Mercurial > hg > CbC > CbC_gcc
annotate gcc/cfgloopmanip.c @ 145:1830386684a0
gcc-9.2.0
author | anatofuz |
---|---|
date | Thu, 13 Feb 2020 11:34:05 +0900 |
parents | 84e7813d76e9 |
children |
rev | line source |
---|---|
0 | 1 /* Loop manipulation code for GNU compiler. |
145 | 2 Copyright (C) 2002-2020 Free Software Foundation, Inc. |
0 | 3 |
4 This file is part of GCC. | |
5 | |
6 GCC is free software; you can redistribute it and/or modify it under | |
7 the terms of the GNU General Public License as published by the Free | |
8 Software Foundation; either version 3, or (at your option) any later | |
9 version. | |
10 | |
11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
12 WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
14 for more details. | |
15 | |
16 You should have received a copy of the GNU General Public License | |
17 along with GCC; see the file COPYING3. If not see | |
18 <http://www.gnu.org/licenses/>. */ | |
19 | |
20 #include "config.h" | |
21 #include "system.h" | |
22 #include "coretypes.h" | |
111 | 23 #include "backend.h" |
0 | 24 #include "rtl.h" |
111 | 25 #include "tree.h" |
26 #include "gimple.h" | |
27 #include "cfghooks.h" | |
28 #include "cfganal.h" | |
0 | 29 #include "cfgloop.h" |
111 | 30 #include "gimple-iterator.h" |
31 #include "gimplify-me.h" | |
32 #include "tree-ssa-loop-manip.h" | |
33 #include "dumpfile.h" | |
0 | 34 |
145 | 35 static void copy_loops_to (class loop **, int, |
36 class loop *); | |
0 | 37 static void loop_redirect_edge (edge, basic_block); |
38 static void remove_bbs (basic_block *, int); | |
39 static bool rpe_enum_p (const_basic_block, const void *); | |
40 static int find_path (edge, basic_block **); | |
145 | 41 static void fix_loop_placements (class loop *, bool *); |
0 | 42 static bool fix_bb_placement (basic_block); |
111 | 43 static void fix_bb_placements (basic_block, bool *, bitmap); |
0 | 44 |
45 /* Checks whether basic block BB is dominated by DATA. */ | |
46 static bool | |
47 rpe_enum_p (const_basic_block bb, const void *data) | |
48 { | |
49 return dominated_by_p (CDI_DOMINATORS, bb, (const_basic_block) data); | |
50 } | |
51 | |
52 /* Remove basic blocks BBS. NBBS is the number of the basic blocks. */ | |
53 | |
54 static void | |
55 remove_bbs (basic_block *bbs, int nbbs) | |
56 { | |
57 int i; | |
58 | |
59 for (i = 0; i < nbbs; i++) | |
60 delete_basic_block (bbs[i]); | |
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. | |
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. */ | |
69 static int | |
70 find_path (edge e, basic_block **bbs) | |
71 { | |
72 gcc_assert (EDGE_COUNT (e->dest->preds) <= 1); | |
73 | |
74 /* Find bbs in the path. */ | |
111 | 75 *bbs = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun)); |
0 | 76 return dfs_enumerate_from (e->dest, 0, rpe_enum_p, *bbs, |
111 | 77 n_basic_blocks_for_fn (cfun), e->dest); |
0 | 78 } |
79 | |
80 /* Fix placement of basic block BB inside loop hierarchy -- | |
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 | |
88 fix_bb_placement (basic_block bb) | |
89 { | |
90 edge e; | |
91 edge_iterator ei; | |
145 | 92 class loop *loop = current_loops->tree_root, *act; |
0 | 93 |
94 FOR_EACH_EDGE (e, ei, bb->succs) | |
95 { | |
111 | 96 if (e->dest == EXIT_BLOCK_PTR_FOR_FN (cfun)) |
0 | 97 continue; |
98 | |
99 act = e->dest->loop_father; | |
100 if (act->header == e->dest) | |
101 act = loop_outer (act); | |
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 | |
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 | |
111 | 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. */ | |
0 | 123 |
124 static bool | |
145 | 125 fix_loop_placement (class loop *loop, bool *irred_invalidated) |
0 | 126 { |
127 unsigned i; | |
128 edge e; | |
111 | 129 vec<edge> exits = get_loop_exit_edges (loop); |
145 | 130 class loop *father = current_loops->tree_root, *act; |
0 | 131 bool ret = false; |
132 | |
111 | 133 FOR_EACH_VEC_ELT (exits, i, e) |
0 | 134 { |
135 act = find_common_loop (loop, e->dest->loop_father); | |
136 if (flow_loop_nested_p (father, act)) | |
137 father = act; | |
138 } | |
139 | |
140 if (father != loop_outer (loop)) | |
141 { | |
142 for (act = loop_outer (loop); act != father; act = loop_outer (act)) | |
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. */ | |
111 | 149 FOR_EACH_VEC_ELT (exits, i, e) |
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 } | |
0 | 156 |
157 ret = true; | |
158 } | |
159 | |
111 | 160 exits.release (); |
0 | 161 return ret; |
162 } | |
163 | |
164 /* Fix placements of basic blocks inside loop hierarchy stored in loops; i.e. | |
111 | 165 enforce condition stated in description of fix_bb_placement. We |
0 | 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 | |
170 to this change; the condition for them is similar, except that instead of | |
171 successors we consider edges coming out of the loops. | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
172 |
0 | 173 If the changes may invalidate the information about irreducible regions, |
111 | 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. */ | |
0 | 178 |
179 static void | |
180 fix_bb_placements (basic_block from, | |
111 | 181 bool *irred_invalidated, |
182 bitmap loop_closed_ssa_invalidated) | |
0 | 183 { |
184 basic_block *queue, *qtop, *qbeg, *qend; | |
145 | 185 class loop *base_loop, *target_loop; |
0 | 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; | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
196 /* If we are already in the outermost loop, the basic blocks cannot be moved |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
197 outside of it. If FROM is the header of the base loop, it cannot be moved |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
198 outside of it, either. In both cases, we can end now. */ |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
199 if (base_loop == current_loops->tree_root |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
200 || from == base_loop->header) |
0 | 201 return; |
202 | |
111 | 203 auto_sbitmap in_queue (last_basic_block_for_fn (cfun)); |
204 bitmap_clear (in_queue); | |
205 bitmap_set_bit (in_queue, from->index); | |
0 | 206 /* Prevent us from going out of the base_loop. */ |
111 | 207 bitmap_set_bit (in_queue, base_loop->header->index); |
0 | 208 |
209 queue = XNEWVEC (basic_block, base_loop->num_nodes + 1); | |
210 qtop = queue + base_loop->num_nodes + 1; | |
211 qbeg = queue; | |
212 qend = queue + 1; | |
213 *qbeg = from; | |
214 | |
215 while (qbeg != qend) | |
216 { | |
217 edge_iterator ei; | |
218 from = *qbeg; | |
219 qbeg++; | |
220 if (qbeg == qtop) | |
221 qbeg = queue; | |
111 | 222 bitmap_clear_bit (in_queue, from->index); |
0 | 223 |
224 if (from->loop_father->header == from) | |
225 { | |
226 /* Subloop header, maybe move the loop upward. */ | |
111 | 227 if (!fix_loop_placement (from->loop_father, irred_invalidated)) |
0 | 228 continue; |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
229 target_loop = loop_outer (from->loop_father); |
111 | 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 } | |
0 | 237 } |
238 else | |
239 { | |
240 /* Ordinary basic block. */ | |
241 if (!fix_bb_placement (from)) | |
242 continue; | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
243 target_loop = from->loop_father; |
111 | 244 if (loop_closed_ssa_invalidated) |
245 bitmap_set_bit (loop_closed_ssa_invalidated, from->index); | |
0 | 246 } |
247 | |
248 FOR_EACH_EDGE (e, ei, from->succs) | |
249 { | |
250 if (e->flags & EDGE_IRREDUCIBLE_LOOP) | |
251 *irred_invalidated = true; | |
252 } | |
253 | |
254 /* Something has changed, insert predecessors into queue. */ | |
255 FOR_EACH_EDGE (e, ei, from->preds) | |
256 { | |
257 basic_block pred = e->src; | |
145 | 258 class loop *nca; |
0 | 259 |
260 if (e->flags & EDGE_IRREDUCIBLE_LOOP) | |
261 *irred_invalidated = true; | |
262 | |
111 | 263 if (bitmap_bit_p (in_queue, pred->index)) |
0 | 264 continue; |
265 | |
266 /* If it is subloop, then it either was not moved, or | |
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; | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
274 else if (!flow_loop_nested_p (target_loop, pred->loop_father)) |
0 | 275 { |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
