]>
Commit | Line | Data |
---|---|---|
3d436d2a | 1 | /* Loop manipulation code for GNU compiler. |
85ec4feb | 2 | Copyright (C) 2002-2018 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 | |
9dcd6f09 | 8 | Software Foundation; either version 3, or (at your option) any later |
3d436d2a ZD |
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 | |
9dcd6f09 NC |
17 | along with GCC; see the file COPYING3. If not see |
18 | <http://www.gnu.org/licenses/>. */ | |
3d436d2a ZD |
19 | |
20 | #include "config.h" | |
21 | #include "system.h" | |
22 | #include "coretypes.h" | |
c7131fb2 | 23 | #include "backend.h" |
957060b5 | 24 | #include "rtl.h" |
c7131fb2 AM |
25 | #include "tree.h" |
26 | #include "gimple.h" | |
957060b5 | 27 | #include "cfghooks.h" |
60393bbc | 28 | #include "cfganal.h" |
3d436d2a | 29 | #include "cfgloop.h" |
5be5c238 | 30 | #include "gimple-iterator.h" |
18f429e2 | 31 | #include "gimplify-me.h" |
e28030cf | 32 | #include "tree-ssa-loop-manip.h" |
7ee2468b | 33 | #include "dumpfile.h" |
3d436d2a | 34 | |
d73be268 | 35 | static void copy_loops_to (struct loop **, int, |
d329e058 AJ |
36 | struct loop *); |
37 | static void loop_redirect_edge (edge, basic_block); | |
d47cc544 | 38 | static void remove_bbs (basic_block *, int); |
ed7a4b4b | 39 | static bool rpe_enum_p (const_basic_block, const void *); |
d47cc544 | 40 | static int find_path (edge, basic_block **); |
d73be268 ZD |
41 | static void fix_loop_placements (struct loop *, bool *); |
42 | static bool fix_bb_placement (basic_block); | |
1a7de201 | 43 | static void fix_bb_placements (basic_block, bool *, bitmap); |
3d436d2a | 44 | |
d47cc544 | 45 | /* Checks whether basic block BB is dominated by DATA. */ |
617b465c | 46 | static bool |
ed7a4b4b | 47 | rpe_enum_p (const_basic_block bb, const void *data) |
617b465c | 48 | { |
ed7a4b4b | 49 | return dominated_by_p (CDI_DOMINATORS, bb, (const_basic_block) data); |
617b465c ZD |
50 | } |
51 | ||
598ec7bd ZD |
52 | /* Remove basic blocks BBS. NBBS is the number of the basic blocks. */ |
53 | ||
617b465c | 54 | static void |
d47cc544 | 55 | remove_bbs (basic_block *bbs, int nbbs) |
617b465c ZD |
56 | { |
57 | int i; | |
58 | ||
59 | for (i = 0; i < nbbs; i++) | |
598ec7bd | 60 | delete_basic_block (bbs[i]); |
617b465c ZD |
61 | } |
62 | ||
63 | /* Find path -- i.e. the basic blocks dominated by edge E and put them | |
64 | into array BBS, that will be allocated large enough to contain them. | |
35b07080 ZD |
65 | E->dest must have exactly one predecessor for this to work (it is |
66 | easy to achieve and we do not put it here because we do not want to | |
67 | alter anything by this function). The number of basic blocks in the | |
68 | path is returned. */ | |
617b465c | 69 | static int |
d47cc544 | 70 | find_path (edge e, basic_block **bbs) |
617b465c | 71 | { |
628f6a4e | 72 | gcc_assert (EDGE_COUNT (e->dest->preds) <= 1); |
617b465c ZD |
73 | |
74 | /* Find bbs in the path. */ | |
0cae8d31 | 75 | *bbs = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun)); |
617b465c | 76 | return dfs_enumerate_from (e->dest, 0, rpe_enum_p, *bbs, |
0cae8d31 | 77 | n_basic_blocks_for_fn (cfun), e->dest); |
617b465c ZD |
78 | } |
79 | ||
d73be268 | 80 | /* Fix placement of basic block BB inside loop hierarchy -- |
617b465c ZD |
81 | Let L be a loop to that BB belongs. Then every successor of BB must either |
82 | 1) belong to some superloop of loop L, or | |
83 | 2) be a header of loop K such that K->outer is superloop of L | |
84 | Returns true if we had to move BB into other loop to enforce this condition, | |
85 | false if the placement of BB was already correct (provided that placements | |
86 | of its successors are correct). */ | |
87 | static bool | |
d73be268 | 88 | fix_bb_placement (basic_block bb) |
617b465c ZD |
89 | { |
90 | edge e; | |
628f6a4e | 91 | edge_iterator ei; |
d73be268 | 92 | struct loop *loop = current_loops->tree_root, *act; |
617b465c | 93 | |
628f6a4e | 94 | FOR_EACH_EDGE (e, ei, bb->succs) |
617b465c | 95 | { |
fefa31b5 | 96 | if (e->dest == EXIT_BLOCK_PTR_FOR_FN (cfun)) |
617b465c ZD |
97 | continue; |
98 | ||
99 | act = e->dest->loop_father; | |
100 | if (act->header == e->dest) | |
9ba025a2 | 101 | act = loop_outer (act); |
617b465c ZD |
102 | |
103 | if (flow_loop_nested_p (loop, act)) | |
104 | loop = act; | |
105 | } | |
106 | ||
107 | if (loop == bb->loop_father) | |
108 | return false; | |
109 | ||
110 | remove_bb_from_loops (bb); | |
111 | add_bb_to_loop (bb, loop); | |
112 | ||
113 | return true; | |
114 | } | |
115 | ||
b4c1c7e3 ZD |
116 | /* Fix placement of LOOP inside loop tree, i.e. find the innermost superloop |
117 | of LOOP to that leads at least one exit edge of LOOP, and set it | |
118 | as the immediate superloop of LOOP. Return true if the immediate superloop | |
1bd3f750 MP |
119 | of LOOP changed. |
120 | ||
121 | IRRED_INVALIDATED is set to true if a change in the loop structures might | |
122 | invalidate the information about irreducible regions. */ | |
b4c1c7e3 ZD |
123 | |
124 | static bool | |
1bd3f750 | 125 | fix_loop_placement (struct loop *loop, bool *irred_invalidated) |
b4c1c7e3 ZD |
126 | { |
127 | unsigned i; | |
128 | edge e; | |
9771b263 | 129 | vec<edge> exits = get_loop_exit_edges (loop); |
b4c1c7e3 ZD |
130 | struct loop *father = current_loops->tree_root, *act; |
131 | bool ret = false; | |
132 | ||
9771b263 | 133 | FOR_EACH_VEC_ELT (exits, i, e) |
b4c1c7e3 ZD |
134 | { |
135 | act = find_common_loop (loop, e->dest->loop_father); | |
136 | if (flow_loop_nested_p (father, act)) | |
137 | father = act; | |
138 | } | |
139 | ||
9ba025a2 | 140 | if (father != loop_outer (loop)) |
b4c1c7e3 | 141 | { |
9ba025a2 | 142 | for (act = loop_outer (loop); act != father; act = loop_outer (act)) |
b4c1c7e3 ZD |
143 | act->num_nodes -= loop->num_nodes; |
144 | flow_loop_tree_node_remove (loop); | |
145 | flow_loop_tree_node_add (father, loop); | |
146 | ||
147 | /* The exit edges of LOOP no longer exits its original immediate | |
148 | superloops; remove them from the appropriate exit lists. */ | |
9771b263 | 149 | FOR_EACH_VEC_ELT (exits, i, e) |
1bd3f750 MP |
150 | { |
151 | /* We may need to recompute irreducible loops. */ | |
152 | if (e->flags & EDGE_IRREDUCIBLE_LOOP) | |
153 | *irred_invalidated = true; | |
154 | rescan_loop_exit (e, false, false); | |
155 | } | |
b4c1c7e3 ZD |
156 | |
157 | ret = true; | |
158 | } | |
159 | ||
9771b263 | 160 | exits.release (); |
b4c1c7e3 ZD |
161 | return ret; |
162 | } | |
163 | ||
617b465c | 164 | /* Fix placements of basic blocks inside loop hierarchy stored in loops; i.e. |
026c3cfd | 165 | enforce condition stated in description of fix_bb_placement. We |
617b465c ZD |
166 | start from basic block FROM that had some of its successors removed, so that |
167 | his placement no longer has to be correct, and iteratively fix placement of | |
168 | its predecessors that may change if placement of FROM changed. Also fix | |
169 | placement of subloops of FROM->loop_father, that might also be altered due | |
4d6922ee | 170 | to this change; the condition for them is similar, except that instead of |
dc14f191 | 171 | successors we consider edges coming out of the loops. |
b8698a0f | 172 | |
dc14f191 | 173 | If the changes may invalidate the information about irreducible regions, |
1a7de201 JH |
174 | IRRED_INVALIDATED is set to true. |
175 | ||
176 | If LOOP_CLOSED_SSA_INVLIDATED is non-zero then all basic blocks with | |
177 | changed loop_father are collected there. */ | |
dc14f191 | 178 | |
617b465c | 179 | static void |
d73be268 | 180 | fix_bb_placements (basic_block from, |
1a7de201 JH |
181 | bool *irred_invalidated, |
182 | bitmap loop_closed_ssa_invalidated) | |
617b465c | 183 | { |
617b465c | 184 | basic_block *queue, *qtop, *qbeg, *qend; |
634ee309 | 185 | struct loop *base_loop, *target_loop; |
617b465c ZD |
186 | edge e; |
187 | ||
188 | /* We pass through blocks back-reachable from FROM, testing whether some | |
189 | of their successors moved to outer loop. It may be necessary to | |
190 | iterate several times, but it is finite, as we stop unless we move | |
191 | the basic block up the loop structure. The whole story is a bit | |
192 | more complicated due to presence of subloops, those are moved using | |
193 | fix_loop_placement. */ | |
194 | ||
195 | base_loop = from->loop_father; | |
cec8ac0b ZD |
196 | /* If we are already in the outermost loop, the basic blocks cannot be moved |
197 | outside of it. If FROM is the header of the base loop, it cannot be moved | |
198 | outside of it, either. In both cases, we can end now. */ | |
199 | if (base_loop == current_loops->tree_root | |
200 | || from == base_loop->header) | |
617b465c ZD |
201 | return; |
202 | ||
7ba9e72d | 203 | auto_sbitmap in_queue (last_basic_block_for_fn (cfun)); |
f61e445a | 204 | bitmap_clear (in_queue); |
d7c028c0 | 205 | bitmap_set_bit (in_queue, from->index); |
617b465c | 206 | /* Prevent us from going out of the base_loop. */ |
d7c028c0 | 207 | bitmap_set_bit (in_queue, base_loop->header->index); |
617b465c | 208 | |
5ed6ace5 | 209 | queue = XNEWVEC (basic_block, base_loop->num_nodes + 1); |
617b465c ZD |
210 | qtop = queue + base_loop->num_nodes + 1; |
211 | qbeg = queue; | |
212 | qend = queue + 1; | |
213 | *qbeg = from; | |
214 | ||
215 | while (qbeg != qend) | |
216 | { | |
