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_get_create_node
763 (builtin_decl_implicit (BUILT_IN_UNREACHABLE
)));
764 e
->inline_failed
= CIF_UNREACHABLE
;
766 cgraph_remove_node_and_inline_clones (callee
, NULL
);
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
=
888 cgraph_function_or_thunk_node (e
->callee
, NULL
);
889 struct inline_summary
*info
= inline_summary (callee
);
890 vec
<tree
> known_vals
= vNULL
;
891 vec
<ipa_agg_jump_function_p
> known_aggs
= vNULL
;
894 *clause_ptr
= inline_p
? 0 : 1 << predicate_not_inlined_condition
;
896 known_vals_ptr
->create (0);
897 if (known_binfos_ptr
)
898 known_binfos_ptr
->create (0);
900 if (ipa_node_params_vector
.exists ()
901 && !e
->call_stmt_cannot_inline_p
902 && ((clause_ptr
&& info
->conds
) || known_vals_ptr
|| known_binfos_ptr
))
904 struct ipa_node_params
*parms_info
;
905 struct ipa_edge_args
*args
= IPA_EDGE_REF (e
);
906 struct inline_edge_summary
*es
= inline_edge_summary (e
);
907 int i
, count
= ipa_get_cs_argument_count (args
);
909 if (e
->caller
->global
.inlined_to
)
910 parms_info
= IPA_NODE_REF (e
->caller
->global
.inlined_to
);
912 parms_info
= IPA_NODE_REF (e
->caller
);
914 if (count
&& (info
->conds
|| known_vals_ptr
))
915 known_vals
.safe_grow_cleared (count
);
916 if (count
&& (info
->conds
|| known_aggs_ptr
))
917 known_aggs
.safe_grow_cleared (count
);
918 if (count
&& known_binfos_ptr
)
919 known_binfos_ptr
->safe_grow_cleared (count
);
921 for (i
= 0; i
< count
; i
++)
923 struct ipa_jump_func
*jf
= ipa_get_ith_jump_func (args
, i
);
924 tree cst
= ipa_value_from_jfunc (parms_info
, jf
);
927 if (known_vals
.exists () && TREE_CODE (cst
) != TREE_BINFO
)
929 else if (known_binfos_ptr
!= NULL
930 && TREE_CODE (cst
) == TREE_BINFO
)
931 (*known_binfos_ptr
)[i
] = cst
;
933 else if (inline_p
&& !es
->param
[i
].change_prob
)
934 known_vals
[i
] = error_mark_node
;
935 /* TODO: When IPA-CP starts propagating and merging aggregate jump
936 functions, use its knowledge of the caller too, just like the
937 scalar case above. */
938 known_aggs
[i
] = &jf
->agg
;
943 *clause_ptr
= evaluate_conditions_for_known_args (callee
, inline_p
,
944 known_vals
, known_aggs
);
947 *known_vals_ptr
= known_vals
;
949 known_vals
.release ();
952 *known_aggs_ptr
= known_aggs
;
954 known_aggs
.release ();
958 /* Allocate the inline summary vector or resize it to cover all cgraph nodes. */
961 inline_summary_alloc (void)
963 if (!node_removal_hook_holder
)
964 node_removal_hook_holder
=
965 cgraph_add_node_removal_hook (&inline_node_removal_hook
, NULL
);
966 if (!edge_removal_hook_holder
)
967 edge_removal_hook_holder
=
968 cgraph_add_edge_removal_hook (&inline_edge_removal_hook
, NULL
);
969 if (!node_duplication_hook_holder
)
970 node_duplication_hook_holder
=
971 cgraph_add_node_duplication_hook (&inline_node_duplication_hook
, NULL
);
972 if (!edge_duplication_hook_holder
)
973 edge_duplication_hook_holder
=
974 cgraph_add_edge_duplication_hook (&inline_edge_duplication_hook
, NULL
);
976 if (vec_safe_length (inline_summary_vec
) <= (unsigned) cgraph_max_uid
)
977 vec_safe_grow_cleared (inline_summary_vec
, cgraph_max_uid
+ 1);
978 if (inline_edge_summary_vec
.length () <= (unsigned) cgraph_edge_max_uid
)
979 inline_edge_summary_vec
.safe_grow_cleared (cgraph_edge_max_uid
+ 1);
980 if (!edge_predicate_pool
)
981 edge_predicate_pool
= create_alloc_pool ("edge predicates",
982 sizeof (struct predicate
), 10);
985 /* We are called multiple time for given function; clear
986 data from previous run so they are not cumulated. */
989 reset_inline_edge_summary (struct cgraph_edge
*e
)
991 if (e
->uid
< (int) inline_edge_summary_vec
.length ())
993 struct inline_edge_summary
*es
= inline_edge_summary (e
);
995 es
->call_stmt_size
= es
->call_stmt_time
= 0;
997 pool_free (edge_predicate_pool
, es
->predicate
);
998 es
->predicate
= NULL
;
999 es
->param
.release ();
1003 /* We are called multiple time for given function; clear
1004 data from previous run so they are not cumulated. */
1007 reset_inline_summary (struct cgraph_node
*node
)
1009 struct inline_summary
*info
= inline_summary (node
);
1010 struct cgraph_edge
*e
;
1012 info
->self_size
= info
->self_time
= 0;
1013 info
->estimated_stack_size
= 0;
1014 info
->estimated_self_stack_size
= 0;
1015 info
->stack_frame_offset
= 0;
1020 if (info
->loop_iterations
)
1022 pool_free (edge_predicate_pool
, info
->loop_iterations
);
1023 info
->loop_iterations
= NULL
;
1025 if (info
->loop_stride
)
1027 pool_free (edge_predicate_pool
, info
->loop_stride
);
1028 info
->loop_stride
= NULL
;
1030 if (info
->array_index
)
1032 pool_free (edge_predicate_pool
, info
->array_index
);
1033 info
->array_index
= NULL
;
1035 vec_free (info
->conds
);
1036 vec_free (info
->entry
);
1037 for (e
= node
->callees
; e
; e
= e
->next_callee
)
1038 reset_inline_edge_summary (e
);
1039 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
1040 reset_inline_edge_summary (e
);
1043 /* Hook that is called by cgraph.c when a node is removed. */
1046 inline_node_removal_hook (struct cgraph_node
*node
,
1047 void *data ATTRIBUTE_UNUSED
)
1049 struct inline_summary
*info
;
1050 if (vec_safe_length (inline_summary_vec
) <= (unsigned) node
->uid
)
1052 info
= inline_summary (node
);
1053 reset_inline_summary (node
);
1054 memset (info
, 0, sizeof (inline_summary_t
));
1057 /* Remap predicate P of former function to be predicate of duplicated function.
1058 POSSIBLE_TRUTHS is clause of possible truths in the duplicated node,
1059 INFO is inline summary of the duplicated node. */
1061 static struct predicate
1062 remap_predicate_after_duplication (struct predicate
*p
,
1063 clause_t possible_truths
,
1064 struct inline_summary
*info
)
1066 struct predicate new_predicate
= true_predicate ();
1068 for (j
= 0; p
->clause
[j
]; j
++)
1069 if (!(possible_truths
& p
->clause
[j
]))
1071 new_predicate
= false_predicate ();
1075 add_clause (info
->conds
, &new_predicate
,
1076 possible_truths
& p
->clause
[j
]);
1077 return new_predicate
;
1080 /* Same as remap_predicate_after_duplication but handle hint predicate *P.
1081 Additionally care about allocating new memory slot for updated predicate
1082 and set it to NULL when it becomes true or false (and thus uninteresting).
1086 remap_hint_predicate_after_duplication (struct predicate
**p
,
1087 clause_t possible_truths
,
1088 struct inline_summary
*info
)
1090 struct predicate new_predicate
;
1095 new_predicate
= remap_predicate_after_duplication (*p
,
1096 possible_truths
, info
);
1097 /* We do not want to free previous predicate; it is used by node origin. */
1099 set_hint_predicate (p
, new_predicate
);
1103 /* Hook that is called by cgraph.c when a node is duplicated. */
1106 inline_node_duplication_hook (struct cgraph_node
*src
,
1107 struct cgraph_node
*dst
,
1108 ATTRIBUTE_UNUSED
void *data
)
1110 struct inline_summary
*info
;
1111 inline_summary_alloc ();
1112 info
= inline_summary (dst
);
1113 memcpy (info
, inline_summary (src
), sizeof (struct inline_summary
));
1114 /* TODO: as an optimization, we may avoid copying conditions
1115 that are known to be false or true. */
1116 info
->conds
= vec_safe_copy (info
->conds
);
1118 /* When there are any replacements in the function body, see if we can figure
1119 out that something was optimized out. */
1120 if (ipa_node_params_vector
.exists () && dst
->clone
.tree_map
)
1122 vec
<size_time_entry
, va_gc
> *entry
= info
->entry
;
1123 /* Use SRC parm info since it may not be copied yet. */
1124 struct ipa_node_params
*parms_info
= IPA_NODE_REF (src
);
1125 vec
<tree
> known_vals
= vNULL
;
1126 int count
= ipa_get_param_count (parms_info
);
1128 clause_t possible_truths
;
1129 struct predicate true_pred
= true_predicate ();
1131 int optimized_out_size
= 0;
1132 bool inlined_to_p
= false;
1133 struct cgraph_edge
*edge
;
1136 known_vals
.safe_grow_cleared (count
);
1137 for (i
= 0; i
< count
; i
++)
1139 struct ipa_replace_map
*r
;
1141 for (j
= 0; vec_safe_iterate (dst
->clone
.tree_map
, j
, &r
); j
++)
1143 if (((!r
->old_tree
&& r
->parm_num
== i
)
1144 || (r
->old_tree
&& r
->old_tree
== ipa_get_param (parms_info
, i
)))
1145 && r
->replace_p
&& !r
->ref_p
)
1147 known_vals
[i
] = r
->new_tree
;
1152 possible_truths
= evaluate_conditions_for_known_args (dst
, false,
1155 known_vals
.release ();
1157 account_size_time (info
, 0, 0, &true_pred
);
1159 /* Remap size_time vectors.
1160 Simplify the predicate by prunning out alternatives that are known
1162 TODO: as on optimization, we can also eliminate conditions known
1164 for (i
= 0; vec_safe_iterate (entry
, i
, &e
); i
++)
1166 struct predicate new_predicate
;
1167 new_predicate
= remap_predicate_after_duplication (&e
->predicate
,
1170 if (false_predicate_p (&new_predicate
))
1171 optimized_out_size
+= e
->size
;
1173 account_size_time (info
, e
->size
, e
->time
, &new_predicate
);
1176 /* Remap edge predicates with the same simplification as above.
1177 Also copy constantness arrays. */
1178 for (edge
= dst
->callees
; edge
; edge
= edge
->next_callee
)
1180 struct predicate new_predicate
;
1181 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
1183 if (!edge
->inline_failed
)
1184 inlined_to_p
= true;
1187 new_predicate
= remap_predicate_after_duplication (es
->predicate
,
1190 if (false_predicate_p (&new_predicate
)
1191 && !false_predicate_p (es
->predicate
))
1193 optimized_out_size
+= es
->call_stmt_size
* INLINE_SIZE_SCALE
;
1194 edge
->frequency
= 0;
1196 edge_set_predicate (edge
, &new_predicate
);
1199 /* Remap indirect edge predicates with the same simplificaiton as above.
1200 Also copy constantness arrays. */
1201 for (edge
= dst
->indirect_calls
; edge
; edge
= edge
->next_callee
)
1203 struct predicate new_predicate
;
1204 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
1206 gcc_checking_assert (edge
->inline_failed
);
1209 new_predicate
= remap_predicate_after_duplication (es
->predicate
,
1212 if (false_predicate_p (&new_predicate
)
1213 && !false_predicate_p (es
->predicate
))
1215 optimized_out_size
+= es
->call_stmt_size
* INLINE_SIZE_SCALE
;
1216 edge
->frequency
= 0;
1218 edge_set_predicate (edge
, &new_predicate
);
1220 remap_hint_predicate_after_duplication (&info
->loop_iterations
,
1221 possible_truths
, info
);
1222 remap_hint_predicate_after_duplication (&info
->loop_stride
,
1223 possible_truths
, info
);
1224 remap_hint_predicate_after_duplication (&info
->array_index
,
1225 possible_truths
, info
);
1227 /* If inliner or someone after inliner will ever start producing
1228 non-trivial clones, we will get trouble with lack of information
1229 about updating self sizes, because size vectors already contains
1230 sizes of the calees. */
1231 gcc_assert (!inlined_to_p
|| !optimized_out_size
);
1235 info
->entry
= vec_safe_copy (info
->entry
);
1236 if (info
->loop_iterations
)
1238 predicate p
= *info
->loop_iterations
;
1239 info
->loop_iterations
= NULL
;
1240 set_hint_predicate (&info
->loop_iterations
, p
);
1242 if (info
->loop_stride
)
1244 predicate p
= *info
->loop_stride
;
1245 info
->loop_stride
= NULL
;
1246 set_hint_predicate (&info
->loop_stride
, p
);
1248 if (info
->array_index
)
1250 predicate p
= *info
->array_index
;
1251 info
->array_index
= NULL
;
1252 set_hint_predicate (&info
->array_index
, p
);
1255 inline_update_overall_summary (dst
);
1259 /* Hook that is called by cgraph.c when a node is duplicated. */
1262 inline_edge_duplication_hook (struct cgraph_edge
*src
,
1263 struct cgraph_edge
*dst
,
1264 ATTRIBUTE_UNUSED
void *data
)
1266 struct inline_edge_summary
*info
;
1267 struct inline_edge_summary
*srcinfo
;
1268 inline_summary_alloc ();
1269 info
= inline_edge_summary (dst
);
1270 srcinfo
= inline_edge_summary (src
);
1271 memcpy (info
, srcinfo
, sizeof (struct inline_edge_summary
));
1272 info
->predicate
= NULL
;
1273 edge_set_predicate (dst
, srcinfo
->predicate
);
1274 info
->param
= srcinfo
->param
.copy ();
1278 /* Keep edge cache consistent across edge removal. */
1281 inline_edge_removal_hook (struct cgraph_edge
*edge
,
1282 void *data ATTRIBUTE_UNUSED
)
1284 if (edge_growth_cache
.exists ())
1285 reset_edge_growth_cache (edge
);
1286 reset_inline_edge_summary (edge
);
1290 /* Initialize growth caches. */
1293 initialize_growth_caches (void)
1295 if (cgraph_edge_max_uid
)
1296 edge_growth_cache
.safe_grow_cleared (cgraph_edge_max_uid
);
1298 node_growth_cache
.safe_grow_cleared (cgraph_max_uid
);
1302 /* Free growth caches. */
1305 free_growth_caches (void)
1307 edge_growth_cache
.release ();
1308 node_growth_cache
.release ();
1312 /* Dump edge summaries associated to NODE and recursively to all clones.
