1 /* Loop manipulation code for GNU compiler.
2 Copyright (C) 2002-2016 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). */
304 basic_block
*rem_bbs
, *bord_bbs
, from
, bb
;
305 vec
<basic_block
> dom_bbs
;
306 int i
, nrem
, n_bord_bbs
;
307 bool irred_invalidated
= false;
311 if (!can_remove_branch_p (e
))
314 /* Keep track of whether we need to update information about irreducible
315 regions. This is the case if the removed area is a part of the
316 irreducible region, or if the set of basic blocks that belong to a loop
317 that is inside an irreducible region is changed, or if such a loop is
319 if (e
->flags
& EDGE_IRREDUCIBLE_LOOP
)
320 irred_invalidated
= true;
322 /* We need to check whether basic blocks are dominated by the edge
323 e, but we only have basic block dominators. This is easy to
324 fix -- when e->dest has exactly one predecessor, this corresponds
325 to blocks dominated by e->dest, if not, split the edge. */
326 if (!single_pred_p (e
->dest
))
327 e
= single_pred_edge (split_edge (e
));
329 /* It may happen that by removing path we remove one or more loops
330 we belong to. In this case first unloop the loops, then proceed
331 normally. We may assume that e->dest is not a header of any loop,
332 as it now has exactly one predecessor. */
333 for (l
= e
->src
->loop_father
; loop_outer (l
); l
= f
)
336 if (dominated_by_p (CDI_DOMINATORS
, l
->latch
, e
->dest
))
337 unloop (l
, &irred_invalidated
, NULL
);
340 /* Identify the path. */
341 nrem
= find_path (e
, &rem_bbs
);
344 bord_bbs
= XNEWVEC (basic_block
, n_basic_blocks_for_fn (cfun
));
345 auto_sbitmap
seen (last_basic_block_for_fn (cfun
));
348 /* Find "border" hexes -- i.e. those with predecessor in removed path. */
349 for (i
= 0; i
< nrem
; i
++)
350 bitmap_set_bit (seen
, rem_bbs
[i
]->index
);
351 if (!irred_invalidated
)
352 FOR_EACH_EDGE (ae
, ei
, e
->src
->succs
)
353 if (ae
!= e
&& ae
->dest
!= EXIT_BLOCK_PTR_FOR_FN (cfun
)
354 && !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_FOR_FN (cfun
)
366 && !bitmap_bit_p (seen
, ae
->dest
->index
))
368 bitmap_set_bit (seen
, ae
->dest
->index
);
369 bord_bbs
[n_bord_bbs
++] = ae
->dest
;
371 if (ae
->flags
& EDGE_IRREDUCIBLE_LOOP
)
372 irred_invalidated
= true;
376 /* Remove the path. */
381 /* Cancel loops contained in the path. */
382 for (i
= 0; i
< nrem
; i
++)
383 if (rem_bbs
[i
]->loop_father
->header
== rem_bbs
[i
])
384 cancel_loop_tree (rem_bbs
[i
]->loop_father
);
386 remove_bbs (rem_bbs
, nrem
);
389 /* Find blocks whose dominators may be affected. */
391 for (i
= 0; i
< n_bord_bbs
; i
++)
395 bb
= get_immediate_dominator (CDI_DOMINATORS
, bord_bbs
[i
]);
396 if (bitmap_bit_p (seen
, bb
->index
))
398 bitmap_set_bit (seen
, bb
->index
);
400 for (ldom
= first_dom_son (CDI_DOMINATORS
, bb
);
402 ldom
= next_dom_son (CDI_DOMINATORS
, ldom
))
403 if (!dominated_by_p (CDI_DOMINATORS
, from
, ldom
))
404 dom_bbs
.safe_push (ldom
);
407 /* Recount dominators. */
408 iterate_fix_dominators (CDI_DOMINATORS
, dom_bbs
, true);
412 /* Fix placements of basic blocks inside loops and the placement of
413 loops in the loop tree. */
414 fix_bb_placements (from
, &irred_invalidated
, NULL
);
415 fix_loop_placements (from
->loop_father
, &irred_invalidated
);
417 if (irred_invalidated
418 && loops_state_satisfies_p (LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS
))
419 mark_irreducible_loops ();
424 /* Creates place for a new LOOP in loops structure of FN. */
427 place_new_loop (struct function
*fn
, struct loop
*loop
)
429 loop
->num
= number_of_loops (fn
);
430 vec_safe_push (loops_for_fn (fn
)->larray
, loop
);
433 /* Given LOOP structure with filled header and latch, find the body of the
434 corresponding loop and add it to loops tree. Insert the LOOP as a son of
438 add_loop (struct loop
*loop
, struct loop
*outer
)
442 struct loop
*subloop
;
446 /* Add it to loop structure. */
447 place_new_loop (cfun
, loop
);
448 flow_loop_tree_node_add (outer
, loop
);
450 /* Find its nodes. */
451 bbs
= XNEWVEC (basic_block
, n_basic_blocks_for_fn (cfun
));
452 n
= get_loop_body_with_size (loop
, bbs
, n_basic_blocks_for_fn (cfun
));
454 for (i
= 0; i
< n
; i
++)
456 if (bbs
[i
]->loop_father
== outer
)
458 remove_bb_from_loops (bbs
[i
]);
459 add_bb_to_loop (bbs
[i
], loop
);
465 /* If we find a direct subloop of OUTER, move it to LOOP. */
466 subloop
= bbs
[i
]->loop_father
;
467 if (loop_outer (subloop
) == outer
468 && subloop
->header
== bbs
[i
])
470 flow_loop_tree_node_remove (subloop
);
471 flow_loop_tree_node_add (loop
, subloop
);
475 /* Update the information about loop exit edges. */
476 for (i
= 0; i
< n
; i
++)
478 FOR_EACH_EDGE (e
, ei
, bbs
[i
]->succs
)
480 rescan_loop_exit (e
, false, false);
487 /* Multiply all frequencies in LOOP by NUM/DEN. */
490 scale_loop_frequencies (struct loop
*loop
, int num
, int den
)
494 bbs
= get_loop_body (loop
);
495 scale_bbs_frequencies_int (bbs
, loop
->num_nodes
, num
, den
);
499 /* Multiply all frequencies in LOOP by SCALE/REG_BR_PROB_BASE.
