Mercurial > hg > CbC > CbC_gcc
annotate gcc/cfgloopmanip.c @ 127:4c56639505ff
fix function.c and add CbC-example Makefile
author | mir3636 |
---|---|
date | Wed, 11 Apr 2018 18:46:58 +0900 |
parents | 04ced10e8804 |
children | 84e7813d76e9 |
rev | line source |
---|---|
0 | 1 /* Loop manipulation code for GNU compiler. |
111 | 2 Copyright (C) 2002-2017 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 |
35 static void copy_loops_to (struct loop **, int, | |
36 struct loop *); | |
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 **); | |
41 static void fix_loop_placements (struct loop *, bool *); | |
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; | |
92 struct loop *loop = current_loops->tree_root, *act; | |
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 | |
111 | 125 fix_loop_placement (struct loop *loop, bool *irred_invalidated) |
0 | 126 { |
127 unsigned i; | |
128 edge e; | |
111 | 129 vec<edge> exits = get_loop_exit_edges (loop); |
0 | 130 struct loop *father = current_loops->tree_root, *act; |
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; | |
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
|
185 struct 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; | |
258 struct loop *nca; | |
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; | |
310 struct loop *l, *f; | |
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 bb = rem_bbs[i]; | |
368 FOR_EACH_EDGE (ae, ei, rem_bbs[i]->succs) | |
111 | 369 if (ae->dest != EXIT_BLOCK_PTR_FOR_FN (cfun) |
370 && !bitmap_bit_p (seen, ae->dest->index)) | |
0 | 371 { |
111 | 372 bitmap_set_bit (seen, ae->dest->index); |
0 | 373 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
|
374 |
0 | 375 if (ae->flags & EDGE_IRREDUCIBLE_LOOP) |
111 | 376 *irred_invalidated = true; |
0 | 377 } |
378 } | |
379 | |
380 /* Remove the path. */ | |
381 from = e->src; | |
382 remove_branch (e); | |
111 | 383 dom_bbs.create (0); |
0 | 384 |
385 /* Cancel loops contained in the path. */ | |
386 for (i = 0; i < nrem; i++) | |
387 if (rem_bbs[i]->loop_father->header == rem_bbs[i]) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
388 cancel_loop_tree (rem_bbs[i]->loop_father); |
0 | 389 |
390 remove_bbs (rem_bbs, nrem); | |
391 free (rem_bbs); | |
392 | |
393 /* Find blocks whose dominators may be affected. */ | |
111 | 394 bitmap_clear (seen); |
0 | 395 for (i = 0; i < n_bord_bbs; i++) |
396 { | |
397 basic_block ldom; | |
398 | |
399 bb = get_immediate_dominator (CDI_DOMINATORS, bord_bbs[i]); | |
111 | 400 if (bitmap_bit_p (seen, bb->index)) |
0 | 401 continue; |
111 | 402 bitmap_set_bit (seen, bb->index); |
0 | 403 |
404 for (ldom = first_dom_son (CDI_DOMINATORS, bb); | |
405 ldom; | |
406 ldom = next_dom_son (CDI_DOMINATORS, ldom)) | |
407 if (!dominated_by_p (CDI_DOMINATORS, from, ldom)) | |
111 | 408 dom_bbs.safe_push (ldom); |
0 | 409 } |
410 | |
411 /* Recount dominators. */ | |
412 iterate_fix_dominators (CDI_DOMINATORS, dom_bbs, true); | |
111 | 413 dom_bbs.release (); |
0 | 414 free (bord_bbs); |
415 | |
416 /* Fix placements of basic blocks inside loops and the placement of | |
417 loops in the loop tree. */ | |
111 | 418 fix_bb_placements (from, irred_invalidated, loop_closed_ssa_invalidated); |
419 fix_loop_placements (from->loop_father, irred_invalidated); | |
0 | 420 |
111 | 421 if (local_irred_invalidated |
0 | 422 && loops_state_satisfies_p (LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS)) |
423 mark_irreducible_loops (); | |
424 | |
425 return true; | |
426 } | |
427 | |
111 | 428 /* Creates place for a new LOOP in loops structure of FN. */ |
0 | 429 |
111 | 430 void |
431 place_new_loop (struct function *fn, struct loop *loop) | |
0 | 432 { |
111 | 433 loop->num = number_of_loops (fn); |
434 vec_safe_push (loops_for_fn (fn)->larray, loop); | |
0 | 435 } |
436 | |
437 /* Given LOOP structure with filled header and latch, find the body of the | |
438 corresponding loop and add it to loops tree. Insert the LOOP as a son of | |
439 outer. */ | |
440 | |
441 void | |
442 add_loop (struct loop *loop, struct loop *outer) | |
443 { | |
444 basic_block *bbs; | |
445 int i, n; | |
446 struct loop *subloop; | |
447 edge e; | |
448 edge_iterator ei; | |
449 | |
450 /* Add it to loop structure. */ | |
111 | 451 place_new_loop (cfun, loop); |
0 | 452 flow_loop_tree_node_add (outer, loop); |
453 | |
454 /* Find its nodes. */ | |
111 | 455 bbs = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun)); |
456 n = get_loop_body_with_size (loop, bbs, n_basic_blocks_for_fn (cfun)); | |
0 | 457 |
458 for (i = 0; i < n; i++) | |
459 { | |
460 if (bbs[i]->loop_father == outer) | |
461 { | |
462 remove_bb_from_loops (bbs[i]); | |
463 add_bb_to_loop (bbs[i], loop); | |
464 continue; | |
465 } | |
466 | |
467 loop->num_nodes++; | |
468 | |
469 /* If we find a direct subloop of OUTER, move it to LOOP. */ | |
470 subloop = bbs[i]->loop_father; | |
471 if (loop_outer (subloop) == outer | |
472 && subloop->header == bbs[i]) | |
473 { | |
474 flow_loop_tree_node_remove (subloop); | |
475 flow_loop_tree_node_add (loop, subloop); | |
476 } | |
477 } | |
478 | |
479 /* Update the information about loop exit edges. */ | |
480 for (i = 0; i < n; i++) | |
481 { | |
482 FOR_EACH_EDGE (e, ei, bbs[i]->succs) | |
483 { | |
484 rescan_loop_exit (e, false, false); | |
485 } | |
486 } | |
487 | |
488 free (bbs); | |
489 } | |
490 | |
111 | 491 /* Scale profile of loop by P. */ |
492 | |
0 | 493 void |
111 | 494 scale_loop_frequencies (struct loop *loop, profile_probability p) |
0 | 495 { |
496 basic_block *bbs; | |
497 | |
498 bbs = get_loop_body (loop); | |
111 | 499 scale_bbs_frequencies (bbs, loop->num_nodes, p); |
0 | 500 free (bbs); |
501 } | |
502 | |
111 | 503 /* Scale profile in LOOP by P. |
504 If ITERATION_BOUND is non-zero, scale even further if loop is predicted | |
505 to iterate too many times. */ | |
506 | |
507 void | |
508 scale_loop_profile (struct loop *loop, profile_probability p, | |
509 gcov_type iteration_bound) | |
510 { | |
511 gcov_type iterations = expected_loop_iterations_unbounded (loop); | |
512 edge e; | |
513 edge_iterator ei; | |
514 | |
515 if (dump_file && (dump_flags & TDF_DETAILS)) | |
516 { | |
