1 /* Inlining decision heuristics.
2 Copyright (C) 2003-2014 Free Software Foundation, Inc.
3 Contributed by Jan Hubicka
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
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"
72 #include "stor-layout.h"
73 #include "stringpool.h"
74 #include "print-tree.h"
75 #include "tree-inline.h"
76 #include "langhooks.h"
78 #include "diagnostic.h"
79 #include "gimple-pretty-print.h"
81 #include "tree-pass.h"
83 #include "basic-block.h"
84 #include "tree-ssa-alias.h"
85 #include "internal-fn.h"
86 #include "gimple-expr.h"
89 #include "gimple-iterator.h"
90 #include "gimple-ssa.h"
92 #include "tree-phinodes.h"
93 #include "ssa-iterators.h"
94 #include "tree-ssanames.h"
95 #include "tree-ssa-loop-niter.h"
96 #include "tree-ssa-loop.h"
98 #include "lto-streamer.h"
99 #include "data-streamer.h"
100 #include "tree-streamer.h"
101 #include "ipa-inline.h"
102 #include "alloc-pool.h"
104 #include "tree-scalar-evolution.h"
105 #include "ipa-utils.h"
107 #include "cfgexpand.h"
109 /* Estimate runtime of function can easilly run into huge numbers with many
110 nested loops. Be sure we can compute time * INLINE_SIZE_SCALE * 2 in an
111 integer. For anything larger we use gcov_type. */
112 #define MAX_TIME 500000
114 /* Number of bits in integer, but we really want to be stable across different
116 #define NUM_CONDITIONS 32
118 enum predicate_conditions
120 predicate_false_condition
= 0,
121 predicate_not_inlined_condition
= 1,
122 predicate_first_dynamic_condition
= 2
125 /* Special condition code we use to represent test that operand is compile time
127 #define IS_NOT_CONSTANT ERROR_MARK
128 /* Special condition code we use to represent test that operand is not changed
129 across invocation of the function. When operand IS_NOT_CONSTANT it is always
130 CHANGED, however i.e. loop invariants can be NOT_CHANGED given percentage
131 of executions even when they are not compile time constants. */
132 #define CHANGED IDENTIFIER_NODE
134 /* Holders of ipa cgraph hooks: */
135 static struct cgraph_node_hook_list
*function_insertion_hook_holder
;
136 static struct cgraph_node_hook_list
*node_removal_hook_holder
;
137 static struct cgraph_2node_hook_list
*node_duplication_hook_holder
;
138 static struct cgraph_2edge_hook_list
*edge_duplication_hook_holder
;
139 static struct cgraph_edge_hook_list
*edge_removal_hook_holder
;
140 static void inline_node_removal_hook (struct cgraph_node
*, void *);
141 static void inline_node_duplication_hook (struct cgraph_node
*,
142 struct cgraph_node
*, void *);
143 static void inline_edge_removal_hook (struct cgraph_edge
*, void *);
144 static void inline_edge_duplication_hook (struct cgraph_edge
*,
145 struct cgraph_edge
*, void *);
147 /* VECtor holding inline summaries.
148 In GGC memory because conditions might point to constant trees. */
149 vec
<inline_summary_t
, va_gc
> *inline_summary_vec
;
150 vec
<inline_edge_summary_t
> inline_edge_summary_vec
;
152 /* Cached node/edge growths. */
153 vec
<int> node_growth_cache
;
154 vec
<edge_growth_cache_entry
> edge_growth_cache
;
156 /* Edge predicates goes here. */
157 static alloc_pool edge_predicate_pool
;
159 /* Return true predicate (tautology).
160 We represent it by empty list of clauses. */
162 static inline struct predicate
163 true_predicate (void)
171 /* Return predicate testing single condition number COND. */
173 static inline struct predicate
174 single_cond_predicate (int cond
)
177 p
.clause
[0] = 1 << cond
;
183 /* Return false predicate. First clause require false condition. */
185 static inline struct predicate
186 false_predicate (void)
188 return single_cond_predicate (predicate_false_condition
);
192 /* Return true if P is (true). */
195 true_predicate_p (struct predicate
*p
)
197 return !p
->clause
[0];
201 /* Return true if P is (false). */
204 false_predicate_p (struct predicate
*p
)
206 if (p
->clause
[0] == (1 << predicate_false_condition
))
208 gcc_checking_assert (!p
->clause
[1]
209 && p
->clause
[0] == 1 << predicate_false_condition
);
216 /* Return predicate that is set true when function is not inlined. */
218 static inline struct predicate
219 not_inlined_predicate (void)
221 return single_cond_predicate (predicate_not_inlined_condition
);
224 /* Simple description of whether a memory load or a condition refers to a load
225 from an aggregate and if so, how and where from in the aggregate.
226 Individual fields have the same meaning like fields with the same name in
229 struct agg_position_info
231 HOST_WIDE_INT offset
;
236 /* Add condition to condition list CONDS. AGGPOS describes whether the used
237 oprand is loaded from an aggregate and where in the aggregate it is. It can
238 be NULL, which means this not a load from an aggregate. */
240 static struct predicate
241 add_condition (struct inline_summary
*summary
, int operand_num
,
242 struct agg_position_info
*aggpos
,
243 enum tree_code code
, tree val
)
247 struct condition new_cond
;
248 HOST_WIDE_INT offset
;
249 bool agg_contents
, by_ref
;
253 offset
= aggpos
->offset
;
254 agg_contents
= aggpos
->agg_contents
;
255 by_ref
= aggpos
->by_ref
;
260 agg_contents
= false;
264 gcc_checking_assert (operand_num
>= 0);
265 for (i
= 0; vec_safe_iterate (summary
->conds
, i
, &c
); i
++)
267 if (c
->operand_num
== operand_num
270 && c
->agg_contents
== agg_contents
271 && (!agg_contents
|| (c
->offset
== offset
&& c
->by_ref
== by_ref
)))
272 return single_cond_predicate (i
+ predicate_first_dynamic_condition
);
274 /* Too many conditions. Give up and return constant true. */
275 if (i
== NUM_CONDITIONS
- predicate_first_dynamic_condition
)
276 return true_predicate ();
278 new_cond
.operand_num
= operand_num
;
279 new_cond
.code
= code
;
281 new_cond
.agg_contents
= agg_contents
;
282 new_cond
.by_ref
= by_ref
;
283 new_cond
.offset
= offset
;
284 vec_safe_push (summary
->conds
, new_cond
);
285 return single_cond_predicate (i
+ predicate_first_dynamic_condition
);
289 /* Add clause CLAUSE into the predicate P. */
292 add_clause (conditions conditions
, struct predicate
*p
, clause_t clause
)
296 int insert_here
= -1;
303 /* False clause makes the whole predicate false. Kill the other variants. */
304 if (clause
== (1 << predicate_false_condition
))
306 p
->clause
[0] = (1 << predicate_false_condition
);
310 if (false_predicate_p (p
))
313 /* No one should be silly enough to add false into nontrivial clauses. */
314 gcc_checking_assert (!(clause
& (1 << predicate_false_condition
)));
316 /* Look where to insert the clause. At the same time prune out
317 clauses of P that are implied by the new clause and thus
319 for (i
= 0, i2
= 0; i
<= MAX_CLAUSES
; i
++)
321 p
->clause
[i2
] = p
->clause
[i
];
326 /* If p->clause[i] implies clause, there is nothing to add. */
327 if ((p
->clause
[i
] & clause
) == p
->clause
[i
])
329 /* We had nothing to add, none of clauses should've become
331 gcc_checking_assert (i
== i2
);
335 if (p
->clause
[i
] < clause
&& insert_here
< 0)
338 /* If clause implies p->clause[i], then p->clause[i] becomes redundant.
339 Otherwise the p->clause[i] has to stay. */
340 if ((p
->clause
[i
] & clause
) != clause
)
344 /* Look for clauses that are obviously true. I.e.
345 op0 == 5 || op0 != 5. */
346 for (c1
= predicate_first_dynamic_condition
; c1
< NUM_CONDITIONS
; c1
++)
349 if (!(clause
& (1 << c1
)))
351 cc1
= &(*conditions
)[c1
- predicate_first_dynamic_condition
];
352 /* We have no way to represent !CHANGED and !IS_NOT_CONSTANT
353 and thus there is no point for looking for them. */
354 if (cc1
->code
== CHANGED
|| cc1
->code
== IS_NOT_CONSTANT
)
356 for (c2
= c1
+ 1; c2
< NUM_CONDITIONS
; c2
++)
357 if (clause
& (1 << c2
))
360 &(*conditions
)[c1
- predicate_first_dynamic_condition
];
362 &(*conditions
)[c2
- predicate_first_dynamic_condition
];
363 if (cc1
->operand_num
== cc2
->operand_num
364 && cc1
->val
== cc2
->val
365 && cc2
->code
!= IS_NOT_CONSTANT
366 && cc2
->code
!= CHANGED
367 && cc1
->code
== invert_tree_comparison
369 HONOR_NANS (TYPE_MODE (TREE_TYPE (cc1
->val
)))))
375 /* We run out of variants. Be conservative in positive direction. */
376 if (i2
== MAX_CLAUSES
)
378 /* Keep clauses in decreasing order. This makes equivalence testing easy. */
379 p
->clause
[i2
+ 1] = 0;
380 if (insert_here
>= 0)
381 for (; i2
> insert_here
; i2
--)
382 p
->clause
[i2
] = p
->clause
[i2
- 1];
385 p
->clause
[insert_here
] = clause
;
391 static struct predicate
392 and_predicates (conditions conditions
,
393 struct predicate
*p
, struct predicate
*p2
)
395 struct predicate out
= *p
;
398 /* Avoid busy work. */
399 if (false_predicate_p (p2
) || true_predicate_p (p
))
401 if (false_predicate_p (p
) || true_predicate_p (p2
))
404 /* See how far predicates match. */
405 for (i
= 0; p
->clause
[i
] && p
->clause
[i
] == p2
->clause
[i
]; i
++)
407 gcc_checking_assert (i
< MAX_CLAUSES
);
410 /* Combine the predicates rest. */
411 for (; p2
->clause
[i
]; i
++)
413 gcc_checking_assert (i
< MAX_CLAUSES
);
414 add_clause (conditions
, &out
, p2
->clause
[i
]);
420 /* Return true if predicates are obviously equal. */
423 predicates_equal_p (struct predicate
*p
, struct predicate
*p2
)
426 for (i
= 0; p
->clause
[i
]; i
++)
428 gcc_checking_assert (i
< MAX_CLAUSES
);
429 gcc_checking_assert (p
->clause
[i
] > p
->clause
[i
+ 1]);
430 gcc_checking_assert (!p2
->clause
[i
]
431 || p2
->clause
[i
] > p2
->clause
[i
+ 1]);
432 if (p
->clause
[i
] != p2
->clause
[i
])
435 return !p2
->clause
[i
];
441 static struct predicate
442 or_predicates (conditions conditions
,
443 struct predicate
*p
, struct predicate
*p2
)
445 struct predicate out
= true_predicate ();
448 /* Avoid busy work. */
449 if (false_predicate_p (p2
) || true_predicate_p (p
))
451 if (false_predicate_p (p
) || true_predicate_p (p2
))
453 if (predicates_equal_p (p
, p2
))
456 /* OK, combine the predicates. */
457 for (i
= 0; p
->clause
[i
]; i
++)
458 for (j
= 0; p2
->clause
[j
]; j
++)
460 gcc_checking_assert (i
< MAX_CLAUSES
&& j
< MAX_CLAUSES
);
461 add_clause (conditions
, &out
, p
->clause
[i
] | p2
->clause
[j
]);
467 /* Having partial truth assignment in POSSIBLE_TRUTHS, return false
468 if predicate P is known to be false. */
471 evaluate_predicate (struct predicate
*p
, clause_t possible_truths
)
475 /* True remains true. */
476 if (true_predicate_p (p
))
479 gcc_assert (!(possible_truths
& (1 << predicate_false_condition
)));
481 /* See if we can find clause we can disprove. */
482 for (i
= 0; p
->clause
[i
]; i
++)
484 gcc_checking_assert (i
< MAX_CLAUSES
);
485 if (!(p
->clause
[i
] & possible_truths
))
491 /* Return the probability in range 0...REG_BR_PROB_BASE that the predicated
492 instruction will be recomputed per invocation of the inlined call. */
495 predicate_probability (conditions conds
,
496 struct predicate
*p
, clause_t possible_truths
,
497 vec
<inline_param_summary
> inline_param_summary
)
500 int combined_prob
= REG_BR_PROB_BASE
;
502 /* True remains true. */
503 if (true_predicate_p (p
))
504 return REG_BR_PROB_BASE
;
506 if (false_predicate_p (p
))
509 gcc_assert (!(possible_truths
& (1 << predicate_false_condition
)));
511 /* See if we can find clause we can disprove. */
512 for (i
= 0; p
->clause
[i
]; i
++)
514 gcc_checking_assert (i
< MAX_CLAUSES
);
515 if (!(p
->clause
[i
] & possible_truths
))
521 if (!inline_param_summary
.exists ())
522 return REG_BR_PROB_BASE
;
523 for (i2
= 0; i2
< NUM_CONDITIONS
; i2
++)
524 if ((p
->clause
[i
] & possible_truths
) & (1 << i2
))
526 if (i2
>= predicate_first_dynamic_condition
)
529 &(*conds
)[i2
- predicate_first_dynamic_condition
];
530 if (c
->code
== CHANGED
532 (int) inline_param_summary
.length ()))
535 inline_param_summary
[c
->operand_num
].change_prob
;
536 this_prob
= MAX (this_prob
, iprob
);
539 this_prob
= REG_BR_PROB_BASE
;
542 this_prob
= REG_BR_PROB_BASE
;
544 combined_prob
= MIN (this_prob
, combined_prob
);
549 return combined_prob
;
553 /* Dump conditional COND. */
556 dump_condition (FILE *f
, conditions conditions
, int cond
)
559 if (cond
== predicate_false_condition
)
560 fprintf (f
, "false");
561 else if (cond
== predicate_not_inlined_condition
)
562 fprintf (f
, "not inlined");
565 c
= &(*conditions
)[cond
- predicate_first_dynamic_condition
];
566 fprintf (f
, "op%i", c
->operand_num
);
568 fprintf (f
, "[%soffset: " HOST_WIDE_INT_PRINT_DEC
"]",
569 c
->by_ref
? "ref " : "", c
->offset
);
570 if (c
->code
== IS_NOT_CONSTANT
)
572 fprintf (f
, " not constant");
575 if (c
->code
== CHANGED
)
577 fprintf (f
, " changed");
580 fprintf (f
, " %s ", op_symbol_code (c
->code
));
581 print_generic_expr (f
, c
->val
, 1);
586 /* Dump clause CLAUSE. */
589 dump_clause (FILE *f
, conditions conds
, clause_t clause
)
596 for (i
= 0; i
< NUM_CONDITIONS
; i
++)
597 if (clause
& (1 << i
))
602 dump_condition (f
, conds
, i
);
608 /* Dump predicate PREDICATE. */
611 dump_predicate (FILE *f
, conditions conds
, struct predicate
*pred
)
614 if (true_predicate_p (pred
))
615 dump_clause (f
, conds
, 0);
617 for (i
= 0; pred
->clause
[i
]; i
++)
621 dump_clause (f
, conds
, pred
->clause
[i
]);
627 /* Dump inline hints. */
629 dump_inline_hints (FILE *f
, inline_hints hints
)
633 fprintf (f
, "inline hints:");
634 if (hints
& INLINE_HINT_indirect_call
)
636 hints
&= ~INLINE_HINT_indirect_call
;
637 fprintf (f
, " indirect_call");
639 if (hints
& INLINE_HINT_loop_iterations
)
641 hints
&= ~INLINE_HINT_loop_iterations
;
642 fprintf (f
, " loop_iterations");
644 if (hints
& INLINE_HINT_loop_stride
)
646 hints
&= ~INLINE_HINT_loop_stride
;
647 fprintf (f
, " loop_stride");
649 if (hints
& INLINE_HINT_same_scc
)
651 hints
&= ~INLINE_HINT_same_scc
;
652 fprintf (f
, " same_scc");
654 if (hints
& INLINE_HINT_in_scc
)
656 hints
&= ~INLINE_HINT_in_scc
;
657 fprintf (f
, " in_scc");
659 if (hints
& INLINE_HINT_cross_module
)
661 hints
&= ~INLINE_HINT_cross_module
;
662 fprintf (f
, " cross_module");
664 if (hints
& INLINE_HINT_declared_inline
)
666 hints
&= ~INLINE_HINT_declared_inline
;
667 fprintf (f
, " declared_inline");
669 if (hints
& INLINE_HINT_array_index
)
671 hints
&= ~INLINE_HINT_array_index
;
672 fprintf (f
, " array_index");
674 if (hints
& INLINE_HINT_known_hot
)
676 hints
&= ~INLINE_HINT_known_hot
;
677 fprintf (f
, " known_hot");
683 /* Record SIZE and TIME under condition PRED into the inline summary. */
686 account_size_time (struct inline_summary
*summary
, int size
, int time
,
687 struct predicate
*pred
)
693 if (false_predicate_p (pred
))
696 /* We need to create initial empty unconitional clause, but otherwie
697 we don't need to account empty times and sizes. */
698 if (!size
&& !time
&& summary
->entry
)
701 /* Watch overflow that might result from insane profiles. */
702 if (time
> MAX_TIME
* INLINE_TIME_SCALE
)
703 time
= MAX_TIME
* INLINE_TIME_SCALE
;
704 gcc_assert (time
>= 0);
706 for (i
= 0; vec_safe_iterate (summary
->entry
, i
, &e
); i
++)
707 if (predicates_equal_p (&e
->predicate
, pred
))
716 e
= &(*summary
->entry
)[0];
717 gcc_assert (!e
->predicate
.clause
[0]);
718 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
720 "\t\tReached limit on number of entries, "
721 "ignoring the predicate.");
723 if (dump_file
&& (dump_flags
& TDF_DETAILS
) && (time
|| size
))
726 "\t\tAccounting size:%3.2f, time:%3.2f on %spredicate:",
727 ((double) size
) / INLINE_SIZE_SCALE
,
728 ((double) time
) / INLINE_TIME_SCALE
, found
? "" : "new ");
729 dump_predicate (dump_file
, summary
->conds
, pred
);
733 struct size_time_entry new_entry
;
734 new_entry
.size
= size
;
735 new_entry
.time
= time
;
736 new_entry
.predicate
= *pred
;
737 vec_safe_push (summary
->entry
, new_entry
);
743 if (e
->time
> MAX_TIME
* INLINE_TIME_SCALE
)
744 e
->time
= MAX_TIME
* INLINE_TIME_SCALE
;
748 /* Set predicate for edge E. */
751 edge_set_predicate (struct cgraph_edge
*e
, struct predicate
*predicate
)
753 struct inline_edge_summary
*es
= inline_edge_summary (e
);
755 /* If the edge is determined to be never executed, redirect it
756 to BUILTIN_UNREACHABLE to save inliner from inlining into it. */
757 if (predicate
&& false_predicate_p (predicate
) && e
->callee
)
759 struct cgraph_node
*callee
= !e
->inline_failed
? e
->callee
: NULL
;
761 cgraph_redirect_edge_callee (e
,
762 cgraph_node::get_create
763 (builtin_decl_implicit (BUILT_IN_UNREACHABLE
)));
764 e
->inline_failed
= CIF_UNREACHABLE
;
766 callee
->remove_symbol_and_inline_clones ();
768 if (predicate
&& !true_predicate_p (predicate
))
771 es
->predicate
= (struct predicate
*) pool_alloc (edge_predicate_pool
);
772 *es
->predicate
= *predicate
;
777 pool_free (edge_predicate_pool
, es
->predicate
);
778 es
->predicate
= NULL
;
782 /* Set predicate for hint *P. */
785 set_hint_predicate (struct predicate
**p
, struct predicate new_predicate
)
787 if (false_predicate_p (&new_predicate
) || true_predicate_p (&new_predicate
))
790 pool_free (edge_predicate_pool
, *p
);
796 *p
= (struct predicate
*) pool_alloc (edge_predicate_pool
);
802 /* KNOWN_VALS is partial mapping of parameters of NODE to constant values.
