]>
Commit | Line | Data |
---|---|---|
3d436d2a | 1 | /* Loop manipulation code for GNU compiler. |
613c5cd0 | 2 | Copyright (C) 2002, 2003, 2004, 2005 Free Software Foundation, Inc. |
3d436d2a ZD |
3 | |
4 | This file is part of GCC. | |
5 | ||
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 2, or (at your option) any later | |
9 | version. | |
10 | ||
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 | |
14 | for more details. | |
15 | ||
16 | You should have received a copy of the GNU General Public License | |
17 | along with GCC; see the file COPYING. If not, write to the Free | |
366ccddb KC |
18 | Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA |
19 | 02110-1301, USA. */ | |
3d436d2a ZD |
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" | |
7932a3db | 27 | #include "obstack.h" |
3d436d2a ZD |
28 | #include "basic-block.h" |
29 | #include "cfgloop.h" | |
30 | #include "cfglayout.h" | |
1cb7dfc3 | 31 | #include "cfghooks.h" |
3d436d2a ZD |
32 | #include "output.h" |
33 | ||
d329e058 AJ |
34 | static void duplicate_subloops (struct loops *, struct loop *, struct loop *); |
35 | static void copy_loops_to (struct loops *, struct loop **, int, | |
36 | struct loop *); | |
37 | static void loop_redirect_edge (edge, basic_block); | |
38 | static bool loop_delete_branch_edge (edge, int); | |
d47cc544 | 39 | static void remove_bbs (basic_block *, int); |
d329e058 | 40 | static bool rpe_enum_p (basic_block, void *); |
d47cc544 | 41 | static int find_path (edge, basic_block **); |
d329e058 AJ |
42 | static bool alp_enum_p (basic_block, void *); |
43 | static void add_loop (struct loops *, struct loop *); | |
dc14f191 | 44 | static void fix_loop_placements (struct loops *, struct loop *, bool *); |
d329e058 | 45 | static bool fix_bb_placement (struct loops *, basic_block); |
dc14f191 | 46 | static void fix_bb_placements (struct loops *, basic_block, bool *); |
d329e058 AJ |
47 | static void place_new_loop (struct loops *, struct loop *); |
48 | static void scale_loop_frequencies (struct loop *, int, int); | |
d47cc544 | 49 | static basic_block create_preheader (struct loop *, int); |
dc14f191 | 50 | static void unloop (struct loops *, struct loop *, bool *); |
3d436d2a | 51 | |
bade3a00 JH |
52 | #define RDIV(X,Y) (((X) + (Y) / 2) / (Y)) |
53 | ||
d47cc544 | 54 | /* Checks whether basic block BB is dominated by DATA. */ |
617b465c | 55 | static bool |
d329e058 | 56 | rpe_enum_p (basic_block bb, void *data) |
617b465c | 57 | { |
d47cc544 | 58 | return dominated_by_p (CDI_DOMINATORS, bb, data); |
617b465c ZD |
59 | } |
60 | ||
61 | /* Remove basic blocks BBS from loop structure and dominance info, | |
62 | and delete them afterwards. */ | |
63 | static void | |
d47cc544 | 64 | remove_bbs (basic_block *bbs, int nbbs) |
617b465c ZD |
65 | { |
66 | int i; | |
67 | ||
68 | for (i = 0; i < nbbs; i++) | |
69 | { | |
70 | remove_bb_from_loops (bbs[i]); | |
f470c378 | 71 | delete_basic_block (bbs[i]); |
617b465c ZD |
72 | } |
73 | } | |
74 | ||
75 | /* Find path -- i.e. the basic blocks dominated by edge E and put them | |
76 | into array BBS, that will be allocated large enough to contain them. | |
35b07080 ZD |
77 | E->dest must have exactly one predecessor for this to work (it is |
78 | easy to achieve and we do not put it here because we do not want to | |
79 | alter anything by this function). The number of basic blocks in the | |
80 | path is returned. */ | |
617b465c | 81 | static int |
d47cc544 | 82 | find_path (edge e, basic_block **bbs) |
617b465c | 83 | { |
628f6a4e | 84 | gcc_assert (EDGE_COUNT (e->dest->preds) <= 1); |
617b465c ZD |
85 | |
86 | /* Find bbs in the path. */ | |
5ed6ace5 | 87 | *bbs = XCNEWVEC (basic_block, n_basic_blocks); |
617b465c | 88 | return dfs_enumerate_from (e->dest, 0, rpe_enum_p, *bbs, |
d47cc544 | 89 | n_basic_blocks, e->dest); |
617b465c ZD |
90 | } |
91 | ||
92 | /* Fix placement of basic block BB inside loop hierarchy stored in LOOPS -- | |
93 | Let L be a loop to that BB belongs. Then every successor of BB must either | |
94 | 1) belong to some superloop of loop L, or | |
95 | 2) be a header of loop K such that K->outer is superloop of L | |
96 | Returns true if we had to move BB into other loop to enforce this condition, | |
97 | false if the placement of BB was already correct (provided that placements | |
98 | of its successors are correct). */ | |
99 | static bool | |
d329e058 | 100 | fix_bb_placement (struct loops *loops, basic_block bb) |
617b465c ZD |
101 | { |
102 | edge e; | |
628f6a4e | 103 | edge_iterator ei; |
617b465c ZD |
104 | struct loop *loop = loops->tree_root, *act; |
105 | ||
628f6a4e | 106 | FOR_EACH_EDGE (e, ei, bb->succs) |
617b465c ZD |
107 | { |
108 | if (e->dest == EXIT_BLOCK_PTR) | |
109 | continue; | |
110 | ||
111 | act = e->dest->loop_father; | |
112 | if (act->header == e->dest) | |
113 | act = act->outer; | |
114 | ||
115 | if (flow_loop_nested_p (loop, act)) | |
116 | loop = act; | |
117 | } | |
118 | ||
119 | if (loop == bb->loop_father) | |
120 | return false; | |
121 | ||
122 | remove_bb_from_loops (bb); | |
123 | add_bb_to_loop (bb, loop); | |
124 | ||
125 | return true; | |
126 | } | |
127 | ||
128 | /* Fix placements of basic blocks inside loop hierarchy stored in loops; i.e. | |
129 | enforce condition condition stated in description of fix_bb_placement. We | |
130 | start from basic block FROM that had some of its successors removed, so that | |
131 | his placement no longer has to be correct, and iteratively fix placement of | |
132 | its predecessors that may change if placement of FROM changed. Also fix | |
133 | placement of subloops of FROM->loop_father, that might also be altered due | |
4d6922ee | 134 | to this change; the condition for them is similar, except that instead of |
dc14f191 ZD |
135 | successors we consider edges coming out of the loops. |
136 | ||
137 | If the changes may invalidate the information about irreducible regions, | |
138 | IRRED_INVALIDATED is set to true. */ | |
139 | ||
617b465c | 140 | static void |
dc14f191 ZD |
141 | fix_bb_placements (struct loops *loops, basic_block from, |
142 | bool *irred_invalidated) | |
617b465c ZD |
143 | { |
144 | sbitmap in_queue; | |
145 | basic_block *queue, *qtop, *qbeg, *qend; | |
146 | struct loop *base_loop; | |
147 | edge e; | |
148 | ||
149 | /* We pass through blocks back-reachable from FROM, testing whether some | |
150 | of their successors moved to outer loop. It may be necessary to | |
151 | iterate several times, but it is finite, as we stop unless we move | |
152 | the basic block up the loop structure. The whole story is a bit | |
153 | more complicated due to presence of subloops, those are moved using | |
154 | fix_loop_placement. */ | |
155 | ||
156 | base_loop = from->loop_father; | |
157 | if (base_loop == loops->tree_root) | |
158 | return; | |
159 | ||
160 | in_queue = sbitmap_alloc (last_basic_block); | |
161 | sbitmap_zero (in_queue); | |
162 | SET_BIT (in_queue, from->index); | |
163 | /* Prevent us from going out of the base_loop. */ | |
164 | SET_BIT (in_queue, base_loop->header->index); | |
165 | ||
5ed6ace5 | 166 | queue = XNEWVEC (basic_block, base_loop->num_nodes + 1); |
617b465c ZD |
167 | qtop = queue + base_loop->num_nodes + 1; |
168 | qbeg = queue; | |
169 | qend = queue + 1; | |
170 | *qbeg = from; | |
171 | ||
172 | while (qbeg != qend) | |
173 | { | |
628f6a4e | 174 | edge_iterator ei; |
617b465c ZD |
175 | from = *qbeg; |
176 | qbeg++; | |
177 | if (qbeg == qtop) | |
178 | qbeg = queue; | |
179 | RESET_BIT (in_queue, from->index); | |
180 | ||
181 | if (from->loop_father->header == from) | |
182 | { | |
183 | /* Subloop header, maybe move the loop upward. */ | |
184 | if (!fix_loop_placement (from->loop_father)) | |
