* ipa-devirt.c (type_pair, default_hashset_traits): New types.
[official-gcc.git] / gcc / ipa-inline-analysis.c
blob38f56d2fcd5cd88533c00426b5624dc0e229cffe
1 /* Inlining decision heuristics.
2 Copyright (C) 2003-2014 Free Software Foundation, Inc.
3 Contributed by Jan Hubicka
5 This file is part of GCC.
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.
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.
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/>. */
21 /* Analysis used by the inliner and other passes limiting code size growth.
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
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).
38 We provide accestor to the inline_summary datastructure and
39 basic logic updating the parameters when inlining is performed.
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.
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.
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.
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. */
67 #include "config.h"
68 #include "system.h"
69 #include "coretypes.h"
70 #include "tm.h"
71 #include "tree.h"
72 #include "stor-layout.h"
73 #include "stringpool.h"
74 #include "print-tree.h"
75 #include "tree-inline.h"
76 #include "langhooks.h"
77 #include "flags.h"
78 #include "diagnostic.h"
79 #include "gimple-pretty-print.h"
80 #include "params.h"
81 #include "tree-pass.h"
82 #include "coverage.h"
83 #include "basic-block.h"
84 #include "tree-ssa-alias.h"
85 #include "internal-fn.h"
86 #include "gimple-expr.h"
87 #include "is-a.h"
88 #include "gimple.h"
89 #include "gimple-iterator.h"
90 #include "gimple-ssa.h"
91 #include "tree-cfg.h"
92 #include "tree-phinodes.h"
93 #include "ssa-iterators.h"
94 #include "tree-ssanames.h"
95 #include "tree-ssa-loop-niter.h"
96 #include "tree-ssa-loop.h"
97 #include "ipa-prop.h"
98 #include "lto-streamer.h"
99 #include "data-streamer.h"
100 #include "tree-streamer.h"
101 #include "ipa-inline.h"
102 #include "alloc-pool.h"
103 #include "cfgloop.h"
104 #include "tree-scalar-evolution.h"
105 #include "ipa-utils.h"
106 #include "cilk.h"
107 #include "cfgexpand.h"
109 /* Estimate runtime of function can easilly run into huge numbers with many
110 nested loops. Be sure we can compute time * INLINE_SIZE_SCALE * 2 in an
111 integer. For anything larger we use gcov_type. */
112 #define MAX_TIME 500000
114 /* Number of bits in integer, but we really want to be stable across different
115 hosts. */
116 #define NUM_CONDITIONS 32
118 enum predicate_conditions
120 predicate_false_condition = 0,
121 predicate_not_inlined_condition = 1,
122 predicate_first_dynamic_condition = 2
125 /* Special condition code we use to represent test that operand is compile time
126 constant. */
127 #define IS_NOT_CONSTANT ERROR_MARK
128 /* Special condition code we use to represent test that operand is not changed
129 across invocation of the function. When operand IS_NOT_CONSTANT it is always
130 CHANGED, however i.e. loop invariants can be NOT_CHANGED given percentage
131 of executions even when they are not compile time constants. */
132 #define CHANGED IDENTIFIER_NODE
134 /* Holders of ipa cgraph hooks: */
135 static struct cgraph_node_hook_list *function_insertion_hook_holder;
136 static struct cgraph_node_hook_list *node_removal_hook_holder;
137 static struct cgraph_2node_hook_list *node_duplication_hook_holder;
138 static struct cgraph_2edge_hook_list *edge_duplication_hook_holder;
139 static struct cgraph_edge_hook_list *edge_removal_hook_holder;
140 static void inline_node_removal_hook (struct cgraph_node *, void *);
141 static void inline_node_duplication_hook (struct cgraph_node *,
142 struct cgraph_node *, void *);
143 static void inline_edge_removal_hook (struct cgraph_edge *, void *);
144 static void inline_edge_duplication_hook (struct cgraph_edge *,
145 struct cgraph_edge *, void *);
147 /* VECtor holding inline summaries.
148 In GGC memory because conditions might point to constant trees. */
149 vec<inline_summary_t, va_gc> *inline_summary_vec;
150 vec<inline_edge_summary_t> inline_edge_summary_vec;
152 /* Cached node/edge growths. */
153 vec<int> node_growth_cache;
154 vec<edge_growth_cache_entry> edge_growth_cache;
156 /* Edge predicates goes here. */
157 static alloc_pool edge_predicate_pool;
159 /* Return true predicate (tautology).
160 We represent it by empty list of clauses. */
162 static inline struct predicate
163 true_predicate (void)
165 struct predicate p;
166 p.clause[0] = 0;
167 return p;
171 /* Return predicate testing single condition number COND. */
173 static inline struct predicate
174 single_cond_predicate (int cond)
176 struct predicate p;
177 p.clause[0] = 1 << cond;
178 p.clause[1] = 0;
179 return p;
183 /* Return false predicate. First clause require false condition. */
185 static inline struct predicate
186 false_predicate (void)
188 return single_cond_predicate (predicate_false_condition);
192 /* Return true if P is (true). */
194 static inline bool
195 true_predicate_p (struct predicate *p)
197 return !p->clause[0];
201 /* Return true if P is (false). */
203 static inline bool
204 false_predicate_p (struct predicate *p)
206 if (p->clause[0] == (1 << predicate_false_condition))
208 gcc_checking_assert (!p->clause[1]
209 && p->clause[0] == 1 << predicate_false_condition);
210 return true;
212 return false;
216 /* Return predicate that is set true when function is not inlined. */
218 static inline struct predicate
219 not_inlined_predicate (void)
221 return single_cond_predicate (predicate_not_inlined_condition);
224 /* Simple description of whether a memory load or a condition refers to a load
225 from an aggregate and if so, how and where from in the aggregate.
226 Individual fields have the same meaning like fields with the same name in
227 struct condition. */
229 struct agg_position_info
231 HOST_WIDE_INT offset;
232 bool agg_contents;
233 bool by_ref;
236 /* Add condition to condition list CONDS. AGGPOS describes whether the used
237 oprand is loaded from an aggregate and where in the aggregate it is. It can
238 be NULL, which means this not a load from an aggregate. */
240 static struct predicate
241 add_condition (struct inline_summary *summary, int operand_num,
242 struct agg_position_info *aggpos,
243 enum tree_code code, tree val)
245 int i;
246 struct condition *c;
247 struct condition new_cond;
248 HOST_WIDE_INT offset;
249 bool agg_contents, by_ref;
251 if (aggpos)
253 offset = aggpos->offset;
254 agg_contents = aggpos->agg_contents;
255 by_ref = aggpos->by_ref;
257 else
259 offset = 0;
260 agg_contents = false;
261 by_ref = false;
264 gcc_checking_assert (operand_num >= 0);
265 for (i = 0; vec_safe_iterate (summary->conds, i, &c); i++)
267 if (c->operand_num == operand_num
268 && c->code == code
269 && c->val == val
270 && c->agg_contents == agg_contents
271 && (!agg_contents || (c->offset == offset && c->by_ref == by_ref)))
272 return single_cond_predicate (i + predicate_first_dynamic_condition);
274 /* Too many conditions. Give up and return constant true. */
275 if (i == NUM_CONDITIONS - predicate_first_dynamic_condition)
276 return true_predicate ();
278 new_cond.operand_num = operand_num;
279 new_cond.code = code;
280 new_cond.val = val;
281 new_cond.agg_contents = agg_contents;
282 new_cond.by_ref = by_ref;
283 new_cond.offset = offset;
284 vec_safe_push (summary->conds, new_cond);
285 return single_cond_predicate (i + predicate_first_dynamic_condition);
289 /* Add clause CLAUSE into the predicate P. */
291 static inline void
292 add_clause (conditions conditions, struct predicate *p, clause_t clause)
294 int i;
295 int i2;
296 int insert_here = -1;
297 int c1, c2;
299 /* True clause. */
300 if (!clause)
301 return;
303 /* False clause makes the whole predicate false. Kill the other variants. */
304 if (clause == (1 << predicate_false_condition))
306 p->clause[0] = (1 << predicate_false_condition);
307 p->clause[1] = 0;
308 return;
310 if (false_predicate_p (p))
311 return;
313 /* No one should be silly enough to add false into nontrivial clauses. */
314 gcc_checking_assert (!(clause & (1 << predicate_false_condition)));
316 /* Look where to insert the clause. At the same time prune out
317 clauses of P that are implied by the new clause and thus
318 redundant. */
319 for (i = 0, i2 = 0; i <= MAX_CLAUSES; i++)
321 p->clause[i2] = p->clause[i];
323 if (!p->clause[i])
324 break;
326 /* If p->clause[i] implies clause, there is nothing to add. */
327 if ((p->clause[i] & clause) == p->clause[i])
329 /* We had nothing to add, none of clauses should've become
330 redundant. */
331 gcc_checking_assert (i == i2);
332 return;
335 if (p->clause[i] < clause && insert_here < 0)
336 insert_here = i2;
338 /* If clause implies p->clause[i], then p->clause[i] becomes redundant.
339 Otherwise the p->clause[i] has to stay. */
340 if ((p->clause[i] & clause) != clause)
341 i2++;
344 /* Look for clauses that are obviously true. I.e.
345 op0 == 5 || op0 != 5. */
346 for (c1 = predicate_first_dynamic_condition; c1 < NUM_CONDITIONS; c1++)
348 condition *cc1;
349 if (!(clause & (1 << c1)))
350 continue;
351 cc1 = &(*conditions)[c1 - predicate_first_dynamic_condition];
352 /* We have no way to represent !CHANGED and !IS_NOT_CONSTANT
353 and thus there is no point for looking for them. */
354 if (cc1->code == CHANGED || cc1->code == IS_NOT_CONSTANT)
355 continue;
356 for (c2 = c1 + 1; c2 < NUM_CONDITIONS; c2++)
357 if (clause & (1 << c2))
359 condition *cc1 =
360 &(*conditions)[c1 - predicate_first_dynamic_condition];
361 condition *cc2 =
362 &(*conditions)[c2 - predicate_first_dynamic_condition];
363 if (cc1->operand_num == cc2->operand_num
364 && cc1->val == cc2->val
365 && cc2->code != IS_NOT_CONSTANT
366 && cc2->code != CHANGED
367 && cc1->code == invert_tree_comparison
368 (cc2->code,
369 HONOR_NANS (TYPE_MODE (TREE_TYPE (cc1->val)))))
370 return;
375 /* We run out of variants. Be conservative in positive direction. */
376 if (i2 == MAX_CLAUSES)
377 return;
378 /* Keep clauses in decreasing order. This makes equivalence testing easy. */
379 p->clause[i2 + 1] = 0;
380 if (insert_here >= 0)
381 for (; i2 > insert_here; i2--)
382 p->clause[i2] = p->clause[i2 - 1];
383 else
384 insert_here = i2;
385 p->clause[insert_here] = clause;
389 /* Return P & P2. */
391 static struct predicate
392 and_predicates (conditions conditions,
393 struct predicate *p, struct predicate *p2)
395 struct predicate out = *p;
396 int i;
398 /* Avoid busy work. */
399 if (false_predicate_p (p2) || true_predicate_p (p))
400 return *p2;
401 if (false_predicate_p (p) || true_predicate_p (p2))
402 return *p;
404 /* See how far predicates match. */
405 for (i = 0; p->clause[i] && p->clause[i] == p2->clause[i]; i++)
407 gcc_checking_assert (i < MAX_CLAUSES);
410 /* Combine the predicates rest. */
411 for (; p2->clause[i]; i++)
413 gcc_checking_assert (i < MAX_CLAUSES);
414 add_clause (conditions, &out, p2->clause[i]);
416 return out;
420 /* Return true if predicates are obviously equal. */
422 static inline bool
423 predicates_equal_p (struct predicate *p, struct predicate *p2)
425 int i;
426 for (i = 0; p->clause[i]; i++)
428 gcc_checking_assert (i < MAX_CLAUSES);
429 gcc_checking_assert (p->clause[i] > p->clause[i + 1]);
430 gcc_checking_assert (!p2->clause[i]
431 || p2->clause[i] > p2->clause[i + 1]);
432 if (p->clause[i] != p2->clause[i])
433 return false;
435 return !p2->clause[i];
439 /* Return P | P2. */
441 static struct predicate
442 or_predicates (conditions conditions,
443 struct predicate *p, struct predicate *p2)
445 struct predicate out = true_predicate ();
446 int i, j;
448 /* Avoid busy work. */
449 if (false_predicate_p (p2) || true_predicate_p (p))
450 return *p;
451 if (false_predicate_p (p) || true_predicate_p (p2))
452 return *p2;
453 if (predicates_equal_p (p, p2))
454 return *p;
456 /* OK, combine the predicates. */
457 for (i = 0; p->clause[i]; i++)
458 for (j = 0; p2->clause[j]; j++)
460 gcc_checking_assert (i < MAX_CLAUSES && j < MAX_CLAUSES);
461 add_clause (conditions, &out, p->clause[i] | p2->clause[j]);
463 return out;
467 /* Having partial truth assignment in POSSIBLE_TRUTHS, return false
468 if predicate P is known to be false. */
470 static bool
471 evaluate_predicate (struct predicate *p, clause_t possible_truths)
473 int i;
475 /* True remains true. */
476 if (true_predicate_p (p))
477 return true;
479 gcc_assert (!(possible_truths & (1 << predicate_false_condition)));
481 /* See if we can find clause we can disprove. */
482 for (i = 0; p->clause[i]; i++)
484 gcc_checking_assert (i < MAX_CLAUSES);
485 if (!(p->clause[i] & possible_truths))
486 return false;
488 return true;
491 /* Return the probability in range 0...REG_BR_PROB_BASE that the predicated
492 instruction will be recomputed per invocation of the inlined call. */
494 static int
495 predicate_probability (conditions conds,
496 struct predicate *p, clause_t possible_truths,
497 vec<inline_param_summary> inline_param_summary)
499 int i;
500 int combined_prob = REG_BR_PROB_BASE;
502 /* True remains true. */
503 if (true_predicate_p (p))
504 return REG_BR_PROB_BASE;
506 if (false_predicate_p (p))
507 return 0;
509 gcc_assert (!(possible_truths & (1 << predicate_false_condition)));
511 /* See if we can find clause we can disprove. */
512 for (i = 0; p->clause[i]; i++)
514 gcc_checking_assert (i < MAX_CLAUSES);
515 if (!(p->clause[i] & possible_truths))
516 return 0;
517 else
519 int this_prob = 0;
520 int i2;
521 if (!inline_param_summary.exists ())
522 return REG_BR_PROB_BASE;
523 for (i2 = 0; i2 < NUM_CONDITIONS; i2++)
524 if ((p->clause[i] & possible_truths) & (1 << i2))
526 if (i2 >= predicate_first_dynamic_condition)
528 condition *c =
529 &(*conds)[i2 - predicate_first_dynamic_condition];
530 if (c->code == CHANGED
531 && (c->operand_num <
532 (int) inline_param_summary.length ()))
534 int iprob =
535 inline_param_summary[c->operand_num].change_prob;
536 this_prob = MAX (this_prob, iprob);
538 else
539 this_prob = REG_BR_PROB_BASE;
541 else
542 this_prob = REG_BR_PROB_BASE;
544 combined_prob = MIN (this_prob, combined_prob);
545 if (!combined_prob)
546 return 0;
549 return combined_prob;
553 /* Dump conditional COND. */
555 static void
556 dump_condition (FILE *f, conditions conditions, int cond)
558 condition *c;
559 if (cond == predicate_false_condition)
560 fprintf (f, "false");
561 else if (cond == predicate_not_inlined_condition)
562 fprintf (f, "not inlined");
563 else
565 c = &(*conditions)[cond - predicate_first_dynamic_condition];
566 fprintf (f, "op%i", c->operand_num);
567 if (c->agg_contents)
568 fprintf (f, "[%soffset: " HOST_WIDE_INT_PRINT_DEC "]",
569 c->by_ref ? "ref " : "", c->offset);
570 if (c->code == IS_NOT_CONSTANT)
572 fprintf (f, " not constant");
573 return;
575 if (c->code == CHANGED)
577 fprintf (f, " changed");
578 return;
580 fprintf (f, " %s ", op_symbol_code (c->code));
581 print_generic_expr (f, c->val, 1);
586 /* Dump clause CLAUSE. */
588 static void
589 dump_clause (FILE *f, conditions conds, clause_t clause)
591 int i;
592 bool found = false;
593 fprintf (f, "(");
594 if (!clause)
595 fprintf (f, "true");
596 for (i = 0; i < NUM_CONDITIONS; i++)
597 if (clause & (1 << i))
599 if (found)
600 fprintf (f, " || ");
601 found = true;
602 dump_condition (f, conds, i);
604 fprintf (f, ")");
608 /* Dump predicate PREDICATE. */
610 static void
611 dump_predicate (FILE *f, conditions conds, struct predicate *pred)
613 int i;
614 if (true_predicate_p (pred))
615 dump_clause (f, conds, 0);
616 else
617 for (i = 0; pred->clause[i]; i++)
619 if (i)
620 fprintf (f, " && ");
621 dump_clause (f, conds, pred->clause[i]);
623 fprintf (f, "\n");
627 /* Dump inline hints. */
628 void
629 dump_inline_hints (FILE *f, inline_hints hints)
631 if (!hints)
632 return;
633 fprintf (f, "inline hints:");
634 if (hints & INLINE_HINT_indirect_call)
636 hints &= ~INLINE_HINT_indirect_call;
637 fprintf (f, " indirect_call");
639 if (hints & INLINE_HINT_loop_iterations)
641 hints &= ~INLINE_HINT_loop_iterations;
642 fprintf (f, " loop_iterations");
644 if (hints & INLINE_HINT_loop_stride)
646 hints &= ~INLINE_HINT_loop_stride;
647 fprintf (f, " loop_stride");
649 if (hints & INLINE_HINT_same_scc)
651 hints &= ~INLINE_HINT_same_scc;
652 fprintf (f, " same_scc");
654 if (hints & INLINE_HINT_in_scc)
656 hints &= ~INLINE_HINT_in_scc;
657 fprintf (f, " in_scc");
659 if (hints & INLINE_HINT_cross_module)
661 hints &= ~INLINE_HINT_cross_module;
662 fprintf (f, " cross_module");
664 if (hints & INLINE_HINT_declared_inline)
666 hints &= ~INLINE_HINT_declared_inline;
667 fprintf (f, " declared_inline");
669 if (hints & INLINE_HINT_array_index)
671 hints &= ~INLINE_HINT_array_index;
672 fprintf (f, " array_index");
674 if (hints & INLINE_HINT_known_hot)
676 hints &= ~INLINE_HINT_known_hot;
677 fprintf (f, " known_hot");
679 gcc_assert (!hints);
683 /* Record SIZE and TIME under condition PRED into the inline summary. */
685 static void
686 account_size_time (struct inline_summary *summary, int size, int time,
687 struct predicate *pred)
689 size_time_entry *e;
690 bool found = false;
691 int i;
693 if (false_predicate_p (pred))
694 return;
696 /* We need to create initial empty unconitional clause, but otherwie
697 we don't need to account empty times and sizes. */
698 if (!size && !time && summary->entry)
699 return;
701 /* Watch overflow that might result from insane profiles. */
702 if (time > MAX_TIME * INLINE_TIME_SCALE)
703 time = MAX_TIME * INLINE_TIME_SCALE;
704 gcc_assert (time >= 0);
706 for (i = 0; vec_safe_iterate (summary->entry, i, &e); i++)
707 if (predicates_equal_p (&e->predicate, pred))
709 found = true;
710 break;
712 if (i == 256)
714 i = 0;
715 found = true;
716 e = &(*summary->entry)[0];
717 gcc_assert (!e->predicate.clause[0]);
718 if (dump_file && (dump_flags & TDF_DETAILS))
719 fprintf (dump_file,
720 "\t\tReached limit on number of entries, "
721 "ignoring the predicate.");
723 if (dump_file && (dump_flags & TDF_DETAILS) && (time || size))
725 fprintf (dump_file,
726 "\t\tAccounting size:%3.2f, time:%3.2f on %spredicate:",
727 ((double) size) / INLINE_SIZE_SCALE,
728 ((double) time) / INLINE_TIME_SCALE, found ? "" : "new ");
729 dump_predicate (dump_file, summary->conds, pred);
731 if (!found)
733 struct size_time_entry new_entry;
734 new_entry.size = size;
735 new_entry.time = time;
736 new_entry.predicate = *pred;
737 vec_safe_push (summary->entry, new_entry);
739 else
741 e->size += size;
742 e->time += time;
743 if (e->time > MAX_TIME * INLINE_TIME_SCALE)
744 e->time = MAX_TIME * INLINE_TIME_SCALE;
748 /* Set predicate for edge E. */
750 static void
751 edge_set_predicate (struct cgraph_edge *e, struct predicate *predicate)
753 struct inline_edge_summary *es = inline_edge_summary (e);
755 /* If the edge is determined to be never executed, redirect it
756 to BUILTIN_UNREACHABLE to save inliner from inlining into it. */
757 if (predicate && false_predicate_p (predicate) && e->callee)
759 struct cgraph_node *callee = !e->inline_failed ? e->callee : NULL;
761 e->redirect_callee (cgraph_node::get_create
762 (builtin_decl_implicit (BUILT_IN_UNREACHABLE)));
763 e->inline_failed = CIF_UNREACHABLE;
764 if (callee)
765 callee->remove_symbol_and_inline_clones ();
767 if (predicate && !true_predicate_p (predicate))
769 if (!es->predicate)
770 es->predicate = (struct predicate *) pool_alloc (edge_predicate_pool);
771 *es->predicate = *predicate;
773 else
775 if (es->predicate)
776 pool_free (edge_predicate_pool, es->predicate);
777 es->predicate = NULL;
781 /* Set predicate for hint *P. */
783 static void
784 set_hint_predicate (struct predicate **p, struct predicate new_predicate)
786 if (false_predicate_p (&new_predicate) || true_predicate_p (&new_predicate))
788 if (*p)
789 pool_free (edge_predicate_pool, *p);
790 *p = NULL;
792 else
794 if (!*p)
795 *p = (struct predicate *) pool_alloc (edge_predicate_pool);
796 **p = new_predicate;
801 /* KNOWN_VALS is partial mapping of parameters of NODE to constant values.