500 If ITERATION_BOUND is non-zero, scale even further if loop is predicted
501 to iterate too many times. */
504 scale_loop_profile (struct loop
*loop
, int scale
, gcov_type iteration_bound
)
506 gcov_type iterations
= expected_loop_iterations_unbounded (loop
);
510 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
511 fprintf (dump_file
, ";; Scaling loop %i with scale %f, "
512 "bounding iterations to %i from guessed %i\n",
513 loop
->num
, (double)scale
/ REG_BR_PROB_BASE
,
514 (int)iteration_bound
, (int)iterations
);
516 /* See if loop is predicted to iterate too many times. */
517 if (iteration_bound
&& iterations
> 0
518 && apply_probability (iterations
, scale
) > iteration_bound
)
520 /* Fixing loop profile for different trip count is not trivial; the exit
521 probabilities has to be updated to match and frequencies propagated down
524 We fully update only the simple case of loop with single exit that is
525 either from the latch or BB just before latch and leads from BB with
526 simple conditional jump. This is OK for use in vectorizer. */
527 e
= single_exit (loop
);
532 gcov_type count_delta
;
534 FOR_EACH_EDGE (other_e
, ei
, e
->src
->succs
)
535 if (!(other_e
->flags
& (EDGE_ABNORMAL
| EDGE_FAKE
))
539 /* Probability of exit must be 1/iterations. */
540 freq_delta
= EDGE_FREQUENCY (e
);
541 e
->probability
= REG_BR_PROB_BASE
/ iteration_bound
;
542 other_e
->probability
= inverse_probability (e
->probability
);
543 freq_delta
-= EDGE_FREQUENCY (e
);
545 /* Adjust counts accordingly. */
546 count_delta
= e
->count
;
547 e
->count
= apply_probability (e
->src
->count
, e
->probability
);
548 other_e
->count
= apply_probability (e
->src
->count
, other_e
->probability
);
549 count_delta
-= e
->count
;
551 /* If latch exists, change its frequency and count, since we changed
552 probability of exit. Theoretically we should update everything from
553 source of exit edge to latch, but for vectorizer this is enough. */
555 && loop
->latch
!= e
->src
)
557 loop
->latch
->frequency
+= freq_delta
;
558 if (loop
->latch
->frequency
< 0)
559 loop
->latch
->frequency
= 0;
560 loop
->latch
->count
+= count_delta
;
561 if (loop
->latch
->count
< 0)
562 loop
->latch
->count
= 0;
566 /* Roughly speaking we want to reduce the loop body profile by the
567 difference of loop iterations. We however can do better if
568 we look at the actual profile, if it is available. */
569 scale
= RDIV (iteration_bound
* scale
, iterations
);
570 if (loop
->header
->count
)
572 gcov_type count_in
= 0;
574 FOR_EACH_EDGE (e
, ei
, loop
->header
->preds
)
575 if (e
->src
!= loop
->latch
)
576 count_in
+= e
->count
;
579 scale
= GCOV_COMPUTE_SCALE (count_in
* iteration_bound
,
580 loop
->header
->count
);
582 else if (loop
->header
->frequency
)
586 FOR_EACH_EDGE (e
, ei
, loop
->header
->preds
)
587 if (e
->src
!= loop
->latch
)
588 freq_in
+= EDGE_FREQUENCY (e
);
591 scale
= GCOV_COMPUTE_SCALE (freq_in
* iteration_bound
,
592 loop
->header
->frequency
);
598 if (scale
== REG_BR_PROB_BASE
)
601 /* Scale the actual probabilities. */
602 scale_loop_frequencies (loop
, scale
, REG_BR_PROB_BASE
);
603 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
604 fprintf (dump_file
, ";; guessed iterations are now %i\n",
605 (int)expected_loop_iterations_unbounded (loop
));
608 /* Recompute dominance information for basic blocks outside LOOP. */
611 update_dominators_in_loop (struct loop
*loop
)
613 vec
<basic_block
> dom_bbs
= vNULL
;
617 auto_sbitmap
seen (last_basic_block_for_fn (cfun
));
619 body
= get_loop_body (loop
);
621 for (i
= 0; i
< loop
->num_nodes
; i
++)
622 bitmap_set_bit (seen
, body
[i
]->index
);
624 for (i
= 0; i
< loop
->num_nodes
; i
++)
628 for (ldom
= first_dom_son (CDI_DOMINATORS
, body
[i
]);
630 ldom
= next_dom_son (CDI_DOMINATORS
, ldom
))
631 if (!bitmap_bit_p (seen
, ldom
->index
))
633 bitmap_set_bit (seen
, ldom
->index
);
634 dom_bbs
.safe_push (ldom
);
638 iterate_fix_dominators (CDI_DOMINATORS
, dom_bbs
, false);
643 /* Creates an if region as shown above. CONDITION is used to create
647 | ------------- -------------
648 | | pred_bb | | pred_bb |
649 | ------------- -------------
653 | | ====> -------------
658 | ------------- e_false / \ e_true
660 | ------------- ----------- -----------
661 | | false_bb | | true_bb |
662 | ----------- -----------
669 | | exit_edge (result)
678 create_empty_if_region_on_edge (edge entry_edge
, tree condition
)
681 basic_block cond_bb
, true_bb
, false_bb
, join_bb
;
682 edge e_true
, e_false
, exit_edge
;
685 gimple_stmt_iterator gsi
;
687 cond_bb
= split_edge (entry_edge
);
689 /* Insert condition in cond_bb. */
690 gsi
= gsi_last_bb (cond_bb
);
692 force_gimple_operand_gsi (&gsi
, condition
, true, NULL
,
693 false, GSI_NEW_STMT
);
694 cond_stmt
= gimple_build_cond_from_tree (simple_cond
, NULL_TREE
, NULL_TREE
);
695 gsi
= gsi_last_bb (cond_bb
);
696 gsi_insert_after (&gsi
, cond_stmt
, GSI_NEW_STMT
);
698 join_bb
= split_edge (single_succ_edge (cond_bb
));
700 e_true
= single_succ_edge (cond_bb
);
701 true_bb
= split_edge (e_true
);
703 e_false
= make_edge (cond_bb
, join_bb
, 0);
704 false_bb
= split_edge (e_false
);
706 e_true
->flags
&= ~EDGE_FALLTHRU
;
707 e_true
->flags
|= EDGE_TRUE_VALUE
;
708 e_false
->flags
&= ~EDGE_FALLTHRU
;
709 e_false
->flags
|= EDGE_FALSE_VALUE
;
711 set_immediate_dominator (CDI_DOMINATORS
, cond_bb
, entry_edge
->src
);
712 set_immediate_dominator (CDI_DOMINATORS
, true_bb
, cond_bb
);
713 set_immediate_dominator (CDI_DOMINATORS
, false_bb
, cond_bb
);
714 set_immediate_dominator (CDI_DOMINATORS
, join_bb
, cond_bb
);
716 exit_edge
= single_succ_edge (join_bb
);
718 if (single_pred_p (exit_edge
->dest
))
719 set_immediate_dominator (CDI_DOMINATORS
, exit_edge
->dest
, join_bb
);
724 /* create_empty_loop_on_edge
726 | - pred_bb - ------ pred_bb ------
727 | | | | iv0 = initial_value |
728 | -----|----- ---------|-----------
729 | | ______ | entry_edge
731 | | ====> | -V---V- loop_header -------------
732 | V | | iv_before = phi (iv0, iv_after) |
733 | - succ_bb - | ---|-----------------------------
735 | ----------- | ---V--- loop_body ---------------
736 | | | iv_after = iv_before + stride |
737 | | | if (iv_before < upper_bound) |
738 | | ---|--------------\--------------
741 | | - loop_latch - V- succ_bb -
743 | | /------------- -----------
746 Creates an empty loop as shown above, the IV_BEFORE is the SSA_NAME
747 that is used before the increment of IV. IV_BEFORE should be used for
748 adding code to the body that uses the IV. OUTER is the outer loop in
749 which the new loop should be inserted.
