1 /* Inlining decision heuristics.
2 Copyright (C) 2003-2013 Free Software Foundation, Inc.
3 Contributed by Jan Hubicka
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
21 /* Analysis used by the inliner and other passes limiting code size growth.
23 We estimate for each function
25 - average function execution time
26 - inlining size benefit (that is how much of function body size
27 and its call sequence is expected to disappear by inlining)
28 - inlining time benefit
31 - call statement size and time
33 inlinie_summary datastructures store above information locally (i.e.
34 parameters of the function itself) and globally (i.e. parameters of
35 the function created by applying all the inline decisions already
36 present in the callgraph).
38 We provide accestor to the inline_summary datastructure and
39 basic logic updating the parameters when inlining is performed.
41 The summaries are context sensitive. Context means
42 1) partial assignment of known constant values of operands
43 2) whether function is inlined into the call or not.
44 It is easy to add more variants. To represent function size and time
45 that depends on context (i.e. it is known to be optimized away when
46 context is known either by inlining or from IP-CP and clonning),
47 we use predicates. Predicates are logical formulas in
48 conjunctive-disjunctive form consisting of clauses. Clauses are bitmaps
49 specifying what conditions must be true. Conditions are simple test
50 of the form described above.
52 In order to make predicate (possibly) true, all of its clauses must
53 be (possibly) true. To make clause (possibly) true, one of conditions
54 it mentions must be (possibly) true. There are fixed bounds on
55 number of clauses and conditions and all the manipulation functions
56 are conservative in positive direction. I.e. we may lose precision
57 by thinking that predicate may be true even when it is not.
59 estimate_edge_size and estimate_edge_growth can be used to query
60 function size/time in the given context. inline_merge_summary merges
61 properties of caller and callee after inlining.
63 Finally pass_inline_parameters is exported. This is used to drive
64 computation of function parameters used by the early inliner. IPA
65 inlined performs analysis via its analyze_function method. */
69 #include "coretypes.h"
72 #include "stor-layout.h"
73 #include "stringpool.h"
74 #include "print-tree.h"
75 #include "tree-inline.h"
76 #include "langhooks.h"
78 #include "diagnostic.h"
79 #include "gimple-pretty-print.h"
81 #include "tree-pass.h"
85 #include "gimple-iterator.h"
86 #include "gimple-ssa.h"
88 #include "tree-phinodes.h"
89 #include "ssa-iterators.h"
90 #include "tree-ssanames.h"
91 #include "tree-ssa-loop-niter.h"
92 #include "tree-ssa-loop.h"
94 #include "lto-streamer.h"
95 #include "data-streamer.h"
96 #include "tree-streamer.h"
97 #include "ipa-inline.h"
98 #include "alloc-pool.h"
100 #include "tree-scalar-evolution.h"
101 #include "ipa-utils.h"
103 #include "cfgexpand.h"
105 /* Estimate runtime of function can easilly run into huge numbers with many
106 nested loops. Be sure we can compute time * INLINE_SIZE_SCALE * 2 in an
107 integer. For anything larger we use gcov_type. */
108 #define MAX_TIME 500000
110 /* Number of bits in integer, but we really want to be stable across different
112 #define NUM_CONDITIONS 32
114 enum predicate_conditions
116 predicate_false_condition
= 0,
117 predicate_not_inlined_condition
= 1,
118 predicate_first_dynamic_condition
= 2
121 /* Special condition code we use to represent test that operand is compile time
123 #define IS_NOT_CONSTANT ERROR_MARK
124 /* Special condition code we use to represent test that operand is not changed
125 across invocation of the function. When operand IS_NOT_CONSTANT it is always
126 CHANGED, however i.e. loop invariants can be NOT_CHANGED given percentage
127 of executions even when they are not compile time constants. */
128 #define CHANGED IDENTIFIER_NODE
130 /* Holders of ipa cgraph hooks: */
131 static struct cgraph_node_hook_list
*function_insertion_hook_holder
;
132 static struct cgraph_node_hook_list
*node_removal_hook_holder
;
133 static struct cgraph_2node_hook_list
*node_duplication_hook_holder
;
134 static struct cgraph_2edge_hook_list
*edge_duplication_hook_holder
;
135 static struct cgraph_edge_hook_list
*edge_removal_hook_holder
;
136 static void inline_node_removal_hook (struct cgraph_node
*, void *);
137 static void inline_node_duplication_hook (struct cgraph_node
*,
138 struct cgraph_node
*, void *);
139 static void inline_edge_removal_hook (struct cgraph_edge
*, void *);
140 static void inline_edge_duplication_hook (struct cgraph_edge
*,
141 struct cgraph_edge
*, void *);
143 /* VECtor holding inline summaries.
144 In GGC memory because conditions might point to constant trees. */
145 vec
<inline_summary_t
, va_gc
> *inline_summary_vec
;
146 vec
<inline_edge_summary_t
> inline_edge_summary_vec
;
148 /* Cached node/edge growths. */
149 vec
<int> node_growth_cache
;
150 vec
<edge_growth_cache_entry
> edge_growth_cache
;
152 /* Edge predicates goes here. */
153 static alloc_pool edge_predicate_pool
;
155 /* Return true predicate (tautology).
156 We represent it by empty list of clauses. */
158 static inline struct predicate
159 true_predicate (void)
167 /* Return predicate testing single condition number COND. */
169 static inline struct predicate
170 single_cond_predicate (int cond
)
173 p
.clause
[0] = 1 << cond
;
179 /* Return false predicate. First clause require false condition. */
181 static inline struct predicate
182 false_predicate (void)
184 return single_cond_predicate (predicate_false_condition
);
188 /* Return true if P is (false). */
191 true_predicate_p (struct predicate
*p
)
193 return !p
->clause
[0];
197 /* Return true if P is (false). */
200 false_predicate_p (struct predicate
*p
)
202 if (p
->clause
[0] == (1 << predicate_false_condition
))
204 gcc_checking_assert (!p
->clause
[1]
205 && p
->clause
[0] == 1 << predicate_false_condition
);
212 /* Return predicate that is set true when function is not inlined. */
214 static inline struct predicate
215 not_inlined_predicate (void)
217 return single_cond_predicate (predicate_not_inlined_condition
);
220 /* Simple description of whether a memory load or a condition refers to a load
221 from an aggregate and if so, how and where from in the aggregate.
222 Individual fields have the same meaning like fields with the same name in
225 struct agg_position_info
227 HOST_WIDE_INT offset
;
232 /* Add condition to condition list CONDS. AGGPOS describes whether the used
233 oprand is loaded from an aggregate and where in the aggregate it is. It can
234 be NULL, which means this not a load from an aggregate. */
236 static struct predicate
237 add_condition (struct inline_summary
*summary
, int operand_num
,
238 struct agg_position_info
*aggpos
,
239 enum tree_code code
, tree val
)
243 struct condition new_cond
;
244 HOST_WIDE_INT offset
;
245 bool agg_contents
, by_ref
;
249 offset
= aggpos
->offset
;
250 agg_contents
= aggpos
->agg_contents
;
251 by_ref
= aggpos
->by_ref
;
256 agg_contents
= false;
260 gcc_checking_assert (operand_num
>= 0);
261 for (i
= 0; vec_safe_iterate (summary
->conds
, i
, &c
); i
++)
263 if (c
->operand_num
== operand_num
266 && c
->agg_contents
== agg_contents
267 && (!agg_contents
|| (c
->offset
== offset
&& c
->by_ref
== by_ref
)))
268 return single_cond_predicate (i
+ predicate_first_dynamic_condition
);
270 /* Too many conditions. Give up and return constant true. */
271 if (i
== NUM_CONDITIONS
- predicate_first_dynamic_condition
)
272 return true_predicate ();
274 new_cond
.operand_num
= operand_num
;
275 new_cond
.code
= code
;
277 new_cond
.agg_contents
= agg_contents
;
278 new_cond
.by_ref
= by_ref
;
279 new_cond
.offset
= offset
;
280 vec_safe_push (summary
->conds
, new_cond
);
281 return single_cond_predicate (i
+ predicate_first_dynamic_condition
);
285 /* Add clause CLAUSE into the predicate P. */
288 add_clause (conditions conditions
, struct predicate
*p
, clause_t clause
)
292 int insert_here
= -1;
299 /* False clause makes the whole predicate false. Kill the other variants. */
300 if (clause
== (1 << predicate_false_condition
))
302 p
->clause
[0] = (1 << predicate_false_condition
);
306 if (false_predicate_p (p
))
309 /* No one should be sily enough to add false into nontrivial clauses. */
310 gcc_checking_assert (!(clause
& (1 << predicate_false_condition
)));
312 /* Look where to insert the clause. At the same time prune out
313 clauses of P that are implied by the new clause and thus
315 for (i
= 0, i2
= 0; i
<= MAX_CLAUSES
; i
++)
317 p
->clause
[i2
] = p
->clause
[i
];
322 /* If p->clause[i] implies clause, there is nothing to add. */
323 if ((p
->clause
[i
] & clause
) == p
->clause
[i
])
325 /* We had nothing to add, none of clauses should've become
327 gcc_checking_assert (i
== i2
);
331 if (p
->clause
[i
] < clause
&& insert_here
< 0)
334 /* If clause implies p->clause[i], then p->clause[i] becomes redundant.
