1 /* Inlining decision heuristics.
2 Copyright (C) 2003-2016 Free Software Foundation, Inc.
3 Contributed by Jan Hubicka
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
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
;
1510 else if (cfun
&& fn_contains_cilk_spawn_p (cfun
))
1511 /* We can't inline if the function is spawing a function. */
1512 e
->inline_failed
= CIF_CILK_SPAWN
;
1514 e
->inline_failed
= CIF_FUNCTION_NOT_CONSIDERED
;
1515 gcc_checking_assert (!e
->call_stmt_cannot_inline_p
1516 || cgraph_inline_failed_type (e
->inline_failed
)
1517 == CIF_FINAL_ERROR
);
1520 /* Callback of walk_aliased_vdefs. Flags that it has been invoked to the
1521 boolean variable pointed to by DATA. */
1524 mark_modified (ao_ref
*ao ATTRIBUTE_UNUSED
, tree vdef ATTRIBUTE_UNUSED
,
1527 bool *b
= (bool *) data
;
1532 /* If OP refers to value of function parameter, return the corresponding
1533 parameter. If non-NULL, the size of the memory load (or the SSA_NAME of the
1534 PARM_DECL) will be stored to *SIZE_P in that case too. */
1537 unmodified_parm_1 (gimple
*stmt
, tree op
, HOST_WIDE_INT
*size_p
)
1539 /* SSA_NAME referring to parm default def? */
1540 if (TREE_CODE (op
) == SSA_NAME
1541 && SSA_NAME_IS_DEFAULT_DEF (op
)
1542 && TREE_CODE (SSA_NAME_VAR (op
)) == PARM_DECL
)
1545 *size_p
= tree_to_shwi (TYPE_SIZE (TREE_TYPE (op
)));
1546 return SSA_NAME_VAR (op
);
1548 /* Non-SSA parm reference? */
1549 if (TREE_CODE (op
) == PARM_DECL
)
1551 bool modified
= false;
1554 ao_ref_init (&refd
, op
);
1555 walk_aliased_vdefs (&refd
, gimple_vuse (stmt
), mark_modified
, &modified
,
1560 *size_p
= tree_to_shwi (TYPE_SIZE (TREE_TYPE (op
)));
1567 /* If OP refers to value of function parameter, return the corresponding
1568 parameter. Also traverse chains of SSA register assignments. If non-NULL,
1569 the size of the memory load (or the SSA_NAME of the PARM_DECL) will be
1570 stored to *SIZE_P in that case too. */
1573 unmodified_parm (gimple
*stmt
, tree op
, HOST_WIDE_INT
*size_p
)
1575 tree res
= unmodified_parm_1 (stmt
, op
, size_p
);
1579 if (TREE_CODE (op
) == SSA_NAME
1580 && !SSA_NAME_IS_DEFAULT_DEF (op
)
1581 && gimple_assign_single_p (SSA_NAME_DEF_STMT (op
)))
1582 return unmodified_parm (SSA_NAME_DEF_STMT (op
),
1583 gimple_assign_rhs1 (SSA_NAME_DEF_STMT (op
)),
1588 /* If OP refers to a value of a function parameter or value loaded from an
1589 aggregate passed to a parameter (either by value or reference), return TRUE
1590 and store the number of the parameter to *INDEX_P, the access size into
1591 *SIZE_P, and information whether and how it has been loaded from an
1592 aggregate into *AGGPOS. INFO describes the function parameters, STMT is the
1593 statement in which OP is used or loaded. */
1596 unmodified_parm_or_parm_agg_item (struct ipa_func_body_info
*fbi
,
1597 gimple
*stmt
, tree op
, int *index_p
,
1598 HOST_WIDE_INT
*size_p
,
1599 struct agg_position_info
*aggpos
)
1601 tree res
= unmodified_parm_1 (stmt
, op
, size_p
);
1603 gcc_checking_assert (aggpos
);
1606 *index_p
= ipa_get_param_decl_index (fbi
->info
, res
);
1609 aggpos
->agg_contents
= false;
1610 aggpos
->by_ref
= false;
1614 if (TREE_CODE (op
) == SSA_NAME
)
1616 if (SSA_NAME_IS_DEFAULT_DEF (op
)
1617 || !gimple_assign_single_p (SSA_NAME_DEF_STMT (op
)))
1619 stmt
= SSA_NAME_DEF_STMT (op
);
1620 op
= gimple_assign_rhs1 (stmt
);
1621 if (!REFERENCE_CLASS_P (op
))
1622 return unmodified_parm_or_parm_agg_item (fbi
, stmt
, op
, index_p
, size_p
,
1626 aggpos
->agg_contents
= true;
1627 return ipa_load_from_parm_agg (fbi
, fbi
->info
->descriptors
,
1628 stmt
, op
, index_p
, &aggpos
->offset
,
1629 size_p
, &aggpos
->by_ref
);
1632 /* See if statement might disappear after inlining.
1633 0 - means not eliminated
1634 1 - half of statements goes away
1635 2 - for sure it is eliminated.
1636 We are not terribly sophisticated, basically looking for simple abstraction
1637 penalty wrappers. */
1640 eliminated_by_inlining_prob (gimple
*stmt
)
1642 enum gimple_code code
= gimple_code (stmt
);
1643 enum tree_code rhs_code
;
1653 if (gimple_num_ops (stmt
) != 2)
1656 rhs_code
= gimple_assign_rhs_code (stmt
);
1658 /* Casts of parameters, loads from parameters passed by reference
1659 and stores to return value or parameters are often free after
1660 inlining dua to SRA and further combining.
1661 Assume that half of statements goes away. */
1662 if (CONVERT_EXPR_CODE_P (rhs_code
)
1663 || rhs_code
== VIEW_CONVERT_EXPR
1664 || rhs_code
== ADDR_EXPR
1665 || gimple_assign_rhs_class (stmt
) == GIMPLE_SINGLE_RHS
)
1667 tree rhs
= gimple_assign_rhs1 (stmt
);
1668 tree lhs
= gimple_assign_lhs (stmt
);
1669 tree inner_rhs
= get_base_address (rhs
);
1670 tree inner_lhs
= get_base_address (lhs
);
1671 bool rhs_free
= false;
1672 bool lhs_free
= false;
1679 /* Reads of parameter are expected to be free. */
1680 if (unmodified_parm (stmt
, inner_rhs
, NULL
))
1682 /* Match expressions of form &this->field. Those will most likely
1683 combine with something upstream after inlining. */
1684 else if (TREE_CODE (inner_rhs
) == ADDR_EXPR
)
1686 tree op
= get_base_address (TREE_OPERAND (inner_rhs
, 0));
1687 if (TREE_CODE (op
) == PARM_DECL
)
1689 else if (TREE_CODE (op
) == MEM_REF
1690 && unmodified_parm (stmt
, TREE_OPERAND (op
, 0), NULL
))
1694 /* When parameter is not SSA register because its address is taken
1695 and it is just copied into one, the statement will be completely
1696 free after inlining (we will copy propagate backward). */
1697 if (rhs_free
&& is_gimple_reg (lhs
))
1700 /* Reads of parameters passed by reference
1701 expected to be free (i.e. optimized out after inlining). */
1702 if (TREE_CODE (inner_rhs
) == MEM_REF
1703 && unmodified_parm (stmt
, TREE_OPERAND (inner_rhs
, 0), NULL
))
1706 /* Copying parameter passed by reference into gimple register is
1707 probably also going to copy propagate, but we can't be quite
1709 if (rhs_free
&& is_gimple_reg (lhs
))
1712 /* Writes to parameters, parameters passed by value and return value
1713 (either dirrectly or passed via invisible reference) are free.
1715 TODO: We ought to handle testcase like
1716 struct a {int a,b;};
1718 retrurnsturct (void)
1724 This translate into:
1739 For that we either need to copy ipa-split logic detecting writes
1741 if (TREE_CODE (inner_lhs
) == PARM_DECL
1742 || TREE_CODE (inner_lhs
) == RESULT_DECL
1743 || (TREE_CODE (inner_lhs
) == MEM_REF
1744 && (unmodified_parm (stmt
, TREE_OPERAND (inner_lhs
, 0), NULL
)
1745 || (TREE_CODE (TREE_OPERAND (inner_lhs
, 0)) == SSA_NAME
1746 && SSA_NAME_VAR (TREE_OPERAND (inner_lhs
, 0))
1747 && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND
1749 0))) == RESULT_DECL
))))
1752 && (is_gimple_reg (rhs
) || is_gimple_min_invariant (rhs
)))
1754 if (lhs_free
&& rhs_free
)
1764 /* If BB ends by a conditional we can turn into predicates, attach corresponding
1765 predicates to the CFG edges. */
1768 set_cond_stmt_execution_predicate (struct ipa_func_body_info
*fbi
,
1769 struct inline_summary
*summary
,
1776 struct agg_position_info aggpos
;
1777 enum tree_code code
, inverted_code
;
1783 last
= last_stmt (bb
);
1784 if (!last
|| gimple_code (last
) != GIMPLE_COND
)
1786 if (!is_gimple_ip_invariant (gimple_cond_rhs (last
)))
1788 op
= gimple_cond_lhs (last
);
1789 /* TODO: handle conditionals like
1792 if (unmodified_parm_or_parm_agg_item (fbi
, last
, op
, &index
, &size
, &aggpos
))
1794 code
= gimple_cond_code (last
);
1795 inverted_code
= invert_tree_comparison (code
, HONOR_NANS (op
));
1797 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1799 enum tree_code this_code
= (e
->flags
& EDGE_TRUE_VALUE
1800 ? code
: inverted_code
);
1801 /* invert_tree_comparison will return ERROR_MARK on FP
1802 comparsions that are not EQ/NE instead of returning proper
1803 unordered one. Be sure it is not confused with NON_CONSTANT. */
1804 if (this_code
!= ERROR_MARK
)
1807 = add_condition (summary
, index
, size
, &aggpos
, this_code
,
1808 unshare_expr_without_location
1809 (gimple_cond_rhs (last
)));
1810 e
->aux
= edge_predicate_pool
.allocate ();
1811 *(struct predicate
*) e
->aux
= p
;
1816 if (TREE_CODE (op
) != SSA_NAME
)
1819 if (builtin_constant_p (op))
1823 Here we can predicate nonconstant_code. We can't
1824 really handle constant_code since we have no predicate
1825 for this and also the constant code is not known to be
1826 optimized away when inliner doen't see operand is constant.
1827 Other optimizers might think otherwise. */
1828 if (gimple_cond_code (last
) != NE_EXPR
1829 || !integer_zerop (gimple_cond_rhs (last
)))
1831 set_stmt
= SSA_NAME_DEF_STMT (op
);
1832 if (!gimple_call_builtin_p (set_stmt
, BUILT_IN_CONSTANT_P
)
1833 || gimple_call_num_args (set_stmt
) != 1)
1835 op2
= gimple_call_arg (set_stmt
, 0);
1836 if (!unmodified_parm_or_parm_agg_item (fbi
, set_stmt
, op2
, &index
, &size
,
1839 FOR_EACH_EDGE (e
, ei
, bb
->succs
) if (e
->flags
& EDGE_FALSE_VALUE
)
1841 struct predicate p
= add_condition (summary
, index
, size
, &aggpos
,
1842 IS_NOT_CONSTANT
, NULL_TREE
);
1843 e
->aux
= edge_predicate_pool
.allocate ();
1844 *(struct predicate
*) e
->aux
= p
;
1849 /* If BB ends by a switch we can turn into predicates, attach corresponding
1850 predicates to the CFG edges. */
1853 set_switch_stmt_execution_predicate (struct ipa_func_body_info
*fbi
,
1854 struct inline_summary
*summary
,
1861 struct agg_position_info aggpos
;
1867 lastg
= last_stmt (bb
);
1868 if (!lastg
|| gimple_code (lastg
) != GIMPLE_SWITCH
)
1870 gswitch
*last
= as_a
<gswitch
*> (lastg
);
1871 op
= gimple_switch_index (last
);
1872 if (!unmodified_parm_or_parm_agg_item (fbi
, last
, op
, &index
, &size
, &aggpos
))
1875 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1877 e
->aux
= edge_predicate_pool
.allocate ();
1878 *(struct predicate
*) e
->aux
= false_predicate ();
1880 n
= gimple_switch_num_labels (last
);
1881 for (case_idx
= 0; case_idx
< n
; ++case_idx
)
1883 tree cl
= gimple_switch_label (last
, case_idx
);
1887 e
= find_edge (bb
, label_to_block (CASE_LABEL (cl
)));
1888 min
= CASE_LOW (cl
);
1889 max
= CASE_HIGH (cl
);
1891 /* For default we might want to construct predicate that none
1892 of cases is met, but it is bit hard to do not having negations
1893 of conditionals handy. */
1895 p
= true_predicate ();
1897 p
= add_condition (summary
, index
, size
, &aggpos
, EQ_EXPR
,
1898 unshare_expr_without_location (min
));
1901 struct predicate p1
, p2
;
1902 p1
= add_condition (summary
, index
, size
, &aggpos
, GE_EXPR
,
1903 unshare_expr_without_location (min
));
1904 p2
= add_condition (summary
, index
, size
, &aggpos
, LE_EXPR
,
1905 unshare_expr_without_location (max
));
1906 p
= and_predicates (summary
->conds
, &p1
, &p2
);
1908 *(struct predicate
*) e
->aux
1909 = or_predicates (summary
->conds
, &p
, (struct predicate
*) e
->aux
);
1914 /* For each BB in NODE attach to its AUX pointer predicate under
1915 which it is executable. */
1918 compute_bb_predicates (struct ipa_func_body_info
*fbi
,
1919 struct cgraph_node
*node
,
1920 struct inline_summary
*summary
)
1922 struct function
*my_function
= DECL_STRUCT_FUNCTION (node
->decl
);
1926 FOR_EACH_BB_FN (bb
, my_function
)
1928 set_cond_stmt_execution_predicate (fbi
, summary
, bb
);
1929 set_switch_stmt_execution_predicate (fbi
, summary
, bb
);
1932 /* Entry block is always executable. */
1933 ENTRY_BLOCK_PTR_FOR_FN (my_function
)->aux
1934 = edge_predicate_pool
.allocate ();
1935 *(struct predicate
*) ENTRY_BLOCK_PTR_FOR_FN (my_function
)->aux
1936 = true_predicate ();
1938 /* A simple dataflow propagation of predicates forward in the CFG.
