Mercurial > hg > CbC > CbC_gcc
annotate gcc/cfgloopmanip.c @ 59:5b5b9ea5b220
fix
author | Shinji KONO <kono@ie.u-ryukyu.ac.jp> |
---|---|
date | Mon, 15 Feb 2010 17:22:24 +0900 |
parents | 77e2b8dfacca |
children | f6334be47118 |
rev | line source |
---|---|
0 | 1 /* Loop manipulation code for GNU compiler. |
2 Copyright (C) 2002, 2003, 2004, 2005, 2007, 2008, 2009 Free Software | |
3 Foundation, Inc. | |
4 | |
5 This file is part of GCC. | |
6 | |
7 GCC is free software; you can redistribute it and/or modify it under | |
8 the terms of the GNU General Public License as published by the Free | |
9 Software Foundation; either version 3, or (at your option) any later | |
10 version. | |
11 | |
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
15 for more details. | |
16 | |
17 You should have received a copy of the GNU General Public License | |
18 along with GCC; see the file COPYING3. If not see | |
19 <http://www.gnu.org/licenses/>. */ | |
20 | |
21 #include "config.h" | |
22 #include "system.h" | |
23 #include "coretypes.h" | |
24 #include "tm.h" | |
25 #include "rtl.h" | |
26 #include "hard-reg-set.h" | |
27 #include "obstack.h" | |
28 #include "basic-block.h" | |
29 #include "cfgloop.h" | |
30 #include "cfglayout.h" | |
31 #include "cfghooks.h" | |
32 #include "output.h" | |
33 #include "tree-flow.h" | |
34 | |
35 static void copy_loops_to (struct loop **, int, | |
36 struct loop *); | |
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 *); | |
44 static void unloop (struct loop *, bool *); | |
45 | |
46 #define RDIV(X,Y) (((X) + (Y) / 2) / (Y)) | |
47 | |
48 /* Checks whether basic block BB is dominated by DATA. */ | |
49 static bool | |
50 rpe_enum_p (const_basic_block bb, const void *data) | |
51 { | |
52 return dominated_by_p (CDI_DOMINATORS, bb, (const_basic_block) data); | |
53 } | |
54 | |
55 /* Remove basic blocks BBS. NBBS is the number of the basic blocks. */ | |
56 | |
57 static void | |
58 remove_bbs (basic_block *bbs, int nbbs) | |
59 { | |
60 int i; | |
61 | |
62 for (i = 0; i < nbbs; i++) | |
63 delete_basic_block (bbs[i]); | |
64 } | |
65 | |
66 /* Find path -- i.e. the basic blocks dominated by edge E and put them | |
67 into array BBS, that will be allocated large enough to contain them. | |
68 E->dest must have exactly one predecessor for this to work (it is | |
69 easy to achieve and we do not put it here because we do not want to | |
70 alter anything by this function). The number of basic blocks in the | |
71 path is returned. */ | |
72 static int | |
73 find_path (edge e, basic_block **bbs) | |
74 { | |
75 gcc_assert (EDGE_COUNT (e->dest->preds) <= 1); | |
76 | |
77 /* Find bbs in the path. */ | |
78 *bbs = XCNEWVEC (basic_block, n_basic_blocks); | |
79 return dfs_enumerate_from (e->dest, 0, rpe_enum_p, *bbs, | |
80 n_basic_blocks, e->dest); | |
81 } | |
82 | |
83 /* Fix placement of basic block BB inside loop hierarchy -- | |
84 Let L be a loop to that BB belongs. Then every successor of BB must either | |
85 1) belong to some superloop of loop L, or | |
86 2) be a header of loop K such that K->outer is superloop of L | |
87 Returns true if we had to move BB into other loop to enforce this condition, | |
88 false if the placement of BB was already correct (provided that placements | |
89 of its successors are correct). */ | |
90 static bool | |
91 fix_bb_placement (basic_block bb) | |
92 { | |
93 edge e; | |
94 edge_iterator ei; | |
95 struct loop *loop = current_loops->tree_root, *act; | |
96 | |
97 FOR_EACH_EDGE (e, ei, bb->succs) | |
98 { | |
99 if (e->dest == EXIT_BLOCK_PTR) | |
100 continue; | |
101 | |
102 act = e->dest->loop_father; | |
103 if (act->header == e->dest) | |
104 act = loop_outer (act); | |
105 | |
106 if (flow_loop_nested_p (loop, act)) | |
107 loop = act; | |
108 } | |
109 | |
110 if (loop == bb->loop_father) | |
111 return false; | |
112 | |
113 remove_bb_from_loops (bb); | |
114 add_bb_to_loop (bb, loop); | |
115 | |
116 return true; | |
117 } | |
118 | |
119 /* Fix placement of LOOP inside loop tree, i.e. find the innermost superloop | |
120 of LOOP to that leads at least one exit edge of LOOP, and set it | |
121 as the immediate superloop of LOOP. Return true if the immediate superloop | |
122 of LOOP changed. */ | |
123 | |
124 static bool | |
125 fix_loop_placement (struct loop *loop) | |
126 { | |
127 unsigned i; | |
128 edge e; | |
129 VEC (edge, heap) *exits = get_loop_exit_edges (loop); | |
130 struct loop *father = current_loops->tree_root, *act; | |
131 bool ret = false; | |
132 | |
133 for (i = 0; VEC_iterate (edge, exits, i, e); i++) | |
134 { | |
135 act = find_common_loop (loop, e->dest->loop_father); | |
136 if (flow_loop_nested_p (father, act)) | |
137 father = act; | |
138 } | |
139 | |
140 if (father != loop_outer (loop)) | |
141 { | |
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); | |
146 | |
147 /* The exit edges of LOOP no longer exits its original immediate | |
148 superloops; remove them from the appropriate exit lists. */ | |
149 for (i = 0; VEC_iterate (edge, exits, i, e); i++) | |
150 rescan_loop_exit (e, false, false); | |
151 | |
152 ret = true; | |
153 } | |
154 | |
155 VEC_free (edge, heap, exits); | |
156 return ret; | |
157 } | |
158 | |
159 /* Fix placements of basic blocks inside loop hierarchy stored in loops; i.e. | |
160 enforce condition condition stated in description of fix_bb_placement. We | |
161 start from basic block FROM that had some of its successors removed, so that | |
162 his placement no longer has to be correct, and iteratively fix placement of | |
163 its predecessors that may change if placement of FROM changed. Also fix | |
164 placement of subloops of FROM->loop_father, that might also be altered due | |
165 to this change; the condition for them is similar, except that instead of | |
166 successors we consider edges coming out of the loops. | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
167 |
0 | 168 If the changes may invalidate the information about irreducible regions, |
169 IRRED_INVALIDATED is set to true. */ | |
170 | |
171 static void | |
172 fix_bb_placements (basic_block from, | |
173 bool *irred_invalidated) | |
174 { | |
175 sbitmap in_queue; | |
176 basic_block *queue, *qtop, *qbeg, *qend; | |
177 struct loop *base_loop; | |
178 edge e; | |
179 | |
180 /* We pass through blocks back-reachable from FROM, testing whether some | |
181 of their successors moved to outer loop. It may be necessary to | |
182 iterate several times, but it is finite, as we stop unless we move | |
183 the basic block up the loop structure. The whole story is a bit | |
184 more complicated due to presence of subloops, those are moved using | |
185 fix_loop_placement. */ | |
186 | |
187 base_loop = from->loop_father; | |
188 if (base_loop == current_loops->tree_root) | |
189 return; | |
190 | |
191 in_queue = sbitmap_alloc (last_basic_block); | |
192 sbitmap_zero (in_queue); | |
193 SET_BIT (in_queue, from->index); | |
194 /* Prevent us from going out of the base_loop. */ | |
195 SET_BIT (in_queue, base_loop->header->index); | |
196 | |
197 queue = XNEWVEC (basic_block, base_loop->num_nodes + 1); | |
198 qtop = queue + base_loop->num_nodes + 1; | |
199 qbeg = queue; | |
200 qend = queue + 1; | |
201 *qbeg = from; | |
202 | |
203 while (qbeg != qend) | |
204 { | |
205 edge_iterator ei; | |
206 from = *qbeg; | |
207 qbeg++; | |
208 if (qbeg == qtop) | |
209 qbeg = queue; | |
210 RESET_BIT (in_queue, from->index); | |
211 | |
212 if (from->loop_father->header == from) | |
213 { | |
214 /* Subloop header, maybe move the loop upward. */ | |
215 if (!fix_loop_placement (from->loop_father)) | |
216 continue; | |
217 } | |
218 else | |
219 { | |
220 /* Ordinary basic block. */ | |
221 if (!fix_bb_placement (from)) | |
222 continue; | |
223 } | |
224 | |
225 FOR_EACH_EDGE (e, ei, from->succs) | |
226 { | |
227 if (e->flags & EDGE_IRREDUCIBLE_LOOP) | |
228 *irred_invalidated = true; | |
229 } | |
