]>
Commit | Line | Data |
---|---|---|
a8046f60 | 1 | /* Thread edges through blocks and update the control flow and SSA graphs. |
fbd26352 | 2 | Copyright (C) 2004-2019 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 | ei_next (&ei); |
307 | } | |
a8046f60 | 308 | } |
5b4ada2a | 309 | |
310 | /* If the remaining edge is a loop exit, there must have | |
311 | a removed edge that was not a loop exit. | |
312 | ||
313 | In that case BB and possibly other blocks were previously | |
314 | in the loop, but are now outside the loop. Thus, we need | |
315 | to update the loop structures. */ | |
316 | if (single_succ_p (bb) | |
317 | && loop_outer (bb->loop_father) | |
318 | && loop_exit_edge_p (bb->loop_father, single_succ_edge (bb))) | |
319 | loops_state_set (LOOPS_NEED_FIXUP); | |
a8046f60 | 320 | } |
321 | ||
11af02d8 | 322 | /* Create a duplicate of BB. Record the duplicate block in an array |
a7ee7309 | 323 | indexed by COUNT stored in RD. */ |
a8046f60 | 324 | |
325 | static void | |
11af02d8 | 326 | create_block_for_threading (basic_block bb, |
327 | struct redirection_data *rd, | |
30e432bb | 328 | unsigned int count, |
a7ee7309 | 329 | bitmap *duplicate_blocks) |
a8046f60 | 330 | { |
eb31063a | 331 | edge_iterator ei; |
332 | edge e; | |
333 | ||
a8046f60 | 334 | /* We can use the generic block duplication code and simply remove |
335 | the stuff we do not need. */ | |
11af02d8 | 336 | rd->dup_blocks[count] = duplicate_block (bb, NULL, NULL); |
a8046f60 | 337 | |
11af02d8 | 338 | FOR_EACH_EDGE (e, ei, rd->dup_blocks[count]->succs) |
be69e68e | 339 | { |
340 | e->aux = NULL; | |
341 | ||
342 | /* If we duplicate a block with an outgoing edge marked as | |
343 | EDGE_IGNORE, we must clear EDGE_IGNORE so that it doesn't | |
344 | leak out of the current pass. | |
345 | ||
346 | It would be better to simplify switch statements and remove | |
347 | the edges before we get here, but the sequencing is nontrivial. */ | |
348 | e->flags &= ~EDGE_IGNORE; | |
349 | } | |
eb31063a | 350 | |
615dd397 | 351 | /* Zero out the profile, since the block is unreachable for now. */ |
db9cef39 | 352 | rd->dup_blocks[count]->count = profile_count::uninitialized (); |
30e432bb | 353 | if (duplicate_blocks) |
354 | bitmap_set_bit (*duplicate_blocks, rd->dup_blocks[count]->index); | |
a8046f60 | 355 | } |
356 | ||
2b15d2ba | 357 | /* Main data structure to hold information for duplicates of BB. */ |
358 | ||
c1f445d2 | 359 | static hash_table<redirection_data> *redirection_data; |
2b15d2ba | 360 | |
778182c1 | 361 | /* Given an outgoing edge E lookup and return its entry in our hash table. |
362 | ||
363 | If INSERT is true, then we insert the entry into the hash table if | |
364 | it is not already present. INCOMING_EDGE is added to the list of incoming | |
365 | edges associated with E in the hash table. */ | |
366 | ||
367 | static struct redirection_data * | |
da81e0c5 | 368 | lookup_redirection_data (edge e, enum insert_option insert) |
778182c1 | 369 | { |
2b15d2ba | 370 | struct redirection_data **slot; |
778182c1 | 371 | struct redirection_data *elt; |
f2981b08 | 372 | vec<jump_thread_edge *> *path = THREAD_PATH (e); |
778182c1 | 373 | |
31e5d72d | 374 | /* Build a hash table element so we can see if E is already |
778182c1 | 375 | in the table. */ |
4c36ffe6 | 376 | elt = XNEW (struct redirection_data); |
5fe6149c | 377 | elt->path = path; |
11af02d8 | 378 | elt->dup_blocks[0] = NULL; |
379 | elt->dup_blocks[1] = NULL; | |
778182c1 | 380 | elt->incoming_edges = NULL; |
381 | ||
c1f445d2 | 382 | slot = redirection_data->find_slot (elt, insert); |
778182c1 | 383 | |
384 | /* This will only happen if INSERT is false and the entry is not | |
385 | in the hash table. */ | |
386 | if (slot == NULL) | |
387 | { | |
388 | free (elt); | |
389 | return NULL; | |
390 | } | |
391 | ||
392 | /* This will only happen if E was not in the hash table and | |
393 | INSERT is true. */ | |
394 | if (*slot == NULL) | |
395 | { | |
2b15d2ba | 396 | *slot = elt; |
4c36ffe6 | 397 | elt->incoming_edges = XNEW (struct el); |
da81e0c5 | 398 | elt->incoming_edges->e = e; |
778182c1 | 399 | elt->incoming_edges->next = NULL; |
400 | return elt; | |
401 | } | |
402 | /* E was in the hash table. */ | |
403 | else | |
404 | { | |
405 | /* Free ELT as we do not need it anymore, we will extract the | |
406 | relevant entry from the hash table itself. */ | |
407 | free (elt); | |
408 | ||
409 | /* Get the entry stored in the hash table. */ | |
2b15d2ba | 410 | elt = *slot; |
778182c1 | 411 | |
412 | /* If insertion was requested, then we need to add INCOMING_EDGE | |
413 | to the list of incoming edges associated with E. */ | |
414 | if (insert) | |
415 | { | |
559685be | 416 | struct el *el = XNEW (struct el); |
778182c1 | 417 | el->next = elt->incoming_edges; |
da81e0c5 | 418 | el->e = e; |
778182c1 | 419 | elt->incoming_edges = el; |
420 | } | |
421 | ||
422 | return elt; | |
423 | } | |
424 | } | |
425 | ||
fc54aba7 | 426 | /* Similar to copy_phi_args, except that the PHI arg exists, it just |
427 | does not have a value associated with it. */ | |
428 | ||
429 | static void | |
430 | copy_phi_arg_into_existing_phi (edge src_e, edge tgt_e) | |
431 | { | |
432 | int src_idx = src_e->dest_idx; | |
433 | int tgt_idx = tgt_e->dest_idx; | |
434 | ||
435 | /* Iterate over each PHI in e->dest. */ | |
1a91d914 | 436 | for (gphi_iterator gsi = gsi_start_phis (src_e->dest), |
437 | gsi2 = gsi_start_phis (tgt_e->dest); | |
fc54aba7 | 438 | !gsi_end_p (gsi); |
439 | gsi_next (&gsi), gsi_next (&gsi2)) | |
440 | { | |
1a91d914 | 441 | gphi *src_phi = gsi.phi (); |
442 | gphi *dest_phi = gsi2.phi (); | |
fc54aba7 | 443 | tree val = gimple_phi_arg_def (src_phi, src_idx); |
be1e7283 | 444 | location_t locus = gimple_phi_arg_location (src_phi, src_idx); |
fc54aba7 | 445 | |
446 | SET_PHI_ARG_DEF (dest_phi, tgt_idx, val); | |
447 | gimple_phi_arg_set_location (dest_phi, tgt_idx, locus); | |
448 | } | |
449 | } | |
450 | ||
1b83c31b | 451 | /* Given ssa_name DEF, backtrack jump threading PATH from node IDX |
452 | to see if it has constant value in a flow sensitive manner. Set | |
453 | LOCUS to location of the constant phi arg and return the value. | |
454 | Return DEF directly if either PATH or idx is ZERO. */ | |
455 | ||
456 | static tree | |
457 | get_value_locus_in_path (tree def, vec<jump_thread_edge *> *path, | |
be1e7283 | 458 | basic_block bb, int idx, location_t *locus) |
1b83c31b | 459 | { |
460 | tree arg; | |
1a91d914 | 461 | gphi *def_phi; |
1b83c31b | 462 | basic_block def_bb; |
463 | ||
464 | if (path == NULL || idx == 0) | |
465 | return def; | |
466 | ||
1a91d914 | 467 | def_phi = dyn_cast <gphi *> (SSA_NAME_DEF_STMT (def)); |
468 | if (!def_phi) | |
1b83c31b | 469 | return def; |
470 | ||
471 | def_bb = gimple_bb (def_phi); | |
472 | /* Don't propagate loop invariants into deeper loops. */ | |
473 | if (!def_bb || bb_loop_depth (def_bb) < bb_loop_depth (bb)) | |
474 | return def; | |
475 | ||
476 | /* Backtrack jump threading path from IDX to see if def has constant | |
477 | value. */ | |
478 | for (int j = idx - 1; j >= 0; j--) | |
479 | { | |
480 | edge e = (*path)[j]->e; | |
481 | if (e->dest == def_bb) | |
482 | { | |
483 | arg = gimple_phi_arg_def (def_phi, e->dest_idx); | |
484 | if (is_gimple_min_invariant (arg)) | |
485 | { | |
486 | *locus = gimple_phi_arg_location (def_phi, e->dest_idx); | |
487 | return arg; | |
488 | } | |
489 | break; | |
490 | } | |
491 | } | |
492 | ||
493 | return def; | |
494 | } | |
495 | ||
496 | /* For each PHI in BB, copy the argument associated with SRC_E to TGT_E. | |
497 | Try to backtrack jump threading PATH from node IDX to see if the arg | |
498 | has constant value, copy constant value instead of argument itself | |
499 | if yes. */ | |
da81e0c5 | 500 | |
501 | static void | |
1b83c31b | 502 | copy_phi_args (basic_block bb, edge src_e, edge tgt_e, |
503 | vec<jump_thread_edge *> *path, int idx) | |
da81e0c5 | 504 | { |
1a91d914 | 505 | gphi_iterator gsi; |
da81e0c5 | 506 | int src_indx = src_e->dest_idx; |
507 | ||
508 | for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
509 | { | |
1a91d914 | 510 | gphi *phi = gsi.phi (); |
1b83c31b | 511 | tree def = gimple_phi_arg_def (phi, src_indx); |
be1e7283 | 512 | location_t locus = gimple_phi_arg_location (phi, src_indx); |
1b83c31b | 513 | |
514 | if (TREE_CODE (def) == SSA_NAME | |
515 | && !virtual_operand_p (gimple_phi_result (phi))) | |
516 | def = get_value_locus_in_path (def, path, bb, idx, &locus); | |
517 | ||
518 | add_phi_arg (phi, def, tgt_e, locus); | |
da81e0c5 | 519 | } |
520 | } | |
521 | ||
522 | /* We have recently made a copy of ORIG_BB, including its outgoing | |
523 | edges. The copy is NEW_BB. Every PHI node in every direct successor of | |
524 | ORIG_BB has a new argument associated with edge from NEW_BB to the | |
525 | successor. Initialize the PHI argument so that it is equal to the PHI | |
1b83c31b | 526 | argument associated with the edge from ORIG_BB to the successor. |
527 | PATH and IDX are used to check if the new PHI argument has constant | |
528 | value in a flow sensitive manner. */ | |
da81e0c5 | 529 | |
530 | static void | |
1b83c31b | 531 | update_destination_phis (basic_block orig_bb, basic_block new_bb, |
532 | vec<jump_thread_edge *> *path, int idx) | |
da81e0c5 | 533 | { |
534 | edge_iterator ei; | |
535 | edge e; | |
536 | ||
537 | FOR_EACH_EDGE (e, ei, orig_bb->succs) | |
538 | { | |
539 | edge e2 = find_edge (new_bb, e->dest); | |
1b83c31b | 540 | copy_phi_args (e->dest, e, e2, path, idx); |
da81e0c5 | 541 | } |
542 | } | |
543 | ||
778182c1 | 544 | /* Given a duplicate block and its single destination (both stored |
545 | in RD). Create an edge between the duplicate and its single | |
546 | destination. | |
547 | ||
548 | Add an additional argument to any PHI nodes at the single | |
1b83c31b | 549 | destination. IDX is the start node in jump threading path |
550 | we start to check to see if the new PHI argument has constant | |
551 | value along the jump threading path. */ | |
778182c1 | 552 | |
553 | static void | |
42b013bc | 554 | create_edge_and_update_destination_phis (struct redirection_data *rd, |
1b83c31b | 555 | basic_block bb, int idx) |
778182c1 | 556 | { |
720cfc43 | 557 | edge e = make_single_succ_edge (bb, rd->path->last ()->e->dest, EDGE_FALLTHRU); |
778182c1 | 558 | |
f9614b84 | 559 | rescan_loop_exit (e, true, false); |
eb31063a | 560 | |
e63988cc | 561 | /* We used to copy the thread path here. That was added in 2007 |
562 | and dutifully updated through the representation changes in 2013. | |
563 | ||
564 | In 2013 we added code to thread from an interior node through | |
565 | the backedge to another interior node. That runs after the code | |
566 | to thread through loop headers from outside the loop. | |
567 | ||
568 | The latter may delete edges in the CFG, including those | |
569 | which appeared in the jump threading path we copied here. Thus | |
570 | we'd end up using a dangling pointer. | |
571 | ||
572 | After reviewing the 2007/2011 code, I can't see how anything | |
573 | depended on copying the AUX field and clearly copying the jump | |
574 | threading path is problematical due to embedded edge pointers. | |
575 | It has been removed. */ | |
576 | e->aux = NULL; | |
421e19dd | 577 | |
778182c1 | 578 | /* If there are any PHI nodes at the destination of the outgoing edge |
579 | from the duplicate block, then we will need to add a new argument | |
580 | to them. The argument should have the same value as the argument | |
581 | associated with the outgoing edge stored in RD. */ | |
1b83c31b | 582 | copy_phi_args (e->dest, rd->path->last ()->e, e, rd->path, idx); |
da81e0c5 | 583 | } |
584 | ||
fc54aba7 | 585 | /* Look through PATH beginning at START and return TRUE if there are |
586 | any additional blocks that need to be duplicated. Otherwise, | |
587 | return FALSE. */ | |
588 | static bool | |
589 | any_remaining_duplicated_blocks (vec<jump_thread_edge *> *path, | |
590 | unsigned int start) | |
591 | { | |
592 | for (unsigned int i = start + 1; i < path->length (); i++) | |