751 Both INITIAL_VALUE and UPPER_BOUND expressions are gimplified and
752 inserted on the loop entry edge. This implies that this function
753 should be used only when the UPPER_BOUND expression is a loop
757 create_empty_loop_on_edge (edge entry_edge
,
759 tree stride
, tree upper_bound
,
765 basic_block loop_header
, loop_latch
, succ_bb
, pred_bb
;
767 gimple_stmt_iterator gsi
;
774 gcc_assert (entry_edge
&& initial_value
&& stride
&& upper_bound
&& iv
);
776 /* Create header, latch and wire up the loop. */
777 pred_bb
= entry_edge
->src
;
778 loop_header
= split_edge (entry_edge
);
779 loop_latch
= split_edge (single_succ_edge (loop_header
));
780 succ_bb
= single_succ (loop_latch
);
781 make_edge (loop_header
, succ_bb
, 0);
782 redirect_edge_succ_nodup (single_succ_edge (loop_latch
), loop_header
);
784 /* Set immediate dominator information. */
785 set_immediate_dominator (CDI_DOMINATORS
, loop_header
, pred_bb
);
786 set_immediate_dominator (CDI_DOMINATORS
, loop_latch
, loop_header
);
787 set_immediate_dominator (CDI_DOMINATORS
, succ_bb
, loop_header
);
789 /* Initialize a loop structure and put it in a loop hierarchy. */
790 loop
= alloc_loop ();
791 loop
->header
= loop_header
;
792 loop
->latch
= loop_latch
;
793 add_loop (loop
, outer
);
795 /* TODO: Fix frequencies and counts. */
796 prob
= REG_BR_PROB_BASE
/ 2;
798 scale_loop_frequencies (loop
, REG_BR_PROB_BASE
- prob
, REG_BR_PROB_BASE
);
800 /* Update dominators. */
801 update_dominators_in_loop (loop
);
803 /* Modify edge flags. */
804 exit_e
= single_exit (loop
);
805 exit_e
->flags
= EDGE_LOOP_EXIT
| EDGE_FALSE_VALUE
;
806 single_pred_edge (loop_latch
)->flags
= EDGE_TRUE_VALUE
;
808 /* Construct IV code in loop. */
809 initial_value
= force_gimple_operand (initial_value
, &stmts
, true, iv
);
812 gsi_insert_seq_on_edge (loop_preheader_edge (loop
), stmts
);
813 gsi_commit_edge_inserts ();
816 upper_bound
= force_gimple_operand (upper_bound
, &stmts
, true, NULL
);
819 gsi_insert_seq_on_edge (loop_preheader_edge (loop
), stmts
);
820 gsi_commit_edge_inserts ();
823 gsi
= gsi_last_bb (loop_header
);
824 create_iv (initial_value
, stride
, iv
, loop
, &gsi
, false,
825 iv_before
, iv_after
);
827 /* Insert loop exit condition. */
828 cond_expr
= gimple_build_cond
829 (LT_EXPR
, *iv_before
, upper_bound
, NULL_TREE
, NULL_TREE
);
831 exit_test
= gimple_cond_lhs (cond_expr
);
832 exit_test
= force_gimple_operand_gsi (&gsi
, exit_test
, true, NULL
,
833 false, GSI_NEW_STMT
);
834 gimple_cond_set_lhs (cond_expr
, exit_test
);
835 gsi
= gsi_last_bb (exit_e
->src
);
836 gsi_insert_after (&gsi
, cond_expr
, GSI_NEW_STMT
);
838 split_block_after_labels (loop_header
);
