1 /* Inlining decision heuristics.
2 Copyright (C) 2003-2017 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
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
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
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
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. */
69 #include "coretypes.h"
73 #include "alloc-pool.h"
74 #include "tree-pass.h"
76 #include "tree-streamer.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"
85 #include "gimple-iterator.h"
87 #include "tree-ssa-loop-niter.h"
88 #include "tree-ssa-loop.h"
89 #include "symbol-summary.h"
91 #include "ipa-inline.h"
93 #include "tree-scalar-evolution.h"
94 #include "ipa-utils.h"
96 #include "cfgexpand.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
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
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)
154 /* Return predicate testing single condition number COND. */
156 static inline struct predicate
157 single_cond_predicate (int cond
)
160 p
.clause
[0] = 1 << cond
;
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). */
178 true_predicate_p (struct predicate
*p
)
180 return !p
->clause
[0];
184 /* Return true if P is (false). */
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
);
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
212 struct agg_position_info
214 HOST_WIDE_INT offset
;
219 /* Add condition to condition list SUMMARY. OPERAND_NUM, SIZE, CODE and VAL
220 correspond to fields of condition structure. AGGPOS describes whether the
221 used operand is loaded from an aggregate and where in the aggregate it is.
222 It can be NULL, which means this not a load from an aggregate. */
224 static struct predicate
225 add_condition (struct inline_summary
*summary
, int operand_num
,
226 HOST_WIDE_INT size
, struct agg_position_info
*aggpos
,
227 enum tree_code code
, tree val
)
231 struct condition new_cond
;
232 HOST_WIDE_INT offset
;
233 bool agg_contents
, by_ref
;
237 offset
= aggpos
->offset
;
238 agg_contents
= aggpos
->agg_contents
;
239 by_ref
= aggpos
->by_ref
;
244 agg_contents
= false;
248 gcc_checking_assert (operand_num
>= 0);
249 for (i
= 0; vec_safe_iterate (summary
->conds
, i
, &c
); i
++)
251 if (c
->operand_num
== operand_num
255 && c
->agg_contents
== agg_contents
256 && (!agg_contents
|| (c
->offset
== offset
&& c
->by_ref
== by_ref
)))
257 return single_cond_predicate (i
+ predicate_first_dynamic_condition
);
259 /* Too many conditions. Give up and return constant true. */
260 if (i
== NUM_CONDITIONS
- predicate_first_dynamic_condition
)
261 return true_predicate ();
263 new_cond
.operand_num
= operand_num
;
264 new_cond
.code
= code
;
266 new_cond
.agg_contents
= agg_contents
;
267 new_cond
.by_ref
= by_ref
;
268 new_cond
.offset
= offset
;
269 new_cond
.size
= size
;
270 vec_safe_push (summary
->conds
, new_cond
);
271 return single_cond_predicate (i
+ predicate_first_dynamic_condition
);
275 /* Add clause CLAUSE into the predicate P. */
278 add_clause (conditions conditions
, struct predicate
*p
, clause_t clause
)
282 int insert_here
= -1;
289 /* False clause makes the whole predicate false. Kill the other variants. */
290 if (clause
== (1 << predicate_false_condition
))
292 p
->clause
[0] = (1 << predicate_false_condition
);
296 if (false_predicate_p (p
))
299 /* No one should be silly enough to add false into nontrivial clauses. */
300 gcc_checking_assert (!(clause
& (1 << predicate_false_condition
)));
302 /* Look where to insert the clause. At the same time prune out
303 clauses of P that are implied by the new clause and thus
305 for (i
= 0, i2
= 0; i
<= MAX_CLAUSES
; i
++)
307 p
->clause
[i2
] = p
->clause
[i
];
312 /* If p->clause[i] implies clause, there is nothing to add. */
313 if ((p
->clause
[i
] & clause
) == p
->clause
[i
])
315 /* We had nothing to add, none of clauses should've become
317 gcc_checking_assert (i
== i2
);
321 if (p
->clause
[i
] < clause
&& insert_here
< 0)
324 /* If clause implies p->clause[i], then p->clause[i] becomes redundant.
325 Otherwise the p->clause[i] has to stay. */
326 if ((p
->clause
[i
] & clause
) != clause
)
330 /* Look for clauses that are obviously true. I.e.
331 op0 == 5 || op0 != 5. */
332 for (c1
= predicate_first_dynamic_condition
; c1
< NUM_CONDITIONS
; c1
++)
335 if (!(clause
& (1 << c1
)))
337 cc1
= &(*conditions
)[c1
- predicate_first_dynamic_condition
];
338 /* We have no way to represent !CHANGED and !IS_NOT_CONSTANT
339 and thus there is no point for looking for them. */
340 if (cc1
->code
== CHANGED
|| cc1
->code
== IS_NOT_CONSTANT
)
342 for (c2
= c1
+ 1; c2
< NUM_CONDITIONS
; c2
++)
343 if (clause
& (1 << c2
))
346 &(*conditions
)[c1
- predicate_first_dynamic_condition
];
348 &(*conditions
)[c2
- predicate_first_dynamic_condition
];
349 if (cc1
->operand_num
== cc2
->operand_num
350 && cc1
->val
== cc2
->val
351 && cc2
->code
!= IS_NOT_CONSTANT
352 && cc2
->code
!= CHANGED
353 && cc1
->code
== invert_tree_comparison (cc2
->code
,
354 HONOR_NANS (cc1
->val
)))
360 /* We run out of variants. Be conservative in positive direction. */
361 if (i2
== MAX_CLAUSES
)
363 /* Keep clauses in decreasing order. This makes equivalence testing easy. */
364 p
->clause
[i2
+ 1] = 0;
365 if (insert_here
>= 0)
366 for (; i2
> insert_here
; i2
--)
367 p
->clause
[i2
] = p
->clause
[i2
- 1];
370 p
->clause
[insert_here
] = clause
;
376 static struct predicate
377 and_predicates (conditions conditions
,
378 struct predicate
*p
, struct predicate
*p2
)
380 struct predicate out
= *p
;
383 /* Avoid busy work. */
384 if (false_predicate_p (p2
) || true_predicate_p (p
))
386 if (false_predicate_p (p
) || true_predicate_p (p2
))
389 /* See how far predicates match. */
390 for (i
= 0; p
->clause
[i
] && p
->clause
[i
] == p2
->clause
[i
]; i
++)
392 gcc_checking_assert (i
< MAX_CLAUSES
);
395 /* Combine the predicates rest. */
396 for (; p2
->clause
[i
]; i
++)
398 gcc_checking_assert (i
< MAX_CLAUSES
);
399 add_clause (conditions
, &out
, p2
->clause
[i
]);
405 /* Return true if predicates are obviously equal. */
408 predicates_equal_p (struct predicate
*p
, struct predicate
*p2
)
411 for (i
= 0; p
->clause
[i
]; i
++)
413 gcc_checking_assert (i
< MAX_CLAUSES
);
414 gcc_checking_assert (p
->clause
[i
] > p
->clause
[i
+ 1]);
415 gcc_checking_assert (!p2
->clause
[i
]
416 || p2
->clause
[i
] > p2
->clause
[i
+ 1]);
417 if (p
->clause
[i
] != p2
->clause
[i
])
420 return !p2
->clause
[i
];
426 static struct predicate
427 or_predicates (conditions conditions
,
428 struct predicate
*p
, struct predicate
*p2
)
430 struct predicate out
= true_predicate ();
433 /* Avoid busy work. */
434 if (false_predicate_p (p2
) || true_predicate_p (p
))
436 if (false_predicate_p (p
) || true_predicate_p (p2
))
438 if (predicates_equal_p (p
, p2
))
441 /* OK, combine the predicates. */
442 for (i
= 0; p
->clause
[i
]; i
++)
443 for (j
= 0; p2
->clause
[j
]; j
++)
445 gcc_checking_assert (i
< MAX_CLAUSES
&& j
< MAX_CLAUSES
);
446 add_clause (conditions
, &out
, p
->clause
[i
] | p2
->clause
[j
]);
452 /* Having partial truth assignment in POSSIBLE_TRUTHS, return false
453 if predicate P is known to be false. */
456 evaluate_predicate (struct predicate
*p
, clause_t possible_truths
)
460 /* True remains true. */
461 if (true_predicate_p (p
))
464 gcc_assert (!(possible_truths
& (1 << predicate_false_condition
)));
466 /* See if we can find clause we can disprove. */
467 for (i
= 0; p
->clause
[i
]; i
++)
469 gcc_checking_assert (i
< MAX_CLAUSES
);
470 if (!(p
->clause
[i
] & possible_truths
))
476 /* Return the probability in range 0...REG_BR_PROB_BASE that the predicated
477 instruction will be recomputed per invocation of the inlined call. */
480 predicate_probability (conditions conds
,
481 struct predicate
*p
, clause_t possible_truths
,
482 vec
<inline_param_summary
> inline_param_summary
)
485 int combined_prob
= REG_BR_PROB_BASE
;
487 /* True remains true. */
488 if (true_predicate_p (p
))
489 return REG_BR_PROB_BASE
;
491 if (false_predicate_p (p
))
494 gcc_assert (!(possible_truths
& (1 << predicate_false_condition
)));
496 /* See if we can find clause we can disprove. */
497 for (i
= 0; p
->clause
[i
]; i
++)
499 gcc_checking_assert (i
< MAX_CLAUSES
);
500 if (!(p
->clause
[i
] & possible_truths
))
506 if (!inline_param_summary
.exists ())
507 return REG_BR_PROB_BASE
;
508 for (i2
= 0; i2
< NUM_CONDITIONS
; i2
++)
509 if ((p
->clause
[i
] & possible_truths
) & (1 << i2
))
511 if (i2
>= predicate_first_dynamic_condition
)
514 &(*conds
)[i2
- predicate_first_dynamic_condition
];
515 if (c
->code
== CHANGED
517 (int) inline_param_summary
.length ()))
520 inline_param_summary
[c
->operand_num
].change_prob
;
521 this_prob
= MAX (this_prob
, iprob
);
524 this_prob
= REG_BR_PROB_BASE
;
527 this_prob
= REG_BR_PROB_BASE
;
529 combined_prob
= MIN (this_prob
, combined_prob
);
534 return combined_prob
;
538 /* Dump conditional COND. */
541 dump_condition (FILE *f
, conditions conditions
, int cond
)
544 if (cond
== predicate_false_condition
)
545 fprintf (f
, "false");
546 else if (cond
== predicate_not_inlined_condition
)
547 fprintf (f
, "not inlined");
550 c
= &(*conditions
)[cond
- predicate_first_dynamic_condition
];
551 fprintf (f
, "op%i", c
->operand_num
);
553 fprintf (f
, "[%soffset: " HOST_WIDE_INT_PRINT_DEC
"]",
554 c
->by_ref
? "ref " : "", c
->offset
);
555 if (c
->code
== IS_NOT_CONSTANT
)
557 fprintf (f
, " not constant");
560 if (c
->code
== CHANGED
)
562 fprintf (f
, " changed");
565 fprintf (f
, " %s ", op_symbol_code (c
->code
));
566 print_generic_expr (f
, c
->val
, 1);
571 /* Dump clause CLAUSE. */
574 dump_clause (FILE *f
, conditions conds
, clause_t clause
)
581 for (i
= 0; i
< NUM_CONDITIONS
; i
++)
582 if (clause
& (1 << i
))
587 dump_condition (f
, conds
, i
);
593 /* Dump predicate PREDICATE. */
596 dump_predicate (FILE *f
, conditions conds
, struct predicate
*pred
)
599 if (true_predicate_p (pred
))
600 dump_clause (f
, conds
, 0);
602 for (i
= 0; pred
->clause
[i
]; i
++)
606 dump_clause (f
, conds
, pred
->clause
[i
]);
612 /* Dump inline hints. */
614 dump_inline_hints (FILE *f
, inline_hints hints
)
618 fprintf (f
, "inline hints:");
619 if (hints
& INLINE_HINT_indirect_call
)
621 hints
&= ~INLINE_HINT_indirect_call
;
622 fprintf (f
, " indirect_call");
624 if (hints
& INLINE_HINT_loop_iterations
)
626 hints
&= ~INLINE_HINT_loop_iterations
;
627 fprintf (f
, " loop_iterations");
629 if (hints
& INLINE_HINT_loop_stride
)
631 hints
&= ~INLINE_HINT_loop_stride
;
632 fprintf (f
, " loop_stride");
634 if (hints
& INLINE_HINT_same_scc
)
636 hints
&= ~INLINE_HINT_same_scc
;
637 fprintf (f
, " same_scc");
639 if (hints
& INLINE_HINT_in_scc
)
641 hints
&= ~INLINE_HINT_in_scc
;
642 fprintf (f
, " in_scc");
644 if (hints
& INLINE_HINT_cross_module
)
646 hints
&= ~INLINE_HINT_cross_module
;
647 fprintf (f
, " cross_module");
649 if (hints
& INLINE_HINT_declared_inline
)
651 hints
&= ~INLINE_HINT_declared_inline
;
652 fprintf (f
, " declared_inline");
654 if (hints
& INLINE_HINT_array_index
)
656 hints
&= ~INLINE_HINT_array_index
;
657 fprintf (f
, " array_index");
659 if (hints
& INLINE_HINT_known_hot
)
661 hints
&= ~INLINE_HINT_known_hot
;
662 fprintf (f
, " known_hot");
668 /* Record SIZE and TIME under condition PRED into the inline summary. */
671 account_size_time (struct inline_summary
*summary
, int size
, int time
,
672 struct predicate
*pred
)
678 if (false_predicate_p (pred
))
681 /* We need to create initial empty unconitional clause, but otherwie
682 we don't need to account empty times and sizes. */
683 if (!size
&& !time
&& summary
->entry
)
686 /* Watch overflow that might result from insane profiles. */
687 if (time
> MAX_TIME
* INLINE_TIME_SCALE
)
688 time
= MAX_TIME
* INLINE_TIME_SCALE
;
689 gcc_assert (time
>= 0);
691 for (i
= 0; vec_safe_iterate (summary
->entry
, i
, &e
); i
++)
692 if (predicates_equal_p (&e
->predicate
, pred
))
701 e
= &(*summary
->entry
)[0];
702 gcc_assert (!e
->predicate
.clause
[0]);
703 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
705 "\t\tReached limit on number of entries, "
706 "ignoring the predicate.");
708 if (dump_file
&& (dump_flags
& TDF_DETAILS
) && (time
|| size
))
711 "\t\tAccounting size:%3.2f, time:%3.2f on %spredicate:",
712 ((double) size
) / INLINE_SIZE_SCALE
,
713 ((double) time
) / INLINE_TIME_SCALE
, found
? "" : "new ");
714 dump_predicate (dump_file
, summary
->conds
, pred
);
718 struct size_time_entry new_entry
;
719 new_entry
.size
= size
;
720 new_entry
.time
= time
;
721 new_entry
.predicate
= *pred
;
722 vec_safe_push (summary
->entry
, new_entry
);
728 if (e
->time
> MAX_TIME
* INLINE_TIME_SCALE
)
729 e
->time
= MAX_TIME
* INLINE_TIME_SCALE
;
733 /* We proved E to be unreachable, redirect it to __bultin_unreachable. */
735 static struct cgraph_edge
*
736 redirect_to_unreachable (struct cgraph_edge
*e
)
738 struct cgraph_node
*callee
= !e
->inline_failed
? e
->callee
: NULL
;
739 struct cgraph_node
*target
= cgraph_node::get_create
740 (builtin_decl_implicit (BUILT_IN_UNREACHABLE
));
743 e
= e
->resolve_speculation (target
->decl
);
745 e
->make_direct (target
);
747 e
->redirect_callee (target
);
748 struct inline_edge_summary
*es
= inline_edge_summary (e
);
749 e
->inline_failed
= CIF_UNREACHABLE
;
752 es
->call_stmt_size
= 0;
753 es
->call_stmt_time
= 0;
755 callee
->remove_symbol_and_inline_clones ();
759 /* Set predicate for edge E. */
762 edge_set_predicate (struct cgraph_edge
*e
, struct predicate
*predicate
)
764 /* If the edge is determined to be never executed, redirect it
765 to BUILTIN_UNREACHABLE to save inliner from inlining into it. */
766 if (predicate
&& false_predicate_p (predicate
)
767 /* When handling speculative edges, we need to do the redirection
768 just once. Do it always on the direct edge, so we do not
769 attempt to resolve speculation while duplicating the edge. */
770 && (!e
->speculative
|| e
->callee
))
771 e
= redirect_to_unreachable (e
);
773 struct inline_edge_summary
*es
= inline_edge_summary (e
);
774 if (predicate
&& !true_predicate_p (predicate
))
777 es
->predicate
= edge_predicate_pool
.allocate ();
778 *es
->predicate
= *predicate
;
783 edge_predicate_pool
.remove (es
->predicate
);
784 es
->predicate
= NULL
;
788 /* Set predicate for hint *P. */
791 set_hint_predicate (struct predicate
**p
, struct predicate new_predicate
)
793 if (false_predicate_p (&new_predicate
) || true_predicate_p (&new_predicate
))
796 edge_predicate_pool
.remove (*p
);
802 *p
= edge_predicate_pool
.allocate ();
808 /* KNOWN_VALS is partial mapping of parameters of NODE to constant values.