335 Otherwise the p->clause[i] has to stay. */
336 if ((p
->clause
[i
] & clause
) != clause
)
340 /* Look for clauses that are obviously true. I.e.
341 op0 == 5 || op0 != 5. */
342 for (c1
= predicate_first_dynamic_condition
; c1
< NUM_CONDITIONS
; c1
++)
345 if (!(clause
& (1 << c1
)))
347 cc1
= &(*conditions
)[c1
- predicate_first_dynamic_condition
];
348 /* We have no way to represent !CHANGED and !IS_NOT_CONSTANT
349 and thus there is no point for looking for them. */
350 if (cc1
->code
== CHANGED
|| cc1
->code
== IS_NOT_CONSTANT
)
352 for (c2
= c1
+ 1; c2
< NUM_CONDITIONS
; c2
++)
353 if (clause
& (1 << c2
))
356 &(*conditions
)[c1
- predicate_first_dynamic_condition
];
358 &(*conditions
)[c2
- predicate_first_dynamic_condition
];
359 if (cc1
->operand_num
== cc2
->operand_num
360 && cc1
->val
== cc2
->val
361 && cc2
->code
!= IS_NOT_CONSTANT
362 && cc2
->code
!= CHANGED
363 && cc1
->code
== invert_tree_comparison
365 HONOR_NANS (TYPE_MODE (TREE_TYPE (cc1
->val
)))))
371 /* We run out of variants. Be conservative in positive direction. */
372 if (i2
== MAX_CLAUSES
)
374 /* Keep clauses in decreasing order. This makes equivalence testing easy. */
375 p
->clause
[i2
+ 1] = 0;
376 if (insert_here
>= 0)
377 for (; i2
> insert_here
; i2
--)
378 p
->clause
[i2
] = p
->clause
[i2
- 1];
381 p
->clause
[insert_here
] = clause
;
387 static struct predicate
388 and_predicates (conditions conditions
,
389 struct predicate
*p
, struct predicate
*p2
)
391 struct predicate out
= *p
;
394 /* Avoid busy work. */
395 if (false_predicate_p (p2
) || true_predicate_p (p
))
397 if (false_predicate_p (p
) || true_predicate_p (p2
))
400 /* See how far predicates match. */
401 for (i
= 0; p
->clause
[i
] && p
->clause
[i
] == p2
->clause
[i
]; i
++)
403 gcc_checking_assert (i
< MAX_CLAUSES
);
406 /* Combine the predicates rest. */
407 for (; p2
->clause
[i
]; i
++)
409 gcc_checking_assert (i
< MAX_CLAUSES
);
410 add_clause (conditions
, &out
, p2
->clause
[i
]);
416 /* Return true if predicates are obviously equal. */
419 predicates_equal_p (struct predicate
*p
, struct predicate
*p2
)
422 for (i
= 0; p
->clause
[i
]; i
++)
424 gcc_checking_assert (i
< MAX_CLAUSES
);
425 gcc_checking_assert (p
->clause
[i
] > p
->clause
[i
+ 1]);
426 gcc_checking_assert (!p2
->clause
[i
]
427 || p2
->clause
[i
] > p2
->clause
[i
+ 1]);
428 if (p
->clause
[i
] != p2
->clause
[i
])
431 return !p2
->clause
[i
];
437 static struct predicate
438 or_predicates (conditions conditions
,
439 struct predicate
*p
, struct predicate
*p2
)
441 struct predicate out
= true_predicate ();
444 /* Avoid busy work. */
445 if (false_predicate_p (p2
) || true_predicate_p (p
))
447 if (false_predicate_p (p
) || true_predicate_p (p2
))
449 if (predicates_equal_p (p
, p2
))
452 /* OK, combine the predicates. */
453 for (i
= 0; p
->clause
[i
]; i
++)
454 for (j
= 0; p2
->clause
[j
]; j
++)
456 gcc_checking_assert (i
< MAX_CLAUSES
&& j
< MAX_CLAUSES
);
457 add_clause (conditions
, &out
, p
->clause
[i
] | p2
->clause
[j
]);
463 /* Having partial truth assignment in POSSIBLE_TRUTHS, return false
464 if predicate P is known to be false. */
467 evaluate_predicate (struct predicate
*p
, clause_t possible_truths
)
471 /* True remains true. */
472 if (true_predicate_p (p
))
475 gcc_assert (!(possible_truths
& (1 << predicate_false_condition
)));
477 /* See if we can find clause we can disprove. */
478 for (i
= 0; p
->clause
[i
]; i
++)
480 gcc_checking_assert (i
< MAX_CLAUSES
);
481 if (!(p
->clause
[i
] & possible_truths
))
487 /* Return the probability in range 0...REG_BR_PROB_BASE that the predicated
488 instruction will be recomputed per invocation of the inlined call. */
491 predicate_probability (conditions conds
,
492 struct predicate
*p
, clause_t possible_truths
,
493 vec
<inline_param_summary_t
> inline_param_summary
)
496 int combined_prob
= REG_BR_PROB_BASE
;
498 /* True remains true. */
499 if (true_predicate_p (p
))
500 return REG_BR_PROB_BASE
;
502 if (false_predicate_p (p
))
505 gcc_assert (!(possible_truths
& (1 << predicate_false_condition
)));
507 /* See if we can find clause we can disprove. */
508 for (i
= 0; p
->clause
[i
]; i
++)
510 gcc_checking_assert (i
< MAX_CLAUSES
);
511 if (!(p
->clause
[i
] & possible_truths
))
517 if (!inline_param_summary
.exists ())
518 return REG_BR_PROB_BASE
;
519 for (i2
= 0; i2
< NUM_CONDITIONS
; i2
++)
520 if ((p
->clause
[i
] & possible_truths
) & (1 << i2
))
522 if (i2
>= predicate_first_dynamic_condition
)
525 &(*conds
)[i2
- predicate_first_dynamic_condition
];
526 if (c
->code
== CHANGED
528 (int) inline_param_summary
.length ()))
531 inline_param_summary
[c
->operand_num
].change_prob
;
532 this_prob
= MAX (this_prob
, iprob
);
535 this_prob
= REG_BR_PROB_BASE
;
538 this_prob
= REG_BR_PROB_BASE
;
540 combined_prob
= MIN (this_prob
, combined_prob
);
545 return combined_prob
;
549 /* Dump conditional COND. */
552 dump_condition (FILE *f
, conditions conditions
, int cond
)
555 if (cond
== predicate_false_condition
)
556 fprintf (f
, "false");
557 else if (cond
== predicate_not_inlined_condition
)
558 fprintf (f
, "not inlined");
561 c
= &(*conditions
)[cond
- predicate_first_dynamic_condition
];
562 fprintf (f
, "op%i", c
->operand_num
);
564 fprintf (f
, "[%soffset: " HOST_WIDE_INT_PRINT_DEC
"]",
565 c
->by_ref
? "ref " : "", c
->offset
);
566 if (c
->code
== IS_NOT_CONSTANT
)
568 fprintf (f
, " not constant");
571 if (c
->code
== CHANGED
)
573 fprintf (f
, " changed");
576 fprintf (f
, " %s ", op_symbol_code (c
->code
));
577 print_generic_expr (f
, c
->val
, 1);
582 /* Dump clause CLAUSE. */
585 dump_clause (FILE *f
, conditions conds
, clause_t clause
)
592 for (i
= 0; i
< NUM_CONDITIONS
; i
++)
593 if (clause
& (1 << i
))
598 dump_condition (f
, conds
, i
);
604 /* Dump predicate PREDICATE. */
607 dump_predicate (FILE *f
, conditions conds
, struct predicate
*pred
)
610 if (true_predicate_p (pred
))
611 dump_clause (f
, conds
, 0);
613 for (i
= 0; pred
->clause
[i
]; i
++)
617 dump_clause (f
, conds
, pred
->clause
[i
]);
623 /* Dump inline hints. */
625 dump_inline_hints (FILE *f
, inline_hints hints
)
629 fprintf (f
, "inline hints:");
630 if (hints
& INLINE_HINT_indirect_call
)
632 hints
&= ~INLINE_HINT_indirect_call
;
633 fprintf (f
, " indirect_call");
635 if (hints
& INLINE_HINT_loop_iterations
)
637 hints
&= ~INLINE_HINT_loop_iterations
;
638 fprintf (f
, " loop_iterations");
640 if (hints
& INLINE_HINT_loop_stride
)
642 hints
&= ~INLINE_HINT_loop_stride
;
643 fprintf (f
, " loop_stride");
645 if (hints
& INLINE_HINT_same_scc
)
647 hints
&= ~INLINE_HINT_same_scc
;
648 fprintf (f
, " same_scc");
650 if (hints
& INLINE_HINT_in_scc
)
652 hints
&= ~INLINE_HINT_in_scc
;
653 fprintf (f
, " in_scc");
655 if (hints
& INLINE_HINT_cross_module
)
657 hints
&= ~INLINE_HINT_cross_module
;
658 fprintf (f
, " cross_module");
660 if (hints
& INLINE_HINT_declared_inline
)
662 hints
&= ~INLINE_HINT_declared_inline
;
663 fprintf (f
, " declared_inline");
665 if (hints
& INLINE_HINT_array_index
)
667 hints
&= ~INLINE_HINT_array_index
;
668 fprintf (f
, " array_index");
674 /* Record SIZE and TIME under condition PRED into the inline summary. */
677 account_size_time (struct inline_summary
*summary
, int size
, int time
,
678 struct predicate
*pred
)
684 if (false_predicate_p (pred
))
687 /* We need to create initial empty unconitional clause, but otherwie
688 we don't need to account empty times and sizes. */
689 if (!size
&& !time
&& summary
->entry
)
692 /* Watch overflow that might result from insane profiles. */
693 if (time
> MAX_TIME
* INLINE_TIME_SCALE
)
694 time
= MAX_TIME
* INLINE_TIME_SCALE
;
695 gcc_assert (time
>= 0);
697 for (i
= 0; vec_safe_iterate (summary
->entry
, i
, &e
); i
++)
698 if (predicates_equal_p (&e
->predicate
, pred
))
707 e
= &(*summary
->entry
)[0];
708 gcc_assert (!e
->predicate
.clause
[0]);
709 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
711 "\t\tReached limit on number of entries, "
712 "ignoring the predicate.");
714 if (dump_file
&& (dump_flags
& TDF_DETAILS
) && (time
|| size
))
717 "\t\tAccounting size:%3.2f, time:%3.2f on %spredicate:",
718 ((double) size
) / INLINE_SIZE_SCALE
,
719 ((double) time
) / INLINE_TIME_SCALE
, found
? "" : "new ");
