]>
Commit | Line | Data |
---|---|---|
a8046f60 | 1 | /* Thread edges through blocks and update the control flow and SSA graphs. |
aad93da1 | 2 | Copyright (C) 2004-2017 Free Software Foundation, Inc. |
a8046f60 | 3 | |
4 | This file is part of GCC. | |
5 | ||
6 | GCC is free software; you can redistribute it and/or modify | |
7 | it under the terms of the GNU General Public License as published by | |
8c4c00c1 | 8 | the Free Software Foundation; either version 3, or (at your option) |
a8046f60 | 9 | any later version. |
10 | ||
11 | GCC is distributed in the hope that it will be useful, | |
12 | but WITHOUT ANY WARRANTY; without even the implied warranty of | |
13 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
14 | GNU General Public License for more details. | |
15 | ||
16 | You should have received a copy of the GNU General Public License | |
8c4c00c1 | 17 | along with GCC; see the file COPYING3. If not see |
18 | <http://www.gnu.org/licenses/>. */ | |
a8046f60 | 19 | |
20 | #include "config.h" | |
21 | #include "system.h" | |
22 | #include "coretypes.h" | |
9ef16211 | 23 | #include "backend.h" |
a8046f60 | 24 | #include "tree.h" |
9ef16211 | 25 | #include "gimple.h" |
7c29e30e | 26 | #include "cfghooks.h" |
27 | #include "tree-pass.h" | |
9ef16211 | 28 | #include "ssa.h" |
b20a8bb4 | 29 | #include "fold-const.h" |
94ea8568 | 30 | #include "cfganal.h" |
dcf1a1ec | 31 | #include "gimple-iterator.h" |
69ee5dbb | 32 | #include "tree-ssa.h" |
0c5b289a | 33 | #include "tree-ssa-threadupdate.h" |
388d1fc1 | 34 | #include "cfgloop.h" |
a3724f9d | 35 | #include "dbgcnt.h" |
ab596744 | 36 | #include "tree-cfg.h" |
325162f2 | 37 | #include "tree-vectorizer.h" |
a8046f60 | 38 | |
39 | /* Given a block B, update the CFG and SSA graph to reflect redirecting | |
40 | one or more in-edges to B to instead reach the destination of an | |
41 | out-edge from B while preserving any side effects in B. | |
42 | ||
0c6d8c36 | 43 | i.e., given A->B and B->C, change A->B to be A->C yet still preserve the |
a8046f60 | 44 | side effects of executing B. |
45 | ||
46 | 1. Make a copy of B (including its outgoing edges and statements). Call | |
47 | the copy B'. Note B' has no incoming edges or PHIs at this time. | |
48 | ||
49 | 2. Remove the control statement at the end of B' and all outgoing edges | |
50 | except B'->C. | |
51 | ||
52 | 3. Add a new argument to each PHI in C with the same value as the existing | |
53 | argument associated with edge B->C. Associate the new PHI arguments | |
54 | with the edge B'->C. | |
55 | ||
56 | 4. For each PHI in B, find or create a PHI in B' with an identical | |
7a635e9c | 57 | PHI_RESULT. Add an argument to the PHI in B' which has the same |
a8046f60 | 58 | value as the PHI in B associated with the edge A->B. Associate |
59 | the new argument in the PHI in B' with the edge A->B. | |
60 | ||
61 | 5. Change the edge A->B to A->B'. | |
62 | ||
63 | 5a. This automatically deletes any PHI arguments associated with the | |
64 | edge A->B in B. | |
65 | ||
66 | 5b. This automatically associates each new argument added in step 4 | |
67 | with the edge A->B'. | |
68 | ||
69 | 6. Repeat for other incoming edges into B. | |
70 | ||
71 | 7. Put the duplicated resources in B and all the B' blocks into SSA form. | |
72 | ||
73 | Note that block duplication can be minimized by first collecting the | |
f0b5f617 | 74 | set of unique destination blocks that the incoming edges should |
255a8494 | 75 | be threaded to. |
76 | ||
afe75331 | 77 | We reduce the number of edges and statements we create by not copying all |
78 | the outgoing edges and the control statement in step #1. We instead create | |
79 | a template block without the outgoing edges and duplicate the template. | |
a8046f60 | 80 | |
afe75331 | 81 | Another case this code handles is threading through a "joiner" block. In |
82 | this case, we do not know the destination of the joiner block, but one | |
83 | of the outgoing edges from the joiner block leads to a threadable path. This | |
84 | case largely works as outlined above, except the duplicate of the joiner | |
85 | block still contains a full set of outgoing edges and its control statement. | |
86 | We just redirect one of its outgoing edges to our jump threading path. */ | |
778182c1 | 87 | |
88 | ||
89 | /* Steps #5 and #6 of the above algorithm are best implemented by walking | |
90 | all the incoming edges which thread to the same destination edge at | |
91 | the same time. That avoids lots of table lookups to get information | |
92 | for the destination edge. | |
93 | ||
94 | To realize that implementation we create a list of incoming edges | |
95 | which thread to the same outgoing edge. Thus to implement steps | |
96 | #5 and #6 we traverse our hash table of outgoing edge information. | |
97 | For each entry we walk the list of incoming edges which thread to | |
98 | the current outgoing edge. */ | |
99 | ||
100 | struct el | |
101 | { | |
102 | edge e; | |
103 | struct el *next; | |
104 | }; | |
a8046f60 | 105 | |
106 | /* Main data structure recording information regarding B's duplicate | |
107 | blocks. */ | |
108 | ||
778182c1 | 109 | /* We need to efficiently record the unique thread destinations of this |
110 | block and specific information associated with those destinations. We | |
111 | may have many incoming edges threaded to the same outgoing edge. This | |
c5d4a10b | 112 | can be naturally implemented with a hash table. */ |
778182c1 | 113 | |
298e7f9a | 114 | struct redirection_data : free_ptr_hash<redirection_data> |
a8046f60 | 115 | { |
11af02d8 | 116 | /* We support wiring up two block duplicates in a jump threading path. |
117 | ||
118 | One is a normal block copy where we remove the control statement | |
119 | and wire up its single remaining outgoing edge to the thread path. | |
120 | ||
121 | The other is a joiner block where we leave the control statement | |
1b83778e | 122 | in place, but wire one of the outgoing edges to a thread path. |
11af02d8 | 123 | |
124 | In theory we could have multiple block duplicates in a jump | |
125 | threading path, but I haven't tried that. | |
126 | ||
127 | The duplicate blocks appear in this array in the same order in | |
128 | which they appear in the jump thread path. */ | |
129 | basic_block dup_blocks[2]; | |
a8046f60 | 130 | |
5fe6149c | 131 | /* The jump threading path. */ |
132 | vec<jump_thread_edge *> *path; | |
778182c1 | 133 | |
5fe6149c | 134 | /* A list of incoming edges which we want to thread to the |
135 | same path. */ | |
778182c1 | 136 | struct el *incoming_edges; |
494bbaae | 137 | |
138 | /* hash_table support. */ | |
9969c043 | 139 | static inline hashval_t hash (const redirection_data *); |
140 | static inline int equal (const redirection_data *, const redirection_data *); | |
a8046f60 | 141 | }; |
142 | ||
b93ba654 | 143 | /* Dump a jump threading path, including annotations about each |
144 | edge in the path. */ | |
145 | ||
146 | static void | |
147 | dump_jump_thread_path (FILE *dump_file, vec<jump_thread_edge *> path, | |
148 | bool registering) | |
149 | { | |
150 | fprintf (dump_file, | |
ded1c768 | 151 | " %s%s jump thread: (%d, %d) incoming edge; ", |
b93ba654 | 152 | (registering ? "Registering" : "Cancelling"), |
ded1c768 | 153 | (path[0]->type == EDGE_FSM_THREAD ? " FSM": ""), |
b93ba654 | 154 | path[0]->e->src->index, path[0]->e->dest->index); |
155 | ||
156 | for (unsigned int i = 1; i < path.length (); i++) | |
157 | { | |
158 | /* We can get paths with a NULL edge when the final destination | |
159 | of a jump thread turns out to be a constant address. We dump | |
160 | those paths when debugging, so we have to be prepared for that | |
161 | possibility here. */ | |
162 | if (path[i]->e == NULL) | |
163 | continue; | |
164 | ||
165 | if (path[i]->type == EDGE_COPY_SRC_JOINER_BLOCK) | |
166 | fprintf (dump_file, " (%d, %d) joiner; ", | |
167 | path[i]->e->src->index, path[i]->e->dest->index); | |
168 | if (path[i]->type == EDGE_COPY_SRC_BLOCK) | |
169 | fprintf (dump_file, " (%d, %d) normal;", | |
170 | path[i]->e->src->index, path[i]->e->dest->index); | |
171 | if (path[i]->type == EDGE_NO_COPY_SRC_BLOCK) | |
172 | fprintf (dump_file, " (%d, %d) nocopy;", | |
173 | path[i]->e->src->index, path[i]->e->dest->index); | |
c5baf1e1 | 174 | if (path[0]->type == EDGE_FSM_THREAD) |
175 | fprintf (dump_file, " (%d, %d) ", | |
176 | path[i]->e->src->index, path[i]->e->dest->index); | |
b93ba654 | 177 | } |
178 | fputc ('\n', dump_file); | |
179 | } | |
180 | ||
5fe6149c | 181 | /* Simple hashing function. For any given incoming edge E, we're going |
182 | to be most concerned with the final destination of its jump thread | |
183 | path. So hash on the block index of the final edge in the path. */ | |
184 | ||
494bbaae | 185 | inline hashval_t |
9969c043 | 186 | redirection_data::hash (const redirection_data *p) |
494bbaae | 187 | { |
5fe6149c | 188 | vec<jump_thread_edge *> *path = p->path; |
189 | return path->last ()->e->dest->index; | |
494bbaae | 190 | } |
191 | ||
5fe6149c | 192 | /* Given two hash table entries, return true if they have the same |
193 | jump threading path. */ | |
494bbaae | 194 | inline int |
9969c043 | 195 | redirection_data::equal (const redirection_data *p1, const redirection_data *p2) |
494bbaae | 196 | { |
5fe6149c | 197 | vec<jump_thread_edge *> *path1 = p1->path; |
198 | vec<jump_thread_edge *> *path2 = p2->path; | |
199 | ||
200 | if (path1->length () != path2->length ()) | |
201 | return false; | |
202 | ||
203 | for (unsigned int i = 1; i < path1->length (); i++) | |
204 | { | |
205 | if ((*path1)[i]->type != (*path2)[i]->type | |
206 | || (*path1)[i]->e != (*path2)[i]->e) | |
207 | return false; | |
208 | } | |
209 | ||
210 | return true; | |
494bbaae | 211 | } |
212 | ||
a27d141e | 213 | /* Rather than search all the edges in jump thread paths each time |
214 | DOM is able to simply if control statement, we build a hash table | |
215 | with the deleted edges. We only care about the address of the edge, | |
216 | not its contents. */ | |
217 | struct removed_edges : nofree_ptr_hash<edge_def> | |
218 | { | |
219 | static hashval_t hash (edge e) { return htab_hash_pointer (e); } | |
220 | static bool equal (edge e1, edge e2) { return e1 == e2; } | |
221 | }; | |
222 | ||
223 | static hash_table<removed_edges> *removed_edges; | |
224 | ||
778182c1 | 225 | /* Data structure of information to pass to hash table traversal routines. */ |
2b15d2ba | 226 | struct ssa_local_info_t |
778182c1 | 227 | { |
228 | /* The current block we are working on. */ | |
229 | basic_block bb; | |
230 | ||
11af02d8 | 231 | /* We only create a template block for the first duplicated block in a |
232 | jump threading path as we may need many duplicates of that block. | |
233 | ||
234 | The second duplicate block in a path is specific to that path. Creating | |
235 | and sharing a template for that block is considerably more difficult. */ | |
778182c1 | 236 | basic_block template_block; |
388d1fc1 | 237 | |
30e432bb | 238 | /* Blocks duplicated for the thread. */ |
239 | bitmap duplicate_blocks; | |
d07cbccc | 240 | |
487798e2 | 241 | /* TRUE if we thread one or more jumps, FALSE otherwise. */ |
242 | bool jumps_threaded; | |
243 | ||
d07cbccc | 244 | /* When we have multiple paths through a joiner which reach different |
245 | final destinations, then we may need to correct for potential | |
246 | profile insanities. */ | |
247 | bool need_profile_correction; | |
778182c1 | 248 | }; |
a3d0fd80 | 249 | |
3cebc9d2 | 250 | /* Passes which use the jump threading code register jump threading |
251 | opportunities as they are discovered. We keep the registered | |
252 | jump threading opportunities in this vector as edge pairs | |
253 | (original_edge, target_edge). */ | |
f2981b08 | 254 | static vec<vec<jump_thread_edge *> *> paths; |
3cebc9d2 | 255 | |
eb31063a | 256 | /* When we start updating the CFG for threading, data necessary for jump |
257 | threading is attached to the AUX field for the incoming edge. Use these | |
258 | macros to access the underlying structure attached to the AUX field. */ | |
f2981b08 | 259 | #define THREAD_PATH(E) ((vec<jump_thread_edge *> *)(E)->aux) |
3cebc9d2 | 260 | |
5236b8bb | 261 | /* Jump threading statistics. */ |
262 | ||
263 | struct thread_stats_d | |
264 | { | |
265 | unsigned long num_threaded_edges; | |
266 | }; | |
267 | ||
268 | struct thread_stats_d thread_stats; | |
269 | ||
270 | ||
f582bb6c | 271 | /* Remove the last statement in block BB if it is a control statement |
272 | Also remove all outgoing edges except the edge which reaches DEST_BB. | |
273 | If DEST_BB is NULL, then remove all outgoing edges. */ | |
a8046f60 | 274 | |
f1344f45 | 275 | void |
f582bb6c | 276 | remove_ctrl_stmt_and_useless_edges (basic_block bb, basic_block dest_bb) |
a8046f60 | 277 | { |
75a70cf9 | 278 | gimple_stmt_iterator gsi; |
cd665a06 | 279 | edge e; |
280 | edge_iterator ei; | |
a8046f60 | 281 | |
75a70cf9 | 282 | gsi = gsi_last_bb (bb); |
a8046f60 | 283 | |
f582bb6c | 284 | /* If the duplicate ends with a control statement, then remove it. |
a8046f60 | 285 | |
f582bb6c | 286 | Note that if we are duplicating the template block rather than the |
287 | original basic block, then the duplicate might not have any real | |
288 | statements in it. */ | |
75a70cf9 | 289 | if (!gsi_end_p (gsi) |
290 | && gsi_stmt (gsi) | |
291 | && (gimple_code (gsi_stmt (gsi)) == GIMPLE_COND | |
292 | || gimple_code (gsi_stmt (gsi)) == GIMPLE_GOTO | |
293 | || gimple_code (gsi_stmt (gsi)) == GIMPLE_SWITCH)) | |
294 | gsi_remove (&gsi, true); | |
a8046f60 | 295 | |
cd665a06 | 296 | for (ei = ei_start (bb->succs); (e = ei_safe_edge (ei)); ) |
a8046f60 | 297 | { |
a8046f60 | 298 | if (e->dest != dest_bb) |
b7cbf36d | 299 | { |
300 | free_dom_edge_info (e); | |
301 | remove_edge (e); | |
302 | } | |
cd665a06 | 303 | else |
acb9fddd | 304 | { |
720cfc43 | 305 | e->probability = profile_probability::always (); |
acb9fddd | 306 | e->count = bb->count; |
307 | ei_next (&ei); | |
308 | } | |
a8046f60 | 309 | } |
5b4ada2a | 310 | |
311 | /* If the remaining edge is a loop exit, there must have | |
312 | a removed edge that was not a loop exit. | |
313 | ||
314 | In that case BB and possibly other blocks were previously | |
315 | in the loop, but are now outside the loop. Thus, we need | |
