PR target/59794
[official-gcc.git] / gcc / ipa-inline-analysis.c
blob6cd5dc1c2e57be49e8670b908b8af213c5f7e0c8
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 gcc_assert (!hints);
678 /* Record SIZE and TIME under condition PRED into the inline summary. */
680 static void
681 account_size_time (struct inline_summary *summary, int size, int time,
682 struct predicate *pred)
684 size_time_entry *e;
685 bool found = false;
686 int i;
688 if (false_predicate_p (pred))
689 return;
691 /* We need to create initial empty unconitional clause, but otherwie
692 we don't need to account empty times and sizes. */
693 if (!size && !time && summary->entry)
694 return;
696 /* Watch overflow that might result from insane profiles. */
697 if (time > MAX_TIME * INLINE_TIME_SCALE)
698 time = MAX_TIME * INLINE_TIME_SCALE;
699 gcc_assert (time >= 0);
701 for (i = 0; vec_safe_iterate (summary->entry, i, &e); i++)
702 if (predicates_equal_p (&e->predicate, pred))
704 found = true;
705 break;
707 if (i == 256)
709 i = 0;
710 found = true;
711 e = &(*summary->entry)[0];
712 gcc_assert (!e->predicate.clause[0]);
713 if (dump_file && (dump_flags & TDF_DETAILS))
714 fprintf (dump_file,
715 "\t\tReached limit on number of entries, "
716 "ignoring the predicate.");
718 if (dump_file && (dump_flags & TDF_DETAILS) && (time || size))
720 fprintf (dump_file,
721 "\t\tAccounting size:%3.2f, time:%3.2f on %spredicate:",
722 ((double) size) / INLINE_SIZE_SCALE,
723 ((double) time) / INLINE_TIME_SCALE, found ? "" : "new ");
724 dump_predicate (dump_file, summary->conds, pred);
726 if (!found)
728 struct size_time_entry new_entry;
729 new_entry.size = size;
730 new_entry.time = time;
731 new_entry.predicate = *pred;
732 vec_safe_push (summary->entry, new_entry);
734 else
736 e->size += size;
737 e->time += time;
738 if (e->time > MAX_TIME * INLINE_TIME_SCALE)
739 e->time = MAX_TIME * INLINE_TIME_SCALE;
743 /* Set predicate for edge E. */
745 static void
746 edge_set_predicate (struct cgraph_edge *e, struct predicate *predicate)
748 struct inline_edge_summary *es = inline_edge_summary (e);
749 if (predicate && !true_predicate_p (predicate))
751 if (!es->predicate)
752 es->predicate = (struct predicate *) pool_alloc (edge_predicate_pool);
753 *es->predicate = *predicate;
755 else
757 if (es->predicate)
758 pool_free (edge_predicate_pool, es->predicate);
759 es->predicate = NULL;
763 /* Set predicate for hint *P. */
765 static void
766 set_hint_predicate (struct predicate **p, struct predicate new_predicate)
768 if (false_predicate_p (&new_predicate) || true_predicate_p (&new_predicate))
770 if (*p)
771 pool_free (edge_predicate_pool, *p);
772 *p = NULL;
774 else
776 if (!*p)
777 *p = (struct predicate *) pool_alloc (edge_predicate_pool);
778 **p = new_predicate;
783 /* KNOWN_VALS is partial mapping of parameters of NODE to constant values.
784 KNOWN_AGGS is a vector of aggreggate jump functions for each parameter.
785 Return clause of possible truths. When INLINE_P is true, assume that we are
786 inlining.
788 ERROR_MARK means compile time invariant. */
790 static clause_t
791 evaluate_conditions_for_known_args (struct cgraph_node *node,
792 bool inline_p,
793 vec<tree> known_vals,
794 vec<ipa_agg_jump_function_p>
795 known_aggs)
797 clause_t clause = inline_p ? 0 : 1 << predicate_not_inlined_condition;
798 struct inline_summary *info = inline_summary (node);
799 int i;
800 struct condition *c;
802 for (i = 0; vec_safe_iterate (info->conds, i, &c); i++)
804 tree val;
805 tree res;
807 /* We allow call stmt to have fewer arguments than the callee function
808 (especially for K&R style programs). So bound check here (we assume
809 known_aggs vector, if non-NULL, has the same length as
810 known_vals). */
811 gcc_checking_assert (!known_aggs.exists ()
812 || (known_vals.length () == known_aggs.length ()));
813 if (c->operand_num >= (int) known_vals.length ())
815 clause |= 1 << (i + predicate_first_dynamic_condition);
816 continue;
819 if (c->agg_contents)
821 struct ipa_agg_jump_function *agg;
823 if (c->code == CHANGED
824 && !c->by_ref
825 && (known_vals[c->operand_num] == error_mark_node))
826 continue;
828 if (known_aggs.exists ())
830 agg = known_aggs[c->operand_num];
831 val = ipa_find_agg_cst_for_param (agg, c->offset, c->by_ref);
833 else
834 val = NULL_TREE;
836 else
838 val = known_vals[c->operand_num];
839 if (val == error_mark_node && c->code != CHANGED)
840 val = NULL_TREE;
843 if (!val)
845 clause |= 1 << (i + predicate_first_dynamic_condition);
846 continue;
848 if (c->code == IS_NOT_CONSTANT || c->code == CHANGED)
849 continue;
850 res = fold_binary_to_constant (c->code, boolean_type_node, val, c->val);
851 if (res && integer_zerop (res))
852 continue;
853 clause |= 1 << (i + predicate_first_dynamic_condition);
855 return clause;
859 /* Work out what conditions might be true at invocation of E. */
861 static void
862 evaluate_properties_for_edge (struct cgraph_edge *e, bool inline_p,
863 clause_t *clause_ptr,
864 vec<tree> *known_vals_ptr,
865 vec<tree> *known_binfos_ptr,
866 vec<ipa_agg_jump_function_p> *known_aggs_ptr)
868 struct cgraph_node *callee =
869 cgraph_function_or_thunk_node (e->callee, NULL);
870 struct inline_summary *info = inline_summary (callee);
871 vec<tree> known_vals = vNULL;
872 vec<ipa_agg_jump_function_p> known_aggs = vNULL;
874 if (clause_ptr)
875 *clause_ptr = inline_p ? 0 : 1 << predicate_not_inlined_condition;
876 if (known_vals_ptr)
877 known_vals_ptr->create (0);
878 if (known_binfos_ptr)
879 known_binfos_ptr->create (0);
881 if (ipa_node_params_vector.exists ()
882 && !e->call_stmt_cannot_inline_p
883 && ((clause_ptr && info->conds) || known_vals_ptr || known_binfos_ptr))
885 struct ipa_node_params *parms_info;
886 struct ipa_edge_args *args = IPA_EDGE_REF (e);
887 struct inline_edge_summary *es = inline_edge_summary (e);
888 int i, count = ipa_get_cs_argument_count (args);
890 if (e->caller->global.inlined_to)
891 parms_info = IPA_NODE_REF (e->caller->global.inlined_to);
892 else
893 parms_info = IPA_NODE_REF (e->caller);
895 if (count && (info->conds || known_vals_ptr))
896 known_vals.safe_grow_cleared (count);
897 if (count && (info->conds || known_aggs_ptr))
898 known_aggs.safe_grow_cleared (count);
899 if (count && known_binfos_ptr)
900 known_binfos_ptr->safe_grow_cleared (count);
902 for (i = 0; i < count; i++)
904 struct ipa_jump_func *jf = ipa_get_ith_jump_func (args, i);
905 tree cst = ipa_value_from_jfunc (parms_info, jf);
906 if (cst)
908 if (known_vals.exists () && TREE_CODE (cst) != TREE_BINFO)
909 known_vals[i] = cst;
910 else if (known_binfos_ptr != NULL
911 && TREE_CODE (cst) == TREE_BINFO)
912 (*known_binfos_ptr)[i] = cst;
914 else if (inline_p && !es->param[i].change_prob)
915 known_vals[i] = error_mark_node;
916 /* TODO: When IPA-CP starts propagating and merging aggregate jump
917 functions, use its knowledge of the caller too, just like the
918 scalar case above. */
919 known_aggs[i] = &jf->agg;
923 if (clause_ptr)
924 *clause_ptr = evaluate_conditions_for_known_args (callee, inline_p,
925 known_vals, known_aggs);
927 if (known_vals_ptr)
928 *known_vals_ptr = known_vals;
929 else
930 known_vals.release ();
932 if (known_aggs_ptr)
933 *known_aggs_ptr = known_aggs;
934 else
935 known_aggs.release ();
939 /* Allocate the inline summary vector or resize it to cover all cgraph nodes. */
941 static void
942 inline_summary_alloc (void)
944 if (!node_removal_hook_holder)
945 node_removal_hook_holder =
946 cgraph_add_node_removal_hook (&inline_node_removal_hook, NULL);
947 if (!edge_removal_hook_holder)
948 edge_removal_hook_holder =
949 cgraph_add_edge_removal_hook (&inline_edge_removal_hook, NULL);
950 if (!node_duplication_hook_holder)
951 node_duplication_hook_holder =
952 cgraph_add_node_duplication_hook (&inline_node_duplication_hook, NULL);
953 if (!edge_duplication_hook_holder)
954 edge_duplication_hook_holder =
955 cgraph_add_edge_duplication_hook (&inline_edge_duplication_hook, NULL);
957 if (vec_safe_length (inline_summary_vec) <= (unsigned) cgraph_max_uid)
958 vec_safe_grow_cleared (inline_summary_vec, cgraph_max_uid + 1);
959 if (inline_edge_summary_vec.length () <= (unsigned) cgraph_edge_max_uid)
960 inline_edge_summary_vec.safe_grow_cleared (cgraph_edge_max_uid + 1);
961 if (!edge_predicate_pool)
962 edge_predicate_pool = create_alloc_pool ("edge predicates",
963 sizeof (struct predicate), 10);
966 /* We are called multiple time for given function; clear
967 data from previous run so they are not cumulated. */
969 static void
970 reset_inline_edge_summary (struct cgraph_edge *e)
972 if (e->uid < (int) inline_edge_summary_vec.length ())
974 struct inline_edge_summary *es = inline_edge_summary (e);
976 es->call_stmt_size = es->call_stmt_time = 0;
977 if (es->predicate)
978 pool_free (edge_predicate_pool, es->predicate);
979 es->predicate = NULL;
980 es->param.release ();
984 /* We are called multiple time for given function; clear
985 data from previous run so they are not cumulated. */
987 static void
988 reset_inline_summary (struct cgraph_node *node)
990 struct inline_summary *info = inline_summary (node);
991 struct cgraph_edge *e;
993 info->self_size = info->self_time = 0;
994 info->estimated_stack_size = 0;
995 info->estimated_self_stack_size = 0;
996 info->stack_frame_offset = 0;
997 info->size = 0;
998 info->time = 0;
999 info->growth = 0;
1000 info->scc_no = 0;
1001 if (info->loop_iterations)
1003 pool_free (edge_predicate_pool, info->loop_iterations);
1004 info->loop_iterations = NULL;
1006 if (info->loop_stride)
1008 pool_free (edge_predicate_pool, info->loop_stride);
1009 info->loop_stride = NULL;
1011 if (info->array_index)
1013 pool_free (edge_predicate_pool, info->array_index);
1014 info->array_index = NULL;
1016 vec_free (info->conds);
1017 vec_free (info->entry);
1018 for (e = node->callees; e; e = e->next_callee)
1019 reset_inline_edge_summary (e);
1020 for (e = node->indirect_calls; e; e = e->next_callee)
1021 reset_inline_edge_summary (e);
1024 /* Hook that is called by cgraph.c when a node is removed. */
1026 static void
1027 inline_node_removal_hook (struct cgraph_node *node,
1028 void *data ATTRIBUTE_UNUSED)
1030 struct inline_summary *info;
1031 if (vec_safe_length (inline_summary_vec) <= (unsigned) node->uid)
1032 return;
1033 info = inline_summary (node);
1034 reset_inline_summary (node);
1035 memset (info, 0, sizeof (inline_summary_t));
1038 /* Remap predicate P of former function to be predicate of duplicated function.
1039 POSSIBLE_TRUTHS is clause of possible truths in the duplicated node,
1040 INFO is inline summary of the duplicated node. */
1042 static struct predicate
1043 remap_predicate_after_duplication (struct predicate *p,
1044 clause_t possible_truths,
1045 struct inline_summary *info)
1047 struct predicate new_predicate = true_predicate ();
1048 int j;
1049 for (j = 0; p->clause[j]; j++)
1050 if (!(possible_truths & p->clause[j]))
1052 new_predicate = false_predicate ();
1053 break;
1055 else
1056 add_clause (info->conds, &new_predicate,
1057 possible_truths & p->clause[j]);
1058 return new_predicate;
1061 /* Same as remap_predicate_after_duplication but handle hint predicate *P.
1062 Additionally care about allocating new memory slot for updated predicate
1063 and set it to NULL when it becomes true or false (and thus uninteresting).
1066 static void
1067 remap_hint_predicate_after_duplication (struct predicate **p,
1068 clause_t possible_truths,
1069 struct inline_summary *info)
1071 struct predicate new_predicate;
1073 if (!*p)
1074 return;
1076 new_predicate = remap_predicate_after_duplication (*p,
1077 possible_truths, info);
1078 /* We do not want to free previous predicate; it is used by node origin. */
1079 *p = NULL;
1080 set_hint_predicate (p, new_predicate);
1084 /* Hook that is called by cgraph.c when a node is duplicated. */
1086 static void
1087 inline_node_duplication_hook (struct cgraph_node *src,
1088 struct cgraph_node *dst,
1089 ATTRIBUTE_UNUSED void *data)
1091 struct inline_summary *info;
1092 inline_summary_alloc ();
1093 info = inline_summary (dst);
1094 memcpy (info, inline_summary (src), sizeof (struct inline_summary));
1095 /* TODO: as an optimization, we may avoid copying conditions
1096 that are known to be false or true. */
1097 info->conds = vec_safe_copy (info->conds);
1099 /* When there are any replacements in the function body, see if we can figure
1100 out that something was optimized out. */
1101 if (ipa_node_params_vector.exists () && dst->clone.tree_map)
1103 vec<size_time_entry, va_gc> *entry = info->entry;
1104 /* Use SRC parm info since it may not be copied yet. */
1105 struct ipa_node_params *parms_info = IPA_NODE_REF (src);
1106 vec<tree> known_vals = vNULL;
1107 int count = ipa_get_param_count (parms_info);
1108 int i, j;
1109 clause_t possible_truths;
1110 struct predicate true_pred = true_predicate ();
1111 size_time_entry *e;
1112 int optimized_out_size = 0;
1113 bool inlined_to_p = false;
1114 struct cgraph_edge *edge;
1116 info->entry = 0;
1117 known_vals.safe_grow_cleared (count);
1118 for (i = 0; i < count; i++)
1120 struct ipa_replace_map *r;
1122 for (j = 0; vec_safe_iterate (dst->clone.tree_map, j, &r); j++)
1124 if (((!r->old_tree && r->parm_num == i)
1125 || (r->old_tree && r->old_tree == ipa_get_param (parms_info, i)))
1126 && r->replace_p && !r->ref_p)
1128 known_vals[i] = r->new_tree;
1129 break;
1133 possible_truths = evaluate_conditions_for_known_args (dst, false,
1134 known_vals,
1135 vNULL);
1136 known_vals.release ();
1138 account_size_time (info, 0, 0, &true_pred);
1140 /* Remap size_time vectors.
