]>
Commit | Line | Data |
---|---|---|
3d436d2a | 1 | /* Loop manipulation code for GNU compiler. |
7adcbafe | 2 | Copyright (C) 2002-2022 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 | |
99b1c316 MS |
35 | static void copy_loops_to (class loop **, int, |
36 | class loop *); | |
d329e058 | 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 **); |
99b1c316 | 41 | static void fix_loop_placements (class loop *, bool *); |
d73be268 | 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; |
99b1c316 | 92 | class 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 | |
99b1c316 | 125 | fix_loop_placement (class loop *loop, bool *irred_invalidated) |
b4c1c7e3 ZD |
126 | { |
127 | unsigned i; | |
128 | edge e; | |
4b9d61f7 | 129 | auto_vec<edge> exits = get_loop_exit_edges (loop); |
99b1c316 | 130 | class loop *father = current_loops->tree_root, *act; |
b4c1c7e3 ZD |
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 | ||
b4c1c7e3 ZD |
160 | return ret; |
161 | } | |
162 | ||
617b465c | 163 | /* Fix placements of basic blocks inside loop hierarchy stored in loops; i.e. |
026c3cfd | 164 | enforce condition stated in description of fix_bb_placement. We |
617b465c ZD |
165 | start from basic block FROM that had some of its successors removed, so that |
166 | his placement no longer has to be correct, and iteratively fix placement of | |
167 | its predecessors that may change if placement of FROM changed. Also fix | |
168 | placement of subloops of FROM->loop_father, that might also be altered due | |
4d6922ee | 169 | to this change; the condition for them is similar, except that instead of |
dc14f191 | 170 | successors we consider edges coming out of the loops. |
b8698a0f | 171 | |
dc14f191 | 172 | If the changes may invalidate the information about irreducible regions, |
1a7de201 JH |
173 | IRRED_INVALIDATED is set to true. |
174 | ||
175 | If LOOP_CLOSED_SSA_INVLIDATED is non-zero then all basic blocks with | |
176 | changed loop_father are collected there. */ | |
dc14f191 | 177 | |
617b465c | 178 | static void |
d73be268 | 179 | fix_bb_placements (basic_block from, |
1a7de201 JH |
180 | bool *irred_invalidated, |
181 | bitmap loop_closed_ssa_invalidated) | |
617b465c | 182 | { |
617b465c | 183 | basic_block *queue, *qtop, *qbeg, *qend; |
99b1c316 | 184 | class loop *base_loop, *target_loop; |
617b465c ZD |
185 | edge e; |
186 | ||
187 | /* We pass through blocks back-reachable from FROM, testing whether some | |
188 | of their successors moved to outer loop. It may be necessary to | |
189 | iterate several times, but it is finite, as we stop unless we move | |
190 | the basic block up the loop structure. The whole story is a bit | |
191 | more complicated due to presence of subloops, those are moved using | |
192 | fix_loop_placement. */ | |
193 | ||
194 | base_loop = from->loop_father; | |
cec8ac0b ZD |
195 | /* If we are already in the outermost loop, the basic blocks cannot be moved |
196 | outside of it. If FROM is the header of the base loop, it cannot be moved | |
197 | outside of it, either. In both cases, we can end now. */ | |
198 | if (base_loop == current_loops->tree_root | |
199 | || from == base_loop->header) | |
617b465c ZD |
200 | return; |
201 | ||
7ba9e72d | 202 | auto_sbitmap in_queue (last_basic_block_for_fn (cfun)); |
f61e445a | 203 | bitmap_clear (in_queue); |
d7c028c0 | 204 | bitmap_set_bit (in_queue, from->index); |
617b465c | 205 | /* Prevent us from going out of the base_loop. */ |
d7c028c0 | 206 | bitmap_set_bit (in_queue, base_loop->header->index); |
617b465c | 207 | |
5ed6ace5 | 208 | queue = XNEWVEC (basic_block, base_loop->num_nodes + 1); |
617b465c ZD |
209 | qtop = queue + base_loop->num_nodes + 1; |
210 | qbeg = queue; | |
211 | qend = queue + 1; | |
212 | *qbeg = from; | |
213 | ||
214 | while (qbeg != qend) | |
215 | { | |
628f6a4e | 216 | edge_iterator ei; |
617b465c ZD |
217 | from = *qbeg; |
218 | qbeg++; | |
219 | if (qbeg == qtop) | |
220 | qbeg = queue; | |
d7c028c0 | 221 | bitmap_clear_bit (in_queue, from->index); |
617b465c ZD |
222 | |
223 | if (from->loop_father->header == from) | |
224 | { | |
225 | /* Subloop header, maybe move the loop upward. */ | |
1bd3f750 | 226 | if (!fix_loop_placement (from->loop_father, irred_invalidated)) |
617b465c | 227 | continue; |
634ee309 | 228 | target_loop = loop_outer (from->loop_father); |
467a3558 RB |
229 | if (loop_closed_ssa_invalidated) |
230 | { | |
231 | basic_block *bbs = get_loop_body (from->loop_father); | |
232 | for (unsigned i = 0; i < from->loop_father->num_nodes; ++i) | |
233 | bitmap_set_bit (loop_closed_ssa_invalidated, bbs[i]->index); | |
234 | free (bbs); | |
235 | } | |
617b465c ZD |
236 | } |
237 | else | |
238 | { | |
239 | /* Ordinary basic block. */ | |
d73be268 | 240 | if (!fix_bb_placement (from)) |
617b465c | 241 | continue; |
467a3558 | 242 | target_loop = from->loop_father; |
1a7de201 JH |
243 | if (loop_closed_ssa_invalidated) |
244 | bitmap_set_bit (loop_closed_ssa_invalidated, from->index); | |
617b465c ZD |
245 | } |
246 | ||
dc14f191 ZD |
247 | FOR_EACH_EDGE (e, ei, from->succs) |
248 | { | |
249 | if (e->flags & EDGE_IRREDUCIBLE_LOOP) | |
250 | *irred_invalidated = true; | |
251 | } | |
252 | ||
617b465c | 253 | /* Something has changed, insert predecessors into queue. */ |
628f6a4e | 254 | FOR_EACH_EDGE (e, ei, from->preds) |
617b465c ZD |
255 | { |
256 | basic_block pred = e->src; | |
99b1c316 | 257 | class loop *nca; |
617b465c | 258 | |
dc14f191 ZD |
259 | if (e->flags & EDGE_IRREDUCIBLE_LOOP) |
260 | *irred_invalidated = true; | |
261 | ||
d7c028c0 | 262 | if (bitmap_bit_p (in_queue, pred->index)) |
617b465c ZD |
263 | continue; |
264 | ||
d329e058 | 265 | /* If it is subloop, then it either was not moved, or |
617b465c ZD |
266 | the path up the loop tree from base_loop do not contain |
267 | it. */ | |
268 | nca = find_common_loop (pred->loop_father, base_loop); | |
269 | if (pred->loop_father != base_loop | |
270 | && (nca == base_loop | |
271 | || nca != pred->loop_father)) | |
272 | pred = pred->loop_father->header; | |
634ee309 | 273 | else if (!flow_loop_nested_p (target_loop, pred->loop_father)) |
617b465c | 274 | { |
634ee309 ZD |
275 | /* If PRED is already higher in the loop hierarchy than the |
276 | TARGET_LOOP to that we moved FROM, the change of the position | |
277 | of FROM does not affect the position of PRED, so there is no | |
278 | point in processing it. */ | |
617b465c ZD |
279 | continue; |
280 | } | |
281 | ||
d7c028c0 | 282 | if (bitmap_bit_p (in_queue, pred->index)) |
617b465c ZD |
283 | continue; |
284 | ||
285 | /* Schedule the basic block. */ | |
286 | *qend = pred; | |
287 | qend++; | |
288 | if (qend == qtop) | |
289 | qend = queue; | |
d7c028c0 | 290 | bitmap_set_bit (in_queue, pred->index); |
617b465c ZD |
291 | } |
292 | } | |
617b465c ZD |
293 | free (queue); |
294 | } | |
295 | ||
296 | /* Removes path beginning at edge E, i.e. remove basic blocks dominated by E | |
d73be268 ZD |
297 | and update loop structures and dominators. Return true if we were able |
298 | to remove the path, false otherwise (and nothing is affected then). */ | |
617b465c | 299 | bool |
eb2afa1a RB |
300 | remove_path (edge e, bool *irred_invalidated, |
301 | bitmap loop_closed_ssa_invalidated) | |
617b465c ZD |
302 | { |
303 | edge ae; | |
66f97d31 | 304 | basic_block *rem_bbs, *bord_bbs, from, bb; |
9771b263 | 305 | vec<basic_block> dom_bbs; |
2f697bc4 | 306 | int i, nrem, n_bord_bbs; |
eb2afa1a | 307 | bool local_irred_invalidated = false; |
06f1716b | 308 | edge_iterator ei; |
99b1c316 | 309 | class loop *l, *f; |
617b465c | 310 | |
eb2afa1a RB |
311 | if (! irred_invalidated) |
312 | irred_invalidated = &local_irred_invalidated; | |
313 | ||
14fa2cc0 | 314 | if (!can_remove_branch_p (e)) |
35b07080 ZD |
315 | return false; |
316 | ||
dc14f191 ZD |
317 | /* Keep track of whether we need to update information about irreducible |
318 | regions. This is the case if the removed area is a part of the | |
319 | irreducible region, or if the set of basic blocks that belong to a loop | |
320 | that is inside an irreducible region is changed, or if such a loop is | |
321 | removed. */ | |
322 | if (e->flags & EDGE_IRREDUCIBLE_LOOP) | |
eb2afa1a | 323 | *irred_invalidated = true; |
dc14f191 | 324 | |
35b07080 ZD |
325 | /* We need to check whether basic blocks are dominated by the edge |
326 | e, but we only have basic block dominators. This is easy to | |