230 | |
231 /* Something has changed, insert predecessors into queue. */ | |
232 FOR_EACH_EDGE (e, ei, from->preds) | |
233 { | |
234 basic_block pred = e->src; | |
235 struct loop *nca; | |
236 | |
237 if (e->flags & EDGE_IRREDUCIBLE_LOOP) | |
238 *irred_invalidated = true; | |
239 | |
240 if (TEST_BIT (in_queue, pred->index)) | |
241 continue; | |
242 | |
243 /* If it is subloop, then it either was not moved, or | |
244 the path up the loop tree from base_loop do not contain | |
245 it. */ | |
246 nca = find_common_loop (pred->loop_father, base_loop); | |
247 if (pred->loop_father != base_loop | |
248 && (nca == base_loop | |
249 || nca != pred->loop_father)) | |
250 pred = pred->loop_father->header; | |
251 else if (!flow_loop_nested_p (from->loop_father, pred->loop_father)) | |
252 { | |
253 /* No point in processing it. */ | |
254 continue; | |
255 } | |
256 | |
257 if (TEST_BIT (in_queue, pred->index)) | |
258 continue; | |
259 | |
260 /* Schedule the basic block. */ | |
261 *qend = pred; | |
262 qend++; | |
263 if (qend == qtop) | |
264 qend = queue; | |
265 SET_BIT (in_queue, pred->index); | |
266 } | |
267 } | |
268 free (in_queue); | |
269 free (queue); | |
270 } | |
271 | |
272 /* Removes path beginning at edge E, i.e. remove basic blocks dominated by E | |
273 and update loop structures and dominators. Return true if we were able | |
274 to remove the path, false otherwise (and nothing is affected then). */ | |
275 bool | |
276 remove_path (edge e) | |
277 { | |
278 edge ae; | |
279 basic_block *rem_bbs, *bord_bbs, from, bb; | |
280 VEC (basic_block, heap) *dom_bbs; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
281 int i, nrem, n_bord_bbs; |
0 | 282 sbitmap seen; |
283 bool irred_invalidated = false; | |
284 | |
285 if (!can_remove_branch_p (e)) | |
286 return false; | |
287 | |
288 /* Keep track of whether we need to update information about irreducible | |
289 regions. This is the case if the removed area is a part of the | |
290 irreducible region, or if the set of basic blocks that belong to a loop | |
291 that is inside an irreducible region is changed, or if such a loop is | |
292 removed. */ | |
293 if (e->flags & EDGE_IRREDUCIBLE_LOOP) | |
294 irred_invalidated = true; | |
295 | |
296 /* We need to check whether basic blocks are dominated by the edge | |
297 e, but we only have basic block dominators. This is easy to | |
298 fix -- when e->dest has exactly one predecessor, this corresponds | |
299 to blocks dominated by e->dest, if not, split the edge. */ | |
300 if (!single_pred_p (e->dest)) | |
301 e = single_pred_edge (split_edge (e)); | |
302 | |
303 /* It may happen that by removing path we remove one or more loops | |
304 we belong to. In this case first unloop the loops, then proceed | |
305 normally. We may assume that e->dest is not a header of any loop, | |
306 as it now has exactly one predecessor. */ | |
307 while (loop_outer (e->src->loop_father) | |
308 && dominated_by_p (CDI_DOMINATORS, | |
309 e->src->loop_father->latch, e->dest)) | |
310 unloop (e->src->loop_father, &irred_invalidated); | |
311 | |
312 /* Identify the path. */ | |
313 nrem = find_path (e, &rem_bbs); | |
314 | |
315 n_bord_bbs = 0; | |
316 bord_bbs = XCNEWVEC (basic_block, n_basic_blocks); | |
317 seen = sbitmap_alloc (last_basic_block); | |
318 sbitmap_zero (seen); | |
319 | |
320 /* Find "border" hexes -- i.e. those with predecessor in removed path. */ | |
321 for (i = 0; i < nrem; i++) | |
322 SET_BIT (seen, rem_bbs[i]->index); | |
323 for (i = 0; i < nrem; i++) | |
324 { | |
325 edge_iterator ei; | |
326 bb = rem_bbs[i]; | |
327 FOR_EACH_EDGE (ae, ei, rem_bbs[i]->succs) | |
328 if (ae->dest != EXIT_BLOCK_PTR && !TEST_BIT (seen, ae->dest->index)) | |
329 { | |
330 SET_BIT (seen, ae->dest->index); | |
331 bord_bbs[n_bord_bbs++] = ae->dest; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
332 |
0 | 333 if (ae->flags & EDGE_IRREDUCIBLE_LOOP) |
334 irred_invalidated = true; | |
335 } | |
336 } | |
337 | |
338 /* Remove the path. */ | |
339 from = e->src; | |
340 remove_branch (e); | |
341 dom_bbs = NULL; | |
342 | |
343 /* Cancel loops contained in the path. */ | |
344 for (i = 0; i < nrem; i++) | |
345 if (rem_bbs[i]->loop_father->header == rem_bbs[i]) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
346 cancel_loop_tree (rem_bbs[i]->loop_father); |
0 | 347 |
348 remove_bbs (rem_bbs, nrem); | |
349 free (rem_bbs); | |
350 | |
351 /* Find blocks whose dominators may be affected. */ | |
352 sbitmap_zero (seen); | |
353 for (i = 0; i < n_bord_bbs; i++) | |
354 { | |
355 basic_block ldom; | |
356 | |
357 bb = get_immediate_dominator (CDI_DOMINATORS, bord_bbs[i]); | |
358 if (TEST_BIT (seen, bb->index)) | |
359 continue; | |
360 SET_BIT (seen, bb->index); | |
361 | |
362 for (ldom = first_dom_son (CDI_DOMINATORS, bb); | |
363 ldom; | |
364 ldom = next_dom_son (CDI_DOMINATORS, ldom)) | |
365 if (!dominated_by_p (CDI_DOMINATORS, from, ldom)) | |
366 VEC_safe_push (basic_block, heap, dom_bbs, ldom); | |
367 } | |
368 | |
369 free (seen); | |
370 | |
371 /* Recount dominators. */ | |
372 iterate_fix_dominators (CDI_DOMINATORS, dom_bbs, true); | |
373 VEC_free (basic_block, heap, dom_bbs); | |
374 free (bord_bbs); | |
375 | |
376 /* Fix placements of basic blocks inside loops and the placement of | |
377 loops in the loop tree. */ | |
378 fix_bb_placements (from, &irred_invalidated); | |
379 fix_loop_placements (from->loop_father, &irred_invalidated); | |
380 | |
381 if (irred_invalidated | |
382 && loops_state_satisfies_p (LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS)) | |
383 mark_irreducible_loops (); | |
384 | |
385 return true; | |
386 } | |
387 | |
388 /* Creates place for a new LOOP in loops structure. */ | |
389 | |
390 static void | |
391 place_new_loop (struct loop *loop) | |
392 { | |
393 loop->num = number_of_loops (); | |
394 VEC_safe_push (loop_p, gc, current_loops->larray, loop); | |
395 } | |
396 | |
397 /* Given LOOP structure with filled header and latch, find the body of the | |
398 corresponding loop and add it to loops tree. Insert the LOOP as a son of | |
399 outer. */ | |
400 | |
401 void | |
402 add_loop (struct loop *loop, struct loop *outer) | |
403 { | |
404 basic_block *bbs; | |
405 int i, n; | |
406 struct loop *subloop; | |
407 edge e; | |
408 edge_iterator ei; | |
409 | |
410 /* Add it to loop structure. */ | |
411 place_new_loop (loop); | |
412 flow_loop_tree_node_add (outer, loop); | |
413 | |
414 /* Find its nodes. */ | |
415 bbs = XNEWVEC (basic_block, n_basic_blocks); | |
416 n = get_loop_body_with_size (loop, bbs, n_basic_blocks); | |
417 | |
418 for (i = 0; i < n; i++) | |
419 { | |
420 if (bbs[i]->loop_father == outer) | |
421 { | |
422 remove_bb_from_loops (bbs[i]); | |
423 add_bb_to_loop (bbs[i], loop); | |
424 continue; | |
425 } | |
426 | |
427 loop->num_nodes++; | |
428 | |
429 /* If we find a direct subloop of OUTER, move it to LOOP. */ | |
430 subloop = bbs[i]->loop_father; | |
431 if (loop_outer (subloop) == outer | |
432 && subloop->header == bbs[i]) | |
433 { | |
434 flow_loop_tree_node_remove (subloop); | |
435 flow_loop_tree_node_add (loop, subloop); | |
436 } | |
437 } | |
438 | |
439 /* Update the information about loop exit edges. */ | |
440 for (i = 0; i < n; i++) | |
441 { | |
442 FOR_EACH_EDGE (e, ei, bbs[i]->succs) | |
443 { | |
444 rescan_loop_exit (e, false, false); | |
445 } | |
446 } | |
447 | |
448 free (bbs); | |
449 } | |
450 | |
451 /* Multiply all frequencies in LOOP by NUM/DEN. */ | |
452 void | |
453 scale_loop_frequencies (struct loop *loop, int num, int den) | |
454 { | |
455 basic_block *bbs; | |
456 | |
457 bbs = get_loop_body (loop); | |
458 scale_bbs_frequencies_int (bbs, loop->num_nodes, num, den); | |
459 free (bbs); | |
460 } | |
461 | |
462 /* Recompute dominance information for basic blocks outside LOOP. */ | |
463 | |
464 static void | |
465 update_dominators_in_loop (struct loop *loop) | |
466 { | |
