2014-04-14 Martin Jambor <mjambor@suse.cz>
[official-gcc.git] / gcc / ipa-inline-analysis.c
blob8e0f5dd898899e98de41c6a4b5a28545d5ef45b1
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);
750 /* If the edge is determined to be never executed, redirect it
751 to BUILTIN_UNREACHABLE to save inliner from inlining into it. */
752 if (predicate && false_predicate_p (predicate) && e->callee)
754 struct cgraph_node *callee = !e->inline_failed ? e->callee : NULL;
756 cgraph_redirect_edge_callee (e,
757 cgraph_get_create_node
758 (builtin_decl_implicit (BUILT_IN_UNREACHABLE)));
759 e->inline_failed = CIF_UNREACHABLE;
760 if (callee)
761 cgraph_remove_node_and_inline_clones (callee, NULL);
763 if (predicate && !true_predicate_p (predicate))
765 if (!es->predicate)
766 es->predicate = (struct predicate *) pool_alloc (edge_predicate_pool);
767 *es->predicate = *predicate;
769 else
771 if (es->predicate)
772 pool_free (edge_predicate_pool, es->predicate);
773 es->predicate = NULL;
777 /* Set predicate for hint *P. */
779 static void
780 set_hint_predicate (struct predicate **p, struct predicate new_predicate)
782 if (false_predicate_p (&new_predicate) || true_predicate_p (&new_predicate))
784 if (*p)
785 pool_free (edge_predicate_pool, *p);
786 *p = NULL;
788 else
790 if (!*p)
791 *p = (struct predicate *) pool_alloc (edge_predicate_pool);
792 **p = new_predicate;
797 /* KNOWN_VALS is partial mapping of parameters of NODE to constant values.
798 KNOWN_AGGS is a vector of aggreggate jump functions for each parameter.
799 Return clause of possible truths. When INLINE_P is true, assume that we are
800 inlining.
802 ERROR_MARK means compile time invariant. */
804 static clause_t
805 evaluate_conditions_for_known_args (struct cgraph_node *node,
806 bool inline_p,
807 vec<tree> known_vals,
808 vec<ipa_agg_jump_function_p>
809 known_aggs)
811 clause_t clause = inline_p ? 0 : 1 << predicate_not_inlined_condition;
812 struct inline_summary *info = inline_summary (node);
813 int i;
814 struct condition *c;
816 for (i = 0; vec_safe_iterate (info->conds, i, &c); i++)
818 tree val;
819 tree res;
821 /* We allow call stmt to have fewer arguments than the callee function
822 (especially for K&R style programs). So bound check here (we assume
823 known_aggs vector, if non-NULL, has the same length as
824 known_vals). */
825 gcc_checking_assert (!known_aggs.exists ()
826 || (known_vals.length () == known_aggs.length ()));
827 if (c->operand_num >= (int) known_vals.length ())
829 clause |= 1 << (i + predicate_first_dynamic_condition);
830 continue;
833 if (c->agg_contents)
835 struct ipa_agg_jump_function *agg;
837 if (c->code == CHANGED
838 && !c->by_ref
839 && (known_vals[c->operand_num] == error_mark_node))
840 continue;
842 if (known_aggs.exists ())
844 agg = known_aggs[c->operand_num];
845 val = ipa_find_agg_cst_for_param (agg, c->offset, c->by_ref);
847 else
848 val = NULL_TREE;
850 else
852 val = known_vals[c->operand_num];
853 if (val == error_mark_node && c->code != CHANGED)
854 val = NULL_TREE;
857 if (!val)
859 clause |= 1 << (i + predicate_first_dynamic_condition);
860 continue;
862 if (c->code == IS_NOT_CONSTANT || c->code == CHANGED)
863 continue;
864 res = fold_binary_to_constant (c->code, boolean_type_node, val, c->val);
865 if (res && integer_zerop (res))
866 continue;
867 clause |= 1 << (i + predicate_first_dynamic_condition);
869 return clause;
873 /* Work out what conditions might be true at invocation of E. */
875 static void
876 evaluate_properties_for_edge (struct cgraph_edge *e, bool inline_p,
877 clause_t *clause_ptr,
878 vec<tree> *known_vals_ptr,
879 vec<tree> *known_binfos_ptr,
880 vec<ipa_agg_jump_function_p> *known_aggs_ptr)
882 struct cgraph_node *callee =
883 cgraph_function_or_thunk_node (e->callee, NULL);
884 struct inline_summary *info = inline_summary (callee);
885 vec<tree> known_vals = vNULL;
886 vec<ipa_agg_jump_function_p> known_aggs = vNULL;
888 if (clause_ptr)
889 *clause_ptr = inline_p ? 0 : 1 << predicate_not_inlined_condition;
890 if (known_vals_ptr)
891 known_vals_ptr->create (0);
892 if (known_binfos_ptr)
893 known_binfos_ptr->create (0);
895 if (ipa_node_params_vector.exists ()
896 && !e->call_stmt_cannot_inline_p
897 && ((clause_ptr && info->conds) || known_vals_ptr || known_binfos_ptr))
899 struct ipa_node_params *parms_info;
900 struct ipa_edge_args *args = IPA_EDGE_REF (e);
901 struct inline_edge_summary *es = inline_edge_summary (e);
902 int i, count = ipa_get_cs_argument_count (args);
904 if (e->caller->global.inlined_to)
905 parms_info = IPA_NODE_REF (e->caller->global.inlined_to);
906 else
907 parms_info = IPA_NODE_REF (e->caller);
909 if (count && (info->conds || known_vals_ptr))
910 known_vals.safe_grow_cleared (count);
911 if (count && (info->conds || known_aggs_ptr))
912 known_aggs.safe_grow_cleared (count);
913 if (count && known_binfos_ptr)
914 known_binfos_ptr->safe_grow_cleared (count);
916 for (i = 0; i < count; i++)
918 struct ipa_jump_func *jf = ipa_get_ith_jump_func (args, i);
919 tree cst = ipa_value_from_jfunc (parms_info, jf);
920 if (cst)
922 if (known_vals.exists () && TREE_CODE (cst) != TREE_BINFO)
923 known_vals[i] = cst;
924 else if (known_binfos_ptr != NULL
925 && TREE_CODE (cst) == TREE_BINFO)
926 (*known_binfos_ptr)[i] = cst;
928 else if (inline_p && !es->param[i].change_prob)
929 known_vals[i] = error_mark_node;
930 /* TODO: When IPA-CP starts propagating and merging aggregate jump
931 functions, use its knowledge of the caller too, just like the
932 scalar case above. */
933 known_aggs[i] = &jf->agg;
937 if (clause_ptr)
938 *clause_ptr = evaluate_conditions_for_known_args (callee, inline_p,
939 known_vals, known_aggs);
941 if (known_vals_ptr)
942 *known_vals_ptr = known_vals;
943 else
944 known_vals.release ();
946 if (known_aggs_ptr)
947 *known_aggs_ptr = known_aggs;
948 else
949 known_aggs.release ();
953 /* Allocate the inline summary vector or resize it to cover all cgraph nodes. */
955 static void
956 inline_summary_alloc (void)
958 if (!node_removal_hook_holder)
959 node_removal_hook_holder =
960 cgraph_add_node_removal_hook (&inline_node_removal_hook, NULL);
961 if (!edge_removal_hook_holder)
962 edge_removal_hook_holder =
963 cgraph_add_edge_removal_hook (&inline_edge_removal_hook, NULL);
964 if (!node_duplication_hook_holder)
965 node_duplication_hook_holder =
966 cgraph_add_node_duplication_hook (&inline_node_duplication_hook, NULL);
967 if (!edge_duplication_hook_holder)
968 edge_duplication_hook_holder =
969 cgraph_add_edge_duplication_hook (&inline_edge_duplication_hook, NULL);
971 if (vec_safe_length (inline_summary_vec) <= (unsigned) cgraph_max_uid)
972 vec_safe_grow_cleared (inline_summary_vec, cgraph_max_uid + 1);
973 if (inline_edge_summary_vec.length () <= (unsigned) cgraph_edge_max_uid)
974 inline_edge_summary_vec.safe_grow_cleared (cgraph_edge_max_uid + 1);
975 if (!edge_predicate_pool)
976 edge_predicate_pool = create_alloc_pool ("edge predicates",
977 sizeof (struct predicate), 10);
980 /* We are called multiple time for given function; clear
981 data from previous run so they are not cumulated. */
983 static void
984 reset_inline_edge_summary (struct cgraph_edge *e)
986 if (e->uid < (int) inline_edge_summary_vec.length ())
988 struct inline_edge_summary *es = inline_edge_summary (e);
990 es->call_stmt_size = es->call_stmt_time = 0;
991 if (es->predicate)
992 pool_free (edge_predicate_pool, es->predicate);
993 es->predicate = NULL;
994 es->param.release ();
998 /* We are called multiple time for given function; clear
999 data from previous run so they are not cumulated. */
1001 static void
1002 reset_inline_summary (struct cgraph_node *node)
1004 struct inline_summary *info = inline_summary (node);
1005 struct cgraph_edge *e;
1007 info->self_size = info->self_time = 0;
1008 info->estimated_stack_size = 0;
1009 info->estimated_self_stack_size = 0;
1010 info->stack_frame_offset = 0;
1011 info->size = 0;
1012 info->time = 0;
1013 info->growth = 0;
1014 info->scc_no = 0;
1015 if (info->loop_iterations)
1017 pool_free (edge_predicate_pool, info->loop_iterations);
1018 info->loop_iterations = NULL;
1020 if (info->loop_stride)
1022 pool_free (edge_predicate_pool, info->loop_stride);
1023 info->loop_stride = NULL;
1025 if (info->array_index)
1027 pool_free (edge_predicate_pool, info->array_index);
1028 info->array_index = NULL;
1030 vec_free (info->conds);
1031 vec_free (info->entry);
1032 for (e = node->callees; e; e = e->next_callee)
1033 reset_inline_edge_summary (e);
1034 for (e = node->indirect_calls; e; e = e->next_callee)
1035 reset_inline_edge_summary (e);
1038 /* Hook that is called by cgraph.c when a node is removed. */
1040 static void
1041 inline_node_removal_hook (struct cgraph_node *node,
1042 void *data ATTRIBUTE_UNUSED)
1044 struct inline_summary *info;
1045 if (vec_safe_length (inline_summary_vec) <= (unsigned) node->uid)
1046 return;
1047 info = inline_summary (node);
1048 reset_inline_summary (node);
1049 memset (info, 0, sizeof (inline_summary_t));
1052 /* Remap predicate P of former function to be predicate of duplicated function.
1053 POSSIBLE_TRUTHS is clause of possible truths in the duplicated node,
1054 INFO is inline summary of the duplicated node. */
1056 static struct predicate
1057 remap_predicate_after_duplication (struct predicate *p,
1058 clause_t possible_truths,
1059 struct inline_summary *info)
1061 struct predicate new_predicate = true_predicate ();
1062 int j;
1063 for (j = 0; p->clause[j]; j++)
1064 if (!(possible_truths & p->clause[j]))
1066 new_predicate = false_predicate ();
1067 break;
1069 else
1070 add_clause (info->conds, &new_predicate,
1071 possible_truths & p->clause[j]);
1072 return new_predicate;
1075 /* Same as remap_predicate_after_duplication but handle hint predicate *P.
1076 Additionally care about allocating new memory slot for updated predicate
1077 and set it to NULL when it becomes true or false (and thus uninteresting).
1080 static void
1081 remap_hint_predicate_after_duplication (struct predicate **p,
1082 clause_t possible_truths,
1083 struct inline_summary *info)
1085 struct predicate new_predicate;
1087 if (!*p)
1088 return;
1090 new_predicate = remap_predicate_after_duplication (*p,
1091 possible_truths, info);
1092 /* We do not want to free previous predicate; it is used by node origin. */
1093 *p = NULL;
1094 set_hint_predicate (p, new_predicate);
1098 /* Hook that is called by cgraph.c when a node is duplicated. */
1100 static void
1101 inline_node_duplication_hook (struct cgraph_node *src,
1102 struct cgraph_node *dst,
1103 ATTRIBUTE_UNUSED void *data)
1105 struct inline_summary *info;
1106 inline_summary_alloc ();
1107 info = inline_summary (dst);
1108 memcpy (info, inline_summary (src), sizeof (struct inline_summary));
1109 /* TODO: as an optimization, we may avoid copying conditions
1110 that are known to be false or true. */
1111 info->conds = vec_safe_copy (info->conds);
1113 /* When there are any replacements in the function body, see if we can figure
1114 out that something was optimized out. */
1115 if (ipa_node_params_vector.exists () && dst->clone.tree_map)
1117 vec<size_time_entry, va_gc> *entry = info->entry;
1118 /* Use SRC parm info since it may not be copied yet. */
1119 struct ipa_node_params *parms_info = IPA_NODE_REF (src);
1120 vec<tree> known_vals = vNULL;
1121 int count = ipa_get_param_count (parms_info);
1122 int i, j;
1123 clause_t possible_truths;
1124 struct predicate true_pred = true_predicate ();
1125 size_time_entry *e;
1126 int optimized_out_size = 0;
1127 bool inlined_to_p = false;
1128 struct cgraph_edge *edge;
1130 info->entry = 0;
1131 known_vals.safe_grow_cleared (count);
1132 for (i = 0; i < count; i++)
1134 struct ipa_replace_map *r;
1136 for (j = 0; vec_safe_iterate (dst->clone.tree_map, j, &r); j++)
1138 if (((!r->old_tree && r->parm_num == i)
1139 || (r->old_tree && r->old_tree == ipa_get_param (parms_info, i)))
1140 && r->replace_p && !r->ref_p)
1142 known_vals[i] = r->new_tree;
1143 break;
1147 possible_truths = evaluate_conditions_for_known_args (dst, false,
1148 known_vals,
1149 vNULL);
1150 known_vals.release ();
1152 account_size_time (info, 0, 0, &true_pred);
1154 /* Remap size_time vectors.
1155 Simplify the predicate by prunning out alternatives that are known
1156 to be false.
1157 TODO: as on optimization, we can also eliminate conditions known
1158 to be true. */
1159 for (i = 0; vec_safe_iterate (entry, i, &e); i++)
1161 struct predicate new_predicate;
1162 new_predicate = remap_predicate_after_duplication (&e->predicate,
1163 possible_truths,
1164 info);
1165 if (false_predicate_p (&new_predicate))
1166 optimized_out_size += e->size;
1167 else
1168 account_size_time (info, e->size, e->time, &new_predicate);
1171 /* Remap edge predicates with the same simplification as above.
1172 Also copy constantness arrays. */
1173 for (edge = dst->callees; edge; edge = edge->next_callee)
1175 struct predicate new_predicate;
1176 struct inline_edge_summary *es = inline_edge_summary (edge);
1178 if (!edge->inline_failed)
1179 inlined_to_p = true;
1180 if (!es->predicate)
1181 continue;
1182 new_predicate = remap_predicate_after_duplication (es->predicate,
1183 possible_truths,
1184 info);
1185 if (false_predicate_p (&new_predicate)
1186 && !false_predicate_p (es->predicate))
1188 optimized_out_size += es->call_stmt_size * INLINE_SIZE_SCALE;
1189 edge->frequency = 0;
1191 edge_set_predicate (edge, &new_predicate);
1194 /* Remap indirect edge predicates with the same simplificaiton as above.