327 | fix -- when e->dest has exactly one predecessor, this corresponds | |
328 | to blocks dominated by e->dest, if not, split the edge. */ | |
c5cbcccf | 329 | if (!single_pred_p (e->dest)) |
598ec7bd | 330 | e = single_pred_edge (split_edge (e)); |
35b07080 ZD |
331 | |
332 | /* It may happen that by removing path we remove one or more loops | |
333 | we belong to. In this case first unloop the loops, then proceed | |
334 | normally. We may assume that e->dest is not a header of any loop, | |
335 | as it now has exactly one predecessor. */ | |
56494762 JH |
336 | for (l = e->src->loop_father; loop_outer (l); l = f) |
337 | { | |
338 | f = loop_outer (l); | |
339 | if (dominated_by_p (CDI_DOMINATORS, l->latch, e->dest)) | |
eb2afa1a | 340 | unloop (l, irred_invalidated, loop_closed_ssa_invalidated); |
56494762 | 341 | } |
d329e058 | 342 | |
35b07080 | 343 | /* Identify the path. */ |
d47cc544 | 344 | nrem = find_path (e, &rem_bbs); |
617b465c ZD |
345 | |
346 | n_bord_bbs = 0; | |
0cae8d31 | 347 | bord_bbs = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun)); |
7ba9e72d | 348 | auto_sbitmap seen (last_basic_block_for_fn (cfun)); |
f61e445a | 349 | bitmap_clear (seen); |
617b465c ZD |
350 | |
351 | /* Find "border" hexes -- i.e. those with predecessor in removed path. */ | |
352 | for (i = 0; i < nrem; i++) | |
d7c028c0 | 353 | bitmap_set_bit (seen, rem_bbs[i]->index); |
eb2afa1a | 354 | if (!*irred_invalidated) |
06f1716b | 355 | FOR_EACH_EDGE (ae, ei, e->src->succs) |
fefa31b5 DM |
356 | if (ae != e && ae->dest != EXIT_BLOCK_PTR_FOR_FN (cfun) |
357 | && !bitmap_bit_p (seen, ae->dest->index) | |
06f1716b | 358 | && ae->flags & EDGE_IRREDUCIBLE_LOOP) |
c7b3b99f | 359 | { |
eb2afa1a | 360 | *irred_invalidated = true; |
c7b3b99f PCC |
361 | break; |
362 | } | |
363 | ||
35b07080 | 364 | for (i = 0; i < nrem; i++) |
617b465c | 365 | { |
628f6a4e | 366 | FOR_EACH_EDGE (ae, ei, rem_bbs[i]->succs) |
fefa31b5 DM |
367 | if (ae->dest != EXIT_BLOCK_PTR_FOR_FN (cfun) |
368 | && !bitmap_bit_p (seen, ae->dest->index)) | |
35b07080 | 369 | { |
d7c028c0 | 370 | bitmap_set_bit (seen, ae->dest->index); |
35b07080 | 371 | bord_bbs[n_bord_bbs++] = ae->dest; |
b8698a0f | 372 | |
dc14f191 | 373 | if (ae->flags & EDGE_IRREDUCIBLE_LOOP) |
eb2afa1a | 374 | *irred_invalidated = true; |
35b07080 | 375 | } |
617b465c | 376 | } |
617b465c ZD |
377 | |
378 | /* Remove the path. */ | |
379 | from = e->src; | |
14fa2cc0 | 380 | remove_branch (e); |
9771b263 | 381 | dom_bbs.create (0); |
617b465c ZD |
382 | |
383 | /* Cancel loops contained in the path. */ | |
384 | for (i = 0; i < nrem; i++) | |
385 | if (rem_bbs[i]->loop_father->header == rem_bbs[i]) | |
2f697bc4 | 386 | cancel_loop_tree (rem_bbs[i]->loop_father); |
598ec7bd | 387 | |
9b43d37b JJ |
388 | remove_bbs (rem_bbs, nrem); |
389 | free (rem_bbs); | |
390 | ||
35b07080 | 391 | /* Find blocks whose dominators may be affected. */ |
f61e445a | 392 | bitmap_clear (seen); |
617b465c ZD |
393 | for (i = 0; i < n_bord_bbs; i++) |
394 | { | |
d47cc544 | 395 | basic_block ldom; |
617b465c | 396 | |
d47cc544 | 397 | bb = get_immediate_dominator (CDI_DOMINATORS, bord_bbs[i]); |
d7c028c0 | 398 | if (bitmap_bit_p (seen, bb->index)) |
617b465c | 399 | continue; |
d7c028c0 | 400 | bitmap_set_bit (seen, bb->index); |
617b465c | 401 | |
d47cc544 SB |
402 | for (ldom = first_dom_son (CDI_DOMINATORS, bb); |
403 | ldom; | |
404 | ldom = next_dom_son (CDI_DOMINATORS, ldom)) | |
405 | if (!dominated_by_p (CDI_DOMINATORS, from, ldom)) | |
9771b263 | 406 | dom_bbs.safe_push (ldom); |
617b465c ZD |
407 | } |
408 | ||
617b465c | 409 | /* Recount dominators. */ |
66f97d31 | 410 | iterate_fix_dominators (CDI_DOMINATORS, dom_bbs, true); |
9771b263 | 411 | dom_bbs.release (); |
35b07080 ZD |
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. */ | |
eb2afa1a RB |
416 | fix_bb_placements (from, irred_invalidated, loop_closed_ssa_invalidated); |
417 | fix_loop_placements (from->loop_father, irred_invalidated); | |
dc14f191 | 418 | |
eb2afa1a | 419 | if (local_irred_invalidated |
f87000d0 | 420 | && loops_state_satisfies_p (LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS)) |
d73be268 | 421 | mark_irreducible_loops (); |
617b465c ZD |
422 | |
423 | return true; | |
424 | } | |
425 | ||
0fc822d0 | 426 | /* Creates place for a new LOOP in loops structure of FN. */ |
89f8f30f | 427 | |
a9e0d843 | 428 | void |
99b1c316 | 429 | place_new_loop (struct function *fn, class loop *loop) |
617b465c | 430 | { |
0fc822d0 RB |
431 | loop->num = number_of_loops (fn); |
432 | vec_safe_push (loops_for_fn (fn)->larray, loop); | |
617b465c ZD |
433 | } |
434 | ||
435 | /* Given LOOP structure with filled header and latch, find the body of the | |
d73be268 | 436 | corresponding loop and add it to loops tree. Insert the LOOP as a son of |
598ec7bd ZD |
437 | outer. */ |
438 | ||
89f8f30f | 439 | void |
99b1c316 | 440 | add_loop (class loop *loop, class loop *outer) |
617b465c ZD |
441 | { |
442 | basic_block *bbs; | |
443 | int i, n; | |
99b1c316 | 444 | class loop *subloop; |
d24a32a1 ZD |
445 | edge e; |
446 | edge_iterator ei; | |
d329e058 | 447 | |
617b465c | 448 | /* Add it to loop structure. */ |
0fc822d0 | 449 | place_new_loop (cfun, loop); |
598ec7bd | 450 | flow_loop_tree_node_add (outer, loop); |
617b465c ZD |
451 | |
452 | /* Find its nodes. */ | |
0cae8d31 DM |
453 | bbs = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun)); |
454 | n = get_loop_body_with_size (loop, bbs, n_basic_blocks_for_fn (cfun)); | |
617b465c ZD |
455 | |
456 | for (i = 0; i < n; i++) | |
598ec7bd | 457 | { |
89f8f30f ZD |
458 | if (bbs[i]->loop_father == outer) |
459 | { | |
460 | remove_bb_from_loops (bbs[i]); | |
461 | add_bb_to_loop (bbs[i], loop); | |
462 | continue; | |
463 | } | |
464 | ||
465 | loop->num_nodes++; | |
466 | ||
467 | /* If we find a direct subloop of OUTER, move it to LOOP. */ | |
468 | subloop = bbs[i]->loop_father; | |
9ba025a2 | 469 | if (loop_outer (subloop) == outer |
89f8f30f ZD |
470 | && subloop->header == bbs[i]) |
471 | { | |
472 | flow_loop_tree_node_remove (subloop); | |
473 | flow_loop_tree_node_add (loop, subloop); | |
474 | } | |
598ec7bd | 475 | } |
617b465c | 476 | |
d24a32a1 ZD |
477 | /* Update the information about loop exit edges. */ |
478 | for (i = 0; i < n; i++) | |
479 | { | |
480 | FOR_EACH_EDGE (e, ei, bbs[i]->succs) | |
481 | { | |
482 | rescan_loop_exit (e, false, false); | |
483 | } | |
484 | } | |
485 | ||
617b465c ZD |
486 | free (bbs); |
487 | } | |
488 | ||
af2bbc51 | 489 | /* Scale profile of loop by P. */ |
6e73e84b | 490 | |
03cb2019 | 491 | void |
99b1c316 | 492 | scale_loop_frequencies (class loop *loop, profile_probability p) |
617b465c ZD |
493 | { |
494 | basic_block *bbs; | |
495 | ||
496 | bbs = get_loop_body (loop); | |
af2bbc51 | 497 | scale_bbs_frequencies (bbs, loop->num_nodes, p); |
617b465c ZD |
498 | free (bbs); |
499 | } | |
500 | ||
af2bbc51 | 501 | /* Scale profile in LOOP by P. |
6e73e84b | 502 | If ITERATION_BOUND is non-zero, scale even further if loop is predicted |
65739a68 JH |
503 | to iterate too many times. |
504 | Before caling this function, preheader block profile should be already | |
505 | scaled to final count. This is necessary because loop iterations are | |
506 | determined by comparing header edge count to latch ege count and thus | |
507 | they need to be scaled synchronously. */ | |
6e73e84b JH |
508 | |
509 | void | |
99b1c316 | 510 | scale_loop_profile (class loop *loop, profile_probability p, |
af2bbc51 | 511 | gcov_type iteration_bound) |
6e73e84b | 512 | { |
65739a68 | 513 | edge e, preheader_e; |
6e73e84b JH |
514 | edge_iterator ei; |
515 | ||
516 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
af2bbc51 JH |
517 | { |
518 | fprintf (dump_file, ";; Scaling loop %i with scale ", | |
519 | loop->num); | |
520 | p.dump (dump_file); | |
65739a68 JH |
521 | fprintf (dump_file, " bounding iterations to %i\n", |
522 | (int)iteration_bound); | |
523 | } | |
524 | ||
525 | /* Scale the probabilities. */ | |
526 | scale_loop_frequencies (loop, p); | |
527 | ||
528 | if (iteration_bound == 0) | |
529 | return; | |
530 | ||
531 | gcov_type iterations = expected_loop_iterations_unbounded (loop, NULL, true); | |
532 | ||
533 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
534 | { | |
535 | fprintf (dump_file, ";; guessed iterations after scaling %i\n", | |
536 | (int)iterations); | |
af2bbc51 | 537 | } |
6e73e84b JH |
538 | |
539 | /* See if loop is predicted to iterate too many times. */ | |
