]>
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" | |
31 | #include "output.h" | |
32 | ||
d329e058 AJ |
33 | static void duplicate_subloops (struct loops *, struct loop *, struct loop *); |
34 | static void copy_loops_to (struct loops *, struct loop **, int, | |
35 | struct loop *); | |
36 | static void loop_redirect_edge (edge, basic_block); | |
37 | static bool loop_delete_branch_edge (edge, int); | |
d47cc544 | 38 | static void remove_bbs (basic_block *, int); |
d329e058 | 39 | static bool rpe_enum_p (basic_block, void *); |
d47cc544 | 40 | static int find_path (edge, basic_block **); |
d329e058 AJ |
41 | static bool alp_enum_p (basic_block, void *); |
42 | static void add_loop (struct loops *, struct loop *); | |
1548580c | 43 | static void fix_loop_placements (struct loops *, struct loop *); |
d329e058 AJ |
44 | static bool fix_bb_placement (struct loops *, basic_block); |
45 | static void fix_bb_placements (struct loops *, basic_block); | |
46 | static void place_new_loop (struct loops *, struct loop *); | |
47 | static void scale_loop_frequencies (struct loop *, int, int); | |
48 | static void scale_bbs_frequencies (basic_block *, int, 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 | ||
e0bb17a8 | 453 | /* Multiply all frequencies of basic blocks in array BBS of length NBBS |
617b465c ZD |
454 | by NUM/DEN. */ |
455 | static void | |
d329e058 | 456 | scale_bbs_frequencies (basic_block *bbs, int nbbs, int num, int den) |
617b465c ZD |
457 | { |
458 | int i; | |
459 | edge e; | |
460 | ||
461 | for (i = 0; i < nbbs; i++) | |
462 | { | |
628f6a4e | 463 | edge_iterator ei; |
617b465c | 464 | bbs[i]->frequency = (bbs[i]->frequency * num) / den; |
bade3a00 | 465 | bbs[i]->count = RDIV (bbs[i]->count * num, den); |
628f6a4e | 466 | FOR_EACH_EDGE (e, ei, bbs[i]->succs) |
617b465c ZD |
467 | e->count = (e->count * num) /den; |
468 | } | |
469 | } | |
470 | ||
471 | /* Multiply all frequencies in LOOP by NUM/DEN. */ | |
472 | static void | |
d329e058 | 473 | scale_loop_frequencies (struct loop *loop, int num, int den) |
617b465c ZD |
474 | { |
475 | basic_block *bbs; | |
476 | ||
477 | bbs = get_loop_body (loop); | |
478 | scale_bbs_frequencies (bbs, loop->num_nodes, num, den); | |
479 | free (bbs); | |
480 | } | |
481 | ||
482 | /* Make area between HEADER_EDGE and LATCH_EDGE a loop by connecting | |
483 | latch to header and update loop tree stored in LOOPS and dominators | |
484 | accordingly. Everything between them plus LATCH_EDGE destination must | |
485 | be dominated by HEADER_EDGE destination, and back-reachable from | |
486 | LATCH_EDGE source. HEADER_EDGE is redirected to basic block SWITCH_BB, | |
5132abc2 KH |
487 | FALSE_EDGE of SWITCH_BB to original destination of HEADER_EDGE and |
488 | TRUE_EDGE of SWITCH_BB to original destination of LATCH_EDGE. | |
617b465c | 489 | Returns newly created loop. */ |
50654f6c | 490 | |
617b465c | 491 | struct loop * |
50654f6c | 492 | loopify (struct loops *loops, edge latch_edge, edge header_edge, |
5132abc2 KH |
493 | basic_block switch_bb, edge true_edge, edge false_edge, |
494 | bool redirect_all_edges) | |
617b465c ZD |
495 | { |
496 | basic_block succ_bb = latch_edge->dest; | |
497 | basic_block pred_bb = header_edge->src; | |
498 | basic_block *dom_bbs, *body; | |
d47cc544 | 499 | unsigned n_dom_bbs, i; |
617b465c ZD |
500 | sbitmap seen; |
501 | struct loop *loop = xcalloc (1, sizeof (struct loop)); | |
502 | struct loop *outer = succ_bb->loop_father->outer; | |
503 | int freq, prob, tot_prob; | |
504 | gcov_type cnt; | |
505 | edge e; | |
628f6a4e | 506 | edge_iterator ei; |
617b465c ZD |
507 | |
508 | loop->header = header_edge->dest; | |
509 | loop->latch = latch_edge->src; | |
510 | ||
511 | freq = EDGE_FREQUENCY (header_edge); | |
512 | cnt = header_edge->count; | |
628f6a4e BE |
513 | prob = EDGE_SUCC (switch_bb, 0)->probability; |
514 | tot_prob = prob + EDGE_SUCC (switch_bb, 1)->probability; | |
617b465c ZD |
515 | if (tot_prob == 0) |
516 | tot_prob = 1; | |
517 | ||
518 | /* Redirect edges. */ | |
519 | loop_redirect_edge (latch_edge, loop->header); | |
5132abc2 | 520 | loop_redirect_edge (true_edge, succ_bb); |
50654f6c | 521 | |
92fc4a2f ZD |
522 | /* During loop versioning, one of the switch_bb edge is already properly |
523 | set. Do not redirect it again unless redirect_all_edges is true. */ | |
524 | if (redirect_all_edges) | |
525 | { | |
526 | loop_redirect_edge (header_edge, switch_bb); | |
5132abc2 | 527 | loop_redirect_edge (false_edge, loop->header); |
92fc4a2f ZD |
528 | |
529 | /* Update dominators. */ | |
530 | set_immediate_dominator (CDI_DOMINATORS, switch_bb, pred_bb); | |
531 | set_immediate_dominator (CDI_DOMINATORS, loop->header, switch_bb); | |
532 | } | |
50654f6c | 533 | |
d47cc544 | 534 | set_immediate_dominator (CDI_DOMINATORS, succ_bb, switch_bb); |
617b465c ZD |
535 | |
536 | /* Compute new loop. */ | |
537 | add_loop (loops, loop); | |
538 | flow_loop_tree_node_add (outer, loop); | |
539 | ||
540 | /* Add switch_bb to appropriate loop. */ | |
541 | add_bb_to_loop (switch_bb, outer); | |
542 | ||
543 | /* Fix frequencies. */ | |
