1 /* Loop manipulation code for GNU compiler.
2 Copyright (C) 2002-2023 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"
36 static void copy_loops_to (class loop
**, int,
38 static void loop_redirect_edge (edge
, basic_block
);
39 static void remove_bbs (basic_block
*, int);
40 static bool rpe_enum_p (const_basic_block
, const void *);
41 static int find_path (edge
, basic_block
**);
42 static void fix_loop_placements (class loop
*, bool *);
43 static bool fix_bb_placement (basic_block
);
44 static void fix_bb_placements (basic_block
, bool *, bitmap
);
46 /* Checks whether basic block BB is dominated by DATA. */
48 rpe_enum_p (const_basic_block bb
, const void *data
)
50 return dominated_by_p (CDI_DOMINATORS
, bb
, (const_basic_block
) data
);
53 /* Remove basic blocks BBS. NBBS is the number of the basic blocks. */
56 remove_bbs (basic_block
*bbs
, int nbbs
)
60 for (i
= 0; i
< nbbs
; i
++)
61 delete_basic_block (bbs
[i
]);
64 /* Find path -- i.e. the basic blocks dominated by edge E and put them
65 into array BBS, that will be allocated large enough to contain them.
66 E->dest must have exactly one predecessor for this to work (it is
67 easy to achieve and we do not put it here because we do not want to
68 alter anything by this function). The number of basic blocks in the
71 find_path (edge e
, basic_block
**bbs
)
73 gcc_assert (EDGE_COUNT (e
->dest
->preds
) <= 1);
75 /* Find bbs in the path. */
76 *bbs
= XNEWVEC (basic_block
, n_basic_blocks_for_fn (cfun
));
77 return dfs_enumerate_from (e
->dest
, 0, rpe_enum_p
, *bbs
,
78 n_basic_blocks_for_fn (cfun
), e
->dest
);
81 /* Fix placement of basic block BB inside loop hierarchy --
82 Let L be a loop to that BB belongs. Then every successor of BB must either
83 1) belong to some superloop of loop L, or
84 2) be a header of loop K such that K->outer is superloop of L
85 Returns true if we had to move BB into other loop to enforce this condition,
86 false if the placement of BB was already correct (provided that placements
87 of its successors are correct). */
89 fix_bb_placement (basic_block bb
)
93 class loop
*loop
= current_loops
->tree_root
, *act
;
95 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
97 if (e
->dest
== EXIT_BLOCK_PTR_FOR_FN (cfun
))
100 act
= e
->dest
->loop_father
;
101 if (act
->header
== e
->dest
)
102 act
= loop_outer (act
);
104 if (flow_loop_nested_p (loop
, act
))
108 if (loop
== bb
->loop_father
)
111 remove_bb_from_loops (bb
);
112 add_bb_to_loop (bb
, loop
);
117 /* Fix placement of LOOP inside loop tree, i.e. find the innermost superloop
118 of LOOP to that leads at least one exit edge of LOOP, and set it
119 as the immediate superloop of LOOP. Return true if the immediate superloop
122 IRRED_INVALIDATED is set to true if a change in the loop structures might
123 invalidate the information about irreducible regions. */
126 fix_loop_placement (class loop
*loop
, bool *irred_invalidated
)
130 auto_vec
<edge
> exits
= get_loop_exit_edges (loop
);
131 class loop
*father
= current_loops
->tree_root
, *act
;
134 FOR_EACH_VEC_ELT (exits
, i
, e
)
136 act
= find_common_loop (loop
, e
->dest
->loop_father
);
137 if (flow_loop_nested_p (father
, act
))
141 if (father
!= loop_outer (loop
))
143 for (act
= loop_outer (loop
); act
!= father
; act
= loop_outer (act
))
144 act
->num_nodes
-= loop
->num_nodes
;
145 flow_loop_tree_node_remove (loop
);
146 flow_loop_tree_node_add (father
, loop
);
148 /* The exit edges of LOOP no longer exits its original immediate
149 superloops; remove them from the appropriate exit lists. */
150 FOR_EACH_VEC_ELT (exits
, i
, e
)
152 /* We may need to recompute irreducible loops. */
153 if (e
->flags
& EDGE_IRREDUCIBLE_LOOP
)
154 *irred_invalidated
= true;
155 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 class 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
++)
367 FOR_EACH_EDGE (ae
, ei
, rem_bbs
[i
]->succs
)
368 if (ae
->dest
!= EXIT_BLOCK_PTR_FOR_FN (cfun
)
369 && !bitmap_bit_p (seen
, ae
->dest
->index
))
371 bitmap_set_bit (seen
, ae
->dest
->index
);
372 bord_bbs
[n_bord_bbs
++] = ae
->dest
;
374 if (ae
->flags
& EDGE_IRREDUCIBLE_LOOP
)
375 *irred_invalidated
= true;
379 /* Remove the path. */
384 /* Cancel loops contained in the path. */
385 for (i
= 0; i
< nrem
; i
++)
386 if (rem_bbs
[i
]->loop_father
->header
== rem_bbs
[i
])
387 cancel_loop_tree (rem_bbs
[i
]->loop_father
);
389 remove_bbs (rem_bbs
, nrem
);
392 /* Find blocks whose dominators may be affected. */
394 for (i
= 0; i
< n_bord_bbs
; i
++)
398 bb
= get_immediate_dominator (CDI_DOMINATORS
, bord_bbs
[i
]);
399 if (bitmap_bit_p (seen
, bb
->index
))
401 bitmap_set_bit (seen
, bb
->index
);
403 for (ldom
= first_dom_son (CDI_DOMINATORS
, bb
);
405 ldom
= next_dom_son (CDI_DOMINATORS
, ldom
))
406 if (!dominated_by_p (CDI_DOMINATORS
, from
, ldom
))
407 dom_bbs
.safe_push (ldom
);
410 /* Recount dominators. */
411 iterate_fix_dominators (CDI_DOMINATORS
, dom_bbs
, true);
415 /* Fix placements of basic blocks inside loops and the placement of
416 loops in the loop tree. */
417 fix_bb_placements (from
, irred_invalidated
, loop_closed_ssa_invalidated
);
418 fix_loop_placements (from
->loop_father
, irred_invalidated
);
420 if (local_irred_invalidated
421 && loops_state_satisfies_p (LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS
))
422 mark_irreducible_loops ();
427 /* Creates place for a new LOOP in loops structure of FN. */
430 place_new_loop (struct function
*fn
, class loop
*loop
)
432 loop
->num
= number_of_loops (fn
);
433 vec_safe_push (loops_for_fn (fn
)->larray
, loop
);
436 /* Given LOOP structure with filled header and latch, find the body of the
437 corresponding loop and add it to loops tree. Insert the LOOP as a son of
441 add_loop (class loop
*loop
, class loop
*outer
)
449 /* Add it to loop structure. */
450 place_new_loop (cfun
, loop
);
451 flow_loop_tree_node_add (outer
, loop
);
453 /* Find its nodes. */
454 bbs
= XNEWVEC (basic_block
, n_basic_blocks_for_fn (cfun
));
455 n
= get_loop_body_with_size (loop
, bbs
, n_basic_blocks_for_fn (cfun
));
457 for (i
= 0; i
< n
; i
++)
459 if (bbs
[i
]->loop_father
== outer
)
461 remove_bb_from_loops (bbs
[i
]);
462 add_bb_to_loop (bbs
[i
], loop
);
468 /* If we find a direct subloop of OUTER, move it to LOOP. */
469 subloop
= bbs
[i
]->loop_father
;
470 if (loop_outer (subloop
) == outer
471 && subloop
->header
== bbs
[i
])
473 flow_loop_tree_node_remove (subloop
);
474 flow_loop_tree_node_add (loop
, subloop
);
478 /* Update the information about loop exit edges. */
479 for (i
= 0; i
< n
; i
++)
481 FOR_EACH_EDGE (e
, ei
, bbs
[i
]->succs
)
483 rescan_loop_exit (e
, false, false);
490 /* Scale profile of loop by P. */
493 scale_loop_frequencies (class loop
*loop
, profile_probability p
)
497 bbs
= get_loop_body (loop
);
498 scale_bbs_frequencies (bbs
, loop
->num_nodes
, p
);
502 /* Scale profile in LOOP by P.