843 /* Make area between HEADER_EDGE and LATCH_EDGE a loop by connecting
844 latch to header and update loop tree and dominators
845 accordingly. Everything between them plus LATCH_EDGE destination must
846 be dominated by HEADER_EDGE destination, and back-reachable from
847 LATCH_EDGE source. HEADER_EDGE is redirected to basic block SWITCH_BB,
848 FALSE_EDGE of SWITCH_BB to original destination of HEADER_EDGE and
849 TRUE_EDGE of SWITCH_BB to original destination of LATCH_EDGE.
850 Returns the newly created loop. Frequencies and counts in the new loop
851 are scaled by FALSE_SCALE and in the old one by TRUE_SCALE. */
854 loopify (edge latch_edge
, edge header_edge
,
855 basic_block switch_bb
, edge true_edge
, edge false_edge
,
856 bool redirect_all_edges
, unsigned true_scale
, unsigned false_scale
)
858 basic_block succ_bb
= latch_edge
->dest
;
859 basic_block pred_bb
= header_edge
->src
;
860 struct loop
*loop
= alloc_loop ();
861 struct loop
*outer
= loop_outer (succ_bb
->loop_father
);
867 loop
->header
= header_edge
->dest
;
868 loop
->latch
= latch_edge
->src
;
870 freq
= EDGE_FREQUENCY (header_edge
);
871 cnt
= header_edge
->count
;
873 /* Redirect edges. */
874 loop_redirect_edge (latch_edge
, loop
->header
);
875 loop_redirect_edge (true_edge
, succ_bb
);
877 /* During loop versioning, one of the switch_bb edge is already properly
878 set. Do not redirect it again unless redirect_all_edges is true. */
879 if (redirect_all_edges
)
881 loop_redirect_edge (header_edge
, switch_bb
);
882 loop_redirect_edge (false_edge
, loop
->header
);
884 /* Update dominators. */
885 set_immediate_dominator (CDI_DOMINATORS
, switch_bb
, pred_bb
);
886 set_immediate_dominator (CDI_DOMINATORS
, loop
->header
, switch_bb
);
889 set_immediate_dominator (CDI_DOMINATORS
, succ_bb
, switch_bb
);
891 /* Compute new loop. */
892 add_loop (loop
, outer
);
894 /* Add switch_bb to appropriate loop. */
895 if (switch_bb
->loop_father
)
896 remove_bb_from_loops (switch_bb
);
897 add_bb_to_loop (switch_bb
, outer
);
899 /* Fix frequencies. */
900 if (redirect_all_edges
)
902 switch_bb
->frequency
= freq
;
903 switch_bb
->count
= cnt
;
904 FOR_EACH_EDGE (e
, ei
, switch_bb
->succs
)
906 e
->count
= apply_probability (switch_bb
->count
, e
->probability
);
909 scale_loop_frequencies (loop
, false_scale
, REG_BR_PROB_BASE
);
910 scale_loop_frequencies (succ_bb
->loop_father
, true_scale
, REG_BR_PROB_BASE
);
911 update_dominators_in_loop (loop
);
916 /* Remove the latch edge of a LOOP and update loops to indicate that
917 the LOOP was removed. After this function, original loop latch will
918 have no successor, which caller is expected to fix somehow.
920 If this may cause the information about irreducible regions to become
921 invalid, IRRED_INVALIDATED is set to true.
923 LOOP_CLOSED_SSA_INVALIDATED, if non-NULL, is a bitmap where we store
924 basic blocks that had non-trivial update on their loop_father.*/
927 unloop (struct loop
*loop
, bool *irred_invalidated
,
928 bitmap loop_closed_ssa_invalidated
)
933 basic_block latch
= loop
->latch
;
936 if (loop_preheader_edge (loop
)->flags
& EDGE_IRREDUCIBLE_LOOP
)
937 *irred_invalidated
= true;
939 /* This is relatively straightforward. The dominators are unchanged, as
940 loop header dominates loop latch, so the only thing we have to care of
941 is the placement of loops and basic blocks inside the loop tree. We
942 move them all to the loop->outer, and then let fix_bb_placements do
945 body
= get_loop_body (loop
);
947 for (i
= 0; i
< n
; i
++)
948 if (body
[i
]->loop_father
== loop
)
950 remove_bb_from_loops (body
[i
]);
951 add_bb_to_loop (body
[i
], loop_outer (loop
));
958 flow_loop_tree_node_remove (ploop
);
959 flow_loop_tree_node_add (loop_outer (loop
), ploop
);
962 /* Remove the loop and free its data. */
965 remove_edge (single_succ_edge (latch
));
967 /* We do not pass IRRED_INVALIDATED to fix_bb_placements here, as even if
968 there is an irreducible region inside the cancelled loop, the flags will
970 fix_bb_placements (latch
, &dummy
, loop_closed_ssa_invalidated
);
973 /* Fix placement of superloops of LOOP inside loop tree, i.e. ensure that
974 condition stated in description of fix_loop_placement holds for them.
975 It is used in case when we removed some edges coming out of LOOP, which
976 may cause the right placement of LOOP inside loop tree to change.
978 IRRED_INVALIDATED is set to true if a change in the loop structures might
979 invalidate the information about irreducible regions. */
982 fix_loop_placements (struct loop
*loop
, bool *irred_invalidated
)
986 while (loop_outer (loop
))
988 outer
= loop_outer (loop
);
989 if (!fix_loop_placement (loop
, irred_invalidated
))
992 /* Changing the placement of a loop in the loop tree may alter the
993 validity of condition 2) of the description of fix_bb_placement
994 for its preheader, because the successor is the header and belongs
995 to the loop. So call fix_bb_placements to fix up the placement
996 of the preheader and (possibly) of its predecessors. */
997 fix_bb_placements (loop_preheader_edge (loop
)->src
,
998 irred_invalidated
, NULL
);
1003 /* Duplicate loop bounds and other information we store about
1004 the loop into its duplicate. */
1007 copy_loop_info (struct loop
*loop
, struct loop
*target
)
1009 gcc_checking_assert (!target
->any_upper_bound
&& !target
->any_estimate
);
1010 target
->any_upper_bound
= loop
->any_upper_bound
;
1011 target