802 KNOWN_AGGS is a vector of aggreggate jump functions for each parameter.
803 Return clause of possible truths. When INLINE_P is true, assume that we are
804 inlining.
806 ERROR_MARK means compile time invariant. */
808 static clause_t
809 evaluate_conditions_for_known_args (struct cgraph_node *node,
810 bool inline_p,
811 vec<tree> known_vals,
812 vec<ipa_agg_jump_function_p>
813 known_aggs)
815 clause_t clause = inline_p ? 0 : 1 << predicate_not_inlined_condition;
816 struct inline_summary *info = inline_summary (node);
817 int i;
818 struct condition *c;
820 for (i = 0; vec_safe_iterate (info->conds, i, &c); i++)
822 tree val;
823 tree res;
825 /* We allow call stmt to have fewer arguments than the callee function
826 (especially for K&R style programs). So bound check here (we assume
827 known_aggs vector, if non-NULL, has the same length as
828 known_vals). */
829 gcc_checking_assert (!known_aggs.exists ()
830 || (known_vals.length () == known_aggs.length ()));
831 if (c->operand_num >= (int) known_vals.length ())
833 clause |= 1 << (i + predicate_first_dynamic_condition);
834 continue;
837 if (c->agg_contents)
839 struct ipa_agg_jump_function *agg;
841 if (c->code == CHANGED
842 && !c->by_ref
843 && (known_vals[c->operand_num] == error_mark_node))
844 continue;
846 if (known_aggs.exists ())
848 agg = known_aggs[c->operand_num];
849 val = ipa_find_agg_cst_for_param (agg, c->offset, c->by_ref);
851 else
852 val = NULL_TREE;
854 else
856 val = known_vals[c->operand_num];
857 if (val == error_mark_node && c->code != CHANGED)
858 val = NULL_TREE;
861 if (!val)
863 clause |= 1 << (i + predicate_first_dynamic_condition);
864 continue;
866 if (c->code == IS_NOT_CONSTANT || c->code == CHANGED)
867 continue;
868 res = fold_binary_to_constant (c->code, boolean_type_node, val, c->val);
869 if (res && integer_zerop (res))
870 continue;
871 clause |= 1 << (i + predicate_first_dynamic_condition);
873 return clause;
877 /* Work out what conditions might be true at invocation of E. */
879 static void
880 evaluate_properties_for_edge (struct cgraph_edge *e, bool inline_p,
881 clause_t *clause_ptr,
882 vec<tree> *known_vals_ptr,
883 vec<tree> *known_binfos_ptr,
884 vec<ipa_agg_jump_function_p> *known_aggs_ptr)
886 struct cgraph_node *callee = e->callee->ultimate_alias_target ();
887 struct inline_summary *info = inline_summary (callee);
888 vec<tree> known_vals = vNULL;
889 vec<ipa_agg_jump_function_p> known_aggs = vNULL;
891 if (clause_ptr)
892 *clause_ptr = inline_p ? 0 : 1 << predicate_not_inlined_condition;
893 if (known_vals_ptr)
894 known_vals_ptr->create (0);
895 if (known_binfos_ptr)
896 known_binfos_ptr->create (0);
898 if (ipa_node_params_vector.exists ()
899 && !e->call_stmt_cannot_inline_p
900 && ((clause_ptr && info->conds) || known_vals_ptr || known_binfos_ptr))
902 struct ipa_node_params *parms_info;
903 struct ipa_edge_args *args = IPA_EDGE_REF (e);
904 struct inline_edge_summary *es = inline_edge_summary (e);
905 int i, count = ipa_get_cs_argument_count (args);
907 if (e->caller->global.inlined_to)
908 parms_info = IPA_NODE_REF (e->caller->global.inlined_to);
909 else
910 parms_info = IPA_NODE_REF (e->caller);
912 if (count && (info->conds || known_vals_ptr))
913 known_vals.safe_grow_cleared (count);
914 if (count && (info->conds || known_aggs_ptr))
915 known_aggs.safe_grow_cleared (count);
916 if (count && known_binfos_ptr)
917 known_binfos_ptr->safe_grow_cleared (count);
919 for (i = 0; i < count; i++)
921 struct ipa_jump_func *jf = ipa_get_ith_jump_func (args, i);
922 tree cst = ipa_value_from_jfunc (parms_info, jf);
923 if (cst)
925 if (known_vals.exists () && TREE_CODE (cst) != TREE_BINFO)
926 known_vals[i] = cst;
927 else if (known_binfos_ptr != NULL
928 && TREE_CODE (cst) == TREE_BINFO)
929 (*known_binfos_ptr)[i] = cst;
931 else if (inline_p && !es->param[i].change_prob)
932 known_vals[i] = error_mark_node;
933 /* TODO: When IPA-CP starts propagating and merging aggregate jump
934 functions, use its knowledge of the caller too, just like the
935 scalar case above. */
936 known_aggs[i] = &jf->agg;
940 if (clause_ptr)
941 *clause_ptr = evaluate_conditions_for_known_args (callee, inline_p,
942 known_vals, known_aggs);
944 if (known_vals_ptr)
945 *known_vals_ptr = known_vals;
946 else
947 known_vals.release ();
949 if (known_aggs_ptr)
950 *known_aggs_ptr = known_aggs;
951 else
952 known_aggs.release ();
956 /* Allocate the inline summary vector or resize it to cover all cgraph nodes. */
958 static void
959 inline_summary_alloc (void)
961 if (!node_removal_hook_holder)
962 node_removal_hook_holder =
963 symtab->add_cgraph_removal_hook (&inline_node_removal_hook, NULL);
964 if (!edge_removal_hook_holder)
965 edge_removal_hook_holder =
966 symtab->add_edge_removal_hook (&inline_edge_removal_hook, NULL);
967 if (!node_duplication_hook_holder)
968 node_duplication_hook_holder =
969 symtab->add_cgraph_duplication_hook (&inline_node_duplication_hook, NULL);
970 if (!edge_duplication_hook_holder)
971 edge_duplication_hook_holder =
972 symtab->add_edge_duplication_hook (&inline_edge_duplication_hook, NULL);
974 if (vec_safe_length (inline_summary_vec) <= (unsigned) symtab->cgraph_max_uid)
975 vec_safe_grow_cleared (inline_summary_vec, symtab->cgraph_max_uid + 1);
976 if (inline_edge_summary_vec.length () <= (unsigned) symtab->edges_max_uid)
977 inline_edge_summary_vec.safe_grow_cleared (symtab->edges_max_uid + 1);
978 if (!edge_predicate_pool)
979 edge_predicate_pool = create_alloc_pool ("edge predicates",
980 sizeof (struct predicate), 10);
983 /* We are called multiple time for given function; clear
984 data from previous run so they are not cumulated. */
986 static void
987 reset_inline_edge_summary (struct cgraph_edge *e)
989 if (e->uid < (int) inline_edge_summary_vec.length ())
991 struct inline_edge_summary *es = inline_edge_summary (e);
993 es->call_stmt_size = es->call_stmt_time = 0;
994 if (es->predicate)
995 pool_free (edge_predicate_pool, es->predicate);
996 es->predicate = NULL;
997 es->param.release ();
1001 /* We are called multiple time for given function; clear
1002 data from previous run so they are not cumulated. */
1004 static void
1005 reset_inline_summary (struct cgraph_node *node)
1007 struct inline_summary *info = inline_summary (node);
1008 struct cgraph_edge *e;
1010 info->self_size = info->self_time = 0;
1011 info->estimated_stack_size = 0;
1012 info->estimated_self_stack_size = 0;
1013 info->stack_frame_offset = 0;
1014 info->size = 0;
1015 info->time = 0;
1016 info->growth = 0;
1017 info->scc_no = 0;
1018 if (info->loop_iterations)
1020 pool_free (edge_predicate_pool, info->loop_iterations);
1021 info->loop_iterations = NULL;
1023 if (info->loop_stride)
1025 pool_free (edge_predicate_pool, info->loop_stride);
1026 info->loop_stride = NULL;
1028 if (info->array_index)
1030 pool_free (edge_predicate_pool, info->array_index);
1031 info->array_index = NULL;
1033 vec_free (info->conds);
1034 vec_free (info->entry);
1035 for (e = node->callees; e; e = e->next_callee)
1036 reset_inline_edge_summary (e);
1037 for (e = node->indirect_calls; e; e = e->next_callee)
1038 reset_inline_edge_summary (e);
1041 /* Hook that is called by cgraph.c when a node is removed. */
1043 static void
1044 inline_node_removal_hook (struct cgraph_node *node,
1045 void *data ATTRIBUTE_UNUSED)
1047 struct inline_summary *info;
1048 if (vec_safe_length (inline_summary_vec) <= (unsigned) node->uid)
1049 return;
1050 info = inline_summary (node);
1051 reset_inline_summary (node);
1052 memset (info, 0, sizeof (inline_summary_t));
1055 /* Remap predicate P of former function to be predicate of duplicated function.
1056 POSSIBLE_TRUTHS is clause of possible truths in the duplicated node,
1057 INFO is inline summary of the duplicated node. */
1059 static struct predicate
1060 remap_predicate_after_duplication (struct predicate *p,
1061 clause_t possible_truths,
1062 struct inline_summary *info)
1064 struct predicate new_predicate = true_predicate ();
1065 int j;
1066 for (j = 0; p->clause[j]; j++)
1067 if (!(possible_truths & p->clause[j]))
1069 new_predicate = false_predicate ();
1070 break;
1072 else
1073 add_clause (info->conds, &new_predicate,
1074 possible_truths & p->clause[j]);
1075 return new_predicate;
1078 /* Same as remap_predicate_after_duplication but handle hint predicate *P.
1079 Additionally care about allocating new memory slot for updated predicate
1080 and set it to NULL when it becomes true or false (and thus uninteresting).
1083 static void
1084 remap_hint_predicate_after_duplication (struct predicate **p,
1085 clause_t possible_truths,
1086 struct inline_summary *info)
1088 struct predicate new_predicate;
1090 if (!*p)
1091 return;
1093 new_predicate = remap_predicate_after_duplication (*p,
1094 possible_truths, info);
1095 /* We do not want to free previous predicate; it is used by node origin. */
1096 *p = NULL;
1097 set_hint_predicate (p, new_predicate);
1101 /* Hook that is called by cgraph.c when a node is duplicated. */
1103 static void
1104 inline_node_duplication_hook (struct cgraph_node *src,
1105 struct cgraph_node *dst,
1106 ATTRIBUTE_UNUSED void *data)
1108 struct inline_summary *info;
1109 inline_summary_alloc ();
1110 info = inline_summary (dst);
1111 memcpy (info, inline_summary (src), sizeof (struct inline_summary));
1112 /* TODO: as an optimization, we may avoid copying conditions
1113 that are known to be false or true. */
1114 info->conds = vec_safe_copy (info->conds);
1116 /* When there are any replacements in the function body, see if we can figure
1117 out that something was optimized out. */
1118 if (ipa_node_params_vector.exists () && dst->clone.tree_map)
1120 vec<size_time_entry, va_gc> *entry = info->entry;
1121 /* Use SRC parm info since it may not be copied yet. */
1122 struct ipa_node_params *parms_info = IPA_NODE_REF (src);
1123 vec<tree> known_vals = vNULL;
1124 int count = ipa_get_param_count (parms_info);
1125 int i, j;
1126 clause_t possible_truths;
1127 struct predicate true_pred = true_predicate ();
1128 size_time_entry *e;
1129 int optimized_out_size = 0;
1130 bool inlined_to_p = false;
1131 struct cgraph_edge *edge;
1133 info->entry = 0;
1134 known_vals.safe_grow_cleared (count);
1135 for (i = 0; i < count; i++)
1137 struct ipa_replace_map *r;
1139 for (j = 0; vec_safe_iterate (dst->clone.tree_map, j, &r); j++)
1141 if (((!r->old_tree && r->parm_num == i)
1142 || (r->old_tree && r->old_tree == ipa_get_param (parms_info, i)))
1143 && r->replace_p && !r->ref_p)
1145 known_vals[i] = r->new_tree;
1146 break;
1150 possible_truths = evaluate_conditions_for_known_args (dst, false,
1151 known_vals,
1152 vNULL);
1153 known_vals.release ();
1155 account_size_time (info, 0, 0, &true_pred);
1157 /* Remap size_time vectors.
1158 Simplify the predicate by prunning out alternatives that are known
1159 to be false.
1160 TODO: as on optimization, we can also eliminate conditions known
1161 to be true. */
1162 for (i = 0; vec_safe_iterate (entry, i, &e); i++)
1164 struct predicate new_predicate;
1165 new_predicate = remap_predicate_after_duplication (&e->predicate,
1166 possible_truths,
1167 info);
1168 if (false_predicate_p (&new_predicate))
1169 optimized_out_size += e->size;
1170 else
1171 account_size_time (info, e->size, e->time, &new_predicate);
1174 /* Remap edge predicates with the same simplification as above.
1175 Also copy constantness arrays. */
1176 for (edge = dst->callees; edge; edge = edge->next_callee)
1178 struct predicate new_predicate;
1179 struct inline_edge_summary *es = inline_edge_summary (edge);
1181 if (!edge->inline_failed)
1182 inlined_to_p = true;
1183 if (!es->predicate)
1184 continue;
1185 new_predicate = remap_predicate_after_duplication (es->predicate,
1186 possible_truths,
1187 info);
1188 if (false_predicate_p (&new_predicate)
1189 && !false_predicate_p (es->predicate))
1191 optimized_out_size += es->call_stmt_size * INLINE_SIZE_SCALE;
1192 edge->frequency = 0;
1194 edge_set_predicate (edge, &new_predicate);
1197 /* Remap indirect edge predicates with the same simplificaiton as above.
