]>
Commit | Line | Data |
---|---|---|
03dfc36d JH |
1 | /* Inlining decision heuristics. |
2 | Copyright (C) 2003, 2004, 2007, 2008, 2009, 2010, 2011 | |
3 | Free Software Foundation, Inc. | |
4 | Contributed by Jan Hubicka | |
5 | ||
6 | This file is part of GCC. | |
7 | ||
8 | GCC is free software; you can redistribute it and/or modify it under | |
9 | the terms of the GNU General Public License as published by the Free | |
10 | Software Foundation; either version 3, or (at your option) any later | |
11 | version. | |
12 | ||
13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
16 | for more details. | |
17 | ||
18 | You should have received a copy of the GNU General Public License | |
19 | along with GCC; see the file COPYING3. If not see | |
20 | <http://www.gnu.org/licenses/>. */ | |
21 | ||
22 | /* Analysis used by the inliner and other passes limiting code size growth. | |
23 | ||
24 | We estimate for each function | |
25 | - function body size | |
10a5dd5d | 26 | - average function execution time |
03dfc36d JH |
27 | - inlining size benefit (that is how much of function body size |
28 | and its call sequence is expected to disappear by inlining) | |
29 | - inlining time benefit | |
30 | - function frame size | |
31 | For each call | |
10a5dd5d | 32 | - call statement size and time |
03dfc36d JH |
33 | |
34 | inlinie_summary datastructures store above information locally (i.e. | |
35 | parameters of the function itself) and globally (i.e. parameters of | |
36 | the function created by applying all the inline decisions already | |
37 | present in the callgraph). | |
38 | ||
632b4f8e | 39 | We provide accestor to the inline_summary datastructure and |
03dfc36d JH |
40 | basic logic updating the parameters when inlining is performed. |
41 | ||
632b4f8e JH |
42 | The summaries are context sensitive. Context means |
43 | 1) partial assignment of known constant values of operands | |
44 | 2) whether function is inlined into the call or not. | |
45 | It is easy to add more variants. To represent function size and time | |
46 | that depends on context (i.e. it is known to be optimized away when | |
47 | context is known either by inlining or from IP-CP and clonning), | |
48 | we use predicates. Predicates are logical formulas in | |
49 | conjunctive-disjunctive form consisting of clauses. Clauses are bitmaps | |
50 | specifying what conditions must be true. Conditions are simple test | |
51 | of the form described above. | |
52 | ||
53 | In order to make predicate (possibly) true, all of its clauses must | |
54 | be (possibly) true. To make clause (possibly) true, one of conditions | |
55 | it mentions must be (possibly) true. There are fixed bounds on | |
56 | number of clauses and conditions and all the manipulation functions | |
57 | are conservative in positive direction. I.e. we may lose precision | |
58 | by thinking that predicate may be true even when it is not. | |
59 | ||
60 | estimate_edge_size and estimate_edge_growth can be used to query | |
61 | function size/time in the given context. inline_merge_summary merges | |
62 | properties of caller and callee after inlining. | |
63 | ||
03dfc36d JH |
64 | Finally pass_inline_parameters is exported. This is used to drive |
65 | computation of function parameters used by the early inliner. IPA | |
66 | inlined performs analysis via its analyze_function method. */ | |
67 | ||
68 | #include "config.h" | |
69 | #include "system.h" | |
70 | #include "coretypes.h" | |
71 | #include "tm.h" | |
72 | #include "tree.h" | |
73 | #include "tree-inline.h" | |
74 | #include "langhooks.h" | |
75 | #include "flags.h" | |
76 | #include "cgraph.h" | |
77 | #include "diagnostic.h" | |
78 | #include "gimple-pretty-print.h" | |
03dfc36d JH |
79 | #include "params.h" |
80 | #include "tree-pass.h" | |
81 | #include "coverage.h" | |
82 | #include "ggc.h" | |
83 | #include "tree-flow.h" | |
84 | #include "ipa-prop.h" | |
10a5dd5d | 85 | #include "lto-streamer.h" |
f0efc7aa DN |
86 | #include "data-streamer.h" |
87 | #include "tree-streamer.h" | |
03dfc36d | 88 | #include "ipa-inline.h" |
991278ab | 89 | #include "alloc-pool.h" |
391886c8 | 90 | #include "cfgloop.h" |
2daffc47 JH |
91 | #include "cfgloop.h" |
92 | #include "tree-scalar-evolution.h" | |
03dfc36d | 93 | |
632b4f8e | 94 | /* Estimate runtime of function can easilly run into huge numbers with many |
93b765d0 RS |
95 | nested loops. Be sure we can compute time * INLINE_SIZE_SCALE * 2 in an |
96 | integer. For anything larger we use gcov_type. */ | |
99e299a8 | 97 | #define MAX_TIME 500000 |
632b4f8e JH |
98 | |
99 | /* Number of bits in integer, but we really want to be stable across different | |
100 | hosts. */ | |
101 | #define NUM_CONDITIONS 32 | |
102 | ||
103 | enum predicate_conditions | |
104 | { | |
105 | predicate_false_condition = 0, | |
106 | predicate_not_inlined_condition = 1, | |
107 | predicate_first_dynamic_condition = 2 | |
108 | }; | |
109 | ||
110 | /* Special condition code we use to represent test that operand is compile time | |
111 | constant. */ | |
112 | #define IS_NOT_CONSTANT ERROR_MARK | |
25837a2f JH |
113 | /* Special condition code we use to represent test that operand is not changed |
114 | across invocation of the function. When operand IS_NOT_CONSTANT it is always | |
115 | CHANGED, however i.e. loop invariants can be NOT_CHANGED given percentage | |
116 | of executions even when they are not compile time constants. */ | |
117 | #define CHANGED IDENTIFIER_NODE | |
03dfc36d JH |
118 | |
119 | /* Holders of ipa cgraph hooks: */ | |
120 | static struct cgraph_node_hook_list *function_insertion_hook_holder; | |
10a5dd5d JH |
121 | static struct cgraph_node_hook_list *node_removal_hook_holder; |
122 | static struct cgraph_2node_hook_list *node_duplication_hook_holder; | |
898b8927 | 123 | static struct cgraph_2edge_hook_list *edge_duplication_hook_holder; |
632b4f8e | 124 | static struct cgraph_edge_hook_list *edge_removal_hook_holder; |
10a5dd5d JH |
125 | static void inline_node_removal_hook (struct cgraph_node *, void *); |
126 | static void inline_node_duplication_hook (struct cgraph_node *, | |
127 | struct cgraph_node *, void *); | |
898b8927 JH |
128 | static void inline_edge_removal_hook (struct cgraph_edge *, void *); |
129 | static void inline_edge_duplication_hook (struct cgraph_edge *, | |
130 | struct cgraph_edge *, | |
131 | void *); | |
10a5dd5d | 132 | |
632b4f8e JH |
133 | /* VECtor holding inline summaries. |
134 | In GGC memory because conditions might point to constant trees. */ | |
135 | VEC(inline_summary_t,gc) *inline_summary_vec; | |
898b8927 | 136 | VEC(inline_edge_summary_t,heap) *inline_edge_summary_vec; |
632b4f8e JH |
137 | |
138 | /* Cached node/edge growths. */ | |
139 | VEC(int,heap) *node_growth_cache; | |
140 | VEC(edge_growth_cache_entry,heap) *edge_growth_cache; | |
141 | ||
991278ab JH |
142 | /* Edge predicates goes here. */ |
143 | static alloc_pool edge_predicate_pool; | |
632b4f8e JH |
144 | |
145 | /* Return true predicate (tautology). | |
146 | We represent it by empty list of clauses. */ | |
147 | ||
148 | static inline struct predicate | |
149 | true_predicate (void) | |
150 | { | |
151 | struct predicate p; | |
152 | p.clause[0]=0; | |
153 | return p; | |
154 | } | |
155 | ||
156 | ||
157 | /* Return predicate testing single condition number COND. */ | |
158 | ||
159 | static inline struct predicate | |
160 | single_cond_predicate (int cond) | |
161 | { | |
162 | struct predicate p; | |
163 | p.clause[0]=1 << cond; | |
164 | p.clause[1]=0; | |
165 | return p; | |
166 | } | |
167 | ||
168 | ||
169 | /* Return false predicate. First clause require false condition. */ | |
170 | ||
171 | static inline struct predicate | |
172 | false_predicate (void) | |
173 | { | |
174 | return single_cond_predicate (predicate_false_condition); | |
175 | } | |
176 | ||
177 | ||
991278ab JH |
178 | /* Return true if P is (false). */ |
179 | ||
180 | static inline bool | |
181 | true_predicate_p (struct predicate *p) | |
182 | { | |
183 | return !p->clause[0]; | |
184 | } | |
185 | ||
186 | ||
187 | /* Return true if P is (false). */ | |
188 | ||
189 | static inline bool | |
190 | false_predicate_p (struct predicate *p) | |
191 | { | |
192 | if (p->clause[0] == (1 << predicate_false_condition)) | |
193 | { | |
194 | gcc_checking_assert (!p->clause[1] | |
195 | && p->clause[0] == 1 << predicate_false_condition); | |
196 | return true; | |
197 | } | |
198 | return false; | |
199 | } | |
200 | ||
201 | ||
632b4f8e JH |
202 | /* Return predicate that is set true when function is not inlined. */ |
203 | static inline struct predicate | |
204 | not_inlined_predicate (void) | |
205 | { | |
206 | return single_cond_predicate (predicate_not_inlined_condition); | |
207 | } | |
208 | ||
8810cc52 MJ |
209 | /* Simple description of whether a memory load or a condition refers to a load |
210 | from an aggregate and if so, how and where from in the aggregate. | |
211 | Individual fields have the same meaning like fields with the same name in | |
212 | struct condition. */ | |
632b4f8e | 213 | |
8810cc52 MJ |
214 | struct agg_position_info |
215 | { | |
216 | HOST_WIDE_INT offset; | |
217 | bool agg_contents; | |
218 | bool by_ref; | |
219 | }; | |
220 | ||
221 | /* Add condition to condition list CONDS. AGGPOS describes whether the used | |
222 | oprand is loaded from an aggregate and where in the aggregate it is. It can | |
223 | be NULL, which means this not a load from an aggregate. */ | |
632b4f8e JH |
224 | |
225 | static struct predicate | |
226 | add_condition (struct inline_summary *summary, int operand_num, | |
8810cc52 | 227 | struct agg_position_info *aggpos, |
632b4f8e JH |
228 | enum tree_code code, tree val) |
229 | { | |
230 | int i; | |
231 | struct condition *c; | |
232 | struct condition new_cond; | |
8810cc52 MJ |
233 | HOST_WIDE_INT offset; |
234 | bool agg_contents, by_ref; | |
632b4f8e | 235 | |
8810cc52 MJ |
236 | if (aggpos) |
237 | { | |
238 | offset = aggpos->offset; | |
239 | agg_contents = aggpos->agg_contents; | |
240 | by_ref = aggpos->by_ref; | |
241 | } | |
242 | else | |
243 | { | |
244 | offset = 0; | |
245 | agg_contents = false; | |
246 | by_ref = false; | |
247 | } | |
248 | ||
249 | gcc_checking_assert (operand_num >= 0); | |
632b4f8e JH |
250 | for (i = 0; VEC_iterate (condition, summary->conds, i, c); i++) |
251 | { | |
252 | if (c->operand_num == operand_num | |
253 | && c->code == code | |
8810cc52 MJ |
254 | && c->val == val |
255 | && c->agg_contents == agg_contents | |
256 | && (!agg_contents || (c->offset == offset && c->by_ref == by_ref))) | |
632b4f8e JH |
257 | return single_cond_predicate (i + predicate_first_dynamic_condition); |
258 | } | |
259 | /* Too many conditions. Give up and return constant true. */ | |
260 | if (i == NUM_CONDITIONS - predicate_first_dynamic_condition) | |
261 | return true_predicate (); | |
262 | ||
263 | new_cond.operand_num = operand_num; | |
264 | new_cond.code = code; | |
265 | new_cond.val = val; | |
8810cc52 MJ |
266 | new_cond.agg_contents = agg_contents; |
267 | new_cond.by_ref = by_ref; | |
268 | new_cond.offset = offset; | |
f32682ca | 269 | VEC_safe_push (condition, gc, summary->conds, new_cond); |
632b4f8e JH |
270 | return single_cond_predicate (i + predicate_first_dynamic_condition); |
271 | } | |
272 | ||
273 | ||
b15c64ee | 274 | /* Add clause CLAUSE into the predicate P. */ |
632b4f8e JH |
275 | |
276 | static inline void | |
a61bd030 | 277 | add_clause (conditions conditions, struct predicate *p, clause_t clause) |
632b4f8e JH |
278 | { |
279 | int i; | |
b15c64ee | 280 | int i2; |
f3181aa2 | 281 | int insert_here = -1; |
a61bd030 | 282 | int c1, c2; |
991278ab | 283 | |
632b4f8e JH |
284 | /* True clause. */ |
285 | if (!clause) | |
286 | return; | |
287 | ||
b15c64ee | 288 | /* False clause makes the whole predicate false. Kill the other variants. */ |
991278ab | 289 | if (clause == (1 << predicate_false_condition)) |
632b4f8e JH |
290 | { |
291 | p->clause[0] = (1 << predicate_false_condition); | |
292 | p->clause[1] = 0; | |
991278ab | 293 | return; |
632b4f8e | 294 | } |
991278ab JH |
295 | if (false_predicate_p (p)) |
296 | return; | |
b15c64ee JH |
297 | |
298 | /* No one should be sily enough to add false into nontrivial clauses. */ | |
299 | gcc_checking_assert (!(clause & (1 << predicate_false_condition))); | |
300 | ||
301 | /* Look where to insert the clause. At the same time prune out | |
302 | clauses of P that are implied by the new clause and thus | |
303 | redundant. */ | |
304 | for (i = 0, i2 = 0; i <= MAX_CLAUSES; i++) | |
632b4f8e | 305 | { |
b15c64ee JH |
306 | p->clause[i2] = p->clause[i]; |
307 | ||
632b4f8e JH |
308 | if (!p->clause[i]) |
309 | break; | |
b15c64ee JH |
310 | |
311 | /* If p->clause[i] implies clause, there is nothing to add. */ | |
312 | if ((p->clause[i] & clause) == p->clause[i]) | |
313 | { | |
9e990d14 JH |
314 | /* We had nothing to add, none of clauses should've become |
315 | redundant. */ | |
b15c64ee JH |
316 | gcc_checking_assert (i == i2); |
317 | return; | |
318 | } | |
319 | ||
320 | if (p->clause[i] < clause && insert_here < 0) | |
321 | insert_here = i2; | |
322 | ||
323 | /* If clause implies p->clause[i], then p->clause[i] becomes redundant. | |
324 | Otherwise the p->clause[i] has to stay. */ | |
325 | if ((p->clause[i] & clause) != clause) | |
326 | i2++; | |
632b4f8e | 327 | } |
a61bd030 JH |
328 | |
329 | /* Look for clauses that are obviously true. I.e. | |
330 | op0 == 5 || op0 != 5. */ | |
331 | for (c1 = predicate_first_dynamic_condition; c1 < NUM_CONDITIONS; c1++) | |
25837a2f JH |
332 | { |
333 | condition *cc1; | |
334 | if (!(clause & (1 << c1))) | |
335 | continue; | |
0823efed DN |
336 | cc1 = &VEC_index (condition, |
337 | conditions, | |
338 | c1 - predicate_first_dynamic_condition); | |
25837a2f JH |
339 | /* We have no way to represent !CHANGED and !IS_NOT_CONSTANT |
340 | and thus there is no point for looking for them. */ | |
341 | if (cc1->code == CHANGED | |
342 | || cc1->code == IS_NOT_CONSTANT) | |
343 | continue; | |
344 | for (c2 = c1 + 1; c2 <= NUM_CONDITIONS; c2++) | |
345 | if (clause & (1 << c2)) | |
346 | { | |
0823efed DN |
347 | condition *cc1 = &VEC_index (condition, |
348 | conditions, | |
349 | c1 - predicate_first_dynamic_condition); | |
350 | condition *cc2 = &VEC_index (condition, | |
351 | conditions, | |
352 | c2 - predicate_first_dynamic_condition); | |
25837a2f JH |
353 | if (cc1->operand_num == cc2->operand_num |
354 | && cc1->val == cc2->val | |
355 | && cc2->code != IS_NOT_CONSTANT | |
356 | && cc2->code != CHANGED | |
357 | && cc1->code == invert_tree_comparison | |
358 | (cc2->code, | |
359 | HONOR_NANS (TYPE_MODE (TREE_TYPE (cc1->val))))) | |
360 | return; | |
361 | } | |
362 | } | |
a61bd030 JH |
363 | |
364 | ||
b15c64ee JH |
365 | /* We run out of variants. Be conservative in positive direction. */ |
366 | if (i2 == MAX_CLAUSES) | |
632b4f8e | 367 | return; |
b15c64ee JH |
368 | /* Keep clauses in decreasing order. This makes equivalence testing easy. */ |
369 | p->clause[i2 + 1] = 0; | |
f3181aa2 | 370 | if (insert_here >= 0) |
b15c64ee JH |
371 | for (;i2 > insert_here; i2--) |
372 | p->clause[i2] = p->clause[i2 - 1]; | |
f3181aa2 | 373 | else |
b15c64ee | 374 | insert_here = i2; |
632b4f8e JH |
375 | p->clause[insert_here] = clause; |
376 | } | |
377 | ||
378 | ||
379 | /* Return P & P2. */ | |
380 | ||
381 | static struct predicate | |
a61bd030 JH |
382 | and_predicates (conditions conditions, |
383 | struct predicate *p, struct predicate *p2) | |
632b4f8e JH |
384 | { |
385 | struct predicate out = *p; | |
386 | int i; | |
991278ab | 387 | |
b15c64ee JH |
388 | /* Avoid busy work. */ |
389 | if (false_predicate_p (p2) || true_predicate_p (p)) | |
390 | return *p2; | |
391 | if (false_predicate_p (p) || true_predicate_p (p2)) | |
392 | return *p; | |
393 | ||
394 | /* See how far predicates match. */ | |
395 | for (i = 0; p->clause[i] && p->clause[i] == p2->clause[i]; i++) | |
396 | { | |
397 | gcc_checking_assert (i < MAX_CLAUSES); | |
398 | } | |
399 | ||
400 | /* Combine the predicates rest. */ | |
401 | for (; p2->clause[i]; i++) | |
f3181aa2 JH |
402 | { |
403 | gcc_checking_assert (i < MAX_CLAUSES); | |
a61bd030 | 404 | add_clause (conditions, &out, p2->clause[i]); |
f3181aa2 | 405 | } |
632b4f8e JH |
406 | return out; |
407 | } | |
408 | ||
409 | ||
b15c64ee JH |
410 | /* Return true if predicates are obviously equal. */ |
411 | ||
412 | static inline bool | |
413 | predicates_equal_p (struct predicate *p, struct predicate *p2) | |
414 | { | |
415 | int i; | |
416 | for (i = 0; p->clause[i]; i++) | |
417 | { | |
418 | gcc_checking_assert (i < MAX_CLAUSES); | |
419 | gcc_checking_assert (p->clause [i] > p->clause[i + 1]); | |
9e990d14 JH |
420 | gcc_checking_assert (!p2->clause[i] |
421 | || p2->clause [i] > p2->clause[i + 1]); | |
b15c64ee JH |
422 | if (p->clause[i] != p2->clause[i]) |
423 | return false; | |
424 | } | |
425 | return !p2->clause[i]; | |
426 | } | |
427 | ||
428 | ||
632b4f8e JH |
429 | /* Return P | P2. */ |
430 | ||
431 | static struct predicate | |
a61bd030 | 432 | or_predicates (conditions conditions, struct predicate *p, struct predicate *p2) |
632b4f8e JH |
433 | { |
434 | struct predicate out = true_predicate (); | |
435 | int i,j; | |
991278ab | 436 | |
b15c64ee JH |
437 | /* Avoid busy work. */ |
438 | if (false_predicate_p (p2) || true_predicate_p (p)) | |
991278ab | 439 | return *p; |
b15c64ee | 440 | if (false_predicate_p (p) || true_predicate_p (p2)) |
991278ab | 441 | return *p2; |
b15c64ee JH |
442 | if (predicates_equal_p (p, p2)) |
443 | return *p; | |
444 | ||
445 | /* OK, combine the predicates. */ | |
632b4f8e JH |
446 | for (i = 0; p->clause[i]; i++) |
447 | for (j = 0; p2->clause[j]; j++) | |
f3181aa2 JH |
448 | { |
449 | gcc_checking_assert (i < MAX_CLAUSES && j < MAX_CLAUSES); | |
a61bd030 | 450 | add_clause (conditions, &out, p->clause[i] | p2->clause[j]); |
f3181aa2 | 451 | } |
632b4f8e JH |
452 | return out; |
453 | } | |
454 | ||
455 | ||
9e990d14 JH |
456 | /* Having partial truth assignment in POSSIBLE_TRUTHS, return false |
457 | if predicate P is known to be false. */ | |
632b4f8e JH |
458 | |
459 | static bool | |
991278ab | 460 | evaluate_predicate (struct predicate *p, clause_t possible_truths) |
632b4f8e JH |
461 | { |
462 | int i; | |
463 | ||
464 | /* True remains true. */ | |
991278ab | 465 | if (true_predicate_p (p)) |
632b4f8e JH |
466 | return true; |
467 | ||
991278ab JH |
468 | gcc_assert (!(possible_truths & (1 << predicate_false_condition))); |
469 | ||
632b4f8e JH |
470 | /* See if we can find clause we can disprove. */ |
471 | for (i = 0; p->clause[i]; i++) | |
f3181aa2 JH |
472 | { |
473 | gcc_checking_assert (i < MAX_CLAUSES); | |
474 | if (!(p->clause[i] & possible_truths)) | |
475 | return false; | |
476 | } | |
632b4f8e JH |
477 | return true; |
478 | } | |
479 | ||
25837a2f JH |
480 | /* Return the probability in range 0...REG_BR_PROB_BASE that the predicated |
481 | instruction will be recomputed per invocation of the inlined call. */ | |
482 | ||
483 | static int | |
484 | predicate_probability (conditions conds, | |
485 | struct predicate *p, clause_t possible_truths, | |
486 | VEC (inline_param_summary_t, heap) *inline_param_summary) | |
487 | { | |
488 | int i; | |
489 | int combined_prob = REG_BR_PROB_BASE; | |
490 | ||
491 | /* True remains true. */ | |
492 | if (true_predicate_p (p)) | |
493 | return REG_BR_PROB_BASE; | |
494 | ||
495 | if (false_predicate_p (p)) | |
496 | return 0; | |
497 | ||
498 | gcc_assert (!(possible_truths & (1 << predicate_false_condition))); | |
499 | ||
500 | /* See if we can find clause we can disprove. */ | |
501 | for (i = 0; p->clause[i]; i++) | |
