PR c++/38611 - missing -Wattributes on a typedef with attribute aligned
[official-gcc.git] / gcc / ipa-inline-analysis.c
blob17b21d17c5f41187712c10f04d095eefdd92618b
1 /* Inlining decision heuristics.
2 Copyright (C) 2003-2016 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 "backend.h"
71 #include "tree.h"
72 #include "gimple.h"
73 #include "alloc-pool.h"
74 #include "tree-pass.h"
75 #include "ssa.h"
76 #include "tree-streamer.h"
77 #include "cgraph.h"
78 #include "diagnostic.h"
79 #include "fold-const.h"
80 #include "print-tree.h"
81 #include "tree-inline.h"
82 #include "gimple-pretty-print.h"
83 #include "params.h"
84 #include "cfganal.h"
85 #include "gimple-iterator.h"
86 #include "tree-cfg.h"
87 #include "tree-ssa-loop-niter.h"
88 #include "tree-ssa-loop.h"
89 #include "symbol-summary.h"
90 #include "ipa-prop.h"
91 #include "ipa-inline.h"
92 #include "cfgloop.h"
93 #include "tree-scalar-evolution.h"
94 #include "ipa-utils.h"
95 #include "cilk.h"
96 #include "cfgexpand.h"
97 #include "gimplify.h"
99 /* Estimate runtime of function can easilly run into huge numbers with many
100 nested loops. Be sure we can compute time * INLINE_SIZE_SCALE * 2 in an
101 integer. For anything larger we use gcov_type. */
102 #define MAX_TIME 500000
104 /* Number of bits in integer, but we really want to be stable across different
105 hosts. */
106 #define NUM_CONDITIONS 32
108 enum predicate_conditions
110 predicate_false_condition = 0,
111 predicate_not_inlined_condition = 1,
112 predicate_first_dynamic_condition = 2
115 /* Special condition code we use to represent test that operand is compile time
116 constant. */
117 #define IS_NOT_CONSTANT ERROR_MARK
118 /* Special condition code we use to represent test that operand is not changed
119 across invocation of the function. When operand IS_NOT_CONSTANT it is always
120 CHANGED, however i.e. loop invariants can be NOT_CHANGED given percentage
121 of executions even when they are not compile time constants. */
122 #define CHANGED IDENTIFIER_NODE
124 /* Holders of ipa cgraph hooks: */
125 static struct cgraph_2edge_hook_list *edge_duplication_hook_holder;
126 static struct cgraph_edge_hook_list *edge_removal_hook_holder;
127 static void inline_edge_removal_hook (struct cgraph_edge *, void *);
128 static void inline_edge_duplication_hook (struct cgraph_edge *,
129 struct cgraph_edge *, void *);
131 /* VECtor holding inline summaries.
132 In GGC memory because conditions might point to constant trees. */
133 function_summary <inline_summary *> *inline_summaries;
134 vec<inline_edge_summary_t> inline_edge_summary_vec;
136 /* Cached node/edge growths. */
137 vec<edge_growth_cache_entry> edge_growth_cache;
139 /* Edge predicates goes here. */
140 static object_allocator<predicate> edge_predicate_pool ("edge predicates");
142 /* Return true predicate (tautology).
143 We represent it by empty list of clauses. */
145 static inline struct predicate
146 true_predicate (void)
148 struct predicate p;
149 p.clause[0] = 0;
150 return p;
154 /* Return predicate testing single condition number COND. */
156 static inline struct predicate
157 single_cond_predicate (int cond)
159 struct predicate p;
160 p.clause[0] = 1 << cond;
161 p.clause[1] = 0;
162 return p;
166 /* Return false predicate. First clause require false condition. */
168 static inline struct predicate
169 false_predicate (void)
171 return single_cond_predicate (predicate_false_condition);
175 /* Return true if P is (true). */
177 static inline bool
178 true_predicate_p (struct predicate *p)
180 return !p->clause[0];
184 /* Return true if P is (false). */
186 static inline bool
187 false_predicate_p (struct predicate *p)
189 if (p->clause[0] == (1 << predicate_false_condition))
191 gcc_checking_assert (!p->clause[1]
192 && p->clause[0] == 1 << predicate_false_condition);
193 return true;
195 return false;
199 /* Return predicate that is set true when function is not inlined. */
201 static inline struct predicate
202 not_inlined_predicate (void)
204 return single_cond_predicate (predicate_not_inlined_condition);
207 /* Simple description of whether a memory load or a condition refers to a load
208 from an aggregate and if so, how and where from in the aggregate.
209 Individual fields have the same meaning like fields with the same name in
210 struct condition. */
212 struct agg_position_info
214 HOST_WIDE_INT offset;
215 bool agg_contents;
216 bool by_ref;
219 /* Add condition to condition list CONDS. AGGPOS describes whether the used
220 oprand is loaded from an aggregate and where in the aggregate it is. It can
221 be NULL, which means this not a load from an aggregate. */
223 static struct predicate
224 add_condition (struct inline_summary *summary, int operand_num,
225 struct agg_position_info *aggpos,
226 enum tree_code code, tree val)
228 int i;
229 struct condition *c;
230 struct condition new_cond;
231 HOST_WIDE_INT offset;
232 bool agg_contents, by_ref;
234 if (aggpos)
236 offset = aggpos->offset;
237 agg_contents = aggpos->agg_contents;
238 by_ref = aggpos->by_ref;
240 else
242 offset = 0;
243 agg_contents = false;
244 by_ref = false;
247 gcc_checking_assert (operand_num >= 0);
248 for (i = 0; vec_safe_iterate (summary->conds, i, &c); i++)
250 if (c->operand_num == operand_num
251 && c->code == code
252 && c->val == val
253 && c->agg_contents == agg_contents
254 && (!agg_contents || (c->offset == offset && c->by_ref == by_ref)))
255 return single_cond_predicate (i + predicate_first_dynamic_condition);
257 /* Too many conditions. Give up and return constant true. */
258 if (i == NUM_CONDITIONS - predicate_first_dynamic_condition)
259 return true_predicate ();
261 new_cond.operand_num = operand_num;
262 new_cond.code = code;
263 new_cond.val = val;
264 new_cond.agg_contents = agg_contents;
265 new_cond.by_ref = by_ref;
266 new_cond.offset = offset;
267 vec_safe_push (summary->conds, new_cond);
268 return single_cond_predicate (i + predicate_first_dynamic_condition);
272 /* Add clause CLAUSE into the predicate P. */
274 static inline void
275 add_clause (conditions conditions, struct predicate *p, clause_t clause)
277 int i;
278 int i2;
279 int insert_here = -1;
280 int c1, c2;
282 /* True clause. */
283 if (!clause)
284 return;
286 /* False clause makes the whole predicate false. Kill the other variants. */
287 if (clause == (1 << predicate_false_condition))
289 p->clause[0] = (1 << predicate_false_condition);
290 p->clause[1] = 0;
291 return;
293 if (false_predicate_p (p))
294 return;
296 /* No one should be silly enough to add false into nontrivial clauses. */
297 gcc_checking_assert (!(clause & (1 << predicate_false_condition)));
299 /* Look where to insert the clause. At the same time prune out
300 clauses of P that are implied by the new clause and thus
301 redundant. */
302 for (i = 0, i2 = 0; i <= MAX_CLAUSES; i++)
304 p->clause[i2] = p->clause[i];
306 if (!p->clause[i])
307 break;
309 /* If p->clause[i] implies clause, there is nothing to add. */
310 if ((p->clause[i] & clause) == p->clause[i])
312 /* We had nothing to add, none of clauses should've become
313 redundant. */
314 gcc_checking_assert (i == i2);
315 return;
318 if (p->clause[i] < clause && insert_here < 0)
319 insert_here = i2;
321 /* If clause implies p->clause[i], then p->clause[i] becomes redundant.
322 Otherwise the p->clause[i] has to stay. */
323 if ((p->clause[i] & clause) != clause)
324 i2++;
327 /* Look for clauses that are obviously true. I.e.
328 op0 == 5 || op0 != 5. */
329 for (c1 = predicate_first_dynamic_condition; c1 < NUM_CONDITIONS; c1++)
331 condition *cc1;
332 if (!(clause & (1 << c1)))
333 continue;
334 cc1 = &(*conditions)[c1 - predicate_first_dynamic_condition];
335 /* We have no way to represent !CHANGED and !IS_NOT_CONSTANT
336 and thus there is no point for looking for them. */
337 if (cc1->code == CHANGED || cc1->code == IS_NOT_CONSTANT)
338 continue;
339 for (c2 = c1 + 1; c2 < NUM_CONDITIONS; c2++)
340 if (clause & (1 << c2))
342 condition *cc1 =
343 &(*conditions)[c1 - predicate_first_dynamic_condition];
344 condition *cc2 =
345 &(*conditions)[c2 - predicate_first_dynamic_condition];
346 if (cc1->operand_num == cc2->operand_num
347 && cc1->val == cc2->val
348 && cc2->code != IS_NOT_CONSTANT
349 && cc2->code != CHANGED
350 && cc1->code == invert_tree_comparison (cc2->code,
351 HONOR_NANS (cc1->val)))
352 return;
357 /* We run out of variants. Be conservative in positive direction. */
358 if (i2 == MAX_CLAUSES)
359 return;
360 /* Keep clauses in decreasing order. This makes equivalence testing easy. */
361 p->clause[i2 + 1] = 0;
362 if (insert_here >= 0)
363 for (; i2 > insert_here; i2--)
364 p->clause[i2] = p->clause[i2 - 1];
365 else
366 insert_here = i2;
367 p->clause[insert_here] = clause;
371 /* Return P & P2. */
373 static struct predicate
374 and_predicates (conditions conditions,
375 struct predicate *p, struct predicate *p2)
377 struct predicate out = *p;
378 int i;
380 /* Avoid busy work. */
381 if (false_predicate_p (p2) || true_predicate_p (p))
382 return *p2;
383 if (false_predicate_p (p) || true_predicate_p (p2))
384 return *p;
386 /* See how far predicates match. */
387 for (i = 0; p->clause[i] && p->clause[i] == p2->clause[i]; i++)
389 gcc_checking_assert (i < MAX_CLAUSES);
392 /* Combine the predicates rest. */
393 for (; p2->clause[i]; i++)
395 gcc_checking_assert (i < MAX_CLAUSES);
396 add_clause (conditions, &out, p2->clause[i]);
398 return out;
402 /* Return true if predicates are obviously equal. */
404 static inline bool
405 predicates_equal_p (struct predicate *p, struct predicate *p2)
407 int i;
408 for (i = 0; p->clause[i]; i++)
410 gcc_checking_assert (i < MAX_CLAUSES);
411 gcc_checking_assert (p->clause[i] > p->clause[i + 1]);
412 gcc_checking_assert (!p2->clause[i]
413 || p2->clause[i] > p2->clause[i + 1]);
414 if (p->clause[i] != p2->clause[i])
415 return false;
417 return !p2->clause[i];
421 /* Return P | P2. */
423 static struct predicate
424 or_predicates (conditions conditions,
425 struct predicate *p, struct predicate *p2)
427 struct predicate out = true_predicate ();
428 int i, j;
430 /* Avoid busy work. */
431 if (false_predicate_p (p2) || true_predicate_p (p))
432 return *p;
433 if (false_predicate_p (p) || true_predicate_p (p2))
434 return *p2;
435 if (predicates_equal_p (p, p2))
436 return *p;
438 /* OK, combine the predicates. */
439 for (i = 0; p->clause[i]; i++)
440 for (j = 0; p2->clause[j]; j++)
442 gcc_checking_assert (i < MAX_CLAUSES && j < MAX_CLAUSES);
443 add_clause (conditions, &out, p->clause[i] | p2->clause[j]);
445 return out;
449 /* Having partial truth assignment in POSSIBLE_TRUTHS, return false
450 if predicate P is known to be false. */
452 static bool
453 evaluate_predicate (struct predicate *p, clause_t possible_truths)
455 int i;
457 /* True remains true. */
458 if (true_predicate_p (p))
459 return true;
461 gcc_assert (!(possible_truths & (1 << predicate_false_condition)));
463 /* See if we can find clause we can disprove. */
464 for (i = 0; p->clause[i]; i++)
466 gcc_checking_assert (i < MAX_CLAUSES);
467 if (!(p->clause[i] & possible_truths))
468 return false;
470 return true;
473 /* Return the probability in range 0...REG_BR_PROB_BASE that the predicated
474 instruction will be recomputed per invocation of the inlined call. */
476 static int
477 predicate_probability (conditions conds,
478 struct predicate *p, clause_t possible_truths,
479 vec<inline_param_summary> inline_param_summary)
481 int i;
482 int combined_prob = REG_BR_PROB_BASE;
484 /* True remains true. */
485 if (true_predicate_p (p))
486 return REG_BR_PROB_BASE;
488 if (false_predicate_p (p))
489 return 0;
491 gcc_assert (!(possible_truths & (1 << predicate_false_condition)));
493 /* See if we can find clause we can disprove. */
494 for (i = 0; p->clause[i]; i++)
496 gcc_checking_assert (i < MAX_CLAUSES);
497 if (!(p->clause[i] & possible_truths))
498 return 0;
499 else
501 int this_prob = 0;
502 int i2;
503 if (!inline_param_summary.exists ())
504 return REG_BR_PROB_BASE;
505 for (i2 = 0; i2 < NUM_CONDITIONS; i2++)
506 if ((p->clause[i] & possible_truths) & (1 << i2))
508 if (i2 >= predicate_first_dynamic_condition)
510 condition *c =
511 &(*conds)[i2 - predicate_first_dynamic_condition];
512 if (c->code == CHANGED
513 && (c->operand_num <
514 (int) inline_param_summary.length ()))
516 int iprob =
517 inline_param_summary[c->operand_num].change_prob;
518 this_prob = MAX (this_prob, iprob);
520 else
521 this_prob = REG_BR_PROB_BASE;
523 else
524 this_prob = REG_BR_PROB_BASE;
526 combined_prob = MIN (this_prob, combined_prob);
527 if (!combined_prob)
528 return 0;
531 return combined_prob;
535 /* Dump conditional COND. */
537 static void
538 dump_condition (FILE *f, conditions conditions, int cond)
540 condition *c;
541 if (cond == predicate_false_condition)
542 fprintf (f, "false");
543 else if (cond == predicate_not_inlined_condition)
544 fprintf (f, "not inlined");
545 else
547 c = &(*conditions)[cond - predicate_first_dynamic_condition];
548 fprintf (f, "op%i", c->operand_num);
549 if (c->agg_contents)
550 fprintf (f, "[%soffset: " HOST_WIDE_INT_PRINT_DEC "]",
551 c->by_ref ? "ref " : "", c->offset);
552 if (c->code == IS_NOT_CONSTANT)
554 fprintf (f, " not constant");
555 return;
557 if (c->code == CHANGED)
559 fprintf (f, " changed");
560 return;
562 fprintf (f, " %s ", op_symbol_code (c->code));
563 print_generic_expr (f, c->val, 1);
568 /* Dump clause CLAUSE. */
570 static void
571 dump_clause (FILE *f, conditions conds, clause_t clause)
573 int i;
574 bool found = false;
575 fprintf (f, "(");
576 if (!clause)
577 fprintf (f, "true");
578 for (i = 0; i < NUM_CONDITIONS; i++)
579 if (clause & (1 << i))
581 if (found)
582 fprintf (f, " || ");
583 found = true;
584 dump_condition (f, conds, i);
586 fprintf (f, ")");
590 /* Dump predicate PREDICATE. */
592 static void
593 dump_predicate (FILE *f, conditions conds, struct predicate *pred)
595 int i;
596 if (true_predicate_p (pred))
597 dump_clause (f, conds, 0);
598 else
599 for (i = 0; pred->clause[i]; i++)
601 if (i)
602 fprintf (f, " && ");
603 dump_clause (f, conds, pred->clause[i]);
605 fprintf (f, "\n");
609 /* Dump inline hints. */
610 void
611 dump_inline_hints (FILE *f, inline_hints hints)
613 if (!hints)
614 return;
615 fprintf (f, "inline hints:");
616 if (hints & INLINE_HINT_indirect_call)
618 hints &= ~INLINE_HINT_indirect_call;
619 fprintf (f, " indirect_call");
621 if (hints & INLINE_HINT_loop_iterations)
623 hints &= ~INLINE_HINT_loop_iterations;
624 fprintf (f, " loop_iterations");
626 if (hints & INLINE_HINT_loop_stride)
628 hints &= ~INLINE_HINT_loop_stride;
629 fprintf (f, " loop_stride");
631 if (hints & INLINE_HINT_same_scc)
633 hints &= ~INLINE_HINT_same_scc;
634 fprintf (f, " same_scc");
636 if (hints & INLINE_HINT_in_scc)
638 hints &= ~INLINE_HINT_in_scc;
639 fprintf (f, " in_scc");
641 if (hints & INLINE_HINT_cross_module)
643 hints &= ~INLINE_HINT_cross_module;
644 fprintf (f, " cross_module");
646 if (hints & INLINE_HINT_declared_inline)
648 hints &= ~INLINE_HINT_declared_inline;
649 fprintf (f, " declared_inline");
651 if (hints & INLINE_HINT_array_index)
653 hints &= ~INLINE_HINT_array_index;
654 fprintf (f, " array_index");
656 if (hints & INLINE_HINT_known_hot)
658 hints &= ~INLINE_HINT_known_hot;
659 fprintf (f, " known_hot");
661 gcc_assert (!hints);
665 /* Record SIZE and TIME under condition PRED into the inline summary. */
667 static void
668 account_size_time (struct inline_summary *summary, int size, int time,
669 struct predicate *pred)
671 size_time_entry *e;
672 bool found = false;
673 int i;
675 if (false_predicate_p (pred))
676 return;
678 /* We need to create initial empty unconitional clause, but otherwie
679 we don't need to account empty times and sizes. */
680 if (!size && !time && summary->entry)
681 return;
683 /* Watch overflow that might result from insane profiles. */
684 if (time > MAX_TIME * INLINE_TIME_SCALE)
685 time = MAX_TIME * INLINE_TIME_SCALE;
686 gcc_assert (time >= 0);
688 for (i = 0; vec_safe_iterate (summary->entry, i, &e); i++)
689 if (predicates_equal_p (&e->predicate, pred))
691 found = true;
692 break;
694 if (i == 256)
696 i = 0;
697 found = true;
698 e = &(*summary->entry)[0];
699 gcc_assert (!e->predicate.clause[0]);
700 if (dump_file && (dump_flags & TDF_DETAILS))
701 fprintf (dump_file,
702 "\t\tReached limit on number of entries, "
703 "ignoring the predicate.");
705 if (dump_file && (dump_flags & TDF_DETAILS) && (time || size))
707 fprintf (dump_file,
708 "\t\tAccounting size:%3.2f, time:%3.2f on %spredicate:",
709 ((double) size) / INLINE_SIZE_SCALE,
710 ((double) time) / INLINE_TIME_SCALE, found ? "" : "new ");
711 dump_predicate (dump_file, summary->conds, pred);
713 if (!found)
715 struct size_time_entry new_entry;
716 new_entry.size = size;
717 new_entry.time = time;
718 new_entry.predicate = *pred;
719 vec_safe_push (summary->entry, new_entry);
721 else
723 e->size += size;
724 e->time += time;
725 if (e->time > MAX_TIME * INLINE_TIME_SCALE)
726 e->time = MAX_TIME * INLINE_TIME_SCALE;
730 /* We proved E to be unreachable, redirect it to __bultin_unreachable. */
732 static struct cgraph_edge *
733 redirect_to_unreachable (struct cgraph_edge *e)
735 struct cgraph_node *callee = !e->inline_failed ? e->callee : NULL;
736 struct cgraph_node *target = cgraph_node::get_create
737 (builtin_decl_implicit (BUILT_IN_UNREACHABLE));
739 if (e->speculative)
740 e = e->resolve_speculation (target->decl);
741 else if (!e->callee)
742 e->make_direct (target);
743 else
744 e->redirect_callee (target);
745 struct inline_edge_summary *es = inline_edge_summary (e);
746 e->inline_failed = CIF_UNREACHABLE;
747 e->frequency = 0;
748 e->count = 0;
749 es->call_stmt_size = 0;
750 es->call_stmt_time = 0;
751 if (callee)
752 callee->remove_symbol_and_inline_clones ();
753 return e;
756 /* Set predicate for edge E. */
758 static void
759 edge_set_predicate (struct cgraph_edge *e, struct predicate *predicate)
761 /* If the edge is determined to be never executed, redirect it
762 to BUILTIN_UNREACHABLE to save inliner from inlining into it. */
763 if (predicate && false_predicate_p (predicate)
764 /* When handling speculative edges, we need to do the redirection
765 just once. Do it always on the direct edge, so we do not
766 attempt to resolve speculation while duplicating the edge. */
767 && (!e->speculative || e->callee))
768 e = redirect_to_unreachable (e);
770 struct inline_edge_summary *es = inline_edge_summary (e);
771 if (predicate && !true_predicate_p (predicate))
773 if (!es->predicate)
774 es->predicate = edge_predicate_pool.allocate ();
775 *es->predicate = *predicate;
777 else
779 if (es->predicate)
780 edge_predicate_pool.remove (es->predicate);
781 es->predicate = NULL;
785 /* Set predicate for hint *P. */
787 static void
788 set_hint_predicate (struct predicate **p, struct predicate new_predicate)
790 if (false_predicate_p (&new_predicate) || true_predicate_p (&new_predicate))
792 if (*p)
793 edge_predicate_pool.remove (*p);
794 *p = NULL;
796 else
798 if (!*p)
799 *p = edge_predicate_pool.allocate ();
800 **p = new_predicate;
805 /* KNOWN_VALS is partial mapping of parameters of NODE to constant values.