544 | switch_bb->frequency = freq; | |
545 | switch_bb->count = cnt; | |
628f6a4e | 546 | FOR_EACH_EDGE (e, ei, switch_bb->succs) |
617b465c ZD |
547 | e->count = (switch_bb->count * e->probability) / REG_BR_PROB_BASE; |
548 | scale_loop_frequencies (loop, prob, tot_prob); | |
549 | scale_loop_frequencies (succ_bb->loop_father, tot_prob - prob, tot_prob); | |
550 | ||
551 | /* Update dominators of blocks outside of LOOP. */ | |
552 | dom_bbs = xcalloc (n_basic_blocks, sizeof (basic_block)); | |
553 | n_dom_bbs = 0; | |
554 | seen = sbitmap_alloc (last_basic_block); | |
555 | sbitmap_zero (seen); | |
556 | body = get_loop_body (loop); | |
557 | ||
558 | for (i = 0; i < loop->num_nodes; i++) | |
559 | SET_BIT (seen, body[i]->index); | |
560 | ||
561 | for (i = 0; i < loop->num_nodes; i++) | |
562 | { | |
d47cc544 | 563 | basic_block ldom; |
617b465c | 564 | |
d47cc544 SB |
565 | for (ldom = first_dom_son (CDI_DOMINATORS, body[i]); |
566 | ldom; | |
567 | ldom = next_dom_son (CDI_DOMINATORS, ldom)) | |
568 | if (!TEST_BIT (seen, ldom->index)) | |
617b465c | 569 | { |
d47cc544 SB |
570 | SET_BIT (seen, ldom->index); |
571 | dom_bbs[n_dom_bbs++] = ldom; | |
617b465c | 572 | } |
617b465c ZD |
573 | } |
574 | ||
d47cc544 | 575 | iterate_fix_dominators (CDI_DOMINATORS, dom_bbs, n_dom_bbs); |
617b465c ZD |
576 | |
577 | free (body); | |
578 | free (seen); | |
579 | free (dom_bbs); | |
580 | ||
581 | return loop; | |
582 | } | |
583 | ||
35b07080 ZD |
584 | /* Remove the latch edge of a LOOP and update LOOPS tree to indicate that |
585 | the LOOP was removed. After this function, original loop latch will | |
586 | have no successor, which caller is expected to fix somehow. */ | |
25a6c68b | 587 | static void |
d329e058 | 588 | unloop (struct loops *loops, struct loop *loop) |
35b07080 ZD |
589 | { |
590 | basic_block *body; | |
591 | struct loop *ploop; | |
592 | unsigned i, n; | |
593 | basic_block latch = loop->latch; | |
594 | edge *edges; | |
595 | unsigned n_edges; | |
596 | ||
e0bb17a8 | 597 | /* This is relatively straightforward. The dominators are unchanged, as |
35b07080 ZD |
598 | loop header dominates loop latch, so the only thing we have to care of |
599 | is the placement of loops and basic blocks inside the loop tree. We | |
600 | move them all to the loop->outer, and then let fix_bb_placements do | |
601 | its work. */ | |
602 | ||
603 | body = get_loop_body (loop); | |
604 | edges = get_loop_exit_edges (loop, &n_edges); | |
605 | n = loop->num_nodes; | |
606 | for (i = 0; i < n; i++) | |
607 | if (body[i]->loop_father == loop) | |
608 | { | |
609 | remove_bb_from_loops (body[i]); | |
610 | add_bb_to_loop (body[i], loop->outer); | |
611 | } | |
612 | free(body); | |
613 | ||
614 | while (loop->inner) | |
615 | { | |
616 | ploop = loop->inner; | |
617 | flow_loop_tree_node_remove (ploop); | |
618 | flow_loop_tree_node_add (loop->outer, ploop); | |
619 | } | |
620 | ||
621 | /* Remove the loop and free its data. */ | |
622 | flow_loop_tree_node_remove (loop); | |
623 | loops->parray[loop->num] = NULL; | |
624 | flow_loop_free (loop); | |
625 | ||
c5cbcccf | 626 | remove_edge (single_succ_edge (latch)); |
35b07080 ZD |
627 | fix_bb_placements (loops, latch); |
628 | ||
629 | /* If the loop was inside an irreducible region, we would have to somehow | |
630 | update the irreducible marks inside its body. While it is certainly | |
631 | possible to do, it is a bit complicated and this situation should be | |
632 | very rare, so we just remark all loops in this case. */ | |
633 | for (i = 0; i < n_edges; i++) | |
634 | if (edges[i]->flags & EDGE_IRREDUCIBLE_LOOP) | |
635 | break; | |
636 | if (i != n_edges) | |
637 | mark_irreducible_loops (loops); | |
638 | free (edges); | |
639 | } | |
640 | ||
617b465c | 641 | /* Fix placement of LOOP inside loop tree, i.e. find the innermost superloop |
2067c116 | 642 | FATHER of LOOP such that all of the edges coming out of LOOP belong to |
617b465c ZD |
643 | FATHER, and set it as outer loop of LOOP. Return 1 if placement of |
644 | LOOP changed. */ | |
645 | int | |
d329e058 | 646 | fix_loop_placement (struct loop *loop) |
617b465c ZD |
647 | { |
648 | basic_block *body; | |
649 | unsigned i; | |
650 | edge e; | |
628f6a4e | 651 | edge_iterator ei; |
617b465c ZD |
652 | struct loop *father = loop->pred[0], *act; |
653 | ||
654 | body = get_loop_body (loop); | |
655 | for (i = 0; i < loop->num_nodes; i++) | |
628f6a4e | 656 | FOR_EACH_EDGE (e, ei, body[i]->succs) |
617b465c ZD |
657 | if (!flow_bb_inside_loop_p (loop, e->dest)) |
658 | { | |
659 | act = find_common_loop (loop, e->dest->loop_father); | |
660 | if (flow_loop_nested_p (father, act)) | |
661 | father = act; | |
662 | } | |
663 | free (body); | |
664 | ||
665 | if (father != loop->outer) | |
666 | { | |
667 | for (act = loop->outer; act != father; act = act->outer) | |
668 | act->num_nodes -= loop->num_nodes; | |
669 | flow_loop_tree_node_remove (loop); | |
670 | flow_loop_tree_node_add (father, loop); | |
671 | return 1; | |
672 | } | |
673 | return 0; | |
674 | } | |
675 | ||
676 | /* Fix placement of superloops of LOOP inside loop tree, i.e. ensure that | |