809 KNOWN_AGGS is a vector of aggreggate jump functions for each parameter.
810 Return clause of possible truths. When INLINE_P is true, assume that we are
813 ERROR_MARK means compile time invariant. */
816 evaluate_conditions_for_known_args (struct cgraph_node
*node
,
818 vec
<tree
> known_vals
,
819 vec
<ipa_agg_jump_function_p
>
822 clause_t clause
= inline_p
? 0 : 1 << predicate_not_inlined_condition
;
823 struct inline_summary
*info
= inline_summaries
->get (node
);
827 for (i
= 0; vec_safe_iterate (info
->conds
, i
, &c
); i
++)
832 /* We allow call stmt to have fewer arguments than the callee function
833 (especially for K&R style programs). So bound check here (we assume
834 known_aggs vector, if non-NULL, has the same length as
836 gcc_checking_assert (!known_aggs
.exists ()
837 || (known_vals
.length () == known_aggs
.length ()));
838 if (c
->operand_num
>= (int) known_vals
.length ())
840 clause
|= 1 << (i
+ predicate_first_dynamic_condition
);
846 struct ipa_agg_jump_function
*agg
;
848 if (c
->code
== CHANGED
850 && (known_vals
[c
->operand_num
] == error_mark_node
))
853 if (known_aggs
.exists ())
855 agg
= known_aggs
[c
->operand_num
];
856 val
= ipa_find_agg_cst_for_param (agg
, known_vals
[c
->operand_num
],
857 c
->offset
, c
->by_ref
);
864 val
= known_vals
[c
->operand_num
];
865 if (val
== error_mark_node
&& c
->code
!= CHANGED
)
871 clause
|= 1 << (i
+ predicate_first_dynamic_condition
);
874 if (c
->code
== CHANGED
)
877 if (tree_to_shwi (TYPE_SIZE (TREE_TYPE (val
))) != c
->size
)
879 clause
|= 1 << (i
+ predicate_first_dynamic_condition
);
882 if (c
->code
== IS_NOT_CONSTANT
)
885 val
= fold_unary (VIEW_CONVERT_EXPR
, TREE_TYPE (c
->val
), val
);
887 ? fold_binary_to_constant (c
->code
, boolean_type_node
, val
, c
->val
)
890 if (res
&& integer_zerop (res
))
893 clause
|= 1 << (i
+ predicate_first_dynamic_condition
);
899 /* Work out what conditions might be true at invocation of E. */
902 evaluate_properties_for_edge (struct cgraph_edge
*e
, bool inline_p
,
903 clause_t
*clause_ptr
,
904 vec
<tree
> *known_vals_ptr
,
905 vec
<ipa_polymorphic_call_context
>
907 vec
<ipa_agg_jump_function_p
> *known_aggs_ptr
)
909 struct cgraph_node
*callee
= e
->callee
->ultimate_alias_target ();
910 struct inline_summary
*info
= inline_summaries
->get (callee
);
911 vec
<tree
> known_vals
= vNULL
;
912 vec
<ipa_agg_jump_function_p
> known_aggs
= vNULL
;
915 *clause_ptr
= inline_p
? 0 : 1 << predicate_not_inlined_condition
;
917 known_vals_ptr
->create (0);
918 if (known_contexts_ptr
)
919 known_contexts_ptr
->create (0);
921 if (ipa_node_params_sum
922 && !e
->call_stmt_cannot_inline_p
923 && ((clause_ptr
&& info
->conds
) || known_vals_ptr
|| known_contexts_ptr
))
925 struct ipa_node_params
*parms_info
;
926 struct ipa_edge_args
*args
= IPA_EDGE_REF (e
);
927 struct inline_edge_summary
*es
= inline_edge_summary (e
);
928 int i
, count
= ipa_get_cs_argument_count (args
);
930 if (e
->caller
->global
.inlined_to
)
931 parms_info
= IPA_NODE_REF (e
->caller
->global
.inlined_to
);
933 parms_info
= IPA_NODE_REF (e
->caller
);
935 if (count
&& (info
->conds
|| known_vals_ptr
))
936 known_vals
.safe_grow_cleared (count
);
937 if (count
&& (info
->conds
|| known_aggs_ptr
))
938 known_aggs
.safe_grow_cleared (count
);
939 if (count
&& known_contexts_ptr
)
940 known_contexts_ptr
->safe_grow_cleared (count
);
942 for (i
= 0; i
< count
; i
++)
944 struct ipa_jump_func
*jf
= ipa_get_ith_jump_func (args
, i
);
945 tree cst
= ipa_value_from_jfunc (parms_info
, jf
);
947 if (!cst
&& e
->call_stmt
948 && i
< (int)gimple_call_num_args (e
->call_stmt
))
950 cst
= gimple_call_arg (e
->call_stmt
, i
);
951 if (!is_gimple_min_invariant (cst
))
956 gcc_checking_assert (TREE_CODE (cst
) != TREE_BINFO
);
957 if (known_vals
.exists ())
960 else if (inline_p
&& !es
->param
[i
].change_prob
)
961 known_vals
[i
] = error_mark_node
;
963 if (known_contexts_ptr
)
964 (*known_contexts_ptr
)[i
] = ipa_context_from_jfunc (parms_info
, e
,
966 /* TODO: When IPA-CP starts propagating and merging aggregate jump
967 functions, use its knowledge of the caller too, just like the
968 scalar case above. */
969 known_aggs
[i
] = &jf
->agg
;
972 else if (e
->call_stmt
&& !e
->call_stmt_cannot_inline_p
973 && ((clause_ptr
&& info
->conds
) || known_vals_ptr
))
975 int i
, count
= (int)gimple_call_num_args (e
->call_stmt
);
977 if (count
&& (info
->conds
|| known_vals_ptr
))
978 known_vals
.safe_grow_cleared (count
);
979 for (i
= 0; i
< count
; i
++)
981 tree cst
= gimple_call_arg (e
->call_stmt
, i
);
982 if (!is_gimple_min_invariant (cst
))
990 *clause_ptr
= evaluate_conditions_for_known_args (callee
, inline_p
,
991 known_vals
, known_aggs
);
994 *known_vals_ptr
= known_vals
;
996 known_vals
.release ();
999 *known_aggs_ptr
= known_aggs
;
1001 known_aggs
.release ();
1005 /* Allocate the inline summary vector or resize it to cover all cgraph nodes. */
1008 inline_summary_alloc (void)
1010 if (!edge_removal_hook_holder
)
1011 edge_removal_hook_holder
=
1012 symtab
->add_edge_removal_hook (&inline_edge_removal_hook
, NULL
);
1013 if (!edge_duplication_hook_holder
)
1014 edge_duplication_hook_holder
=
1015 symtab
->add_edge_duplication_hook (&inline_edge_duplication_hook
, NULL
);
1017 if (!inline_summaries
)
1018 inline_summaries
= (inline_summary_t
*) inline_summary_t::create_ggc (symtab
);
1020 if (inline_edge_summary_vec
.length () <= (unsigned) symtab
->edges_max_uid
)
1021 inline_edge_summary_vec
.safe_grow_cleared (symtab
->edges_max_uid
+ 1);
1024 /* We are called multiple time for given function; clear
1025 data from previous run so they are not cumulated. */
1028 reset_inline_edge_summary (struct cgraph_edge
*e
)
1030 if (e
->uid
< (int) inline_edge_summary_vec
.length ())
1032 struct inline_edge_summary
*es
= inline_edge_summary (e
);
1034 es
->call_stmt_size
= es
->call_stmt_time
= 0;
1036 edge_predicate_pool
.remove (es
->predicate
);
1037 es
->predicate
= NULL
;
1038 es
->param
.release ();
1042 /* We are called multiple time for given function; clear
1043 data from previous run so they are not cumulated. */
1046 reset_inline_summary (struct cgraph_node
*node
,
1047 inline_summary
*info
)
1049 struct cgraph_edge
*e
;
1051 info
->self_size
= info
->self_time
= 0;
1052 info
->estimated_stack_size
= 0;
1053 info
->estimated_self_stack_size
= 0;
1054 info
->stack_frame_offset
= 0;
1059 if (info
->loop_iterations
)
1061 edge_predicate_pool
.remove (info
->loop_iterations
);
1062 info
->loop_iterations
= NULL
;
1064 if (info
->loop_stride
)
1066 edge_predicate_pool
.remove (info
->loop_stride
);
1067 info
->loop_stride
= NULL
;
1069 if (info
->array_index
)
1071 edge_predicate_pool
.remove (info
->array_index
);
1072 info
->array_index
= NULL
;
1074 vec_free (info
->conds
);
1075 vec_free (info
->entry
);
1076 for (e
= node
->callees
; e
; e
= e
->next_callee
)
1077 reset_inline_edge_summary (e
);
1078 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
1079 reset_inline_edge_summary (e
);
1080 info
->fp_expressions
= false;
1083 /* Hook that is called by cgraph.c when a node is removed. */
1086 inline_summary_t::remove (cgraph_node
*node
, inline_summary
*info
)
1088 reset_inline_summary (node
, info
);
1091 /* Remap predicate P of former function to be predicate of duplicated function.
1092 POSSIBLE_TRUTHS is clause of possible truths in the duplicated node,
1093 INFO is inline summary of the duplicated node. */
1095 static struct predicate
1096 remap_predicate_after_duplication (struct predicate
*p
,
1097 clause_t possible_truths
,
1098 struct inline_summary
*info
)
1100 struct predicate new_predicate
= true_predicate ();
1102 for (j
= 0; p
->clause
[j
]; j
++)
1103 if (!(possible_truths
& p
->clause
[j
]))
1105 new_predicate
= false_predicate ();
1109 add_clause (info
->conds
, &new_predicate
,
1110 possible_truths
& p
->clause
[j
]);
1111 return new_predicate
;
1114 /* Same as remap_predicate_after_duplication but handle hint predicate *P.
1115 Additionally care about allocating new memory slot for updated predicate
1116 and set it to NULL when it becomes true or false (and thus uninteresting).
1120 remap_hint_predicate_after_duplication (struct predicate
**p
,
1121 clause_t possible_truths
,
1122 struct inline_summary
*info
)
1124 struct predicate new_predicate
;
1129 new_predicate
= remap_predicate_after_duplication (*p
,
1130 possible_truths
, info
);
1131 /* We do not want to free previous predicate; it is used by node origin. */
1133 set_hint_predicate (p
, new_predicate
);
1137 /* Hook that is called by cgraph.c when a node is duplicated. */
1139 inline_summary_t::duplicate (cgraph_node
*src
,
1142 inline_summary
*info
)
1144 inline_summary_alloc ();
1145 memcpy (info
, inline_summaries
->get (src
), sizeof (inline_summary
));
1146 /* TODO: as an optimization, we may avoid copying conditions
1147 that are known to be false or true. */
1148 info
->conds
= vec_safe_copy (info
->conds
);
1150 /* When there are any replacements in the function body, see if we can figure
1151 out that something was optimized out. */
1152 if (ipa_node_params_sum
&& dst
->clone
.tree_map
)
1154 vec
<size_time_entry
, va_gc
> *entry
= info
->entry
;
1155 /* Use SRC parm info since it may not be copied yet. */
1156 struct ipa_node_params
*parms_info
= IPA_NODE_REF (src
);
1157 vec
<tree
> known_vals
= vNULL
;
1158 int count
= ipa_get_param_count (parms_info
);
1160 clause_t possible_truths
;
1161 struct predicate true_pred
= true_predicate ();
1163 int optimized_out_size
= 0;
1164 bool inlined_to_p
= false;
1165 struct cgraph_edge
*edge
, *next
;
1168 known_vals
.safe_grow_cleared (count
);
1169 for (i
= 0; i
< count
; i
++)
1171 struct ipa_replace_map
*r
;
1173 for (j
= 0; vec_safe_iterate (dst
->clone
.tree_map
, j
, &r
); j
++)
1175 if (((!r
->old_tree
&& r
->parm_num
== i
)
1176 || (r
->old_tree
&& r
->old_tree
== ipa_get_param (parms_info
, i
)))
1177 && r
->replace_p
&& !r
->ref_p
)
1179 known_vals
[i
] = r
->new_tree
;
1184 possible_truths
= evaluate_conditions_for_known_args (dst
, false,
1187 known_vals
.release ();
1189 account_size_time (info
, 0, 0, &true_pred
);
1191 /* Remap size_time vectors.
1192 Simplify the predicate by prunning out alternatives that are known
1194 TODO: as on optimization, we can also eliminate conditions known
1196 for (i
= 0; vec_safe_iterate (entry
, i
, &e
); i
++)
1198 struct predicate new_predicate
;
1199 new_predicate
= remap_predicate_after_duplication (&e
->predicate
,
1202 if (false_predicate_p (&new_predicate
))
1203 optimized_out_size
+= e
->size
;
1205 account_size_time (info
, e
->size
, e
->time
, &new_predicate
);
1208 /* Remap edge predicates with the same simplification as above.
1209 Also copy constantness arrays. */
1210 for (edge
= dst
->callees
; edge
; edge
= next
)
1212 struct predicate new_predicate
;
1213 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
1214 next
= edge
->next_callee
;
1216 if (!edge
->inline_failed
)
1217 inlined_to_p
= true;
1220 new_predicate
= remap_predicate_after_duplication (es
->predicate
,
1223 if (false_predicate_p (&new_predicate
)
1224 && !false_predicate_p (es
->predicate
))
1225 optimized_out_size
+= es
->call_stmt_size
* INLINE_SIZE_SCALE
;
1226 edge_set_predicate (edge
, &new_predicate
);
1229 /* Remap indirect edge predicates with the same simplificaiton as above.
1230 Also copy constantness arrays. */
1231 for (edge
= dst
->indirect_calls
; edge
; edge
= next
)
1233 struct predicate new_predicate
;
1234 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
1235 next
= edge
->next_callee
;
1237 gcc_checking_assert (edge
->inline_failed
);
1240 new_predicate
= remap_predicate_after_duplication (es
->predicate
,
1243 if (false_predicate_p (&new_predicate
)
1244 && !false_predicate_p (es
->predicate
))
1245 optimized_out_size
+= es
->call_stmt_size
* INLINE_SIZE_SCALE
;
1246 edge_set_predicate (edge
, &new_predicate
);
1248 remap_hint_predicate_after_duplication (&info
->loop_iterations
,
1249 possible_truths
, info
);
1250 remap_hint_predicate_after_duplication (&info
->loop_stride
,
1251 possible_truths
, info
);
1252 remap_hint_predicate_after_duplication (&info
->array_index
,
1253 possible_truths
, info
);
1255 /* If inliner or someone after inliner will ever start producing
1256 non-trivial clones, we will get trouble with lack of information
1257 about updating self sizes, because size vectors already contains
1258 sizes of the calees. */
1259 gcc_assert (!inlined_to_p
|| !optimized_out_size
);
1263 info
->entry
= vec_safe_copy (info
->entry
);
1264 if (info
->loop_iterations
)
1266 predicate p
= *info
->loop_iterations
;
1267 info
->loop_iterations
= NULL
;
1268 set_hint_predicate (&info
->loop_iterations
, p
);
1270 if (info
->loop_stride
)
1272 predicate p
= *info
->loop_stride
;
1273 info
->loop_stride
= NULL
;
1274 set_hint_predicate (&info
->loop_stride
, p
);
1276 if (info
->array_index
)
1278 predicate p
= *info
->array_index
;
1279 info
->array_index
= NULL
;
1280 set_hint_predicate (&info
->array_index
, p
);
1283 if (!dst
->global
.inlined_to
)
1284 inline_update_overall_summary (dst
);
1288 /* Hook that is called by cgraph.c when a node is duplicated. */
1291 inline_edge_duplication_hook (struct cgraph_edge
*src
,
1292 struct cgraph_edge
*dst
,
1293 ATTRIBUTE_UNUSED
void *data
)
1295 struct inline_edge_summary
*info
;
1296 struct inline_edge_summary
*srcinfo
;
1297 inline_summary_alloc ();
1298 info
= inline_edge_summary (dst
);
1299 srcinfo
= inline_edge_summary (src
);
1300 memcpy (info
, srcinfo
, sizeof (struct inline_edge_summary
));
1301 info
->predicate
= NULL
;
1302 edge_set_predicate (dst
, srcinfo
->predicate
);
1303 info
->param
= srcinfo
->param
.copy ();
1304 if (!dst
->indirect_unknown_callee
&& src
->indirect_unknown_callee
)
1306 info
->call_stmt_size
-= (eni_size_weights
.indirect_call_cost
1307 - eni_size_weights
.call_cost
);
1308 info
->call_stmt_time
-= (eni_time_weights
.indirect_call_cost
1309 - eni_time_weights
.call_cost
);
1314 /* Keep edge cache consistent across edge removal. */
1317 inline_edge_removal_hook (struct cgraph_edge
*edge
,
1318 void *data ATTRIBUTE_UNUSED
)
1320 if (edge_growth_cache
.exists ())
1321 reset_edge_growth_cache (edge
);
1322 reset_inline_edge_summary (edge
);
1326 /* Initialize growth caches. */
1329 initialize_growth_caches (void)
1331 if (symtab
->edges_max_uid
)
1332 edge_growth_cache
.safe_grow_cleared (symtab
->edges_max_uid
);
1336 /* Free growth caches. */
1339 free_growth_caches (void)
1341 edge_growth_cache
.release ();
1345 /* Dump edge summaries associated to NODE and recursively to all clones.