503 If ITERATION_BOUND is not -1, scale even further if loop is predicted
504 to iterate too many times.
505 Before caling this function, preheader block profile should be already
506 scaled to final count. This is necessary because loop iterations are
507 determined by comparing header edge count to latch ege count and thus
508 they need to be scaled synchronously. */
511 scale_loop_profile (class loop
*loop
, profile_probability p
,
512 gcov_type iteration_bound
)
514 if (!(p
== profile_probability::always ()))
516 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
518 fprintf (dump_file
, ";; Scaling loop %i with scale ",
521 fprintf (dump_file
, "\n");
524 /* Scale the probabilities. */
525 scale_loop_frequencies (loop
, p
);
528 if (iteration_bound
== -1)
532 if (!expected_loop_iterations_by_profile (loop
, &iterations
))
535 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
538 ";; guessed iterations of loop %i:%f new upper bound %i:\n",
540 iterations
.to_double (),
541 (int)iteration_bound
);
544 /* See if loop is predicted to iterate too many times. */
545 if (iterations
<= iteration_bound
)
548 /* Compute number of invocations of the loop. */
549 profile_count count_in
= profile_count::zero ();
552 bool found_latch
= false;
553 FOR_EACH_EDGE (e
, ei
, loop
->header
->preds
)
554 if (e
->src
!= loop
->latch
)
555 count_in
+= e
->count ();
558 gcc_checking_assert (found_latch
);
560 /* Now scale the loop body so header count is
561 count_in * (iteration_bound + 1) */
562 profile_probability scale_prob
563 = (count_in
* (iteration_bound
+ 1)).probability_in (loop
->header
->count
);
564 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
566 fprintf (dump_file
, ";; Scaling loop %i with scale ",
568 scale_prob
.dump (dump_file
);
569 fprintf (dump_file
, " to reach upper bound %i\n",
570 (int)iteration_bound
);
572 /* Finally attempt to fix exit edge probability. */
573 auto_vec
<edge
> exits
= get_loop_exit_edges (loop
);
574 edge exit_edge
= single_likely_exit (loop
, exits
);
576 /* In a consistent profile unadjusted_exit_count should be same as count_in,
577 however to preserve as much of the original info, avoid recomputing
579 profile_count unadjusted_exit_count
;
581 unadjusted_exit_count
= exit_edge
->count ();
582 scale_loop_frequencies (loop
, scale_prob
);
584 if (exit_edge
&& exit_edge
->src
->loop_father
!= loop
)
586 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
588 ";; Loop exit is in inner loop;"
589 " will leave exit probabilities inconsistent\n");
593 profile_count old_exit_count
= exit_edge
->count ();
594 profile_probability new_probability
;
595 if (iteration_bound
> 0)
597 /* It may happen that the source basic block of the exit edge is
598 inside in-loop condition:
609 If B2 count is smaller than desired exit edge count
610 the profile was inconsistent with the newly discovered upper bound.
611 Probablity of edge B1->B2 is too low. We do not attempt to fix
612 that (as it is hard in general) but we want to avoid dropping
613 count of edge B2->B3 to zero may confuse later optimizations. */
614 if (unadjusted_exit_count
.apply_scale (7, 8) > exit_edge
->src
->count
)
616 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
618 ";; Source basic block of loop exit count is too small;"
619 " will leave exit probabilities inconsistent\n");
620 exit_edge
->probability
= exit_edge
->probability
.guessed ();
624 = unadjusted_exit_count
.probability_in (exit_edge
->src
->count
);
627 new_probability
= profile_probability::always ();
628 set_edge_probability_and_rescale_others (exit_edge
, new_probability
);
629 profile_count new_exit_count
= exit_edge
->count ();
631 /* Rescale the remaining edge probabilities and see if there is only
633 edge other_edge
= NULL
;
635 FOR_EACH_EDGE (e
, ei
, exit_edge
->src
->succs
)
636 if (!(e
->flags
& EDGE_FAKE
)
637 && !loop_exit_edge_p (loop
, e
))
647 /* If there is only loop latch after other edge,
648 update its profile. */
649 if (other_edge
&& other_edge
->dest
== loop
->latch
)
650 loop
->latch
->count
-= new_exit_count
- old_exit_count
;
653 basic_block
*body
= get_loop_body (loop
);
654 profile_count new_count
= exit_edge
->src
->count
- new_exit_count
;
655 profile_count old_count
= exit_edge
->src
->count
- old_exit_count
;
657 for (unsigned int i
= 0; i
< loop
->num_nodes
; i
++)
658 if (body
[i
] != exit_edge
->src
659 && dominated_by_p (CDI_DOMINATORS
, body
[i
], exit_edge
->src
))
660 body
[i
]->count
= body
[i
]->count
.apply_scale (new_count
,
666 else if (dump_file
&& (dump_flags
& TDF_DETAILS
))
669 ";; Loop has mulitple exits;"
670 " will leave exit probabilities inconsistent\n");
674 /* Recompute dominance information for basic blocks outside LOOP. */
677 update_dominators_in_loop (class loop
*loop
)
679 vec
<basic_block
> dom_bbs
= vNULL
;
683 auto_sbitmap
seen (last_basic_block_for_fn (cfun
));
685 body
= get_loop_body (loop
);
687 for (i
= 0; i
< loop
->num_nodes
; i
++)
688 bitmap_set_bit (seen
, body
[i
]->index
);
690 for (i
= 0; i
< loop
->num_nodes
; i
++)
694 for (ldom
= first_dom_son (CDI_DOMINATORS
, body
[i
]);
696 ldom
= next_dom_son (CDI_DOMINATORS
, ldom
))
697 if (!bitmap_bit_p (seen
, ldom
->index
))
699 bitmap_set_bit (seen
, ldom
->index
);
700 dom_bbs
.safe_push (ldom
);
704 iterate_fix_dominators (CDI_DOMINATORS
, dom_bbs
, false);
709 /* Creates an if region as shown above. CONDITION is used to create