628f6a4e | 217 | edge_iterator ei; |
617b465c ZD |
218 | from = *qbeg; |
219 | qbeg++; | |
220 | if (qbeg == qtop) | |
221 | qbeg = queue; | |
d7c028c0 | 222 | bitmap_clear_bit (in_queue, from->index); |
617b465c ZD |
223 | |
224 | if (from->loop_father->header == from) | |
225 | { | |
226 | /* Subloop header, maybe move the loop upward. */ | |
1bd3f750 | 227 | if (!fix_loop_placement (from->loop_father, irred_invalidated)) |
617b465c | 228 | continue; |
634ee309 | 229 | target_loop = loop_outer (from->loop_father); |
467a3558 RB |
230 | if (loop_closed_ssa_invalidated) |
231 | { | |
232 | basic_block *bbs = get_loop_body (from->loop_father); | |
233 | for (unsigned i = 0; i < from->loop_father->num_nodes; ++i) | |
234 | bitmap_set_bit (loop_closed_ssa_invalidated, bbs[i]->index); | |
235 | free (bbs); | |
236 | } | |
617b465c ZD |
237 | } |
238 | else | |
239 | { | |
240 | /* Ordinary basic block. */ | |
d73be268 | 241 | if (!fix_bb_placement (from)) |
617b465c | 242 | continue; |
467a3558 | 243 | target_loop = from->loop_father; |
1a7de201 JH |
244 | if (loop_closed_ssa_invalidated) |
245 | bitmap_set_bit (loop_closed_ssa_invalidated, from->index); | |
617b465c ZD |
246 | } |
247 | ||
dc14f191 ZD |
248 | FOR_EACH_EDGE (e, ei, from->succs) |
249 | { | |
250 | if (e->flags & EDGE_IRREDUCIBLE_LOOP) | |
251 | *irred_invalidated = true; | |
252 | } | |
253 | ||
617b465c | 254 | /* Something has changed, insert predecessors into queue. */ |
628f6a4e | 255 | FOR_EACH_EDGE (e, ei, from->preds) |
617b465c ZD |
256 | { |
257 | basic_block pred = e->src; | |
258 | struct loop *nca; | |
259 | ||
dc14f191 ZD |
260 | if (e->flags & EDGE_IRREDUCIBLE_LOOP) |
261 | *irred_invalidated = true; | |
262 | ||
d7c028c0 | 263 | if (bitmap_bit_p (in_queue, pred->index)) |
617b465c ZD |
264 | continue; |
265 | ||
d329e058 | 266 | /* If it is subloop, then it either was not moved, or |
617b465c ZD |
267 | the path up the loop tree from base_loop do not contain |
268 | it. */ | |
269 | nca = find_common_loop (pred->loop_father, base_loop); | |
270 | if (pred->loop_father != base_loop | |
271 | && (nca == base_loop | |
272 | || nca != pred->loop_father)) | |
273 | pred = pred->loop_father->header; | |
634ee309 | 274 | else if (!flow_loop_nested_p (target_loop, pred->loop_father)) |
617b465c | 275 | { |
634ee309 ZD |
276 | /* If PRED is already higher in the loop hierarchy than the |
277 | TARGET_LOOP to that we moved FROM, the change of the position | |
278 | of FROM does not affect the position of PRED, so there is no | |
279 | point in processing it. */ | |
617b465c ZD |
280 | continue; |
281 | } | |
282 | ||
d7c028c0 | 283 | if (bitmap_bit_p (in_queue, pred->index)) |
617b465c ZD |
284 | continue; |
285 | ||
286 | /* Schedule the basic block. */ | |
287 | *qend = pred; | |
288 | qend++; | |
289 | if (qend == qtop) | |
290 | qend = queue; | |
d7c028c0 | 291 | bitmap_set_bit (in_queue, pred->index); |
617b465c ZD |
292 | } |
293 | } | |
617b465c ZD |
294 | free (queue); |
295 | } | |
296 | ||
297 | /* Removes path beginning at edge E, i.e. remove basic blocks dominated by E | |
d73be268 ZD |
298 | and update loop structures and dominators. Return true if we were able |
299 | to remove the path, false otherwise (and nothing is affected then). */ | |
617b465c | 300 | bool |
eb2afa1a RB |
301 | remove_path (edge e, bool *irred_invalidated, |
302 | bitmap loop_closed_ssa_invalidated) | |
617b465c ZD |
303 | { |
304 | edge ae; | |
66f97d31 | 305 | basic_block *rem_bbs, *bord_bbs, from, bb; |
9771b263 | 306 | vec<basic_block> dom_bbs; |
2f697bc4 | 307 | int i, nrem, n_bord_bbs; |
eb2afa1a | 308 | bool local_irred_invalidated = false; |
06f1716b | 309 | edge_iterator ei; |
56494762 | 310 | struct loop *l, *f; |
617b465c | 311 | |
eb2afa1a RB |
312 | if (! irred_invalidated) |
313 | irred_invalidated = &local_irred_invalidated; | |
314 | ||
14fa2cc0 | 315 | if (!can_remove_branch_p (e)) |
35b07080 ZD |
316 | return false; |
317 | ||
dc14f191 ZD |
318 | /* Keep track of whether we need to update information about irreducible |
319 | regions. This is the case if the removed area is a part of the | |
320 | irreducible region, or if the set of basic blocks that belong to a loop | |
321 | that is inside an irreducible region is changed, or if such a loop is | |
322 | removed. */ | |
323 | if (e->flags & EDGE_IRREDUCIBLE_LOOP) | |
eb2afa1a | 324 | *irred_invalidated = true; |
dc14f191 | 325 | |
35b07080 ZD |
326 | /* We need to check whether basic blocks are dominated by the edge |
327 | e, but we only have basic block dominators. This is easy to | |
328 | fix -- when e->dest has exactly one predecessor, this corresponds | |
329 | to blocks dominated by e->dest, if not, split the edge. */ | |
c5cbcccf | 330 | if (!single_pred_p (e->dest)) |
598ec7bd | 331 | e = single_pred_edge (split_edge (e)); |
35b07080 ZD |
332 | |
333 | /* It may happen that by removing path we remove one or more loops | |
334 | we belong to. In this case first unloop the loops, then proceed | |
335 | normally. We may assume that e->dest is not a header of any loop, | |
336 | as it now has exactly one predecessor. */ | |
56494762 JH |
337 | for (l = e->src->loop_father; loop_outer (l); l = f) |
338 | { | |
339 | f = loop_outer (l); | |
340 | if (dominated_by_p (CDI_DOMINATORS, l->latch, e->dest)) | |
eb2afa1a | 341 | unloop (l, irred_invalidated, loop_closed_ssa_invalidated); |
56494762 | 342 | } |
d329e058 | 343 | |
35b07080 | 344 | /* Identify the path. */ |
d47cc544 | 345 | nrem = find_path (e, &rem_bbs); |
617b465c ZD |
346 | |
347 | n_bord_bbs = 0; | |
0cae8d31 | 348 | bord_bbs = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun)); |
7ba9e72d | 349 | auto_sbitmap seen (last_basic_block_for_fn (cfun)); |
f61e445a | 350 | bitmap_clear (seen); |
617b465c ZD |
351 | |
352 | /* Find "border" hexes -- i.e. those with predecessor in removed path. */ | |
353 | for (i = 0; i < nrem; i++) | |
d7c028c0 | 354 | bitmap_set_bit (seen, rem_bbs[i]->index); |
eb2afa1a | 355 | if (!*irred_invalidated) |
06f1716b | 356 | FOR_EACH_EDGE (ae, ei, e->src->succs) |
fefa31b5 DM |
357 | if (ae != e && ae->dest != EXIT_BLOCK_PTR_FOR_FN (cfun) |
358 | && !bitmap_bit_p (seen, ae->dest->index) | |
06f1716b | 359 | && ae->flags & EDGE_IRREDUCIBLE_LOOP) |
c7b3b99f | 360 | { |
eb2afa1a | 361 | *irred_invalidated = true; |
c7b3b99f PCC |
362 | break; |
363 | } | |
364 | ||
35b07080 | 365 | for (i = 0; i < nrem; i++) |
617b465c | 366 | { |
35b07080 | 367 | bb = rem_bbs[i]; |
628f6a4e | 368 | FOR_EACH_EDGE (ae, ei, rem_bbs[i]->succs) |
fefa31b5 DM |
369 | if (ae->dest != EXIT_BLOCK_PTR_FOR_FN (cfun) |
370 | && !bitmap_bit_p (seen, ae->dest->index)) | |
35b07080 | 371 | { |
d7c028c0 | 372 | bitmap_set_bit (seen, ae->dest->index); |
35b07080 | 373 | bord_bbs[n_bord_bbs++] = ae->dest; |
b8698a0f | 374 | |
dc14f191 | 375 | if (ae->flags & EDGE_IRREDUCIBLE_LOOP) |
eb2afa1a | 376 | *irred_invalidated = true; |
35b07080 | 377 | } |
617b465c | 378 | } |
617b465c ZD |
379 | |
380 | /* Remove the path. */ | |
381 | from = e->src; | |
14fa2cc0 | 382 | remove_branch (e); |
9771b263 | 383 | dom_bbs.create (0); |
617b465c ZD |
384 | |
385 | /* Cancel loops contained in the path. */ | |
386 | for (i = 0; i < nrem; i++) | |
387 | if (rem_bbs[i]->loop_father->header == rem_bbs[i]) | |
2f697bc4 | 388 | cancel_loop_tree (rem_bbs[i]->loop_father); |
598ec7bd | 389 | |
9b43d37b JJ |
390 | remove_bbs (rem_bbs, nrem); |
391 | free (rem_bbs); | |
392 | ||
35b07080 | 393 | /* Find blocks whose dominators may be affected. */ |
f61e445a | 394 | bitmap_clear (seen); |
617b465c ZD |
395 | for (i = 0; i < n_bord_bbs; i++) |
396 | { | |
d47cc544 | 397 | basic_block ldom; |
617b465c | 398 | |
d47cc544 | 399 | bb = get_immediate_dominator (CDI_DOMINATORS, bord_bbs[i]); |
d7c028c0 | 400 | if (bitmap_bit_p (seen, bb->index)) |
617b465c | 401 | continue; |
d7c028c0 | 402 | bitmap_set_bit (seen, bb->index); |
617b465c | 403 | |
d47cc544 SB |
404 | for (ldom = first_dom_son (CDI_DOMINATORS, bb); |
405 | ldom; | |
406 | ldom = next_dom_son (CDI_DOMINATORS, ldom)) | |
407 | if (!dominated_by_p (CDI_DOMINATORS, from, ldom)) | |
9771b263 | 408 | dom_bbs.safe_push (ldom); |
617b465c ZD |
409 | } |
410 | ||
617b465c | 411 | /* Recount dominators. */ |
66f97d31 | 412 | iterate_fix_dominators (CDI_DOMINATORS, dom_bbs, true); |
9771b263 | 413 | dom_bbs.release (); |
35b07080 ZD |
414 | free (bord_bbs); |
415 | ||
617b465c ZD |
416 | /* Fix placements of basic blocks inside loops and the placement of |
417 | loops in the loop tree. */ | |
eb2afa1a RB |
418 | fix_bb_placements (from, irred_invalidated, loop_closed_ssa_invalidated); |
419 | fix_loop_placements (from->loop_father, irred_invalidated); | |
dc14f191 | 420 | |
eb2afa1a | 421 | if (local_irred_invalidated |
f87000d0 | 422 | && loops_state_satisfies_p (LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS)) |
d73be268 | 423 | mark_irreducible_loops (); |
617b465c ZD |
424 | |
425 | return true; | |
426 | } | |
427 | ||
0fc822d0 | 428 | /* Creates place for a new LOOP in loops structure of FN. */ |
89f8f30f | 429 | |
a9e0d843 | 430 | void |
0fc822d0 | 431 | place_new_loop (struct function *fn, struct loop *loop) |
617b465c | 432 | { |
0fc822d0 RB |
433 | loop->num = number_of_loops (fn); |
434 | vec_safe_push (loops_for_fn (fn)->larray, loop); | |