1346 Indent by INDENT. */
1349 dump_inline_edge_summary (FILE *f
, int indent
, struct cgraph_node
*node
,
1350 struct inline_summary
*info
)
1352 struct cgraph_edge
*edge
;
1353 for (edge
= node
->callees
; edge
; edge
= edge
->next_callee
)
1355 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
1356 struct cgraph_node
*callee
= edge
->callee
->ultimate_alias_target ();
1360 "%*s%s/%i %s\n%*s loop depth:%2i freq:%4i size:%2i"
1361 " time: %2i callee size:%2i stack:%2i",
1362 indent
, "", callee
->name (), callee
->order
,
1363 !edge
->inline_failed
1364 ? "inlined" : cgraph_inline_failed_string (edge
-> inline_failed
),
1365 indent
, "", es
->loop_depth
, edge
->frequency
,
1366 es
->call_stmt_size
, es
->call_stmt_time
,
1367 (int) inline_summaries
->get (callee
)->size
/ INLINE_SIZE_SCALE
,
1368 (int) inline_summaries
->get (callee
)->estimated_stack_size
);
1372 fprintf (f
, " predicate: ");
1373 dump_predicate (f
, info
->conds
, es
->predicate
);
1377 if (es
->param
.exists ())
1378 for (i
= 0; i
< (int) es
->param
.length (); i
++)
1380 int prob
= es
->param
[i
].change_prob
;
1383 fprintf (f
, "%*s op%i is compile time invariant\n",
1385 else if (prob
!= REG_BR_PROB_BASE
)
1386 fprintf (f
, "%*s op%i change %f%% of time\n", indent
+ 2, "", i
,
1387 prob
* 100.0 / REG_BR_PROB_BASE
);
1389 if (!edge
->inline_failed
)
1391 fprintf (f
, "%*sStack frame offset %i, callee self size %i,"
1392 " callee size %i\n",
1394 (int) inline_summaries
->get (callee
)->stack_frame_offset
,
1395 (int) inline_summaries
->get (callee
)->estimated_self_stack_size
,
1396 (int) inline_summaries
->get (callee
)->estimated_stack_size
);
1397 dump_inline_edge_summary (f
, indent
+ 2, callee
, info
);
1400 for (edge
= node
->indirect_calls
; edge
; edge
= edge
->next_callee
)
1402 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
1403 fprintf (f
, "%*sindirect call loop depth:%2i freq:%4i size:%2i"
1407 edge
->frequency
, es
->call_stmt_size
, es
->call_stmt_time
);
1410 fprintf (f
, "predicate: ");
1411 dump_predicate (f
, info
->conds
, es
->predicate
);
1420 dump_inline_summary (FILE *f
, struct cgraph_node
*node
)
1422 if (node
->definition
)
1424 struct inline_summary
*s
= inline_summaries
->get (node
);
1427 fprintf (f
, "Inline summary for %s/%i", node
->name (),
1429 if (DECL_DISREGARD_INLINE_LIMITS (node
->decl
))
1430 fprintf (f
, " always_inline");
1432 fprintf (f
, " inlinable");
1433 if (s
->contains_cilk_spawn
)
1434 fprintf (f
, " contains_cilk_spawn");
1435 if (s
->fp_expressions
)
1436 fprintf (f
, " fp_expression");
1437 fprintf (f
, "\n self time: %i\n", s
->self_time
);
1438 fprintf (f
, " global time: %i\n", s
->time
);
1439 fprintf (f
, " self size: %i\n", s
->self_size
);
1440 fprintf (f
, " global size: %i\n", s
->size
);
1441 fprintf (f
, " min size: %i\n", s
->min_size
);
1442 fprintf (f
, " self stack: %i\n",
1443 (int) s
->estimated_self_stack_size
);
1444 fprintf (f
, " global stack: %i\n", (int) s
->estimated_stack_size
);
1446 fprintf (f
, " estimated growth:%i\n", (int) s
->growth
);
1448 fprintf (f
, " In SCC: %i\n", (int) s
->scc_no
);
1449 for (i
= 0; vec_safe_iterate (s
->entry
, i
, &e
); i
++)
1451 fprintf (f
, " size:%f, time:%f, predicate:",
1452 (double) e
->size
/ INLINE_SIZE_SCALE
,
1453 (double) e
->time
/ INLINE_TIME_SCALE
);
1454 dump_predicate (f
, s
->conds
, &e
->predicate
);
1456 if (s
->loop_iterations
)
1458 fprintf (f
, " loop iterations:");
1459 dump_predicate (f
, s
->conds
, s
->loop_iterations
);
1463 fprintf (f
, " loop stride:");
1464 dump_predicate (f
, s
->conds
, s
->loop_stride
);
1468 fprintf (f
, " array index:");
1469 dump_predicate (f
, s
->conds
, s
->array_index
);
1471 fprintf (f
, " calls:\n");
1472 dump_inline_edge_summary (f
, 4, node
, s
);
1478 debug_inline_summary (struct cgraph_node
*node
)
1480 dump_inline_summary (stderr
, node
);
1484 dump_inline_summaries (FILE *f
)
1486 struct cgraph_node
*node
;
1488 FOR_EACH_DEFINED_FUNCTION (node
)
1489 if (!node
->global
.inlined_to
)
1490 dump_inline_summary (f
, node
);
1493 /* Give initial reasons why inlining would fail on EDGE. This gets either
1494 nullified or usually overwritten by more precise reasons later. */
1497 initialize_inline_failed (struct cgraph_edge
*e
)
1499 struct cgraph_node
*callee
= e
->callee
;
1501 if (e
->inline_failed
&& e
->inline_failed
!= CIF_BODY_NOT_AVAILABLE
1502 && cgraph_inline_failed_type (e
->inline_failed
) == CIF_FINAL_ERROR
)
1504 else if (e
->indirect_unknown_callee
)
1505 e
->inline_failed
= CIF_INDIRECT_UNKNOWN_CALL
;
1506 else if (!callee
->definition
)
1507 e
->inline_failed
= CIF_BODY_NOT_AVAILABLE
;
1508 else if (callee
->local
.redefined_extern_inline
)
1509 e
->inline_failed
= CIF_REDEFINED_EXTERN_INLINE
;
1511 e
->inline_failed
= CIF_FUNCTION_NOT_CONSIDERED
;
1512 gcc_checking_assert (!e
->call_stmt_cannot_inline_p
1513 || cgraph_inline_failed_type (e
->inline_failed
)
1514 == CIF_FINAL_ERROR
);
1517 /* Callback of walk_aliased_vdefs. Flags that it has been invoked to the
1518 boolean variable pointed to by DATA. */
1521 mark_modified (ao_ref
*ao ATTRIBUTE_UNUSED
, tree vdef ATTRIBUTE_UNUSED
,
1524 bool *b
= (bool *) data
;
1529 /* If OP refers to value of function parameter, return the corresponding
1530 parameter. If non-NULL, the size of the memory load (or the SSA_NAME of the
1531 PARM_DECL) will be stored to *SIZE_P in that case too. */
1534 unmodified_parm_1 (gimple
*stmt
, tree op
, HOST_WIDE_INT
*size_p
)
1536 /* SSA_NAME referring to parm default def? */
1537 if (TREE_CODE (op
) == SSA_NAME
1538 && SSA_NAME_IS_DEFAULT_DEF (op
)
1539 && TREE_CODE (SSA_NAME_VAR (op
)) == PARM_DECL
)
1542 *size_p
= tree_to_shwi (TYPE_SIZE (TREE_TYPE (op
)));
1543 return SSA_NAME_VAR (op
);
1545 /* Non-SSA parm reference? */
1546 if (TREE_CODE (op
) == PARM_DECL
)
1548 bool modified
= false;
1551 ao_ref_init (&refd
, op
);
1552 walk_aliased_vdefs (&refd
, gimple_vuse (stmt
), mark_modified
, &modified
,
1557 *size_p
= tree_to_shwi (TYPE_SIZE (TREE_TYPE (op
)));
1564 /* If OP refers to value of function parameter, return the corresponding
1565 parameter. Also traverse chains of SSA register assignments. If non-NULL,
1566 the size of the memory load (or the SSA_NAME of the PARM_DECL) will be
1567 stored to *SIZE_P in that case too. */
1570 unmodified_parm (gimple
*stmt
, tree op
, HOST_WIDE_INT
*size_p
)
1572 tree res
= unmodified_parm_1 (stmt
, op
, size_p
);
1576 if (TREE_CODE (op
) == SSA_NAME
1577 && !SSA_NAME_IS_DEFAULT_DEF (op
)
1578 && gimple_assign_single_p (SSA_NAME_DEF_STMT (op
)))
1579 return unmodified_parm (SSA_NAME_DEF_STMT (op
),
1580 gimple_assign_rhs1 (SSA_NAME_DEF_STMT (op
)),
1585 /* If OP refers to a value of a function parameter or value loaded from an
1586 aggregate passed to a parameter (either by value or reference), return TRUE
1587 and store the number of the parameter to *INDEX_P, the access size into
1588 *SIZE_P, and information whether and how it has been loaded from an
1589 aggregate into *AGGPOS. INFO describes the function parameters, STMT is the
1590 statement in which OP is used or loaded. */
1593 unmodified_parm_or_parm_agg_item (struct ipa_func_body_info
*fbi
,
1594 gimple
*stmt
, tree op
, int *index_p
,
1595 HOST_WIDE_INT
*size_p
,
1596 struct agg_position_info
*aggpos
)
1598 tree res
= unmodified_parm_1 (stmt
, op
, size_p
);
1600 gcc_checking_assert (aggpos
);
1603 *index_p
= ipa_get_param_decl_index (fbi
->info
, res
);
1606 aggpos
->agg_contents
= false;
1607 aggpos
->by_ref
= false;
1611 if (TREE_CODE (op
) == SSA_NAME
)
1613 if (SSA_NAME_IS_DEFAULT_DEF (op
)
1614 || !gimple_assign_single_p (SSA_NAME_DEF_STMT (op
)))
1616 stmt
= SSA_NAME_DEF_STMT (op
);
1617 op
= gimple_assign_rhs1 (stmt
);
1618 if (!REFERENCE_CLASS_P (op
))
1619 return unmodified_parm_or_parm_agg_item (fbi
, stmt
, op
, index_p
, size_p
,
1623 aggpos
->agg_contents
= true;
1624 return ipa_load_from_parm_agg (fbi
, fbi
->info
->descriptors
,
1625 stmt
, op
, index_p
, &aggpos
->offset
,
1626 size_p
, &aggpos
->by_ref
);
1629 /* See if statement might disappear after inlining.
1630 0 - means not eliminated
1631 1 - half of statements goes away
1632 2 - for sure it is eliminated.
1633 We are not terribly sophisticated, basically looking for simple abstraction
1634 penalty wrappers. */
1637 eliminated_by_inlining_prob (gimple
*stmt
)
1639 enum gimple_code code
= gimple_code (stmt
);
1640 enum tree_code rhs_code
;
1650 if (gimple_num_ops (stmt
) != 2)
1653 rhs_code
= gimple_assign_rhs_code (stmt
);
1655 /* Casts of parameters, loads from parameters passed by reference
1656 and stores to return value or parameters are often free after
1657 inlining dua to SRA and further combining.
1658 Assume that half of statements goes away. */
1659 if (CONVERT_EXPR_CODE_P (rhs_code
)
1660 || rhs_code
== VIEW_CONVERT_EXPR
1661 || rhs_code
== ADDR_EXPR
1662 || gimple_assign_rhs_class (stmt
) == GIMPLE_SINGLE_RHS
)
1664 tree rhs
= gimple_assign_rhs1 (stmt
);
1665 tree lhs
= gimple_assign_lhs (stmt
);
1666 tree inner_rhs
= get_base_address (rhs
);
1667 tree inner_lhs
= get_base_address (lhs
);
1668 bool rhs_free
= false;
1669 bool lhs_free
= false;
1676 /* Reads of parameter are expected to be free. */
1677 if (unmodified_parm (stmt
, inner_rhs
, NULL
))
1679 /* Match expressions of form &this->field. Those will most likely
1680 combine with something upstream after inlining. */
1681 else if (TREE_CODE (inner_rhs
) == ADDR_EXPR
)
1683 tree op
= get_base_address (TREE_OPERAND (inner_rhs
, 0));
1684 if (TREE_CODE (op
) == PARM_DECL
)
1686 else if (TREE_CODE (op
) == MEM_REF
1687 && unmodified_parm (stmt
, TREE_OPERAND (op
, 0), NULL
))
1691 /* When parameter is not SSA register because its address is taken
1692 and it is just copied into one, the statement will be completely
1693 free after inlining (we will copy propagate backward). */
1694 if (rhs_free
&& is_gimple_reg (lhs
))
1697 /* Reads of parameters passed by reference
1698 expected to be free (i.e. optimized out after inlining). */
1699 if (TREE_CODE (inner_rhs
) == MEM_REF
1700 && unmodified_parm (stmt
, TREE_OPERAND (inner_rhs
, 0), NULL
))
1703 /* Copying parameter passed by reference into gimple register is
1704 probably also going to copy propagate, but we can't be quite
1706 if (rhs_free
&& is_gimple_reg (lhs
))
1709 /* Writes to parameters, parameters passed by value and return value
1710 (either dirrectly or passed via invisible reference) are free.
1712 TODO: We ought to handle testcase like
1713 struct a {int a,b;};
1715 retrurnsturct (void)
1721 This translate into:
1736 For that we either need to copy ipa-split logic detecting writes
1738 if (TREE_CODE (inner_lhs
) == PARM_DECL
1739 || TREE_CODE (inner_lhs
) == RESULT_DECL
1740 || (TREE_CODE (inner_lhs
) == MEM_REF
1741 && (unmodified_parm (stmt
, TREE_OPERAND (inner_lhs
, 0), NULL
)
1742 || (TREE_CODE (TREE_OPERAND (inner_lhs
, 0)) == SSA_NAME
1743 && SSA_NAME_VAR (TREE_OPERAND (inner_lhs
, 0))
1744 && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND
1746 0))) == RESULT_DECL
))))
1749 && (is_gimple_reg (rhs
) || is_gimple_min_invariant (rhs
)))
1751 if (lhs_free
&& rhs_free
)
1761 /* If BB ends by a conditional we can turn into predicates, attach corresponding
1762 predicates to the CFG edges. */
1765 set_cond_stmt_execution_predicate (struct ipa_func_body_info
*fbi
,
1766 struct inline_summary
*summary
,
1773 struct agg_position_info aggpos
;
1774 enum tree_code code
, inverted_code
;
1780 last
= last_stmt (bb
);
1781 if (!last
|| gimple_code (last
) != GIMPLE_COND
)
1783 if (!is_gimple_ip_invariant (gimple_cond_rhs (last
)))
1785 op
= gimple_cond_lhs (last
);
1786 /* TODO: handle conditionals like
1789 if (unmodified_parm_or_parm_agg_item (fbi
, last
, op
, &index
, &size
, &aggpos
))
1791 code
= gimple_cond_code (last
);
1792 inverted_code
= invert_tree_comparison (code
, HONOR_NANS (op
));
1794 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1796 enum tree_code this_code
= (e
->flags
& EDGE_TRUE_VALUE
1797 ? code
: inverted_code
);
1798 /* invert_tree_comparison will return ERROR_MARK on FP
1799 comparsions that are not EQ/NE instead of returning proper
1800 unordered one. Be sure it is not confused with NON_CONSTANT. */
1801 if (this_code
!= ERROR_MARK
)
1804 = add_condition (summary
, index
, size
, &aggpos
, this_code
,
1805 unshare_expr_without_location
1806 (gimple_cond_rhs (last
)));
1807 e
->aux
= edge_predicate_pool
.allocate ();
1808 *(struct predicate
*) e
->aux
= p
;
1813 if (TREE_CODE (op
) != SSA_NAME
)
1816 if (builtin_constant_p (op))
1820 Here we can predicate nonconstant_code. We can't
1821 really handle constant_code since we have no predicate
1822 for this and also the constant code is not known to be
1823 optimized away when inliner doen't see operand is constant.
1824 Other optimizers might think otherwise. */
1825 if (gimple_cond_code (last
) != NE_EXPR
1826 || !integer_zerop (gimple_cond_rhs (last
)))
1828 set_stmt
= SSA_NAME_DEF_STMT (op
);
1829 if (!gimple_call_builtin_p (set_stmt
, BUILT_IN_CONSTANT_P
)
1830 || gimple_call_num_args (set_stmt
) != 1)
1832 op2
= gimple_call_arg (set_stmt
, 0);
1833 if (!unmodified_parm_or_parm_agg_item (fbi
, set_stmt
, op2
, &index
, &size
,
1836 FOR_EACH_EDGE (e
, ei
, bb
->succs
) if (e
->flags
& EDGE_FALSE_VALUE
)
1838 struct predicate p
= add_condition (summary
, index
, size
, &aggpos
,
1839 IS_NOT_CONSTANT
, NULL_TREE
);
1840 e
->aux
= edge_predicate_pool
.allocate ();
1841 *(struct predicate
*) e
->aux
= p
;
1846 /* If BB ends by a switch we can turn into predicates, attach corresponding
1847 predicates to the CFG edges. */
1850 set_switch_stmt_execution_predicate (struct ipa_func_body_info
*fbi
,
1851 struct inline_summary
*summary
,
1858 struct agg_position_info aggpos
;
1864 lastg
= last_stmt (bb
);
1865 if (!lastg
|| gimple_code (lastg
) != GIMPLE_SWITCH
)
1867 gswitch
*last
= as_a
<gswitch
*> (lastg
);
1868 op
= gimple_switch_index (last
);
1869 if (!unmodified_parm_or_parm_agg_item (fbi
, last
, op
, &index
, &size
, &aggpos
))
1872 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1874 e
->aux
= edge_predicate_pool
.allocate ();
1875 *(struct predicate
*) e
->aux
= false_predicate ();
1877 n
= gimple_switch_num_labels (last
);
1878 for (case_idx
= 0; case_idx
< n
; ++case_idx
)
1880 tree cl
= gimple_switch_label (last
, case_idx
);
1884 e
= find_edge (bb
, label_to_block (CASE_LABEL (cl
)));
1885 min
= CASE_LOW (cl
);
1886 max
= CASE_HIGH (cl
);
1888 /* For default we might want to construct predicate that none
1889 of cases is met, but it is bit hard to do not having negations
1890 of conditionals handy. */
1892 p
= true_predicate ();
1894 p
= add_condition (summary
, index
, size
, &aggpos
, EQ_EXPR
,
1895 unshare_expr_without_location (min
));
1898 struct predicate p1
, p2
;
1899 p1
= add_condition (summary
, index
, size
, &aggpos
, GE_EXPR
,
1900 unshare_expr_without_location (min
));
1901 p2
= add_condition (summary
, index
, size
, &aggpos
, LE_EXPR
,
1902 unshare_expr_without_location (max
));
1903 p
= and_predicates (summary
->conds
, &p1
, &p2
);
1905 *(struct predicate
*) e
->aux
1906 = or_predicates (summary
->conds
, &p
, (struct predicate
*) e
->aux
);
1911 /* For each BB in NODE attach to its AUX pointer predicate under
1912 which it is executable. */
1915 compute_bb_predicates (struct ipa_func_body_info
*fbi
,
1916 struct cgraph_node
*node
,
1917 struct inline_summary
*summary
)
1919 struct function
*my_function
= DECL_STRUCT_FUNCTION (node
->decl
);
1923 FOR_EACH_BB_FN (bb
, my_function
)
1925 set_cond_stmt_execution_predicate (fbi
, summary
, bb
);
1926 set_switch_stmt_execution_predicate (fbi
, summary
, bb
);
1929 /* Entry block is always executable. */
1930 ENTRY_BLOCK_PTR_FOR_FN (my_function
)->aux
1931 = edge_predicate_pool
.allocate ();
1932 *(struct predicate
*) ENTRY_BLOCK_PTR_FOR_FN (my_function
)->aux
1933 = true_predicate ();
1935 /* A simple dataflow propagation of predicates forward in the CFG.
