]>
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 | |
18 | Software Foundation, 59 Temple Place - Suite 330, Boston, MA | |
19 | 02111-1307, USA. */ | |
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 *); | |
1548580c | 44 | static void fix_loop_placements (struct loops *, struct loop *); |
d329e058 AJ |
45 | static bool fix_bb_placement (struct loops *, basic_block); |
46 | static void fix_bb_placements (struct loops *, basic_block); | |
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); |
d329e058 | 50 | static void fix_irreducible_loops (basic_block); |
25a6c68b | 51 | static void unloop (struct loops *, struct loop *); |
3d436d2a | 52 | |
bade3a00 JH |
53 | #define RDIV(X,Y) (((X) + (Y) / 2) / (Y)) |
54 | ||
617b465c ZD |
55 | /* Splits basic block BB after INSN, returns created edge. Updates loops |
56 | and dominators. */ | |
57 | edge | |
92fc4a2f | 58 | split_loop_bb (basic_block bb, void *insn) |
617b465c ZD |
59 | { |
60 | edge e; | |
617b465c ZD |
61 | |
62 | /* Split the block. */ | |
63 | e = split_block (bb, insn); | |
64 | ||
65 | /* Add dest to loop. */ | |
66 | add_bb_to_loop (e->dest, e->src->loop_father); | |
67 | ||
617b465c ZD |
68 | return e; |
69 | } | |
70 | ||
d47cc544 | 71 | /* Checks whether basic block BB is dominated by DATA. */ |
617b465c | 72 | static bool |
d329e058 | 73 | rpe_enum_p (basic_block bb, void *data) |
617b465c | 74 | { |
d47cc544 | 75 | return dominated_by_p (CDI_DOMINATORS, bb, data); |
617b465c ZD |
76 | } |
77 | ||
78 | /* Remove basic blocks BBS from loop structure and dominance info, | |
79 | and delete them afterwards. */ | |
80 | static void | |
d47cc544 | 81 | remove_bbs (basic_block *bbs, int nbbs) |
617b465c ZD |
82 | { |
83 | int i; | |
84 | ||
85 | for (i = 0; i < nbbs; i++) | |
86 | { | |
87 | remove_bb_from_loops (bbs[i]); | |
f470c378 | 88 | delete_basic_block (bbs[i]); |
617b465c ZD |
89 | } |
90 | } | |
91 | ||
92 | /* Find path -- i.e. the basic blocks dominated by edge E and put them | |
93 | into array BBS, that will be allocated large enough to contain them. | |
35b07080 ZD |
94 | E->dest must have exactly one predecessor for this to work (it is |
95 | easy to achieve and we do not put it here because we do not want to | |
96 | alter anything by this function). The number of basic blocks in the | |
97 | path is returned. */ | |
617b465c | 98 | static int |
d47cc544 | 99 | find_path (edge e, basic_block **bbs) |
617b465c | 100 | { |
628f6a4e | 101 | gcc_assert (EDGE_COUNT (e->dest->preds) <= 1); |
617b465c ZD |
102 | |
103 | /* Find bbs in the path. */ | |
617b465c ZD |
104 | *bbs = xcalloc (n_basic_blocks, sizeof (basic_block)); |
105 | return dfs_enumerate_from (e->dest, 0, rpe_enum_p, *bbs, | |
d47cc544 | 106 | n_basic_blocks, e->dest); |
617b465c ZD |
107 | } |
108 | ||
109 | /* Fix placement of basic block BB inside loop hierarchy stored in LOOPS -- | |
110 | Let L be a loop to that BB belongs. Then every successor of BB must either | |
111 | 1) belong to some superloop of loop L, or | |
112 | 2) be a header of loop K such that K->outer is superloop of L | |
113 | Returns true if we had to move BB into other loop to enforce this condition, | |
114 | false if the placement of BB was already correct (provided that placements | |
115 | of its successors are correct). */ | |
116 | static bool | |
d329e058 | 117 | fix_bb_placement (struct loops *loops, basic_block bb) |
617b465c ZD |
118 | { |
119 | edge e; | |
628f6a4e | 120 | edge_iterator ei; |
617b465c ZD |
121 | struct loop *loop = loops->tree_root, *act; |
122 | ||
628f6a4e | 123 | FOR_EACH_EDGE (e, ei, bb->succs) |
617b465c ZD |
124 | { |
125 | if (e->dest == EXIT_BLOCK_PTR) | |
126 | continue; | |
127 | ||
128 | act = e->dest->loop_father; | |
129 | if (act->header == e->dest) | |
130 | act = act->outer; | |
131 | ||
132 | if (flow_loop_nested_p (loop, act)) | |
133 | loop = act; | |
134 | } | |
135 | ||
136 | if (loop == bb->loop_father) | |
137 | return false; | |
138 | ||
139 | remove_bb_from_loops (bb); | |
140 | add_bb_to_loop (bb, loop); | |
141 | ||
142 | return true; | |
143 | } | |
144 | ||
145 | /* Fix placements of basic blocks inside loop hierarchy stored in loops; i.e. | |
146 | enforce condition condition stated in description of fix_bb_placement. We | |
147 | start from basic block FROM that had some of its successors removed, so that | |
148 | his placement no longer has to be correct, and iteratively fix placement of | |
149 | its predecessors that may change if placement of FROM changed. Also fix | |
150 | placement of subloops of FROM->loop_father, that might also be altered due | |
4d6922ee | 151 | to this change; the condition for them is similar, except that instead of |
617b465c ZD |
152 | successors we consider edges coming out of the loops. */ |
153 | static void | |
d329e058 | 154 | fix_bb_placements (struct loops *loops, basic_block from) |
617b465c ZD |
155 | { |
156 | sbitmap in_queue; | |
157 | basic_block *queue, *qtop, *qbeg, *qend; | |
158 | struct loop *base_loop; | |
159 | edge e; | |
160 | ||
161 | /* We pass through blocks back-reachable from FROM, testing whether some | |
162 | of their successors moved to outer loop. It may be necessary to | |
163 | iterate several times, but it is finite, as we stop unless we move | |
164 | the basic block up the loop structure. The whole story is a bit | |
165 | more complicated due to presence of subloops, those are moved using | |
166 | fix_loop_placement. */ | |
167 | ||
168 | base_loop = from->loop_father; | |
169 | if (base_loop == loops->tree_root) | |
170 | return; | |
171 | ||
172 | in_queue = sbitmap_alloc (last_basic_block); | |
173 | sbitmap_zero (in_queue); | |
174 | SET_BIT (in_queue, from->index); | |
175 | /* Prevent us from going out of the base_loop. */ | |
176 | SET_BIT (in_queue, base_loop->header->index); | |
177 | ||
35b07080 | 178 | queue = xmalloc ((base_loop->num_nodes + 1) * sizeof (basic_block)); |
617b465c ZD |
179 | qtop = queue + base_loop->num_nodes + 1; |
180 | qbeg = queue; | |
181 | qend = queue + 1; | |
182 | *qbeg = from; | |
183 | ||
184 | while (qbeg != qend) | |
185 | { | |
628f6a4e | 186 | edge_iterator ei; |
617b465c ZD |
187 | from = *qbeg; |
188 | qbeg++; | |
189 | if (qbeg == qtop) | |
190 | qbeg = queue; | |
191 | RESET_BIT (in_queue, from->index); | |
192 | ||
193 | if (from->loop_father->header == from) | |
194 | { | |
195 | /* Subloop header, maybe move the loop upward. */ | |
196 | if (!fix_loop_placement (from->loop_father)) | |
197 | continue; | |
198 | } | |
199 | else | |
200 | { | |
201 | /* Ordinary basic block. */ | |
202 | if (!fix_bb_placement (loops, from)) | |
203 | continue; | |
204 | } | |
205 | ||
206 | /* Something has changed, insert predecessors into queue. */ | |
628f6a4e | 207 | FOR_EACH_EDGE (e, ei, from->preds) |
617b465c ZD |
208 | { |
209 | basic_block pred = e->src; | |
210 | struct loop *nca; | |
211 | ||
212 | if (TEST_BIT (in_queue, pred->index)) | |
213 | continue; | |
214 | ||
d329e058 | 215 | /* If it is subloop, then it either was not moved, or |
617b465c ZD |
216 | the path up the loop tree from base_loop do not contain |
217 | it. */ | |
218 | nca = find_common_loop (pred->loop_father, base_loop); | |
219 | if (pred->loop_father != base_loop | |
220 | && (nca == base_loop | |
221 | || nca != pred->loop_father)) | |
222 | pred = pred->loop_father->header; | |
223 | else if (!flow_loop_nested_p (from->loop_father, pred->loop_father)) | |
224 | { | |
225 | /* No point in processing it. */ | |
226 | continue; | |