713 | ------------- -------------
714 | | pred_bb | | pred_bb |
715 | ------------- -------------
719 | | ====> -------------
724 | ------------- e_false / \ e_true
726 | ------------- ----------- -----------
727 | | false_bb | | true_bb |
728 | ----------- -----------
735 | | exit_edge (result)
744 create_empty_if_region_on_edge (edge entry_edge
, tree condition
)
747 basic_block cond_bb
, true_bb
, false_bb
, join_bb
;
748 edge e_true
, e_false
, exit_edge
;
751 gimple_stmt_iterator gsi
;
753 cond_bb
= split_edge (entry_edge
);
755 /* Insert condition in cond_bb. */
756 gsi
= gsi_last_bb (cond_bb
);
758 force_gimple_operand_gsi (&gsi
, condition
, true, NULL
,
759 false, GSI_NEW_STMT
);
760 cond_stmt
= gimple_build_cond_from_tree (simple_cond
, NULL_TREE
, NULL_TREE
);
761 gsi
= gsi_last_bb (cond_bb
);
762 gsi_insert_after (&gsi
, cond_stmt
, GSI_NEW_STMT
);
764 join_bb
= split_edge (single_succ_edge (cond_bb
));
766 e_true
= single_succ_edge (cond_bb
);
767 true_bb
= split_edge (e_true
);
769 e_false
= make_edge (cond_bb
, join_bb
, 0);
770 false_bb
= split_edge (e_false
);
772 e_true
->flags
&= ~EDGE_FALLTHRU
;
773 e_true
->flags
|= EDGE_TRUE_VALUE
;
774 e_false
->flags
&= ~EDGE_FALLTHRU
;
775 e_false
->flags
|= EDGE_FALSE_VALUE
;
777 set_immediate_dominator (CDI_DOMINATORS
, cond_bb
, entry_edge
->src
);
778 set_immediate_dominator (CDI_DOMINATORS
, true_bb
, cond_bb
);
779 set_immediate_dominator (CDI_DOMINATORS
, false_bb
, cond_bb
);
780 set_immediate_dominator (CDI_DOMINATORS
, join_bb
, cond_bb
);
782 exit_edge
= single_succ_edge (join_bb
);
784 if (single_pred_p (exit_edge
->dest
))
785 set_immediate_dominator (CDI_DOMINATORS
, exit_edge
->dest
, join_bb
);
790 /* create_empty_loop_on_edge
792 | - pred_bb - ------ pred_bb ------
793 | | | | iv0 = initial_value |
794 | -----|----- ---------|-----------
795 | | ______ | entry_edge
797 | | ====> | -V---V- loop_header -------------
798 | V | | iv_before = phi (iv0, iv_after) |
799 | - succ_bb - | ---|-----------------------------
801 | ----------- | ---V--- loop_body ---------------
802 | | | iv_after = iv_before + stride |
803 | | | if (iv_before < upper_bound) |
804 | | ---|--------------\--------------
807 | | - loop_latch - V- succ_bb -
809 | | /------------- -----------
812 Creates an empty loop as shown above, the IV_BEFORE is the SSA_NAME
813 that is used before the increment of IV. IV_BEFORE should be used for
814 adding code to the body that uses the IV. OUTER is the outer loop in
815 which the new loop should be inserted.
817 Both INITIAL_VALUE and UPPER_BOUND expressions are gimplified and
818 inserted on the loop entry edge. This implies that this function
819 should be used only when the UPPER_BOUND expression is a loop
823 create_empty_loop_on_edge (edge entry_edge
,
825 tree stride
, tree upper_bound
,
831 basic_block loop_header
, loop_latch
, succ_bb
, pred_bb
;
833 gimple_stmt_iterator gsi
;
839 gcc_assert (entry_edge
&& initial_value
&& stride
&& upper_bound
&& iv
);
841 /* Create header, latch and wire up the loop. */
842 pred_bb
= entry_edge
->src
;
843 loop_header
= split_edge (entry_edge
);
844 loop_latch
= split_edge (single_succ_edge (loop_header
));
845 succ_bb
= single_succ (loop_latch
);
846 make_edge (loop_header
, succ_bb
, 0);
847 redirect_edge_succ_nodup (single_succ_edge (loop_latch
), loop_header
);
849 /* Set immediate dominator information. */
850 set_immediate_dominator (CDI_DOMINATORS
, loop_header
, pred_bb
);
851 set_immediate_dominator (CDI_DOMINATORS
, loop_latch
, loop_header
);
852 set_immediate_dominator (CDI_DOMINATORS
, succ_bb
, loop_header
);
854 /* Initialize a loop structure and put it in a loop hierarchy. */
855 loop
= alloc_loop ();
856 loop
->header
= loop_header
;
857 loop
->latch
= loop_latch
;
858 add_loop (loop
, outer
);
860 /* TODO: Fix counts. */
861 scale_loop_frequencies (loop
, profile_probability::even ());
863 /* Update dominators. */
864 update_dominators_in_loop (loop
);
866 /* Modify edge flags. */
867 exit_e
= single_exit (loop
);
868 exit_e
->flags
= EDGE_LOOP_EXIT
| EDGE_FALSE_VALUE
;
869 single_pred_edge (loop_latch
)->flags
= EDGE_TRUE_VALUE
;
871 /* Construct IV code in loop. */
872 initial_value
= force_gimple_operand (initial_value
, &stmts
, true, iv
);
875 gsi_insert_seq_on_edge (loop_preheader_edge (loop
), stmts
);
876 gsi_commit_edge_inserts ();
879 upper_bound
= force_gimple_operand (upper_bound
, &stmts
, true, NULL
);
882 gsi_insert_seq_on_edge (loop_preheader_edge (loop
), stmts
);
883 gsi_commit_edge_inserts ();
886 gsi
= gsi_last_bb (loop_header
);
887 create_iv (initial_value
, PLUS_EXPR
, stride
, iv
, loop
, &gsi
, false,
888 iv_before
, iv_after
);
890 /* Insert loop exit condition. */
891 cond_expr
= gimple_build_cond
892 (LT_EXPR
, *iv_before
, upper_bound
, NULL_TREE
, NULL_TREE
);
894 exit_test
= gimple_cond_lhs (cond_expr
);
895 exit_test
= force_gimple_operand_gsi (&gsi
, exit_test
, true, NULL
,
896 false, GSI_NEW_STMT
);
897 gimple_cond_set_lhs (cond_expr
, exit_test
);
898 gsi
= gsi_last_bb (exit_e
->src
);
899 gsi_insert_after (&gsi
, cond_expr
, GSI_NEW_STMT
);
901 split_block_after_labels (loop_header
);
906 /* Remove the latch edge of a LOOP and update loops to indicate that
907 the LOOP was removed. After this function, original loop latch will
908 have no successor, which caller is expected to fix somehow.