617b465c ZD |
435 | } |
436 | ||
437 | /* Given LOOP structure with filled header and latch, find the body of the | |
d73be268 | 438 | corresponding loop and add it to loops tree. Insert the LOOP as a son of |
598ec7bd ZD |
439 | outer. */ |
440 | ||
89f8f30f | 441 | void |
d73be268 | 442 | add_loop (struct loop *loop, struct loop *outer) |
617b465c ZD |
443 | { |
444 | basic_block *bbs; | |
445 | int i, n; | |
89f8f30f | 446 | struct loop *subloop; |
d24a32a1 ZD |
447 | edge e; |
448 | edge_iterator ei; | |
d329e058 | 449 | |
617b465c | 450 | /* Add it to loop structure. */ |
0fc822d0 | 451 | place_new_loop (cfun, loop); |
598ec7bd | 452 | flow_loop_tree_node_add (outer, loop); |
617b465c ZD |
453 | |
454 | /* Find its nodes. */ | |
0cae8d31 DM |
455 | bbs = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun)); |
456 | n = get_loop_body_with_size (loop, bbs, n_basic_blocks_for_fn (cfun)); | |
617b465c ZD |
457 | |
458 | for (i = 0; i < n; i++) | |
598ec7bd | 459 | { |
89f8f30f ZD |
460 | if (bbs[i]->loop_father == outer) |
461 | { | |
462 | remove_bb_from_loops (bbs[i]); | |
463 | add_bb_to_loop (bbs[i], loop); | |
464 | continue; | |
465 | } | |
466 | ||
467 | loop->num_nodes++; | |
468 | ||
469 | /* If we find a direct subloop of OUTER, move it to LOOP. */ | |
470 | subloop = bbs[i]->loop_father; | |
9ba025a2 | 471 | if (loop_outer (subloop) == outer |
89f8f30f ZD |
472 | && subloop->header == bbs[i]) |
473 | { | |
474 | flow_loop_tree_node_remove (subloop); | |
475 | flow_loop_tree_node_add (loop, subloop); | |
476 | } | |
598ec7bd | 477 | } |
617b465c | 478 | |
d24a32a1 ZD |
479 | /* Update the information about loop exit edges. */ |
480 | for (i = 0; i < n; i++) | |
481 | { | |
482 | FOR_EACH_EDGE (e, ei, bbs[i]->succs) | |
483 | { | |
484 | rescan_loop_exit (e, false, false); | |
485 | } | |
486 | } | |
487 | ||
617b465c ZD |
488 | free (bbs); |
489 | } | |
490 | ||
af2bbc51 | 491 | /* Scale profile of loop by P. */ |
6e73e84b | 492 | |
03cb2019 | 493 | void |
af2bbc51 | 494 | scale_loop_frequencies (struct loop *loop, profile_probability p) |
617b465c ZD |
495 | { |
496 | basic_block *bbs; | |
497 | ||
498 | bbs = get_loop_body (loop); | |
af2bbc51 | 499 | scale_bbs_frequencies (bbs, loop->num_nodes, p); |
617b465c ZD |
500 | free (bbs); |
501 | } | |
502 | ||
af2bbc51 | 503 | /* Scale profile in LOOP by P. |
6e73e84b JH |
504 | If ITERATION_BOUND is non-zero, scale even further if loop is predicted |
505 | to iterate too many times. */ | |
506 | ||
507 | void | |
af2bbc51 JH |
508 | scale_loop_profile (struct loop *loop, profile_probability p, |
509 | gcov_type iteration_bound) | |
6e73e84b JH |
510 | { |
511 | gcov_type iterations = expected_loop_iterations_unbounded (loop); | |
512 | edge e; | |
513 | edge_iterator ei; | |
514 | ||
515 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
af2bbc51 JH |
516 | { |
517 | fprintf (dump_file, ";; Scaling loop %i with scale ", | |
518 | loop->num); | |
519 | p.dump (dump_file); | |
520 | fprintf (dump_file, " bounding iterations to %i from guessed %i\n", | |
521 | (int)iteration_bound, (int)iterations); | |
522 | } | |
6e73e84b JH |
523 | |
524 | /* See if loop is predicted to iterate too many times. */ | |
525 | if (iteration_bound && iterations > 0 | |
af2bbc51 | 526 | && p.apply (iterations) > iteration_bound) |
6e73e84b JH |
527 | { |
528 | /* Fixing loop profile for different trip count is not trivial; the exit | |
529 | probabilities has to be updated to match and frequencies propagated down | |
530 | to the loop body. | |
531 | ||
532 | We fully update only the simple case of loop with single exit that is | |
533 | either from the latch or BB just before latch and leads from BB with | |
534 | simple conditional jump. This is OK for use in vectorizer. */ | |
535 | e = single_exit (loop); | |
536 | if (e) | |
537 | { | |
538 | edge other_e; | |
3995f3a2 | 539 | profile_count count_delta; |
6e73e84b JH |
540 | |
541 | FOR_EACH_EDGE (other_e, ei, e->src->succs) | |
542 | if (!(other_e->flags & (EDGE_ABNORMAL | EDGE_FAKE)) | |
543 | && e != other_e) | |
544 | break; | |
545 | ||
546 | /* Probability of exit must be 1/iterations. */ | |
ef30ab83 | 547 | count_delta = e->count (); |
b2ff44a8 JH |
548 | e->probability = profile_probability::always () |
549 | .apply_scale (1, iteration_bound); | |
357067f2 | 550 | other_e->probability = e->probability.invert (); |
ef30ab83 | 551 | count_delta -= e->count (); |
6e73e84b | 552 | |
e7a74006 | 553 | /* If latch exists, change its count, since we changed |
6e73e84b JH |
554 | probability of exit. Theoretically we should update everything from |
555 | source of exit edge to latch, but for vectorizer this is enough. */ | |
556 | if (loop->latch | |
557 | && loop->latch != e->src) | |
558 | { | |
6e73e84b | 559 | loop->latch->count += count_delta; |
6e73e84b JH |
560 | } |
561 | } | |
562 | ||
563 | /* Roughly speaking we want to reduce the loop body profile by the | |
6af801f5 | 564 | difference of loop iterations. We however can do better if |
6e73e84b | 565 | we look at the actual profile, if it is available. */ |
af2bbc51 | 566 | p = p.apply_scale (iteration_bound, iterations); |
3995f3a2 | 567 | |
3995f3a2 | 568 | if (loop->header->count.initialized_p ()) |
6e73e84b | 569 | { |
3995f3a2 | 570 | profile_count count_in = profile_count::zero (); |
6e73e84b JH |
571 | |
572 | FOR_EACH_EDGE (e, ei, loop->header->preds) | |
573 | if (e->src != loop->latch) | |
ef30ab83 | 574 | count_in += e->count (); |
6e73e84b | 575 | |
3995f3a2 JH |
576 | if (count_in > profile_count::zero () ) |
577 | { | |
af2bbc51 JH |
578 | p = count_in.probability_in (loop->header->count.apply_scale |
579 | (iteration_bound, 1)); | |
3995f3a2 | 580 | } |
6e73e84b | 581 | } |
af2bbc51 JH |
582 | if (!(p > profile_probability::never ())) |
583 | p = profile_probability::very_unlikely (); | |
6e73e84b JH |
584 | } |
585 | ||
af2bbc51 JH |
586 | if (p >= profile_probability::always () |
587 | || !p.initialized_p ()) | |
6e73e84b JH |
588 | return; |
589 | ||
590 | /* Scale the actual probabilities. */ | |
af2bbc51 | 591 | scale_loop_frequencies (loop, p); |
6e73e84b JH |
592 | if (dump_file && (dump_flags & TDF_DETAILS)) |
593 | fprintf (dump_file, ";; guessed iterations are now %i\n", | |
594 | (int)expected_loop_iterations_unbounded (loop)); | |
595 | } | |
596 | ||
f8bf9252 SP |
597 | /* Recompute dominance information for basic blocks outside LOOP. */ |
598 | ||
599 | static void | |
600 | update_dominators_in_loop (struct loop *loop) | |
601 | { | |
6e1aa848 | 602 | vec<basic_block> dom_bbs = vNULL; |
f8bf9252 SP |
603 | basic_block *body; |
604 | unsigned i; | |
605 | ||
7ba9e72d | 606 | auto_sbitmap seen (last_basic_block_for_fn (cfun)); |
f61e445a | 607 | bitmap_clear (seen); |
f8bf9252 SP |
608 | body = get_loop_body (loop); |
609 | ||
610 | for (i = 0; i < loop->num_nodes; i++) | |
d7c028c0 | 611 | bitmap_set_bit (seen, body[i]->index); |
f8bf9252 SP |
612 | |
613 | for (i = 0; i < loop->num_nodes; i++) | |
614 | { | |
615 | basic_block ldom; | |
616 | ||
617 | for (ldom = first_dom_son (CDI_DOMINATORS, body[i]); | |
618 | ldom; | |
619 | ldom = next_dom_son (CDI_DOMINATORS, ldom)) | |
d7c028c0 | 620 | if (!bitmap_bit_p (seen, ldom->index)) |
f8bf9252 | 621 | { |
d7c028c0 | 622 | bitmap_set_bit (seen, ldom->index); |
9771b263 | 623 | dom_bbs.safe_push (ldom); |
f8bf9252 SP |
624 | } |
625 | } | |
626 | ||
627 | iterate_fix_dominators (CDI_DOMINATORS, dom_bbs, false); | |
628 | free (body); | |
9771b263 | 629 | dom_bbs.release (); |
f8bf9252 SP |
630 | } |
631 | ||
632 | /* Creates an if region as shown above. CONDITION is used to create | |
b8698a0f | 633 | the test for the if. |
f8bf9252 SP |
634 | |
635 | | | |
636 | | ------------- ------------- | |
637 | | | pred_bb | | pred_bb | | |
638 | | ------------- ------------- | |
639 | | | | | |
640 | | | | ENTRY_EDGE | |
641 | | | ENTRY_EDGE V | |
642 | | | ====> ------------- | |
643 | | | | cond_bb | | |
644 | | | | CONDITION | | |
645 | | | ------------- | |
646 | | V / \ | |
647 | | ------------- e_false / \ e_true | |
648 | | | succ_bb | V V | |
649 | | ------------- ----------- ----------- | |
650 | | | false_bb | | true_bb | | |
651 | | ----------- ----------- | |
652 | | \ / | |
653 | | \ / | |
654 | | V V | |
655 | | ------------- | |
656 | | | join_bb | | |
657 | | ------------- | |
658 | | | exit_edge (result) | |
659 | | V | |
660 | | ----------- | |
661 | | | succ_bb | | |
662 | | ----------- | |
663 | | | |
664 | */ | |
665 | ||
666 | edge | |
667 | create_empty_if_region_on_edge (edge entry_edge, tree condition) | |
668 | { | |
669 | ||
f1ed99cd | 670 | basic_block cond_bb, true_bb, false_bb, join_bb; |
f8bf9252 | 671 | edge e_true, e_false, exit_edge; |
538dd0b7 | 672 | gcond *cond_stmt; |
f8bf9252 SP |
673 | tree simple_cond; |
674 | gimple_stmt_iterator gsi; | |
675 | ||
f8bf9252 | 676 | cond_bb = split_edge (entry_edge); |
b8698a0f | 677 | |
f8bf9252 SP |
678 | /* Insert condition in cond_bb. */ |
679 | gsi = gsi_last_bb (cond_bb); | |
680 | simple_cond = | |
681 | force_gimple_operand_gsi (&gsi, condition, true, NULL, | |
682 | false, GSI_NEW_STMT); | |
683 | cond_stmt = gimple_build_cond_from_tree (simple_cond, NULL_TREE, NULL_TREE); | |
684 | gsi = gsi_last_bb (cond_bb); | |
685 | gsi_insert_after (&gsi, cond_stmt, GSI_NEW_STMT); | |