502 | { | |
503 | gcc_checking_assert (i < MAX_CLAUSES); | |
504 | if (!(p->clause[i] & possible_truths)) | |
505 | return 0; | |
506 | else | |
507 | { | |
508 | int this_prob = 0; | |
509 | int i2; | |
510 | if (!inline_param_summary) | |
511 | return REG_BR_PROB_BASE; | |
512 | for (i2 = 0; i2 < NUM_CONDITIONS; i2++) | |
513 | if ((p->clause[i] & possible_truths) & (1 << i2)) | |
514 | { | |
515 | if (i2 >= predicate_first_dynamic_condition) | |
516 | { | |
0823efed | 517 | condition *c = &VEC_index |
25837a2f JH |
518 | (condition, conds, |
519 | i2 - predicate_first_dynamic_condition); | |
520 | if (c->code == CHANGED | |
521 | && (c->operand_num | |
0578e417 RS |
522 | < (int) VEC_length (inline_param_summary_t, |
523 | inline_param_summary))) | |
25837a2f JH |
524 | { |
525 | int iprob = VEC_index (inline_param_summary_t, | |
526 | inline_param_summary, | |
0823efed | 527 | c->operand_num).change_prob; |
25837a2f JH |
528 | this_prob = MAX (this_prob, iprob); |
529 | } | |
530 | else | |
531 | this_prob = REG_BR_PROB_BASE; | |
532 | } | |
533 | else | |
534 | this_prob = REG_BR_PROB_BASE; | |
535 | } | |
536 | combined_prob = MIN (this_prob, combined_prob); | |
537 | if (!combined_prob) | |
538 | return 0; | |
539 | } | |
540 | } | |
541 | return combined_prob; | |
542 | } | |
543 | ||
632b4f8e JH |
544 | |
545 | /* Dump conditional COND. */ | |
546 | ||
547 | static void | |
548 | dump_condition (FILE *f, conditions conditions, int cond) | |
549 | { | |
550 | condition *c; | |
551 | if (cond == predicate_false_condition) | |
552 | fprintf (f, "false"); | |
553 | else if (cond == predicate_not_inlined_condition) | |
554 | fprintf (f, "not inlined"); | |
555 | else | |
556 | { | |
0823efed DN |
557 | c = &VEC_index (condition, conditions, |
558 | cond - predicate_first_dynamic_condition); | |
632b4f8e | 559 | fprintf (f, "op%i", c->operand_num); |
8810cc52 MJ |
560 | if (c->agg_contents) |
561 | fprintf (f, "[%soffset: " HOST_WIDE_INT_PRINT_DEC "]", | |
562 | c->by_ref ? "ref " : "", c->offset); | |
632b4f8e JH |
563 | if (c->code == IS_NOT_CONSTANT) |
564 | { | |
565 | fprintf (f, " not constant"); | |
566 | return; | |
567 | } | |
25837a2f JH |
568 | if (c->code == CHANGED) |
569 | { | |
570 | fprintf (f, " changed"); | |
571 | return; | |
572 | } | |
632b4f8e JH |
573 | fprintf (f, " %s ", op_symbol_code (c->code)); |
574 | print_generic_expr (f, c->val, 1); | |
575 | } | |
576 | } | |
577 | ||
578 | ||
579 | /* Dump clause CLAUSE. */ | |
580 | ||
581 | static void | |
582 | dump_clause (FILE *f, conditions conds, clause_t clause) | |
583 | { | |
584 | int i; | |
585 | bool found = false; | |
586 | fprintf (f, "("); | |
587 | if (!clause) | |
588 | fprintf (f, "true"); | |
589 | for (i = 0; i < NUM_CONDITIONS; i++) | |
590 | if (clause & (1 << i)) | |
591 | { | |
592 | if (found) | |
593 | fprintf (f, " || "); | |
594 | found = true; | |
595 | dump_condition (f, conds, i); | |
596 | } | |
597 | fprintf (f, ")"); | |
598 | } | |
599 | ||
600 | ||
601 | /* Dump predicate PREDICATE. */ | |
602 | ||
603 | static void | |
604 | dump_predicate (FILE *f, conditions conds, struct predicate *pred) | |
605 | { | |
606 | int i; | |
991278ab | 607 | if (true_predicate_p (pred)) |
632b4f8e JH |
608 | dump_clause (f, conds, 0); |
609 | else | |
610 | for (i = 0; pred->clause[i]; i++) | |
611 | { | |
612 | if (i) | |
613 | fprintf (f, " && "); | |
614 | dump_clause (f, conds, pred->clause[i]); | |
615 | } | |
616 | fprintf (f, "\n"); | |
617 | } | |
618 | ||
619 | ||
37678631 JH |
620 | /* Dump inline hints. */ |
621 | void | |
622 | dump_inline_hints (FILE *f, inline_hints hints) | |
623 | { | |
624 | if (!hints) | |
625 | return; | |
626 | fprintf (f, "inline hints:"); | |
627 | if (hints & INLINE_HINT_indirect_call) | |
628 | { | |
629 | hints &= ~INLINE_HINT_indirect_call; | |
630 | fprintf (f, " indirect_call"); | |
631 | } | |
2daffc47 JH |
632 | if (hints & INLINE_HINT_loop_iterations) |
633 | { | |
634 | hints &= ~INLINE_HINT_loop_iterations; | |
635 | fprintf (f, " loop_iterations"); | |
636 | } | |
128e0d89 JH |
637 | if (hints & INLINE_HINT_loop_stride) |
638 | { | |
639 | hints &= ~INLINE_HINT_loop_stride; | |
640 | fprintf (f, " loop_stride"); | |
641 | } | |
b48ccf0d JH |
642 | if (hints & INLINE_HINT_same_scc) |
643 | { | |
644 | hints &= ~INLINE_HINT_same_scc; | |
645 | fprintf (f, " same_scc"); | |
646 | } | |
647 | if (hints & INLINE_HINT_in_scc) | |
648 | { | |
649 | hints &= ~INLINE_HINT_in_scc; | |
650 | fprintf (f, " in_scc"); | |
651 | } | |
37678631 JH |
652 | gcc_assert (!hints); |
653 | } | |
654 | ||
655 | ||
632b4f8e JH |
656 | /* Record SIZE and TIME under condition PRED into the inline summary. */ |
657 | ||
658 | static void | |
9e990d14 JH |
659 | account_size_time (struct inline_summary *summary, int size, int time, |
660 | struct predicate *pred) | |
632b4f8e JH |
661 | { |
662 | size_time_entry *e; | |
663 | bool found = false; | |
664 | int i; | |
665 | ||
991278ab | 666 | if (false_predicate_p (pred)) |
632b4f8e JH |
667 | return; |
668 | ||
669 | /* We need to create initial empty unconitional clause, but otherwie | |
670 | we don't need to account empty times and sizes. */ | |
74605a11 | 671 | if (!size && !time && summary->entry) |
632b4f8e JH |
672 | return; |
673 | ||
674 | /* Watch overflow that might result from insane profiles. */ | |
675 | if (time > MAX_TIME * INLINE_TIME_SCALE) | |
676 | time = MAX_TIME * INLINE_TIME_SCALE; | |
677 | gcc_assert (time >= 0); | |
678 | ||
679 | for (i = 0; VEC_iterate (size_time_entry, summary->entry, i, e); i++) | |
680 | if (predicates_equal_p (&e->predicate, pred)) | |
681 | { | |
682 | found = true; | |
683 | break; | |
684 | } | |
685 | if (i == 32) | |
686 | { | |
687 | i = 0; | |
688 | found = true; | |
0823efed | 689 | e = &VEC_index (size_time_entry, summary->entry, 0); |
632b4f8e JH |
690 | gcc_assert (!e->predicate.clause[0]); |
691 | } | |
692 | if (dump_file && (dump_flags & TDF_DETAILS) && (time || size)) | |
693 | { | |
694 | fprintf (dump_file, "\t\tAccounting size:%3.2f, time:%3.2f on %spredicate:", | |
9e990d14 JH |
695 | ((double)size) / INLINE_SIZE_SCALE, |
696 | ((double)time) / INLINE_TIME_SCALE, | |
632b4f8e JH |
697 | found ? "" : "new "); |
698 | dump_predicate (dump_file, summary->conds, pred); | |
699 | } | |
700 | if (!found) | |
701 | { | |
702 | struct size_time_entry new_entry; | |
703 | new_entry.size = size; | |
704 | new_entry.time = time; | |
705 | new_entry.predicate = *pred; | |
f32682ca | 706 | VEC_safe_push (size_time_entry, gc, summary->entry, new_entry); |
632b4f8e JH |
707 | } |
708 | else | |
709 | { | |
710 | e->size += size; | |
711 | e->time += time; | |
712 | if (e->time > MAX_TIME * INLINE_TIME_SCALE) | |
713 | e->time = MAX_TIME * INLINE_TIME_SCALE; | |
714 | } | |
715 | } | |
716 | ||
991278ab JH |
717 | /* Set predicate for edge E. */ |
718 | ||
719 | static void | |
720 | edge_set_predicate (struct cgraph_edge *e, struct predicate *predicate) | |
721 | { | |
722 | struct inline_edge_summary *es = inline_edge_summary (e); | |
723 | if (predicate && !true_predicate_p (predicate)) | |
724 | { | |
725 | if (!es->predicate) | |
726 | es->predicate = (struct predicate *)pool_alloc (edge_predicate_pool); | |
727 | *es->predicate = *predicate; | |
728 | } | |
729 | else | |
730 | { | |
731 | if (es->predicate) | |
732 | pool_free (edge_predicate_pool, es->predicate); | |
733 | es->predicate = NULL; | |
734 | } | |
735 | } | |
736 | ||
128e0d89 JH |
737 | /* Set predicate for hint *P. */ |
738 | ||
739 | static void | |
740 | set_hint_predicate (struct predicate **p, struct predicate new_predicate) | |
741 | { | |
742 | if (false_predicate_p (&new_predicate) | |
743 | || true_predicate_p (&new_predicate)) | |
744 | { | |
745 | if (*p) | |
746 | pool_free (edge_predicate_pool, *p); | |
747 | *p = NULL; | |
748 | } | |
749 | else | |
750 | { | |
751 | if (!*p) | |
752 | *p = (struct predicate *)pool_alloc (edge_predicate_pool); | |
753 | **p = new_predicate; | |
754 | } | |
755 | } | |
756 | ||
632b4f8e | 757 | |
74605a11 | 758 | /* KNOWN_VALS is partial mapping of parameters of NODE to constant values. |
8810cc52 MJ |
759 | KNOWN_AGGS is a vector of aggreggate jump functions for each parameter. |
760 | Return clause of possible truths. When INLINE_P is true, assume that we are | |
761 | inlining. | |
25837a2f JH |
762 | |
763 | ERROR_MARK means compile time invariant. */ | |
74605a11 JH |
764 | |
765 | static clause_t | |
766 | evaluate_conditions_for_known_args (struct cgraph_node *node, | |
8810cc52 MJ |
767 | bool inline_p, |
768 | VEC (tree, heap) *known_vals, | |
769 | VEC (ipa_agg_jump_function_p, heap) *known_aggs) | |
74605a11 JH |
770 | { |
771 | clause_t clause = inline_p ? 0 : 1 << predicate_not_inlined_condition; | |
772 | struct inline_summary *info = inline_summary (node); | |
773 | int i; | |
774 | struct condition *c; | |
775 | ||
776 | for (i = 0; VEC_iterate (condition, info->conds, i, c); i++) | |
777 | { | |
a45c0557 | 778 | tree val; |
74605a11 JH |
779 | tree res; |
780 | ||
8810cc52 MJ |
781 | /* We allow call stmt to have fewer arguments than the callee function |
782 | (especially for K&R style programs). So bound check here (we assume | |
783 | known_aggs vector, if non-NULL, has the same length as | |
784 | known_vals). */ | |
785 | gcc_checking_assert (!known_aggs | |
786 | || (VEC_length (tree, known_vals) | |
787 | == VEC_length (ipa_agg_jump_function_p, | |
788 | known_aggs))); | |
789 | if (c->operand_num >= (int) VEC_length (tree, known_vals)) | |
790 | { | |
791 | clause |= 1 << (i + predicate_first_dynamic_condition); | |
792 | continue; | |
793 | } | |
a45c0557 | 794 | |
8810cc52 MJ |
795 | if (c->agg_contents) |
796 | { | |
797 | struct ipa_agg_jump_function *agg; | |
798 | ||
799 | if (c->code == CHANGED | |
800 | && !c->by_ref | |
801 | && (VEC_index (tree, known_vals, c->operand_num) | |
802 | == error_mark_node)) | |
803 | continue; | |
804 | ||
805 | if (known_aggs) | |
806 | { | |
807 | agg = VEC_index (ipa_agg_jump_function_p, known_aggs, | |
808 | c->operand_num); | |
809 | val = ipa_find_agg_cst_for_param (agg, c->offset, c->by_ref); | |
810 | } | |
811 | else | |
812 | val = NULL_TREE; | |
813 | } | |
814 | else | |
815 | { | |
816 | val = VEC_index (tree, known_vals, c->operand_num); | |
817 | if (val == error_mark_node && c->code != CHANGED) | |
818 | val = NULL_TREE; | |
819 | } | |
25837a2f | 820 | |
74605a11 JH |
821 | if (!val) |
822 | { | |
823 | clause |= 1 << (i + predicate_first_dynamic_condition); | |
824 | continue; | |
825 | } | |
25837a2f | 826 | if (c->code == IS_NOT_CONSTANT || c->code == CHANGED) |
74605a11 JH |
827 | continue; |
828 | res = fold_binary_to_constant (c->code, boolean_type_node, val, c->val); | |
829 | if (res | |
830 | && integer_zerop (res)) | |
831 | continue; | |
832 | clause |= 1 << (i + predicate_first_dynamic_condition); | |
833 | } | |
834 | return clause; | |
835 | } | |
836 | ||
837 | ||
632b4f8e JH |
838 | /* Work out what conditions might be true at invocation of E. */ |
839 | ||
d2d668fb MK |
840 | static void |
841 | evaluate_properties_for_edge (struct cgraph_edge *e, bool inline_p, | |
8810cc52 MJ |
842 | clause_t *clause_ptr, |
843 | VEC (tree, heap) **known_vals_ptr, | |
844 | VEC (tree, heap) **known_binfos_ptr, | |
845 | VEC (ipa_agg_jump_function_p, heap) **known_aggs_ptr) | |
632b4f8e | 846 | { |
a5b1779f JH |
847 | struct cgraph_node *callee = cgraph_function_or_thunk_node (e->callee, NULL); |
848 | struct inline_summary *info = inline_summary (callee); | |
d028561e | 849 | VEC (tree, heap) *known_vals = NULL; |
8810cc52 | 850 | VEC (ipa_agg_jump_function_p, heap) *known_aggs = NULL; |
632b4f8e | 851 | |
d2d668fb MK |
852 | if (clause_ptr) |
853 | *clause_ptr = inline_p ? 0 : 1 << predicate_not_inlined_condition; | |
854 | if (known_vals_ptr) | |
855 | *known_vals_ptr = NULL; | |
856 | if (known_binfos_ptr) | |
857 | *known_binfos_ptr = NULL; | |
858 | ||
859 | if (ipa_node_params_vector | |
d028561e | 860 | && !e->call_stmt_cannot_inline_p |
d2d668fb | 861 | && ((clause_ptr && info->conds) || known_vals_ptr || known_binfos_ptr)) |
632b4f8e JH |
862 | { |
863 | struct ipa_node_params *parms_info; | |
864 | struct ipa_edge_args *args = IPA_EDGE_REF (e); | |
25837a2f | 865 | struct inline_edge_summary *es = inline_edge_summary (e); |
632b4f8e | 866 | int i, count = ipa_get_cs_argument_count (args); |
632b4f8e JH |
867 | |
868 | if (e->caller->global.inlined_to) | |
869 | parms_info = IPA_NODE_REF (e->caller->global.inlined_to); | |
870 | else | |
871 | parms_info = IPA_NODE_REF (e->caller); | |
872 | ||
d2d668fb MK |
873 | if (count && (info->conds || known_vals_ptr)) |
874 | VEC_safe_grow_cleared (tree, heap, known_vals, count); | |
8810cc52 MJ |
875 | if (count && (info->conds || known_aggs_ptr)) |
876 | VEC_safe_grow_cleared (ipa_agg_jump_function_p, heap, known_aggs, | |
877 | count); | |
d2d668fb MK |
878 | if (count && known_binfos_ptr) |
879 | VEC_safe_grow_cleared (tree, heap, *known_binfos_ptr, count); | |
880 | ||
632b4f8e JH |
881 | for (i = 0; i < count; i++) |
882 | { | |
8810cc52 MJ |
883 | struct ipa_jump_func *jf = ipa_get_ith_jump_func (args, i); |
884 | tree cst = ipa_value_from_jfunc (parms_info, jf); | |
411a20d6 | 885 | if (cst) |
d2d668fb | 886 | { |
d028561e | 887 | if (known_vals && TREE_CODE (cst) != TREE_BINFO) |
d2d668fb | 888 | VEC_replace (tree, known_vals, i, cst); |
d028561e | 889 | else if (known_binfos_ptr != NULL && TREE_CODE (cst) == TREE_BINFO) |
d2d668fb MK |
890 | VEC_replace (tree, *known_binfos_ptr, i, cst); |
891 | } | |
25837a2f JH |
892 | else if (inline_p |
893 | && !VEC_index (inline_param_summary_t, | |
894 | es->param, | |
0823efed | 895 | i).change_prob) |
25837a2f | 896 | VEC_replace (tree, known_vals, i, error_mark_node); |
8810cc52 MJ |
897 | /* TODO: When IPA-CP starts propagating and merging aggregate jump |
898 | functions, use its knowledge of the caller too, just like the | |
899 | scalar case above. */ | |
900 | VEC_replace (ipa_agg_jump_function_p, known_aggs, i, &jf->agg); | |
632b4f8e | 901 | } |
632b4f8e | 902 | } |
632b4f8e | 903 | |
d028561e JH |
904 | if (clause_ptr) |
905 | *clause_ptr = evaluate_conditions_for_known_args (callee, inline_p, | |
8810cc52 | 906 | known_vals, known_aggs); |
d028561e JH |
907 | |
908 | if (known_vals_ptr) | |
909 | *known_vals_ptr = known_vals; | |
910 | else | |
911 | VEC_free (tree, heap, known_vals); | |
8810cc52 MJ |
912 | |
913 | if (known_aggs_ptr) | |
914 | *known_aggs_ptr = known_aggs; | |
915 | else | |
916 | VEC_free (ipa_agg_jump_function_p, heap, known_aggs); | |
632b4f8e JH |
917 | } |
918 | ||
10a5dd5d JH |
919 | |
920 | /* Allocate the inline summary vector or resize it to cover all cgraph nodes. */ | |
921 | ||
922 | static void | |
923 | inline_summary_alloc (void) | |
924 | { | |
925 | if (!node_removal_hook_holder) | |
926 | node_removal_hook_holder = | |
927 | cgraph_add_node_removal_hook (&inline_node_removal_hook, NULL); | |
898b8927 JH |
928 | if (!edge_removal_hook_holder) |
929 | edge_removal_hook_holder = | |
930 | cgraph_add_edge_removal_hook (&inline_edge_removal_hook, NULL); | |
10a5dd5d JH |
931 | if (!node_duplication_hook_holder) |
932 | node_duplication_hook_holder = | |
933 | cgraph_add_node_duplication_hook (&inline_node_duplication_hook, NULL); | |
898b8927 JH |
934 | if (!edge_duplication_hook_holder) |
935 | edge_duplication_hook_holder = | |
936 | cgraph_add_edge_duplication_hook (&inline_edge_duplication_hook, NULL); | |
10a5dd5d JH |
937 | |
938 | if (VEC_length (inline_summary_t, inline_summary_vec) | |
939 | <= (unsigned) cgraph_max_uid) | |
632b4f8e | 940 | VEC_safe_grow_cleared (inline_summary_t, gc, |
10a5dd5d | 941 | inline_summary_vec, cgraph_max_uid + 1); |
898b8927 JH |
942 | if (VEC_length (inline_edge_summary_t, inline_edge_summary_vec) |
943 | <= (unsigned) cgraph_edge_max_uid) | |
944 | VEC_safe_grow_cleared (inline_edge_summary_t, heap, | |
945 | inline_edge_summary_vec, cgraph_edge_max_uid + 1); | |
991278ab | 946 | if (!edge_predicate_pool) |
9e990d14 JH |
947 | edge_predicate_pool = create_alloc_pool ("edge predicates", |
948 | sizeof (struct predicate), | |
991278ab | 949 | 10); |
10a5dd5d JH |
950 | } |
951 | ||
1c52c601 JH |
952 | /* We are called multiple time for given function; clear |
953 | data from previous run so they are not cumulated. */ | |
954 | ||
955 | static void | |
956 | reset_inline_edge_summary (struct cgraph_edge *e) | |
957 | { | |
78e5ce9f JH |
958 | if (e->uid |
959 | < (int)VEC_length (inline_edge_summary_t, inline_edge_summary_vec)) | |
960 | { | |
961 | struct inline_edge_summary *es = inline_edge_summary (e); | |
1c52c601 | 962 | |
78e5ce9f JH |
963 | es->call_stmt_size = es->call_stmt_time =0; |
964 | if (es->predicate) | |
965 | pool_free (edge_predicate_pool, es->predicate); | |
966 | es->predicate = NULL; | |
967 | VEC_free (inline_param_summary_t, heap, es->param); | |
968 | } | |
1c52c601 JH |
969 | } |
970 | ||
971 | /* We are called multiple time for given function; clear | |
972 | data from previous run so they are not cumulated. */ | |
973 | ||
974 | static void | |
975 | reset_inline_summary (struct cgraph_node *node) | |
976 | { | |
977 | struct inline_summary *info = inline_summary (node); | |
978 | struct cgraph_edge *e; | |
979 | ||
980 | info->self_size = info->self_time = 0; | |
981 | info->estimated_stack_size = 0; | |
982 | info->estimated_self_stack_size = 0; | |
983 | info->stack_frame_offset = 0; | |
984 | info->size = 0; | |
985 | info->time = 0; | |
b48ccf0d | 986 | info->scc_no = 0; |
2daffc47 JH |
987 | if (info->loop_iterations) |
988 | { | |
989 | pool_free (edge_predicate_pool, info->loop_iterations); | |
990 | info->loop_iterations = NULL; | |
991 | } | |
128e0d89 JH |
992 | if (info->loop_stride) |
993 | { | |
994 | pool_free (edge_predicate_pool, info->loop_stride); | |
995 | info->loop_stride = NULL; | |
996 | } | |
1c52c601 JH |
997 | VEC_free (condition, gc, info->conds); |
998 | VEC_free (size_time_entry,gc, info->entry); | |
999 | for (e = node->callees; e; e = e->next_callee) | |
1000 | reset_inline_edge_summary (e); | |
1001 | for (e = node->indirect_calls; e; e = e->next_callee) | |
1002 | reset_inline_edge_summary (e); | |
1003 | } | |
1004 | ||
10a5dd5d JH |
1005 | /* Hook that is called by cgraph.c when a node is removed. */ |
1006 | ||
1007 | static void | |
1008 | inline_node_removal_hook (struct cgraph_node *node, void *data ATTRIBUTE_UNUSED) | |
1009 | { | |
e7f23018 | 1010 | struct inline_summary *info; |
10a5dd5d JH |
1011 | if (VEC_length (inline_summary_t, inline_summary_vec) |
1012 | <= (unsigned)node->uid) | |
1013 | return; | |
e7f23018 | 1014 | info = inline_summary (node); |
1c52c601 | 1015 | reset_inline_summary (node); |
e7f23018 | 1016 | memset (info, 0, sizeof (inline_summary_t)); |
10a5dd5d JH |
1017 | } |
1018 | ||
128e0d89 JH |
1019 | /* Remap predicate P of former function to be predicate of duplicated functoin. |
1020 | POSSIBLE_TRUTHS is clause of possible truths in the duplicated node, | |
1021 | INFO is inline summary of the duplicated node. */ | |
1022 | ||
1023 | static struct predicate | |
1024 | remap_predicate_after_duplication (struct predicate *p, | |
1025 | clause_t possible_truths, | |
1026 | struct inline_summary *info) | |
1027 | { | |
1028 | struct predicate new_predicate = true_predicate (); | |
1029 | int j; | |
1030 | for (j = 0; p->clause[j]; j++) | |
1031 | if (!(possible_truths & p->clause[j])) | |
1032 | { | |
1033 | new_predicate = false_predicate (); | |
1034 | break; | |
1035 | } | |
1036 | else | |
1037 | add_clause (info->conds, &new_predicate, | |
1038 | possible_truths & p->clause[j]); | |
1039 | return new_predicate; | |