677 | condition stated in description of fix_loop_placement holds for them. | |
678 | It is used in case when we removed some edges coming out of LOOP, which | |
679 | may cause the right placement of LOOP inside loop tree to change. */ | |
680 | static void | |
1548580c | 681 | fix_loop_placements (struct loops *loops, struct loop *loop) |
617b465c ZD |
682 | { |
683 | struct loop *outer; | |
684 | ||
685 | while (loop->outer) | |
686 | { | |
687 | outer = loop->outer; | |
688 | if (!fix_loop_placement (loop)) | |
689 | break; | |
1548580c EB |
690 | |
691 | /* Changing the placement of a loop in the loop tree may alter the | |
692 | validity of condition 2) of the description of fix_bb_placement | |
693 | for its preheader, because the successor is the header and belongs | |
694 | to the loop. So call fix_bb_placements to fix up the placement | |
695 | of the preheader and (possibly) of its predecessors. */ | |
696 | fix_bb_placements (loops, loop_preheader_edge (loop)->src); | |
617b465c ZD |
697 | loop = outer; |
698 | } | |
699 | } | |
700 | ||
701 | /* Creates place for a new LOOP in LOOPS structure. */ | |
702 | static void | |
d329e058 | 703 | place_new_loop (struct loops *loops, struct loop *loop) |
617b465c ZD |
704 | { |
705 | loops->parray = | |
706 | xrealloc (loops->parray, (loops->num + 1) * sizeof (struct loop *)); | |
707 | loops->parray[loops->num] = loop; | |
708 | ||
709 | loop->num = loops->num++; | |
710 | } | |
711 | ||
712 | /* Copies copy of LOOP as subloop of TARGET loop, placing newly | |
713 | created loop into LOOPS structure. */ | |
f67d92e9 | 714 | struct loop * |
d329e058 | 715 | duplicate_loop (struct loops *loops, struct loop *loop, struct loop *target) |
617b465c ZD |
716 | { |
717 | struct loop *cloop; | |
718 | cloop = xcalloc (1, sizeof (struct loop)); | |
719 | place_new_loop (loops, cloop); | |
720 | ||
721 | /* Initialize copied loop. */ | |
722 | cloop->level = loop->level; | |
723 | ||
724 | /* Set it as copy of loop. */ | |
725 | loop->copy = cloop; | |
726 | ||
727 | /* Add it to target. */ | |
728 | flow_loop_tree_node_add (target, cloop); | |
729 | ||
730 | return cloop; | |
731 | } | |
732 | ||
733 | /* Copies structure of subloops of LOOP into TARGET loop, placing | |
734 | newly created loops into loop tree stored in LOOPS. */ | |
d329e058 AJ |
735 | static void |
736 | duplicate_subloops (struct loops *loops, struct loop *loop, struct loop *target) | |
617b465c ZD |
737 | { |
738 | struct loop *aloop, *cloop; | |
739 | ||
740 | for (aloop = loop->inner; aloop; aloop = aloop->next) | |
741 | { | |
742 | cloop = duplicate_loop (loops, aloop, target); | |
743 | duplicate_subloops (loops, aloop, cloop); | |
744 | } | |
745 | } | |
746 | ||
747 | /* Copies structure of subloops of N loops, stored in array COPIED_LOOPS, | |
748 | into TARGET loop, placing newly created loops into loop tree LOOPS. */ | |
d329e058 AJ |
749 | static void |
750 | copy_loops_to (struct loops *loops, struct loop **copied_loops, int n, struct loop *target) | |
617b465c ZD |
751 | { |
752 | struct loop *aloop; | |
753 | int i; | |
754 | ||
755 | for (i = 0; i < n; i++) | |
756 | { | |
757 | aloop = duplicate_loop (loops, copied_loops[i], target); | |
758 | duplicate_subloops (loops, copied_loops[i], aloop); | |
759 | } | |
760 | } | |
761 | ||
762 | /* Redirects edge E to basic block DEST. */ | |
763 | static void | |
d329e058 | 764 | loop_redirect_edge (edge e, basic_block dest) |
617b465c ZD |
765 | { |
766 | if (e->dest == dest) | |
767 | return; | |
768 | ||
9ee634e3 | 769 | redirect_edge_and_branch_force (e, dest); |
617b465c ZD |
770 | } |
771 | ||
35b07080 ZD |
772 | /* Deletes edge E from a branch if possible. Unless REALLY_DELETE is set, |
773 | just test whether it is possible to remove the edge. */ | |
617b465c | 774 | static bool |
d329e058 | 775 | loop_delete_branch_edge (edge e, int really_delete) |
617b465c ZD |
776 | { |
777 | basic_block src = e->src; | |
341c100f | 778 | basic_block newdest; |
35b07080 ZD |
779 | int irr; |
780 | edge snd; | |
617b465c | 781 | |
628f6a4e | 782 | gcc_assert (EDGE_COUNT (src->succs) > 1); |
341c100f NS |
783 | |
784 | /* Cannot handle more than two exit edges. */ | |
628f6a4e | 785 | if (EDGE_COUNT (src->succs) > 2) |
341c100f NS |
786 | return false; |
787 | /* And it must be just a simple branch. */ | |
788 | if (!any_condjump_p (BB_END (src))) | |
789 | return false; | |
35b07080 | 790 | |
628f6a4e | 791 | snd = e == EDGE_SUCC (src, 0) ? EDGE_SUCC (src, 1) : EDGE_SUCC (src, 0); |
341c100f NS |
792 | newdest = snd->dest; |
793 | if (newdest == EXIT_BLOCK_PTR) | |
794 | return false; | |
d329e058 | 795 | |
341c100f NS |
796 | /* Hopefully the above conditions should suffice. */ |
797 | if (!really_delete) | |
798 | return true; | |
35b07080 | 799 | |
341c100f NS |
800 | /* Redirecting behaves wrongly wrto this flag. */ |
801 | irr = snd->flags & EDGE_IRREDUCIBLE_LOOP; | |
617b465c | 802 | |
341c100f NS |
803 | if (!redirect_edge_and_branch (e, newdest)) |
804 | return false; | |
c5cbcccf ZD |
805 | single_succ_edge (src)->flags &= ~EDGE_IRREDUCIBLE_LOOP; |