65739a68 JH |
540 | if (iterations <= iteration_bound) |
541 | return; | |
542 | ||
543 | preheader_e = loop_preheader_edge (loop); | |
544 | ||
545 | /* We could handle also loops without preheaders, but bounding is | |
546 | currently used only by optimizers that have preheaders constructed. */ | |
547 | gcc_checking_assert (preheader_e); | |
548 | profile_count count_in = preheader_e->count (); | |
549 | ||
550 | if (count_in > profile_count::zero () | |
551 | && loop->header->count.initialized_p ()) | |
6e73e84b | 552 | { |
65739a68 | 553 | profile_count count_delta = profile_count::zero (); |
6e73e84b | 554 | |
6e73e84b JH |
555 | e = single_exit (loop); |
556 | if (e) | |
557 | { | |
558 | edge other_e; | |
65739a68 | 559 | FOR_EACH_EDGE (other_e, ei, e->src->succs) |
6e73e84b JH |
560 | if (!(other_e->flags & (EDGE_ABNORMAL | EDGE_FAKE)) |
561 | && e != other_e) | |
562 | break; | |
563 | ||
564 | /* Probability of exit must be 1/iterations. */ | |
ef30ab83 | 565 | count_delta = e->count (); |
b2ff44a8 | 566 | e->probability = profile_probability::always () |
65739a68 | 567 | .apply_scale (1, iteration_bound); |
357067f2 | 568 | other_e->probability = e->probability.invert (); |
6e73e84b | 569 | |
65739a68 JH |
570 | /* In code below we only handle the following two updates. */ |
571 | if (other_e->dest != loop->header | |
572 | && other_e->dest != loop->latch | |
573 | && (dump_file && (dump_flags & TDF_DETAILS))) | |
6e73e84b | 574 | { |
65739a68 JH |
575 | fprintf (dump_file, ";; giving up on update of paths from " |
576 | "exit condition to latch\n"); | |
6e73e84b JH |
577 | } |
578 | } | |
65739a68 JH |
579 | else |
580 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
581 | fprintf (dump_file, ";; Loop has multiple exit edges; " | |
582 | "giving up on exit condition update\n"); | |
6e73e84b JH |
583 | |
584 | /* Roughly speaking we want to reduce the loop body profile by the | |
6af801f5 | 585 | difference of loop iterations. We however can do better if |
6e73e84b | 586 | we look at the actual profile, if it is available. */ |
65739a68 | 587 | p = profile_probability::always (); |
6e73e84b | 588 | |
65739a68 JH |
589 | count_in = count_in.apply_scale (iteration_bound, 1); |
590 | p = count_in.probability_in (loop->header->count); | |
af2bbc51 JH |
591 | if (!(p > profile_probability::never ())) |
592 | p = profile_probability::very_unlikely (); | |
6e73e84b | 593 | |
65739a68 JH |
594 | if (p == profile_probability::always () |
595 | || !p.initialized_p ()) | |
596 | return; | |
6e73e84b | 597 | |
65739a68 JH |
598 | /* If latch exists, change its count, since we changed |
599 | probability of exit. Theoretically we should update everything from | |
600 | source of exit edge to latch, but for vectorizer this is enough. */ | |
601 | if (loop->latch && loop->latch != e->src) | |
602 | loop->latch->count += count_delta; | |
603 | ||
604 | /* Scale the probabilities. */ | |
605 | scale_loop_frequencies (loop, p); | |
606 | ||
607 | /* Change latch's count back. */ | |
608 | if (loop->latch && loop->latch != e->src) | |
609 | loop->latch->count -= count_delta; | |
610 | ||
611 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
612 | fprintf (dump_file, ";; guessed iterations are now %i\n", | |
613 | (int)expected_loop_iterations_unbounded (loop, NULL, true)); | |
614 | } | |
6e73e84b JH |
615 | } |
616 | ||
f8bf9252 SP |
617 | /* Recompute dominance information for basic blocks outside LOOP. */ |
618 | ||
619 | static void | |
99b1c316 | 620 | update_dominators_in_loop (class loop *loop) |
f8bf9252 | 621 | { |
6e1aa848 | 622 | vec<basic_block> dom_bbs = vNULL; |
f8bf9252 SP |
623 | basic_block *body; |
624 | unsigned i; | |
625 | ||
7ba9e72d | 626 | auto_sbitmap seen (last_basic_block_for_fn (cfun)); |
f61e445a | 627 | bitmap_clear (seen); |
f8bf9252 SP |
628 | body = get_loop_body (loop); |
629 | ||
630 | for (i = 0; i < loop->num_nodes; i++) | |
d7c028c0 | 631 | bitmap_set_bit (seen, body[i]->index); |
f8bf9252 SP |
632 | |
633 | for (i = 0; i < loop->num_nodes; i++) | |
634 | { | |
635 | basic_block ldom; | |
636 | ||
637 | for (ldom = first_dom_son (CDI_DOMINATORS, body[i]); | |
638 | ldom; | |
639 | ldom = next_dom_son (CDI_DOMINATORS, ldom)) | |
d7c028c0 | 640 | if (!bitmap_bit_p (seen, ldom->index)) |
f8bf9252 | 641 | { |
d7c028c0 | 642 | bitmap_set_bit (seen, ldom->index); |
9771b263 | 643 | dom_bbs.safe_push (ldom); |
f8bf9252 SP |
644 | } |
645 | } | |
646 | ||
647 | iterate_fix_dominators (CDI_DOMINATORS, dom_bbs, false); | |
648 | free (body); | |
9771b263 | 649 | dom_bbs.release (); |
f8bf9252 SP |
650 | } |
651 | ||
652 | /* Creates an if region as shown above. CONDITION is used to create | |
b8698a0f | 653 | the test for the if. |
f8bf9252 SP |
654 | |
655 | | | |
656 | | ------------- ------------- | |
657 | | | pred_bb | | pred_bb | | |
658 | | ------------- ------------- | |
659 | | | | | |
660 | | | | ENTRY_EDGE | |
661 | | | ENTRY_EDGE V | |
662 | | | ====> ------------- | |
663 | | | | cond_bb | | |
664 | | | | CONDITION | | |
665 | | | ------------- | |
666 | | V / \ | |
667 | | ------------- e_false / \ e_true | |
668 | | | succ_bb | V V | |
669 | | ------------- ----------- ----------- | |
670 | | | false_bb | | true_bb | | |
671 | | ----------- ----------- | |
672 | | \ / | |
673 | | \ / | |
674 | | V V | |
675 | | ------------- | |
676 | | | join_bb | | |
677 | | ------------- | |
678 | | | exit_edge (result) | |
679 | | V | |
680 | | ----------- | |
681 | | | succ_bb | | |
682 | | ----------- | |
683 | | | |
684 | */ | |
685 | ||
686 | edge | |
687 | create_empty_if_region_on_edge (edge entry_edge, tree condition) | |
688 | { | |
689 | ||
f1ed99cd | 690 | basic_block cond_bb, true_bb, false_bb, join_bb; |
f8bf9252 | 691 | edge e_true, e_false, exit_edge; |
538dd0b7 | 692 | gcond *cond_stmt; |
f8bf9252 SP |
693 | tree simple_cond; |
694 | gimple_stmt_iterator gsi; | |
695 | ||
f8bf9252 | 696 | cond_bb = split_edge (entry_edge); |
b8698a0f | 697 | |
f8bf9252 SP |
698 | /* Insert condition in cond_bb. */ |
699 | gsi = gsi_last_bb (cond_bb); | |
700 | simple_cond = | |
701 | force_gimple_operand_gsi (&gsi, condition, true, NULL, | |
702 | false, GSI_NEW_STMT); | |
703 | cond_stmt = gimple_build_cond_from_tree (simple_cond, NULL_TREE, NULL_TREE); | |
704 | gsi = gsi_last_bb (cond_bb); | |
705 | gsi_insert_after (&gsi, cond_stmt, GSI_NEW_STMT); | |
b8698a0f | 706 | |
f8bf9252 SP |
707 | join_bb = split_edge (single_succ_edge (cond_bb)); |
708 | ||
709 | e_true = single_succ_edge (cond_bb); | |
710 | true_bb = split_edge (e_true); | |
711 | ||
712 | e_false = make_edge (cond_bb, join_bb, 0); | |
713 | false_bb = split_edge (e_false); | |
714 | ||
715 | e_true->flags &= ~EDGE_FALLTHRU; | |
716 | e_true->flags |= EDGE_TRUE_VALUE; | |
717 | e_false->flags &= ~EDGE_FALLTHRU; | |
718 | e_false->flags |= EDGE_FALSE_VALUE; | |
719 | ||
720 | set_immediate_dominator (CDI_DOMINATORS, cond_bb, entry_edge->src); | |
721 | set_immediate_dominator (CDI_DOMINATORS, true_bb, cond_bb); | |
722 | set_immediate_dominator (CDI_DOMINATORS, false_bb, cond_bb); | |
723 | set_immediate_dominator (CDI_DOMINATORS, join_bb, cond_bb); | |
724 | ||
725 | exit_edge = single_succ_edge (join_bb); | |
726 | ||
727 | if (single_pred_p (exit_edge->dest)) | |
728 | set_immediate_dominator (CDI_DOMINATORS, exit_edge->dest, join_bb); | |
729 | ||
730 | return exit_edge; | |
731 | } | |
732 | ||
733 | /* create_empty_loop_on_edge | |
734 | | | |
8e74b397 SP |
735 | | - pred_bb - ------ pred_bb ------ |
736 | | | | | iv0 = initial_value | | |
737 | | -----|----- ---------|----------- | |
738 | | | ______ | entry_edge | |
739 | | | entry_edge / | | | |
740 | | | ====> | -V---V- loop_header ------------- | |
741 | | V | | iv_before = phi (iv0, iv_after) | | |
742 | | - succ_bb - | ---|----------------------------- | |
743 | | | | | | | |
744 | | ----------- | ---V--- loop_body --------------- | |
745 | | | | iv_after = iv_before + stride | | |
45e76e9f | 746 | | | | if (iv_before < upper_bound) | |
8e74b397 SP |
747 | | | ---|--------------\-------------- |
748 | | | | \ exit_e | |
749 | | | V \ | |
750 | | | - loop_latch - V- succ_bb - | |
751 | | | | | | | | |
752 | | | /------------- ----------- | |
753 | | \ ___ / | |
f8bf9252 SP |
754 | |
755 | Creates an empty loop as shown above, the IV_BEFORE is the SSA_NAME | |
45e76e9f | 756 | that is used before the increment of IV. IV_BEFORE should be used for |
