]>
Commit | Line | Data |
---|---|---|
99c67f24 | 1 | /* Inlining decision heuristics. |
f1717362 | 2 | Copyright (C) 2003-2016 Free Software Foundation, Inc. |
99c67f24 | 3 | Contributed by Jan Hubicka |
4 | ||
5 | This file is part of GCC. | |
6 | ||
7 | GCC is free software; you can redistribute it and/or modify it under | |
8 | the terms of the GNU General Public License as published by the Free | |
9 | Software Foundation; either version 3, or (at your option) any later | |
10 | version. | |
11 | ||
12 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
13 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
14 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
15 | for more details. | |
16 | ||
17 | You should have received a copy of the GNU General Public License | |
18 | along with GCC; see the file COPYING3. If not see | |
19 | <http://www.gnu.org/licenses/>. */ | |
20 | ||
21 | /* Analysis used by the inliner and other passes limiting code size growth. | |
22 | ||
23 | We estimate for each function | |
24 | - function body size | |
c7b2cc59 | 25 | - average function execution time |
99c67f24 | 26 | - inlining size benefit (that is how much of function body size |
27 | and its call sequence is expected to disappear by inlining) | |
28 | - inlining time benefit | |
29 | - function frame size | |
30 | For each call | |
c7b2cc59 | 31 | - call statement size and time |
99c67f24 | 32 | |
33 | inlinie_summary datastructures store above information locally (i.e. | |
34 | parameters of the function itself) and globally (i.e. parameters of | |
35 | the function created by applying all the inline decisions already | |
36 | present in the callgraph). | |
37 | ||
a41f2a28 | 38 | We provide accestor to the inline_summary datastructure and |
99c67f24 | 39 | basic logic updating the parameters when inlining is performed. |
40 | ||
a41f2a28 | 41 | The summaries are context sensitive. Context means |
42 | 1) partial assignment of known constant values of operands | |
43 | 2) whether function is inlined into the call or not. | |
44 | It is easy to add more variants. To represent function size and time | |
45 | that depends on context (i.e. it is known to be optimized away when | |
46 | context is known either by inlining or from IP-CP and clonning), | |
47 | we use predicates. Predicates are logical formulas in | |
48 | conjunctive-disjunctive form consisting of clauses. Clauses are bitmaps | |
49 | specifying what conditions must be true. Conditions are simple test | |
50 | of the form described above. | |
51 | ||
52 | In order to make predicate (possibly) true, all of its clauses must | |
53 | be (possibly) true. To make clause (possibly) true, one of conditions | |
54 | it mentions must be (possibly) true. There are fixed bounds on | |
55 | number of clauses and conditions and all the manipulation functions | |
56 | are conservative in positive direction. I.e. we may lose precision | |
57 | by thinking that predicate may be true even when it is not. | |
58 | ||
59 | estimate_edge_size and estimate_edge_growth can be used to query | |
60 | function size/time in the given context. inline_merge_summary merges | |
61 | properties of caller and callee after inlining. | |
62 | ||
99c67f24 | 63 | Finally pass_inline_parameters is exported. This is used to drive |
64 | computation of function parameters used by the early inliner. IPA | |
65 | inlined performs analysis via its analyze_function method. */ | |
66 | ||
67 | #include "config.h" | |
68 | #include "system.h" | |
69 | #include "coretypes.h" | |
9ef16211 | 70 | #include "backend.h" |
99c67f24 | 71 | #include "tree.h" |
9ef16211 | 72 | #include "gimple.h" |
7c29e30e | 73 | #include "alloc-pool.h" |
74 | #include "tree-pass.h" | |
9ef16211 | 75 | #include "ssa.h" |
7c29e30e | 76 | #include "tree-streamer.h" |
77 | #include "cgraph.h" | |
7c29e30e | 78 | #include "diagnostic.h" |
b20a8bb4 | 79 | #include "fold-const.h" |
9ed99284 | 80 | #include "print-tree.h" |
99c67f24 | 81 | #include "tree-inline.h" |
99c67f24 | 82 | #include "gimple-pretty-print.h" |
99c67f24 | 83 | #include "params.h" |
94ea8568 | 84 | #include "cfganal.h" |
dcf1a1ec | 85 | #include "gimple-iterator.h" |
073c1fd5 | 86 | #include "tree-cfg.h" |
05d9c18a | 87 | #include "tree-ssa-loop-niter.h" |
073c1fd5 | 88 | #include "tree-ssa-loop.h" |
2cc80ac3 | 89 | #include "symbol-summary.h" |
99c67f24 | 90 | #include "ipa-prop.h" |
91 | #include "ipa-inline.h" | |
6b42039a | 92 | #include "cfgloop.h" |
7c07aa3d | 93 | #include "tree-scalar-evolution.h" |
6eaf903b | 94 | #include "ipa-utils.h" |
d037099f | 95 | #include "cilk.h" |
e797f49f | 96 | #include "cfgexpand.h" |
83caf9ff | 97 | #include "gimplify.h" |
99c67f24 | 98 | |
a41f2a28 | 99 | /* Estimate runtime of function can easilly run into huge numbers with many |
5aebd70d | 100 | nested loops. Be sure we can compute time * INLINE_SIZE_SCALE * 2 in an |
101 | integer. For anything larger we use gcov_type. */ | |
39273ba1 | 102 | #define MAX_TIME 500000 |
a41f2a28 | 103 | |
104 | /* Number of bits in integer, but we really want to be stable across different | |
105 | hosts. */ | |
106 | #define NUM_CONDITIONS 32 | |
107 | ||
108 | enum predicate_conditions | |
109 | { | |
110 | predicate_false_condition = 0, | |
111 | predicate_not_inlined_condition = 1, | |
112 | predicate_first_dynamic_condition = 2 | |
113 | }; | |
114 | ||
115 | /* Special condition code we use to represent test that operand is compile time | |
116 | constant. */ | |
117 | #define IS_NOT_CONSTANT ERROR_MARK | |
eb4ae064 | 118 | /* Special condition code we use to represent test that operand is not changed |
119 | across invocation of the function. When operand IS_NOT_CONSTANT it is always | |
120 | CHANGED, however i.e. loop invariants can be NOT_CHANGED given percentage | |
121 | of executions even when they are not compile time constants. */ | |
122 | #define CHANGED IDENTIFIER_NODE | |
99c67f24 | 123 | |
124 | /* Holders of ipa cgraph hooks: */ | |
0835ad03 | 125 | static struct cgraph_2edge_hook_list *edge_duplication_hook_holder; |
a41f2a28 | 126 | static struct cgraph_edge_hook_list *edge_removal_hook_holder; |
0835ad03 | 127 | static void inline_edge_removal_hook (struct cgraph_edge *, void *); |
128 | static void inline_edge_duplication_hook (struct cgraph_edge *, | |
e876d531 | 129 | struct cgraph_edge *, void *); |
c7b2cc59 | 130 | |
a41f2a28 | 131 | /* VECtor holding inline summaries. |
132 | In GGC memory because conditions might point to constant trees. */ | |
b4bae7a0 | 133 | function_summary <inline_summary *> *inline_summaries; |
f1f41a6c | 134 | vec<inline_edge_summary_t> inline_edge_summary_vec; |
a41f2a28 | 135 | |
136 | /* Cached node/edge growths. */ | |
f1f41a6c | 137 | vec<edge_growth_cache_entry> edge_growth_cache; |
a41f2a28 | 138 | |
6a18c0be | 139 | /* Edge predicates goes here. */ |
1dc6c44d | 140 | static object_allocator<predicate> edge_predicate_pool ("edge predicates"); |
a41f2a28 | 141 | |
142 | /* Return true predicate (tautology). | |
143 | We represent it by empty list of clauses. */ | |
144 | ||
145 | static inline struct predicate | |
146 | true_predicate (void) | |
147 | { | |
148 | struct predicate p; | |
563fae60 | 149 | p.clause[0] = 0; |
a41f2a28 | 150 | return p; |
151 | } | |
152 | ||
153 | ||
154 | /* Return predicate testing single condition number COND. */ | |
155 | ||
156 | static inline struct predicate | |
157 | single_cond_predicate (int cond) | |
158 | { | |
159 | struct predicate p; | |
563fae60 | 160 | p.clause[0] = 1 << cond; |
161 | p.clause[1] = 0; | |
a41f2a28 | 162 | return p; |
163 | } | |
164 | ||
165 | ||
166 | /* Return false predicate. First clause require false condition. */ | |
167 | ||
168 | static inline struct predicate | |
169 | false_predicate (void) | |
170 | { | |
171 | return single_cond_predicate (predicate_false_condition); | |
172 | } | |
173 | ||
174 | ||
007315c1 | 175 | /* Return true if P is (true). */ |
6a18c0be | 176 | |
177 | static inline bool | |
178 | true_predicate_p (struct predicate *p) | |
179 | { | |
180 | return !p->clause[0]; | |
181 | } | |
182 | ||
183 | ||
184 | /* Return true if P is (false). */ | |
185 | ||
186 | static inline bool | |
187 | false_predicate_p (struct predicate *p) | |
188 | { | |
189 | if (p->clause[0] == (1 << predicate_false_condition)) | |
190 | { | |
191 | gcc_checking_assert (!p->clause[1] | |
192 | && p->clause[0] == 1 << predicate_false_condition); | |
193 | return true; | |
194 | } | |
195 | return false; | |
196 | } | |
197 | ||
198 | ||
a41f2a28 | 199 | /* Return predicate that is set true when function is not inlined. */ |
e876d531 | 200 | |
a41f2a28 | 201 | static inline struct predicate |
202 | not_inlined_predicate (void) | |
203 | { | |
204 | return single_cond_predicate (predicate_not_inlined_condition); | |
205 | } | |
206 | ||
a4f60e55 | 207 | /* Simple description of whether a memory load or a condition refers to a load |
208 | from an aggregate and if so, how and where from in the aggregate. | |
209 | Individual fields have the same meaning like fields with the same name in | |
210 | struct condition. */ | |
a41f2a28 | 211 | |
a4f60e55 | 212 | struct agg_position_info |
213 | { | |
214 | HOST_WIDE_INT offset; | |
215 | bool agg_contents; | |
216 | bool by_ref; | |
217 | }; | |
218 | ||
219 | /* Add condition to condition list CONDS. AGGPOS describes whether the used | |
220 | oprand is loaded from an aggregate and where in the aggregate it is. It can | |
221 | be NULL, which means this not a load from an aggregate. */ | |
a41f2a28 | 222 | |
223 | static struct predicate | |
224 | add_condition (struct inline_summary *summary, int operand_num, | |
a4f60e55 | 225 | struct agg_position_info *aggpos, |
a41f2a28 | 226 | enum tree_code code, tree val) |
227 | { | |
228 | int i; | |
229 | struct condition *c; | |
230 | struct condition new_cond; | |
a4f60e55 | 231 | HOST_WIDE_INT offset; |
232 | bool agg_contents, by_ref; | |
a41f2a28 | 233 | |
a4f60e55 | 234 | if (aggpos) |
235 | { | |
236 | offset = aggpos->offset; | |
237 | agg_contents = aggpos->agg_contents; | |
238 | by_ref = aggpos->by_ref; | |
239 | } | |
240 | else | |
241 | { | |
242 | offset = 0; | |
243 | agg_contents = false; | |
244 | by_ref = false; | |
245 | } | |
246 | ||
247 | gcc_checking_assert (operand_num >= 0); | |
f1f41a6c | 248 | for (i = 0; vec_safe_iterate (summary->conds, i, &c); i++) |
a41f2a28 | 249 | { |
250 | if (c->operand_num == operand_num | |
251 | && c->code == code | |
a4f60e55 | 252 | && c->val == val |
253 | && c->agg_contents == agg_contents | |
254 | && (!agg_contents || (c->offset == offset && c->by_ref == by_ref))) | |
e876d531 | 255 | return single_cond_predicate (i + predicate_first_dynamic_condition); |
a41f2a28 | 256 | } |
257 | /* Too many conditions. Give up and return constant true. */ | |
258 | if (i == NUM_CONDITIONS - predicate_first_dynamic_condition) | |
259 | return true_predicate (); | |
260 | ||
261 | new_cond.operand_num = operand_num; | |
262 | new_cond.code = code; | |
263 | new_cond.val = val; | |
a4f60e55 | 264 | new_cond.agg_contents = agg_contents; |
265 | new_cond.by_ref = by_ref; | |
266 | new_cond.offset = offset; | |
f1f41a6c | 267 | vec_safe_push (summary->conds, new_cond); |
a41f2a28 | 268 | return single_cond_predicate (i + predicate_first_dynamic_condition); |
269 | } | |
270 | ||
271 | ||
905aa3bd | 272 | /* Add clause CLAUSE into the predicate P. */ |
a41f2a28 | 273 | |
274 | static inline void | |
94646c9c | 275 | add_clause (conditions conditions, struct predicate *p, clause_t clause) |
a41f2a28 | 276 | { |
277 | int i; | |
905aa3bd | 278 | int i2; |
5cb1b112 | 279 | int insert_here = -1; |
94646c9c | 280 | int c1, c2; |
6a18c0be | 281 | |
a41f2a28 | 282 | /* True clause. */ |
283 | if (!clause) | |
284 | return; | |
285 | ||
905aa3bd | 286 | /* False clause makes the whole predicate false. Kill the other variants. */ |
6a18c0be | 287 | if (clause == (1 << predicate_false_condition)) |
a41f2a28 | 288 | { |
289 | p->clause[0] = (1 << predicate_false_condition); | |
290 | p->clause[1] = 0; | |
6a18c0be | 291 | return; |
a41f2a28 | 292 | } |
6a18c0be | 293 | if (false_predicate_p (p)) |
294 | return; | |
905aa3bd | 295 | |
74d6d072 | 296 | /* No one should be silly enough to add false into nontrivial clauses. */ |
905aa3bd | 297 | gcc_checking_assert (!(clause & (1 << predicate_false_condition))); |
298 | ||
299 | /* Look where to insert the clause. At the same time prune out | |
300 | clauses of P that are implied by the new clause and thus | |
301 | redundant. */ | |
302 | for (i = 0, i2 = 0; i <= MAX_CLAUSES; i++) | |
a41f2a28 | 303 | { |
905aa3bd | 304 | p->clause[i2] = p->clause[i]; |
305 | ||
a41f2a28 | 306 | if (!p->clause[i]) |
307 | break; | |
905aa3bd | 308 | |
309 | /* If p->clause[i] implies clause, there is nothing to add. */ | |
310 | if ((p->clause[i] & clause) == p->clause[i]) | |
311 | { | |
fb3c587e | 312 | /* We had nothing to add, none of clauses should've become |
313 | redundant. */ | |
905aa3bd | 314 | gcc_checking_assert (i == i2); |
315 | return; | |
316 | } | |
317 | ||
318 | if (p->clause[i] < clause && insert_here < 0) | |
319 | insert_here = i2; | |
320 | ||
321 | /* If clause implies p->clause[i], then p->clause[i] becomes redundant. | |
e876d531 | 322 | Otherwise the p->clause[i] has to stay. */ |
905aa3bd | 323 | if ((p->clause[i] & clause) != clause) |
324 | i2++; | |
a41f2a28 | 325 | } |
94646c9c | 326 | |
327 | /* Look for clauses that are obviously true. I.e. | |
328 | op0 == 5 || op0 != 5. */ | |
329 | for (c1 = predicate_first_dynamic_condition; c1 < NUM_CONDITIONS; c1++) | |
eb4ae064 | 330 | { |
331 | condition *cc1; | |
332 | if (!(clause & (1 << c1))) | |
333 | continue; | |
f1f41a6c | 334 | cc1 = &(*conditions)[c1 - predicate_first_dynamic_condition]; |
eb4ae064 | 335 | /* We have no way to represent !CHANGED and !IS_NOT_CONSTANT |
e876d531 | 336 | and thus there is no point for looking for them. */ |
337 | if (cc1->code == CHANGED || cc1->code == IS_NOT_CONSTANT) | |
eb4ae064 | 338 | continue; |
8249d427 | 339 | for (c2 = c1 + 1; c2 < NUM_CONDITIONS; c2++) |
eb4ae064 | 340 | if (clause & (1 << c2)) |
341 | { | |
e876d531 | 342 | condition *cc1 = |
343 | &(*conditions)[c1 - predicate_first_dynamic_condition]; | |
344 | condition *cc2 = | |
345 | &(*conditions)[c2 - predicate_first_dynamic_condition]; | |
eb4ae064 | 346 | if (cc1->operand_num == cc2->operand_num |
347 | && cc1->val == cc2->val | |
348 | && cc2->code != IS_NOT_CONSTANT | |
349 | && cc2->code != CHANGED | |
93633022 | 350 | && cc1->code == invert_tree_comparison (cc2->code, |
351 | HONOR_NANS (cc1->val))) | |
eb4ae064 | 352 | return; |
353 | } | |
354 | } | |
e876d531 | 355 | |
94646c9c | 356 | |
905aa3bd | 357 | /* We run out of variants. Be conservative in positive direction. */ |
358 | if (i2 == MAX_CLAUSES) | |
a41f2a28 | 359 | return; |
905aa3bd | 360 | /* Keep clauses in decreasing order. This makes equivalence testing easy. */ |
361 | p->clause[i2 + 1] = 0; | |
5cb1b112 | 362 | if (insert_here >= 0) |
e876d531 | 363 | for (; i2 > insert_here; i2--) |
905aa3bd | 364 | p->clause[i2] = p->clause[i2 - 1]; |
5cb1b112 | 365 | else |
905aa3bd | 366 | insert_here = i2; |
a41f2a28 | 367 | p->clause[insert_here] = clause; |
368 | } | |
369 | ||
370 | ||
371 | /* Return P & P2. */ | |
372 | ||
373 | static struct predicate | |
94646c9c | 374 | and_predicates (conditions conditions, |
375 | struct predicate *p, struct predicate *p2) | |
a41f2a28 | 376 | { |
377 | struct predicate out = *p; | |
378 | int i; | |
6a18c0be | 379 | |
905aa3bd | 380 | /* Avoid busy work. */ |
381 | if (false_predicate_p (p2) || true_predicate_p (p)) | |
382 | return *p2; | |
383 | if (false_predicate_p (p) || true_predicate_p (p2)) | |
384 | return *p; | |
385 | ||
386 | /* See how far predicates match. */ | |
387 | for (i = 0; p->clause[i] && p->clause[i] == p2->clause[i]; i++) | |
388 | { | |
389 | gcc_checking_assert (i < MAX_CLAUSES); | |
390 | } | |
e876d531 | 391 | |
905aa3bd | 392 | /* Combine the predicates rest. */ |
393 | for (; p2->clause[i]; i++) | |
5cb1b112 | 394 | { |
395 | gcc_checking_assert (i < MAX_CLAUSES); | |
94646c9c | 396 | add_clause (conditions, &out, p2->clause[i]); |
5cb1b112 | 397 | } |
a41f2a28 | 398 | return out; |
399 | } | |
400 | ||
401 | ||
905aa3bd | 402 | /* Return true if predicates are obviously equal. */ |
403 | ||
404 | static inline bool | |
405 | predicates_equal_p (struct predicate *p, struct predicate *p2) | |
406 | { | |
407 | int i; | |
408 | for (i = 0; p->clause[i]; i++) | |
409 | { | |
410 | gcc_checking_assert (i < MAX_CLAUSES); | |
e876d531 | 411 | gcc_checking_assert (p->clause[i] > p->clause[i + 1]); |
fb3c587e | 412 | gcc_checking_assert (!p2->clause[i] |
e876d531 | 413 | || p2->clause[i] > p2->clause[i + 1]); |
905aa3bd | 414 | if (p->clause[i] != p2->clause[i]) |
e876d531 | 415 | return false; |
905aa3bd | 416 | } |
417 | return !p2->clause[i]; | |
418 | } | |
419 | ||
420 | ||
a41f2a28 | 421 | /* Return P | P2. */ |
422 | ||
423 | static struct predicate | |
e876d531 | 424 | or_predicates (conditions conditions, |
425 | struct predicate *p, struct predicate *p2) | |
a41f2a28 | 426 | { |
427 | struct predicate out = true_predicate (); | |
e876d531 | 428 | int i, j; |
6a18c0be | 429 | |
905aa3bd | 430 | /* Avoid busy work. */ |
431 | if (false_predicate_p (p2) || true_predicate_p (p)) | |
6a18c0be | 432 | return *p; |
905aa3bd | 433 | if (false_predicate_p (p) || true_predicate_p (p2)) |
6a18c0be | 434 | return *p2; |
905aa3bd | 435 | if (predicates_equal_p (p, p2)) |
436 | return *p; | |
437 | ||
438 | /* OK, combine the predicates. */ | |
a41f2a28 | 439 | for (i = 0; p->clause[i]; i++) |
440 | for (j = 0; p2->clause[j]; j++) | |
5cb1b112 | 441 | { |
e876d531 | 442 | gcc_checking_assert (i < MAX_CLAUSES && j < MAX_CLAUSES); |
443 | add_clause (conditions, &out, p->clause[i] | p2->clause[j]); | |
5cb1b112 | 444 | } |
a41f2a28 | 445 | return out; |
446 | } | |
447 | ||
448 | ||
fb3c587e | 449 | /* Having partial truth assignment in POSSIBLE_TRUTHS, return false |
450 | if predicate P is known to be false. */ | |
a41f2a28 | 451 | |
452 | static bool | |
6a18c0be | 453 | evaluate_predicate (struct predicate *p, clause_t possible_truths) |
a41f2a28 | 454 | { |
455 | int i; | |
456 | ||
457 | /* True remains true. */ | |
6a18c0be | 458 | if (true_predicate_p (p)) |
a41f2a28 | 459 | return true; |
460 | ||
6a18c0be | 461 | gcc_assert (!(possible_truths & (1 << predicate_false_condition))); |
462 | ||
a41f2a28 | 463 | /* See if we can find clause we can disprove. */ |
464 | for (i = 0; p->clause[i]; i++) | |
5cb1b112 | 465 | { |
466 | gcc_checking_assert (i < MAX_CLAUSES); | |
467 | if (!(p->clause[i] & possible_truths)) | |
e876d531 | 468 | return false; |
5cb1b112 | 469 | } |
a41f2a28 | 470 | return true; |
471 | } | |
472 | ||
eb4ae064 | 473 | /* Return the probability in range 0...REG_BR_PROB_BASE that the predicated |
474 | instruction will be recomputed per invocation of the inlined call. */ | |
475 | ||
476 | static int | |
477 | predicate_probability (conditions conds, | |
478 | struct predicate *p, clause_t possible_truths, | |
b3e7c666 | 479 | vec<inline_param_summary> inline_param_summary) |
eb4ae064 | 480 | { |
481 | int i; | |
482 | int combined_prob = REG_BR_PROB_BASE; | |
483 | ||
484 | /* True remains true. */ | |
485 | if (true_predicate_p (p)) | |
486 | return REG_BR_PROB_BASE; | |
487 | ||
488 | if (false_predicate_p (p)) | |
489 | return 0; | |
490 | ||
491 | gcc_assert (!(possible_truths & (1 << predicate_false_condition))); | |
492 | ||
493 | /* See if we can find clause we can disprove. */ | |
494 | for (i = 0; p->clause[i]; i++) | |
495 | { | |
496 | gcc_checking_assert (i < MAX_CLAUSES); | |
497 | if (!(p->clause[i] & possible_truths)) | |
498 | return 0; | |
499 | else | |
500 | { | |
501 | int this_prob = 0; | |
502 | int i2; | |
f1f41a6c | 503 | if (!inline_param_summary.exists ()) |
eb4ae064 | 504 | return REG_BR_PROB_BASE; |
505 | for (i2 = 0; i2 < NUM_CONDITIONS; i2++) | |
506 | if ((p->clause[i] & possible_truths) & (1 << i2)) | |
507 | { | |
508 | if (i2 >= predicate_first_dynamic_condition) | |
509 | { | |
e876d531 | 510 | condition *c = |
511 | &(*conds)[i2 - predicate_first_dynamic_condition]; | |
eb4ae064 | 512 | if (c->code == CHANGED |
e876d531 | 513 | && (c->operand_num < |
514 | (int) inline_param_summary.length ())) | |
eb4ae064 | 515 | { |
e876d531 | 516 | int iprob = |
517 | inline_param_summary[c->operand_num].change_prob; | |
eb4ae064 | 518 | this_prob = MAX (this_prob, iprob); |
519 | } | |
520 | else | |
521 | this_prob = REG_BR_PROB_BASE; | |
e876d531 | 522 | } |
523 | else | |
524 | this_prob = REG_BR_PROB_BASE; | |
eb4ae064 | 525 | } |
526 | combined_prob = MIN (this_prob, combined_prob); | |
527 | if (!combined_prob) | |
e876d531 | 528 | return 0; |
eb4ae064 | 529 | } |
530 | } | |
531 | return combined_prob; | |
532 | } | |
533 | ||
a41f2a28 | 534 | |
535 | /* Dump conditional COND. */ | |
536 | ||
537 | static void | |
538 | dump_condition (FILE *f, conditions conditions, int cond) | |
539 | { | |
540 | condition *c; | |
541 | if (cond == predicate_false_condition) | |
542 | fprintf (f, "false"); | |
543 | else if (cond == predicate_not_inlined_condition) | |
544 | fprintf (f, "not inlined"); | |
545 | else | |
546 | { | |
f1f41a6c | 547 | c = &(*conditions)[cond - predicate_first_dynamic_condition]; |
a41f2a28 | 548 | fprintf (f, "op%i", c->operand_num); |
a4f60e55 | 549 | if (c->agg_contents) |
550 | fprintf (f, "[%soffset: " HOST_WIDE_INT_PRINT_DEC "]", | |
551 | c->by_ref ? "ref " : "", c->offset); | |
a41f2a28 | 552 | if (c->code == IS_NOT_CONSTANT) |
553 | { | |
554 | fprintf (f, " not constant"); | |
555 | return; | |
556 | } | |
eb4ae064 | 557 | if (c->code == CHANGED) |
558 | { | |
559 | fprintf (f, " changed"); | |
560 | return; | |
561 | } | |
a41f2a28 | 562 | fprintf (f, " %s ", op_symbol_code (c->code)); |
563 | print_generic_expr (f, c->val, 1); | |
564 | } | |
565 | } | |
566 | ||
567 | ||
568 | /* Dump clause CLAUSE. */ | |
569 | ||
570 | static void | |
571 | dump_clause (FILE *f, conditions conds, clause_t clause) | |
572 | { | |
573 | int i; | |
574 | bool found = false; | |
575 | fprintf (f, "("); | |
576 | if (!clause) | |
577 | fprintf (f, "true"); | |
578 | for (i = 0; i < NUM_CONDITIONS; i++) | |
579 | if (clause & (1 << i)) | |
580 | { | |
581 | if (found) | |
582 | fprintf (f, " || "); | |
583 | found = true; | |
e876d531 | 584 | dump_condition (f, conds, i); |