316 | to update the loop structures. */ | |
317 | if (single_succ_p (bb) | |
318 | && loop_outer (bb->loop_father) | |
319 | && loop_exit_edge_p (bb->loop_father, single_succ_edge (bb))) | |
320 | loops_state_set (LOOPS_NEED_FIXUP); | |
a8046f60 | 321 | } |
322 | ||
11af02d8 | 323 | /* Create a duplicate of BB. Record the duplicate block in an array |
a7ee7309 | 324 | indexed by COUNT stored in RD. */ |
a8046f60 | 325 | |
326 | static void | |
11af02d8 | 327 | create_block_for_threading (basic_block bb, |
328 | struct redirection_data *rd, | |
30e432bb | 329 | unsigned int count, |
a7ee7309 | 330 | bitmap *duplicate_blocks) |
a8046f60 | 331 | { |
eb31063a | 332 | edge_iterator ei; |
333 | edge e; | |
334 | ||
a8046f60 | 335 | /* We can use the generic block duplication code and simply remove |
336 | the stuff we do not need. */ | |
11af02d8 | 337 | rd->dup_blocks[count] = duplicate_block (bb, NULL, NULL); |
a8046f60 | 338 | |
11af02d8 | 339 | FOR_EACH_EDGE (e, ei, rd->dup_blocks[count]->succs) |
eb31063a | 340 | e->aux = NULL; |
341 | ||
615dd397 | 342 | /* Zero out the profile, since the block is unreachable for now. */ |
11af02d8 | 343 | rd->dup_blocks[count]->frequency = 0; |
db9cef39 | 344 | rd->dup_blocks[count]->count = profile_count::uninitialized (); |
30e432bb | 345 | if (duplicate_blocks) |
346 | bitmap_set_bit (*duplicate_blocks, rd->dup_blocks[count]->index); | |
a8046f60 | 347 | } |
348 | ||
2b15d2ba | 349 | /* Main data structure to hold information for duplicates of BB. */ |
350 | ||
c1f445d2 | 351 | static hash_table<redirection_data> *redirection_data; |
2b15d2ba | 352 | |
778182c1 | 353 | /* Given an outgoing edge E lookup and return its entry in our hash table. |
354 | ||
355 | If INSERT is true, then we insert the entry into the hash table if | |
356 | it is not already present. INCOMING_EDGE is added to the list of incoming | |
357 | edges associated with E in the hash table. */ | |
358 | ||
359 | static struct redirection_data * | |
da81e0c5 | 360 | lookup_redirection_data (edge e, enum insert_option insert) |
778182c1 | 361 | { |
2b15d2ba | 362 | struct redirection_data **slot; |
778182c1 | 363 | struct redirection_data *elt; |
f2981b08 | 364 | vec<jump_thread_edge *> *path = THREAD_PATH (e); |
778182c1 | 365 | |
31e5d72d | 366 | /* Build a hash table element so we can see if E is already |
778182c1 | 367 | in the table. */ |
4c36ffe6 | 368 | elt = XNEW (struct redirection_data); |
5fe6149c | 369 | elt->path = path; |
11af02d8 | 370 | elt->dup_blocks[0] = NULL; |
371 | elt->dup_blocks[1] = NULL; | |
778182c1 | 372 | elt->incoming_edges = NULL; |
373 | ||
c1f445d2 | 374 | slot = redirection_data->find_slot (elt, insert); |
778182c1 | 375 | |
376 | /* This will only happen if INSERT is false and the entry is not | |
377 | in the hash table. */ | |
378 | if (slot == NULL) | |
379 | { | |
380 | free (elt); | |
381 | return NULL; | |
382 | } | |
383 | ||
384 | /* This will only happen if E was not in the hash table and | |
385 | INSERT is true. */ | |
386 | if (*slot == NULL) | |
387 | { | |
2b15d2ba | 388 | *slot = elt; |
4c36ffe6 | 389 | elt->incoming_edges = XNEW (struct el); |
da81e0c5 | 390 | elt->incoming_edges->e = e; |
778182c1 | 391 | elt->incoming_edges->next = NULL; |
392 | return elt; | |
393 | } | |
394 | /* E was in the hash table. */ | |
395 | else | |
396 | { | |
397 | /* Free ELT as we do not need it anymore, we will extract the | |
398 | relevant entry from the hash table itself. */ | |
399 | free (elt); | |
400 | ||
401 | /* Get the entry stored in the hash table. */ | |
2b15d2ba | 402 | elt = *slot; |
778182c1 | 403 | |
404 | /* If insertion was requested, then we need to add INCOMING_EDGE | |
405 | to the list of incoming edges associated with E. */ | |
406 | if (insert) | |
407 | { | |
559685be | 408 | struct el *el = XNEW (struct el); |
778182c1 | 409 | el->next = elt->incoming_edges; |
da81e0c5 | 410 | el->e = e; |
778182c1 | 411 | elt->incoming_edges = el; |
412 | } | |
413 | ||
414 | return elt; | |
415 | } | |
416 | } | |
417 | ||
fc54aba7 | 418 | /* Similar to copy_phi_args, except that the PHI arg exists, it just |
419 | does not have a value associated with it. */ | |
420 | ||
421 | static void | |
422 | copy_phi_arg_into_existing_phi (edge src_e, edge tgt_e) | |
423 | { | |
424 | int src_idx = src_e->dest_idx; | |
425 | int tgt_idx = tgt_e->dest_idx; | |
426 | ||
427 | /* Iterate over each PHI in e->dest. */ | |
1a91d914 | 428 | for (gphi_iterator gsi = gsi_start_phis (src_e->dest), |
429 | gsi2 = gsi_start_phis (tgt_e->dest); | |
fc54aba7 | 430 | !gsi_end_p (gsi); |
431 | gsi_next (&gsi), gsi_next (&gsi2)) | |
432 | { | |
1a91d914 | 433 | gphi *src_phi = gsi.phi (); |
434 | gphi *dest_phi = gsi2.phi (); | |
fc54aba7 | 435 | tree val = gimple_phi_arg_def (src_phi, src_idx); |
436 | source_location locus = gimple_phi_arg_location (src_phi, src_idx); | |
437 | ||
438 | SET_PHI_ARG_DEF (dest_phi, tgt_idx, val); | |
439 | gimple_phi_arg_set_location (dest_phi, tgt_idx, locus); | |
440 | } | |
441 | } | |
442 | ||
1b83c31b | 443 | /* Given ssa_name DEF, backtrack jump threading PATH from node IDX |
444 | to see if it has constant value in a flow sensitive manner. Set | |
445 | LOCUS to location of the constant phi arg and return the value. | |
446 | Return DEF directly if either PATH or idx is ZERO. */ | |
447 | ||
448 | static tree | |
449 | get_value_locus_in_path (tree def, vec<jump_thread_edge *> *path, | |
450 | basic_block bb, int idx, source_location *locus) | |
451 | { | |
452 | tree arg; | |
1a91d914 | 453 | gphi *def_phi; |
1b83c31b | 454 | basic_block def_bb; |
455 | ||
456 | if (path == NULL || idx == 0) | |
457 | return def; | |
458 | ||
1a91d914 | 459 | def_phi = dyn_cast <gphi *> (SSA_NAME_DEF_STMT (def)); |
460 | if (!def_phi) | |
1b83c31b | 461 | return def; |
462 | ||
463 | def_bb = gimple_bb (def_phi); | |
464 | /* Don't propagate loop invariants into deeper loops. */ | |
465 | if (!def_bb || bb_loop_depth (def_bb) < bb_loop_depth (bb)) | |
466 | return def; | |
467 | ||
468 | /* Backtrack jump threading path from IDX to see if def has constant | |
469 | value. */ | |
470 | for (int j = idx - 1; j >= 0; j--) | |
471 | { | |
472 | edge e = (*path)[j]->e; | |
473 | if (e->dest == def_bb) | |
474 | { | |
475 | arg = gimple_phi_arg_def (def_phi, e->dest_idx); | |
476 | if (is_gimple_min_invariant (arg)) | |
477 | { | |
478 | *locus = gimple_phi_arg_location (def_phi, e->dest_idx); | |
479 | return arg; | |
480 | } | |
481 | break; | |
482 | } | |
483 | } | |
484 | ||
485 | return def; | |
486 | } | |
487 | ||
488 | /* For each PHI in BB, copy the argument associated with SRC_E to TGT_E. | |
489 | Try to backtrack jump threading PATH from node IDX to see if the arg | |
490 | has constant value, copy constant value instead of argument itself | |
491 | if yes. */ | |
da81e0c5 | 492 | |
493 | static void | |
1b83c31b | 494 | copy_phi_args (basic_block bb, edge src_e, edge tgt_e, |
495 | vec<jump_thread_edge *> *path, int idx) | |
da81e0c5 | 496 | { |
1a91d914 | 497 | gphi_iterator gsi; |
da81e0c5 | 498 | int src_indx = src_e->dest_idx; |
499 | ||
500 | for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
501 | { | |
1a91d914 | 502 | gphi *phi = gsi.phi (); |
1b83c31b | 503 | tree def = gimple_phi_arg_def (phi, src_indx); |
da81e0c5 | 504 | source_location locus = gimple_phi_arg_location (phi, src_indx); |
1b83c31b | 505 | |
506 | if (TREE_CODE (def) == SSA_NAME | |
507 | && !virtual_operand_p (gimple_phi_result (phi))) | |
508 | def = get_value_locus_in_path (def, path, bb, idx, &locus); | |
509 | ||
510 | add_phi_arg (phi, def, tgt_e, locus); | |
da81e0c5 | 511 | } |
512 | } | |
513 | ||
514 | /* We have recently made a copy of ORIG_BB, including its outgoing | |
515 | edges. The copy is NEW_BB. Every PHI node in every direct successor of | |
516 | ORIG_BB has a new argument associated with edge from NEW_BB to the | |
517 | successor. Initialize the PHI argument so that it is equal to the PHI | |
1b83c31b | 518 | argument associated with the edge from ORIG_BB to the successor. |
519 | PATH and IDX are used to check if the new PHI argument has constant | |
520 | value in a flow sensitive manner. */ | |
da81e0c5 | 521 | |
522 | static void | |
1b83c31b | 523 | update_destination_phis (basic_block orig_bb, basic_block new_bb, |
524 | vec<jump_thread_edge *> *path, int idx) | |
da81e0c5 | 525 | { |
526 | edge_iterator ei; | |
527 | edge e; | |
528 | ||
529 | FOR_EACH_EDGE (e, ei, orig_bb->succs) | |
530 | { | |
531 | edge e2 = find_edge (new_bb, e->dest); | |
1b83c31b | 532 | copy_phi_args (e->dest, e, e2, path, idx); |
da81e0c5 | 533 | } |
534 | } | |
535 | ||
778182c1 | 536 | /* Given a duplicate block and its single destination (both stored |
537 | in RD). Create an edge between the duplicate and its single | |
538 | destination. | |
539 | ||
540 | Add an additional argument to any PHI nodes at the single | |
1b83c31b | 541 | destination. IDX is the start node in jump threading path |
542 | we start to check to see if the new PHI argument has constant | |
543 | value along the jump threading path. */ | |
778182c1 | 544 | |
545 | static void | |
42b013bc | 546 | create_edge_and_update_destination_phis (struct redirection_data *rd, |
1b83c31b | 547 | basic_block bb, int idx) |
778182c1 | 548 | { |
720cfc43 | 549 | edge e = make_single_succ_edge (bb, rd->path->last ()->e->dest, EDGE_FALLTHRU); |
778182c1 | 550 | |
f9614b84 | 551 | rescan_loop_exit (e, true, false); |
eb31063a | 552 | |
e63988cc | 553 | /* We used to copy the thread path here. That was added in 2007 |
554 | and dutifully updated through the representation changes in 2013. | |
555 | ||
556 | In 2013 we added code to thread from an interior node through | |
557 | the backedge to another interior node. That runs after the code | |
558 | to thread through loop headers from outside the loop. | |
559 | ||
560 | The latter may delete edges in the CFG, including those | |
561 | which appeared in the jump threading path we copied here. Thus | |
562 | we'd end up using a dangling pointer. | |
563 | ||
564 | After reviewing the 2007/2011 code, I can't see how anything | |
565 | depended on copying the AUX field and clearly copying the jump | |
566 | threading path is problematical due to embedded edge pointers. | |
567 | It has been removed. */ | |
568 | e->aux = NULL; | |
421e19dd | 569 | |
778182c1 | 570 | /* If there are any PHI nodes at the destination of the outgoing edge |
571 | from the duplicate block, then we will need to add a new argument | |
572 | to them. The argument should have the same value as the argument | |
573 | associated with the outgoing edge stored in RD. */ | |
1b83c31b | 574 | copy_phi_args (e->dest, rd->path->last ()->e, e, rd->path, idx); |
da81e0c5 | 575 | } |
576 | ||
fc54aba7 | 577 | /* Look through PATH beginning at START and return TRUE if there are |
578 | any additional blocks that need to be duplicated. Otherwise, | |
579 | return FALSE. */ | |
580 | static bool | |
581 | any_remaining_duplicated_blocks (vec<jump_thread_edge *> *path, | |
582 | unsigned int start) | |
583 | { | |
584 | for (unsigned int i = start + 1; i < path->length (); i++) | |
585 | { | |
586 | if ((*path)[i]->type == EDGE_COPY_SRC_JOINER_BLOCK | |
587 | || (*path)[i]->type == EDGE_COPY_SRC_BLOCK) | |
588 | return true; | |
589 | } | |
590 | return false; | |
591 | } | |
592 | ||
30e432bb | 593 | |
594 | /* Compute the amount of profile count/frequency coming into the jump threading | |
595 | path stored in RD that we are duplicating, returned in PATH_IN_COUNT_PTR and | |
596 | PATH_IN_FREQ_PTR, as well as the amount of counts flowing out of the | |
597 | duplicated path, returned in PATH_OUT_COUNT_PTR. LOCAL_INFO is used to | |
598 | identify blocks duplicated for jump threading, which have duplicated | |
599 | edges that need to be ignored in the analysis. Return true if path contains | |
600 | a joiner, false otherwise. | |
601 | ||
602 | In the non-joiner case, this is straightforward - all the counts/frequency | |
603 | flowing into the jump threading path should flow through the duplicated | |
604 | block and out of the duplicated path. | |
605 | ||
606 | In the joiner case, it is very tricky. Some of the counts flowing into | |
607 | the original path go offpath at the joiner. The problem is that while | |
608 | we know how much total count goes off-path in the original control flow, | |
609 | we don't know how many of the counts corresponding to just the jump | |
610 | threading path go offpath at the joiner. | |
611 | ||
612 | For example, assume we have the following control flow and identified | |
613 | jump threading paths: | |
614 | ||
9943b198 | 615 | A B C |
616 | \ | / | |
617 | Ea \ |Eb / Ec | |
618 | \ | / | |
619 | v v v | |
620 | J <-- Joiner | |
621 | / \ | |
622 | Eoff/ \Eon | |
623 | / \ | |
624 | v v | |
625 | Soff Son <--- Normal | |
626 | /\ | |
627 | Ed/ \ Ee | |
628 | / \ | |
629 | v v | |
630 | D E | |
631 | ||
632 | Jump threading paths: A -> J -> Son -> D (path 1) | |
633 | C -> J -> Son -> E (path 2) | |
30e432bb | 634 | |
635 | Note that the control flow could be more complicated: | |
636 | - Each jump threading path may have more than one incoming edge. I.e. A and | |
637 | Ea could represent multiple incoming blocks/edges that are included in | |
638 | path 1. | |
639 | - There could be EDGE_NO_COPY_SRC_BLOCK edges after the joiner (either | |
640 | before or after the "normal" copy block). These are not duplicated onto | |
641 | the jump threading path, as they are single-successor. | |
642 | - Any of the blocks along the path may have other incoming edges that | |
643 | are not part of any jump threading path, but add profile counts along | |
644 | the path. | |
645 | ||
31e5d72d | 646 | In the above example, after all jump threading is complete, we will |
30e432bb | 647 | end up with the following control flow: |
648 | ||
31e5d72d | 649 | A B C |
650 | | | | | |
651 | Ea| |Eb |Ec | |
652 | | | | | |
653 | v v v | |
654 | Ja J Jc | |
655 | / \ / \Eon' / \ | |
9943b198 | 656 | Eona/ \ ---/---\-------- \Eonc |
31e5d72d | 657 | / \ / / \ \ |
9943b198 | 658 | v v v v v |
659 | Sona Soff Son Sonc | |
31e5d72d | 660 | \ /\ / |
9943b198 | 661 | \___________ / \ _____/ |
662 | \ / \/ | |
663 | vv v | |
664 | D E | |