1198 Also copy constantness arrays. */
1199 for (edge = dst->indirect_calls; edge; edge = edge->next_callee)
1201 struct predicate new_predicate;
1202 struct inline_edge_summary *es = inline_edge_summary (edge);
1204 gcc_checking_assert (edge->inline_failed);
1205 if (!es->predicate)
1206 continue;
1207 new_predicate = remap_predicate_after_duplication (es->predicate,
1208 possible_truths,
1209 info);
1210 if (false_predicate_p (&new_predicate)
1211 && !false_predicate_p (es->predicate))
1213 optimized_out_size += es->call_stmt_size * INLINE_SIZE_SCALE;
1214 edge->frequency = 0;
1216 edge_set_predicate (edge, &new_predicate);
1218 remap_hint_predicate_after_duplication (&info->loop_iterations,
1219 possible_truths, info);
1220 remap_hint_predicate_after_duplication (&info->loop_stride,
1221 possible_truths, info);
1222 remap_hint_predicate_after_duplication (&info->array_index,
1223 possible_truths, info);
1225 /* If inliner or someone after inliner will ever start producing
1226 non-trivial clones, we will get trouble with lack of information
1227 about updating self sizes, because size vectors already contains
1228 sizes of the calees. */
1229 gcc_assert (!inlined_to_p || !optimized_out_size);
1231 else
1233 info->entry = vec_safe_copy (info->entry);
1234 if (info->loop_iterations)
1236 predicate p = *info->loop_iterations;
1237 info->loop_iterations = NULL;
1238 set_hint_predicate (&info->loop_iterations, p);
1240 if (info->loop_stride)
1242 predicate p = *info->loop_stride;
1243 info->loop_stride = NULL;
1244 set_hint_predicate (&info->loop_stride, p);
1246 if (info->array_index)
1248 predicate p = *info->array_index;
1249 info->array_index = NULL;
1250 set_hint_predicate (&info->array_index, p);
1253 inline_update_overall_summary (dst);
1257 /* Hook that is called by cgraph.c when a node is duplicated. */
1259 static void
1260 inline_edge_duplication_hook (struct cgraph_edge *src,
1261 struct cgraph_edge *dst,
1262 ATTRIBUTE_UNUSED void *data)
1264 struct inline_edge_summary *info;
1265 struct inline_edge_summary *srcinfo;
1266 inline_summary_alloc ();
1267 info = inline_edge_summary (dst);
1268 srcinfo = inline_edge_summary (src);
1269 memcpy (info, srcinfo, sizeof (struct inline_edge_summary));
1270 info->predicate = NULL;
1271 edge_set_predicate (dst, srcinfo->predicate);
1272 info->param = srcinfo->param.copy ();
1276 /* Keep edge cache consistent across edge removal. */
1278 static void
1279 inline_edge_removal_hook (struct cgraph_edge *edge,
1280 void *data ATTRIBUTE_UNUSED)
1282 if (edge_growth_cache.exists ())
1283 reset_edge_growth_cache (edge);
1284 reset_inline_edge_summary (edge);
1288 /* Initialize growth caches. */
1290 void
1291 initialize_growth_caches (void)
1293 if (symtab->edges_max_uid)
1294 edge_growth_cache.safe_grow_cleared (symtab->edges_max_uid);
1295 if (symtab->cgraph_max_uid)
1296 node_growth_cache.safe_grow_cleared (symtab->cgraph_max_uid);
1300 /* Free growth caches. */
1302 void
1303 free_growth_caches (void)
1305 edge_growth_cache.release ();
1306 node_growth_cache.release ();
1310 /* Dump edge summaries associated to NODE and recursively to all clones.
1311 Indent by INDENT. */
1313 static void
1314 dump_inline_edge_summary (FILE *f, int indent, struct cgraph_node *node,
1315 struct inline_summary *info)
1317 struct cgraph_edge *edge;
1318 for (edge = node->callees; edge; edge = edge->next_callee)
1320 struct inline_edge_summary *es = inline_edge_summary (edge);
1321 struct cgraph_node *callee = edge->callee->ultimate_alias_target ();
1322 int i;
1324 fprintf (f,
1325 "%*s%s/%i %s\n%*s loop depth:%2i freq:%4i size:%2i"
1326 " time: %2i callee size:%2i stack:%2i",
1327 indent, "", callee->name (), callee->order,
1328 !edge->inline_failed
1329 ? "inlined" : cgraph_inline_failed_string (edge-> inline_failed),
1330 indent, "", es->loop_depth, edge->frequency,
1331 es->call_stmt_size, es->call_stmt_time,
1332 (int) inline_summary (callee)->size / INLINE_SIZE_SCALE,
1333 (int) inline_summary (callee)->estimated_stack_size);
1335 if (es->predicate)
1337 fprintf (f, " predicate: ");
1338 dump_predicate (f, info->conds, es->predicate);
1340 else
1341 fprintf (f, "\n");
1342 if (es->param.exists ())
1343 for (i = 0; i < (int) es->param.length (); i++)
1345 int prob = es->param[i].change_prob;
1347 if (!prob)
1348 fprintf (f, "%*s op%i is compile time invariant\n",
1349 indent + 2, "", i);
1350 else if (prob != REG_BR_PROB_BASE)
1351 fprintf (f, "%*s op%i change %f%% of time\n", indent + 2, "", i,
1352 prob * 100.0 / REG_BR_PROB_BASE);
1354 if (!edge->inline_failed)
1356 fprintf (f, "%*sStack frame offset %i, callee self size %i,"
1357 " callee size %i\n",
1358 indent + 2, "",
1359 (int) inline_summary (callee)->stack_frame_offset,
1360 (int) inline_summary (callee)->estimated_self_stack_size,
1361 (int) inline_summary (callee)->estimated_stack_size);
1362 dump_inline_edge_summary (f, indent + 2, callee, info);
1365 for (edge = node->indirect_calls; edge; edge = edge->next_callee)
1367 struct inline_edge_summary *es = inline_edge_summary (edge);
1368 fprintf (f, "%*sindirect call loop depth:%2i freq:%4i size:%2i"
1369 " time: %2i",
1370 indent, "",
1371 es->loop_depth,
1372 edge->frequency, es->call_stmt_size, es->call_stmt_time);
1373 if (es->predicate)
1375 fprintf (f, "predicate: ");
1376 dump_predicate (f, info->conds, es->predicate);
1378 else
1379 fprintf (f, "\n");
1384 void
1385 dump_inline_summary (FILE *f, struct cgraph_node *node)
1387 if (node->definition)
1389 struct inline_summary *s = inline_summary (node);
1390 size_time_entry *e;
1391 int i;
1392 fprintf (f, "Inline summary for %s/%i", node->name (),
1393 node->order);
1394 if (DECL_DISREGARD_INLINE_LIMITS (node->decl))
1395 fprintf (f, " always_inline");
1396 if (s->inlinable)
1397 fprintf (f, " inlinable");
1398 fprintf (f, "\n self time: %i\n", s->self_time);
1399 fprintf (f, " global time: %i\n", s->time);
1400 fprintf (f, " self size: %i\n", s->self_size);
1401 fprintf (f, " global size: %i\n", s->size);
1402 fprintf (f, " min size: %i\n", s->min_size);
1403 fprintf (f, " self stack: %i\n",
1404 (int) s->estimated_self_stack_size);
1405 fprintf (f, " global stack: %i\n", (int) s->estimated_stack_size);
1406 if (s->growth)
1407 fprintf (f, " estimated growth:%i\n", (int) s->growth);
1408 if (s->scc_no)
1409 fprintf (f, " In SCC: %i\n", (int) s->scc_no);
1410 for (i = 0; vec_safe_iterate (s->entry, i, &e); i++)
1412 fprintf (f, " size:%f, time:%f, predicate:",
1413 (double) e->size / INLINE_SIZE_SCALE,
1414 (double) e->time / INLINE_TIME_SCALE);
1415 dump_predicate (f, s->conds, &e->predicate);
1417 if (s->loop_iterations)
1419 fprintf (f, " loop iterations:");
1420 dump_predicate (f, s->conds, s->loop_iterations);
1422 if (s->loop_stride)
1424 fprintf (f, " loop stride:");
1425 dump_predicate (f, s->conds, s->loop_stride);
1427 if (s->array_index)
1429 fprintf (f, " array index:");
1430 dump_predicate (f, s->conds, s->array_index);
1432 fprintf (f, " calls:\n");
1433 dump_inline_edge_summary (f, 4, node, s);
1434 fprintf (f, "\n");
1438 DEBUG_FUNCTION void
1439 debug_inline_summary (struct cgraph_node *node)
1441 dump_inline_summary (stderr, node);
1444 void
1445 dump_inline_summaries (FILE *f)
1447 struct cgraph_node *node;
1449 FOR_EACH_DEFINED_FUNCTION (node)
1450 if (!node->global.inlined_to)
1451 dump_inline_summary (f, node);
1454 /* Give initial reasons why inlining would fail on EDGE. This gets either
1455 nullified or usually overwritten by more precise reasons later. */
1457 void
1458 initialize_inline_failed (struct cgraph_edge *e)
1460 struct cgraph_node *callee = e->callee;
1462 if (e->indirect_unknown_callee)
1463 e->inline_failed = CIF_INDIRECT_UNKNOWN_CALL;
1464 else if (!callee->definition)
1465 e->inline_failed = CIF_BODY_NOT_AVAILABLE;
1466 else if (callee->local.redefined_extern_inline)
1467 e->inline_failed = CIF_REDEFINED_EXTERN_INLINE;
1468 else if (e->call_stmt_cannot_inline_p)
1469 e->inline_failed = CIF_MISMATCHED_ARGUMENTS;
1470 else if (cfun && fn_contains_cilk_spawn_p (cfun))
1471 /* We can't inline if the function is spawing a function. */
1472 e->inline_failed = CIF_FUNCTION_NOT_INLINABLE;
1473 else
1474 e->inline_failed = CIF_FUNCTION_NOT_CONSIDERED;
1477 /* Callback of walk_aliased_vdefs. Flags that it has been invoked to the
1478 boolean variable pointed to by DATA. */
1480 static bool
1481 mark_modified (ao_ref *ao ATTRIBUTE_UNUSED, tree vdef ATTRIBUTE_UNUSED,
1482 void *data)
1484 bool *b = (bool *) data;
1485 *b = true;
1486 return true;
1489 /* If OP refers to value of function parameter, return the corresponding
1490 parameter. */
1492 static tree
1493 unmodified_parm_1 (gimple stmt, tree op)
1495 /* SSA_NAME referring to parm default def? */
1496 if (TREE_CODE (op) == SSA_NAME
1497 && SSA_NAME_IS_DEFAULT_DEF (op)
1498 && TREE_CODE (SSA_NAME_VAR (op)) == PARM_DECL)
1499 return SSA_NAME_VAR (op);
1500 /* Non-SSA parm reference? */
1501 if (TREE_CODE (op) == PARM_DECL)
1503 bool modified = false;
1505 ao_ref refd;
1506 ao_ref_init (&refd, op);
1507 walk_aliased_vdefs (&refd, gimple_vuse (stmt), mark_modified, &modified,
1508 NULL);
1509 if (!modified)
1510 return op;
1512 return NULL_TREE;
1515 /* If OP refers to value of function parameter, return the corresponding
1516 parameter. Also traverse chains of SSA register assignments. */
1518 static tree
1519 unmodified_parm (gimple stmt, tree op)
1521 tree res = unmodified_parm_1 (stmt, op);
1522 if (res)
1523 return res;
1525 if (TREE_CODE (op) == SSA_NAME
1526 && !SSA_NAME_IS_DEFAULT_DEF (op)
1527 && gimple_assign_single_p (SSA_NAME_DEF_STMT (op)))
1528 return unmodified_parm (SSA_NAME_DEF_STMT (op),
1529 gimple_assign_rhs1 (SSA_NAME_DEF_STMT (op)));
1530 return NULL_TREE;
1533 /* If OP refers to a value of a function parameter or value loaded from an
1534 aggregate passed to a parameter (either by value or reference), return TRUE
1535 and store the number of the parameter to *INDEX_P and information whether
1536 and how it has been loaded from an aggregate into *AGGPOS. INFO describes
1537 the function parameters, STMT is the statement in which OP is used or
1538 loaded. */
1540 static bool
1541 unmodified_parm_or_parm_agg_item (struct ipa_node_params *info,
1542 gimple stmt, tree op, int *index_p,
1543 struct agg_position_info *aggpos)
1545 tree res = unmodified_parm_1 (stmt, op);
1547 gcc_checking_assert (aggpos);
1548 if (res)
1550 *index_p = ipa_get_param_decl_index (info, res);
1551 if (*index_p < 0)
1552 return false;
1553 aggpos->agg_contents = false;
1554 aggpos->by_ref = false;
1555 return true;
1558 if (TREE_CODE (op) == SSA_NAME)
1560 if (SSA_NAME_IS_DEFAULT_DEF (op)
1561 || !gimple_assign_single_p (SSA_NAME_DEF_STMT (op)))
1562 return false;
1563 stmt = SSA_NAME_DEF_STMT (op);
1564 op = gimple_assign_rhs1 (stmt);
1565 if (!REFERENCE_CLASS_P (op))
1566 return unmodified_parm_or_parm_agg_item (info, stmt, op, index_p,
1567 aggpos);
1570 aggpos->agg_contents = true;
1571 return ipa_load_from_parm_agg (info, stmt, op, index_p, &aggpos->offset,
1572 &aggpos->by_ref);
1575 /* See if statement might disappear after inlining.
1576 0 - means not eliminated
1577 1 - half of statements goes away
1578 2 - for sure it is eliminated.
1579 We are not terribly sophisticated, basically looking for simple abstraction
1580 penalty wrappers. */
1582 static int
1583 eliminated_by_inlining_prob (gimple stmt)
1585 enum gimple_code code = gimple_code (stmt);
1586 enum tree_code rhs_code;
1588 if (!optimize)
1589 return 0;
1591 switch (code)
1593 case GIMPLE_RETURN:
1594 return 2;
1595 case GIMPLE_ASSIGN:
1596 if (gimple_num_ops (stmt) != 2)
1597 return 0;
1599 rhs_code = gimple_assign_rhs_code (stmt);
1601 /* Casts of parameters, loads from parameters passed by reference
1602 and stores to return value or parameters are often free after
1603 inlining dua to SRA and further combining.
1604 Assume that half of statements goes away. */
1605 if (rhs_code == CONVERT_EXPR
1606 || rhs_code == NOP_EXPR
1607 || rhs_code == VIEW_CONVERT_EXPR
1608 || rhs_code == ADDR_EXPR
1609 || gimple_assign_rhs_class (stmt) == GIMPLE_SINGLE_RHS)
1611 tree rhs = gimple_assign_rhs1 (stmt);
1612 tree lhs = gimple_assign_lhs (stmt);
1613 tree inner_rhs = get_base_address (rhs);
1614 tree inner_lhs = get_base_address (lhs);
1615 bool rhs_free = false;
1616 bool lhs_free = false;
1618 if (!inner_rhs)
1619 inner_rhs = rhs;
1620 if (!inner_lhs)
1621 inner_lhs = lhs;
1623 /* Reads of parameter are expected to be free. */
1624 if (unmodified_parm (stmt, inner_rhs))
1625 rhs_free = true;
1626 /* Match expressions of form &this->field. Those will most likely
1627 combine with something upstream after inlining. */
1628 else if (TREE_CODE (inner_rhs) == ADDR_EXPR)
1630 tree op = get_base_address (TREE_OPERAND (inner_rhs, 0));
1631 if (TREE_CODE (op) == PARM_DECL)
1632 rhs_free = true;
1633 else if (TREE_CODE (op) == MEM_REF
1634 && unmodified_parm (stmt, TREE_OPERAND (op, 0)))
1635 rhs_free = true;
1638 /* When parameter is not SSA register because its address is taken
1639 and it is just copied into one, the statement will be completely
1640 free after inlining (we will copy propagate backward). */
1641 if (rhs_free && is_gimple_reg (lhs))
1642 return 2;
1644 /* Reads of parameters passed by reference
1645 expected to be free (i.e. optimized out after inlining). */
1646 if (TREE_CODE (inner_rhs) == MEM_REF
1647 && unmodified_parm (stmt, TREE_OPERAND (inner_rhs, 0)))
1648 rhs_free = true;
1650 /* Copying parameter passed by reference into gimple register is
1651 probably also going to copy propagate, but we can't be quite
1652 sure. */
1653 if (rhs_free && is_gimple_reg (lhs))
1654 lhs_free = true;
1656 /* Writes to parameters, parameters passed by value and return value
1657 (either dirrectly or passed via invisible reference) are free.
1659 TODO: We ought to handle testcase like
1660 struct a {int a,b;};
1661 struct a
1662 retrurnsturct (void)
1664 struct a a ={1,2};
1665 return a;
1668 This translate into:
1670 retrurnsturct ()
1672 int a$b;
1673 int a$a;
1674 struct a a;
1675 struct a D.2739;
1677 <bb 2>:
1678 D.2739.a = 1;
1679 D.2739.b = 2;
1680 return D.2739;
1683 For that we either need to copy ipa-split logic detecting writes
1684 to return value. */
1685 if (TREE_CODE (inner_lhs) == PARM_DECL
1686 || TREE_CODE (inner_lhs) == RESULT_DECL
1687 || (TREE_CODE (inner_lhs) == MEM_REF
1688 && (unmodified_parm (stmt, TREE_OPERAND (inner_lhs, 0))
1689 || (TREE_CODE (TREE_OPERAND (inner_lhs, 0)) == SSA_NAME
1690 && SSA_NAME_VAR (TREE_OPERAND (inner_lhs, 0))
1691 && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND
1692 (inner_lhs,
1693 0))) == RESULT_DECL))))
1694 lhs_free = true;
1695 if (lhs_free
1696 && (is_gimple_reg (rhs) || is_gimple_min_invariant (rhs)))
1697 rhs_free = true;
1698 if (lhs_free && rhs_free)
1699 return 1;
1701 return 0;
1702 default:
1703 return 0;
1708 /* If BB ends by a conditional we can turn into predicates, attach corresponding
1709 predicates to the CFG edges. */
1711 static void
1712 set_cond_stmt_execution_predicate (struct ipa_node_params *info,
1713 struct inline_summary *summary,
1714 basic_block bb)
1716 gimple last;
1717 tree op;
1718 int index;
1719 struct agg_position_info aggpos;
1720 enum tree_code code, inverted_code;
1721 edge e;
1722 edge_iterator ei;
1723 gimple set_stmt;
1724 tree op2;
1726 last = last_stmt (bb);
1727 if (!last || gimple_code (last) != GIMPLE_COND)
1728 return;
1729 if (!is_gimple_ip_invariant (gimple_cond_rhs (last)))
1730 return;
1731 op = gimple_cond_lhs (last);
1732 /* TODO: handle conditionals like
1733 var = op0 < 4;
1734 if (var != 0). */
1735 if (unmodified_parm_or_parm_agg_item (info, last, op, &index, &aggpos))
1737 code = gimple_cond_code (last);
1738 inverted_code
1739 = invert_tree_comparison (code,
1740 HONOR_NANS (TYPE_MODE (TREE_TYPE (op))));
1742 FOR_EACH_EDGE (e, ei, bb->succs)
1744 enum tree_code this_code = (e->flags & EDGE_TRUE_VALUE
1745 ? code : inverted_code);
1746 /* invert_tree_comparison will return ERROR_MARK on FP
1747 comparsions that are not EQ/NE instead of returning proper
1748 unordered one. Be sure it is not confused with NON_CONSTANT. */
1749 if (this_code != ERROR_MARK)
1751 struct predicate p = add_condition (summary, index, &aggpos,
1752 this_code,
1753 gimple_cond_rhs (last));
1754 e->aux = pool_alloc (edge_predicate_pool);
1755 *(struct predicate *) e->aux = p;
1760 if (TREE_CODE (op) != SSA_NAME)
1761 return;
1762 /* Special case
1763 if (builtin_constant_p (op))
1764 constant_code
1765 else
1766 nonconstant_code.