806 | single_succ_edge (src)->flags |= irr; | |
341c100f NS |
807 | |
808 | return true; | |
617b465c ZD |
809 | } |
810 | ||
617b465c ZD |
811 | /* Check whether LOOP's body can be duplicated. */ |
812 | bool | |
d329e058 | 813 | can_duplicate_loop_p (struct loop *loop) |
617b465c | 814 | { |
8d28e87d ZD |
815 | int ret; |
816 | basic_block *bbs = get_loop_body (loop); | |
617b465c | 817 | |
8d28e87d | 818 | ret = can_copy_bbs_p (bbs, loop->num_nodes); |
617b465c | 819 | free (bbs); |
8d28e87d ZD |
820 | |
821 | return ret; | |
617b465c ZD |
822 | } |
823 | ||
82b85a85 ZD |
824 | /* The NBBS blocks in BBS will get duplicated and the copies will be placed |
825 | to LOOP. Update the single_exit information in superloops of LOOP. */ | |
826 | ||
25a6c68b | 827 | static void |
82b85a85 ZD |
828 | update_single_exits_after_duplication (basic_block *bbs, unsigned nbbs, |
829 | struct loop *loop) | |
830 | { | |
831 | unsigned i; | |
832 | ||
833 | for (i = 0; i < nbbs; i++) | |
834 | bbs[i]->rbi->duplicated = 1; | |
835 | ||
836 | for (; loop->outer; loop = loop->outer) | |
837 | { | |
838 | if (!loop->single_exit) | |
839 | continue; | |
840 | ||
841 | if (loop->single_exit->src->rbi->duplicated) | |
842 | loop->single_exit = NULL; | |
843 | } | |
844 | ||
845 | for (i = 0; i < nbbs; i++) | |
846 | bbs[i]->rbi->duplicated = 0; | |
847 | } | |
848 | ||
8d28e87d ZD |
849 | /* Duplicates body of LOOP to given edge E NDUPL times. Takes care of updating |
850 | LOOPS structure and dominators. E's destination must be LOOP header for | |
851 | this to work, i.e. it must be entry or latch edge of this loop; these are | |
852 | unique, as the loops must have preheaders for this function to work | |
853 | correctly (in case E is latch, the function unrolls the loop, if E is entry | |
854 | edge, it peels the loop). Store edges created by copying ORIG edge from | |
855 | copies corresponding to set bits in WONT_EXIT bitmap (bit 0 corresponds to | |
856 | original LOOP body, the other copies are numbered in order given by control | |
857 | flow through them) into TO_REMOVE array. Returns false if duplication is | |
858 | impossible. */ | |
617b465c | 859 | int |
d329e058 AJ |
860 | duplicate_loop_to_header_edge (struct loop *loop, edge e, struct loops *loops, |
861 | unsigned int ndupl, sbitmap wont_exit, | |
862 | edge orig, edge *to_remove, | |
863 | unsigned int *n_to_remove, int flags) | |
617b465c ZD |
864 | { |
865 | struct loop *target, *aloop; | |
866 | struct loop **orig_loops; | |
867 | unsigned n_orig_loops; | |
868 | basic_block header = loop->header, latch = loop->latch; | |
869 | basic_block *new_bbs, *bbs, *first_active; | |
870 | basic_block new_bb, bb, first_active_latch = NULL; | |
8d28e87d ZD |
871 | edge ae, latch_edge; |
872 | edge spec_edges[2], new_spec_edges[2]; | |
873 | #define SE_LATCH 0 | |
874 | #define SE_ORIG 1 | |
617b465c ZD |
875 | unsigned i, j, n; |
876 | int is_latch = (latch == e->src); | |
877 | int scale_act = 0, *scale_step = NULL, scale_main = 0; | |
878 | int p, freq_in, freq_le, freq_out_orig; | |
879 | int prob_pass_thru, prob_pass_wont_exit, prob_pass_main; | |
880 | int add_irreducible_flag; | |
881 | ||
341c100f NS |
882 | gcc_assert (e->dest == loop->header); |
883 | gcc_assert (ndupl > 0); | |
617b465c ZD |
884 | |
885 | if (orig) | |
886 | { | |
887 | /* Orig must be edge out of the loop. */ | |
341c100f NS |
888 | gcc_assert (flow_bb_inside_loop_p (loop, orig->src)); |
889 | gcc_assert (!flow_bb_inside_loop_p (loop, orig->dest)); | |
617b465c ZD |
890 | } |
891 | ||
892 | bbs = get_loop_body (loop); | |
893 | ||
894 | /* Check whether duplication is possible. */ | |
8d28e87d | 895 | if (!can_copy_bbs_p (bbs, loop->num_nodes)) |
617b465c | 896 | { |
8d28e87d ZD |
897 | free (bbs); |
898 | return false; | |
617b465c | 899 | } |
8d28e87d | 900 | new_bbs = xmalloc (sizeof (basic_block) * loop->num_nodes); |
617b465c | 901 | |
8d28e87d ZD |
902 | /* In case we are doing loop peeling and the loop is in the middle of |
903 | irreducible region, the peeled copies will be inside it too. */ | |
904 | add_irreducible_flag = e->flags & EDGE_IRREDUCIBLE_LOOP; | |
341c100f | 905 | gcc_assert (!is_latch || !add_irreducible_flag); |
617b465c ZD |
906 | |
907 | /* Find edge from latch. */ | |
908 | latch_edge = loop_latch_edge (loop); | |
909 | ||
910 | if (flags & DLTHE_FLAG_UPDATE_FREQ) | |
911 | { | |
912 | /* Calculate coefficients by that we have to scale frequencies | |
913 | of duplicated loop bodies. */ | |
914 | freq_in = header->frequency; | |
915 | freq_le = EDGE_FREQUENCY (latch_edge); | |
916 | if (freq_in == 0) | |
917 | freq_in = 1; | |
918 | if (freq_in < freq_le) | |
919 | freq_in = freq_le; | |
920 | freq_out_orig = orig ? EDGE_FREQUENCY (orig) : freq_in - freq_le; | |
921 | if (freq_out_orig > freq_in - freq_le) | |
922 | freq_out_orig = freq_in - freq_le; | |
923 | prob_pass_thru = RDIV (REG_BR_PROB_BASE * freq_le, freq_in); | |
924 | prob_pass_wont_exit = | |
925 | RDIV (REG_BR_PROB_BASE * (freq_le + freq_out_orig), freq_in); | |
926 | ||
927 | scale_step = xmalloc (ndupl * sizeof (int)); | |
928 | ||
929 | for (i = 1; i <= ndupl; i++) | |