1040 | } | |
1041 | ||
1042 | /* Same as remap_predicate_after_duplication but handle hint predicate *P. | |
1043 | Additionally care about allocating new memory slot for updated predicate | |
1044 | and set it to NULL when it becomes true or false (and thus uninteresting). | |
1045 | */ | |
1046 | ||
1047 | static void | |
1048 | remap_hint_predicate_after_duplication (struct predicate **p, | |
1049 | clause_t possible_truths, | |
1050 | struct inline_summary *info) | |
1051 | { | |
1052 | struct predicate new_predicate; | |
1053 | ||
1054 | if (!*p) | |
1055 | return; | |
1056 | ||
1057 | new_predicate = remap_predicate_after_duplication (*p, | |
1058 | possible_truths, | |
1059 | info); | |
1060 | /* We do not want to free previous predicate; it is used by node origin. */ | |
1061 | *p = NULL; | |
1062 | set_hint_predicate (p, new_predicate); | |
1063 | } | |
1064 | ||
898b8927 | 1065 | |
10a5dd5d JH |
1066 | /* Hook that is called by cgraph.c when a node is duplicated. */ |
1067 | ||
1068 | static void | |
1069 | inline_node_duplication_hook (struct cgraph_node *src, struct cgraph_node *dst, | |
1070 | ATTRIBUTE_UNUSED void *data) | |
1071 | { | |
e7f23018 | 1072 | struct inline_summary *info; |
10a5dd5d | 1073 | inline_summary_alloc (); |
e7f23018 JH |
1074 | info = inline_summary (dst); |
1075 | memcpy (info, inline_summary (src), | |
10a5dd5d | 1076 | sizeof (struct inline_summary)); |
74605a11 JH |
1077 | /* TODO: as an optimization, we may avoid copying conditions |
1078 | that are known to be false or true. */ | |
632b4f8e | 1079 | info->conds = VEC_copy (condition, gc, info->conds); |
74605a11 JH |
1080 | |
1081 | /* When there are any replacements in the function body, see if we can figure | |
1082 | out that something was optimized out. */ | |
1083 | if (ipa_node_params_vector && dst->clone.tree_map) | |
1084 | { | |
1085 | VEC(size_time_entry,gc) *entry = info->entry; | |
1086 | /* Use SRC parm info since it may not be copied yet. */ | |
1087 | struct ipa_node_params *parms_info = IPA_NODE_REF (src); | |
1088 | VEC (tree, heap) *known_vals = NULL; | |
1089 | int count = ipa_get_param_count (parms_info); | |
1090 | int i,j; | |
1091 | clause_t possible_truths; | |
1092 | struct predicate true_pred = true_predicate (); | |
1093 | size_time_entry *e; | |
1094 | int optimized_out_size = 0; | |
74605a11 JH |
1095 | bool inlined_to_p = false; |
1096 | struct cgraph_edge *edge; | |
1097 | ||
267ffce3 | 1098 | info->entry = 0; |
74605a11 JH |
1099 | VEC_safe_grow_cleared (tree, heap, known_vals, count); |
1100 | for (i = 0; i < count; i++) | |
1101 | { | |
1102 | tree t = ipa_get_param (parms_info, i); | |
1103 | struct ipa_replace_map *r; | |
1104 | ||
1105 | for (j = 0; | |
1106 | VEC_iterate (ipa_replace_map_p, dst->clone.tree_map, j, r); | |
1107 | j++) | |
1108 | { | |
1109 | if (r->old_tree == t | |
1110 | && r->replace_p | |
1111 | && !r->ref_p) | |
1112 | { | |
1113 | VEC_replace (tree, known_vals, i, r->new_tree); | |
1114 | break; | |
1115 | } | |
1116 | } | |
1117 | } | |
8810cc52 MJ |
1118 | possible_truths = evaluate_conditions_for_known_args (dst, false, |
1119 | known_vals, NULL); | |
74605a11 JH |
1120 | VEC_free (tree, heap, known_vals); |
1121 | ||
1122 | account_size_time (info, 0, 0, &true_pred); | |
1123 | ||
1124 | /* Remap size_time vectors. | |
1125 | Simplify the predicate by prunning out alternatives that are known | |
1126 | to be false. | |
9e990d14 JH |
1127 | TODO: as on optimization, we can also eliminate conditions known |
1128 | to be true. */ | |
74605a11 JH |
1129 | for (i = 0; VEC_iterate (size_time_entry, entry, i, e); i++) |
1130 | { | |
128e0d89 JH |
1131 | struct predicate new_predicate; |
1132 | new_predicate = remap_predicate_after_duplication (&e->predicate, | |
1133 | possible_truths, | |
1134 | info); | |
74605a11 | 1135 | if (false_predicate_p (&new_predicate)) |
0f378cb5 | 1136 | optimized_out_size += e->size; |
74605a11 JH |
1137 | else |
1138 | account_size_time (info, e->size, e->time, &new_predicate); | |
1139 | } | |
1140 | ||
9e990d14 JH |
1141 | /* Remap edge predicates with the same simplification as above. |
1142 | Also copy constantness arrays. */ | |
74605a11 JH |
1143 | for (edge = dst->callees; edge; edge = edge->next_callee) |
1144 | { | |
128e0d89 | 1145 | struct predicate new_predicate; |
74605a11 JH |
1146 | struct inline_edge_summary *es = inline_edge_summary (edge); |
1147 | ||
1148 | if (!edge->inline_failed) | |
1149 | inlined_to_p = true; | |
1150 | if (!es->predicate) | |
1151 | continue; | |
128e0d89 JH |
1152 | new_predicate = remap_predicate_after_duplication (es->predicate, |
1153 | possible_truths, | |
1154 | info); | |
74605a11 JH |
1155 | if (false_predicate_p (&new_predicate) |
1156 | && !false_predicate_p (es->predicate)) | |
1157 | { | |
1158 | optimized_out_size += es->call_stmt_size * INLINE_SIZE_SCALE; | |
74605a11 JH |
1159 | edge->frequency = 0; |
1160 | } | |
2daffc47 | 1161 | edge_set_predicate (edge, &new_predicate); |
74605a11 JH |
1162 | } |
1163 | ||
9e990d14 JH |
1164 | /* Remap indirect edge predicates with the same simplificaiton as above. |
1165 | Also copy constantness arrays. */ | |
74605a11 JH |
1166 | for (edge = dst->indirect_calls; edge; edge = edge->next_callee) |
1167 | { | |
128e0d89 | 1168 | struct predicate new_predicate; |
74605a11 JH |
1169 | struct inline_edge_summary *es = inline_edge_summary (edge); |
1170 | ||
128e0d89 | 1171 | gcc_checking_assert (edge->inline_failed); |
74605a11 JH |
1172 | if (!es->predicate) |
1173 | continue; | |
128e0d89 JH |
1174 | new_predicate = remap_predicate_after_duplication (es->predicate, |
1175 | possible_truths, | |
1176 | info); | |
74605a11 JH |
1177 | if (false_predicate_p (&new_predicate) |
1178 | && !false_predicate_p (es->predicate)) | |
1179 | { | |
1180 | optimized_out_size += es->call_stmt_size * INLINE_SIZE_SCALE; | |
74605a11 JH |
1181 | edge->frequency = 0; |
1182 | } | |
2daffc47 JH |
1183 | edge_set_predicate (edge, &new_predicate); |
1184 | } | |
128e0d89 JH |
1185 | remap_hint_predicate_after_duplication (&info->loop_iterations, |
1186 | possible_truths, | |
1187 | info); | |
1188 | remap_hint_predicate_after_duplication (&info->loop_stride, | |
1189 | possible_truths, | |
1190 | info); | |
74605a11 JH |
1191 | |
1192 | /* If inliner or someone after inliner will ever start producing | |
1193 | non-trivial clones, we will get trouble with lack of information | |
1194 | about updating self sizes, because size vectors already contains | |
1195 | sizes of the calees. */ | |
1196 | gcc_assert (!inlined_to_p | |
0f378cb5 | 1197 | || !optimized_out_size); |
74605a11 JH |
1198 | } |
1199 | else | |
2daffc47 JH |
1200 | { |
1201 | info->entry = VEC_copy (size_time_entry, gc, info->entry); | |
1202 | if (info->loop_iterations) | |
1203 | { | |
1204 | predicate p = *info->loop_iterations; | |
128e0d89 JH |
1205 | info->loop_iterations = NULL; |
1206 | set_hint_predicate (&info->loop_iterations, p); | |
1207 | } | |
1208 | if (info->loop_stride) | |
1209 | { | |
1210 | predicate p = *info->loop_stride; | |
1211 | info->loop_stride = NULL; | |
1212 | set_hint_predicate (&info->loop_stride, p); | |
2daffc47 JH |
1213 | } |
1214 | } | |
0f378cb5 | 1215 | inline_update_overall_summary (dst); |
632b4f8e JH |
1216 | } |
1217 | ||
1218 | ||
898b8927 JH |
1219 | /* Hook that is called by cgraph.c when a node is duplicated. */ |
1220 | ||
1221 | static void | |
1222 | inline_edge_duplication_hook (struct cgraph_edge *src, struct cgraph_edge *dst, | |
1223 | ATTRIBUTE_UNUSED void *data) | |
1224 | { | |
1225 | struct inline_edge_summary *info; | |
991278ab | 1226 | struct inline_edge_summary *srcinfo; |
898b8927 JH |
1227 | inline_summary_alloc (); |
1228 | info = inline_edge_summary (dst); | |
991278ab JH |
1229 | srcinfo = inline_edge_summary (src); |
1230 | memcpy (info, srcinfo, | |
898b8927 | 1231 | sizeof (struct inline_edge_summary)); |
991278ab JH |
1232 | info->predicate = NULL; |
1233 | edge_set_predicate (dst, srcinfo->predicate); | |
25837a2f | 1234 | info->param = VEC_copy (inline_param_summary_t, heap, srcinfo->param); |
898b8927 JH |
1235 | } |
1236 | ||
1237 | ||
632b4f8e JH |
1238 | /* Keep edge cache consistent across edge removal. */ |
1239 | ||
1240 | static void | |
1241 | inline_edge_removal_hook (struct cgraph_edge *edge, void *data ATTRIBUTE_UNUSED) | |
1242 | { | |
898b8927 JH |
1243 | if (edge_growth_cache) |
1244 | reset_edge_growth_cache (edge); | |
78e5ce9f | 1245 | reset_inline_edge_summary (edge); |
632b4f8e JH |
1246 | } |
1247 | ||
1248 | ||
1249 | /* Initialize growth caches. */ | |
1250 | ||
1251 | void | |
1252 | initialize_growth_caches (void) | |
1253 | { | |
632b4f8e JH |
1254 | if (cgraph_edge_max_uid) |
1255 | VEC_safe_grow_cleared (edge_growth_cache_entry, heap, edge_growth_cache, | |
1256 | cgraph_edge_max_uid); | |
1257 | if (cgraph_max_uid) | |
1258 | VEC_safe_grow_cleared (int, heap, node_growth_cache, cgraph_max_uid); | |
1259 | } | |
1260 | ||
1261 | ||
1262 | /* Free growth caches. */ | |
1263 | ||
1264 | void | |
1265 | free_growth_caches (void) | |
1266 | { | |
632b4f8e JH |
1267 | VEC_free (edge_growth_cache_entry, heap, edge_growth_cache); |
1268 | edge_growth_cache = 0; | |
1269 | VEC_free (int, heap, node_growth_cache); | |
1270 | node_growth_cache = 0; | |
10a5dd5d JH |
1271 | } |
1272 | ||
632b4f8e | 1273 | |
898b8927 JH |
1274 | /* Dump edge summaries associated to NODE and recursively to all clones. |
1275 | Indent by INDENT. */ | |
1276 | ||
1277 | static void | |
991278ab JH |
1278 | dump_inline_edge_summary (FILE * f, int indent, struct cgraph_node *node, |
1279 | struct inline_summary *info) | |
898b8927 JH |
1280 | { |
1281 | struct cgraph_edge *edge; | |
1282 | for (edge = node->callees; edge; edge = edge->next_callee) | |
1283 | { | |
1284 | struct inline_edge_summary *es = inline_edge_summary (edge); | |
a5b1779f | 1285 | struct cgraph_node *callee = cgraph_function_or_thunk_node (edge->callee, NULL); |
25837a2f JH |
1286 | int i; |
1287 | ||
09dfe187 | 1288 | fprintf (f, "%*s%s/%i %s\n%*s loop depth:%2i freq:%4i size:%2i time: %2i callee size:%2i stack:%2i", |
a5b1779f JH |
1289 | indent, "", cgraph_node_name (callee), |
1290 | callee->uid, | |
991278ab | 1291 | !edge->inline_failed ? "inlined" |
898b8927 JH |
1292 | : cgraph_inline_failed_string (edge->inline_failed), |
1293 | indent, "", | |
1294 | es->loop_depth, | |
1295 | edge->frequency, | |
1296 | es->call_stmt_size, | |
09dfe187 | 1297 | es->call_stmt_time, |
25837a2f | 1298 | (int)inline_summary (callee)->size / INLINE_SIZE_SCALE, |
a5b1779f | 1299 | (int)inline_summary (callee)->estimated_stack_size); |
25837a2f | 1300 | |
991278ab JH |
1301 | if (es->predicate) |
1302 | { | |
1303 | fprintf (f, " predicate: "); | |
1304 | dump_predicate (f, info->conds, es->predicate); | |
1305 | } | |
1306 | else | |
1307 | fprintf (f, "\n"); | |
25837a2f JH |
1308 | if (es->param) |
1309 | for (i = 0; i < (int)VEC_length (inline_param_summary_t, es->param); | |
1310 | i++) | |
1311 | { | |
1312 | int prob = VEC_index (inline_param_summary_t, | |
0823efed | 1313 | es->param, i).change_prob; |
25837a2f JH |
1314 | |
1315 | if (!prob) | |
1316 | fprintf (f, "%*s op%i is compile time invariant\n", | |
1317 | indent + 2, "", i); | |
1318 | else if (prob != REG_BR_PROB_BASE) | |
1319 | fprintf (f, "%*s op%i change %f%% of time\n", indent + 2, "", i, | |
1320 | prob * 100.0 / REG_BR_PROB_BASE); | |
1321 | } | |
898b8927 | 1322 | if (!edge->inline_failed) |
09dfe187 | 1323 | { |
25837a2f JH |
1324 | fprintf (f, "%*sStack frame offset %i, callee self size %i," |
1325 | " callee size %i\n", | |
09dfe187 | 1326 | indent+2, "", |
a5b1779f JH |
1327 | (int)inline_summary (callee)->stack_frame_offset, |
1328 | (int)inline_summary (callee)->estimated_self_stack_size, | |
1329 | (int)inline_summary (callee)->estimated_stack_size); | |
1330 | dump_inline_edge_summary (f, indent+2, callee, info); | |
09dfe187 | 1331 | } |
898b8927 JH |
1332 | } |
1333 | for (edge = node->indirect_calls; edge; edge = edge->next_callee) | |
1334 | { | |
1335 | struct inline_edge_summary *es = inline_edge_summary (edge); | |
9e990d14 | 1336 | fprintf (f, "%*sindirect call loop depth:%2i freq:%4i size:%2i" |
25837a2f | 1337 | " time: %2i", |
898b8927 JH |
1338 | indent, "", |
1339 | es->loop_depth, | |
1340 | edge->frequency, | |
1341 | es->call_stmt_size, | |
1342 | es->call_stmt_time); | |
991278ab JH |
1343 | if (es->predicate) |
1344 | { | |
1345 | fprintf (f, "predicate: "); | |
1346 | dump_predicate (f, info->conds, es->predicate); | |
1347 | } | |
1348 | else | |
9e990d14 | 1349 | fprintf (f, "\n"); |
898b8927 JH |
1350 | } |
1351 | } | |
1352 | ||
1353 | ||
09dfe187 | 1354 | void |
632b4f8e | 1355 | dump_inline_summary (FILE * f, struct cgraph_node *node) |
10a5dd5d JH |
1356 | { |
1357 | if (node->analyzed) | |
1358 | { | |
1359 | struct inline_summary *s = inline_summary (node); | |
632b4f8e JH |
1360 | size_time_entry *e; |
1361 | int i; | |
e7f23018 | 1362 | fprintf (f, "Inline summary for %s/%i", cgraph_node_name (node), |
10a5dd5d | 1363 | node->uid); |
960bfb69 | 1364 | if (DECL_DISREGARD_INLINE_LIMITS (node->symbol.decl)) |
e7f23018 JH |
1365 | fprintf (f, " always_inline"); |
1366 | if (s->inlinable) | |
1367 | fprintf (f, " inlinable"); | |
632b4f8e JH |
1368 | fprintf (f, "\n self time: %i\n", |
1369 | s->self_time); | |
e7f23018 | 1370 | fprintf (f, " global time: %i\n", s->time); |
632b4f8e JH |
1371 | fprintf (f, " self size: %i\n", |
1372 | s->self_size); | |
4c0f7679 | 1373 | fprintf (f, " global size: %i\n", s->size); |
10a5dd5d | 1374 | fprintf (f, " self stack: %i\n", |
632b4f8e JH |
1375 | (int) s->estimated_self_stack_size); |
1376 | fprintf (f, " global stack: %i\n", | |
1377 | (int) s->estimated_stack_size); | |
bf3f6510 JH |
1378 | if (s->scc_no) |
1379 | fprintf (f, " In SCC: %i\n", | |
1380 | (int) s->scc_no); | |
632b4f8e JH |
1381 | for (i = 0; |
1382 | VEC_iterate (size_time_entry, s->entry, i, e); | |
1383 | i++) | |
1384 | { | |
1385 | fprintf (f, " size:%f, time:%f, predicate:", | |
1386 | (double) e->size / INLINE_SIZE_SCALE, | |
1387 | (double) e->time / INLINE_TIME_SCALE); | |
1388 | dump_predicate (f, s->conds, &e->predicate); | |
1389 | } | |
2daffc47 JH |
1390 | if (s->loop_iterations) |
1391 | { | |
1392 | fprintf (f, " loop iterations:"); | |
1393 | dump_predicate (f, s->conds, s->loop_iterations); | |
1394 | } | |
128e0d89 JH |
1395 | if (s->loop_stride) |
1396 | { | |
1397 | fprintf (f, " loop stride:"); | |
1398 | dump_predicate (f, s->conds, s->loop_stride); | |
1399 | } | |
898b8927 | 1400 | fprintf (f, " calls:\n"); |
991278ab | 1401 | dump_inline_edge_summary (f, 4, node, s); |
632b4f8e | 1402 | fprintf (f, "\n"); |
10a5dd5d JH |
1403 | } |
1404 | } | |
1405 | ||
09dfe187 | 1406 | DEBUG_FUNCTION void |
10a5dd5d JH |
1407 | debug_inline_summary (struct cgraph_node *node) |
1408 | { | |
1409 | dump_inline_summary (stderr, node); | |
1410 | } | |
1411 | ||
1412 | void | |
1413 | dump_inline_summaries (FILE *f) | |
1414 | { | |
1415 | struct cgraph_node *node; | |
1416 | ||
65c70e6b JH |
1417 | FOR_EACH_DEFINED_FUNCTION (node) |
1418 | if (!node->global.inlined_to) | |
10a5dd5d JH |
1419 | dump_inline_summary (f, node); |
1420 | } | |
03dfc36d | 1421 | |
e7f23018 JH |
1422 | /* Give initial reasons why inlining would fail on EDGE. This gets either |
1423 | nullified or usually overwritten by more precise reasons later. */ | |
1424 | ||
1425 | void | |
1426 | initialize_inline_failed (struct cgraph_edge *e) | |
1427 | { | |
1428 | struct cgraph_node *callee = e->callee; | |
1429 | ||
1430 | if (e->indirect_unknown_callee) | |
1431 | e->inline_failed = CIF_INDIRECT_UNKNOWN_CALL; | |
1432 | else if (!callee->analyzed) | |
1433 | e->inline_failed = CIF_BODY_NOT_AVAILABLE; | |
1434 | else if (callee->local.redefined_extern_inline) | |
1435 | e->inline_failed = CIF_REDEFINED_EXTERN_INLINE; | |
89faf322 | 1436 | else if (e->call_stmt_cannot_inline_p) |
e7f23018 JH |
1437 | e->inline_failed = CIF_MISMATCHED_ARGUMENTS; |
1438 | else | |
1439 | e->inline_failed = CIF_FUNCTION_NOT_CONSIDERED; | |
1440 | } | |
1441 | ||
a61bd030 JH |
1442 | /* Callback of walk_aliased_vdefs. Flags that it has been invoked to the |
1443 | boolean variable pointed to by DATA. */ | |
1444 | ||
1445 | static bool | |
1446 | mark_modified (ao_ref *ao ATTRIBUTE_UNUSED, tree vdef ATTRIBUTE_UNUSED, | |
1447 | void *data) | |
1448 | { | |
1449 | bool *b = (bool *) data; | |
1450 | *b = true; | |
1451 | return true; | |
1452 | } | |
1453 | ||
8810cc52 MJ |
1454 | /* If OP refers to value of function parameter, return the corresponding |
1455 | parameter. */ | |
a61bd030 JH |
1456 | |
1457 | static tree | |
8810cc52 | 1458 | unmodified_parm_1 (gimple stmt, tree op) |
a61bd030 JH |
1459 | { |
1460 | /* SSA_NAME referring to parm default def? */ | |
1461 | if (TREE_CODE (op) == SSA_NAME | |
1462 | && SSA_NAME_IS_DEFAULT_DEF (op) | |
1463 | && TREE_CODE (SSA_NAME_VAR (op)) == PARM_DECL) | |
1464 | return SSA_NAME_VAR (op); | |
1465 | /* Non-SSA parm reference? */ | |
1466 | if (TREE_CODE (op) == PARM_DECL) | |
1467 | { | |
1468 | bool modified = false; | |
1469 | ||
1470 | ao_ref refd; | |
1471 | ao_ref_init (&refd, op); | |
1472 | walk_aliased_vdefs (&refd, gimple_vuse (stmt), mark_modified, &modified, | |
1473 | NULL); | |
1474 | if (!modified) | |
1475 | return op; | |
1476 | } | |
8810cc52 MJ |
1477 | return NULL_TREE; |
1478 | } | |
1479 | ||
1480 | /* If OP refers to value of function parameter, return the corresponding | |
1481 | parameter. Also traverse chains of SSA register assignments. */ | |
1482 | ||
1483 | static tree | |
1484 | unmodified_parm (gimple stmt, tree op) | |
1485 | { | |
1486 | tree res = unmodified_parm_1 (stmt, op); | |
1487 | if (res) | |
1488 | return res; | |
1489 | ||
a61bd030 JH |
1490 | if (TREE_CODE (op) == SSA_NAME |
1491 | && !SSA_NAME_IS_DEFAULT_DEF (op) | |
1492 | && gimple_assign_single_p (SSA_NAME_DEF_STMT (op))) | |
1493 | return unmodified_parm (SSA_NAME_DEF_STMT (op), | |
1494 | gimple_assign_rhs1 (SSA_NAME_DEF_STMT (op))); | |
8810cc52 MJ |
1495 | return NULL_TREE; |
1496 | } | |
1497 | ||
1498 | /* If OP refers to a value of a function parameter or value loaded from an | |
1499 | aggregate passed to a parameter (either by value or reference), return TRUE | |
1500 | and store the number of the parameter to *INDEX_P and information whether | |
1501 | and how it has been loaded from an aggregate into *AGGPOS. INFO describes | |
1502 | the function parameters, STMT is the statement in which OP is used or | |
1503 | loaded. */ | |
1504 | ||
1505 | static bool | |
1506 | unmodified_parm_or_parm_agg_item (struct ipa_node_params *info, | |
1507 | gimple stmt, tree op, int *index_p, | |
1508 | struct agg_position_info *aggpos) | |
1509 | { | |
1510 | tree res = unmodified_parm_1 (stmt, op); | |
1511 | ||
1512 | gcc_checking_assert (aggpos); | |
1513 | if (res) | |
1514 | { | |
1515 | *index_p = ipa_get_param_decl_index (info, res); | |
1516 | if (*index_p < 0) | |
1517 | return false; | |
1518 | aggpos->agg_contents = false; | |
1519 | aggpos->by_ref = false; | |
1520 | return true; | |
1521 | } | |
1522 | ||
1523 | if (TREE_CODE (op) == SSA_NAME) | |
1524 | { | |
1525 | if (SSA_NAME_IS_DEFAULT_DEF (op) | |
1526 | || !gimple_assign_single_p (SSA_NAME_DEF_STMT (op))) | |
1527 | return false; | |
1528 | stmt = SSA_NAME_DEF_STMT (op); | |
1529 | op = gimple_assign_rhs1 (stmt); | |
1530 | if (!REFERENCE_CLASS_P (op)) | |
1531 | return unmodified_parm_or_parm_agg_item (info, stmt, op, index_p, | |
1532 | aggpos); | |
1533 | } | |
1534 | ||
1535 | aggpos->agg_contents = true; | |
1536 | return ipa_load_from_parm_agg (info, stmt, op, index_p, &aggpos->offset, | |