b8698a0f | 686 | |
f8bf9252 SP |
687 | join_bb = split_edge (single_succ_edge (cond_bb)); |
688 | ||
689 | e_true = single_succ_edge (cond_bb); | |
690 | true_bb = split_edge (e_true); | |
691 | ||
692 | e_false = make_edge (cond_bb, join_bb, 0); | |
693 | false_bb = split_edge (e_false); | |
694 | ||
695 | e_true->flags &= ~EDGE_FALLTHRU; | |
696 | e_true->flags |= EDGE_TRUE_VALUE; | |
697 | e_false->flags &= ~EDGE_FALLTHRU; | |
698 | e_false->flags |= EDGE_FALSE_VALUE; | |
699 | ||
700 | set_immediate_dominator (CDI_DOMINATORS, cond_bb, entry_edge->src); | |
701 | set_immediate_dominator (CDI_DOMINATORS, true_bb, cond_bb); | |
702 | set_immediate_dominator (CDI_DOMINATORS, false_bb, cond_bb); | |
703 | set_immediate_dominator (CDI_DOMINATORS, join_bb, cond_bb); | |
704 | ||
705 | exit_edge = single_succ_edge (join_bb); | |
706 | ||
707 | if (single_pred_p (exit_edge->dest)) | |
708 | set_immediate_dominator (CDI_DOMINATORS, exit_edge->dest, join_bb); | |
709 | ||
710 | return exit_edge; | |
711 | } | |
712 | ||
713 | /* create_empty_loop_on_edge | |
714 | | | |
8e74b397 SP |
715 | | - pred_bb - ------ pred_bb ------ |
716 | | | | | iv0 = initial_value | | |
717 | | -----|----- ---------|----------- | |
718 | | | ______ | entry_edge | |
719 | | | entry_edge / | | | |
720 | | | ====> | -V---V- loop_header ------------- | |
721 | | V | | iv_before = phi (iv0, iv_after) | | |
722 | | - succ_bb - | ---|----------------------------- | |
723 | | | | | | | |
724 | | ----------- | ---V--- loop_body --------------- | |
725 | | | | iv_after = iv_before + stride | | |
45e76e9f | 726 | | | | if (iv_before < upper_bound) | |
8e74b397 SP |
727 | | | ---|--------------\-------------- |
728 | | | | \ exit_e | |
729 | | | V \ | |
730 | | | - loop_latch - V- succ_bb - | |
731 | | | | | | | | |
732 | | | /------------- ----------- | |
733 | | \ ___ / | |
f8bf9252 SP |
734 | |
735 | Creates an empty loop as shown above, the IV_BEFORE is the SSA_NAME | |
45e76e9f | 736 | that is used before the increment of IV. IV_BEFORE should be used for |
f8bf9252 | 737 | adding code to the body that uses the IV. OUTER is the outer loop in |
45e76e9f | 738 | which the new loop should be inserted. |
8e74b397 SP |
739 | |
740 | Both INITIAL_VALUE and UPPER_BOUND expressions are gimplified and | |
741 | inserted on the loop entry edge. This implies that this function | |
742 | should be used only when the UPPER_BOUND expression is a loop | |
743 | invariant. */ | |
f8bf9252 SP |
744 | |
745 | struct loop * | |
45e76e9f | 746 | create_empty_loop_on_edge (edge entry_edge, |
f8bf9252 SP |
747 | tree initial_value, |
748 | tree stride, tree upper_bound, | |
749 | tree iv, | |
750 | tree *iv_before, | |
8e74b397 | 751 | tree *iv_after, |
f8bf9252 SP |
752 | struct loop *outer) |
753 | { | |
754 | basic_block loop_header, loop_latch, succ_bb, pred_bb; | |
755 | struct loop *loop; | |
f8bf9252 | 756 | gimple_stmt_iterator gsi; |
f8bf9252 | 757 | gimple_seq stmts; |
538dd0b7 | 758 | gcond *cond_expr; |
f8bf9252 SP |
759 | tree exit_test; |
760 | edge exit_e; | |
45e76e9f | 761 | |
f8bf9252 SP |
762 | gcc_assert (entry_edge && initial_value && stride && upper_bound && iv); |
763 | ||
764 | /* Create header, latch and wire up the loop. */ | |
765 | pred_bb = entry_edge->src; | |
766 | loop_header = split_edge (entry_edge); | |
767 | loop_latch = split_edge (single_succ_edge (loop_header)); | |
768 | succ_bb = single_succ (loop_latch); | |
769 | make_edge (loop_header, succ_bb, 0); | |
770 | redirect_edge_succ_nodup (single_succ_edge (loop_latch), loop_header); | |
771 | ||
772 | /* Set immediate dominator information. */ | |
773 | set_immediate_dominator (CDI_DOMINATORS, loop_header, pred_bb); | |
774 | set_immediate_dominator (CDI_DOMINATORS, loop_latch, loop_header); | |
775 | set_immediate_dominator (CDI_DOMINATORS, succ_bb, loop_header); | |
776 | ||
777 | /* Initialize a loop structure and put it in a loop hierarchy. */ | |
778 | loop = alloc_loop (); | |
779 | loop->header = loop_header; | |
780 | loop->latch = loop_latch; | |
781 | add_loop (loop, outer); | |
782 | ||
e7a74006 | 783 | /* TODO: Fix counts. */ |
af2bbc51 | 784 | scale_loop_frequencies (loop, profile_probability::even ()); |
f8bf9252 SP |
785 | |
786 | /* Update dominators. */ | |
787 | update_dominators_in_loop (loop); | |
788 | ||
8e74b397 SP |
789 | /* Modify edge flags. */ |
790 | exit_e = single_exit (loop); | |
791 | exit_e->flags = EDGE_LOOP_EXIT | EDGE_FALSE_VALUE; | |
792 | single_pred_edge (loop_latch)->flags = EDGE_TRUE_VALUE; | |
793 | ||
f8bf9252 SP |
794 | /* Construct IV code in loop. */ |
795 | initial_value = force_gimple_operand (initial_value, &stmts, true, iv); | |
796 | if (stmts) | |
797 | { | |
798 | gsi_insert_seq_on_edge (loop_preheader_edge (loop), stmts); | |
799 | gsi_commit_edge_inserts (); | |
800 | } | |
801 | ||
8e74b397 SP |
802 | upper_bound = force_gimple_operand (upper_bound, &stmts, true, NULL); |
803 | if (stmts) | |
804 | { | |
805 | gsi_insert_seq_on_edge (loop_preheader_edge (loop), stmts); | |
806 | gsi_commit_edge_inserts (); | |
807 | } | |
f8bf9252 | 808 | |
8e74b397 SP |
809 | gsi = gsi_last_bb (loop_header); |
810 | create_iv (initial_value, stride, iv, loop, &gsi, false, | |
811 | iv_before, iv_after); | |
f8bf9252 | 812 | |
8e74b397 SP |
813 | /* Insert loop exit condition. */ |
814 | cond_expr = gimple_build_cond | |
45e76e9f | 815 | (LT_EXPR, *iv_before, upper_bound, NULL_TREE, NULL_TREE); |
f8bf9252 SP |
816 | |
817 | exit_test = gimple_cond_lhs (cond_expr); | |
818 | exit_test = force_gimple_operand_gsi (&gsi, exit_test, true, NULL, | |
819 | false, GSI_NEW_STMT); | |
820 | gimple_cond_set_lhs (cond_expr, exit_test); | |
821 | gsi = gsi_last_bb (exit_e->src); | |
822 | gsi_insert_after (&gsi, cond_expr, GSI_NEW_STMT); | |
823 | ||
8e74b397 SP |
824 | split_block_after_labels (loop_header); |
825 | ||
f8bf9252 SP |
826 | return loop; |
827 | } | |
828 | ||
617b465c | 829 | /* Make area between HEADER_EDGE and LATCH_EDGE a loop by connecting |
d73be268 | 830 | latch to header and update loop tree and dominators |
617b465c ZD |
831 | accordingly. Everything between them plus LATCH_EDGE destination must |
832 | be dominated by HEADER_EDGE destination, and back-reachable from | |
833 | LATCH_EDGE source. HEADER_EDGE is redirected to basic block SWITCH_BB, | |
5132abc2 KH |
834 | FALSE_EDGE of SWITCH_BB to original destination of HEADER_EDGE and |
835 | TRUE_EDGE of SWITCH_BB to original destination of LATCH_EDGE. | |
03cb2019 ZD |
836 | Returns the newly created loop. Frequencies and counts in the new loop |
837 | are scaled by FALSE_SCALE and in the old one by TRUE_SCALE. */ | |
50654f6c | 838 | |
617b465c | 839 | struct loop * |
d73be268 | 840 | loopify (edge latch_edge, edge header_edge, |
5132abc2 | 841 | basic_block switch_bb, edge true_edge, edge false_edge, |
af2bbc51 JH |
842 | bool redirect_all_edges, profile_probability true_scale, |
843 | profile_probability false_scale) | |
617b465c ZD |
844 | { |
845 | basic_block succ_bb = latch_edge->dest; | |
846 | basic_block pred_bb = header_edge->src; | |
6270df4c | 847 | struct loop *loop = alloc_loop (); |
9ba025a2 | 848 | struct loop *outer = loop_outer (succ_bb->loop_father); |
3995f3a2 | 849 | profile_count cnt; |
617b465c ZD |
850 | |
851 | loop->header = header_edge->dest; | |
852 | loop->latch = latch_edge->src; | |
853 | ||
ef30ab83 | 854 | cnt = header_edge->count (); |
617b465c ZD |
855 | |
856 | /* Redirect edges. */ | |
857 | loop_redirect_edge (latch_edge, loop->header); | |
5132abc2 | 858 | loop_redirect_edge (true_edge, succ_bb); |
50654f6c | 859 | |
92fc4a2f ZD |
860 | /* During loop versioning, one of the switch_bb edge is already properly |
861 | set. Do not redirect it again unless redirect_all_edges is true. */ | |
862 | if (redirect_all_edges) | |
863 | { | |
864 | loop_redirect_edge (header_edge, switch_bb); | |
c22cacf3 MS |
865 | loop_redirect_edge (false_edge, loop->header); |
866 | ||
92fc4a2f ZD |
867 | /* Update dominators. */ |
868 | set_immediate_dominator (CDI_DOMINATORS, switch_bb, pred_bb); | |
869 | set_immediate_dominator (CDI_DOMINATORS, loop->header, switch_bb); | |
870 | } | |
50654f6c | 871 | |
d47cc544 | 872 | set_immediate_dominator (CDI_DOMINATORS, succ_bb, switch_bb); |
617b465c ZD |
873 | |
874 | /* Compute new loop. */ | |
d73be268 | 875 | add_loop (loop, outer); |
617b465c ZD |
876 | |
877 | /* Add switch_bb to appropriate loop. */ | |
598ec7bd ZD |
878 | if (switch_bb->loop_father) |
879 | remove_bb_from_loops (switch_bb); | |
617b465c ZD |
880 | add_bb_to_loop (switch_bb, outer); |
881 | ||
e7a74006 | 882 | /* Fix counts. */ |
03cb2019 ZD |
883 | if (redirect_all_edges) |
884 | { | |
03cb2019 | 885 | switch_bb->count = cnt; |
03cb2019 | 886 | } |
af2bbc51 JH |
887 | scale_loop_frequencies (loop, false_scale); |
888 | scale_loop_frequencies (succ_bb->loop_father, true_scale); | |
f8bf9252 | 889 | update_dominators_in_loop (loop); |
617b465c ZD |
890 | |
891 | return loop; | |
892 | } | |
893 | ||
d73be268 | 894 | /* Remove the latch edge of a LOOP and update loops to indicate that |
35b07080 | 895 | the LOOP was removed. After this function, original loop latch will |
dc14f191 ZD |
896 | have no successor, which caller is expected to fix somehow. |
897 | ||
