1 /* Loop manipulation code for GNU compiler.
2 Copyright (C) 2002-2017 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"
30 #include "gimple-iterator.h"
31 #include "gimplify-me.h"
32 #include "tree-ssa-loop-manip.h"
35 static void copy_loops_to (struct loop
**, int,
37 static void loop_redirect_edge (edge
, basic_block
);
38 static void remove_bbs (basic_block
*, int);
39 static bool rpe_enum_p (const_basic_block
, const void *);
40 static int find_path (edge
, basic_block
**);
41 static void fix_loop_placements (struct loop
*, bool *);
42 static bool fix_bb_placement (basic_block
);
43 static void fix_bb_placements (basic_block
, bool *, bitmap
);
45 /* Checks whether basic block BB is dominated by DATA. */
47 rpe_enum_p (const_basic_block bb
, const void *data
)
49 return dominated_by_p (CDI_DOMINATORS
, bb
, (const_basic_block
) data
);
52 /* Remove basic blocks BBS. NBBS is the number of the basic blocks. */
55 remove_bbs (basic_block
*bbs
, int nbbs
)
59 for (i
= 0; i
< nbbs
; i
++)
60 delete_basic_block (bbs
[i
]);
63 /* Find path -- i.e. the basic blocks dominated by edge E and put them
64 into array BBS, that will be allocated large enough to contain them.
65 E->dest must have exactly one predecessor for this to work (it is
66 easy to achieve and we do not put it here because we do not want to
67 alter anything by this function). The number of basic blocks in the
70 find_path (edge e
, basic_block
**bbs
)
72 gcc_assert (EDGE_COUNT (e
->dest
->preds
) <= 1);
74 /* Find bbs in the path. */
75 *bbs
= XNEWVEC (basic_block
, n_basic_blocks_for_fn (cfun
));
76 return dfs_enumerate_from (e
->dest
, 0, rpe_enum_p
, *bbs
,
77 n_basic_blocks_for_fn (cfun
), e
->dest
);
80 /* Fix placement of basic block BB inside loop hierarchy --
81 Let L be a loop to that BB belongs. Then every successor of BB must either
82 1) belong to some superloop of loop L, or
83 2) be a header of loop K such that K->outer is superloop of L
84 Returns true if we had to move BB into other loop to enforce this condition,
85 false if the placement of BB was already correct (provided that placements
86 of its successors are correct). */
88 fix_bb_placement (basic_block bb
)
92 struct loop
*loop
= current_loops
->tree_root
, *act
;
94 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
96 if (e
->dest
== EXIT_BLOCK_PTR_FOR_FN (cfun
))
99 act
= e
->dest
->loop_father
;
100 if (act
->header
== e
->dest
)
101 act
= loop_outer (act
);
103 if (flow_loop_nested_p (loop
, act
))
107 if (loop
== bb
->loop_father
)
110 remove_bb_from_loops (bb
);
111 add_bb_to_loop (bb
, loop
);
116 /* Fix placement of LOOP inside loop tree, i.e. find the innermost superloop
117 of LOOP to that leads at least one exit edge of LOOP, and set it
118 as the immediate superloop of LOOP. Return true if the immediate superloop
121 IRRED_INVALIDATED is set to true if a change in the loop structures might
122 invalidate the information about irreducible regions. */
125 fix_loop_placement (struct loop
*loop
, bool *irred_invalidated
)
129 vec
<edge
> exits
= get_loop_exit_edges (loop
);
130 struct loop
*father
= current_loops
->tree_root
, *act
;
133 FOR_EACH_VEC_ELT (exits
, i
, e
)
135 act
= find_common_loop (loop
, e
->dest
->loop_father
);
136 if (flow_loop_nested_p (father
, act
))
140 if (father
!= loop_outer (loop
))
142 for (act
= loop_outer (loop
); act
!= father
; act
= loop_outer (act
))
143 act
->num_nodes
-= loop
->num_nodes
;
144 flow_loop_tree_node_remove (loop
);
145 flow_loop_tree_node_add (father
, loop
);
147 /* The exit edges of LOOP no longer exits its original immediate
148 superloops; remove them from the appropriate exit lists. */
149 FOR_EACH_VEC_ELT (exits
, i
, e
)
151 /* We may need to recompute irreducible loops. */
152 if (e
->flags
& EDGE_IRREDUCIBLE_LOOP
)
153 *irred_invalidated
= true;
154 rescan_loop_exit (e
, false, false);
164 /* Fix placements of basic blocks inside loop hierarchy stored in loops; i.e.
165 enforce condition stated in description of fix_bb_placement. We
166 start from basic block FROM that had some of its successors removed, so that
167 his placement no longer has to be correct, and iteratively fix placement of
168 its predecessors that may change if placement of FROM changed. Also fix
169 placement of subloops of FROM->loop_father, that might also be altered due
170 to this change; the condition for them is similar, except that instead of
171 successors we consider edges coming out of the loops.
173 If the changes may invalidate the information about irreducible regions,
174 IRRED_INVALIDATED is set to true.
176 If LOOP_CLOSED_SSA_INVLIDATED is non-zero then all basic blocks with
177 changed loop_father are collected there. */
180 fix_bb_placements (basic_block from
,
181 bool *irred_invalidated
,
182 bitmap loop_closed_ssa_invalidated
)
184 basic_block
*queue
, *qtop
, *qbeg
, *qend
;
185 struct loop
*base_loop
, *target_loop
;
188 /* We pass through blocks back-reachable from FROM, testing whether some
189 of their successors moved to outer loop. It may be necessary to
190 iterate several times, but it is finite, as we stop unless we move
191 the basic block up the loop structure. The whole story is a bit
192 more complicated due to presence of subloops, those are moved using
193 fix_loop_placement. */
195 base_loop
= from
->loop_father
;
196 /* If we are already in the outermost loop, the basic blocks cannot be moved
197 outside of it. If FROM is the header of the base loop, it cannot be moved
198 outside of it, either. In both cases, we can end now. */
199 if (base_loop
== current_loops
->tree_root
200 || from
== base_loop
->header
)
203 auto_sbitmap
in_queue (last_basic_block_for_fn (cfun
));
204 bitmap_clear (in_queue
);
205 bitmap_set_bit (in_queue
, from
->index
);
206 /* Prevent us from going out of the base_loop. */
207 bitmap_set_bit (in_queue
, base_loop
->header
->index
);
209 queue
= XNEWVEC (basic_block
, base_loop
->num_nodes
+ 1);
210 qtop
= queue
+ base_loop
->num_nodes
+ 1;
222 bitmap_clear_bit (in_queue
, from
->index
);
224 if (from
->loop_father
->header
== from
)
226 /* Subloop header, maybe move the loop upward. */
227 if (!fix_loop_placement (from
->loop_father
, irred_invalidated
))
229 target_loop
= loop_outer (from
->loop_father
);
230 if (loop_closed_ssa_invalidated
)
232 basic_block
*bbs
= get_loop_body (from
->loop_father
);
233 for (unsigned i
= 0; i
< from
->loop_father
->num_nodes
; ++i
)
234 bitmap_set_bit (loop_closed_ssa_invalidated
, bbs
[i
]->index
);
240 /* Ordinary basic block. */
241 if (!fix_bb_placement (from
))