806 KNOWN_AGGS is a vector of aggreggate jump functions for each parameter.
807 Return clause of possible truths. When INLINE_P is true, assume that we are
808 inlining.
810 ERROR_MARK means compile time invariant. */
812 static clause_t
813 evaluate_conditions_for_known_args (struct cgraph_node *node,
814 bool inline_p,
815 vec<tree> known_vals,
816 vec<ipa_agg_jump_function_p>
817 known_aggs)
819 clause_t clause = inline_p ? 0 : 1 << predicate_not_inlined_condition;
820 struct inline_summary *info = inline_summaries->get (node);
821 int i;
822 struct condition *c;
824 for (i = 0; vec_safe_iterate (info->conds, i, &c); i++)
826 tree val;
827 tree res;
829 /* We allow call stmt to have fewer arguments than the callee function
830 (especially for K&R style programs). So bound check here (we assume
831 known_aggs vector, if non-NULL, has the same length as
832 known_vals). */
833 gcc_checking_assert (!known_aggs.exists ()
834 || (known_vals.length () == known_aggs.length ()));
835 if (c->operand_num >= (int) known_vals.length ())
837 clause |= 1 << (i + predicate_first_dynamic_condition);
838 continue;
841 if (c->agg_contents)
843 struct ipa_agg_jump_function *agg;
845 if (c->code == CHANGED
846 && !c->by_ref
847 && (known_vals[c->operand_num] == error_mark_node))
848 continue;
850 if (known_aggs.exists ())
852 agg = known_aggs[c->operand_num];
853 val = ipa_find_agg_cst_for_param (agg, c->offset, c->by_ref);
855 else
856 val = NULL_TREE;
858 else
860 val = known_vals[c->operand_num];
861 if (val == error_mark_node && c->code != CHANGED)
862 val = NULL_TREE;
865 if (!val)
867 clause |= 1 << (i + predicate_first_dynamic_condition);
868 continue;
870 if (c->code == IS_NOT_CONSTANT || c->code == CHANGED)
871 continue;
873 if (operand_equal_p (TYPE_SIZE (TREE_TYPE (c->val)),
874 TYPE_SIZE (TREE_TYPE (val)), 0))
876 val = fold_unary (VIEW_CONVERT_EXPR, TREE_TYPE (c->val), val);
878 res = val
879 ? fold_binary_to_constant (c->code, boolean_type_node, val, c->val)
880 : NULL;
882 if (res && integer_zerop (res))
883 continue;
885 clause |= 1 << (i + predicate_first_dynamic_condition);
887 return clause;
891 /* Work out what conditions might be true at invocation of E. */
893 static void
894 evaluate_properties_for_edge (struct cgraph_edge *e, bool inline_p,
895 clause_t *clause_ptr,
896 vec<tree> *known_vals_ptr,
897 vec<ipa_polymorphic_call_context>
898 *known_contexts_ptr,
899 vec<ipa_agg_jump_function_p> *known_aggs_ptr)
901 struct cgraph_node *callee = e->callee->ultimate_alias_target ();
902 struct inline_summary *info = inline_summaries->get (callee);
903 vec<tree> known_vals = vNULL;
904 vec<ipa_agg_jump_function_p> known_aggs = vNULL;
906 if (clause_ptr)
907 *clause_ptr = inline_p ? 0 : 1 << predicate_not_inlined_condition;
908 if (known_vals_ptr)
909 known_vals_ptr->create (0);
910 if (known_contexts_ptr)
911 known_contexts_ptr->create (0);
913 if (ipa_node_params_sum
914 && !e->call_stmt_cannot_inline_p
915 && ((clause_ptr && info->conds) || known_vals_ptr || known_contexts_ptr))
917 struct ipa_node_params *parms_info;
918 struct ipa_edge_args *args = IPA_EDGE_REF (e);
919 struct inline_edge_summary *es = inline_edge_summary (e);
920 int i, count = ipa_get_cs_argument_count (args);
922 if (e->caller->global.inlined_to)
923 parms_info = IPA_NODE_REF (e->caller->global.inlined_to);
924 else
925 parms_info = IPA_NODE_REF (e->caller);
927 if (count && (info->conds || known_vals_ptr))
928 known_vals.safe_grow_cleared (count);
929 if (count && (info->conds || known_aggs_ptr))
930 known_aggs.safe_grow_cleared (count);
931 if (count && known_contexts_ptr)
932 known_contexts_ptr->safe_grow_cleared (count);
934 for (i = 0; i < count; i++)
936 struct ipa_jump_func *jf = ipa_get_ith_jump_func (args, i);
937 tree cst = ipa_value_from_jfunc (parms_info, jf);
939 if (!cst && e->call_stmt
940 && i < (int)gimple_call_num_args (e->call_stmt))
942 cst = gimple_call_arg (e->call_stmt, i);
943 if (!is_gimple_min_invariant (cst))
944 cst = NULL;
946 if (cst)
948 gcc_checking_assert (TREE_CODE (cst) != TREE_BINFO);
949 if (known_vals.exists ())
950 known_vals[i] = cst;
952 else if (inline_p && !es->param[i].change_prob)
953 known_vals[i] = error_mark_node;
955 if (known_contexts_ptr)
956 (*known_contexts_ptr)[i] = ipa_context_from_jfunc (parms_info, e,
957 i, jf);
958 /* TODO: When IPA-CP starts propagating and merging aggregate jump
959 functions, use its knowledge of the caller too, just like the
960 scalar case above. */
961 known_aggs[i] = &jf->agg;
964 else if (e->call_stmt && !e->call_stmt_cannot_inline_p
965 && ((clause_ptr && info->conds) || known_vals_ptr))
967 int i, count = (int)gimple_call_num_args (e->call_stmt);
969 if (count && (info->conds || known_vals_ptr))
970 known_vals.safe_grow_cleared (count);
971 for (i = 0; i < count; i++)
973 tree cst = gimple_call_arg (e->call_stmt, i);
974 if (!is_gimple_min_invariant (cst))
975 cst = NULL;
976 if (cst)
977 known_vals[i] = cst;
981 if (clause_ptr)
982 *clause_ptr = evaluate_conditions_for_known_args (callee, inline_p,
983 known_vals, known_aggs);
985 if (known_vals_ptr)
986 *known_vals_ptr = known_vals;
987 else
988 known_vals.release ();
990 if (known_aggs_ptr)
991 *known_aggs_ptr = known_aggs;
992 else
993 known_aggs.release ();
997 /* Allocate the inline summary vector or resize it to cover all cgraph nodes. */
999 static void
1000 inline_summary_alloc (void)
1002 if (!edge_removal_hook_holder)
1003 edge_removal_hook_holder =
1004 symtab->add_edge_removal_hook (&inline_edge_removal_hook, NULL);
1005 if (!edge_duplication_hook_holder)
1006 edge_duplication_hook_holder =
1007 symtab->add_edge_duplication_hook (&inline_edge_duplication_hook, NULL);
1009 if (!inline_summaries)
1010 inline_summaries = (inline_summary_t*) inline_summary_t::create_ggc (symtab);
1012 if (inline_edge_summary_vec.length () <= (unsigned) symtab->edges_max_uid)
1013 inline_edge_summary_vec.safe_grow_cleared (symtab->edges_max_uid + 1);
1016 /* We are called multiple time for given function; clear
1017 data from previous run so they are not cumulated. */
1019 static void
1020 reset_inline_edge_summary (struct cgraph_edge *e)
1022 if (e->uid < (int) inline_edge_summary_vec.length ())
1024 struct inline_edge_summary *es = inline_edge_summary (e);
1026 es->call_stmt_size = es->call_stmt_time = 0;
1027 if (es->predicate)
1028 edge_predicate_pool.remove (es->predicate);
1029 es->predicate = NULL;
1030 es->param.release ();
1034 /* We are called multiple time for given function; clear
1035 data from previous run so they are not cumulated. */
1037 static void
1038 reset_inline_summary (struct cgraph_node *node,
1039 inline_summary *info)
1041 struct cgraph_edge *e;
1043 info->self_size = info->self_time = 0;
1044 info->estimated_stack_size = 0;
1045 info->estimated_self_stack_size = 0;
1046 info->stack_frame_offset = 0;
1047 info->size = 0;
1048 info->time = 0;
1049 info->growth = 0;
1050 info->scc_no = 0;
1051 if (info->loop_iterations)
1053 edge_predicate_pool.remove (info->loop_iterations);
1054 info->loop_iterations = NULL;
1056 if (info->loop_stride)
1058 edge_predicate_pool.remove (info->loop_stride);
1059 info->loop_stride = NULL;
1061 if (info->array_index)
1063 edge_predicate_pool.remove (info->array_index);
1064 info->array_index = NULL;
1066 vec_free (info->conds);
1067 vec_free (info->entry);
1068 for (e = node->callees; e; e = e->next_callee)
1069 reset_inline_edge_summary (e);
1070 for (e = node->indirect_calls; e; e = e->next_callee)
1071 reset_inline_edge_summary (e);
1072 info->fp_expressions = false;
1075 /* Hook that is called by cgraph.c when a node is removed. */
1077 void
1078 inline_summary_t::remove (cgraph_node *node, inline_summary *info)
1080 reset_inline_summary (node, info);
1083 /* Remap predicate P of former function to be predicate of duplicated function.
1084 POSSIBLE_TRUTHS is clause of possible truths in the duplicated node,
1085 INFO is inline summary of the duplicated node. */
1087 static struct predicate
1088 remap_predicate_after_duplication (struct predicate *p,
1089 clause_t possible_truths,
1090 struct inline_summary *info)
1092 struct predicate new_predicate = true_predicate ();
1093 int j;
1094 for (j = 0; p->clause[j]; j++)
1095 if (!(possible_truths & p->clause[j]))
1097 new_predicate = false_predicate ();
1098 break;
1100 else
1101 add_clause (info->conds, &new_predicate,
1102 possible_truths & p->clause[j]);
1103 return new_predicate;
1106 /* Same as remap_predicate_after_duplication but handle hint predicate *P.
1107 Additionally care about allocating new memory slot for updated predicate
1108 and set it to NULL when it becomes true or false (and thus uninteresting).
1111 static void
1112 remap_hint_predicate_after_duplication (struct predicate **p,
1113 clause_t possible_truths,
1114 struct inline_summary *info)
1116 struct predicate new_predicate;
1118 if (!*p)
1119 return;
1121 new_predicate = remap_predicate_after_duplication (*p,
1122 possible_truths, info);
1123 /* We do not want to free previous predicate; it is used by node origin. */
1124 *p = NULL;
1125 set_hint_predicate (p, new_predicate);
1129 /* Hook that is called by cgraph.c when a node is duplicated. */
1130 void
1131 inline_summary_t::duplicate (cgraph_node *src,
1132 cgraph_node *dst,
1133 inline_summary *,
1134 inline_summary *info)
1136 inline_summary_alloc ();
1137 memcpy (info, inline_summaries->get (src), sizeof (inline_summary));
1138 /* TODO: as an optimization, we may avoid copying conditions
1139 that are known to be false or true. */
1140 info->conds = vec_safe_copy (info->conds);
1142 /* When there are any replacements in the function body, see if we can figure
1143 out that something was optimized out. */
1144 if (ipa_node_params_sum && dst->clone.tree_map)
1146 vec<size_time_entry, va_gc> *entry = info->entry;
1147 /* Use SRC parm info since it may not be copied yet. */
1148 struct ipa_node_params *parms_info = IPA_NODE_REF (src);
1149 vec<tree> known_vals = vNULL;
1150 int count = ipa_get_param_count (parms_info);
1151 int i, j;
1152 clause_t possible_truths;
1153 struct predicate true_pred = true_predicate ();
1154 size_time_entry *e;
1155 int optimized_out_size = 0;
1156 bool inlined_to_p = false;
1157 struct cgraph_edge *edge, *next;
1159 info->entry = 0;
1160 known_vals.safe_grow_cleared (count);
1161 for (i = 0; i < count; i++)
1163 struct ipa_replace_map *r;
1165 for (j = 0; vec_safe_iterate (dst->clone.tree_map, j, &r); j++)
1167 if (((!r->old_tree && r->parm_num == i)
1168 || (r->old_tree && r->old_tree == ipa_get_param (parms_info, i)))
1169 && r->replace_p && !r->ref_p)
1171 known_vals[i] = r->new_tree;
1172 break;
1176 possible_truths = evaluate_conditions_for_known_args (dst, false,
1177 known_vals,
1178 vNULL);
1179 known_vals.release ();
1181 account_size_time (info, 0, 0, &true_pred);
1183 /* Remap size_time vectors.
1184 Simplify the predicate by prunning out alternatives that are known
1185 to be false.
1186 TODO: as on optimization, we can also eliminate conditions known
1187 to be true. */
1188 for (i = 0; vec_safe_iterate (entry, i, &e); i++)
1190 struct predicate new_predicate;
1191 new_predicate = remap_predicate_after_duplication (&e->predicate,
1192 possible_truths,
1193 info);
1194 if (false_predicate_p (&new_predicate))
1195 optimized_out_size += e->size;
1196 else
1197 account_size_time (info, e->size, e->time, &new_predicate);
1200 /* Remap edge predicates with the same simplification as above.
1201 Also copy constantness arrays. */
1202 for (edge = dst->callees; edge; edge = next)
1204 struct predicate new_predicate;
1205 struct inline_edge_summary *es = inline_edge_summary (edge);
1206 next = edge->next_callee;
1208 if (!edge->inline_failed)
1209 inlined_to_p = true;
1210 if (!es->predicate)
1211 continue;
1212 new_predicate = remap_predicate_after_duplication (es->predicate,
1213 possible_truths,
1214 info);
1215 if (false_predicate_p (&new_predicate)
1216 && !false_predicate_p (es->predicate))
1217 optimized_out_size += es->call_stmt_size * INLINE_SIZE_SCALE;
1218 edge_set_predicate (edge, &new_predicate);
1221 /* Remap indirect edge predicates with the same simplificaiton as above.
1222 Also copy constantness arrays. */
1223 for (edge = dst->indirect_calls; edge; edge = next)
1225 struct predicate new_predicate;
1226 struct inline_edge_summary *es = inline_edge_summary (edge);
1227 next = edge->next_callee;
1229 gcc_checking_assert (edge->inline_failed);
1230 if (!es->predicate)
1231 continue;
1232 new_predicate = remap_predicate_after_duplication (es->predicate,
1233 possible_truths,
1234 info);
1235 if (false_predicate_p (&new_predicate)
1236 && !false_predicate_p (es->predicate))
1237 optimized_out_size += es->call_stmt_size * INLINE_SIZE_SCALE;
1238 edge_set_predicate (edge, &new_predicate);
1240 remap_hint_predicate_after_duplication (&info->loop_iterations,
1241 possible_truths, info);
1242 remap_hint_predicate_after_duplication (&info->loop_stride,
1243 possible_truths, info);
1244 remap_hint_predicate_after_duplication (&info->array_index,
1245 possible_truths, info);
1247 /* If inliner or someone after inliner will ever start producing
1248 non-trivial clones, we will get trouble with lack of information
1249 about updating self sizes, because size vectors already contains
1250 sizes of the calees. */
1251 gcc_assert (!inlined_to_p || !optimized_out_size);
1253 else
1255 info->entry = vec_safe_copy (info->entry);
1256 if (info->loop_iterations)
1258 predicate p = *info->loop_iterations;
1259 info->loop_iterations = NULL;
1260 set_hint_predicate (&info->loop_iterations, p);
1262 if (info->loop_stride)
1264 predicate p = *info->loop_stride;
1265 info->loop_stride = NULL;
1266 set_hint_predicate (&info->loop_stride, p);
1268 if (info->array_index)
1270 predicate p = *info->array_index;
1271 info->array_index = NULL;
1272 set_hint_predicate (&info->array_index, p);
1275 if (!dst->global.inlined_to)
1276 inline_update_overall_summary (dst);
1280 /* Hook that is called by cgraph.c when a node is duplicated. */
1282 static void
1283 inline_edge_duplication_hook (struct cgraph_edge *src,
1284 struct cgraph_edge *dst,
1285 ATTRIBUTE_UNUSED void *data)
1287 struct inline_edge_summary *info;
1288 struct inline_edge_summary *srcinfo;
1289 inline_summary_alloc ();
1290 info = inline_edge_summary (dst);
1291 srcinfo = inline_edge_summary (src);
1292 memcpy (info, srcinfo, sizeof (struct inline_edge_summary));
1293 info->predicate = NULL;
1294 edge_set_predicate (dst, srcinfo->predicate);
1295 info->param = srcinfo->param.copy ();
1296 if (!dst->indirect_unknown_callee && src->indirect_unknown_callee)
1298 info->call_stmt_size -= (eni_size_weights.indirect_call_cost
1299 - eni_size_weights.call_cost);
1300 info->call_stmt_time -= (eni_time_weights.indirect_call_cost
1301 - eni_time_weights.call_cost);
1306 /* Keep edge cache consistent across edge removal. */
1308 static void
1309 inline_edge_removal_hook (struct cgraph_edge *edge,
1310 void *data ATTRIBUTE_UNUSED)
1312 if (edge_growth_cache.exists ())
1313 reset_edge_growth_cache (edge);
1314 reset_inline_edge_summary (edge);
1318 /* Initialize growth caches. */
1320 void
1321 initialize_growth_caches (void)
1323 if (symtab->edges_max_uid)
1324 edge_growth_cache.safe_grow_cleared (symtab->edges_max_uid);
1328 /* Free growth caches. */
1330 void
1331 free_growth_caches (void)
1333 edge_growth_cache.release ();
1337 /* Dump edge summaries associated to NODE and recursively to all clones.