1313 Indent by INDENT. */
1316 dump_inline_edge_summary (FILE *f
, int indent
, struct cgraph_node
*node
,
1317 struct inline_summary
*info
)
1319 struct cgraph_edge
*edge
;
1320 for (edge
= node
->callees
; edge
; edge
= edge
->next_callee
)
1322 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
1323 struct cgraph_node
*callee
=
1324 cgraph_function_or_thunk_node (edge
->callee
, NULL
);
1328 "%*s%s/%i %s\n%*s loop depth:%2i freq:%4i size:%2i"
1329 " time: %2i callee size:%2i stack:%2i",
1330 indent
, "", callee
->name (), callee
->order
,
1331 !edge
->inline_failed
1332 ? "inlined" : cgraph_inline_failed_string (edge
-> inline_failed
),
1333 indent
, "", es
->loop_depth
, edge
->frequency
,
1334 es
->call_stmt_size
, es
->call_stmt_time
,
1335 (int) inline_summary (callee
)->size
/ INLINE_SIZE_SCALE
,
1336 (int) inline_summary (callee
)->estimated_stack_size
);
1340 fprintf (f
, " predicate: ");
1341 dump_predicate (f
, info
->conds
, es
->predicate
);
1345 if (es
->param
.exists ())
1346 for (i
= 0; i
< (int) es
->param
.length (); i
++)
1348 int prob
= es
->param
[i
].change_prob
;
1351 fprintf (f
, "%*s op%i is compile time invariant\n",
1353 else if (prob
!= REG_BR_PROB_BASE
)
1354 fprintf (f
, "%*s op%i change %f%% of time\n", indent
+ 2, "", i
,
1355 prob
* 100.0 / REG_BR_PROB_BASE
);
1357 if (!edge
->inline_failed
)
1359 fprintf (f
, "%*sStack frame offset %i, callee self size %i,"
1360 " callee size %i\n",
1362 (int) inline_summary (callee
)->stack_frame_offset
,
1363 (int) inline_summary (callee
)->estimated_self_stack_size
,
1364 (int) inline_summary (callee
)->estimated_stack_size
);
1365 dump_inline_edge_summary (f
, indent
+ 2, callee
, info
);
1368 for (edge
= node
->indirect_calls
; edge
; edge
= edge
->next_callee
)
1370 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
1371 fprintf (f
, "%*sindirect call loop depth:%2i freq:%4i size:%2i"
1375 edge
->frequency
, es
->call_stmt_size
, es
->call_stmt_time
);
1378 fprintf (f
, "predicate: ");
1379 dump_predicate (f
, info
->conds
, es
->predicate
);
1388 dump_inline_summary (FILE *f
, struct cgraph_node
*node
)
1390 if (node
->definition
)
1392 struct inline_summary
*s
= inline_summary (node
);
1395 fprintf (f
, "Inline summary for %s/%i", node
->name (),
1397 if (DECL_DISREGARD_INLINE_LIMITS (node
->decl
))
1398 fprintf (f
, " always_inline");
1400 fprintf (f
, " inlinable");
1401 fprintf (f
, "\n self time: %i\n", s
->self_time
);
1402 fprintf (f
, " global time: %i\n", s
->time
);
1403 fprintf (f
, " self size: %i\n", s
->self_size
);
1404 fprintf (f
, " global size: %i\n", s
->size
);
1405 fprintf (f
, " min size: %i\n", s
->min_size
);
1406 fprintf (f
, " self stack: %i\n",
1407 (int) s
->estimated_self_stack_size
);
1408 fprintf (f
, " global stack: %i\n", (int) s
->estimated_stack_size
);
1410 fprintf (f
, " estimated growth:%i\n", (int) s
->growth
);
1412 fprintf (f
, " In SCC: %i\n", (int) s
->scc_no
);
1413 for (i
= 0; vec_safe_iterate (s
->entry
, i
, &e
); i
++)
1415 fprintf (f
, " size:%f, time:%f, predicate:",
1416 (double) e
->size
/ INLINE_SIZE_SCALE
,
1417 (double) e
->time
/ INLINE_TIME_SCALE
);
1418 dump_predicate (f
, s
->conds
, &e
->predicate
);
1420 if (s
->loop_iterations
)
1422 fprintf (f
, " loop iterations:");
1423 dump_predicate (f
, s
->conds
, s
->loop_iterations
);
1427 fprintf (f
, " loop stride:");
1428 dump_predicate (f
, s
->conds
, s
->loop_stride
);
1432 fprintf (f
, " array index:");
1433 dump_predicate (f
, s
->conds
, s
->array_index
);
1435 fprintf (f
, " calls:\n");
1436 dump_inline_edge_summary (f
, 4, node
, s
);
1442 debug_inline_summary (struct cgraph_node
*node
)
1444 dump_inline_summary (stderr
, node
);
1448 dump_inline_summaries (FILE *f
)
1450 struct cgraph_node
*node
;
1452 FOR_EACH_DEFINED_FUNCTION (node
)
1453 if (!node
->global
.inlined_to
)
1454 dump_inline_summary (f
, node
);
1457 /* Give initial reasons why inlining would fail on EDGE. This gets either
1458 nullified or usually overwritten by more precise reasons later. */
1461 initialize_inline_failed (struct cgraph_edge
*e
)
1463 struct cgraph_node
*callee
= e
->callee
;
1465 if (e
->indirect_unknown_callee
)
1466 e
->inline_failed
= CIF_INDIRECT_UNKNOWN_CALL
;
1467 else if (!callee
->definition
)
1468 e
->inline_failed
= CIF_BODY_NOT_AVAILABLE
;
1469 else if (callee
->local
.redefined_extern_inline
)
1470 e
->inline_failed
= CIF_REDEFINED_EXTERN_INLINE
;
1471 else if (e
->call_stmt_cannot_inline_p
)
1472 e
->inline_failed
= CIF_MISMATCHED_ARGUMENTS
;
1473 else if (cfun
&& fn_contains_cilk_spawn_p (cfun
))
1474 /* We can't inline if the function is spawing a function. */
1475 e
->inline_failed
= CIF_FUNCTION_NOT_INLINABLE
;
1477 e
->inline_failed
= CIF_FUNCTION_NOT_CONSIDERED
;
1480 /* Callback of walk_aliased_vdefs. Flags that it has been invoked to the
1481 boolean variable pointed to by DATA. */
1484 mark_modified (ao_ref
*ao ATTRIBUTE_UNUSED
, tree vdef ATTRIBUTE_UNUSED
,
1487 bool *b
= (bool *) data
;
1492 /* If OP refers to value of function parameter, return the corresponding
1496 unmodified_parm_1 (gimple stmt
, tree op
)
1498 /* SSA_NAME referring to parm default def? */
1499 if (TREE_CODE (op
) == SSA_NAME
1500 && SSA_NAME_IS_DEFAULT_DEF (op
)
1501 && TREE_CODE (SSA_NAME_VAR (op
)) == PARM_DECL
)
1502 return SSA_NAME_VAR (op
);
1503 /* Non-SSA parm reference? */
1504 if (TREE_CODE (op
) == PARM_DECL
)
1506 bool modified
= false;
1509 ao_ref_init (&refd
, op
);
1510 walk_aliased_vdefs (&refd
, gimple_vuse (stmt
), mark_modified
, &modified
,
1518 /* If OP refers to value of function parameter, return the corresponding
1519 parameter. Also traverse chains of SSA register assignments. */
1522 unmodified_parm (gimple stmt
, tree op
)
1524 tree res
= unmodified_parm_1 (stmt
, op
);
1528 if (TREE_CODE (op
) == SSA_NAME
1529 && !SSA_NAME_IS_DEFAULT_DEF (op
)
1530 && gimple_assign_single_p (SSA_NAME_DEF_STMT (op
)))
1531 return unmodified_parm (SSA_NAME_DEF_STMT (op
),
1532 gimple_assign_rhs1 (SSA_NAME_DEF_STMT (op
)));
1536 /* If OP refers to a value of a function parameter or value loaded from an
1537 aggregate passed to a parameter (either by value or reference), return TRUE
1538 and store the number of the parameter to *INDEX_P and information whether
1539 and how it has been loaded from an aggregate into *AGGPOS. INFO describes
1540 the function parameters, STMT is the statement in which OP is used or
1544 unmodified_parm_or_parm_agg_item (struct ipa_node_params
*info
,
1545 gimple stmt
, tree op
, int *index_p
,
1546 struct agg_position_info
*aggpos
)
1548 tree res
= unmodified_parm_1 (stmt
, op
);
1550 gcc_checking_assert (aggpos
);
1553 *index_p
= ipa_get_param_decl_index (info
, res
);
1556 aggpos
->agg_contents
= false;
1557 aggpos
->by_ref
= false;
1561 if (TREE_CODE (op
) == SSA_NAME
)
1563 if (SSA_NAME_IS_DEFAULT_DEF (op
)
1564 || !gimple_assign_single_p (SSA_NAME_DEF_STMT (op
)))
1566 stmt
= SSA_NAME_DEF_STMT (op
);
1567 op
= gimple_assign_rhs1 (stmt
);
1568 if (!REFERENCE_CLASS_P (op
))
1569 return unmodified_parm_or_parm_agg_item (info
, stmt
, op
, index_p
,
1573 aggpos
->agg_contents
= true;
1574 return ipa_load_from_parm_agg (info
, stmt
, op
, index_p
, &aggpos
->offset
,
1578 /* See if statement might disappear after inlining.
1579 0 - means not eliminated
1580 1 - half of statements goes away
1581 2 - for sure it is eliminated.
1582 We are not terribly sophisticated, basically looking for simple abstraction
1583 penalty wrappers. */
1586 eliminated_by_inlining_prob (gimple stmt
)
1588 enum gimple_code code
= gimple_code (stmt
);
1589 enum tree_code rhs_code
;
1599 if (gimple_num_ops (stmt
) != 2)
1602 rhs_code
= gimple_assign_rhs_code (stmt
);
1604 /* Casts of parameters, loads from parameters passed by reference
1605 and stores to return value or parameters are often free after
1606 inlining dua to SRA and further combining.
1607 Assume that half of statements goes away. */
1608 if (rhs_code
== CONVERT_EXPR
1609 || rhs_code
== NOP_EXPR
1610 || rhs_code
== VIEW_CONVERT_EXPR
1611 || rhs_code
== ADDR_EXPR
1612 || gimple_assign_rhs_class (stmt
) == GIMPLE_SINGLE_RHS
)
1614 tree rhs
= gimple_assign_rhs1 (stmt
);
1615 tree lhs
= gimple_assign_lhs (stmt
);
1616 tree inner_rhs
= get_base_address (rhs
);
1617 tree inner_lhs
= get_base_address (lhs
);
1618 bool rhs_free
= false;
1619 bool lhs_free
= false;
1626 /* Reads of parameter are expected to be free. */
1627 if (unmodified_parm (stmt
, inner_rhs
))
1629 /* Match expressions of form &this->field. Those will most likely
1630 combine with something upstream after inlining. */
1631 else if (TREE_CODE (inner_rhs
) == ADDR_EXPR
)
1633 tree op
= get_base_address (TREE_OPERAND (inner_rhs
, 0));
1634 if (TREE_CODE (op
) == PARM_DECL
)
1636 else if (TREE_CODE (op
) == MEM_REF
1637 && unmodified_parm (stmt
, TREE_OPERAND (op
, 0)))
1641 /* When parameter is not SSA register because its address is taken
1642 and it is just copied into one, the statement will be completely
1643 free after inlining (we will copy propagate backward). */
1644 if (rhs_free
&& is_gimple_reg (lhs
))
1647 /* Reads of parameters passed by reference
1648 expected to be free (i.e. optimized out after inlining). */
1649 if (TREE_CODE (inner_rhs
) == MEM_REF
1650 && unmodified_parm (stmt
, TREE_OPERAND (inner_rhs
, 0)))
1653 /* Copying parameter passed by reference into gimple register is
1654 probably also going to copy propagate, but we can't be quite
1656 if (rhs_free
&& is_gimple_reg (lhs
))
1659 /* Writes to parameters, parameters passed by value and return value
1660 (either dirrectly or passed via invisible reference) are free.
