]>
Commit | Line | Data |
---|---|---|
3d436d2a | 1 | /* Loop manipulation code for GNU compiler. |
fa10beec RW |
2 | Copyright (C) 2002, 2003, 2004, 2005, 2007, 2008 Free Software |
3 | Foundation, Inc. | |
3d436d2a ZD |
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 | |
9dcd6f09 | 9 | Software Foundation; either version 3, or (at your option) any later |
3d436d2a ZD |
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 | |
9dcd6f09 NC |
18 | along with GCC; see the file COPYING3. If not see |
19 | <http://www.gnu.org/licenses/>. */ | |
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 | ||
d73be268 ZD |
34 | static void duplicate_subloops (struct loop *, struct loop *); |
35 | static void copy_loops_to (struct loop **, int, | |
d329e058 AJ |
36 | struct loop *); |
37 | static void loop_redirect_edge (edge, basic_block); | |
d47cc544 | 38 | static void remove_bbs (basic_block *, int); |
ed7a4b4b | 39 | static bool rpe_enum_p (const_basic_block, const void *); |
d47cc544 | 40 | static int find_path (edge, basic_block **); |
d73be268 ZD |
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 *); | |
d73be268 | 44 | static void unloop (struct loop *, bool *); |
3d436d2a | 45 | |
bade3a00 JH |
46 | #define RDIV(X,Y) (((X) + (Y) / 2) / (Y)) |
47 | ||
d47cc544 | 48 | /* Checks whether basic block BB is dominated by DATA. */ |
617b465c | 49 | static bool |
ed7a4b4b | 50 | rpe_enum_p (const_basic_block bb, const void *data) |
617b465c | 51 | { |
ed7a4b4b | 52 | return dominated_by_p (CDI_DOMINATORS, bb, (const_basic_block) data); |
617b465c ZD |
53 | } |
54 | ||
598ec7bd ZD |
55 | /* Remove basic blocks BBS. NBBS is the number of the basic blocks. */ |
56 | ||
617b465c | 57 | static void |
d47cc544 | 58 | remove_bbs (basic_block *bbs, int nbbs) |
617b465c ZD |
59 | { |
60 | int i; | |
61 | ||
62 | for (i = 0; i < nbbs; i++) | |
598ec7bd | 63 | delete_basic_block (bbs[i]); |
617b465c ZD |
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. | |
35b07080 ZD |
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. */ | |
617b465c | 72 | static int |
d47cc544 | 73 | find_path (edge e, basic_block **bbs) |
617b465c | 74 | { |
628f6a4e | 75 | gcc_assert (EDGE_COUNT (e->dest->preds) <= 1); |
617b465c ZD |
76 | |
77 | /* Find bbs in the path. */ | |
5ed6ace5 | 78 | *bbs = XCNEWVEC (basic_block, n_basic_blocks); |
617b465c | 79 | return dfs_enumerate_from (e->dest, 0, rpe_enum_p, *bbs, |
d47cc544 | 80 | n_basic_blocks, e->dest); |
617b465c ZD |
81 | } |
82 | ||
d73be268 | 83 | /* Fix placement of basic block BB inside loop hierarchy -- |
617b465c ZD |
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 | |
d73be268 | 91 | fix_bb_placement (basic_block bb) |
617b465c ZD |
92 | { |
93 | edge e; | |
628f6a4e | 94 | edge_iterator ei; |
d73be268 | 95 | struct loop *loop = current_loops->tree_root, *act; |
617b465c | 96 | |
628f6a4e | 97 | FOR_EACH_EDGE (e, ei, bb->succs) |
617b465c ZD |
98 | { |
99 | if (e->dest == EXIT_BLOCK_PTR) | |
100 | continue; | |
101 | ||
102 | act = e->dest->loop_father; | |
103 | if (act->header == e->dest) | |
9ba025a2 | 104 | act = loop_outer (act); |
617b465c ZD |
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 | ||
b4c1c7e3 ZD |
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 | ||
9ba025a2 | 140 | if (father != loop_outer (loop)) |
b4c1c7e3 | 141 | { |
9ba025a2 | 142 | for (act = loop_outer (loop); act != father; act = loop_outer (act)) |
b4c1c7e3 ZD |
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 | ||
617b465c ZD |
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 | |
4d6922ee | 165 | to this change; the condition for them is similar, except that instead of |
dc14f191 ZD |
166 | successors we consider edges coming out of the loops. |
167 | ||
168 | If the changes may invalidate the information about irreducible regions, | |
169 | IRRED_INVALIDATED is set to true. */ | |
170 | ||
617b465c | 171 | static void |
d73be268 | 172 | fix_bb_placements (basic_block from, |
dc14f191 | 173 | bool *irred_invalidated) |
617b465c ZD |
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; | |
d73be268 | 188 | if (base_loop == current_loops->tree_root) |
617b465c ZD |
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 | ||
5ed6ace5 | 197 | queue = XNEWVEC (basic_block, base_loop->num_nodes + 1); |
617b465c ZD |
198 | qtop = queue + base_loop->num_nodes + 1; |
199 | qbeg = queue; | |
200 | qend = queue + 1; | |
201 | *qbeg = from; | |
202 | ||
203 | while (qbeg != qend) | |
204 | { | |
628f6a4e | 205 | edge_iterator ei; |
617b465c ZD |
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. */ | |
d73be268 | 221 | if (!fix_bb_placement (from)) |
617b465c ZD |
222 | continue; |
223 | } | |
224 | ||
dc14f191 ZD |
225 | FOR_EACH_EDGE (e, ei, from->succs) |
226 | { | |
227 | if (e->flags & EDGE_IRREDUCIBLE_LOOP) | |
228 | *irred_invalidated = true; | |
229 | } | |
230 | ||
617b465c | 231 | /* Something has changed, insert predecessors into queue. */ |
628f6a4e | 232 | FOR_EACH_EDGE (e, ei, from->preds) |
617b465c ZD |
233 | { |
234 | basic_block pred = e->src; | |
235 | struct loop *nca; | |
236 | ||
dc14f191 ZD |
237 | if (e->flags & EDGE_IRREDUCIBLE_LOOP) |
238 | *irred_invalidated = true; | |
239 | ||
617b465c ZD |
240 | if (TEST_BIT (in_queue, pred->index)) |
241 | continue; | |
242 | ||
d329e058 | 243 | /* If it is subloop, then it either was not moved, or |
617b465c ZD |
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 | |
d73be268 ZD |
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). */ | |
617b465c | 275 | bool |
d73be268 | 276 | remove_path (edge e) |
617b465c ZD |
277 | { |
278 | edge ae; | |
66f97d31 ZD |
279 | basic_block *rem_bbs, *bord_bbs, from, bb; |
280 | VEC (basic_block, heap) *dom_bbs; | |
281 | int i, nrem, n_bord_bbs, nreml; | |
617b465c | 282 | sbitmap seen; |
14fa2cc0 | 283 | bool irred_invalidated = false; |
598ec7bd | 284 | struct loop **deleted_loop; |
617b465c | 285 | |
14fa2cc0 | 286 | if (!can_remove_branch_p (e)) |
35b07080 ZD |
287 | return false; |
288 | ||
dc14f191 ZD |
289 | /* Keep track of whether we need to update information about irreducible |
290 | regions. This is the case if the removed area is a part of the | |
291 | irreducible region, or if the set of basic blocks that belong to a loop | |
292 | that is inside an irreducible region is changed, or if such a loop is | |
293 | removed. */ | |
294 | if (e->flags & EDGE_IRREDUCIBLE_LOOP) | |
295 | irred_invalidated = true; | |
296 | ||
35b07080 ZD |
297 | /* We need to check whether basic blocks are dominated by the edge |
298 | e, but we only have basic block dominators. This is easy to | |
299 | fix -- when e->dest has exactly one predecessor, this corresponds | |
300 | to blocks dominated by e->dest, if not, split the edge. */ | |
c5cbcccf | 301 | if (!single_pred_p (e->dest)) |
598ec7bd | 302 | e = single_pred_edge (split_edge (e)); |
35b07080 ZD |
303 | |
304 | /* It may happen that by removing path we remove one or more loops | |