f8bf9252 | 757 | adding code to the body that uses the IV. OUTER is the outer loop in |
45e76e9f | 758 | which the new loop should be inserted. |
8e74b397 SP |
759 | |
760 | Both INITIAL_VALUE and UPPER_BOUND expressions are gimplified and | |
761 | inserted on the loop entry edge. This implies that this function | |
762 | should be used only when the UPPER_BOUND expression is a loop | |
763 | invariant. */ | |
f8bf9252 | 764 | |
99b1c316 | 765 | class loop * |
45e76e9f | 766 | create_empty_loop_on_edge (edge entry_edge, |
f8bf9252 SP |
767 | tree initial_value, |
768 | tree stride, tree upper_bound, | |
769 | tree iv, | |
770 | tree *iv_before, | |
8e74b397 | 771 | tree *iv_after, |
99b1c316 | 772 | class loop *outer) |
f8bf9252 SP |
773 | { |
774 | basic_block loop_header, loop_latch, succ_bb, pred_bb; | |
99b1c316 | 775 | class loop *loop; |
f8bf9252 | 776 | gimple_stmt_iterator gsi; |
f8bf9252 | 777 | gimple_seq stmts; |
538dd0b7 | 778 | gcond *cond_expr; |
f8bf9252 SP |
779 | tree exit_test; |
780 | edge exit_e; | |
45e76e9f | 781 | |
f8bf9252 SP |
782 | gcc_assert (entry_edge && initial_value && stride && upper_bound && iv); |
783 | ||
784 | /* Create header, latch and wire up the loop. */ | |
785 | pred_bb = entry_edge->src; | |
786 | loop_header = split_edge (entry_edge); | |
787 | loop_latch = split_edge (single_succ_edge (loop_header)); | |
788 | succ_bb = single_succ (loop_latch); | |
789 | make_edge (loop_header, succ_bb, 0); | |
790 | redirect_edge_succ_nodup (single_succ_edge (loop_latch), loop_header); | |
791 | ||
792 | /* Set immediate dominator information. */ | |
793 | set_immediate_dominator (CDI_DOMINATORS, loop_header, pred_bb); | |
794 | set_immediate_dominator (CDI_DOMINATORS, loop_latch, loop_header); | |
795 | set_immediate_dominator (CDI_DOMINATORS, succ_bb, loop_header); | |
796 | ||
797 | /* Initialize a loop structure and put it in a loop hierarchy. */ | |
798 | loop = alloc_loop (); | |
799 | loop->header = loop_header; | |
800 | loop->latch = loop_latch; | |
801 | add_loop (loop, outer); | |
802 | ||
e7a74006 | 803 | /* TODO: Fix counts. */ |
af2bbc51 | 804 | scale_loop_frequencies (loop, profile_probability::even ()); |
f8bf9252 SP |
805 | |
806 | /* Update dominators. */ | |
807 | update_dominators_in_loop (loop); | |
808 | ||
8e74b397 SP |
809 | /* Modify edge flags. */ |
810 | exit_e = single_exit (loop); | |
811 | exit_e->flags = EDGE_LOOP_EXIT | EDGE_FALSE_VALUE; | |
812 | single_pred_edge (loop_latch)->flags = EDGE_TRUE_VALUE; | |
813 | ||
f8bf9252 SP |
814 | /* Construct IV code in loop. */ |
815 | initial_value = force_gimple_operand (initial_value, &stmts, true, iv); | |
816 | if (stmts) | |
817 | { | |
818 | gsi_insert_seq_on_edge (loop_preheader_edge (loop), stmts); | |
819 | gsi_commit_edge_inserts (); | |
820 | } | |
821 | ||
8e74b397 SP |
822 | upper_bound = force_gimple_operand (upper_bound, &stmts, true, NULL); |
823 | if (stmts) | |
824 | { | |
825 | gsi_insert_seq_on_edge (loop_preheader_edge (loop), stmts); | |
826 | gsi_commit_edge_inserts (); | |
827 | } | |
f8bf9252 | 828 | |
8e74b397 SP |
829 | gsi = gsi_last_bb (loop_header); |
830 | create_iv (initial_value, stride, iv, loop, &gsi, false, | |
831 | iv_before, iv_after); | |
f8bf9252 | 832 | |
8e74b397 SP |
833 | /* Insert loop exit condition. */ |
834 | cond_expr = gimple_build_cond | |
45e76e9f | 835 | (LT_EXPR, *iv_before, upper_bound, NULL_TREE, NULL_TREE); |
f8bf9252 SP |
836 | |
837 | exit_test = gimple_cond_lhs (cond_expr); | |
838 | exit_test = force_gimple_operand_gsi (&gsi, exit_test, true, NULL, | |
839 | false, GSI_NEW_STMT); | |
840 | gimple_cond_set_lhs (cond_expr, exit_test); | |
841 | gsi = gsi_last_bb (exit_e->src); | |
842 | gsi_insert_after (&gsi, cond_expr, GSI_NEW_STMT); | |
843 | ||
8e74b397 SP |
844 | split_block_after_labels (loop_header); |
845 | ||
f8bf9252 SP |
846 | return loop; |
847 | } | |
848 | ||
d73be268 | 849 | /* Remove the latch edge of a LOOP and update loops to indicate that |
35b07080 | 850 | the LOOP was removed. After this function, original loop latch will |
dc14f191 ZD |
851 | have no successor, which caller is expected to fix somehow. |
852 | ||
853 | If this may cause the information about irreducible regions to become | |
1a7de201 JH |
854 | invalid, IRRED_INVALIDATED is set to true. |
855 | ||
856 | LOOP_CLOSED_SSA_INVALIDATED, if non-NULL, is a bitmap where we store | |
857 | basic blocks that had non-trivial update on their loop_father.*/ | |
dc14f191 | 858 | |
b7442c2f | 859 | void |
99b1c316 | 860 | unloop (class loop *loop, bool *irred_invalidated, |
1a7de201 | 861 | bitmap loop_closed_ssa_invalidated) |
35b07080 ZD |
862 | { |
863 | basic_block *body; | |
99b1c316 | 864 | class loop *ploop; |
35b07080 ZD |
865 | unsigned i, n; |
866 | basic_block latch = loop->latch; | |
dc14f191 ZD |
867 | bool dummy = false; |
868 | ||
869 | if (loop_preheader_edge (loop)->flags & EDGE_IRREDUCIBLE_LOOP) | |
870 | *irred_invalidated = true; | |
35b07080 | 871 | |
e0bb17a8 | 872 | /* This is relatively straightforward. The dominators are unchanged, as |
35b07080 ZD |
873 | loop header dominates loop latch, so the only thing we have to care of |
874 | is the placement of loops and basic blocks inside the loop tree. We | |
875 | move them all to the loop->outer, and then let fix_bb_placements do | |
876 | its work. */ | |
877 | ||
878 | body = get_loop_body (loop); | |
35b07080 ZD |
879 | n = loop->num_nodes; |
880 | for (i = 0; i < n; i++) | |
881 | if (body[i]->loop_father == loop) | |
882 | { | |
883 | remove_bb_from_loops (body[i]); | |
9ba025a2 | 884 | add_bb_to_loop (body[i], loop_outer (loop)); |
35b07080 | 885 | } |
c3284718 | 886 | free (body); |
35b07080 ZD |
887 | |
888 | while (loop->inner) | |
889 | { | |
890 | ploop = loop->inner; | |
891 | flow_loop_tree_node_remove (ploop); | |
9ba025a2 | 892 | flow_loop_tree_node_add (loop_outer (loop), ploop); |
35b07080 ZD |
893 | } |
894 | ||
895 | /* Remove the loop and free its data. */ | |
42fd6772 | 896 | delete_loop (loop); |
35b07080 | 897 | |
c5cbcccf | 898 | remove_edge (single_succ_edge (latch)); |
dc14f191 ZD |
899 | |
900 | /* We do not pass IRRED_INVALIDATED to fix_bb_placements here, as even if | |
901 | there is an irreducible region inside the cancelled loop, the flags will | |
902 | be still correct. */ | |
1a7de201 | 903 | fix_bb_placements (latch, &dummy, loop_closed_ssa_invalidated); |
35b07080 ZD |
904 | } |
905 | ||
617b465c ZD |
906 | /* Fix placement of superloops of LOOP inside loop tree, i.e. ensure that |
907 | condition stated in description of fix_loop_placement holds for them. | |
908 | It is used in case when we removed some edges coming out of LOOP, which | |
dc14f191 | 909 | may cause the right placement of LOOP inside loop tree to change. |
b8698a0f | 910 | |
dc14f191 ZD |
911 | IRRED_INVALIDATED is set to true if a change in the loop structures might |
912 | invalidate the information about irreducible regions. */ | |
913 | ||
617b465c | 914 | static void |
99b1c316 | 915 | fix_loop_placements (class loop *loop, bool *irred_invalidated) |
617b465c | 916 | { |
99b1c316 | 917 | class loop *outer; |
617b465c | 918 | |
9ba025a2 | 919 | while (loop_outer (loop)) |
617b465c | 920 | { |
9ba025a2 | 921 | outer = loop_outer (loop); |
1bd3f750 | 922 | if (!fix_loop_placement (loop, irred_invalidated)) |
c22cacf3 | 923 | break; |
1548580c EB |
924 | |
925 | /* Changing the placement of a loop in the loop tree may alter the | |
926 | validity of condition 2) of the description of fix_bb_placement | |
927 | for its preheader, because the successor is the header and belongs | |
928 | to the loop. So call fix_bb_placements to fix up the placement | |
929 | of the preheader and (possibly) of its predecessors. */ | |
d73be268 | 930 | fix_bb_placements (loop_preheader_edge (loop)->src, |
1a7de201 | 931 | irred_invalidated, NULL); |
617b465c ZD |
932 | loop = outer; |
933 | } | |
934 | } | |
935 | ||
bf45c4c0 JH |
936 | /* Duplicate loop bounds and other information we store about |
937 | the loop into its duplicate. */ | |
938 | ||
939 | void | |
99b1c316 | 940 | copy_loop_info (class loop *loop, class loop *target) |
bf45c4c0 JH |
941 | { |
942 | gcc_checking_assert (!target->any_upper_bound && !target->any_estimate); | |
943 | target->any_upper_bound = loop->any_upper_bound; | |
944 | target->nb_iterations_upper_bound = loop->nb_iterations_upper_bound; | |
105e29c5 JH |
945 | target->any_likely_upper_bound = loop->any_likely_upper_bound; |