910 If this may cause the information about irreducible regions to become
911 invalid, IRRED_INVALIDATED is set to true.
913 LOOP_CLOSED_SSA_INVALIDATED, if non-NULL, is a bitmap where we store
914 basic blocks that had non-trivial update on their loop_father.*/
917 unloop (class loop
*loop
, bool *irred_invalidated
,
918 bitmap loop_closed_ssa_invalidated
)
923 basic_block latch
= loop
->latch
;
926 if (loop_preheader_edge (loop
)->flags
& EDGE_IRREDUCIBLE_LOOP
)
927 *irred_invalidated
= true;
929 /* This is relatively straightforward. The dominators are unchanged, as
930 loop header dominates loop latch, so the only thing we have to care of
931 is the placement of loops and basic blocks inside the loop tree. We
932 move them all to the loop->outer, and then let fix_bb_placements do
935 body
= get_loop_body (loop
);
937 for (i
= 0; i
< n
; i
++)
938 if (body
[i
]->loop_father
== loop
)
940 remove_bb_from_loops (body
[i
]);
941 add_bb_to_loop (body
[i
], loop_outer (loop
));
948 flow_loop_tree_node_remove (ploop
);
949 flow_loop_tree_node_add (loop_outer (loop
), ploop
);
952 /* Remove the loop and free its data. */
955 remove_edge (single_succ_edge (latch
));
957 /* We do not pass IRRED_INVALIDATED to fix_bb_placements here, as even if
958 there is an irreducible region inside the cancelled loop, the flags will
960 fix_bb_placements (latch
, &dummy
, loop_closed_ssa_invalidated
);
963 /* Fix placement of superloops of LOOP inside loop tree, i.e. ensure that
964 condition stated in description of fix_loop_placement holds for them.
965 It is used in case when we removed some edges coming out of LOOP, which
966 may cause the right placement of LOOP inside loop tree to change.
968 IRRED_INVALIDATED is set to true if a change in the loop structures might
969 invalidate the information about irreducible regions. */
972 fix_loop_placements (class loop
*loop
, bool *irred_invalidated
)
976 while (loop_outer (loop
))
978 outer
= loop_outer (loop
);
979 if (!fix_loop_placement (loop
, irred_invalidated
))
982 /* Changing the placement of a loop in the loop tree may alter the
983 validity of condition 2) of the description of fix_bb_placement
984 for its preheader, because the successor is the header and belongs
985 to the loop. So call fix_bb_placements to fix up the placement
986 of the preheader and (possibly) of its predecessors. */
987 fix_bb_placements (loop_preheader_edge (loop
)->src
,
988 irred_invalidated
, NULL
);
993 /* Duplicate loop bounds and other information we store about
994 the loop into its duplicate. */
997 copy_loop_info (class loop
*loop
, class loop
*target
)
999 gcc_checking_assert (!target
->any_upper_bound
&& !target
->any_estimate
);
1000 target
->any_upper_bound
= loop
->any_upper_bound
;
1001 target
->nb_iterations_upper_bound
= loop
->nb_iterations_upper_bound
;
1002 target
->any_likely_upper_bound
= loop
->any_likely_upper_bound
;
1003 target
->nb_iterations_likely_upper_bound
1004 = loop
->nb_iterations_likely_upper_bound
;
1005 target
->any_estimate
= loop
->any_estimate
;
1006 target
->nb_iterations_estimate
= loop
->nb_iterations_estimate
;
1007 target
->estimate_state
= loop
->estimate_state
;
1008 target
->safelen
= loop
->safelen
;
1009 target
->simdlen
= loop
->simdlen
;
1010 target
->constraints
= loop
->constraints
;
1011 target
->can_be_parallel
= loop
->can_be_parallel
;
1012 target
->warned_aggressive_loop_optimizations
1013 |= loop
->warned_aggressive_loop_optimizations
;
1014 target
->dont_vectorize
= loop
->dont_vectorize
;
1015 target
->force_vectorize
= loop
->force_vectorize
;
1016 target
->in_oacc_kernels_region
= loop
->in_oacc_kernels_region
;
1017 target
->finite_p
= loop
->finite_p
;
1018 target
->unroll
= loop
->unroll
;
1019 target
->owned_clique
= loop
->owned_clique
;
1022 /* Copies copy of LOOP as subloop of TARGET loop, placing newly
1023 created loop into loops structure. If AFTER is non-null
1024 the new loop is added at AFTER->next, otherwise in front of TARGETs
1027 duplicate_loop (class loop
*loop
, class loop
*target
, class loop
*after
)
1030 cloop
= alloc_loop ();
1031 place_new_loop (cfun
, cloop
);
1033 copy_loop_info (loop
, cloop
);
1035 /* Mark the new loop as copy of LOOP. */
1036 set_loop_copy (loop
, cloop
);
1038 /* Add it to target. */
1039 flow_loop_tree_node_add (target
, cloop
, after
);
1044 /* Copies structure of subloops of LOOP into TARGET loop, placing
1045 newly created loops into loop tree at the end of TARGETs sibling
1046 list in the original order. */
1048 duplicate_subloops (class loop
*loop
, class loop
*target
)
1050 class loop
*aloop
, *cloop
, *tail
;
1052 for (tail
= target
->inner
; tail
&& tail
->next
; tail
= tail
->next
)
1054 for (aloop
= loop
->inner
; aloop
; aloop
= aloop
->next
)
1056 cloop
= duplicate_loop (aloop
, target
, tail
);
1058 gcc_assert(!tail
->next
);
1059 duplicate_subloops (aloop
, cloop
);
1063 /* Copies structure of subloops of N loops, stored in array COPIED_LOOPS,
1064 into TARGET loop, placing newly created loops into loop tree adding
1065 them to TARGETs sibling list at the end in order. */
1067 copy_loops_to (class loop
**copied_loops
, int n
, class loop
*target
)
1069 class loop
*aloop
, *tail
;
1072 for (tail
= target
->inner
; tail
&& tail
->next
; tail
= tail
->next
)
1074 for (i
= 0; i
< n
; i
++)
1076 aloop
= duplicate_loop (copied_loops
[i
], target
, tail
);
1078 gcc_assert(!tail
->next
);
1079 duplicate_subloops (copied_loops
[i
], aloop
);
1083 /* Redirects edge E to basic block DEST. */
1085 loop_redirect_edge (edge e
, basic_block dest
)
1087 if (e
->dest
== dest
)
1090 redirect_edge_and_branch_force (e
, dest
);
1093 /* Check whether LOOP's body can be duplicated. */
1095 can_duplicate_loop_p (const class loop
*loop
)
1098 basic_block
*bbs
= get_loop_body (loop
);
1100 ret
= can_copy_bbs_p (bbs
, loop
->num_nodes
);
1106 /* Duplicates body of LOOP to given edge E NDUPL times. Takes care of updating
1107 loop structure and dominators (order of inner subloops is retained).
