 irreducible region, or if the set of basic blocks that belong to a loop
318 that is inside an irreducible region is changed, or if such a loop is
320 if (e
->flags
& EDGE_IRREDUCIBLE_LOOP
)
321 irred_invalidated
= true;
323 /* We need to check whether basic blocks are dominated by the edge
324 e, but we only have basic block dominators. This is easy to
325 fix -- when e->dest has exactly one predecessor, this corresponds
326 to blocks dominated by e->dest, if not, split the edge. */
327 if (!single_pred_p (e
->dest
))
328 e
= single_pred_edge (split_edge (e
));
330 /* It may happen that by removing path we remove one or more loops
331 we belong to. In this case first unloop the loops, then proceed
332 normally. We may assume that e->dest is not a header of any loop,
333 as it now has exactly one predecessor. */
334 for (l
= e
->src
->loop_father
; loop_outer (l
); l
= f
)
337 if (dominated_by_p (CDI_DOMINATORS
, l
->latch
, e
->dest
))
338 unloop (l
, &irred_invalidated
, NULL
);
341 /* Identify the path. */
342 nrem
= find_path (e
, &rem_bbs
);
345 bord_bbs
= XNEWVEC (basic_block
, n_basic_blocks
);
346 seen
= sbitmap_alloc (last_basic_block
);
349 /* Find "border" hexes -- i.e. those with predecessor in removed path. */
350 for (i
= 0; i
< nrem
; i
++)
351 bitmap_set_bit (seen
, rem_bbs
[i
]->index
);
352 if (!irred_invalidated
)
353 FOR_EACH_EDGE (ae
, ei
, e
->src
->succs
)
354 if (ae
!= e
&& ae
->dest
!= EXIT_BLOCK_PTR
&& !bitmap_bit_p (seen
, ae
->dest
->index
)
355 && ae
->flags
& EDGE_IRREDUCIBLE_LOOP
)
357 irred_invalidated
= true;
361 for (i
= 0; i
< nrem
; i
++)
364 FOR_EACH_EDGE (ae
, ei
, rem_bbs
[i
]->succs
)
365 if (ae
->dest
!= EXIT_BLOCK_PTR
&& !bitmap_bit_p (seen
, ae
->dest
->index
))
367 bitmap_set_bit (seen
, ae
->dest
->index
);
368 bord_bbs
[n_bord_bbs
++] = ae
->dest
;
370 if (ae
->flags
& EDGE_IRREDUCIBLE_LOOP
)
371 irred_invalidated
= true;
375 /* Remove the path. */
380 /* Cancel loops contained in the path. */
381 for (i
= 0; i
< nrem
; i
++)
382 if (rem_bbs
[i
]->loop_father
->header
== rem_bbs
[i
])
383 cancel_loop_tree (rem_bbs
[i
]->loop_father
);
385 remove_bbs (rem_bbs
, nrem
);
388 /* Find blocks whose dominators may be affected. */
390 for (i
= 0; i
< n_bord_bbs
; i
++)
394 bb
= get_immediate_dominator (CDI_DOMINATORS
, bord_bbs
[i
]);
395 if (bitmap_bit_p (seen
, bb
->index
))
397 bitmap_set_bit (seen
, bb
->index
);
399 for (ldom
= first_dom_son (CDI_DOMINATORS
, bb
);
401 ldom
= next_dom_son (CDI_DOMINATORS
, ldom
))
402 if (!dominated_by_p (CDI_DOMINATORS
, from
, ldom
))
403 dom_bbs
.safe_push (ldom
);
408 /* Recount dominators. */
409 iterate_fix_dominators (CDI_DOMINATORS
, dom_bbs
, true);
413 /* Fix placements of basic blocks inside loops and the placement of
414 loops in the loop tree. */
415 fix_bb_placements (from
, &irred_invalidated
, NULL
);
416 fix_loop_placements (from
->loop_father
, &irred_invalidated
);
418 if (irred_invalidated
419 && loops_state_satisfies_p (LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS
))
420 mark_irreducible_loops ();
425 /* Creates place for a new LOOP in loops structure of FN. */
428 place_new_loop (struct function
*fn
, struct loop
*loop
)
430 loop
->num
= number_of_loops (fn
);
431 vec_safe_push (loops_for_fn (fn
)->larray
, loop
);
434 /* Given LOOP structure with filled header and latch, find the body of the
435 corresponding loop and add it to loops tree. Insert the LOOP as a son of
439 add_loop (struct loop
*loop
, struct loop
*outer
)
443 struct loop
*subloop
;
447 /* Add it to loop structure. */
448 place_new_loop (cfun
, loop
);
449 flow_loop_tree_node_add (outer
, loop
);
451 /* Find its nodes. */
452 bbs
= XNEWVEC (basic_block
, n_basic_blocks
);
453 n
= get_loop_body_with_size (loop
, bbs
, n_basic_blocks
);
455 for (i
= 0; i
< n
; i
++)
457 if (bbs
[i
]->loop_father
== outer
)
459 remove_bb_from_loops (bbs
[i
]);
460 add_bb_to_loop (bbs
[i
], loop
);
466 /* If we find a direct subloop of OUTER, move it to LOOP. */
467 subloop
= bbs
[i
]->loop_father
;
468 if (loop_outer (subloop
) == outer
469 && subloop
->header
== bbs
[i
])
471 flow_loop_tree_node_remove (subloop
);
472 flow_loop_tree_node_add (loop
, subloop
);
476 /* Update the information about loop exit edges. */
477 for (i
= 0; i
< n
; i
++)
479 FOR_EACH_EDGE (e
, ei
, bbs
[i
]->succs
)
481 rescan_loop_exit (e
, false, false);
488 /* Multiply all frequencies in LOOP by NUM/DEN. */
491 scale_loop_frequencies (struct loop
*loop
, int num
, int den
)
495 bbs
= get_loop_body (loop
);
496 scale_bbs_frequencies_int (bbs
, loop
->num_nodes
, num
, den
);
500 /* Multiply all frequencies in LOOP by SCALE/REG_BR_PROB_BASE.