1767 Here we can predicate nonconstant_code. We can't
1768 really handle constant_code since we have no predicate
1769 for this and also the constant code is not known to be
1770 optimized away when inliner doen't see operand is constant.
1771 Other optimizers might think otherwise. */
1772 if (gimple_cond_code (last) != NE_EXPR
1773 || !integer_zerop (gimple_cond_rhs (last)))
1774 return;
1775 set_stmt = SSA_NAME_DEF_STMT (op);
1776 if (!gimple_call_builtin_p (set_stmt, BUILT_IN_CONSTANT_P)
1777 || gimple_call_num_args (set_stmt) != 1)
1778 return;
1779 op2 = gimple_call_arg (set_stmt, 0);
1780 if (!unmodified_parm_or_parm_agg_item
1781 (info, set_stmt, op2, &index, &aggpos))
1782 return;
1783 FOR_EACH_EDGE (e, ei, bb->succs) if (e->flags & EDGE_FALSE_VALUE)
1785 struct predicate p = add_condition (summary, index, &aggpos,
1786 IS_NOT_CONSTANT, NULL_TREE);
1787 e->aux = pool_alloc (edge_predicate_pool);
1788 *(struct predicate *) e->aux = p;
1793 /* If BB ends by a switch we can turn into predicates, attach corresponding
1794 predicates to the CFG edges. */
1796 static void
1797 set_switch_stmt_execution_predicate (struct ipa_node_params *info,
1798 struct inline_summary *summary,
1799 basic_block bb)
1801 gimple last;
1802 tree op;
1803 int index;
1804 struct agg_position_info aggpos;
1805 edge e;
1806 edge_iterator ei;
1807 size_t n;
1808 size_t case_idx;
1810 last = last_stmt (bb);
1811 if (!last || gimple_code (last) != GIMPLE_SWITCH)
1812 return;
1813 op = gimple_switch_index (last);
1814 if (!unmodified_parm_or_parm_agg_item (info, last, op, &index, &aggpos))
1815 return;
1817 FOR_EACH_EDGE (e, ei, bb->succs)
1819 e->aux = pool_alloc (edge_predicate_pool);
1820 *(struct predicate *) e->aux = false_predicate ();
1822 n = gimple_switch_num_labels (last);
1823 for (case_idx = 0; case_idx < n; ++case_idx)
1825 tree cl = gimple_switch_label (last, case_idx);
1826 tree min, max;
1827 struct predicate p;
1829 e = find_edge (bb, label_to_block (CASE_LABEL (cl)));
1830 min = CASE_LOW (cl);
1831 max = CASE_HIGH (cl);
1833 /* For default we might want to construct predicate that none
1834 of cases is met, but it is bit hard to do not having negations
1835 of conditionals handy. */
1836 if (!min && !max)
1837 p = true_predicate ();
1838 else if (!max)
1839 p = add_condition (summary, index, &aggpos, EQ_EXPR, min);
1840 else
1842 struct predicate p1, p2;
1843 p1 = add_condition (summary, index, &aggpos, GE_EXPR, min);
1844 p2 = add_condition (summary, index, &aggpos, LE_EXPR, max);
1845 p = and_predicates (summary->conds, &p1, &p2);
1847 *(struct predicate *) e->aux
1848 = or_predicates (summary->conds, &p, (struct predicate *) e->aux);
1853 /* For each BB in NODE attach to its AUX pointer predicate under
1854 which it is executable. */
1856 static void
1857 compute_bb_predicates (struct cgraph_node *node,
1858 struct ipa_node_params *parms_info,
1859 struct inline_summary *summary)
1861 struct function *my_function = DECL_STRUCT_FUNCTION (node->decl);
1862 bool done = false;
1863 basic_block bb;
1865 FOR_EACH_BB_FN (bb, my_function)
1867 set_cond_stmt_execution_predicate (parms_info, summary, bb);
1868 set_switch_stmt_execution_predicate (parms_info, summary, bb);
1871 /* Entry block is always executable. */
1872 ENTRY_BLOCK_PTR_FOR_FN (my_function)->aux
1873 = pool_alloc (edge_predicate_pool);
1874 *(struct predicate *) ENTRY_BLOCK_PTR_FOR_FN (my_function)->aux
1875 = true_predicate ();
1877 /* A simple dataflow propagation of predicates forward in the CFG.
1878 TODO: work in reverse postorder. */
1879 while (!done)
1881 done = true;
1882 FOR_EACH_BB_FN (bb, my_function)
1884 struct predicate p = false_predicate ();
1885 edge e;
1886 edge_iterator ei;
1887 FOR_EACH_EDGE (e, ei, bb->preds)
1889 if (e->src->aux)
1891 struct predicate this_bb_predicate
1892 = *(struct predicate *) e->src->aux;
1893 if (e->aux)
1894 this_bb_predicate
1895 = and_predicates (summary->conds, &this_bb_predicate,
1896 (struct predicate *) e->aux);
1897 p = or_predicates (summary->conds, &p, &this_bb_predicate);
1898 if (true_predicate_p (&p))
1899 break;
1902 if (false_predicate_p (&p))
1903 gcc_assert (!bb->aux);
1904 else
1906 if (!bb->aux)
1908 done = false;
1909 bb->aux = pool_alloc (edge_predicate_pool);
1910 *((struct predicate *) bb->aux) = p;
1912 else if (!predicates_equal_p (&p, (struct predicate *) bb->aux))
1914 /* This OR operation is needed to ensure monotonous data flow
1915 in the case we hit the limit on number of clauses and the
1916 and/or operations above give approximate answers. */
1917 p = or_predicates (summary->conds, &p, (struct predicate *)bb->aux);
1918 if (!predicates_equal_p (&p, (struct predicate *) bb->aux))
1920 done = false;
1921 *((struct predicate *) bb->aux) = p;
1930 /* We keep info about constantness of SSA names. */
1932 typedef struct predicate predicate_t;
1933 /* Return predicate specifying when the STMT might have result that is not
1934 a compile time constant. */
1936 static struct predicate
1937 will_be_nonconstant_expr_predicate (struct ipa_node_params *info,
1938 struct inline_summary *summary,
1939 tree expr,
1940 vec<predicate_t> nonconstant_names)
1942 tree parm;
1943 int index;
1945 while (UNARY_CLASS_P (expr))
1946 expr = TREE_OPERAND (expr, 0);
1948 parm = unmodified_parm (NULL, expr);
1949 if (parm && (index = ipa_get_param_decl_index (info, parm)) >= 0)
1950 return add_condition (summary, index, NULL, CHANGED, NULL_TREE);
1951 if (is_gimple_min_invariant (expr))
1952 return false_predicate ();
1953 if (TREE_CODE (expr) == SSA_NAME)
1954 return nonconstant_names[SSA_NAME_VERSION (expr)];
1955 if (BINARY_CLASS_P (expr) || COMPARISON_CLASS_P (expr))
1957 struct predicate p1 = will_be_nonconstant_expr_predicate
1958 (info, summary, TREE_OPERAND (expr, 0),
1959 nonconstant_names);
1960 struct predicate p2;
1961 if (true_predicate_p (&p1))
1962 return p1;
1963 p2 = will_be_nonconstant_expr_predicate (info, summary,
1964 TREE_OPERAND (expr, 1),
1965 nonconstant_names);
1966 return or_predicates (summary->conds, &p1, &p2);
1968 else if (TREE_CODE (expr) == COND_EXPR)
1970 struct predicate p1 = will_be_nonconstant_expr_predicate
1971 (info, summary, TREE_OPERAND (expr, 0),
1972 nonconstant_names);
1973 struct predicate p2;
1974 if (true_predicate_p (&p1))
1975 return p1;
1976 p2 = will_be_nonconstant_expr_predicate (info, summary,
1977 TREE_OPERAND (expr, 1),
1978 nonconstant_names);
1979 if (true_predicate_p (&p2))
1980 return p2;
1981 p1 = or_predicates (summary->conds, &p1, &p2);
1982 p2 = will_be_nonconstant_expr_predicate (info, summary,
1983 TREE_OPERAND (expr, 2),
1984 nonconstant_names);
1985 return or_predicates (summary->conds, &p1, &p2);
1987 else
1989 debug_tree (expr);
1990 gcc_unreachable ();
1992 return false_predicate ();
1996 /* Return predicate specifying when the STMT might have result that is not
1997 a compile time constant. */
1999 static struct predicate
2000 will_be_nonconstant_predicate (struct ipa_node_params *info,
2001 struct inline_summary *summary,
2002 gimple stmt,
2003 vec<predicate_t> nonconstant_names)
2005 struct predicate p = true_predicate ();
2006 ssa_op_iter iter;
2007 tree use;
2008 struct predicate op_non_const;
2009 bool is_load;
2010 int base_index;
2011 struct agg_position_info aggpos;
2013 /* What statments might be optimized away
2014 when their arguments are constant
2015 TODO: also trivial builtins.
2016 builtin_constant_p is already handled later. */
2017 if (gimple_code (stmt) != GIMPLE_ASSIGN
2018 && gimple_code (stmt) != GIMPLE_COND
2019 && gimple_code (stmt) != GIMPLE_SWITCH)
2020 return p;
2022 /* Stores will stay anyway. */
2023 if (gimple_store_p (stmt))
2024 return p;
2026 is_load = gimple_assign_load_p (stmt);
2028 /* Loads can be optimized when the value is known. */
2029 if (is_load)
2031 tree op;
2032 gcc_assert (gimple_assign_single_p (stmt));
2033 op = gimple_assign_rhs1 (stmt);
2034 if (!unmodified_parm_or_parm_agg_item (info, stmt, op, &base_index,
2035 &aggpos))
2036 return p;
2038 else
2039 base_index = -1;
2041 /* See if we understand all operands before we start
2042 adding conditionals. */
2043 FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE)
2045 tree parm = unmodified_parm (stmt, use);
2046 /* For arguments we can build a condition. */
2047 if (parm && ipa_get_param_decl_index (info, parm) >= 0)
2048 continue;
2049 if (TREE_CODE (use) != SSA_NAME)
2050 return p;
2051 /* If we know when operand is constant,
2052 we still can say something useful. */
2053 if (!true_predicate_p (&nonconstant_names[SSA_NAME_VERSION (use)]))
2054 continue;
2055 return p;
2058 if (is_load)
2059 op_non_const =
2060 add_condition (summary, base_index, &aggpos, CHANGED, NULL);
2061 else
2062 op_non_const = false_predicate ();
2063 FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE)
2065 tree parm = unmodified_parm (stmt, use);
2066 int index;
2068 if (parm && (index = ipa_get_param_decl_index (info, parm)) >= 0)
2070 if (index != base_index)
2071 p = add_condition (summary, index, NULL, CHANGED, NULL_TREE);
2072 else
2073 continue;
2075 else
2076 p = nonconstant_names[SSA_NAME_VERSION (use)];
2077 op_non_const = or_predicates (summary->conds, &p, &op_non_const);
2079 if (gimple_code (stmt) == GIMPLE_ASSIGN
2080 && TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME)
2081 nonconstant_names[SSA_NAME_VERSION (gimple_assign_lhs (stmt))]
2082 = op_non_const;
2083 return op_non_const;
2086 struct record_modified_bb_info
2088 bitmap bb_set;
2089 gimple stmt;
2092 /* Callback of walk_aliased_vdefs. Records basic blocks where the value may be
2093 set except for info->stmt. */
2095 static bool
2096 record_modified (ao_ref *ao ATTRIBUTE_UNUSED, tree vdef, void *data)
2098 struct record_modified_bb_info *info =
2099 (struct record_modified_bb_info *) data;
2100 if (SSA_NAME_DEF_STMT (vdef) == info->stmt)
2101 return false;
2102 bitmap_set_bit (info->bb_set,
2103 SSA_NAME_IS_DEFAULT_DEF (vdef)
2104 ? ENTRY_BLOCK_PTR_FOR_FN (cfun)->index
2105 : gimple_bb (SSA_NAME_DEF_STMT (vdef))->index);
2106 return false;
2109 /* Return probability (based on REG_BR_PROB_BASE) that I-th parameter of STMT
2110 will change since last invocation of STMT.
2112 Value 0 is reserved for compile time invariants.
2113 For common parameters it is REG_BR_PROB_BASE. For loop invariants it
2114 ought to be REG_BR_PROB_BASE / estimated_iters. */
2116 static int
2117 param_change_prob (gimple stmt, int i)
2119 tree op = gimple_call_arg (stmt, i);
2120 basic_block bb = gimple_bb (stmt);
2121 tree base;
2123 /* Global invariants neve change. */
2124 if (is_gimple_min_invariant (op))
2125 return 0;
2126 /* We would have to do non-trivial analysis to really work out what
2127 is the probability of value to change (i.e. when init statement
2128 is in a sibling loop of the call).
2130 We do an conservative estimate: when call is executed N times more often
2131 than the statement defining value, we take the frequency 1/N. */
2132 if (TREE_CODE (op) == SSA_NAME)
2134 int init_freq;
2136 if (!bb->frequency)
2137 return REG_BR_PROB_BASE;
2139 if (SSA_NAME_IS_DEFAULT_DEF (op))
2140 init_freq = ENTRY_BLOCK_PTR_FOR_FN (cfun)->frequency;
2141 else
2142 init_freq = gimple_bb (SSA_NAME_DEF_STMT (op))->frequency;
2144 if (!init_freq)
2145 init_freq = 1;
2146 if (init_freq < bb->frequency)
2147 return MAX (GCOV_COMPUTE_SCALE (init_freq, bb->frequency), 1);
2148 else
2149 return REG_BR_PROB_BASE;
2152 base = get_base_address (op);
2153 if (base)
2155 ao_ref refd;
2156 int max;
2157 struct record_modified_bb_info info;
2158 bitmap_iterator bi;
2159 unsigned index;
2160 tree init = ctor_for_folding (base);
2162 if (init != error_mark_node)
2163 return 0;
2164 if (!bb->frequency)
2165 return REG_BR_PROB_BASE;
2166 ao_ref_init (&refd, op);
2167 info.stmt = stmt;
2168 info.bb_set = BITMAP_ALLOC (NULL);
2169 walk_aliased_vdefs (&refd, gimple_vuse (stmt), record_modified, &info,
2170 NULL);
2171 if (bitmap_bit_p (info.bb_set, bb->index))
2173 BITMAP_FREE (info.bb_set);
2174 return REG_BR_PROB_BASE;
2177 /* Assume that every memory is initialized at entry.
2178 TODO: Can we easilly determine if value is always defined
2179 and thus we may skip entry block? */
2180 if (ENTRY_BLOCK_PTR_FOR_FN (cfun)->frequency)
2181 max = ENTRY_BLOCK_PTR_FOR_FN (cfun)->frequency;
2182 else
2183 max = 1;
2185 EXECUTE_IF_SET_IN_BITMAP (info.bb_set, 0, index, bi)
2186 max = MIN (max, BASIC_BLOCK_FOR_FN (cfun, index)->frequency);
2188 BITMAP_FREE (info.bb_set);
2189 if (max < bb->frequency)
2190 return MAX (GCOV_COMPUTE_SCALE (max, bb->frequency), 1);
2191 else
2192 return REG_BR_PROB_BASE;
2194 return REG_BR_PROB_BASE;
2197 /* Find whether a basic block BB is the final block of a (half) diamond CFG
2198 sub-graph and if the predicate the condition depends on is known. If so,
2199 return true and store the pointer the predicate in *P. */
2201 static bool
2202 phi_result_unknown_predicate (struct ipa_node_params *info,
2203 struct inline_summary *summary, basic_block bb,
2204 struct predicate *p,
2205 vec<predicate_t> nonconstant_names)
2207 edge e;
2208 edge_iterator ei;
2209 basic_block first_bb = NULL;
2210 gimple stmt;
2212 if (single_pred_p (bb))
2214 *p = false_predicate ();
2215 return true;
2218 FOR_EACH_EDGE (e, ei, bb->preds)
2220 if (single_succ_p (e->src))
2222 if (!single_pred_p (e->src))
2223 return false;
2224 if (!first_bb)
2225 first_bb = single_pred (e->src);
2226 else if (single_pred (e->src) != first_bb)
2227 return false;
2229 else
2231 if (!first_bb)
2232 first_bb = e->src;
2233 else if (e->src != first_bb)
2234 return false;
2238 if (!first_bb)
2239 return false;
2241 stmt = last_stmt (first_bb);
2242 if (!stmt
2243 || gimple_code (stmt) != GIMPLE_COND
2244 || !is_gimple_ip_invariant (gimple_cond_rhs (stmt)))
2245 return false;
2247 *p = will_be_nonconstant_expr_predicate (info, summary,
2248 gimple_cond_lhs (stmt),
2249 nonconstant_names);
2250 if (true_predicate_p (p))
2251 return false;
2252 else
2253 return true;
2256 /* Given a PHI statement in a function described by inline properties SUMMARY
2257 and *P being the predicate describing whether the selected PHI argument is
2258 known, store a predicate for the result of the PHI statement into
2259 NONCONSTANT_NAMES, if possible. */
2261 static void
2262 predicate_for_phi_result (struct inline_summary *summary, gimple phi,
2263 struct predicate *p,
2264 vec<predicate_t> nonconstant_names)
2266 unsigned i;
2268 for (i = 0; i < gimple_phi_num_args (phi); i++)
2270 tree arg = gimple_phi_arg (phi, i)->def;
2271 if (!is_gimple_min_invariant (arg))
2273 gcc_assert (TREE_CODE (arg) == SSA_NAME);
2274 *p = or_predicates (summary->conds, p,
2275 &nonconstant_names[SSA_NAME_VERSION (arg)]);
2276 if (true_predicate_p (p))
2277 return;
2281 if (dump_file && (dump_flags & TDF_DETAILS))
2283 fprintf (dump_file, "\t\tphi predicate: ");
2284 dump_predicate (dump_file, summary->conds, p);
2286 nonconstant_names[SSA_NAME_VERSION (gimple_phi_result (phi))] = *p;
2289 /* Return predicate specifying when array index in access OP becomes non-constant. */
2291 static struct predicate
2292 array_index_predicate (struct inline_summary *info,
2293 vec< predicate_t> nonconstant_names, tree op)
2295 struct predicate p = false_predicate ();
2296 while (handled_component_p (op))
2298 if (TREE_CODE (op) == ARRAY_REF || TREE_CODE (op) == ARRAY_RANGE_REF)
2300 if (TREE_CODE (TREE_OPERAND (op, 1)) == SSA_NAME)
2301 p = or_predicates (info->conds, &p,
2302 &nonconstant_names[SSA_NAME_VERSION
2303 (TREE_OPERAND (op, 1))]);
2305 op = TREE_OPERAND (op, 0);
2307 return p;
2310 /* For a typical usage of __builtin_expect (a<b, 1), we
2311 may introduce an extra relation stmt:
2312 With the builtin, we have
2313 t1 = a <= b;
2314 t2 = (long int) t1;
2315 t3 = __builtin_expect (t2, 1);
2316 if (t3 != 0)
2317 goto ...