276 /* If PRED is already higher in the loop hierarchy than the |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
277 TARGET_LOOP to that we moved FROM, the change of the position |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
278 of FROM does not affect the position of PRED, so there is no |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
279 point in processing it. */ |
0 | 280 continue; |
281 } | |
282 | |
111 | 283 if (bitmap_bit_p (in_queue, pred->index)) |
0 | 284 continue; |
285 | |
286 /* Schedule the basic block. */ | |
287 *qend = pred; | |
288 qend++; | |
289 if (qend == qtop) | |
290 qend = queue; | |
111 | 291 bitmap_set_bit (in_queue, pred->index); |
0 | 292 } |
293 } | |
294 free (queue); | |
295 } | |
296 | |
297 /* Removes path beginning at edge E, i.e. remove basic blocks dominated by E | |
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). */ | |
300 bool | |
111 | 301 remove_path (edge e, bool *irred_invalidated, |
302 bitmap loop_closed_ssa_invalidated) | |
0 | 303 { |
304 edge ae; | |
305 basic_block *rem_bbs, *bord_bbs, from, bb; | |
111 | 306 vec<basic_block> dom_bbs; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
307 int i, nrem, n_bord_bbs; |
111 | 308 bool local_irred_invalidated = false; |
309 edge_iterator ei; | |
145 | 310 class loop *l, *f; |
111 | 311 |
312 if (! irred_invalidated) | |
313 irred_invalidated = &local_irred_invalidated; | |
0 | 314 |
315 if (!can_remove_branch_p (e)) | |
316 return false; | |
317 | |
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) | |
111 | 324 *irred_invalidated = true; |
0 | 325 |
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. */ | |
330 if (!single_pred_p (e->dest)) | |
331 e = single_pred_edge (split_edge (e)); | |
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. */ | |
111 | 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)) | |
341 unloop (l, irred_invalidated, loop_closed_ssa_invalidated); | |
342 } | |
0 | 343 |
344 /* Identify the path. */ | |
345 nrem = find_path (e, &rem_bbs); | |
346 | |
347 n_bord_bbs = 0; | |
111 | 348 bord_bbs = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun)); |
349 auto_sbitmap seen (last_basic_block_for_fn (cfun)); | |
350 bitmap_clear (seen); | |
0 | 351 |
352 /* Find "border" hexes -- i.e. those with predecessor in removed path. */ | |
353 for (i = 0; i < nrem; i++) | |
111 | 354 bitmap_set_bit (seen, rem_bbs[i]->index); |
355 if (!*irred_invalidated) | |
356 FOR_EACH_EDGE (ae, ei, e->src->succs) | |
357 if (ae != e && ae->dest != EXIT_BLOCK_PTR_FOR_FN (cfun) | |
358 && !bitmap_bit_p (seen, ae->dest->index) | |
359 && ae->flags & EDGE_IRREDUCIBLE_LOOP) | |
360 { | |
361 *irred_invalidated = true; | |
362 break; | |
363 } | |
364 | |
0 | 365 for (i = 0; i < nrem; i++) |
366 { | |
367 FOR_EACH_EDGE (ae, ei, rem_bbs[i]->succs) | |
111 | 368 if (ae->dest != EXIT_BLOCK_PTR_FOR_FN (cfun) |
369 && !bitmap_bit_p (seen, ae->dest->index)) | |
0 | 370 { |
111 | 371 bitmap_set_bit (seen, ae->dest->index); |
0 | 372 bord_bbs[n_bord_bbs++] = ae->dest; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
373 |
0 | 374 if (ae->flags & EDGE_IRREDUCIBLE_LOOP) |
111 | 375 *irred_invalidated = true; |
0 | 376 } |
377 } | |
378 | |
379 /* Remove the path. */ | |
380 from = e->src; | |
381 remove_branch (e); | |
111 | 382 dom_bbs.create (0); |
0 | 383 |
384 /* Cancel loops contained in the path. */ | |
385 for (i = 0; i < nrem; i++) | |
386 if (rem_bbs[i]->loop_father->header == rem_bbs[i]) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
387 cancel_loop_tree (rem_bbs[i]->loop_father); |
0 | 388 |
389 remove_bbs (rem_bbs, nrem); | |
390 free (rem_bbs); | |
391 | |
392 /* Find blocks whose dominators may be affected. */ | |
111 | 393 bitmap_clear (seen); |
0 | 394 for (i = 0; i < n_bord_bbs; i++) |
395 { | |
396 basic_block ldom; | |
397 | |
398 bb = get_immediate_dominator (CDI_DOMINATORS, bord_bbs[i]); | |
111 | 399 if (bitmap_bit_p (seen, bb->index)) |
0 | 400 continue; |
111 | 401 bitmap_set_bit (seen, bb->index); |
0 | 402 |
403 for (ldom = first_dom_son (CDI_DOMINATORS, bb); | |
404 ldom; | |
405 ldom = next_dom_son (CDI_DOMINATORS, ldom)) | |
406 if (!dominated_by_p (CDI_DOMINATORS, from, ldom)) | |
111 | 407 dom_bbs.safe_push (ldom); |
0 | 408 } |
409 | |
410 /* Recount dominators. */ | |
411 iterate_fix_dominators (CDI_DOMINATORS, dom_bbs, true); | |
111 | 412 dom_bbs.release (); |
0 | 413 free (bord_bbs); |
414 | |
415 /* Fix placements of basic blocks inside loops and the placement of | |
416 loops in the loop tree. */ | |
111 | 417 fix_bb_placements (from, irred_invalidated, loop_closed_ssa_invalidated); |
418 fix_loop_placements (from->loop_father, irred_invalidated); | |
0 | 419 |
111 | 420 if (local_irred_invalidated |
0 | 421 && loops_state_satisfies_p (LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS)) |
422 mark_irreducible_loops (); | |
423 | |
424 return true; | |
425 } | |
426 | |
111 | 427 /* Creates place for a new LOOP in loops structure of FN. */ |
0 | 428 |
111 | 429 void |
145 | 430 place_new_loop (struct function *fn, class loop *loop) |
0 | 431 { |
111 | 432 loop->num = number_of_loops (fn); |
433 vec_safe_push (loops_for_fn (fn)->larray, loop); | |
0 | 434 } |
435 | |
436 /* Given LOOP structure with filled header and latch, find the body of the | |
437 corresponding loop and add it to loops tree. Insert the LOOP as a son of | |
438 outer. */ | |
439 | |
440 void | |
145 | 441 add_loop (class loop *loop, class loop *outer) |
0 | 442 { |
443 basic_block *bbs; | |
444 int i, n; | |
145 | 445 class loop *subloop; |
0 | 446 edge e; |
447 edge_iterator ei; | |
448 | |
449 /* Add it to loop structure. */ | |
111 | 450 place_new_loop (cfun, loop); |
0 | 451 flow_loop_tree_node_add (outer, loop); |
452 | |
453 /* Find its nodes. */ | |
111 | 454 bbs = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun)); |
455 n = get_loop_body_with_size (loop, bbs, n_basic_blocks_for_fn (cfun)); | |
0 | 456 |
457 for (i = 0; i < n; i++) | |
458 { | |
459 if (bbs[i]->loop_father == outer) | |
460 { | |
461 remove_bb_from_loops (bbs[i]); | |
462 add_bb_to_loop (bbs[i], loop); | |
463 continue; | |
464 } | |
465 | |
466 loop->num_nodes++; | |
467 | |
468 /* If we find a direct subloop of OUTER, move it to LOOP. */ | |
469 subloop = bbs[i]->loop_father; | |
470 if (loop_outer (subloop) == outer | |
471 && subloop->header == bbs[i]) | |
472 { | |
473 flow_loop_tree_node_remove (subloop); | |
474 flow_loop_tree_node_add (loop, subloop); | |
475 } | |
476 } | |
477 | |
478 /* Update the information about loop exit edges. */ | |
479 for (i = 0; i < n; i++) | |
480 { | |
481 FOR_EACH_EDGE (e, ei, bbs[i]->succs) | |
482 { | |
483 rescan_loop_exit (e, false, false); | |
484 } | |
485 } | |
486 | |
487 free (bbs); | |
488 } | |
489 | |
111 | 490 /* Scale profile of loop by P. */ |
491 | |
0 | 492 void |
145 | 493 scale_loop_frequencies (class loop *loop, profile_probability p) |
0 | 494 { |
495 basic_block *bbs; | |
496 | |
497 bbs = get_loop_body (loop); | |
111 | 498 scale_bbs_frequencies (bbs, loop->num_nodes, p); |
0 | 499 free (bbs); |
500 } | |
501 | |
111 | 502 /* Scale profile in LOOP by P. |
503 If ITERATION_BOUND is non-zero, scale even further if loop is predicted | |
131 | 504 to iterate too many times. |
505 Before caling this function, preheader block profile should be already | |
506 scaled to final count. This is necessary because loop iterations are | |
507 determined by comparing header edge count to latch ege count and thus | |
508 they need to be scaled synchronously. */ | |
111 | 509 |
510 void | |
145 | 511 scale_loop_profile (class loop *loop, profile_probability p, |
111 | 512 gcov_type iteration_bound) |
513 { | |
131 | 514 edge e, preheader_e; |
111 | 515 edge_iterator ei; |
516 | |
517 if (dump_file && (dump_flags & TDF_DETAILS)) | |
518 { | |
519 fprintf (dump_file, ";; Scaling loop %i with scale ", | |
520 loop->num); | |
521 p.dump (dump_file); | |
131 | 522 fprintf (dump_file, " bounding iterations to %i\n", |
523 (int)iteration_bound); | |
524 } | |
525 | |
526 /* Scale the probabilities. */ | |
527 scale_loop_frequencies (loop, p); | |
528 | |
529 if (iteration_bound == 0) | |
530 return; | |
531 | |
532 gcov_type iterations = expected_loop_iterations_unbounded (loop, NULL, true); | |