1537 | &aggpos->by_ref); | |
a61bd030 JH |
1538 | } |
1539 | ||
03dfc36d JH |
1540 | /* See if statement might disappear after inlining. |
1541 | 0 - means not eliminated | |
1542 | 1 - half of statements goes away | |
1543 | 2 - for sure it is eliminated. | |
1544 | We are not terribly sophisticated, basically looking for simple abstraction | |
1545 | penalty wrappers. */ | |
1546 | ||
1547 | static int | |
1548 | eliminated_by_inlining_prob (gimple stmt) | |
1549 | { | |
1550 | enum gimple_code code = gimple_code (stmt); | |
63cf7260 | 1551 | enum tree_code rhs_code; |
a61bd030 JH |
1552 | |
1553 | if (!optimize) | |
1554 | return 0; | |
1555 | ||
03dfc36d JH |
1556 | switch (code) |
1557 | { | |
1558 | case GIMPLE_RETURN: | |
1559 | return 2; | |
1560 | case GIMPLE_ASSIGN: | |
1561 | if (gimple_num_ops (stmt) != 2) | |
1562 | return 0; | |
1563 | ||
63cf7260 JH |
1564 | rhs_code = gimple_assign_rhs_code (stmt); |
1565 | ||
03dfc36d JH |
1566 | /* Casts of parameters, loads from parameters passed by reference |
1567 | and stores to return value or parameters are often free after | |
1568 | inlining dua to SRA and further combining. | |
1569 | Assume that half of statements goes away. */ | |
63cf7260 JH |
1570 | if (rhs_code == CONVERT_EXPR |
1571 | || rhs_code == NOP_EXPR | |
1572 | || rhs_code == VIEW_CONVERT_EXPR | |
1573 | || rhs_code == ADDR_EXPR | |
03dfc36d JH |
1574 | || gimple_assign_rhs_class (stmt) == GIMPLE_SINGLE_RHS) |
1575 | { | |
1576 | tree rhs = gimple_assign_rhs1 (stmt); | |
1577 | tree lhs = gimple_assign_lhs (stmt); | |
116b9c07 JH |
1578 | tree inner_rhs = get_base_address (rhs); |
1579 | tree inner_lhs = get_base_address (lhs); | |
03dfc36d JH |
1580 | bool rhs_free = false; |
1581 | bool lhs_free = false; | |
1582 | ||
116b9c07 JH |
1583 | if (!inner_rhs) |
1584 | inner_rhs = rhs; | |
1585 | if (!inner_lhs) | |
1586 | inner_lhs = lhs; | |
03dfc36d | 1587 | |
2ceb2339 | 1588 | /* Reads of parameter are expected to be free. */ |
a61bd030 | 1589 | if (unmodified_parm (stmt, inner_rhs)) |
03dfc36d | 1590 | rhs_free = true; |
a7dbc1cb JH |
1591 | /* Match expressions of form &this->field. Those will most likely |
1592 | combine with something upstream after inlining. */ | |
1593 | else if (TREE_CODE (inner_rhs) == ADDR_EXPR) | |
1594 | { | |
1595 | tree op = get_base_address (TREE_OPERAND (inner_rhs, 0)); | |
1596 | if (TREE_CODE (op) == PARM_DECL) | |
1597 | rhs_free = true; | |
63cf7260 JH |
1598 | else if (TREE_CODE (op) == MEM_REF |
1599 | && unmodified_parm (stmt, TREE_OPERAND (op, 0))) | |
1600 | rhs_free = true; | |
a7dbc1cb | 1601 | } |
2ceb2339 JH |
1602 | |
1603 | /* When parameter is not SSA register because its address is taken | |
1604 | and it is just copied into one, the statement will be completely | |
1605 | free after inlining (we will copy propagate backward). */ | |
1606 | if (rhs_free && is_gimple_reg (lhs)) | |
1607 | return 2; | |
1608 | ||
1609 | /* Reads of parameters passed by reference | |
1610 | expected to be free (i.e. optimized out after inlining). */ | |
1611 | if (TREE_CODE(inner_rhs) == MEM_REF | |
1612 | && unmodified_parm (stmt, TREE_OPERAND (inner_rhs, 0))) | |
1613 | rhs_free = true; | |
1614 | ||
1615 | /* Copying parameter passed by reference into gimple register is | |
1616 | probably also going to copy propagate, but we can't be quite | |
1617 | sure. */ | |
03dfc36d JH |
1618 | if (rhs_free && is_gimple_reg (lhs)) |
1619 | lhs_free = true; | |
2ceb2339 JH |
1620 | |
1621 | /* Writes to parameters, parameters passed by value and return value | |
1622 | (either dirrectly or passed via invisible reference) are free. | |
1623 | ||
1624 | TODO: We ought to handle testcase like | |
1625 | struct a {int a,b;}; | |
1626 | struct a | |
1627 | retrurnsturct (void) | |
1628 | { | |
1629 | struct a a ={1,2}; | |
1630 | return a; | |
1631 | } | |
1632 | ||
1633 | This translate into: | |
1634 | ||
1635 | retrurnsturct () | |
1636 | { | |
1637 | int a$b; | |
1638 | int a$a; | |
1639 | struct a a; | |
1640 | struct a D.2739; | |
1641 | ||
1642 | <bb 2>: | |
1643 | D.2739.a = 1; | |
1644 | D.2739.b = 2; | |
1645 | return D.2739; | |
1646 | ||
1647 | } | |
1648 | For that we either need to copy ipa-split logic detecting writes | |
1649 | to return value. */ | |
1650 | if (TREE_CODE (inner_lhs) == PARM_DECL | |
1651 | || TREE_CODE (inner_lhs) == RESULT_DECL | |
1652 | || (TREE_CODE(inner_lhs) == MEM_REF | |
1653 | && (unmodified_parm (stmt, TREE_OPERAND (inner_lhs, 0)) | |
1654 | || (TREE_CODE (TREE_OPERAND (inner_lhs, 0)) == SSA_NAME | |
70b5e7dc RG |
1655 | && SSA_NAME_VAR (TREE_OPERAND (inner_lhs, 0)) |
1656 | && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND | |
1657 | (inner_lhs, 0))) == RESULT_DECL)))) | |
03dfc36d JH |
1658 | lhs_free = true; |
1659 | if (lhs_free | |
1660 | && (is_gimple_reg (rhs) || is_gimple_min_invariant (rhs))) | |
1661 | rhs_free = true; | |
1662 | if (lhs_free && rhs_free) | |
1663 | return 1; | |
1664 | } | |
1665 | return 0; | |
1666 | default: | |
1667 | return 0; | |
1668 | } | |
1669 | } | |
1670 | ||
1671 | ||
b15c64ee JH |
1672 | /* If BB ends by a conditional we can turn into predicates, attach corresponding |
1673 | predicates to the CFG edges. */ | |
632b4f8e | 1674 | |
b15c64ee JH |
1675 | static void |
1676 | set_cond_stmt_execution_predicate (struct ipa_node_params *info, | |
1677 | struct inline_summary *summary, | |
1678 | basic_block bb) | |
632b4f8e | 1679 | { |
632b4f8e JH |
1680 | gimple last; |
1681 | tree op; | |
1682 | int index; | |
8810cc52 | 1683 | struct agg_position_info aggpos; |
b15c64ee JH |
1684 | enum tree_code code, inverted_code; |
1685 | edge e; | |
1686 | edge_iterator ei; | |
1687 | gimple set_stmt; | |
1688 | tree op2; | |
632b4f8e | 1689 | |
b15c64ee | 1690 | last = last_stmt (bb); |
632b4f8e JH |
1691 | if (!last |
1692 | || gimple_code (last) != GIMPLE_COND) | |
b15c64ee | 1693 | return; |
632b4f8e | 1694 | if (!is_gimple_ip_invariant (gimple_cond_rhs (last))) |
b15c64ee | 1695 | return; |
632b4f8e JH |
1696 | op = gimple_cond_lhs (last); |
1697 | /* TODO: handle conditionals like | |
1698 | var = op0 < 4; | |
b15c64ee | 1699 | if (var != 0). */ |
8810cc52 | 1700 | if (unmodified_parm_or_parm_agg_item (info, last, op, &index, &aggpos)) |
b15c64ee | 1701 | { |
b15c64ee | 1702 | code = gimple_cond_code (last); |
9e990d14 JH |
1703 | inverted_code |
1704 | = invert_tree_comparison (code, | |
1705 | HONOR_NANS (TYPE_MODE (TREE_TYPE (op)))); | |
b15c64ee JH |
1706 | |
1707 | FOR_EACH_EDGE (e, ei, bb->succs) | |
1708 | { | |
8810cc52 | 1709 | struct predicate p = add_condition (summary, index, &aggpos, |
b15c64ee JH |
1710 | e->flags & EDGE_TRUE_VALUE |
1711 | ? code : inverted_code, | |
1712 | gimple_cond_rhs (last)); | |
1713 | e->aux = pool_alloc (edge_predicate_pool); | |
1714 | *(struct predicate *)e->aux = p; | |
1715 | } | |
1716 | } | |
1717 | ||
a61bd030 JH |
1718 | if (TREE_CODE (op) != SSA_NAME) |
1719 | return; | |
b15c64ee JH |
1720 | /* Special case |
1721 | if (builtin_constant_p (op)) | |
1722 | constant_code | |
1723 | else | |
1724 | nonconstant_code. | |
1725 | Here we can predicate nonconstant_code. We can't | |
1726 | really handle constant_code since we have no predicate | |
1727 | for this and also the constant code is not known to be | |
1728 | optimized away when inliner doen't see operand is constant. | |
1729 | Other optimizers might think otherwise. */ | |
8810cc52 MJ |
1730 | if (gimple_cond_code (last) != NE_EXPR |
1731 | || !integer_zerop (gimple_cond_rhs (last))) | |
1732 | return; | |
b15c64ee JH |
1733 | set_stmt = SSA_NAME_DEF_STMT (op); |
1734 | if (!gimple_call_builtin_p (set_stmt, BUILT_IN_CONSTANT_P) | |
1735 | || gimple_call_num_args (set_stmt) != 1) | |
1736 | return; | |
1737 | op2 = gimple_call_arg (set_stmt, 0); | |
8810cc52 | 1738 | if (!unmodified_parm_or_parm_agg_item (info, set_stmt, op2, &index, &aggpos)) |
b15c64ee JH |
1739 | return; |
1740 | FOR_EACH_EDGE (e, ei, bb->succs) | |
1741 | if (e->flags & EDGE_FALSE_VALUE) | |
1742 | { | |
8810cc52 MJ |
1743 | struct predicate p = add_condition (summary, index, &aggpos, |
1744 | IS_NOT_CONSTANT, NULL_TREE); | |
b15c64ee JH |
1745 | e->aux = pool_alloc (edge_predicate_pool); |
1746 | *(struct predicate *)e->aux = p; | |
1747 | } | |
1748 | } | |
1749 | ||
1750 | ||
1751 | /* If BB ends by a switch we can turn into predicates, attach corresponding | |
1752 | predicates to the CFG edges. */ | |
1753 | ||
1754 | static void | |
1755 | set_switch_stmt_execution_predicate (struct ipa_node_params *info, | |
1756 | struct inline_summary *summary, | |
1757 | basic_block bb) | |
1758 | { | |
1759 | gimple last; | |
1760 | tree op; | |
1761 | int index; | |
8810cc52 | 1762 | struct agg_position_info aggpos; |
b15c64ee JH |
1763 | edge e; |
1764 | edge_iterator ei; | |
1765 | size_t n; | |
1766 | size_t case_idx; | |
1767 | ||
1768 | last = last_stmt (bb); | |
1769 | if (!last | |
1770 | || gimple_code (last) != GIMPLE_SWITCH) | |
1771 | return; | |
1772 | op = gimple_switch_index (last); | |
8810cc52 | 1773 | if (!unmodified_parm_or_parm_agg_item (info, last, op, &index, &aggpos)) |
b15c64ee | 1774 | return; |
632b4f8e | 1775 | |
b15c64ee JH |
1776 | FOR_EACH_EDGE (e, ei, bb->succs) |
1777 | { | |
1778 | e->aux = pool_alloc (edge_predicate_pool); | |
1779 | *(struct predicate *)e->aux = false_predicate (); | |
1780 | } | |
1781 | n = gimple_switch_num_labels(last); | |
1782 | for (case_idx = 0; case_idx < n; ++case_idx) | |
1783 | { | |
1784 | tree cl = gimple_switch_label (last, case_idx); | |
1785 | tree min, max; | |
1786 | struct predicate p; | |
632b4f8e | 1787 | |
b15c64ee JH |
1788 | e = find_edge (bb, label_to_block (CASE_LABEL (cl))); |
1789 | min = CASE_LOW (cl); | |
1790 | max = CASE_HIGH (cl); | |
1791 | ||
1792 | /* For default we might want to construct predicate that none | |
1793 | of cases is met, but it is bit hard to do not having negations | |
1794 | of conditionals handy. */ | |
1795 | if (!min && !max) | |
1796 | p = true_predicate (); | |
1797 | else if (!max) | |
8810cc52 | 1798 | p = add_condition (summary, index, &aggpos, EQ_EXPR, min); |
b15c64ee JH |
1799 | else |
1800 | { | |
1801 | struct predicate p1, p2; | |
8810cc52 MJ |
1802 | p1 = add_condition (summary, index, &aggpos, GE_EXPR, min); |
1803 | p2 = add_condition (summary, index, &aggpos, LE_EXPR, max); | |
a61bd030 | 1804 | p = and_predicates (summary->conds, &p1, &p2); |
b15c64ee JH |
1805 | } |
1806 | *(struct predicate *)e->aux | |
a61bd030 | 1807 | = or_predicates (summary->conds, &p, (struct predicate *)e->aux); |
b15c64ee JH |
1808 | } |
1809 | } | |
1810 | ||
1811 | ||
1812 | /* For each BB in NODE attach to its AUX pointer predicate under | |
1813 | which it is executable. */ | |
1814 | ||
1815 | static void | |
1816 | compute_bb_predicates (struct cgraph_node *node, | |
1817 | struct ipa_node_params *parms_info, | |
1818 | struct inline_summary *summary) | |
1819 | { | |
960bfb69 | 1820 | struct function *my_function = DECL_STRUCT_FUNCTION (node->symbol.decl); |
b15c64ee JH |
1821 | bool done = false; |
1822 | basic_block bb; | |
1823 | ||
1824 | FOR_EACH_BB_FN (bb, my_function) | |
1825 | { | |
1826 | set_cond_stmt_execution_predicate (parms_info, summary, bb); | |
1827 | set_switch_stmt_execution_predicate (parms_info, summary, bb); | |
1828 | } | |
1829 | ||
1830 | /* Entry block is always executable. */ | |
9e990d14 JH |
1831 | ENTRY_BLOCK_PTR_FOR_FUNCTION (my_function)->aux |
1832 | = pool_alloc (edge_predicate_pool); | |
b15c64ee JH |
1833 | *(struct predicate *)ENTRY_BLOCK_PTR_FOR_FUNCTION (my_function)->aux |
1834 | = true_predicate (); | |
1835 | ||
1836 | /* A simple dataflow propagation of predicates forward in the CFG. | |
1837 | TODO: work in reverse postorder. */ | |
1838 | while (!done) | |
1839 | { | |
1840 | done = true; | |
1841 | FOR_EACH_BB_FN (bb, my_function) | |
1842 | { | |
1843 | struct predicate p = false_predicate (); | |
1844 | edge e; | |
1845 | edge_iterator ei; | |
1846 | FOR_EACH_EDGE (e, ei, bb->preds) | |
1847 | { | |
1848 | if (e->src->aux) | |
1849 | { | |
9e990d14 JH |
1850 | struct predicate this_bb_predicate |
1851 | = *(struct predicate *)e->src->aux; | |
b15c64ee | 1852 | if (e->aux) |
9e990d14 JH |
1853 | this_bb_predicate |
1854 | = and_predicates (summary->conds, &this_bb_predicate, | |
1855 | (struct predicate *)e->aux); | |
a61bd030 | 1856 | p = or_predicates (summary->conds, &p, &this_bb_predicate); |
b15c64ee JH |
1857 | if (true_predicate_p (&p)) |
1858 | break; | |
1859 | } | |
1860 | } | |
1861 | if (false_predicate_p (&p)) | |
1862 | gcc_assert (!bb->aux); | |
1863 | else | |
1864 | { | |
1865 | if (!bb->aux) | |
1866 | { | |
1867 | done = false; | |
1868 | bb->aux = pool_alloc (edge_predicate_pool); | |
1869 | *((struct predicate *)bb->aux) = p; | |
1870 | } | |
1871 | else if (!predicates_equal_p (&p, (struct predicate *)bb->aux)) | |
1872 | { | |
1873 | done = false; | |
1874 | *((struct predicate *)bb->aux) = p; | |
1875 | } | |
1876 | } | |
1877 | } | |
1878 | } | |
632b4f8e JH |
1879 | } |
1880 | ||
970dabbd JH |
1881 | |
1882 | /* We keep info about constantness of SSA names. */ | |
1883 | ||
1884 | typedef struct predicate predicate_t; | |
1885 | DEF_VEC_O (predicate_t); | |
1886 | DEF_VEC_ALLOC_O (predicate_t, heap); | |
2daffc47 JH |
1887 | /* Return predicate specifying when the STMT might have result that is not |
1888 | a compile time constant. */ | |
1889 | ||
1890 | static struct predicate | |
1891 | will_be_nonconstant_expr_predicate (struct ipa_node_params *info, | |
1892 | struct inline_summary *summary, | |
1893 | tree expr, | |
1894 | VEC (predicate_t, heap) *nonconstant_names) | |
1895 | { | |
1896 | tree parm; | |
1897 | int index; | |
1898 | ||
1899 | while (UNARY_CLASS_P (expr)) | |
1900 | expr = TREE_OPERAND (expr, 0); | |
1901 | ||
1902 | parm = unmodified_parm (NULL, expr); | |
1903 | if (parm | |
1904 | && (index = ipa_get_param_decl_index (info, parm)) >= 0) | |
1905 | return add_condition (summary, index, NULL, CHANGED, NULL_TREE); | |
1906 | if (is_gimple_min_invariant (expr)) | |
1907 | return false_predicate (); | |
1908 | if (TREE_CODE (expr) == SSA_NAME) | |
1909 | return VEC_index (predicate_t, nonconstant_names, | |
1910 | SSA_NAME_VERSION (expr)); | |
128e0d89 JH |
1911 | if (BINARY_CLASS_P (expr) |
1912 | || COMPARISON_CLASS_P (expr)) | |
1913 | { | |
1914 | struct predicate p1 = will_be_nonconstant_expr_predicate | |
1915 | (info, summary, TREE_OPERAND (expr, 0), | |
1916 | nonconstant_names); | |
1917 | struct predicate p2; | |
1918 | if (true_predicate_p (&p1)) | |
1919 | return p1; | |
1920 | p2 = will_be_nonconstant_expr_predicate (info, summary, | |
1921 | TREE_OPERAND (expr, 1), | |
1922 | nonconstant_names); | |
1923 | return or_predicates (summary->conds, &p1, &p2); | |
1924 | } | |
1925 | else if (TREE_CODE (expr) == COND_EXPR) | |
2daffc47 | 1926 | { |
128e0d89 JH |
1927 | struct predicate p1 = will_be_nonconstant_expr_predicate |
1928 | (info, summary, TREE_OPERAND (expr, 0), | |
1929 | nonconstant_names); | |
2daffc47 JH |
1930 | struct predicate p2; |
1931 | if (true_predicate_p (&p1)) | |
1932 | return p1; | |
128e0d89 JH |
1933 | p2 = will_be_nonconstant_expr_predicate (info, summary, |
1934 | TREE_OPERAND (expr, 1), | |
1935 | nonconstant_names); | |
1936 | if (true_predicate_p (&p2)) | |
1937 | return p2; | |
1938 | p1 = or_predicates (summary->conds, &p1, &p2); | |
1939 | p2 = will_be_nonconstant_expr_predicate (info, summary, | |
1940 | TREE_OPERAND (expr, 2), | |
1941 | nonconstant_names); | |
2daffc47 JH |
1942 | return or_predicates (summary->conds, &p1, &p2); |
1943 | } | |
1944 | else | |
1945 | { | |
1946 | debug_tree (expr); | |
1947 | gcc_unreachable (); | |
1948 | } | |
1949 | return false_predicate (); | |
1950 | } | |
970dabbd JH |
1951 | |
1952 | ||
9e990d14 JH |
1953 | /* Return predicate specifying when the STMT might have result that is not |
1954 | a compile time constant. */ | |
970dabbd | 1955 | |
632b4f8e JH |
1956 | static struct predicate |
1957 | will_be_nonconstant_predicate (struct ipa_node_params *info, | |
1958 | struct inline_summary *summary, | |
970dabbd JH |
1959 | gimple stmt, |
1960 | VEC (predicate_t, heap) *nonconstant_names) | |
632b4f8e JH |
1961 | { |
1962 | struct predicate p = true_predicate (); | |
1963 | ssa_op_iter iter; | |
1964 | tree use; | |
1965 | struct predicate op_non_const; | |
5f9f3517 | 1966 | bool is_load; |
8810cc52 MJ |
1967 | int base_index; |
1968 | struct agg_position_info aggpos; | |
632b4f8e JH |
1969 | |
1970 | /* What statments might be optimized away | |
1971 | when their arguments are constant | |
b15c64ee JH |
1972 | TODO: also trivial builtins. |
1973 | builtin_constant_p is already handled later. */ | |
632b4f8e JH |
1974 | if (gimple_code (stmt) != GIMPLE_ASSIGN |
1975 | && gimple_code (stmt) != GIMPLE_COND | |
1976 | && gimple_code (stmt) != GIMPLE_SWITCH) | |
1977 | return p; | |
1978 | ||
5f9f3517 JH |
1979 | /* Stores will stay anyway. */ |
1980 | if (gimple_vdef (stmt)) | |
632b4f8e JH |
1981 | return p; |
1982 | ||
5f9f3517 | 1983 | is_load = gimple_vuse (stmt) != NULL; |
5f9f3517 JH |
1984 | /* Loads can be optimized when the value is known. */ |
1985 | if (is_load) | |
1986 | { | |
8810cc52 | 1987 | tree op; |
5f9f3517 | 1988 | gcc_assert (gimple_assign_single_p (stmt)); |
8810cc52 MJ |
1989 | op = gimple_assign_rhs1 (stmt); |
1990 | if (!unmodified_parm_or_parm_agg_item (info, stmt, op, &base_index, | |
1991 | &aggpos)) | |
5f9f3517 JH |
1992 | return p; |
1993 | } | |
8810cc52 MJ |
1994 | else |
1995 | base_index = -1; | |
5f9f3517 | 1996 | |
632b4f8e JH |
1997 | /* See if we understand all operands before we start |
1998 | adding conditionals. */ | |
1999 | FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE) | |
2000 | { | |
a61bd030 | 2001 | tree parm = unmodified_parm (stmt, use); |
970dabbd | 2002 | /* For arguments we can build a condition. */ |
a61bd030 | 2003 | if (parm && ipa_get_param_decl_index (info, parm) >= 0) |
970dabbd | 2004 | continue; |
a61bd030 JH |
2005 | if (TREE_CODE (use) != SSA_NAME) |
2006 | return p; | |
970dabbd JH |
2007 | /* If we know when operand is constant, |
2008 | we still can say something useful. */ | |
0823efed DN |
2009 | if (!true_predicate_p (&VEC_index (predicate_t, nonconstant_names, |
2010 | SSA_NAME_VERSION (use)))) | |
970dabbd JH |
2011 | continue; |
2012 | return p; | |
632b4f8e | 2013 | } |
8810cc52 | 2014 | |
5f9f3517 | 2015 | if (is_load) |
8810cc52 MJ |
2016 | op_non_const = add_condition (summary, base_index, &aggpos, CHANGED, NULL); |
2017 | else | |
2018 | op_non_const = false_predicate (); | |
632b4f8e JH |
2019 | FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE) |
2020 | { | |
a61bd030 | 2021 | tree parm = unmodified_parm (stmt, use); |
8810cc52 MJ |
2022 | int index; |
2023 | ||
2024 | if (parm | |
2025 | && (index = ipa_get_param_decl_index (info, parm)) >= 0) | |
2026 | { | |
2027 | if (index != base_index) | |
2028 | p = add_condition (summary, index, NULL, CHANGED, NULL_TREE); | |
2029 | else | |
2030 | continue; | |
2031 | } | |
970dabbd | 2032 | else |
0823efed DN |
2033 | p = VEC_index (predicate_t, nonconstant_names, |
2034 | SSA_NAME_VERSION (use)); | |
a61bd030 | 2035 | op_non_const = or_predicates (summary->conds, &p, &op_non_const); |
632b4f8e | 2036 | } |
970dabbd JH |
2037 | if (gimple_code (stmt) == GIMPLE_ASSIGN |
2038 | && TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) | |
2039 | VEC_replace (predicate_t, nonconstant_names, | |
0823efed | 2040 | SSA_NAME_VERSION (gimple_assign_lhs (stmt)), op_non_const); |
632b4f8e JH |
2041 | return op_non_const; |
2042 | } | |
2043 | ||
25837a2f JH |
2044 | struct record_modified_bb_info |