1141 Simplify the predicate by prunning out alternatives that are known
1142 to be false.
1143 TODO: as on optimization, we can also eliminate conditions known
1144 to be true. */
1145 for (i = 0; vec_safe_iterate (entry, i, &e); i++)
1147 struct predicate new_predicate;
1148 new_predicate = remap_predicate_after_duplication (&e->predicate,
1149 possible_truths,
1150 info);
1151 if (false_predicate_p (&new_predicate))
1152 optimized_out_size += e->size;
1153 else
1154 account_size_time (info, e->size, e->time, &new_predicate);
1157 /* Remap edge predicates with the same simplification as above.
1158 Also copy constantness arrays. */
1159 for (edge = dst->callees; edge; edge = edge->next_callee)
1161 struct predicate new_predicate;
1162 struct inline_edge_summary *es = inline_edge_summary (edge);
1164 if (!edge->inline_failed)
1165 inlined_to_p = true;
1166 if (!es->predicate)
1167 continue;
1168 new_predicate = remap_predicate_after_duplication (es->predicate,
1169 possible_truths,
1170 info);
1171 if (false_predicate_p (&new_predicate)
1172 && !false_predicate_p (es->predicate))
1174 optimized_out_size += es->call_stmt_size * INLINE_SIZE_SCALE;
1175 edge->frequency = 0;
1177 edge_set_predicate (edge, &new_predicate);
1180 /* Remap indirect edge predicates with the same simplificaiton as above.
1181 Also copy constantness arrays. */
1182 for (edge = dst->indirect_calls; edge; edge = edge->next_callee)
1184 struct predicate new_predicate;
1185 struct inline_edge_summary *es = inline_edge_summary (edge);
1187 gcc_checking_assert (edge->inline_failed);
1188 if (!es->predicate)
1189 continue;
1190 new_predicate = remap_predicate_after_duplication (es->predicate,
1191 possible_truths,
1192 info);
1193 if (false_predicate_p (&new_predicate)
1194 && !false_predicate_p (es->predicate))
1196 optimized_out_size += es->call_stmt_size * INLINE_SIZE_SCALE;
1197 edge->frequency = 0;
1199 edge_set_predicate (edge, &new_predicate);
1201 remap_hint_predicate_after_duplication (&info->loop_iterations,
1202 possible_truths, info);
1203 remap_hint_predicate_after_duplication (&info->loop_stride,
1204 possible_truths, info);
1205 remap_hint_predicate_after_duplication (&info->array_index,
1206 possible_truths, info);
1208 /* If inliner or someone after inliner will ever start producing
1209 non-trivial clones, we will get trouble with lack of information
1210 about updating self sizes, because size vectors already contains
1211 sizes of the calees. */
1212 gcc_assert (!inlined_to_p || !optimized_out_size);
1214 else
1216 info->entry = vec_safe_copy (info->entry);
1217 if (info->loop_iterations)
1219 predicate p = *info->loop_iterations;
1220 info->loop_iterations = NULL;
1221 set_hint_predicate (&info->loop_iterations, p);
1223 if (info->loop_stride)
1225 predicate p = *info->loop_stride;
1226 info->loop_stride = NULL;
1227 set_hint_predicate (&info->loop_stride, p);
1229 if (info->array_index)
1231 predicate p = *info->array_index;
1232 info->array_index = NULL;
1233 set_hint_predicate (&info->array_index, p);
1236 inline_update_overall_summary (dst);
1240 /* Hook that is called by cgraph.c when a node is duplicated. */
1242 static void
1243 inline_edge_duplication_hook (struct cgraph_edge *src,
1244 struct cgraph_edge *dst,
1245 ATTRIBUTE_UNUSED void *data)
1247 struct inline_edge_summary *info;
1248 struct inline_edge_summary *srcinfo;
1249 inline_summary_alloc ();
1250 info = inline_edge_summary (dst);
1251 srcinfo = inline_edge_summary (src);
1252 memcpy (info, srcinfo, sizeof (struct inline_edge_summary));
1253 info->predicate = NULL;
1254 edge_set_predicate (dst, srcinfo->predicate);
1255 info->param = srcinfo->param.copy ();
1259 /* Keep edge cache consistent across edge removal. */
1261 static void
1262 inline_edge_removal_hook (struct cgraph_edge *edge,
1263 void *data ATTRIBUTE_UNUSED)
1265 if (edge_growth_cache.exists ())
1266 reset_edge_growth_cache (edge);
1267 reset_inline_edge_summary (edge);
1271 /* Initialize growth caches. */
1273 void
1274 initialize_growth_caches (void)
1276 if (cgraph_edge_max_uid)
1277 edge_growth_cache.safe_grow_cleared (cgraph_edge_max_uid);
1278 if (cgraph_max_uid)
1279 node_growth_cache.safe_grow_cleared (cgraph_max_uid);
1283 /* Free growth caches. */
1285 void
1286 free_growth_caches (void)
1288 edge_growth_cache.release ();
1289 node_growth_cache.release ();
1293 /* Dump edge summaries associated to NODE and recursively to all clones.
1294 Indent by INDENT. */
1296 static void
1297 dump_inline_edge_summary (FILE *f, int indent, struct cgraph_node *node,
1298 struct inline_summary *info)
1300 struct cgraph_edge *edge;
1301 for (edge = node->callees; edge; edge = edge->next_callee)
1303 struct inline_edge_summary *es = inline_edge_summary (edge);
1304 struct cgraph_node *callee =
1305 cgraph_function_or_thunk_node (edge->callee, NULL);
1306 int i;
1308 fprintf (f,
1309 "%*s%s/%i %s\n%*s loop depth:%2i freq:%4i size:%2i"
1310 " time: %2i callee size:%2i stack:%2i",
1311 indent, "", callee->name (), callee->order,
1312 !edge->inline_failed
1313 ? "inlined" : cgraph_inline_failed_string (edge-> inline_failed),
1314 indent, "", es->loop_depth, edge->frequency,
1315 es->call_stmt_size, es->call_stmt_time,
1316 (int) inline_summary (callee)->size / INLINE_SIZE_SCALE,
1317 (int) inline_summary (callee)->estimated_stack_size);
1319 if (es->predicate)
1321 fprintf (f, " predicate: ");
1322 dump_predicate (f, info->conds, es->predicate);
1324 else
1325 fprintf (f, "\n");
1326 if (es->param.exists ())
1327 for (i = 0; i < (int) es->param.length (); i++)
1329 int prob = es->param[i].change_prob;
1331 if (!prob)
1332 fprintf (f, "%*s op%i is compile time invariant\n",
1333 indent + 2, "", i);
1334 else if (prob != REG_BR_PROB_BASE)
1335 fprintf (f, "%*s op%i change %f%% of time\n", indent + 2, "", i,
1336 prob * 100.0 / REG_BR_PROB_BASE);
1338 if (!edge->inline_failed)
1340 fprintf (f, "%*sStack frame offset %i, callee self size %i,"
1341 " callee size %i\n",
1342 indent + 2, "",
1343 (int) inline_summary (callee)->stack_frame_offset,
1344 (int) inline_summary (callee)->estimated_self_stack_size,
1345 (int) inline_summary (callee)->estimated_stack_size);
1346 dump_inline_edge_summary (f, indent + 2, callee, info);
1349 for (edge = node->indirect_calls; edge; edge = edge->next_callee)
1351 struct inline_edge_summary *es = inline_edge_summary (edge);
1352 fprintf (f, "%*sindirect call loop depth:%2i freq:%4i size:%2i"
1353 " time: %2i",
1354 indent, "",
1355 es->loop_depth,
1356 edge->frequency, es->call_stmt_size, es->call_stmt_time);
1357 if (es->predicate)
1359 fprintf (f, "predicate: ");
1360 dump_predicate (f, info->conds, es->predicate);
1362 else
1363 fprintf (f, "\n");
1368 void
1369 dump_inline_summary (FILE *f, struct cgraph_node *node)
1371 if (node->definition)
1373 struct inline_summary *s = inline_summary (node);
1374 size_time_entry *e;
1375 int i;
1376 fprintf (f, "Inline summary for %s/%i", node->name (),
1377 node->order);
1378 if (DECL_DISREGARD_INLINE_LIMITS (node->decl))
1379 fprintf (f, " always_inline");
1380 if (s->inlinable)
1381 fprintf (f, " inlinable");
1382 fprintf (f, "\n self time: %i\n", s->self_time);
1383 fprintf (f, " global time: %i\n", s->time);
1384 fprintf (f, " self size: %i\n", s->self_size);
1385 fprintf (f, " global size: %i\n", s->size);
1386 fprintf (f, " self stack: %i\n",
1387 (int) s->estimated_self_stack_size);
1388 fprintf (f, " global stack: %i\n", (int) s->estimated_stack_size);
1389 if (s->growth)
1390 fprintf (f, " estimated growth:%i\n", (int) s->growth);
1391 if (s->scc_no)
1392 fprintf (f, " In SCC: %i\n", (int) s->scc_no);
1393 for (i = 0; vec_safe_iterate (s->entry, i, &e); i++)
1395 fprintf (f, " size:%f, time:%f, predicate:",
1396 (double) e->size / INLINE_SIZE_SCALE,
1397 (double) e->time / INLINE_TIME_SCALE);
1398 dump_predicate (f, s->conds, &e->predicate);
1400 if (s->loop_iterations)
1402 fprintf (f, " loop iterations:");
1403 dump_predicate (f, s->conds, s->loop_iterations);
1405 if (s->loop_stride)
1407 fprintf (f, " loop stride:");
1408 dump_predicate (f, s->conds, s->loop_stride);
1410 if (s->array_index)
1412 fprintf (f, " array index:");
1413 dump_predicate (f, s->conds, s->array_index);
1415 fprintf (f, " calls:\n");
1416 dump_inline_edge_summary (f, 4, node, s);
1417 fprintf (f, "\n");
1421 DEBUG_FUNCTION void
1422 debug_inline_summary (struct cgraph_node *node)
1424 dump_inline_summary (stderr, node);
1427 void
1428 dump_inline_summaries (FILE *f)
1430 struct cgraph_node *node;
1432 FOR_EACH_DEFINED_FUNCTION (node)
1433 if (!node->global.inlined_to)
1434 dump_inline_summary (f, node);
1437 /* Give initial reasons why inlining would fail on EDGE. This gets either
1438 nullified or usually overwritten by more precise reasons later. */
1440 void
1441 initialize_inline_failed (struct cgraph_edge *e)
1443 struct cgraph_node *callee = e->callee;
1445 if (e->indirect_unknown_callee)
1446 e->inline_failed = CIF_INDIRECT_UNKNOWN_CALL;
1447 else if (!callee->definition)
1448 e->inline_failed = CIF_BODY_NOT_AVAILABLE;
1449 else if (callee->local.redefined_extern_inline)
1450 e->inline_failed = CIF_REDEFINED_EXTERN_INLINE;
1451 else if (e->call_stmt_cannot_inline_p)
1452 e->inline_failed = CIF_MISMATCHED_ARGUMENTS;
1453 else if (cfun && fn_contains_cilk_spawn_p (cfun))
1454 /* We can't inline if the function is spawing a function. */
1455 e->inline_failed = CIF_FUNCTION_NOT_INLINABLE;
1456 else
1457 e->inline_failed = CIF_FUNCTION_NOT_CONSIDERED;
1460 /* Callback of walk_aliased_vdefs. Flags that it has been invoked to the
1461 boolean variable pointed to by DATA. */
1463 static bool
1464 mark_modified (ao_ref *ao ATTRIBUTE_UNUSED, tree vdef ATTRIBUTE_UNUSED,
1465 void *data)
1467 bool *b = (bool *) data;
1468 *b = true;
1469 return true;
1472 /* If OP refers to value of function parameter, return the corresponding
1473 parameter. */
1475 static tree
1476 unmodified_parm_1 (gimple stmt, tree op)
1478 /* SSA_NAME referring to parm default def? */
1479 if (TREE_CODE (op) == SSA_NAME
1480 && SSA_NAME_IS_DEFAULT_DEF (op)
1481 && TREE_CODE (SSA_NAME_VAR (op)) == PARM_DECL)
1482 return SSA_NAME_VAR (op);
1483 /* Non-SSA parm reference? */
1484 if (TREE_CODE (op) == PARM_DECL)
1486 bool modified = false;
1488 ao_ref refd;
1489 ao_ref_init (&refd, op);
1490 walk_aliased_vdefs (&refd, gimple_vuse (stmt), mark_modified, &modified,
1491 NULL);
1492 if (!modified)
1493 return op;
1495 return NULL_TREE;
1498 /* If OP refers to value of function parameter, return the corresponding
1499 parameter. Also traverse chains of SSA register assignments. */
1501 static tree
1502 unmodified_parm (gimple stmt, tree op)
1504 tree res = unmodified_parm_1 (stmt, op);
1505 if (res)
1506 return res;
1508 if (TREE_CODE (op) == SSA_NAME
1509 && !SSA_NAME_IS_DEFAULT_DEF (op)
1510 && gimple_assign_single_p (SSA_NAME_DEF_STMT (op)))
1511 return unmodified_parm (SSA_NAME_DEF_STMT (op),
1512 gimple_assign_rhs1 (SSA_NAME_DEF_STMT (op)));
1513 return NULL_TREE;
1516 /* If OP refers to a value of a function parameter or value loaded from an
1517 aggregate passed to a parameter (either by value or reference), return TRUE
1518 and store the number of the parameter to *INDEX_P and information whether
1519 and how it has been loaded from an aggregate into *AGGPOS. INFO describes
1520 the function parameters, STMT is the statement in which OP is used or
1521 loaded. */
1523 static bool
1524 unmodified_parm_or_parm_agg_item (struct ipa_node_params *info,
1525 gimple stmt, tree op, int *index_p,
1526 struct agg_position_info *aggpos)
1528 tree res = unmodified_parm_1 (stmt, op);
1530 gcc_checking_assert (aggpos);
1531 if (res)
1533 *index_p = ipa_get_param_decl_index (info, res);
1534 if (*index_p < 0)
1535 return false;
1536 aggpos->agg_contents = false;
1537 aggpos->by_ref = false;
1538 return true;
1541 if (TREE_CODE (op) == SSA_NAME)
1543 if (SSA_NAME_IS_DEFAULT_DEF (op)
1544 || !gimple_assign_single_p (SSA_NAME_DEF_STMT (op)))
1545 return false;
1546 stmt = SSA_NAME_DEF_STMT (op);
1547 op = gimple_assign_rhs1 (stmt);
1548 if (!REFERENCE_CLASS_P (op))
1549 return unmodified_parm_or_parm_agg_item (info, stmt, op, index_p,
1550 aggpos);
1553 aggpos->agg_contents = true;
1554 return ipa_load_from_parm_agg (info, stmt, op, index_p, &aggpos->offset,
1555 &aggpos->by_ref);
1558 /* See if statement might disappear after inlining.
1559 0 - means not eliminated
1560 1 - half of statements goes away
1561 2 - for sure it is eliminated.
1562 We are not terribly sophisticated, basically looking for simple abstraction
1563 penalty wrappers. */
1565 static int
1566 eliminated_by_inlining_prob (gimple stmt)
1568 enum gimple_code code = gimple_code (stmt);
1569 enum tree_code rhs_code;
1571 if (!optimize)
1572 return 0;
1574 switch (code)
1576 case GIMPLE_RETURN:
1577 return 2;
1578 case GIMPLE_ASSIGN:
1579 if (gimple_num_ops (stmt) != 2)
1580 return 0;
1582 rhs_code = gimple_assign_rhs_code (stmt);
1584 /* Casts of parameters, loads from parameters passed by reference
1585 and stores to return value or parameters are often free after
1586 inlining dua to SRA and further combining.