227 | } | |
228 | ||
229 | if (TEST_BIT (in_queue, pred->index)) | |
230 | continue; | |
231 | ||
232 | /* Schedule the basic block. */ | |
233 | *qend = pred; | |
234 | qend++; | |
235 | if (qend == qtop) | |
236 | qend = queue; | |
237 | SET_BIT (in_queue, pred->index); | |
238 | } | |
239 | } | |
240 | free (in_queue); | |
241 | free (queue); | |
242 | } | |
243 | ||
35b07080 ZD |
244 | /* Basic block from has lost one or more of its predecessors, so it might |
245 | mo longer be part irreducible loop. Fix it and proceed recursively | |
246 | for its successors if needed. */ | |
247 | static void | |
d329e058 | 248 | fix_irreducible_loops (basic_block from) |
35b07080 ZD |
249 | { |
250 | basic_block bb; | |
251 | basic_block *stack; | |
252 | int stack_top; | |
253 | sbitmap on_stack; | |
254 | edge *edges, e; | |
255 | unsigned n_edges, i; | |
256 | ||
257 | if (!(from->flags & BB_IRREDUCIBLE_LOOP)) | |
258 | return; | |
259 | ||
260 | on_stack = sbitmap_alloc (last_basic_block); | |
261 | sbitmap_zero (on_stack); | |
262 | SET_BIT (on_stack, from->index); | |
263 | stack = xmalloc (from->loop_father->num_nodes * sizeof (basic_block)); | |
264 | stack[0] = from; | |
265 | stack_top = 1; | |
266 | ||
267 | while (stack_top) | |
268 | { | |
628f6a4e | 269 | edge_iterator ei; |
35b07080 ZD |
270 | bb = stack[--stack_top]; |
271 | RESET_BIT (on_stack, bb->index); | |
272 | ||
628f6a4e | 273 | FOR_EACH_EDGE (e, ei, bb->preds) |
35b07080 ZD |
274 | if (e->flags & EDGE_IRREDUCIBLE_LOOP) |
275 | break; | |
276 | if (e) | |
277 | continue; | |
278 | ||
279 | bb->flags &= ~BB_IRREDUCIBLE_LOOP; | |
280 | if (bb->loop_father->header == bb) | |
281 | edges = get_loop_exit_edges (bb->loop_father, &n_edges); | |
282 | else | |
283 | { | |
628f6a4e | 284 | n_edges = EDGE_COUNT (bb->succs); |
35b07080 | 285 | edges = xmalloc (n_edges * sizeof (edge)); |
628f6a4e BE |
286 | FOR_EACH_EDGE (e, ei, bb->succs) |
287 | edges[ei.index] = e; | |
35b07080 | 288 | } |
d329e058 | 289 | |
35b07080 | 290 | for (i = 0; i < n_edges; i++) |
3ed961a0 ZD |
291 | { |
292 | e = edges[i]; | |
35b07080 | 293 | |
3ed961a0 ZD |
294 | if (e->flags & EDGE_IRREDUCIBLE_LOOP) |
295 | { | |
296 | if (!flow_bb_inside_loop_p (from->loop_father, e->dest)) | |
297 | continue; | |
298 | ||
299 | e->flags &= ~EDGE_IRREDUCIBLE_LOOP; | |
300 | if (TEST_BIT (on_stack, e->dest->index)) | |
301 | continue; | |
35b07080 | 302 | |
3ed961a0 ZD |
303 | SET_BIT (on_stack, e->dest->index); |
304 | stack[stack_top++] = e->dest; | |
305 | } | |
306 | } | |
35b07080 ZD |
307 | free (edges); |
308 | } | |
309 | ||
310 | free (on_stack); | |
311 | free (stack); | |
312 | } | |
313 | ||
617b465c ZD |
314 | /* Removes path beginning at edge E, i.e. remove basic blocks dominated by E |
315 | and update loop structure stored in LOOPS and dominators. Return true if | |
316 | we were able to remove the path, false otherwise (and nothing is affected | |
317 | then). */ | |
318 | bool | |
d329e058 | 319 | remove_path (struct loops *loops, edge e) |
617b465c ZD |
320 | { |
321 | edge ae; | |
322 | basic_block *rem_bbs, *bord_bbs, *dom_bbs, from, bb; | |
323 | int i, nrem, n_bord_bbs, n_dom_bbs; | |
324 | sbitmap seen; | |
341c100f | 325 | bool deleted; |
617b465c | 326 | |
35b07080 ZD |
327 | if (!loop_delete_branch_edge (e, 0)) |
328 | return false; | |
329 | ||
330 | /* We need to check whether basic blocks are dominated by the edge | |
331 | e, but we only have basic block dominators. This is easy to | |
332 | fix -- when e->dest has exactly one predecessor, this corresponds | |
333 | to blocks dominated by e->dest, if not, split the edge. */ | |
c5cbcccf ZD |
334 | if (!single_pred_p (e->dest)) |
335 | e = single_pred_edge (loop_split_edge_with (e, NULL_RTX)); | |
35b07080 ZD |
336 | |
337 | /* It may happen that by removing path we remove one or more loops | |
338 | we belong to. In this case first unloop the loops, then proceed | |
339 | normally. We may assume that e->dest is not a header of any loop, | |
340 | as it now has exactly one predecessor. */ | |
341 | while (e->src->loop_father->outer | |
d47cc544 | 342 | && dominated_by_p (CDI_DOMINATORS, |
35b07080 ZD |
343 | e->src->loop_father->latch, e->dest)) |
344 | unloop (loops, e->src->loop_father); | |
d329e058 | 345 | |
35b07080 | 346 | /* Identify the path. */ |
d47cc544 | 347 | nrem = find_path (e, &rem_bbs); |
617b465c ZD |
348 | |
349 | n_bord_bbs = 0; | |
350 | bord_bbs = xcalloc (n_basic_blocks, sizeof (basic_block)); | |
351 | seen = sbitmap_alloc (last_basic_block); | |
352 | sbitmap_zero (seen); | |
353 | ||
354 | /* Find "border" hexes -- i.e. those with predecessor in removed path. */ | |
355 | for (i = 0; i < nrem; i++) | |
356 | SET_BIT (seen, rem_bbs[i]->index); | |
35b07080 | 357 | for (i = 0; i < nrem; i++) |
617b465c | 358 | { |
628f6a4e | 359 | edge_iterator ei; |
35b07080 | 360 | bb = rem_bbs[i]; |
628f6a4e | 361 | FOR_EACH_EDGE (ae, ei, rem_bbs[i]->succs) |
35b07080 ZD |
362 | if (ae->dest != EXIT_BLOCK_PTR && !TEST_BIT (seen, ae->dest->index)) |
363 | { | |
364 | SET_BIT (seen, ae->dest->index); | |
365 | bord_bbs[n_bord_bbs++] = ae->dest; | |
366 | } | |
617b465c | 367 | } |
617b465c ZD |
368 | |
369 | /* Remove the path. */ | |
370 | from = e->src; | |
341c100f NS |
371 | deleted = loop_delete_branch_edge (e, 1); |
372 | gcc_assert (deleted); | |
617b465c ZD |
373 | dom_bbs = xcalloc (n_basic_blocks, sizeof (basic_block)); |
374 | ||
375 | /* Cancel loops contained in the path. */ | |
376 | for (i = 0; i < nrem; i++) | |
377 | if (rem_bbs[i]->loop_father->header == rem_bbs[i]) | |
378 | cancel_loop_tree (loops, rem_bbs[i]->loop_father); | |
379 | ||
d47cc544 | 380 | remove_bbs (rem_bbs, nrem); |
617b465c ZD |
381 | free (rem_bbs); |
382 | ||
35b07080 | 383 | /* Find blocks whose dominators may be affected. */ |
617b465c ZD |
384 | n_dom_bbs = 0; |
385 | sbitmap_zero (seen); | |
386 | for (i = 0; i < n_bord_bbs; i++) | |
387 | { | |
d47cc544 | 388 | basic_block ldom; |
617b465c | 389 | |
d47cc544 | 390 | bb = get_immediate_dominator (CDI_DOMINATORS, bord_bbs[i]); |
617b465c ZD |
391 | if (TEST_BIT (seen, bb->index)) |
392 | continue; | |
393 | SET_BIT (seen, bb->index); | |
394 | ||
d47cc544 SB |
395 | for (ldom = first_dom_son (CDI_DOMINATORS, bb); |
396 | ldom; | |
397 | ldom = next_dom_son (CDI_DOMINATORS, ldom)) | |
398 | if (!dominated_by_p (CDI_DOMINATORS, from, ldom)) | |
399 | dom_bbs[n_dom_bbs++] = ldom; | |
617b465c ZD |
400 | } |
401 | ||
617b465c ZD |
402 | free (seen); |
403 | ||
404 | /* Recount dominators. */ | |
d47cc544 | 405 | iterate_fix_dominators (CDI_DOMINATORS, dom_bbs, n_dom_bbs); |
617b465c ZD |
406 | free (dom_bbs); |
407 | ||
35b07080 ZD |
408 | /* These blocks have lost some predecessor(s), thus their irreducible |
409 | status could be changed. */ | |
410 | for (i = 0; i < n_bord_bbs; i++) | |
411 | fix_irreducible_loops (bord_bbs[i]); | |
412 | free (bord_bbs); | |
413 | ||
617b465c ZD |
414 | /* Fix placements of basic blocks inside loops and the placement of |
415 | loops in the loop tree. */ | |
416 | fix_bb_placements (loops, from); | |
1548580c | 417 | fix_loop_placements (loops, from->loop_father); |
617b465c ZD |
418 | |
419 | return true; | |
420 | } | |
421 | ||
422 | /* Predicate for enumeration in add_loop. */ | |
423 | static bool | |
d329e058 | 424 | alp_enum_p (basic_block bb, void *alp_header) |
617b465c ZD |
425 | { |
426 | return bb != (basic_block) alp_header; | |
427 | } | |
428 | ||
429 | /* Given LOOP structure with filled header and latch, find the body of the | |