185 | continue; | |
186 | } | |
187 | else | |
188 | { | |
189 | /* Ordinary basic block. */ | |
190 | if (!fix_bb_placement (loops, from)) | |
191 | continue; | |
192 | } | |
193 | ||
dc14f191 ZD |
194 | FOR_EACH_EDGE (e, ei, from->succs) |
195 | { | |
196 | if (e->flags & EDGE_IRREDUCIBLE_LOOP) | |
197 | *irred_invalidated = true; | |
198 | } | |
199 | ||
617b465c | 200 | /* Something has changed, insert predecessors into queue. */ |
628f6a4e | 201 | FOR_EACH_EDGE (e, ei, from->preds) |
617b465c ZD |
202 | { |
203 | basic_block pred = e->src; | |
204 | struct loop *nca; | |
205 | ||
dc14f191 ZD |
206 | if (e->flags & EDGE_IRREDUCIBLE_LOOP) |
207 | *irred_invalidated = true; | |
208 | ||
617b465c ZD |
209 | if (TEST_BIT (in_queue, pred->index)) |
210 | continue; | |
211 | ||
d329e058 | 212 | /* If it is subloop, then it either was not moved, or |
617b465c ZD |
213 | the path up the loop tree from base_loop do not contain |
214 | it. */ | |
215 | nca = find_common_loop (pred->loop_father, base_loop); | |
216 | if (pred->loop_father != base_loop | |
217 | && (nca == base_loop | |
218 | || nca != pred->loop_father)) | |
219 | pred = pred->loop_father->header; | |
220 | else if (!flow_loop_nested_p (from->loop_father, pred->loop_father)) | |
221 | { | |
222 | /* No point in processing it. */ | |
223 | continue; | |
224 | } | |
225 | ||
226 | if (TEST_BIT (in_queue, pred->index)) | |
227 | continue; | |
228 | ||
229 | /* Schedule the basic block. */ | |
230 | *qend = pred; | |
231 | qend++; | |
232 | if (qend == qtop) | |
233 | qend = queue; | |
234 | SET_BIT (in_queue, pred->index); | |
235 | } | |
236 | } | |
237 | free (in_queue); | |
238 | free (queue); | |
239 | } | |
240 | ||
241 | /* Removes path beginning at edge E, i.e. remove basic blocks dominated by E | |
242 | and update loop structure stored in LOOPS and dominators. Return true if | |
243 | we were able to remove the path, false otherwise (and nothing is affected | |
244 | then). */ | |
245 | bool | |
d329e058 | 246 | remove_path (struct loops *loops, edge e) |
617b465c ZD |
247 | { |
248 | edge ae; | |
249 | basic_block *rem_bbs, *bord_bbs, *dom_bbs, from, bb; | |
250 | int i, nrem, n_bord_bbs, n_dom_bbs; | |
251 | sbitmap seen; | |
dc14f191 | 252 | bool deleted, irred_invalidated = false; |
617b465c | 253 | |
35b07080 ZD |
254 | if (!loop_delete_branch_edge (e, 0)) |
255 | return false; | |
256 | ||
dc14f191 ZD |
257 | /* Keep track of whether we need to update information about irreducible |
258 | regions. This is the case if the removed area is a part of the | |
259 | irreducible region, or if the set of basic blocks that belong to a loop | |
260 | that is inside an irreducible region is changed, or if such a loop is | |
261 | removed. */ | |
262 | if (e->flags & EDGE_IRREDUCIBLE_LOOP) | |
263 | irred_invalidated = true; | |
264 | ||
35b07080 ZD |
265 | /* We need to check whether basic blocks are dominated by the edge |
266 | e, but we only have basic block dominators. This is easy to | |
267 | fix -- when e->dest has exactly one predecessor, this corresponds | |
268 | to blocks dominated by e->dest, if not, split the edge. */ | |
c5cbcccf ZD |
269 | if (!single_pred_p (e->dest)) |
270 | e = single_pred_edge (loop_split_edge_with (e, NULL_RTX)); | |
35b07080 ZD |
271 | |
272 | /* It may happen that by removing path we remove one or more loops | |
273 | we belong to. In this case first unloop the loops, then proceed | |
274 | normally. We may assume that e->dest is not a header of any loop, | |
275 | as it now has exactly one predecessor. */ | |
276 | while (e->src->loop_father->outer | |
d47cc544 | 277 | && dominated_by_p (CDI_DOMINATORS, |
35b07080 | 278 | e->src->loop_father->latch, e->dest)) |
dc14f191 | 279 | unloop (loops, e->src->loop_father, &irred_invalidated); |
d329e058 | 280 | |
35b07080 | 281 | /* Identify the path. */ |
d47cc544 | 282 | nrem = find_path (e, &rem_bbs); |
617b465c ZD |
283 | |
284 | n_bord_bbs = 0; | |
5ed6ace5 | 285 | bord_bbs = XCNEWVEC (basic_block, n_basic_blocks); |
617b465c ZD |
286 | seen = sbitmap_alloc (last_basic_block); |
287 | sbitmap_zero (seen); | |
288 | ||
289 | /* Find "border" hexes -- i.e. those with predecessor in removed path. */ | |
290 | for (i = 0; i < nrem; i++) | |
291 | SET_BIT (seen, rem_bbs[i]->index); | |
35b07080 | 292 | for (i = 0; i < nrem; i++) |
617b465c | 293 | { |
628f6a4e | 294 | edge_iterator ei; |
35b07080 | 295 | bb = rem_bbs[i]; |
628f6a4e | 296 | FOR_EACH_EDGE (ae, ei, rem_bbs[i]->succs) |
35b07080 ZD |
297 | if (ae->dest != EXIT_BLOCK_PTR && !TEST_BIT (seen, ae->dest->index)) |
298 | { | |
299 | SET_BIT (seen, ae->dest->index); | |
300 | bord_bbs[n_bord_bbs++] = ae->dest; | |
dc14f191 ZD |
301 | |
302 | if (ae->flags & EDGE_IRREDUCIBLE_LOOP) | |
303 | irred_invalidated = true; | |
35b07080 | 304 | } |
617b465c | 305 | } |
617b465c ZD |
306 | |
307 | /* Remove the path. */ | |
308 | from = e->src; | |
341c100f NS |
309 | deleted = loop_delete_branch_edge (e, 1); |
310 | gcc_assert (deleted); | |
5ed6ace5 | 311 | dom_bbs = XCNEWVEC (basic_block, n_basic_blocks); |
617b465c ZD |
312 | |
313 | /* Cancel loops contained in the path. */ | |
314 | for (i = 0; i < nrem; i++) | |
315 | if (rem_bbs[i]->loop_father->header == rem_bbs[i]) | |
316 | cancel_loop_tree (loops, rem_bbs[i]->loop_father); | |
317 | ||
d47cc544 | 318 | remove_bbs (rem_bbs, nrem); |
617b465c ZD |
319 | free (rem_bbs); |
320 | ||
35b07080 | 321 | /* Find blocks whose dominators may be affected. */ |
617b465c ZD |
322 | n_dom_bbs = 0; |
323 | sbitmap_zero (seen); | |
324 | for (i = 0; i < n_bord_bbs; i++) | |
325 | { | |
d47cc544 | 326 | basic_block ldom; |
617b465c | 327 | |
d47cc544 | 328 | bb = get_immediate_dominator (CDI_DOMINATORS, bord_bbs[i]); |
617b465c ZD |
329 | if (TEST_BIT (seen, bb->index)) |
330 | continue; | |
331 | SET_BIT (seen, bb->index); | |
332 | ||
d47cc544 SB |
333 | for (ldom = first_dom_son (CDI_DOMINATORS, bb); |
334 | ldom; | |
335 | ldom = next_dom_son (CDI_DOMINATORS, ldom)) | |
336 | if (!dominated_by_p (CDI_DOMINATORS, from, ldom)) | |
337 | dom_bbs[n_dom_bbs++] = ldom; | |
617b465c ZD |
338 | } |
339 | ||
617b465c ZD |
340 | free (seen); |
341 | ||
342 | /* Recount dominators. */ | |
d47cc544 | 343 | iterate_fix_dominators (CDI_DOMINATORS, dom_bbs, n_dom_bbs); |
617b465c | 344 | free (dom_bbs); |
35b07080 ZD |
345 | free (bord_bbs); |
346 | ||
617b465c ZD |
347 | /* Fix placements of basic blocks inside loops and the placement of |
348 | loops in the loop tree. */ | |
dc14f191 ZD |
349 | fix_bb_placements (loops, from, &irred_invalidated); |
350 | fix_loop_placements (loops, from->loop_father, &irred_invalidated); | |
351 | ||
352 | if (irred_invalidated | |
353 | && (loops->state & LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS) != 0) | |
354 | mark_irreducible_loops (loops); | |
617b465c ZD |
355 | |
356 | return true; | |
357 | } | |
358 | ||
359 | /* Predicate for enumeration in add_loop. */ | |
360 | static bool | |
d329e058 | 361 | alp_enum_p (basic_block bb, void *alp_header) |
617b465c ZD |
362 | { |
363 | return bb != (basic_block) alp_header; | |
364 | } | |
365 | ||
366 | /* Given LOOP structure with filled header and latch, find the body of the | |
367 | corresponding loop and add it to LOOPS tree. */ | |
368 | static void | |
d329e058 | 369 | add_loop (struct loops *loops, struct loop *loop) |
617b465c ZD |
370 | { |
371 | basic_block *bbs; | |
372 | int i, n; | |
d329e058 | 373 | |
617b465c ZD |
374 | /* Add it to loop structure. */ |
375 | place_new_loop (loops, loop); | |
376 | loop->level = 1; | |
377 | ||
378 | /* Find its nodes. */ | |
5ed6ace5 | 379 | bbs = XCNEWVEC (basic_block, n_basic_blocks); |
617b465c ZD |
380 | n = dfs_enumerate_from (loop->latch, 1, alp_enum_p, |