1936 TODO: work in reverse postorder. */
1940 FOR_EACH_BB_FN (bb
, my_function
)
1942 struct predicate p
= false_predicate ();
1945 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
1949 struct predicate this_bb_predicate
1950 = *(struct predicate
*) e
->src
->aux
;
1953 = and_predicates (summary
->conds
, &this_bb_predicate
,
1954 (struct predicate
*) e
->aux
);
1955 p
= or_predicates (summary
->conds
, &p
, &this_bb_predicate
);
1956 if (true_predicate_p (&p
))
1960 if (false_predicate_p (&p
))
1961 gcc_assert (!bb
->aux
);
1967 bb
->aux
= edge_predicate_pool
.allocate ();
1968 *((struct predicate
*) bb
->aux
) = p
;
1970 else if (!predicates_equal_p (&p
, (struct predicate
*) bb
->aux
))
1972 /* This OR operation is needed to ensure monotonous data flow
1973 in the case we hit the limit on number of clauses and the
1974 and/or operations above give approximate answers. */
1975 p
= or_predicates (summary
->conds
, &p
, (struct predicate
*)bb
->aux
);
1976 if (!predicates_equal_p (&p
, (struct predicate
*) bb
->aux
))
1979 *((struct predicate
*) bb
->aux
) = p
;
1988 /* We keep info about constantness of SSA names. */
1990 typedef struct predicate predicate_t
;
1991 /* Return predicate specifying when the STMT might have result that is not
1992 a compile time constant. */
1994 static struct predicate
1995 will_be_nonconstant_expr_predicate (struct ipa_node_params
*info
,
1996 struct inline_summary
*summary
,
1998 vec
<predicate_t
> nonconstant_names
)
2004 while (UNARY_CLASS_P (expr
))
2005 expr
= TREE_OPERAND (expr
, 0);
2007 parm
= unmodified_parm (NULL
, expr
, &size
);
2008 if (parm
&& (index
= ipa_get_param_decl_index (info
, parm
)) >= 0)
2009 return add_condition (summary
, index
, size
, NULL
, CHANGED
, NULL_TREE
);
2010 if (is_gimple_min_invariant (expr
))
2011 return false_predicate ();
2012 if (TREE_CODE (expr
) == SSA_NAME
)
2013 return nonconstant_names
[SSA_NAME_VERSION (expr
)];
2014 if (BINARY_CLASS_P (expr
) || COMPARISON_CLASS_P (expr
))
2016 struct predicate p1
= will_be_nonconstant_expr_predicate
2017 (info
, summary
, TREE_OPERAND (expr
, 0),
2019 struct predicate p2
;
2020 if (true_predicate_p (&p1
))
2022 p2
= will_be_nonconstant_expr_predicate (info
, summary
,
2023 TREE_OPERAND (expr
, 1),
2025 return or_predicates (summary
->conds
, &p1
, &p2
);
2027 else if (TREE_CODE (expr
) == COND_EXPR
)
2029 struct predicate p1
= will_be_nonconstant_expr_predicate
2030 (info
, summary
, TREE_OPERAND (expr
, 0),
2032 struct predicate p2
;
2033 if (true_predicate_p (&p1
))
2035 p2
= will_be_nonconstant_expr_predicate (info
, summary
,
2036 TREE_OPERAND (expr
, 1),
2038 if (true_predicate_p (&p2
))
2040 p1
= or_predicates (summary
->conds
, &p1
, &p2
);
2041 p2
= will_be_nonconstant_expr_predicate (info
, summary
,
2042 TREE_OPERAND (expr
, 2),
2044 return or_predicates (summary
->conds
, &p1
, &p2
);
2051 return false_predicate ();
2055 /* Return predicate specifying when the STMT might have result that is not
2056 a compile time constant. */
2058 static struct predicate
2059 will_be_nonconstant_predicate (struct ipa_func_body_info
*fbi
,
2060 struct inline_summary
*summary
,
2062 vec
<predicate_t
> nonconstant_names
)
2064 struct predicate p
= true_predicate ();
2067 struct predicate op_non_const
;
2071 struct agg_position_info aggpos
;
2073 /* What statments might be optimized away
2074 when their arguments are constant. */
2075 if (gimple_code (stmt
) != GIMPLE_ASSIGN
2076 && gimple_code (stmt
) != GIMPLE_COND
2077 && gimple_code (stmt
) != GIMPLE_SWITCH
2078 && (gimple_code (stmt
) != GIMPLE_CALL
2079 || !(gimple_call_flags (stmt
) & ECF_CONST
)))
2082 /* Stores will stay anyway. */
2083 if (gimple_store_p (stmt
))
2086 is_load
= gimple_assign_load_p (stmt
);
2088 /* Loads can be optimized when the value is known. */
2092 gcc_assert (gimple_assign_single_p (stmt
));
2093 op
= gimple_assign_rhs1 (stmt
);
2094 if (!unmodified_parm_or_parm_agg_item (fbi
, stmt
, op
, &base_index
, &size
,
2101 /* See if we understand all operands before we start
2102 adding conditionals. */
2103 FOR_EACH_SSA_TREE_OPERAND (use
, stmt
, iter
, SSA_OP_USE
)
2105 tree parm
= unmodified_parm (stmt
, use
, NULL
);
2106 /* For arguments we can build a condition. */
2107 if (parm
&& ipa_get_param_decl_index (fbi
->info
, parm
) >= 0)
2109 if (TREE_CODE (use
) != SSA_NAME
)
2111 /* If we know when operand is constant,
2112 we still can say something useful. */
2113 if (!true_predicate_p (&nonconstant_names
[SSA_NAME_VERSION (use
)]))
2120 add_condition (summary
, base_index
, size
, &aggpos
, CHANGED
, NULL
);
2122 op_non_const
= false_predicate ();
2123 FOR_EACH_SSA_TREE_OPERAND (use
, stmt
, iter
, SSA_OP_USE
)
2126 tree parm
= unmodified_parm (stmt
, use
, &size
);
2129 if (parm
&& (index
= ipa_get_param_decl_index (fbi
->info
, parm
)) >= 0)
2131 if (index
!= base_index
)
2132 p
= add_condition (summary
, index
, size
, NULL
, CHANGED
, NULL_TREE
);
2137 p
= nonconstant_names
[SSA_NAME_VERSION (use
)];
2138 op_non_const
= or_predicates (summary
->conds
, &p
, &op_non_const
);
2140 if ((gimple_code (stmt
) == GIMPLE_ASSIGN
|| gimple_code (stmt
) == GIMPLE_CALL
)
2141 && gimple_op (stmt
, 0)
2142 && TREE_CODE (gimple_op (stmt
, 0)) == SSA_NAME
)
2143 nonconstant_names
[SSA_NAME_VERSION (gimple_op (stmt
, 0))]
2145 return op_non_const
;
2148 struct record_modified_bb_info
2154 /* Callback of walk_aliased_vdefs. Records basic blocks where the value may be
2155 set except for info->stmt. */
2158 record_modified (ao_ref
*ao ATTRIBUTE_UNUSED
, tree vdef
, void *data
)
2160 struct record_modified_bb_info
*info
=
2161 (struct record_modified_bb_info
*) data
;
2162 if (SSA_NAME_DEF_STMT (vdef
) == info
->stmt
)
2164 bitmap_set_bit (info
->bb_set
,
2165 SSA_NAME_IS_DEFAULT_DEF (vdef
)
2166 ? ENTRY_BLOCK_PTR_FOR_FN (cfun
)->index
2167 : gimple_bb (SSA_NAME_DEF_STMT (vdef
))->index
);
2171 /* Return probability (based on REG_BR_PROB_BASE) that I-th parameter of STMT
2172 will change since last invocation of STMT.
2174 Value 0 is reserved for compile time invariants.
2175 For common parameters it is REG_BR_PROB_BASE. For loop invariants it
2176 ought to be REG_BR_PROB_BASE / estimated_iters. */
2179 param_change_prob (gimple
*stmt
, int i
)
2181 tree op
= gimple_call_arg (stmt
, i
);
2182 basic_block bb
= gimple_bb (stmt
);
2184 if (TREE_CODE (op
) == WITH_SIZE_EXPR
)
2185 op
= TREE_OPERAND (op
, 0);
2187 tree base
= get_base_address (op
);
2189 /* Global invariants never change. */
2190 if (is_gimple_min_invariant (base
))
2193 /* We would have to do non-trivial analysis to really work out what
2194 is the probability of value to change (i.e. when init statement
2195 is in a sibling loop of the call).
2197 We do an conservative estimate: when call is executed N times more often
2198 than the statement defining value, we take the frequency 1/N. */
2199 if (TREE_CODE (base
) == SSA_NAME
)
2204 return REG_BR_PROB_BASE
;
2206 if (SSA_NAME_IS_DEFAULT_DEF (base
))
2207 init_freq
= ENTRY_BLOCK_PTR_FOR_FN (cfun
)->frequency
;
2209 init_freq
= gimple_bb (SSA_NAME_DEF_STMT (base
))->frequency
;
2213 if (init_freq
< bb
->frequency
)
2214 return MAX (GCOV_COMPUTE_SCALE (init_freq
, bb
->frequency
), 1);
2216 return REG_BR_PROB_BASE
;
2222 struct record_modified_bb_info info
;
2225 tree init
= ctor_for_folding (base
);
2227 if (init
!= error_mark_node
)
2230 return REG_BR_PROB_BASE
;
2231 ao_ref_init (&refd
, op
);
2233 info
.bb_set
= BITMAP_ALLOC (NULL
);
2234 walk_aliased_vdefs (&refd
, gimple_vuse (stmt
), record_modified
, &info
,
2236 if (bitmap_bit_p (info
.bb_set
, bb
->index
))
2238 BITMAP_FREE (info
.bb_set
);
2239 return REG_BR_PROB_BASE
;
2242 /* Assume that every memory is initialized at entry.
2243 TODO: Can we easilly determine if value is always defined
2244 and thus we may skip entry block? */
2245 if (ENTRY_BLOCK_PTR_FOR_FN (cfun
)->frequency
)
2246 max
= ENTRY_BLOCK_PTR_FOR_FN (cfun
)->frequency
;
2250 EXECUTE_IF_SET_IN_BITMAP (info
.bb_set
, 0, index
, bi
)
2251 max
= MIN (max
, BASIC_BLOCK_FOR_FN (cfun
, index
)->frequency
);
2253 BITMAP_FREE (info
.bb_set
);
2254 if (max
< bb
->frequency
)
2255 return MAX (GCOV_COMPUTE_SCALE (max
, bb
->frequency
), 1);
2257 return REG_BR_PROB_BASE
;
2261 /* Find whether a basic block BB is the final block of a (half) diamond CFG
2262 sub-graph and if the predicate the condition depends on is known. If so,
2263 return true and store the pointer the predicate in *P. */
2266 phi_result_unknown_predicate (struct ipa_node_params
*info
,
2267 inline_summary
*summary
, basic_block bb
,
2268 struct predicate
*p
,
2269 vec
<predicate_t
> nonconstant_names
)
2273 basic_block first_bb
= NULL
;
2276 if (single_pred_p (bb
))
2278 *p
= false_predicate ();
2282 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
2284 if (single_succ_p (e
->src
))
2286 if (!single_pred_p (e
->src
))
2289 first_bb
= single_pred (e
->src
);
2290 else if (single_pred (e
->src
) != first_bb
)
2297 else if (e
->src
!= first_bb
)
2305 stmt
= last_stmt (first_bb
);
2307 || gimple_code (stmt
) != GIMPLE_COND
2308 || !is_gimple_ip_invariant (gimple_cond_rhs (stmt
)))
2311 *p
= will_be_nonconstant_expr_predicate (info
, summary
,
2312 gimple_cond_lhs (stmt
),
2314 if (true_predicate_p (p
))
2320 /* Given a PHI statement in a function described by inline properties SUMMARY
2321 and *P being the predicate describing whether the selected PHI argument is
2322 known, store a predicate for the result of the PHI statement into
2323 NONCONSTANT_NAMES, if possible. */
2326 predicate_for_phi_result (struct inline_summary
*summary
, gphi
*phi
,
2327 struct predicate
*p
,
2328 vec
<predicate_t
> nonconstant_names
)
2332 for (i
= 0; i
< gimple_phi_num_args (phi
); i
++)
2334 tree arg
= gimple_phi_arg (phi
, i
)->def
;
2335 if (!is_gimple_min_invariant (arg
))
2337 gcc_assert (TREE_CODE (arg
) == SSA_NAME
);
2338 *p
= or_predicates (summary
->conds
, p
,
2339 &nonconstant_names
[SSA_NAME_VERSION (arg
)]);
2340 if (true_predicate_p (p
))
2345 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2347 fprintf (dump_file
, "\t\tphi predicate: ");
2348 dump_predicate (dump_file
, summary
->conds
, p
);
2350 nonconstant_names
[SSA_NAME_VERSION (gimple_phi_result (phi
))] = *p
;
2353 /* Return predicate specifying when array index in access OP becomes non-constant. */
2355 static struct predicate
2356 array_index_predicate (inline_summary
*info
,
2357 vec
< predicate_t
> nonconstant_names
, tree op
)
2359 struct predicate p
= false_predicate ();
2360 while (handled_component_p (op
))
2362 if (TREE_CODE (op
) == ARRAY_REF
|| TREE_CODE (op
) == ARRAY_RANGE_REF
)
2364 if (TREE_CODE (TREE_OPERAND (op
, 1)) == SSA_NAME
)
2365 p
= or_predicates (info
->conds
, &p
,
2366 &nonconstant_names
[SSA_NAME_VERSION
2367 (TREE_OPERAND (op
, 1))]);
2369 op
= TREE_OPERAND (op
, 0);
2374 /* For a typical usage of __builtin_expect (a<b, 1), we
2375 may introduce an extra relation stmt:
2376 With the builtin, we have
2379 t3 = __builtin_expect (t2, 1);
2382 Without the builtin, we have
2385 This affects the size/time estimation and may have
2386 an impact on the earlier inlining.
2387 Here find this pattern and fix it up later. */
2390 find_foldable_builtin_expect (basic_block bb
)
2392 gimple_stmt_iterator bsi
;
2394 for (bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
2396 gimple
*stmt
= gsi_stmt (bsi
);
2397 if (gimple_call_builtin_p (stmt
, BUILT_IN_EXPECT
)
2398 || gimple_call_internal_p (stmt
, IFN_BUILTIN_EXPECT
))
2400 tree var
= gimple_call_lhs (stmt
);
2401 tree arg
= gimple_call_arg (stmt
, 0);
2402 use_operand_p use_p
;
2409 gcc_assert (TREE_CODE (var
) == SSA_NAME
);
2411 while (TREE_CODE (arg
) == SSA_NAME
)
2413 gimple
*stmt_tmp
= SSA_NAME_DEF_STMT (arg
);
2414 if (!is_gimple_assign (stmt_tmp
))
2416 switch (gimple_assign_rhs_code (stmt_tmp
))
2435 arg
= gimple_assign_rhs1 (stmt_tmp
);
2438 if (match
&& single_imm_use (var
, &use_p
, &use_stmt
)
2439 && gimple_code (use_stmt
) == GIMPLE_COND
)
2446 /* Return true when the basic blocks contains only clobbers followed by RESX.