1939 TODO: work in reverse postorder. */
1943 FOR_EACH_BB_FN (bb
, my_function
)
1945 struct predicate p
= false_predicate ();
1948 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
1952 struct predicate this_bb_predicate
1953 = *(struct predicate
*) e
->src
->aux
;
1956 = and_predicates (summary
->conds
, &this_bb_predicate
,
1957 (struct predicate
*) e
->aux
);
1958 p
= or_predicates (summary
->conds
, &p
, &this_bb_predicate
);
1959 if (true_predicate_p (&p
))
1963 if (false_predicate_p (&p
))
1964 gcc_assert (!bb
->aux
);
1970 bb
->aux
= edge_predicate_pool
.allocate ();
1971 *((struct predicate
*) bb
->aux
) = p
;
1973 else if (!predicates_equal_p (&p
, (struct predicate
*) bb
->aux
))
1975 /* This OR operation is needed to ensure monotonous data flow
1976 in the case we hit the limit on number of clauses and the
1977 and/or operations above give approximate answers. */
1978 p
= or_predicates (summary
->conds
, &p
, (struct predicate
*)bb
->aux
);
1979 if (!predicates_equal_p (&p
, (struct predicate
*) bb
->aux
))
1982 *((struct predicate
*) bb
->aux
) = p
;
1991 /* We keep info about constantness of SSA names. */
1993 typedef struct predicate predicate_t
;
1994 /* Return predicate specifying when the STMT might have result that is not
1995 a compile time constant. */
1997 static struct predicate
1998 will_be_nonconstant_expr_predicate (struct ipa_node_params
*info
,
1999 struct inline_summary
*summary
,
2001 vec
<predicate_t
> nonconstant_names
)
2007 while (UNARY_CLASS_P (expr
))
2008 expr
= TREE_OPERAND (expr
, 0);
2010 parm
= unmodified_parm (NULL
, expr
, &size
);
2011 if (parm
&& (index
= ipa_get_param_decl_index (info
, parm
)) >= 0)
2012 return add_condition (summary
, index
, size
, NULL
, CHANGED
, NULL_TREE
);
2013 if (is_gimple_min_invariant (expr
))
2014 return false_predicate ();
2015 if (TREE_CODE (expr
) == SSA_NAME
)
2016 return nonconstant_names
[SSA_NAME_VERSION (expr
)];
2017 if (BINARY_CLASS_P (expr
) || COMPARISON_CLASS_P (expr
))
2019 struct predicate p1
= will_be_nonconstant_expr_predicate
2020 (info
, summary
, TREE_OPERAND (expr
, 0),
2022 struct predicate p2
;
2023 if (true_predicate_p (&p1
))
2025 p2
= will_be_nonconstant_expr_predicate (info
, summary
,
2026 TREE_OPERAND (expr
, 1),
2028 return or_predicates (summary
->conds
, &p1
, &p2
);
2030 else if (TREE_CODE (expr
) == COND_EXPR
)
2032 struct predicate p1
= will_be_nonconstant_expr_predicate
2033 (info
, summary
, TREE_OPERAND (expr
, 0),
2035 struct predicate p2
;
2036 if (true_predicate_p (&p1
))
2038 p2
= will_be_nonconstant_expr_predicate (info
, summary
,
2039 TREE_OPERAND (expr
, 1),
2041 if (true_predicate_p (&p2
))
2043 p1
= or_predicates (summary
->conds
, &p1
, &p2
);
2044 p2
= will_be_nonconstant_expr_predicate (info
, summary
,
2045 TREE_OPERAND (expr
, 2),
2047 return or_predicates (summary
->conds
, &p1
, &p2
);
2054 return false_predicate ();
2058 /* Return predicate specifying when the STMT might have result that is not
2059 a compile time constant. */
2061 static struct predicate
2062 will_be_nonconstant_predicate (struct ipa_func_body_info
*fbi
,
2063 struct inline_summary
*summary
,
2065 vec
<predicate_t
> nonconstant_names
)
2067 struct predicate p
= true_predicate ();
2070 struct predicate op_non_const
;
2074 struct agg_position_info aggpos
;
2076 /* What statments might be optimized away
2077 when their arguments are constant. */
2078 if (gimple_code (stmt
) != GIMPLE_ASSIGN
2079 && gimple_code (stmt
) != GIMPLE_COND
2080 && gimple_code (stmt
) != GIMPLE_SWITCH
2081 && (gimple_code (stmt
) != GIMPLE_CALL
2082 || !(gimple_call_flags (stmt
) & ECF_CONST
)))
2085 /* Stores will stay anyway. */
2086 if (gimple_store_p (stmt
))
2089 is_load
= gimple_assign_load_p (stmt
);
2091 /* Loads can be optimized when the value is known. */
2095 gcc_assert (gimple_assign_single_p (stmt
));
2096 op
= gimple_assign_rhs1 (stmt
);
2097 if (!unmodified_parm_or_parm_agg_item (fbi
, stmt
, op
, &base_index
, &size
,
2104 /* See if we understand all operands before we start
2105 adding conditionals. */
2106 FOR_EACH_SSA_TREE_OPERAND (use
, stmt
, iter
, SSA_OP_USE
)
2108 tree parm
= unmodified_parm (stmt
, use
, NULL
);
2109 /* For arguments we can build a condition. */
2110 if (parm
&& ipa_get_param_decl_index (fbi
->info
, parm
) >= 0)
2112 if (TREE_CODE (use
) != SSA_NAME
)
2114 /* If we know when operand is constant,
2115 we still can say something useful. */
2116 if (!true_predicate_p (&nonconstant_names
[SSA_NAME_VERSION (use
)]))
2123 add_condition (summary
, base_index
, size
, &aggpos
, CHANGED
, NULL
);
2125 op_non_const
= false_predicate ();
2126 FOR_EACH_SSA_TREE_OPERAND (use
, stmt
, iter
, SSA_OP_USE
)
2129 tree parm
= unmodified_parm (stmt
, use
, &size
);
2132 if (parm
&& (index
= ipa_get_param_decl_index (fbi
->info
, parm
)) >= 0)
2134 if (index
!= base_index
)
2135 p
= add_condition (summary
, index
, size
, NULL
, CHANGED
, NULL_TREE
);
2140 p
= nonconstant_names
[SSA_NAME_VERSION (use
)];
2141 op_non_const
= or_predicates (summary
->conds
, &p
, &op_non_const
);
2143 if ((gimple_code (stmt
) == GIMPLE_ASSIGN
|| gimple_code (stmt
) == GIMPLE_CALL
)
2144 && gimple_op (stmt
, 0)
2145 && TREE_CODE (gimple_op (stmt
, 0)) == SSA_NAME
)
2146 nonconstant_names
[SSA_NAME_VERSION (gimple_op (stmt
, 0))]
2148 return op_non_const
;
2151 struct record_modified_bb_info
2157 /* Callback of walk_aliased_vdefs. Records basic blocks where the value may be
2158 set except for info->stmt. */
2161 record_modified (ao_ref
*ao ATTRIBUTE_UNUSED
, tree vdef
, void *data
)
2163 struct record_modified_bb_info
*info
=
2164 (struct record_modified_bb_info
*) data
;
2165 if (SSA_NAME_DEF_STMT (vdef
) == info
->stmt
)
2167 bitmap_set_bit (info
->bb_set
,
2168 SSA_NAME_IS_DEFAULT_DEF (vdef
)
2169 ? ENTRY_BLOCK_PTR_FOR_FN (cfun
)->index
2170 : gimple_bb (SSA_NAME_DEF_STMT (vdef
))->index
);
2174 /* Return probability (based on REG_BR_PROB_BASE) that I-th parameter of STMT
2175 will change since last invocation of STMT.
2177 Value 0 is reserved for compile time invariants.
2178 For common parameters it is REG_BR_PROB_BASE. For loop invariants it
2179 ought to be REG_BR_PROB_BASE / estimated_iters. */
2182 param_change_prob (gimple
*stmt
, int i
)
2184 tree op
= gimple_call_arg (stmt
, i
);
2185 basic_block bb
= gimple_bb (stmt
);
2188 /* Global invariants neve change. */
2189 if (is_gimple_min_invariant (op
))
2191 /* We would have to do non-trivial analysis to really work out what
2192 is the probability of value to change (i.e. when init statement
2193 is in a sibling loop of the call).
2195 We do an conservative estimate: when call is executed N times more often
2196 than the statement defining value, we take the frequency 1/N. */
2197 if (TREE_CODE (op
) == SSA_NAME
)
2202 return REG_BR_PROB_BASE
;
2204 if (SSA_NAME_IS_DEFAULT_DEF (op
))
2205 init_freq
= ENTRY_BLOCK_PTR_FOR_FN (cfun
)->frequency
;
2207 init_freq
= gimple_bb (SSA_NAME_DEF_STMT (op
))->frequency
;
2211 if (init_freq
< bb
->frequency
)
2212 return MAX (GCOV_COMPUTE_SCALE (init_freq
, bb
->frequency
), 1);
2214 return REG_BR_PROB_BASE
;
2217 base
= get_base_address (op
);
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
;
2259 return REG_BR_PROB_BASE
;
2262 /* Find whether a basic block BB is the final block of a (half) diamond CFG
2263 sub-graph and if the predicate the condition depends on is known. If so,
2264 return true and store the pointer the predicate in *P. */
2267 phi_result_unknown_predicate (struct ipa_node_params
*info
,
2268 inline_summary
*summary
, basic_block bb
,
2269 struct predicate
*p
,
2270 vec
<predicate_t
> nonconstant_names
)
2274 basic_block first_bb
= NULL
;
2277 if (single_pred_p (bb
))
2279 *p
= false_predicate ();
2283 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
2285 if (single_succ_p (e
->src
))
2287 if (!single_pred_p (e
->src
))
2290 first_bb
= single_pred (e
->src
);
2291 else if (single_pred (e
->src
) != first_bb
)
2298 else if (e
->src
!= first_bb
)
2306 stmt
= last_stmt (first_bb
);
2308 || gimple_code (stmt
) != GIMPLE_COND
2309 || !is_gimple_ip_invariant (gimple_cond_rhs (stmt
)))
2312 *p
= will_be_nonconstant_expr_predicate (info
, summary
,
2313 gimple_cond_lhs (stmt
),
2315 if (true_predicate_p (p
))
2321 /* Given a PHI statement in a function described by inline properties SUMMARY
2322 and *P being the predicate describing whether the selected PHI argument is
2323 known, store a predicate for the result of the PHI statement into
2324 NONCONSTANT_NAMES, if possible. */
2327 predicate_for_phi_result (struct inline_summary
*summary
, gphi
*phi
,
2328 struct predicate
*p
,
2329 vec
<predicate_t
> nonconstant_names
)
2333 for (i
= 0; i
< gimple_phi_num_args (phi
); i
++)
2335 tree arg
= gimple_phi_arg (phi
, i
)->def
;
2336 if (!is_gimple_min_invariant (arg
))
2338 gcc_assert (TREE_CODE (arg
) == SSA_NAME
);
2339 *p
= or_predicates (summary
->conds
, p
,
2340 &nonconstant_names
[SSA_NAME_VERSION (arg
)]);
2341 if (true_predicate_p (p
))
2346 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2348 fprintf (dump_file
, "\t\tphi predicate: ");
2349 dump_predicate (dump_file
, summary
->conds
, p
);
2351 nonconstant_names
[SSA_NAME_VERSION (gimple_phi_result (phi
))] = *p
;
2354 /* Return predicate specifying when array index in access OP becomes non-constant. */
2356 static struct predicate
2357 array_index_predicate (inline_summary
*info
,
2358 vec
< predicate_t
> nonconstant_names
, tree op
)
2360 struct predicate p
= false_predicate ();
2361 while (handled_component_p (op
))
2363 if (TREE_CODE (op
) == ARRAY_REF
|| TREE_CODE (op
) == ARRAY_RANGE_REF
)
2365 if (TREE_CODE (TREE_OPERAND (op
, 1)) == SSA_NAME
)
2366 p
= or_predicates (info
->conds
, &p
,
2367 &nonconstant_names
[SSA_NAME_VERSION
2368 (TREE_OPERAND (op
, 1))]);
2370 op
= TREE_OPERAND (op
, 0);
2375 /* For a typical usage of __builtin_expect (a<b, 1), we
2376 may introduce an extra relation stmt:
2377 With the builtin, we have
2380 t3 = __builtin_expect (t2, 1);
2383 Without the builtin, we have
2386 This affects the size/time estimation and may have
2387 an impact on the earlier inlining.