898 | If this may cause the information about irreducible regions to become | |
1a7de201 JH |
899 | invalid, IRRED_INVALIDATED is set to true. |
900 | ||
901 | LOOP_CLOSED_SSA_INVALIDATED, if non-NULL, is a bitmap where we store | |
902 | basic blocks that had non-trivial update on their loop_father.*/ | |
dc14f191 | 903 | |
b7442c2f | 904 | void |
1a7de201 JH |
905 | unloop (struct loop *loop, bool *irred_invalidated, |
906 | bitmap loop_closed_ssa_invalidated) | |
35b07080 ZD |
907 | { |
908 | basic_block *body; | |
909 | struct loop *ploop; | |
910 | unsigned i, n; | |
911 | basic_block latch = loop->latch; | |
dc14f191 ZD |
912 | bool dummy = false; |
913 | ||
914 | if (loop_preheader_edge (loop)->flags & EDGE_IRREDUCIBLE_LOOP) | |
915 | *irred_invalidated = true; | |
35b07080 | 916 | |
e0bb17a8 | 917 | /* This is relatively straightforward. The dominators are unchanged, as |
35b07080 ZD |
918 | loop header dominates loop latch, so the only thing we have to care of |
919 | is the placement of loops and basic blocks inside the loop tree. We | |
920 | move them all to the loop->outer, and then let fix_bb_placements do | |
921 | its work. */ | |
922 | ||
923 | body = get_loop_body (loop); | |
35b07080 ZD |
924 | n = loop->num_nodes; |
925 | for (i = 0; i < n; i++) | |
926 | if (body[i]->loop_father == loop) | |
927 | { | |
928 | remove_bb_from_loops (body[i]); | |
9ba025a2 | 929 | add_bb_to_loop (body[i], loop_outer (loop)); |
35b07080 | 930 | } |
c3284718 | 931 | free (body); |
35b07080 ZD |
932 | |
933 | while (loop->inner) | |
934 | { | |
935 | ploop = loop->inner; | |
936 | flow_loop_tree_node_remove (ploop); | |
9ba025a2 | 937 | flow_loop_tree_node_add (loop_outer (loop), ploop); |
35b07080 ZD |
938 | } |
939 | ||
940 | /* Remove the loop and free its data. */ | |
42fd6772 | 941 | delete_loop (loop); |
35b07080 | 942 | |
c5cbcccf | 943 | remove_edge (single_succ_edge (latch)); |
dc14f191 ZD |
944 | |
945 | /* We do not pass IRRED_INVALIDATED to fix_bb_placements here, as even if | |
946 | there is an irreducible region inside the cancelled loop, the flags will | |
947 | be still correct. */ | |
1a7de201 | 948 | fix_bb_placements (latch, &dummy, loop_closed_ssa_invalidated); |
35b07080 ZD |
949 | } |
950 | ||
617b465c ZD |
951 | /* Fix placement of superloops of LOOP inside loop tree, i.e. ensure that |
952 | condition stated in description of fix_loop_placement holds for them. | |
953 | It is used in case when we removed some edges coming out of LOOP, which | |
dc14f191 | 954 | may cause the right placement of LOOP inside loop tree to change. |
b8698a0f | 955 | |
dc14f191 ZD |
956 | IRRED_INVALIDATED is set to true if a change in the loop structures might |
957 | invalidate the information about irreducible regions. */ | |
958 | ||
617b465c | 959 | static void |
d73be268 | 960 | fix_loop_placements (struct loop *loop, bool *irred_invalidated) |
617b465c ZD |
961 | { |
962 | struct loop *outer; | |
963 | ||
9ba025a2 | 964 | while (loop_outer (loop)) |
617b465c | 965 | { |
9ba025a2 | 966 | outer = loop_outer (loop); |
1bd3f750 | 967 | if (!fix_loop_placement (loop, irred_invalidated)) |
c22cacf3 | 968 | break; |
1548580c EB |
969 | |
970 | /* Changing the placement of a loop in the loop tree may alter the | |
971 | validity of condition 2) of the description of fix_bb_placement | |
972 | for its preheader, because the successor is the header and belongs | |
973 | to the loop. So call fix_bb_placements to fix up the placement | |
974 | of the preheader and (possibly) of its predecessors. */ | |
d73be268 | 975 | fix_bb_placements (loop_preheader_edge (loop)->src, |
1a7de201 | 976 | irred_invalidated, NULL); |
617b465c ZD |
977 | loop = outer; |
978 | } | |
979 | } | |
980 | ||
bf45c4c0 JH |
981 | /* Duplicate loop bounds and other information we store about |
982 | the loop into its duplicate. */ | |
983 | ||
984 | void | |
985 | copy_loop_info (struct loop *loop, struct loop *target) | |
986 | { | |
987 | gcc_checking_assert (!target->any_upper_bound && !target->any_estimate); | |
988 | target->any_upper_bound = loop->any_upper_bound; | |
989 | target->nb_iterations_upper_bound = loop->nb_iterations_upper_bound; | |
105e29c5 JH |
990 | target->any_likely_upper_bound = loop->any_likely_upper_bound; |
991 | target->nb_iterations_likely_upper_bound | |
992 | = loop->nb_iterations_likely_upper_bound; | |
bf45c4c0 JH |
993 | target->any_estimate = loop->any_estimate; |
994 | target->nb_iterations_estimate = loop->nb_iterations_estimate; | |
995 | target->estimate_state = loop->estimate_state; | |
18767ebc | 996 | target->constraints = loop->constraints; |
2105be5a JJ |
997 | target->warned_aggressive_loop_optimizations |
998 | |= loop->warned_aggressive_loop_optimizations; | |
886c388d | 999 | target->in_oacc_kernels_region = loop->in_oacc_kernels_region; |
bf45c4c0 JH |
1000 | } |
1001 | ||
617b465c | 1002 | /* Copies copy of LOOP as subloop of TARGET loop, placing newly |
1cc521f1 MM |
1003 | created loop into loops structure. If AFTER is non-null |
1004 | the new loop is added at AFTER->next, otherwise in front of TARGETs | |
1005 | sibling list. */ | |
f67d92e9 | 1006 | struct loop * |
1cc521f1 | 1007 | duplicate_loop (struct loop *loop, struct loop *target, struct loop *after) |
617b465c ZD |
1008 | { |
1009 | struct loop *cloop; | |
6270df4c | 1010 | cloop = alloc_loop (); |
0fc822d0 | 1011 | place_new_loop (cfun, cloop); |
bf45c4c0 JH |
1012 | |
1013 | copy_loop_info (loop, cloop); | |
617b465c | 1014 | |
99f8a411 | 1015 | /* Mark the new loop as copy of LOOP. */ |
561e8a90 | 1016 | set_loop_copy (loop, cloop); |
617b465c ZD |
1017 | |
1018 | /* Add it to target. */ | |
1cc521f1 | 1019 | flow_loop_tree_node_add (target, cloop, after); |
617b465c ZD |
1020 | |
1021 | return cloop; | |
1022 | } | |
1023 | ||
1024 | /* Copies structure of subloops of LOOP into TARGET loop, placing | |
1cc521f1 MM |
1025 | newly created loops into loop tree at the end of TARGETs sibling |
1026 | list in the original order. */ | |
48710229 | 1027 | void |
d73be268 | 1028 | duplicate_subloops (struct loop *loop, struct loop *target) |
617b465c | 1029 | { |
1cc521f1 | 1030 | struct loop *aloop, *cloop, *tail; |
617b465c | 1031 | |
1cc521f1 MM |
1032 | for (tail = target->inner; tail && tail->next; tail = tail->next) |
1033 | ; | |
617b465c ZD |
1034 | for (aloop = loop->inner; aloop; aloop = aloop->next) |
1035 | { | |
1cc521f1 MM |
1036 | cloop = duplicate_loop (aloop, target, tail); |
1037 | tail = cloop; | |
1038 | gcc_assert(!tail->next); | |
d73be268 | 1039 | duplicate_subloops (aloop, cloop); |
617b465c ZD |
1040 | } |
1041 | } | |
1042 | ||
1043 | /* Copies structure of subloops of N loops, stored in array COPIED_LOOPS, | |
1cc521f1 MM |
1044 | into TARGET loop, placing newly created loops into loop tree adding |
1045 | them to TARGETs sibling list at the end in order. */ | |
d329e058 | 1046 | static void |
d73be268 | 1047 | copy_loops_to (struct loop **copied_loops, int n, struct loop *target) |
617b465c | 1048 | { |
1cc521f1 | 1049 | struct loop *aloop, *tail; |
617b465c ZD |
1050 | int i; |
1051 | ||
1cc521f1 MM |
1052 | for (tail = target->inner; tail && tail->next; tail = tail->next) |
1053 | ; | |
617b465c ZD |
1054 | for (i = 0; i < n; i++) |
1055 | { | |
1cc521f1 MM |
1056 | aloop = duplicate_loop (copied_loops[i], target, tail); |
1057 | tail = aloop; | |
1058 | gcc_assert(!tail->next); | |
d73be268 | 1059 | duplicate_subloops (copied_loops[i], aloop); |
617b465c ZD |
1060 | } |
1061 | } | |
1062 | ||
1063 | /* Redirects edge E to basic block DEST. */ | |
1064 | static void | |
d329e058 | 1065 | loop_redirect_edge (edge e, basic_block dest) |
617b465c ZD |
1066 | { |
1067 | if (e->dest == dest) | |
1068 | return; | |
1069 | ||
9ee634e3 | 1070 | redirect_edge_and_branch_force (e, dest); |
617b465c ZD |
1071 | } |
1072 | ||
617b465c ZD |
1073 | /* Check whether LOOP's body can be duplicated. */ |
1074 | bool | |
ed7a4b4b | 1075 | can_duplicate_loop_p (const struct loop *loop) |
617b465c | 1076 | { |
8d28e87d ZD |
1077 | int ret; |
1078 | basic_block *bbs = get_loop_body (loop); | |
617b465c | 1079 | |
8d28e87d | 1080 | ret = can_copy_bbs_p (bbs, loop->num_nodes); |
617b465c | 1081 | free (bbs); |
c22cacf3 | 1082 | |
8d28e87d | 1083 | return ret; |
617b465c ZD |
1084 | } |
1085 | ||
8d28e87d | 1086 | /* Duplicates body of LOOP to given edge E NDUPL times. Takes care of updating |
1cc521f1 MM |
1087 | loop structure and dominators (order of inner subloops is retained). |
1088 | E's destination must be LOOP header for this to work, i.e. it must be entry | |
1089 | or latch edge of this loop; these are unique, as the loops must have | |
1090 | preheaders for this function to work correctly (in case E is latch, the | |
1091 | function unrolls the loop, if E is entry edge, it peels the loop). Store | |
1092 | edges created by copying ORIG edge from copies corresponding to set bits in | |
1093 | WONT_EXIT bitmap (bit 0 corresponds to original LOOP body, the other copies | |
1094 | are numbered in order given by control flow through them) into TO_REMOVE | |
1095 | array. Returns false if duplication is | |
8d28e87d | 1096 | impossible. */ |
ee8c1b05 | 1097 | |
1cb7dfc3 | 1098 | bool |
d73be268 | 1099 | duplicate_loop_to_header_edge (struct loop *loop, edge e, |
d329e058 | 1100 | unsigned int ndupl, sbitmap wont_exit, |