467 VEC (basic_block, heap) *dom_bbs = NULL; | |
468 sbitmap seen; | |
469 basic_block *body; | |
470 unsigned i; | |
471 | |
472 seen = sbitmap_alloc (last_basic_block); | |
473 sbitmap_zero (seen); | |
474 body = get_loop_body (loop); | |
475 | |
476 for (i = 0; i < loop->num_nodes; i++) | |
477 SET_BIT (seen, body[i]->index); | |
478 | |
479 for (i = 0; i < loop->num_nodes; i++) | |
480 { | |
481 basic_block ldom; | |
482 | |
483 for (ldom = first_dom_son (CDI_DOMINATORS, body[i]); | |
484 ldom; | |
485 ldom = next_dom_son (CDI_DOMINATORS, ldom)) | |
486 if (!TEST_BIT (seen, ldom->index)) | |
487 { | |
488 SET_BIT (seen, ldom->index); | |
489 VEC_safe_push (basic_block, heap, dom_bbs, ldom); | |
490 } | |
491 } | |
492 | |
493 iterate_fix_dominators (CDI_DOMINATORS, dom_bbs, false); | |
494 free (body); | |
495 free (seen); | |
496 VEC_free (basic_block, heap, dom_bbs); | |
497 } | |
498 | |
499 /* Creates an if region as shown above. CONDITION is used to create | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
500 the test for the if. |
0 | 501 |
502 | | |
503 | ------------- ------------- | |
504 | | pred_bb | | pred_bb | | |
505 | ------------- ------------- | |
506 | | | | |
507 | | | ENTRY_EDGE | |
508 | | ENTRY_EDGE V | |
509 | | ====> ------------- | |
510 | | | cond_bb | | |
511 | | | CONDITION | | |
512 | | ------------- | |
513 | V / \ | |
514 | ------------- e_false / \ e_true | |
515 | | succ_bb | V V | |
516 | ------------- ----------- ----------- | |
517 | | false_bb | | true_bb | | |
518 | ----------- ----------- | |
519 | \ / | |
520 | \ / | |
521 | V V | |
522 | ------------- | |
523 | | join_bb | | |
524 | ------------- | |
525 | | exit_edge (result) | |
526 | V | |
527 | ----------- | |
528 | | succ_bb | | |
529 | ----------- | |
530 | | |
531 */ | |
532 | |
533 edge | |
534 create_empty_if_region_on_edge (edge entry_edge, tree condition) | |
535 { | |
536 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
537 basic_block cond_bb, true_bb, false_bb, join_bb; |
0 | 538 edge e_true, e_false, exit_edge; |
539 gimple cond_stmt; | |
540 tree simple_cond; | |
541 gimple_stmt_iterator gsi; | |
542 | |
543 cond_bb = split_edge (entry_edge); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
544 |
0 | 545 /* Insert condition in cond_bb. */ |
546 gsi = gsi_last_bb (cond_bb); | |
547 simple_cond = | |
548 force_gimple_operand_gsi (&gsi, condition, true, NULL, | |
549 false, GSI_NEW_STMT); | |
550 cond_stmt = gimple_build_cond_from_tree (simple_cond, NULL_TREE, NULL_TREE); | |
551 gsi = gsi_last_bb (cond_bb); | |
552 gsi_insert_after (&gsi, cond_stmt, GSI_NEW_STMT); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
553 |
0 | 554 join_bb = split_edge (single_succ_edge (cond_bb)); |
555 | |
556 e_true = single_succ_edge (cond_bb); | |
557 true_bb = split_edge (e_true); | |
558 | |
559 e_false = make_edge (cond_bb, join_bb, 0); | |
560 false_bb = split_edge (e_false); | |
561 | |
562 e_true->flags &= ~EDGE_FALLTHRU; | |
563 e_true->flags |= EDGE_TRUE_VALUE; | |
564 e_false->flags &= ~EDGE_FALLTHRU; | |
565 e_false->flags |= EDGE_FALSE_VALUE; | |
566 | |
567 set_immediate_dominator (CDI_DOMINATORS, cond_bb, entry_edge->src); | |
568 set_immediate_dominator (CDI_DOMINATORS, true_bb, cond_bb); | |
569 set_immediate_dominator (CDI_DOMINATORS, false_bb, cond_bb); | |
570 set_immediate_dominator (CDI_DOMINATORS, join_bb, cond_bb); | |
571 | |
572 exit_edge = single_succ_edge (join_bb); | |
573 | |
574 if (single_pred_p (exit_edge->dest)) | |
575 set_immediate_dominator (CDI_DOMINATORS, exit_edge->dest, join_bb); | |
576 | |
577 return exit_edge; | |
578 } | |
579 | |
580 /* create_empty_loop_on_edge | |
581 | | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
582 | - pred_bb - ------ pred_bb ------ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
583 | | | | iv0 = initial_value | |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
584 | -----|----- ---------|----------- |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
585 | | ______ | entry_edge |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
586 | | entry_edge / | | |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
587 | | ====> | -V---V- loop_header ------------- |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
588 | V | | iv_before = phi (iv0, iv_after) | |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
589 | - succ_bb - | ---|----------------------------- |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
590 | | | | | |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
591 | ----------- | ---V--- loop_body --------------- |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
592 | | | iv_after = iv_before + stride | |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
593 | | | if (iv_before < upper_bound) | |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
594 | | ---|--------------\-------------- |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
595 | | | \ exit_e |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
596 | | V \ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
597 | | - loop_latch - V- succ_bb - |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
598 | | | | | | |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
599 | | /------------- ----------- |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
600 | \ ___ / |
0 | 601 |
602 Creates an empty loop as shown above, the IV_BEFORE is the SSA_NAME | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
603 that is used before the increment of IV. IV_BEFORE should be used for |
0 | 604 adding code to the body that uses the IV. OUTER is the outer loop in |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
605 which the new loop should be inserted. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
606 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
607 Both INITIAL_VALUE and UPPER_BOUND expressions are gimplified and |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
608 inserted on the loop entry edge. This implies that this function |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
609 should be used only when the UPPER_BOUND expression is a loop |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
610 invariant. */ |
0 | 611 |
612 struct loop * | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
613 create_empty_loop_on_edge (edge entry_edge, |
0 | 614 tree initial_value, |
615 tree stride, tree upper_bound, | |
616 tree iv, | |
617 tree *iv_before, | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
618 tree *iv_after, |
0 | 619 struct loop *outer) |
620 { | |
621 basic_block loop_header, loop_latch, succ_bb, pred_bb; | |
622 struct loop *loop; | |
623 gimple_stmt_iterator gsi; | |
624 gimple_seq stmts; | |
625 gimple cond_expr; | |
626 tree exit_test; | |
627 edge exit_e; | |
628 int prob; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
629 |
0 | 630 gcc_assert (entry_edge && initial_value && stride && upper_bound && iv); |
631 | |
632 /* Create header, latch and wire up the loop. */ | |
633 pred_bb = entry_edge->src; | |
634 loop_header = split_edge (entry_edge); | |
635 loop_latch = split_edge (single_succ_edge (loop_header)); | |
636 succ_bb = single_succ (loop_latch); | |
637 make_edge (loop_header, succ_bb, 0); | |
638 redirect_edge_succ_nodup (single_succ_edge (loop_latch), loop_header); | |
639 | |
640 /* Set immediate dominator information. */ | |
641 set_immediate_dominator (CDI_DOMINATORS, loop_header, pred_bb); | |
642 set_immediate_dominator (CDI_DOMINATORS, loop_latch, loop_header); | |
643 set_immediate_dominator (CDI_DOMINATORS, succ_bb, loop_header); | |
644 | |
645 /* Initialize a loop structure and put it in a loop hierarchy. */ | |
646 loop = alloc_loop (); | |
647 loop->header = loop_header; | |
648 loop->latch = loop_latch; | |
649 add_loop (loop, outer); | |
650 | |
651 /* TODO: Fix frequencies and counts. */ | |
652 prob = REG_BR_PROB_BASE / 2; | |