30e432bb | 665 | |
666 | The main issue to notice here is that when we are processing path 1 | |
667 | (A->J->Son->D) we need to figure out the outgoing edge weights to | |
668 | the duplicated edges Ja->Sona and Ja->Soff, while ensuring that the | |
669 | sum of the incoming weights to D remain Ed. The problem with simply | |
670 | assuming that Ja (and Jc when processing path 2) has the same outgoing | |
671 | probabilities to its successors as the original block J, is that after | |
672 | all paths are processed and other edges/counts removed (e.g. none | |
673 | of Ec will reach D after processing path 2), we may end up with not | |
674 | enough count flowing along duplicated edge Sona->D. | |
675 | ||
676 | Therefore, in the case of a joiner, we keep track of all counts | |
677 | coming in along the current path, as well as from predecessors not | |
678 | on any jump threading path (Eb in the above example). While we | |
679 | first assume that the duplicated Eona for Ja->Sona has the same | |
680 | probability as the original, we later compensate for other jump | |
681 | threading paths that may eliminate edges. We do that by keep track | |
682 | of all counts coming into the original path that are not in a jump | |
683 | thread (Eb in the above example, but as noted earlier, there could | |
684 | be other predecessors incoming to the path at various points, such | |
685 | as at Son). Call this cumulative non-path count coming into the path | |
686 | before D as Enonpath. We then ensure that the count from Sona->D is as at | |
687 | least as big as (Ed - Enonpath), but no bigger than the minimum | |
688 | weight along the jump threading path. The probabilities of both the | |
689 | original and duplicated joiner block J and Ja will be adjusted | |
690 | accordingly after the updates. */ | |
691 | ||
692 | static bool | |
693 | compute_path_counts (struct redirection_data *rd, | |
9943b198 | 694 | ssa_local_info_t *local_info, |
db9cef39 | 695 | profile_count *path_in_count_ptr, |
696 | profile_count *path_out_count_ptr, | |
9943b198 | 697 | int *path_in_freq_ptr) |
30e432bb | 698 | { |
699 | edge e = rd->incoming_edges->e; | |
700 | vec<jump_thread_edge *> *path = THREAD_PATH (e); | |
701 | edge elast = path->last ()->e; | |
db9cef39 | 702 | profile_count nonpath_count = profile_count::zero (); |
30e432bb | 703 | bool has_joiner = false; |
db9cef39 | 704 | profile_count path_in_count = profile_count::zero (); |
30e432bb | 705 | int path_in_freq = 0; |
706 | ||
707 | /* Start by accumulating incoming edge counts to the path's first bb | |
708 | into a couple buckets: | |
9943b198 | 709 | path_in_count: total count of incoming edges that flow into the |
710 | current path. | |
711 | nonpath_count: total count of incoming edges that are not | |
712 | flowing along *any* path. These are the counts | |
713 | that will still flow along the original path after | |
714 | all path duplication is done by potentially multiple | |
715 | calls to this routine. | |
30e432bb | 716 | (any other incoming edge counts are for a different jump threading |
717 | path that will be handled by a later call to this routine.) | |
718 | To make this easier, start by recording all incoming edges that flow into | |
719 | the current path in a bitmap. We could add up the path's incoming edge | |
720 | counts here, but we still need to walk all the first bb's incoming edges | |
721 | below to add up the counts of the other edges not included in this jump | |
722 | threading path. */ | |
723 | struct el *next, *el; | |
035def86 | 724 | auto_bitmap in_edge_srcs; |
30e432bb | 725 | for (el = rd->incoming_edges; el; el = next) |
726 | { | |
727 | next = el->next; | |
728 | bitmap_set_bit (in_edge_srcs, el->e->src->index); | |
729 | } | |
730 | edge ein; | |
731 | edge_iterator ei; | |
732 | FOR_EACH_EDGE (ein, ei, e->dest->preds) | |
733 | { | |
734 | vec<jump_thread_edge *> *ein_path = THREAD_PATH (ein); | |
735 | /* Simply check the incoming edge src against the set captured above. */ | |
736 | if (ein_path | |
9943b198 | 737 | && bitmap_bit_p (in_edge_srcs, (*ein_path)[0]->e->src->index)) |
738 | { | |
739 | /* It is necessary but not sufficient that the last path edges | |
740 | are identical. There may be different paths that share the | |
741 | same last path edge in the case where the last edge has a nocopy | |
742 | source block. */ | |
743 | gcc_assert (ein_path->last ()->e == elast); | |
744 | path_in_count += ein->count; | |
745 | path_in_freq += EDGE_FREQUENCY (ein); | |
746 | } | |
30e432bb | 747 | else if (!ein_path) |
9943b198 | 748 | { |
749 | /* Keep track of the incoming edges that are not on any jump-threading | |
750 | path. These counts will still flow out of original path after all | |
751 | jump threading is complete. */ | |
752 | nonpath_count += ein->count; | |
753 | } | |
30e432bb | 754 | } |
664dd751 | 755 | |
756 | /* This is needed due to insane incoming frequencies. */ | |
757 | if (path_in_freq > BB_FREQ_MAX) | |
758 | path_in_freq = BB_FREQ_MAX; | |
759 | ||
30e432bb | 760 | /* Now compute the fraction of the total count coming into the first |
761 | path bb that is from the current threading path. */ | |
db9cef39 | 762 | profile_count total_count = e->dest->count; |
30e432bb | 763 | /* Handle incoming profile insanities. */ |
764 | if (total_count < path_in_count) | |
765 | path_in_count = total_count; | |
720cfc43 | 766 | int onpath_scale |
767 | = path_in_count.probability_in (total_count).to_reg_br_prob_base (); | |
30e432bb | 768 | |
769 | /* Walk the entire path to do some more computation in order to estimate | |
770 | how much of the path_in_count will flow out of the duplicated threading | |
771 | path. In the non-joiner case this is straightforward (it should be | |
772 | the same as path_in_count, although we will handle incoming profile | |
773 | insanities by setting it equal to the minimum count along the path). | |
774 | ||
775 | In the joiner case, we need to estimate how much of the path_in_count | |
776 | will stay on the threading path after the joiner's conditional branch. | |
777 | We don't really know for sure how much of the counts | |
778 | associated with this path go to each successor of the joiner, but we'll | |
779 | estimate based on the fraction of the total count coming into the path | |
780 | bb was from the threading paths (computed above in onpath_scale). | |
781 | Afterwards, we will need to do some fixup to account for other threading | |
782 | paths and possible profile insanities. | |
783 | ||
784 | In order to estimate the joiner case's counts we also need to update | |
785 | nonpath_count with any additional counts coming into the path. Other | |
786 | blocks along the path may have additional predecessors from outside | |
787 | the path. */ | |
db9cef39 | 788 | profile_count path_out_count = path_in_count; |
789 | profile_count min_path_count = path_in_count; | |
30e432bb | 790 | for (unsigned int i = 1; i < path->length (); i++) |
791 | { | |
792 | edge epath = (*path)[i]->e; | |
db9cef39 | 793 | profile_count cur_count = epath->count; |
30e432bb | 794 | if ((*path)[i]->type == EDGE_COPY_SRC_JOINER_BLOCK) |
9943b198 | 795 | { |
796 | has_joiner = true; | |
db9cef39 | 797 | cur_count = cur_count.apply_probability (onpath_scale); |
9943b198 | 798 | } |
30e432bb | 799 | /* In the joiner case we need to update nonpath_count for any edges |
9943b198 | 800 | coming into the path that will contribute to the count flowing |
801 | into the path successor. */ | |
30e432bb | 802 | if (has_joiner && epath != elast) |
31e5d72d | 803 | { |
804 | /* Look for other incoming edges after joiner. */ | |
805 | FOR_EACH_EDGE (ein, ei, epath->dest->preds) | |
806 | { | |
807 | if (ein != epath | |
808 | /* Ignore in edges from blocks we have duplicated for a | |
809 | threading path, which have duplicated edge counts until | |
810 | they are redirected by an invocation of this routine. */ | |
811 | && !bitmap_bit_p (local_info->duplicate_blocks, | |
812 | ein->src->index)) | |
813 | nonpath_count += ein->count; | |
814 | } | |
815 | } | |
30e432bb | 816 | if (cur_count < path_out_count) |
9943b198 | 817 | path_out_count = cur_count; |
30e432bb | 818 | if (epath->count < min_path_count) |
9943b198 | 819 | min_path_count = epath->count; |
30e432bb | 820 | } |
821 | ||
822 | /* We computed path_out_count above assuming that this path targeted | |
823 | the joiner's on-path successor with the same likelihood as it | |
824 | reached the joiner. However, other thread paths through the joiner | |
825 | may take a different path through the normal copy source block | |
826 | (i.e. they have a different elast), meaning that they do not | |
827 | contribute any counts to this path's elast. As a result, it may | |
828 | turn out that this path must have more count flowing to the on-path | |
829 | successor of the joiner. Essentially, all of this path's elast | |
830 | count must be contributed by this path and any nonpath counts | |
831 | (since any path through the joiner with a different elast will not | |
832 | include a copy of this elast in its duplicated path). | |
833 | So ensure that this path's path_out_count is at least the | |
834 | difference between elast->count and nonpath_count. Otherwise the edge | |
835 | counts after threading will not be sane. */ | |
d07cbccc | 836 | if (local_info->need_profile_correction |
837 | && has_joiner && path_out_count < elast->count - nonpath_count) | |
31e5d72d | 838 | { |
839 | path_out_count = elast->count - nonpath_count; | |
840 | /* But neither can we go above the minimum count along the path | |
841 | we are duplicating. This can be an issue due to profile | |
842 | insanities coming in to this pass. */ | |
843 | if (path_out_count > min_path_count) | |
844 | path_out_count = min_path_count; | |
845 | } | |
30e432bb | 846 | |
847 | *path_in_count_ptr = path_in_count; | |
848 | *path_out_count_ptr = path_out_count; | |
849 | *path_in_freq_ptr = path_in_freq; | |
850 | return has_joiner; | |
851 | } | |
852 | ||
853 | ||
854 | /* Update the counts and frequencies for both an original path | |
855 | edge EPATH and its duplicate EDUP. The duplicate source block | |
856 | will get a count/frequency of PATH_IN_COUNT and PATH_IN_FREQ, | |
857 | and the duplicate edge EDUP will have a count of PATH_OUT_COUNT. */ | |
858 | static void | |
db9cef39 | 859 | update_profile (edge epath, edge edup, profile_count path_in_count, |
860 | profile_count path_out_count, int path_in_freq) | |
30e432bb | 861 | { |
862 | ||
863 | /* First update the duplicated block's count / frequency. */ | |
864 | if (edup) | |
865 | { | |
866 | basic_block dup_block = edup->src; | |
db9cef39 | 867 | gcc_assert (!dup_block->count.initialized_p ()); |
30e432bb | 868 | gcc_assert (dup_block->frequency == 0); |
869 | dup_block->count = path_in_count; | |
870 | dup_block->frequency = path_in_freq; | |
871 | } | |
872 | ||
873 | /* Now update the original block's count and frequency in the | |
874 | opposite manner - remove the counts/freq that will flow | |
875 | into the duplicated block. Handle underflow due to precision/ | |
876 | rounding issues. */ | |
877 | epath->src->count -= path_in_count; | |
30e432bb | 878 | epath->src->frequency -= path_in_freq; |
879 | if (epath->src->frequency < 0) | |
880 | epath->src->frequency = 0; | |
881 | ||
882 | /* Next update this path edge's original and duplicated counts. We know | |
883 | that the duplicated path will have path_out_count flowing | |
884 | out of it (in the joiner case this is the count along the duplicated path | |
885 | out of the duplicated joiner). This count can then be removed from the | |
886 | original path edge. */ | |
887 | if (edup) | |
888 | edup->count = path_out_count; | |
889 | epath->count -= path_out_count; | |
db9cef39 | 890 | /* FIXME: can epath->count be legally uninitialized here? */ |
30e432bb | 891 | } |
892 | ||
893 | ||
894 | /* The duplicate and original joiner blocks may end up with different | |
895 | probabilities (different from both the original and from each other). | |
896 | Recompute the probabilities here once we have updated the edge | |
897 | counts and frequencies. */ | |
898 | ||
899 | static void | |
900 | recompute_probabilities (basic_block bb) | |
901 | { | |
902 | edge esucc; | |
903 | edge_iterator ei; | |
904 | FOR_EACH_EDGE (esucc, ei, bb->succs) | |
905 | { | |
db9cef39 | 906 | if (!(bb->count > 0)) |
9943b198 | 907 | continue; |
bdd367a0 | 908 | |
909 | /* Prevent overflow computation due to insane profiles. */ | |
910 | if (esucc->count < bb->count) | |
4bb697cd | 911 | esucc->probability = esucc->count.probability_in (bb->count).guessed (); |
bdd367a0 | 912 | else |
9943b198 | 913 | /* Can happen with missing/guessed probabilities, since we |
914 | may determine that more is flowing along duplicated | |
915 | path than joiner succ probabilities allowed. | |
916 | Counts and freqs will be insane after jump threading, | |
917 | at least make sure probability is sane or we will | |
918 | get a flow verification error. | |
919 | Not much we can do to make counts/freqs sane without | |
920 | redoing the profile estimation. */ | |
720cfc43 | 921 | esucc->probability = profile_probability::guessed_always (); |
30e432bb | 922 | } |
923 | } | |
924 | ||
925 | ||
926 | /* Update the counts of the original and duplicated edges from a joiner | |
927 | that go off path, given that we have already determined that the | |
928 | duplicate joiner DUP_BB has incoming count PATH_IN_COUNT and | |
929 | outgoing count along the path PATH_OUT_COUNT. The original (on-)path | |
930 | edge from joiner is EPATH. */ | |
931 | ||
932 | static void | |
933 | update_joiner_offpath_counts (edge epath, basic_block dup_bb, | |
db9cef39 | 934 | profile_count path_in_count, |
935 | profile_count path_out_count) | |
30e432bb | 936 | { |
937 | /* Compute the count that currently flows off path from the joiner. | |
938 | In other words, the total count of joiner's out edges other than | |
939 | epath. Compute this by walking the successors instead of | |
940 | subtracting epath's count from the joiner bb count, since there | |
941 | are sometimes slight insanities where the total out edge count is | |
942 | larger than the bb count (possibly due to rounding/truncation | |