803 KNOWN_AGGS is a vector of aggreggate jump functions for each parameter.
804 Return clause of possible truths. When INLINE_P is true, assume that we are
807 ERROR_MARK means compile time invariant. */
810 evaluate_conditions_for_known_args (struct cgraph_node
*node
,
812 vec
<tree
> known_vals
,
813 vec
<ipa_agg_jump_function_p
>
816 clause_t clause
= inline_p
? 0 : 1 << predicate_not_inlined_condition
;
817 struct inline_summary
*info
= inline_summary (node
);
821 for (i
= 0; vec_safe_iterate (info
->conds
, i
, &c
); i
++)
826 /* We allow call stmt to have fewer arguments than the callee function
827 (especially for K&R style programs). So bound check here (we assume
828 known_aggs vector, if non-NULL, has the same length as
830 gcc_checking_assert (!known_aggs
.exists ()
831 || (known_vals
.length () == known_aggs
.length ()));
832 if (c
->operand_num
>= (int) known_vals
.length ())
834 clause
|= 1 << (i
+ predicate_first_dynamic_condition
);
840 struct ipa_agg_jump_function
*agg
;
842 if (c
->code
== CHANGED
844 && (known_vals
[c
->operand_num
] == error_mark_node
))
847 if (known_aggs
.exists ())
849 agg
= known_aggs
[c
->operand_num
];
850 val
= ipa_find_agg_cst_for_param (agg
, c
->offset
, c
->by_ref
);
857 val
= known_vals
[c
->operand_num
];
858 if (val
== error_mark_node
&& c
->code
!= CHANGED
)
864 clause
|= 1 << (i
+ predicate_first_dynamic_condition
);
867 if (c
->code
== IS_NOT_CONSTANT
|| c
->code
== CHANGED
)
869 res
= fold_binary_to_constant (c
->code
, boolean_type_node
, val
, c
->val
);
870 if (res
&& integer_zerop (res
))
872 clause
|= 1 << (i
+ predicate_first_dynamic_condition
);
878 /* Work out what conditions might be true at invocation of E. */
881 evaluate_properties_for_edge (struct cgraph_edge
*e
, bool inline_p
,
882 clause_t
*clause_ptr
,
883 vec
<tree
> *known_vals_ptr
,
884 vec
<tree
> *known_binfos_ptr
,
885 vec
<ipa_agg_jump_function_p
> *known_aggs_ptr
)
887 struct cgraph_node
*callee
= e
->callee
->ultimate_alias_target ();
888 struct inline_summary
*info
= inline_summary (callee
);
889 vec
<tree
> known_vals
= vNULL
;
890 vec
<ipa_agg_jump_function_p
> known_aggs
= vNULL
;
893 *clause_ptr
= inline_p
? 0 : 1 << predicate_not_inlined_condition
;
895 known_vals_ptr
->create (0);
896 if (known_binfos_ptr
)
897 known_binfos_ptr
->create (0);
899 if (ipa_node_params_vector
.exists ()
900 && !e
->call_stmt_cannot_inline_p
901 && ((clause_ptr
&& info
->conds
) || known_vals_ptr
|| known_binfos_ptr
))
903 struct ipa_node_params
*parms_info
;
904 struct ipa_edge_args
*args
= IPA_EDGE_REF (e
);
905 struct inline_edge_summary
*es
= inline_edge_summary (e
);
906 int i
, count
= ipa_get_cs_argument_count (args
);
908 if (e
->caller
->global
.inlined_to
)
909 parms_info
= IPA_NODE_REF (e
->caller
->global
.inlined_to
);
911 parms_info
= IPA_NODE_REF (e
->caller
);
913 if (count
&& (info
->conds
|| known_vals_ptr
))
914 known_vals
.safe_grow_cleared (count
);
915 if (count
&& (info
->conds
|| known_aggs_ptr
))
916 known_aggs
.safe_grow_cleared (count
);
917 if (count
&& known_binfos_ptr
)
918 known_binfos_ptr
->safe_grow_cleared (count
);
920 for (i
= 0; i
< count
; i
++)
922 struct ipa_jump_func
*jf
= ipa_get_ith_jump_func (args
, i
);
923 tree cst
= ipa_value_from_jfunc (parms_info
, jf
);
926 if (known_vals
.exists () && TREE_CODE (cst
) != TREE_BINFO
)
928 else if (known_binfos_ptr
!= NULL
929 && TREE_CODE (cst
) == TREE_BINFO
)
930 (*known_binfos_ptr
)[i
] = cst
;
932 else if (inline_p
&& !es
->param
[i
].change_prob
)
933 known_vals
[i
] = error_mark_node
;
934 /* TODO: When IPA-CP starts propagating and merging aggregate jump
935 functions, use its knowledge of the caller too, just like the
936 scalar case above. */
937 known_aggs
[i
] = &jf
->agg
;
942 *clause_ptr
= evaluate_conditions_for_known_args (callee
, inline_p
,
943 known_vals
, known_aggs
);
946 *known_vals_ptr
= known_vals
;
948 known_vals
.release ();
951 *known_aggs_ptr
= known_aggs
;
953 known_aggs
.release ();
957 /* Allocate the inline summary vector or resize it to cover all cgraph nodes. */
960 inline_summary_alloc (void)
962 if (!node_removal_hook_holder
)
963 node_removal_hook_holder
=
964 cgraph_add_node_removal_hook (&inline_node_removal_hook
, NULL
);
965 if (!edge_removal_hook_holder
)
966 edge_removal_hook_holder
=
967 cgraph_add_edge_removal_hook (&inline_edge_removal_hook
, NULL
);
968 if (!node_duplication_hook_holder
)
969 node_duplication_hook_holder
=
970 cgraph_add_node_duplication_hook (&inline_node_duplication_hook
, NULL
);
971 if (!edge_duplication_hook_holder
)
972 edge_duplication_hook_holder
=
973 cgraph_add_edge_duplication_hook (&inline_edge_duplication_hook
, NULL
);
975 if (vec_safe_length (inline_summary_vec
) <= (unsigned) cgraph_max_uid
)
976 vec_safe_grow_cleared (inline_summary_vec
, cgraph_max_uid
+ 1);
977 if (inline_edge_summary_vec
.length () <= (unsigned) cgraph_edge_max_uid
)
978 inline_edge_summary_vec
.safe_grow_cleared (cgraph_edge_max_uid
+ 1);
979 if (!edge_predicate_pool
)
980 edge_predicate_pool
= create_alloc_pool ("edge predicates",
981 sizeof (struct predicate
), 10);
984 /* We are called multiple time for given function; clear
985 data from previous run so they are not cumulated. */
988 reset_inline_edge_summary (struct cgraph_edge
*e
)
990 if (e
->uid
< (int) inline_edge_summary_vec
.length ())
992 struct inline_edge_summary
*es
= inline_edge_summary (e
);
994 es
->call_stmt_size
= es
->call_stmt_time
= 0;
996 pool_free (edge_predicate_pool
, es
->predicate
);
997 es
->predicate
= NULL
;
998 es
->param
.release ();
1002 /* We are called multiple time for given function; clear
1003 data from previous run so they are not cumulated. */
1006 reset_inline_summary (struct cgraph_node
*node
)
1008 struct inline_summary
*info
= inline_summary (node
);
1009 struct cgraph_edge
*e
;
1011 info
->self_size
= info
->self_time
= 0;
1012 info
->estimated_stack_size
= 0;
1013 info
->estimated_self_stack_size
= 0;
1014 info
->stack_frame_offset
= 0;
1019 if (info
->loop_iterations
)
1021 pool_free (edge_predicate_pool
, info
->loop_iterations
);
1022 info
->loop_iterations
= NULL
;
1024 if (info
->loop_stride
)
1026 pool_free (edge_predicate_pool
, info
->loop_stride
);
1027 info
->loop_stride
= NULL
;
1029 if (info
->array_index
)
1031 pool_free (edge_predicate_pool
, info
->array_index
);
1032 info
->array_index
= NULL
;
1034 vec_free (info
->conds
);
1035 vec_free (info
->entry
);
1036 for (e
= node
->callees
; e
; e
= e
->next_callee
)
1037 reset_inline_edge_summary (e
);
1038 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
1039 reset_inline_edge_summary (e
);
1042 /* Hook that is called by cgraph.c when a node is removed. */
1045 inline_node_removal_hook (struct cgraph_node
*node
,
1046 void *data ATTRIBUTE_UNUSED
)
1048 struct inline_summary
*info
;
1049 if (vec_safe_length (inline_summary_vec
) <= (unsigned) node
->uid
)
1051 info
= inline_summary (node
);
1052 reset_inline_summary (node
);
1053 memset (info
, 0, sizeof (inline_summary_t
));
1056 /* Remap predicate P of former function to be predicate of duplicated function.
1057 POSSIBLE_TRUTHS is clause of possible truths in the duplicated node,
1058 INFO is inline summary of the duplicated node. */
1060 static struct predicate
1061 remap_predicate_after_duplication (struct predicate
*p
,
1062 clause_t possible_truths
,
1063 struct inline_summary
*info
)
1065 struct predicate new_predicate
= true_predicate ();
1067 for (j
= 0; p
->clause
[j
]; j
++)
1068 if (!(possible_truths
& p
->clause
[j
]))
1070 new_predicate
= false_predicate ();
1074 add_clause (info
->conds
, &new_predicate
,
1075 possible_truths
& p
->clause
[j
]);
1076 return new_predicate
;
1079 /* Same as remap_predicate_after_duplication but handle hint predicate *P.
1080 Additionally care about allocating new memory slot for updated predicate
1081 and set it to NULL when it becomes true or false (and thus uninteresting).
1085 remap_hint_predicate_after_duplication (struct predicate
**p
,
1086 clause_t possible_truths
,
1087 struct inline_summary
*info
)
1089 struct predicate new_predicate
;
1094 new_predicate
= remap_predicate_after_duplication (*p
,
1095 possible_truths
, info
);
1096 /* We do not want to free previous predicate; it is used by node origin. */
1098 set_hint_predicate (p
, new_predicate
);
1102 /* Hook that is called by cgraph.c when a node is duplicated. */
1105 inline_node_duplication_hook (struct cgraph_node
*src
,
1106 struct cgraph_node
*dst
,
1107 ATTRIBUTE_UNUSED
void *data
)
1109 struct inline_summary
*info
;
1110 inline_summary_alloc ();
1111 info
= inline_summary (dst
);
1112 memcpy (info
, inline_summary (src
), sizeof (struct inline_summary
));
1113 /* TODO: as an optimization, we may avoid copying conditions
1114 that are known to be false or true. */
1115 info
->conds
= vec_safe_copy (info
->conds
);
1117 /* When there are any replacements in the function body, see if we can figure
1118 out that something was optimized out. */
1119 if (ipa_node_params_vector
.exists () && dst
->clone
.tree_map
)
1121 vec
<size_time_entry
, va_gc
> *entry
= info
->entry
;
1122 /* Use SRC parm info since it may not be copied yet. */
1123 struct ipa_node_params
*parms_info
= IPA_NODE_REF (src
);
1124 vec
<tree
> known_vals
= vNULL
;
1125 int count
= ipa_get_param_count (parms_info
);
1127 clause_t possible_truths
;
1128 struct predicate true_pred
= true_predicate ();
1130 int optimized_out_size
= 0;
1131 bool inlined_to_p
= false;
1132 struct cgraph_edge
*edge
;
1135 known_vals
.safe_grow_cleared (count
);
1136 for (i
= 0; i
< count
; i
++)
1138 struct ipa_replace_map
*r
;
1140 for (j
= 0; vec_safe_iterate (dst
->clone
.tree_map
, j
, &r
); j
++)
1142 if (((!r
->old_tree
&& r
->parm_num
== i
)
1143 || (r
->old_tree
&& r
->old_tree
== ipa_get_param (parms_info
, i
)))
1144 && r
->replace_p
&& !r
->ref_p
)
1146 known_vals
[i
] = r
->new_tree
;
1151 possible_truths
= evaluate_conditions_for_known_args (dst
, false,
1154 known_vals
.release ();
1156 account_size_time (info
, 0, 0, &true_pred
);
1158 /* Remap size_time vectors.
1159 Simplify the predicate by prunning out alternatives that are known
1161 TODO: as on optimization, we can also eliminate conditions known
1163 for (i
= 0; vec_safe_iterate (entry
, i
, &e
); i
++)
1165 struct predicate new_predicate
;
1166 new_predicate
= remap_predicate_after_duplication (&e
->predicate
,
1169 if (false_predicate_p (&new_predicate
))
1170 optimized_out_size
+= e
->size
;
1172 account_size_time (info
, e
->size
, e
->time
, &new_predicate
);
1175 /* Remap edge predicates with the same simplification as above.
1176 Also copy constantness arrays. */
1177 for (edge
= dst
->callees
; edge
; edge
= edge
->next_callee
)
1179 struct predicate new_predicate
;
1180 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
1182 if (!edge
->inline_failed
)
1183 inlined_to_p
= true;
1186 new_predicate
= remap_predicate_after_duplication (es
->predicate
,
1189 if (false_predicate_p (&new_predicate
)
1190 && !false_predicate_p (es
->predicate
))
1192 optimized_out_size
+= es
->call_stmt_size
* INLINE_SIZE_SCALE
;
1193 edge
->frequency
= 0;
1195 edge_set_predicate (edge
, &new_predicate
);
1198 /* Remap indirect edge predicates with the same simplificaiton as above.