->nb_iterations_upper_bound
= loop
->nb_iterations_upper_bound
;
1012 target
->any_likely_upper_bound
= loop
->any_likely_upper_bound
;
1013 target
->nb_iterations_likely_upper_bound
1014 = loop
->nb_iterations_likely_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
;
1018 target
->warned_aggressive_loop_optimizations
1019 |= loop
->warned_aggressive_loop_optimizations
;
1020 target
->in_oacc_kernels_region
= loop
->in_oacc_kernels_region
;
1023 /* Copies copy of LOOP as subloop of TARGET loop, placing newly
1024 created loop into loops structure. */
1026 duplicate_loop (struct loop
*loop
, struct loop
*target
)
1029 cloop
= alloc_loop ();
1030 place_new_loop (cfun
, cloop
);
1032 copy_loop_info (loop
, cloop
);
1034 /* Mark the new loop as copy of LOOP. */
1035 set_loop_copy (loop
, cloop
);
1037 /* Add it to target. */
1038 flow_loop_tree_node_add (target
, cloop
);
1043 /* Copies structure of subloops of LOOP into TARGET loop, placing
1044 newly created loops into loop tree. */
1046 duplicate_subloops (struct loop
*loop
, struct loop
*target
)
1048 struct loop
*aloop
, *cloop
;
1050 for (aloop
= loop
->inner
; aloop
; aloop
= aloop
->next
)
1052 cloop
= duplicate_loop (aloop
, target
);
1053 duplicate_subloops (aloop
, cloop
);
1057 /* Copies structure of subloops of N loops, stored in array COPIED_LOOPS,
1058 into TARGET loop, placing newly created loops into loop tree. */
1060 copy_loops_to (struct loop
**copied_loops
, int n
, struct loop
*target
)
1065 for (i
= 0; i
< n
; i
++)
1067 aloop
= duplicate_loop (copied_loops
[i
], target
);
1068 duplicate_subloops (copied_loops
[i
], aloop
);
1072 /* Redirects edge E to basic block DEST. */
1074 loop_redirect_edge (edge e
, basic_block dest
)
1076 if (e
->dest
== dest
)
1079 redirect_edge_and_branch_force (e
, dest
);
1082 /* Check whether LOOP's body can be duplicated. */
1084 can_duplicate_loop_p (const struct loop
*loop
)
1087 basic_block
*bbs
= get_loop_body (loop
);
1089 ret
= can_copy_bbs_p (bbs
, loop
->num_nodes
);
1095 /* Sets probability and count of edge E to zero. The probability and count
1096 is redistributed evenly to the remaining edges coming from E->src. */
1099 set_zero_probability (edge e
)
1101 basic_block bb
= e
->src
;
1103 edge ae
, last
= NULL
;
1104 unsigned n
= EDGE_COUNT (bb
->succs
);
1105 gcov_type cnt
= e
->count
, cnt1
;
1106 unsigned prob
= e
->probability
, prob1
;
1109 cnt1
= cnt
/ (n
- 1);
1110 prob1
= prob
/ (n
- 1);
1112 FOR_EACH_EDGE (ae
, ei
, bb
->succs
)
1117 ae
->probability
+= prob1
;
1122 /* Move the rest to one of the edges. */
1123 last
->probability
+= prob
% (n
- 1);
1124 last
->count
+= cnt
% (n
- 1);
1130 /* Duplicates body of LOOP to given edge E NDUPL times. Takes care of updating
1131 loop structure and dominators. E's destination must be LOOP header for
1132 this to work, i.e. it must be entry or latch edge of this loop; these are
1133 unique, as the loops must have preheaders for this function to work
1134 correctly (in case E is latch, the function unrolls the loop, if E is entry
1135 edge, it peels the loop). Store edges created by copying ORIG edge from
1136 copies corresponding to set bits in WONT_EXIT bitmap (bit 0 corresponds to
1137 original LOOP body, the other copies are numbered in order given by control
1138 flow through them) into TO_REMOVE array. Returns false if duplication is
1142 duplicate_loop_to_header_edge (struct loop
*loop
, edge e
,
1143 unsigned int ndupl
, sbitmap wont_exit
,
1144 edge orig
, vec
<edge
> *to_remove
,
1147 struct loop
*target
, *aloop
;
1148 struct loop
**orig_loops
;
1149 unsigned n_orig_loops
;
1150 basic_block header
= loop
->header
, latch
= loop
->latch
;
1151 basic_block
*new_bbs
, *bbs
, *first_active
;
1152 basic_block new_bb
, bb
, first_active_latch
= NULL
;
1153 edge ae
, latch_edge
;
1154 edge spec_edges
[2], new_spec_edges
[2];
1158 int is_latch
= (latch
== e
->src
);
1159 int scale_act
= 0, *scale_step
= NULL
, scale_main
= 0;
1160 int scale_after_exit
= 0;
1161 int p
, freq_in
, freq_le
, freq_out_orig
;
1162 int prob_pass_thru
, prob_pass_wont_exit
, prob_pass_main
;
1163 int add_irreducible_flag
;
1164 basic_block place_after
;
1165 bitmap bbs_to_scale
= NULL
;
1168 gcc_assert (e
->dest
== loop
->header
);
1169 gcc_assert (ndupl
> 0);
1173 /* Orig must be edge out of the loop. */
1174 gcc_assert (flow_bb_inside_loop_p (loop
, orig
->src
));
1175 gcc_assert (!flow_bb_inside_loop_p (loop
, orig
->dest
));
1178 n
= loop
->num_nodes
;
1179 bbs
= get_loop_body_in_dom_order (loop
);
1180 gcc_assert (bbs
[0] == loop
->header
);
1181 gcc_assert (bbs
[n
- 1] == loop
->latch
);
1183 /* Check whether duplication is possible. */
1184 if (!can_copy_bbs_p (bbs
, loop
->num_nodes
))
1189 new_bbs
= XNEWVEC (basic_block
, loop
->num_nodes
);
1191 /* In case we are doing loop peeling and the loop is in the middle of
1192 irreducible region, the peeled copies will be inside it too. */
1193 add_irreducible_flag
= e
->flags
& EDGE_IRREDUCIBLE_LOOP
;
1194 gcc_assert (!is_latch
|| !add_irreducible_flag
);
1196 /* Find edge from latch. */
1197 latch_edge
= loop_latch_edge (loop
);
1199 if (flags
& DLTHE_FLAG_UPDATE_FREQ
)
1201 /* Calculate coefficients by that we have to scale frequencies
1202 of duplicated loop bodies. */
1203 freq_in
= header
->frequency
;
1204 freq_le
= EDGE_FREQUENCY (latch_edge
);
1207 if (freq_in
< freq_le
)
1209 freq_out_orig
= orig
? EDGE_FREQUENCY (orig
) : freq_in
- freq_le