1587 Assume that half of statements goes away. */
1588 if (rhs_code == CONVERT_EXPR
1589 || rhs_code == NOP_EXPR
1590 || rhs_code == VIEW_CONVERT_EXPR
1591 || rhs_code == ADDR_EXPR
1592 || gimple_assign_rhs_class (stmt) == GIMPLE_SINGLE_RHS)
1594 tree rhs = gimple_assign_rhs1 (stmt);
1595 tree lhs = gimple_assign_lhs (stmt);
1596 tree inner_rhs = get_base_address (rhs);
1597 tree inner_lhs = get_base_address (lhs);
1598 bool rhs_free = false;
1599 bool lhs_free = false;
1601 if (!inner_rhs)
1602 inner_rhs = rhs;
1603 if (!inner_lhs)
1604 inner_lhs = lhs;
1606 /* Reads of parameter are expected to be free. */
1607 if (unmodified_parm (stmt, inner_rhs))
1608 rhs_free = true;
1609 /* Match expressions of form &this->field. Those will most likely
1610 combine with something upstream after inlining. */
1611 else if (TREE_CODE (inner_rhs) == ADDR_EXPR)
1613 tree op = get_base_address (TREE_OPERAND (inner_rhs, 0));
1614 if (TREE_CODE (op) == PARM_DECL)
1615 rhs_free = true;
1616 else if (TREE_CODE (op) == MEM_REF
1617 && unmodified_parm (stmt, TREE_OPERAND (op, 0)))
1618 rhs_free = true;
1621 /* When parameter is not SSA register because its address is taken
1622 and it is just copied into one, the statement will be completely
1623 free after inlining (we will copy propagate backward). */
1624 if (rhs_free && is_gimple_reg (lhs))
1625 return 2;
1627 /* Reads of parameters passed by reference
1628 expected to be free (i.e. optimized out after inlining). */
1629 if (TREE_CODE (inner_rhs) == MEM_REF
1630 && unmodified_parm (stmt, TREE_OPERAND (inner_rhs, 0)))
1631 rhs_free = true;
1633 /* Copying parameter passed by reference into gimple register is
1634 probably also going to copy propagate, but we can't be quite
1635 sure. */
1636 if (rhs_free && is_gimple_reg (lhs))
1637 lhs_free = true;
1639 /* Writes to parameters, parameters passed by value and return value
1640 (either dirrectly or passed via invisible reference) are free.
1642 TODO: We ought to handle testcase like
1643 struct a {int a,b;};
1644 struct a
1645 retrurnsturct (void)
1647 struct a a ={1,2};
1648 return a;
1651 This translate into:
1653 retrurnsturct ()
1655 int a$b;
1656 int a$a;
1657 struct a a;
1658 struct a D.2739;
1660 <bb 2>:
1661 D.2739.a = 1;
1662 D.2739.b = 2;
1663 return D.2739;
1666 For that we either need to copy ipa-split logic detecting writes
1667 to return value. */
1668 if (TREE_CODE (inner_lhs) == PARM_DECL
1669 || TREE_CODE (inner_lhs) == RESULT_DECL
1670 || (TREE_CODE (inner_lhs) == MEM_REF
1671 && (unmodified_parm (stmt, TREE_OPERAND (inner_lhs, 0))
1672 || (TREE_CODE (TREE_OPERAND (inner_lhs, 0)) == SSA_NAME
1673 && SSA_NAME_VAR (TREE_OPERAND (inner_lhs, 0))
1674 && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND
1675 (inner_lhs,
1676 0))) == RESULT_DECL))))
1677 lhs_free = true;
1678 if (lhs_free
1679 && (is_gimple_reg (rhs) || is_gimple_min_invariant (rhs)))
1680 rhs_free = true;
1681 if (lhs_free && rhs_free)
1682 return 1;
1684 return 0;
1685 default:
1686 return 0;
1691 /* If BB ends by a conditional we can turn into predicates, attach corresponding
1692 predicates to the CFG edges. */
1694 static void
1695 set_cond_stmt_execution_predicate (struct ipa_node_params *info,
1696 struct inline_summary *summary,
1697 basic_block bb)
1699 gimple last;
1700 tree op;
1701 int index;
1702 struct agg_position_info aggpos;
1703 enum tree_code code, inverted_code;
1704 edge e;
1705 edge_iterator ei;
1706 gimple set_stmt;
1707 tree op2;
1709 last = last_stmt (bb);
1710 if (!last || gimple_code (last) != GIMPLE_COND)
1711 return;
1712 if (!is_gimple_ip_invariant (gimple_cond_rhs (last)))
1713 return;
1714 op = gimple_cond_lhs (last);
1715 /* TODO: handle conditionals like
1716 var = op0 < 4;
1717 if (var != 0). */
1718 if (unmodified_parm_or_parm_agg_item (info, last, op, &index, &aggpos))
1720 code = gimple_cond_code (last);
1721 inverted_code
1722 = invert_tree_comparison (code,
1723 HONOR_NANS (TYPE_MODE (TREE_TYPE (op))));
1725 FOR_EACH_EDGE (e, ei, bb->succs)
1727 struct predicate p = add_condition (summary, index, &aggpos,
1728 e->flags & EDGE_TRUE_VALUE
1729 ? code : inverted_code,
1730 gimple_cond_rhs (last));
1731 e->aux = pool_alloc (edge_predicate_pool);
1732 *(struct predicate *) e->aux = p;
1736 if (TREE_CODE (op) != SSA_NAME)
1737 return;
1738 /* Special case
1739 if (builtin_constant_p (op))
1740 constant_code
1741 else
1742 nonconstant_code.
1743 Here we can predicate nonconstant_code. We can't
1744 really handle constant_code since we have no predicate
1745 for this and also the constant code is not known to be
1746 optimized away when inliner doen't see operand is constant.
1747 Other optimizers might think otherwise. */
1748 if (gimple_cond_code (last) != NE_EXPR
1749 || !integer_zerop (gimple_cond_rhs (last)))
1750 return;
1751 set_stmt = SSA_NAME_DEF_STMT (op);
1752 if (!gimple_call_builtin_p (set_stmt, BUILT_IN_CONSTANT_P)
1753 || gimple_call_num_args (set_stmt) != 1)
1754 return;
1755 op2 = gimple_call_arg (set_stmt, 0);
1756 if (!unmodified_parm_or_parm_agg_item
1757 (info, set_stmt, op2, &index, &aggpos))
1758 return;
1759 FOR_EACH_EDGE (e, ei, bb->succs) if (e->flags & EDGE_FALSE_VALUE)
1761 struct predicate p = add_condition (summary, index, &aggpos,
1762 IS_NOT_CONSTANT, NULL_TREE);
1763 e->aux = pool_alloc (edge_predicate_pool);
1764 *(struct predicate *) e->aux = p;
1769 /* If BB ends by a switch we can turn into predicates, attach corresponding
1770 predicates to the CFG edges. */
1772 static void
1773 set_switch_stmt_execution_predicate (struct ipa_node_params *info,
1774 struct inline_summary *summary,
1775 basic_block bb)
1777 gimple last;
1778 tree op;
1779 int index;
1780 struct agg_position_info aggpos;
1781 edge e;
1782 edge_iterator ei;
1783 size_t n;
1784 size_t case_idx;
1786 last = last_stmt (bb);
1787 if (!last || gimple_code (last) != GIMPLE_SWITCH)
1788 return;
1789 op = gimple_switch_index (last);
1790 if (!unmodified_parm_or_parm_agg_item (info, last, op, &index, &aggpos))
1791 return;
1793 FOR_EACH_EDGE (e, ei, bb->succs)
1795 e->aux = pool_alloc (edge_predicate_pool);
1796 *(struct predicate *) e->aux = false_predicate ();
1798 n = gimple_switch_num_labels (last);
1799 for (case_idx = 0; case_idx < n; ++case_idx)
1801 tree cl = gimple_switch_label (last, case_idx);
1802 tree min, max;
1803 struct predicate p;
1805 e = find_edge (bb, label_to_block (CASE_LABEL (cl)));
1806 min = CASE_LOW (cl);
1807 max = CASE_HIGH (cl);
1809 /* For default we might want to construct predicate that none
1810 of cases is met, but it is bit hard to do not having negations
1811 of conditionals handy. */
1812 if (!min && !max)
1813 p = true_predicate ();
1814 else if (!max)
1815 p = add_condition (summary, index, &aggpos, EQ_EXPR, min);
1816 else
1818 struct predicate p1, p2;
1819 p1 = add_condition (summary, index, &aggpos, GE_EXPR, min);
1820 p2 = add_condition (summary, index, &aggpos, LE_EXPR, max);
1821 p = and_predicates (summary->conds, &p1, &p2);
1823 *(struct predicate *) e->aux
1824 = or_predicates (summary->conds, &p, (struct predicate *) e->aux);
1829 /* For each BB in NODE attach to its AUX pointer predicate under
1830 which it is executable. */
1832 static void
1833 compute_bb_predicates (struct cgraph_node *node,
1834 struct ipa_node_params *parms_info,
1835 struct inline_summary *summary)
1837 struct function *my_function = DECL_STRUCT_FUNCTION (node->decl);
1838 bool done = false;
1839 basic_block bb;
1841 FOR_EACH_BB_FN (bb, my_function)
1843 set_cond_stmt_execution_predicate (parms_info, summary, bb);
1844 set_switch_stmt_execution_predicate (parms_info, summary, bb);
1847 /* Entry block is always executable. */
1848 ENTRY_BLOCK_PTR_FOR_FN (my_function)->aux
1849 = pool_alloc (edge_predicate_pool);
1850 *(struct predicate *) ENTRY_BLOCK_PTR_FOR_FN (my_function)->aux
1851 = true_predicate ();
1853 /* A simple dataflow propagation of predicates forward in the CFG.
1854 TODO: work in reverse postorder. */
1855 while (!done)
1857 done = true;
1858 FOR_EACH_BB_FN (bb, my_function)
1860 struct predicate p = false_predicate ();
1861 edge e;
1862 edge_iterator ei;
1863 FOR_EACH_EDGE (e, ei, bb->preds)
1865 if (e->src->aux)
1867 struct predicate this_bb_predicate
1868 = *(struct predicate *) e->src->aux;
1869 if (e->aux)
1870 this_bb_predicate
1871 = and_predicates (summary->conds, &this_bb_predicate,
1872 (struct predicate *) e->aux);
1873 p = or_predicates (summary->conds, &p, &this_bb_predicate);
1874 if (true_predicate_p (&p))
1875 break;
1878 if (false_predicate_p (&p))
1879 gcc_assert (!bb->aux);
1880 else
1882 if (!bb->aux)
1884 done = false;
1885 bb->aux = pool_alloc (edge_predicate_pool);
1886 *((struct predicate *) bb->aux) = p;
1888 else if (!predicates_equal_p (&p, (struct predicate *) bb->aux))
1890 /* This OR operation is needed to ensure monotonous data flow
1891 in the case we hit the limit on number of clauses and the
1892 and/or operations above give approximate answers. */
1893 p = or_predicates (summary->conds, &p, (struct predicate *)bb->aux);
1894 if (!predicates_equal_p (&p, (struct predicate *) bb->aux))
1896 done = false;
1897 *((struct predicate *) bb->aux) = p;
1906 /* We keep info about constantness of SSA names. */
1908 typedef struct predicate predicate_t;
1909 /* Return predicate specifying when the STMT might have result that is not
1910 a compile time constant. */
1912 static struct predicate
1913 will_be_nonconstant_expr_predicate (struct ipa_node_params *info,
1914 struct inline_summary *summary,
1915 tree expr,
1916 vec<predicate_t> nonconstant_names)
1918 tree parm;
1919 int index;
1921 while (UNARY_CLASS_P (expr))
1922 expr = TREE_OPERAND (expr, 0);
1924 parm = unmodified_parm (NULL, expr);
1925 if (parm && (index = ipa_get_param_decl_index (info, parm)) >= 0)
1926 return add_condition (summary, index, NULL, CHANGED, NULL_TREE);
1927 if (is_gimple_min_invariant (expr))
1928 return false_predicate ();
1929 if (TREE_CODE (expr) == SSA_NAME)
1930 return nonconstant_names[SSA_NAME_VERSION (expr)];
1931 if (BINARY_CLASS_P (expr) || COMPARISON_CLASS_P (expr))
1933 struct predicate p1 = will_be_nonconstant_expr_predicate
1934 (info, summary, TREE_OPERAND (expr, 0),
1935 nonconstant_names);
1936 struct predicate p2;
1937 if (true_predicate_p (&p1))
1938 return p1;
1939 p2 = will_be_nonconstant_expr_predicate (info, summary,
1940 TREE_OPERAND (expr, 1),
1941 nonconstant_names);
1942 return or_predicates (summary->conds, &p1, &p2);
1944 else if (TREE_CODE (expr) == COND_EXPR)
1946 struct predicate p1 = will_be_nonconstant_expr_predicate
1947 (info, summary, TREE_OPERAND (expr, 0),
1948 nonconstant_names);
1949 struct predicate p2;
1950 if (true_predicate_p (&p1))
1951 return p1;
1952 p2 = will_be_nonconstant_expr_predicate (info, summary,
1953 TREE_OPERAND (expr, 1),
1954 nonconstant_names);
1955 if (true_predicate_p (&p2))
1956 return p2;
1957 p1 = or_predicates (summary->conds, &p1, &p2);
1958 p2 = will_be_nonconstant_expr_predicate (info, summary,
1959 TREE_OPERAND (expr, 2),
1960 nonconstant_names);
1961 return or_predicates (summary->conds, &p1, &p2);
1963 else
1965 debug_tree (expr);
1966 gcc_unreachable ();
1968 return false_predicate ();
1972 /* Return predicate specifying when the STMT might have result that is not
1973 a compile time constant. */
1975 static struct predicate
1976 will_be_nonconstant_predicate (struct ipa_node_params *info,
1977 struct inline_summary *summary,
1978 gimple stmt,
1979 vec<predicate_t> nonconstant_names)
1981 struct predicate p = true_predicate ();
1982 ssa_op_iter iter;
1983 tree use;
1984 struct predicate op_non_const;
1985 bool is_load;
1986 int base_index;
1987 struct agg_position_info aggpos;