2388 Here find this pattern and fix it up later. */
2391 find_foldable_builtin_expect (basic_block bb
)
2393 gimple_stmt_iterator bsi
;
2395 for (bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
2397 gimple
*stmt
= gsi_stmt (bsi
);
2398 if (gimple_call_builtin_p (stmt
, BUILT_IN_EXPECT
)
2399 || (is_gimple_call (stmt
)
2400 && gimple_call_internal_p (stmt
)
2401 && gimple_call_internal_fn (stmt
) == IFN_BUILTIN_EXPECT
))
2403 tree var
= gimple_call_lhs (stmt
);
2404 tree arg
= gimple_call_arg (stmt
, 0);
2405 use_operand_p use_p
;
2412 gcc_assert (TREE_CODE (var
) == SSA_NAME
);
2414 while (TREE_CODE (arg
) == SSA_NAME
)
2416 gimple
*stmt_tmp
= SSA_NAME_DEF_STMT (arg
);
2417 if (!is_gimple_assign (stmt_tmp
))
2419 switch (gimple_assign_rhs_code (stmt_tmp
))
2438 arg
= gimple_assign_rhs1 (stmt_tmp
);
2441 if (match
&& single_imm_use (var
, &use_p
, &use_stmt
)
2442 && gimple_code (use_stmt
) == GIMPLE_COND
)
2449 /* Return true when the basic blocks contains only clobbers followed by RESX.
2450 Such BBs are kept around to make removal of dead stores possible with
2451 presence of EH and will be optimized out by optimize_clobbers later in the
2454 NEED_EH is used to recurse in case the clobber has non-EH predecestors
2455 that can be clobber only, too.. When it is false, the RESX is not necessary
2456 on the end of basic block. */
2459 clobber_only_eh_bb_p (basic_block bb
, bool need_eh
= true)
2461 gimple_stmt_iterator gsi
= gsi_last_bb (bb
);
2467 if (gsi_end_p (gsi
))
2469 if (gimple_code (gsi_stmt (gsi
)) != GIMPLE_RESX
)
2473 else if (!single_succ_p (bb
))
2476 for (; !gsi_end_p (gsi
); gsi_prev (&gsi
))
2478 gimple
*stmt
= gsi_stmt (gsi
);
2479 if (is_gimple_debug (stmt
))
2481 if (gimple_clobber_p (stmt
))
2483 if (gimple_code (stmt
) == GIMPLE_LABEL
)
2488 /* See if all predecestors are either throws or clobber only BBs. */
2489 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
2490 if (!(e
->flags
& EDGE_EH
)
2491 && !clobber_only_eh_bb_p (e
->src
, false))
2497 /* Return true if STMT compute a floating point expression that may be affected
2498 by -ffast-math and similar flags. */
2501 fp_expression_p (gimple
*stmt
)
2506 FOR_EACH_SSA_TREE_OPERAND (op
, stmt
, i
, SSA_OP_DEF
|SSA_OP_USE
)
2507 if (FLOAT_TYPE_P (TREE_TYPE (op
)))
2512 /* Compute function body size parameters for NODE.
2513 When EARLY is true, we compute only simple summaries without
2514 non-trivial predicates to drive the early inliner. */
2517 estimate_function_body_sizes (struct cgraph_node
*node
, bool early
)
2520 /* Estimate static overhead for function prologue/epilogue and alignment. */
2522 /* Benefits are scaled by probability of elimination that is in range
2525 struct function
*my_function
= DECL_STRUCT_FUNCTION (node
->decl
);
2527 struct inline_summary
*info
= inline_summaries
->get (node
);
2528 struct predicate bb_predicate
;
2529 struct ipa_func_body_info fbi
;
2530 vec
<predicate_t
> nonconstant_names
= vNULL
;
2533 predicate array_index
= true_predicate ();
2534 gimple
*fix_builtin_expect_stmt
;
2536 gcc_assert (my_function
&& my_function
->cfg
);
2537 gcc_assert (cfun
== my_function
);
2539 memset(&fbi
, 0, sizeof(fbi
));
2543 /* When optimizing and analyzing for IPA inliner, initialize loop optimizer
2544 so we can produce proper inline hints.
2546 When optimizing and analyzing for early inliner, initialize node params
2547 so we can produce correct BB predicates. */
2549 if (opt_for_fn (node
->decl
, optimize
))
2551 calculate_dominance_info (CDI_DOMINATORS
);
2553 loop_optimizer_init (LOOPS_NORMAL
| LOOPS_HAVE_RECORDED_EXITS
);
2556 ipa_check_create_node_params ();
2557 ipa_initialize_node_params (node
);
2560 if (ipa_node_params_sum
)
2563 fbi
.info
= IPA_NODE_REF (node
);
2564 fbi
.bb_infos
= vNULL
;
2565 fbi
.bb_infos
.safe_grow_cleared (last_basic_block_for_fn (cfun
));
2566 fbi
.param_count
= count_formal_params(node
->decl
);
2567 nonconstant_names
.safe_grow_cleared
2568 (SSANAMES (my_function
)->length ());
2573 fprintf (dump_file
, "\nAnalyzing function body size: %s\n",
2576 /* When we run into maximal number of entries, we assign everything to the
2577 constant truth case. Be sure to have it in list. */
2578 bb_predicate
= true_predicate ();
2579 account_size_time (info
, 0, 0, &bb_predicate
);
2581 bb_predicate
= not_inlined_predicate ();
2582 account_size_time (info
, 2 * INLINE_SIZE_SCALE
, 0, &bb_predicate
);
2585 compute_bb_predicates (&fbi
, node
, info
);
2586 order
= XNEWVEC (int, n_basic_blocks_for_fn (cfun
));
2587 nblocks
= pre_and_rev_post_order_compute (NULL
, order
, false);
2588 for (n
= 0; n
< nblocks
; n
++)
2590 bb
= BASIC_BLOCK_FOR_FN (cfun
, order
[n
]);
2591 freq
= compute_call_stmt_bb_frequency (node
->decl
, bb
);
2592 if (clobber_only_eh_bb_p (bb
))
2594 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2595 fprintf (dump_file
, "\n Ignoring BB %i;"
2596 " it will be optimized away by cleanup_clobbers\n",
2601 /* TODO: Obviously predicates can be propagated down across CFG. */
2605 bb_predicate
= *(struct predicate
*) bb
->aux
;
2607 bb_predicate
= false_predicate ();
2610 bb_predicate
= true_predicate ();
2612 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2614 fprintf (dump_file
, "\n BB %i predicate:", bb
->index
);
2615 dump_predicate (dump_file
, info
->conds
, &bb_predicate
);
2618 if (fbi
.info
&& nonconstant_names
.exists ())
2620 struct predicate phi_predicate
;
2621 bool first_phi
= true;
2623 for (gphi_iterator bsi
= gsi_start_phis (bb
); !gsi_end_p (bsi
);
2627 && !phi_result_unknown_predicate (fbi
.info
, info
, bb
,
2632 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2634 fprintf (dump_file
, " ");
2635 print_gimple_stmt (dump_file
, gsi_stmt (bsi
), 0, 0);
2637 predicate_for_phi_result (info
, bsi
.phi (), &phi_predicate
,
2642 fix_builtin_expect_stmt
= find_foldable_builtin_expect (bb
);
2644 for (gimple_stmt_iterator bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
);
2647 gimple
*stmt
= gsi_stmt (bsi
);
2648 int this_size
= estimate_num_insns (stmt
, &eni_size_weights
);
2649 int this_time
= estimate_num_insns (stmt
, &eni_time_weights
);
2651 struct predicate will_be_nonconstant
;
2653 /* This relation stmt should be folded after we remove
2654 buildin_expect call. Adjust the cost here. */
2655 if (stmt
== fix_builtin_expect_stmt
)
2661 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2663 fprintf (dump_file
, " ");
2664 print_gimple_stmt (dump_file
, stmt
, 0, 0);
2665 fprintf (dump_file
, "\t\tfreq:%3.2f size:%3i time:%3i\n",
2666 ((double) freq
) / CGRAPH_FREQ_BASE
, this_size
,
2670 if (gimple_assign_load_p (stmt
) && nonconstant_names
.exists ())
2672 struct predicate this_array_index
;
2674 array_index_predicate (info
, nonconstant_names
,
2675 gimple_assign_rhs1 (stmt
));
2676 if (!false_predicate_p (&this_array_index
))
2678 and_predicates (info
->conds
, &array_index
,
2681 if (gimple_store_p (stmt
) && nonconstant_names
.exists ())
2683 struct predicate this_array_index
;
2685 array_index_predicate (info
, nonconstant_names
,
2686 gimple_get_lhs (stmt
));
2687 if (!false_predicate_p (&this_array_index
))
2689 and_predicates (info
->conds
, &array_index
,
2694 if (is_gimple_call (stmt
)
2695 && !gimple_call_internal_p (stmt
))
2697 struct cgraph_edge
*edge
= node
->get_edge (stmt
);
2698 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
2700 /* Special case: results of BUILT_IN_CONSTANT_P will be always
2701 resolved as constant. We however don't want to optimize
2702 out the cgraph edges. */
2703 if (nonconstant_names
.exists ()
2704 && gimple_call_builtin_p (stmt
, BUILT_IN_CONSTANT_P
)
2705 && gimple_call_lhs (stmt
)
2706 && TREE_CODE (gimple_call_lhs (stmt
)) == SSA_NAME
)
2708 struct predicate false_p
= false_predicate ();
2709 nonconstant_names
[SSA_NAME_VERSION (gimple_call_lhs (stmt
))]
2712 if (ipa_node_params_sum
)
2714 int count
= gimple_call_num_args (stmt
);
2718 es
->param
.safe_grow_cleared (count
);
2719 for (i
= 0; i
< count
; i
++)
2721 int prob
= param_change_prob (stmt
, i
);
2722 gcc_assert (prob
>= 0 && prob
<= REG_BR_PROB_BASE
);
2723 es
->param
[i
].change_prob
= prob
;
2727 es
->call_stmt_size
= this_size
;
2728 es
->call_stmt_time
= this_time
;
2729 es
->loop_depth
= bb_loop_depth (bb
);
2730 edge_set_predicate (edge
, &bb_predicate
);
2733 /* TODO: When conditional jump or swithc is known to be constant, but
2734 we did not translate it into the predicates, we really can account
2735 just maximum of the possible paths. */
2738 = will_be_nonconstant_predicate (&fbi
, info
,
2739 stmt
, nonconstant_names
);
2740 if (this_time
|| this_size
)
2746 prob
= eliminated_by_inlining_prob (stmt
);
2747 if (prob
== 1 && dump_file
&& (dump_flags
& TDF_DETAILS
))
2749 "\t\t50%% will be eliminated by inlining\n");
2750 if (prob
== 2 && dump_file
&& (dump_flags
& TDF_DETAILS
))
2751 fprintf (dump_file
, "\t\tWill be eliminated by inlining\n");
2754 p
= and_predicates (info
->conds
, &bb_predicate
,
2755 &will_be_nonconstant
);
2757 p
= true_predicate ();
2759 if (!false_predicate_p (&p
)
2760 || (is_gimple_call (stmt
)
2761 && !false_predicate_p (&bb_predicate
)))
2765 if (time
> MAX_TIME
* INLINE_TIME_SCALE
)
2766 time
= MAX_TIME
* INLINE_TIME_SCALE
;
2769 /* We account everything but the calls. Calls have their own
2770 size/time info attached to cgraph edges. This is necessary
2771 in order to make the cost disappear after inlining. */
2772 if (!is_gimple_call (stmt
))
2776 struct predicate ip
= not_inlined_predicate ();
2777 ip
= and_predicates (info
->conds
, &ip
, &p
);
2778 account_size_time (info
, this_size
* prob
,
2779 this_time
* prob
, &ip
);
2782 account_size_time (info
, this_size
* (2 - prob
),
2783 this_time
* (2 - prob
), &p
);
2786 if (!info
->fp_expressions
&& fp_expression_p (stmt
))
2788 info
->fp_expressions
= true;
2790 fprintf (dump_file
, " fp_expression set\n");