501 If ITERATION_BOUND is non-zero, scale even further if loop is predicted
502 to iterate too many times. */
505 scale_loop_profile (struct loop
*loop
, int scale
, gcov_type iteration_bound
)
507 gcov_type iterations
= expected_loop_iterations_unbounded (loop
);
511 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
512 fprintf (dump_file
, ";; Scaling loop %i with scale %f, "
513 "bounding iterations to %i from guessed %i\n",
514 loop
->num
, (double)scale
/ REG_BR_PROB_BASE
,
515 (int)iteration_bound
, (int)iterations
);
517 /* See if loop is predicted to iterate too many times. */
518 if (iteration_bound
&& iterations
> 0
519 && apply_probability (iterations
, scale
) > iteration_bound
)
521 /* Fixing loop profile for different trip count is not trivial; the exit
522 probabilities has to be updated to match and frequencies propagated down
525 We fully update only the simple case of loop with single exit that is
526 either from the latch or BB just before latch and leads from BB with
527 simple conditional jump. This is OK for use in vectorizer. */
528 e
= single_exit (loop
);
533 gcov_type count_delta
;
535 FOR_EACH_EDGE (other_e
, ei
, e
->src
->succs
)
536 if (!(other_e
->flags
& (EDGE_ABNORMAL
| EDGE_FAKE
))
540 /* Probability of exit must be 1/iterations. */
541 freq_delta
= EDGE_FREQUENCY (e
);
542 e
->probability
= REG_BR_PROB_BASE
/ iteration_bound
;
543 other_e
->probability
= inverse_probability (e
->probability
);
544 freq_delta
-= EDGE_FREQUENCY (e
);
546 /* Adjust counts accordingly. */
547 count_delta
= e
->count
;
548 e
->count
= apply_probability (e
->src
->count
, e
->probability
);
549 other_e
->count
= apply_probability (e
->src
->count
, other_e
->probability
);
550 count_delta
-= e
->count
;
552 /* If latch exists, change its frequency and count, since we changed
553 probability of exit. Theoretically we should update everything from
554 source of exit edge to latch, but for vectorizer this is enough. */
556 && loop
->latch
!= e
->src
)
558 loop
->latch
->frequency
+= freq_delta
;
559 if (loop
->latch
->frequency
< 0)
560 loop
->latch
->frequency
= 0;
561 loop
->latch
->count
+= count_delta
;
562 if (loop
->latch
->count
< 0)
563 loop
->latch
->count
= 0;
567 /* Roughly speaking we want to reduce the loop body profile by the
568 the difference of loop iterations. We however can do better if
569 we look at the actual profile, if it is available. */
570 scale
= RDIV (iteration_bound
* scale
, iterations
);
571 if (loop
->header
->count
)
573 gcov_type count_in
= 0;
575 FOR_EACH_EDGE (e
, ei
, loop
->header
->preds
)
576 if (e
->src
!= loop
->latch
)
577 count_in
+= e
->count
;
580 scale
= GCOV_COMPUTE_SCALE (count_in
* iteration_bound
,
581 loop
->header
->count
);
583 else if (loop
->header
->frequency
)
587 FOR_EACH_EDGE (e
, ei
, loop
->header
->preds
)
588 if (e
->src
!= loop
->latch
)
589 freq_in
+= EDGE_FREQUENCY (e
);
592 scale
= GCOV_COMPUTE_SCALE (freq_in
* iteration_bound
,
593 loop
->header
->frequency
);
599 if (scale
== REG_BR_PROB_BASE
)
602 /* Scale the actual probabilities. */
603 scale_loop_frequencies (loop
, scale
, REG_BR_PROB_BASE
);
604 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
605 fprintf (dump_file
, ";; guessed iterations are now %i\n",
606 (int)expected_loop_iterations_unbounded (loop
));
609 /* Recompute dominance information for basic blocks outside LOOP. */
612 update_dominators_in_loop (struct loop
*loop
)
614 vec
<basic_block
> dom_bbs
= vNULL
;
619 seen
= sbitmap_alloc (last_basic_block
);
621 body
= get_loop_body (loop
);
623 for (i
= 0; i
< loop
->num_nodes
; i
++)
624 bitmap_set_bit (seen
, body
[i
]->index
);
626 for (i
= 0; i
< loop
->num_nodes
; i
++)
630 for (ldom
= first_dom_son (CDI_DOMINATORS
, body
[i
]);
632 ldom
= next_dom_son (CDI_DOMINATORS
, ldom
))
633 if (!bitmap_bit_p (seen
, ldom
->index
))
635 bitmap_set_bit (seen
, ldom
->index
);
636 dom_bbs
.safe_push (ldom
);
640 iterate_fix_dominators (CDI_DOMINATORS
, dom_bbs
, false);
646 /* Creates an if region as shown above. CONDITION is used to create
650 | ------------- -------------
651 | | pred_bb | | pred_bb |
652 | ------------- -------------
656 | | ====> -------------
661 | ------------- e_false / \ e_true
663 | ------------- ----------- -----------
664 | | false_bb | | true_bb |
665 | ----------- -----------
672 | | exit_edge (result)
681 create_empty_if_region_on_edge (edge entry_edge
, tree condition
)
684 basic_block cond_bb
, true_bb
, false_bb
, join_bb
;
685 edge e_true
, e_false
, exit_edge
;
688 gimple_stmt_iterator gsi
;
690 cond_bb
= split_edge (entry_edge
);
692 /* Insert condition in cond_bb. */
693 gsi
= gsi_last_bb (cond_bb
);
695 force_gimple_operand_gsi (&gsi
, condition
, true, NULL
,
696 false, GSI_NEW_STMT
);
697 cond_stmt
= gimple_build_cond_from_tree (simple_cond
, NULL_TREE
, NULL_TREE
);
698 gsi
= gsi_last_bb (cond_bb
);
699 gsi_insert_after (&gsi
, cond_stmt
, GSI_NEW_STMT
);
701 join_bb
= split_edge (single_succ_edge (cond_bb
));
703 e_true
= single_succ_edge (cond_bb
);
704 true_bb
= split_edge (e_true
);
706 e_false
= make_edge (cond_bb
, join_bb
, 0);
707 false_bb
= split_edge (e_false
);
709 e_true
->flags
&= ~EDGE_FALLTHRU
;
710 e_true
->flags
|= EDGE_TRUE_VALUE
;
711 e_false
->flags
&= ~EDGE_FALLTHRU
;
712 e_false
->flags
|= EDGE_FALSE_VALUE
;
714 set_immediate_dominator (CDI_DOMINATORS
, cond_bb
, entry_edge
->src
);
715 set_immediate_dominator (CDI_DOMINATORS
, true_bb
, cond_bb
);
716 set_immediate_dominator (CDI_DOMINATORS
, false_bb
, cond_bb
);
717 set_immediate_dominator (CDI_DOMINATORS
, join_bb
, cond_bb
);
719 exit_edge
= single_succ_edge (join_bb
);
721 if (single_pred_p (exit_edge
->dest
))
722 set_immediate_dominator (CDI_DOMINATORS
, exit_edge
->dest
, join_bb
);
727 /* create_empty_loop_on_edge
729 | - pred_bb - ------ pred_bb ------
730 | | | | iv0 = initial_value |
731 | -----|----- ---------|-----------
732 | | ______ | entry_edge
734 | | ====> | -V---V- loop_header -------------
735 | V | | iv_before = phi (iv0, iv_after) |
736 | - succ_bb - | ---|-----------------------------
738 | ----------- | ---V--- loop_body ---------------
739 | | | iv_after = iv_before + stride |
740 | | | if (iv_before < upper_bound) |
741 | | ---|--------------\--------------
744 | | - loop_latch - V- succ_bb -
746 | | /------------- -----------
749 Creates an empty loop as shown above, the IV_BEFORE is the SSA_NAME
750 that is used before the increment of IV. IV_BEFORE should be used for
751 adding code to the body that uses the IV. OUTER is the outer loop in
752 which the new loop should be inserted.