243 target_loop
= from
->loop_father
;
244 if (loop_closed_ssa_invalidated
)
245 bitmap_set_bit (loop_closed_ssa_invalidated
, from
->index
);
248 FOR_EACH_EDGE (e
, ei
, from
->succs
)
250 if (e
->flags
& EDGE_IRREDUCIBLE_LOOP
)
251 *irred_invalidated
= true;
254 /* Something has changed, insert predecessors into queue. */
255 FOR_EACH_EDGE (e
, ei
, from
->preds
)
257 basic_block pred
= e
->src
;
260 if (e
->flags
& EDGE_IRREDUCIBLE_LOOP
)
261 *irred_invalidated
= true;
263 if (bitmap_bit_p (in_queue
, pred
->index
))
266 /* If it is subloop, then it either was not moved, or
267 the path up the loop tree from base_loop do not contain
269 nca
= find_common_loop (pred
->loop_father
, base_loop
);
270 if (pred
->loop_father
!= base_loop
272 || nca
!= pred
->loop_father
))
273 pred
= pred
->loop_father
->header
;
274 else if (!flow_loop_nested_p (target_loop
, pred
->loop_father
))
276 /* If PRED is already higher in the loop hierarchy than the
277 TARGET_LOOP to that we moved FROM, the change of the position
278 of FROM does not affect the position of PRED, so there is no
279 point in processing it. */
283 if (bitmap_bit_p (in_queue
, pred
->index
))
286 /* Schedule the basic block. */
291 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). */
301 remove_path (edge e
, bool *irred_invalidated
,
302 bitmap loop_closed_ssa_invalidated
)
305 basic_block
*rem_bbs
, *bord_bbs
, from
, bb
;
306 vec
<basic_block
> dom_bbs
;
307 int i
, nrem
, n_bord_bbs
;
308 bool local_irred_invalidated
= false;
312 if (! irred_invalidated
)
313 irred_invalidated
= &local_irred_invalidated
;
315 if (!can_remove_branch_p (e
))
318 /* Keep track of whether we need to update information about irreducible
319 regions. This is the case if the removed area is a part of the
320 irreducible region, or if the set of basic blocks that belong to a loop
321 that is inside an irreducible region is changed, or if such a loop is
323 if (e
->flags
& EDGE_IRREDUCIBLE_LOOP
)
324 *irred_invalidated
= true;
326 /* We need to check whether basic blocks are dominated by the edge
327 e, but we only have basic block dominators. This is easy to
328 fix -- when e->dest has exactly one predecessor, this corresponds
329 to blocks dominated by e->dest, if not, split the edge. */
330 if (!single_pred_p (e
->dest
))
331 e
= single_pred_edge (split_edge (e
));
333 /* It may happen that by removing path we remove one or more loops
334 we belong to. In this case first unloop the loops, then proceed
335 normally. We may assume that e->dest is not a header of any loop,
336 as it now has exactly one predecessor. */
337 for (l
= e
->src
->loop_father
; loop_outer (l
); l
= f
)
340 if (dominated_by_p (CDI_DOMINATORS
, l
->latch
, e
->dest
))
341 unloop (l
, irred_invalidated
, loop_closed_ssa_invalidated
);
344 /* Identify the path. */
345 nrem
= find_path (e
, &rem_bbs
);
348 bord_bbs
= XNEWVEC (basic_block
, n_basic_blocks_for_fn (cfun
));
349 auto_sbitmap
seen (last_basic_block_for_fn (cfun
));
352 /* Find "border" hexes -- i.e. those with predecessor in removed path. */
353 for (i
= 0; i
< nrem
; i
++)
354 bitmap_set_bit (seen
, rem_bbs
[i
]->index
);
355 if (!*irred_invalidated
)
356 FOR_EACH_EDGE (ae
, ei
, e
->src
->succs
)
357 if (ae
!= e
&& ae
->dest
!= EXIT_BLOCK_PTR_FOR_FN (cfun
)
358 && !bitmap_bit_p (seen
, ae
->dest
->index
)
359 && ae
->flags
& EDGE_IRREDUCIBLE_LOOP
)
361 *irred_invalidated
= true;
365 for (i
= 0; i
< nrem
; i
++)
368 FOR_EACH_EDGE (ae
, ei
, rem_bbs
[i
]->succs
)
369 if (ae
->dest
!= EXIT_BLOCK_PTR_FOR_FN (cfun
)
370 && !bitmap_bit_p (seen
, ae
->dest
->index
))
372 bitmap_set_bit (seen
, ae
->dest
->index
);
373 bord_bbs
[n_bord_bbs
++] = ae
->dest
;
375 if (ae
->flags
& EDGE_IRREDUCIBLE_LOOP
)
376 *irred_invalidated
= true;
380 /* Remove the path. */
385 /* Cancel loops contained in the path. */
386 for (i
= 0; i
< nrem
; i
++)
387 if (rem_bbs
[i
]->loop_father
->header
== rem_bbs
[i
])
388 cancel_loop_tree (rem_bbs
[i
]->loop_father
);
390 remove_bbs (rem_bbs
, nrem
);
393 /* Find blocks whose dominators may be affected. */
395 for (i
= 0; i
< n_bord_bbs
; i
++)
399 bb
= get_immediate_dominator (CDI_DOMINATORS
, bord_bbs
[i
]);
400 if (bitmap_bit_p (seen
, bb
->index
))
402 bitmap_set_bit (seen
, bb
->index
);
404 for (ldom
= first_dom_son (CDI_DOMINATORS
, bb
);
406 ldom
= next_dom_son (CDI_DOMINATORS
, ldom
))
407 if (!dominated_by_p (CDI_DOMINATORS
, from
, ldom
))
408 dom_bbs
.safe_push (ldom
);
411 /* Recount dominators. */
412 iterate_fix_dominators (CDI_DOMINATORS
, dom_bbs
, true);
416 /* Fix placements of basic blocks inside loops and the placement of
417 loops in the loop tree. */
418 fix_bb_placements (from
, irred_invalidated
, loop_closed_ssa_invalidated
);
419 fix_loop_placements (from
->loop_father
, irred_invalidated
);
421 if (local_irred_invalidated
422 && loops_state_satisfies_p (LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS
))
423 mark_irreducible_loops ();
428 /* Creates place for a new LOOP in loops structure of FN. */
431 place_new_loop (struct function
*fn
, struct loop
*loop
)
433 loop
->num
= number_of_loops (fn
);
434 vec_safe_push (loops_for_fn (fn
)->larray
, loop
);
437 /* Given LOOP structure with filled header and latch, find the body of the
438 corresponding loop and add it to loops tree. Insert the LOOP as a son of
442 add_loop (struct loop
*loop
, struct loop
*outer
)
446 struct loop
*subloop
;
450 /* Add it to loop structure. */
451 place_new_loop (cfun
, loop
);
452 flow_loop_tree_node_add (outer
, loop
);
454 /* Find its nodes. */
455 bbs
= XNEWVEC (basic_block
, n_basic_blocks_for_fn (cfun
));
456 n
= get_loop_body_with_size (loop
, bbs
, n_basic_blocks_for_fn (cfun
));
458 for (i
= 0; i
< n
; i
++)
460 if (bbs
[i
]->loop_father
== outer
)
462 remove_bb_from_loops (bbs
[i
]);
463 add_bb_to_loop (bbs
[i
], loop
);
469 /* If we find a direct subloop of OUTER, move it to LOOP. */
470 subloop
= bbs
[i
]->loop_father
;
471 if (loop_outer (subloop
) == outer
472 && subloop
->header
== bbs
[i
])
474 flow_loop_tree_node_remove (subloop
);
475 flow_loop_tree_node_add (loop
, subloop
);
479 /* Update the information about loop exit edges. */
480 for (i
= 0; i
< n
; i
++)
482 FOR_EACH_EDGE (e
, ei
, bbs
[i
]->succs
)
484 rescan_loop_exit (e
, false, false);
491 /* Multiply all frequencies in LOOP by NUM/DEN. */
494 scale_loop_frequencies (struct loop
*loop
, int num
, int den
)
498 bbs
= get_loop_body (loop
);
499 scale_bbs_frequencies_int (bbs
, loop
->num_nodes
, num
, den
);
503 /* Multiply all frequencies in LOOP by SCALE/REG_BR_PROB_BASE.