305 | we belong to. In this case first unloop the loops, then proceed | |
306 | normally. We may assume that e->dest is not a header of any loop, | |
307 | as it now has exactly one predecessor. */ | |
9ba025a2 | 308 | while (loop_outer (e->src->loop_father) |
d47cc544 | 309 | && dominated_by_p (CDI_DOMINATORS, |
35b07080 | 310 | e->src->loop_father->latch, e->dest)) |
d73be268 | 311 | unloop (e->src->loop_father, &irred_invalidated); |
d329e058 | 312 | |
35b07080 | 313 | /* Identify the path. */ |
d47cc544 | 314 | nrem = find_path (e, &rem_bbs); |
617b465c ZD |
315 | |
316 | n_bord_bbs = 0; | |
5ed6ace5 | 317 | bord_bbs = XCNEWVEC (basic_block, n_basic_blocks); |
617b465c ZD |
318 | seen = sbitmap_alloc (last_basic_block); |
319 | sbitmap_zero (seen); | |
320 | ||
321 | /* Find "border" hexes -- i.e. those with predecessor in removed path. */ | |
322 | for (i = 0; i < nrem; i++) | |
323 | SET_BIT (seen, rem_bbs[i]->index); | |
35b07080 | 324 | for (i = 0; i < nrem; i++) |
617b465c | 325 | { |
628f6a4e | 326 | edge_iterator ei; |
35b07080 | 327 | bb = rem_bbs[i]; |
628f6a4e | 328 | FOR_EACH_EDGE (ae, ei, rem_bbs[i]->succs) |
35b07080 ZD |
329 | if (ae->dest != EXIT_BLOCK_PTR && !TEST_BIT (seen, ae->dest->index)) |
330 | { | |
331 | SET_BIT (seen, ae->dest->index); | |
332 | bord_bbs[n_bord_bbs++] = ae->dest; | |
dc14f191 ZD |
333 | |
334 | if (ae->flags & EDGE_IRREDUCIBLE_LOOP) | |
335 | irred_invalidated = true; | |
35b07080 | 336 | } |
617b465c | 337 | } |
617b465c ZD |
338 | |
339 | /* Remove the path. */ | |
340 | from = e->src; | |
14fa2cc0 | 341 | remove_branch (e); |
66f97d31 | 342 | dom_bbs = NULL; |
617b465c ZD |
343 | |
344 | /* Cancel loops contained in the path. */ | |
598ec7bd ZD |
345 | deleted_loop = XNEWVEC (struct loop *, nrem); |
346 | nreml = 0; | |
617b465c ZD |
347 | for (i = 0; i < nrem; i++) |
348 | if (rem_bbs[i]->loop_father->header == rem_bbs[i]) | |
598ec7bd | 349 | deleted_loop[nreml++] = rem_bbs[i]->loop_father; |
617b465c | 350 | |
d47cc544 | 351 | remove_bbs (rem_bbs, nrem); |
617b465c ZD |
352 | free (rem_bbs); |
353 | ||
598ec7bd | 354 | for (i = 0; i < nreml; i++) |
d73be268 | 355 | cancel_loop_tree (deleted_loop[i]); |
598ec7bd ZD |
356 | free (deleted_loop); |
357 | ||
35b07080 | 358 | /* Find blocks whose dominators may be affected. */ |
617b465c ZD |
359 | sbitmap_zero (seen); |
360 | for (i = 0; i < n_bord_bbs; i++) | |
361 | { | |
d47cc544 | 362 | basic_block ldom; |
617b465c | 363 | |
d47cc544 | 364 | bb = get_immediate_dominator (CDI_DOMINATORS, bord_bbs[i]); |
617b465c ZD |
365 | if (TEST_BIT (seen, bb->index)) |
366 | continue; | |
367 | SET_BIT (seen, bb->index); | |
368 | ||
d47cc544 SB |
369 | for (ldom = first_dom_son (CDI_DOMINATORS, bb); |
370 | ldom; | |
371 | ldom = next_dom_son (CDI_DOMINATORS, ldom)) | |
372 | if (!dominated_by_p (CDI_DOMINATORS, from, ldom)) | |
66f97d31 | 373 | VEC_safe_push (basic_block, heap, dom_bbs, ldom); |
617b465c ZD |
374 | } |
375 | ||
617b465c ZD |
376 | free (seen); |
377 | ||
378 | /* Recount dominators. */ | |
66f97d31 ZD |
379 | iterate_fix_dominators (CDI_DOMINATORS, dom_bbs, true); |
380 | VEC_free (basic_block, heap, dom_bbs); | |
35b07080 ZD |
381 | free (bord_bbs); |
382 | ||
617b465c ZD |
383 | /* Fix placements of basic blocks inside loops and the placement of |
384 | loops in the loop tree. */ | |
d73be268 ZD |
385 | fix_bb_placements (from, &irred_invalidated); |
386 | fix_loop_placements (from->loop_father, &irred_invalidated); | |
dc14f191 ZD |
387 | |
388 | if (irred_invalidated | |
f87000d0 | 389 | && loops_state_satisfies_p (LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS)) |
d73be268 | 390 | mark_irreducible_loops (); |
617b465c ZD |
391 | |
392 | return true; | |
393 | } | |
394 | ||
89f8f30f ZD |
395 | /* Creates place for a new LOOP in loops structure. */ |
396 | ||
397 | static void | |
398 | place_new_loop (struct loop *loop) | |
617b465c | 399 | { |
89f8f30f | 400 | loop->num = number_of_loops (); |
9e2f83a5 | 401 | VEC_safe_push (loop_p, gc, current_loops->larray, loop); |
617b465c ZD |
402 | } |
403 | ||
404 | /* Given LOOP structure with filled header and latch, find the body of the | |
d73be268 | 405 | corresponding loop and add it to loops tree. Insert the LOOP as a son of |
598ec7bd ZD |
406 | outer. */ |
407 | ||
89f8f30f | 408 | void |
d73be268 | 409 | add_loop (struct loop *loop, struct loop *outer) |
617b465c ZD |
410 | { |
411 | basic_block *bbs; | |
412 | int i, n; | |
89f8f30f | 413 | struct loop *subloop; |
d24a32a1 ZD |
414 | edge e; |
415 | edge_iterator ei; | |
d329e058 | 416 | |
617b465c | 417 | /* Add it to loop structure. */ |
d73be268 | 418 | place_new_loop (loop); |
598ec7bd | 419 | flow_loop_tree_node_add (outer, loop); |
617b465c ZD |
420 | |
421 | /* Find its nodes. */ | |
89f8f30f ZD |
422 | bbs = XNEWVEC (basic_block, n_basic_blocks); |
423 | n = get_loop_body_with_size (loop, bbs, n_basic_blocks); | |
617b465c ZD |
424 | |
425 | for (i = 0; i < n; i++) | |
598ec7bd | 426 | { |
89f8f30f ZD |
427 | if (bbs[i]->loop_father == outer) |
428 | { | |
429 | remove_bb_from_loops (bbs[i]); | |
430 | add_bb_to_loop (bbs[i], loop); | |
431 | continue; | |
432 | } | |
433 | ||
434 | loop->num_nodes++; | |
435 | ||
436 | /* If we find a direct subloop of OUTER, move it to LOOP. */ | |
437 | subloop = bbs[i]->loop_father; | |
9ba025a2 | 438 | if (loop_outer (subloop) == outer |
89f8f30f ZD |
439 | && subloop->header == bbs[i]) |
440 | { | |
441 | flow_loop_tree_node_remove (subloop); | |
442 | flow_loop_tree_node_add (loop, subloop); | |
443 | } | |
598ec7bd | 444 | } |
617b465c | 445 | |
d24a32a1 ZD |
446 | /* Update the information about loop exit edges. */ |
447 | for (i = 0; i < n; i++) | |
448 | { | |
449 | FOR_EACH_EDGE (e, ei, bbs[i]->succs) | |
450 | { | |
451 | rescan_loop_exit (e, false, false); | |
452 | } | |
453 | } | |
454 | ||
617b465c ZD |
455 | free (bbs); |
456 | } | |
457 | ||
617b465c | 458 | /* Multiply all frequencies in LOOP by NUM/DEN. */ |
03cb2019 | 459 | void |
d329e058 | 460 | scale_loop_frequencies (struct loop *loop, int num, int den) |
617b465c ZD |
461 | { |
462 | basic_block *bbs; | |
463 | ||
464 | bbs = get_loop_body (loop); | |
33156717 | 465 | scale_bbs_frequencies_int (bbs, loop->num_nodes, num, den); |
617b465c ZD |
466 | free (bbs); |
467 | } | |
468 | ||
469 | /* Make area between HEADER_EDGE and LATCH_EDGE a loop by connecting | |
d73be268 | 470 | latch to header and update loop tree and dominators |
617b465c ZD |
471 | accordingly. Everything between them plus LATCH_EDGE destination must |
472 | be dominated by HEADER_EDGE destination, and back-reachable from | |
473 | LATCH_EDGE source. HEADER_EDGE is redirected to basic block SWITCH_BB, | |
5132abc2 KH |
474 | FALSE_EDGE of SWITCH_BB to original destination of HEADER_EDGE and |
475 | TRUE_EDGE of SWITCH_BB to original destination of LATCH_EDGE. | |
03cb2019 ZD |
476 | Returns the newly created loop. Frequencies and counts in the new loop |