2447 Such BBs are kept around to make removal of dead stores possible with
2448 presence of EH and will be optimized out by optimize_clobbers later in the
2451 NEED_EH is used to recurse in case the clobber has non-EH predecestors
2452 that can be clobber only, too.. When it is false, the RESX is not necessary
2453 on the end of basic block. */
2456 clobber_only_eh_bb_p (basic_block bb
, bool need_eh
= true)
2458 gimple_stmt_iterator gsi
= gsi_last_bb (bb
);
2464 if (gsi_end_p (gsi
))
2466 if (gimple_code (gsi_stmt (gsi
)) != GIMPLE_RESX
)
2470 else if (!single_succ_p (bb
))
2473 for (; !gsi_end_p (gsi
); gsi_prev (&gsi
))
2475 gimple
*stmt
= gsi_stmt (gsi
);
2476 if (is_gimple_debug (stmt
))
2478 if (gimple_clobber_p (stmt
))
2480 if (gimple_code (stmt
) == GIMPLE_LABEL
)
2485 /* See if all predecestors are either throws or clobber only BBs. */
2486 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
2487 if (!(e
->flags
& EDGE_EH
)
2488 && !clobber_only_eh_bb_p (e
->src
, false))
2494 /* Return true if STMT compute a floating point expression that may be affected
2495 by -ffast-math and similar flags. */
2498 fp_expression_p (gimple
*stmt
)
2503 FOR_EACH_SSA_TREE_OPERAND (op
, stmt
, i
, SSA_OP_DEF
|SSA_OP_USE
)
2504 if (FLOAT_TYPE_P (TREE_TYPE (op
)))
2509 /* Compute function body size parameters for NODE.
2510 When EARLY is true, we compute only simple summaries without
2511 non-trivial predicates to drive the early inliner. */
2514 estimate_function_body_sizes (struct cgraph_node
*node
, bool early
)
2517 /* Estimate static overhead for function prologue/epilogue and alignment. */
2519 /* Benefits are scaled by probability of elimination that is in range
2522 struct function
*my_function
= DECL_STRUCT_FUNCTION (node
->decl
);
2524 struct inline_summary
*info
= inline_summaries
->get (node
);
2525 struct predicate bb_predicate
;
2526 struct ipa_func_body_info fbi
;
2527 vec
<predicate_t
> nonconstant_names
= vNULL
;
2530 predicate array_index
= true_predicate ();
2531 gimple
*fix_builtin_expect_stmt
;
2533 gcc_assert (my_function
&& my_function
->cfg
);
2534 gcc_assert (cfun
== my_function
);
2536 memset(&fbi
, 0, sizeof(fbi
));
2540 /* When optimizing and analyzing for IPA inliner, initialize loop optimizer
2541 so we can produce proper inline hints.
2543 When optimizing and analyzing for early inliner, initialize node params
2544 so we can produce correct BB predicates. */
2546 if (opt_for_fn (node
->decl
, optimize
))
2548 calculate_dominance_info (CDI_DOMINATORS
);
2550 loop_optimizer_init (LOOPS_NORMAL
| LOOPS_HAVE_RECORDED_EXITS
);
2553 ipa_check_create_node_params ();
2554 ipa_initialize_node_params (node
);
2557 if (ipa_node_params_sum
)
2560 fbi
.info
= IPA_NODE_REF (node
);
2561 fbi
.bb_infos
= vNULL
;
2562 fbi
.bb_infos
.safe_grow_cleared (last_basic_block_for_fn (cfun
));
2563 fbi
.param_count
= count_formal_params(node
->decl
);
2564 nonconstant_names
.safe_grow_cleared
2565 (SSANAMES (my_function
)->length ());
2570 fprintf (dump_file
, "\nAnalyzing function body size: %s\n",
2573 /* When we run into maximal number of entries, we assign everything to the
2574 constant truth case. Be sure to have it in list. */
2575 bb_predicate
= true_predicate ();
2576 account_size_time (info
, 0, 0, &bb_predicate
);
2578 bb_predicate
= not_inlined_predicate ();
2579 account_size_time (info
, 2 * INLINE_SIZE_SCALE
, 0, &bb_predicate
);
2582 compute_bb_predicates (&fbi
, node
, info
);
2583 order
= XNEWVEC (int, n_basic_blocks_for_fn (cfun
));
2584 nblocks
= pre_and_rev_post_order_compute (NULL
, order
, false);
2585 for (n
= 0; n
< nblocks
; n
++)
2587 bb
= BASIC_BLOCK_FOR_FN (cfun
, order
[n
]);
2588 freq
= compute_call_stmt_bb_frequency (node
->decl
, bb
);
2589 if (clobber_only_eh_bb_p (bb
))
2591 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2592 fprintf (dump_file
, "\n Ignoring BB %i;"
2593 " it will be optimized away by cleanup_clobbers\n",
2598 /* TODO: Obviously predicates can be propagated down across CFG. */
2602 bb_predicate
= *(struct predicate
*) bb
->aux
;
2604 bb_predicate
= false_predicate ();
2607 bb_predicate
= true_predicate ();
2609 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2611 fprintf (dump_file
, "\n BB %i predicate:", bb
->index
);
2612 dump_predicate (dump_file
, info
->conds
, &bb_predicate
);
2615 if (fbi
.info
&& nonconstant_names
.exists ())
2617 struct predicate phi_predicate
;
2618 bool first_phi
= true;
2620 for (gphi_iterator bsi
= gsi_start_phis (bb
); !gsi_end_p (bsi
);
2624 && !phi_result_unknown_predicate (fbi
.info
, info
, bb
,
2629 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2631 fprintf (dump_file
, " ");
2632 print_gimple_stmt (dump_file
, gsi_stmt (bsi
), 0, 0);
2634 predicate_for_phi_result (info
, bsi
.phi (), &phi_predicate
,
2639 fix_builtin_expect_stmt
= find_foldable_builtin_expect (bb
);
2641 for (gimple_stmt_iterator bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
);
2644 gimple
*stmt
= gsi_stmt (bsi
);
2645 int this_size
= estimate_num_insns (stmt
, &eni_size_weights
);
2646 int this_time
= estimate_num_insns (stmt
, &eni_time_weights
);
2648 struct predicate will_be_nonconstant
;
2650 /* This relation stmt should be folded after we remove
2651 buildin_expect call. Adjust the cost here. */
2652 if (stmt
== fix_builtin_expect_stmt
)
2658 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2660 fprintf (dump_file
, " ");
2661 print_gimple_stmt (dump_file
, stmt
, 0, 0);
2662 fprintf (dump_file
, "\t\tfreq:%3.2f size:%3i time:%3i\n",
2663 ((double) freq
) / CGRAPH_FREQ_BASE
, this_size
,
2667 if (gimple_assign_load_p (stmt
) && nonconstant_names
.exists ())
2669 struct predicate this_array_index
;
2671 array_index_predicate (info
, nonconstant_names
,
2672 gimple_assign_rhs1 (stmt
));
2673 if (!false_predicate_p (&this_array_index
))
2675 and_predicates (info
->conds
, &array_index
,
2678 if (gimple_store_p (stmt
) && nonconstant_names
.exists ())
2680 struct predicate this_array_index
;
2682 array_index_predicate (info
, nonconstant_names
,
2683 gimple_get_lhs (stmt
));
2684 if (!false_predicate_p (&this_array_index
))
2686 and_predicates (info
->conds
, &array_index
,
2691 if (is_gimple_call (stmt
)
2692 && !gimple_call_internal_p (stmt
))
2694 struct cgraph_edge
*edge
= node
->get_edge (stmt
);
2695 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
2697 /* Special case: results of BUILT_IN_CONSTANT_P will be always
2698 resolved as constant. We however don't want to optimize
2699 out the cgraph edges. */
2700 if (nonconstant_names
.exists ()
2701 && gimple_call_builtin_p (stmt
, BUILT_IN_CONSTANT_P
)
2702 && gimple_call_lhs (stmt
)
2703 && TREE_CODE (gimple_call_lhs (stmt
)) == SSA_NAME
)
2705 struct predicate false_p
= false_predicate ();
2706 nonconstant_names
[SSA_NAME_VERSION (gimple_call_lhs (stmt
))]
2709 if (ipa_node_params_sum
)
2711 int count
= gimple_call_num_args (stmt
);
2715 es
->param
.safe_grow_cleared (count
);
2716 for (i
= 0; i
< count
; i
++)
2718 int prob
= param_change_prob (stmt
, i
);
2719 gcc_assert (prob
>= 0 && prob
<= REG_BR_PROB_BASE
);
2720 es
->param
[i
].change_prob
= prob
;
2724 es
->call_stmt_size
= this_size
;
2725 es
->call_stmt_time
= this_time
;
2726 es
->loop_depth
= bb_loop_depth (bb
);
2727 edge_set_predicate (edge
, &bb_predicate
);
2730 /* TODO: When conditional jump or swithc is known to be constant, but
2731 we did not translate it into the predicates, we really can account
2732 just maximum of the possible paths. */
2735 = will_be_nonconstant_predicate (&fbi
, info
,
2736 stmt
, nonconstant_names
);
2737 if (this_time
|| this_size
)
2743 prob
= eliminated_by_inlining_prob (stmt
);
2744 if (prob
== 1 && dump_file
&& (dump_flags
& TDF_DETAILS
))
2746 "\t\t50%% will be eliminated by inlining\n");
2747 if (prob
== 2 && dump_file
&& (dump_flags
& TDF_DETAILS
))
2748 fprintf (dump_file
, "\t\tWill be eliminated by inlining\n");
2751 p
= and_predicates (info
->conds
, &bb_predicate
,
2752 &will_be_nonconstant
);
2754 p
= true_predicate ();
2756 if (!false_predicate_p (&p
)
2757 || (is_gimple_call (stmt
)
2758 && !false_predicate_p (&bb_predicate
)))
2762 if (time
> MAX_TIME
* INLINE_TIME_SCALE
)
2763 time
= MAX_TIME
* INLINE_TIME_SCALE
;
2766 /* We account everything but the calls. Calls have their own
2767 size/time info attached to cgraph edges. This is necessary
2768 in order to make the cost disappear after inlining. */
2769 if (!is_gimple_call (stmt
))
2773 struct predicate ip
= not_inlined_predicate ();
2774 ip
= and_predicates (info
->conds
, &ip
, &p
);
2775 account_size_time (info
, this_size
* prob
,
2776 this_time
* prob
, &ip
);
2779 account_size_time (info
, this_size
* (2 - prob
),
2780 this_time
* (2 - prob
), &p
);
2783 if (!info
->fp_expressions
&& fp_expression_p (stmt
))
2785 info
->fp_expressions
= true;
2787 fprintf (dump_file
, " fp_expression set\n");
2790 gcc_assert (time
>= 0);
2791 gcc_assert (size
>= 0);
2795 set_hint_predicate (&inline_summaries
->get (node
)->array_index
, array_index
);
2796 time
= (time
+ CGRAPH_FREQ_BASE
/ 2) / CGRAPH_FREQ_BASE
;
2797 if (time
> MAX_TIME
)
2801 if (nonconstant_names
.exists () && !early
)
2804 predicate loop_iterations
= true_predicate ();
2805 predicate loop_stride
= true_predicate ();
2807 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2808 flow_loops_dump (dump_file
, NULL
, 0);
2810 FOR_EACH_LOOP (loop
, 0)
2815 struct tree_niter_desc niter_desc
;
2816 bb_predicate
= *(struct predicate
*) loop
->header
->aux
;
2818 exits
= get_loop_exit_edges (loop
);
2819 FOR_EACH_VEC_ELT (exits
, j
, ex
)
2820 if (number_of_iterations_exit (loop
, ex
, &niter_desc
, false)
2821 && !is_gimple_min_invariant (niter_desc
.niter
))
2823 predicate will_be_nonconstant
2824 = will_be_nonconstant_expr_predicate (fbi
.info
, info
,
2827 if (!true_predicate_p (&will_be_nonconstant
))
2828 will_be_nonconstant
= and_predicates (info
->conds
,
2830 &will_be_nonconstant
);
2831 if (!true_predicate_p (&will_be_nonconstant
)
2832 && !false_predicate_p (&will_be_nonconstant
))
2833 /* This is slightly inprecise. We may want to represent each
2834 loop with independent predicate. */
2836 and_predicates (info
->conds
, &loop_iterations
,
2837 &will_be_nonconstant
);
2842 /* To avoid quadratic behavior we analyze stride predicates only
2843 with respect to the containing loop. Thus we simply iterate
2844 over all defs in the outermost loop body. */
2845 for (loop
= loops_for_fn (cfun
)->tree_root
->inner
;
2846 loop
!= NULL
; loop
= loop
->next
)
2848 basic_block
*body
= get_loop_body (loop
);
2849 for (unsigned i
= 0; i
< loop
->num_nodes
; i
++)
2851 gimple_stmt_iterator gsi
;
2852 bb_predicate
= *(struct predicate
*) body
[i
]->aux
;
2853 for (gsi
= gsi_start_bb (body
[i
]); !gsi_end_p (gsi
);
2856 gimple
*stmt
= gsi_stmt (gsi
);
2858 if (!is_gimple_assign (stmt
))
2861 tree def
= gimple_assign_lhs (stmt
);
2862 if (TREE_CODE (def
) != SSA_NAME
)
2866 if (!simple_iv (loop_containing_stmt (stmt
),
2867 loop_containing_stmt (stmt
),
2869 || is_gimple_min_invariant (iv
.step
))
2872 predicate will_be_nonconstant
2873 = will_be_nonconstant_expr_predicate (fbi
.info
, info
,
2876 if (!true_predicate_p (&will_be_nonconstant
))
2878 = and_predicates (info
->conds
, &bb_predicate
,
2879 &will_be_nonconstant
);
2880 if (!true_predicate_p (&will_be_nonconstant
)
2881 && !false_predicate_p (&will_be_nonconstant
))
2882 /* This is slightly inprecise. We may want to represent
2883 each loop with independent predicate. */
2884 loop_stride
= and_predicates (info
->conds
, &loop_stride
,
2885 &will_be_nonconstant
);
2890 set_hint_predicate (&inline_summaries
->get (node
)->loop_iterations
,
2892 set_hint_predicate (&inline_summaries
->get (node
)->loop_stride
,
2896 FOR_ALL_BB_FN (bb
, my_function
)
2902 edge_predicate_pool
.remove ((predicate
*)bb
->aux
);
2904 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
2907 edge_predicate_pool
.remove ((predicate
*) e
->aux
);
2911 inline_summaries
->get (node
)->self_time
= time
;
2912 inline_summaries
->get (node
)->self_size
= size
;
2913 nonconstant_names
.release ();
2914 ipa_release_body_info (&fbi
);
2915 if (opt_for_fn (node
->decl
, optimize
))
2918 loop_optimizer_finalize ();
2919 else if (!ipa_edge_args_vector
)
2920 ipa_free_all_node_params ();
2921 free_dominance_info (CDI_DOMINATORS
);
2925 fprintf (dump_file
, "\n");
2926 dump_inline_summary (dump_file
, node
);
2931 /* Compute parameters of functions used by inliner.