533 | |
534 if (dump_file && (dump_flags & TDF_DETAILS)) | |
535 { | |
536 fprintf (dump_file, ";; guessed iterations after scaling %i\n", | |
537 (int)iterations); | |
111 | 538 } |
539 | |
540 /* See if loop is predicted to iterate too many times. */ | |
131 | 541 if (iterations <= iteration_bound) |
542 return; | |
543 | |
544 preheader_e = loop_preheader_edge (loop); | |
545 | |
546 /* We could handle also loops without preheaders, but bounding is | |
547 currently used only by optimizers that have preheaders constructed. */ | |
548 gcc_checking_assert (preheader_e); | |
549 profile_count count_in = preheader_e->count (); | |
550 | |
551 if (count_in > profile_count::zero () | |
552 && loop->header->count.initialized_p ()) | |
111 | 553 { |
131 | 554 profile_count count_delta = profile_count::zero (); |
111 | 555 |
556 e = single_exit (loop); | |
557 if (e) | |
558 { | |
559 edge other_e; | |
131 | 560 FOR_EACH_EDGE (other_e, ei, e->src->succs) |
111 | 561 if (!(other_e->flags & (EDGE_ABNORMAL | EDGE_FAKE)) |
562 && e != other_e) | |
563 break; | |
564 | |
565 /* Probability of exit must be 1/iterations. */ | |
566 count_delta = e->count (); | |
567 e->probability = profile_probability::always () | |
131 | 568 .apply_scale (1, iteration_bound); |
111 | 569 other_e->probability = e->probability.invert (); |
570 | |
131 | 571 /* In code below we only handle the following two updates. */ |
572 if (other_e->dest != loop->header | |
573 && other_e->dest != loop->latch | |
574 && (dump_file && (dump_flags & TDF_DETAILS))) | |
111 | 575 { |
131 | 576 fprintf (dump_file, ";; giving up on update of paths from " |
577 "exit condition to latch\n"); | |
111 | 578 } |
579 } | |
131 | 580 else |
581 if (dump_file && (dump_flags & TDF_DETAILS)) | |
582 fprintf (dump_file, ";; Loop has multiple exit edges; " | |
583 "giving up on exit condition update\n"); | |
111 | 584 |
585 /* Roughly speaking we want to reduce the loop body profile by the | |
586 difference of loop iterations. We however can do better if | |
587 we look at the actual profile, if it is available. */ | |
131 | 588 p = profile_probability::always (); |
111 | 589 |
131 | 590 count_in = count_in.apply_scale (iteration_bound, 1); |
591 p = count_in.probability_in (loop->header->count); | |
111 | 592 if (!(p > profile_probability::never ())) |
593 p = profile_probability::very_unlikely (); | |
131 | 594 |
595 if (p == profile_probability::always () | |
596 || !p.initialized_p ()) | |
597 return; | |
111 | 598 |
131 | 599 /* If latch exists, change its count, since we changed |
600 probability of exit. Theoretically we should update everything from | |
601 source of exit edge to latch, but for vectorizer this is enough. */ | |
602 if (loop->latch && loop->latch != e->src) | |
603 loop->latch->count += count_delta; | |
111 | 604 |
131 | 605 /* Scale the probabilities. */ |
606 scale_loop_frequencies (loop, p); | |
607 | |
608 /* Change latch's count back. */ | |
609 if (loop->latch && loop->latch != e->src) | |
610 loop->latch->count -= count_delta; | |
611 | |
612 if (dump_file && (dump_flags & TDF_DETAILS)) | |
613 fprintf (dump_file, ";; guessed iterations are now %i\n", | |
614 (int)expected_loop_iterations_unbounded (loop, NULL, true)); | |
615 } | |
111 | 616 } |
617 | |
0 | 618 /* Recompute dominance information for basic blocks outside LOOP. */ |
619 | |
620 static void | |
145 | 621 update_dominators_in_loop (class loop *loop) |
0 | 622 { |
111 | 623 vec<basic_block> dom_bbs = vNULL; |
0 | 624 basic_block *body; |
625 unsigned i; | |
626 | |
111 | 627 auto_sbitmap seen (last_basic_block_for_fn (cfun)); |
628 bitmap_clear (seen); | |
0 | 629 body = get_loop_body (loop); |
630 | |
631 for (i = 0; i < loop->num_nodes; i++) | |
111 | 632 bitmap_set_bit (seen, body[i]->index); |
0 | 633 |
634 for (i = 0; i < loop->num_nodes; i++) | |
635 { | |
636 basic_block ldom; | |
637 | |
638 for (ldom = first_dom_son (CDI_DOMINATORS, body[i]); | |
639 ldom; | |
640 ldom = next_dom_son (CDI_DOMINATORS, ldom)) | |
111 | 641 if (!bitmap_bit_p (seen, ldom->index)) |
0 | 642 { |
111 | 643 bitmap_set_bit (seen, ldom->index); |
644 dom_bbs.safe_push (ldom); | |
0 | 645 } |
646 } | |
647 | |
648 iterate_fix_dominators (CDI_DOMINATORS, dom_bbs, false); | |
649 free (body); | |
111 | 650 dom_bbs.release (); |
0 | 651 } |
652 | |
653 /* Creates an if region as shown above. CONDITION is used to create | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
654 the test for the if. |
0 | 655 |
656 | | |
657 | ------------- ------------- | |
658 | | pred_bb | | pred_bb | | |
659 | ------------- ------------- | |
660 | | | | |
661 | | | ENTRY_EDGE | |
662 | | ENTRY_EDGE V | |
663 | | ====> ------------- | |
664 | | | cond_bb | | |
665 | | | CONDITION | | |
666 | | ------------- | |
667 | V / \ | |
668 | ------------- e_false / \ e_true | |
669 | | succ_bb | V V | |
670 | ------------- ----------- ----------- | |
671 | | false_bb | | true_bb | | |
672 | ----------- ----------- | |
673 | \ / | |
674 | \ / | |
675 | V V | |
676 | ------------- | |
677 | | join_bb | | |
678 | ------------- | |
679 | | exit_edge (result) | |
680 | V | |
681 | ----------- | |
682 | | succ_bb | | |
683 | ----------- | |
684 | | |
685 */ | |
686 | |
687 edge | |
688 create_empty_if_region_on_edge (edge entry_edge, tree condition) | |
689 { | |
690 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
691 basic_block cond_bb, true_bb, false_bb, join_bb; |
0 | 692 edge e_true, e_false, exit_edge; |
111 | 693 gcond *cond_stmt; |
0 | 694 tree simple_cond; |
695 gimple_stmt_iterator gsi; | |
696 | |
697 cond_bb = split_edge (entry_edge); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
698 |
0 | 699 /* Insert condition in cond_bb. */ |
700 gsi = gsi_last_bb (cond_bb); | |
701 simple_cond = | |
702 force_gimple_operand_gsi (&gsi, condition, true, NULL, | |
703 false, GSI_NEW_STMT); | |
704 cond_stmt = gimple_build_cond_from_tree (simple_cond, NULL_TREE, NULL_TREE); | |
705 gsi = gsi_last_bb (cond_bb); | |
706 gsi_insert_after (&gsi, cond_stmt, GSI_NEW_STMT); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
707 |
0 | 708 join_bb = split_edge (single_succ_edge (cond_bb)); |
709 | |
710 e_true = single_succ_edge (cond_bb); | |
711 true_bb = split_edge (e_true); | |
712 | |
713 e_false = make_edge (cond_bb, join_bb, 0); | |
714 false_bb = split_edge (e_false); | |
715 | |
716 e_true->flags &= ~EDGE_FALLTHRU; | |
717 e_true->flags |= EDGE_TRUE_VALUE; | |
718 e_false->flags &= ~EDGE_FALLTHRU; | |
719 e_false->flags |= EDGE_FALSE_VALUE; | |
720 | |
721 set_immediate_dominator (CDI_DOMINATORS, cond_bb, entry_edge->src); | |
722 set_immediate_dominator (CDI_DOMINATORS, true_bb, cond_bb); | |
723 set_immediate_dominator (CDI_DOMINATORS, false_bb, cond_bb); | |
724 set_immediate_dominator (CDI_DOMINATORS, join_bb, cond_bb); | |
725 | |
726 exit_edge = single_succ_edge (join_bb); | |
727 | |
728 if (single_pred_p (exit_edge->dest)) | |
729 set_immediate_dominator (CDI_DOMINATORS, exit_edge->dest, join_bb); | |
730 | |
731 return exit_edge; | |
732 } | |
733 | |
734 /* create_empty_loop_on_edge | |
735 | | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
736 | - pred_bb - ------ pred_bb ------ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
737 | | | | iv0 = initial_value | |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
738 | -----|----- ---------|----------- |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
739 | | ______ | entry_edge |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
740 | | entry_edge / | | |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
741 | | ====> | -V---V- loop_header ------------- |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
742 | V | | iv_before = phi (iv0, iv_after) | |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
743 | - succ_bb - | ---|----------------------------- |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
744 | | | | | |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
745 | ----------- | ---V--- loop_body --------------- |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
746 | | | iv_after = iv_before + stride | |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