1338 Indent by INDENT. */
1340 static void
1341 dump_inline_edge_summary (FILE *f, int indent, struct cgraph_node *node,
1342 struct inline_summary *info)
1344 struct cgraph_edge *edge;
1345 for (edge = node->callees; edge; edge = edge->next_callee)
1347 struct inline_edge_summary *es = inline_edge_summary (edge);
1348 struct cgraph_node *callee = edge->callee->ultimate_alias_target ();
1349 int i;
1351 fprintf (f,
1352 "%*s%s/%i %s\n%*s loop depth:%2i freq:%4i size:%2i"
1353 " time: %2i callee size:%2i stack:%2i",
1354 indent, "", callee->name (), callee->order,
1355 !edge->inline_failed
1356 ? "inlined" : cgraph_inline_failed_string (edge-> inline_failed),
1357 indent, "", es->loop_depth, edge->frequency,
1358 es->call_stmt_size, es->call_stmt_time,
1359 (int) inline_summaries->get (callee)->size / INLINE_SIZE_SCALE,
1360 (int) inline_summaries->get (callee)->estimated_stack_size);
1362 if (es->predicate)
1364 fprintf (f, " predicate: ");
1365 dump_predicate (f, info->conds, es->predicate);
1367 else
1368 fprintf (f, "\n");
1369 if (es->param.exists ())
1370 for (i = 0; i < (int) es->param.length (); i++)
1372 int prob = es->param[i].change_prob;
1374 if (!prob)
1375 fprintf (f, "%*s op%i is compile time invariant\n",
1376 indent + 2, "", i);
1377 else if (prob != REG_BR_PROB_BASE)
1378 fprintf (f, "%*s op%i change %f%% of time\n", indent + 2, "", i,
1379 prob * 100.0 / REG_BR_PROB_BASE);
1381 if (!edge->inline_failed)
1383 fprintf (f, "%*sStack frame offset %i, callee self size %i,"
1384 " callee size %i\n",
1385 indent + 2, "",
1386 (int) inline_summaries->get (callee)->stack_frame_offset,
1387 (int) inline_summaries->get (callee)->estimated_self_stack_size,
1388 (int) inline_summaries->get (callee)->estimated_stack_size);
1389 dump_inline_edge_summary (f, indent + 2, callee, info);
1392 for (edge = node->indirect_calls; edge; edge = edge->next_callee)
1394 struct inline_edge_summary *es = inline_edge_summary (edge);
1395 fprintf (f, "%*sindirect call loop depth:%2i freq:%4i size:%2i"
1396 " time: %2i",
1397 indent, "",
1398 es->loop_depth,
1399 edge->frequency, es->call_stmt_size, es->call_stmt_time);
1400 if (es->predicate)
1402 fprintf (f, "predicate: ");
1403 dump_predicate (f, info->conds, es->predicate);
1405 else
1406 fprintf (f, "\n");
1411 void
1412 dump_inline_summary (FILE *f, struct cgraph_node *node)
1414 if (node->definition)
1416 struct inline_summary *s = inline_summaries->get (node);
1417 size_time_entry *e;
1418 int i;
1419 fprintf (f, "Inline summary for %s/%i", node->name (),
1420 node->order);
1421 if (DECL_DISREGARD_INLINE_LIMITS (node->decl))
1422 fprintf (f, " always_inline");
1423 if (s->inlinable)
1424 fprintf (f, " inlinable");
1425 if (s->contains_cilk_spawn)
1426 fprintf (f, " contains_cilk_spawn");
1427 if (s->fp_expressions)
1428 fprintf (f, " fp_expression");
1429 fprintf (f, "\n self time: %i\n", s->self_time);
1430 fprintf (f, " global time: %i\n", s->time);
1431 fprintf (f, " self size: %i\n", s->self_size);
1432 fprintf (f, " global size: %i\n", s->size);
1433 fprintf (f, " min size: %i\n", s->min_size);
1434 fprintf (f, " self stack: %i\n",
1435 (int) s->estimated_self_stack_size);
1436 fprintf (f, " global stack: %i\n", (int) s->estimated_stack_size);
1437 if (s->growth)
1438 fprintf (f, " estimated growth:%i\n", (int) s->growth);
1439 if (s->scc_no)
1440 fprintf (f, " In SCC: %i\n", (int) s->scc_no);
1441 for (i = 0; vec_safe_iterate (s->entry, i, &e); i++)
1443 fprintf (f, " size:%f, time:%f, predicate:",
1444 (double) e->size / INLINE_SIZE_SCALE,
1445 (double) e->time / INLINE_TIME_SCALE);
1446 dump_predicate (f, s->conds, &e->predicate);
1448 if (s->loop_iterations)
1450 fprintf (f, " loop iterations:");
1451 dump_predicate (f, s->conds, s->loop_iterations);
1453 if (s->loop_stride)
1455 fprintf (f, " loop stride:");
1456 dump_predicate (f, s->conds, s->loop_stride);
1458 if (s->array_index)
1460 fprintf (f, " array index:");
1461 dump_predicate (f, s->conds, s->array_index);
1463 fprintf (f, " calls:\n");
1464 dump_inline_edge_summary (f, 4, node, s);
1465 fprintf (f, "\n");
1469 DEBUG_FUNCTION void
1470 debug_inline_summary (struct cgraph_node *node)
1472 dump_inline_summary (stderr, node);
1475 void
1476 dump_inline_summaries (FILE *f)
1478 struct cgraph_node *node;
1480 FOR_EACH_DEFINED_FUNCTION (node)
1481 if (!node->global.inlined_to)
1482 dump_inline_summary (f, node);
1485 /* Give initial reasons why inlining would fail on EDGE. This gets either
1486 nullified or usually overwritten by more precise reasons later. */
1488 void
1489 initialize_inline_failed (struct cgraph_edge *e)
1491 struct cgraph_node *callee = e->callee;
1493 if (e->inline_failed && e->inline_failed != CIF_BODY_NOT_AVAILABLE
1494 && cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR)
1496 else if (e->indirect_unknown_callee)
1497 e->inline_failed = CIF_INDIRECT_UNKNOWN_CALL;
1498 else if (!callee->definition)
1499 e->inline_failed = CIF_BODY_NOT_AVAILABLE;
1500 else if (callee->local.redefined_extern_inline)
1501 e->inline_failed = CIF_REDEFINED_EXTERN_INLINE;
1502 else if (cfun && fn_contains_cilk_spawn_p (cfun))
1503 /* We can't inline if the function is spawing a function. */
1504 e->inline_failed = CIF_CILK_SPAWN;
1505 else
1506 e->inline_failed = CIF_FUNCTION_NOT_CONSIDERED;
1507 gcc_checking_assert (!e->call_stmt_cannot_inline_p
1508 || cgraph_inline_failed_type (e->inline_failed)
1509 == CIF_FINAL_ERROR);
1512 /* Callback of walk_aliased_vdefs. Flags that it has been invoked to the
1513 boolean variable pointed to by DATA. */
1515 static bool
1516 mark_modified (ao_ref *ao ATTRIBUTE_UNUSED, tree vdef ATTRIBUTE_UNUSED,
1517 void *data)
1519 bool *b = (bool *) data;
1520 *b = true;
1521 return true;
1524 /* If OP refers to value of function parameter, return the corresponding
1525 parameter. */
1527 static tree
1528 unmodified_parm_1 (gimple *stmt, tree op)
1530 /* SSA_NAME referring to parm default def? */
1531 if (TREE_CODE (op) == SSA_NAME
1532 && SSA_NAME_IS_DEFAULT_DEF (op)
1533 && TREE_CODE (SSA_NAME_VAR (op)) == PARM_DECL)
1534 return SSA_NAME_VAR (op);
1535 /* Non-SSA parm reference? */
1536 if (TREE_CODE (op) == PARM_DECL)
1538 bool modified = false;
1540 ao_ref refd;
1541 ao_ref_init (&refd, op);
1542 walk_aliased_vdefs (&refd, gimple_vuse (stmt), mark_modified, &modified,
1543 NULL);
1544 if (!modified)
1545 return op;
1547 return NULL_TREE;
1550 /* If OP refers to value of function parameter, return the corresponding
1551 parameter. Also traverse chains of SSA register assignments. */
1553 static tree
1554 unmodified_parm (gimple *stmt, tree op)
1556 tree res = unmodified_parm_1 (stmt, op);
1557 if (res)
1558 return res;
1560 if (TREE_CODE (op) == SSA_NAME
1561 && !SSA_NAME_IS_DEFAULT_DEF (op)
1562 && gimple_assign_single_p (SSA_NAME_DEF_STMT (op)))
1563 return unmodified_parm (SSA_NAME_DEF_STMT (op),
1564 gimple_assign_rhs1 (SSA_NAME_DEF_STMT (op)));
1565 return NULL_TREE;
1568 /* If OP refers to a value of a function parameter or value loaded from an
1569 aggregate passed to a parameter (either by value or reference), return TRUE
1570 and store the number of the parameter to *INDEX_P and information whether
1571 and how it has been loaded from an aggregate into *AGGPOS. INFO describes
1572 the function parameters, STMT is the statement in which OP is used or
1573 loaded. */
1575 static bool
1576 unmodified_parm_or_parm_agg_item (struct ipa_func_body_info *fbi,
1577 gimple *stmt, tree op, int *index_p,
1578 struct agg_position_info *aggpos)
1580 tree res = unmodified_parm_1 (stmt, op);
1582 gcc_checking_assert (aggpos);
1583 if (res)
1585 *index_p = ipa_get_param_decl_index (fbi->info, res);
1586 if (*index_p < 0)
1587 return false;
1588 aggpos->agg_contents = false;
1589 aggpos->by_ref = false;
1590 return true;
1593 if (TREE_CODE (op) == SSA_NAME)
1595 if (SSA_NAME_IS_DEFAULT_DEF (op)
1596 || !gimple_assign_single_p (SSA_NAME_DEF_STMT (op)))
1597 return false;
1598 stmt = SSA_NAME_DEF_STMT (op);
1599 op = gimple_assign_rhs1 (stmt);
1600 if (!REFERENCE_CLASS_P (op))
1601 return unmodified_parm_or_parm_agg_item (fbi, stmt, op, index_p,
1602 aggpos);
1605 aggpos->agg_contents = true;
1606 return ipa_load_from_parm_agg (fbi, fbi->info->descriptors,
1607 stmt, op, index_p, &aggpos->offset,
1608 NULL, &aggpos->by_ref);
1611 /* See if statement might disappear after inlining.
1612 0 - means not eliminated
1613 1 - half of statements goes away
1614 2 - for sure it is eliminated.
1615 We are not terribly sophisticated, basically looking for simple abstraction
1616 penalty wrappers. */
1618 static int
1619 eliminated_by_inlining_prob (gimple *stmt)
1621 enum gimple_code code = gimple_code (stmt);
1622 enum tree_code rhs_code;
1624 if (!optimize)
1625 return 0;
1627 switch (code)
1629 case GIMPLE_RETURN:
1630 return 2;
1631 case GIMPLE_ASSIGN:
1632 if (gimple_num_ops (stmt) != 2)
1633 return 0;
1635 rhs_code = gimple_assign_rhs_code (stmt);
1637 /* Casts of parameters, loads from parameters passed by reference
1638 and stores to return value or parameters are often free after
1639 inlining dua to SRA and further combining.
1640 Assume that half of statements goes away. */
1641 if (CONVERT_EXPR_CODE_P (rhs_code)
1642 || rhs_code == VIEW_CONVERT_EXPR
1643 || rhs_code == ADDR_EXPR
1644 || gimple_assign_rhs_class (stmt) == GIMPLE_SINGLE_RHS)
1646 tree rhs = gimple_assign_rhs1 (stmt);
1647 tree lhs = gimple_assign_lhs (stmt);
1648 tree inner_rhs = get_base_address (rhs);
1649 tree inner_lhs = get_base_address (lhs);
1650 bool rhs_free = false;
1651 bool lhs_free = false;
1653 if (!inner_rhs)
1654 inner_rhs = rhs;
1655 if (!inner_lhs)
1656 inner_lhs = lhs;
1658 /* Reads of parameter are expected to be free. */
1659 if (unmodified_parm (stmt, inner_rhs))
1660 rhs_free = true;
1661 /* Match expressions of form &this->field. Those will most likely
1662 combine with something upstream after inlining. */
1663 else if (TREE_CODE (inner_rhs) == ADDR_EXPR)
1665 tree op = get_base_address (TREE_OPERAND (inner_rhs, 0));
1666 if (TREE_CODE (op) == PARM_DECL)
1667 rhs_free = true;
1668 else if (TREE_CODE (op) == MEM_REF
1669 && unmodified_parm (stmt, TREE_OPERAND (op, 0)))
1670 rhs_free = true;
1673 /* When parameter is not SSA register because its address is taken
1674 and it is just copied into one, the statement will be completely
1675 free after inlining (we will copy propagate backward). */
1676 if (rhs_free && is_gimple_reg (lhs))
1677 return 2;
1679 /* Reads of parameters passed by reference
1680 expected to be free (i.e. optimized out after inlining). */
1681 if (TREE_CODE (inner_rhs) == MEM_REF
1682 && unmodified_parm (stmt, TREE_OPERAND (inner_rhs, 0)))
1683 rhs_free = true;
1685 /* Copying parameter passed by reference into gimple register is
1686 probably also going to copy propagate, but we can't be quite
1687 sure. */
1688 if (rhs_free && is_gimple_reg (lhs))
1689 lhs_free = true;
1691 /* Writes to parameters, parameters passed by value and return value
1692 (either dirrectly or passed via invisible reference) are free.
1694 TODO: We ought to handle testcase like
1695 struct a {int a,b;};
1696 struct a
1697 retrurnsturct (void)
1699 struct a a ={1,2};
1700 return a;
1703 This translate into:
1705 retrurnsturct ()
1707 int a$b;
1708 int a$a;
1709 struct a a;
1710 struct a D.2739;
1712 <bb 2>:
1713 D.2739.a = 1;
1714 D.2739.b = 2;
1715 return D.2739;
1718 For that we either need to copy ipa-split logic detecting writes
1719 to return value. */
1720 if (TREE_CODE (inner_lhs) == PARM_DECL
1721 || TREE_CODE (inner_lhs) == RESULT_DECL
1722 || (TREE_CODE (inner_lhs) == MEM_REF
1723 && (unmodified_parm (stmt, TREE_OPERAND (inner_lhs, 0))
1724 || (TREE_CODE (TREE_OPERAND (inner_lhs, 0)) == SSA_NAME
1725 && SSA_NAME_VAR (TREE_OPERAND (inner_lhs, 0))
1726 && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND
1727 (inner_lhs,
1728 0))) == RESULT_DECL))))
1729 lhs_free = true;
1730 if (lhs_free
1731 && (is_gimple_reg (rhs) || is_gimple_min_invariant (rhs)))
1732 rhs_free = true;
1733 if (lhs_free && rhs_free)
1734 return 1;
1736 return 0;
1737 default:
1738 return 0;
1743 /* If BB ends by a conditional we can turn into predicates, attach corresponding
1744 predicates to the CFG edges. */
1746 static void
1747 set_cond_stmt_execution_predicate (struct ipa_func_body_info *fbi,
1748 struct inline_summary *summary,
1749 basic_block bb)
1751 gimple *last;
1752 tree op;
1753 int index;
1754 struct agg_position_info aggpos;
1755 enum tree_code code, inverted_code;
1756 edge e;
1757 edge_iterator ei;
1758 gimple *set_stmt;
1759 tree op2;
1761 last = last_stmt (bb);
1762 if (!last || gimple_code (last) != GIMPLE_COND)
1763 return;
1764 if (!is_gimple_ip_invariant (gimple_cond_rhs (last)))
1765 return;
1766 op = gimple_cond_lhs (last);
1767 /* TODO: handle conditionals like
1768 var = op0 < 4;
1769 if (var != 0). */
1770 if (unmodified_parm_or_parm_agg_item (fbi, last, op, &index, &aggpos))
1772 code = gimple_cond_code (last);
1773 inverted_code = invert_tree_comparison (code, HONOR_NANS (op));
1775 FOR_EACH_EDGE (e, ei, bb->succs)
1777 enum tree_code this_code = (e->flags & EDGE_TRUE_VALUE
1778 ? code : inverted_code);
1779 /* invert_tree_comparison will return ERROR_MARK on FP
1780 comparsions that are not EQ/NE instead of returning proper
1781 unordered one. Be sure it is not confused with NON_CONSTANT. */
1782 if (this_code != ERROR_MARK)
1784 struct predicate p = add_condition
1785 (summary, index, &aggpos, this_code,
1786 unshare_expr_without_location (gimple_cond_rhs (last)));
1787 e->aux = edge_predicate_pool.allocate ();
1788 *(struct predicate *) e->aux = p;
1793 if (TREE_CODE (op) != SSA_NAME)
1794 return;
1795 /* Special case
1796 if (builtin_constant_p (op))
1797 constant_code
1798 else
1799 nonconstant_code.
1800 Here we can predicate nonconstant_code. We can't
1801 really handle constant_code since we have no predicate
1802 for this and also the constant code is not known to be
1803 optimized away when inliner doen't see operand is constant.