1108 E's destination must be LOOP header for this to work, i.e. it must be entry
1109 or latch edge of this loop; these are unique, as the loops must have
1110 preheaders for this function to work correctly (in case E is latch, the
1111 function unrolls the loop, if E is entry edge, it peels the loop). Store
1112 edges created by copying ORIG edge from copies corresponding to set bits in
1113 WONT_EXIT bitmap (bit 0 corresponds to original LOOP body, the other copies
1114 are numbered in order given by control flow through them) into TO_REMOVE
1115 array. Returns false if duplication is
1119 duplicate_loop_body_to_header_edge (class loop
*loop
, edge e
,
1120 unsigned int ndupl
, sbitmap wont_exit
,
1121 edge orig
, vec
<edge
> *to_remove
, int flags
)
1123 class loop
*target
, *aloop
;
1124 class loop
**orig_loops
;
1125 unsigned n_orig_loops
;
1126 basic_block header
= loop
->header
, latch
= loop
->latch
;
1127 basic_block
*new_bbs
, *bbs
, *first_active
;
1128 basic_block new_bb
, bb
, first_active_latch
= NULL
;
1129 edge ae
, latch_edge
;
1130 edge spec_edges
[2], new_spec_edges
[2];
1131 const int SE_LATCH
= 0;
1132 const int SE_ORIG
= 1;
1134 int is_latch
= (latch
== e
->src
);
1135 profile_probability
*scale_step
= NULL
;
1136 profile_probability scale_main
= profile_probability::always ();
1137 profile_probability scale_act
= profile_probability::always ();
1138 profile_count after_exit_num
= profile_count::zero (),
1139 after_exit_den
= profile_count::zero ();
1140 bool scale_after_exit
= false;
1141 int add_irreducible_flag
;
1142 basic_block place_after
;
1143 bitmap bbs_to_scale
= NULL
;
1146 gcc_assert (e
->dest
== loop
->header
);
1147 gcc_assert (ndupl
> 0);
1151 /* Orig must be edge out of the loop. */
1152 gcc_assert (flow_bb_inside_loop_p (loop
, orig
->src
));
1153 gcc_assert (!flow_bb_inside_loop_p (loop
, orig
->dest
));
1156 n
= loop
->num_nodes
;
1157 bbs
= get_loop_body_in_dom_order (loop
);
1158 gcc_assert (bbs
[0] == loop
->header
);
1159 gcc_assert (bbs
[n
- 1] == loop
->latch
);
1161 /* Check whether duplication is possible. */
1162 if (!can_copy_bbs_p (bbs
, loop
->num_nodes
))
1167 new_bbs
= XNEWVEC (basic_block
, loop
->num_nodes
);
1169 /* In case we are doing loop peeling and the loop is in the middle of
1170 irreducible region, the peeled copies will be inside it too. */
1171 add_irreducible_flag
= e
->flags
& EDGE_IRREDUCIBLE_LOOP
;
1172 gcc_assert (!is_latch
|| !add_irreducible_flag
);
1174 /* Find edge from latch. */
1175 latch_edge
= loop_latch_edge (loop
);
1177 if (flags
& DLTHE_FLAG_UPDATE_FREQ
)
1179 /* Calculate coefficients by that we have to scale counts
1180 of duplicated loop bodies. */
1181 profile_count count_in
= header
->count
;
1182 profile_count count_le
= latch_edge
->count ();
1183 profile_count count_out_orig
= orig
? orig
->count () : count_in
- count_le
;
1184 profile_probability prob_pass_thru
= count_le
.probability_in (count_in
);
1185 profile_count new_count_le
= count_le
+ count_out_orig
;
1187 if (orig
&& orig
->probability
.initialized_p ()
1188 && !(orig
->probability
== profile_probability::always ()))
1190 /* The blocks that are dominated by a removed exit edge ORIG have
1191 frequencies scaled by this. */
1192 if (orig
->count ().initialized_p ())
1194 after_exit_num
= orig
->src
->count
;
1195 after_exit_den
= after_exit_num
- orig
->count ();
1196 scale_after_exit
= true;
1198 bbs_to_scale
= BITMAP_ALLOC (NULL
);
1199 for (i
= 0; i
< n
; i
++)
1201 if (bbs
[i
] != orig
->src
1202 && dominated_by_p (CDI_DOMINATORS
, bbs
[i
], orig
->src
))
1203 bitmap_set_bit (bbs_to_scale
, i
);
1205 /* Since we will scale up all basic blocks dominated by orig, exits
1206 will become more likely; compensate for that. */
1207 if (after_exit_den
.nonzero_p ())
1209 auto_vec
<edge
> exits
= get_loop_exit_edges (loop
);
1210 for (edge ex
: exits
)
1212 && dominated_by_p (CDI_DOMINATORS
, ex
->src
, orig
->src
))
1213 new_count_le
-= ex
->count ().apply_scale (after_exit_num
1218 profile_probability prob_pass_wont_exit
=
1219 new_count_le
.probability_in (count_in
);
1221 scale_step
= XNEWVEC (profile_probability
, ndupl
);
1223 for (i
= 1; i
<= ndupl
; i
++)
1224 scale_step
[i
- 1] = bitmap_bit_p (wont_exit
, i
)
1225 ? prob_pass_wont_exit
1228 /* Complete peeling is special as the probability of exit in last
1230 if (flags
& DLTHE_FLAG_COMPLETTE_PEEL
)
1232 profile_count wanted_count
= e
->count ();
1234 gcc_assert (!is_latch
);
1235 /* First copy has count of incoming edge. Each subsequent
1236 count should be reduced by prob_pass_wont_exit. Caller
1237 should've managed the flags so all except for original loop
1238 has won't exist set. */
1239 scale_act
= wanted_count
.probability_in (count_in
);
1240 /* Now simulate the duplication adjustments and compute header
1241 frequency of the last copy. */
1242 for (i
= 0; i
< ndupl
; i
++)
1243 wanted_count
= wanted_count
.apply_probability (scale_step
[i
]);
1244 scale_main
= wanted_count
.probability_in (count_in
);
1246 /* Here we insert loop bodies inside the loop itself (for loop unrolling).