593 | { | |
594 | if ((*path)[i]->type == EDGE_COPY_SRC_JOINER_BLOCK | |
595 | || (*path)[i]->type == EDGE_COPY_SRC_BLOCK) | |
596 | return true; | |
597 | } | |
598 | return false; | |
599 | } | |
600 | ||
30e432bb | 601 | |
205ce1aa | 602 | /* Compute the amount of profile count coming into the jump threading |
30e432bb | 603 | path stored in RD that we are duplicating, returned in PATH_IN_COUNT_PTR and |
604 | PATH_IN_FREQ_PTR, as well as the amount of counts flowing out of the | |
605 | duplicated path, returned in PATH_OUT_COUNT_PTR. LOCAL_INFO is used to | |
606 | identify blocks duplicated for jump threading, which have duplicated | |
607 | edges that need to be ignored in the analysis. Return true if path contains | |
608 | a joiner, false otherwise. | |
609 | ||
205ce1aa | 610 | In the non-joiner case, this is straightforward - all the counts |
30e432bb | 611 | flowing into the jump threading path should flow through the duplicated |
612 | block and out of the duplicated path. | |
613 | ||
614 | In the joiner case, it is very tricky. Some of the counts flowing into | |
615 | the original path go offpath at the joiner. The problem is that while | |
616 | we know how much total count goes off-path in the original control flow, | |
617 | we don't know how many of the counts corresponding to just the jump | |
618 | threading path go offpath at the joiner. | |
619 | ||
620 | For example, assume we have the following control flow and identified | |
621 | jump threading paths: | |
622 | ||
9943b198 | 623 | A B C |
624 | \ | / | |
625 | Ea \ |Eb / Ec | |
626 | \ | / | |
627 | v v v | |
628 | J <-- Joiner | |
629 | / \ | |
630 | Eoff/ \Eon | |
631 | / \ | |
632 | v v | |
633 | Soff Son <--- Normal | |
634 | /\ | |
635 | Ed/ \ Ee | |
636 | / \ | |
637 | v v | |
638 | D E | |
639 | ||
640 | Jump threading paths: A -> J -> Son -> D (path 1) | |
641 | C -> J -> Son -> E (path 2) | |
30e432bb | 642 | |
643 | Note that the control flow could be more complicated: | |
644 | - Each jump threading path may have more than one incoming edge. I.e. A and | |
645 | Ea could represent multiple incoming blocks/edges that are included in | |
646 | path 1. | |
647 | - There could be EDGE_NO_COPY_SRC_BLOCK edges after the joiner (either | |
648 | before or after the "normal" copy block). These are not duplicated onto | |
649 | the jump threading path, as they are single-successor. | |
650 | - Any of the blocks along the path may have other incoming edges that | |
651 | are not part of any jump threading path, but add profile counts along | |
652 | the path. | |
653 | ||
31e5d72d | 654 | In the above example, after all jump threading is complete, we will |
30e432bb | 655 | end up with the following control flow: |
656 | ||
31e5d72d | 657 | A B C |
658 | | | | | |
659 | Ea| |Eb |Ec | |
660 | | | | | |
661 | v v v | |
662 | Ja J Jc | |
663 | / \ / \Eon' / \ | |
9943b198 | 664 | Eona/ \ ---/---\-------- \Eonc |
31e5d72d | 665 | / \ / / \ \ |
9943b198 | 666 | v v v v v |
667 | Sona Soff Son Sonc | |
31e5d72d | 668 | \ /\ / |
9943b198 | 669 | \___________ / \ _____/ |
670 | \ / \/ | |
671 | vv v | |
672 | D E | |
30e432bb | 673 | |
674 | The main issue to notice here is that when we are processing path 1 | |
675 | (A->J->Son->D) we need to figure out the outgoing edge weights to | |
676 | the duplicated edges Ja->Sona and Ja->Soff, while ensuring that the | |
677 | sum of the incoming weights to D remain Ed. The problem with simply | |
678 | assuming that Ja (and Jc when processing path 2) has the same outgoing | |
679 | probabilities to its successors as the original block J, is that after | |
680 | all paths are processed and other edges/counts removed (e.g. none | |
681 | of Ec will reach D after processing path 2), we may end up with not | |
682 | enough count flowing along duplicated edge Sona->D. | |
683 | ||
684 | Therefore, in the case of a joiner, we keep track of all counts | |
685 | coming in along the current path, as well as from predecessors not | |
686 | on any jump threading path (Eb in the above example). While we | |
687 | first assume that the duplicated Eona for Ja->Sona has the same | |
688 | probability as the original, we later compensate for other jump | |
689 | threading paths that may eliminate edges. We do that by keep track | |
690 | of all counts coming into the original path that are not in a jump | |
691 | thread (Eb in the above example, but as noted earlier, there could | |
692 | be other predecessors incoming to the path at various points, such | |
693 | as at Son). Call this cumulative non-path count coming into the path | |
694 | before D as Enonpath. We then ensure that the count from Sona->D is as at | |
695 | least as big as (Ed - Enonpath), but no bigger than the minimum | |
696 | weight along the jump threading path. The probabilities of both the | |
697 | original and duplicated joiner block J and Ja will be adjusted | |
698 | accordingly after the updates. */ | |
699 | ||
700 | static bool | |
701 | compute_path_counts (struct redirection_data *rd, | |
9943b198 | 702 | ssa_local_info_t *local_info, |
db9cef39 | 703 | profile_count *path_in_count_ptr, |
36b59a3c | 704 | profile_count *path_out_count_ptr) |
30e432bb | 705 | { |
706 | edge e = rd->incoming_edges->e; | |
707 | vec<jump_thread_edge *> *path = THREAD_PATH (e); | |
708 | edge elast = path->last ()->e; | |
db9cef39 | 709 | profile_count nonpath_count = profile_count::zero (); |
30e432bb | 710 | bool has_joiner = false; |
db9cef39 | 711 | profile_count path_in_count = profile_count::zero (); |
30e432bb | 712 | |
713 | /* Start by accumulating incoming edge counts to the path's first bb | |
714 | into a couple buckets: | |
9943b198 | 715 | path_in_count: total count of incoming edges that flow into the |
716 | current path. | |
717 | nonpath_count: total count of incoming edges that are not | |
718 | flowing along *any* path. These are the counts | |
719 | that will still flow along the original path after | |
720 | all path duplication is done by potentially multiple | |
721 | calls to this routine. | |
30e432bb | 722 | (any other incoming edge counts are for a different jump threading |
723 | path that will be handled by a later call to this routine.) | |
724 | To make this easier, start by recording all incoming edges that flow into | |
725 | the current path in a bitmap. We could add up the path's incoming edge | |
726 | counts here, but we still need to walk all the first bb's incoming edges | |
727 | below to add up the counts of the other edges not included in this jump | |
728 | threading path. */ | |
729 | struct el *next, *el; | |
035def86 | 730 | auto_bitmap in_edge_srcs; |
30e432bb | 731 | for (el = rd->incoming_edges; el; el = next) |
732 | { | |
733 | next = el->next; | |
734 | bitmap_set_bit (in_edge_srcs, el->e->src->index); | |
735 | } | |
736 | edge ein; | |
737 | edge_iterator ei; | |
738 | FOR_EACH_EDGE (ein, ei, e->dest->preds) | |
739 | { | |
740 | vec<jump_thread_edge *> *ein_path = THREAD_PATH (ein); | |
741 | /* Simply check the incoming edge src against the set captured above. */ | |
742 | if (ein_path | |
9943b198 | 743 | && bitmap_bit_p (in_edge_srcs, (*ein_path)[0]->e->src->index)) |
744 | { | |
745 | /* It is necessary but not sufficient that the last path edges | |
746 | are identical. There may be different paths that share the | |
747 | same last path edge in the case where the last edge has a nocopy | |
748 | source block. */ | |
749 | gcc_assert (ein_path->last ()->e == elast); | |
ea5d3981 | 750 | path_in_count += ein->count (); |
9943b198 | 751 | } |
30e432bb | 752 | else if (!ein_path) |
9943b198 | 753 | { |
754 | /* Keep track of the incoming edges that are not on any jump-threading | |
755 | path. These counts will still flow out of original path after all | |
756 | jump threading is complete. */ | |
ea5d3981 | 757 | nonpath_count += ein->count (); |
9943b198 | 758 | } |
30e432bb | 759 | } |
664dd751 | 760 | |
30e432bb | 761 | /* Now compute the fraction of the total count coming into the first |
762 | path bb that is from the current threading path. */ | |
db9cef39 | 763 | profile_count total_count = e->dest->count; |
30e432bb | 764 | /* Handle incoming profile insanities. */ |
765 | if (total_count < path_in_count) | |
766 | path_in_count = total_count; | |
c3f62dc3 | 767 | profile_probability onpath_scale = path_in_count.probability_in (total_count); |
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; | |
ea5d3981 | 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)) | |
ea5d3981 | 813 | nonpath_count += ein->count (); |
31e5d72d | 814 | } |
815 | } | |
30e432bb | 816 | if (cur_count < path_out_count) |
9943b198 | 817 | path_out_count = cur_count; |
ea5d3981 | 818 | if (epath->count () < min_path_count) |
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 | |
ea5d3981 | 834 | difference between elast->count () and nonpath_count. Otherwise the edge |
30e432bb | 835 | counts after threading will not be sane. */ |
d07cbccc | 836 | if (local_info->need_profile_correction |
ea5d3981 | 837 | && has_joiner && path_out_count < elast->count () - nonpath_count) |
31e5d72d | 838 | { |
ea5d3981 | 839 | path_out_count = elast->count () - nonpath_count; |
31e5d72d | 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; | |
30e432bb | 849 | return has_joiner; |
850 | } | |
851 | ||
852 | ||
853 | /* Update the counts and frequencies for both an original path | |
854 | edge EPATH and its duplicate EDUP. The duplicate source block | |
205ce1aa | 855 | will get a count of PATH_IN_COUNT and PATH_IN_FREQ, |
30e432bb | 856 | and the duplicate edge EDUP will have a count of PATH_OUT_COUNT. */ |
857 | static void | |
db9cef39 | 858 | update_profile (edge epath, edge edup, profile_count path_in_count, |
205ce1aa | 859 | profile_count path_out_count) |
30e432bb | 860 | { |
861 | ||
205ce1aa | 862 | /* First update the duplicated block's count. */ |
30e432bb | 863 | if (edup) |
864 | { | |
865 | basic_block dup_block = edup->src; | |
ea5d3981 | 866 | |
867 | /* Edup's count is reduced by path_out_count. We need to redistribute | |
868 | probabilities to the remaining edges. */ | |
869 | ||
870 | edge esucc; | |
871 | edge_iterator ei; | |
872 | profile_probability edup_prob | |
873 | = path_out_count.probability_in (path_in_count); | |
874 | ||
875 | /* Either scale up or down the remaining edges. | |
876 | probabilities are always in range <0,1> and thus we can't do | |
877 | both by same loop. */ | |
878 | if (edup->probability > edup_prob) | |
879 | { | |
880 | profile_probability rev_scale | |
881 | = (profile_probability::always () - edup->probability) | |
882 | / (profile_probability::always () - edup_prob); | |
883 | FOR_EACH_EDGE (esucc, ei, dup_block->succs) | |
884 | if (esucc != edup) | |
885 | esucc->probability /= rev_scale; | |
886 | } | |
887 | else if (edup->probability < edup_prob) | |
888 | { | |
889 | profile_probability scale | |
890 | = (profile_probability::always () - edup_prob) | |
891 | / (profile_probability::always () - edup->probability); | |
892 | FOR_EACH_EDGE (esucc, ei, dup_block->succs) | |
893 | if (esucc != edup) | |
894 | esucc->probability *= scale; | |
895 | } | |
205ce1aa | 896 | if (edup_prob.initialized_p ()) |
897 | edup->probability = edup_prob; | |
ea5d3981 | 898 | |
205ce1aa | 899 | gcc_assert (!dup_block->count.initialized_p ()); |
30e432bb | 900 | dup_block->count = path_in_count; |
30e432bb | 901 | } |
902 | ||
205ce1aa | 903 | if (path_in_count == profile_count::zero ()) |
904 | return; | |
905 | ||
ea5d3981 | 906 | profile_count final_count = epath->count () - path_out_count; |
907 | ||
205ce1aa | 908 | /* Now update the original block's count in the |
30e432bb | 909 | opposite manner - remove the counts/freq that will flow |
910 | into the duplicated block. Handle underflow due to precision/ | |
911 | rounding issues. */ | |
912 | epath->src->count -= path_in_count; | |
30e432bb | 913 | |
914 | /* Next update this path edge's original and duplicated counts. We know | |
915 | that the duplicated path will have path_out_count flowing | |