1804 Other optimizers might think otherwise. */
1805 if (gimple_cond_code (last) != NE_EXPR
1806 || !integer_zerop (gimple_cond_rhs (last)))
1807 return;
1808 set_stmt = SSA_NAME_DEF_STMT (op);
1809 if (!gimple_call_builtin_p (set_stmt, BUILT_IN_CONSTANT_P)
1810 || gimple_call_num_args (set_stmt) != 1)
1811 return;
1812 op2 = gimple_call_arg (set_stmt, 0);
1813 if (!unmodified_parm_or_parm_agg_item (fbi, set_stmt, op2, &index, &aggpos))
1814 return;
1815 FOR_EACH_EDGE (e, ei, bb->succs) if (e->flags & EDGE_FALSE_VALUE)
1817 struct predicate p = add_condition (summary, index, &aggpos,
1818 IS_NOT_CONSTANT, NULL_TREE);
1819 e->aux = edge_predicate_pool.allocate ();
1820 *(struct predicate *) e->aux = p;
1825 /* If BB ends by a switch we can turn into predicates, attach corresponding
1826 predicates to the CFG edges. */
1828 static void
1829 set_switch_stmt_execution_predicate (struct ipa_func_body_info *fbi,
1830 struct inline_summary *summary,
1831 basic_block bb)
1833 gimple *lastg;
1834 tree op;
1835 int index;
1836 struct agg_position_info aggpos;
1837 edge e;
1838 edge_iterator ei;
1839 size_t n;
1840 size_t case_idx;
1842 lastg = last_stmt (bb);
1843 if (!lastg || gimple_code (lastg) != GIMPLE_SWITCH)
1844 return;
1845 gswitch *last = as_a <gswitch *> (lastg);
1846 op = gimple_switch_index (last);
1847 if (!unmodified_parm_or_parm_agg_item (fbi, last, op, &index, &aggpos))
1848 return;
1850 FOR_EACH_EDGE (e, ei, bb->succs)
1852 e->aux = edge_predicate_pool.allocate ();
1853 *(struct predicate *) e->aux = false_predicate ();
1855 n = gimple_switch_num_labels (last);
1856 for (case_idx = 0; case_idx < n; ++case_idx)
1858 tree cl = gimple_switch_label (last, case_idx);
1859 tree min, max;
1860 struct predicate p;
1862 e = find_edge (bb, label_to_block (CASE_LABEL (cl)));
1863 min = CASE_LOW (cl);
1864 max = CASE_HIGH (cl);
1866 /* For default we might want to construct predicate that none
1867 of cases is met, but it is bit hard to do not having negations
1868 of conditionals handy. */
1869 if (!min && !max)
1870 p = true_predicate ();
1871 else if (!max)
1872 p = add_condition (summary, index, &aggpos, EQ_EXPR,
1873 unshare_expr_without_location (min));
1874 else
1876 struct predicate p1, p2;
1877 p1 = add_condition (summary, index, &aggpos, GE_EXPR,
1878 unshare_expr_without_location (min));
1879 p2 = add_condition (summary, index, &aggpos, LE_EXPR,
1880 unshare_expr_without_location (max));
1881 p = and_predicates (summary->conds, &p1, &p2);
1883 *(struct predicate *) e->aux
1884 = or_predicates (summary->conds, &p, (struct predicate *) e->aux);
1889 /* For each BB in NODE attach to its AUX pointer predicate under
1890 which it is executable. */
1892 static void
1893 compute_bb_predicates (struct ipa_func_body_info *fbi,
1894 struct cgraph_node *node,
1895 struct inline_summary *summary)
1897 struct function *my_function = DECL_STRUCT_FUNCTION (node->decl);
1898 bool done = false;
1899 basic_block bb;
1901 FOR_EACH_BB_FN (bb, my_function)
1903 set_cond_stmt_execution_predicate (fbi, summary, bb);
1904 set_switch_stmt_execution_predicate (fbi, summary, bb);
1907 /* Entry block is always executable. */
1908 ENTRY_BLOCK_PTR_FOR_FN (my_function)->aux
1909 = edge_predicate_pool.allocate ();
1910 *(struct predicate *) ENTRY_BLOCK_PTR_FOR_FN (my_function)->aux
1911 = true_predicate ();
1913 /* A simple dataflow propagation of predicates forward in the CFG.
1914 TODO: work in reverse postorder. */
1915 while (!done)
1917 done = true;
1918 FOR_EACH_BB_FN (bb, my_function)
1920 struct predicate p = false_predicate ();
1921 edge e;
1922 edge_iterator ei;
1923 FOR_EACH_EDGE (e, ei, bb->preds)
1925 if (e->src->aux)
1927 struct predicate this_bb_predicate
1928 = *(struct predicate *) e->src->aux;
1929 if (e->aux)
1930 this_bb_predicate
1931 = and_predicates (summary->conds, &this_bb_predicate,
1932 (struct predicate *) e->aux);
1933 p = or_predicates (summary->conds, &p, &this_bb_predicate);
1934 if (true_predicate_p (&p))
1935 break;
1938 if (false_predicate_p (&p))
1939 gcc_assert (!bb->aux);
1940 else
1942 if (!bb->aux)
1944 done = false;
1945 bb->aux = edge_predicate_pool.allocate ();
1946 *((struct predicate *) bb->aux) = p;
1948 else if (!predicates_equal_p (&p, (struct predicate *) bb->aux))
1950 /* This OR operation is needed to ensure monotonous data flow
1951 in the case we hit the limit on number of clauses and the
1952 and/or operations above give approximate answers. */
1953 p = or_predicates (summary->conds, &p, (struct predicate *)bb->aux);
1954 if (!predicates_equal_p (&p, (struct predicate *) bb->aux))
1956 done = false;
1957 *((struct predicate *) bb->aux) = p;
1966 /* We keep info about constantness of SSA names. */
1968 typedef struct predicate predicate_t;
1969 /* Return predicate specifying when the STMT might have result that is not
1970 a compile time constant. */
1972 static struct predicate
1973 will_be_nonconstant_expr_predicate (struct ipa_node_params *info,
1974 struct inline_summary *summary,
1975 tree expr,
1976 vec<predicate_t> nonconstant_names)
1978 tree parm;
1979 int index;
1981 while (UNARY_CLASS_P (expr))
1982 expr = TREE_OPERAND (expr, 0);
1984 parm = unmodified_parm (NULL, expr);
1985 if (parm && (index = ipa_get_param_decl_index (info, parm)) >= 0)
1986 return add_condition (summary, index, NULL, CHANGED, NULL_TREE);
1987 if (is_gimple_min_invariant (expr))
1988 return false_predicate ();
1989 if (TREE_CODE (expr) == SSA_NAME)
1990 return nonconstant_names[SSA_NAME_VERSION (expr)];
1991 if (BINARY_CLASS_P (expr) || COMPARISON_CLASS_P (expr))
1993 struct predicate p1 = will_be_nonconstant_expr_predicate
1994 (info, summary, TREE_OPERAND (expr, 0),
1995 nonconstant_names);
1996 struct predicate p2;
1997 if (true_predicate_p (&p1))
1998 return p1;
1999 p2 = will_be_nonconstant_expr_predicate (info, summary,
2000 TREE_OPERAND (expr, 1),
2001 nonconstant_names);
2002 return or_predicates (summary->conds, &p1, &p2);
2004 else if (TREE_CODE (expr) == COND_EXPR)
2006 struct predicate p1 = will_be_nonconstant_expr_predicate
2007 (info, summary, TREE_OPERAND (expr, 0),
2008 nonconstant_names);
2009 struct predicate p2;
2010 if (true_predicate_p (&p1))
2011 return p1;
2012 p2 = will_be_nonconstant_expr_predicate (info, summary,
2013 TREE_OPERAND (expr, 1),
2014 nonconstant_names);
2015 if (true_predicate_p (&p2))
2016 return p2;
2017 p1 = or_predicates (summary->conds, &p1, &p2);
2018 p2 = will_be_nonconstant_expr_predicate (info, summary,
2019 TREE_OPERAND (expr, 2),
2020 nonconstant_names);
2021 return or_predicates (summary->conds, &p1, &p2);
2023 else
2025 debug_tree (expr);
2026 gcc_unreachable ();
2028 return false_predicate ();
2032 /* Return predicate specifying when the STMT might have result that is not
2033 a compile time constant. */
2035 static struct predicate
2036 will_be_nonconstant_predicate (struct ipa_func_body_info *fbi,
2037 struct inline_summary *summary,
2038 gimple *stmt,
2039 vec<predicate_t> nonconstant_names)
2041 struct predicate p = true_predicate ();
2042 ssa_op_iter iter;
2043 tree use;
2044 struct predicate op_non_const;
2045 bool is_load;
2046 int base_index;
2047 struct agg_position_info aggpos;
2049 /* What statments might be optimized away
2050 when their arguments are constant. */
2051 if (gimple_code (stmt) != GIMPLE_ASSIGN
2052 && gimple_code (stmt) != GIMPLE_COND
2053 && gimple_code (stmt) != GIMPLE_SWITCH
2054 && (gimple_code (stmt) != GIMPLE_CALL
2055 || !(gimple_call_flags (stmt) & ECF_CONST)))
2056 return p;
2058 /* Stores will stay anyway. */
2059 if (gimple_store_p (stmt))
2060 return p;
2062 is_load = gimple_assign_load_p (stmt);
2064 /* Loads can be optimized when the value is known. */
2065 if (is_load)
2067 tree op;
2068 gcc_assert (gimple_assign_single_p (stmt));
2069 op = gimple_assign_rhs1 (stmt);
2070 if (!unmodified_parm_or_parm_agg_item (fbi, stmt, op, &base_index,
2071 &aggpos))
2072 return p;
2074 else
2075 base_index = -1;
2077 /* See if we understand all operands before we start
2078 adding conditionals. */
2079 FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE)
2081 tree parm = unmodified_parm (stmt, use);
2082 /* For arguments we can build a condition. */
2083 if (parm && ipa_get_param_decl_index (fbi->info, parm) >= 0)
2084 continue;
2085 if (TREE_CODE (use) != SSA_NAME)
2086 return p;
2087 /* If we know when operand is constant,
2088 we still can say something useful. */
2089 if (!true_predicate_p (&nonconstant_names[SSA_NAME_VERSION (use)]))
2090 continue;
2091 return p;
2094 if (is_load)
2095 op_non_const =
2096 add_condition (summary, base_index, &aggpos, CHANGED, NULL);
2097 else
2098 op_non_const = false_predicate ();
2099 FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE)
2101 tree parm = unmodified_parm (stmt, use);
2102 int index;
2104 if (parm && (index = ipa_get_param_decl_index (fbi->info, parm)) >= 0)
2106 if (index != base_index)
2107 p = add_condition (summary, index, NULL, CHANGED, NULL_TREE);
2108 else
2109 continue;
2111 else
2112 p = nonconstant_names[SSA_NAME_VERSION (use)];
2113 op_non_const = or_predicates (summary->conds, &p, &op_non_const);
2115 if ((gimple_code (stmt) == GIMPLE_ASSIGN || gimple_code (stmt) == GIMPLE_CALL)
2116 && gimple_op (stmt, 0)
2117 && TREE_CODE (gimple_op (stmt, 0)) == SSA_NAME)
2118 nonconstant_names[SSA_NAME_VERSION (gimple_op (stmt, 0))]
2119 = op_non_const;
2120 return op_non_const;
2123 struct record_modified_bb_info
2125 bitmap bb_set;
2126 gimple *stmt;
2129 /* Callback of walk_aliased_vdefs. Records basic blocks where the value may be
2130 set except for info->stmt. */
2132 static bool
2133 record_modified (ao_ref *ao ATTRIBUTE_UNUSED, tree vdef, void *data)
2135 struct record_modified_bb_info *info =
2136 (struct record_modified_bb_info *) data;
2137 if (SSA_NAME_DEF_STMT (vdef) == info->stmt)
2138 return false;
2139 bitmap_set_bit (info->bb_set,
2140 SSA_NAME_IS_DEFAULT_DEF (vdef)
2141 ? ENTRY_BLOCK_PTR_FOR_FN (cfun)->index
2142 : gimple_bb (SSA_NAME_DEF_STMT (vdef))->index);
2143 return false;
2146 /* Return probability (based on REG_BR_PROB_BASE) that I-th parameter of STMT
2147 will change since last invocation of STMT.
2149 Value 0 is reserved for compile time invariants.
2150 For common parameters it is REG_BR_PROB_BASE. For loop invariants it
2151 ought to be REG_BR_PROB_BASE / estimated_iters. */
2153 static int
2154 param_change_prob (gimple *stmt, int i)
2156 tree op = gimple_call_arg (stmt, i);
2157 basic_block bb = gimple_bb (stmt);
2158 tree base;
2160 /* Global invariants neve change. */
2161 if (is_gimple_min_invariant (op))
2162 return 0;
2163 /* We would have to do non-trivial analysis to really work out what
2164 is the probability of value to change (i.e. when init statement
2165 is in a sibling loop of the call).
2167 We do an conservative estimate: when call is executed N times more often
2168 than the statement defining value, we take the frequency 1/N. */
2169 if (TREE_CODE (op) == SSA_NAME)
2171 int init_freq;
2173 if (!bb->frequency)
2174 return REG_BR_PROB_BASE;
2176 if (SSA_NAME_IS_DEFAULT_DEF (op))
2177 init_freq = ENTRY_BLOCK_PTR_FOR_FN (cfun)->frequency;
2178 else
2179 init_freq = gimple_bb (SSA_NAME_DEF_STMT (op))->frequency;
2181 if (!init_freq)
2182 init_freq = 1;
2183 if (init_freq < bb->frequency)
2184 return MAX (GCOV_COMPUTE_SCALE (init_freq, bb->frequency), 1);
2185 else
2186 return REG_BR_PROB_BASE;
2189 base = get_base_address (op);
2190 if (base)
2192 ao_ref refd;
2193 int max;
2194 struct record_modified_bb_info info;
2195 bitmap_iterator bi;
2196 unsigned index;
2197 tree init = ctor_for_folding (base);
2199 if (init != error_mark_node)
2200 return 0;
2201 if (!bb->frequency)
2202 return REG_BR_PROB_BASE;
2203 ao_ref_init (&refd, op);
2204 info.stmt = stmt;
2205 info.bb_set = BITMAP_ALLOC (NULL);
2206 walk_aliased_vdefs (&refd, gimple_vuse (stmt), record_modified, &info,
2207 NULL);
2208 if (bitmap_bit_p (info.bb_set, bb->index))
2210 BITMAP_FREE (info.bb_set);
2211 return REG_BR_PROB_BASE;
2214 /* Assume that every memory is initialized at entry.
2215 TODO: Can we easilly determine if value is always defined
2216 and thus we may skip entry block? */
2217 if (ENTRY_BLOCK_PTR_FOR_FN (cfun)->frequency)
2218 max = ENTRY_BLOCK_PTR_FOR_FN (cfun)->frequency;
2219 else
2220 max = 1;
2222 EXECUTE_IF_SET_IN_BITMAP (info.bb_set, 0, index, bi)
2223 max = MIN (max, BASIC_BLOCK_FOR_FN (cfun, index)->frequency);
2225 BITMAP_FREE (info.bb_set);
2226 if (max < bb->frequency)
2227 return MAX (GCOV_COMPUTE_SCALE (max, bb->frequency), 1);
2228 else
2229 return REG_BR_PROB_BASE;
2231 return REG_BR_PROB_BASE;
2234 /* Find whether a basic block BB is the final block of a (half) diamond CFG
2235 sub-graph and if the predicate the condition depends on is known. If so,
2236 return true and store the pointer the predicate in *P. */
2238 static bool
2239 phi_result_unknown_predicate (struct ipa_node_params *info,
2240 inline_summary *summary, basic_block bb,
2241 struct predicate *p,
2242 vec<predicate_t> nonconstant_names)
2244 edge e;
2245 edge_iterator ei;
2246 basic_block first_bb = NULL;
2247 gimple *stmt;
2249 if (single_pred_p (bb))
2251 *p = false_predicate ();
2252 return true;
2255 FOR_EACH_EDGE (e, ei, bb->preds)
2257 if (single_succ_p (e->src))
2259 if (!single_pred_p (e->src))
2260 return false;
2261 if (!first_bb)
2262 first_bb = single_pred (e->src);
2263 else if (single_pred (e->src) != first_bb)
2264 return false;
2266 else
2268 if (!first_bb)
2269 first_bb = e->src;
2270 else if (e->src != first_bb)
2271 return false;
2275 if (!first_bb)
2276 return false;
2278 stmt = last_stmt (first_bb);
2279 if (!stmt
2280 || gimple_code (stmt) != GIMPLE_COND
2281 || !is_gimple_ip_invariant (gimple_cond_rhs (stmt)))
2282 return false;
2284 *p = will_be_nonconstant_expr_predicate (info, summary,
2285 gimple_cond_lhs (stmt),
2286 nonconstant_names);
2287 if (true_predicate_p (p))
2288 return false;
2289 else
2290 return true;
2293 /* Given a PHI statement in a function described by inline properties SUMMARY
2294 and *P being the predicate describing whether the selected PHI argument is
2295 known, store a predicate for the result of the PHI statement into
2296 NONCONSTANT_NAMES, if possible. */
2298 static void
2299 predicate_for_phi_result (struct inline_summary *summary, gphi *phi,
2300 struct predicate *p,
2301 vec<predicate_t> nonconstant_names)
2303 unsigned i;
2305 for (i = 0; i < gimple_phi_num_args (phi); i++)
2307 tree arg = gimple_phi_arg (phi, i)->def;
2308 if (!is_gimple_min_invariant (arg))
2310 gcc_assert (TREE_CODE (arg) == SSA_NAME);
2311 *p = or_predicates (summary->conds, p,
2312 &nonconstant_names[SSA_NAME_VERSION (arg)]);
2313 if (true_predicate_p (p))
2314 return;
2318 if (dump_file && (dump_flags & TDF_DETAILS))
2320 fprintf (dump_file, "\t\tphi predicate: ");
2321 dump_predicate (dump_file, summary->conds, p);
2323 nonconstant_names[SSA_NAME_VERSION (gimple_phi_result (phi))] = *p;
2326 /* Return predicate specifying when array index in access OP becomes non-constant. */
2328 static struct predicate
2329 array_index_predicate (inline_summary *info,
2330 vec< predicate_t> nonconstant_names, tree op)
2332 struct predicate p = false_predicate ();
2333 while (handled_component_p (op))
2335 if (TREE_CODE (op) == ARRAY_REF || TREE_CODE (op) == ARRAY_RANGE_REF)
2337 if (TREE_CODE (TREE_OPERAND (op, 1)) == SSA_NAME)
2338 p = or_predicates (info->conds, &p,
2339 &nonconstant_names[SSA_NAME_VERSION
2340 (TREE_OPERAND (op, 1))]);
2342 op = TREE_OPERAND (op, 0);
2344 return p;
2347 /* For a typical usage of __builtin_expect (a<b, 1), we
2348 may introduce an extra relation stmt:
2349 With the builtin, we have
2350 t1 = a <= b;
2351 t2 = (long int) t1;
2352 t3 = __builtin_expect (t2, 1);
2353 if (t3 != 0)
2354 goto ...
