1 /* Inlining decision heuristics.
2 Copyright (C) 2003-2015 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"
74 #include "double-int.h"
82 #include "fold-const.h"
83 #include "stor-layout.h"
84 #include "stringpool.h"
85 #include "print-tree.h"
86 #include "tree-inline.h"
87 #include "langhooks.h"
89 #include "diagnostic.h"
90 #include "gimple-pretty-print.h"
92 #include "tree-pass.h"
95 #include "hard-reg-set.h"
98 #include "dominance.h"
101 #include "basic-block.h"
102 #include "tree-ssa-alias.h"
103 #include "internal-fn.h"
104 #include "gimple-expr.h"
107 #include "gimple-iterator.h"
108 #include "gimple-ssa.h"
109 #include "tree-cfg.h"
110 #include "tree-phinodes.h"
111 #include "ssa-iterators.h"
112 #include "tree-ssanames.h"
113 #include "tree-ssa-loop-niter.h"
114 #include "tree-ssa-loop.h"
115 #include "hash-map.h"
116 #include "plugin-api.h"
119 #include "alloc-pool.h"
120 #include "symbol-summary.h"
121 #include "ipa-prop.h"
122 #include "lto-streamer.h"
123 #include "data-streamer.h"
124 #include "tree-streamer.h"
125 #include "ipa-inline.h"
127 #include "tree-scalar-evolution.h"
128 #include "ipa-utils.h"
130 #include "cfgexpand.h"
132 /* Estimate runtime of function can easilly run into huge numbers with many
133 nested loops. Be sure we can compute time * INLINE_SIZE_SCALE * 2 in an
134 integer. For anything larger we use gcov_type. */
135 #define MAX_TIME 500000
137 /* Number of bits in integer, but we really want to be stable across different
139 #define NUM_CONDITIONS 32
141 enum predicate_conditions
143 predicate_false_condition
= 0,
144 predicate_not_inlined_condition
= 1,
145 predicate_first_dynamic_condition
= 2
148 /* Special condition code we use to represent test that operand is compile time
150 #define IS_NOT_CONSTANT ERROR_MARK
151 /* Special condition code we use to represent test that operand is not changed
152 across invocation of the function. When operand IS_NOT_CONSTANT it is always
153 CHANGED, however i.e. loop invariants can be NOT_CHANGED given percentage
154 of executions even when they are not compile time constants. */
155 #define CHANGED IDENTIFIER_NODE
157 /* Holders of ipa cgraph hooks: */
158 static struct cgraph_2edge_hook_list
*edge_duplication_hook_holder
;
159 static struct cgraph_edge_hook_list
*edge_removal_hook_holder
;
160 static void inline_edge_removal_hook (struct cgraph_edge
*, void *);
161 static void inline_edge_duplication_hook (struct cgraph_edge
*,
162 struct cgraph_edge
*, void *);
164 /* VECtor holding inline summaries.
165 In GGC memory because conditions might point to constant trees. */
166 function_summary
<inline_summary
*> *inline_summaries
;
167 vec
<inline_edge_summary_t
> inline_edge_summary_vec
;
169 /* Cached node/edge growths. */
170 vec
<edge_growth_cache_entry
> edge_growth_cache
;
172 /* Edge predicates goes here. */
173 static alloc_pool edge_predicate_pool
;
175 /* Return true predicate (tautology).
176 We represent it by empty list of clauses. */
178 static inline struct predicate
179 true_predicate (void)
187 /* Return predicate testing single condition number COND. */
189 static inline struct predicate
190 single_cond_predicate (int cond
)
193 p
.clause
[0] = 1 << cond
;
199 /* Return false predicate. First clause require false condition. */
201 static inline struct predicate
202 false_predicate (void)
204 return single_cond_predicate (predicate_false_condition
);
208 /* Return true if P is (true). */
211 true_predicate_p (struct predicate
*p
)
213 return !p
->clause
[0];
217 /* Return true if P is (false). */
220 false_predicate_p (struct predicate
*p
)
222 if (p
->clause
[0] == (1 << predicate_false_condition
))
224 gcc_checking_assert (!p
->clause
[1]
225 && p
->clause
[0] == 1 << predicate_false_condition
);
232 /* Return predicate that is set true when function is not inlined. */
234 static inline struct predicate
235 not_inlined_predicate (void)
237 return single_cond_predicate (predicate_not_inlined_condition
);
240 /* Simple description of whether a memory load or a condition refers to a load
241 from an aggregate and if so, how and where from in the aggregate.
242 Individual fields have the same meaning like fields with the same name in
245 struct agg_position_info
247 HOST_WIDE_INT offset
;
252 /* Add condition to condition list CONDS. AGGPOS describes whether the used
253 oprand is loaded from an aggregate and where in the aggregate it is. It can
254 be NULL, which means this not a load from an aggregate. */
256 static struct predicate
257 add_condition (struct inline_summary
*summary
, int operand_num
,
258 struct agg_position_info
*aggpos
,
259 enum tree_code code
, tree val
)
263 struct condition new_cond
;
264 HOST_WIDE_INT offset
;
265 bool agg_contents
, by_ref
;
269 offset
= aggpos
->offset
;
270 agg_contents
= aggpos
->agg_contents
;
271 by_ref
= aggpos
->by_ref
;
276 agg_contents
= false;
280 gcc_checking_assert (operand_num
>= 0);
281 for (i
= 0; vec_safe_iterate (summary
->conds
, i
, &c
); i
++)
283 if (c
->operand_num
== operand_num
286 && c
->agg_contents
== agg_contents
287 && (!agg_contents
|| (c
->offset
== offset
&& c
->by_ref
== by_ref
)))
288 return single_cond_predicate (i
+ predicate_first_dynamic_condition
);
290 /* Too many conditions. Give up and return constant true. */
291 if (i
== NUM_CONDITIONS
- predicate_first_dynamic_condition
)
292 return true_predicate ();
294 new_cond
.operand_num
= operand_num
;
295 new_cond
.code
= code
;
297 new_cond
.agg_contents
= agg_contents
;
298 new_cond
.by_ref
= by_ref
;
299 new_cond
.offset
= offset
;
300 vec_safe_push (summary
->conds
, new_cond
);
301 return single_cond_predicate (i
+ predicate_first_dynamic_condition
);
305 /* Add clause CLAUSE into the predicate P. */
308 add_clause (conditions conditions
, struct predicate
*p
, clause_t clause
)
312 int insert_here
= -1;
319 /* False clause makes the whole predicate false. Kill the other variants. */
320 if (clause
== (1 << predicate_false_condition
))
322 p
->clause
[0] = (1 << predicate_false_condition
);
326 if (false_predicate_p (p
))
329 /* No one should be silly enough to add false into nontrivial clauses. */
330 gcc_checking_assert (!(clause
& (1 << predicate_false_condition
)));
332 /* Look where to insert the clause. At the same time prune out
333 clauses of P that are implied by the new clause and thus
335 for (i
= 0, i2
= 0; i
<= MAX_CLAUSES
; i
++)
337 p
->clause
[i2
] = p
->clause
[i
];
342 /* If p->clause[i] implies clause, there is nothing to add. */
343 if ((p
->clause
[i
] & clause
) == p
->clause
[i
])
345 /* We had nothing to add, none of clauses should've become
347 gcc_checking_assert (i
== i2
);
351 if (p
->clause
[i
] < clause
&& insert_here
< 0)
354 /* If clause implies p->clause[i], then p->clause[i] becomes redundant.
355 Otherwise the p->clause[i] has to stay. */
356 if ((p
->clause
[i
] & clause
) != clause
)
360 /* Look for clauses that are obviously true. I.e.
361 op0 == 5 || op0 != 5. */
362 for (c1
= predicate_first_dynamic_condition
; c1
< NUM_CONDITIONS
; c1
++)
365 if (!(clause
& (1 << c1
)))
367 cc1
= &(*conditions
)[c1
- predicate_first_dynamic_condition
];
368 /* We have no way to represent !CHANGED and !IS_NOT_CONSTANT
369 and thus there is no point for looking for them. */
370 if (cc1
->code
== CHANGED
|| cc1
->code
== IS_NOT_CONSTANT
)
372 for (c2
= c1
+ 1; c2
< NUM_CONDITIONS
; c2
++)
373 if (clause
& (1 << c2
))
376 &(*conditions
)[c1
- predicate_first_dynamic_condition
];
378 &(*conditions
)[c2
- predicate_first_dynamic_condition
];
379 if (cc1
->operand_num
== cc2
->operand_num
380 && cc1
->val
== cc2
->val
381 && cc2
->code
!= IS_NOT_CONSTANT
382 && cc2
->code
!= CHANGED
383 && cc1
->code
== invert_tree_comparison (cc2
->code
,
384 HONOR_NANS (cc1
->val
)))
390 /* We run out of variants. Be conservative in positive direction. */
391 if (i2
== MAX_CLAUSES
)
393 /* Keep clauses in decreasing order. This makes equivalence testing easy. */
394 p
->clause
[i2
+ 1] = 0;
395 if (insert_here
>= 0)
396 for (; i2
> insert_here
; i2
--)
397 p
->clause
[i2
] = p
->clause
[i2
- 1];
400 p
->clause
[insert_here
] = clause
;
406 static struct predicate
407 and_predicates (conditions conditions
,
408 struct predicate
*p
, struct predicate
*p2
)
410 struct predicate out
= *p
;
413 /* Avoid busy work. */
414 if (false_predicate_p (p2
) || true_predicate_p (p
))
416 if (false_predicate_p (p
) || true_predicate_p (p2
))
419 /* See how far predicates match. */
420 for (i
= 0; p
->clause
[i
] && p
->clause
[i
] == p2
->clause
[i
]; i
++)
422 gcc_checking_assert (i
< MAX_CLAUSES
);
425 /* Combine the predicates rest. */
426 for (; p2
->clause
[i
]; i
++)
428 gcc_checking_assert (i
< MAX_CLAUSES
);
429 add_clause (conditions
, &out
, p2
->clause
[i
]);
435 /* Return true if predicates are obviously equal. */
438 predicates_equal_p (struct predicate
*p
, struct predicate
*p2
)
441 for (i
= 0; p
->clause
[i
]; i
++)
443 gcc_checking_assert (i
< MAX_CLAUSES
);
444 gcc_checking_assert (p
->clause
[i
] > p
->clause
[i
+ 1]);
445 gcc_checking_assert (!p2
->clause
[i
]
446 || p2
->clause
[i
] > p2
->clause
[i
+ 1]);
447 if (p
->clause
[i
] != p2
->clause
[i
])
450 return !p2
->clause
[i
];
456 static struct predicate
457 or_predicates (conditions conditions
,
458 struct predicate
*p
, struct predicate
*p2
)
460 struct predicate out
= true_predicate ();
463 /* Avoid busy work. */
464 if (false_predicate_p (p2
) || true_predicate_p (p
))
466 if (false_predicate_p (p
) || true_predicate_p (p2
))
468 if (predicates_equal_p (p
, p2
))
471 /* OK, combine the predicates. */
472 for (i
= 0; p
->clause
[i
]; i
++)
473 for (j
= 0; p2
->clause
[j
]; j
++)
475 gcc_checking_assert (i
< MAX_CLAUSES
&& j
< MAX_CLAUSES
);
476 add_clause (conditions
, &out
, p
->clause
[i
] | p2
->clause
[j
]);
482 /* Having partial truth assignment in POSSIBLE_TRUTHS, return false
483 if predicate P is known to be false. */
486 evaluate_predicate (struct predicate
*p
, clause_t possible_truths
)
490 /* True remains true. */
491 if (true_predicate_p (p
))
494 gcc_assert (!(possible_truths
& (1 << predicate_false_condition
)));
496 /* See if we can find clause we can disprove. */
497 for (i
= 0; p
->clause
[i
]; i
++)
499 gcc_checking_assert (i
< MAX_CLAUSES
);
500 if (!(p
->clause
[i
] & possible_truths
))
506 /* Return the probability in range 0...REG_BR_PROB_BASE that the predicated
507 instruction will be recomputed per invocation of the inlined call. */
510 predicate_probability (conditions conds
,
511 struct predicate
*p
, clause_t possible_truths
,
512 vec
<inline_param_summary
> inline_param_summary
)
515 int combined_prob
= REG_BR_PROB_BASE
;
517 /* True remains true. */
518 if (true_predicate_p (p
))
519 return REG_BR_PROB_BASE
;
521 if (false_predicate_p (p
))
524 gcc_assert (!(possible_truths
& (1 << predicate_false_condition
)));
526 /* See if we can find clause we can disprove. */
527 for (i
= 0; p
->clause
[i
]; i
++)
529 gcc_checking_assert (i
< MAX_CLAUSES
);
530 if (!(p
->clause
[i
] & possible_truths
))
536 if (!inline_param_summary
.exists ())
537 return REG_BR_PROB_BASE
;
538 for (i2
= 0; i2
< NUM_CONDITIONS
; i2
++)
539 if ((p
->clause
[i
] & possible_truths
) & (1 << i2
))
541 if (i2
>= predicate_first_dynamic_condition
)
544 &(*conds
)[i2
- predicate_first_dynamic_condition
];
545 if (c
->code
== CHANGED
547 (int) inline_param_summary
.length ()))
550 inline_param_summary
[c
->operand_num
].change_prob
;
551 this_prob
= MAX (this_prob
, iprob
);
554 this_prob
= REG_BR_PROB_BASE
;
557 this_prob
= REG_BR_PROB_BASE
;
559 combined_prob
= MIN (this_prob
, combined_prob
);
564 return combined_prob
;
568 /* Dump conditional COND. */
571 dump_condition (FILE *f
, conditions conditions
, int cond
)
574 if (cond
== predicate_false_condition
)
575 fprintf (f
, "false");
576 else if (cond
== predicate_not_inlined_condition
)
577 fprintf (f
, "not inlined");
580 c
= &(*conditions
)[cond
- predicate_first_dynamic_condition
];
581 fprintf (f
, "op%i", c
->operand_num
);
583 fprintf (f
, "[%soffset: " HOST_WIDE_INT_PRINT_DEC
"]",
584 c
->by_ref
? "ref " : "", c
->offset
);
585 if (c
->code
== IS_NOT_CONSTANT
)
587 fprintf (f
, " not constant");
590 if (c
->code
== CHANGED
)
592 fprintf (f
, " changed");
595 fprintf (f
, " %s ", op_symbol_code (c
->code
));
596 print_generic_expr (f
, c
->val
, 1);
601 /* Dump clause CLAUSE. */
604 dump_clause (FILE *f
, conditions conds
, clause_t clause
)
611 for (i
= 0; i
< NUM_CONDITIONS
; i
++)
612 if (clause
& (1 << i
))
617 dump_condition (f
, conds
, i
);
623 /* Dump predicate PREDICATE. */
626 dump_predicate (FILE *f
, conditions conds
, struct predicate
*pred
)
629 if (true_predicate_p (pred
))
630 dump_clause (f
, conds
, 0);
632 for (i
= 0; pred
->clause
[i
]; i
++)
636 dump_clause (f
, conds
, pred
->clause
[i
]);
642 /* Dump inline hints. */
644 dump_inline_hints (FILE *f
, inline_hints hints
)
648 fprintf (f
, "inline hints:");
649 if (hints
& INLINE_HINT_indirect_call
)
651 hints
&= ~INLINE_HINT_indirect_call
;
652 fprintf (f
, " indirect_call");
654 if (hints
& INLINE_HINT_loop_iterations
)
656 hints
&= ~INLINE_HINT_loop_iterations
;
657 fprintf (f
, " loop_iterations");
659 if (hints
& INLINE_HINT_loop_stride
)
661 hints
&= ~INLINE_HINT_loop_stride
;
662 fprintf (f
, " loop_stride");
664 if (hints
& INLINE_HINT_same_scc
)
666 hints
&= ~INLINE_HINT_same_scc
;
667 fprintf (f
, " same_scc");
669 if (hints
& INLINE_HINT_in_scc
)
671 hints
&= ~INLINE_HINT_in_scc
;
672 fprintf (f
, " in_scc");
674 if (hints
& INLINE_HINT_cross_module
)
676 hints
&= ~INLINE_HINT_cross_module
;
677 fprintf (f
, " cross_module");
679 if (hints
& INLINE_HINT_declared_inline
)
681 hints
&= ~INLINE_HINT_declared_inline
;
682 fprintf (f
, " declared_inline");
684 if (hints
& INLINE_HINT_array_index
)
686 hints
&= ~INLINE_HINT_array_index
;
687 fprintf (f
, " array_index");
689 if (hints
& INLINE_HINT_known_hot
)
691 hints
&= ~INLINE_HINT_known_hot
;
692 fprintf (f
, " known_hot");
698 /* Record SIZE and TIME under condition PRED into the inline summary. */
701 account_size_time (struct inline_summary
*summary
, int size
, int time
,
702 struct predicate
*pred
)
708 if (false_predicate_p (pred
))
711 /* We need to create initial empty unconitional clause, but otherwie
712 we don't need to account empty times and sizes. */
713 if (!size
&& !time
&& summary
->entry
)
716 /* Watch overflow that might result from insane profiles. */
717 if (time
> MAX_TIME
* INLINE_TIME_SCALE
)
718 time
= MAX_TIME
* INLINE_TIME_SCALE
;
719 gcc_assert (time
>= 0);
721 for (i
= 0; vec_safe_iterate (summary
->entry
, i
, &e
); i
++)
722 if (predicates_equal_p (&e
->predicate
, pred
))
731 e
= &(*summary
->entry
)[0];
732 gcc_assert (!e
->predicate
.clause
[0]);
733 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
735 "\t\tReached limit on number of entries, "
736 "ignoring the predicate.");
738 if (dump_file
&& (dump_flags
& TDF_DETAILS
) && (time
|| size
))
741 "\t\tAccounting size:%3.2f, time:%3.2f on %spredicate:",
742 ((double) size
) / INLINE_SIZE_SCALE
,
743 ((double) time
) / INLINE_TIME_SCALE
, found
? "" : "new ");
744 dump_predicate (dump_file
, summary
->conds
, pred
);
748 struct size_time_entry new_entry
;
749 new_entry
.size
= size
;
750 new_entry
.time
= time
;
751 new_entry
.predicate
= *pred
;
752 vec_safe_push (summary
->entry
, new_entry
);
758 if (e
->time
> MAX_TIME
* INLINE_TIME_SCALE
)
759 e
->time
= MAX_TIME
* INLINE_TIME_SCALE
;
763 /* Set predicate for edge E. */
766 edge_set_predicate (struct cgraph_edge
*e
, struct predicate
*predicate
)
768 struct inline_edge_summary
*es
= inline_edge_summary (e
);
770 /* If the edge is determined to be never executed, redirect it
771 to BUILTIN_UNREACHABLE to save inliner from inlining into it. */
772 if (predicate
&& false_predicate_p (predicate
) && e
->callee
)
774 struct cgraph_node
*callee
= !e
->inline_failed
? e
->callee
: NULL
;
776 e
->redirect_callee (cgraph_node::get_create
777 (builtin_decl_implicit (BUILT_IN_UNREACHABLE
)));
778 e
->inline_failed
= CIF_UNREACHABLE
;
779 es
->call_stmt_size
= 0;
780 es
->call_stmt_time
= 0;
782 callee
->remove_symbol_and_inline_clones ();
784 if (predicate
&& !true_predicate_p (predicate
))
787 es
->predicate
= (struct predicate
*) pool_alloc (edge_predicate_pool
);
788 *es
->predicate
= *predicate
;
793 pool_free (edge_predicate_pool
, es
->predicate
);
794 es
->predicate
= NULL
;
798 /* Set predicate for hint *P. */
801 set_hint_predicate (struct predicate
**p
, struct predicate new_predicate
)
803 if (false_predicate_p (&new_predicate
) || true_predicate_p (&new_predicate
))
806 pool_free (edge_predicate_pool
, *p
);
812 *p
= (struct predicate
*) pool_alloc (edge_predicate_pool
);
818 /* KNOWN_VALS is partial mapping of parameters of NODE to constant values.