946 | target->nb_iterations_likely_upper_bound | |
947 | = loop->nb_iterations_likely_upper_bound; | |
bf45c4c0 JH |
948 | target->any_estimate = loop->any_estimate; |
949 | target->nb_iterations_estimate = loop->nb_iterations_estimate; | |
950 | target->estimate_state = loop->estimate_state; | |
5161ffa4 | 951 | target->safelen = loop->safelen; |
f63445e5 | 952 | target->simdlen = loop->simdlen; |
18767ebc | 953 | target->constraints = loop->constraints; |
5161ffa4 | 954 | target->can_be_parallel = loop->can_be_parallel; |
2105be5a JJ |
955 | target->warned_aggressive_loop_optimizations |
956 | |= loop->warned_aggressive_loop_optimizations; | |
5161ffa4 RB |
957 | target->dont_vectorize = loop->dont_vectorize; |
958 | target->force_vectorize = loop->force_vectorize; | |
886c388d | 959 | target->in_oacc_kernels_region = loop->in_oacc_kernels_region; |
75efe9cb | 960 | target->finite_p = loop->finite_p; |
5161ffa4 | 961 | target->unroll = loop->unroll; |
94ec37a9 | 962 | target->owned_clique = loop->owned_clique; |
bf45c4c0 JH |
963 | } |
964 | ||
617b465c | 965 | /* Copies copy of LOOP as subloop of TARGET loop, placing newly |
1cc521f1 MM |
966 | created loop into loops structure. If AFTER is non-null |
967 | the new loop is added at AFTER->next, otherwise in front of TARGETs | |
968 | sibling list. */ | |
99b1c316 MS |
969 | class loop * |
970 | duplicate_loop (class loop *loop, class loop *target, class loop *after) | |
617b465c | 971 | { |
99b1c316 | 972 | class loop *cloop; |
6270df4c | 973 | cloop = alloc_loop (); |
0fc822d0 | 974 | place_new_loop (cfun, cloop); |
bf45c4c0 JH |
975 | |
976 | copy_loop_info (loop, cloop); | |
617b465c | 977 | |
99f8a411 | 978 | /* Mark the new loop as copy of LOOP. */ |
561e8a90 | 979 | set_loop_copy (loop, cloop); |
617b465c ZD |
980 | |
981 | /* Add it to target. */ | |
1cc521f1 | 982 | flow_loop_tree_node_add (target, cloop, after); |
617b465c ZD |
983 | |
984 | return cloop; | |
985 | } | |
986 | ||
987 | /* Copies structure of subloops of LOOP into TARGET loop, placing | |
1cc521f1 MM |
988 | newly created loops into loop tree at the end of TARGETs sibling |
989 | list in the original order. */ | |
48710229 | 990 | void |
99b1c316 | 991 | duplicate_subloops (class loop *loop, class loop *target) |
617b465c | 992 | { |
99b1c316 | 993 | class loop *aloop, *cloop, *tail; |
617b465c | 994 | |
1cc521f1 MM |
995 | for (tail = target->inner; tail && tail->next; tail = tail->next) |
996 | ; | |
617b465c ZD |
997 | for (aloop = loop->inner; aloop; aloop = aloop->next) |
998 | { | |
1cc521f1 MM |
999 | cloop = duplicate_loop (aloop, target, tail); |
1000 | tail = cloop; | |
1001 | gcc_assert(!tail->next); | |
d73be268 | 1002 | duplicate_subloops (aloop, cloop); |
617b465c ZD |
1003 | } |
1004 | } | |
1005 | ||
1006 | /* Copies structure of subloops of N loops, stored in array COPIED_LOOPS, | |
1cc521f1 MM |
1007 | into TARGET loop, placing newly created loops into loop tree adding |
1008 | them to TARGETs sibling list at the end in order. */ | |
d329e058 | 1009 | static void |
99b1c316 | 1010 | copy_loops_to (class loop **copied_loops, int n, class loop *target) |
617b465c | 1011 | { |
99b1c316 | 1012 | class loop *aloop, *tail; |
617b465c ZD |
1013 | int i; |
1014 | ||
1cc521f1 MM |
1015 | for (tail = target->inner; tail && tail->next; tail = tail->next) |
1016 | ; | |
617b465c ZD |
1017 | for (i = 0; i < n; i++) |
1018 | { | |
1cc521f1 MM |
1019 | aloop = duplicate_loop (copied_loops[i], target, tail); |
1020 | tail = aloop; | |
1021 | gcc_assert(!tail->next); | |
d73be268 | 1022 | duplicate_subloops (copied_loops[i], aloop); |
617b465c ZD |
1023 | } |
1024 | } | |
1025 | ||
1026 | /* Redirects edge E to basic block DEST. */ | |
1027 | static void | |
d329e058 | 1028 | loop_redirect_edge (edge e, basic_block dest) |
617b465c ZD |
1029 | { |
1030 | if (e->dest == dest) | |
1031 | return; | |
1032 | ||
9ee634e3 | 1033 | redirect_edge_and_branch_force (e, dest); |
617b465c ZD |
1034 | } |
1035 | ||
617b465c ZD |
1036 | /* Check whether LOOP's body can be duplicated. */ |
1037 | bool | |
99b1c316 | 1038 | can_duplicate_loop_p (const class loop *loop) |
617b465c | 1039 | { |
8d28e87d ZD |
1040 | int ret; |
1041 | basic_block *bbs = get_loop_body (loop); | |
617b465c | 1042 | |
8d28e87d | 1043 | ret = can_copy_bbs_p (bbs, loop->num_nodes); |
617b465c | 1044 | free (bbs); |
c22cacf3 | 1045 | |
8d28e87d | 1046 | return ret; |
617b465c ZD |
1047 | } |
1048 | ||
8d28e87d | 1049 | /* Duplicates body of LOOP to given edge E NDUPL times. Takes care of updating |
1cc521f1 MM |
1050 | loop structure and dominators (order of inner subloops is retained). |
1051 | E's destination must be LOOP header for this to work, i.e. it must be entry | |
1052 | or latch edge of this loop; these are unique, as the loops must have | |
1053 | preheaders for this function to work correctly (in case E is latch, the | |
1054 | function unrolls the loop, if E is entry edge, it peels the loop). Store | |
1055 | edges created by copying ORIG edge from copies corresponding to set bits in | |
1056 | WONT_EXIT bitmap (bit 0 corresponds to original LOOP body, the other copies | |
1057 | are numbered in order given by control flow through them) into TO_REMOVE | |
1058 | array. Returns false if duplication is | |
8d28e87d | 1059 | impossible. */ |
ee8c1b05 | 1060 | |
1cb7dfc3 | 1061 | bool |
4851c80c XL |
1062 | duplicate_loop_body_to_header_edge (class loop *loop, edge e, |
1063 | unsigned int ndupl, sbitmap wont_exit, | |
1064 | edge orig, vec<edge> *to_remove, int flags) | |
617b465c | 1065 | { |
99b1c316 MS |
1066 | class loop *target, *aloop; |
1067 | class loop **orig_loops; | |
617b465c ZD |
1068 | unsigned n_orig_loops; |
1069 | basic_block header = loop->header, latch = loop->latch; | |
1070 | basic_block *new_bbs, *bbs, *first_active; | |
1071 | basic_block new_bb, bb, first_active_latch = NULL; | |
8d28e87d ZD |
1072 | edge ae, latch_edge; |
1073 | edge spec_edges[2], new_spec_edges[2]; | |
e7b655e8 JH |
1074 | const int SE_LATCH = 0; |
1075 | const int SE_ORIG = 1; | |
617b465c ZD |
1076 | unsigned i, j, n; |
1077 | int is_latch = (latch == e->src); | |
e7b655e8 JH |
1078 | profile_probability *scale_step = NULL; |
1079 | profile_probability scale_main = profile_probability::always (); | |
1080 | profile_probability scale_act = profile_probability::always (); | |
1081 | profile_count after_exit_num = profile_count::zero (), | |
1082 | after_exit_den = profile_count::zero (); | |
1083 | bool scale_after_exit = false; | |
617b465c | 1084 | int add_irreducible_flag; |
b9a66240 | 1085 | basic_block place_after; |
03cb2019 ZD |
1086 | bitmap bbs_to_scale = NULL; |
1087 | bitmap_iterator bi; | |
617b465c | 1088 | |
341c100f NS |
1089 | gcc_assert (e->dest == loop->header); |
1090 | gcc_assert (ndupl > 0); | |
617b465c ZD |
1091 | |
1092 | if (orig) | |
1093 | { | |
1094 | /* Orig must be edge out of the loop. */ | |
341c100f NS |
1095 | gcc_assert (flow_bb_inside_loop_p (loop, orig->src)); |
1096 | gcc_assert (!flow_bb_inside_loop_p (loop, orig->dest)); | |
617b465c ZD |
1097 | } |
1098 | ||
b9a66240 ZD |
1099 | n = loop->num_nodes; |
1100 | bbs = get_loop_body_in_dom_order (loop); | |
1101 | gcc_assert (bbs[0] == loop->header); | |
1102 | gcc_assert (bbs[n - 1] == loop->latch); | |
617b465c ZD |
1103 | |
1104 | /* Check whether duplication is possible. */ | |
8d28e87d | 1105 | if (!can_copy_bbs_p (bbs, loop->num_nodes)) |
617b465c | 1106 | { |
8d28e87d ZD |
1107 | free (bbs); |
1108 | return false; | |
617b465c | 1109 | } |
5ed6ace5 | 1110 | new_bbs = XNEWVEC (basic_block, loop->num_nodes); |
617b465c | 1111 | |
8d28e87d ZD |
1112 | /* In case we are doing loop peeling and the loop is in the middle of |
1113 | irreducible region, the peeled copies will be inside it too. */ | |
1114 | add_irreducible_flag = e->flags & EDGE_IRREDUCIBLE_LOOP; | |
341c100f | 1115 | gcc_assert (!is_latch || !add_irreducible_flag); |
617b465c ZD |
1116 | |
1117 | /* Find edge from latch. */ | |
1118 | latch_edge = loop_latch_edge (loop); | |
1119 | ||
1120 | if (flags & DLTHE_FLAG_UPDATE_FREQ) | |
1121 | { | |
e7b655e8 | 1122 | /* Calculate coefficients by that we have to scale counts |
617b465c | 1123 | of duplicated loop bodies. */ |
e7b655e8 JH |
1124 | profile_count count_in = header->count; |
1125 | profile_count count_le = latch_edge->count (); | |
1126 | profile_count count_out_orig = orig ? orig->count () : count_in - count_le; | |