1199 Also copy constantness arrays. */
1200 for (edge
= dst
->indirect_calls
; edge
; edge
= edge
->next_callee
)
1202 struct predicate new_predicate
;
1203 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
1205 gcc_checking_assert (edge
->inline_failed
);
1208 new_predicate
= remap_predicate_after_duplication (es
->predicate
,
1211 if (false_predicate_p (&new_predicate
)
1212 && !false_predicate_p (es
->predicate
))
1214 optimized_out_size
+= es
->call_stmt_size
* INLINE_SIZE_SCALE
;
1215 edge
->frequency
= 0;
1217 edge_set_predicate (edge
, &new_predicate
);
1219 remap_hint_predicate_after_duplication (&info
->loop_iterations
,
1220 possible_truths
, info
);
1221 remap_hint_predicate_after_duplication (&info
->loop_stride
,
1222 possible_truths
, info
);
1223 remap_hint_predicate_after_duplication (&info
->array_index
,
1224 possible_truths
, info
);
1226 /* If inliner or someone after inliner will ever start producing
1227 non-trivial clones, we will get trouble with lack of information
1228 about updating self sizes, because size vectors already contains
1229 sizes of the calees. */
1230 gcc_assert (!inlined_to_p
|| !optimized_out_size
);
1234 info
->entry
= vec_safe_copy (info
->entry
);
1235 if (info
->loop_iterations
)
1237 predicate p
= *info
->loop_iterations
;
1238 info
->loop_iterations
= NULL
;
1239 set_hint_predicate (&info
->loop_iterations
, p
);
1241 if (info
->loop_stride
)
1243 predicate p
= *info
->loop_stride
;
1244 info
->loop_stride
= NULL
;
1245 set_hint_predicate (&info
->loop_stride
, p
);
1247 if (info
->array_index
)
1249 predicate p
= *info
->array_index
;
1250 info
->array_index
= NULL
;
1251 set_hint_predicate (&info
->array_index
, p
);
1254 inline_update_overall_summary (dst
);
1258 /* Hook that is called by cgraph.c when a node is duplicated. */
1261 inline_edge_duplication_hook (struct cgraph_edge
*src
,
1262 struct cgraph_edge
*dst
,
1263 ATTRIBUTE_UNUSED
void *data
)
1265 struct inline_edge_summary
*info
;
1266 struct inline_edge_summary
*srcinfo
;
1267 inline_summary_alloc ();
1268 info
= inline_edge_summary (dst
);
1269 srcinfo
= inline_edge_summary (src
);
1270 memcpy (info
, srcinfo
, sizeof (struct inline_edge_summary
));
1271 info
->predicate
= NULL
;
1272 edge_set_predicate (dst
, srcinfo
->predicate
);
1273 info
->param
= srcinfo
->param
.copy ();
1277 /* Keep edge cache consistent across edge removal. */
1280 inline_edge_removal_hook (struct cgraph_edge
*edge
,
1281 void *data ATTRIBUTE_UNUSED
)
1283 if (edge_growth_cache
.exists ())
1284 reset_edge_growth_cache (edge
);
1285 reset_inline_edge_summary (edge
);
1289 /* Initialize growth caches. */
1292 initialize_growth_caches (void)
1294 if (cgraph_edge_max_uid
)
1295 edge_growth_cache
.safe_grow_cleared (cgraph_edge_max_uid
);
1297 node_growth_cache
.safe_grow_cleared (cgraph_max_uid
);
1301 /* Free growth caches. */
1304 free_growth_caches (void)
1306 edge_growth_cache
.release ();
1307 node_growth_cache
.release ();
1311 /* Dump edge summaries associated to NODE and recursively to all clones.
1312 Indent by INDENT. */
1315 dump_inline_edge_summary (FILE *f
, int indent
, struct cgraph_node
*node
,
1316 struct inline_summary
*info
)
1318 struct cgraph_edge
*edge
;
1319 for (edge
= node
->callees
; edge
; edge
= edge
->next_callee
)
1321 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
1322 struct cgraph_node
*callee
= edge
->callee
->ultimate_alias_target ();
1326 "%*s%s/%i %s\n%*s loop depth:%2i freq:%4i size:%2i"
1327 " time: %2i callee size:%2i stack:%2i",
1328 indent
, "", callee
->name (), callee
->order
,
1329 !edge
->inline_failed
1330 ? "inlined" : cgraph_inline_failed_string (edge
-> inline_failed
),
1331 indent
, "", es
->loop_depth
, edge
->frequency
,
1332 es
->call_stmt_size
, es
->call_stmt_time
,
1333 (int) inline_summary (callee
)->size
/ INLINE_SIZE_SCALE
,
1334 (int) inline_summary (callee
)->estimated_stack_size
);
1338 fprintf (f
, " predicate: ");
1339 dump_predicate (f
, info
->conds
, es
->predicate
);
1343 if (es
->param
.exists ())
1344 for (i
= 0; i
< (int) es
->param
.length (); i
++)
1346 int prob
= es
->param
[i
].change_prob
;
1349 fprintf (f
, "%*s op%i is compile time invariant\n",
1351 else if (prob
!= REG_BR_PROB_BASE
)
1352 fprintf (f
, "%*s op%i change %f%% of time\n", indent
+ 2, "", i
,
1353 prob
* 100.0 / REG_BR_PROB_BASE
);
1355 if (!edge
->inline_failed
)
1357 fprintf (f
, "%*sStack frame offset %i, callee self size %i,"
1358 " callee size %i\n",
1360 (int) inline_summary (callee
)->stack_frame_offset
,
1361 (int) inline_summary (callee
)->estimated_self_stack_size
,
1362 (int) inline_summary (callee
)->estimated_stack_size
);
1363 dump_inline_edge_summary (f
, indent
+ 2, callee
, info
);
1366 for (edge
= node
->indirect_calls
; edge
; edge
= edge
->next_callee
)
1368 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
1369 fprintf (f
, "%*sindirect call loop depth:%2i freq:%4i size:%2i"
1373 edge
->frequency
, es
->call_stmt_size
, es
->call_stmt_time
);
1376 fprintf (f
, "predicate: ");
1377 dump_predicate (f
, info
->conds
, es
->predicate
);
1386 dump_inline_summary (FILE *f
, struct cgraph_node
*node
)
1388 if (node
->definition
)
1390 struct inline_summary
*s
= inline_summary (node
);
1393 fprintf (f
, "Inline summary for %s/%i", node
->name (),
1395 if (DECL_DISREGARD_INLINE_LIMITS (node
->decl
))
1396 fprintf (f
, " always_inline");
1398 fprintf (f
, " inlinable");
1399 fprintf (f
, "\n self time: %i\n", s
->self_time
);
1400 fprintf (f
, " global time: %i\n", s
->time
);
1401 fprintf (f
, " self size: %i\n", s
->self_size
);
1402 fprintf (f
, " global size: %i\n", s
->size
);
1403 fprintf (f
, " min size: %i\n", s
->min_size
);
1404 fprintf (f
, " self stack: %i\n",
1405 (int) s
->estimated_self_stack_size
);
1406 fprintf (f
, " global stack: %i\n", (int) s
->estimated_stack_size
);
1408 fprintf (f
, " estimated growth:%i\n", (int) s
->growth
);
1410 fprintf (f
, " In SCC: %i\n", (int) s
->scc_no
);
1411 for (i
= 0; vec_safe_iterate (s
->entry
, i
, &e
); i
++)
1413 fprintf (f
, " size:%f, time:%f, predicate:",
1414 (double) e
->size
/ INLINE_SIZE_SCALE
,
1415 (double) e
->time
/ INLINE_TIME_SCALE
);
1416 dump_predicate (f
, s
->conds
, &e
->predicate
);
1418 if (s
->loop_iterations
)
1420 fprintf (f
, " loop iterations:");
1421 dump_predicate (f
, s
->conds
, s
->loop_iterations
);
1425 fprintf (f
, " loop stride:");
1426 dump_predicate (f
, s
->conds
, s
->loop_stride
);
1430 fprintf (f
, " array index:");
1431 dump_predicate (f
, s
->conds
, s
->array_index
);
1433 fprintf (f
, " calls:\n");
1434 dump_inline_edge_summary (f
, 4, node
, s
);
1440 debug_inline_summary (struct cgraph_node
*node
)
1442 dump_inline_summary (stderr
, node
);
1446 dump_inline_summaries (FILE *f
)
1448 struct cgraph_node
*node
;
1450 FOR_EACH_DEFINED_FUNCTION (node
)
1451 if (!node
->global
.inlined_to
)
1452 dump_inline_summary (f
, node
);
1455 /* Give initial reasons why inlining would fail on EDGE. This gets either
1456 nullified or usually overwritten by more precise reasons later. */
1459 initialize_inline_failed (struct cgraph_edge
*e
)
1461 struct cgraph_node
*callee
= e
->callee
;
1463 if (e
->indirect_unknown_callee
)
1464 e
->inline_failed
= CIF_INDIRECT_UNKNOWN_CALL
;
1465 else if (!callee
->definition
)
1466 e
->inline_failed
= CIF_BODY_NOT_AVAILABLE
;
1467 else if (callee
->local
.redefined_extern_inline
)
1468 e
->inline_failed
= CIF_REDEFINED_EXTERN_INLINE
;
1469 else if (e
->call_stmt_cannot_inline_p
)
1470 e
->inline_failed
= CIF_MISMATCHED_ARGUMENTS
;
1471 else if (cfun
&& fn_contains_cilk_spawn_p (cfun
))
1472 /* We can't inline if the function is spawing a function. */
1473 e
->inline_failed
= CIF_FUNCTION_NOT_INLINABLE
;
1475 e
->inline_failed
= CIF_FUNCTION_NOT_CONSIDERED
;
1478 /* Callback of walk_aliased_vdefs. Flags that it has been invoked to the
1479 boolean variable pointed to by DATA. */
1482 mark_modified (ao_ref
*ao ATTRIBUTE_UNUSED
, tree vdef ATTRIBUTE_UNUSED
,
1485 bool *b
= (bool *) data
;
1490 /* If OP refers to value of function parameter, return the corresponding
1494 unmodified_parm_1 (gimple stmt
, tree op
)
1496 /* SSA_NAME referring to parm default def? */
1497 if (TREE_CODE (op
) == SSA_NAME
1498 && SSA_NAME_IS_DEFAULT_DEF (op
)
1499 && TREE_CODE (SSA_NAME_VAR (op
)) == PARM_DECL
)
1500 return SSA_NAME_VAR (op
);
1501 /* Non-SSA parm reference? */
1502 if (TREE_CODE (op
) == PARM_DECL
)
1504 bool modified
= false;
1507 ao_ref_init (&refd
, op
);
1508 walk_aliased_vdefs (&refd
, gimple_vuse (stmt
), mark_modified
, &modified
,
1516 /* If OP refers to value of function parameter, return the corresponding
1517 parameter. Also traverse chains of SSA register assignments. */
1520 unmodified_parm (gimple stmt
, tree op
)
1522 tree res
= unmodified_parm_1 (stmt
, op
);
1526 if (TREE_CODE (op
) == SSA_NAME
1527 && !SSA_NAME_IS_DEFAULT_DEF (op
)
1528 && gimple_assign_single_p (SSA_NAME_DEF_STMT (op
)))
1529 return unmodified_parm (SSA_NAME_DEF_STMT (op
),
1530 gimple_assign_rhs1 (SSA_NAME_DEF_STMT (op
)));
1534 /* If OP refers to a value of a function parameter or value loaded from an
1535 aggregate passed to a parameter (either by value or reference), return TRUE
1536 and store the number of the parameter to *INDEX_P and information whether
1537 and how it has been loaded from an aggregate into *AGGPOS. INFO describes
1538 the function parameters, STMT is the statement in which OP is used or
1542 unmodified_parm_or_parm_agg_item (struct ipa_node_params
*info
,
1543 gimple stmt
, tree op
, int *index_p
,
1544 struct agg_position_info
*aggpos
)
1546 tree res
= unmodified_parm_1 (stmt
, op
);
1548 gcc_checking_assert (aggpos
);
1551 *index_p
= ipa_get_param_decl_index (info
, res
);
1554 aggpos
->agg_contents
= false;
1555 aggpos
->by_ref
= false;
1559 if (TREE_CODE (op
) == SSA_NAME
)
1561 if (SSA_NAME_IS_DEFAULT_DEF (op
)
1562 || !gimple_assign_single_p (SSA_NAME_DEF_STMT (op
)))
1564 stmt
= SSA_NAME_DEF_STMT (op
);
1565 op
= gimple_assign_rhs1 (stmt
);
1566 if (!REFERENCE_CLASS_P (op
))
1567 return unmodified_parm_or_parm_agg_item (info
, stmt
, op
, index_p
,
1571 aggpos
->agg_contents
= true;
1572 return ipa_load_from_parm_agg (info
, stmt
, op
, index_p
, &aggpos
->offset
,
1576 /* See if statement might disappear after inlining.
1577 0 - means not eliminated
1578 1 - half of statements goes away
1579 2 - for sure it is eliminated.
1580 We are not terribly sophisticated, basically looking for simple abstraction
1581 penalty wrappers. */
1584 eliminated_by_inlining_prob (gimple stmt
)
1586 enum gimple_code code
= gimple_code (stmt
);
1587 enum tree_code rhs_code
;
1597 if (gimple_num_ops (stmt
) != 2)
1600 rhs_code
= gimple_assign_rhs_code (stmt
);
1602 /* Casts of parameters, loads from parameters passed by reference
1603 and stores to return value or parameters are often free after
1604 inlining dua to SRA and further combining.
1605 Assume that half of statements goes away. */
1606 if (rhs_code
== CONVERT_EXPR
1607 || rhs_code
== NOP_EXPR
1608 || rhs_code
== VIEW_CONVERT_EXPR
1609 || rhs_code
== ADDR_EXPR
1610 || gimple_assign_rhs_class (stmt
) == GIMPLE_SINGLE_RHS
)
1612 tree rhs
= gimple_assign_rhs1 (stmt
);
1613 tree lhs
= gimple_assign_lhs (stmt
);
1614 tree inner_rhs
= get_base_address (rhs
);
1615 tree inner_lhs
= get_base_address (lhs
);
1616 bool rhs_free
= false;
1617 bool lhs_free
= false;
1624 /* Reads of parameter are expected to be free. */
1625 if (unmodified_parm (stmt
, inner_rhs
))
1627 /* Match expressions of form &this->field. Those will most likely
1628 combine with something upstream after inlining. */
1629 else if (TREE_CODE (inner_rhs
) == ADDR_EXPR
)
1631 tree op
= get_base_address (TREE_OPERAND (inner_rhs
, 0));
1632 if (TREE_CODE (op
) == PARM_DECL
)
1634 else if (TREE_CODE (op
) == MEM_REF
1635 && unmodified_parm (stmt
, TREE_OPERAND (op
, 0)))
1639 /* When parameter is not SSA register because its address is taken
1640 and it is just copied into one, the statement will be completely
1641 free after inlining (we will copy propagate backward). */
1642 if (rhs_free
&& is_gimple_reg (lhs
))
1645 /* Reads of parameters passed by reference
1646 expected to be free (i.e. optimized out after inlining). */
1647 if (TREE_CODE (inner_rhs
) == MEM_REF
1648 && unmodified_parm (stmt
, TREE_OPERAND (inner_rhs
, 0)))
1651 /* Copying parameter passed by reference into gimple register is
1652 probably also going to copy propagate, but we can't be quite
1654 if (rhs_free
&& is_gimple_reg (lhs
))
1657 /* Writes to parameters, parameters passed by value and return value
1658 (either dirrectly or passed via invisible reference) are free.
1660 TODO: We ought to handle testcase like
1661 struct a {int a,b;};
1663 retrurnsturct (void)
1669 This translate into:
1684 For that we either need to copy ipa-split logic detecting writes
1686 if (TREE_CODE (inner_lhs
) == PARM_DECL
1687 || TREE_CODE (inner_lhs
) == RESULT_DECL
1688 || (TREE_CODE (inner_lhs
) == MEM_REF
1689 && (unmodified_parm (stmt
, TREE_OPERAND (inner_lhs
, 0))
1690 || (TREE_CODE (TREE_OPERAND (inner_lhs
, 0)) == SSA_NAME
1691 && SSA_NAME_VAR (TREE_OPERAND (inner_lhs
, 0))
1692 && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND
1694 0))) == RESULT_DECL
))))
1697 && (is_gimple_reg (rhs
) || is_gimple_min_invariant (rhs
)))
1699 if (lhs_free
&& rhs_free
)
1709 /* If BB ends by a conditional we can turn into predicates, attach corresponding
1710 predicates to the CFG edges. */
1713 set_cond_stmt_execution_predicate (struct ipa_node_params
*info
,
1714 struct inline_summary
*summary
,
1720 struct agg_position_info aggpos
;
1721 enum tree_code code
, inverted_code
;
1727 last
= last_stmt (bb
);
1728 if (!last
|| gimple_code (last
) != GIMPLE_COND
)
1730 if (!is_gimple_ip_invariant (gimple_cond_rhs (last
)))
1732 op
= gimple_cond_lhs (last
);
1733 /* TODO: handle conditionals like
1736 if (unmodified_parm_or_parm_agg_item (info
, last
, op
, &index
, &aggpos
))
1738 code
= gimple_cond_code (last
);
1740 = invert_tree_comparison (code
,
1741 HONOR_NANS (TYPE_MODE (TREE_TYPE (op
))));
1743 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1745 enum tree_code this_code
= (e
->flags
& EDGE_TRUE_VALUE
1746 ? code
: inverted_code
);
1747 /* invert_tree_comparison will return ERROR_MARK on FP
1748 comparsions that are not EQ/NE instead of returning proper
1749 unordered one. Be sure it is not confused with NON_CONSTANT. */
1750 if (this_code
!= ERROR_MARK
)
1752 struct predicate p
= add_condition (summary
, index
, &aggpos
,
1754 gimple_cond_rhs (last
));
1755 e
->aux
= pool_alloc (edge_predicate_pool
);
1756 *(struct predicate
*) e
->aux
= p
;
1761 if (TREE_CODE (op
) != SSA_NAME
)
1764 if (builtin_constant_p (op))
1768 Here we can predicate nonconstant_code. We can't
1769 really handle constant_code since we have no predicate
1770 for this and also the constant code is not known to be
1771 optimized away when inliner doen't see operand is constant.