9771b263 | 1101 | edge orig, vec<edge> *to_remove, |
ee8c1b05 | 1102 | int flags) |
617b465c ZD |
1103 | { |
1104 | struct loop *target, *aloop; | |
1105 | struct loop **orig_loops; | |
1106 | unsigned n_orig_loops; | |
1107 | basic_block header = loop->header, latch = loop->latch; | |
1108 | basic_block *new_bbs, *bbs, *first_active; | |
1109 | basic_block new_bb, bb, first_active_latch = NULL; | |
8d28e87d ZD |
1110 | edge ae, latch_edge; |
1111 | edge spec_edges[2], new_spec_edges[2]; | |
e7b655e8 JH |
1112 | const int SE_LATCH = 0; |
1113 | const int SE_ORIG = 1; | |
617b465c ZD |
1114 | unsigned i, j, n; |
1115 | int is_latch = (latch == e->src); | |
e7b655e8 JH |
1116 | profile_probability *scale_step = NULL; |
1117 | profile_probability scale_main = profile_probability::always (); | |
1118 | profile_probability scale_act = profile_probability::always (); | |
1119 | profile_count after_exit_num = profile_count::zero (), | |
1120 | after_exit_den = profile_count::zero (); | |
1121 | bool scale_after_exit = false; | |
617b465c | 1122 | int add_irreducible_flag; |
b9a66240 | 1123 | basic_block place_after; |
03cb2019 ZD |
1124 | bitmap bbs_to_scale = NULL; |
1125 | bitmap_iterator bi; | |
617b465c | 1126 | |
341c100f NS |
1127 | gcc_assert (e->dest == loop->header); |
1128 | gcc_assert (ndupl > 0); | |
617b465c ZD |
1129 | |
1130 | if (orig) | |
1131 | { | |
1132 | /* Orig must be edge out of the loop. */ | |
341c100f NS |
1133 | gcc_assert (flow_bb_inside_loop_p (loop, orig->src)); |
1134 | gcc_assert (!flow_bb_inside_loop_p (loop, orig->dest)); | |
617b465c ZD |
1135 | } |
1136 | ||
b9a66240 ZD |
1137 | n = loop->num_nodes; |
1138 | bbs = get_loop_body_in_dom_order (loop); | |
1139 | gcc_assert (bbs[0] == loop->header); | |
1140 | gcc_assert (bbs[n - 1] == loop->latch); | |
617b465c ZD |
1141 | |
1142 | /* Check whether duplication is possible. */ | |
8d28e87d | 1143 | if (!can_copy_bbs_p (bbs, loop->num_nodes)) |
617b465c | 1144 | { |
8d28e87d ZD |
1145 | free (bbs); |
1146 | return false; | |
617b465c | 1147 | } |
5ed6ace5 | 1148 | new_bbs = XNEWVEC (basic_block, loop->num_nodes); |
617b465c | 1149 | |
8d28e87d ZD |
1150 | /* In case we are doing loop peeling and the loop is in the middle of |
1151 | irreducible region, the peeled copies will be inside it too. */ | |
1152 | add_irreducible_flag = e->flags & EDGE_IRREDUCIBLE_LOOP; | |
341c100f | 1153 | gcc_assert (!is_latch || !add_irreducible_flag); |
617b465c ZD |
1154 | |
1155 | /* Find edge from latch. */ | |
1156 | latch_edge = loop_latch_edge (loop); | |
1157 | ||
1158 | if (flags & DLTHE_FLAG_UPDATE_FREQ) | |
1159 | { | |
e7b655e8 | 1160 | /* Calculate coefficients by that we have to scale counts |
617b465c | 1161 | of duplicated loop bodies. */ |
e7b655e8 JH |
1162 | profile_count count_in = header->count; |
1163 | profile_count count_le = latch_edge->count (); | |
1164 | profile_count count_out_orig = orig ? orig->count () : count_in - count_le; | |
1165 | profile_probability prob_pass_thru = count_le.probability_in (count_in); | |
1166 | profile_probability prob_pass_wont_exit = | |
1167 | (count_le + count_out_orig).probability_in (count_in); | |
617b465c | 1168 | |
357067f2 JH |
1169 | if (orig && orig->probability.initialized_p () |
1170 | && !(orig->probability == profile_probability::always ())) | |
03cb2019 ZD |
1171 | { |
1172 | /* The blocks that are dominated by a removed exit edge ORIG have | |
1173 | frequencies scaled by this. */ | |
e7b655e8 JH |
1174 | if (orig->count ().initialized_p ()) |
1175 | { | |
1176 | after_exit_num = orig->src->count; | |
1177 | after_exit_den = after_exit_num - orig->count (); | |
1178 | scale_after_exit = true; | |
1179 | } | |
03cb2019 ZD |
1180 | bbs_to_scale = BITMAP_ALLOC (NULL); |
1181 | for (i = 0; i < n; i++) | |
1182 | { | |
1183 | if (bbs[i] != orig->src | |
1184 | && dominated_by_p (CDI_DOMINATORS, bbs[i], orig->src)) | |
1185 | bitmap_set_bit (bbs_to_scale, i); | |
1186 | } | |
1187 | } | |
1188 | ||
e7b655e8 | 1189 | scale_step = XNEWVEC (profile_probability, ndupl); |
617b465c | 1190 | |
03cb2019 | 1191 | for (i = 1; i <= ndupl; i++) |
d7c028c0 | 1192 | scale_step[i - 1] = bitmap_bit_p (wont_exit, i) |
617b465c ZD |
1193 | ? prob_pass_wont_exit |
1194 | : prob_pass_thru; | |
1195 | ||
a4d05547 | 1196 | /* Complete peeling is special as the probability of exit in last |
c22cacf3 | 1197 | copy becomes 1. */ |
178df94f JH |
1198 | if (flags & DLTHE_FLAG_COMPLETTE_PEEL) |
1199 | { | |
e7b655e8 | 1200 | profile_count wanted_count = e->count (); |
178df94f JH |
1201 | |
1202 | gcc_assert (!is_latch); | |
e7b655e8 JH |
1203 | /* First copy has count of incoming edge. Each subsequent |
1204 | count should be reduced by prob_pass_wont_exit. Caller | |
178df94f JH |
1205 | should've managed the flags so all except for original loop |
1206 | has won't exist set. */ | |
e7b655e8 | 1207 | scale_act = wanted_count.probability_in (count_in); |
178df94f JH |
1208 | /* Now simulate the duplication adjustments and compute header |
1209 | frequency of the last copy. */ | |
1210 | for (i = 0; i < ndupl; i++) | |
e7b655e8 JH |
1211 | wanted_count = wanted_count.apply_probability (scale_step [i]); |
1212 | scale_main = wanted_count.probability_in (count_in); | |
178df94f | 1213 | } |
e7b655e8 JH |
1214 | /* Here we insert loop bodies inside the loop itself (for loop unrolling). |
1215 | First iteration will be original loop followed by duplicated bodies. | |
1216 | It is necessary to scale down the original so we get right overall | |
1217 | number of iterations. */ | |
178df94f | 1218 | else if (is_latch) |
617b465c | 1219 | { |
e7b655e8 JH |
1220 | profile_probability prob_pass_main = bitmap_bit_p (wont_exit, 0) |
1221 | ? prob_pass_wont_exit | |
1222 | : prob_pass_thru; | |
1223 | profile_probability p = prob_pass_main; | |
1224 | profile_count scale_main_den = count_in; | |
617b465c ZD |
1225 | for (i = 0; i < ndupl; i++) |
1226 | { | |
e7b655e8 JH |
1227 | scale_main_den += count_in.apply_probability (p); |
1228 | p = p * scale_step[i]; | |
617b465c | 1229 | } |
e7b655e8 JH |
1230 | /* If original loop is executed COUNT_IN times, the unrolled |
1231 | loop will account SCALE_MAIN_DEN times. */ | |
1232 | scale_main = count_in.probability_in (scale_main_den); | |
1233 | scale_act = scale_main * prob_pass_main; | |
617b465c ZD |
1234 | } |
1235 | else | |
1236 | { | |
e7b655e8 | 1237 | profile_count preheader_count = e->count (); |
617b465c | 1238 | for (i = 0; i < ndupl; i++) |
e7b655e8 JH |
1239 | scale_main = scale_main * scale_step[i]; |
1240 | scale_act = preheader_count.probability_in (count_in); | |
617b465c | 1241 | } |
617b465c ZD |
1242 | } |
1243 | ||
1244 | /* Loop the new bbs will belong to. */ | |
8d28e87d | 1245 | target = e->src->loop_father; |
617b465c ZD |
1246 | |
1247 | /* Original loops. */ | |
1248 | n_orig_loops = 0; | |
1249 | for (aloop = loop->inner; aloop; aloop = aloop->next) | |
1250 | n_orig_loops++; | |
c302207e | 1251 | orig_loops = XNEWVEC (struct loop *, n_orig_loops); |
617b465c ZD |
1252 | for (aloop = loop->inner, i = 0; aloop; aloop = aloop->next, i++) |
1253 | orig_loops[i] = aloop; | |
1254 | ||
561e8a90 | 1255 | set_loop_copy (loop, target); |
d329e058 | 1256 | |
5ed6ace5 | 1257 | first_active = XNEWVEC (basic_block, n); |
617b465c ZD |
1258 | if (is_latch) |
1259 | { | |
1260 | memcpy (first_active, bbs, n * sizeof (basic_block)); | |
1261 | first_active_latch = latch; | |
1262 | } | |
1263 | ||
8d28e87d ZD |
1264 | spec_edges[SE_ORIG] = orig; |
1265 | spec_edges[SE_LATCH] = latch_edge; | |
d329e058 | 1266 | |
b9a66240 | 1267 | place_after = e->src; |
617b465c ZD |
1268 | for (j = 0; j < ndupl; j++) |
1269 | { | |
1270 | /* Copy loops. */ | |
d73be268 | 1271 | copy_loops_to (orig_loops, n_orig_loops, target); |
617b465c ZD |
1272 | |
1273 | /* Copy bbs. */ | |
b9a66240 | 1274 | copy_bbs (bbs, n, new_bbs, spec_edges, 2, new_spec_edges, loop, |
f14540b6 | 1275 | place_after, true); |
b9a66240 | 1276 | place_after = new_spec_edges[SE_LATCH]->src; |
8d28e87d | 1277 | |
7f7b1718 JH |
1278 | if (flags & DLTHE_RECORD_COPY_NUMBER) |
1279 | for (i = 0; i < n; i++) | |
1280 | { | |
1281 | gcc_assert (!new_bbs[i]->aux); | |
1282 | new_bbs[i]->aux = (void *)(size_t)(j + 1); | |
1283 | } | |
113d659a | 1284 | |
84d45ad1 ZD |
1285 | /* Note whether the blocks and edges belong to an irreducible loop. */ |
1286 | if (add_irreducible_flag) | |
1287 | { | |
1288 | for (i = 0; i < n; i++) | |
6580ee77 | 1289 | new_bbs[i]->flags |= BB_DUPLICATED; |
84d45ad1 ZD |
1290 | for (i = 0; i < n; i++) |
1291 | { | |
628f6a4e | 1292 | edge_iterator ei; |
84d45ad1 ZD |
1293 | new_bb = new_bbs[i]; |
1294 | if (new_bb->loop_father == target) | |
1295 | new_bb->flags |= BB_IRREDUCIBLE_LOOP; | |
1296 | ||
628f6a4e | 1297 | FOR_EACH_EDGE (ae, ei, new_bb->succs) |
6580ee77 | 1298 | if ((ae->dest->flags & BB_DUPLICATED) |
84d45ad1 ZD |
1299 | && (ae->src->loop_father == target |
1300 | || ae->dest->loop_father == target)) | |
1301 | ae->flags |= EDGE_IRREDUCIBLE_LOOP; | |
1302 | } | |
1303 | for (i = 0; i < n; i++) | |
6580ee77 | 1304 | new_bbs[i]->flags &= ~BB_DUPLICATED; |
84d45ad1 ZD |
1305 | } |
1306 | ||
8d28e87d | 1307 | /* Redirect the special edges. */ |
617b465c | 1308 | if (is_latch) |
8d28e87d ZD |
1309 | { |