1127 | profile_probability prob_pass_thru = count_le.probability_in (count_in); | |
1128 | profile_probability prob_pass_wont_exit = | |
1129 | (count_le + count_out_orig).probability_in (count_in); | |
617b465c | 1130 | |
357067f2 JH |
1131 | if (orig && orig->probability.initialized_p () |
1132 | && !(orig->probability == profile_probability::always ())) | |
03cb2019 ZD |
1133 | { |
1134 | /* The blocks that are dominated by a removed exit edge ORIG have | |
1135 | frequencies scaled by this. */ | |
e7b655e8 JH |
1136 | if (orig->count ().initialized_p ()) |
1137 | { | |
1138 | after_exit_num = orig->src->count; | |
1139 | after_exit_den = after_exit_num - orig->count (); | |
1140 | scale_after_exit = true; | |
1141 | } | |
03cb2019 ZD |
1142 | bbs_to_scale = BITMAP_ALLOC (NULL); |
1143 | for (i = 0; i < n; i++) | |
1144 | { | |
1145 | if (bbs[i] != orig->src | |
1146 | && dominated_by_p (CDI_DOMINATORS, bbs[i], orig->src)) | |
1147 | bitmap_set_bit (bbs_to_scale, i); | |
1148 | } | |
1149 | } | |
1150 | ||
e7b655e8 | 1151 | scale_step = XNEWVEC (profile_probability, ndupl); |
617b465c | 1152 | |
03cb2019 | 1153 | for (i = 1; i <= ndupl; i++) |
d7c028c0 | 1154 | scale_step[i - 1] = bitmap_bit_p (wont_exit, i) |
617b465c ZD |
1155 | ? prob_pass_wont_exit |
1156 | : prob_pass_thru; | |
1157 | ||
a4d05547 | 1158 | /* Complete peeling is special as the probability of exit in last |
c22cacf3 | 1159 | copy becomes 1. */ |
178df94f JH |
1160 | if (flags & DLTHE_FLAG_COMPLETTE_PEEL) |
1161 | { | |
e7b655e8 | 1162 | profile_count wanted_count = e->count (); |
178df94f JH |
1163 | |
1164 | gcc_assert (!is_latch); | |
e7b655e8 JH |
1165 | /* First copy has count of incoming edge. Each subsequent |
1166 | count should be reduced by prob_pass_wont_exit. Caller | |
178df94f JH |
1167 | should've managed the flags so all except for original loop |
1168 | has won't exist set. */ | |
e7b655e8 | 1169 | scale_act = wanted_count.probability_in (count_in); |
178df94f JH |
1170 | /* Now simulate the duplication adjustments and compute header |
1171 | frequency of the last copy. */ | |
1172 | for (i = 0; i < ndupl; i++) | |
e7b655e8 JH |
1173 | wanted_count = wanted_count.apply_probability (scale_step [i]); |
1174 | scale_main = wanted_count.probability_in (count_in); | |
178df94f | 1175 | } |
e7b655e8 JH |
1176 | /* Here we insert loop bodies inside the loop itself (for loop unrolling). |
1177 | First iteration will be original loop followed by duplicated bodies. | |
1178 | It is necessary to scale down the original so we get right overall | |
1179 | number of iterations. */ | |
178df94f | 1180 | else if (is_latch) |
617b465c | 1181 | { |
e7b655e8 JH |
1182 | profile_probability prob_pass_main = bitmap_bit_p (wont_exit, 0) |
1183 | ? prob_pass_wont_exit | |
1184 | : prob_pass_thru; | |
1185 | profile_probability p = prob_pass_main; | |
1186 | profile_count scale_main_den = count_in; | |
617b465c ZD |
1187 | for (i = 0; i < ndupl; i++) |
1188 | { | |
e7b655e8 JH |
1189 | scale_main_den += count_in.apply_probability (p); |
1190 | p = p * scale_step[i]; | |
617b465c | 1191 | } |
e7b655e8 JH |
1192 | /* If original loop is executed COUNT_IN times, the unrolled |
1193 | loop will account SCALE_MAIN_DEN times. */ | |
1194 | scale_main = count_in.probability_in (scale_main_den); | |
1195 | scale_act = scale_main * prob_pass_main; | |
617b465c ZD |
1196 | } |
1197 | else | |
1198 | { | |
e7b655e8 | 1199 | profile_count preheader_count = e->count (); |
617b465c | 1200 | for (i = 0; i < ndupl; i++) |
e7b655e8 JH |
1201 | scale_main = scale_main * scale_step[i]; |
1202 | scale_act = preheader_count.probability_in (count_in); | |
617b465c | 1203 | } |
617b465c ZD |
1204 | } |
1205 | ||
1206 | /* Loop the new bbs will belong to. */ | |
8d28e87d | 1207 | target = e->src->loop_father; |
617b465c ZD |
1208 | |
1209 | /* Original loops. */ | |
1210 | n_orig_loops = 0; | |
1211 | for (aloop = loop->inner; aloop; aloop = aloop->next) | |
1212 | n_orig_loops++; | |
99b1c316 | 1213 | orig_loops = XNEWVEC (class loop *, n_orig_loops); |
617b465c ZD |
1214 | for (aloop = loop->inner, i = 0; aloop; aloop = aloop->next, i++) |
1215 | orig_loops[i] = aloop; | |
1216 | ||
561e8a90 | 1217 | set_loop_copy (loop, target); |
d329e058 | 1218 | |
5ed6ace5 | 1219 | first_active = XNEWVEC (basic_block, n); |
617b465c ZD |
1220 | if (is_latch) |
1221 | { | |
1222 | memcpy (first_active, bbs, n * sizeof (basic_block)); | |
1223 | first_active_latch = latch; | |
1224 | } | |
1225 | ||
8d28e87d ZD |
1226 | spec_edges[SE_ORIG] = orig; |
1227 | spec_edges[SE_LATCH] = latch_edge; | |
d329e058 | 1228 | |
b9a66240 | 1229 | place_after = e->src; |
617b465c ZD |
1230 | for (j = 0; j < ndupl; j++) |
1231 | { | |
1232 | /* Copy loops. */ | |
d73be268 | 1233 | copy_loops_to (orig_loops, n_orig_loops, target); |
617b465c ZD |
1234 | |
1235 | /* Copy bbs. */ | |
b9a66240 | 1236 | copy_bbs (bbs, n, new_bbs, spec_edges, 2, new_spec_edges, loop, |
f14540b6 | 1237 | place_after, true); |
b9a66240 | 1238 | place_after = new_spec_edges[SE_LATCH]->src; |
8d28e87d | 1239 | |
7f7b1718 JH |
1240 | if (flags & DLTHE_RECORD_COPY_NUMBER) |
1241 | for (i = 0; i < n; i++) | |
1242 | { | |
1243 | gcc_assert (!new_bbs[i]->aux); | |
1244 | new_bbs[i]->aux = (void *)(size_t)(j + 1); | |
1245 | } | |
113d659a | 1246 | |
84d45ad1 ZD |
1247 | /* Note whether the blocks and edges belong to an irreducible loop. */ |
1248 | if (add_irreducible_flag) | |
1249 | { | |
1250 | for (i = 0; i < n; i++) | |
6580ee77 | 1251 | new_bbs[i]->flags |= BB_DUPLICATED; |
84d45ad1 ZD |
1252 | for (i = 0; i < n; i++) |
1253 | { | |
628f6a4e | 1254 | edge_iterator ei; |
84d45ad1 ZD |
1255 | new_bb = new_bbs[i]; |
1256 | if (new_bb->loop_father == target) | |
1257 | new_bb->flags |= BB_IRREDUCIBLE_LOOP; | |
1258 | ||
628f6a4e | 1259 | FOR_EACH_EDGE (ae, ei, new_bb->succs) |
6580ee77 | 1260 | if ((ae->dest->flags & BB_DUPLICATED) |
84d45ad1 ZD |
1261 | && (ae->src->loop_father == target |
1262 | || ae->dest->loop_father == target)) | |
1263 | ae->flags |= EDGE_IRREDUCIBLE_LOOP; | |
1264 | } | |
1265 | for (i = 0; i < n; i++) | |
6580ee77 | 1266 | new_bbs[i]->flags &= ~BB_DUPLICATED; |
84d45ad1 ZD |
1267 | } |
1268 | ||
8d28e87d | 1269 | /* Redirect the special edges. */ |
617b465c | 1270 | if (is_latch) |
8d28e87d ZD |
1271 | { |
1272 | redirect_edge_and_branch_force (latch_edge, new_bbs[0]); | |
1273 | redirect_edge_and_branch_force (new_spec_edges[SE_LATCH], | |
1274 | loop->header); | |
d47cc544 | 1275 | set_immediate_dominator (CDI_DOMINATORS, new_bbs[0], latch); |
b9a66240 | 1276 | latch = loop->latch = new_bbs[n - 1]; |
8d28e87d ZD |
1277 | e = latch_edge = new_spec_edges[SE_LATCH]; |
1278 | } | |
1279 | else | |
1280 | { | |
1281 | redirect_edge_and_branch_force (new_spec_edges[SE_LATCH], | |
1282 | loop->header); | |
1283 | redirect_edge_and_branch_force (e, new_bbs[0]); | |
d47cc544 | 1284 | set_immediate_dominator (CDI_DOMINATORS, new_bbs[0], e->src); |
8d28e87d ZD |
1285 | e = new_spec_edges[SE_LATCH]; |
1286 | } | |
617b465c | 1287 | |
8d28e87d | 1288 | /* Record exit edge in this copy. */ |
d7c028c0 | 1289 | if (orig && bitmap_bit_p (wont_exit, j + 1)) |
ee8c1b05 ZD |
1290 | { |
1291 | if (to_remove) | |
9771b263 | 1292 | to_remove->safe_push (new_spec_edges[SE_ORIG]); |
357067f2 | 1293 | force_edge_cold (new_spec_edges[SE_ORIG], true); |
03cb2019 ZD |
1294 | |
1295 | /* Scale the frequencies of the blocks dominated by the exit. */ | |
e7b655e8 | 1296 | if (bbs_to_scale && scale_after_exit) |
03cb2019 ZD |
1297 | { |
1298 | EXECUTE_IF_SET_IN_BITMAP (bbs_to_scale, 0, i, bi) | |
e7b655e8 JH |
1299 | scale_bbs_frequencies_profile_count (new_bbs + i, 1, after_exit_num, |
1300 | after_exit_den); | |
03cb2019 | 1301 | } |
ee8c1b05 | 1302 | } |
d329e058 | 1303 | |
8d28e87d ZD |
1304 | /* Record the first copy in the control flow order if it is not |
1305 | the original loop (i.e. in case of peeling). */ | |
617b465c ZD |
1306 | if (!first_active_latch) |
1307 | { | |
1308 | memcpy (first_active, new_bbs, n * sizeof (basic_block)); | |
b9a66240 | 1309 | first_active_latch = new_bbs[n - 1]; |
617b465c | 1310 | } |
d329e058 | 1311 | |
8d28e87d ZD |
1312 | /* Set counts and frequencies. */ |
1313 | if (flags & DLTHE_FLAG_UPDATE_FREQ) | |
617b465c | 1314 | { |
e7b655e8 JH |
1315 | scale_bbs_frequencies (new_bbs, n, scale_act); |
1316 | scale_act = scale_act * scale_step[j]; | |
617b465c ZD |
1317 | } |
1318 | } | |