819 KNOWN_AGGS is a vector of aggreggate jump functions for each parameter.
820 Return clause of possible truths. When INLINE_P is true, assume that we are
823 ERROR_MARK means compile time invariant. */
826 evaluate_conditions_for_known_args (struct cgraph_node
*node
,
828 vec
<tree
> known_vals
,
829 vec
<ipa_agg_jump_function_p
>
832 clause_t clause
= inline_p
? 0 : 1 << predicate_not_inlined_condition
;
833 struct inline_summary
*info
= inline_summaries
->get (node
);
837 for (i
= 0; vec_safe_iterate (info
->conds
, i
, &c
); i
++)
842 /* We allow call stmt to have fewer arguments than the callee function
843 (especially for K&R style programs). So bound check here (we assume
844 known_aggs vector, if non-NULL, has the same length as
846 gcc_checking_assert (!known_aggs
.exists ()
847 || (known_vals
.length () == known_aggs
.length ()));
848 if (c
->operand_num
>= (int) known_vals
.length ())
850 clause
|= 1 << (i
+ predicate_first_dynamic_condition
);
856 struct ipa_agg_jump_function
*agg
;
858 if (c
->code
== CHANGED
860 && (known_vals
[c
->operand_num
] == error_mark_node
))
863 if (known_aggs
.exists ())
865 agg
= known_aggs
[c
->operand_num
];
866 val
= ipa_find_agg_cst_for_param (agg
, c
->offset
, c
->by_ref
);
873 val
= known_vals
[c
->operand_num
];
874 if (val
== error_mark_node
&& c
->code
!= CHANGED
)
880 clause
|= 1 << (i
+ predicate_first_dynamic_condition
);
883 if (c
->code
== IS_NOT_CONSTANT
|| c
->code
== CHANGED
)
886 if (operand_equal_p (TYPE_SIZE (TREE_TYPE (c
->val
)),
887 TYPE_SIZE (TREE_TYPE (val
)), 0))
889 val
= fold_unary (VIEW_CONVERT_EXPR
, TREE_TYPE (c
->val
), val
);
892 ? fold_binary_to_constant (c
->code
, boolean_type_node
, val
, c
->val
)
895 if (res
&& integer_zerop (res
))
898 clause
|= 1 << (i
+ predicate_first_dynamic_condition
);
904 /* Work out what conditions might be true at invocation of E. */
907 evaluate_properties_for_edge (struct cgraph_edge
*e
, bool inline_p
,
908 clause_t
*clause_ptr
,
909 vec
<tree
> *known_vals_ptr
,
910 vec
<ipa_polymorphic_call_context
>
912 vec
<ipa_agg_jump_function_p
> *known_aggs_ptr
)
914 struct cgraph_node
*callee
= e
->callee
->ultimate_alias_target ();
915 struct inline_summary
*info
= inline_summaries
->get (callee
);
916 vec
<tree
> known_vals
= vNULL
;
917 vec
<ipa_agg_jump_function_p
> known_aggs
= vNULL
;
920 *clause_ptr
= inline_p
? 0 : 1 << predicate_not_inlined_condition
;
922 known_vals_ptr
->create (0);
923 if (known_contexts_ptr
)
924 known_contexts_ptr
->create (0);
926 if (ipa_node_params_sum
927 && !e
->call_stmt_cannot_inline_p
928 && ((clause_ptr
&& info
->conds
) || known_vals_ptr
|| known_contexts_ptr
))
930 struct ipa_node_params
*parms_info
;
931 struct ipa_edge_args
*args
= IPA_EDGE_REF (e
);
932 struct inline_edge_summary
*es
= inline_edge_summary (e
);
933 int i
, count
= ipa_get_cs_argument_count (args
);
935 if (e
->caller
->global
.inlined_to
)
936 parms_info
= IPA_NODE_REF (e
->caller
->global
.inlined_to
);
938 parms_info
= IPA_NODE_REF (e
->caller
);
940 if (count
&& (info
->conds
|| known_vals_ptr
))
941 known_vals
.safe_grow_cleared (count
);
942 if (count
&& (info
->conds
|| known_aggs_ptr
))
943 known_aggs
.safe_grow_cleared (count
);
944 if (count
&& known_contexts_ptr
)
945 known_contexts_ptr
->safe_grow_cleared (count
);
947 for (i
= 0; i
< count
; i
++)
949 struct ipa_jump_func
*jf
= ipa_get_ith_jump_func (args
, i
);
950 tree cst
= ipa_value_from_jfunc (parms_info
, jf
);
952 if (!cst
&& e
->call_stmt
953 && i
< (int)gimple_call_num_args (e
->call_stmt
))
955 cst
= gimple_call_arg (e
->call_stmt
, i
);
956 if (!is_gimple_min_invariant (cst
))
961 gcc_checking_assert (TREE_CODE (cst
) != TREE_BINFO
);
962 if (known_vals
.exists ())
965 else if (inline_p
&& !es
->param
[i
].change_prob
)
966 known_vals
[i
] = error_mark_node
;
968 if (known_contexts_ptr
)
969 (*known_contexts_ptr
)[i
] = ipa_context_from_jfunc (parms_info
, e
,
971 /* TODO: When IPA-CP starts propagating and merging aggregate jump
972 functions, use its knowledge of the caller too, just like the
973 scalar case above. */
974 known_aggs
[i
] = &jf
->agg
;
977 else if (e
->call_stmt
&& !e
->call_stmt_cannot_inline_p
978 && ((clause_ptr
&& info
->conds
) || known_vals_ptr
))
980 int i
, count
= (int)gimple_call_num_args (e
->call_stmt
);
982 if (count
&& (info
->conds
|| known_vals_ptr
))
983 known_vals
.safe_grow_cleared (count
);
984 for (i
= 0; i
< count
; i
++)
986 tree cst
= gimple_call_arg (e
->call_stmt
, i
);
987 if (!is_gimple_min_invariant (cst
))
995 *clause_ptr
= evaluate_conditions_for_known_args (callee
, inline_p
,
996 known_vals
, known_aggs
);
999 *known_vals_ptr
= known_vals
;
1001 known_vals
.release ();
1004 *known_aggs_ptr
= known_aggs
;
1006 known_aggs
.release ();
1010 /* Allocate the inline summary vector or resize it to cover all cgraph nodes. */
1013 inline_summary_alloc (void)
1015 if (!edge_removal_hook_holder
)
1016 edge_removal_hook_holder
=
1017 symtab
->add_edge_removal_hook (&inline_edge_removal_hook
, NULL
);
1018 if (!edge_duplication_hook_holder
)
1019 edge_duplication_hook_holder
=
1020 symtab
->add_edge_duplication_hook (&inline_edge_duplication_hook
, NULL
);
1022 if (!inline_summaries
)
1023 inline_summaries
= (inline_summary_t
*) inline_summary_t::create_ggc (symtab
);
1025 if (inline_edge_summary_vec
.length () <= (unsigned) symtab
->edges_max_uid
)
1026 inline_edge_summary_vec
.safe_grow_cleared (symtab
->edges_max_uid
+ 1);
1027 if (!edge_predicate_pool
)
1028 edge_predicate_pool
= create_alloc_pool ("edge predicates",
1029 sizeof (struct predicate
), 10);
1032 /* We are called multiple time for given function; clear
1033 data from previous run so they are not cumulated. */
1036 reset_inline_edge_summary (struct cgraph_edge
*e
)
1038 if (e
->uid
< (int) inline_edge_summary_vec
.length ())
1040 struct inline_edge_summary
*es
= inline_edge_summary (e
);
1042 es
->call_stmt_size
= es
->call_stmt_time
= 0;
1044 pool_free (edge_predicate_pool
, es
->predicate
);
1045 es
->predicate
= NULL
;
1046 es
->param
.release ();
1050 /* We are called multiple time for given function; clear
1051 data from previous run so they are not cumulated. */
1054 reset_inline_summary (struct cgraph_node
*node
,
1055 inline_summary
*info
)
1057 struct cgraph_edge
*e
;
1059 info
->self_size
= info
->self_time
= 0;
1060 info
->estimated_stack_size
= 0;
1061 info
->estimated_self_stack_size
= 0;
1062 info
->stack_frame_offset
= 0;
1067 if (info
->loop_iterations
)
1069 pool_free (edge_predicate_pool
, info
->loop_iterations
);
1070 info
->loop_iterations
= NULL
;
1072 if (info
->loop_stride
)
1074 pool_free (edge_predicate_pool
, info
->loop_stride
);
1075 info
->loop_stride
= NULL
;
1077 if (info
->array_index
)
1079 pool_free (edge_predicate_pool
, info
->array_index
);
1080 info
->array_index
= NULL
;
1082 vec_free (info
->conds
);
1083 vec_free (info
->entry
);
1084 for (e
= node
->callees
; e
; e
= e
->next_callee
)
1085 reset_inline_edge_summary (e
);
1086 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
1087 reset_inline_edge_summary (e
);
1090 /* Hook that is called by cgraph.c when a node is removed. */
1093 inline_summary_t::remove (cgraph_node
*node
, inline_summary
*info
)
1095 reset_inline_summary (node
, info
);
1098 /* Remap predicate P of former function to be predicate of duplicated function.
1099 POSSIBLE_TRUTHS is clause of possible truths in the duplicated node,
1100 INFO is inline summary of the duplicated node. */
1102 static struct predicate
1103 remap_predicate_after_duplication (struct predicate
*p
,
1104 clause_t possible_truths
,
1105 struct inline_summary
*info
)
1107 struct predicate new_predicate
= true_predicate ();
1109 for (j
= 0; p
->clause
[j
]; j
++)
1110 if (!(possible_truths
& p
->clause
[j
]))
1112 new_predicate
= false_predicate ();
1116 add_clause (info
->conds
, &new_predicate
,
1117 possible_truths
& p
->clause
[j
]);
1118 return new_predicate
;
1121 /* Same as remap_predicate_after_duplication but handle hint predicate *P.
1122 Additionally care about allocating new memory slot for updated predicate
1123 and set it to NULL when it becomes true or false (and thus uninteresting).
1127 remap_hint_predicate_after_duplication (struct predicate
**p
,
1128 clause_t possible_truths
,
1129 struct inline_summary
*info
)
1131 struct predicate new_predicate
;
1136 new_predicate
= remap_predicate_after_duplication (*p
,
1137 possible_truths
, info
);
1138 /* We do not want to free previous predicate; it is used by node origin. */
1140 set_hint_predicate (p
, new_predicate
);
1144 /* Hook that is called by cgraph.c when a node is duplicated. */
1146 inline_summary_t::duplicate (cgraph_node
*src
,
1149 inline_summary
*info
)
1151 inline_summary_alloc ();
1152 memcpy (info
, inline_summaries
->get (src
), sizeof (inline_summary
));
1153 /* TODO: as an optimization, we may avoid copying conditions
1154 that are known to be false or true. */
1155 info
->conds
= vec_safe_copy (info
->conds
);
1157 /* When there are any replacements in the function body, see if we can figure
1158 out that something was optimized out. */
1159 if (ipa_node_params_sum
&& dst
->clone
.tree_map
)
1161 vec
<size_time_entry
, va_gc
> *entry
= info
->entry
;
1162 /* Use SRC parm info since it may not be copied yet. */
1163 struct ipa_node_params
*parms_info
= IPA_NODE_REF (src
);
1164 vec
<tree
> known_vals
= vNULL
;
1165 int count
= ipa_get_param_count (parms_info
);
1167 clause_t possible_truths
;
1168 struct predicate true_pred
= true_predicate ();
1170 int optimized_out_size
= 0;
1171 bool inlined_to_p
= false;
1172 struct cgraph_edge
*edge
;
1175 known_vals
.safe_grow_cleared (count
);
1176 for (i
= 0; i
< count
; i
++)
1178 struct ipa_replace_map
*r
;
1180 for (j
= 0; vec_safe_iterate (dst
->clone
.tree_map
, j
, &r
); j
++)
1182 if (((!r
->old_tree
&& r
->parm_num
== i
)
1183 || (r
->old_tree
&& r
->old_tree
== ipa_get_param (parms_info
, i
)))
1184 && r
->replace_p
&& !r
->ref_p
)
1186 known_vals
[i
] = r
->new_tree
;
1191 possible_truths
= evaluate_conditions_for_known_args (dst
, false,
1194 known_vals
.release ();
1196 account_size_time (info
, 0, 0, &true_pred
);
1198 /* Remap size_time vectors.
1199 Simplify the predicate by prunning out alternatives that are known
1201 TODO: as on optimization, we can also eliminate conditions known
1203 for (i
= 0; vec_safe_iterate (entry
, i
, &e
); i
++)
1205 struct predicate new_predicate
;
1206 new_predicate
= remap_predicate_after_duplication (&e
->predicate
,
1209 if (false_predicate_p (&new_predicate
))
1210 optimized_out_size
+= e
->size
;
1212 account_size_time (info
, e
->size
, e
->time
, &new_predicate
);
1215 /* Remap edge predicates with the same simplification as above.
1216 Also copy constantness arrays. */
1217 for (edge
= dst
->callees
; edge
; edge
= edge
->next_callee
)
1219 struct predicate new_predicate
;
1220 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
1222 if (!edge
->inline_failed
)
1223 inlined_to_p
= true;
1226 new_predicate
= remap_predicate_after_duplication (es
->predicate
,
1229 if (false_predicate_p (&new_predicate
)
1230 && !false_predicate_p (es
->predicate
))
1232 optimized_out_size
+= es
->call_stmt_size
* INLINE_SIZE_SCALE
;
1233 edge
->frequency
= 0;
1235 edge_set_predicate (edge
, &new_predicate
);
1238 /* Remap indirect edge predicates with the same simplificaiton as above.