943 | errors). */ | |
db9cef39 | 944 | profile_count total_orig_off_path_count = profile_count::zero (); |
30e432bb | 945 | edge enonpath; |
946 | edge_iterator ei; | |
947 | FOR_EACH_EDGE (enonpath, ei, epath->src->succs) | |
948 | { | |
949 | if (enonpath == epath) | |
9943b198 | 950 | continue; |
30e432bb | 951 | total_orig_off_path_count += enonpath->count; |
952 | } | |
953 | ||
954 | /* For the path that we are duplicating, the amount that will flow | |
955 | off path from the duplicated joiner is the delta between the | |
956 | path's cumulative in count and the portion of that count we | |
957 | estimated above as flowing from the joiner along the duplicated | |
958 | path. */ | |
db9cef39 | 959 | profile_count total_dup_off_path_count = path_in_count - path_out_count; |
30e432bb | 960 | |
961 | /* Now do the actual updates of the off-path edges. */ | |
962 | FOR_EACH_EDGE (enonpath, ei, epath->src->succs) | |
963 | { | |
964 | /* Look for edges going off of the threading path. */ | |
965 | if (enonpath == epath) | |
9943b198 | 966 | continue; |
30e432bb | 967 | |
968 | /* Find the corresponding edge out of the duplicated joiner. */ | |
969 | edge enonpathdup = find_edge (dup_bb, enonpath->dest); | |
970 | gcc_assert (enonpathdup); | |
971 | ||
972 | /* We can't use the original probability of the joiner's out | |
9943b198 | 973 | edges, since the probabilities of the original branch |
974 | and the duplicated branches may vary after all threading is | |
975 | complete. But apportion the duplicated joiner's off-path | |
976 | total edge count computed earlier (total_dup_off_path_count) | |
977 | among the duplicated off-path edges based on their original | |
978 | ratio to the full off-path count (total_orig_off_path_count). | |
979 | */ | |
720cfc43 | 980 | int scale = enonpath->count.probability_in (total_orig_off_path_count) |
981 | .to_reg_br_prob_base (); | |
30e432bb | 982 | /* Give the duplicated offpath edge a portion of the duplicated |
9943b198 | 983 | total. */ |
db9cef39 | 984 | enonpathdup->count = total_dup_off_path_count.apply_probability (scale); |
30e432bb | 985 | /* Now update the original offpath edge count, handling underflow |
9943b198 | 986 | due to rounding errors. */ |
30e432bb | 987 | enonpath->count -= enonpathdup->count; |
30e432bb | 988 | } |
989 | } | |
990 | ||
991 | ||
f1ce4e72 | 992 | /* Check if the paths through RD all have estimated frequencies but zero |
993 | profile counts. This is more accurate than checking the entry block | |
994 | for a zero profile count, since profile insanities sometimes creep in. */ | |
995 | ||
996 | static bool | |
997 | estimated_freqs_path (struct redirection_data *rd) | |
998 | { | |
999 | edge e = rd->incoming_edges->e; | |
1000 | vec<jump_thread_edge *> *path = THREAD_PATH (e); | |
1001 | edge ein; | |
1002 | edge_iterator ei; | |
1003 | bool non_zero_freq = false; | |
1004 | FOR_EACH_EDGE (ein, ei, e->dest->preds) | |
1005 | { | |
db9cef39 | 1006 | if (ein->count > 0) |
9943b198 | 1007 | return false; |
f1ce4e72 | 1008 | non_zero_freq |= ein->src->frequency != 0; |
1009 | } | |
1010 | ||
1011 | for (unsigned int i = 1; i < path->length (); i++) | |
1012 | { | |
1013 | edge epath = (*path)[i]->e; | |
db9cef39 | 1014 | if (epath->src->count > 0) |
9943b198 | 1015 | return false; |
f1ce4e72 | 1016 | non_zero_freq |= epath->src->frequency != 0; |
1017 | edge esucc; | |
1018 | FOR_EACH_EDGE (esucc, ei, epath->src->succs) | |
9943b198 | 1019 | { |
db9cef39 | 1020 | if (esucc->count > 0) |
9943b198 | 1021 | return false; |
1022 | non_zero_freq |= esucc->src->frequency != 0; | |
1023 | } | |
f1ce4e72 | 1024 | } |
1025 | return non_zero_freq; | |
1026 | } | |
1027 | ||
1028 | ||
30e432bb | 1029 | /* Invoked for routines that have guessed frequencies and no profile |
1030 | counts to record the block and edge frequencies for paths through RD | |
1031 | in the profile count fields of those blocks and edges. This is because | |
1032 | ssa_fix_duplicate_block_edges incrementally updates the block and | |
1033 | edge counts as edges are redirected, and it is difficult to do that | |
1034 | for edge frequencies which are computed on the fly from the source | |
1035 | block frequency and probability. When a block frequency is updated | |
1036 | its outgoing edge frequencies are affected and become difficult to | |
1037 | adjust. */ | |
1038 | ||
1039 | static void | |
1040 | freqs_to_counts_path (struct redirection_data *rd) | |
1041 | { | |
1042 | edge e = rd->incoming_edges->e; | |
1043 | vec<jump_thread_edge *> *path = THREAD_PATH (e); | |
1044 | edge ein; | |
1045 | edge_iterator ei; | |
1046 | FOR_EACH_EDGE (ein, ei, e->dest->preds) | |
e8038c32 | 1047 | { |
1048 | /* Scale up the frequency by REG_BR_PROB_BASE, to avoid rounding | |
9943b198 | 1049 | errors applying the probability when the frequencies are very |
1050 | small. */ | |
720cfc43 | 1051 | if (ein->probability.initialized_p ()) |
1052 | ein->count = profile_count::from_gcov_type | |
1053 | (apply_probability (ein->src->frequency * REG_BR_PROB_BASE, | |
4bb697cd | 1054 | ein->probability |
1055 | .to_reg_br_prob_base ())).guessed (); | |
720cfc43 | 1056 | else |
1057 | /* FIXME: this is hack; we should track uninitialized values. */ | |
1058 | ein->count = profile_count::zero (); | |
e8038c32 | 1059 | } |
30e432bb | 1060 | |
1061 | for (unsigned int i = 1; i < path->length (); i++) | |
1062 | { | |
1063 | edge epath = (*path)[i]->e; | |
30e432bb | 1064 | edge esucc; |
e8038c32 | 1065 | /* Scale up the frequency by REG_BR_PROB_BASE, to avoid rounding |
9943b198 | 1066 | errors applying the edge probability when the frequencies are very |
1067 | small. */ | |
db9cef39 | 1068 | epath->src->count = |
1069 | profile_count::from_gcov_type | |
1070 | (epath->src->frequency * REG_BR_PROB_BASE); | |
30e432bb | 1071 | FOR_EACH_EDGE (esucc, ei, epath->src->succs) |
db9cef39 | 1072 | esucc->count = |
1073 | esucc->src->count.apply_probability (esucc->probability); | |
30e432bb | 1074 | } |
1075 | } | |
1076 | ||
1077 | ||
1078 | /* For routines that have guessed frequencies and no profile counts, where we | |
1079 | used freqs_to_counts_path to record block and edge frequencies for paths | |
1080 | through RD, we clear the counts after completing all updates for RD. | |
1081 | The updates in ssa_fix_duplicate_block_edges are based off the count fields, | |
1082 | but the block frequencies and edge probabilities were updated as well, | |
1083 | so we can simply clear the count fields. */ | |
1084 | ||
1085 | static void | |
1086 | clear_counts_path (struct redirection_data *rd) | |
1087 | { | |
1088 | edge e = rd->incoming_edges->e; | |
1089 | vec<jump_thread_edge *> *path = THREAD_PATH (e); | |
1090 | edge ein, esucc; | |
1091 | edge_iterator ei; | |
25d2128b | 1092 | profile_count val = profile_count::uninitialized (); |
1093 | if (profile_status_for_fn (cfun) == PROFILE_READ) | |
1094 | val = profile_count::zero (); | |
1095 | ||
30e432bb | 1096 | FOR_EACH_EDGE (ein, ei, e->dest->preds) |
25d2128b | 1097 | ein->count = val; |
30e432bb | 1098 | |
1099 | /* First clear counts along original path. */ | |
1100 | for (unsigned int i = 1; i < path->length (); i++) | |
1101 | { | |
1102 | edge epath = (*path)[i]->e; | |
1103 | FOR_EACH_EDGE (esucc, ei, epath->src->succs) | |
25d2128b | 1104 | esucc->count = val; |
1105 | epath->src->count = val; | |
30e432bb | 1106 | } |
1107 | /* Also need to clear the counts along duplicated path. */ | |
1108 | for (unsigned int i = 0; i < 2; i++) | |
1109 | { | |
1110 | basic_block dup = rd->dup_blocks[i]; | |
1111 | if (!dup) | |
9943b198 | 1112 | continue; |
30e432bb | 1113 | FOR_EACH_EDGE (esucc, ei, dup->succs) |
25d2128b | 1114 | esucc->count = val; |
1115 | dup->count = val; | |
30e432bb | 1116 | } |
1117 | } | |
1118 | ||
fc54aba7 | 1119 | /* Wire up the outgoing edges from the duplicate blocks and |
30e432bb | 1120 | update any PHIs as needed. Also update the profile counts |
1121 | on the original and duplicate blocks and edges. */ | |
2b15d2ba | 1122 | void |
1123 | ssa_fix_duplicate_block_edges (struct redirection_data *rd, | |
1124 | ssa_local_info_t *local_info) | |
da81e0c5 | 1125 | { |
1b83c31b | 1126 | bool multi_incomings = (rd->incoming_edges->next != NULL); |
f2981b08 | 1127 | edge e = rd->incoming_edges->e; |
1128 | vec<jump_thread_edge *> *path = THREAD_PATH (e); | |
30e432bb | 1129 | edge elast = path->last ()->e; |
db9cef39 | 1130 | profile_count path_in_count = profile_count::zero (); |
1131 | profile_count path_out_count = profile_count::zero (); | |
30e432bb | 1132 | int path_in_freq = 0; |
1133 | ||
1134 | /* This routine updates profile counts, frequencies, and probabilities | |
1135 | incrementally. Since it is difficult to do the incremental updates | |
1136 | using frequencies/probabilities alone, for routines without profile | |
1137 | data we first take a snapshot of the existing block and edge frequencies | |
1138 | by copying them into the empty profile count fields. These counts are | |
1139 | then used to do the incremental updates, and cleared at the end of this | |
f1ce4e72 | 1140 | routine. If the function is marked as having a profile, we still check |
1141 | to see if the paths through RD are using estimated frequencies because | |
1142 | the routine had zero profile counts. */ | |
30e432bb | 1143 | bool do_freqs_to_counts = (profile_status_for_fn (cfun) != PROFILE_READ |
9943b198 | 1144 | || estimated_freqs_path (rd)); |
30e432bb | 1145 | if (do_freqs_to_counts) |
1146 | freqs_to_counts_path (rd); | |
1147 | ||
1148 | /* First determine how much profile count to move from original | |
1149 | path to the duplicate path. This is tricky in the presence of | |
1150 | a joiner (see comments for compute_path_counts), where some portion | |
1151 | of the path's counts will flow off-path from the joiner. In the | |
1152 | non-joiner case the path_in_count and path_out_count should be the | |
1153 | same. */ | |
1154 | bool has_joiner = compute_path_counts (rd, local_info, | |
9943b198 | 1155 | &path_in_count, &path_out_count, |
1156 | &path_in_freq); | |
30e432bb | 1157 | |
1158 | int cur_path_freq = path_in_freq; | |
fc54aba7 | 1159 | for (unsigned int count = 0, i = 1; i < path->length (); i++) |
1b83778e | 1160 | { |
30e432bb | 1161 | edge epath = (*path)[i]->e; |
1162 | ||
fc54aba7 | 1163 | /* If we were threading through an joiner block, then we want |
1164 | to keep its control statement and redirect an outgoing edge. | |
1165 | Else we want to remove the control statement & edges, then create | |
1166 | a new outgoing edge. In both cases we may need to update PHIs. */ | |
1167 | if ((*path)[i]->type == EDGE_COPY_SRC_JOINER_BLOCK) | |
1168 | { | |
1169 | edge victim; | |
1170 | edge e2; | |
1171 | ||
9943b198 | 1172 | gcc_assert (has_joiner); |
30e432bb | 1173 | |
fc54aba7 | 1174 | /* This updates the PHIs at the destination of the duplicate |
1b83c31b | 1175 | block. Pass 0 instead of i if we are threading a path which |
1176 | has multiple incoming edges. */ | |
1177 | update_destination_phis (local_info->bb, rd->dup_blocks[count], | |
1178 | path, multi_incomings ? 0 : i); | |
fc54aba7 | 1179 | |
1180 | /* Find the edge from the duplicate block to the block we're | |
1181 | threading through. That's the edge we want to redirect. */ | |
1182 | victim = find_edge (rd->dup_blocks[count], (*path)[i]->e->dest); | |
1183 | ||
1184 | /* If there are no remaining blocks on the path to duplicate, | |
1185 | then redirect VICTIM to the final destination of the jump | |
1186 | threading path. */ | |
1187 | if (!any_remaining_duplicated_blocks (path, i)) | |
1188 | { | |
30e432bb | 1189 | e2 = redirect_edge_and_branch (victim, elast->dest); |
fc54aba7 | 1190 | /* If we redirected the edge, then we need to copy PHI arguments |
559685be | 1191 | at the target. If the edge already existed (e2 != victim |
fc54aba7 | 1192 | case), then the PHIs in the target already have the correct |
1193 | arguments. */ | |
1194 | if (e2 == victim) | |
30e432bb | 1195 | copy_phi_args (e2->dest, elast, e2, |
1b83c31b | 1196 | path, multi_incomings ? 0 : i); |
fc54aba7 | 1197 | } |
1198 | else | |
1199 | { | |
1200 | /* Redirect VICTIM to the next duplicated block in the path. */ | |
1201 | e2 = redirect_edge_and_branch (victim, rd->dup_blocks[count + 1]); | |
1202 | ||
1203 | /* We need to update the PHIs in the next duplicated block. We | |
1204 | want the new PHI args to have the same value as they had | |
1205 | in the source of the next duplicate block. | |
1206 | ||
1207 | Thus, we need to know which edge we traversed into the | |
1208 | source of the duplicate. Furthermore, we may have | |
1209 | traversed many edges to reach the source of the duplicate. | |
1210 | ||
1211 | Walk through the path starting at element I until we | |
1212 | hit an edge marked with EDGE_COPY_SRC_BLOCK. We want | |
1213 | the edge from the prior element. */ | |
1214 | for (unsigned int j = i + 1; j < path->length (); j++) | |
1215 | { | |
1216 | if ((*path)[j]->type == EDGE_COPY_SRC_BLOCK) | |
1217 | { | |
1218 | copy_phi_arg_into_existing_phi ((*path)[j - 1]->e, e2); | |
1219 | break; | |
1220 | } | |
1221 | } | |
1222 | } | |
30e432bb | 1223 | |
1224 | /* Update the counts and frequency of both the original block | |
1225 | and path edge, and the duplicates. The path duplicate's | |
1226 | incoming count and frequency are the totals for all edges | |
1227 | incoming to this jump threading path computed earlier. | |
1228 | And we know that the duplicated path will have path_out_count | |
1229 | flowing out of it (i.e. along the duplicated path out of the | |
1230 | duplicated joiner). */ | |
1231 | update_profile (epath, e2, path_in_count, path_out_count, | |
1232 | path_in_freq); | |
1233 | ||
1234 | /* Next we need to update the counts of the original and duplicated | |
1235 | edges from the joiner that go off path. */ | |
1236 | update_joiner_offpath_counts (epath, e2->src, path_in_count, | |
9943b198 | 1237 | path_out_count); |
30e432bb | 1238 | |
1239 | /* Finally, we need to set the probabilities on the duplicated | |
1240 | edges out of the duplicated joiner (e2->src). The probabilities | |
1241 | along the original path will all be updated below after we finish | |
1242 | processing the whole path. */ | |
1243 | recompute_probabilities (e2->src); | |