477 | are scaled by FALSE_SCALE and in the old one by TRUE_SCALE. */ | |
50654f6c | 478 | |
617b465c | 479 | struct loop * |
d73be268 | 480 | loopify (edge latch_edge, edge header_edge, |
5132abc2 | 481 | basic_block switch_bb, edge true_edge, edge false_edge, |
03cb2019 | 482 | bool redirect_all_edges, unsigned true_scale, unsigned false_scale) |
617b465c ZD |
483 | { |
484 | basic_block succ_bb = latch_edge->dest; | |
485 | basic_block pred_bb = header_edge->src; | |
66f97d31 ZD |
486 | basic_block *body; |
487 | VEC (basic_block, heap) *dom_bbs; | |
488 | unsigned i; | |
617b465c | 489 | sbitmap seen; |
6270df4c | 490 | struct loop *loop = alloc_loop (); |
9ba025a2 | 491 | struct loop *outer = loop_outer (succ_bb->loop_father); |
03cb2019 | 492 | int freq; |
617b465c ZD |
493 | gcov_type cnt; |
494 | edge e; | |
628f6a4e | 495 | edge_iterator ei; |
617b465c ZD |
496 | |
497 | loop->header = header_edge->dest; | |
498 | loop->latch = latch_edge->src; | |
499 | ||
500 | freq = EDGE_FREQUENCY (header_edge); | |
501 | cnt = header_edge->count; | |
617b465c ZD |
502 | |
503 | /* Redirect edges. */ | |
504 | loop_redirect_edge (latch_edge, loop->header); | |
5132abc2 | 505 | loop_redirect_edge (true_edge, succ_bb); |
50654f6c | 506 | |
92fc4a2f ZD |
507 | /* During loop versioning, one of the switch_bb edge is already properly |
508 | set. Do not redirect it again unless redirect_all_edges is true. */ | |
509 | if (redirect_all_edges) | |
510 | { | |
511 | loop_redirect_edge (header_edge, switch_bb); | |
c22cacf3 MS |
512 | loop_redirect_edge (false_edge, loop->header); |
513 | ||
92fc4a2f ZD |
514 | /* Update dominators. */ |
515 | set_immediate_dominator (CDI_DOMINATORS, switch_bb, pred_bb); | |
516 | set_immediate_dominator (CDI_DOMINATORS, loop->header, switch_bb); | |
517 | } | |
50654f6c | 518 | |
d47cc544 | 519 | set_immediate_dominator (CDI_DOMINATORS, succ_bb, switch_bb); |
617b465c ZD |
520 | |
521 | /* Compute new loop. */ | |
d73be268 | 522 | add_loop (loop, outer); |
617b465c ZD |
523 | |
524 | /* Add switch_bb to appropriate loop. */ | |
598ec7bd ZD |
525 | if (switch_bb->loop_father) |
526 | remove_bb_from_loops (switch_bb); | |
617b465c ZD |
527 | add_bb_to_loop (switch_bb, outer); |
528 | ||
529 | /* Fix frequencies. */ | |
03cb2019 ZD |
530 | if (redirect_all_edges) |
531 | { | |
532 | switch_bb->frequency = freq; | |
533 | switch_bb->count = cnt; | |
534 | FOR_EACH_EDGE (e, ei, switch_bb->succs) | |
535 | { | |
536 | e->count = (switch_bb->count * e->probability) / REG_BR_PROB_BASE; | |
537 | } | |
538 | } | |
539 | scale_loop_frequencies (loop, false_scale, REG_BR_PROB_BASE); | |
540 | scale_loop_frequencies (succ_bb->loop_father, true_scale, REG_BR_PROB_BASE); | |
617b465c ZD |
541 | |
542 | /* Update dominators of blocks outside of LOOP. */ | |
66f97d31 | 543 | dom_bbs = NULL; |
617b465c ZD |
544 | seen = sbitmap_alloc (last_basic_block); |
545 | sbitmap_zero (seen); | |
546 | body = get_loop_body (loop); | |
547 | ||
548 | for (i = 0; i < loop->num_nodes; i++) | |
549 | SET_BIT (seen, body[i]->index); | |
550 | ||
551 | for (i = 0; i < loop->num_nodes; i++) | |
552 | { | |
d47cc544 | 553 | basic_block ldom; |
617b465c | 554 | |
d47cc544 SB |
555 | for (ldom = first_dom_son (CDI_DOMINATORS, body[i]); |
556 | ldom; | |
557 | ldom = next_dom_son (CDI_DOMINATORS, ldom)) | |
558 | if (!TEST_BIT (seen, ldom->index)) | |
617b465c | 559 | { |
d47cc544 | 560 | SET_BIT (seen, ldom->index); |
66f97d31 | 561 | VEC_safe_push (basic_block, heap, dom_bbs, ldom); |
617b465c | 562 | } |
617b465c ZD |
563 | } |
564 | ||
66f97d31 | 565 | iterate_fix_dominators (CDI_DOMINATORS, dom_bbs, false); |
617b465c ZD |
566 | |
567 | free (body); | |
568 | free (seen); | |
66f97d31 | 569 | VEC_free (basic_block, heap, dom_bbs); |
617b465c ZD |
570 | |
571 | return loop; | |
572 | } | |
573 | ||
d73be268 | 574 | /* Remove the latch edge of a LOOP and update loops to indicate that |
35b07080 | 575 | the LOOP was removed. After this function, original loop latch will |
dc14f191 ZD |
576 | have no successor, which caller is expected to fix somehow. |
577 | ||
578 | If this may cause the information about irreducible regions to become | |
579 | invalid, IRRED_INVALIDATED is set to true. */ | |
580 | ||
25a6c68b | 581 | static void |
d73be268 | 582 | unloop (struct loop *loop, bool *irred_invalidated) |
35b07080 ZD |
583 | { |
584 | basic_block *body; | |
585 | struct loop *ploop; | |
586 | unsigned i, n; | |
587 | basic_block latch = loop->latch; | |
dc14f191 ZD |
588 | bool dummy = false; |
589 | ||
590 | if (loop_preheader_edge (loop)->flags & EDGE_IRREDUCIBLE_LOOP) | |
591 | *irred_invalidated = true; | |
35b07080 | 592 | |
e0bb17a8 | 593 | /* This is relatively straightforward. The dominators are unchanged, as |
35b07080 ZD |
594 | loop header dominates loop latch, so the only thing we have to care of |
595 | is the placement of loops and basic blocks inside the loop tree. We | |
596 | move them all to the loop->outer, and then let fix_bb_placements do | |
597 | its work. */ | |
598 | ||
599 | body = get_loop_body (loop); | |
35b07080 ZD |
600 | n = loop->num_nodes; |
601 | for (i = 0; i < n; i++) | |
602 | if (body[i]->loop_father == loop) | |
603 | { | |
604 | remove_bb_from_loops (body[i]); | |
9ba025a2 | 605 | add_bb_to_loop (body[i], loop_outer (loop)); |
35b07080 ZD |
606 | } |
607 | free(body); | |
608 | ||
609 | while (loop->inner) | |
610 | { | |
611 | ploop = loop->inner; | |
612 | flow_loop_tree_node_remove (ploop); | |
9ba025a2 | 613 | flow_loop_tree_node_add (loop_outer (loop), ploop); |
35b07080 ZD |
614 | } |
615 | ||
616 | /* Remove the loop and free its data. */ | |
42fd6772 | 617 | delete_loop (loop); |
35b07080 | 618 | |
c5cbcccf | 619 | remove_edge (single_succ_edge (latch)); |
dc14f191 ZD |
620 | |
621 | /* We do not pass IRRED_INVALIDATED to fix_bb_placements here, as even if | |
622 | there is an irreducible region inside the cancelled loop, the flags will | |
623 | be still correct. */ | |
d73be268 | 624 | fix_bb_placements (latch, &dummy); |
35b07080 ZD |
625 | } |
626 | ||
617b465c ZD |
627 | /* Fix placement of superloops of LOOP inside loop tree, i.e. ensure that |
628 | condition stated in description of fix_loop_placement holds for them. | |
629 | It is used in case when we removed some edges coming out of LOOP, which | |
dc14f191 ZD |
630 | may cause the right placement of LOOP inside loop tree to change. |
631 | ||
632 | IRRED_INVALIDATED is set to true if a change in the loop structures might | |
633 | invalidate the information about irreducible regions. */ | |
634 | ||
617b465c | 635 | static void |
d73be268 | 636 | fix_loop_placements (struct loop *loop, bool *irred_invalidated) |
617b465c ZD |
637 | { |
638 | struct loop *outer; | |
639 | ||
9ba025a2 | 640 | while (loop_outer (loop)) |
617b465c | 641 | { |
9ba025a2 | 642 | outer = loop_outer (loop); |
617b465c | 643 | if (!fix_loop_placement (loop)) |
c22cacf3 | 644 | break; |
1548580c EB |
645 | |