;
1210 if (freq_out_orig
> freq_in
- freq_le
)
1211 freq_out_orig
= freq_in
- freq_le
;
1212 prob_pass_thru
= RDIV (REG_BR_PROB_BASE
* freq_le
, freq_in
);
1213 prob_pass_wont_exit
=
1214 RDIV (REG_BR_PROB_BASE
* (freq_le
+ freq_out_orig
), freq_in
);
1217 && REG_BR_PROB_BASE
- orig
->probability
!= 0)
1219 /* The blocks that are dominated by a removed exit edge ORIG have
1220 frequencies scaled by this. */
1222 = GCOV_COMPUTE_SCALE (REG_BR_PROB_BASE
,
1223 REG_BR_PROB_BASE
- orig
->probability
);
1224 bbs_to_scale
= BITMAP_ALLOC (NULL
);
1225 for (i
= 0; i
< n
; i
++)
1227 if (bbs
[i
] != orig
->src
1228 && dominated_by_p (CDI_DOMINATORS
, bbs
[i
], orig
->src
))
1229 bitmap_set_bit (bbs_to_scale
, i
);
1233 scale_step
= XNEWVEC (int, ndupl
);
1235 for (i
= 1; i
<= ndupl
; i
++)
1236 scale_step
[i
- 1] = bitmap_bit_p (wont_exit
, i
)
1237 ? prob_pass_wont_exit
1240 /* Complete peeling is special as the probability of exit in last
1242 if (flags
& DLTHE_FLAG_COMPLETTE_PEEL
)
1244 int wanted_freq
= EDGE_FREQUENCY (e
);
1246 if (wanted_freq
> freq_in
)
1247 wanted_freq
= freq_in
;
1249 gcc_assert (!is_latch
);
1250 /* First copy has frequency of incoming edge. Each subsequent
1251 frequency should be reduced by prob_pass_wont_exit. Caller
1252 should've managed the flags so all except for original loop
1253 has won't exist set. */
1254 scale_act
= GCOV_COMPUTE_SCALE (wanted_freq
, freq_in
);
1255 /* Now simulate the duplication adjustments and compute header
1256 frequency of the last copy. */
1257 for (i
= 0; i
< ndupl
; i
++)
1258 wanted_freq
= combine_probabilities (wanted_freq
, scale_step
[i
]);
1259 scale_main
= GCOV_COMPUTE_SCALE (wanted_freq
, freq_in
);
1263 prob_pass_main
= bitmap_bit_p (wont_exit
, 0)
1264 ? prob_pass_wont_exit
1267 scale_main
= REG_BR_PROB_BASE
;
1268 for (i
= 0; i
< ndupl
; i
++)
1271 p
= combine_probabilities (p
, scale_step
[i
]);
1273 scale_main
= GCOV_COMPUTE_SCALE (REG_BR_PROB_BASE
, scale_main
);
1274 scale_act
= combine_probabilities (scale_main
, prob_pass_main
);
1278 scale_main
= REG_BR_PROB_BASE
;
1279 for (i
= 0; i
< ndupl
; i
++)
1280 scale_main
= combine_probabilities (scale_main
, scale_step
[i
]);
1281 scale_act
= REG_BR_PROB_BASE
- prob_pass_thru
;
1283 for (i
= 0; i
< ndupl
; i
++)
1284 gcc_assert (scale_step
[i
] >= 0 && scale_step
[i
] <= REG_BR_PROB_BASE
);
1285 gcc_assert (scale_main
>= 0 && scale_main
<= REG_BR_PROB_BASE
1286 && scale_act
>= 0 && scale_act
<= REG_BR_PROB_BASE
);
1289 /* Loop the new bbs will belong to. */
1290 target
= e
->src
->loop_father
;
1292 /* Original loops. */
1294 for (aloop
= loop
->inner
; aloop
; aloop
= aloop
->next
)
1296 orig_loops
= XNEWVEC (struct loop
*, n_orig_loops
);
1297 for (aloop
= loop
->inner
, i
= 0; aloop
; aloop
= aloop
->next
, i
++)
1298 orig_loops
[i
] = aloop
;
1300 set_loop_copy (loop
, target
);
1302 first_active
= XNEWVEC (basic_block
, n
);
1305 memcpy (first_active
, bbs
, n
* sizeof (basic_block
));
1306 first_active_latch
= latch
;
1309 spec_edges
[SE_ORIG
] = orig
;
1310 spec_edges
[SE_LATCH
] = latch_edge
;
1312 place_after
= e
->src
;
1313 for (j
= 0; j
< ndupl
; j
++)
1316 copy_loops_to (orig_loops
, n_orig_loops
, target
);
1319 copy_bbs (bbs
, n
, new_bbs
, spec_edges
, 2, new_spec_edges
, loop
,
1321 place_after
= new_spec_edges
[SE_LATCH
]->src
;
1323 if (flags
& DLTHE_RECORD_COPY_NUMBER
)
1324 for (i
= 0; i
< n
; i
++)
1326 gcc_assert (!new_bbs
[i
]->aux
);
1327 new_bbs
[i
]->aux
= (void *)(size_t)(j
+ 1);
1330 /* Note whether the blocks and edges belong to an irreducible loop. */
1331 if (add_irreducible_flag
)
1333 for (i
= 0; i
< n
; i
++)
1334 new_bbs
[i
]->flags
|= BB_DUPLICATED
;
1335 for (i
= 0; i
< n
; i
++)
1338 new_bb
= new_bbs
[i
];
1339 if (new_bb
->loop_father
== target
)
1340 new_bb
->flags
|= BB_IRREDUCIBLE_LOOP
;
1342 FOR_EACH_EDGE (ae
, ei
, new_bb
->succs
)
1343 if ((ae
->dest
->flags
& BB_DUPLICATED
)
1344 && (ae
->src
->loop_father
== target
1345 || ae
->dest
->loop_father
== target
))
1346 ae
->flags
|= EDGE_IRREDUCIBLE_LOOP
;
1348 for (i
= 0; i
< n
; i
++)
1349 new_bbs
[i
]->flags
&= ~BB_DUPLICATED
;
1352 /* Redirect the special edges. */
1355 redirect_edge_and_branch_force (latch_edge
, new_bbs
[0]);
1356 redirect_edge_and_branch_force (new_spec_edges
[SE_LATCH
],
1358 set_immediate_dominator (CDI_DOMINATORS
, new_bbs
[0], latch
);
1359 latch
= loop
->latch
= new_bbs
[n
- 1];
1360 e
= latch_edge
= new_spec_edges
[SE_LATCH
];
1364 redirect_edge_and_branch_force (new_spec_edges
[SE_LATCH
],
1366 redirect_edge_and_branch_force (e
, new_bbs
[0]);
1367 set_immediate_dominator (CDI_DOMINATORS
, new_bbs
[0], e
->src
);
1368 e
= new_spec_edges
[SE_LATCH
];
1371 /* Record exit edge in this copy. */
1372 if (orig
&& bitmap_bit_p (wont_exit
, j
+ 1))
1375 to_remove
->safe_push (new_spec_edges
[SE_ORIG
]);
1376 set_zero_probability (new_spec_edges
[SE_ORIG
]);
1378 /* Scale the frequencies of the blocks dominated by the exit. */
1381 EXECUTE_IF_SET_IN_BITMAP (bbs_to_scale
, 0, i
, bi
)
1383 scale_bbs_frequencies_int (new_bbs
+ i
, 1, scale_after_exit
,
1389 /* Record the first copy in the control flow order if it is not
1390 the original loop (i.e. in case of peeling). */
1391 if (!first_active_latch
)
1393 memcpy (first_active
, new_bbs
, n
* sizeof (basic_block
));
1394 first_active_latch
= new_bbs
[n
- 1];
1397 /* Set counts and frequencies. */
1398 if (flags
& DLTHE_FLAG_UPDATE_FREQ
)
1400 scale_bbs_frequencies_int (new_bbs