2793 gcc_assert (time
>= 0);
2794 gcc_assert (size
>= 0);
2798 set_hint_predicate (&inline_summaries
->get (node
)->array_index
, array_index
);
2799 time
= (time
+ CGRAPH_FREQ_BASE
/ 2) / CGRAPH_FREQ_BASE
;
2800 if (time
> MAX_TIME
)
2804 if (nonconstant_names
.exists () && !early
)
2807 predicate loop_iterations
= true_predicate ();
2808 predicate loop_stride
= true_predicate ();
2810 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2811 flow_loops_dump (dump_file
, NULL
, 0);
2813 FOR_EACH_LOOP (loop
, 0)
2818 struct tree_niter_desc niter_desc
;
2819 bb_predicate
= *(struct predicate
*) loop
->header
->aux
;
2821 exits
= get_loop_exit_edges (loop
);
2822 FOR_EACH_VEC_ELT (exits
, j
, ex
)
2823 if (number_of_iterations_exit (loop
, ex
, &niter_desc
, false)
2824 && !is_gimple_min_invariant (niter_desc
.niter
))
2826 predicate will_be_nonconstant
2827 = will_be_nonconstant_expr_predicate (fbi
.info
, info
,
2830 if (!true_predicate_p (&will_be_nonconstant
))
2831 will_be_nonconstant
= and_predicates (info
->conds
,
2833 &will_be_nonconstant
);
2834 if (!true_predicate_p (&will_be_nonconstant
)
2835 && !false_predicate_p (&will_be_nonconstant
))
2836 /* This is slightly inprecise. We may want to represent each
2837 loop with independent predicate. */
2839 and_predicates (info
->conds
, &loop_iterations
,
2840 &will_be_nonconstant
);
2845 /* To avoid quadratic behavior we analyze stride predicates only
2846 with respect to the containing loop. Thus we simply iterate
2847 over all defs in the outermost loop body. */
2848 for (loop
= loops_for_fn (cfun
)->tree_root
->inner
;
2849 loop
!= NULL
; loop
= loop
->next
)
2851 basic_block
*body
= get_loop_body (loop
);
2852 for (unsigned i
= 0; i
< loop
->num_nodes
; i
++)
2854 gimple_stmt_iterator gsi
;
2855 bb_predicate
= *(struct predicate
*) body
[i
]->aux
;
2856 for (gsi
= gsi_start_bb (body
[i
]); !gsi_end_p (gsi
);
2859 gimple
*stmt
= gsi_stmt (gsi
);
2861 if (!is_gimple_assign (stmt
))
2864 tree def
= gimple_assign_lhs (stmt
);
2865 if (TREE_CODE (def
) != SSA_NAME
)
2869 if (!simple_iv (loop_containing_stmt (stmt
),
2870 loop_containing_stmt (stmt
),
2872 || is_gimple_min_invariant (iv
.step
))
2875 predicate will_be_nonconstant
2876 = will_be_nonconstant_expr_predicate (fbi
.info
, info
,
2879 if (!true_predicate_p (&will_be_nonconstant
))
2881 = and_predicates (info
->conds
, &bb_predicate
,
2882 &will_be_nonconstant
);
2883 if (!true_predicate_p (&will_be_nonconstant
)
2884 && !false_predicate_p (&will_be_nonconstant
))
2885 /* This is slightly inprecise. We may want to represent
2886 each loop with independent predicate. */
2887 loop_stride
= and_predicates (info
->conds
, &loop_stride
,
2888 &will_be_nonconstant
);
2893 set_hint_predicate (&inline_summaries
->get (node
)->loop_iterations
,
2895 set_hint_predicate (&inline_summaries
->get (node
)->loop_stride
,
2899 FOR_ALL_BB_FN (bb
, my_function
)
2905 edge_predicate_pool
.remove ((predicate
*)bb
->aux
);
2907 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
2910 edge_predicate_pool
.remove ((predicate
*) e
->aux
);
2914 inline_summaries
->get (node
)->self_time
= time
;
2915 inline_summaries
->get (node
)->self_size
= size
;
2916 nonconstant_names
.release ();
2917 ipa_release_body_info (&fbi
);
2918 if (opt_for_fn (node
->decl
, optimize
))
2921 loop_optimizer_finalize ();
2922 else if (!ipa_edge_args_vector
)
2923 ipa_free_all_node_params ();
2924 free_dominance_info (CDI_DOMINATORS
);
2928 fprintf (dump_file
, "\n");
2929 dump_inline_summary (dump_file
, node
);
2934 /* Compute parameters of functions used by inliner.
2935 EARLY is true when we compute parameters for the early inliner */
2938 compute_inline_parameters (struct cgraph_node
*node
, bool early
)
2940 HOST_WIDE_INT self_stack_size
;
2941 struct cgraph_edge
*e
;
2942 struct inline_summary
*info
;
2944 gcc_assert (!node
->global
.inlined_to
);
2946 inline_summary_alloc ();
2948 info
= inline_summaries
->get (node
);
2949 reset_inline_summary (node
, info
);
2951 /* Estimate the stack size for the function if we're optimizing. */
2952 self_stack_size
= optimize
&& !node
->thunk
.thunk_p
2953 ? estimated_stack_frame_size (node
) : 0;
2954 info
->estimated_self_stack_size
= self_stack_size
;
2955 info
->estimated_stack_size
= self_stack_size
;
2956 info
->stack_frame_offset
= 0;
2958 if (node
->thunk
.thunk_p
)
2960 struct inline_edge_summary
*es
= inline_edge_summary (node
->callees
);
2961 struct predicate t
= true_predicate ();
2963 node
->local
.can_change_signature
= false;
2964 es
->call_stmt_size
= eni_size_weights
.call_cost
;
2965 es
->call_stmt_time
= eni_time_weights
.call_cost
;
2966 account_size_time (info
, INLINE_SIZE_SCALE
* 2,
2967 INLINE_TIME_SCALE
* 2, &t
);
2968 t
= not_inlined_predicate ();
2969 account_size_time (info
, 2 * INLINE_SIZE_SCALE
, 0, &t
);
2970 inline_update_overall_summary (node
);
2971 info
->self_size
= info
->size
;
2972 info
->self_time
= info
->time
;
2973 /* We can not inline instrumetnation clones. */
2974 if (node
->thunk
.add_pointer_bounds_args
)
2976 info
->inlinable
= false;
2977 node
->callees
->inline_failed
= CIF_CHKP
;
2980 info
->inlinable
= true;
2984 /* Even is_gimple_min_invariant rely on current_function_decl. */
2985 push_cfun (DECL_STRUCT_FUNCTION (node
->decl
));
2987 /* Can this function be inlined at all? */
2988 if (!opt_for_fn (node
->decl
, optimize
)
2989 && !lookup_attribute ("always_inline",
2990 DECL_ATTRIBUTES (node
->decl
)))
2991 info
->inlinable
= false;
2993 info
->inlinable
= tree_inlinable_function_p (node
->decl
);
2995 info
->contains_cilk_spawn
= fn_contains_cilk_spawn_p (cfun
);
2997 /* Type attributes can use parameter indices to describe them. */
2998 if (TYPE_ATTRIBUTES (TREE_TYPE (node
->decl
)))
2999 node
->local
.can_change_signature
= false;
3002 /* Otherwise, inlinable functions always can change signature. */
3003 if (info
->inlinable
)
3004 node
->local
.can_change_signature
= true;
3007 /* Functions calling builtin_apply can not change signature. */
3008 for (e
= node
->callees
; e
; e
= e
->next_callee
)
3010 tree
cdecl = e
->callee
->decl
;
3011 if (DECL_BUILT_IN (cdecl)
3012 && DECL_BUILT_IN_CLASS (cdecl) == BUILT_IN_NORMAL
3013 && (DECL_FUNCTION_CODE (cdecl) == BUILT_IN_APPLY_ARGS
3014 || DECL_FUNCTION_CODE (cdecl) == BUILT_IN_VA_START
))
3017 node
->local
.can_change_signature
= !e
;
3020 /* Functions called by instrumentation thunk can't change signature
3021 because instrumentation thunk modification is not supported. */
3022 if (node
->local
.can_change_signature
)
3023 for (e
= node
->callers
; e
; e
= e
->next_caller
)
3024 if (e
->caller
->thunk
.thunk_p
3025 && e
->caller
->thunk
.add_pointer_bounds_args
)
3027 node
->local
.can_change_signature
= false;
3030 estimate_function_body_sizes (node
, early
);
3033 for (e
= node
->callees
; e
; e
= e
->next_callee
)
3034 if (e
->callee
->comdat_local_p ())
3036 node
->calls_comdat_local
= (e
!= NULL
);
3038 /* Inlining characteristics are maintained by the cgraph_mark_inline. */
3039 info
->time
= info
->self_time
;
3040 info
->size
= info
->self_size
;
3041 info
->stack_frame_offset
= 0;
3042 info
->estimated_stack_size
= info
->estimated_self_stack_size
;
3045 inline_update_overall_summary (node
);
3046 gcc_assert (info
->time
== info
->self_time
3047 && info
->size
== info
->self_size
);
3052 /* Compute parameters of functions used by inliner using
3053 current_function_decl. */
3056 compute_inline_parameters_for_current (void)
3058 compute_inline_parameters (cgraph_node::get (current_function_decl
), true);
3064 const pass_data pass_data_inline_parameters
=
3066 GIMPLE_PASS
, /* type */
3067 "inline_param", /* name */
3068 OPTGROUP_INLINE
, /* optinfo_flags */
3069 TV_INLINE_PARAMETERS
, /* tv_id */
3070 0, /* properties_required */
3071 0, /* properties_provided */
3072 0, /* properties_destroyed */
3073 0, /* todo_flags_start */
3074 0, /* todo_flags_finish */
3077 class pass_inline_parameters
: public gimple_opt_pass
3080 pass_inline_parameters (gcc::context
*ctxt
)
3081 : gimple_opt_pass (pass_data_inline_parameters
, ctxt
)
3084 /* opt_pass methods: */
3085 opt_pass
* clone () { return new pass_inline_parameters (m_ctxt
); }
3086 virtual unsigned int execute (function
*)
3088 return compute_inline_parameters_for_current ();
3091 }; // class pass_inline_parameters
3096 make_pass_inline_parameters (gcc::context
*ctxt
)
3098 return new pass_inline_parameters (ctxt
);
3102 /* Estimate benefit devirtualizing indirect edge IE, provided KNOWN_VALS,
3103 KNOWN_CONTEXTS and KNOWN_AGGS. */
3106 estimate_edge_devirt_benefit (struct cgraph_edge
*ie
,
3107 int *size
, int *time
,
3108 vec
<tree
> known_vals
,
3109 vec
<ipa_polymorphic_call_context
> known_contexts
,
3110 vec
<ipa_agg_jump_function_p
> known_aggs
)
3113 struct cgraph_node
*callee
;
3114 struct inline_summary
*isummary
;
3115 enum availability avail
;
3118 if (!known_vals
.exists () && !known_contexts
.exists ())
3120 if (!opt_for_fn (ie
->caller
->decl
, flag_indirect_inlining
))
3123 target
= ipa_get_indirect_edge_target (ie
, known_vals
, known_contexts
,
3124 known_aggs
, &speculative
);
3125 if (!target
|| speculative
)
3128 /* Account for difference in cost between indirect and direct calls. */
3129 *size
-= (eni_size_weights
.indirect_call_cost
- eni_size_weights
.call_cost
);
3130 *time
-= (eni_time_weights
.indirect_call_cost
- eni_time_weights
.call_cost
);
3131 gcc_checking_assert (*time
>= 0);
3132 gcc_checking_assert (*size
>= 0);
3134 callee
= cgraph_node::get (target
);
3135 if (!callee
|| !callee
->definition
)
3137 callee
= callee
->function_symbol (&avail
);
3138 if (avail
< AVAIL_AVAILABLE
)
3140 isummary
= inline_summaries
->get (callee
);
3141 return isummary
->inlinable
;
3144 /* Increase SIZE, MIN_SIZE (if non-NULL) and TIME for size and time needed to
3145 handle edge E with probability PROB.