653 | |
654 scale_loop_frequencies (loop, REG_BR_PROB_BASE - prob, REG_BR_PROB_BASE); | |
655 | |
656 /* Update dominators. */ | |
657 update_dominators_in_loop (loop); | |
658 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
659 /* Modify edge flags. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
660 exit_e = single_exit (loop); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
661 exit_e->flags = EDGE_LOOP_EXIT | EDGE_FALSE_VALUE; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
662 single_pred_edge (loop_latch)->flags = EDGE_TRUE_VALUE; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
663 |
0 | 664 /* Construct IV code in loop. */ |
665 initial_value = force_gimple_operand (initial_value, &stmts, true, iv); | |
666 if (stmts) | |
667 { | |
668 gsi_insert_seq_on_edge (loop_preheader_edge (loop), stmts); | |
669 gsi_commit_edge_inserts (); | |
670 } | |
671 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
672 upper_bound = force_gimple_operand (upper_bound, &stmts, true, NULL); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
673 if (stmts) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
674 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
675 gsi_insert_seq_on_edge (loop_preheader_edge (loop), stmts); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
676 gsi_commit_edge_inserts (); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
677 } |
0 | 678 |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
679 gsi = gsi_last_bb (loop_header); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
680 create_iv (initial_value, stride, iv, loop, &gsi, false, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
681 iv_before, iv_after); |
0 | 682 |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
683 /* Insert loop exit condition. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
684 cond_expr = gimple_build_cond |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
685 (LT_EXPR, *iv_before, upper_bound, NULL_TREE, NULL_TREE); |
0 | 686 |
687 exit_test = gimple_cond_lhs (cond_expr); | |
688 exit_test = force_gimple_operand_gsi (&gsi, exit_test, true, NULL, | |
689 false, GSI_NEW_STMT); | |
690 gimple_cond_set_lhs (cond_expr, exit_test); | |
691 gsi = gsi_last_bb (exit_e->src); | |
692 gsi_insert_after (&gsi, cond_expr, GSI_NEW_STMT); | |
693 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
694 split_block_after_labels (loop_header); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
695 |
0 | 696 return loop; |
697 } | |
698 | |
699 /* Make area between HEADER_EDGE and LATCH_EDGE a loop by connecting | |
700 latch to header and update loop tree and dominators | |
701 accordingly. Everything between them plus LATCH_EDGE destination must | |
702 be dominated by HEADER_EDGE destination, and back-reachable from | |
703 LATCH_EDGE source. HEADER_EDGE is redirected to basic block SWITCH_BB, | |
704 FALSE_EDGE of SWITCH_BB to original destination of HEADER_EDGE and | |
705 TRUE_EDGE of SWITCH_BB to original destination of LATCH_EDGE. | |
706 Returns the newly created loop. Frequencies and counts in the new loop | |
707 are scaled by FALSE_SCALE and in the old one by TRUE_SCALE. */ | |
708 | |
709 struct loop * | |
710 loopify (edge latch_edge, edge header_edge, | |
711 basic_block switch_bb, edge true_edge, edge false_edge, | |
712 bool redirect_all_edges, unsigned true_scale, unsigned false_scale) | |
713 { | |
714 basic_block succ_bb = latch_edge->dest; | |
715 basic_block pred_bb = header_edge->src; | |
716 struct loop *loop = alloc_loop (); | |
717 struct loop *outer = loop_outer (succ_bb->loop_father); | |
718 int freq; | |
719 gcov_type cnt; | |
720 edge e; | |
721 edge_iterator ei; | |
722 | |
723 loop->header = header_edge->dest; | |
724 loop->latch = latch_edge->src; | |
725 | |
726 freq = EDGE_FREQUENCY (header_edge); | |
727 cnt = header_edge->count; | |
728 | |
729 /* Redirect edges. */ | |
730 loop_redirect_edge (latch_edge, loop->header); | |
731 loop_redirect_edge (true_edge, succ_bb); | |
732 | |
733 /* During loop versioning, one of the switch_bb edge is already properly | |
734 set. Do not redirect it again unless redirect_all_edges is true. */ | |
735 if (redirect_all_edges) | |
736 { | |
737 loop_redirect_edge (header_edge, switch_bb); | |
738 loop_redirect_edge (false_edge, loop->header); | |
739 | |
740 /* Update dominators. */ | |
741 set_immediate_dominator (CDI_DOMINATORS, switch_bb, pred_bb); | |
742 set_immediate_dominator (CDI_DOMINATORS, loop->header, switch_bb); | |
743 } | |
744 | |
745 set_immediate_dominator (CDI_DOMINATORS, succ_bb, switch_bb); | |
746 | |
747 /* Compute new loop. */ | |
748 add_loop (loop, outer); | |
749 | |
750 /* Add switch_bb to appropriate loop. */ | |
751 if (switch_bb->loop_father) | |
752 remove_bb_from_loops (switch_bb); | |
753 add_bb_to_loop (switch_bb, outer); | |
754 | |
755 /* Fix frequencies. */ | |
756 if (redirect_all_edges) | |
757 { | |
758 switch_bb->frequency = freq; | |
759 switch_bb->count = cnt; | |
760 FOR_EACH_EDGE (e, ei, switch_bb->succs) | |
761 { | |
762 e->count = (switch_bb->count * e->probability) / REG_BR_PROB_BASE; | |
763 } | |
764 } | |
765 scale_loop_frequencies (loop, false_scale, REG_BR_PROB_BASE); | |
766 scale_loop_frequencies (succ_bb->loop_father, true_scale, REG_BR_PROB_BASE); | |
767 update_dominators_in_loop (loop); | |
768 | |
769 return loop; | |
770 } | |
771 | |
772 /* Remove the latch edge of a LOOP and update loops to indicate that | |
773 the LOOP was removed. After this function, original loop latch will | |
774 have no successor, which caller is expected to fix somehow. | |
775 | |
776 If this may cause the information about irreducible regions to become | |
777 invalid, IRRED_INVALIDATED is set to true. */ | |
778 | |
779 static void | |
780 unloop (struct loop *loop, bool *irred_invalidated) | |
781 { | |
782 basic_block *body; | |
783 struct loop *ploop; | |
784 unsigned i, n; | |
785 basic_block latch = loop->latch; | |
786 bool dummy = false; | |
787 | |
788 if (loop_preheader_edge (loop)->flags & EDGE_IRREDUCIBLE_LOOP) | |
789 *irred_invalidated = true; | |
790 | |
791 /* This is relatively straightforward. The dominators are unchanged, as | |
792 loop header dominates loop latch, so the only thing we have to care of | |
793 is the placement of loops and basic blocks inside the loop tree. We | |
794 move them all to the loop->outer, and then let fix_bb_placements do | |
795 its work. */ | |
796 | |
797 body = get_loop_body (loop); | |
798 n = loop->num_nodes; | |
799 for (i = 0; i < n; i++) | |
800 if (body[i]->loop_father == loop) | |
801 { | |
802 remove_bb_from_loops (body[i]); | |
803 add_bb_to_loop (body[i], loop_outer (loop)); | |
804 } | |
805 free(body); | |
806 | |
807 while (loop->inner) | |
808 { | |
809 ploop = loop->inner; | |
810 flow_loop_tree_node_remove (ploop); | |
811 flow_loop_tree_node_add (loop_outer (loop), ploop); | |
812 } | |
813 | |
814 /* Remove the loop and free its data. */ | |
815 delete_loop (loop); | |
816 | |
817 remove_edge (single_succ_edge (latch)); | |
818 | |
819 /* We do not pass IRRED_INVALIDATED to fix_bb_placements here, as even if | |
820 there is an irreducible region inside the cancelled loop, the flags will | |
821 be still correct. */ | |
822 fix_bb_placements (latch, &dummy); | |
823 } | |
824 | |
825 /* Fix placement of superloops of LOOP inside loop tree, i.e. ensure that | |
826 condition stated in description of fix_loop_placement holds for them. | |
827 It is used in case when we removed some edges coming out of LOOP, which | |
828 may cause the right placement of LOOP inside loop tree to change. | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
829 |
0 | 830 IRRED_INVALIDATED is set to true if a change in the loop structures might |
831 invalidate the information about irreducible regions. */ | |
832 | |
833 static void | |
834 fix_loop_placements (struct loop *loop, bool *irred_invalidated) | |