504 If ITERATION_BOUND is non-zero, scale even further if loop is predicted
505 to iterate too many times. */
508 scale_loop_profile (struct loop
*loop
, int scale
, gcov_type iteration_bound
)
510 gcov_type iterations
= expected_loop_iterations_unbounded (loop
);
514 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
515 fprintf (dump_file
, ";; Scaling loop %i with scale %f, "
516 "bounding iterations to %i from guessed %i\n",
517 loop
->num
, (double)scale
/ REG_BR_PROB_BASE
,
518 (int)iteration_bound
, (int)iterations
);
520 /* See if loop is predicted to iterate too many times. */
521 if (iteration_bound
&& iterations
> 0
522 && apply_probability (iterations
, scale
) > iteration_bound
)
524 /* Fixing loop profile for different trip count is not trivial; the exit
525 probabilities has to be updated to match and frequencies propagated down
528 We fully update only the simple case of loop with single exit that is
529 either from the latch or BB just before latch and leads from BB with
530 simple conditional jump. This is OK for use in vectorizer. */
531 e
= single_exit (loop
);
536 profile_count count_delta
;
538 FOR_EACH_EDGE (other_e
, ei
, e
->src
->succs
)
539 if (!(other_e
->flags
& (EDGE_ABNORMAL
| EDGE_FAKE
))
543 /* Probability of exit must be 1/iterations. */
544 freq_delta
= EDGE_FREQUENCY (e
);
545 e
->probability
= REG_BR_PROB_BASE
/ iteration_bound
;
546 other_e
->probability
= inverse_probability (e
->probability
);
547 freq_delta
-= EDGE_FREQUENCY (e
);
549 /* Adjust counts accordingly. */
550 count_delta
= e
->count
;
551 e
->count
= e
->src
->count
.apply_probability (e
->probability
);
552 other_e
->count
= e
->src
->count
.apply_probability (other_e
->probability
);
553 count_delta
-= e
->count
;
555 /* If latch exists, change its frequency and count, since we changed
556 probability of exit. Theoretically we should update everything from
557 source of exit edge to latch, but for vectorizer this is enough. */
559 && loop
->latch
!= e
->src
)
561 loop
->latch
->frequency
+= freq_delta
;
562 if (loop
->latch
->frequency
< 0)
563 loop
->latch
->frequency
= 0;
564 loop
->latch
->count
+= count_delta
;
568 /* Roughly speaking we want to reduce the loop body profile by the
569 difference of loop iterations. We however can do better if
570 we look at the actual profile, if it is available. */
571 scale
= RDIV (iteration_bound
* scale
, iterations
);
573 bool determined
= false;
574 if (loop
->header
->count
.initialized_p ())
576 profile_count count_in
= profile_count::zero ();
578 FOR_EACH_EDGE (e
, ei
, loop
->header
->preds
)
579 if (e
->src
!= loop
->latch
)
580 count_in
+= e
->count
;
582 if (count_in
> profile_count::zero () )
584 scale
= GCOV_COMPUTE_SCALE (count_in
.to_gcov_type ()
586 loop
->header
->count
.to_gcov_type ());
594 FOR_EACH_EDGE (e
, ei
, loop
->header
->preds
)
595 if (e
->src
!= loop
->latch
)
596 freq_in
+= EDGE_FREQUENCY (e
);
599 scale
= GCOV_COMPUTE_SCALE (freq_in
* iteration_bound
,
600 loop
->header
->frequency
);
606 if (scale
== REG_BR_PROB_BASE
)
609 /* Scale the actual probabilities. */
610 scale_loop_frequencies (loop
, scale
, REG_BR_PROB_BASE
);
611 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
612 fprintf (dump_file
, ";; guessed iterations are now %i\n",
613 (int)expected_loop_iterations_unbounded (loop
));
616 /* Recompute dominance information for basic blocks outside LOOP. */
619 update_dominators_in_loop (struct loop
*loop
)
621 vec
<basic_block
> dom_bbs
= vNULL
;
625 auto_sbitmap
seen (last_basic_block_for_fn (cfun
));
627 body
= get_loop_body (loop
);
629 for (i
= 0; i
< loop
->num_nodes
; i
++)
630 bitmap_set_bit (seen
, body
[i
]->index
);
632 for (i
= 0; i
< loop
->num_nodes
; i
++)
636 for (ldom
= first_dom_son (CDI_DOMINATORS
, body
[i
]);
638 ldom
= next_dom_son (CDI_DOMINATORS
, ldom
))
639 if (!bitmap_bit_p (seen
, ldom
->index
))
641 bitmap_set_bit (seen
, ldom
->index
);
642 dom_bbs
.safe_push (ldom
);
646 iterate_fix_dominators (CDI_DOMINATORS
, dom_bbs
, false);
651 /* Creates an if region as shown above. CONDITION is used to create
655 | ------------- -------------
656 | | pred_bb | | pred_bb |
657 | ------------- -------------
661 | | ====> -------------
666 | ------------- e_false / \ e_true
668 | ------------- ----------- -----------
669 | | false_bb | | true_bb |
670 | ----------- -----------
677 | | exit_edge (result)
686 create_empty_if_region_on_edge (edge entry_edge
, tree condition
)
689 basic_block cond_bb
, true_bb
, false_bb
, join_bb
;
690 edge e_true
, e_false
, exit_edge
;
693 gimple_stmt_iterator gsi
;
695 cond_bb
= split_edge (entry_edge
);
697 /* Insert condition in cond_bb. */
698 gsi
= gsi_last_bb (cond_bb
);
700 force_gimple_operand_gsi (&gsi
, condition
, true, NULL
,
701 false, GSI_NEW_STMT
);
702 cond_stmt
= gimple_build_cond_from_tree (simple_cond
, NULL_TREE
, NULL_TREE
);
703 gsi
= gsi_last_bb (cond_bb
);
704 gsi_insert_after (&gsi
, cond_stmt
, GSI_NEW_STMT
);
706 join_bb
= split_edge (single_succ_edge (cond_bb
));
708 e_true
= single_succ_edge (cond_bb
);
709 true_bb
= split_edge (e_true
);
711 e_false
= make_edge (cond_bb
, join_bb
, 0);
712 false_bb
= split_edge (e_false
);
714 e_true
->flags
&= ~EDGE_FALLTHRU
;
715 e_true
->flags
|= EDGE_TRUE_VALUE
;
716 e_false
->flags
&= ~EDGE_FALLTHRU
;
717 e_false
->flags
|= EDGE_FALSE_VALUE
;
719 set_immediate_dominator (CDI_DOMINATORS
, cond_bb
, entry_edge
->src
);
720 set_immediate_dominator (CDI_DOMINATORS
, true_bb
, cond_bb
);
721 set_immediate_dominator (CDI_DOMINATORS
, false_bb
, cond_bb
);
722 set_immediate_dominator (CDI_DOMINATORS
, join_bb
, cond_bb
);
724 exit_edge
= single_succ_edge (join_bb
);
726 if (single_pred_p (exit_edge
->dest
))
727 set_immediate_dominator (CDI_DOMINATORS
, exit_edge
->dest
, join_bb
);
732 /* create_empty_loop_on_edge
734 | - pred_bb - ------ pred_bb ------
735 | | | | iv0 = initial_value |
736 | -----|----- ---------|-----------
737 | | ______ | entry_edge
739 | | ====> | -V---V- loop_header -------------
740 | V | | iv_before = phi (iv0, iv_after) |
741 | - succ_bb - | ---|-----------------------------
743 | ----------- | ---V--- loop_body ---------------
744 | | | iv_after = iv_before + stride |
745 | | | if (iv_before < upper_bound) |
746 | | ---|--------------\--------------
749 | | - loop_latch - V- succ_bb -
751 | | /------------- -----------
754 Creates an empty loop as shown above, the IV_BEFORE is the SSA_NAME
755 that is used before the increment of IV. IV_BEFORE should be used for
756 adding code to the body that uses the IV. OUTER is the outer loop in
757 which the new loop should be inserted.