430 | corresponding loop and add it to LOOPS tree. */ | |
431 | static void | |
d329e058 | 432 | add_loop (struct loops *loops, struct loop *loop) |
617b465c ZD |
433 | { |
434 | basic_block *bbs; | |
435 | int i, n; | |
d329e058 | 436 | |
617b465c ZD |
437 | /* Add it to loop structure. */ |
438 | place_new_loop (loops, loop); | |
439 | loop->level = 1; | |
440 | ||
441 | /* Find its nodes. */ | |
442 | bbs = xcalloc (n_basic_blocks, sizeof (basic_block)); | |
443 | n = dfs_enumerate_from (loop->latch, 1, alp_enum_p, | |
444 | bbs, n_basic_blocks, loop->header); | |
445 | ||
446 | for (i = 0; i < n; i++) | |
447 | add_bb_to_loop (bbs[i], loop); | |
448 | add_bb_to_loop (loop->header, loop); | |
449 | ||
450 | free (bbs); | |
451 | } | |
452 | ||
617b465c ZD |
453 | /* Multiply all frequencies in LOOP by NUM/DEN. */ |
454 | static void | |
d329e058 | 455 | scale_loop_frequencies (struct loop *loop, int num, int den) |
617b465c ZD |
456 | { |
457 | basic_block *bbs; | |
458 | ||
459 | bbs = get_loop_body (loop); | |
33156717 | 460 | scale_bbs_frequencies_int (bbs, loop->num_nodes, num, den); |
617b465c ZD |
461 | free (bbs); |
462 | } | |
463 | ||
464 | /* Make area between HEADER_EDGE and LATCH_EDGE a loop by connecting | |
465 | latch to header and update loop tree stored in LOOPS and dominators | |
466 | accordingly. Everything between them plus LATCH_EDGE destination must | |
467 | be dominated by HEADER_EDGE destination, and back-reachable from | |
468 | LATCH_EDGE source. HEADER_EDGE is redirected to basic block SWITCH_BB, | |
5132abc2 KH |
469 | FALSE_EDGE of SWITCH_BB to original destination of HEADER_EDGE and |
470 | TRUE_EDGE of SWITCH_BB to original destination of LATCH_EDGE. | |
617b465c | 471 | Returns newly created loop. */ |
50654f6c | 472 | |
617b465c | 473 | struct loop * |
50654f6c | 474 | loopify (struct loops *loops, edge latch_edge, edge header_edge, |
5132abc2 KH |
475 | basic_block switch_bb, edge true_edge, edge false_edge, |
476 | bool redirect_all_edges) | |
617b465c ZD |
477 | { |
478 | basic_block succ_bb = latch_edge->dest; | |
479 | basic_block pred_bb = header_edge->src; | |
480 | basic_block *dom_bbs, *body; | |
d47cc544 | 481 | unsigned n_dom_bbs, i; |
617b465c ZD |
482 | sbitmap seen; |
483 | struct loop *loop = xcalloc (1, sizeof (struct loop)); | |
484 | struct loop *outer = succ_bb->loop_father->outer; | |
485 | int freq, prob, tot_prob; | |
486 | gcov_type cnt; | |
487 | edge e; | |
628f6a4e | 488 | edge_iterator ei; |
617b465c ZD |
489 | |
490 | loop->header = header_edge->dest; | |
491 | loop->latch = latch_edge->src; | |
492 | ||
493 | freq = EDGE_FREQUENCY (header_edge); | |
494 | cnt = header_edge->count; | |
628f6a4e BE |
495 | prob = EDGE_SUCC (switch_bb, 0)->probability; |
496 | tot_prob = prob + EDGE_SUCC (switch_bb, 1)->probability; | |
617b465c ZD |
497 | if (tot_prob == 0) |
498 | tot_prob = 1; | |
499 | ||
500 | /* Redirect edges. */ | |
501 | loop_redirect_edge (latch_edge, loop->header); | |
5132abc2 | 502 | loop_redirect_edge (true_edge, succ_bb); |
50654f6c | 503 | |
92fc4a2f ZD |
504 | /* During loop versioning, one of the switch_bb edge is already properly |
505 | set. Do not redirect it again unless redirect_all_edges is true. */ | |
506 | if (redirect_all_edges) | |
507 | { | |
508 | loop_redirect_edge (header_edge, switch_bb); | |
5132abc2 | 509 | loop_redirect_edge (false_edge, loop->header); |
92fc4a2f ZD |
510 | |
511 | /* Update dominators. */ | |
512 | set_immediate_dominator (CDI_DOMINATORS, switch_bb, pred_bb); | |
513 | set_immediate_dominator (CDI_DOMINATORS, loop->header, switch_bb); | |
514 | } | |
50654f6c | 515 | |
d47cc544 | 516 | set_immediate_dominator (CDI_DOMINATORS, succ_bb, switch_bb); |
617b465c ZD |
517 | |
518 | /* Compute new loop. */ | |
519 | add_loop (loops, loop); | |
520 | flow_loop_tree_node_add (outer, loop); | |
521 | ||
522 | /* Add switch_bb to appropriate loop. */ | |
523 | add_bb_to_loop (switch_bb, outer); | |
524 | ||
525 | /* Fix frequencies. */ | |
526 | switch_bb->frequency = freq; | |
527 | switch_bb->count = cnt; | |
628f6a4e | 528 | FOR_EACH_EDGE (e, ei, switch_bb->succs) |
617b465c ZD |
529 | e->count = (switch_bb->count * e->probability) / REG_BR_PROB_BASE; |
530 | scale_loop_frequencies (loop, prob, tot_prob); | |
531 | scale_loop_frequencies (succ_bb->loop_father, tot_prob - prob, tot_prob); | |
532 | ||
533 | /* Update dominators of blocks outside of LOOP. */ | |
534 | dom_bbs = xcalloc (n_basic_blocks, sizeof (basic_block)); | |
535 | n_dom_bbs = 0; | |
536 | seen = sbitmap_alloc (last_basic_block); | |
537 | sbitmap_zero (seen); | |
538 | body = get_loop_body (loop); | |
539 | ||
540 | for (i = 0; i < loop->num_nodes; i++) | |
541 | SET_BIT (seen, body[i]->index); | |
542 | ||
543 | for (i = 0; i < loop->num_nodes; i++) | |
544 | { | |
d47cc544 | 545 | basic_block ldom; |
617b465c | 546 | |
d47cc544 SB |
547 | for (ldom = first_dom_son (CDI_DOMINATORS, body[i]); |
548 | ldom; | |
549 | ldom = next_dom_son (CDI_DOMINATORS, ldom)) | |
550 | if (!TEST_BIT (seen, ldom->index)) | |
617b465c | 551 | { |
d47cc544 SB |
552 | SET_BIT (seen, ldom->index); |
553 | dom_bbs[n_dom_bbs++] = ldom; | |
617b465c | 554 | } |
617b465c ZD |
555 | } |
556 | ||
d47cc544 | 557 | iterate_fix_dominators (CDI_DOMINATORS, dom_bbs, n_dom_bbs); |
617b465c ZD |
558 | |
559 | free (body); | |
560 | free (seen); | |
561 | free (dom_bbs); | |
562 | ||
563 | return loop; | |
564 | } | |
565 | ||
35b07080 ZD |
566 | /* Remove the latch edge of a LOOP and update LOOPS tree to indicate that |
567 | the LOOP was removed. After this function, original loop latch will | |
568 | have no successor, which caller is expected to fix somehow. */ | |
25a6c68b | 569 | static void |
d329e058 | 570 | unloop (struct loops *loops, struct loop *loop) |
35b07080 ZD |
571 | { |
572 | basic_block *body; | |
573 | struct loop *ploop; | |
574 | unsigned i, n; | |
575 | basic_block latch = loop->latch; | |
576 | edge *edges; | |
577 | unsigned n_edges; | |
578 | ||
e0bb17a8 | 579 | /* This is relatively straightforward. The dominators are unchanged, as |
35b07080 ZD |
580 | loop header dominates loop latch, so the only thing we have to care of |
581 | is the placement of loops and basic blocks inside the loop tree. We | |
582 | move them all to the loop->outer, and then let fix_bb_placements do | |
583 | its work. */ | |
584 | ||
585 | body = get_loop_body (loop); | |
586 | edges = get_loop_exit_edges (loop, &n_edges); | |
587 | n = loop->num_nodes; | |
588 | for (i = 0; i < n; i++) | |
589 | if (body[i]->loop_father == loop) | |
590 | { | |
591 | remove_bb_from_loops (body[i]); | |
592 | add_bb_to_loop (body[i], loop->outer); | |
593 | } | |
594 | free(body); | |
595 | ||
596 | while (loop->inner) | |
597 | { | |
598 | ploop = loop->inner; | |
599 | flow_loop_tree_node_remove (ploop); | |
600 | flow_loop_tree_node_add (loop->outer, ploop); | |
601 | } | |
602 | ||
603 | /* Remove the loop and free its data. */ | |
604 | flow_loop_tree_node_remove (loop); | |
605 | loops->parray[loop->num] = NULL; | |
606 | flow_loop_free (loop); | |
607 | ||
c5cbcccf | 608 | remove_edge (single_succ_edge (latch)); |
35b07080 ZD |
609 | fix_bb_placements (loops, latch); |
610 | ||
611 | /* If the loop was inside an irreducible region, we would have to somehow | |
612 | update the irreducible marks inside its body. While it is certainly | |
613 | possible to do, it is a bit complicated and this situation should be | |