835 { | |
836 struct loop *outer; | |
837 | |
838 while (loop_outer (loop)) | |
839 { | |
840 outer = loop_outer (loop); | |
841 if (!fix_loop_placement (loop)) | |
842 break; | |
843 | |
844 /* Changing the placement of a loop in the loop tree may alter the | |
845 validity of condition 2) of the description of fix_bb_placement | |
846 for its preheader, because the successor is the header and belongs | |
847 to the loop. So call fix_bb_placements to fix up the placement | |
848 of the preheader and (possibly) of its predecessors. */ | |
849 fix_bb_placements (loop_preheader_edge (loop)->src, | |
850 irred_invalidated); | |
851 loop = outer; | |
852 } | |
853 } | |
854 | |
855 /* Copies copy of LOOP as subloop of TARGET loop, placing newly | |
856 created loop into loops structure. */ | |
857 struct loop * | |
858 duplicate_loop (struct loop *loop, struct loop *target) | |
859 { | |
860 struct loop *cloop; | |
861 cloop = alloc_loop (); | |
862 place_new_loop (cloop); | |
863 | |
864 /* Mark the new loop as copy of LOOP. */ | |
865 set_loop_copy (loop, cloop); | |
866 | |
867 /* Add it to target. */ | |
868 flow_loop_tree_node_add (target, cloop); | |
869 | |
870 return cloop; | |
871 } | |
872 | |
873 /* Copies structure of subloops of LOOP into TARGET loop, placing | |
874 newly created loops into loop tree. */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
875 void |
0 | 876 duplicate_subloops (struct loop *loop, struct loop *target) |
877 { | |
878 struct loop *aloop, *cloop; | |
879 | |
880 for (aloop = loop->inner; aloop; aloop = aloop->next) | |
881 { | |
882 cloop = duplicate_loop (aloop, target); | |
883 duplicate_subloops (aloop, cloop); | |
884 } | |
885 } | |
886 | |
887 /* Copies structure of subloops of N loops, stored in array COPIED_LOOPS, | |
888 into TARGET loop, placing newly created loops into loop tree. */ | |
889 static void | |
890 copy_loops_to (struct loop **copied_loops, int n, struct loop *target) | |
891 { | |
892 struct loop *aloop; | |
893 int i; | |
894 | |
895 for (i = 0; i < n; i++) | |
896 { | |
897 aloop = duplicate_loop (copied_loops[i], target); | |
898 duplicate_subloops (copied_loops[i], aloop); | |
899 } | |
900 } | |
901 | |
902 /* Redirects edge E to basic block DEST. */ | |
903 static void | |
904 loop_redirect_edge (edge e, basic_block dest) | |
905 { | |
906 if (e->dest == dest) | |
907 return; | |
908 | |
909 redirect_edge_and_branch_force (e, dest); | |
910 } | |
911 | |
912 /* Check whether LOOP's body can be duplicated. */ | |
913 bool | |
914 can_duplicate_loop_p (const struct loop *loop) | |
915 { | |
916 int ret; | |
917 basic_block *bbs = get_loop_body (loop); | |
918 | |
919 ret = can_copy_bbs_p (bbs, loop->num_nodes); | |
920 free (bbs); | |
921 | |
922 return ret; | |
923 } | |
924 | |
925 /* Sets probability and count of edge E to zero. The probability and count | |
926 is redistributed evenly to the remaining edges coming from E->src. */ | |
927 | |
928 static void | |
929 set_zero_probability (edge e) | |
930 { | |
931 basic_block bb = e->src; | |
932 edge_iterator ei; | |
933 edge ae, last = NULL; | |
934 unsigned n = EDGE_COUNT (bb->succs); | |
935 gcov_type cnt = e->count, cnt1; | |
936 unsigned prob = e->probability, prob1; | |
937 | |
938 gcc_assert (n > 1); | |
939 cnt1 = cnt / (n - 1); | |
940 prob1 = prob / (n - 1); | |
941 | |
942 FOR_EACH_EDGE (ae, ei, bb->succs) | |
943 { | |
944 if (ae == e) | |
945 continue; | |
946 | |
947 ae->probability += prob1; | |
948 ae->count += cnt1; | |
949 last = ae; | |
950 } | |
951 | |
952 /* Move the rest to one of the edges. */ | |
953 last->probability += prob % (n - 1); | |
954 last->count += cnt % (n - 1); | |
955 | |
956 e->probability = 0; | |
957 e->count = 0; | |
958 } | |
959 | |
960 /* Duplicates body of LOOP to given edge E NDUPL times. Takes care of updating | |
961 loop structure and dominators. E's destination must be LOOP header for | |
962 this to work, i.e. it must be entry or latch edge of this loop; these are | |
963 unique, as the loops must have preheaders for this function to work | |
964 correctly (in case E is latch, the function unrolls the loop, if E is entry | |
965 edge, it peels the loop). Store edges created by copying ORIG edge from | |
966 copies corresponding to set bits in WONT_EXIT bitmap (bit 0 corresponds to | |
967 original LOOP body, the other copies are numbered in order given by control | |
968 flow through them) into TO_REMOVE array. Returns false if duplication is | |
969 impossible. */ | |
970 | |
971 bool | |
972 duplicate_loop_to_header_edge (struct loop *loop, edge e, | |
973 unsigned int ndupl, sbitmap wont_exit, | |
974 edge orig, VEC (edge, heap) **to_remove, | |
975 int flags) | |
976 { | |
977 struct loop *target, *aloop; | |
978 struct loop **orig_loops; | |
979 unsigned n_orig_loops; | |
980 basic_block header = loop->header, latch = loop->latch; | |
981 basic_block *new_bbs, *bbs, *first_active; | |
982 basic_block new_bb, bb, first_active_latch = NULL; | |
983 edge ae, latch_edge; | |
984 edge spec_edges[2], new_spec_edges[2]; | |
985 #define SE_LATCH 0 | |
986 #define SE_ORIG 1 | |
987 unsigned i, j, n; | |
988 int is_latch = (latch == e->src); | |
989 int scale_act = 0, *scale_step = NULL, scale_main = 0; | |
990 int scale_after_exit = 0; | |
991 int p, freq_in, freq_le, freq_out_orig; | |
992 int prob_pass_thru, prob_pass_wont_exit, prob_pass_main; | |
993 int add_irreducible_flag; | |
994 basic_block place_after; | |
995 bitmap bbs_to_scale = NULL; | |
996 bitmap_iterator bi; | |
997 | |
998 gcc_assert (e->dest == loop->header); | |
999 gcc_assert (ndupl > 0); | |
1000 | |
1001 if (orig) | |
1002 { | |
1003 /* Orig must be edge out of the loop. */ | |
1004 gcc_assert (flow_bb_inside_loop_p (loop, orig->src)); | |
1005 gcc_assert (!flow_bb_inside_loop_p (loop, orig->dest)); | |
1006 } | |
1007 | |
1008 n = loop->num_nodes; | |
1009 bbs = get_loop_body_in_dom_order (loop); | |
1010 gcc_assert (bbs[0] == loop->header); | |
1011 gcc_assert (bbs[n - 1] == loop->latch); | |
1012 | |
1013 /* Check whether duplication is possible. */ | |
1014 if (!can_copy_bbs_p (bbs, loop->num_nodes)) | |
1015 { | |
1016 free (bbs); | |
1017 return false; | |
1018 } | |
1019 new_bbs = XNEWVEC (basic_block, loop->num_nodes); | |
1020 | |
1021 /* In case we are doing loop peeling and the loop is in the middle of | |
1022 irreducible region, the peeled copies will be inside it too. */ | |
1023 add_irreducible_flag = e->flags & EDGE_IRREDUCIBLE_LOOP; | |
1024 gcc_assert (!is_latch || !add_irreducible_flag); | |
1025 | |
1026 /* Find edge from latch. */ | |
1027 latch_edge = loop_latch_edge (loop); | |
1028 | |
1029 if (flags & DLTHE_FLAG_UPDATE_FREQ) | |
1030 { | |
1031 /* Calculate coefficients by that we have to scale frequencies | |
1032 of duplicated loop bodies. */ | |
1033 freq_in = header->frequency; | |
1034 freq_le = EDGE_FREQUENCY (latch_edge); | |
1035 if (freq_in == 0) | |
1036 freq_in = 1; | |
1037 if (freq_in < freq_le) | |
1038 freq_in = freq_le; | |
1039 freq_out_orig = orig ? EDGE_FREQUENCY (orig) : freq_in - freq_le; | |
1040 if (freq_out_orig > freq_in - freq_le) | |
1041 freq_out_orig = freq_in - freq_le; | |
1042 prob_pass_thru = RDIV (REG_BR_PROB_BASE * freq_le, freq_in); | |
1043 prob_pass_wont_exit = | |
1044 RDIV (REG_BR_PROB_BASE * (freq_le + freq_out_orig), freq_in); | |
1045 | |
1046 if (orig | |
1047 && REG_BR_PROB_BASE - orig->probability != 0) | |
1048 { | |
1049 /* The blocks that are dominated by a removed exit edge ORIG have | |
1050 frequencies scaled by this. */ | |
1051 scale_after_exit = RDIV (REG_BR_PROB_BASE * REG_BR_PROB_BASE, | |
1052 REG_BR_PROB_BASE - orig->probability); | |
1053 bbs_to_scale = BITMAP_ALLOC (NULL); | |
1054 for (i = 0; i < n; i++) | |
1055 { | |
1056 if (bbs[i] != orig->src | |
1057 && dominated_by_p (CDI_DOMINATORS, bbs[i], orig->src)) | |
1058 bitmap_set_bit (bbs_to_scale, i); | |
1059 } | |