1662 TODO: We ought to handle testcase like
1663 struct a {int a,b;};
1665 retrurnsturct (void)
1671 This translate into:
1686 For that we either need to copy ipa-split logic detecting writes
1688 if (TREE_CODE (inner_lhs
) == PARM_DECL
1689 || TREE_CODE (inner_lhs
) == RESULT_DECL
1690 || (TREE_CODE (inner_lhs
) == MEM_REF
1691 && (unmodified_parm (stmt
, TREE_OPERAND (inner_lhs
, 0))
1692 || (TREE_CODE (TREE_OPERAND (inner_lhs
, 0)) == SSA_NAME
1693 && SSA_NAME_VAR (TREE_OPERAND (inner_lhs
, 0))
1694 && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND
1696 0))) == RESULT_DECL
))))
1699 && (is_gimple_reg (rhs
) || is_gimple_min_invariant (rhs
)))
1701 if (lhs_free
&& rhs_free
)
1711 /* If BB ends by a conditional we can turn into predicates, attach corresponding
1712 predicates to the CFG edges. */
1715 set_cond_stmt_execution_predicate (struct ipa_node_params
*info
,
1716 struct inline_summary
*summary
,
1722 struct agg_position_info aggpos
;
1723 enum tree_code code
, inverted_code
;
1729 last
= last_stmt (bb
);
1730 if (!last
|| gimple_code (last
) != GIMPLE_COND
)
1732 if (!is_gimple_ip_invariant (gimple_cond_rhs (last
)))
1734 op
= gimple_cond_lhs (last
);
1735 /* TODO: handle conditionals like
1738 if (unmodified_parm_or_parm_agg_item (info
, last
, op
, &index
, &aggpos
))
1740 code
= gimple_cond_code (last
);
1742 = invert_tree_comparison (code
,
1743 HONOR_NANS (TYPE_MODE (TREE_TYPE (op
))));
1745 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1747 enum tree_code this_code
= (e
->flags
& EDGE_TRUE_VALUE
1748 ? code
: inverted_code
);
1749 /* invert_tree_comparison will return ERROR_MARK on FP
1750 comparsions that are not EQ/NE instead of returning proper
1751 unordered one. Be sure it is not confused with NON_CONSTANT. */
1752 if (this_code
!= ERROR_MARK
)
1754 struct predicate p
= add_condition (summary
, index
, &aggpos
,
1756 gimple_cond_rhs (last
));
1757 e
->aux
= pool_alloc (edge_predicate_pool
);
1758 *(struct predicate
*) e
->aux
= p
;
1763 if (TREE_CODE (op
) != SSA_NAME
)
1766 if (builtin_constant_p (op))
1770 Here we can predicate nonconstant_code. We can't
1771 really handle constant_code since we have no predicate
1772 for this and also the constant code is not known to be
1773 optimized away when inliner doen't see operand is constant.
1774 Other optimizers might think otherwise. */
1775 if (gimple_cond_code (last
) != NE_EXPR
1776 || !integer_zerop (gimple_cond_rhs (last
)))
1778 set_stmt
= SSA_NAME_DEF_STMT (op
);
1779 if (!gimple_call_builtin_p (set_stmt
, BUILT_IN_CONSTANT_P
)
1780 || gimple_call_num_args (set_stmt
) != 1)
1782 op2
= gimple_call_arg (set_stmt
, 0);
1783 if (!unmodified_parm_or_parm_agg_item
1784 (info
, set_stmt
, op2
, &index
, &aggpos
))
1786 FOR_EACH_EDGE (e
, ei
, bb
->succs
) if (e
->flags
& EDGE_FALSE_VALUE
)
1788 struct predicate p
= add_condition (summary
, index
, &aggpos
,
1789 IS_NOT_CONSTANT
, NULL_TREE
);
1790 e
->aux
= pool_alloc (edge_predicate_pool
);
1791 *(struct predicate
*) e
->aux
= p
;
1796 /* If BB ends by a switch we can turn into predicates, attach corresponding
1797 predicates to the CFG edges. */
1800 set_switch_stmt_execution_predicate (struct ipa_node_params
*info
,
1801 struct inline_summary
*summary
,
1807 struct agg_position_info aggpos
;
1813 last
= last_stmt (bb
);
1814 if (!last
|| gimple_code (last
) != GIMPLE_SWITCH
)
1816 op
= gimple_switch_index (last
);
1817 if (!unmodified_parm_or_parm_agg_item (info
, last
, op
, &index
, &aggpos
))
1820 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1822 e
->aux
= pool_alloc (edge_predicate_pool
);
1823 *(struct predicate
*) e
->aux
= false_predicate ();
1825 n
= gimple_switch_num_labels (last
);
1826 for (case_idx
= 0; case_idx
< n
; ++case_idx
)
1828 tree cl
= gimple_switch_label (last
, case_idx
);
1832 e
= find_edge (bb
, label_to_block (CASE_LABEL (cl
)));
1833 min
= CASE_LOW (cl
);
1834 max
= CASE_HIGH (cl
);
1836 /* For default we might want to construct predicate that none
1837 of cases is met, but it is bit hard to do not having negations
1838 of conditionals handy. */
1840 p
= true_predicate ();
1842 p
= add_condition (summary
, index
, &aggpos
, EQ_EXPR
, min
);
1845 struct predicate p1
, p2
;
1846 p1
= add_condition (summary
, index
, &aggpos
, GE_EXPR
, min
);
1847 p2
= add_condition (summary
, index
, &aggpos
, LE_EXPR
, max
);
1848 p
= and_predicates (summary
->conds
, &p1
, &p2
);
1850 *(struct predicate
*) e
->aux
1851 = or_predicates (summary
->conds
, &p
, (struct predicate
*) e
->aux
);
1856 /* For each BB in NODE attach to its AUX pointer predicate under
1857 which it is executable. */
1860 compute_bb_predicates (struct cgraph_node
*node
,
1861 struct ipa_node_params
*parms_info
,
1862 struct inline_summary
*summary
)
1864 struct function
*my_function
= DECL_STRUCT_FUNCTION (node
->decl
);
1868 FOR_EACH_BB_FN (bb
, my_function
)
1870 set_cond_stmt_execution_predicate (parms_info
, summary
, bb
);
1871 set_switch_stmt_execution_predicate (parms_info
, summary
, bb
);
1874 /* Entry block is always executable. */
1875 ENTRY_BLOCK_PTR_FOR_FN (my_function
)->aux
1876 = pool_alloc (edge_predicate_pool
);
1877 *(struct predicate
*) ENTRY_BLOCK_PTR_FOR_FN (my_function
)->aux
1878 = true_predicate ();
1880 /* A simple dataflow propagation of predicates forward in the CFG.
1881 TODO: work in reverse postorder. */
1885 FOR_EACH_BB_FN (bb
, my_function
)
1887 struct predicate p
= false_predicate ();
1890 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
1894 struct predicate this_bb_predicate
1895 = *(struct predicate
*) e
->src
->aux
;
1898 = and_predicates (summary
->conds
, &this_bb_predicate
,
1899 (struct predicate
*) e
->aux
);
1900 p
= or_predicates (summary
->conds
, &p
, &this_bb_predicate
);
1901 if (true_predicate_p (&p
))
1905 if (false_predicate_p (&p
))
1906 gcc_assert (!bb
->aux
);
1912 bb
->aux
= pool_alloc (edge_predicate_pool
);
1913 *((struct predicate
*) bb
->aux
) = p
;
1915 else if (!predicates_equal_p (&p
, (struct predicate
*) bb
->aux
))
1917 /* This OR operation is needed to ensure monotonous data flow
1918 in the case we hit the limit on number of clauses and the
1919 and/or operations above give approximate answers. */
1920 p
= or_predicates (summary
->conds
, &p
, (struct predicate
*)bb
->aux
);
1921 if (!predicates_equal_p (&p
, (struct predicate
*) bb
->aux
))
1924 *((struct predicate
*) bb
->aux
) = p
;
1933 /* We keep info about constantness of SSA names. */
1935 typedef struct predicate predicate_t
;
1936 /* Return predicate specifying when the STMT might have result that is not
1937 a compile time constant. */
1939 static struct predicate
1940 will_be_nonconstant_expr_predicate (struct ipa_node_params
*info
,
1941 struct inline_summary
*summary
,
1943 vec
<predicate_t
> nonconstant_names
)
1948 while (UNARY_CLASS_P (expr
))
1949 expr
= TREE_OPERAND (expr
, 0);
1951 parm
= unmodified_parm (NULL
, expr
);
1952 if (parm
&& (index
= ipa_get_param_decl_index (info
, parm
)) >= 0)
1953 return add_condition (summary
, index
, NULL
, CHANGED
, NULL_TREE
);
1954 if (is_gimple_min_invariant (expr
))
1955 return false_predicate ();
1956 if (TREE_CODE (expr
) == SSA_NAME
)
1957 return nonconstant_names
[SSA_NAME_VERSION (expr
)];
1958 if (BINARY_CLASS_P (expr
) || COMPARISON_CLASS_P (expr
))
1960 struct predicate p1
= will_be_nonconstant_expr_predicate
1961 (info
, summary
, TREE_OPERAND (expr
, 0),
1963 struct predicate p2
;
1964 if (true_predicate_p (&p1
))
1966 p2
= will_be_nonconstant_expr_predicate (info
, summary
,
1967 TREE_OPERAND (expr
, 1),
1969 return or_predicates (summary
->conds
, &p1
, &p2
);
1971 else if (TREE_CODE (expr
) == COND_EXPR
)
1973 struct predicate p1
= will_be_nonconstant_expr_predicate
1974 (info
, summary
, TREE_OPERAND (expr
, 0),
1976 struct predicate p2
;
1977 if (true_predicate_p (&p1
))
1979 p2
= will_be_nonconstant_expr_predicate (info
, summary
,
1980 TREE_OPERAND (expr
, 1),
1982 if (true_predicate_p (&p2
))
1984 p1
= or_predicates (summary
->conds
, &p1
, &p2
);
1985 p2
= will_be_nonconstant_expr_predicate (info
, summary
,
1986 TREE_OPERAND (expr
, 2),
1988 return or_predicates (summary
->conds
, &p1
, &p2
);
1995 return false_predicate ();
1999 /* Return predicate specifying when the STMT might have result that is not
2000 a compile time constant. */
2002 static struct predicate
2003 will_be_nonconstant_predicate (struct ipa_node_params
*info
,
2004 struct inline_summary
*summary
,
2006 vec
<predicate_t
> nonconstant_names
)
2008 struct predicate p
= true_predicate ();
2011 struct predicate op_non_const
;
2014 struct agg_position_info aggpos
;
2016 /* What statments might be optimized away
2017 when their arguments are constant
2018 TODO: also trivial builtins.
2019 builtin_constant_p is already handled later. */
2020 if (gimple_code (stmt
) != GIMPLE_ASSIGN
2021 && gimple_code (stmt
) != GIMPLE_COND
2022 && gimple_code (stmt
) != GIMPLE_SWITCH
)
2025 /* Stores will stay anyway. */
2026 if (gimple_store_p (stmt
))
2029 is_load
= gimple_assign_load_p (stmt
);
2031 /* Loads can be optimized when the value is known. */
2035 gcc_assert (gimple_assign_single_p (stmt
));
2036 op
= gimple_assign_rhs1 (stmt
);
2037 if (!unmodified_parm_or_parm_agg_item (info
, stmt
, op
, &base_index
,
2044 /* See if we understand all operands before we start
2045 adding conditionals. */
2046 FOR_EACH_SSA_TREE_OPERAND (use
, stmt
, iter
, SSA_OP_USE
)
2048 tree parm
= unmodified_parm (stmt
, use
);
2049 /* For arguments we can build a condition. */
2050 if (parm
&& ipa_get_param_decl_index (info
, parm
) >= 0)
2052 if (TREE_CODE (use
) != SSA_NAME
)
2054 /* If we know when operand is constant,
2055 we still can say something useful. */
2056 if (!true_predicate_p (&nonconstant_names
[SSA_NAME_VERSION (use
)]))
2063 add_condition (summary
, base_index
, &aggpos
, CHANGED
, NULL
);
2065 op_non_const
= false_predicate ();
2066 FOR_EACH_SSA_TREE_OPERAND (use
, stmt
, iter
, SSA_OP_USE
)
2068 tree parm
= unmodified_parm (stmt
, use
);
2071 if (parm
&& (index
= ipa_get_param_decl_index (info
, parm
)) >= 0)
2073 if (index
!= base_index
)
2074 p
= add_condition (summary
, index
, NULL
, CHANGED
, NULL_TREE
);
2079 p
= nonconstant_names
[SSA_NAME_VERSION (use
)];
2080 op_non_const
= or_predicates (summary
->conds
, &p
, &op_non_const
);
2082 if (gimple_code (stmt
) == GIMPLE_ASSIGN
2083 && TREE_CODE (gimple_assign_lhs (stmt
)) == SSA_NAME
)
2084 nonconstant_names
[SSA_NAME_VERSION (gimple_assign_lhs (stmt
))]
2086 return op_non_const
;
2089 struct record_modified_bb_info
2095 /* Callback of walk_aliased_vdefs. Records basic blocks where the value may be
2096 set except for info->stmt. */
2099 record_modified (ao_ref
*ao ATTRIBUTE_UNUSED
, tree vdef
, void *data
)
2101 struct record_modified_bb_info
*info
=
2102 (struct record_modified_bb_info
*) data
;
2103 if (SSA_NAME_DEF_STMT (vdef
) == info
->stmt
)
2105 bitmap_set_bit (info
->bb_set
,
2106 SSA_NAME_IS_DEFAULT_DEF (vdef
)
2107 ? ENTRY_BLOCK_PTR_FOR_FN (cfun
)->index
2108 : gimple_bb (SSA_NAME_DEF_STMT (vdef
))->index
);
2112 /* Return probability (based on REG_BR_PROB_BASE) that I-th parameter of STMT
2113 will change since last invocation of STMT.