2355 Without the builtin, we have
2356 if (a<=b)
2357 goto...
2358 This affects the size/time estimation and may have
2359 an impact on the earlier inlining.
2360 Here find this pattern and fix it up later. */
2362 static gimple *
2363 find_foldable_builtin_expect (basic_block bb)
2365 gimple_stmt_iterator bsi;
2367 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
2369 gimple *stmt = gsi_stmt (bsi);
2370 if (gimple_call_builtin_p (stmt, BUILT_IN_EXPECT)
2371 || (is_gimple_call (stmt)
2372 && gimple_call_internal_p (stmt)
2373 && gimple_call_internal_fn (stmt) == IFN_BUILTIN_EXPECT))
2375 tree var = gimple_call_lhs (stmt);
2376 tree arg = gimple_call_arg (stmt, 0);
2377 use_operand_p use_p;
2378 gimple *use_stmt;
2379 bool match = false;
2380 bool done = false;
2382 if (!var || !arg)
2383 continue;
2384 gcc_assert (TREE_CODE (var) == SSA_NAME);
2386 while (TREE_CODE (arg) == SSA_NAME)
2388 gimple *stmt_tmp = SSA_NAME_DEF_STMT (arg);
2389 if (!is_gimple_assign (stmt_tmp))
2390 break;
2391 switch (gimple_assign_rhs_code (stmt_tmp))
2393 case LT_EXPR:
2394 case LE_EXPR:
2395 case GT_EXPR:
2396 case GE_EXPR:
2397 case EQ_EXPR:
2398 case NE_EXPR:
2399 match = true;
2400 done = true;
2401 break;
2402 CASE_CONVERT:
2403 break;
2404 default:
2405 done = true;
2406 break;
2408 if (done)
2409 break;
2410 arg = gimple_assign_rhs1 (stmt_tmp);
2413 if (match && single_imm_use (var, &use_p, &use_stmt)
2414 && gimple_code (use_stmt) == GIMPLE_COND)
2415 return use_stmt;
2418 return NULL;
2421 /* Return true when the basic blocks contains only clobbers followed by RESX.
2422 Such BBs are kept around to make removal of dead stores possible with
2423 presence of EH and will be optimized out by optimize_clobbers later in the
2424 game.
2426 NEED_EH is used to recurse in case the clobber has non-EH predecestors
2427 that can be clobber only, too.. When it is false, the RESX is not necessary
2428 on the end of basic block. */
2430 static bool
2431 clobber_only_eh_bb_p (basic_block bb, bool need_eh = true)
2433 gimple_stmt_iterator gsi = gsi_last_bb (bb);
2434 edge_iterator ei;
2435 edge e;
2437 if (need_eh)
2439 if (gsi_end_p (gsi))
2440 return false;
2441 if (gimple_code (gsi_stmt (gsi)) != GIMPLE_RESX)
2442 return false;
2443 gsi_prev (&gsi);
2445 else if (!single_succ_p (bb))
2446 return false;
2448 for (; !gsi_end_p (gsi); gsi_prev (&gsi))
2450 gimple *stmt = gsi_stmt (gsi);
2451 if (is_gimple_debug (stmt))
2452 continue;
2453 if (gimple_clobber_p (stmt))
2454 continue;
2455 if (gimple_code (stmt) == GIMPLE_LABEL)
2456 break;
2457 return false;
2460 /* See if all predecestors are either throws or clobber only BBs. */
2461 FOR_EACH_EDGE (e, ei, bb->preds)
2462 if (!(e->flags & EDGE_EH)
2463 && !clobber_only_eh_bb_p (e->src, false))
2464 return false;
2466 return true;
2469 /* Return true if STMT compute a floating point expression that may be affected
2470 by -ffast-math and similar flags. */
2472 static bool
2473 fp_expression_p (gimple *stmt)
2475 ssa_op_iter i;
2476 tree op;
2478 FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_DEF|SSA_OP_USE)
2479 if (FLOAT_TYPE_P (TREE_TYPE (op)))
2480 return true;
2481 return false;
2484 /* Compute function body size parameters for NODE.
2485 When EARLY is true, we compute only simple summaries without
2486 non-trivial predicates to drive the early inliner. */
2488 static void
2489 estimate_function_body_sizes (struct cgraph_node *node, bool early)
2491 gcov_type time = 0;
2492 /* Estimate static overhead for function prologue/epilogue and alignment. */
2493 int size = 2;
2494 /* Benefits are scaled by probability of elimination that is in range
2495 <0,2>. */
2496 basic_block bb;
2497 struct function *my_function = DECL_STRUCT_FUNCTION (node->decl);
2498 int freq;
2499 struct inline_summary *info = inline_summaries->get (node);
2500 struct predicate bb_predicate;
2501 struct ipa_func_body_info fbi;
2502 vec<predicate_t> nonconstant_names = vNULL;
2503 int nblocks, n;
2504 int *order;
2505 predicate array_index = true_predicate ();
2506 gimple *fix_builtin_expect_stmt;
2508 gcc_assert (my_function && my_function->cfg);
2509 gcc_assert (cfun == my_function);
2511 memset(&fbi, 0, sizeof(fbi));
2512 info->conds = NULL;
2513 info->entry = NULL;
2515 /* When optimizing and analyzing for IPA inliner, initialize loop optimizer
2516 so we can produce proper inline hints.
2518 When optimizing and analyzing for early inliner, initialize node params
2519 so we can produce correct BB predicates. */
2521 if (opt_for_fn (node->decl, optimize))
2523 calculate_dominance_info (CDI_DOMINATORS);
2524 if (!early)
2525 loop_optimizer_init (LOOPS_NORMAL | LOOPS_HAVE_RECORDED_EXITS);
2526 else
2528 ipa_check_create_node_params ();
2529 ipa_initialize_node_params (node);
2532 if (ipa_node_params_sum)
2534 fbi.node = node;
2535 fbi.info = IPA_NODE_REF (node);
2536 fbi.bb_infos = vNULL;
2537 fbi.bb_infos.safe_grow_cleared (last_basic_block_for_fn (cfun));
2538 fbi.param_count = count_formal_params(node->decl);
2539 nonconstant_names.safe_grow_cleared
2540 (SSANAMES (my_function)->length ());
2544 if (dump_file)
2545 fprintf (dump_file, "\nAnalyzing function body size: %s\n",
2546 node->name ());
2548 /* When we run into maximal number of entries, we assign everything to the
2549 constant truth case. Be sure to have it in list. */
2550 bb_predicate = true_predicate ();
2551 account_size_time (info, 0, 0, &bb_predicate);
2553 bb_predicate = not_inlined_predicate ();
2554 account_size_time (info, 2 * INLINE_SIZE_SCALE, 0, &bb_predicate);
2556 if (fbi.info)
2557 compute_bb_predicates (&fbi, node, info);
2558 order = XNEWVEC (int, n_basic_blocks_for_fn (cfun));
2559 nblocks = pre_and_rev_post_order_compute (NULL, order, false);
2560 for (n = 0; n < nblocks; n++)
2562 bb = BASIC_BLOCK_FOR_FN (cfun, order[n]);
2563 freq = compute_call_stmt_bb_frequency (node->decl, bb);
2564 if (clobber_only_eh_bb_p (bb))
2566 if (dump_file && (dump_flags & TDF_DETAILS))
2567 fprintf (dump_file, "\n Ignoring BB %i;"
2568 " it will be optimized away by cleanup_clobbers\n",
2569 bb->index);
2570 continue;
2573 /* TODO: Obviously predicates can be propagated down across CFG. */
2574 if (fbi.info)
2576 if (bb->aux)
2577 bb_predicate = *(struct predicate *) bb->aux;
2578 else
2579 bb_predicate = false_predicate ();
2581 else
2582 bb_predicate = true_predicate ();
2584 if (dump_file && (dump_flags & TDF_DETAILS))
2586 fprintf (dump_file, "\n BB %i predicate:", bb->index);
2587 dump_predicate (dump_file, info->conds, &bb_predicate);
2590 if (fbi.info && nonconstant_names.exists ())
2592 struct predicate phi_predicate;
2593 bool first_phi = true;
2595 for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi);
2596 gsi_next (&bsi))
2598 if (first_phi
2599 && !phi_result_unknown_predicate (fbi.info, info, bb,
2600 &phi_predicate,
2601 nonconstant_names))
2602 break;
2603 first_phi = false;
2604 if (dump_file && (dump_flags & TDF_DETAILS))
2606 fprintf (dump_file, " ");
2607 print_gimple_stmt (dump_file, gsi_stmt (bsi), 0, 0);
2609 predicate_for_phi_result (info, bsi.phi (), &phi_predicate,
2610 nonconstant_names);
2614 fix_builtin_expect_stmt = find_foldable_builtin_expect (bb);
2616 for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi);
2617 gsi_next (&bsi))
2619 gimple *stmt = gsi_stmt (bsi);
2620 int this_size = estimate_num_insns (stmt, &eni_size_weights);
2621 int this_time = estimate_num_insns (stmt, &eni_time_weights);
2622 int prob;
2623 struct predicate will_be_nonconstant;
2625 /* This relation stmt should be folded after we remove
2626 buildin_expect call. Adjust the cost here. */
2627 if (stmt == fix_builtin_expect_stmt)
2629 this_size--;
2630 this_time--;
2633 if (dump_file && (dump_flags & TDF_DETAILS))
2635 fprintf (dump_file, " ");
2636 print_gimple_stmt (dump_file, stmt, 0, 0);
2637 fprintf (dump_file, "\t\tfreq:%3.2f size:%3i time:%3i\n",
2638 ((double) freq) / CGRAPH_FREQ_BASE, this_size,
2639 this_time);
2642 if (gimple_assign_load_p (stmt) && nonconstant_names.exists ())
2644 struct predicate this_array_index;
2645 this_array_index =
2646 array_index_predicate (info, nonconstant_names,
2647 gimple_assign_rhs1 (stmt));
2648 if (!false_predicate_p (&this_array_index))
2649 array_index =
2650 and_predicates (info->conds, &array_index,
2651 &this_array_index);
2653 if (gimple_store_p (stmt) && nonconstant_names.exists ())
2655 struct predicate this_array_index;
2656 this_array_index =
2657 array_index_predicate (info, nonconstant_names,
2658 gimple_get_lhs (stmt));
2659 if (!false_predicate_p (&this_array_index))
2660 array_index =
2661 and_predicates (info->conds, &array_index,
2662 &this_array_index);
2666 if (is_gimple_call (stmt)
2667 && !gimple_call_internal_p (stmt))
2669 struct cgraph_edge *edge = node->get_edge (stmt);
2670 struct inline_edge_summary *es = inline_edge_summary (edge);
2672 /* Special case: results of BUILT_IN_CONSTANT_P will be always
2673 resolved as constant. We however don't want to optimize
2674 out the cgraph edges. */
2675 if (nonconstant_names.exists ()
2676 && gimple_call_builtin_p (stmt, BUILT_IN_CONSTANT_P)
2677 && gimple_call_lhs (stmt)
2678 && TREE_CODE (gimple_call_lhs (stmt)) == SSA_NAME)
2680 struct predicate false_p = false_predicate ();
2681 nonconstant_names[SSA_NAME_VERSION (gimple_call_lhs (stmt))]
2682 = false_p;
2684 if (ipa_node_params_sum)
2686 int count = gimple_call_num_args (stmt);
2687 int i;
2689 if (count)
2690 es->param.safe_grow_cleared (count);
2691 for (i = 0; i < count; i++)
2693 int prob = param_change_prob (stmt, i);
2694 gcc_assert (prob >= 0 && prob <= REG_BR_PROB_BASE);
2695 es->param[i].change_prob = prob;
2699 es->call_stmt_size = this_size;
2700 es->call_stmt_time = this_time;
2701 es->loop_depth = bb_loop_depth (bb);
2702 edge_set_predicate (edge, &bb_predicate);
2705 /* TODO: When conditional jump or swithc is known to be constant, but
2706 we did not translate it into the predicates, we really can account
2707 just maximum of the possible paths. */
2708 if (fbi.info)
2709 will_be_nonconstant
2710 = will_be_nonconstant_predicate (&fbi, info,
2711 stmt, nonconstant_names);
2712 if (this_time || this_size)
2714 struct predicate p;
2716 this_time *= freq;
2718 prob = eliminated_by_inlining_prob (stmt);
2719 if (prob == 1 && dump_file && (dump_flags & TDF_DETAILS))
2720 fprintf (dump_file,
2721 "\t\t50%% will be eliminated by inlining\n");
2722 if (prob == 2 && dump_file && (dump_flags & TDF_DETAILS))
2723 fprintf (dump_file, "\t\tWill be eliminated by inlining\n");
2725 if (fbi.info)
2726 p = and_predicates (info->conds, &bb_predicate,
2727 &will_be_nonconstant);
2728 else
2729 p = true_predicate ();
2731 if (!false_predicate_p (&p)
2732 || (is_gimple_call (stmt)
2733 && !false_predicate_p (&bb_predicate)))
2735 time += this_time;
2736 size += this_size;
2737 if (time > MAX_TIME * INLINE_TIME_SCALE)
2738 time = MAX_TIME * INLINE_TIME_SCALE;
2741 /* We account everything but the calls. Calls have their own
2742 size/time info attached to cgraph edges. This is necessary
2743 in order to make the cost disappear after inlining. */
2744 if (!is_gimple_call (stmt))
2746 if (prob)
2748 struct predicate ip = not_inlined_predicate ();
2749 ip = and_predicates (info->conds, &ip, &p);
2750 account_size_time (info, this_size * prob,
2751 this_time * prob, &ip);
2753 if (prob != 2)
2754 account_size_time (info, this_size * (2 - prob),
2755 this_time * (2 - prob), &p);
2758 if (!info->fp_expressions && fp_expression_p (stmt))
2760 info->fp_expressions = true;
2761 if (dump_file)
2762 fprintf (dump_file, " fp_expression set\n");
2765 gcc_assert (time >= 0);
2766 gcc_assert (size >= 0);
2770 set_hint_predicate (&inline_summaries->get (node)->array_index, array_index);
2771 time = (time + CGRAPH_FREQ_BASE / 2) / CGRAPH_FREQ_BASE;
2772 if (time > MAX_TIME)
2773 time = MAX_TIME;
2774 free (order);
2776 if (nonconstant_names.exists () && !early)
2778 struct loop *loop;
2779 predicate loop_iterations = true_predicate ();
2780 predicate loop_stride = true_predicate ();
2782 if (dump_file && (dump_flags & TDF_DETAILS))
2783 flow_loops_dump (dump_file, NULL, 0);
2784 scev_initialize ();
2785 FOR_EACH_LOOP (loop, 0)
2787 vec<edge> exits;
2788 edge ex;
2789 unsigned int j;
2790 struct tree_niter_desc niter_desc;
2791 bb_predicate = *(struct predicate *) loop->header->aux;
2793 exits = get_loop_exit_edges (loop);
2794 FOR_EACH_VEC_ELT (exits, j, ex)
2795 if (number_of_iterations_exit (loop, ex, &niter_desc, false)
2796 && !is_gimple_min_invariant (niter_desc.niter))
2798 predicate will_be_nonconstant
2799 = will_be_nonconstant_expr_predicate (fbi.info, info,
2800 niter_desc.niter,
2801 nonconstant_names);
2802 if (!true_predicate_p (&will_be_nonconstant))
2803 will_be_nonconstant = and_predicates (info->conds,
2804 &bb_predicate,
2805 &will_be_nonconstant);
2806 if (!true_predicate_p (&will_be_nonconstant)
2807 && !false_predicate_p (&will_be_nonconstant))
2808 /* This is slightly inprecise. We may want to represent each
2809 loop with independent predicate. */
2810 loop_iterations =
2811 and_predicates (info->conds, &loop_iterations,
2812 &will_be_nonconstant);
2814 exits.release ();
2817 /* To avoid quadratic behavior we analyze stride predicates only
2818 with respect to the containing loop. Thus we simply iterate
2819 over all defs in the outermost loop body. */
2820 for (loop = loops_for_fn (cfun)->tree_root->inner;
2821 loop != NULL; loop = loop->next)
2823 basic_block *body = get_loop_body (loop);
2824 for (unsigned i = 0; i < loop->num_nodes; i++)
2826 gimple_stmt_iterator gsi;
2827 bb_predicate = *(struct predicate *) body[i]->aux;
2828 for (gsi = gsi_start_bb (body[i]); !gsi_end_p (gsi);
2829 gsi_next (&gsi))
2831 gimple *stmt = gsi_stmt (gsi);
2833 if (!is_gimple_assign (stmt))
2834 continue;
2836 tree def = gimple_assign_lhs (stmt);
2837 if (TREE_CODE (def) != SSA_NAME)
2838 continue;
2840 affine_iv iv;
2841 if (!simple_iv (loop_containing_stmt (stmt),
2842 loop_containing_stmt (stmt),
2843 def, &iv, true)
2844 || is_gimple_min_invariant (iv.step))
2845 continue;
2847 predicate will_be_nonconstant
2848 = will_be_nonconstant_expr_predicate (fbi.info, info,
2849 iv.step,
2850 nonconstant_names);
2851 if (!true_predicate_p (&will_be_nonconstant))
2852 will_be_nonconstant
2853 = and_predicates (info->conds, &bb_predicate,
2854 &will_be_nonconstant);
2855 if (!true_predicate_p (&will_be_nonconstant)
2856 && !false_predicate_p (&will_be_nonconstant))
2857 /* This is slightly inprecise. We may want to represent
2858 each loop with independent predicate. */
2859 loop_stride = and_predicates (info->conds, &loop_stride,
2860 &will_be_nonconstant);
2863 free (body);
2865 set_hint_predicate (&inline_summaries->get (node)->loop_iterations,
2866 loop_iterations);
2867 set_hint_predicate (&inline_summaries->get (node)->loop_stride,
2868 loop_stride);
2869 scev_finalize ();
2871 FOR_ALL_BB_FN (bb, my_function)
2873 edge e;
2874 edge_iterator ei;
2876 if (bb->aux)
2877 edge_predicate_pool.remove ((predicate *)bb->aux);
2878 bb->aux = NULL;
2879 FOR_EACH_EDGE (e, ei, bb->succs)
2881 if (e->aux)
2882 edge_predicate_pool.remove ((predicate *) e->aux);
2883 e->aux = NULL;
2886 inline_summaries->get (node)->self_time = time;
2887 inline_summaries->get (node)->self_size = size;
2888 nonconstant_names.release ();
2889 ipa_release_body_info (&fbi);
2890 if (opt_for_fn (node->decl, optimize))
2892 if (!early)
2893 loop_optimizer_finalize ();
2894 else if (!ipa_edge_args_vector)
2895 ipa_free_all_node_params ();
2896 free_dominance_info (CDI_DOMINATORS);
2898 if (dump_file)
2900 fprintf (dump_file, "\n");
2901 dump_inline_summary (dump_file, node);
2906 /* Compute parameters of functions used by inliner.