759 Both INITIAL_VALUE and UPPER_BOUND expressions are gimplified and
760 inserted on the loop entry edge. This implies that this function
761 should be used only when the UPPER_BOUND expression is a loop
765 create_empty_loop_on_edge (edge entry_edge
,
767 tree stride
, tree upper_bound
,
773 basic_block loop_header
, loop_latch
, succ_bb
, pred_bb
;
775 gimple_stmt_iterator gsi
;
782 gcc_assert (entry_edge
&& initial_value
&& stride
&& upper_bound
&& iv
);
784 /* Create header, latch and wire up the loop. */
785 pred_bb
= entry_edge
->src
;
786 loop_header
= split_edge (entry_edge
);
787 loop_latch
= split_edge (single_succ_edge (loop_header
));
788 succ_bb
= single_succ (loop_latch
);
789 make_edge (loop_header
, succ_bb
, 0);
790 redirect_edge_succ_nodup (single_succ_edge (loop_latch
), loop_header
);
792 /* Set immediate dominator information. */
793 set_immediate_dominator (CDI_DOMINATORS
, loop_header
, pred_bb
);
794 set_immediate_dominator (CDI_DOMINATORS
, loop_latch
, loop_header
);
795 set_immediate_dominator (CDI_DOMINATORS
, succ_bb
, loop_header
);
797 /* Initialize a loop structure and put it in a loop hierarchy. */
798 loop
= alloc_loop ();
799 loop
->header
= loop_header
;
800 loop
->latch
= loop_latch
;
801 add_loop (loop
, outer
);
803 /* TODO: Fix frequencies and counts. */
804 prob
= REG_BR_PROB_BASE
/ 2;
806 scale_loop_frequencies (loop
, REG_BR_PROB_BASE
- prob
, REG_BR_PROB_BASE
);
808 /* Update dominators. */
809 update_dominators_in_loop (loop
);
811 /* Modify edge flags. */
812 exit_e
= single_exit (loop
);
813 exit_e
->flags
= EDGE_LOOP_EXIT
| EDGE_FALSE_VALUE
;
814 single_pred_edge (loop_latch
)->flags
= EDGE_TRUE_VALUE
;
816 /* Construct IV code in loop. */
817 initial_value
= force_gimple_operand (initial_value
, &stmts
, true, iv
);
820 gsi_insert_seq_on_edge (loop_preheader_edge (loop
), stmts
);
821 gsi_commit_edge_inserts ();
824 upper_bound
= force_gimple_operand (upper_bound
, &stmts
, true, NULL
);
827 gsi_insert_seq_on_edge (loop_preheader_edge (loop
), stmts
);
828 gsi_commit_edge_inserts ();
831 gsi
= gsi_last_bb (loop_header
);
832 create_iv (initial_value
, stride
, iv
, loop
, &gsi
, false,
833 iv_before
, iv_after
);
835 /* Insert loop exit condition. */
836 cond_expr
= gimple_build_cond
837 (LT_EXPR
, *iv_before
, upper_bound
, NULL_TREE
, NULL_TREE
);
839 exit_test
= gimple_cond_lhs (cond_expr
);
840 exit_test
= force_gimple_operand_gsi (&gsi
, exit_test
, true, NULL
,
841 false, GSI_NEW_STMT
);
842 gimple_cond_set_lhs (cond_expr
, exit_test
);
843 gsi
= gsi_last_bb (exit_e
->src
);
844 gsi_insert_after (&gsi
, cond_expr
, GSI_NEW_STMT
);
846 split_block_after_labels (loop_header
);
851 /* Make area between HEADER_EDGE and LATCH_EDGE a loop by connecting
852 latch to header and update loop tree and dominators
853 accordingly. Everything between them plus LATCH_EDGE destination must
854 be dominated by HEADER_EDGE destination, and back-reachable from
855 LATCH_EDGE source. HEADER_EDGE is redirected to basic block SWITCH_BB,
856 FALSE_EDGE of SWITCH_BB to original destination of HEADER_EDGE and
857 TRUE_EDGE of SWITCH_BB to original destination of LATCH_EDGE.
858 Returns the newly created loop. Frequencies and counts in the new loop
859 are scaled by FALSE_SCALE and in the old one by TRUE_SCALE. */
862 loopify (edge latch_edge
, edge header_edge
,
863 basic_block switch_bb
, edge true_edge
, edge false_edge
,
864 bool redirect_all_edges
, unsigned true_scale
, unsigned false_scale
)
866 basic_block succ_bb
= latch_edge
->dest
;
867 basic_block pred_bb
= header_edge
->src
;
868 struct loop
*loop
= alloc_loop ();
869 struct loop
*outer
= loop_outer (succ_bb
->loop_father
);
875 loop
->header
= header_edge
->dest
;
876 loop
->latch
= latch_edge
->src
;
878 freq
= EDGE_FREQUENCY (header_edge
);
879 cnt
= header_edge
->count
;
881 /* Redirect edges. */
882 loop_redirect_edge (latch_edge
, loop
->header
);
883 loop_redirect_edge (true_edge
, succ_bb
);
885 /* During loop versioning, one of the switch_bb edge is already properly
886 set. Do not redirect it again unless redirect_all_edges is true. */
887 if (redirect_all_edges
)
889 loop_redirect_edge (header_edge
, switch_bb
);
890 loop_redirect_edge (false_edge
, loop
->header
);
892 /* Update dominators. */
893 set_immediate_dominator (CDI_DOMINATORS
, switch_bb
, pred_bb
);
894 set_immediate_dominator (CDI_DOMINATORS
, loop
->header
, switch_bb
);
897 set_immediate_dominator (CDI_DOMINATORS
, succ_bb
, switch_bb
);
899 /* Compute new loop. */
900 add_loop (loop
, outer
);
902 /* Add switch_bb to appropriate loop. */
903 if (switch_bb
->loop_father
)
904 remove_bb_from_loops (switch_bb
);
905 add_bb_to_loop (switch_bb
, outer
);
907 /* Fix frequencies. */
908 if (redirect_all_edges
)
910 switch_bb
->frequency
= freq
;
911 switch_bb
->count
= cnt
;
912 FOR_EACH_EDGE (e
, ei
, switch_bb
->succs
)
914 e
->count
= switch_bb
->count
.apply_probability (e
->probability
);
917 scale_loop_frequencies (loop
, false_scale
, REG_BR_PROB_BASE
);
918 scale_loop_frequencies (succ_bb
->loop_father
, true_scale
, REG_BR_PROB_BASE
);
919 update_dominators_in_loop (loop
);
924 /* Remove the latch edge of a LOOP and update loops to indicate that
925 the LOOP was removed. After this function, original loop latch will
926 have no successor, which caller is expected to fix somehow.
928 If this may cause the information about irreducible regions to become
929 invalid, IRRED_INVALIDATED is set to true.