1195 Also copy constantness arrays. */
1196 for (edge = dst->indirect_calls; edge; edge = edge->next_callee)
1198 struct predicate new_predicate;
1199 struct inline_edge_summary *es = inline_edge_summary (edge);
1201 gcc_checking_assert (edge->inline_failed);
1202 if (!es->predicate)
1203 continue;
1204 new_predicate = remap_predicate_after_duplication (es->predicate,
1205 possible_truths,
1206 info);
1207 if (false_predicate_p (&new_predicate)
1208 && !false_predicate_p (es->predicate))
1210 optimized_out_size += es->call_stmt_size * INLINE_SIZE_SCALE;
1211 edge->frequency = 0;
1213 edge_set_predicate (edge, &new_predicate);
1215 remap_hint_predicate_after_duplication (&info->loop_iterations,
1216 possible_truths, info);
1217 remap_hint_predicate_after_duplication (&info->loop_stride,
1218 possible_truths, info);
1219 remap_hint_predicate_after_duplication (&info->array_index,
1220 possible_truths, info);
1222 /* If inliner or someone after inliner will ever start producing
1223 non-trivial clones, we will get trouble with lack of information
1224 about updating self sizes, because size vectors already contains
1225 sizes of the calees. */
1226 gcc_assert (!inlined_to_p || !optimized_out_size);
1228 else
1230 info->entry = vec_safe_copy (info->entry);
1231 if (info->loop_iterations)
1233 predicate p = *info->loop_iterations;
1234 info->loop_iterations = NULL;
1235 set_hint_predicate (&info->loop_iterations, p);
1237 if (info->loop_stride)
1239 predicate p = *info->loop_stride;
1240 info->loop_stride = NULL;
1241 set_hint_predicate (&info->loop_stride, p);
1243 if (info->array_index)
1245 predicate p = *info->array_index;
1246 info->array_index = NULL;
1247 set_hint_predicate (&info->array_index, p);
1250 inline_update_overall_summary (dst);
1254 /* Hook that is called by cgraph.c when a node is duplicated. */
1256 static void
1257 inline_edge_duplication_hook (struct cgraph_edge *src,
1258 struct cgraph_edge *dst,
1259 ATTRIBUTE_UNUSED void *data)
1261 struct inline_edge_summary *info;
1262 struct inline_edge_summary *srcinfo;
1263 inline_summary_alloc ();
1264 info = inline_edge_summary (dst);
1265 srcinfo = inline_edge_summary (src);
1266 memcpy (info, srcinfo, sizeof (struct inline_edge_summary));
1267 info->predicate = NULL;
1268 edge_set_predicate (dst, srcinfo->predicate);
1269 info->param = srcinfo->param.copy ();
1273 /* Keep edge cache consistent across edge removal. */
1275 static void
1276 inline_edge_removal_hook (struct cgraph_edge *edge,
1277 void *data ATTRIBUTE_UNUSED)
1279 if (edge_growth_cache.exists ())
1280 reset_edge_growth_cache (edge);
1281 reset_inline_edge_summary (edge);
1285 /* Initialize growth caches. */
1287 void
1288 initialize_growth_caches (void)
1290 if (cgraph_edge_max_uid)
1291 edge_growth_cache.safe_grow_cleared (cgraph_edge_max_uid);
1292 if (cgraph_max_uid)
1293 node_growth_cache.safe_grow_cleared (cgraph_max_uid);
1297 /* Free growth caches. */
1299 void
1300 free_growth_caches (void)
1302 edge_growth_cache.release ();
1303 node_growth_cache.release ();
1307 /* Dump edge summaries associated to NODE and recursively to all clones.
1308 Indent by INDENT. */
1310 static void
1311 dump_inline_edge_summary (FILE *f, int indent, struct cgraph_node *node,
1312 struct inline_summary *info)
1314 struct cgraph_edge *edge;
1315 for (edge = node->callees; edge; edge = edge->next_callee)
1317 struct inline_edge_summary *es = inline_edge_summary (edge);
1318 struct cgraph_node *callee =
1319 cgraph_function_or_thunk_node (edge->callee, NULL);
1320 int i;
1322 fprintf (f,
1323 "%*s%s/%i %s\n%*s loop depth:%2i freq:%4i size:%2i"
1324 " time: %2i callee size:%2i stack:%2i",
1325 indent, "", callee->name (), callee->order,
1326 !edge->inline_failed
1327 ? "inlined" : cgraph_inline_failed_string (edge-> inline_failed),
1328 indent, "", es->loop_depth, edge->frequency,
1329 es->call_stmt_size, es->call_stmt_time,
1330 (int) inline_summary (callee)->size / INLINE_SIZE_SCALE,
1331 (int) inline_summary (callee)->estimated_stack_size);
1333 if (es->predicate)
1335 fprintf (f, " predicate: ");
1336 dump_predicate (f, info->conds, es->predicate);
1338 else
1339 fprintf (f, "\n");
1340 if (es->param.exists ())
1341 for (i = 0; i < (int) es->param.length (); i++)
1343 int prob = es->param[i].change_prob;
1345 if (!prob)
1346 fprintf (f, "%*s op%i is compile time invariant\n",
1347 indent + 2, "", i);
1348 else if (prob != REG_BR_PROB_BASE)
1349 fprintf (f, "%*s op%i change %f%% of time\n", indent + 2, "", i,
1350 prob * 100.0 / REG_BR_PROB_BASE);
1352 if (!edge->inline_failed)
1354 fprintf (f, "%*sStack frame offset %i, callee self size %i,"
1355 " callee size %i\n",
1356 indent + 2, "",
1357 (int) inline_summary (callee)->stack_frame_offset,
1358 (int) inline_summary (callee)->estimated_self_stack_size,
1359 (int) inline_summary (callee)->estimated_stack_size);
1360 dump_inline_edge_summary (f, indent + 2, callee, info);
1363 for (edge = node->indirect_calls; edge; edge = edge->next_callee)
1365 struct inline_edge_summary *es = inline_edge_summary (edge);
1366 fprintf (f, "%*sindirect call loop depth:%2i freq:%4i size:%2i"
1367 " time: %2i",
1368 indent, "",
1369 es->loop_depth,
1370 edge->frequency, es->call_stmt_size, es->call_stmt_time);
1371 if (es->predicate)
1373 fprintf (f, "predicate: ");
1374 dump_predicate (f, info->conds, es->predicate);
1376 else
1377 fprintf (f, "\n");
1382 void
1383 dump_inline_summary (FILE *f, struct cgraph_node *node)
1385 if (node->definition)
1387 struct inline_summary *s = inline_summary (node);
1388 size_time_entry *e;
1389 int i;
1390 fprintf (f, "Inline summary for %s/%i", node->name (),
1391 node->order);
1392 if (DECL_DISREGARD_INLINE_LIMITS (node->decl))
1393 fprintf (f, " always_inline");
1394 if (s->inlinable)
1395 fprintf (f, " inlinable");
1396 fprintf (f, "\n self time: %i\n", s->self_time);
1397 fprintf (f, " global time: %i\n", s->time);
1398 fprintf (f, " self size: %i\n", s->self_size);
1399 fprintf (f, " global size: %i\n", s->size);
1400 fprintf (f, " min size: %i\n", s->min_size);
1401 fprintf (f, " self stack: %i\n",
1402 (int) s->estimated_self_stack_size);
1403 fprintf (f, " global stack: %i\n", (int) s->estimated_stack_size);
1404 if (s->growth)
1405 fprintf (f, " estimated growth:%i\n", (int) s->growth);
1406 if (s->scc_no)
1407 fprintf (f, " In SCC: %i\n", (int) s->scc_no);
1408 for (i = 0; vec_safe_iterate (s->entry, i, &e); i++)
1410 fprintf (f, " size:%f, time:%f, predicate:",
1411 (double) e->size / INLINE_SIZE_SCALE,
1412 (double) e->time / INLINE_TIME_SCALE);
1413 dump_predicate (f, s->conds, &e->predicate);
1415 if (s->loop_iterations)
1417 fprintf (f, " loop iterations:");
1418 dump_predicate (f, s->conds, s->loop_iterations);
1420 if (s->loop_stride)
1422 fprintf (f, " loop stride:");
1423 dump_predicate (f, s->conds, s->loop_stride);
1425 if (s->array_index)
1427 fprintf (f, " array index:");
1428 dump_predicate (f, s->conds, s->array_index);
1430 fprintf (f, " calls:\n");
1431 dump_inline_edge_summary (f, 4, node, s);
1432 fprintf (f, "\n");
1436 DEBUG_FUNCTION void
1437 debug_inline_summary (struct cgraph_node *node)
1439 dump_inline_summary (stderr, node);
1442 void
1443 dump_inline_summaries (FILE *f)
1445 struct cgraph_node *node;
1447 FOR_EACH_DEFINED_FUNCTION (node)
1448 if (!node->global.inlined_to)
1449 dump_inline_summary (f, node);
1452 /* Give initial reasons why inlining would fail on EDGE. This gets either
1453 nullified or usually overwritten by more precise reasons later. */
1455 void
1456 initialize_inline_failed (struct cgraph_edge *e)
1458 struct cgraph_node *callee = e->callee;
1460 if (e->indirect_unknown_callee)
1461 e->inline_failed = CIF_INDIRECT_UNKNOWN_CALL;
1462 else if (!callee->definition)
1463 e->inline_failed = CIF_BODY_NOT_AVAILABLE;
1464 else if (callee->local.redefined_extern_inline)
1465 e->inline_failed = CIF_REDEFINED_EXTERN_INLINE;
1466 else if (e->call_stmt_cannot_inline_p)
1467 e->inline_failed = CIF_MISMATCHED_ARGUMENTS;
1468 else if (cfun && fn_contains_cilk_spawn_p (cfun))
1469 /* We can't inline if the function is spawing a function. */
1470 e->inline_failed = CIF_FUNCTION_NOT_INLINABLE;
1471 else
1472 e->inline_failed = CIF_FUNCTION_NOT_CONSIDERED;
1475 /* Callback of walk_aliased_vdefs. Flags that it has been invoked to the
1476 boolean variable pointed to by DATA. */
1478 static bool
1479 mark_modified (ao_ref *ao ATTRIBUTE_UNUSED, tree vdef ATTRIBUTE_UNUSED,
1480 void *data)
1482 bool *b = (bool *) data;
1483 *b = true;
1484 return true;
1487 /* If OP refers to value of function parameter, return the corresponding
1488 parameter. */
1490 static tree
1491 unmodified_parm_1 (gimple stmt, tree op)
1493 /* SSA_NAME referring to parm default def? */
1494 if (TREE_CODE (op) == SSA_NAME
1495 && SSA_NAME_IS_DEFAULT_DEF (op)
1496 && TREE_CODE (SSA_NAME_VAR (op)) == PARM_DECL)
1497 return SSA_NAME_VAR (op);
1498 /* Non-SSA parm reference? */
1499 if (TREE_CODE (op) == PARM_DECL)
1501 bool modified = false;
1503 ao_ref refd;
1504 ao_ref_init (&refd, op);
1505 walk_aliased_vdefs (&refd, gimple_vuse (stmt), mark_modified, &modified,
1506 NULL);
1507 if (!modified)
1508 return op;
1510 return NULL_TREE;
1513 /* If OP refers to value of function parameter, return the corresponding
1514 parameter. Also traverse chains of SSA register assignments. */
1516 static tree
1517 unmodified_parm (gimple stmt, tree op)
1519 tree res = unmodified_parm_1 (stmt, op);
1520 if (res)
1521 return res;
1523 if (TREE_CODE (op) == SSA_NAME
1524 && !SSA_NAME_IS_DEFAULT_DEF (op)
1525 && gimple_assign_single_p (SSA_NAME_DEF_STMT (op)))
1526 return unmodified_parm (SSA_NAME_DEF_STMT (op),
1527 gimple_assign_rhs1 (SSA_NAME_DEF_STMT (op)));
1528 return NULL_TREE;
1531 /* If OP refers to a value of a function parameter or value loaded from an
1532 aggregate passed to a parameter (either by value or reference), return TRUE
1533 and store the number of the parameter to *INDEX_P and information whether
1534 and how it has been loaded from an aggregate into *AGGPOS. INFO describes
1535 the function parameters, STMT is the statement in which OP is used or
1536 loaded. */
1538 static bool
1539 unmodified_parm_or_parm_agg_item (struct ipa_node_params *info,
1540 gimple stmt, tree op, int *index_p,
1541 struct agg_position_info *aggpos)
1543 tree res = unmodified_parm_1 (stmt, op);
1545 gcc_checking_assert (aggpos);
1546 if (res)
1548 *index_p = ipa_get_param_decl_index (info, res);
1549 if (*index_p < 0)
1550 return false;
1551 aggpos->agg_contents = false;
1552 aggpos->by_ref = false;
1553 return true;
1556 if (TREE_CODE (op) == SSA_NAME)
1558 if (SSA_NAME_IS_DEFAULT_DEF (op)
1559 || !gimple_assign_single_p (SSA_NAME_DEF_STMT (op)))
1560 return false;
1561 stmt = SSA_NAME_DEF_STMT (op);
1562 op = gimple_assign_rhs1 (stmt);
1563 if (!REFERENCE_CLASS_P (op))
1564 return unmodified_parm_or_parm_agg_item (info, stmt, op, index_p,
1565 aggpos);
1568 aggpos->agg_contents = true;
1569 return ipa_load_from_parm_agg (info, stmt, op, index_p, &aggpos->offset,
1570 &aggpos->by_ref);
1573 /* See if statement might disappear after inlining.
1574 0 - means not eliminated
1575 1 - half of statements goes away
1576 2 - for sure it is eliminated.
1577 We are not terribly sophisticated, basically looking for simple abstraction
1578 penalty wrappers. */
1580 static int
1581 eliminated_by_inlining_prob (gimple stmt)
1583 enum gimple_code code = gimple_code (stmt);
1584 enum tree_code rhs_code;
1586 if (!optimize)
1587 return 0;
1589 switch (code)
1591 case GIMPLE_RETURN:
1592 return 2;
1593 case GIMPLE_ASSIGN:
1594 if (gimple_num_ops (stmt) != 2)
1595 return 0;
1597 rhs_code = gimple_assign_rhs_code (stmt);
1599 /* Casts of parameters, loads from parameters passed by reference
1600 and stores to return value or parameters are often free after
1601 inlining dua to SRA and further combining.
1602 Assume that half of statements goes away. */
1603 if (rhs_code == CONVERT_EXPR
1604 || rhs_code == NOP_EXPR
1605 || rhs_code == VIEW_CONVERT_EXPR
1606 || rhs_code == ADDR_EXPR
1607 || gimple_assign_rhs_class (stmt) == GIMPLE_SINGLE_RHS)
1609 tree rhs = gimple_assign_rhs1 (stmt);
1610 tree lhs = gimple_assign_lhs (stmt);
1611 tree inner_rhs = get_base_address (rhs);
1612 tree inner_lhs = get_base_address (lhs);
1613 bool rhs_free = false;
1614 bool lhs_free = false;
1616 if (!inner_rhs)
1617 inner_rhs = rhs;
1618 if (!inner_lhs)
1619 inner_lhs = lhs;
1621 /* Reads of parameter are expected to be free. */
1622 if (unmodified_parm (stmt, inner_rhs))
1623 rhs_free = true;
1624 /* Match expressions of form &this->field. Those will most likely
1625 combine with something upstream after inlining. */
1626 else if (TREE_CODE (inner_rhs) == ADDR_EXPR)
1628 tree op = get_base_address (TREE_OPERAND (inner_rhs, 0));
1629 if (TREE_CODE (op) == PARM_DECL)
1630 rhs_free = true;
1631 else if (TREE_CODE (op) == MEM_REF
1632 && unmodified_parm (stmt, TREE_OPERAND (op, 0)))
1633 rhs_free = true;
1636 /* When parameter is not SSA register because its address is taken
1637 and it is just copied into one, the statement will be completely
1638 free after inlining (we will copy propagate backward). */
1639 if (rhs_free && is_gimple_reg (lhs))
1640 return 2;
1642 /* Reads of parameters passed by reference
1643 expected to be free (i.e. optimized out after inlining). */
1644 if (TREE_CODE (inner_rhs) == MEM_REF
1645 && unmodified_parm (stmt, TREE_OPERAND (inner_rhs, 0)))
1646 rhs_free = true;
1648 /* Copying parameter passed by reference into gimple register is
1649 probably also going to copy propagate, but we can't be quite
1650 sure. */
1651 if (rhs_free && is_gimple_reg (lhs))
1652 lhs_free = true;
1654 /* Writes to parameters, parameters passed by value and return value
1655 (either dirrectly or passed via invisible reference) are free.