2907 EARLY is true when we compute parameters for the early inliner */
2909 void
2910 compute_inline_parameters (struct cgraph_node *node, bool early)
2912 HOST_WIDE_INT self_stack_size;
2913 struct cgraph_edge *e;
2914 struct inline_summary *info;
2916 gcc_assert (!node->global.inlined_to);
2918 inline_summary_alloc ();
2920 info = inline_summaries->get (node);
2921 reset_inline_summary (node, info);
2923 /* Estimate the stack size for the function if we're optimizing. */
2924 self_stack_size = optimize && !node->thunk.thunk_p
2925 ? estimated_stack_frame_size (node) : 0;
2926 info->estimated_self_stack_size = self_stack_size;
2927 info->estimated_stack_size = self_stack_size;
2928 info->stack_frame_offset = 0;
2930 if (node->thunk.thunk_p)
2932 struct inline_edge_summary *es = inline_edge_summary (node->callees);
2933 struct predicate t = true_predicate ();
2935 node->callees->inline_failed = CIF_THUNK;
2936 node->local.can_change_signature = false;
2937 es->call_stmt_size = INLINE_SIZE_SCALE;
2938 es->call_stmt_time = INLINE_TIME_SCALE;
2939 account_size_time (info, INLINE_SIZE_SCALE * 2, INLINE_TIME_SCALE * 2, &t);
2940 inline_update_overall_summary (node);
2941 info->self_size = info->size;
2942 info->self_time = info->time;
2943 /* We can not inline instrumetnation clones. */
2944 info->inlinable = !node->thunk.add_pointer_bounds_args;
2946 else
2948 /* Even is_gimple_min_invariant rely on current_function_decl. */
2949 push_cfun (DECL_STRUCT_FUNCTION (node->decl));
2951 /* Can this function be inlined at all? */
2952 if (!opt_for_fn (node->decl, optimize)
2953 && !lookup_attribute ("always_inline",
2954 DECL_ATTRIBUTES (node->decl)))
2955 info->inlinable = false;
2956 else
2957 info->inlinable = tree_inlinable_function_p (node->decl);
2959 info->contains_cilk_spawn = fn_contains_cilk_spawn_p (cfun);
2961 /* Type attributes can use parameter indices to describe them. */
2962 if (TYPE_ATTRIBUTES (TREE_TYPE (node->decl)))
2963 node->local.can_change_signature = false;
2964 else
2966 /* Otherwise, inlinable functions always can change signature. */
2967 if (info->inlinable)
2968 node->local.can_change_signature = true;
2969 else
2971 /* Functions calling builtin_apply can not change signature. */
2972 for (e = node->callees; e; e = e->next_callee)
2974 tree cdecl = e->callee->decl;
2975 if (DECL_BUILT_IN (cdecl)
2976 && DECL_BUILT_IN_CLASS (cdecl) == BUILT_IN_NORMAL
2977 && (DECL_FUNCTION_CODE (cdecl) == BUILT_IN_APPLY_ARGS
2978 || DECL_FUNCTION_CODE (cdecl) == BUILT_IN_VA_START))
2979 break;
2981 node->local.can_change_signature = !e;
2984 estimate_function_body_sizes (node, early);
2985 pop_cfun ();
2987 for (e = node->callees; e; e = e->next_callee)
2988 if (e->callee->comdat_local_p ())
2989 break;
2990 node->calls_comdat_local = (e != NULL);
2992 /* Inlining characteristics are maintained by the cgraph_mark_inline. */
2993 info->time = info->self_time;
2994 info->size = info->self_size;
2995 info->stack_frame_offset = 0;
2996 info->estimated_stack_size = info->estimated_self_stack_size;
2997 if (flag_checking)
2999 inline_update_overall_summary (node);
3000 gcc_assert (info->time == info->self_time
3001 && info->size == info->self_size);
3006 /* Compute parameters of functions used by inliner using
3007 current_function_decl. */
3009 static unsigned int
3010 compute_inline_parameters_for_current (void)
3012 compute_inline_parameters (cgraph_node::get (current_function_decl), true);
3013 return 0;
3016 namespace {
3018 const pass_data pass_data_inline_parameters =
3020 GIMPLE_PASS, /* type */
3021 "inline_param", /* name */
3022 OPTGROUP_INLINE, /* optinfo_flags */
3023 TV_INLINE_PARAMETERS, /* tv_id */
3024 0, /* properties_required */
3025 0, /* properties_provided */
3026 0, /* properties_destroyed */
3027 0, /* todo_flags_start */
3028 0, /* todo_flags_finish */
3031 class pass_inline_parameters : public gimple_opt_pass
3033 public:
3034 pass_inline_parameters (gcc::context *ctxt)
3035 : gimple_opt_pass (pass_data_inline_parameters, ctxt)
3038 /* opt_pass methods: */
3039 opt_pass * clone () { return new pass_inline_parameters (m_ctxt); }
3040 virtual unsigned int execute (function *)
3042 return compute_inline_parameters_for_current ();
3045 }; // class pass_inline_parameters
3047 } // anon namespace
3049 gimple_opt_pass *
3050 make_pass_inline_parameters (gcc::context *ctxt)
3052 return new pass_inline_parameters (ctxt);
3056 /* Estimate benefit devirtualizing indirect edge IE, provided KNOWN_VALS,
3057 KNOWN_CONTEXTS and KNOWN_AGGS. */
3059 static bool
3060 estimate_edge_devirt_benefit (struct cgraph_edge *ie,
3061 int *size, int *time,
3062 vec<tree> known_vals,
3063 vec<ipa_polymorphic_call_context> known_contexts,
3064 vec<ipa_agg_jump_function_p> known_aggs)
3066 tree target;
3067 struct cgraph_node *callee;
3068 struct inline_summary *isummary;
3069 enum availability avail;
3070 bool speculative;
3072 if (!known_vals.exists () && !known_contexts.exists ())
3073 return false;
3074 if (!opt_for_fn (ie->caller->decl, flag_indirect_inlining))
3075 return false;
3077 target = ipa_get_indirect_edge_target (ie, known_vals, known_contexts,
3078 known_aggs, &speculative);
3079 if (!target || speculative)
3080 return false;
3082 /* Account for difference in cost between indirect and direct calls. */
3083 *size -= (eni_size_weights.indirect_call_cost - eni_size_weights.call_cost);
3084 *time -= (eni_time_weights.indirect_call_cost - eni_time_weights.call_cost);
3085 gcc_checking_assert (*time >= 0);
3086 gcc_checking_assert (*size >= 0);
3088 callee = cgraph_node::get (target);
3089 if (!callee || !callee->definition)
3090 return false;
3091 callee = callee->function_symbol (&avail);
3092 if (avail < AVAIL_AVAILABLE)
3093 return false;
3094 isummary = inline_summaries->get (callee);
3095 return isummary->inlinable;
3098 /* Increase SIZE, MIN_SIZE (if non-NULL) and TIME for size and time needed to
3099 handle edge E with probability PROB.
3100 Set HINTS if edge may be devirtualized.
3101 KNOWN_VALS, KNOWN_AGGS and KNOWN_CONTEXTS describe context of the call
3102 site. */
3104 static inline void
3105 estimate_edge_size_and_time (struct cgraph_edge *e, int *size, int *min_size,
3106 int *time,
3107 int prob,
3108 vec<tree> known_vals,
3109 vec<ipa_polymorphic_call_context> known_contexts,
3110 vec<ipa_agg_jump_function_p> known_aggs,
3111 inline_hints *hints)
3113 struct inline_edge_summary *es = inline_edge_summary (e);
3114 int call_size = es->call_stmt_size;
3115 int call_time = es->call_stmt_time;
3116 int cur_size;
3117 if (!e->callee
3118 && estimate_edge_devirt_benefit (e, &call_size, &call_time,
3119 known_vals, known_contexts, known_aggs)
3120 && hints && e->maybe_hot_p ())
3121 *hints |= INLINE_HINT_indirect_call;
3122 cur_size = call_size * INLINE_SIZE_SCALE;
3123 *size += cur_size;
3124 if (min_size)
3125 *min_size += cur_size;
3126 *time += apply_probability ((gcov_type) call_time, prob)
3127 * e->frequency * (INLINE_TIME_SCALE / CGRAPH_FREQ_BASE);
3128 if (*time > MAX_TIME * INLINE_TIME_SCALE)
3129 *time = MAX_TIME * INLINE_TIME_SCALE;
3134 /* Increase SIZE, MIN_SIZE and TIME for size and time needed to handle all
3135 calls in NODE. POSSIBLE_TRUTHS, KNOWN_VALS, KNOWN_AGGS and KNOWN_CONTEXTS
3136 describe context of the call site. */
3138 static void
3139 estimate_calls_size_and_time (struct cgraph_node *node, int *size,
3140 int *min_size, int *time,
3141 inline_hints *hints,
3142 clause_t possible_truths,
3143 vec<tree> known_vals,
3144 vec<ipa_polymorphic_call_context> known_contexts,
3145 vec<ipa_agg_jump_function_p> known_aggs)
3147 struct cgraph_edge *e;
3148 for (e = node->callees; e; e = e->next_callee)
3150 if (inline_edge_summary_vec.length () <= (unsigned) e->uid)
3151 continue;
3153 struct inline_edge_summary *es = inline_edge_summary (e);
3155 /* Do not care about zero sized builtins. */
3156 if (e->inline_failed && !es->call_stmt_size)
3158 gcc_checking_assert (!es->call_stmt_time);
3159 continue;
3161 if (!es->predicate
3162 || evaluate_predicate (es->predicate, possible_truths))
3164 if (e->inline_failed)
3166 /* Predicates of calls shall not use NOT_CHANGED codes,
3167 sowe do not need to compute probabilities. */
3168 estimate_edge_size_and_time (e, size,
3169 es->predicate ? NULL : min_size,
3170 time, REG_BR_PROB_BASE,
3171 known_vals, known_contexts,
3172 known_aggs, hints);
3174 else
3175 estimate_calls_size_and_time (e->callee, size, min_size, time,
3176 hints,
3177 possible_truths,
3178 known_vals, known_contexts,
3179 known_aggs);
3182 for (e = node->indirect_calls; e; e = e->next_callee)
3184 if (inline_edge_summary_vec.length () <= (unsigned) e->uid)
3185 continue;
3187 struct inline_edge_summary *es = inline_edge_summary (e);
3188 if (!es->predicate
3189 || evaluate_predicate (es->predicate, possible_truths))
3190 estimate_edge_size_and_time (e, size,
3191 es->predicate ? NULL : min_size,
3192 time, REG_BR_PROB_BASE,
3193 known_vals, known_contexts, known_aggs,
3194 hints);
3199 /* Estimate size and time needed to execute NODE assuming
3200 POSSIBLE_TRUTHS clause, and KNOWN_VALS, KNOWN_AGGS and KNOWN_CONTEXTS
3201 information about NODE's arguments. If non-NULL use also probability
3202 information present in INLINE_PARAM_SUMMARY vector.
3203 Additionally detemine hints determined by the context. Finally compute
3204 minimal size needed for the call that is independent on the call context and
3205 can be used for fast estimates. Return the values in RET_SIZE,
3206 RET_MIN_SIZE, RET_TIME and RET_HINTS. */
3208 static void
3209 estimate_node_size_and_time (struct cgraph_node *node,
3210 clause_t possible_truths,
3211 vec<tree> known_vals,
3212 vec<ipa_polymorphic_call_context> known_contexts,
3213 vec<ipa_agg_jump_function_p> known_aggs,
3214 int *ret_size, int *ret_min_size, int *ret_time,
3215 inline_hints *ret_hints,
3216 vec<inline_param_summary>
3217 inline_param_summary)
3219 struct inline_summary *info = inline_summaries->get (node);
3220 size_time_entry *e;
3221 int size = 0;
3222 int time = 0;
3223 int min_size = 0;
3224 inline_hints hints = 0;
3225 int i;
3227 if (dump_file && (dump_flags & TDF_DETAILS))
3229 bool found = false;
3230 fprintf (dump_file, " Estimating body: %s/%i\n"
3231 " Known to be false: ", node->name (),
3232 node->order);
3234 for (i = predicate_not_inlined_condition;
3235 i < (predicate_first_dynamic_condition
3236 + (int) vec_safe_length (info->conds)); i++)
3237 if (!(possible_truths & (1 << i)))
3239 if (found)
3240 fprintf (dump_file, ", ");
3241 found = true;
3242 dump_condition (dump_file, info->conds, i);
3246 for (i = 0; vec_safe_iterate (info->entry, i, &e); i++)
3247 if (evaluate_predicate (&e->predicate, possible_truths))
3249 size += e->size;
3250 gcc_checking_assert (e->time >= 0);
3251 gcc_checking_assert (time >= 0);
3252 if (!inline_param_summary.exists ())
3253 time += e->time;
3254 else
3256 int prob = predicate_probability (info->conds,
3257 &e->predicate,
3258 possible_truths,
3259 inline_param_summary);
3260 gcc_checking_assert (prob >= 0);
3261 gcc_checking_assert (prob <= REG_BR_PROB_BASE);
3262 time += apply_probability ((gcov_type) e->time, prob);
3264 if (time > MAX_TIME * INLINE_TIME_SCALE)
3265 time = MAX_TIME * INLINE_TIME_SCALE;
3266 gcc_checking_assert (time >= 0);
3269 gcc_checking_assert (true_predicate_p (&(*info->entry)[0].predicate));
3270 min_size = (*info->entry)[0].size;
3271 gcc_checking_assert (size >= 0);
3272 gcc_checking_assert (time >= 0);
3274 if (info->loop_iterations
3275 && !evaluate_predicate (info->loop_iterations, possible_truths))
3276 hints |= INLINE_HINT_loop_iterations;
3277 if (info->loop_stride
3278 && !evaluate_predicate (info->loop_stride, possible_truths))
3279 hints |= INLINE_HINT_loop_stride;
3280 if (info->array_index
3281 && !evaluate_predicate (info->array_index, possible_truths))
3282 hints |= INLINE_HINT_array_index;
3283 if (info->scc_no)
3284 hints |= INLINE_HINT_in_scc;
3285 if (DECL_DECLARED_INLINE_P (node->decl))
3286 hints |= INLINE_HINT_declared_inline;
3288 estimate_calls_size_and_time (node, &size, &min_size, &time, &hints, possible_truths,
3289 known_vals, known_contexts, known_aggs);
3290 gcc_checking_assert (size >= 0);
3291 gcc_checking_assert (time >= 0);
3292 time = RDIV (time, INLINE_TIME_SCALE);
3293 size = RDIV (size, INLINE_SIZE_SCALE);
3294 min_size = RDIV (min_size, INLINE_SIZE_SCALE);
3296 if (dump_file && (dump_flags & TDF_DETAILS))
3297 fprintf (dump_file, "\n size:%i time:%i\n", (int) size, (int) time);
3298 if (ret_time)
3299 *ret_time = time;
3300 if (ret_size)
3301 *ret_size = size;
3302 if (ret_min_size)
3303 *ret_min_size = min_size;
3304 if (ret_hints)
3305 *ret_hints = hints;
3306 return;
3310 /* Estimate size and time needed to execute callee of EDGE assuming that
3311 parameters known to be constant at caller of EDGE are propagated.
3312 KNOWN_VALS and KNOWN_CONTEXTS are vectors of assumed known constant values
3313 and types for parameters. */
3315 void
3316 estimate_ipcp_clone_size_and_time (struct cgraph_node *node,
3317 vec<tree> known_vals,
3318 vec<ipa_polymorphic_call_context>
3319 known_contexts,
3320 vec<ipa_agg_jump_function_p> known_aggs,
3321 int *ret_size, int *ret_time,
3322 inline_hints *hints)
3324 clause_t clause;
3326 clause = evaluate_conditions_for_known_args (node, false, known_vals,
3327 known_aggs);
3328 estimate_node_size_and_time (node, clause, known_vals, known_contexts,
3329 known_aggs, ret_size, NULL, ret_time, hints, vNULL);
3332 /* Translate all conditions from callee representation into caller
3333 representation and symbolically evaluate predicate P into new predicate.
3335 INFO is inline_summary of function we are adding predicate into, CALLEE_INFO
3336 is summary of function predicate P is from. OPERAND_MAP is array giving
3337 callee formal IDs the caller formal IDs. POSSSIBLE_TRUTHS is clausule of all
3338 callee conditions that may be true in caller context. TOPLEV_PREDICATE is
3339 predicate under which callee is executed. OFFSET_MAP is an array of of
3340 offsets that need to be added to conditions, negative offset means that
3341 conditions relying on values passed by reference have to be discarded
3342 because they might not be preserved (and should be considered offset zero
3343 for other purposes). */
3345 static struct predicate
3346 remap_predicate (struct inline_summary *info,
3347 struct inline_summary *callee_info,
3348 struct predicate *p,
3349 vec<int> operand_map,
3350 vec<int> offset_map,
3351 clause_t possible_truths, struct predicate *toplev_predicate)
3353 int i;
3354 struct predicate out = true_predicate ();
3356 /* True predicate is easy. */
3357 if (true_predicate_p (p))
3358 return *toplev_predicate;
3359 for (i = 0; p->clause[i]; i++)
3361 clause_t clause = p->clause[i];
3362 int cond;
3363 struct predicate clause_predicate = false_predicate ();
3365 gcc_assert (i < MAX_CLAUSES);
3367 for (cond = 0; cond < NUM_CONDITIONS; cond++)
3368 /* Do we have condition we can't disprove? */
3369 if (clause & possible_truths & (1 << cond))
3371 struct predicate cond_predicate;
3372 /* Work out if the condition can translate to predicate in the
3373 inlined function. */
3374 if (cond >= predicate_first_dynamic_condition)
3376 struct condition *c;
3378 c = &(*callee_info->conds)[cond
3380 predicate_first_dynamic_condition];
3381 /* See if we can remap condition operand to caller's operand.