931 LOOP_CLOSED_SSA_INVALIDATED, if non-NULL, is a bitmap where we store
932 basic blocks that had non-trivial update on their loop_father.*/
935 unloop (struct loop
*loop
, bool *irred_invalidated
,
936 bitmap loop_closed_ssa_invalidated
)
941 basic_block latch
= loop
->latch
;
944 if (loop_preheader_edge (loop
)->flags
& EDGE_IRREDUCIBLE_LOOP
)
945 *irred_invalidated
= true;
947 /* This is relatively straightforward. The dominators are unchanged, as
948 loop header dominates loop latch, so the only thing we have to care of
949 is the placement of loops and basic blocks inside the loop tree. We
950 move them all to the loop->outer, and then let fix_bb_placements do
953 body
= get_loop_body (loop
);
955 for (i
= 0; i
< n
; i
++)
956 if (body
[i
]->loop_father
== loop
)
958 remove_bb_from_loops (body
[i
]);
959 add_bb_to_loop (body
[i
], loop_outer (loop
));
966 flow_loop_tree_node_remove (ploop
);
967 flow_loop_tree_node_add (loop_outer (loop
), ploop
);
970 /* Remove the loop and free its data. */
973 remove_edge (single_succ_edge (latch
));
975 /* We do not pass IRRED_INVALIDATED to fix_bb_placements here, as even if
976 there is an irreducible region inside the cancelled loop, the flags will
978 fix_bb_placements (latch
, &dummy
, loop_closed_ssa_invalidated
);
981 /* Fix placement of superloops of LOOP inside loop tree, i.e. ensure that
982 condition stated in description of fix_loop_placement holds for them.
983 It is used in case when we removed some edges coming out of LOOP, which
984 may cause the right placement of LOOP inside loop tree to change.
986 IRRED_INVALIDATED is set to true if a change in the loop structures might
987 invalidate the information about irreducible regions. */
990 fix_loop_placements (struct loop
*loop
, bool *irred_invalidated
)
994 while (loop_outer (loop
))
996 outer
= loop_outer (loop
);
997 if (!fix_loop_placement (loop
, irred_invalidated
))
1000 /* Changing the placement of a loop in the loop tree may alter the
1001 validity of condition 2) of the description of fix_bb_placement
1002 for its preheader, because the successor is the header and belongs
1003 to the loop. So call fix_bb_placements to fix up the placement
1004 of the preheader and (possibly) of its predecessors. */
1005 fix_bb_placements (loop_preheader_edge (loop
)->src
,
1006 irred_invalidated
, NULL
);
1011 /* Duplicate loop bounds and other information we store about
1012 the loop into its duplicate. */
1015 copy_loop_info (struct loop
*loop
, struct loop
*target
)
1017 gcc_checking_assert (!target
->any_upper_bound
&& !target
->any_estimate
);
1018 target
->any_upper_bound
= loop
->any_upper_bound
;
1019 target
->nb_iterations_upper_bound
= loop
->nb_iterations_upper_bound
;
1020 target
->any_likely_upper_bound
= loop
->any_likely_upper_bound
;
1021 target
->nb_iterations_likely_upper_bound
1022 = loop
->nb_iterations_likely_upper_bound
;
1023 target
->any_estimate
= loop
->any_estimate
;
1024 target
->nb_iterations_estimate
= loop
->nb_iterations_estimate
;
1025 target
->estimate_state
= loop
->estimate_state
;
1026 target
->constraints
= loop
->constraints
;
1027 target
->warned_aggressive_loop_optimizations
1028 |= loop
->warned_aggressive_loop_optimizations
;
1029 target
->in_oacc_kernels_region
= loop
->in_oacc_kernels_region
;
1032 /* Copies copy of LOOP as subloop of TARGET loop, placing newly
1033 created loop into loops structure. */
1035 duplicate_loop (struct loop
*loop
, struct loop
*target
)
1038 cloop
= alloc_loop ();
1039 place_new_loop (cfun
, cloop
);
1041 copy_loop_info (loop
, cloop
);
1043 /* Mark the new loop as copy of LOOP. */
1044 set_loop_copy (loop
, cloop
);
1046 /* Add it to target. */
1047 flow_loop_tree_node_add (target
, cloop
);
1052 /* Copies structure of subloops of LOOP into TARGET loop, placing
1053 newly created loops into loop tree. */
1055 duplicate_subloops (struct loop
*loop
, struct loop
*target
)
1057 struct loop
*aloop
, *cloop
;
1059 for (aloop
= loop
->inner
; aloop
; aloop
= aloop
->next
)
1061 cloop
= duplicate_loop (aloop
, target
);
1062 duplicate_subloops (aloop
, cloop
);
1066 /* Copies structure of subloops of N loops, stored in array COPIED_LOOPS,
1067 into TARGET loop, placing newly created loops into loop tree. */
1069 copy_loops_to (struct loop
**copied_loops
, int n
, struct loop
*target
)
1074 for (i
= 0; i
< n
; i
++)
1076 aloop
= duplicate_loop (copied_loops
[i
], target
);
1077 duplicate_subloops (copied_loops
[i
], aloop
);
1081 /* Redirects edge E to basic block DEST. */
1083 loop_redirect_edge (edge e
, basic_block dest
)
1085 if (e
->dest
== dest
)
1088 redirect_edge_and_branch_force (e
, dest
);
1091 /* Check whether LOOP's body can be duplicated. */
1093 can_duplicate_loop_p (const struct loop
*loop
)
1096 basic_block
*bbs
= get_loop_body (loop
);
1098 ret
= can_copy_bbs_p (bbs
, loop
->num_nodes
);
1104 /* Sets probability and count of edge E to zero. The probability and count
1105 is redistributed evenly to the remaining edges coming from E->src. */
1108 set_zero_probability (edge e
)
1110 basic_block bb
= e
->src
;
1112 edge ae
, last
= NULL
;
1113 unsigned n
= EDGE_COUNT (bb
->succs
);
1114 profile_count cnt
= e
->count
, cnt1
;
1115 unsigned prob
= e
->probability
, prob1
;
1118 cnt1
= cnt
.apply_scale (1, (n
- 1));
1119 prob1
= prob
/ (n
- 1);
1121 FOR_EACH_EDGE (ae
, ei
, bb
->succs
)
1126 ae
->probability
+= prob1
;
1131 /* Move the rest to one of the edges. */
1132 last
->probability
+= prob
% (n
- 1);
1133 /* TODO: Remove once we have fractional counts. */
1134 if (cnt
.initialized_p ())
1135 last
->count
+= profile_count::from_gcov_type (cnt
.to_gcov_type () % (n
- 1));
1138 e
->count
= profile_count::zero ();
1141 /* Duplicates body of LOOP to given edge E NDUPL times. Takes care of updating
1142 loop structure and dominators. E's destination must be LOOP header for
1143 this to work, i.e. it must be entry or latch edge of this loop; these are
1144 unique, as the loops must have preheaders for this function to work