2045 | { | |
2046 | bitmap bb_set; | |
2047 | gimple stmt; | |
2048 | }; | |
2049 | ||
2050 | /* Callback of walk_aliased_vdefs. Records basic blocks where the value may be | |
2051 | set except for info->stmt. */ | |
2052 | ||
2053 | static bool | |
2054 | record_modified (ao_ref *ao ATTRIBUTE_UNUSED, tree vdef, | |
2055 | void *data) | |
2056 | { | |
2057 | struct record_modified_bb_info *info = (struct record_modified_bb_info *) data; | |
2058 | if (SSA_NAME_DEF_STMT (vdef) == info->stmt) | |
2059 | return false; | |
2060 | bitmap_set_bit (info->bb_set, | |
2061 | SSA_NAME_IS_DEFAULT_DEF (vdef) | |
2062 | ? ENTRY_BLOCK_PTR->index : gimple_bb (SSA_NAME_DEF_STMT (vdef))->index); | |
2063 | return false; | |
2064 | } | |
2065 | ||
2066 | /* Return probability (based on REG_BR_PROB_BASE) that I-th parameter of STMT | |
2067 | will change since last invocation of STMT. | |
2068 | ||
2069 | Value 0 is reserved for compile time invariants. | |
2070 | For common parameters it is REG_BR_PROB_BASE. For loop invariants it | |
2071 | ought to be REG_BR_PROB_BASE / estimated_iters. */ | |
2072 | ||
2073 | static int | |
2074 | param_change_prob (gimple stmt, int i) | |
2075 | { | |
2076 | tree op = gimple_call_arg (stmt, i); | |
2077 | basic_block bb = gimple_bb (stmt); | |
2078 | tree base; | |
2079 | ||
2080 | if (is_gimple_min_invariant (op)) | |
2081 | return 0; | |
2082 | /* We would have to do non-trivial analysis to really work out what | |
2083 | is the probability of value to change (i.e. when init statement | |
2084 | is in a sibling loop of the call). | |
2085 | ||
2086 | We do an conservative estimate: when call is executed N times more often | |
2087 | than the statement defining value, we take the frequency 1/N. */ | |
2088 | if (TREE_CODE (op) == SSA_NAME) | |
2089 | { | |
2090 | int init_freq; | |
2091 | ||
2092 | if (!bb->frequency) | |
2093 | return REG_BR_PROB_BASE; | |
2094 | ||
2095 | if (SSA_NAME_IS_DEFAULT_DEF (op)) | |
2096 | init_freq = ENTRY_BLOCK_PTR->frequency; | |
2097 | else | |
2098 | init_freq = gimple_bb (SSA_NAME_DEF_STMT (op))->frequency; | |
2099 | ||
2100 | if (!init_freq) | |
2101 | init_freq = 1; | |
2102 | if (init_freq < bb->frequency) | |
2103 | return MAX ((init_freq * REG_BR_PROB_BASE + | |
2104 | bb->frequency / 2) / bb->frequency, 1); | |
2105 | else | |
2106 | return REG_BR_PROB_BASE; | |
2107 | } | |
2108 | ||
2109 | base = get_base_address (op); | |
2110 | if (base) | |
2111 | { | |
2112 | ao_ref refd; | |
2113 | int max; | |
2114 | struct record_modified_bb_info info; | |
2115 | bitmap_iterator bi; | |
2116 | unsigned index; | |
2117 | ||
2118 | if (const_value_known_p (base)) | |
2119 | return 0; | |
2120 | if (!bb->frequency) | |
2121 | return REG_BR_PROB_BASE; | |
2122 | ao_ref_init (&refd, op); | |
2123 | info.stmt = stmt; | |
2124 | info.bb_set = BITMAP_ALLOC (NULL); | |
2125 | walk_aliased_vdefs (&refd, gimple_vuse (stmt), record_modified, &info, | |
2126 | NULL); | |
2127 | if (bitmap_bit_p (info.bb_set, bb->index)) | |
2128 | { | |
2129 | BITMAP_FREE (info.bb_set); | |
2130 | return REG_BR_PROB_BASE; | |
2131 | } | |
2132 | ||
2133 | /* Assume that every memory is initialized at entry. | |
2134 | TODO: Can we easilly determine if value is always defined | |
2135 | and thus we may skip entry block? */ | |
2136 | if (ENTRY_BLOCK_PTR->frequency) | |
2137 | max = ENTRY_BLOCK_PTR->frequency; | |
2138 | else | |
2139 | max = 1; | |
2140 | ||
2141 | EXECUTE_IF_SET_IN_BITMAP (info.bb_set, 0, index, bi) | |
2142 | max = MIN (max, BASIC_BLOCK (index)->frequency); | |
2143 | ||
2144 | BITMAP_FREE (info.bb_set); | |
2145 | if (max < bb->frequency) | |
2146 | return MAX ((max * REG_BR_PROB_BASE + | |
2147 | bb->frequency / 2) / bb->frequency, 1); | |
2148 | else | |
2149 | return REG_BR_PROB_BASE; | |
2150 | } | |
2151 | return REG_BR_PROB_BASE; | |
2152 | } | |
2153 | ||
48679f6e MJ |
2154 | /* Find whether a basic block BB is the final block of a (half) diamond CFG |
2155 | sub-graph and if the predicate the condition depends on is known. If so, | |
2156 | return true and store the pointer the predicate in *P. */ | |
2157 | ||
2158 | static bool | |
2159 | phi_result_unknown_predicate (struct ipa_node_params *info, | |
2160 | struct inline_summary *summary, basic_block bb, | |
2161 | struct predicate *p, | |
2162 | VEC (predicate_t, heap) *nonconstant_names) | |
2163 | { | |
2164 | edge e; | |
2165 | edge_iterator ei; | |
2166 | basic_block first_bb = NULL; | |
2167 | gimple stmt; | |
2168 | ||
2169 | if (single_pred_p (bb)) | |
2170 | { | |
2171 | *p = false_predicate (); | |
2172 | return true; | |
2173 | } | |
2174 | ||
2175 | FOR_EACH_EDGE (e, ei, bb->preds) | |
2176 | { | |
2177 | if (single_succ_p (e->src)) | |
2178 | { | |
2179 | if (!single_pred_p (e->src)) | |
2180 | return false; | |
2181 | if (!first_bb) | |
2182 | first_bb = single_pred (e->src); | |
2183 | else if (single_pred (e->src) != first_bb) | |
2184 | return false; | |
2185 | } | |
2186 | else | |
2187 | { | |
2188 | if (!first_bb) | |
2189 | first_bb = e->src; | |
2190 | else if (e->src != first_bb) | |
2191 | return false; | |
2192 | } | |
2193 | } | |
2194 | ||
2195 | if (!first_bb) | |
2196 | return false; | |
2197 | ||
2198 | stmt = last_stmt (first_bb); | |
2199 | if (!stmt | |
2200 | || gimple_code (stmt) != GIMPLE_COND | |
2201 | || !is_gimple_ip_invariant (gimple_cond_rhs (stmt))) | |
2202 | return false; | |
2203 | ||
2204 | *p = will_be_nonconstant_expr_predicate (info, summary, | |
2205 | gimple_cond_lhs (stmt), | |
2206 | nonconstant_names); | |
2207 | if (true_predicate_p (p)) | |
2208 | return false; | |
2209 | else | |
2210 | return true; | |
2211 | } | |
2212 | ||
2213 | /* Given a PHI statement in a function described by inline properties SUMMARY | |
2214 | and *P being the predicate describing whether the selected PHI argument is | |
2215 | known, store a predicate for the result of the PHI statement into | |
2216 | NONCONSTANT_NAMES, if possible. */ | |
2217 | ||
2218 | static void | |
2219 | predicate_for_phi_result (struct inline_summary *summary, gimple phi, | |
2220 | struct predicate *p, | |
2221 | VEC (predicate_t, heap) *nonconstant_names) | |
2222 | { | |
2223 | unsigned i; | |
2224 | ||
2225 | for (i = 0; i < gimple_phi_num_args (phi); i++) | |
2226 | { | |
2227 | tree arg = gimple_phi_arg (phi, i)->def; | |
2228 | if (!is_gimple_min_invariant (arg)) | |
2229 | { | |
2230 | gcc_assert (TREE_CODE (arg) == SSA_NAME); | |
2231 | *p = or_predicates (summary->conds, p, | |
2232 | &VEC_index (predicate_t, nonconstant_names, | |
2233 | SSA_NAME_VERSION (arg))); | |
2234 | if (true_predicate_p (p)) | |
2235 | return; | |
2236 | } | |
2237 | } | |
2238 | ||
2239 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2240 | { | |
2241 | fprintf (dump_file, "\t\tphi predicate: "); | |
2242 | dump_predicate (dump_file, summary->conds, p); | |
2243 | } | |
2244 | VEC_replace (predicate_t, nonconstant_names, | |
2245 | SSA_NAME_VERSION (gimple_phi_result (phi)), *p); | |
2246 | } | |
632b4f8e JH |
2247 | |
2248 | /* Compute function body size parameters for NODE. | |
2249 | When EARLY is true, we compute only simple summaries without | |
2250 | non-trivial predicates to drive the early inliner. */ | |
03dfc36d JH |
2251 | |
2252 | static void | |
632b4f8e | 2253 | estimate_function_body_sizes (struct cgraph_node *node, bool early) |
03dfc36d JH |
2254 | { |
2255 | gcov_type time = 0; | |
03dfc36d JH |
2256 | /* Estimate static overhead for function prologue/epilogue and alignment. */ |
2257 | int size = 2; | |
2258 | /* Benefits are scaled by probability of elimination that is in range | |
2259 | <0,2>. */ | |
03dfc36d JH |
2260 | basic_block bb; |
2261 | gimple_stmt_iterator bsi; | |
960bfb69 | 2262 | struct function *my_function = DECL_STRUCT_FUNCTION (node->symbol.decl); |
03dfc36d | 2263 | int freq; |
632b4f8e JH |
2264 | struct inline_summary *info = inline_summary (node); |
2265 | struct predicate bb_predicate; | |
970dabbd JH |
2266 | struct ipa_node_params *parms_info = NULL; |
2267 | VEC (predicate_t, heap) *nonconstant_names = NULL; | |
632b4f8e | 2268 | |
632b4f8e JH |
2269 | info->conds = 0; |
2270 | info->entry = 0; | |
2271 | ||
172e74fa MJ |
2272 | if (optimize && !early) |
2273 | { | |
2274 | calculate_dominance_info (CDI_DOMINATORS); | |
2275 | loop_optimizer_init (LOOPS_NORMAL | LOOPS_HAVE_RECORDED_EXITS); | |
be279f86 MJ |
2276 | |
2277 | if (ipa_node_params_vector) | |
2278 | { | |
2279 | parms_info = IPA_NODE_REF (node); | |
2280 | VEC_safe_grow_cleared (predicate_t, heap, nonconstant_names, | |
2281 | VEC_length (tree, SSANAMES (my_function))); | |
2282 | } | |
172e74fa | 2283 | } |
03dfc36d JH |
2284 | |
2285 | if (dump_file) | |
632b4f8e | 2286 | fprintf (dump_file, "\nAnalyzing function body size: %s\n", |
03dfc36d JH |
2287 | cgraph_node_name (node)); |
2288 | ||
632b4f8e JH |
2289 | /* When we run into maximal number of entries, we assign everything to the |
2290 | constant truth case. Be sure to have it in list. */ | |
2291 | bb_predicate = true_predicate (); | |
2292 | account_size_time (info, 0, 0, &bb_predicate); | |
2293 | ||
2294 | bb_predicate = not_inlined_predicate (); | |
2295 | account_size_time (info, 2 * INLINE_SIZE_SCALE, 0, &bb_predicate); | |
2296 | ||
03dfc36d | 2297 | gcc_assert (my_function && my_function->cfg); |
b15c64ee JH |
2298 | if (parms_info) |
2299 | compute_bb_predicates (node, parms_info, info); | |
03dfc36d JH |
2300 | FOR_EACH_BB_FN (bb, my_function) |
2301 | { | |
960bfb69 | 2302 | freq = compute_call_stmt_bb_frequency (node->symbol.decl, bb); |
632b4f8e JH |
2303 | |
2304 | /* TODO: Obviously predicates can be propagated down across CFG. */ | |
2305 | if (parms_info) | |
2306 | { | |
b15c64ee JH |
2307 | if (bb->aux) |
2308 | bb_predicate = *(struct predicate *)bb->aux; | |
2309 | else | |
2310 | bb_predicate = false_predicate (); | |
632b4f8e JH |
2311 | } |
2312 | else | |
2313 | bb_predicate = true_predicate (); | |
2314 | ||
2315 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2316 | { | |
2317 | fprintf (dump_file, "\n BB %i predicate:", bb->index); | |
2318 | dump_predicate (dump_file, info->conds, &bb_predicate); | |
2319 | } | |
48679f6e MJ |
2320 | |
2321 | if (parms_info && nonconstant_names) | |
2322 | { | |
2323 | struct predicate phi_predicate; | |
2324 | bool first_phi = true; | |
2325 | ||
2326 | for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
2327 | { | |
2328 | if (first_phi | |
2329 | && !phi_result_unknown_predicate (parms_info, info, bb, | |
2330 | &phi_predicate, | |
2331 | nonconstant_names)) | |
2332 | break; | |
2333 | first_phi = false; | |
2334 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2335 | { | |
2336 | fprintf (dump_file, " "); | |
2337 | print_gimple_stmt (dump_file, gsi_stmt (bsi), 0, 0); | |
2338 | } | |
2339 | predicate_for_phi_result (info, gsi_stmt (bsi), &phi_predicate, | |
2340 | nonconstant_names); | |
2341 | } | |
2342 | } | |
2343 | ||
03dfc36d JH |
2344 | for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) |
2345 | { | |
2346 | gimple stmt = gsi_stmt (bsi); | |
2347 | int this_size = estimate_num_insns (stmt, &eni_size_weights); | |
2348 | int this_time = estimate_num_insns (stmt, &eni_time_weights); | |
2349 | int prob; | |
b15c64ee | 2350 | struct predicate will_be_nonconstant; |
03dfc36d JH |
2351 | |
2352 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2353 | { | |
632b4f8e | 2354 | fprintf (dump_file, " "); |
03dfc36d | 2355 | print_gimple_stmt (dump_file, stmt, 0, 0); |
632b4f8e JH |
2356 | fprintf (dump_file, "\t\tfreq:%3.2f size:%3i time:%3i\n", |
2357 | ((double)freq)/CGRAPH_FREQ_BASE, this_size, this_time); | |
03dfc36d | 2358 | } |
10a5dd5d JH |
2359 | |
2360 | if (is_gimple_call (stmt)) | |
2361 | { | |
2362 | struct cgraph_edge *edge = cgraph_edge (node, stmt); | |
898b8927 JH |
2363 | struct inline_edge_summary *es = inline_edge_summary (edge); |
2364 | ||
970dabbd JH |
2365 | /* Special case: results of BUILT_IN_CONSTANT_P will be always |
2366 | resolved as constant. We however don't want to optimize | |
2367 | out the cgraph edges. */ | |
2368 | if (nonconstant_names | |
2369 | && gimple_call_builtin_p (stmt, BUILT_IN_CONSTANT_P) | |
2370 | && gimple_call_lhs (stmt) | |
2371 | && TREE_CODE (gimple_call_lhs (stmt)) == SSA_NAME) | |
2372 | { | |
2373 | struct predicate false_p = false_predicate (); | |
2374 | VEC_replace (predicate_t, nonconstant_names, | |
25837a2f | 2375 | SSA_NAME_VERSION (gimple_call_lhs (stmt)), |
0823efed | 2376 | false_p); |
25837a2f JH |
2377 | } |
2378 | if (ipa_node_params_vector) | |
2379 | { | |
2380 | int count = gimple_call_num_args (stmt); | |
2381 | int i; | |
2382 | ||
2383 | if (count) | |
2384 | VEC_safe_grow_cleared (inline_param_summary_t, heap, | |
2385 | es->param, count); | |
2386 | for (i = 0; i < count; i++) | |
2387 | { | |
2388 | int prob = param_change_prob (stmt, i); | |
2389 | gcc_assert (prob >= 0 && prob <= REG_BR_PROB_BASE); | |
2390 | VEC_index (inline_param_summary_t, | |
0823efed | 2391 | es->param, i).change_prob = prob; |
25837a2f | 2392 | } |
970dabbd JH |
2393 | } |
2394 | ||
898b8927 JH |
2395 | es->call_stmt_size = this_size; |
2396 | es->call_stmt_time = this_time; | |
391886c8 | 2397 | es->loop_depth = bb_loop_depth (bb); |
991278ab | 2398 | edge_set_predicate (edge, &bb_predicate); |
10a5dd5d JH |
2399 | } |
2400 | ||
b15c64ee JH |
2401 | /* TODO: When conditional jump or swithc is known to be constant, but |
2402 | we did not translate it into the predicates, we really can account | |
2403 | just maximum of the possible paths. */ | |
2404 | if (parms_info) | |
2405 | will_be_nonconstant | |
2406 | = will_be_nonconstant_predicate (parms_info, info, | |
2407 | stmt, nonconstant_names); | |
632b4f8e JH |
2408 | if (this_time || this_size) |
2409 | { | |
632b4f8e JH |
2410 | struct predicate p; |
2411 | ||
2412 | this_time *= freq; | |
10a5dd5d | 2413 | |
632b4f8e JH |
2414 | prob = eliminated_by_inlining_prob (stmt); |
2415 | if (prob == 1 && dump_file && (dump_flags & TDF_DETAILS)) | |
2416 | fprintf (dump_file, "\t\t50%% will be eliminated by inlining\n"); | |
2417 | if (prob == 2 && dump_file && (dump_flags & TDF_DETAILS)) | |
2ceb2339 | 2418 | fprintf (dump_file, "\t\tWill be eliminated by inlining\n"); |
632b4f8e JH |
2419 | |
2420 | if (parms_info) | |
9e990d14 JH |
2421 | p = and_predicates (info->conds, &bb_predicate, |
2422 | &will_be_nonconstant); | |
632b4f8e JH |
2423 | else |
2424 | p = true_predicate (); | |
10a5dd5d | 2425 | |
0f378cb5 JH |
2426 | if (!false_predicate_p (&p)) |
2427 | { | |
2428 | time += this_time; | |
2429 | size += this_size; | |
2430 | } | |
2431 | ||
632b4f8e | 2432 | /* We account everything but the calls. Calls have their own |
073a8998 | 2433 | size/time info attached to cgraph edges. This is necessary |
632b4f8e JH |
2434 | in order to make the cost disappear after inlining. */ |
2435 | if (!is_gimple_call (stmt)) | |
2436 | { | |
2437 | if (prob) | |
2438 | { | |
2439 | struct predicate ip = not_inlined_predicate (); | |
a61bd030 | 2440 | ip = and_predicates (info->conds, &ip, &p); |
632b4f8e JH |
2441 | account_size_time (info, this_size * prob, |
2442 | this_time * prob, &ip); | |
2443 | } | |
2444 | if (prob != 2) | |
2445 | account_size_time (info, this_size * (2 - prob), | |
2446 | this_time * (2 - prob), &p); | |
2447 | } | |
10a5dd5d | 2448 | |
632b4f8e JH |
2449 | gcc_assert (time >= 0); |
2450 | gcc_assert (size >= 0); | |
2451 | } | |
03dfc36d JH |
2452 | } |
2453 | } | |
b15c64ee JH |
2454 | FOR_ALL_BB_FN (bb, my_function) |
2455 | { | |
2456 | edge e; | |
2457 | edge_iterator ei; | |
2458 | ||
2459 | if (bb->aux) | |
2460 | pool_free (edge_predicate_pool, bb->aux); | |
2461 | bb->aux = NULL; | |
2462 | FOR_EACH_EDGE (e, ei, bb->succs) | |
2463 | { | |
2464 | if (e->aux) | |
2465 | pool_free (edge_predicate_pool, e->aux); | |
2466 | e->aux = NULL; | |
2467 | } | |
2468 | } | |
03dfc36d | 2469 | time = (time + CGRAPH_FREQ_BASE / 2) / CGRAPH_FREQ_BASE; |
03dfc36d JH |
2470 | if (time > MAX_TIME) |
2471 | time = MAX_TIME; | |
2daffc47 JH |
2472 | |
2473 | if (!early && nonconstant_names) | |
2474 | { | |
2475 | struct loop *loop; | |
2476 | loop_iterator li; | |
2477 | predicate loop_iterations = true_predicate (); | |
128e0d89 | 2478 | predicate loop_stride = true_predicate (); |
2daffc47 | 2479 | |
2daffc47 JH |
2480 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2481 | flow_loops_dump (dump_file, NULL, 0); | |
2482 | scev_initialize (); | |
2483 | FOR_EACH_LOOP (li, loop, 0) | |
2484 | { | |
2485 | VEC (edge, heap) *exits; | |
2486 | edge ex; | |
128e0d89 | 2487 | unsigned int j, i; |
2daffc47 | 2488 | struct tree_niter_desc niter_desc; |
128e0d89 | 2489 | basic_block *body = get_loop_body (loop); |
2daffc47 JH |
2490 | |
2491 | exits = get_loop_exit_edges (loop); | |
2492 | FOR_EACH_VEC_ELT (edge, exits, j, ex) | |
2493 | if (number_of_iterations_exit (loop, ex, &niter_desc, false) | |
2494 | && !is_gimple_min_invariant (niter_desc.niter)) | |
2495 | { | |
2496 | predicate will_be_nonconstant | |
2497 | = will_be_nonconstant_expr_predicate (parms_info, info, | |
2498 | niter_desc.niter, nonconstant_names); | |
2499 | if (!true_predicate_p (&will_be_nonconstant) | |
2500 | && !false_predicate_p (&will_be_nonconstant)) | |
2501 | /* This is slightly inprecise. We may want to represent each loop with | |
2502 | independent predicate. */ | |
2503 | loop_iterations = and_predicates (info->conds, &loop_iterations, &will_be_nonconstant); | |
2504 | } | |
2505 | VEC_free (edge, heap, exits); | |
128e0d89 JH |
2506 | |
2507 | for (i = 0; i < loop->num_nodes; i++) | |
2508 | { | |
2509 | gimple_stmt_iterator gsi; | |
2510 | for (gsi = gsi_start_bb (body[i]); !gsi_end_p (gsi); gsi_next (&gsi)) | |
2511 | { | |
2512 | gimple stmt = gsi_stmt (gsi); | |
2513 | affine_iv iv; | |
2514 | ssa_op_iter iter; | |
2515 | tree use; | |
2516 | ||
2517 | FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE) | |
2518 | { | |
2519 | predicate will_be_nonconstant; | |
2520 | ||
2521 | if (!simple_iv (loop, loop_containing_stmt (stmt), use, &iv, true) | |
2522 | || is_gimple_min_invariant (iv.step)) | |
2523 | continue; | |
2524 | will_be_nonconstant | |
2525 | = will_be_nonconstant_expr_predicate (parms_info, info, | |
2526 | iv.step, nonconstant_names); | |
2527 | if (!true_predicate_p (&will_be_nonconstant) | |
2528 | && !false_predicate_p (&will_be_nonconstant)) | |
2529 | /* This is slightly inprecise. We may want to represent each loop with | |
2530 | independent predicate. */ | |
2531 | loop_stride = and_predicates (info->conds, &loop_stride, &will_be_nonconstant); | |
2532 | } | |
2533 | } | |
2534 | } | |
2535 | free (body); | |
2daffc47 | 2536 | } |
128e0d89 JH |
2537 | set_hint_predicate (&inline_summary (node)->loop_iterations, loop_iterations); |
2538 | set_hint_predicate (&inline_summary (node)->loop_stride, loop_stride); | |
2daffc47 | 2539 | scev_finalize (); |
2daffc47 | 2540 | } |
03dfc36d JH |
2541 | inline_summary (node)->self_time = time; |
2542 | inline_summary (node)->self_size = size; | |
970dabbd | 2543 | VEC_free (predicate_t, heap, nonconstant_names); |
172e74fa MJ |
2544 | if (optimize && !early) |
2545 | { | |
2546 | loop_optimizer_finalize (); | |
2547 | free_dominance_info (CDI_DOMINATORS); | |
2548 | } | |
632b4f8e JH |
2549 | if (dump_file) |
2550 | { | |
2551 | fprintf (dump_file, "\n"); | |
2552 | dump_inline_summary (dump_file, node); | |
2553 | } | |
03dfc36d JH |
2554 | } |
2555 | ||
2556 | ||
632b4f8e JH |
2557 | /* Compute parameters of functions used by inliner. |
2558 | EARLY is true when we compute parameters for the early inliner */ | |
03dfc36d JH |
2559 | |
2560 | void | |