a41f2a28 | 585 | } |
586 | fprintf (f, ")"); | |
587 | } | |
588 | ||
589 | ||
590 | /* Dump predicate PREDICATE. */ | |
591 | ||
592 | static void | |
593 | dump_predicate (FILE *f, conditions conds, struct predicate *pred) | |
594 | { | |
595 | int i; | |
6a18c0be | 596 | if (true_predicate_p (pred)) |
a41f2a28 | 597 | dump_clause (f, conds, 0); |
598 | else | |
599 | for (i = 0; pred->clause[i]; i++) | |
600 | { | |
601 | if (i) | |
602 | fprintf (f, " && "); | |
e876d531 | 603 | dump_clause (f, conds, pred->clause[i]); |
a41f2a28 | 604 | } |
605 | fprintf (f, "\n"); | |
606 | } | |
607 | ||
608 | ||
eb7c606e | 609 | /* Dump inline hints. */ |
610 | void | |
611 | dump_inline_hints (FILE *f, inline_hints hints) | |
612 | { | |
613 | if (!hints) | |
614 | return; | |
615 | fprintf (f, "inline hints:"); | |
616 | if (hints & INLINE_HINT_indirect_call) | |
617 | { | |
618 | hints &= ~INLINE_HINT_indirect_call; | |
619 | fprintf (f, " indirect_call"); | |
620 | } | |
7c07aa3d | 621 | if (hints & INLINE_HINT_loop_iterations) |
622 | { | |
623 | hints &= ~INLINE_HINT_loop_iterations; | |
624 | fprintf (f, " loop_iterations"); | |
625 | } | |
3716ee8f | 626 | if (hints & INLINE_HINT_loop_stride) |
627 | { | |
628 | hints &= ~INLINE_HINT_loop_stride; | |
629 | fprintf (f, " loop_stride"); | |
630 | } | |
41d39f38 | 631 | if (hints & INLINE_HINT_same_scc) |
632 | { | |
633 | hints &= ~INLINE_HINT_same_scc; | |
634 | fprintf (f, " same_scc"); | |
635 | } | |
636 | if (hints & INLINE_HINT_in_scc) | |
637 | { | |
638 | hints &= ~INLINE_HINT_in_scc; | |
639 | fprintf (f, " in_scc"); | |
640 | } | |
3172b7bf | 641 | if (hints & INLINE_HINT_cross_module) |
642 | { | |
643 | hints &= ~INLINE_HINT_cross_module; | |
644 | fprintf (f, " cross_module"); | |
645 | } | |
646 | if (hints & INLINE_HINT_declared_inline) | |
647 | { | |
648 | hints &= ~INLINE_HINT_declared_inline; | |
649 | fprintf (f, " declared_inline"); | |
650 | } | |
be343a9c | 651 | if (hints & INLINE_HINT_array_index) |
652 | { | |
653 | hints &= ~INLINE_HINT_array_index; | |
654 | fprintf (f, " array_index"); | |
655 | } | |
3072aa32 | 656 | if (hints & INLINE_HINT_known_hot) |
657 | { | |
658 | hints &= ~INLINE_HINT_known_hot; | |
659 | fprintf (f, " known_hot"); | |
660 | } | |
eb7c606e | 661 | gcc_assert (!hints); |
662 | } | |
663 | ||
664 | ||
a41f2a28 | 665 | /* Record SIZE and TIME under condition PRED into the inline summary. */ |
666 | ||
667 | static void | |
fb3c587e | 668 | account_size_time (struct inline_summary *summary, int size, int time, |
669 | struct predicate *pred) | |
a41f2a28 | 670 | { |
671 | size_time_entry *e; | |
672 | bool found = false; | |
673 | int i; | |
674 | ||
6a18c0be | 675 | if (false_predicate_p (pred)) |
a41f2a28 | 676 | return; |
677 | ||
678 | /* We need to create initial empty unconitional clause, but otherwie | |
679 | we don't need to account empty times and sizes. */ | |
8bae3ea4 | 680 | if (!size && !time && summary->entry) |
a41f2a28 | 681 | return; |
682 | ||
683 | /* Watch overflow that might result from insane profiles. */ | |
684 | if (time > MAX_TIME * INLINE_TIME_SCALE) | |
685 | time = MAX_TIME * INLINE_TIME_SCALE; | |
686 | gcc_assert (time >= 0); | |
687 | ||
f1f41a6c | 688 | for (i = 0; vec_safe_iterate (summary->entry, i, &e); i++) |
a41f2a28 | 689 | if (predicates_equal_p (&e->predicate, pred)) |
690 | { | |
691 | found = true; | |
e876d531 | 692 | break; |
a41f2a28 | 693 | } |
563fae60 | 694 | if (i == 256) |
a41f2a28 | 695 | { |
696 | i = 0; | |
697 | found = true; | |
f1f41a6c | 698 | e = &(*summary->entry)[0]; |
a41f2a28 | 699 | gcc_assert (!e->predicate.clause[0]); |
563fae60 | 700 | if (dump_file && (dump_flags & TDF_DETAILS)) |
e876d531 | 701 | fprintf (dump_file, |
702 | "\t\tReached limit on number of entries, " | |
703 | "ignoring the predicate."); | |
a41f2a28 | 704 | } |
705 | if (dump_file && (dump_flags & TDF_DETAILS) && (time || size)) | |
706 | { | |
e876d531 | 707 | fprintf (dump_file, |
708 | "\t\tAccounting size:%3.2f, time:%3.2f on %spredicate:", | |
709 | ((double) size) / INLINE_SIZE_SCALE, | |
710 | ((double) time) / INLINE_TIME_SCALE, found ? "" : "new "); | |
a41f2a28 | 711 | dump_predicate (dump_file, summary->conds, pred); |
712 | } | |
713 | if (!found) | |
714 | { | |
715 | struct size_time_entry new_entry; | |
716 | new_entry.size = size; | |
717 | new_entry.time = time; | |
718 | new_entry.predicate = *pred; | |
f1f41a6c | 719 | vec_safe_push (summary->entry, new_entry); |
a41f2a28 | 720 | } |
721 | else | |
722 | { | |
723 | e->size += size; | |
724 | e->time += time; | |
725 | if (e->time > MAX_TIME * INLINE_TIME_SCALE) | |
726 | e->time = MAX_TIME * INLINE_TIME_SCALE; | |
727 | } | |
728 | } | |
729 | ||
e806c56f | 730 | /* We proved E to be unreachable, redirect it to __bultin_unreachable. */ |
731 | ||
99fe8cbd | 732 | static struct cgraph_edge * |
e806c56f | 733 | redirect_to_unreachable (struct cgraph_edge *e) |
734 | { | |
735 | struct cgraph_node *callee = !e->inline_failed ? e->callee : NULL; | |
bdbe7747 | 736 | struct cgraph_node *target = cgraph_node::get_create |
737 | (builtin_decl_implicit (BUILT_IN_UNREACHABLE)); | |
e806c56f | 738 | |
739 | if (e->speculative) | |
bdbe7747 | 740 | e = e->resolve_speculation (target->decl); |
99fe8cbd | 741 | else if (!e->callee) |
bdbe7747 | 742 | e->make_direct (target); |
e806c56f | 743 | else |
bdbe7747 | 744 | e->redirect_callee (target); |
99fe8cbd | 745 | struct inline_edge_summary *es = inline_edge_summary (e); |
e806c56f | 746 | e->inline_failed = CIF_UNREACHABLE; |
747 | e->frequency = 0; | |
748 | e->count = 0; | |
749 | es->call_stmt_size = 0; | |
750 | es->call_stmt_time = 0; | |
751 | if (callee) | |
752 | callee->remove_symbol_and_inline_clones (); | |
99fe8cbd | 753 | return e; |
e806c56f | 754 | } |
755 | ||
6a18c0be | 756 | /* Set predicate for edge E. */ |
757 | ||
758 | static void | |
759 | edge_set_predicate (struct cgraph_edge *e, struct predicate *predicate) | |
760 | { | |
5a7ad253 | 761 | /* If the edge is determined to be never executed, redirect it |
762 | to BUILTIN_UNREACHABLE to save inliner from inlining into it. */ | |
93b6a5b4 | 763 | if (predicate && false_predicate_p (predicate) |
764 | /* When handling speculative edges, we need to do the redirection | |
765 | just once. Do it always on the direct edge, so we do not | |
766 | attempt to resolve speculation while duplicating the edge. */ | |
767 | && (!e->speculative || e->callee)) | |
99fe8cbd | 768 | e = redirect_to_unreachable (e); |
769 | ||
770 | struct inline_edge_summary *es = inline_edge_summary (e); | |
6a18c0be | 771 | if (predicate && !true_predicate_p (predicate)) |
772 | { | |
773 | if (!es->predicate) | |
2a5261f7 | 774 | es->predicate = edge_predicate_pool.allocate (); |
6a18c0be | 775 | *es->predicate = *predicate; |
776 | } | |
777 | else | |
778 | { | |
779 | if (es->predicate) | |
2a5261f7 | 780 | edge_predicate_pool.remove (es->predicate); |
6a18c0be | 781 | es->predicate = NULL; |
782 | } | |
783 | } | |
784 | ||
3716ee8f | 785 | /* Set predicate for hint *P. */ |
786 | ||
787 | static void | |
788 | set_hint_predicate (struct predicate **p, struct predicate new_predicate) | |
789 | { | |
e876d531 | 790 | if (false_predicate_p (&new_predicate) || true_predicate_p (&new_predicate)) |
3716ee8f | 791 | { |
792 | if (*p) | |
2a5261f7 | 793 | edge_predicate_pool.remove (*p); |
3716ee8f | 794 | *p = NULL; |
795 | } | |
796 | else | |
797 | { | |
798 | if (!*p) | |
2a5261f7 | 799 | *p = edge_predicate_pool.allocate (); |
3716ee8f | 800 | **p = new_predicate; |
801 | } | |
802 | } | |
803 | ||
a41f2a28 | 804 | |
8bae3ea4 | 805 | /* KNOWN_VALS is partial mapping of parameters of NODE to constant values. |
a4f60e55 | 806 | KNOWN_AGGS is a vector of aggreggate jump functions for each parameter. |
807 | Return clause of possible truths. When INLINE_P is true, assume that we are | |
808 | inlining. | |
eb4ae064 | 809 | |
810 | ERROR_MARK means compile time invariant. */ | |
8bae3ea4 | 811 | |
812 | static clause_t | |
813 | evaluate_conditions_for_known_args (struct cgraph_node *node, | |
e876d531 | 814 | bool inline_p, |
815 | vec<tree> known_vals, | |
816 | vec<ipa_agg_jump_function_p> | |
817 | known_aggs) | |
8bae3ea4 | 818 | { |
819 | clause_t clause = inline_p ? 0 : 1 << predicate_not_inlined_condition; | |
b4bae7a0 | 820 | struct inline_summary *info = inline_summaries->get (node); |
8bae3ea4 | 821 | int i; |
822 | struct condition *c; | |
823 | ||
f1f41a6c | 824 | for (i = 0; vec_safe_iterate (info->conds, i, &c); i++) |
8bae3ea4 | 825 | { |
92ceb887 | 826 | tree val; |
8bae3ea4 | 827 | tree res; |
828 | ||
a4f60e55 | 829 | /* We allow call stmt to have fewer arguments than the callee function |
e876d531 | 830 | (especially for K&R style programs). So bound check here (we assume |
831 | known_aggs vector, if non-NULL, has the same length as | |
832 | known_vals). */ | |
f1f41a6c | 833 | gcc_checking_assert (!known_aggs.exists () |
834 | || (known_vals.length () == known_aggs.length ())); | |
835 | if (c->operand_num >= (int) known_vals.length ()) | |
a4f60e55 | 836 | { |
837 | clause |= 1 << (i + predicate_first_dynamic_condition); | |
838 | continue; | |
839 | } | |
92ceb887 | 840 | |
a4f60e55 | 841 | if (c->agg_contents) |
842 | { | |
843 | struct ipa_agg_jump_function *agg; | |
844 | ||
845 | if (c->code == CHANGED | |
846 | && !c->by_ref | |
e876d531 | 847 | && (known_vals[c->operand_num] == error_mark_node)) |
a4f60e55 | 848 | continue; |
849 | ||
f1f41a6c | 850 | if (known_aggs.exists ()) |
a4f60e55 | 851 | { |
f1f41a6c | 852 | agg = known_aggs[c->operand_num]; |
a4f60e55 | 853 | val = ipa_find_agg_cst_for_param (agg, c->offset, c->by_ref); |
854 | } | |
855 | else | |
856 | val = NULL_TREE; | |
857 | } | |
858 | else | |
859 | { | |
f1f41a6c | 860 | val = known_vals[c->operand_num]; |
a4f60e55 | 861 | if (val == error_mark_node && c->code != CHANGED) |
862 | val = NULL_TREE; | |
863 | } | |
eb4ae064 | 864 | |
8bae3ea4 | 865 | if (!val) |
866 | { | |
867 | clause |= 1 << (i + predicate_first_dynamic_condition); | |
868 | continue; | |
869 | } | |
eb4ae064 | 870 | if (c->code == IS_NOT_CONSTANT || c->code == CHANGED) |
8bae3ea4 | 871 | continue; |
0b6455f7 | 872 | |
873 | if (operand_equal_p (TYPE_SIZE (TREE_TYPE (c->val)), | |
874 | TYPE_SIZE (TREE_TYPE (val)), 0)) | |
875 | { | |
876 | val = fold_unary (VIEW_CONVERT_EXPR, TREE_TYPE (c->val), val); | |
877 | ||
878 | res = val | |
879 | ? fold_binary_to_constant (c->code, boolean_type_node, val, c->val) | |
880 | : NULL; | |
881 | ||
882 | if (res && integer_zerop (res)) | |
883 | continue; | |
884 | } | |
8bae3ea4 | 885 | clause |= 1 << (i + predicate_first_dynamic_condition); |
886 | } | |
887 | return clause; | |
888 | } | |
889 | ||
890 | ||
a41f2a28 | 891 | /* Work out what conditions might be true at invocation of E. */ |
892 | ||
20da2013 | 893 | static void |
894 | evaluate_properties_for_edge (struct cgraph_edge *e, bool inline_p, | |
e876d531 | 895 | clause_t *clause_ptr, |
896 | vec<tree> *known_vals_ptr, | |
245ab191 | 897 | vec<ipa_polymorphic_call_context> |
898 | *known_contexts_ptr, | |
e876d531 | 899 | vec<ipa_agg_jump_function_p> *known_aggs_ptr) |
a41f2a28 | 900 | { |
415d1b9a | 901 | struct cgraph_node *callee = e->callee->ultimate_alias_target (); |
b4bae7a0 | 902 | struct inline_summary *info = inline_summaries->get (callee); |
1e094109 | 903 | vec<tree> known_vals = vNULL; |
904 | vec<ipa_agg_jump_function_p> known_aggs = vNULL; | |
a41f2a28 | 905 | |
20da2013 | 906 | if (clause_ptr) |
907 | *clause_ptr = inline_p ? 0 : 1 << predicate_not_inlined_condition; | |
908 | if (known_vals_ptr) | |
f1f41a6c | 909 | known_vals_ptr->create (0); |
245ab191 | 910 | if (known_contexts_ptr) |
911 | known_contexts_ptr->create (0); | |
20da2013 | 912 | |
2cc80ac3 | 913 | if (ipa_node_params_sum |
e67e73bd | 914 | && !e->call_stmt_cannot_inline_p |
245ab191 | 915 | && ((clause_ptr && info->conds) || known_vals_ptr || known_contexts_ptr)) |
a41f2a28 | 916 | { |
917 | struct ipa_node_params *parms_info; | |
918 | struct ipa_edge_args *args = IPA_EDGE_REF (e); | |
eb4ae064 | 919 | struct inline_edge_summary *es = inline_edge_summary (e); |
a41f2a28 | 920 | int i, count = ipa_get_cs_argument_count (args); |
a41f2a28 | 921 | |
922 | if (e->caller->global.inlined_to) | |
e876d531 | 923 | parms_info = IPA_NODE_REF (e->caller->global.inlined_to); |
a41f2a28 | 924 | else |
e876d531 | 925 | parms_info = IPA_NODE_REF (e->caller); |
a41f2a28 | 926 | |
20da2013 | 927 | if (count && (info->conds || known_vals_ptr)) |
f1f41a6c | 928 | known_vals.safe_grow_cleared (count); |
a4f60e55 | 929 | if (count && (info->conds || known_aggs_ptr)) |
f1f41a6c | 930 | known_aggs.safe_grow_cleared (count); |
245ab191 | 931 | if (count && known_contexts_ptr) |
932 | known_contexts_ptr->safe_grow_cleared (count); | |
20da2013 | 933 | |
a41f2a28 | 934 | for (i = 0; i < count; i++) |
935 | { | |
a4f60e55 | 936 | struct ipa_jump_func *jf = ipa_get_ith_jump_func (args, i); |
937 | tree cst = ipa_value_from_jfunc (parms_info, jf); | |
f4e523eb | 938 | |
939 | if (!cst && e->call_stmt | |
940 | && i < (int)gimple_call_num_args (e->call_stmt)) | |
941 | { | |
942 | cst = gimple_call_arg (e->call_stmt, i); | |
943 | if (!is_gimple_min_invariant (cst)) | |
944 | cst = NULL; | |
945 | } | |
93f713da | 946 | if (cst) |
20da2013 | 947 | { |
245ab191 | 948 | gcc_checking_assert (TREE_CODE (cst) != TREE_BINFO); |
949 | if (known_vals.exists ()) | |
f1f41a6c | 950 | known_vals[i] = cst; |
20da2013 | 951 | } |
f1f41a6c | 952 | else if (inline_p && !es->param[i].change_prob) |
953 | known_vals[i] = error_mark_node; | |
245ab191 | 954 | |
955 | if (known_contexts_ptr) | |
956 | (*known_contexts_ptr)[i] = ipa_context_from_jfunc (parms_info, e, | |
957 | i, jf); | |
a4f60e55 | 958 | /* TODO: When IPA-CP starts propagating and merging aggregate jump |
959 | functions, use its knowledge of the caller too, just like the | |
960 | scalar case above. */ | |
f1f41a6c | 961 | known_aggs[i] = &jf->agg; |
a41f2a28 | 962 | } |
a41f2a28 | 963 | } |
f4e523eb | 964 | else if (e->call_stmt && !e->call_stmt_cannot_inline_p |
965 | && ((clause_ptr && info->conds) || known_vals_ptr)) | |
966 | { | |
967 | int i, count = (int)gimple_call_num_args (e->call_stmt); | |
968 | ||
969 | if (count && (info->conds || known_vals_ptr)) | |
970 | known_vals.safe_grow_cleared (count); | |
971 | for (i = 0; i < count; i++) | |
972 | { | |
973 | tree cst = gimple_call_arg (e->call_stmt, i); | |
974 | if (!is_gimple_min_invariant (cst)) | |
975 | cst = NULL; | |
976 | if (cst) | |
977 | known_vals[i] = cst; | |
978 | } | |
979 | } | |
a41f2a28 | 980 | |
e67e73bd | 981 | if (clause_ptr) |
982 | *clause_ptr = evaluate_conditions_for_known_args (callee, inline_p, | |
a4f60e55 | 983 | known_vals, known_aggs); |
e67e73bd | 984 | |
985 | if (known_vals_ptr) | |
986 | *known_vals_ptr = known_vals; | |
987 | else | |
f1f41a6c | 988 | known_vals.release (); |
a4f60e55 | 989 | |
990 | if (known_aggs_ptr) | |
991 | *known_aggs_ptr = known_aggs; | |
992 | else | |
f1f41a6c | 993 | known_aggs.release (); |
a41f2a28 | 994 | } |
995 | ||
c7b2cc59 | 996 | |
997 | /* Allocate the inline summary vector or resize it to cover all cgraph nodes. */ | |
998 | ||
999 | static void | |
1000 | inline_summary_alloc (void) | |
1001 | { | |
0835ad03 | 1002 | if (!edge_removal_hook_holder) |
1003 | edge_removal_hook_holder = | |
35ee1c66 | 1004 | symtab->add_edge_removal_hook (&inline_edge_removal_hook, NULL); |
0835ad03 | 1005 | if (!edge_duplication_hook_holder) |
1006 | edge_duplication_hook_holder = | |
35ee1c66 | 1007 | symtab->add_edge_duplication_hook (&inline_edge_duplication_hook, NULL); |
c7b2cc59 | 1008 | |
b4bae7a0 | 1009 | if (!inline_summaries) |
1010 | inline_summaries = (inline_summary_t*) inline_summary_t::create_ggc (symtab); | |
1011 | ||
35ee1c66 | 1012 | if (inline_edge_summary_vec.length () <= (unsigned) symtab->edges_max_uid) |
1013 | inline_edge_summary_vec.safe_grow_cleared (symtab->edges_max_uid + 1); | |
c7b2cc59 | 1014 | } |
1015 | ||
3b9dd281 | 1016 | /* We are called multiple time for given function; clear |
1017 | data from previous run so they are not cumulated. */ | |
1018 | ||
1019 | static void | |
1020 | reset_inline_edge_summary (struct cgraph_edge *e) | |
1021 | { | |
e876d531 | 1022 | if (e->uid < (int) inline_edge_summary_vec.length ()) |
d10a25bb | 1023 | { |
1024 | struct inline_edge_summary *es = inline_edge_summary (e); | |
3b9dd281 | 1025 | |
563fae60 | 1026 | es->call_stmt_size = es->call_stmt_time = 0; |
d10a25bb | 1027 | if (es->predicate) |
2a5261f7 | 1028 | edge_predicate_pool.remove (es->predicate); |
d10a25bb | 1029 | es->predicate = NULL; |
f1f41a6c | 1030 | es->param.release (); |
d10a25bb | 1031 | } |
3b9dd281 | 1032 | } |
1033 | ||
1034 | /* We are called multiple time for given function; clear | |
1035 | data from previous run so they are not cumulated. */ | |
1036 | ||
1037 | static void | |
b4bae7a0 | 1038 | reset_inline_summary (struct cgraph_node *node, |
1039 | inline_summary *info) | |
3b9dd281 | 1040 | { |
3b9dd281 | 1041 | struct cgraph_edge *e; |
1042 | ||
1043 | info->self_size = info->self_time = 0; | |
1044 | info->estimated_stack_size = 0; | |
1045 | info->estimated_self_stack_size = 0; | |
1046 | info->stack_frame_offset = 0; | |
1047 | info->size = 0; | |
1048 | info->time = 0; | |
3172b7bf | 1049 | info->growth = 0; |
41d39f38 | 1050 | info->scc_no = 0; |
7c07aa3d | 1051 | if (info->loop_iterations) |
1052 | { | |
2a5261f7 | 1053 | edge_predicate_pool.remove (info->loop_iterations); |
7c07aa3d | 1054 | info->loop_iterations = NULL; |
1055 | } | |
3716ee8f | 1056 | if (info->loop_stride) |
1057 | { | |
2a5261f7 | 1058 | edge_predicate_pool.remove (info->loop_stride); |
3716ee8f | 1059 | info->loop_stride = NULL; |
1060 | } | |
be343a9c | 1061 | if (info->array_index) |
1062 | { | |
2a5261f7 | 1063 | edge_predicate_pool.remove (info->array_index); |
be343a9c | 1064 | info->array_index = NULL; |
1065 | } | |
f1f41a6c | 1066 | vec_free (info->conds); |
1067 | vec_free (info->entry); | |
3b9dd281 | 1068 | for (e = node->callees; e; e = e->next_callee) |
1069 | reset_inline_edge_summary (e); | |
1070 | for (e = node->indirect_calls; e; e = e->next_callee) | |
1071 | reset_inline_edge_summary (e); | |
1072 | } | |
1073 | ||
c7b2cc59 | 1074 | /* Hook that is called by cgraph.c when a node is removed. */ |
1075 | ||
b4bae7a0 | 1076 | void |
1077 | inline_summary_t::remove (cgraph_node *node, inline_summary *info) | |
c7b2cc59 | 1078 | { |
b4bae7a0 | 1079 | reset_inline_summary (node, info); |
c7b2cc59 | 1080 | } |
1081 | ||
74d6d072 | 1082 | /* Remap predicate P of former function to be predicate of duplicated function. |
3716ee8f | 1083 | POSSIBLE_TRUTHS is clause of possible truths in the duplicated node, |
1084 | INFO is inline summary of the duplicated node. */ | |
1085 | ||
1086 | static struct predicate | |
1087 | remap_predicate_after_duplication (struct predicate *p, | |
1088 | clause_t possible_truths, | |
1089 | struct inline_summary *info) | |
1090 | { | |
1091 | struct predicate new_predicate = true_predicate (); | |
1092 | int j; | |
1093 | for (j = 0; p->clause[j]; j++) | |
1094 | if (!(possible_truths & p->clause[j])) | |
1095 | { | |
1096 | new_predicate = false_predicate (); | |
1097 | break; | |
1098 | } | |
1099 | else | |
1100 | add_clause (info->conds, &new_predicate, | |
1101 | possible_truths & p->clause[j]); | |
1102 | return new_predicate; | |
1103 | } | |
1104 | ||
1105 | /* Same as remap_predicate_after_duplication but handle hint predicate *P. | |
1106 | Additionally care about allocating new memory slot for updated predicate | |
1107 | and set it to NULL when it becomes true or false (and thus uninteresting). | |
1108 | */ | |
1109 | ||
1110 | static void | |
1111 | remap_hint_predicate_after_duplication (struct predicate **p, | |
1112 | clause_t possible_truths, | |
1113 | struct inline_summary *info) | |
1114 | { | |
1115 | struct predicate new_predicate; | |
1116 | ||
1117 | if (!*p) | |
1118 | return; | |
1119 | ||
1120 | new_predicate = remap_predicate_after_duplication (*p, | |
e876d531 | 1121 | possible_truths, info); |
3716ee8f | 1122 | /* We do not want to free previous predicate; it is used by node origin. */ |
1123 | *p = NULL; | |
1124 | set_hint_predicate (p, new_predicate); | |
1125 | } | |
1126 | ||
0835ad03 | 1127 | |
c7b2cc59 | 1128 | /* Hook that is called by cgraph.c when a node is duplicated. */ |
b4bae7a0 | 1129 | void |
1130 | inline_summary_t::duplicate (cgraph_node *src, | |
1131 | cgraph_node *dst, | |
1132 | inline_summary *, | |
1133 | inline_summary *info) | |
c7b2cc59 | 1134 | { |
1135 | inline_summary_alloc (); | |
b4bae7a0 | 1136 | memcpy (info, inline_summaries->get (src), sizeof (inline_summary)); |
8bae3ea4 | 1137 | /* TODO: as an optimization, we may avoid copying conditions |
1138 | that are known to be false or true. */ | |
f1f41a6c | 1139 | info->conds = vec_safe_copy (info->conds); |
8bae3ea4 | 1140 | |
1141 | /* When there are any replacements in the function body, see if we can figure | |
1142 | out that something was optimized out. */ | |
2cc80ac3 | 1143 | if (ipa_node_params_sum && dst->clone.tree_map) |
8bae3ea4 | 1144 | { |
f1f41a6c | 1145 | vec<size_time_entry, va_gc> *entry = info->entry; |
8bae3ea4 | 1146 | /* Use SRC parm info since it may not be copied yet. */ |
1147 | struct ipa_node_params *parms_info = IPA_NODE_REF (src); | |
1e094109 | 1148 | vec<tree> known_vals = vNULL; |
8bae3ea4 | 1149 | int count = ipa_get_param_count (parms_info); |
e876d531 | 1150 | int i, j; |
8bae3ea4 | 1151 | clause_t possible_truths; |
1152 | struct predicate true_pred = true_predicate (); | |
1153 | size_time_entry *e; | |
1154 | int optimized_out_size = 0; | |
8bae3ea4 | 1155 | bool inlined_to_p = false; |
99fe8cbd | 1156 | struct cgraph_edge *edge, *next; |
8bae3ea4 | 1157 | |
839c5aac | 1158 | info->entry = 0; |
f1f41a6c | 1159 | known_vals.safe_grow_cleared (count); |
8bae3ea4 | 1160 | for (i = 0; i < count; i++) |
e876d531 | 1161 | { |
8bae3ea4 | 1162 | struct ipa_replace_map *r; |
1163 | ||
f1f41a6c | 1164 | for (j = 0; vec_safe_iterate (dst->clone.tree_map, j, &r); j++) |
8bae3ea4 | 1165 | { |
09ab6335 | 1166 | if (((!r->old_tree && r->parm_num == i) |
1167 | || (r->old_tree && r->old_tree == ipa_get_param (parms_info, i))) | |
1168 | && r->replace_p && !r->ref_p) | |
8bae3ea4 | 1169 | { |
f1f41a6c | 1170 | known_vals[i] = r->new_tree; |
8bae3ea4 | 1171 | break; |
1172 | } | |
1173 | } | |