916 | out of it (in the joiner case this is the count along the duplicated path | |
917 | out of the duplicated joiner). This count can then be removed from the | |
918 | original path edge. */ | |
25d2128b | 919 | |
205ce1aa | 920 | edge esucc; |
ea5d3981 | 921 | edge_iterator ei; |
205ce1aa | 922 | profile_probability epath_prob = final_count.probability_in (epath->src->count); |
ea5d3981 | 923 | |
205ce1aa | 924 | if (epath->probability > epath_prob) |
30e432bb | 925 | { |
205ce1aa | 926 | profile_probability rev_scale |
927 | = (profile_probability::always () - epath->probability) | |
928 | / (profile_probability::always () - epath_prob); | |
929 | FOR_EACH_EDGE (esucc, ei, epath->src->succs) | |
930 | if (esucc != epath) | |
931 | esucc->probability /= rev_scale; | |
30e432bb | 932 | } |
205ce1aa | 933 | else if (epath->probability < epath_prob) |
30e432bb | 934 | { |
205ce1aa | 935 | profile_probability scale |
936 | = (profile_probability::always () - epath_prob) | |
937 | / (profile_probability::always () - epath->probability); | |
938 | FOR_EACH_EDGE (esucc, ei, epath->src->succs) | |
939 | if (esucc != epath) | |
940 | esucc->probability *= scale; | |
30e432bb | 941 | } |
205ce1aa | 942 | if (epath_prob.initialized_p ()) |
943 | epath->probability = epath_prob; | |
30e432bb | 944 | } |
945 | ||
fc54aba7 | 946 | /* Wire up the outgoing edges from the duplicate blocks and |
30e432bb | 947 | update any PHIs as needed. Also update the profile counts |
948 | on the original and duplicate blocks and edges. */ | |
2b15d2ba | 949 | void |
950 | ssa_fix_duplicate_block_edges (struct redirection_data *rd, | |
951 | ssa_local_info_t *local_info) | |
da81e0c5 | 952 | { |
1b83c31b | 953 | bool multi_incomings = (rd->incoming_edges->next != NULL); |
f2981b08 | 954 | edge e = rd->incoming_edges->e; |
955 | vec<jump_thread_edge *> *path = THREAD_PATH (e); | |
30e432bb | 956 | edge elast = path->last ()->e; |
db9cef39 | 957 | profile_count path_in_count = profile_count::zero (); |
958 | profile_count path_out_count = profile_count::zero (); | |
30e432bb | 959 | |
30e432bb | 960 | /* First determine how much profile count to move from original |
961 | path to the duplicate path. This is tricky in the presence of | |
962 | a joiner (see comments for compute_path_counts), where some portion | |
963 | of the path's counts will flow off-path from the joiner. In the | |
964 | non-joiner case the path_in_count and path_out_count should be the | |
965 | same. */ | |
966 | bool has_joiner = compute_path_counts (rd, local_info, | |
36b59a3c | 967 | &path_in_count, &path_out_count); |
30e432bb | 968 | |
fc54aba7 | 969 | for (unsigned int count = 0, i = 1; i < path->length (); i++) |
1b83778e | 970 | { |
30e432bb | 971 | edge epath = (*path)[i]->e; |
972 | ||
fc54aba7 | 973 | /* If we were threading through an joiner block, then we want |
974 | to keep its control statement and redirect an outgoing edge. | |
975 | Else we want to remove the control statement & edges, then create | |
976 | a new outgoing edge. In both cases we may need to update PHIs. */ | |
977 | if ((*path)[i]->type == EDGE_COPY_SRC_JOINER_BLOCK) | |
978 | { | |
979 | edge victim; | |
980 | edge e2; | |
981 | ||
9943b198 | 982 | gcc_assert (has_joiner); |
30e432bb | 983 | |
fc54aba7 | 984 | /* This updates the PHIs at the destination of the duplicate |
1b83c31b | 985 | block. Pass 0 instead of i if we are threading a path which |
986 | has multiple incoming edges. */ | |
987 | update_destination_phis (local_info->bb, rd->dup_blocks[count], | |
988 | path, multi_incomings ? 0 : i); | |
fc54aba7 | 989 | |
990 | /* Find the edge from the duplicate block to the block we're | |
991 | threading through. That's the edge we want to redirect. */ | |
992 | victim = find_edge (rd->dup_blocks[count], (*path)[i]->e->dest); | |
993 | ||
994 | /* If there are no remaining blocks on the path to duplicate, | |
995 | then redirect VICTIM to the final destination of the jump | |
996 | threading path. */ | |
997 | if (!any_remaining_duplicated_blocks (path, i)) | |
998 | { | |
30e432bb | 999 | e2 = redirect_edge_and_branch (victim, elast->dest); |
fc54aba7 | 1000 | /* If we redirected the edge, then we need to copy PHI arguments |
559685be | 1001 | at the target. If the edge already existed (e2 != victim |
fc54aba7 | 1002 | case), then the PHIs in the target already have the correct |
1003 | arguments. */ | |
1004 | if (e2 == victim) | |
30e432bb | 1005 | copy_phi_args (e2->dest, elast, e2, |
1b83c31b | 1006 | path, multi_incomings ? 0 : i); |
fc54aba7 | 1007 | } |
1008 | else | |
1009 | { | |
1010 | /* Redirect VICTIM to the next duplicated block in the path. */ | |
1011 | e2 = redirect_edge_and_branch (victim, rd->dup_blocks[count + 1]); | |
1012 | ||
1013 | /* We need to update the PHIs in the next duplicated block. We | |
1014 | want the new PHI args to have the same value as they had | |
1015 | in the source of the next duplicate block. | |
1016 | ||
1017 | Thus, we need to know which edge we traversed into the | |
1018 | source of the duplicate. Furthermore, we may have | |
1019 | traversed many edges to reach the source of the duplicate. | |
1020 | ||
1021 | Walk through the path starting at element I until we | |
1022 | hit an edge marked with EDGE_COPY_SRC_BLOCK. We want | |
1023 | the edge from the prior element. */ | |
1024 | for (unsigned int j = i + 1; j < path->length (); j++) | |
1025 | { | |
1026 | if ((*path)[j]->type == EDGE_COPY_SRC_BLOCK) | |
1027 | { | |
1028 | copy_phi_arg_into_existing_phi ((*path)[j - 1]->e, e2); | |
1029 | break; | |
1030 | } | |
1031 | } | |
1032 | } | |
30e432bb | 1033 | |
205ce1aa | 1034 | /* Update the counts of both the original block |
30e432bb | 1035 | and path edge, and the duplicates. The path duplicate's |
205ce1aa | 1036 | incoming count are the totals for all edges |
30e432bb | 1037 | incoming to this jump threading path computed earlier. |
1038 | And we know that the duplicated path will have path_out_count | |
1039 | flowing out of it (i.e. along the duplicated path out of the | |
1040 | duplicated joiner). */ | |
205ce1aa | 1041 | update_profile (epath, e2, path_in_count, path_out_count); |
fc54aba7 | 1042 | } |
1043 | else if ((*path)[i]->type == EDGE_COPY_SRC_BLOCK) | |
1044 | { | |
1045 | remove_ctrl_stmt_and_useless_edges (rd->dup_blocks[count], NULL); | |
1b83c31b | 1046 | create_edge_and_update_destination_phis (rd, rd->dup_blocks[count], |
1047 | multi_incomings ? 0 : i); | |
fc54aba7 | 1048 | if (count == 1) |
1049 | single_succ_edge (rd->dup_blocks[1])->aux = NULL; | |
30e432bb | 1050 | |
205ce1aa | 1051 | /* Update the counts of both the original block |
30e432bb | 1052 | and path edge, and the duplicates. Since we are now after |
1053 | any joiner that may have existed on the path, the count | |
1054 | flowing along the duplicated threaded path is path_out_count. | |
1055 | If we didn't have a joiner, then cur_path_freq was the sum | |
1056 | of the total frequencies along all incoming edges to the | |
1057 | thread path (path_in_freq). If we had a joiner, it would have | |
1058 | been updated at the end of that handling to the edge frequency | |
1059 | along the duplicated joiner path edge. */ | |
1060 | update_profile (epath, EDGE_SUCC (rd->dup_blocks[count], 0), | |
205ce1aa | 1061 | path_out_count, path_out_count); |
fc54aba7 | 1062 | } |
30e432bb | 1063 | else |
9943b198 | 1064 | { |
30e432bb | 1065 | /* No copy case. In this case we don't have an equivalent block |
1066 | on the duplicated thread path to update, but we do need | |
1067 | to remove the portion of the counts/freqs that were moved | |
1068 | to the duplicated path from the counts/freqs flowing through | |
1069 | this block on the original path. Since all the no-copy edges | |
1070 | are after any joiner, the removed count is the same as | |
1071 | path_out_count. | |
1072 | ||
1073 | If we didn't have a joiner, then cur_path_freq was the sum | |
1074 | of the total frequencies along all incoming edges to the | |
1075 | thread path (path_in_freq). If we had a joiner, it would have | |
1076 | been updated at the end of that handling to the edge frequency | |
1077 | along the duplicated joiner path edge. */ | |
205ce1aa | 1078 | update_profile (epath, NULL, path_out_count, path_out_count); |
30e432bb | 1079 | } |
1080 | ||
1081 | /* Increment the index into the duplicated path when we processed | |
9943b198 | 1082 | a duplicated block. */ |
30e432bb | 1083 | if ((*path)[i]->type == EDGE_COPY_SRC_JOINER_BLOCK |
9943b198 | 1084 | || (*path)[i]->type == EDGE_COPY_SRC_BLOCK) |
31e5d72d | 1085 | { |
30e432bb | 1086 | count++; |
31e5d72d | 1087 | } |
30e432bb | 1088 | } |
778182c1 | 1089 | } |
fc54aba7 | 1090 | |
778182c1 | 1091 | /* Hash table traversal callback routine to create duplicate blocks. */ |
1092 | ||
2b15d2ba | 1093 | int |
1094 | ssa_create_duplicates (struct redirection_data **slot, | |
1095 | ssa_local_info_t *local_info) | |
778182c1 | 1096 | { |
2b15d2ba | 1097 | struct redirection_data *rd = *slot; |
778182c1 | 1098 | |
11af02d8 | 1099 | /* The second duplicated block in a jump threading path is specific |
1b83778e | 1100 | to the path. So it gets stored in RD rather than in LOCAL_DATA. |
559685be | 1101 | |
11af02d8 | 1102 | Each time we're called, we have to look through the path and see |
1b83778e | 1103 | if a second block needs to be duplicated. |
11af02d8 | 1104 | |
1105 | Note the search starts with the third edge on the path. The first | |
1106 | edge is the incoming edge, the second edge always has its source | |
1107 | duplicated. Thus we start our search with the third edge. */ | |
a7ee7309 | 1108 | vec<jump_thread_edge *> *path = rd->path; |
11af02d8 | 1109 | for (unsigned int i = 2; i < path->length (); i++) |
1110 | { | |
1111 | if ((*path)[i]->type == EDGE_COPY_SRC_BLOCK | |
1112 | || (*path)[i]->type == EDGE_COPY_SRC_JOINER_BLOCK) | |
1113 | { | |
30e432bb | 1114 | create_block_for_threading ((*path)[i]->e->src, rd, 1, |
a7ee7309 | 1115 | &local_info->duplicate_blocks); |
11af02d8 | 1116 | break; |
1117 | } | |
1118 | } | |
1b83778e | 1119 | |
778182c1 | 1120 | /* Create a template block if we have not done so already. Otherwise |
1121 | use the template to create a new block. */ | |
1122 | if (local_info->template_block == NULL) | |
1123 | { | |
30e432bb | 1124 | create_block_for_threading ((*path)[1]->e->src, rd, 0, |
a7ee7309 | 1125 | &local_info->duplicate_blocks); |
11af02d8 | 1126 | local_info->template_block = rd->dup_blocks[0]; |
778182c1 | 1127 | |
1128 | /* We do not create any outgoing edges for the template. We will | |
1129 | take care of that in a later traversal. That way we do not | |
1130 | create edges that are going to just be deleted. */ | |
1131 | } | |
1132 | else | |
1133 | { | |
30e432bb | 1134 | create_block_for_threading (local_info->template_block, rd, 0, |
a7ee7309 | 1135 | &local_info->duplicate_blocks); |
778182c1 | 1136 | |
1137 | /* Go ahead and wire up outgoing edges and update PHIs for the duplicate | |
da81e0c5 | 1138 | block. */ |
2b15d2ba | 1139 | ssa_fix_duplicate_block_edges (rd, local_info); |
778182c1 | 1140 | } |
1141 | ||
1142 | /* Keep walking the hash table. */ | |
1143 | return 1; | |
1144 | } | |
1145 | ||
1146 | /* We did not create any outgoing edges for the template block during | |
1147 | block creation. This hash table traversal callback creates the | |
1148 | outgoing edge for the template block. */ | |
1149 | ||
2b15d2ba | 1150 | inline int |
1151 | ssa_fixup_template_block (struct redirection_data **slot, | |
1152 | ssa_local_info_t *local_info) | |
778182c1 | 1153 | { |
2b15d2ba | 1154 | struct redirection_data *rd = *slot; |
778182c1 | 1155 | |
da81e0c5 | 1156 | /* If this is the template block halt the traversal after updating |
1157 | it appropriately. | |
1158 | ||
1159 | If we were threading through an joiner block, then we want | |
1160 | to keep its control statement and redirect an outgoing edge. | |
1161 | Else we want to remove the control statement & edges, then create | |
1162 | a new outgoing edge. In both cases we may need to update PHIs. */ | |