1239 Also copy constantness arrays. */
1240 for (edge
= dst
->indirect_calls
; edge
; edge
= edge
->next_callee
)
1242 struct predicate new_predicate
;
1243 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
1245 gcc_checking_assert (edge
->inline_failed
);
1248 new_predicate
= remap_predicate_after_duplication (es
->predicate
,
1251 if (false_predicate_p (&new_predicate
)
1252 && !false_predicate_p (es
->predicate
))
1254 optimized_out_size
+= es
->call_stmt_size
* INLINE_SIZE_SCALE
;
1255 edge
->frequency
= 0;
1257 edge_set_predicate (edge
, &new_predicate
);
1259 remap_hint_predicate_after_duplication (&info
->loop_iterations
,
1260 possible_truths
, info
);
1261 remap_hint_predicate_after_duplication (&info
->loop_stride
,
1262 possible_truths
, info
);
1263 remap_hint_predicate_after_duplication (&info
->array_index
,
1264 possible_truths
, info
);
1266 /* If inliner or someone after inliner will ever start producing
1267 non-trivial clones, we will get trouble with lack of information
1268 about updating self sizes, because size vectors already contains
1269 sizes of the calees. */
1270 gcc_assert (!inlined_to_p
|| !optimized_out_size
);
1274 info
->entry
= vec_safe_copy (info
->entry
);
1275 if (info
->loop_iterations
)
1277 predicate p
= *info
->loop_iterations
;
1278 info
->loop_iterations
= NULL
;
1279 set_hint_predicate (&info
->loop_iterations
, p
);
1281 if (info
->loop_stride
)
1283 predicate p
= *info
->loop_stride
;
1284 info
->loop_stride
= NULL
;
1285 set_hint_predicate (&info
->loop_stride
, p
);
1287 if (info
->array_index
)
1289 predicate p
= *info
->array_index
;
1290 info
->array_index
= NULL
;
1291 set_hint_predicate (&info
->array_index
, p
);
1294 if (!dst
->global
.inlined_to
)
1295 inline_update_overall_summary (dst
);
1299 /* Hook that is called by cgraph.c when a node is duplicated. */
1302 inline_edge_duplication_hook (struct cgraph_edge
*src
,
1303 struct cgraph_edge
*dst
,
1304 ATTRIBUTE_UNUSED
void *data
)
1306 struct inline_edge_summary
*info
;
1307 struct inline_edge_summary
*srcinfo
;
1308 inline_summary_alloc ();
1309 info
= inline_edge_summary (dst
);
1310 srcinfo
= inline_edge_summary (src
);
1311 memcpy (info
, srcinfo
, sizeof (struct inline_edge_summary
));
1312 info
->predicate
= NULL
;
1313 edge_set_predicate (dst
, srcinfo
->predicate
);
1314 info
->param
= srcinfo
->param
.copy ();
1315 if (!dst
->indirect_unknown_callee
&& src
->indirect_unknown_callee
)
1317 info
->call_stmt_size
-= (eni_size_weights
.indirect_call_cost
1318 - eni_size_weights
.call_cost
);
1319 info
->call_stmt_time
-= (eni_time_weights
.indirect_call_cost
1320 - eni_time_weights
.call_cost
);
1325 /* Keep edge cache consistent across edge removal. */
1328 inline_edge_removal_hook (struct cgraph_edge
*edge
,
1329 void *data ATTRIBUTE_UNUSED
)
1331 if (edge_growth_cache
.exists ())
1332 reset_edge_growth_cache (edge
);
1333 reset_inline_edge_summary (edge
);
1337 /* Initialize growth caches. */
1340 initialize_growth_caches (void)
1342 if (symtab
->edges_max_uid
)
1343 edge_growth_cache
.safe_grow_cleared (symtab
->edges_max_uid
);
1347 /* Free growth caches. */
1350 free_growth_caches (void)
1352 edge_growth_cache
.release ();
1356 /* Dump edge summaries associated to NODE and recursively to all clones.
1357 Indent by INDENT. */
1360 dump_inline_edge_summary (FILE *f
, int indent
, struct cgraph_node
*node
,
1361 struct inline_summary
*info
)
1363 struct cgraph_edge
*edge
;
1364 for (edge
= node
->callees
; edge
; edge
= edge
->next_callee
)
1366 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
1367 struct cgraph_node
*callee
= edge
->callee
->ultimate_alias_target ();
1371 "%*s%s/%i %s\n%*s loop depth:%2i freq:%4i size:%2i"
1372 " time: %2i callee size:%2i stack:%2i",
1373 indent
, "", callee
->name (), callee
->order
,
1374 !edge
->inline_failed
1375 ? "inlined" : cgraph_inline_failed_string (edge
-> inline_failed
),
1376 indent
, "", es
->loop_depth
, edge
->frequency
,
1377 es
->call_stmt_size
, es
->call_stmt_time
,
1378 (int) inline_summaries
->get (callee
)->size
/ INLINE_SIZE_SCALE
,
1379 (int) inline_summaries
->get (callee
)->estimated_stack_size
);
1383 fprintf (f
, " predicate: ");
1384 dump_predicate (f
, info
->conds
, es
->predicate
);
1388 if (es
->param
.exists ())
1389 for (i
= 0; i
< (int) es
->param
.length (); i
++)
1391 int prob
= es
->param
[i
].change_prob
;
1394 fprintf (f
, "%*s op%i is compile time invariant\n",
1396 else if (prob
!= REG_BR_PROB_BASE
)
1397 fprintf (f
, "%*s op%i change %f%% of time\n", indent
+ 2, "", i
,
1398 prob
* 100.0 / REG_BR_PROB_BASE
);
1400 if (!edge
->inline_failed
)
1402 fprintf (f
, "%*sStack frame offset %i, callee self size %i,"
1403 " callee size %i\n",
1405 (int) inline_summaries
->get (callee
)->stack_frame_offset
,
1406 (int) inline_summaries
->get (callee
)->estimated_self_stack_size
,
1407 (int) inline_summaries
->get (callee
)->estimated_stack_size
);
1408 dump_inline_edge_summary (f
, indent
+ 2, callee
, info
);
1411 for (edge
= node
->indirect_calls
; edge
; edge
= edge
->next_callee
)
1413 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
1414 fprintf (f
, "%*sindirect call loop depth:%2i freq:%4i size:%2i"
1418 edge
->frequency
, es
->call_stmt_size
, es
->call_stmt_time
);
1421 fprintf (f
, "predicate: ");
1422 dump_predicate (f
, info
->conds
, es
->predicate
);
1431 dump_inline_summary (FILE *f
, struct cgraph_node
*node
)
1433 if (node
->definition
)
1435 struct inline_summary
*s
= inline_summaries
->get (node
);
1438 fprintf (f
, "Inline summary for %s/%i", node
->name (),
1440 if (DECL_DISREGARD_INLINE_LIMITS (node
->decl
))
1441 fprintf (f
, " always_inline");
1443 fprintf (f
, " inlinable");
1444 fprintf (f
, "\n self time: %i\n", s
->self_time
);
1445 fprintf (f
, " global time: %i\n", s
->time
);
1446 fprintf (f
, " self size: %i\n", s
->self_size
);
1447 fprintf (f
, " global size: %i\n", s
->size
);
1448 fprintf (f
, " min size: %i\n", s
->min_size
);
1449 fprintf (f
, " self stack: %i\n",
1450 (int) s
->estimated_self_stack_size
);
1451 fprintf (f
, " global stack: %i\n", (int) s
->estimated_stack_size
);
1453 fprintf (f
, " estimated growth:%i\n", (int) s
->growth
);
1455 fprintf (f
, " In SCC: %i\n", (int) s
->scc_no
);
1456 for (i
= 0; vec_safe_iterate (s
->entry
, i
, &e
); i
++)
1458 fprintf (f
, " size:%f, time:%f, predicate:",
1459 (double) e
->size
/ INLINE_SIZE_SCALE
,
1460 (double) e
->time
/ INLINE_TIME_SCALE
);
1461 dump_predicate (f
, s
->conds
, &e
->predicate
);
1463 if (s
->loop_iterations
)
1465 fprintf (f
, " loop iterations:");
1466 dump_predicate (f
, s
->conds
, s
->loop_iterations
);
1470 fprintf (f
, " loop stride:");
1471 dump_predicate (f
, s
->conds
, s
->loop_stride
);
1475 fprintf (f
, " array index:");
1476 dump_predicate (f
, s
->conds
, s
->array_index
);
1478 fprintf (f
, " calls:\n");
1479 dump_inline_edge_summary (f
, 4, node
, s
);
1485 debug_inline_summary (struct cgraph_node
*node
)
1487 dump_inline_summary (stderr
, node
);
1491 dump_inline_summaries (FILE *f
)
1493 struct cgraph_node
*node
;
1495 FOR_EACH_DEFINED_FUNCTION (node
)
1496 if (!node
->global
.inlined_to
)
1497 dump_inline_summary (f
, node
);
1500 /* Give initial reasons why inlining would fail on EDGE. This gets either
1501 nullified or usually overwritten by more precise reasons later. */
1504 initialize_inline_failed (struct cgraph_edge
*e
)
1506 struct cgraph_node
*callee
= e
->callee
;
1508 if (e
->indirect_unknown_callee
)
1509 e
->inline_failed
= CIF_INDIRECT_UNKNOWN_CALL
;
1510 else if (!callee
->definition
)
1511 e
->inline_failed
= CIF_BODY_NOT_AVAILABLE
;
1512 else if (callee
->local
.redefined_extern_inline
)
1513 e
->inline_failed
= CIF_REDEFINED_EXTERN_INLINE
;
1514 else if (e
->call_stmt_cannot_inline_p
)
1515 e
->inline_failed
= CIF_MISMATCHED_ARGUMENTS
;
1516 else if (cfun
&& fn_contains_cilk_spawn_p (cfun
))
1517 /* We can't inline if the function is spawing a function. */
1518 e
->inline_failed
= CIF_FUNCTION_NOT_INLINABLE
;
1520 e
->inline_failed
= CIF_FUNCTION_NOT_CONSIDERED
;
1523 /* Callback of walk_aliased_vdefs. Flags that it has been invoked to the
1524 boolean variable pointed to by DATA. */
1527 mark_modified (ao_ref
*ao ATTRIBUTE_UNUSED
, tree vdef ATTRIBUTE_UNUSED
,
1530 bool *b
= (bool *) data
;
1535 /* If OP refers to value of function parameter, return the corresponding
1539 unmodified_parm_1 (gimple stmt
, tree op
)
1541 /* SSA_NAME referring to parm default def? */
1542 if (TREE_CODE (op
) == SSA_NAME
1543 && SSA_NAME_IS_DEFAULT_DEF (op
)
1544 && TREE_CODE (SSA_NAME_VAR (op
)) == PARM_DECL
)
1545 return SSA_NAME_VAR (op
);
1546 /* Non-SSA parm reference? */
1547 if (TREE_CODE (op
) == PARM_DECL
)
1549 bool modified
= false;
1552 ao_ref_init (&refd
, op
);
1553 walk_aliased_vdefs (&refd
, gimple_vuse (stmt
), mark_modified
, &modified
,
1561 /* If OP refers to value of function parameter, return the corresponding
1562 parameter. Also traverse chains of SSA register assignments. */
1565 unmodified_parm (gimple stmt
, tree op
)
1567 tree res
= unmodified_parm_1 (stmt
, op
);
1571 if (TREE_CODE (op
) == SSA_NAME
1572 && !SSA_NAME_IS_DEFAULT_DEF (op
)
1573 && gimple_assign_single_p (SSA_NAME_DEF_STMT (op
)))
1574 return unmodified_parm (SSA_NAME_DEF_STMT (op
),
1575 gimple_assign_rhs1 (SSA_NAME_DEF_STMT (op
)));
1579 /* If OP refers to a value of a function parameter or value loaded from an
1580 aggregate passed to a parameter (either by value or reference), return TRUE
1581 and store the number of the parameter to *INDEX_P and information whether
1582 and how it has been loaded from an aggregate into *AGGPOS. INFO describes
1583 the function parameters, STMT is the statement in which OP is used or
1587 unmodified_parm_or_parm_agg_item (struct ipa_node_params
*info
,
1588 gimple stmt
, tree op
, int *index_p
,
1589 struct agg_position_info
*aggpos
)
1591 tree res
= unmodified_parm_1 (stmt
, op
);
1593 gcc_checking_assert (aggpos
);
1596 *index_p
= ipa_get_param_decl_index (info
, res
);
1599 aggpos
->agg_contents
= false;
1600 aggpos
->by_ref
= false;
1604 if (TREE_CODE (op
) == SSA_NAME
)
1606 if (SSA_NAME_IS_DEFAULT_DEF (op
)
1607 || !gimple_assign_single_p (SSA_NAME_DEF_STMT (op
)))
1609 stmt
= SSA_NAME_DEF_STMT (op
);
1610 op
= gimple_assign_rhs1 (stmt
);
1611 if (!REFERENCE_CLASS_P (op
))
1612 return unmodified_parm_or_parm_agg_item (info
, stmt
, op
, index_p
,
1616 aggpos
->agg_contents
= true;
1617 return ipa_load_from_parm_agg (info
, stmt
, op
, index_p
, &aggpos
->offset
,
1621 /* See if statement might disappear after inlining.
1622 0 - means not eliminated
1623 1 - half of statements goes away
1624 2 - for sure it is eliminated.
1625 We are not terribly sophisticated, basically looking for simple abstraction
1626 penalty wrappers. */
1629 eliminated_by_inlining_prob (gimple stmt
)
1631 enum gimple_code code
= gimple_code (stmt
);
1632 enum tree_code rhs_code
;
1642 if (gimple_num_ops (stmt
) != 2)
1645 rhs_code
= gimple_assign_rhs_code (stmt
);
1647 /* Casts of parameters, loads from parameters passed by reference
1648 and stores to return value or parameters are often free after
1649 inlining dua to SRA and further combining.
1650 Assume that half of statements goes away. */
1651 if (CONVERT_EXPR_CODE_P (rhs_code
)
1652 || rhs_code
== VIEW_CONVERT_EXPR
1653 || rhs_code
== ADDR_EXPR
1654 || gimple_assign_rhs_class (stmt
) == GIMPLE_SINGLE_RHS
)
1656 tree rhs
= gimple_assign_rhs1 (stmt
);
1657 tree lhs
= gimple_assign_lhs (stmt
);
1658 tree inner_rhs
= get_base_address (rhs
);
1659 tree inner_lhs
= get_base_address (lhs
);
1660 bool rhs_free
= false;
1661 bool lhs_free
= false;
1668 /* Reads of parameter are expected to be free. */
1669 if (unmodified_parm (stmt
, inner_rhs
))
1671 /* Match expressions of form &this->field. Those will most likely
1672 combine with something upstream after inlining. */
1673 else if (TREE_CODE (inner_rhs
) == ADDR_EXPR
)
1675 tree op
= get_base_address (TREE_OPERAND (inner_rhs
, 0));
1676 if (TREE_CODE (op
) == PARM_DECL
)
1678 else if (TREE_CODE (op
) == MEM_REF
1679 && unmodified_parm (stmt
, TREE_OPERAND (op
, 0)))
1683 /* When parameter is not SSA register because its address is taken
1684 and it is just copied into one, the statement will be completely
1685 free after inlining (we will copy propagate backward). */
1686 if (rhs_free
&& is_gimple_reg (lhs
))
1689 /* Reads of parameters passed by reference
1690 expected to be free (i.e. optimized out after inlining). */
1691 if (TREE_CODE (inner_rhs
) == MEM_REF
1692 && unmodified_parm (stmt
, TREE_OPERAND (inner_rhs
, 0)))
1695 /* Copying parameter passed by reference into gimple register is
1696 probably also going to copy propagate, but we can't be quite
1698 if (rhs_free
&& is_gimple_reg (lhs
))
1701 /* Writes to parameters, parameters passed by value and return value
1702 (either dirrectly or passed via invisible reference) are free.
1704 TODO: We ought to handle testcase like
1705 struct a {int a,b;};
1707 retrurnsturct (void)
1713 This translate into:
1728 For that we either need to copy ipa-split logic detecting writes
1730 if (TREE_CODE (inner_lhs
) == PARM_DECL
1731 || TREE_CODE (inner_lhs
) == RESULT_DECL
1732 || (TREE_CODE (inner_lhs
) == MEM_REF
1733 && (unmodified_parm (stmt
, TREE_OPERAND (inner_lhs
, 0))
1734 || (TREE_CODE (TREE_OPERAND (inner_lhs
, 0)) == SSA_NAME
1735 && SSA_NAME_VAR (TREE_OPERAND (inner_lhs
, 0))
1736 && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND
1738 0))) == RESULT_DECL
))))
1741 && (is_gimple_reg (rhs
) || is_gimple_min_invariant (rhs
)))
1743 if (lhs_free
&& rhs_free
)
1753 /* If BB ends by a conditional we can turn into predicates, attach corresponding
1754 predicates to the CFG edges. */
1757 set_cond_stmt_execution_predicate (struct ipa_node_params
*info
,
1758 struct inline_summary
*summary
,
1764 struct agg_position_info aggpos
;
1765 enum tree_code code
, inverted_code
;
1771 last
= last_stmt (bb
);
1772 if (!last
|| gimple_code (last
) != GIMPLE_COND
)
1774 if (!is_gimple_ip_invariant (gimple_cond_rhs (last
)))
1776 op
= gimple_cond_lhs (last
);
1777 /* TODO: handle conditionals like
1780 if (unmodified_parm_or_parm_agg_item (info
, last
, op
, &index
, &aggpos
))
1782 code
= gimple_cond_code (last
);
1783 inverted_code
= invert_tree_comparison (code
, HONOR_NANS (op
));
1785 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1787 enum tree_code this_code
= (e
->flags
& EDGE_TRUE_VALUE
1788 ? code
: inverted_code
);
1789 /* invert_tree_comparison will return ERROR_MARK on FP
1790 comparsions that are not EQ/NE instead of returning proper
1791 unordered one. Be sure it is not confused with NON_CONSTANT. */
1792 if (this_code
!= ERROR_MARK
)
1794 struct predicate p
= add_condition (summary
, index
, &aggpos
,
1796 gimple_cond_rhs (last
));
1797 e
->aux
= pool_alloc (edge_predicate_pool
);
1798 *(struct predicate
*) e
->aux
= p
;
1803 if (TREE_CODE (op
) != SSA_NAME
)
1806 if (builtin_constant_p (op))
1810 Here we can predicate nonconstant_code. We can't
1811 really handle constant_code since we have no predicate
1812 for this and also the constant code is not known to be
1813 optimized away when inliner doen't see operand is constant.