720 dump_predicate (dump_file
, summary
->conds
, pred
);
724 struct size_time_entry new_entry
;
725 new_entry
.size
= size
;
726 new_entry
.time
= time
;
727 new_entry
.predicate
= *pred
;
728 vec_safe_push (summary
->entry
, new_entry
);
734 if (e
->time
> MAX_TIME
* INLINE_TIME_SCALE
)
735 e
->time
= MAX_TIME
* INLINE_TIME_SCALE
;
739 /* Set predicate for edge E. */
742 edge_set_predicate (struct cgraph_edge
*e
, struct predicate
*predicate
)
744 struct inline_edge_summary
*es
= inline_edge_summary (e
);
745 if (predicate
&& !true_predicate_p (predicate
))
748 es
->predicate
= (struct predicate
*) pool_alloc (edge_predicate_pool
);
749 *es
->predicate
= *predicate
;
754 pool_free (edge_predicate_pool
, es
->predicate
);
755 es
->predicate
= NULL
;
759 /* Set predicate for hint *P. */
762 set_hint_predicate (struct predicate
**p
, struct predicate new_predicate
)
764 if (false_predicate_p (&new_predicate
) || true_predicate_p (&new_predicate
))
767 pool_free (edge_predicate_pool
, *p
);
773 *p
= (struct predicate
*) pool_alloc (edge_predicate_pool
);
779 /* KNOWN_VALS is partial mapping of parameters of NODE to constant values.
780 KNOWN_AGGS is a vector of aggreggate jump functions for each parameter.
781 Return clause of possible truths. When INLINE_P is true, assume that we are
784 ERROR_MARK means compile time invariant. */
787 evaluate_conditions_for_known_args (struct cgraph_node
*node
,
789 vec
<tree
> known_vals
,
790 vec
<ipa_agg_jump_function_p
>
793 clause_t clause
= inline_p
? 0 : 1 << predicate_not_inlined_condition
;
794 struct inline_summary
*info
= inline_summary (node
);
798 for (i
= 0; vec_safe_iterate (info
->conds
, i
, &c
); i
++)
803 /* We allow call stmt to have fewer arguments than the callee function
804 (especially for K&R style programs). So bound check here (we assume
805 known_aggs vector, if non-NULL, has the same length as
807 gcc_checking_assert (!known_aggs
.exists ()
808 || (known_vals
.length () == known_aggs
.length ()));
809 if (c
->operand_num
>= (int) known_vals
.length ())
811 clause
|= 1 << (i
+ predicate_first_dynamic_condition
);
817 struct ipa_agg_jump_function
*agg
;
819 if (c
->code
== CHANGED
821 && (known_vals
[c
->operand_num
] == error_mark_node
))
824 if (known_aggs
.exists ())
826 agg
= known_aggs
[c
->operand_num
];
827 val
= ipa_find_agg_cst_for_param (agg
, c
->offset
, c
->by_ref
);
834 val
= known_vals
[c
->operand_num
];
835 if (val
== error_mark_node
&& c
->code
!= CHANGED
)
841 clause
|= 1 << (i
+ predicate_first_dynamic_condition
);
844 if (c
->code
== IS_NOT_CONSTANT
|| c
->code
== CHANGED
)
846 res
= fold_binary_to_constant (c
->code
, boolean_type_node
, val
, c
->val
);
847 if (res
&& integer_zerop (res
))
849 clause
|= 1 << (i
+ predicate_first_dynamic_condition
);
855 /* Work out what conditions might be true at invocation of E. */
858 evaluate_properties_for_edge (struct cgraph_edge
*e
, bool inline_p
,
859 clause_t
*clause_ptr
,
860 vec
<tree
> *known_vals_ptr
,
861 vec
<tree
> *known_binfos_ptr
,
862 vec
<ipa_agg_jump_function_p
> *known_aggs_ptr
)
864 struct cgraph_node
*callee
=
865 cgraph_function_or_thunk_node (e
->callee
, NULL
);
866 struct inline_summary
*info
= inline_summary (callee
);
867 vec
<tree
> known_vals
= vNULL
;
868 vec
<ipa_agg_jump_function_p
> known_aggs
= vNULL
;
871 *clause_ptr
= inline_p
? 0 : 1 << predicate_not_inlined_condition
;
873 known_vals_ptr
->create (0);
874 if (known_binfos_ptr
)
875 known_binfos_ptr
->create (0);
877 if (ipa_node_params_vector
.exists ()
878 && !e
->call_stmt_cannot_inline_p
879 && ((clause_ptr
&& info
->conds
) || known_vals_ptr
|| known_binfos_ptr
))
881 struct ipa_node_params
*parms_info
;
882 struct ipa_edge_args
*args
= IPA_EDGE_REF (e
);
883 struct inline_edge_summary
*es
= inline_edge_summary (e
);
884 int i
, count
= ipa_get_cs_argument_count (args
);
886 if (e
->caller
->global
.inlined_to
)
887 parms_info
= IPA_NODE_REF (e
->caller
->global
.inlined_to
);
889 parms_info
= IPA_NODE_REF (e
->caller
);
891 if (count
&& (info
->conds
|| known_vals_ptr
))
892 known_vals
.safe_grow_cleared (count
);
893 if (count
&& (info
->conds
|| known_aggs_ptr
))
894 known_aggs
.safe_grow_cleared (count
);
895 if (count
&& known_binfos_ptr
)
896 known_binfos_ptr
->safe_grow_cleared (count
);
898 for (i
= 0; i
< count
; i
++)
900 struct ipa_jump_func
*jf
= ipa_get_ith_jump_func (args
, i
);
901 tree cst
= ipa_value_from_jfunc (parms_info
, jf
);
904 if (known_vals
.exists () && TREE_CODE (cst
) != TREE_BINFO
)
906 else if (known_binfos_ptr
!= NULL
907 && TREE_CODE (cst
) == TREE_BINFO
)
908 (*known_binfos_ptr
)[i
] = cst
;
910 else if (inline_p
&& !es
->param
[i
].change_prob
)
911 known_vals
[i
] = error_mark_node
;
912 /* TODO: When IPA-CP starts propagating and merging aggregate jump
913 functions, use its knowledge of the caller too, just like the
914 scalar case above. */
915 known_aggs
[i
] = &jf
->agg
;
920 *clause_ptr
= evaluate_conditions_for_known_args (callee
, inline_p
,
921 known_vals
, known_aggs
);
924 *known_vals_ptr
= known_vals
;
926 known_vals
.release ();
929 *known_aggs_ptr
= known_aggs
;
931 known_aggs
.release ();
935 /* Allocate the inline summary vector or resize it to cover all cgraph nodes. */
938 inline_summary_alloc (void)
940 if (!node_removal_hook_holder
)
941 node_removal_hook_holder
=
942 cgraph_add_node_removal_hook (&inline_node_removal_hook
, NULL
);
943 if (!edge_removal_hook_holder
)
944 edge_removal_hook_holder
=
945 cgraph_add_edge_removal_hook (&inline_edge_removal_hook
, NULL
);
946 if (!node_duplication_hook_holder
)
947 node_duplication_hook_holder
=
948 cgraph_add_node_duplication_hook (&inline_node_duplication_hook
, NULL
);
949 if (!edge_duplication_hook_holder
)
950 edge_duplication_hook_holder
=
951 cgraph_add_edge_duplication_hook (&inline_edge_duplication_hook
, NULL
);
953 if (vec_safe_length (inline_summary_vec
) <= (unsigned) cgraph_max_uid
)
954 vec_safe_grow_cleared (inline_summary_vec
, cgraph_max_uid
+ 1);
955 if (inline_edge_summary_vec
.length () <= (unsigned) cgraph_edge_max_uid
)
956 inline_edge_summary_vec
.safe_grow_cleared (cgraph_edge_max_uid
+ 1);
957 if (!edge_predicate_pool
)
958 edge_predicate_pool
= create_alloc_pool ("edge predicates",
959 sizeof (struct predicate
), 10);
962 /* We are called multiple time for given function; clear
963 data from previous run so they are not cumulated. */
966 reset_inline_edge_summary (struct cgraph_edge
*e
)
968 if (e
->uid
< (int) inline_edge_summary_vec
.length ())
970 struct inline_edge_summary
*es
= inline_edge_summary (e
);
972 es
->call_stmt_size
= es
->call_stmt_time
= 0;
974 pool_free (edge_predicate_pool
, es
->predicate
);
975 es
->predicate
= NULL
;
976 es
->param
.release ();
980 /* We are called multiple time for given function; clear
981 data from previous run so they are not cumulated. */
984 reset_inline_summary (struct cgraph_node
*node
)
986 struct inline_summary
*info
= inline_summary (node
);
987 struct cgraph_edge
*e
;
989 info
->self_size
= info
->self_time
= 0;
990 info
->estimated_stack_size
= 0;
991 info
->estimated_self_stack_size
= 0;
992 info
->stack_frame_offset
= 0;
997 if (info
->loop_iterations
)
999 pool_free (edge_predicate_pool
, info
->loop_iterations
);
1000 info
->loop_iterations
= NULL
;
1002 if (info
->loop_stride
)
1004 pool_free (edge_predicate_pool
, info
->loop_stride
);
1005 info
->loop_stride
= NULL
;
1007 if (info
->array_index
)
1009 pool_free (edge_predicate_pool
, info
->array_index
);
1010 info
->array_index
= NULL
;
1012 vec_free (info
->conds
);
1013 vec_free (info
->entry
);
1014 for (e
= node
->callees
; e
; e
= e
->next_callee
)
1015 reset_inline_edge_summary (e
);
1016 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
1017 reset_inline_edge_summary (e
);
1020 /* Hook that is called by cgraph.c when a node is removed. */
1023 inline_node_removal_hook (struct cgraph_node
*node
,
1024 void *data ATTRIBUTE_UNUSED
)
1026 struct inline_summary
*info
;
1027 if (vec_safe_length (inline_summary_vec
) <= (unsigned) node
->uid
)
1029 info
= inline_summary (node
);
1030 reset_inline_summary (node
);
1031 memset (info
, 0, sizeof (inline_summary_t
));
1034 /* Remap predicate P of former function to be predicate of duplicated functoin.