747 | | | if (iv_before < upper_bound) | |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
748 | | ---|--------------\-------------- |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
749 | | | \ exit_e |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
750 | | V \ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
751 | | - loop_latch - V- succ_bb - |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
752 | | | | | | |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
753 | | /------------- ----------- |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
754 | \ ___ / |
0 | 755 |
756 Creates an empty loop as shown above, the IV_BEFORE is the SSA_NAME | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
757 that is used before the increment of IV. IV_BEFORE should be used for |
0 | 758 adding code to the body that uses the IV. OUTER is the outer loop in |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
759 which the new loop should be inserted. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
760 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
761 Both INITIAL_VALUE and UPPER_BOUND expressions are gimplified and |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
762 inserted on the loop entry edge. This implies that this function |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
763 should be used only when the UPPER_BOUND expression is a loop |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
764 invariant. */ |
0 | 765 |
145 | 766 class loop * |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
767 create_empty_loop_on_edge (edge entry_edge, |
0 | 768 tree initial_value, |
769 tree stride, tree upper_bound, | |
770 tree iv, | |
771 tree *iv_before, | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
772 tree *iv_after, |
145 | 773 class loop *outer) |
0 | 774 { |
775 basic_block loop_header, loop_latch, succ_bb, pred_bb; | |
145 | 776 class loop *loop; |
0 | 777 gimple_stmt_iterator gsi; |
778 gimple_seq stmts; | |
111 | 779 gcond *cond_expr; |
0 | 780 tree exit_test; |
781 edge exit_e; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
782 |
0 | 783 gcc_assert (entry_edge && initial_value && stride && upper_bound && iv); |
784 | |
785 /* Create header, latch and wire up the loop. */ | |
786 pred_bb = entry_edge->src; | |
787 loop_header = split_edge (entry_edge); | |
788 loop_latch = split_edge (single_succ_edge (loop_header)); | |
789 succ_bb = single_succ (loop_latch); | |
790 make_edge (loop_header, succ_bb, 0); | |
791 redirect_edge_succ_nodup (single_succ_edge (loop_latch), loop_header); | |
792 | |
793 /* Set immediate dominator information. */ | |
794 set_immediate_dominator (CDI_DOMINATORS, loop_header, pred_bb); | |
795 set_immediate_dominator (CDI_DOMINATORS, loop_latch, loop_header); | |
796 set_immediate_dominator (CDI_DOMINATORS, succ_bb, loop_header); | |
797 | |
798 /* Initialize a loop structure and put it in a loop hierarchy. */ | |
799 loop = alloc_loop (); | |
800 loop->header = loop_header; | |
801 loop->latch = loop_latch; | |
802 add_loop (loop, outer); | |
803 | |
131 | 804 /* TODO: Fix counts. */ |
111 | 805 scale_loop_frequencies (loop, profile_probability::even ()); |
0 | 806 |
807 /* Update dominators. */ | |
808 update_dominators_in_loop (loop); | |
809 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
810 /* Modify edge flags. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
811 exit_e = single_exit (loop); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
812 exit_e->flags = EDGE_LOOP_EXIT | EDGE_FALSE_VALUE; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
813 single_pred_edge (loop_latch)->flags = EDGE_TRUE_VALUE; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
814 |
0 | 815 /* Construct IV code in loop. */ |
816 initial_value = force_gimple_operand (initial_value, &stmts, true, iv); | |
817 if (stmts) | |
818 { | |
819 gsi_insert_seq_on_edge (loop_preheader_edge (loop), stmts); | |
820 gsi_commit_edge_inserts (); | |
821 } | |
822 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
823 upper_bound = force_gimple_operand (upper_bound, &stmts, true, NULL); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
824 if (stmts) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
825 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
826 gsi_insert_seq_on_edge (loop_preheader_edge (loop), stmts); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
827 gsi_commit_edge_inserts (); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
828 } |
0 | 829 |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
830 gsi = gsi_last_bb (loop_header); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
831 create_iv (initial_value, stride, iv, loop, &gsi, false, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
832 iv_before, iv_after); |
0 | 833 |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
834 /* Insert loop exit condition. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
835 cond_expr = gimple_build_cond |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
836 (LT_EXPR, *iv_before, upper_bound, NULL_TREE, NULL_TREE); |
0 | 837 |
838 exit_test = gimple_cond_lhs (cond_expr); | |
839 exit_test = force_gimple_operand_gsi (&gsi, exit_test, true, NULL, | |
840 false, GSI_NEW_STMT); | |
841 gimple_cond_set_lhs (cond_expr, exit_test); | |
842 gsi = gsi_last_bb (exit_e->src); | |
843 gsi_insert_after (&gsi, cond_expr, GSI_NEW_STMT); | |
844 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
845 split_block_after_labels (loop_header); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
846 |
0 | 847 return loop; |
848 } | |
849 | |
850 /* Make area between HEADER_EDGE and LATCH_EDGE a loop by connecting | |
851 latch to header and update loop tree and dominators | |
852 accordingly. Everything between them plus LATCH_EDGE destination must | |
853 be dominated by HEADER_EDGE destination, and back-reachable from | |
854 LATCH_EDGE source. HEADER_EDGE is redirected to basic block SWITCH_BB, | |
855 FALSE_EDGE of SWITCH_BB to original destination of HEADER_EDGE and | |
856 TRUE_EDGE of SWITCH_BB to original destination of LATCH_EDGE. | |
857 Returns the newly created loop. Frequencies and counts in the new loop | |
858 are scaled by FALSE_SCALE and in the old one by TRUE_SCALE. */ | |
859 | |
145 | 860 class loop * |
0 | 861 loopify (edge latch_edge, edge header_edge, |
862 basic_block switch_bb, edge true_edge, edge false_edge, | |
111 | 863 bool redirect_all_edges, profile_probability true_scale, |
864 profile_probability false_scale) | |
0 | 865 { |
866 basic_block succ_bb = latch_edge->dest; | |
867 basic_block pred_bb = header_edge->src; | |
145 | 868 class loop *loop = alloc_loop (); |
869 class loop *outer = loop_outer (succ_bb->loop_father); | |
111 | 870 profile_count cnt; |
0 | 871 |
872 loop->header = header_edge->dest; | |
873 loop->latch = latch_edge->src; | |
874 | |
111 | 875 cnt = header_edge->count (); |
0 | 876 |
877 /* Redirect edges. */ | |
878 loop_redirect_edge (latch_edge, loop->header); | |
879 loop_redirect_edge (true_edge, succ_bb); | |
880 | |
881 /* During loop versioning, one of the switch_bb edge is already properly | |
882 set. Do not redirect it again unless redirect_all_edges is true. */ | |
883 if (redirect_all_edges) | |
884 { | |
885 loop_redirect_edge (header_edge, switch_bb); | |
886 loop_redirect_edge (false_edge, loop->header); | |
887 | |
888 /* Update dominators. */ | |
889 set_immediate_dominator (CDI_DOMINATORS, switch_bb, pred_bb); | |
890 set_immediate_dominator (CDI_DOMINATORS, loop->header, switch_bb); | |
891 } | |
892 | |
893 set_immediate_dominator (CDI_DOMINATORS, succ_bb, switch_bb); | |
894 | |
895 /* Compute new loop. */ | |
896 add_loop (loop, outer); | |
897 | |
898 /* Add switch_bb to appropriate loop. */ | |
899 if (switch_bb->loop_father) | |
900 remove_bb_from_loops (switch_bb); | |
901 add_bb_to_loop (switch_bb, outer); | |
902 | |
131 | 903 /* Fix counts. */ |
0 | 904 if (redirect_all_edges) |
905 { | |
906 switch_bb->count = cnt; | |
907 } | |
111 | 908 scale_loop_frequencies (loop, false_scale); |
909 scale_loop_frequencies (succ_bb->loop_father, true_scale); | |
0 | 910 update_dominators_in_loop (loop); |
911 | |
912 return loop; | |
913 } | |
914 | |