381 | bbs, n_basic_blocks, loop->header); | |
382 | ||
383 | for (i = 0; i < n; i++) | |
384 | add_bb_to_loop (bbs[i], loop); | |
385 | add_bb_to_loop (loop->header, loop); | |
386 | ||
387 | free (bbs); | |
388 | } | |
389 | ||
617b465c ZD |
390 | /* Multiply all frequencies in LOOP by NUM/DEN. */ |
391 | static void | |
d329e058 | 392 | scale_loop_frequencies (struct loop *loop, int num, int den) |
617b465c ZD |
393 | { |
394 | basic_block *bbs; | |
395 | ||
396 | bbs = get_loop_body (loop); | |
33156717 | 397 | scale_bbs_frequencies_int (bbs, loop->num_nodes, num, den); |
617b465c ZD |
398 | free (bbs); |
399 | } | |
400 | ||
401 | /* Make area between HEADER_EDGE and LATCH_EDGE a loop by connecting | |
402 | latch to header and update loop tree stored in LOOPS and dominators | |
403 | accordingly. Everything between them plus LATCH_EDGE destination must | |
404 | be dominated by HEADER_EDGE destination, and back-reachable from | |
405 | LATCH_EDGE source. HEADER_EDGE is redirected to basic block SWITCH_BB, | |
5132abc2 KH |
406 | FALSE_EDGE of SWITCH_BB to original destination of HEADER_EDGE and |
407 | TRUE_EDGE of SWITCH_BB to original destination of LATCH_EDGE. | |
617b465c | 408 | Returns newly created loop. */ |
50654f6c | 409 | |
617b465c | 410 | struct loop * |
c22cacf3 | 411 | loopify (struct loops *loops, edge latch_edge, edge header_edge, |
5132abc2 KH |
412 | basic_block switch_bb, edge true_edge, edge false_edge, |
413 | bool redirect_all_edges) | |
617b465c ZD |
414 | { |
415 | basic_block succ_bb = latch_edge->dest; | |
416 | basic_block pred_bb = header_edge->src; | |
417 | basic_block *dom_bbs, *body; | |
d47cc544 | 418 | unsigned n_dom_bbs, i; |
617b465c | 419 | sbitmap seen; |
5ed6ace5 | 420 | struct loop *loop = XCNEW (struct loop); |
617b465c ZD |
421 | struct loop *outer = succ_bb->loop_father->outer; |
422 | int freq, prob, tot_prob; | |
423 | gcov_type cnt; | |
424 | edge e; | |
628f6a4e | 425 | edge_iterator ei; |
617b465c ZD |
426 | |
427 | loop->header = header_edge->dest; | |
428 | loop->latch = latch_edge->src; | |
429 | ||
430 | freq = EDGE_FREQUENCY (header_edge); | |
431 | cnt = header_edge->count; | |
628f6a4e BE |
432 | prob = EDGE_SUCC (switch_bb, 0)->probability; |
433 | tot_prob = prob + EDGE_SUCC (switch_bb, 1)->probability; | |
617b465c ZD |
434 | if (tot_prob == 0) |
435 | tot_prob = 1; | |
436 | ||
437 | /* Redirect edges. */ | |
438 | loop_redirect_edge (latch_edge, loop->header); | |
5132abc2 | 439 | loop_redirect_edge (true_edge, succ_bb); |
50654f6c | 440 | |
92fc4a2f ZD |
441 | /* During loop versioning, one of the switch_bb edge is already properly |
442 | set. Do not redirect it again unless redirect_all_edges is true. */ | |
443 | if (redirect_all_edges) | |
444 | { | |
445 | loop_redirect_edge (header_edge, switch_bb); | |
c22cacf3 MS |
446 | loop_redirect_edge (false_edge, loop->header); |
447 | ||
92fc4a2f ZD |
448 | /* Update dominators. */ |
449 | set_immediate_dominator (CDI_DOMINATORS, switch_bb, pred_bb); | |
450 | set_immediate_dominator (CDI_DOMINATORS, loop->header, switch_bb); | |
451 | } | |
50654f6c | 452 | |
d47cc544 | 453 | set_immediate_dominator (CDI_DOMINATORS, succ_bb, switch_bb); |
617b465c ZD |
454 | |
455 | /* Compute new loop. */ | |
456 | add_loop (loops, loop); | |
457 | flow_loop_tree_node_add (outer, loop); | |
458 | ||
459 | /* Add switch_bb to appropriate loop. */ | |
460 | add_bb_to_loop (switch_bb, outer); | |
461 | ||
462 | /* Fix frequencies. */ | |
463 | switch_bb->frequency = freq; | |
464 | switch_bb->count = cnt; | |
628f6a4e | 465 | FOR_EACH_EDGE (e, ei, switch_bb->succs) |
617b465c ZD |
466 | e->count = (switch_bb->count * e->probability) / REG_BR_PROB_BASE; |
467 | scale_loop_frequencies (loop, prob, tot_prob); | |
468 | scale_loop_frequencies (succ_bb->loop_father, tot_prob - prob, tot_prob); | |
469 | ||
470 | /* Update dominators of blocks outside of LOOP. */ | |
5ed6ace5 | 471 | dom_bbs = XCNEWVEC (basic_block, n_basic_blocks); |
617b465c ZD |
472 | n_dom_bbs = 0; |
473 | seen = sbitmap_alloc (last_basic_block); | |
474 | sbitmap_zero (seen); | |
475 | body = get_loop_body (loop); | |
476 | ||
477 | for (i = 0; i < loop->num_nodes; i++) | |
478 | SET_BIT (seen, body[i]->index); | |
479 | ||
480 | for (i = 0; i < loop->num_nodes; i++) | |
481 | { | |
d47cc544 | 482 | basic_block ldom; |
617b465c | 483 | |
d47cc544 SB |
484 | for (ldom = first_dom_son (CDI_DOMINATORS, body[i]); |
485 | ldom; | |
486 | ldom = next_dom_son (CDI_DOMINATORS, ldom)) | |
487 | if (!TEST_BIT (seen, ldom->index)) | |
617b465c | 488 | { |
d47cc544 SB |
489 | SET_BIT (seen, ldom->index); |
490 | dom_bbs[n_dom_bbs++] = ldom; | |
617b465c | 491 | } |
617b465c ZD |
492 | } |
493 | ||
d47cc544 | 494 | iterate_fix_dominators (CDI_DOMINATORS, dom_bbs, n_dom_bbs); |
617b465c ZD |
495 | |
496 | free (body); | |
497 | free (seen); | |
498 | free (dom_bbs); | |
499 | ||
500 | return loop; | |
501 | } | |
502 | ||
35b07080 ZD |
503 | /* Remove the latch edge of a LOOP and update LOOPS tree to indicate that |
504 | the LOOP was removed. After this function, original loop latch will | |
dc14f191 ZD |
505 | have no successor, which caller is expected to fix somehow. |
506 | ||
507 | If this may cause the information about irreducible regions to become | |
508 | invalid, IRRED_INVALIDATED is set to true. */ | |
509 | ||
25a6c68b | 510 | static void |
dc14f191 | 511 | unloop (struct loops *loops, struct loop *loop, bool *irred_invalidated) |
35b07080 ZD |
512 | { |
513 | basic_block *body; | |
514 | struct loop *ploop; | |
515 | unsigned i, n; | |
516 | basic_block latch = loop->latch; | |
dc14f191 ZD |
517 | bool dummy = false; |
518 | ||
519 | if (loop_preheader_edge (loop)->flags & EDGE_IRREDUCIBLE_LOOP) | |
520 | *irred_invalidated = true; | |
35b07080 | 521 | |
e0bb17a8 | 522 | /* This is relatively straightforward. The dominators are unchanged, as |
35b07080 ZD |
523 | loop header dominates loop latch, so the only thing we have to care of |
524 | is the placement of loops and basic blocks inside the loop tree. We | |
525 | move them all to the loop->outer, and then let fix_bb_placements do | |
526 | its work. */ | |
527 | ||
528 | body = get_loop_body (loop); | |
35b07080 ZD |
529 | n = loop->num_nodes; |
530 | for (i = 0; i < n; i++) | |
531 | if (body[i]->loop_father == loop) | |
532 | { | |
533 | remove_bb_from_loops (body[i]); | |
534 | add_bb_to_loop (body[i], loop->outer); | |
535 | } | |
536 | free(body); | |
537 | ||
538 | while (loop->inner) | |
539 | { | |
540 | ploop = loop->inner; | |
541 | flow_loop_tree_node_remove (ploop); | |
542 | flow_loop_tree_node_add (loop->outer, ploop); | |
543 | } | |
544 | ||
545 | /* Remove the loop and free its data. */ | |
546 | flow_loop_tree_node_remove (loop); | |
547 | loops->parray[loop->num] = NULL; | |
548 | flow_loop_free (loop); | |
549 | ||
c5cbcccf | 550 | remove_edge (single_succ_edge (latch)); |
dc14f191 ZD |
551 | |
552 | /* We do not pass IRRED_INVALIDATED to fix_bb_placements here, as even if | |
553 | there is an irreducible region inside the cancelled loop, the flags will | |
554 | be still correct. */ | |
555 | fix_bb_placements (loops, latch, &dummy); | |
35b07080 ZD |
556 | } |
557 | ||
617b465c | 558 | /* Fix placement of LOOP inside loop tree, i.e. find the innermost superloop |
2067c116 | 559 | FATHER of LOOP such that all of the edges coming out of LOOP belong to |
dc14f191 | 560 | FATHER, and set it as outer loop of LOOP. Return true if placement of |
617b465c | 561 | LOOP changed. */ |
dc14f191 | 562 | |
617b465c | 563 | int |
d329e058 | 564 | fix_loop_placement (struct loop *loop) |
617b465c ZD |
565 | { |
566 | basic_block *body; | |
567 | unsigned i; | |