1035 POSSIBLE_TRUTHS is clause of possible truths in the duplicated node,
1036 INFO is inline summary of the duplicated node. */
1038 static struct predicate
1039 remap_predicate_after_duplication (struct predicate
*p
,
1040 clause_t possible_truths
,
1041 struct inline_summary
*info
)
1043 struct predicate new_predicate
= true_predicate ();
1045 for (j
= 0; p
->clause
[j
]; j
++)
1046 if (!(possible_truths
& p
->clause
[j
]))
1048 new_predicate
= false_predicate ();
1052 add_clause (info
->conds
, &new_predicate
,
1053 possible_truths
& p
->clause
[j
]);
1054 return new_predicate
;
1057 /* Same as remap_predicate_after_duplication but handle hint predicate *P.
1058 Additionally care about allocating new memory slot for updated predicate
1059 and set it to NULL when it becomes true or false (and thus uninteresting).
1063 remap_hint_predicate_after_duplication (struct predicate
**p
,
1064 clause_t possible_truths
,
1065 struct inline_summary
*info
)
1067 struct predicate new_predicate
;
1072 new_predicate
= remap_predicate_after_duplication (*p
,
1073 possible_truths
, info
);
1074 /* We do not want to free previous predicate; it is used by node origin. */
1076 set_hint_predicate (p
, new_predicate
);
1080 /* Hook that is called by cgraph.c when a node is duplicated. */
1083 inline_node_duplication_hook (struct cgraph_node
*src
,
1084 struct cgraph_node
*dst
,
1085 ATTRIBUTE_UNUSED
void *data
)
1087 struct inline_summary
*info
;
1088 inline_summary_alloc ();
1089 info
= inline_summary (dst
);
1090 memcpy (info
, inline_summary (src
), sizeof (struct inline_summary
));
1091 /* TODO: as an optimization, we may avoid copying conditions
1092 that are known to be false or true. */
1093 info
->conds
= vec_safe_copy (info
->conds
);
1095 /* When there are any replacements in the function body, see if we can figure
1096 out that something was optimized out. */
1097 if (ipa_node_params_vector
.exists () && dst
->clone
.tree_map
)
1099 vec
<size_time_entry
, va_gc
> *entry
= info
->entry
;
1100 /* Use SRC parm info since it may not be copied yet. */
1101 struct ipa_node_params
*parms_info
= IPA_NODE_REF (src
);
1102 vec
<tree
> known_vals
= vNULL
;
1103 int count
= ipa_get_param_count (parms_info
);
1105 clause_t possible_truths
;
1106 struct predicate true_pred
= true_predicate ();
1108 int optimized_out_size
= 0;
1109 bool inlined_to_p
= false;
1110 struct cgraph_edge
*edge
;
1113 known_vals
.safe_grow_cleared (count
);
1114 for (i
= 0; i
< count
; i
++)
1116 struct ipa_replace_map
*r
;
1118 for (j
= 0; vec_safe_iterate (dst
->clone
.tree_map
, j
, &r
); j
++)
1120 if (((!r
->old_tree
&& r
->parm_num
== i
)
1121 || (r
->old_tree
&& r
->old_tree
== ipa_get_param (parms_info
, i
)))
1122 && r
->replace_p
&& !r
->ref_p
)
1124 known_vals
[i
] = r
->new_tree
;
1129 possible_truths
= evaluate_conditions_for_known_args (dst
, false,
1132 known_vals
.release ();
1134 account_size_time (info
, 0, 0, &true_pred
);
1136 /* Remap size_time vectors.
1137 Simplify the predicate by prunning out alternatives that are known
1139 TODO: as on optimization, we can also eliminate conditions known
1141 for (i
= 0; vec_safe_iterate (entry
, i
, &e
); i
++)
1143 struct predicate new_predicate
;
1144 new_predicate
= remap_predicate_after_duplication (&e
->predicate
,
1147 if (false_predicate_p (&new_predicate
))
1148 optimized_out_size
+= e
->size
;
1150 account_size_time (info
, e
->size
, e
->time
, &new_predicate
);
1153 /* Remap edge predicates with the same simplification as above.
1154 Also copy constantness arrays. */
1155 for (edge
= dst
->callees
; edge
; edge
= edge
->next_callee
)
1157 struct predicate new_predicate
;
1158 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
1160 if (!edge
->inline_failed
)
1161 inlined_to_p
= true;
1164 new_predicate
= remap_predicate_after_duplication (es
->predicate
,
1167 if (false_predicate_p (&new_predicate
)
1168 && !false_predicate_p (es
->predicate
))
1170 optimized_out_size
+= es
->call_stmt_size
* INLINE_SIZE_SCALE
;
1171 edge
->frequency
= 0;
1173 edge_set_predicate (edge
, &new_predicate
);
1176 /* Remap indirect edge predicates with the same simplificaiton as above.
1177 Also copy constantness arrays. */
1178 for (edge
= dst
->indirect_calls
; edge
; edge
= edge
->next_callee
)
1180 struct predicate new_predicate
;
1181 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
1183 gcc_checking_assert (edge
->inline_failed
);
1186 new_predicate
= remap_predicate_after_duplication (es
->predicate
,
1189 if (false_predicate_p (&new_predicate
)
1190 && !false_predicate_p (es
->predicate
))
1192 optimized_out_size
+= es
->call_stmt_size
* INLINE_SIZE_SCALE
;
1193 edge
->frequency
= 0;
1195 edge_set_predicate (edge
, &new_predicate
);
1197 remap_hint_predicate_after_duplication (&info
->loop_iterations
,
1198 possible_truths
, info
);
1199 remap_hint_predicate_after_duplication (&info
->loop_stride
,
1200 possible_truths
, info
);
1201 remap_hint_predicate_after_duplication (&info
->array_index
,
1202 possible_truths
, info
);
1204 /* If inliner or someone after inliner will ever start producing
1205 non-trivial clones, we will get trouble with lack of information
1206 about updating self sizes, because size vectors already contains
1207 sizes of the calees. */
1208 gcc_assert (!inlined_to_p
|| !optimized_out_size
);
1212 info
->entry
= vec_safe_copy (info
->entry
);
1213 if (info
->loop_iterations
)
1215 predicate p
= *info
->loop_iterations
;
1216 info
->loop_iterations
= NULL
;
1217 set_hint_predicate (&info
->loop_iterations
, p
);
1219 if (info
->loop_stride
)
1221 predicate p
= *info
->loop_stride
;
1222 info
->loop_stride
= NULL
;
1223 set_hint_predicate (&info
->loop_stride
, p
);
1225 if (info
->array_index
)
1227 predicate p
= *info
->array_index
;
1228 info
->array_index
= NULL
;
1229 set_hint_predicate (&info
->array_index
, p
);
1232 inline_update_overall_summary (dst
);
1236 /* Hook that is called by cgraph.c when a node is duplicated. */
1239 inline_edge_duplication_hook (struct cgraph_edge
*src
,
1240 struct cgraph_edge
*dst
,
1241 ATTRIBUTE_UNUSED
void *data
)
1243 struct inline_edge_summary
*info
;
1244 struct inline_edge_summary
*srcinfo
;
1245 inline_summary_alloc ();
1246 info
= inline_edge_summary (dst
);
1247 srcinfo
= inline_edge_summary (src
);
1248 memcpy (info
, srcinfo
, sizeof (struct inline_edge_summary
));
1249 info
->predicate
= NULL
;
1250 edge_set_predicate (dst
, srcinfo
->predicate
);
1251 info
->param
= srcinfo
->param
.copy ();
1255 /* Keep edge cache consistent across edge removal. */
1258 inline_edge_removal_hook (struct cgraph_edge
*edge
,
1259 void *data ATTRIBUTE_UNUSED
)
1261 if (edge_growth_cache
.exists ())
1262 reset_edge_growth_cache (edge
);
1263 reset_inline_edge_summary (edge
);
1267 /* Initialize growth caches. */
1270 initialize_growth_caches (void)
1272 if (cgraph_edge_max_uid
)
1273 edge_growth_cache
.safe_grow_cleared (cgraph_edge_max_uid
);
1275 node_growth_cache
.safe_grow_cleared (cgraph_max_uid
);
1279 /* Free growth caches. */
1282 free_growth_caches (void)
1284 edge_growth_cache
.release ();
1285 node_growth_cache
.release ();
1289 /* Dump edge summaries associated to NODE and recursively to all clones.