915 /* Remove the latch edge of a LOOP and update loops to indicate that | |
916 the LOOP was removed. After this function, original loop latch will | |
917 have no successor, which caller is expected to fix somehow. | |
918 | |
919 If this may cause the information about irreducible regions to become | |
111 | 920 invalid, IRRED_INVALIDATED is set to true. |
0 | 921 |
111 | 922 LOOP_CLOSED_SSA_INVALIDATED, if non-NULL, is a bitmap where we store |
923 basic blocks that had non-trivial update on their loop_father.*/ | |
924 | |
925 void | |
145 | 926 unloop (class loop *loop, bool *irred_invalidated, |
111 | 927 bitmap loop_closed_ssa_invalidated) |
0 | 928 { |
929 basic_block *body; | |
145 | 930 class loop *ploop; |
0 | 931 unsigned i, n; |
932 basic_block latch = loop->latch; | |
933 bool dummy = false; | |
934 | |
935 if (loop_preheader_edge (loop)->flags & EDGE_IRREDUCIBLE_LOOP) | |
936 *irred_invalidated = true; | |
937 | |
938 /* This is relatively straightforward. The dominators are unchanged, as | |
939 loop header dominates loop latch, so the only thing we have to care of | |
940 is the placement of loops and basic blocks inside the loop tree. We | |
941 move them all to the loop->outer, and then let fix_bb_placements do | |
942 its work. */ | |
943 | |
944 body = get_loop_body (loop); | |
945 n = loop->num_nodes; | |
946 for (i = 0; i < n; i++) | |
947 if (body[i]->loop_father == loop) | |
948 { | |
949 remove_bb_from_loops (body[i]); | |
950 add_bb_to_loop (body[i], loop_outer (loop)); | |
951 } | |
111 | 952 free (body); |
0 | 953 |
954 while (loop->inner) | |
955 { | |
956 ploop = loop->inner; | |
957 flow_loop_tree_node_remove (ploop); | |
958 flow_loop_tree_node_add (loop_outer (loop), ploop); | |
959 } | |
960 | |
961 /* Remove the loop and free its data. */ | |
962 delete_loop (loop); | |
963 | |
964 remove_edge (single_succ_edge (latch)); | |
965 | |
966 /* We do not pass IRRED_INVALIDATED to fix_bb_placements here, as even if | |
967 there is an irreducible region inside the cancelled loop, the flags will | |
968 be still correct. */ | |
111 | 969 fix_bb_placements (latch, &dummy, loop_closed_ssa_invalidated); |
0 | 970 } |
971 | |
972 /* Fix placement of superloops of LOOP inside loop tree, i.e. ensure that | |
973 condition stated in description of fix_loop_placement holds for them. | |
974 It is used in case when we removed some edges coming out of LOOP, which | |
975 may cause the right placement of LOOP inside loop tree to change. | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
976 |
0 | 977 IRRED_INVALIDATED is set to true if a change in the loop structures might |
978 invalidate the information about irreducible regions. */ | |
979 | |
980 static void | |
145 | 981 fix_loop_placements (class loop *loop, bool *irred_invalidated) |
0 | 982 { |
145 | 983 class loop *outer; |
0 | 984 |
985 while (loop_outer (loop)) | |
986 { | |
987 outer = loop_outer (loop); | |
111 | 988 if (!fix_loop_placement (loop, irred_invalidated)) |
0 | 989 break; |
990 | |
991 /* Changing the placement of a loop in the loop tree may alter the | |
992 validity of condition 2) of the description of fix_bb_placement | |
993 for its preheader, because the successor is the header and belongs | |
994 to the loop. So call fix_bb_placements to fix up the placement | |
995 of the preheader and (possibly) of its predecessors. */ | |
996 fix_bb_placements (loop_preheader_edge (loop)->src, | |
111 | 997 irred_invalidated, NULL); |
0 | 998 loop = outer; |
999 } | |
1000 } | |
1001 | |
111 | 1002 /* Duplicate loop bounds and other information we store about |
1003 the loop into its duplicate. */ | |
1004 | |
1005 void | |
145 | 1006 copy_loop_info (class loop *loop, class loop *target) |
111 | 1007 { |
1008 gcc_checking_assert (!target->any_upper_bound && !target->any_estimate); | |
1009 target->any_upper_bound = loop->any_upper_bound; | |
1010 target->nb_iterations_upper_bound = loop->nb_iterations_upper_bound; | |
1011 target->any_likely_upper_bound = loop->any_likely_upper_bound; | |
1012 target->nb_iterations_likely_upper_bound | |
1013 = loop->nb_iterations_likely_upper_bound; | |
1014 target->any_estimate = loop->any_estimate; | |
1015 target->nb_iterations_estimate = loop->nb_iterations_estimate; | |
1016 target->estimate_state = loop->estimate_state; | |
145 | 1017 target->safelen = loop->safelen; |
1018 target->simdlen = loop->simdlen; | |
111 | 1019 target->constraints = loop->constraints; |
145 | 1020 target->can_be_parallel = loop->can_be_parallel; |
111 | 1021 target->warned_aggressive_loop_optimizations |
1022 |= loop->warned_aggressive_loop_optimizations; | |
145 | 1023 target->dont_vectorize = loop->dont_vectorize; |
1024 target->force_vectorize = loop->force_vectorize; | |
111 | 1025 target->in_oacc_kernels_region = loop->in_oacc_kernels_region; |
145 | 1026 target->unroll = loop->unroll; |
1027 target->owned_clique = loop->owned_clique; | |
111 | 1028 } |
1029 | |
0 | 1030 /* Copies copy of LOOP as subloop of TARGET loop, placing newly |
131 | 1031 created loop into loops structure. If AFTER is non-null |
1032 the new loop is added at AFTER->next, otherwise in front of TARGETs | |
1033 sibling list. */ | |
145 | 1034 class loop * |
1035 duplicate_loop (class loop *loop, class loop *target, class loop *after) | |
0 | 1036 { |
145 | 1037 class loop *cloop; |
0 | 1038 cloop = alloc_loop (); |
111 | 1039 place_new_loop (cfun, cloop); |
1040 | |
1041 copy_loop_info (loop, cloop); | |
0 | 1042 |
1043 /* Mark the new loop as copy of LOOP. */ | |
1044 set_loop_copy (loop, cloop); | |
1045 | |
1046 /* Add it to target. */ | |
131 | 1047 flow_loop_tree_node_add (target, cloop, after); |
0 | 1048 |
1049 return cloop; | |
1050 } | |
1051 | |
1052 /* Copies structure of subloops of LOOP into TARGET loop, placing | |
131 | 1053 newly created loops into loop tree at the end of TARGETs sibling |
1054 list in the original order. */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1055 void |
145 | 1056 duplicate_subloops (class loop *loop, class loop *target) |
0 | 1057 { |
145 | 1058 class loop *aloop, *cloop, *tail; |
0 | 1059 |
131 | 1060 for (tail = target->inner; tail && tail->next; tail = tail->next) |
1061 ; | |
0 | 1062 for (aloop = loop->inner; aloop; aloop = aloop->next) |
1063 { | |
131 | 1064 cloop = duplicate_loop (aloop, target, tail); |
1065 tail = cloop; | |
1066 gcc_assert(!tail->next); | |
0 | 1067 duplicate_subloops (aloop, cloop); |
1068 } | |
1069 } | |
1070 | |
1071 /* Copies structure of subloops of N loops, stored in array COPIED_LOOPS, | |
131 | 1072 into TARGET loop, placing newly created loops into loop tree adding |
1073 them to TARGETs sibling list at the end in order. */ | |
0 | 1074 static void |
145 | 1075 copy_loops_to (class loop **copied_loops, int n, class loop *target) |
0 | 1076 { |
145 | 1077 class loop *aloop, *tail; |
0 | 1078 int i; |
1079 | |
131 | 1080 for (tail = target->inner; tail && tail->next; tail = tail->next) |
1081 ; | |
0 | 1082 for (i = 0; i < n; i++) |
1083 { | |
131 | 1084 aloop = duplicate_loop (copied_loops[i], target, tail); |
1085 tail = aloop; | |
1086 gcc_assert(!tail->next); | |
0 | 1087 duplicate_subloops (copied_loops[i], aloop); |
1088 } | |
1089 } | |
1090 | |
1091 /* Redirects edge E to basic block DEST. */ | |
1092 static void | |
1093 loop_redirect_edge (edge e, basic_block dest) | |
1094 { | |
1095 if (e->dest == dest) | |
1096 return; | |
1097 | |
1098 redirect_edge_and_branch_force (e, dest); | |
1099 } | |
1100 | |
1101 /* Check whether LOOP's body can be duplicated. */ | |
1102 bool | |
145 | 1103 can_duplicate_loop_p (const class loop *loop) |
0 | 1104 { |
1105 int ret; | |
1106 basic_block *bbs = get_loop_body (loop); | |
1107 | |
1108 ret = can_copy_bbs_p (bbs, loop->num_nodes); | |
1109 free (bbs); | |
1110 | |
1111 return ret; | |
1112 } | |
1113 | |
1114 /* Duplicates body of LOOP to given edge E NDUPL times. Takes care of updating | |
131 | 1115 loop structure and dominators (order of inner subloops is retained). |
1116 E's destination must be LOOP header for this to work, i.e. it must be entry | |
1117 or latch edge of this loop; these are unique, as the loops must have | |
1118 preheaders for this function to work correctly (in case E is latch, the | |