1989 /* What statments might be optimized away
1990 when their arguments are constant
1991 TODO: also trivial builtins.
1992 builtin_constant_p is already handled later. */
1993 if (gimple_code (stmt) != GIMPLE_ASSIGN
1994 && gimple_code (stmt) != GIMPLE_COND
1995 && gimple_code (stmt) != GIMPLE_SWITCH)
1996 return p;
1998 /* Stores will stay anyway. */
1999 if (gimple_store_p (stmt))
2000 return p;
2002 is_load = gimple_assign_load_p (stmt);
2004 /* Loads can be optimized when the value is known. */
2005 if (is_load)
2007 tree op;
2008 gcc_assert (gimple_assign_single_p (stmt));
2009 op = gimple_assign_rhs1 (stmt);
2010 if (!unmodified_parm_or_parm_agg_item (info, stmt, op, &base_index,
2011 &aggpos))
2012 return p;
2014 else
2015 base_index = -1;
2017 /* See if we understand all operands before we start
2018 adding conditionals. */
2019 FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE)
2021 tree parm = unmodified_parm (stmt, use);
2022 /* For arguments we can build a condition. */
2023 if (parm && ipa_get_param_decl_index (info, parm) >= 0)
2024 continue;
2025 if (TREE_CODE (use) != SSA_NAME)
2026 return p;
2027 /* If we know when operand is constant,
2028 we still can say something useful. */
2029 if (!true_predicate_p (&nonconstant_names[SSA_NAME_VERSION (use)]))
2030 continue;
2031 return p;
2034 if (is_load)
2035 op_non_const =
2036 add_condition (summary, base_index, &aggpos, CHANGED, NULL);
2037 else
2038 op_non_const = false_predicate ();
2039 FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE)
2041 tree parm = unmodified_parm (stmt, use);
2042 int index;
2044 if (parm && (index = ipa_get_param_decl_index (info, parm)) >= 0)
2046 if (index != base_index)
2047 p = add_condition (summary, index, NULL, CHANGED, NULL_TREE);
2048 else
2049 continue;
2051 else
2052 p = nonconstant_names[SSA_NAME_VERSION (use)];
2053 op_non_const = or_predicates (summary->conds, &p, &op_non_const);
2055 if (gimple_code (stmt) == GIMPLE_ASSIGN
2056 && TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME)
2057 nonconstant_names[SSA_NAME_VERSION (gimple_assign_lhs (stmt))]
2058 = op_non_const;
2059 return op_non_const;
2062 struct record_modified_bb_info
2064 bitmap bb_set;
2065 gimple stmt;
2068 /* Callback of walk_aliased_vdefs. Records basic blocks where the value may be
2069 set except for info->stmt. */
2071 static bool
2072 record_modified (ao_ref *ao ATTRIBUTE_UNUSED, tree vdef, void *data)
2074 struct record_modified_bb_info *info =
2075 (struct record_modified_bb_info *) data;
2076 if (SSA_NAME_DEF_STMT (vdef) == info->stmt)
2077 return false;
2078 bitmap_set_bit (info->bb_set,
2079 SSA_NAME_IS_DEFAULT_DEF (vdef)
2080 ? ENTRY_BLOCK_PTR_FOR_FN (cfun)->index
2081 : gimple_bb (SSA_NAME_DEF_STMT (vdef))->index);
2082 return false;
2085 /* Return probability (based on REG_BR_PROB_BASE) that I-th parameter of STMT
2086 will change since last invocation of STMT.
2088 Value 0 is reserved for compile time invariants.
2089 For common parameters it is REG_BR_PROB_BASE. For loop invariants it
2090 ought to be REG_BR_PROB_BASE / estimated_iters. */
2092 static int
2093 param_change_prob (gimple stmt, int i)
2095 tree op = gimple_call_arg (stmt, i);
2096 basic_block bb = gimple_bb (stmt);
2097 tree base;
2099 /* Global invariants neve change. */
2100 if (is_gimple_min_invariant (op))
2101 return 0;
2102 /* We would have to do non-trivial analysis to really work out what
2103 is the probability of value to change (i.e. when init statement
2104 is in a sibling loop of the call).
2106 We do an conservative estimate: when call is executed N times more often
2107 than the statement defining value, we take the frequency 1/N. */
2108 if (TREE_CODE (op) == SSA_NAME)
2110 int init_freq;
2112 if (!bb->frequency)
2113 return REG_BR_PROB_BASE;
2115 if (SSA_NAME_IS_DEFAULT_DEF (op))
2116 init_freq = ENTRY_BLOCK_PTR_FOR_FN (cfun)->frequency;
2117 else
2118 init_freq = gimple_bb (SSA_NAME_DEF_STMT (op))->frequency;
2120 if (!init_freq)
2121 init_freq = 1;
2122 if (init_freq < bb->frequency)
2123 return MAX (GCOV_COMPUTE_SCALE (init_freq, bb->frequency), 1);
2124 else
2125 return REG_BR_PROB_BASE;
2128 base = get_base_address (op);
2129 if (base)
2131 ao_ref refd;
2132 int max;
2133 struct record_modified_bb_info info;
2134 bitmap_iterator bi;
2135 unsigned index;
2136 tree init = ctor_for_folding (base);
2138 if (init != error_mark_node)
2139 return 0;
2140 if (!bb->frequency)
2141 return REG_BR_PROB_BASE;
2142 ao_ref_init (&refd, op);
2143 info.stmt = stmt;
2144 info.bb_set = BITMAP_ALLOC (NULL);
2145 walk_aliased_vdefs (&refd, gimple_vuse (stmt), record_modified, &info,
2146 NULL);
2147 if (bitmap_bit_p (info.bb_set, bb->index))
2149 BITMAP_FREE (info.bb_set);
2150 return REG_BR_PROB_BASE;
2153 /* Assume that every memory is initialized at entry.
2154 TODO: Can we easilly determine if value is always defined
2155 and thus we may skip entry block? */
2156 if (ENTRY_BLOCK_PTR_FOR_FN (cfun)->frequency)
2157 max = ENTRY_BLOCK_PTR_FOR_FN (cfun)->frequency;
2158 else
2159 max = 1;
2161 EXECUTE_IF_SET_IN_BITMAP (info.bb_set, 0, index, bi)
2162 max = MIN (max, BASIC_BLOCK_FOR_FN (cfun, index)->frequency);
2164 BITMAP_FREE (info.bb_set);
2165 if (max < bb->frequency)
2166 return MAX (GCOV_COMPUTE_SCALE (max, bb->frequency), 1);
2167 else
2168 return REG_BR_PROB_BASE;
2170 return REG_BR_PROB_BASE;
2173 /* Find whether a basic block BB is the final block of a (half) diamond CFG
2174 sub-graph and if the predicate the condition depends on is known. If so,
2175 return true and store the pointer the predicate in *P. */
2177 static bool
2178 phi_result_unknown_predicate (struct ipa_node_params *info,
2179 struct inline_summary *summary, basic_block bb,
2180 struct predicate *p,
2181 vec<predicate_t> nonconstant_names)
2183 edge e;
2184 edge_iterator ei;
2185 basic_block first_bb = NULL;
2186 gimple stmt;
2188 if (single_pred_p (bb))
2190 *p = false_predicate ();
2191 return true;
2194 FOR_EACH_EDGE (e, ei, bb->preds)
2196 if (single_succ_p (e->src))
2198 if (!single_pred_p (e->src))
2199 return false;
2200 if (!first_bb)
2201 first_bb = single_pred (e->src);
2202 else if (single_pred (e->src) != first_bb)
2203 return false;
2205 else
2207 if (!first_bb)
2208 first_bb = e->src;
2209 else if (e->src != first_bb)
2210 return false;
2214 if (!first_bb)
2215 return false;
2217 stmt = last_stmt (first_bb);
2218 if (!stmt
2219 || gimple_code (stmt) != GIMPLE_COND
2220 || !is_gimple_ip_invariant (gimple_cond_rhs (stmt)))
2221 return false;
2223 *p = will_be_nonconstant_expr_predicate (info, summary,
2224 gimple_cond_lhs (stmt),
2225 nonconstant_names);
2226 if (true_predicate_p (p))
2227 return false;
2228 else
2229 return true;
2232 /* Given a PHI statement in a function described by inline properties SUMMARY
2233 and *P being the predicate describing whether the selected PHI argument is
2234 known, store a predicate for the result of the PHI statement into
2235 NONCONSTANT_NAMES, if possible. */
2237 static void
2238 predicate_for_phi_result (struct inline_summary *summary, gimple phi,
2239 struct predicate *p,
2240 vec<predicate_t> nonconstant_names)
2242 unsigned i;
2244 for (i = 0; i < gimple_phi_num_args (phi); i++)
2246 tree arg = gimple_phi_arg (phi, i)->def;
2247 if (!is_gimple_min_invariant (arg))
2249 gcc_assert (TREE_CODE (arg) == SSA_NAME);
2250 *p = or_predicates (summary->conds, p,
2251 &nonconstant_names[SSA_NAME_VERSION (arg)]);
2252 if (true_predicate_p (p))
2253 return;
2257 if (dump_file && (dump_flags & TDF_DETAILS))
2259 fprintf (dump_file, "\t\tphi predicate: ");
2260 dump_predicate (dump_file, summary->conds, p);
2262 nonconstant_names[SSA_NAME_VERSION (gimple_phi_result (phi))] = *p;
2265 /* Return predicate specifying when array index in access OP becomes non-constant. */
2267 static struct predicate
2268 array_index_predicate (struct inline_summary *info,
2269 vec< predicate_t> nonconstant_names, tree op)
2271 struct predicate p = false_predicate ();
2272 while (handled_component_p (op))
2274 if (TREE_CODE (op) == ARRAY_REF || TREE_CODE (op) == ARRAY_RANGE_REF)
2276 if (TREE_CODE (TREE_OPERAND (op, 1)) == SSA_NAME)
2277 p = or_predicates (info->conds, &p,
2278 &nonconstant_names[SSA_NAME_VERSION
2279 (TREE_OPERAND (op, 1))]);
2281 op = TREE_OPERAND (op, 0);
2283 return p;
2286 /* For a typical usage of __builtin_expect (a<b, 1), we
2287 may introduce an extra relation stmt:
2288 With the builtin, we have
2289 t1 = a <= b;
2290 t2 = (long int) t1;
2291 t3 = __builtin_expect (t2, 1);
2292 if (t3 != 0)
2293 goto ...
2294 Without the builtin, we have
2295 if (a<=b)
2296 goto...
2297 This affects the size/time estimation and may have
2298 an impact on the earlier inlining.
2299 Here find this pattern and fix it up later. */
2301 static gimple
2302 find_foldable_builtin_expect (basic_block bb)
2304 gimple_stmt_iterator bsi;
2306 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
2308 gimple stmt = gsi_stmt (bsi);
2309 if (gimple_call_builtin_p (stmt, BUILT_IN_EXPECT))
2311 tree var = gimple_call_lhs (stmt);
2312 tree arg = gimple_call_arg (stmt, 0);
2313 use_operand_p use_p;
2314 gimple use_stmt;
2315 bool match = false;
2316 bool done = false;
2318 if (!var || !arg)
2319 continue;
2320 gcc_assert (TREE_CODE (var) == SSA_NAME);
2322 while (TREE_CODE (arg) == SSA_NAME)
2324 gimple stmt_tmp = SSA_NAME_DEF_STMT (arg);
2325 if (!is_gimple_assign (stmt_tmp))
2326 break;
2327 switch (gimple_assign_rhs_code (stmt_tmp))
2329 case LT_EXPR:
2330 case LE_EXPR:
2331 case GT_EXPR:
2332 case GE_EXPR:
2333 case EQ_EXPR:
2334 case NE_EXPR:
2335 match = true;
2336 done = true;
2337 break;
2338 case NOP_EXPR:
2339 break;
2340 default:
2341 done = true;
2342 break;
2344 if (done)
2345 break;
2346 arg = gimple_assign_rhs1 (stmt_tmp);
2349 if (match && single_imm_use (var, &use_p, &use_stmt)
2350 && gimple_code (use_stmt) == GIMPLE_COND)
2351 return use_stmt;
2354 return NULL;
2357 /* Return true when the basic blocks contains only clobbers followed by RESX.
2358 Such BBs are kept around to make removal of dead stores possible with
2359 presence of EH and will be optimized out by optimize_clobbers later in the
2360 game.
2362 NEED_EH is used to recurse in case the clobber has non-EH predecestors
2363 that can be clobber only, too.. When it is false, the RESX is not necessary
2364 on the end of basic block. */
2366 static bool
2367 clobber_only_eh_bb_p (basic_block bb, bool need_eh = true)
2369 gimple_stmt_iterator gsi = gsi_last_bb (bb);
2370 edge_iterator ei;
2371 edge e;
2373 if (need_eh)
2375 if (gsi_end_p (gsi))
2376 return false;
2377 if (gimple_code (gsi_stmt (gsi)) != GIMPLE_RESX)
2378 return false;
2379 gsi_prev (&gsi);
2381 else if (!single_succ_p (bb))
2382 return false;
2384 for (; !gsi_end_p (gsi); gsi_prev (&gsi))
2386 gimple stmt = gsi_stmt (gsi);
2387 if (is_gimple_debug (stmt))
2388 continue;
2389 if (gimple_clobber_p (stmt))
2390 continue;
2391 if (gimple_code (stmt) == GIMPLE_LABEL)
2392 break;
2393 return false;
2396 /* See if all predecestors are either throws or clobber only BBs. */
2397 FOR_EACH_EDGE (e, ei, bb->preds)
2398 if (!(e->flags & EDGE_EH)
2399 && !clobber_only_eh_bb_p (e->src, false))
2400 return false;
2402 return true;
2405 /* Compute function body size parameters for NODE.