2318 Without the builtin, we have
2319 if (a<=b)
2320 goto...
2321 This affects the size/time estimation and may have
2322 an impact on the earlier inlining.
2323 Here find this pattern and fix it up later. */
2325 static gimple
2326 find_foldable_builtin_expect (basic_block bb)
2328 gimple_stmt_iterator bsi;
2330 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
2332 gimple stmt = gsi_stmt (bsi);
2333 if (gimple_call_builtin_p (stmt, BUILT_IN_EXPECT)
2334 || (is_gimple_call (stmt)
2335 && gimple_call_internal_p (stmt)
2336 && gimple_call_internal_fn (stmt) == IFN_BUILTIN_EXPECT))
2338 tree var = gimple_call_lhs (stmt);
2339 tree arg = gimple_call_arg (stmt, 0);
2340 use_operand_p use_p;
2341 gimple use_stmt;
2342 bool match = false;
2343 bool done = false;
2345 if (!var || !arg)
2346 continue;
2347 gcc_assert (TREE_CODE (var) == SSA_NAME);
2349 while (TREE_CODE (arg) == SSA_NAME)
2351 gimple stmt_tmp = SSA_NAME_DEF_STMT (arg);
2352 if (!is_gimple_assign (stmt_tmp))
2353 break;
2354 switch (gimple_assign_rhs_code (stmt_tmp))
2356 case LT_EXPR:
2357 case LE_EXPR:
2358 case GT_EXPR:
2359 case GE_EXPR:
2360 case EQ_EXPR:
2361 case NE_EXPR:
2362 match = true;
2363 done = true;
2364 break;
2365 case NOP_EXPR:
2366 break;
2367 default:
2368 done = true;
2369 break;
2371 if (done)
2372 break;
2373 arg = gimple_assign_rhs1 (stmt_tmp);
2376 if (match && single_imm_use (var, &use_p, &use_stmt)
2377 && gimple_code (use_stmt) == GIMPLE_COND)
2378 return use_stmt;
2381 return NULL;
2384 /* Return true when the basic blocks contains only clobbers followed by RESX.
2385 Such BBs are kept around to make removal of dead stores possible with
2386 presence of EH and will be optimized out by optimize_clobbers later in the
2387 game.
2389 NEED_EH is used to recurse in case the clobber has non-EH predecestors
2390 that can be clobber only, too.. When it is false, the RESX is not necessary
2391 on the end of basic block. */
2393 static bool
2394 clobber_only_eh_bb_p (basic_block bb, bool need_eh = true)
2396 gimple_stmt_iterator gsi = gsi_last_bb (bb);
2397 edge_iterator ei;
2398 edge e;
2400 if (need_eh)
2402 if (gsi_end_p (gsi))
2403 return false;
2404 if (gimple_code (gsi_stmt (gsi)) != GIMPLE_RESX)
2405 return false;
2406 gsi_prev (&gsi);
2408 else if (!single_succ_p (bb))
2409 return false;
2411 for (; !gsi_end_p (gsi); gsi_prev (&gsi))
2413 gimple stmt = gsi_stmt (gsi);
2414 if (is_gimple_debug (stmt))
2415 continue;
2416 if (gimple_clobber_p (stmt))
2417 continue;
2418 if (gimple_code (stmt) == GIMPLE_LABEL)
2419 break;
2420 return false;
2423 /* See if all predecestors are either throws or clobber only BBs. */
2424 FOR_EACH_EDGE (e, ei, bb->preds)
2425 if (!(e->flags & EDGE_EH)
2426 && !clobber_only_eh_bb_p (e->src, false))
2427 return false;
2429 return true;
2432 /* Compute function body size parameters for NODE.
2433 When EARLY is true, we compute only simple summaries without
2434 non-trivial predicates to drive the early inliner. */
2436 static void
2437 estimate_function_body_sizes (struct cgraph_node *node, bool early)
2439 gcov_type time = 0;
2440 /* Estimate static overhead for function prologue/epilogue and alignment. */
2441 int size = 2;
2442 /* Benefits are scaled by probability of elimination that is in range
2443 <0,2>. */
2444 basic_block bb;
2445 gimple_stmt_iterator bsi;
2446 struct function *my_function = DECL_STRUCT_FUNCTION (node->decl);
2447 int freq;
2448 struct inline_summary *info = inline_summary (node);
2449 struct predicate bb_predicate;
2450 struct ipa_node_params *parms_info = NULL;
2451 vec<predicate_t> nonconstant_names = vNULL;
2452 int nblocks, n;
2453 int *order;
2454 predicate array_index = true_predicate ();
2455 gimple fix_builtin_expect_stmt;
2457 info->conds = NULL;
2458 info->entry = NULL;
2460 if (optimize && !early)
2462 calculate_dominance_info (CDI_DOMINATORS);
2463 loop_optimizer_init (LOOPS_NORMAL | LOOPS_HAVE_RECORDED_EXITS);
2465 if (ipa_node_params_vector.exists ())
2467 parms_info = IPA_NODE_REF (node);
2468 nonconstant_names.safe_grow_cleared
2469 (SSANAMES (my_function)->length ());
2473 if (dump_file)
2474 fprintf (dump_file, "\nAnalyzing function body size: %s\n",
2475 node->name ());
2477 /* When we run into maximal number of entries, we assign everything to the
2478 constant truth case. Be sure to have it in list. */
2479 bb_predicate = true_predicate ();
2480 account_size_time (info, 0, 0, &bb_predicate);
2482 bb_predicate = not_inlined_predicate ();
2483 account_size_time (info, 2 * INLINE_SIZE_SCALE, 0, &bb_predicate);
2485 gcc_assert (my_function && my_function->cfg);
2486 if (parms_info)
2487 compute_bb_predicates (node, parms_info, info);
2488 gcc_assert (cfun == my_function);
2489 order = XNEWVEC (int, n_basic_blocks_for_fn (cfun));
2490 nblocks = pre_and_rev_post_order_compute (NULL, order, false);
2491 for (n = 0; n < nblocks; n++)
2493 bb = BASIC_BLOCK_FOR_FN (cfun, order[n]);
2494 freq = compute_call_stmt_bb_frequency (node->decl, bb);
2495 if (clobber_only_eh_bb_p (bb))
2497 if (dump_file && (dump_flags & TDF_DETAILS))
2498 fprintf (dump_file, "\n Ignoring BB %i;"
2499 " it will be optimized away by cleanup_clobbers\n",
2500 bb->index);
2501 continue;
2504 /* TODO: Obviously predicates can be propagated down across CFG. */
2505 if (parms_info)
2507 if (bb->aux)
2508 bb_predicate = *(struct predicate *) bb->aux;
2509 else
2510 bb_predicate = false_predicate ();
2512 else
2513 bb_predicate = true_predicate ();
2515 if (dump_file && (dump_flags & TDF_DETAILS))
2517 fprintf (dump_file, "\n BB %i predicate:", bb->index);
2518 dump_predicate (dump_file, info->conds, &bb_predicate);
2521 if (parms_info && nonconstant_names.exists ())
2523 struct predicate phi_predicate;
2524 bool first_phi = true;
2526 for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi))
2528 if (first_phi
2529 && !phi_result_unknown_predicate (parms_info, info, bb,
2530 &phi_predicate,
2531 nonconstant_names))
2532 break;
2533 first_phi = false;
2534 if (dump_file && (dump_flags & TDF_DETAILS))
2536 fprintf (dump_file, " ");
2537 print_gimple_stmt (dump_file, gsi_stmt (bsi), 0, 0);
2539 predicate_for_phi_result (info, gsi_stmt (bsi), &phi_predicate,
2540 nonconstant_names);
2544 fix_builtin_expect_stmt = find_foldable_builtin_expect (bb);
2546 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
2548 gimple stmt = gsi_stmt (bsi);
2549 int this_size = estimate_num_insns (stmt, &eni_size_weights);
2550 int this_time = estimate_num_insns (stmt, &eni_time_weights);
2551 int prob;
2552 struct predicate will_be_nonconstant;
2554 /* This relation stmt should be folded after we remove
2555 buildin_expect call. Adjust the cost here. */
2556 if (stmt == fix_builtin_expect_stmt)
2558 this_size--;
2559 this_time--;
2562 if (dump_file && (dump_flags & TDF_DETAILS))
2564 fprintf (dump_file, " ");
2565 print_gimple_stmt (dump_file, stmt, 0, 0);
2566 fprintf (dump_file, "\t\tfreq:%3.2f size:%3i time:%3i\n",
2567 ((double) freq) / CGRAPH_FREQ_BASE, this_size,
2568 this_time);
2571 if (gimple_assign_load_p (stmt) && nonconstant_names.exists ())
2573 struct predicate this_array_index;
2574 this_array_index =
2575 array_index_predicate (info, nonconstant_names,
2576 gimple_assign_rhs1 (stmt));
2577 if (!false_predicate_p (&this_array_index))
2578 array_index =
2579 and_predicates (info->conds, &array_index,
2580 &this_array_index);
2582 if (gimple_store_p (stmt) && nonconstant_names.exists ())
2584 struct predicate this_array_index;
2585 this_array_index =
2586 array_index_predicate (info, nonconstant_names,
2587 gimple_get_lhs (stmt));
2588 if (!false_predicate_p (&this_array_index))
2589 array_index =
2590 and_predicates (info->conds, &array_index,
2591 &this_array_index);
2595 if (is_gimple_call (stmt)
2596 && !gimple_call_internal_p (stmt))
2598 struct cgraph_edge *edge = node->get_edge (stmt);
2599 struct inline_edge_summary *es = inline_edge_summary (edge);
2601 /* Special case: results of BUILT_IN_CONSTANT_P will be always
2602 resolved as constant. We however don't want to optimize
2603 out the cgraph edges. */
2604 if (nonconstant_names.exists ()
2605 && gimple_call_builtin_p (stmt, BUILT_IN_CONSTANT_P)
2606 && gimple_call_lhs (stmt)
2607 && TREE_CODE (gimple_call_lhs (stmt)) == SSA_NAME)
2609 struct predicate false_p = false_predicate ();
2610 nonconstant_names[SSA_NAME_VERSION (gimple_call_lhs (stmt))]
2611 = false_p;
2613 if (ipa_node_params_vector.exists ())
2615 int count = gimple_call_num_args (stmt);
2616 int i;
2618 if (count)
2619 es->param.safe_grow_cleared (count);
2620 for (i = 0; i < count; i++)
2622 int prob = param_change_prob (stmt, i);
2623 gcc_assert (prob >= 0 && prob <= REG_BR_PROB_BASE);
2624 es->param[i].change_prob = prob;
2628 es->call_stmt_size = this_size;
2629 es->call_stmt_time = this_time;
2630 es->loop_depth = bb_loop_depth (bb);
2631 edge_set_predicate (edge, &bb_predicate);
2634 /* TODO: When conditional jump or swithc is known to be constant, but
2635 we did not translate it into the predicates, we really can account
2636 just maximum of the possible paths. */
2637 if (parms_info)
2638 will_be_nonconstant
2639 = will_be_nonconstant_predicate (parms_info, info,
2640 stmt, nonconstant_names);
2641 if (this_time || this_size)
2643 struct predicate p;
2645 this_time *= freq;
2647 prob = eliminated_by_inlining_prob (stmt);
2648 if (prob == 1 && dump_file && (dump_flags & TDF_DETAILS))
2649 fprintf (dump_file,
2650 "\t\t50%% will be eliminated by inlining\n");
2651 if (prob == 2 && dump_file && (dump_flags & TDF_DETAILS))
2652 fprintf (dump_file, "\t\tWill be eliminated by inlining\n");
2654 if (parms_info)
2655 p = and_predicates (info->conds, &bb_predicate,
2656 &will_be_nonconstant);
2657 else
2658 p = true_predicate ();
2660 if (!false_predicate_p (&p))
2662 time += this_time;
2663 size += this_size;
2664 if (time > MAX_TIME * INLINE_TIME_SCALE)
2665 time = MAX_TIME * INLINE_TIME_SCALE;
2668 /* We account everything but the calls. Calls have their own
2669 size/time info attached to cgraph edges. This is necessary
2670 in order to make the cost disappear after inlining. */
2671 if (!is_gimple_call (stmt))
2673 if (prob)
2675 struct predicate ip = not_inlined_predicate ();
2676 ip = and_predicates (info->conds, &ip, &p);
2677 account_size_time (info, this_size * prob,
2678 this_time * prob, &ip);
2680 if (prob != 2)
2681 account_size_time (info, this_size * (2 - prob),
2682 this_time * (2 - prob), &p);
2685 gcc_assert (time >= 0);
2686 gcc_assert (size >= 0);
2690 set_hint_predicate (&inline_summary (node)->array_index, array_index);
2691 time = (time + CGRAPH_FREQ_BASE / 2) / CGRAPH_FREQ_BASE;
2692 if (time > MAX_TIME)
2693 time = MAX_TIME;
2694 free (order);
2696 if (!early && nonconstant_names.exists ())
2698 struct loop *loop;
2699 predicate loop_iterations = true_predicate ();
2700 predicate loop_stride = true_predicate ();
2702 if (dump_file && (dump_flags & TDF_DETAILS))
2703 flow_loops_dump (dump_file, NULL, 0);
2704 scev_initialize ();
2705 FOR_EACH_LOOP (loop, 0)
2707 vec<edge> exits;
2708 edge ex;
2709 unsigned int j, i;
2710 struct tree_niter_desc niter_desc;
2711 basic_block *body = get_loop_body (loop);
2712 bb_predicate = *(struct predicate *) loop->header->aux;
2714 exits = get_loop_exit_edges (loop);
2715 FOR_EACH_VEC_ELT (exits, j, ex)
2716 if (number_of_iterations_exit (loop, ex, &niter_desc, false)
2717 && !is_gimple_min_invariant (niter_desc.niter))
2719 predicate will_be_nonconstant
2720 = will_be_nonconstant_expr_predicate (parms_info, info,
2721 niter_desc.niter,
2722 nonconstant_names);
2723 if (!true_predicate_p (&will_be_nonconstant))
2724 will_be_nonconstant = and_predicates (info->conds,
2725 &bb_predicate,
2726 &will_be_nonconstant);
2727 if (!true_predicate_p (&will_be_nonconstant)
2728 && !false_predicate_p (&will_be_nonconstant))
2729 /* This is slightly inprecise. We may want to represent each
2730 loop with independent predicate. */
2731 loop_iterations =
2732 and_predicates (info->conds, &loop_iterations,
2733 &will_be_nonconstant);
2735 exits.release ();
2737 for (i = 0; i < loop->num_nodes; i++)
2739 gimple_stmt_iterator gsi;
2740 bb_predicate = *(struct predicate *) body[i]->aux;
2741 for (gsi = gsi_start_bb (body[i]); !gsi_end_p (gsi);
2742 gsi_next (&gsi))
2744 gimple stmt = gsi_stmt (gsi);
2745 affine_iv iv;
2746 ssa_op_iter iter;
2747 tree use;
2749 FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE)
2751 predicate will_be_nonconstant;
2753 if (!simple_iv
2754 (loop, loop_containing_stmt (stmt), use, &iv, true)
2755 || is_gimple_min_invariant (iv.step))
2756 continue;
2757 will_be_nonconstant
2758 = will_be_nonconstant_expr_predicate (parms_info, info,
2759 iv.step,
2760 nonconstant_names);
2761 if (!true_predicate_p (&will_be_nonconstant))
2762 will_be_nonconstant
2763 = and_predicates (info->conds,
2764 &bb_predicate,
2765 &will_be_nonconstant);
2766 if (!true_predicate_p (&will_be_nonconstant)
2767 && !false_predicate_p (&will_be_nonconstant))
2768 /* This is slightly inprecise. We may want to represent
2769 each loop with independent predicate. */
2770 loop_stride =
2771 and_predicates (info->conds, &loop_stride,
2772 &will_be_nonconstant);
2776 free (body);
2778 set_hint_predicate (&inline_summary (node)->loop_iterations,
2779 loop_iterations);
2780 set_hint_predicate (&inline_summary (node)->loop_stride, loop_stride);
2781 scev_finalize ();
2783 FOR_ALL_BB_FN (bb, my_function)
2785 edge e;
2786 edge_iterator ei;
2788 if (bb->aux)
2789 pool_free (edge_predicate_pool, bb->aux);
2790 bb->aux = NULL;
2791 FOR_EACH_EDGE (e, ei, bb->succs)
2793 if (e->aux)
2794 pool_free (edge_predicate_pool, e->aux);
2795 e->aux = NULL;
2798 inline_summary (node)->self_time = time;
2799 inline_summary (node)->self_size = size;
2800 nonconstant_names.release ();
2801 if (optimize && !early)
2803 loop_optimizer_finalize ();
2804 free_dominance_info (CDI_DOMINATORS);
2806 if (dump_file)
2808 fprintf (dump_file, "\n");
2809 dump_inline_summary (dump_file, node);
2814 /* Compute parameters of functions used by inliner.