517 fprintf (dump_file, ";; Scaling loop %i with scale ", | |
518 loop->num); | |
519 p.dump (dump_file); | |
520 fprintf (dump_file, " bounding iterations to %i from guessed %i\n", | |
521 (int)iteration_bound, (int)iterations); | |
522 } | |
523 | |
524 /* See if loop is predicted to iterate too many times. */ | |
525 if (iteration_bound && iterations > 0 | |
526 && p.apply (iterations) > iteration_bound) | |
527 { | |
528 /* Fixing loop profile for different trip count is not trivial; the exit | |
529 probabilities has to be updated to match and frequencies propagated down | |
530 to the loop body. | |
531 | |
532 We fully update only the simple case of loop with single exit that is | |
533 either from the latch or BB just before latch and leads from BB with | |
534 simple conditional jump. This is OK for use in vectorizer. */ | |
535 e = single_exit (loop); | |
536 if (e) | |
537 { | |
538 edge other_e; | |
539 int freq_delta; | |
540 profile_count count_delta; | |
541 | |
542 FOR_EACH_EDGE (other_e, ei, e->src->succs) | |
543 if (!(other_e->flags & (EDGE_ABNORMAL | EDGE_FAKE)) | |
544 && e != other_e) | |
545 break; | |
546 | |
547 /* Probability of exit must be 1/iterations. */ | |
548 freq_delta = EDGE_FREQUENCY (e); | |
549 count_delta = e->count (); | |
550 e->probability = profile_probability::always () | |
551 .apply_scale (1, iteration_bound); | |
552 other_e->probability = e->probability.invert (); | |
553 freq_delta -= EDGE_FREQUENCY (e); | |
554 count_delta -= e->count (); | |
555 | |
556 /* If latch exists, change its frequency and count, since we changed | |
557 probability of exit. Theoretically we should update everything from | |
558 source of exit edge to latch, but for vectorizer this is enough. */ | |
559 if (loop->latch | |
560 && loop->latch != e->src) | |
561 { | |
562 loop->latch->frequency += freq_delta; | |
563 if (loop->latch->frequency < 0) | |
564 loop->latch->frequency = 0; | |
565 loop->latch->count += count_delta; | |
566 } | |
567 } | |
568 | |
569 /* Roughly speaking we want to reduce the loop body profile by the | |
570 difference of loop iterations. We however can do better if | |
571 we look at the actual profile, if it is available. */ | |
572 p = p.apply_scale (iteration_bound, iterations); | |
573 | |
574 bool determined = false; | |
575 if (loop->header->count.initialized_p ()) | |
576 { | |
577 profile_count count_in = profile_count::zero (); | |
578 | |
579 FOR_EACH_EDGE (e, ei, loop->header->preds) | |
580 if (e->src != loop->latch) | |
581 count_in += e->count (); | |
582 | |
583 if (count_in > profile_count::zero () ) | |
584 { | |
585 p = count_in.probability_in (loop->header->count.apply_scale | |
586 (iteration_bound, 1)); | |
587 determined = true; | |
588 } | |
589 } | |
590 if (!determined && loop->header->frequency) | |
591 { | |
592 int freq_in = 0; | |
593 | |
594 FOR_EACH_EDGE (e, ei, loop->header->preds) | |
595 if (e->src != loop->latch) | |
596 freq_in += EDGE_FREQUENCY (e); | |
597 | |
598 if (freq_in != 0) | |
599 p = profile_probability::probability_in_gcov_type | |
600 (freq_in * iteration_bound, loop->header->frequency); | |
601 } | |
602 if (!(p > profile_probability::never ())) | |
603 p = profile_probability::very_unlikely (); | |
604 } | |
605 | |
606 if (p >= profile_probability::always () | |
607 || !p.initialized_p ()) | |
608 return; | |
609 | |
610 /* Scale the actual probabilities. */ | |
611 scale_loop_frequencies (loop, p); | |
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)); | |
615 } | |
616 | |
0 | 617 /* Recompute dominance information for basic blocks outside LOOP. */ |
618 | |
619 static void | |
620 update_dominators_in_loop (struct loop *loop) | |
621 { | |
111 | 622 vec<basic_block> dom_bbs = vNULL; |
0 | 623 basic_block *body; |
624 unsigned i; | |
625 | |
111 | 626 auto_sbitmap seen (last_basic_block_for_fn (cfun)); |
627 bitmap_clear (seen); | |
0 | 628 body = get_loop_body (loop); |
629 | |
630 for (i = 0; i < loop->num_nodes; i++) | |
111 | 631 bitmap_set_bit (seen, body[i]->index); |
0 | 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)) | |
111 | 640 if (!bitmap_bit_p (seen, ldom->index)) |
0 | 641 { |
111 | 642 bitmap_set_bit (seen, ldom->index); |
643 dom_bbs.safe_push (ldom); | |
0 | 644 } |
645 } | |
646 | |
647 iterate_fix_dominators (CDI_DOMINATORS, dom_bbs, false); | |
648 free (body); | |
111 | 649 dom_bbs.release (); |
0 | 650 } |
651 | |
652 /* 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
|
653 the test for the if. |
0 | 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 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
690 basic_block cond_bb, true_bb, false_bb, join_bb; |
0 | 691 edge e_true, e_false, exit_edge; |
111 | 692 gcond *cond_stmt; |
0 | 693 tree simple_cond; |
694 gimple_stmt_iterator gsi; | |
695 | |
696 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
|
697 |
0 | 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); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
706 |
0 | 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 | | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
735 | - 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
|
736 | | | | 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
|
737 | -----|----- ---------|----------- |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
738 | | ______ | entry_edge |
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 | | ====> | -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
|
741 | 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
|
742 | - succ_bb - | ---|----------------------------- |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
743 | | | | | |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
744 | ----------- | ---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
|
745 | | | 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
|
746 | | | 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
|
747 | | ---|--------------\-------------- |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
748 | | | \ exit_e |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
749 | | V \ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
750 | | - 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