1145 correctly (in case E is latch, the function unrolls the loop, if E is entry
1146 edge, it peels the loop). Store edges created by copying ORIG edge from
1147 copies corresponding to set bits in WONT_EXIT bitmap (bit 0 corresponds to
1148 original LOOP body, the other copies are numbered in order given by control
1149 flow through them) into TO_REMOVE array. Returns false if duplication is
1153 duplicate_loop_to_header_edge (struct loop
*loop
, edge e
,
1154 unsigned int ndupl
, sbitmap wont_exit
,
1155 edge orig
, vec
<edge
> *to_remove
,
1158 struct loop
*target
, *aloop
;
1159 struct loop
**orig_loops
;
1160 unsigned n_orig_loops
;
1161 basic_block header
= loop
->header
, latch
= loop
->latch
;
1162 basic_block
*new_bbs
, *bbs
, *first_active
;
1163 basic_block new_bb
, bb
, first_active_latch
= NULL
;
1164 edge ae
, latch_edge
;
1165 edge spec_edges
[2], new_spec_edges
[2];
1169 int is_latch
= (latch
== e
->src
);
1170 int scale_act
= 0, *scale_step
= NULL
, scale_main
= 0;
1171 int scale_after_exit
= 0;
1172 int p
, freq_in
, freq_le
, freq_out_orig
;
1173 int prob_pass_thru
, prob_pass_wont_exit
, prob_pass_main
;
1174 int add_irreducible_flag
;
1175 basic_block place_after
;
1176 bitmap bbs_to_scale
= NULL
;
1179 gcc_assert (e
->dest
== loop
->header
);
1180 gcc_assert (ndupl
> 0);
1184 /* Orig must be edge out of the loop. */
1185 gcc_assert (flow_bb_inside_loop_p (loop
, orig
->src
));
1186 gcc_assert (!flow_bb_inside_loop_p (loop
, orig
->dest
));
1189 n
= loop
->num_nodes
;
1190 bbs
= get_loop_body_in_dom_order (loop
);
1191 gcc_assert (bbs
[0] == loop
->header
);
1192 gcc_assert (bbs
[n
- 1] == loop
->latch
);
1194 /* Check whether duplication is possible. */
1195 if (!can_copy_bbs_p (bbs
, loop
->num_nodes
))
1200 new_bbs
= XNEWVEC (basic_block
, loop
->num_nodes
);
1202 /* In case we are doing loop peeling and the loop is in the middle of
1203 irreducible region, the peeled copies will be inside it too. */
1204 add_irreducible_flag
= e
->flags
& EDGE_IRREDUCIBLE_LOOP
;
1205 gcc_assert (!is_latch
|| !add_irreducible_flag
);
1207 /* Find edge from latch. */
1208 latch_edge
= loop_latch_edge (loop
);
1210 if (flags
& DLTHE_FLAG_UPDATE_FREQ
)
1212 /* Calculate coefficients by that we have to scale frequencies
1213 of duplicated loop bodies. */
1214 freq_in
= header
->frequency
;
1215 freq_le
= EDGE_FREQUENCY (latch_edge
);
1218 if (freq_in
< freq_le
)
1220 freq_out_orig
= orig
? EDGE_FREQUENCY (orig
) : freq_in
- freq_le
;
1221 if (freq_out_orig
> freq_in
- freq_le
)
1222 freq_out_orig
= freq_in
- freq_le
;
1223 prob_pass_thru
= RDIV (REG_BR_PROB_BASE
* freq_le
, freq_in
);
1224 prob_pass_wont_exit
=
1225 RDIV (REG_BR_PROB_BASE
* (freq_le
+ freq_out_orig
), freq_in
);
1228 && REG_BR_PROB_BASE
- orig
->probability
!= 0)
1230 /* The blocks that are dominated by a removed exit edge ORIG have
1231 frequencies scaled by this. */
1233 = GCOV_COMPUTE_SCALE (REG_BR_PROB_BASE
,
1234 REG_BR_PROB_BASE
- orig
->probability
);
1235 bbs_to_scale
= BITMAP_ALLOC (NULL
);
1236 for (i
= 0; i
< n
; i
++)
1238 if (bbs
[i
] != orig
->src
1239 && dominated_by_p (CDI_DOMINATORS
, bbs
[i
], orig
->src
))
1240 bitmap_set_bit (bbs_to_scale
, i
);
1244 scale_step
= XNEWVEC (int, ndupl
);
1246 for (i
= 1; i
<= ndupl
; i
++)
1247 scale_step
[i
- 1] = bitmap_bit_p (wont_exit
, i
)
1248 ? prob_pass_wont_exit
1251 /* Complete peeling is special as the probability of exit in last
1253 if (flags
& DLTHE_FLAG_COMPLETTE_PEEL
)
1255 int wanted_freq
= EDGE_FREQUENCY (e
);
1257 if (wanted_freq
> freq_in
)
1258 wanted_freq
= freq_in
;
1260 gcc_assert (!is_latch
);
1261 /* First copy has frequency of incoming edge. Each subsequent
1262 frequency should be reduced by prob_pass_wont_exit. Caller
1263 should've managed the flags so all except for original loop
1264 has won't exist set. */
1265 scale_act
= GCOV_COMPUTE_SCALE (wanted_freq
, freq_in
);
1266 /* Now simulate the duplication adjustments and compute header
1267 frequency of the last copy. */
1268 for (i
= 0; i
< ndupl
; i
++)
1269 wanted_freq
= combine_probabilities (wanted_freq
, scale_step
[i
]);
1270 scale_main
= GCOV_COMPUTE_SCALE (wanted_freq
, freq_in
);
1274 prob_pass_main
= bitmap_bit_p (wont_exit
, 0)
1275 ? prob_pass_wont_exit
1278 scale_main
= REG_BR_PROB_BASE
;
1279 for (i
= 0; i
< ndupl
; i
++)
1282 p
= combine_probabilities (p
, scale_step
[i
]);
1284 scale_main
= GCOV_COMPUTE_SCALE (REG_BR_PROB_BASE
, scale_main
);
1285 scale_act
= combine_probabilities (scale_main
, prob_pass_main
);
1289 int preheader_freq
= EDGE_FREQUENCY (e
);
1290 scale_main
= REG_BR_PROB_BASE
;
1291 for (i
= 0; i
< ndupl
; i
++)
1292 scale_main
= combine_probabilities (scale_main
, scale_step
[i
]);
1293 if (preheader_freq
> freq_in
)
1294 preheader_freq
= freq_in
;
1295 scale_act
= GCOV_COMPUTE_SCALE (preheader_freq
, freq_in
);
1297 for (i
= 0; i
< ndupl
; i
++)
1298 gcc_assert (scale_step
[i
] >= 0 && scale_step
[i
] <= REG_BR_PROB_BASE
);
1299 gcc_assert (scale_main
>= 0 && scale_main
<= REG_BR_PROB_BASE
1300 && scale_act
>= 0 && scale_act
<= REG_BR_PROB_BASE
);
1303 /* Loop the new bbs will belong to. */
1304 target
= e
->src
->loop_father
;
1306 /* Original loops. */
1308 for (aloop
= loop
->inner
; aloop
; aloop
= aloop
->next
)
1310 orig_loops
= XNEWVEC (struct loop
*, n_orig_loops
);
1311 for (aloop
= loop
->inner
, i
= 0; aloop
; aloop
= aloop
->next
, i
++)
1312 orig_loops
[i
] = aloop
;
1314 set_loop_copy (loop
, target
);
1316 first_active
= XNEWVEC (basic_block
, n
);
1319 memcpy (first_active
, bbs
, n
* sizeof (basic_block
));
1320 first_active_latch
= latch
;
1323 spec_edges
[SE_ORIG
] = orig
;
1324 spec_edges
[SE_LATCH
] = latch_edge
;
1326 place_after
= e
->src
;
1327 for (j
= 0; j
< ndupl
; j
++)
1330 copy_loops_to (orig_loops
, n_orig_loops
, target
);
1333 copy_bbs (bbs
, n
, new_bbs
, spec_edges
, 2, new_spec_edges
, loop
,
1335 place_after
= new_spec_edges
[SE_LATCH
]->src
;
1337 if (flags
& DLTHE_RECORD_COPY_NUMBER
)
1338 for (i
= 0; i
< n
; i
++)
1340 gcc_assert (!new_bbs
[i
]->aux
);
1341 new_bbs