2406 When EARLY is true, we compute only simple summaries without
2407 non-trivial predicates to drive the early inliner. */
2409 static void
2410 estimate_function_body_sizes (struct cgraph_node *node, bool early)
2412 gcov_type time = 0;
2413 /* Estimate static overhead for function prologue/epilogue and alignment. */
2414 int size = 2;
2415 /* Benefits are scaled by probability of elimination that is in range
2416 <0,2>. */
2417 basic_block bb;
2418 gimple_stmt_iterator bsi;
2419 struct function *my_function = DECL_STRUCT_FUNCTION (node->decl);
2420 int freq;
2421 struct inline_summary *info = inline_summary (node);
2422 struct predicate bb_predicate;
2423 struct ipa_node_params *parms_info = NULL;
2424 vec<predicate_t> nonconstant_names = vNULL;
2425 int nblocks, n;
2426 int *order;
2427 predicate array_index = true_predicate ();
2428 gimple fix_builtin_expect_stmt;
2430 info->conds = NULL;
2431 info->entry = NULL;
2433 if (optimize && !early)
2435 calculate_dominance_info (CDI_DOMINATORS);
2436 loop_optimizer_init (LOOPS_NORMAL | LOOPS_HAVE_RECORDED_EXITS);
2438 if (ipa_node_params_vector.exists ())
2440 parms_info = IPA_NODE_REF (node);
2441 nonconstant_names.safe_grow_cleared
2442 (SSANAMES (my_function)->length ());
2446 if (dump_file)
2447 fprintf (dump_file, "\nAnalyzing function body size: %s\n",
2448 node->name ());
2450 /* When we run into maximal number of entries, we assign everything to the
2451 constant truth case. Be sure to have it in list. */
2452 bb_predicate = true_predicate ();
2453 account_size_time (info, 0, 0, &bb_predicate);
2455 bb_predicate = not_inlined_predicate ();
2456 account_size_time (info, 2 * INLINE_SIZE_SCALE, 0, &bb_predicate);
2458 gcc_assert (my_function && my_function->cfg);
2459 if (parms_info)
2460 compute_bb_predicates (node, parms_info, info);
2461 gcc_assert (cfun == my_function);
2462 order = XNEWVEC (int, n_basic_blocks_for_fn (cfun));
2463 nblocks = pre_and_rev_post_order_compute (NULL, order, false);
2464 for (n = 0; n < nblocks; n++)
2466 bb = BASIC_BLOCK_FOR_FN (cfun, order[n]);
2467 freq = compute_call_stmt_bb_frequency (node->decl, bb);
2468 if (clobber_only_eh_bb_p (bb))
2470 if (dump_file && (dump_flags & TDF_DETAILS))
2471 fprintf (dump_file, "\n Ignoring BB %i;"
2472 " it will be optimized away by cleanup_clobbers\n",
2473 bb->index);
2474 continue;
2477 /* TODO: Obviously predicates can be propagated down across CFG. */
2478 if (parms_info)
2480 if (bb->aux)
2481 bb_predicate = *(struct predicate *) bb->aux;
2482 else
2483 bb_predicate = false_predicate ();
2485 else
2486 bb_predicate = true_predicate ();
2488 if (dump_file && (dump_flags & TDF_DETAILS))
2490 fprintf (dump_file, "\n BB %i predicate:", bb->index);
2491 dump_predicate (dump_file, info->conds, &bb_predicate);
2494 if (parms_info && nonconstant_names.exists ())
2496 struct predicate phi_predicate;
2497 bool first_phi = true;
2499 for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi))
2501 if (first_phi
2502 && !phi_result_unknown_predicate (parms_info, info, bb,
2503 &phi_predicate,
2504 nonconstant_names))
2505 break;
2506 first_phi = false;
2507 if (dump_file && (dump_flags & TDF_DETAILS))
2509 fprintf (dump_file, " ");
2510 print_gimple_stmt (dump_file, gsi_stmt (bsi), 0, 0);
2512 predicate_for_phi_result (info, gsi_stmt (bsi), &phi_predicate,
2513 nonconstant_names);
2517 fix_builtin_expect_stmt = find_foldable_builtin_expect (bb);
2519 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
2521 gimple stmt = gsi_stmt (bsi);
2522 int this_size = estimate_num_insns (stmt, &eni_size_weights);
2523 int this_time = estimate_num_insns (stmt, &eni_time_weights);
2524 int prob;
2525 struct predicate will_be_nonconstant;
2527 /* This relation stmt should be folded after we remove
2528 buildin_expect call. Adjust the cost here. */
2529 if (stmt == fix_builtin_expect_stmt)
2531 this_size--;
2532 this_time--;
2535 if (dump_file && (dump_flags & TDF_DETAILS))
2537 fprintf (dump_file, " ");
2538 print_gimple_stmt (dump_file, stmt, 0, 0);
2539 fprintf (dump_file, "\t\tfreq:%3.2f size:%3i time:%3i\n",
2540 ((double) freq) / CGRAPH_FREQ_BASE, this_size,
2541 this_time);
2544 if (gimple_assign_load_p (stmt) && nonconstant_names.exists ())
2546 struct predicate this_array_index;
2547 this_array_index =
2548 array_index_predicate (info, nonconstant_names,
2549 gimple_assign_rhs1 (stmt));
2550 if (!false_predicate_p (&this_array_index))
2551 array_index =
2552 and_predicates (info->conds, &array_index,
2553 &this_array_index);
2555 if (gimple_store_p (stmt) && nonconstant_names.exists ())
2557 struct predicate this_array_index;
2558 this_array_index =
2559 array_index_predicate (info, nonconstant_names,
2560 gimple_get_lhs (stmt));
2561 if (!false_predicate_p (&this_array_index))
2562 array_index =
2563 and_predicates (info->conds, &array_index,
2564 &this_array_index);
2568 if (is_gimple_call (stmt)
2569 && !gimple_call_internal_p (stmt))
2571 struct cgraph_edge *edge = cgraph_edge (node, stmt);
2572 struct inline_edge_summary *es = inline_edge_summary (edge);
2574 /* Special case: results of BUILT_IN_CONSTANT_P will be always
2575 resolved as constant. We however don't want to optimize
2576 out the cgraph edges. */
2577 if (nonconstant_names.exists ()
2578 && gimple_call_builtin_p (stmt, BUILT_IN_CONSTANT_P)
2579 && gimple_call_lhs (stmt)
2580 && TREE_CODE (gimple_call_lhs (stmt)) == SSA_NAME)
2582 struct predicate false_p = false_predicate ();
2583 nonconstant_names[SSA_NAME_VERSION (gimple_call_lhs (stmt))]
2584 = false_p;
2586 if (ipa_node_params_vector.exists ())
2588 int count = gimple_call_num_args (stmt);
2589 int i;
2591 if (count)
2592 es->param.safe_grow_cleared (count);
2593 for (i = 0; i < count; i++)
2595 int prob = param_change_prob (stmt, i);
2596 gcc_assert (prob >= 0 && prob <= REG_BR_PROB_BASE);
2597 es->param[i].change_prob = prob;
2601 es->call_stmt_size = this_size;
2602 es->call_stmt_time = this_time;
2603 es->loop_depth = bb_loop_depth (bb);
2604 edge_set_predicate (edge, &bb_predicate);
2607 /* TODO: When conditional jump or swithc is known to be constant, but
2608 we did not translate it into the predicates, we really can account
2609 just maximum of the possible paths. */
2610 if (parms_info)
2611 will_be_nonconstant
2612 = will_be_nonconstant_predicate (parms_info, info,
2613 stmt, nonconstant_names);
2614 if (this_time || this_size)
2616 struct predicate p;
2618 this_time *= freq;
2620 prob = eliminated_by_inlining_prob (stmt);
2621 if (prob == 1 && dump_file && (dump_flags & TDF_DETAILS))
2622 fprintf (dump_file,
2623 "\t\t50%% will be eliminated by inlining\n");
2624 if (prob == 2 && dump_file && (dump_flags & TDF_DETAILS))
2625 fprintf (dump_file, "\t\tWill be eliminated by inlining\n");
2627 if (parms_info)
2628 p = and_predicates (info->conds, &bb_predicate,
2629 &will_be_nonconstant);
2630 else
2631 p = true_predicate ();
2633 if (!false_predicate_p (&p))
2635 time += this_time;
2636 size += this_size;
2637 if (time > MAX_TIME * INLINE_TIME_SCALE)
2638 time = MAX_TIME * INLINE_TIME_SCALE;
2641 /* We account everything but the calls. Calls have their own
2642 size/time info attached to cgraph edges. This is necessary
2643 in order to make the cost disappear after inlining. */
2644 if (!is_gimple_call (stmt))
2646 if (prob)
2648 struct predicate ip = not_inlined_predicate ();
2649 ip = and_predicates (info->conds, &ip, &p);
2650 account_size_time (info, this_size * prob,
2651 this_time * prob, &ip);
2653 if (prob != 2)
2654 account_size_time (info, this_size * (2 - prob),
2655 this_time * (2 - prob), &p);
2658 gcc_assert (time >= 0);
2659 gcc_assert (size >= 0);
2663 set_hint_predicate (&inline_summary (node)->array_index, array_index);
2664 time = (time + CGRAPH_FREQ_BASE / 2) / CGRAPH_FREQ_BASE;
2665 if (time > MAX_TIME)
2666 time = MAX_TIME;
2667 free (order);
2669 if (!early && nonconstant_names.exists ())
2671 struct loop *loop;
2672 predicate loop_iterations = true_predicate ();
2673 predicate loop_stride = true_predicate ();
2675 if (dump_file && (dump_flags & TDF_DETAILS))
2676 flow_loops_dump (dump_file, NULL, 0);
2677 scev_initialize ();
2678 FOR_EACH_LOOP (loop, 0)
2680 vec<edge> exits;
2681 edge ex;
2682 unsigned int j, i;
2683 struct tree_niter_desc niter_desc;
2684 basic_block *body = get_loop_body (loop);
2685 bb_predicate = *(struct predicate *) loop->header->aux;
2687 exits = get_loop_exit_edges (loop);
2688 FOR_EACH_VEC_ELT (exits, j, ex)
2689 if (number_of_iterations_exit (loop, ex, &niter_desc, false)
2690 && !is_gimple_min_invariant (niter_desc.niter))
2692 predicate will_be_nonconstant
2693 = will_be_nonconstant_expr_predicate (parms_info, info,
2694 niter_desc.niter,
2695 nonconstant_names);
2696 if (!true_predicate_p (&will_be_nonconstant))
2697 will_be_nonconstant = and_predicates (info->conds,
2698 &bb_predicate,
2699 &will_be_nonconstant);
2700 if (!true_predicate_p (&will_be_nonconstant)
2701 && !false_predicate_p (&will_be_nonconstant))
2702 /* This is slightly inprecise. We may want to represent each
2703 loop with independent predicate. */
2704 loop_iterations =
2705 and_predicates (info->conds, &loop_iterations,
2706 &will_be_nonconstant);
2708 exits.release ();
2710 for (i = 0; i < loop->num_nodes; i++)
2712 gimple_stmt_iterator gsi;
2713 bb_predicate = *(struct predicate *) body[i]->aux;
2714 for (gsi = gsi_start_bb (body[i]); !gsi_end_p (gsi);
2715 gsi_next (&gsi))
2717 gimple stmt = gsi_stmt (gsi);
2718 affine_iv iv;
2719 ssa_op_iter iter;
2720 tree use;
2722 FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE)
2724 predicate will_be_nonconstant;
2726 if (!simple_iv
2727 (loop, loop_containing_stmt (stmt), use, &iv, true)
2728 || is_gimple_min_invariant (iv.step))
2729 continue;
2730 will_be_nonconstant
2731 = will_be_nonconstant_expr_predicate (parms_info, info,
2732 iv.step,
2733 nonconstant_names);
2734 if (!true_predicate_p (&will_be_nonconstant))
2735 will_be_nonconstant
2736 = and_predicates (info->conds,
2737 &bb_predicate,
2738 &will_be_nonconstant);
2739 if (!true_predicate_p (&will_be_nonconstant)
2740 && !false_predicate_p (&will_be_nonconstant))
2741 /* This is slightly inprecise. We may want to represent
2742 each loop with independent predicate. */
2743 loop_stride =
2744 and_predicates (info->conds, &loop_stride,
2745 &will_be_nonconstant);
2749 free (body);
2751 set_hint_predicate (&inline_summary (node)->loop_iterations,
2752 loop_iterations);
2753 set_hint_predicate (&inline_summary (node)->loop_stride, loop_stride);
2754 scev_finalize ();
2756 FOR_ALL_BB_FN (bb, my_function)
2758 edge e;
2759 edge_iterator ei;
2761 if (bb->aux)
2762 pool_free (edge_predicate_pool, bb->aux);
2763 bb->aux = NULL;
2764 FOR_EACH_EDGE (e, ei, bb->succs)
2766 if (e->aux)
2767 pool_free (edge_predicate_pool, e->aux);
2768 e->aux = NULL;
2771 inline_summary (node)->self_time = time;
2772 inline_summary (node)->self_size = size;
2773 nonconstant_names.release ();
2774 if (optimize && !early)
2776 loop_optimizer_finalize ();
2777 free_dominance_info (CDI_DOMINATORS);
2779 if (dump_file)
2781 fprintf (dump_file, "\n");
2782 dump_inline_summary (dump_file, node);
2787 /* Compute parameters of functions used by inliner.
2788 EARLY is true when we compute parameters for the early inliner */
2790 void
2791 compute_inline_parameters (struct cgraph_node *node, bool early)
2793 HOST_WIDE_INT self_stack_size;
2794 struct cgraph_edge *e;
2795 struct inline_summary *info;
2797 gcc_assert (!node->global.inlined_to);
2799 inline_summary_alloc ();
2801 info = inline_summary (node);
2802 reset_inline_summary (node);
2804 /* FIXME: Thunks are inlinable, but tree-inline don't know how to do that.
2805 Once this happen, we will need to more curefully predict call
2806 statement size. */
2807 if (node->thunk.thunk_p)
2809 struct inline_edge_summary *es = inline_edge_summary (node->callees);
2810 struct predicate t = true_predicate ();
2812 info->inlinable = 0;
2813 node->callees->call_stmt_cannot_inline_p = true;
2814 node->local.can_change_signature = false;
2815 es->call_stmt_time = 1;
2816 es->call_stmt_size = 1;
2817 account_size_time (info, 0, 0, &t);
2818 return;
2821 /* Even is_gimple_min_invariant rely on current_function_decl. */
2822 push_cfun (DECL_STRUCT_FUNCTION (node->decl));
2824 /* Estimate the stack size for the function if we're optimizing. */
2825 self_stack_size = optimize ? estimated_stack_frame_size (node) : 0;
2826 info->estimated_self_stack_size = self_stack_size;
2827 info->estimated_stack_size = self_stack_size;
2828 info->stack_frame_offset = 0;
2830 /* Can this function be inlined at all? */
2831 if (!optimize && !lookup_attribute ("always_inline",
2832 DECL_ATTRIBUTES (node->decl)))
2833 info->inlinable = false;
2834 else
2835 info->inlinable = tree_inlinable_function_p (node->decl);
2837 /* Type attributes can use parameter indices to describe them. */
2838 if (TYPE_ATTRIBUTES (TREE_TYPE (node->decl)))
2839 node->local.can_change_signature = false;
2840 else
2842 /* Otherwise, inlinable functions always can change signature. */
2843 if (info->inlinable)
2844 node->local.can_change_signature = true;
2845 else
2847 /* Functions calling builtin_apply can not change signature. */
2848 for (e = node->callees; e; e = e->next_callee)
2850 tree cdecl = e->callee->decl;
2851 if (DECL_BUILT_IN (cdecl)
2852 && DECL_BUILT_IN_CLASS (cdecl) == BUILT_IN_NORMAL
2853 && (DECL_FUNCTION_CODE (cdecl) == BUILT_IN_APPLY_ARGS
2854 || DECL_FUNCTION_CODE (cdecl) == BUILT_IN_VA_START))
2855 break;
2857 node->local.can_change_signature = !e;
2860 estimate_function_body_sizes (node, early);
2862 for (e = node->callees; e; e = e->next_callee)
2863 if (symtab_comdat_local_p (e->callee))
2864 break;
2865 node->calls_comdat_local = (e != NULL);
2867 /* Inlining characteristics are maintained by the cgraph_mark_inline. */
2868 info->time = info->self_time;
2869 info->size = info->self_size;
2870 info->stack_frame_offset = 0;
2871 info->estimated_stack_size = info->estimated_self_stack_size;
2872 #ifdef ENABLE_CHECKING
2873 inline_update_overall_summary (node);
2874 gcc_assert (info->time == info->self_time && info->size == info->self_size);
2875 #endif
2877 pop_cfun ();
2881 /* Compute parameters of functions used by inliner using
2882 current_function_decl. */
2884 static unsigned int
2885 compute_inline_parameters_for_current (void)
2887 compute_inline_parameters (cgraph_get_node (current_function_decl), true);
2888 return 0;
2891 namespace {
2893 const pass_data pass_data_inline_parameters =
2895 GIMPLE_PASS, /* type */
2896 "inline_param", /* name */
2897 OPTGROUP_INLINE, /* optinfo_flags */
2898 false, /* has_gate */
2899 true, /* has_execute */
2900 TV_INLINE_PARAMETERS, /* tv_id */
2901 0, /* properties_required */
2902 0, /* properties_provided */
2903 0, /* properties_destroyed */
2904 0, /* todo_flags_start */
2905 0, /* todo_flags_finish */
2908 class pass_inline_parameters : public gimple_opt_pass
2910 public:
2911 pass_inline_parameters (gcc::context *ctxt)
2912 : gimple_opt_pass (pass_data_inline_parameters, ctxt)
2915 /* opt_pass methods: */
2916 opt_pass * clone () { return new pass_inline_parameters (m_ctxt); }
2917 unsigned int execute () {
2918 return compute_inline_parameters_for_current ();
2921 }; // class pass_inline_parameters
2923 } // anon namespace
2925 gimple_opt_pass *
2926 make_pass_inline_parameters (gcc::context *ctxt)
2928 return new pass_inline_parameters (ctxt);
2932 /* Estimate benefit devirtualizing indirect edge IE, provided KNOWN_VALS and
2933 KNOWN_BINFOS. */
2935 static bool
2936 estimate_edge_devirt_benefit (struct cgraph_edge *ie,
2937 int *size, int *time,
2938 vec<tree> known_vals,
2939 vec<tree> known_binfos,
2940 vec<ipa_agg_jump_function_p> known_aggs)
2942 tree target;
2943 struct cgraph_node *callee;
2944 struct inline_summary *isummary;
2946 if (!known_vals.exists () && !known_binfos.exists ())
2947 return false;
2948 if (!flag_indirect_inlining)
2949 return false;
2951 target = ipa_get_indirect_edge_target (ie, known_vals, known_binfos,
2952 known_aggs);
2953 if (!target)
2954 return false;
2956 /* Account for difference in cost between indirect and direct calls. */
2957 *size -= (eni_size_weights.indirect_call_cost - eni_size_weights.call_cost);
2958 *time -= (eni_time_weights.indirect_call_cost - eni_time_weights.call_cost);
2959 gcc_checking_assert (*time >= 0);
2960 gcc_checking_assert (*size >= 0);
2962 callee = cgraph_get_node (target);
2963 if (!callee || !callee->definition)
2964 return false;
2965 isummary = inline_summary (callee);
2966 return isummary->inlinable;
2969 /* Increase SIZE and TIME for size and time needed to handle edge E. */
2971 static inline void
2972 estimate_edge_size_and_time (struct cgraph_edge *e, int *size, int *time,
2973 int prob,
2974 vec<tree> known_vals,
2975 vec<tree> known_binfos,
2976 vec<ipa_agg_jump_function_p> known_aggs,
2977 inline_hints *hints)
2979 struct inline_edge_summary *es = inline_edge_summary (e);
2980 int call_size = es->call_stmt_size;
2981 int call_time = es->call_stmt_time;
2982 if (!e->callee
2983 && estimate_edge_devirt_benefit (e, &call_size, &call_time,
2984 known_vals, known_binfos, known_aggs)
2985 && hints && cgraph_maybe_hot_edge_p (e))
2986 *hints |= INLINE_HINT_indirect_call;
2987 *size += call_size * INLINE_SIZE_SCALE;
2988 *time += apply_probability ((gcov_type) call_time, prob)
2989 * e->frequency * (INLINE_TIME_SCALE / CGRAPH_FREQ_BASE);
2990 if (*time > MAX_TIME * INLINE_TIME_SCALE)
2991 *time = MAX_TIME * INLINE_TIME_SCALE;
2996 /* Increase SIZE and TIME for size and time needed to handle all calls in NODE.