|
751 | | | | | | |
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 | \ ___ / |
0 | 754 |
755 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
|
756 that is used before the increment of IV. IV_BEFORE should be used for |
0 | 757 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
|
758 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
|
759 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
760 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
|
761 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
|
762 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
|
763 invariant. */ |
0 | 764 |
765 struct loop * | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
766 create_empty_loop_on_edge (edge entry_edge, |
0 | 767 tree initial_value, |
768 tree stride, tree upper_bound, | |
769 tree iv, | |
770 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
|
771 tree *iv_after, |
0 | 772 struct loop *outer) |
773 { | |
774 basic_block loop_header, loop_latch, succ_bb, pred_bb; | |
775 struct loop *loop; | |
776 gimple_stmt_iterator gsi; | |
777 gimple_seq stmts; | |
111 | 778 gcond *cond_expr; |
0 | 779 tree exit_test; |
780 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
|
781 |
0 | 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 | |
803 /* TODO: Fix frequencies and counts. */ | |
111 | 804 scale_loop_frequencies (loop, profile_probability::even ()); |
0 | 805 |
806 /* Update dominators. */ | |
807 update_dominators_in_loop (loop); | |
808 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
809 /* 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
|
810 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
|
811 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
|
812 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
|
813 |
0 | 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 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
822 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
|
823 if (stmts) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
824 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
825 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
|
826 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
|
827 } |
0 | 828 |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
829 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
|
830 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
|
831 iv_before, iv_after); |
0 | 832 |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
833 /* 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
|
834 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
|
835 (LT_EXPR, *iv_before, upper_bound, NULL_TREE, NULL_TREE); |
0 | 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 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
844 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
|
845 |
0 | 846 return loop; |
847 } | |
848 | |
849 /* Make area between HEADER_EDGE and LATCH_EDGE a loop by connecting | |
850 latch to header and update loop tree and dominators | |
851 accordingly. Everything between them plus LATCH_EDGE destination must | |
852 be dominated by HEADER_EDGE destination, and back-reachable from | |
853 LATCH_EDGE source. HEADER_EDGE is redirected to basic block SWITCH_BB, | |
854 FALSE_EDGE of SWITCH_BB to original destination of HEADER_EDGE and | |
855 TRUE_EDGE of SWITCH_BB to original destination of LATCH_EDGE. | |
856 Returns the newly created loop. Frequencies and counts in the new loop | |
857 are scaled by FALSE_SCALE and in the old one by TRUE_SCALE. */ | |
858 | |
859 struct loop * | |
860 loopify (edge latch_edge, edge header_edge, | |
861 basic_block switch_bb, edge true_edge, edge false_edge, | |
111 | 862 bool redirect_all_edges, profile_probability true_scale, |
863 profile_probability false_scale) | |
0 | 864 { |
865 basic_block succ_bb = latch_edge->dest; | |
866 basic_block pred_bb = header_edge->src; | |
867 struct loop *loop = alloc_loop (); | |
868 struct loop *outer = loop_outer (succ_bb->loop_father); | |
869 int freq; | |
111 | 870 profile_count cnt; |
0 | 871 |
872 loop->header = header_edge->dest; | |
873 loop->latch = latch_edge->src; | |
874 | |
875 freq = EDGE_FREQUENCY (header_edge); | |
111 | 876 cnt = header_edge->count (); |
0 | 877 |
878 /* Redirect edges. */ | |
879 loop_redirect_edge (latch_edge, loop->header); | |
880 loop_redirect_edge (true_edge, succ_bb); | |
881 | |
882 /* During loop versioning, one of the switch_bb edge is already properly | |
883 set. Do not redirect it again unless redirect_all_edges is true. */ | |
884 if (redirect_all_edges) | |
885 { | |
886 loop_redirect_edge (header_edge, switch_bb); | |
887 loop_redirect_edge (false_edge, loop->header); | |
888 | |
889 /* Update dominators. */ | |
890 set_immediate_dominator (CDI_DOMINATORS, switch_bb, pred_bb); | |
891 set_immediate_dominator (CDI_DOMINATORS, loop->header, switch_bb); | |
892 } | |
893 | |
894 set_immediate_dominator (CDI_DOMINATORS, succ_bb, switch_bb); | |
895 | |
896 /* Compute new loop. */ | |
897 add_loop (loop, outer); | |
898 | |
899 /* Add switch_bb to appropriate loop. */ | |
900 if (switch_bb->loop_father) | |
901 remove_bb_from_loops (switch_bb); | |
902 add_bb_to_loop (switch_bb, outer); | |
903 | |
904 /* Fix frequencies. */ | |
905 if (redirect_all_edges) | |
906 { | |
907 switch_bb->frequency = freq; | |
908 switch_bb->count = cnt; | |
909 } | |
111 | 910 scale_loop_frequencies (loop, false_scale); |
911 scale_loop_frequencies (succ_bb->loop_father, true_scale); | |
0 | 912 update_dominators_in_loop (loop); |
913 | |
914 return loop; | |
915 } | |
916 | |
917 /* Remove the latch edge of a LOOP and update loops to indicate that | |
918 the LOOP was removed. After this function, original loop latch will | |
919 have no successor, which caller is expected to fix somehow. | |
920 | |
921 If this may cause the information about irreducible regions to become | |
111 | 922 invalid, IRRED_INVALIDATED is set to true. |
0 | 923 |
111 | 924 LOOP_CLOSED_SSA_INVALIDATED, if non-NULL, is a bitmap where we store |
925 basic blocks that had non-trivial update on their loop_father.*/ | |
926 | |
927 void | |