11af02d8 | 1163 | if (rd->dup_blocks[0] && rd->dup_blocks[0] == local_info->template_block) |
778182c1 | 1164 | { |
2b15d2ba | 1165 | ssa_fix_duplicate_block_edges (rd, local_info); |
778182c1 | 1166 | return 0; |
1167 | } | |
1168 | ||
1169 | return 1; | |
1170 | } | |
1171 | ||
1172 | /* Hash table traversal callback to redirect each incoming edge | |
1173 | associated with this hash table element to its new destination. */ | |
1174 | ||
2b15d2ba | 1175 | int |
1176 | ssa_redirect_edges (struct redirection_data **slot, | |
1177 | ssa_local_info_t *local_info) | |
778182c1 | 1178 | { |
2b15d2ba | 1179 | struct redirection_data *rd = *slot; |
778182c1 | 1180 | struct el *next, *el; |
1181 | ||
47ae02b7 | 1182 | /* Walk over all the incoming edges associated with this hash table |
1183 | entry. */ | |
778182c1 | 1184 | for (el = rd->incoming_edges; el; el = next) |
1185 | { | |
1186 | edge e = el->e; | |
f2981b08 | 1187 | vec<jump_thread_edge *> *path = THREAD_PATH (e); |
778182c1 | 1188 | |
1189 | /* Go ahead and free this element from the list. Doing this now | |
1190 | avoids the need for another list walk when we destroy the hash | |
1191 | table. */ | |
1192 | next = el->next; | |
1193 | free (el); | |
1194 | ||
5236b8bb | 1195 | thread_stats.num_threaded_edges++; |
1196 | ||
11af02d8 | 1197 | if (rd->dup_blocks[0]) |
778182c1 | 1198 | { |
1199 | edge e2; | |
1200 | ||
1201 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1202 | fprintf (dump_file, " Threaded jump %d --> %d to %d\n", | |
11af02d8 | 1203 | e->src->index, e->dest->index, rd->dup_blocks[0]->index); |
778182c1 | 1204 | |
353f9f16 | 1205 | /* Redirect the incoming edge (possibly to the joiner block) to the |
1206 | appropriate duplicate block. */ | |
11af02d8 | 1207 | e2 = redirect_edge_and_branch (e, rd->dup_blocks[0]); |
7e0311ae | 1208 | gcc_assert (e == e2); |
778182c1 | 1209 | flush_pending_stmts (e2); |
778182c1 | 1210 | } |
eb31063a | 1211 | |
1212 | /* Go ahead and clear E->aux. It's not needed anymore and failure | |
559685be | 1213 | to clear it will cause all kinds of unpleasant problems later. */ |
6d1fdbf9 | 1214 | delete_jump_thread_path (path); |
eb31063a | 1215 | e->aux = NULL; |
1216 | ||
778182c1 | 1217 | } |
388d1fc1 | 1218 | |
1219 | /* Indicate that we actually threaded one or more jumps. */ | |
1220 | if (rd->incoming_edges) | |
1221 | local_info->jumps_threaded = true; | |
1222 | ||
778182c1 | 1223 | return 1; |
1224 | } | |
1225 | ||
aed95130 | 1226 | /* Return true if this block has no executable statements other than |
1227 | a simple ctrl flow instruction. When the number of outgoing edges | |
1228 | is one, this is equivalent to a "forwarder" block. */ | |
1229 | ||
1230 | static bool | |
47aaf6e6 | 1231 | redirection_block_p (basic_block bb) |
aed95130 | 1232 | { |
75a70cf9 | 1233 | gimple_stmt_iterator gsi; |
aed95130 | 1234 | |
1235 | /* Advance to the first executable statement. */ | |
75a70cf9 | 1236 | gsi = gsi_start_bb (bb); |
1237 | while (!gsi_end_p (gsi) | |
559685be | 1238 | && (gimple_code (gsi_stmt (gsi)) == GIMPLE_LABEL |
9845d120 | 1239 | || is_gimple_debug (gsi_stmt (gsi)) |
c7566ddf | 1240 | || gimple_nop_p (gsi_stmt (gsi)) |
581c1a3c | 1241 | || gimple_clobber_p (gsi_stmt (gsi)))) |
75a70cf9 | 1242 | gsi_next (&gsi); |
48e1416a | 1243 | |
aed95130 | 1244 | /* Check if this is an empty block. */ |
75a70cf9 | 1245 | if (gsi_end_p (gsi)) |
aed95130 | 1246 | return true; |
1247 | ||
1248 | /* Test that we've reached the terminating control statement. */ | |
75a70cf9 | 1249 | return gsi_stmt (gsi) |
559685be | 1250 | && (gimple_code (gsi_stmt (gsi)) == GIMPLE_COND |
1251 | || gimple_code (gsi_stmt (gsi)) == GIMPLE_GOTO | |
1252 | || gimple_code (gsi_stmt (gsi)) == GIMPLE_SWITCH); | |
aed95130 | 1253 | } |
1254 | ||
a8046f60 | 1255 | /* BB is a block which ends with a COND_EXPR or SWITCH_EXPR and when BB |
1256 | is reached via one or more specific incoming edges, we know which | |
1257 | outgoing edge from BB will be traversed. | |
1258 | ||
778182c1 | 1259 | We want to redirect those incoming edges to the target of the |
a8046f60 | 1260 | appropriate outgoing edge. Doing so avoids a conditional branch |
1261 | and may expose new optimization opportunities. Note that we have | |
1262 | to update dominator tree and SSA graph after such changes. | |
1263 | ||
597ff315 | 1264 | The key to keeping the SSA graph update manageable is to duplicate |
91275768 | 1265 | the side effects occurring in BB so that those side effects still |
a8046f60 | 1266 | occur on the paths which bypass BB after redirecting edges. |
1267 | ||
1268 | We accomplish this by creating duplicates of BB and arranging for | |
1269 | the duplicates to unconditionally pass control to one specific | |
1270 | successor of BB. We then revector the incoming edges into BB to | |
1271 | the appropriate duplicate of BB. | |
1272 | ||
7e0311ae | 1273 | If NOLOOP_ONLY is true, we only perform the threading as long as it |
1b83778e | 1274 | does not affect the structure of the loops in a nontrivial way. |
ed4feca1 | 1275 | |
1276 | If JOINERS is true, then thread through joiner blocks as well. */ | |
a8046f60 | 1277 | |
388d1fc1 | 1278 | static bool |
ed4feca1 | 1279 | thread_block_1 (basic_block bb, bool noloop_only, bool joiners) |
a8046f60 | 1280 | { |
1281 | /* E is an incoming edge into BB that we may or may not want to | |
1282 | redirect to a duplicate of BB. */ | |
7e0311ae | 1283 | edge e, e2; |
cd665a06 | 1284 | edge_iterator ei; |
2b15d2ba | 1285 | ssa_local_info_t local_info; |
388d1fc1 | 1286 | |
30e432bb | 1287 | local_info.duplicate_blocks = BITMAP_ALLOC (NULL); |
d07cbccc | 1288 | local_info.need_profile_correction = false; |
30e432bb | 1289 | |
778182c1 | 1290 | /* To avoid scanning a linear array for the element we need we instead |
c5d4a10b | 1291 | use a hash table. For normal code there should be no noticeable |
778182c1 | 1292 | difference. However, if we have a block with a large number of |
1293 | incoming and outgoing edges such linear searches can get expensive. */ | |
c1f445d2 | 1294 | redirection_data |
1295 | = new hash_table<struct redirection_data> (EDGE_COUNT (bb->succs)); | |
778182c1 | 1296 | |
1297 | /* Record each unique threaded destination into a hash table for | |
1298 | efficient lookups. */ | |
d07cbccc | 1299 | edge last = NULL; |
cd665a06 | 1300 | FOR_EACH_EDGE (e, ei, bb->preds) |
a8046f60 | 1301 | { |
eb31063a | 1302 | if (e->aux == NULL) |
1303 | continue; | |
1304 | ||
f2981b08 | 1305 | vec<jump_thread_edge *> *path = THREAD_PATH (e); |
ed4feca1 | 1306 | |
1307 | if (((*path)[1]->type == EDGE_COPY_SRC_JOINER_BLOCK && !joiners) | |
1308 | || ((*path)[1]->type == EDGE_COPY_SRC_BLOCK && joiners)) | |
1309 | continue; | |
1310 | ||
f2981b08 | 1311 | e2 = path->last ()->e; |
e2b72d6c | 1312 | if (!e2 || noloop_only) |
1313 | { | |
7e0311ae | 1314 | /* If NOLOOP_ONLY is true, we only allow threading through the |
559685be | 1315 | header of a loop to exit edges. */ |
e2b72d6c | 1316 | |
559685be | 1317 | /* One case occurs when there was loop header buried in a jump |
1318 | threading path that crosses loop boundaries. We do not try | |
1319 | and thread this elsewhere, so just cancel the jump threading | |
1320 | request by clearing the AUX field now. */ | |
9f3abcb8 | 1321 | if (bb->loop_father != e2->src->loop_father |
9a56d038 | 1322 | && (!loop_exit_edge_p (e2->src->loop_father, e2) |
1323 | || flow_loop_nested_p (bb->loop_father, | |
1324 | e2->dest->loop_father))) | |
e2b72d6c | 1325 | { |
1326 | /* Since this case is not handled by our special code | |
1327 | to thread through a loop header, we must explicitly | |
1328 | cancel the threading request here. */ | |
6d1fdbf9 | 1329 | delete_jump_thread_path (path); |
e2b72d6c | 1330 | e->aux = NULL; |
1331 | continue; | |
1332 | } | |
559685be | 1333 | |
1334 | /* Another case occurs when trying to thread through our | |
ab596744 | 1335 | own loop header, possibly from inside the loop. We will |
1336 | thread these later. */ | |
559685be | 1337 | unsigned int i; |
1338 | for (i = 1; i < path->length (); i++) | |
1339 | { | |
1340 | if ((*path)[i]->e->src == bb->loop_father->header | |
1341 | && (!loop_exit_edge_p (bb->loop_father, e2) | |
1342 | || (*path)[1]->type == EDGE_COPY_SRC_JOINER_BLOCK)) | |
ab596744 | 1343 | break; |
559685be | 1344 | } |
1345 | ||
1346 | if (i != path->length ()) | |
1347 | continue; | |
76cf4406 | 1348 | |
1349 | /* Loop parallelization can be confused by the result of | |
1350 | threading through the loop exit test back into the loop. | |
1351 | However, theading those jumps seems to help other codes. | |
1352 | ||
1353 | I have been unable to find anything related to the shape of | |
1354 | the CFG, the contents of the affected blocks, etc which would | |
1355 | allow a more sensible test than what we're using below which | |
1356 | merely avoids the optimization when parallelizing loops. */ | |
1357 | if (flag_tree_parallelize_loops > 1) | |
1358 | { | |
1359 | for (i = 1; i < path->length (); i++) | |
1360 | if (bb->loop_father == e2->src->loop_father | |
1361 | && loop_exits_from_bb_p (bb->loop_father, | |
1362 | (*path)[i]->e->src) | |
1363 | && !loop_exit_edge_p (bb->loop_father, e2)) | |
1364 | break; | |
1365 | ||
1366 | if (i != path->length ()) | |
1367 | { | |
1368 | delete_jump_thread_path (path); | |
1369 | e->aux = NULL; | |
1370 | continue; | |
1371 | } | |
1372 | } | |
e2b72d6c | 1373 | } |
778182c1 | 1374 | |
7e0311ae | 1375 | /* Insert the outgoing edge into the hash table if it is not |
1376 | already in the hash table. */ | |
da81e0c5 | 1377 | lookup_redirection_data (e, INSERT); |
d07cbccc | 1378 | |
1379 | /* When we have thread paths through a common joiner with different | |
1380 | final destinations, then we may need corrections to deal with | |
1381 | profile insanities. See the big comment before compute_path_counts. */ | |
1382 | if ((*path)[1]->type == EDGE_COPY_SRC_JOINER_BLOCK) | |
1383 | { | |
1384 | if (!last) | |
1385 | last = e2; | |
1386 | else if (e2 != last) | |
1387 | local_info.need_profile_correction = true; | |
1388 | } | |
a8046f60 | 1389 | } |
1390 | ||
3f9439d7 | 1391 | /* We do not update dominance info. */ |
1392 | free_dominance_info (CDI_DOMINATORS); | |
1393 | ||
d906930c | 1394 | /* We know we only thread through the loop header to loop exits. |
1395 | Let the basic block duplication hook know we are not creating | |
1396 | a multiple entry loop. */ | |
1397 | if (noloop_only | |
1398 | && bb == bb->loop_father->header) | |
1399 | set_loop_copy (bb->loop_father, loop_outer (bb->loop_father)); | |
1400 | ||
778182c1 | 1401 | /* Now create duplicates of BB. |
f582bb6c | 1402 | |
1403 | Note that for a block with a high outgoing degree we can waste | |
1404 | a lot of time and memory creating and destroying useless edges. | |
1405 | ||
1406 | So we first duplicate BB and remove the control structure at the | |
1407 | tail of the duplicate as well as all outgoing edges from the | |
1408 | duplicate. We then use that duplicate block as a template for | |
1409 | the rest of the duplicates. */ | |
778182c1 | 1410 | local_info.template_block = NULL; |
1411 | local_info.bb = bb; | |
388d1fc1 | 1412 | local_info.jumps_threaded = false; |
c1f445d2 | 1413 | redirection_data->traverse <ssa_local_info_t *, ssa_create_duplicates> |
2b15d2ba | 1414 | (&local_info); |
f582bb6c | 1415 | |
778182c1 | 1416 | /* The template does not have an outgoing edge. Create that outgoing |
1417 | edge and update PHI nodes as the edge's target as necessary. | |
f582bb6c | 1418 | |
778182c1 | 1419 | We do this after creating all the duplicates to avoid creating |
1420 | unnecessary edges. */ | |
c1f445d2 | 1421 | redirection_data->traverse <ssa_local_info_t *, ssa_fixup_template_block> |
2b15d2ba | 1422 | (&local_info); |
f582bb6c | 1423 | |
778182c1 | 1424 | /* The hash table traversals above created the duplicate blocks (and the |
1425 | statements within the duplicate blocks). This loop creates PHI nodes for | |
1426 | the duplicated blocks and redirects the incoming edges into BB to reach | |
1427 | the duplicates of BB. */ | |