d329e058 | 930 | scale_step[i - 1] = TEST_BIT (wont_exit, i) |
617b465c ZD |
931 | ? prob_pass_wont_exit |
932 | : prob_pass_thru; | |
933 | ||
934 | if (is_latch) | |
935 | { | |
936 | prob_pass_main = TEST_BIT (wont_exit, 0) | |
937 | ? prob_pass_wont_exit | |
938 | : prob_pass_thru; | |
939 | p = prob_pass_main; | |
940 | scale_main = REG_BR_PROB_BASE; | |
941 | for (i = 0; i < ndupl; i++) | |
942 | { | |
943 | scale_main += p; | |
944 | p = RDIV (p * scale_step[i], REG_BR_PROB_BASE); | |
945 | } | |
946 | scale_main = RDIV (REG_BR_PROB_BASE * REG_BR_PROB_BASE, scale_main); | |
947 | scale_act = RDIV (scale_main * prob_pass_main, REG_BR_PROB_BASE); | |
948 | } | |
949 | else | |
950 | { | |
951 | scale_main = REG_BR_PROB_BASE; | |
952 | for (i = 0; i < ndupl; i++) | |
953 | scale_main = RDIV (scale_main * scale_step[i], REG_BR_PROB_BASE); | |
954 | scale_act = REG_BR_PROB_BASE - prob_pass_thru; | |
955 | } | |
956 | for (i = 0; i < ndupl; i++) | |
341c100f NS |
957 | gcc_assert (scale_step[i] >= 0 && scale_step[i] <= REG_BR_PROB_BASE); |
958 | gcc_assert (scale_main >= 0 && scale_main <= REG_BR_PROB_BASE | |
959 | && scale_act >= 0 && scale_act <= REG_BR_PROB_BASE); | |
617b465c ZD |
960 | } |
961 | ||
962 | /* Loop the new bbs will belong to. */ | |
8d28e87d | 963 | target = e->src->loop_father; |
617b465c ZD |
964 | |
965 | /* Original loops. */ | |
966 | n_orig_loops = 0; | |
967 | for (aloop = loop->inner; aloop; aloop = aloop->next) | |
968 | n_orig_loops++; | |
969 | orig_loops = xcalloc (n_orig_loops, sizeof (struct loop *)); | |
970 | for (aloop = loop->inner, i = 0; aloop; aloop = aloop->next, i++) | |
971 | orig_loops[i] = aloop; | |
972 | ||
973 | loop->copy = target; | |
d329e058 | 974 | |
617b465c ZD |
975 | n = loop->num_nodes; |
976 | ||
8d28e87d | 977 | first_active = xmalloc (n * sizeof (basic_block)); |
617b465c ZD |
978 | if (is_latch) |
979 | { | |
980 | memcpy (first_active, bbs, n * sizeof (basic_block)); | |
981 | first_active_latch = latch; | |
982 | } | |
983 | ||
82b85a85 ZD |
984 | /* Update the information about single exits. */ |
985 | if (loops->state & LOOPS_HAVE_MARKED_SINGLE_EXITS) | |
986 | update_single_exits_after_duplication (bbs, n, target); | |
987 | ||
8d28e87d ZD |
988 | /* Record exit edge in original loop body. */ |
989 | if (orig && TEST_BIT (wont_exit, 0)) | |
990 | to_remove[(*n_to_remove)++] = orig; | |
991 | ||
992 | spec_edges[SE_ORIG] = orig; | |
993 | spec_edges[SE_LATCH] = latch_edge; | |
d329e058 | 994 | |
617b465c ZD |
995 | for (j = 0; j < ndupl; j++) |
996 | { | |
997 | /* Copy loops. */ | |
998 | copy_loops_to (loops, orig_loops, n_orig_loops, target); | |
999 | ||
1000 | /* Copy bbs. */ | |
d47cc544 | 1001 | copy_bbs (bbs, n, new_bbs, spec_edges, 2, new_spec_edges, loop); |
8d28e87d | 1002 | |
113d659a ZD |
1003 | for (i = 0; i < n; i++) |
1004 | new_bbs[i]->rbi->copy_number = j + 1; | |
1005 | ||
84d45ad1 ZD |
1006 | /* Note whether the blocks and edges belong to an irreducible loop. */ |
1007 | if (add_irreducible_flag) | |
1008 | { | |
1009 | for (i = 0; i < n; i++) | |
1010 | new_bbs[i]->rbi->duplicated = 1; | |
1011 | for (i = 0; i < n; i++) | |
1012 | { | |
628f6a4e | 1013 | edge_iterator ei; |
84d45ad1 ZD |
1014 | new_bb = new_bbs[i]; |
1015 | if (new_bb->loop_father == target) | |
1016 | new_bb->flags |= BB_IRREDUCIBLE_LOOP; | |
1017 | ||
628f6a4e | 1018 | FOR_EACH_EDGE (ae, ei, new_bb->succs) |
84d45ad1 ZD |
1019 | if (ae->dest->rbi->duplicated |
1020 | && (ae->src->loop_father == target | |
1021 | || ae->dest->loop_father == target)) | |
1022 | ae->flags |= EDGE_IRREDUCIBLE_LOOP; | |
1023 | } | |
1024 | for (i = 0; i < n; i++) | |
1025 | new_bbs[i]->rbi->duplicated = 0; | |
1026 | } | |
1027 | ||
8d28e87d | 1028 | /* Redirect the special edges. */ |
617b465c | 1029 | if (is_latch) |
8d28e87d ZD |
1030 | { |
1031 | redirect_edge_and_branch_force (latch_edge, new_bbs[0]); | |
1032 | redirect_edge_and_branch_force (new_spec_edges[SE_LATCH], | |
1033 | loop->header); | |
d47cc544 | 1034 | set_immediate_dominator (CDI_DOMINATORS, new_bbs[0], latch); |
8d28e87d ZD |
1035 | latch = loop->latch = new_bbs[1]; |
1036 | e = latch_edge = new_spec_edges[SE_LATCH]; | |
1037 | } | |
1038 | else | |
1039 | { | |
1040 | redirect_edge_and_branch_force (new_spec_edges[SE_LATCH], | |
1041 | loop->header); | |
1042 | redirect_edge_and_branch_force (e, new_bbs[0]); | |
d47cc544 | 1043 | set_immediate_dominator (CDI_DOMINATORS, new_bbs[0], e->src); |
8d28e87d ZD |
1044 | e = new_spec_edges[SE_LATCH]; |
1045 | } | |
617b465c | 1046 | |
8d28e87d ZD |
1047 | /* Record exit edge in this copy. */ |
1048 | if (orig && TEST_BIT (wont_exit, j + 1)) | |
1049 | to_remove[(*n_to_remove)++] = new_spec_edges[SE_ORIG]; | |
d329e058 | 1050 | |
8d28e87d ZD |
1051 | /* Record the first copy in the control flow order if it is not |
1052 | the original loop (i.e. in case of peeling). */ | |
617b465c ZD |
1053 | if (!first_active_latch) |
1054 | { | |
1055 | memcpy (first_active, new_bbs, n * sizeof (basic_block)); | |
8d28e87d | 1056 | first_active_latch = new_bbs[1]; |
617b465c | 1057 | } |
d329e058 | 1058 | |