754 Both INITIAL_VALUE and UPPER_BOUND expressions are gimplified and
755 inserted on the loop entry edge. This implies that this function
756 should be used only when the UPPER_BOUND expression is a loop
760 create_empty_loop_on_edge (edge entry_edge
,
762 tree stride
, tree upper_bound
,
768 basic_block loop_header
, loop_latch
, succ_bb
, pred_bb
;
770 gimple_stmt_iterator gsi
;
777 gcc_assert (entry_edge
&& initial_value
&& stride
&& upper_bound
&& iv
);
779 /* Create header, latch and wire up the loop. */
780 pred_bb
= entry_edge
->src
;
781 loop_header
= split_edge (entry_edge
);
782 loop_latch
= split_edge (single_succ_edge (loop_header
));
783 succ_bb
= single_succ (loop_latch
);
784 make_edge (loop_header
, succ_bb
, 0);
785 redirect_edge_succ_nodup (single_succ_edge (loop_latch
), loop_header
);
787 /* Set immediate dominator information. */
788 set_immediate_dominator (CDI_DOMINATORS
, loop_header
, pred_bb
);
789 set_immediate_dominator (CDI_DOMINATORS
, loop_latch
, loop_header
);
790 set_immediate_dominator (CDI_DOMINATORS
, succ_bb
, loop_header
);
792 /* Initialize a loop structure and put it in a loop hierarchy. */
793 loop
= alloc_loop ();
794 loop
->header
= loop_header
;
795 loop
->latch
= loop_latch
;
796 add_loop (loop
, outer
);
798 /* TODO: Fix frequencies and counts. */
799 prob
= REG_BR_PROB_BASE
/ 2;
801 scale_loop_frequencies (loop
, REG_BR_PROB_BASE
- prob
, REG_BR_PROB_BASE
);
803 /* Update dominators. */
804 update_dominators_in_loop (loop
);
806 /* Modify edge flags. */
807 exit_e
= single_exit (loop
);
808 exit_e
->flags
= EDGE_LOOP_EXIT
| EDGE_FALSE_VALUE
;
809 single_pred_edge (loop_latch
)->flags
= EDGE_TRUE_VALUE
;
811 /* Construct IV code in loop. */
812 initial_value
= force_gimple_operand (initial_value
, &stmts
, true, iv
);
815 gsi_insert_seq_on_edge (loop_preheader_edge (loop
), stmts
);
816 gsi_commit_edge_inserts ();
819 upper_bound
= force_gimple_operand (upper_bound
, &stmts
, true, NULL
);
822 gsi_insert_seq_on_edge (loop_preheader_edge (loop
), stmts
);
823 gsi_commit_edge_inserts ();
826 gsi
= gsi_last_bb (loop_header
);
827 create_iv (initial_value
, stride
, iv
, loop
, &gsi
, false,
828 iv_before
, iv_after
);
830 /* Insert loop exit condition. */
831 cond_expr
= gimple_build_cond
832 (LT_EXPR
, *iv_before
, upper_bound
, NULL_TREE
, NULL_TREE
);
834 exit_test
= gimple_cond_lhs (cond_expr
);
835 exit_test
= force_gimple_operand_gsi (&gsi
, exit_test
, true, NULL
,
836 false, GSI_NEW_STMT
);
837 gimple_cond_set_lhs (cond_expr
, exit_test
);
838 gsi
= gsi_last_bb (exit_e
->src
);
839 gsi_insert_after (&gsi
, cond_expr
, GSI_NEW_STMT
);
841 split_block_after_labels (loop_header
);