2115 Value 0 is reserved for compile time invariants.
2116 For common parameters it is REG_BR_PROB_BASE. For loop invariants it
2117 ought to be REG_BR_PROB_BASE / estimated_iters. */
2120 param_change_prob (gimple stmt
, int i
)
2122 tree op
= gimple_call_arg (stmt
, i
);
2123 basic_block bb
= gimple_bb (stmt
);
2126 /* Global invariants neve change. */
2127 if (is_gimple_min_invariant (op
))
2129 /* We would have to do non-trivial analysis to really work out what
2130 is the probability of value to change (i.e. when init statement
2131 is in a sibling loop of the call).
2133 We do an conservative estimate: when call is executed N times more often
2134 than the statement defining value, we take the frequency 1/N. */
2135 if (TREE_CODE (op
) == SSA_NAME
)
2140 return REG_BR_PROB_BASE
;
2142 if (SSA_NAME_IS_DEFAULT_DEF (op
))
2143 init_freq
= ENTRY_BLOCK_PTR_FOR_FN (cfun
)->frequency
;
2145 init_freq
= gimple_bb (SSA_NAME_DEF_STMT (op
))->frequency
;
2149 if (init_freq
< bb
->frequency
)
2150 return MAX (GCOV_COMPUTE_SCALE (init_freq
, bb
->frequency
), 1);
2152 return REG_BR_PROB_BASE
;
2155 base
= get_base_address (op
);
2160 struct record_modified_bb_info info
;
2163 tree init
= ctor_for_folding (base
);
2165 if (init
!= error_mark_node
)
2168 return REG_BR_PROB_BASE
;
2169 ao_ref_init (&refd
, op
);
2171 info
.bb_set
= BITMAP_ALLOC (NULL
);
2172 walk_aliased_vdefs (&refd
, gimple_vuse (stmt
), record_modified
, &info
,
2174 if (bitmap_bit_p (info
.bb_set
, bb
->index
))
2176 BITMAP_FREE (info
.bb_set
);
2177 return REG_BR_PROB_BASE
;
2180 /* Assume that every memory is initialized at entry.
2181 TODO: Can we easilly determine if value is always defined
2182 and thus we may skip entry block? */
2183 if (ENTRY_BLOCK_PTR_FOR_FN (cfun
)->frequency
)
2184 max
= ENTRY_BLOCK_PTR_FOR_FN (cfun
)->frequency
;
2188 EXECUTE_IF_SET_IN_BITMAP (info
.bb_set
, 0, index
, bi
)
2189 max
= MIN (max
, BASIC_BLOCK_FOR_FN (cfun
, index
)->frequency
);
2191 BITMAP_FREE (info
.bb_set
);
2192 if (max
< bb
->frequency
)
2193 return MAX (GCOV_COMPUTE_SCALE (max
, bb
->frequency
), 1);
2195 return REG_BR_PROB_BASE
;
2197 return REG_BR_PROB_BASE
;
2200 /* Find whether a basic block BB is the final block of a (half) diamond CFG
2201 sub-graph and if the predicate the condition depends on is known. If so,
2202 return true and store the pointer the predicate in *P. */
2205 phi_result_unknown_predicate (struct ipa_node_params
*info
,
2206 struct inline_summary
*summary
, basic_block bb
,
2207 struct predicate
*p
,
2208 vec
<predicate_t
> nonconstant_names
)
2212 basic_block first_bb
= NULL
;
2215 if (single_pred_p (bb
))
2217 *p
= false_predicate ();
2221 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
2223 if (single_succ_p (e
->src
))
2225 if (!single_pred_p (e
->src
))
2228 first_bb
= single_pred (e
->src
);
2229 else if (single_pred (e
->src
) != first_bb
)
2236 else if (e
->src
!= first_bb
)
2244 stmt
= last_stmt (first_bb
);
2246 || gimple_code (stmt
) != GIMPLE_COND
2247 || !is_gimple_ip_invariant (gimple_cond_rhs (stmt
)))
2250 *p
= will_be_nonconstant_expr_predicate (info
, summary
,
2251 gimple_cond_lhs (stmt
),
2253 if (true_predicate_p (p
))
2259 /* Given a PHI statement in a function described by inline properties SUMMARY
2260 and *P being the predicate describing whether the selected PHI argument is
2261 known, store a predicate for the result of the PHI statement into
2262 NONCONSTANT_NAMES, if possible. */
2265 predicate_for_phi_result (struct inline_summary
*summary
, gimple phi
,
2266 struct predicate
*p
,
2267 vec
<predicate_t
> nonconstant_names
)
2271 for (i
= 0; i
< gimple_phi_num_args (phi
); i
++)
2273 tree arg
= gimple_phi_arg (phi
, i
)->def
;
2274 if (!is_gimple_min_invariant (arg
))
2276 gcc_assert (TREE_CODE (arg
) == SSA_NAME
);
2277 *p
= or_predicates (summary
->conds
, p
,
2278 &nonconstant_names
[SSA_NAME_VERSION (arg
)]);
2279 if (true_predicate_p (p
))
2284 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2286 fprintf (dump_file
, "\t\tphi predicate: ");
2287 dump_predicate (dump_file
, summary
->conds
, p
);
2289 nonconstant_names
[SSA_NAME_VERSION (gimple_phi_result (phi
))] = *p
;
2292 /* Return predicate specifying when array index in access OP becomes non-constant. */
2294 static struct predicate
2295 array_index_predicate (struct inline_summary
*info
,
2296 vec
< predicate_t
> nonconstant_names
, tree op
)
2298 struct predicate p
= false_predicate ();
2299 while (handled_component_p (op
))
2301 if (TREE_CODE (op
) == ARRAY_REF
|| TREE_CODE (op
) == ARRAY_RANGE_REF
)
2303 if (TREE_CODE (TREE_OPERAND (op
, 1)) == SSA_NAME
)
2304 p
= or_predicates (info
->conds
, &p
,
2305 &nonconstant_names
[SSA_NAME_VERSION
2306 (TREE_OPERAND (op
, 1))]);
2308 op
= TREE_OPERAND (op
, 0);
2313 /* For a typical usage of __builtin_expect (a<b, 1), we
2314 may introduce an extra relation stmt:
2315 With the builtin, we have
2318 t3 = __builtin_expect (t2, 1);
2321 Without the builtin, we have
2324 This affects the size/time estimation and may have
2325 an impact on the earlier inlining.
2326 Here find this pattern and fix it up later. */
2329 find_foldable_builtin_expect (basic_block bb
)
2331 gimple_stmt_iterator bsi
;
2333 for (bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
2335 gimple stmt
= gsi_stmt (bsi
);
2336 if (gimple_call_builtin_p (stmt
, BUILT_IN_EXPECT
)
2337 || (is_gimple_call (stmt
)
2338 && gimple_call_internal_p (stmt
)
2339 && gimple_call_internal_fn (stmt
) == IFN_BUILTIN_EXPECT
))
2341 tree var
= gimple_call_lhs (stmt
);
2342 tree arg
= gimple_call_arg (stmt
, 0);
2343 use_operand_p use_p
;
2350 gcc_assert (TREE_CODE (var
) == SSA_NAME
);
2352 while (TREE_CODE (arg
) == SSA_NAME
)
2354 gimple stmt_tmp
= SSA_NAME_DEF_STMT (arg
);
2355 if (!is_gimple_assign (stmt_tmp
))
2357 switch (gimple_assign_rhs_code (stmt_tmp
))
2376 arg
= gimple_assign_rhs1 (stmt_tmp
);
2379 if (match
&& single_imm_use (var
, &use_p
, &use_stmt
)
2380 && gimple_code (use_stmt
) == GIMPLE_COND
)
2387 /* Return true when the basic blocks contains only clobbers followed by RESX.
2388 Such BBs are kept around to make removal of dead stores possible with
2389 presence of EH and will be optimized out by optimize_clobbers later in the
2392 NEED_EH is used to recurse in case the clobber has non-EH predecestors
2393 that can be clobber only, too.. When it is false, the RESX is not necessary
2394 on the end of basic block. */
2397 clobber_only_eh_bb_p (basic_block bb
, bool need_eh
= true)
2399 gimple_stmt_iterator gsi
= gsi_last_bb (bb
);
2405 if (gsi_end_p (gsi
))
2407 if (gimple_code (gsi_stmt (gsi
)) != GIMPLE_RESX
)
2411 else if (!single_succ_p (bb
))
2414 for (; !gsi_end_p (gsi
); gsi_prev (&gsi
))
2416 gimple stmt
= gsi_stmt (gsi
);
2417 if (is_gimple_debug (stmt
))
2419 if (gimple_clobber_p (stmt
))
2421 if (gimple_code (stmt
) == GIMPLE_LABEL
)
2426 /* See if all predecestors are either throws or clobber only BBs. */
2427 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
2428 if (!(e
->flags
& EDGE_EH
)
2429 && !clobber_only_eh_bb_p (e
->src
, false))
2435 /* Compute function body size parameters for NODE.