1060 } | |
1061 | |
1062 scale_step = XNEWVEC (int, ndupl); | |
1063 | |
1064 for (i = 1; i <= ndupl; i++) | |
1065 scale_step[i - 1] = TEST_BIT (wont_exit, i) | |
1066 ? prob_pass_wont_exit | |
1067 : prob_pass_thru; | |
1068 | |
1069 /* Complete peeling is special as the probability of exit in last | |
1070 copy becomes 1. */ | |
1071 if (flags & DLTHE_FLAG_COMPLETTE_PEEL) | |
1072 { | |
1073 int wanted_freq = EDGE_FREQUENCY (e); | |
1074 | |
1075 if (wanted_freq > freq_in) | |
1076 wanted_freq = freq_in; | |
1077 | |
1078 gcc_assert (!is_latch); | |
1079 /* First copy has frequency of incoming edge. Each subsequent | |
1080 frequency should be reduced by prob_pass_wont_exit. Caller | |
1081 should've managed the flags so all except for original loop | |
1082 has won't exist set. */ | |
1083 scale_act = RDIV (wanted_freq * REG_BR_PROB_BASE, freq_in); | |
1084 /* Now simulate the duplication adjustments and compute header | |
1085 frequency of the last copy. */ | |
1086 for (i = 0; i < ndupl; i++) | |
1087 wanted_freq = RDIV (wanted_freq * scale_step[i], REG_BR_PROB_BASE); | |
1088 scale_main = RDIV (wanted_freq * REG_BR_PROB_BASE, freq_in); | |
1089 } | |
1090 else if (is_latch) | |
1091 { | |
1092 prob_pass_main = TEST_BIT (wont_exit, 0) | |
1093 ? prob_pass_wont_exit | |
1094 : prob_pass_thru; | |
1095 p = prob_pass_main; | |
1096 scale_main = REG_BR_PROB_BASE; | |
1097 for (i = 0; i < ndupl; i++) | |
1098 { | |
1099 scale_main += p; | |
1100 p = RDIV (p * scale_step[i], REG_BR_PROB_BASE); | |
1101 } | |
1102 scale_main = RDIV (REG_BR_PROB_BASE * REG_BR_PROB_BASE, scale_main); | |
1103 scale_act = RDIV (scale_main * prob_pass_main, REG_BR_PROB_BASE); | |
1104 } | |
1105 else | |
1106 { | |
1107 scale_main = REG_BR_PROB_BASE; | |
1108 for (i = 0; i < ndupl; i++) | |
1109 scale_main = RDIV (scale_main * scale_step[i], REG_BR_PROB_BASE); | |
1110 scale_act = REG_BR_PROB_BASE - prob_pass_thru; | |
1111 } | |
1112 for (i = 0; i < ndupl; i++) | |
1113 gcc_assert (scale_step[i] >= 0 && scale_step[i] <= REG_BR_PROB_BASE); | |
1114 gcc_assert (scale_main >= 0 && scale_main <= REG_BR_PROB_BASE | |
1115 && scale_act >= 0 && scale_act <= REG_BR_PROB_BASE); | |
1116 } | |
1117 | |
1118 /* Loop the new bbs will belong to. */ | |
1119 target = e->src->loop_father; | |
1120 | |
1121 /* Original loops. */ | |
1122 n_orig_loops = 0; | |
1123 for (aloop = loop->inner; aloop; aloop = aloop->next) | |
1124 n_orig_loops++; | |
1125 orig_loops = XCNEWVEC (struct loop *, n_orig_loops); | |
1126 for (aloop = loop->inner, i = 0; aloop; aloop = aloop->next, i++) | |
1127 orig_loops[i] = aloop; | |
1128 | |
1129 set_loop_copy (loop, target); | |
1130 | |
1131 first_active = XNEWVEC (basic_block, n); | |
1132 if (is_latch) | |
1133 { | |
1134 memcpy (first_active, bbs, n * sizeof (basic_block)); | |
1135 first_active_latch = latch; | |
1136 } | |
1137 | |
1138 spec_edges[SE_ORIG] = orig; | |
1139 spec_edges[SE_LATCH] = latch_edge; | |
1140 | |
1141 place_after = e->src; | |
1142 for (j = 0; j < ndupl; j++) | |
1143 { | |
1144 /* Copy loops. */ | |
1145 copy_loops_to (orig_loops, n_orig_loops, target); | |
1146 | |
1147 /* Copy bbs. */ | |
1148 copy_bbs (bbs, n, new_bbs, spec_edges, 2, new_spec_edges, loop, | |
1149 place_after); | |
1150 place_after = new_spec_edges[SE_LATCH]->src; | |
1151 | |
1152 if (flags & DLTHE_RECORD_COPY_NUMBER) | |
1153 for (i = 0; i < n; i++) | |
1154 { | |
1155 gcc_assert (!new_bbs[i]->aux); | |
1156 new_bbs[i]->aux = (void *)(size_t)(j + 1); | |
1157 } | |
1158 | |
1159 /* Note whether the blocks and edges belong to an irreducible loop. */ | |
1160 if (add_irreducible_flag) | |
1161 { | |
1162 for (i = 0; i < n; i++) | |
1163 new_bbs[i]->flags |= BB_DUPLICATED; | |
1164 for (i = 0; i < n; i++) | |
1165 { | |
1166 edge_iterator ei; | |
1167 new_bb = new_bbs[i]; | |
1168 if (new_bb->loop_father == target) | |
1169 new_bb->flags |= BB_IRREDUCIBLE_LOOP; | |
1170 | |
1171 FOR_EACH_EDGE (ae, ei, new_bb->succs) | |
1172 if ((ae->dest->flags & BB_DUPLICATED) | |
1173 && (ae->src->loop_father == target | |
1174 || ae->dest->loop_father == target)) | |
1175 ae->flags |= EDGE_IRREDUCIBLE_LOOP; | |
1176 } | |
1177 for (i = 0; i < n; i++) | |
1178 new_bbs[i]->flags &= ~BB_DUPLICATED; | |
1179 } | |
1180 | |
1181 /* Redirect the special edges. */ | |
1182 if (is_latch) | |
1183 { | |
1184 redirect_edge_and_branch_force (latch_edge, new_bbs[0]); | |
1185 redirect_edge_and_branch_force (new_spec_edges[SE_LATCH], | |
1186 loop->header); | |
1187 set_immediate_dominator (CDI_DOMINATORS, new_bbs[0], latch); | |
1188 latch = loop->latch = new_bbs[n - 1]; | |
1189 e = latch_edge = new_spec_edges[SE_LATCH]; | |
1190 } | |
1191 else | |
1192 { | |
1193 redirect_edge_and_branch_force (new_spec_edges[SE_LATCH], | |
1194 loop->header); | |
1195 redirect_edge_and_branch_force (e, new_bbs[0]); | |
1196 set_immediate_dominator (CDI_DOMINATORS, new_bbs[0], e->src); | |
1197 e = new_spec_edges[SE_LATCH]; | |
1198 } | |
1199 | |
1200 /* Record exit edge in this copy. */ | |
1201 if (orig && TEST_BIT (wont_exit, j + 1)) | |
1202 { | |
1203 if (to_remove) | |
1204 VEC_safe_push (edge, heap, *to_remove, new_spec_edges[SE_ORIG]); | |
1205 set_zero_probability (new_spec_edges[SE_ORIG]); | |
1206 | |
1207 /* Scale the frequencies of the blocks dominated by the exit. */ | |
1208 if (bbs_to_scale) | |
1209 { | |
1210 EXECUTE_IF_SET_IN_BITMAP (bbs_to_scale, 0, i, bi) | |
1211 { | |
1212 scale_bbs_frequencies_int (new_bbs + i, 1, scale_after_exit, | |
1213 REG_BR_PROB_BASE); | |
1214 } | |
1215 } | |
1216 } | |
1217 | |
1218 /* Record the first copy in the control flow order if it is not | |
1219 the original loop (i.e. in case of peeling). */ | |
1220 if (!first_active_latch) | |
1221 { | |
1222 memcpy (first_active, new_bbs, n * sizeof (basic_block)); | |
1223 first_active_latch = new_bbs[n - 1]; | |
1224 } | |
1225 | |
1226 /* Set counts and frequencies. */ | |
1227 if (flags & DLTHE_FLAG_UPDATE_FREQ) | |
1228 { | |
1229 scale_bbs_frequencies_int (new_bbs, n, scale_act, REG_BR_PROB_BASE); | |
1230 scale_act = RDIV (scale_act * scale_step[j], REG_BR_PROB_BASE); | |
1231 } | |
1232 } | |
1233 free (new_bbs); | |
1234 free (orig_loops); | |
1235 | |
1236 /* Record the exit edge in the original loop body, and update the frequencies. */ | |
1237 if (orig && TEST_BIT (wont_exit, 0)) | |
1238 { | |
1239 if (to_remove) | |
1240 VEC_safe_push (edge, heap, *to_remove, orig); | |
1241 set_zero_probability (orig); | |
1242 | |
1243 /* Scale the frequencies of the blocks dominated by the exit. */ | |
1244 if (bbs_to_scale) | |
1245 { | |
1246 EXECUTE_IF_SET_IN_BITMAP (bbs_to_scale, 0, i, bi) | |
1247 { | |
1248 scale_bbs_frequencies_int (bbs + i, 1, scale_after_exit, | |
1249 REG_BR_PROB_BASE); | |
1250 } | |
1251 } | |
1252 } | |
1253 | |
1254 /* Update the original loop. */ | |
1255 if (!is_latch) | |
1256 set_immediate_dominator (CDI_DOMINATORS, e->dest, e->src); | |
1257 if (flags & DLTHE_FLAG_UPDATE_FREQ) | |
1258 { | |
1259 scale_bbs_frequencies_int (bbs, n, scale_main, REG_BR_PROB_BASE); | |
1260 free (scale_step); | |
1261 } | |
1262 | |
1263 /* Update dominators of outer blocks if affected. */ | |
1264 for (i = 0; i < n; i++) | |
1265 { | |
1266 basic_block dominated, dom_bb; | |
1267 VEC (basic_block, heap) *dom_bbs; | |
1268 unsigned j; | |
1269 | |
1270 bb = bbs[i]; | |
1271 bb->aux = 0; | |
1272 | |
1273 dom_bbs = get_dominated_by (CDI_DOMINATORS, bb); | |
1274 for (j = 0; VEC_iterate (basic_block, dom_bbs, j, dominated); j++) | |
1275 { | |
1276 if (flow_bb_inside_loop_p (loop, dominated)) | |
1277 continue; | |
1278 dom_bb = nearest_common_dominator ( | |
1279 CDI_DOMINATORS, first_active[i], first_active_latch); | |
1280 set_immediate_dominator (CDI_DOMINATORS, dominated, dom_bb); | |
1281 } | |
1282 VEC_free (basic_block, heap, dom_bbs); | |
1283 } | |
1284 free (first_active); | |
1285 | |
1286 free (bbs); | |
1287 BITMAP_FREE (bbs_to_scale); | |