8d28e87d ZD |
1319 | free (new_bbs); |
1320 | free (orig_loops); | |
c22cacf3 | 1321 | |
ee8c1b05 | 1322 | /* Record the exit edge in the original loop body, and update the frequencies. */ |
d7c028c0 | 1323 | if (orig && bitmap_bit_p (wont_exit, 0)) |
ee8c1b05 ZD |
1324 | { |
1325 | if (to_remove) | |
9771b263 | 1326 | to_remove->safe_push (orig); |
357067f2 | 1327 | force_edge_cold (orig, true); |
03cb2019 ZD |
1328 | |
1329 | /* Scale the frequencies of the blocks dominated by the exit. */ | |
e7b655e8 | 1330 | if (bbs_to_scale && scale_after_exit) |
03cb2019 ZD |
1331 | { |
1332 | EXECUTE_IF_SET_IN_BITMAP (bbs_to_scale, 0, i, bi) | |
e7b655e8 JH |
1333 | scale_bbs_frequencies_profile_count (bbs + i, 1, after_exit_num, |
1334 | after_exit_den); | |
03cb2019 | 1335 | } |
ee8c1b05 ZD |
1336 | } |
1337 | ||
8d28e87d ZD |
1338 | /* Update the original loop. */ |
1339 | if (!is_latch) | |
d47cc544 | 1340 | set_immediate_dominator (CDI_DOMINATORS, e->dest, e->src); |
617b465c ZD |
1341 | if (flags & DLTHE_FLAG_UPDATE_FREQ) |
1342 | { | |
e7b655e8 | 1343 | scale_bbs_frequencies (bbs, n, scale_main); |
617b465c ZD |
1344 | free (scale_step); |
1345 | } | |
617b465c | 1346 | |
8d28e87d | 1347 | /* Update dominators of outer blocks if affected. */ |
617b465c ZD |
1348 | for (i = 0; i < n; i++) |
1349 | { | |
66f97d31 | 1350 | basic_block dominated, dom_bb; |
66f97d31 | 1351 | unsigned j; |
617b465c ZD |
1352 | |
1353 | bb = bbs[i]; | |
113d659a | 1354 | |
4f899c42 | 1355 | auto_vec<basic_block> dom_bbs = get_dominated_by (CDI_DOMINATORS, bb); |
9771b263 | 1356 | FOR_EACH_VEC_ELT (dom_bbs, j, dominated) |
617b465c | 1357 | { |
617b465c ZD |
1358 | if (flow_bb_inside_loop_p (loop, dominated)) |
1359 | continue; | |
1360 | dom_bb = nearest_common_dominator ( | |
d47cc544 | 1361 | CDI_DOMINATORS, first_active[i], first_active_latch); |
c22cacf3 | 1362 | set_immediate_dominator (CDI_DOMINATORS, dominated, dom_bb); |
617b465c | 1363 | } |
617b465c ZD |
1364 | } |
1365 | free (first_active); | |
1366 | ||
1367 | free (bbs); | |
03cb2019 | 1368 | BITMAP_FREE (bbs_to_scale); |
617b465c ZD |
1369 | |
1370 | return true; | |
1371 | } | |
1372 | ||
f470c378 ZD |
1373 | /* A callback for make_forwarder block, to redirect all edges except for |
1374 | MFB_KJ_EDGE to the entry part. E is the edge for that we should decide | |
1375 | whether to redirect it. */ | |
1376 | ||
b02b9b53 ZD |
1377 | edge mfb_kj_edge; |
1378 | bool | |
f470c378 ZD |
1379 | mfb_keep_just (edge e) |
1380 | { | |
1381 | return e != mfb_kj_edge; | |
1382 | } | |
1383 | ||
e855c69d AB |
1384 | /* True when a candidate preheader BLOCK has predecessors from LOOP. */ |
1385 | ||
1386 | static bool | |
99b1c316 | 1387 | has_preds_from_loop (basic_block block, class loop *loop) |
e855c69d AB |
1388 | { |
1389 | edge e; | |
1390 | edge_iterator ei; | |
b8698a0f | 1391 | |
e855c69d AB |
1392 | FOR_EACH_EDGE (e, ei, block->preds) |
1393 | if (e->src->loop_father == loop) | |
1394 | return true; | |
1395 | return false; | |
1396 | } | |
1397 | ||
3d436d2a ZD |
1398 | /* Creates a pre-header for a LOOP. Returns newly created block. Unless |
1399 | CP_SIMPLE_PREHEADERS is set in FLAGS, we only force LOOP to have single | |
1400 | entry; otherwise we also force preheader block to have only one successor. | |
e855c69d | 1401 | When CP_FALLTHRU_PREHEADERS is set in FLAGS, we force the preheader block |
b8698a0f | 1402 | to be a fallthru predecessor to the loop header and to have only |
e855c69d | 1403 | predecessors from outside of the loop. |
f470c378 ZD |
1404 | The function also updates dominators. */ |
1405 | ||
b02b9b53 | 1406 | basic_block |
99b1c316 | 1407 | create_preheader (class loop *loop, int flags) |
3d436d2a | 1408 | { |
183ac6d0 | 1409 | edge e; |
3d436d2a | 1410 | basic_block dummy; |
3d436d2a | 1411 | int nentry = 0; |
f470c378 | 1412 | bool irred = false; |
c15bc84b | 1413 | bool latch_edge_was_fallthru; |
c7b852c8 | 1414 | edge one_succ_pred = NULL, single_entry = NULL; |
628f6a4e | 1415 | edge_iterator ei; |
3d436d2a | 1416 | |
628f6a4e | 1417 | FOR_EACH_EDGE (e, ei, loop->header->preds) |
3d436d2a ZD |
1418 | { |
1419 | if (e->src == loop->latch) | |
1420 | continue; | |
f470c378 | 1421 | irred |= (e->flags & EDGE_IRREDUCIBLE_LOOP) != 0; |
3d436d2a | 1422 | nentry++; |
c7b852c8 | 1423 | single_entry = e; |
c5cbcccf | 1424 | if (single_succ_p (e->src)) |
c15bc84b | 1425 | one_succ_pred = e; |
3d436d2a | 1426 | } |
341c100f | 1427 | gcc_assert (nentry); |
3d436d2a ZD |
1428 | if (nentry == 1) |
1429 | { | |
e855c69d | 1430 | bool need_forwarder_block = false; |
b8698a0f | 1431 | |
e855c69d | 1432 | /* We do not allow entry block to be the loop preheader, since we |
89f8f30f | 1433 | cannot emit code there. */ |
fefa31b5 | 1434 | if (single_entry->src == ENTRY_BLOCK_PTR_FOR_FN (cfun)) |
e855c69d AB |
1435 | need_forwarder_block = true; |
1436 | else | |
1437 | { | |
1438 | /* If we want simple preheaders, also force the preheader to have | |
6fc00b41 | 1439 | just a single successor and a normal edge. */ |
e855c69d | 1440 | if ((flags & CP_SIMPLE_PREHEADERS) |
6fc00b41 RB |
1441 | && ((single_entry->flags & EDGE_COMPLEX) |
1442 | || !single_succ_p (single_entry->src))) | |
e855c69d AB |
1443 | need_forwarder_block = true; |
1444 | /* If we want fallthru preheaders, also create forwarder block when | |
1445 | preheader ends with a jump or has predecessors from loop. */ | |
1446 | else if ((flags & CP_FALLTHRU_PREHEADERS) | |
1447 | && (JUMP_P (BB_END (single_entry->src)) | |
1448 | || has_preds_from_loop (single_entry->src, loop))) | |
1449 | need_forwarder_block = true; | |
1450 | } | |
1451 | if (! need_forwarder_block) | |
3d436d2a ZD |
1452 | return NULL; |
1453 | } | |
1454 | ||
f470c378 | 1455 | mfb_kj_edge = loop_latch_edge (loop); |
c15bc84b | 1456 | latch_edge_was_fallthru = (mfb_kj_edge->flags & EDGE_FALLTHRU) != 0; |
49574486 JJ |
1457 | if (nentry == 1 |
1458 | && ((flags & CP_FALLTHRU_PREHEADERS) == 0 | |
1459 | || (single_entry->flags & EDGE_CROSSING) == 0)) | |
183ac6d0 RB |
1460 | dummy = split_edge (single_entry); |
1461 | else | |
1462 | { | |
1463 | edge fallthru = make_forwarder_block (loop->header, mfb_keep_just, NULL); | |
1464 | dummy = fallthru->src; | |
1465 | loop->header = fallthru->dest; | |
1466 | } | |
3d436d2a | 1467 | |
c15bc84b EB |
1468 | /* Try to be clever in placing the newly created preheader. The idea is to |
1469 | avoid breaking any "fallthruness" relationship between blocks. | |
1470 | ||
1471 | The preheader was created just before the header and all incoming edges | |
1472 | to the header were redirected to the preheader, except the latch edge. | |
1473 | So the only problematic case is when this latch edge was a fallthru | |
1474 | edge: it is not anymore after the preheader creation so we have broken | |
1475 | the fallthruness. We're therefore going to look for a better place. */ | |
1476 | if (latch_edge_was_fallthru) | |
1477 | { | |
1478 | if (one_succ_pred) | |
1479 | e = one_succ_pred; | |
1480 | else | |
1481 | e = EDGE_PRED (dummy, 0); | |
1482 | ||
1483 | move_block_after (dummy, e->src); | |
1484 | } | |
f470c378 | 1485 | |
f470c378 | 1486 | if (irred) |
3d436d2a | 1487 | { |
f470c378 | 1488 | dummy->flags |= BB_IRREDUCIBLE_LOOP; |
c5cbcccf | 1489 | single_succ_edge (dummy)->flags |= EDGE_IRREDUCIBLE_LOOP; |
3d436d2a ZD |
1490 | } |
1491 | ||
c263766c RH |
1492 | if (dump_file) |
1493 | fprintf (dump_file, "Created preheader block for loop %i\n", | |
3d436d2a | 1494 | loop->num); |
b8698a0f | 1495 | |
e855c69d AB |
1496 | if (flags & CP_FALLTHRU_PREHEADERS) |
1497 | gcc_assert ((single_succ_edge (dummy)->flags & EDGE_FALLTHRU) | |
1498 | && !JUMP_P (BB_END (dummy))); | |
3d436d2a ZD |
1499 | |
1500 | return dummy; | |
1501 | } | |
1502 | ||
d73be268 ZD |
1503 | /* Create preheaders for each loop; for meaning of FLAGS see create_preheader. */ |
1504 | ||
3d436d2a | 1505 | void |
d73be268 | 1506 | create_preheaders (int flags) |
3d436d2a | 1507 | { |
c7b852c8 ZD |
1508 | if (!current_loops) |
1509 | return; | |
1510 | ||
e41ba804 | 1511 | for (auto loop : loops_list (cfun, 0)) |
42fd6772 | 1512 | create_preheader (loop, flags); |
f87000d0 | 1513 | loops_state_set (LOOPS_HAVE_PREHEADERS); |
3d436d2a ZD |
1514 | } |
1515 | ||
d73be268 ZD |
1516 | /* Forces all loop latches to have only single successor. */ |
1517 | ||
3d436d2a | 1518 | void |
d73be268 | 1519 | force_single_succ_latches (void) |
3d436d2a | 1520 | { |