632b4f8e | 2561 | compute_inline_parameters (struct cgraph_node *node, bool early) |
03dfc36d JH |
2562 | { |
2563 | HOST_WIDE_INT self_stack_size; | |
2564 | struct cgraph_edge *e; | |
e7f23018 | 2565 | struct inline_summary *info; |
03dfc36d JH |
2566 | |
2567 | gcc_assert (!node->global.inlined_to); | |
2568 | ||
10a5dd5d JH |
2569 | inline_summary_alloc (); |
2570 | ||
e7f23018 | 2571 | info = inline_summary (node); |
1c52c601 | 2572 | reset_inline_summary (node); |
e7f23018 | 2573 | |
c47d0034 JH |
2574 | /* FIXME: Thunks are inlinable, but tree-inline don't know how to do that. |
2575 | Once this happen, we will need to more curefully predict call | |
2576 | statement size. */ | |
2577 | if (node->thunk.thunk_p) | |
2578 | { | |
2579 | struct inline_edge_summary *es = inline_edge_summary (node->callees); | |
2580 | struct predicate t = true_predicate (); | |
2581 | ||
124f1be6 | 2582 | info->inlinable = 0; |
c47d0034 JH |
2583 | node->callees->call_stmt_cannot_inline_p = true; |
2584 | node->local.can_change_signature = false; | |
2585 | es->call_stmt_time = 1; | |
2586 | es->call_stmt_size = 1; | |
2587 | account_size_time (info, 0, 0, &t); | |
2588 | return; | |
2589 | } | |
2590 | ||
5f9f3517 | 2591 | /* Even is_gimple_min_invariant rely on current_function_decl. */ |
960bfb69 | 2592 | push_cfun (DECL_STRUCT_FUNCTION (node->symbol.decl)); |
5f9f3517 | 2593 | |
03dfc36d JH |
2594 | /* Estimate the stack size for the function if we're optimizing. */ |
2595 | self_stack_size = optimize ? estimated_stack_frame_size (node) : 0; | |
e7f23018 JH |
2596 | info->estimated_self_stack_size = self_stack_size; |
2597 | info->estimated_stack_size = self_stack_size; | |
2598 | info->stack_frame_offset = 0; | |
03dfc36d JH |
2599 | |
2600 | /* Can this function be inlined at all? */ | |
960bfb69 | 2601 | info->inlinable = tree_inlinable_function_p (node->symbol.decl); |
03dfc36d | 2602 | |
201176d3 | 2603 | /* Type attributes can use parameter indices to describe them. */ |
960bfb69 | 2604 | if (TYPE_ATTRIBUTES (TREE_TYPE (node->symbol.decl))) |
201176d3 | 2605 | node->local.can_change_signature = false; |
03dfc36d JH |
2606 | else |
2607 | { | |
201176d3 MJ |
2608 | /* Otherwise, inlinable functions always can change signature. */ |
2609 | if (info->inlinable) | |
2610 | node->local.can_change_signature = true; | |
2611 | else | |
2612 | { | |
2613 | /* Functions calling builtin_apply can not change signature. */ | |
2614 | for (e = node->callees; e; e = e->next_callee) | |
2615 | { | |
960bfb69 | 2616 | tree cdecl = e->callee->symbol.decl; |
201176d3 MJ |
2617 | if (DECL_BUILT_IN (cdecl) |
2618 | && DECL_BUILT_IN_CLASS (cdecl) == BUILT_IN_NORMAL | |
2619 | && (DECL_FUNCTION_CODE (cdecl) == BUILT_IN_APPLY_ARGS | |
2620 | || DECL_FUNCTION_CODE (cdecl) == BUILT_IN_VA_START)) | |
2621 | break; | |
2622 | } | |
2623 | node->local.can_change_signature = !e; | |
2624 | } | |
03dfc36d | 2625 | } |
632b4f8e | 2626 | estimate_function_body_sizes (node, early); |
10a5dd5d | 2627 | |
03dfc36d | 2628 | /* Inlining characteristics are maintained by the cgraph_mark_inline. */ |
e7f23018 JH |
2629 | info->time = info->self_time; |
2630 | info->size = info->self_size; | |
e7f23018 JH |
2631 | info->stack_frame_offset = 0; |
2632 | info->estimated_stack_size = info->estimated_self_stack_size; | |
0f378cb5 JH |
2633 | #ifdef ENABLE_CHECKING |
2634 | inline_update_overall_summary (node); | |
2635 | gcc_assert (info->time == info->self_time | |
2636 | && info->size == info->self_size); | |
2637 | #endif | |
2638 | ||
5f9f3517 | 2639 | pop_cfun (); |
03dfc36d JH |
2640 | } |
2641 | ||
2642 | ||
2643 | /* Compute parameters of functions used by inliner using | |
2644 | current_function_decl. */ | |
2645 | ||
2646 | static unsigned int | |
2647 | compute_inline_parameters_for_current (void) | |
2648 | { | |
632b4f8e | 2649 | compute_inline_parameters (cgraph_get_node (current_function_decl), true); |
03dfc36d JH |
2650 | return 0; |
2651 | } | |
2652 | ||
2653 | struct gimple_opt_pass pass_inline_parameters = | |
2654 | { | |
2655 | { | |
2656 | GIMPLE_PASS, | |
2657 | "inline_param", /* name */ | |
2b4e6bf1 | 2658 | OPTGROUP_INLINE, /* optinfo_flags */ |
03dfc36d JH |
2659 | NULL, /* gate */ |
2660 | compute_inline_parameters_for_current,/* execute */ | |
2661 | NULL, /* sub */ | |
2662 | NULL, /* next */ | |
2663 | 0, /* static_pass_number */ | |
63ef63bf | 2664 | TV_INLINE_PARAMETERS, /* tv_id */ |
03dfc36d JH |
2665 | 0, /* properties_required */ |
2666 | 0, /* properties_provided */ | |
2667 | 0, /* properties_destroyed */ | |
2668 | 0, /* todo_flags_start */ | |
2669 | 0 /* todo_flags_finish */ | |
2670 | } | |
2671 | }; | |
2672 | ||
2673 | ||
d2d668fb MK |
2674 | /* Estimate benefit devirtualizing indirect edge IE, provided KNOWN_VALS and |
2675 | KNOWN_BINFOS. */ | |
2676 | ||
37678631 | 2677 | static bool |
d2d668fb | 2678 | estimate_edge_devirt_benefit (struct cgraph_edge *ie, |
0f378cb5 | 2679 | int *size, int *time, |
d2d668fb | 2680 | VEC (tree, heap) *known_vals, |
8810cc52 MJ |
2681 | VEC (tree, heap) *known_binfos, |
2682 | VEC (ipa_agg_jump_function_p, heap) *known_aggs) | |
d2d668fb MK |
2683 | { |
2684 | tree target; | |
37678631 JH |
2685 | struct cgraph_node *callee; |
2686 | struct inline_summary *isummary; | |
d2d668fb | 2687 | |
f45b2a8a | 2688 | if (!known_vals && !known_binfos) |
37678631 | 2689 | return false; |
0f378cb5 JH |
2690 | if (!flag_indirect_inlining) |
2691 | return false; | |
d2d668fb | 2692 | |
8810cc52 MJ |
2693 | target = ipa_get_indirect_edge_target (ie, known_vals, known_binfos, |
2694 | known_aggs); | |
d2d668fb | 2695 | if (!target) |
37678631 | 2696 | return false; |
d2d668fb MK |
2697 | |
2698 | /* Account for difference in cost between indirect and direct calls. */ | |
0f378cb5 JH |
2699 | *size -= (eni_size_weights.indirect_call_cost - eni_size_weights.call_cost); |
2700 | *time -= (eni_time_weights.indirect_call_cost - eni_time_weights.call_cost); | |
2701 | gcc_checking_assert (*time >= 0); | |
2702 | gcc_checking_assert (*size >= 0); | |
f45b2a8a | 2703 | |
d2d668fb MK |
2704 | callee = cgraph_get_node (target); |
2705 | if (!callee || !callee->analyzed) | |
37678631 | 2706 | return false; |
d2d668fb | 2707 | isummary = inline_summary (callee); |
37678631 | 2708 | return isummary->inlinable; |
d2d668fb MK |
2709 | } |
2710 | ||
0f378cb5 JH |
2711 | /* Increase SIZE and TIME for size and time needed to handle edge E. */ |
2712 | ||
2713 | static inline void | |
2714 | estimate_edge_size_and_time (struct cgraph_edge *e, int *size, int *time, | |
2715 | int prob, | |
2716 | VEC (tree, heap) *known_vals, | |
2717 | VEC (tree, heap) *known_binfos, | |
2718 | VEC (ipa_agg_jump_function_p, heap) *known_aggs, | |
2719 | inline_hints *hints) | |
2720 | ||
2721 | { | |
2722 | struct inline_edge_summary *es = inline_edge_summary (e); | |
2723 | int call_size = es->call_stmt_size; | |
2724 | int call_time = es->call_stmt_time; | |
2725 | if (!e->callee | |
2726 | && estimate_edge_devirt_benefit (e, &call_size, &call_time, | |
2727 | known_vals, known_binfos, known_aggs) | |
2728 | && hints | |
2729 | && cgraph_maybe_hot_edge_p (e)) | |
2730 | *hints |= INLINE_HINT_indirect_call; | |
2731 | *size += call_size * INLINE_SIZE_SCALE; | |
2732 | *time += call_time * prob / REG_BR_PROB_BASE | |
2733 | * e->frequency * (INLINE_TIME_SCALE / CGRAPH_FREQ_BASE); | |
2734 | if (*time > MAX_TIME * INLINE_TIME_SCALE) | |
2735 | *time = MAX_TIME * INLINE_TIME_SCALE; | |
2736 | } | |
2737 | ||
2738 | ||
d2d668fb MK |
2739 | |
2740 | /* Increase SIZE and TIME for size and time needed to handle all calls in NODE. | |
2741 | POSSIBLE_TRUTHS, KNOWN_VALS and KNOWN_BINFOS describe context of the call | |
2742 | site. */ | |
632b4f8e JH |
2743 | |
2744 | static void | |
991278ab | 2745 | estimate_calls_size_and_time (struct cgraph_node *node, int *size, int *time, |
37678631 | 2746 | inline_hints *hints, |
d2d668fb MK |
2747 | clause_t possible_truths, |
2748 | VEC (tree, heap) *known_vals, | |
8810cc52 MJ |
2749 | VEC (tree, heap) *known_binfos, |
2750 | VEC (ipa_agg_jump_function_p, heap) *known_aggs) | |
632b4f8e JH |
2751 | { |
2752 | struct cgraph_edge *e; | |
2753 | for (e = node->callees; e; e = e->next_callee) | |
991278ab JH |
2754 | { |
2755 | struct inline_edge_summary *es = inline_edge_summary (e); | |
2756 | if (!es->predicate || evaluate_predicate (es->predicate, possible_truths)) | |
2757 | { | |
2758 | if (e->inline_failed) | |
25837a2f JH |
2759 | { |
2760 | /* Predicates of calls shall not use NOT_CHANGED codes, | |
2761 | sowe do not need to compute probabilities. */ | |
0f378cb5 JH |
2762 | estimate_edge_size_and_time (e, size, time, REG_BR_PROB_BASE, |
2763 | known_vals, known_binfos, known_aggs, | |
2764 | hints); | |
25837a2f | 2765 | } |
991278ab | 2766 | else |
37678631 | 2767 | estimate_calls_size_and_time (e->callee, size, time, hints, |
d2d668fb | 2768 | possible_truths, |
8810cc52 | 2769 | known_vals, known_binfos, known_aggs); |
991278ab JH |
2770 | } |
2771 | } | |
632b4f8e | 2772 | for (e = node->indirect_calls; e; e = e->next_callee) |
991278ab JH |
2773 | { |
2774 | struct inline_edge_summary *es = inline_edge_summary (e); | |
2775 | if (!es->predicate || evaluate_predicate (es->predicate, possible_truths)) | |
0f378cb5 JH |
2776 | estimate_edge_size_and_time (e, size, time, REG_BR_PROB_BASE, |
2777 | known_vals, known_binfos, known_aggs, | |
2778 | hints); | |
991278ab | 2779 | } |
632b4f8e JH |
2780 | } |
2781 | ||
2782 | ||
74605a11 | 2783 | /* Estimate size and time needed to execute NODE assuming |
d2d668fb MK |
2784 | POSSIBLE_TRUTHS clause, and KNOWN_VALS and KNOWN_BINFOS information |
2785 | about NODE's arguments. */ | |
03dfc36d | 2786 | |
632b4f8e | 2787 | static void |
74605a11 JH |
2788 | estimate_node_size_and_time (struct cgraph_node *node, |
2789 | clause_t possible_truths, | |
d2d668fb MK |
2790 | VEC (tree, heap) *known_vals, |
2791 | VEC (tree, heap) *known_binfos, | |
8810cc52 | 2792 | VEC (ipa_agg_jump_function_p, heap) *known_aggs, |
25837a2f | 2793 | int *ret_size, int *ret_time, |
37678631 | 2794 | inline_hints *ret_hints, |
25837a2f JH |
2795 | VEC (inline_param_summary_t, heap) |
2796 | *inline_param_summary) | |
03dfc36d | 2797 | { |
74605a11 | 2798 | struct inline_summary *info = inline_summary (node); |
632b4f8e | 2799 | size_time_entry *e; |
0f378cb5 JH |
2800 | int size = 0; |
2801 | int time = 0; | |
37678631 | 2802 | inline_hints hints = 0; |
632b4f8e JH |
2803 | int i; |
2804 | ||
2805 | if (dump_file | |
2806 | && (dump_flags & TDF_DETAILS)) | |
2807 | { | |
2808 | bool found = false; | |
74605a11 | 2809 | fprintf (dump_file, " Estimating body: %s/%i\n" |
632b4f8e | 2810 | " Known to be false: ", |
74605a11 JH |
2811 | cgraph_node_name (node), |
2812 | node->uid); | |
632b4f8e JH |
2813 | |
2814 | for (i = predicate_not_inlined_condition; | |
2815 | i < (predicate_first_dynamic_condition | |
2816 | + (int)VEC_length (condition, info->conds)); i++) | |
74605a11 | 2817 | if (!(possible_truths & (1 << i))) |
632b4f8e JH |
2818 | { |
2819 | if (found) | |
2820 | fprintf (dump_file, ", "); | |
2821 | found = true; | |
2822 | dump_condition (dump_file, info->conds, i); | |
2823 | } | |
2824 | } | |
2825 | ||
2826 | for (i = 0; VEC_iterate (size_time_entry, info->entry, i, e); i++) | |
74605a11 | 2827 | if (evaluate_predicate (&e->predicate, possible_truths)) |
25837a2f JH |
2828 | { |
2829 | size += e->size; | |
0f378cb5 JH |
2830 | gcc_checking_assert (e->time >= 0); |
2831 | gcc_checking_assert (time >= 0); | |
25837a2f JH |
2832 | if (!inline_param_summary) |
2833 | time += e->time; | |
2834 | else | |
2835 | { | |
2836 | int prob = predicate_probability (info->conds, | |
2837 | &e->predicate, | |
2838 | possible_truths, | |
2839 | inline_param_summary); | |
0f378cb5 JH |
2840 | gcc_checking_assert (prob >= 0); |
2841 | gcc_checking_assert (prob <= REG_BR_PROB_BASE); | |
2842 | time += ((gcov_type)e->time * prob) / REG_BR_PROB_BASE; | |
25837a2f | 2843 | } |
0f378cb5 JH |
2844 | if (time > MAX_TIME * INLINE_TIME_SCALE) |
2845 | time = MAX_TIME * INLINE_TIME_SCALE; | |
2846 | gcc_checking_assert (time >= 0); | |
25837a2f JH |
2847 | |
2848 | } | |
0f378cb5 JH |
2849 | gcc_checking_assert (size >= 0); |
2850 | gcc_checking_assert (time >= 0); | |
e7f23018 | 2851 | |
2daffc47 JH |
2852 | if (info->loop_iterations |
2853 | && !evaluate_predicate (info->loop_iterations, possible_truths)) | |
2854 | hints |=INLINE_HINT_loop_iterations; | |
128e0d89 JH |
2855 | if (info->loop_stride |
2856 | && !evaluate_predicate (info->loop_stride, possible_truths)) | |
2857 | hints |=INLINE_HINT_loop_stride; | |
b48ccf0d JH |
2858 | if (info->scc_no) |
2859 | hints |= INLINE_HINT_in_scc; | |
632b4f8e | 2860 | |
37678631 | 2861 | estimate_calls_size_and_time (node, &size, &time, &hints, possible_truths, |
8810cc52 | 2862 | known_vals, known_binfos, known_aggs); |
0f378cb5 JH |
2863 | gcc_checking_assert (size >= 0); |
2864 | gcc_checking_assert (time >= 0); | |
2865 | time = RDIV (time, INLINE_TIME_SCALE); | |
2866 | size = RDIV (size, INLINE_SIZE_SCALE); | |
632b4f8e JH |
2867 | |
2868 | ||
2869 | if (dump_file | |
2870 | && (dump_flags & TDF_DETAILS)) | |
0f378cb5 | 2871 | fprintf (dump_file, "\n size:%i time:%i\n", (int)size, (int)time); |
632b4f8e JH |
2872 | if (ret_time) |
2873 | *ret_time = time; | |
2874 | if (ret_size) | |
2875 | *ret_size = size; | |
37678631 JH |
2876 | if (ret_hints) |
2877 | *ret_hints = hints; | |
632b4f8e JH |
2878 | return; |
2879 | } | |
2880 | ||
2881 | ||
411a20d6 MJ |
2882 | /* Estimate size and time needed to execute callee of EDGE assuming that |
2883 | parameters known to be constant at caller of EDGE are propagated. | |
d2d668fb MK |
2884 | KNOWN_VALS and KNOWN_BINFOS are vectors of assumed known constant values |
2885 | and types for parameters. */ | |
74605a11 JH |
2886 | |
2887 | void | |
2888 | estimate_ipcp_clone_size_and_time (struct cgraph_node *node, | |
411a20d6 | 2889 | VEC (tree, heap) *known_vals, |
d2d668fb | 2890 | VEC (tree, heap) *known_binfos, |
74605a11 JH |
2891 | int *ret_size, int *ret_time) |
2892 | { | |
411a20d6 MJ |
2893 | clause_t clause; |
2894 | ||
8810cc52 MJ |
2895 | clause = evaluate_conditions_for_known_args (node, false, known_vals, NULL); |
2896 | estimate_node_size_and_time (node, clause, known_vals, known_binfos, NULL, | |
37678631 | 2897 | ret_size, ret_time, NULL, |
25837a2f | 2898 | NULL); |
74605a11 JH |
2899 | } |
2900 | ||
25837a2f JH |
2901 | /* Translate all conditions from callee representation into caller |
2902 | representation and symbolically evaluate predicate P into new predicate. | |
991278ab | 2903 | |
8810cc52 MJ |
2904 | INFO is inline_summary of function we are adding predicate into, CALLEE_INFO |
2905 | is summary of function predicate P is from. OPERAND_MAP is array giving | |
2906 | callee formal IDs the caller formal IDs. POSSSIBLE_TRUTHS is clausule of all | |
2907 | callee conditions that may be true in caller context. TOPLEV_PREDICATE is | |
2908 | predicate under which callee is executed. OFFSET_MAP is an array of of | |
2909 | offsets that need to be added to conditions, negative offset means that | |
2910 | conditions relying on values passed by reference have to be discarded | |
2911 | because they might not be preserved (and should be considered offset zero | |
2912 | for other purposes). */ | |
632b4f8e JH |
2913 | |
2914 | static struct predicate | |
25837a2f JH |
2915 | remap_predicate (struct inline_summary *info, |
2916 | struct inline_summary *callee_info, | |
632b4f8e JH |
2917 | struct predicate *p, |
2918 | VEC (int, heap) *operand_map, | |
8810cc52 | 2919 | VEC (int, heap) *offset_map, |
991278ab JH |
2920 | clause_t possible_truths, |
2921 | struct predicate *toplev_predicate) | |
632b4f8e JH |
2922 | { |
2923 | int i; | |
2924 | struct predicate out = true_predicate (); | |
2925 | ||
2926 | /* True predicate is easy. */ | |
991278ab JH |
2927 | if (true_predicate_p (p)) |
2928 | return *toplev_predicate; | |
632b4f8e JH |
2929 | for (i = 0; p->clause[i]; i++) |
2930 | { | |
2931 | clause_t clause = p->clause[i]; | |
2932 | int cond; | |
2933 | struct predicate clause_predicate = false_predicate (); | |
2934 | ||
f3181aa2 JH |
2935 | gcc_assert (i < MAX_CLAUSES); |
2936 | ||
632b4f8e JH |
2937 | for (cond = 0; cond < NUM_CONDITIONS; cond ++) |
2938 | /* Do we have condition we can't disprove? */ | |
2939 | if (clause & possible_truths & (1 << cond)) | |
2940 | { | |
2941 | struct predicate cond_predicate; | |
2942 | /* Work out if the condition can translate to predicate in the | |
2943 | inlined function. */ | |
2944 | if (cond >= predicate_first_dynamic_condition) | |
2945 | { | |
2946 | struct condition *c; | |
2947 | ||
0823efed DN |
2948 | c = &VEC_index (condition, callee_info->conds, |
2949 | cond - predicate_first_dynamic_condition); | |
632b4f8e JH |
2950 | /* See if we can remap condition operand to caller's operand. |
2951 | Otherwise give up. */ | |
2952 | if (!operand_map | |
daae4650 | 2953 | || (int)VEC_length (int, operand_map) <= c->operand_num |
8810cc52 | 2954 | || VEC_index (int, operand_map, c->operand_num) == -1 |
03360965 MJ |
2955 | /* TODO: For non-aggregate conditions, adding an offset is |
2956 | basically an arithmetic jump function processing which | |
2957 | we should support in future. */ | |
2958 | || ((!c->agg_contents || !c->by_ref) | |
2959 | && VEC_index (int, offset_map, c->operand_num) > 0) | |
8810cc52 MJ |
2960 | || (c->agg_contents && c->by_ref |
2961 | && VEC_index (int, offset_map, c->operand_num) < 0)) | |
632b4f8e JH |
2962 | cond_predicate = true_predicate (); |
2963 | else | |
8810cc52 MJ |
2964 | { |
2965 | struct agg_position_info ap; | |
2966 | HOST_WIDE_INT offset_delta = VEC_index (int, offset_map, | |
2967 | c->operand_num); | |
2968 | if (offset_delta < 0) | |
2969 | { | |
2970 | gcc_checking_assert (!c->agg_contents || !c->by_ref); | |
2971 | offset_delta = 0; | |
2972 | } | |
2973 | gcc_assert (!c->agg_contents | |
2974 | || c->by_ref | |
2975 | || offset_delta == 0); | |
2976 | ap.offset = c->offset + offset_delta; | |
2977 | ap.agg_contents = c->agg_contents; | |
2978 | ap.by_ref = c->by_ref; | |
2979 | cond_predicate = add_condition (info, | |
2980 | VEC_index (int, | |
2981 | operand_map, | |
2982 | c->operand_num), | |
2983 | &ap, c->code, c->val); | |
2984 | } | |
632b4f8e JH |
2985 | } |
2986 | /* Fixed conditions remains same, construct single | |
2987 | condition predicate. */ | |
2988 | else | |
2989 | { | |
2990 | cond_predicate.clause[0] = 1 << cond; | |
2991 | cond_predicate.clause[1] = 0; | |
2992 | } | |
a61bd030 JH |
2993 | clause_predicate = or_predicates (info->conds, &clause_predicate, |
2994 | &cond_predicate); | |
632b4f8e | 2995 | } |
a61bd030 | 2996 | out = and_predicates (info->conds, &out, &clause_predicate); |
632b4f8e | 2997 | } |
a61bd030 | 2998 | return and_predicates (info->conds, &out, toplev_predicate); |
632b4f8e JH |
2999 | } |
3000 | ||
3001 | ||
898b8927 JH |
3002 | /* Update summary information of inline clones after inlining. |
3003 | Compute peak stack usage. */ | |
3004 | ||
3005 | static void | |
3006 | inline_update_callee_summaries (struct cgraph_node *node, | |
3007 | int depth) | |
3008 | { | |
3009 | struct cgraph_edge *e; | |
3010 | struct inline_summary *callee_info = inline_summary (node); | |
3011 | struct inline_summary *caller_info = inline_summary (node->callers->caller); | |
3012 | HOST_WIDE_INT peak; | |
3013 | ||
3014 | callee_info->stack_frame_offset | |
3015 | = caller_info->stack_frame_offset | |
3016 | + caller_info->estimated_self_stack_size; | |
3017 | peak = callee_info->stack_frame_offset | |
3018 | + callee_info->estimated_self_stack_size; | |