1310 | redirect_edge_and_branch_force (latch_edge, new_bbs[0]); | |
1311 | redirect_edge_and_branch_force (new_spec_edges[SE_LATCH], | |
1312 | loop->header); | |
d47cc544 | 1313 | set_immediate_dominator (CDI_DOMINATORS, new_bbs[0], latch); |
b9a66240 | 1314 | latch = loop->latch = new_bbs[n - 1]; |
8d28e87d ZD |
1315 | e = latch_edge = new_spec_edges[SE_LATCH]; |
1316 | } | |
1317 | else | |
1318 | { | |
1319 | redirect_edge_and_branch_force (new_spec_edges[SE_LATCH], | |
1320 | loop->header); | |
1321 | redirect_edge_and_branch_force (e, new_bbs[0]); | |
d47cc544 | 1322 | set_immediate_dominator (CDI_DOMINATORS, new_bbs[0], e->src); |
8d28e87d ZD |
1323 | e = new_spec_edges[SE_LATCH]; |
1324 | } | |
617b465c | 1325 | |
8d28e87d | 1326 | /* Record exit edge in this copy. */ |
d7c028c0 | 1327 | if (orig && bitmap_bit_p (wont_exit, j + 1)) |
ee8c1b05 ZD |
1328 | { |
1329 | if (to_remove) | |
9771b263 | 1330 | to_remove->safe_push (new_spec_edges[SE_ORIG]); |
357067f2 | 1331 | force_edge_cold (new_spec_edges[SE_ORIG], true); |
03cb2019 ZD |
1332 | |
1333 | /* Scale the frequencies of the blocks dominated by the exit. */ | |
e7b655e8 | 1334 | if (bbs_to_scale && scale_after_exit) |
03cb2019 ZD |
1335 | { |
1336 | EXECUTE_IF_SET_IN_BITMAP (bbs_to_scale, 0, i, bi) | |
e7b655e8 JH |
1337 | scale_bbs_frequencies_profile_count (new_bbs + i, 1, after_exit_num, |
1338 | after_exit_den); | |
03cb2019 | 1339 | } |
ee8c1b05 | 1340 | } |
d329e058 | 1341 | |
8d28e87d ZD |
1342 | /* Record the first copy in the control flow order if it is not |
1343 | the original loop (i.e. in case of peeling). */ | |
617b465c ZD |
1344 | if (!first_active_latch) |
1345 | { | |
1346 | memcpy (first_active, new_bbs, n * sizeof (basic_block)); | |
b9a66240 | 1347 | first_active_latch = new_bbs[n - 1]; |
617b465c | 1348 | } |
d329e058 | 1349 | |
8d28e87d ZD |
1350 | /* Set counts and frequencies. */ |
1351 | if (flags & DLTHE_FLAG_UPDATE_FREQ) | |
617b465c | 1352 | { |
e7b655e8 JH |
1353 | scale_bbs_frequencies (new_bbs, n, scale_act); |
1354 | scale_act = scale_act * scale_step[j]; | |
617b465c ZD |
1355 | } |
1356 | } | |
8d28e87d ZD |
1357 | free (new_bbs); |
1358 | free (orig_loops); | |
c22cacf3 | 1359 | |
ee8c1b05 | 1360 | /* Record the exit edge in the original loop body, and update the frequencies. */ |
d7c028c0 | 1361 | if (orig && bitmap_bit_p (wont_exit, 0)) |
ee8c1b05 ZD |
1362 | { |
1363 | if (to_remove) | |
9771b263 | 1364 | to_remove->safe_push (orig); |
357067f2 | 1365 | force_edge_cold (orig, true); |
03cb2019 ZD |
1366 | |
1367 | /* Scale the frequencies of the blocks dominated by the exit. */ | |
e7b655e8 | 1368 | if (bbs_to_scale && scale_after_exit) |
03cb2019 ZD |
1369 | { |
1370 | EXECUTE_IF_SET_IN_BITMAP (bbs_to_scale, 0, i, bi) | |
e7b655e8 JH |
1371 | scale_bbs_frequencies_profile_count (bbs + i, 1, after_exit_num, |
1372 | after_exit_den); | |
03cb2019 | 1373 | } |
ee8c1b05 ZD |
1374 | } |
1375 | ||
8d28e87d ZD |
1376 | /* Update the original loop. */ |
1377 | if (!is_latch) | |
d47cc544 | 1378 | set_immediate_dominator (CDI_DOMINATORS, e->dest, e->src); |
617b465c ZD |
1379 | if (flags & DLTHE_FLAG_UPDATE_FREQ) |
1380 | { | |
e7b655e8 | 1381 | scale_bbs_frequencies (bbs, n, scale_main); |
617b465c ZD |
1382 | free (scale_step); |
1383 | } | |
617b465c | 1384 | |
8d28e87d | 1385 | /* Update dominators of outer blocks if affected. */ |
617b465c ZD |
1386 | for (i = 0; i < n; i++) |
1387 | { | |
66f97d31 | 1388 | basic_block dominated, dom_bb; |
9771b263 | 1389 | vec<basic_block> dom_bbs; |
66f97d31 | 1390 | unsigned j; |
617b465c ZD |
1391 | |
1392 | bb = bbs[i]; | |
6580ee77 | 1393 | bb->aux = 0; |
113d659a | 1394 | |
66f97d31 | 1395 | dom_bbs = get_dominated_by (CDI_DOMINATORS, bb); |
9771b263 | 1396 | FOR_EACH_VEC_ELT (dom_bbs, j, dominated) |
617b465c | 1397 | { |
617b465c ZD |
1398 | if (flow_bb_inside_loop_p (loop, dominated)) |
1399 | continue; | |
1400 | dom_bb = nearest_common_dominator ( | |
d47cc544 | 1401 | CDI_DOMINATORS, first_active[i], first_active_latch); |
c22cacf3 | 1402 | set_immediate_dominator (CDI_DOMINATORS, dominated, dom_bb); |
617b465c | 1403 | } |
9771b263 | 1404 | dom_bbs.release (); |
617b465c ZD |
1405 | } |
1406 | free (first_active); | |
1407 | ||
1408 | free (bbs); | |
03cb2019 | 1409 | BITMAP_FREE (bbs_to_scale); |
617b465c ZD |
1410 | |
1411 | return true; | |
1412 | } | |
1413 | ||
f470c378 ZD |
1414 | /* A callback for make_forwarder block, to redirect all edges except for |
1415 | MFB_KJ_EDGE to the entry part. E is the edge for that we should decide | |
1416 | whether to redirect it. */ | |
1417 | ||
b02b9b53 ZD |
1418 | edge mfb_kj_edge; |
1419 | bool | |
f470c378 ZD |
1420 | mfb_keep_just (edge e) |
1421 | { | |
1422 | return e != mfb_kj_edge; | |
1423 | } | |
1424 | ||
e855c69d AB |
1425 | /* True when a candidate preheader BLOCK has predecessors from LOOP. */ |
1426 | ||
1427 | static bool | |
1428 | has_preds_from_loop (basic_block block, struct loop *loop) | |
1429 | { | |
1430 | edge e; | |
1431 | edge_iterator ei; | |
b8698a0f | 1432 | |
e855c69d AB |
1433 | FOR_EACH_EDGE (e, ei, block->preds) |
1434 | if (e->src->loop_father == loop) | |
1435 | return true; | |
1436 | return false; | |
1437 | } | |
1438 | ||
3d436d2a ZD |
1439 | /* Creates a pre-header for a LOOP. Returns newly created block. Unless |
1440 | CP_SIMPLE_PREHEADERS is set in FLAGS, we only force LOOP to have single | |
1441 | entry; otherwise we also force preheader block to have only one successor. | |
e855c69d | 1442 | When CP_FALLTHRU_PREHEADERS is set in FLAGS, we force the preheader block |
b8698a0f | 1443 | to be a fallthru predecessor to the loop header and to have only |
e855c69d | 1444 | predecessors from outside of the loop. |
f470c378 ZD |
1445 | The function also updates dominators. */ |
1446 | ||
b02b9b53 | 1447 | basic_block |
d47cc544 | 1448 | create_preheader (struct loop *loop, int flags) |
3d436d2a | 1449 | { |
183ac6d0 | 1450 | edge e; |
3d436d2a | 1451 | basic_block dummy; |
3d436d2a | 1452 | int nentry = 0; |
f470c378 | 1453 | bool irred = false; |
c15bc84b | 1454 | bool latch_edge_was_fallthru; |
c7b852c8 | 1455 | edge one_succ_pred = NULL, single_entry = NULL; |
628f6a4e | 1456 | edge_iterator ei; |
3d436d2a | 1457 | |
628f6a4e | 1458 | FOR_EACH_EDGE (e, ei, loop->header->preds) |
3d436d2a ZD |
1459 | { |
1460 | if (e->src == loop->latch) | |
1461 | continue; | |
f470c378 | 1462 | irred |= (e->flags & EDGE_IRREDUCIBLE_LOOP) != 0; |
3d436d2a | 1463 | nentry++; |
c7b852c8 | 1464 | single_entry = e; |
c5cbcccf | 1465 | if (single_succ_p (e->src)) |
c15bc84b | 1466 | one_succ_pred = e; |
3d436d2a | 1467 | } |
341c100f | 1468 | gcc_assert (nentry); |
3d436d2a ZD |
1469 | if (nentry == 1) |
1470 | { | |
e855c69d | 1471 | bool need_forwarder_block = false; |
b8698a0f | 1472 | |
e855c69d | 1473 | /* We do not allow entry block to be the loop preheader, since we |
89f8f30f | 1474 | cannot emit code there. */ |
fefa31b5 | 1475 | if (single_entry->src == ENTRY_BLOCK_PTR_FOR_FN (cfun)) |
e855c69d AB |
1476 | need_forwarder_block = true; |
1477 | else | |
1478 | { | |
1479 | /* If we want simple preheaders, also force the preheader to have | |
1480 | just a single successor. */ | |
1481 | if ((flags & CP_SIMPLE_PREHEADERS) | |
1482 | && !single_succ_p (single_entry->src)) | |
1483 | need_forwarder_block = true; | |
1484 | /* If we want fallthru preheaders, also create forwarder block when | |
1485 | preheader ends with a jump or has predecessors from loop. */ | |
1486 | else if ((flags & CP_FALLTHRU_PREHEADERS) | |
1487 | && (JUMP_P (BB_END (single_entry->src)) | |
1488 | || has_preds_from_loop (single_entry->src, loop))) | |
1489 | need_forwarder_block = true; | |
1490 | } | |
1491 | if (! need_forwarder_block) | |
3d436d2a ZD |
1492 | return NULL; |
1493 | } | |
1494 | ||
f470c378 | 1495 | mfb_kj_edge = loop_latch_edge (loop); |
c15bc84b | 1496 | latch_edge_was_fallthru = (mfb_kj_edge->flags & EDGE_FALLTHRU) != 0; |
49574486 JJ |
1497 | if (nentry == 1 |
1498 | && ((flags & CP_FALLTHRU_PREHEADERS) == 0 | |
1499 | || (single_entry->flags & EDGE_CROSSING) == 0)) | |
183ac6d0 RB |
1500 | dummy = split_edge (single_entry); |
1501 | else | |
1502 | { | |
1503 | edge fallthru = make_forwarder_block (loop->header, mfb_keep_just, NULL); | |
1504 | dummy = fallthru->src; | |
1505 | loop->header = fallthru->dest; | |
1506 | } | |
3d436d2a | 1507 | |
c15bc84b EB |
1508 | /* Try to be clever in placing the newly created preheader. The idea is to |
1509 | avoid breaking any "fallthruness" relationship between blocks. | |
1510 | ||
1511 | The preheader was created just before the header and all incoming edges | |
1512 | to the header were redirected to the preheader, except the latch edge. | |
1513 | So the only problematic case is when this latch edge was a fallthru | |
1514 | edge: it is not anymore after the preheader creation so we have broken | |
1515 | the fallthruness. We're therefore going to look for a better place. */ | |
1516 | if (latch_edge_was_fallthru) | |
1517 | { | |
1518 | if (one_succ_pred) | |
1519 | e = one_succ_pred; | |
1520 | else | |
1521 | e = EDGE_PRED (dummy, 0); | |
1522 | ||
1523 | move_block_after (dummy, e->src); | |
1524 | } | |
f470c378 | 1525 | |
f470c378 | 1526 | if (irred) |
3d436d2a | 1527 | { |
f470c378 | 1528 | dummy->flags |= BB_IRREDUCIBLE_LOOP; |