2932 EARLY is true when we compute parameters for the early inliner */
2935 compute_inline_parameters (struct cgraph_node
*node
, bool early
)
2937 HOST_WIDE_INT self_stack_size
;
2938 struct cgraph_edge
*e
;
2939 struct inline_summary
*info
;
2941 gcc_assert (!node
->global
.inlined_to
);
2943 inline_summary_alloc ();
2945 info
= inline_summaries
->get (node
);
2946 reset_inline_summary (node
, info
);
2948 /* Estimate the stack size for the function if we're optimizing. */
2949 self_stack_size
= optimize
&& !node
->thunk
.thunk_p
2950 ? estimated_stack_frame_size (node
) : 0;
2951 info
->estimated_self_stack_size
= self_stack_size
;
2952 info
->estimated_stack_size
= self_stack_size
;
2953 info
->stack_frame_offset
= 0;
2955 if (node
->thunk
.thunk_p
)
2957 struct inline_edge_summary
*es
= inline_edge_summary (node
->callees
);
2958 struct predicate t
= true_predicate ();
2960 node
->local
.can_change_signature
= false;
2961 es
->call_stmt_size
= eni_size_weights
.call_cost
;
2962 es
->call_stmt_time
= eni_time_weights
.call_cost
;
2963 account_size_time (info
, INLINE_SIZE_SCALE
* 2,
2964 INLINE_TIME_SCALE
* 2, &t
);
2965 t
= not_inlined_predicate ();
2966 account_size_time (info
, 2 * INLINE_SIZE_SCALE
, 0, &t
);
2967 inline_update_overall_summary (node
);
2968 info
->self_size
= info
->size
;
2969 info
->self_time
= info
->time
;
2970 /* We can not inline instrumentation clones. */
2971 if (node
->thunk
.add_pointer_bounds_args
)
2973 info
->inlinable
= false;
2974 node
->callees
->inline_failed
= CIF_CHKP
;
2977 info
->inlinable
= true;
2981 /* Even is_gimple_min_invariant rely on current_function_decl. */
2982 push_cfun (DECL_STRUCT_FUNCTION (node
->decl
));
2984 /* Can this function be inlined at all? */
2985 if (!opt_for_fn (node
->decl
, optimize
)
2986 && !lookup_attribute ("always_inline",
2987 DECL_ATTRIBUTES (node
->decl
)))
2988 info
->inlinable
= false;
2990 info
->inlinable
= tree_inlinable_function_p (node
->decl
);
2992 info
->contains_cilk_spawn
= fn_contains_cilk_spawn_p (cfun
);
2994 /* Type attributes can use parameter indices to describe them. */
2995 if (TYPE_ATTRIBUTES (TREE_TYPE (node
->decl
)))
2996 node
->local
.can_change_signature
= false;
2999 /* Otherwise, inlinable functions always can change signature. */
3000 if (info
->inlinable
)
3001 node
->local
.can_change_signature
= true;
3004 /* Functions calling builtin_apply can not change signature. */
3005 for (e
= node
->callees
; e
; e
= e
->next_callee
)
3007 tree
cdecl = e
->callee
->decl
;
3008 if (DECL_BUILT_IN (cdecl)
3009 && DECL_BUILT_IN_CLASS (cdecl) == BUILT_IN_NORMAL
3010 && (DECL_FUNCTION_CODE (cdecl) == BUILT_IN_APPLY_ARGS
3011 || DECL_FUNCTION_CODE (cdecl) == BUILT_IN_VA_START
))
3014 node
->local
.can_change_signature
= !e
;
3017 /* Functions called by instrumentation thunk can't change signature
3018 because instrumentation thunk modification is not supported. */
3019 if (node
->local
.can_change_signature
)
3020 for (e
= node
->callers
; e
; e
= e
->next_caller
)
3021 if (e
->caller
->thunk
.thunk_p
3022 && e
->caller
->thunk
.add_pointer_bounds_args
)
3024 node
->local
.can_change_signature
= false;
3027 estimate_function_body_sizes (node
, early
);
3030 for (e
= node
->callees
; e
; e
= e
->next_callee
)
3031 if (e
->callee
->comdat_local_p ())
3033 node
->calls_comdat_local
= (e
!= NULL
);
3035 /* Inlining characteristics are maintained by the cgraph_mark_inline. */
3036 info
->time
= info
->self_time
;
3037 info
->size
= info
->self_size
;
3038 info
->stack_frame_offset
= 0;
3039 info
->estimated_stack_size
= info
->estimated_self_stack_size
;
3042 inline_update_overall_summary (node
);
3043 gcc_assert (info
->time
== info
->self_time
3044 && info
->size
== info
->self_size
);
3049 /* Compute parameters of functions used by inliner using
3050 current_function_decl. */
3053 compute_inline_parameters_for_current (void)
3055 compute_inline_parameters (cgraph_node::get (current_function_decl
), true);
3061 const pass_data pass_data_inline_parameters
=
3063 GIMPLE_PASS
, /* type */
3064 "inline_param", /* name */
3065 OPTGROUP_INLINE
, /* optinfo_flags */
3066 TV_INLINE_PARAMETERS
, /* tv_id */
3067 0, /* properties_required */
3068 0, /* properties_provided */
3069 0, /* properties_destroyed */
3070 0, /* todo_flags_start */
3071 0, /* todo_flags_finish */
3074 class pass_inline_parameters
: public gimple_opt_pass
3077 pass_inline_parameters (gcc::context
*ctxt
)
3078 : gimple_opt_pass (pass_data_inline_parameters
, ctxt
)
3081 /* opt_pass methods: */
3082 opt_pass
* clone () { return new pass_inline_parameters (m_ctxt
); }
3083 virtual unsigned int execute (function
*)
3085 return compute_inline_parameters_for_current ();
3088 }; // class pass_inline_parameters
3093 make_pass_inline_parameters (gcc::context
*ctxt
)
3095 return new pass_inline_parameters (ctxt
);
3099 /* Estimate benefit devirtualizing indirect edge IE, provided KNOWN_VALS,
3100 KNOWN_CONTEXTS and KNOWN_AGGS. */
3103 estimate_edge_devirt_benefit (struct cgraph_edge
*ie
,
3104 int *size
, int *time
,
3105 vec
<tree
> known_vals
,
3106 vec
<ipa_polymorphic_call_context
> known_contexts
,
3107 vec
<ipa_agg_jump_function_p
> known_aggs
)
3110 struct cgraph_node
*callee
;
3111 struct inline_summary
*isummary
;
3112 enum availability avail
;
3115 if (!known_vals
.exists () && !known_contexts
.exists ())
3117 if (!opt_for_fn (ie
->caller
->decl
, flag_indirect_inlining
))
3120 target
= ipa_get_indirect_edge_target (ie
, known_vals
, known_contexts
,
3121 known_aggs
, &speculative
);
3122 if (!target
|| speculative
)
3125 /* Account for difference in cost between indirect and direct calls. */
3126 *size
-= (eni_size_weights
.indirect_call_cost
- eni_size_weights
.call_cost
);
3127 *time
-= (eni_time_weights
.indirect_call_cost
- eni_time_weights
.call_cost
);
3128 gcc_checking_assert (*time
>= 0);
3129 gcc_checking_assert (*size
>= 0);
3131 callee
= cgraph_node::get (target
);
3132 if (!callee
|| !callee
->definition
)
3134 callee
= callee
->function_symbol (&avail
);
3135 if (avail
< AVAIL_AVAILABLE
)
3137 isummary
= inline_summaries
->get (callee
);
3138 return isummary
->inlinable
;
3141 /* Increase SIZE, MIN_SIZE (if non-NULL) and TIME for size and time needed to
3142 handle edge E with probability PROB.
3143 Set HINTS if edge may be devirtualized.
3144 KNOWN_VALS, KNOWN_AGGS and KNOWN_CONTEXTS describe context of the call
3148 estimate_edge_size_and_time (struct cgraph_edge
*e
, int *size
, int *min_size
,
3151 vec
<tree
> known_vals
,
3152 vec
<ipa_polymorphic_call_context
> known_contexts
,
3153 vec
<ipa_agg_jump_function_p
> known_aggs
,
3154 inline_hints
*hints
)
3156 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3157 int call_size
= es
->call_stmt_size
;
3158 int call_time
= es
->call_stmt_time
;
3161 && estimate_edge_devirt_benefit (e
, &call_size
, &call_time
,
3162 known_vals
, known_contexts
, known_aggs
)
3163 && hints
&& e
->maybe_hot_p ())
3164 *hints
|= INLINE_HINT_indirect_call
;
3165 cur_size
= call_size
* INLINE_SIZE_SCALE
;
3168 *min_size
+= cur_size
;
3169 *time
+= apply_probability ((gcov_type
) call_time
, prob
)
3170 * e
->frequency
* (INLINE_TIME_SCALE
/ CGRAPH_FREQ_BASE
);
3171 if (*time
> MAX_TIME
* INLINE_TIME_SCALE
)
3172 *time
= MAX_TIME
* INLINE_TIME_SCALE
;
3177 /* Increase SIZE, MIN_SIZE and TIME for size and time needed to handle all
3178 calls in NODE. POSSIBLE_TRUTHS, KNOWN_VALS, KNOWN_AGGS and KNOWN_CONTEXTS
3179 describe context of the call site. */
3182 estimate_calls_size_and_time (struct cgraph_node
*node
, int *size
,
3183 int *min_size
, int *time
,
3184 inline_hints
*hints
,
3185 clause_t possible_truths
,
3186 vec
<tree
> known_vals
,
3187 vec
<ipa_polymorphic_call_context
> known_contexts
,
3188 vec
<ipa_agg_jump_function_p
> known_aggs
)
3190 struct cgraph_edge
*e
;
3191 for (e
= node
->callees
; e
; e
= e
->next_callee
)
3193 if (inline_edge_summary_vec
.length () <= (unsigned) e
->uid
)
3196 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3198 /* Do not care about zero sized builtins. */
3199 if (e
->inline_failed
&& !es
->call_stmt_size
)
3201 gcc_checking_assert (!es
->call_stmt_time
);
3205 || evaluate_predicate (es
->predicate
, possible_truths
))
3207 if (e
->inline_failed
)
3209 /* Predicates of calls shall not use NOT_CHANGED codes,
3210 sowe do not need to compute probabilities. */
3211 estimate_edge_size_and_time (e
, size
,
3212 es
->predicate
? NULL
: min_size
,
3213 time
, REG_BR_PROB_BASE
,
3214 known_vals
, known_contexts
,
3218 estimate_calls_size_and_time (e
->callee
, size
, min_size
, time
,
3221 known_vals
, known_contexts
,
3225 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
3227 if (inline_edge_summary_vec
.length () <= (unsigned) e
->uid
)
3230 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3232 || evaluate_predicate (es
->predicate
, possible_truths
))
3233 estimate_edge_size_and_time (e
, size
,
3234 es
->predicate
? NULL
: min_size
,
3235 time
, REG_BR_PROB_BASE
,
3236 known_vals
, known_contexts
, known_aggs
,
3242 /* Estimate size and time needed to execute NODE assuming
3243 POSSIBLE_TRUTHS clause, and KNOWN_VALS, KNOWN_AGGS and KNOWN_CONTEXTS
3244 information about NODE's arguments. If non-NULL use also probability
3245 information present in INLINE_PARAM_SUMMARY vector.
3246 Additionally detemine hints determined by the context. Finally compute
3247 minimal size needed for the call that is independent on the call context and
3248 can be used for fast estimates. Return the values in RET_SIZE,
3249 RET_MIN_SIZE, RET_TIME and RET_HINTS. */
3252 estimate_node_size_and_time (struct cgraph_node
*node
,
3253 clause_t possible_truths
,
3254 vec
<tree
> known_vals
,
3255 vec
<ipa_polymorphic_call_context
> known_contexts
,
3256 vec
<ipa_agg_jump_function_p
> known_aggs
,
3257 int *ret_size
, int *ret_min_size
, int *ret_time
,
3258 inline_hints
*ret_hints
,
3259 vec
<inline_param_summary
>
3260 inline_param_summary
)
3262 struct inline_summary
*info
= inline_summaries
->get (node
);
3267 inline_hints hints
= 0;
3270 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3273 fprintf (dump_file
, " Estimating body: %s/%i\n"
3274 " Known to be false: ", node
->name (),
3277 for (i
= predicate_not_inlined_condition
;
3278 i
< (predicate_first_dynamic_condition
3279 + (int) vec_safe_length (info
->conds
)); i
++)
3280 if (!(possible_truths
& (1 << i
)))
3283 fprintf (dump_file
, ", ");
3285 dump_condition (dump_file
, info
->conds
, i
);
3289 for (i
= 0; vec_safe_iterate (info
->entry
, i
, &e
); i
++)
3290 if (evaluate_predicate (&e
->predicate
, possible_truths
))
3293 gcc_checking_assert (e
->time
>= 0);
3294 gcc_checking_assert (time
>= 0);
3295 if (!inline_param_summary
.exists ())
3299 int prob
= predicate_probability (info
->conds
,
3302 inline_param_summary
);
3303 gcc_checking_assert (prob
>= 0);
3304 gcc_checking_assert (prob
<= REG_BR_PROB_BASE
);
3305 time
+= apply_probability ((gcov_type
) e
->time
, prob
);
3307 if (time
> MAX_TIME
* INLINE_TIME_SCALE
)
3308 time
= MAX_TIME
* INLINE_TIME_SCALE
;
3309 gcc_checking_assert (time
>= 0);
3312 gcc_checking_assert (true_predicate_p (&(*info
->entry
)[0].predicate
));
3313 min_size
= (*info
->entry
)[0].size
;
3314 gcc_checking_assert (size
>= 0);
3315 gcc_checking_assert (time
>= 0);
3317 if (info
->loop_iterations
3318 && !evaluate_predicate (info
->loop_iterations
, possible_truths
))
3319 hints
|= INLINE_HINT_loop_iterations
;
3320 if (info
->loop_stride
3321 && !evaluate_predicate (info
->loop_stride
, possible_truths
))
3322 hints
|= INLINE_HINT_loop_stride
;
3323 if (info
->array_index
3324 && !evaluate_predicate (info
->array_index
, possible_truths
))
3325 hints
|= INLINE_HINT_array_index
;
3327 hints
|= INLINE_HINT_in_scc
;
3328 if (DECL_DECLARED_INLINE_P (node
->decl
))
3329 hints
|= INLINE_HINT_declared_inline
;
3331 estimate_calls_size_and_time (node
, &size
, &min_size
, &time
, &hints
, possible_truths
,
3332 known_vals
, known_contexts
, known_aggs
);
3333 gcc_checking_assert (size
>= 0);
3334 gcc_checking_assert (time
>= 0);
3335 time
= RDIV (time
, INLINE_TIME_SCALE
);
3336 size
= RDIV (size
, INLINE_SIZE_SCALE
);
3337 min_size
= RDIV (min_size
, INLINE_SIZE_SCALE
);
3339 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3340 fprintf (dump_file
, "\n size:%i time:%i\n", (int) size
, (int) time
);
3346 *ret_min_size
= min_size
;
3353 /* Estimate size and time needed to execute callee of EDGE assuming that
3354 parameters known to be constant at caller of EDGE are propagated.
3355 KNOWN_VALS and KNOWN_CONTEXTS are vectors of assumed known constant values
3356 and types for parameters. */
3359 estimate_ipcp_clone_size_and_time (struct cgraph_node
*node
,
3360 vec
<tree
> known_vals
,
3361 vec
<ipa_polymorphic_call_context
>
3363 vec
<ipa_agg_jump_function_p
> known_aggs
,
3364 int *ret_size
, int *ret_time
,
3365 inline_hints
*hints
)
3369 clause
= evaluate_conditions_for_known_args (node
, false, known_vals
,
3371 estimate_node_size_and_time (node
, clause
, known_vals
, known_contexts
,
3372 known_aggs
, ret_size
, NULL
, ret_time
, hints
, vNULL
);
3375 /* Translate all conditions from callee representation into caller
3376 representation and symbolically evaluate predicate P into new predicate.
3378 INFO is inline_summary of function we are adding predicate into, CALLEE_INFO
3379 is summary of function predicate P is from. OPERAND_MAP is array giving
3380 callee formal IDs the caller formal IDs. POSSSIBLE_TRUTHS is clausule of all
3381 callee conditions that may be true in caller context. TOPLEV_PREDICATE is
3382 predicate under which callee is executed. OFFSET_MAP is an array of of
3383 offsets that need to be added to conditions, negative offset means that
3384 conditions relying on values passed by reference have to be discarded
3385 because they might not be preserved (and should be considered offset zero
3386 for other purposes). */
3388 static struct predicate
3389 remap_predicate (struct inline_summary
*info
,
3390 struct inline_summary
*callee_info
,
3391 struct predicate
*p
,
3392 vec
<int> operand_map
,
3393 vec
<int> offset_map
,
3394 clause_t possible_truths
, struct predicate
*toplev_predicate
)
3397 struct predicate out
= true_predicate ();
3399 /* True predicate is easy. */
3400 if (true_predicate_p (p
))
3401 return *toplev_predicate
;
3402 for (i
= 0; p
->clause
[i
]; i
++)
3404 clause_t clause
= p
->clause
[i
];
3406 struct predicate clause_predicate
= false_predicate ();
3408 gcc_assert (i
< MAX_CLAUSES
);
3410 for (cond
= 0; cond
< NUM_CONDITIONS
; cond
++)
3411 /* Do we have condition we can't disprove? */
3412 if (clause
& possible_truths
& (1 << cond
))
3414 struct predicate cond_predicate
;
3415 /* Work out if the condition can translate to predicate in the
3416 inlined function. */
3417 if (cond
>= predicate_first_dynamic_condition
)
3419 struct condition
*c
;
3421 c
= &(*callee_info
->conds
)[cond
3423 predicate_first_dynamic_condition
];
3424 /* See if we can remap condition operand to caller's operand.
3425 Otherwise give up. */
3426 if (!operand_map
.exists ()
3427 || (int) operand_map
.length () <= c
->operand_num
3428 || operand_map
[c
->operand_num
] == -1
3429 /* TODO: For non-aggregate conditions, adding an offset is
3430 basically an arithmetic jump function processing which
3431 we should support in future. */
3432 || ((!c
->agg_contents
|| !c
->by_ref
)
3433 && offset_map
[c
->operand_num
] > 0)
3434 || (c
->agg_contents
&& c
->by_ref
3435 && offset_map
[c
->operand_num
] < 0))
3436 cond_predicate
= true_predicate ();
3439 struct agg_position_info ap
;
3440 HOST_WIDE_INT offset_delta
= offset_map
[c
->operand_num
];
3441 if (offset_delta
< 0)
3443 gcc_checking_assert (!c
->agg_contents
|| !c
->by_ref
);
3446 gcc_assert (!c
->agg_contents
3447 || c
->by_ref
|| offset_delta
== 0);
3448 ap
.offset
= c
->offset
+ offset_delta
;
3449 ap
.agg_contents
= c
->agg_contents
;
3450 ap
.by_ref
= c
->by_ref
;
3451 cond_predicate
= add_condition (info
,
3452 operand_map
[c
->operand_num
],
3453 c
->size
, &ap
, c
->code
,
3457 /* Fixed conditions remains same, construct single
3458 condition predicate. */
3461 cond_predicate
.clause
[0] = 1 << cond
;
3462 cond_predicate
.clause
[1] = 0;
3464 clause_predicate
= or_predicates (info
->conds
, &clause_predicate
,
3467 out
= and_predicates (info
->conds
, &out
, &clause_predicate
);
3469 return and_predicates (info
->conds
, &out
, toplev_predicate
);
3473 /* Update summary information of inline clones after inlining.
3474 Compute peak stack usage. */
3477 inline_update_callee_summaries (struct cgraph_node
*node
, int depth
)
3479 struct cgraph_edge
*e
;
3480 struct inline_summary
*callee_info
= inline_summaries
->get (node
);
3481 struct inline_summary
*caller_info
= inline_summaries
->get (node
->callers
->caller
);
3484 callee_info
->stack_frame_offset
3485 = caller_info
->stack_frame_offset
3486 + caller_info
->estimated_self_stack_size
;
3487 peak
= callee_info
->stack_frame_offset
3488 + callee_info
->estimated_self_stack_size
;
3489 if (inline_summaries
->get (node
->global
.inlined_to
)->estimated_stack_size
< peak
)
3490 inline_summaries
->get (node
->global
.inlined_to
)->estimated_stack_size
= peak
;
3491 ipa_propagate_frequency (node
);
3492 for (e
= node
->callees
; e
; e
= e
->next_callee
)
3494 if (!e
->inline_failed
)
3495 inline_update_callee_summaries (e
->callee
, depth
);
3496 inline_edge_summary (e
)->loop_depth
+= depth
;
3498 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
3499 inline_edge_summary (e
)->loop_depth
+= depth
;
3502 /* Update change_prob of EDGE after INLINED_EDGE has been inlined.