646 | /* Changing the placement of a loop in the loop tree may alter the | |
647 | validity of condition 2) of the description of fix_bb_placement | |
648 | for its preheader, because the successor is the header and belongs | |
649 | to the loop. So call fix_bb_placements to fix up the placement | |
650 | of the preheader and (possibly) of its predecessors. */ | |
d73be268 | 651 | fix_bb_placements (loop_preheader_edge (loop)->src, |
dc14f191 | 652 | irred_invalidated); |
617b465c ZD |
653 | loop = outer; |
654 | } | |
655 | } | |
656 | ||
617b465c | 657 | /* Copies copy of LOOP as subloop of TARGET loop, placing newly |
d73be268 | 658 | created loop into loops structure. */ |
f67d92e9 | 659 | struct loop * |
d73be268 | 660 | duplicate_loop (struct loop *loop, struct loop *target) |
617b465c ZD |
661 | { |
662 | struct loop *cloop; | |
6270df4c | 663 | cloop = alloc_loop (); |
d73be268 | 664 | place_new_loop (cloop); |
617b465c | 665 | |
99f8a411 | 666 | /* Mark the new loop as copy of LOOP. */ |
561e8a90 | 667 | set_loop_copy (loop, cloop); |
617b465c ZD |
668 | |
669 | /* Add it to target. */ | |
670 | flow_loop_tree_node_add (target, cloop); | |
671 | ||
672 | return cloop; | |
673 | } | |
674 | ||
675 | /* Copies structure of subloops of LOOP into TARGET loop, placing | |
d73be268 | 676 | newly created loops into loop tree. */ |
d329e058 | 677 | static void |
d73be268 | 678 | duplicate_subloops (struct loop *loop, struct loop *target) |
617b465c ZD |
679 | { |
680 | struct loop *aloop, *cloop; | |
681 | ||
682 | for (aloop = loop->inner; aloop; aloop = aloop->next) | |
683 | { | |
d73be268 ZD |
684 | cloop = duplicate_loop (aloop, target); |
685 | duplicate_subloops (aloop, cloop); | |
617b465c ZD |
686 | } |
687 | } | |
688 | ||
689 | /* Copies structure of subloops of N loops, stored in array COPIED_LOOPS, | |
d73be268 | 690 | into TARGET loop, placing newly created loops into loop tree. */ |
d329e058 | 691 | static void |
d73be268 | 692 | copy_loops_to (struct loop **copied_loops, int n, struct loop *target) |
617b465c ZD |
693 | { |
694 | struct loop *aloop; | |
695 | int i; | |
696 | ||
697 | for (i = 0; i < n; i++) | |
698 | { | |
d73be268 ZD |
699 | aloop = duplicate_loop (copied_loops[i], target); |
700 | duplicate_subloops (copied_loops[i], aloop); | |
617b465c ZD |
701 | } |
702 | } | |
703 | ||
704 | /* Redirects edge E to basic block DEST. */ | |
705 | static void | |
d329e058 | 706 | loop_redirect_edge (edge e, basic_block dest) |
617b465c ZD |
707 | { |
708 | if (e->dest == dest) | |
709 | return; | |
710 | ||
9ee634e3 | 711 | redirect_edge_and_branch_force (e, dest); |
617b465c ZD |
712 | } |
713 | ||
617b465c ZD |
714 | /* Check whether LOOP's body can be duplicated. */ |
715 | bool | |
ed7a4b4b | 716 | can_duplicate_loop_p (const struct loop *loop) |
617b465c | 717 | { |
8d28e87d ZD |
718 | int ret; |
719 | basic_block *bbs = get_loop_body (loop); | |
617b465c | 720 | |
8d28e87d | 721 | ret = can_copy_bbs_p (bbs, loop->num_nodes); |
617b465c | 722 | free (bbs); |
c22cacf3 | 723 | |
8d28e87d | 724 | return ret; |
617b465c ZD |
725 | } |
726 | ||
03cb2019 ZD |
727 | /* Sets probability and count of edge E to zero. The probability and count |
728 | is redistributed evenly to the remaining edges coming from E->src. */ | |
729 | ||
730 | static void | |
731 | set_zero_probability (edge e) | |
732 | { | |
733 | basic_block bb = e->src; | |
734 | edge_iterator ei; | |
735 | edge ae, last = NULL; | |
736 | unsigned n = EDGE_COUNT (bb->succs); | |
737 | gcov_type cnt = e->count, cnt1; | |
738 | unsigned prob = e->probability, prob1; | |
739 | ||
740 | gcc_assert (n > 1); | |
741 | cnt1 = cnt / (n - 1); | |
742 | prob1 = prob / (n - 1); | |
743 | ||
744 | FOR_EACH_EDGE (ae, ei, bb->succs) | |
745 | { | |
746 | if (ae == e) | |
747 | continue; | |
748 | ||
749 | ae->probability += prob1; | |
750 | ae->count += cnt1; | |
751 | last = ae; | |
752 | } | |
753 | ||
754 | /* Move the rest to one of the edges. */ | |
755 | last->probability += prob % (n - 1); | |
756 | last->count += cnt % (n - 1); | |
757 | ||
758 | e->probability = 0; | |
759 | e->count = 0; | |
760 | } | |
761 | ||
8d28e87d | 762 | /* Duplicates body of LOOP to given edge E NDUPL times. Takes care of updating |
d73be268 | 763 | loop structure and dominators. E's destination must be LOOP header for |
8d28e87d ZD |
764 | this to work, i.e. it must be entry or latch edge of this loop; these are |
765 | unique, as the loops must have preheaders for this function to work | |
766 | correctly (in case E is latch, the function unrolls the loop, if E is entry | |
767 | edge, it peels the loop). Store edges created by copying ORIG edge from | |
768 | copies corresponding to set bits in WONT_EXIT bitmap (bit 0 corresponds to | |
769 | original LOOP body, the other copies are numbered in order given by control | |
770 | flow through them) into TO_REMOVE array. Returns false if duplication is | |
771 | impossible. */ | |
ee8c1b05 | 772 | |
1cb7dfc3 | 773 | bool |
d73be268 | 774 | duplicate_loop_to_header_edge (struct loop *loop, edge e, |
d329e058 | 775 | unsigned int ndupl, sbitmap wont_exit, |
ee8c1b05 ZD |
776 | edge orig, VEC (edge, heap) **to_remove, |
777 | int flags) | |
617b465c ZD |
778 | { |
779 | struct loop *target, *aloop; | |
780 | struct loop **orig_loops; | |
781 | unsigned n_orig_loops; | |
782 | basic_block header = loop->header, latch = loop->latch; | |
783 | basic_block *new_bbs, *bbs, *first_active; | |
784 | basic_block new_bb, bb, first_active_latch = NULL; | |
8d28e87d ZD |
785 | edge ae, latch_edge; |
786 | edge spec_edges[2], new_spec_edges[2]; | |
787 | #define SE_LATCH 0 | |
788 | #define SE_ORIG 1 | |
617b465c ZD |
789 | unsigned i, j, n; |
790 | int is_latch = (latch == e->src); | |
791 | int scale_act = 0, *scale_step = NULL, scale_main = 0; | |
03cb2019 | 792 | int scale_after_exit = 0; |
617b465c ZD |
793 | int p, freq_in, freq_le, freq_out_orig; |
794 | int prob_pass_thru, prob_pass_wont_exit, prob_pass_main; | |
795 | int add_irreducible_flag; | |
b9a66240 | 796 | basic_block place_after; |
03cb2019 ZD |
797 | bitmap bbs_to_scale = NULL; |
798 | bitmap_iterator bi; | |
617b465c | 799 | |
341c100f NS |
800 | gcc_assert (e->dest == loop->header); |
801 | gcc_assert (ndupl > 0); | |
617b465c ZD |
802 | |
803 | if (orig) | |
804 | { | |
805 | /* Orig must be edge out of the loop. */ | |
341c100f NS |
806 | gcc_assert (flow_bb_inside_loop_p (loop, orig->src)); |
807 | gcc_assert (!flow_bb_inside_loop_p (loop, orig->dest)); | |
617b465c ZD |
808 | } |
809 | ||
b9a66240 ZD |
810 | n = loop->num_nodes; |
811 | bbs = get_loop_body_in_dom_order (loop); | |
812 | gcc_assert (bbs[0] == loop->header); | |
813 | gcc_assert (bbs[n - 1] == loop->latch); | |
617b465c ZD |
814 | |
815 | /* Check whether duplication is possible. */ | |
8d28e87d | 816 | if (!can_copy_bbs_p (bbs, loop->num_nodes)) |
617b465c | 817 | { |
8d28e87d ZD |
818 | free (bbs); |
819 | return false; | |
617b465c | 820 | } |