3019 | if (inline_summary (node->global.inlined_to)->estimated_stack_size | |
3020 | < peak) | |
3021 | inline_summary (node->global.inlined_to)->estimated_stack_size = peak; | |
3022 | cgraph_propagate_frequency (node); | |
3023 | for (e = node->callees; e; e = e->next_callee) | |
3024 | { | |
3025 | if (!e->inline_failed) | |
3026 | inline_update_callee_summaries (e->callee, depth); | |
3027 | inline_edge_summary (e)->loop_depth += depth; | |
3028 | } | |
3029 | for (e = node->indirect_calls; e; e = e->next_callee) | |
3030 | inline_edge_summary (e)->loop_depth += depth; | |
3031 | } | |
3032 | ||
25837a2f JH |
3033 | /* Update change_prob of EDGE after INLINED_EDGE has been inlined. |
3034 | When functoin A is inlined in B and A calls C with parameter that | |
3035 | changes with probability PROB1 and C is known to be passthroug | |
3036 | of argument if B that change with probability PROB2, the probability | |
3037 | of change is now PROB1*PROB2. */ | |
3038 | ||
3039 | static void | |
3040 | remap_edge_change_prob (struct cgraph_edge *inlined_edge, | |
3041 | struct cgraph_edge *edge) | |
3042 | { | |
3043 | if (ipa_node_params_vector) | |
3044 | { | |
3045 | int i; | |
3046 | struct ipa_edge_args *args = IPA_EDGE_REF (edge); | |
3047 | struct inline_edge_summary *es = inline_edge_summary (edge); | |
3048 | struct inline_edge_summary *inlined_es | |
3049 | = inline_edge_summary (inlined_edge); | |
3050 | ||
3051 | for (i = 0; i < ipa_get_cs_argument_count (args); i++) | |
3052 | { | |
3053 | struct ipa_jump_func *jfunc = ipa_get_ith_jump_func (args, i); | |
3054 | if (jfunc->type == IPA_JF_PASS_THROUGH | |
7b872d9e | 3055 | && (ipa_get_jf_pass_through_formal_id (jfunc) |
0578e417 | 3056 | < (int) VEC_length (inline_param_summary_t, |
7b872d9e | 3057 | inlined_es->param))) |
25837a2f | 3058 | { |
7b872d9e | 3059 | int jf_formal_id = ipa_get_jf_pass_through_formal_id (jfunc); |
25837a2f | 3060 | int prob1 = VEC_index (inline_param_summary_t, |
0823efed | 3061 | es->param, i).change_prob; |
25837a2f JH |
3062 | int prob2 = VEC_index |
3063 | (inline_param_summary_t, | |
0823efed | 3064 | inlined_es->param, jf_formal_id).change_prob; |
25837a2f JH |
3065 | int prob = ((prob1 * prob2 + REG_BR_PROB_BASE / 2) |
3066 | / REG_BR_PROB_BASE); | |
3067 | ||
3068 | if (prob1 && prob2 && !prob) | |
3069 | prob = 1; | |
3070 | ||
3071 | VEC_index (inline_param_summary_t, | |
0823efed | 3072 | es->param, i).change_prob = prob; |
25837a2f JH |
3073 | } |
3074 | } | |
3075 | } | |
3076 | } | |
3077 | ||
3078 | /* Update edge summaries of NODE after INLINED_EDGE has been inlined. | |
3079 | ||
3080 | Remap predicates of callees of NODE. Rest of arguments match | |
3081 | remap_predicate. | |
898b8927 | 3082 | |
25837a2f | 3083 | Also update change probabilities. */ |
991278ab JH |
3084 | |
3085 | static void | |
25837a2f JH |
3086 | remap_edge_summaries (struct cgraph_edge *inlined_edge, |
3087 | struct cgraph_node *node, | |
991278ab JH |
3088 | struct inline_summary *info, |
3089 | struct inline_summary *callee_info, | |
3090 | VEC (int, heap) *operand_map, | |
8810cc52 | 3091 | VEC (int, heap) *offset_map, |
991278ab JH |
3092 | clause_t possible_truths, |
3093 | struct predicate *toplev_predicate) | |
3094 | { | |
3095 | struct cgraph_edge *e; | |
3096 | for (e = node->callees; e; e = e->next_callee) | |
3097 | { | |
3098 | struct inline_edge_summary *es = inline_edge_summary (e); | |
3099 | struct predicate p; | |
25837a2f | 3100 | |
5ee53a06 | 3101 | if (e->inline_failed) |
991278ab | 3102 | { |
25837a2f JH |
3103 | remap_edge_change_prob (inlined_edge, e); |
3104 | ||
5ee53a06 | 3105 | if (es->predicate) |
991278ab | 3106 | { |
5ee53a06 | 3107 | p = remap_predicate (info, callee_info, |
8810cc52 MJ |
3108 | es->predicate, operand_map, offset_map, |
3109 | possible_truths, | |
5ee53a06 JH |
3110 | toplev_predicate); |
3111 | edge_set_predicate (e, &p); | |
25837a2f JH |
3112 | /* TODO: We should remove the edge for code that will be |
3113 | optimized out, but we need to keep verifiers and tree-inline | |
3114 | happy. Make it cold for now. */ | |
5ee53a06 JH |
3115 | if (false_predicate_p (&p)) |
3116 | { | |
3117 | e->count = 0; | |
3118 | e->frequency = 0; | |
3119 | } | |
991278ab | 3120 | } |
5ee53a06 JH |
3121 | else |
3122 | edge_set_predicate (e, toplev_predicate); | |
991278ab | 3123 | } |
5ee53a06 | 3124 | else |
25837a2f | 3125 | remap_edge_summaries (inlined_edge, e->callee, info, callee_info, |
8810cc52 MJ |
3126 | operand_map, offset_map, possible_truths, |
3127 | toplev_predicate); | |
991278ab JH |
3128 | } |
3129 | for (e = node->indirect_calls; e; e = e->next_callee) | |
3130 | { | |
3131 | struct inline_edge_summary *es = inline_edge_summary (e); | |
3132 | struct predicate p; | |
25837a2f JH |
3133 | |
3134 | remap_edge_change_prob (inlined_edge, e); | |
991278ab JH |
3135 | if (es->predicate) |
3136 | { | |
3137 | p = remap_predicate (info, callee_info, | |
8810cc52 MJ |
3138 | es->predicate, operand_map, offset_map, |
3139 | possible_truths, toplev_predicate); | |
991278ab | 3140 | edge_set_predicate (e, &p); |
25837a2f JH |
3141 | /* TODO: We should remove the edge for code that will be optimized |
3142 | out, but we need to keep verifiers and tree-inline happy. | |
991278ab JH |
3143 | Make it cold for now. */ |
3144 | if (false_predicate_p (&p)) | |
3145 | { | |
3146 | e->count = 0; | |
3147 | e->frequency = 0; | |
3148 | } | |
3149 | } | |
e3195c52 JH |
3150 | else |
3151 | edge_set_predicate (e, toplev_predicate); | |
991278ab JH |
3152 | } |
3153 | } | |
3154 | ||
128e0d89 JH |
3155 | /* Same as remap_predicate, but set result into hint *HINT. */ |
3156 | ||
3157 | static void | |
3158 | remap_hint_predicate (struct inline_summary *info, | |
3159 | struct inline_summary *callee_info, | |
3160 | struct predicate **hint, | |
3161 | VEC (int, heap) *operand_map, | |
3162 | VEC (int, heap) *offset_map, | |
3163 | clause_t possible_truths, | |
3164 | struct predicate *toplev_predicate) | |
3165 | { | |
3166 | predicate p; | |
3167 | ||
3168 | if (!*hint) | |
3169 | return; | |
3170 | p = remap_predicate (info, callee_info, | |
3171 | *hint, | |
3172 | operand_map, offset_map, | |
3173 | possible_truths, | |
3174 | toplev_predicate); | |
3175 | if (!false_predicate_p (&p) | |
3176 | && !true_predicate_p (&p)) | |
3177 | { | |
3178 | if (!*hint) | |
3179 | set_hint_predicate (hint, p); | |
3180 | else | |
3181 | **hint = and_predicates (info->conds, | |
3182 | *hint, | |
3183 | &p); | |
3184 | } | |
3185 | } | |
991278ab | 3186 | |
632b4f8e JH |
3187 | /* We inlined EDGE. Update summary of the function we inlined into. */ |
3188 | ||
3189 | void | |
3190 | inline_merge_summary (struct cgraph_edge *edge) | |
3191 | { | |
3192 | struct inline_summary *callee_info = inline_summary (edge->callee); | |
3193 | struct cgraph_node *to = (edge->caller->global.inlined_to | |
3194 | ? edge->caller->global.inlined_to : edge->caller); | |
3195 | struct inline_summary *info = inline_summary (to); | |
3196 | clause_t clause = 0; /* not_inline is known to be false. */ | |
3197 | size_time_entry *e; | |
3198 | VEC (int, heap) *operand_map = NULL; | |
8810cc52 | 3199 | VEC (int, heap) *offset_map = NULL; |
632b4f8e | 3200 | int i; |
991278ab | 3201 | struct predicate toplev_predicate; |
5ee53a06 | 3202 | struct predicate true_p = true_predicate (); |
991278ab JH |
3203 | struct inline_edge_summary *es = inline_edge_summary (edge); |
3204 | ||
3205 | if (es->predicate) | |
3206 | toplev_predicate = *es->predicate; | |
3207 | else | |
3208 | toplev_predicate = true_predicate (); | |
632b4f8e | 3209 | |
5ee53a06 | 3210 | if (ipa_node_params_vector && callee_info->conds) |
632b4f8e JH |
3211 | { |
3212 | struct ipa_edge_args *args = IPA_EDGE_REF (edge); | |
3213 | int count = ipa_get_cs_argument_count (args); | |
3214 | int i; | |
3215 | ||
8810cc52 | 3216 | evaluate_properties_for_edge (edge, true, &clause, NULL, NULL, NULL); |
5ee53a06 | 3217 | if (count) |
8810cc52 MJ |
3218 | { |
3219 | VEC_safe_grow_cleared (int, heap, operand_map, count); | |
3220 | VEC_safe_grow_cleared (int, heap, offset_map, count); | |
3221 | } | |
632b4f8e JH |
3222 | for (i = 0; i < count; i++) |
3223 | { | |
3224 | struct ipa_jump_func *jfunc = ipa_get_ith_jump_func (args, i); | |
3225 | int map = -1; | |
8810cc52 | 3226 | |
632b4f8e | 3227 | /* TODO: handle non-NOPs when merging. */ |
8810cc52 MJ |
3228 | if (jfunc->type == IPA_JF_PASS_THROUGH) |
3229 | { | |
3230 | if (ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR) | |
3231 | map = ipa_get_jf_pass_through_formal_id (jfunc); | |
3232 | if (!ipa_get_jf_pass_through_agg_preserved (jfunc)) | |
3233 | VEC_replace (int, offset_map, i, -1); | |
3234 | } | |
3235 | else if (jfunc->type == IPA_JF_ANCESTOR) | |
3236 | { | |
3237 | HOST_WIDE_INT offset = ipa_get_jf_ancestor_offset (jfunc); | |
3238 | if (offset >= 0 && offset < INT_MAX) | |
3239 | { | |
3240 | map = ipa_get_jf_ancestor_formal_id (jfunc); | |
3241 | if (!ipa_get_jf_ancestor_agg_preserved (jfunc)) | |
3242 | offset = -1; | |
3243 | VEC_replace (int, offset_map, i, offset); | |
3244 | } | |
3245 | } | |
632b4f8e | 3246 | VEC_replace (int, operand_map, i, map); |
f3181aa2 | 3247 | gcc_assert (map < ipa_get_param_count (IPA_NODE_REF (to))); |
632b4f8e JH |
3248 | } |
3249 | } | |
3250 | for (i = 0; VEC_iterate (size_time_entry, callee_info->entry, i, e); i++) | |
3251 | { | |
3252 | struct predicate p = remap_predicate (info, callee_info, | |
8810cc52 MJ |
3253 | &e->predicate, operand_map, |
3254 | offset_map, clause, | |
991278ab | 3255 | &toplev_predicate); |
25837a2f JH |
3256 | if (!false_predicate_p (&p)) |
3257 | { | |
3258 | gcov_type add_time = ((gcov_type)e->time * edge->frequency | |
3259 | + CGRAPH_FREQ_BASE / 2) / CGRAPH_FREQ_BASE; | |
3260 | int prob = predicate_probability (callee_info->conds, | |
3261 | &e->predicate, | |
3262 | clause, es->param); | |
0f378cb5 | 3263 | add_time = ((gcov_type)add_time * prob) / REG_BR_PROB_BASE; |
25837a2f JH |
3264 | if (add_time > MAX_TIME * INLINE_TIME_SCALE) |
3265 | add_time = MAX_TIME * INLINE_TIME_SCALE; | |
3266 | if (prob != REG_BR_PROB_BASE | |
3267 | && dump_file && (dump_flags & TDF_DETAILS)) | |
3268 | { | |
3269 | fprintf (dump_file, "\t\tScaling time by probability:%f\n", | |
3270 | (double)prob / REG_BR_PROB_BASE); | |
3271 | } | |
3272 | account_size_time (info, e->size, add_time, &p); | |
3273 | } | |
3274 | } | |
3275 | remap_edge_summaries (edge, edge->callee, info, callee_info, operand_map, | |
8810cc52 | 3276 | offset_map, clause, &toplev_predicate); |
128e0d89 JH |
3277 | remap_hint_predicate (info, callee_info, |
3278 | &callee_info->loop_iterations, | |
3279 | operand_map, offset_map, | |
3280 | clause, &toplev_predicate); | |
3281 | remap_hint_predicate (info, callee_info, | |
3282 | &callee_info->loop_stride, | |
3283 | operand_map, offset_map, | |
3284 | clause, &toplev_predicate); | |
898b8927 JH |
3285 | |
3286 | inline_update_callee_summaries (edge->callee, | |
3287 | inline_edge_summary (edge)->loop_depth); | |
3288 | ||
5ee53a06 JH |
3289 | /* We do not maintain predicates of inlined edges, free it. */ |
3290 | edge_set_predicate (edge, &true_p); | |
25837a2f JH |
3291 | /* Similarly remove param summaries. */ |
3292 | VEC_free (inline_param_summary_t, heap, es->param); | |
fb909888 | 3293 | VEC_free (int, heap, operand_map); |
8810cc52 | 3294 | VEC_free (int, heap, offset_map); |
c170d40f JH |
3295 | } |
3296 | ||
3297 | /* For performance reasons inline_merge_summary is not updating overall size | |
3298 | and time. Recompute it. */ | |
5ee53a06 | 3299 | |
c170d40f JH |
3300 | void |
3301 | inline_update_overall_summary (struct cgraph_node *node) | |
3302 | { | |
3303 | struct inline_summary *info = inline_summary (node); | |
3304 | size_time_entry *e; | |
3305 | int i; | |
3306 | ||
3307 | info->size = 0; | |
3308 | info->time = 0; | |
3309 | for (i = 0; VEC_iterate (size_time_entry, info->entry, i, e); i++) | |
3310 | info->size += e->size, info->time += e->time; | |
37678631 | 3311 | estimate_calls_size_and_time (node, &info->size, &info->time, NULL, |
c170d40f | 3312 | ~(clause_t)(1 << predicate_false_condition), |
8810cc52 | 3313 | NULL, NULL, NULL); |
632b4f8e JH |
3314 | info->time = (info->time + INLINE_TIME_SCALE / 2) / INLINE_TIME_SCALE; |
3315 | info->size = (info->size + INLINE_SIZE_SCALE / 2) / INLINE_SIZE_SCALE; | |
3316 | } | |
3317 | ||
632b4f8e JH |
3318 | /* Estimate the time cost for the caller when inlining EDGE. |
3319 | Only to be called via estimate_edge_time, that handles the | |
3320 | caching mechanism. | |
3321 | ||
3322 | When caching, also update the cache entry. Compute both time and | |
3323 | size, since we always need both metrics eventually. */ | |
3324 | ||
3325 | int | |
3326 | do_estimate_edge_time (struct cgraph_edge *edge) | |
3327 | { | |
3328 | int time; | |
3329 | int size; | |
37678631 | 3330 | inline_hints hints; |
632b4f8e | 3331 | gcov_type ret; |
d2d668fb MK |
3332 | struct cgraph_node *callee; |
3333 | clause_t clause; | |
3334 | VEC (tree, heap) *known_vals; | |
3335 | VEC (tree, heap) *known_binfos; | |
8810cc52 | 3336 | VEC (ipa_agg_jump_function_p, heap) *known_aggs; |
898b8927 | 3337 | struct inline_edge_summary *es = inline_edge_summary (edge); |
632b4f8e | 3338 | |
d2d668fb MK |
3339 | callee = cgraph_function_or_thunk_node (edge->callee, NULL); |
3340 | ||
632b4f8e | 3341 | gcc_checking_assert (edge->inline_failed); |
d2d668fb | 3342 | evaluate_properties_for_edge (edge, true, |
8810cc52 MJ |
3343 | &clause, &known_vals, &known_binfos, |
3344 | &known_aggs); | |
d2d668fb | 3345 | estimate_node_size_and_time (callee, clause, known_vals, known_binfos, |
37678631 | 3346 | known_aggs, &size, &time, &hints, es->param); |
d2d668fb MK |
3347 | VEC_free (tree, heap, known_vals); |
3348 | VEC_free (tree, heap, known_binfos); | |
8810cc52 | 3349 | VEC_free (ipa_agg_jump_function_p, heap, known_aggs); |
632b4f8e | 3350 | |
ed901e4c JH |
3351 | ret = RDIV ((gcov_type)time * edge->frequency, |
3352 | CGRAPH_FREQ_BASE); | |
632b4f8e JH |
3353 | |
3354 | /* When caching, update the cache entry. */ | |
3355 | if (edge_growth_cache) | |
3356 | { | |
b48ccf0d JH |
3357 | struct cgraph_node *to = (edge->caller->global.inlined_to |
3358 | ? edge->caller->global.inlined_to | |
3359 | : edge->caller); | |
632b4f8e JH |
3360 | if ((int)VEC_length (edge_growth_cache_entry, edge_growth_cache) |
3361 | <= edge->uid) | |
3362 | VEC_safe_grow_cleared (edge_growth_cache_entry, heap, edge_growth_cache, | |
3363 | cgraph_edge_max_uid); | |
0823efed | 3364 | VEC_index (edge_growth_cache_entry, edge_growth_cache, edge->uid).time |
632b4f8e JH |
3365 | = ret + (ret >= 0); |
3366 | ||
0823efed | 3367 | VEC_index (edge_growth_cache_entry, edge_growth_cache, edge->uid).size |
ed901e4c | 3368 | = size + (size >= 0); |
b48ccf0d | 3369 | if (inline_summary (to)->scc_no |
bf3f6510 JH |
3370 | && inline_summary (to)->scc_no == inline_summary (callee)->scc_no |
3371 | && !cgraph_edge_recursive_p (edge)) | |
b48ccf0d | 3372 | hints |= INLINE_HINT_same_scc; |
37678631 JH |
3373 | VEC_index (edge_growth_cache_entry, edge_growth_cache, edge->uid).hints |
3374 | = hints + 1; | |
632b4f8e JH |
3375 | } |
3376 | return ret; | |
3377 | } | |
3378 | ||
3379 | ||
ed901e4c | 3380 | /* Return estimated callee growth after inlining EDGE. |
632b4f8e JH |
3381 | Only to be called via estimate_edge_size. */ |
3382 | ||
3383 | int | |
ed901e4c | 3384 | do_estimate_edge_size (struct cgraph_edge *edge) |
632b4f8e JH |
3385 | { |
3386 | int size; | |
a5b1779f | 3387 | struct cgraph_node *callee; |
d2d668fb MK |
3388 | clause_t clause; |
3389 | VEC (tree, heap) *known_vals; | |
3390 | VEC (tree, heap) *known_binfos; | |
8810cc52 | 3391 | VEC (ipa_agg_jump_function_p, heap) *known_aggs; |
632b4f8e JH |
3392 | |
3393 | /* When we do caching, use do_estimate_edge_time to populate the entry. */ | |
3394 | ||
3395 | if (edge_growth_cache) | |
3396 | { | |
3397 | do_estimate_edge_time (edge); | |
3398 | size = VEC_index (edge_growth_cache_entry, | |
3399 | edge_growth_cache, | |
0823efed | 3400 | edge->uid).size; |
632b4f8e JH |
3401 | gcc_checking_assert (size); |
3402 | return size - (size > 0); | |
3403 | } | |
d2d668fb | 3404 | |
a5b1779f | 3405 | callee = cgraph_function_or_thunk_node (edge->callee, NULL); |
632b4f8e JH |
3406 | |
3407 | /* Early inliner runs without caching, go ahead and do the dirty work. */ | |
3408 | gcc_checking_assert (edge->inline_failed); | |
d2d668fb | 3409 | evaluate_properties_for_edge (edge, true, |
8810cc52 MJ |
3410 | &clause, &known_vals, &known_binfos, |
3411 | &known_aggs); | |
d2d668fb | 3412 | estimate_node_size_and_time (callee, clause, known_vals, known_binfos, |
37678631 | 3413 | known_aggs, &size, NULL, NULL, NULL); |
d2d668fb MK |
3414 | VEC_free (tree, heap, known_vals); |
3415 | VEC_free (tree, heap, known_binfos); | |
8810cc52 | 3416 | VEC_free (ipa_agg_jump_function_p, heap, known_aggs); |
ed901e4c | 3417 | return size; |
03dfc36d JH |
3418 | } |
3419 | ||
3420 | ||
37678631 JH |
3421 | /* Estimate the growth of the caller when inlining EDGE. |
3422 | Only to be called via estimate_edge_size. */ | |
3423 | ||
3424 | inline_hints | |
3425 | do_estimate_edge_hints (struct cgraph_edge *edge) | |
3426 | { | |
3427 | inline_hints hints; | |
3428 | struct cgraph_node *callee; | |
3429 | clause_t clause; | |
3430 | VEC (tree, heap) *known_vals; | |
3431 | VEC (tree, heap) *known_binfos; | |
3432 | VEC (ipa_agg_jump_function_p, heap) *known_aggs; | |
b48ccf0d JH |
3433 | struct cgraph_node *to = (edge->caller->global.inlined_to |
3434 | ? edge->caller->global.inlined_to | |
3435 | : edge->caller); | |
37678631 JH |
3436 | |
3437 | /* When we do caching, use do_estimate_edge_time to populate the entry. */ | |
3438 | ||
3439 | if (edge_growth_cache) | |
3440 | { | |
3441 | do_estimate_edge_time (edge); | |
3442 | hints = VEC_index (edge_growth_cache_entry, | |
3443 | edge_growth_cache, | |
3444 | edge->uid).hints; | |
3445 | gcc_checking_assert (hints); | |
3446 | return hints - 1; | |
3447 | } | |
3448 | ||
3449 | callee = cgraph_function_or_thunk_node (edge->callee, NULL); | |
3450 | ||
3451 | /* Early inliner runs without caching, go ahead and do the dirty work. */ | |
3452 | gcc_checking_assert (edge->inline_failed); | |
3453 | evaluate_properties_for_edge (edge, true, | |
3454 | &clause, &known_vals, &known_binfos, | |
3455 | &known_aggs); | |
3456 | estimate_node_size_and_time (callee, clause, known_vals, known_binfos, | |
3457 | known_aggs, NULL, NULL, &hints, NULL); | |
3458 | VEC_free (tree, heap, known_vals); | |
3459 | VEC_free (tree, heap, known_binfos); | |
3460 | VEC_free (ipa_agg_jump_function_p, heap, known_aggs); | |
b48ccf0d | 3461 | if (inline_summary (to)->scc_no |
bf3f6510 JH |
3462 | && inline_summary (to)->scc_no == inline_summary (callee)->scc_no |
3463 | && !cgraph_edge_recursive_p (edge)) | |
b48ccf0d | 3464 | hints |= INLINE_HINT_same_scc; |
37678631 JH |
3465 | return hints; |
3466 | } | |
3467 | ||
3468 | ||
03dfc36d JH |
3469 | /* Estimate self time of the function NODE after inlining EDGE. */ |
3470 | ||
3471 | int | |
3472 | estimate_time_after_inlining (struct cgraph_node *node, | |
3473 | struct cgraph_edge *edge) | |
3474 | { | |
b15c64ee JH |
3475 | struct inline_edge_summary *es = inline_edge_summary (edge); |
3476 | if (!es->predicate || !false_predicate_p (es->predicate)) | |
3477 | { | |
3478 | gcov_type time = inline_summary (node)->time + estimate_edge_time (edge); | |
3479 | if (time < 0) | |
3480 | time = 0; | |
3481 | if (time > MAX_TIME) | |
3482 | time = MAX_TIME; | |
3483 | return time; | |
3484 | } | |
3485 | return inline_summary (node)->time; | |
03dfc36d JH |
3486 | } |
3487 | ||