1244 | ||
1245 | /* Record the frequency flowing to the downstream duplicated | |
1246 | path blocks. */ | |
1247 | cur_path_freq = EDGE_FREQUENCY (e2); | |
fc54aba7 | 1248 | } |
1249 | else if ((*path)[i]->type == EDGE_COPY_SRC_BLOCK) | |
1250 | { | |
1251 | remove_ctrl_stmt_and_useless_edges (rd->dup_blocks[count], NULL); | |
1b83c31b | 1252 | create_edge_and_update_destination_phis (rd, rd->dup_blocks[count], |
1253 | multi_incomings ? 0 : i); | |
fc54aba7 | 1254 | if (count == 1) |
1255 | single_succ_edge (rd->dup_blocks[1])->aux = NULL; | |
30e432bb | 1256 | |
1257 | /* Update the counts and frequency of both the original block | |
1258 | and path edge, and the duplicates. Since we are now after | |
1259 | any joiner that may have existed on the path, the count | |
1260 | flowing along the duplicated threaded path is path_out_count. | |
1261 | If we didn't have a joiner, then cur_path_freq was the sum | |
1262 | of the total frequencies along all incoming edges to the | |
1263 | thread path (path_in_freq). If we had a joiner, it would have | |
1264 | been updated at the end of that handling to the edge frequency | |
1265 | along the duplicated joiner path edge. */ | |
1266 | update_profile (epath, EDGE_SUCC (rd->dup_blocks[count], 0), | |
1267 | path_out_count, path_out_count, | |
1268 | cur_path_freq); | |
fc54aba7 | 1269 | } |
30e432bb | 1270 | else |
9943b198 | 1271 | { |
30e432bb | 1272 | /* No copy case. In this case we don't have an equivalent block |
1273 | on the duplicated thread path to update, but we do need | |
1274 | to remove the portion of the counts/freqs that were moved | |
1275 | to the duplicated path from the counts/freqs flowing through | |
1276 | this block on the original path. Since all the no-copy edges | |
1277 | are after any joiner, the removed count is the same as | |
1278 | path_out_count. | |
1279 | ||
1280 | If we didn't have a joiner, then cur_path_freq was the sum | |
1281 | of the total frequencies along all incoming edges to the | |
1282 | thread path (path_in_freq). If we had a joiner, it would have | |
1283 | been updated at the end of that handling to the edge frequency | |
1284 | along the duplicated joiner path edge. */ | |
31e5d72d | 1285 | update_profile (epath, NULL, path_out_count, path_out_count, |
1286 | cur_path_freq); | |
30e432bb | 1287 | } |
1288 | ||
1289 | /* Increment the index into the duplicated path when we processed | |
9943b198 | 1290 | a duplicated block. */ |
30e432bb | 1291 | if ((*path)[i]->type == EDGE_COPY_SRC_JOINER_BLOCK |
9943b198 | 1292 | || (*path)[i]->type == EDGE_COPY_SRC_BLOCK) |
31e5d72d | 1293 | { |
30e432bb | 1294 | count++; |
31e5d72d | 1295 | } |
30e432bb | 1296 | } |
1297 | ||
1298 | /* Now walk orig blocks and update their probabilities, since the | |
1299 | counts and freqs should be updated properly by above loop. */ | |
1300 | for (unsigned int i = 1; i < path->length (); i++) | |
1301 | { | |
1302 | edge epath = (*path)[i]->e; | |
1303 | recompute_probabilities (epath->src); | |
778182c1 | 1304 | } |
30e432bb | 1305 | |
1306 | /* Done with all profile and frequency updates, clear counts if they | |
1307 | were copied. */ | |
1308 | if (do_freqs_to_counts) | |
1309 | clear_counts_path (rd); | |
778182c1 | 1310 | } |
fc54aba7 | 1311 | |
778182c1 | 1312 | /* Hash table traversal callback routine to create duplicate blocks. */ |
1313 | ||
2b15d2ba | 1314 | int |
1315 | ssa_create_duplicates (struct redirection_data **slot, | |
1316 | ssa_local_info_t *local_info) | |
778182c1 | 1317 | { |
2b15d2ba | 1318 | struct redirection_data *rd = *slot; |
778182c1 | 1319 | |
11af02d8 | 1320 | /* The second duplicated block in a jump threading path is specific |
1b83778e | 1321 | to the path. So it gets stored in RD rather than in LOCAL_DATA. |
559685be | 1322 | |
11af02d8 | 1323 | Each time we're called, we have to look through the path and see |
1b83778e | 1324 | if a second block needs to be duplicated. |
11af02d8 | 1325 | |
1326 | Note the search starts with the third edge on the path. The first | |
1327 | edge is the incoming edge, the second edge always has its source | |
1328 | duplicated. Thus we start our search with the third edge. */ | |
a7ee7309 | 1329 | vec<jump_thread_edge *> *path = rd->path; |
11af02d8 | 1330 | for (unsigned int i = 2; i < path->length (); i++) |
1331 | { | |
1332 | if ((*path)[i]->type == EDGE_COPY_SRC_BLOCK | |
1333 | || (*path)[i]->type == EDGE_COPY_SRC_JOINER_BLOCK) | |
1334 | { | |
30e432bb | 1335 | create_block_for_threading ((*path)[i]->e->src, rd, 1, |
a7ee7309 | 1336 | &local_info->duplicate_blocks); |
11af02d8 | 1337 | break; |
1338 | } | |
1339 | } | |
1b83778e | 1340 | |
778182c1 | 1341 | /* Create a template block if we have not done so already. Otherwise |
1342 | use the template to create a new block. */ | |
1343 | if (local_info->template_block == NULL) | |
1344 | { | |
30e432bb | 1345 | create_block_for_threading ((*path)[1]->e->src, rd, 0, |
a7ee7309 | 1346 | &local_info->duplicate_blocks); |
11af02d8 | 1347 | local_info->template_block = rd->dup_blocks[0]; |
778182c1 | 1348 | |
1349 | /* We do not create any outgoing edges for the template. We will | |
1350 | take care of that in a later traversal. That way we do not | |
1351 | create edges that are going to just be deleted. */ | |
1352 | } | |
1353 | else | |
1354 | { | |
30e432bb | 1355 | create_block_for_threading (local_info->template_block, rd, 0, |
a7ee7309 | 1356 | &local_info->duplicate_blocks); |
778182c1 | 1357 | |
1358 | /* Go ahead and wire up outgoing edges and update PHIs for the duplicate | |
da81e0c5 | 1359 | block. */ |
2b15d2ba | 1360 | ssa_fix_duplicate_block_edges (rd, local_info); |
778182c1 | 1361 | } |
1362 | ||
1363 | /* Keep walking the hash table. */ | |
1364 | return 1; | |
1365 | } | |
1366 | ||
1367 | /* We did not create any outgoing edges for the template block during | |
1368 | block creation. This hash table traversal callback creates the | |
1369 | outgoing edge for the template block. */ | |
1370 | ||
2b15d2ba | 1371 | inline int |
1372 | ssa_fixup_template_block (struct redirection_data **slot, | |
1373 | ssa_local_info_t *local_info) | |
778182c1 | 1374 | { |
2b15d2ba | 1375 | struct redirection_data *rd = *slot; |
778182c1 | 1376 | |
da81e0c5 | 1377 | /* If this is the template block halt the traversal after updating |
1378 | it appropriately. | |
1379 | ||
1380 | If we were threading through an joiner block, then we want | |
1381 | to keep its control statement and redirect an outgoing edge. | |
1382 | Else we want to remove the control statement & edges, then create | |
1383 | a new outgoing edge. In both cases we may need to update PHIs. */ | |
11af02d8 | 1384 | if (rd->dup_blocks[0] && rd->dup_blocks[0] == local_info->template_block) |
778182c1 | 1385 | { |
2b15d2ba | 1386 | ssa_fix_duplicate_block_edges (rd, local_info); |
778182c1 | 1387 | return 0; |
1388 | } | |
1389 | ||
1390 | return 1; | |
1391 | } | |
1392 | ||
1393 | /* Hash table traversal callback to redirect each incoming edge | |
1394 | associated with this hash table element to its new destination. */ | |
1395 | ||
2b15d2ba | 1396 | int |
1397 | ssa_redirect_edges (struct redirection_data **slot, | |
1398 | ssa_local_info_t *local_info) | |
778182c1 | 1399 | { |
2b15d2ba | 1400 | struct redirection_data *rd = *slot; |
778182c1 | 1401 | struct el *next, *el; |
1402 | ||
47ae02b7 | 1403 | /* Walk over all the incoming edges associated with this hash table |
1404 | entry. */ | |
778182c1 | 1405 | for (el = rd->incoming_edges; el; el = next) |
1406 | { | |
1407 | edge e = el->e; | |
f2981b08 | 1408 | vec<jump_thread_edge *> *path = THREAD_PATH (e); |
778182c1 | 1409 | |
1410 | /* Go ahead and free this element from the list. Doing this now | |
1411 | avoids the need for another list walk when we destroy the hash | |
1412 | table. */ | |
1413 | next = el->next; | |
1414 | free (el); | |
1415 | ||
5236b8bb | 1416 | thread_stats.num_threaded_edges++; |
1417 | ||
11af02d8 | 1418 | if (rd->dup_blocks[0]) |
778182c1 | 1419 | { |
1420 | edge e2; | |
1421 | ||
1422 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1423 | fprintf (dump_file, " Threaded jump %d --> %d to %d\n", | |
11af02d8 | 1424 | e->src->index, e->dest->index, rd->dup_blocks[0]->index); |
778182c1 | 1425 | |
353f9f16 | 1426 | /* Redirect the incoming edge (possibly to the joiner block) to the |
1427 | appropriate duplicate block. */ | |
11af02d8 | 1428 | e2 = redirect_edge_and_branch (e, rd->dup_blocks[0]); |
7e0311ae | 1429 | gcc_assert (e == e2); |
778182c1 | 1430 | flush_pending_stmts (e2); |
778182c1 | 1431 | } |
eb31063a | 1432 | |
1433 | /* Go ahead and clear E->aux. It's not needed anymore and failure | |
559685be | 1434 | to clear it will cause all kinds of unpleasant problems later. */ |
6d1fdbf9 | 1435 | delete_jump_thread_path (path); |
eb31063a | 1436 | e->aux = NULL; |
1437 | ||
778182c1 | 1438 | } |
388d1fc1 | 1439 | |
1440 | /* Indicate that we actually threaded one or more jumps. */ | |
1441 | if (rd->incoming_edges) | |
1442 | local_info->jumps_threaded = true; | |
1443 | ||
778182c1 | 1444 | return 1; |
1445 | } | |
1446 | ||
aed95130 | 1447 | /* Return true if this block has no executable statements other than |
1448 | a simple ctrl flow instruction. When the number of outgoing edges | |
1449 | is one, this is equivalent to a "forwarder" block. */ | |
1450 | ||
1451 | static bool | |
47aaf6e6 | 1452 | redirection_block_p (basic_block bb) |
aed95130 | 1453 | { |
75a70cf9 | 1454 | gimple_stmt_iterator gsi; |
aed95130 | 1455 | |
1456 | /* Advance to the first executable statement. */ | |
75a70cf9 | 1457 | gsi = gsi_start_bb (bb); |
1458 | while (!gsi_end_p (gsi) | |
559685be | 1459 | && (gimple_code (gsi_stmt (gsi)) == GIMPLE_LABEL |
9845d120 | 1460 | || is_gimple_debug (gsi_stmt (gsi)) |
c7566ddf | 1461 | || gimple_nop_p (gsi_stmt (gsi)) |
581c1a3c | 1462 | || gimple_clobber_p (gsi_stmt (gsi)))) |
75a70cf9 | 1463 | gsi_next (&gsi); |
48e1416a | 1464 | |
aed95130 | 1465 | /* Check if this is an empty block. */ |
75a70cf9 | 1466 | if (gsi_end_p (gsi)) |
aed95130 | 1467 | return true; |
1468 | ||
1469 | /* Test that we've reached the terminating control statement. */ | |
75a70cf9 | 1470 | return gsi_stmt (gsi) |
559685be | 1471 | && (gimple_code (gsi_stmt (gsi)) == GIMPLE_COND |
1472 | || gimple_code (gsi_stmt (gsi)) == GIMPLE_GOTO | |
1473 | || gimple_code (gsi_stmt (gsi)) == GIMPLE_SWITCH); | |
aed95130 | 1474 | } |
1475 | ||
a8046f60 | 1476 | /* BB is a block which ends with a COND_EXPR or SWITCH_EXPR and when BB |
1477 | is reached via one or more specific incoming edges, we know which | |
1478 | outgoing edge from BB will be traversed. | |
1479 | ||
778182c1 | 1480 | We want to redirect those incoming edges to the target of the |
a8046f60 | 1481 | appropriate outgoing edge. Doing so avoids a conditional branch |
1482 | and may expose new optimization opportunities. Note that we have | |
1483 | to update dominator tree and SSA graph after such changes. | |
1484 | ||
597ff315 | 1485 | The key to keeping the SSA graph update manageable is to duplicate |
91275768 | 1486 | the side effects occurring in BB so that those side effects still |
a8046f60 | 1487 | occur on the paths which bypass BB after redirecting edges. |
1488 | ||
1489 | We accomplish this by creating duplicates of BB and arranging for | |
1490 | the duplicates to unconditionally pass control to one specific | |
1491 | successor of BB. We then revector the incoming edges into BB to | |
1492 | the appropriate duplicate of BB. | |
1493 | ||
7e0311ae | 1494 | If NOLOOP_ONLY is true, we only perform the threading as long as it |
1b83778e | 1495 | does not affect the structure of the loops in a nontrivial way. |
ed4feca1 | 1496 | |
1497 | If JOINERS is true, then thread through joiner blocks as well. */ | |
a8046f60 | 1498 | |
388d1fc1 | 1499 | static bool |
ed4feca1 | 1500 | thread_block_1 (basic_block bb, bool noloop_only, bool joiners) |
a8046f60 | 1501 | { |
1502 | /* E is an incoming edge into BB that we may or may not want to | |
1503 | redirect to a duplicate of BB. */ | |
7e0311ae | 1504 | edge e, e2; |
cd665a06 | 1505 | edge_iterator ei; |
2b15d2ba | 1506 | ssa_local_info_t local_info; |
388d1fc1 | 1507 | |
30e432bb | 1508 | local_info.duplicate_blocks = BITMAP_ALLOC (NULL); |
d07cbccc | 1509 | local_info.need_profile_correction = false; |
30e432bb | 1510 | |
778182c1 | 1511 | /* To avoid scanning a linear array for the element we need we instead |
c5d4a10b | 1512 | use a hash table. For normal code there should be no noticeable |
778182c1 | 1513 | difference. However, if we have a block with a large number of |
1514 | incoming and outgoing edges such linear searches can get expensive. */ | |
c1f445d2 | 1515 | redirection_data |
1516 | = new hash_table<struct redirection_data> (EDGE_COUNT (bb->succs)); | |
778182c1 | 1517 | |
1518 | /* Record each unique threaded destination into a hash table for | |
1519 | efficient lookups. */ | |
d07cbccc | 1520 | edge last = NULL; |
cd665a06 | 1521 | FOR_EACH_EDGE (e, ei, bb->preds) |
a8046f60 | 1522 | { |
eb31063a | 1523 | if (e->aux == NULL) |
1524 | continue; | |
1525 | ||
f2981b08 | 1526 | vec<jump_thread_edge *> *path = THREAD_PATH (e); |
ed4feca1 | 1527 | |
1528 | if (((*path)[1]->type == EDGE_COPY_SRC_JOINER_BLOCK && !joiners) | |
1529 | || ((*path)[1]->type == EDGE_COPY_SRC_BLOCK && joiners)) | |
1530 | continue; | |
1531 | ||
f2981b08 | 1532 | e2 = path->last ()->e; |
e2b72d6c | 1533 | if (!e2 || noloop_only) |
1534 | { | |
7e0311ae | 1535 | /* If NOLOOP_ONLY is true, we only allow threading through the |
559685be | 1536 | header of a loop to exit edges. */ |
e2b72d6c | 1537 | |
559685be | 1538 | /* One case occurs when there was loop header buried in a jump |
1539 | threading path that crosses loop boundaries. We do not try | |
1540 | and thread this elsewhere, so just cancel the jump threading | |
1541 | request by clearing the AUX field now. */ | |
9f3abcb8 | 1542 | if (bb->loop_father != e2->src->loop_father |
1543 | && !loop_exit_edge_p (e2->src->loop_father, e2)) | |
e2b72d6c | 1544 | { |
1545 | /* Since this case is not handled by our special code | |
1546 | to thread through a loop header, we must explicitly | |
1547 | cancel the threading request here. */ | |
6d1fdbf9 | 1548 | delete_jump_thread_path (path); |