1174 | } | |
a4f60e55 | 1175 | possible_truths = evaluate_conditions_for_known_args (dst, false, |
e876d531 | 1176 | known_vals, |
1177 | vNULL); | |
f1f41a6c | 1178 | known_vals.release (); |
8bae3ea4 | 1179 | |
1180 | account_size_time (info, 0, 0, &true_pred); | |
1181 | ||
1182 | /* Remap size_time vectors. | |
e876d531 | 1183 | Simplify the predicate by prunning out alternatives that are known |
1184 | to be false. | |
1185 | TODO: as on optimization, we can also eliminate conditions known | |
1186 | to be true. */ | |
f1f41a6c | 1187 | for (i = 0; vec_safe_iterate (entry, i, &e); i++) |
8bae3ea4 | 1188 | { |
3716ee8f | 1189 | struct predicate new_predicate; |
1190 | new_predicate = remap_predicate_after_duplication (&e->predicate, | |
1191 | possible_truths, | |
1192 | info); | |
8bae3ea4 | 1193 | if (false_predicate_p (&new_predicate)) |
18b64b34 | 1194 | optimized_out_size += e->size; |
8bae3ea4 | 1195 | else |
1196 | account_size_time (info, e->size, e->time, &new_predicate); | |
1197 | } | |
1198 | ||
fb3c587e | 1199 | /* Remap edge predicates with the same simplification as above. |
e876d531 | 1200 | Also copy constantness arrays. */ |
99fe8cbd | 1201 | for (edge = dst->callees; edge; edge = next) |
8bae3ea4 | 1202 | { |
3716ee8f | 1203 | struct predicate new_predicate; |
8bae3ea4 | 1204 | struct inline_edge_summary *es = inline_edge_summary (edge); |
99fe8cbd | 1205 | next = edge->next_callee; |
8bae3ea4 | 1206 | |
1207 | if (!edge->inline_failed) | |
1208 | inlined_to_p = true; | |
1209 | if (!es->predicate) | |
1210 | continue; | |
3716ee8f | 1211 | new_predicate = remap_predicate_after_duplication (es->predicate, |
1212 | possible_truths, | |
1213 | info); | |
8bae3ea4 | 1214 | if (false_predicate_p (&new_predicate) |
1215 | && !false_predicate_p (es->predicate)) | |
e806c56f | 1216 | optimized_out_size += es->call_stmt_size * INLINE_SIZE_SCALE; |
7c07aa3d | 1217 | edge_set_predicate (edge, &new_predicate); |
8bae3ea4 | 1218 | } |
1219 | ||
fb3c587e | 1220 | /* Remap indirect edge predicates with the same simplificaiton as above. |
e876d531 | 1221 | Also copy constantness arrays. */ |
99fe8cbd | 1222 | for (edge = dst->indirect_calls; edge; edge = next) |
8bae3ea4 | 1223 | { |
3716ee8f | 1224 | struct predicate new_predicate; |
8bae3ea4 | 1225 | struct inline_edge_summary *es = inline_edge_summary (edge); |
99fe8cbd | 1226 | next = edge->next_callee; |
8bae3ea4 | 1227 | |
3716ee8f | 1228 | gcc_checking_assert (edge->inline_failed); |
8bae3ea4 | 1229 | if (!es->predicate) |
1230 | continue; | |
3716ee8f | 1231 | new_predicate = remap_predicate_after_duplication (es->predicate, |
1232 | possible_truths, | |
1233 | info); | |
8bae3ea4 | 1234 | if (false_predicate_p (&new_predicate) |
1235 | && !false_predicate_p (es->predicate)) | |
e806c56f | 1236 | optimized_out_size += es->call_stmt_size * INLINE_SIZE_SCALE; |
7c07aa3d | 1237 | edge_set_predicate (edge, &new_predicate); |
1238 | } | |
3716ee8f | 1239 | remap_hint_predicate_after_duplication (&info->loop_iterations, |
e876d531 | 1240 | possible_truths, info); |
3716ee8f | 1241 | remap_hint_predicate_after_duplication (&info->loop_stride, |
e876d531 | 1242 | possible_truths, info); |
be343a9c | 1243 | remap_hint_predicate_after_duplication (&info->array_index, |
e876d531 | 1244 | possible_truths, info); |
8bae3ea4 | 1245 | |
1246 | /* If inliner or someone after inliner will ever start producing | |
e876d531 | 1247 | non-trivial clones, we will get trouble with lack of information |
1248 | about updating self sizes, because size vectors already contains | |
1249 | sizes of the calees. */ | |
1250 | gcc_assert (!inlined_to_p || !optimized_out_size); | |
8bae3ea4 | 1251 | } |
1252 | else | |
7c07aa3d | 1253 | { |
f1f41a6c | 1254 | info->entry = vec_safe_copy (info->entry); |
7c07aa3d | 1255 | if (info->loop_iterations) |
1256 | { | |
1257 | predicate p = *info->loop_iterations; | |
3716ee8f | 1258 | info->loop_iterations = NULL; |
1259 | set_hint_predicate (&info->loop_iterations, p); | |
1260 | } | |
1261 | if (info->loop_stride) | |
1262 | { | |
1263 | predicate p = *info->loop_stride; | |
1264 | info->loop_stride = NULL; | |
1265 | set_hint_predicate (&info->loop_stride, p); | |
7c07aa3d | 1266 | } |
be343a9c | 1267 | if (info->array_index) |
1268 | { | |
1269 | predicate p = *info->array_index; | |
1270 | info->array_index = NULL; | |
1271 | set_hint_predicate (&info->array_index, p); | |
1272 | } | |
7c07aa3d | 1273 | } |
9dcc8702 | 1274 | if (!dst->global.inlined_to) |
1275 | inline_update_overall_summary (dst); | |
a41f2a28 | 1276 | } |
1277 | ||
1278 | ||
0835ad03 | 1279 | /* Hook that is called by cgraph.c when a node is duplicated. */ |
1280 | ||
1281 | static void | |
e876d531 | 1282 | inline_edge_duplication_hook (struct cgraph_edge *src, |
1283 | struct cgraph_edge *dst, | |
0835ad03 | 1284 | ATTRIBUTE_UNUSED void *data) |
1285 | { | |
1286 | struct inline_edge_summary *info; | |
6a18c0be | 1287 | struct inline_edge_summary *srcinfo; |
0835ad03 | 1288 | inline_summary_alloc (); |
1289 | info = inline_edge_summary (dst); | |
6a18c0be | 1290 | srcinfo = inline_edge_summary (src); |
e876d531 | 1291 | memcpy (info, srcinfo, sizeof (struct inline_edge_summary)); |
6a18c0be | 1292 | info->predicate = NULL; |
1293 | edge_set_predicate (dst, srcinfo->predicate); | |
f1f41a6c | 1294 | info->param = srcinfo->param.copy (); |
1a92a535 | 1295 | if (!dst->indirect_unknown_callee && src->indirect_unknown_callee) |
1296 | { | |
1297 | info->call_stmt_size -= (eni_size_weights.indirect_call_cost | |
1298 | - eni_size_weights.call_cost); | |
1299 | info->call_stmt_time -= (eni_time_weights.indirect_call_cost | |
1300 | - eni_time_weights.call_cost); | |
1301 | } | |
0835ad03 | 1302 | } |
1303 | ||
1304 | ||
a41f2a28 | 1305 | /* Keep edge cache consistent across edge removal. */ |
1306 | ||
1307 | static void | |
e876d531 | 1308 | inline_edge_removal_hook (struct cgraph_edge *edge, |
1309 | void *data ATTRIBUTE_UNUSED) | |
a41f2a28 | 1310 | { |
f1f41a6c | 1311 | if (edge_growth_cache.exists ()) |
0835ad03 | 1312 | reset_edge_growth_cache (edge); |
d10a25bb | 1313 | reset_inline_edge_summary (edge); |
a41f2a28 | 1314 | } |
1315 | ||
1316 | ||
1317 | /* Initialize growth caches. */ | |
1318 | ||
1319 | void | |
1320 | initialize_growth_caches (void) | |
1321 | { | |
35ee1c66 | 1322 | if (symtab->edges_max_uid) |
1323 | edge_growth_cache.safe_grow_cleared (symtab->edges_max_uid); | |
a41f2a28 | 1324 | } |
1325 | ||
1326 | ||
1327 | /* Free growth caches. */ | |
1328 | ||
1329 | void | |
1330 | free_growth_caches (void) | |
1331 | { | |
f1f41a6c | 1332 | edge_growth_cache.release (); |
c7b2cc59 | 1333 | } |
1334 | ||
a41f2a28 | 1335 | |
0835ad03 | 1336 | /* Dump edge summaries associated to NODE and recursively to all clones. |
1337 | Indent by INDENT. */ | |
1338 | ||
1339 | static void | |
e876d531 | 1340 | dump_inline_edge_summary (FILE *f, int indent, struct cgraph_node *node, |
6a18c0be | 1341 | struct inline_summary *info) |
0835ad03 | 1342 | { |
1343 | struct cgraph_edge *edge; | |
1344 | for (edge = node->callees; edge; edge = edge->next_callee) | |
1345 | { | |
1346 | struct inline_edge_summary *es = inline_edge_summary (edge); | |
415d1b9a | 1347 | struct cgraph_node *callee = edge->callee->ultimate_alias_target (); |
eb4ae064 | 1348 | int i; |
1349 | ||
e876d531 | 1350 | fprintf (f, |
1351 | "%*s%s/%i %s\n%*s loop depth:%2i freq:%4i size:%2i" | |
1352 | " time: %2i callee size:%2i stack:%2i", | |
f1c8b4d7 | 1353 | indent, "", callee->name (), callee->order, |
e876d531 | 1354 | !edge->inline_failed |
1355 | ? "inlined" : cgraph_inline_failed_string (edge-> inline_failed), | |
1356 | indent, "", es->loop_depth, edge->frequency, | |
1357 | es->call_stmt_size, es->call_stmt_time, | |
b4bae7a0 | 1358 | (int) inline_summaries->get (callee)->size / INLINE_SIZE_SCALE, |
1359 | (int) inline_summaries->get (callee)->estimated_stack_size); | |
eb4ae064 | 1360 | |
6a18c0be | 1361 | if (es->predicate) |
1362 | { | |
1363 | fprintf (f, " predicate: "); | |
1364 | dump_predicate (f, info->conds, es->predicate); | |
1365 | } | |
1366 | else | |
e876d531 | 1367 | fprintf (f, "\n"); |
f1f41a6c | 1368 | if (es->param.exists ()) |
e876d531 | 1369 | for (i = 0; i < (int) es->param.length (); i++) |
eb4ae064 | 1370 | { |
f1f41a6c | 1371 | int prob = es->param[i].change_prob; |
eb4ae064 | 1372 | |
1373 | if (!prob) | |
1374 | fprintf (f, "%*s op%i is compile time invariant\n", | |
1375 | indent + 2, "", i); | |
1376 | else if (prob != REG_BR_PROB_BASE) | |
1377 | fprintf (f, "%*s op%i change %f%% of time\n", indent + 2, "", i, | |
1378 | prob * 100.0 / REG_BR_PROB_BASE); | |
1379 | } | |
0835ad03 | 1380 | if (!edge->inline_failed) |
0a0ca4d6 | 1381 | { |
e876d531 | 1382 | fprintf (f, "%*sStack frame offset %i, callee self size %i," |
eb4ae064 | 1383 | " callee size %i\n", |
e876d531 | 1384 | indent + 2, "", |
b4bae7a0 | 1385 | (int) inline_summaries->get (callee)->stack_frame_offset, |
1386 | (int) inline_summaries->get (callee)->estimated_self_stack_size, | |
1387 | (int) inline_summaries->get (callee)->estimated_stack_size); | |
e876d531 | 1388 | dump_inline_edge_summary (f, indent + 2, callee, info); |
0a0ca4d6 | 1389 | } |
0835ad03 | 1390 | } |
1391 | for (edge = node->indirect_calls; edge; edge = edge->next_callee) | |
1392 | { | |
1393 | struct inline_edge_summary *es = inline_edge_summary (edge); | |
fb3c587e | 1394 | fprintf (f, "%*sindirect call loop depth:%2i freq:%4i size:%2i" |
eb4ae064 | 1395 | " time: %2i", |
0835ad03 | 1396 | indent, "", |
e876d531 | 1397 | es->loop_depth, |
1398 | edge->frequency, es->call_stmt_size, es->call_stmt_time); | |
6a18c0be | 1399 | if (es->predicate) |
1400 | { | |
1401 | fprintf (f, "predicate: "); | |
1402 | dump_predicate (f, info->conds, es->predicate); | |
1403 | } | |
1404 | else | |
fb3c587e | 1405 | fprintf (f, "\n"); |
0835ad03 | 1406 | } |
1407 | } | |
1408 | ||
1409 | ||
0a0ca4d6 | 1410 | void |
e876d531 | 1411 | dump_inline_summary (FILE *f, struct cgraph_node *node) |
c7b2cc59 | 1412 | { |
02774f2d | 1413 | if (node->definition) |
c7b2cc59 | 1414 | { |
b4bae7a0 | 1415 | struct inline_summary *s = inline_summaries->get (node); |
a41f2a28 | 1416 | size_time_entry *e; |
1417 | int i; | |
f1c8b4d7 | 1418 | fprintf (f, "Inline summary for %s/%i", node->name (), |
02774f2d | 1419 | node->order); |
1420 | if (DECL_DISREGARD_INLINE_LIMITS (node->decl)) | |
cbd7f5a0 | 1421 | fprintf (f, " always_inline"); |
1422 | if (s->inlinable) | |
1423 | fprintf (f, " inlinable"); | |
e806c56f | 1424 | if (s->contains_cilk_spawn) |
1425 | fprintf (f, " contains_cilk_spawn"); | |
e876d531 | 1426 | fprintf (f, "\n self time: %i\n", s->self_time); |
cbd7f5a0 | 1427 | fprintf (f, " global time: %i\n", s->time); |
e876d531 | 1428 | fprintf (f, " self size: %i\n", s->self_size); |
4869c23f | 1429 | fprintf (f, " global size: %i\n", s->size); |
db197f90 | 1430 | fprintf (f, " min size: %i\n", s->min_size); |
c7b2cc59 | 1431 | fprintf (f, " self stack: %i\n", |
a41f2a28 | 1432 | (int) s->estimated_self_stack_size); |
e876d531 | 1433 | fprintf (f, " global stack: %i\n", (int) s->estimated_stack_size); |
3172b7bf | 1434 | if (s->growth) |
e876d531 | 1435 | fprintf (f, " estimated growth:%i\n", (int) s->growth); |
db2db13c | 1436 | if (s->scc_no) |
e876d531 | 1437 | fprintf (f, " In SCC: %i\n", (int) s->scc_no); |
f1f41a6c | 1438 | for (i = 0; vec_safe_iterate (s->entry, i, &e); i++) |
a41f2a28 | 1439 | { |
1440 | fprintf (f, " size:%f, time:%f, predicate:", | |
1441 | (double) e->size / INLINE_SIZE_SCALE, | |
1442 | (double) e->time / INLINE_TIME_SCALE); | |
1443 | dump_predicate (f, s->conds, &e->predicate); | |
1444 | } | |
7c07aa3d | 1445 | if (s->loop_iterations) |
1446 | { | |
1447 | fprintf (f, " loop iterations:"); | |
1448 | dump_predicate (f, s->conds, s->loop_iterations); | |
1449 | } | |
3716ee8f | 1450 | if (s->loop_stride) |
1451 | { | |
1452 | fprintf (f, " loop stride:"); | |
1453 | dump_predicate (f, s->conds, s->loop_stride); | |
1454 | } | |
be343a9c | 1455 | if (s->array_index) |
1456 | { | |
1457 | fprintf (f, " array index:"); | |
1458 | dump_predicate (f, s->conds, s->array_index); | |
1459 | } | |
0835ad03 | 1460 | fprintf (f, " calls:\n"); |
6a18c0be | 1461 | dump_inline_edge_summary (f, 4, node, s); |
a41f2a28 | 1462 | fprintf (f, "\n"); |
c7b2cc59 | 1463 | } |
1464 | } | |
1465 | ||
0a0ca4d6 | 1466 | DEBUG_FUNCTION void |
c7b2cc59 | 1467 | debug_inline_summary (struct cgraph_node *node) |
1468 | { | |
1469 | dump_inline_summary (stderr, node); | |
1470 | } | |
1471 | ||
1472 | void | |
1473 | dump_inline_summaries (FILE *f) | |
1474 | { | |
1475 | struct cgraph_node *node; | |
1476 | ||
7c455d87 | 1477 | FOR_EACH_DEFINED_FUNCTION (node) |
1478 | if (!node->global.inlined_to) | |
c7b2cc59 | 1479 | dump_inline_summary (f, node); |
1480 | } | |
99c67f24 | 1481 | |
cbd7f5a0 | 1482 | /* Give initial reasons why inlining would fail on EDGE. This gets either |
1483 | nullified or usually overwritten by more precise reasons later. */ | |
1484 | ||
1485 | void | |
1486 | initialize_inline_failed (struct cgraph_edge *e) | |
1487 | { | |
1488 | struct cgraph_node *callee = e->callee; | |
1489 | ||
1490 | if (e->indirect_unknown_callee) | |
1491 | e->inline_failed = CIF_INDIRECT_UNKNOWN_CALL; | |
02774f2d | 1492 | else if (!callee->definition) |
cbd7f5a0 | 1493 | e->inline_failed = CIF_BODY_NOT_AVAILABLE; |
1494 | else if (callee->local.redefined_extern_inline) | |
1495 | e->inline_failed = CIF_REDEFINED_EXTERN_INLINE; | |
f883da84 | 1496 | else if (e->call_stmt_cannot_inline_p) |
cbd7f5a0 | 1497 | e->inline_failed = CIF_MISMATCHED_ARGUMENTS; |
d037099f | 1498 | else if (cfun && fn_contains_cilk_spawn_p (cfun)) |
1499 | /* We can't inline if the function is spawing a function. */ | |
1500 | e->inline_failed = CIF_FUNCTION_NOT_INLINABLE; | |
cbd7f5a0 | 1501 | else |
1502 | e->inline_failed = CIF_FUNCTION_NOT_CONSIDERED; | |
1503 | } | |
1504 | ||
94646c9c | 1505 | /* Callback of walk_aliased_vdefs. Flags that it has been invoked to the |
1506 | boolean variable pointed to by DATA. */ | |
1507 | ||
1508 | static bool | |
1509 | mark_modified (ao_ref *ao ATTRIBUTE_UNUSED, tree vdef ATTRIBUTE_UNUSED, | |
e876d531 | 1510 | void *data) |
94646c9c | 1511 | { |
1512 | bool *b = (bool *) data; | |
1513 | *b = true; | |
1514 | return true; | |
1515 | } | |
1516 | ||
a4f60e55 | 1517 | /* If OP refers to value of function parameter, return the corresponding |
1518 | parameter. */ | |
94646c9c | 1519 | |
1520 | static tree | |
42acab1c | 1521 | unmodified_parm_1 (gimple *stmt, tree op) |
94646c9c | 1522 | { |
1523 | /* SSA_NAME referring to parm default def? */ | |
1524 | if (TREE_CODE (op) == SSA_NAME | |
1525 | && SSA_NAME_IS_DEFAULT_DEF (op) | |
1526 | && TREE_CODE (SSA_NAME_VAR (op)) == PARM_DECL) | |
1527 | return SSA_NAME_VAR (op); | |
1528 | /* Non-SSA parm reference? */ | |
1529 | if (TREE_CODE (op) == PARM_DECL) | |
1530 | { | |
1531 | bool modified = false; | |
1532 | ||
1533 | ao_ref refd; | |
1534 | ao_ref_init (&refd, op); | |
1535 | walk_aliased_vdefs (&refd, gimple_vuse (stmt), mark_modified, &modified, | |
1536 | NULL); | |
1537 | if (!modified) | |
1538 | return op; | |
1539 | } | |
a4f60e55 | 1540 | return NULL_TREE; |
1541 | } | |
1542 | ||
1543 | /* If OP refers to value of function parameter, return the corresponding | |
1544 | parameter. Also traverse chains of SSA register assignments. */ | |
1545 | ||
1546 | static tree | |
42acab1c | 1547 | unmodified_parm (gimple *stmt, tree op) |
a4f60e55 | 1548 | { |
1549 | tree res = unmodified_parm_1 (stmt, op); | |
1550 | if (res) | |
1551 | return res; | |
1552 | ||
94646c9c | 1553 | if (TREE_CODE (op) == SSA_NAME |
1554 | && !SSA_NAME_IS_DEFAULT_DEF (op) | |
1555 | && gimple_assign_single_p (SSA_NAME_DEF_STMT (op))) | |
1556 | return unmodified_parm (SSA_NAME_DEF_STMT (op), | |
1557 | gimple_assign_rhs1 (SSA_NAME_DEF_STMT (op))); | |
a4f60e55 | 1558 | return NULL_TREE; |
1559 | } | |
1560 | ||
1561 | /* If OP refers to a value of a function parameter or value loaded from an | |
1562 | aggregate passed to a parameter (either by value or reference), return TRUE | |
1563 | and store the number of the parameter to *INDEX_P and information whether | |
1564 | and how it has been loaded from an aggregate into *AGGPOS. INFO describes | |
1565 | the function parameters, STMT is the statement in which OP is used or | |
1566 | loaded. */ | |
1567 | ||
1568 | static bool | |
9ea91b78 | 1569 | unmodified_parm_or_parm_agg_item (struct ipa_func_body_info *fbi, |
42acab1c | 1570 | gimple *stmt, tree op, int *index_p, |
a4f60e55 | 1571 | struct agg_position_info *aggpos) |
1572 | { | |
1573 | tree res = unmodified_parm_1 (stmt, op); | |
1574 | ||
1575 | gcc_checking_assert (aggpos); | |
1576 | if (res) | |
1577 | { | |
1a673ff0 | 1578 | *index_p = ipa_get_param_decl_index (fbi->info, res); |
a4f60e55 | 1579 | if (*index_p < 0) |
1580 | return false; | |
1581 | aggpos->agg_contents = false; | |
1582 | aggpos->by_ref = false; | |
1583 | return true; | |
1584 | } | |
1585 | ||
1586 | if (TREE_CODE (op) == SSA_NAME) | |
1587 | { | |
1588 | if (SSA_NAME_IS_DEFAULT_DEF (op) | |
1589 | || !gimple_assign_single_p (SSA_NAME_DEF_STMT (op))) | |
1590 | return false; | |
1591 | stmt = SSA_NAME_DEF_STMT (op); | |
1592 | op = gimple_assign_rhs1 (stmt); | |
1593 | if (!REFERENCE_CLASS_P (op)) | |
1a673ff0 | 1594 | return unmodified_parm_or_parm_agg_item (fbi, stmt, op, index_p, |
a4f60e55 | 1595 | aggpos); |
1596 | } | |
1597 | ||
1598 | aggpos->agg_contents = true; | |
1a673ff0 | 1599 | return ipa_load_from_parm_agg (fbi, fbi->info->descriptors, |
1600 | stmt, op, index_p, &aggpos->offset, | |
1601 | NULL, &aggpos->by_ref); | |
94646c9c | 1602 | } |
1603 | ||
99c67f24 | 1604 | /* See if statement might disappear after inlining. |
1605 | 0 - means not eliminated | |
1606 | 1 - half of statements goes away | |
1607 | 2 - for sure it is eliminated. | |
1608 | We are not terribly sophisticated, basically looking for simple abstraction | |
1609 | penalty wrappers. */ | |
1610 | ||
1611 | static int | |
42acab1c | 1612 | eliminated_by_inlining_prob (gimple *stmt) |
99c67f24 | 1613 | { |
1614 | enum gimple_code code = gimple_code (stmt); | |
11f20fba | 1615 | enum tree_code rhs_code; |
94646c9c | 1616 | |
1617 | if (!optimize) | |
1618 | return 0; | |
1619 | ||
99c67f24 | 1620 | switch (code) |
1621 | { | |
e876d531 | 1622 | case GIMPLE_RETURN: |
1623 | return 2; | |
1624 | case GIMPLE_ASSIGN: | |
1625 | if (gimple_num_ops (stmt) != 2) | |
99c67f24 | 1626 | return 0; |
e876d531 | 1627 | |
1628 | rhs_code = gimple_assign_rhs_code (stmt); | |
1629 | ||
1630 | /* Casts of parameters, loads from parameters passed by reference | |
1631 | and stores to return value or parameters are often free after | |
1632 | inlining dua to SRA and further combining. | |
1633 | Assume that half of statements goes away. */ | |
d09ef31a | 1634 | if (CONVERT_EXPR_CODE_P (rhs_code) |
e876d531 | 1635 | || rhs_code == VIEW_CONVERT_EXPR |
1636 | || rhs_code == ADDR_EXPR | |
1637 | || gimple_assign_rhs_class (stmt) == GIMPLE_SINGLE_RHS) | |
1638 | { | |
1639 | tree rhs = gimple_assign_rhs1 (stmt); | |
1640 | tree lhs = gimple_assign_lhs (stmt); | |
1641 | tree inner_rhs = get_base_address (rhs); | |
1642 | tree inner_lhs = get_base_address (lhs); | |
1643 | bool rhs_free = false; | |
1644 | bool lhs_free = false; | |
1645 | ||
1646 | if (!inner_rhs) | |
1647 | inner_rhs = rhs; | |
1648 | if (!inner_lhs) | |
1649 | inner_lhs = lhs; | |
1650 | ||
1651 | /* Reads of parameter are expected to be free. */ | |
1652 | if (unmodified_parm (stmt, inner_rhs)) | |
1653 | rhs_free = true; | |
1654 | /* Match expressions of form &this->field. Those will most likely | |
1655 | combine with something upstream after inlining. */ | |
1656 | else if (TREE_CODE (inner_rhs) == ADDR_EXPR) | |
1657 | { | |
1658 | tree op = get_base_address (TREE_OPERAND (inner_rhs, 0)); | |
1659 | if (TREE_CODE (op) == PARM_DECL) | |
1660 | rhs_free = true; | |
1661 | else if (TREE_CODE (op) == MEM_REF | |
1662 | && unmodified_parm (stmt, TREE_OPERAND (op, 0))) | |
1663 | rhs_free = true; | |
1664 | } | |
1665 | ||
1666 | /* When parameter is not SSA register because its address is taken | |
1667 | and it is just copied into one, the statement will be completely | |
1668 | free after inlining (we will copy propagate backward). */ | |
1669 | if (rhs_free && is_gimple_reg (lhs)) | |
1670 | return 2; | |
1671 | ||
1672 | /* Reads of parameters passed by reference | |
1673 | expected to be free (i.e. optimized out after inlining). */ | |
1674 | if (TREE_CODE (inner_rhs) == MEM_REF | |
1675 | && unmodified_parm (stmt, TREE_OPERAND (inner_rhs, 0))) | |
1676 | rhs_free = true; | |
1677 | ||
1678 | /* Copying parameter passed by reference into gimple register is | |
1679 | probably also going to copy propagate, but we can't be quite | |
1680 | sure. */ | |
1681 | if (rhs_free && is_gimple_reg (lhs)) | |
1682 | lhs_free = true; | |
1683 | ||
1684 | /* Writes to parameters, parameters passed by value and return value | |
1685 | (either dirrectly or passed via invisible reference) are free. | |
1686 | ||
1687 | TODO: We ought to handle testcase like | |
1688 | struct a {int a,b;}; | |
1689 | struct a | |
1690 | retrurnsturct (void) | |
1691 | { | |
1692 | struct a a ={1,2}; | |
1693 | return a; | |
1694 | } | |
1695 | ||
1696 | This translate into: | |
1697 | ||
1698 | retrurnsturct () | |
1699 | { | |
1700 | int a$b; | |
1701 | int a$a; | |
1702 | struct a a; | |
1703 | struct a D.2739; | |
1704 | ||
1705 | <bb 2>: | |
1706 | D.2739.a = 1; | |
1707 | D.2739.b = 2; | |
1708 | return D.2739; | |
1709 | ||
1710 | } | |
1711 | For that we either need to copy ipa-split logic detecting writes | |
1712 | to return value. */ | |
1713 | if (TREE_CODE (inner_lhs) == PARM_DECL | |
1714 | || TREE_CODE (inner_lhs) == RESULT_DECL | |
1715 | || (TREE_CODE (inner_lhs) == MEM_REF | |
1716 | && (unmodified_parm (stmt, TREE_OPERAND (inner_lhs, 0)) | |
1717 | || (TREE_CODE (TREE_OPERAND (inner_lhs, 0)) == SSA_NAME | |
1718 | && SSA_NAME_VAR (TREE_OPERAND (inner_lhs, 0)) | |
1719 | && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND | |
1720 | (inner_lhs, | |
1721 | 0))) == RESULT_DECL)))) | |
1722 | lhs_free = true; | |
1723 | if (lhs_free | |
1724 | && (is_gimple_reg (rhs) || is_gimple_min_invariant (rhs))) | |
1725 | rhs_free = true; | |
1726 | if (lhs_free && rhs_free) | |
1727 | return 1; | |
1728 | } | |
1729 | return 0; | |
1730 | default: | |