1657 TODO: We ought to handle testcase like
1658 struct a {int a,b;};
1659 struct a
1660 retrurnsturct (void)
1662 struct a a ={1,2};
1663 return a;
1666 This translate into:
1668 retrurnsturct ()
1670 int a$b;
1671 int a$a;
1672 struct a a;
1673 struct a D.2739;
1675 <bb 2>:
1676 D.2739.a = 1;
1677 D.2739.b = 2;
1678 return D.2739;
1681 For that we either need to copy ipa-split logic detecting writes
1682 to return value. */
1683 if (TREE_CODE (inner_lhs) == PARM_DECL
1684 || TREE_CODE (inner_lhs) == RESULT_DECL
1685 || (TREE_CODE (inner_lhs) == MEM_REF
1686 && (unmodified_parm (stmt, TREE_OPERAND (inner_lhs, 0))
1687 || (TREE_CODE (TREE_OPERAND (inner_lhs, 0)) == SSA_NAME
1688 && SSA_NAME_VAR (TREE_OPERAND (inner_lhs, 0))
1689 && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND
1690 (inner_lhs,
1691 0))) == RESULT_DECL))))
1692 lhs_free = true;
1693 if (lhs_free
1694 && (is_gimple_reg (rhs) || is_gimple_min_invariant (rhs)))
1695 rhs_free = true;
1696 if (lhs_free && rhs_free)
1697 return 1;
1699 return 0;
1700 default:
1701 return 0;
1706 /* If BB ends by a conditional we can turn into predicates, attach corresponding
1707 predicates to the CFG edges. */
1709 static void
1710 set_cond_stmt_execution_predicate (struct ipa_node_params *info,
1711 struct inline_summary *summary,
1712 basic_block bb)
1714 gimple last;
1715 tree op;
1716 int index;
1717 struct agg_position_info aggpos;
1718 enum tree_code code, inverted_code;
1719 edge e;
1720 edge_iterator ei;
1721 gimple set_stmt;
1722 tree op2;
1724 last = last_stmt (bb);
1725 if (!last || gimple_code (last) != GIMPLE_COND)
1726 return;
1727 if (!is_gimple_ip_invariant (gimple_cond_rhs (last)))
1728 return;
1729 op = gimple_cond_lhs (last);
1730 /* TODO: handle conditionals like
1731 var = op0 < 4;
1732 if (var != 0). */
1733 if (unmodified_parm_or_parm_agg_item (info, last, op, &index, &aggpos))
1735 code = gimple_cond_code (last);
1736 inverted_code
1737 = invert_tree_comparison (code,
1738 HONOR_NANS (TYPE_MODE (TREE_TYPE (op))));
1740 FOR_EACH_EDGE (e, ei, bb->succs)
1742 enum tree_code this_code = (e->flags & EDGE_TRUE_VALUE
1743 ? code : inverted_code);
1744 /* invert_tree_comparison will return ERROR_MARK on FP
1745 comparsions that are not EQ/NE instead of returning proper
1746 unordered one. Be sure it is not confused with NON_CONSTANT. */
1747 if (this_code != ERROR_MARK)
1749 struct predicate p = add_condition (summary, index, &aggpos,
1750 this_code,
1751 gimple_cond_rhs (last));
1752 e->aux = pool_alloc (edge_predicate_pool);
1753 *(struct predicate *) e->aux = p;
1758 if (TREE_CODE (op) != SSA_NAME)
1759 return;
1760 /* Special case
1761 if (builtin_constant_p (op))
1762 constant_code
1763 else
1764 nonconstant_code.
1765 Here we can predicate nonconstant_code. We can't
1766 really handle constant_code since we have no predicate
1767 for this and also the constant code is not known to be
1768 optimized away when inliner doen't see operand is constant.
1769 Other optimizers might think otherwise. */
1770 if (gimple_cond_code (last) != NE_EXPR
1771 || !integer_zerop (gimple_cond_rhs (last)))
1772 return;
1773 set_stmt = SSA_NAME_DEF_STMT (op);
1774 if (!gimple_call_builtin_p (set_stmt, BUILT_IN_CONSTANT_P)
1775 || gimple_call_num_args (set_stmt) != 1)
1776 return;
1777 op2 = gimple_call_arg (set_stmt, 0);
1778 if (!unmodified_parm_or_parm_agg_item
1779 (info, set_stmt, op2, &index, &aggpos))
1780 return;
1781 FOR_EACH_EDGE (e, ei, bb->succs) if (e->flags & EDGE_FALSE_VALUE)
1783 struct predicate p = add_condition (summary, index, &aggpos,
1784 IS_NOT_CONSTANT, NULL_TREE);
1785 e->aux = pool_alloc (edge_predicate_pool);
1786 *(struct predicate *) e->aux = p;
1791 /* If BB ends by a switch we can turn into predicates, attach corresponding
1792 predicates to the CFG edges. */
1794 static void
1795 set_switch_stmt_execution_predicate (struct ipa_node_params *info,
1796 struct inline_summary *summary,
1797 basic_block bb)
1799 gimple last;
1800 tree op;
1801 int index;
1802 struct agg_position_info aggpos;
1803 edge e;
1804 edge_iterator ei;
1805 size_t n;
1806 size_t case_idx;
1808 last = last_stmt (bb);
1809 if (!last || gimple_code (last) != GIMPLE_SWITCH)
1810 return;
1811 op = gimple_switch_index (last);
1812 if (!unmodified_parm_or_parm_agg_item (info, last, op, &index, &aggpos))
1813 return;
1815 FOR_EACH_EDGE (e, ei, bb->succs)
1817 e->aux = pool_alloc (edge_predicate_pool);
1818 *(struct predicate *) e->aux = false_predicate ();
1820 n = gimple_switch_num_labels (last);
1821 for (case_idx = 0; case_idx < n; ++case_idx)
1823 tree cl = gimple_switch_label (last, case_idx);
1824 tree min, max;
1825 struct predicate p;
1827 e = find_edge (bb, label_to_block (CASE_LABEL (cl)));
1828 min = CASE_LOW (cl);
1829 max = CASE_HIGH (cl);
1831 /* For default we might want to construct predicate that none
1832 of cases is met, but it is bit hard to do not having negations
1833 of conditionals handy. */
1834 if (!min && !max)
1835 p = true_predicate ();
1836 else if (!max)
1837 p = add_condition (summary, index, &aggpos, EQ_EXPR, min);
1838 else
1840 struct predicate p1, p2;
1841 p1 = add_condition (summary, index, &aggpos, GE_EXPR, min);
1842 p2 = add_condition (summary, index, &aggpos, LE_EXPR, max);
1843 p = and_predicates (summary->conds, &p1, &p2);
1845 *(struct predicate *) e->aux
1846 = or_predicates (summary->conds, &p, (struct predicate *) e->aux);
1851 /* For each BB in NODE attach to its AUX pointer predicate under
1852 which it is executable. */
1854 static void
1855 compute_bb_predicates (struct cgraph_node *node,
1856 struct ipa_node_params *parms_info,
1857 struct inline_summary *summary)
1859 struct function *my_function = DECL_STRUCT_FUNCTION (node->decl);
1860 bool done = false;
1861 basic_block bb;
1863 FOR_EACH_BB_FN (bb, my_function)
1865 set_cond_stmt_execution_predicate (parms_info, summary, bb);
1866 set_switch_stmt_execution_predicate (parms_info, summary, bb);
1869 /* Entry block is always executable. */
1870 ENTRY_BLOCK_PTR_FOR_FN (my_function)->aux
1871 = pool_alloc (edge_predicate_pool);
1872 *(struct predicate *) ENTRY_BLOCK_PTR_FOR_FN (my_function)->aux
1873 = true_predicate ();
1875 /* A simple dataflow propagation of predicates forward in the CFG.
1876 TODO: work in reverse postorder. */
1877 while (!done)
1879 done = true;
1880 FOR_EACH_BB_FN (bb, my_function)
1882 struct predicate p = false_predicate ();
1883 edge e;
1884 edge_iterator ei;
1885 FOR_EACH_EDGE (e, ei, bb->preds)
1887 if (e->src->aux)
1889 struct predicate this_bb_predicate
1890 = *(struct predicate *) e->src->aux;
1891 if (e->aux)
1892 this_bb_predicate
1893 = and_predicates (summary->conds, &this_bb_predicate,
1894 (struct predicate *) e->aux);
1895 p = or_predicates (summary->conds, &p, &this_bb_predicate);
1896 if (true_predicate_p (&p))
1897 break;
1900 if (false_predicate_p (&p))
1901 gcc_assert (!bb->aux);
1902 else
1904 if (!bb->aux)
1906 done = false;
1907 bb->aux = pool_alloc (edge_predicate_pool);
1908 *((struct predicate *) bb->aux) = p;
1910 else if (!predicates_equal_p (&p, (struct predicate *) bb->aux))
1912 /* This OR operation is needed to ensure monotonous data flow
1913 in the case we hit the limit on number of clauses and the
1914 and/or operations above give approximate answers. */
1915 p = or_predicates (summary->conds, &p, (struct predicate *)bb->aux);
1916 if (!predicates_equal_p (&p, (struct predicate *) bb->aux))
1918 done = false;
1919 *((struct predicate *) bb->aux) = p;
1928 /* We keep info about constantness of SSA names. */
1930 typedef struct predicate predicate_t;
1931 /* Return predicate specifying when the STMT might have result that is not
1932 a compile time constant. */
1934 static struct predicate
1935 will_be_nonconstant_expr_predicate (struct ipa_node_params *info,
1936 struct inline_summary *summary,
1937 tree expr,
1938 vec<predicate_t> nonconstant_names)
1940 tree parm;
1941 int index;
1943 while (UNARY_CLASS_P (expr))
1944 expr = TREE_OPERAND (expr, 0);
1946 parm = unmodified_parm (NULL, expr);
1947 if (parm && (index = ipa_get_param_decl_index (info, parm)) >= 0)
1948 return add_condition (summary, index, NULL, CHANGED, NULL_TREE);
1949 if (is_gimple_min_invariant (expr))
1950 return false_predicate ();
1951 if (TREE_CODE (expr) == SSA_NAME)
1952 return nonconstant_names[SSA_NAME_VERSION (expr)];
1953 if (BINARY_CLASS_P (expr) || COMPARISON_CLASS_P (expr))
1955 struct predicate p1 = will_be_nonconstant_expr_predicate
1956 (info, summary, TREE_OPERAND (expr, 0),
1957 nonconstant_names);
1958 struct predicate p2;
1959 if (true_predicate_p (&p1))
1960 return p1;
1961 p2 = will_be_nonconstant_expr_predicate (info, summary,
1962 TREE_OPERAND (expr, 1),
1963 nonconstant_names);
1964 return or_predicates (summary->conds, &p1, &p2);
1966 else if (TREE_CODE (expr) == COND_EXPR)
1968 struct predicate p1 = will_be_nonconstant_expr_predicate
1969 (info, summary, TREE_OPERAND (expr, 0),
1970 nonconstant_names);
1971 struct predicate p2;
1972 if (true_predicate_p (&p1))
1973 return p1;
1974 p2 = will_be_nonconstant_expr_predicate (info, summary,
1975 TREE_OPERAND (expr, 1),
1976 nonconstant_names);
1977 if (true_predicate_p (&p2))
1978 return p2;
1979 p1 = or_predicates (summary->conds, &p1, &p2);
1980 p2 = will_be_nonconstant_expr_predicate (info, summary,
1981 TREE_OPERAND (expr, 2),
1982 nonconstant_names);
1983 return or_predicates (summary->conds, &p1, &p2);
1985 else
1987 debug_tree (expr);
1988 gcc_unreachable ();
1990 return false_predicate ();
1994 /* Return predicate specifying when the STMT might have result that is not
1995 a compile time constant. */
1997 static struct predicate
1998 will_be_nonconstant_predicate (struct ipa_node_params *info,
1999 struct inline_summary *summary,
2000 gimple stmt,
2001 vec<predicate_t> nonconstant_names)
2003 struct predicate p = true_predicate ();
2004 ssa_op_iter iter;
2005 tree use;
2006 struct predicate op_non_const;
2007 bool is_load;
2008 int base_index;
2009 struct agg_position_info aggpos;
2011 /* What statments might be optimized away
2012 when their arguments are constant
2013 TODO: also trivial builtins.
2014 builtin_constant_p is already handled later. */
2015 if (gimple_code (stmt) != GIMPLE_ASSIGN
2016 && gimple_code (stmt) != GIMPLE_COND
2017 && gimple_code (stmt) != GIMPLE_SWITCH)
2018 return p;
2020 /* Stores will stay anyway. */
2021 if (gimple_store_p (stmt))
2022 return p;
2024 is_load = gimple_assign_load_p (stmt);
2026 /* Loads can be optimized when the value is known. */
2027 if (is_load)
2029 tree op;
2030 gcc_assert (gimple_assign_single_p (stmt));
2031 op = gimple_assign_rhs1 (stmt);
2032 if (!unmodified_parm_or_parm_agg_item (info, stmt, op, &base_index,
2033 &aggpos))
2034 return p;
2036 else
2037 base_index = -1;
2039 /* See if we understand all operands before we start
2040 adding conditionals. */
2041 FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE)
2043 tree parm = unmodified_parm (stmt, use);
2044 /* For arguments we can build a condition. */
2045 if (parm && ipa_get_param_decl_index (info, parm) >= 0)
2046 continue;
2047 if (TREE_CODE (use) != SSA_NAME)
2048 return p;
2049 /* If we know when operand is constant,
2050 we still can say something useful. */
2051 if (!true_predicate_p (&nonconstant_names[SSA_NAME_VERSION (use)]))
2052 continue;
2053 return p;
2056 if (is_load)
2057 op_non_const =
2058 add_condition (summary, base_index, &aggpos, CHANGED, NULL);
2059 else
2060 op_non_const = false_predicate ();
2061 FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE)
2063 tree parm = unmodified_parm (stmt, use);
2064 int index;
2066 if (parm && (index = ipa_get_param_decl_index (info, parm)) >= 0)
2068 if (index != base_index)
2069 p = add_condition (summary, index, NULL, CHANGED, NULL_TREE);
2070 else
2071 continue;
2073 else
2074 p = nonconstant_names[SSA_NAME_VERSION (use)];
2075 op_non_const = or_predicates (summary->conds, &p, &op_non_const);
2077 if (gimple_code (stmt) == GIMPLE_ASSIGN
2078 && TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME)
2079 nonconstant_names[SSA_NAME_VERSION (gimple_assign_lhs (stmt))]
2080 = op_non_const;
2081 return op_non_const;
2084 struct record_modified_bb_info
2086 bitmap bb_set;
2087 gimple stmt;
2090 /* Callback of walk_aliased_vdefs. Records basic blocks where the value may be
2091 set except for info->stmt. */
2093 static bool
2094 record_modified (ao_ref *ao ATTRIBUTE_UNUSED, tree vdef, void *data)
2096 struct record_modified_bb_info *info =
2097 (struct record_modified_bb_info *) data;
2098 if (SSA_NAME_DEF_STMT (vdef) == info->stmt)
2099 return false;
2100 bitmap_set_bit (info->bb_set,
2101 SSA_NAME_IS_DEFAULT_DEF (vdef)
2102 ? ENTRY_BLOCK_PTR_FOR_FN (cfun)->index
2103 : gimple_bb (SSA_NAME_DEF_STMT (vdef))->index);
2104 return false;
2107 /* Return probability (based on REG_BR_PROB_BASE) that I-th parameter of STMT
2108 will change since last invocation of STMT.