3503 When functoin A is inlined in B and A calls C with parameter that
3504 changes with probability PROB1 and C is known to be passthroug
3505 of argument if B that change with probability PROB2, the probability
3506 of change is now PROB1*PROB2. */
3509 remap_edge_change_prob (struct cgraph_edge
*inlined_edge
,
3510 struct cgraph_edge
*edge
)
3512 if (ipa_node_params_sum
)
3515 struct ipa_edge_args
*args
= IPA_EDGE_REF (edge
);
3516 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3517 struct inline_edge_summary
*inlined_es
3518 = inline_edge_summary (inlined_edge
);
3520 for (i
= 0; i
< ipa_get_cs_argument_count (args
); i
++)
3522 struct ipa_jump_func
*jfunc
= ipa_get_ith_jump_func (args
, i
);
3523 if (jfunc
->type
== IPA_JF_PASS_THROUGH
3524 && (ipa_get_jf_pass_through_formal_id (jfunc
)
3525 < (int) inlined_es
->param
.length ()))
3527 int jf_formal_id
= ipa_get_jf_pass_through_formal_id (jfunc
);
3528 int prob1
= es
->param
[i
].change_prob
;
3529 int prob2
= inlined_es
->param
[jf_formal_id
].change_prob
;
3530 int prob
= combine_probabilities (prob1
, prob2
);
3532 if (prob1
&& prob2
&& !prob
)
3535 es
->param
[i
].change_prob
= prob
;
3541 /* Update edge summaries of NODE after INLINED_EDGE has been inlined.
3543 Remap predicates of callees of NODE. Rest of arguments match
3546 Also update change probabilities. */
3549 remap_edge_summaries (struct cgraph_edge
*inlined_edge
,
3550 struct cgraph_node
*node
,
3551 struct inline_summary
*info
,
3552 struct inline_summary
*callee_info
,
3553 vec
<int> operand_map
,
3554 vec
<int> offset_map
,
3555 clause_t possible_truths
,
3556 struct predicate
*toplev_predicate
)
3558 struct cgraph_edge
*e
, *next
;
3559 for (e
= node
->callees
; e
; e
= next
)
3561 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3563 next
= e
->next_callee
;
3565 if (e
->inline_failed
)
3567 remap_edge_change_prob (inlined_edge
, e
);
3571 p
= remap_predicate (info
, callee_info
,
3572 es
->predicate
, operand_map
, offset_map
,
3573 possible_truths
, toplev_predicate
);
3574 edge_set_predicate (e
, &p
);
3577 edge_set_predicate (e
, toplev_predicate
);
3580 remap_edge_summaries (inlined_edge
, e
->callee
, info
, callee_info
,
3581 operand_map
, offset_map
, possible_truths
,
3584 for (e
= node
->indirect_calls
; e
; e
= next
)
3586 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3588 next
= e
->next_callee
;
3590 remap_edge_change_prob (inlined_edge
, e
);
3593 p
= remap_predicate (info
, callee_info
,
3594 es
->predicate
, operand_map
, offset_map
,
3595 possible_truths
, toplev_predicate
);
3596 edge_set_predicate (e
, &p
);
3599 edge_set_predicate (e
, toplev_predicate
);
3603 /* Same as remap_predicate, but set result into hint *HINT. */
3606 remap_hint_predicate (struct inline_summary
*info
,
3607 struct inline_summary
*callee_info
,
3608 struct predicate
**hint
,
3609 vec
<int> operand_map
,
3610 vec
<int> offset_map
,
3611 clause_t possible_truths
,
3612 struct predicate
*toplev_predicate
)
3618 p
= remap_predicate (info
, callee_info
,
3620 operand_map
, offset_map
,
3621 possible_truths
, toplev_predicate
);
3622 if (!false_predicate_p (&p
) && !true_predicate_p (&p
))
3625 set_hint_predicate (hint
, p
);
3627 **hint
= and_predicates (info
->conds
, *hint
, &p
);
3631 /* We inlined EDGE. Update summary of the function we inlined into. */
3634 inline_merge_summary (struct cgraph_edge
*edge
)
3636 struct inline_summary
*callee_info
= inline_summaries
->get (edge
->callee
);
3637 struct cgraph_node
*to
= (edge
->caller
->global
.inlined_to
3638 ? edge
->caller
->global
.inlined_to
: edge
->caller
);
3639 struct inline_summary
*info
= inline_summaries
->get (to
);
3640 clause_t clause
= 0; /* not_inline is known to be false. */
3642 vec
<int> operand_map
= vNULL
;
3643 vec
<int> offset_map
= vNULL
;
3645 struct predicate toplev_predicate
;
3646 struct predicate true_p
= true_predicate ();
3647 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3650 toplev_predicate
= *es
->predicate
;
3652 toplev_predicate
= true_predicate ();
3654 info
->fp_expressions
|= callee_info
->fp_expressions
;
3656 if (callee_info
->conds
)
3657 evaluate_properties_for_edge (edge
, true, &clause
, NULL
, NULL
, NULL
);
3658 if (ipa_node_params_sum
&& callee_info
->conds
)
3660 struct ipa_edge_args
*args
= IPA_EDGE_REF (edge
);
3661 int count
= ipa_get_cs_argument_count (args
);
3666 operand_map
.safe_grow_cleared (count
);
3667 offset_map
.safe_grow_cleared (count
);
3669 for (i
= 0; i
< count
; i
++)
3671 struct ipa_jump_func
*jfunc
= ipa_get_ith_jump_func (args
, i
);
3674 /* TODO: handle non-NOPs when merging. */
3675 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
3677 if (ipa_get_jf_pass_through_operation (jfunc
) == NOP_EXPR
)
3678 map
= ipa_get_jf_pass_through_formal_id (jfunc
);
3679 if (!ipa_get_jf_pass_through_agg_preserved (jfunc
))
3682 else if (jfunc
->type
== IPA_JF_ANCESTOR
)
3684 HOST_WIDE_INT offset
= ipa_get_jf_ancestor_offset (jfunc
);
3685 if (offset
>= 0 && offset
< INT_MAX
)
3687 map
= ipa_get_jf_ancestor_formal_id (jfunc
);
3688 if (!ipa_get_jf_ancestor_agg_preserved (jfunc
))
3690 offset_map
[i
] = offset
;
3693 operand_map
[i
] = map
;
3694 gcc_assert (map
< ipa_get_param_count (IPA_NODE_REF (to
)));
3697 for (i
= 0; vec_safe_iterate (callee_info
->entry
, i
, &e
); i
++)
3699 struct predicate p
= remap_predicate (info
, callee_info
,
3700 &e
->predicate
, operand_map
,
3703 if (!false_predicate_p (&p
))
3705 gcov_type add_time
= ((gcov_type
) e
->time
* edge
->frequency
3706 + CGRAPH_FREQ_BASE
/ 2) / CGRAPH_FREQ_BASE
;
3707 int prob
= predicate_probability (callee_info
->conds
,
3710 add_time
= apply_probability ((gcov_type
) add_time
, prob
);
3711 if (add_time
> MAX_TIME
* INLINE_TIME_SCALE
)
3712 add_time
= MAX_TIME
* INLINE_TIME_SCALE
;
3713 if (prob
!= REG_BR_PROB_BASE
3714 && dump_file
&& (dump_flags
& TDF_DETAILS
))
3716 fprintf (dump_file
, "\t\tScaling time by probability:%f\n",
3717 (double) prob
/ REG_BR_PROB_BASE
);
3719 account_size_time (info
, e
->size
, add_time
, &p
);
3722 remap_edge_summaries (edge
, edge
->callee
, info
, callee_info
, operand_map
,
3723 offset_map
, clause
, &toplev_predicate
);
3724 remap_hint_predicate (info
, callee_info
,
3725 &callee_info
->loop_iterations
,
3726 operand_map
, offset_map
, clause
, &toplev_predicate
);
3727 remap_hint_predicate (info
, callee_info
,
3728 &callee_info
->loop_stride
,
3729 operand_map
, offset_map
, clause
, &toplev_predicate
);
3730 remap_hint_predicate (info
, callee_info
,
3731 &callee_info
->array_index
,
3732 operand_map
, offset_map
, clause
, &toplev_predicate
);
3734 inline_update_callee_summaries (edge
->callee
,
3735 inline_edge_summary (edge
)->loop_depth
);
3737 /* We do not maintain predicates of inlined edges, free it. */
3738 edge_set_predicate (edge
, &true_p
);
3739 /* Similarly remove param summaries. */
3740 es
->param
.release ();
3741 operand_map
.release ();
3742 offset_map
.release ();
3745 /* For performance reasons inline_merge_summary is not updating overall size
3746 and time. Recompute it. */
3749 inline_update_overall_summary (struct cgraph_node
*node
)
3751 struct inline_summary
*info
= inline_summaries
->get (node
);
3757 for (i
= 0; vec_safe_iterate (info
->entry
, i
, &e
); i
++)
3759 info
->size
+= e
->size
, info
->time
+= e
->time
;
3760 if (info
->time
> MAX_TIME
* INLINE_TIME_SCALE
)
3761 info
->time
= MAX_TIME
* INLINE_TIME_SCALE
;
3763 estimate_calls_size_and_time (node
, &info
->size
, &info
->min_size
,
3765 ~(clause_t
) (1 << predicate_false_condition
),
3766 vNULL
, vNULL
, vNULL
);
3767 info
->time
= (info
->time
+ INLINE_TIME_SCALE
/ 2) / INLINE_TIME_SCALE
;
3768 info
->size
= (info
->size
+ INLINE_SIZE_SCALE
/ 2) / INLINE_SIZE_SCALE
;
3771 /* Return hints derrived from EDGE. */
3773 simple_edge_hints (struct cgraph_edge
*edge
)
3776 struct cgraph_node
*to
= (edge
->caller
->global
.inlined_to
3777 ? edge
->caller
->global
.inlined_to
: edge
->caller
);
3778 struct cgraph_node
*callee
= edge
->callee
->ultimate_alias_target ();
3779 if (inline_summaries
->get (to
)->scc_no
3780 && inline_summaries
->get (to
)->scc_no
3781 == inline_summaries
->get (callee
)->scc_no
3782 && !edge
->recursive_p ())
3783 hints
|= INLINE_HINT_same_scc
;
3785 if (callee
->lto_file_data
&& edge
->caller
->lto_file_data
3786 && edge
->caller
->lto_file_data
!= callee
->lto_file_data
3787 && !callee
->merged_comdat
&& !callee
->icf_merged
)
3788 hints
|= INLINE_HINT_cross_module
;
3793 /* Estimate the time cost for the caller when inlining EDGE.
3794 Only to be called via estimate_edge_time, that handles the
3797 When caching, also update the cache entry. Compute both time and
3798 size, since we always need both metrics eventually. */
3801 do_estimate_edge_time (struct cgraph_edge
*edge
)
3806 struct cgraph_node
*callee
;
3808 vec
<tree
> known_vals
;
3809 vec
<ipa_polymorphic_call_context
> known_contexts
;
3810 vec
<ipa_agg_jump_function_p
> known_aggs
;
3811 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3814 callee
= edge
->callee
->ultimate_alias_target ();
3816 gcc_checking_assert (edge
->inline_failed
);
3817 evaluate_properties_for_edge (edge
, true,
3818 &clause
, &known_vals
, &known_contexts
,
3820 estimate_node_size_and_time (callee
, clause
, known_vals
, known_contexts
,
3821 known_aggs
, &size
, &min_size
, &time
, &hints
, es
->param
);
3823 /* When we have profile feedback, we can quite safely identify hot
3824 edges and for those we disable size limits. Don't do that when
3825 probability that caller will call the callee is low however, since it
3826 may hurt optimization of the caller's hot path. */
3827 if (edge
->count
&& edge
->maybe_hot_p ()
3829 > (edge
->caller
->global
.inlined_to
3830 ? edge
->caller
->global
.inlined_to
->count
: edge
->caller
->count
)))
3831 hints
|= INLINE_HINT_known_hot
;
3833 known_vals
.release ();
3834 known_contexts
.release ();
3835 known_aggs
.release ();
3836 gcc_checking_assert (size
>= 0);
3837 gcc_checking_assert (time
>= 0);
3839 /* When caching, update the cache entry. */
3840 if (edge_growth_cache
.exists ())
3842 inline_summaries
->get (edge
->callee
)->min_size
= min_size
;
3843 if ((int) edge_growth_cache
.length () <= edge
->uid
)
3844 edge_growth_cache
.safe_grow_cleared (symtab
->edges_max_uid
);
3845 edge_growth_cache
[edge
->uid
].time
= time
+ (time
>= 0);
3847 edge_growth_cache
[edge
->uid
].size
= size
+ (size
>= 0);
3848 hints
|= simple_edge_hints (edge
);
3849 edge_growth_cache
[edge
->uid
].hints
= hints
+ 1;
3855 /* Return estimated callee growth after inlining EDGE.
3856 Only to be called via estimate_edge_size. */
3859 do_estimate_edge_size (struct cgraph_edge
*edge
)
3862 struct cgraph_node
*callee
;
3864 vec
<tree
> known_vals
;
3865 vec
<ipa_polymorphic_call_context
> known_contexts
;
3866 vec
<ipa_agg_jump_function_p
> known_aggs
;
3868 /* When we do caching, use do_estimate_edge_time to populate the entry. */
3870 if (edge_growth_cache
.exists ())
3872 do_estimate_edge_time (edge
);
3873 size
= edge_growth_cache
[edge
->uid
].size
;
3874 gcc_checking_assert (size
);
3875 return size
- (size
> 0);
3878 callee
= edge
->callee
->ultimate_alias_target ();
3880 /* Early inliner runs without caching, go ahead and do the dirty work. */
3881 gcc_checking_assert (edge
->inline_failed
);
3882 evaluate_properties_for_edge (edge
, true,
3883 &clause
, &known_vals
, &known_contexts
,
3885 estimate_node_size_and_time (callee
, clause
, known_vals
, known_contexts
,
3886 known_aggs
, &size
, NULL
, NULL
, NULL
, vNULL
);
3887 known_vals
.release ();
3888 known_contexts
.release ();
3889 known_aggs
.release ();
3894 /* Estimate the growth of the caller when inlining EDGE.
3895 Only to be called via estimate_edge_size. */
3898 do_estimate_edge_hints (struct cgraph_edge
*edge
)
3901 struct cgraph_node
*callee
;
3903 vec
<tree
> known_vals
;
3904 vec
<ipa_polymorphic_call_context
> known_contexts
;
3905 vec
<ipa_agg_jump_function_p
> known_aggs
;
3907 /* When we do caching, use do_estimate_edge_time to populate the entry. */
3909 if (edge_growth_cache
.exists ())
3911 do_estimate_edge_time (edge
);
3912 hints
= edge_growth_cache
[edge
->uid
].hints
;
3913 gcc_checking_assert (hints
);
3917 callee
= edge
->callee
->ultimate_alias_target ();
3919 /* Early inliner runs without caching, go ahead and do the dirty work. */
3920 gcc_checking_assert (edge
->inline_failed
);
3921 evaluate_properties_for_edge (edge
, true,
3922 &clause
, &known_vals
, &known_contexts
,
3924 estimate_node_size_and_time (callee
, clause
, known_vals
, known_contexts
,
3925 known_aggs
, NULL
, NULL
, NULL
, &hints
, vNULL
);
3926 known_vals
.release ();
3927 known_contexts
.release ();
3928 known_aggs
.release ();
3929 hints
|= simple_edge_hints (edge
);
3934 /* Estimate self time of the function NODE after inlining EDGE. */
3937 estimate_time_after_inlining (struct cgraph_node
*node
,
3938 struct cgraph_edge
*edge
)
3940 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3941 if (!es
->predicate
|| !false_predicate_p (es
->predicate
))
3944 inline_summaries
->get (node
)->time
+ estimate_edge_time (edge
);
3947 if (time
> MAX_TIME
)
3951 return inline_summaries
->get (node
)->time
;
3955 /* Estimate the size of NODE after inlining EDGE which should be an
3956 edge to either NODE or a call inlined into NODE. */
3959 estimate_size_after_inlining (struct cgraph_node
*node
,
3960 struct cgraph_edge
*edge
)
3962 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3963 if (!es
->predicate
|| !false_predicate_p (es
->predicate
))
3965 int size
= inline_summaries
->get (node
)->size
+ estimate_edge_growth (edge
);
3966 gcc_assert (size
>= 0);
3969 return inline_summaries
->get (node
)->size
;
3975 struct cgraph_node
*node
;
3976 bool self_recursive
;
3982 /* Worker for do_estimate_growth. Collect growth for all callers. */
3985 do_estimate_growth_1 (struct cgraph_node
*node
, void *data
)
3987 struct cgraph_edge
*e
;
3988 struct growth_data
*d
= (struct growth_data
*) data
;
3990 for (e
= node
->callers
; e
; e
= e
->next_caller
)
3992 gcc_checking_assert (e
->inline_failed
);
3994 if (cgraph_inline_failed_type (e
->inline_failed
) == CIF_FINAL_ERROR
)
3996 d
->uninlinable
= true;
4000 if (e
->recursive_p ())
4002 d
->self_recursive
= true;
4005 d
->growth
+= estimate_edge_growth (e
);
4011 /* Estimate the growth caused by inlining NODE into all callees. */
4014 estimate_growth (struct cgraph_node
*node
)
4016 struct growth_data d
= { node
, false, false, 0 };
4017 struct inline_summary
*info
= inline_summaries
->get (node
);
4019 node
->call_for_symbol_and_aliases (do_estimate_growth_1
, &d
, true);
4021 /* For self recursive functions the growth estimation really should be
4022 infinity. We don't want to return very large values because the growth
4023 plays various roles in badness computation fractions. Be sure to not
4024 return zero or negative growths. */
4025 if (d
.self_recursive
)
4026 d
.growth
= d
.growth
< info
->size
? info
->size
: d
.growth
;
4027 else if (DECL_EXTERNAL (node
->decl
) || d
.uninlinable
)
4031 if (node
->will_be_removed_from_program_if_no_direct_calls_p ())
4032 d
.growth
-= info
->size
;
4033 /* COMDAT functions are very often not shared across multiple units
4034 since they come from various template instantiations.