1731 | return 0; | |
99c67f24 | 1732 | } |
1733 | } | |
1734 | ||
1735 | ||
905aa3bd | 1736 | /* If BB ends by a conditional we can turn into predicates, attach corresponding |
1737 | predicates to the CFG edges. */ | |
a41f2a28 | 1738 | |
905aa3bd | 1739 | static void |
9ea91b78 | 1740 | set_cond_stmt_execution_predicate (struct ipa_func_body_info *fbi, |
e876d531 | 1741 | struct inline_summary *summary, |
1742 | basic_block bb) | |
a41f2a28 | 1743 | { |
42acab1c | 1744 | gimple *last; |
a41f2a28 | 1745 | tree op; |
1746 | int index; | |
a4f60e55 | 1747 | struct agg_position_info aggpos; |
905aa3bd | 1748 | enum tree_code code, inverted_code; |
1749 | edge e; | |
1750 | edge_iterator ei; | |
42acab1c | 1751 | gimple *set_stmt; |
905aa3bd | 1752 | tree op2; |
a41f2a28 | 1753 | |
905aa3bd | 1754 | last = last_stmt (bb); |
e876d531 | 1755 | if (!last || gimple_code (last) != GIMPLE_COND) |
905aa3bd | 1756 | return; |
a41f2a28 | 1757 | if (!is_gimple_ip_invariant (gimple_cond_rhs (last))) |
905aa3bd | 1758 | return; |
a41f2a28 | 1759 | op = gimple_cond_lhs (last); |
1760 | /* TODO: handle conditionals like | |
1761 | var = op0 < 4; | |
905aa3bd | 1762 | if (var != 0). */ |
1a673ff0 | 1763 | if (unmodified_parm_or_parm_agg_item (fbi, last, op, &index, &aggpos)) |
905aa3bd | 1764 | { |
905aa3bd | 1765 | code = gimple_cond_code (last); |
93633022 | 1766 | inverted_code = invert_tree_comparison (code, HONOR_NANS (op)); |
905aa3bd | 1767 | |
1768 | FOR_EACH_EDGE (e, ei, bb->succs) | |
1769 | { | |
5a7ad253 | 1770 | enum tree_code this_code = (e->flags & EDGE_TRUE_VALUE |
1771 | ? code : inverted_code); | |
1772 | /* invert_tree_comparison will return ERROR_MARK on FP | |
1773 | comparsions that are not EQ/NE instead of returning proper | |
1774 | unordered one. Be sure it is not confused with NON_CONSTANT. */ | |
1775 | if (this_code != ERROR_MARK) | |
1776 | { | |
83caf9ff | 1777 | struct predicate p = add_condition |
1778 | (summary, index, &aggpos, this_code, | |
1779 | unshare_expr_without_location (gimple_cond_rhs (last))); | |
2a5261f7 | 1780 | e->aux = edge_predicate_pool.allocate (); |
5a7ad253 | 1781 | *(struct predicate *) e->aux = p; |
1782 | } | |
905aa3bd | 1783 | } |
1784 | } | |
1785 | ||
94646c9c | 1786 | if (TREE_CODE (op) != SSA_NAME) |
1787 | return; | |
905aa3bd | 1788 | /* Special case |
1789 | if (builtin_constant_p (op)) | |
e876d531 | 1790 | constant_code |
905aa3bd | 1791 | else |
e876d531 | 1792 | nonconstant_code. |
905aa3bd | 1793 | Here we can predicate nonconstant_code. We can't |
1794 | really handle constant_code since we have no predicate | |
1795 | for this and also the constant code is not known to be | |
1796 | optimized away when inliner doen't see operand is constant. | |
1797 | Other optimizers might think otherwise. */ | |
a4f60e55 | 1798 | if (gimple_cond_code (last) != NE_EXPR |
1799 | || !integer_zerop (gimple_cond_rhs (last))) | |
1800 | return; | |
905aa3bd | 1801 | set_stmt = SSA_NAME_DEF_STMT (op); |
1802 | if (!gimple_call_builtin_p (set_stmt, BUILT_IN_CONSTANT_P) | |
1803 | || gimple_call_num_args (set_stmt) != 1) | |
1804 | return; | |
1805 | op2 = gimple_call_arg (set_stmt, 0); | |
1a673ff0 | 1806 | if (!unmodified_parm_or_parm_agg_item (fbi, set_stmt, op2, &index, &aggpos)) |
905aa3bd | 1807 | return; |
e876d531 | 1808 | FOR_EACH_EDGE (e, ei, bb->succs) if (e->flags & EDGE_FALSE_VALUE) |
1809 | { | |
1810 | struct predicate p = add_condition (summary, index, &aggpos, | |
1811 | IS_NOT_CONSTANT, NULL_TREE); | |
2a5261f7 | 1812 | e->aux = edge_predicate_pool.allocate (); |
e876d531 | 1813 | *(struct predicate *) e->aux = p; |
1814 | } | |
905aa3bd | 1815 | } |
1816 | ||
1817 | ||
1818 | /* If BB ends by a switch we can turn into predicates, attach corresponding | |
1819 | predicates to the CFG edges. */ | |
1820 | ||
1821 | static void | |
9ea91b78 | 1822 | set_switch_stmt_execution_predicate (struct ipa_func_body_info *fbi, |
e876d531 | 1823 | struct inline_summary *summary, |
1824 | basic_block bb) | |
905aa3bd | 1825 | { |
42acab1c | 1826 | gimple *lastg; |
905aa3bd | 1827 | tree op; |
1828 | int index; | |
a4f60e55 | 1829 | struct agg_position_info aggpos; |
905aa3bd | 1830 | edge e; |
1831 | edge_iterator ei; | |
1832 | size_t n; | |
1833 | size_t case_idx; | |
1834 | ||
1a91d914 | 1835 | lastg = last_stmt (bb); |
1836 | if (!lastg || gimple_code (lastg) != GIMPLE_SWITCH) | |
905aa3bd | 1837 | return; |
1a91d914 | 1838 | gswitch *last = as_a <gswitch *> (lastg); |
905aa3bd | 1839 | op = gimple_switch_index (last); |
1a673ff0 | 1840 | if (!unmodified_parm_or_parm_agg_item (fbi, last, op, &index, &aggpos)) |
905aa3bd | 1841 | return; |
a41f2a28 | 1842 | |
905aa3bd | 1843 | FOR_EACH_EDGE (e, ei, bb->succs) |
1844 | { | |
2a5261f7 | 1845 | e->aux = edge_predicate_pool.allocate (); |
e876d531 | 1846 | *(struct predicate *) e->aux = false_predicate (); |
905aa3bd | 1847 | } |
e876d531 | 1848 | n = gimple_switch_num_labels (last); |
905aa3bd | 1849 | for (case_idx = 0; case_idx < n; ++case_idx) |
1850 | { | |
1851 | tree cl = gimple_switch_label (last, case_idx); | |
1852 | tree min, max; | |
1853 | struct predicate p; | |
a41f2a28 | 1854 | |
905aa3bd | 1855 | e = find_edge (bb, label_to_block (CASE_LABEL (cl))); |
1856 | min = CASE_LOW (cl); | |
1857 | max = CASE_HIGH (cl); | |
1858 | ||
1859 | /* For default we might want to construct predicate that none | |
e876d531 | 1860 | of cases is met, but it is bit hard to do not having negations |
1861 | of conditionals handy. */ | |
905aa3bd | 1862 | if (!min && !max) |
1863 | p = true_predicate (); | |
1864 | else if (!max) | |
83caf9ff | 1865 | p = add_condition (summary, index, &aggpos, EQ_EXPR, |
1866 | unshare_expr_without_location (min)); | |
905aa3bd | 1867 | else |
1868 | { | |
1869 | struct predicate p1, p2; | |
83caf9ff | 1870 | p1 = add_condition (summary, index, &aggpos, GE_EXPR, |
1871 | unshare_expr_without_location (min)); | |
1872 | p2 = add_condition (summary, index, &aggpos, LE_EXPR, | |
1873 | unshare_expr_without_location (max)); | |
94646c9c | 1874 | p = and_predicates (summary->conds, &p1, &p2); |
905aa3bd | 1875 | } |
e876d531 | 1876 | *(struct predicate *) e->aux |
1877 | = or_predicates (summary->conds, &p, (struct predicate *) e->aux); | |
905aa3bd | 1878 | } |
1879 | } | |
1880 | ||
1881 | ||
1882 | /* For each BB in NODE attach to its AUX pointer predicate under | |
1883 | which it is executable. */ | |
1884 | ||
1885 | static void | |
9ea91b78 | 1886 | compute_bb_predicates (struct ipa_func_body_info *fbi, |
1a673ff0 | 1887 | struct cgraph_node *node, |
905aa3bd | 1888 | struct inline_summary *summary) |
1889 | { | |
02774f2d | 1890 | struct function *my_function = DECL_STRUCT_FUNCTION (node->decl); |
905aa3bd | 1891 | bool done = false; |
1892 | basic_block bb; | |
1893 | ||
1894 | FOR_EACH_BB_FN (bb, my_function) | |
1895 | { | |
1a673ff0 | 1896 | set_cond_stmt_execution_predicate (fbi, summary, bb); |
1897 | set_switch_stmt_execution_predicate (fbi, summary, bb); | |
905aa3bd | 1898 | } |
1899 | ||
1900 | /* Entry block is always executable. */ | |
34154e27 | 1901 | ENTRY_BLOCK_PTR_FOR_FN (my_function)->aux |
2a5261f7 | 1902 | = edge_predicate_pool.allocate (); |
34154e27 | 1903 | *(struct predicate *) ENTRY_BLOCK_PTR_FOR_FN (my_function)->aux |
905aa3bd | 1904 | = true_predicate (); |
1905 | ||
1906 | /* A simple dataflow propagation of predicates forward in the CFG. | |
1907 | TODO: work in reverse postorder. */ | |
1908 | while (!done) | |
1909 | { | |
1910 | done = true; | |
1911 | FOR_EACH_BB_FN (bb, my_function) | |
1912 | { | |
e876d531 | 1913 | struct predicate p = false_predicate (); |
1914 | edge e; | |
1915 | edge_iterator ei; | |
905aa3bd | 1916 | FOR_EACH_EDGE (e, ei, bb->preds) |
1917 | { | |
1918 | if (e->src->aux) | |
1919 | { | |
fb3c587e | 1920 | struct predicate this_bb_predicate |
e876d531 | 1921 | = *(struct predicate *) e->src->aux; |
905aa3bd | 1922 | if (e->aux) |
fb3c587e | 1923 | this_bb_predicate |
e876d531 | 1924 | = and_predicates (summary->conds, &this_bb_predicate, |
1925 | (struct predicate *) e->aux); | |
94646c9c | 1926 | p = or_predicates (summary->conds, &p, &this_bb_predicate); |
905aa3bd | 1927 | if (true_predicate_p (&p)) |
1928 | break; | |
1929 | } | |
1930 | } | |
1931 | if (false_predicate_p (&p)) | |
1932 | gcc_assert (!bb->aux); | |
1933 | else | |
1934 | { | |
1935 | if (!bb->aux) | |
1936 | { | |
1937 | done = false; | |
2a5261f7 | 1938 | bb->aux = edge_predicate_pool.allocate (); |
e876d531 | 1939 | *((struct predicate *) bb->aux) = p; |
905aa3bd | 1940 | } |
e876d531 | 1941 | else if (!predicates_equal_p (&p, (struct predicate *) bb->aux)) |
905aa3bd | 1942 | { |
74d6d072 | 1943 | /* This OR operation is needed to ensure monotonous data flow |
1944 | in the case we hit the limit on number of clauses and the | |
1945 | and/or operations above give approximate answers. */ | |
1946 | p = or_predicates (summary->conds, &p, (struct predicate *)bb->aux); | |
1947 | if (!predicates_equal_p (&p, (struct predicate *) bb->aux)) | |
1948 | { | |
1949 | done = false; | |
1950 | *((struct predicate *) bb->aux) = p; | |
1951 | } | |
905aa3bd | 1952 | } |
1953 | } | |
1954 | } | |
1955 | } | |
a41f2a28 | 1956 | } |
1957 | ||
0b50fa0e | 1958 | |
1959 | /* We keep info about constantness of SSA names. */ | |
1960 | ||
1961 | typedef struct predicate predicate_t; | |
7c07aa3d | 1962 | /* Return predicate specifying when the STMT might have result that is not |
1963 | a compile time constant. */ | |
1964 | ||
1965 | static struct predicate | |
1966 | will_be_nonconstant_expr_predicate (struct ipa_node_params *info, | |
e876d531 | 1967 | struct inline_summary *summary, |
1968 | tree expr, | |
1969 | vec<predicate_t> nonconstant_names) | |
7c07aa3d | 1970 | { |
1971 | tree parm; | |
1972 | int index; | |
1973 | ||
1974 | while (UNARY_CLASS_P (expr)) | |
1975 | expr = TREE_OPERAND (expr, 0); | |
1976 | ||
1977 | parm = unmodified_parm (NULL, expr); | |
e876d531 | 1978 | if (parm && (index = ipa_get_param_decl_index (info, parm)) >= 0) |
7c07aa3d | 1979 | return add_condition (summary, index, NULL, CHANGED, NULL_TREE); |
1980 | if (is_gimple_min_invariant (expr)) | |
1981 | return false_predicate (); | |
1982 | if (TREE_CODE (expr) == SSA_NAME) | |
f1f41a6c | 1983 | return nonconstant_names[SSA_NAME_VERSION (expr)]; |
e876d531 | 1984 | if (BINARY_CLASS_P (expr) || COMPARISON_CLASS_P (expr)) |
3716ee8f | 1985 | { |
1986 | struct predicate p1 = will_be_nonconstant_expr_predicate | |
e876d531 | 1987 | (info, summary, TREE_OPERAND (expr, 0), |
1988 | nonconstant_names); | |
3716ee8f | 1989 | struct predicate p2; |
1990 | if (true_predicate_p (&p1)) | |
1991 | return p1; | |
1992 | p2 = will_be_nonconstant_expr_predicate (info, summary, | |
1993 | TREE_OPERAND (expr, 1), | |
1994 | nonconstant_names); | |
1995 | return or_predicates (summary->conds, &p1, &p2); | |
1996 | } | |
1997 | else if (TREE_CODE (expr) == COND_EXPR) | |
7c07aa3d | 1998 | { |
3716ee8f | 1999 | struct predicate p1 = will_be_nonconstant_expr_predicate |
e876d531 | 2000 | (info, summary, TREE_OPERAND (expr, 0), |
2001 | nonconstant_names); | |
7c07aa3d | 2002 | struct predicate p2; |
2003 | if (true_predicate_p (&p1)) | |
2004 | return p1; | |
3716ee8f | 2005 | p2 = will_be_nonconstant_expr_predicate (info, summary, |
2006 | TREE_OPERAND (expr, 1), | |
2007 | nonconstant_names); | |
2008 | if (true_predicate_p (&p2)) | |
2009 | return p2; | |
2010 | p1 = or_predicates (summary->conds, &p1, &p2); | |
2011 | p2 = will_be_nonconstant_expr_predicate (info, summary, | |
2012 | TREE_OPERAND (expr, 2), | |
2013 | nonconstant_names); | |
7c07aa3d | 2014 | return or_predicates (summary->conds, &p1, &p2); |
2015 | } | |
2016 | else | |
2017 | { | |
2018 | debug_tree (expr); | |
2019 | gcc_unreachable (); | |
2020 | } | |
2021 | return false_predicate (); | |
2022 | } | |
0b50fa0e | 2023 | |
2024 | ||
fb3c587e | 2025 | /* Return predicate specifying when the STMT might have result that is not |
2026 | a compile time constant. */ | |
0b50fa0e | 2027 | |
a41f2a28 | 2028 | static struct predicate |
9ea91b78 | 2029 | will_be_nonconstant_predicate (struct ipa_func_body_info *fbi, |
a41f2a28 | 2030 | struct inline_summary *summary, |
42acab1c | 2031 | gimple *stmt, |
f1f41a6c | 2032 | vec<predicate_t> nonconstant_names) |
a41f2a28 | 2033 | { |
2034 | struct predicate p = true_predicate (); | |
2035 | ssa_op_iter iter; | |
2036 | tree use; | |
2037 | struct predicate op_non_const; | |
8e22665e | 2038 | bool is_load; |
a4f60e55 | 2039 | int base_index; |
2040 | struct agg_position_info aggpos; | |
a41f2a28 | 2041 | |
2042 | /* What statments might be optimized away | |
1d4ee63f | 2043 | when their arguments are constant. */ |
a41f2a28 | 2044 | if (gimple_code (stmt) != GIMPLE_ASSIGN |
2045 | && gimple_code (stmt) != GIMPLE_COND | |
1d4ee63f | 2046 | && gimple_code (stmt) != GIMPLE_SWITCH |
2047 | && (gimple_code (stmt) != GIMPLE_CALL | |
2048 | || !(gimple_call_flags (stmt) & ECF_CONST))) | |
a41f2a28 | 2049 | return p; |
2050 | ||
8e22665e | 2051 | /* Stores will stay anyway. */ |
3172b7bf | 2052 | if (gimple_store_p (stmt)) |
a41f2a28 | 2053 | return p; |
2054 | ||
3172b7bf | 2055 | is_load = gimple_assign_load_p (stmt); |
2056 | ||
8e22665e | 2057 | /* Loads can be optimized when the value is known. */ |
2058 | if (is_load) | |
2059 | { | |
a4f60e55 | 2060 | tree op; |
8e22665e | 2061 | gcc_assert (gimple_assign_single_p (stmt)); |
a4f60e55 | 2062 | op = gimple_assign_rhs1 (stmt); |
1a673ff0 | 2063 | if (!unmodified_parm_or_parm_agg_item (fbi, stmt, op, &base_index, |
a4f60e55 | 2064 | &aggpos)) |
8e22665e | 2065 | return p; |
2066 | } | |
a4f60e55 | 2067 | else |
2068 | base_index = -1; | |
8e22665e | 2069 | |
a41f2a28 | 2070 | /* See if we understand all operands before we start |
2071 | adding conditionals. */ | |
2072 | FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE) | |
2073 | { | |
94646c9c | 2074 | tree parm = unmodified_parm (stmt, use); |
0b50fa0e | 2075 | /* For arguments we can build a condition. */ |
1a673ff0 | 2076 | if (parm && ipa_get_param_decl_index (fbi->info, parm) >= 0) |
0b50fa0e | 2077 | continue; |
94646c9c | 2078 | if (TREE_CODE (use) != SSA_NAME) |
2079 | return p; | |
0b50fa0e | 2080 | /* If we know when operand is constant, |
2081 | we still can say something useful. */ | |
f1f41a6c | 2082 | if (!true_predicate_p (&nonconstant_names[SSA_NAME_VERSION (use)])) |
0b50fa0e | 2083 | continue; |
2084 | return p; | |
a41f2a28 | 2085 | } |
a4f60e55 | 2086 | |
8e22665e | 2087 | if (is_load) |
e876d531 | 2088 | op_non_const = |
2089 | add_condition (summary, base_index, &aggpos, CHANGED, NULL); | |
a4f60e55 | 2090 | else |
2091 | op_non_const = false_predicate (); | |
a41f2a28 | 2092 | FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE) |
2093 | { | |
94646c9c | 2094 | tree parm = unmodified_parm (stmt, use); |
a4f60e55 | 2095 | int index; |
2096 | ||
1a673ff0 | 2097 | if (parm && (index = ipa_get_param_decl_index (fbi->info, parm)) >= 0) |
a4f60e55 | 2098 | { |
2099 | if (index != base_index) | |
2100 | p = add_condition (summary, index, NULL, CHANGED, NULL_TREE); | |
2101 | else | |
2102 | continue; | |
2103 | } | |
0b50fa0e | 2104 | else |
f1f41a6c | 2105 | p = nonconstant_names[SSA_NAME_VERSION (use)]; |
94646c9c | 2106 | op_non_const = or_predicates (summary->conds, &p, &op_non_const); |
a41f2a28 | 2107 | } |
1d4ee63f | 2108 | if ((gimple_code (stmt) == GIMPLE_ASSIGN || gimple_code (stmt) == GIMPLE_CALL) |
2109 | && gimple_op (stmt, 0) | |
2110 | && TREE_CODE (gimple_op (stmt, 0)) == SSA_NAME) | |
2111 | nonconstant_names[SSA_NAME_VERSION (gimple_op (stmt, 0))] | |
e876d531 | 2112 | = op_non_const; |
a41f2a28 | 2113 | return op_non_const; |
2114 | } | |
2115 | ||
eb4ae064 | 2116 | struct record_modified_bb_info |
2117 | { | |
2118 | bitmap bb_set; | |
42acab1c | 2119 | gimple *stmt; |
eb4ae064 | 2120 | }; |
2121 | ||
2122 | /* Callback of walk_aliased_vdefs. Records basic blocks where the value may be | |
2123 | set except for info->stmt. */ | |
2124 | ||
2125 | static bool | |
e876d531 | 2126 | record_modified (ao_ref *ao ATTRIBUTE_UNUSED, tree vdef, void *data) |
eb4ae064 | 2127 | { |
e876d531 | 2128 | struct record_modified_bb_info *info = |
2129 | (struct record_modified_bb_info *) data; | |
eb4ae064 | 2130 | if (SSA_NAME_DEF_STMT (vdef) == info->stmt) |
2131 | return false; | |
2132 | bitmap_set_bit (info->bb_set, | |
2133 | SSA_NAME_IS_DEFAULT_DEF (vdef) | |
34154e27 | 2134 | ? ENTRY_BLOCK_PTR_FOR_FN (cfun)->index |
e876d531 | 2135 | : gimple_bb (SSA_NAME_DEF_STMT (vdef))->index); |
eb4ae064 | 2136 | return false; |
2137 | } | |
2138 | ||
2139 | /* Return probability (based on REG_BR_PROB_BASE) that I-th parameter of STMT | |
2140 | will change since last invocation of STMT. | |
2141 | ||
2142 | Value 0 is reserved for compile time invariants. | |
2143 | For common parameters it is REG_BR_PROB_BASE. For loop invariants it | |
2144 | ought to be REG_BR_PROB_BASE / estimated_iters. */ | |
2145 | ||
2146 | static int | |
42acab1c | 2147 | param_change_prob (gimple *stmt, int i) |
eb4ae064 | 2148 | { |
2149 | tree op = gimple_call_arg (stmt, i); | |
2150 | basic_block bb = gimple_bb (stmt); | |
2151 | tree base; | |
2152 | ||
e876d531 | 2153 | /* Global invariants neve change. */ |
eb4ae064 | 2154 | if (is_gimple_min_invariant (op)) |
2155 | return 0; | |
2156 | /* We would have to do non-trivial analysis to really work out what | |
2157 | is the probability of value to change (i.e. when init statement | |
2158 | is in a sibling loop of the call). | |
2159 | ||
2160 | We do an conservative estimate: when call is executed N times more often | |
2161 | than the statement defining value, we take the frequency 1/N. */ | |
2162 | if (TREE_CODE (op) == SSA_NAME) | |
2163 | { | |
2164 | int init_freq; | |
2165 | ||
2166 | if (!bb->frequency) | |
2167 | return REG_BR_PROB_BASE; | |
2168 | ||
2169 | if (SSA_NAME_IS_DEFAULT_DEF (op)) | |
34154e27 | 2170 | init_freq = ENTRY_BLOCK_PTR_FOR_FN (cfun)->frequency; |
eb4ae064 | 2171 | else |
2172 | init_freq = gimple_bb (SSA_NAME_DEF_STMT (op))->frequency; | |
2173 | ||
2174 | if (!init_freq) | |
2175 | init_freq = 1; | |
2176 | if (init_freq < bb->frequency) | |
f9d4b7f4 | 2177 | return MAX (GCOV_COMPUTE_SCALE (init_freq, bb->frequency), 1); |
eb4ae064 | 2178 | else |
e876d531 | 2179 | return REG_BR_PROB_BASE; |
eb4ae064 | 2180 | } |
2181 | ||
2182 | base = get_base_address (op); | |
2183 | if (base) | |
2184 | { | |
2185 | ao_ref refd; | |
2186 | int max; | |
2187 | struct record_modified_bb_info info; | |
2188 | bitmap_iterator bi; | |
2189 | unsigned index; | |
df8d3e89 | 2190 | tree init = ctor_for_folding (base); |
eb4ae064 | 2191 | |
df8d3e89 | 2192 | if (init != error_mark_node) |
eb4ae064 | 2193 | return 0; |
2194 | if (!bb->frequency) | |
2195 | return REG_BR_PROB_BASE; | |
2196 | ao_ref_init (&refd, op); | |
2197 | info.stmt = stmt; | |
2198 | info.bb_set = BITMAP_ALLOC (NULL); | |
2199 | walk_aliased_vdefs (&refd, gimple_vuse (stmt), record_modified, &info, | |
2200 | NULL); | |
2201 | if (bitmap_bit_p (info.bb_set, bb->index)) | |
2202 | { | |
e876d531 | 2203 | BITMAP_FREE (info.bb_set); |
eb4ae064 | 2204 | return REG_BR_PROB_BASE; |
2205 | } | |
2206 | ||
2207 | /* Assume that every memory is initialized at entry. | |
e876d531 | 2208 | TODO: Can we easilly determine if value is always defined |
2209 | and thus we may skip entry block? */ | |
34154e27 | 2210 | if (ENTRY_BLOCK_PTR_FOR_FN (cfun)->frequency) |
2211 | max = ENTRY_BLOCK_PTR_FOR_FN (cfun)->frequency; | |
eb4ae064 | 2212 | else |
2213 | max = 1; | |
2214 | ||
2215 | EXECUTE_IF_SET_IN_BITMAP (info.bb_set, 0, index, bi) | |
f5a6b05f | 2216 | max = MIN (max, BASIC_BLOCK_FOR_FN (cfun, index)->frequency); |
e876d531 | 2217 | |
eb4ae064 | 2218 | BITMAP_FREE (info.bb_set); |
2219 | if (max < bb->frequency) | |
f9d4b7f4 | 2220 | return MAX (GCOV_COMPUTE_SCALE (max, bb->frequency), 1); |
eb4ae064 | 2221 | else |
e876d531 | 2222 | return REG_BR_PROB_BASE; |
eb4ae064 | 2223 | } |
2224 | return REG_BR_PROB_BASE; | |
2225 | } | |
2226 | ||
32691791 | 2227 | /* Find whether a basic block BB is the final block of a (half) diamond CFG |
2228 | sub-graph and if the predicate the condition depends on is known. If so, | |
2229 | return true and store the pointer the predicate in *P. */ | |
2230 | ||
2231 | static bool | |
2232 | phi_result_unknown_predicate (struct ipa_node_params *info, | |
b4bae7a0 | 2233 | inline_summary *summary, basic_block bb, |
32691791 | 2234 | struct predicate *p, |
f1f41a6c | 2235 | vec<predicate_t> nonconstant_names) |
32691791 | 2236 | { |
2237 | edge e; | |
2238 | edge_iterator ei; | |
2239 | basic_block first_bb = NULL; | |
42acab1c | 2240 | gimple *stmt; |
32691791 | 2241 | |
2242 | if (single_pred_p (bb)) | |
2243 | { | |
2244 | *p = false_predicate (); | |
2245 | return true; | |
2246 | } | |
2247 | ||
2248 | FOR_EACH_EDGE (e, ei, bb->preds) | |
2249 | { | |
2250 | if (single_succ_p (e->src)) | |
2251 | { | |
2252 | if (!single_pred_p (e->src)) | |
2253 | return false; | |
2254 | if (!first_bb) | |
2255 | first_bb = single_pred (e->src); | |
2256 | else if (single_pred (e->src) != first_bb) | |
2257 | return false; | |
2258 | } | |
2259 | else | |
2260 | { | |
2261 | if (!first_bb) | |
2262 | first_bb = e->src; | |
2263 | else if (e->src != first_bb) | |
2264 | return false; | |
2265 | } | |
2266 | } | |
2267 | ||
2268 | if (!first_bb) | |
2269 | return false; | |
2270 | ||
2271 | stmt = last_stmt (first_bb); | |
2272 | if (!stmt | |
2273 | || gimple_code (stmt) != GIMPLE_COND | |
2274 | || !is_gimple_ip_invariant (gimple_cond_rhs (stmt))) | |
2275 | return false; | |
2276 | ||
2277 | *p = will_be_nonconstant_expr_predicate (info, summary, | |
2278 | gimple_cond_lhs (stmt), | |
2279 | nonconstant_names); | |
2280 | if (true_predicate_p (p)) | |
2281 | return false; | |
2282 | else | |
2283 | return true; | |
2284 | } | |
2285 | ||
2286 | /* Given a PHI statement in a function described by inline properties SUMMARY | |
2287 | and *P being the predicate describing whether the selected PHI argument is | |
2288 | known, store a predicate for the result of the PHI statement into | |
2289 | NONCONSTANT_NAMES, if possible. */ | |
2290 | ||
2291 | static void | |
1a91d914 | 2292 | predicate_for_phi_result (struct inline_summary *summary, gphi *phi, |
32691791 | 2293 | struct predicate *p, |
f1f41a6c | 2294 | vec<predicate_t> nonconstant_names) |
32691791 | 2295 | { |
2296 | unsigned i; | |
2297 | ||