2436 When EARLY is true, we compute only simple summaries without
2437 non-trivial predicates to drive the early inliner. */
2440 estimate_function_body_sizes (struct cgraph_node
*node
, bool early
)
2443 /* Estimate static overhead for function prologue/epilogue and alignment. */
2445 /* Benefits are scaled by probability of elimination that is in range
2448 gimple_stmt_iterator bsi
;
2449 struct function
*my_function
= DECL_STRUCT_FUNCTION (node
->decl
);
2451 struct inline_summary
*info
= inline_summary (node
);
2452 struct predicate bb_predicate
;
2453 struct ipa_node_params
*parms_info
= NULL
;
2454 vec
<predicate_t
> nonconstant_names
= vNULL
;
2457 predicate array_index
= true_predicate ();
2458 gimple fix_builtin_expect_stmt
;
2463 if (optimize
&& !early
)
2465 calculate_dominance_info (CDI_DOMINATORS
);
2466 loop_optimizer_init (LOOPS_NORMAL
| LOOPS_HAVE_RECORDED_EXITS
);
2468 if (ipa_node_params_vector
.exists ())
2470 parms_info
= IPA_NODE_REF (node
);
2471 nonconstant_names
.safe_grow_cleared
2472 (SSANAMES (my_function
)->length ());
2477 fprintf (dump_file
, "\nAnalyzing function body size: %s\n",
2480 /* When we run into maximal number of entries, we assign everything to the
2481 constant truth case. Be sure to have it in list. */
2482 bb_predicate
= true_predicate ();
2483 account_size_time (info
, 0, 0, &bb_predicate
);
2485 bb_predicate
= not_inlined_predicate ();
2486 account_size_time (info
, 2 * INLINE_SIZE_SCALE
, 0, &bb_predicate
);
2488 gcc_assert (my_function
&& my_function
->cfg
);
2490 compute_bb_predicates (node
, parms_info
, info
);
2491 gcc_assert (cfun
== my_function
);
2492 order
= XNEWVEC (int, n_basic_blocks_for_fn (cfun
));
2493 nblocks
= pre_and_rev_post_order_compute (NULL
, order
, false);
2494 for (n
= 0; n
< nblocks
; n
++)
2496 bb
= BASIC_BLOCK_FOR_FN (cfun
, order
[n
]);
2497 freq
= compute_call_stmt_bb_frequency (node
->decl
, bb
);
2498 if (clobber_only_eh_bb_p (bb
))
2500 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2501 fprintf (dump_file
, "\n Ignoring BB %i;"
2502 " it will be optimized away by cleanup_clobbers\n",
2507 /* TODO: Obviously predicates can be propagated down across CFG. */
2511 bb_predicate
= *(struct predicate
*) bb
->aux
;
2513 bb_predicate
= false_predicate ();
2516 bb_predicate
= true_predicate ();
2518 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2520 fprintf (dump_file
, "\n BB %i predicate:", bb
->index
);
2521 dump_predicate (dump_file
, info
->conds
, &bb_predicate
);
2524 if (parms_info
&& nonconstant_names
.exists ())
2526 struct predicate phi_predicate
;
2527 bool first_phi
= true;
2529 for (bsi
= gsi_start_phis (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
2532 && !phi_result_unknown_predicate (parms_info
, info
, bb
,
2537 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2539 fprintf (dump_file
, " ");
2540 print_gimple_stmt (dump_file
, gsi_stmt (bsi
), 0, 0);
2542 predicate_for_phi_result (info
, gsi_stmt (bsi
), &phi_predicate
,
2547 fix_builtin_expect_stmt
= find_foldable_builtin_expect (bb
);
2549 for (bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
2551 gimple stmt
= gsi_stmt (bsi
);
2552 int this_size
= estimate_num_insns (stmt
, &eni_size_weights
);
2553 int this_time
= estimate_num_insns (stmt
, &eni_time_weights
);
2555 struct predicate will_be_nonconstant
;
2557 /* This relation stmt should be folded after we remove
2558 buildin_expect call. Adjust the cost here. */
2559 if (stmt
== fix_builtin_expect_stmt
)
2565 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2567 fprintf (dump_file
, " ");
2568 print_gimple_stmt (dump_file
, stmt
, 0, 0);
2569 fprintf (dump_file
, "\t\tfreq:%3.2f size:%3i time:%3i\n",
2570 ((double) freq
) / CGRAPH_FREQ_BASE
, this_size
,
2574 if (gimple_assign_load_p (stmt
) && nonconstant_names
.exists ())
2576 struct predicate this_array_index
;
2578 array_index_predicate (info
, nonconstant_names
,
2579 gimple_assign_rhs1 (stmt
));
2580 if (!false_predicate_p (&this_array_index
))
2582 and_predicates (info
->conds
, &array_index
,
2585 if (gimple_store_p (stmt
) && nonconstant_names
.exists ())
2587 struct predicate this_array_index
;
2589 array_index_predicate (info
, nonconstant_names
,
2590 gimple_get_lhs (stmt
));
2591 if (!false_predicate_p (&this_array_index
))
2593 and_predicates (info
->conds
, &array_index
,
2598 if (is_gimple_call (stmt
)
2599 && !gimple_call_internal_p (stmt
))
2601 struct cgraph_edge
*edge
= cgraph_edge (node
, stmt
);
2602 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
2604 /* Special case: results of BUILT_IN_CONSTANT_P will be always
2605 resolved as constant. We however don't want to optimize
2606 out the cgraph edges. */
2607 if (nonconstant_names
.exists ()
2608 && gimple_call_builtin_p (stmt
, BUILT_IN_CONSTANT_P
)
2609 && gimple_call_lhs (stmt
)
2610 && TREE_CODE (gimple_call_lhs (stmt
)) == SSA_NAME
)
2612 struct predicate false_p
= false_predicate ();
2613 nonconstant_names
[SSA_NAME_VERSION (gimple_call_lhs (stmt
))]
2616 if (ipa_node_params_vector
.exists ())
2618 int count
= gimple_call_num_args (stmt
);
2622 es
->param
.safe_grow_cleared (count
);
2623 for (i
= 0; i
< count
; i
++)
2625 int prob
= param_change_prob (stmt
, i
);
2626 gcc_assert (prob
>= 0 && prob
<= REG_BR_PROB_BASE
);
2627 es
->param
[i
].change_prob
= prob
;
2631 es
->call_stmt_size
= this_size
;
2632 es
->call_stmt_time
= this_time
;
2633 es
->loop_depth
= bb_loop_depth (bb
);
2634 edge_set_predicate (edge
, &bb_predicate
);
2637 /* TODO: When conditional jump or swithc is known to be constant, but
2638 we did not translate it into the predicates, we really can account
2639 just maximum of the possible paths. */
2642 = will_be_nonconstant_predicate (parms_info
, info
,
2643 stmt
, nonconstant_names
);
2644 if (this_time
|| this_size
)
2650 prob
= eliminated_by_inlining_prob (stmt
);
2651 if (prob
== 1 && dump_file
&& (dump_flags
& TDF_DETAILS
))
2653 "\t\t50%% will be eliminated by inlining\n");
2654 if (prob
== 2 && dump_file
&& (dump_flags
& TDF_DETAILS
))
2655 fprintf (dump_file
, "\t\tWill be eliminated by inlining\n");
2658 p
= and_predicates (info
->conds
, &bb_predicate
,
2659 &will_be_nonconstant
);
2661 p
= true_predicate ();
2663 if (!false_predicate_p (&p
))
2667 if (time
> MAX_TIME
* INLINE_TIME_SCALE
)
2668 time
= MAX_TIME
* INLINE_TIME_SCALE
;
2671 /* We account everything but the calls. Calls have their own
2672 size/time info attached to cgraph edges. This is necessary
2673 in order to make the cost disappear after inlining. */
2674 if (!is_gimple_call (stmt
))
2678 struct predicate ip
= not_inlined_predicate ();
2679 ip
= and_predicates (info
->conds
, &ip
, &p
);
2680 account_size_time (info
, this_size
* prob
,
2681 this_time
* prob
, &ip
);
2684 account_size_time (info
, this_size
* (2 - prob
),
2685 this_time
* (2 - prob
), &p
);
2688 gcc_assert (time
>= 0);
2689 gcc_assert (size
>= 0);
2693 set_hint_predicate (&inline_summary (node
)->array_index
, array_index
);
2694 time
= (time
+ CGRAPH_FREQ_BASE
/ 2) / CGRAPH_FREQ_BASE
;
2695 if (time
> MAX_TIME
)
2699 if (!early
&& nonconstant_names
.exists ())
2702 predicate loop_iterations
= true_predicate ();
2703 predicate loop_stride
= true_predicate ();
2705 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2706 flow_loops_dump (dump_file
, NULL
, 0);
2708 FOR_EACH_LOOP (loop
, 0)
2713 struct tree_niter_desc niter_desc
;
2714 basic_block
*body
= get_loop_body (loop
);
2715 bb_predicate
= *(struct predicate
*) loop
->header
->aux
;
2717 exits
= get_loop_exit_edges (loop
);
2718 FOR_EACH_VEC_ELT (exits
, j
, ex
)
2719 if (number_of_iterations_exit (loop
, ex
, &niter_desc
, false)
2720 && !is_gimple_min_invariant (niter_desc
.niter
))
2722 predicate will_be_nonconstant
2723 = will_be_nonconstant_expr_predicate (parms_info
, info
,
2726 if (!true_predicate_p (&will_be_nonconstant
))
2727 will_be_nonconstant
= and_predicates (info
->conds
,
2729 &will_be_nonconstant
);
2730 if (!true_predicate_p (&will_be_nonconstant
)
2731 && !false_predicate_p (&will_be_nonconstant
))
2732 /* This is slightly inprecise. We may want to represent each
2733 loop with independent predicate. */
2735 and_predicates (info
->conds
, &loop_iterations
,
2736 &will_be_nonconstant
);
2740 for (i
= 0; i
< loop
->num_nodes
; i
++)
2742 gimple_stmt_iterator gsi
;
2743 bb_predicate
= *(struct predicate
*) body
[i
]->aux
;
2744 for (gsi
= gsi_start_bb (body
[i
]); !gsi_end_p (gsi
);
2747 gimple stmt
= gsi_stmt (gsi
);
2752 FOR_EACH_SSA_TREE_OPERAND (use
, stmt
, iter
, SSA_OP_USE
)
2754 predicate will_be_nonconstant
;
2757 (loop
, loop_containing_stmt (stmt
), use
, &iv
, true)
2758 || is_gimple_min_invariant (iv
.step
))
2761 = will_be_nonconstant_expr_predicate (parms_info
, info
,
2764 if (!true_predicate_p (&will_be_nonconstant
))
2766 = and_predicates (info
->conds
,
2768 &will_be_nonconstant
);
2769 if (!true_predicate_p (&will_be_nonconstant
)
2770 && !false_predicate_p (&will_be_nonconstant
))
2771 /* This is slightly inprecise. We may want to represent
2772 each loop with independent predicate. */
2774 and_predicates (info
->conds
, &loop_stride
,
2775 &will_be_nonconstant
);
2781 set_hint_predicate (&inline_summary (node
)->loop_iterations
,
2783 set_hint_predicate (&inline_summary (node
)->loop_stride
, loop_stride
);
2786 FOR_ALL_BB_FN (bb
, my_function
)
2792 pool_free (edge_predicate_pool
, bb
->aux
);
2794 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
2797 pool_free (edge_predicate_pool
, e
->aux
);
2801 inline_summary (node
)->self_time
= time
;
2802 inline_summary (node
)->self_size
= size
;
2803 nonconstant_names
.release ();
2804 if (optimize
&& !early
)
2806 loop_optimizer_finalize ();
2807 free_dominance_info (CDI_DOMINATORS
);
2811 fprintf (dump_file
, "\n");
2812 dump_inline_summary (dump_file
, node
);
2817 /* Compute parameters of functions used by inliner.
2818 EARLY is true when we compute parameters for the early inliner */
2821 compute_inline_parameters (struct cgraph_node
*node
, bool early
)
2823 HOST_WIDE_INT self_stack_size
;
2824 struct cgraph_edge
*e
;
2825 struct inline_summary
*info
;
2827 gcc_assert (!node
->global
.inlined_to
);
2829 inline_summary_alloc ();
2831 info
= inline_summary (node
);
2832 reset_inline_summary (node
);
2834 /* FIXME: Thunks are inlinable, but tree-inline don't know how to do that.
2835 Once this happen, we will need to more curefully predict call
2837 if (node
->thunk
.thunk_p
)
2839 struct inline_edge_summary
*es
= inline_edge_summary (node
->callees
);
2840 struct predicate t
= true_predicate ();
2842 info
->inlinable
= 0;
2843 node
->callees
->call_stmt_cannot_inline_p
= true;
2844 node
->local
.can_change_signature
= false;
2845 es
->call_stmt_time
= 1;
2846 es
->call_stmt_size
= 1;
2847 account_size_time (info
, 0, 0, &t
);
2851 /* Even is_gimple_min_invariant rely on current_function_decl. */
2852 push_cfun (DECL_STRUCT_FUNCTION (node
->decl
));
2854 /* Estimate the stack size for the function if we're optimizing. */
2855 self_stack_size
= optimize
? estimated_stack_frame_size (node
) : 0;
2856 info
->estimated_self_stack_size
= self_stack_size
;
2857 info
->estimated_stack_size
= self_stack_size
;
2858 info
->stack_frame_offset
= 0;
2860 /* Can this function be inlined at all? */
2861 if (!optimize
&& !lookup_attribute ("always_inline",
2862 DECL_ATTRIBUTES (node
->decl
)))
2863 info
->inlinable
= false;
2865 info
->inlinable
= tree_inlinable_function_p (node
->decl
);
2867 /* Type attributes can use parameter indices to describe them. */
2868 if (TYPE_ATTRIBUTES (TREE_TYPE (node
->decl
)))
2869 node
->local
.can_change_signature
= false;
2872 /* Otherwise, inlinable functions always can change signature. */
2873 if (info
->inlinable
)
2874 node
->local
.can_change_signature
= true;
2877 /* Functions calling builtin_apply can not change signature. */
2878 for (e
= node
->callees
; e
; e
= e
->next_callee
)
2880 tree
cdecl = e
->callee
->decl
;
2881 if (DECL_BUILT_IN (cdecl)
2882 && DECL_BUILT_IN_CLASS (cdecl) == BUILT_IN_NORMAL
2883 && (DECL_FUNCTION_CODE (cdecl) == BUILT_IN_APPLY_ARGS
2884 || DECL_FUNCTION_CODE (cdecl) == BUILT_IN_VA_START
))
2887 node
->local
.can_change_signature
= !e
;
2890 estimate_function_body_sizes (node
, early
);
2892 for (e
= node
->callees
; e
; e
= e
->next_callee
)
2893 if (symtab_comdat_local_p (e
->callee
))
2895 node
->calls_comdat_local
= (e
!= NULL
);
2897 /* Inlining characteristics are maintained by the cgraph_mark_inline. */
2898 info
->time
= info
->self_time
;
2899 info
->size
= info
->self_size
;
2900 info
->stack_frame_offset
= 0;
2901 info
->estimated_stack_size
= info
->estimated_self_stack_size
;
2902 #ifdef ENABLE_CHECKING
2903 inline_update_overall_summary (node
);
2904 gcc_assert (info
->time
== info
->self_time
&& info
->size
== info
->self_size
);
2911 /* Compute parameters of functions used by inliner using
2912 current_function_decl. */
2915 compute_inline_parameters_for_current (void)
2917 compute_inline_parameters (cgraph_get_node (current_function_decl
), true);
2923 const pass_data pass_data_inline_parameters
=
2925 GIMPLE_PASS
, /* type */
2926 "inline_param", /* name */
2927 OPTGROUP_INLINE
, /* optinfo_flags */
2928 true, /* has_execute */
2929 TV_INLINE_PARAMETERS
, /* tv_id */
2930 0, /* properties_required */
2931 0, /* properties_provided */
2932 0, /* properties_destroyed */
2933 0, /* todo_flags_start */
2934 0, /* todo_flags_finish */
2937 class pass_inline_parameters
: public gimple_opt_pass
2940 pass_inline_parameters (gcc::context
*ctxt
)
2941 : gimple_opt_pass (pass_data_inline_parameters
, ctxt
)
2944 /* opt_pass methods: */
2945 opt_pass
* clone () { return new pass_inline_parameters (m_ctxt
); }
2946 virtual unsigned int execute (function
*)
2948 return compute_inline_parameters_for_current ();
2951 }; // class pass_inline_parameters
2956 make_pass_inline_parameters (gcc::context
*ctxt
)
2958 return new pass_inline_parameters (ctxt
);
2962 /* Estimate benefit devirtualizing indirect edge IE, provided KNOWN_VALS and
2966 estimate_edge_devirt_benefit (struct cgraph_edge
*ie
,
2967 int *size
, int *time
,
2968 vec
<tree
> known_vals
,
2969 vec
<tree
> known_binfos
,
2970 vec
<ipa_agg_jump_function_p
> known_aggs
)
2973 struct cgraph_node
*callee
;
2974 struct inline_summary
*isummary
;
2976 if (!known_vals
.exists () && !known_binfos
.exists ())
2978 if (!flag_indirect_inlining
)
2981 target
= ipa_get_indirect_edge_target (ie
, known_vals
, known_binfos
,
2986 /* Account for difference in cost between indirect and direct calls. */
2987 *size
-= (eni_size_weights
.indirect_call_cost
- eni_size_weights
.call_cost
);
2988 *time
-= (eni_time_weights
.indirect_call_cost
- eni_time_weights
.call_cost
);
2989 gcc_checking_assert (*time
>= 0);
2990 gcc_checking_assert (*size
>= 0);
2992 callee
= cgraph_get_node (target
);
2993 if (!callee
|| !callee
->definition
)
2995 isummary
= inline_summary (callee
);
2996 return isummary
->inlinable
;
2999 /* Increase SIZE, MIN_SIZE (if non-NULL) and TIME for size and time needed to
3000 handle edge E with probability PROB.