2110 Value 0 is reserved for compile time invariants.
2111 For common parameters it is REG_BR_PROB_BASE. For loop invariants it
2112 ought to be REG_BR_PROB_BASE / estimated_iters. */
2114 static int
2115 param_change_prob (gimple stmt, int i)
2117 tree op = gimple_call_arg (stmt, i);
2118 basic_block bb = gimple_bb (stmt);
2119 tree base;
2121 /* Global invariants neve change. */
2122 if (is_gimple_min_invariant (op))
2123 return 0;
2124 /* We would have to do non-trivial analysis to really work out what
2125 is the probability of value to change (i.e. when init statement
2126 is in a sibling loop of the call).
2128 We do an conservative estimate: when call is executed N times more often
2129 than the statement defining value, we take the frequency 1/N. */
2130 if (TREE_CODE (op) == SSA_NAME)
2132 int init_freq;
2134 if (!bb->frequency)
2135 return REG_BR_PROB_BASE;
2137 if (SSA_NAME_IS_DEFAULT_DEF (op))
2138 init_freq = ENTRY_BLOCK_PTR_FOR_FN (cfun)->frequency;
2139 else
2140 init_freq = gimple_bb (SSA_NAME_DEF_STMT (op))->frequency;
2142 if (!init_freq)
2143 init_freq = 1;
2144 if (init_freq < bb->frequency)
2145 return MAX (GCOV_COMPUTE_SCALE (init_freq, bb->frequency), 1);
2146 else
2147 return REG_BR_PROB_BASE;
2150 base = get_base_address (op);
2151 if (base)
2153 ao_ref refd;
2154 int max;
2155 struct record_modified_bb_info info;
2156 bitmap_iterator bi;
2157 unsigned index;
2158 tree init = ctor_for_folding (base);
2160 if (init != error_mark_node)
2161 return 0;
2162 if (!bb->frequency)
2163 return REG_BR_PROB_BASE;
2164 ao_ref_init (&refd, op);
2165 info.stmt = stmt;
2166 info.bb_set = BITMAP_ALLOC (NULL);
2167 walk_aliased_vdefs (&refd, gimple_vuse (stmt), record_modified, &info,
2168 NULL);
2169 if (bitmap_bit_p (info.bb_set, bb->index))
2171 BITMAP_FREE (info.bb_set);
2172 return REG_BR_PROB_BASE;
2175 /* Assume that every memory is initialized at entry.
2176 TODO: Can we easilly determine if value is always defined
2177 and thus we may skip entry block? */
2178 if (ENTRY_BLOCK_PTR_FOR_FN (cfun)->frequency)
2179 max = ENTRY_BLOCK_PTR_FOR_FN (cfun)->frequency;
2180 else
2181 max = 1;
2183 EXECUTE_IF_SET_IN_BITMAP (info.bb_set, 0, index, bi)
2184 max = MIN (max, BASIC_BLOCK_FOR_FN (cfun, index)->frequency);
2186 BITMAP_FREE (info.bb_set);
2187 if (max < bb->frequency)
2188 return MAX (GCOV_COMPUTE_SCALE (max, bb->frequency), 1);
2189 else
2190 return REG_BR_PROB_BASE;
2192 return REG_BR_PROB_BASE;
2195 /* Find whether a basic block BB is the final block of a (half) diamond CFG
2196 sub-graph and if the predicate the condition depends on is known. If so,
2197 return true and store the pointer the predicate in *P. */
2199 static bool
2200 phi_result_unknown_predicate (struct ipa_node_params *info,
2201 struct inline_summary *summary, basic_block bb,
2202 struct predicate *p,
2203 vec<predicate_t> nonconstant_names)
2205 edge e;
2206 edge_iterator ei;
2207 basic_block first_bb = NULL;
2208 gimple stmt;
2210 if (single_pred_p (bb))
2212 *p = false_predicate ();
2213 return true;
2216 FOR_EACH_EDGE (e, ei, bb->preds)
2218 if (single_succ_p (e->src))
2220 if (!single_pred_p (e->src))
2221 return false;
2222 if (!first_bb)
2223 first_bb = single_pred (e->src);
2224 else if (single_pred (e->src) != first_bb)
2225 return false;
2227 else
2229 if (!first_bb)
2230 first_bb = e->src;
2231 else if (e->src != first_bb)
2232 return false;
2236 if (!first_bb)
2237 return false;
2239 stmt = last_stmt (first_bb);
2240 if (!stmt
2241 || gimple_code (stmt) != GIMPLE_COND
2242 || !is_gimple_ip_invariant (gimple_cond_rhs (stmt)))
2243 return false;
2245 *p = will_be_nonconstant_expr_predicate (info, summary,
2246 gimple_cond_lhs (stmt),
2247 nonconstant_names);
2248 if (true_predicate_p (p))
2249 return false;
2250 else
2251 return true;
2254 /* Given a PHI statement in a function described by inline properties SUMMARY
2255 and *P being the predicate describing whether the selected PHI argument is
2256 known, store a predicate for the result of the PHI statement into
2257 NONCONSTANT_NAMES, if possible. */
2259 static void
2260 predicate_for_phi_result (struct inline_summary *summary, gimple phi,
2261 struct predicate *p,
2262 vec<predicate_t> nonconstant_names)
2264 unsigned i;
2266 for (i = 0; i < gimple_phi_num_args (phi); i++)
2268 tree arg = gimple_phi_arg (phi, i)->def;
2269 if (!is_gimple_min_invariant (arg))
2271 gcc_assert (TREE_CODE (arg) == SSA_NAME);
2272 *p = or_predicates (summary->conds, p,
2273 &nonconstant_names[SSA_NAME_VERSION (arg)]);
2274 if (true_predicate_p (p))
2275 return;
2279 if (dump_file && (dump_flags & TDF_DETAILS))
2281 fprintf (dump_file, "\t\tphi predicate: ");
2282 dump_predicate (dump_file, summary->conds, p);
2284 nonconstant_names[SSA_NAME_VERSION (gimple_phi_result (phi))] = *p;
2287 /* Return predicate specifying when array index in access OP becomes non-constant. */
2289 static struct predicate
2290 array_index_predicate (struct inline_summary *info,
2291 vec< predicate_t> nonconstant_names, tree op)
2293 struct predicate p = false_predicate ();
2294 while (handled_component_p (op))
2296 if (TREE_CODE (op) == ARRAY_REF || TREE_CODE (op) == ARRAY_RANGE_REF)
2298 if (TREE_CODE (TREE_OPERAND (op, 1)) == SSA_NAME)
2299 p = or_predicates (info->conds, &p,
2300 &nonconstant_names[SSA_NAME_VERSION
2301 (TREE_OPERAND (op, 1))]);
2303 op = TREE_OPERAND (op, 0);
2305 return p;
2308 /* For a typical usage of __builtin_expect (a<b, 1), we
2309 may introduce an extra relation stmt:
2310 With the builtin, we have
2311 t1 = a <= b;
2312 t2 = (long int) t1;
2313 t3 = __builtin_expect (t2, 1);
2314 if (t3 != 0)
2315 goto ...
2316 Without the builtin, we have
2317 if (a<=b)
2318 goto...
2319 This affects the size/time estimation and may have
2320 an impact on the earlier inlining.
2321 Here find this pattern and fix it up later. */
2323 static gimple
2324 find_foldable_builtin_expect (basic_block bb)
2326 gimple_stmt_iterator bsi;
2328 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
2330 gimple stmt = gsi_stmt (bsi);
2331 if (gimple_call_builtin_p (stmt, BUILT_IN_EXPECT)
2332 || (is_gimple_call (stmt)
2333 && gimple_call_internal_p (stmt)
2334 && gimple_call_internal_fn (stmt) == IFN_BUILTIN_EXPECT))
2336 tree var = gimple_call_lhs (stmt);
2337 tree arg = gimple_call_arg (stmt, 0);
2338 use_operand_p use_p;
2339 gimple use_stmt;
2340 bool match = false;
2341 bool done = false;
2343 if (!var || !arg)
2344 continue;
2345 gcc_assert (TREE_CODE (var) == SSA_NAME);
2347 while (TREE_CODE (arg) == SSA_NAME)
2349 gimple stmt_tmp = SSA_NAME_DEF_STMT (arg);
2350 if (!is_gimple_assign (stmt_tmp))
2351 break;
2352 switch (gimple_assign_rhs_code (stmt_tmp))
2354 case LT_EXPR:
2355 case LE_EXPR:
2356 case GT_EXPR:
2357 case GE_EXPR:
2358 case EQ_EXPR:
2359 case NE_EXPR:
2360 match = true;
2361 done = true;
2362 break;
2363 case NOP_EXPR:
2364 break;
2365 default:
2366 done = true;
2367 break;
2369 if (done)
2370 break;
2371 arg = gimple_assign_rhs1 (stmt_tmp);
2374 if (match && single_imm_use (var, &use_p, &use_stmt)
2375 && gimple_code (use_stmt) == GIMPLE_COND)
2376 return use_stmt;
2379 return NULL;
2382 /* Return true when the basic blocks contains only clobbers followed by RESX.
2383 Such BBs are kept around to make removal of dead stores possible with
2384 presence of EH and will be optimized out by optimize_clobbers later in the
2385 game.
2387 NEED_EH is used to recurse in case the clobber has non-EH predecestors
2388 that can be clobber only, too.. When it is false, the RESX is not necessary
2389 on the end of basic block. */
2391 static bool
2392 clobber_only_eh_bb_p (basic_block bb, bool need_eh = true)
2394 gimple_stmt_iterator gsi = gsi_last_bb (bb);
2395 edge_iterator ei;
2396 edge e;
2398 if (need_eh)
2400 if (gsi_end_p (gsi))
2401 return false;
2402 if (gimple_code (gsi_stmt (gsi)) != GIMPLE_RESX)
2403 return false;
2404 gsi_prev (&gsi);
2406 else if (!single_succ_p (bb))
2407 return false;
2409 for (; !gsi_end_p (gsi); gsi_prev (&gsi))
2411 gimple stmt = gsi_stmt (gsi);
2412 if (is_gimple_debug (stmt))
2413 continue;
2414 if (gimple_clobber_p (stmt))
2415 continue;
2416 if (gimple_code (stmt) == GIMPLE_LABEL)
2417 break;
2418 return false;
2421 /* See if all predecestors are either throws or clobber only BBs. */
2422 FOR_EACH_EDGE (e, ei, bb->preds)
2423 if (!(e->flags & EDGE_EH)
2424 && !clobber_only_eh_bb_p (e->src, false))
2425 return false;
2427 return true;
2430 /* Compute function body size parameters for NODE.
2431 When EARLY is true, we compute only simple summaries without
2432 non-trivial predicates to drive the early inliner. */
2434 static void
2435 estimate_function_body_sizes (struct cgraph_node *node, bool early)
2437 gcov_type time = 0;
2438 /* Estimate static overhead for function prologue/epilogue and alignment. */
2439 int size = 2;
2440 /* Benefits are scaled by probability of elimination that is in range
2441 <0,2>. */
2442 basic_block bb;
2443 gimple_stmt_iterator bsi;
2444 struct function *my_function = DECL_STRUCT_FUNCTION (node->decl);
2445 int freq;
2446 struct inline_summary *info = inline_summary (node);
2447 struct predicate bb_predicate;
2448 struct ipa_node_params *parms_info = NULL;
2449 vec<predicate_t> nonconstant_names = vNULL;
2450 int nblocks, n;
2451 int *order;
2452 predicate array_index = true_predicate ();
2453 gimple fix_builtin_expect_stmt;
2455 info->conds = NULL;
2456 info->entry = NULL;
2458 if (optimize && !early)
2460 calculate_dominance_info (CDI_DOMINATORS);
2461 loop_optimizer_init (LOOPS_NORMAL | LOOPS_HAVE_RECORDED_EXITS);
2463 if (ipa_node_params_vector.exists ())
2465 parms_info = IPA_NODE_REF (node);
2466 nonconstant_names.safe_grow_cleared
2467 (SSANAMES (my_function)->length ());
2471 if (dump_file)
2472 fprintf (dump_file, "\nAnalyzing function body size: %s\n",
2473 node->name ());
2475 /* When we run into maximal number of entries, we assign everything to the
2476 constant truth case. Be sure to have it in list. */
2477 bb_predicate = true_predicate ();
2478 account_size_time (info, 0, 0, &bb_predicate);
2480 bb_predicate = not_inlined_predicate ();
2481 account_size_time (info, 2 * INLINE_SIZE_SCALE, 0, &bb_predicate);
2483 gcc_assert (my_function && my_function->cfg);
2484 if (parms_info)
2485 compute_bb_predicates (node, parms_info, info);
2486 gcc_assert (cfun == my_function);
2487 order = XNEWVEC (int, n_basic_blocks_for_fn (cfun));
2488 nblocks = pre_and_rev_post_order_compute (NULL, order, false);
2489 for (n = 0; n < nblocks; n++)
2491 bb = BASIC_BLOCK_FOR_FN (cfun, order[n]);
2492 freq = compute_call_stmt_bb_frequency (node->decl, bb);
2493 if (clobber_only_eh_bb_p (bb))
2495 if (dump_file && (dump_flags & TDF_DETAILS))
2496 fprintf (dump_file, "\n Ignoring BB %i;"
2497 " it will be optimized away by cleanup_clobbers\n",
2498 bb->index);
2499 continue;
2502 /* TODO: Obviously predicates can be propagated down across CFG. */
2503 if (parms_info)
2505 if (bb->aux)
2506 bb_predicate = *(struct predicate *) bb->aux;
2507 else
2508 bb_predicate = false_predicate ();
2510 else
2511 bb_predicate = true_predicate ();
2513 if (dump_file && (dump_flags & TDF_DETAILS))
2515 fprintf (dump_file, "\n BB %i predicate:", bb->index);
2516 dump_predicate (dump_file, info->conds, &bb_predicate);
2519 if (parms_info && nonconstant_names.exists ())
2521 struct predicate phi_predicate;
2522 bool first_phi = true;
2524 for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi))
2526 if (first_phi
2527 && !phi_result_unknown_predicate (parms_info, info, bb,
2528 &phi_predicate,
2529 nonconstant_names))
2530 break;
2531 first_phi = false;
2532 if (dump_file && (dump_flags & TDF_DETAILS))
2534 fprintf (dump_file, " ");
2535 print_gimple_stmt (dump_file, gsi_stmt (bsi), 0, 0);
2537 predicate_for_phi_result (info, gsi_stmt (bsi), &phi_predicate,
2538 nonconstant_names);
2542 fix_builtin_expect_stmt = find_foldable_builtin_expect (bb);
2544 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
2546 gimple stmt = gsi_stmt (bsi);
2547 int this_size = estimate_num_insns (stmt, &eni_size_weights);
2548 int this_time = estimate_num_insns (stmt, &eni_time_weights);
2549 int prob;
2550 struct predicate will_be_nonconstant;
2552 /* This relation stmt should be folded after we remove
2553 buildin_expect call. Adjust the cost here. */
2554 if (stmt == fix_builtin_expect_stmt)
2556 this_size--;
2557 this_time--;
2560 if (dump_file && (dump_flags & TDF_DETAILS))
2562 fprintf (dump_file, " ");
2563 print_gimple_stmt (dump_file, stmt, 0, 0);
2564 fprintf (dump_file, "\t\tfreq:%3.2f size:%3i time:%3i\n",
2565 ((double) freq) / CGRAPH_FREQ_BASE, this_size,
2566 this_time);
2569 if (gimple_assign_load_p (stmt) && nonconstant_names.exists ())
2571 struct predicate this_array_index;
2572 this_array_index =
2573 array_index_predicate (info, nonconstant_names,
2574 gimple_assign_rhs1 (stmt));
2575 if (!false_predicate_p (&this_array_index))
2576 array_index =
2577 and_predicates (info->conds, &array_index,
2578 &this_array_index);
2580 if (gimple_store_p (stmt) && nonconstant_names.exists ())
2582 struct predicate this_array_index;
2583 this_array_index =
2584 array_index_predicate (info, nonconstant_names,
2585 gimple_get_lhs (stmt));