2298 | for (i = 0; i < gimple_phi_num_args (phi); i++) | |
2299 | { | |
2300 | tree arg = gimple_phi_arg (phi, i)->def; | |
2301 | if (!is_gimple_min_invariant (arg)) | |
2302 | { | |
2303 | gcc_assert (TREE_CODE (arg) == SSA_NAME); | |
2304 | *p = or_predicates (summary->conds, p, | |
f1f41a6c | 2305 | &nonconstant_names[SSA_NAME_VERSION (arg)]); |
32691791 | 2306 | if (true_predicate_p (p)) |
2307 | return; | |
2308 | } | |
2309 | } | |
2310 | ||
2311 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2312 | { | |
2313 | fprintf (dump_file, "\t\tphi predicate: "); | |
2314 | dump_predicate (dump_file, summary->conds, p); | |
2315 | } | |
f1f41a6c | 2316 | nonconstant_names[SSA_NAME_VERSION (gimple_phi_result (phi))] = *p; |
32691791 | 2317 | } |
a41f2a28 | 2318 | |
be343a9c | 2319 | /* Return predicate specifying when array index in access OP becomes non-constant. */ |
2320 | ||
2321 | static struct predicate | |
b4bae7a0 | 2322 | array_index_predicate (inline_summary *info, |
e876d531 | 2323 | vec< predicate_t> nonconstant_names, tree op) |
be343a9c | 2324 | { |
2325 | struct predicate p = false_predicate (); | |
2326 | while (handled_component_p (op)) | |
2327 | { | |
e876d531 | 2328 | if (TREE_CODE (op) == ARRAY_REF || TREE_CODE (op) == ARRAY_RANGE_REF) |
2329 | { | |
be343a9c | 2330 | if (TREE_CODE (TREE_OPERAND (op, 1)) == SSA_NAME) |
e876d531 | 2331 | p = or_predicates (info->conds, &p, |
2332 | &nonconstant_names[SSA_NAME_VERSION | |
2333 | (TREE_OPERAND (op, 1))]); | |
2334 | } | |
be343a9c | 2335 | op = TREE_OPERAND (op, 0); |
2336 | } | |
2337 | return p; | |
2338 | } | |
2339 | ||
caa0f772 | 2340 | /* For a typical usage of __builtin_expect (a<b, 1), we |
2341 | may introduce an extra relation stmt: | |
2342 | With the builtin, we have | |
2343 | t1 = a <= b; | |
2344 | t2 = (long int) t1; | |
2345 | t3 = __builtin_expect (t2, 1); | |
2346 | if (t3 != 0) | |
2347 | goto ... | |
2348 | Without the builtin, we have | |
2349 | if (a<=b) | |
2350 | goto... | |
2351 | This affects the size/time estimation and may have | |
2352 | an impact on the earlier inlining. | |
2353 | Here find this pattern and fix it up later. */ | |
2354 | ||
42acab1c | 2355 | static gimple * |
caa0f772 | 2356 | find_foldable_builtin_expect (basic_block bb) |
2357 | { | |
2358 | gimple_stmt_iterator bsi; | |
2359 | ||
2360 | for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
2361 | { | |
42acab1c | 2362 | gimple *stmt = gsi_stmt (bsi); |
c83059be | 2363 | if (gimple_call_builtin_p (stmt, BUILT_IN_EXPECT) |
2364 | || (is_gimple_call (stmt) | |
2365 | && gimple_call_internal_p (stmt) | |
2366 | && gimple_call_internal_fn (stmt) == IFN_BUILTIN_EXPECT)) | |
caa0f772 | 2367 | { |
2368 | tree var = gimple_call_lhs (stmt); | |
2369 | tree arg = gimple_call_arg (stmt, 0); | |
2370 | use_operand_p use_p; | |
42acab1c | 2371 | gimple *use_stmt; |
caa0f772 | 2372 | bool match = false; |
2373 | bool done = false; | |
2374 | ||
2375 | if (!var || !arg) | |
2376 | continue; | |
2377 | gcc_assert (TREE_CODE (var) == SSA_NAME); | |
2378 | ||
2379 | while (TREE_CODE (arg) == SSA_NAME) | |
2380 | { | |
42acab1c | 2381 | gimple *stmt_tmp = SSA_NAME_DEF_STMT (arg); |
caa0f772 | 2382 | if (!is_gimple_assign (stmt_tmp)) |
2383 | break; | |
2384 | switch (gimple_assign_rhs_code (stmt_tmp)) | |
2385 | { | |
2386 | case LT_EXPR: | |
2387 | case LE_EXPR: | |
2388 | case GT_EXPR: | |
2389 | case GE_EXPR: | |
2390 | case EQ_EXPR: | |
2391 | case NE_EXPR: | |
2392 | match = true; | |
2393 | done = true; | |
2394 | break; | |
d09ef31a | 2395 | CASE_CONVERT: |
caa0f772 | 2396 | break; |
2397 | default: | |
2398 | done = true; | |
2399 | break; | |
2400 | } | |
2401 | if (done) | |
2402 | break; | |
2403 | arg = gimple_assign_rhs1 (stmt_tmp); | |
2404 | } | |
2405 | ||
2406 | if (match && single_imm_use (var, &use_p, &use_stmt) | |
2407 | && gimple_code (use_stmt) == GIMPLE_COND) | |
2408 | return use_stmt; | |
2409 | } | |
2410 | } | |
2411 | return NULL; | |
2412 | } | |
2413 | ||
4c54368f | 2414 | /* Return true when the basic blocks contains only clobbers followed by RESX. |
2415 | Such BBs are kept around to make removal of dead stores possible with | |
2416 | presence of EH and will be optimized out by optimize_clobbers later in the | |
2417 | game. | |
2418 | ||
2419 | NEED_EH is used to recurse in case the clobber has non-EH predecestors | |
2420 | that can be clobber only, too.. When it is false, the RESX is not necessary | |
2421 | on the end of basic block. */ | |
2422 | ||
2423 | static bool | |
2424 | clobber_only_eh_bb_p (basic_block bb, bool need_eh = true) | |
2425 | { | |
2426 | gimple_stmt_iterator gsi = gsi_last_bb (bb); | |
2427 | edge_iterator ei; | |
2428 | edge e; | |
2429 | ||
2430 | if (need_eh) | |
2431 | { | |
2432 | if (gsi_end_p (gsi)) | |
2433 | return false; | |
2434 | if (gimple_code (gsi_stmt (gsi)) != GIMPLE_RESX) | |
2435 | return false; | |
2436 | gsi_prev (&gsi); | |
2437 | } | |
2438 | else if (!single_succ_p (bb)) | |
2439 | return false; | |
2440 | ||
2441 | for (; !gsi_end_p (gsi); gsi_prev (&gsi)) | |
2442 | { | |
42acab1c | 2443 | gimple *stmt = gsi_stmt (gsi); |
4c54368f | 2444 | if (is_gimple_debug (stmt)) |
2445 | continue; | |
2446 | if (gimple_clobber_p (stmt)) | |
2447 | continue; | |
2448 | if (gimple_code (stmt) == GIMPLE_LABEL) | |
2449 | break; | |
2450 | return false; | |
2451 | } | |
2452 | ||
2453 | /* See if all predecestors are either throws or clobber only BBs. */ | |
2454 | FOR_EACH_EDGE (e, ei, bb->preds) | |
2455 | if (!(e->flags & EDGE_EH) | |
2456 | && !clobber_only_eh_bb_p (e->src, false)) | |
2457 | return false; | |
2458 | ||
2459 | return true; | |
2460 | } | |
2461 | ||
a41f2a28 | 2462 | /* Compute function body size parameters for NODE. |
2463 | When EARLY is true, we compute only simple summaries without | |
2464 | non-trivial predicates to drive the early inliner. */ | |
99c67f24 | 2465 | |
2466 | static void | |
a41f2a28 | 2467 | estimate_function_body_sizes (struct cgraph_node *node, bool early) |
99c67f24 | 2468 | { |
2469 | gcov_type time = 0; | |
99c67f24 | 2470 | /* Estimate static overhead for function prologue/epilogue and alignment. */ |
2471 | int size = 2; | |
2472 | /* Benefits are scaled by probability of elimination that is in range | |
2473 | <0,2>. */ | |
99c67f24 | 2474 | basic_block bb; |
02774f2d | 2475 | struct function *my_function = DECL_STRUCT_FUNCTION (node->decl); |
99c67f24 | 2476 | int freq; |
b4bae7a0 | 2477 | struct inline_summary *info = inline_summaries->get (node); |
a41f2a28 | 2478 | struct predicate bb_predicate; |
9ea91b78 | 2479 | struct ipa_func_body_info fbi; |
1e094109 | 2480 | vec<predicate_t> nonconstant_names = vNULL; |
563fae60 | 2481 | int nblocks, n; |
2482 | int *order; | |
be343a9c | 2483 | predicate array_index = true_predicate (); |
42acab1c | 2484 | gimple *fix_builtin_expect_stmt; |
a41f2a28 | 2485 | |
1a673ff0 | 2486 | gcc_assert (my_function && my_function->cfg); |
2487 | gcc_assert (cfun == my_function); | |
2488 | ||
2489 | memset(&fbi, 0, sizeof(fbi)); | |
f1f41a6c | 2490 | info->conds = NULL; |
2491 | info->entry = NULL; | |
a41f2a28 | 2492 | |
f4e523eb | 2493 | /* When optimizing and analyzing for IPA inliner, initialize loop optimizer |
2494 | so we can produce proper inline hints. | |
2495 | ||
2496 | When optimizing and analyzing for early inliner, initialize node params | |
2497 | so we can produce correct BB predicates. */ | |
2498 | ||
2499 | if (opt_for_fn (node->decl, optimize)) | |
4b209fe7 | 2500 | { |
2501 | calculate_dominance_info (CDI_DOMINATORS); | |
f4e523eb | 2502 | if (!early) |
2503 | loop_optimizer_init (LOOPS_NORMAL | LOOPS_HAVE_RECORDED_EXITS); | |
2504 | else | |
2505 | { | |
2506 | ipa_check_create_node_params (); | |
2507 | ipa_initialize_node_params (node); | |
2508 | } | |
f228b1b9 | 2509 | |
2cc80ac3 | 2510 | if (ipa_node_params_sum) |
f228b1b9 | 2511 | { |
1a673ff0 | 2512 | fbi.node = node; |
2513 | fbi.info = IPA_NODE_REF (node); | |
2514 | fbi.bb_infos = vNULL; | |
2515 | fbi.bb_infos.safe_grow_cleared (last_basic_block_for_fn (cfun)); | |
2516 | fbi.param_count = count_formal_params(node->decl); | |
e876d531 | 2517 | nonconstant_names.safe_grow_cleared |
2518 | (SSANAMES (my_function)->length ()); | |
f228b1b9 | 2519 | } |
4b209fe7 | 2520 | } |
99c67f24 | 2521 | |
2522 | if (dump_file) | |
a41f2a28 | 2523 | fprintf (dump_file, "\nAnalyzing function body size: %s\n", |
f1c8b4d7 | 2524 | node->name ()); |
99c67f24 | 2525 | |
a41f2a28 | 2526 | /* When we run into maximal number of entries, we assign everything to the |
2527 | constant truth case. Be sure to have it in list. */ | |
2528 | bb_predicate = true_predicate (); | |
2529 | account_size_time (info, 0, 0, &bb_predicate); | |
2530 | ||
2531 | bb_predicate = not_inlined_predicate (); | |
2532 | account_size_time (info, 2 * INLINE_SIZE_SCALE, 0, &bb_predicate); | |
2533 | ||
1a673ff0 | 2534 | if (fbi.info) |
2535 | compute_bb_predicates (&fbi, node, info); | |
a28770e1 | 2536 | order = XNEWVEC (int, n_basic_blocks_for_fn (cfun)); |
563fae60 | 2537 | nblocks = pre_and_rev_post_order_compute (NULL, order, false); |
2538 | for (n = 0; n < nblocks; n++) | |
99c67f24 | 2539 | { |
f5a6b05f | 2540 | bb = BASIC_BLOCK_FOR_FN (cfun, order[n]); |
02774f2d | 2541 | freq = compute_call_stmt_bb_frequency (node->decl, bb); |
4c54368f | 2542 | if (clobber_only_eh_bb_p (bb)) |
2543 | { | |
2544 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2545 | fprintf (dump_file, "\n Ignoring BB %i;" | |
2546 | " it will be optimized away by cleanup_clobbers\n", | |
2547 | bb->index); | |
2548 | continue; | |
2549 | } | |
a41f2a28 | 2550 | |
2551 | /* TODO: Obviously predicates can be propagated down across CFG. */ | |
1a673ff0 | 2552 | if (fbi.info) |
a41f2a28 | 2553 | { |
905aa3bd | 2554 | if (bb->aux) |
e876d531 | 2555 | bb_predicate = *(struct predicate *) bb->aux; |
905aa3bd | 2556 | else |
2557 | bb_predicate = false_predicate (); | |
a41f2a28 | 2558 | } |
2559 | else | |
2560 | bb_predicate = true_predicate (); | |
2561 | ||
2562 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2563 | { | |
2564 | fprintf (dump_file, "\n BB %i predicate:", bb->index); | |
2565 | dump_predicate (dump_file, info->conds, &bb_predicate); | |
2566 | } | |
32691791 | 2567 | |
1a673ff0 | 2568 | if (fbi.info && nonconstant_names.exists ()) |
32691791 | 2569 | { |
2570 | struct predicate phi_predicate; | |
2571 | bool first_phi = true; | |
2572 | ||
1a91d914 | 2573 | for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi); |
2574 | gsi_next (&bsi)) | |
32691791 | 2575 | { |
2576 | if (first_phi | |
1a673ff0 | 2577 | && !phi_result_unknown_predicate (fbi.info, info, bb, |
32691791 | 2578 | &phi_predicate, |
2579 | nonconstant_names)) | |
2580 | break; | |
2581 | first_phi = false; | |
2582 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2583 | { | |
2584 | fprintf (dump_file, " "); | |
2585 | print_gimple_stmt (dump_file, gsi_stmt (bsi), 0, 0); | |
2586 | } | |
1a91d914 | 2587 | predicate_for_phi_result (info, bsi.phi (), &phi_predicate, |
32691791 | 2588 | nonconstant_names); |
2589 | } | |
2590 | } | |
2591 | ||
caa0f772 | 2592 | fix_builtin_expect_stmt = find_foldable_builtin_expect (bb); |
2593 | ||
1a91d914 | 2594 | for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi); |
2595 | gsi_next (&bsi)) | |
99c67f24 | 2596 | { |
42acab1c | 2597 | gimple *stmt = gsi_stmt (bsi); |
99c67f24 | 2598 | int this_size = estimate_num_insns (stmt, &eni_size_weights); |
2599 | int this_time = estimate_num_insns (stmt, &eni_time_weights); | |
2600 | int prob; | |
905aa3bd | 2601 | struct predicate will_be_nonconstant; |
99c67f24 | 2602 | |
caa0f772 | 2603 | /* This relation stmt should be folded after we remove |
2604 | buildin_expect call. Adjust the cost here. */ | |
2605 | if (stmt == fix_builtin_expect_stmt) | |
2606 | { | |
2607 | this_size--; | |
2608 | this_time--; | |
2609 | } | |
2610 | ||
99c67f24 | 2611 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2612 | { | |
a41f2a28 | 2613 | fprintf (dump_file, " "); |
99c67f24 | 2614 | print_gimple_stmt (dump_file, stmt, 0, 0); |
a41f2a28 | 2615 | fprintf (dump_file, "\t\tfreq:%3.2f size:%3i time:%3i\n", |
e876d531 | 2616 | ((double) freq) / CGRAPH_FREQ_BASE, this_size, |
2617 | this_time); | |
99c67f24 | 2618 | } |
c7b2cc59 | 2619 | |
f1f41a6c | 2620 | if (gimple_assign_load_p (stmt) && nonconstant_names.exists ()) |
be343a9c | 2621 | { |
2622 | struct predicate this_array_index; | |
e876d531 | 2623 | this_array_index = |
2624 | array_index_predicate (info, nonconstant_names, | |
2625 | gimple_assign_rhs1 (stmt)); | |
be343a9c | 2626 | if (!false_predicate_p (&this_array_index)) |
e876d531 | 2627 | array_index = |
2628 | and_predicates (info->conds, &array_index, | |
2629 | &this_array_index); | |
be343a9c | 2630 | } |
f1f41a6c | 2631 | if (gimple_store_p (stmt) && nonconstant_names.exists ()) |
be343a9c | 2632 | { |
2633 | struct predicate this_array_index; | |
e876d531 | 2634 | this_array_index = |
2635 | array_index_predicate (info, nonconstant_names, | |
2636 | gimple_get_lhs (stmt)); | |
be343a9c | 2637 | if (!false_predicate_p (&this_array_index)) |
e876d531 | 2638 | array_index = |
2639 | and_predicates (info->conds, &array_index, | |
2640 | &this_array_index); | |
be343a9c | 2641 | } |
e876d531 | 2642 | |
be343a9c | 2643 | |
9978ce6d | 2644 | if (is_gimple_call (stmt) |
2645 | && !gimple_call_internal_p (stmt)) | |
c7b2cc59 | 2646 | { |
415d1b9a | 2647 | struct cgraph_edge *edge = node->get_edge (stmt); |
0835ad03 | 2648 | struct inline_edge_summary *es = inline_edge_summary (edge); |
2649 | ||
0b50fa0e | 2650 | /* Special case: results of BUILT_IN_CONSTANT_P will be always |
e876d531 | 2651 | resolved as constant. We however don't want to optimize |
2652 | out the cgraph edges. */ | |
f1f41a6c | 2653 | if (nonconstant_names.exists () |
0b50fa0e | 2654 | && gimple_call_builtin_p (stmt, BUILT_IN_CONSTANT_P) |
2655 | && gimple_call_lhs (stmt) | |
2656 | && TREE_CODE (gimple_call_lhs (stmt)) == SSA_NAME) | |
2657 | { | |
2658 | struct predicate false_p = false_predicate (); | |
f1f41a6c | 2659 | nonconstant_names[SSA_NAME_VERSION (gimple_call_lhs (stmt))] |
e876d531 | 2660 | = false_p; |
eb4ae064 | 2661 | } |
2cc80ac3 | 2662 | if (ipa_node_params_sum) |
eb4ae064 | 2663 | { |
e876d531 | 2664 | int count = gimple_call_num_args (stmt); |
eb4ae064 | 2665 | int i; |
2666 | ||
2667 | if (count) | |
f1f41a6c | 2668 | es->param.safe_grow_cleared (count); |
eb4ae064 | 2669 | for (i = 0; i < count; i++) |
2670 | { | |
2671 | int prob = param_change_prob (stmt, i); | |
2672 | gcc_assert (prob >= 0 && prob <= REG_BR_PROB_BASE); | |
f1f41a6c | 2673 | es->param[i].change_prob = prob; |
eb4ae064 | 2674 | } |
0b50fa0e | 2675 | } |
2676 | ||
0835ad03 | 2677 | es->call_stmt_size = this_size; |
2678 | es->call_stmt_time = this_time; | |
6b42039a | 2679 | es->loop_depth = bb_loop_depth (bb); |
6a18c0be | 2680 | edge_set_predicate (edge, &bb_predicate); |
c7b2cc59 | 2681 | } |
2682 | ||
905aa3bd | 2683 | /* TODO: When conditional jump or swithc is known to be constant, but |
e876d531 | 2684 | we did not translate it into the predicates, we really can account |
905aa3bd | 2685 | just maximum of the possible paths. */ |
1a673ff0 | 2686 | if (fbi.info) |
905aa3bd | 2687 | will_be_nonconstant |
1a673ff0 | 2688 | = will_be_nonconstant_predicate (&fbi, info, |
e876d531 | 2689 | stmt, nonconstant_names); |
a41f2a28 | 2690 | if (this_time || this_size) |
2691 | { | |
a41f2a28 | 2692 | struct predicate p; |
2693 | ||
2694 | this_time *= freq; | |
c7b2cc59 | 2695 | |
a41f2a28 | 2696 | prob = eliminated_by_inlining_prob (stmt); |
2697 | if (prob == 1 && dump_file && (dump_flags & TDF_DETAILS)) | |
e876d531 | 2698 | fprintf (dump_file, |
2699 | "\t\t50%% will be eliminated by inlining\n"); | |
a41f2a28 | 2700 | if (prob == 2 && dump_file && (dump_flags & TDF_DETAILS)) |
d4db9dfd | 2701 | fprintf (dump_file, "\t\tWill be eliminated by inlining\n"); |
a41f2a28 | 2702 | |
1a673ff0 | 2703 | if (fbi.info) |
fb3c587e | 2704 | p = and_predicates (info->conds, &bb_predicate, |
2705 | &will_be_nonconstant); | |
a41f2a28 | 2706 | else |
2707 | p = true_predicate (); | |
c7b2cc59 | 2708 | |
1d4ee63f | 2709 | if (!false_predicate_p (&p) |
2710 | || (is_gimple_call (stmt) | |
2711 | && !false_predicate_p (&bb_predicate))) | |
18b64b34 | 2712 | { |
2713 | time += this_time; | |
2714 | size += this_size; | |
3c49142d | 2715 | if (time > MAX_TIME * INLINE_TIME_SCALE) |
2716 | time = MAX_TIME * INLINE_TIME_SCALE; | |
18b64b34 | 2717 | } |
2718 | ||
a41f2a28 | 2719 | /* We account everything but the calls. Calls have their own |
e876d531 | 2720 | size/time info attached to cgraph edges. This is necessary |
2721 | in order to make the cost disappear after inlining. */ | |
a41f2a28 | 2722 | if (!is_gimple_call (stmt)) |
2723 | { | |
2724 | if (prob) | |
2725 | { | |
2726 | struct predicate ip = not_inlined_predicate (); | |
94646c9c | 2727 | ip = and_predicates (info->conds, &ip, &p); |
a41f2a28 | 2728 | account_size_time (info, this_size * prob, |
2729 | this_time * prob, &ip); | |
2730 | } | |
2731 | if (prob != 2) | |
2732 | account_size_time (info, this_size * (2 - prob), | |
2733 | this_time * (2 - prob), &p); | |
2734 | } | |
c7b2cc59 | 2735 | |
a41f2a28 | 2736 | gcc_assert (time >= 0); |
2737 | gcc_assert (size >= 0); | |
2738 | } | |
99c67f24 | 2739 | } |
2740 | } | |
b4bae7a0 | 2741 | set_hint_predicate (&inline_summaries->get (node)->array_index, array_index); |
99c67f24 | 2742 | time = (time + CGRAPH_FREQ_BASE / 2) / CGRAPH_FREQ_BASE; |
99c67f24 | 2743 | if (time > MAX_TIME) |
2744 | time = MAX_TIME; | |
563fae60 | 2745 | free (order); |
7c07aa3d | 2746 | |
f4e523eb | 2747 | if (nonconstant_names.exists () && !early) |
7c07aa3d | 2748 | { |
2749 | struct loop *loop; | |
7c07aa3d | 2750 | predicate loop_iterations = true_predicate (); |
3716ee8f | 2751 | predicate loop_stride = true_predicate (); |
7c07aa3d | 2752 | |
7c07aa3d | 2753 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2754 | flow_loops_dump (dump_file, NULL, 0); | |
2755 | scev_initialize (); | |
f21d4d00 | 2756 | FOR_EACH_LOOP (loop, 0) |
7c07aa3d | 2757 | { |
e876d531 | 2758 | vec<edge> exits; |
2759 | edge ex; | |
73f86fed | 2760 | unsigned int j; |
7c07aa3d | 2761 | struct tree_niter_desc niter_desc; |
e876d531 | 2762 | bb_predicate = *(struct predicate *) loop->header->aux; |
7c07aa3d | 2763 | |
2764 | exits = get_loop_exit_edges (loop); | |
e876d531 | 2765 | FOR_EACH_VEC_ELT (exits, j, ex) |
7c07aa3d | 2766 | if (number_of_iterations_exit (loop, ex, &niter_desc, false) |
2767 | && !is_gimple_min_invariant (niter_desc.niter)) | |
e876d531 | 2768 | { |
2769 | predicate will_be_nonconstant | |
1a673ff0 | 2770 | = will_be_nonconstant_expr_predicate (fbi.info, info, |
e876d531 | 2771 | niter_desc.niter, |
2772 | nonconstant_names); | |
2773 | if (!true_predicate_p (&will_be_nonconstant)) | |
2774 | will_be_nonconstant = and_predicates (info->conds, | |
2775 | &bb_predicate, | |
2776 | &will_be_nonconstant); | |
2777 | if (!true_predicate_p (&will_be_nonconstant) | |
2778 | && !false_predicate_p (&will_be_nonconstant)) | |
2779 | /* This is slightly inprecise. We may want to represent each | |
2780 | loop with independent predicate. */ | |
2781 | loop_iterations = | |
2782 | and_predicates (info->conds, &loop_iterations, | |
2783 | &will_be_nonconstant); | |
2784 | } | |
2785 | exits.release (); | |
0a65bce6 | 2786 | } |
3716ee8f | 2787 | |
0a65bce6 | 2788 | /* To avoid quadratic behavior we analyze stride predicates only |
2789 | with respect to the containing loop. Thus we simply iterate | |
2790 | over all defs in the outermost loop body. */ | |
2791 | for (loop = loops_for_fn (cfun)->tree_root->inner; | |
2792 | loop != NULL; loop = loop->next) | |
2793 | { | |
2794 | basic_block *body = get_loop_body (loop); | |
2795 | for (unsigned i = 0; i < loop->num_nodes; i++) | |
3716ee8f | 2796 | { |
0a65bce6 | 2797 | gimple_stmt_iterator gsi; |
2798 | bb_predicate = *(struct predicate *) body[i]->aux; | |
2799 | for (gsi = gsi_start_bb (body[i]); !gsi_end_p (gsi); | |
2800 | gsi_next (&gsi)) | |
2801 | { | |
2802 | gimple *stmt = gsi_stmt (gsi); | |
2803 | ||
2804 | if (!is_gimple_assign (stmt)) | |
2805 | continue; | |
2806 | ||
2807 | tree def = gimple_assign_lhs (stmt); | |
2808 | if (TREE_CODE (def) != SSA_NAME) | |
2809 | continue; | |
2810 | ||
2811 | affine_iv iv; | |
2812 | if (!simple_iv (loop_containing_stmt (stmt), | |
2813 | loop_containing_stmt (stmt), | |
2814 | def, &iv, true) | |
2815 | || is_gimple_min_invariant (iv.step)) | |
2816 | continue; | |
2817 | ||
2818 | predicate will_be_nonconstant | |
2819 | = will_be_nonconstant_expr_predicate (fbi.info, info, | |
2820 | iv.step, | |
2821 | nonconstant_names); | |
2822 | if (!true_predicate_p (&will_be_nonconstant)) | |
2823 | will_be_nonconstant | |
2824 | = and_predicates (info->conds, &bb_predicate, | |
2825 | &will_be_nonconstant); | |
2826 | if (!true_predicate_p (&will_be_nonconstant) | |
2827 | && !false_predicate_p (&will_be_nonconstant)) | |
2828 | /* This is slightly inprecise. We may want to represent | |
2829 | each loop with independent predicate. */ | |
2830 | loop_stride = and_predicates (info->conds, &loop_stride, | |
2831 | &will_be_nonconstant); | |
2832 | } | |
3716ee8f | 2833 | } |
0a65bce6 | 2834 | free (body); |
7c07aa3d | 2835 | } |
b4bae7a0 | 2836 | set_hint_predicate (&inline_summaries->get (node)->loop_iterations, |
e876d531 | 2837 | loop_iterations); |
0a65bce6 | 2838 | set_hint_predicate (&inline_summaries->get (node)->loop_stride, |
2839 | loop_stride); | |
7c07aa3d | 2840 | scev_finalize (); |
7c07aa3d | 2841 | } |
be343a9c | 2842 | FOR_ALL_BB_FN (bb, my_function) |
2843 | { | |
2844 | edge e; | |
2845 | edge_iterator ei; | |
2846 | ||
2847 | if (bb->aux) | |
2a5261f7 | 2848 | edge_predicate_pool.remove ((predicate *)bb->aux); |
be343a9c | 2849 | bb->aux = NULL; |
2850 | FOR_EACH_EDGE (e, ei, bb->succs) | |
2851 | { | |
2852 | if (e->aux) | |
2a5261f7 | 2853 | edge_predicate_pool.remove ((predicate *) e->aux); |
be343a9c | 2854 | e->aux = NULL; |
2855 | } | |
2856 | } | |
b4bae7a0 | 2857 | inline_summaries->get (node)->self_time = time; |
2858 | inline_summaries->get (node)->self_size = size; | |
f1f41a6c | 2859 | nonconstant_names.release (); |
73bd7d5f | 2860 | ipa_release_body_info (&fbi); |
f4e523eb | 2861 | if (opt_for_fn (node->decl, optimize)) |
4b209fe7 | 2862 | { |
f4e523eb | 2863 | if (!early) |
2864 | loop_optimizer_finalize (); | |