1119 function unrolls the loop, if E is entry edge, it peels the loop). Store | |
1120 edges created by copying ORIG edge from copies corresponding to set bits in | |
1121 WONT_EXIT bitmap (bit 0 corresponds to original LOOP body, the other copies | |
1122 are numbered in order given by control flow through them) into TO_REMOVE | |
1123 array. Returns false if duplication is | |
0 | 1124 impossible. */ |
1125 | |
1126 bool | |
145 | 1127 duplicate_loop_to_header_edge (class loop *loop, edge e, |
0 | 1128 unsigned int ndupl, sbitmap wont_exit, |
111 | 1129 edge orig, vec<edge> *to_remove, |
0 | 1130 int flags) |
1131 { | |
145 | 1132 class loop *target, *aloop; |
1133 class loop **orig_loops; | |
0 | 1134 unsigned n_orig_loops; |
1135 basic_block header = loop->header, latch = loop->latch; | |
1136 basic_block *new_bbs, *bbs, *first_active; | |
1137 basic_block new_bb, bb, first_active_latch = NULL; | |
1138 edge ae, latch_edge; | |
1139 edge spec_edges[2], new_spec_edges[2]; | |
131 | 1140 const int SE_LATCH = 0; |
1141 const int SE_ORIG = 1; | |
0 | 1142 unsigned i, j, n; |
1143 int is_latch = (latch == e->src); | |
131 | 1144 profile_probability *scale_step = NULL; |
1145 profile_probability scale_main = profile_probability::always (); | |
1146 profile_probability scale_act = profile_probability::always (); | |
1147 profile_count after_exit_num = profile_count::zero (), | |
1148 after_exit_den = profile_count::zero (); | |
1149 bool scale_after_exit = false; | |
0 | 1150 int add_irreducible_flag; |
1151 basic_block place_after; | |
1152 bitmap bbs_to_scale = NULL; | |
1153 bitmap_iterator bi; | |
1154 | |
1155 gcc_assert (e->dest == loop->header); | |
1156 gcc_assert (ndupl > 0); | |
1157 | |
1158 if (orig) | |
1159 { | |
1160 /* Orig must be edge out of the loop. */ | |
1161 gcc_assert (flow_bb_inside_loop_p (loop, orig->src)); | |
1162 gcc_assert (!flow_bb_inside_loop_p (loop, orig->dest)); | |
1163 } | |
1164 | |
1165 n = loop->num_nodes; | |
1166 bbs = get_loop_body_in_dom_order (loop); | |
1167 gcc_assert (bbs[0] == loop->header); | |
1168 gcc_assert (bbs[n - 1] == loop->latch); | |
1169 | |
1170 /* Check whether duplication is possible. */ | |
1171 if (!can_copy_bbs_p (bbs, loop->num_nodes)) | |
1172 { | |
1173 free (bbs); | |
1174 return false; | |
1175 } | |
1176 new_bbs = XNEWVEC (basic_block, loop->num_nodes); | |
1177 | |
1178 /* In case we are doing loop peeling and the loop is in the middle of | |
1179 irreducible region, the peeled copies will be inside it too. */ | |
1180 add_irreducible_flag = e->flags & EDGE_IRREDUCIBLE_LOOP; | |
1181 gcc_assert (!is_latch || !add_irreducible_flag); | |
1182 | |
1183 /* Find edge from latch. */ | |
1184 latch_edge = loop_latch_edge (loop); | |
1185 | |
1186 if (flags & DLTHE_FLAG_UPDATE_FREQ) | |
1187 { | |
131 | 1188 /* Calculate coefficients by that we have to scale counts |
0 | 1189 of duplicated loop bodies. */ |
131 | 1190 profile_count count_in = header->count; |
1191 profile_count count_le = latch_edge->count (); | |
1192 profile_count count_out_orig = orig ? orig->count () : count_in - count_le; | |
1193 profile_probability prob_pass_thru = count_le.probability_in (count_in); | |
1194 profile_probability prob_pass_wont_exit = | |
1195 (count_le + count_out_orig).probability_in (count_in); | |
0 | 1196 |
111 | 1197 if (orig && orig->probability.initialized_p () |
1198 && !(orig->probability == profile_probability::always ())) | |
0 | 1199 { |
1200 /* The blocks that are dominated by a removed exit edge ORIG have | |
1201 frequencies scaled by this. */ | |
131 | 1202 if (orig->count ().initialized_p ()) |
1203 { | |
1204 after_exit_num = orig->src->count; | |
1205 after_exit_den = after_exit_num - orig->count (); | |
1206 scale_after_exit = true; | |
1207 } | |
0 | 1208 bbs_to_scale = BITMAP_ALLOC (NULL); |
1209 for (i = 0; i < n; i++) | |
1210 { | |
1211 if (bbs[i] != orig->src | |
1212 && dominated_by_p (CDI_DOMINATORS, bbs[i], orig->src)) | |
1213 bitmap_set_bit (bbs_to_scale, i); | |
1214 } | |
1215 } | |
1216 | |
131 | 1217 scale_step = XNEWVEC (profile_probability, ndupl); |
0 | 1218 |
1219 for (i = 1; i <= ndupl; i++) | |
111 | 1220 scale_step[i - 1] = bitmap_bit_p (wont_exit, i) |
0 | 1221 ? prob_pass_wont_exit |
1222 : prob_pass_thru; | |
1223 | |
1224 /* Complete peeling is special as the probability of exit in last | |
1225 copy becomes 1. */ | |
1226 if (flags & DLTHE_FLAG_COMPLETTE_PEEL) | |
1227 { | |
131 | 1228 profile_count wanted_count = e->count (); |
0 | 1229 |
1230 gcc_assert (!is_latch); | |
131 | 1231 /* First copy has count of incoming edge. Each subsequent |
1232 count should be reduced by prob_pass_wont_exit. Caller | |
0 | 1233 should've managed the flags so all except for original loop |
1234 has won't exist set. */ | |
131 | 1235 scale_act = wanted_count.probability_in (count_in); |
0 | 1236 /* Now simulate the duplication adjustments and compute header |
1237 frequency of the last copy. */ | |
1238 for (i = 0; i < ndupl; i++) | |
131 | 1239 wanted_count = wanted_count.apply_probability (scale_step [i]); |
1240 scale_main = wanted_count.probability_in (count_in); | |
0 | 1241 } |
131 | 1242 /* Here we insert loop bodies inside the loop itself (for loop unrolling). |
1243 First iteration will be original loop followed by duplicated bodies. | |
1244 It is necessary to scale down the original so we get right overall | |
1245 number of iterations. */ | |
0 | 1246 else if (is_latch) |
1247 { | |
131 | 1248 profile_probability prob_pass_main = bitmap_bit_p (wont_exit, 0) |
1249 ? prob_pass_wont_exit | |
1250 : prob_pass_thru; | |
1251 profile_probability p = prob_pass_main; | |
1252 profile_count scale_main_den = count_in; | |
0 | 1253 for (i = 0; i < ndupl; i++) |
1254 { | |
131 | 1255 scale_main_den += count_in.apply_probability (p); |
1256 p = p * scale_step[i]; | |
0 | 1257 } |
131 | 1258 /* If original loop is executed COUNT_IN times, the unrolled |
1259 loop will account SCALE_MAIN_DEN times. */ | |
1260 scale_main = count_in.probability_in (scale_main_den); | |
1261 scale_act = scale_main * prob_pass_main; | |
0 | 1262 } |
1263 else | |
1264 { | |
131 | 1265 profile_count preheader_count = e->count (); |
0 | 1266 for (i = 0; i < ndupl; i++) |
131 | 1267 scale_main = scale_main * scale_step[i]; |
1268 scale_act = preheader_count.probability_in (count_in); | |
0 | 1269 } |
1270 } | |
1271 | |
1272 /* Loop the new bbs will belong to. */ | |
1273 target = e->src->loop_father; | |
1274 | |
1275 /* Original loops. */ | |
1276 n_orig_loops = 0; | |
1277 for (aloop = loop->inner; aloop; aloop = aloop->next) | |
1278 n_orig_loops++; | |
145 | 1279 orig_loops = XNEWVEC (class loop *, n_orig_loops); |
0 | 1280 for (aloop = loop->inner, i = 0; aloop; aloop = aloop->next, i++) |
1281 orig_loops[i] = aloop; | |
1282 | |
1283 set_loop_copy (loop, target); | |
1284 | |
1285 first_active = XNEWVEC (basic_block, n); | |
1286 if (is_latch) | |
1287 { | |
1288 memcpy (first_active, bbs, n * sizeof (basic_block)); | |
1289 first_active_latch = latch; | |
1290 } | |
1291 | |
1292 spec_edges[SE_ORIG] = orig; | |
1293 spec_edges[SE_LATCH] = latch_edge; | |
1294 | |
1295 place_after = e->src; | |
1296 for (j = 0; j < ndupl; j++) | |
1297 { | |
1298 /* Copy loops. */ | |
1299 copy_loops_to (orig_loops, n_orig_loops, target); | |
1300 | |
1301 /* Copy bbs. */ | |
1302 copy_bbs (bbs, n, new_bbs, spec_edges, 2, new_spec_edges, loop, | |
111 | 1303 place_after, true); |
0 | 1304 place_after = new_spec_edges[SE_LATCH]->src; |
1305 | |
1306 if (flags & DLTHE_RECORD_COPY_NUMBER) | |
1307 for (i = 0; i < n; i++) | |
1308 { | |
1309 gcc_assert (!new_bbs[i]->aux); | |
1310 new_bbs[i]->aux = (void *)(size_t)(j + 1); | |
1311 } | |
1312 | |
1313 /* Note whether the blocks and edges belong to an irreducible loop. */ | |
1314 if (add_irreducible_flag) | |
1315 { | |
1316 for (i = 0; i < n; i++) | |
1317 new_bbs[i]->flags |= BB_DUPLICATED; | |
1318 for (i = 0; i < n; i++) | |
1319 { | |
1320 edge_iterator ei; | |
1321 new_bb = new_bbs[i]; | |
1322 if (new_bb->loop_father == target) | |
1323 new_bb->flags |= BB_IRREDUCIBLE_LOOP; | |
1324 | |
1325 FOR_EACH_EDGE (ae, ei, new_bb->succs) | |
1326 if ((ae->dest->flags & BB_DUPLICATED) | |
1327 && (ae->src->loop_father == target | |