1814 Other optimizers might think otherwise. */
1815 if (gimple_cond_code (last
) != NE_EXPR
1816 || !integer_zerop (gimple_cond_rhs (last
)))
1818 set_stmt
= SSA_NAME_DEF_STMT (op
);
1819 if (!gimple_call_builtin_p (set_stmt
, BUILT_IN_CONSTANT_P
)
1820 || gimple_call_num_args (set_stmt
) != 1)
1822 op2
= gimple_call_arg (set_stmt
, 0);
1823 if (!unmodified_parm_or_parm_agg_item
1824 (info
, set_stmt
, op2
, &index
, &aggpos
))
1826 FOR_EACH_EDGE (e
, ei
, bb
->succs
) if (e
->flags
& EDGE_FALSE_VALUE
)
1828 struct predicate p
= add_condition (summary
, index
, &aggpos
,
1829 IS_NOT_CONSTANT
, NULL_TREE
);
1830 e
->aux
= pool_alloc (edge_predicate_pool
);
1831 *(struct predicate
*) e
->aux
= p
;
1836 /* If BB ends by a switch we can turn into predicates, attach corresponding
1837 predicates to the CFG edges. */
1840 set_switch_stmt_execution_predicate (struct ipa_node_params
*info
,
1841 struct inline_summary
*summary
,
1847 struct agg_position_info aggpos
;
1853 lastg
= last_stmt (bb
);
1854 if (!lastg
|| gimple_code (lastg
) != GIMPLE_SWITCH
)
1856 gswitch
*last
= as_a
<gswitch
*> (lastg
);
1857 op
= gimple_switch_index (last
);
1858 if (!unmodified_parm_or_parm_agg_item (info
, last
, op
, &index
, &aggpos
))
1861 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1863 e
->aux
= pool_alloc (edge_predicate_pool
);
1864 *(struct predicate
*) e
->aux
= false_predicate ();
1866 n
= gimple_switch_num_labels (last
);
1867 for (case_idx
= 0; case_idx
< n
; ++case_idx
)
1869 tree cl
= gimple_switch_label (last
, case_idx
);
1873 e
= find_edge (bb
, label_to_block (CASE_LABEL (cl
)));
1874 min
= CASE_LOW (cl
);
1875 max
= CASE_HIGH (cl
);
1877 /* For default we might want to construct predicate that none
1878 of cases is met, but it is bit hard to do not having negations
1879 of conditionals handy. */
1881 p
= true_predicate ();
1883 p
= add_condition (summary
, index
, &aggpos
, EQ_EXPR
, min
);
1886 struct predicate p1
, p2
;
1887 p1
= add_condition (summary
, index
, &aggpos
, GE_EXPR
, min
);
1888 p2
= add_condition (summary
, index
, &aggpos
, LE_EXPR
, max
);
1889 p
= and_predicates (summary
->conds
, &p1
, &p2
);
1891 *(struct predicate
*) e
->aux
1892 = or_predicates (summary
->conds
, &p
, (struct predicate
*) e
->aux
);
1897 /* For each BB in NODE attach to its AUX pointer predicate under
1898 which it is executable. */
1901 compute_bb_predicates (struct cgraph_node
*node
,
1902 struct ipa_node_params
*parms_info
,
1903 struct inline_summary
*summary
)
1905 struct function
*my_function
= DECL_STRUCT_FUNCTION (node
->decl
);
1909 FOR_EACH_BB_FN (bb
, my_function
)
1911 set_cond_stmt_execution_predicate (parms_info
, summary
, bb
);
1912 set_switch_stmt_execution_predicate (parms_info
, summary
, bb
);
1915 /* Entry block is always executable. */
1916 ENTRY_BLOCK_PTR_FOR_FN (my_function
)->aux
1917 = pool_alloc (edge_predicate_pool
);
1918 *(struct predicate
*) ENTRY_BLOCK_PTR_FOR_FN (my_function
)->aux
1919 = true_predicate ();
1921 /* A simple dataflow propagation of predicates forward in the CFG.
1922 TODO: work in reverse postorder. */
1926 FOR_EACH_BB_FN (bb
, my_function
)
1928 struct predicate p
= false_predicate ();
1931 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
1935 struct predicate this_bb_predicate
1936 = *(struct predicate
*) e
->src
->aux
;
1939 = and_predicates (summary
->conds
, &this_bb_predicate
,
1940 (struct predicate
*) e
->aux
);
1941 p
= or_predicates (summary
->conds
, &p
, &this_bb_predicate
);
1942 if (true_predicate_p (&p
))
1946 if (false_predicate_p (&p
))
1947 gcc_assert (!bb
->aux
);
1953 bb
->aux
= pool_alloc (edge_predicate_pool
);
1954 *((struct predicate
*) bb
->aux
) = p
;
1956 else if (!predicates_equal_p (&p
, (struct predicate
*) bb
->aux
))
1958 /* This OR operation is needed to ensure monotonous data flow
1959 in the case we hit the limit on number of clauses and the
1960 and/or operations above give approximate answers. */
1961 p
= or_predicates (summary
->conds
, &p
, (struct predicate
*)bb
->aux
);
1962 if (!predicates_equal_p (&p
, (struct predicate
*) bb
->aux
))
1965 *((struct predicate
*) bb
->aux
) = p
;
1974 /* We keep info about constantness of SSA names. */
1976 typedef struct predicate predicate_t
;
1977 /* Return predicate specifying when the STMT might have result that is not
1978 a compile time constant. */
1980 static struct predicate
1981 will_be_nonconstant_expr_predicate (struct ipa_node_params
*info
,
1982 struct inline_summary
*summary
,
1984 vec
<predicate_t
> nonconstant_names
)
1989 while (UNARY_CLASS_P (expr
))
1990 expr
= TREE_OPERAND (expr
, 0);
1992 parm
= unmodified_parm (NULL
, expr
);
1993 if (parm
&& (index
= ipa_get_param_decl_index (info
, parm
)) >= 0)
1994 return add_condition (summary
, index
, NULL
, CHANGED
, NULL_TREE
);
1995 if (is_gimple_min_invariant (expr
))
1996 return false_predicate ();
1997 if (TREE_CODE (expr
) == SSA_NAME
)
1998 return nonconstant_names
[SSA_NAME_VERSION (expr
)];
1999 if (BINARY_CLASS_P (expr
) || COMPARISON_CLASS_P (expr
))
2001 struct predicate p1
= will_be_nonconstant_expr_predicate
2002 (info
, summary
, TREE_OPERAND (expr
, 0),
2004 struct predicate p2
;
2005 if (true_predicate_p (&p1
))
2007 p2
= will_be_nonconstant_expr_predicate (info
, summary
,
2008 TREE_OPERAND (expr
, 1),
2010 return or_predicates (summary
->conds
, &p1
, &p2
);
2012 else if (TREE_CODE (expr
) == COND_EXPR
)
2014 struct predicate p1
= will_be_nonconstant_expr_predicate
2015 (info
, summary
, TREE_OPERAND (expr
, 0),
2017 struct predicate p2
;
2018 if (true_predicate_p (&p1
))
2020 p2
= will_be_nonconstant_expr_predicate (info
, summary
,
2021 TREE_OPERAND (expr
, 1),
2023 if (true_predicate_p (&p2
))
2025 p1
= or_predicates (summary
->conds
, &p1
, &p2
);
2026 p2
= will_be_nonconstant_expr_predicate (info
, summary
,
2027 TREE_OPERAND (expr
, 2),
2029 return or_predicates (summary
->conds
, &p1
, &p2
);
2036 return false_predicate ();
2040 /* Return predicate specifying when the STMT might have result that is not
2041 a compile time constant. */
2043 static struct predicate
2044 will_be_nonconstant_predicate (struct ipa_node_params
*info
,
2045 struct inline_summary
*summary
,
2047 vec
<predicate_t
> nonconstant_names
)
2049 struct predicate p
= true_predicate ();
2052 struct predicate op_non_const
;
2055 struct agg_position_info aggpos
;
2057 /* What statments might be optimized away
2058 when their arguments are constant. */
2059 if (gimple_code (stmt
) != GIMPLE_ASSIGN
2060 && gimple_code (stmt
) != GIMPLE_COND
2061 && gimple_code (stmt
) != GIMPLE_SWITCH
2062 && (gimple_code (stmt
) != GIMPLE_CALL
2063 || !(gimple_call_flags (stmt
) & ECF_CONST
)))
2066 /* Stores will stay anyway. */
2067 if (gimple_store_p (stmt
))
2070 is_load
= gimple_assign_load_p (stmt
);
2072 /* Loads can be optimized when the value is known. */
2076 gcc_assert (gimple_assign_single_p (stmt
));
2077 op
= gimple_assign_rhs1 (stmt
);
2078 if (!unmodified_parm_or_parm_agg_item (info
, stmt
, op
, &base_index
,
2085 /* See if we understand all operands before we start
2086 adding conditionals. */
2087 FOR_EACH_SSA_TREE_OPERAND (use
, stmt
, iter
, SSA_OP_USE
)
2089 tree parm
= unmodified_parm (stmt
, use
);
2090 /* For arguments we can build a condition. */
2091 if (parm
&& ipa_get_param_decl_index (info
, parm
) >= 0)
2093 if (TREE_CODE (use
) != SSA_NAME
)
2095 /* If we know when operand is constant,
2096 we still can say something useful. */
2097 if (!true_predicate_p (&nonconstant_names
[SSA_NAME_VERSION (use
)]))
2104 add_condition (summary
, base_index
, &aggpos
, CHANGED
, NULL
);
2106 op_non_const
= false_predicate ();
2107 FOR_EACH_SSA_TREE_OPERAND (use
, stmt
, iter
, SSA_OP_USE
)
2109 tree parm
= unmodified_parm (stmt
, use
);
2112 if (parm
&& (index
= ipa_get_param_decl_index (info
, parm
)) >= 0)
2114 if (index
!= base_index
)
2115 p
= add_condition (summary
, index
, NULL
, CHANGED
, NULL_TREE
);
2120 p
= nonconstant_names
[SSA_NAME_VERSION (use
)];
2121 op_non_const
= or_predicates (summary
->conds
, &p
, &op_non_const
);
2123 if ((gimple_code (stmt
) == GIMPLE_ASSIGN
|| gimple_code (stmt
) == GIMPLE_CALL
)
2124 && gimple_op (stmt
, 0)
2125 && TREE_CODE (gimple_op (stmt
, 0)) == SSA_NAME
)
2126 nonconstant_names
[SSA_NAME_VERSION (gimple_op (stmt
, 0))]
2128 return op_non_const
;
2131 struct record_modified_bb_info
2137 /* Callback of walk_aliased_vdefs. Records basic blocks where the value may be
2138 set except for info->stmt. */
2141 record_modified (ao_ref
*ao ATTRIBUTE_UNUSED
, tree vdef
, void *data
)
2143 struct record_modified_bb_info
*info
=
2144 (struct record_modified_bb_info
*) data
;
2145 if (SSA_NAME_DEF_STMT (vdef
) == info
->stmt
)
2147 bitmap_set_bit (info
->bb_set
,
2148 SSA_NAME_IS_DEFAULT_DEF (vdef
)
2149 ? ENTRY_BLOCK_PTR_FOR_FN (cfun
)->index
2150 : gimple_bb (SSA_NAME_DEF_STMT (vdef
))->index
);
2154 /* Return probability (based on REG_BR_PROB_BASE) that I-th parameter of STMT
2155 will change since last invocation of STMT.
2157 Value 0 is reserved for compile time invariants.
2158 For common parameters it is REG_BR_PROB_BASE. For loop invariants it
2159 ought to be REG_BR_PROB_BASE / estimated_iters. */
2162 param_change_prob (gimple stmt
, int i
)
2164 tree op
= gimple_call_arg (stmt
, i
);
2165 basic_block bb
= gimple_bb (stmt
);
2168 /* Global invariants neve change. */
2169 if (is_gimple_min_invariant (op
))
2171 /* We would have to do non-trivial analysis to really work out what
2172 is the probability of value to change (i.e. when init statement
2173 is in a sibling loop of the call).
2175 We do an conservative estimate: when call is executed N times more often
2176 than the statement defining value, we take the frequency 1/N. */
2177 if (TREE_CODE (op
) == SSA_NAME
)
2182 return REG_BR_PROB_BASE
;
2184 if (SSA_NAME_IS_DEFAULT_DEF (op
))
2185 init_freq
= ENTRY_BLOCK_PTR_FOR_FN (cfun
)->frequency
;
2187 init_freq
= gimple_bb (SSA_NAME_DEF_STMT (op
))->frequency
;
2191 if (init_freq
< bb
->frequency
)
2192 return MAX (GCOV_COMPUTE_SCALE (init_freq
, bb
->frequency
), 1);
2194 return REG_BR_PROB_BASE
;
2197 base
= get_base_address (op
);
2202 struct record_modified_bb_info info
;
2205 tree init
= ctor_for_folding (base
);
2207 if (init
!= error_mark_node
)
2210 return REG_BR_PROB_BASE
;
2211 ao_ref_init (&refd
, op
);
2213 info
.bb_set
= BITMAP_ALLOC (NULL
);
2214 walk_aliased_vdefs (&refd
, gimple_vuse (stmt
), record_modified
, &info
,
2216 if (bitmap_bit_p (info
.bb_set
, bb
->index
))
2218 BITMAP_FREE (info
.bb_set
);
2219 return REG_BR_PROB_BASE
;
2222 /* Assume that every memory is initialized at entry.
2223 TODO: Can we easilly determine if value is always defined
2224 and thus we may skip entry block? */
2225 if (ENTRY_BLOCK_PTR_FOR_FN (cfun
)->frequency
)
2226 max
= ENTRY_BLOCK_PTR_FOR_FN (cfun
)->frequency
;
2230 EXECUTE_IF_SET_IN_BITMAP (info
.bb_set
, 0, index
, bi
)
2231 max
= MIN (max
, BASIC_BLOCK_FOR_FN (cfun
, index
)->frequency
);
2233 BITMAP_FREE (info
.bb_set
);
2234 if (max
< bb
->frequency
)
2235 return MAX (GCOV_COMPUTE_SCALE (max
, bb
->frequency
), 1);
2237 return REG_BR_PROB_BASE
;
2239 return REG_BR_PROB_BASE
;
2242 /* Find whether a basic block BB is the final block of a (half) diamond CFG
2243 sub-graph and if the predicate the condition depends on is known. If so,
2244 return true and store the pointer the predicate in *P. */
2247 phi_result_unknown_predicate (struct ipa_node_params
*info
,
2248 inline_summary
*summary
, basic_block bb
,
2249 struct predicate
*p
,
2250 vec
<predicate_t
> nonconstant_names
)
2254 basic_block first_bb
= NULL
;
2257 if (single_pred_p (bb
))
2259 *p
= false_predicate ();
2263 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
2265 if (single_succ_p (e
->src
))
2267 if (!single_pred_p (e
->src
))
2270 first_bb
= single_pred (e
->src
);
2271 else if (single_pred (e
->src
) != first_bb
)
2278 else if (e
->src
!= first_bb
)
2286 stmt
= last_stmt (first_bb
);
2288 || gimple_code (stmt
) != GIMPLE_COND
2289 || !is_gimple_ip_invariant (gimple_cond_rhs (stmt
)))
2292 *p
= will_be_nonconstant_expr_predicate (info
, summary
,
2293 gimple_cond_lhs (stmt
),
2295 if (true_predicate_p (p
))
2301 /* Given a PHI statement in a function described by inline properties SUMMARY
2302 and *P being the predicate describing whether the selected PHI argument is
2303 known, store a predicate for the result of the PHI statement into
2304 NONCONSTANT_NAMES, if possible. */
2307 predicate_for_phi_result (struct inline_summary
*summary
, gphi
*phi
,
2308 struct predicate
*p
,
2309 vec
<predicate_t
> nonconstant_names
)
2313 for (i
= 0; i
< gimple_phi_num_args (phi
); i
++)
2315 tree arg
= gimple_phi_arg (phi
, i
)->def
;
2316 if (!is_gimple_min_invariant (arg
))
2318 gcc_assert (TREE_CODE (arg
) == SSA_NAME
);
2319 *p
= or_predicates (summary
->conds
, p
,
2320 &nonconstant_names
[SSA_NAME_VERSION (arg
)]);
2321 if (true_predicate_p (p
))
2326 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2328 fprintf (dump_file
, "\t\tphi predicate: ");
2329 dump_predicate (dump_file
, summary
->conds
, p
);
2331 nonconstant_names
[SSA_NAME_VERSION (gimple_phi_result (phi
))] = *p
;
2334 /* Return predicate specifying when array index in access OP becomes non-constant. */
2336 static struct predicate
2337 array_index_predicate (inline_summary
*info
,
2338 vec
< predicate_t
> nonconstant_names
, tree op
)
2340 struct predicate p
= false_predicate ();
2341 while (handled_component_p (op
))
2343 if (TREE_CODE (op
) == ARRAY_REF
|| TREE_CODE (op
) == ARRAY_RANGE_REF
)
2345 if (TREE_CODE (TREE_OPERAND (op
, 1)) == SSA_NAME
)
2346 p
= or_predicates (info
->conds
, &p
,
2347 &nonconstant_names
[SSA_NAME_VERSION
2348 (TREE_OPERAND (op
, 1))]);
2350 op
= TREE_OPERAND (op
, 0);
2355 /* For a typical usage of __builtin_expect (a<b, 1), we
2356 may introduce an extra relation stmt:
2357 With the builtin, we have
2360 t3 = __builtin_expect (t2, 1);
2363 Without the builtin, we have
2366 This affects the size/time estimation and may have
2367 an impact on the earlier inlining.