1772 Other optimizers might think otherwise. */
1773 if (gimple_cond_code (last
) != NE_EXPR
1774 || !integer_zerop (gimple_cond_rhs (last
)))
1776 set_stmt
= SSA_NAME_DEF_STMT (op
);
1777 if (!gimple_call_builtin_p (set_stmt
, BUILT_IN_CONSTANT_P
)
1778 || gimple_call_num_args (set_stmt
) != 1)
1780 op2
= gimple_call_arg (set_stmt
, 0);
1781 if (!unmodified_parm_or_parm_agg_item
1782 (info
, set_stmt
, op2
, &index
, &aggpos
))
1784 FOR_EACH_EDGE (e
, ei
, bb
->succs
) if (e
->flags
& EDGE_FALSE_VALUE
)
1786 struct predicate p
= add_condition (summary
, index
, &aggpos
,
1787 IS_NOT_CONSTANT
, NULL_TREE
);
1788 e
->aux
= pool_alloc (edge_predicate_pool
);
1789 *(struct predicate
*) e
->aux
= p
;
1794 /* If BB ends by a switch we can turn into predicates, attach corresponding
1795 predicates to the CFG edges. */
1798 set_switch_stmt_execution_predicate (struct ipa_node_params
*info
,
1799 struct inline_summary
*summary
,
1805 struct agg_position_info aggpos
;
1811 last
= last_stmt (bb
);
1812 if (!last
|| gimple_code (last
) != GIMPLE_SWITCH
)
1814 op
= gimple_switch_index (last
);
1815 if (!unmodified_parm_or_parm_agg_item (info
, last
, op
, &index
, &aggpos
))
1818 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1820 e
->aux
= pool_alloc (edge_predicate_pool
);
1821 *(struct predicate
*) e
->aux
= false_predicate ();
1823 n
= gimple_switch_num_labels (last
);
1824 for (case_idx
= 0; case_idx
< n
; ++case_idx
)
1826 tree cl
= gimple_switch_label (last
, case_idx
);
1830 e
= find_edge (bb
, label_to_block (CASE_LABEL (cl
)));
1831 min
= CASE_LOW (cl
);
1832 max
= CASE_HIGH (cl
);
1834 /* For default we might want to construct predicate that none
1835 of cases is met, but it is bit hard to do not having negations
1836 of conditionals handy. */
1838 p
= true_predicate ();
1840 p
= add_condition (summary
, index
, &aggpos
, EQ_EXPR
, min
);
1843 struct predicate p1
, p2
;
1844 p1
= add_condition (summary
, index
, &aggpos
, GE_EXPR
, min
);
1845 p2
= add_condition (summary
, index
, &aggpos
, LE_EXPR
, max
);
1846 p
= and_predicates (summary
->conds
, &p1
, &p2
);
1848 *(struct predicate
*) e
->aux
1849 = or_predicates (summary
->conds
, &p
, (struct predicate
*) e
->aux
);
1854 /* For each BB in NODE attach to its AUX pointer predicate under
1855 which it is executable. */
1858 compute_bb_predicates (struct cgraph_node
*node
,
1859 struct ipa_node_params
*parms_info
,
1860 struct inline_summary
*summary
)
1862 struct function
*my_function
= DECL_STRUCT_FUNCTION (node
->decl
);
1866 FOR_EACH_BB_FN (bb
, my_function
)
1868 set_cond_stmt_execution_predicate (parms_info
, summary
, bb
);
1869 set_switch_stmt_execution_predicate (parms_info
, summary
, bb
);
1872 /* Entry block is always executable. */
1873 ENTRY_BLOCK_PTR_FOR_FN (my_function
)->aux
1874 = pool_alloc (edge_predicate_pool
);
1875 *(struct predicate
*) ENTRY_BLOCK_PTR_FOR_FN (my_function
)->aux
1876 = true_predicate ();
1878 /* A simple dataflow propagation of predicates forward in the CFG.
1879 TODO: work in reverse postorder. */
1883 FOR_EACH_BB_FN (bb
, my_function
)
1885 struct predicate p
= false_predicate ();
1888 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
1892 struct predicate this_bb_predicate
1893 = *(struct predicate
*) e
->src
->aux
;
1896 = and_predicates (summary
->conds
, &this_bb_predicate
,
1897 (struct predicate
*) e
->aux
);
1898 p
= or_predicates (summary
->conds
, &p
, &this_bb_predicate
);
1899 if (true_predicate_p (&p
))
1903 if (false_predicate_p (&p
))
1904 gcc_assert (!bb
->aux
);
1910 bb
->aux
= pool_alloc (edge_predicate_pool
);
1911 *((struct predicate
*) bb
->aux
) = p
;
1913 else if (!predicates_equal_p (&p
, (struct predicate
*) bb
->aux
))
1915 /* This OR operation is needed to ensure monotonous data flow
1916 in the case we hit the limit on number of clauses and the
1917 and/or operations above give approximate answers. */
1918 p
= or_predicates (summary
->conds
, &p
, (struct predicate
*)bb
->aux
);
1919 if (!predicates_equal_p (&p
, (struct predicate
*) bb
->aux
))
1922 *((struct predicate
*) bb
->aux
) = p
;
1931 /* We keep info about constantness of SSA names. */
1933 typedef struct predicate predicate_t
;
1934 /* Return predicate specifying when the STMT might have result that is not
1935 a compile time constant. */
1937 static struct predicate
1938 will_be_nonconstant_expr_predicate (struct ipa_node_params
*info
,
1939 struct inline_summary
*summary
,
1941 vec
<predicate_t
> nonconstant_names
)
1946 while (UNARY_CLASS_P (expr
))
1947 expr
= TREE_OPERAND (expr
, 0);
1949 parm
= unmodified_parm (NULL
, expr
);
1950 if (parm
&& (index
= ipa_get_param_decl_index (info
, parm
)) >= 0)
1951 return add_condition (summary
, index
, NULL
, CHANGED
, NULL_TREE
);
1952 if (is_gimple_min_invariant (expr
))
1953 return false_predicate ();
1954 if (TREE_CODE (expr
) == SSA_NAME
)
1955 return nonconstant_names
[SSA_NAME_VERSION (expr
)];
1956 if (BINARY_CLASS_P (expr
) || COMPARISON_CLASS_P (expr
))
1958 struct predicate p1
= will_be_nonconstant_expr_predicate
1959 (info
, summary
, TREE_OPERAND (expr
, 0),
1961 struct predicate p2
;
1962 if (true_predicate_p (&p1
))
1964 p2
= will_be_nonconstant_expr_predicate (info
, summary
,
1965 TREE_OPERAND (expr
, 1),
1967 return or_predicates (summary
->conds
, &p1
, &p2
);
1969 else if (TREE_CODE (expr
) == COND_EXPR
)
1971 struct predicate p1
= will_be_nonconstant_expr_predicate
1972 (info
, summary
, TREE_OPERAND (expr
, 0),
1974 struct predicate p2
;
1975 if (true_predicate_p (&p1
))
1977 p2
= will_be_nonconstant_expr_predicate (info
, summary
,
1978 TREE_OPERAND (expr
, 1),
1980 if (true_predicate_p (&p2
))
1982 p1
= or_predicates (summary
->conds
, &p1
, &p2
);
1983 p2
= will_be_nonconstant_expr_predicate (info
, summary
,
1984 TREE_OPERAND (expr
, 2),
1986 return or_predicates (summary
->conds
, &p1
, &p2
);
1993 return false_predicate ();
1997 /* Return predicate specifying when the STMT might have result that is not
1998 a compile time constant. */
2000 static struct predicate
2001 will_be_nonconstant_predicate (struct ipa_node_params
*info
,
2002 struct inline_summary
*summary
,
2004 vec
<predicate_t
> nonconstant_names
)
2006 struct predicate p
= true_predicate ();
2009 struct predicate op_non_const
;
2012 struct agg_position_info aggpos
;
2014 /* What statments might be optimized away
2015 when their arguments are constant
2016 TODO: also trivial builtins.
2017 builtin_constant_p is already handled later. */
2018 if (gimple_code (stmt
) != GIMPLE_ASSIGN
2019 && gimple_code (stmt
) != GIMPLE_COND
2020 && gimple_code (stmt
) != GIMPLE_SWITCH
)
2023 /* Stores will stay anyway. */
2024 if (gimple_store_p (stmt
))
2027 is_load
= gimple_assign_load_p (stmt
);
2029 /* Loads can be optimized when the value is known. */
2033 gcc_assert (gimple_assign_single_p (stmt
));
2034 op
= gimple_assign_rhs1 (stmt
);
2035 if (!unmodified_parm_or_parm_agg_item (info
, stmt
, op
, &base_index
,
2042 /* See if we understand all operands before we start
2043 adding conditionals. */
2044 FOR_EACH_SSA_TREE_OPERAND (use
, stmt
, iter
, SSA_OP_USE
)
2046 tree parm
= unmodified_parm (stmt
, use
);
2047 /* For arguments we can build a condition. */
2048 if (parm
&& ipa_get_param_decl_index (info
, parm
) >= 0)
2050 if (TREE_CODE (use
) != SSA_NAME
)
2052 /* If we know when operand is constant,
2053 we still can say something useful. */
2054 if (!true_predicate_p (&nonconstant_names
[SSA_NAME_VERSION (use
)]))
2061 add_condition (summary
, base_index
, &aggpos
, CHANGED
, NULL
);
2063 op_non_const
= false_predicate ();
2064 FOR_EACH_SSA_TREE_OPERAND (use
, stmt
, iter
, SSA_OP_USE
)
2066 tree parm
= unmodified_parm (stmt
, use
);
2069 if (parm
&& (index
= ipa_get_param_decl_index (info
, parm
)) >= 0)
2071 if (index
!= base_index
)
2072 p
= add_condition (summary
, index
, NULL
, CHANGED
, NULL_TREE
);
2077 p
= nonconstant_names
[SSA_NAME_VERSION (use
)];
2078 op_non_const
= or_predicates (summary
->conds
, &p
, &op_non_const
);
2080 if (gimple_code (stmt
) == GIMPLE_ASSIGN
2081 && TREE_CODE (gimple_assign_lhs (stmt
)) == SSA_NAME
)
2082 nonconstant_names
[SSA_NAME_VERSION (gimple_assign_lhs (stmt
))]
2084 return op_non_const
;
2087 struct record_modified_bb_info
2093 /* Callback of walk_aliased_vdefs. Records basic blocks where the value may be
2094 set except for info->stmt. */
2097 record_modified (ao_ref
*ao ATTRIBUTE_UNUSED
, tree vdef
, void *data
)
2099 struct record_modified_bb_info
*info
=
2100 (struct record_modified_bb_info
*) data
;
2101 if (SSA_NAME_DEF_STMT (vdef
) == info
->stmt
)
2103 bitmap_set_bit (info
->bb_set
,
2104 SSA_NAME_IS_DEFAULT_DEF (vdef
)
2105 ? ENTRY_BLOCK_PTR_FOR_FN (cfun
)->index
2106 : gimple_bb (SSA_NAME_DEF_STMT (vdef
))->index
);
2110 /* Return probability (based on REG_BR_PROB_BASE) that I-th parameter of STMT
2111 will change since last invocation of STMT.
2113 Value 0 is reserved for compile time invariants.
2114 For common parameters it is REG_BR_PROB_BASE. For loop invariants it
2115 ought to be REG_BR_PROB_BASE / estimated_iters. */
2118 param_change_prob (gimple stmt
, int i
)
2120 tree op
= gimple_call_arg (stmt
, i
);
2121 basic_block bb
= gimple_bb (stmt
);
2124 /* Global invariants neve change. */
2125 if (is_gimple_min_invariant (op
))
2127 /* We would have to do non-trivial analysis to really work out what
2128 is the probability of value to change (i.e. when init statement
2129 is in a sibling loop of the call).
2131 We do an conservative estimate: when call is executed N times more often
2132 than the statement defining value, we take the frequency 1/N. */
2133 if (TREE_CODE (op
) == SSA_NAME
)
2138 return REG_BR_PROB_BASE
;
2140 if (SSA_NAME_IS_DEFAULT_DEF (op
))
2141 init_freq
= ENTRY_BLOCK_PTR_FOR_FN (cfun
)->frequency
;
2143 init_freq
= gimple_bb (SSA_NAME_DEF_STMT (op
))->frequency
;
2147 if (init_freq
< bb
->frequency
)
2148 return MAX (GCOV_COMPUTE_SCALE (init_freq
, bb
->frequency
), 1);
2150 return REG_BR_PROB_BASE
;
2153 base
= get_base_address (op
);
2158 struct record_modified_bb_info info
;
2161 tree init
= ctor_for_folding (base
);
2163 if (init
!= error_mark_node
)
2166 return REG_BR_PROB_BASE
;
2167 ao_ref_init (&refd
, op
);
2169 info
.bb_set
= BITMAP_ALLOC (NULL
);
2170 walk_aliased_vdefs (&refd
, gimple_vuse (stmt
), record_modified
, &info
,
2172 if (bitmap_bit_p (info
.bb_set
, bb
->index
))
2174 BITMAP_FREE (info
.bb_set
);
2175 return REG_BR_PROB_BASE
;
2178 /* Assume that every memory is initialized at entry.
2179 TODO: Can we easilly determine if value is always defined
2180 and thus we may skip entry block? */
2181 if (ENTRY_BLOCK_PTR_FOR_FN (cfun
)->frequency
)
2182 max
= ENTRY_BLOCK_PTR_FOR_FN (cfun
)->frequency
;
2186 EXECUTE_IF_SET_IN_BITMAP (info
.bb_set
, 0, index
, bi
)
2187 max
= MIN (max
, BASIC_BLOCK_FOR_FN (cfun
, index
)->frequency
);
2189 BITMAP_FREE (info
.bb_set
);
2190 if (max
< bb
->frequency
)
2191 return MAX (GCOV_COMPUTE_SCALE (max
, bb
->frequency
), 1);
2193 return REG_BR_PROB_BASE
;
2195 return REG_BR_PROB_BASE
;
2198 /* Find whether a basic block BB is the final block of a (half) diamond CFG
2199 sub-graph and if the predicate the condition depends on is known. If so,
2200 return true and store the pointer the predicate in *P. */
2203 phi_result_unknown_predicate (struct ipa_node_params
*info
,
2204 struct inline_summary
*summary
, basic_block bb
,
2205 struct predicate
*p
,
2206 vec
<predicate_t
> nonconstant_names
)
2210 basic_block first_bb
= NULL
;
2213 if (single_pred_p (bb
))
2215 *p
= false_predicate ();
2219 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
2221 if (single_succ_p (e
->src
))
2223 if (!single_pred_p (e
->src
))
2226 first_bb
= single_pred (e
->src
);
2227 else if (single_pred (e
->src
) != first_bb
)
2234 else if (e
->src
!= first_bb
)
2242 stmt
= last_stmt (first_bb
);
2244 || gimple_code (stmt
) != GIMPLE_COND
2245 || !is_gimple_ip_invariant (gimple_cond_rhs (stmt
)))
2248 *p
= will_be_nonconstant_expr_predicate (info
, summary
,
2249 gimple_cond_lhs (stmt
),
2251 if (true_predicate_p (p
))
2257 /* Given a PHI statement in a function described by inline properties SUMMARY
2258 and *P being the predicate describing whether the selected PHI argument is
2259 known, store a predicate for the result of the PHI statement into
2260 NONCONSTANT_NAMES, if possible. */
2263 predicate_for_phi_result (struct inline_summary
*summary
, gimple phi
,
2264 struct predicate
*p
,
2265 vec
<predicate_t
> nonconstant_names
)
2269 for (i
= 0; i
< gimple_phi_num_args (phi
); i
++)
2271 tree arg
= gimple_phi_arg (phi
, i
)->def
;
2272 if (!is_gimple_min_invariant (arg
))
2274 gcc_assert (TREE_CODE (arg
) == SSA_NAME
);
2275 *p
= or_predicates (summary
->conds
, p
,
2276 &nonconstant_names
[SSA_NAME_VERSION (arg
)]);
2277 if (true_predicate_p (p
))
2282 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2284 fprintf (dump_file
, "\t\tphi predicate: ");
2285 dump_predicate (dump_file
, summary
->conds
, p
);
2287 nonconstant_names
[SSA_NAME_VERSION (gimple_phi_result (phi
))] = *p
;
2290 /* Return predicate specifying when array index in access OP becomes non-constant. */
2292 static struct predicate
2293 array_index_predicate (struct inline_summary
*info
,
2294 vec
< predicate_t
> nonconstant_names
, tree op
)
2296 struct predicate p
= false_predicate ();
2297 while (handled_component_p (op
))
2299 if (TREE_CODE (op
) == ARRAY_REF
|| TREE_CODE (op
) == ARRAY_RANGE_REF
)
2301 if (TREE_CODE (TREE_OPERAND (op
, 1)) == SSA_NAME
)
2302 p
= or_predicates (info
->conds
, &p
,
2303 &nonconstant_names
[SSA_NAME_VERSION
2304 (TREE_OPERAND (op
, 1))]);
2306 op
= TREE_OPERAND (op
, 0);
2311 /* For a typical usage of __builtin_expect (a<b, 1), we
2312 may introduce an extra relation stmt:
2313 With the builtin, we have
2316 t3 = __builtin_expect (t2, 1);
2319 Without the builtin, we have
2322 This affects the size/time estimation and may have
2323 an impact on the earlier inlining.