4035 Take this into account. */
4036 else if (DECL_COMDAT (node
->decl
)
4037 && node
->can_remove_if_no_direct_calls_p ())
4038 d
.growth
-= (info
->size
4039 * (100 - PARAM_VALUE (PARAM_COMDAT_SHARING_PROBABILITY
))
4046 /* Verify if there are fewer than MAX_CALLERS. */
4049 check_callers (cgraph_node
*node
, int *max_callers
)
4053 if (!node
->can_remove_if_no_direct_calls_and_refs_p ())
4056 for (cgraph_edge
*e
= node
->callers
; e
; e
= e
->next_caller
)
4060 || cgraph_inline_failed_type (e
->inline_failed
) == CIF_FINAL_ERROR
)
4064 FOR_EACH_ALIAS (node
, ref
)
4065 if (check_callers (dyn_cast
<cgraph_node
*> (ref
->referring
), max_callers
))
4072 /* Make cheap estimation if growth of NODE is likely positive knowing
4073 EDGE_GROWTH of one particular edge.
4074 We assume that most of other edges will have similar growth
4075 and skip computation if there are too many callers. */
4078 growth_likely_positive (struct cgraph_node
*node
,
4082 struct cgraph_edge
*e
;
4083 gcc_checking_assert (edge_growth
> 0);
4085 /* First quickly check if NODE is removable at all. */
4086 if (DECL_EXTERNAL (node
->decl
))
4088 if (!node
->can_remove_if_no_direct_calls_and_refs_p ()
4089 || node
->address_taken
)
4092 max_callers
= inline_summaries
->get (node
)->size
* 4 / edge_growth
+ 2;
4094 for (e
= node
->callers
; e
; e
= e
->next_caller
)
4098 || cgraph_inline_failed_type (e
->inline_failed
) == CIF_FINAL_ERROR
)
4103 FOR_EACH_ALIAS (node
, ref
)
4104 if (check_callers (dyn_cast
<cgraph_node
*> (ref
->referring
), &max_callers
))
4107 /* Unlike for functions called once, we play unsafe with
4108 COMDATs. We can allow that since we know functions
4109 in consideration are small (and thus risk is small) and
4110 moreover grow estimates already accounts that COMDAT
4111 functions may or may not disappear when eliminated from
4112 current unit. With good probability making aggressive
4113 choice in all units is going to make overall program
4115 if (DECL_COMDAT (node
->decl
))
4117 if (!node
->can_remove_if_no_direct_calls_p ())
4120 else if (!node
->will_be_removed_from_program_if_no_direct_calls_p ())
4123 return estimate_growth (node
) > 0;
4127 /* This function performs intraprocedural analysis in NODE that is required to
4128 inline indirect calls. */
4131 inline_indirect_intraprocedural_analysis (struct cgraph_node
*node
)
4133 ipa_analyze_node (node
);
4134 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4136 ipa_print_node_params (dump_file
, node
);
4137 ipa_print_node_jump_functions (dump_file
, node
);
4142 /* Note function body size. */
4145 inline_analyze_function (struct cgraph_node
*node
)
4147 push_cfun (DECL_STRUCT_FUNCTION (node
->decl
));
4150 fprintf (dump_file
, "\nAnalyzing function: %s/%u\n",
4151 node
->name (), node
->order
);
4152 if (opt_for_fn (node
->decl
, optimize
) && !node
->thunk
.thunk_p
)
4153 inline_indirect_intraprocedural_analysis (node
);
4154 compute_inline_parameters (node
, false);
4157 struct cgraph_edge
*e
;
4158 for (e
= node
->callees
; e
; e
= e
->next_callee
)
4159 e
->inline_failed
= CIF_FUNCTION_NOT_OPTIMIZED
;
4160 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
4161 e
->inline_failed
= CIF_FUNCTION_NOT_OPTIMIZED
;
4168 /* Called when new function is inserted to callgraph late. */
4171 inline_summary_t::insert (struct cgraph_node
*node
, inline_summary
*)
4173 inline_analyze_function (node
);
4176 /* Note function body size. */
4179 inline_generate_summary (void)
4181 struct cgraph_node
*node
;
4183 FOR_EACH_DEFINED_FUNCTION (node
)
4184 if (DECL_STRUCT_FUNCTION (node
->decl
))
4185 node
->local
.versionable
= tree_versionable_function_p (node
->decl
);
4187 /* When not optimizing, do not bother to analyze. Inlining is still done
4188 because edge redirection needs to happen there. */
4189 if (!optimize
&& !flag_generate_lto
&& !flag_generate_offload
&& !flag_wpa
)
4192 if (!inline_summaries
)
4193 inline_summaries
= (inline_summary_t
*) inline_summary_t::create_ggc (symtab
);
4195 inline_summaries
->enable_insertion_hook ();
4197 ipa_register_cgraph_hooks ();
4198 inline_free_summary ();
4200 FOR_EACH_DEFINED_FUNCTION (node
)
4202 inline_analyze_function (node
);
4206 /* Read predicate from IB. */
4208 static struct predicate
4209 read_predicate (struct lto_input_block
*ib
)
4211 struct predicate out
;
4217 gcc_assert (k
<= MAX_CLAUSES
);
4218 clause
= out
.clause
[k
++] = streamer_read_uhwi (ib
);
4222 /* Zero-initialize the remaining clauses in OUT. */
4223 while (k
<= MAX_CLAUSES
)
4224 out
.clause
[k
++] = 0;
4230 /* Write inline summary for edge E to OB. */
4233 read_inline_edge_summary (struct lto_input_block
*ib
, struct cgraph_edge
*e
)
4235 struct inline_edge_summary
*es
= inline_edge_summary (e
);
4239 es
->call_stmt_size
= streamer_read_uhwi (ib
);
4240 es
->call_stmt_time
= streamer_read_uhwi (ib
);
4241 es
->loop_depth
= streamer_read_uhwi (ib
);
4242 p
= read_predicate (ib
);
4243 edge_set_predicate (e
, &p
);
4244 length
= streamer_read_uhwi (ib
);
4247 es
->param
.safe_grow_cleared (length
);
4248 for (i
= 0; i
< length
; i
++)
4249 es
->param
[i
].change_prob
= streamer_read_uhwi (ib
);
4254 /* Stream in inline summaries from the section. */
4257 inline_read_section (struct lto_file_decl_data
*file_data
, const char *data
,
4260 const struct lto_function_header
*header
=
4261 (const struct lto_function_header
*) data
;
4262 const int cfg_offset
= sizeof (struct lto_function_header
);
4263 const int main_offset
= cfg_offset
+ header
->cfg_size
;
4264 const int string_offset
= main_offset
+ header
->main_size
;
4265 struct data_in
*data_in
;
4266 unsigned int i
, count2
, j
;
4267 unsigned int f_count
;
4269 lto_input_block
ib ((const char *) data
+ main_offset
, header
->main_size
,
4270 file_data
->mode_table
);
4273 lto_data_in_create (file_data
, (const char *) data
+ string_offset
,
4274 header
->string_size
, vNULL
);
4275 f_count
= streamer_read_uhwi (&ib
);
4276 for (i
= 0; i
< f_count
; i
++)
4279 struct cgraph_node
*node
;
4280 struct inline_summary
*info
;
4281 lto_symtab_encoder_t encoder
;
4282 struct bitpack_d bp
;
4283 struct cgraph_edge
*e
;
4286 index
= streamer_read_uhwi (&ib
);
4287 encoder
= file_data
->symtab_node_encoder
;
4288 node
= dyn_cast
<cgraph_node
*> (lto_symtab_encoder_deref (encoder
,
4290 info
= inline_summaries
->get (node
);
4292 info
->estimated_stack_size
4293 = info
->estimated_self_stack_size
= streamer_read_uhwi (&ib
);
4294 info
->size
= info
->self_size
= streamer_read_uhwi (&ib
);
4295 info
->time
= info
->self_time
= streamer_read_uhwi (&ib
);
4297 bp
= streamer_read_bitpack (&ib
);
4298 info
->inlinable
= bp_unpack_value (&bp
, 1);
4299 info
->contains_cilk_spawn
= bp_unpack_value (&bp
, 1);
4300 info
->fp_expressions
= bp_unpack_value (&bp
, 1);
4302 count2
= streamer_read_uhwi (&ib
);
4303 gcc_assert (!info
->conds
);
4304 for (j
= 0; j
< count2
; j
++)
4307 c
.operand_num
= streamer_read_uhwi (&ib
);
4308 c
.size
= streamer_read_uhwi (&ib
);
4309 c
.code
= (enum tree_code
) streamer_read_uhwi (&ib
);
4310 c
.val
= stream_read_tree (&ib
, data_in
);
4311 bp
= streamer_read_bitpack (&ib
);
4312 c
.agg_contents
= bp_unpack_value (&bp
, 1);
4313 c
.by_ref
= bp_unpack_value (&bp
, 1);
4315 c
.offset
= streamer_read_uhwi (&ib
);
4316 vec_safe_push (info
->conds
, c
);
4318 count2
= streamer_read_uhwi (&ib
);
4319 gcc_assert (!info
->entry
);
4320 for (j
= 0; j
< count2
; j
++)
4322 struct size_time_entry e
;
4324 e
.size
= streamer_read_uhwi (&ib
);
4325 e
.time
= streamer_read_uhwi (&ib
);
4326 e
.predicate
= read_predicate (&ib
);
4328 vec_safe_push (info
->entry
, e
);
4331 p
= read_predicate (&ib
);
4332 set_hint_predicate (&info
->loop_iterations
, p
);
4333 p
= read_predicate (&ib
);
4334 set_hint_predicate (&info
->loop_stride
, p
);
4335 p
= read_predicate (&ib
);
4336 set_hint_predicate (&info
->array_index
, p
);
4337 for (e
= node
->callees
; e
; e
= e
->next_callee
)
4338 read_inline_edge_summary (&ib
, e
);
4339 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
4340 read_inline_edge_summary (&ib
, e
);
4343 lto_free_section_data (file_data
, LTO_section_inline_summary
, NULL
, data
,
4345 lto_data_in_delete (data_in
);
4349 /* Read inline summary. Jump functions are shared among ipa-cp
4350 and inliner, so when ipa-cp is active, we don't need to write them
4354 inline_read_summary (void)
4356 struct lto_file_decl_data
**file_data_vec
= lto_get_file_decl_data ();
4357 struct lto_file_decl_data
*file_data
;
4360 inline_summary_alloc ();
4362 while ((file_data
= file_data_vec
[j
++]))
4365 const char *data
= lto_get_section_data (file_data
,
4366 LTO_section_inline_summary
,
4369 inline_read_section (file_data
, data
, len
);
4371 /* Fatal error here. We do not want to support compiling ltrans units
4372 with different version of compiler or different flags than the WPA
4373 unit, so this should never happen. */
4374 fatal_error (input_location
,
4375 "ipa inline summary is missing in input file");
4379 ipa_register_cgraph_hooks ();
4381 ipa_prop_read_jump_functions ();
4384 gcc_assert (inline_summaries
);
4385 inline_summaries
->enable_insertion_hook ();
4389 /* Write predicate P to OB. */
4392 write_predicate (struct output_block
*ob
, struct predicate
*p
)
4396 for (j
= 0; p
->clause
[j
]; j
++)
4398 gcc_assert (j
< MAX_CLAUSES
);
4399 streamer_write_uhwi (ob
, p
->clause
[j
]);
4401 streamer_write_uhwi (ob
, 0);
4405 /* Write inline summary for edge E to OB. */
4408 write_inline_edge_summary (struct output_block
*ob
, struct cgraph_edge
*e
)
4410 struct inline_edge_summary
*es
= inline_edge_summary (e
);
4413 streamer_write_uhwi (ob
, es
->call_stmt_size
);
4414 streamer_write_uhwi (ob
, es
->call_stmt_time
);
4415 streamer_write_uhwi (ob
, es
->loop_depth
);
4416 write_predicate (ob
, es
->predicate
);
4417 streamer_write_uhwi (ob
, es
->param
.length ());
4418 for (i
= 0; i
< (int) es
->param
.length (); i
++)
4419 streamer_write_uhwi (ob
, es
->param
[i
].change_prob
);
4423 /* Write inline summary for node in SET.
4424 Jump functions are shared among ipa-cp and inliner, so when ipa-cp is
4425 active, we don't need to write them twice. */
4428 inline_write_summary (void)
4430 struct output_block
*ob
= create_output_block (LTO_section_inline_summary
);
4431 lto_symtab_encoder_t encoder
= ob
->decl_state
->symtab_node_encoder
;
4432 unsigned int count
= 0;
4435 for (i
= 0; i
< lto_symtab_encoder_size (encoder
); i
++)
4437 symtab_node
*snode
= lto_symtab_encoder_deref (encoder
, i
);
4438 cgraph_node
*cnode
= dyn_cast
<cgraph_node
*> (snode
);
4439 if (cnode
&& cnode
->definition
&& !cnode
->alias
)
4442 streamer_write_uhwi (ob
, count
);
4444 for (i
= 0; i
< lto_symtab_encoder_size (encoder
); i
++)
4446 symtab_node
*snode
= lto_symtab_encoder_deref (encoder
, i
);
4447 cgraph_node
*cnode
= dyn_cast
<cgraph_node
*> (snode
);
4448 if (cnode
&& cnode
->definition
&& !cnode
->alias
)
4450 struct inline_summary
*info
= inline_summaries
->get (cnode
);
4451 struct bitpack_d bp
;
4452 struct cgraph_edge
*edge
;
4455 struct condition
*c
;
4457 streamer_write_uhwi (ob
, lto_symtab_encoder_encode (encoder
, cnode
));
4458 streamer_write_hwi (ob
, info
->estimated_self_stack_size
);
4459 streamer_write_hwi (ob
, info
->self_size
);
4460 streamer_write_hwi (ob
, info
->self_time
);
4461 bp
= bitpack_create (ob
->main_stream
);
4462 bp_pack_value (&bp
, info
->inlinable
, 1);
4463 bp_pack_value (&bp
, info
->contains_cilk_spawn
, 1);
4464 bp_pack_value (&bp
, info
->fp_expressions
, 1);
4465 streamer_write_bitpack (&bp
);
4466 streamer_write_uhwi (ob
, vec_safe_length (info
->conds
));
4467 for (i
= 0; vec_safe_iterate (info
->conds
, i
, &c
); i
++)
4469 streamer_write_uhwi (ob
, c
->operand_num
);
4470 streamer_write_uhwi (ob
, c
->size
);
4471 streamer_write_uhwi (ob
, c
->code
);
4472 stream_write_tree (ob
, c
->val
, true);
4473 bp
= bitpack_create (ob
->main_stream
);
4474 bp_pack_value (&bp
, c
->agg_contents
, 1);
4475 bp_pack_value (&bp
, c
->by_ref
, 1);
4476 streamer_write_bitpack (&bp
);
4477 if (c
->agg_contents
)
4478 streamer_write_uhwi (ob
, c
->offset
);
4480 streamer_write_uhwi (ob
, vec_safe_length (info
->entry
));
4481 for (i
= 0; vec_safe_iterate (info
->entry
, i
, &e
); i
++)
4483 streamer_write_uhwi (ob
, e
->size
);
4484 streamer_write_uhwi (ob
, e
->time
);
4485 write_predicate (ob
, &e
->predicate
);
4487 write_predicate (ob
, info
->loop_iterations
);
4488 write_predicate (ob
, info
->loop_stride
);
4489 write_predicate (ob
, info
->array_index
);
4490 for (edge
= cnode
->callees
; edge
; edge
= edge
->next_callee
)
4491 write_inline_edge_summary (ob
, edge
);
4492 for (edge
= cnode
->indirect_calls
; edge
; edge
= edge
->next_callee
)
4493 write_inline_edge_summary (ob
, edge
);
4496 streamer_write_char_stream (ob
->main_stream
, 0);
4497 produce_asm (ob
, NULL
);
4498 destroy_output_block (ob
);
4500 if (optimize
&& !flag_ipa_cp
)
4501 ipa_prop_write_jump_functions ();
4505 /* Release inline summary. */
4508 inline_free_summary (void)
4510 struct cgraph_node
*node
;
4511 if (edge_removal_hook_holder
)
4512 symtab
->remove_edge_removal_hook (edge_removal_hook_holder
);
4513 edge_removal_hook_holder
= NULL
;
4514 if (edge_duplication_hook_holder
)
4515 symtab
->remove_edge_duplication_hook (edge_duplication_hook_holder
);
4516 edge_duplication_hook_holder
= NULL
;
4517 if (!inline_edge_summary_vec
.exists ())
4519 FOR_EACH_DEFINED_FUNCTION (node
)
4521 reset_inline_summary (node
, inline_summaries
->get (node
));
4522 inline_summaries
->release ();
4523 inline_summaries
= NULL
;
4524 inline_edge_summary_vec
.release ();
4525 edge_predicate_pool
.release ();