614 | very rare, so we just remark all loops in this case. */ | |
615 | for (i = 0; i < n_edges; i++) | |
616 | if (edges[i]->flags & EDGE_IRREDUCIBLE_LOOP) | |
617 | break; | |
618 | if (i != n_edges) | |
619 | mark_irreducible_loops (loops); | |
620 | free (edges); | |
621 | } | |
622 | ||
617b465c | 623 | /* Fix placement of LOOP inside loop tree, i.e. find the innermost superloop |
2067c116 | 624 | FATHER of LOOP such that all of the edges coming out of LOOP belong to |
617b465c ZD |
625 | FATHER, and set it as outer loop of LOOP. Return 1 if placement of |
626 | LOOP changed. */ | |
627 | int | |
d329e058 | 628 | fix_loop_placement (struct loop *loop) |
617b465c ZD |
629 | { |
630 | basic_block *body; | |
631 | unsigned i; | |
632 | edge e; | |
628f6a4e | 633 | edge_iterator ei; |
617b465c ZD |
634 | struct loop *father = loop->pred[0], *act; |
635 | ||
636 | body = get_loop_body (loop); | |
637 | for (i = 0; i < loop->num_nodes; i++) | |
628f6a4e | 638 | FOR_EACH_EDGE (e, ei, body[i]->succs) |
617b465c ZD |
639 | if (!flow_bb_inside_loop_p (loop, e->dest)) |
640 | { | |
641 | act = find_common_loop (loop, e->dest->loop_father); | |
642 | if (flow_loop_nested_p (father, act)) | |
643 | father = act; | |
644 | } | |
645 | free (body); | |
646 | ||
647 | if (father != loop->outer) | |
648 | { | |
649 | for (act = loop->outer; act != father; act = act->outer) | |
650 | act->num_nodes -= loop->num_nodes; | |
651 | flow_loop_tree_node_remove (loop); | |
652 | flow_loop_tree_node_add (father, loop); | |
653 | return 1; | |
654 | } | |
655 | return 0; | |
656 | } | |
657 | ||
658 | /* Fix placement of superloops of LOOP inside loop tree, i.e. ensure that | |
659 | condition stated in description of fix_loop_placement holds for them. | |
660 | It is used in case when we removed some edges coming out of LOOP, which | |
661 | may cause the right placement of LOOP inside loop tree to change. */ | |
662 | static void | |
1548580c | 663 | fix_loop_placements (struct loops *loops, struct loop *loop) |
617b465c ZD |
664 | { |
665 | struct loop *outer; | |
666 | ||
667 | while (loop->outer) | |
668 | { | |
669 | outer = loop->outer; | |
670 | if (!fix_loop_placement (loop)) | |
671 | break; | |
1548580c EB |
672 | |
673 | /* Changing the placement of a loop in the loop tree may alter the | |
674 | validity of condition 2) of the description of fix_bb_placement | |
675 | for its preheader, because the successor is the header and belongs | |
676 | to the loop. So call fix_bb_placements to fix up the placement | |
677 | of the preheader and (possibly) of its predecessors. */ | |
678 | fix_bb_placements (loops, loop_preheader_edge (loop)->src); | |
617b465c ZD |
679 | loop = outer; |
680 | } | |
681 | } | |
682 | ||
683 | /* Creates place for a new LOOP in LOOPS structure. */ | |
684 | static void | |
d329e058 | 685 | place_new_loop (struct loops *loops, struct loop *loop) |
617b465c ZD |
686 | { |
687 | loops->parray = | |
688 | xrealloc (loops->parray, (loops->num + 1) * sizeof (struct loop *)); | |
689 | loops->parray[loops->num] = loop; | |
690 | ||
691 | loop->num = loops->num++; | |
692 | } | |
693 | ||
694 | /* Copies copy of LOOP as subloop of TARGET loop, placing newly | |
695 | created loop into LOOPS structure. */ | |
f67d92e9 | 696 | struct loop * |
d329e058 | 697 | duplicate_loop (struct loops *loops, struct loop *loop, struct loop *target) |
617b465c ZD |
698 | { |
699 | struct loop *cloop; | |
700 | cloop = xcalloc (1, sizeof (struct loop)); | |
701 | place_new_loop (loops, cloop); | |
702 | ||
703 | /* Initialize copied loop. */ | |
704 | cloop->level = loop->level; | |
705 | ||
706 | /* Set it as copy of loop. */ | |
707 | loop->copy = cloop; | |
708 | ||
709 | /* Add it to target. */ | |
710 | flow_loop_tree_node_add (target, cloop); | |
711 | ||
712 | return cloop; | |
713 | } | |
714 | ||
715 | /* Copies structure of subloops of LOOP into TARGET loop, placing | |
716 | newly created loops into loop tree stored in LOOPS. */ | |
d329e058 AJ |
717 | static void |
718 | duplicate_subloops (struct loops *loops, struct loop *loop, struct loop *target) | |
617b465c ZD |
719 | { |
720 | struct loop *aloop, *cloop; | |
721 | ||
722 | for (aloop = loop->inner; aloop; aloop = aloop->next) | |
723 | { | |
724 | cloop = duplicate_loop (loops, aloop, target); | |
725 | duplicate_subloops (loops, aloop, cloop); | |
726 | } | |
727 | } | |
728 | ||
729 | /* Copies structure of subloops of N loops, stored in array COPIED_LOOPS, | |
730 | into TARGET loop, placing newly created loops into loop tree LOOPS. */ | |
d329e058 AJ |
731 | static void |
732 | copy_loops_to (struct loops *loops, struct loop **copied_loops, int n, struct loop *target) | |
617b465c ZD |
733 | { |
734 | struct loop *aloop; | |
735 | int i; | |
736 | ||
737 | for (i = 0; i < n; i++) | |
738 | { | |
739 | aloop = duplicate_loop (loops, copied_loops[i], target); | |
740 | duplicate_subloops (loops, copied_loops[i], aloop); | |
741 | } | |
742 | } | |
743 | ||
744 | /* Redirects edge E to basic block DEST. */ | |
745 | static void | |
d329e058 | 746 | loop_redirect_edge (edge e, basic_block dest) |
617b465c ZD |
747 | { |
748 | if (e->dest == dest) | |
749 | return; | |
750 | ||
9ee634e3 | 751 | redirect_edge_and_branch_force (e, dest); |
617b465c ZD |
752 | } |
753 | ||
35b07080 ZD |
754 | /* Deletes edge E from a branch if possible. Unless REALLY_DELETE is set, |
755 | just test whether it is possible to remove the edge. */ | |
617b465c | 756 | static bool |
d329e058 | 757 | loop_delete_branch_edge (edge e, int really_delete) |
617b465c ZD |
758 | { |
759 | basic_block src = e->src; | |
341c100f | 760 | basic_block newdest; |
35b07080 ZD |
761 | int irr; |
762 | edge snd; | |
617b465c | 763 | |
628f6a4e | 764 | gcc_assert (EDGE_COUNT (src->succs) > 1); |
341c100f NS |
765 | |
766 | /* Cannot handle more than two exit edges. */ | |
628f6a4e | 767 | if (EDGE_COUNT (src->succs) > 2) |
341c100f NS |
768 | return false; |
769 | /* And it must be just a simple branch. */ | |
770 | if (!any_condjump_p (BB_END (src))) | |
771 | return false; | |
35b07080 | 772 | |
628f6a4e | 773 | snd = e == EDGE_SUCC (src, 0) ? EDGE_SUCC (src, 1) : EDGE_SUCC (src, 0); |
341c100f NS |
774 | newdest = snd->dest; |
775 | if (newdest == EXIT_BLOCK_PTR) | |
776 | return false; | |
d329e058 | 777 | |
341c100f NS |
778 | /* Hopefully the above conditions should suffice. */ |
779 | if (!really_delete) | |
780 | return true; | |
35b07080 | 781 | |
341c100f NS |
782 | /* Redirecting behaves wrongly wrto this flag. */ |
783 | irr = snd->flags & EDGE_IRREDUCIBLE_LOOP; | |
617b465c | 784 | |
341c100f NS |
785 | if (!redirect_edge_and_branch (e, newdest)) |
786 | return false; | |
c5cbcccf ZD |
787 | single_succ_edge (src)->flags &= ~EDGE_IRREDUCIBLE_LOOP; |
788 | single_succ_edge (src)->flags |= irr; | |
341c100f NS |
789 | |
790 | return true; | |
617b465c ZD |
791 | } |
792 | ||
617b465c ZD |
793 | /* Check whether LOOP's body can be duplicated. */ |
794 | bool | |
d329e058 | 795 | can_duplicate_loop_p (struct loop *loop) |
617b465c | 796 | { |
8d28e87d ZD |
797 | int ret; |
798 | basic_block *bbs = get_loop_body (loop); | |
617b465c | 799 | |
8d28e87d | 800 | ret = can_copy_bbs_p (bbs, loop->num_nodes); |
617b465c | 801 | free (bbs); |
8d28e87d ZD |
802 | |
803 | return ret; | |
617b465c ZD |
804 | } |
805 | ||
82b85a85 ZD |
806 | /* The NBBS blocks in BBS will get duplicated and the copies will be placed |
807 | to LOOP. Update the single_exit information in superloops of LOOP. */ | |