2815 EARLY is true when we compute parameters for the early inliner */
2817 void
2818 compute_inline_parameters (struct cgraph_node *node, bool early)
2820 HOST_WIDE_INT self_stack_size;
2821 struct cgraph_edge *e;
2822 struct inline_summary *info;
2824 gcc_assert (!node->global.inlined_to);
2826 inline_summary_alloc ();
2828 info = inline_summary (node);
2829 reset_inline_summary (node);
2831 /* FIXME: Thunks are inlinable, but tree-inline don't know how to do that.
2832 Once this happen, we will need to more curefully predict call
2833 statement size. */
2834 if (node->thunk.thunk_p)
2836 struct inline_edge_summary *es = inline_edge_summary (node->callees);
2837 struct predicate t = true_predicate ();
2839 info->inlinable = 0;
2840 node->callees->call_stmt_cannot_inline_p = true;
2841 node->local.can_change_signature = false;
2842 es->call_stmt_time = 1;
2843 es->call_stmt_size = 1;
2844 account_size_time (info, 0, 0, &t);
2845 return;
2848 /* Even is_gimple_min_invariant rely on current_function_decl. */
2849 push_cfun (DECL_STRUCT_FUNCTION (node->decl));
2851 /* Estimate the stack size for the function if we're optimizing. */
2852 self_stack_size = optimize ? estimated_stack_frame_size (node) : 0;
2853 info->estimated_self_stack_size = self_stack_size;
2854 info->estimated_stack_size = self_stack_size;
2855 info->stack_frame_offset = 0;
2857 /* Can this function be inlined at all? */
2858 if (!optimize && !lookup_attribute ("always_inline",
2859 DECL_ATTRIBUTES (node->decl)))
2860 info->inlinable = false;
2861 else
2862 info->inlinable = tree_inlinable_function_p (node->decl);
2864 /* Type attributes can use parameter indices to describe them. */
2865 if (TYPE_ATTRIBUTES (TREE_TYPE (node->decl)))
2866 node->local.can_change_signature = false;
2867 else
2869 /* Otherwise, inlinable functions always can change signature. */
2870 if (info->inlinable)
2871 node->local.can_change_signature = true;
2872 else
2874 /* Functions calling builtin_apply can not change signature. */
2875 for (e = node->callees; e; e = e->next_callee)
2877 tree cdecl = e->callee->decl;
2878 if (DECL_BUILT_IN (cdecl)
2879 && DECL_BUILT_IN_CLASS (cdecl) == BUILT_IN_NORMAL
2880 && (DECL_FUNCTION_CODE (cdecl) == BUILT_IN_APPLY_ARGS
2881 || DECL_FUNCTION_CODE (cdecl) == BUILT_IN_VA_START))
2882 break;
2884 node->local.can_change_signature = !e;
2887 estimate_function_body_sizes (node, early);
2889 for (e = node->callees; e; e = e->next_callee)
2890 if (e->callee->comdat_local_p ())
2891 break;
2892 node->calls_comdat_local = (e != NULL);
2894 /* Inlining characteristics are maintained by the cgraph_mark_inline. */
2895 info->time = info->self_time;
2896 info->size = info->self_size;
2897 info->stack_frame_offset = 0;
2898 info->estimated_stack_size = info->estimated_self_stack_size;
2899 #ifdef ENABLE_CHECKING
2900 inline_update_overall_summary (node);
2901 gcc_assert (info->time == info->self_time && info->size == info->self_size);
2902 #endif
2904 pop_cfun ();
2908 /* Compute parameters of functions used by inliner using
2909 current_function_decl. */
2911 static unsigned int
2912 compute_inline_parameters_for_current (void)
2914 compute_inline_parameters (cgraph_node::get (current_function_decl), true);
2915 return 0;
2918 namespace {
2920 const pass_data pass_data_inline_parameters =
2922 GIMPLE_PASS, /* type */
2923 "inline_param", /* name */
2924 OPTGROUP_INLINE, /* optinfo_flags */
2925 TV_INLINE_PARAMETERS, /* tv_id */
2926 0, /* properties_required */
2927 0, /* properties_provided */
2928 0, /* properties_destroyed */
2929 0, /* todo_flags_start */
2930 0, /* todo_flags_finish */
2933 class pass_inline_parameters : public gimple_opt_pass
2935 public:
2936 pass_inline_parameters (gcc::context *ctxt)
2937 : gimple_opt_pass (pass_data_inline_parameters, ctxt)
2940 /* opt_pass methods: */
2941 opt_pass * clone () { return new pass_inline_parameters (m_ctxt); }
2942 virtual unsigned int execute (function *)
2944 return compute_inline_parameters_for_current ();
2947 }; // class pass_inline_parameters
2949 } // anon namespace
2951 gimple_opt_pass *
2952 make_pass_inline_parameters (gcc::context *ctxt)
2954 return new pass_inline_parameters (ctxt);
2958 /* Estimate benefit devirtualizing indirect edge IE, provided KNOWN_VALS and
2959 KNOWN_BINFOS. */
2961 static bool
2962 estimate_edge_devirt_benefit (struct cgraph_edge *ie,
2963 int *size, int *time,
2964 vec<tree> known_vals,
2965 vec<tree> known_binfos,
2966 vec<ipa_agg_jump_function_p> known_aggs)
2968 tree target;
2969 struct cgraph_node *callee;
2970 struct inline_summary *isummary;
2971 enum availability avail;
2973 if (!known_vals.exists () && !known_binfos.exists ())
2974 return false;
2975 if (!flag_indirect_inlining)
2976 return false;
2978 target = ipa_get_indirect_edge_target (ie, known_vals, known_binfos,
2979 known_aggs);
2980 if (!target)
2981 return false;
2983 /* Account for difference in cost between indirect and direct calls. */
2984 *size -= (eni_size_weights.indirect_call_cost - eni_size_weights.call_cost);
2985 *time -= (eni_time_weights.indirect_call_cost - eni_time_weights.call_cost);
2986 gcc_checking_assert (*time >= 0);
2987 gcc_checking_assert (*size >= 0);
2989 callee = cgraph_node::get (target);
2990 if (!callee || !callee->definition)
2991 return false;
2992 callee = callee->function_symbol (&avail);
2993 if (avail < AVAIL_AVAILABLE)
2994 return false;
2995 isummary = inline_summary (callee);
2996 return isummary->inlinable;
2999 /* Increase SIZE, MIN_SIZE (if non-NULL) and TIME for size and time needed to
3000 handle edge E with probability PROB.
3001 Set HINTS if edge may be devirtualized.
3002 KNOWN_VALS, KNOWN_AGGS and KNOWN_BINFOS describe context of the call
3003 site. */
3005 static inline void
3006 estimate_edge_size_and_time (struct cgraph_edge *e, int *size, int *min_size,
3007 int *time,
3008 int prob,
3009 vec<tree> known_vals,
3010 vec<tree> known_binfos,
3011 vec<ipa_agg_jump_function_p> known_aggs,
3012 inline_hints *hints)
3014 struct inline_edge_summary *es = inline_edge_summary (e);
3015 int call_size = es->call_stmt_size;
3016 int call_time = es->call_stmt_time;
3017 int cur_size;
3018 if (!e->callee
3019 && estimate_edge_devirt_benefit (e, &call_size, &call_time,
3020 known_vals, known_binfos, known_aggs)
3021 && hints && e->maybe_hot_p ())
3022 *hints |= INLINE_HINT_indirect_call;
3023 cur_size = call_size * INLINE_SIZE_SCALE;
3024 *size += cur_size;
3025 if (min_size)
3026 *min_size += cur_size;
3027 *time += apply_probability ((gcov_type) call_time, prob)
3028 * e->frequency * (INLINE_TIME_SCALE / CGRAPH_FREQ_BASE);
3029 if (*time > MAX_TIME * INLINE_TIME_SCALE)
3030 *time = MAX_TIME * INLINE_TIME_SCALE;
3035 /* Increase SIZE, MIN_SIZE and TIME for size and time needed to handle all
3036 calls in NODE.
3037 POSSIBLE_TRUTHS, KNOWN_VALS, KNOWN_AGGS and KNOWN_BINFOS describe context of
3038 the call site. */
3040 static void
3041 estimate_calls_size_and_time (struct cgraph_node *node, int *size,
3042 int *min_size, int *time,
3043 inline_hints *hints,
3044 clause_t possible_truths,
3045 vec<tree> known_vals,
3046 vec<tree> known_binfos,
3047 vec<ipa_agg_jump_function_p> known_aggs)
3049 struct cgraph_edge *e;
3050 for (e = node->callees; e; e = e->next_callee)
3052 struct inline_edge_summary *es = inline_edge_summary (e);
3053 if (!es->predicate
3054 || evaluate_predicate (es->predicate, possible_truths))
3056 if (e->inline_failed)
3058 /* Predicates of calls shall not use NOT_CHANGED codes,
3059 sowe do not need to compute probabilities. */
3060 estimate_edge_size_and_time (e, size,
3061 es->predicate ? NULL : min_size,
3062 time, REG_BR_PROB_BASE,
3063 known_vals, known_binfos,
3064 known_aggs, hints);
3066 else
3067 estimate_calls_size_and_time (e->callee, size, min_size, time,
3068 hints,
3069 possible_truths,
3070 known_vals, known_binfos,
3071 known_aggs);
3074 for (e = node->indirect_calls; e; e = e->next_callee)
3076 struct inline_edge_summary *es = inline_edge_summary (e);
3077 if (!es->predicate
3078 || evaluate_predicate (es->predicate, possible_truths))
3079 estimate_edge_size_and_time (e, size,
3080 es->predicate ? NULL : min_size,
3081 time, REG_BR_PROB_BASE,
3082 known_vals, known_binfos, known_aggs,
3083 hints);
3088 /* Estimate size and time needed to execute NODE assuming
3089 POSSIBLE_TRUTHS clause, and KNOWN_VALS, KNOWN_AGGS and KNOWN_BINFOS
3090 information about NODE's arguments. If non-NULL use also probability
3091 information present in INLINE_PARAM_SUMMARY vector.
3092 Additionally detemine hints determined by the context. Finally compute
3093 minimal size needed for the call that is independent on the call context and
3094 can be used for fast estimates. Return the values in RET_SIZE,
3095 RET_MIN_SIZE, RET_TIME and RET_HINTS. */
3097 static void
3098 estimate_node_size_and_time (struct cgraph_node *node,
3099 clause_t possible_truths,
3100 vec<tree> known_vals,
3101 vec<tree> known_binfos,
3102 vec<ipa_agg_jump_function_p> known_aggs,
3103 int *ret_size, int *ret_min_size, int *ret_time,
3104 inline_hints *ret_hints,
3105 vec<inline_param_summary>
3106 inline_param_summary)
3108 struct inline_summary *info = inline_summary (node);
3109 size_time_entry *e;
3110 int size = 0;
3111 int time = 0;
3112 int min_size = 0;
3113 inline_hints hints = 0;
3114 int i;
3116 if (dump_file && (dump_flags & TDF_DETAILS))
3118 bool found = false;
3119 fprintf (dump_file, " Estimating body: %s/%i\n"
3120 " Known to be false: ", node->name (),
3121 node->order);
3123 for (i = predicate_not_inlined_condition;
3124 i < (predicate_first_dynamic_condition
3125 + (int) vec_safe_length (info->conds)); i++)
3126 if (!(possible_truths & (1 << i)))
3128 if (found)
3129 fprintf (dump_file, ", ");
3130 found = true;
3131 dump_condition (dump_file, info->conds, i);
3135 for (i = 0; vec_safe_iterate (info->entry, i, &e); i++)
3136 if (evaluate_predicate (&e->predicate, possible_truths))
3138 size += e->size;
3139 gcc_checking_assert (e->time >= 0);
3140 gcc_checking_assert (time >= 0);
3141 if (!inline_param_summary.exists ())
3142 time += e->time;
3143 else
3145 int prob = predicate_probability (info->conds,
3146 &e->predicate,
3147 possible_truths,
3148 inline_param_summary);
3149 gcc_checking_assert (prob >= 0);
3150 gcc_checking_assert (prob <= REG_BR_PROB_BASE);
3151 time += apply_probability ((gcov_type) e->time, prob);
3153 if (time > MAX_TIME * INLINE_TIME_SCALE)
3154 time = MAX_TIME * INLINE_TIME_SCALE;
3155 gcc_checking_assert (time >= 0);
3158 gcc_checking_assert (true_predicate_p (&(*info->entry)[0].predicate));
3159 min_size = (*info->entry)[0].size;
3160 gcc_checking_assert (size >= 0);
3161 gcc_checking_assert (time >= 0);
3163 if (info->loop_iterations
3164 && !evaluate_predicate (info->loop_iterations, possible_truths))
3165 hints |= INLINE_HINT_loop_iterations;
3166 if (info->loop_stride
3167 && !evaluate_predicate (info->loop_stride, possible_truths))
3168 hints |= INLINE_HINT_loop_stride;
3169 if (info->array_index
3170 && !evaluate_predicate (info->array_index, possible_truths))
3171 hints |= INLINE_HINT_array_index;
3172 if (info->scc_no)
3173 hints |= INLINE_HINT_in_scc;
3174 if (DECL_DECLARED_INLINE_P (node->decl))
3175 hints |= INLINE_HINT_declared_inline;
3177 estimate_calls_size_and_time (node, &size, &min_size, &time, &hints, possible_truths,
3178 known_vals, known_binfos, known_aggs);
3179 gcc_checking_assert (size >= 0);
3180 gcc_checking_assert (time >= 0);
3181 time = RDIV (time, INLINE_TIME_SCALE);
3182 size = RDIV (size, INLINE_SIZE_SCALE);
3183 min_size = RDIV (min_size, INLINE_SIZE_SCALE);
3185 if (dump_file && (dump_flags & TDF_DETAILS))
3186 fprintf (dump_file, "\n size:%i time:%i\n", (int) size, (int) time);
3187 if (ret_time)
3188 *ret_time = time;
3189 if (ret_size)
3190 *ret_size = size;
3191 if (ret_min_size)
3192 *ret_min_size = min_size;
3193 if (ret_hints)
3194 *ret_hints = hints;
3195 return;
3199 /* Estimate size and time needed to execute callee of EDGE assuming that
3200 parameters known to be constant at caller of EDGE are propagated.
3201 KNOWN_VALS and KNOWN_BINFOS are vectors of assumed known constant values
3202 and types for parameters. */
3204 void
3205 estimate_ipcp_clone_size_and_time (struct cgraph_node *node,
3206 vec<tree> known_vals,
3207 vec<tree> known_binfos,
3208 vec<ipa_agg_jump_function_p> known_aggs,
3209 int *ret_size, int *ret_time,
3210 inline_hints *hints)
3212 clause_t clause;
3214 clause = evaluate_conditions_for_known_args (node, false, known_vals,
3215 known_aggs);
3216 estimate_node_size_and_time (node, clause, known_vals, known_binfos,
3217 known_aggs, ret_size, NULL, ret_time, hints, vNULL);
3220 /* Translate all conditions from callee representation into caller
3221 representation and symbolically evaluate predicate P into new predicate.
3223 INFO is inline_summary of function we are adding predicate into, CALLEE_INFO
3224 is summary of function predicate P is from. OPERAND_MAP is array giving
3225 callee formal IDs the caller formal IDs. POSSSIBLE_TRUTHS is clausule of all
3226 callee conditions that may be true in caller context. TOPLEV_PREDICATE is
3227 predicate under which callee is executed. OFFSET_MAP is an array of of
3228 offsets that need to be added to conditions, negative offset means that
3229 conditions relying on values passed by reference have to be discarded
3230 because they might not be preserved (and should be considered offset zero
3231 for other purposes). */
3233 static struct predicate
3234 remap_predicate (struct inline_summary *info,
3235 struct inline_summary *callee_info,
3236 struct predicate *p,
3237 vec<int> operand_map,
3238 vec<int> offset_map,
3239 clause_t possible_truths, struct predicate *toplev_predicate)
3241 int i;
3242 struct predicate out = true_predicate ();
3244 /* True predicate is easy. */
3245 if (true_predicate_p (p))
3246 return *toplev_predicate;
3247 for (i = 0; p->clause[i]; i++)
3249 clause_t clause = p->clause[i];
3250 int cond;
3251 struct predicate clause_predicate = false_predicate ();
3253 gcc_assert (i < MAX_CLAUSES);
3255 for (cond = 0; cond < NUM_CONDITIONS; cond++)
3256 /* Do we have condition we can't disprove? */
3257 if (clause & possible_truths & (1 << cond))
3259 struct predicate cond_predicate;
3260 /* Work out if the condition can translate to predicate in the
3261 inlined function. */
3262 if (cond >= predicate_first_dynamic_condition)
3264 struct condition *c;
3266 c = &(*callee_info->conds)[cond
3268 predicate_first_dynamic_condition];
3269 /* See if we can remap condition operand to caller's operand.
3270 Otherwise give up. */
3271 if (!operand_map.exists ()
3272 || (int) operand_map.length () <= c->operand_num
3273 || operand_map[c->operand_num] == -1
3274 /* TODO: For non-aggregate conditions, adding an offset is
3275 basically an arithmetic jump function processing which
3276 we should support in future. */
3277 || ((!c->agg_contents || !c->by_ref)
3278 && offset_map[c->operand_num] > 0)
3279 || (c->agg_contents && c->by_ref
3280 && offset_map[c->operand_num] < 0))
3281 cond_predicate = true_predicate ();
3282 else
3284 struct agg_position_info ap;
3285 HOST_WIDE_INT offset_delta = offset_map[c->operand_num];
3286 if (offset_delta < 0)
3288 gcc_checking_assert (!c->agg_contents || !c->by_ref);
3289 offset_delta = 0;
3291 gcc_assert (!c->agg_contents
3292 || c->by_ref || offset_delta == 0);
3293 ap.offset = c->offset + offset_delta;
3294 ap.agg_contents = c->agg_contents;
3295 ap.by_ref = c->by_ref;
3296 cond_predicate = add_condition (info,
3297 operand_map[c->operand_num],
3298 &ap, c->code, c->val);
3301 /* Fixed conditions remains same, construct single
3302 condition predicate. */
3303 else
3305 cond_predicate.clause[0] = 1 << cond;
3306 cond_predicate.clause[1] = 0;
3308 clause_predicate = or_predicates (info->conds, &clause_predicate,
3309 &cond_predicate);
3311 out = and_predicates (info->conds, &out, &clause_predicate);
3313 return and_predicates (info->conds, &out, toplev_predicate);
3317 /* Update summary information of inline clones after inlining.