2586 if (!false_predicate_p (&this_array_index))
2587 array_index =
2588 and_predicates (info->conds, &array_index,
2589 &this_array_index);
2593 if (is_gimple_call (stmt)
2594 && !gimple_call_internal_p (stmt))
2596 struct cgraph_edge *edge = cgraph_edge (node, stmt);
2597 struct inline_edge_summary *es = inline_edge_summary (edge);
2599 /* Special case: results of BUILT_IN_CONSTANT_P will be always
2600 resolved as constant. We however don't want to optimize
2601 out the cgraph edges. */
2602 if (nonconstant_names.exists ()
2603 && gimple_call_builtin_p (stmt, BUILT_IN_CONSTANT_P)
2604 && gimple_call_lhs (stmt)
2605 && TREE_CODE (gimple_call_lhs (stmt)) == SSA_NAME)
2607 struct predicate false_p = false_predicate ();
2608 nonconstant_names[SSA_NAME_VERSION (gimple_call_lhs (stmt))]
2609 = false_p;
2611 if (ipa_node_params_vector.exists ())
2613 int count = gimple_call_num_args (stmt);
2614 int i;
2616 if (count)
2617 es->param.safe_grow_cleared (count);
2618 for (i = 0; i < count; i++)
2620 int prob = param_change_prob (stmt, i);
2621 gcc_assert (prob >= 0 && prob <= REG_BR_PROB_BASE);
2622 es->param[i].change_prob = prob;
2626 es->call_stmt_size = this_size;
2627 es->call_stmt_time = this_time;
2628 es->loop_depth = bb_loop_depth (bb);
2629 edge_set_predicate (edge, &bb_predicate);
2632 /* TODO: When conditional jump or swithc is known to be constant, but
2633 we did not translate it into the predicates, we really can account
2634 just maximum of the possible paths. */
2635 if (parms_info)
2636 will_be_nonconstant
2637 = will_be_nonconstant_predicate (parms_info, info,
2638 stmt, nonconstant_names);
2639 if (this_time || this_size)
2641 struct predicate p;
2643 this_time *= freq;
2645 prob = eliminated_by_inlining_prob (stmt);
2646 if (prob == 1 && dump_file && (dump_flags & TDF_DETAILS))
2647 fprintf (dump_file,
2648 "\t\t50%% will be eliminated by inlining\n");
2649 if (prob == 2 && dump_file && (dump_flags & TDF_DETAILS))
2650 fprintf (dump_file, "\t\tWill be eliminated by inlining\n");
2652 if (parms_info)
2653 p = and_predicates (info->conds, &bb_predicate,
2654 &will_be_nonconstant);
2655 else
2656 p = true_predicate ();
2658 if (!false_predicate_p (&p))
2660 time += this_time;
2661 size += this_size;
2662 if (time > MAX_TIME * INLINE_TIME_SCALE)
2663 time = MAX_TIME * INLINE_TIME_SCALE;
2666 /* We account everything but the calls. Calls have their own
2667 size/time info attached to cgraph edges. This is necessary
2668 in order to make the cost disappear after inlining. */
2669 if (!is_gimple_call (stmt))
2671 if (prob)
2673 struct predicate ip = not_inlined_predicate ();
2674 ip = and_predicates (info->conds, &ip, &p);
2675 account_size_time (info, this_size * prob,
2676 this_time * prob, &ip);
2678 if (prob != 2)
2679 account_size_time (info, this_size * (2 - prob),
2680 this_time * (2 - prob), &p);
2683 gcc_assert (time >= 0);
2684 gcc_assert (size >= 0);
2688 set_hint_predicate (&inline_summary (node)->array_index, array_index);
2689 time = (time + CGRAPH_FREQ_BASE / 2) / CGRAPH_FREQ_BASE;
2690 if (time > MAX_TIME)
2691 time = MAX_TIME;
2692 free (order);
2694 if (!early && nonconstant_names.exists ())
2696 struct loop *loop;
2697 predicate loop_iterations = true_predicate ();
2698 predicate loop_stride = true_predicate ();
2700 if (dump_file && (dump_flags & TDF_DETAILS))
2701 flow_loops_dump (dump_file, NULL, 0);
2702 scev_initialize ();
2703 FOR_EACH_LOOP (loop, 0)
2705 vec<edge> exits;
2706 edge ex;
2707 unsigned int j, i;
2708 struct tree_niter_desc niter_desc;
2709 basic_block *body = get_loop_body (loop);
2710 bb_predicate = *(struct predicate *) loop->header->aux;
2712 exits = get_loop_exit_edges (loop);
2713 FOR_EACH_VEC_ELT (exits, j, ex)
2714 if (number_of_iterations_exit (loop, ex, &niter_desc, false)
2715 && !is_gimple_min_invariant (niter_desc.niter))
2717 predicate will_be_nonconstant
2718 = will_be_nonconstant_expr_predicate (parms_info, info,
2719 niter_desc.niter,
2720 nonconstant_names);
2721 if (!true_predicate_p (&will_be_nonconstant))
2722 will_be_nonconstant = and_predicates (info->conds,
2723 &bb_predicate,
2724 &will_be_nonconstant);
2725 if (!true_predicate_p (&will_be_nonconstant)
2726 && !false_predicate_p (&will_be_nonconstant))
2727 /* This is slightly inprecise. We may want to represent each
2728 loop with independent predicate. */
2729 loop_iterations =
2730 and_predicates (info->conds, &loop_iterations,
2731 &will_be_nonconstant);
2733 exits.release ();
2735 for (i = 0; i < loop->num_nodes; i++)
2737 gimple_stmt_iterator gsi;
2738 bb_predicate = *(struct predicate *) body[i]->aux;
2739 for (gsi = gsi_start_bb (body[i]); !gsi_end_p (gsi);
2740 gsi_next (&gsi))
2742 gimple stmt = gsi_stmt (gsi);
2743 affine_iv iv;
2744 ssa_op_iter iter;
2745 tree use;
2747 FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE)
2749 predicate will_be_nonconstant;
2751 if (!simple_iv
2752 (loop, loop_containing_stmt (stmt), use, &iv, true)
2753 || is_gimple_min_invariant (iv.step))
2754 continue;
2755 will_be_nonconstant
2756 = will_be_nonconstant_expr_predicate (parms_info, info,
2757 iv.step,
2758 nonconstant_names);
2759 if (!true_predicate_p (&will_be_nonconstant))
2760 will_be_nonconstant
2761 = and_predicates (info->conds,
2762 &bb_predicate,
2763 &will_be_nonconstant);
2764 if (!true_predicate_p (&will_be_nonconstant)
2765 && !false_predicate_p (&will_be_nonconstant))
2766 /* This is slightly inprecise. We may want to represent
2767 each loop with independent predicate. */
2768 loop_stride =
2769 and_predicates (info->conds, &loop_stride,
2770 &will_be_nonconstant);
2774 free (body);
2776 set_hint_predicate (&inline_summary (node)->loop_iterations,
2777 loop_iterations);
2778 set_hint_predicate (&inline_summary (node)->loop_stride, loop_stride);
2779 scev_finalize ();
2781 FOR_ALL_BB_FN (bb, my_function)
2783 edge e;
2784 edge_iterator ei;
2786 if (bb->aux)
2787 pool_free (edge_predicate_pool, bb->aux);
2788 bb->aux = NULL;
2789 FOR_EACH_EDGE (e, ei, bb->succs)
2791 if (e->aux)
2792 pool_free (edge_predicate_pool, e->aux);
2793 e->aux = NULL;
2796 inline_summary (node)->self_time = time;
2797 inline_summary (node)->self_size = size;
2798 nonconstant_names.release ();
2799 if (optimize && !early)
2801 loop_optimizer_finalize ();
2802 free_dominance_info (CDI_DOMINATORS);
2804 if (dump_file)
2806 fprintf (dump_file, "\n");
2807 dump_inline_summary (dump_file, node);
2812 /* Compute parameters of functions used by inliner.
2813 EARLY is true when we compute parameters for the early inliner */
2815 void
2816 compute_inline_parameters (struct cgraph_node *node, bool early)
2818 HOST_WIDE_INT self_stack_size;
2819 struct cgraph_edge *e;
2820 struct inline_summary *info;
2822 gcc_assert (!node->global.inlined_to);
2824 inline_summary_alloc ();
2826 info = inline_summary (node);
2827 reset_inline_summary (node);
2829 /* FIXME: Thunks are inlinable, but tree-inline don't know how to do that.
2830 Once this happen, we will need to more curefully predict call
2831 statement size. */
2832 if (node->thunk.thunk_p)
2834 struct inline_edge_summary *es = inline_edge_summary (node->callees);
2835 struct predicate t = true_predicate ();
2837 info->inlinable = 0;
2838 node->callees->call_stmt_cannot_inline_p = true;
2839 node->local.can_change_signature = false;
2840 es->call_stmt_time = 1;
2841 es->call_stmt_size = 1;
2842 account_size_time (info, 0, 0, &t);
2843 return;
2846 /* Even is_gimple_min_invariant rely on current_function_decl. */
2847 push_cfun (DECL_STRUCT_FUNCTION (node->decl));
2849 /* Estimate the stack size for the function if we're optimizing. */
2850 self_stack_size = optimize ? estimated_stack_frame_size (node) : 0;
2851 info->estimated_self_stack_size = self_stack_size;
2852 info->estimated_stack_size = self_stack_size;
2853 info->stack_frame_offset = 0;
2855 /* Can this function be inlined at all? */
2856 if (!optimize && !lookup_attribute ("always_inline",
2857 DECL_ATTRIBUTES (node->decl)))
2858 info->inlinable = false;
2859 else
2860 info->inlinable = tree_inlinable_function_p (node->decl);
2862 /* Type attributes can use parameter indices to describe them. */
2863 if (TYPE_ATTRIBUTES (TREE_TYPE (node->decl)))
2864 node->local.can_change_signature = false;
2865 else
2867 /* Otherwise, inlinable functions always can change signature. */
2868 if (info->inlinable)
2869 node->local.can_change_signature = true;
2870 else
2872 /* Functions calling builtin_apply can not change signature. */
2873 for (e = node->callees; e; e = e->next_callee)
2875 tree cdecl = e->callee->decl;
2876 if (DECL_BUILT_IN (cdecl)
2877 && DECL_BUILT_IN_CLASS (cdecl) == BUILT_IN_NORMAL
2878 && (DECL_FUNCTION_CODE (cdecl) == BUILT_IN_APPLY_ARGS
2879 || DECL_FUNCTION_CODE (cdecl) == BUILT_IN_VA_START))
2880 break;
2882 node->local.can_change_signature = !e;
2885 estimate_function_body_sizes (node, early);
2887 for (e = node->callees; e; e = e->next_callee)
2888 if (symtab_comdat_local_p (e->callee))
2889 break;
2890 node->calls_comdat_local = (e != NULL);
2892 /* Inlining characteristics are maintained by the cgraph_mark_inline. */
2893 info->time = info->self_time;
2894 info->size = info->self_size;
2895 info->stack_frame_offset = 0;
2896 info->estimated_stack_size = info->estimated_self_stack_size;
2897 #ifdef ENABLE_CHECKING
2898 inline_update_overall_summary (node);
2899 gcc_assert (info->time == info->self_time && info->size == info->self_size);
2900 #endif
2902 pop_cfun ();
2906 /* Compute parameters of functions used by inliner using
2907 current_function_decl. */
2909 static unsigned int
2910 compute_inline_parameters_for_current (void)
2912 compute_inline_parameters (cgraph_get_node (current_function_decl), true);
2913 return 0;
2916 namespace {
2918 const pass_data pass_data_inline_parameters =
2920 GIMPLE_PASS, /* type */
2921 "inline_param", /* name */
2922 OPTGROUP_INLINE, /* optinfo_flags */
2923 false, /* has_gate */
2924 true, /* has_execute */
2925 TV_INLINE_PARAMETERS, /* tv_id */
2926 0, /* properties_required */
2927 0, /* properties_provided */
2928 0, /* properties_destroyed */
2929 0, /* todo_flags_start */
2930 0, /* todo_flags_finish */
2933 class pass_inline_parameters : public gimple_opt_pass
2935 public:
2936 pass_inline_parameters (gcc::context *ctxt)
2937 : gimple_opt_pass (pass_data_inline_parameters, ctxt)
2940 /* opt_pass methods: */
2941 opt_pass * clone () { return new pass_inline_parameters (m_ctxt); }
2942 unsigned int execute () {
2943 return compute_inline_parameters_for_current ();
2946 }; // class pass_inline_parameters
2948 } // anon namespace
2950 gimple_opt_pass *
2951 make_pass_inline_parameters (gcc::context *ctxt)
2953 return new pass_inline_parameters (ctxt);
2957 /* Estimate benefit devirtualizing indirect edge IE, provided KNOWN_VALS and
2958 KNOWN_BINFOS. */
2960 static bool
2961 estimate_edge_devirt_benefit (struct cgraph_edge *ie,
2962 int *size, int *time,
2963 vec<tree> known_vals,
2964 vec<tree> known_binfos,
2965 vec<ipa_agg_jump_function_p> known_aggs)
2967 tree target;
2968 struct cgraph_node *callee;
2969 struct inline_summary *isummary;
2971 if (!known_vals.exists () && !known_binfos.exists ())
2972 return false;
2973 if (!flag_indirect_inlining)
2974 return false;
2976 target = ipa_get_indirect_edge_target (ie, known_vals, known_binfos,
2977 known_aggs);
2978 if (!target)
2979 return false;
2981 /* Account for difference in cost between indirect and direct calls. */
2982 *size -= (eni_size_weights.indirect_call_cost - eni_size_weights.call_cost);
2983 *time -= (eni_time_weights.indirect_call_cost - eni_time_weights.call_cost);
2984 gcc_checking_assert (*time >= 0);
2985 gcc_checking_assert (*size >= 0);
2987 callee = cgraph_get_node (target);
2988 if (!callee || !callee->definition)
2989 return false;
2990 isummary = inline_summary (callee);
2991 return isummary->inlinable;
2994 /* Increase SIZE, MIN_SIZE (if non-NULL) and TIME for size and time needed to
2995 handle edge E with probability PROB.
2996 Set HINTS if edge may be devirtualized.
2997 KNOWN_VALS, KNOWN_AGGS and KNOWN_BINFOS describe context of the call
2998 site. */
3000 static inline void
3001 estimate_edge_size_and_time (struct cgraph_edge *e, int *size, int *min_size,
3002 int *time,
3003 int prob,
3004 vec<tree> known_vals,
3005 vec<tree> known_binfos,
3006 vec<ipa_agg_jump_function_p> known_aggs,
3007 inline_hints *hints)
3009 struct inline_edge_summary *es = inline_edge_summary (e);
3010 int call_size = es->call_stmt_size;
3011 int call_time = es->call_stmt_time;
3012 int cur_size;
3013 if (!e->callee
3014 && estimate_edge_devirt_benefit (e, &call_size, &call_time,
3015 known_vals, known_binfos, known_aggs)
3016 && hints && cgraph_maybe_hot_edge_p (e))
3017 *hints |= INLINE_HINT_indirect_call;
3018 cur_size = call_size * INLINE_SIZE_SCALE;
3019 *size += cur_size;
3020 if (min_size)
3021 *min_size += cur_size;
3022 *time += apply_probability ((gcov_type) call_time, prob)
3023 * e->frequency * (INLINE_TIME_SCALE / CGRAPH_FREQ_BASE);
3024 if (*time > MAX_TIME * INLINE_TIME_SCALE)
3025 *time = MAX_TIME * INLINE_TIME_SCALE;
3030 /* Increase SIZE, MIN_SIZE and TIME for size and time needed to handle all
3031 calls in NODE.