2368 Here find this pattern and fix it up later. */
2371 find_foldable_builtin_expect (basic_block bb
)
2373 gimple_stmt_iterator bsi
;
2375 for (bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
2377 gimple stmt
= gsi_stmt (bsi
);
2378 if (gimple_call_builtin_p (stmt
, BUILT_IN_EXPECT
)
2379 || (is_gimple_call (stmt
)
2380 && gimple_call_internal_p (stmt
)
2381 && gimple_call_internal_fn (stmt
) == IFN_BUILTIN_EXPECT
))
2383 tree var
= gimple_call_lhs (stmt
);
2384 tree arg
= gimple_call_arg (stmt
, 0);
2385 use_operand_p use_p
;
2392 gcc_assert (TREE_CODE (var
) == SSA_NAME
);
2394 while (TREE_CODE (arg
) == SSA_NAME
)
2396 gimple stmt_tmp
= SSA_NAME_DEF_STMT (arg
);
2397 if (!is_gimple_assign (stmt_tmp
))
2399 switch (gimple_assign_rhs_code (stmt_tmp
))
2418 arg
= gimple_assign_rhs1 (stmt_tmp
);
2421 if (match
&& single_imm_use (var
, &use_p
, &use_stmt
)
2422 && gimple_code (use_stmt
) == GIMPLE_COND
)
2429 /* Return true when the basic blocks contains only clobbers followed by RESX.
2430 Such BBs are kept around to make removal of dead stores possible with
2431 presence of EH and will be optimized out by optimize_clobbers later in the
2434 NEED_EH is used to recurse in case the clobber has non-EH predecestors
2435 that can be clobber only, too.. When it is false, the RESX is not necessary
2436 on the end of basic block. */
2439 clobber_only_eh_bb_p (basic_block bb
, bool need_eh
= true)
2441 gimple_stmt_iterator gsi
= gsi_last_bb (bb
);
2447 if (gsi_end_p (gsi
))
2449 if (gimple_code (gsi_stmt (gsi
)) != GIMPLE_RESX
)
2453 else if (!single_succ_p (bb
))
2456 for (; !gsi_end_p (gsi
); gsi_prev (&gsi
))
2458 gimple stmt
= gsi_stmt (gsi
);
2459 if (is_gimple_debug (stmt
))
2461 if (gimple_clobber_p (stmt
))
2463 if (gimple_code (stmt
) == GIMPLE_LABEL
)
2468 /* See if all predecestors are either throws or clobber only BBs. */
2469 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
2470 if (!(e
->flags
& EDGE_EH
)
2471 && !clobber_only_eh_bb_p (e
->src
, false))
2477 /* Compute function body size parameters for NODE.
2478 When EARLY is true, we compute only simple summaries without
2479 non-trivial predicates to drive the early inliner. */
2482 estimate_function_body_sizes (struct cgraph_node
*node
, bool early
)
2485 /* Estimate static overhead for function prologue/epilogue and alignment. */
2487 /* Benefits are scaled by probability of elimination that is in range
2490 struct function
*my_function
= DECL_STRUCT_FUNCTION (node
->decl
);
2492 struct inline_summary
*info
= inline_summaries
->get (node
);
2493 struct predicate bb_predicate
;
2494 struct ipa_node_params
*parms_info
= NULL
;
2495 vec
<predicate_t
> nonconstant_names
= vNULL
;
2498 predicate array_index
= true_predicate ();
2499 gimple fix_builtin_expect_stmt
;
2504 /* When optimizing and analyzing for IPA inliner, initialize loop optimizer
2505 so we can produce proper inline hints.
2507 When optimizing and analyzing for early inliner, initialize node params
2508 so we can produce correct BB predicates. */
2510 if (opt_for_fn (node
->decl
, optimize
))
2512 calculate_dominance_info (CDI_DOMINATORS
);
2514 loop_optimizer_init (LOOPS_NORMAL
| LOOPS_HAVE_RECORDED_EXITS
);
2517 ipa_check_create_node_params ();
2518 ipa_initialize_node_params (node
);
2521 if (ipa_node_params_sum
)
2523 parms_info
= IPA_NODE_REF (node
);
2524 nonconstant_names
.safe_grow_cleared
2525 (SSANAMES (my_function
)->length ());
2530 fprintf (dump_file
, "\nAnalyzing function body size: %s\n",
2533 /* When we run into maximal number of entries, we assign everything to the
2534 constant truth case. Be sure to have it in list. */
2535 bb_predicate
= true_predicate ();
2536 account_size_time (info
, 0, 0, &bb_predicate
);
2538 bb_predicate
= not_inlined_predicate ();
2539 account_size_time (info
, 2 * INLINE_SIZE_SCALE
, 0, &bb_predicate
);
2541 gcc_assert (my_function
&& my_function
->cfg
);
2543 compute_bb_predicates (node
, parms_info
, info
);
2544 gcc_assert (cfun
== my_function
);
2545 order
= XNEWVEC (int, n_basic_blocks_for_fn (cfun
));
2546 nblocks
= pre_and_rev_post_order_compute (NULL
, order
, false);
2547 for (n
= 0; n
< nblocks
; n
++)
2549 bb
= BASIC_BLOCK_FOR_FN (cfun
, order
[n
]);
2550 freq
= compute_call_stmt_bb_frequency (node
->decl
, bb
);
2551 if (clobber_only_eh_bb_p (bb
))
2553 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2554 fprintf (dump_file
, "\n Ignoring BB %i;"
2555 " it will be optimized away by cleanup_clobbers\n",
2560 /* TODO: Obviously predicates can be propagated down across CFG. */
2564 bb_predicate
= *(struct predicate
*) bb
->aux
;
2566 bb_predicate
= false_predicate ();
2569 bb_predicate
= true_predicate ();
2571 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2573 fprintf (dump_file
, "\n BB %i predicate:", bb
->index
);
2574 dump_predicate (dump_file
, info
->conds
, &bb_predicate
);
2577 if (parms_info
&& nonconstant_names
.exists ())
2579 struct predicate phi_predicate
;
2580 bool first_phi
= true;
2582 for (gphi_iterator bsi
= gsi_start_phis (bb
); !gsi_end_p (bsi
);
2586 && !phi_result_unknown_predicate (parms_info
, info
, bb
,
2591 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2593 fprintf (dump_file
, " ");
2594 print_gimple_stmt (dump_file
, gsi_stmt (bsi
), 0, 0);
2596 predicate_for_phi_result (info
, bsi
.phi (), &phi_predicate
,
2601 fix_builtin_expect_stmt
= find_foldable_builtin_expect (bb
);
2603 for (gimple_stmt_iterator bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
);
2606 gimple stmt
= gsi_stmt (bsi
);
2607 int this_size
= estimate_num_insns (stmt
, &eni_size_weights
);
2608 int this_time
= estimate_num_insns (stmt
, &eni_time_weights
);
2610 struct predicate will_be_nonconstant
;
2612 /* This relation stmt should be folded after we remove
2613 buildin_expect call. Adjust the cost here. */
2614 if (stmt
== fix_builtin_expect_stmt
)
2620 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2622 fprintf (dump_file
, " ");
2623 print_gimple_stmt (dump_file
, stmt
, 0, 0);
2624 fprintf (dump_file
, "\t\tfreq:%3.2f size:%3i time:%3i\n",
2625 ((double) freq
) / CGRAPH_FREQ_BASE
, this_size
,
2629 if (gimple_assign_load_p (stmt
) && nonconstant_names
.exists ())
2631 struct predicate this_array_index
;
2633 array_index_predicate (info
, nonconstant_names
,
2634 gimple_assign_rhs1 (stmt
));
2635 if (!false_predicate_p (&this_array_index
))
2637 and_predicates (info
->conds
, &array_index
,
2640 if (gimple_store_p (stmt
) && nonconstant_names
.exists ())
2642 struct predicate this_array_index
;
2644 array_index_predicate (info
, nonconstant_names
,
2645 gimple_get_lhs (stmt
));
2646 if (!false_predicate_p (&this_array_index
))
2648 and_predicates (info
->conds
, &array_index
,
2653 if (is_gimple_call (stmt
)
2654 && !gimple_call_internal_p (stmt
))
2656 struct cgraph_edge
*edge
= node
->get_edge (stmt
);
2657 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
2659 /* Special case: results of BUILT_IN_CONSTANT_P will be always
2660 resolved as constant. We however don't want to optimize
2661 out the cgraph edges. */
2662 if (nonconstant_names
.exists ()
2663 && gimple_call_builtin_p (stmt
, BUILT_IN_CONSTANT_P
)
2664 && gimple_call_lhs (stmt
)
2665 && TREE_CODE (gimple_call_lhs (stmt
)) == SSA_NAME
)
2667 struct predicate false_p
= false_predicate ();
2668 nonconstant_names
[SSA_NAME_VERSION (gimple_call_lhs (stmt
))]
2671 if (ipa_node_params_sum
)
2673 int count
= gimple_call_num_args (stmt
);
2677 es
->param
.safe_grow_cleared (count
);
2678 for (i
= 0; i
< count
; i
++)
2680 int prob
= param_change_prob (stmt
, i
);
2681 gcc_assert (prob
>= 0 && prob
<= REG_BR_PROB_BASE
);
2682 es
->param
[i
].change_prob
= prob
;
2686 es
->call_stmt_size
= this_size
;
2687 es
->call_stmt_time
= this_time
;
2688 es
->loop_depth
= bb_loop_depth (bb
);
2689 edge_set_predicate (edge
, &bb_predicate
);
2692 /* TODO: When conditional jump or swithc is known to be constant, but
2693 we did not translate it into the predicates, we really can account
2694 just maximum of the possible paths. */
2697 = will_be_nonconstant_predicate (parms_info
, info
,
2698 stmt
, nonconstant_names
);
2699 if (this_time
|| this_size
)
2705 prob
= eliminated_by_inlining_prob (stmt
);
2706 if (prob
== 1 && dump_file
&& (dump_flags
& TDF_DETAILS
))
2708 "\t\t50%% will be eliminated by inlining\n");
2709 if (prob
== 2 && dump_file
&& (dump_flags
& TDF_DETAILS
))
2710 fprintf (dump_file
, "\t\tWill be eliminated by inlining\n");
2713 p
= and_predicates (info
->conds
, &bb_predicate
,
2714 &will_be_nonconstant
);
2716 p
= true_predicate ();
2718 if (!false_predicate_p (&p
)
2719 || (is_gimple_call (stmt
)
2720 && !false_predicate_p (&bb_predicate
)))
2724 if (time
> MAX_TIME
* INLINE_TIME_SCALE
)
2725 time
= MAX_TIME
* INLINE_TIME_SCALE
;
2728 /* We account everything but the calls. Calls have their own
2729 size/time info attached to cgraph edges. This is necessary
2730 in order to make the cost disappear after inlining. */
2731 if (!is_gimple_call (stmt
))
2735 struct predicate ip
= not_inlined_predicate ();
2736 ip
= and_predicates (info
->conds
, &ip
, &p
);
2737 account_size_time (info
, this_size
* prob
,
2738 this_time
* prob
, &ip
);
2741 account_size_time (info
, this_size
* (2 - prob
),
2742 this_time
* (2 - prob
), &p
);
2745 gcc_assert (time
>= 0);
2746 gcc_assert (size
>= 0);
2750 set_hint_predicate (&inline_summaries
->get (node
)->array_index
, array_index
);
2751 time
= (time
+ CGRAPH_FREQ_BASE
/ 2) / CGRAPH_FREQ_BASE
;
2752 if (time
> MAX_TIME
)
2756 if (nonconstant_names
.exists () && !early
)
2759 predicate loop_iterations
= true_predicate ();
2760 predicate loop_stride
= true_predicate ();
2762 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2763 flow_loops_dump (dump_file
, NULL
, 0);
2765 FOR_EACH_LOOP (loop
, 0)
2770 struct tree_niter_desc niter_desc
;
2771 basic_block
*body
= get_loop_body (loop
);
2772 bb_predicate
= *(struct predicate
*) loop
->header
->aux
;
2774 exits
= get_loop_exit_edges (loop
);
2775 FOR_EACH_VEC_ELT (exits
, j
, ex
)
2776 if (number_of_iterations_exit (loop
, ex
, &niter_desc
, false)
2777 && !is_gimple_min_invariant (niter_desc
.niter
))
2779 predicate will_be_nonconstant
2780 = will_be_nonconstant_expr_predicate (parms_info
, info
,
2783 if (!true_predicate_p (&will_be_nonconstant
))
2784 will_be_nonconstant
= and_predicates (info
->conds
,
2786 &will_be_nonconstant
);
2787 if (!true_predicate_p (&will_be_nonconstant
)
2788 && !false_predicate_p (&will_be_nonconstant
))
2789 /* This is slightly inprecise. We may want to represent each
2790 loop with independent predicate. */
2792 and_predicates (info
->conds
, &loop_iterations
,
2793 &will_be_nonconstant
);
2797 for (i
= 0; i
< loop
->num_nodes
; i
++)
2799 gimple_stmt_iterator gsi
;
2800 bb_predicate
= *(struct predicate
*) body
[i
]->aux
;
2801 for (gsi
= gsi_start_bb (body
[i
]); !gsi_end_p (gsi
);
2804 gimple stmt
= gsi_stmt (gsi
);
2809 FOR_EACH_SSA_TREE_OPERAND (use
, stmt
, iter
, SSA_OP_USE
)
2811 predicate will_be_nonconstant
;
2814 (loop
, loop_containing_stmt (stmt
), use
, &iv
, true)
2815 || is_gimple_min_invariant (iv
.step
))
2818 = will_be_nonconstant_expr_predicate (parms_info
, info
,
2821 if (!true_predicate_p (&will_be_nonconstant
))
2823 = and_predicates (info
->conds
,
2825 &will_be_nonconstant
);
2826 if (!true_predicate_p (&will_be_nonconstant
)
2827 && !false_predicate_p (&will_be_nonconstant
))
2828 /* This is slightly inprecise. We may want to represent
2829 each loop with independent predicate. */
2831 and_predicates (info
->conds
, &loop_stride
,
2832 &will_be_nonconstant
);
2838 set_hint_predicate (&inline_summaries
->get (node
)->loop_iterations
,
2840 set_hint_predicate (&inline_summaries
->get (node
)->loop_stride
, loop_stride
);
2843 FOR_ALL_BB_FN (bb
, my_function
)
2849 pool_free (edge_predicate_pool
, bb
->aux
);
2851 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
2854 pool_free (edge_predicate_pool
, e
->aux
);
2858 inline_summaries
->get (node
)->self_time
= time
;
2859 inline_summaries
->get (node
)->self_size
= size
;
2860 nonconstant_names
.release ();
2861 if (opt_for_fn (node
->decl
, optimize
))
2864 loop_optimizer_finalize ();
2865 else if (!ipa_edge_args_vector
)
2866 ipa_free_all_node_params ();
2867 free_dominance_info (CDI_DOMINATORS
);
2871 fprintf (dump_file
, "\n");
2872 dump_inline_summary (dump_file
, node
);
2877 /* Compute parameters of functions used by inliner.
2878 EARLY is true when we compute parameters for the early inliner */
2881 compute_inline_parameters (struct cgraph_node
*node
, bool early
)
2883 HOST_WIDE_INT self_stack_size
;
2884 struct cgraph_edge
*e
;
2885 struct inline_summary
*info
;
2887 gcc_assert (!node
->global
.inlined_to
);
2889 inline_summary_alloc ();
2891 info
= inline_summaries
->get (node
);
2892 reset_inline_summary (node
, info
);
2894 /* FIXME: Thunks are inlinable, but tree-inline don't know how to do that.