3146 Set HINTS if edge may be devirtualized.
3147 KNOWN_VALS, KNOWN_AGGS and KNOWN_CONTEXTS describe context of the call
3151 estimate_edge_size_and_time (struct cgraph_edge
*e
, int *size
, int *min_size
,
3154 vec
<tree
> known_vals
,
3155 vec
<ipa_polymorphic_call_context
> known_contexts
,
3156 vec
<ipa_agg_jump_function_p
> known_aggs
,
3157 inline_hints
*hints
)
3159 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3160 int call_size
= es
->call_stmt_size
;
3161 int call_time
= es
->call_stmt_time
;
3164 && estimate_edge_devirt_benefit (e
, &call_size
, &call_time
,
3165 known_vals
, known_contexts
, known_aggs
)
3166 && hints
&& e
->maybe_hot_p ())
3167 *hints
|= INLINE_HINT_indirect_call
;
3168 cur_size
= call_size
* INLINE_SIZE_SCALE
;
3171 *min_size
+= cur_size
;
3172 *time
+= apply_probability ((gcov_type
) call_time
, prob
)
3173 * e
->frequency
* (INLINE_TIME_SCALE
/ CGRAPH_FREQ_BASE
);
3174 if (*time
> MAX_TIME
* INLINE_TIME_SCALE
)
3175 *time
= MAX_TIME
* INLINE_TIME_SCALE
;
3180 /* Increase SIZE, MIN_SIZE and TIME for size and time needed to handle all
3181 calls in NODE. POSSIBLE_TRUTHS, KNOWN_VALS, KNOWN_AGGS and KNOWN_CONTEXTS
3182 describe context of the call site. */
3185 estimate_calls_size_and_time (struct cgraph_node
*node
, int *size
,
3186 int *min_size
, int *time
,
3187 inline_hints
*hints
,
3188 clause_t possible_truths
,
3189 vec
<tree
> known_vals
,
3190 vec
<ipa_polymorphic_call_context
> known_contexts
,
3191 vec
<ipa_agg_jump_function_p
> known_aggs
)
3193 struct cgraph_edge
*e
;
3194 for (e
= node
->callees
; e
; e
= e
->next_callee
)
3196 if (inline_edge_summary_vec
.length () <= (unsigned) e
->uid
)
3199 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3201 /* Do not care about zero sized builtins. */
3202 if (e
->inline_failed
&& !es
->call_stmt_size
)
3204 gcc_checking_assert (!es
->call_stmt_time
);
3208 || evaluate_predicate (es
->predicate
, possible_truths
))
3210 if (e
->inline_failed
)
3212 /* Predicates of calls shall not use NOT_CHANGED codes,
3213 sowe do not need to compute probabilities. */
3214 estimate_edge_size_and_time (e
, size
,
3215 es
->predicate
? NULL
: min_size
,
3216 time
, REG_BR_PROB_BASE
,
3217 known_vals
, known_contexts
,
3221 estimate_calls_size_and_time (e
->callee
, size
, min_size
, time
,
3224 known_vals
, known_contexts
,
3228 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
3230 if (inline_edge_summary_vec
.length () <= (unsigned) e
->uid
)
3233 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3235 || evaluate_predicate (es
->predicate
, possible_truths
))
3236 estimate_edge_size_and_time (e
, size
,
3237 es
->predicate
? NULL
: min_size
,
3238 time
, REG_BR_PROB_BASE
,
3239 known_vals
, known_contexts
, known_aggs
,
3245 /* Estimate size and time needed to execute NODE assuming
3246 POSSIBLE_TRUTHS clause, and KNOWN_VALS, KNOWN_AGGS and KNOWN_CONTEXTS
3247 information about NODE's arguments. If non-NULL use also probability
3248 information present in INLINE_PARAM_SUMMARY vector.
3249 Additionally detemine hints determined by the context. Finally compute
3250 minimal size needed for the call that is independent on the call context and
3251 can be used for fast estimates. Return the values in RET_SIZE,
3252 RET_MIN_SIZE, RET_TIME and RET_HINTS. */
3255 estimate_node_size_and_time (struct cgraph_node
*node
,
3256 clause_t possible_truths
,
3257 vec
<tree
> known_vals
,
3258 vec
<ipa_polymorphic_call_context
> known_contexts
,
3259 vec
<ipa_agg_jump_function_p
> known_aggs
,
3260 int *ret_size
, int *ret_min_size
, int *ret_time
,
3261 inline_hints
*ret_hints
,
3262 vec
<inline_param_summary
>
3263 inline_param_summary
)
3265 struct inline_summary
*info
= inline_summaries
->get (node
);
3270 inline_hints hints
= 0;
3273 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3276 fprintf (dump_file
, " Estimating body: %s/%i\n"
3277 " Known to be false: ", node
->name (),
3280 for (i
= predicate_not_inlined_condition
;
3281 i
< (predicate_first_dynamic_condition
3282 + (int) vec_safe_length (info
->conds
)); i
++)
3283 if (!(possible_truths
& (1 << i
)))
3286 fprintf (dump_file
, ", ");
3288 dump_condition (dump_file
, info
->conds
, i
);
3292 for (i
= 0; vec_safe_iterate (info
->entry
, i
, &e
); i
++)
3293 if (evaluate_predicate (&e
->predicate
, possible_truths
))
3296 gcc_checking_assert (e
->time
>= 0);
3297 gcc_checking_assert (time
>= 0);
3298 if (!inline_param_summary
.exists ())
3302 int prob
= predicate_probability (info
->conds
,
3305 inline_param_summary
);
3306 gcc_checking_assert (prob
>= 0);
3307 gcc_checking_assert (prob
<= REG_BR_PROB_BASE
);
3308 time
+= apply_probability ((gcov_type
) e
->time
, prob
);
3310 if (time
> MAX_TIME
* INLINE_TIME_SCALE
)
3311 time
= MAX_TIME
* INLINE_TIME_SCALE
;
3312 gcc_checking_assert (time
>= 0);
3315 gcc_checking_assert (true_predicate_p (&(*info
->entry
)[0].predicate
));
3316 min_size
= (*info
->entry
)[0].size
;
3317 gcc_checking_assert (size
>= 0);
3318 gcc_checking_assert (time
>= 0);
3320 if (info
->loop_iterations
3321 && !evaluate_predicate (info
->loop_iterations
, possible_truths
))
3322 hints
|= INLINE_HINT_loop_iterations
;
3323 if (info
->loop_stride
3324 && !evaluate_predicate (info
->loop_stride
, possible_truths
))
3325 hints
|= INLINE_HINT_loop_stride
;
3326 if (info
->array_index
3327 && !evaluate_predicate (info
->array_index
, possible_truths
))
3328 hints
|= INLINE_HINT_array_index
;
3330 hints
|= INLINE_HINT_in_scc
;
3331 if (DECL_DECLARED_INLINE_P (node
->decl
))
3332 hints
|= INLINE_HINT_declared_inline
;
3334 estimate_calls_size_and_time (node
, &size
, &min_size
, &time
, &hints
, possible_truths
,
3335 known_vals
, known_contexts
, known_aggs
);
3336 gcc_checking_assert (size
>= 0);
3337 gcc_checking_assert (time
>= 0);
3338 time
= RDIV (time
, INLINE_TIME_SCALE
);
3339 size
= RDIV (size
, INLINE_SIZE_SCALE
);
3340 min_size
= RDIV (min_size
, INLINE_SIZE_SCALE
);
3342 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3343 fprintf (dump_file
, "\n size:%i time:%i\n", (int) size
, (int) time
);
3349 *ret_min_size
= min_size
;
3356 /* Estimate size and time needed to execute callee of EDGE assuming that
3357 parameters known to be constant at caller of EDGE are propagated.
3358 KNOWN_VALS and KNOWN_CONTEXTS are vectors of assumed known constant values
3359 and types for parameters. */
3362 estimate_ipcp_clone_size_and_time (struct cgraph_node
*node
,
3363 vec
<tree
> known_vals
,
3364 vec
<ipa_polymorphic_call_context
>
3366 vec
<ipa_agg_jump_function_p
> known_aggs
,
3367 int *ret_size
, int *ret_time
,
3368 inline_hints
*hints
)
3372 clause
= evaluate_conditions_for_known_args (node
, false, known_vals
,
3374 estimate_node_size_and_time (node
, clause
, known_vals
, known_contexts
,
3375 known_aggs
, ret_size
, NULL
, ret_time
, hints
, vNULL
);
3378 /* Translate all conditions from callee representation into caller
3379 representation and symbolically evaluate predicate P into new predicate.
3381 INFO is inline_summary of function we are adding predicate into, CALLEE_INFO
3382 is summary of function predicate P is from. OPERAND_MAP is array giving
3383 callee formal IDs the caller formal IDs. POSSSIBLE_TRUTHS is clausule of all
3384 callee conditions that may be true in caller context. TOPLEV_PREDICATE is
3385 predicate under which callee is executed. OFFSET_MAP is an array of of
3386 offsets that need to be added to conditions, negative offset means that
3387 conditions relying on values passed by reference have to be discarded
3388 because they might not be preserved (and should be considered offset zero
3389 for other purposes). */
3391 static struct predicate
3392 remap_predicate (struct inline_summary
*info
,
3393 struct inline_summary
*callee_info
,
3394 struct predicate
*p
,
3395 vec
<int> operand_map
,
3396 vec
<int> offset_map
,
3397 clause_t possible_truths
, struct predicate
*toplev_predicate
)
3400 struct predicate out
= true_predicate ();
3402 /* True predicate is easy. */
3403 if (true_predicate_p (p
))
3404 return *toplev_predicate
;
3405 for (i
= 0; p
->clause
[i
]; i
++)
3407 clause_t clause
= p
->clause
[i
];
3409 struct predicate clause_predicate
= false_predicate ();
3411 gcc_assert (i
< MAX_CLAUSES
);
3413 for (cond
= 0; cond
< NUM_CONDITIONS
; cond
++)
3414 /* Do we have condition we can't disprove? */
3415 if (clause
& possible_truths
& (1 << cond
))
3417 struct predicate cond_predicate
;
3418 /* Work out if the condition can translate to predicate in the
3419 inlined function. */
3420 if (cond
>= predicate_first_dynamic_condition
)
3422 struct condition
*c
;
3424 c
= &(*callee_info
->conds
)[cond
3426 predicate_first_dynamic_condition
];
3427 /* See if we can remap condition operand to caller's operand.
3428 Otherwise give up. */
3429 if (!operand_map
.exists ()
3430 || (int) operand_map
.length () <= c
->operand_num
3431 || operand_map
[c
->operand_num
] == -1
3432 /* TODO: For non-aggregate conditions, adding an offset is
3433 basically an arithmetic jump function processing which
3434 we should support in future. */
3435 || ((!c
->agg_contents
|| !c
->by_ref
)
3436 && offset_map
[c
->operand_num
] > 0)
3437 || (c
->agg_contents
&& c
->by_ref
3438 && offset_map
[c
->operand_num
] < 0))
3439 cond_predicate
= true_predicate ();
3442 struct agg_position_info ap
;
3443 HOST_WIDE_INT offset_delta
= offset_map
[c
->operand_num
];
3444 if (offset_delta
< 0)
3446 gcc_checking_assert (!c
->agg_contents
|| !c
->by_ref
);
3449 gcc_assert (!c
->agg_contents
3450 || c
->by_ref
|| offset_delta
== 0);
3451 ap
.offset
= c
->offset
+ offset_delta
;
3452 ap
.agg_contents
= c
->agg_contents
;
3453 ap
.by_ref
= c
->by_ref
;
3454 cond_predicate
= add_condition (info
,
3455 operand_map
[c
->operand_num
],
3456 c
->size
, &ap
, c
->code
,
3460 /* Fixed conditions remains same, construct single
3461 condition predicate. */
3464 cond_predicate
.clause
[0] = 1 << cond
;
3465 cond_predicate
.clause
[1] = 0;
3467 clause_predicate
= or_predicates (info
->conds
, &clause_predicate
,
3470 out
= and_predicates (info
->conds
, &out
, &clause_predicate
);
3472 return and_predicates (info
->conds
, &out
, toplev_predicate
);
3476 /* Update summary information of inline clones after inlining.