2324 Here find this pattern and fix it up later. */
2327 find_foldable_builtin_expect (basic_block bb
)
2329 gimple_stmt_iterator bsi
;
2331 for (bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
2333 gimple stmt
= gsi_stmt (bsi
);
2334 if (gimple_call_builtin_p (stmt
, BUILT_IN_EXPECT
)
2335 || (is_gimple_call (stmt
)
2336 && gimple_call_internal_p (stmt
)
2337 && gimple_call_internal_fn (stmt
) == IFN_BUILTIN_EXPECT
))
2339 tree var
= gimple_call_lhs (stmt
);
2340 tree arg
= gimple_call_arg (stmt
, 0);
2341 use_operand_p use_p
;
2348 gcc_assert (TREE_CODE (var
) == SSA_NAME
);
2350 while (TREE_CODE (arg
) == SSA_NAME
)
2352 gimple stmt_tmp
= SSA_NAME_DEF_STMT (arg
);
2353 if (!is_gimple_assign (stmt_tmp
))
2355 switch (gimple_assign_rhs_code (stmt_tmp
))
2374 arg
= gimple_assign_rhs1 (stmt_tmp
);
2377 if (match
&& single_imm_use (var
, &use_p
, &use_stmt
)
2378 && gimple_code (use_stmt
) == GIMPLE_COND
)
2385 /* Return true when the basic blocks contains only clobbers followed by RESX.
2386 Such BBs are kept around to make removal of dead stores possible with
2387 presence of EH and will be optimized out by optimize_clobbers later in the
2390 NEED_EH is used to recurse in case the clobber has non-EH predecestors
2391 that can be clobber only, too.. When it is false, the RESX is not necessary
2392 on the end of basic block. */
2395 clobber_only_eh_bb_p (basic_block bb
, bool need_eh
= true)
2397 gimple_stmt_iterator gsi
= gsi_last_bb (bb
);
2403 if (gsi_end_p (gsi
))
2405 if (gimple_code (gsi_stmt (gsi
)) != GIMPLE_RESX
)
2409 else if (!single_succ_p (bb
))
2412 for (; !gsi_end_p (gsi
); gsi_prev (&gsi
))
2414 gimple stmt
= gsi_stmt (gsi
);
2415 if (is_gimple_debug (stmt
))
2417 if (gimple_clobber_p (stmt
))
2419 if (gimple_code (stmt
) == GIMPLE_LABEL
)
2424 /* See if all predecestors are either throws or clobber only BBs. */
2425 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
2426 if (!(e
->flags
& EDGE_EH
)
2427 && !clobber_only_eh_bb_p (e
->src
, false))
2433 /* Compute function body size parameters for NODE.
2434 When EARLY is true, we compute only simple summaries without
2435 non-trivial predicates to drive the early inliner. */
2438 estimate_function_body_sizes (struct cgraph_node
*node
, bool early
)
2441 /* Estimate static overhead for function prologue/epilogue and alignment. */
2443 /* Benefits are scaled by probability of elimination that is in range
2446 gimple_stmt_iterator bsi
;
2447 struct function
*my_function
= DECL_STRUCT_FUNCTION (node
->decl
);
2449 struct inline_summary
*info
= inline_summary (node
);
2450 struct predicate bb_predicate
;
2451 struct ipa_node_params
*parms_info
= NULL
;
2452 vec
<predicate_t
> nonconstant_names
= vNULL
;
2455 predicate array_index
= true_predicate ();
2456 gimple fix_builtin_expect_stmt
;
2461 if (optimize
&& !early
)
2463 calculate_dominance_info (CDI_DOMINATORS
);
2464 loop_optimizer_init (LOOPS_NORMAL
| LOOPS_HAVE_RECORDED_EXITS
);
2466 if (ipa_node_params_vector
.exists ())
2468 parms_info
= IPA_NODE_REF (node
);
2469 nonconstant_names
.safe_grow_cleared
2470 (SSANAMES (my_function
)->length ());
2475 fprintf (dump_file
, "\nAnalyzing function body size: %s\n",
2478 /* When we run into maximal number of entries, we assign everything to the
2479 constant truth case. Be sure to have it in list. */
2480 bb_predicate
= true_predicate ();
2481 account_size_time (info
, 0, 0, &bb_predicate
);
2483 bb_predicate
= not_inlined_predicate ();
2484 account_size_time (info
, 2 * INLINE_SIZE_SCALE
, 0, &bb_predicate
);
2486 gcc_assert (my_function
&& my_function
->cfg
);
2488 compute_bb_predicates (node
, parms_info
, info
);
2489 gcc_assert (cfun
== my_function
);
2490 order
= XNEWVEC (int, n_basic_blocks_for_fn (cfun
));
2491 nblocks
= pre_and_rev_post_order_compute (NULL
, order
, false);
2492 for (n
= 0; n
< nblocks
; n
++)
2494 bb
= BASIC_BLOCK_FOR_FN (cfun
, order
[n
]);
2495 freq
= compute_call_stmt_bb_frequency (node
->decl
, bb
);
2496 if (clobber_only_eh_bb_p (bb
))
2498 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2499 fprintf (dump_file
, "\n Ignoring BB %i;"
2500 " it will be optimized away by cleanup_clobbers\n",
2505 /* TODO: Obviously predicates can be propagated down across CFG. */
2509 bb_predicate
= *(struct predicate
*) bb
->aux
;
2511 bb_predicate
= false_predicate ();
2514 bb_predicate
= true_predicate ();
2516 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2518 fprintf (dump_file
, "\n BB %i predicate:", bb
->index
);
2519 dump_predicate (dump_file
, info
->conds
, &bb_predicate
);
2522 if (parms_info
&& nonconstant_names
.exists ())
2524 struct predicate phi_predicate
;
2525 bool first_phi
= true;
2527 for (bsi
= gsi_start_phis (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
2530 && !phi_result_unknown_predicate (parms_info
, info
, bb
,
2535 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2537 fprintf (dump_file
, " ");
2538 print_gimple_stmt (dump_file
, gsi_stmt (bsi
), 0, 0);
2540 predicate_for_phi_result (info
, gsi_stmt (bsi
), &phi_predicate
,
2545 fix_builtin_expect_stmt
= find_foldable_builtin_expect (bb
);
2547 for (bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
2549 gimple stmt
= gsi_stmt (bsi
);
2550 int this_size
= estimate_num_insns (stmt
, &eni_size_weights
);
2551 int this_time
= estimate_num_insns (stmt
, &eni_time_weights
);
2553 struct predicate will_be_nonconstant
;
2555 /* This relation stmt should be folded after we remove
2556 buildin_expect call. Adjust the cost here. */
2557 if (stmt
== fix_builtin_expect_stmt
)
2563 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2565 fprintf (dump_file
, " ");
2566 print_gimple_stmt (dump_file
, stmt
, 0, 0);
2567 fprintf (dump_file
, "\t\tfreq:%3.2f size:%3i time:%3i\n",
2568 ((double) freq
) / CGRAPH_FREQ_BASE
, this_size
,
2572 if (gimple_assign_load_p (stmt
) && nonconstant_names
.exists ())
2574 struct predicate this_array_index
;
2576 array_index_predicate (info
, nonconstant_names
,
2577 gimple_assign_rhs1 (stmt
));
2578 if (!false_predicate_p (&this_array_index
))
2580 and_predicates (info
->conds
, &array_index
,
2583 if (gimple_store_p (stmt
) && nonconstant_names
.exists ())
2585 struct predicate this_array_index
;
2587 array_index_predicate (info
, nonconstant_names
,
2588 gimple_get_lhs (stmt
));
2589 if (!false_predicate_p (&this_array_index
))
2591 and_predicates (info
->conds
, &array_index
,
2596 if (is_gimple_call (stmt
)
2597 && !gimple_call_internal_p (stmt
))
2599 struct cgraph_edge
*edge
= node
->get_edge (stmt
);
2600 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
2602 /* Special case: results of BUILT_IN_CONSTANT_P will be always
2603 resolved as constant. We however don't want to optimize
2604 out the cgraph edges. */
2605 if (nonconstant_names
.exists ()
2606 && gimple_call_builtin_p (stmt
, BUILT_IN_CONSTANT_P
)
2607 && gimple_call_lhs (stmt
)
2608 && TREE_CODE (gimple_call_lhs (stmt
)) == SSA_NAME
)
2610 struct predicate false_p
= false_predicate ();
2611 nonconstant_names
[SSA_NAME_VERSION (gimple_call_lhs (stmt
))]
2614 if (ipa_node_params_vector
.exists ())
2616 int count
= gimple_call_num_args (stmt
);
2620 es
->param
.safe_grow_cleared (count
);
2621 for (i
= 0; i
< count
; i
++)
2623 int prob
= param_change_prob (stmt
, i
);
2624 gcc_assert (prob
>= 0 && prob
<= REG_BR_PROB_BASE
);
2625 es
->param
[i
].change_prob
= prob
;
2629 es
->call_stmt_size
= this_size
;
2630 es
->call_stmt_time
= this_time
;
2631 es
->loop_depth
= bb_loop_depth (bb
);
2632 edge_set_predicate (edge
, &bb_predicate
);
2635 /* TODO: When conditional jump or swithc is known to be constant, but
2636 we did not translate it into the predicates, we really can account
2637 just maximum of the possible paths. */
2640 = will_be_nonconstant_predicate (parms_info
, info
,
2641 stmt
, nonconstant_names
);
2642 if (this_time
|| this_size
)
2648 prob
= eliminated_by_inlining_prob (stmt
);
2649 if (prob
== 1 && dump_file
&& (dump_flags
& TDF_DETAILS
))
2651 "\t\t50%% will be eliminated by inlining\n");
2652 if (prob
== 2 && dump_file
&& (dump_flags
& TDF_DETAILS
))
2653 fprintf (dump_file
, "\t\tWill be eliminated by inlining\n");
2656 p
= and_predicates (info
->conds
, &bb_predicate
,
2657 &will_be_nonconstant
);
2659 p
= true_predicate ();
2661 if (!false_predicate_p (&p
))
2665 if (time
> MAX_TIME
* INLINE_TIME_SCALE
)
2666 time
= MAX_TIME
* INLINE_TIME_SCALE
;
2669 /* We account everything but the calls. Calls have their own
2670 size/time info attached to cgraph edges. This is necessary
2671 in order to make the cost disappear after inlining. */
2672 if (!is_gimple_call (stmt
))
2676 struct predicate ip
= not_inlined_predicate ();
2677 ip
= and_predicates (info
->conds
, &ip
, &p
);
2678 account_size_time (info
, this_size
* prob
,
2679 this_time
* prob
, &ip
);
2682 account_size_time (info
, this_size
* (2 - prob
),
2683 this_time
* (2 - prob
), &p
);
2686 gcc_assert (time
>= 0);
2687 gcc_assert (size
>= 0);
2691 set_hint_predicate (&inline_summary (node
)->array_index
, array_index
);
2692 time
= (time
+ CGRAPH_FREQ_BASE
/ 2) / CGRAPH_FREQ_BASE
;
2693 if (time
> MAX_TIME
)
2697 if (!early
&& nonconstant_names
.exists ())
2700 predicate loop_iterations
= true_predicate ();
2701 predicate loop_stride
= true_predicate ();
2703 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2704 flow_loops_dump (dump_file
, NULL
, 0);
2706 FOR_EACH_LOOP (loop
, 0)
2711 struct tree_niter_desc niter_desc
;
2712 basic_block
*body
= get_loop_body (loop
);
2713 bb_predicate
= *(struct predicate
*) loop
->header
->aux
;
2715 exits
= get_loop_exit_edges (loop
);
2716 FOR_EACH_VEC_ELT (exits
, j
, ex
)
2717 if (number_of_iterations_exit (loop
, ex
, &niter_desc
, false)
2718 && !is_gimple_min_invariant (niter_desc
.niter
))
2720 predicate will_be_nonconstant
2721 = will_be_nonconstant_expr_predicate (parms_info
, info
,
2724 if (!true_predicate_p (&will_be_nonconstant
))
2725 will_be_nonconstant
= and_predicates (info
->conds
,
2727 &will_be_nonconstant
);
2728 if (!true_predicate_p (&will_be_nonconstant
)
2729 && !false_predicate_p (&will_be_nonconstant
))
2730 /* This is slightly inprecise. We may want to represent each
2731 loop with independent predicate. */
2733 and_predicates (info
->conds
, &loop_iterations
,
2734 &will_be_nonconstant
);
2738 for (i
= 0; i
< loop
->num_nodes
; i
++)
2740 gimple_stmt_iterator gsi
;
2741 bb_predicate
= *(struct predicate
*) body
[i
]->aux
;
2742 for (gsi
= gsi_start_bb (body
[i
]); !gsi_end_p (gsi
);
2745 gimple stmt
= gsi_stmt (gsi
);
2750 FOR_EACH_SSA_TREE_OPERAND (use
, stmt
, iter
, SSA_OP_USE
)
2752 predicate will_be_nonconstant
;
2755 (loop
, loop_containing_stmt (stmt
), use
, &iv
, true)
2756 || is_gimple_min_invariant (iv
.step
))
2759 = will_be_nonconstant_expr_predicate (parms_info
, info
,
2762 if (!true_predicate_p (&will_be_nonconstant
))
2764 = and_predicates (info
->conds
,
2766 &will_be_nonconstant
);
2767 if (!true_predicate_p (&will_be_nonconstant
)
2768 && !false_predicate_p (&will_be_nonconstant
))
2769 /* This is slightly inprecise. We may want to represent
2770 each loop with independent predicate. */
2772 and_predicates (info
->conds
, &loop_stride
,
2773 &will_be_nonconstant
);
2779 set_hint_predicate (&inline_summary (node
)->loop_iterations
,
2781 set_hint_predicate (&inline_summary (node
)->loop_stride
, loop_stride
);
2784 FOR_ALL_BB_FN (bb
, my_function
)
2790 pool_free (edge_predicate_pool
, bb
->aux
);
2792 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
2795 pool_free (edge_predicate_pool
, e
->aux
);
2799 inline_summary (node
)->self_time
= time
;
2800 inline_summary (node
)->self_size
= size
;
2801 nonconstant_names
.release ();
2802 if (optimize
&& !early
)
2804 loop_optimizer_finalize ();
2805 free_dominance_info (CDI_DOMINATORS
);
2809 fprintf (dump_file
, "\n");
2810 dump_inline_summary (dump_file
, node
);
2815 /* Compute parameters of functions used by inliner.
2816 EARLY is true when we compute parameters for the early inliner */
2819 compute_inline_parameters (struct cgraph_node
*node
, bool early
)
2821 HOST_WIDE_INT self_stack_size
;
2822 struct cgraph_edge
*e
;
2823 struct inline_summary
*info
;
2825 gcc_assert (!node
->global
.inlined_to
);
2827 inline_summary_alloc ();
2829 info
= inline_summary (node
);
2830 reset_inline_summary (node
);
2832 /* FIXME: Thunks are inlinable, but tree-inline don't know how to do that.