808 | ||
25a6c68b | 809 | static void |
82b85a85 ZD |
810 | update_single_exits_after_duplication (basic_block *bbs, unsigned nbbs, |
811 | struct loop *loop) | |
812 | { | |
813 | unsigned i; | |
814 | ||
815 | for (i = 0; i < nbbs; i++) | |
816 | bbs[i]->rbi->duplicated = 1; | |
817 | ||
818 | for (; loop->outer; loop = loop->outer) | |
819 | { | |
820 | if (!loop->single_exit) | |
821 | continue; | |
822 | ||
823 | if (loop->single_exit->src->rbi->duplicated) | |
824 | loop->single_exit = NULL; | |
825 | } | |
826 | ||
827 | for (i = 0; i < nbbs; i++) | |
828 | bbs[i]->rbi->duplicated = 0; | |
829 | } | |
830 | ||
8d28e87d ZD |
831 | /* Duplicates body of LOOP to given edge E NDUPL times. Takes care of updating |
832 | LOOPS structure and dominators. E's destination must be LOOP header for | |
833 | this to work, i.e. it must be entry or latch edge of this loop; these are | |
834 | unique, as the loops must have preheaders for this function to work | |
835 | correctly (in case E is latch, the function unrolls the loop, if E is entry | |
836 | edge, it peels the loop). Store edges created by copying ORIG edge from | |
837 | copies corresponding to set bits in WONT_EXIT bitmap (bit 0 corresponds to | |
838 | original LOOP body, the other copies are numbered in order given by control | |
839 | flow through them) into TO_REMOVE array. Returns false if duplication is | |
840 | impossible. */ | |
1cb7dfc3 | 841 | bool |
d329e058 AJ |
842 | duplicate_loop_to_header_edge (struct loop *loop, edge e, struct loops *loops, |
843 | unsigned int ndupl, sbitmap wont_exit, | |
844 | edge orig, edge *to_remove, | |
845 | unsigned int *n_to_remove, int flags) | |
617b465c ZD |
846 | { |
847 | struct loop *target, *aloop; | |
848 | struct loop **orig_loops; | |
849 | unsigned n_orig_loops; | |
850 | basic_block header = loop->header, latch = loop->latch; | |
851 | basic_block *new_bbs, *bbs, *first_active; | |
852 | basic_block new_bb, bb, first_active_latch = NULL; | |
8d28e87d ZD |
853 | edge ae, latch_edge; |
854 | edge spec_edges[2], new_spec_edges[2]; | |
855 | #define SE_LATCH 0 | |
856 | #define SE_ORIG 1 | |
617b465c ZD |
857 | unsigned i, j, n; |
858 | int is_latch = (latch == e->src); | |
859 | int scale_act = 0, *scale_step = NULL, scale_main = 0; | |
860 | int p, freq_in, freq_le, freq_out_orig; | |
861 | int prob_pass_thru, prob_pass_wont_exit, prob_pass_main; | |
862 | int add_irreducible_flag; | |
863 | ||
341c100f NS |
864 | gcc_assert (e->dest == loop->header); |
865 | gcc_assert (ndupl > 0); | |
617b465c ZD |
866 | |
867 | if (orig) | |
868 | { | |
869 | /* Orig must be edge out of the loop. */ | |
341c100f NS |
870 | gcc_assert (flow_bb_inside_loop_p (loop, orig->src)); |
871 | gcc_assert (!flow_bb_inside_loop_p (loop, orig->dest)); | |
617b465c ZD |
872 | } |
873 | ||
874 | bbs = get_loop_body (loop); | |
875 | ||
876 | /* Check whether duplication is possible. */ | |
8d28e87d | 877 | if (!can_copy_bbs_p (bbs, loop->num_nodes)) |
617b465c | 878 | { |
8d28e87d ZD |
879 | free (bbs); |
880 | return false; | |
617b465c | 881 | } |
8d28e87d | 882 | new_bbs = xmalloc (sizeof (basic_block) * loop->num_nodes); |
617b465c | 883 | |
8d28e87d ZD |
884 | /* In case we are doing loop peeling and the loop is in the middle of |
885 | irreducible region, the peeled copies will be inside it too. */ | |
886 | add_irreducible_flag = e->flags & EDGE_IRREDUCIBLE_LOOP; | |
341c100f | 887 | gcc_assert (!is_latch || !add_irreducible_flag); |
617b465c ZD |
888 | |
889 | /* Find edge from latch. */ | |
890 | latch_edge = loop_latch_edge (loop); | |
891 | ||
892 | if (flags & DLTHE_FLAG_UPDATE_FREQ) | |
893 | { | |
894 | /* Calculate coefficients by that we have to scale frequencies | |
895 | of duplicated loop bodies. */ | |
896 | freq_in = header->frequency; | |
897 | freq_le = EDGE_FREQUENCY (latch_edge); | |
898 | if (freq_in == 0) | |
899 | freq_in = 1; | |
900 | if (freq_in < freq_le) | |
901 | freq_in = freq_le; | |
902 | freq_out_orig = orig ? EDGE_FREQUENCY (orig) : freq_in - freq_le; | |
903 | if (freq_out_orig > freq_in - freq_le) | |
904 | freq_out_orig = freq_in - freq_le; | |
905 | prob_pass_thru = RDIV (REG_BR_PROB_BASE * freq_le, freq_in); | |
906 | prob_pass_wont_exit = | |
907 | RDIV (REG_BR_PROB_BASE * (freq_le + freq_out_orig), freq_in); | |
908 | ||
909 | scale_step = xmalloc (ndupl * sizeof (int)); | |
910 | ||
911 | for (i = 1; i <= ndupl; i++) | |
d329e058 | 912 | scale_step[i - 1] = TEST_BIT (wont_exit, i) |
617b465c ZD |
913 | ? prob_pass_wont_exit |
914 | : prob_pass_thru; | |
915 | ||
916 | if (is_latch) | |
917 | { | |
918 | prob_pass_main = TEST_BIT (wont_exit, 0) | |
919 | ? prob_pass_wont_exit | |
920 | : prob_pass_thru; | |
921 | p = prob_pass_main; | |
922 | scale_main = REG_BR_PROB_BASE; | |
923 | for (i = 0; i < ndupl; i++) | |
924 | { | |
925 | scale_main += p; | |
926 | p = RDIV (p * scale_step[i], REG_BR_PROB_BASE); | |
927 | } | |
928 | scale_main = RDIV (REG_BR_PROB_BASE * REG_BR_PROB_BASE, scale_main); | |
929 | scale_act = RDIV (scale_main * prob_pass_main, REG_BR_PROB_BASE); | |
930 | } | |
931 | else | |
932 | { | |
933 | scale_main = REG_BR_PROB_BASE; | |
934 | for (i = 0; i < ndupl; i++) | |
935 | scale_main = RDIV (scale_main * scale_step[i], REG_BR_PROB_BASE); | |
936 | scale_act = REG_BR_PROB_BASE - prob_pass_thru; | |
937 | } | |
938 | for (i = 0; i < ndupl; i++) | |
341c100f NS |
939 | gcc_assert (scale_step[i] >= 0 && scale_step[i] <= REG_BR_PROB_BASE); |
940 | gcc_assert (scale_main >= 0 && scale_main <= REG_BR_PROB_BASE | |
941 | && scale_act >= 0 && scale_act <= REG_BR_PROB_BASE); | |
617b465c ZD |
942 | } |
943 | ||
944 | /* Loop the new bbs will belong to. */ | |
8d28e87d | 945 | target = e->src->loop_father; |
617b465c ZD |
946 | |
947 | /* Original loops. */ | |
948 | n_orig_loops = 0; | |
949 | for (aloop = loop->inner; aloop; aloop = aloop->next) | |
950 | n_orig_loops++; | |
951 | orig_loops = xcalloc (n_orig_loops, sizeof (struct loop *)); | |
952 | for (aloop = loop->inner, i = 0; aloop; aloop = aloop->next, i++) | |
953 | orig_loops[i] = aloop; | |
954 | ||
955 | loop->copy = target; | |
d329e058 | 956 | |
617b465c ZD |
957 | n = loop->num_nodes; |
958 | ||
8d28e87d | 959 | first_active = xmalloc (n * sizeof (basic_block)); |
617b465c ZD |
960 | if (is_latch) |
961 | { | |
962 | memcpy (first_active, bbs, n * sizeof (basic_block)); | |
963 | first_active_latch = latch; | |
964 | } | |
965 | ||
82b85a85 ZD |
966 | /* Update the information about single exits. */ |
967 | if (loops->state & LOOPS_HAVE_MARKED_SINGLE_EXITS) | |
968 | update_single_exits_after_duplication (bbs, n, target); | |
969 | ||
8d28e87d ZD |
970 | /* Record exit edge in original loop body. */ |
971 | if (orig && TEST_BIT (wont_exit, 0)) | |
972 | to_remove[(*n_to_remove)++] = orig; | |
973 | ||
974 | spec_edges[SE_ORIG] = orig; | |
975 | spec_edges[SE_LATCH] = latch_edge; | |
d329e058 | 976 | |
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. */ | |
d47cc544 | 983 | copy_bbs (bbs, n, new_bbs, spec_edges, 2, new_spec_edges, loop); |
8d28e87d | 984 | |
113d659a ZD |
985 | for (i = 0; i < n; i++) |
986 | new_bbs[i]->rbi->copy_number = j + 1; | |
987 | ||
84d45ad1 ZD |
988 | /* Note whether the blocks and edges belong to an irreducible loop. */ |
989 | if (add_irreducible_flag) | |
990 | { | |
991 | for (i = 0; i < n; i++) | |
992 | new_bbs[i]->rbi->duplicated = 1; | |
993 | for (i = 0; i < n; i++) | |
994 | { | |
628f6a4e | 995 | edge_iterator ei; |
84d45ad1 ZD |
996 | new_bb = new_bbs[i]; |