2895 Once this happen, we will need to more curefully predict call
2897 if (node
->thunk
.thunk_p
)
2899 struct inline_edge_summary
*es
= inline_edge_summary (node
->callees
);
2900 struct predicate t
= true_predicate ();
2902 info
->inlinable
= 0;
2903 node
->callees
->call_stmt_cannot_inline_p
= true;
2904 node
->local
.can_change_signature
= false;
2905 es
->call_stmt_time
= 1;
2906 es
->call_stmt_size
= 1;
2907 account_size_time (info
, 0, 0, &t
);
2911 /* Even is_gimple_min_invariant rely on current_function_decl. */
2912 push_cfun (DECL_STRUCT_FUNCTION (node
->decl
));
2914 /* Estimate the stack size for the function if we're optimizing. */
2915 self_stack_size
= optimize
? estimated_stack_frame_size (node
) : 0;
2916 info
->estimated_self_stack_size
= self_stack_size
;
2917 info
->estimated_stack_size
= self_stack_size
;
2918 info
->stack_frame_offset
= 0;
2920 /* Can this function be inlined at all? */
2921 if (!opt_for_fn (node
->decl
, optimize
)
2922 && !lookup_attribute ("always_inline",
2923 DECL_ATTRIBUTES (node
->decl
)))
2924 info
->inlinable
= false;
2926 info
->inlinable
= tree_inlinable_function_p (node
->decl
);
2928 /* Type attributes can use parameter indices to describe them. */
2929 if (TYPE_ATTRIBUTES (TREE_TYPE (node
->decl
)))
2930 node
->local
.can_change_signature
= false;
2933 /* Otherwise, inlinable functions always can change signature. */
2934 if (info
->inlinable
)
2935 node
->local
.can_change_signature
= true;
2938 /* Functions calling builtin_apply can not change signature. */
2939 for (e
= node
->callees
; e
; e
= e
->next_callee
)
2941 tree
cdecl = e
->callee
->decl
;
2942 if (DECL_BUILT_IN (cdecl)
2943 && DECL_BUILT_IN_CLASS (cdecl) == BUILT_IN_NORMAL
2944 && (DECL_FUNCTION_CODE (cdecl) == BUILT_IN_APPLY_ARGS
2945 || DECL_FUNCTION_CODE (cdecl) == BUILT_IN_VA_START
))
2948 node
->local
.can_change_signature
= !e
;
2951 estimate_function_body_sizes (node
, early
);
2953 for (e
= node
->callees
; e
; e
= e
->next_callee
)
2954 if (e
->callee
->comdat_local_p ())
2956 node
->calls_comdat_local
= (e
!= NULL
);
2958 /* Inlining characteristics are maintained by the cgraph_mark_inline. */
2959 info
->time
= info
->self_time
;
2960 info
->size
= info
->self_size
;
2961 info
->stack_frame_offset
= 0;
2962 info
->estimated_stack_size
= info
->estimated_self_stack_size
;
2963 #ifdef ENABLE_CHECKING
2964 inline_update_overall_summary (node
);
2965 gcc_assert (info
->time
== info
->self_time
&& info
->size
== info
->self_size
);
2972 /* Compute parameters of functions used by inliner using
2973 current_function_decl. */
2976 compute_inline_parameters_for_current (void)
2978 compute_inline_parameters (cgraph_node::get (current_function_decl
), true);
2984 const pass_data pass_data_inline_parameters
=
2986 GIMPLE_PASS
, /* type */
2987 "inline_param", /* name */
2988 OPTGROUP_INLINE
, /* optinfo_flags */
2989 TV_INLINE_PARAMETERS
, /* tv_id */
2990 0, /* properties_required */
2991 0, /* properties_provided */
2992 0, /* properties_destroyed */
2993 0, /* todo_flags_start */
2994 0, /* todo_flags_finish */
2997 class pass_inline_parameters
: public gimple_opt_pass
3000 pass_inline_parameters (gcc::context
*ctxt
)
3001 : gimple_opt_pass (pass_data_inline_parameters
, ctxt
)
3004 /* opt_pass methods: */
3005 opt_pass
* clone () { return new pass_inline_parameters (m_ctxt
); }
3006 virtual unsigned int execute (function
*)
3008 return compute_inline_parameters_for_current ();
3011 }; // class pass_inline_parameters
3016 make_pass_inline_parameters (gcc::context
*ctxt
)
3018 return new pass_inline_parameters (ctxt
);
3022 /* Estimate benefit devirtualizing indirect edge IE, provided KNOWN_VALS,
3023 KNOWN_CONTEXTS and KNOWN_AGGS. */
3026 estimate_edge_devirt_benefit (struct cgraph_edge
*ie
,
3027 int *size
, int *time
,
3028 vec
<tree
> known_vals
,
3029 vec
<ipa_polymorphic_call_context
> known_contexts
,
3030 vec
<ipa_agg_jump_function_p
> known_aggs
)
3033 struct cgraph_node
*callee
;
3034 struct inline_summary
*isummary
;
3035 enum availability avail
;
3038 if (!known_vals
.exists () && !known_contexts
.exists ())
3040 if (!opt_for_fn (ie
->caller
->decl
, flag_indirect_inlining
))
3043 target
= ipa_get_indirect_edge_target (ie
, known_vals
, known_contexts
,
3044 known_aggs
, &speculative
);
3045 if (!target
|| speculative
)
3048 /* Account for difference in cost between indirect and direct calls. */
3049 *size
-= (eni_size_weights
.indirect_call_cost
- eni_size_weights
.call_cost
);
3050 *time
-= (eni_time_weights
.indirect_call_cost
- eni_time_weights
.call_cost
);
3051 gcc_checking_assert (*time
>= 0);
3052 gcc_checking_assert (*size
>= 0);
3054 callee
= cgraph_node::get (target
);
3055 if (!callee
|| !callee
->definition
)
3057 callee
= callee
->function_symbol (&avail
);
3058 if (avail
< AVAIL_AVAILABLE
)
3060 isummary
= inline_summaries
->get (callee
);
3061 return isummary
->inlinable
;
3064 /* Increase SIZE, MIN_SIZE (if non-NULL) and TIME for size and time needed to
3065 handle edge E with probability PROB.
3066 Set HINTS if edge may be devirtualized.
3067 KNOWN_VALS, KNOWN_AGGS and KNOWN_CONTEXTS describe context of the call
3071 estimate_edge_size_and_time (struct cgraph_edge
*e
, int *size
, int *min_size
,
3074 vec
<tree
> known_vals
,
3075 vec
<ipa_polymorphic_call_context
> known_contexts
,
3076 vec
<ipa_agg_jump_function_p
> known_aggs
,
3077 inline_hints
*hints
)
3079 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3080 int call_size
= es
->call_stmt_size
;
3081 int call_time
= es
->call_stmt_time
;
3084 && estimate_edge_devirt_benefit (e
, &call_size
, &call_time
,
3085 known_vals
, known_contexts
, known_aggs
)
3086 && hints
&& e
->maybe_hot_p ())
3087 *hints
|= INLINE_HINT_indirect_call
;
3088 cur_size
= call_size
* INLINE_SIZE_SCALE
;
3091 *min_size
+= cur_size
;
3092 *time
+= apply_probability ((gcov_type
) call_time
, prob
)
3093 * e
->frequency
* (INLINE_TIME_SCALE
/ CGRAPH_FREQ_BASE
);
3094 if (*time
> MAX_TIME
* INLINE_TIME_SCALE
)
3095 *time
= MAX_TIME
* INLINE_TIME_SCALE
;
3100 /* Increase SIZE, MIN_SIZE and TIME for size and time needed to handle all
3101 calls in NODE. POSSIBLE_TRUTHS, KNOWN_VALS, KNOWN_AGGS and KNOWN_CONTEXTS
3102 describe context of the call site. */
3105 estimate_calls_size_and_time (struct cgraph_node
*node
, int *size
,
3106 int *min_size
, int *time
,
3107 inline_hints
*hints
,
3108 clause_t possible_truths
,
3109 vec
<tree
> known_vals
,
3110 vec
<ipa_polymorphic_call_context
> known_contexts
,
3111 vec
<ipa_agg_jump_function_p
> known_aggs
)
3113 struct cgraph_edge
*e
;
3114 for (e
= node
->callees
; e
; e
= e
->next_callee
)
3116 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3118 /* Do not care about zero sized builtins. */
3119 if (e
->inline_failed
&& !es
->call_stmt_size
)
3121 gcc_checking_assert (!es
->call_stmt_time
);
3125 || evaluate_predicate (es
->predicate
, possible_truths
))
3127 if (e
->inline_failed
)
3129 /* Predicates of calls shall not use NOT_CHANGED codes,
3130 sowe do not need to compute probabilities. */
3131 estimate_edge_size_and_time (e
, size
,
3132 es
->predicate
? NULL
: min_size
,
3133 time
, REG_BR_PROB_BASE
,
3134 known_vals
, known_contexts
,
3138 estimate_calls_size_and_time (e
->callee
, size
, min_size
, time
,
3141 known_vals
, known_contexts
,
3145 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
3147 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3149 || evaluate_predicate (es
->predicate
, possible_truths
))
3150 estimate_edge_size_and_time (e
, size
,
3151 es
->predicate
? NULL
: min_size
,
3152 time
, REG_BR_PROB_BASE
,
3153 known_vals
, known_contexts
, known_aggs
,
3159 /* Estimate size and time needed to execute NODE assuming
3160 POSSIBLE_TRUTHS clause, and KNOWN_VALS, KNOWN_AGGS and KNOWN_CONTEXTS
3161 information about NODE's arguments. If non-NULL use also probability
3162 information present in INLINE_PARAM_SUMMARY vector.
3163 Additionally detemine hints determined by the context. Finally compute
3164 minimal size needed for the call that is independent on the call context and
3165 can be used for fast estimates. Return the values in RET_SIZE,
3166 RET_MIN_SIZE, RET_TIME and RET_HINTS. */
3169 estimate_node_size_and_time (struct cgraph_node
*node
,
3170 clause_t possible_truths
,
3171 vec
<tree
> known_vals
,
3172 vec
<ipa_polymorphic_call_context
> known_contexts
,
3173 vec
<ipa_agg_jump_function_p
> known_aggs
,
3174 int *ret_size
, int *ret_min_size
, int *ret_time
,
3175 inline_hints
*ret_hints
,
3176 vec
<inline_param_summary
>
3177 inline_param_summary
)
3179 struct inline_summary
*info
= inline_summaries
->get (node
);
3184 inline_hints hints
= 0;
3187 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3190 fprintf (dump_file
, " Estimating body: %s/%i\n"
3191 " Known to be false: ", node
->name (),
3194 for (i
= predicate_not_inlined_condition
;
3195 i
< (predicate_first_dynamic_condition
3196 + (int) vec_safe_length (info
->conds
)); i
++)
3197 if (!(possible_truths
& (1 << i
)))
3200 fprintf (dump_file
, ", ");
3202 dump_condition (dump_file
, info
->conds
, i
);
3206 for (i
= 0; vec_safe_iterate (info
->entry
, i
, &e
); i
++)
3207 if (evaluate_predicate (&e
->predicate
, possible_truths
))
3210 gcc_checking_assert (e
->time
>= 0);
3211 gcc_checking_assert (time
>= 0);
3212 if (!inline_param_summary
.exists ())
3216 int prob
= predicate_probability (info
->conds
,
3219 inline_param_summary
);
3220 gcc_checking_assert (prob
>= 0);
3221 gcc_checking_assert (prob
<= REG_BR_PROB_BASE
);
3222 time
+= apply_probability ((gcov_type
) e
->time
, prob
);
3224 if (time
> MAX_TIME
* INLINE_TIME_SCALE
)
3225 time
= MAX_TIME
* INLINE_TIME_SCALE
;
3226 gcc_checking_assert (time
>= 0);
3229 gcc_checking_assert (true_predicate_p (&(*info
->entry
)[0].predicate
));
3230 min_size
= (*info
->entry
)[0].size
;
3231 gcc_checking_assert (size
>= 0);
3232 gcc_checking_assert (time
>= 0);
3234 if (info
->loop_iterations
3235 && !evaluate_predicate (info
->loop_iterations
, possible_truths
))
3236 hints
|= INLINE_HINT_loop_iterations
;
3237 if (info
->loop_stride
3238 && !evaluate_predicate (info
->loop_stride
, possible_truths
))
3239 hints
|= INLINE_HINT_loop_stride
;
3240 if (info
->array_index
3241 && !evaluate_predicate (info
->array_index
, possible_truths
))
3242 hints
|= INLINE_HINT_array_index
;
3244 hints
|= INLINE_HINT_in_scc
;
3245 if (DECL_DECLARED_INLINE_P (node
->decl
))
3246 hints
|= INLINE_HINT_declared_inline
;
3248 estimate_calls_size_and_time (node
, &size
, &min_size
, &time
, &hints
, possible_truths
,
3249 known_vals
, known_contexts
, known_aggs
);
3250 gcc_checking_assert (size
>= 0);
3251 gcc_checking_assert (time
>= 0);
3252 time
= RDIV (time
, INLINE_TIME_SCALE
);
3253 size
= RDIV (size
, INLINE_SIZE_SCALE
);
3254 min_size
= RDIV (min_size
, INLINE_SIZE_SCALE
);
3256 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3257 fprintf (dump_file
, "\n size:%i time:%i\n", (int) size
, (int) time
);
3263 *ret_min_size
= min_size
;
3270 /* Estimate size and time needed to execute callee of EDGE assuming that
3271 parameters known to be constant at caller of EDGE are propagated.
3272 KNOWN_VALS and KNOWN_CONTEXTS are vectors of assumed known constant values
3273 and types for parameters. */
3276 estimate_ipcp_clone_size_and_time (struct cgraph_node
*node
,
3277 vec
<tree
> known_vals
,
3278 vec
<ipa_polymorphic_call_context
>
3280 vec
<ipa_agg_jump_function_p
> known_aggs
,
3281 int *ret_size
, int *ret_time
,
3282 inline_hints
*hints
)
3286 clause
= evaluate_conditions_for_known_args (node
, false, known_vals
,
3288 estimate_node_size_and_time (node
, clause
, known_vals
, known_contexts
,
3289 known_aggs
, ret_size
, NULL
, ret_time
, hints
, vNULL
);
3292 /* Translate all conditions from callee representation into caller
3293 representation and symbolically evaluate predicate P into new predicate.
3295 INFO is inline_summary of function we are adding predicate into, CALLEE_INFO
3296 is summary of function predicate P is from. OPERAND_MAP is array giving
3297 callee formal IDs the caller formal IDs. POSSSIBLE_TRUTHS is clausule of all
3298 callee conditions that may be true in caller context. TOPLEV_PREDICATE is
3299 predicate under which callee is executed. OFFSET_MAP is an array of of
3300 offsets that need to be added to conditions, negative offset means that
3301 conditions relying on values passed by reference have to be discarded
3302 because they might not be preserved (and should be considered offset zero
3303 for other purposes). */
3305 static struct predicate
3306 remap_predicate (struct inline_summary
*info
,
3307 struct inline_summary
*callee_info
,
3308 struct predicate
*p
,
3309 vec
<int> operand_map
,
3310 vec
<int> offset_map
,
3311 clause_t possible_truths
, struct predicate
*toplev_predicate
)
3314 struct predicate out
= true_predicate ();
3316 /* True predicate is easy. */
3317 if (true_predicate_p (p
))
3318 return *toplev_predicate
;
3319 for (i
= 0; p
->clause
[i
]; i
++)
3321 clause_t clause
= p
->clause
[i
];
3323 struct predicate clause_predicate
= false_predicate ();
3325 gcc_assert (i
< MAX_CLAUSES
);
3327 for (cond
= 0; cond
< NUM_CONDITIONS
; cond
++)
3328 /* Do we have condition we can't disprove? */
3329 if (clause
& possible_truths
& (1 << cond
))
3331 struct predicate cond_predicate
;
3332 /* Work out if the condition can translate to predicate in the
3333 inlined function. */
3334 if (cond
>= predicate_first_dynamic_condition
)
3336 struct condition
*c
;
3338 c
= &(*callee_info
->conds
)[cond
3340 predicate_first_dynamic_condition
];
3341 /* See if we can remap condition operand to caller's operand.
3342 Otherwise give up. */
3343 if (!operand_map
.exists ()
3344 || (int) operand_map
.length () <= c
->operand_num
3345 || operand_map
[c
->operand_num
] == -1
3346 /* TODO: For non-aggregate conditions, adding an offset is
3347 basically an arithmetic jump function processing which
3348 we should support in future. */
3349 || ((!c
->agg_contents
|| !c
->by_ref
)
3350 && offset_map
[c
->operand_num
] > 0)
3351 || (c
->agg_contents
&& c
->by_ref
3352 && offset_map
[c
->operand_num
] < 0))
3353 cond_predicate
= true_predicate ();
3356 struct agg_position_info ap
;
3357 HOST_WIDE_INT offset_delta
= offset_map
[c
->operand_num
];
3358 if (offset_delta
< 0)
3360 gcc_checking_assert (!c
->agg_contents
|| !c
->by_ref
);
3363 gcc_assert (!c
->agg_contents
3364 || c
->by_ref
|| offset_delta
== 0);
3365 ap
.offset
= c
->offset
+ offset_delta
;
3366 ap
.agg_contents
= c
->agg_contents
;
3367 ap
.by_ref
= c
->by_ref
;
3368 cond_predicate
= add_condition (info
,
3369 operand_map
[c
->operand_num
],
3370 &ap
, c
->code
, c
->val
);
3373 /* Fixed conditions remains same, construct single
3374 condition predicate. */
3377 cond_predicate
.clause
[0] = 1 << cond
;
3378 cond_predicate
.clause
[1] = 0;
3380 clause_predicate
= or_predicates (info
->conds
, &clause_predicate
,
3383 out
= and_predicates (info
->conds
, &out
, &clause_predicate
);
3385 return and_predicates (info
->conds
, &out
, toplev_predicate
);
3389 /* Update summary information of inline clones after inlining.
3390 Compute peak stack usage. */
3393 inline_update_callee_summaries (struct cgraph_node
*node
, int depth
)
3395 struct cgraph_edge
*e
;
3396 struct inline_summary
*callee_info
= inline_summaries
->get (node
);
3397 struct inline_summary
*caller_info
= inline_summaries
->get (node
->callers
->caller
);
3400 callee_info
->stack_frame_offset
3401 = caller_info
->stack_frame_offset
3402 + caller_info
->estimated_self_stack_size
;
3403 peak
= callee_info
->stack_frame_offset
3404 + callee_info
->estimated_self_stack_size
;
3405 if (inline_summaries
->get (node
->global
.inlined_to
)->estimated_stack_size
< peak
)
3406 inline_summaries
->get (node
->global
.inlined_to
)->estimated_stack_size
= peak
;
3407 ipa_propagate_frequency (node
);
3408 for (e
= node
->callees
; e
; e
= e
->next_callee
)
3410 if (!e
->inline_failed
)
3411 inline_update_callee_summaries (e
->callee
, depth
);
3412 inline_edge_summary (e
)->loop_depth
+= depth
;
3414 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
3415 inline_edge_summary (e
)->loop_depth
+= depth
;
3418 /* Update change_prob of EDGE after INLINED_EDGE has been inlined.