2833 Once this happen, we will need to more curefully predict call
2835 if (node
->thunk
.thunk_p
)
2837 struct inline_edge_summary
*es
= inline_edge_summary (node
->callees
);
2838 struct predicate t
= true_predicate ();
2840 info
->inlinable
= 0;
2841 node
->callees
->call_stmt_cannot_inline_p
= true;
2842 node
->local
.can_change_signature
= false;
2843 es
->call_stmt_time
= 1;
2844 es
->call_stmt_size
= 1;
2845 account_size_time (info
, 0, 0, &t
);
2849 /* Even is_gimple_min_invariant rely on current_function_decl. */
2850 push_cfun (DECL_STRUCT_FUNCTION (node
->decl
));
2852 /* Estimate the stack size for the function if we're optimizing. */
2853 self_stack_size
= optimize
? estimated_stack_frame_size (node
) : 0;
2854 info
->estimated_self_stack_size
= self_stack_size
;
2855 info
->estimated_stack_size
= self_stack_size
;
2856 info
->stack_frame_offset
= 0;
2858 /* Can this function be inlined at all? */
2859 if (!optimize
&& !lookup_attribute ("always_inline",
2860 DECL_ATTRIBUTES (node
->decl
)))
2861 info
->inlinable
= false;
2863 info
->inlinable
= tree_inlinable_function_p (node
->decl
);
2865 /* Type attributes can use parameter indices to describe them. */
2866 if (TYPE_ATTRIBUTES (TREE_TYPE (node
->decl
)))
2867 node
->local
.can_change_signature
= false;
2870 /* Otherwise, inlinable functions always can change signature. */
2871 if (info
->inlinable
)
2872 node
->local
.can_change_signature
= true;
2875 /* Functions calling builtin_apply can not change signature. */
2876 for (e
= node
->callees
; e
; e
= e
->next_callee
)
2878 tree
cdecl = e
->callee
->decl
;
2879 if (DECL_BUILT_IN (cdecl)
2880 && DECL_BUILT_IN_CLASS (cdecl) == BUILT_IN_NORMAL
2881 && (DECL_FUNCTION_CODE (cdecl) == BUILT_IN_APPLY_ARGS
2882 || DECL_FUNCTION_CODE (cdecl) == BUILT_IN_VA_START
))
2885 node
->local
.can_change_signature
= !e
;
2888 estimate_function_body_sizes (node
, early
);
2890 for (e
= node
->callees
; e
; e
= e
->next_callee
)
2891 if (e
->callee
->comdat_local_p ())
2893 node
->calls_comdat_local
= (e
!= NULL
);
2895 /* Inlining characteristics are maintained by the cgraph_mark_inline. */
2896 info
->time
= info
->self_time
;
2897 info
->size
= info
->self_size
;
2898 info
->stack_frame_offset
= 0;
2899 info
->estimated_stack_size
= info
->estimated_self_stack_size
;
2900 #ifdef ENABLE_CHECKING
2901 inline_update_overall_summary (node
);
2902 gcc_assert (info
->time
== info
->self_time
&& info
->size
== info
->self_size
);
2909 /* Compute parameters of functions used by inliner using
2910 current_function_decl. */
2913 compute_inline_parameters_for_current (void)
2915 compute_inline_parameters (cgraph_node::get (current_function_decl
), true);
2921 const pass_data pass_data_inline_parameters
=
2923 GIMPLE_PASS
, /* type */
2924 "inline_param", /* name */
2925 OPTGROUP_INLINE
, /* optinfo_flags */
2926 TV_INLINE_PARAMETERS
, /* tv_id */
2927 0, /* properties_required */
2928 0, /* properties_provided */
2929 0, /* properties_destroyed */
2930 0, /* todo_flags_start */
2931 0, /* todo_flags_finish */
2934 class pass_inline_parameters
: public gimple_opt_pass
2937 pass_inline_parameters (gcc::context
*ctxt
)
2938 : gimple_opt_pass (pass_data_inline_parameters
, ctxt
)
2941 /* opt_pass methods: */
2942 opt_pass
* clone () { return new pass_inline_parameters (m_ctxt
); }
2943 virtual unsigned int execute (function
*)
2945 return compute_inline_parameters_for_current ();
2948 }; // class pass_inline_parameters
2953 make_pass_inline_parameters (gcc::context
*ctxt
)
2955 return new pass_inline_parameters (ctxt
);
2959 /* Estimate benefit devirtualizing indirect edge IE, provided KNOWN_VALS and
2963 estimate_edge_devirt_benefit (struct cgraph_edge
*ie
,
2964 int *size
, int *time
,
2965 vec
<tree
> known_vals
,
2966 vec
<tree
> known_binfos
,
2967 vec
<ipa_agg_jump_function_p
> known_aggs
)
2970 struct cgraph_node
*callee
;
2971 struct inline_summary
*isummary
;
2972 enum availability avail
;
2974 if (!known_vals
.exists () && !known_binfos
.exists ())
2976 if (!flag_indirect_inlining
)
2979 target
= ipa_get_indirect_edge_target (ie
, known_vals
, known_binfos
,
2984 /* Account for difference in cost between indirect and direct calls. */
2985 *size
-= (eni_size_weights
.indirect_call_cost
- eni_size_weights
.call_cost
);
2986 *time
-= (eni_time_weights
.indirect_call_cost
- eni_time_weights
.call_cost
);
2987 gcc_checking_assert (*time
>= 0);
2988 gcc_checking_assert (*size
>= 0);
2990 callee
= cgraph_node::get (target
);
2991 if (!callee
|| !callee
->definition
)
2993 callee
= callee
->function_symbol (&avail
);
2994 if (avail
< AVAIL_AVAILABLE
)
2996 isummary
= inline_summary (callee
);
2997 return isummary
->inlinable
;
3000 /* Increase SIZE, MIN_SIZE (if non-NULL) and TIME for size and time needed to
3001 handle edge E with probability PROB.
3002 Set HINTS if edge may be devirtualized.
3003 KNOWN_VALS, KNOWN_AGGS and KNOWN_BINFOS describe context of the call
3007 estimate_edge_size_and_time (struct cgraph_edge
*e
, int *size
, int *min_size
,
3010 vec
<tree
> known_vals
,
3011 vec
<tree
> known_binfos
,
3012 vec
<ipa_agg_jump_function_p
> known_aggs
,
3013 inline_hints
*hints
)
3015 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3016 int call_size
= es
->call_stmt_size
;
3017 int call_time
= es
->call_stmt_time
;
3020 && estimate_edge_devirt_benefit (e
, &call_size
, &call_time
,
3021 known_vals
, known_binfos
, known_aggs
)
3022 && hints
&& cgraph_maybe_hot_edge_p (e
))
3023 *hints
|= INLINE_HINT_indirect_call
;
3024 cur_size
= call_size
* INLINE_SIZE_SCALE
;
3027 *min_size
+= cur_size
;
3028 *time
+= apply_probability ((gcov_type
) call_time
, prob
)
3029 * e
->frequency
* (INLINE_TIME_SCALE
/ CGRAPH_FREQ_BASE
);
3030 if (*time
> MAX_TIME
* INLINE_TIME_SCALE
)
3031 *time
= MAX_TIME
* INLINE_TIME_SCALE
;
3036 /* Increase SIZE, MIN_SIZE and TIME for size and time needed to handle all
3038 POSSIBLE_TRUTHS, KNOWN_VALS, KNOWN_AGGS and KNOWN_BINFOS describe context of
3042 estimate_calls_size_and_time (struct cgraph_node
*node
, int *size
,
3043 int *min_size
, int *time
,
3044 inline_hints
*hints
,
3045 clause_t possible_truths
,
3046 vec
<tree
> known_vals
,
3047 vec
<tree
> known_binfos
,
3048 vec
<ipa_agg_jump_function_p
> known_aggs
)
3050 struct cgraph_edge
*e
;
3051 for (e
= node
->callees
; e
; e
= e
->next_callee
)
3053 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3055 || evaluate_predicate (es
->predicate
, possible_truths
))
3057 if (e
->inline_failed
)
3059 /* Predicates of calls shall not use NOT_CHANGED codes,
3060 sowe do not need to compute probabilities. */
3061 estimate_edge_size_and_time (e
, size
,
3062 es
->predicate
? NULL
: min_size
,
3063 time
, REG_BR_PROB_BASE
,
3064 known_vals
, known_binfos
,
3068 estimate_calls_size_and_time (e
->callee
, size
, min_size
, time
,
3071 known_vals
, known_binfos
,
3075 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
3077 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3079 || evaluate_predicate (es
->predicate
, possible_truths
))
3080 estimate_edge_size_and_time (e
, size
,
3081 es
->predicate
? NULL
: min_size
,
3082 time
, REG_BR_PROB_BASE
,
3083 known_vals
, known_binfos
, known_aggs
,
3089 /* Estimate size and time needed to execute NODE assuming
3090 POSSIBLE_TRUTHS clause, and KNOWN_VALS, KNOWN_AGGS and KNOWN_BINFOS
3091 information about NODE's arguments. If non-NULL use also probability
3092 information present in INLINE_PARAM_SUMMARY vector.
3093 Additionally detemine hints determined by the context. Finally compute
3094 minimal size needed for the call that is independent on the call context and
3095 can be used for fast estimates. Return the values in RET_SIZE,
3096 RET_MIN_SIZE, RET_TIME and RET_HINTS. */
3099 estimate_node_size_and_time (struct cgraph_node
*node
,
3100 clause_t possible_truths
,
3101 vec
<tree
> known_vals
,
3102 vec
<tree
> known_binfos
,
3103 vec
<ipa_agg_jump_function_p
> known_aggs
,
3104 int *ret_size
, int *ret_min_size
, int *ret_time
,
3105 inline_hints
*ret_hints
,
3106 vec
<inline_param_summary
>
3107 inline_param_summary
)
3109 struct inline_summary
*info
= inline_summary (node
);
3114 inline_hints hints
= 0;
3117 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3120 fprintf (dump_file
, " Estimating body: %s/%i\n"
3121 " Known to be false: ", node
->name (),
3124 for (i
= predicate_not_inlined_condition
;
3125 i
< (predicate_first_dynamic_condition
3126 + (int) vec_safe_length (info
->conds
)); i
++)
3127 if (!(possible_truths
& (1 << i
)))
3130 fprintf (dump_file
, ", ");
3132 dump_condition (dump_file
, info
->conds
, i
);
3136 for (i
= 0; vec_safe_iterate (info
->entry
, i
, &e
); i
++)
3137 if (evaluate_predicate (&e
->predicate
, possible_truths
))
3140 gcc_checking_assert (e
->time
>= 0);
3141 gcc_checking_assert (time
>= 0);
3142 if (!inline_param_summary
.exists ())
3146 int prob
= predicate_probability (info
->conds
,
3149 inline_param_summary
);
3150 gcc_checking_assert (prob
>= 0);
3151 gcc_checking_assert (prob
<= REG_BR_PROB_BASE
);
3152 time
+= apply_probability ((gcov_type
) e
->time
, prob
);
3154 if (time
> MAX_TIME
* INLINE_TIME_SCALE
)
3155 time
= MAX_TIME
* INLINE_TIME_SCALE
;
3156 gcc_checking_assert (time
>= 0);
3159 gcc_checking_assert (true_predicate_p (&(*info
->entry
)[0].predicate
));
3160 min_size
= (*info
->entry
)[0].size
;
3161 gcc_checking_assert (size
>= 0);
3162 gcc_checking_assert (time
>= 0);
3164 if (info
->loop_iterations
3165 && !evaluate_predicate (info
->loop_iterations
, possible_truths
))
3166 hints
|= INLINE_HINT_loop_iterations
;
3167 if (info
->loop_stride
3168 && !evaluate_predicate (info
->loop_stride
, possible_truths
))
3169 hints
|= INLINE_HINT_loop_stride
;
3170 if (info
->array_index
3171 && !evaluate_predicate (info
->array_index
, possible_truths
))
3172 hints
|= INLINE_HINT_array_index
;
3174 hints
|= INLINE_HINT_in_scc
;
3175 if (DECL_DECLARED_INLINE_P (node
->decl
))
3176 hints
|= INLINE_HINT_declared_inline
;
3178 estimate_calls_size_and_time (node
, &size
, &min_size
, &time
, &hints
, possible_truths
,
3179 known_vals
, known_binfos
, known_aggs
);
3180 gcc_checking_assert (size
>= 0);
3181 gcc_checking_assert (time
>= 0);
3182 time
= RDIV (time
, INLINE_TIME_SCALE
);
3183 size
= RDIV (size
, INLINE_SIZE_SCALE
);
3184 min_size
= RDIV (min_size
, INLINE_SIZE_SCALE
);
3186 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3187 fprintf (dump_file
, "\n size:%i time:%i\n", (int) size
, (int) time
);
3193 *ret_min_size
= min_size
;
3200 /* Estimate size and time needed to execute callee of EDGE assuming that
3201 parameters known to be constant at caller of EDGE are propagated.
3202 KNOWN_VALS and KNOWN_BINFOS are vectors of assumed known constant values
3203 and types for parameters. */
3206 estimate_ipcp_clone_size_and_time (struct cgraph_node
*node
,
3207 vec
<tree
> known_vals
,
3208 vec
<tree
> known_binfos
,
3209 vec
<ipa_agg_jump_function_p
> known_aggs
,
3210 int *ret_size
, int *ret_time
,
3211 inline_hints
*hints
)
3215 clause
= evaluate_conditions_for_known_args (node
, false, known_vals
,
3217 estimate_node_size_and_time (node
, clause
, known_vals
, known_binfos
,
3218 known_aggs
, ret_size
, NULL
, ret_time
, hints
, vNULL
);
3221 /* Translate all conditions from callee representation into caller
3222 representation and symbolically evaluate predicate P into new predicate.
3224 INFO is inline_summary of function we are adding predicate into, CALLEE_INFO
3225 is summary of function predicate P is from. OPERAND_MAP is array giving
3226 callee formal IDs the caller formal IDs. POSSSIBLE_TRUTHS is clausule of all
3227 callee conditions that may be true in caller context. TOPLEV_PREDICATE is
3228 predicate under which callee is executed. OFFSET_MAP is an array of of
3229 offsets that need to be added to conditions, negative offset means that
3230 conditions relying on values passed by reference have to be discarded
3231 because they might not be preserved (and should be considered offset zero
3232 for other purposes). */
3234 static struct predicate
3235 remap_predicate (struct inline_summary
*info
,
3236 struct inline_summary
*callee_info
,
3237 struct predicate
*p
,
3238 vec
<int> operand_map
,
3239 vec
<int> offset_map
,
3240 clause_t possible_truths
, struct predicate
*toplev_predicate
)
3243 struct predicate out
= true_predicate ();
3245 /* True predicate is easy. */
3246 if (true_predicate_p (p
))
3247 return *toplev_predicate
;
3248 for (i
= 0; p
->clause
[i
]; i
++)
3250 clause_t clause
= p
->clause
[i
];
3252 struct predicate clause_predicate
= false_predicate ();
3254 gcc_assert (i
< MAX_CLAUSES
);
3256 for (cond
= 0; cond
< NUM_CONDITIONS
; cond
++)
3257 /* Do we have condition we can't disprove? */
3258 if (clause
& possible_truths
& (1 << cond
))
3260 struct predicate cond_predicate
;
3261 /* Work out if the condition can translate to predicate in the
3262 inlined function. */
3263 if (cond
>= predicate_first_dynamic_condition
)
3265 struct condition
*c
;
3267 c
= &(*callee_info
->conds
)[cond
3269 predicate_first_dynamic_condition
];
3270 /* See if we can remap condition operand to caller's operand.
3271 Otherwise give up. */
3272 if (!operand_map
.exists ()
3273 || (int) operand_map
.length () <= c
->operand_num
3274 || operand_map
[c
->operand_num
] == -1
3275 /* TODO: For non-aggregate conditions, adding an offset is
3276 basically an arithmetic jump function processing which
3277 we should support in future. */
3278 || ((!c
->agg_contents
|| !c
->by_ref
)
3279 && offset_map
[c
->operand_num
] > 0)
3280 || (c
->agg_contents
&& c
->by_ref
3281 && offset_map
[c
->operand_num
] < 0))
3282 cond_predicate
= true_predicate ();
3285 struct agg_position_info ap
;
3286 HOST_WIDE_INT offset_delta
= offset_map
[c
->operand_num
];
3287 if (offset_delta
< 0)
3289 gcc_checking_assert (!c
->agg_contents
|| !c
->by_ref
);
3292 gcc_assert (!c
->agg_contents
3293 || c
->by_ref
|| offset_delta
== 0);
3294 ap
.offset
= c
->offset
+ offset_delta
;
3295 ap
.agg_contents
= c
->agg_contents
;
3296 ap
.by_ref
= c
->by_ref
;
3297 cond_predicate
= add_condition (info
,
3298 operand_map
[c
->operand_num
],
3299 &ap
, c
->code
, c
->val
);
3302 /* Fixed conditions remains same, construct single
3303 condition predicate. */
3306 cond_predicate
.clause
[0] = 1 << cond
;
3307 cond_predicate
.clause
[1] = 0;
3309 clause_predicate
= or_predicates (info
->conds
, &clause_predicate
,
3312 out
= and_predicates (info
->conds
, &out
, &clause_predicate
);
3314 return and_predicates (info
->conds
, &out
, toplev_predicate
);
3318 /* Update summary information of inline clones after inlining.
3319 Compute peak stack usage. */
3322 inline_update_callee_summaries (struct cgraph_node
*node
, int depth
)
3324 struct cgraph_edge
*e
;
3325 struct inline_summary
*callee_info
= inline_summary (node
);
3326 struct inline_summary
*caller_info
= inline_summary (node
->callers
->caller
);
3329 callee_info
->stack_frame_offset
3330 = caller_info
->stack_frame_offset
3331 + caller_info
->estimated_self_stack_size
;
3332 peak
= callee_info
->stack_frame_offset
3333 + callee_info
->estimated_self_stack_size
;
3334 if (inline_summary (node
->global
.inlined_to
)->estimated_stack_size
< peak
)
3335 inline_summary (node
->global
.inlined_to
)->estimated_stack_size
= peak
;
3336 ipa_propagate_frequency (node
);
3337 for (e
= node
->callees
; e
; e
= e
->next_callee
)
3339 if (!e
->inline_failed
)
3340 inline_update_callee_summaries (e
->callee
, depth
);
3341 inline_edge_summary (e
)->loop_depth
+= depth
;
3343 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
3344 inline_edge_summary (e
)->loop_depth
+= depth
;
3347 /* Update change_prob of EDGE after INLINED_EDGE has been inlined.