3318 Compute peak stack usage. */
3320 static void
3321 inline_update_callee_summaries (struct cgraph_node *node, int depth)
3323 struct cgraph_edge *e;
3324 struct inline_summary *callee_info = inline_summary (node);
3325 struct inline_summary *caller_info = inline_summary (node->callers->caller);
3326 HOST_WIDE_INT peak;
3328 callee_info->stack_frame_offset
3329 = caller_info->stack_frame_offset
3330 + caller_info->estimated_self_stack_size;
3331 peak = callee_info->stack_frame_offset
3332 + callee_info->estimated_self_stack_size;
3333 if (inline_summary (node->global.inlined_to)->estimated_stack_size < peak)
3334 inline_summary (node->global.inlined_to)->estimated_stack_size = peak;
3335 ipa_propagate_frequency (node);
3336 for (e = node->callees; e; e = e->next_callee)
3338 if (!e->inline_failed)
3339 inline_update_callee_summaries (e->callee, depth);
3340 inline_edge_summary (e)->loop_depth += depth;
3342 for (e = node->indirect_calls; e; e = e->next_callee)
3343 inline_edge_summary (e)->loop_depth += depth;
3346 /* Update change_prob of EDGE after INLINED_EDGE has been inlined.
3347 When functoin A is inlined in B and A calls C with parameter that
3348 changes with probability PROB1 and C is known to be passthroug
3349 of argument if B that change with probability PROB2, the probability
3350 of change is now PROB1*PROB2. */
3352 static void
3353 remap_edge_change_prob (struct cgraph_edge *inlined_edge,
3354 struct cgraph_edge *edge)
3356 if (ipa_node_params_vector.exists ())
3358 int i;
3359 struct ipa_edge_args *args = IPA_EDGE_REF (edge);
3360 struct inline_edge_summary *es = inline_edge_summary (edge);
3361 struct inline_edge_summary *inlined_es
3362 = inline_edge_summary (inlined_edge);
3364 for (i = 0; i < ipa_get_cs_argument_count (args); i++)
3366 struct ipa_jump_func *jfunc = ipa_get_ith_jump_func (args, i);
3367 if (jfunc->type == IPA_JF_PASS_THROUGH
3368 && (ipa_get_jf_pass_through_formal_id (jfunc)
3369 < (int) inlined_es->param.length ()))
3371 int jf_formal_id = ipa_get_jf_pass_through_formal_id (jfunc);
3372 int prob1 = es->param[i].change_prob;
3373 int prob2 = inlined_es->param[jf_formal_id].change_prob;
3374 int prob = combine_probabilities (prob1, prob2);
3376 if (prob1 && prob2 && !prob)
3377 prob = 1;
3379 es->param[i].change_prob = prob;
3385 /* Update edge summaries of NODE after INLINED_EDGE has been inlined.
3387 Remap predicates of callees of NODE. Rest of arguments match
3388 remap_predicate.
3390 Also update change probabilities. */
3392 static void
3393 remap_edge_summaries (struct cgraph_edge *inlined_edge,
3394 struct cgraph_node *node,
3395 struct inline_summary *info,
3396 struct inline_summary *callee_info,
3397 vec<int> operand_map,
3398 vec<int> offset_map,
3399 clause_t possible_truths,
3400 struct predicate *toplev_predicate)
3402 struct cgraph_edge *e;
3403 for (e = node->callees; e; e = e->next_callee)
3405 struct inline_edge_summary *es = inline_edge_summary (e);
3406 struct predicate p;
3408 if (e->inline_failed)
3410 remap_edge_change_prob (inlined_edge, e);
3412 if (es->predicate)
3414 p = remap_predicate (info, callee_info,
3415 es->predicate, operand_map, offset_map,
3416 possible_truths, toplev_predicate);
3417 edge_set_predicate (e, &p);
3418 /* TODO: We should remove the edge for code that will be
3419 optimized out, but we need to keep verifiers and tree-inline
3420 happy. Make it cold for now. */
3421 if (false_predicate_p (&p))
3423 e->count = 0;
3424 e->frequency = 0;
3427 else
3428 edge_set_predicate (e, toplev_predicate);
3430 else
3431 remap_edge_summaries (inlined_edge, e->callee, info, callee_info,
3432 operand_map, offset_map, possible_truths,
3433 toplev_predicate);
3435 for (e = node->indirect_calls; e; e = e->next_callee)
3437 struct inline_edge_summary *es = inline_edge_summary (e);
3438 struct predicate p;
3440 remap_edge_change_prob (inlined_edge, e);
3441 if (es->predicate)
3443 p = remap_predicate (info, callee_info,
3444 es->predicate, operand_map, offset_map,
3445 possible_truths, toplev_predicate);
3446 edge_set_predicate (e, &p);
3447 /* TODO: We should remove the edge for code that will be optimized
3448 out, but we need to keep verifiers and tree-inline happy.
3449 Make it cold for now. */
3450 if (false_predicate_p (&p))
3452 e->count = 0;
3453 e->frequency = 0;
3456 else
3457 edge_set_predicate (e, toplev_predicate);
3461 /* Same as remap_predicate, but set result into hint *HINT. */
3463 static void
3464 remap_hint_predicate (struct inline_summary *info,
3465 struct inline_summary *callee_info,
3466 struct predicate **hint,
3467 vec<int> operand_map,
3468 vec<int> offset_map,
3469 clause_t possible_truths,
3470 struct predicate *toplev_predicate)
3472 predicate p;
3474 if (!*hint)
3475 return;
3476 p = remap_predicate (info, callee_info,
3477 *hint,
3478 operand_map, offset_map,
3479 possible_truths, toplev_predicate);
3480 if (!false_predicate_p (&p) && !true_predicate_p (&p))
3482 if (!*hint)
3483 set_hint_predicate (hint, p);
3484 else
3485 **hint = and_predicates (info->conds, *hint, &p);
3489 /* We inlined EDGE. Update summary of the function we inlined into. */
3491 void
3492 inline_merge_summary (struct cgraph_edge *edge)
3494 struct inline_summary *callee_info = inline_summary (edge->callee);
3495 struct cgraph_node *to = (edge->caller->global.inlined_to
3496 ? edge->caller->global.inlined_to : edge->caller);
3497 struct inline_summary *info = inline_summary (to);
3498 clause_t clause = 0; /* not_inline is known to be false. */
3499 size_time_entry *e;
3500 vec<int> operand_map = vNULL;
3501 vec<int> offset_map = vNULL;
3502 int i;
3503 struct predicate toplev_predicate;
3504 struct predicate true_p = true_predicate ();
3505 struct inline_edge_summary *es = inline_edge_summary (edge);
3507 if (es->predicate)
3508 toplev_predicate = *es->predicate;
3509 else
3510 toplev_predicate = true_predicate ();
3512 if (ipa_node_params_vector.exists () && callee_info->conds)
3514 struct ipa_edge_args *args = IPA_EDGE_REF (edge);
3515 int count = ipa_get_cs_argument_count (args);
3516 int i;
3518 evaluate_properties_for_edge (edge, true, &clause, NULL, NULL, NULL);
3519 if (count)
3521 operand_map.safe_grow_cleared (count);
3522 offset_map.safe_grow_cleared (count);
3524 for (i = 0; i < count; i++)
3526 struct ipa_jump_func *jfunc = ipa_get_ith_jump_func (args, i);
3527 int map = -1;
3529 /* TODO: handle non-NOPs when merging. */
3530 if (jfunc->type == IPA_JF_PASS_THROUGH)
3532 if (ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
3533 map = ipa_get_jf_pass_through_formal_id (jfunc);
3534 if (!ipa_get_jf_pass_through_agg_preserved (jfunc))
3535 offset_map[i] = -1;
3537 else if (jfunc->type == IPA_JF_ANCESTOR)
3539 HOST_WIDE_INT offset = ipa_get_jf_ancestor_offset (jfunc);
3540 if (offset >= 0 && offset < INT_MAX)
3542 map = ipa_get_jf_ancestor_formal_id (jfunc);
3543 if (!ipa_get_jf_ancestor_agg_preserved (jfunc))
3544 offset = -1;
3545 offset_map[i] = offset;
3548 operand_map[i] = map;
3549 gcc_assert (map < ipa_get_param_count (IPA_NODE_REF (to)));
3552 for (i = 0; vec_safe_iterate (callee_info->entry, i, &e); i++)
3554 struct predicate p = remap_predicate (info, callee_info,
3555 &e->predicate, operand_map,
3556 offset_map, clause,
3557 &toplev_predicate);
3558 if (!false_predicate_p (&p))
3560 gcov_type add_time = ((gcov_type) e->time * edge->frequency
3561 + CGRAPH_FREQ_BASE / 2) / CGRAPH_FREQ_BASE;
3562 int prob = predicate_probability (callee_info->conds,
3563 &e->predicate,
3564 clause, es->param);
3565 add_time = apply_probability ((gcov_type) add_time, prob);
3566 if (add_time > MAX_TIME * INLINE_TIME_SCALE)
3567 add_time = MAX_TIME * INLINE_TIME_SCALE;
3568 if (prob != REG_BR_PROB_BASE
3569 && dump_file && (dump_flags & TDF_DETAILS))
3571 fprintf (dump_file, "\t\tScaling time by probability:%f\n",
3572 (double) prob / REG_BR_PROB_BASE);
3574 account_size_time (info, e->size, add_time, &p);
3577 remap_edge_summaries (edge, edge->callee, info, callee_info, operand_map,
3578 offset_map, clause, &toplev_predicate);
3579 remap_hint_predicate (info, callee_info,
3580 &callee_info->loop_iterations,
3581 operand_map, offset_map, clause, &toplev_predicate);
3582 remap_hint_predicate (info, callee_info,
3583 &callee_info->loop_stride,
3584 operand_map, offset_map, clause, &toplev_predicate);
3585 remap_hint_predicate (info, callee_info,
3586 &callee_info->array_index,
3587 operand_map, offset_map, clause, &toplev_predicate);
3589 inline_update_callee_summaries (edge->callee,
3590 inline_edge_summary (edge)->loop_depth);
3592 /* We do not maintain predicates of inlined edges, free it. */
3593 edge_set_predicate (edge, &true_p);
3594 /* Similarly remove param summaries. */
3595 es->param.release ();
3596 operand_map.release ();
3597 offset_map.release ();
3600 /* For performance reasons inline_merge_summary is not updating overall size
3601 and time. Recompute it. */
3603 void
3604 inline_update_overall_summary (struct cgraph_node *node)
3606 struct inline_summary *info = inline_summary (node);
3607 size_time_entry *e;
3608 int i;
3610 info->size = 0;
3611 info->time = 0;
3612 for (i = 0; vec_safe_iterate (info->entry, i, &e); i++)
3614 info->size += e->size, info->time += e->time;
3615 if (info->time > MAX_TIME * INLINE_TIME_SCALE)
3616 info->time = MAX_TIME * INLINE_TIME_SCALE;
3618 estimate_calls_size_and_time (node, &info->size, &info->min_size,
3619 &info->time, NULL,
3620 ~(clause_t) (1 << predicate_false_condition),
3621 vNULL, vNULL, vNULL);
3622 info->time = (info->time + INLINE_TIME_SCALE / 2) / INLINE_TIME_SCALE;
3623 info->size = (info->size + INLINE_SIZE_SCALE / 2) / INLINE_SIZE_SCALE;
3626 /* Return hints derrived from EDGE. */
3628 simple_edge_hints (struct cgraph_edge *edge)
3630 int hints = 0;
3631 struct cgraph_node *to = (edge->caller->global.inlined_to
3632 ? edge->caller->global.inlined_to : edge->caller);
3633 if (inline_summary (to)->scc_no
3634 && inline_summary (to)->scc_no == inline_summary (edge->callee)->scc_no
3635 && !edge->recursive_p ())
3636 hints |= INLINE_HINT_same_scc;
3638 if (to->lto_file_data && edge->callee->lto_file_data
3639 && to->lto_file_data != edge->callee->lto_file_data)
3640 hints |= INLINE_HINT_cross_module;
3642 return hints;
3645 /* Estimate the time cost for the caller when inlining EDGE.
3646 Only to be called via estimate_edge_time, that handles the
3647 caching mechanism.
3649 When caching, also update the cache entry. Compute both time and
3650 size, since we always need both metrics eventually. */
3653 do_estimate_edge_time (struct cgraph_edge *edge)
3655 int time;
3656 int size;
3657 inline_hints hints;
3658 struct cgraph_node *callee;
3659 clause_t clause;
3660 vec<tree> known_vals;
3661 vec<tree> known_binfos;
3662 vec<ipa_agg_jump_function_p> known_aggs;
3663 struct inline_edge_summary *es = inline_edge_summary (edge);
3664 int min_size;
3666 callee = edge->callee->ultimate_alias_target ();
3668 gcc_checking_assert (edge->inline_failed);
3669 evaluate_properties_for_edge (edge, true,
3670 &clause, &known_vals, &known_binfos,
3671 &known_aggs);
3672 estimate_node_size_and_time (callee, clause, known_vals, known_binfos,
3673 known_aggs, &size, &min_size, &time, &hints, es->param);
3675 /* When we have profile feedback, we can quite safely identify hot
3676 edges and for those we disable size limits. Don't do that when
3677 probability that caller will call the callee is low however, since it
3678 may hurt optimization of the caller's hot path. */
3679 if (edge->count && edge->maybe_hot_p ()
3680 && (edge->count * 2
3681 > (edge->caller->global.inlined_to
3682 ? edge->caller->global.inlined_to->count : edge->caller->count)))
3683 hints |= INLINE_HINT_known_hot;
3685 known_vals.release ();
3686 known_binfos.release ();
3687 known_aggs.release ();
3688 gcc_checking_assert (size >= 0);
3689 gcc_checking_assert (time >= 0);
3691 /* When caching, update the cache entry. */
3692 if (edge_growth_cache.exists ())
3694 inline_summary (edge->callee)->min_size = min_size;
3695 if ((int) edge_growth_cache.length () <= edge->uid)
3696 edge_growth_cache.safe_grow_cleared (symtab->edges_max_uid);
3697 edge_growth_cache[edge->uid].time = time + (time >= 0);
3699 edge_growth_cache[edge->uid].size = size + (size >= 0);
3700 hints |= simple_edge_hints (edge);
3701 edge_growth_cache[edge->uid].hints = hints + 1;
3703 return time;
3707 /* Return estimated callee growth after inlining EDGE.
3708 Only to be called via estimate_edge_size. */
3711 do_estimate_edge_size (struct cgraph_edge *edge)
3713 int size;
3714 struct cgraph_node *callee;
3715 clause_t clause;
3716 vec<tree> known_vals;
3717 vec<tree> known_binfos;
3718 vec<ipa_agg_jump_function_p> known_aggs;
3720 /* When we do caching, use do_estimate_edge_time to populate the entry. */
3722 if (edge_growth_cache.exists ())
3724 do_estimate_edge_time (edge);
3725 size = edge_growth_cache[edge->uid].size;
3726 gcc_checking_assert (size);
3727 return size - (size > 0);
3730 callee = edge->callee->ultimate_alias_target ();
3732 /* Early inliner runs without caching, go ahead and do the dirty work. */
3733 gcc_checking_assert (edge->inline_failed);
3734 evaluate_properties_for_edge (edge, true,
3735 &clause, &known_vals, &known_binfos,
3736 &known_aggs);
3737 estimate_node_size_and_time (callee, clause, known_vals, known_binfos,
3738 known_aggs, &size, NULL, NULL, NULL, vNULL);
3739 known_vals.release ();
3740 known_binfos.release ();
3741 known_aggs.release ();
3742 return size;
3746 /* Estimate the growth of the caller when inlining EDGE.
3747 Only to be called via estimate_edge_size. */
3749 inline_hints
3750 do_estimate_edge_hints (struct cgraph_edge *edge)
3752 inline_hints hints;
3753 struct cgraph_node *callee;
3754 clause_t clause;
3755 vec<tree> known_vals;
3756 vec<tree> known_binfos;
3757 vec<ipa_agg_jump_function_p> known_aggs;
3759 /* When we do caching, use do_estimate_edge_time to populate the entry. */
3761 if (edge_growth_cache.exists ())
3763 do_estimate_edge_time (edge);
3764 hints = edge_growth_cache[edge->uid].hints;
3765 gcc_checking_assert (hints);
3766 return hints - 1;
3769 callee = edge->callee->ultimate_alias_target ();
3771 /* Early inliner runs without caching, go ahead and do the dirty work. */
3772 gcc_checking_assert (edge->inline_failed);
3773 evaluate_properties_for_edge (edge, true,
3774 &clause, &known_vals, &known_binfos,
3775 &known_aggs);
3776 estimate_node_size_and_time (callee, clause, known_vals, known_binfos,
3777 known_aggs, NULL, NULL, NULL, &hints, vNULL);
3778 known_vals.release ();
3779 known_binfos.release ();
3780 known_aggs.release ();
3781 hints |= simple_edge_hints (edge);
3782 return hints;
3786 /* Estimate self time of the function NODE after inlining EDGE. */
3789 estimate_time_after_inlining (struct cgraph_node *node,
3790 struct cgraph_edge *edge)
3792 struct inline_edge_summary *es = inline_edge_summary (edge);
3793 if (!es->predicate || !false_predicate_p (es->predicate))
3795 gcov_type time =
3796 inline_summary (node)->time + estimate_edge_time (edge);
3797 if (time < 0)
3798 time = 0;
3799 if (time > MAX_TIME)
3800 time = MAX_TIME;
3801 return time;
3803 return inline_summary (node)->time;
3807 /* Estimate the size of NODE after inlining EDGE which should be an
3808 edge to either NODE or a call inlined into NODE. */
3811 estimate_size_after_inlining (struct cgraph_node *node,
3812 struct cgraph_edge *edge)
3814 struct inline_edge_summary *es = inline_edge_summary (edge);
3815 if (!es->predicate || !false_predicate_p (es->predicate))
3817 int size = inline_summary (node)->size + estimate_edge_growth (edge);
3818 gcc_assert (size >= 0);
3819 return size;
3821 return inline_summary (node)->size;
3825 struct growth_data
3827 struct cgraph_node *node;
3828 bool self_recursive;
3829 int growth;
3833 /* Worker for do_estimate_growth. Collect growth for all callers. */
3835 static bool
3836 do_estimate_growth_1 (struct cgraph_node *node, void *data)
3838 struct cgraph_edge *e;
3839 struct growth_data *d = (struct growth_data *) data;
3841 for (e = node->callers; e; e = e->next_caller)
3843 gcc_checking_assert (e->inline_failed);
3845 if (e->caller == d->node
3846 || (e->caller->global.inlined_to
3847 && e->caller->global.inlined_to == d->node))
3848 d->self_recursive = true;
3849 d->growth += estimate_edge_growth (e);
3851 return false;
3855 /* Estimate the growth caused by inlining NODE into all callees. */
3858 do_estimate_growth (struct cgraph_node *node)
3860 struct growth_data d = { node, 0, false };
3861 struct inline_summary *info = inline_summary (node);
3863 node->call_for_symbol_thunks_and_aliases (do_estimate_growth_1, &d, true);
3865 /* For self recursive functions the growth estimation really should be
3866 infinity. We don't want to return very large values because the growth
3867 plays various roles in badness computation fractions. Be sure to not
3868 return zero or negative growths. */
3869 if (d.self_recursive)
3870 d.growth = d.growth < info->size ? info->size : d.growth;
3871 else if (DECL_EXTERNAL (node->decl))
3873 else
3875 if (node->will_be_removed_from_program_if_no_direct_calls_p ())
3876 d.growth -= info->size;
3877 /* COMDAT functions are very often not shared across multiple units
3878 since they come from various template instantiations.