997 | if (new_bb->loop_father == target) | |
998 | new_bb->flags |= BB_IRREDUCIBLE_LOOP; | |
999 | ||
628f6a4e | 1000 | FOR_EACH_EDGE (ae, ei, new_bb->succs) |
84d45ad1 ZD |
1001 | if (ae->dest->rbi->duplicated |
1002 | && (ae->src->loop_father == target | |
1003 | || ae->dest->loop_father == target)) | |
1004 | ae->flags |= EDGE_IRREDUCIBLE_LOOP; | |
1005 | } | |
1006 | for (i = 0; i < n; i++) | |
1007 | new_bbs[i]->rbi->duplicated = 0; | |
1008 | } | |
1009 | ||
8d28e87d | 1010 | /* Redirect the special edges. */ |
617b465c | 1011 | if (is_latch) |
8d28e87d ZD |
1012 | { |
1013 | redirect_edge_and_branch_force (latch_edge, new_bbs[0]); | |
1014 | redirect_edge_and_branch_force (new_spec_edges[SE_LATCH], | |
1015 | loop->header); | |
d47cc544 | 1016 | set_immediate_dominator (CDI_DOMINATORS, new_bbs[0], latch); |
8d28e87d ZD |
1017 | latch = loop->latch = new_bbs[1]; |
1018 | e = latch_edge = new_spec_edges[SE_LATCH]; | |
1019 | } | |
1020 | else | |
1021 | { | |
1022 | redirect_edge_and_branch_force (new_spec_edges[SE_LATCH], | |
1023 | loop->header); | |
1024 | redirect_edge_and_branch_force (e, new_bbs[0]); | |
d47cc544 | 1025 | set_immediate_dominator (CDI_DOMINATORS, new_bbs[0], e->src); |
8d28e87d ZD |
1026 | e = new_spec_edges[SE_LATCH]; |
1027 | } | |
617b465c | 1028 | |
8d28e87d ZD |
1029 | /* Record exit edge in this copy. */ |
1030 | if (orig && TEST_BIT (wont_exit, j + 1)) | |
1031 | to_remove[(*n_to_remove)++] = new_spec_edges[SE_ORIG]; | |
d329e058 | 1032 | |
8d28e87d ZD |
1033 | /* Record the first copy in the control flow order if it is not |
1034 | the original loop (i.e. in case of peeling). */ | |
617b465c ZD |
1035 | if (!first_active_latch) |
1036 | { | |
1037 | memcpy (first_active, new_bbs, n * sizeof (basic_block)); | |
8d28e87d | 1038 | first_active_latch = new_bbs[1]; |
617b465c | 1039 | } |
d329e058 | 1040 | |
8d28e87d ZD |
1041 | /* Set counts and frequencies. */ |
1042 | if (flags & DLTHE_FLAG_UPDATE_FREQ) | |
617b465c | 1043 | { |
33156717 | 1044 | scale_bbs_frequencies_int (new_bbs, n, scale_act, REG_BR_PROB_BASE); |
8d28e87d | 1045 | scale_act = RDIV (scale_act * scale_step[j], REG_BR_PROB_BASE); |
617b465c ZD |
1046 | } |
1047 | } | |
8d28e87d ZD |
1048 | free (new_bbs); |
1049 | free (orig_loops); | |
1050 | ||
1051 | /* Update the original loop. */ | |
1052 | if (!is_latch) | |
d47cc544 | 1053 | set_immediate_dominator (CDI_DOMINATORS, e->dest, e->src); |
617b465c ZD |
1054 | if (flags & DLTHE_FLAG_UPDATE_FREQ) |
1055 | { | |
33156717 | 1056 | scale_bbs_frequencies_int (bbs, n, scale_main, REG_BR_PROB_BASE); |
617b465c ZD |
1057 | free (scale_step); |
1058 | } | |
617b465c | 1059 | |
8d28e87d | 1060 | /* Update dominators of outer blocks if affected. */ |
617b465c ZD |
1061 | for (i = 0; i < n; i++) |
1062 | { | |
1063 | basic_block dominated, dom_bb, *dom_bbs; | |
1064 | int n_dom_bbs,j; | |
1065 | ||
1066 | bb = bbs[i]; | |
113d659a ZD |
1067 | bb->rbi->copy_number = 0; |
1068 | ||
d47cc544 | 1069 | n_dom_bbs = get_dominated_by (CDI_DOMINATORS, bb, &dom_bbs); |
617b465c ZD |
1070 | for (j = 0; j < n_dom_bbs; j++) |
1071 | { | |
1072 | dominated = dom_bbs[j]; | |
1073 | if (flow_bb_inside_loop_p (loop, dominated)) | |
1074 | continue; | |
1075 | dom_bb = nearest_common_dominator ( | |
d47cc544 SB |
1076 | CDI_DOMINATORS, first_active[i], first_active_latch); |
1077 | set_immediate_dominator (CDI_DOMINATORS, dominated, dom_bb); | |
617b465c ZD |
1078 | } |
1079 | free (dom_bbs); | |
1080 | } | |
1081 | free (first_active); | |
1082 | ||
1083 | free (bbs); | |
1084 | ||
1085 | return true; | |
1086 | } | |
1087 | ||
f470c378 ZD |
1088 | /* A callback for make_forwarder block, to redirect all edges except for |
1089 | MFB_KJ_EDGE to the entry part. E is the edge for that we should decide | |
1090 | whether to redirect it. */ | |
1091 | ||
1092 | static edge mfb_kj_edge; | |
1093 | static bool | |
1094 | mfb_keep_just (edge e) | |
1095 | { | |
1096 | return e != mfb_kj_edge; | |
1097 | } | |
1098 | ||
1099 | /* A callback for make_forwarder block, to update data structures for a basic | |
1100 | block JUMP created by redirecting an edge (only the latch edge is being | |
1101 | redirected). */ | |
1102 | ||
1103 | static void | |
1104 | mfb_update_loops (basic_block jump) | |
1105 | { | |
c5cbcccf | 1106 | struct loop *loop = single_succ (jump)->loop_father; |
f470c378 ZD |
1107 | |
1108 | if (dom_computed[CDI_DOMINATORS]) | |
c5cbcccf | 1109 | set_immediate_dominator (CDI_DOMINATORS, jump, single_pred (jump)); |
f470c378 ZD |
1110 | add_bb_to_loop (jump, loop); |
1111 | loop->latch = jump; | |
1112 | } | |
1113 | ||
3d436d2a ZD |
1114 | /* Creates a pre-header for a LOOP. Returns newly created block. Unless |
1115 | CP_SIMPLE_PREHEADERS is set in FLAGS, we only force LOOP to have single | |
1116 | entry; otherwise we also force preheader block to have only one successor. | |
f470c378 ZD |
1117 | The function also updates dominators. */ |
1118 | ||
3d436d2a | 1119 | static basic_block |
d47cc544 | 1120 | create_preheader (struct loop *loop, int flags) |
3d436d2a ZD |
1121 | { |
1122 | edge e, fallthru; | |
1123 | basic_block dummy; | |
3d436d2a ZD |
1124 | struct loop *cloop, *ploop; |
1125 | int nentry = 0; | |
f470c378 | 1126 | bool irred = false; |
c15bc84b EB |
1127 | bool latch_edge_was_fallthru; |
1128 | edge one_succ_pred = 0; | |
628f6a4e | 1129 | edge_iterator ei; |
3d436d2a ZD |
1130 | |
1131 | cloop = loop->outer; | |
1132 | ||
628f6a4e | 1133 | FOR_EACH_EDGE (e, ei, loop->header->preds) |
3d436d2a ZD |
1134 | { |
1135 | if (e->src == loop->latch) | |
1136 | continue; | |
f470c378 | 1137 | irred |= (e->flags & EDGE_IRREDUCIBLE_LOOP) != 0; |
3d436d2a | 1138 | nentry++; |
c5cbcccf | 1139 | if (single_succ_p (e->src)) |
c15bc84b | 1140 | one_succ_pred = e; |
3d436d2a | 1141 | } |
341c100f | 1142 | gcc_assert (nentry); |
3d436d2a ZD |
1143 | if (nentry == 1) |
1144 | { | |
b057216d KH |
1145 | /* Get an edge that is different from the one from loop->latch |
1146 | to loop->header. */ | |
1147 | e = EDGE_PRED (loop->header, | |
1148 | EDGE_PRED (loop->header, 0)->src == loop->latch); | |
628f6a4e | 1149 | |
c5cbcccf | 1150 | if (!(flags & CP_SIMPLE_PREHEADERS) || single_succ_p (e->src)) |
3d436d2a ZD |
1151 | return NULL; |
1152 | } | |
1153 | ||
f470c378 | 1154 | mfb_kj_edge = loop_latch_edge (loop); |
c15bc84b | 1155 | latch_edge_was_fallthru = (mfb_kj_edge->flags & EDGE_FALLTHRU) != 0; |
f470c378 ZD |
1156 | fallthru = make_forwarder_block (loop->header, mfb_keep_just, |
1157 | mfb_update_loops); | |
3d436d2a ZD |
1158 | dummy = fallthru->src; |
1159 | loop->header = fallthru->dest; | |
1160 | ||
1161 | /* The header could be a latch of some superloop(s); due to design of | |
1162 | split_block, it would now move to fallthru->dest. */ | |
1163 | for (ploop = loop; ploop; ploop = ploop->outer) | |
1164 | if (ploop->latch == dummy) | |
1165 | ploop->latch = fallthru->dest; | |
1166 | ||
c15bc84b EB |
1167 | /* Try to be clever in placing the newly created preheader. The idea is to |
1168 | avoid breaking any "fallthruness" relationship between blocks. | |
1169 | ||
1170 | The preheader was created just before the header and all incoming edges | |
1171 | to the header were redirected to the preheader, except the latch edge. | |
1172 | So the only problematic case is when this latch edge was a fallthru | |
1173 | edge: it is not anymore after the preheader creation so we have broken | |
1174 | the fallthruness. We're therefore going to look for a better place. */ | |