568 | edge e; | |
628f6a4e | 569 | edge_iterator ei; |
617b465c ZD |
570 | struct loop *father = loop->pred[0], *act; |
571 | ||
572 | body = get_loop_body (loop); | |
573 | for (i = 0; i < loop->num_nodes; i++) | |
628f6a4e | 574 | FOR_EACH_EDGE (e, ei, body[i]->succs) |
617b465c ZD |
575 | if (!flow_bb_inside_loop_p (loop, e->dest)) |
576 | { | |
577 | act = find_common_loop (loop, e->dest->loop_father); | |
578 | if (flow_loop_nested_p (father, act)) | |
579 | father = act; | |
580 | } | |
581 | free (body); | |
582 | ||
583 | if (father != loop->outer) | |
584 | { | |
585 | for (act = loop->outer; act != father; act = act->outer) | |
586 | act->num_nodes -= loop->num_nodes; | |
587 | flow_loop_tree_node_remove (loop); | |
588 | flow_loop_tree_node_add (father, loop); | |
589 | return 1; | |
590 | } | |
591 | return 0; | |
592 | } | |
593 | ||
594 | /* Fix placement of superloops of LOOP inside loop tree, i.e. ensure that | |
595 | condition stated in description of fix_loop_placement holds for them. | |
596 | It is used in case when we removed some edges coming out of LOOP, which | |
dc14f191 ZD |
597 | may cause the right placement of LOOP inside loop tree to change. |
598 | ||
599 | IRRED_INVALIDATED is set to true if a change in the loop structures might | |
600 | invalidate the information about irreducible regions. */ | |
601 | ||
617b465c | 602 | static void |
dc14f191 ZD |
603 | fix_loop_placements (struct loops *loops, struct loop *loop, |
604 | bool *irred_invalidated) | |
617b465c ZD |
605 | { |
606 | struct loop *outer; | |
607 | ||
608 | while (loop->outer) | |
609 | { | |
610 | outer = loop->outer; | |
611 | if (!fix_loop_placement (loop)) | |
c22cacf3 | 612 | break; |
1548580c EB |
613 | |
614 | /* Changing the placement of a loop in the loop tree may alter the | |
615 | validity of condition 2) of the description of fix_bb_placement | |
616 | for its preheader, because the successor is the header and belongs | |
617 | to the loop. So call fix_bb_placements to fix up the placement | |
618 | of the preheader and (possibly) of its predecessors. */ | |
dc14f191 ZD |
619 | fix_bb_placements (loops, loop_preheader_edge (loop)->src, |
620 | irred_invalidated); | |
617b465c ZD |
621 | loop = outer; |
622 | } | |
623 | } | |
624 | ||
625 | /* Creates place for a new LOOP in LOOPS structure. */ | |
626 | static void | |
d329e058 | 627 | place_new_loop (struct loops *loops, struct loop *loop) |
617b465c ZD |
628 | { |
629 | loops->parray = | |
630 | xrealloc (loops->parray, (loops->num + 1) * sizeof (struct loop *)); | |
631 | loops->parray[loops->num] = loop; | |
632 | ||
633 | loop->num = loops->num++; | |
634 | } | |
635 | ||
636 | /* Copies copy of LOOP as subloop of TARGET loop, placing newly | |
637 | created loop into LOOPS structure. */ | |
f67d92e9 | 638 | struct loop * |
d329e058 | 639 | duplicate_loop (struct loops *loops, struct loop *loop, struct loop *target) |
617b465c ZD |
640 | { |
641 | struct loop *cloop; | |
5ed6ace5 | 642 | cloop = XCNEW (struct loop); |
617b465c ZD |
643 | place_new_loop (loops, cloop); |
644 | ||
645 | /* Initialize copied loop. */ | |
646 | cloop->level = loop->level; | |
647 | ||
648 | /* Set it as copy of loop. */ | |
649 | loop->copy = cloop; | |
650 | ||
651 | /* Add it to target. */ | |
652 | flow_loop_tree_node_add (target, cloop); | |
653 | ||
654 | return cloop; | |
655 | } | |
656 | ||
657 | /* Copies structure of subloops of LOOP into TARGET loop, placing | |
658 | newly created loops into loop tree stored in LOOPS. */ | |
d329e058 AJ |
659 | static void |
660 | duplicate_subloops (struct loops *loops, struct loop *loop, struct loop *target) | |
617b465c ZD |
661 | { |
662 | struct loop *aloop, *cloop; | |
663 | ||
664 | for (aloop = loop->inner; aloop; aloop = aloop->next) | |
665 | { | |
666 | cloop = duplicate_loop (loops, aloop, target); | |
667 | duplicate_subloops (loops, aloop, cloop); | |
668 | } | |
669 | } | |
670 | ||
671 | /* Copies structure of subloops of N loops, stored in array COPIED_LOOPS, | |
672 | into TARGET loop, placing newly created loops into loop tree LOOPS. */ | |
d329e058 AJ |
673 | static void |
674 | copy_loops_to (struct loops *loops, struct loop **copied_loops, int n, struct loop *target) | |
617b465c ZD |
675 | { |
676 | struct loop *aloop; | |
677 | int i; | |
678 | ||
679 | for (i = 0; i < n; i++) | |
680 | { | |
681 | aloop = duplicate_loop (loops, copied_loops[i], target); | |
682 | duplicate_subloops (loops, copied_loops[i], aloop); | |
683 | } | |
684 | } | |
685 | ||
686 | /* Redirects edge E to basic block DEST. */ | |
687 | static void | |
d329e058 | 688 | loop_redirect_edge (edge e, basic_block dest) |
617b465c ZD |
689 | { |
690 | if (e->dest == dest) | |
691 | return; | |
692 | ||
9ee634e3 | 693 | redirect_edge_and_branch_force (e, dest); |
617b465c ZD |
694 | } |
695 | ||
35b07080 ZD |
696 | /* Deletes edge E from a branch if possible. Unless REALLY_DELETE is set, |
697 | just test whether it is possible to remove the edge. */ | |
617b465c | 698 | static bool |
d329e058 | 699 | loop_delete_branch_edge (edge e, int really_delete) |
617b465c ZD |
700 | { |
701 | basic_block src = e->src; | |
341c100f | 702 | basic_block newdest; |
35b07080 ZD |
703 | int irr; |
704 | edge snd; | |
617b465c | 705 | |
628f6a4e | 706 | gcc_assert (EDGE_COUNT (src->succs) > 1); |
c22cacf3 | 707 | |
341c100f | 708 | /* Cannot handle more than two exit edges. */ |
628f6a4e | 709 | if (EDGE_COUNT (src->succs) > 2) |
341c100f NS |
710 | return false; |
711 | /* And it must be just a simple branch. */ | |
712 | if (!any_condjump_p (BB_END (src))) | |
713 | return false; | |
35b07080 | 714 | |
628f6a4e | 715 | snd = e == EDGE_SUCC (src, 0) ? EDGE_SUCC (src, 1) : EDGE_SUCC (src, 0); |
341c100f NS |
716 | newdest = snd->dest; |
717 | if (newdest == EXIT_BLOCK_PTR) | |
718 | return false; | |
d329e058 | 719 | |
341c100f NS |
720 | /* Hopefully the above conditions should suffice. */ |
721 | if (!really_delete) | |
722 | return true; | |
35b07080 | 723 | |
341c100f NS |
724 | /* Redirecting behaves wrongly wrto this flag. */ |
725 | irr = snd->flags & EDGE_IRREDUCIBLE_LOOP; | |
617b465c | 726 | |
341c100f NS |
727 | if (!redirect_edge_and_branch (e, newdest)) |
728 | return false; | |
c5cbcccf ZD |
729 | single_succ_edge (src)->flags &= ~EDGE_IRREDUCIBLE_LOOP; |
730 | single_succ_edge (src)->flags |= irr; | |
c22cacf3 | 731 | |
341c100f | 732 | return true; |
617b465c ZD |
733 | } |
734 | ||
617b465c ZD |
735 | /* Check whether LOOP's body can be duplicated. */ |
736 | bool | |
d329e058 | 737 | can_duplicate_loop_p (struct loop *loop) |
617b465c | 738 | { |
8d28e87d ZD |
739 | int ret; |
740 | basic_block *bbs = get_loop_body (loop); | |
617b465c | 741 | |
8d28e87d | 742 | ret = can_copy_bbs_p (bbs, loop->num_nodes); |
617b465c | 743 | free (bbs); |
c22cacf3 | 744 | |
8d28e87d | 745 | return ret; |
617b465c ZD |
746 | } |
747 | ||
82b85a85 ZD |
748 | /* The NBBS blocks in BBS will get duplicated and the copies will be placed |
749 | to LOOP. Update the single_exit information in superloops of LOOP. */ | |
750 | ||
25a6c68b | 751 | static void |
82b85a85 ZD |
752 | update_single_exits_after_duplication (basic_block *bbs, unsigned nbbs, |
753 | struct loop *loop) | |
754 | { | |
755 | unsigned i; | |
756 | ||
757 | for (i = 0; i < nbbs; i++) | |
6580ee77 | 758 | bbs[i]->flags |= BB_DUPLICATED; |
82b85a85 ZD |
759 | |
760 | for (; loop->outer; loop = loop->outer) | |
761 | { | |
762 | if (!loop->single_exit) | |
763 | continue; | |
764 | ||
6580ee77 | 765 | if (loop->single_exit->src->flags & BB_DUPLICATED) |
82b85a85 ZD |
766 | loop->single_exit = NULL; |
767 | } | |
768 | ||
769 | for (i = 0; i < nbbs; i++) | |