c1f445d2 | 1428 | redirection_data->traverse <ssa_local_info_t *, ssa_redirect_edges> |
2b15d2ba | 1429 | (&local_info); |
a8046f60 | 1430 | |
a3d0fd80 | 1431 | /* Done with this block. Clear REDIRECTION_DATA. */ |
c1f445d2 | 1432 | delete redirection_data; |
1433 | redirection_data = NULL; | |
388d1fc1 | 1434 | |
d906930c | 1435 | if (noloop_only |
1436 | && bb == bb->loop_father->header) | |
1437 | set_loop_copy (bb->loop_father, NULL); | |
1438 | ||
30e432bb | 1439 | BITMAP_FREE (local_info.duplicate_blocks); |
1440 | local_info.duplicate_blocks = NULL; | |
1441 | ||
388d1fc1 | 1442 | /* Indicate to our caller whether or not any jumps were threaded. */ |
1443 | return local_info.jumps_threaded; | |
a8046f60 | 1444 | } |
1445 | ||
ed4feca1 | 1446 | /* Wrapper for thread_block_1 so that we can first handle jump |
1447 | thread paths which do not involve copying joiner blocks, then | |
1448 | handle jump thread paths which have joiner blocks. | |
1449 | ||
1450 | By doing things this way we can be as aggressive as possible and | |
1451 | not worry that copying a joiner block will create a jump threading | |
1452 | opportunity. */ | |
1b83778e | 1453 | |
ed4feca1 | 1454 | static bool |
1455 | thread_block (basic_block bb, bool noloop_only) | |
1456 | { | |
1457 | bool retval; | |
1458 | retval = thread_block_1 (bb, noloop_only, false); | |
1459 | retval |= thread_block_1 (bb, noloop_only, true); | |
1460 | return retval; | |
1461 | } | |
1462 | ||
7e0311ae | 1463 | /* Callback for dfs_enumerate_from. Returns true if BB is different |
1464 | from STOP and DBDS_CE_STOP. */ | |
1465 | ||
1466 | static basic_block dbds_ce_stop; | |
1467 | static bool | |
7ecb5bb2 | 1468 | dbds_continue_enumeration_p (const_basic_block bb, const void *stop) |
7e0311ae | 1469 | { |
7ecb5bb2 | 1470 | return (bb != (const_basic_block) stop |
7e0311ae | 1471 | && bb != dbds_ce_stop); |
1472 | } | |
1473 | ||
1474 | /* Evaluates the dominance relationship of latch of the LOOP and BB, and | |
1475 | returns the state. */ | |
1476 | ||
1477 | enum bb_dom_status | |
7e0311ae | 1478 | determine_bb_domination_status (struct loop *loop, basic_block bb) |
1479 | { | |
1480 | basic_block *bblocks; | |
1481 | unsigned nblocks, i; | |
1482 | bool bb_reachable = false; | |
1483 | edge_iterator ei; | |
1484 | edge e; | |
1485 | ||
42b013bc | 1486 | /* This function assumes BB is a successor of LOOP->header. |
1487 | If that is not the case return DOMST_NONDOMINATING which | |
1488 | is always safe. */ | |
7e0311ae | 1489 | { |
1490 | bool ok = false; | |
1491 | ||
1492 | FOR_EACH_EDGE (e, ei, bb->preds) | |
1493 | { | |
1494 | if (e->src == loop->header) | |
1495 | { | |
1496 | ok = true; | |
1497 | break; | |
1498 | } | |
1499 | } | |
1500 | ||
42b013bc | 1501 | if (!ok) |
1502 | return DOMST_NONDOMINATING; | |
7e0311ae | 1503 | } |
7e0311ae | 1504 | |
1505 | if (bb == loop->latch) | |
1506 | return DOMST_DOMINATING; | |
1507 | ||
1508 | /* Check that BB dominates LOOP->latch, and that it is back-reachable | |
1509 | from it. */ | |
1510 | ||
1511 | bblocks = XCNEWVEC (basic_block, loop->num_nodes); | |
1512 | dbds_ce_stop = loop->header; | |
1513 | nblocks = dfs_enumerate_from (loop->latch, 1, dbds_continue_enumeration_p, | |
1514 | bblocks, loop->num_nodes, bb); | |
1515 | for (i = 0; i < nblocks; i++) | |
1516 | FOR_EACH_EDGE (e, ei, bblocks[i]->preds) | |
1517 | { | |
1518 | if (e->src == loop->header) | |
1519 | { | |
1520 | free (bblocks); | |
1521 | return DOMST_NONDOMINATING; | |
1522 | } | |
1523 | if (e->src == bb) | |
1524 | bb_reachable = true; | |
1525 | } | |
1526 | ||
1527 | free (bblocks); | |
1528 | return (bb_reachable ? DOMST_DOMINATING : DOMST_LOOP_BROKEN); | |
1529 | } | |
1530 | ||
1531 | /* Thread jumps through the header of LOOP. Returns true if cfg changes. | |
1532 | If MAY_PEEL_LOOP_HEADERS is false, we avoid threading from entry edges | |
1533 | to the inside of the loop. */ | |
1534 | ||
1535 | static bool | |
1536 | thread_through_loop_header (struct loop *loop, bool may_peel_loop_headers) | |
1537 | { | |
1538 | basic_block header = loop->header; | |
1539 | edge e, tgt_edge, latch = loop_latch_edge (loop); | |
1540 | edge_iterator ei; | |
1541 | basic_block tgt_bb, atgt_bb; | |
1542 | enum bb_dom_status domst; | |
1543 | ||
1544 | /* We have already threaded through headers to exits, so all the threading | |
1545 | requests now are to the inside of the loop. We need to avoid creating | |
1546 | irreducible regions (i.e., loops with more than one entry block), and | |
1547 | also loop with several latch edges, or new subloops of the loop (although | |
1548 | there are cases where it might be appropriate, it is difficult to decide, | |
1549 | and doing it wrongly may confuse other optimizers). | |
1550 | ||
1551 | We could handle more general cases here. However, the intention is to | |
1552 | preserve some information about the loop, which is impossible if its | |
1553 | structure changes significantly, in a way that is not well understood. | |
1554 | Thus we only handle few important special cases, in which also updating | |
1555 | of the loop-carried information should be feasible: | |
1556 | ||
1557 | 1) Propagation of latch edge to a block that dominates the latch block | |
1558 | of a loop. This aims to handle the following idiom: | |
1559 | ||
1560 | first = 1; | |
1561 | while (1) | |
1562 | { | |
1563 | if (first) | |
1564 | initialize; | |
1565 | first = 0; | |
1566 | body; | |
1567 | } | |
1568 | ||
1569 | After threading the latch edge, this becomes | |
1570 | ||
1571 | first = 1; | |
1572 | if (first) | |
1573 | initialize; | |
1574 | while (1) | |
1575 | { | |
1576 | first = 0; | |
1577 | body; | |
1578 | } | |
1579 | ||
1580 | The original header of the loop is moved out of it, and we may thread | |
1581 | the remaining edges through it without further constraints. | |
1582 | ||
1583 | 2) All entry edges are propagated to a single basic block that dominates | |
1584 | the latch block of the loop. This aims to handle the following idiom | |
1585 | (normally created for "for" loops): | |
1586 | ||
1587 | i = 0; | |
1588 | while (1) | |
1589 | { | |
1590 | if (i >= 100) | |
1591 | break; | |
1592 | body; | |
1593 | i++; | |
1594 | } | |
1595 | ||
1596 | This becomes | |
1597 | ||
1598 | i = 0; | |
1599 | while (1) | |
1600 | { | |
1601 | body; | |
1602 | i++; | |
1603 | if (i >= 100) | |
1604 | break; | |
1605 | } | |
1606 | */ | |
1607 | ||
1608 | /* Threading through the header won't improve the code if the header has just | |
1609 | one successor. */ | |
1610 | if (single_succ_p (header)) | |
1611 | goto fail; | |
1612 | ||
d29932c9 | 1613 | if (!may_peel_loop_headers && !redirection_block_p (loop->header)) |
7e0311ae | 1614 | goto fail; |
1615 | else | |
1616 | { | |
1617 | tgt_bb = NULL; | |
1618 | tgt_edge = NULL; | |
1619 | FOR_EACH_EDGE (e, ei, header->preds) | |
1620 | { | |
1621 | if (!e->aux) | |
1622 | { | |
1623 | if (e == latch) | |
1624 | continue; | |
1625 | ||
1626 | /* If latch is not threaded, and there is a header | |
1627 | edge that is not threaded, we would create loop | |
1628 | with multiple entries. */ | |
1629 | goto fail; | |
1630 | } | |
1631 | ||
f2981b08 | 1632 | vec<jump_thread_edge *> *path = THREAD_PATH (e); |
1633 | ||
1634 | if ((*path)[1]->type == EDGE_COPY_SRC_JOINER_BLOCK) | |
da81e0c5 | 1635 | goto fail; |
f2981b08 | 1636 | tgt_edge = (*path)[1]->e; |
7e0311ae | 1637 | atgt_bb = tgt_edge->dest; |
1638 | if (!tgt_bb) | |
1639 | tgt_bb = atgt_bb; | |
1640 | /* Two targets of threading would make us create loop | |
1641 | with multiple entries. */ | |
1642 | else if (tgt_bb != atgt_bb) | |
1643 | goto fail; | |
1644 | } | |
1645 | ||
1646 | if (!tgt_bb) | |
1647 | { | |
1648 | /* There are no threading requests. */ | |
1649 | return false; | |
1650 | } | |
1651 | ||
1652 | /* Redirecting to empty loop latch is useless. */ | |
1653 | if (tgt_bb == loop->latch | |
1654 | && empty_block_p (loop->latch)) | |
1655 | goto fail; | |
1656 | } | |
1657 | ||
1658 | /* The target block must dominate the loop latch, otherwise we would be | |
1659 | creating a subloop. */ | |
1660 | domst = determine_bb_domination_status (loop, tgt_bb); | |
1661 | if (domst == DOMST_NONDOMINATING) | |
1662 | goto fail; | |
1663 | if (domst == DOMST_LOOP_BROKEN) | |
1664 | { | |
1665 | /* If the loop ceased to exist, mark it as such, and thread through its | |
1666 | original header. */ | |
d25159cc | 1667 | mark_loop_for_removal (loop); |
7e0311ae | 1668 | return thread_block (header, false); |
1669 | } | |
1670 | ||
1671 | if (tgt_bb->loop_father->header == tgt_bb) | |
1672 | { | |
1673 | /* If the target of the threading is a header of a subloop, we need | |
1674 | to create a preheader for it, so that the headers of the two loops | |
1675 | do not merge. */ | |
1676 | if (EDGE_COUNT (tgt_bb->preds) > 2) | |
1677 | { | |
1678 | tgt_bb = create_preheader (tgt_bb->loop_father, 0); | |
1679 | gcc_assert (tgt_bb != NULL); | |
1680 | } | |
1681 | else | |
1682 | tgt_bb = split_edge (tgt_edge); | |
1683 | } | |
48e1416a | 1684 | |
d29932c9 | 1685 | basic_block new_preheader; |
6eb99d8a | 1686 | |
d29932c9 | 1687 | /* Now consider the case entry edges are redirected to the new entry |
1688 | block. Remember one entry edge, so that we can find the new | |
1689 | preheader (its destination after threading). */ | |
1690 | FOR_EACH_EDGE (e, ei, header->preds) | |
7e0311ae | 1691 | { |
d29932c9 | 1692 | if (e->aux) |
1693 | break; | |
1694 | } | |
7e0311ae | 1695 | |
d29932c9 | 1696 | /* The duplicate of the header is the new preheader of the loop. Ensure |
1697 | that it is placed correctly in the loop hierarchy. */ | |
1698 | set_loop_copy (loop, loop_outer (loop)); | |
7e0311ae | 1699 | |
d29932c9 | 1700 | thread_block (header, false); |
1701 | set_loop_copy (loop, NULL); | |
1702 | new_preheader = e->dest; | |
48e1416a | 1703 | |
d29932c9 | 1704 | /* Create the new latch block. This is always necessary, as the latch |
1705 | must have only a single successor, but the original header had at | |
1706 | least two successors. */ | |
1707 | loop->latch = NULL; | |
1708 | mfb_kj_edge = single_succ_edge (new_preheader); | |
1709 | loop->header = mfb_kj_edge->dest; | |
1710 | latch = make_forwarder_block (tgt_bb, mfb_keep_just, NULL); | |
1711 | loop->header = latch->dest; | |
1712 | loop->latch = latch->src; | |
7e0311ae | 1713 | return true; |
1714 | ||
1715 | fail: | |
1716 | /* We failed to thread anything. Cancel the requests. */ | |
1717 | FOR_EACH_EDGE (e, ei, header->preds) | |
1718 | { | |
f2981b08 | 1719 | vec<jump_thread_edge *> *path = THREAD_PATH (e); |
1720 | ||
1721 | if (path) | |
1722 | { | |
6d1fdbf9 | 1723 | delete_jump_thread_path (path); |
f2981b08 | 1724 | e->aux = NULL; |
1725 | } | |
7e0311ae | 1726 | } |
1727 | return false; | |
1728 | } | |
1729 | ||
b99a7d6d | 1730 | /* E1 and E2 are edges into the same basic block. Return TRUE if the |
1731 | PHI arguments associated with those edges are equal or there are no | |
1732 | PHI arguments, otherwise return FALSE. */ | |
1733 | ||
1734 | static bool | |
1735 | phi_args_equal_on_edges (edge e1, edge e2) | |
1736 | { | |
1a91d914 | 1737 | gphi_iterator gsi; |
b99a7d6d | 1738 | int indx1 = e1->dest_idx; |
1739 | int indx2 = e2->dest_idx; | |
1740 | ||
1741 | for (gsi = gsi_start_phis (e1->dest); !gsi_end_p (gsi); gsi_next (&gsi)) | |
1742 | { | |
1a91d914 | 1743 | gphi *phi = gsi.phi (); |
b99a7d6d | 1744 | |
1745 | if (!operand_equal_p (gimple_phi_arg_def (phi, indx1), | |
1746 | gimple_phi_arg_def (phi, indx2), 0)) | |
1747 | return false; | |
1748 | } | |
1749 | return true; | |
1750 | } | |
1751 | ||
5f885dd1 | 1752 | /* Return the number of non-debug statements and non-virtual PHIs in a |
1753 | block. */ | |
1754 | ||
1755 | static unsigned int | |
1756 | count_stmts_and_phis_in_block (basic_block bb) | |
1757 | { | |
1758 | unsigned int num_stmts = 0; | |
1759 | ||
1760 | gphi_iterator gpi; | |
1761 | for (gpi = gsi_start_phis (bb); !gsi_end_p (gpi); gsi_next (&gpi)) | |
1762 | if (!virtual_operand_p (PHI_RESULT (gpi.phi ()))) | |
1763 | num_stmts++; | |