1175 | if (latch_edge_was_fallthru) | |
1176 | { | |
1177 | if (one_succ_pred) | |
1178 | e = one_succ_pred; | |
1179 | else | |
1180 | e = EDGE_PRED (dummy, 0); | |
1181 | ||
1182 | move_block_after (dummy, e->src); | |
1183 | } | |
f470c378 ZD |
1184 | |
1185 | loop->header->loop_father = loop; | |
1186 | add_bb_to_loop (dummy, cloop); | |
3d436d2a | 1187 | |
f470c378 | 1188 | if (irred) |
3d436d2a | 1189 | { |
f470c378 | 1190 | dummy->flags |= BB_IRREDUCIBLE_LOOP; |
c5cbcccf | 1191 | single_succ_edge (dummy)->flags |= EDGE_IRREDUCIBLE_LOOP; |
3d436d2a ZD |
1192 | } |
1193 | ||
c263766c RH |
1194 | if (dump_file) |
1195 | fprintf (dump_file, "Created preheader block for loop %i\n", | |
3d436d2a ZD |
1196 | loop->num); |
1197 | ||
1198 | return dummy; | |
1199 | } | |
1200 | ||
617b465c ZD |
1201 | /* Create preheaders for each loop from loop tree stored in LOOPS; for meaning |
1202 | of FLAGS see create_preheader. */ | |
3d436d2a | 1203 | void |
d329e058 | 1204 | create_preheaders (struct loops *loops, int flags) |
3d436d2a ZD |
1205 | { |
1206 | unsigned i; | |
1207 | for (i = 1; i < loops->num; i++) | |
d47cc544 | 1208 | create_preheader (loops->parray[i], flags); |
3d436d2a ZD |
1209 | loops->state |= LOOPS_HAVE_PREHEADERS; |
1210 | } | |
1211 | ||
617b465c ZD |
1212 | /* Forces all loop latches of loops from loop tree LOOPS to have only single |
1213 | successor. */ | |
3d436d2a | 1214 | void |
d329e058 | 1215 | force_single_succ_latches (struct loops *loops) |
3d436d2a ZD |
1216 | { |
1217 | unsigned i; | |
1218 | struct loop *loop; | |
1219 | edge e; | |
1220 | ||
1221 | for (i = 1; i < loops->num; i++) | |
1222 | { | |
1223 | loop = loops->parray[i]; | |
c5cbcccf | 1224 | if (loop->latch != loop->header && single_succ_p (loop->latch)) |
3d436d2a | 1225 | continue; |
d329e058 | 1226 | |
9ff3d2de | 1227 | e = find_edge (loop->latch, loop->header); |
bc810602 | 1228 | |
d47cc544 | 1229 | loop_split_edge_with (e, NULL_RTX); |
3d436d2a ZD |
1230 | } |
1231 | loops->state |= LOOPS_HAVE_SIMPLE_LATCHES; | |
1232 | } | |
1233 | ||
617b465c ZD |
1234 | /* A quite stupid function to put INSNS on edge E. They are supposed to form |
1235 | just one basic block. Jumps in INSNS are not handled, so cfg do not have to | |
1236 | be ok after this function. The created block is placed on correct place | |
1237 | in LOOPS structure and its dominator is set. */ | |
3d436d2a | 1238 | basic_block |
d47cc544 | 1239 | loop_split_edge_with (edge e, rtx insns) |
3d436d2a ZD |
1240 | { |
1241 | basic_block src, dest, new_bb; | |
1242 | struct loop *loop_c; | |
d329e058 | 1243 | |
3d436d2a ZD |
1244 | src = e->src; |
1245 | dest = e->dest; | |
1246 | ||
1247 | loop_c = find_common_loop (src->loop_father, dest->loop_father); | |
1248 | ||
1249 | /* Create basic block for it. */ | |
1250 | ||
bc35512f | 1251 | new_bb = split_edge (e); |
3d436d2a | 1252 | add_bb_to_loop (new_bb, loop_c); |
a746fd8c | 1253 | new_bb->flags |= (insns ? BB_SUPERBLOCK : 0); |
3d436d2a | 1254 | |
3d436d2a | 1255 | if (insns) |
a813c111 | 1256 | emit_insn_after (insns, BB_END (new_bb)); |
3d436d2a | 1257 | |
3d436d2a ZD |
1258 | if (dest->loop_father->latch == src) |
1259 | dest->loop_father->latch = new_bb; | |
d329e058 | 1260 | |
3d436d2a ZD |
1261 | return new_bb; |
1262 | } | |
6de9cd9a DN |
1263 | |
1264 | /* Uses the natural loop discovery to recreate loop notes. */ | |
1265 | void | |
1266 | create_loop_notes (void) | |
1267 | { | |
1268 | rtx insn, head, end; | |
1269 | struct loops loops; | |
1270 | struct loop *loop; | |
1271 | basic_block *first, *last, bb, pbb; | |
1272 | struct loop **stack, **top; | |
1273 | ||
1274 | #ifdef ENABLE_CHECKING | |
1275 | /* Verify that there really are no loop notes. */ | |
1276 | for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) | |
341c100f NS |
1277 | gcc_assert (!NOTE_P (insn) || |
1278 | NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG); | |
6de9cd9a DN |
1279 | #endif |
1280 | ||
70388d94 | 1281 | flow_loops_find (&loops); |
6de9cd9a DN |
1282 | free_dominance_info (CDI_DOMINATORS); |
1283 | if (loops.num > 1) | |
1284 | { | |
1285 | last = xcalloc (loops.num, sizeof (basic_block)); | |
1286 | ||
1287 | FOR_EACH_BB (bb) | |
1288 | { | |
1289 | for (loop = bb->loop_father; loop->outer; loop = loop->outer) | |
1290 | last[loop->num] = bb; | |
1291 | } | |
1292 | ||
1293 | first = xcalloc (loops.num, sizeof (basic_block)); | |
1294 | stack = xcalloc (loops.num, sizeof (struct loop *)); | |
1295 | top = stack; | |
1296 | ||
1297 | FOR_EACH_BB (bb) | |
1298 | { | |
1299 | for (loop = bb->loop_father; loop->outer; loop = loop->outer) | |
1300 | { | |
1301 | if (!first[loop->num]) | |
1302 | { | |
1303 | *top++ = loop; | |
1304 | first[loop->num] = bb; | |
1305 | } | |
1306 | ||
1307 | if (bb == last[loop->num]) | |
1308 | { | |
1309 | /* Prevent loops from overlapping. */ | |
1310 | while (*--top != loop) | |
1311 | last[(*top)->num] = EXIT_BLOCK_PTR; | |
1312 | ||
1313 | /* If loop starts with jump into it, place the note in | |
1314 | front of the jump. */ | |
1315 | insn = PREV_INSN (BB_HEAD (first[loop->num])); | |
1316 | if (insn | |
4b4bf941 | 1317 | && BARRIER_P (insn)) |
6de9cd9a DN |
1318 | insn = PREV_INSN (insn); |
1319 | ||
1320 | if (insn | |
4b4bf941 | 1321 | && JUMP_P (insn) |
6de9cd9a DN |
1322 | && any_uncondjump_p (insn) |
1323 | && onlyjump_p (insn)) | |
1324 | { | |
1325 | pbb = BLOCK_FOR_INSN (insn); | |
c5cbcccf | 1326 | gcc_assert (pbb && single_succ_p (pbb)); |
6de9cd9a | 1327 | |
c5cbcccf | 1328 | if (!flow_bb_inside_loop_p (loop, single_succ (pbb))) |
6de9cd9a DN |
1329 | insn = BB_HEAD (first[loop->num]); |
1330 | } | |
1331 | else | |
1332 | insn = BB_HEAD (first[loop->num]); | |
1333 | ||
1334 | head = BB_HEAD (first[loop->num]); | |
1335 | emit_note_before (NOTE_INSN_LOOP_BEG, insn); | |
1336 | BB_HEAD (first[loop->num]) = head; | |
1337 | ||
1338 | /* Position the note correctly wrto barrier. */ | |
1339 | insn = BB_END (last[loop->num]); | |
1340 | if (NEXT_INSN (insn) | |
4b4bf941 | 1341 | && BARRIER_P (NEXT_INSN (insn))) |
6de9cd9a DN |
1342 | insn = NEXT_INSN (insn); |
1343 | ||
1344 | end = BB_END (last[loop->num]); | |
1345 | emit_note_after (NOTE_INSN_LOOP_END, insn); | |
1346 | BB_END (last[loop->num]) = end; | |
1347 | } | |
1348 | } | |
1349 | } | |
1350 | ||
1351 | free (first); | |
1352 | free (last); | |
1353 | free (stack); | |
1354 | } | |
1355 | flow_loops_free (&loops); | |
1356 | } | |
2b271002 | 1357 | |
1cb7dfc3 MH |
1358 | /* This function is called from loop_version. It splits the entry edge |
1359 | of the loop we want to version, adds the versioning condition, and | |
1360 | adjust the edges to the two versions of the loop appropriately. | |
1361 | e is an incoming edge. Returns the basic block containing the | |
1362 | condition. | |
1363 | ||
1364 | --- edge e ---- > [second_head] | |
1365 | ||
1366 | Split it and insert new conditional expression and adjust edges. | |
1367 | ||
1368 | --- edge e ---> [cond expr] ---> [first_head] | |
1369 | | | |
1370 | +---------> [second_head] | |
1371 | */ | |
1372 | ||
1373 | static basic_block | |
1374 | lv_adjust_loop_entry_edge (basic_block first_head, | |
1375 | basic_block second_head, | |
1376 | edge e, | |
1377 | tree cond_expr) | |
1378 | { | |
1379 | basic_block new_head = NULL; | |
1380 | edge e1; | |
1381 | ||
1382 | gcc_assert (e->dest == second_head); | |
1383 | ||
1384 | /* Split edge 'e'. This will create a new basic block, where we can | |
1385 | insert conditional expr. */ | |