5ed6ace5 | 821 | new_bbs = XNEWVEC (basic_block, loop->num_nodes); |
617b465c | 822 | |
8d28e87d ZD |
823 | /* In case we are doing loop peeling and the loop is in the middle of |
824 | irreducible region, the peeled copies will be inside it too. */ | |
825 | add_irreducible_flag = e->flags & EDGE_IRREDUCIBLE_LOOP; | |
341c100f | 826 | gcc_assert (!is_latch || !add_irreducible_flag); |
617b465c ZD |
827 | |
828 | /* Find edge from latch. */ | |
829 | latch_edge = loop_latch_edge (loop); | |
830 | ||
831 | if (flags & DLTHE_FLAG_UPDATE_FREQ) | |
832 | { | |
833 | /* Calculate coefficients by that we have to scale frequencies | |
834 | of duplicated loop bodies. */ | |
835 | freq_in = header->frequency; | |
836 | freq_le = EDGE_FREQUENCY (latch_edge); | |
837 | if (freq_in == 0) | |
838 | freq_in = 1; | |
839 | if (freq_in < freq_le) | |
840 | freq_in = freq_le; | |
841 | freq_out_orig = orig ? EDGE_FREQUENCY (orig) : freq_in - freq_le; | |
842 | if (freq_out_orig > freq_in - freq_le) | |
843 | freq_out_orig = freq_in - freq_le; | |
844 | prob_pass_thru = RDIV (REG_BR_PROB_BASE * freq_le, freq_in); | |
845 | prob_pass_wont_exit = | |
846 | RDIV (REG_BR_PROB_BASE * (freq_le + freq_out_orig), freq_in); | |
847 | ||
03cb2019 ZD |
848 | if (orig |
849 | && REG_BR_PROB_BASE - orig->probability != 0) | |
850 | { | |
851 | /* The blocks that are dominated by a removed exit edge ORIG have | |
852 | frequencies scaled by this. */ | |
853 | scale_after_exit = RDIV (REG_BR_PROB_BASE * REG_BR_PROB_BASE, | |
854 | REG_BR_PROB_BASE - orig->probability); | |
855 | bbs_to_scale = BITMAP_ALLOC (NULL); | |
856 | for (i = 0; i < n; i++) | |
857 | { | |
858 | if (bbs[i] != orig->src | |
859 | && dominated_by_p (CDI_DOMINATORS, bbs[i], orig->src)) | |
860 | bitmap_set_bit (bbs_to_scale, i); | |
861 | } | |
862 | } | |
863 | ||
5ed6ace5 | 864 | scale_step = XNEWVEC (int, ndupl); |
617b465c | 865 | |
03cb2019 ZD |
866 | for (i = 1; i <= ndupl; i++) |
867 | scale_step[i - 1] = TEST_BIT (wont_exit, i) | |
617b465c ZD |
868 | ? prob_pass_wont_exit |
869 | : prob_pass_thru; | |
870 | ||
a4d05547 | 871 | /* Complete peeling is special as the probability of exit in last |
c22cacf3 | 872 | copy becomes 1. */ |
178df94f JH |
873 | if (flags & DLTHE_FLAG_COMPLETTE_PEEL) |
874 | { | |
875 | int wanted_freq = EDGE_FREQUENCY (e); | |
876 | ||
877 | if (wanted_freq > freq_in) | |
878 | wanted_freq = freq_in; | |
879 | ||
880 | gcc_assert (!is_latch); | |
a4d05547 | 881 | /* First copy has frequency of incoming edge. Each subsequent |
178df94f JH |
882 | frequency should be reduced by prob_pass_wont_exit. Caller |
883 | should've managed the flags so all except for original loop | |
884 | has won't exist set. */ | |
885 | scale_act = RDIV (wanted_freq * REG_BR_PROB_BASE, freq_in); | |
886 | /* Now simulate the duplication adjustments and compute header | |
887 | frequency of the last copy. */ | |
888 | for (i = 0; i < ndupl; i++) | |
c22cacf3 | 889 | wanted_freq = RDIV (wanted_freq * scale_step[i], REG_BR_PROB_BASE); |
178df94f JH |
890 | scale_main = RDIV (wanted_freq * REG_BR_PROB_BASE, freq_in); |
891 | } | |
892 | else if (is_latch) | |
617b465c ZD |
893 | { |
894 | prob_pass_main = TEST_BIT (wont_exit, 0) | |
895 | ? prob_pass_wont_exit | |
896 | : prob_pass_thru; | |
897 | p = prob_pass_main; | |
898 | scale_main = REG_BR_PROB_BASE; | |
899 | for (i = 0; i < ndupl; i++) | |
900 | { | |
901 | scale_main += p; | |
902 | p = RDIV (p * scale_step[i], REG_BR_PROB_BASE); | |
903 | } | |
904 | scale_main = RDIV (REG_BR_PROB_BASE * REG_BR_PROB_BASE, scale_main); | |
905 | scale_act = RDIV (scale_main * prob_pass_main, REG_BR_PROB_BASE); | |
906 | } | |
907 | else | |
908 | { | |
909 | scale_main = REG_BR_PROB_BASE; | |
910 | for (i = 0; i < ndupl; i++) | |
911 | scale_main = RDIV (scale_main * scale_step[i], REG_BR_PROB_BASE); | |
912 | scale_act = REG_BR_PROB_BASE - prob_pass_thru; | |
913 | } | |
914 | for (i = 0; i < ndupl; i++) | |
341c100f NS |
915 | gcc_assert (scale_step[i] >= 0 && scale_step[i] <= REG_BR_PROB_BASE); |
916 | gcc_assert (scale_main >= 0 && scale_main <= REG_BR_PROB_BASE | |
917 | && scale_act >= 0 && scale_act <= REG_BR_PROB_BASE); | |
617b465c ZD |
918 | } |
919 | ||
920 | /* Loop the new bbs will belong to. */ | |
8d28e87d | 921 | target = e->src->loop_father; |
617b465c ZD |
922 | |
923 | /* Original loops. */ | |
924 | n_orig_loops = 0; | |
925 | for (aloop = loop->inner; aloop; aloop = aloop->next) | |
926 | n_orig_loops++; | |
5ed6ace5 | 927 | orig_loops = XCNEWVEC (struct loop *, n_orig_loops); |
617b465c ZD |
928 | for (aloop = loop->inner, i = 0; aloop; aloop = aloop->next, i++) |
929 | orig_loops[i] = aloop; | |
930 | ||
561e8a90 | 931 | set_loop_copy (loop, target); |
d329e058 | 932 | |
5ed6ace5 | 933 | first_active = XNEWVEC (basic_block, n); |
617b465c ZD |
934 | if (is_latch) |
935 | { | |
936 | memcpy (first_active, bbs, n * sizeof (basic_block)); | |
937 | first_active_latch = latch; | |
938 | } | |
939 | ||
8d28e87d ZD |
940 | spec_edges[SE_ORIG] = orig; |
941 | spec_edges[SE_LATCH] = latch_edge; | |
d329e058 | 942 | |
b9a66240 | 943 | place_after = e->src; |
617b465c ZD |
944 | for (j = 0; j < ndupl; j++) |
945 | { | |
946 | /* Copy loops. */ | |
d73be268 | 947 | copy_loops_to (orig_loops, n_orig_loops, target); |
617b465c ZD |
948 | |
949 | /* Copy bbs. */ | |
b9a66240 ZD |
950 | copy_bbs (bbs, n, new_bbs, spec_edges, 2, new_spec_edges, loop, |
951 | place_after); | |
952 | place_after = new_spec_edges[SE_LATCH]->src; | |
8d28e87d | 953 | |
7f7b1718 JH |
954 | if (flags & DLTHE_RECORD_COPY_NUMBER) |
955 | for (i = 0; i < n; i++) | |
956 | { | |
957 | gcc_assert (!new_bbs[i]->aux); | |
958 | new_bbs[i]->aux = (void *)(size_t)(j + 1); | |
959 | } | |
113d659a | 960 | |
84d45ad1 ZD |
961 | /* Note whether the blocks and edges belong to an irreducible loop. */ |
962 | if (add_irreducible_flag) | |
963 | { | |
964 | for (i = 0; i < n; i++) | |
6580ee77 | 965 | new_bbs[i]->flags |= BB_DUPLICATED; |
84d45ad1 ZD |
966 | for (i = 0; i < n; i++) |
967 | { | |
628f6a4e | 968 | edge_iterator ei; |
84d45ad1 ZD |
969 | new_bb = new_bbs[i]; |
970 | if (new_bb->loop_father == target) | |
971 | new_bb->flags |= BB_IRREDUCIBLE_LOOP; | |
972 | ||
628f6a4e | 973 | FOR_EACH_EDGE (ae, ei, new_bb->succs) |
6580ee77 | 974 | if ((ae->dest->flags & BB_DUPLICATED) |
84d45ad1 ZD |
975 | && (ae->src->loop_father == target |
976 | || ae->dest->loop_father == target)) | |
977 | ae->flags |= EDGE_IRREDUCIBLE_LOOP; | |
978 | } | |
979 | for (i = 0; i < n; i++) | |
6580ee77 | 980 | new_bbs[i]->flags &= ~BB_DUPLICATED; |
84d45ad1 ZD |
981 | } |
982 | ||
8d28e87d | 983 | /* Redirect the special edges. */ |
617b465c | 984 | if (is_latch) |
8d28e87d ZD |
985 | { |
986 | redirect_edge_and_branch_force (latch_edge, new_bbs[0]); | |
987 | redirect_edge_and_branch_force (new_spec_edges[SE_LATCH], | |
988 | loop->header); | |
d47cc544 | 989 | set_immediate_dominator (CDI_DOMINATORS, new_bbs[0], latch); |