928 unloop (struct loop *loop, bool *irred_invalidated, | |
929 bitmap loop_closed_ssa_invalidated) | |
0 | 930 { |
931 basic_block *body; | |
932 struct loop *ploop; | |
933 unsigned i, n; | |
934 basic_block latch = loop->latch; | |
935 bool dummy = false; | |
936 | |
937 if (loop_preheader_edge (loop)->flags & EDGE_IRREDUCIBLE_LOOP) | |
938 *irred_invalidated = true; | |
939 | |
940 /* This is relatively straightforward. The dominators are unchanged, as | |
941 loop header dominates loop latch, so the only thing we have to care of | |
942 is the placement of loops and basic blocks inside the loop tree. We | |
943 move them all to the loop->outer, and then let fix_bb_placements do | |
944 its work. */ | |
945 | |
946 body = get_loop_body (loop); | |
947 n = loop->num_nodes; | |
948 for (i = 0; i < n; i++) | |
949 if (body[i]->loop_father == loop) | |
950 { | |
951 remove_bb_from_loops (body[i]); | |
952 add_bb_to_loop (body[i], loop_outer (loop)); | |
953 } | |
111 | 954 free (body); |
0 | 955 |
956 while (loop->inner) | |
957 { | |
958 ploop = loop->inner; | |
959 flow_loop_tree_node_remove (ploop); | |
960 flow_loop_tree_node_add (loop_outer (loop), ploop); | |
961 } | |
962 | |
963 /* Remove the loop and free its data. */ | |
964 delete_loop (loop); | |
965 | |
966 remove_edge (single_succ_edge (latch)); | |
967 | |
968 /* We do not pass IRRED_INVALIDATED to fix_bb_placements here, as even if | |
969 there is an irreducible region inside the cancelled loop, the flags will | |
970 be still correct. */ | |
111 | 971 fix_bb_placements (latch, &dummy, loop_closed_ssa_invalidated); |
0 | 972 } |
973 | |
974 /* Fix placement of superloops of LOOP inside loop tree, i.e. ensure that | |
975 condition stated in description of fix_loop_placement holds for them. | |
976 It is used in case when we removed some edges coming out of LOOP, which | |
977 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
|
978 |
0 | 979 IRRED_INVALIDATED is set to true if a change in the loop structures might |
980 invalidate the information about irreducible regions. */ | |
981 | |
982 static void | |
983 fix_loop_placements (struct loop *loop, bool *irred_invalidated) | |
984 { | |
985 struct loop *outer; | |
986 | |
987 while (loop_outer (loop)) | |
988 { | |
989 outer = loop_outer (loop); | |
111 | 990 if (!fix_loop_placement (loop, irred_invalidated)) |
0 | 991 break; |
992 | |
993 /* Changing the placement of a loop in the loop tree may alter the | |
994 validity of condition 2) of the description of fix_bb_placement | |
995 for its preheader, because the successor is the header and belongs | |
996 to the loop. So call fix_bb_placements to fix up the placement | |
997 of the preheader and (possibly) of its predecessors. */ | |
998 fix_bb_placements (loop_preheader_edge (loop)->src, | |
111 | 999 irred_invalidated, NULL); |
0 | 1000 loop = outer; |
1001 } | |
1002 } | |
1003 | |
111 | 1004 /* Duplicate loop bounds and other information we store about |
1005 the loop into its duplicate. */ | |
1006 | |
1007 void | |
1008 copy_loop_info (struct loop *loop, struct loop *target) | |
1009 { | |
1010 gcc_checking_assert (!target->any_upper_bound && !target->any_estimate); | |
1011 target->any_upper_bound = loop->any_upper_bound; | |
1012 target->nb_iterations_upper_bound = loop->nb_iterations_upper_bound; | |
1013 target->any_likely_upper_bound = loop->any_likely_upper_bound; | |
1014 target->nb_iterations_likely_upper_bound | |
1015 = loop->nb_iterations_likely_upper_bound; | |
1016 target->any_estimate = loop->any_estimate; | |
1017 target->nb_iterations_estimate = loop->nb_iterations_estimate; | |
1018 target->estimate_state = loop->estimate_state; | |
1019 target->constraints = loop->constraints; | |
1020 target->warned_aggressive_loop_optimizations | |
1021 |= loop->warned_aggressive_loop_optimizations; | |
1022 target->in_oacc_kernels_region = loop->in_oacc_kernels_region; | |
1023 } | |
1024 | |
0 | 1025 /* Copies copy of LOOP as subloop of TARGET loop, placing newly |
1026 created loop into loops structure. */ | |
1027 struct loop * | |
1028 duplicate_loop (struct loop *loop, struct loop *target) | |
1029 { | |
1030 struct loop *cloop; | |
1031 cloop = alloc_loop (); | |
111 | 1032 place_new_loop (cfun, cloop); |
1033 | |
1034 copy_loop_info (loop, cloop); | |
0 | 1035 |
1036 /* Mark the new loop as copy of LOOP. */ | |
1037 set_loop_copy (loop, cloop); | |
1038 | |
1039 /* Add it to target. */ | |
1040 flow_loop_tree_node_add (target, cloop); | |
1041 | |
1042 return cloop; | |
1043 } | |
1044 | |
1045 /* Copies structure of subloops of LOOP into TARGET loop, placing | |
1046 newly created loops into loop tree. */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1047 void |
0 | 1048 duplicate_subloops (struct loop *loop, struct loop *target) |
1049 { | |
1050 struct loop *aloop, *cloop; | |
1051 | |
1052 for (aloop = loop->inner; aloop; aloop = aloop->next) | |
1053 { | |
1054 cloop = duplicate_loop (aloop, target); | |
1055 duplicate_subloops (aloop, cloop); | |
1056 } | |
1057 } | |
1058 | |
1059 /* Copies structure of subloops of N loops, stored in array COPIED_LOOPS, | |
1060 into TARGET loop, placing newly created loops into loop tree. */ | |
1061 static void | |
1062 copy_loops_to (struct loop **copied_loops, int n, struct loop *target) | |
1063 { | |
1064 struct loop *aloop; | |
1065 int i; | |
1066 | |
1067 for (i = 0; i < n; i++) | |
1068 { | |
1069 aloop = duplicate_loop (copied_loops[i], target); | |
1070 duplicate_subloops (copied_loops[i], aloop); | |
1071 } | |
1072 } | |
1073 | |
1074 /* Redirects edge E to basic block DEST. */ | |
1075 static void | |
1076 loop_redirect_edge (edge e, basic_block dest) | |
1077 { | |
1078 if (e->dest == dest) | |
1079 return; | |
1080 | |
1081 redirect_edge_and_branch_force (e, dest); | |
1082 } | |
1083 | |
1084 /* Check whether LOOP's body can be duplicated. */ | |
1085 bool | |
1086 can_duplicate_loop_p (const struct loop *loop) | |
1087 { | |
1088 int ret; | |
1089 basic_block *bbs = get_loop_body (loop); | |
1090 | |
1091 ret = can_copy_bbs_p (bbs, loop->num_nodes); | |
1092 free (bbs); | |
1093 | |
1094 return ret; | |
1095 } | |
1096 | |
1097 /* Duplicates body of LOOP to given edge E NDUPL times. Takes care of updating | |