, n
, scale_act
, REG_BR_PROB_BASE
);
1401 scale_act
= combine_probabilities (scale_act
, scale_step
[j
]);
1407 /* Record the exit edge in the original loop body, and update the frequencies. */
1408 if (orig
&& bitmap_bit_p (wont_exit
, 0))
1411 to_remove
->safe_push (orig
);
1412 set_zero_probability (orig
);
1414 /* Scale the frequencies of the blocks dominated by the exit. */
1417 EXECUTE_IF_SET_IN_BITMAP (bbs_to_scale
, 0, i
, bi
)
1419 scale_bbs_frequencies_int (bbs
+ i
, 1, scale_after_exit
,
1425 /* Update the original loop. */
1427 set_immediate_dominator (CDI_DOMINATORS
, e
->dest
, e
->src
);
1428 if (flags
& DLTHE_FLAG_UPDATE_FREQ
)
1430 scale_bbs_frequencies_int (bbs
, n
, scale_main
, REG_BR_PROB_BASE
);
1434 /* Update dominators of outer blocks if affected. */
1435 for (i
= 0; i
< n
; i
++)
1437 basic_block dominated
, dom_bb
;
1438 vec
<basic_block
> dom_bbs
;
1444 dom_bbs
= get_dominated_by (CDI_DOMINATORS
, bb
);
1445 FOR_EACH_VEC_ELT (dom_bbs
, j
, dominated
)
1447 if (flow_bb_inside_loop_p (loop
, dominated
))
1449 dom_bb
= nearest_common_dominator (
1450 CDI_DOMINATORS
, first_active
[i
], first_active_latch
);
1451 set_immediate_dominator (CDI_DOMINATORS
, dominated
, dom_bb
);
1455 free (first_active
);
1458 BITMAP_FREE (bbs_to_scale
);
1463 /* A callback for make_forwarder block, to redirect all edges except for
1464 MFB_KJ_EDGE to the entry part. E is the edge for that we should decide
1465 whether to redirect it. */
1469 mfb_keep_just (edge e
)
1471 return e
!= mfb_kj_edge
;
1474 /* True when a candidate preheader BLOCK has predecessors from LOOP. */
1477 has_preds_from_loop (basic_block block
, struct loop
*loop
)
1482 FOR_EACH_EDGE (e
, ei
, block
->preds
)
1483 if (e
->src
->loop_father
== loop
)
1488 /* Creates a pre-header for a LOOP. Returns newly created block. Unless
1489 CP_SIMPLE_PREHEADERS is set in FLAGS, we only force LOOP to have single
1490 entry; otherwise we also force preheader block to have only one successor.
1491 When CP_FALLTHRU_PREHEADERS is set in FLAGS, we force the preheader block
1492 to be a fallthru predecessor to the loop header and to have only
1493 predecessors from outside of the loop.
1494 The function also updates dominators. */
1497 create_preheader (struct loop
*loop
, int flags
)
1503 bool latch_edge_was_fallthru
;
1504 edge one_succ_pred
= NULL
, single_entry
= NULL
;
1507 FOR_EACH_EDGE (e
, ei
, loop
->header
->preds
)
1509 if (e
->src
== loop
->latch
)
1511 irred
|= (e
->flags
& EDGE_IRREDUCIBLE_LOOP
) != 0;
1514 if (single_succ_p (e
->src
))
1517 gcc_assert (nentry
);
1520 bool need_forwarder_block
= false;
1522 /* We do not allow entry block to be the loop preheader, since we
1523 cannot emit code there. */
1524 if (single_entry
->src
== ENTRY_BLOCK_PTR_FOR_FN (cfun
))
1525 need_forwarder_block
= true;
1528 /* If we want simple preheaders, also force the preheader to have
1529 just a single successor. */
1530 if ((flags
& CP_SIMPLE_PREHEADERS
)
1531 && !single_succ_p (single_entry
->src
))
1532 need_forwarder_block
= true;
1533 /* If we want fallthru preheaders, also create forwarder block when
1534 preheader ends with a jump or has predecessors from loop. */
1535 else if ((flags
& CP_FALLTHRU_PREHEADERS
)
1536 && (JUMP_P (BB_END (single_entry
->src
))
1537 || has_preds_from_loop (single_entry
->src
, loop
)))
1538 need_forwarder_block
= true;
1540 if (! need_forwarder_block
)
1544 mfb_kj_edge
= loop_latch_edge (loop
);
1545 latch_edge_was_fallthru
= (mfb_kj_edge
->flags
& EDGE_FALLTHRU
) != 0;
1546 fallthru
= make_forwarder_block (loop
->header
, mfb_keep_just
, NULL
);
1547 dummy
= fallthru
->src
;
1548 loop
->header
= fallthru
->dest
;
1550 /* Try to be clever in placing the newly created preheader. The idea is to
1551 avoid breaking any "fallthruness" relationship between blocks.
1553 The preheader was created just before the header and all incoming edges
1554 to the header were redirected to the preheader, except the latch edge.
1555 So the only problematic case is when this latch edge was a fallthru
1556 edge: it is not anymore after the preheader creation so we have broken
1557 the fallthruness. We're therefore going to look for a better place. */
1558 if (latch_edge_was_fallthru
)
1563 e
= EDGE_PRED (dummy
, 0);
1565 move_block_after (dummy
, e
->src
);
1570 dummy
->flags
|= BB_IRREDUCIBLE_LOOP
;
1571 single_succ_edge (dummy
)->flags
|= EDGE_IRREDUCIBLE_LOOP
;
1575 fprintf (dump_file
, "Created preheader block for loop %i\n",
1578 if (flags
& CP_FALLTHRU_PREHEADERS
)
1579 gcc_assert ((single_succ_edge (dummy
)->flags
& EDGE_FALLTHRU
)
1580 && !JUMP_P (BB_END (dummy
)));
1585 /* Create preheaders for each loop; for meaning of FLAGS see create_preheader. */
1588 create_preheaders (int flags
)
1595 FOR_EACH_LOOP (loop
, 0)
1596 create_preheader (loop
, flags
);
1597 loops_state_set (LOOPS_HAVE_PREHEADERS
);
1600 /* Forces all loop latches to have only single successor. */
1603 force_single_succ_latches (void)
1608 FOR_EACH_LOOP (loop
, 0)
1610 if (loop
->latch
!= loop
->header
&& single_succ_p (loop
->latch
))
1613 e
= find_edge (loop
->latch
, loop
->header
);
1614 gcc_checking_assert (e
!= NULL
);
1618 loops_state_set (LOOPS_HAVE_SIMPLE_LATCHES
);
1621 /* This function is called from loop_version. It splits the entry edge
1622 of the loop we want to version, adds the versioning condition, and
1623 adjust the edges to the two versions of the loop appropriately.