c5cbcccf | 1529 | single_succ_edge (dummy)->flags |= EDGE_IRREDUCIBLE_LOOP; |
3d436d2a ZD |
1530 | } |
1531 | ||
c263766c RH |
1532 | if (dump_file) |
1533 | fprintf (dump_file, "Created preheader block for loop %i\n", | |
3d436d2a | 1534 | loop->num); |
b8698a0f | 1535 | |
e855c69d AB |
1536 | if (flags & CP_FALLTHRU_PREHEADERS) |
1537 | gcc_assert ((single_succ_edge (dummy)->flags & EDGE_FALLTHRU) | |
1538 | && !JUMP_P (BB_END (dummy))); | |
3d436d2a ZD |
1539 | |
1540 | return dummy; | |
1541 | } | |
1542 | ||
d73be268 ZD |
1543 | /* Create preheaders for each loop; for meaning of FLAGS see create_preheader. */ |
1544 | ||
3d436d2a | 1545 | void |
d73be268 | 1546 | create_preheaders (int flags) |
3d436d2a | 1547 | { |
42fd6772 ZD |
1548 | struct loop *loop; |
1549 | ||
c7b852c8 ZD |
1550 | if (!current_loops) |
1551 | return; | |
1552 | ||
f0bd40b1 | 1553 | FOR_EACH_LOOP (loop, 0) |
42fd6772 | 1554 | create_preheader (loop, flags); |
f87000d0 | 1555 | loops_state_set (LOOPS_HAVE_PREHEADERS); |
3d436d2a ZD |
1556 | } |
1557 | ||
d73be268 ZD |
1558 | /* Forces all loop latches to have only single successor. */ |
1559 | ||
3d436d2a | 1560 | void |
d73be268 | 1561 | force_single_succ_latches (void) |
3d436d2a | 1562 | { |
3d436d2a ZD |
1563 | struct loop *loop; |
1564 | edge e; | |
1565 | ||
f0bd40b1 | 1566 | FOR_EACH_LOOP (loop, 0) |
3d436d2a | 1567 | { |
c5cbcccf | 1568 | if (loop->latch != loop->header && single_succ_p (loop->latch)) |
3d436d2a | 1569 | continue; |
d329e058 | 1570 | |
9ff3d2de | 1571 | e = find_edge (loop->latch, loop->header); |
9dcbb380 | 1572 | gcc_checking_assert (e != NULL); |
bc810602 | 1573 | |
598ec7bd | 1574 | split_edge (e); |
3d436d2a | 1575 | } |
f87000d0 | 1576 | loops_state_set (LOOPS_HAVE_SIMPLE_LATCHES); |
3d436d2a ZD |
1577 | } |
1578 | ||
1cb7dfc3 MH |
1579 | /* This function is called from loop_version. It splits the entry edge |
1580 | of the loop we want to version, adds the versioning condition, and | |
1581 | adjust the edges to the two versions of the loop appropriately. | |
1582 | e is an incoming edge. Returns the basic block containing the | |
1583 | condition. | |
1584 | ||
1585 | --- edge e ---- > [second_head] | |
1586 | ||
1587 | Split it and insert new conditional expression and adjust edges. | |
1588 | ||
1589 | --- edge e ---> [cond expr] ---> [first_head] | |
c22cacf3 MS |
1590 | | |
1591 | +---------> [second_head] | |
03cb2019 | 1592 | |
5d3ebb71 JH |
1593 | THEN_PROB is the probability of then branch of the condition. |
1594 | ELSE_PROB is the probability of else branch. Note that they may be both | |
542e7230 BC |
1595 | REG_BR_PROB_BASE when condition is IFN_LOOP_VECTORIZED or |
1596 | IFN_LOOP_DIST_ALIAS. */ | |
1cb7dfc3 MH |
1597 | |
1598 | static basic_block | |
03cb2019 | 1599 | lv_adjust_loop_entry_edge (basic_block first_head, basic_block second_head, |
357067f2 JH |
1600 | edge e, void *cond_expr, |
1601 | profile_probability then_prob, | |
1602 | profile_probability else_prob) | |
1cb7dfc3 MH |
1603 | { |
1604 | basic_block new_head = NULL; | |
1605 | edge e1; | |
1606 | ||
1607 | gcc_assert (e->dest == second_head); | |
1608 | ||
1609 | /* Split edge 'e'. This will create a new basic block, where we can | |
1610 | insert conditional expr. */ | |
1611 | new_head = split_edge (e); | |
1612 | ||
1cb7dfc3 MH |
1613 | lv_add_condition_to_bb (first_head, second_head, new_head, |
1614 | cond_expr); | |
1615 | ||
766613a4 | 1616 | /* Don't set EDGE_TRUE_VALUE in RTL mode, as it's invalid there. */ |
03cb2019 | 1617 | e = single_succ_edge (new_head); |
52bca999 SB |
1618 | e1 = make_edge (new_head, first_head, |
1619 | current_ir_type () == IR_GIMPLE ? EDGE_TRUE_VALUE : 0); | |
03cb2019 | 1620 | e1->probability = then_prob; |
5d3ebb71 | 1621 | e->probability = else_prob; |
03cb2019 | 1622 | |
1cb7dfc3 MH |
1623 | set_immediate_dominator (CDI_DOMINATORS, first_head, new_head); |
1624 | set_immediate_dominator (CDI_DOMINATORS, second_head, new_head); | |
1625 | ||
1626 | /* Adjust loop header phi nodes. */ | |
1627 | lv_adjust_loop_header_phi (first_head, second_head, new_head, e1); | |
1628 | ||
1629 | return new_head; | |
1630 | } | |
1631 | ||
1632 | /* Main entry point for Loop Versioning transformation. | |
c22cacf3 | 1633 | |
b9a66240 ZD |
1634 | This transformation given a condition and a loop, creates |
1635 | -if (condition) { loop_copy1 } else { loop_copy2 }, | |
1636 | where loop_copy1 is the loop transformed in one way, and loop_copy2 | |
5daaf2d5 | 1637 | is the loop transformed in another way (or unchanged). COND_EXPR |
b9a66240 ZD |
1638 | may be a run time test for things that were not resolved by static |
1639 | analysis (overlapping ranges (anti-aliasing), alignment, etc.). | |
1640 | ||
5daaf2d5 AH |
1641 | If non-NULL, CONDITION_BB is set to the basic block containing the |
1642 | condition. | |
1643 | ||
03cb2019 ZD |
1644 | THEN_PROB is the probability of the then edge of the if. THEN_SCALE |
1645 | is the ratio by that the frequencies in the original loop should | |
1646 | be scaled. ELSE_SCALE is the ratio by that the frequencies in the | |
1647 | new loop should be scaled. | |
b8698a0f | 1648 | |
b9a66240 ZD |
1649 | If PLACE_AFTER is true, we place the new loop after LOOP in the |
1650 | instruction stream, otherwise it is placed before LOOP. */ | |
1cb7dfc3 MH |
1651 | |
1652 | struct loop * | |
d73be268 | 1653 | loop_version (struct loop *loop, |
b9a66240 | 1654 | void *cond_expr, basic_block *condition_bb, |
357067f2 | 1655 | profile_probability then_prob, profile_probability else_prob, |
af2bbc51 | 1656 | profile_probability then_scale, profile_probability else_scale, |
b9a66240 | 1657 | bool place_after) |
1cb7dfc3 MH |
1658 | { |
1659 | basic_block first_head, second_head; | |
6270df4c | 1660 | edge entry, latch_edge, true_edge, false_edge; |
1cb7dfc3 MH |
1661 | int irred_flag; |
1662 | struct loop *nloop; | |
b9a66240 | 1663 | basic_block cond_bb; |
1cb7dfc3 | 1664 | |
1cb7dfc3 MH |
1665 | /* Record entry and latch edges for the loop */ |
1666 | entry = loop_preheader_edge (loop); | |
1667 | irred_flag = entry->flags & EDGE_IRREDUCIBLE_LOOP; | |
1668 | entry->flags &= ~EDGE_IRREDUCIBLE_LOOP; | |
c22cacf3 | 1669 | |
1cb7dfc3 MH |
1670 | /* Note down head of loop as first_head. */ |
1671 | first_head = entry->dest; | |
1672 | ||
1673 | /* Duplicate loop. */ | |
d73be268 | 1674 | if (!cfg_hook_duplicate_loop_to_header_edge (loop, entry, 1, |
ee8c1b05 | 1675 | NULL, NULL, NULL, 0)) |
71056fef JJ |
1676 | { |
1677 | entry->flags |= irred_flag; | |
1678 | return NULL; | |
1679 | } | |
1cb7dfc3 MH |
1680 | |
1681 | /* After duplication entry edge now points to new loop head block. | |
1682 | Note down new head as second_head. */ | |
1683 | second_head = entry->dest; | |
1684 | ||
1685 | /* Split loop entry edge and insert new block with cond expr. */ | |
b9a66240 | 1686 | cond_bb = lv_adjust_loop_entry_edge (first_head, second_head, |
5d3ebb71 | 1687 | entry, cond_expr, then_prob, else_prob); |
b9a66240 ZD |
1688 | if (condition_bb) |
1689 | *condition_bb = cond_bb; | |
1690 | ||
1691 | if (!cond_bb) | |
1cb7dfc3 MH |
1692 | { |
1693 | entry->flags |= irred_flag; | |
1694 | return NULL; | |
1695 | } | |
1696 | ||
6580ee77 | 1697 | latch_edge = single_succ_edge (get_bb_copy (loop->latch)); |
c22cacf3 | 1698 | |
b9a66240 | 1699 | extract_cond_bb_edges (cond_bb, &true_edge, &false_edge); |
d73be268 | 1700 | nloop = loopify (latch_edge, |
6580ee77 | 1701 | single_pred_edge (get_bb_copy (loop->header)), |
b9a66240 | 1702 | cond_bb, true_edge, false_edge, |
03cb2019 ZD |
1703 | false /* Do not redirect all edges. */, |
1704 | then_scale, else_scale); | |
1cb7dfc3 | 1705 | |
bf45c4c0 JH |
1706 | copy_loop_info (loop, nloop); |
1707 | ||
c22cacf3 | 1708 | /* loopify redirected latch_edge. Update its PENDING_STMTS. */ |
1cb7dfc3 MH |
1709 | lv_flush_pending_stmts (latch_edge); |
1710 | ||
c22cacf3 | 1711 | /* loopify redirected condition_bb's succ edge. Update its PENDING_STMTS. */ |
b9a66240 | 1712 | extract_cond_bb_edges (cond_bb, &true_edge, &false_edge); |
1cb7dfc3 MH |
1713 | lv_flush_pending_stmts (false_edge); |
1714 | /* Adjust irreducible flag. */ | |
1715 | if (irred_flag) | |
1716 | { | |
b9a66240 | 1717 | cond_bb->flags |= BB_IRREDUCIBLE_LOOP; |
1cb7dfc3 MH |
1718 | loop_preheader_edge (loop)->flags |= EDGE_IRREDUCIBLE_LOOP; |
1719 | loop_preheader_edge (nloop)->flags |= EDGE_IRREDUCIBLE_LOOP; | |
b9a66240 ZD |
1720 | single_pred_edge (cond_bb)->flags |= EDGE_IRREDUCIBLE_LOOP; |
1721 | } | |
1722 | ||
1723 | if (place_after) | |
1724 | { | |
1725 | basic_block *bbs = get_loop_body_in_dom_order (nloop), after; | |
1726 | unsigned i; | |
1727 | ||
1728 | after = loop->latch; | |
1729 | ||
1730 | for (i = 0; i < nloop->num_nodes; i++) | |
1731 | { | |
1732 | move_block_after (bbs[i], after); | |
1733 | after = bbs[i]; | |
1734 | } | |
1735 | free (bbs); | |
1cb7dfc3 MH |
1736 | } |
1737 | ||
fa10beec RW |
1738 | /* At this point condition_bb is loop preheader with two successors, |
1739 | first_head and second_head. Make sure that loop preheader has only | |
1cb7dfc3 | 1740 | one successor. */ |
598ec7bd ZD |
1741 | split_edge (loop_preheader_edge (loop)); |
1742 | split_edge (loop_preheader_edge (nloop)); | |
1cb7dfc3 MH |
1743 | |
1744 | return nloop; | |
1745 | } |