3d436d2a ZD |
1521 | edge e; |
1522 | ||
e41ba804 | 1523 | for (auto loop : loops_list (cfun, 0)) |
3d436d2a | 1524 | { |
c5cbcccf | 1525 | if (loop->latch != loop->header && single_succ_p (loop->latch)) |
3d436d2a | 1526 | continue; |
d329e058 | 1527 | |
9ff3d2de | 1528 | e = find_edge (loop->latch, loop->header); |
9dcbb380 | 1529 | gcc_checking_assert (e != NULL); |
bc810602 | 1530 | |
598ec7bd | 1531 | split_edge (e); |
3d436d2a | 1532 | } |
f87000d0 | 1533 | loops_state_set (LOOPS_HAVE_SIMPLE_LATCHES); |
3d436d2a ZD |
1534 | } |
1535 | ||
1cb7dfc3 MH |
1536 | /* This function is called from loop_version. It splits the entry edge |
1537 | of the loop we want to version, adds the versioning condition, and | |
1538 | adjust the edges to the two versions of the loop appropriately. | |
1539 | e is an incoming edge. Returns the basic block containing the | |
1540 | condition. | |
1541 | ||
1542 | --- edge e ---- > [second_head] | |
1543 | ||
1544 | Split it and insert new conditional expression and adjust edges. | |
1545 | ||
1546 | --- edge e ---> [cond expr] ---> [first_head] | |
c22cacf3 MS |
1547 | | |
1548 | +---------> [second_head] | |
03cb2019 | 1549 | |
5d3ebb71 JH |
1550 | THEN_PROB is the probability of then branch of the condition. |
1551 | ELSE_PROB is the probability of else branch. Note that they may be both | |
542e7230 BC |
1552 | REG_BR_PROB_BASE when condition is IFN_LOOP_VECTORIZED or |
1553 | IFN_LOOP_DIST_ALIAS. */ | |
1cb7dfc3 MH |
1554 | |
1555 | static basic_block | |
03cb2019 | 1556 | lv_adjust_loop_entry_edge (basic_block first_head, basic_block second_head, |
357067f2 JH |
1557 | edge e, void *cond_expr, |
1558 | profile_probability then_prob, | |
1559 | profile_probability else_prob) | |
1cb7dfc3 MH |
1560 | { |
1561 | basic_block new_head = NULL; | |
1562 | edge e1; | |
1563 | ||
1564 | gcc_assert (e->dest == second_head); | |
1565 | ||
1566 | /* Split edge 'e'. This will create a new basic block, where we can | |
1567 | insert conditional expr. */ | |
1568 | new_head = split_edge (e); | |
1569 | ||
1cb7dfc3 MH |
1570 | lv_add_condition_to_bb (first_head, second_head, new_head, |
1571 | cond_expr); | |
1572 | ||
766613a4 | 1573 | /* Don't set EDGE_TRUE_VALUE in RTL mode, as it's invalid there. */ |
03cb2019 | 1574 | e = single_succ_edge (new_head); |
52bca999 SB |
1575 | e1 = make_edge (new_head, first_head, |
1576 | current_ir_type () == IR_GIMPLE ? EDGE_TRUE_VALUE : 0); | |
03cb2019 | 1577 | e1->probability = then_prob; |
5d3ebb71 | 1578 | e->probability = else_prob; |
03cb2019 | 1579 | |
1cb7dfc3 MH |
1580 | set_immediate_dominator (CDI_DOMINATORS, first_head, new_head); |
1581 | set_immediate_dominator (CDI_DOMINATORS, second_head, new_head); | |
1582 | ||
1583 | /* Adjust loop header phi nodes. */ | |
1584 | lv_adjust_loop_header_phi (first_head, second_head, new_head, e1); | |
1585 | ||
1586 | return new_head; | |
1587 | } | |
1588 | ||
1589 | /* Main entry point for Loop Versioning transformation. | |
c22cacf3 | 1590 | |
b9a66240 ZD |
1591 | This transformation given a condition and a loop, creates |
1592 | -if (condition) { loop_copy1 } else { loop_copy2 }, | |
1593 | where loop_copy1 is the loop transformed in one way, and loop_copy2 | |
5daaf2d5 | 1594 | is the loop transformed in another way (or unchanged). COND_EXPR |
b9a66240 ZD |
1595 | may be a run time test for things that were not resolved by static |
1596 | analysis (overlapping ranges (anti-aliasing), alignment, etc.). | |
1597 | ||
5daaf2d5 AH |
1598 | If non-NULL, CONDITION_BB is set to the basic block containing the |
1599 | condition. | |
1600 | ||
03cb2019 ZD |
1601 | THEN_PROB is the probability of the then edge of the if. THEN_SCALE |
1602 | is the ratio by that the frequencies in the original loop should | |
1603 | be scaled. ELSE_SCALE is the ratio by that the frequencies in the | |
1604 | new loop should be scaled. | |
b8698a0f | 1605 | |
b9a66240 ZD |
1606 | If PLACE_AFTER is true, we place the new loop after LOOP in the |
1607 | instruction stream, otherwise it is placed before LOOP. */ | |
1cb7dfc3 | 1608 | |
99b1c316 MS |
1609 | class loop * |
1610 | loop_version (class loop *loop, | |
b9a66240 | 1611 | void *cond_expr, basic_block *condition_bb, |
357067f2 | 1612 | profile_probability then_prob, profile_probability else_prob, |
af2bbc51 | 1613 | profile_probability then_scale, profile_probability else_scale, |
b9a66240 | 1614 | bool place_after) |
1cb7dfc3 MH |
1615 | { |
1616 | basic_block first_head, second_head; | |
f35af8df | 1617 | edge entry, latch_edge; |
1cb7dfc3 | 1618 | int irred_flag; |
99b1c316 | 1619 | class loop *nloop; |
b9a66240 | 1620 | basic_block cond_bb; |
1cb7dfc3 | 1621 | |
1cb7dfc3 MH |
1622 | /* Record entry and latch edges for the loop */ |
1623 | entry = loop_preheader_edge (loop); | |
1624 | irred_flag = entry->flags & EDGE_IRREDUCIBLE_LOOP; | |
1625 | entry->flags &= ~EDGE_IRREDUCIBLE_LOOP; | |
c22cacf3 | 1626 | |
1cb7dfc3 MH |
1627 | /* Note down head of loop as first_head. */ |
1628 | first_head = entry->dest; | |
1629 | ||
f35af8df | 1630 | /* 1) Duplicate loop on the entry edge. */ |
4851c80c XL |
1631 | if (!cfg_hook_duplicate_loop_body_to_header_edge (loop, entry, 1, NULL, NULL, |
1632 | NULL, 0)) | |
71056fef JJ |
1633 | { |
1634 | entry->flags |= irred_flag; | |
1635 | return NULL; | |
1636 | } | |
1cb7dfc3 | 1637 | |
f35af8df XL |
1638 | /* 2) loopify the duplicated new loop. */ |
1639 | latch_edge = single_succ_edge (get_bb_copy (loop->latch)); | |
1640 | nloop = alloc_loop (); | |
1641 | class loop *outer = loop_outer (latch_edge->dest->loop_father); | |
1642 | edge new_header_edge = single_pred_edge (get_bb_copy (loop->header)); | |
1643 | nloop->header = new_header_edge->dest; | |
1644 | nloop->latch = latch_edge->src; | |
1645 | loop_redirect_edge (latch_edge, nloop->header); | |
1646 | ||
1647 | /* Compute new loop. */ | |
1648 | add_loop (nloop, outer); | |
1649 | copy_loop_info (loop, nloop); | |
1650 | set_loop_copy (loop, nloop); | |
1651 | ||
1652 | /* loopify redirected latch_edge. Update its PENDING_STMTS. */ | |
1653 | lv_flush_pending_stmts (latch_edge); | |
1654 | ||
1cb7dfc3 MH |
1655 | /* After duplication entry edge now points to new loop head block. |
1656 | Note down new head as second_head. */ | |
1657 | second_head = entry->dest; | |
1658 | ||
f35af8df | 1659 | /* 3) Split loop entry edge and insert new block with cond expr. */ |
b9a66240 | 1660 | cond_bb = lv_adjust_loop_entry_edge (first_head, second_head, |
5d3ebb71 | 1661 | entry, cond_expr, then_prob, else_prob); |
b9a66240 ZD |
1662 | if (condition_bb) |
1663 | *condition_bb = cond_bb; | |
1664 | ||
1665 | if (!cond_bb) | |
1cb7dfc3 MH |
1666 | { |
1667 | entry->flags |= irred_flag; | |
1668 | return NULL; | |
1669 | } | |
1670 | ||
f35af8df XL |
1671 | /* Add cond_bb to appropriate loop. */ |
1672 | if (cond_bb->loop_father) | |
1673 | remove_bb_from_loops (cond_bb); | |
1674 | add_bb_to_loop (cond_bb, outer); | |
bf45c4c0 | 1675 | |
f35af8df XL |
1676 | /* 4) Scale the original loop and new loop frequency. */ |
1677 | scale_loop_frequencies (loop, then_scale); | |
1678 | scale_loop_frequencies (nloop, else_scale); | |
1679 | update_dominators_in_loop (loop); | |
1680 | update_dominators_in_loop (nloop); | |
1cb7dfc3 | 1681 | |
1cb7dfc3 MH |
1682 | /* Adjust irreducible flag. */ |
1683 | if (irred_flag) | |
1684 | { | |
b9a66240 | 1685 | cond_bb->flags |= BB_IRREDUCIBLE_LOOP; |
1cb7dfc3 MH |
1686 | loop_preheader_edge (loop)->flags |= EDGE_IRREDUCIBLE_LOOP; |
1687 | loop_preheader_edge (nloop)->flags |= EDGE_IRREDUCIBLE_LOOP; | |
b9a66240 ZD |
1688 | single_pred_edge (cond_bb)->flags |= EDGE_IRREDUCIBLE_LOOP; |
1689 | } | |
1690 | ||
1691 | if (place_after) | |
1692 | { | |
1693 | basic_block *bbs = get_loop_body_in_dom_order (nloop), after; | |
1694 | unsigned i; | |
1695 | ||
1696 | after = loop->latch; | |
1697 | ||
1698 | for (i = 0; i < nloop->num_nodes; i++) | |
1699 | { | |
1700 | move_block_after (bbs[i], after); | |
1701 | after = bbs[i]; | |
1702 | } | |
1703 | free (bbs); | |
1cb7dfc3 MH |
1704 | } |
1705 | ||
fa10beec RW |
1706 | /* At this point condition_bb is loop preheader with two successors, |
1707 | first_head and second_head. Make sure that loop preheader has only | |
1cb7dfc3 | 1708 | one successor. */ |
598ec7bd ZD |
1709 | split_edge (loop_preheader_edge (loop)); |
1710 | split_edge (loop_preheader_edge (nloop)); | |
1cb7dfc3 MH |
1711 | |
1712 | return nloop; | |
1713 | } |