3382 Otherwise give up. */
3383 if (!operand_map.exists ()
3384 || (int) operand_map.length () <= c->operand_num
3385 || operand_map[c->operand_num] == -1
3386 /* TODO: For non-aggregate conditions, adding an offset is
3387 basically an arithmetic jump function processing which
3388 we should support in future. */
3389 || ((!c->agg_contents || !c->by_ref)
3390 && offset_map[c->operand_num] > 0)
3391 || (c->agg_contents && c->by_ref
3392 && offset_map[c->operand_num] < 0))
3393 cond_predicate = true_predicate ();
3394 else
3396 struct agg_position_info ap;
3397 HOST_WIDE_INT offset_delta = offset_map[c->operand_num];
3398 if (offset_delta < 0)
3400 gcc_checking_assert (!c->agg_contents || !c->by_ref);
3401 offset_delta = 0;
3403 gcc_assert (!c->agg_contents
3404 || c->by_ref || offset_delta == 0);
3405 ap.offset = c->offset + offset_delta;
3406 ap.agg_contents = c->agg_contents;
3407 ap.by_ref = c->by_ref;
3408 cond_predicate = add_condition (info,
3409 operand_map[c->operand_num],
3410 &ap, c->code, c->val);
3413 /* Fixed conditions remains same, construct single
3414 condition predicate. */
3415 else
3417 cond_predicate.clause[0] = 1 << cond;
3418 cond_predicate.clause[1] = 0;
3420 clause_predicate = or_predicates (info->conds, &clause_predicate,
3421 &cond_predicate);
3423 out = and_predicates (info->conds, &out, &clause_predicate);
3425 return and_predicates (info->conds, &out, toplev_predicate);
3429 /* Update summary information of inline clones after inlining.
3430 Compute peak stack usage. */
3432 static void
3433 inline_update_callee_summaries (struct cgraph_node *node, int depth)
3435 struct cgraph_edge *e;
3436 struct inline_summary *callee_info = inline_summaries->get (node);
3437 struct inline_summary *caller_info = inline_summaries->get (node->callers->caller);
3438 HOST_WIDE_INT peak;
3440 callee_info->stack_frame_offset
3441 = caller_info->stack_frame_offset
3442 + caller_info->estimated_self_stack_size;
3443 peak = callee_info->stack_frame_offset
3444 + callee_info->estimated_self_stack_size;
3445 if (inline_summaries->get (node->global.inlined_to)->estimated_stack_size < peak)
3446 inline_summaries->get (node->global.inlined_to)->estimated_stack_size = peak;
3447 ipa_propagate_frequency (node);
3448 for (e = node->callees; e; e = e->next_callee)
3450 if (!e->inline_failed)
3451 inline_update_callee_summaries (e->callee, depth);
3452 inline_edge_summary (e)->loop_depth += depth;
3454 for (e = node->indirect_calls; e; e = e->next_callee)
3455 inline_edge_summary (e)->loop_depth += depth;
3458 /* Update change_prob of EDGE after INLINED_EDGE has been inlined.
3459 When functoin A is inlined in B and A calls C with parameter that
3460 changes with probability PROB1 and C is known to be passthroug
3461 of argument if B that change with probability PROB2, the probability
3462 of change is now PROB1*PROB2. */
3464 static void
3465 remap_edge_change_prob (struct cgraph_edge *inlined_edge,
3466 struct cgraph_edge *edge)
3468 if (ipa_node_params_sum)
3470 int i;
3471 struct ipa_edge_args *args = IPA_EDGE_REF (edge);
3472 struct inline_edge_summary *es = inline_edge_summary (edge);
3473 struct inline_edge_summary *inlined_es
3474 = inline_edge_summary (inlined_edge);
3476 for (i = 0; i < ipa_get_cs_argument_count (args); i++)
3478 struct ipa_jump_func *jfunc = ipa_get_ith_jump_func (args, i);
3479 if (jfunc->type == IPA_JF_PASS_THROUGH
3480 && (ipa_get_jf_pass_through_formal_id (jfunc)
3481 < (int) inlined_es->param.length ()))
3483 int jf_formal_id = ipa_get_jf_pass_through_formal_id (jfunc);
3484 int prob1 = es->param[i].change_prob;
3485 int prob2 = inlined_es->param[jf_formal_id].change_prob;
3486 int prob = combine_probabilities (prob1, prob2);
3488 if (prob1 && prob2 && !prob)
3489 prob = 1;
3491 es->param[i].change_prob = prob;
3497 /* Update edge summaries of NODE after INLINED_EDGE has been inlined.
3499 Remap predicates of callees of NODE. Rest of arguments match
3500 remap_predicate.
3502 Also update change probabilities. */
3504 static void
3505 remap_edge_summaries (struct cgraph_edge *inlined_edge,
3506 struct cgraph_node *node,
3507 struct inline_summary *info,
3508 struct inline_summary *callee_info,
3509 vec<int> operand_map,
3510 vec<int> offset_map,
3511 clause_t possible_truths,
3512 struct predicate *toplev_predicate)
3514 struct cgraph_edge *e, *next;
3515 for (e = node->callees; e; e = next)
3517 struct inline_edge_summary *es = inline_edge_summary (e);
3518 struct predicate p;
3519 next = e->next_callee;
3521 if (e->inline_failed)
3523 remap_edge_change_prob (inlined_edge, e);
3525 if (es->predicate)
3527 p = remap_predicate (info, callee_info,
3528 es->predicate, operand_map, offset_map,
3529 possible_truths, toplev_predicate);
3530 edge_set_predicate (e, &p);
3532 else
3533 edge_set_predicate (e, toplev_predicate);
3535 else
3536 remap_edge_summaries (inlined_edge, e->callee, info, callee_info,
3537 operand_map, offset_map, possible_truths,
3538 toplev_predicate);
3540 for (e = node->indirect_calls; e; e = next)
3542 struct inline_edge_summary *es = inline_edge_summary (e);
3543 struct predicate p;
3544 next = e->next_callee;
3546 remap_edge_change_prob (inlined_edge, e);
3547 if (es->predicate)
3549 p = remap_predicate (info, callee_info,
3550 es->predicate, operand_map, offset_map,
3551 possible_truths, toplev_predicate);
3552 edge_set_predicate (e, &p);
3554 else
3555 edge_set_predicate (e, toplev_predicate);
3559 /* Same as remap_predicate, but set result into hint *HINT. */
3561 static void
3562 remap_hint_predicate (struct inline_summary *info,
3563 struct inline_summary *callee_info,
3564 struct predicate **hint,
3565 vec<int> operand_map,
3566 vec<int> offset_map,
3567 clause_t possible_truths,
3568 struct predicate *toplev_predicate)
3570 predicate p;
3572 if (!*hint)
3573 return;
3574 p = remap_predicate (info, callee_info,
3575 *hint,
3576 operand_map, offset_map,
3577 possible_truths, toplev_predicate);
3578 if (!false_predicate_p (&p) && !true_predicate_p (&p))
3580 if (!*hint)
3581 set_hint_predicate (hint, p);
3582 else
3583 **hint = and_predicates (info->conds, *hint, &p);
3587 /* We inlined EDGE. Update summary of the function we inlined into. */
3589 void
3590 inline_merge_summary (struct cgraph_edge *edge)
3592 struct inline_summary *callee_info = inline_summaries->get (edge->callee);
3593 struct cgraph_node *to = (edge->caller->global.inlined_to
3594 ? edge->caller->global.inlined_to : edge->caller);
3595 struct inline_summary *info = inline_summaries->get (to);
3596 clause_t clause = 0; /* not_inline is known to be false. */
3597 size_time_entry *e;
3598 vec<int> operand_map = vNULL;
3599 vec<int> offset_map = vNULL;
3600 int i;
3601 struct predicate toplev_predicate;
3602 struct predicate true_p = true_predicate ();
3603 struct inline_edge_summary *es = inline_edge_summary (edge);
3605 if (es->predicate)
3606 toplev_predicate = *es->predicate;
3607 else
3608 toplev_predicate = true_predicate ();
3610 info->fp_expressions |= callee_info->fp_expressions;
3612 if (callee_info->conds)
3613 evaluate_properties_for_edge (edge, true, &clause, NULL, NULL, NULL);
3614 if (ipa_node_params_sum && callee_info->conds)
3616 struct ipa_edge_args *args = IPA_EDGE_REF (edge);
3617 int count = ipa_get_cs_argument_count (args);
3618 int i;
3620 if (count)
3622 operand_map.safe_grow_cleared (count);
3623 offset_map.safe_grow_cleared (count);
3625 for (i = 0; i < count; i++)
3627 struct ipa_jump_func *jfunc = ipa_get_ith_jump_func (args, i);
3628 int map = -1;
3630 /* TODO: handle non-NOPs when merging. */
3631 if (jfunc->type == IPA_JF_PASS_THROUGH)
3633 if (ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
3634 map = ipa_get_jf_pass_through_formal_id (jfunc);
3635 if (!ipa_get_jf_pass_through_agg_preserved (jfunc))
3636 offset_map[i] = -1;
3638 else if (jfunc->type == IPA_JF_ANCESTOR)
3640 HOST_WIDE_INT offset = ipa_get_jf_ancestor_offset (jfunc);
3641 if (offset >= 0 && offset < INT_MAX)
3643 map = ipa_get_jf_ancestor_formal_id (jfunc);
3644 if (!ipa_get_jf_ancestor_agg_preserved (jfunc))
3645 offset = -1;
3646 offset_map[i] = offset;
3649 operand_map[i] = map;
3650 gcc_assert (map < ipa_get_param_count (IPA_NODE_REF (to)));
3653 for (i = 0; vec_safe_iterate (callee_info->entry, i, &e); i++)
3655 struct predicate p = remap_predicate (info, callee_info,
3656 &e->predicate, operand_map,
3657 offset_map, clause,
3658 &toplev_predicate);
3659 if (!false_predicate_p (&p))
3661 gcov_type add_time = ((gcov_type) e->time * edge->frequency
3662 + CGRAPH_FREQ_BASE / 2) / CGRAPH_FREQ_BASE;
3663 int prob = predicate_probability (callee_info->conds,
3664 &e->predicate,
3665 clause, es->param);
3666 add_time = apply_probability ((gcov_type) add_time, prob);
3667 if (add_time > MAX_TIME * INLINE_TIME_SCALE)
3668 add_time = MAX_TIME * INLINE_TIME_SCALE;
3669 if (prob != REG_BR_PROB_BASE
3670 && dump_file && (dump_flags & TDF_DETAILS))
3672 fprintf (dump_file, "\t\tScaling time by probability:%f\n",
3673 (double) prob / REG_BR_PROB_BASE);
3675 account_size_time (info, e->size, add_time, &p);
3678 remap_edge_summaries (edge, edge->callee, info, callee_info, operand_map,
3679 offset_map, clause, &toplev_predicate);
3680 remap_hint_predicate (info, callee_info,
3681 &callee_info->loop_iterations,
3682 operand_map, offset_map, clause, &toplev_predicate);
3683 remap_hint_predicate (info, callee_info,
3684 &callee_info->loop_stride,
3685 operand_map, offset_map, clause, &toplev_predicate);
3686 remap_hint_predicate (info, callee_info,
3687 &callee_info->array_index,
3688 operand_map, offset_map, clause, &toplev_predicate);
3690 inline_update_callee_summaries (edge->callee,
3691 inline_edge_summary (edge)->loop_depth);
3693 /* We do not maintain predicates of inlined edges, free it. */
3694 edge_set_predicate (edge, &true_p);
3695 /* Similarly remove param summaries. */
3696 es->param.release ();
3697 operand_map.release ();
3698 offset_map.release ();
3701 /* For performance reasons inline_merge_summary is not updating overall size
3702 and time. Recompute it. */
3704 void
3705 inline_update_overall_summary (struct cgraph_node *node)
3707 struct inline_summary *info = inline_summaries->get (node);
3708 size_time_entry *e;
3709 int i;
3711 info->size = 0;
3712 info->time = 0;
3713 for (i = 0; vec_safe_iterate (info->entry, i, &e); i++)
3715 info->size += e->size, info->time += e->time;
3716 if (info->time > MAX_TIME * INLINE_TIME_SCALE)
3717 info->time = MAX_TIME * INLINE_TIME_SCALE;
3719 estimate_calls_size_and_time (node, &info->size, &info->min_size,
3720 &info->time, NULL,
3721 ~(clause_t) (1 << predicate_false_condition),
3722 vNULL, vNULL, vNULL);
3723 info->time = (info->time + INLINE_TIME_SCALE / 2) / INLINE_TIME_SCALE;
3724 info->size = (info->size + INLINE_SIZE_SCALE / 2) / INLINE_SIZE_SCALE;
3727 /* Return hints derrived from EDGE. */
3729 simple_edge_hints (struct cgraph_edge *edge)
3731 int hints = 0;
3732 struct cgraph_node *to = (edge->caller->global.inlined_to
3733 ? edge->caller->global.inlined_to : edge->caller);
3734 struct cgraph_node *callee = edge->callee->ultimate_alias_target ();
3735 if (inline_summaries->get (to)->scc_no
3736 && inline_summaries->get (to)->scc_no
3737 == inline_summaries->get (callee)->scc_no
3738 && !edge->recursive_p ())
3739 hints |= INLINE_HINT_same_scc;
3741 if (callee->lto_file_data && edge->caller->lto_file_data
3742 && edge->caller->lto_file_data != callee->lto_file_data
3743 && !callee->merged_comdat && !callee->icf_merged)
3744 hints |= INLINE_HINT_cross_module;
3746 return hints;
3749 /* Estimate the time cost for the caller when inlining EDGE.
3750 Only to be called via estimate_edge_time, that handles the
3751 caching mechanism.
3753 When caching, also update the cache entry. Compute both time and
3754 size, since we always need both metrics eventually. */
3757 do_estimate_edge_time (struct cgraph_edge *edge)
3759 int time;
3760 int size;
3761 inline_hints hints;
3762 struct cgraph_node *callee;
3763 clause_t clause;
3764 vec<tree> known_vals;
3765 vec<ipa_polymorphic_call_context> known_contexts;
3766 vec<ipa_agg_jump_function_p> known_aggs;
3767 struct inline_edge_summary *es = inline_edge_summary (edge);
3768 int min_size;
3770 callee = edge->callee->ultimate_alias_target ();
3772 gcc_checking_assert (edge->inline_failed);
3773 evaluate_properties_for_edge (edge, true,
3774 &clause, &known_vals, &known_contexts,
3775 &known_aggs);
3776 estimate_node_size_and_time (callee, clause, known_vals, known_contexts,
3777 known_aggs, &size, &min_size, &time, &hints, es->param);
3779 /* When we have profile feedback, we can quite safely identify hot
3780 edges and for those we disable size limits. Don't do that when
3781 probability that caller will call the callee is low however, since it
3782 may hurt optimization of the caller's hot path. */
3783 if (edge->count && edge->maybe_hot_p ()
3784 && (edge->count * 2
3785 > (edge->caller->global.inlined_to
3786 ? edge->caller->global.inlined_to->count : edge->caller->count)))
3787 hints |= INLINE_HINT_known_hot;
3789 known_vals.release ();
3790 known_contexts.release ();
3791 known_aggs.release ();
3792 gcc_checking_assert (size >= 0);
3793 gcc_checking_assert (time >= 0);
3795 /* When caching, update the cache entry. */
3796 if (edge_growth_cache.exists ())
3798 inline_summaries->get (edge->callee)->min_size = min_size;
3799 if ((int) edge_growth_cache.length () <= edge->uid)
3800 edge_growth_cache.safe_grow_cleared (symtab->edges_max_uid);
3801 edge_growth_cache[edge->uid].time = time + (time >= 0);
3803 edge_growth_cache[edge->uid].size = size + (size >= 0);
3804 hints |= simple_edge_hints (edge);
3805 edge_growth_cache[edge->uid].hints = hints + 1;
3807 return time;
3811 /* Return estimated callee growth after inlining EDGE.
3812 Only to be called via estimate_edge_size. */
3815 do_estimate_edge_size (struct cgraph_edge *edge)
3817 int size;
3818 struct cgraph_node *callee;
3819 clause_t clause;
3820 vec<tree> known_vals;
3821 vec<ipa_polymorphic_call_context> known_contexts;
3822 vec<ipa_agg_jump_function_p> known_aggs;
3824 /* When we do caching, use do_estimate_edge_time to populate the entry. */
3826 if (edge_growth_cache.exists ())
3828 do_estimate_edge_time (edge);
3829 size = edge_growth_cache[edge->uid].size;
3830 gcc_checking_assert (size);
3831 return size - (size > 0);
3834 callee = edge->callee->ultimate_alias_target ();
3836 /* Early inliner runs without caching, go ahead and do the dirty work. */
3837 gcc_checking_assert (edge->inline_failed);
3838 evaluate_properties_for_edge (edge, true,
3839 &clause, &known_vals, &known_contexts,
3840 &known_aggs);
3841 estimate_node_size_and_time (callee, clause, known_vals, known_contexts,
3842 known_aggs, &size, NULL, NULL, NULL, vNULL);
3843 known_vals.release ();
3844 known_contexts.release ();
3845 known_aggs.release ();
3846 return size;
3850 /* Estimate the growth of the caller when inlining EDGE.
3851 Only to be called via estimate_edge_size. */
3853 inline_hints
3854 do_estimate_edge_hints (struct cgraph_edge *edge)
3856 inline_hints hints;
3857 struct cgraph_node *callee;
3858 clause_t clause;
3859 vec<tree> known_vals;
3860 vec<ipa_polymorphic_call_context> known_contexts;
3861 vec<ipa_agg_jump_function_p> known_aggs;
3863 /* When we do caching, use do_estimate_edge_time to populate the entry. */
3865 if (edge_growth_cache.exists ())
3867 do_estimate_edge_time (edge);
3868 hints = edge_growth_cache[edge->uid].hints;
3869 gcc_checking_assert (hints);
3870 return hints - 1;
3873 callee = edge->callee->ultimate_alias_target ();
3875 /* Early inliner runs without caching, go ahead and do the dirty work. */
3876 gcc_checking_assert (edge->inline_failed);
3877 evaluate_properties_for_edge (edge, true,
3878 &clause, &known_vals, &known_contexts,
3879 &known_aggs);
3880 estimate_node_size_and_time (callee, clause, known_vals, known_contexts,
3881 known_aggs, NULL, NULL, NULL, &hints, vNULL);
3882 known_vals.release ();
3883 known_contexts.release ();
3884 known_aggs.release ();
3885 hints |= simple_edge_hints (edge);
3886 return hints;
3890 /* Estimate self time of the function NODE after inlining EDGE. */
3893 estimate_time_after_inlining (struct cgraph_node *node,
3894 struct cgraph_edge *edge)
3896 struct inline_edge_summary *es = inline_edge_summary (edge);
3897 if (!es->predicate || !false_predicate_p (es->predicate))
3899 gcov_type time =
3900 inline_summaries->get (node)->time + estimate_edge_time (edge);
3901 if (time < 0)
3902 time = 0;
3903 if (time > MAX_TIME)
3904 time = MAX_TIME;
3905 return time;
3907 return inline_summaries->get (node)->time;
3911 /* Estimate the size of NODE after inlining EDGE which should be an
3912 edge to either NODE or a call inlined into NODE. */
3915 estimate_size_after_inlining (struct cgraph_node *node,
3916 struct cgraph_edge *edge)
3918 struct inline_edge_summary *es = inline_edge_summary (edge);
3919 if (!es->predicate || !false_predicate_p (es->predicate))
3921 int size = inline_summaries->get (node)->size + estimate_edge_growth (edge);
3922 gcc_assert (size >= 0);
3923 return size;
3925 return inline_summaries->get (node)->size;
3929 struct growth_data
3931 struct cgraph_node *node;
3932 bool self_recursive;
3933 bool uninlinable;
3934 int growth;
3938 /* Worker for do_estimate_growth. Collect growth for all callers. */
3940 static bool
3941 do_estimate_growth_1 (struct cgraph_node *node, void *data)
3943 struct cgraph_edge *e;
3944 struct growth_data *d = (struct growth_data *) data;
3946 for (e = node->callers; e; e = e->next_caller)
3948 gcc_checking_assert (e->inline_failed);
3950 if (cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR)
3952 d->uninlinable = true;
3953 continue;
3956 if (e->recursive_p ())
3958 d->self_recursive = true;
3959 continue;
3961 d->growth += estimate_edge_growth (e);
3963 return false;
3967 /* Estimate the growth caused by inlining NODE into all callees. */
3970 estimate_growth (struct cgraph_node *node)
3972 struct growth_data d = { node, false, false, 0 };
3973 struct inline_summary *info = inline_summaries->get (node);
3975 node->call_for_symbol_and_aliases (do_estimate_growth_1, &d, true);
3977 /* For self recursive functions the growth estimation really should be
3978 infinity. We don't want to return very large values because the growth
3979 plays various roles in badness computation fractions. Be sure to not
3980 return zero or negative growths. */
3981 if (d.self_recursive)
3982 d.growth = d.growth < info->size ? info->size : d.growth;
3983 else if (DECL_EXTERNAL (node->decl) || d.uninlinable)
3985 else
3987 if (node->will_be_removed_from_program_if_no_direct_calls_p ())
3988 d.growth -= info->size;
3989 /* COMDAT functions are very often not shared across multiple units
3990 since they come from various template instantiations.