1386 | new_head = split_edge (e); | |
1387 | ||
1388 | ||
1389 | lv_add_condition_to_bb (first_head, second_head, new_head, | |
1390 | cond_expr); | |
1391 | ||
1392 | e1 = make_edge (new_head, first_head, EDGE_TRUE_VALUE); | |
1393 | set_immediate_dominator (CDI_DOMINATORS, first_head, new_head); | |
1394 | set_immediate_dominator (CDI_DOMINATORS, second_head, new_head); | |
1395 | ||
1396 | /* Adjust loop header phi nodes. */ | |
1397 | lv_adjust_loop_header_phi (first_head, second_head, new_head, e1); | |
1398 | ||
1399 | return new_head; | |
1400 | } | |
1401 | ||
1402 | /* Main entry point for Loop Versioning transformation. | |
1403 | ||
1404 | This transformation given a condition and a loop, creates | |
1405 | -if (condition) { loop_copy1 } else { loop_copy2 }, | |
1406 | where loop_copy1 is the loop transformed in one way, and loop_copy2 | |
1407 | is the loop transformed in another way (or unchanged). 'condition' | |
1408 | may be a run time test for things that were not resolved by static | |
1409 | analysis (overlapping ranges (anti-aliasing), alignment, etc.). */ | |
1410 | ||
1411 | struct loop * | |
1412 | loop_version (struct loops *loops, struct loop * loop, | |
1413 | void *cond_expr, basic_block *condition_bb) | |
1414 | { | |
1415 | basic_block first_head, second_head; | |
1416 | edge entry, latch_edge, exit, true_edge, false_edge; | |
1417 | int irred_flag; | |
1418 | struct loop *nloop; | |
1419 | ||
1420 | /* CHECKME: Loop versioning does not handle nested loop at this point. */ | |
1421 | if (loop->inner) | |
1422 | return NULL; | |
1423 | ||
1424 | /* Record entry and latch edges for the loop */ | |
1425 | entry = loop_preheader_edge (loop); | |
1426 | irred_flag = entry->flags & EDGE_IRREDUCIBLE_LOOP; | |
1427 | entry->flags &= ~EDGE_IRREDUCIBLE_LOOP; | |
1428 | ||
1429 | /* Note down head of loop as first_head. */ | |
1430 | first_head = entry->dest; | |
1431 | ||
1432 | /* Duplicate loop. */ | |
1433 | if (!cfg_hook_duplicate_loop_to_header_edge (loop, entry, loops, 1, | |
1434 | NULL, NULL, NULL, NULL, 0)) | |
1435 | return NULL; | |
1436 | ||
1437 | /* After duplication entry edge now points to new loop head block. | |
1438 | Note down new head as second_head. */ | |
1439 | second_head = entry->dest; | |
1440 | ||
1441 | /* Split loop entry edge and insert new block with cond expr. */ | |
1442 | *condition_bb = lv_adjust_loop_entry_edge (first_head, second_head, | |
1443 | entry, cond_expr); | |
1444 | if (!*condition_bb) | |
1445 | { | |
1446 | entry->flags |= irred_flag; | |
1447 | return NULL; | |
1448 | } | |
1449 | ||
1450 | latch_edge = single_succ_edge (loop->latch->rbi->copy); | |
1451 | ||
1452 | extract_cond_bb_edges (*condition_bb, &true_edge, &false_edge); | |
1453 | nloop = loopify (loops, | |
1454 | latch_edge, | |
1455 | single_pred_edge (loop->header->rbi->copy), | |
1456 | *condition_bb, true_edge, false_edge, | |
1457 | false /* Do not redirect all edges. */); | |
1458 | ||
1459 | exit = loop->single_exit; | |
1460 | if (exit) | |
1461 | nloop->single_exit = find_edge (exit->src->rbi->copy, exit->dest); | |
1462 | ||
1463 | /* loopify redirected latch_edge. Update its PENDING_STMTS. */ | |
1464 | lv_flush_pending_stmts (latch_edge); | |
1465 | ||
1466 | /* loopify redirected condition_bb's succ edge. Update its PENDING_STMTS. */ | |
1467 | extract_cond_bb_edges (*condition_bb, &true_edge, &false_edge); | |
1468 | lv_flush_pending_stmts (false_edge); | |
1469 | /* Adjust irreducible flag. */ | |
1470 | if (irred_flag) | |
1471 | { | |
1472 | (*condition_bb)->flags |= BB_IRREDUCIBLE_LOOP; | |
1473 | loop_preheader_edge (loop)->flags |= EDGE_IRREDUCIBLE_LOOP; | |
1474 | loop_preheader_edge (nloop)->flags |= EDGE_IRREDUCIBLE_LOOP; | |
1475 | single_pred_edge ((*condition_bb))->flags |= EDGE_IRREDUCIBLE_LOOP; | |
1476 | } | |
1477 | ||
1478 | /* At this point condition_bb is loop predheader with two successors, | |
1479 | first_head and second_head. Make sure that loop predheader has only | |
1480 | one successor. */ | |
1481 | loop_split_edge_with (loop_preheader_edge (loop), NULL); | |
1482 | loop_split_edge_with (loop_preheader_edge (nloop), NULL); | |
1483 | ||
1484 | return nloop; | |
1485 | } | |
1486 | ||
2b271002 ZD |
1487 | /* The structure of LOOPS might have changed. Some loops might get removed |
1488 | (and their headers and latches were set to NULL), loop exists might get | |
1489 | removed (thus the loop nesting may be wrong), and some blocks and edges | |
1490 | were changed (so the information about bb --> loop mapping does not have | |
1491 | to be correct). But still for the remaining loops the header dominates | |
1492 | the latch, and loops did not get new subloobs (new loops might possibly | |
1493 | get created, but we are not interested in them). Fix up the mess. | |
1494 | ||
1495 | If CHANGED_BBS is not NULL, basic blocks whose loop has changed are | |
1496 | marked in it. */ | |
1497 | ||
1498 | void | |
1499 | fix_loop_structure (struct loops *loops, bitmap changed_bbs) | |
1500 | { | |
1501 | basic_block bb; | |
1502 | struct loop *loop, *ploop; | |
1503 | unsigned i; | |
1504 | ||
1505 | /* Remove the old bb -> loop mapping. */ | |
1506 | FOR_EACH_BB (bb) | |
1507 | { | |
1508 | bb->aux = (void *) (size_t) bb->loop_father->depth; | |
1509 | bb->loop_father = loops->tree_root; | |
1510 | } | |
1511 | ||
1512 | /* Remove the dead loops from structures. */ | |
1513 | loops->tree_root->num_nodes = n_basic_blocks + 2; | |
1514 | for (i = 1; i < loops->num; i++) | |
1515 | { | |
1516 | loop = loops->parray[i]; | |
1517 | if (!loop) | |
1518 | continue; | |
1519 | ||
1520 | loop->num_nodes = 0; | |
1521 | if (loop->header) | |
1522 | continue; | |
1523 | ||
1524 | while (loop->inner) | |
1525 | { | |
1526 | ploop = loop->inner; | |
1527 | flow_loop_tree_node_remove (ploop); | |
1528 | flow_loop_tree_node_add (loop->outer, ploop); | |
1529 | } | |
1530 | ||
1531 | /* Remove the loop and free its data. */ | |
1532 | flow_loop_tree_node_remove (loop); | |
1533 | loops->parray[loop->num] = NULL; | |
1534 | flow_loop_free (loop); | |
1535 | } | |
1536 | ||
1537 | /* Rescan the bodies of loops, starting from the outermost. */ | |
1538 | loop = loops->tree_root; | |
1539 | while (1) | |
1540 | { | |
1541 | if (loop->inner) | |
1542 | loop = loop->inner; | |
1543 | else | |
1544 | { | |
1545 | while (!loop->next | |
1546 | && loop != loops->tree_root) | |
1547 | loop = loop->outer; | |
1548 | if (loop == loops->tree_root) | |
1549 | break; | |
1550 | ||
1551 | loop = loop->next; | |
1552 | } | |
1553 | ||
1554 | loop->num_nodes = flow_loop_nodes_find (loop->header, loop); | |
1555 | } | |
1556 | ||
1557 | /* Now fix the loop nesting. */ | |
1558 | for (i = 1; i < loops->num; i++) | |
1559 | { | |
1560 | loop = loops->parray[i]; | |
1561 | if (!loop) | |
1562 | continue; | |
1563 | ||
1564 | bb = loop_preheader_edge (loop)->src; | |
1565 | if (bb->loop_father != loop->outer) | |
1566 | { | |
1567 | flow_loop_tree_node_remove (loop); | |
1568 | flow_loop_tree_node_add (bb->loop_father, loop); | |
1569 | } | |
1570 | } | |
1571 | ||
1572 | /* Mark the blocks whose loop has changed. */ | |
1573 | FOR_EACH_BB (bb) | |
1574 | { | |
1575 | if (changed_bbs | |
1576 | && (void *) (size_t) bb->loop_father->depth != bb->aux) | |
1577 | bitmap_set_bit (changed_bbs, bb->index); | |
1578 | ||
1579 | bb->aux = NULL; | |
1580 | } | |
1581 | ||
1582 | mark_single_exit_loops (loops); | |
1583 | mark_irreducible_loops (loops); | |
1584 | } |