1247 First iteration will be original loop followed by duplicated bodies.
1248 It is necessary to scale down the original so we get right overall
1249 number of iterations. */
1252 profile_probability prob_pass_main
= bitmap_bit_p (wont_exit
, 0)
1253 ? prob_pass_wont_exit
1255 profile_probability p
= prob_pass_main
;
1256 profile_count scale_main_den
= count_in
;
1257 for (i
= 0; i
< ndupl
; i
++)
1259 scale_main_den
+= count_in
.apply_probability (p
);
1260 p
= p
* scale_step
[i
];
1262 /* If original loop is executed COUNT_IN times, the unrolled
1263 loop will account SCALE_MAIN_DEN times. */
1264 scale_main
= count_in
.probability_in (scale_main_den
);
1265 scale_act
= scale_main
* prob_pass_main
;
1269 profile_count preheader_count
= e
->count ();
1270 for (i
= 0; i
< ndupl
; i
++)
1271 scale_main
= scale_main
* scale_step
[i
];
1272 scale_act
= preheader_count
.probability_in (count_in
);
1276 /* Loop the new bbs will belong to. */
1277 target
= e
->src
->loop_father
;
1279 /* Original loops. */
1281 for (aloop
= loop
->inner
; aloop
; aloop
= aloop
->next
)
1283 orig_loops
= XNEWVEC (class loop
*, n_orig_loops
);
1284 for (aloop
= loop
->inner
, i
= 0; aloop
; aloop
= aloop
->next
, i
++)
1285 orig_loops
[i
] = aloop
;
1287 set_loop_copy (loop
, target
);
1289 first_active
= XNEWVEC (basic_block
, n
);
1292 memcpy (first_active
, bbs
, n
* sizeof (basic_block
));
1293 first_active_latch
= latch
;
1296 spec_edges
[SE_ORIG
] = orig
;
1297 spec_edges
[SE_LATCH
] = latch_edge
;
1299 place_after
= e
->src
;
1300 for (j
= 0; j
< ndupl
; j
++)
1303 copy_loops_to (orig_loops
, n_orig_loops
, target
);
1306 copy_bbs (bbs
, n
, new_bbs
, spec_edges
, 2, new_spec_edges
, loop
,
1308 place_after
= new_spec_edges
[SE_LATCH
]->src
;
1310 if (flags
& DLTHE_RECORD_COPY_NUMBER
)
1311 for (i
= 0; i
< n
; i
++)
1313 gcc_assert (!new_bbs
[i
]->aux
);
1314 new_bbs
[i
]->aux
= (void *)(size_t)(j
+ 1);
1317 /* Note whether the blocks and edges belong to an irreducible loop. */
1318 if (add_irreducible_flag
)
1320 for (i
= 0; i
< n
; i
++)
1321 new_bbs
[i
]->flags
|= BB_DUPLICATED
;
1322 for (i
= 0; i
< n
; i
++)
1325 new_bb
= new_bbs
[i
];
1326 if (new_bb
->loop_father
== target
)
1327 new_bb
->flags
|= BB_IRREDUCIBLE_LOOP
;
1329 FOR_EACH_EDGE (ae
, ei
, new_bb
->succs
)
1330 if ((ae
->dest
->flags
& BB_DUPLICATED
)
1331 && (ae
->src
->loop_father
== target
1332 || ae
->dest
->loop_father
== target
))
1333 ae
->flags
|= EDGE_IRREDUCIBLE_LOOP
;
1335 for (i
= 0; i
< n
; i
++)
1336 new_bbs
[i
]->flags
&= ~BB_DUPLICATED
;
1339 /* Redirect the special edges. */
1342 redirect_edge_and_branch_force (latch_edge
, new_bbs
[0]);
1343 redirect_edge_and_branch_force (new_spec_edges
[SE_LATCH
],
1345 set_immediate_dominator (CDI_DOMINATORS
, new_bbs
[0], latch
);
1346 latch
= loop
->latch
= new_bbs
[n
- 1];
1347 e
= latch_edge
= new_spec_edges
[SE_LATCH
];
1351 redirect_edge_and_branch_force (new_spec_edges
[SE_LATCH
],
1353 redirect_edge_and_branch_force (e
, new_bbs
[0]);
1354 set_immediate_dominator (CDI_DOMINATORS
, new_bbs
[0], e
->src
);
1355 e
= new_spec_edges
[SE_LATCH
];
1358 /* Record exit edge in this copy. */
1359 if (orig
&& bitmap_bit_p (wont_exit
, j
+ 1))
1362 to_remove
->safe_push (new_spec_edges
[SE_ORIG
]);
1363 force_edge_cold (new_spec_edges
[SE_ORIG
], true);
1365 /* Scale the frequencies of the blocks dominated by the exit. */
1366 if (bbs_to_scale
&& scale_after_exit
)
1368 EXECUTE_IF_SET_IN_BITMAP (bbs_to_scale
, 0, i
, bi
)
1369 scale_bbs_frequencies_profile_count (new_bbs
+ i
, 1, after_exit_num
,
1374 /* Record the first copy in the control flow order if it is not
1375 the original loop (i.e. in case of peeling). */
1376 if (!first_active_latch
)
1378 memcpy (first_active
, new_bbs
, n
* sizeof (basic_block
));
1379 first_active_latch
= new_bbs
[n
- 1];
1382 /* Set counts and frequencies. */
1383 if (flags
& DLTHE_FLAG_UPDATE_FREQ
)
1385 scale_bbs_frequencies (new_bbs
, n
, scale_act
);
1386 scale_act
= scale_act
* scale_step
[j
];
1392 /* Record the exit edge in the original loop body, and update the frequencies. */
1393 if (orig
&& bitmap_bit_p (wont_exit
, 0))
1396 to_remove
->safe_push (orig
);
1397 force_edge_cold (orig
, true);
1399 /* Scale the frequencies of the blocks dominated by the exit. */
1400 if (bbs_to_scale
&& scale_after_exit
)
1402 EXECUTE_IF_SET_IN_BITMAP (bbs_to_scale
, 0, i
, bi
)
1403 scale_bbs_frequencies_profile_count (bbs
+ i
, 1, after_exit_num
,
1408 /* Update the original loop. */
1410 set_immediate_dominator (CDI_DOMINATORS
, e
->dest
, e
->src
);
1411 if (flags
& DLTHE_FLAG_UPDATE_FREQ
)
1413 scale_bbs_frequencies (bbs
, n
, scale_main
);
1417 /* Update dominators of outer blocks if affected. */
1418 for (i
= 0; i
< n
; i
++)
1420 basic_block dominated
, dom_bb
;
1425 auto_vec