1290 Indent by INDENT. */
1293 dump_inline_edge_summary (FILE *f
, int indent
, struct cgraph_node
*node
,
1294 struct inline_summary
*info
)
1296 struct cgraph_edge
*edge
;
1297 for (edge
= node
->callees
; edge
; edge
= edge
->next_callee
)
1299 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
1300 struct cgraph_node
*callee
=
1301 cgraph_function_or_thunk_node (edge
->callee
, NULL
);
1305 "%*s%s/%i %s\n%*s loop depth:%2i freq:%4i size:%2i"
1306 " time: %2i callee size:%2i stack:%2i",
1307 indent
, "", callee
->name (), callee
->order
,
1308 !edge
->inline_failed
1309 ? "inlined" : cgraph_inline_failed_string (edge
-> inline_failed
),
1310 indent
, "", es
->loop_depth
, edge
->frequency
,
1311 es
->call_stmt_size
, es
->call_stmt_time
,
1312 (int) inline_summary (callee
)->size
/ INLINE_SIZE_SCALE
,
1313 (int) inline_summary (callee
)->estimated_stack_size
);
1317 fprintf (f
, " predicate: ");
1318 dump_predicate (f
, info
->conds
, es
->predicate
);
1322 if (es
->param
.exists ())
1323 for (i
= 0; i
< (int) es
->param
.length (); i
++)
1325 int prob
= es
->param
[i
].change_prob
;
1328 fprintf (f
, "%*s op%i is compile time invariant\n",
1330 else if (prob
!= REG_BR_PROB_BASE
)
1331 fprintf (f
, "%*s op%i change %f%% of time\n", indent
+ 2, "", i
,
1332 prob
* 100.0 / REG_BR_PROB_BASE
);
1334 if (!edge
->inline_failed
)
1336 fprintf (f
, "%*sStack frame offset %i, callee self size %i,"
1337 " callee size %i\n",
1339 (int) inline_summary (callee
)->stack_frame_offset
,
1340 (int) inline_summary (callee
)->estimated_self_stack_size
,
1341 (int) inline_summary (callee
)->estimated_stack_size
);
1342 dump_inline_edge_summary (f
, indent
+ 2, callee
, info
);
1345 for (edge
= node
->indirect_calls
; edge
; edge
= edge
->next_callee
)
1347 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
1348 fprintf (f
, "%*sindirect call loop depth:%2i freq:%4i size:%2i"
1352 edge
->frequency
, es
->call_stmt_size
, es
->call_stmt_time
);
1355 fprintf (f
, "predicate: ");
1356 dump_predicate (f
, info
->conds
, es
->predicate
);
1365 dump_inline_summary (FILE *f
, struct cgraph_node
*node
)
1367 if (node
->definition
)
1369 struct inline_summary
*s
= inline_summary (node
);
1372 fprintf (f
, "Inline summary for %s/%i", node
->name (),
1374 if (DECL_DISREGARD_INLINE_LIMITS (node
->decl
))
1375 fprintf (f
, " always_inline");
1377 fprintf (f
, " inlinable");
1378 fprintf (f
, "\n self time: %i\n", s
->self_time
);
1379 fprintf (f
, " global time: %i\n", s
->time
);
1380 fprintf (f
, " self size: %i\n", s
->self_size
);
1381 fprintf (f
, " global size: %i\n", s
->size
);
1382 fprintf (f
, " self stack: %i\n",
1383 (int) s
->estimated_self_stack_size
);
1384 fprintf (f
, " global stack: %i\n", (int) s
->estimated_stack_size
);
1386 fprintf (f
, " estimated growth:%i\n", (int) s
->growth
);
1388 fprintf (f
, " In SCC: %i\n", (int) s
->scc_no
);
1389 for (i
= 0; vec_safe_iterate (s
->entry
, i
, &e
); i
++)
1391 fprintf (f
, " size:%f, time:%f, predicate:",
1392 (double) e
->size
/ INLINE_SIZE_SCALE
,
1393 (double) e
->time
/ INLINE_TIME_SCALE
);
1394 dump_predicate (f
, s
->conds
, &e
->predicate
);
1396 if (s
->loop_iterations
)
1398 fprintf (f
, " loop iterations:");
1399 dump_predicate (f
, s
->conds
, s
->loop_iterations
);
1403 fprintf (f
, " loop stride:");
1404 dump_predicate (f
, s
->conds
, s
->loop_stride
);
1408 fprintf (f
, " array index:");
1409 dump_predicate (f
, s
->conds
, s
->array_index
);
1411 fprintf (f
, " calls:\n");
1412 dump_inline_edge_summary (f
, 4, node
, s
);
1418 debug_inline_summary (struct cgraph_node
*node
)
1420 dump_inline_summary (stderr
, node
);
1424 dump_inline_summaries (FILE *f
)
1426 struct cgraph_node
*node
;
1428 FOR_EACH_DEFINED_FUNCTION (node
)
1429 if (!node
->global
.inlined_to
)
1430 dump_inline_summary (f
, node
);
1433 /* Give initial reasons why inlining would fail on EDGE. This gets either
1434 nullified or usually overwritten by more precise reasons later. */
1437 initialize_inline_failed (struct cgraph_edge
*e
)
1439 struct cgraph_node
*callee
= e
->callee
;
1441 if (e
->indirect_unknown_callee
)
1442 e
->inline_failed
= CIF_INDIRECT_UNKNOWN_CALL
;
1443 else if (!callee
->definition
)
1444 e
->inline_failed
= CIF_BODY_NOT_AVAILABLE
;
1445 else if (callee
->local
.redefined_extern_inline
)
1446 e
->inline_failed
= CIF_REDEFINED_EXTERN_INLINE
;
1447 else if (e
->call_stmt_cannot_inline_p
)
1448 e
->inline_failed
= CIF_MISMATCHED_ARGUMENTS
;
1449 else if (cfun
&& fn_contains_cilk_spawn_p (cfun
))
1450 /* We can't inline if the function is spawing a function. */
1451 e
->inline_failed
= CIF_FUNCTION_NOT_INLINABLE
;
1453 e
->inline_failed
= CIF_FUNCTION_NOT_CONSIDERED
;
1456 /* Callback of walk_aliased_vdefs. Flags that it has been invoked to the
1457 boolean variable pointed to by DATA. */
1460 mark_modified (ao_ref
*ao ATTRIBUTE_UNUSED
, tree vdef ATTRIBUTE_UNUSED
,
1463 bool *b
= (bool *) data
;
1468 /* If OP refers to value of function parameter, return the corresponding
1472 unmodified_parm_1 (gimple stmt
, tree op
)
1474 /* SSA_NAME referring to parm default def? */
1475 if (TREE_CODE (op
) == SSA_NAME
1476 && SSA_NAME_IS_DEFAULT_DEF (op
)
1477 && TREE_CODE (SSA_NAME_VAR (op
)) == PARM_DECL
)
1478 return SSA_NAME_VAR (op
);
1479 /* Non-SSA parm reference? */
1480 if (TREE_CODE (op
) == PARM_DECL
)
1482 bool modified
= false;
1485 ao_ref_init (&refd
, op
);
1486 walk_aliased_vdefs (&refd
, gimple_vuse (stmt
), mark_modified
, &modified
,
1494 /* If OP refers to value of function parameter, return the corresponding
1495 parameter. Also traverse chains of SSA register assignments. */
1498 unmodified_parm (gimple stmt
, tree op
)
1500 tree res
= unmodified_parm_1 (stmt
, op
);
1504 if (TREE_CODE (op
) == SSA_NAME
1505 && !SSA_NAME_IS_DEFAULT_DEF (op
)
1506 && gimple_assign_single_p (SSA_NAME_DEF_STMT (op
)))
1507 return unmodified_parm (SSA_NAME_DEF_STMT (op
),
1508 gimple_assign_rhs1 (SSA_NAME_DEF_STMT (op
)));
1512 /* If OP refers to a value of a function parameter or value loaded from an
1513 aggregate passed to a parameter (either by value or reference), return TRUE
1514 and store the number of the parameter to *INDEX_P and information whether
1515 and how it has been loaded from an aggregate into *AGGPOS. INFO describes
1516 the function parameters, STMT is the statement in which OP is used or
1520 unmodified_parm_or_parm_agg_item (struct ipa_node_params
*info
,
1521 gimple stmt
, tree op
, int *index_p
,
1522 struct agg_position_info
*aggpos
)
1524 tree res
= unmodified_parm_1 (stmt
, op
);
1526 gcc_checking_assert (aggpos
);
1529 *index_p
= ipa_get_param_decl_index (info
, res
);
1532 aggpos
->agg_contents
= false;
1533 aggpos
->by_ref
= false;
1537 if (TREE_CODE (op
) == SSA_NAME
)
1539 if (SSA_NAME_IS_DEFAULT_DEF (op
)
1540 || !gimple_assign_single_p (SSA_NAME_DEF_STMT (op
)))
1542 stmt
= SSA_NAME_DEF_STMT (op
);
1543 op
= gimple_assign_rhs1 (stmt
);
1544 if (!REFERENCE_CLASS_P (op
))
1545 return unmodified_parm_or_parm_agg_item (info
, stmt
, op
, index_p
,
1549 aggpos
->agg_contents
= true;
1550 return ipa_load_from_parm_agg (info
, stmt
, op
, index_p
, &aggpos
->offset
,
1554 /* See if statement might disappear after inlining.