1328 || ae->dest->loop_father == target)) | |
1329 ae->flags |= EDGE_IRREDUCIBLE_LOOP; | |
1330 } | |
1331 for (i = 0; i < n; i++) | |
1332 new_bbs[i]->flags &= ~BB_DUPLICATED; | |
1333 } | |
1334 | |
1335 /* Redirect the special edges. */ | |
1336 if (is_latch) | |
1337 { | |
1338 redirect_edge_and_branch_force (latch_edge, new_bbs[0]); | |
1339 redirect_edge_and_branch_force (new_spec_edges[SE_LATCH], | |
1340 loop->header); | |
1341 set_immediate_dominator (CDI_DOMINATORS, new_bbs[0], latch); | |
1342 latch = loop->latch = new_bbs[n - 1]; | |
1343 e = latch_edge = new_spec_edges[SE_LATCH]; | |
1344 } | |
1345 else | |
1346 { | |
1347 redirect_edge_and_branch_force (new_spec_edges[SE_LATCH], | |
1348 loop->header); | |
1349 redirect_edge_and_branch_force (e, new_bbs[0]); | |
1350 set_immediate_dominator (CDI_DOMINATORS, new_bbs[0], e->src); | |
1351 e = new_spec_edges[SE_LATCH]; | |
1352 } | |
1353 | |
1354 /* Record exit edge in this copy. */ | |
111 | 1355 if (orig && bitmap_bit_p (wont_exit, j + 1)) |
0 | 1356 { |
1357 if (to_remove) | |
111 | 1358 to_remove->safe_push (new_spec_edges[SE_ORIG]); |
1359 force_edge_cold (new_spec_edges[SE_ORIG], true); | |
0 | 1360 |
1361 /* Scale the frequencies of the blocks dominated by the exit. */ | |
131 | 1362 if (bbs_to_scale && scale_after_exit) |
0 | 1363 { |
1364 EXECUTE_IF_SET_IN_BITMAP (bbs_to_scale, 0, i, bi) | |
131 | 1365 scale_bbs_frequencies_profile_count (new_bbs + i, 1, after_exit_num, |
1366 after_exit_den); | |
0 | 1367 } |
1368 } | |
1369 | |
1370 /* Record the first copy in the control flow order if it is not | |
1371 the original loop (i.e. in case of peeling). */ | |
1372 if (!first_active_latch) | |
1373 { | |
1374 memcpy (first_active, new_bbs, n * sizeof (basic_block)); | |
1375 first_active_latch = new_bbs[n - 1]; | |
1376 } | |
1377 | |
1378 /* Set counts and frequencies. */ | |
1379 if (flags & DLTHE_FLAG_UPDATE_FREQ) | |
1380 { | |
131 | 1381 scale_bbs_frequencies (new_bbs, n, scale_act); |
1382 scale_act = scale_act * scale_step[j]; | |
0 | 1383 } |
1384 } | |
1385 free (new_bbs); | |
1386 free (orig_loops); | |
1387 | |
1388 /* Record the exit edge in the original loop body, and update the frequencies. */ | |
111 | 1389 if (orig && bitmap_bit_p (wont_exit, 0)) |
0 | 1390 { |
1391 if (to_remove) | |
111 | 1392 to_remove->safe_push (orig); |
1393 force_edge_cold (orig, true); | |
0 | 1394 |
1395 /* Scale the frequencies of the blocks dominated by the exit. */ | |
131 | 1396 if (bbs_to_scale && scale_after_exit) |
0 | 1397 { |
1398 EXECUTE_IF_SET_IN_BITMAP (bbs_to_scale, 0, i, bi) | |
131 | 1399 scale_bbs_frequencies_profile_count (bbs + i, 1, after_exit_num, |
1400 after_exit_den); | |
0 | 1401 } |
1402 } | |
1403 | |
1404 /* Update the original loop. */ | |
1405 if (!is_latch) | |
1406 set_immediate_dominator (CDI_DOMINATORS, e->dest, e->src); | |
1407 if (flags & DLTHE_FLAG_UPDATE_FREQ) | |
1408 { | |
131 | 1409 scale_bbs_frequencies (bbs, n, scale_main); |
0 | 1410 free (scale_step); |
1411 } | |
1412 | |
1413 /* Update dominators of outer blocks if affected. */ | |
1414 for (i = 0; i < n; i++) | |
1415 { | |
1416 basic_block dominated, dom_bb; | |
111 | 1417 vec<basic_block> dom_bbs; |
0 | 1418 unsigned j; |
1419 | |
1420 bb = bbs[i]; | |
1421 bb->aux = 0; | |
1422 | |
1423 dom_bbs = get_dominated_by (CDI_DOMINATORS, bb); | |
111 | 1424 FOR_EACH_VEC_ELT (dom_bbs, j, dominated) |
0 | 1425 { |
1426 if (flow_bb_inside_loop_p (loop, dominated)) | |
1427 continue; | |
1428 dom_bb = nearest_common_dominator ( | |
1429 CDI_DOMINATORS, first_active[i], first_active_latch); | |
1430 set_immediate_dominator (CDI_DOMINATORS, dominated, dom_bb); | |
1431 } | |
111 | 1432 dom_bbs.release (); |
0 | 1433 } |
1434 free (first_active); | |
1435 | |
1436 free (bbs); | |
1437 BITMAP_FREE (bbs_to_scale); | |
1438 | |
1439 return true; | |
1440 } | |
1441 | |
1442 /* A callback for make_forwarder block, to redirect all edges except for | |
1443 MFB_KJ_EDGE to the entry part. E is the edge for that we should decide | |
1444 whether to redirect it. */ | |
1445 | |
1446 edge mfb_kj_edge; | |
1447 bool | |
1448 mfb_keep_just (edge e) | |
1449 { | |
1450 return e != mfb_kj_edge; | |
1451 } | |
1452 | |
1453 /* True when a candidate preheader BLOCK has predecessors from LOOP. */ | |
1454 | |
1455 static bool | |
145 | 1456 has_preds_from_loop (basic_block block, class loop *loop) |
0 | 1457 { |
1458 edge e; | |
1459 edge_iterator ei; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1460 |
0 | 1461 FOR_EACH_EDGE (e, ei, block->preds) |
1462 if (e->src->loop_father == loop) | |
1463 return true; | |
1464 return false; | |
1465 } | |
1466 | |
1467 /* Creates a pre-header for a LOOP. Returns newly created block. Unless | |
1468 CP_SIMPLE_PREHEADERS is set in FLAGS, we only force LOOP to have single | |
1469 entry; otherwise we also force preheader block to have only one successor. | |
1470 When CP_FALLTHRU_PREHEADERS is set in FLAGS, we force the preheader block | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1471 to be a fallthru predecessor to the loop header and to have only |
0 | 1472 predecessors from outside of the loop. |
1473 The function also updates dominators. */ | |
1474 | |
1475 basic_block | |
145 | 1476 create_preheader (class loop *loop, int flags) |
0 | 1477 { |
111 | 1478 edge e; |
0 | 1479 basic_block dummy; |
1480 int nentry = 0; | |
1481 bool irred = false; | |
1482 bool latch_edge_was_fallthru; | |
1483 edge one_succ_pred = NULL, single_entry = NULL; | |
1484 edge_iterator ei; | |
1485 | |
1486 FOR_EACH_EDGE (e, ei, loop->header->preds) | |
1487 { | |
1488 if (e->src == loop->latch) | |
1489 continue; | |
1490 irred |= (e->flags & EDGE_IRREDUCIBLE_LOOP) != 0; | |
1491 nentry++; | |
1492 single_entry = e; | |
1493 if (single_succ_p (e->src)) | |
1494 one_succ_pred = e; | |
1495 } | |
1496 gcc_assert (nentry); | |
1497 if (nentry == 1) | |
1498 { | |
1499 bool need_forwarder_block = false; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1500 |
0 | 1501 /* We do not allow entry block to be the loop preheader, since we |
1502 cannot emit code there. */ | |
111 | 1503 if (single_entry->src == ENTRY_BLOCK_PTR_FOR_FN (cfun)) |
0 | 1504 need_forwarder_block = true; |
1505 else | |
1506 { | |
1507 /* If we want simple preheaders, also force the preheader to have | |
1508 just a single successor. */ | |
1509 if ((flags & CP_SIMPLE_PREHEADERS) | |
1510 && !single_succ_p (single_entry->src)) | |
1511 need_forwarder_block = true; | |
1512 /* If we want fallthru preheaders, also create forwarder block when | |
1513 preheader ends with a jump or has predecessors from loop. */ | |
1514 else if ((flags & CP_FALLTHRU_PREHEADERS) | |
1515 && (JUMP_P (BB_END (single_entry->src)) | |
1516 || has_preds_from_loop (single_entry->src, loop))) | |
1517 need_forwarder_block = true; | |
1518 } | |
1519 if (! need_forwarder_block) | |
1520 return NULL; | |
1521 } | |
1522 | |
1523 mfb_kj_edge = loop_latch_edge (loop); | |
1524 latch_edge_was_fallthru = (mfb_kj_edge->flags & EDGE_FALLTHRU) != 0; | |
131 | 1525 if (nentry == 1 |
1526 && ((flags & CP_FALLTHRU_PREHEADERS) == 0 | |
1527 || (single_entry->flags & EDGE_CROSSING) == 0)) | |
111 | 1528 dummy = split_edge (single_entry); |
1529 else | |
1530 { | |
1531 edge fallthru = make_forwarder_block (loop->header, mfb_keep_just, NULL); | |
1532 dummy = fallthru->src; | |
1533 loop->header = fallthru->dest; | |
1534 } | |
0 | 1535 |
1536 /* Try to be clever in placing the newly created preheader. The idea is to | |
1537 avoid breaking any "fallthruness" relationship between blocks. | |
1538 | |
1539 The preheader was created just before the header and all incoming edges | |
1540 to the header were redirected to the preheader, except the latch edge. | |
1541 So the only problematic case is when this latch edge was a fallthru | |
1542 edge: it is not anymore after the preheader creation so we have broken | |
1543 the fallthruness. We're therefore going to look for a better place. */ | |
1544 if (latch_edge_was_fallthru) | |
1545 { | |
1546 if (one_succ_pred) | |
1547 e = one_succ_pred; | |
1548 else | |
1549 e = EDGE_PRED (dummy, 0); | |
1550 | |
1551 move_block_after (dummy, e->src); | |
1552 } | |
1553 | |
1554 if (irred) | |
1555 { | |