1098 loop structure and dominators. E's destination must be LOOP header for | |
1099 this to work, i.e. it must be entry or latch edge of this loop; these are | |
1100 unique, as the loops must have preheaders for this function to work | |
1101 correctly (in case E is latch, the function unrolls the loop, if E is entry | |
1102 edge, it peels the loop). Store edges created by copying ORIG edge from | |
1103 copies corresponding to set bits in WONT_EXIT bitmap (bit 0 corresponds to | |
1104 original LOOP body, the other copies are numbered in order given by control | |
1105 flow through them) into TO_REMOVE array. Returns false if duplication is | |
1106 impossible. */ | |
1107 | |
1108 bool | |
1109 duplicate_loop_to_header_edge (struct loop *loop, edge e, | |
1110 unsigned int ndupl, sbitmap wont_exit, | |
111 | 1111 edge orig, vec<edge> *to_remove, |
0 | 1112 int flags) |
1113 { | |
1114 struct loop *target, *aloop; | |
1115 struct loop **orig_loops; | |
1116 unsigned n_orig_loops; | |
1117 basic_block header = loop->header, latch = loop->latch; | |
1118 basic_block *new_bbs, *bbs, *first_active; | |
1119 basic_block new_bb, bb, first_active_latch = NULL; | |
1120 edge ae, latch_edge; | |
1121 edge spec_edges[2], new_spec_edges[2]; | |
1122 #define SE_LATCH 0 | |
1123 #define SE_ORIG 1 | |
1124 unsigned i, j, n; | |
1125 int is_latch = (latch == e->src); | |
1126 int scale_act = 0, *scale_step = NULL, scale_main = 0; | |
1127 int scale_after_exit = 0; | |
1128 int p, freq_in, freq_le, freq_out_orig; | |
1129 int prob_pass_thru, prob_pass_wont_exit, prob_pass_main; | |
1130 int add_irreducible_flag; | |
1131 basic_block place_after; | |
1132 bitmap bbs_to_scale = NULL; | |
1133 bitmap_iterator bi; | |
1134 | |
1135 gcc_assert (e->dest == loop->header); | |
1136 gcc_assert (ndupl > 0); | |
1137 | |
1138 if (orig) | |
1139 { | |
1140 /* Orig must be edge out of the loop. */ | |
1141 gcc_assert (flow_bb_inside_loop_p (loop, orig->src)); | |
1142 gcc_assert (!flow_bb_inside_loop_p (loop, orig->dest)); | |
1143 } | |
1144 | |
1145 n = loop->num_nodes; | |
1146 bbs = get_loop_body_in_dom_order (loop); | |
1147 gcc_assert (bbs[0] == loop->header); | |
1148 gcc_assert (bbs[n - 1] == loop->latch); | |
1149 | |
1150 /* Check whether duplication is possible. */ | |
1151 if (!can_copy_bbs_p (bbs, loop->num_nodes)) | |
1152 { | |
1153 free (bbs); | |
1154 return false; | |
1155 } | |
1156 new_bbs = XNEWVEC (basic_block, loop->num_nodes); | |
1157 | |
1158 /* In case we are doing loop peeling and the loop is in the middle of | |
1159 irreducible region, the peeled copies will be inside it too. */ | |
1160 add_irreducible_flag = e->flags & EDGE_IRREDUCIBLE_LOOP; | |
1161 gcc_assert (!is_latch || !add_irreducible_flag); | |
1162 | |
1163 /* Find edge from latch. */ | |
1164 latch_edge = loop_latch_edge (loop); | |
1165 | |
1166 if (flags & DLTHE_FLAG_UPDATE_FREQ) | |
1167 { | |
1168 /* Calculate coefficients by that we have to scale frequencies | |
1169 of duplicated loop bodies. */ | |
1170 freq_in = header->frequency; | |
1171 freq_le = EDGE_FREQUENCY (latch_edge); | |
1172 if (freq_in == 0) | |
1173 freq_in = 1; | |
1174 if (freq_in < freq_le) | |
1175 freq_in = freq_le; | |
1176 freq_out_orig = orig ? EDGE_FREQUENCY (orig) : freq_in - freq_le; | |
1177 if (freq_out_orig > freq_in - freq_le) | |
1178 freq_out_orig = freq_in - freq_le; | |
1179 prob_pass_thru = RDIV (REG_BR_PROB_BASE * freq_le, freq_in); | |
1180 prob_pass_wont_exit = | |
1181 RDIV (REG_BR_PROB_BASE * (freq_le + freq_out_orig), freq_in); | |
1182 | |
111 | 1183 if (orig && orig->probability.initialized_p () |
1184 && !(orig->probability == profile_probability::always ())) | |
0 | 1185 { |
1186 /* The blocks that are dominated by a removed exit edge ORIG have | |
1187 frequencies scaled by this. */ | |
111 | 1188 if (orig->probability.initialized_p ()) |
1189 scale_after_exit | |
1190 = GCOV_COMPUTE_SCALE (REG_BR_PROB_BASE, | |
1191 REG_BR_PROB_BASE | |
1192 - orig->probability.to_reg_br_prob_base ()); | |
1193 else | |
1194 scale_after_exit = REG_BR_PROB_BASE; | |
0 | 1195 bbs_to_scale = BITMAP_ALLOC (NULL); |
1196 for (i = 0; i < n; i++) | |
1197 { | |
1198 if (bbs[i] != orig->src | |
1199 && dominated_by_p (CDI_DOMINATORS, bbs[i], orig->src)) | |
1200 bitmap_set_bit (bbs_to_scale, i); | |
1201 } | |
1202 } | |
1203 | |
1204 scale_step = XNEWVEC (int, ndupl); | |
1205 | |
1206 for (i = 1; i <= ndupl; i++) | |
111 | 1207 scale_step[i - 1] = bitmap_bit_p (wont_exit, i) |
0 | 1208 ? prob_pass_wont_exit |
1209 : prob_pass_thru; | |
1210 | |
1211 /* Complete peeling is special as the probability of exit in last | |
1212 copy becomes 1. */ | |
1213 if (flags & DLTHE_FLAG_COMPLETTE_PEEL) | |
1214 { | |
1215 int wanted_freq = EDGE_FREQUENCY (e); | |
1216 | |
1217 if (wanted_freq > freq_in) | |
1218 wanted_freq = freq_in; | |
1219 | |
1220 gcc_assert (!is_latch); | |
1221 /* First copy has frequency of incoming edge. Each subsequent | |
1222 frequency should be reduced by prob_pass_wont_exit. Caller | |
1223 should've managed the flags so all except for original loop | |
1224 has won't exist set. */ | |
111 | 1225 scale_act = GCOV_COMPUTE_SCALE (wanted_freq, freq_in); |
0 | 1226 /* Now simulate the duplication adjustments and compute header |
1227 frequency of the last copy. */ | |
1228 for (i = 0; i < ndupl; i++) | |
111 | 1229 wanted_freq = combine_probabilities (wanted_freq, scale_step[i]); |
1230 scale_main = GCOV_COMPUTE_SCALE (wanted_freq, freq_in); | |
0 | 1231 } |
1232 else if (is_latch) | |
1233 { | |
111 | 1234 prob_pass_main = bitmap_bit_p (wont_exit, 0) |
0 | 1235 ? prob_pass_wont_exit |
1236 : prob_pass_thru; | |
1237 p = prob_pass_main; | |
1238 scale_main = REG_BR_PROB_BASE; | |
1239 for (i = 0; i < ndupl; i++) | |
1240 { | |
1241 scale_main += p; | |
111 | 1242 p = combine_probabilities (p, scale_step[i]); |
0 | 1243 } |
111 | 1244 scale_main = GCOV_COMPUTE_SCALE (REG_BR_PROB_BASE, scale_main); |
1245 scale_act = combine_probabilities (scale_main, prob_pass_main); | |
0 | 1246 } |
1247 else | |
1248 { | |
111 | 1249 int preheader_freq = EDGE_FREQUENCY (e); |
0 | 1250 scale_main = REG_BR_PROB_BASE; |