5e8fd8b6 | 2865 | else if (!ipa_edge_args_vector) |
f4e523eb | 2866 | ipa_free_all_node_params (); |
4b209fe7 | 2867 | free_dominance_info (CDI_DOMINATORS); |
2868 | } | |
a41f2a28 | 2869 | if (dump_file) |
2870 | { | |
2871 | fprintf (dump_file, "\n"); | |
2872 | dump_inline_summary (dump_file, node); | |
2873 | } | |
99c67f24 | 2874 | } |
2875 | ||
2876 | ||
a41f2a28 | 2877 | /* Compute parameters of functions used by inliner. |
2878 | EARLY is true when we compute parameters for the early inliner */ | |
99c67f24 | 2879 | |
2880 | void | |
a41f2a28 | 2881 | compute_inline_parameters (struct cgraph_node *node, bool early) |
99c67f24 | 2882 | { |
2883 | HOST_WIDE_INT self_stack_size; | |
2884 | struct cgraph_edge *e; | |
cbd7f5a0 | 2885 | struct inline_summary *info; |
99c67f24 | 2886 | |
2887 | gcc_assert (!node->global.inlined_to); | |
2888 | ||
c7b2cc59 | 2889 | inline_summary_alloc (); |
2890 | ||
b4bae7a0 | 2891 | info = inline_summaries->get (node); |
2892 | reset_inline_summary (node, info); | |
cbd7f5a0 | 2893 | |
91bf9d9a | 2894 | /* FIXME: Thunks are inlinable, but tree-inline don't know how to do that. |
2895 | Once this happen, we will need to more curefully predict call | |
2896 | statement size. */ | |
2897 | if (node->thunk.thunk_p) | |
2898 | { | |
2899 | struct inline_edge_summary *es = inline_edge_summary (node->callees); | |
2900 | struct predicate t = true_predicate (); | |
2901 | ||
c8d92fc1 | 2902 | info->inlinable = 0; |
91bf9d9a | 2903 | node->callees->call_stmt_cannot_inline_p = true; |
2904 | node->local.can_change_signature = false; | |
2905 | es->call_stmt_time = 1; | |
2906 | es->call_stmt_size = 1; | |
2907 | account_size_time (info, 0, 0, &t); | |
2908 | return; | |
2909 | } | |
2910 | ||
8e22665e | 2911 | /* Even is_gimple_min_invariant rely on current_function_decl. */ |
02774f2d | 2912 | push_cfun (DECL_STRUCT_FUNCTION (node->decl)); |
8e22665e | 2913 | |
99c67f24 | 2914 | /* Estimate the stack size for the function if we're optimizing. */ |
2915 | self_stack_size = optimize ? estimated_stack_frame_size (node) : 0; | |
cbd7f5a0 | 2916 | info->estimated_self_stack_size = self_stack_size; |
2917 | info->estimated_stack_size = self_stack_size; | |
2918 | info->stack_frame_offset = 0; | |
99c67f24 | 2919 | |
2920 | /* Can this function be inlined at all? */ | |
d1f68cd8 | 2921 | if (!opt_for_fn (node->decl, optimize) |
2922 | && !lookup_attribute ("always_inline", | |
2923 | DECL_ATTRIBUTES (node->decl))) | |
26051fcf | 2924 | info->inlinable = false; |
2925 | else | |
02774f2d | 2926 | info->inlinable = tree_inlinable_function_p (node->decl); |
99c67f24 | 2927 | |
e806c56f | 2928 | info->contains_cilk_spawn = fn_contains_cilk_spawn_p (cfun); |
2929 | ||
982ffd8d | 2930 | /* Type attributes can use parameter indices to describe them. */ |
02774f2d | 2931 | if (TYPE_ATTRIBUTES (TREE_TYPE (node->decl))) |
982ffd8d | 2932 | node->local.can_change_signature = false; |
99c67f24 | 2933 | else |
2934 | { | |
982ffd8d | 2935 | /* Otherwise, inlinable functions always can change signature. */ |
2936 | if (info->inlinable) | |
2937 | node->local.can_change_signature = true; | |
2938 | else | |
2939 | { | |
2940 | /* Functions calling builtin_apply can not change signature. */ | |
2941 | for (e = node->callees; e; e = e->next_callee) | |
2942 | { | |
02774f2d | 2943 | tree cdecl = e->callee->decl; |
982ffd8d | 2944 | if (DECL_BUILT_IN (cdecl) |
2945 | && DECL_BUILT_IN_CLASS (cdecl) == BUILT_IN_NORMAL | |
2946 | && (DECL_FUNCTION_CODE (cdecl) == BUILT_IN_APPLY_ARGS | |
2947 | || DECL_FUNCTION_CODE (cdecl) == BUILT_IN_VA_START)) | |
2948 | break; | |
2949 | } | |
2950 | node->local.can_change_signature = !e; | |
2951 | } | |
99c67f24 | 2952 | } |
a41f2a28 | 2953 | estimate_function_body_sizes (node, early); |
c7b2cc59 | 2954 | |
468088ac | 2955 | for (e = node->callees; e; e = e->next_callee) |
415d1b9a | 2956 | if (e->callee->comdat_local_p ()) |
468088ac | 2957 | break; |
2958 | node->calls_comdat_local = (e != NULL); | |
2959 | ||
99c67f24 | 2960 | /* Inlining characteristics are maintained by the cgraph_mark_inline. */ |
cbd7f5a0 | 2961 | info->time = info->self_time; |
2962 | info->size = info->self_size; | |
cbd7f5a0 | 2963 | info->stack_frame_offset = 0; |
2964 | info->estimated_stack_size = info->estimated_self_stack_size; | |
382ecba7 | 2965 | if (flag_checking) |
2966 | { | |
2967 | inline_update_overall_summary (node); | |
2968 | gcc_assert (info->time == info->self_time | |
2969 | && info->size == info->self_size); | |
2970 | } | |
18b64b34 | 2971 | |
8e22665e | 2972 | pop_cfun (); |
99c67f24 | 2973 | } |
2974 | ||
2975 | ||
2976 | /* Compute parameters of functions used by inliner using | |
2977 | current_function_decl. */ | |
2978 | ||
2979 | static unsigned int | |
2980 | compute_inline_parameters_for_current (void) | |
2981 | { | |
415d1b9a | 2982 | compute_inline_parameters (cgraph_node::get (current_function_decl), true); |
99c67f24 | 2983 | return 0; |
2984 | } | |
2985 | ||
cbe8bda8 | 2986 | namespace { |
2987 | ||
2988 | const pass_data pass_data_inline_parameters = | |
99c67f24 | 2989 | { |
cbe8bda8 | 2990 | GIMPLE_PASS, /* type */ |
2991 | "inline_param", /* name */ | |
2992 | OPTGROUP_INLINE, /* optinfo_flags */ | |
cbe8bda8 | 2993 | TV_INLINE_PARAMETERS, /* tv_id */ |
2994 | 0, /* properties_required */ | |
2995 | 0, /* properties_provided */ | |
2996 | 0, /* properties_destroyed */ | |
2997 | 0, /* todo_flags_start */ | |
2998 | 0, /* todo_flags_finish */ | |
99c67f24 | 2999 | }; |
3000 | ||
cbe8bda8 | 3001 | class pass_inline_parameters : public gimple_opt_pass |
3002 | { | |
3003 | public: | |
9af5ce0c | 3004 | pass_inline_parameters (gcc::context *ctxt) |
3005 | : gimple_opt_pass (pass_data_inline_parameters, ctxt) | |
cbe8bda8 | 3006 | {} |
3007 | ||
3008 | /* opt_pass methods: */ | |
ae84f584 | 3009 | opt_pass * clone () { return new pass_inline_parameters (m_ctxt); } |
65b0537f | 3010 | virtual unsigned int execute (function *) |
3011 | { | |
3012 | return compute_inline_parameters_for_current (); | |
3013 | } | |
cbe8bda8 | 3014 | |
3015 | }; // class pass_inline_parameters | |
3016 | ||
3017 | } // anon namespace | |
3018 | ||
3019 | gimple_opt_pass * | |
3020 | make_pass_inline_parameters (gcc::context *ctxt) | |
3021 | { | |
3022 | return new pass_inline_parameters (ctxt); | |
3023 | } | |
3024 | ||
99c67f24 | 3025 | |
245ab191 | 3026 | /* Estimate benefit devirtualizing indirect edge IE, provided KNOWN_VALS, |
3027 | KNOWN_CONTEXTS and KNOWN_AGGS. */ | |
20da2013 | 3028 | |
eb7c606e | 3029 | static bool |
20da2013 | 3030 | estimate_edge_devirt_benefit (struct cgraph_edge *ie, |
18b64b34 | 3031 | int *size, int *time, |
f1f41a6c | 3032 | vec<tree> known_vals, |
245ab191 | 3033 | vec<ipa_polymorphic_call_context> known_contexts, |
f1f41a6c | 3034 | vec<ipa_agg_jump_function_p> known_aggs) |
20da2013 | 3035 | { |
3036 | tree target; | |
eb7c606e | 3037 | struct cgraph_node *callee; |
3038 | struct inline_summary *isummary; | |
f28422b6 | 3039 | enum availability avail; |
f21a87d8 | 3040 | bool speculative; |
20da2013 | 3041 | |
245ab191 | 3042 | if (!known_vals.exists () && !known_contexts.exists ()) |
eb7c606e | 3043 | return false; |
d1f68cd8 | 3044 | if (!opt_for_fn (ie->caller->decl, flag_indirect_inlining)) |
18b64b34 | 3045 | return false; |
20da2013 | 3046 | |
245ab191 | 3047 | target = ipa_get_indirect_edge_target (ie, known_vals, known_contexts, |
f21a87d8 | 3048 | known_aggs, &speculative); |
3049 | if (!target || speculative) | |
eb7c606e | 3050 | return false; |
20da2013 | 3051 | |
3052 | /* Account for difference in cost between indirect and direct calls. */ | |
18b64b34 | 3053 | *size -= (eni_size_weights.indirect_call_cost - eni_size_weights.call_cost); |
3054 | *time -= (eni_time_weights.indirect_call_cost - eni_time_weights.call_cost); | |
3055 | gcc_checking_assert (*time >= 0); | |
3056 | gcc_checking_assert (*size >= 0); | |
5dcaa672 | 3057 | |
415d1b9a | 3058 | callee = cgraph_node::get (target); |
02774f2d | 3059 | if (!callee || !callee->definition) |
eb7c606e | 3060 | return false; |
415d1b9a | 3061 | callee = callee->function_symbol (&avail); |
f28422b6 | 3062 | if (avail < AVAIL_AVAILABLE) |
3063 | return false; | |
b4bae7a0 | 3064 | isummary = inline_summaries->get (callee); |
eb7c606e | 3065 | return isummary->inlinable; |
20da2013 | 3066 | } |
3067 | ||
db197f90 | 3068 | /* Increase SIZE, MIN_SIZE (if non-NULL) and TIME for size and time needed to |
3069 | handle edge E with probability PROB. | |
3070 | Set HINTS if edge may be devirtualized. | |
245ab191 | 3071 | KNOWN_VALS, KNOWN_AGGS and KNOWN_CONTEXTS describe context of the call |
db197f90 | 3072 | site. */ |
18b64b34 | 3073 | |
3074 | static inline void | |
db197f90 | 3075 | estimate_edge_size_and_time (struct cgraph_edge *e, int *size, int *min_size, |
3076 | int *time, | |
18b64b34 | 3077 | int prob, |
f1f41a6c | 3078 | vec<tree> known_vals, |
245ab191 | 3079 | vec<ipa_polymorphic_call_context> known_contexts, |
f1f41a6c | 3080 | vec<ipa_agg_jump_function_p> known_aggs, |
18b64b34 | 3081 | inline_hints *hints) |
18b64b34 | 3082 | { |
3083 | struct inline_edge_summary *es = inline_edge_summary (e); | |
3084 | int call_size = es->call_stmt_size; | |
3085 | int call_time = es->call_stmt_time; | |
db197f90 | 3086 | int cur_size; |
18b64b34 | 3087 | if (!e->callee |
3088 | && estimate_edge_devirt_benefit (e, &call_size, &call_time, | |
245ab191 | 3089 | known_vals, known_contexts, known_aggs) |
35ee1c66 | 3090 | && hints && e->maybe_hot_p ()) |
18b64b34 | 3091 | *hints |= INLINE_HINT_indirect_call; |
db197f90 | 3092 | cur_size = call_size * INLINE_SIZE_SCALE; |
3093 | *size += cur_size; | |
3094 | if (min_size) | |
3095 | *min_size += cur_size; | |
70074000 | 3096 | *time += apply_probability ((gcov_type) call_time, prob) |
e876d531 | 3097 | * e->frequency * (INLINE_TIME_SCALE / CGRAPH_FREQ_BASE); |
18b64b34 | 3098 | if (*time > MAX_TIME * INLINE_TIME_SCALE) |
3099 | *time = MAX_TIME * INLINE_TIME_SCALE; | |
3100 | } | |
3101 | ||
3102 | ||
20da2013 | 3103 | |
db197f90 | 3104 | /* Increase SIZE, MIN_SIZE and TIME for size and time needed to handle all |
245ab191 | 3105 | calls in NODE. POSSIBLE_TRUTHS, KNOWN_VALS, KNOWN_AGGS and KNOWN_CONTEXTS |
3106 | describe context of the call site. */ | |
a41f2a28 | 3107 | |
3108 | static void | |
db197f90 | 3109 | estimate_calls_size_and_time (struct cgraph_node *node, int *size, |
3110 | int *min_size, int *time, | |
eb7c606e | 3111 | inline_hints *hints, |
20da2013 | 3112 | clause_t possible_truths, |
f1f41a6c | 3113 | vec<tree> known_vals, |
245ab191 | 3114 | vec<ipa_polymorphic_call_context> known_contexts, |
f1f41a6c | 3115 | vec<ipa_agg_jump_function_p> known_aggs) |
a41f2a28 | 3116 | { |
3117 | struct cgraph_edge *e; | |
3118 | for (e = node->callees; e; e = e->next_callee) | |
6a18c0be | 3119 | { |
b0343e00 | 3120 | if (inline_edge_summary_vec.length () <= (unsigned) e->uid) |
3121 | continue; | |
3122 | ||
6a18c0be | 3123 | struct inline_edge_summary *es = inline_edge_summary (e); |
f4e523eb | 3124 | |
3125 | /* Do not care about zero sized builtins. */ | |
3126 | if (e->inline_failed && !es->call_stmt_size) | |
3127 | { | |
3128 | gcc_checking_assert (!es->call_stmt_time); | |
3129 | continue; | |
3130 | } | |
e876d531 | 3131 | if (!es->predicate |
3132 | || evaluate_predicate (es->predicate, possible_truths)) | |
6a18c0be | 3133 | { |
3134 | if (e->inline_failed) | |
eb4ae064 | 3135 | { |
3136 | /* Predicates of calls shall not use NOT_CHANGED codes, | |
e876d531 | 3137 | sowe do not need to compute probabilities. */ |
db197f90 | 3138 | estimate_edge_size_and_time (e, size, |
3139 | es->predicate ? NULL : min_size, | |
3140 | time, REG_BR_PROB_BASE, | |
245ab191 | 3141 | known_vals, known_contexts, |
e876d531 | 3142 | known_aggs, hints); |
eb4ae064 | 3143 | } |
6a18c0be | 3144 | else |
db197f90 | 3145 | estimate_calls_size_and_time (e->callee, size, min_size, time, |
3146 | hints, | |
20da2013 | 3147 | possible_truths, |
245ab191 | 3148 | known_vals, known_contexts, |
e876d531 | 3149 | known_aggs); |
6a18c0be | 3150 | } |
3151 | } | |
a41f2a28 | 3152 | for (e = node->indirect_calls; e; e = e->next_callee) |
6a18c0be | 3153 | { |
b0343e00 | 3154 | if (inline_edge_summary_vec.length () <= (unsigned) e->uid) |
3155 | continue; | |
3156 | ||
6a18c0be | 3157 | struct inline_edge_summary *es = inline_edge_summary (e); |
e876d531 | 3158 | if (!es->predicate |
3159 | || evaluate_predicate (es->predicate, possible_truths)) | |
db197f90 | 3160 | estimate_edge_size_and_time (e, size, |
3161 | es->predicate ? NULL : min_size, | |
3162 | time, REG_BR_PROB_BASE, | |
245ab191 | 3163 | known_vals, known_contexts, known_aggs, |
18b64b34 | 3164 | hints); |
6a18c0be | 3165 | } |
a41f2a28 | 3166 | } |
3167 | ||
3168 | ||
8bae3ea4 | 3169 | /* Estimate size and time needed to execute NODE assuming |
245ab191 | 3170 | POSSIBLE_TRUTHS clause, and KNOWN_VALS, KNOWN_AGGS and KNOWN_CONTEXTS |
db197f90 | 3171 | information about NODE's arguments. If non-NULL use also probability |
3172 | information present in INLINE_PARAM_SUMMARY vector. | |
3173 | Additionally detemine hints determined by the context. Finally compute | |
3174 | minimal size needed for the call that is independent on the call context and | |
3175 | can be used for fast estimates. Return the values in RET_SIZE, | |
3176 | RET_MIN_SIZE, RET_TIME and RET_HINTS. */ | |
99c67f24 | 3177 | |
a41f2a28 | 3178 | static void |
8bae3ea4 | 3179 | estimate_node_size_and_time (struct cgraph_node *node, |
3180 | clause_t possible_truths, | |
f1f41a6c | 3181 | vec<tree> known_vals, |
245ab191 | 3182 | vec<ipa_polymorphic_call_context> known_contexts, |
f1f41a6c | 3183 | vec<ipa_agg_jump_function_p> known_aggs, |
db197f90 | 3184 | int *ret_size, int *ret_min_size, int *ret_time, |
eb7c606e | 3185 | inline_hints *ret_hints, |
b3e7c666 | 3186 | vec<inline_param_summary> |
e876d531 | 3187 | inline_param_summary) |
99c67f24 | 3188 | { |
b4bae7a0 | 3189 | struct inline_summary *info = inline_summaries->get (node); |
a41f2a28 | 3190 | size_time_entry *e; |
18b64b34 | 3191 | int size = 0; |
3192 | int time = 0; | |
db197f90 | 3193 | int min_size = 0; |
eb7c606e | 3194 | inline_hints hints = 0; |
a41f2a28 | 3195 | int i; |
3196 | ||
e876d531 | 3197 | if (dump_file && (dump_flags & TDF_DETAILS)) |
a41f2a28 | 3198 | { |
3199 | bool found = false; | |
8bae3ea4 | 3200 | fprintf (dump_file, " Estimating body: %s/%i\n" |
f1c8b4d7 | 3201 | " Known to be false: ", node->name (), |
02774f2d | 3202 | node->order); |
a41f2a28 | 3203 | |
3204 | for (i = predicate_not_inlined_condition; | |
3205 | i < (predicate_first_dynamic_condition | |
e876d531 | 3206 | + (int) vec_safe_length (info->conds)); i++) |
8bae3ea4 | 3207 | if (!(possible_truths & (1 << i))) |
a41f2a28 | 3208 | { |
3209 | if (found) | |
3210 | fprintf (dump_file, ", "); | |
3211 | found = true; | |
e876d531 | 3212 | dump_condition (dump_file, info->conds, i); |
a41f2a28 | 3213 | } |
3214 | } | |
3215 | ||
f1f41a6c | 3216 | for (i = 0; vec_safe_iterate (info->entry, i, &e); i++) |
8bae3ea4 | 3217 | if (evaluate_predicate (&e->predicate, possible_truths)) |
eb4ae064 | 3218 | { |
3219 | size += e->size; | |
18b64b34 | 3220 | gcc_checking_assert (e->time >= 0); |
e876d531 | 3221 | gcc_checking_assert (time >= 0); |
f1f41a6c | 3222 | if (!inline_param_summary.exists ()) |
eb4ae064 | 3223 | time += e->time; |
3224 | else | |
3225 | { | |
3226 | int prob = predicate_probability (info->conds, | |
3227 | &e->predicate, | |
3228 | possible_truths, | |
3229 | inline_param_summary); | |
18b64b34 | 3230 | gcc_checking_assert (prob >= 0); |
3231 | gcc_checking_assert (prob <= REG_BR_PROB_BASE); | |
70074000 | 3232 | time += apply_probability ((gcov_type) e->time, prob); |
eb4ae064 | 3233 | } |
e876d531 | 3234 | if (time > MAX_TIME * INLINE_TIME_SCALE) |
3235 | time = MAX_TIME * INLINE_TIME_SCALE; | |
3236 | gcc_checking_assert (time >= 0); | |
3237 | ||
eb4ae064 | 3238 | } |
db197f90 | 3239 | gcc_checking_assert (true_predicate_p (&(*info->entry)[0].predicate)); |
3240 | min_size = (*info->entry)[0].size; | |
18b64b34 | 3241 | gcc_checking_assert (size >= 0); |
3242 | gcc_checking_assert (time >= 0); | |
cbd7f5a0 | 3243 | |
7c07aa3d | 3244 | if (info->loop_iterations |
3245 | && !evaluate_predicate (info->loop_iterations, possible_truths)) | |
e876d531 | 3246 | hints |= INLINE_HINT_loop_iterations; |
3716ee8f | 3247 | if (info->loop_stride |
3248 | && !evaluate_predicate (info->loop_stride, possible_truths)) | |
e876d531 | 3249 | hints |= INLINE_HINT_loop_stride; |
be343a9c | 3250 | if (info->array_index |
3251 | && !evaluate_predicate (info->array_index, possible_truths)) | |
e876d531 | 3252 | hints |= INLINE_HINT_array_index; |
41d39f38 | 3253 | if (info->scc_no) |
3254 | hints |= INLINE_HINT_in_scc; | |
02774f2d | 3255 | if (DECL_DECLARED_INLINE_P (node->decl)) |
3172b7bf | 3256 | hints |= INLINE_HINT_declared_inline; |
a41f2a28 | 3257 | |
db197f90 | 3258 | estimate_calls_size_and_time (node, &size, &min_size, &time, &hints, possible_truths, |
245ab191 | 3259 | known_vals, known_contexts, known_aggs); |
18b64b34 | 3260 | gcc_checking_assert (size >= 0); |
3261 | gcc_checking_assert (time >= 0); | |
3262 | time = RDIV (time, INLINE_TIME_SCALE); | |
3263 | size = RDIV (size, INLINE_SIZE_SCALE); | |
db197f90 | 3264 | min_size = RDIV (min_size, INLINE_SIZE_SCALE); |
a41f2a28 | 3265 | |
e876d531 | 3266 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3267 | fprintf (dump_file, "\n size:%i time:%i\n", (int) size, (int) time); | |
a41f2a28 | 3268 | if (ret_time) |
3269 | *ret_time = time; | |
3270 | if (ret_size) | |
3271 | *ret_size = size; | |
db197f90 | 3272 | if (ret_min_size) |
3273 | *ret_min_size = min_size; | |
eb7c606e | 3274 | if (ret_hints) |
3275 | *ret_hints = hints; | |
a41f2a28 | 3276 | return; |
3277 | } | |
3278 | ||
3279 | ||
93f713da | 3280 | /* Estimate size and time needed to execute callee of EDGE assuming that |
3281 | parameters known to be constant at caller of EDGE are propagated. | |
245ab191 | 3282 | KNOWN_VALS and KNOWN_CONTEXTS are vectors of assumed known constant values |
20da2013 | 3283 | and types for parameters. */ |
8bae3ea4 | 3284 | |
3285 | void | |
3286 | estimate_ipcp_clone_size_and_time (struct cgraph_node *node, | |
e876d531 | 3287 | vec<tree> known_vals, |
245ab191 | 3288 | vec<ipa_polymorphic_call_context> |
3289 | known_contexts, | |
e876d531 | 3290 | vec<ipa_agg_jump_function_p> known_aggs, |
3291 | int *ret_size, int *ret_time, | |
3292 | inline_hints *hints) | |
8bae3ea4 | 3293 | { |
93f713da | 3294 | clause_t clause; |
3295 | ||
803a7988 | 3296 | clause = evaluate_conditions_for_known_args (node, false, known_vals, |
3297 | known_aggs); | |
245ab191 | 3298 | estimate_node_size_and_time (node, clause, known_vals, known_contexts, |
db197f90 | 3299 | known_aggs, ret_size, NULL, ret_time, hints, vNULL); |
8bae3ea4 | 3300 | } |
3301 | ||
eb4ae064 | 3302 | /* Translate all conditions from callee representation into caller |
3303 | representation and symbolically evaluate predicate P into new predicate. | |
6a18c0be | 3304 | |
a4f60e55 | 3305 | INFO is inline_summary of function we are adding predicate into, CALLEE_INFO |
3306 | is summary of function predicate P is from. OPERAND_MAP is array giving | |
3307 | callee formal IDs the caller formal IDs. POSSSIBLE_TRUTHS is clausule of all | |
3308 | callee conditions that may be true in caller context. TOPLEV_PREDICATE is | |
3309 | predicate under which callee is executed. OFFSET_MAP is an array of of | |
3310 | offsets that need to be added to conditions, negative offset means that | |
3311 | conditions relying on values passed by reference have to be discarded | |
3312 | because they might not be preserved (and should be considered offset zero | |
3313 | for other purposes). */ | |
a41f2a28 | 3314 | |
3315 | static struct predicate | |
eb4ae064 | 3316 | remap_predicate (struct inline_summary *info, |
3317 | struct inline_summary *callee_info, | |
a41f2a28 | 3318 | struct predicate *p, |
f1f41a6c | 3319 | vec<int> operand_map, |
3320 | vec<int> offset_map, | |
e876d531 | 3321 | clause_t possible_truths, struct predicate *toplev_predicate) |
a41f2a28 | 3322 | { |
3323 | int i; | |
3324 | struct predicate out = true_predicate (); | |
3325 | ||
3326 | /* True predicate is easy. */ | |
6a18c0be | 3327 | if (true_predicate_p (p)) |
3328 | return *toplev_predicate; | |
a41f2a28 | 3329 | for (i = 0; p->clause[i]; i++) |
3330 | { | |
3331 | clause_t clause = p->clause[i]; | |
3332 | int cond; | |
3333 | struct predicate clause_predicate = false_predicate (); | |
3334 | ||
5cb1b112 | 3335 | gcc_assert (i < MAX_CLAUSES); |
3336 | ||
e876d531 | 3337 | for (cond = 0; cond < NUM_CONDITIONS; cond++) |
a41f2a28 | 3338 | /* Do we have condition we can't disprove? */ |
3339 | if (clause & possible_truths & (1 << cond)) | |
3340 | { | |
3341 | struct predicate cond_predicate; | |
3342 | /* Work out if the condition can translate to predicate in the | |
3343 | inlined function. */ | |
3344 | if (cond >= predicate_first_dynamic_condition) | |
3345 | { | |
e876d531 | 3346 | struct condition *c; |
3347 | ||
3348 | c = &(*callee_info->conds)[cond | |
3349 | - | |
3350 | predicate_first_dynamic_condition]; | |
3351 | /* See if we can remap condition operand to caller's operand. | |
3352 | Otherwise give up. */ | |
3353 | if (!operand_map.exists () | |
3354 | || (int) operand_map.length () <= c->operand_num | |
3355 | || operand_map[c->operand_num] == -1 | |
3356 | /* TODO: For non-aggregate conditions, adding an offset is | |
3357 | basically an arithmetic jump function processing which | |
3358 | we should support in future. */ | |
3359 | || ((!c->agg_contents || !c->by_ref) | |
3360 | && offset_map[c->operand_num] > 0) | |
3361 | || (c->agg_contents && c->by_ref | |
3362 | && offset_map[c->operand_num] < 0)) | |
3363 | cond_predicate = true_predicate (); | |
3364 | else | |
3365 | { | |
3366 | struct agg_position_info ap; | |
3367 | HOST_WIDE_INT offset_delta = offset_map[c->operand_num]; | |
3368 | if (offset_delta < 0) | |
3369 | { | |
3370 | gcc_checking_assert (!c->agg_contents || !c->by_ref); | |
3371 | offset_delta = 0; | |
3372 | } | |
3373 | gcc_assert (!c->agg_contents | |
3374 | || c->by_ref || offset_delta == 0); | |
3375 | ap.offset = c->offset + offset_delta; | |
3376 | ap.agg_contents = c->agg_contents; | |
3377 | ap.by_ref = c->by_ref; | |
3378 | cond_predicate = add_condition (info, | |
3379 | operand_map[c->operand_num], | |
3380 | &ap, c->code, c->val); | |