1764 | ||
1765 | gimple_stmt_iterator gsi; | |
1766 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
1767 | { | |
1768 | gimple *stmt = gsi_stmt (gsi); | |
1769 | if (!is_gimple_debug (stmt)) | |
1770 | num_stmts++; | |
1771 | } | |
1772 | ||
1773 | return num_stmts; | |
1774 | } | |
1775 | ||
1776 | ||
3cebc9d2 | 1777 | /* Walk through the registered jump threads and convert them into a |
334ec2d8 | 1778 | form convenient for this pass. |
3cebc9d2 | 1779 | |
1780 | Any block which has incoming edges threaded to outgoing edges | |
1781 | will have its entry in THREADED_BLOCK set. | |
a8046f60 | 1782 | |
3cebc9d2 | 1783 | Any threaded edge will have its new outgoing edge stored in the |
1784 | original edge's AUX field. | |
a8046f60 | 1785 | |
3cebc9d2 | 1786 | This form avoids the need to walk all the edges in the CFG to |
1787 | discover blocks which need processing and avoids unnecessary | |
1788 | hash table lookups to map from threaded edge to new target. */ | |
a8046f60 | 1789 | |
3cebc9d2 | 1790 | static void |
1791 | mark_threaded_blocks (bitmap threaded_blocks) | |
1792 | { | |
1793 | unsigned int i; | |
7e0311ae | 1794 | bitmap_iterator bi; |
035def86 | 1795 | auto_bitmap tmp; |
7e0311ae | 1796 | basic_block bb; |
1797 | edge e; | |
1798 | edge_iterator ei; | |
3cebc9d2 | 1799 | |
b93ba654 | 1800 | /* It is possible to have jump threads in which one is a subpath |
1801 | of the other. ie, (A, B), (B, C), (C, D) where B is a joiner | |
1802 | block and (B, C), (C, D) where no joiner block exists. | |
1803 | ||
1804 | When this occurs ignore the jump thread request with the joiner | |
1805 | block. It's totally subsumed by the simpler jump thread request. | |
1806 | ||
1807 | This results in less block copying, simpler CFGs. More importantly, | |
1808 | when we duplicate the joiner block, B, in this case we will create | |
1809 | a new threading opportunity that we wouldn't be able to optimize | |
1810 | until the next jump threading iteration. | |
1811 | ||
1812 | So first convert the jump thread requests which do not require a | |
1813 | joiner block. */ | |
f2981b08 | 1814 | for (i = 0; i < paths.length (); i++) |
3cebc9d2 | 1815 | { |
f2981b08 | 1816 | vec<jump_thread_edge *> *path = paths[i]; |
b93ba654 | 1817 | |
ff92f40d | 1818 | if (path->length () > 1 |
1819 | && (*path)[1]->type != EDGE_COPY_SRC_JOINER_BLOCK) | |
b93ba654 | 1820 | { |
1821 | edge e = (*path)[0]->e; | |
1822 | e->aux = (void *)path; | |
1823 | bitmap_set_bit (tmp, e->dest->index); | |
1824 | } | |
1f3976e7 | 1825 | } |
1826 | ||
b93ba654 | 1827 | /* Now iterate again, converting cases where we want to thread |
1828 | through a joiner block, but only if no other edge on the path | |
f1ce4e72 | 1829 | already has a jump thread attached to it. We do this in two passes, |
1830 | to avoid situations where the order in the paths vec can hide overlapping | |
1831 | threads (the path is recorded on the incoming edge, so we would miss | |
1832 | cases where the second path starts at a downstream edge on the same | |
1833 | path). First record all joiner paths, deleting any in the unexpected | |
1834 | case where there is already a path for that incoming edge. */ | |
51ea8bc6 | 1835 | for (i = 0; i < paths.length ();) |
b93ba654 | 1836 | { |
1837 | vec<jump_thread_edge *> *path = paths[i]; | |
1838 | ||
ff92f40d | 1839 | if (path->length () > 1 |
1840 | && (*path)[1]->type == EDGE_COPY_SRC_JOINER_BLOCK) | |
9943b198 | 1841 | { |
f1ce4e72 | 1842 | /* Attach the path to the starting edge if none is yet recorded. */ |
9943b198 | 1843 | if ((*path)[0]->e->aux == NULL) |
5d293e79 | 1844 | { |
9943b198 | 1845 | (*path)[0]->e->aux = path; |
51ea8bc6 | 1846 | i++; |
5d293e79 | 1847 | } |
1848 | else | |
1849 | { | |
1850 | paths.unordered_remove (i); | |
1851 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
9943b198 | 1852 | dump_jump_thread_path (dump_file, *path, false); |
5d293e79 | 1853 | delete_jump_thread_path (path); |
1854 | } | |
9943b198 | 1855 | } |
51ea8bc6 | 1856 | else |
1857 | { | |
1858 | i++; | |
1859 | } | |
f1ce4e72 | 1860 | } |
51ea8bc6 | 1861 | |
f1ce4e72 | 1862 | /* Second, look for paths that have any other jump thread attached to |
1863 | them, and either finish converting them or cancel them. */ | |
51ea8bc6 | 1864 | for (i = 0; i < paths.length ();) |
f1ce4e72 | 1865 | { |
1866 | vec<jump_thread_edge *> *path = paths[i]; | |
1867 | edge e = (*path)[0]->e; | |
1868 | ||
ff92f40d | 1869 | if (path->length () > 1 |
1870 | && (*path)[1]->type == EDGE_COPY_SRC_JOINER_BLOCK && e->aux == path) | |
b93ba654 | 1871 | { |
1872 | unsigned int j; | |
f1ce4e72 | 1873 | for (j = 1; j < path->length (); j++) |
b93ba654 | 1874 | if ((*path)[j]->e->aux != NULL) |
1875 | break; | |
1876 | ||
1877 | /* If we iterated through the entire path without exiting the loop, | |
f1ce4e72 | 1878 | then we are good to go, record it. */ |
b93ba654 | 1879 | if (j == path->length ()) |
51ea8bc6 | 1880 | { |
1881 | bitmap_set_bit (tmp, e->dest->index); | |
1882 | i++; | |
1883 | } | |
f1ce4e72 | 1884 | else |
b93ba654 | 1885 | { |
f1ce4e72 | 1886 | e->aux = NULL; |
5d293e79 | 1887 | paths.unordered_remove (i); |
f1ce4e72 | 1888 | if (dump_file && (dump_flags & TDF_DETAILS)) |
9943b198 | 1889 | dump_jump_thread_path (dump_file, *path, false); |
5d293e79 | 1890 | delete_jump_thread_path (path); |
b93ba654 | 1891 | } |
1892 | } | |
51ea8bc6 | 1893 | else |
1894 | { | |
1895 | i++; | |
1896 | } | |
b93ba654 | 1897 | } |
b99a7d6d | 1898 | |
5f885dd1 | 1899 | /* When optimizing for size, prune all thread paths where statement |
1900 | duplication is necessary. | |
1901 | ||
1902 | We walk the jump thread path looking for copied blocks. There's | |
1903 | two types of copied blocks. | |
1904 | ||
1905 | EDGE_COPY_SRC_JOINER_BLOCK is always copied and thus we will | |
1906 | cancel the jump threading request when optimizing for size. | |
1907 | ||
1908 | EDGE_COPY_SRC_BLOCK which is copied, but some of its statements | |
1909 | will be killed by threading. If threading does not kill all of | |
1910 | its statements, then we should cancel the jump threading request | |
1911 | when optimizing for size. */ | |
0bfd8d5c | 1912 | if (optimize_function_for_size_p (cfun)) |
7e0311ae | 1913 | { |
1914 | EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi) | |
1915 | { | |
5f885dd1 | 1916 | FOR_EACH_EDGE (e, ei, BASIC_BLOCK_FOR_FN (cfun, i)->preds) |
1917 | if (e->aux) | |
1918 | { | |
1919 | vec<jump_thread_edge *> *path = THREAD_PATH (e); | |
1920 | ||
1921 | unsigned int j; | |
1922 | for (j = 1; j < path->length (); j++) | |
1923 | { | |
1924 | bb = (*path)[j]->e->src; | |
1925 | if (redirection_block_p (bb)) | |
1926 | ; | |
1927 | else if ((*path)[j]->type == EDGE_COPY_SRC_JOINER_BLOCK | |
1928 | || ((*path)[j]->type == EDGE_COPY_SRC_BLOCK | |
1929 | && (count_stmts_and_phis_in_block (bb) | |
1930 | != estimate_threading_killed_stmts (bb)))) | |
1931 | break; | |
1932 | } | |
1933 | ||
1934 | if (j != path->length ()) | |
1935 | { | |
1936 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1937 | dump_jump_thread_path (dump_file, *path, 0); | |
1938 | delete_jump_thread_path (path); | |
1939 | e->aux = NULL; | |
1940 | } | |
1941 | else | |
1942 | bitmap_set_bit (threaded_blocks, i); | |
1943 | } | |
7e0311ae | 1944 | } |
3cebc9d2 | 1945 | } |
7e0311ae | 1946 | else |
1947 | bitmap_copy (threaded_blocks, tmp); | |
1948 | ||
af6b6631 | 1949 | /* If we have a joiner block (J) which has two successors S1 and S2 and |
1950 | we are threading though S1 and the final destination of the thread | |
1951 | is S2, then we must verify that any PHI nodes in S2 have the same | |
1952 | PHI arguments for the edge J->S2 and J->S1->...->S2. | |
1953 | ||
1954 | We used to detect this prior to registering the jump thread, but | |
1955 | that prohibits propagation of edge equivalences into non-dominated | |
1956 | PHI nodes as the equivalency test might occur before propagation. | |
1957 | ||
1958 | This must also occur after we truncate any jump threading paths | |
1959 | as this scenario may only show up after truncation. | |
1960 | ||
1961 | This works for now, but will need improvement as part of the FSA | |
1962 | optimization. | |
1963 | ||
1964 | Note since we've moved the thread request data to the edges, | |
1965 | we have to iterate on those rather than the threaded_edges vector. */ | |
1966 | EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi) | |
1967 | { | |
f5a6b05f | 1968 | bb = BASIC_BLOCK_FOR_FN (cfun, i); |
af6b6631 | 1969 | FOR_EACH_EDGE (e, ei, bb->preds) |
1970 | { | |
1971 | if (e->aux) | |
1972 | { | |
1973 | vec<jump_thread_edge *> *path = THREAD_PATH (e); | |
1974 | bool have_joiner = ((*path)[1]->type == EDGE_COPY_SRC_JOINER_BLOCK); | |
1975 | ||
1976 | if (have_joiner) | |
1977 | { | |
1978 | basic_block joiner = e->dest; | |
1979 | edge final_edge = path->last ()->e; | |
1980 | basic_block final_dest = final_edge->dest; | |
1981 | edge e2 = find_edge (joiner, final_dest); | |
1982 | ||
1983 | if (e2 && !phi_args_equal_on_edges (e2, final_edge)) | |
1984 | { | |
1985 | delete_jump_thread_path (path); | |
1986 | e->aux = NULL; | |
1987 | } | |
1988 | } | |
1989 | } | |
1990 | } | |
1991 | } | |
1992 | ||
0afc9b47 | 1993 | /* Look for jump threading paths which cross multiple loop headers. |
1994 | ||
1995 | The code to thread through loop headers will change the CFG in ways | |
1996 | that invalidate the cached loop iteration information. So we must | |
1997 | detect that case and wipe the cached information. */ | |
1998 | EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi) | |
1999 | { | |
2000 | basic_block bb = BASIC_BLOCK_FOR_FN (cfun, i); | |
2001 | FOR_EACH_EDGE (e, ei, bb->preds) | |
2002 | { | |
2003 | if (e->aux) | |
2004 | { | |
2005 | vec<jump_thread_edge *> *path = THREAD_PATH (e); | |
2006 | ||
2007 | for (unsigned int i = 0, crossed_headers = 0; | |
2008 | i < path->length (); | |
2009 | i++) | |
2010 | { | |
2011 | basic_block dest = (*path)[i]->e->dest; | |
2012 | basic_block src = (*path)[i]->e->src; | |
2013 | /* If we enter a loop. */ | |
2014 | if (flow_loop_nested_p (src->loop_father, dest->loop_father)) | |
2015 | ++crossed_headers; | |
2016 | /* If we step from a block outside an irreducible region | |
2017 | to a block inside an irreducible region, then we have | |
2018 | crossed into a loop. */ | |
2019 | else if (! (src->flags & BB_IRREDUCIBLE_LOOP) | |
2020 | && (dest->flags & BB_IRREDUCIBLE_LOOP)) | |
2021 | ++crossed_headers; | |
2022 | if (crossed_headers > 1) | |
2023 | { | |
2024 | vect_free_loop_info_assumptions | |
2025 | ((*path)[path->length () - 1]->e->dest->loop_father); | |
2026 | break; | |
2027 | } | |
2028 | } | |
2029 | } | |
2030 | } | |
2031 | } | |
3cebc9d2 | 2032 | } |
2033 | ||
2034 | ||
9e0d85a7 | 2035 | /* Verify that the REGION is a valid jump thread. A jump thread is a special |
2036 | case of SEME Single Entry Multiple Exits region in which all nodes in the | |
2037 | REGION have exactly one incoming edge. The only exception is the first block | |
2038 | that may not have been connected to the rest of the cfg yet. */ | |
ded1c768 | 2039 | |
2040 | DEBUG_FUNCTION void | |
9e0d85a7 | 2041 | verify_jump_thread (basic_block *region, unsigned n_region) |
ded1c768 | 2042 | { |
ded1c768 | 2043 | for (unsigned i = 0; i < n_region; i++) |
9e0d85a7 | 2044 | gcc_assert (EDGE_COUNT (region[i]->preds) <= 1); |
2045 | } | |
ded1c768 | 2046 | |
9e0d85a7 | 2047 | /* Return true when BB is one of the first N items in BBS. */ |
ded1c768 | 2048 | |
9e0d85a7 | 2049 | static inline bool |
2050 | bb_in_bbs (basic_block bb, basic_block *bbs, int n) | |
2051 | { | |
2052 | for (int i = 0; i < n; i++) | |
2053 | if (bb == bbs[i]) | |
2054 | return true; | |
ded1c768 | 2055 | |
9e0d85a7 | 2056 | return false; |
ded1c768 | 2057 | } |
2058 | ||
ff92f40d | 2059 | DEBUG_FUNCTION void |
2060 | debug_path (FILE *dump_file, int pathno) | |
2061 | { | |
2062 | vec<jump_thread_edge *> *p = paths[pathno]; | |
2063 | fprintf (dump_file, "path: "); | |
2064 | for (unsigned i = 0; i < p->length (); ++i) | |