1555 0 - means not eliminated
1556 1 - half of statements goes away
1557 2 - for sure it is eliminated.
1558 We are not terribly sophisticated, basically looking for simple abstraction
1559 penalty wrappers. */
1562 eliminated_by_inlining_prob (gimple stmt
)
1564 enum gimple_code code
= gimple_code (stmt
);
1565 enum tree_code rhs_code
;
1575 if (gimple_num_ops (stmt
) != 2)
1578 rhs_code
= gimple_assign_rhs_code (stmt
);
1580 /* Casts of parameters, loads from parameters passed by reference
1581 and stores to return value or parameters are often free after
1582 inlining dua to SRA and further combining.
1583 Assume that half of statements goes away. */
1584 if (rhs_code
== CONVERT_EXPR
1585 || rhs_code
== NOP_EXPR
1586 || rhs_code
== VIEW_CONVERT_EXPR
1587 || rhs_code
== ADDR_EXPR
1588 || gimple_assign_rhs_class (stmt
) == GIMPLE_SINGLE_RHS
)
1590 tree rhs
= gimple_assign_rhs1 (stmt
);
1591 tree lhs
= gimple_assign_lhs (stmt
);
1592 tree inner_rhs
= get_base_address (rhs
);
1593 tree inner_lhs
= get_base_address (lhs
);
1594 bool rhs_free
= false;
1595 bool lhs_free
= false;
1602 /* Reads of parameter are expected to be free. */
1603 if (unmodified_parm (stmt
, inner_rhs
))
1605 /* Match expressions of form &this->field. Those will most likely
1606 combine with something upstream after inlining. */
1607 else if (TREE_CODE (inner_rhs
) == ADDR_EXPR
)
1609 tree op
= get_base_address (TREE_OPERAND (inner_rhs
, 0));
1610 if (TREE_CODE (op
) == PARM_DECL
)
1612 else if (TREE_CODE (op
) == MEM_REF
1613 && unmodified_parm (stmt
, TREE_OPERAND (op
, 0)))
1617 /* When parameter is not SSA register because its address is taken
1618 and it is just copied into one, the statement will be completely
1619 free after inlining (we will copy propagate backward). */
1620 if (rhs_free
&& is_gimple_reg (lhs
))
1623 /* Reads of parameters passed by reference
1624 expected to be free (i.e. optimized out after inlining). */
1625 if (TREE_CODE (inner_rhs
) == MEM_REF
1626 && unmodified_parm (stmt
, TREE_OPERAND (inner_rhs
, 0)))
1629 /* Copying parameter passed by reference into gimple register is
1630 probably also going to copy propagate, but we can't be quite
1632 if (rhs_free
&& is_gimple_reg (lhs
))
1635 /* Writes to parameters, parameters passed by value and return value
1636 (either dirrectly or passed via invisible reference) are free.
1638 TODO: We ought to handle testcase like
1639 struct a {int a,b;};
1641 retrurnsturct (void)
1647 This translate into:
1662 For that we either need to copy ipa-split logic detecting writes
1664 if (TREE_CODE (inner_lhs
) == PARM_DECL
1665 || TREE_CODE (inner_lhs
) == RESULT_DECL
1666 || (TREE_CODE (inner_lhs
) == MEM_REF
1667 && (unmodified_parm (stmt
, TREE_OPERAND (inner_lhs
, 0))
1668 || (TREE_CODE (TREE_OPERAND (inner_lhs
, 0)) == SSA_NAME
1669 && SSA_NAME_VAR (TREE_OPERAND (inner_lhs
, 0))
1670 && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND
1672 0))) == RESULT_DECL
))))
1675 && (is_gimple_reg (rhs
) || is_gimple_min_invariant (rhs
)))
1677 if (lhs_free
&& rhs_free
)
1687 /* If BB ends by a conditional we can turn into predicates, attach corresponding
1688 predicates to the CFG edges. */
1691 set_cond_stmt_execution_predicate (struct ipa_node_params
*info
,
1692 struct inline_summary
*summary
,
1698 struct agg_position_info aggpos
;
1699 enum tree_code code
, inverted_code
;
1705 last
= last_stmt (bb
);
1706 if (!last
|| gimple_code (last
) != GIMPLE_COND
)
1708 if (!is_gimple_ip_invariant (gimple_cond_rhs (last
)))
1710 op
= gimple_cond_lhs (last
);
1711 /* TODO: handle conditionals like
1714 if (unmodified_parm_or_parm_agg_item (info
, last
, op
, &index
, &aggpos
))
1716 code
= gimple_cond_code (last
);
1718 = invert_tree_comparison (code
,
1719 HONOR_NANS (TYPE_MODE (TREE_TYPE (op
))));
1721 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1723 struct predicate p
= add_condition (summary
, index
, &aggpos
,
1724 e
->flags
& EDGE_TRUE_VALUE
1725 ? code
: inverted_code
,
1726 gimple_cond_rhs (last
));
1727 e
->aux
= pool_alloc (edge_predicate_pool
);
1728 *(struct predicate
*) e
->aux
= p
;
1732 if (TREE_CODE (op
) != SSA_NAME
)
1735 if (builtin_constant_p (op))
1739 Here we can predicate nonconstant_code. We can't
1740 really handle constant_code since we have no predicate
1741 for this and also the constant code is not known to be
1742 optimized away when inliner doen't see operand is constant.
1743 Other optimizers might think otherwise. */
1744 if (gimple_cond_code (last
) != NE_EXPR
1745 || !integer_zerop (gimple_cond_rhs (last
)))
1747 set_stmt
= SSA_NAME_DEF_STMT (op
);
1748 if (!gimple_call_builtin_p (set_stmt
, BUILT_IN_CONSTANT_P
)
1749 || gimple_call_num_args (set_stmt
) != 1)
1751 op2
= gimple_call_arg (set_stmt
, 0);
1752 if (!unmodified_parm_or_parm_agg_item
1753 (info
, set_stmt
, op2
, &index
, &aggpos
))
1755 FOR_EACH_EDGE (e
, ei
, bb
->succs
) if (e
->flags
& EDGE_FALSE_VALUE
)
1757 struct predicate p
= add_condition (summary
, index
, &aggpos
,
1758 IS_NOT_CONSTANT
, NULL_TREE
);
1759 e
->aux
= pool_alloc (edge_predicate_pool
);
1760 *(struct predicate
*) e
->aux
= p
;
1765 /* If BB ends by a switch we can turn into predicates, attach corresponding
1766 predicates to the CFG edges. */
1769 set_switch_stmt_execution_predicate (struct ipa_node_params
*info
,
1770 struct inline_summary
*summary
,
1776 struct agg_position_info aggpos
;
1782 last
= last_stmt (bb
);
1783 if (!last
|| gimple_code (last
) != GIMPLE_SWITCH
)
1785 op
= gimple_switch_index (last
);
1786 if (!unmodified_parm_or_parm_agg_item (info
, last
, op
, &index
, &aggpos
))
1789 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1791 e
->aux
= pool_alloc (edge_predicate_pool
);
1792 *(struct predicate
*) e
->aux
= false_predicate ();
1794 n
= gimple_switch_num_labels (last
);
1795 for (case_idx
= 0; case_idx
< n
; ++case_idx
)
1797 tree cl
= gimple_switch_label (last
, case_idx
);
1801 e
= find_edge (bb
, label_to_block (CASE_LABEL (cl
)));
1802 min
= CASE_LOW (cl
);
1803 max
= CASE_HIGH (cl
);
1805 /* For default we might want to construct predicate that none
1806 of cases is met, but it is bit hard to do not having negations
1807 of conditionals handy. */
1809 p
= true_predicate ();
1811 p
= add_condition (summary
, index
, &aggpos
, EQ_EXPR
, min
);
1814 struct predicate p1
, p2
;
1815 p1
= add_condition (summary
, index
, &aggpos
, GE_EXPR
, min
);
1816 p2
= add_condition (summary
, index
, &aggpos
, LE_EXPR
, max
);
1817 p
= and_predicates (summary
->conds
, &p1
, &p2
);
1819 *(struct predicate
*) e
->aux
1820 = or_predicates (summary
->conds
, &p
, (struct predicate
*) e
->aux
);
1825 /* For each BB in NODE attach to its AUX pointer predicate under
1826 which it is executable. */
1829 compute_bb_predicates (struct cgraph_node
*node
,
1830 struct ipa_node_params
*parms_info
,
1831 struct inline_summary
*summary
)
1833 struct function
*my_function
= DECL_STRUCT_FUNCTION (node
->decl
);
1837 FOR_EACH_BB_FN (bb
, my_function
)
1839 set_cond_stmt_execution_predicate (parms_info
, summary
, bb
);
1840 set_switch_stmt_execution_predicate (parms_info
, summary
, bb
);
1843 /* Entry block is always executable. */
1844 ENTRY_BLOCK_PTR_FOR_FUNCTION (my_function
)->aux
1845 = pool_alloc (edge_predicate_pool
);
1846 *(struct predicate
*) ENTRY_BLOCK_PTR_FOR_FUNCTION (my_function
)->aux
1847 = true_predicate ();
1849 /* A simple dataflow propagation of predicates forward in the CFG.