3001 Set HINTS if edge may be devirtualized.
3002 KNOWN_VALS, KNOWN_AGGS and KNOWN_BINFOS describe context of the call
3006 estimate_edge_size_and_time (struct cgraph_edge
*e
, int *size
, int *min_size
,
3009 vec
<tree
> known_vals
,
3010 vec
<tree
> known_binfos
,
3011 vec
<ipa_agg_jump_function_p
> known_aggs
,
3012 inline_hints
*hints
)
3014 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3015 int call_size
= es
->call_stmt_size
;
3016 int call_time
= es
->call_stmt_time
;
3019 && estimate_edge_devirt_benefit (e
, &call_size
, &call_time
,
3020 known_vals
, known_binfos
, known_aggs
)
3021 && hints
&& cgraph_maybe_hot_edge_p (e
))
3022 *hints
|= INLINE_HINT_indirect_call
;
3023 cur_size
= call_size
* INLINE_SIZE_SCALE
;
3026 *min_size
+= cur_size
;
3027 *time
+= apply_probability ((gcov_type
) call_time
, prob
)
3028 * e
->frequency
* (INLINE_TIME_SCALE
/ CGRAPH_FREQ_BASE
);
3029 if (*time
> MAX_TIME
* INLINE_TIME_SCALE
)
3030 *time
= MAX_TIME
* INLINE_TIME_SCALE
;
3035 /* Increase SIZE, MIN_SIZE and TIME for size and time needed to handle all
3037 POSSIBLE_TRUTHS, KNOWN_VALS, KNOWN_AGGS and KNOWN_BINFOS describe context of
3041 estimate_calls_size_and_time (struct cgraph_node
*node
, int *size
,
3042 int *min_size
, int *time
,
3043 inline_hints
*hints
,
3044 clause_t possible_truths
,
3045 vec
<tree
> known_vals
,
3046 vec
<tree
> known_binfos
,
3047 vec
<ipa_agg_jump_function_p
> known_aggs
)
3049 struct cgraph_edge
*e
;
3050 for (e
= node
->callees
; e
; e
= e
->next_callee
)
3052 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3054 || evaluate_predicate (es
->predicate
, possible_truths
))
3056 if (e
->inline_failed
)
3058 /* Predicates of calls shall not use NOT_CHANGED codes,
3059 sowe do not need to compute probabilities. */
3060 estimate_edge_size_and_time (e
, size
,
3061 es
->predicate
? NULL
: min_size
,
3062 time
, REG_BR_PROB_BASE
,
3063 known_vals
, known_binfos
,
3067 estimate_calls_size_and_time (e
->callee
, size
, min_size
, time
,
3070 known_vals
, known_binfos
,
3074 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
3076 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3078 || evaluate_predicate (es
->predicate
, possible_truths
))
3079 estimate_edge_size_and_time (e
, size
,
3080 es
->predicate
? NULL
: min_size
,
3081 time
, REG_BR_PROB_BASE
,
3082 known_vals
, known_binfos
, known_aggs
,
3088 /* Estimate size and time needed to execute NODE assuming
3089 POSSIBLE_TRUTHS clause, and KNOWN_VALS, KNOWN_AGGS and KNOWN_BINFOS
3090 information about NODE's arguments. If non-NULL use also probability
3091 information present in INLINE_PARAM_SUMMARY vector.
3092 Additionally detemine hints determined by the context. Finally compute
3093 minimal size needed for the call that is independent on the call context and
3094 can be used for fast estimates. Return the values in RET_SIZE,
3095 RET_MIN_SIZE, RET_TIME and RET_HINTS. */
3098 estimate_node_size_and_time (struct cgraph_node
*node
,
3099 clause_t possible_truths
,
3100 vec
<tree
> known_vals
,
3101 vec
<tree
> known_binfos
,
3102 vec
<ipa_agg_jump_function_p
> known_aggs
,
3103 int *ret_size
, int *ret_min_size
, int *ret_time
,
3104 inline_hints
*ret_hints
,
3105 vec
<inline_param_summary
>
3106 inline_param_summary
)
3108 struct inline_summary
*info
= inline_summary (node
);
3113 inline_hints hints
= 0;
3116 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3119 fprintf (dump_file
, " Estimating body: %s/%i\n"
3120 " Known to be false: ", node
->name (),
3123 for (i
= predicate_not_inlined_condition
;
3124 i
< (predicate_first_dynamic_condition
3125 + (int) vec_safe_length (info
->conds
)); i
++)
3126 if (!(possible_truths
& (1 << i
)))
3129 fprintf (dump_file
, ", ");
3131 dump_condition (dump_file
, info
->conds
, i
);
3135 for (i
= 0; vec_safe_iterate (info
->entry
, i
, &e
); i
++)
3136 if (evaluate_predicate (&e
->predicate
, possible_truths
))
3139 gcc_checking_assert (e
->time
>= 0);
3140 gcc_checking_assert (time
>= 0);
3141 if (!inline_param_summary
.exists ())
3145 int prob
= predicate_probability (info
->conds
,
3148 inline_param_summary
);
3149 gcc_checking_assert (prob
>= 0);
3150 gcc_checking_assert (prob
<= REG_BR_PROB_BASE
);
3151 time
+= apply_probability ((gcov_type
) e
->time
, prob
);
3153 if (time
> MAX_TIME
* INLINE_TIME_SCALE
)
3154 time
= MAX_TIME
* INLINE_TIME_SCALE
;
3155 gcc_checking_assert (time
>= 0);
3158 gcc_checking_assert (true_predicate_p (&(*info
->entry
)[0].predicate
));
3159 min_size
= (*info
->entry
)[0].size
;
3160 gcc_checking_assert (size
>= 0);
3161 gcc_checking_assert (time
>= 0);
3163 if (info
->loop_iterations
3164 && !evaluate_predicate (info
->loop_iterations
, possible_truths
))
3165 hints
|= INLINE_HINT_loop_iterations
;
3166 if (info
->loop_stride
3167 && !evaluate_predicate (info
->loop_stride
, possible_truths
))
3168 hints
|= INLINE_HINT_loop_stride
;
3169 if (info
->array_index
3170 && !evaluate_predicate (info
->array_index
, possible_truths
))
3171 hints
|= INLINE_HINT_array_index
;
3173 hints
|= INLINE_HINT_in_scc
;
3174 if (DECL_DECLARED_INLINE_P (node
->decl
))
3175 hints
|= INLINE_HINT_declared_inline
;
3177 estimate_calls_size_and_time (node
, &size
, &min_size
, &time
, &hints
, possible_truths
,
3178 known_vals
, known_binfos
, known_aggs
);
3179 gcc_checking_assert (size
>= 0);
3180 gcc_checking_assert (time
>= 0);
3181 time
= RDIV (time
, INLINE_TIME_SCALE
);
3182 size
= RDIV (size
, INLINE_SIZE_SCALE
);
3183 min_size
= RDIV (min_size
, INLINE_SIZE_SCALE
);
3185 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3186 fprintf (dump_file
, "\n size:%i time:%i\n", (int) size
, (int) time
);
3192 *ret_min_size
= min_size
;
3199 /* Estimate size and time needed to execute callee of EDGE assuming that
3200 parameters known to be constant at caller of EDGE are propagated.
3201 KNOWN_VALS and KNOWN_BINFOS are vectors of assumed known constant values
3202 and types for parameters. */
3205 estimate_ipcp_clone_size_and_time (struct cgraph_node
*node
,
3206 vec
<tree
> known_vals
,
3207 vec
<tree
> known_binfos
,
3208 vec
<ipa_agg_jump_function_p
> known_aggs
,
3209 int *ret_size
, int *ret_time
,
3210 inline_hints
*hints
)
3214 clause
= evaluate_conditions_for_known_args (node
, false, known_vals
,
3216 estimate_node_size_and_time (node
, clause
, known_vals
, known_binfos
,
3217 known_aggs
, ret_size
, NULL
, ret_time
, hints
, vNULL
);
3220 /* Translate all conditions from callee representation into caller
3221 representation and symbolically evaluate predicate P into new predicate.
3223 INFO is inline_summary of function we are adding predicate into, CALLEE_INFO
3224 is summary of function predicate P is from. OPERAND_MAP is array giving
3225 callee formal IDs the caller formal IDs. POSSSIBLE_TRUTHS is clausule of all
3226 callee conditions that may be true in caller context. TOPLEV_PREDICATE is
3227 predicate under which callee is executed. OFFSET_MAP is an array of of
3228 offsets that need to be added to conditions, negative offset means that
3229 conditions relying on values passed by reference have to be discarded
3230 because they might not be preserved (and should be considered offset zero
3231 for other purposes). */
3233 static struct predicate
3234 remap_predicate (struct inline_summary
*info
,
3235 struct inline_summary
*callee_info
,
3236 struct predicate
*p
,
3237 vec
<int> operand_map
,
3238 vec
<int> offset_map
,
3239 clause_t possible_truths
, struct predicate
*toplev_predicate
)
3242 struct predicate out
= true_predicate ();
3244 /* True predicate is easy. */
3245 if (true_predicate_p (p
))
3246 return *toplev_predicate
;
3247 for (i
= 0; p
->clause
[i
]; i
++)
3249 clause_t clause
= p
->clause
[i
];
3251 struct predicate clause_predicate
= false_predicate ();
3253 gcc_assert (i
< MAX_CLAUSES
);
3255 for (cond
= 0; cond
< NUM_CONDITIONS
; cond
++)
3256 /* Do we have condition we can't disprove? */
3257 if (clause
& possible_truths
& (1 << cond
))
3259 struct predicate cond_predicate
;
3260 /* Work out if the condition can translate to predicate in the
3261 inlined function. */
3262 if (cond
>= predicate_first_dynamic_condition
)
3264 struct condition
*c
;
3266 c
= &(*callee_info
->conds
)[cond
3268 predicate_first_dynamic_condition
];
3269 /* See if we can remap condition operand to caller's operand.
3270 Otherwise give up. */
3271 if (!operand_map
.exists ()
3272 || (int) operand_map
.length () <= c
->operand_num
3273 || operand_map
[c
->operand_num
] == -1
3274 /* TODO: For non-aggregate conditions, adding an offset is
3275 basically an arithmetic jump function processing which
3276 we should support in future. */
3277 || ((!c
->agg_contents
|| !c
->by_ref
)
3278 && offset_map
[c
->operand_num
] > 0)
3279 || (c
->agg_contents
&& c
->by_ref
3280 && offset_map
[c
->operand_num
] < 0))
3281 cond_predicate
= true_predicate ();
3284 struct agg_position_info ap
;
3285 HOST_WIDE_INT offset_delta
= offset_map
[c
->operand_num
];
3286 if (offset_delta
< 0)
3288 gcc_checking_assert (!c
->agg_contents
|| !c
->by_ref
);
3291 gcc_assert (!c
->agg_contents
3292 || c
->by_ref
|| offset_delta
== 0);
3293 ap
.offset
= c
->offset
+ offset_delta
;
3294 ap
.agg_contents
= c
->agg_contents
;
3295 ap
.by_ref
= c
->by_ref
;
3296 cond_predicate
= add_condition (info
,
3297 operand_map
[c
->operand_num
],
3298 &ap
, c
->code
, c
->val
);
3301 /* Fixed conditions remains same, construct single
3302 condition predicate. */
3305 cond_predicate
.clause
[0] = 1 << cond
;
3306 cond_predicate
.clause
[1] = 0;
3308 clause_predicate
= or_predicates (info
->conds
, &clause_predicate
,
3311 out
= and_predicates (info
->conds
, &out
, &clause_predicate
);
3313 return and_predicates (info
->conds
, &out
, toplev_predicate
);
3317 /* Update summary information of inline clones after inlining.