846 /* Make area between HEADER_EDGE and LATCH_EDGE a loop by connecting
847 latch to header and update loop tree and dominators
848 accordingly. Everything between them plus LATCH_EDGE destination must
849 be dominated by HEADER_EDGE destination, and back-reachable from
850 LATCH_EDGE source. HEADER_EDGE is redirected to basic block SWITCH_BB,
851 FALSE_EDGE of SWITCH_BB to original destination of HEADER_EDGE and
852 TRUE_EDGE of SWITCH_BB to original destination of LATCH_EDGE.
853 Returns the newly created loop. Frequencies and counts in the new loop
854 are scaled by FALSE_SCALE and in the old one by TRUE_SCALE. */
857 loopify (edge latch_edge
, edge header_edge
,
858 basic_block switch_bb
, edge true_edge
, edge false_edge
,
859 bool redirect_all_edges
, unsigned true_scale
, unsigned false_scale
)
861 basic_block succ_bb
= latch_edge
->dest
;
862 basic_block pred_bb
= header_edge
->src
;
863 struct loop
*loop
= alloc_loop ();
864 struct loop
*outer
= loop_outer (succ_bb
->loop_father
);
870 loop
->header
= header_edge
->dest
;
871 loop
->latch
= latch_edge
->src
;
873 freq
= EDGE_FREQUENCY (header_edge
);
874 cnt
= header_edge
->count
;
876 /* Redirect edges. */
877 loop_redirect_edge (latch_edge
, loop
->header
);
878 loop_redirect_edge (true_edge
, succ_bb
);
880 /* During loop versioning, one of the switch_bb edge is already properly
881 set. Do not redirect it again unless redirect_all_edges is true. */
882 if (redirect_all_edges
)
884 loop_redirect_edge (header_edge
, switch_bb
);
885 loop_redirect_edge (false_edge
, loop
->header
);
887 /* Update dominators. */
888 set_immediate_dominator (CDI_DOMINATORS
, switch_bb
, pred_bb
);
889 set_immediate_dominator (CDI_DOMINATORS
, loop
->header
, switch_bb
);
892 set_immediate_dominator (CDI_DOMINATORS
, succ_bb
, switch_bb
);
894 /* Compute new loop. */
895 add_loop (loop
, outer
);
897 /* Add switch_bb to appropriate loop. */
898 if (switch_bb
->loop_father
)
899 remove_bb_from_loops (switch_bb
);
900 add_bb_to_loop (switch_bb
, outer
);
902 /* Fix frequencies. */
903 if (redirect_all_edges
)
905 switch_bb
->frequency
= freq
;
906 switch_bb
->count
= cnt
;
907 FOR_EACH_EDGE (e
, ei
, switch_bb
->succs
)
909 e
->count
= apply_probability (switch_bb
->count
, e
->probability
);
912 scale_loop_frequencies (loop
, false_scale
, REG_BR_PROB_BASE
);
913 scale_loop_frequencies (succ_bb
->loop_father
, true_scale
, REG_BR_PROB_BASE
);
914 update_dominators_in_loop (loop
);
919 /* Remove the latch edge of a LOOP and update loops to indicate that
920 the LOOP was removed. After this function, original loop latch will
921 have no successor, which caller is expected to fix somehow.
923 If this may cause the information about irreducible regions to become
924 invalid, IRRED_INVALIDATED is set to true.
926 LOOP_CLOSED_SSA_INVALIDATED, if non-NULL, is a bitmap where we store
927 basic blocks that had non-trivial update on their loop_father.*/
930 unloop (struct loop
*loop
, bool *irred_invalidated
,
931 bitmap loop_closed_ssa_invalidated
)
936 basic_block latch
= loop
->latch
;
939 if (loop_preheader_edge (loop
)->flags
& EDGE_IRREDUCIBLE_LOOP
)
940 *irred_invalidated
= true;
942 /* This is relatively straightforward. The dominators are unchanged, as
943 loop header dominates loop latch, so the only thing we have to care of
944 is the placement of loops and basic blocks inside the loop tree. We
945 move them all to the loop->outer, and then let fix_bb_placements do
948 body
= get_loop_body (loop
);
950 for (i
= 0; i
< n
; i
++)
951 if (body
[i
]->loop_father
== loop
)
953 remove_bb_from_loops (body
[i
]);
954 add_bb_to_loop (body
[i
], loop_outer (loop
));
961 flow_loop_tree_node_remove (ploop
);
962 flow_loop_tree_node_add (loop_outer (loop
), ploop
);
965 /* Remove the loop and free its data. */
968 remove_edge (single_succ_edge (latch
));
970 /* We do not pass IRRED_INVALIDATED to fix_bb_placements here, as even if
971 there is an irreducible region inside the cancelled loop, the flags will
973 fix_bb_placements (latch
, &dummy
, loop_closed_ssa_invalidated
);
976 /* Fix placement of superloops of LOOP inside loop tree, i.e. ensure that
977 condition stated in description of fix_loop_placement holds for them.
978 It is used in case when we removed some edges coming out of LOOP, which
979 may cause the right placement of LOOP inside loop tree to change.
981 IRRED_INVALIDATED is set to true if a change in the loop structures might
982 invalidate the information about irreducible regions. */
985 fix_loop_placements (struct loop
*loop
, bool *irred_invalidated
)
989 while (loop_outer (loop
))
991 outer
= loop_outer (loop
);
992 if (!fix_loop_placement (loop
, irred_invalidated
))
995 /* Changing the placement of a loop in the loop tree may alter the
996 validity of condition 2) of the description of fix_bb_placement
997 for its preheader, because the successor is the header and belongs
998 to the loop. So call fix_bb_placements to fix up the placement
999 of the preheader and (possibly) of its predecessors. */
1000 fix_bb_placements (loop_preheader_edge (loop
)->src
,
1001 irred_invalidated
, NULL
);
1006 /* Duplicate loop bounds and other information we store about
1007 the loop into its duplicate. */
1010 copy_loop_info (struct loop
*loop
, struct loop
*target
)
1012 gcc_checking_assert (!target
->any_upper_bound
&& !target
->any_estimate
);
1013 target