6580ee77 | 770 | bbs[i]->flags &= ~BB_DUPLICATED; |
82b85a85 ZD |
771 | } |
772 | ||
c0493b13 ZD |
773 | /* Updates single exit information for the copy of LOOP. */ |
774 | ||
775 | static void | |
776 | update_single_exit_for_duplicated_loop (struct loop *loop) | |
777 | { | |
778 | struct loop *copy = loop->copy; | |
779 | basic_block src, dest; | |
780 | edge exit = loop->single_exit; | |
781 | ||
782 | if (!exit) | |
783 | return; | |
784 | ||
785 | src = get_bb_copy (exit->src); | |
786 | dest = exit->dest; | |
787 | if (dest->flags & BB_DUPLICATED) | |
788 | dest = get_bb_copy (dest); | |
789 | ||
790 | exit = find_edge (src, dest); | |
791 | gcc_assert (exit != NULL); | |
792 | copy->single_exit = exit; | |
793 | } | |
794 | ||
795 | /* Updates single exit information for copies of ORIG_LOOPS and their subloops. | |
796 | N is the number of the loops in the ORIG_LOOPS array. */ | |
797 | ||
798 | static void | |
799 | update_single_exit_for_duplicated_loops (struct loop *orig_loops[], unsigned n) | |
800 | { | |
801 | unsigned i; | |
802 | ||
803 | for (i = 0; i < n; i++) | |
804 | update_single_exit_for_duplicated_loop (orig_loops[i]); | |
805 | } | |
806 | ||
8d28e87d ZD |
807 | /* Duplicates body of LOOP to given edge E NDUPL times. Takes care of updating |
808 | LOOPS structure and dominators. E's destination must be LOOP header for | |
809 | this to work, i.e. it must be entry or latch edge of this loop; these are | |
810 | unique, as the loops must have preheaders for this function to work | |
811 | correctly (in case E is latch, the function unrolls the loop, if E is entry | |
812 | edge, it peels the loop). Store edges created by copying ORIG edge from | |
813 | copies corresponding to set bits in WONT_EXIT bitmap (bit 0 corresponds to | |
814 | original LOOP body, the other copies are numbered in order given by control | |
815 | flow through them) into TO_REMOVE array. Returns false if duplication is | |
816 | impossible. */ | |
1cb7dfc3 | 817 | bool |
d329e058 AJ |
818 | duplicate_loop_to_header_edge (struct loop *loop, edge e, struct loops *loops, |
819 | unsigned int ndupl, sbitmap wont_exit, | |
820 | edge orig, edge *to_remove, | |
821 | unsigned int *n_to_remove, int flags) | |
617b465c ZD |
822 | { |
823 | struct loop *target, *aloop; | |
824 | struct loop **orig_loops; | |
825 | unsigned n_orig_loops; | |
826 | basic_block header = loop->header, latch = loop->latch; | |
827 | basic_block *new_bbs, *bbs, *first_active; | |
828 | basic_block new_bb, bb, first_active_latch = NULL; | |
8d28e87d ZD |
829 | edge ae, latch_edge; |
830 | edge spec_edges[2], new_spec_edges[2]; | |
831 | #define SE_LATCH 0 | |
832 | #define SE_ORIG 1 | |
617b465c ZD |
833 | unsigned i, j, n; |
834 | int is_latch = (latch == e->src); | |
835 | int scale_act = 0, *scale_step = NULL, scale_main = 0; | |
836 | int p, freq_in, freq_le, freq_out_orig; | |
837 | int prob_pass_thru, prob_pass_wont_exit, prob_pass_main; | |
838 | int add_irreducible_flag; | |
b9a66240 | 839 | basic_block place_after; |
617b465c | 840 | |
341c100f NS |
841 | gcc_assert (e->dest == loop->header); |
842 | gcc_assert (ndupl > 0); | |
617b465c ZD |
843 | |
844 | if (orig) | |
845 | { | |
846 | /* Orig must be edge out of the loop. */ | |
341c100f NS |
847 | gcc_assert (flow_bb_inside_loop_p (loop, orig->src)); |
848 | gcc_assert (!flow_bb_inside_loop_p (loop, orig->dest)); | |
617b465c ZD |
849 | } |
850 | ||
b9a66240 ZD |
851 | n = loop->num_nodes; |
852 | bbs = get_loop_body_in_dom_order (loop); | |
853 | gcc_assert (bbs[0] == loop->header); | |
854 | gcc_assert (bbs[n - 1] == loop->latch); | |
617b465c ZD |
855 | |
856 | /* Check whether duplication is possible. */ | |
8d28e87d | 857 | if (!can_copy_bbs_p (bbs, loop->num_nodes)) |
617b465c | 858 | { |
8d28e87d ZD |
859 | free (bbs); |
860 | return false; | |
617b465c | 861 | } |
5ed6ace5 | 862 | new_bbs = XNEWVEC (basic_block, loop->num_nodes); |
617b465c | 863 | |
8d28e87d ZD |
864 | /* In case we are doing loop peeling and the loop is in the middle of |
865 | irreducible region, the peeled copies will be inside it too. */ | |
866 | add_irreducible_flag = e->flags & EDGE_IRREDUCIBLE_LOOP; | |
341c100f | 867 | gcc_assert (!is_latch || !add_irreducible_flag); |
617b465c ZD |
868 | |
869 | /* Find edge from latch. */ | |
870 | latch_edge = loop_latch_edge (loop); | |
871 | ||
872 | if (flags & DLTHE_FLAG_UPDATE_FREQ) | |
873 | { | |
874 | /* Calculate coefficients by that we have to scale frequencies | |
875 | of duplicated loop bodies. */ | |
876 | freq_in = header->frequency; | |
877 | freq_le = EDGE_FREQUENCY (latch_edge); | |
878 | if (freq_in == 0) | |
879 | freq_in = 1; | |
880 | if (freq_in < freq_le) | |
881 | freq_in = freq_le; | |
882 | freq_out_orig = orig ? EDGE_FREQUENCY (orig) : freq_in - freq_le; | |
883 | if (freq_out_orig > freq_in - freq_le) | |
884 | freq_out_orig = freq_in - freq_le; | |
885 | prob_pass_thru = RDIV (REG_BR_PROB_BASE * freq_le, freq_in); | |
886 | prob_pass_wont_exit = | |
887 | RDIV (REG_BR_PROB_BASE * (freq_le + freq_out_orig), freq_in); | |
888 | ||
5ed6ace5 | 889 | scale_step = XNEWVEC (int, ndupl); |
617b465c ZD |
890 | |
891 | for (i = 1; i <= ndupl; i++) | |
d329e058 | 892 | scale_step[i - 1] = TEST_BIT (wont_exit, i) |
617b465c ZD |
893 | ? prob_pass_wont_exit |
894 | : prob_pass_thru; | |
895 | ||
a4d05547 | 896 | /* Complete peeling is special as the probability of exit in last |
c22cacf3 | 897 | copy becomes 1. */ |
178df94f JH |
898 | if (flags & DLTHE_FLAG_COMPLETTE_PEEL) |
899 | { | |
900 | int wanted_freq = EDGE_FREQUENCY (e); | |
901 | ||
902 | if (wanted_freq > freq_in) | |
903 | wanted_freq = freq_in; | |
904 | ||
905 | gcc_assert (!is_latch); | |
a4d05547 | 906 | /* First copy has frequency of incoming edge. Each subsequent |
178df94f JH |
907 | frequency should be reduced by prob_pass_wont_exit. Caller |
908 | should've managed the flags so all except for original loop | |
909 | has won't exist set. */ | |
910 | scale_act = RDIV (wanted_freq * REG_BR_PROB_BASE, freq_in); | |
911 | /* Now simulate the duplication adjustments and compute header | |
912 | frequency of the last copy. */ | |
913 | for (i = 0; i < ndupl; i++) | |
c22cacf3 | 914 | wanted_freq = RDIV (wanted_freq * scale_step[i], REG_BR_PROB_BASE); |
178df94f JH |
915 | scale_main = RDIV (wanted_freq * REG_BR_PROB_BASE, freq_in); |
916 | } | |
917 | else if (is_latch) | |
617b465c ZD |
918 | { |
919 | prob_pass_main = TEST_BIT (wont_exit, 0) | |
920 | ? prob_pass_wont_exit | |
921 | : prob_pass_thru; | |
922 | p = prob_pass_main; | |
923 | scale_main = REG_BR_PROB_BASE; | |
924 | for (i = 0; i < ndupl; i++) | |
925 | { | |
926 | scale_main += p; | |
927 | p = RDIV (p * scale_step[i], REG_BR_PROB_BASE); | |
928 | } | |
929 | scale_main = RDIV (REG_BR_PROB_BASE * REG_BR_PROB_BASE, scale_main); | |
930 | scale_act = RDIV (scale_main * prob_pass_main, REG_BR_PROB_BASE); | |
931 | } | |
932 | else | |
933 | { | |
934 | scale_main = REG_BR_PROB_BASE; | |
935 | for (i = 0; i < ndupl; i++) | |
936 | scale_main = RDIV (scale_main * scale_step[i], REG_BR_PROB_BASE); | |
937 | scale_act = REG_BR_PROB_BASE - prob_pass_thru; | |
938 | } | |
939 | for (i = 0; i < ndupl; i++) | |
341c100f NS |
940 | gcc_assert (scale_step[i] >= 0 && scale_step[i] <= REG_BR_PROB_BASE); |
941 | gcc_assert (scale_main >= 0 && scale_main <= REG_BR_PROB_BASE | |
942 | && scale_act >= 0 && scale_act <= REG_BR_PROB_BASE); | |
617b465c ZD |
943 | } |
944 | ||
945 | /* Loop the new bbs will belong to. */ | |
8d28e87d | 946 | target = e->src->loop_father; |
617b465c ZD |
947 | |
948 | /* Original loops. */ | |
949 | n_orig_loops = 0; | |
950 | for (aloop = loop->inner; aloop; aloop = aloop->next) | |
951 | n_orig_loops++; | |