1288 | |
1289 return true; | |
1290 } | |
1291 | |
1292 /* A callback for make_forwarder block, to redirect all edges except for | |
1293 MFB_KJ_EDGE to the entry part. E is the edge for that we should decide | |
1294 whether to redirect it. */ | |
1295 | |
1296 edge mfb_kj_edge; | |
1297 bool | |
1298 mfb_keep_just (edge e) | |
1299 { | |
1300 return e != mfb_kj_edge; | |
1301 } | |
1302 | |
1303 /* True when a candidate preheader BLOCK has predecessors from LOOP. */ | |
1304 | |
1305 static bool | |
1306 has_preds_from_loop (basic_block block, struct loop *loop) | |
1307 { | |
1308 edge e; | |
1309 edge_iterator ei; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1310 |
0 | 1311 FOR_EACH_EDGE (e, ei, block->preds) |
1312 if (e->src->loop_father == loop) | |
1313 return true; | |
1314 return false; | |
1315 } | |
1316 | |
1317 /* Creates a pre-header for a LOOP. Returns newly created block. Unless | |
1318 CP_SIMPLE_PREHEADERS is set in FLAGS, we only force LOOP to have single | |
1319 entry; otherwise we also force preheader block to have only one successor. | |
1320 When CP_FALLTHRU_PREHEADERS is set in FLAGS, we force the preheader block | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1321 to be a fallthru predecessor to the loop header and to have only |
0 | 1322 predecessors from outside of the loop. |
1323 The function also updates dominators. */ | |
1324 | |
1325 basic_block | |
1326 create_preheader (struct loop *loop, int flags) | |
1327 { | |
1328 edge e, fallthru; | |
1329 basic_block dummy; | |
1330 int nentry = 0; | |
1331 bool irred = false; | |
1332 bool latch_edge_was_fallthru; | |
1333 edge one_succ_pred = NULL, single_entry = NULL; | |
1334 edge_iterator ei; | |
1335 | |
1336 FOR_EACH_EDGE (e, ei, loop->header->preds) | |
1337 { | |
1338 if (e->src == loop->latch) | |
1339 continue; | |
1340 irred |= (e->flags & EDGE_IRREDUCIBLE_LOOP) != 0; | |
1341 nentry++; | |
1342 single_entry = e; | |
1343 if (single_succ_p (e->src)) | |
1344 one_succ_pred = e; | |
1345 } | |
1346 gcc_assert (nentry); | |
1347 if (nentry == 1) | |
1348 { | |
1349 bool need_forwarder_block = false; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1350 |
0 | 1351 /* We do not allow entry block to be the loop preheader, since we |
1352 cannot emit code there. */ | |
1353 if (single_entry->src == ENTRY_BLOCK_PTR) | |
1354 need_forwarder_block = true; | |
1355 else | |
1356 { | |
1357 /* If we want simple preheaders, also force the preheader to have | |
1358 just a single successor. */ | |
1359 if ((flags & CP_SIMPLE_PREHEADERS) | |
1360 && !single_succ_p (single_entry->src)) | |
1361 need_forwarder_block = true; | |
1362 /* If we want fallthru preheaders, also create forwarder block when | |
1363 preheader ends with a jump or has predecessors from loop. */ | |
1364 else if ((flags & CP_FALLTHRU_PREHEADERS) | |
1365 && (JUMP_P (BB_END (single_entry->src)) | |
1366 || has_preds_from_loop (single_entry->src, loop))) | |
1367 need_forwarder_block = true; | |
1368 } | |
1369 if (! need_forwarder_block) | |
1370 return NULL; | |
1371 } | |
1372 | |
1373 mfb_kj_edge = loop_latch_edge (loop); | |
1374 latch_edge_was_fallthru = (mfb_kj_edge->flags & EDGE_FALLTHRU) != 0; | |
1375 fallthru = make_forwarder_block (loop->header, mfb_keep_just, NULL); | |
1376 dummy = fallthru->src; | |
1377 loop->header = fallthru->dest; | |
1378 | |
1379 /* Try to be clever in placing the newly created preheader. The idea is to | |
1380 avoid breaking any "fallthruness" relationship between blocks. | |
1381 | |
1382 The preheader was created just before the header and all incoming edges | |
1383 to the header were redirected to the preheader, except the latch edge. | |
1384 So the only problematic case is when this latch edge was a fallthru | |
1385 edge: it is not anymore after the preheader creation so we have broken | |
1386 the fallthruness. We're therefore going to look for a better place. */ | |
1387 if (latch_edge_was_fallthru) | |
1388 { | |
1389 if (one_succ_pred) | |
1390 e = one_succ_pred; | |
1391 else | |
1392 e = EDGE_PRED (dummy, 0); | |
1393 | |
1394 move_block_after (dummy, e->src); | |
1395 } | |
1396 | |
1397 if (irred) | |
1398 { | |
1399 dummy->flags |= BB_IRREDUCIBLE_LOOP; | |
1400 single_succ_edge (dummy)->flags |= EDGE_IRREDUCIBLE_LOOP; | |
1401 } | |
1402 | |
1403 if (dump_file) | |
1404 fprintf (dump_file, "Created preheader block for loop %i\n", | |
1405 loop->num); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1406 |
0 | 1407 if (flags & CP_FALLTHRU_PREHEADERS) |
1408 gcc_assert ((single_succ_edge (dummy)->flags & EDGE_FALLTHRU) | |
1409 && !JUMP_P (BB_END (dummy))); | |
1410 | |
1411 return dummy; | |
1412 } | |
1413 | |
1414 /* Create preheaders for each loop; for meaning of FLAGS see create_preheader. */ | |
1415 | |
1416 void | |
1417 create_preheaders (int flags) | |
1418 { | |
1419 loop_iterator li; | |
1420 struct loop *loop; | |
1421 | |
1422 if (!current_loops) | |
1423 return; | |
1424 | |
1425 FOR_EACH_LOOP (li, loop, 0) | |
1426 create_preheader (loop, flags); | |
1427 loops_state_set (LOOPS_HAVE_PREHEADERS); | |
1428 } | |
1429 | |
1430 /* Forces all loop latches to have only single successor. */ | |
1431 | |
1432 void | |
1433 force_single_succ_latches (void) | |
1434 { | |
1435 loop_iterator li; | |
1436 struct loop *loop; | |
1437 edge e; | |
1438 | |
1439 FOR_EACH_LOOP (li, loop, 0) | |
1440 { | |
1441 if (loop->latch != loop->header && single_succ_p (loop->latch)) | |
1442 continue; | |
1443 | |
1444 e = find_edge (loop->latch, loop->header); | |
1445 | |
1446 split_edge (e); | |
1447 } | |
1448 loops_state_set (LOOPS_HAVE_SIMPLE_LATCHES); | |
1449 } | |
1450 | |
1451 /* This function is called from loop_version. It splits the entry edge | |
1452 of the loop we want to version, adds the versioning condition, and | |
1453 adjust the edges to the two versions of the loop appropriately. | |
1454 e is an incoming edge. Returns the basic block containing the | |
1455 condition. | |
1456 | |
1457 --- edge e ---- > [second_head] | |
1458 | |
1459 Split it and insert new conditional expression and adjust edges. | |
1460 | |
1461 --- edge e ---> [cond expr] ---> [first_head] | |
1462 | | |
1463 +---------> [second_head] | |
1464 | |
1465 THEN_PROB is the probability of then branch of the condition. */ | |
1466 | |
1467 static basic_block | |
1468 lv_adjust_loop_entry_edge (basic_block first_head, basic_block second_head, | |
1469 edge e, void *cond_expr, unsigned then_prob) | |
1470 { | |
1471 basic_block new_head = NULL; | |
1472 edge e1; | |
1473 | |
1474 gcc_assert (e->dest == second_head); | |
1475 | |
1476 /* Split edge 'e'. This will create a new basic block, where we can | |
1477 insert conditional expr. */ | |
1478 new_head = split_edge (e); | |
1479 | |
1480 lv_add_condition_to_bb (first_head, second_head, new_head, | |
1481 cond_expr); | |
1482 | |
1483 /* Don't set EDGE_TRUE_VALUE in RTL mode, as it's invalid there. */ | |
1484 e = single_succ_edge (new_head); | |
1485 e1 = make_edge (new_head, first_head, | |
1486 current_ir_type () == IR_GIMPLE ? EDGE_TRUE_VALUE : 0); | |
1487 e1->probability = then_prob; | |
1488 e->probability = REG_BR_PROB_BASE - then_prob; | |
1489 e1->count = RDIV (e->count * e1->probability, REG_BR_PROB_BASE); | |
1490 e->count = RDIV (e->count * e->probability, REG_BR_PROB_BASE); | |
1491 | |
1492 set_immediate_dominator (CDI_DOMINATORS, first_head, new_head); | |
1493 set_immediate_dominator (CDI_DOMINATORS, second_head, new_head); | |
1494 | |
1495 /* Adjust loop header phi nodes. */ | |
1496 lv_adjust_loop_header_phi (first_head, second_head, new_head, e1); | |
1497 | |
1498 return new_head; | |
1499 } | |
1500 | |
1501 /* Main entry point for Loop Versioning transformation. | |
1502 | |
1503 This transformation given a condition and a loop, creates | |
1504 -if (condition) { loop_copy1 } else { loop_copy2 }, | |