2997 POSSIBLE_TRUTHS, KNOWN_VALS and KNOWN_BINFOS describe context of the call
2998 site. */
3000 static void
3001 estimate_calls_size_and_time (struct cgraph_node *node, int *size, int *time,
3002 inline_hints *hints,
3003 clause_t possible_truths,
3004 vec<tree> known_vals,
3005 vec<tree> known_binfos,
3006 vec<ipa_agg_jump_function_p> known_aggs)
3008 struct cgraph_edge *e;
3009 for (e = node->callees; e; e = e->next_callee)
3011 struct inline_edge_summary *es = inline_edge_summary (e);
3012 if (!es->predicate
3013 || evaluate_predicate (es->predicate, possible_truths))
3015 if (e->inline_failed)
3017 /* Predicates of calls shall not use NOT_CHANGED codes,
3018 sowe do not need to compute probabilities. */
3019 estimate_edge_size_and_time (e, size, time, REG_BR_PROB_BASE,
3020 known_vals, known_binfos,
3021 known_aggs, hints);
3023 else
3024 estimate_calls_size_and_time (e->callee, size, time, hints,
3025 possible_truths,
3026 known_vals, known_binfos,
3027 known_aggs);
3030 for (e = node->indirect_calls; e; e = e->next_callee)
3032 struct inline_edge_summary *es = inline_edge_summary (e);
3033 if (!es->predicate
3034 || evaluate_predicate (es->predicate, possible_truths))
3035 estimate_edge_size_and_time (e, size, time, REG_BR_PROB_BASE,
3036 known_vals, known_binfos, known_aggs,
3037 hints);
3042 /* Estimate size and time needed to execute NODE assuming
3043 POSSIBLE_TRUTHS clause, and KNOWN_VALS and KNOWN_BINFOS information
3044 about NODE's arguments. */
3046 static void
3047 estimate_node_size_and_time (struct cgraph_node *node,
3048 clause_t possible_truths,
3049 vec<tree> known_vals,
3050 vec<tree> known_binfos,
3051 vec<ipa_agg_jump_function_p> known_aggs,
3052 int *ret_size, int *ret_time,
3053 inline_hints *ret_hints,
3054 vec<inline_param_summary>
3055 inline_param_summary)
3057 struct inline_summary *info = inline_summary (node);
3058 size_time_entry *e;
3059 int size = 0;
3060 int time = 0;
3061 inline_hints hints = 0;
3062 int i;
3064 if (dump_file && (dump_flags & TDF_DETAILS))
3066 bool found = false;
3067 fprintf (dump_file, " Estimating body: %s/%i\n"
3068 " Known to be false: ", node->name (),
3069 node->order);
3071 for (i = predicate_not_inlined_condition;
3072 i < (predicate_first_dynamic_condition
3073 + (int) vec_safe_length (info->conds)); i++)
3074 if (!(possible_truths & (1 << i)))
3076 if (found)
3077 fprintf (dump_file, ", ");
3078 found = true;
3079 dump_condition (dump_file, info->conds, i);
3083 for (i = 0; vec_safe_iterate (info->entry, i, &e); i++)
3084 if (evaluate_predicate (&e->predicate, possible_truths))
3086 size += e->size;
3087 gcc_checking_assert (e->time >= 0);
3088 gcc_checking_assert (time >= 0);
3089 if (!inline_param_summary.exists ())
3090 time += e->time;
3091 else
3093 int prob = predicate_probability (info->conds,
3094 &e->predicate,
3095 possible_truths,
3096 inline_param_summary);
3097 gcc_checking_assert (prob >= 0);
3098 gcc_checking_assert (prob <= REG_BR_PROB_BASE);
3099 time += apply_probability ((gcov_type) e->time, prob);
3101 if (time > MAX_TIME * INLINE_TIME_SCALE)
3102 time = MAX_TIME * INLINE_TIME_SCALE;
3103 gcc_checking_assert (time >= 0);
3106 gcc_checking_assert (size >= 0);
3107 gcc_checking_assert (time >= 0);
3109 if (info->loop_iterations
3110 && !evaluate_predicate (info->loop_iterations, possible_truths))
3111 hints |= INLINE_HINT_loop_iterations;
3112 if (info->loop_stride
3113 && !evaluate_predicate (info->loop_stride, possible_truths))
3114 hints |= INLINE_HINT_loop_stride;
3115 if (info->array_index
3116 && !evaluate_predicate (info->array_index, possible_truths))
3117 hints |= INLINE_HINT_array_index;
3118 if (info->scc_no)
3119 hints |= INLINE_HINT_in_scc;
3120 if (DECL_DECLARED_INLINE_P (node->decl))
3121 hints |= INLINE_HINT_declared_inline;
3123 estimate_calls_size_and_time (node, &size, &time, &hints, possible_truths,
3124 known_vals, known_binfos, known_aggs);
3125 gcc_checking_assert (size >= 0);
3126 gcc_checking_assert (time >= 0);
3127 time = RDIV (time, INLINE_TIME_SCALE);
3128 size = RDIV (size, INLINE_SIZE_SCALE);
3130 if (dump_file && (dump_flags & TDF_DETAILS))
3131 fprintf (dump_file, "\n size:%i time:%i\n", (int) size, (int) time);
3132 if (ret_time)
3133 *ret_time = time;
3134 if (ret_size)
3135 *ret_size = size;
3136 if (ret_hints)
3137 *ret_hints = hints;
3138 return;
3142 /* Estimate size and time needed to execute callee of EDGE assuming that
3143 parameters known to be constant at caller of EDGE are propagated.
3144 KNOWN_VALS and KNOWN_BINFOS are vectors of assumed known constant values
3145 and types for parameters. */
3147 void
3148 estimate_ipcp_clone_size_and_time (struct cgraph_node *node,
3149 vec<tree> known_vals,
3150 vec<tree> known_binfos,
3151 vec<ipa_agg_jump_function_p> known_aggs,
3152 int *ret_size, int *ret_time,
3153 inline_hints *hints)
3155 clause_t clause;
3157 clause = evaluate_conditions_for_known_args (node, false, known_vals,
3158 known_aggs);
3159 estimate_node_size_and_time (node, clause, known_vals, known_binfos,
3160 known_aggs, ret_size, ret_time, hints, vNULL);
3163 /* Translate all conditions from callee representation into caller
3164 representation and symbolically evaluate predicate P into new predicate.
3166 INFO is inline_summary of function we are adding predicate into, CALLEE_INFO
3167 is summary of function predicate P is from. OPERAND_MAP is array giving
3168 callee formal IDs the caller formal IDs. POSSSIBLE_TRUTHS is clausule of all
3169 callee conditions that may be true in caller context. TOPLEV_PREDICATE is
3170 predicate under which callee is executed. OFFSET_MAP is an array of of
3171 offsets that need to be added to conditions, negative offset means that
3172 conditions relying on values passed by reference have to be discarded
3173 because they might not be preserved (and should be considered offset zero
3174 for other purposes). */
3176 static struct predicate
3177 remap_predicate (struct inline_summary *info,
3178 struct inline_summary *callee_info,
3179 struct predicate *p,
3180 vec<int> operand_map,
3181 vec<int> offset_map,
3182 clause_t possible_truths, struct predicate *toplev_predicate)
3184 int i;
3185 struct predicate out = true_predicate ();
3187 /* True predicate is easy. */
3188 if (true_predicate_p (p))
3189 return *toplev_predicate;
3190 for (i = 0; p->clause[i]; i++)
3192 clause_t clause = p->clause[i];
3193 int cond;
3194 struct predicate clause_predicate = false_predicate ();
3196 gcc_assert (i < MAX_CLAUSES);
3198 for (cond = 0; cond < NUM_CONDITIONS; cond++)
3199 /* Do we have condition we can't disprove? */
3200 if (clause & possible_truths & (1 << cond))
3202 struct predicate cond_predicate;
3203 /* Work out if the condition can translate to predicate in the
3204 inlined function. */
3205 if (cond >= predicate_first_dynamic_condition)
3207 struct condition *c;
3209 c = &(*callee_info->conds)[cond
3211 predicate_first_dynamic_condition];
3212 /* See if we can remap condition operand to caller's operand.
3213 Otherwise give up. */
3214 if (!operand_map.exists ()
3215 || (int) operand_map.length () <= c->operand_num
3216 || operand_map[c->operand_num] == -1
3217 /* TODO: For non-aggregate conditions, adding an offset is
3218 basically an arithmetic jump function processing which
3219 we should support in future. */
3220 || ((!c->agg_contents || !c->by_ref)
3221 && offset_map[c->operand_num] > 0)
3222 || (c->agg_contents && c->by_ref
3223 && offset_map[c->operand_num] < 0))
3224 cond_predicate = true_predicate ();
3225 else
3227 struct agg_position_info ap;
3228 HOST_WIDE_INT offset_delta = offset_map[c->operand_num];
3229 if (offset_delta < 0)
3231 gcc_checking_assert (!c->agg_contents || !c->by_ref);
3232 offset_delta = 0;
3234 gcc_assert (!c->agg_contents
3235 || c->by_ref || offset_delta == 0);
3236 ap.offset = c->offset + offset_delta;
3237 ap.agg_contents = c->agg_contents;
3238 ap.by_ref = c->by_ref;
3239 cond_predicate = add_condition (info,
3240 operand_map[c->operand_num],
3241 &ap, c->code, c->val);
3244 /* Fixed conditions remains same, construct single
3245 condition predicate. */
3246 else
3248 cond_predicate.clause[0] = 1 << cond;
3249 cond_predicate.clause[1] = 0;
3251 clause_predicate = or_predicates (info->conds, &clause_predicate,
3252 &cond_predicate);
3254 out = and_predicates (info->conds, &out, &clause_predicate);
3256 return and_predicates (info->conds, &out, toplev_predicate);
3260 /* Update summary information of inline clones after inlining.
3261 Compute peak stack usage. */
3263 static void
3264 inline_update_callee_summaries (struct cgraph_node *node, int depth)
3266 struct cgraph_edge *e;
3267 struct inline_summary *callee_info = inline_summary (node);
3268 struct inline_summary *caller_info = inline_summary (node->callers->caller);
3269 HOST_WIDE_INT peak;
3271 callee_info->stack_frame_offset
3272 = caller_info->stack_frame_offset
3273 + caller_info->estimated_self_stack_size;
3274 peak = callee_info->stack_frame_offset
3275 + callee_info->estimated_self_stack_size;
3276 if (inline_summary (node->global.inlined_to)->estimated_stack_size < peak)
3277 inline_summary (node->global.inlined_to)->estimated_stack_size = peak;
3278 ipa_propagate_frequency (node);
3279 for (e = node->callees; e; e = e->next_callee)
3281 if (!e->inline_failed)
3282 inline_update_callee_summaries (e->callee, depth);
3283 inline_edge_summary (e)->loop_depth += depth;
3285 for (e = node->indirect_calls; e; e = e->next_callee)
3286 inline_edge_summary (e)->loop_depth += depth;
3289 /* Update change_prob of EDGE after INLINED_EDGE has been inlined.
3290 When functoin A is inlined in B and A calls C with parameter that
3291 changes with probability PROB1 and C is known to be passthroug
3292 of argument if B that change with probability PROB2, the probability
3293 of change is now PROB1*PROB2. */
3295 static void
3296 remap_edge_change_prob (struct cgraph_edge *inlined_edge,
3297 struct cgraph_edge *edge)
3299 if (ipa_node_params_vector.exists ())
3301 int i;
3302 struct ipa_edge_args *args = IPA_EDGE_REF (edge);
3303 struct inline_edge_summary *es = inline_edge_summary (edge);
3304 struct inline_edge_summary *inlined_es
3305 = inline_edge_summary (inlined_edge);
3307 for (i = 0; i < ipa_get_cs_argument_count (args); i++)
3309 struct ipa_jump_func *jfunc = ipa_get_ith_jump_func (args, i);
3310 if (jfunc->type == IPA_JF_PASS_THROUGH
3311 && (ipa_get_jf_pass_through_formal_id (jfunc)
3312 < (int) inlined_es->param.length ()))
3314 int jf_formal_id = ipa_get_jf_pass_through_formal_id (jfunc);
3315 int prob1 = es->param[i].change_prob;
3316 int prob2 = inlined_es->param[jf_formal_id].change_prob;
3317 int prob = combine_probabilities (prob1, prob2);
3319 if (prob1 && prob2 && !prob)
3320 prob = 1;
3322 es->param[i].change_prob = prob;
3328 /* Update edge summaries of NODE after INLINED_EDGE has been inlined.
3330 Remap predicates of callees of NODE. Rest of arguments match
3331 remap_predicate.