5ed6ace5 | 952 | orig_loops = XCNEWVEC (struct loop *, n_orig_loops); |
617b465c ZD |
953 | for (aloop = loop->inner, i = 0; aloop; aloop = aloop->next, i++) |
954 | orig_loops[i] = aloop; | |
955 | ||
956 | loop->copy = target; | |
d329e058 | 957 | |
5ed6ace5 | 958 | first_active = XNEWVEC (basic_block, n); |
617b465c ZD |
959 | if (is_latch) |
960 | { | |
961 | memcpy (first_active, bbs, n * sizeof (basic_block)); | |
962 | first_active_latch = latch; | |
963 | } | |
964 | ||
82b85a85 ZD |
965 | /* Update the information about single exits. */ |
966 | if (loops->state & LOOPS_HAVE_MARKED_SINGLE_EXITS) | |
967 | update_single_exits_after_duplication (bbs, n, target); | |
968 | ||
8d28e87d ZD |
969 | /* Record exit edge in original loop body. */ |
970 | if (orig && TEST_BIT (wont_exit, 0)) | |
971 | to_remove[(*n_to_remove)++] = orig; | |
972 | ||
973 | spec_edges[SE_ORIG] = orig; | |
974 | spec_edges[SE_LATCH] = latch_edge; | |
d329e058 | 975 | |
b9a66240 | 976 | place_after = e->src; |
617b465c ZD |
977 | for (j = 0; j < ndupl; j++) |
978 | { | |
979 | /* Copy loops. */ | |
980 | copy_loops_to (loops, orig_loops, n_orig_loops, target); | |
981 | ||
982 | /* Copy bbs. */ | |
b9a66240 ZD |
983 | copy_bbs (bbs, n, new_bbs, spec_edges, 2, new_spec_edges, loop, |
984 | place_after); | |
985 | place_after = new_spec_edges[SE_LATCH]->src; | |
8d28e87d | 986 | |
c0493b13 ZD |
987 | if (loops->state & LOOPS_HAVE_MARKED_SINGLE_EXITS) |
988 | { | |
989 | for (i = 0; i < n; i++) | |
990 | bbs[i]->flags |= BB_DUPLICATED; | |
991 | update_single_exit_for_duplicated_loops (orig_loops, n_orig_loops); | |
992 | for (i = 0; i < n; i++) | |
993 | bbs[i]->flags &= ~BB_DUPLICATED; | |
994 | } | |
995 | ||
7f7b1718 JH |
996 | if (flags & DLTHE_RECORD_COPY_NUMBER) |
997 | for (i = 0; i < n; i++) | |
998 | { | |
999 | gcc_assert (!new_bbs[i]->aux); | |
1000 | new_bbs[i]->aux = (void *)(size_t)(j + 1); | |
1001 | } | |
113d659a | 1002 | |
84d45ad1 ZD |
1003 | /* Note whether the blocks and edges belong to an irreducible loop. */ |
1004 | if (add_irreducible_flag) | |
1005 | { | |
1006 | for (i = 0; i < n; i++) | |
6580ee77 | 1007 | new_bbs[i]->flags |= BB_DUPLICATED; |
84d45ad1 ZD |
1008 | for (i = 0; i < n; i++) |
1009 | { | |
628f6a4e | 1010 | edge_iterator ei; |
84d45ad1 ZD |
1011 | new_bb = new_bbs[i]; |
1012 | if (new_bb->loop_father == target) | |
1013 | new_bb->flags |= BB_IRREDUCIBLE_LOOP; | |
1014 | ||
628f6a4e | 1015 | FOR_EACH_EDGE (ae, ei, new_bb->succs) |
6580ee77 | 1016 | if ((ae->dest->flags & BB_DUPLICATED) |
84d45ad1 ZD |
1017 | && (ae->src->loop_father == target |
1018 | || ae->dest->loop_father == target)) | |
1019 | ae->flags |= EDGE_IRREDUCIBLE_LOOP; | |
1020 | } | |
1021 | for (i = 0; i < n; i++) | |
6580ee77 | 1022 | new_bbs[i]->flags &= ~BB_DUPLICATED; |
84d45ad1 ZD |
1023 | } |
1024 | ||
8d28e87d | 1025 | /* Redirect the special edges. */ |
617b465c | 1026 | if (is_latch) |
8d28e87d ZD |
1027 | { |
1028 | redirect_edge_and_branch_force (latch_edge, new_bbs[0]); | |
1029 | redirect_edge_and_branch_force (new_spec_edges[SE_LATCH], | |
1030 | loop->header); | |
d47cc544 | 1031 | set_immediate_dominator (CDI_DOMINATORS, new_bbs[0], latch); |
b9a66240 | 1032 | latch = loop->latch = new_bbs[n - 1]; |
8d28e87d ZD |
1033 | e = latch_edge = new_spec_edges[SE_LATCH]; |
1034 | } | |
1035 | else | |
1036 | { | |
1037 | redirect_edge_and_branch_force (new_spec_edges[SE_LATCH], | |
1038 | loop->header); | |
1039 | redirect_edge_and_branch_force (e, new_bbs[0]); | |
d47cc544 | 1040 | set_immediate_dominator (CDI_DOMINATORS, new_bbs[0], e->src); |
8d28e87d ZD |
1041 | e = new_spec_edges[SE_LATCH]; |
1042 | } | |
617b465c | 1043 | |
8d28e87d ZD |
1044 | /* Record exit edge in this copy. */ |
1045 | if (orig && TEST_BIT (wont_exit, j + 1)) | |
1046 | to_remove[(*n_to_remove)++] = new_spec_edges[SE_ORIG]; | |
d329e058 | 1047 | |
8d28e87d ZD |
1048 | /* Record the first copy in the control flow order if it is not |
1049 | the original loop (i.e. in case of peeling). */ | |
617b465c ZD |
1050 | if (!first_active_latch) |
1051 | { | |
1052 | memcpy (first_active, new_bbs, n * sizeof (basic_block)); | |
b9a66240 | 1053 | first_active_latch = new_bbs[n - 1]; |
617b465c | 1054 | } |
d329e058 | 1055 | |
8d28e87d ZD |
1056 | /* Set counts and frequencies. */ |
1057 | if (flags & DLTHE_FLAG_UPDATE_FREQ) | |
617b465c | 1058 | { |
33156717 | 1059 | scale_bbs_frequencies_int (new_bbs, n, scale_act, REG_BR_PROB_BASE); |
8d28e87d | 1060 | scale_act = RDIV (scale_act * scale_step[j], REG_BR_PROB_BASE); |
617b465c ZD |
1061 | } |
1062 | } | |
8d28e87d ZD |
1063 | free (new_bbs); |
1064 | free (orig_loops); | |
c22cacf3 | 1065 | |
8d28e87d ZD |
1066 | /* Update the original loop. */ |
1067 | if (!is_latch) | |
d47cc544 | 1068 | set_immediate_dominator (CDI_DOMINATORS, e->dest, e->src); |
617b465c ZD |
1069 | if (flags & DLTHE_FLAG_UPDATE_FREQ) |
1070 | { | |
33156717 | 1071 | scale_bbs_frequencies_int (bbs, n, scale_main, REG_BR_PROB_BASE); |
617b465c ZD |
1072 | free (scale_step); |
1073 | } | |
617b465c | 1074 | |
8d28e87d | 1075 | /* Update dominators of outer blocks if affected. */ |
617b465c ZD |
1076 | for (i = 0; i < n; i++) |
1077 | { | |
1078 | basic_block dominated, dom_bb, *dom_bbs; | |
1079 | int n_dom_bbs,j; | |
1080 | ||
1081 | bb = bbs[i]; | |
6580ee77 | 1082 | bb->aux = 0; |
113d659a | 1083 | |
d47cc544 | 1084 | n_dom_bbs = get_dominated_by (CDI_DOMINATORS, bb, &dom_bbs); |
617b465c ZD |
1085 | for (j = 0; j < n_dom_bbs; j++) |
1086 | { | |
1087 | dominated = dom_bbs[j]; | |
1088 | if (flow_bb_inside_loop_p (loop, dominated)) | |
1089 | continue; | |
1090 | dom_bb = nearest_common_dominator ( | |
d47cc544 | 1091 | CDI_DOMINATORS, first_active[i], first_active_latch); |
c22cacf3 | 1092 | set_immediate_dominator (CDI_DOMINATORS, dominated, dom_bb); |
617b465c ZD |
1093 | } |
1094 | free (dom_bbs); | |
1095 | } | |
1096 | free (first_active); | |
1097 | ||
1098 | free (bbs); | |
1099 | ||
1100 | return true; | |
1101 | } | |
1102 | ||
f470c378 ZD |
1103 | /* A callback for make_forwarder block, to redirect all edges except for |
1104 | MFB_KJ_EDGE to the entry part. E is the edge for that we should decide | |
1105 | whether to redirect it. */ | |
1106 | ||
1107 | static edge mfb_kj_edge; | |
1108 | static bool | |
1109 | mfb_keep_just (edge e) | |
1110 | { | |
1111 | return e != mfb_kj_edge; | |
1112 | } | |
1113 | ||
1114 | /* A callback for make_forwarder block, to update data structures for a basic | |
1115 | block JUMP created by redirecting an edge (only the latch edge is being | |
1116 | redirected). */ | |
1117 | ||
1118 | static void | |
1119 | mfb_update_loops (basic_block jump) | |
1120 | { | |
c5cbcccf | 1121 | struct loop *loop = single_succ (jump)->loop_father; |
f470c378 ZD |
1122 | |
1123 | if (dom_computed[CDI_DOMINATORS]) | |
c5cbcccf | 1124 | set_immediate_dominator (CDI_DOMINATORS, jump, single_pred (jump)); |
f470c378 ZD |
1125 | add_bb_to_loop (jump, loop); |
1126 | loop->latch = jump; | |
1127 | } | |
1128 | ||
3d436d2a ZD |
1129 | /* Creates a pre-header for a LOOP. Returns newly created block. Unless |
1130 | CP_SIMPLE_PREHEADERS is set in FLAGS, we only force LOOP to have single | |
1131 | entry; otherwise we also force preheader block to have only one successor. | |
f470c378 ZD |
1132 | The function also updates dominators. */ |
1133 | ||
3d436d2a | 1134 | static basic_block |
d47cc544 | 1135 | create_preheader (struct loop *loop, int flags) |
3d436d2a ZD |
1136 | { |
1137 | edge e, fallthru; | |
1138 | basic_block dummy; | |
3d436d2a ZD |
1139 | struct loop *cloop, *ploop; |
1140 | int nentry = 0; | |
f470c378 | 1141 | bool irred = false; |
c15bc84b EB |
1142 | bool latch_edge_was_fallthru; |