3032 POSSIBLE_TRUTHS, KNOWN_VALS, KNOWN_AGGS and KNOWN_BINFOS describe context of
3033 the call site. */
3035 static void
3036 estimate_calls_size_and_time (struct cgraph_node *node, int *size,
3037 int *min_size, int *time,
3038 inline_hints *hints,
3039 clause_t possible_truths,
3040 vec<tree> known_vals,
3041 vec<tree> known_binfos,
3042 vec<ipa_agg_jump_function_p> known_aggs)
3044 struct cgraph_edge *e;
3045 for (e = node->callees; e; e = e->next_callee)
3047 struct inline_edge_summary *es = inline_edge_summary (e);
3048 if (!es->predicate
3049 || evaluate_predicate (es->predicate, possible_truths))
3051 if (e->inline_failed)
3053 /* Predicates of calls shall not use NOT_CHANGED codes,
3054 sowe do not need to compute probabilities. */
3055 estimate_edge_size_and_time (e, size,
3056 es->predicate ? NULL : min_size,
3057 time, REG_BR_PROB_BASE,
3058 known_vals, known_binfos,
3059 known_aggs, hints);
3061 else
3062 estimate_calls_size_and_time (e->callee, size, min_size, time,
3063 hints,
3064 possible_truths,
3065 known_vals, known_binfos,
3066 known_aggs);
3069 for (e = node->indirect_calls; e; e = e->next_callee)
3071 struct inline_edge_summary *es = inline_edge_summary (e);
3072 if (!es->predicate
3073 || evaluate_predicate (es->predicate, possible_truths))
3074 estimate_edge_size_and_time (e, size,
3075 es->predicate ? NULL : min_size,
3076 time, REG_BR_PROB_BASE,
3077 known_vals, known_binfos, known_aggs,
3078 hints);
3083 /* Estimate size and time needed to execute NODE assuming
3084 POSSIBLE_TRUTHS clause, and KNOWN_VALS, KNOWN_AGGS and KNOWN_BINFOS
3085 information about NODE's arguments. If non-NULL use also probability
3086 information present in INLINE_PARAM_SUMMARY vector.
3087 Additionally detemine hints determined by the context. Finally compute
3088 minimal size needed for the call that is independent on the call context and
3089 can be used for fast estimates. Return the values in RET_SIZE,
3090 RET_MIN_SIZE, RET_TIME and RET_HINTS. */
3092 static void
3093 estimate_node_size_and_time (struct cgraph_node *node,
3094 clause_t possible_truths,
3095 vec<tree> known_vals,
3096 vec<tree> known_binfos,
3097 vec<ipa_agg_jump_function_p> known_aggs,
3098 int *ret_size, int *ret_min_size, int *ret_time,
3099 inline_hints *ret_hints,
3100 vec<inline_param_summary>
3101 inline_param_summary)
3103 struct inline_summary *info = inline_summary (node);
3104 size_time_entry *e;
3105 int size = 0;
3106 int time = 0;
3107 int min_size = 0;
3108 inline_hints hints = 0;
3109 int i;
3111 if (dump_file && (dump_flags & TDF_DETAILS))
3113 bool found = false;
3114 fprintf (dump_file, " Estimating body: %s/%i\n"
3115 " Known to be false: ", node->name (),
3116 node->order);
3118 for (i = predicate_not_inlined_condition;
3119 i < (predicate_first_dynamic_condition
3120 + (int) vec_safe_length (info->conds)); i++)
3121 if (!(possible_truths & (1 << i)))
3123 if (found)
3124 fprintf (dump_file, ", ");
3125 found = true;
3126 dump_condition (dump_file, info->conds, i);
3130 for (i = 0; vec_safe_iterate (info->entry, i, &e); i++)
3131 if (evaluate_predicate (&e->predicate, possible_truths))
3133 size += e->size;
3134 gcc_checking_assert (e->time >= 0);
3135 gcc_checking_assert (time >= 0);
3136 if (!inline_param_summary.exists ())
3137 time += e->time;
3138 else
3140 int prob = predicate_probability (info->conds,
3141 &e->predicate,
3142 possible_truths,
3143 inline_param_summary);
3144 gcc_checking_assert (prob >= 0);
3145 gcc_checking_assert (prob <= REG_BR_PROB_BASE);
3146 time += apply_probability ((gcov_type) e->time, prob);
3148 if (time > MAX_TIME * INLINE_TIME_SCALE)
3149 time = MAX_TIME * INLINE_TIME_SCALE;
3150 gcc_checking_assert (time >= 0);
3153 gcc_checking_assert (true_predicate_p (&(*info->entry)[0].predicate));
3154 min_size = (*info->entry)[0].size;
3155 gcc_checking_assert (size >= 0);
3156 gcc_checking_assert (time >= 0);
3158 if (info->loop_iterations
3159 && !evaluate_predicate (info->loop_iterations, possible_truths))
3160 hints |= INLINE_HINT_loop_iterations;
3161 if (info->loop_stride
3162 && !evaluate_predicate (info->loop_stride, possible_truths))
3163 hints |= INLINE_HINT_loop_stride;
3164 if (info->array_index
3165 && !evaluate_predicate (info->array_index, possible_truths))
3166 hints |= INLINE_HINT_array_index;
3167 if (info->scc_no)
3168 hints |= INLINE_HINT_in_scc;
3169 if (DECL_DECLARED_INLINE_P (node->decl))
3170 hints |= INLINE_HINT_declared_inline;
3172 estimate_calls_size_and_time (node, &size, &min_size, &time, &hints, possible_truths,
3173 known_vals, known_binfos, known_aggs);
3174 gcc_checking_assert (size >= 0);
3175 gcc_checking_assert (time >= 0);
3176 time = RDIV (time, INLINE_TIME_SCALE);
3177 size = RDIV (size, INLINE_SIZE_SCALE);
3178 min_size = RDIV (min_size, INLINE_SIZE_SCALE);
3180 if (dump_file && (dump_flags & TDF_DETAILS))
3181 fprintf (dump_file, "\n size:%i time:%i\n", (int) size, (int) time);
3182 if (ret_time)
3183 *ret_time = time;
3184 if (ret_size)
3185 *ret_size = size;
3186 if (ret_min_size)
3187 *ret_min_size = min_size;
3188 if (ret_hints)
3189 *ret_hints = hints;
3190 return;
3194 /* Estimate size and time needed to execute callee of EDGE assuming that
3195 parameters known to be constant at caller of EDGE are propagated.
3196 KNOWN_VALS and KNOWN_BINFOS are vectors of assumed known constant values
3197 and types for parameters. */
3199 void
3200 estimate_ipcp_clone_size_and_time (struct cgraph_node *node,
3201 vec<tree> known_vals,
3202 vec<tree> known_binfos,
3203 vec<ipa_agg_jump_function_p> known_aggs,
3204 int *ret_size, int *ret_time,
3205 inline_hints *hints)
3207 clause_t clause;
3209 clause = evaluate_conditions_for_known_args (node, false, known_vals,
3210 known_aggs);
3211 estimate_node_size_and_time (node, clause, known_vals, known_binfos,
3212 known_aggs, ret_size, NULL, ret_time, hints, vNULL);
3215 /* Translate all conditions from callee representation into caller
3216 representation and symbolically evaluate predicate P into new predicate.
3218 INFO is inline_summary of function we are adding predicate into, CALLEE_INFO
3219 is summary of function predicate P is from. OPERAND_MAP is array giving
3220 callee formal IDs the caller formal IDs. POSSSIBLE_TRUTHS is clausule of all
3221 callee conditions that may be true in caller context. TOPLEV_PREDICATE is
3222 predicate under which callee is executed. OFFSET_MAP is an array of of
3223 offsets that need to be added to conditions, negative offset means that
3224 conditions relying on values passed by reference have to be discarded
3225 because they might not be preserved (and should be considered offset zero
3226 for other purposes). */
3228 static struct predicate
3229 remap_predicate (struct inline_summary *info,
3230 struct inline_summary *callee_info,
3231 struct predicate *p,
3232 vec<int> operand_map,
3233 vec<int> offset_map,
3234 clause_t possible_truths, struct predicate *toplev_predicate)
3236 int i;
3237 struct predicate out = true_predicate ();
3239 /* True predicate is easy. */
3240 if (true_predicate_p (p))
3241 return *toplev_predicate;
3242 for (i = 0; p->clause[i]; i++)
3244 clause_t clause = p->clause[i];
3245 int cond;
3246 struct predicate clause_predicate = false_predicate ();
3248 gcc_assert (i < MAX_CLAUSES);
3250 for (cond = 0; cond < NUM_CONDITIONS; cond++)
3251 /* Do we have condition we can't disprove? */
3252 if (clause & possible_truths & (1 << cond))
3254 struct predicate cond_predicate;
3255 /* Work out if the condition can translate to predicate in the
3256 inlined function. */
3257 if (cond >= predicate_first_dynamic_condition)
3259 struct condition *c;
3261 c = &(*callee_info->conds)[cond
3263 predicate_first_dynamic_condition];
3264 /* See if we can remap condition operand to caller's operand.
3265 Otherwise give up. */
3266 if (!operand_map.exists ()
3267 || (int) operand_map.length () <= c->operand_num
3268 || operand_map[c->operand_num] == -1
3269 /* TODO: For non-aggregate conditions, adding an offset is
3270 basically an arithmetic jump function processing which
3271 we should support in future. */
3272 || ((!c->agg_contents || !c->by_ref)
3273 && offset_map[c->operand_num] > 0)
3274 || (c->agg_contents && c->by_ref
3275 && offset_map[c->operand_num] < 0))
3276 cond_predicate = true_predicate ();
3277 else
3279 struct agg_position_info ap;
3280 HOST_WIDE_INT offset_delta = offset_map[c->operand_num];
3281 if (offset_delta < 0)
3283 gcc_checking_assert (!c->agg_contents || !c->by_ref);
3284 offset_delta = 0;
3286 gcc_assert (!c->agg_contents
3287 || c->by_ref || offset_delta == 0);
3288 ap.offset = c->offset + offset_delta;
3289 ap.agg_contents = c->agg_contents;
3290 ap.by_ref = c->by_ref;
3291 cond_predicate = add_condition (info,
3292 operand_map[c->operand_num],
3293 &ap, c->code, c->val);
3296 /* Fixed conditions remains same, construct single
3297 condition predicate. */
3298 else
3300 cond_predicate.clause[0] = 1 << cond;
3301 cond_predicate.clause[1] = 0;
3303 clause_predicate = or_predicates (info->conds, &clause_predicate,
3304 &cond_predicate);
3306 out = and_predicates (info->conds, &out, &clause_predicate);
3308 return and_predicates (info->conds, &out, toplev_predicate);
3312 /* Update summary information of inline clones after inlining.
3313 Compute peak stack usage. */
3315 static void
3316 inline_update_callee_summaries (struct cgraph_node *node, int depth)
3318 struct cgraph_edge *e;
3319 struct inline_summary *callee_info = inline_summary (node);
3320 struct inline_summary *caller_info = inline_summary (node->callers->caller);
3321 HOST_WIDE_INT peak;
3323 callee_info->stack_frame_offset
3324 = caller_info->stack_frame_offset
3325 + caller_info->estimated_self_stack_size;
3326 peak = callee_info->stack_frame_offset
3327 + callee_info->estimated_self_stack_size;
3328 if (inline_summary (node->global.inlined_to)->estimated_stack_size < peak)
3329 inline_summary (node->global.inlined_to)->estimated_stack_size = peak;
3330 ipa_propagate_frequency (node);
3331 for (e = node->callees; e; e = e->next_callee)
3333 if (!e->inline_failed)
3334 inline_update_callee_summaries (e->callee, depth);
3335 inline_edge_summary (e)->loop_depth += depth;
3337 for (e = node->indirect_calls; e; e = e->next_callee)
3338 inline_edge_summary (e)->loop_depth += depth;
3341 /* Update change_prob of EDGE after INLINED_EDGE has been inlined.
3342 When functoin A is inlined in B and A calls C with parameter that
3343 changes with probability PROB1 and C is known to be passthroug
3344 of argument if B that change with probability PROB2, the probability
3345 of change is now PROB1*PROB2. */
3347 static void
3348 remap_edge_change_prob (struct cgraph_edge *inlined_edge,
3349 struct cgraph_edge *edge)
3351 if (ipa_node_params_vector.exists ())
3353 int i;
3354 struct ipa_edge_args *args = IPA_EDGE_REF (edge);
3355 struct inline_edge_summary *es = inline_edge_summary (edge);
3356 struct inline_edge_summary *inlined_es
3357 = inline_edge_summary (inlined_edge);
3359 for (i = 0; i < ipa_get_cs_argument_count (args); i++)
3361 struct ipa_jump_func *jfunc = ipa_get_ith_jump_func (args, i);
3362 if (jfunc->type == IPA_JF_PASS_THROUGH
3363 && (ipa_get_jf_pass_through_formal_id (jfunc)
3364 < (int) inlined_es->param.length ()))
3366 int jf_formal_id = ipa_get_jf_pass_through_formal_id (jfunc);
3367 int prob1 = es->param[i].change_prob;
3368 int prob2 = inlined_es->param[jf_formal_id].change_prob;
3369 int prob = combine_probabilities (prob1, prob2);
3371 if (prob1 && prob2 && !prob)
3372 prob = 1;
3374 es->param[i].change_prob = prob;
3380 /* Update edge summaries of NODE after INLINED_EDGE has been inlined.
3382 Remap predicates of callees of NODE. Rest of arguments match
3383 remap_predicate.
3385 Also update change probabilities. */
3387 static void
3388 remap_edge_summaries (struct cgraph_edge *inlined_edge,
3389 struct cgraph_node *node,
3390 struct inline_summary *info,
3391 struct inline_summary *callee_info,
3392 vec<int> operand_map,
3393 vec<int> offset_map,
3394 clause_t possible_truths,
3395 struct predicate *toplev_predicate)
3397 struct cgraph_edge *e;
3398 for (e = node->callees; e; e = e->next_callee)
3400 struct inline_edge_summary *es = inline_edge_summary (e);
3401 struct predicate p;
3403 if (e->inline_failed)
3405 remap_edge_change_prob (inlined_edge, e);
3407 if (es->predicate)
3409 p = remap_predicate (info, callee_info,
3410 es->predicate, operand_map, offset_map,
3411 possible_truths, toplev_predicate);
3412 edge_set_predicate (e, &p);
3413 /* TODO: We should remove the edge for code that will be
3414 optimized out, but we need to keep verifiers and tree-inline
3415 happy. Make it cold for now. */
3416 if (false_predicate_p (&p))
3418 e->count = 0;
3419 e->frequency = 0;
3422 else
3423 edge_set_predicate (e, toplev_predicate);
3425 else
3426 remap_edge_summaries (inlined_edge, e->callee, info, callee_info,
3427 operand_map, offset_map, possible_truths,
3428 toplev_predicate);
3430 for (e = node->indirect_calls; e; e = e->next_callee)
3432 struct inline_edge_summary *es = inline_edge_summary (e);
3433 struct predicate p;
3435 remap_edge_change_prob (inlined_edge, e);
3436 if (es->predicate)
3438 p = remap_predicate (info, callee_info,
3439 es->predicate, operand_map, offset_map,
3440 possible_truths, toplev_predicate);
3441 edge_set_predicate (e, &p);
3442 /* TODO: We should remove the edge for code that will be optimized
3443 out, but we need to keep verifiers and tree-inline happy.