3318 Compute peak stack usage. */
3321 inline_update_callee_summaries (struct cgraph_node
*node
, int depth
)
3323 struct cgraph_edge
*e
;
3324 struct inline_summary
*callee_info
= inline_summary (node
);
3325 struct inline_summary
*caller_info
= inline_summary (node
->callers
->caller
);
3328 callee_info
->stack_frame_offset
3329 = caller_info
->stack_frame_offset
3330 + caller_info
->estimated_self_stack_size
;
3331 peak
= callee_info
->stack_frame_offset
3332 + callee_info
->estimated_self_stack_size
;
3333 if (inline_summary (node
->global
.inlined_to
)->estimated_stack_size
< peak
)
3334 inline_summary (node
->global
.inlined_to
)->estimated_stack_size
= peak
;
3335 ipa_propagate_frequency (node
);
3336 for (e
= node
->callees
; e
; e
= e
->next_callee
)
3338 if (!e
->inline_failed
)
3339 inline_update_callee_summaries (e
->callee
, depth
);
3340 inline_edge_summary (e
)->loop_depth
+= depth
;
3342 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
3343 inline_edge_summary (e
)->loop_depth
+= depth
;
3346 /* Update change_prob of EDGE after INLINED_EDGE has been inlined.
3347 When functoin A is inlined in B and A calls C with parameter that
3348 changes with probability PROB1 and C is known to be passthroug
3349 of argument if B that change with probability PROB2, the probability
3350 of change is now PROB1*PROB2. */
3353 remap_edge_change_prob (struct cgraph_edge
*inlined_edge
,
3354 struct cgraph_edge
*edge
)
3356 if (ipa_node_params_vector
.exists ())
3359 struct ipa_edge_args
*args
= IPA_EDGE_REF (edge
);
3360 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3361 struct inline_edge_summary
*inlined_es
3362 = inline_edge_summary (inlined_edge
);
3364 for (i
= 0; i
< ipa_get_cs_argument_count (args
); i
++)
3366 struct ipa_jump_func
*jfunc
= ipa_get_ith_jump_func (args
, i
);
3367 if (jfunc
->type
== IPA_JF_PASS_THROUGH
3368 && (ipa_get_jf_pass_through_formal_id (jfunc
)
3369 < (int) inlined_es
->param
.length ()))
3371 int jf_formal_id
= ipa_get_jf_pass_through_formal_id (jfunc
);
3372 int prob1
= es
->param
[i
].change_prob
;
3373 int prob2
= inlined_es
->param
[jf_formal_id
].change_prob
;
3374 int prob
= combine_probabilities (prob1
, prob2
);
3376 if (prob1
&& prob2
&& !prob
)
3379 es
->param
[i
].change_prob
= prob
;
3385 /* Update edge summaries of NODE after INLINED_EDGE has been inlined.
3387 Remap predicates of callees of NODE. Rest of arguments match
3390 Also update change probabilities. */
3393 remap_edge_summaries (struct cgraph_edge
*inlined_edge
,
3394 struct cgraph_node
*node
,
3395 struct inline_summary
*info
,
3396 struct inline_summary
*callee_info
,
3397 vec
<int> operand_map
,
3398 vec
<int> offset_map
,
3399 clause_t possible_truths
,
3400 struct predicate
*toplev_predicate
)
3402 struct cgraph_edge
*e
;
3403 for (e
= node
->callees
; e
; e
= e
->next_callee
)
3405 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3408 if (e
->inline_failed
)
3410 remap_edge_change_prob (inlined_edge
, e
);
3414 p
= remap_predicate (info
, callee_info
,
3415 es
->predicate
, operand_map
, offset_map
,
3416 possible_truths
, toplev_predicate
);
3417 edge_set_predicate (e
, &p
);
3418 /* TODO: We should remove the edge for code that will be
3419 optimized out, but we need to keep verifiers and tree-inline
3420 happy. Make it cold for now. */
3421 if (false_predicate_p (&p
))
3428 edge_set_predicate (e
, toplev_predicate
);
3431 remap_edge_summaries (inlined_edge
, e
->callee
, info
, callee_info
,
3432 operand_map
, offset_map
, possible_truths
,
3435 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
3437 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3440 remap_edge_change_prob (inlined_edge
, e
);
3443 p
= remap_predicate (info
, callee_info
,
3444 es
->predicate
, operand_map
, offset_map
,
3445 possible_truths
, toplev_predicate
);
3446 edge_set_predicate (e
, &p
);
3447 /* TODO: We should remove the edge for code that will be optimized
3448 out, but we need to keep verifiers and tree-inline happy.
3449 Make it cold for now. */
3450 if (false_predicate_p (&p
))
3457 edge_set_predicate (e
, toplev_predicate
);
3461 /* Same as remap_predicate, but set result into hint *HINT. */
3464 remap_hint_predicate (struct inline_summary
*info
,
3465 struct inline_summary
*callee_info
,
3466 struct predicate
**hint
,
3467 vec
<int> operand_map
,
3468 vec
<int> offset_map
,
3469 clause_t possible_truths
,
3470 struct predicate
*toplev_predicate
)
3476 p
= remap_predicate (info
, callee_info
,
3478 operand_map
, offset_map
,
3479 possible_truths
, toplev_predicate
);
3480 if (!false_predicate_p (&p
) && !true_predicate_p (&p
))
3483 set_hint_predicate (hint
, p
);
3485 **hint
= and_predicates (info
->conds
, *hint
, &p
);
3489 /* We inlined EDGE. Update summary of the function we inlined into. */
3492 inline_merge_summary (struct cgraph_edge
*edge
)
3494 struct inline_summary
*callee_info
= inline_summary (edge
->callee
);
3495 struct cgraph_node
*to
= (edge
->caller
->global
.inlined_to
3496 ? edge
->caller
->global
.inlined_to
: edge
->caller
);
3497 struct inline_summary
*info
= inline_summary (to
);
3498 clause_t clause
= 0; /* not_inline is known to be false. */
3500 vec
<int> operand_map
= vNULL
;
3501 vec
<int> offset_map
= vNULL
;
3503 struct predicate toplev_predicate
;
3504 struct predicate true_p
= true_predicate ();
3505 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3508 toplev_predicate
= *es
->predicate
;
3510 toplev_predicate
= true_predicate ();
3512 if (ipa_node_params_vector
.exists () && callee_info
->conds
)
3514 struct ipa_edge_args
*args
= IPA_EDGE_REF (edge
);
3515 int count
= ipa_get_cs_argument_count (args
);
3518 evaluate_properties_for_edge (edge
, true, &clause
, NULL
, NULL
, NULL
);
3521 operand_map
.safe_grow_cleared (count
);
3522 offset_map
.safe_grow_cleared (count
);
3524 for (i
= 0; i
< count
; i
++)
3526 struct ipa_jump_func
*jfunc
= ipa_get_ith_jump_func (args
, i
);
3529 /* TODO: handle non-NOPs when merging. */
3530 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
3532 if (ipa_get_jf_pass_through_operation (jfunc
) == NOP_EXPR
)
3533 map
= ipa_get_jf_pass_through_formal_id (jfunc
);
3534 if (!ipa_get_jf_pass_through_agg_preserved (jfunc
))
3537 else if (jfunc
->type
== IPA_JF_ANCESTOR
)
3539 HOST_WIDE_INT offset
= ipa_get_jf_ancestor_offset (jfunc
);
3540 if (offset
>= 0 && offset
< INT_MAX
)
3542 map
= ipa_get_jf_ancestor_formal_id (jfunc
);
3543 if (!ipa_get_jf_ancestor_agg_preserved (jfunc
))
3545 offset_map
[i
] = offset
;
3548 operand_map
[i
] = map
;
3549 gcc_assert (map
< ipa_get_param_count (IPA_NODE_REF (to
)));
3552 for (i
= 0; vec_safe_iterate (callee_info
->entry
, i
, &e
); i
++)
3554 struct predicate p
= remap_predicate (info
, callee_info
,
3555 &e
->predicate
, operand_map
,
3558 if (!false_predicate_p (&p
))
3560 gcov_type add_time
= ((gcov_type
) e
->time
* edge
->frequency
3561 + CGRAPH_FREQ_BASE
/ 2) / CGRAPH_FREQ_BASE
;
3562 int prob
= predicate_probability (callee_info
->conds
,
3565 add_time
= apply_probability ((gcov_type
) add_time
, prob
);
3566 if (add_time
> MAX_TIME
* INLINE_TIME_SCALE
)
3567 add_time
= MAX_TIME
* INLINE_TIME_SCALE
;
3568 if (prob
!= REG_BR_PROB_BASE
3569 && dump_file
&& (dump_flags
& TDF_DETAILS
))
3571 fprintf (dump_file
, "\t\tScaling time by probability:%f\n",
3572 (double) prob
/ REG_BR_PROB_BASE
);
3574 account_size_time (info
, e
->size
, add_time
, &p
);
3577 remap_edge_summaries (edge
, edge
->callee
, info
, callee_info
, operand_map
,
3578 offset_map
, clause
, &toplev_predicate
);
3579 remap_hint_predicate (info
, callee_info
,
3580 &callee_info
->loop_iterations
,
3581 operand_map
, offset_map
, clause
, &toplev_predicate
);
3582 remap_hint_predicate (info
, callee_info
,
3583 &callee_info
->loop_stride
,
3584 operand_map
, offset_map
, clause
, &toplev_predicate
);
3585 remap_hint_predicate (info
, callee_info
,
3586 &callee_info
->array_index
,
3587 operand_map
, offset_map
, clause
, &toplev_predicate
);
3589 inline_update_callee_summaries (edge
->callee
,
3590 inline_edge_summary (edge
)->loop_depth
);
3592 /* We do not maintain predicates of inlined edges, free it. */
3593 edge_set_predicate (edge
, &true_p
);
3594 /* Similarly remove param summaries. */
3595 es
->param
.release ();
3596 operand_map
.release ();
3597 offset_map
.release ();
3600 /* For performance reasons inline_merge_summary is not updating overall size
3601 and time. Recompute it. */
3604 inline_update_overall_summary (struct cgraph_node
*node
)
3606 struct inline_summary
*info
= inline_summary (node
);
3612 for (i
= 0; vec_safe_iterate (info
->entry
, i
, &e
); i
++)
3614 info
->size
+= e
->size
, info
->time
+= e
->time
;
3615 if (info
->time
> MAX_TIME
* INLINE_TIME_SCALE
)
3616 info
->time
= MAX_TIME
* INLINE_TIME_SCALE
;
3618 estimate_calls_size_and_time (node
, &info
->size
, &info
->min_size
,
3620 ~(clause_t
) (1 << predicate_false_condition
),
3621 vNULL
, vNULL
, vNULL
);
3622 info
->time
= (info
->time
+ INLINE_TIME_SCALE
/ 2) / INLINE_TIME_SCALE
;
3623 info
->size
= (info
->size
+ INLINE_SIZE_SCALE
/ 2) / INLINE_SIZE_SCALE
;
3626 /* Return hints derrived from EDGE. */
3628 simple_edge_hints (struct cgraph_edge
*edge
)
3631 struct cgraph_node
*to
= (edge
->caller
->global
.inlined_to
3632 ? edge
->caller
->global
.inlined_to
: edge
->caller
);
3633 if (inline_summary (to
)->scc_no
3634 && inline_summary (to
)->scc_no
== inline_summary (edge
->callee
)->scc_no
3635 && !cgraph_edge_recursive_p (edge
))
3636 hints
|= INLINE_HINT_same_scc
;
3638 if (to
->lto_file_data
&& edge
->callee
->lto_file_data
3639 && to
->lto_file_data
!= edge
->callee
->lto_file_data
)
3640 hints
|= INLINE_HINT_cross_module
;
3645 /* Estimate the time cost for the caller when inlining EDGE.
3646 Only to be called via estimate_edge_time, that handles the
3649 When caching, also update the cache entry. Compute both time and
3650 size, since we always need both metrics eventually. */
3653 do_estimate_edge_time (struct cgraph_edge
*edge
)
3658 struct cgraph_node
*callee
;
3660 vec
<tree
> known_vals
;
3661 vec
<tree
> known_binfos
;
3662 vec
<ipa_agg_jump_function_p
> known_aggs
;
3663 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3666 callee
= cgraph_function_or_thunk_node (edge
->callee
, NULL
);
3668 gcc_checking_assert (edge
->inline_failed
);
3669 evaluate_properties_for_edge (edge
, true,
3670 &clause
, &known_vals
, &known_binfos
,
3672 estimate_node_size_and_time (callee
, clause
, known_vals
, known_binfos
,
3673 known_aggs
, &size
, &min_size
, &time
, &hints
, es
->param
);
3675 /* When we have profile feedback, we can quite safely identify hot
3676 edges and for those we disable size limits. Don't do that when
3677 probability that caller will call the callee is low however, since it
3678 may hurt optimization of the caller's hot path. */
3679 if (edge
->count
&& cgraph_maybe_hot_edge_p (edge
)
3681 > (edge
->caller
->global
.inlined_to
3682 ? edge
->caller
->global
.inlined_to
->count
: edge
->caller
->count
)))
3683 hints
|= INLINE_HINT_known_hot
;
3685 known_vals
.release ();
3686 known_binfos
.release ();
3687 known_aggs
.release ();
3688 gcc_checking_assert (size
>= 0);
3689 gcc_checking_assert (time
>= 0);
3691 /* When caching, update the cache entry. */
3692 if (edge_growth_cache
.exists ())
3694 inline_summary (edge
->callee
)->min_size
= min_size
;
3695 if ((int) edge_growth_cache
.length () <= edge
->uid
)
3696 edge_growth_cache
.safe_grow_cleared (cgraph_edge_max_uid
);
3697 edge_growth_cache
[edge
->uid
].time
= time
+ (time
>= 0);
3699 edge_growth_cache
[edge
->uid
].size
= size
+ (size
>= 0);
3700 hints
|= simple_edge_hints (edge
);
3701 edge_growth_cache
[edge
->uid
].hints
= hints
+ 1;
3707 /* Return estimated callee growth after inlining EDGE.