2065 | fprintf (dump_file, "%d -> %d, ", | |
2066 | (*p)[i]->e->src->index, (*p)[i]->e->dest->index); | |
2067 | fprintf (dump_file, "\n"); | |
2068 | } | |
2069 | ||
2070 | DEBUG_FUNCTION void | |
2071 | debug_all_paths () | |
2072 | { | |
2073 | for (unsigned i = 0; i < paths.length (); ++i) | |
2074 | debug_path (stderr, i); | |
2075 | } | |
2076 | ||
2077 | /* Rewire a jump_thread_edge so that the source block is now a | |
2078 | threaded source block. | |
2079 | ||
2080 | PATH_NUM is an index into the global path table PATHS. | |
2081 | EDGE_NUM is the jump thread edge number into said path. | |
2082 | ||
2083 | Returns TRUE if we were able to successfully rewire the edge. */ | |
2084 | ||
2085 | static bool | |
2086 | rewire_first_differing_edge (unsigned path_num, unsigned edge_num) | |
2087 | { | |
2088 | vec<jump_thread_edge *> *path = paths[path_num]; | |
2089 | edge &e = (*path)[edge_num]->e; | |
2090 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2091 | fprintf (dump_file, "rewiring edge candidate: %d -> %d\n", | |
2092 | e->src->index, e->dest->index); | |
2093 | basic_block src_copy = get_bb_copy (e->src); | |
2094 | if (src_copy == NULL) | |
2095 | { | |
2096 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2097 | fprintf (dump_file, "ignoring candidate: there is no src COPY\n"); | |
2098 | return false; | |
2099 | } | |
2100 | edge new_edge = find_edge (src_copy, e->dest); | |
2101 | /* If the previously threaded paths created a flow graph where we | |
2102 | can no longer figure out where to go, give up. */ | |
2103 | if (new_edge == NULL) | |
2104 | { | |
2105 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2106 | fprintf (dump_file, "ignoring candidate: we lost our way\n"); | |
2107 | return false; | |
2108 | } | |
2109 | e = new_edge; | |
2110 | return true; | |
2111 | } | |
2112 | ||
2113 | /* After an FSM path has been jump threaded, adjust the remaining FSM | |
2114 | paths that are subsets of this path, so these paths can be safely | |
2115 | threaded within the context of the new threaded path. | |
2116 | ||
2117 | For example, suppose we have just threaded: | |
2118 | ||
2119 | 5 -> 6 -> 7 -> 8 -> 12 => 5 -> 6' -> 7' -> 8' -> 12' | |
2120 | ||
2121 | And we have an upcoming threading candidate: | |
2122 | 5 -> 6 -> 7 -> 8 -> 15 -> 20 | |
2123 | ||
2124 | This function adjusts the upcoming path into: | |
2125 | 8' -> 15 -> 20 | |
2126 | ||
2127 | CURR_PATH_NUM is an index into the global paths table. It | |
2128 | specifies the path that was just threaded. */ | |
2129 | ||
2130 | static void | |
2131 | adjust_paths_after_duplication (unsigned curr_path_num) | |
2132 | { | |
2133 | vec<jump_thread_edge *> *curr_path = paths[curr_path_num]; | |
2134 | gcc_assert ((*curr_path)[0]->type == EDGE_FSM_THREAD); | |
2135 | ||
2136 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2137 | { | |
2138 | fprintf (dump_file, "just threaded: "); | |
2139 | debug_path (dump_file, curr_path_num); | |
2140 | } | |
2141 | ||
2142 | /* Iterate through all the other paths and adjust them. */ | |
2143 | for (unsigned cand_path_num = 0; cand_path_num < paths.length (); ) | |
2144 | { | |
2145 | if (cand_path_num == curr_path_num) | |
2146 | { | |
2147 | ++cand_path_num; | |
2148 | continue; | |
2149 | } | |
2150 | /* Make sure the candidate to adjust starts with the same path | |
2151 | as the recently threaded path and is an FSM thread. */ | |
2152 | vec<jump_thread_edge *> *cand_path = paths[cand_path_num]; | |
2153 | if ((*cand_path)[0]->type != EDGE_FSM_THREAD | |
2154 | || (*cand_path)[0]->e != (*curr_path)[0]->e) | |
2155 | { | |
2156 | ++cand_path_num; | |
2157 | continue; | |
2158 | } | |
2159 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2160 | { | |
2161 | fprintf (dump_file, "adjusting candidate: "); | |
2162 | debug_path (dump_file, cand_path_num); | |
2163 | } | |
2164 | ||
2165 | /* Chop off from the candidate path any prefix it shares with | |
2166 | the recently threaded path. */ | |
2167 | unsigned minlength = MIN (curr_path->length (), cand_path->length ()); | |
2168 | unsigned j; | |
2169 | for (j = 0; j < minlength; ++j) | |
2170 | { | |
2171 | edge cand_edge = (*cand_path)[j]->e; | |
2172 | edge curr_edge = (*curr_path)[j]->e; | |
2173 | ||
2174 | /* Once the prefix no longer matches, adjust the first | |
2175 | non-matching edge to point from an adjusted edge to | |
2176 | wherever it was going. */ | |
2177 | if (cand_edge != curr_edge) | |
2178 | { | |
2179 | gcc_assert (cand_edge->src == curr_edge->src); | |
2180 | if (!rewire_first_differing_edge (cand_path_num, j)) | |
2181 | goto remove_candidate_from_list; | |
2182 | break; | |
2183 | } | |
2184 | } | |
2185 | if (j == minlength) | |
2186 | { | |
2187 | /* If we consumed the max subgraph we could look at, and | |
2188 | still didn't find any different edges, it's the | |
2189 | last edge after MINLENGTH. */ | |
2190 | if (cand_path->length () > minlength) | |
2191 | { | |
2192 | if (!rewire_first_differing_edge (cand_path_num, j)) | |
2193 | goto remove_candidate_from_list; | |
2194 | } | |
2195 | else if (dump_file && (dump_flags & TDF_DETAILS)) | |
2196 | fprintf (dump_file, "adjusting first edge after MINLENGTH.\n"); | |
2197 | } | |
2198 | if (j > 0) | |
2199 | { | |
2200 | /* If we are removing everything, delete the entire candidate. */ | |
2201 | if (j == cand_path->length ()) | |
2202 | { | |
2203 | remove_candidate_from_list: | |
2204 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2205 | fprintf (dump_file, "adjusted candidate: [EMPTY]\n"); | |
2206 | delete_jump_thread_path (cand_path); | |
2207 | paths.unordered_remove (cand_path_num); | |
2208 | continue; | |
2209 | } | |
2210 | /* Otherwise, just remove the redundant sub-path. */ | |
2211 | cand_path->block_remove (0, j); | |
2212 | } | |
2213 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2214 | { | |
2215 | fprintf (dump_file, "adjusted candidate: "); | |
2216 | debug_path (dump_file, cand_path_num); | |
2217 | } | |
2218 | ++cand_path_num; | |
2219 | } | |
2220 | } | |
2221 | ||
9e0d85a7 | 2222 | /* Duplicates a jump-thread path of N_REGION basic blocks. |
2223 | The ENTRY edge is redirected to the duplicate of the region. | |
ded1c768 | 2224 | |
2225 | Remove the last conditional statement in the last basic block in the REGION, | |
2226 | and create a single fallthru edge pointing to the same destination as the | |
2227 | EXIT edge. | |
2228 | ||
ff92f40d | 2229 | CURRENT_PATH_NO is an index into the global paths[] table |
2230 | specifying the jump-thread path. | |
2231 | ||
ded1c768 | 2232 | Returns false if it is unable to copy the region, true otherwise. */ |
2233 | ||
2234 | static bool | |
26300b20 | 2235 | duplicate_thread_path (edge entry, edge exit, basic_block *region, |
ff92f40d | 2236 | unsigned n_region, unsigned current_path_no) |
ded1c768 | 2237 | { |
2238 | unsigned i; | |
ded1c768 | 2239 | struct loop *loop = entry->dest->loop_father; |
2240 | edge exit_copy; | |
2241 | edge redirected; | |
db9cef39 | 2242 | profile_count curr_count; |
ded1c768 | 2243 | |
2244 | if (!can_copy_bbs_p (region, n_region)) | |
2245 | return false; | |
2246 | ||
ff92f40d | 2247 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2248 | { | |
2249 | fprintf (dump_file, "\nabout to thread: "); | |
2250 | debug_path (dump_file, current_path_no); | |
2251 | } | |
2252 | ||
ded1c768 | 2253 | /* Some sanity checking. Note that we do not check for all possible |
2254 | missuses of the functions. I.e. if you ask to copy something weird, | |
2255 | it will work, but the state of structures probably will not be | |
2256 | correct. */ | |
2257 | for (i = 0; i < n_region; i++) | |
2258 | { | |
2259 | /* We do not handle subloops, i.e. all the blocks must belong to the | |
2260 | same loop. */ | |
2261 | if (region[i]->loop_father != loop) | |
2262 | return false; | |
2263 | } | |
2264 | ||
2265 | initialize_original_copy_tables (); | |
2266 | ||
af5f6a93 | 2267 | set_loop_copy (loop, loop); |
ded1c768 | 2268 | |
26300b20 | 2269 | basic_block *region_copy = XNEWVEC (basic_block, n_region); |
ded1c768 | 2270 | copy_bbs (region, n_region, region_copy, &exit, 1, &exit_copy, loop, |
9e0d85a7 | 2271 | split_edge_bb_loc (entry), false); |
2272 | ||
2273 | /* Fix up: copy_bbs redirects all edges pointing to copied blocks. The | |
2274 | following code ensures that all the edges exiting the jump-thread path are | |
2275 | redirected back to the original code: these edges are exceptions | |
2276 | invalidating the property that is propagated by executing all the blocks of | |
2277 | the jump-thread path in order. */ | |
2278 | ||
ea5d3981 | 2279 | curr_count = entry->count (); |
acb9fddd | 2280 | |
9e0d85a7 | 2281 | for (i = 0; i < n_region; i++) |
2282 | { | |
2283 | edge e; | |
2284 | edge_iterator ei; | |
2285 | basic_block bb = region_copy[i]; | |
2286 | ||
acb9fddd | 2287 | /* Watch inconsistent profile. */ |
2288 | if (curr_count > region[i]->count) | |
2289 | curr_count = region[i]->count; | |
acb9fddd | 2290 | /* Scale current BB. */ |
205ce1aa | 2291 | if (region[i]->count.nonzero_p () && curr_count.initialized_p ()) |
acb9fddd | 2292 | { |
2293 | /* In the middle of the path we only scale the frequencies. | |
2294 | In last BB we need to update probabilities of outgoing edges | |
2295 | because we know which one is taken at the threaded path. */ | |
2296 | if (i + 1 != n_region) | |
db9cef39 | 2297 | scale_bbs_frequencies_profile_count (region + i, 1, |
2298 | region[i]->count - curr_count, | |
2299 | region[i]->count); | |
acb9fddd | 2300 | else |
2301 | update_bb_profile_for_threading (region[i], | |
205ce1aa | 2302 | curr_count, |
acb9fddd | 2303 | exit); |
db9cef39 | 2304 | scale_bbs_frequencies_profile_count (region_copy + i, 1, curr_count, |
2305 | region_copy[i]->count); | |
acb9fddd | 2306 | } |
acb9fddd | 2307 | |
9e0d85a7 | 2308 | if (single_succ_p (bb)) |
2309 | { | |
2310 | /* Make sure the successor is the next node in the path. */ | |
2311 | gcc_assert (i + 1 == n_region | |
2312 | || region_copy[i + 1] == single_succ_edge (bb)->dest); | |
acb9fddd | 2313 | if (i + 1 != n_region) |
2314 | { | |
ea5d3981 | 2315 | curr_count = single_succ_edge (bb)->count (); |
acb9fddd | 2316 | } |
9e0d85a7 | 2317 | continue; |
2318 | } | |
2319 | ||
2320 | /* Special case the last block on the path: make sure that it does not | |
5a653985 | 2321 | jump back on the copied path, including back to itself. */ |
9e0d85a7 | 2322 | if (i + 1 == n_region) |
2323 | { | |
2324 | FOR_EACH_EDGE (e, ei, bb->succs) | |
5a653985 | 2325 | if (bb_in_bbs (e->dest, region_copy, n_region)) |
9e0d85a7 | 2326 | { |
2327 | basic_block orig = get_bb_original (e->dest); | |
2328 | if (orig) | |
2329 | redirect_edge_and_branch_force (e, orig); | |
2330 | } | |
2331 | continue; | |
2332 | } | |
2333 | ||
2334 | /* Redirect all other edges jumping to non-adjacent blocks back to the | |
2335 | original code. */ | |
2336 | FOR_EACH_EDGE (e, ei, bb->succs) | |
2337 | if (region_copy[i + 1] != e->dest) | |
2338 | { | |
2339 | basic_block orig = get_bb_original (e->dest); | |
2340 | if (orig) | |
2341 | redirect_edge_and_branch_force (e, orig); | |
2342 | } | |
acb9fddd | 2343 | else |
2344 | { | |
ea5d3981 | 2345 | curr_count = e->count (); |
acb9fddd | 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 | } |
ded1c768 | 2368 | |
2369 | /* Redirect the entry and add the phi node arguments. */ | |
af5f6a93 | 2370 | if (entry->dest == loop->header) |
2371 | mark_loop_for_removal (loop); | |
ded1c768 | 2372 | redirected = redirect_edge_and_branch (entry, get_bb_copy (entry->dest)); |
2373 | gcc_assert (redirected != NULL); | |
2374 | flush_pending_stmts (entry); | |
2375 | ||
2376 | /* Add the other PHI node arguments. */ | |
2377 | add_phi_args_after_copy (region_copy, n_region, NULL); | |
2378 | ||
26300b20 | 2379 | free (region_copy); |
ded1c768 | 2380 | |
ff92f40d | 2381 | adjust_paths_after_duplication (current_path_no); |
2382 | ||
ded1c768 | 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; |
7e0311ae | 2439 | struct loop *loop; |
035def86 | 2440 | auto_bitmap threaded_blocks; |