3381 | } | |
a41f2a28 | 3382 | } |
3383 | /* Fixed conditions remains same, construct single | |
3384 | condition predicate. */ | |
3385 | else | |
3386 | { | |
3387 | cond_predicate.clause[0] = 1 << cond; | |
3388 | cond_predicate.clause[1] = 0; | |
3389 | } | |
94646c9c | 3390 | clause_predicate = or_predicates (info->conds, &clause_predicate, |
3391 | &cond_predicate); | |
a41f2a28 | 3392 | } |
94646c9c | 3393 | out = and_predicates (info->conds, &out, &clause_predicate); |
a41f2a28 | 3394 | } |
94646c9c | 3395 | return and_predicates (info->conds, &out, toplev_predicate); |
a41f2a28 | 3396 | } |
3397 | ||
3398 | ||
0835ad03 | 3399 | /* Update summary information of inline clones after inlining. |
3400 | Compute peak stack usage. */ | |
3401 | ||
3402 | static void | |
e876d531 | 3403 | inline_update_callee_summaries (struct cgraph_node *node, int depth) |
0835ad03 | 3404 | { |
3405 | struct cgraph_edge *e; | |
b4bae7a0 | 3406 | struct inline_summary *callee_info = inline_summaries->get (node); |
3407 | struct inline_summary *caller_info = inline_summaries->get (node->callers->caller); | |
0835ad03 | 3408 | HOST_WIDE_INT peak; |
3409 | ||
3410 | callee_info->stack_frame_offset | |
3411 | = caller_info->stack_frame_offset | |
e876d531 | 3412 | + caller_info->estimated_self_stack_size; |
0835ad03 | 3413 | peak = callee_info->stack_frame_offset |
e876d531 | 3414 | + callee_info->estimated_self_stack_size; |
b4bae7a0 | 3415 | if (inline_summaries->get (node->global.inlined_to)->estimated_stack_size < peak) |
3416 | inline_summaries->get (node->global.inlined_to)->estimated_stack_size = peak; | |
6eaf903b | 3417 | ipa_propagate_frequency (node); |
0835ad03 | 3418 | for (e = node->callees; e; e = e->next_callee) |
3419 | { | |
3420 | if (!e->inline_failed) | |
3421 | inline_update_callee_summaries (e->callee, depth); | |
3422 | inline_edge_summary (e)->loop_depth += depth; | |
3423 | } | |
3424 | for (e = node->indirect_calls; e; e = e->next_callee) | |
3425 | inline_edge_summary (e)->loop_depth += depth; | |
3426 | } | |
3427 | ||
eb4ae064 | 3428 | /* Update change_prob of EDGE after INLINED_EDGE has been inlined. |
3429 | When functoin A is inlined in B and A calls C with parameter that | |
3430 | changes with probability PROB1 and C is known to be passthroug | |
3431 | of argument if B that change with probability PROB2, the probability | |
3432 | of change is now PROB1*PROB2. */ | |
3433 | ||
3434 | static void | |
3435 | remap_edge_change_prob (struct cgraph_edge *inlined_edge, | |
3436 | struct cgraph_edge *edge) | |
3437 | { | |
2cc80ac3 | 3438 | if (ipa_node_params_sum) |
eb4ae064 | 3439 | { |
3440 | int i; | |
3441 | struct ipa_edge_args *args = IPA_EDGE_REF (edge); | |
3442 | struct inline_edge_summary *es = inline_edge_summary (edge); | |
3443 | struct inline_edge_summary *inlined_es | |
e876d531 | 3444 | = inline_edge_summary (inlined_edge); |
eb4ae064 | 3445 | |
3446 | for (i = 0; i < ipa_get_cs_argument_count (args); i++) | |
3447 | { | |
3448 | struct ipa_jump_func *jfunc = ipa_get_ith_jump_func (args, i); | |
3449 | if (jfunc->type == IPA_JF_PASS_THROUGH | |
4fa83f96 | 3450 | && (ipa_get_jf_pass_through_formal_id (jfunc) |
f1f41a6c | 3451 | < (int) inlined_es->param.length ())) |
eb4ae064 | 3452 | { |
4fa83f96 | 3453 | int jf_formal_id = ipa_get_jf_pass_through_formal_id (jfunc); |
f1f41a6c | 3454 | int prob1 = es->param[i].change_prob; |
3455 | int prob2 = inlined_es->param[jf_formal_id].change_prob; | |
f9d4b7f4 | 3456 | int prob = combine_probabilities (prob1, prob2); |
eb4ae064 | 3457 | |
3458 | if (prob1 && prob2 && !prob) | |
3459 | prob = 1; | |
3460 | ||
f1f41a6c | 3461 | es->param[i].change_prob = prob; |
eb4ae064 | 3462 | } |
3463 | } | |
e876d531 | 3464 | } |
eb4ae064 | 3465 | } |
3466 | ||
3467 | /* Update edge summaries of NODE after INLINED_EDGE has been inlined. | |
3468 | ||
3469 | Remap predicates of callees of NODE. Rest of arguments match | |
3470 | remap_predicate. | |
0835ad03 | 3471 | |
eb4ae064 | 3472 | Also update change probabilities. */ |
6a18c0be | 3473 | |
3474 | static void | |
e876d531 | 3475 | remap_edge_summaries (struct cgraph_edge *inlined_edge, |
3476 | struct cgraph_node *node, | |
3477 | struct inline_summary *info, | |
3478 | struct inline_summary *callee_info, | |
3479 | vec<int> operand_map, | |
3480 | vec<int> offset_map, | |
3481 | clause_t possible_truths, | |
3482 | struct predicate *toplev_predicate) | |
6a18c0be | 3483 | { |
99fe8cbd | 3484 | struct cgraph_edge *e, *next; |
3485 | for (e = node->callees; e; e = next) | |
6a18c0be | 3486 | { |
3487 | struct inline_edge_summary *es = inline_edge_summary (e); | |
3488 | struct predicate p; | |
99fe8cbd | 3489 | next = e->next_callee; |
eb4ae064 | 3490 | |
a226c368 | 3491 | if (e->inline_failed) |
6a18c0be | 3492 | { |
eb4ae064 | 3493 | remap_edge_change_prob (inlined_edge, e); |
3494 | ||
a226c368 | 3495 | if (es->predicate) |
6a18c0be | 3496 | { |
a226c368 | 3497 | p = remap_predicate (info, callee_info, |
a4f60e55 | 3498 | es->predicate, operand_map, offset_map, |
e876d531 | 3499 | possible_truths, toplev_predicate); |
a226c368 | 3500 | edge_set_predicate (e, &p); |
6a18c0be | 3501 | } |
a226c368 | 3502 | else |
3503 | edge_set_predicate (e, toplev_predicate); | |
6a18c0be | 3504 | } |
a226c368 | 3505 | else |
eb4ae064 | 3506 | remap_edge_summaries (inlined_edge, e->callee, info, callee_info, |
a4f60e55 | 3507 | operand_map, offset_map, possible_truths, |
3508 | toplev_predicate); | |
6a18c0be | 3509 | } |
99fe8cbd | 3510 | for (e = node->indirect_calls; e; e = next) |
6a18c0be | 3511 | { |
3512 | struct inline_edge_summary *es = inline_edge_summary (e); | |
3513 | struct predicate p; | |
99fe8cbd | 3514 | next = e->next_callee; |
eb4ae064 | 3515 | |
3516 | remap_edge_change_prob (inlined_edge, e); | |
6a18c0be | 3517 | if (es->predicate) |
3518 | { | |
3519 | p = remap_predicate (info, callee_info, | |
a4f60e55 | 3520 | es->predicate, operand_map, offset_map, |
3521 | possible_truths, toplev_predicate); | |
6a18c0be | 3522 | edge_set_predicate (e, &p); |
6a18c0be | 3523 | } |
144eea3a | 3524 | else |
3525 | edge_set_predicate (e, toplev_predicate); | |
6a18c0be | 3526 | } |
3527 | } | |
3528 | ||
3716ee8f | 3529 | /* Same as remap_predicate, but set result into hint *HINT. */ |
3530 | ||
3531 | static void | |
3532 | remap_hint_predicate (struct inline_summary *info, | |
3533 | struct inline_summary *callee_info, | |
3534 | struct predicate **hint, | |
f1f41a6c | 3535 | vec<int> operand_map, |
3536 | vec<int> offset_map, | |
3716ee8f | 3537 | clause_t possible_truths, |
3538 | struct predicate *toplev_predicate) | |
3539 | { | |
3540 | predicate p; | |
3541 | ||
3542 | if (!*hint) | |
3543 | return; | |
3544 | p = remap_predicate (info, callee_info, | |
3545 | *hint, | |
3546 | operand_map, offset_map, | |
e876d531 | 3547 | possible_truths, toplev_predicate); |
3548 | if (!false_predicate_p (&p) && !true_predicate_p (&p)) | |
3716ee8f | 3549 | { |
3550 | if (!*hint) | |
3551 | set_hint_predicate (hint, p); | |
3552 | else | |
e876d531 | 3553 | **hint = and_predicates (info->conds, *hint, &p); |
3716ee8f | 3554 | } |
3555 | } | |
6a18c0be | 3556 | |
a41f2a28 | 3557 | /* We inlined EDGE. Update summary of the function we inlined into. */ |
3558 | ||
3559 | void | |
3560 | inline_merge_summary (struct cgraph_edge *edge) | |
3561 | { | |
b4bae7a0 | 3562 | struct inline_summary *callee_info = inline_summaries->get (edge->callee); |
a41f2a28 | 3563 | struct cgraph_node *to = (edge->caller->global.inlined_to |
3564 | ? edge->caller->global.inlined_to : edge->caller); | |
b4bae7a0 | 3565 | struct inline_summary *info = inline_summaries->get (to); |
a41f2a28 | 3566 | clause_t clause = 0; /* not_inline is known to be false. */ |
3567 | size_time_entry *e; | |
1e094109 | 3568 | vec<int> operand_map = vNULL; |
3569 | vec<int> offset_map = vNULL; | |
a41f2a28 | 3570 | int i; |
6a18c0be | 3571 | struct predicate toplev_predicate; |
a226c368 | 3572 | struct predicate true_p = true_predicate (); |
6a18c0be | 3573 | struct inline_edge_summary *es = inline_edge_summary (edge); |
3574 | ||
3575 | if (es->predicate) | |
3576 | toplev_predicate = *es->predicate; | |
3577 | else | |
3578 | toplev_predicate = true_predicate (); | |
a41f2a28 | 3579 | |
f4e523eb | 3580 | if (callee_info->conds) |
3581 | evaluate_properties_for_edge (edge, true, &clause, NULL, NULL, NULL); | |
2cc80ac3 | 3582 | if (ipa_node_params_sum && callee_info->conds) |
a41f2a28 | 3583 | { |
3584 | struct ipa_edge_args *args = IPA_EDGE_REF (edge); | |
3585 | int count = ipa_get_cs_argument_count (args); | |
3586 | int i; | |
3587 | ||
a226c368 | 3588 | if (count) |
a4f60e55 | 3589 | { |
f1f41a6c | 3590 | operand_map.safe_grow_cleared (count); |
3591 | offset_map.safe_grow_cleared (count); | |
a4f60e55 | 3592 | } |
a41f2a28 | 3593 | for (i = 0; i < count; i++) |
3594 | { | |
3595 | struct ipa_jump_func *jfunc = ipa_get_ith_jump_func (args, i); | |
3596 | int map = -1; | |
a4f60e55 | 3597 | |
a41f2a28 | 3598 | /* TODO: handle non-NOPs when merging. */ |
a4f60e55 | 3599 | if (jfunc->type == IPA_JF_PASS_THROUGH) |
3600 | { | |
3601 | if (ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR) | |
3602 | map = ipa_get_jf_pass_through_formal_id (jfunc); | |
3603 | if (!ipa_get_jf_pass_through_agg_preserved (jfunc)) | |
f1f41a6c | 3604 | offset_map[i] = -1; |
a4f60e55 | 3605 | } |
3606 | else if (jfunc->type == IPA_JF_ANCESTOR) | |
3607 | { | |
3608 | HOST_WIDE_INT offset = ipa_get_jf_ancestor_offset (jfunc); | |
3609 | if (offset >= 0 && offset < INT_MAX) | |
3610 | { | |
3611 | map = ipa_get_jf_ancestor_formal_id (jfunc); | |
3612 | if (!ipa_get_jf_ancestor_agg_preserved (jfunc)) | |
3613 | offset = -1; | |
f1f41a6c | 3614 | offset_map[i] = offset; |
a4f60e55 | 3615 | } |
3616 | } | |
f1f41a6c | 3617 | operand_map[i] = map; |
5cb1b112 | 3618 | gcc_assert (map < ipa_get_param_count (IPA_NODE_REF (to))); |
a41f2a28 | 3619 | } |
3620 | } | |
f1f41a6c | 3621 | for (i = 0; vec_safe_iterate (callee_info->entry, i, &e); i++) |
a41f2a28 | 3622 | { |
3623 | struct predicate p = remap_predicate (info, callee_info, | |
a4f60e55 | 3624 | &e->predicate, operand_map, |
3625 | offset_map, clause, | |
6a18c0be | 3626 | &toplev_predicate); |
eb4ae064 | 3627 | if (!false_predicate_p (&p)) |
3628 | { | |
e876d531 | 3629 | gcov_type add_time = ((gcov_type) e->time * edge->frequency |
eb4ae064 | 3630 | + CGRAPH_FREQ_BASE / 2) / CGRAPH_FREQ_BASE; |
3631 | int prob = predicate_probability (callee_info->conds, | |
3632 | &e->predicate, | |
3633 | clause, es->param); | |
70074000 | 3634 | add_time = apply_probability ((gcov_type) add_time, prob); |
eb4ae064 | 3635 | if (add_time > MAX_TIME * INLINE_TIME_SCALE) |
3636 | add_time = MAX_TIME * INLINE_TIME_SCALE; | |
3637 | if (prob != REG_BR_PROB_BASE | |
3638 | && dump_file && (dump_flags & TDF_DETAILS)) | |
3639 | { | |
3640 | fprintf (dump_file, "\t\tScaling time by probability:%f\n", | |
e876d531 | 3641 | (double) prob / REG_BR_PROB_BASE); |
eb4ae064 | 3642 | } |
3643 | account_size_time (info, e->size, add_time, &p); | |
3644 | } | |
3645 | } | |
3646 | remap_edge_summaries (edge, edge->callee, info, callee_info, operand_map, | |
a4f60e55 | 3647 | offset_map, clause, &toplev_predicate); |
3716ee8f | 3648 | remap_hint_predicate (info, callee_info, |
3649 | &callee_info->loop_iterations, | |
e876d531 | 3650 | operand_map, offset_map, clause, &toplev_predicate); |
3716ee8f | 3651 | remap_hint_predicate (info, callee_info, |
3652 | &callee_info->loop_stride, | |
e876d531 | 3653 | operand_map, offset_map, clause, &toplev_predicate); |
be343a9c | 3654 | remap_hint_predicate (info, callee_info, |
3655 | &callee_info->array_index, | |
e876d531 | 3656 | operand_map, offset_map, clause, &toplev_predicate); |
0835ad03 | 3657 | |
3658 | inline_update_callee_summaries (edge->callee, | |
3659 | inline_edge_summary (edge)->loop_depth); | |
3660 | ||
a226c368 | 3661 | /* We do not maintain predicates of inlined edges, free it. */ |
3662 | edge_set_predicate (edge, &true_p); | |
eb4ae064 | 3663 | /* Similarly remove param summaries. */ |
f1f41a6c | 3664 | es->param.release (); |
3665 | operand_map.release (); | |
3666 | offset_map.release (); | |
6331b6fa | 3667 | } |
3668 | ||
3669 | /* For performance reasons inline_merge_summary is not updating overall size | |
3670 | and time. Recompute it. */ | |
a226c368 | 3671 | |
6331b6fa | 3672 | void |
3673 | inline_update_overall_summary (struct cgraph_node *node) | |
3674 | { | |
b4bae7a0 | 3675 | struct inline_summary *info = inline_summaries->get (node); |
6331b6fa | 3676 | size_time_entry *e; |
3677 | int i; | |
3678 | ||
3679 | info->size = 0; | |
3680 | info->time = 0; | |
f1f41a6c | 3681 | for (i = 0; vec_safe_iterate (info->entry, i, &e); i++) |
3c49142d | 3682 | { |
3683 | info->size += e->size, info->time += e->time; | |
3684 | if (info->time > MAX_TIME * INLINE_TIME_SCALE) | |
e876d531 | 3685 | info->time = MAX_TIME * INLINE_TIME_SCALE; |
3c49142d | 3686 | } |
db197f90 | 3687 | estimate_calls_size_and_time (node, &info->size, &info->min_size, |
3688 | &info->time, NULL, | |
e876d531 | 3689 | ~(clause_t) (1 << predicate_false_condition), |
1e094109 | 3690 | vNULL, vNULL, vNULL); |
a41f2a28 | 3691 | info->time = (info->time + INLINE_TIME_SCALE / 2) / INLINE_TIME_SCALE; |
3692 | info->size = (info->size + INLINE_SIZE_SCALE / 2) / INLINE_SIZE_SCALE; | |
3693 | } | |
3694 | ||
3172b7bf | 3695 | /* Return hints derrived from EDGE. */ |
3696 | int | |
3697 | simple_edge_hints (struct cgraph_edge *edge) | |
3698 | { | |
3699 | int hints = 0; | |
3700 | struct cgraph_node *to = (edge->caller->global.inlined_to | |
e876d531 | 3701 | ? edge->caller->global.inlined_to : edge->caller); |
57e20c4a | 3702 | struct cgraph_node *callee = edge->callee->ultimate_alias_target (); |
b4bae7a0 | 3703 | if (inline_summaries->get (to)->scc_no |
57e20c4a | 3704 | && inline_summaries->get (to)->scc_no |
3705 | == inline_summaries->get (callee)->scc_no | |
35ee1c66 | 3706 | && !edge->recursive_p ()) |
3172b7bf | 3707 | hints |= INLINE_HINT_same_scc; |
3708 | ||
57e20c4a | 3709 | if (callee->lto_file_data && edge->caller->lto_file_data |
3710 | && edge->caller->lto_file_data != callee->lto_file_data | |
7d38b7bc | 3711 | && !callee->merged_comdat && !callee->icf_merged) |
3172b7bf | 3712 | hints |= INLINE_HINT_cross_module; |
3713 | ||
3714 | return hints; | |
3715 | } | |
3716 | ||
a41f2a28 | 3717 | /* Estimate the time cost for the caller when inlining EDGE. |
3718 | Only to be called via estimate_edge_time, that handles the | |
3719 | caching mechanism. | |
3720 | ||
3721 | When caching, also update the cache entry. Compute both time and | |
3722 | size, since we always need both metrics eventually. */ | |
3723 | ||
3724 | int | |
3725 | do_estimate_edge_time (struct cgraph_edge *edge) | |
3726 | { | |
3727 | int time; | |
3728 | int size; | |
eb7c606e | 3729 | inline_hints hints; |
20da2013 | 3730 | struct cgraph_node *callee; |
3731 | clause_t clause; | |
f1f41a6c | 3732 | vec<tree> known_vals; |
245ab191 | 3733 | vec<ipa_polymorphic_call_context> known_contexts; |
f1f41a6c | 3734 | vec<ipa_agg_jump_function_p> known_aggs; |
0835ad03 | 3735 | struct inline_edge_summary *es = inline_edge_summary (edge); |
db197f90 | 3736 | int min_size; |
a41f2a28 | 3737 | |
415d1b9a | 3738 | callee = edge->callee->ultimate_alias_target (); |
20da2013 | 3739 | |
a41f2a28 | 3740 | gcc_checking_assert (edge->inline_failed); |
20da2013 | 3741 | evaluate_properties_for_edge (edge, true, |
245ab191 | 3742 | &clause, &known_vals, &known_contexts, |
a4f60e55 | 3743 | &known_aggs); |
245ab191 | 3744 | estimate_node_size_and_time (callee, clause, known_vals, known_contexts, |
db197f90 | 3745 | known_aggs, &size, &min_size, &time, &hints, es->param); |
3072aa32 | 3746 | |
3747 | /* When we have profile feedback, we can quite safely identify hot | |
3748 | edges and for those we disable size limits. Don't do that when | |
3749 | probability that caller will call the callee is low however, since it | |
3750 | may hurt optimization of the caller's hot path. */ | |
35ee1c66 | 3751 | if (edge->count && edge->maybe_hot_p () |
3072aa32 | 3752 | && (edge->count * 2 |
3753 | > (edge->caller->global.inlined_to | |
3754 | ? edge->caller->global.inlined_to->count : edge->caller->count))) | |
3755 | hints |= INLINE_HINT_known_hot; | |
3756 | ||
f1f41a6c | 3757 | known_vals.release (); |
245ab191 | 3758 | known_contexts.release (); |
f1f41a6c | 3759 | known_aggs.release (); |
3172b7bf | 3760 | gcc_checking_assert (size >= 0); |
3761 | gcc_checking_assert (time >= 0); | |
a41f2a28 | 3762 | |
3763 | /* When caching, update the cache entry. */ | |
f1f41a6c | 3764 | if (edge_growth_cache.exists ()) |
a41f2a28 | 3765 | { |
b4bae7a0 | 3766 | inline_summaries->get (edge->callee)->min_size = min_size; |
e876d531 | 3767 | if ((int) edge_growth_cache.length () <= edge->uid) |
35ee1c66 | 3768 | edge_growth_cache.safe_grow_cleared (symtab->edges_max_uid); |
f1f41a6c | 3769 | edge_growth_cache[edge->uid].time = time + (time >= 0); |
a41f2a28 | 3770 | |
f1f41a6c | 3771 | edge_growth_cache[edge->uid].size = size + (size >= 0); |
3172b7bf | 3772 | hints |= simple_edge_hints (edge); |
f1f41a6c | 3773 | edge_growth_cache[edge->uid].hints = hints + 1; |
a41f2a28 | 3774 | } |
3172b7bf | 3775 | return time; |
a41f2a28 | 3776 | } |
3777 | ||
3778 | ||
6c2c7775 | 3779 | /* Return estimated callee growth after inlining EDGE. |
a41f2a28 | 3780 | Only to be called via estimate_edge_size. */ |
3781 | ||
3782 | int | |
6c2c7775 | 3783 | do_estimate_edge_size (struct cgraph_edge *edge) |
a41f2a28 | 3784 | { |
3785 | int size; | |
82626cb0 | 3786 | struct cgraph_node *callee; |
20da2013 | 3787 | clause_t clause; |
f1f41a6c | 3788 | vec<tree> known_vals; |
245ab191 | 3789 | vec<ipa_polymorphic_call_context> known_contexts; |
f1f41a6c | 3790 | vec<ipa_agg_jump_function_p> known_aggs; |
a41f2a28 | 3791 | |
3792 | /* When we do caching, use do_estimate_edge_time to populate the entry. */ | |
3793 | ||
f1f41a6c | 3794 | if (edge_growth_cache.exists ()) |
a41f2a28 | 3795 | { |
3796 | do_estimate_edge_time (edge); | |
f1f41a6c | 3797 | size = edge_growth_cache[edge->uid].size; |
a41f2a28 | 3798 | gcc_checking_assert (size); |
3799 | return size - (size > 0); | |
3800 | } | |
20da2013 | 3801 | |
415d1b9a | 3802 | callee = edge->callee->ultimate_alias_target (); |
a41f2a28 | 3803 | |
3804 | /* Early inliner runs without caching, go ahead and do the dirty work. */ | |
3805 | gcc_checking_assert (edge->inline_failed); | |
20da2013 | 3806 | evaluate_properties_for_edge (edge, true, |
245ab191 | 3807 | &clause, &known_vals, &known_contexts, |
a4f60e55 | 3808 | &known_aggs); |
245ab191 | 3809 | estimate_node_size_and_time (callee, clause, known_vals, known_contexts, |
db197f90 | 3810 | known_aggs, &size, NULL, NULL, NULL, vNULL); |
f1f41a6c | 3811 | known_vals.release (); |
245ab191 | 3812 | known_contexts.release (); |
f1f41a6c | 3813 | known_aggs.release (); |
6c2c7775 | 3814 | return size; |
99c67f24 | 3815 | } |
3816 | ||
3817 | ||
eb7c606e | 3818 | /* Estimate the growth of the caller when inlining EDGE. |
3819 | Only to be called via estimate_edge_size. */ | |
3820 | ||
3821 | inline_hints | |
3822 | do_estimate_edge_hints (struct cgraph_edge *edge) | |
3823 | { | |
3824 | inline_hints hints; | |
3825 | struct cgraph_node *callee; | |
3826 | clause_t clause; | |
f1f41a6c | 3827 | vec<tree> known_vals; |
245ab191 | 3828 | vec<ipa_polymorphic_call_context> known_contexts; |
f1f41a6c | 3829 | vec<ipa_agg_jump_function_p> known_aggs; |
eb7c606e | 3830 | |
3831 | /* When we do caching, use do_estimate_edge_time to populate the entry. */ | |
3832 | ||
f1f41a6c | 3833 | if (edge_growth_cache.exists ()) |
eb7c606e | 3834 | { |
3835 | do_estimate_edge_time (edge); | |
f1f41a6c | 3836 | hints = edge_growth_cache[edge->uid].hints; |
eb7c606e | 3837 | gcc_checking_assert (hints); |
3838 | return hints - 1; | |
3839 | } | |
3840 | ||
415d1b9a | 3841 | callee = edge->callee->ultimate_alias_target (); |
eb7c606e | 3842 | |
3843 | /* Early inliner runs without caching, go ahead and do the dirty work. */ | |
3844 | gcc_checking_assert (edge->inline_failed); | |
3845 | evaluate_properties_for_edge (edge, true, | |
245ab191 | 3846 | &clause, &known_vals, &known_contexts, |
eb7c606e | 3847 | &known_aggs); |
245ab191 | 3848 | estimate_node_size_and_time (callee, clause, known_vals, known_contexts, |
db197f90 | 3849 | known_aggs, NULL, NULL, NULL, &hints, vNULL); |
f1f41a6c | 3850 | known_vals.release (); |
245ab191 | 3851 | known_contexts.release (); |
f1f41a6c | 3852 | known_aggs.release (); |
3172b7bf | 3853 | hints |= simple_edge_hints (edge); |
eb7c606e | 3854 | return hints; |
3855 | } | |
3856 | ||
3857 | ||
99c67f24 | 3858 | /* Estimate self time of the function NODE after inlining EDGE. */ |
3859 | ||
3860 | int | |
3861 | estimate_time_after_inlining (struct cgraph_node *node, | |
3862 | struct cgraph_edge *edge) | |
3863 | { | |
905aa3bd | 3864 | struct inline_edge_summary *es = inline_edge_summary (edge); |
3865 | if (!es->predicate || !false_predicate_p (es->predicate)) | |
3866 | { | |
e876d531 | 3867 | gcov_type time = |
b4bae7a0 | 3868 | inline_summaries->get (node)->time + estimate_edge_time (edge); |
905aa3bd | 3869 | if (time < 0) |
3870 | time = 0; | |
3871 | if (time > MAX_TIME) | |
3872 | time = MAX_TIME; | |
3873 | return time; | |
3874 | } | |
b4bae7a0 | 3875 | return inline_summaries->get (node)->time; |
99c67f24 | 3876 | } |
3877 | ||
3878 | ||
3879 | /* Estimate the size of NODE after inlining EDGE which should be an | |
3880 | edge to either NODE or a call inlined into NODE. */ | |
3881 | ||
3882 | int | |
3883 | estimate_size_after_inlining (struct cgraph_node *node, | |
c7b2cc59 | 3884 | struct cgraph_edge *edge) |
99c67f24 | 3885 | { |
905aa3bd | 3886 | struct inline_edge_summary *es = inline_edge_summary (edge); |
3887 | if (!es->predicate || !false_predicate_p (es->predicate)) | |
3888 | { | |
b4bae7a0 | 3889 | int size = inline_summaries->get (node)->size + estimate_edge_growth (edge); |
905aa3bd | 3890 | gcc_assert (size >= 0); |
3891 | return size; | |
3892 | } | |
b4bae7a0 | 3893 | return inline_summaries->get (node)->size; |
99c67f24 | 3894 | } |
3895 | ||
3896 | ||
82626cb0 | 3897 | struct growth_data |
3898 | { | |
ce1f57d6 | 3899 | struct cgraph_node *node; |
82626cb0 | 3900 | bool self_recursive; |
6a044047 | 3901 | bool uninlinable; |
82626cb0 | 3902 | int growth; |
3903 | }; | |
99c67f24 | 3904 | |