b9a66240 | 990 | latch = loop->latch = new_bbs[n - 1]; |
8d28e87d ZD |
991 | e = latch_edge = new_spec_edges[SE_LATCH]; |
992 | } | |
993 | else | |
994 | { | |
995 | redirect_edge_and_branch_force (new_spec_edges[SE_LATCH], | |
996 | loop->header); | |
997 | redirect_edge_and_branch_force (e, new_bbs[0]); | |
d47cc544 | 998 | set_immediate_dominator (CDI_DOMINATORS, new_bbs[0], e->src); |
8d28e87d ZD |
999 | e = new_spec_edges[SE_LATCH]; |
1000 | } | |
617b465c | 1001 | |
8d28e87d ZD |
1002 | /* Record exit edge in this copy. */ |
1003 | if (orig && TEST_BIT (wont_exit, j + 1)) | |
ee8c1b05 ZD |
1004 | { |
1005 | if (to_remove) | |
1006 | VEC_safe_push (edge, heap, *to_remove, new_spec_edges[SE_ORIG]); | |
03cb2019 ZD |
1007 | set_zero_probability (new_spec_edges[SE_ORIG]); |
1008 | ||
1009 | /* Scale the frequencies of the blocks dominated by the exit. */ | |
1010 | if (bbs_to_scale) | |
1011 | { | |
1012 | EXECUTE_IF_SET_IN_BITMAP (bbs_to_scale, 0, i, bi) | |
1013 | { | |
1014 | scale_bbs_frequencies_int (new_bbs + i, 1, scale_after_exit, | |
1015 | REG_BR_PROB_BASE); | |
1016 | } | |
1017 | } | |
ee8c1b05 | 1018 | } |
d329e058 | 1019 | |
8d28e87d ZD |
1020 | /* Record the first copy in the control flow order if it is not |
1021 | the original loop (i.e. in case of peeling). */ | |
617b465c ZD |
1022 | if (!first_active_latch) |
1023 | { | |
1024 | memcpy (first_active, new_bbs, n * sizeof (basic_block)); | |
b9a66240 | 1025 | first_active_latch = new_bbs[n - 1]; |
617b465c | 1026 | } |
d329e058 | 1027 | |
8d28e87d ZD |
1028 | /* Set counts and frequencies. */ |
1029 | if (flags & DLTHE_FLAG_UPDATE_FREQ) | |
617b465c | 1030 | { |
33156717 | 1031 | scale_bbs_frequencies_int (new_bbs, n, scale_act, REG_BR_PROB_BASE); |
8d28e87d | 1032 | scale_act = RDIV (scale_act * scale_step[j], REG_BR_PROB_BASE); |
617b465c ZD |
1033 | } |
1034 | } | |
8d28e87d ZD |
1035 | free (new_bbs); |
1036 | free (orig_loops); | |
c22cacf3 | 1037 | |
ee8c1b05 ZD |
1038 | /* Record the exit edge in the original loop body, and update the frequencies. */ |
1039 | if (orig && TEST_BIT (wont_exit, 0)) | |
1040 | { | |
1041 | if (to_remove) | |
1042 | VEC_safe_push (edge, heap, *to_remove, orig); | |
03cb2019 ZD |
1043 | set_zero_probability (orig); |
1044 | ||
1045 | /* Scale the frequencies of the blocks dominated by the exit. */ | |
1046 | if (bbs_to_scale) | |
1047 | { | |
1048 | EXECUTE_IF_SET_IN_BITMAP (bbs_to_scale, 0, i, bi) | |
1049 | { | |
1050 | scale_bbs_frequencies_int (bbs + i, 1, scale_after_exit, | |
1051 | REG_BR_PROB_BASE); | |
1052 | } | |
1053 | } | |
ee8c1b05 ZD |
1054 | } |
1055 | ||
8d28e87d ZD |
1056 | /* Update the original loop. */ |
1057 | if (!is_latch) | |
d47cc544 | 1058 | set_immediate_dominator (CDI_DOMINATORS, e->dest, e->src); |
617b465c ZD |
1059 | if (flags & DLTHE_FLAG_UPDATE_FREQ) |
1060 | { | |
33156717 | 1061 | scale_bbs_frequencies_int (bbs, n, scale_main, REG_BR_PROB_BASE); |
617b465c ZD |
1062 | free (scale_step); |
1063 | } | |
617b465c | 1064 | |
8d28e87d | 1065 | /* Update dominators of outer blocks if affected. */ |
617b465c ZD |
1066 | for (i = 0; i < n; i++) |
1067 | { | |
66f97d31 ZD |
1068 | basic_block dominated, dom_bb; |
1069 | VEC (basic_block, heap) *dom_bbs; | |
1070 | unsigned j; | |
617b465c ZD |
1071 | |
1072 | bb = bbs[i]; | |
6580ee77 | 1073 | bb->aux = 0; |
113d659a | 1074 | |
66f97d31 ZD |
1075 | dom_bbs = get_dominated_by (CDI_DOMINATORS, bb); |
1076 | for (j = 0; VEC_iterate (basic_block, dom_bbs, j, dominated); j++) | |
617b465c | 1077 | { |
617b465c ZD |
1078 | if (flow_bb_inside_loop_p (loop, dominated)) |
1079 | continue; | |
1080 | dom_bb = nearest_common_dominator ( | |
d47cc544 | 1081 | CDI_DOMINATORS, first_active[i], first_active_latch); |
c22cacf3 | 1082 | set_immediate_dominator (CDI_DOMINATORS, dominated, dom_bb); |
617b465c | 1083 | } |
66f97d31 | 1084 | VEC_free (basic_block, heap, dom_bbs); |
617b465c ZD |
1085 | } |
1086 | free (first_active); | |
1087 | ||
1088 | free (bbs); | |
03cb2019 | 1089 | BITMAP_FREE (bbs_to_scale); |
617b465c ZD |
1090 | |
1091 | return true; | |
1092 | } | |
1093 | ||
f470c378 ZD |
1094 | /* A callback for make_forwarder block, to redirect all edges except for |
1095 | MFB_KJ_EDGE to the entry part. E is the edge for that we should decide | |
1096 | whether to redirect it. */ | |
1097 | ||
b02b9b53 ZD |
1098 | edge mfb_kj_edge; |
1099 | bool | |
f470c378 ZD |
1100 | mfb_keep_just (edge e) |
1101 | { | |
1102 | return e != mfb_kj_edge; | |
1103 | } | |
1104 | ||
3d436d2a ZD |
1105 | /* Creates a pre-header for a LOOP. Returns newly created block. Unless |
1106 | CP_SIMPLE_PREHEADERS is set in FLAGS, we only force LOOP to have single | |
1107 | entry; otherwise we also force preheader block to have only one successor. | |
f470c378 ZD |
1108 | The function also updates dominators. */ |
1109 | ||
b02b9b53 | 1110 | basic_block |
d47cc544 | 1111 | create_preheader (struct loop *loop, int flags) |
3d436d2a ZD |
1112 | { |
1113 | edge e, fallthru; | |
1114 | basic_block dummy; | |
3d436d2a | 1115 | int nentry = 0; |
f470c378 | 1116 | bool irred = false; |
c15bc84b | 1117 | bool latch_edge_was_fallthru; |
c7b852c8 | 1118 | edge one_succ_pred = NULL, single_entry = NULL; |
628f6a4e | 1119 | edge_iterator ei; |
3d436d2a | 1120 | |
628f6a4e | 1121 | FOR_EACH_EDGE (e, ei, loop->header->preds) |
3d436d2a ZD |
1122 | { |
1123 | if (e->src == loop->latch) | |
1124 | continue; | |
f470c378 | 1125 | irred |= (e->flags & EDGE_IRREDUCIBLE_LOOP) != 0; |
3d436d2a | 1126 | nentry++; |
c7b852c8 | 1127 | single_entry = e; |
c5cbcccf | 1128 | if (single_succ_p (e->src)) |
c15bc84b | 1129 | one_succ_pred = e; |
3d436d2a | 1130 | } |
341c100f | 1131 | gcc_assert (nentry); |
3d436d2a ZD |
1132 | if (nentry == 1) |
1133 | { | |
89f8f30f ZD |
1134 | if (/* We do not allow entry block to be the loop preheader, since we |
1135 | cannot emit code there. */ | |
c7b852c8 | 1136 | single_entry->src != ENTRY_BLOCK_PTR |
89f8f30f ZD |
1137 | /* If we want simple preheaders, also force the preheader to have |
1138 | just a single successor. */ | |
1139 | && !((flags & CP_SIMPLE_PREHEADERS) | |
c7b852c8 | 1140 | && !single_succ_p (single_entry->src))) |
3d436d2a ZD |
1141 | return NULL; |
1142 | } | |
1143 | ||
f470c378 | 1144 | mfb_kj_edge = loop_latch_edge (loop); |
c15bc84b | 1145 | latch_edge_was_fallthru = (mfb_kj_edge->flags & EDGE_FALLTHRU) != 0; |
598ec7bd | 1146 | fallthru = make_forwarder_block (loop->header, mfb_keep_just, NULL); |
3d436d2a ZD |
1147 | dummy = fallthru->src; |
1148 | loop->header = fallthru->dest; | |
1149 | ||
c15bc84b EB |
1150 | /* Try to be clever in placing the newly created preheader. The idea is to |
1151 | avoid breaking any "fallthruness" relationship between blocks. | |
1152 | ||
1153 | The preheader was created just before the header and all incoming edges | |
1154 | to the header were redirected to the preheader, except the latch edge. | |
1155 | So the only problematic case is when this latch edge was a fallthru | |