->any_upper_bound
= loop
->any_upper_bound
;
1014 target
->nb_iterations_upper_bound
= loop
->nb_iterations_upper_bound
;
1015 target
->any_estimate
= loop
->any_estimate
;
1016 target
->nb_iterations_estimate
= loop
->nb_iterations_estimate
;
1017 target
->estimate_state
= loop
->estimate_state
;
1020 /* Copies copy of LOOP as subloop of TARGET loop, placing newly
1021 created loop into loops structure. */
1023 duplicate_loop (struct loop
*loop
, struct loop
*target
)
1026 cloop
= alloc_loop ();
1027 place_new_loop (cfun
, cloop
);
1029 copy_loop_info (loop
, cloop
);
1031 /* Mark the new loop as copy of LOOP. */
1032 set_loop_copy (loop
, cloop
);
1034 /* Add it to target. */
1035 flow_loop_tree_node_add (target
, cloop
);
1040 /* Copies structure of subloops of LOOP into TARGET loop, placing
1041 newly created loops into loop tree. */
1043 duplicate_subloops (struct loop
*loop
, struct loop
*target
)
1045 struct loop
*aloop
, *cloop
;
1047 for (aloop
= loop
->inner
; aloop
; aloop
= aloop
->next
)
1049 cloop
= duplicate_loop (aloop
, target
);
1050 duplicate_subloops (aloop
, cloop
);
1054 /* Copies structure of subloops of N loops, stored in array COPIED_LOOPS,
1055 into TARGET loop, placing newly created loops into loop tree. */
1057 copy_loops_to (struct loop
**copied_loops
, int n
, struct loop
*target
)
1062 for (i
= 0; i
< n
; i
++)
1064 aloop
= duplicate_loop (copied_loops
[i
], target
);
1065 duplicate_subloops (copied_loops
[i
], aloop
);
1069 /* Redirects edge E to basic block DEST. */
1071 loop_redirect_edge (edge e
, basic_block dest
)
1073 if (e
->dest
== dest
)
1076 redirect_edge_and_branch_force (e
, dest
);
1079 /* Check whether LOOP's body can be duplicated. */
1081 can_duplicate_loop_p (const struct loop
*loop
)
1084 basic_block
*bbs
= get_loop_body (loop
);
1086 ret
= can_copy_bbs_p (bbs
, loop
->num_nodes
);
1092 /* Sets probability and count of edge E to zero. The probability and count
1093 is redistributed evenly to the remaining edges coming from E->src. */
1096 set_zero_probability (edge e
)
1098 basic_block bb
= e
->src
;
1100 edge ae
, last
= NULL
;
1101 unsigned n
= EDGE_COUNT (bb
->succs
);
1102 gcov_type cnt
= e
->count
, cnt1
;
1103 unsigned prob
= e
->probability
, prob1
;
1106 cnt1
= cnt
/ (n
- 1);
1107 prob1
= prob
/ (n
- 1);
1109 FOR_EACH_EDGE (ae
, ei
, bb
->succs
)
1114 ae
->probability
+= prob1
;
1119 /* Move the rest to one of the edges. */
1120 last
->probability
+= prob
% (n
- 1);
1121 last
->count
+= cnt
% (n
- 1);
1127 /* Duplicates body of LOOP to given edge E NDUPL times. Takes care of updating
1128 loop structure and dominators. E's destination must be LOOP header for
1129 this to work, i.e. it must be entry or latch edge of this loop; these are
1130 unique, as the loops must have preheaders for this function to work
1131 correctly (in case E is latch, the function unrolls the loop, if E is entry
1132 edge, it peels the loop). Store edges created by copying ORIG edge from
1133 copies corresponding to set bits in WONT_EXIT bitmap (bit 0 corresponds to
1134 original LOOP body, the other copies are numbered in order given by control
1135 flow through them) into TO_REMOVE array. Returns false if duplication is
1139 duplicate_loop_to_header_edge (struct loop
*loop
, edge e
,
1140 unsigned int ndupl
, sbitmap wont_exit
,
1141 edge orig
, vec
<edge
> *to_remove
,
1144 struct loop
*target
, *aloop
;
1145 struct loop
**orig_loops
;
1146 unsigned n_orig_loops
;
1147 basic_block header
= loop
->header
, latch
= loop
->latch
;
1148 basic_block
*new_bbs
, *bbs
, *first_active
;
1149 basic_block new_bb
, bb
, first_active_latch
= NULL
;
1150 edge ae
, latch_edge
;
1151 edge spec_edges
[2], new_spec_edges
[2];
1155 int is_latch
= (latch
== e
->src
);
1156 int scale_act
= 0, *scale_step
= NULL
, scale_main
= 0;
1157 int scale_after_exit
= 0;
1158 int p
, freq_in
, freq_le
, freq_out_orig
;
1159 int prob_pass_thru
, prob_pass_wont_exit
, prob_pass_main
;
1160 int add_irreducible_flag
;
1161 basic_block place_after
;
1162 bitmap bbs_to_scale
= NULL
;
1165 gcc_assert (e
->dest
== loop
->header
);
1166 gcc_assert (ndupl
> 0);
1170 /* Orig must be edge out of the loop. */
1171 gcc_assert (flow_bb_inside_loop_p (loop
, orig
->src
));
1172 gcc_assert (!flow_bb_inside_loop_p (loop
, orig
->dest
));
1175 n
= loop
->num_nodes
;
1176 bbs
= get_loop_body_in_dom_order (loop
);
1177 gcc_assert (bbs
[0] == loop
->header
);
1178 gcc_assert (bbs
[n
- 1] == loop
->latch
);
1180 /* Check whether duplication is possible. */
1181 if (!can_copy_bbs_p (bbs
, loop
->num_nodes
))
1186 new_bbs
= XNEWVEC (basic_block
, loop
->num_nodes
);
1188 /* In case we are doing loop peeling and the loop is in the middle of
1189 irreducible region, the peeled copies will be inside it too. */
1190 add_irreducible_flag
= e
->flags
& EDGE_IRREDUCIBLE_LOOP
;
1191 gcc_assert (!is_latch
|| !add_irreducible_flag
);
1193 /* Find edge from latch. */
1194 latch_edge
= loop_latch_edge (loop
);
1196 if (flags
& DLTHE_FLAG_UPDATE_FREQ
)
1198 /* Calculate coefficients by that we have to scale frequencies
1199 of duplicated loop bodies. */
1200 freq_in
= header
->frequency
;
1201 freq_le
= EDGE_FREQUENCY (latch_edge
);
1204 if (freq_in
< freq_le
)
1206 freq_out_orig
= orig
? EDGE_FREQUENCY (orig
) : freq_in
- freq_le
;
1207 if (freq_out_orig
> freq_in
- freq_le
)
1208 freq_out_orig
= freq_in
- freq_le
;
1209 prob_pass_thru