1850 TODO: work in reverse postorder. */
1854 FOR_EACH_BB_FN (bb
, my_function
)
1856 struct predicate p
= false_predicate ();
1859 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
1863 struct predicate this_bb_predicate
1864 = *(struct predicate
*) e
->src
->aux
;
1867 = and_predicates (summary
->conds
, &this_bb_predicate
,
1868 (struct predicate
*) e
->aux
);
1869 p
= or_predicates (summary
->conds
, &p
, &this_bb_predicate
);
1870 if (true_predicate_p (&p
))
1874 if (false_predicate_p (&p
))
1875 gcc_assert (!bb
->aux
);
1881 bb
->aux
= pool_alloc (edge_predicate_pool
);
1882 *((struct predicate
*) bb
->aux
) = p
;
1884 else if (!predicates_equal_p (&p
, (struct predicate
*) bb
->aux
))
1887 *((struct predicate
*) bb
->aux
) = p
;
1895 /* We keep info about constantness of SSA names. */
1897 typedef struct predicate predicate_t
;
1898 /* Return predicate specifying when the STMT might have result that is not
1899 a compile time constant. */
1901 static struct predicate
1902 will_be_nonconstant_expr_predicate (struct ipa_node_params
*info
,
1903 struct inline_summary
*summary
,
1905 vec
<predicate_t
> nonconstant_names
)
1910 while (UNARY_CLASS_P (expr
))
1911 expr
= TREE_OPERAND (expr
, 0);
1913 parm
= unmodified_parm (NULL
, expr
);
1914 if (parm
&& (index
= ipa_get_param_decl_index (info
, parm
)) >= 0)
1915 return add_condition (summary
, index
, NULL
, CHANGED
, NULL_TREE
);
1916 if (is_gimple_min_invariant (expr
))
1917 return false_predicate ();
1918 if (TREE_CODE (expr
) == SSA_NAME
)
1919 return nonconstant_names
[SSA_NAME_VERSION (expr
)];
1920 if (BINARY_CLASS_P (expr
) || COMPARISON_CLASS_P (expr
))
1922 struct predicate p1
= will_be_nonconstant_expr_predicate
1923 (info
, summary
, TREE_OPERAND (expr
, 0),
1925 struct predicate p2
;
1926 if (true_predicate_p (&p1
))
1928 p2
= will_be_nonconstant_expr_predicate (info
, summary
,
1929 TREE_OPERAND (expr
, 1),
1931 return or_predicates (summary
->conds
, &p1
, &p2
);
1933 else if (TREE_CODE (expr
) == COND_EXPR
)
1935 struct predicate p1
= will_be_nonconstant_expr_predicate
1936 (info
, summary
, TREE_OPERAND (expr
, 0),
1938 struct predicate p2
;
1939 if (true_predicate_p (&p1
))
1941 p2
= will_be_nonconstant_expr_predicate (info
, summary
,
1942 TREE_OPERAND (expr
, 1),
1944 if (true_predicate_p (&p2
))
1946 p1
= or_predicates (summary
->conds
, &p1
, &p2
);
1947 p2
= will_be_nonconstant_expr_predicate (info
, summary
,
1948 TREE_OPERAND (expr
, 2),
1950 return or_predicates (summary
->conds
, &p1
, &p2
);
1957 return false_predicate ();
1961 /* Return predicate specifying when the STMT might have result that is not
1962 a compile time constant. */
1964 static struct predicate
1965 will_be_nonconstant_predicate (struct ipa_node_params
*info
,
1966 struct inline_summary
*summary
,
1968 vec
<predicate_t
> nonconstant_names
)
1970 struct predicate p
= true_predicate ();
1973 struct predicate op_non_const
;
1976 struct agg_position_info aggpos
;
1978 /* What statments might be optimized away
1979 when their arguments are constant
1980 TODO: also trivial builtins.
1981 builtin_constant_p is already handled later. */
1982 if (gimple_code (stmt
) != GIMPLE_ASSIGN
1983 && gimple_code (stmt
) != GIMPLE_COND
1984 && gimple_code (stmt
) != GIMPLE_SWITCH
)
1987 /* Stores will stay anyway. */
1988 if (gimple_store_p (stmt
))
1991 is_load
= gimple_assign_load_p (stmt
);
1993 /* Loads can be optimized when the value is known. */
1997 gcc_assert (gimple_assign_single_p (stmt
));
1998 op
= gimple_assign_rhs1 (stmt
);
1999 if (!unmodified_parm_or_parm_agg_item (info
, stmt
, op
, &base_index
,
2006 /* See if we understand all operands before we start
2007 adding conditionals. */
2008 FOR_EACH_SSA_TREE_OPERAND (use
, stmt
, iter
, SSA_OP_USE
)
2010 tree parm
= unmodified_parm (stmt
, use
);
2011 /* For arguments we can build a condition. */
2012 if (parm
&& ipa_get_param_decl_index (info
, parm
) >= 0)
2014 if (TREE_CODE (use
) != SSA_NAME
)
2016 /* If we know when operand is constant,
2017 we still can say something useful. */
2018 if (!true_predicate_p (&nonconstant_names
[SSA_NAME_VERSION (use
)]))
2025 add_condition (summary
, base_index
, &aggpos
, CHANGED
, NULL
);
2027 op_non_const
= false_predicate ();
2028 FOR_EACH_SSA_TREE_OPERAND (use
, stmt
, iter
, SSA_OP_USE
)
2030 tree parm
= unmodified_parm (stmt
, use
);
2033 if (parm
&& (index
= ipa_get_param_decl_index (info
, parm
)) >= 0)
2035 if (index
!= base_index
)
2036 p
= add_condition (summary
, index
, NULL
, CHANGED
, NULL_TREE
);
2041 p
= nonconstant_names
[SSA_NAME_VERSION (use
)];
2042 op_non_const
= or_predicates (summary
->conds
, &p
, &op_non_const
);
2044 if (gimple_code (stmt
) == GIMPLE_ASSIGN
2045 && TREE_CODE (gimple_assign_lhs (stmt
)) == SSA_NAME
)
2046 nonconstant_names
[SSA_NAME_VERSION (gimple_assign_lhs (stmt
))]
2048 return op_non_const
;
2051 struct record_modified_bb_info
2057 /* Callback of walk_aliased_vdefs. Records basic blocks where the value may be
2058 set except for info->stmt. */
2061 record_modified (ao_ref
*ao ATTRIBUTE_UNUSED
, tree vdef
, void *data
)
2063 struct record_modified_bb_info
*info
=
2064 (struct record_modified_bb_info
*) data
;
2065 if (SSA_NAME_DEF_STMT (vdef
) == info
->stmt
)
2067 bitmap_set_bit (info
->bb_set
,
2068 SSA_NAME_IS_DEFAULT_DEF (vdef
)
2069 ? ENTRY_BLOCK_PTR
->index
2070 : gimple_bb (SSA_NAME_DEF_STMT (vdef
))->index
);
2074 /* Return probability (based on REG_BR_PROB_BASE) that I-th parameter of STMT
2075 will change since last invocation of STMT.
2077 Value 0 is reserved for compile time invariants.
2078 For common parameters it is REG_BR_PROB_BASE. For loop invariants it
2079 ought to be REG_BR_PROB_BASE / estimated_iters. */
2082 param_change_prob (gimple stmt
, int i
)
2084 tree op
= gimple_call_arg (stmt
, i
);
2085 basic_block bb
= gimple_bb (stmt
);
2088 /* Global invariants neve change. */
2089 if (is_gimple_min_invariant (op
))
2091 /* We would have to do non-trivial analysis to really work out what
2092 is the probability of value to change (i.e. when init statement
2093 is in a sibling loop of the call).