1251 for (i = 0; i < ndupl; i++) | |
111 | 1252 scale_main = combine_probabilities (scale_main, scale_step[i]); |
1253 if (preheader_freq > freq_in) | |
1254 preheader_freq = freq_in; | |
1255 scale_act = GCOV_COMPUTE_SCALE (preheader_freq, freq_in); | |
0 | 1256 } |
1257 for (i = 0; i < ndupl; i++) | |
1258 gcc_assert (scale_step[i] >= 0 && scale_step[i] <= REG_BR_PROB_BASE); | |
1259 gcc_assert (scale_main >= 0 && scale_main <= REG_BR_PROB_BASE | |
1260 && scale_act >= 0 && scale_act <= REG_BR_PROB_BASE); | |
1261 } | |
1262 | |
1263 /* Loop the new bbs will belong to. */ | |
1264 target = e->src->loop_father; | |
1265 | |
1266 /* Original loops. */ | |
1267 n_orig_loops = 0; | |
1268 for (aloop = loop->inner; aloop; aloop = aloop->next) | |
1269 n_orig_loops++; | |
111 | 1270 orig_loops = XNEWVEC (struct loop *, n_orig_loops); |
0 | 1271 for (aloop = loop->inner, i = 0; aloop; aloop = aloop->next, i++) |
1272 orig_loops[i] = aloop; | |
1273 | |
1274 set_loop_copy (loop, target); | |
1275 | |
1276 first_active = XNEWVEC (basic_block, n); | |
1277 if (is_latch) | |
1278 { | |
1279 memcpy (first_active, bbs, n * sizeof (basic_block)); | |
1280 first_active_latch = latch; | |
1281 } | |
1282 | |
1283 spec_edges[SE_ORIG] = orig; | |
1284 spec_edges[SE_LATCH] = latch_edge; | |
1285 | |
1286 place_after = e->src; | |
1287 for (j = 0; j < ndupl; j++) | |
1288 { | |
1289 /* Copy loops. */ | |
1290 copy_loops_to (orig_loops, n_orig_loops, target); | |
1291 | |
1292 /* Copy bbs. */ | |
1293 copy_bbs (bbs, n, new_bbs, spec_edges, 2, new_spec_edges, loop, | |
111 | 1294 place_after, true); |
0 | 1295 place_after = new_spec_edges[SE_LATCH]->src; |
1296 | |
1297 if (flags & DLTHE_RECORD_COPY_NUMBER) | |
1298 for (i = 0; i < n; i++) | |
1299 { | |
1300 gcc_assert (!new_bbs[i]->aux); | |
1301 new_bbs[i]->aux = (void *)(size_t)(j + 1); | |
1302 } | |
1303 | |
1304 /* Note whether the blocks and edges belong to an irreducible loop. */ | |
1305 if (add_irreducible_flag) | |
1306 { | |
1307 for (i = 0; i < n; i++) | |
1308 new_bbs[i]->flags |= BB_DUPLICATED; | |
1309 for (i = 0; i < n; i++) | |
1310 { | |
1311 edge_iterator ei; | |
1312 new_bb = new_bbs[i]; | |
1313 if (new_bb->loop_father == target) | |
1314 new_bb->flags |= BB_IRREDUCIBLE_LOOP; | |
1315 | |
1316 FOR_EACH_EDGE (ae, ei, new_bb->succs) | |
1317 if ((ae->dest->flags & BB_DUPLICATED) | |
1318 && (ae->src->loop_father == target | |
1319 || ae->dest->loop_father == target)) | |
1320 ae->flags |= EDGE_IRREDUCIBLE_LOOP; | |
1321 } | |
1322 for (i = 0; i < n; i++) | |
1323 new_bbs[i]->flags &= ~BB_DUPLICATED; | |
1324 } | |
1325 | |
1326 /* Redirect the special edges. */ | |
1327 if (is_latch) | |
1328 { | |
1329 redirect_edge_and_branch_force (latch_edge, new_bbs[0]); | |
1330 redirect_edge_and_branch_force (new_spec_edges[SE_LATCH], | |
1331 loop->header); | |
1332 set_immediate_dominator (CDI_DOMINATORS, new_bbs[0], latch); | |
1333 latch = loop->latch = new_bbs[n - 1]; | |
1334 e = latch_edge = new_spec_edges[SE_LATCH]; | |
1335 } | |
1336 else | |
1337 { | |
1338 redirect_edge_and_branch_force (new_spec_edges[SE_LATCH], | |
1339 loop->header); | |
1340 redirect_edge_and_branch_force (e, new_bbs[0]); | |
1341 set_immediate_dominator (CDI_DOMINATORS, new_bbs[0], e->src); | |
1342 e = new_spec_edges[SE_LATCH]; | |
1343 } | |
1344 | |
1345 /* Record exit edge in this copy. */ | |
111 | 1346 if (orig && bitmap_bit_p (wont_exit, j + 1)) |
0 | 1347 { |
1348 if (to_remove) | |
111 | 1349 to_remove->safe_push (new_spec_edges[SE_ORIG]); |
1350 force_edge_cold (new_spec_edges[SE_ORIG], true); | |
0 | 1351 |
1352 /* Scale the frequencies of the blocks dominated by the exit. */ | |
1353 if (bbs_to_scale) | |
1354 { | |
1355 EXECUTE_IF_SET_IN_BITMAP (bbs_to_scale, 0, i, bi) | |
1356 { | |
1357 scale_bbs_frequencies_int (new_bbs + i, 1, scale_after_exit, | |
1358 REG_BR_PROB_BASE); | |
1359 } | |
1360 } | |
1361 } | |
1362 | |
1363 /* Record the first copy in the control flow order if it is not | |
1364 the original loop (i.e. in case of peeling). */ | |
1365 if (!first_active_latch) | |
1366 { | |
1367 memcpy (first_active, new_bbs, n * sizeof (basic_block)); | |
1368 first_active_latch = new_bbs[n - 1]; | |
1369 } | |
1370 | |
1371 /* Set counts and frequencies. */ | |
1372 if (flags & DLTHE_FLAG_UPDATE_FREQ) | |
1373 { | |
1374 scale_bbs_frequencies_int (new_bbs, n, scale_act, REG_BR_PROB_BASE); | |
111 | 1375 scale_act = combine_probabilities (scale_act, scale_step[j]); |
0 | 1376 } |
1377 } | |
1378 free (new_bbs); | |
1379 free (orig_loops); | |
1380 | |
1381 /* Record the exit edge in the original loop body, and update the frequencies. */ | |
111 | 1382 if (orig && bitmap_bit_p (wont_exit, 0)) |
0 | 1383 { |
1384 if (to_remove) | |
111 | 1385 to_remove->safe_push (orig); |
1386 force_edge_cold (orig, true); | |
0 | 1387 |
1388 /* Scale the frequencies of the blocks dominated by the exit. */ | |
1389 if (bbs_to_scale) | |
1390 { | |
1391 EXECUTE_IF_SET_IN_BITMAP (bbs_to_scale, 0, i, bi) | |
1392 { | |
1393 scale_bbs_frequencies_int (bbs + i, 1, scale_after_exit, | |
1394 REG_BR_PROB_BASE); | |
1395 } | |
1396 } | |
1397 } | |
1398 | |
1399 /* Update the original loop. */ | |
1400 if (!is_latch) | |
1401 set_immediate_dominator (CDI_DOMINATORS, e->dest, e->src); | |
1402 if (flags & DLTHE_FLAG_UPDATE_FREQ) | |
1403 { | |
1404 scale_bbs_frequencies_int (bbs, n, scale_main, REG_BR_PROB_BASE); | |
1405 free (scale_step); | |
1406 } | |
1407 | |
1408 /* Update dominators of outer blocks if affected. */ | |
1409 for (i = 0; i < n; i++) | |
1410 { | |
1411 basic_block dominated, dom_bb; | |
111 | 1412 vec<basic_block> dom_bbs; |
0 | 1413 unsigned j; |
1414 | |
1415 bb = bbs[i]; | |
1416 bb->aux = 0; | |
1417 | |
1418 dom_bbs = get_dominated_by (CDI_DOMINATORS, bb); | |
111 | 1419 FOR_EACH_VEC_ELT (dom_bbs, j, dominated) |
0 | 1420 { |
1421 if (flow_bb_inside_loop_p (loop, dominated)) | |
1422 continue; | |
1423 dom_bb = nearest_common_dominator ( | |
1424 CDI_DOMINATORS, first_active[i], first_active_latch); | |
1425 set_immediate_dominator (CDI_DOMINATORS, dominated, dom_bb); | |
1426 } | |
111 | 1427 dom_bbs.release (); |
0 | 1428 } |
1429 free (first_active); | |
1430 | |
1431 free (bbs); | |