82626cb0 | 3905 | |
3906 | /* Worker for do_estimate_growth. Collect growth for all callers. */ | |
3907 | ||
3908 | static bool | |
3909 | do_estimate_growth_1 (struct cgraph_node *node, void *data) | |
99c67f24 | 3910 | { |
99c67f24 | 3911 | struct cgraph_edge *e; |
82626cb0 | 3912 | struct growth_data *d = (struct growth_data *) data; |
99c67f24 | 3913 | |
99c67f24 | 3914 | for (e = node->callers; e; e = e->next_caller) |
3915 | { | |
4869c23f | 3916 | gcc_checking_assert (e->inline_failed); |
3917 | ||
6a044047 | 3918 | if (cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR) |
3919 | { | |
3920 | d->uninlinable = true; | |
3921 | continue; | |
3922 | } | |
3923 | ||
9dcc8702 | 3924 | if (e->recursive_p ()) |
3925 | { | |
3926 | d->self_recursive = true; | |
3927 | continue; | |
3928 | } | |
82626cb0 | 3929 | d->growth += estimate_edge_growth (e); |
4869c23f | 3930 | } |
82626cb0 | 3931 | return false; |
3932 | } | |
3933 | ||
3934 | ||
3935 | /* Estimate the growth caused by inlining NODE into all callees. */ | |
3936 | ||
3937 | int | |
bc42c20c | 3938 | estimate_growth (struct cgraph_node *node) |
82626cb0 | 3939 | { |
6a044047 | 3940 | struct growth_data d = { node, false, false, 0 }; |
b4bae7a0 | 3941 | struct inline_summary *info = inline_summaries->get (node); |
82626cb0 | 3942 | |
6a044047 | 3943 | node->call_for_symbol_and_aliases (do_estimate_growth_1, &d, true); |
4869c23f | 3944 | |
3945 | /* For self recursive functions the growth estimation really should be | |
3946 | infinity. We don't want to return very large values because the growth | |
3947 | plays various roles in badness computation fractions. Be sure to not | |
3948 | return zero or negative growths. */ | |
82626cb0 | 3949 | if (d.self_recursive) |
3950 | d.growth = d.growth < info->size ? info->size : d.growth; | |
6a044047 | 3951 | else if (DECL_EXTERNAL (node->decl) || d.uninlinable) |
3172b7bf | 3952 | ; |
4869c23f | 3953 | else |
3954 | { | |
415d1b9a | 3955 | if (node->will_be_removed_from_program_if_no_direct_calls_p ()) |
82626cb0 | 3956 | d.growth -= info->size; |
fb3c587e | 3957 | /* COMDAT functions are very often not shared across multiple units |
e876d531 | 3958 | since they come from various template instantiations. |
3959 | Take this into account. */ | |
aca3df3b | 3960 | else if (DECL_COMDAT (node->decl) |
415d1b9a | 3961 | && node->can_remove_if_no_direct_calls_p ()) |
82626cb0 | 3962 | d.growth -= (info->size |
fb3c587e | 3963 | * (100 - PARAM_VALUE (PARAM_COMDAT_SHARING_PROBABILITY)) |
3964 | + 50) / 100; | |
99c67f24 | 3965 | } |
99c67f24 | 3966 | |
82626cb0 | 3967 | return d.growth; |
99c67f24 | 3968 | } |
3969 | ||
6a044047 | 3970 | /* Verify if there are fewer than MAX_CALLERS. */ |
3971 | ||
3972 | static bool | |
3973 | check_callers (cgraph_node *node, int *max_callers) | |
3974 | { | |
3975 | ipa_ref *ref; | |
3976 | ||
b12b920e | 3977 | if (!node->can_remove_if_no_direct_calls_and_refs_p ()) |
3978 | return true; | |
3979 | ||
6a044047 | 3980 | for (cgraph_edge *e = node->callers; e; e = e->next_caller) |
3981 | { | |
3982 | (*max_callers)--; | |
3983 | if (!*max_callers | |
3984 | || cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR) | |
3985 | return true; | |
3986 | } | |
3987 | ||
3988 | FOR_EACH_ALIAS (node, ref) | |
3989 | if (check_callers (dyn_cast <cgraph_node *> (ref->referring), max_callers)) | |
3990 | return true; | |
3991 | ||
3992 | return false; | |
3993 | } | |
3994 | ||
c7b2cc59 | 3995 | |
db197f90 | 3996 | /* Make cheap estimation if growth of NODE is likely positive knowing |
3997 | EDGE_GROWTH of one particular edge. | |
3998 | We assume that most of other edges will have similar growth | |
3999 | and skip computation if there are too many callers. */ | |
4000 | ||
4001 | bool | |
6a044047 | 4002 | growth_likely_positive (struct cgraph_node *node, |
4003 | int edge_growth) | |
db197f90 | 4004 | { |
4005 | int max_callers; | |
db197f90 | 4006 | struct cgraph_edge *e; |
4007 | gcc_checking_assert (edge_growth > 0); | |
4008 | ||
b12b920e | 4009 | /* First quickly check if NODE is removable at all. */ |
e88fecaf | 4010 | if (DECL_EXTERNAL (node->decl)) |
4011 | return true; | |
b12b920e | 4012 | if (!node->can_remove_if_no_direct_calls_and_refs_p () |
4013 | || node->address_taken) | |
db197f90 | 4014 | return true; |
b12b920e | 4015 | |
b4bae7a0 | 4016 | max_callers = inline_summaries->get (node)->size * 4 / edge_growth + 2; |
db197f90 | 4017 | |
4018 | for (e = node->callers; e; e = e->next_caller) | |
4019 | { | |
4020 | max_callers--; | |
6a044047 | 4021 | if (!max_callers |
4022 | || cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR) | |
db197f90 | 4023 | return true; |
4024 | } | |
6a044047 | 4025 | |
4026 | ipa_ref *ref; | |
4027 | FOR_EACH_ALIAS (node, ref) | |
4028 | if (check_callers (dyn_cast <cgraph_node *> (ref->referring), &max_callers)) | |
4029 | return true; | |
4030 | ||
b12b920e | 4031 | /* Unlike for functions called once, we play unsafe with |
4032 | COMDATs. We can allow that since we know functions | |
4033 | in consideration are small (and thus risk is small) and | |
4034 | moreover grow estimates already accounts that COMDAT | |
4035 | functions may or may not disappear when eliminated from | |
4036 | current unit. With good probability making aggressive | |
4037 | choice in all units is going to make overall program | |
4038 | smaller. */ | |
4039 | if (DECL_COMDAT (node->decl)) | |
4040 | { | |
4041 | if (!node->can_remove_if_no_direct_calls_p ()) | |
4042 | return true; | |
4043 | } | |
4044 | else if (!node->will_be_removed_from_program_if_no_direct_calls_p ()) | |
4045 | return true; | |
4046 | ||
db197f90 | 4047 | return estimate_growth (node) > 0; |
4048 | } | |
4049 | ||
4050 | ||
99c67f24 | 4051 | /* This function performs intraprocedural analysis in NODE that is required to |
4052 | inline indirect calls. */ | |
c7b2cc59 | 4053 | |
99c67f24 | 4054 | static void |
4055 | inline_indirect_intraprocedural_analysis (struct cgraph_node *node) | |
4056 | { | |
4057 | ipa_analyze_node (node); | |
4058 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
4059 | { | |
4060 | ipa_print_node_params (dump_file, node); | |
4061 | ipa_print_node_jump_functions (dump_file, node); | |
4062 | } | |
4063 | } | |
4064 | ||
4065 | ||
4066 | /* Note function body size. */ | |
4067 | ||
cb8994e9 | 4068 | void |
99c67f24 | 4069 | inline_analyze_function (struct cgraph_node *node) |
4070 | { | |
02774f2d | 4071 | push_cfun (DECL_STRUCT_FUNCTION (node->decl)); |
99c67f24 | 4072 | |
a41f2a28 | 4073 | if (dump_file) |
4074 | fprintf (dump_file, "\nAnalyzing function: %s/%u\n", | |
f1c8b4d7 | 4075 | node->name (), node->order); |
d1f68cd8 | 4076 | if (opt_for_fn (node->decl, optimize) && !node->thunk.thunk_p) |
99c67f24 | 4077 | inline_indirect_intraprocedural_analysis (node); |
a41f2a28 | 4078 | compute_inline_parameters (node, false); |
26051fcf | 4079 | if (!optimize) |
4080 | { | |
4081 | struct cgraph_edge *e; | |
4082 | for (e = node->callees; e; e = e->next_callee) | |
4083 | { | |
4084 | if (e->inline_failed == CIF_FUNCTION_NOT_CONSIDERED) | |
4085 | e->inline_failed = CIF_FUNCTION_NOT_OPTIMIZED; | |
4086 | e->call_stmt_cannot_inline_p = true; | |
4087 | } | |
4088 | for (e = node->indirect_calls; e; e = e->next_callee) | |
4089 | { | |
4090 | if (e->inline_failed == CIF_FUNCTION_NOT_CONSIDERED) | |
4091 | e->inline_failed = CIF_FUNCTION_NOT_OPTIMIZED; | |
4092 | e->call_stmt_cannot_inline_p = true; | |
4093 | } | |
4094 | } | |
99c67f24 | 4095 | |
99c67f24 | 4096 | pop_cfun (); |
4097 | } | |
4098 | ||
4099 | ||
4100 | /* Called when new function is inserted to callgraph late. */ | |
4101 | ||
b4bae7a0 | 4102 | void |
4103 | inline_summary_t::insert (struct cgraph_node *node, inline_summary *) | |
99c67f24 | 4104 | { |
4105 | inline_analyze_function (node); | |
4106 | } | |
4107 | ||
99c67f24 | 4108 | /* Note function body size. */ |
4109 | ||
4110 | void | |
4111 | inline_generate_summary (void) | |
4112 | { | |
4113 | struct cgraph_node *node; | |
4114 | ||
b8681788 | 4115 | FOR_EACH_DEFINED_FUNCTION (node) |
4116 | if (DECL_STRUCT_FUNCTION (node->decl)) | |
4117 | node->local.versionable = tree_versionable_function_p (node->decl); | |
4118 | ||
2fe870c5 | 4119 | /* When not optimizing, do not bother to analyze. Inlining is still done |
4120 | because edge redirection needs to happen there. */ | |
9f28dc4c | 4121 | if (!optimize && !flag_generate_lto && !flag_generate_offload && !flag_wpa) |
2fe870c5 | 4122 | return; |
4123 | ||
b4bae7a0 | 4124 | if (!inline_summaries) |
4125 | inline_summaries = (inline_summary_t*) inline_summary_t::create_ggc (symtab); | |
4126 | ||
4127 | inline_summaries->enable_insertion_hook (); | |
99c67f24 | 4128 | |
a226c368 | 4129 | ipa_register_cgraph_hooks (); |
3b9dd281 | 4130 | inline_free_summary (); |
99c67f24 | 4131 | |
91bf9d9a | 4132 | FOR_EACH_DEFINED_FUNCTION (node) |
02774f2d | 4133 | if (!node->alias) |
99c67f24 | 4134 | inline_analyze_function (node); |
99c67f24 | 4135 | } |
4136 | ||
4137 | ||
6a18c0be | 4138 | /* Read predicate from IB. */ |
4139 | ||
4140 | static struct predicate | |
4141 | read_predicate (struct lto_input_block *ib) | |
4142 | { | |
4143 | struct predicate out; | |
4144 | clause_t clause; | |
4145 | int k = 0; | |
4146 | ||
e876d531 | 4147 | do |
6a18c0be | 4148 | { |
905aa3bd | 4149 | gcc_assert (k <= MAX_CLAUSES); |
7f385784 | 4150 | clause = out.clause[k++] = streamer_read_uhwi (ib); |
6a18c0be | 4151 | } |
4152 | while (clause); | |
72cb6720 | 4153 | |
4154 | /* Zero-initialize the remaining clauses in OUT. */ | |
4155 | while (k <= MAX_CLAUSES) | |
4156 | out.clause[k++] = 0; | |
4157 | ||
6a18c0be | 4158 | return out; |
4159 | } | |
4160 | ||
4161 | ||
0835ad03 | 4162 | /* Write inline summary for edge E to OB. */ |
4163 | ||
4164 | static void | |
4165 | read_inline_edge_summary (struct lto_input_block *ib, struct cgraph_edge *e) | |
4166 | { | |
4167 | struct inline_edge_summary *es = inline_edge_summary (e); | |
6a18c0be | 4168 | struct predicate p; |
eb4ae064 | 4169 | int length, i; |
6a18c0be | 4170 | |
7f385784 | 4171 | es->call_stmt_size = streamer_read_uhwi (ib); |
4172 | es->call_stmt_time = streamer_read_uhwi (ib); | |
4173 | es->loop_depth = streamer_read_uhwi (ib); | |
6a18c0be | 4174 | p = read_predicate (ib); |
4175 | edge_set_predicate (e, &p); | |
eb4ae064 | 4176 | length = streamer_read_uhwi (ib); |
4177 | if (length) | |
4178 | { | |
f1f41a6c | 4179 | es->param.safe_grow_cleared (length); |
eb4ae064 | 4180 | for (i = 0; i < length; i++) |
e876d531 | 4181 | es->param[i].change_prob = streamer_read_uhwi (ib); |
eb4ae064 | 4182 | } |
0835ad03 | 4183 | } |
4184 | ||
4185 | ||
a41f2a28 | 4186 | /* Stream in inline summaries from the section. */ |
4187 | ||
4188 | static void | |
4189 | inline_read_section (struct lto_file_decl_data *file_data, const char *data, | |
4190 | size_t len) | |
4191 | { | |
4192 | const struct lto_function_header *header = | |
4193 | (const struct lto_function_header *) data; | |
949e5786 | 4194 | const int cfg_offset = sizeof (struct lto_function_header); |
4195 | const int main_offset = cfg_offset + header->cfg_size; | |
4196 | const int string_offset = main_offset + header->main_size; | |
a41f2a28 | 4197 | struct data_in *data_in; |
a41f2a28 | 4198 | unsigned int i, count2, j; |
4199 | unsigned int f_count; | |
4200 | ||
2e971afd | 4201 | lto_input_block ib ((const char *) data + main_offset, header->main_size, |
4202 | file_data->mode_table); | |
a41f2a28 | 4203 | |
4204 | data_in = | |
4205 | lto_data_in_create (file_data, (const char *) data + string_offset, | |
1e094109 | 4206 | header->string_size, vNULL); |
7f385784 | 4207 | f_count = streamer_read_uhwi (&ib); |
a41f2a28 | 4208 | for (i = 0; i < f_count; i++) |
4209 | { | |
4210 | unsigned int index; | |
4211 | struct cgraph_node *node; | |
4212 | struct inline_summary *info; | |
70225339 | 4213 | lto_symtab_encoder_t encoder; |
a41f2a28 | 4214 | struct bitpack_d bp; |
0835ad03 | 4215 | struct cgraph_edge *e; |
7c07aa3d | 4216 | predicate p; |
a41f2a28 | 4217 | |
7f385784 | 4218 | index = streamer_read_uhwi (&ib); |
70225339 | 4219 | encoder = file_data->symtab_node_encoder; |
415d1b9a | 4220 | node = dyn_cast<cgraph_node *> (lto_symtab_encoder_deref (encoder, |
4221 | index)); | |
b4bae7a0 | 4222 | info = inline_summaries->get (node); |
a41f2a28 | 4223 | |
4224 | info->estimated_stack_size | |
7f385784 | 4225 | = info->estimated_self_stack_size = streamer_read_uhwi (&ib); |
4226 | info->size = info->self_size = streamer_read_uhwi (&ib); | |
4227 | info->time = info->self_time = streamer_read_uhwi (&ib); | |
a41f2a28 | 4228 | |
7f385784 | 4229 | bp = streamer_read_bitpack (&ib); |
a41f2a28 | 4230 | info->inlinable = bp_unpack_value (&bp, 1); |
e806c56f | 4231 | info->contains_cilk_spawn = bp_unpack_value (&bp, 1); |
a41f2a28 | 4232 | |
7f385784 | 4233 | count2 = streamer_read_uhwi (&ib); |
a41f2a28 | 4234 | gcc_assert (!info->conds); |
4235 | for (j = 0; j < count2; j++) | |
4236 | { | |
4237 | struct condition c; | |
7f385784 | 4238 | c.operand_num = streamer_read_uhwi (&ib); |
4239 | c.code = (enum tree_code) streamer_read_uhwi (&ib); | |
515cf651 | 4240 | c.val = stream_read_tree (&ib, data_in); |
a4f60e55 | 4241 | bp = streamer_read_bitpack (&ib); |
4242 | c.agg_contents = bp_unpack_value (&bp, 1); | |
4243 | c.by_ref = bp_unpack_value (&bp, 1); | |
4244 | if (c.agg_contents) | |
4245 | c.offset = streamer_read_uhwi (&ib); | |
f1f41a6c | 4246 | vec_safe_push (info->conds, c); |
a41f2a28 | 4247 | } |
7f385784 | 4248 | count2 = streamer_read_uhwi (&ib); |
a41f2a28 | 4249 | gcc_assert (!info->entry); |
4250 | for (j = 0; j < count2; j++) | |
4251 | { | |
4252 | struct size_time_entry e; | |
a41f2a28 | 4253 | |
7f385784 | 4254 | e.size = streamer_read_uhwi (&ib); |
4255 | e.time = streamer_read_uhwi (&ib); | |
6a18c0be | 4256 | e.predicate = read_predicate (&ib); |
a41f2a28 | 4257 | |
f1f41a6c | 4258 | vec_safe_push (info->entry, e); |
a41f2a28 | 4259 | } |
e876d531 | 4260 | |
7c07aa3d | 4261 | p = read_predicate (&ib); |
3716ee8f | 4262 | set_hint_predicate (&info->loop_iterations, p); |
4263 | p = read_predicate (&ib); | |
4264 | set_hint_predicate (&info->loop_stride, p); | |
be343a9c | 4265 | p = read_predicate (&ib); |
4266 | set_hint_predicate (&info->array_index, p); | |
0835ad03 | 4267 | for (e = node->callees; e; e = e->next_callee) |
4268 | read_inline_edge_summary (&ib, e); | |
4269 | for (e = node->indirect_calls; e; e = e->next_callee) | |
4270 | read_inline_edge_summary (&ib, e); | |
a41f2a28 | 4271 | } |
4272 | ||
4273 | lto_free_section_data (file_data, LTO_section_inline_summary, NULL, data, | |
4274 | len); | |
4275 | lto_data_in_delete (data_in); | |
4276 | } | |
4277 | ||
4278 | ||
99c67f24 | 4279 | /* Read inline summary. Jump functions are shared among ipa-cp |
4280 | and inliner, so when ipa-cp is active, we don't need to write them | |
4281 | twice. */ | |
4282 | ||
4283 | void | |
4284 | inline_read_summary (void) | |
4285 | { | |
c7b2cc59 | 4286 | struct lto_file_decl_data **file_data_vec = lto_get_file_decl_data (); |
4287 | struct lto_file_decl_data *file_data; | |
4288 | unsigned int j = 0; | |
4289 | ||
4290 | inline_summary_alloc (); | |
4291 | ||
4292 | while ((file_data = file_data_vec[j++])) | |
4293 | { | |
4294 | size_t len; | |
eb4ae064 | 4295 | const char *data = lto_get_section_data (file_data, |
4296 | LTO_section_inline_summary, | |
4297 | NULL, &len); | |
a41f2a28 | 4298 | if (data) |
e876d531 | 4299 | inline_read_section (file_data, data, len); |
c7b2cc59 | 4300 | else |
eb4ae064 | 4301 | /* Fatal error here. We do not want to support compiling ltrans units |
4302 | with different version of compiler or different flags than the WPA | |
4303 | unit, so this should never happen. */ | |
c05be867 | 4304 | fatal_error (input_location, |
4305 | "ipa inline summary is missing in input file"); | |
c7b2cc59 | 4306 | } |
a226c368 | 4307 | if (optimize) |
99c67f24 | 4308 | { |
4309 | ipa_register_cgraph_hooks (); | |
4310 | if (!flag_ipa_cp) | |
e876d531 | 4311 | ipa_prop_read_jump_functions (); |
99c67f24 | 4312 | } |
b4bae7a0 | 4313 | |
4314 | gcc_assert (inline_summaries); | |
4315 | inline_summaries->enable_insertion_hook (); | |
99c67f24 | 4316 | } |
4317 | ||
6a18c0be | 4318 | |
4319 | /* Write predicate P to OB. */ | |
4320 | ||
4321 | static void | |
4322 | write_predicate (struct output_block *ob, struct predicate *p) | |
4323 | { | |
4324 | int j; | |
4325 | if (p) | |
4326 | for (j = 0; p->clause[j]; j++) | |
4327 | { | |
e876d531 | 4328 | gcc_assert (j < MAX_CLAUSES); |
4329 | streamer_write_uhwi (ob, p->clause[j]); | |
6a18c0be | 4330 | } |
7f385784 | 4331 | streamer_write_uhwi (ob, 0); |
6a18c0be | 4332 | } |
4333 | ||
4334 | ||
0835ad03 | 4335 | /* Write inline summary for edge E to OB. */ |
4336 | ||
4337 | static void | |
4338 | write_inline_edge_summary (struct output_block *ob, struct cgraph_edge *e) | |
4339 | { | |
4340 | struct inline_edge_summary *es = inline_edge_summary (e); | |
eb4ae064 | 4341 | int i; |
4342 | ||
7f385784 | 4343 | streamer_write_uhwi (ob, es->call_stmt_size); |
4344 | streamer_write_uhwi (ob, es->call_stmt_time); | |
4345 | streamer_write_uhwi (ob, es->loop_depth); | |
6a18c0be | 4346 | write_predicate (ob, es->predicate); |
f1f41a6c | 4347 | streamer_write_uhwi (ob, es->param.length ()); |
e876d531 | 4348 | for (i = 0; i < (int) es->param.length (); i++) |
f1f41a6c | 4349 | streamer_write_uhwi (ob, es->param[i].change_prob); |
0835ad03 | 4350 | } |
4351 | ||
99c67f24 | 4352 | |
4353 | /* Write inline summary for node in SET. | |
4354 | Jump functions are shared among ipa-cp and inliner, so when ipa-cp is | |
4355 | active, we don't need to write them twice. */ | |
4356 | ||
4357 | void | |
eab36a5a | 4358 | inline_write_summary (void) |
99c67f24 | 4359 | { |
c7b2cc59 | 4360 | struct cgraph_node *node; |
a41f2a28 | 4361 | struct output_block *ob = create_output_block (LTO_section_inline_summary); |
70225339 | 4362 | lto_symtab_encoder_t encoder = ob->decl_state->symtab_node_encoder; |
c7b2cc59 | 4363 | unsigned int count = 0; |
4364 | int i; | |
4365 | ||
70225339 | 4366 | for (i = 0; i < lto_symtab_encoder_size (encoder); i++) |
2dc9831f | 4367 | { |
452659af | 4368 | symtab_node *snode = lto_symtab_encoder_deref (encoder, i); |
13cbeaac | 4369 | cgraph_node *cnode = dyn_cast <cgraph_node *> (snode); |
02774f2d | 4370 | if (cnode && cnode->definition && !cnode->alias) |
2dc9831f | 4371 | count++; |
4372 | } | |
7f385784 | 4373 | streamer_write_uhwi (ob, count); |
c7b2cc59 | 4374 | |
70225339 | 4375 | for (i = 0; i < lto_symtab_encoder_size (encoder); i++) |
c7b2cc59 | 4376 | { |
452659af | 4377 | symtab_node *snode = lto_symtab_encoder_deref (encoder, i); |
13cbeaac | 4378 | cgraph_node *cnode = dyn_cast <cgraph_node *> (snode); |
02774f2d | 4379 | if (cnode && (node = cnode)->definition && !node->alias) |
c7b2cc59 | 4380 | { |
b4bae7a0 | 4381 | struct inline_summary *info = inline_summaries->get (node); |
cbd7f5a0 | 4382 | struct bitpack_d bp; |
0835ad03 | 4383 | struct cgraph_edge *edge; |
a41f2a28 | 4384 | int i; |
4385 | size_time_entry *e; | |
4386 | struct condition *c; | |
e876d531 | 4387 | |
4388 | streamer_write_uhwi (ob, | |
4389 | lto_symtab_encoder_encode (encoder, | |
02774f2d | 4390 | |
e876d531 | 4391 | node)); |
7f385784 | 4392 | streamer_write_hwi (ob, info->estimated_self_stack_size); |
4393 | streamer_write_hwi (ob, info->self_size); | |
4394 | streamer_write_hwi (ob, info->self_time); | |
cbd7f5a0 | 4395 | bp = bitpack_create (ob->main_stream); |
4396 | bp_pack_value (&bp, info->inlinable, 1); | |
e806c56f | 4397 | bp_pack_value (&bp, info->contains_cilk_spawn, 1); |
7f385784 | 4398 | streamer_write_bitpack (&bp); |
f1f41a6c | 4399 | streamer_write_uhwi (ob, vec_safe_length (info->conds)); |
4400 | for (i = 0; vec_safe_iterate (info->conds, i, &c); i++) | |
a41f2a28 | 4401 | { |
7f385784 | 4402 | streamer_write_uhwi (ob, c->operand_num); |
4403 | streamer_write_uhwi (ob, c->code); | |
515cf651 | 4404 | stream_write_tree (ob, c->val, true); |
a4f60e55 | 4405 | bp = bitpack_create (ob->main_stream); |
4406 | bp_pack_value (&bp, c->agg_contents, 1); | |
4407 | bp_pack_value (&bp, c->by_ref, 1); | |
4408 | streamer_write_bitpack (&bp); | |
4409 | if (c->agg_contents) | |
e876d531 | 4410 | streamer_write_uhwi (ob, c->offset); |
a41f2a28 | 4411 | } |
f1f41a6c | 4412 | streamer_write_uhwi (ob, vec_safe_length (info->entry)); |
4413 | for (i = 0; vec_safe_iterate (info->entry, i, &e); i++) | |
a41f2a28 | 4414 | { |
7f385784 | 4415 | streamer_write_uhwi (ob, e->size); |
4416 | streamer_write_uhwi (ob, e->time); | |
6a18c0be | 4417 | write_predicate (ob, &e->predicate); |
a41f2a28 | 4418 | } |
7c07aa3d | 4419 | write_predicate (ob, info->loop_iterations); |
3716ee8f | 4420 | write_predicate (ob, info->loop_stride); |
be343a9c | 4421 | write_predicate (ob, info->array_index); |
0835ad03 | 4422 | for (edge = node->callees; edge; edge = edge->next_callee) |
4423 | write_inline_edge_summary (ob, edge); | |
4424 | for (edge = node->indirect_calls; edge; edge = edge->next_callee) | |
4425 | write_inline_edge_summary (ob, edge); | |
c7b2cc59 | 4426 | } |
4427 | } | |
7f385784 | 4428 | streamer_write_char_stream (ob->main_stream, 0); |
a41f2a28 | 4429 | produce_asm (ob, NULL); |
4430 | destroy_output_block (ob); | |
c7b2cc59 | 4431 | |
a226c368 | 4432 | if (optimize && !flag_ipa_cp) |
eab36a5a | 4433 | ipa_prop_write_jump_functions (); |
99c67f24 | 4434 | } |
4435 | ||
c7b2cc59 | 4436 | |
99c67f24 | 4437 | /* Release inline summary. */ |
4438 | ||
4439 | void | |
4440 | inline_free_summary (void) | |
4441 | { | |
3b9dd281 | 4442 | struct cgraph_node *node; |
0835ad03 | 4443 | if (edge_removal_hook_holder) |
35ee1c66 | 4444 | symtab->remove_edge_removal_hook (edge_removal_hook_holder); |
3b9dd281 | 4445 | edge_removal_hook_holder = NULL; |
0835ad03 | 4446 | if (edge_duplication_hook_holder) |
35ee1c66 | 4447 | symtab->remove_edge_duplication_hook (edge_duplication_hook_holder); |
3b9dd281 | 4448 | edge_duplication_hook_holder = NULL; |
b0c5e347 | 4449 | if (!inline_edge_summary_vec.exists ()) |
4450 | return; | |
4451 | FOR_EACH_DEFINED_FUNCTION (node) | |
4452 | if (!node->alias) | |
b4bae7a0 | 4453 | reset_inline_summary (node, inline_summaries->get (node)); |
4454 | inline_summaries->release (); | |
4455 | inline_summaries = NULL; | |
f1f41a6c | 4456 | inline_edge_summary_vec.release (); |
2a5261f7 | 4457 | edge_predicate_pool.release (); |
99c67f24 | 4458 | } |