= RDIV (REG_BR_PROB_BASE
* freq_le
, freq_in
);
1210 prob_pass_wont_exit
=
1211 RDIV (REG_BR_PROB_BASE
* (freq_le
+ freq_out_orig
), freq_in
);
1214 && REG_BR_PROB_BASE
- orig
->probability
!= 0)
1216 /* The blocks that are dominated by a removed exit edge ORIG have
1217 frequencies scaled by this. */
1219 = GCOV_COMPUTE_SCALE (REG_BR_PROB_BASE
,
1220 REG_BR_PROB_BASE
- orig
->probability
);
1221 bbs_to_scale
= BITMAP_ALLOC (NULL
);
1222 for (i
= 0; i
< n
; i
++)
1224 if (bbs
[i
] != orig
->src
1225 && dominated_by_p (CDI_DOMINATORS
, bbs
[i
], orig
->src
))
1226 bitmap_set_bit (bbs_to_scale
, i
);
1230 scale_step
= XNEWVEC (int, ndupl
);
1232 for (i
= 1; i
<= ndupl
; i
++)
1233 scale_step
[i
- 1] = bitmap_bit_p (wont_exit
, i
)
1234 ? prob_pass_wont_exit
1237 /* Complete peeling is special as the probability of exit in last
1239 if (flags
& DLTHE_FLAG_COMPLETTE_PEEL
)
1241 int wanted_freq
= EDGE_FREQUENCY (e
);
1243 if (wanted_freq
> freq_in
)
1244 wanted_freq
= freq_in
;
1246 gcc_assert (!is_latch
);
1247 /* First copy has frequency of incoming edge. Each subsequent
1248 frequency should be reduced by prob_pass_wont_exit. Caller
1249 should've managed the flags so all except for original loop
1250 has won't exist set. */
1251 scale_act
= GCOV_COMPUTE_SCALE (wanted_freq
, freq_in
);
1252 /* Now simulate the duplication adjustments and compute header
1253 frequency of the last copy. */
1254 for (i
= 0; i
< ndupl
; i
++)
1255 wanted_freq
= combine_probabilities (wanted_freq
, scale_step
[i
]);
1256 scale_main
= GCOV_COMPUTE_SCALE (wanted_freq
, freq_in
);
1260 prob_pass_main
= bitmap_bit_p (wont_exit
, 0)
1261 ? prob_pass_wont_exit
1264 scale_main
= REG_BR_PROB_BASE
;
1265 for (i
= 0; i
< ndupl
; i
++)
1268 p
= combine_probabilities (p
, scale_step
[i
]);
1270 scale_main
= GCOV_COMPUTE_SCALE (REG_BR_PROB_BASE
, scale_main
);
1271 scale_act
= combine_probabilities (scale_main
, prob_pass_main
);
1275 scale_main
= REG_BR_PROB_BASE
;
1276 for (i
= 0; i
< ndupl
; i
++)
1277 scale_main
= combine_probabilities (scale_main
, scale_step
[i
]);
1278 scale_act
= REG_BR_PROB_BASE
- prob_pass_thru
;
1280 for (i
= 0; i
< ndupl
; i
++)
1281 gcc_assert (scale_step
[i
] >= 0 && scale_step
[i
] <= REG_BR_PROB_BASE
);
1282 gcc_assert (scale_main
>= 0 && scale_main
<= REG_BR_PROB_BASE
1283 && scale_act
>= 0 && scale_act
<= REG_BR_PROB_BASE
);
1286 /* Loop the new bbs will belong to. */
1287 target
= e
->src
->loop_father
;
1289 /* Original loops. */
1291 for (aloop
= loop
->inner
; aloop
; aloop
= aloop
->next
)
1293 orig_loops
= XNEWVEC (struct loop
*, n_orig_loops
);
1294 for (aloop
= loop
->inner
, i
= 0; aloop
; aloop
= aloop
->next
, i
++)
1295 orig_loops
[i
] = aloop
;
1297 set_loop_copy (loop
, target
);
1299 first_active
= XNEWVEC (basic_block
, n
);
1302 memcpy (first_active
, bbs
, n
* sizeof (basic_block
));
1303 first_active_latch
= latch
;
1306 spec_edges
[SE_ORIG
] = orig
;
1307 spec_edges
[SE_LATCH
] = latch_edge
;
1309 place_after
= e
->src
;
1310 for (j
= 0; j
< ndupl
; j
++)
1313 copy_loops_to (orig_loops
, n_orig_loops
, target
);
1316 copy_bbs (bbs
, n
, new_bbs
, spec_edges
, 2, new_spec_edges
, loop
,
1318 place_after
= new_spec_edges
[SE_LATCH
]->src
;
1320 if (flags
& DLTHE_RECORD_COPY_NUMBER
)
1321 for (i
= 0; i
< n
; i
++)
1323 gcc_assert (!new_bbs
[i
]->aux
);
1324 new_bbs
[i
]->aux
= (void *)(size_t)(j
+ 1);
1327 /* Note whether the blocks and edges belong to an irreducible loop. */
1328 if (add_irreducible_flag
)
1330 for (i
= 0; i
< n
; i
++)
1331 new_bbs
[i
]->flags
|= BB_DUPLICATED
;
1332 for (i
= 0; i
< n
; i
++)
1335 new_bb
= new_bbs
[i
];
1336 if (new_bb
->loop_father
== target
)
1337 new_bb
->flags
|= BB_IRREDUCIBLE_LOOP
;
1339 FOR_EACH_EDGE (ae
, ei
, new_bb
->succs
)
1340 if ((ae
->dest
->flags
& BB_DUPLICATED
)
1341 && (ae
->src
->loop_father
== target
1342 || ae
->dest
->loop_father
== target
))
1343 ae
->flags
|= EDGE_IRREDUCIBLE_LOOP
;
1345 for (i
= 0; i
< n
; i
++)
1346 new_bbs
[i
]->flags
&= ~BB_DUPLICATED
;
1349 /* Redirect the special edges. */
1352 redirect_edge_and_branch_force (latch_edge
, new_bbs
[0]);
1353 redirect_edge_and_branch_force (new_spec_edges
[SE_LATCH
],
1355 set_immediate_dominator (CDI_DOMINATORS
, new_bbs
[0], latch
);
1356 latch
= loop
->latch
= new_bbs
[n
- 1];
1357 e
= latch_edge
= new_spec_edges
[SE_LATCH
];
1361 redirect_edge_and_branch_force (new_spec_edges
[SE_LATCH
],
1363 redirect_edge_and_branch_force (e
, new_bbs
[0]);
1364 set_immediate_dominator (CDI_DOMINATORS
, new_bbs
[0], e
->src
);
1365 e
= new_spec_edges
[SE_LATCH
];
1368 /* Record exit edge in this copy. */
1369 if (orig
&& bitmap_bit_p (wont_exit
, j
+ 1))
1372 to_remove
->safe_push (new_spec_edges
[SE_ORIG
]);
1373 set_zero_probability (new_spec_edges
[SE_ORIG
]);
1375 /* Scale the frequencies of the blocks dominated by the exit. */
1378 EXECUTE_IF_SET_IN_BITMAP (bbs_to_scale
, 0, i
, bi
)
1380 scale_bbs_frequencies_int (new_bbs
+ i
, 1, scale_after_exit
,
1386 /* Record the first copy in the control flow order if it is not
1387 the original loop (i.e. in case of peeling). */
1388 if (!first_active_latch
)
1390 memcpy (first_active
, new_bbs
, n
* sizeof (basic_block
));
1391 first_active_latch
= new_bbs
[n
- 1];
1394 /* Set counts and frequencies. */
1395 if (flags
& DLTHE_FLAG_UPDATE_FREQ
)
1397 scale_bbs_frequencies_int (new_bbs
, n
, scale_act
, REG_BR_PROB_BASE
);