8d28e87d ZD |
1059 | /* Set counts and frequencies. */ |
1060 | if (flags & DLTHE_FLAG_UPDATE_FREQ) | |
617b465c | 1061 | { |
8d28e87d ZD |
1062 | scale_bbs_frequencies (new_bbs, n, scale_act, REG_BR_PROB_BASE); |
1063 | scale_act = RDIV (scale_act * scale_step[j], REG_BR_PROB_BASE); | |
617b465c ZD |
1064 | } |
1065 | } | |
8d28e87d ZD |
1066 | free (new_bbs); |
1067 | free (orig_loops); | |
1068 | ||
1069 | /* Update the original loop. */ | |
1070 | if (!is_latch) | |
d47cc544 | 1071 | set_immediate_dominator (CDI_DOMINATORS, e->dest, e->src); |
617b465c ZD |
1072 | if (flags & DLTHE_FLAG_UPDATE_FREQ) |
1073 | { | |
8d28e87d | 1074 | scale_bbs_frequencies (bbs, n, scale_main, REG_BR_PROB_BASE); |
617b465c ZD |
1075 | free (scale_step); |
1076 | } | |
617b465c | 1077 | |
8d28e87d | 1078 | /* Update dominators of outer blocks if affected. */ |
617b465c ZD |
1079 | for (i = 0; i < n; i++) |
1080 | { | |
1081 | basic_block dominated, dom_bb, *dom_bbs; | |
1082 | int n_dom_bbs,j; | |
1083 | ||
1084 | bb = bbs[i]; | |
113d659a ZD |
1085 | bb->rbi->copy_number = 0; |
1086 | ||
d47cc544 | 1087 | n_dom_bbs = get_dominated_by (CDI_DOMINATORS, bb, &dom_bbs); |
617b465c ZD |
1088 | for (j = 0; j < n_dom_bbs; j++) |
1089 | { | |
1090 | dominated = dom_bbs[j]; | |
1091 | if (flow_bb_inside_loop_p (loop, dominated)) | |
1092 | continue; | |
1093 | dom_bb = nearest_common_dominator ( | |
d47cc544 SB |
1094 | CDI_DOMINATORS, first_active[i], first_active_latch); |
1095 | set_immediate_dominator (CDI_DOMINATORS, dominated, dom_bb); | |
617b465c ZD |
1096 | } |
1097 | free (dom_bbs); | |
1098 | } | |
1099 | free (first_active); | |
1100 | ||
1101 | free (bbs); | |
1102 | ||
1103 | return true; | |
1104 | } | |
1105 | ||
f470c378 ZD |
1106 | /* A callback for make_forwarder block, to redirect all edges except for |
1107 | MFB_KJ_EDGE to the entry part. E is the edge for that we should decide | |
1108 | whether to redirect it. */ | |
1109 | ||
1110 | static edge mfb_kj_edge; | |
1111 | static bool | |
1112 | mfb_keep_just (edge e) | |
1113 | { | |
1114 | return e != mfb_kj_edge; | |
1115 | } | |
1116 | ||
1117 | /* A callback for make_forwarder block, to update data structures for a basic | |
1118 | block JUMP created by redirecting an edge (only the latch edge is being | |
1119 | redirected). */ | |
1120 | ||
1121 | static void | |
1122 | mfb_update_loops (basic_block jump) | |
1123 | { | |
c5cbcccf | 1124 | struct loop *loop = single_succ (jump)->loop_father; |
f470c378 ZD |
1125 | |
1126 | if (dom_computed[CDI_DOMINATORS]) | |
c5cbcccf | 1127 | set_immediate_dominator (CDI_DOMINATORS, jump, single_pred (jump)); |
f470c378 ZD |
1128 | add_bb_to_loop (jump, loop); |
1129 | loop->latch = jump; | |
1130 | } | |
1131 | ||
3d436d2a ZD |
1132 | /* Creates a pre-header for a LOOP. Returns newly created block. Unless |
1133 | CP_SIMPLE_PREHEADERS is set in FLAGS, we only force LOOP to have single | |
1134 | entry; otherwise we also force preheader block to have only one successor. | |
f470c378 ZD |
1135 | The function also updates dominators. */ |
1136 | ||
3d436d2a | 1137 | static basic_block |
d47cc544 | 1138 | create_preheader (struct loop *loop, int flags) |
3d436d2a ZD |
1139 | { |
1140 | edge e, fallthru; | |
1141 | basic_block dummy; | |
3d436d2a ZD |
1142 | struct loop *cloop, *ploop; |
1143 | int nentry = 0; | |
f470c378 | 1144 | bool irred = false; |
c15bc84b EB |
1145 | bool latch_edge_was_fallthru; |
1146 | edge one_succ_pred = 0; | |
628f6a4e | 1147 | edge_iterator ei; |
3d436d2a ZD |
1148 | |
1149 | cloop = loop->outer; | |
1150 | ||
628f6a4e | 1151 | FOR_EACH_EDGE (e, ei, loop->header->preds) |
3d436d2a ZD |
1152 | { |
1153 | if (e->src == loop->latch) | |
1154 | continue; | |
f470c378 | 1155 | irred |= (e->flags & EDGE_IRREDUCIBLE_LOOP) != 0; |
3d436d2a | 1156 | nentry++; |
c5cbcccf | 1157 | if (single_succ_p (e->src)) |
c15bc84b | 1158 | one_succ_pred = e; |
3d436d2a | 1159 | } |
341c100f | 1160 | gcc_assert (nentry); |
3d436d2a ZD |
1161 | if (nentry == 1) |
1162 | { | |
b057216d KH |
1163 | /* Get an edge that is different from the one from loop->latch |
1164 | to loop->header. */ | |
1165 | e = EDGE_PRED (loop->header, | |
1166 | EDGE_PRED (loop->header, 0)->src == loop->latch); | |
628f6a4e | 1167 | |
c5cbcccf | 1168 | if (!(flags & CP_SIMPLE_PREHEADERS) || single_succ_p (e->src)) |
3d436d2a ZD |
1169 | return NULL; |
1170 | } | |
1171 | ||
f470c378 | 1172 | mfb_kj_edge = loop_latch_edge (loop); |
c15bc84b | 1173 | latch_edge_was_fallthru = (mfb_kj_edge->flags & EDGE_FALLTHRU) != 0; |
f470c378 ZD |
1174 | fallthru = make_forwarder_block (loop->header, mfb_keep_just, |
1175 | mfb_update_loops); | |
3d436d2a ZD |
1176 | dummy = fallthru->src; |
1177 | loop->header = fallthru->dest; | |
1178 | ||
1179 | /* The header could be a latch of some superloop(s); due to design of | |
1180 | split_block, it would now move to fallthru->dest. */ | |
1181 | for (ploop = loop; ploop; ploop = ploop->outer) | |
1182 | if (ploop->latch == dummy) | |
1183 | ploop->latch = fallthru->dest; | |
1184 | ||
c15bc84b EB |
1185 | /* Try to be clever in placing the newly created preheader. The idea is to |