1156 | edge: it is not anymore after the preheader creation so we have broken | |
1157 | the fallthruness. We're therefore going to look for a better place. */ | |
1158 | if (latch_edge_was_fallthru) | |
1159 | { | |
1160 | if (one_succ_pred) | |
1161 | e = one_succ_pred; | |
1162 | else | |
1163 | e = EDGE_PRED (dummy, 0); | |
1164 | ||
1165 | move_block_after (dummy, e->src); | |
1166 | } | |
f470c378 | 1167 | |
f470c378 | 1168 | if (irred) |
3d436d2a | 1169 | { |
f470c378 | 1170 | dummy->flags |= BB_IRREDUCIBLE_LOOP; |
c5cbcccf | 1171 | single_succ_edge (dummy)->flags |= EDGE_IRREDUCIBLE_LOOP; |
3d436d2a ZD |
1172 | } |
1173 | ||
c263766c RH |
1174 | if (dump_file) |
1175 | fprintf (dump_file, "Created preheader block for loop %i\n", | |
3d436d2a ZD |
1176 | loop->num); |
1177 | ||
1178 | return dummy; | |
1179 | } | |
1180 | ||
d73be268 ZD |
1181 | /* Create preheaders for each loop; for meaning of FLAGS see create_preheader. */ |
1182 | ||
3d436d2a | 1183 | void |
d73be268 | 1184 | create_preheaders (int flags) |
3d436d2a | 1185 | { |
42fd6772 ZD |
1186 | loop_iterator li; |
1187 | struct loop *loop; | |
1188 | ||
c7b852c8 ZD |
1189 | if (!current_loops) |
1190 | return; | |
1191 | ||
42fd6772 ZD |
1192 | FOR_EACH_LOOP (li, loop, 0) |
1193 | create_preheader (loop, flags); | |
f87000d0 | 1194 | loops_state_set (LOOPS_HAVE_PREHEADERS); |
3d436d2a ZD |
1195 | } |
1196 | ||
d73be268 ZD |
1197 | /* Forces all loop latches to have only single successor. */ |
1198 | ||
3d436d2a | 1199 | void |
d73be268 | 1200 | force_single_succ_latches (void) |
3d436d2a | 1201 | { |
42fd6772 | 1202 | loop_iterator li; |
3d436d2a ZD |
1203 | struct loop *loop; |
1204 | edge e; | |
1205 | ||
42fd6772 | 1206 | FOR_EACH_LOOP (li, loop, 0) |
3d436d2a | 1207 | { |
c5cbcccf | 1208 | if (loop->latch != loop->header && single_succ_p (loop->latch)) |
3d436d2a | 1209 | continue; |
d329e058 | 1210 | |
9ff3d2de | 1211 | e = find_edge (loop->latch, loop->header); |
bc810602 | 1212 | |
598ec7bd | 1213 | split_edge (e); |
3d436d2a | 1214 | } |
f87000d0 | 1215 | loops_state_set (LOOPS_HAVE_SIMPLE_LATCHES); |
3d436d2a ZD |
1216 | } |
1217 | ||
1cb7dfc3 MH |
1218 | /* This function is called from loop_version. It splits the entry edge |
1219 | of the loop we want to version, adds the versioning condition, and | |
1220 | adjust the edges to the two versions of the loop appropriately. | |
1221 | e is an incoming edge. Returns the basic block containing the | |
1222 | condition. | |
1223 | ||
1224 | --- edge e ---- > [second_head] | |
1225 | ||
1226 | Split it and insert new conditional expression and adjust edges. | |
1227 | ||
1228 | --- edge e ---> [cond expr] ---> [first_head] | |
c22cacf3 MS |
1229 | | |
1230 | +---------> [second_head] | |
03cb2019 ZD |
1231 | |
1232 | THEN_PROB is the probability of then branch of the condition. */ | |
1cb7dfc3 MH |
1233 | |
1234 | static basic_block | |
03cb2019 ZD |
1235 | lv_adjust_loop_entry_edge (basic_block first_head, basic_block second_head, |
1236 | edge e, void *cond_expr, unsigned then_prob) | |
1cb7dfc3 MH |
1237 | { |
1238 | basic_block new_head = NULL; | |
1239 | edge e1; | |
1240 | ||
1241 | gcc_assert (e->dest == second_head); | |
1242 | ||
1243 | /* Split edge 'e'. This will create a new basic block, where we can | |
1244 | insert conditional expr. */ | |
1245 | new_head = split_edge (e); | |
1246 | ||
1cb7dfc3 MH |
1247 | lv_add_condition_to_bb (first_head, second_head, new_head, |
1248 | cond_expr); | |
1249 | ||
766613a4 | 1250 | /* Don't set EDGE_TRUE_VALUE in RTL mode, as it's invalid there. */ |
03cb2019 | 1251 | e = single_succ_edge (new_head); |
52bca999 SB |
1252 | e1 = make_edge (new_head, first_head, |
1253 | current_ir_type () == IR_GIMPLE ? EDGE_TRUE_VALUE : 0); | |
03cb2019 ZD |
1254 | e1->probability = then_prob; |
1255 | e->probability = REG_BR_PROB_BASE - then_prob; | |
1256 | e1->count = RDIV (e->count * e1->probability, REG_BR_PROB_BASE); | |
1257 | e->count = RDIV (e->count * e->probability, REG_BR_PROB_BASE); | |
1258 | ||
1cb7dfc3 MH |
1259 | set_immediate_dominator (CDI_DOMINATORS, first_head, new_head); |
1260 | set_immediate_dominator (CDI_DOMINATORS, second_head, new_head); | |
1261 | ||
1262 | /* Adjust loop header phi nodes. */ | |
1263 | lv_adjust_loop_header_phi (first_head, second_head, new_head, e1); | |
1264 | ||
1265 | return new_head; | |
1266 | } | |
1267 | ||
1268 | /* Main entry point for Loop Versioning transformation. | |
c22cacf3 | 1269 | |
b9a66240 ZD |
1270 | This transformation given a condition and a loop, creates |
1271 | -if (condition) { loop_copy1 } else { loop_copy2 }, | |
1272 | where loop_copy1 is the loop transformed in one way, and loop_copy2 | |
1273 | is the loop transformed in another way (or unchanged). 'condition' | |
1274 | may be a run time test for things that were not resolved by static | |
1275 | analysis (overlapping ranges (anti-aliasing), alignment, etc.). | |
1276 | ||
03cb2019 ZD |
1277 | THEN_PROB is the probability of the then edge of the if. THEN_SCALE |
1278 | is the ratio by that the frequencies in the original loop should | |
1279 | be scaled. ELSE_SCALE is the ratio by that the frequencies in the | |
1280 | new loop should be scaled. | |
1281 | ||
b9a66240 ZD |
1282 | If PLACE_AFTER is true, we place the new loop after LOOP in the |
1283 | instruction stream, otherwise it is placed before LOOP. */ | |
1cb7dfc3 MH |
1284 | |
1285 | struct loop * | |
d73be268 | 1286 | loop_version (struct loop *loop, |
b9a66240 | 1287 | void *cond_expr, basic_block *condition_bb, |
03cb2019 | 1288 | unsigned then_prob, unsigned then_scale, unsigned else_scale, |
b9a66240 | 1289 | bool place_after) |
1cb7dfc3 MH |
1290 | { |
1291 | basic_block first_head, second_head; | |
6270df4c | 1292 | edge entry, latch_edge, true_edge, false_edge; |
1cb7dfc3 MH |
1293 | int irred_flag; |
1294 | struct loop *nloop; | |
b9a66240 | 1295 | basic_block cond_bb; |
1cb7dfc3 | 1296 | |
1cb7dfc3 MH |
1297 | /* Record entry and latch edges for the loop */ |
1298 | entry = loop_preheader_edge (loop); | |
1299 | irred_flag = entry->flags & EDGE_IRREDUCIBLE_LOOP; | |
1300 | entry->flags &= ~EDGE_IRREDUCIBLE_LOOP; | |
c22cacf3 | 1301 | |
1cb7dfc3 MH |
1302 | /* Note down head of loop as first_head. */ |
1303 | first_head = entry->dest; | |
1304 | ||
1305 | /* Duplicate loop. */ | |
d73be268 | 1306 | if (!cfg_hook_duplicate_loop_to_header_edge (loop, entry, 1, |
ee8c1b05 | 1307 | NULL, NULL, NULL, 0)) |
1cb7dfc3 MH |
1308 | return NULL; |
1309 | ||
1310 | /* After duplication entry edge now points to new loop head block. | |
1311 | Note down new head as second_head. */ | |
1312 | second_head = entry->dest; | |
1313 | ||
1314 | /* Split loop entry edge and insert new block with cond expr. */ | |
b9a66240 | 1315 | cond_bb = lv_adjust_loop_entry_edge (first_head, second_head, |
03cb2019 | 1316 | entry, cond_expr, then_prob); |
b9a66240 ZD |
1317 | if (condition_bb) |
1318 | *condition_bb = cond_bb; | |