3333 Also update change probabilities. */
3335 static void
3336 remap_edge_summaries (struct cgraph_edge *inlined_edge,
3337 struct cgraph_node *node,
3338 struct inline_summary *info,
3339 struct inline_summary *callee_info,
3340 vec<int> operand_map,
3341 vec<int> offset_map,
3342 clause_t possible_truths,
3343 struct predicate *toplev_predicate)
3345 struct cgraph_edge *e;
3346 for (e = node->callees; e; e = e->next_callee)
3348 struct inline_edge_summary *es = inline_edge_summary (e);
3349 struct predicate p;
3351 if (e->inline_failed)
3353 remap_edge_change_prob (inlined_edge, e);
3355 if (es->predicate)
3357 p = remap_predicate (info, callee_info,
3358 es->predicate, operand_map, offset_map,
3359 possible_truths, toplev_predicate);
3360 edge_set_predicate (e, &p);
3361 /* TODO: We should remove the edge for code that will be
3362 optimized out, but we need to keep verifiers and tree-inline
3363 happy. Make it cold for now. */
3364 if (false_predicate_p (&p))
3366 e->count = 0;
3367 e->frequency = 0;
3370 else
3371 edge_set_predicate (e, toplev_predicate);
3373 else
3374 remap_edge_summaries (inlined_edge, e->callee, info, callee_info,
3375 operand_map, offset_map, possible_truths,
3376 toplev_predicate);
3378 for (e = node->indirect_calls; e; e = e->next_callee)
3380 struct inline_edge_summary *es = inline_edge_summary (e);
3381 struct predicate p;
3383 remap_edge_change_prob (inlined_edge, e);
3384 if (es->predicate)
3386 p = remap_predicate (info, callee_info,
3387 es->predicate, operand_map, offset_map,
3388 possible_truths, toplev_predicate);
3389 edge_set_predicate (e, &p);
3390 /* TODO: We should remove the edge for code that will be optimized
3391 out, but we need to keep verifiers and tree-inline happy.
3392 Make it cold for now. */
3393 if (false_predicate_p (&p))
3395 e->count = 0;
3396 e->frequency = 0;
3399 else
3400 edge_set_predicate (e, toplev_predicate);
3404 /* Same as remap_predicate, but set result into hint *HINT. */
3406 static void
3407 remap_hint_predicate (struct inline_summary *info,
3408 struct inline_summary *callee_info,
3409 struct predicate **hint,
3410 vec<int> operand_map,
3411 vec<int> offset_map,
3412 clause_t possible_truths,
3413 struct predicate *toplev_predicate)
3415 predicate p;
3417 if (!*hint)
3418 return;
3419 p = remap_predicate (info, callee_info,
3420 *hint,
3421 operand_map, offset_map,
3422 possible_truths, toplev_predicate);
3423 if (!false_predicate_p (&p) && !true_predicate_p (&p))
3425 if (!*hint)
3426 set_hint_predicate (hint, p);
3427 else
3428 **hint = and_predicates (info->conds, *hint, &p);
3432 /* We inlined EDGE. Update summary of the function we inlined into. */
3434 void
3435 inline_merge_summary (struct cgraph_edge *edge)
3437 struct inline_summary *callee_info = inline_summary (edge->callee);
3438 struct cgraph_node *to = (edge->caller->global.inlined_to
3439 ? edge->caller->global.inlined_to : edge->caller);
3440 struct inline_summary *info = inline_summary (to);
3441 clause_t clause = 0; /* not_inline is known to be false. */
3442 size_time_entry *e;
3443 vec<int> operand_map = vNULL;
3444 vec<int> offset_map = vNULL;
3445 int i;
3446 struct predicate toplev_predicate;
3447 struct predicate true_p = true_predicate ();
3448 struct inline_edge_summary *es = inline_edge_summary (edge);
3450 if (es->predicate)
3451 toplev_predicate = *es->predicate;
3452 else
3453 toplev_predicate = true_predicate ();
3455 if (ipa_node_params_vector.exists () && callee_info->conds)
3457 struct ipa_edge_args *args = IPA_EDGE_REF (edge);
3458 int count = ipa_get_cs_argument_count (args);
3459 int i;
3461 evaluate_properties_for_edge (edge, true, &clause, NULL, NULL, NULL);
3462 if (count)
3464 operand_map.safe_grow_cleared (count);
3465 offset_map.safe_grow_cleared (count);
3467 for (i = 0; i < count; i++)
3469 struct ipa_jump_func *jfunc = ipa_get_ith_jump_func (args, i);
3470 int map = -1;
3472 /* TODO: handle non-NOPs when merging. */
3473 if (jfunc->type == IPA_JF_PASS_THROUGH)
3475 if (ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
3476 map = ipa_get_jf_pass_through_formal_id (jfunc);
3477 if (!ipa_get_jf_pass_through_agg_preserved (jfunc))
3478 offset_map[i] = -1;
3480 else if (jfunc->type == IPA_JF_ANCESTOR)
3482 HOST_WIDE_INT offset = ipa_get_jf_ancestor_offset (jfunc);
3483 if (offset >= 0 && offset < INT_MAX)
3485 map = ipa_get_jf_ancestor_formal_id (jfunc);
3486 if (!ipa_get_jf_ancestor_agg_preserved (jfunc))
3487 offset = -1;
3488 offset_map[i] = offset;
3491 operand_map[i] = map;
3492 gcc_assert (map < ipa_get_param_count (IPA_NODE_REF (to)));
3495 for (i = 0; vec_safe_iterate (callee_info->entry, i, &e); i++)
3497 struct predicate p = remap_predicate (info, callee_info,
3498 &e->predicate, operand_map,
3499 offset_map, clause,
3500 &toplev_predicate);
3501 if (!false_predicate_p (&p))
3503 gcov_type add_time = ((gcov_type) e->time * edge->frequency
3504 + CGRAPH_FREQ_BASE / 2) / CGRAPH_FREQ_BASE;
3505 int prob = predicate_probability (callee_info->conds,
3506 &e->predicate,
3507 clause, es->param);
3508 add_time = apply_probability ((gcov_type) add_time, prob);
3509 if (add_time > MAX_TIME * INLINE_TIME_SCALE)
3510 add_time = MAX_TIME * INLINE_TIME_SCALE;
3511 if (prob != REG_BR_PROB_BASE
3512 && dump_file && (dump_flags & TDF_DETAILS))
3514 fprintf (dump_file, "\t\tScaling time by probability:%f\n",
3515 (double) prob / REG_BR_PROB_BASE);
3517 account_size_time (info, e->size, add_time, &p);
3520 remap_edge_summaries (edge, edge->callee, info, callee_info, operand_map,
3521 offset_map, clause, &toplev_predicate);
3522 remap_hint_predicate (info, callee_info,
3523 &callee_info->loop_iterations,
3524 operand_map, offset_map, clause, &toplev_predicate);
3525 remap_hint_predicate (info, callee_info,
3526 &callee_info->loop_stride,
3527 operand_map, offset_map, clause, &toplev_predicate);
3528 remap_hint_predicate (info, callee_info,
3529 &callee_info->array_index,
3530 operand_map, offset_map, clause, &toplev_predicate);
3532 inline_update_callee_summaries (edge->callee,
3533 inline_edge_summary (edge)->loop_depth);
3535 /* We do not maintain predicates of inlined edges, free it. */
3536 edge_set_predicate (edge, &true_p);
3537 /* Similarly remove param summaries. */
3538 es->param.release ();
3539 operand_map.release ();
3540 offset_map.release ();
3543 /* For performance reasons inline_merge_summary is not updating overall size
3544 and time. Recompute it. */
3546 void
3547 inline_update_overall_summary (struct cgraph_node *node)
3549 struct inline_summary *info = inline_summary (node);
3550 size_time_entry *e;
3551 int i;
3553 info->size = 0;
3554 info->time = 0;
3555 for (i = 0; vec_safe_iterate (info->entry, i, &e); i++)
3557 info->size += e->size, info->time += e->time;
3558 if (info->time > MAX_TIME * INLINE_TIME_SCALE)
3559 info->time = MAX_TIME * INLINE_TIME_SCALE;
3561 estimate_calls_size_and_time (node, &info->size, &info->time, NULL,
3562 ~(clause_t) (1 << predicate_false_condition),
3563 vNULL, vNULL, vNULL);
3564 info->time = (info->time + INLINE_TIME_SCALE / 2) / INLINE_TIME_SCALE;
3565 info->size = (info->size + INLINE_SIZE_SCALE / 2) / INLINE_SIZE_SCALE;
3568 /* Return hints derrived from EDGE. */
3570 simple_edge_hints (struct cgraph_edge *edge)
3572 int hints = 0;
3573 struct cgraph_node *to = (edge->caller->global.inlined_to
3574 ? edge->caller->global.inlined_to : edge->caller);
3575 if (inline_summary (to)->scc_no
3576 && inline_summary (to)->scc_no == inline_summary (edge->callee)->scc_no
3577 && !cgraph_edge_recursive_p (edge))
3578 hints |= INLINE_HINT_same_scc;
3580 if (to->lto_file_data && edge->callee->lto_file_data
3581 && to->lto_file_data != edge->callee->lto_file_data)
3582 hints |= INLINE_HINT_cross_module;
3584 return hints;
3587 /* Estimate the time cost for the caller when inlining EDGE.
3588 Only to be called via estimate_edge_time, that handles the
3589 caching mechanism.
3591 When caching, also update the cache entry. Compute both time and
3592 size, since we always need both metrics eventually. */
3595 do_estimate_edge_time (struct cgraph_edge *edge)
3597 int time;
3598 int size;
3599 inline_hints hints;
3600 struct cgraph_node *callee;
3601 clause_t clause;
3602 vec<tree> known_vals;
3603 vec<tree> known_binfos;
3604 vec<ipa_agg_jump_function_p> known_aggs;
3605 struct inline_edge_summary *es = inline_edge_summary (edge);
3607 callee = cgraph_function_or_thunk_node (edge->callee, NULL);
3609 gcc_checking_assert (edge->inline_failed);
3610 evaluate_properties_for_edge (edge, true,
3611 &clause, &known_vals, &known_binfos,
3612 &known_aggs);
3613 estimate_node_size_and_time (callee, clause, known_vals, known_binfos,
3614 known_aggs, &size, &time, &hints, es->param);
3615 known_vals.release ();
3616 known_binfos.release ();
3617 known_aggs.release ();
3618 gcc_checking_assert (size >= 0);
3619 gcc_checking_assert (time >= 0);
3621 /* When caching, update the cache entry. */
3622 if (edge_growth_cache.exists ())
3624 if ((int) edge_growth_cache.length () <= edge->uid)
3625 edge_growth_cache.safe_grow_cleared (cgraph_edge_max_uid);
3626 edge_growth_cache[edge->uid].time = time + (time >= 0);
3628 edge_growth_cache[edge->uid].size = size + (size >= 0);
3629 hints |= simple_edge_hints (edge);
3630 edge_growth_cache[edge->uid].hints = hints + 1;
3632 return time;
3636 /* Return estimated callee growth after inlining EDGE.
3637 Only to be called via estimate_edge_size. */
3640 do_estimate_edge_size (struct cgraph_edge *edge)
3642 int size;
3643 struct cgraph_node *callee;
3644 clause_t clause;
3645 vec<tree> known_vals;
3646 vec<tree> known_binfos;
3647 vec<ipa_agg_jump_function_p> known_aggs;
3649 /* When we do caching, use do_estimate_edge_time to populate the entry. */
3651 if (edge_growth_cache.exists ())
3653 do_estimate_edge_time (edge);
3654 size = edge_growth_cache[edge->uid].size;
3655 gcc_checking_assert (size);
3656 return size - (size > 0);
3659 callee = cgraph_function_or_thunk_node (edge->callee, NULL);
3661 /* Early inliner runs without caching, go ahead and do the dirty work. */
3662 gcc_checking_assert (edge->inline_failed);
3663 evaluate_properties_for_edge (edge, true,
3664 &clause, &known_vals, &known_binfos,
3665 &known_aggs);
3666 estimate_node_size_and_time (callee, clause, known_vals, known_binfos,
3667 known_aggs, &size, NULL, NULL, vNULL);
3668 known_vals.release ();
3669 known_binfos.release ();
3670 known_aggs.release ();
3671 return size;
3675 /* Estimate the growth of the caller when inlining EDGE.
3676 Only to be called via estimate_edge_size. */
3678 inline_hints
3679 do_estimate_edge_hints (struct cgraph_edge *edge)
3681 inline_hints hints;
3682 struct cgraph_node *callee;
3683 clause_t clause;
3684 vec<tree> known_vals;
3685 vec<tree> known_binfos;
3686 vec<ipa_agg_jump_function_p> known_aggs;
3688 /* When we do caching, use do_estimate_edge_time to populate the entry. */
3690 if (edge_growth_cache.exists ())
3692 do_estimate_edge_time (edge);
3693 hints = edge_growth_cache[edge->uid].hints;
3694 gcc_checking_assert (hints);
3695 return hints - 1;
3698 callee = cgraph_function_or_thunk_node (edge->callee, NULL);
3700 /* Early inliner runs without caching, go ahead and do the dirty work. */
3701 gcc_checking_assert (edge->inline_failed);
3702 evaluate_properties_for_edge (edge, true,
3703 &clause, &known_vals, &known_binfos,
3704 &known_aggs);
3705 estimate_node_size_and_time (callee, clause, known_vals, known_binfos,
3706 known_aggs, NULL, NULL, &hints, vNULL);
3707 known_vals.release ();
3708 known_binfos.release ();
3709 known_aggs.release ();
3710 hints |= simple_edge_hints (edge);
3711 return hints;
3715 /* Estimate self time of the function NODE after inlining EDGE. */
3718 estimate_time_after_inlining (struct cgraph_node *node,
3719 struct cgraph_edge *edge)
3721 struct inline_edge_summary *es = inline_edge_summary (edge);
3722 if (!es->predicate || !false_predicate_p (es->predicate))
3724 gcov_type time =
3725 inline_summary (node)->time + estimate_edge_time (edge);
3726 if (time < 0)
3727 time = 0;
3728 if (time > MAX_TIME)
3729 time = MAX_TIME;
3730 return time;
3732 return inline_summary (node)->time;
3736 /* Estimate the size of NODE after inlining EDGE which should be an
3737 edge to either NODE or a call inlined into NODE. */
3740 estimate_size_after_inlining (struct cgraph_node *node,
3741 struct cgraph_edge *edge)
3743 struct inline_edge_summary *es = inline_edge_summary (edge);
3744 if (!es->predicate || !false_predicate_p (es->predicate))
3746 int size = inline_summary (node)->size + estimate_edge_growth (edge);
3747 gcc_assert (size >= 0);
3748 return size;
3750 return inline_summary (node)->size;
3754 struct growth_data
3756 struct cgraph_node *node;
3757 bool self_recursive;
3758 int growth;
3762 /* Worker for do_estimate_growth. Collect growth for all callers. */
3764 static bool
3765 do_estimate_growth_1 (struct cgraph_node *node, void *data)
3767 struct cgraph_edge *e;
3768 struct growth_data *d = (struct growth_data *) data;
3770 for (e = node->callers; e; e = e->next_caller)
3772 gcc_checking_assert (e->inline_failed);
3774 if (e->caller == d->node
3775 || (e->caller->global.inlined_to
3776 && e->caller->global.inlined_to == d->node))
3777 d->self_recursive = true;
3778 d->growth += estimate_edge_growth (e);
3780 return false;
3784 /* Estimate the growth caused by inlining NODE into all callees. */
3787 do_estimate_growth (struct cgraph_node *node)
3789 struct growth_data d = { node, 0, false };
3790 struct inline_summary *info = inline_summary (node);
3792 cgraph_for_node_and_aliases (node, do_estimate_growth_1, &d, true);
3794 /* For self recursive functions the growth estimation really should be
3795 infinity. We don't want to return very large values because the growth
3796 plays various roles in badness computation fractions. Be sure to not
3797 return zero or negative growths. */
3798 if (d.self_recursive)
3799 d.growth = d.growth < info->size ? info->size : d.growth;
3800 else if (DECL_EXTERNAL (node->decl))
3802 else
3804 if (cgraph_will_be_removed_from_program_if_no_direct_calls (node))
3805 d.growth -= info->size;
3806 /* COMDAT functions are very often not shared across multiple units
3807 since they come from various template instantiations.