1186 | avoid breaking any "fallthruness" relationship between blocks. | |
1187 | ||
1188 | The preheader was created just before the header and all incoming edges | |
1189 | to the header were redirected to the preheader, except the latch edge. | |
1190 | So the only problematic case is when this latch edge was a fallthru | |
1191 | edge: it is not anymore after the preheader creation so we have broken | |
1192 | the fallthruness. We're therefore going to look for a better place. */ | |
1193 | if (latch_edge_was_fallthru) | |
1194 | { | |
1195 | if (one_succ_pred) | |
1196 | e = one_succ_pred; | |
1197 | else | |
1198 | e = EDGE_PRED (dummy, 0); | |
1199 | ||
1200 | move_block_after (dummy, e->src); | |
1201 | } | |
f470c378 ZD |
1202 | |
1203 | loop->header->loop_father = loop; | |
1204 | add_bb_to_loop (dummy, cloop); | |
3d436d2a | 1205 | |
f470c378 | 1206 | if (irred) |
3d436d2a | 1207 | { |
f470c378 | 1208 | dummy->flags |= BB_IRREDUCIBLE_LOOP; |
c5cbcccf | 1209 | single_succ_edge (dummy)->flags |= EDGE_IRREDUCIBLE_LOOP; |
3d436d2a ZD |
1210 | } |
1211 | ||
c263766c RH |
1212 | if (dump_file) |
1213 | fprintf (dump_file, "Created preheader block for loop %i\n", | |
3d436d2a ZD |
1214 | loop->num); |
1215 | ||
1216 | return dummy; | |
1217 | } | |
1218 | ||
617b465c ZD |
1219 | /* Create preheaders for each loop from loop tree stored in LOOPS; for meaning |
1220 | of FLAGS see create_preheader. */ | |
3d436d2a | 1221 | void |
d329e058 | 1222 | create_preheaders (struct loops *loops, int flags) |
3d436d2a ZD |
1223 | { |
1224 | unsigned i; | |
1225 | for (i = 1; i < loops->num; i++) | |
d47cc544 | 1226 | create_preheader (loops->parray[i], flags); |
3d436d2a ZD |
1227 | loops->state |= LOOPS_HAVE_PREHEADERS; |
1228 | } | |
1229 | ||
617b465c ZD |
1230 | /* Forces all loop latches of loops from loop tree LOOPS to have only single |
1231 | successor. */ | |
3d436d2a | 1232 | void |
d329e058 | 1233 | force_single_succ_latches (struct loops *loops) |
3d436d2a ZD |
1234 | { |
1235 | unsigned i; | |
1236 | struct loop *loop; | |
1237 | edge e; | |
1238 | ||
1239 | for (i = 1; i < loops->num; i++) | |
1240 | { | |
1241 | loop = loops->parray[i]; | |
c5cbcccf | 1242 | if (loop->latch != loop->header && single_succ_p (loop->latch)) |
3d436d2a | 1243 | continue; |
d329e058 | 1244 | |
9ff3d2de | 1245 | e = find_edge (loop->latch, loop->header); |
bc810602 | 1246 | |
d47cc544 | 1247 | loop_split_edge_with (e, NULL_RTX); |
3d436d2a ZD |
1248 | } |
1249 | loops->state |= LOOPS_HAVE_SIMPLE_LATCHES; | |
1250 | } | |
1251 | ||
617b465c ZD |
1252 | /* A quite stupid function to put INSNS on edge E. They are supposed to form |
1253 | just one basic block. Jumps in INSNS are not handled, so cfg do not have to | |
1254 | be ok after this function. The created block is placed on correct place | |
1255 | in LOOPS structure and its dominator is set. */ | |
3d436d2a | 1256 | basic_block |
d47cc544 | 1257 | loop_split_edge_with (edge e, rtx insns) |
3d436d2a ZD |
1258 | { |
1259 | basic_block src, dest, new_bb; | |
1260 | struct loop *loop_c; | |
d329e058 | 1261 | |
3d436d2a ZD |
1262 | src = e->src; |
1263 | dest = e->dest; | |
1264 | ||
1265 | loop_c = find_common_loop (src->loop_father, dest->loop_father); | |
1266 | ||
1267 | /* Create basic block for it. */ | |
1268 | ||
bc35512f | 1269 | new_bb = split_edge (e); |
3d436d2a | 1270 | add_bb_to_loop (new_bb, loop_c); |
a746fd8c | 1271 | new_bb->flags |= (insns ? BB_SUPERBLOCK : 0); |
3d436d2a | 1272 | |
3d436d2a | 1273 | if (insns) |
a813c111 | 1274 | emit_insn_after (insns, BB_END (new_bb)); |
3d436d2a | 1275 | |
3d436d2a ZD |
1276 | if (dest->loop_father->latch == src) |
1277 | dest->loop_father->latch = new_bb; | |
d329e058 | 1278 | |
3d436d2a ZD |
1279 | return new_bb; |
1280 | } | |
6de9cd9a DN |
1281 | |
1282 | /* Uses the natural loop discovery to recreate loop notes. */ | |
1283 | void | |
1284 | create_loop_notes (void) | |
1285 | { | |
1286 | rtx insn, head, end; | |
1287 | struct loops loops; | |
1288 | struct loop *loop; | |
1289 | basic_block *first, *last, bb, pbb; | |
1290 | struct loop **stack, **top; | |
1291 | ||
1292 | #ifdef ENABLE_CHECKING | |
1293 | /* Verify that there really are no loop notes. */ | |
1294 | for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) | |
341c100f NS |
1295 | gcc_assert (!NOTE_P (insn) || |
1296 | NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG); | |
6de9cd9a DN |
1297 | #endif |
1298 | ||
70388d94 | 1299 | flow_loops_find (&loops); |
6de9cd9a DN |
1300 | free_dominance_info (CDI_DOMINATORS); |
1301 | if (loops.num > 1) | |
1302 | { | |
1303 | last = xcalloc (loops.num, sizeof (basic_block)); | |
1304 | ||
1305 | FOR_EACH_BB (bb) | |
1306 | { | |
1307 | for (loop = bb->loop_father; loop->outer; loop = loop->outer) | |
1308 | last[loop->num] = bb; | |
1309 | } | |
1310 | ||
1311 | first = xcalloc (loops.num, sizeof (basic_block)); | |
1312 | stack = xcalloc (loops.num, sizeof (struct loop *)); | |
1313 | top = stack; | |
1314 | ||
1315 | FOR_EACH_BB (bb) | |
1316 | { | |
1317 | for (loop = bb->loop_father; loop->outer; loop = loop->outer) | |