3444 Make it cold for now. */
3445 if (false_predicate_p (&p))
3447 e->count = 0;
3448 e->frequency = 0;
3451 else
3452 edge_set_predicate (e, toplev_predicate);
3456 /* Same as remap_predicate, but set result into hint *HINT. */
3458 static void
3459 remap_hint_predicate (struct inline_summary *info,
3460 struct inline_summary *callee_info,
3461 struct predicate **hint,
3462 vec<int> operand_map,
3463 vec<int> offset_map,
3464 clause_t possible_truths,
3465 struct predicate *toplev_predicate)
3467 predicate p;
3469 if (!*hint)
3470 return;
3471 p = remap_predicate (info, callee_info,
3472 *hint,
3473 operand_map, offset_map,
3474 possible_truths, toplev_predicate);
3475 if (!false_predicate_p (&p) && !true_predicate_p (&p))
3477 if (!*hint)
3478 set_hint_predicate (hint, p);
3479 else
3480 **hint = and_predicates (info->conds, *hint, &p);
3484 /* We inlined EDGE. Update summary of the function we inlined into. */
3486 void
3487 inline_merge_summary (struct cgraph_edge *edge)
3489 struct inline_summary *callee_info = inline_summary (edge->callee);
3490 struct cgraph_node *to = (edge->caller->global.inlined_to
3491 ? edge->caller->global.inlined_to : edge->caller);
3492 struct inline_summary *info = inline_summary (to);
3493 clause_t clause = 0; /* not_inline is known to be false. */
3494 size_time_entry *e;
3495 vec<int> operand_map = vNULL;
3496 vec<int> offset_map = vNULL;
3497 int i;
3498 struct predicate toplev_predicate;
3499 struct predicate true_p = true_predicate ();
3500 struct inline_edge_summary *es = inline_edge_summary (edge);
3502 if (es->predicate)
3503 toplev_predicate = *es->predicate;
3504 else
3505 toplev_predicate = true_predicate ();
3507 if (ipa_node_params_vector.exists () && callee_info->conds)
3509 struct ipa_edge_args *args = IPA_EDGE_REF (edge);
3510 int count = ipa_get_cs_argument_count (args);
3511 int i;
3513 evaluate_properties_for_edge (edge, true, &clause, NULL, NULL, NULL);
3514 if (count)
3516 operand_map.safe_grow_cleared (count);
3517 offset_map.safe_grow_cleared (count);
3519 for (i = 0; i < count; i++)
3521 struct ipa_jump_func *jfunc = ipa_get_ith_jump_func (args, i);
3522 int map = -1;
3524 /* TODO: handle non-NOPs when merging. */
3525 if (jfunc->type == IPA_JF_PASS_THROUGH)
3527 if (ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
3528 map = ipa_get_jf_pass_through_formal_id (jfunc);
3529 if (!ipa_get_jf_pass_through_agg_preserved (jfunc))
3530 offset_map[i] = -1;
3532 else if (jfunc->type == IPA_JF_ANCESTOR)
3534 HOST_WIDE_INT offset = ipa_get_jf_ancestor_offset (jfunc);
3535 if (offset >= 0 && offset < INT_MAX)
3537 map = ipa_get_jf_ancestor_formal_id (jfunc);
3538 if (!ipa_get_jf_ancestor_agg_preserved (jfunc))
3539 offset = -1;
3540 offset_map[i] = offset;
3543 operand_map[i] = map;
3544 gcc_assert (map < ipa_get_param_count (IPA_NODE_REF (to)));
3547 for (i = 0; vec_safe_iterate (callee_info->entry, i, &e); i++)
3549 struct predicate p = remap_predicate (info, callee_info,
3550 &e->predicate, operand_map,
3551 offset_map, clause,
3552 &toplev_predicate);
3553 if (!false_predicate_p (&p))
3555 gcov_type add_time = ((gcov_type) e->time * edge->frequency
3556 + CGRAPH_FREQ_BASE / 2) / CGRAPH_FREQ_BASE;
3557 int prob = predicate_probability (callee_info->conds,
3558 &e->predicate,
3559 clause, es->param);
3560 add_time = apply_probability ((gcov_type) add_time, prob);
3561 if (add_time > MAX_TIME * INLINE_TIME_SCALE)
3562 add_time = MAX_TIME * INLINE_TIME_SCALE;
3563 if (prob != REG_BR_PROB_BASE
3564 && dump_file && (dump_flags & TDF_DETAILS))
3566 fprintf (dump_file, "\t\tScaling time by probability:%f\n",
3567 (double) prob / REG_BR_PROB_BASE);
3569 account_size_time (info, e->size, add_time, &p);
3572 remap_edge_summaries (edge, edge->callee, info, callee_info, operand_map,
3573 offset_map, clause, &toplev_predicate);
3574 remap_hint_predicate (info, callee_info,
3575 &callee_info->loop_iterations,
3576 operand_map, offset_map, clause, &toplev_predicate);
3577 remap_hint_predicate (info, callee_info,
3578 &callee_info->loop_stride,
3579 operand_map, offset_map, clause, &toplev_predicate);
3580 remap_hint_predicate (info, callee_info,
3581 &callee_info->array_index,
3582 operand_map, offset_map, clause, &toplev_predicate);
3584 inline_update_callee_summaries (edge->callee,
3585 inline_edge_summary (edge)->loop_depth);
3587 /* We do not maintain predicates of inlined edges, free it. */
3588 edge_set_predicate (edge, &true_p);
3589 /* Similarly remove param summaries. */
3590 es->param.release ();
3591 operand_map.release ();
3592 offset_map.release ();
3595 /* For performance reasons inline_merge_summary is not updating overall size
3596 and time. Recompute it. */
3598 void
3599 inline_update_overall_summary (struct cgraph_node *node)
3601 struct inline_summary *info = inline_summary (node);
3602 size_time_entry *e;
3603 int i;
3605 info->size = 0;
3606 info->time = 0;
3607 for (i = 0; vec_safe_iterate (info->entry, i, &e); i++)
3609 info->size += e->size, info->time += e->time;
3610 if (info->time > MAX_TIME * INLINE_TIME_SCALE)
3611 info->time = MAX_TIME * INLINE_TIME_SCALE;
3613 estimate_calls_size_and_time (node, &info->size, &info->min_size,
3614 &info->time, NULL,
3615 ~(clause_t) (1 << predicate_false_condition),
3616 vNULL, vNULL, vNULL);
3617 info->time = (info->time + INLINE_TIME_SCALE / 2) / INLINE_TIME_SCALE;
3618 info->size = (info->size + INLINE_SIZE_SCALE / 2) / INLINE_SIZE_SCALE;
3621 /* Return hints derrived from EDGE. */
3623 simple_edge_hints (struct cgraph_edge *edge)
3625 int hints = 0;
3626 struct cgraph_node *to = (edge->caller->global.inlined_to
3627 ? edge->caller->global.inlined_to : edge->caller);
3628 if (inline_summary (to)->scc_no
3629 && inline_summary (to)->scc_no == inline_summary (edge->callee)->scc_no
3630 && !cgraph_edge_recursive_p (edge))
3631 hints |= INLINE_HINT_same_scc;
3633 if (to->lto_file_data && edge->callee->lto_file_data
3634 && to->lto_file_data != edge->callee->lto_file_data)
3635 hints |= INLINE_HINT_cross_module;
3637 return hints;
3640 /* Estimate the time cost for the caller when inlining EDGE.
3641 Only to be called via estimate_edge_time, that handles the
3642 caching mechanism.
3644 When caching, also update the cache entry. Compute both time and
3645 size, since we always need both metrics eventually. */
3648 do_estimate_edge_time (struct cgraph_edge *edge)
3650 int time;
3651 int size;
3652 inline_hints hints;
3653 struct cgraph_node *callee;
3654 clause_t clause;
3655 vec<tree> known_vals;
3656 vec<tree> known_binfos;
3657 vec<ipa_agg_jump_function_p> known_aggs;
3658 struct inline_edge_summary *es = inline_edge_summary (edge);
3659 int min_size;
3661 callee = cgraph_function_or_thunk_node (edge->callee, NULL);
3663 gcc_checking_assert (edge->inline_failed);
3664 evaluate_properties_for_edge (edge, true,
3665 &clause, &known_vals, &known_binfos,
3666 &known_aggs);
3667 estimate_node_size_and_time (callee, clause, known_vals, known_binfos,
3668 known_aggs, &size, &min_size, &time, &hints, es->param);
3669 known_vals.release ();
3670 known_binfos.release ();
3671 known_aggs.release ();
3672 gcc_checking_assert (size >= 0);
3673 gcc_checking_assert (time >= 0);
3675 /* When caching, update the cache entry. */
3676 if (edge_growth_cache.exists ())
3678 inline_summary (edge->callee)->min_size = min_size;
3679 if ((int) edge_growth_cache.length () <= edge->uid)
3680 edge_growth_cache.safe_grow_cleared (cgraph_edge_max_uid);
3681 edge_growth_cache[edge->uid].time = time + (time >= 0);
3683 edge_growth_cache[edge->uid].size = size + (size >= 0);
3684 hints |= simple_edge_hints (edge);
3685 edge_growth_cache[edge->uid].hints = hints + 1;
3687 return time;
3691 /* Return estimated callee growth after inlining EDGE.
3692 Only to be called via estimate_edge_size. */
3695 do_estimate_edge_size (struct cgraph_edge *edge)
3697 int size;
3698 struct cgraph_node *callee;
3699 clause_t clause;
3700 vec<tree> known_vals;
3701 vec<tree> known_binfos;
3702 vec<ipa_agg_jump_function_p> known_aggs;
3704 /* When we do caching, use do_estimate_edge_time to populate the entry. */
3706 if (edge_growth_cache.exists ())
3708 do_estimate_edge_time (edge);
3709 size = edge_growth_cache[edge->uid].size;
3710 gcc_checking_assert (size);
3711 return size - (size > 0);
3714 callee = cgraph_function_or_thunk_node (edge->callee, NULL);
3716 /* Early inliner runs without caching, go ahead and do the dirty work. */
3717 gcc_checking_assert (edge->inline_failed);
3718 evaluate_properties_for_edge (edge, true,
3719 &clause, &known_vals, &known_binfos,
3720 &known_aggs);
3721 estimate_node_size_and_time (callee, clause, known_vals, known_binfos,
3722 known_aggs, &size, NULL, NULL, NULL, vNULL);
3723 known_vals.release ();
3724 known_binfos.release ();
3725 known_aggs.release ();
3726 return size;
3730 /* Estimate the growth of the caller when inlining EDGE.
3731 Only to be called via estimate_edge_size. */
3733 inline_hints
3734 do_estimate_edge_hints (struct cgraph_edge *edge)
3736 inline_hints hints;
3737 struct cgraph_node *callee;
3738 clause_t clause;
3739 vec<tree> known_vals;
3740 vec<tree> known_binfos;
3741 vec<ipa_agg_jump_function_p> known_aggs;
3743 /* When we do caching, use do_estimate_edge_time to populate the entry. */
3745 if (edge_growth_cache.exists ())
3747 do_estimate_edge_time (edge);
3748 hints = edge_growth_cache[edge->uid].hints;
3749 gcc_checking_assert (hints);
3750 return hints - 1;
3753 callee = cgraph_function_or_thunk_node (edge->callee, NULL);
3755 /* Early inliner runs without caching, go ahead and do the dirty work. */
3756 gcc_checking_assert (edge->inline_failed);
3757 evaluate_properties_for_edge (edge, true,
3758 &clause, &known_vals, &known_binfos,
3759 &known_aggs);
3760 estimate_node_size_and_time (callee, clause, known_vals, known_binfos,
3761 known_aggs, NULL, NULL, NULL, &hints, vNULL);
3762 known_vals.release ();
3763 known_binfos.release ();
3764 known_aggs.release ();
3765 hints |= simple_edge_hints (edge);
3766 return hints;
3770 /* Estimate self time of the function NODE after inlining EDGE. */
3773 estimate_time_after_inlining (struct cgraph_node *node,
3774 struct cgraph_edge *edge)
3776 struct inline_edge_summary *es = inline_edge_summary (edge);
3777 if (!es->predicate || !false_predicate_p (es->predicate))
3779 gcov_type time =
3780 inline_summary (node)->time + estimate_edge_time (edge);
3781 if (time < 0)
3782 time = 0;
3783 if (time > MAX_TIME)
3784 time = MAX_TIME;
3785 return time;
3787 return inline_summary (node)->time;
3791 /* Estimate the size of NODE after inlining EDGE which should be an
3792 edge to either NODE or a call inlined into NODE. */
3795 estimate_size_after_inlining (struct cgraph_node *node,
3796 struct cgraph_edge *edge)
3798 struct inline_edge_summary *es = inline_edge_summary (edge);
3799 if (!es->predicate || !false_predicate_p (es->predicate))
3801 int size = inline_summary (node)->size + estimate_edge_growth (edge);
3802 gcc_assert (size >= 0);
3803 return size;
3805 return inline_summary (node)->size;
3809 struct growth_data
3811 struct cgraph_node *node;
3812 bool self_recursive;
3813 int growth;
3817 /* Worker for do_estimate_growth. Collect growth for all callers. */
3819 static bool
3820 do_estimate_growth_1 (struct cgraph_node *node, void *data)
3822 struct cgraph_edge *e;
3823 struct growth_data *d = (struct growth_data *) data;
3825 for (e = node->callers; e; e = e->next_caller)
3827 gcc_checking_assert (e->inline_failed);
3829 if (e->caller == d->node
3830 || (e->caller->global.inlined_to
3831 && e->caller->global.inlined_to == d->node))
3832 d->self_recursive = true;
3833 d->growth += estimate_edge_growth (e);
3835 return false;
3839 /* Estimate the growth caused by inlining NODE into all callees. */
3842 do_estimate_growth (struct cgraph_node *node)
3844 struct growth_data d = { node, 0, false };
3845 struct inline_summary *info = inline_summary (node);
3847 cgraph_for_node_and_aliases (node, do_estimate_growth_1, &d, true);
3849 /* For self recursive functions the growth estimation really should be
3850 infinity. We don't want to return very large values because the growth
3851 plays various roles in badness computation fractions. Be sure to not
3852 return zero or negative growths. */
3853 if (d.self_recursive)
3854 d.growth = d.growth < info->size ? info->size : d.growth;
3855 else if (DECL_EXTERNAL (node->decl))
3857 else
3859 if (cgraph_will_be_removed_from_program_if_no_direct_calls (node))
3860 d.growth -= info->size;
3861 /* COMDAT functions are very often not shared across multiple units
3862 since they come from various template instantiations.
3863 Take this into account. */
3864 else if (DECL_COMDAT (node->decl)
3865 && cgraph_can_remove_if_no_direct_calls_p (node))
3866 d.growth -= (info->size
3867 * (100 - PARAM_VALUE (PARAM_COMDAT_SHARING_PROBABILITY))
3868 + 50) / 100;
3871 if (node_growth_cache.exists ())
3873 if ((int) node_growth_cache.length () <= node->uid)
3874 node_growth_cache.safe_grow_cleared (cgraph_max_uid);
3875 node_growth_cache[node->uid] = d.growth + (d.growth >= 0);
3877 return d.growth;
3881 /* Make cheap estimation if growth of NODE is likely positive knowing
3882 EDGE_GROWTH of one particular edge.
3883 We assume that most of other edges will have similar growth
3884 and skip computation if there are too many callers. */
3886 bool
3887 growth_likely_positive (struct cgraph_node *node, int edge_growth ATTRIBUTE_UNUSED)
3889 int max_callers;
3890 int ret;
3891 struct cgraph_edge *e;
3892 gcc_checking_assert (edge_growth > 0);
3894 /* Unlike for functions called once, we play unsafe with
3895 COMDATs. We can allow that since we know functions
3896 in consideration are small (and thus risk is small) and
3897 moreover grow estimates already accounts that COMDAT
3898 functions may or may not disappear when eliminated from
3899 current unit. With good probability making aggressive
3900 choice in all units is going to make overall program
3901 smaller.