e2b72d6c | 1549 | e->aux = NULL; |
1550 | continue; | |
1551 | } | |
559685be | 1552 | |
1553 | /* Another case occurs when trying to thread through our | |
ab596744 | 1554 | own loop header, possibly from inside the loop. We will |
1555 | thread these later. */ | |
559685be | 1556 | unsigned int i; |
1557 | for (i = 1; i < path->length (); i++) | |
1558 | { | |
1559 | if ((*path)[i]->e->src == bb->loop_father->header | |
1560 | && (!loop_exit_edge_p (bb->loop_father, e2) | |
1561 | || (*path)[1]->type == EDGE_COPY_SRC_JOINER_BLOCK)) | |
ab596744 | 1562 | break; |
559685be | 1563 | } |
1564 | ||
1565 | if (i != path->length ()) | |
1566 | continue; | |
e2b72d6c | 1567 | } |
778182c1 | 1568 | |
7e0311ae | 1569 | /* Insert the outgoing edge into the hash table if it is not |
1570 | already in the hash table. */ | |
da81e0c5 | 1571 | lookup_redirection_data (e, INSERT); |
d07cbccc | 1572 | |
1573 | /* When we have thread paths through a common joiner with different | |
1574 | final destinations, then we may need corrections to deal with | |
1575 | profile insanities. See the big comment before compute_path_counts. */ | |
1576 | if ((*path)[1]->type == EDGE_COPY_SRC_JOINER_BLOCK) | |
1577 | { | |
1578 | if (!last) | |
1579 | last = e2; | |
1580 | else if (e2 != last) | |
1581 | local_info.need_profile_correction = true; | |
1582 | } | |
a8046f60 | 1583 | } |
1584 | ||
3f9439d7 | 1585 | /* We do not update dominance info. */ |
1586 | free_dominance_info (CDI_DOMINATORS); | |
1587 | ||
d906930c | 1588 | /* We know we only thread through the loop header to loop exits. |
1589 | Let the basic block duplication hook know we are not creating | |
1590 | a multiple entry loop. */ | |
1591 | if (noloop_only | |
1592 | && bb == bb->loop_father->header) | |
1593 | set_loop_copy (bb->loop_father, loop_outer (bb->loop_father)); | |
1594 | ||
778182c1 | 1595 | /* Now create duplicates of BB. |
f582bb6c | 1596 | |
1597 | Note that for a block with a high outgoing degree we can waste | |
1598 | a lot of time and memory creating and destroying useless edges. | |
1599 | ||
1600 | So we first duplicate BB and remove the control structure at the | |
1601 | tail of the duplicate as well as all outgoing edges from the | |
1602 | duplicate. We then use that duplicate block as a template for | |
1603 | the rest of the duplicates. */ | |
778182c1 | 1604 | local_info.template_block = NULL; |
1605 | local_info.bb = bb; | |
388d1fc1 | 1606 | local_info.jumps_threaded = false; |
c1f445d2 | 1607 | redirection_data->traverse <ssa_local_info_t *, ssa_create_duplicates> |
2b15d2ba | 1608 | (&local_info); |
f582bb6c | 1609 | |
778182c1 | 1610 | /* The template does not have an outgoing edge. Create that outgoing |
1611 | edge and update PHI nodes as the edge's target as necessary. | |
f582bb6c | 1612 | |
778182c1 | 1613 | We do this after creating all the duplicates to avoid creating |
1614 | unnecessary edges. */ | |
c1f445d2 | 1615 | redirection_data->traverse <ssa_local_info_t *, ssa_fixup_template_block> |
2b15d2ba | 1616 | (&local_info); |
f582bb6c | 1617 | |
778182c1 | 1618 | /* The hash table traversals above created the duplicate blocks (and the |
1619 | statements within the duplicate blocks). This loop creates PHI nodes for | |
1620 | the duplicated blocks and redirects the incoming edges into BB to reach | |
1621 | the duplicates of BB. */ | |
c1f445d2 | 1622 | redirection_data->traverse <ssa_local_info_t *, ssa_redirect_edges> |
2b15d2ba | 1623 | (&local_info); |
a8046f60 | 1624 | |
a3d0fd80 | 1625 | /* Done with this block. Clear REDIRECTION_DATA. */ |
c1f445d2 | 1626 | delete redirection_data; |
1627 | redirection_data = NULL; | |
388d1fc1 | 1628 | |
d906930c | 1629 | if (noloop_only |
1630 | && bb == bb->loop_father->header) | |
1631 | set_loop_copy (bb->loop_father, NULL); | |
1632 | ||
30e432bb | 1633 | BITMAP_FREE (local_info.duplicate_blocks); |
1634 | local_info.duplicate_blocks = NULL; | |
1635 | ||
388d1fc1 | 1636 | /* Indicate to our caller whether or not any jumps were threaded. */ |
1637 | return local_info.jumps_threaded; | |
a8046f60 | 1638 | } |
1639 | ||
ed4feca1 | 1640 | /* Wrapper for thread_block_1 so that we can first handle jump |
1641 | thread paths which do not involve copying joiner blocks, then | |
1642 | handle jump thread paths which have joiner blocks. | |
1643 | ||
1644 | By doing things this way we can be as aggressive as possible and | |
1645 | not worry that copying a joiner block will create a jump threading | |
1646 | opportunity. */ | |
1b83778e | 1647 | |
ed4feca1 | 1648 | static bool |
1649 | thread_block (basic_block bb, bool noloop_only) | |
1650 | { | |
1651 | bool retval; | |
1652 | retval = thread_block_1 (bb, noloop_only, false); | |
1653 | retval |= thread_block_1 (bb, noloop_only, true); | |
1654 | return retval; | |
1655 | } | |
1656 | ||
7e0311ae | 1657 | /* Callback for dfs_enumerate_from. Returns true if BB is different |
1658 | from STOP and DBDS_CE_STOP. */ | |
1659 | ||
1660 | static basic_block dbds_ce_stop; | |
1661 | static bool | |
7ecb5bb2 | 1662 | dbds_continue_enumeration_p (const_basic_block bb, const void *stop) |
7e0311ae | 1663 | { |
7ecb5bb2 | 1664 | return (bb != (const_basic_block) stop |
7e0311ae | 1665 | && bb != dbds_ce_stop); |
1666 | } | |
1667 | ||
1668 | /* Evaluates the dominance relationship of latch of the LOOP and BB, and | |
1669 | returns the state. */ | |
1670 | ||
1671 | enum bb_dom_status | |
7e0311ae | 1672 | determine_bb_domination_status (struct loop *loop, basic_block bb) |
1673 | { | |
1674 | basic_block *bblocks; | |
1675 | unsigned nblocks, i; | |
1676 | bool bb_reachable = false; | |
1677 | edge_iterator ei; | |
1678 | edge e; | |
1679 | ||
42b013bc | 1680 | /* This function assumes BB is a successor of LOOP->header. |
1681 | If that is not the case return DOMST_NONDOMINATING which | |
1682 | is always safe. */ | |
7e0311ae | 1683 | { |
1684 | bool ok = false; | |
1685 | ||
1686 | FOR_EACH_EDGE (e, ei, bb->preds) | |
1687 | { | |
1688 | if (e->src == loop->header) | |
1689 | { | |
1690 | ok = true; | |
1691 | break; | |
1692 | } | |
1693 | } | |
1694 | ||
42b013bc | 1695 | if (!ok) |
1696 | return DOMST_NONDOMINATING; | |
7e0311ae | 1697 | } |
7e0311ae | 1698 | |
1699 | if (bb == loop->latch) | |
1700 | return DOMST_DOMINATING; | |
1701 | ||
1702 | /* Check that BB dominates LOOP->latch, and that it is back-reachable | |
1703 | from it. */ | |
1704 | ||
1705 | bblocks = XCNEWVEC (basic_block, loop->num_nodes); | |
1706 | dbds_ce_stop = loop->header; | |
1707 | nblocks = dfs_enumerate_from (loop->latch, 1, dbds_continue_enumeration_p, | |
1708 | bblocks, loop->num_nodes, bb); | |
1709 | for (i = 0; i < nblocks; i++) | |
1710 | FOR_EACH_EDGE (e, ei, bblocks[i]->preds) | |
1711 | { | |
1712 | if (e->src == loop->header) | |
1713 | { | |
1714 | free (bblocks); | |
1715 | return DOMST_NONDOMINATING; | |
1716 | } | |
1717 | if (e->src == bb) | |
1718 | bb_reachable = true; | |
1719 | } | |
1720 | ||
1721 | free (bblocks); | |
1722 | return (bb_reachable ? DOMST_DOMINATING : DOMST_LOOP_BROKEN); | |
1723 | } | |
1724 | ||
1725 | /* Thread jumps through the header of LOOP. Returns true if cfg changes. | |
1726 | If MAY_PEEL_LOOP_HEADERS is false, we avoid threading from entry edges | |
1727 | to the inside of the loop. */ | |
1728 | ||
1729 | static bool | |
1730 | thread_through_loop_header (struct loop *loop, bool may_peel_loop_headers) | |
1731 | { | |
1732 | basic_block header = loop->header; | |
1733 | edge e, tgt_edge, latch = loop_latch_edge (loop); | |
1734 | edge_iterator ei; | |
1735 | basic_block tgt_bb, atgt_bb; | |
1736 | enum bb_dom_status domst; | |
1737 | ||
1738 | /* We have already threaded through headers to exits, so all the threading | |
1739 | requests now are to the inside of the loop. We need to avoid creating | |
1740 | irreducible regions (i.e., loops with more than one entry block), and | |
1741 | also loop with several latch edges, or new subloops of the loop (although | |
1742 | there are cases where it might be appropriate, it is difficult to decide, | |
1743 | and doing it wrongly may confuse other optimizers). | |
1744 | ||
1745 | We could handle more general cases here. However, the intention is to | |
1746 | preserve some information about the loop, which is impossible if its | |
1747 | structure changes significantly, in a way that is not well understood. | |
1748 | Thus we only handle few important special cases, in which also updating | |
1749 | of the loop-carried information should be feasible: | |
1750 | ||
1751 | 1) Propagation of latch edge to a block that dominates the latch block | |
1752 | of a loop. This aims to handle the following idiom: | |
1753 | ||
1754 | first = 1; | |
1755 | while (1) | |
1756 | { | |
1757 | if (first) | |
1758 | initialize; | |
1759 | first = 0; | |
1760 | body; | |
1761 | } | |
1762 | ||
1763 | After threading the latch edge, this becomes | |
1764 | ||
1765 | first = 1; | |
1766 | if (first) | |
1767 | initialize; | |
1768 | while (1) | |
1769 | { | |
1770 | first = 0; | |
1771 | body; | |
1772 | } | |
1773 | ||
1774 | The original header of the loop is moved out of it, and we may thread | |
1775 | the remaining edges through it without further constraints. | |
1776 | ||
1777 | 2) All entry edges are propagated to a single basic block that dominates | |
1778 | the latch block of the loop. This aims to handle the following idiom | |
1779 | (normally created for "for" loops): | |
1780 | ||
1781 | i = 0; | |
1782 | while (1) | |
1783 | { | |
1784 | if (i >= 100) | |
1785 | break; | |
1786 | body; | |
1787 | i++; | |
1788 | } | |
1789 | ||
1790 | This becomes | |
1791 | ||
1792 | i = 0; | |
1793 | while (1) | |
1794 | { | |
1795 | body; | |
1796 | i++; | |
1797 | if (i >= 100) | |
1798 | break; | |
1799 | } | |
1800 | */ | |
1801 | ||
1802 | /* Threading through the header won't improve the code if the header has just | |
1803 | one successor. */ | |
1804 | if (single_succ_p (header)) | |
1805 | goto fail; | |
1806 | ||
d29932c9 | 1807 | if (!may_peel_loop_headers && !redirection_block_p (loop->header)) |
7e0311ae | 1808 | goto fail; |
1809 | else | |
1810 | { | |
1811 | tgt_bb = NULL; | |
1812 | tgt_edge = NULL; | |
1813 | FOR_EACH_EDGE (e, ei, header->preds) | |
1814 | { | |
1815 | if (!e->aux) | |
1816 | { | |
1817 | if (e == latch) | |
1818 | continue; | |
1819 | ||
1820 | /* If latch is not threaded, and there is a header | |
1821 | edge that is not threaded, we would create loop | |
1822 | with multiple entries. */ | |
1823 | goto fail; | |
1824 | } | |
1825 | ||
f2981b08 | 1826 | vec<jump_thread_edge *> *path = THREAD_PATH (e); |
1827 | ||
1828 | if ((*path)[1]->type == EDGE_COPY_SRC_JOINER_BLOCK) | |
da81e0c5 | 1829 | goto fail; |
f2981b08 | 1830 | tgt_edge = (*path)[1]->e; |
7e0311ae | 1831 | atgt_bb = tgt_edge->dest; |
1832 | if (!tgt_bb) | |
1833 | tgt_bb = atgt_bb; | |
1834 | /* Two targets of threading would make us create loop | |
1835 | with multiple entries. */ | |
1836 | else if (tgt_bb != atgt_bb) | |
1837 | goto fail; | |
1838 | } | |
1839 | ||
1840 | if (!tgt_bb) | |
1841 | { | |
1842 | /* There are no threading requests. */ | |
1843 | return false; | |
1844 | } | |
1845 | ||
1846 | /* Redirecting to empty loop latch is useless. */ | |
1847 | if (tgt_bb == loop->latch | |
1848 | && empty_block_p (loop->latch)) | |
1849 | goto fail; | |
1850 | } | |
1851 | ||
1852 | /* The target block must dominate the loop latch, otherwise we would be | |
1853 | creating a subloop. */ | |
1854 | domst = determine_bb_domination_status (loop, tgt_bb); | |
1855 | if (domst == DOMST_NONDOMINATING) | |
1856 | goto fail; | |
1857 | if (domst == DOMST_LOOP_BROKEN) | |
1858 | { | |
1859 | /* If the loop ceased to exist, mark it as such, and thread through its | |
1860 | original header. */ | |
d25159cc | 1861 | mark_loop_for_removal (loop); |
7e0311ae | 1862 | return thread_block (header, false); |
1863 | } | |
1864 | ||
1865 | if (tgt_bb->loop_father->header == tgt_bb) | |
1866 | { | |
1867 | /* If the target of the threading is a header of a subloop, we need | |
1868 | to create a preheader for it, so that the headers of the two loops | |
1869 | do not merge. */ | |
1870 | if (EDGE_COUNT (tgt_bb->preds) > 2) | |
1871 | { | |
1872 | tgt_bb = create_preheader (tgt_bb->loop_father, 0); | |
1873 | gcc_assert (tgt_bb != NULL); | |
1874 | } | |
1875 | else | |
1876 | tgt_bb = split_edge (tgt_edge); | |
1877 | } | |
48e1416a | 1878 | |
d29932c9 | 1879 | basic_block new_preheader; |
6eb99d8a | 1880 | |
d29932c9 | 1881 | /* Now consider the case entry edges are redirected to the new entry |
1882 | block. Remember one entry edge, so that we can find the new | |
1883 | preheader (its destination after threading). */ | |
1884 | FOR_EACH_EDGE (e, ei, header->preds) | |
7e0311ae | 1885 | { |
d29932c9 | 1886 | if (e->aux) |
1887 | break; | |
1888 | } | |
7e0311ae | 1889 | |
d29932c9 | 1890 | /* The duplicate of the header is the new preheader of the loop. Ensure |
1891 | that it is placed correctly in the loop hierarchy. */ | |
1892 | set_loop_copy (loop, loop_outer (loop)); | |
7e0311ae | 1893 | |
d29932c9 | 1894 | thread_block (header, false); |
1895 | set_loop_copy (loop, NULL); | |
1896 | new_preheader = e->dest; | |
48e1416a | 1897 | |
d29932c9 | 1898 | /* Create the new latch block. This is always necessary, as the latch |
1899 | must have only a single successor, but the original header had at | |
1900 | least two successors. */ | |
1901 | loop->latch = NULL; | |
1902 | mfb_kj_edge = single_succ_edge (new_preheader); | |
1903 | loop->header = mfb_kj_edge->dest; | |
1904 | latch = make_forwarder_block (tgt_bb, mfb_keep_just, NULL); | |
1905 | loop->header = latch->dest; | |
1906 | loop->latch = latch->src; | |
7e0311ae | 1907 | return true; |
1908 | ||
1909 | fail: | |
1910 | /* We failed to thread anything. Cancel the requests. */ | |
1911 | FOR_EACH_EDGE (e, ei, header->preds) | |
1912 | { | |
f2981b08 | 1913 | vec<jump_thread_edge *> *path = THREAD_PATH (e); |
1914 | ||
1915 | if (path) | |
1916 | { | |
6d1fdbf9 | 1917 | delete_jump_thread_path (path); |
f2981b08 | 1918 | e->aux = NULL; |
1919 | } | |
7e0311ae | 1920 | } |
1921 | return false; | |
1922 | } | |
1923 | ||
b99a7d6d | 1924 | /* E1 and E2 are edges into the same basic block. Return TRUE if the |