1318 | { | |
1319 | if (!first[loop->num]) | |
1320 | { | |
1321 | *top++ = loop; | |
1322 | first[loop->num] = bb; | |
1323 | } | |
1324 | ||
1325 | if (bb == last[loop->num]) | |
1326 | { | |
1327 | /* Prevent loops from overlapping. */ | |
1328 | while (*--top != loop) | |
1329 | last[(*top)->num] = EXIT_BLOCK_PTR; | |
1330 | ||
1331 | /* If loop starts with jump into it, place the note in | |
1332 | front of the jump. */ | |
1333 | insn = PREV_INSN (BB_HEAD (first[loop->num])); | |
1334 | if (insn | |
4b4bf941 | 1335 | && BARRIER_P (insn)) |
6de9cd9a DN |
1336 | insn = PREV_INSN (insn); |
1337 | ||
1338 | if (insn | |
4b4bf941 | 1339 | && JUMP_P (insn) |
6de9cd9a DN |
1340 | && any_uncondjump_p (insn) |
1341 | && onlyjump_p (insn)) | |
1342 | { | |
1343 | pbb = BLOCK_FOR_INSN (insn); | |
c5cbcccf | 1344 | gcc_assert (pbb && single_succ_p (pbb)); |
6de9cd9a | 1345 | |
c5cbcccf | 1346 | if (!flow_bb_inside_loop_p (loop, single_succ (pbb))) |
6de9cd9a DN |
1347 | insn = BB_HEAD (first[loop->num]); |
1348 | } | |
1349 | else | |
1350 | insn = BB_HEAD (first[loop->num]); | |
1351 | ||
1352 | head = BB_HEAD (first[loop->num]); | |
1353 | emit_note_before (NOTE_INSN_LOOP_BEG, insn); | |
1354 | BB_HEAD (first[loop->num]) = head; | |
1355 | ||
1356 | /* Position the note correctly wrto barrier. */ | |
1357 | insn = BB_END (last[loop->num]); | |
1358 | if (NEXT_INSN (insn) | |
4b4bf941 | 1359 | && BARRIER_P (NEXT_INSN (insn))) |
6de9cd9a DN |
1360 | insn = NEXT_INSN (insn); |
1361 | ||
1362 | end = BB_END (last[loop->num]); | |
1363 | emit_note_after (NOTE_INSN_LOOP_END, insn); | |
1364 | BB_END (last[loop->num]) = end; | |
1365 | } | |
1366 | } | |
1367 | } | |
1368 | ||
1369 | free (first); | |
1370 | free (last); | |
1371 | free (stack); | |
1372 | } | |
1373 | flow_loops_free (&loops); | |
1374 | } | |
2b271002 ZD |
1375 | |
1376 | /* The structure of LOOPS might have changed. Some loops might get removed | |
1377 | (and their headers and latches were set to NULL), loop exists might get | |
1378 | removed (thus the loop nesting may be wrong), and some blocks and edges | |
1379 | were changed (so the information about bb --> loop mapping does not have | |
1380 | to be correct). But still for the remaining loops the header dominates | |
1381 | the latch, and loops did not get new subloobs (new loops might possibly | |
1382 | get created, but we are not interested in them). Fix up the mess. | |
1383 | ||
1384 | If CHANGED_BBS is not NULL, basic blocks whose loop has changed are | |
1385 | marked in it. */ | |
1386 | ||
1387 | void | |
1388 | fix_loop_structure (struct loops *loops, bitmap changed_bbs) | |
1389 | { | |
1390 | basic_block bb; | |
1391 | struct loop *loop, *ploop; | |
1392 | unsigned i; | |
1393 | ||
1394 | /* Remove the old bb -> loop mapping. */ | |
1395 | FOR_EACH_BB (bb) | |
1396 | { | |
1397 | bb->aux = (void *) (size_t) bb->loop_father->depth; | |
1398 | bb->loop_father = loops->tree_root; | |
1399 | } | |
1400 | ||
1401 | /* Remove the dead loops from structures. */ | |
1402 | loops->tree_root->num_nodes = n_basic_blocks + 2; | |
1403 | for (i = 1; i < loops->num; i++) | |
1404 | { | |
1405 | loop = loops->parray[i]; | |
1406 | if (!loop) | |
1407 | continue; | |
1408 | ||
1409 | loop->num_nodes = 0; | |
1410 | if (loop->header) | |
1411 | continue; | |
1412 | ||
1413 | while (loop->inner) | |
1414 | { | |
1415 | ploop = loop->inner; | |
1416 | flow_loop_tree_node_remove (ploop); | |
1417 | flow_loop_tree_node_add (loop->outer, ploop); | |
1418 | } | |
1419 | ||
1420 | /* Remove the loop and free its data. */ | |
1421 | flow_loop_tree_node_remove (loop); | |
1422 | loops->parray[loop->num] = NULL; | |
1423 | flow_loop_free (loop); | |
1424 | } | |
1425 | ||
1426 | /* Rescan the bodies of loops, starting from the outermost. */ | |
1427 | loop = loops->tree_root; | |
1428 | while (1) | |
1429 | { | |
1430 | if (loop->inner) | |
1431 | loop = loop->inner; | |
1432 | else | |
1433 | { | |
1434 | while (!loop->next | |
1435 | && loop != loops->tree_root) | |
1436 | loop = loop->outer; | |
1437 | if (loop == loops->tree_root) | |
1438 | break; | |
1439 | ||
1440 | loop = loop->next; | |
1441 | } | |
1442 | ||
1443 | loop->num_nodes = flow_loop_nodes_find (loop->header, loop); | |
1444 | } | |
1445 | ||
1446 | /* Now fix the loop nesting. */ | |
1447 | for (i = 1; i < loops->num; i++) | |
1448 | { | |
1449 | loop = loops->parray[i]; | |
1450 | if (!loop) | |
1451 | continue; | |
1452 | ||
1453 | bb = loop_preheader_edge (loop)->src; | |
1454 | if (bb->loop_father != loop->outer) | |
1455 | { | |
1456 | flow_loop_tree_node_remove (loop); | |
1457 | flow_loop_tree_node_add (bb->loop_father, loop); | |
1458 | } | |
1459 | } | |
1460 | ||
1461 | /* Mark the blocks whose loop has changed. */ | |
1462 | FOR_EACH_BB (bb) | |
1463 | { | |
1464 | if (changed_bbs | |
1465 | && (void *) (size_t) bb->loop_father->depth != bb->aux) | |
1466 | bitmap_set_bit (changed_bbs, bb->index); | |
1467 | ||
1468 | bb->aux = NULL; | |
1469 | } | |
1470 | ||
1471 | mark_single_exit_loops (loops); | |
1472 | mark_irreducible_loops (loops); | |
1473 | } |