3903 Consequently we ask cgraph_can_remove_if_no_direct_calls_p
3904 instead of
3905 cgraph_will_be_removed_from_program_if_no_direct_calls */
3906 if (DECL_EXTERNAL (node->decl)
3907 || !cgraph_can_remove_if_no_direct_calls_p (node))
3908 return true;
3910 /* If there is cached value, just go ahead. */
3911 if ((int)node_growth_cache.length () > node->uid
3912 && (ret = node_growth_cache[node->uid]))
3913 return ret > 0;
3914 if (!cgraph_will_be_removed_from_program_if_no_direct_calls (node)
3915 && (!DECL_COMDAT (node->decl)
3916 || !cgraph_can_remove_if_no_direct_calls_p (node)))
3917 return true;
3918 max_callers = inline_summary (node)->size * 4 / edge_growth + 2;
3920 for (e = node->callers; e; e = e->next_caller)
3922 max_callers--;
3923 if (!max_callers)
3924 return true;
3926 return estimate_growth (node) > 0;
3930 /* This function performs intraprocedural analysis in NODE that is required to
3931 inline indirect calls. */
3933 static void
3934 inline_indirect_intraprocedural_analysis (struct cgraph_node *node)
3936 ipa_analyze_node (node);
3937 if (dump_file && (dump_flags & TDF_DETAILS))
3939 ipa_print_node_params (dump_file, node);
3940 ipa_print_node_jump_functions (dump_file, node);
3945 /* Note function body size. */
3947 static void
3948 inline_analyze_function (struct cgraph_node *node)
3950 push_cfun (DECL_STRUCT_FUNCTION (node->decl));
3952 if (dump_file)
3953 fprintf (dump_file, "\nAnalyzing function: %s/%u\n",
3954 node->name (), node->order);
3955 if (optimize && !node->thunk.thunk_p)
3956 inline_indirect_intraprocedural_analysis (node);
3957 compute_inline_parameters (node, false);
3958 if (!optimize)
3960 struct cgraph_edge *e;
3961 for (e = node->callees; e; e = e->next_callee)
3963 if (e->inline_failed == CIF_FUNCTION_NOT_CONSIDERED)
3964 e->inline_failed = CIF_FUNCTION_NOT_OPTIMIZED;
3965 e->call_stmt_cannot_inline_p = true;
3967 for (e = node->indirect_calls; e; e = e->next_callee)
3969 if (e->inline_failed == CIF_FUNCTION_NOT_CONSIDERED)
3970 e->inline_failed = CIF_FUNCTION_NOT_OPTIMIZED;
3971 e->call_stmt_cannot_inline_p = true;
3975 pop_cfun ();
3979 /* Called when new function is inserted to callgraph late. */
3981 static void
3982 add_new_function (struct cgraph_node *node, void *data ATTRIBUTE_UNUSED)
3984 inline_analyze_function (node);
3988 /* Note function body size. */
3990 void
3991 inline_generate_summary (void)
3993 struct cgraph_node *node;
3995 /* When not optimizing, do not bother to analyze. Inlining is still done
3996 because edge redirection needs to happen there. */
3997 if (!optimize && !flag_lto && !flag_wpa)
3998 return;
4000 function_insertion_hook_holder =
4001 cgraph_add_function_insertion_hook (&add_new_function, NULL);
4003 ipa_register_cgraph_hooks ();
4004 inline_free_summary ();
4006 FOR_EACH_DEFINED_FUNCTION (node)
4007 if (!node->alias)
4008 inline_analyze_function (node);
4012 /* Read predicate from IB. */
4014 static struct predicate
4015 read_predicate (struct lto_input_block *ib)
4017 struct predicate out;
4018 clause_t clause;
4019 int k = 0;
4023 gcc_assert (k <= MAX_CLAUSES);
4024 clause = out.clause[k++] = streamer_read_uhwi (ib);
4026 while (clause);
4028 /* Zero-initialize the remaining clauses in OUT. */
4029 while (k <= MAX_CLAUSES)
4030 out.clause[k++] = 0;
4032 return out;
4036 /* Write inline summary for edge E to OB. */
4038 static void
4039 read_inline_edge_summary (struct lto_input_block *ib, struct cgraph_edge *e)
4041 struct inline_edge_summary *es = inline_edge_summary (e);
4042 struct predicate p;
4043 int length, i;
4045 es->call_stmt_size = streamer_read_uhwi (ib);
4046 es->call_stmt_time = streamer_read_uhwi (ib);
4047 es->loop_depth = streamer_read_uhwi (ib);
4048 p = read_predicate (ib);
4049 edge_set_predicate (e, &p);
4050 length = streamer_read_uhwi (ib);
4051 if (length)
4053 es->param.safe_grow_cleared (length);
4054 for (i = 0; i < length; i++)
4055 es->param[i].change_prob = streamer_read_uhwi (ib);
4060 /* Stream in inline summaries from the section. */
4062 static void
4063 inline_read_section (struct lto_file_decl_data *file_data, const char *data,
4064 size_t len)
4066 const struct lto_function_header *header =
4067 (const struct lto_function_header *) data;
4068 const int cfg_offset = sizeof (struct lto_function_header);
4069 const int main_offset = cfg_offset + header->cfg_size;
4070 const int string_offset = main_offset + header->main_size;
4071 struct data_in *data_in;
4072 struct lto_input_block ib;
4073 unsigned int i, count2, j;
4074 unsigned int f_count;
4076 LTO_INIT_INPUT_BLOCK (ib, (const char *) data + main_offset, 0,
4077 header->main_size);
4079 data_in =
4080 lto_data_in_create (file_data, (const char *) data + string_offset,
4081 header->string_size, vNULL);
4082 f_count = streamer_read_uhwi (&ib);
4083 for (i = 0; i < f_count; i++)
4085 unsigned int index;
4086 struct cgraph_node *node;
4087 struct inline_summary *info;
4088 lto_symtab_encoder_t encoder;
4089 struct bitpack_d bp;
4090 struct cgraph_edge *e;
4091 predicate p;
4093 index = streamer_read_uhwi (&ib);
4094 encoder = file_data->symtab_node_encoder;
4095 node = cgraph (lto_symtab_encoder_deref (encoder, index));
4096 info = inline_summary (node);
4098 info->estimated_stack_size
4099 = info->estimated_self_stack_size = streamer_read_uhwi (&ib);
4100 info->size = info->self_size = streamer_read_uhwi (&ib);
4101 info->time = info->self_time = streamer_read_uhwi (&ib);
4103 bp = streamer_read_bitpack (&ib);
4104 info->inlinable = bp_unpack_value (&bp, 1);
4106 count2 = streamer_read_uhwi (&ib);
4107 gcc_assert (!info->conds);
4108 for (j = 0; j < count2; j++)
4110 struct condition c;
4111 c.operand_num = streamer_read_uhwi (&ib);
4112 c.code = (enum tree_code) streamer_read_uhwi (&ib);
4113 c.val = stream_read_tree (&ib, data_in);
4114 bp = streamer_read_bitpack (&ib);
4115 c.agg_contents = bp_unpack_value (&bp, 1);
4116 c.by_ref = bp_unpack_value (&bp, 1);
4117 if (c.agg_contents)
4118 c.offset = streamer_read_uhwi (&ib);
4119 vec_safe_push (info->conds, c);
4121 count2 = streamer_read_uhwi (&ib);
4122 gcc_assert (!info->entry);
4123 for (j = 0; j < count2; j++)
4125 struct size_time_entry e;
4127 e.size = streamer_read_uhwi (&ib);
4128 e.time = streamer_read_uhwi (&ib);
4129 e.predicate = read_predicate (&ib);
4131 vec_safe_push (info->entry, e);
4134 p = read_predicate (&ib);
4135 set_hint_predicate (&info->loop_iterations, p);
4136 p = read_predicate (&ib);
4137 set_hint_predicate (&info->loop_stride, p);
4138 p = read_predicate (&ib);
4139 set_hint_predicate (&info->array_index, p);
4140 for (e = node->callees; e; e = e->next_callee)
4141 read_inline_edge_summary (&ib, e);
4142 for (e = node->indirect_calls; e; e = e->next_callee)
4143 read_inline_edge_summary (&ib, e);
4146 lto_free_section_data (file_data, LTO_section_inline_summary, NULL, data,
4147 len);
4148 lto_data_in_delete (data_in);
4152 /* Read inline summary. Jump functions are shared among ipa-cp
4153 and inliner, so when ipa-cp is active, we don't need to write them
4154 twice. */
4156 void
4157 inline_read_summary (void)
4159 struct lto_file_decl_data **file_data_vec = lto_get_file_decl_data ();
4160 struct lto_file_decl_data *file_data;
4161 unsigned int j = 0;
4163 inline_summary_alloc ();
4165 while ((file_data = file_data_vec[j++]))
4167 size_t len;
4168 const char *data = lto_get_section_data (file_data,
4169 LTO_section_inline_summary,
4170 NULL, &len);
4171 if (data)
4172 inline_read_section (file_data, data, len);
4173 else
4174 /* Fatal error here. We do not want to support compiling ltrans units
4175 with different version of compiler or different flags than the WPA
4176 unit, so this should never happen. */
4177 fatal_error ("ipa inline summary is missing in input file");
4179 if (optimize)
4181 ipa_register_cgraph_hooks ();
4182 if (!flag_ipa_cp)
4183 ipa_prop_read_jump_functions ();
4185 function_insertion_hook_holder =
4186 cgraph_add_function_insertion_hook (&add_new_function, NULL);
4190 /* Write predicate P to OB. */
4192 static void
4193 write_predicate (struct output_block *ob, struct predicate *p)
4195 int j;
4196 if (p)
4197 for (j = 0; p->clause[j]; j++)
4199 gcc_assert (j < MAX_CLAUSES);
4200 streamer_write_uhwi (ob, p->clause[j]);
4202 streamer_write_uhwi (ob, 0);
4206 /* Write inline summary for edge E to OB. */
4208 static void
4209 write_inline_edge_summary (struct output_block *ob, struct cgraph_edge *e)
4211 struct inline_edge_summary *es = inline_edge_summary (e);
4212 int i;
4214 streamer_write_uhwi (ob, es->call_stmt_size);
4215 streamer_write_uhwi (ob, es->call_stmt_time);
4216 streamer_write_uhwi (ob, es->loop_depth);
4217 write_predicate (ob, es->predicate);
4218 streamer_write_uhwi (ob, es->param.length ());
4219 for (i = 0; i < (int) es->param.length (); i++)
4220 streamer_write_uhwi (ob, es->param[i].change_prob);
4224 /* Write inline summary for node in SET.
4225 Jump functions are shared among ipa-cp and inliner, so when ipa-cp is
4226 active, we don't need to write them twice. */
4228 void
4229 inline_write_summary (void)
4231 struct cgraph_node *node;
4232 struct output_block *ob = create_output_block (LTO_section_inline_summary);
4233 lto_symtab_encoder_t encoder = ob->decl_state->symtab_node_encoder;
4234 unsigned int count = 0;
4235 int i;
4237 for (i = 0; i < lto_symtab_encoder_size (encoder); i++)
4239 symtab_node *snode = lto_symtab_encoder_deref (encoder, i);
4240 cgraph_node *cnode = dyn_cast <cgraph_node> (snode);
4241 if (cnode && cnode->definition && !cnode->alias)
4242 count++;
4244 streamer_write_uhwi (ob, count);
4246 for (i = 0; i < lto_symtab_encoder_size (encoder); i++)
4248 symtab_node *snode = lto_symtab_encoder_deref (encoder, i);
4249 cgraph_node *cnode = dyn_cast <cgraph_node> (snode);
4250 if (cnode && (node = cnode)->definition && !node->alias)
4252 struct inline_summary *info = inline_summary (node);
4253 struct bitpack_d bp;
4254 struct cgraph_edge *edge;
4255 int i;
4256 size_time_entry *e;
4257 struct condition *c;
4259 streamer_write_uhwi (ob,
4260 lto_symtab_encoder_encode (encoder,
4262 node));
4263 streamer_write_hwi (ob, info->estimated_self_stack_size);
4264 streamer_write_hwi (ob, info->self_size);
4265 streamer_write_hwi (ob, info->self_time);
4266 bp = bitpack_create (ob->main_stream);
4267 bp_pack_value (&bp, info->inlinable, 1);
4268 streamer_write_bitpack (&bp);
4269 streamer_write_uhwi (ob, vec_safe_length (info->conds));
4270 for (i = 0; vec_safe_iterate (info->conds, i, &c); i++)
4272 streamer_write_uhwi (ob, c->operand_num);
4273 streamer_write_uhwi (ob, c->code);
4274 stream_write_tree (ob, c->val, true);
4275 bp = bitpack_create (ob->main_stream);
4276 bp_pack_value (&bp, c->agg_contents, 1);
4277 bp_pack_value (&bp, c->by_ref, 1);
4278 streamer_write_bitpack (&bp);
4279 if (c->agg_contents)
4280 streamer_write_uhwi (ob, c->offset);
4282 streamer_write_uhwi (ob, vec_safe_length (info->entry));
4283 for (i = 0; vec_safe_iterate (info->entry, i, &e); i++)
4285 streamer_write_uhwi (ob, e->size);
4286 streamer_write_uhwi (ob, e->time);
4287 write_predicate (ob, &e->predicate);
4289 write_predicate (ob, info->loop_iterations);
4290 write_predicate (ob, info->loop_stride);
4291 write_predicate (ob, info->array_index);
4292 for (edge = node->callees; edge; edge = edge->next_callee)
4293 write_inline_edge_summary (ob, edge);
4294 for (edge = node->indirect_calls; edge; edge = edge->next_callee)
4295 write_inline_edge_summary (ob, edge);
4298 streamer_write_char_stream (ob->main_stream, 0);
4299 produce_asm (ob, NULL);
4300 destroy_output_block (ob);
4302 if (optimize && !flag_ipa_cp)
4303 ipa_prop_write_jump_functions ();
4307 /* Release inline summary. */
4309 void
4310 inline_free_summary (void)
4312 struct cgraph_node *node;
4313 if (!inline_edge_summary_vec.exists ())
4314 return;
4315 FOR_EACH_DEFINED_FUNCTION (node)
4316 if (!node->alias)
4317 reset_inline_summary (node);
4318 if (function_insertion_hook_holder)
4319 cgraph_remove_function_insertion_hook (function_insertion_hook_holder);
4320 function_insertion_hook_holder = NULL;
4321 if (node_removal_hook_holder)
4322 cgraph_remove_node_removal_hook (node_removal_hook_holder);
4323 node_removal_hook_holder = NULL;
4324 if (edge_removal_hook_holder)
4325 cgraph_remove_edge_removal_hook (edge_removal_hook_holder);
4326 edge_removal_hook_holder = NULL;
4327 if (node_duplication_hook_holder)
4328 cgraph_remove_node_duplication_hook (node_duplication_hook_holder);
4329 node_duplication_hook_holder = NULL;
4330 if (edge_duplication_hook_holder)
4331 cgraph_remove_edge_duplication_hook (edge_duplication_hook_holder);
4332 edge_duplication_hook_holder = NULL;
4333 vec_free (inline_summary_vec);
4334 inline_edge_summary_vec.release ();
4335 if (edge_predicate_pool)
4336 free_alloc_pool (edge_predicate_pool);
4337 edge_predicate_pool = 0;