3419 When functoin A is inlined in B and A calls C with parameter that
3420 changes with probability PROB1 and C is known to be passthroug
3421 of argument if B that change with probability PROB2, the probability
3422 of change is now PROB1*PROB2. */
3425 remap_edge_change_prob (struct cgraph_edge
*inlined_edge
,
3426 struct cgraph_edge
*edge
)
3428 if (ipa_node_params_sum
)
3431 struct ipa_edge_args
*args
= IPA_EDGE_REF (edge
);
3432 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3433 struct inline_edge_summary
*inlined_es
3434 = inline_edge_summary (inlined_edge
);
3436 for (i
= 0; i
< ipa_get_cs_argument_count (args
); i
++)
3438 struct ipa_jump_func
*jfunc
= ipa_get_ith_jump_func (args
, i
);
3439 if (jfunc
->type
== IPA_JF_PASS_THROUGH
3440 && (ipa_get_jf_pass_through_formal_id (jfunc
)
3441 < (int) inlined_es
->param
.length ()))
3443 int jf_formal_id
= ipa_get_jf_pass_through_formal_id (jfunc
);
3444 int prob1
= es
->param
[i
].change_prob
;
3445 int prob2
= inlined_es
->param
[jf_formal_id
].change_prob
;
3446 int prob
= combine_probabilities (prob1
, prob2
);
3448 if (prob1
&& prob2
&& !prob
)
3451 es
->param
[i
].change_prob
= prob
;
3457 /* Update edge summaries of NODE after INLINED_EDGE has been inlined.
3459 Remap predicates of callees of NODE. Rest of arguments match
3462 Also update change probabilities. */
3465 remap_edge_summaries (struct cgraph_edge
*inlined_edge
,
3466 struct cgraph_node
*node
,
3467 struct inline_summary
*info
,
3468 struct inline_summary
*callee_info
,
3469 vec
<int> operand_map
,
3470 vec
<int> offset_map
,
3471 clause_t possible_truths
,
3472 struct predicate
*toplev_predicate
)
3474 struct cgraph_edge
*e
;
3475 for (e
= node
->callees
; e
; e
= e
->next_callee
)
3477 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3480 if (e
->inline_failed
)
3482 remap_edge_change_prob (inlined_edge
, e
);
3486 p
= remap_predicate (info
, callee_info
,
3487 es
->predicate
, operand_map
, offset_map
,
3488 possible_truths
, toplev_predicate
);
3489 edge_set_predicate (e
, &p
);
3490 /* TODO: We should remove the edge for code that will be
3491 optimized out, but we need to keep verifiers and tree-inline
3492 happy. Make it cold for now. */
3493 if (false_predicate_p (&p
))
3500 edge_set_predicate (e
, toplev_predicate
);
3503 remap_edge_summaries (inlined_edge
, e
->callee
, info
, callee_info
,
3504 operand_map
, offset_map
, possible_truths
,
3507 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
3509 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3512 remap_edge_change_prob (inlined_edge
, e
);
3515 p
= remap_predicate (info
, callee_info
,
3516 es
->predicate
, operand_map
, offset_map
,
3517 possible_truths
, toplev_predicate
);
3518 edge_set_predicate (e
, &p
);
3519 /* TODO: We should remove the edge for code that will be optimized
3520 out, but we need to keep verifiers and tree-inline happy.
3521 Make it cold for now. */
3522 if (false_predicate_p (&p
))
3529 edge_set_predicate (e
, toplev_predicate
);
3533 /* Same as remap_predicate, but set result into hint *HINT. */
3536 remap_hint_predicate (struct inline_summary
*info
,
3537 struct inline_summary
*callee_info
,
3538 struct predicate
**hint
,
3539 vec
<int> operand_map
,
3540 vec
<int> offset_map
,
3541 clause_t possible_truths
,
3542 struct predicate
*toplev_predicate
)
3548 p
= remap_predicate (info
, callee_info
,
3550 operand_map
, offset_map
,
3551 possible_truths
, toplev_predicate
);
3552 if (!false_predicate_p (&p
) && !true_predicate_p (&p
))
3555 set_hint_predicate (hint
, p
);
3557 **hint
= and_predicates (info
->conds
, *hint
, &p
);
3561 /* We inlined EDGE. Update summary of the function we inlined into. */
3564 inline_merge_summary (struct cgraph_edge
*edge
)
3566 struct inline_summary
*callee_info
= inline_summaries
->get (edge
->callee
);
3567 struct cgraph_node
*to
= (edge
->caller
->global
.inlined_to
3568 ? edge
->caller
->global
.inlined_to
: edge
->caller
);
3569 struct inline_summary
*info
= inline_summaries
->get (to
);
3570 clause_t clause
= 0; /* not_inline is known to be false. */
3572 vec
<int> operand_map
= vNULL
;
3573 vec
<int> offset_map
= vNULL
;
3575 struct predicate toplev_predicate
;
3576 struct predicate true_p
= true_predicate ();
3577 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3580 toplev_predicate
= *es
->predicate
;
3582 toplev_predicate
= true_predicate ();
3584 if (callee_info
->conds
)
3585 evaluate_properties_for_edge (edge
, true, &clause
, NULL
, NULL
, NULL
);
3586 if (ipa_node_params_sum
&& callee_info
->conds
)
3588 struct ipa_edge_args
*args
= IPA_EDGE_REF (edge
);
3589 int count
= ipa_get_cs_argument_count (args
);
3594 operand_map
.safe_grow_cleared (count
);
3595 offset_map
.safe_grow_cleared (count
);
3597 for (i
= 0; i
< count
; i
++)
3599 struct ipa_jump_func
*jfunc
= ipa_get_ith_jump_func (args
, i
);
3602 /* TODO: handle non-NOPs when merging. */
3603 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
3605 if (ipa_get_jf_pass_through_operation (jfunc
) == NOP_EXPR
)
3606 map
= ipa_get_jf_pass_through_formal_id (jfunc
);
3607 if (!ipa_get_jf_pass_through_agg_preserved (jfunc
))
3610 else if (jfunc
->type
== IPA_JF_ANCESTOR
)
3612 HOST_WIDE_INT offset
= ipa_get_jf_ancestor_offset (jfunc
);
3613 if (offset
>= 0 && offset
< INT_MAX
)
3615 map
= ipa_get_jf_ancestor_formal_id (jfunc
);
3616 if (!ipa_get_jf_ancestor_agg_preserved (jfunc
))
3618 offset_map
[i
] = offset
;
3621 operand_map
[i
] = map
;
3622 gcc_assert (map
< ipa_get_param_count (IPA_NODE_REF (to
)));
3625 for (i
= 0; vec_safe_iterate (callee_info
->entry
, i
, &e
); i
++)
3627 struct predicate p
= remap_predicate (info
, callee_info
,
3628 &e
->predicate
, operand_map
,
3631 if (!false_predicate_p (&p
))
3633 gcov_type add_time
= ((gcov_type
) e
->time
* edge
->frequency
3634 + CGRAPH_FREQ_BASE
/ 2) / CGRAPH_FREQ_BASE
;
3635 int prob
= predicate_probability (callee_info
->conds
,
3638 add_time
= apply_probability ((gcov_type
) add_time
, prob
);
3639 if (add_time
> MAX_TIME
* INLINE_TIME_SCALE
)
3640 add_time
= MAX_TIME
* INLINE_TIME_SCALE
;
3641 if (prob
!= REG_BR_PROB_BASE
3642 && dump_file
&& (dump_flags
& TDF_DETAILS
))
3644 fprintf (dump_file
, "\t\tScaling time by probability:%f\n",
3645 (double) prob
/ REG_BR_PROB_BASE
);
3647 account_size_time (info
, e
->size
, add_time
, &p
);
3650 remap_edge_summaries (edge
, edge
->callee
, info
, callee_info
, operand_map
,
3651 offset_map
, clause
, &toplev_predicate
);
3652 remap_hint_predicate (info
, callee_info
,
3653 &callee_info
->loop_iterations
,
3654 operand_map
, offset_map
, clause
, &toplev_predicate
);
3655 remap_hint_predicate (info
, callee_info
,
3656 &callee_info
->loop_stride
,
3657 operand_map
, offset_map
, clause
, &toplev_predicate
);
3658 remap_hint_predicate (info
, callee_info
,
3659 &callee_info
->array_index
,
3660 operand_map
, offset_map
, clause
, &toplev_predicate
);
3662 inline_update_callee_summaries (edge
->callee
,
3663 inline_edge_summary (edge
)->loop_depth
);
3665 /* We do not maintain predicates of inlined edges, free it. */
3666 edge_set_predicate (edge
, &true_p
);
3667 /* Similarly remove param summaries. */
3668 es
->param
.release ();
3669 operand_map
.release ();
3670 offset_map
.release ();
3673 /* For performance reasons inline_merge_summary is not updating overall size
3674 and time. Recompute it. */
3677 inline_update_overall_summary (struct cgraph_node
*node
)
3679 struct inline_summary
*info
= inline_summaries
->get (node
);
3685 for (i
= 0; vec_safe_iterate (info
->entry
, i
, &e
); i
++)
3687 info
->size
+= e
->size
, info
->time
+= e
->time
;
3688 if (info
->time
> MAX_TIME
* INLINE_TIME_SCALE
)
3689 info
->time
= MAX_TIME
* INLINE_TIME_SCALE
;
3691 estimate_calls_size_and_time (node
, &info
->size
, &info
->min_size
,
3693 ~(clause_t
) (1 << predicate_false_condition
),
3694 vNULL
, vNULL
, vNULL
);
3695 info
->time
= (info
->time
+ INLINE_TIME_SCALE
/ 2) / INLINE_TIME_SCALE
;
3696 info
->size
= (info
->size
+ INLINE_SIZE_SCALE
/ 2) / INLINE_SIZE_SCALE
;
3699 /* Return hints derrived from EDGE. */
3701 simple_edge_hints (struct cgraph_edge
*edge
)
3704 struct cgraph_node
*to
= (edge
->caller
->global
.inlined_to
3705 ? edge
->caller
->global
.inlined_to
: edge
->caller
);
3706 struct cgraph_node
*callee
= edge
->callee
->ultimate_alias_target ();
3707 if (inline_summaries
->get (to
)->scc_no
3708 && inline_summaries
->get (to
)->scc_no
3709 == inline_summaries
->get (callee
)->scc_no
3710 && !edge
->recursive_p ())
3711 hints
|= INLINE_HINT_same_scc
;
3713 if (callee
->lto_file_data
&& edge
->caller
->lto_file_data
3714 && edge
->caller
->lto_file_data
!= callee
->lto_file_data
3716 hints
|= INLINE_HINT_cross_module
;
3721 /* Estimate the time cost for the caller when inlining EDGE.
3722 Only to be called via estimate_edge_time, that handles the
3725 When caching, also update the cache entry. Compute both time and
3726 size, since we always need both metrics eventually. */
3729 do_estimate_edge_time (struct cgraph_edge
*edge
)
3734 struct cgraph_node
*callee
;
3736 vec
<tree
> known_vals
;
3737 vec
<ipa_polymorphic_call_context
> known_contexts
;
3738 vec
<ipa_agg_jump_function_p
> known_aggs
;
3739 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3742 callee
= edge
->callee
->ultimate_alias_target ();
3744 gcc_checking_assert (edge
->inline_failed
);
3745 evaluate_properties_for_edge (edge
, true,
3746 &clause
, &known_vals
, &known_contexts
,
3748 estimate_node_size_and_time (callee
, clause
, known_vals
, known_contexts
,
3749 known_aggs
, &size
, &min_size
, &time
, &hints
, es
->param
);
3751 /* When we have profile feedback, we can quite safely identify hot
3752 edges and for those we disable size limits. Don't do that when
3753 probability that caller will call the callee is low however, since it
3754 may hurt optimization of the caller's hot path. */
3755 if (edge
->count
&& edge
->maybe_hot_p ()
3757 > (edge
->caller
->global
.inlined_to
3758 ? edge
->caller
->global
.inlined_to
->count
: edge
->caller
->count
)))
3759 hints
|= INLINE_HINT_known_hot
;
3761 known_vals
.release ();
3762 known_contexts
.release ();
3763 known_aggs
.release ();
3764 gcc_checking_assert (size
>= 0);
3765 gcc_checking_assert (time
>= 0);
3767 /* When caching, update the cache entry. */
3768 if (edge_growth_cache
.exists ())
3770 inline_summaries
->get (edge
->callee
)->min_size
= min_size
;
3771 if ((int) edge_growth_cache
.length () <= edge
->uid
)
3772 edge_growth_cache
.safe_grow_cleared (symtab
->edges_max_uid
);
3773 edge_growth_cache
[edge
->uid
].time
= time
+ (time
>= 0);
3775 edge_growth_cache
[edge
->uid
].size
= size
+ (size
>= 0);
3776 hints
|= simple_edge_hints (edge
);
3777 edge_growth_cache
[edge
->uid
].hints
= hints
+ 1;
3783 /* Return estimated callee growth after inlining EDGE.
3784 Only to be called via estimate_edge_size. */
3787 do_estimate_edge_size (struct cgraph_edge
*edge
)
3790 struct cgraph_node
*callee
;
3792 vec
<tree
> known_vals
;
3793 vec
<ipa_polymorphic_call_context
> known_contexts
;
3794 vec
<ipa_agg_jump_function_p
> known_aggs
;
3796 /* When we do caching, use do_estimate_edge_time to populate the entry. */
3798 if (edge_growth_cache
.exists ())
3800 do_estimate_edge_time (edge
);
3801 size
= edge_growth_cache
[edge
->uid
].size
;
3802 gcc_checking_assert (size
);
3803 return size
- (size
> 0);
3806 callee
= edge
->callee
->ultimate_alias_target ();
3808 /* Early inliner runs without caching, go ahead and do the dirty work. */
3809 gcc_checking_assert (edge
->inline_failed
);
3810 evaluate_properties_for_edge (edge
, true,
3811 &clause
, &known_vals
, &known_contexts
,
3813 estimate_node_size_and_time (callee
, clause
, known_vals
, known_contexts
,
3814 known_aggs
, &size
, NULL
, NULL
, NULL
, vNULL
);
3815 known_vals
.release ();
3816 known_contexts
.release ();
3817 known_aggs
.release ();
3822 /* Estimate the growth of the caller when inlining EDGE.
3823 Only to be called via estimate_edge_size. */
3826 do_estimate_edge_hints (struct cgraph_edge
*edge
)
3829 struct cgraph_node
*callee
;
3831 vec
<tree
> known_vals
;
3832 vec
<ipa_polymorphic_call_context
> known_contexts
;
3833 vec
<ipa_agg_jump_function_p
> known_aggs
;
3835 /* When we do caching, use do_estimate_edge_time to populate the entry. */
3837 if (edge_growth_cache
.exists ())
3839 do_estimate_edge_time (edge
);
3840 hints
= edge_growth_cache
[edge
->uid
].hints
;
3841 gcc_checking_assert (hints
);
3845 callee
= edge
->callee
->ultimate_alias_target ();
3847 /* Early inliner runs without caching, go ahead and do the dirty work. */
3848 gcc_checking_assert (edge
->inline_failed
);
3849 evaluate_properties_for_edge (edge
, true,
3850 &clause
, &known_vals
, &known_contexts
,
3852 estimate_node_size_and_time (callee
, clause
, known_vals
, known_contexts
,
3853 known_aggs
, NULL
, NULL
, NULL
, &hints
, vNULL
);
3854 known_vals
.release ();
3855 known_contexts
.release ();
3856 known_aggs
.release ();
3857 hints
|= simple_edge_hints (edge
);
3862 /* Estimate self time of the function NODE after inlining EDGE. */
3865 estimate_time_after_inlining (struct cgraph_node
*node
,
3866 struct cgraph_edge
*edge
)
3868 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3869 if (!es
->predicate
|| !false_predicate_p (es
->predicate
))
3872 inline_summaries
->get (node
)->time
+ estimate_edge_time (edge
);
3875 if (time
> MAX_TIME
)
3879 return inline_summaries
->get (node
)->time
;
3883 /* Estimate the size of NODE after inlining EDGE which should be an
3884 edge to either NODE or a call inlined into NODE. */
3887 estimate_size_after_inlining (struct cgraph_node
*node
,
3888 struct cgraph_edge
*edge
)
3890 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3891 if (!es
->predicate
|| !false_predicate_p (es
->predicate
))
3893 int size
= inline_summaries
->get (node
)->size
+ estimate_edge_growth (edge
);
3894 gcc_assert (size
>= 0);
3897 return inline_summaries
->get (node
)->size
;
3903 struct cgraph_node
*node
;
3904 bool self_recursive
;
3910 /* Worker for do_estimate_growth. Collect growth for all callers. */
3913 do_estimate_growth_1 (struct cgraph_node
*node
, void *data
)
3915 struct cgraph_edge
*e
;
3916 struct growth_data
*d
= (struct growth_data
*) data
;
3918 for (e
= node
->callers
; e
; e
= e
->next_caller
)
3920 gcc_checking_assert (e
->inline_failed
);
3922 if (cgraph_inline_failed_type (e
->inline_failed
) == CIF_FINAL_ERROR
)
3924 d
->uninlinable
= true;
3928 if (e
->recursive_p ())
3930 d
->self_recursive
= true;
3933 d
->growth
+= estimate_edge_growth (e
);
3939 /* Estimate the growth caused by inlining NODE into all callees. */
3942 estimate_growth (struct cgraph_node
*node
)
3944 struct growth_data d
= { node
, false, false, 0 };
3945 struct inline_summary
*info
= inline_summaries
->get (node
);
3947 node
->call_for_symbol_and_aliases (do_estimate_growth_1
, &d
, true);
3949 /* For self recursive functions the growth estimation really should be
3950 infinity. We don't want to return very large values because the growth
3951 plays various roles in badness computation fractions. Be sure to not
3952 return zero or negative growths. */
3953 if (d
.self_recursive
)
3954 d
.growth
= d
.growth
< info
->size
? info
->size
: d
.growth
;
3955 else if (DECL_EXTERNAL (node
->decl
) || d
.uninlinable
)
3959 if (node
->will_be_removed_from_program_if_no_direct_calls_p ())
3960 d
.growth
-= info
->size
;
3961 /* COMDAT functions are very often not shared across multiple units
3962 since they come from various template instantiations.