3808 Take this into account. */
3809 else if (DECL_COMDAT (node->decl)
3810 && cgraph_can_remove_if_no_direct_calls_p (node))
3811 d.growth -= (info->size
3812 * (100 - PARAM_VALUE (PARAM_COMDAT_SHARING_PROBABILITY))
3813 + 50) / 100;
3816 if (node_growth_cache.exists ())
3818 if ((int) node_growth_cache.length () <= node->uid)
3819 node_growth_cache.safe_grow_cleared (cgraph_max_uid);
3820 node_growth_cache[node->uid] = d.growth + (d.growth >= 0);
3822 return d.growth;
3826 /* This function performs intraprocedural analysis in NODE that is required to
3827 inline indirect calls. */
3829 static void
3830 inline_indirect_intraprocedural_analysis (struct cgraph_node *node)
3832 ipa_analyze_node (node);
3833 if (dump_file && (dump_flags & TDF_DETAILS))
3835 ipa_print_node_params (dump_file, node);
3836 ipa_print_node_jump_functions (dump_file, node);
3841 /* Note function body size. */
3843 static void
3844 inline_analyze_function (struct cgraph_node *node)
3846 push_cfun (DECL_STRUCT_FUNCTION (node->decl));
3848 if (dump_file)
3849 fprintf (dump_file, "\nAnalyzing function: %s/%u\n",
3850 node->name (), node->order);
3851 if (optimize && !node->thunk.thunk_p)
3852 inline_indirect_intraprocedural_analysis (node);
3853 compute_inline_parameters (node, false);
3854 if (!optimize)
3856 struct cgraph_edge *e;
3857 for (e = node->callees; e; e = e->next_callee)
3859 if (e->inline_failed == CIF_FUNCTION_NOT_CONSIDERED)
3860 e->inline_failed = CIF_FUNCTION_NOT_OPTIMIZED;
3861 e->call_stmt_cannot_inline_p = true;
3863 for (e = node->indirect_calls; e; e = e->next_callee)
3865 if (e->inline_failed == CIF_FUNCTION_NOT_CONSIDERED)
3866 e->inline_failed = CIF_FUNCTION_NOT_OPTIMIZED;
3867 e->call_stmt_cannot_inline_p = true;
3871 pop_cfun ();
3875 /* Called when new function is inserted to callgraph late. */
3877 static void
3878 add_new_function (struct cgraph_node *node, void *data ATTRIBUTE_UNUSED)
3880 inline_analyze_function (node);
3884 /* Note function body size. */
3886 void
3887 inline_generate_summary (void)
3889 struct cgraph_node *node;
3891 /* When not optimizing, do not bother to analyze. Inlining is still done
3892 because edge redirection needs to happen there. */
3893 if (!optimize && !flag_lto && !flag_wpa)
3894 return;
3896 function_insertion_hook_holder =
3897 cgraph_add_function_insertion_hook (&add_new_function, NULL);
3899 ipa_register_cgraph_hooks ();
3900 inline_free_summary ();
3902 FOR_EACH_DEFINED_FUNCTION (node)
3903 if (!node->alias)
3904 inline_analyze_function (node);
3908 /* Read predicate from IB. */
3910 static struct predicate
3911 read_predicate (struct lto_input_block *ib)
3913 struct predicate out;
3914 clause_t clause;
3915 int k = 0;
3919 gcc_assert (k <= MAX_CLAUSES);
3920 clause = out.clause[k++] = streamer_read_uhwi (ib);
3922 while (clause);
3924 /* Zero-initialize the remaining clauses in OUT. */
3925 while (k <= MAX_CLAUSES)
3926 out.clause[k++] = 0;
3928 return out;
3932 /* Write inline summary for edge E to OB. */
3934 static void
3935 read_inline_edge_summary (struct lto_input_block *ib, struct cgraph_edge *e)
3937 struct inline_edge_summary *es = inline_edge_summary (e);
3938 struct predicate p;
3939 int length, i;
3941 es->call_stmt_size = streamer_read_uhwi (ib);
3942 es->call_stmt_time = streamer_read_uhwi (ib);
3943 es->loop_depth = streamer_read_uhwi (ib);
3944 p = read_predicate (ib);
3945 edge_set_predicate (e, &p);
3946 length = streamer_read_uhwi (ib);
3947 if (length)
3949 es->param.safe_grow_cleared (length);
3950 for (i = 0; i < length; i++)
3951 es->param[i].change_prob = streamer_read_uhwi (ib);
3956 /* Stream in inline summaries from the section. */
3958 static void
3959 inline_read_section (struct lto_file_decl_data *file_data, const char *data,
3960 size_t len)
3962 const struct lto_function_header *header =
3963 (const struct lto_function_header *) data;
3964 const int cfg_offset = sizeof (struct lto_function_header);
3965 const int main_offset = cfg_offset + header->cfg_size;
3966 const int string_offset = main_offset + header->main_size;
3967 struct data_in *data_in;
3968 struct lto_input_block ib;
3969 unsigned int i, count2, j;
3970 unsigned int f_count;
3972 LTO_INIT_INPUT_BLOCK (ib, (const char *) data + main_offset, 0,
3973 header->main_size);
3975 data_in =
3976 lto_data_in_create (file_data, (const char *) data + string_offset,
3977 header->string_size, vNULL);
3978 f_count = streamer_read_uhwi (&ib);
3979 for (i = 0; i < f_count; i++)
3981 unsigned int index;
3982 struct cgraph_node *node;
3983 struct inline_summary *info;
3984 lto_symtab_encoder_t encoder;
3985 struct bitpack_d bp;
3986 struct cgraph_edge *e;
3987 predicate p;
3989 index = streamer_read_uhwi (&ib);
3990 encoder = file_data->symtab_node_encoder;
3991 node = cgraph (lto_symtab_encoder_deref (encoder, index));
3992 info = inline_summary (node);
3994 info->estimated_stack_size
3995 = info->estimated_self_stack_size = streamer_read_uhwi (&ib);
3996 info->size = info->self_size = streamer_read_uhwi (&ib);
3997 info->time = info->self_time = streamer_read_uhwi (&ib);
3999 bp = streamer_read_bitpack (&ib);
4000 info->inlinable = bp_unpack_value (&bp, 1);
4002 count2 = streamer_read_uhwi (&ib);
4003 gcc_assert (!info->conds);
4004 for (j = 0; j < count2; j++)
4006 struct condition c;
4007 c.operand_num = streamer_read_uhwi (&ib);
4008 c.code = (enum tree_code) streamer_read_uhwi (&ib);
4009 c.val = stream_read_tree (&ib, data_in);
4010 bp = streamer_read_bitpack (&ib);
4011 c.agg_contents = bp_unpack_value (&bp, 1);
4012 c.by_ref = bp_unpack_value (&bp, 1);
4013 if (c.agg_contents)
4014 c.offset = streamer_read_uhwi (&ib);
4015 vec_safe_push (info->conds, c);
4017 count2 = streamer_read_uhwi (&ib);
4018 gcc_assert (!info->entry);
4019 for (j = 0; j < count2; j++)
4021 struct size_time_entry e;
4023 e.size = streamer_read_uhwi (&ib);
4024 e.time = streamer_read_uhwi (&ib);
4025 e.predicate = read_predicate (&ib);
4027 vec_safe_push (info->entry, e);
4030 p = read_predicate (&ib);
4031 set_hint_predicate (&info->loop_iterations, p);
4032 p = read_predicate (&ib);
4033 set_hint_predicate (&info->loop_stride, p);
4034 p = read_predicate (&ib);
4035 set_hint_predicate (&info->array_index, p);
4036 for (e = node->callees; e; e = e->next_callee)
4037 read_inline_edge_summary (&ib, e);
4038 for (e = node->indirect_calls; e; e = e->next_callee)
4039 read_inline_edge_summary (&ib, e);
4042 lto_free_section_data (file_data, LTO_section_inline_summary, NULL, data,
4043 len);
4044 lto_data_in_delete (data_in);
4048 /* Read inline summary. Jump functions are shared among ipa-cp
4049 and inliner, so when ipa-cp is active, we don't need to write them
4050 twice. */
4052 void
4053 inline_read_summary (void)
4055 struct lto_file_decl_data **file_data_vec = lto_get_file_decl_data ();
4056 struct lto_file_decl_data *file_data;
4057 unsigned int j = 0;
4059 inline_summary_alloc ();
4061 while ((file_data = file_data_vec[j++]))
4063 size_t len;
4064 const char *data = lto_get_section_data (file_data,
4065 LTO_section_inline_summary,
4066 NULL, &len);
4067 if (data)
4068 inline_read_section (file_data, data, len);
4069 else
4070 /* Fatal error here. We do not want to support compiling ltrans units
4071 with different version of compiler or different flags than the WPA
4072 unit, so this should never happen. */
4073 fatal_error ("ipa inline summary is missing in input file");
4075 if (optimize)
4077 ipa_register_cgraph_hooks ();
4078 if (!flag_ipa_cp)
4079 ipa_prop_read_jump_functions ();
4081 function_insertion_hook_holder =
4082 cgraph_add_function_insertion_hook (&add_new_function, NULL);
4086 /* Write predicate P to OB. */
4088 static void
4089 write_predicate (struct output_block *ob, struct predicate *p)
4091 int j;
4092 if (p)
4093 for (j = 0; p->clause[j]; j++)
4095 gcc_assert (j < MAX_CLAUSES);
4096 streamer_write_uhwi (ob, p->clause[j]);
4098 streamer_write_uhwi (ob, 0);
4102 /* Write inline summary for edge E to OB. */
4104 static void
4105 write_inline_edge_summary (struct output_block *ob, struct cgraph_edge *e)
4107 struct inline_edge_summary *es = inline_edge_summary (e);
4108 int i;
4110 streamer_write_uhwi (ob, es->call_stmt_size);
4111 streamer_write_uhwi (ob, es->call_stmt_time);
4112 streamer_write_uhwi (ob, es->loop_depth);
4113 write_predicate (ob, es->predicate);
4114 streamer_write_uhwi (ob, es->param.length ());
4115 for (i = 0; i < (int) es->param.length (); i++)
4116 streamer_write_uhwi (ob, es->param[i].change_prob);
4120 /* Write inline summary for node in SET.
4121 Jump functions are shared among ipa-cp and inliner, so when ipa-cp is
4122 active, we don't need to write them twice. */
4124 void
4125 inline_write_summary (void)
4127 struct cgraph_node *node;
4128 struct output_block *ob = create_output_block (LTO_section_inline_summary);
4129 lto_symtab_encoder_t encoder = ob->decl_state->symtab_node_encoder;
4130 unsigned int count = 0;
4131 int i;
4133 for (i = 0; i < lto_symtab_encoder_size (encoder); i++)
4135 symtab_node *snode = lto_symtab_encoder_deref (encoder, i);
4136 cgraph_node *cnode = dyn_cast <cgraph_node> (snode);
4137 if (cnode && cnode->definition && !cnode->alias)
4138 count++;
4140 streamer_write_uhwi (ob, count);
4142 for (i = 0; i < lto_symtab_encoder_size (encoder); i++)
4144 symtab_node *snode = lto_symtab_encoder_deref (encoder, i);
4145 cgraph_node *cnode = dyn_cast <cgraph_node> (snode);
4146 if (cnode && (node = cnode)->definition && !node->alias)
4148 struct inline_summary *info = inline_summary (node);
4149 struct bitpack_d bp;
4150 struct cgraph_edge *edge;
4151 int i;
4152 size_time_entry *e;
4153 struct condition *c;
4155 streamer_write_uhwi (ob,
4156 lto_symtab_encoder_encode (encoder,
4158 node));
4159 streamer_write_hwi (ob, info->estimated_self_stack_size);
4160 streamer_write_hwi (ob, info->self_size);
4161 streamer_write_hwi (ob, info->self_time);
4162 bp = bitpack_create (ob->main_stream);
4163 bp_pack_value (&bp, info->inlinable, 1);
4164 streamer_write_bitpack (&bp);
4165 streamer_write_uhwi (ob, vec_safe_length (info->conds));
4166 for (i = 0; vec_safe_iterate (info->conds, i, &c); i++)
4168 streamer_write_uhwi (ob, c->operand_num);
4169 streamer_write_uhwi (ob, c->code);
4170 stream_write_tree (ob, c->val, true);
4171 bp = bitpack_create (ob->main_stream);
4172 bp_pack_value (&bp, c->agg_contents, 1);
4173 bp_pack_value (&bp, c->by_ref, 1);
4174 streamer_write_bitpack (&bp);
4175 if (c->agg_contents)
4176 streamer_write_uhwi (ob, c->offset);
4178 streamer_write_uhwi (ob, vec_safe_length (info->entry));
4179 for (i = 0; vec_safe_iterate (info->entry, i, &e); i++)
4181 streamer_write_uhwi (ob, e->size);
4182 streamer_write_uhwi (ob, e->time);
4183 write_predicate (ob, &e->predicate);
4185 write_predicate (ob, info->loop_iterations);
4186 write_predicate (ob, info->loop_stride);
4187 write_predicate (ob, info->array_index);
4188 for (edge = node->callees; edge; edge = edge->next_callee)
4189 write_inline_edge_summary (ob, edge);
4190 for (edge = node->indirect_calls; edge; edge = edge->next_callee)
4191 write_inline_edge_summary (ob, edge);
4194 streamer_write_char_stream (ob->main_stream, 0);
4195 produce_asm (ob, NULL);
4196 destroy_output_block (ob);
4198 if (optimize && !flag_ipa_cp)
4199 ipa_prop_write_jump_functions ();
4203 /* Release inline summary. */
4205 void
4206 inline_free_summary (void)
4208 struct cgraph_node *node;
4209 if (!inline_edge_summary_vec.exists ())
4210 return;
4211 FOR_EACH_DEFINED_FUNCTION (node)
4212 if (!node->alias)
4213 reset_inline_summary (node);
4214 if (function_insertion_hook_holder)
4215 cgraph_remove_function_insertion_hook (function_insertion_hook_holder);
4216 function_insertion_hook_holder = NULL;
4217 if (node_removal_hook_holder)
4218 cgraph_remove_node_removal_hook (node_removal_hook_holder);
4219 node_removal_hook_holder = NULL;
4220 if (edge_removal_hook_holder)
4221 cgraph_remove_edge_removal_hook (edge_removal_hook_holder);
4222 edge_removal_hook_holder = NULL;
4223 if (node_duplication_hook_holder)
4224 cgraph_remove_node_duplication_hook (node_duplication_hook_holder);
4225 node_duplication_hook_holder = NULL;
4226 if (edge_duplication_hook_holder)
4227 cgraph_remove_edge_duplication_hook (edge_duplication_hook_holder);
4228 edge_duplication_hook_holder = NULL;
4229 vec_free (inline_summary_vec);
4230 inline_edge_summary_vec.release ();
4231 if (edge_predicate_pool)
4232 free_alloc_pool (edge_predicate_pool);
4233 edge_predicate_pool = 0;