1925 | PHI arguments associated with those edges are equal or there are no | |
1926 | PHI arguments, otherwise return FALSE. */ | |
1927 | ||
1928 | static bool | |
1929 | phi_args_equal_on_edges (edge e1, edge e2) | |
1930 | { | |
1a91d914 | 1931 | gphi_iterator gsi; |
b99a7d6d | 1932 | int indx1 = e1->dest_idx; |
1933 | int indx2 = e2->dest_idx; | |
1934 | ||
1935 | for (gsi = gsi_start_phis (e1->dest); !gsi_end_p (gsi); gsi_next (&gsi)) | |
1936 | { | |
1a91d914 | 1937 | gphi *phi = gsi.phi (); |
b99a7d6d | 1938 | |
1939 | if (!operand_equal_p (gimple_phi_arg_def (phi, indx1), | |
1940 | gimple_phi_arg_def (phi, indx2), 0)) | |
1941 | return false; | |
1942 | } | |
1943 | return true; | |
1944 | } | |
1945 | ||
3cebc9d2 | 1946 | /* Walk through the registered jump threads and convert them into a |
334ec2d8 | 1947 | form convenient for this pass. |
3cebc9d2 | 1948 | |
1949 | Any block which has incoming edges threaded to outgoing edges | |
1950 | will have its entry in THREADED_BLOCK set. | |
a8046f60 | 1951 | |
3cebc9d2 | 1952 | Any threaded edge will have its new outgoing edge stored in the |
1953 | original edge's AUX field. | |
a8046f60 | 1954 | |
3cebc9d2 | 1955 | This form avoids the need to walk all the edges in the CFG to |
1956 | discover blocks which need processing and avoids unnecessary | |
1957 | hash table lookups to map from threaded edge to new target. */ | |
a8046f60 | 1958 | |
3cebc9d2 | 1959 | static void |
1960 | mark_threaded_blocks (bitmap threaded_blocks) | |
1961 | { | |
1962 | unsigned int i; | |
7e0311ae | 1963 | bitmap_iterator bi; |
035def86 | 1964 | auto_bitmap tmp; |
7e0311ae | 1965 | basic_block bb; |
1966 | edge e; | |
1967 | edge_iterator ei; | |
3cebc9d2 | 1968 | |
b93ba654 | 1969 | /* It is possible to have jump threads in which one is a subpath |
1970 | of the other. ie, (A, B), (B, C), (C, D) where B is a joiner | |
1971 | block and (B, C), (C, D) where no joiner block exists. | |
1972 | ||
1973 | When this occurs ignore the jump thread request with the joiner | |
1974 | block. It's totally subsumed by the simpler jump thread request. | |
1975 | ||
1976 | This results in less block copying, simpler CFGs. More importantly, | |
1977 | when we duplicate the joiner block, B, in this case we will create | |
1978 | a new threading opportunity that we wouldn't be able to optimize | |
1979 | until the next jump threading iteration. | |
1980 | ||
1981 | So first convert the jump thread requests which do not require a | |
1982 | joiner block. */ | |
f2981b08 | 1983 | for (i = 0; i < paths.length (); i++) |
3cebc9d2 | 1984 | { |
f2981b08 | 1985 | vec<jump_thread_edge *> *path = paths[i]; |
b93ba654 | 1986 | |
1987 | if ((*path)[1]->type != EDGE_COPY_SRC_JOINER_BLOCK) | |
1988 | { | |
1989 | edge e = (*path)[0]->e; | |
1990 | e->aux = (void *)path; | |
1991 | bitmap_set_bit (tmp, e->dest->index); | |
1992 | } | |
1f3976e7 | 1993 | } |
1994 | ||
b93ba654 | 1995 | /* Now iterate again, converting cases where we want to thread |
1996 | through a joiner block, but only if no other edge on the path | |
f1ce4e72 | 1997 | already has a jump thread attached to it. We do this in two passes, |
1998 | to avoid situations where the order in the paths vec can hide overlapping | |
1999 | threads (the path is recorded on the incoming edge, so we would miss | |
2000 | cases where the second path starts at a downstream edge on the same | |
2001 | path). First record all joiner paths, deleting any in the unexpected | |
2002 | case where there is already a path for that incoming edge. */ | |
51ea8bc6 | 2003 | for (i = 0; i < paths.length ();) |
b93ba654 | 2004 | { |
2005 | vec<jump_thread_edge *> *path = paths[i]; | |
2006 | ||
2007 | if ((*path)[1]->type == EDGE_COPY_SRC_JOINER_BLOCK) | |
9943b198 | 2008 | { |
f1ce4e72 | 2009 | /* Attach the path to the starting edge if none is yet recorded. */ |
9943b198 | 2010 | if ((*path)[0]->e->aux == NULL) |
5d293e79 | 2011 | { |
9943b198 | 2012 | (*path)[0]->e->aux = path; |
51ea8bc6 | 2013 | i++; |
5d293e79 | 2014 | } |
2015 | else | |
2016 | { | |
2017 | paths.unordered_remove (i); | |
2018 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
9943b198 | 2019 | dump_jump_thread_path (dump_file, *path, false); |
5d293e79 | 2020 | delete_jump_thread_path (path); |
2021 | } | |
9943b198 | 2022 | } |
51ea8bc6 | 2023 | else |
2024 | { | |
2025 | i++; | |
2026 | } | |
f1ce4e72 | 2027 | } |
51ea8bc6 | 2028 | |
f1ce4e72 | 2029 | /* Second, look for paths that have any other jump thread attached to |
2030 | them, and either finish converting them or cancel them. */ | |
51ea8bc6 | 2031 | for (i = 0; i < paths.length ();) |
f1ce4e72 | 2032 | { |
2033 | vec<jump_thread_edge *> *path = paths[i]; | |
2034 | edge e = (*path)[0]->e; | |
2035 | ||
2036 | if ((*path)[1]->type == EDGE_COPY_SRC_JOINER_BLOCK && e->aux == path) | |
b93ba654 | 2037 | { |
2038 | unsigned int j; | |
f1ce4e72 | 2039 | for (j = 1; j < path->length (); j++) |
b93ba654 | 2040 | if ((*path)[j]->e->aux != NULL) |
2041 | break; | |
2042 | ||
2043 | /* If we iterated through the entire path without exiting the loop, | |
f1ce4e72 | 2044 | then we are good to go, record it. */ |
b93ba654 | 2045 | if (j == path->length ()) |
51ea8bc6 | 2046 | { |
2047 | bitmap_set_bit (tmp, e->dest->index); | |
2048 | i++; | |
2049 | } | |
f1ce4e72 | 2050 | else |
b93ba654 | 2051 | { |
f1ce4e72 | 2052 | e->aux = NULL; |
5d293e79 | 2053 | paths.unordered_remove (i); |
f1ce4e72 | 2054 | if (dump_file && (dump_flags & TDF_DETAILS)) |
9943b198 | 2055 | dump_jump_thread_path (dump_file, *path, false); |
5d293e79 | 2056 | delete_jump_thread_path (path); |
b93ba654 | 2057 | } |
2058 | } | |
51ea8bc6 | 2059 | else |
2060 | { | |
2061 | i++; | |
2062 | } | |
b93ba654 | 2063 | } |
b99a7d6d | 2064 | |
7e0311ae | 2065 | /* If optimizing for size, only thread through block if we don't have |
2066 | to duplicate it or it's an otherwise empty redirection block. */ | |
0bfd8d5c | 2067 | if (optimize_function_for_size_p (cfun)) |
7e0311ae | 2068 | { |
2069 | EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi) | |
2070 | { | |
f5a6b05f | 2071 | bb = BASIC_BLOCK_FOR_FN (cfun, i); |
7e0311ae | 2072 | if (EDGE_COUNT (bb->preds) > 1 |
2073 | && !redirection_block_p (bb)) | |
2074 | { | |
2075 | FOR_EACH_EDGE (e, ei, bb->preds) | |
eb31063a | 2076 | { |
f2981b08 | 2077 | if (e->aux) |
2078 | { | |
2079 | vec<jump_thread_edge *> *path = THREAD_PATH (e); | |
6d1fdbf9 | 2080 | delete_jump_thread_path (path); |
f2981b08 | 2081 | e->aux = NULL; |
2082 | } | |
eb31063a | 2083 | } |
7e0311ae | 2084 | } |
2085 | else | |
2086 | bitmap_set_bit (threaded_blocks, i); | |
2087 | } | |
3cebc9d2 | 2088 | } |
7e0311ae | 2089 | else |
2090 | bitmap_copy (threaded_blocks, tmp); | |
2091 | ||
af6b6631 | 2092 | /* If we have a joiner block (J) which has two successors S1 and S2 and |
2093 | we are threading though S1 and the final destination of the thread | |
2094 | is S2, then we must verify that any PHI nodes in S2 have the same | |
2095 | PHI arguments for the edge J->S2 and J->S1->...->S2. | |
2096 | ||
2097 | We used to detect this prior to registering the jump thread, but | |
2098 | that prohibits propagation of edge equivalences into non-dominated | |
2099 | PHI nodes as the equivalency test might occur before propagation. | |
2100 | ||
2101 | This must also occur after we truncate any jump threading paths | |
2102 | as this scenario may only show up after truncation. | |
2103 | ||
2104 | This works for now, but will need improvement as part of the FSA | |
2105 | optimization. | |
2106 | ||
2107 | Note since we've moved the thread request data to the edges, | |
2108 | we have to iterate on those rather than the threaded_edges vector. */ | |
2109 | EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi) | |
2110 | { | |
f5a6b05f | 2111 | bb = BASIC_BLOCK_FOR_FN (cfun, i); |
af6b6631 | 2112 | FOR_EACH_EDGE (e, ei, bb->preds) |
2113 | { | |
2114 | if (e->aux) | |
2115 | { | |
2116 | vec<jump_thread_edge *> *path = THREAD_PATH (e); | |
2117 | bool have_joiner = ((*path)[1]->type == EDGE_COPY_SRC_JOINER_BLOCK); | |
2118 | ||
2119 | if (have_joiner) | |
2120 | { | |
2121 | basic_block joiner = e->dest; | |
2122 | edge final_edge = path->last ()->e; | |
2123 | basic_block final_dest = final_edge->dest; | |
2124 | edge e2 = find_edge (joiner, final_dest); | |
2125 | ||
2126 | if (e2 && !phi_args_equal_on_edges (e2, final_edge)) | |
2127 | { | |
2128 | delete_jump_thread_path (path); | |
2129 | e->aux = NULL; | |
2130 | } | |
2131 | } | |
2132 | } | |
2133 | } | |
2134 | } | |
2135 | ||
0afc9b47 | 2136 | /* Look for jump threading paths which cross multiple loop headers. |
2137 | ||
2138 | The code to thread through loop headers will change the CFG in ways | |
2139 | that invalidate the cached loop iteration information. So we must | |
2140 | detect that case and wipe the cached information. */ | |
2141 | EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi) | |
2142 | { | |
2143 | basic_block bb = BASIC_BLOCK_FOR_FN (cfun, i); | |
2144 | FOR_EACH_EDGE (e, ei, bb->preds) | |
2145 | { | |
2146 | if (e->aux) | |
2147 | { | |
2148 | vec<jump_thread_edge *> *path = THREAD_PATH (e); | |
2149 | ||
2150 | for (unsigned int i = 0, crossed_headers = 0; | |
2151 | i < path->length (); | |
2152 | i++) | |
2153 | { | |
2154 | basic_block dest = (*path)[i]->e->dest; | |
2155 | basic_block src = (*path)[i]->e->src; | |
2156 | /* If we enter a loop. */ | |
2157 | if (flow_loop_nested_p (src->loop_father, dest->loop_father)) | |
2158 | ++crossed_headers; | |
2159 | /* If we step from a block outside an irreducible region | |
2160 | to a block inside an irreducible region, then we have | |
2161 | crossed into a loop. */ | |
2162 | else if (! (src->flags & BB_IRREDUCIBLE_LOOP) | |
2163 | && (dest->flags & BB_IRREDUCIBLE_LOOP)) | |
2164 | ++crossed_headers; | |
2165 | if (crossed_headers > 1) | |
2166 | { | |
2167 | vect_free_loop_info_assumptions | |
2168 | ((*path)[path->length () - 1]->e->dest->loop_father); | |
2169 | break; | |
2170 | } | |
2171 | } | |
2172 | } | |
2173 | } | |
2174 | } | |
3cebc9d2 | 2175 | } |
2176 | ||
2177 | ||
9e0d85a7 | 2178 | /* Verify that the REGION is a valid jump thread. A jump thread is a special |
2179 | case of SEME Single Entry Multiple Exits region in which all nodes in the | |
2180 | REGION have exactly one incoming edge. The only exception is the first block | |
2181 | that may not have been connected to the rest of the cfg yet. */ | |
ded1c768 | 2182 | |
2183 | DEBUG_FUNCTION void | |
9e0d85a7 | 2184 | verify_jump_thread (basic_block *region, unsigned n_region) |
ded1c768 | 2185 | { |
ded1c768 | 2186 | for (unsigned i = 0; i < n_region; i++) |
9e0d85a7 | 2187 | gcc_assert (EDGE_COUNT (region[i]->preds) <= 1); |
2188 | } | |
ded1c768 | 2189 | |
9e0d85a7 | 2190 | /* Return true when BB is one of the first N items in BBS. */ |
ded1c768 | 2191 | |
9e0d85a7 | 2192 | static inline bool |
2193 | bb_in_bbs (basic_block bb, basic_block *bbs, int n) | |
2194 | { | |
2195 | for (int i = 0; i < n; i++) | |
2196 | if (bb == bbs[i]) | |
2197 | return true; | |
ded1c768 | 2198 | |
9e0d85a7 | 2199 | return false; |
ded1c768 | 2200 | } |
2201 | ||
9e0d85a7 | 2202 | /* Duplicates a jump-thread path of N_REGION basic blocks. |
2203 | The ENTRY edge is redirected to the duplicate of the region. | |
ded1c768 | 2204 | |
2205 | Remove the last conditional statement in the last basic block in the REGION, | |
2206 | and create a single fallthru edge pointing to the same destination as the | |
2207 | EXIT edge. | |
2208 | ||
2209 | The new basic blocks are stored to REGION_COPY in the same order as they had | |
2210 | in REGION, provided that REGION_COPY is not NULL. | |
2211 | ||
2212 | Returns false if it is unable to copy the region, true otherwise. */ | |
2213 | ||
2214 | static bool | |
9e0d85a7 | 2215 | duplicate_thread_path (edge entry, edge exit, |
ded1c768 | 2216 | basic_block *region, unsigned n_region, |
2217 | basic_block *region_copy) | |
2218 | { | |
2219 | unsigned i; | |
af5f6a93 | 2220 | bool free_region_copy = false; |
ded1c768 | 2221 | struct loop *loop = entry->dest->loop_father; |
2222 | edge exit_copy; | |
2223 | edge redirected; | |
acb9fddd | 2224 | int curr_freq; |
db9cef39 | 2225 | profile_count curr_count; |
ded1c768 | 2226 | |
2227 | if (!can_copy_bbs_p (region, n_region)) | |
2228 | return false; | |
2229 | ||
2230 | /* Some sanity checking. Note that we do not check for all possible | |
2231 | missuses of the functions. I.e. if you ask to copy something weird, | |
2232 | it will work, but the state of structures probably will not be | |
2233 | correct. */ | |
2234 | for (i = 0; i < n_region; i++) | |
2235 | { | |
2236 | /* We do not handle subloops, i.e. all the blocks must belong to the | |
2237 | same loop. */ | |
2238 | if (region[i]->loop_father != loop) | |
2239 | return false; | |
2240 | } | |
2241 | ||
2242 | initialize_original_copy_tables (); | |
2243 | ||
af5f6a93 | 2244 | set_loop_copy (loop, loop); |
ded1c768 | 2245 | |
2246 | if (!region_copy) | |
2247 | { | |
2248 | region_copy = XNEWVEC (basic_block, n_region); | |
2249 | free_region_copy = true; | |
2250 | } | |
2251 | ||
ded1c768 | 2252 | copy_bbs (region, n_region, region_copy, &exit, 1, &exit_copy, loop, |
9e0d85a7 | 2253 | split_edge_bb_loc (entry), false); |
2254 | ||
2255 | /* Fix up: copy_bbs redirects all edges pointing to copied blocks. The | |
2256 | following code ensures that all the edges exiting the jump-thread path are | |
2257 | redirected back to the original code: these edges are exceptions | |
2258 | invalidating the property that is propagated by executing all the blocks of | |
2259 | the jump-thread path in order. */ | |
2260 | ||
acb9fddd | 2261 | curr_count = entry->count; |
2262 | curr_freq = EDGE_FREQUENCY (entry); | |
2263 | ||
9e0d85a7 | 2264 | for (i = 0; i < n_region; i++) |
2265 | { | |
2266 | edge e; | |
2267 | edge_iterator ei; | |
2268 | basic_block bb = region_copy[i]; | |
2269 | ||
acb9fddd | 2270 | /* Watch inconsistent profile. */ |
2271 | if (curr_count > region[i]->count) | |
2272 | curr_count = region[i]->count; | |
2273 | if (curr_freq > region[i]->frequency) | |
2274 | curr_freq = region[i]->frequency; | |
2275 | /* Scale current BB. */ | |
db9cef39 | 2276 | if (region[i]->count > 0 && curr_count.initialized_p ()) |