3879 Take this into account. */
3880 else if (DECL_COMDAT (node->decl)
3881 && node->can_remove_if_no_direct_calls_p ())
3882 d.growth -= (info->size
3883 * (100 - PARAM_VALUE (PARAM_COMDAT_SHARING_PROBABILITY))
3884 + 50) / 100;
3887 if (node_growth_cache.exists ())
3889 if ((int) node_growth_cache.length () <= node->uid)
3890 node_growth_cache.safe_grow_cleared (symtab->cgraph_max_uid);
3891 node_growth_cache[node->uid] = d.growth + (d.growth >= 0);
3893 return d.growth;
3897 /* Make cheap estimation if growth of NODE is likely positive knowing
3898 EDGE_GROWTH of one particular edge.
3899 We assume that most of other edges will have similar growth
3900 and skip computation if there are too many callers. */
3902 bool
3903 growth_likely_positive (struct cgraph_node *node, int edge_growth ATTRIBUTE_UNUSED)
3905 int max_callers;
3906 int ret;
3907 struct cgraph_edge *e;
3908 gcc_checking_assert (edge_growth > 0);
3910 /* Unlike for functions called once, we play unsafe with
3911 COMDATs. We can allow that since we know functions
3912 in consideration are small (and thus risk is small) and
3913 moreover grow estimates already accounts that COMDAT
3914 functions may or may not disappear when eliminated from
3915 current unit. With good probability making aggressive
3916 choice in all units is going to make overall program
3917 smaller.
3919 Consequently we ask cgraph_can_remove_if_no_direct_calls_p
3920 instead of
3921 cgraph_will_be_removed_from_program_if_no_direct_calls */
3922 if (DECL_EXTERNAL (node->decl)
3923 || !node->can_remove_if_no_direct_calls_p ())
3924 return true;
3926 /* If there is cached value, just go ahead. */
3927 if ((int)node_growth_cache.length () > node->uid
3928 && (ret = node_growth_cache[node->uid]))
3929 return ret > 0;
3930 if (!node->will_be_removed_from_program_if_no_direct_calls_p ()
3931 && (!DECL_COMDAT (node->decl)
3932 || !node->can_remove_if_no_direct_calls_p ()))
3933 return true;
3934 max_callers = inline_summary (node)->size * 4 / edge_growth + 2;
3936 for (e = node->callers; e; e = e->next_caller)
3938 max_callers--;
3939 if (!max_callers)
3940 return true;
3942 return estimate_growth (node) > 0;
3946 /* This function performs intraprocedural analysis in NODE that is required to
3947 inline indirect calls. */
3949 static void
3950 inline_indirect_intraprocedural_analysis (struct cgraph_node *node)
3952 ipa_analyze_node (node);
3953 if (dump_file && (dump_flags & TDF_DETAILS))
3955 ipa_print_node_params (dump_file, node);
3956 ipa_print_node_jump_functions (dump_file, node);
3961 /* Note function body size. */
3963 void
3964 inline_analyze_function (struct cgraph_node *node)
3966 push_cfun (DECL_STRUCT_FUNCTION (node->decl));
3968 if (dump_file)
3969 fprintf (dump_file, "\nAnalyzing function: %s/%u\n",
3970 node->name (), node->order);
3971 if (optimize && !node->thunk.thunk_p)
3972 inline_indirect_intraprocedural_analysis (node);
3973 compute_inline_parameters (node, false);
3974 if (!optimize)
3976 struct cgraph_edge *e;
3977 for (e = node->callees; e; e = e->next_callee)
3979 if (e->inline_failed == CIF_FUNCTION_NOT_CONSIDERED)
3980 e->inline_failed = CIF_FUNCTION_NOT_OPTIMIZED;
3981 e->call_stmt_cannot_inline_p = true;
3983 for (e = node->indirect_calls; e; e = e->next_callee)
3985 if (e->inline_failed == CIF_FUNCTION_NOT_CONSIDERED)
3986 e->inline_failed = CIF_FUNCTION_NOT_OPTIMIZED;
3987 e->call_stmt_cannot_inline_p = true;
3991 pop_cfun ();
3995 /* Called when new function is inserted to callgraph late. */
3997 static void
3998 add_new_function (struct cgraph_node *node, void *data ATTRIBUTE_UNUSED)
4000 inline_analyze_function (node);
4004 /* Note function body size. */
4006 void
4007 inline_generate_summary (void)
4009 struct cgraph_node *node;
4011 /* When not optimizing, do not bother to analyze. Inlining is still done
4012 because edge redirection needs to happen there. */
4013 if (!optimize && !flag_lto && !flag_wpa)
4014 return;
4016 function_insertion_hook_holder =
4017 symtab->add_cgraph_insertion_hook (&add_new_function, NULL);
4019 ipa_register_cgraph_hooks ();
4020 inline_free_summary ();
4022 FOR_EACH_DEFINED_FUNCTION (node)
4023 if (!node->alias)
4024 inline_analyze_function (node);
4028 /* Read predicate from IB. */
4030 static struct predicate
4031 read_predicate (struct lto_input_block *ib)
4033 struct predicate out;
4034 clause_t clause;
4035 int k = 0;
4039 gcc_assert (k <= MAX_CLAUSES);
4040 clause = out.clause[k++] = streamer_read_uhwi (ib);
4042 while (clause);
4044 /* Zero-initialize the remaining clauses in OUT. */
4045 while (k <= MAX_CLAUSES)
4046 out.clause[k++] = 0;
4048 return out;
4052 /* Write inline summary for edge E to OB. */
4054 static void
4055 read_inline_edge_summary (struct lto_input_block *ib, struct cgraph_edge *e)
4057 struct inline_edge_summary *es = inline_edge_summary (e);
4058 struct predicate p;
4059 int length, i;
4061 es->call_stmt_size = streamer_read_uhwi (ib);
4062 es->call_stmt_time = streamer_read_uhwi (ib);
4063 es->loop_depth = streamer_read_uhwi (ib);
4064 p = read_predicate (ib);
4065 edge_set_predicate (e, &p);
4066 length = streamer_read_uhwi (ib);
4067 if (length)
4069 es->param.safe_grow_cleared (length);
4070 for (i = 0; i < length; i++)
4071 es->param[i].change_prob = streamer_read_uhwi (ib);
4076 /* Stream in inline summaries from the section. */
4078 static void
4079 inline_read_section (struct lto_file_decl_data *file_data, const char *data,
4080 size_t len)
4082 const struct lto_function_header *header =
4083 (const struct lto_function_header *) data;
4084 const int cfg_offset = sizeof (struct lto_function_header);
4085 const int main_offset = cfg_offset + header->cfg_size;
4086 const int string_offset = main_offset + header->main_size;
4087 struct data_in *data_in;
4088 unsigned int i, count2, j;
4089 unsigned int f_count;
4091 lto_input_block ib ((const char *) data + main_offset, header->main_size);
4093 data_in =
4094 lto_data_in_create (file_data, (const char *) data + string_offset,
4095 header->string_size, vNULL);
4096 f_count = streamer_read_uhwi (&ib);
4097 for (i = 0; i < f_count; i++)
4099 unsigned int index;
4100 struct cgraph_node *node;
4101 struct inline_summary *info;
4102 lto_symtab_encoder_t encoder;
4103 struct bitpack_d bp;
4104 struct cgraph_edge *e;
4105 predicate p;
4107 index = streamer_read_uhwi (&ib);
4108 encoder = file_data->symtab_node_encoder;
4109 node = dyn_cast<cgraph_node *> (lto_symtab_encoder_deref (encoder,
4110 index));
4111 info = inline_summary (node);
4113 info->estimated_stack_size
4114 = info->estimated_self_stack_size = streamer_read_uhwi (&ib);
4115 info->size = info->self_size = streamer_read_uhwi (&ib);
4116 info->time = info->self_time = streamer_read_uhwi (&ib);
4118 bp = streamer_read_bitpack (&ib);
4119 info->inlinable = bp_unpack_value (&bp, 1);
4121 count2 = streamer_read_uhwi (&ib);
4122 gcc_assert (!info->conds);
4123 for (j = 0; j < count2; j++)
4125 struct condition c;
4126 c.operand_num = streamer_read_uhwi (&ib);
4127 c.code = (enum tree_code) streamer_read_uhwi (&ib);
4128 c.val = stream_read_tree (&ib, data_in);
4129 bp = streamer_read_bitpack (&ib);
4130 c.agg_contents = bp_unpack_value (&bp, 1);
4131 c.by_ref = bp_unpack_value (&bp, 1);
4132 if (c.agg_contents)
4133 c.offset = streamer_read_uhwi (&ib);
4134 vec_safe_push (info->conds, c);
4136 count2 = streamer_read_uhwi (&ib);
4137 gcc_assert (!info->entry);
4138 for (j = 0; j < count2; j++)
4140 struct size_time_entry e;
4142 e.size = streamer_read_uhwi (&ib);
4143 e.time = streamer_read_uhwi (&ib);
4144 e.predicate = read_predicate (&ib);
4146 vec_safe_push (info->entry, e);
4149 p = read_predicate (&ib);
4150 set_hint_predicate (&info->loop_iterations, p);
4151 p = read_predicate (&ib);
4152 set_hint_predicate (&info->loop_stride, p);
4153 p = read_predicate (&ib);
4154 set_hint_predicate (&info->array_index, p);
4155 for (e = node->callees; e; e = e->next_callee)
4156 read_inline_edge_summary (&ib, e);
4157 for (e = node->indirect_calls; e; e = e->next_callee)
4158 read_inline_edge_summary (&ib, e);
4161 lto_free_section_data (file_data, LTO_section_inline_summary, NULL, data,
4162 len);
4163 lto_data_in_delete (data_in);
4167 /* Read inline summary. Jump functions are shared among ipa-cp
4168 and inliner, so when ipa-cp is active, we don't need to write them
4169 twice. */
4171 void
4172 inline_read_summary (void)
4174 struct lto_file_decl_data **file_data_vec = lto_get_file_decl_data ();
4175 struct lto_file_decl_data *file_data;
4176 unsigned int j = 0;
4178 inline_summary_alloc ();
4180 while ((file_data = file_data_vec[j++]))
4182 size_t len;
4183 const char *data = lto_get_section_data (file_data,
4184 LTO_section_inline_summary,
4185 NULL, &len);
4186 if (data)
4187 inline_read_section (file_data, data, len);
4188 else
4189 /* Fatal error here. We do not want to support compiling ltrans units
4190 with different version of compiler or different flags than the WPA
4191 unit, so this should never happen. */
4192 fatal_error ("ipa inline summary is missing in input file");
4194 if (optimize)
4196 ipa_register_cgraph_hooks ();
4197 if (!flag_ipa_cp)
4198 ipa_prop_read_jump_functions ();
4200 function_insertion_hook_holder =
4201 symtab->add_cgraph_insertion_hook (&add_new_function, NULL);
4205 /* Write predicate P to OB. */
4207 static void
4208 write_predicate (struct output_block *ob, struct predicate *p)
4210 int j;
4211 if (p)
4212 for (j = 0; p->clause[j]; j++)
4214 gcc_assert (j < MAX_CLAUSES);
4215 streamer_write_uhwi (ob, p->clause[j]);
4217 streamer_write_uhwi (ob, 0);
4221 /* Write inline summary for edge E to OB. */
4223 static void
4224 write_inline_edge_summary (struct output_block *ob, struct cgraph_edge *e)
4226 struct inline_edge_summary *es = inline_edge_summary (e);
4227 int i;
4229 streamer_write_uhwi (ob, es->call_stmt_size);
4230 streamer_write_uhwi (ob, es->call_stmt_time);
4231 streamer_write_uhwi (ob, es->loop_depth);
4232 write_predicate (ob, es->predicate);
4233 streamer_write_uhwi (ob, es->param.length ());
4234 for (i = 0; i < (int) es->param.length (); i++)
4235 streamer_write_uhwi (ob, es->param[i].change_prob);
4239 /* Write inline summary for node in SET.
4240 Jump functions are shared among ipa-cp and inliner, so when ipa-cp is
4241 active, we don't need to write them twice. */
4243 void
4244 inline_write_summary (void)
4246 struct cgraph_node *node;
4247 struct output_block *ob = create_output_block (LTO_section_inline_summary);
4248 lto_symtab_encoder_t encoder = ob->decl_state->symtab_node_encoder;
4249 unsigned int count = 0;
4250 int i;
4252 for (i = 0; i < lto_symtab_encoder_size (encoder); i++)
4254 symtab_node *snode = lto_symtab_encoder_deref (encoder, i);
4255 cgraph_node *cnode = dyn_cast <cgraph_node *> (snode);
4256 if (cnode && cnode->definition && !cnode->alias)
4257 count++;
4259 streamer_write_uhwi (ob, count);
4261 for (i = 0; i < lto_symtab_encoder_size (encoder); i++)
4263 symtab_node *snode = lto_symtab_encoder_deref (encoder, i);
4264 cgraph_node *cnode = dyn_cast <cgraph_node *> (snode);
4265 if (cnode && (node = cnode)->definition && !node->alias)
4267 struct inline_summary *info = inline_summary (node);
4268 struct bitpack_d bp;
4269 struct cgraph_edge *edge;
4270 int i;
4271 size_time_entry *e;
4272 struct condition *c;
4274 streamer_write_uhwi (ob,
4275 lto_symtab_encoder_encode (encoder,
4277 node));
4278 streamer_write_hwi (ob, info->estimated_self_stack_size);
4279 streamer_write_hwi (ob, info->self_size);
4280 streamer_write_hwi (ob, info->self_time);
4281 bp = bitpack_create (ob->main_stream);
4282 bp_pack_value (&bp, info->inlinable, 1);
4283 streamer_write_bitpack (&bp);
4284 streamer_write_uhwi (ob, vec_safe_length (info->conds));
4285 for (i = 0; vec_safe_iterate (info->conds, i, &c); i++)
4287 streamer_write_uhwi (ob, c->operand_num);
4288 streamer_write_uhwi (ob, c->code);
4289 stream_write_tree (ob, c->val, true);
4290 bp = bitpack_create (ob->main_stream);
4291 bp_pack_value (&bp, c->agg_contents, 1);
4292 bp_pack_value (&bp, c->by_ref, 1);
4293 streamer_write_bitpack (&bp);
4294 if (c->agg_contents)
4295 streamer_write_uhwi (ob, c->offset);
4297 streamer_write_uhwi (ob, vec_safe_length (info->entry));
4298 for (i = 0; vec_safe_iterate (info->entry, i, &e); i++)
4300 streamer_write_uhwi (ob, e->size);
4301 streamer_write_uhwi (ob, e->time);
4302 write_predicate (ob, &e->predicate);
4304 write_predicate (ob, info->loop_iterations);
4305 write_predicate (ob, info->loop_stride);
4306 write_predicate (ob, info->array_index);
4307 for (edge = node->callees; edge; edge = edge->next_callee)
4308 write_inline_edge_summary (ob, edge);
4309 for (edge = node->indirect_calls; edge; edge = edge->next_callee)
4310 write_inline_edge_summary (ob, edge);
4313 streamer_write_char_stream (ob->main_stream, 0);
4314 produce_asm (ob, NULL);
4315 destroy_output_block (ob);
4317 if (optimize && !flag_ipa_cp)
4318 ipa_prop_write_jump_functions ();
4322 /* Release inline summary. */
4324 void
4325 inline_free_summary (void)
4327 struct cgraph_node *node;
4328 if (!inline_edge_summary_vec.exists ())
4329 return;
4330 FOR_EACH_DEFINED_FUNCTION (node)
4331 if (!node->alias)
4332 reset_inline_summary (node);
4333 if (function_insertion_hook_holder)
4334 symtab->remove_cgraph_insertion_hook (function_insertion_hook_holder);
4335 function_insertion_hook_holder = NULL;
4336 if (node_removal_hook_holder)
4337 symtab->remove_cgraph_removal_hook (node_removal_hook_holder);
4338 node_removal_hook_holder = NULL;
4339 if (edge_removal_hook_holder)
4340 symtab->remove_edge_removal_hook (edge_removal_hook_holder);
4341 edge_removal_hook_holder = NULL;
4342 if (node_duplication_hook_holder)
4343 symtab->remove_cgraph_duplication_hook (node_duplication_hook_holder);
4344 node_duplication_hook_holder = NULL;
4345 if (edge_duplication_hook_holder)
4346 symtab->remove_edge_duplication_hook (edge_duplication_hook_holder);
4347 edge_duplication_hook_holder = NULL;
4348 vec_free (inline_summary_vec);
4349 inline_edge_summary_vec.release ();
4350 if (edge_predicate_pool)
4351 free_alloc_pool (edge_predicate_pool);
4352 edge_predicate_pool = 0;