1143 | edge one_succ_pred = 0; | |
628f6a4e | 1144 | edge_iterator ei; |
3d436d2a ZD |
1145 | |
1146 | cloop = loop->outer; | |
1147 | ||
628f6a4e | 1148 | FOR_EACH_EDGE (e, ei, loop->header->preds) |
3d436d2a ZD |
1149 | { |
1150 | if (e->src == loop->latch) | |
1151 | continue; | |
f470c378 | 1152 | irred |= (e->flags & EDGE_IRREDUCIBLE_LOOP) != 0; |
3d436d2a | 1153 | nentry++; |
c5cbcccf | 1154 | if (single_succ_p (e->src)) |
c15bc84b | 1155 | one_succ_pred = e; |
3d436d2a | 1156 | } |
341c100f | 1157 | gcc_assert (nentry); |
3d436d2a ZD |
1158 | if (nentry == 1) |
1159 | { | |
b057216d KH |
1160 | /* Get an edge that is different from the one from loop->latch |
1161 | to loop->header. */ | |
1162 | e = EDGE_PRED (loop->header, | |
1163 | EDGE_PRED (loop->header, 0)->src == loop->latch); | |
628f6a4e | 1164 | |
c5cbcccf | 1165 | if (!(flags & CP_SIMPLE_PREHEADERS) || single_succ_p (e->src)) |
3d436d2a ZD |
1166 | return NULL; |
1167 | } | |
1168 | ||
f470c378 | 1169 | mfb_kj_edge = loop_latch_edge (loop); |
c15bc84b | 1170 | latch_edge_was_fallthru = (mfb_kj_edge->flags & EDGE_FALLTHRU) != 0; |
f470c378 ZD |
1171 | fallthru = make_forwarder_block (loop->header, mfb_keep_just, |
1172 | mfb_update_loops); | |
3d436d2a ZD |
1173 | dummy = fallthru->src; |
1174 | loop->header = fallthru->dest; | |
1175 | ||
1176 | /* The header could be a latch of some superloop(s); due to design of | |
1177 | split_block, it would now move to fallthru->dest. */ | |
1178 | for (ploop = loop; ploop; ploop = ploop->outer) | |
1179 | if (ploop->latch == dummy) | |
1180 | ploop->latch = fallthru->dest; | |
1181 | ||
c15bc84b EB |
1182 | /* Try to be clever in placing the newly created preheader. The idea is to |
1183 | avoid breaking any "fallthruness" relationship between blocks. | |
1184 | ||
1185 | The preheader was created just before the header and all incoming edges | |
1186 | to the header were redirected to the preheader, except the latch edge. | |
1187 | So the only problematic case is when this latch edge was a fallthru | |
1188 | edge: it is not anymore after the preheader creation so we have broken | |
1189 | the fallthruness. We're therefore going to look for a better place. */ | |
1190 | if (latch_edge_was_fallthru) | |
1191 | { | |
1192 | if (one_succ_pred) | |
1193 | e = one_succ_pred; | |
1194 | else | |
1195 | e = EDGE_PRED (dummy, 0); | |
1196 | ||
1197 | move_block_after (dummy, e->src); | |
1198 | } | |
f470c378 ZD |
1199 | |
1200 | loop->header->loop_father = loop; | |
1201 | add_bb_to_loop (dummy, cloop); | |
3d436d2a | 1202 | |
f470c378 | 1203 | if (irred) |
3d436d2a | 1204 | { |
f470c378 | 1205 | dummy->flags |= BB_IRREDUCIBLE_LOOP; |
c5cbcccf | 1206 | single_succ_edge (dummy)->flags |= EDGE_IRREDUCIBLE_LOOP; |
3d436d2a ZD |
1207 | } |
1208 | ||
c263766c RH |
1209 | if (dump_file) |
1210 | fprintf (dump_file, "Created preheader block for loop %i\n", | |
3d436d2a ZD |
1211 | loop->num); |
1212 | ||
1213 | return dummy; | |
1214 | } | |
1215 | ||
617b465c ZD |
1216 | /* Create preheaders for each loop from loop tree stored in LOOPS; for meaning |
1217 | of FLAGS see create_preheader. */ | |
3d436d2a | 1218 | void |
d329e058 | 1219 | create_preheaders (struct loops *loops, int flags) |
3d436d2a ZD |
1220 | { |
1221 | unsigned i; | |
1222 | for (i = 1; i < loops->num; i++) | |
d47cc544 | 1223 | create_preheader (loops->parray[i], flags); |
3d436d2a ZD |
1224 | loops->state |= LOOPS_HAVE_PREHEADERS; |
1225 | } | |
1226 | ||
617b465c ZD |
1227 | /* Forces all loop latches of loops from loop tree LOOPS to have only single |
1228 | successor. */ | |
3d436d2a | 1229 | void |
d329e058 | 1230 | force_single_succ_latches (struct loops *loops) |
3d436d2a ZD |
1231 | { |
1232 | unsigned i; | |
1233 | struct loop *loop; | |
1234 | edge e; | |
1235 | ||
1236 | for (i = 1; i < loops->num; i++) | |
1237 | { | |
1238 | loop = loops->parray[i]; | |
c5cbcccf | 1239 | if (loop->latch != loop->header && single_succ_p (loop->latch)) |
3d436d2a | 1240 | continue; |
d329e058 | 1241 | |
9ff3d2de | 1242 | e = find_edge (loop->latch, loop->header); |
bc810602 | 1243 | |
d47cc544 | 1244 | loop_split_edge_with (e, NULL_RTX); |
3d436d2a ZD |
1245 | } |
1246 | loops->state |= LOOPS_HAVE_SIMPLE_LATCHES; | |
1247 | } | |
1248 | ||
617b465c ZD |
1249 | /* A quite stupid function to put INSNS on edge E. They are supposed to form |
1250 | just one basic block. Jumps in INSNS are not handled, so cfg do not have to | |
1251 | be ok after this function. The created block is placed on correct place | |
1252 | in LOOPS structure and its dominator is set. */ | |
3d436d2a | 1253 | basic_block |
d47cc544 | 1254 | loop_split_edge_with (edge e, rtx insns) |
3d436d2a ZD |
1255 | { |
1256 | basic_block src, dest, new_bb; | |
1257 | struct loop *loop_c; | |
d329e058 | 1258 | |
3d436d2a ZD |
1259 | src = e->src; |
1260 | dest = e->dest; | |
1261 | ||
1262 | loop_c = find_common_loop (src->loop_father, dest->loop_father); | |
1263 | ||
1264 | /* Create basic block for it. */ | |
1265 | ||
bc35512f | 1266 | new_bb = split_edge (e); |
3d436d2a | 1267 | add_bb_to_loop (new_bb, loop_c); |
a746fd8c | 1268 | new_bb->flags |= (insns ? BB_SUPERBLOCK : 0); |
3d436d2a | 1269 | |
3d436d2a | 1270 | if (insns) |
a813c111 | 1271 | emit_insn_after (insns, BB_END (new_bb)); |
3d436d2a | 1272 | |
3d436d2a ZD |
1273 | if (dest->loop_father->latch == src) |
1274 | dest->loop_father->latch = new_bb; | |
d329e058 | 1275 | |
3d436d2a ZD |
1276 | return new_bb; |
1277 | } | |
6de9cd9a | 1278 | |
1cb7dfc3 MH |
1279 | /* This function is called from loop_version. It splits the entry edge |
1280 | of the loop we want to version, adds the versioning condition, and | |
1281 | adjust the edges to the two versions of the loop appropriately. | |
1282 | e is an incoming edge. Returns the basic block containing the | |
1283 | condition. | |
1284 | ||
1285 | --- edge e ---- > [second_head] | |
1286 | ||
1287 | Split it and insert new conditional expression and adjust edges. | |
1288 | ||
1289 | --- edge e ---> [cond expr] ---> [first_head] | |
c22cacf3 MS |
1290 | | |
1291 | +---------> [second_head] | |
1cb7dfc3 MH |
1292 | */ |
1293 | ||
1294 | static basic_block | |
1295 | lv_adjust_loop_entry_edge (basic_block first_head, | |
1296 | basic_block second_head, | |
1297 | edge e, | |
ee39680c | 1298 | void *cond_expr) |
1cb7dfc3 MH |
1299 | { |
1300 | basic_block new_head = NULL; | |
1301 | edge e1; | |
1302 | ||
1303 | gcc_assert (e->dest == second_head); | |
1304 | ||
1305 | /* Split edge 'e'. This will create a new basic block, where we can | |
1306 | insert conditional expr. */ | |
1307 | new_head = split_edge (e); | |
1308 | ||
1309 | ||
1310 | lv_add_condition_to_bb (first_head, second_head, new_head, | |
1311 | cond_expr); | |
1312 | ||
766613a4 | 1313 | /* Don't set EDGE_TRUE_VALUE in RTL mode, as it's invalid there. */ |
52bca999 SB |
1314 | e1 = make_edge (new_head, first_head, |
1315 | current_ir_type () == IR_GIMPLE ? EDGE_TRUE_VALUE : 0); | |
1cb7dfc3 MH |
1316 | set_immediate_dominator (CDI_DOMINATORS, first_head, new_head); |
1317 | set_immediate_dominator (CDI_DOMINATORS, second_head, new_head); | |
1318 | ||
1319 | /* Adjust loop header phi nodes. */ | |
1320 | lv_adjust_loop_header_phi (first_head, second_head, new_head, e1); | |
1321 | ||
1322 | return new_head; | |
1323 | } | |
1324 | ||
1325 | /* Main entry point for Loop Versioning transformation. | |
c22cacf3 | 1326 | |
b9a66240 ZD |
1327 | This transformation given a condition and a loop, creates |
1328 | -if (condition) { loop_copy1 } else { loop_copy2 }, | |
1329 | where loop_copy1 is the loop transformed in one way, and loop_copy2 | |
1330 | is the loop transformed in another way (or unchanged). 'condition' | |