<basic_block
> dom_bbs
= get_dominated_by (CDI_DOMINATORS
, bb
);
1426 FOR_EACH_VEC_ELT (dom_bbs
, j
, dominated
)
1428 if (flow_bb_inside_loop_p (loop
, dominated
))
1430 dom_bb
= nearest_common_dominator (
1431 CDI_DOMINATORS
, first_active
[i
], first_active_latch
);
1432 set_immediate_dominator (CDI_DOMINATORS
, dominated
, dom_bb
);
1435 free (first_active
);
1438 BITMAP_FREE (bbs_to_scale
);
1443 /* A callback for make_forwarder block, to redirect all edges except for
1444 MFB_KJ_EDGE to the entry part. E is the edge for that we should decide
1445 whether to redirect it. */
1449 mfb_keep_just (edge e
)
1451 return e
!= mfb_kj_edge
;
1454 /* True when a candidate preheader BLOCK has predecessors from LOOP. */
1457 has_preds_from_loop (basic_block block
, class loop
*loop
)
1462 FOR_EACH_EDGE (e
, ei
, block
->preds
)
1463 if (e
->src
->loop_father
== loop
)
1468 /* Creates a pre-header for a LOOP. Returns newly created block. Unless
1469 CP_SIMPLE_PREHEADERS is set in FLAGS, we only force LOOP to have single
1470 entry; otherwise we also force preheader block to have only one successor.
1471 When CP_FALLTHRU_PREHEADERS is set in FLAGS, we force the preheader block
1472 to be a fallthru predecessor to the loop header and to have only
1473 predecessors from outside of the loop.
1474 The function also updates dominators. */
1477 create_preheader (class loop
*loop
, int flags
)
1483 bool latch_edge_was_fallthru
;
1484 edge one_succ_pred
= NULL
, single_entry
= NULL
;
1487 FOR_EACH_EDGE (e
, ei
, loop
->header
->preds
)
1489 if (e
->src
== loop
->latch
)
1491 irred
|= (e
->flags
& EDGE_IRREDUCIBLE_LOOP
) != 0;
1494 if (single_succ_p (e
->src
))
1497 gcc_assert (nentry
);
1500 bool need_forwarder_block
= false;
1502 /* We do not allow entry block to be the loop preheader, since we
1503 cannot emit code there. */
1504 if (single_entry
->src
== ENTRY_BLOCK_PTR_FOR_FN (cfun
))
1505 need_forwarder_block
= true;
1508 /* If we want simple preheaders, also force the preheader to have
1509 just a single successor and a normal edge. */
1510 if ((flags
& CP_SIMPLE_PREHEADERS
)
1511 && ((single_entry
->flags
& EDGE_COMPLEX
)
1512 || !single_succ_p (single_entry
->src
)))
1513 need_forwarder_block
= true;
1514 /* If we want fallthru preheaders, also create forwarder block when
1515 preheader ends with a jump or has predecessors from loop. */
1516 else if ((flags
& CP_FALLTHRU_PREHEADERS
)
1517 && (JUMP_P (BB_END (single_entry
->src
))
1518 || has_preds_from_loop (single_entry
->src
, loop
)))
1519 need_forwarder_block
= true;
1521 if (! need_forwarder_block
)
1525 mfb_kj_edge
= loop_latch_edge (loop
);
1526 latch_edge_was_fallthru
= (mfb_kj_edge
->flags
& EDGE_FALLTHRU
) != 0;
1528 && ((flags
& CP_FALLTHRU_PREHEADERS
) == 0
1529 || (single_entry
->flags
& EDGE_CROSSING
) == 0))
1530 dummy
= split_edge (single_entry
);
1533 edge fallthru
= make_forwarder_block (loop
->header
, mfb_keep_just
, NULL
);
1534 dummy
= fallthru
->src
;
1535 loop
->header
= fallthru
->dest
;
1538 /* Try to be clever in placing the newly created preheader. The idea is to
1539 avoid breaking any "fallthruness" relationship between blocks.
1541 The preheader was created just before the header and all incoming edges
1542 to the header were redirected to the preheader, except the latch edge.
1543 So the only problematic case is when this latch edge was a fallthru
1544 edge: it is not anymore after the preheader creation so we have broken
1545 the fallthruness. We're therefore going to look for a better place. */
1546 if (latch_edge_was_fallthru
)
1551 e
= EDGE_PRED (dummy
, 0);
1553 move_block_after (dummy
, e
->src
);
1558 dummy
->flags
|= BB_IRREDUCIBLE_LOOP
;
1559 single_succ_edge (dummy
)->flags
|= EDGE_IRREDUCIBLE_LOOP
;
1563 fprintf (dump_file
, "Created preheader block for loop %i\n",
1566 if (flags
& CP_FALLTHRU_PREHEADERS
)
1567 gcc_assert ((single_succ_edge (dummy
)->flags
& EDGE_FALLTHRU
)
1568 && !JUMP_P (BB_END (dummy
)));
1573 /* Create preheaders for each loop; for meaning of FLAGS see create_preheader. */
1576 create_preheaders (int flags
)
1581 for (auto loop
: loops_list (cfun
, 0))
1582 create_preheader (loop
, flags
);
1583 loops_state_set (LOOPS_HAVE_PREHEADERS
);
1586 /* Forces all loop latches to have only single successor. */
1589 force_single_succ_latches (void)
1593 for (auto loop
: loops_list (cfun
, 0))
1595 if (loop
->latch
!= loop
->header
&& single_succ_p (loop
->latch
))
1598 e
= find_edge (loop
->latch
, loop
->header
);
1599 gcc_checking_assert (e
!= NULL
);
1603 loops_state_set (LOOPS_HAVE_SIMPLE_LATCHES
);
1606 /* This function is called from loop_version. It splits the entry edge
1607 of the loop we want to version, adds the versioning condition, and
1608 adjust the edges to the two versions of the loop appropriately.