1432 BITMAP_FREE (bbs_to_scale); | |
1433 | |
1434 return true; | |
1435 } | |
1436 | |
1437 /* A callback for make_forwarder block, to redirect all edges except for | |
1438 MFB_KJ_EDGE to the entry part. E is the edge for that we should decide | |
1439 whether to redirect it. */ | |
1440 | |
1441 edge mfb_kj_edge; | |
1442 bool | |
1443 mfb_keep_just (edge e) | |
1444 { | |
1445 return e != mfb_kj_edge; | |
1446 } | |
1447 | |
1448 /* True when a candidate preheader BLOCK has predecessors from LOOP. */ | |
1449 | |
1450 static bool | |
1451 has_preds_from_loop (basic_block block, struct loop *loop) | |
1452 { | |
1453 edge e; | |
1454 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
|
1455 |
0 | 1456 FOR_EACH_EDGE (e, ei, block->preds) |
1457 if (e->src->loop_father == loop) | |
1458 return true; | |
1459 return false; | |
1460 } | |
1461 | |
1462 /* Creates a pre-header for a LOOP. Returns newly created block. Unless | |
1463 CP_SIMPLE_PREHEADERS is set in FLAGS, we only force LOOP to have single | |
1464 entry; otherwise we also force preheader block to have only one successor. | |
1465 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
|
1466 to be a fallthru predecessor to the loop header and to have only |
0 | 1467 predecessors from outside of the loop. |
1468 The function also updates dominators. */ | |
1469 | |
1470 basic_block | |
1471 create_preheader (struct loop *loop, int flags) | |
1472 { | |
111 | 1473 edge e; |
0 | 1474 basic_block dummy; |
1475 int nentry = 0; | |
1476 bool irred = false; | |
1477 bool latch_edge_was_fallthru; | |
1478 edge one_succ_pred = NULL, single_entry = NULL; | |
1479 edge_iterator ei; | |
1480 | |
1481 FOR_EACH_EDGE (e, ei, loop->header->preds) | |
1482 { | |
1483 if (e->src == loop->latch) | |
1484 continue; | |
1485 irred |= (e->flags & EDGE_IRREDUCIBLE_LOOP) != 0; | |
1486 nentry++; | |
1487 single_entry = e; | |
1488 if (single_succ_p (e->src)) | |
1489 one_succ_pred = e; | |
1490 } | |
1491 gcc_assert (nentry); | |
1492 if (nentry == 1) | |
1493 { | |
1494 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
|
1495 |
0 | 1496 /* We do not allow entry block to be the loop preheader, since we |
1497 cannot emit code there. */ | |
111 | 1498 if (single_entry->src == ENTRY_BLOCK_PTR_FOR_FN (cfun)) |
0 | 1499 need_forwarder_block = true; |
1500 else | |
1501 { | |
1502 /* If we want simple preheaders, also force the preheader to have | |
1503 just a single successor. */ | |
1504 if ((flags & CP_SIMPLE_PREHEADERS) | |
1505 && !single_succ_p (single_entry->src)) | |
1506 need_forwarder_block = true; | |
1507 /* If we want fallthru preheaders, also create forwarder block when | |
1508 preheader ends with a jump or has predecessors from loop. */ | |
1509 else if ((flags & CP_FALLTHRU_PREHEADERS) | |
1510 && (JUMP_P (BB_END (single_entry->src)) | |
1511 || has_preds_from_loop (single_entry->src, loop))) | |
1512 need_forwarder_block = true; | |
1513 } | |
1514 if (! need_forwarder_block) | |
1515 return NULL; | |
1516 } | |
1517 | |
1518 mfb_kj_edge = loop_latch_edge (loop); | |
1519 latch_edge_was_fallthru = (mfb_kj_edge->flags & EDGE_FALLTHRU) != 0; | |
111 | 1520 if (nentry == 1) |
1521 dummy = split_edge (single_entry); | |
1522 else | |
1523 { | |
1524 edge fallthru = make_forwarder_block (loop->header, mfb_keep_just, NULL); | |
1525 dummy = fallthru->src; | |
1526 loop->header = fallthru->dest; | |
1527 } | |
0 | 1528 |
1529 /* Try to be clever in placing the newly created preheader. The idea is to | |
1530 avoid breaking any "fallthruness" relationship between blocks. | |
1531 | |
1532 The preheader was created just before the header and all incoming edges | |
1533 to the header were redirected to the preheader, except the latch edge. | |
1534 So the only problematic case is when this latch edge was a fallthru | |
1535 edge: it is not anymore after the preheader creation so we have broken | |
1536 the fallthruness. We're therefore going to look for a better place. */ | |
1537 if (latch_edge_was_fallthru) | |
1538 { | |
1539 if (one_succ_pred) | |
1540 e = one_succ_pred; | |
1541 else | |
1542 e = EDGE_PRED (dummy, 0); | |
1543 | |
1544 move_block_after (dummy, e->src); | |
1545 } | |
1546 | |
1547 if (irred) | |
1548 { | |
1549 dummy->flags |= BB_IRREDUCIBLE_LOOP; | |
1550 single_succ_edge (dummy)->flags |= EDGE_IRREDUCIBLE_LOOP; | |
1551 } | |
1552 | |
1553 if (dump_file) | |
1554 fprintf (dump_file, "Created preheader block for loop %i\n", | |
1555 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
|
1556 |
0 | 1557 if (flags & CP_FALLTHRU_PREHEADERS) |
1558 gcc_assert ((single_succ_edge (dummy)->flags & EDGE_FALLTHRU) | |
1559 && !JUMP_P (BB_END (dummy))); | |
1560 | |
1561 return dummy; | |
1562 } | |
1563 | |
1564 /* Create preheaders for each loop; for meaning of FLAGS see create_preheader. */ | |
1565 | |
1566 void | |
1567 create_preheaders (int flags) | |
1568 { | |
1569 struct loop *loop; | |
1570 | |
1571 if (!current_loops) | |
1572 return; | |
1573 | |
111 | 1574 FOR_EACH_LOOP (loop, 0) |
0 | 1575 create_preheader (loop, flags); |
1576 loops_state_set (LOOPS_HAVE_PREHEADERS); | |
1577 } | |
1578 | |
1579 /* Forces all loop latches to have only single successor. */ | |
1580 | |
1581 void | |
1582 force_single_succ_latches (void) | |
1583 { | |
1584 struct loop *loop; | |
1585 edge e; | |
1586 | |
111 | 1587 FOR_EACH_LOOP (loop, 0) |
0 | 1588 { |
1589 if (loop->latch != loop->header && single_succ_p (loop->latch)) | |
1590 continue; | |
1591 | |
1592 e = find_edge (loop->latch, loop->header); | |
111 | 1593 gcc_checking_assert (e != NULL); |
0 | 1594 |
1595 split_edge (e); | |
1596 } | |
1597 loops_state_set (LOOPS_HAVE_SIMPLE_LATCHES); | |
1598 } | |
1599 | |
1600 /* This function is called from loop_version. It splits the entry edge | |
1601 of the loop we want to version, adds the versioning condition, and | |
1602 adjust the edges to the two versions of the loop appropriately. | |
1603 e is an incoming edge. Returns the basic block containing the | |
1604 condition. | |
1605 | |
1606 --- edge e ---- > [second_head] | |
1607 | |
1608 Split it and insert new conditional expression and adjust edges. | |
1609 | |