3477 Compute peak stack usage. */
3480 inline_update_callee_summaries (struct cgraph_node
*node
, int depth
)
3482 struct cgraph_edge
*e
;
3483 struct inline_summary
*callee_info
= inline_summaries
->get (node
);
3484 struct inline_summary
*caller_info
= inline_summaries
->get (node
->callers
->caller
);
3487 callee_info
->stack_frame_offset
3488 = caller_info
->stack_frame_offset
3489 + caller_info
->estimated_self_stack_size
;
3490 peak
= callee_info
->stack_frame_offset
3491 + callee_info
->estimated_self_stack_size
;
3492 if (inline_summaries
->get (node
->global
.inlined_to
)->estimated_stack_size
< peak
)
3493 inline_summaries
->get (node
->global
.inlined_to
)->estimated_stack_size
= peak
;
3494 ipa_propagate_frequency (node
);
3495 for (e
= node
->callees
; e
; e
= e
->next_callee
)
3497 if (!e
->inline_failed
)
3498 inline_update_callee_summaries (e
->callee
, depth
);
3499 inline_edge_summary (e
)->loop_depth
+= depth
;
3501 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
3502 inline_edge_summary (e
)->loop_depth
+= depth
;
3505 /* Update change_prob of EDGE after INLINED_EDGE has been inlined.
3506 When functoin A is inlined in B and A calls C with parameter that
3507 changes with probability PROB1 and C is known to be passthroug
3508 of argument if B that change with probability PROB2, the probability
3509 of change is now PROB1*PROB2. */
3512 remap_edge_change_prob (struct cgraph_edge
*inlined_edge
,
3513 struct cgraph_edge
*edge
)
3515 if (ipa_node_params_sum
)
3518 struct ipa_edge_args
*args
= IPA_EDGE_REF (edge
);
3519 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3520 struct inline_edge_summary
*inlined_es
3521 = inline_edge_summary (inlined_edge
);
3523 for (i
= 0; i
< ipa_get_cs_argument_count (args
); i
++)
3525 struct ipa_jump_func
*jfunc
= ipa_get_ith_jump_func (args
, i
);
3526 if (jfunc
->type
== IPA_JF_PASS_THROUGH
3527 && (ipa_get_jf_pass_through_formal_id (jfunc
)
3528 < (int) inlined_es
->param
.length ()))
3530 int jf_formal_id
= ipa_get_jf_pass_through_formal_id (jfunc
);
3531 int prob1
= es
->param
[i
].change_prob
;
3532 int prob2
= inlined_es
->param
[jf_formal_id
].change_prob
;
3533 int prob
= combine_probabilities (prob1
, prob2
);
3535 if (prob1
&& prob2
&& !prob
)
3538 es
->param
[i
].change_prob
= prob
;
3544 /* Update edge summaries of NODE after INLINED_EDGE has been inlined.
3546 Remap predicates of callees of NODE. Rest of arguments match
3549 Also update change probabilities. */
3552 remap_edge_summaries (struct cgraph_edge
*inlined_edge
,
3553 struct cgraph_node
*node
,
3554 struct inline_summary
*info
,
3555 struct inline_summary
*callee_info
,
3556 vec
<int> operand_map
,
3557 vec
<int> offset_map
,
3558 clause_t possible_truths
,
3559 struct predicate
*toplev_predicate
)
3561 struct cgraph_edge
*e
, *next
;
3562 for (e
= node
->callees
; e
; e
= next
)
3564 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3566 next
= e
->next_callee
;
3568 if (e
->inline_failed
)
3570 remap_edge_change_prob (inlined_edge
, e
);
3574 p
= remap_predicate (info
, callee_info
,
3575 es
->predicate
, operand_map
, offset_map
,
3576 possible_truths
, toplev_predicate
);
3577 edge_set_predicate (e
, &p
);
3580 edge_set_predicate (e
, toplev_predicate
);
3583 remap_edge_summaries (inlined_edge
, e
->callee
, info
, callee_info
,
3584 operand_map
, offset_map
, possible_truths
,
3587 for (e
= node
->indirect_calls
; e
; e
= next
)
3589 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3591 next
= e
->next_callee
;
3593 remap_edge_change_prob (inlined_edge
, e
);
3596 p
= remap_predicate (info
, callee_info
,
3597 es
->predicate
, operand_map
, offset_map
,
3598 possible_truths
, toplev_predicate
);
3599 edge_set_predicate (e
, &p
);
3602 edge_set_predicate (e
, toplev_predicate
);
3606 /* Same as remap_predicate, but set result into hint *HINT. */
3609 remap_hint_predicate (struct inline_summary
*info
,
3610 struct inline_summary
*callee_info
,
3611 struct predicate
**hint
,
3612 vec
<int> operand_map
,
3613 vec
<int> offset_map
,
3614 clause_t possible_truths
,
3615 struct predicate
*toplev_predicate
)
3621 p
= remap_predicate (info
, callee_info
,
3623 operand_map
, offset_map
,
3624 possible_truths
, toplev_predicate
);
3625 if (!false_predicate_p (&p
) && !true_predicate_p (&p
))
3628 set_hint_predicate (hint
, p
);
3630 **hint
= and_predicates (info
->conds
, *hint
, &p
);
3634 /* We inlined EDGE. Update summary of the function we inlined into. */
3637 inline_merge_summary (struct cgraph_edge
*edge
)
3639 struct inline_summary
*callee_info
= inline_summaries
->get (edge
->callee
);
3640 struct cgraph_node
*to
= (edge
->caller
->global
.inlined_to
3641 ? edge
->caller
->global
.inlined_to
: edge
->caller
);
3642 struct inline_summary
*info
= inline_summaries
->get (to
);
3643 clause_t clause
= 0; /* not_inline is known to be false. */
3645 vec
<int> operand_map
= vNULL
;
3646 vec
<int> offset_map
= vNULL
;
3648 struct predicate toplev_predicate
;
3649 struct predicate true_p
= true_predicate ();
3650 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3653 toplev_predicate
= *es
->predicate
;
3655 toplev_predicate
= true_predicate ();
3657 info
->fp_expressions
|= callee_info
->fp_expressions
;
3659 if (callee_info
->conds
)
3660 evaluate_properties_for_edge (edge
, true, &clause
, NULL
, NULL
, NULL
);
3661 if (ipa_node_params_sum
&& callee_info
->conds
)
3663 struct ipa_edge_args
*args
= IPA_EDGE_REF (edge
);
3664 int count
= ipa_get_cs_argument_count (args
);
3669 operand_map
.safe_grow_cleared (count
);
3670 offset_map
.safe_grow_cleared (count
);
3672 for (i
= 0; i
< count
; i
++)
3674 struct ipa_jump_func
*jfunc
= ipa_get_ith_jump_func (args
, i
);
3677 /* TODO: handle non-NOPs when merging. */
3678 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
3680 if (ipa_get_jf_pass_through_operation (jfunc
) == NOP_EXPR
)
3681 map
= ipa_get_jf_pass_through_formal_id (jfunc
);
3682 if (!ipa_get_jf_pass_through_agg_preserved (jfunc
))
3685 else if (jfunc
->type
== IPA_JF_ANCESTOR
)
3687 HOST_WIDE_INT offset
= ipa_get_jf_ancestor_offset (jfunc
);
3688 if (offset
>= 0 && offset
< INT_MAX
)
3690 map
= ipa_get_jf_ancestor_formal_id (jfunc
);
3691 if (!ipa_get_jf_ancestor_agg_preserved (jfunc
))
3693 offset_map
[i
] = offset
;
3696 operand_map
[i
] = map
;
3697 gcc_assert (map
< ipa_get_param_count (IPA_NODE_REF (to
)));
3700 for (i
= 0; vec_safe_iterate (callee_info
->entry
, i
, &e
); i
++)
3702 struct predicate p
= remap_predicate (info
, callee_info
,
3703 &e
->predicate
, operand_map
,
3706 if (!false_predicate_p (&p
))
3708 gcov_type add_time
= ((gcov_type
) e
->time
* edge
->frequency
3709 + CGRAPH_FREQ_BASE
/ 2) / CGRAPH_FREQ_BASE
;
3710 int prob
= predicate_probability (callee_info
->conds
,
3713 add_time
= apply_probability ((gcov_type
) add_time
, prob
);
3714 if (add_time
> MAX_TIME
* INLINE_TIME_SCALE
)
3715 add_time
= MAX_TIME
* INLINE_TIME_SCALE
;
3716 if (prob
!= REG_BR_PROB_BASE
3717 && dump_file
&& (dump_flags
& TDF_DETAILS
))
3719 fprintf (dump_file
, "\t\tScaling time by probability:%f\n",
3720 (double) prob
/ REG_BR_PROB_BASE
);
3722 account_size_time (info
, e
->size
, add_time
, &p
);
3725 remap_edge_summaries (edge
, edge
->callee
, info
, callee_info
, operand_map
,
3726 offset_map
, clause
, &toplev_predicate
);
3727 remap_hint_predicate (info
, callee_info
,
3728 &callee_info
->loop_iterations
,
3729 operand_map
, offset_map
, clause
, &toplev_predicate
);
3730 remap_hint_predicate (info
, callee_info
,
3731 &callee_info
->loop_stride
,
3732 operand_map
, offset_map
, clause
, &toplev_predicate
);
3733 remap_hint_predicate (info
, callee_info
,
3734 &callee_info
->array_index
,
3735 operand_map
, offset_map
, clause
, &toplev_predicate
);
3737 inline_update_callee_summaries (edge
->callee
,
3738 inline_edge_summary (edge
)->loop_depth
);
3740 /* We do not maintain predicates of inlined edges, free it. */
3741 edge_set_predicate (edge
, &true_p
);
3742 /* Similarly remove param summaries. */
3743 es
->param
.release ();
3744 operand_map
.release ();
3745 offset_map
.release ();
3748 /* For performance reasons inline_merge_summary is not updating overall size
3749 and time. Recompute it. */
3752 inline_update_overall_summary (struct cgraph_node
*node
)
3754 struct inline_summary
*info
= inline_summaries
->get (node
);
3760 for (i
= 0; vec_safe_iterate (info
->entry
, i
, &e
); i
++)
3762 info
->size
+= e
->size
, info
->time
+= e
->time
;
3763 if (info
->time
> MAX_TIME
* INLINE_TIME_SCALE
)
3764 info
->time
= MAX_TIME
* INLINE_TIME_SCALE
;
3766 estimate_calls_size_and_time (node
, &info
->size
, &info
->min_size
,
3768 ~(clause_t
) (1 << predicate_false_condition
),
3769 vNULL
, vNULL
, vNULL
);
3770 info
->time
= (info
->time
+ INLINE_TIME_SCALE
/ 2) / INLINE_TIME_SCALE
;
3771 info
->size
= (info
->size
+ INLINE_SIZE_SCALE
/ 2) / INLINE_SIZE_SCALE
;
3774 /* Return hints derrived from EDGE. */
3776 simple_edge_hints (struct cgraph_edge
*edge
)
3779 struct cgraph_node
*to
= (edge
->caller
->global
.inlined_to
3780 ? edge
->caller
->global
.inlined_to
: edge
->caller
);
3781 struct cgraph_node
*callee
= edge
->callee
->ultimate_alias_target ();
3782 if (inline_summaries
->get (to
)->scc_no
3783 && inline_summaries
->get (to
)->scc_no
3784 == inline_summaries
->get (callee
)->scc_no
3785 && !edge
->recursive_p ())
3786 hints
|= INLINE_HINT_same_scc
;
3788 if (callee
->lto_file_data
&& edge
->caller
->lto_file_data
3789 && edge
->caller
->lto_file_data
!= callee
->lto_file_data
3790 && !callee
->merged_comdat
&& !callee
->icf_merged
)
3791 hints
|= INLINE_HINT_cross_module
;
3796 /* Estimate the time cost for the caller when inlining EDGE.