1505 where loop_copy1 is the loop transformed in one way, and loop_copy2 | |
1506 is the loop transformed in another way (or unchanged). 'condition' | |
1507 may be a run time test for things that were not resolved by static | |
1508 analysis (overlapping ranges (anti-aliasing), alignment, etc.). | |
1509 | |
1510 THEN_PROB is the probability of the then edge of the if. THEN_SCALE | |
1511 is the ratio by that the frequencies in the original loop should | |
1512 be scaled. ELSE_SCALE is the ratio by that the frequencies in the | |
1513 new loop should be scaled. | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1514 |
0 | 1515 If PLACE_AFTER is true, we place the new loop after LOOP in the |
1516 instruction stream, otherwise it is placed before LOOP. */ | |
1517 | |
1518 struct loop * | |
1519 loop_version (struct loop *loop, | |
1520 void *cond_expr, basic_block *condition_bb, | |
1521 unsigned then_prob, unsigned then_scale, unsigned else_scale, | |
1522 bool place_after) | |
1523 { | |
1524 basic_block first_head, second_head; | |
1525 edge entry, latch_edge, true_edge, false_edge; | |
1526 int irred_flag; | |
1527 struct loop *nloop; | |
1528 basic_block cond_bb; | |
1529 | |
1530 /* Record entry and latch edges for the loop */ | |
1531 entry = loop_preheader_edge (loop); | |
1532 irred_flag = entry->flags & EDGE_IRREDUCIBLE_LOOP; | |
1533 entry->flags &= ~EDGE_IRREDUCIBLE_LOOP; | |
1534 | |
1535 /* Note down head of loop as first_head. */ | |
1536 first_head = entry->dest; | |
1537 | |
1538 /* Duplicate loop. */ | |
1539 if (!cfg_hook_duplicate_loop_to_header_edge (loop, entry, 1, | |
1540 NULL, NULL, NULL, 0)) | |
1541 return NULL; | |
1542 | |
1543 /* After duplication entry edge now points to new loop head block. | |
1544 Note down new head as second_head. */ | |
1545 second_head = entry->dest; | |
1546 | |
1547 /* Split loop entry edge and insert new block with cond expr. */ | |
1548 cond_bb = lv_adjust_loop_entry_edge (first_head, second_head, | |
1549 entry, cond_expr, then_prob); | |
1550 if (condition_bb) | |
1551 *condition_bb = cond_bb; | |
1552 | |
1553 if (!cond_bb) | |
1554 { | |
1555 entry->flags |= irred_flag; | |
1556 return NULL; | |
1557 } | |
1558 | |
1559 latch_edge = single_succ_edge (get_bb_copy (loop->latch)); | |
1560 | |
1561 extract_cond_bb_edges (cond_bb, &true_edge, &false_edge); | |
1562 nloop = loopify (latch_edge, | |
1563 single_pred_edge (get_bb_copy (loop->header)), | |
1564 cond_bb, true_edge, false_edge, | |
1565 false /* Do not redirect all edges. */, | |
1566 then_scale, else_scale); | |
1567 | |
1568 /* loopify redirected latch_edge. Update its PENDING_STMTS. */ | |
1569 lv_flush_pending_stmts (latch_edge); | |
1570 | |
1571 /* loopify redirected condition_bb's succ edge. Update its PENDING_STMTS. */ | |
1572 extract_cond_bb_edges (cond_bb, &true_edge, &false_edge); | |
1573 lv_flush_pending_stmts (false_edge); | |
1574 /* Adjust irreducible flag. */ | |
1575 if (irred_flag) | |
1576 { | |
1577 cond_bb->flags |= BB_IRREDUCIBLE_LOOP; | |
1578 loop_preheader_edge (loop)->flags |= EDGE_IRREDUCIBLE_LOOP; | |
1579 loop_preheader_edge (nloop)->flags |= EDGE_IRREDUCIBLE_LOOP; | |
1580 single_pred_edge (cond_bb)->flags |= EDGE_IRREDUCIBLE_LOOP; | |
1581 } | |
1582 | |
1583 if (place_after) | |
1584 { | |
1585 basic_block *bbs = get_loop_body_in_dom_order (nloop), after; | |
1586 unsigned i; | |
1587 | |
1588 after = loop->latch; | |
1589 | |
1590 for (i = 0; i < nloop->num_nodes; i++) | |
1591 { | |
1592 move_block_after (bbs[i], after); | |
1593 after = bbs[i]; | |
1594 } | |
1595 free (bbs); | |
1596 } | |
1597 | |
1598 /* At this point condition_bb is loop preheader with two successors, | |
1599 first_head and second_head. Make sure that loop preheader has only | |
1600 one successor. */ | |
1601 split_edge (loop_preheader_edge (loop)); | |
1602 split_edge (loop_preheader_edge (nloop)); | |
1603 | |
1604 return nloop; | |
1605 } | |
1606 | |
1607 /* The structure of loops might have changed. Some loops might get removed | |
1608 (and their headers and latches were set to NULL), loop exists might get | |
1609 removed (thus the loop nesting may be wrong), and some blocks and edges | |
1610 were changed (so the information about bb --> loop mapping does not have | |
1611 to be correct). But still for the remaining loops the header dominates | |
1612 the latch, and loops did not get new subloops (new loops might possibly | |
1613 get created, but we are not interested in them). Fix up the mess. | |
1614 | |
1615 If CHANGED_BBS is not NULL, basic blocks whose loop has changed are | |
1616 marked in it. */ | |
1617 | |
1618 void | |
1619 fix_loop_structure (bitmap changed_bbs) | |
1620 { | |
1621 basic_block bb; | |
1622 struct loop *loop, *ploop; | |
1623 loop_iterator li; | |
1624 bool record_exits = false; | |
1625 struct loop **superloop = XNEWVEC (struct loop *, number_of_loops ()); | |
1626 | |
1627 /* Remove the old bb -> loop mapping. Remember the depth of the blocks in | |
1628 the loop hierarchy, so that we can recognize blocks whose loop nesting | |
1629 relationship has changed. */ | |
1630 FOR_EACH_BB (bb) | |
1631 { | |
1632 if (changed_bbs) | |
1633 bb->aux = (void *) (size_t) loop_depth (bb->loop_father); | |
1634 bb->loop_father = current_loops->tree_root; | |
1635 } | |
1636 | |
1637 if (loops_state_satisfies_p (LOOPS_HAVE_RECORDED_EXITS)) | |
1638 { | |
1639 release_recorded_exits (); | |
1640 record_exits = true; | |
1641 } | |
1642 | |
1643 /* Remove the dead loops from structures. We start from the innermost | |
1644 loops, so that when we remove the loops, we know that the loops inside | |
1645 are preserved, and do not waste time relinking loops that will be | |
1646 removed later. */ | |
1647 FOR_EACH_LOOP (li, loop, LI_FROM_INNERMOST) | |
1648 { | |
1649 if (loop->header) | |
1650 continue; | |
1651 | |
1652 while (loop->inner) | |
1653 { | |
1654 ploop = loop->inner; | |
1655 flow_loop_tree_node_remove (ploop); | |
1656 flow_loop_tree_node_add (loop_outer (loop), ploop); | |
1657 } | |
1658 | |
1659 /* Remove the loop and free its data. */ | |
1660 delete_loop (loop); | |
1661 } | |
1662 | |
1663 /* Rescan the bodies of loops, starting from the outermost ones. We assume | |
1664 that no optimization interchanges the order of the loops, i.e., it cannot | |
1665 happen that L1 was superloop of L2 before and it is subloop of L2 now | |
1666 (without explicitly updating loop information). At the same time, we also | |
1667 determine the new loop structure. */ | |
1668 current_loops->tree_root->num_nodes = n_basic_blocks; | |
1669 FOR_EACH_LOOP (li, loop, 0) | |
1670 { | |
1671 superloop[loop->num] = loop->header->loop_father; | |
1672 loop->num_nodes = flow_loop_nodes_find (loop->header, loop); | |
1673 } | |
1674 | |
1675 /* Now fix the loop nesting. */ | |
1676 FOR_EACH_LOOP (li, loop, 0) | |
1677 { | |
1678 ploop = superloop[loop->num]; | |
1679 if (ploop != loop_outer (loop)) | |
1680 { | |
1681 flow_loop_tree_node_remove (loop); | |
1682 flow_loop_tree_node_add (ploop, loop); | |
1683 } | |
1684 } | |
1685 free (superloop); | |
1686 | |
1687 /* Mark the blocks whose loop has changed. */ | |
1688 if (changed_bbs) | |
1689 { | |
1690 FOR_EACH_BB (bb) | |
1691 { | |
1692 if ((void *) (size_t) loop_depth (bb->loop_father) != bb->aux) | |
1693 bitmap_set_bit (changed_bbs, bb->index); | |
1694 | |
1695 bb->aux = NULL; | |
1696 } | |
1697 } | |
1698 | |
1699 if (loops_state_satisfies_p (LOOPS_HAVE_PREHEADERS)) | |
1700 create_preheaders (CP_SIMPLE_PREHEADERS); | |
1701 | |
1702 if (loops_state_satisfies_p (LOOPS_HAVE_SIMPLE_LATCHES)) | |
1703 force_single_succ_latches (); | |
1704 | |
1705 if (loops_state_satisfies_p (LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS)) | |
1706 mark_irreducible_loops (); | |
1707 | |
1708 if (record_exits) | |
1709 record_loop_exits (); | |
1710 | |
1711 #ifdef ENABLE_CHECKING | |
1712 verify_loop_structure (); | |
1713 #endif | |
1714 } |