3708 Only to be called via estimate_edge_size. */
3711 do_estimate_edge_size (struct cgraph_edge
*edge
)
3714 struct cgraph_node
*callee
;
3716 vec
<tree
> known_vals
;
3717 vec
<tree
> known_binfos
;
3718 vec
<ipa_agg_jump_function_p
> known_aggs
;
3720 /* When we do caching, use do_estimate_edge_time to populate the entry. */
3722 if (edge_growth_cache
.exists ())
3724 do_estimate_edge_time (edge
);
3725 size
= edge_growth_cache
[edge
->uid
].size
;
3726 gcc_checking_assert (size
);
3727 return size
- (size
> 0);
3730 callee
= cgraph_function_or_thunk_node (edge
->callee
, NULL
);
3732 /* Early inliner runs without caching, go ahead and do the dirty work. */
3733 gcc_checking_assert (edge
->inline_failed
);
3734 evaluate_properties_for_edge (edge
, true,
3735 &clause
, &known_vals
, &known_binfos
,
3737 estimate_node_size_and_time (callee
, clause
, known_vals
, known_binfos
,
3738 known_aggs
, &size
, NULL
, NULL
, NULL
, vNULL
);
3739 known_vals
.release ();
3740 known_binfos
.release ();
3741 known_aggs
.release ();
3746 /* Estimate the growth of the caller when inlining EDGE.
3747 Only to be called via estimate_edge_size. */
3750 do_estimate_edge_hints (struct cgraph_edge
*edge
)
3753 struct cgraph_node
*callee
;
3755 vec
<tree
> known_vals
;
3756 vec
<tree
> known_binfos
;
3757 vec
<ipa_agg_jump_function_p
> known_aggs
;
3759 /* When we do caching, use do_estimate_edge_time to populate the entry. */
3761 if (edge_growth_cache
.exists ())
3763 do_estimate_edge_time (edge
);
3764 hints
= edge_growth_cache
[edge
->uid
].hints
;
3765 gcc_checking_assert (hints
);
3769 callee
= cgraph_function_or_thunk_node (edge
->callee
, NULL
);
3771 /* Early inliner runs without caching, go ahead and do the dirty work. */
3772 gcc_checking_assert (edge
->inline_failed
);
3773 evaluate_properties_for_edge (edge
, true,
3774 &clause
, &known_vals
, &known_binfos
,
3776 estimate_node_size_and_time (callee
, clause
, known_vals
, known_binfos
,
3777 known_aggs
, NULL
, NULL
, NULL
, &hints
, vNULL
);
3778 known_vals
.release ();
3779 known_binfos
.release ();
3780 known_aggs
.release ();
3781 hints
|= simple_edge_hints (edge
);
3786 /* Estimate self time of the function NODE after inlining EDGE. */
3789 estimate_time_after_inlining (struct cgraph_node
*node
,
3790 struct cgraph_edge
*edge
)
3792 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3793 if (!es
->predicate
|| !false_predicate_p (es
->predicate
))
3796 inline_summary (node
)->time
+ estimate_edge_time (edge
);
3799 if (time
> MAX_TIME
)
3803 return inline_summary (node
)->time
;
3807 /* Estimate the size of NODE after inlining EDGE which should be an
3808 edge to either NODE or a call inlined into NODE. */
3811 estimate_size_after_inlining (struct cgraph_node
*node
,
3812 struct cgraph_edge
*edge
)
3814 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3815 if (!es
->predicate
|| !false_predicate_p (es
->predicate
))
3817 int size
= inline_summary (node
)->size
+ estimate_edge_growth (edge
);
3818 gcc_assert (size
>= 0);
3821 return inline_summary (node
)->size
;
3827 struct cgraph_node
*node
;
3828 bool self_recursive
;
3833 /* Worker for do_estimate_growth. Collect growth for all callers. */
3836 do_estimate_growth_1 (struct cgraph_node
*node
, void *data
)
3838 struct cgraph_edge
*e
;
3839 struct growth_data
*d
= (struct growth_data
*) data
;
3841 for (e
= node
->callers
; e
; e
= e
->next_caller
)
3843 gcc_checking_assert (e
->inline_failed
);
3845 if (e
->caller
== d
->node
3846 || (e
->caller
->global
.inlined_to
3847 && e
->caller
->global
.inlined_to
== d
->node
))
3848 d
->self_recursive
= true;
3849 d
->growth
+= estimate_edge_growth (e
);
3855 /* Estimate the growth caused by inlining NODE into all callees. */
3858 do_estimate_growth (struct cgraph_node
*node
)
3860 struct growth_data d
= { node
, 0, false };
3861 struct inline_summary
*info
= inline_summary (node
);
3863 cgraph_for_node_and_aliases (node
, do_estimate_growth_1
, &d
, true);
3865 /* For self recursive functions the growth estimation really should be
3866 infinity. We don't want to return very large values because the growth
3867 plays various roles in badness computation fractions. Be sure to not
3868 return zero or negative growths. */
3869 if (d
.self_recursive
)
3870 d
.growth
= d
.growth
< info
->size
? info
->size
: d
.growth
;
3871 else if (DECL_EXTERNAL (node
->decl
))
3875 if (cgraph_will_be_removed_from_program_if_no_direct_calls (node
))
3876 d
.growth
-= info
->size
;
3877 /* COMDAT functions are very often not shared across multiple units
3878 since they come from various template instantiations.
3879 Take this into account. */
3880 else if (node
->externally_visible
&& node
->get_comdat_group ()
3881 && cgraph_can_remove_if_no_direct_calls_p (node
))
3882 d
.growth
-= (info
->size
3883 * (100 - PARAM_VALUE (PARAM_COMDAT_SHARING_PROBABILITY
))
3887 if (node_growth_cache
.exists ())
3889 if ((int) node_growth_cache
.length () <= node
->uid
)
3890 node_growth_cache
.safe_grow_cleared (cgraph_max_uid
);
3891 node_growth_cache
[node
->uid
] = d
.growth
+ (d
.growth
>= 0);
3897 /* Make cheap estimation if growth of NODE is likely positive knowing
3898 EDGE_GROWTH of one particular edge.
3899 We assume that most of other edges will have similar growth
3900 and skip computation if there are too many callers. */
3903 growth_likely_positive (struct cgraph_node
*node
, int edge_growth ATTRIBUTE_UNUSED
)
3907 struct cgraph_edge
*e
;
3908 gcc_checking_assert (edge_growth
> 0);
3910 /* Unlike for functions called once, we play unsafe with
3911 COMDATs. We can allow that since we know functions
3912 in consideration are small (and thus risk is small) and
3913 moreover grow estimates already accounts that COMDAT
3914 functions may or may not disappear when eliminated from
3915 current unit. With good probability making aggressive
3916 choice in all units is going to make overall program
3919 Consequently we ask cgraph_can_remove_if_no_direct_calls_p
3921 cgraph_will_be_removed_from_program_if_no_direct_calls */
3922 if (DECL_EXTERNAL (node
->decl
)
3923 || !cgraph_can_remove_if_no_direct_calls_p (node
))
3926 /* If there is cached value, just go ahead. */
3927 if ((int)node_growth_cache
.length () > node
->uid
3928 && (ret
= node_growth_cache
[node
->uid
]))
3930 if (!cgraph_will_be_removed_from_program_if_no_direct_calls (node
)
3931 && (!node
->externally_visible
|| !node
->get_comdat_group ()
3932 || !cgraph_can_remove_if_no_direct_calls_p (node
)))
3934 max_callers
= inline_summary (node
)->size
* 4 / edge_growth
+ 2;
3936 for (e
= node
->callers
; e
; e
= e
->next_caller
)
3942 return estimate_growth (node
) > 0;
3946 /* This function performs intraprocedural analysis in NODE that is required to
3947 inline indirect calls. */
3950 inline_indirect_intraprocedural_analysis (struct cgraph_node
*node
)
3952 ipa_analyze_node (node
);
3953 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3955 ipa_print_node_params (dump_file
, node
);
3956 ipa_print_node_jump_functions (dump_file
, node
);
3961 /* Note function body size. */
3964 inline_analyze_function (struct cgraph_node
*node
)
3966 push_cfun (DECL_STRUCT_FUNCTION (node
->decl
));
3969 fprintf (dump_file
, "\nAnalyzing function: %s/%u\n",
3970 node
->name (), node
->order
);
3971 if (optimize
&& !node
->thunk
.thunk_p
)
3972 inline_indirect_intraprocedural_analysis (node
);
3973 compute_inline_parameters (node
, false);
3976 struct cgraph_edge
*e
;
3977 for (e
= node
->callees
; e
; e
= e
->next_callee
)
3979 if (e
->inline_failed
== CIF_FUNCTION_NOT_CONSIDERED
)
3980 e
->inline_failed
= CIF_FUNCTION_NOT_OPTIMIZED
;
3981 e
->call_stmt_cannot_inline_p
= true;
3983 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
3985 if (e
->inline_failed
== CIF_FUNCTION_NOT_CONSIDERED
)
3986 e
->inline_failed
= CIF_FUNCTION_NOT_OPTIMIZED
;
3987 e
->call_stmt_cannot_inline_p
= true;
3995 /* Called when new function is inserted to callgraph late. */
3998 add_new_function (struct cgraph_node
*node
, void *data ATTRIBUTE_UNUSED
)
4000 inline_analyze_function (node
);
4004 /* Note function body size. */
4007 inline_generate_summary (void)
4009 struct cgraph_node
*node
;
4011 /* When not optimizing, do not bother to analyze. Inlining is still done
4012 because edge redirection needs to happen there. */
4013 if (!optimize
&& !flag_lto
&& !flag_wpa
)
4016 function_insertion_hook_holder
=
4017 cgraph_add_function_insertion_hook (&add_new_function
, NULL
);
4019 ipa_register_cgraph_hooks ();
4020 inline_free_summary ();
4022 FOR_EACH_DEFINED_FUNCTION (node
)
4024 inline_analyze_function (node
);
4028 /* Read predicate from IB. */
4030 static struct predicate
4031 read_predicate (struct lto_input_block
*ib
)
4033 struct predicate out
;
4039 gcc_assert (k
<= MAX_CLAUSES
);
4040 clause
= out
.clause
[k
++] = streamer_read_uhwi (ib
);
4044 /* Zero-initialize the remaining clauses in OUT. */
4045 while (k
<= MAX_CLAUSES
)
4046 out
.clause
[k
++] = 0;
4052 /* Write inline summary for edge E to OB. */
4055 read_inline_edge_summary (struct lto_input_block
*ib
, struct cgraph_edge
*e
)
4057 struct inline_edge_summary
*es
= inline_edge_summary (e
);
4061 es
->call_stmt_size
= streamer_read_uhwi (ib
);
4062 es
->call_stmt_time
= streamer_read_uhwi (ib
);
4063 es
->loop_depth
= streamer_read_uhwi (ib
);
4064 p
= read_predicate (ib
);
4065 edge_set_predicate (e
, &p
);
4066 length
= streamer_read_uhwi (ib
);
4069 es
->param
.safe_grow_cleared (length
);
4070 for (i
= 0; i
< length
; i
++)
4071 es
->param
[i
].change_prob
= streamer_read_uhwi (ib
);
4076 /* Stream in inline summaries from the section. */
4079 inline_read_section (struct lto_file_decl_data
*file_data
, const char *data
,
4082 const struct lto_function_header
*header
=
4083 (const struct lto_function_header
*) data
;
4084 const int cfg_offset
= sizeof (struct lto_function_header
);
4085 const int main_offset
= cfg_offset
+ header
->cfg_size
;
4086 const int string_offset
= main_offset
+ header
->main_size
;
4087 struct data_in
*data_in
;
4088 struct lto_input_block ib
;
4089 unsigned int i
, count2
, j
;
4090 unsigned int f_count
;
4092 LTO_INIT_INPUT_BLOCK (ib
, (const char *) data
+ main_offset
, 0,
4096 lto_data_in_create (file_data
, (const char *) data
+ string_offset
,
4097 header
->string_size
, vNULL
);
4098 f_count
= streamer_read_uhwi (&ib
);
4099 for (i
= 0; i
< f_count
; i
++)
4102 struct cgraph_node
*node
;
4103 struct inline_summary
*info
;
4104 lto_symtab_encoder_t encoder
;
4105 struct bitpack_d bp
;
4106 struct cgraph_edge
*e
;
4109 index
= streamer_read_uhwi (&ib
);
4110 encoder
= file_data
->symtab_node_encoder
;
4111 node
= cgraph (lto_symtab_encoder_deref (encoder
, index
));
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;