3797 Only to be called via estimate_edge_time, that handles the
3800 When caching, also update the cache entry. Compute both time and
3801 size, since we always need both metrics eventually. */
3804 do_estimate_edge_time (struct cgraph_edge
*edge
)
3809 struct cgraph_node
*callee
;
3811 vec
<tree
> known_vals
;
3812 vec
<ipa_polymorphic_call_context
> known_contexts
;
3813 vec
<ipa_agg_jump_function_p
> known_aggs
;
3814 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3817 callee
= edge
->callee
->ultimate_alias_target ();
3819 gcc_checking_assert (edge
->inline_failed
);
3820 evaluate_properties_for_edge (edge
, true,
3821 &clause
, &known_vals
, &known_contexts
,
3823 estimate_node_size_and_time (callee
, clause
, known_vals
, known_contexts
,
3824 known_aggs
, &size
, &min_size
, &time
, &hints
, es
->param
);
3826 /* When we have profile feedback, we can quite safely identify hot
3827 edges and for those we disable size limits. Don't do that when
3828 probability that caller will call the callee is low however, since it
3829 may hurt optimization of the caller's hot path. */
3830 if (edge
->count
&& edge
->maybe_hot_p ()
3832 > (edge
->caller
->global
.inlined_to
3833 ? edge
->caller
->global
.inlined_to
->count
: edge
->caller
->count
)))
3834 hints
|= INLINE_HINT_known_hot
;
3836 known_vals
.release ();
3837 known_contexts
.release ();
3838 known_aggs
.release ();
3839 gcc_checking_assert (size
>= 0);
3840 gcc_checking_assert (time
>= 0);
3842 /* When caching, update the cache entry. */
3843 if (edge_growth_cache
.exists ())
3845 inline_summaries
->get (edge
->callee
)->min_size
= min_size
;
3846 if ((int) edge_growth_cache
.length () <= edge
->uid
)
3847 edge_growth_cache
.safe_grow_cleared (symtab
->edges_max_uid
);
3848 edge_growth_cache
[edge
->uid
].time
= time
+ (time
>= 0);
3850 edge_growth_cache
[edge
->uid
].size
= size
+ (size
>= 0);
3851 hints
|= simple_edge_hints (edge
);
3852 edge_growth_cache
[edge
->uid
].hints
= hints
+ 1;
3858 /* Return estimated callee growth after inlining EDGE.
3859 Only to be called via estimate_edge_size. */
3862 do_estimate_edge_size (struct cgraph_edge
*edge
)
3865 struct cgraph_node
*callee
;
3867 vec
<tree
> known_vals
;
3868 vec
<ipa_polymorphic_call_context
> known_contexts
;
3869 vec
<ipa_agg_jump_function_p
> known_aggs
;
3871 /* When we do caching, use do_estimate_edge_time to populate the entry. */
3873 if (edge_growth_cache
.exists ())
3875 do_estimate_edge_time (edge
);
3876 size
= edge_growth_cache
[edge
->uid
].size
;
3877 gcc_checking_assert (size
);
3878 return size
- (size
> 0);
3881 callee
= edge
->callee
->ultimate_alias_target ();
3883 /* Early inliner runs without caching, go ahead and do the dirty work. */
3884 gcc_checking_assert (edge
->inline_failed
);
3885 evaluate_properties_for_edge (edge
, true,
3886 &clause
, &known_vals
, &known_contexts
,
3888 estimate_node_size_and_time (callee
, clause
, known_vals
, known_contexts
,
3889 known_aggs
, &size
, NULL
, NULL
, NULL
, vNULL
);
3890 known_vals
.release ();
3891 known_contexts
.release ();
3892 known_aggs
.release ();
3897 /* Estimate the growth of the caller when inlining EDGE.
3898 Only to be called via estimate_edge_size. */
3901 do_estimate_edge_hints (struct cgraph_edge
*edge
)
3904 struct cgraph_node
*callee
;
3906 vec
<tree
> known_vals
;
3907 vec
<ipa_polymorphic_call_context
> known_contexts
;
3908 vec
<ipa_agg_jump_function_p
> known_aggs
;
3910 /* When we do caching, use do_estimate_edge_time to populate the entry. */
3912 if (edge_growth_cache
.exists ())
3914 do_estimate_edge_time (edge
);
3915 hints
= edge_growth_cache
[edge
->uid
].hints
;
3916 gcc_checking_assert (hints
);
3920 callee
= edge
->callee
->ultimate_alias_target ();
3922 /* Early inliner runs without caching, go ahead and do the dirty work. */
3923 gcc_checking_assert (edge
->inline_failed
);
3924 evaluate_properties_for_edge (edge
, true,
3925 &clause
, &known_vals
, &known_contexts
,
3927 estimate_node_size_and_time (callee
, clause
, known_vals
, known_contexts
,
3928 known_aggs
, NULL
, NULL
, NULL
, &hints
, vNULL
);
3929 known_vals
.release ();
3930 known_contexts
.release ();
3931 known_aggs
.release ();
3932 hints
|= simple_edge_hints (edge
);
3937 /* Estimate self time of the function NODE after inlining EDGE. */
3940 estimate_time_after_inlining (struct cgraph_node
*node
,
3941 struct cgraph_edge
*edge
)
3943 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3944 if (!es
->predicate
|| !false_predicate_p (es
->predicate
))
3947 inline_summaries
->get (node
)->time
+ estimate_edge_time (edge
);
3950 if (time
> MAX_TIME
)
3954 return inline_summaries
->get (node
)->time
;
3958 /* Estimate the size of NODE after inlining EDGE which should be an
3959 edge to either NODE or a call inlined into NODE. */
3962 estimate_size_after_inlining (struct cgraph_node
*node
,
3963 struct cgraph_edge
*edge
)
3965 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3966 if (!es
->predicate
|| !false_predicate_p (es
->predicate
))
3968 int size
= inline_summaries
->get (node
)->size
+ estimate_edge_growth (edge
);
3969 gcc_assert (size
>= 0);
3972 return inline_summaries
->get (node
)->size
;
3978 struct cgraph_node
*node
;
3979 bool self_recursive
;
3985 /* Worker for do_estimate_growth. Collect growth for all callers. */
3988 do_estimate_growth_1 (struct cgraph_node
*node
, void *data
)
3990 struct cgraph_edge
*e
;
3991 struct growth_data
*d
= (struct growth_data
*) data
;
3993 for (e
= node
->callers
; e
; e
= e
->next_caller
)
3995 gcc_checking_assert (e
->inline_failed
);
3997 if (cgraph_inline_failed_type (e
->inline_failed
) == CIF_FINAL_ERROR
)
3999 d
->uninlinable
= true;
4003 if (e
->recursive_p ())
4005 d
->self_recursive
= true;
4008 d
->growth
+= estimate_edge_growth (e
);
4014 /* Estimate the growth caused by inlining NODE into all callees. */
4017 estimate_growth (struct cgraph_node
*node
)
4019 struct growth_data d
= { node
, false, false, 0 };
4020 struct inline_summary
*info
= inline_summaries
->get (node
);
4022 node
->call_for_symbol_and_aliases (do_estimate_growth_1
, &d
, true);
4024 /* For self recursive functions the growth estimation really should be
4025 infinity. We don't want to return very large values because the growth
4026 plays various roles in badness computation fractions. Be sure to not
4027 return zero or negative growths. */
4028 if (d
.self_recursive
)
4029 d
.growth
= d
.growth
< info
->size
? info
->size
: d
.growth
;
4030 else if (DECL_EXTERNAL (node
->decl
) || d
.uninlinable
)
4034 if (node
->will_be_removed_from_program_if_no_direct_calls_p ())
4035 d
.growth
-= info
->size
;
4036 /* COMDAT functions are very often not shared across multiple units
4037 since they come from various template instantiations.
4038 Take this into account. */
4039 else if (DECL_COMDAT (node
->decl
)
4040 && node
->can_remove_if_no_direct_calls_p ())
4041 d
.growth
-= (info
->size
4042 * (100 - PARAM_VALUE (PARAM_COMDAT_SHARING_PROBABILITY
))
4049 /* Verify if there are fewer than MAX_CALLERS. */
4052 check_callers (cgraph_node
*node
, int *max_callers
)
4056 if (!node
->can_remove_if_no_direct_calls_and_refs_p ())
4059 for (cgraph_edge
*e
= node
->callers
; e
; e
= e
->next_caller
)
4063 || cgraph_inline_failed_type (e
->inline_failed
) == CIF_FINAL_ERROR
)
4067 FOR_EACH_ALIAS (node
, ref
)
4068 if (check_callers (dyn_cast
<cgraph_node
*> (ref
->referring
), max_callers
))