8095249e | 2441 | hash_set<edge> visited_starting_edges; |
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 | ||
8095249e | 2487 | /* Do not jump-thread twice from the same starting edge. |
2488 | ||
2489 | Previously we only checked that we weren't threading twice | |
2490 | from the same BB, but that was too restrictive. Imagine a | |
2491 | path that starts from GIMPLE_COND(x_123 == 0,...), where both | |
2492 | edges out of this conditional yield paths that can be | |
2493 | threaded (for example, both lead to an x_123==0 or x_123!=0 | |
2494 | conditional further down the line. */ | |
2495 | if (visited_starting_edges.contains (entry) | |
2496 | /* We may not want to realize this jump thread path for | |
2497 | various reasons. So check it first. */ | |
b9903eb3 | 2498 | || !valid_jump_thread_path (path)) |
2499 | { | |
2500 | /* Remove invalid FSM jump-thread paths. */ | |
2501 | delete_jump_thread_path (path); | |
2502 | paths.unordered_remove (i); | |
2503 | continue; | |
2504 | } | |
ded1c768 | 2505 | |
2506 | unsigned len = path->length (); | |
2507 | edge exit = (*path)[len - 1]->e; | |
2508 | basic_block *region = XNEWVEC (basic_block, len - 1); | |
2509 | ||
2510 | for (unsigned int j = 0; j < len - 1; j++) | |
2511 | region[j] = (*path)[j]->e->dest; | |
2512 | ||
ff92f40d | 2513 | if (duplicate_thread_path (entry, exit, region, len - 1, i)) |
ded1c768 | 2514 | { |
2515 | /* We do not update dominance info. */ | |
2516 | free_dominance_info (CDI_DOMINATORS); | |
8095249e | 2517 | visited_starting_edges.add (entry); |
ded1c768 | 2518 | retval = true; |
7998c0b5 | 2519 | thread_stats.num_threaded_edges++; |
ded1c768 | 2520 | } |
2521 | ||
2522 | delete_jump_thread_path (path); | |
2523 | paths.unordered_remove (i); | |
b7cbf36d | 2524 | free (region); |
ded1c768 | 2525 | } |
2526 | ||
2527 | /* Remove from PATHS all the jump-threads starting with an edge already | |
2528 | jump-threaded. */ | |
2529 | for (i = 0; i < paths.length ();) | |
2530 | { | |
2531 | vec<jump_thread_edge *> *path = paths[i]; | |
2532 | edge entry = (*path)[0]->e; | |
2533 | ||
2534 | /* Do not jump-thread twice from the same block. */ | |
8095249e | 2535 | if (visited_starting_edges.contains (entry)) |
ded1c768 | 2536 | { |
2537 | delete_jump_thread_path (path); | |
2538 | paths.unordered_remove (i); | |
2539 | } | |
2540 | else | |
2541 | i++; | |
2542 | } | |
2543 | ||
3cebc9d2 | 2544 | mark_threaded_blocks (threaded_blocks); |
2545 | ||
96c90e5e | 2546 | initialize_original_copy_tables (); |
7e0311ae | 2547 | |
b0915eb6 | 2548 | /* The order in which we process jump threads can be important. |
2549 | ||
2550 | Consider if we have two jump threading paths A and B. If the | |
2551 | target edge of A is the starting edge of B and we thread path A | |
2552 | first, then we create an additional incoming edge into B->dest that | |
f4d3c071 | 2553 | we cannot discover as a jump threading path on this iteration. |
b0915eb6 | 2554 | |
2555 | If we instead thread B first, then the edge into B->dest will have | |
2556 | already been redirected before we process path A and path A will | |
2557 | natually, with no further work, target the redirected path for B. | |
7ea47fbd | 2558 | |
b0915eb6 | 2559 | An post-order is sufficient here. Compute the ordering first, then |
2560 | process the blocks. */ | |
2561 | if (!bitmap_empty_p (threaded_blocks)) | |
2562 | { | |
2563 | int *postorder = XNEWVEC (int, n_basic_blocks_for_fn (cfun)); | |
2564 | unsigned int postorder_num = post_order_compute (postorder, false, false); | |
2565 | for (unsigned int i = 0; i < postorder_num; i++) | |
2566 | { | |
2567 | unsigned int indx = postorder[i]; | |
2568 | if (bitmap_bit_p (threaded_blocks, indx)) | |
2569 | { | |
2570 | basic_block bb = BASIC_BLOCK_FOR_FN (cfun, indx); | |
2571 | retval |= thread_block (bb, true); | |
2572 | } | |
2573 | } | |
2574 | free (postorder); | |
7e0311ae | 2575 | } |
2576 | ||
2577 | /* Then perform the threading through loop headers. We start with the | |
2578 | innermost loop, so that the changes in cfg we perform won't affect | |
2579 | further threading. */ | |
f21d4d00 | 2580 | FOR_EACH_LOOP (loop, LI_FROM_INNERMOST) |
7e0311ae | 2581 | { |
7a3bf727 | 2582 | if (!loop->header |
2583 | || !bitmap_bit_p (threaded_blocks, loop->header->index)) | |
2584 | continue; | |
7e0311ae | 2585 | |
7a3bf727 | 2586 | retval |= thread_through_loop_header (loop, may_peel_loop_headers); |
a8046f60 | 2587 | } |
388d1fc1 | 2588 | |
c3dbcc09 | 2589 | /* All jump threading paths should have been resolved at this |
2590 | point. Verify that is the case. */ | |
ed4feca1 | 2591 | basic_block bb; |
fc00614f | 2592 | FOR_EACH_BB_FN (bb, cfun) |
ed4feca1 | 2593 | { |
c3dbcc09 | 2594 | edge_iterator ei; |
2595 | edge e; | |
2596 | FOR_EACH_EDGE (e, ei, bb->preds) | |
2597 | gcc_assert (e->aux == NULL); | |
ed4feca1 | 2598 | } |
2599 | ||
581f8050 | 2600 | statistics_counter_event (cfun, "Jumps threaded", |
2601 | thread_stats.num_threaded_edges); | |
5236b8bb | 2602 | |
96c90e5e | 2603 | free_original_copy_tables (); |
2604 | ||
f2981b08 | 2605 | paths.release (); |
7e0311ae | 2606 | |
396c773e | 2607 | if (retval) |
f24ec26f | 2608 | loops_state_set (LOOPS_NEED_FIXUP); |
eb2a640e | 2609 | |
a27d141e | 2610 | out: |
2611 | delete removed_edges; | |
2612 | removed_edges = NULL; | |
a8046f60 | 2613 | return retval; |
2614 | } | |
3cebc9d2 | 2615 | |
73b43bd5 | 2616 | /* Delete the jump threading path PATH. We have to explicitly delete |
6d1fdbf9 | 2617 | each entry in the vector, then the container. */ |
2618 | ||
2619 | void | |
2620 | delete_jump_thread_path (vec<jump_thread_edge *> *path) | |
2621 | { | |
2622 | for (unsigned int i = 0; i < path->length (); i++) | |
2623 | delete (*path)[i]; | |
2624 | path->release(); | |
9b5a88db | 2625 | delete path; |
6d1fdbf9 | 2626 | } |
2627 | ||
3cebc9d2 | 2628 | /* Register a jump threading opportunity. We queue up all the jump |
2629 | threading opportunities discovered by a pass and update the CFG | |
2630 | and SSA form all at once. | |
2631 | ||
f0b5f617 | 2632 | E is the edge we can thread, E2 is the new target edge, i.e., we |
3cebc9d2 | 2633 | are effectively recording that E->dest can be changed to E2->dest |
2634 | after fixing the SSA graph. */ | |
2635 | ||
2636 | void | |
f2981b08 | 2637 | register_jump_thread (vec<jump_thread_edge *> *path) |
3cebc9d2 | 2638 | { |
a3724f9d | 2639 | if (!dbg_cnt (registered_jump_thread)) |
2640 | { | |
6d1fdbf9 | 2641 | delete_jump_thread_path (path); |
a3724f9d | 2642 | return; |
2643 | } | |
2644 | ||
0c5b289a | 2645 | /* First make sure there are no NULL outgoing edges on the jump threading |
2646 | path. That can happen for jumping to a constant address. */ | |
f2981b08 | 2647 | for (unsigned int i = 0; i < path->length (); i++) |
d29932c9 | 2648 | { |
2649 | if ((*path)[i]->e == NULL) | |
31e5d72d | 2650 | { |
d29932c9 | 2651 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2652 | { | |
2653 | fprintf (dump_file, | |
2654 | "Found NULL edge in jump threading path. Cancelling jump thread:\n"); | |
2655 | dump_jump_thread_path (dump_file, *path, false); | |
2656 | } | |
f2981b08 | 2657 | |
d29932c9 | 2658 | delete_jump_thread_path (path); |
2659 | return; | |
31e5d72d | 2660 | } |
d29932c9 | 2661 | |
2662 | /* Only the FSM threader is allowed to thread across | |
2663 | backedges in the CFG. */ | |
2664 | if (flag_checking | |
2665 | && (*path)[0]->type != EDGE_FSM_THREAD) | |
2666 | gcc_assert (((*path)[i]->e->flags & EDGE_DFS_BACK) == 0); | |
2667 | } | |
5411af4e | 2668 | |
631d940c | 2669 | if (dump_file && (dump_flags & TDF_DETAILS)) |
b93ba654 | 2670 | dump_jump_thread_path (dump_file, *path, true); |
631d940c | 2671 | |
f2981b08 | 2672 | if (!paths.exists ()) |
2673 | paths.create (5); | |
631d940c | 2674 | |
f2981b08 | 2675 | paths.safe_push (path); |
3cebc9d2 | 2676 | } |
a8855004 | 2677 | |
2678 | /* Return how many uses of T there are within BB, as long as there | |
2679 | aren't any uses outside BB. If there are any uses outside BB, | |
2680 | return -1 if there's at most one use within BB, or -2 if there is | |
2681 | more than one use within BB. */ | |
2682 | ||
2683 | static int | |
2684 | uses_in_bb (tree t, basic_block bb) | |
2685 | { | |
2686 | int uses = 0; | |
2687 | bool outside_bb = false; | |
2688 | ||
2689 | imm_use_iterator iter; | |
2690 | use_operand_p use_p; | |
2691 | FOR_EACH_IMM_USE_FAST (use_p, iter, t) | |
2692 | { | |
2693 | if (is_gimple_debug (USE_STMT (use_p))) | |
2694 | continue; | |
2695 | ||
2696 | if (gimple_bb (USE_STMT (use_p)) != bb) | |
2697 | outside_bb = true; | |
2698 | else | |
2699 | uses++; | |
2700 | ||
2701 | if (outside_bb && uses > 1) | |
2702 | return -2; | |
2703 | } | |
2704 | ||
2705 | if (outside_bb) | |
2706 | return -1; | |
2707 | ||
2708 | return uses; | |
2709 | } | |
2710 | ||
2711 | /* Starting from the final control flow stmt in BB, assuming it will | |
2712 | be removed, follow uses in to-be-removed stmts back to their defs | |
2713 | and count how many defs are to become dead and be removed as | |
2714 | well. */ | |
2715 | ||
2716 | unsigned int | |
2717 | estimate_threading_killed_stmts (basic_block bb) | |
2718 | { | |
2719 | int killed_stmts = 0; | |
2720 | hash_map<tree, int> ssa_remaining_uses; | |
2721 | auto_vec<gimple *, 4> dead_worklist; | |
2722 | ||
2723 | /* If the block has only two predecessors, threading will turn phi | |
2724 | dsts into either src, so count them as dead stmts. */ | |
2725 | bool drop_all_phis = EDGE_COUNT (bb->preds) == 2; | |
2726 | ||
2727 | if (drop_all_phis) | |
2728 | for (gphi_iterator gsi = gsi_start_phis (bb); | |
2729 | !gsi_end_p (gsi); gsi_next (&gsi)) | |
2730 | { | |
2731 | gphi *phi = gsi.phi (); | |
2732 | tree dst = gimple_phi_result (phi); | |
2733 | ||
2734 | /* We don't count virtual PHIs as stmts in | |
2735 | record_temporary_equivalences_from_phis. */ | |
2736 | if (virtual_operand_p (dst)) | |
2737 | continue; | |
2738 | ||
2739 | killed_stmts++; | |
2740 | } | |
2741 | ||
2742 | if (gsi_end_p (gsi_last_bb (bb))) | |
2743 | return killed_stmts; | |
2744 | ||
2745 | gimple *stmt = gsi_stmt (gsi_last_bb (bb)); | |
2746 | if (gimple_code (stmt) != GIMPLE_COND | |
2747 | && gimple_code (stmt) != GIMPLE_GOTO | |
2748 | && gimple_code (stmt) != GIMPLE_SWITCH) | |
2749 | return killed_stmts; | |
2750 | ||
2751 | /* The control statement is always dead. */ | |
2752 | killed_stmts++; | |
2753 | dead_worklist.quick_push (stmt); | |
2754 | while (!dead_worklist.is_empty ()) | |
2755 | { | |
2756 | stmt = dead_worklist.pop (); | |
2757 | ||
2758 | ssa_op_iter iter; | |
2759 | use_operand_p use_p; | |
2760 | FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE) | |
2761 | { | |
2762 | tree t = USE_FROM_PTR (use_p); | |
2763 | gimple *def = SSA_NAME_DEF_STMT (t); | |
2764 | ||
2765 | if (gimple_bb (def) == bb | |
2766 | && (gimple_code (def) != GIMPLE_PHI | |
2767 | || !drop_all_phis) | |
2768 | && !gimple_has_side_effects (def)) | |
2769 | { | |
2770 | int *usesp = ssa_remaining_uses.get (t); | |
2771 | int uses; | |
2772 | ||
2773 | if (usesp) | |
2774 | uses = *usesp; | |
2775 | else | |
2776 | uses = uses_in_bb (t, bb); | |
2777 | ||
2778 | gcc_assert (uses); | |
2779 | ||
2780 | /* Don't bother recording the expected use count if we | |
2781 | won't find any further uses within BB. */ | |
2782 | if (!usesp && (uses < -1 || uses > 1)) | |
2783 | { | |
2784 | usesp = &ssa_remaining_uses.get_or_insert (t); | |
2785 | *usesp = uses; | |
2786 | } | |
2787 | ||
2788 | if (uses < 0) | |
2789 | continue; | |
2790 | ||
2791 | --uses; | |
2792 | if (usesp) | |
2793 | *usesp = uses; | |
2794 | ||
2795 | if (!uses) | |
2796 | { | |
2797 | killed_stmts++; | |
2798 | if (usesp) | |
2799 | ssa_remaining_uses.remove (t); | |
2800 | if (gimple_code (def) != GIMPLE_PHI) | |
2801 | dead_worklist.safe_push (def); | |
2802 | } | |
2803 | } | |
2804 | } | |
2805 | } | |
2806 | ||
2807 | if (dump_file) | |
2808 | fprintf (dump_file, "threading bb %i kills %i stmts\n", | |
2809 | bb->index, killed_stmts); | |
2810 | ||
2811 | return killed_stmts; | |
2812 | } |