3348 When functoin A is inlined in B and A calls C with parameter that
3349 changes with probability PROB1 and C is known to be passthroug
3350 of argument if B that change with probability PROB2, the probability
3351 of change is now PROB1*PROB2. */
3354 remap_edge_change_prob (struct cgraph_edge
*inlined_edge
,
3355 struct cgraph_edge
*edge
)
3357 if (ipa_node_params_vector
.exists ())
3360 struct ipa_edge_args
*args
= IPA_EDGE_REF (edge
);
3361 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3362 struct inline_edge_summary
*inlined_es
3363 = inline_edge_summary (inlined_edge
);
3365 for (i
= 0; i
< ipa_get_cs_argument_count (args
); i
++)
3367 struct ipa_jump_func
*jfunc
= ipa_get_ith_jump_func (args
, i
);
3368 if (jfunc
->type
== IPA_JF_PASS_THROUGH
3369 && (ipa_get_jf_pass_through_formal_id (jfunc
)
3370 < (int) inlined_es
->param
.length ()))
3372 int jf_formal_id
= ipa_get_jf_pass_through_formal_id (jfunc
);
3373 int prob1
= es
->param
[i
].change_prob
;
3374 int prob2
= inlined_es
->param
[jf_formal_id
].change_prob
;
3375 int prob
= combine_probabilities (prob1
, prob2
);
3377 if (prob1
&& prob2
&& !prob
)
3380 es
->param
[i
].change_prob
= prob
;
3386 /* Update edge summaries of NODE after INLINED_EDGE has been inlined.
3388 Remap predicates of callees of NODE. Rest of arguments match
3391 Also update change probabilities. */
3394 remap_edge_summaries (struct cgraph_edge
*inlined_edge
,
3395 struct cgraph_node
*node
,
3396 struct inline_summary
*info
,
3397 struct inline_summary
*callee_info
,
3398 vec
<int> operand_map
,
3399 vec
<int> offset_map
,
3400 clause_t possible_truths
,
3401 struct predicate
*toplev_predicate
)
3403 struct cgraph_edge
*e
;
3404 for (e
= node
->callees
; e
; e
= e
->next_callee
)
3406 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3409 if (e
->inline_failed
)
3411 remap_edge_change_prob (inlined_edge
, e
);
3415 p
= remap_predicate (info
, callee_info
,
3416 es
->predicate
, operand_map
, offset_map
,
3417 possible_truths
, toplev_predicate
);
3418 edge_set_predicate (e
, &p
);
3419 /* TODO: We should remove the edge for code that will be
3420 optimized out, but we need to keep verifiers and tree-inline
3421 happy. Make it cold for now. */
3422 if (false_predicate_p (&p
))
3429 edge_set_predicate (e
, toplev_predicate
);
3432 remap_edge_summaries (inlined_edge
, e
->callee
, info
, callee_info
,
3433 operand_map
, offset_map
, possible_truths
,
3436 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
3438 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3441 remap_edge_change_prob (inlined_edge
, e
);
3444 p
= remap_predicate (info
, callee_info
,
3445 es
->predicate
, operand_map
, offset_map
,
3446 possible_truths
, toplev_predicate
);
3447 edge_set_predicate (e
, &p
);
3448 /* TODO: We should remove the edge for code that will be optimized
3449 out, but we need to keep verifiers and tree-inline happy.
3450 Make it cold for now. */
3451 if (false_predicate_p (&p
))
3458 edge_set_predicate (e
, toplev_predicate
);
3462 /* Same as remap_predicate, but set result into hint *HINT. */
3465 remap_hint_predicate (struct inline_summary
*info
,
3466 struct inline_summary
*callee_info
,
3467 struct predicate
**hint
,
3468 vec
<int> operand_map
,
3469 vec
<int> offset_map
,
3470 clause_t possible_truths
,
3471 struct predicate
*toplev_predicate
)
3477 p
= remap_predicate (info
, callee_info
,
3479 operand_map
, offset_map
,
3480 possible_truths
, toplev_predicate
);
3481 if (!false_predicate_p (&p
) && !true_predicate_p (&p
))
3484 set_hint_predicate (hint
, p
);
3486 **hint
= and_predicates (info
->conds
, *hint
, &p
);
3490 /* We inlined EDGE. Update summary of the function we inlined into. */
3493 inline_merge_summary (struct cgraph_edge
*edge
)
3495 struct inline_summary
*callee_info
= inline_summary (edge
->callee
);
3496 struct cgraph_node
*to
= (edge
->caller
->global
.inlined_to
3497 ? edge
->caller
->global
.inlined_to
: edge
->caller
);
3498 struct inline_summary
*info
= inline_summary (to
);
3499 clause_t clause
= 0; /* not_inline is known to be false. */
3501 vec
<int> operand_map
= vNULL
;
3502 vec
<int> offset_map
= vNULL
;
3504 struct predicate toplev_predicate
;
3505 struct predicate true_p
= true_predicate ();
3506 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3509 toplev_predicate
= *es
->predicate
;
3511 toplev_predicate
= true_predicate ();
3513 if (ipa_node_params_vector
.exists () && callee_info
->conds
)
3515 struct ipa_edge_args
*args
= IPA_EDGE_REF (edge
);
3516 int count
= ipa_get_cs_argument_count (args
);
3519 evaluate_properties_for_edge (edge
, true, &clause
, NULL
, NULL
, NULL
);
3522 operand_map
.safe_grow_cleared (count
);
3523 offset_map
.safe_grow_cleared (count
);
3525 for (i
= 0; i
< count
; i
++)
3527 struct ipa_jump_func
*jfunc
= ipa_get_ith_jump_func (args
, i
);
3530 /* TODO: handle non-NOPs when merging. */
3531 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
3533 if (ipa_get_jf_pass_through_operation (jfunc
) == NOP_EXPR
)
3534 map
= ipa_get_jf_pass_through_formal_id (jfunc
);
3535 if (!ipa_get_jf_pass_through_agg_preserved (jfunc
))
3538 else if (jfunc
->type
== IPA_JF_ANCESTOR
)
3540 HOST_WIDE_INT offset
= ipa_get_jf_ancestor_offset (jfunc
);
3541 if (offset
>= 0 && offset
< INT_MAX
)
3543 map
= ipa_get_jf_ancestor_formal_id (jfunc
);
3544 if (!ipa_get_jf_ancestor_agg_preserved (jfunc
))
3546 offset_map
[i
] = offset
;
3549 operand_map
[i
] = map
;
3550 gcc_assert (map
< ipa_get_param_count (IPA_NODE_REF (to
)));
3553 for (i
= 0; vec_safe_iterate (callee_info
->entry
, i
, &e
); i
++)
3555 struct predicate p
= remap_predicate (info
, callee_info
,
3556 &e
->predicate
, operand_map
,
3559 if (!false_predicate_p (&p
))
3561 gcov_type add_time
= ((gcov_type
) e
->time
* edge
->frequency
3562 + CGRAPH_FREQ_BASE
/ 2) / CGRAPH_FREQ_BASE
;
3563 int prob
= predicate_probability (callee_info
->conds
,
3566 add_time
= apply_probability ((gcov_type
) add_time
, prob
);
3567 if (add_time
> MAX_TIME
* INLINE_TIME_SCALE
)
3568 add_time
= MAX_TIME
* INLINE_TIME_SCALE
;
3569 if (prob
!= REG_BR_PROB_BASE
3570 && dump_file
&& (dump_flags
& TDF_DETAILS
))
3572 fprintf (dump_file
, "\t\tScaling time by probability:%f\n",
3573 (double) prob
/ REG_BR_PROB_BASE
);
3575 account_size_time (info
, e
->size
, add_time
, &p
);
3578 remap_edge_summaries (edge
, edge
->callee
, info
, callee_info
, operand_map
,
3579 offset_map
, clause
, &toplev_predicate
);
3580 remap_hint_predicate (info
, callee_info
,
3581 &callee_info
->loop_iterations
,
3582 operand_map
, offset_map
, clause
, &toplev_predicate
);
3583 remap_hint_predicate (info
, callee_info
,
3584 &callee_info
->loop_stride
,
3585 operand_map
, offset_map
, clause
, &toplev_predicate
);
3586 remap_hint_predicate (info
, callee_info
,
3587 &callee_info
->array_index
,
3588 operand_map
, offset_map
, clause
, &toplev_predicate
);
3590 inline_update_callee_summaries (edge
->callee
,
3591 inline_edge_summary (edge
)->loop_depth
);
3593 /* We do not maintain predicates of inlined edges, free it. */
3594 edge_set_predicate (edge
, &true_p
);
3595 /* Similarly remove param summaries. */
3596 es
->param
.release ();
3597 operand_map
.release ();
3598 offset_map
.release ();
3601 /* For performance reasons inline_merge_summary is not updating overall size
3602 and time. Recompute it. */
3605 inline_update_overall_summary (struct cgraph_node
*node
)
3607 struct inline_summary
*info
= inline_summary (node
);
3613 for (i
= 0; vec_safe_iterate (info
->entry
, i
, &e
); i
++)
3615 info
->size
+= e
->size
, info
->time
+= e
->time
;
3616 if (info
->time
> MAX_TIME
* INLINE_TIME_SCALE
)
3617 info
->time
= MAX_TIME
* INLINE_TIME_SCALE
;
3619 estimate_calls_size_and_time (node
, &info
->size
, &info
->min_size
,
3621 ~(clause_t
) (1 << predicate_false_condition
),
3622 vNULL
, vNULL
, vNULL
);
3623 info
->time
= (info
->time
+ INLINE_TIME_SCALE
/ 2) / INLINE_TIME_SCALE
;
3624 info
->size
= (info
->size
+ INLINE_SIZE_SCALE
/ 2) / INLINE_SIZE_SCALE
;
3627 /* Return hints derrived from EDGE. */
3629 simple_edge_hints (struct cgraph_edge
*edge
)
3632 struct cgraph_node
*to
= (edge
->caller
->global
.inlined_to
3633 ? edge
->caller
->global
.inlined_to
: edge
->caller
);
3634 if (inline_summary (to
)->scc_no
3635 && inline_summary (to
)->scc_no
== inline_summary (edge
->callee
)->scc_no
3636 && !cgraph_edge_recursive_p (edge
))
3637 hints
|= INLINE_HINT_same_scc
;
3639 if (to
->lto_file_data
&& edge
->callee
->lto_file_data
3640 && to
->lto_file_data
!= edge
->callee
->lto_file_data
)
3641 hints
|= INLINE_HINT_cross_module
;
3646 /* Estimate the time cost for the caller when inlining EDGE.
3647 Only to be called via estimate_edge_time, that handles the
3650 When caching, also update the cache entry. Compute both time and
3651 size, since we always need both metrics eventually. */
3654 do_estimate_edge_time (struct cgraph_edge
*edge
)
3659 struct cgraph_node
*callee
;
3661 vec
<tree
> known_vals
;
3662 vec
<tree
> known_binfos
;
3663 vec
<ipa_agg_jump_function_p
> known_aggs
;
3664 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3667 callee
= edge
->callee
->ultimate_alias_target ();
3669 gcc_checking_assert (edge
->inline_failed
);
3670 evaluate_properties_for_edge (edge
, true,
3671 &clause
, &known_vals
, &known_binfos
,
3673 estimate_node_size_and_time (callee
, clause
, known_vals
, known_binfos
,
3674 known_aggs
, &size
, &min_size
, &time
, &hints
, es
->param
);
3676 /* When we have profile feedback, we can quite safely identify hot
3677 edges and for those we disable size limits. Don't do that when
3678 probability that caller will call the callee is low however, since it
3679 may hurt optimization of the caller's hot path. */
3680 if (edge
->count
&& cgraph_maybe_hot_edge_p (edge
)
3682 > (edge
->caller
->global
.inlined_to
3683 ? edge
->caller
->global
.inlined_to
->count
: edge
->caller
->count
)))
3684 hints
|= INLINE_HINT_known_hot
;
3686 known_vals
.release ();
3687 known_binfos
.release ();
3688 known_aggs
.release ();
3689 gcc_checking_assert (size
>= 0);
3690 gcc_checking_assert (time
>= 0);
3692 /* When caching, update the cache entry. */
3693 if (edge_growth_cache
.exists ())
3695 inline_summary (edge
->callee
)->min_size
= min_size
;
3696 if ((int) edge_growth_cache
.length () <= edge
->uid
)
3697 edge_growth_cache
.safe_grow_cleared (cgraph_edge_max_uid
);
3698 edge_growth_cache
[edge
->uid
].time
= time
+ (time
>= 0);
3700 edge_growth_cache
[edge
->uid
].size
= size
+ (size
>= 0);
3701 hints
|= simple_edge_hints (edge
);
3702 edge_growth_cache
[edge
->uid
].hints
= hints
+ 1;
3708 /* Return estimated callee growth after inlining EDGE.
3709 Only to be called via estimate_edge_size. */
3712 do_estimate_edge_size (struct cgraph_edge
*edge
)
3715 struct cgraph_node
*callee
;
3717 vec
<tree
> known_vals
;
3718 vec
<tree
> known_binfos
;
3719 vec
<ipa_agg_jump_function_p
> known_aggs
;
3721 /* When we do caching, use do_estimate_edge_time to populate the entry. */
3723 if (edge_growth_cache
.exists ())
3725 do_estimate_edge_time (edge
);
3726 size
= edge_growth_cache
[edge
->uid
].size
;
3727 gcc_checking_assert (size
);
3728 return size
- (size
> 0);
3731 callee
= edge
->callee
->ultimate_alias_target ();
3733 /* Early inliner runs without caching, go ahead and do the dirty work. */
3734 gcc_checking_assert (edge
->inline_failed
);
3735 evaluate_properties_for_edge (edge
, true,
3736 &clause
, &known_vals
, &known_binfos
,
3738 estimate_node_size_and_time (callee
, clause
, known_vals
, known_binfos
,
3739 known_aggs
, &size
, NULL
, NULL
, NULL
, vNULL
);
3740 known_vals
.release ();
3741 known_binfos
.release ();
3742 known_aggs
.release ();
3747 /* Estimate the growth of the caller when inlining EDGE.
3748 Only to be called via estimate_edge_size. */
3751 do_estimate_edge_hints (struct cgraph_edge
*edge
)
3754 struct cgraph_node
*callee
;
3756 vec
<tree
> known_vals
;
3757 vec
<tree
> known_binfos
;
3758 vec
<ipa_agg_jump_function_p
> known_aggs
;
3760 /* When we do caching, use do_estimate_edge_time to populate the entry. */
3762 if (edge_growth_cache
.exists ())
3764 do_estimate_edge_time (edge
);
3765 hints
= edge_growth_cache
[edge
->uid
].hints
;
3766 gcc_checking_assert (hints
);
3770 callee
= edge
->callee
->ultimate_alias_target ();
3772 /* Early inliner runs without caching, go ahead and do the dirty work. */
3773 gcc_checking_assert (edge
->inline_failed
);
3774 evaluate_properties_for_edge (edge
, true,
3775 &clause
, &known_vals
, &known_binfos
,
3777 estimate_node_size_and_time (callee
, clause
, known_vals
, known_binfos
,
3778 known_aggs
, NULL
, NULL
, NULL
, &hints
, vNULL
);
3779 known_vals
.release ();
3780 known_binfos
.release ();
3781 known_aggs
.release ();
3782 hints
|= simple_edge_hints (edge
);
3787 /* Estimate self time of the function NODE after inlining EDGE. */
3790 estimate_time_after_inlining (struct cgraph_node
*node
,
3791 struct cgraph_edge
*edge
)
3793 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3794 if (!es
->predicate
|| !false_predicate_p (es
->predicate
))
3797 inline_summary (node
)->time
+ estimate_edge_time (edge
);
3800 if (time
> MAX_TIME
)
3804 return inline_summary (node
)->time
;
3808 /* Estimate the size of NODE after inlining EDGE which should be an
3809 edge to either NODE or a call inlined into NODE. */
3812 estimate_size_after_inlining (struct cgraph_node
*node
,
3813 struct cgraph_edge
*edge
)
3815 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3816 if (!es
->predicate
|| !false_predicate_p (es
->predicate
))
3818 int size
= inline_summary (node
)->size
+ estimate_edge_growth (edge
);
3819 gcc_assert (size
>= 0);
3822 return inline_summary (node
)->size
;
3828 struct cgraph_node
*node
;
3829 bool self_recursive
;
3834 /* Worker for do_estimate_growth. Collect growth for all callers. */
3837 do_estimate_growth_1 (struct cgraph_node
*node
, void *data
)
3839 struct cgraph_edge
*e
;
3840 struct growth_data
*d
= (struct growth_data
*) data
;
3842 for (e
= node
->callers
; e
; e
= e
->next_caller
)
3844 gcc_checking_assert (e
->inline_failed
);
3846 if (e
->caller
== d
->node
3847 || (e
->caller
->global
.inlined_to
3848 && e
->caller
->global
.inlined_to
== d
->node
))
3849 d
->self_recursive
= true;
3850 d
->growth
+= estimate_edge_growth (e
);
3856 /* Estimate the growth caused by inlining NODE into all callees. */
3859 do_estimate_growth (struct cgraph_node
*node
)
3861 struct growth_data d
= { node
, 0, false };
3862 struct inline_summary
*info
= inline_summary (node
);
3864 node
->call_for_symbol_thunks_and_aliases (do_estimate_growth_1
, &d
, true);
3866 /* For self recursive functions the growth estimation really should be
3867 infinity. We don't want to return very large values because the growth
3868 plays various roles in badness computation fractions. Be sure to not
3869 return zero or negative growths. */
3870 if (d
.self_recursive
)
3871 d
.growth
= d
.growth
< info
->size
? info
->size
: d
.growth
;
3872 else if (DECL_EXTERNAL (node
->decl
))
3876 if (node
->will_be_removed_from_program_if_no_direct_calls_p ())
3877 d
.growth
-= info
->size
;
3878 /* COMDAT functions are very often not shared across multiple units
3879 since they come from various template instantiations.