1398 scale_act
= combine_probabilities (scale_act
, scale_step
[j
]);
1404 /* Record the exit edge in the original loop body, and update the frequencies. */
1405 if (orig
&& bitmap_bit_p (wont_exit
, 0))
1408 to_remove
->safe_push (orig
);
1409 set_zero_probability (orig
);
1411 /* Scale the frequencies of the blocks dominated by the exit. */
1414 EXECUTE_IF_SET_IN_BITMAP (bbs_to_scale
, 0, i
, bi
)
1416 scale_bbs_frequencies_int (bbs
+ i
, 1, scale_after_exit
,
1422 /* Update the original loop. */
1424 set_immediate_dominator (CDI_DOMINATORS
, e
->dest
, e
->src
);
1425 if (flags
& DLTHE_FLAG_UPDATE_FREQ
)
1427 scale_bbs_frequencies_int (bbs
, n
, scale_main
, REG_BR_PROB_BASE
);
1431 /* Update dominators of outer blocks if affected. */
1432 for (i
= 0; i
< n
; i
++)
1434 basic_block dominated
, dom_bb
;
1435 vec
<basic_block
> dom_bbs
;
1441 dom_bbs
= get_dominated_by (CDI_DOMINATORS
, bb
);
1442 FOR_EACH_VEC_ELT (dom_bbs
, j
, dominated
)
1444 if (flow_bb_inside_loop_p (loop
, dominated
))
1446 dom_bb
= nearest_common_dominator (
1447 CDI_DOMINATORS
, first_active
[i
], first_active_latch
);
1448 set_immediate_dominator (CDI_DOMINATORS
, dominated
, dom_bb
);
1452 free (first_active
);
1455 BITMAP_FREE (bbs_to_scale
);
1460 /* A callback for make_forwarder block, to redirect all edges except for
1461 MFB_KJ_EDGE to the entry part. E is the edge for that we should decide
1462 whether to redirect it. */
1466 mfb_keep_just (edge e
)
1468 return e
!= mfb_kj_edge
;
1471 /* True when a candidate preheader BLOCK has predecessors from LOOP. */
1474 has_preds_from_loop (basic_block block
, struct loop
*loop
)
1479 FOR_EACH_EDGE (e
, ei
, block
->preds
)
1480 if (e
->src
->loop_father
== loop
)
1485 /* Creates a pre-header for a LOOP. Returns newly created block. Unless
1486 CP_SIMPLE_PREHEADERS is set in FLAGS, we only force LOOP to have single
1487 entry; otherwise we also force preheader block to have only one successor.
1488 When CP_FALLTHRU_PREHEADERS is set in FLAGS, we force the preheader block
1489 to be a fallthru predecessor to the loop header and to have only
1490 predecessors from outside of the loop.
1491 The function also updates dominators. */
1494 create_preheader (struct loop
*loop
, int flags
)
1500 bool latch_edge_was_fallthru
;
1501 edge one_succ_pred
= NULL
, single_entry
= NULL
;
1504 FOR_EACH_EDGE (e
, ei
, loop
->header
->preds
)
1506 if (e
->src
== loop
->latch
)
1508 irred
|= (e
->flags
& EDGE_IRREDUCIBLE_LOOP
) != 0;
1511 if (single_succ_p (e
->src
))
1514 gcc_assert (nentry
);
1517 bool need_forwarder_block
= false;
1519 /* We do not allow entry block to be the loop preheader, since we
1520 cannot emit code there. */
1521 if (single_entry
->src
== ENTRY_BLOCK_PTR
)
1522 need_forwarder_block
= true;
1525 /* If we want simple preheaders, also force the preheader to have
1526 just a single successor. */
1527 if ((flags
& CP_SIMPLE_PREHEADERS
)
1528 && !single_succ_p (single_entry
->src
))
1529 need_forwarder_block
= true;
1530 /* If we want fallthru preheaders, also create forwarder block when
1531 preheader ends with a jump or has predecessors from loop. */
1532 else if ((flags
& CP_FALLTHRU_PREHEADERS
)
1533 && (JUMP_P (BB_END (single_entry
->src
))
1534 || has_preds_from_loop (single_entry
->src
, loop
)))
1535 need_forwarder_block
= true;
1537 if (! need_forwarder_block
)
1541 mfb_kj_edge
= loop_latch_edge (loop
);
1542 latch_edge_was_fallthru
= (mfb_kj_edge
->flags
& EDGE_FALLTHRU
) != 0;
1543 fallthru
= make_forwarder_block (loop
->header
, mfb_keep_just
, NULL
);
1544 dummy
= fallthru
->src
;
1545 loop
->header
= fallthru
->dest
;
1547 /* Try to be clever in placing the newly created preheader. The idea is to
1548 avoid breaking any "fallthruness" relationship between blocks.
1550 The preheader was created just before the header and all incoming edges
1551 to the header were redirected to the preheader, except the latch edge.
1552 So the only problematic case is when this latch edge was a fallthru
1553 edge: it is not anymore after the preheader creation so we have broken
1554 the fallthruness. We're therefore going to look for a better place. */
1555 if (latch_edge_was_fallthru
)
1560 e
= EDGE_PRED (dummy
, 0);
1562 move_block_after (dummy
, e
->src
);
1567 dummy
->flags
|= BB_IRREDUCIBLE_LOOP
;
1568 single_succ_edge (dummy
)->flags
|= EDGE_IRREDUCIBLE_LOOP
;
1572 fprintf (dump_file
, "Created preheader block for loop %i\n",
1575 if (flags
& CP_FALLTHRU_PREHEADERS
)
1576 gcc_assert ((single_succ_edge (dummy
)->flags
& EDGE_FALLTHRU
)
1577 && !JUMP_P (BB_END (dummy
)));
1582 /* Create preheaders for each loop; for meaning of FLAGS see create_preheader. */
1585 create_preheaders (int flags
)
1593 FOR_EACH_LOOP (li
, loop
, 0)
1594 create_preheader (loop
, flags
);
1595 loops_state_set (LOOPS_HAVE_PREHEADERS
);
1598 /* Forces all loop latches to have only single successor. */
1601 force_single_succ_latches (void)
1607 FOR_EACH_LOOP (li
, loop
, 0)
1609 if (loop
->latch
!= loop
->header
&& single_succ_p (loop
->latch
))
1612 e
= find_edge (loop
->latch
, loop
->header
);
1613 gcc_checking_assert (e
!= NULL
);
1617 loops_state_set (LOOPS_HAVE_SIMPLE_LATCHES
);
1620 /* This function is called from loop_version. It splits the entry edge
1621 of the loop we want to version, adds the versioning condition, and
1622 adjust the edges to the two versions of the loop appropriately.