1610 --- edge e ---> [cond expr] ---> [first_head] | |
1611 | | |
1612 +---------> [second_head] | |
1613 | |
111 | 1614 THEN_PROB is the probability of then branch of the condition. |
1615 ELSE_PROB is the probability of else branch. Note that they may be both | |
1616 REG_BR_PROB_BASE when condition is IFN_LOOP_VECTORIZED or | |
1617 IFN_LOOP_DIST_ALIAS. */ | |
0 | 1618 |
1619 static basic_block | |
1620 lv_adjust_loop_entry_edge (basic_block first_head, basic_block second_head, | |
111 | 1621 edge e, void *cond_expr, |
1622 profile_probability then_prob, | |
1623 profile_probability else_prob) | |
0 | 1624 { |
1625 basic_block new_head = NULL; | |
1626 edge e1; | |
1627 | |
1628 gcc_assert (e->dest == second_head); | |
1629 | |
1630 /* Split edge 'e'. This will create a new basic block, where we can | |
1631 insert conditional expr. */ | |
1632 new_head = split_edge (e); | |
1633 | |
1634 lv_add_condition_to_bb (first_head, second_head, new_head, | |
1635 cond_expr); | |
1636 | |
1637 /* Don't set EDGE_TRUE_VALUE in RTL mode, as it's invalid there. */ | |
1638 e = single_succ_edge (new_head); | |
1639 e1 = make_edge (new_head, first_head, | |
1640 current_ir_type () == IR_GIMPLE ? EDGE_TRUE_VALUE : 0); | |
1641 e1->probability = then_prob; | |
111 | 1642 e->probability = else_prob; |
0 | 1643 |
1644 set_immediate_dominator (CDI_DOMINATORS, first_head, new_head); | |
1645 set_immediate_dominator (CDI_DOMINATORS, second_head, new_head); | |
1646 | |
1647 /* Adjust loop header phi nodes. */ | |
1648 lv_adjust_loop_header_phi (first_head, second_head, new_head, e1); | |
1649 | |
1650 return new_head; | |
1651 } | |
1652 | |
1653 /* Main entry point for Loop Versioning transformation. | |
1654 | |
1655 This transformation given a condition and a loop, creates | |
1656 -if (condition) { loop_copy1 } else { loop_copy2 }, | |
1657 where loop_copy1 is the loop transformed in one way, and loop_copy2 | |
111 | 1658 is the loop transformed in another way (or unchanged). COND_EXPR |
0 | 1659 may be a run time test for things that were not resolved by static |
1660 analysis (overlapping ranges (anti-aliasing), alignment, etc.). | |
1661 | |
111 | 1662 If non-NULL, CONDITION_BB is set to the basic block containing the |
1663 condition. | |
1664 | |
0 | 1665 THEN_PROB is the probability of the then edge of the if. THEN_SCALE |
1666 is the ratio by that the frequencies in the original loop should | |
1667 be scaled. ELSE_SCALE is the ratio by that the frequencies in the | |
1668 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
|
1669 |
0 | 1670 If PLACE_AFTER is true, we place the new loop after LOOP in the |
1671 instruction stream, otherwise it is placed before LOOP. */ | |
1672 | |
1673 struct loop * | |
1674 loop_version (struct loop *loop, | |
1675 void *cond_expr, basic_block *condition_bb, | |
111 | 1676 profile_probability then_prob, profile_probability else_prob, |
1677 profile_probability then_scale, profile_probability else_scale, | |
0 | 1678 bool place_after) |
1679 { | |
1680 basic_block first_head, second_head; | |
1681 edge entry, latch_edge, true_edge, false_edge; | |
1682 int irred_flag; | |
1683 struct loop *nloop; | |
1684 basic_block cond_bb; | |
1685 | |
1686 /* Record entry and latch edges for the loop */ | |
1687 entry = loop_preheader_edge (loop); | |
1688 irred_flag = entry->flags & EDGE_IRREDUCIBLE_LOOP; | |
1689 entry->flags &= ~EDGE_IRREDUCIBLE_LOOP; | |
1690 | |
1691 /* Note down head of loop as first_head. */ | |
1692 first_head = entry->dest; | |
1693 | |
1694 /* Duplicate loop. */ | |
1695 if (!cfg_hook_duplicate_loop_to_header_edge (loop, entry, 1, | |
1696 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
|
1697 { |
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
|
1698 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
|
1699 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
|
1700 } |
0 | 1701 |
1702 /* After duplication entry edge now points to new loop head block. | |
1703 Note down new head as second_head. */ | |
1704 second_head = entry->dest; | |
1705 | |
1706 /* Split loop entry edge and insert new block with cond expr. */ | |
1707 cond_bb = lv_adjust_loop_entry_edge (first_head, second_head, | |
111 | 1708 entry, cond_expr, then_prob, else_prob); |
0 | 1709 if (condition_bb) |
1710 *condition_bb = cond_bb; | |
1711 | |
1712 if (!cond_bb) | |
1713 { | |
1714 entry->flags |= irred_flag; | |
1715 return NULL; | |
1716 } | |
1717 | |
1718 latch_edge = single_succ_edge (get_bb_copy (loop->latch)); | |
1719 | |
1720 extract_cond_bb_edges (cond_bb, &true_edge, &false_edge); | |
1721 nloop = loopify (latch_edge, | |
1722 single_pred_edge (get_bb_copy (loop->header)), | |
1723 cond_bb, true_edge, false_edge, | |
1724 false /* Do not redirect all edges. */, | |
1725 then_scale, else_scale); | |
1726 | |
111 | 1727 copy_loop_info (loop, nloop); |
1728 | |
0 | 1729 /* loopify redirected latch_edge. Update its PENDING_STMTS. */ |
1730 lv_flush_pending_stmts (latch_edge); | |
1731 | |
1732 /* loopify redirected condition_bb's succ edge. Update its PENDING_STMTS. */ | |
1733 extract_cond_bb_edges (cond_bb, &true_edge, &false_edge); | |
1734 lv_flush_pending_stmts (false_edge); | |
1735 /* Adjust irreducible flag. */ | |
1736 if (irred_flag) | |
1737 { | |
1738 cond_bb->flags |= BB_IRREDUCIBLE_LOOP; | |
1739 loop_preheader_edge (loop)->flags |= EDGE_IRREDUCIBLE_LOOP; | |
1740 loop_preheader_edge (nloop)->flags |= EDGE_IRREDUCIBLE_LOOP; | |
1741 single_pred_edge (cond_bb)->flags |= EDGE_IRREDUCIBLE_LOOP; | |
1742 } | |
1743 | |
1744 if (place_after) | |
1745 { | |
1746 basic_block *bbs = get_loop_body_in_dom_order (nloop), after; | |
1747 unsigned i; | |
1748 | |
1749 after = loop->latch; | |
1750 | |
1751 for (i = 0; i < nloop->num_nodes; i++) | |
1752 { | |
1753 move_block_after (bbs[i], after); | |
1754 after = bbs[i]; | |
1755 } | |
1756 free (bbs); | |
1757 } | |
1758 | |
1759 /* At this point condition_bb is loop preheader with two successors, | |
1760 first_head and second_head. Make sure that loop preheader has only | |
1761 one successor. */ | |
1762 split_edge (loop_preheader_edge (loop)); | |
1763 split_edge (loop_preheader_edge (nloop)); | |
1764 | |
1765 return nloop; | |
1766 } |