1319 | ||
1320 | if (!cond_bb) | |
1cb7dfc3 MH |
1321 | { |
1322 | entry->flags |= irred_flag; | |
1323 | return NULL; | |
1324 | } | |
1325 | ||
6580ee77 | 1326 | latch_edge = single_succ_edge (get_bb_copy (loop->latch)); |
c22cacf3 | 1327 | |
b9a66240 | 1328 | extract_cond_bb_edges (cond_bb, &true_edge, &false_edge); |
d73be268 | 1329 | nloop = loopify (latch_edge, |
6580ee77 | 1330 | single_pred_edge (get_bb_copy (loop->header)), |
b9a66240 | 1331 | cond_bb, true_edge, false_edge, |
03cb2019 ZD |
1332 | false /* Do not redirect all edges. */, |
1333 | then_scale, else_scale); | |
1cb7dfc3 | 1334 | |
c22cacf3 | 1335 | /* loopify redirected latch_edge. Update its PENDING_STMTS. */ |
1cb7dfc3 MH |
1336 | lv_flush_pending_stmts (latch_edge); |
1337 | ||
c22cacf3 | 1338 | /* loopify redirected condition_bb's succ edge. Update its PENDING_STMTS. */ |
b9a66240 | 1339 | extract_cond_bb_edges (cond_bb, &true_edge, &false_edge); |
1cb7dfc3 MH |
1340 | lv_flush_pending_stmts (false_edge); |
1341 | /* Adjust irreducible flag. */ | |
1342 | if (irred_flag) | |
1343 | { | |
b9a66240 | 1344 | cond_bb->flags |= BB_IRREDUCIBLE_LOOP; |
1cb7dfc3 MH |
1345 | loop_preheader_edge (loop)->flags |= EDGE_IRREDUCIBLE_LOOP; |
1346 | loop_preheader_edge (nloop)->flags |= EDGE_IRREDUCIBLE_LOOP; | |
b9a66240 ZD |
1347 | single_pred_edge (cond_bb)->flags |= EDGE_IRREDUCIBLE_LOOP; |
1348 | } | |
1349 | ||
1350 | if (place_after) | |
1351 | { | |
1352 | basic_block *bbs = get_loop_body_in_dom_order (nloop), after; | |
1353 | unsigned i; | |
1354 | ||
1355 | after = loop->latch; | |
1356 | ||
1357 | for (i = 0; i < nloop->num_nodes; i++) | |
1358 | { | |
1359 | move_block_after (bbs[i], after); | |
1360 | after = bbs[i]; | |
1361 | } | |
1362 | free (bbs); | |
1cb7dfc3 MH |
1363 | } |
1364 | ||
fa10beec RW |
1365 | /* At this point condition_bb is loop preheader with two successors, |
1366 | first_head and second_head. Make sure that loop preheader has only | |
1cb7dfc3 | 1367 | one successor. */ |
598ec7bd ZD |
1368 | split_edge (loop_preheader_edge (loop)); |
1369 | split_edge (loop_preheader_edge (nloop)); | |
1cb7dfc3 MH |
1370 | |
1371 | return nloop; | |
1372 | } | |
1373 | ||
d73be268 | 1374 | /* The structure of loops might have changed. Some loops might get removed |
2b271002 ZD |
1375 | (and their headers and latches were set to NULL), loop exists might get |
1376 | removed (thus the loop nesting may be wrong), and some blocks and edges | |
1377 | were changed (so the information about bb --> loop mapping does not have | |
1378 | to be correct). But still for the remaining loops the header dominates | |
fa10beec | 1379 | the latch, and loops did not get new subloops (new loops might possibly |
2b271002 | 1380 | get created, but we are not interested in them). Fix up the mess. |
c22cacf3 | 1381 | |
2b271002 ZD |
1382 | If CHANGED_BBS is not NULL, basic blocks whose loop has changed are |
1383 | marked in it. */ | |
1384 | ||
1385 | void | |
d73be268 | 1386 | fix_loop_structure (bitmap changed_bbs) |
2b271002 ZD |
1387 | { |
1388 | basic_block bb; | |
1389 | struct loop *loop, *ploop; | |
42fd6772 | 1390 | loop_iterator li; |
c7b852c8 ZD |
1391 | bool record_exits = false; |
1392 | struct loop **superloop = XNEWVEC (struct loop *, number_of_loops ()); | |
2b271002 | 1393 | |
c7b852c8 ZD |
1394 | /* Remove the old bb -> loop mapping. Remember the depth of the blocks in |
1395 | the loop hierarchy, so that we can recognize blocks whose loop nesting | |
1396 | relationship has changed. */ | |
2b271002 ZD |
1397 | FOR_EACH_BB (bb) |
1398 | { | |
c7b852c8 | 1399 | if (changed_bbs) |
9ba025a2 | 1400 | bb->aux = (void *) (size_t) loop_depth (bb->loop_father); |
d73be268 | 1401 | bb->loop_father = current_loops->tree_root; |
2b271002 ZD |
1402 | } |
1403 | ||
f87000d0 | 1404 | if (loops_state_satisfies_p (LOOPS_HAVE_RECORDED_EXITS)) |
c7b852c8 ZD |
1405 | { |
1406 | release_recorded_exits (); | |
1407 | record_exits = true; | |
1408 | } | |
1409 | ||
1410 | /* Remove the dead loops from structures. We start from the innermost | |
1411 | loops, so that when we remove the loops, we know that the loops inside | |
1412 | are preserved, and do not waste time relinking loops that will be | |
1413 | removed later. */ | |
1414 | FOR_EACH_LOOP (li, loop, LI_FROM_INNERMOST) | |
2b271002 | 1415 | { |
2b271002 ZD |
1416 | if (loop->header) |
1417 | continue; | |
1418 | ||
1419 | while (loop->inner) | |
1420 | { | |
1421 | ploop = loop->inner; | |
1422 | flow_loop_tree_node_remove (ploop); | |
9ba025a2 | 1423 | flow_loop_tree_node_add (loop_outer (loop), ploop); |
2b271002 ZD |
1424 | } |
1425 | ||
1426 | /* Remove the loop and free its data. */ | |
42fd6772 | 1427 | delete_loop (loop); |
2b271002 ZD |
1428 | } |
1429 | ||
c7b852c8 ZD |
1430 | /* Rescan the bodies of loops, starting from the outermost ones. We assume |
1431 | that no optimization interchanges the order of the loops, i.e., it cannot | |
1432 | happen that L1 was superloop of L2 before and it is subloop of L2 now | |
1433 | (without explicitly updating loop information). At the same time, we also | |
1434 | determine the new loop structure. */ | |
1435 | current_loops->tree_root->num_nodes = n_basic_blocks; | |
42fd6772 | 1436 | FOR_EACH_LOOP (li, loop, 0) |
2b271002 | 1437 | { |
c7b852c8 | 1438 | superloop[loop->num] = loop->header->loop_father; |
2b271002 ZD |
1439 | loop->num_nodes = flow_loop_nodes_find (loop->header, loop); |
1440 | } | |
1441 | ||
1442 | /* Now fix the loop nesting. */ | |
42fd6772 | 1443 | FOR_EACH_LOOP (li, loop, 0) |
2b271002 | 1444 | { |
c7b852c8 | 1445 | ploop = superloop[loop->num]; |
9ba025a2 | 1446 | if (ploop != loop_outer (loop)) |
2b271002 ZD |
1447 | { |
1448 | flow_loop_tree_node_remove (loop); | |
c7b852c8 | 1449 | flow_loop_tree_node_add (ploop, loop); |
2b271002 ZD |
1450 | } |
1451 | } | |
c7b852c8 | 1452 | free (superloop); |
2b271002 ZD |
1453 | |
1454 | /* Mark the blocks whose loop has changed. */ | |
c7b852c8 | 1455 | if (changed_bbs) |
2b271002 | 1456 | { |
c7b852c8 ZD |
1457 | FOR_EACH_BB (bb) |
1458 | { | |
9ba025a2 | 1459 | if ((void *) (size_t) loop_depth (bb->loop_father) != bb->aux) |
c7b852c8 | 1460 | bitmap_set_bit (changed_bbs, bb->index); |
2b271002 | 1461 | |
c7b852c8 ZD |
1462 | bb->aux = NULL; |
1463 | } | |
2b271002 ZD |
1464 | } |
1465 | ||
f87000d0 | 1466 | if (loops_state_satisfies_p (LOOPS_HAVE_PREHEADERS)) |
c7b852c8 ZD |
1467 | create_preheaders (CP_SIMPLE_PREHEADERS); |
1468 | ||
f87000d0 | 1469 | if (loops_state_satisfies_p (LOOPS_HAVE_SIMPLE_LATCHES)) |
af0b10e5 ZD |
1470 | force_single_succ_latches (); |
1471 | ||
f87000d0 | 1472 | if (loops_state_satisfies_p (LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS)) |
d73be268 | 1473 | mark_irreducible_loops (); |
6270df4c | 1474 | |
c7b852c8 ZD |
1475 | if (record_exits) |
1476 | record_loop_exits (); | |
1477 | ||
1478 | #ifdef ENABLE_CHECKING | |
1479 | verify_loop_structure (); | |
1480 | #endif | |
2b271002 | 1481 | } |