1623 e is an incoming edge. Returns the basic block containing the
1626 --- edge e ---- > [second_head]
1628 Split it and insert new conditional expression and adjust edges.
1630 --- edge e ---> [cond expr] ---> [first_head]
1632 +---------> [second_head]
1634 THEN_PROB is the probability of then branch of the condition. */
1637 lv_adjust_loop_entry_edge (basic_block first_head
, basic_block second_head
,
1638 edge e
, void *cond_expr
, unsigned then_prob
)
1640 basic_block new_head
= NULL
;
1643 gcc_assert (e
->dest
== second_head
);
1645 /* Split edge 'e'. This will create a new basic block, where we can
1646 insert conditional expr. */
1647 new_head
= split_edge (e
);
1649 lv_add_condition_to_bb (first_head
, second_head
, new_head
,
1652 /* Don't set EDGE_TRUE_VALUE in RTL mode, as it's invalid there. */
1653 e
= single_succ_edge (new_head
);
1654 e1
= make_edge (new_head
, first_head
,
1655 current_ir_type () == IR_GIMPLE
? EDGE_TRUE_VALUE
: 0);
1656 e1
->probability
= then_prob
;
1657 e
->probability
= REG_BR_PROB_BASE
- then_prob
;
1658 e1
->count
= apply_probability (e
->count
, e1
->probability
);
1659 e
->count
= apply_probability (e
->count
, e
->probability
);
1661 set_immediate_dominator (CDI_DOMINATORS
, first_head
, new_head
);
1662 set_immediate_dominator (CDI_DOMINATORS
, second_head
, new_head
);
1664 /* Adjust loop header phi nodes. */
1665 lv_adjust_loop_header_phi (first_head
, second_head
, new_head
, e1
);
1670 /* Main entry point for Loop Versioning transformation.
1672 This transformation given a condition and a loop, creates
1673 -if (condition) { loop_copy1 } else { loop_copy2 },
1674 where loop_copy1 is the loop transformed in one way, and loop_copy2
1675 is the loop transformed in another way (or unchanged). 'condition'
1676 may be a run time test for things that were not resolved by static
1677 analysis (overlapping ranges (anti-aliasing), alignment, etc.).
1679 THEN_PROB is the probability of the then edge of the if. THEN_SCALE
1680 is the ratio by that the frequencies in the original loop should
1681 be scaled. ELSE_SCALE is the ratio by that the frequencies in the
1682 new loop should be scaled.
1684 If PLACE_AFTER is true, we place the new loop after LOOP in the
1685 instruction stream, otherwise it is placed before LOOP. */
1688 loop_version (struct loop
*loop
,
1689 void *cond_expr
, basic_block
*condition_bb
,
1690 unsigned then_prob
, unsigned then_scale
, unsigned else_scale
,
1693 basic_block first_head
, second_head
;
1694 edge entry
, latch_edge
, true_edge
, false_edge
;
1697 basic_block cond_bb
;
1699 /* Record entry and latch edges for the loop */
1700 entry
= loop_preheader_edge (loop
);
1701 irred_flag
= entry
->flags
& EDGE_IRREDUCIBLE_LOOP
;
1702 entry
->flags
&= ~EDGE_IRREDUCIBLE_LOOP
;
1704 /* Note down head of loop as first_head. */
1705 first_head
= entry
->dest
;
1707 /* Duplicate loop. */
1708 if (!cfg_hook_duplicate_loop_to_header_edge (loop
, entry
, 1,
1709 NULL
, NULL
, NULL
, 0))
1711 entry
->flags
|= irred_flag
;
1715 /* After duplication entry edge now points to new loop head block.
1716 Note down new head as second_head. */
1717 second_head
= entry
->dest
;
1719 /* Split loop entry edge and insert new block with cond expr. */
1720 cond_bb
= lv_adjust_loop_entry_edge (first_head
, second_head
,
1721 entry
, cond_expr
, then_prob
);
1723 *condition_bb
= cond_bb
;
1727 entry
->flags
|= irred_flag
;
1731 latch_edge
= single_succ_edge (get_bb_copy (loop
->latch
));
1733 extract_cond_bb_edges (cond_bb
, &true_edge
, &false_edge
);
1734 nloop
= loopify (latch_edge
,
1735 single_pred_edge (get_bb_copy (loop
->header
)),
1736 cond_bb
, true_edge
, false_edge
,
1737 false /* Do not redirect all edges. */,
1738 then_scale
, else_scale
);
1740 copy_loop_info (loop
, nloop
);
1742 /* loopify redirected latch_edge. Update its PENDING_STMTS. */
1743 lv_flush_pending_stmts (latch_edge
);
1745 /* loopify redirected condition_bb's succ edge. Update its PENDING_STMTS. */
1746 extract_cond_bb_edges (cond_bb
, &true_edge
, &false_edge
);
1747 lv_flush_pending_stmts (false_edge
);
1748 /* Adjust irreducible flag. */
1751 cond_bb
->flags
|= BB_IRREDUCIBLE_LOOP
;
1752 loop_preheader_edge (loop
)->flags
|= EDGE_IRREDUCIBLE_LOOP
;
1753 loop_preheader_edge (nloop
)->flags
|= EDGE_IRREDUCIBLE_LOOP
;
1754 single_pred_edge (cond_bb
)->flags
|= EDGE_IRREDUCIBLE_LOOP
;
1759 basic_block
*bbs
= get_loop_body_in_dom_order (nloop
), after
;
1762 after
= loop
->latch
;
1764 for (i
= 0; i
< nloop
->num_nodes
; i
++)
1766 move_block_after (bbs
[i
], after
);
1772 /* At this point condition_bb is loop preheader with two successors,
1773 first_head and second_head. Make sure that loop preheader has only
1775 split_edge (loop_preheader_edge (loop
));
1776 split_edge (loop_preheader_edge (nloop
));