3991 Take this into account. */
3992 else if (DECL_COMDAT (node->decl)
3993 && node->can_remove_if_no_direct_calls_p ())
3994 d.growth -= (info->size
3995 * (100 - PARAM_VALUE (PARAM_COMDAT_SHARING_PROBABILITY))
3996 + 50) / 100;
3999 return d.growth;
4002 /* Verify if there are fewer than MAX_CALLERS. */
4004 static bool
4005 check_callers (cgraph_node *node, int *max_callers)
4007 ipa_ref *ref;
4009 if (!node->can_remove_if_no_direct_calls_and_refs_p ())
4010 return true;
4012 for (cgraph_edge *e = node->callers; e; e = e->next_caller)
4014 (*max_callers)--;
4015 if (!*max_callers
4016 || cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR)
4017 return true;
4020 FOR_EACH_ALIAS (node, ref)
4021 if (check_callers (dyn_cast <cgraph_node *> (ref->referring), max_callers))
4022 return true;
4024 return false;
4028 /* Make cheap estimation if growth of NODE is likely positive knowing
4029 EDGE_GROWTH of one particular edge.
4030 We assume that most of other edges will have similar growth
4031 and skip computation if there are too many callers. */
4033 bool
4034 growth_likely_positive (struct cgraph_node *node,
4035 int edge_growth)
4037 int max_callers;
4038 struct cgraph_edge *e;
4039 gcc_checking_assert (edge_growth > 0);
4041 /* First quickly check if NODE is removable at all. */
4042 if (DECL_EXTERNAL (node->decl))
4043 return true;
4044 if (!node->can_remove_if_no_direct_calls_and_refs_p ()
4045 || node->address_taken)
4046 return true;
4048 max_callers = inline_summaries->get (node)->size * 4 / edge_growth + 2;
4050 for (e = node->callers; e; e = e->next_caller)
4052 max_callers--;
4053 if (!max_callers
4054 || cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR)
4055 return true;
4058 ipa_ref *ref;
4059 FOR_EACH_ALIAS (node, ref)
4060 if (check_callers (dyn_cast <cgraph_node *> (ref->referring), &max_callers))
4061 return true;
4063 /* Unlike for functions called once, we play unsafe with
4064 COMDATs. We can allow that since we know functions
4065 in consideration are small (and thus risk is small) and
4066 moreover grow estimates already accounts that COMDAT
4067 functions may or may not disappear when eliminated from
4068 current unit. With good probability making aggressive
4069 choice in all units is going to make overall program
4070 smaller. */
4071 if (DECL_COMDAT (node->decl))
4073 if (!node->can_remove_if_no_direct_calls_p ())
4074 return true;
4076 else if (!node->will_be_removed_from_program_if_no_direct_calls_p ())
4077 return true;
4079 return estimate_growth (node) > 0;
4083 /* This function performs intraprocedural analysis in NODE that is required to
4084 inline indirect calls. */
4086 static void
4087 inline_indirect_intraprocedural_analysis (struct cgraph_node *node)
4089 ipa_analyze_node (node);
4090 if (dump_file && (dump_flags & TDF_DETAILS))
4092 ipa_print_node_params (dump_file, node);
4093 ipa_print_node_jump_functions (dump_file, node);
4098 /* Note function body size. */
4100 void
4101 inline_analyze_function (struct cgraph_node *node)
4103 push_cfun (DECL_STRUCT_FUNCTION (node->decl));
4105 if (dump_file)
4106 fprintf (dump_file, "\nAnalyzing function: %s/%u\n",
4107 node->name (), node->order);
4108 if (opt_for_fn (node->decl, optimize) && !node->thunk.thunk_p)
4109 inline_indirect_intraprocedural_analysis (node);
4110 compute_inline_parameters (node, false);
4111 if (!optimize)
4113 struct cgraph_edge *e;
4114 for (e = node->callees; e; e = e->next_callee)
4115 e->inline_failed = CIF_FUNCTION_NOT_OPTIMIZED;
4116 for (e = node->indirect_calls; e; e = e->next_callee)
4117 e->inline_failed = CIF_FUNCTION_NOT_OPTIMIZED;
4120 pop_cfun ();
4124 /* Called when new function is inserted to callgraph late. */
4126 void
4127 inline_summary_t::insert (struct cgraph_node *node, inline_summary *)
4129 inline_analyze_function (node);
4132 /* Note function body size. */
4134 void
4135 inline_generate_summary (void)
4137 struct cgraph_node *node;
4139 FOR_EACH_DEFINED_FUNCTION (node)
4140 if (DECL_STRUCT_FUNCTION (node->decl))
4141 node->local.versionable = tree_versionable_function_p (node->decl);
4143 /* When not optimizing, do not bother to analyze. Inlining is still done
4144 because edge redirection needs to happen there. */
4145 if (!optimize && !flag_generate_lto && !flag_generate_offload && !flag_wpa)
4146 return;
4148 if (!inline_summaries)
4149 inline_summaries = (inline_summary_t*) inline_summary_t::create_ggc (symtab);
4151 inline_summaries->enable_insertion_hook ();
4153 ipa_register_cgraph_hooks ();
4154 inline_free_summary ();
4156 FOR_EACH_DEFINED_FUNCTION (node)
4157 if (!node->alias)
4158 inline_analyze_function (node);
4162 /* Read predicate from IB. */
4164 static struct predicate
4165 read_predicate (struct lto_input_block *ib)
4167 struct predicate out;
4168 clause_t clause;
4169 int k = 0;
4173 gcc_assert (k <= MAX_CLAUSES);
4174 clause = out.clause[k++] = streamer_read_uhwi (ib);
4176 while (clause);
4178 /* Zero-initialize the remaining clauses in OUT. */
4179 while (k <= MAX_CLAUSES)
4180 out.clause[k++] = 0;
4182 return out;
4186 /* Write inline summary for edge E to OB. */
4188 static void
4189 read_inline_edge_summary (struct lto_input_block *ib, struct cgraph_edge *e)
4191 struct inline_edge_summary *es = inline_edge_summary (e);
4192 struct predicate p;
4193 int length, i;
4195 es->call_stmt_size = streamer_read_uhwi (ib);
4196 es->call_stmt_time = streamer_read_uhwi (ib);
4197 es->loop_depth = streamer_read_uhwi (ib);
4198 p = read_predicate (ib);
4199 edge_set_predicate (e, &p);
4200 length = streamer_read_uhwi (ib);
4201 if (length)
4203 es->param.safe_grow_cleared (length);
4204 for (i = 0; i < length; i++)
4205 es->param[i].change_prob = streamer_read_uhwi (ib);
4210 /* Stream in inline summaries from the section. */
4212 static void
4213 inline_read_section (struct lto_file_decl_data *file_data, const char *data,
4214 size_t len)
4216 const struct lto_function_header *header =
4217 (const struct lto_function_header *) data;
4218 const int cfg_offset = sizeof (struct lto_function_header);
4219 const int main_offset = cfg_offset + header->cfg_size;
4220 const int string_offset = main_offset + header->main_size;
4221 struct data_in *data_in;
4222 unsigned int i, count2, j;
4223 unsigned int f_count;
4225 lto_input_block ib ((const char *) data + main_offset, header->main_size,
4226 file_data->mode_table);
4228 data_in =
4229 lto_data_in_create (file_data, (const char *) data + string_offset,
4230 header->string_size, vNULL);
4231 f_count = streamer_read_uhwi (&ib);
4232 for (i = 0; i < f_count; i++)
4234 unsigned int index;
4235 struct cgraph_node *node;
4236 struct inline_summary *info;
4237 lto_symtab_encoder_t encoder;
4238 struct bitpack_d bp;
4239 struct cgraph_edge *e;
4240 predicate p;
4242 index = streamer_read_uhwi (&ib);
4243 encoder = file_data->symtab_node_encoder;
4244 node = dyn_cast<cgraph_node *> (lto_symtab_encoder_deref (encoder,
4245 index));
4246 info = inline_summaries->get (node);
4248 info->estimated_stack_size
4249 = info->estimated_self_stack_size = streamer_read_uhwi (&ib);
4250 info->size = info->self_size = streamer_read_uhwi (&ib);
4251 info->time = info->self_time = streamer_read_uhwi (&ib);
4253 bp = streamer_read_bitpack (&ib);
4254 info->inlinable = bp_unpack_value (&bp, 1);
4255 info->contains_cilk_spawn = bp_unpack_value (&bp, 1);
4256 info->fp_expressions = bp_unpack_value (&bp, 1);
4258 count2 = streamer_read_uhwi (&ib);
4259 gcc_assert (!info->conds);
4260 for (j = 0; j < count2; j++)
4262 struct condition c;
4263 c.operand_num = streamer_read_uhwi (&ib);
4264 c.code = (enum tree_code) streamer_read_uhwi (&ib);
4265 c.val = stream_read_tree (&ib, data_in);
4266 bp = streamer_read_bitpack (&ib);
4267 c.agg_contents = bp_unpack_value (&bp, 1);
4268 c.by_ref = bp_unpack_value (&bp, 1);
4269 if (c.agg_contents)
4270 c.offset = streamer_read_uhwi (&ib);
4271 vec_safe_push (info->conds, c);
4273 count2 = streamer_read_uhwi (&ib);
4274 gcc_assert (!info->entry);
4275 for (j = 0; j < count2; j++)
4277 struct size_time_entry e;
4279 e.size = streamer_read_uhwi (&ib);
4280 e.time = streamer_read_uhwi (&ib);
4281 e.predicate = read_predicate (&ib);
4283 vec_safe_push (info->entry, e);
4286 p = read_predicate (&ib);
4287 set_hint_predicate (&info->loop_iterations, p);
4288 p = read_predicate (&ib);
4289 set_hint_predicate (&info->loop_stride, p);
4290 p = read_predicate (&ib);
4291 set_hint_predicate (&info->array_index, p);
4292 for (e = node->callees; e; e = e->next_callee)
4293 read_inline_edge_summary (&ib, e);
4294 for (e = node->indirect_calls; e; e = e->next_callee)
4295 read_inline_edge_summary (&ib, e);
4298 lto_free_section_data (file_data, LTO_section_inline_summary, NULL, data,
4299 len);
4300 lto_data_in_delete (data_in);
4304 /* Read inline summary. Jump functions are shared among ipa-cp
4305 and inliner, so when ipa-cp is active, we don't need to write them
4306 twice. */
4308 void
4309 inline_read_summary (void)
4311 struct lto_file_decl_data **file_data_vec = lto_get_file_decl_data ();
4312 struct lto_file_decl_data *file_data;
4313 unsigned int j = 0;
4315 inline_summary_alloc ();
4317 while ((file_data = file_data_vec[j++]))
4319 size_t len;
4320 const char *data = lto_get_section_data (file_data,
4321 LTO_section_inline_summary,
4322 NULL, &len);
4323 if (data)
4324 inline_read_section (file_data, data, len);
4325 else
4326 /* Fatal error here. We do not want to support compiling ltrans units
4327 with different version of compiler or different flags than the WPA
4328 unit, so this should never happen. */
4329 fatal_error (input_location,
4330 "ipa inline summary is missing in input file");
4332 if (optimize)
4334 ipa_register_cgraph_hooks ();
4335 if (!flag_ipa_cp)
4336 ipa_prop_read_jump_functions ();
4339 gcc_assert (inline_summaries);
4340 inline_summaries->enable_insertion_hook ();
4344 /* Write predicate P to OB. */
4346 static void
4347 write_predicate (struct output_block *ob, struct predicate *p)
4349 int j;
4350 if (p)
4351 for (j = 0; p->clause[j]; j++)
4353 gcc_assert (j < MAX_CLAUSES);
4354 streamer_write_uhwi (ob, p->clause[j]);
4356 streamer_write_uhwi (ob, 0);
4360 /* Write inline summary for edge E to OB. */
4362 static void
4363 write_inline_edge_summary (struct output_block *ob, struct cgraph_edge *e)
4365 struct inline_edge_summary *es = inline_edge_summary (e);
4366 int i;
4368 streamer_write_uhwi (ob, es->call_stmt_size);
4369 streamer_write_uhwi (ob, es->call_stmt_time);
4370 streamer_write_uhwi (ob, es->loop_depth);
4371 write_predicate (ob, es->predicate);
4372 streamer_write_uhwi (ob, es->param.length ());
4373 for (i = 0; i < (int) es->param.length (); i++)
4374 streamer_write_uhwi (ob, es->param[i].change_prob);
4378 /* Write inline summary for node in SET.
4379 Jump functions are shared among ipa-cp and inliner, so when ipa-cp is
4380 active, we don't need to write them twice. */
4382 void
4383 inline_write_summary (void)
4385 struct cgraph_node *node;
4386 struct output_block *ob = create_output_block (LTO_section_inline_summary);
4387 lto_symtab_encoder_t encoder = ob->decl_state->symtab_node_encoder;
4388 unsigned int count = 0;
4389 int i;
4391 for (i = 0; i < lto_symtab_encoder_size (encoder); i++)
4393 symtab_node *snode = lto_symtab_encoder_deref (encoder, i);
4394 cgraph_node *cnode = dyn_cast <cgraph_node *> (snode);
4395 if (cnode && cnode->definition && !cnode->alias)
4396 count++;
4398 streamer_write_uhwi (ob, count);
4400 for (i = 0; i < lto_symtab_encoder_size (encoder); i++)
4402 symtab_node *snode = lto_symtab_encoder_deref (encoder, i);
4403 cgraph_node *cnode = dyn_cast <cgraph_node *> (snode);
4404 if (cnode && (node = cnode)->definition && !node->alias)
4406 struct inline_summary *info = inline_summaries->get (node);
4407 struct bitpack_d bp;
4408 struct cgraph_edge *edge;
4409 int i;
4410 size_time_entry *e;
4411 struct condition *c;
4413 streamer_write_uhwi (ob,
4414 lto_symtab_encoder_encode (encoder,
4416 node));
4417 streamer_write_hwi (ob, info->estimated_self_stack_size);
4418 streamer_write_hwi (ob, info->self_size);
4419 streamer_write_hwi (ob, info->self_time);
4420 bp = bitpack_create (ob->main_stream);
4421 bp_pack_value (&bp, info->inlinable, 1);
4422 bp_pack_value (&bp, info->contains_cilk_spawn, 1);
4423 bp_pack_value (&bp, info->fp_expressions, 1);
4424 streamer_write_bitpack (&bp);
4425 streamer_write_uhwi (ob, vec_safe_length (info->conds));
4426 for (i = 0; vec_safe_iterate (info->conds, i, &c); i++)
4428 streamer_write_uhwi (ob, c->operand_num);
4429 streamer_write_uhwi (ob, c->code);
4430 stream_write_tree (ob, c->val, true);
4431 bp = bitpack_create (ob->main_stream);
4432 bp_pack_value (&bp, c->agg_contents, 1);
4433 bp_pack_value (&bp, c->by_ref, 1);
4434 streamer_write_bitpack (&bp);
4435 if (c->agg_contents)
4436 streamer_write_uhwi (ob, c->offset);
4438 streamer_write_uhwi (ob, vec_safe_length (info->entry));
4439 for (i = 0; vec_safe_iterate (info->entry, i, &e); i++)
4441 streamer_write_uhwi (ob, e->size);
4442 streamer_write_uhwi (ob, e->time);
4443 write_predicate (ob, &e->predicate);
4445 write_predicate (ob, info->loop_iterations);
4446 write_predicate (ob, info->loop_stride);
4447 write_predicate (ob, info->array_index);
4448 for (edge = node->callees; edge; edge = edge->next_callee)
4449 write_inline_edge_summary (ob, edge);
4450 for (edge = node->indirect_calls; edge; edge = edge->next_callee)
4451 write_inline_edge_summary (ob, edge);
4454 streamer_write_char_stream (ob->main_stream, 0);
4455 produce_asm (ob, NULL);
4456 destroy_output_block (ob);
4458 if (optimize && !flag_ipa_cp)
4459 ipa_prop_write_jump_functions ();
4463 /* Release inline summary. */
4465 void
4466 inline_free_summary (void)
4468 struct cgraph_node *node;
4469 if (edge_removal_hook_holder)
4470 symtab->remove_edge_removal_hook (edge_removal_hook_holder);
4471 edge_removal_hook_holder = NULL;
4472 if (edge_duplication_hook_holder)
4473 symtab->remove_edge_duplication_hook (edge_duplication_hook_holder);
4474 edge_duplication_hook_holder = NULL;
4475 if (!inline_edge_summary_vec.exists ())
4476 return;
4477 FOR_EACH_DEFINED_FUNCTION (node)
4478 if (!node->alias)
4479 reset_inline_summary (node, inline_summaries->get (node));
4480 inline_summaries->release ();
4481 inline_summaries = NULL;
4482 inline_edge_summary_vec.release ();
4483 edge_predicate_pool.release ();