4075 /* Make cheap estimation if growth of NODE is likely positive knowing
4076 EDGE_GROWTH of one particular edge.
4077 We assume that most of other edges will have similar growth
4078 and skip computation if there are too many callers. */
4081 growth_likely_positive (struct cgraph_node
*node
,
4085 struct cgraph_edge
*e
;
4086 gcc_checking_assert (edge_growth
> 0);
4088 /* First quickly check if NODE is removable at all. */
4089 if (DECL_EXTERNAL (node
->decl
))
4091 if (!node
->can_remove_if_no_direct_calls_and_refs_p ()
4092 || node
->address_taken
)
4095 max_callers
= inline_summaries
->get (node
)->size
* 4 / edge_growth
+ 2;
4097 for (e
= node
->callers
; e
; e
= e
->next_caller
)
4101 || cgraph_inline_failed_type (e
->inline_failed
) == CIF_FINAL_ERROR
)
4106 FOR_EACH_ALIAS (node
, ref
)
4107 if (check_callers (dyn_cast
<cgraph_node
*> (ref
->referring
), &max_callers
))
4110 /* Unlike for functions called once, we play unsafe with
4111 COMDATs. We can allow that since we know functions
4112 in consideration are small (and thus risk is small) and
4113 moreover grow estimates already accounts that COMDAT
4114 functions may or may not disappear when eliminated from
4115 current unit. With good probability making aggressive
4116 choice in all units is going to make overall program
4118 if (DECL_COMDAT (node
->decl
))
4120 if (!node
->can_remove_if_no_direct_calls_p ())
4123 else if (!node
->will_be_removed_from_program_if_no_direct_calls_p ())
4126 return estimate_growth (node
) > 0;
4130 /* This function performs intraprocedural analysis in NODE that is required to
4131 inline indirect calls. */
4134 inline_indirect_intraprocedural_analysis (struct cgraph_node
*node
)
4136 ipa_analyze_node (node
);
4137 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4139 ipa_print_node_params (dump_file
, node
);
4140 ipa_print_node_jump_functions (dump_file
, node
);
4145 /* Note function body size. */
4148 inline_analyze_function (struct cgraph_node
*node
)
4150 push_cfun (DECL_STRUCT_FUNCTION (node
->decl
));
4153 fprintf (dump_file
, "\nAnalyzing function: %s/%u\n",
4154 node
->name (), node
->order
);
4155 if (opt_for_fn (node
->decl
, optimize
) && !node
->thunk
.thunk_p
)
4156 inline_indirect_intraprocedural_analysis (node
);
4157 compute_inline_parameters (node
, false);
4160 struct cgraph_edge
*e
;
4161 for (e
= node
->callees
; e
; e
= e
->next_callee
)
4162 e
->inline_failed
= CIF_FUNCTION_NOT_OPTIMIZED
;
4163 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
4164 e
->inline_failed
= CIF_FUNCTION_NOT_OPTIMIZED
;
4171 /* Called when new function is inserted to callgraph late. */
4174 inline_summary_t::insert (struct cgraph_node
*node
, inline_summary
*)
4176 inline_analyze_function (node
);
4179 /* Note function body size. */
4182 inline_generate_summary (void)
4184 struct cgraph_node
*node
;
4186 FOR_EACH_DEFINED_FUNCTION (node
)
4187 if (DECL_STRUCT_FUNCTION (node
->decl
))
4188 node
->local
.versionable
= tree_versionable_function_p (node
->decl
);
4190 /* When not optimizing, do not bother to analyze. Inlining is still done
4191 because edge redirection needs to happen there. */
4192 if (!optimize
&& !flag_generate_lto
&& !flag_generate_offload
&& !flag_wpa
)
4195 if (!inline_summaries
)
4196 inline_summaries
= (inline_summary_t
*) inline_summary_t::create_ggc (symtab
);
4198 inline_summaries
->enable_insertion_hook ();
4200 ipa_register_cgraph_hooks ();
4201 inline_free_summary ();
4203 FOR_EACH_DEFINED_FUNCTION (node
)
4205 inline_analyze_function (node
);
4209 /* Read predicate from IB. */
4211 static struct predicate
4212 read_predicate (struct lto_input_block
*ib
)
4214 struct predicate out
;
4220 gcc_assert (k
<= MAX_CLAUSES
);
4221 clause
= out
.clause
[k
++] = streamer_read_uhwi (ib
);
4225 /* Zero-initialize the remaining clauses in OUT. */
4226 while (k
<= MAX_CLAUSES
)
4227 out
.clause
[k
++] = 0;
4233 /* Write inline summary for edge E to OB. */
4236 read_inline_edge_summary (struct lto_input_block
*ib
, struct cgraph_edge
*e
)
4238 struct inline_edge_summary
*es
= inline_edge_summary (e
);
4242 es
->call_stmt_size
= streamer_read_uhwi (ib
);
4243 es
->call_stmt_time
= streamer_read_uhwi (ib
);
4244 es
->loop_depth
= streamer_read_uhwi (ib
);
4245 p
= read_predicate (ib
);
4246 edge_set_predicate (e
, &p
);
4247 length
= streamer_read_uhwi (ib
);
4250 es
->param
.safe_grow_cleared (length
);
4251 for (i
= 0; i
< length
; i
++)
4252 es
->param
[i
].change_prob
= streamer_read_uhwi (ib
);
4257 /* Stream in inline summaries from the section. */
4260 inline_read_section (struct lto_file_decl_data
*file_data
, const char *data
,
4263 const struct lto_function_header
*header
=
4264 (const struct lto_function_header
*) data
;
4265 const int cfg_offset
= sizeof (struct lto_function_header
);
4266 const int main_offset
= cfg_offset
+ header
->cfg_size
;
4267 const int string_offset
= main_offset
+ header
->main_size
;
4268 struct data_in
*data_in
;
4269 unsigned int i
, count2
, j
;
4270 unsigned int f_count
;
4272 lto_input_block
ib ((const char *) data
+ main_offset
, header
->main_size
,
4273 file_data
->mode_table
);
4276 lto_data_in_create (file_data
, (const char *) data
+ string_offset
,
4277 header
->string_size
, vNULL
);
4278 f_count
= streamer_read_uhwi (&ib
);
4279 for (i
= 0; i
< f_count
; i
++)
4282 struct cgraph_node
*node
;
4283 struct inline_summary
*info
;
4284 lto_symtab_encoder_t encoder
;
4285 struct bitpack_d bp
;
4286 struct cgraph_edge
*e
;
4289 index
= streamer_read_uhwi (&ib
);
4290 encoder
= file_data
->symtab_node_encoder
;
4291 node
= dyn_cast
<cgraph_node
*> (lto_symtab_encoder_deref (encoder
,
4293 info
= inline_summaries
->get (node
);
4295 info
->estimated_stack_size
4296 = info
->estimated_self_stack_size
= streamer_read_uhwi (&ib
);
4297 info
->size
= info
->self_size
= streamer_read_uhwi (&ib
);
4298 info
->time
= info
->self_time
= streamer_read_uhwi (&ib
);
4300 bp
= streamer_read_bitpack (&ib
);
4301 info
->inlinable
= bp_unpack_value (&bp
, 1);
4302 info
->contains_cilk_spawn
= bp_unpack_value (&bp
, 1);
4303 info
->fp_expressions
= bp_unpack_value (&bp
, 1);
4305 count2
= streamer_read_uhwi (&ib
);
4306 gcc_assert (!info
->conds
);
4307 for (j
= 0; j
< count2
; j
++)
4310 c
.operand_num
= streamer_read_uhwi (&ib
);
4311 c
.size
= streamer_read_uhwi (&ib
);
4312 c
.code
= (enum tree_code
) streamer_read_uhwi (&ib
);
4313 c
.val
= stream_read_tree (&ib
, data_in
);
4314 bp
= streamer_read_bitpack (&ib
);
4315 c
.agg_contents
= bp_unpack_value (&bp
, 1);
4316 c
.by_ref
= bp_unpack_value (&bp
, 1);
4318 c
.offset
= streamer_read_uhwi (&ib
);
4319 vec_safe_push (info
->conds
, c
);
4321 count2
= streamer_read_uhwi (&ib
);
4322 gcc_assert (!info
->entry
);
4323 for (j
= 0; j
< count2
; j
++)
4325 struct size_time_entry e
;
4327 e
.size
= streamer_read_uhwi (&ib
);
4328 e
.time
= streamer_read_uhwi (&ib
);
4329 e
.predicate
= read_predicate (&ib
);
4331 vec_safe_push (info
->entry
, e
);
4334 p
= read_predicate (&ib
);
4335 set_hint_predicate (&info
->loop_iterations
, p
);
4336 p
= read_predicate (&ib
);
4337 set_hint_predicate (&info
->loop_stride
, p
);
4338 p
= read_predicate (&ib
);
4339 set_hint_predicate (&info
->array_index
, p
);
4340 for (e
= node
->callees
; e
; e
= e
->next_callee
)
4341 read_inline_edge_summary (&ib
, e
);
4342 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
4343 read_inline_edge_summary (&ib
, e
);
4346 lto_free_section_data (file_data
, LTO_section_inline_summary
, NULL
, data
,
4348 lto_data_in_delete (data_in
);
4352 /* Read inline summary. Jump functions are shared among ipa-cp
4353 and inliner, so when ipa-cp is active, we don't need to write them
4357 inline_read_summary (void)
4359 struct lto_file_decl_data
**file_data_vec
= lto_get_file_decl_data ();
4360 struct lto_file_decl_data
*file_data
;
4363 inline_summary_alloc ();
4365 while ((file_data
= file_data_vec
[j
++]))
4368 const char *data
= lto_get_section_data (file_data
,
4369 LTO_section_inline_summary
,
4372 inline_read_section (file_data
, data
, len
);
4374 /* Fatal error here. We do not want to support compiling ltrans units
4375 with different version of compiler or different flags than the WPA
4376 unit, so this should never happen. */
4377 fatal_error (input_location
,
4378 "ipa inline summary is missing in input file");
4382 ipa_register_cgraph_hooks ();
4384 ipa_prop_read_jump_functions ();
4387 gcc_assert (inline_summaries
);
4388 inline_summaries
->enable_insertion_hook ();
4392 /* Write predicate P to OB. */
4395 write_predicate (struct output_block
*ob
, struct predicate
*p
)
4399 for (j
= 0; p
->clause
[j
]; j
++)
4401 gcc_assert (j
< MAX_CLAUSES
);
4402 streamer_write_uhwi (ob
, p
->clause
[j
]);
4404 streamer_write_uhwi (ob
, 0);
4408 /* Write inline summary for edge E to OB. */
4411 write_inline_edge_summary (struct output_block
*ob
, struct cgraph_edge
*e
)
4413 struct inline_edge_summary
*es
= inline_edge_summary (e
);
4416 streamer_write_uhwi (ob
, es
->call_stmt_size
);
4417 streamer_write_uhwi (ob
, es
->call_stmt_time
);
4418 streamer_write_uhwi (ob
, es
->loop_depth
);
4419 write_predicate (ob
, es
->predicate
);
4420 streamer_write_uhwi (ob
, es
->param
.length ());
4421 for (i
= 0; i
< (int) es
->param
.length (); i
++)
4422 streamer_write_uhwi (ob
, es
->param
[i
].change_prob
);
4426 /* Write inline summary for node in SET.
4427 Jump functions are shared among ipa-cp and inliner, so when ipa-cp is
4428 active, we don't need to write them twice. */
4431 inline_write_summary (void)
4433 struct cgraph_node
*node
;
4434 struct output_block
*ob
= create_output_block (LTO_section_inline_summary
);
4435 lto_symtab_encoder_t encoder
= ob
->decl_state
->symtab_node_encoder
;
4436 unsigned int count
= 0;
4439 for (i
= 0; i
< lto_symtab_encoder_size (encoder
); i
++)
4441 symtab_node
*snode
= lto_symtab_encoder_deref (encoder
, i
);
4442 cgraph_node
*cnode
= dyn_cast
<cgraph_node
*> (snode
);
4443 if (cnode
&& cnode
->definition
&& !cnode
->alias
)
4446 streamer_write_uhwi (ob
, count
);
4448 for (i
= 0; i
< lto_symtab_encoder_size (encoder
); i
++)
4450 symtab_node
*snode
= lto_symtab_encoder_deref (encoder
, i
);
4451 cgraph_node
*cnode
= dyn_cast
<cgraph_node
*> (snode
);
4452 if (cnode
&& (node
= cnode
)->definition
&& !node
->alias
)
4454 struct inline_summary
*info
= inline_summaries
->get (node
);
4455 struct bitpack_d bp
;
4456 struct cgraph_edge
*edge
;
4459 struct condition
*c
;
4461 streamer_write_uhwi (ob
,
4462 lto_symtab_encoder_encode (encoder
,
4465 streamer_write_hwi (ob
, info
->estimated_self_stack_size
);
4466 streamer_write_hwi (ob
, info
->self_size
);
4467 streamer_write_hwi (ob
, info
->self_time
);
4468 bp
= bitpack_create (ob
->main_stream
);
4469 bp_pack_value (&bp
, info
->inlinable
, 1);
4470 bp_pack_value (&bp
, info
->contains_cilk_spawn
, 1);
4471 bp_pack_value (&bp
, info
->fp_expressions
, 1);
4472 streamer_write_bitpack (&bp
);
4473 streamer_write_uhwi (ob
, vec_safe_length (info
->conds
));
4474 for (i
= 0; vec_safe_iterate (info
->conds
, i
, &c
); i
++)
4476 streamer_write_uhwi (ob
, c
->operand_num
);
4477 streamer_write_uhwi (ob
, c
->size
);
4478 streamer_write_uhwi (ob
, c
->code
);
4479 stream_write_tree (ob
, c
->val
, true);
4480 bp
= bitpack_create (ob
->main_stream
);
4481 bp_pack_value (&bp
, c
->agg_contents
, 1);
4482 bp_pack_value (&bp
, c
->by_ref
, 1);
4483 streamer_write_bitpack (&bp
);
4484 if (c
->agg_contents
)
4485 streamer_write_uhwi (ob
, c
->offset
);
4487 streamer_write_uhwi (ob
, vec_safe_length (info
->entry
));
4488 for (i
= 0; vec_safe_iterate (info
->entry
, i
, &e
); i
++)
4490 streamer_write_uhwi (ob
, e
->size
);
4491 streamer_write_uhwi (ob
, e
->time
);
4492 write_predicate (ob
, &e
->predicate
);
4494 write_predicate (ob
, info
->loop_iterations
);
4495 write_predicate (ob
, info
->loop_stride
);
4496 write_predicate (ob
, info
->array_index
);
4497 for (edge
= node
->callees
; edge
; edge
= edge
->next_callee
)
4498 write_inline_edge_summary (ob
, edge
);
4499 for (edge
= node
->indirect_calls
; edge
; edge
= edge
->next_callee
)
4500 write_inline_edge_summary (ob
, edge
);
4503 streamer_write_char_stream (ob
->main_stream
, 0);
4504 produce_asm (ob
, NULL
);
4505 destroy_output_block (ob
);
4507 if (optimize
&& !flag_ipa_cp
)
4508 ipa_prop_write_jump_functions ();
4512 /* Release inline summary. */
4515 inline_free_summary (void)
4517 struct cgraph_node
*node
;
4518 if (edge_removal_hook_holder
)
4519 symtab
->remove_edge_removal_hook (edge_removal_hook_holder
);
4520 edge_removal_hook_holder
= NULL
;
4521 if (edge_duplication_hook_holder
)
4522 symtab
->remove_edge_duplication_hook (edge_duplication_hook_holder
);
4523 edge_duplication_hook_holder
= NULL
;
4524 if (!inline_edge_summary_vec
.exists ())
4526 FOR_EACH_DEFINED_FUNCTION (node
)
4528 reset_inline_summary (node
, inline_summaries
->get (node
));
4529 inline_summaries
->release ();
4530 inline_summaries
= NULL
;
4531 inline_edge_summary_vec
.release ();
4532 edge_predicate_pool
.release ();