[i
]->aux
= (void *)(size_t)(j
+ 1);
1344 /* Note whether the blocks and edges belong to an irreducible loop. */
1345 if (add_irreducible_flag
)
1347 for (i
= 0; i
< n
; i
++)
1348 new_bbs
[i
]->flags
|= BB_DUPLICATED
;
1349 for (i
= 0; i
< n
; i
++)
1352 new_bb
= new_bbs
[i
];
1353 if (new_bb
->loop_father
== target
)
1354 new_bb
->flags
|= BB_IRREDUCIBLE_LOOP
;
1356 FOR_EACH_EDGE (ae
, ei
, new_bb
->succs
)
1357 if ((ae
->dest
->flags
& BB_DUPLICATED
)
1358 && (ae
->src
->loop_father
== target
1359 || ae
->dest
->loop_father
== target
))
1360 ae
->flags
|= EDGE_IRREDUCIBLE_LOOP
;
1362 for (i
= 0; i
< n
; i
++)
1363 new_bbs
[i
]->flags
&= ~BB_DUPLICATED
;
1366 /* Redirect the special edges. */
1369 redirect_edge_and_branch_force (latch_edge
, new_bbs
[0]);
1370 redirect_edge_and_branch_force (new_spec_edges
[SE_LATCH
],
1372 set_immediate_dominator (CDI_DOMINATORS
, new_bbs
[0], latch
);
1373 latch
= loop
->latch
= new_bbs
[n
- 1];
1374 e
= latch_edge
= new_spec_edges
[SE_LATCH
];
1378 redirect_edge_and_branch_force (new_spec_edges
[SE_LATCH
],
1380 redirect_edge_and_branch_force (e
, new_bbs
[0]);
1381 set_immediate_dominator (CDI_DOMINATORS
, new_bbs
[0], e
->src
);
1382 e
= new_spec_edges
[SE_LATCH
];
1385 /* Record exit edge in this copy. */
1386 if (orig
&& bitmap_bit_p (wont_exit
, j
+ 1))
1389 to_remove
->safe_push (new_spec_edges
[SE_ORIG
]);
1390 set_zero_probability (new_spec_edges
[SE_ORIG
]);
1392 /* Scale the frequencies of the blocks dominated by the exit. */
1395 EXECUTE_IF_SET_IN_BITMAP (bbs_to_scale
, 0, i
, bi
)
1397 scale_bbs_frequencies_int (new_bbs
+ i
, 1, scale_after_exit
,
1403 /* Record the first copy in the control flow order if it is not
1404 the original loop (i.e. in case of peeling). */
1405 if (!first_active_latch
)
1407 memcpy (first_active
, new_bbs
, n
* sizeof (basic_block
));
1408 first_active_latch
= new_bbs
[n
- 1];
1411 /* Set counts and frequencies. */
1412 if (flags
& DLTHE_FLAG_UPDATE_FREQ
)
1414 scale_bbs_frequencies_int (new_bbs
, n
, scale_act
, REG_BR_PROB_BASE
);
1415 scale_act
= combine_probabilities (scale_act
, scale_step
[j
]);
1421 /* Record the exit edge in the original loop body, and update the frequencies. */
1422 if (orig
&& bitmap_bit_p (wont_exit
, 0))
1425 to_remove
->safe_push (orig
);
1426 set_zero_probability (orig
);
1428 /* Scale the frequencies of the blocks dominated by the exit. */
1431 EXECUTE_IF_SET_IN_BITMAP (bbs_to_scale
, 0, i
, bi
)
1433 scale_bbs_frequencies_int (bbs
+ i
, 1, scale_after_exit
,
1439 /* Update the original loop. */
1441 set_immediate_dominator (CDI_DOMINATORS
, e
->dest
, e
->src
);
1442 if (flags
& DLTHE_FLAG_UPDATE_FREQ
)
1444 scale_bbs_frequencies_int (bbs
, n
, scale_main
, REG_BR_PROB_BASE
);
1448 /* Update dominators of outer blocks if affected. */
1449 for (i
= 0; i
< n
; i
++)
1451 basic_block dominated
, dom_bb
;
1452 vec
<basic_block
> dom_bbs
;
1458 dom_bbs
= get_dominated_by (CDI_DOMINATORS
, bb
);
1459 FOR_EACH_VEC_ELT (dom_bbs
, j
, dominated
)
1461 if (flow_bb_inside_loop_p (loop
, dominated
))
1463 dom_bb
= nearest_common_dominator (
1464 CDI_DOMINATORS
, first_active
[i
], first_active_latch
);
1465 set_immediate_dominator (CDI_DOMINATORS
, dominated
, dom_bb
);
1469 free (first_active
);
1472 BITMAP_FREE (bbs_to_scale
);
1477 /* A callback for make_forwarder block, to redirect all edges except for
1478 MFB_KJ_EDGE to the entry part. E is the edge for that we should decide
1479 whether to redirect it. */
1483 mfb_keep_just (edge e
)
1485 return e
!= mfb_kj_edge
;
1488 /* True when a candidate preheader BLOCK has predecessors from LOOP. */
1491 has_preds_from_loop (basic_block block
, struct loop
*loop
)
1496 FOR_EACH_EDGE (e
, ei
, block
->preds
)
1497 if (e
->src
->loop_father
== loop
)
1502 /* Creates a pre-header for a LOOP. Returns newly created block. Unless
1503 CP_SIMPLE_PREHEADERS is set in FLAGS, we only force LOOP to have single
1504 entry; otherwise we also force preheader block to have only one successor.
1505 When CP_FALLTHRU_PREHEADERS is set in FLAGS, we force the preheader block
1506 to be a fallthru predecessor to the loop header and to have only
1507 predecessors from outside of the loop.
1508 The function also updates dominators. */
1511 create_preheader (struct loop
*loop
, int flags
)
1517 bool latch_edge_was_fallthru
;
1518 edge one_succ_pred
= NULL
, single_entry
= NULL
;
1521 FOR_EACH_EDGE (e
, ei
, loop
->header
->preds
)
1523 if (e
->src
== loop
->latch
)
1525 irred
|= (e
->flags
& EDGE_IRREDUCIBLE_LOOP
) != 0;
1528 if (single_succ_p (e
->src
))
1531 gcc_assert (nentry
);
1534 bool need_forwarder_block
= false;
1536 /* We do not allow entry block to be the loop preheader, since we
1537 cannot emit code there. */
1538 if (single_entry
->src
== ENTRY_BLOCK_PTR_FOR_FN (cfun
))
1539 need_forwarder_block
= true;
1542 /* If we want simple preheaders, also force the preheader to have
1543 just a single successor. */
1544 if ((flags
& CP_SIMPLE_PREHEADERS
)
1545 && !single_succ_p (single_entry
->src
))
1546 need_forwarder_block
= true;
1547 /* If we want fallthru preheaders, also create forwarder block when
1548 preheader ends with a jump or has predecessors from loop. */
1549 else if ((flags
& CP_FALLTHRU_PREHEADERS
)
1550 && (JUMP_P (BB_END (single_entry
->src
))
1551 || has_preds_from_loop (single_entry
->src
, loop
)))
1552 need_forwarder_block
= true;
1554 if (! need_forwarder_block
)
1558 mfb_kj_edge
= loop_latch_edge (loop
);
1559 latch_edge_was_fallthru
= (mfb_kj_edge
->flags
& EDGE_FALLTHRU
) != 0;
1561 dummy
= split_edge (single_entry
);
1564 edge fallthru
= make_forwarder_block (loop
->header
, mfb_keep_just
, NULL
);
1565 dummy
= fallthru
->src
;
1566 loop
->header
= fallthru
->dest
;
1569 /* Try to be clever in placing the newly created preheader. The idea is to
1570 avoid breaking any "fallthruness" relationship between blocks.
1572 The preheader was created just before the header and all incoming edges
1573 to the header were redirected to the preheader, except the latch edge.