3963 Take this into account. */
3964 else if (DECL_COMDAT (node
->decl
)
3965 && node
->can_remove_if_no_direct_calls_p ())
3966 d
.growth
-= (info
->size
3967 * (100 - PARAM_VALUE (PARAM_COMDAT_SHARING_PROBABILITY
))
3974 /* Verify if there are fewer than MAX_CALLERS. */
3977 check_callers (cgraph_node
*node
, int *max_callers
)
3981 if (!node
->can_remove_if_no_direct_calls_and_refs_p ())
3984 for (cgraph_edge
*e
= node
->callers
; e
; e
= e
->next_caller
)
3988 || cgraph_inline_failed_type (e
->inline_failed
) == CIF_FINAL_ERROR
)
3992 FOR_EACH_ALIAS (node
, ref
)
3993 if (check_callers (dyn_cast
<cgraph_node
*> (ref
->referring
), max_callers
))
4000 /* Make cheap estimation if growth of NODE is likely positive knowing
4001 EDGE_GROWTH of one particular edge.
4002 We assume that most of other edges will have similar growth
4003 and skip computation if there are too many callers. */
4006 growth_likely_positive (struct cgraph_node
*node
,
4010 struct cgraph_edge
*e
;
4011 gcc_checking_assert (edge_growth
> 0);
4013 /* First quickly check if NODE is removable at all. */
4014 if (DECL_EXTERNAL (node
->decl
))
4016 if (!node
->can_remove_if_no_direct_calls_and_refs_p ()
4017 || node
->address_taken
)
4020 max_callers
= inline_summaries
->get (node
)->size
* 4 / edge_growth
+ 2;
4022 for (e
= node
->callers
; e
; e
= e
->next_caller
)
4026 || cgraph_inline_failed_type (e
->inline_failed
) == CIF_FINAL_ERROR
)
4031 FOR_EACH_ALIAS (node
, ref
)
4032 if (check_callers (dyn_cast
<cgraph_node
*> (ref
->referring
), &max_callers
))
4035 /* Unlike for functions called once, we play unsafe with
4036 COMDATs. We can allow that since we know functions
4037 in consideration are small (and thus risk is small) and
4038 moreover grow estimates already accounts that COMDAT
4039 functions may or may not disappear when eliminated from
4040 current unit. With good probability making aggressive
4041 choice in all units is going to make overall program
4043 if (DECL_COMDAT (node
->decl
))
4045 if (!node
->can_remove_if_no_direct_calls_p ())
4048 else if (!node
->will_be_removed_from_program_if_no_direct_calls_p ())
4051 return estimate_growth (node
) > 0;
4055 /* This function performs intraprocedural analysis in NODE that is required to
4056 inline indirect calls. */
4059 inline_indirect_intraprocedural_analysis (struct cgraph_node
*node
)
4061 ipa_analyze_node (node
);
4062 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4064 ipa_print_node_params (dump_file
, node
);
4065 ipa_print_node_jump_functions (dump_file
, node
);
4070 /* Note function body size. */
4073 inline_analyze_function (struct cgraph_node
*node
)
4075 push_cfun (DECL_STRUCT_FUNCTION (node
->decl
));
4078 fprintf (dump_file
, "\nAnalyzing function: %s/%u\n",
4079 node
->name (), node
->order
);
4080 if (opt_for_fn (node
->decl
, optimize
) && !node
->thunk
.thunk_p
)
4081 inline_indirect_intraprocedural_analysis (node
);
4082 compute_inline_parameters (node
, false);
4085 struct cgraph_edge
*e
;
4086 for (e
= node
->callees
; e
; e
= e
->next_callee
)
4088 if (e
->inline_failed
== CIF_FUNCTION_NOT_CONSIDERED
)
4089 e
->inline_failed
= CIF_FUNCTION_NOT_OPTIMIZED
;
4090 e
->call_stmt_cannot_inline_p
= true;
4092 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
4094 if (e
->inline_failed
== CIF_FUNCTION_NOT_CONSIDERED
)
4095 e
->inline_failed
= CIF_FUNCTION_NOT_OPTIMIZED
;
4096 e
->call_stmt_cannot_inline_p
= true;
4104 /* Called when new function is inserted to callgraph late. */
4107 inline_summary_t::insert (struct cgraph_node
*node
, inline_summary
*)
4109 inline_analyze_function (node
);
4112 /* Note function body size. */
4115 inline_generate_summary (void)
4117 struct cgraph_node
*node
;
4119 /* When not optimizing, do not bother to analyze. Inlining is still done
4120 because edge redirection needs to happen there. */
4121 if (!optimize
&& !flag_generate_lto
&& !flag_generate_offload
&& !flag_wpa
)
4124 if (!inline_summaries
)
4125 inline_summaries
= (inline_summary_t
*) inline_summary_t::create_ggc (symtab
);
4127 inline_summaries
->enable_insertion_hook ();
4129 ipa_register_cgraph_hooks ();
4130 inline_free_summary ();
4132 FOR_EACH_DEFINED_FUNCTION (node
)
4134 inline_analyze_function (node
);
4138 /* Read predicate from IB. */
4140 static struct predicate
4141 read_predicate (struct lto_input_block
*ib
)
4143 struct predicate out
;
4149 gcc_assert (k
<= MAX_CLAUSES
);
4150 clause
= out
.clause
[k
++] = streamer_read_uhwi (ib
);
4154 /* Zero-initialize the remaining clauses in OUT. */
4155 while (k
<= MAX_CLAUSES
)
4156 out
.clause
[k
++] = 0;
4162 /* Write inline summary for edge E to OB. */
4165 read_inline_edge_summary (struct lto_input_block
*ib
, struct cgraph_edge
*e
)
4167 struct inline_edge_summary
*es
= inline_edge_summary (e
);
4171 es
->call_stmt_size
= streamer_read_uhwi (ib
);
4172 es
->call_stmt_time
= streamer_read_uhwi (ib
);
4173 es
->loop_depth
= streamer_read_uhwi (ib
);
4174 p
= read_predicate (ib
);
4175 edge_set_predicate (e
, &p
);
4176 length
= streamer_read_uhwi (ib
);
4179 es
->param
.safe_grow_cleared (length
);
4180 for (i
= 0; i
< length
; i
++)
4181 es
->param
[i
].change_prob
= streamer_read_uhwi (ib
);
4186 /* Stream in inline summaries from the section. */
4189 inline_read_section (struct lto_file_decl_data
*file_data
, const char *data
,
4192 const struct lto_function_header
*header
=
4193 (const struct lto_function_header
*) data
;
4194 const int cfg_offset
= sizeof (struct lto_function_header
);
4195 const int main_offset
= cfg_offset
+ header
->cfg_size
;
4196 const int string_offset
= main_offset
+ header
->main_size
;
4197 struct data_in
*data_in
;
4198 unsigned int i
, count2
, j
;
4199 unsigned int f_count
;
4201 lto_input_block
ib ((const char *) data
+ main_offset
, header
->main_size
,
4202 file_data
->mode_table
);
4205 lto_data_in_create (file_data
, (const char *) data
+ string_offset
,
4206 header
->string_size
, vNULL
);
4207 f_count
= streamer_read_uhwi (&ib
);
4208 for (i
= 0; i
< f_count
; i
++)
4211 struct cgraph_node
*node
;
4212 struct inline_summary
*info
;
4213 lto_symtab_encoder_t encoder
;
4214 struct bitpack_d bp
;
4215 struct cgraph_edge
*e
;
4218 index
= streamer_read_uhwi (&ib
);
4219 encoder
= file_data
->symtab_node_encoder
;
4220 node
= dyn_cast
<cgraph_node
*> (lto_symtab_encoder_deref (encoder
,
4222 info
= inline_summaries
->get (node
);
4224 info
->estimated_stack_size
4225 = info
->estimated_self_stack_size
= streamer_read_uhwi (&ib
);
4226 info
->size
= info
->self_size
= streamer_read_uhwi (&ib
);
4227 info
->time
= info
->self_time
= streamer_read_uhwi (&ib
);
4229 bp
= streamer_read_bitpack (&ib
);
4230 info
->inlinable
= bp_unpack_value (&bp
, 1);
4232 count2
= streamer_read_uhwi (&ib
);
4233 gcc_assert (!info
->conds
);
4234 for (j
= 0; j
< count2
; j
++)
4237 c
.operand_num
= streamer_read_uhwi (&ib
);
4238 c
.code
= (enum tree_code
) streamer_read_uhwi (&ib
);
4239 c
.val
= stream_read_tree (&ib
, data_in
);
4240 bp
= streamer_read_bitpack (&ib
);
4241 c
.agg_contents
= bp_unpack_value (&bp
, 1);
4242 c
.by_ref
= bp_unpack_value (&bp
, 1);
4244 c
.offset
= streamer_read_uhwi (&ib
);
4245 vec_safe_push (info
->conds
, c
);
4247 count2
= streamer_read_uhwi (&ib
);
4248 gcc_assert (!info
->entry
);
4249 for (j
= 0; j
< count2
; j
++)
4251 struct size_time_entry e
;
4253 e
.size
= streamer_read_uhwi (&ib
);
4254 e
.time
= streamer_read_uhwi (&ib
);
4255 e
.predicate
= read_predicate (&ib
);
4257 vec_safe_push (info
->entry
, e
);
4260 p
= read_predicate (&ib
);
4261 set_hint_predicate (&info
->loop_iterations
, p
);
4262 p
= read_predicate (&ib
);
4263 set_hint_predicate (&info
->loop_stride
, p
);
4264 p
= read_predicate (&ib
);
4265 set_hint_predicate (&info
->array_index
, p
);
4266 for (e
= node
->callees
; e
; e
= e
->next_callee
)
4267 read_inline_edge_summary (&ib
, e
);
4268 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
4269 read_inline_edge_summary (&ib
, e
);
4272 lto_free_section_data (file_data
, LTO_section_inline_summary
, NULL
, data
,
4274 lto_data_in_delete (data_in
);
4278 /* Read inline summary. Jump functions are shared among ipa-cp
4279 and inliner, so when ipa-cp is active, we don't need to write them
4283 inline_read_summary (void)
4285 struct lto_file_decl_data
**file_data_vec
= lto_get_file_decl_data ();
4286 struct lto_file_decl_data
*file_data
;
4289 inline_summary_alloc ();
4291 while ((file_data
= file_data_vec
[j
++]))
4294 const char *data
= lto_get_section_data (file_data
,
4295 LTO_section_inline_summary
,
4298 inline_read_section (file_data
, data
, len
);
4300 /* Fatal error here. We do not want to support compiling ltrans units
4301 with different version of compiler or different flags than the WPA
4302 unit, so this should never happen. */
4303 fatal_error (input_location
,
4304 "ipa inline summary is missing in input file");
4308 ipa_register_cgraph_hooks ();
4310 ipa_prop_read_jump_functions ();
4313 gcc_assert (inline_summaries
);
4314 inline_summaries
->enable_insertion_hook ();
4318 /* Write predicate P to OB. */
4321 write_predicate (struct output_block
*ob
, struct predicate
*p
)
4325 for (j
= 0; p
->clause
[j
]; j
++)
4327 gcc_assert (j
< MAX_CLAUSES
);
4328 streamer_write_uhwi (ob
, p
->clause
[j
]);
4330 streamer_write_uhwi (ob
, 0);
4334 /* Write inline summary for edge E to OB. */
4337 write_inline_edge_summary (struct output_block
*ob
, struct cgraph_edge
*e
)
4339 struct inline_edge_summary
*es
= inline_edge_summary (e
);
4342 streamer_write_uhwi (ob
, es
->call_stmt_size
);
4343 streamer_write_uhwi (ob
, es
->call_stmt_time
);
4344 streamer_write_uhwi (ob
, es
->loop_depth
);
4345 write_predicate (ob
, es
->predicate
);
4346 streamer_write_uhwi (ob
, es
->param
.length ());
4347 for (i
= 0; i
< (int) es
->param
.length (); i
++)
4348 streamer_write_uhwi (ob
, es
->param
[i
].change_prob
);
4352 /* Write inline summary for node in SET.
4353 Jump functions are shared among ipa-cp and inliner, so when ipa-cp is
4354 active, we don't need to write them twice. */
4357 inline_write_summary (void)
4359 struct cgraph_node
*node
;
4360 struct output_block
*ob
= create_output_block (LTO_section_inline_summary
);
4361 lto_symtab_encoder_t encoder
= ob
->decl_state
->symtab_node_encoder
;
4362 unsigned int count
= 0;
4365 for (i
= 0; i
< lto_symtab_encoder_size (encoder
); i
++)
4367 symtab_node
*snode
= lto_symtab_encoder_deref (encoder
, i
);
4368 cgraph_node
*cnode
= dyn_cast
<cgraph_node
*> (snode
);
4369 if (cnode
&& cnode
->definition
&& !cnode
->alias
)
4372 streamer_write_uhwi (ob
, count
);
4374 for (i
= 0; i
< lto_symtab_encoder_size (encoder
); i
++)
4376 symtab_node
*snode
= lto_symtab_encoder_deref (encoder
, i
);
4377 cgraph_node
*cnode
= dyn_cast
<cgraph_node
*> (snode
);
4378 if (cnode
&& (node
= cnode
)->definition
&& !node
->alias
)
4380 struct inline_summary
*info
= inline_summaries
->get (node
);
4381 struct bitpack_d bp
;
4382 struct cgraph_edge
*edge
;
4385 struct condition
*c
;
4387 streamer_write_uhwi (ob
,
4388 lto_symtab_encoder_encode (encoder
,
4391 streamer_write_hwi (ob
, info
->estimated_self_stack_size
);
4392 streamer_write_hwi (ob
, info
->self_size
);
4393 streamer_write_hwi (ob
, info
->self_time
);
4394 bp
= bitpack_create (ob
->main_stream
);
4395 bp_pack_value (&bp
, info
->inlinable
, 1);
4396 streamer_write_bitpack (&bp
);
4397 streamer_write_uhwi (ob
, vec_safe_length (info
->conds
));
4398 for (i
= 0; vec_safe_iterate (info
->conds
, i
, &c
); i
++)
4400 streamer_write_uhwi (ob
, c
->operand_num
);
4401 streamer_write_uhwi (ob
, c
->code
);
4402 stream_write_tree (ob
, c
->val
, true);
4403 bp
= bitpack_create (ob
->main_stream
);
4404 bp_pack_value (&bp
, c
->agg_contents
, 1);
4405 bp_pack_value (&bp
, c
->by_ref
, 1);
4406 streamer_write_bitpack (&bp
);
4407 if (c
->agg_contents
)
4408 streamer_write_uhwi (ob
, c
->offset
);
4410 streamer_write_uhwi (ob
, vec_safe_length (info
->entry
));
4411 for (i
= 0; vec_safe_iterate (info
->entry
, i
, &e
); i
++)
4413 streamer_write_uhwi (ob
, e
->size
);
4414 streamer_write_uhwi (ob
, e
->time
);
4415 write_predicate (ob
, &e
->predicate
);
4417 write_predicate (ob
, info
->loop_iterations
);
4418 write_predicate (ob
, info
->loop_stride
);
4419 write_predicate (ob
, info
->array_index
);
4420 for (edge
= node
->callees
; edge
; edge
= edge
->next_callee
)
4421 write_inline_edge_summary (ob
, edge
);
4422 for (edge
= node
->indirect_calls
; edge
; edge
= edge
->next_callee
)
4423 write_inline_edge_summary (ob
, edge
);
4426 streamer_write_char_stream (ob
->main_stream
, 0);
4427 produce_asm (ob
, NULL
);
4428 destroy_output_block (ob
);
4430 if (optimize
&& !flag_ipa_cp
)
4431 ipa_prop_write_jump_functions ();
4435 /* Release inline summary. */
4438 inline_free_summary (void)
4440 struct cgraph_node
*node
;
4441 if (edge_removal_hook_holder
)
4442 symtab
->remove_edge_removal_hook (edge_removal_hook_holder
);
4443 edge_removal_hook_holder
= NULL
;
4444 if (edge_duplication_hook_holder
)
4445 symtab
->remove_edge_duplication_hook (edge_duplication_hook_holder
);
4446 edge_duplication_hook_holder
= NULL
;
4447 if (!inline_edge_summary_vec
.exists ())
4449 FOR_EACH_DEFINED_FUNCTION (node
)
4451 reset_inline_summary (node
, inline_summaries
->get (node
));
4452 inline_summaries
->release ();
4453 inline_summaries
= NULL
;
4454 inline_edge_summary_vec
.release ();
4455 if (edge_predicate_pool
)
4456 free_alloc_pool (edge_predicate_pool
);
4457 edge_predicate_pool
= 0;