acb9fddd | 2277 | { |
2278 | /* In the middle of the path we only scale the frequencies. | |
2279 | In last BB we need to update probabilities of outgoing edges | |
2280 | because we know which one is taken at the threaded path. */ | |
2281 | if (i + 1 != n_region) | |
db9cef39 | 2282 | scale_bbs_frequencies_profile_count (region + i, 1, |
2283 | region[i]->count - curr_count, | |
2284 | region[i]->count); | |
acb9fddd | 2285 | else |
2286 | update_bb_profile_for_threading (region[i], | |
2287 | curr_freq, curr_count, | |
2288 | exit); | |
db9cef39 | 2289 | scale_bbs_frequencies_profile_count (region_copy + i, 1, curr_count, |
2290 | region_copy[i]->count); | |
acb9fddd | 2291 | } |
2292 | else if (region[i]->frequency) | |
2293 | { | |
2294 | if (i + 1 != n_region) | |
2295 | scale_bbs_frequencies_int (region + i, 1, | |
2296 | region[i]->frequency - curr_freq, | |
2297 | region[i]->frequency); | |
2298 | else | |
2299 | update_bb_profile_for_threading (region[i], | |
2300 | curr_freq, curr_count, | |
2301 | exit); | |
2302 | scale_bbs_frequencies_int (region_copy + i, 1, curr_freq, | |
2303 | region_copy[i]->frequency); | |
2304 | } | |
2305 | ||
9e0d85a7 | 2306 | if (single_succ_p (bb)) |
2307 | { | |
2308 | /* Make sure the successor is the next node in the path. */ | |
2309 | gcc_assert (i + 1 == n_region | |
2310 | || region_copy[i + 1] == single_succ_edge (bb)->dest); | |
acb9fddd | 2311 | if (i + 1 != n_region) |
2312 | { | |
2313 | curr_freq = EDGE_FREQUENCY (single_succ_edge (bb)); | |
2314 | curr_count = single_succ_edge (bb)->count; | |
2315 | } | |
9e0d85a7 | 2316 | continue; |
2317 | } | |
2318 | ||
2319 | /* Special case the last block on the path: make sure that it does not | |
5a653985 | 2320 | jump back on the copied path, including back to itself. */ |
9e0d85a7 | 2321 | if (i + 1 == n_region) |
2322 | { | |
2323 | FOR_EACH_EDGE (e, ei, bb->succs) | |
5a653985 | 2324 | if (bb_in_bbs (e->dest, region_copy, n_region)) |
9e0d85a7 | 2325 | { |
2326 | basic_block orig = get_bb_original (e->dest); | |
2327 | if (orig) | |
2328 | redirect_edge_and_branch_force (e, orig); | |
2329 | } | |
2330 | continue; | |
2331 | } | |
2332 | ||
2333 | /* Redirect all other edges jumping to non-adjacent blocks back to the | |
2334 | original code. */ | |
2335 | FOR_EACH_EDGE (e, ei, bb->succs) | |
2336 | if (region_copy[i + 1] != e->dest) | |
2337 | { | |
2338 | basic_block orig = get_bb_original (e->dest); | |
2339 | if (orig) | |
2340 | redirect_edge_and_branch_force (e, orig); | |
2341 | } | |
acb9fddd | 2342 | else |
2343 | { | |
2344 | curr_freq = EDGE_FREQUENCY (e); | |
2345 | curr_count = e->count; | |
2346 | } | |
9e0d85a7 | 2347 | } |
2348 | ||
ded1c768 | 2349 | |
382ecba7 | 2350 | if (flag_checking) |
2351 | verify_jump_thread (region_copy, n_region); | |
ded1c768 | 2352 | |
2353 | /* Remove the last branch in the jump thread path. */ | |
2354 | remove_ctrl_stmt_and_useless_edges (region_copy[n_region - 1], exit->dest); | |
7729459f | 2355 | |
2356 | /* And fixup the flags on the single remaining edge. */ | |
2357 | edge fix_e = find_edge (region_copy[n_region - 1], exit->dest); | |
2358 | fix_e->flags &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE | EDGE_ABNORMAL); | |
2359 | fix_e->flags |= EDGE_FALLTHRU; | |
2360 | ||
ded1c768 | 2361 | edge e = make_edge (region_copy[n_region - 1], exit->dest, EDGE_FALLTHRU); |
2362 | ||
acb9fddd | 2363 | if (e) |
2364 | { | |
2365 | rescan_loop_exit (e, true, false); | |
720cfc43 | 2366 | e->probability = profile_probability::always (); |
acb9fddd | 2367 | e->count = region_copy[n_region - 1]->count; |
2368 | } | |
ded1c768 | 2369 | |
2370 | /* Redirect the entry and add the phi node arguments. */ | |
af5f6a93 | 2371 | if (entry->dest == loop->header) |
2372 | mark_loop_for_removal (loop); | |
ded1c768 | 2373 | redirected = redirect_edge_and_branch (entry, get_bb_copy (entry->dest)); |
2374 | gcc_assert (redirected != NULL); | |
2375 | flush_pending_stmts (entry); | |
2376 | ||
2377 | /* Add the other PHI node arguments. */ | |
2378 | add_phi_args_after_copy (region_copy, n_region, NULL); | |
2379 | ||
2380 | if (free_region_copy) | |
2381 | free (region_copy); | |
2382 | ||
2383 | free_original_copy_tables (); | |
2384 | return true; | |
2385 | } | |
2386 | ||
b9903eb3 | 2387 | /* Return true when PATH is a valid jump-thread path. */ |
2388 | ||
2389 | static bool | |
2390 | valid_jump_thread_path (vec<jump_thread_edge *> *path) | |
2391 | { | |
2392 | unsigned len = path->length (); | |
2393 | ||
105cb8d7 | 2394 | /* Check that the path is connected. */ |
b9903eb3 | 2395 | for (unsigned int j = 0; j < len - 1; j++) |
8587f7ea | 2396 | { |
f497d674 | 2397 | edge e = (*path)[j]->e; |
f497d674 | 2398 | if (e->dest != (*path)[j+1]->e->src) |
31e5d72d | 2399 | return false; |
8587f7ea | 2400 | } |
b9903eb3 | 2401 | return true; |
2402 | } | |
2403 | ||
a27d141e | 2404 | /* Remove any queued jump threads that include edge E. |
2405 | ||
2406 | We don't actually remove them here, just record the edges into ax | |
2407 | hash table. That way we can do the search once per iteration of | |
2408 | DOM/VRP rather than for every case where DOM optimizes away a COND_EXPR. */ | |
f1344f45 | 2409 | |
2410 | void | |
a27d141e | 2411 | remove_jump_threads_including (edge_def *e) |
f1344f45 | 2412 | { |
2413 | if (!paths.exists ()) | |
2414 | return; | |
2415 | ||
a27d141e | 2416 | if (!removed_edges) |
2417 | removed_edges = new hash_table<struct removed_edges> (17); | |
f1344f45 | 2418 | |
a27d141e | 2419 | edge *slot = removed_edges->find_slot (e, INSERT); |
2420 | *slot = e; | |
f1344f45 | 2421 | } |
2422 | ||
3cebc9d2 | 2423 | /* Walk through all blocks and thread incoming edges to the appropriate |
2424 | outgoing edge for each edge pair recorded in THREADED_EDGES. | |
a8046f60 | 2425 | |
2426 | It is the caller's responsibility to fix the dominance information | |
2427 | and rewrite duplicated SSA_NAMEs back into SSA form. | |
2428 | ||
7e0311ae | 2429 | If MAY_PEEL_LOOP_HEADERS is false, we avoid threading edges through |
2430 | loop headers if it does not simplify the loop. | |
2431 | ||
dac49aa5 | 2432 | Returns true if one or more edges were threaded, false otherwise. */ |
a8046f60 | 2433 | |
2434 | bool | |
7e0311ae | 2435 | thread_through_all_blocks (bool may_peel_loop_headers) |
a8046f60 | 2436 | { |
a8046f60 | 2437 | bool retval = false; |
7ea47fbd | 2438 | unsigned int i; |
2439 | bitmap_iterator bi; | |
7e0311ae | 2440 | struct loop *loop; |
035def86 | 2441 | auto_bitmap threaded_blocks; |
3cebc9d2 | 2442 | |
f2981b08 | 2443 | if (!paths.exists ()) |
a27d141e | 2444 | { |
2445 | retval = false; | |
2446 | goto out; | |
2447 | } | |
a8046f60 | 2448 | |
5236b8bb | 2449 | memset (&thread_stats, 0, sizeof (thread_stats)); |
388d1fc1 | 2450 | |
a27d141e | 2451 | /* Remove any paths that referenced removed edges. */ |
2452 | if (removed_edges) | |
2453 | for (i = 0; i < paths.length (); ) | |
2454 | { | |
2455 | unsigned int j; | |
2456 | vec<jump_thread_edge *> *path = paths[i]; | |
2457 | ||
2458 | for (j = 0; j < path->length (); j++) | |
2459 | { | |
2460 | edge e = (*path)[j]->e; | |
2461 | if (removed_edges->find_slot (e, NO_INSERT)) | |
2462 | break; | |
2463 | } | |
2464 | ||
2465 | if (j != path->length ()) | |
2466 | { | |
2467 | delete_jump_thread_path (path); | |
2468 | paths.unordered_remove (i); | |
2469 | continue; | |
2470 | } | |
2471 | i++; | |
2472 | } | |
2473 | ||
ded1c768 | 2474 | /* Jump-thread all FSM threads before other jump-threads. */ |
2475 | for (i = 0; i < paths.length ();) | |
2476 | { | |
2477 | vec<jump_thread_edge *> *path = paths[i]; | |
2478 | edge entry = (*path)[0]->e; | |
2479 | ||
b9903eb3 | 2480 | /* Only code-generate FSM jump-threads in this loop. */ |
2481 | if ((*path)[0]->type != EDGE_FSM_THREAD) | |
2482 | { | |
2483 | i++; | |
2484 | continue; | |
2485 | } | |
2486 | ||
2487 | /* Do not jump-thread twice from the same block. */ | |
2488 | if (bitmap_bit_p (threaded_blocks, entry->src->index) | |
545d4a44 | 2489 | /* We may not want to realize this jump thread path |
2490 | for various reasons. So check it first. */ | |
b9903eb3 | 2491 | || !valid_jump_thread_path (path)) |
2492 | { | |
2493 | /* Remove invalid FSM jump-thread paths. */ | |
2494 | delete_jump_thread_path (path); | |
2495 | paths.unordered_remove (i); | |
2496 | continue; | |
2497 | } | |
ded1c768 | 2498 | |
2499 | unsigned len = path->length (); | |
2500 | edge exit = (*path)[len - 1]->e; | |
2501 | basic_block *region = XNEWVEC (basic_block, len - 1); | |
2502 | ||
2503 | for (unsigned int j = 0; j < len - 1; j++) | |
2504 | region[j] = (*path)[j]->e->dest; | |
2505 | ||
9e0d85a7 | 2506 | if (duplicate_thread_path (entry, exit, region, len - 1, NULL)) |
ded1c768 | 2507 | { |
2508 | /* We do not update dominance info. */ | |
2509 | free_dominance_info (CDI_DOMINATORS); | |
2510 | bitmap_set_bit (threaded_blocks, entry->src->index); | |
2511 | retval = true; | |
7998c0b5 | 2512 | thread_stats.num_threaded_edges++; |
ded1c768 | 2513 | } |
2514 | ||
2515 | delete_jump_thread_path (path); | |
2516 | paths.unordered_remove (i); | |
b7cbf36d | 2517 | free (region); |
ded1c768 | 2518 | } |
2519 | ||
2520 | /* Remove from PATHS all the jump-threads starting with an edge already | |
2521 | jump-threaded. */ | |
2522 | for (i = 0; i < paths.length ();) | |
2523 | { | |
2524 | vec<jump_thread_edge *> *path = paths[i]; | |
2525 | edge entry = (*path)[0]->e; | |
2526 | ||
2527 | /* Do not jump-thread twice from the same block. */ | |
2528 | if (bitmap_bit_p (threaded_blocks, entry->src->index)) | |
2529 | { | |
2530 | delete_jump_thread_path (path); | |
2531 | paths.unordered_remove (i); | |
2532 | } | |
2533 | else | |
2534 | i++; | |
2535 | } | |
2536 | ||
2537 | bitmap_clear (threaded_blocks); | |
2538 | ||
3cebc9d2 | 2539 | mark_threaded_blocks (threaded_blocks); |
2540 | ||
96c90e5e | 2541 | initialize_original_copy_tables (); |
7e0311ae | 2542 | |
2543 | /* First perform the threading requests that do not affect | |
2544 | loop structure. */ | |
7ea47fbd | 2545 | EXECUTE_IF_SET_IN_BITMAP (threaded_blocks, 0, i, bi) |
a8046f60 | 2546 | { |
f5a6b05f | 2547 | basic_block bb = BASIC_BLOCK_FOR_FN (cfun, i); |
7ea47fbd | 2548 | |
2549 | if (EDGE_COUNT (bb->preds) > 0) | |
7e0311ae | 2550 | retval |= thread_block (bb, true); |
2551 | } | |
2552 | ||
2553 | /* Then perform the threading through loop headers. We start with the | |
2554 | innermost loop, so that the changes in cfg we perform won't affect | |
2555 | further threading. */ | |
f21d4d00 | 2556 | FOR_EACH_LOOP (loop, LI_FROM_INNERMOST) |
7e0311ae | 2557 | { |
7a3bf727 | 2558 | if (!loop->header |
2559 | || !bitmap_bit_p (threaded_blocks, loop->header->index)) | |
2560 | continue; | |
7e0311ae | 2561 | |
7a3bf727 | 2562 | retval |= thread_through_loop_header (loop, may_peel_loop_headers); |
a8046f60 | 2563 | } |
388d1fc1 | 2564 | |
c3dbcc09 | 2565 | /* All jump threading paths should have been resolved at this |
2566 | point. Verify that is the case. */ | |
ed4feca1 | 2567 | basic_block bb; |
fc00614f | 2568 | FOR_EACH_BB_FN (bb, cfun) |
ed4feca1 | 2569 | { |
c3dbcc09 | 2570 | edge_iterator ei; |
2571 | edge e; | |
2572 | FOR_EACH_EDGE (e, ei, bb->preds) | |
2573 | gcc_assert (e->aux == NULL); | |
ed4feca1 | 2574 | } |
2575 | ||
581f8050 | 2576 | statistics_counter_event (cfun, "Jumps threaded", |
2577 | thread_stats.num_threaded_edges); | |
5236b8bb | 2578 | |
96c90e5e | 2579 | free_original_copy_tables (); |
2580 | ||
f2981b08 | 2581 | paths.release (); |
7e0311ae | 2582 | |
396c773e | 2583 | if (retval) |
f24ec26f | 2584 | loops_state_set (LOOPS_NEED_FIXUP); |
eb2a640e | 2585 | |
a27d141e | 2586 | out: |
2587 | delete removed_edges; | |
2588 | removed_edges = NULL; | |
a8046f60 | 2589 | return retval; |
2590 | } | |
3cebc9d2 | 2591 | |
6d1fdbf9 | 2592 | /* Delete the jump threading path PATH. We have to explcitly delete |
2593 | each entry in the vector, then the container. */ | |
2594 | ||
2595 | void | |
2596 | delete_jump_thread_path (vec<jump_thread_edge *> *path) | |
2597 | { | |
2598 | for (unsigned int i = 0; i < path->length (); i++) | |
2599 | delete (*path)[i]; | |
2600 | path->release(); | |
9b5a88db | 2601 | delete path; |
6d1fdbf9 | 2602 | } |
2603 | ||
3cebc9d2 | 2604 | /* Register a jump threading opportunity. We queue up all the jump |
2605 | threading opportunities discovered by a pass and update the CFG | |
2606 | and SSA form all at once. | |
2607 | ||
f0b5f617 | 2608 | E is the edge we can thread, E2 is the new target edge, i.e., we |
3cebc9d2 | 2609 | are effectively recording that E->dest can be changed to E2->dest |
2610 | after fixing the SSA graph. */ | |
2611 | ||
2612 | void | |
f2981b08 | 2613 | register_jump_thread (vec<jump_thread_edge *> *path) |
3cebc9d2 | 2614 | { |
a3724f9d | 2615 | if (!dbg_cnt (registered_jump_thread)) |
2616 | { | |
6d1fdbf9 | 2617 | delete_jump_thread_path (path); |
a3724f9d | 2618 | return; |
2619 | } | |
2620 | ||
0c5b289a | 2621 | /* First make sure there are no NULL outgoing edges on the jump threading |
2622 | path. That can happen for jumping to a constant address. */ | |
f2981b08 | 2623 | for (unsigned int i = 0; i < path->length (); i++) |
d29932c9 | 2624 | { |
2625 | if ((*path)[i]->e == NULL) | |
31e5d72d | 2626 | { |
d29932c9 | 2627 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2628 | { | |
2629 | fprintf (dump_file, | |
2630 | "Found NULL edge in jump threading path. Cancelling jump thread:\n"); | |
2631 | dump_jump_thread_path (dump_file, *path, false); | |
2632 | } | |
f2981b08 | 2633 | |
d29932c9 | 2634 | delete_jump_thread_path (path); |
2635 | return; | |
31e5d72d | 2636 | } |
d29932c9 | 2637 | |
2638 | /* Only the FSM threader is allowed to thread across | |
2639 | backedges in the CFG. */ | |
2640 | if (flag_checking | |
2641 | && (*path)[0]->type != EDGE_FSM_THREAD) | |
2642 | gcc_assert (((*path)[i]->e->flags & EDGE_DFS_BACK) == 0); | |
2643 | } | |
5411af4e | 2644 | |
631d940c | 2645 | if (dump_file && (dump_flags & TDF_DETAILS)) |
b93ba654 | 2646 | dump_jump_thread_path (dump_file, *path, true); |
631d940c | 2647 | |
f2981b08 | 2648 | if (!paths.exists ()) |
2649 | paths.create (5); | |
631d940c | 2650 | |
f2981b08 | 2651 | paths.safe_push (path); |
3cebc9d2 | 2652 | } |