1574 So the only problematic case is when this latch edge was a fallthru
1575 edge: it is not anymore after the preheader creation so we have broken
1576 the fallthruness. We're therefore going to look for a better place. */
1577 if (latch_edge_was_fallthru
)
1582 e
= EDGE_PRED (dummy
, 0);
1584 move_block_after (dummy
, e
->src
);
1589 dummy
->flags
|= BB_IRREDUCIBLE_LOOP
;
1590 single_succ_edge (dummy
)->flags
|= EDGE_IRREDUCIBLE_LOOP
;
1594 fprintf (dump_file
, "Created preheader block for loop %i\n",
1597 if (flags
& CP_FALLTHRU_PREHEADERS
)
1598 gcc_assert ((single_succ_edge (dummy
)->flags
& EDGE_FALLTHRU
)
1599 && !JUMP_P (BB_END (dummy
)));
1604 /* Create preheaders for each loop; for meaning of FLAGS see create_preheader. */
1607 create_preheaders (int flags
)
1614 FOR_EACH_LOOP (loop
, 0)
1615 create_preheader (loop
, flags
);
1616 loops_state_set (LOOPS_HAVE_PREHEADERS
);
1619 /* Forces all loop latches to have only single successor. */
1622 force_single_succ_latches (void)
1627 FOR_EACH_LOOP (loop
, 0)
1629 if (loop
->latch
!= loop
->header
&& single_succ_p (loop
->latch
))
1632 e
= find_edge (loop
->latch
, loop
->header
);
1633 gcc_checking_assert (e
!= NULL
);
1637 loops_state_set (LOOPS_HAVE_SIMPLE_LATCHES
);
1640 /* This function is called from loop_version. It splits the entry edge
1641 of the loop we want to version, adds the versioning condition, and
1642 adjust the edges to the two versions of the loop appropriately.
1643 e is an incoming edge. Returns the basic block containing the
1646 --- edge e ---- > [second_head]
1648 Split it and insert new conditional expression and adjust edges.
1650 --- edge e ---> [cond expr] ---> [first_head]
1652 +---------> [second_head]
1654 THEN_PROB is the probability of then branch of the condition.
1655 ELSE_PROB is the probability of else branch. Note that they may be both
1656 REG_BR_PROB_BASE when condition is IFN_LOOP_VECTORIZED. */
1659 lv_adjust_loop_entry_edge (basic_block first_head
, basic_block second_head
,
1660 edge e
, void *cond_expr
, unsigned then_prob
,
1663 basic_block new_head
= NULL
;
1666 gcc_assert (e
->dest
== second_head
);
1668 /* Split edge 'e'. This will create a new basic block, where we can
1669 insert conditional expr. */
1670 new_head
= split_edge (e
);
1672 lv_add_condition_to_bb (first_head
, second_head
, new_head
,
1675 /* Don't set EDGE_TRUE_VALUE in RTL mode, as it's invalid there. */
1676 e
= single_succ_edge (new_head
);
1677 e1
= make_edge (new_head
, first_head
,
1678 current_ir_type () == IR_GIMPLE
? EDGE_TRUE_VALUE
: 0);
1679 e1
->probability
= then_prob
;
1680 e
->probability
= else_prob
;
1681 e1
->count
= e
->count
.apply_probability (e1
->probability
);
1682 e
->count
= e
->count
.apply_probability (e
->probability
);
1684 set_immediate_dominator (CDI_DOMINATORS
, first_head
, new_head
);
1685 set_immediate_dominator (CDI_DOMINATORS
, second_head
, new_head
);
1687 /* Adjust loop header phi nodes. */
1688 lv_adjust_loop_header_phi (first_head
, second_head
, new_head
, e1
);
1693 /* Main entry point for Loop Versioning transformation.
1695 This transformation given a condition and a loop, creates
1696 -if (condition) { loop_copy1 } else { loop_copy2 },
1697 where loop_copy1 is the loop transformed in one way, and loop_copy2
1698 is the loop transformed in another way (or unchanged). COND_EXPR
1699 may be a run time test for things that were not resolved by static
1700 analysis (overlapping ranges (anti-aliasing), alignment, etc.).
1702 If non-NULL, CONDITION_BB is set to the basic block containing the
1705 THEN_PROB is the probability of the then edge of the if. THEN_SCALE
1706 is the ratio by that the frequencies in the original loop should
1707 be scaled. ELSE_SCALE is the ratio by that the frequencies in the
1708 new loop should be scaled.
1710 If PLACE_AFTER is true, we place the new loop after LOOP in the
1711 instruction stream, otherwise it is placed before LOOP. */
1714 loop_version (struct loop
*loop
,
1715 void *cond_expr
, basic_block
*condition_bb
,
1716 unsigned then_prob
, unsigned else_prob
,
1717 unsigned then_scale
, unsigned else_scale
,
1720 basic_block first_head
, second_head
;
1721 edge entry
, latch_edge
, true_edge
, false_edge
;
1724 basic_block cond_bb
;
1726 /* Record entry and latch edges for the loop */
1727 entry
= loop_preheader_edge (loop
);
1728 irred_flag
= entry
->flags
& EDGE_IRREDUCIBLE_LOOP
;
1729 entry
->flags
&= ~EDGE_IRREDUCIBLE_LOOP
;
1731 /* Note down head of loop as first_head. */
1732 first_head
= entry
->dest
;
1734 /* Duplicate loop. */
1735 if (!cfg_hook_duplicate_loop_to_header_edge (loop
, entry
, 1,
1736 NULL
, NULL
, NULL
, 0))
1738 entry
->flags
|= irred_flag
;
1742 /* After duplication entry edge now points to new loop head block.
1743 Note down new head as second_head. */
1744 second_head
= entry
->dest
;
1746 /* Split loop entry edge and insert new block with cond expr. */
1747 cond_bb
= lv_adjust_loop_entry_edge (first_head
, second_head
,
1748 entry
, cond_expr
, then_prob
, else_prob
);
1750 *condition_bb
= cond_bb
;
1754 entry
->flags
|= irred_flag
;
1758 latch_edge
= single_succ_edge (get_bb_copy (loop
->latch
));
1760 extract_cond_bb_edges (cond_bb
, &true_edge
, &false_edge
);
1761 nloop
= loopify (latch_edge
,
1762 single_pred_edge (get_bb_copy (loop
->header
)),
1763 cond_bb
, true_edge
, false_edge
,
1764 false /* Do not redirect all edges. */,
1765 then_scale
, else_scale
);
1767 copy_loop_info (loop
, nloop
);
1769 /* loopify redirected latch_edge. Update its PENDING_STMTS. */
1770 lv_flush_pending_stmts (latch_edge
);
1772 /* loopify redirected condition_bb's succ edge. Update its PENDING_STMTS. */
1773 extract_cond_bb_edges (cond_bb
, &true_edge
, &false_edge
);
1774 lv_flush_pending_stmts (false_edge
);
1775 /* Adjust irreducible flag. */
1778 cond_bb
->flags
|= BB_IRREDUCIBLE_LOOP
;
1779 loop_preheader_edge (loop
)->flags
|= EDGE_IRREDUCIBLE_LOOP
;
1780 loop_preheader_edge (nloop
)->flags
|= EDGE_IRREDUCIBLE_LOOP
;
1781 single_pred_edge (cond_bb
)->flags
|= EDGE_IRREDUCIBLE_LOOP
;
1786 basic_block
*bbs
= get_loop_body_in_dom_order (nloop
), after
;
1789 after
= loop
->latch
;
1791 for (i
= 0; i
< nloop
->num_nodes
; i
++)
1793 move_block_after (bbs
[i
], after
);
1799 /* At this point condition_bb is loop preheader with two successors,
1800 first_head and second_head. Make sure that loop preheader has only
1802 split_edge (loop_preheader_edge (loop
));
1803 split_edge (loop_preheader_edge (nloop
));