1556 dummy->flags |= BB_IRREDUCIBLE_LOOP; | |
1557 single_succ_edge (dummy)->flags |= EDGE_IRREDUCIBLE_LOOP; | |
1558 } | |
1559 | |
1560 if (dump_file) | |
1561 fprintf (dump_file, "Created preheader block for loop %i\n", | |
1562 loop->num); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1563 |
0 | 1564 if (flags & CP_FALLTHRU_PREHEADERS) |
1565 gcc_assert ((single_succ_edge (dummy)->flags & EDGE_FALLTHRU) | |
1566 && !JUMP_P (BB_END (dummy))); | |
1567 | |
1568 return dummy; | |
1569 } | |
1570 | |
1571 /* Create preheaders for each loop; for meaning of FLAGS see create_preheader. */ | |
1572 | |
1573 void | |
1574 create_preheaders (int flags) | |
1575 { | |
145 | 1576 class loop *loop; |
0 | 1577 |
1578 if (!current_loops) | |
1579 return; | |
1580 | |
111 | 1581 FOR_EACH_LOOP (loop, 0) |
0 | 1582 create_preheader (loop, flags); |
1583 loops_state_set (LOOPS_HAVE_PREHEADERS); | |
1584 } | |
1585 | |
1586 /* Forces all loop latches to have only single successor. */ | |
1587 | |
1588 void | |
1589 force_single_succ_latches (void) | |
1590 { | |
145 | 1591 class loop *loop; |
0 | 1592 edge e; |
1593 | |
111 | 1594 FOR_EACH_LOOP (loop, 0) |
0 | 1595 { |
1596 if (loop->latch != loop->header && single_succ_p (loop->latch)) | |
1597 continue; | |
1598 | |
1599 e = find_edge (loop->latch, loop->header); | |
111 | 1600 gcc_checking_assert (e != NULL); |
0 | 1601 |
1602 split_edge (e); | |
1603 } | |
1604 loops_state_set (LOOPS_HAVE_SIMPLE_LATCHES); | |
1605 } | |
1606 | |
1607 /* This function is called from loop_version. It splits the entry edge | |
1608 of the loop we want to version, adds the versioning condition, and | |
1609 adjust the edges to the two versions of the loop appropriately. | |
1610 e is an incoming edge. Returns the basic block containing the | |
1611 condition. | |
1612 | |
1613 --- edge e ---- > [second_head] | |
1614 | |
1615 Split it and insert new conditional expression and adjust edges. | |
1616 | |
1617 --- edge e ---> [cond expr] ---> [first_head] | |
1618 | | |
1619 +---------> [second_head] | |
1620 | |
111 | 1621 THEN_PROB is the probability of then branch of the condition. |
1622 ELSE_PROB is the probability of else branch. Note that they may be both | |
1623 REG_BR_PROB_BASE when condition is IFN_LOOP_VECTORIZED or | |
1624 IFN_LOOP_DIST_ALIAS. */ | |
0 | 1625 |
1626 static basic_block | |
1627 lv_adjust_loop_entry_edge (basic_block first_head, basic_block second_head, | |
111 | 1628 edge e, void *cond_expr, |
1629 profile_probability then_prob, | |
1630 profile_probability else_prob) | |
0 | 1631 { |
1632 basic_block new_head = NULL; | |
1633 edge e1; | |
1634 | |
1635 gcc_assert (e->dest == second_head); | |
1636 | |
1637 /* Split edge 'e'. This will create a new basic block, where we can | |
1638 insert conditional expr. */ | |
1639 new_head = split_edge (e); | |
1640 | |
1641 lv_add_condition_to_bb (first_head, second_head, new_head, | |
1642 cond_expr); | |
1643 | |
1644 /* Don't set EDGE_TRUE_VALUE in RTL mode, as it's invalid there. */ | |
1645 e = single_succ_edge (new_head); | |
1646 e1 = make_edge (new_head, first_head, | |
1647 current_ir_type () == IR_GIMPLE ? EDGE_TRUE_VALUE : 0); | |
1648 e1->probability = then_prob; | |
111 | 1649 e->probability = else_prob; |
0 | 1650 |
1651 set_immediate_dominator (CDI_DOMINATORS, first_head, new_head); | |
1652 set_immediate_dominator (CDI_DOMINATORS, second_head, new_head); | |
1653 | |
1654 /* Adjust loop header phi nodes. */ | |
1655 lv_adjust_loop_header_phi (first_head, second_head, new_head, e1); | |
1656 | |
1657 return new_head; | |
1658 } | |
1659 | |
1660 /* Main entry point for Loop Versioning transformation. | |
1661 | |
1662 This transformation given a condition and a loop, creates | |
1663 -if (condition) { loop_copy1 } else { loop_copy2 }, | |
1664 where loop_copy1 is the loop transformed in one way, and loop_copy2 | |
111 | 1665 is the loop transformed in another way (or unchanged). COND_EXPR |
0 | 1666 may be a run time test for things that were not resolved by static |
1667 analysis (overlapping ranges (anti-aliasing), alignment, etc.). | |
1668 | |
111 | 1669 If non-NULL, CONDITION_BB is set to the basic block containing the |
1670 condition. | |
1671 | |
0 | 1672 THEN_PROB is the probability of the then edge of the if. THEN_SCALE |
1673 is the ratio by that the frequencies in the original loop should | |
1674 be scaled. ELSE_SCALE is the ratio by that the frequencies in the | |
1675 new loop should be scaled. | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1676 |
0 | 1677 If PLACE_AFTER is true, we place the new loop after LOOP in the |
1678 instruction stream, otherwise it is placed before LOOP. */ | |
1679 | |
145 | 1680 class loop * |
1681 loop_version (class loop *loop, | |
0 | 1682 void *cond_expr, basic_block *condition_bb, |
111 | 1683 profile_probability then_prob, profile_probability else_prob, |
1684 profile_probability then_scale, profile_probability else_scale, | |
0 | 1685 bool place_after) |
1686 { | |
1687 basic_block first_head, second_head; | |
1688 edge entry, latch_edge, true_edge, false_edge; | |
1689 int irred_flag; | |
145 | 1690 class loop *nloop; |
0 | 1691 basic_block cond_bb; |
1692 | |
1693 /* Record entry and latch edges for the loop */ | |
1694 entry = loop_preheader_edge (loop); | |
1695 irred_flag = entry->flags & EDGE_IRREDUCIBLE_LOOP; | |
1696 entry->flags &= ~EDGE_IRREDUCIBLE_LOOP; | |
1697 | |
1698 /* Note down head of loop as first_head. */ | |
1699 first_head = entry->dest; | |
1700 | |
1701 /* Duplicate loop. */ | |
1702 if (!cfg_hook_duplicate_loop_to_header_edge (loop, entry, 1, | |
1703 NULL, NULL, NULL, 0)) | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1704 { |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1705 entry->flags |= irred_flag; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1706 return NULL; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1707 } |
0 | 1708 |
1709 /* After duplication entry edge now points to new loop head block. | |
1710 Note down new head as second_head. */ | |
1711 second_head = entry->dest; | |
1712 | |
1713 /* Split loop entry edge and insert new block with cond expr. */ | |
1714 cond_bb = lv_adjust_loop_entry_edge (first_head, second_head, | |
111 | 1715 entry, cond_expr, then_prob, else_prob); |
0 | 1716 if (condition_bb) |
1717 *condition_bb = cond_bb; | |
1718 | |
1719 if (!cond_bb) | |
1720 { | |
1721 entry->flags |= irred_flag; | |
1722 return NULL; | |
1723 } | |
1724 | |
1725 latch_edge = single_succ_edge (get_bb_copy (loop->latch)); | |
1726 | |
1727 extract_cond_bb_edges (cond_bb, &true_edge, &false_edge); | |
1728 nloop = loopify (latch_edge, | |
1729 single_pred_edge (get_bb_copy (loop->header)), | |
1730 cond_bb, true_edge, false_edge, | |
1731 false /* Do not redirect all edges. */, | |
1732 then_scale, else_scale); | |
1733 | |
111 | 1734 copy_loop_info (loop, nloop); |
1735 | |
0 | 1736 /* loopify redirected latch_edge. Update its PENDING_STMTS. */ |
1737 lv_flush_pending_stmts (latch_edge); | |
1738 | |
1739 /* loopify redirected condition_bb's succ edge. Update its PENDING_STMTS. */ | |
1740 extract_cond_bb_edges (cond_bb, &true_edge, &false_edge); | |
1741 lv_flush_pending_stmts (false_edge); | |
1742 /* Adjust irreducible flag. */ | |
1743 if (irred_flag) | |
1744 { | |
1745 cond_bb->flags |= BB_IRREDUCIBLE_LOOP; | |
1746 loop_preheader_edge (loop)->flags |= EDGE_IRREDUCIBLE_LOOP; | |
1747 loop_preheader_edge (nloop)->flags |= EDGE_IRREDUCIBLE_LOOP; | |
1748 single_pred_edge (cond_bb)->flags |= EDGE_IRREDUCIBLE_LOOP; | |
1749 } | |
1750 | |
1751 if (place_after) | |
1752 { | |
1753 basic_block *bbs = get_loop_body_in_dom_order (nloop), after; | |
1754 unsigned i; | |
1755 | |
1756 after = loop->latch; | |
1757 | |
1758 for (i = 0; i < nloop->num_nodes; i++) | |
1759 { | |
1760 move_block_after (bbs[i], after); | |
1761 after = bbs[i]; | |
1762 } | |
1763 free (bbs); | |
1764 } | |
1765 | |
1766 /* At this point condition_bb is loop preheader with two successors, | |
1767 first_head and second_head. Make sure that loop preheader has only | |
1768 one successor. */ | |
1769 split_edge (loop_preheader_edge (loop)); | |
1770 split_edge (loop_preheader_edge (nloop)); | |
1771 | |
1772 return nloop; | |
1773 } |