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