3488 | ||
3489 | /* Estimate the size of NODE after inlining EDGE which should be an | |
3490 | edge to either NODE or a call inlined into NODE. */ | |
3491 | ||
3492 | int | |
3493 | estimate_size_after_inlining (struct cgraph_node *node, | |
10a5dd5d | 3494 | struct cgraph_edge *edge) |
03dfc36d | 3495 | { |
b15c64ee JH |
3496 | struct inline_edge_summary *es = inline_edge_summary (edge); |
3497 | if (!es->predicate || !false_predicate_p (es->predicate)) | |
3498 | { | |
3499 | int size = inline_summary (node)->size + estimate_edge_growth (edge); | |
3500 | gcc_assert (size >= 0); | |
3501 | return size; | |
3502 | } | |
3503 | return inline_summary (node)->size; | |
03dfc36d JH |
3504 | } |
3505 | ||
3506 | ||
a5b1779f JH |
3507 | struct growth_data |
3508 | { | |
3509 | bool self_recursive; | |
3510 | int growth; | |
3511 | }; | |
03dfc36d | 3512 | |
a5b1779f JH |
3513 | |
3514 | /* Worker for do_estimate_growth. Collect growth for all callers. */ | |
3515 | ||
3516 | static bool | |
3517 | do_estimate_growth_1 (struct cgraph_node *node, void *data) | |
03dfc36d | 3518 | { |
03dfc36d | 3519 | struct cgraph_edge *e; |
a5b1779f | 3520 | struct growth_data *d = (struct growth_data *) data; |
03dfc36d | 3521 | |
03dfc36d JH |
3522 | for (e = node->callers; e; e = e->next_caller) |
3523 | { | |
4c0f7679 JH |
3524 | gcc_checking_assert (e->inline_failed); |
3525 | ||
3526 | if (e->caller == node | |
3527 | || (e->caller->global.inlined_to | |
3528 | && e->caller->global.inlined_to == node)) | |
a5b1779f JH |
3529 | d->self_recursive = true; |
3530 | d->growth += estimate_edge_growth (e); | |
4c0f7679 | 3531 | } |
a5b1779f JH |
3532 | return false; |
3533 | } | |
3534 | ||
3535 | ||
3536 | /* Estimate the growth caused by inlining NODE into all callees. */ | |
3537 | ||
3538 | int | |
3539 | do_estimate_growth (struct cgraph_node *node) | |
3540 | { | |
3541 | struct growth_data d = {0, false}; | |
3542 | struct inline_summary *info = inline_summary (node); | |
3543 | ||
3544 | cgraph_for_node_and_aliases (node, do_estimate_growth_1, &d, true); | |
4c0f7679 JH |
3545 | |
3546 | /* For self recursive functions the growth estimation really should be | |
3547 | infinity. We don't want to return very large values because the growth | |
3548 | plays various roles in badness computation fractions. Be sure to not | |
3549 | return zero or negative growths. */ | |
a5b1779f JH |
3550 | if (d.self_recursive) |
3551 | d.growth = d.growth < info->size ? info->size : d.growth; | |
4c0f7679 JH |
3552 | else |
3553 | { | |
960bfb69 | 3554 | if (!DECL_EXTERNAL (node->symbol.decl) |
317a0646 | 3555 | && cgraph_will_be_removed_from_program_if_no_direct_calls (node)) |
a5b1779f | 3556 | d.growth -= info->size; |
9e990d14 JH |
3557 | /* COMDAT functions are very often not shared across multiple units |
3558 | since they come from various template instantiations. | |
3559 | Take this into account. */ | |
960bfb69 | 3560 | else if (DECL_COMDAT (node->symbol.decl) |
4c0f7679 | 3561 | && cgraph_can_remove_if_no_direct_calls_p (node)) |
a5b1779f | 3562 | d.growth -= (info->size |
9e990d14 JH |
3563 | * (100 - PARAM_VALUE (PARAM_COMDAT_SHARING_PROBABILITY)) |
3564 | + 50) / 100; | |
03dfc36d | 3565 | } |
03dfc36d | 3566 | |
632b4f8e JH |
3567 | if (node_growth_cache) |
3568 | { | |
3569 | if ((int)VEC_length (int, node_growth_cache) <= node->uid) | |
3570 | VEC_safe_grow_cleared (int, heap, node_growth_cache, cgraph_max_uid); | |
9e990d14 JH |
3571 | VEC_replace (int, node_growth_cache, node->uid, |
3572 | d.growth + (d.growth >= 0)); | |
632b4f8e | 3573 | } |
a5b1779f | 3574 | return d.growth; |
03dfc36d JH |
3575 | } |
3576 | ||
10a5dd5d | 3577 | |
03dfc36d JH |
3578 | /* This function performs intraprocedural analysis in NODE that is required to |
3579 | inline indirect calls. */ | |
10a5dd5d | 3580 | |
03dfc36d JH |
3581 | static void |
3582 | inline_indirect_intraprocedural_analysis (struct cgraph_node *node) | |
3583 | { | |
3584 | ipa_analyze_node (node); | |
3585 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3586 | { | |
3587 | ipa_print_node_params (dump_file, node); | |
3588 | ipa_print_node_jump_functions (dump_file, node); | |
3589 | } | |
3590 | } | |
3591 | ||
3592 | ||
3593 | /* Note function body size. */ | |
3594 | ||
3595 | static void | |
3596 | inline_analyze_function (struct cgraph_node *node) | |
3597 | { | |
960bfb69 | 3598 | push_cfun (DECL_STRUCT_FUNCTION (node->symbol.decl)); |
03dfc36d | 3599 | |
632b4f8e JH |
3600 | if (dump_file) |
3601 | fprintf (dump_file, "\nAnalyzing function: %s/%u\n", | |
3602 | cgraph_node_name (node), node->uid); | |
5ee53a06 | 3603 | if (optimize && !node->thunk.thunk_p) |
03dfc36d | 3604 | inline_indirect_intraprocedural_analysis (node); |
632b4f8e | 3605 | compute_inline_parameters (node, false); |
03dfc36d | 3606 | |
03dfc36d JH |
3607 | pop_cfun (); |
3608 | } | |
3609 | ||
3610 | ||
3611 | /* Called when new function is inserted to callgraph late. */ | |
3612 | ||
3613 | static void | |
3614 | add_new_function (struct cgraph_node *node, void *data ATTRIBUTE_UNUSED) | |
3615 | { | |
3616 | inline_analyze_function (node); | |
3617 | } | |
3618 | ||
3619 | ||
3620 | /* Note function body size. */ | |
3621 | ||
3622 | void | |
3623 | inline_generate_summary (void) | |
3624 | { | |
3625 | struct cgraph_node *node; | |
3626 | ||
3627 | function_insertion_hook_holder = | |
3628 | cgraph_add_function_insertion_hook (&add_new_function, NULL); | |
3629 | ||
5ee53a06 | 3630 | ipa_register_cgraph_hooks (); |
1c52c601 | 3631 | inline_free_summary (); |
03dfc36d | 3632 | |
c47d0034 | 3633 | FOR_EACH_DEFINED_FUNCTION (node) |
a5b1779f | 3634 | if (!node->alias) |
03dfc36d | 3635 | inline_analyze_function (node); |
03dfc36d JH |
3636 | } |
3637 | ||
3638 | ||
991278ab JH |
3639 | /* Read predicate from IB. */ |
3640 | ||
3641 | static struct predicate | |
3642 | read_predicate (struct lto_input_block *ib) | |
3643 | { | |
3644 | struct predicate out; | |
3645 | clause_t clause; | |
3646 | int k = 0; | |
3647 | ||
3648 | do | |
3649 | { | |
b15c64ee | 3650 | gcc_assert (k <= MAX_CLAUSES); |
412288f1 | 3651 | clause = out.clause[k++] = streamer_read_uhwi (ib); |
991278ab JH |
3652 | } |
3653 | while (clause); | |
f75e1f1e AO |
3654 | |
3655 | /* Zero-initialize the remaining clauses in OUT. */ | |
3656 | while (k <= MAX_CLAUSES) | |
3657 | out.clause[k++] = 0; | |
3658 | ||
991278ab JH |
3659 | return out; |
3660 | } | |
3661 | ||
3662 | ||
898b8927 JH |
3663 | /* Write inline summary for edge E to OB. */ |
3664 | ||
3665 | static void | |
3666 | read_inline_edge_summary (struct lto_input_block *ib, struct cgraph_edge *e) | |
3667 | { | |
3668 | struct inline_edge_summary *es = inline_edge_summary (e); | |
991278ab | 3669 | struct predicate p; |
25837a2f | 3670 | int length, i; |
991278ab | 3671 | |
412288f1 DN |
3672 | es->call_stmt_size = streamer_read_uhwi (ib); |
3673 | es->call_stmt_time = streamer_read_uhwi (ib); | |
3674 | es->loop_depth = streamer_read_uhwi (ib); | |
991278ab JH |
3675 | p = read_predicate (ib); |
3676 | edge_set_predicate (e, &p); | |
25837a2f JH |
3677 | length = streamer_read_uhwi (ib); |
3678 | if (length) | |
3679 | { | |
3680 | VEC_safe_grow_cleared (inline_param_summary_t, heap, es->param, length); | |
3681 | for (i = 0; i < length; i++) | |
0823efed | 3682 | VEC_index (inline_param_summary_t, es->param, i).change_prob |
25837a2f JH |
3683 | = streamer_read_uhwi (ib); |
3684 | } | |
898b8927 JH |
3685 | } |
3686 | ||
3687 | ||
632b4f8e JH |
3688 | /* Stream in inline summaries from the section. */ |
3689 | ||
3690 | static void | |
3691 | inline_read_section (struct lto_file_decl_data *file_data, const char *data, | |
3692 | size_t len) | |
3693 | { | |
3694 | const struct lto_function_header *header = | |
3695 | (const struct lto_function_header *) data; | |
4ad9a9de EB |
3696 | const int cfg_offset = sizeof (struct lto_function_header); |
3697 | const int main_offset = cfg_offset + header->cfg_size; | |
3698 | const int string_offset = main_offset + header->main_size; | |
632b4f8e JH |
3699 | struct data_in *data_in; |
3700 | struct lto_input_block ib; | |
3701 | unsigned int i, count2, j; | |
3702 | unsigned int f_count; | |
3703 | ||
3704 | LTO_INIT_INPUT_BLOCK (ib, (const char *) data + main_offset, 0, | |
3705 | header->main_size); | |
3706 | ||
3707 | data_in = | |
3708 | lto_data_in_create (file_data, (const char *) data + string_offset, | |
3709 | header->string_size, NULL); | |
412288f1 | 3710 | f_count = streamer_read_uhwi (&ib); |
632b4f8e JH |
3711 | for (i = 0; i < f_count; i++) |
3712 | { | |
3713 | unsigned int index; | |
3714 | struct cgraph_node *node; | |
3715 | struct inline_summary *info; | |
7380e6ef | 3716 | lto_symtab_encoder_t encoder; |
632b4f8e | 3717 | struct bitpack_d bp; |
898b8927 | 3718 | struct cgraph_edge *e; |
2daffc47 | 3719 | predicate p; |
632b4f8e | 3720 | |
412288f1 | 3721 | index = streamer_read_uhwi (&ib); |
7380e6ef JH |
3722 | encoder = file_data->symtab_node_encoder; |
3723 | node = cgraph (lto_symtab_encoder_deref (encoder, index)); | |
632b4f8e JH |
3724 | info = inline_summary (node); |
3725 | ||
3726 | info->estimated_stack_size | |
412288f1 DN |
3727 | = info->estimated_self_stack_size = streamer_read_uhwi (&ib); |
3728 | info->size = info->self_size = streamer_read_uhwi (&ib); | |
3729 | info->time = info->self_time = streamer_read_uhwi (&ib); | |
632b4f8e | 3730 | |
412288f1 | 3731 | bp = streamer_read_bitpack (&ib); |
632b4f8e | 3732 | info->inlinable = bp_unpack_value (&bp, 1); |
632b4f8e | 3733 | |
412288f1 | 3734 | count2 = streamer_read_uhwi (&ib); |
632b4f8e JH |
3735 | gcc_assert (!info->conds); |
3736 | for (j = 0; j < count2; j++) | |
3737 | { | |
3738 | struct condition c; | |
412288f1 DN |
3739 | c.operand_num = streamer_read_uhwi (&ib); |
3740 | c.code = (enum tree_code) streamer_read_uhwi (&ib); | |
b9393656 | 3741 | c.val = stream_read_tree (&ib, data_in); |
8810cc52 MJ |
3742 | bp = streamer_read_bitpack (&ib); |
3743 | c.agg_contents = bp_unpack_value (&bp, 1); | |
3744 | c.by_ref = bp_unpack_value (&bp, 1); | |
3745 | if (c.agg_contents) | |
3746 | c.offset = streamer_read_uhwi (&ib); | |
f32682ca | 3747 | VEC_safe_push (condition, gc, info->conds, c); |
632b4f8e | 3748 | } |
412288f1 | 3749 | count2 = streamer_read_uhwi (&ib); |
632b4f8e JH |
3750 | gcc_assert (!info->entry); |
3751 | for (j = 0; j < count2; j++) | |
3752 | { | |
3753 | struct size_time_entry e; | |
632b4f8e | 3754 | |
412288f1 DN |
3755 | e.size = streamer_read_uhwi (&ib); |
3756 | e.time = streamer_read_uhwi (&ib); | |
991278ab | 3757 | e.predicate = read_predicate (&ib); |
632b4f8e | 3758 | |
f32682ca | 3759 | VEC_safe_push (size_time_entry, gc, info->entry, e); |
632b4f8e | 3760 | } |
2daffc47 JH |
3761 | |
3762 | p = read_predicate (&ib); | |
128e0d89 JH |
3763 | set_hint_predicate (&info->loop_iterations, p); |
3764 | p = read_predicate (&ib); | |
3765 | set_hint_predicate (&info->loop_stride, p); | |
898b8927 JH |
3766 | for (e = node->callees; e; e = e->next_callee) |
3767 | read_inline_edge_summary (&ib, e); | |
3768 | for (e = node->indirect_calls; e; e = e->next_callee) | |
3769 | read_inline_edge_summary (&ib, e); | |
632b4f8e JH |
3770 | } |
3771 | ||
3772 | lto_free_section_data (file_data, LTO_section_inline_summary, NULL, data, | |
3773 | len); | |
3774 | lto_data_in_delete (data_in); | |
3775 | } | |
3776 | ||
3777 | ||
03dfc36d JH |
3778 | /* Read inline summary. Jump functions are shared among ipa-cp |
3779 | and inliner, so when ipa-cp is active, we don't need to write them | |
3780 | twice. */ | |
3781 | ||
3782 | void | |
3783 | inline_read_summary (void) | |
3784 | { | |
10a5dd5d JH |
3785 | struct lto_file_decl_data **file_data_vec = lto_get_file_decl_data (); |
3786 | struct lto_file_decl_data *file_data; | |
3787 | unsigned int j = 0; | |
3788 | ||
3789 | inline_summary_alloc (); | |
3790 | ||
3791 | while ((file_data = file_data_vec[j++])) | |
3792 | { | |
3793 | size_t len; | |
25837a2f JH |
3794 | const char *data = lto_get_section_data (file_data, |
3795 | LTO_section_inline_summary, | |
3796 | NULL, &len); | |
632b4f8e JH |
3797 | if (data) |
3798 | inline_read_section (file_data, data, len); | |
10a5dd5d | 3799 | else |
25837a2f JH |
3800 | /* Fatal error here. We do not want to support compiling ltrans units |
3801 | with different version of compiler or different flags than the WPA | |
3802 | unit, so this should never happen. */ | |
10a5dd5d JH |
3803 | fatal_error ("ipa inline summary is missing in input file"); |
3804 | } | |
5ee53a06 | 3805 | if (optimize) |
03dfc36d JH |
3806 | { |
3807 | ipa_register_cgraph_hooks (); | |
3808 | if (!flag_ipa_cp) | |
3809 | ipa_prop_read_jump_functions (); | |
3810 | } | |
3811 | function_insertion_hook_holder = | |
3812 | cgraph_add_function_insertion_hook (&add_new_function, NULL); | |
3813 | } | |
3814 | ||
991278ab JH |
3815 | |
3816 | /* Write predicate P to OB. */ | |
3817 | ||
3818 | static void | |
3819 | write_predicate (struct output_block *ob, struct predicate *p) | |
3820 | { | |
3821 | int j; | |
3822 | if (p) | |
3823 | for (j = 0; p->clause[j]; j++) | |
3824 | { | |
3825 | gcc_assert (j < MAX_CLAUSES); | |
412288f1 | 3826 | streamer_write_uhwi (ob, p->clause[j]); |
991278ab | 3827 | } |
412288f1 | 3828 | streamer_write_uhwi (ob, 0); |
991278ab JH |
3829 | } |
3830 | ||
3831 | ||
898b8927 JH |
3832 | /* Write inline summary for edge E to OB. */ |
3833 | ||
3834 | static void | |
3835 | write_inline_edge_summary (struct output_block *ob, struct cgraph_edge *e) | |
3836 | { | |
3837 | struct inline_edge_summary *es = inline_edge_summary (e); | |
25837a2f JH |
3838 | int i; |
3839 | ||
412288f1 DN |
3840 | streamer_write_uhwi (ob, es->call_stmt_size); |
3841 | streamer_write_uhwi (ob, es->call_stmt_time); | |
3842 | streamer_write_uhwi (ob, es->loop_depth); | |
991278ab | 3843 | write_predicate (ob, es->predicate); |
25837a2f JH |
3844 | streamer_write_uhwi (ob, VEC_length (inline_param_summary_t, es->param)); |
3845 | for (i = 0; i < (int)VEC_length (inline_param_summary_t, es->param); i++) | |
3846 | streamer_write_uhwi (ob, VEC_index (inline_param_summary_t, | |
0823efed | 3847 | es->param, i).change_prob); |
898b8927 JH |
3848 | } |
3849 | ||
03dfc36d JH |
3850 | |
3851 | /* Write inline summary for node in SET. | |
3852 | Jump functions are shared among ipa-cp and inliner, so when ipa-cp is | |
3853 | active, we don't need to write them twice. */ | |
3854 | ||
3855 | void | |
f27c1867 | 3856 | inline_write_summary (void) |
03dfc36d | 3857 | { |
10a5dd5d | 3858 | struct cgraph_node *node; |
632b4f8e | 3859 | struct output_block *ob = create_output_block (LTO_section_inline_summary); |
7380e6ef | 3860 | lto_symtab_encoder_t encoder = ob->decl_state->symtab_node_encoder; |
10a5dd5d JH |
3861 | unsigned int count = 0; |
3862 | int i; | |
3863 | ||
7380e6ef | 3864 | for (i = 0; i < lto_symtab_encoder_size (encoder); i++) |
5d59b5e1 LC |
3865 | { |
3866 | symtab_node snode = lto_symtab_encoder_deref (encoder, i); | |
3867 | cgraph_node *cnode = dyn_cast <cgraph_node> (snode); | |
3868 | if (cnode && cnode->analyzed) | |
3869 | count++; | |
3870 | } | |
412288f1 | 3871 | streamer_write_uhwi (ob, count); |
10a5dd5d | 3872 | |
7380e6ef | 3873 | for (i = 0; i < lto_symtab_encoder_size (encoder); i++) |
10a5dd5d | 3874 | { |
5d59b5e1 LC |
3875 | symtab_node snode = lto_symtab_encoder_deref (encoder, i); |
3876 | cgraph_node *cnode = dyn_cast <cgraph_node> (snode); | |
3877 | if (cnode && (node = cnode)->analyzed) | |
10a5dd5d JH |
3878 | { |
3879 | struct inline_summary *info = inline_summary (node); | |
e7f23018 | 3880 | struct bitpack_d bp; |
898b8927 | 3881 | struct cgraph_edge *edge; |
632b4f8e JH |
3882 | int i; |
3883 | size_time_entry *e; | |
3884 | struct condition *c; | |
5d59b5e1 | 3885 | |
7380e6ef | 3886 | streamer_write_uhwi (ob, lto_symtab_encoder_encode (encoder, (symtab_node)node)); |
412288f1 DN |
3887 | streamer_write_hwi (ob, info->estimated_self_stack_size); |
3888 | streamer_write_hwi (ob, info->self_size); | |
3889 | streamer_write_hwi (ob, info->self_time); | |
e7f23018 JH |
3890 | bp = bitpack_create (ob->main_stream); |
3891 | bp_pack_value (&bp, info->inlinable, 1); | |
412288f1 DN |
3892 | streamer_write_bitpack (&bp); |
3893 | streamer_write_uhwi (ob, VEC_length (condition, info->conds)); | |
632b4f8e JH |
3894 | for (i = 0; VEC_iterate (condition, info->conds, i, c); i++) |
3895 | { | |
412288f1 DN |
3896 | streamer_write_uhwi (ob, c->operand_num); |
3897 | streamer_write_uhwi (ob, c->code); | |
b9393656 | 3898 | stream_write_tree (ob, c->val, true); |
8810cc52 MJ |
3899 | bp = bitpack_create (ob->main_stream); |
3900 | bp_pack_value (&bp, c->agg_contents, 1); | |
3901 | bp_pack_value (&bp, c->by_ref, 1); | |
3902 | streamer_write_bitpack (&bp); | |
3903 | if (c->agg_contents) | |
5d59b5e1 | 3904 | streamer_write_uhwi (ob, c->offset); |
632b4f8e | 3905 | } |
412288f1 | 3906 | streamer_write_uhwi (ob, VEC_length (size_time_entry, info->entry)); |
632b4f8e JH |
3907 | for (i = 0; |
3908 | VEC_iterate (size_time_entry, info->entry, i, e); | |
3909 | i++) | |
3910 | { | |
412288f1 DN |
3911 | streamer_write_uhwi (ob, e->size); |
3912 | streamer_write_uhwi (ob, e->time); | |
991278ab | 3913 | write_predicate (ob, &e->predicate); |
632b4f8e | 3914 | } |
2daffc47 | 3915 | write_predicate (ob, info->loop_iterations); |
128e0d89 | 3916 | write_predicate (ob, info->loop_stride); |
898b8927 JH |
3917 | for (edge = node->callees; edge; edge = edge->next_callee) |
3918 | write_inline_edge_summary (ob, edge); | |
3919 | for (edge = node->indirect_calls; edge; edge = edge->next_callee) | |
3920 | write_inline_edge_summary (ob, edge); | |
10a5dd5d JH |
3921 | } |
3922 | } | |
412288f1 | 3923 | streamer_write_char_stream (ob->main_stream, 0); |
632b4f8e JH |
3924 | produce_asm (ob, NULL); |
3925 | destroy_output_block (ob); | |
10a5dd5d | 3926 | |
5ee53a06 | 3927 | if (optimize && !flag_ipa_cp) |
f27c1867 | 3928 | ipa_prop_write_jump_functions (); |
03dfc36d JH |
3929 | } |
3930 | ||
10a5dd5d | 3931 | |
03dfc36d JH |
3932 | /* Release inline summary. */ |
3933 | ||
3934 | void | |
3935 | inline_free_summary (void) | |
3936 | { | |
1c52c601 | 3937 | struct cgraph_node *node; |
a8da72b8 L |
3938 | if (inline_edge_summary_vec == NULL) |
3939 | return; | |
1c52c601 JH |
3940 | FOR_EACH_DEFINED_FUNCTION (node) |
3941 | reset_inline_summary (node); | |
10a5dd5d JH |
3942 | if (function_insertion_hook_holder) |
3943 | cgraph_remove_function_insertion_hook (function_insertion_hook_holder); | |
3944 | function_insertion_hook_holder = NULL; | |
3945 | if (node_removal_hook_holder) | |
3946 | cgraph_remove_node_removal_hook (node_removal_hook_holder); | |
1c52c601 | 3947 | node_removal_hook_holder = NULL; |
898b8927 JH |
3948 | if (edge_removal_hook_holder) |
3949 | cgraph_remove_edge_removal_hook (edge_removal_hook_holder); | |
1c52c601 | 3950 | edge_removal_hook_holder = NULL; |
10a5dd5d JH |
3951 | if (node_duplication_hook_holder) |
3952 | cgraph_remove_node_duplication_hook (node_duplication_hook_holder); | |
1c52c601 | 3953 | node_duplication_hook_holder = NULL; |
898b8927 JH |
3954 | if (edge_duplication_hook_holder) |
3955 | cgraph_remove_edge_duplication_hook (edge_duplication_hook_holder); | |
1c52c601 | 3956 | edge_duplication_hook_holder = NULL; |
632b4f8e JH |
3957 | VEC_free (inline_summary_t, gc, inline_summary_vec); |
3958 | inline_summary_vec = NULL; | |
991278ab JH |
3959 | VEC_free (inline_edge_summary_t, heap, inline_edge_summary_vec); |
3960 | inline_edge_summary_vec = NULL; | |
3961 | if (edge_predicate_pool) | |
3962 | free_alloc_pool (edge_predicate_pool); | |
3963 | edge_predicate_pool = 0; | |
03dfc36d | 3964 | } |