1609 e is an incoming edge. Returns the basic block containing the
1612 --- edge e ---- > [second_head]
1614 Split it and insert new conditional expression and adjust edges.
1616 --- edge e ---> [cond expr] ---> [first_head]
1618 +---------> [second_head]
1620 THEN_PROB is the probability of then branch of the condition.
1621 ELSE_PROB is the probability of else branch. Note that they may be both
1622 REG_BR_PROB_BASE when condition is IFN_LOOP_VECTORIZED or
1623 IFN_LOOP_DIST_ALIAS. */
1626 lv_adjust_loop_entry_edge (basic_block first_head
, basic_block second_head
,
1627 edge e
, void *cond_expr
,
1628 profile_probability then_prob
,
1629 profile_probability else_prob
)
1631 basic_block new_head
= NULL
;
1634 gcc_assert (e
->dest
== second_head
);
1636 /* Split edge 'e'. This will create a new basic block, where we can
1637 insert conditional expr. */
1638 new_head
= split_edge (e
);
1640 lv_add_condition_to_bb (first_head
, second_head
, new_head
,
1643 /* Don't set EDGE_TRUE_VALUE in RTL mode, as it's invalid there. */
1644 e
= single_succ_edge (new_head
);
1645 e1
= make_edge (new_head
, first_head
,
1646 current_ir_type () == IR_GIMPLE
? EDGE_TRUE_VALUE
: 0);
1647 e1
->probability
= then_prob
;
1648 e
->probability
= else_prob
;
1650 set_immediate_dominator (CDI_DOMINATORS
, first_head
, new_head
);
1651 set_immediate_dominator (CDI_DOMINATORS
, second_head
, new_head
);
1653 /* Adjust loop header phi nodes. */
1654 lv_adjust_loop_header_phi (first_head
, second_head
, new_head
, e1
);
1659 /* Main entry point for Loop Versioning transformation.
1661 This transformation given a condition and a loop, creates
1662 -if (condition) { loop_copy1 } else { loop_copy2 },
1663 where loop_copy1 is the loop transformed in one way, and loop_copy2
1664 is the loop transformed in another way (or unchanged). COND_EXPR
1665 may be a run time test for things that were not resolved by static
1666 analysis (overlapping ranges (anti-aliasing), alignment, etc.).
1668 If non-NULL, CONDITION_BB is set to the basic block containing the
1671 THEN_PROB is the probability of the then edge of the if. THEN_SCALE
1672 is the ratio by that the frequencies in the original loop should
1673 be scaled. ELSE_SCALE is the ratio by that the frequencies in the
1674 new loop should be scaled.
1676 If PLACE_AFTER is true, we place the new loop after LOOP in the
1677 instruction stream, otherwise it is placed before LOOP. */
1680 loop_version (class loop
*loop
,
1681 void *cond_expr
, basic_block
*condition_bb
,
1682 profile_probability then_prob
, profile_probability else_prob
,
1683 profile_probability then_scale
, profile_probability else_scale
,
1686 basic_block first_head
, second_head
;
1687 edge entry
, latch_edge
;
1690 basic_block cond_bb
;
1692 /* Record entry and latch edges for the loop */
1693 entry
= loop_preheader_edge (loop
);
1694 irred_flag
= entry
->flags
& EDGE_IRREDUCIBLE_LOOP
;
1695 entry
->flags
&= ~EDGE_IRREDUCIBLE_LOOP
;
1697 /* Note down head of loop as first_head. */
1698 first_head
= entry
->dest
;
1700 /* 1) Duplicate loop on the entry edge. */
1701 if (!cfg_hook_duplicate_loop_body_to_header_edge (loop
, entry
, 1, NULL
, NULL
,
1704 entry
->flags
|= irred_flag
;
1708 /* 2) loopify the duplicated new loop. */
1709 latch_edge
= single_succ_edge (get_bb_copy (loop
->latch
));
1710 nloop
= alloc_loop ();
1711 class loop
*outer
= loop_outer (latch_edge
->dest
->loop_father
);
1712 edge new_header_edge
= single_pred_edge (get_bb_copy (loop
->header
));
1713 nloop
->header
= new_header_edge
->dest
;
1714 nloop
->latch
= latch_edge
->src
;
1715 loop_redirect_edge (latch_edge
, nloop
->header
);
1717 /* Compute new loop. */
1718 add_loop (nloop
, outer
);
1719 copy_loop_info (loop
, nloop
);
1720 set_loop_copy (loop
, nloop
);
1722 /* loopify redirected latch_edge. Update its PENDING_STMTS. */
1723 lv_flush_pending_stmts (latch_edge
);
1725 /* After duplication entry edge now points to new loop head block.
1726 Note down new head as second_head. */
1727 second_head
= entry
->dest
;
1729 /* 3) Split loop entry edge and insert new block with cond expr. */
1730 cond_bb
= lv_adjust_loop_entry_edge (first_head
, second_head
,
1731 entry
, cond_expr
, then_prob
, else_prob
);
1733 *condition_bb
= cond_bb
;
1737 entry
->flags
|= irred_flag
;
1741 /* Add cond_bb to appropriate loop. */
1742 if (cond_bb
->loop_father
)
1743 remove_bb_from_loops (cond_bb
);
1744 add_bb_to_loop (cond_bb
, outer
);
1746 /* 4) Scale the original loop and new loop frequency. */
1747 scale_loop_frequencies (loop
, then_scale
);
1748 scale_loop_frequencies (nloop
, else_scale
);
1749 update_dominators_in_loop (loop
);
1750 update_dominators_in_loop (nloop
);
1752 /* Adjust irreducible flag. */
1755 cond_bb
->flags
|= BB_IRREDUCIBLE_LOOP
;
1756 loop_preheader_edge (loop
)->flags
|= EDGE_IRREDUCIBLE_LOOP
;
1757 loop_preheader_edge (nloop
)->flags
|= EDGE_IRREDUCIBLE_LOOP
;
1758 single_pred_edge (cond_bb
)->flags
|= EDGE_IRREDUCIBLE_LOOP
;
1763 basic_block
*bbs
= get_loop_body_in_dom_order (nloop
), after
;
1766 after
= loop
->latch
;
1768 for (i
= 0; i
< nloop
->num_nodes
; i
++)
1770 move_block_after (bbs
[i
], after
);
1776 /* At this point condition_bb is loop preheader with two successors,
1777 first_head and second_head. Make sure that loop preheader has only
1779 split_edge (loop_preheader_edge (loop
));
1780 split_edge (loop_preheader_edge (nloop
));