1331 | may be a run time test for things that were not resolved by static | |
1332 | analysis (overlapping ranges (anti-aliasing), alignment, etc.). | |
1333 | ||
1334 | If PLACE_AFTER is true, we place the new loop after LOOP in the | |
1335 | instruction stream, otherwise it is placed before LOOP. */ | |
1cb7dfc3 MH |
1336 | |
1337 | struct loop * | |
c22cacf3 | 1338 | loop_version (struct loops *loops, struct loop * loop, |
b9a66240 ZD |
1339 | void *cond_expr, basic_block *condition_bb, |
1340 | bool place_after) | |
1cb7dfc3 MH |
1341 | { |
1342 | basic_block first_head, second_head; | |
1343 | edge entry, latch_edge, exit, true_edge, false_edge; | |
1344 | int irred_flag; | |
1345 | struct loop *nloop; | |
b9a66240 | 1346 | basic_block cond_bb; |
1cb7dfc3 MH |
1347 | |
1348 | /* CHECKME: Loop versioning does not handle nested loop at this point. */ | |
1349 | if (loop->inner) | |
1350 | return NULL; | |
1351 | ||
1352 | /* Record entry and latch edges for the loop */ | |
1353 | entry = loop_preheader_edge (loop); | |
1354 | irred_flag = entry->flags & EDGE_IRREDUCIBLE_LOOP; | |
1355 | entry->flags &= ~EDGE_IRREDUCIBLE_LOOP; | |
c22cacf3 | 1356 | |
1cb7dfc3 MH |
1357 | /* Note down head of loop as first_head. */ |
1358 | first_head = entry->dest; | |
1359 | ||
1360 | /* Duplicate loop. */ | |
1361 | if (!cfg_hook_duplicate_loop_to_header_edge (loop, entry, loops, 1, | |
c22cacf3 | 1362 | NULL, NULL, NULL, NULL, 0)) |
1cb7dfc3 MH |
1363 | return NULL; |
1364 | ||
1365 | /* After duplication entry edge now points to new loop head block. | |
1366 | Note down new head as second_head. */ | |
1367 | second_head = entry->dest; | |
1368 | ||
1369 | /* Split loop entry edge and insert new block with cond expr. */ | |
b9a66240 ZD |
1370 | cond_bb = lv_adjust_loop_entry_edge (first_head, second_head, |
1371 | entry, cond_expr); | |
1372 | if (condition_bb) | |
1373 | *condition_bb = cond_bb; | |
1374 | ||
1375 | if (!cond_bb) | |
1cb7dfc3 MH |
1376 | { |
1377 | entry->flags |= irred_flag; | |
1378 | return NULL; | |
1379 | } | |
1380 | ||
6580ee77 | 1381 | latch_edge = single_succ_edge (get_bb_copy (loop->latch)); |
c22cacf3 | 1382 | |
b9a66240 | 1383 | extract_cond_bb_edges (cond_bb, &true_edge, &false_edge); |
1cb7dfc3 MH |
1384 | nloop = loopify (loops, |
1385 | latch_edge, | |
6580ee77 | 1386 | single_pred_edge (get_bb_copy (loop->header)), |
b9a66240 | 1387 | cond_bb, true_edge, false_edge, |
1cb7dfc3 MH |
1388 | false /* Do not redirect all edges. */); |
1389 | ||
1390 | exit = loop->single_exit; | |
1391 | if (exit) | |
6580ee77 | 1392 | nloop->single_exit = find_edge (get_bb_copy (exit->src), exit->dest); |
1cb7dfc3 | 1393 | |
c22cacf3 | 1394 | /* loopify redirected latch_edge. Update its PENDING_STMTS. */ |
1cb7dfc3 MH |
1395 | lv_flush_pending_stmts (latch_edge); |
1396 | ||
c22cacf3 | 1397 | /* loopify redirected condition_bb's succ edge. Update its PENDING_STMTS. */ |
b9a66240 | 1398 | extract_cond_bb_edges (cond_bb, &true_edge, &false_edge); |
1cb7dfc3 MH |
1399 | lv_flush_pending_stmts (false_edge); |
1400 | /* Adjust irreducible flag. */ | |
1401 | if (irred_flag) | |
1402 | { | |
b9a66240 | 1403 | cond_bb->flags |= BB_IRREDUCIBLE_LOOP; |
1cb7dfc3 MH |
1404 | loop_preheader_edge (loop)->flags |= EDGE_IRREDUCIBLE_LOOP; |
1405 | loop_preheader_edge (nloop)->flags |= EDGE_IRREDUCIBLE_LOOP; | |
b9a66240 ZD |
1406 | single_pred_edge (cond_bb)->flags |= EDGE_IRREDUCIBLE_LOOP; |
1407 | } | |
1408 | ||
1409 | if (place_after) | |
1410 | { | |
1411 | basic_block *bbs = get_loop_body_in_dom_order (nloop), after; | |
1412 | unsigned i; | |
1413 | ||
1414 | after = loop->latch; | |
1415 | ||
1416 | for (i = 0; i < nloop->num_nodes; i++) | |
1417 | { | |
1418 | move_block_after (bbs[i], after); | |
1419 | after = bbs[i]; | |
1420 | } | |
1421 | free (bbs); | |
1cb7dfc3 MH |
1422 | } |
1423 | ||
c22cacf3 MS |
1424 | /* At this point condition_bb is loop predheader with two successors, |
1425 | first_head and second_head. Make sure that loop predheader has only | |
1cb7dfc3 MH |
1426 | one successor. */ |
1427 | loop_split_edge_with (loop_preheader_edge (loop), NULL); | |
1428 | loop_split_edge_with (loop_preheader_edge (nloop), NULL); | |
1429 | ||
1430 | return nloop; | |
1431 | } | |
1432 | ||
2b271002 ZD |
1433 | /* The structure of LOOPS might have changed. Some loops might get removed |
1434 | (and their headers and latches were set to NULL), loop exists might get | |
1435 | removed (thus the loop nesting may be wrong), and some blocks and edges | |
1436 | were changed (so the information about bb --> loop mapping does not have | |
1437 | to be correct). But still for the remaining loops the header dominates | |
1438 | the latch, and loops did not get new subloobs (new loops might possibly | |
1439 | get created, but we are not interested in them). Fix up the mess. | |
c22cacf3 | 1440 | |
2b271002 ZD |
1441 | If CHANGED_BBS is not NULL, basic blocks whose loop has changed are |
1442 | marked in it. */ | |
1443 | ||
1444 | void | |
1445 | fix_loop_structure (struct loops *loops, bitmap changed_bbs) | |
1446 | { | |
1447 | basic_block bb; | |
1448 | struct loop *loop, *ploop; | |
1449 | unsigned i; | |
1450 | ||
1451 | /* Remove the old bb -> loop mapping. */ | |
1452 | FOR_EACH_BB (bb) | |
1453 | { | |
1454 | bb->aux = (void *) (size_t) bb->loop_father->depth; | |
1455 | bb->loop_father = loops->tree_root; | |
1456 | } | |
1457 | ||
1458 | /* Remove the dead loops from structures. */ | |
c22cacf3 | 1459 | loops->tree_root->num_nodes = n_basic_blocks; |
2b271002 ZD |
1460 | for (i = 1; i < loops->num; i++) |
1461 | { | |
1462 | loop = loops->parray[i]; | |
1463 | if (!loop) | |
1464 | continue; | |
1465 | ||
1466 | loop->num_nodes = 0; | |
1467 | if (loop->header) | |
1468 | continue; | |
1469 | ||
1470 | while (loop->inner) | |
1471 | { | |
1472 | ploop = loop->inner; | |
1473 | flow_loop_tree_node_remove (ploop); | |
1474 | flow_loop_tree_node_add (loop->outer, ploop); | |
1475 | } | |
1476 | ||
1477 | /* Remove the loop and free its data. */ | |
1478 | flow_loop_tree_node_remove (loop); | |
1479 | loops->parray[loop->num] = NULL; | |
1480 | flow_loop_free (loop); | |
1481 | } | |
1482 | ||
1483 | /* Rescan the bodies of loops, starting from the outermost. */ | |
1484 | loop = loops->tree_root; | |
1485 | while (1) | |
1486 | { | |
1487 | if (loop->inner) | |
1488 | loop = loop->inner; | |
1489 | else | |
1490 | { | |
1491 | while (!loop->next | |
1492 | && loop != loops->tree_root) | |
1493 | loop = loop->outer; | |
1494 | if (loop == loops->tree_root) | |
1495 | break; | |
1496 | ||
1497 | loop = loop->next; | |
1498 | } | |
1499 | ||
1500 | loop->num_nodes = flow_loop_nodes_find (loop->header, loop); | |
1501 | } | |
1502 | ||
1503 | /* Now fix the loop nesting. */ | |
1504 | for (i = 1; i < loops->num; i++) | |
1505 | { | |
1506 | loop = loops->parray[i]; | |
1507 | if (!loop) | |
1508 | continue; | |
1509 | ||
1510 | bb = loop_preheader_edge (loop)->src; | |
1511 | if (bb->loop_father != loop->outer) | |
1512 | { | |
1513 | flow_loop_tree_node_remove (loop); | |
1514 | flow_loop_tree_node_add (bb->loop_father, loop); | |
1515 | } | |
1516 | } | |
1517 | ||
1518 | /* Mark the blocks whose loop has changed. */ | |
1519 | FOR_EACH_BB (bb) | |
1520 | { | |
1521 | if (changed_bbs | |
1522 | && (void *) (size_t) bb->loop_father->depth != bb->aux) | |
1523 | bitmap_set_bit (changed_bbs, bb->index); | |
1524 | ||
1525 | bb->aux = NULL; | |
1526 | } | |
1527 | ||
d78f3f78 ZD |
1528 | if (loops->state & LOOPS_HAVE_MARKED_SINGLE_EXITS) |
1529 | mark_single_exit_loops (loops); | |
1530 | if (loops->state & LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS) | |
1531 | mark_irreducible_loops (loops); | |
2b271002 | 1532 | } |