1624 e is an incoming edge. Returns the basic block containing the
1627 --- edge e ---- > [second_head]
1629 Split it and insert new conditional expression and adjust edges.
1631 --- edge e ---> [cond expr] ---> [first_head]
1633 +---------> [second_head]
1635 THEN_PROB is the probability of then branch of the condition. */
1638 lv_adjust_loop_entry_edge (basic_block first_head
, basic_block second_head
,
1639 edge e
, void *cond_expr
, unsigned then_prob
)
1641 basic_block new_head
= NULL
;
1644 gcc_assert (e
->dest
== second_head
);
1646 /* Split edge 'e'. This will create a new basic block, where we can
1647 insert conditional expr. */
1648 new_head
= split_edge (e
);
1650 lv_add_condition_to_bb (first_head
, second_head
, new_head
,
1653 /* Don't set EDGE_TRUE_VALUE in RTL mode, as it's invalid there. */
1654 e
= single_succ_edge (new_head
);
1655 e1
= make_edge (new_head
, first_head
,
1656 current_ir_type () == IR_GIMPLE
? EDGE_TRUE_VALUE
: 0);
1657 e1
->probability
= then_prob
;
1658 e
->probability
= REG_BR_PROB_BASE
- then_prob
;
1659 e1
->count
= apply_probability (e
->count
, e1
->probability
);
1660 e
->count
= apply_probability (e
->count
, e
->probability
);
1662 set_immediate_dominator (CDI_DOMINATORS
, first_head
, new_head
);
1663 set_immediate_dominator (CDI_DOMINATORS
, second_head
, new_head
);
1665 /* Adjust loop header phi nodes. */
1666 lv_adjust_loop_header_phi (first_head
, second_head
, new_head
, e1
);
1671 /* Main entry point for Loop Versioning transformation.
1673 This transformation given a condition and a loop, creates
1674 -if (condition) { loop_copy1 } else { loop_copy2 },
1675 where loop_copy1 is the loop transformed in one way, and loop_copy2
1676 is the loop transformed in another way (or unchanged). 'condition'
1677 may be a run time test for things that were not resolved by static
1678 analysis (overlapping ranges (anti-aliasing), alignment, etc.).
1680 THEN_PROB is the probability of the then edge of the if. THEN_SCALE
1681 is the ratio by that the frequencies in the original loop should
1682 be scaled. ELSE_SCALE is the ratio by that the frequencies in the
1683 new loop should be scaled.
1685 If PLACE_AFTER is true, we place the new loop after LOOP in the
1686 instruction stream, otherwise it is placed before LOOP. */
1689 loop_version (struct loop
*loop
,
1690 void *cond_expr
, basic_block
*condition_bb
,
1691 unsigned then_prob
, unsigned then_scale
, unsigned else_scale
,
1694 basic_block first_head
, second_head
;
1695 edge entry
, latch_edge
, true_edge
, false_edge
;
1698 basic_block cond_bb
;
1700 /* Record entry and latch edges for the loop */
1701 entry
= loop_preheader_edge (loop
);
1702 irred_flag
= entry
->flags
& EDGE_IRREDUCIBLE_LOOP
;
1703 entry
->flags
&= ~EDGE_IRREDUCIBLE_LOOP
;
1705 /* Note down head of loop as first_head. */
1706 first_head
= entry
->dest
;
1708 /* Duplicate loop. */
1709 if (!cfg_hook_duplicate_loop_to_header_edge (loop
, entry
, 1,
1710 NULL
, NULL
, NULL
, 0))
1712 entry
->flags
|= irred_flag
;
1716 /* After duplication entry edge now points to new loop head block.
1717 Note down new head as second_head. */
1718 second_head
= entry
->dest
;
1720 /* Split loop entry edge and insert new block with cond expr. */
1721 cond_bb
= lv_adjust_loop_entry_edge (first_head
, second_head
,
1722 entry
, cond_expr
, then_prob
);
1724 *condition_bb
= cond_bb
;
1728 entry
->flags
|= irred_flag
;
1732 latch_edge
= single_succ_edge (get_bb_copy (loop
->latch
));
1734 extract_cond_bb_edges (cond_bb
, &true_edge
, &false_edge
);
1735 nloop
= loopify (latch_edge
,
1736 single_pred_edge (get_bb_copy (loop
->header
)),
1737 cond_bb
, true_edge
, false_edge
,
1738 false /* Do not redirect all edges. */,
1739 then_scale
, else_scale
);
1741 copy_loop_info (loop
, nloop
);
1743 /* loopify redirected latch_edge. Update its PENDING_STMTS. */
1744 lv_flush_pending_stmts (latch_edge
);
1746 /* loopify redirected condition_bb's succ edge. Update its PENDING_STMTS. */
1747 extract_cond_bb_edges (cond_bb
, &true_edge
, &false_edge
);
1748 lv_flush_pending_stmts (false_edge
);
1749 /* Adjust irreducible flag. */
1752 cond_bb
->flags
|= BB_IRREDUCIBLE_LOOP
;
1753 loop_preheader_edge (loop
)->flags
|= EDGE_IRREDUCIBLE_LOOP
;
1754 loop_preheader_edge (nloop
)->flags
|= EDGE_IRREDUCIBLE_LOOP
;
1755 single_pred_edge (cond_bb
)->flags
|= EDGE_IRREDUCIBLE_LOOP
;
1760 basic_block
*bbs
= get_loop_body_in_dom_order (nloop
), after
;
1763 after
= loop
->latch
;
1765 for (i
= 0; i
< nloop
->num_nodes
; i
++)
1767 move_block_after (bbs
[i
], after
);
1773 /* At this point condition_bb is loop preheader with two successors,
1774 first_head and second_head. Make sure that loop preheader has only
1776 split_edge (loop_preheader_edge (loop
));
1777 split_edge (loop_preheader_edge (nloop
));