3880 Take this into account. */
3881 else if (DECL_COMDAT (node
->decl
)
3882 && node
->can_remove_if_no_direct_calls_p ())
3883 d
.growth
-= (info
->size
3884 * (100 - PARAM_VALUE (PARAM_COMDAT_SHARING_PROBABILITY
))
3888 if (node_growth_cache
.exists ())
3890 if ((int) node_growth_cache
.length () <= node
->uid
)
3891 node_growth_cache
.safe_grow_cleared (cgraph_max_uid
);
3892 node_growth_cache
[node
->uid
] = d
.growth
+ (d
.growth
>= 0);
3898 /* Make cheap estimation if growth of NODE is likely positive knowing
3899 EDGE_GROWTH of one particular edge.
3900 We assume that most of other edges will have similar growth
3901 and skip computation if there are too many callers. */
3904 growth_likely_positive (struct cgraph_node
*node
, int edge_growth ATTRIBUTE_UNUSED
)
3908 struct cgraph_edge
*e
;
3909 gcc_checking_assert (edge_growth
> 0);
3911 /* Unlike for functions called once, we play unsafe with
3912 COMDATs. We can allow that since we know functions
3913 in consideration are small (and thus risk is small) and
3914 moreover grow estimates already accounts that COMDAT
3915 functions may or may not disappear when eliminated from
3916 current unit. With good probability making aggressive
3917 choice in all units is going to make overall program
3920 Consequently we ask cgraph_can_remove_if_no_direct_calls_p
3922 cgraph_will_be_removed_from_program_if_no_direct_calls */
3923 if (DECL_EXTERNAL (node
->decl
)
3924 || !node
->can_remove_if_no_direct_calls_p ())
3927 /* If there is cached value, just go ahead. */
3928 if ((int)node_growth_cache
.length () > node
->uid
3929 && (ret
= node_growth_cache
[node
->uid
]))
3931 if (!node
->will_be_removed_from_program_if_no_direct_calls_p ()
3932 && (!DECL_COMDAT (node
->decl
)
3933 || !node
->can_remove_if_no_direct_calls_p ()))
3935 max_callers
= inline_summary (node
)->size
* 4 / edge_growth
+ 2;
3937 for (e
= node
->callers
; e
; e
= e
->next_caller
)
3943 return estimate_growth (node
) > 0;
3947 /* This function performs intraprocedural analysis in NODE that is required to
3948 inline indirect calls. */
3951 inline_indirect_intraprocedural_analysis (struct cgraph_node
*node
)
3953 ipa_analyze_node (node
);
3954 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3956 ipa_print_node_params (dump_file
, node
);
3957 ipa_print_node_jump_functions (dump_file
, node
);
3962 /* Note function body size. */
3965 inline_analyze_function (struct cgraph_node
*node
)
3967 push_cfun (DECL_STRUCT_FUNCTION (node
->decl
));
3970 fprintf (dump_file
, "\nAnalyzing function: %s/%u\n",
3971 node
->name (), node
->order
);
3972 if (optimize
&& !node
->thunk
.thunk_p
)
3973 inline_indirect_intraprocedural_analysis (node
);
3974 compute_inline_parameters (node
, false);
3977 struct cgraph_edge
*e
;
3978 for (e
= node
->callees
; e
; e
= e
->next_callee
)
3980 if (e
->inline_failed
== CIF_FUNCTION_NOT_CONSIDERED
)
3981 e
->inline_failed
= CIF_FUNCTION_NOT_OPTIMIZED
;
3982 e
->call_stmt_cannot_inline_p
= true;
3984 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
3986 if (e
->inline_failed
== CIF_FUNCTION_NOT_CONSIDERED
)
3987 e
->inline_failed
= CIF_FUNCTION_NOT_OPTIMIZED
;
3988 e
->call_stmt_cannot_inline_p
= true;
3996 /* Called when new function is inserted to callgraph late. */
3999 add_new_function (struct cgraph_node
*node
, void *data ATTRIBUTE_UNUSED
)
4001 inline_analyze_function (node
);
4005 /* Note function body size. */
4008 inline_generate_summary (void)
4010 struct cgraph_node
*node
;
4012 /* When not optimizing, do not bother to analyze. Inlining is still done
4013 because edge redirection needs to happen there. */
4014 if (!optimize
&& !flag_lto
&& !flag_wpa
)
4017 function_insertion_hook_holder
=
4018 cgraph_add_function_insertion_hook (&add_new_function
, NULL
);
4020 ipa_register_cgraph_hooks ();
4021 inline_free_summary ();
4023 FOR_EACH_DEFINED_FUNCTION (node
)
4025 inline_analyze_function (node
);
4029 /* Read predicate from IB. */
4031 static struct predicate
4032 read_predicate (struct lto_input_block
*ib
)
4034 struct predicate out
;
4040 gcc_assert (k
<= MAX_CLAUSES
);
4041 clause
= out
.clause
[k
++] = streamer_read_uhwi (ib
);
4045 /* Zero-initialize the remaining clauses in OUT. */
4046 while (k
<= MAX_CLAUSES
)
4047 out
.clause
[k
++] = 0;
4053 /* Write inline summary for edge E to OB. */
4056 read_inline_edge_summary (struct lto_input_block
*ib
, struct cgraph_edge
*e
)
4058 struct inline_edge_summary
*es
= inline_edge_summary (e
);
4062 es
->call_stmt_size
= streamer_read_uhwi (ib
);
4063 es
->call_stmt_time
= streamer_read_uhwi (ib
);
4064 es
->loop_depth
= streamer_read_uhwi (ib
);
4065 p
= read_predicate (ib
);
4066 edge_set_predicate (e
, &p
);
4067 length
= streamer_read_uhwi (ib
);
4070 es
->param
.safe_grow_cleared (length
);
4071 for (i
= 0; i
< length
; i
++)
4072 es
->param
[i
].change_prob
= streamer_read_uhwi (ib
);
4077 /* Stream in inline summaries from the section. */
4080 inline_read_section (struct lto_file_decl_data
*file_data
, const char *data
,
4083 const struct lto_function_header
*header
=
4084 (const struct lto_function_header
*) data
;
4085 const int cfg_offset
= sizeof (struct lto_function_header
);
4086 const int main_offset
= cfg_offset
+ header
->cfg_size
;
4087 const int string_offset
= main_offset
+ header
->main_size
;
4088 struct data_in
*data_in
;
4089 unsigned int i
, count2
, j
;
4090 unsigned int f_count
;
4092 lto_input_block
ib ((const char *) data
+ main_offset
, header
->main_size
);
4095 lto_data_in_create (file_data
, (const char *) data
+ string_offset
,
4096 header
->string_size
, vNULL
);
4097 f_count
= streamer_read_uhwi (&ib
);
4098 for (i
= 0; i
< f_count
; i
++)
4101 struct cgraph_node
*node
;
4102 struct inline_summary
*info
;
4103 lto_symtab_encoder_t encoder
;
4104 struct bitpack_d bp
;
4105 struct cgraph_edge
*e
;
4108 index
= streamer_read_uhwi (&ib
);
4109 encoder
= file_data
->symtab_node_encoder
;
4110 node
= dyn_cast
<cgraph_node
*> (lto_symtab_encoder_deref (encoder
,
4112 info
= inline_summary (node
);
4114 info
->estimated_stack_size
4115 = info
->estimated_self_stack_size
= streamer_read_uhwi (&ib
);
4116 info
->size
= info
->self_size
= streamer_read_uhwi (&ib
);
4117 info
->time
= info
->self_time
= streamer_read_uhwi (&ib
);
4119 bp
= streamer_read_bitpack (&ib
);
4120 info
->inlinable
= bp_unpack_value (&bp
, 1);
4122 count2
= streamer_read_uhwi (&ib
);
4123 gcc_assert (!info
->conds
);
4124 for (j
= 0; j
< count2
; j
++)
4127 c
.operand_num
= streamer_read_uhwi (&ib
);
4128 c
.code
= (enum tree_code
) streamer_read_uhwi (&ib
);
4129 c
.val
= stream_read_tree (&ib
, data_in
);
4130 bp
= streamer_read_bitpack (&ib
);
4131 c
.agg_contents
= bp_unpack_value (&bp
, 1);
4132 c
.by_ref
= bp_unpack_value (&bp
, 1);
4134 c
.offset
= streamer_read_uhwi (&ib
);
4135 vec_safe_push (info
->conds
, c
);
4137 count2
= streamer_read_uhwi (&ib
);
4138 gcc_assert (!info
->entry
);
4139 for (j
= 0; j
< count2
; j
++)
4141 struct size_time_entry e
;
4143 e
.size
= streamer_read_uhwi (&ib
);
4144 e
.time
= streamer_read_uhwi (&ib
);
4145 e
.predicate
= read_predicate (&ib
);
4147 vec_safe_push (info
->entry
, e
);
4150 p
= read_predicate (&ib
);
4151 set_hint_predicate (&info
->loop_iterations
, p
);
4152 p
= read_predicate (&ib
);
4153 set_hint_predicate (&info
->loop_stride
, p
);
4154 p
= read_predicate (&ib
);
4155 set_hint_predicate (&info
->array_index
, p
);
4156 for (e
= node
->callees
; e
; e
= e
->next_callee
)
4157 read_inline_edge_summary (&ib
, e
);
4158 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
4159 read_inline_edge_summary (&ib
, e
);
4162 lto_free_section_data (file_data
, LTO_section_inline_summary
, NULL
, data
,
4164 lto_data_in_delete (data_in
);
4168 /* Read inline summary. Jump functions are shared among ipa-cp
4169 and inliner, so when ipa-cp is active, we don't need to write them
4173 inline_read_summary (void)
4175 struct lto_file_decl_data
**file_data_vec
= lto_get_file_decl_data ();
4176 struct lto_file_decl_data
*file_data
;
4179 inline_summary_alloc ();
4181 while ((file_data
= file_data_vec
[j
++]))
4184 const char *data
= lto_get_section_data (file_data
,
4185 LTO_section_inline_summary
,
4188 inline_read_section (file_data
, data
, len
);
4190 /* Fatal error here. We do not want to support compiling ltrans units
4191 with different version of compiler or different flags than the WPA
4192 unit, so this should never happen. */
4193 fatal_error ("ipa inline summary is missing in input file");
4197 ipa_register_cgraph_hooks ();
4199 ipa_prop_read_jump_functions ();
4201 function_insertion_hook_holder
=
4202 cgraph_add_function_insertion_hook (&add_new_function
, NULL
);
4206 /* Write predicate P to OB. */
4209 write_predicate (struct output_block
*ob
, struct predicate
*p
)
4213 for (j
= 0; p
->clause
[j
]; j
++)
4215 gcc_assert (j
< MAX_CLAUSES
);
4216 streamer_write_uhwi (ob
, p
->clause
[j
]);
4218 streamer_write_uhwi (ob
, 0);
4222 /* Write inline summary for edge E to OB. */
4225 write_inline_edge_summary (struct output_block
*ob
, struct cgraph_edge
*e
)
4227 struct inline_edge_summary
*es
= inline_edge_summary (e
);
4230 streamer_write_uhwi (ob
, es
->call_stmt_size
);
4231 streamer_write_uhwi (ob
, es
->call_stmt_time
);
4232 streamer_write_uhwi (ob
, es
->loop_depth
);
4233 write_predicate (ob
, es
->predicate
);
4234 streamer_write_uhwi (ob
, es
->param
.length ());
4235 for (i
= 0; i
< (int) es
->param
.length (); i
++)
4236 streamer_write_uhwi (ob
, es
->param
[i
].change_prob
);
4240 /* Write inline summary for node in SET.
4241 Jump functions are shared among ipa-cp and inliner, so when ipa-cp is
4242 active, we don't need to write them twice. */
4245 inline_write_summary (void)
4247 struct cgraph_node
*node
;
4248 struct output_block
*ob
= create_output_block (LTO_section_inline_summary
);
4249 lto_symtab_encoder_t encoder
= ob
->decl_state
->symtab_node_encoder
;
4250 unsigned int count
= 0;
4253 for (i
= 0; i
< lto_symtab_encoder_size (encoder
); i
++)
4255 symtab_node
*snode
= lto_symtab_encoder_deref (encoder
, i
);
4256 cgraph_node
*cnode
= dyn_cast
<cgraph_node
*> (snode
);
4257 if (cnode
&& cnode
->definition
&& !cnode
->alias
)
4260 streamer_write_uhwi (ob
, count
);
4262 for (i
= 0; i
< lto_symtab_encoder_size (encoder
); i
++)
4264 symtab_node
*snode
= lto_symtab_encoder_deref (encoder
, i
);
4265 cgraph_node
*cnode
= dyn_cast
<cgraph_node
*> (snode
);
4266 if (cnode
&& (node
= cnode
)->definition
&& !node
->alias
)
4268 struct inline_summary
*info
= inline_summary (node
);
4269 struct bitpack_d bp
;
4270 struct cgraph_edge
*edge
;
4273 struct condition
*c
;
4275 streamer_write_uhwi (ob
,
4276 lto_symtab_encoder_encode (encoder
,
4279 streamer_write_hwi (ob
, info
->estimated_self_stack_size
);
4280 streamer_write_hwi (ob
, info
->self_size
);
4281 streamer_write_hwi (ob
, info
->self_time
);
4282 bp
= bitpack_create (ob
->main_stream
);
4283 bp_pack_value (&bp
, info
->inlinable
, 1);
4284 streamer_write_bitpack (&bp
);
4285 streamer_write_uhwi (ob
, vec_safe_length (info
->conds
));
4286 for (i
= 0; vec_safe_iterate (info
->conds
, i
, &c
); i
++)
4288 streamer_write_uhwi (ob
, c
->operand_num
);
4289 streamer_write_uhwi (ob
, c
->code
);
4290 stream_write_tree (ob
, c
->val
, true);
4291 bp
= bitpack_create (ob
->main_stream
);
4292 bp_pack_value (&bp
, c
->agg_contents
, 1);
4293 bp_pack_value (&bp
, c
->by_ref
, 1);
4294 streamer_write_bitpack (&bp
);
4295 if (c
->agg_contents
)
4296 streamer_write_uhwi (ob
, c
->offset
);
4298 streamer_write_uhwi (ob
, vec_safe_length (info
->entry
));
4299 for (i
= 0; vec_safe_iterate (info
->entry
, i
, &e
); i
++)
4301 streamer_write_uhwi (ob
, e
->size
);
4302 streamer_write_uhwi (ob
, e
->time
);
4303 write_predicate (ob
, &e
->predicate
);
4305 write_predicate (ob
, info
->loop_iterations
);
4306 write_predicate (ob
, info
->loop_stride
);
4307 write_predicate (ob
, info
->array_index
);
4308 for (edge
= node
->callees
; edge
; edge
= edge
->next_callee
)
4309 write_inline_edge_summary (ob
, edge
);
4310 for (edge
= node
->indirect_calls
; edge
; edge
= edge
->next_callee
)
4311 write_inline_edge_summary (ob
, edge
);
4314 streamer_write_char_stream (ob
->main_stream
, 0);
4315 produce_asm (ob
, NULL
);
4316 destroy_output_block (ob
);
4318 if (optimize
&& !flag_ipa_cp
)
4319 ipa_prop_write_jump_functions ();
4323 /* Release inline summary. */
4326 inline_free_summary (void)
4328 struct cgraph_node
*node
;
4329 if (!inline_edge_summary_vec
.exists ())
4331 FOR_EACH_DEFINED_FUNCTION (node
)
4333 reset_inline_summary (node
);
4334 if (function_insertion_hook_holder
)
4335 cgraph_remove_function_insertion_hook (function_insertion_hook_holder
);
4336 function_insertion_hook_holder
= NULL
;
4337 if (node_removal_hook_holder
)
4338 cgraph_remove_node_removal_hook (node_removal_hook_holder
);
4339 node_removal_hook_holder
= NULL
;
4340 if (edge_removal_hook_holder
)
4341 cgraph_remove_edge_removal_hook (edge_removal_hook_holder
);
4342 edge_removal_hook_holder
= NULL
;
4343 if (node_duplication_hook_holder
)
4344 cgraph_remove_node_duplication_hook (node_duplication_hook_holder
);
4345 node_duplication_hook_holder
= NULL
;
4346 if (edge_duplication_hook_holder
)
4347 cgraph_remove_edge_duplication_hook (edge_duplication_hook_holder
);
4348 edge_duplication_hook_holder
= NULL
;
4349 vec_free (inline_summary_vec
);
4350 inline_edge_summary_vec
.release ();
4351 if (edge_predicate_pool
)
4352 free_alloc_pool (edge_predicate_pool
);
4353 edge_predicate_pool
= 0;