2095 We do an conservative estimate: when call is executed N times more often
2096 than the statement defining value, we take the frequency 1/N. */
2097 if (TREE_CODE (op
) == SSA_NAME
)
2102 return REG_BR_PROB_BASE
;
2104 if (SSA_NAME_IS_DEFAULT_DEF (op
))
2105 init_freq
= ENTRY_BLOCK_PTR
->frequency
;
2107 init_freq
= gimple_bb (SSA_NAME_DEF_STMT (op
))->frequency
;
2111 if (init_freq
< bb
->frequency
)
2112 return MAX (GCOV_COMPUTE_SCALE (init_freq
, bb
->frequency
), 1);
2114 return REG_BR_PROB_BASE
;
2117 base
= get_base_address (op
);
2122 struct record_modified_bb_info info
;
2125 tree init
= ctor_for_folding (base
);
2127 if (init
!= error_mark_node
)
2130 return REG_BR_PROB_BASE
;
2131 ao_ref_init (&refd
, op
);
2133 info
.bb_set
= BITMAP_ALLOC (NULL
);
2134 walk_aliased_vdefs (&refd
, gimple_vuse (stmt
), record_modified
, &info
,
2136 if (bitmap_bit_p (info
.bb_set
, bb
->index
))
2138 BITMAP_FREE (info
.bb_set
);
2139 return REG_BR_PROB_BASE
;
2142 /* Assume that every memory is initialized at entry.
2143 TODO: Can we easilly determine if value is always defined
2144 and thus we may skip entry block? */
2145 if (ENTRY_BLOCK_PTR
->frequency
)
2146 max
= ENTRY_BLOCK_PTR
->frequency
;
2150 EXECUTE_IF_SET_IN_BITMAP (info
.bb_set
, 0, index
, bi
)
2151 max
= MIN (max
, BASIC_BLOCK (index
)->frequency
);
2153 BITMAP_FREE (info
.bb_set
);
2154 if (max
< bb
->frequency
)
2155 return MAX (GCOV_COMPUTE_SCALE (max
, bb
->frequency
), 1);
2157 return REG_BR_PROB_BASE
;
2159 return REG_BR_PROB_BASE
;
2162 /* Find whether a basic block BB is the final block of a (half) diamond CFG
2163 sub-graph and if the predicate the condition depends on is known. If so,
2164 return true and store the pointer the predicate in *P. */
2167 phi_result_unknown_predicate (struct ipa_node_params
*info
,
2168 struct inline_summary
*summary
, basic_block bb
,
2169 struct predicate
*p
,
2170 vec
<predicate_t
> nonconstant_names
)
2174 basic_block first_bb
= NULL
;
2177 if (single_pred_p (bb
))
2179 *p
= false_predicate ();
2183 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
2185 if (single_succ_p (e
->src
))
2187 if (!single_pred_p (e
->src
))
2190 first_bb
= single_pred (e
->src
);
2191 else if (single_pred (e
->src
) != first_bb
)
2198 else if (e
->src
!= first_bb
)
2206 stmt
= last_stmt (first_bb
);
2208 || gimple_code (stmt
) != GIMPLE_COND
2209 || !is_gimple_ip_invariant (gimple_cond_rhs (stmt
)))
2212 *p
= will_be_nonconstant_expr_predicate (info
, summary
,
2213 gimple_cond_lhs (stmt
),
2215 if (true_predicate_p (p
))
2221 /* Given a PHI statement in a function described by inline properties SUMMARY
2222 and *P being the predicate describing whether the selected PHI argument is
2223 known, store a predicate for the result of the PHI statement into
2224 NONCONSTANT_NAMES, if possible. */
2227 predicate_for_phi_result (struct inline_summary
*summary
, gimple phi
,
2228 struct predicate
*p
,
2229 vec
<predicate_t
> nonconstant_names
)
2233 for (i
= 0; i
< gimple_phi_num_args (phi
); i
++)
2235 tree arg
= gimple_phi_arg (phi
, i
)->def
;
2236 if (!is_gimple_min_invariant (arg
))
2238 gcc_assert (TREE_CODE (arg
) == SSA_NAME
);
2239 *p
= or_predicates (summary
->conds
, p
,
2240 &nonconstant_names
[SSA_NAME_VERSION (arg
)]);
2241 if (true_predicate_p (p
))
2246 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2248 fprintf (dump_file
, "\t\tphi predicate: ");
2249 dump_predicate (dump_file
, summary
->conds
, p
);
2251 nonconstant_names
[SSA_NAME_VERSION (gimple_phi_result (phi
))] = *p
;
2254 /* Return predicate specifying when array index in access OP becomes non-constant. */
2256 static struct predicate
2257 array_index_predicate (struct inline_summary
*info
,
2258 vec
< predicate_t
> nonconstant_names
, tree op
)
2260 struct predicate p
= false_predicate ();
2261 while (handled_component_p (op
))
2263 if (TREE_CODE (op
) == ARRAY_REF
|| TREE_CODE (op
) == ARRAY_RANGE_REF
)
2265 if (TREE_CODE (TREE_OPERAND (op
, 1)) == SSA_NAME
)
2266 p
= or_predicates (info
->conds
, &p
,
2267 &nonconstant_names
[SSA_NAME_VERSION
2268 (TREE_OPERAND (op
, 1))]);
2270 op
= TREE_OPERAND (op
, 0);
2275 /* For a typical usage of __builtin_expect (a<b, 1), we
2276 may introduce an extra relation stmt:
2277 With the builtin, we have
2280 t3 = __builtin_expect (t2, 1);
2283 Without the builtin, we have
2286 This affects the size/time estimation and may have
2287 an impact on the earlier inlining.
2288 Here find this pattern and fix it up later. */
2291 find_foldable_builtin_expect (basic_block bb
)
2293 gimple_stmt_iterator bsi
;
2295 for (bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
2297 gimple stmt
= gsi_stmt (bsi
);
2298 if (gimple_call_builtin_p (stmt
, BUILT_IN_EXPECT
))
2300 tree var
= gimple_call_lhs (stmt
);
2301 tree arg
= gimple_call_arg (stmt
, 0);
2302 use_operand_p use_p
;
2309 gcc_assert (TREE_CODE (var
) == SSA_NAME
);
2311 while (TREE_CODE (arg
) == SSA_NAME
)
2313 gimple stmt_tmp
= SSA_NAME_DEF_STMT (arg
);
2314 if (!is_gimple_assign (stmt_tmp
))
2316 switch (gimple_assign_rhs_code (stmt_tmp
))
2335 arg
= gimple_assign_rhs1 (stmt_tmp
);
2338 if (match
&& single_imm_use (var
, &use_p
, &use_stmt
)
2339 && gimple_code (use_stmt
) == GIMPLE_COND
)
2346 /* Compute function body size parameters for NODE.
2347 When EARLY is true, we compute only simple summaries without
2348 non-trivial predicates to drive the early inliner. */
2351 estimate_function_body_sizes (struct cgraph_node
*node
, bool early
)
2354 /* Estimate static overhead for function prologue/epilogue and alignment. */
2356 /* Benefits are scaled by probability of elimination that is in range
2359 gimple_stmt_iterator bsi
;
2360 struct function
*my_function
= DECL_STRUCT_FUNCTION (node
->decl
);
2362 struct inline_summary
*info
= inline_summary (node
);
2363 struct predicate bb_predicate
;
2364 struct ipa_node_params
*parms_info
= NULL
;
2365 vec
<predicate_t
> nonconstant_names
= vNULL
;
2368 predicate array_index
= true_predicate ();
2369 gimple fix_builtin_expect_stmt
;
2374 if (optimize
&& !early
)
2376 calculate_dominance_info (CDI_DOMINATORS
);
2377 loop_optimizer_init (LOOPS_NORMAL
| LOOPS_HAVE_RECORDED_EXITS
);
2379 if (ipa_node_params_vector
.exists ())
2381 parms_info
= IPA_NODE_REF (node
);
2382 nonconstant_names
.safe_grow_cleared
2383 (SSANAMES (my_function
)->length ());
2388 fprintf (dump_file
, "\nAnalyzing function body size: %s\n",
2391 /* When we run into maximal number of entries, we assign everything to the
2392 constant truth case. Be sure to have it in list. */
2393 bb_predicate
= true_predicate ();
2394 account_size_time (info
, 0, 0, &bb_predicate
);
2396 bb_predicate
= not_inlined_predicate ();
2397 account_size_time (info
, 2 * INLINE_SIZE_SCALE
, 0, &bb_predicate
);
2399 gcc_assert (my_function
&& my_function
->cfg
);
2401 compute_bb_predicates (node
, parms_info
, info
);
2402 gcc_assert (cfun
== my_function
);
2403 order
= XNEWVEC (int, n_basic_blocks_for_fn (cfun
));
2404 nblocks
= pre_and_rev_post_order_compute (NULL
, order
, false);
2405 for (n
= 0; n
< nblocks
; n
++)
2407 bb
= BASIC_BLOCK (order
[n
]);
2408 freq
= compute_call_stmt_bb_frequency (node
->decl
, bb
);
2410 /* TODO: Obviously predicates can be propagated down across CFG. */
2414 bb_predicate
= *(struct predicate
*) bb
->aux
;
2416 bb_predicate
= false_predicate ();
2419 bb_predicate
= true_predicate ();
2421 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2423 fprintf (dump_file
, "\n BB %i predicate:", bb
->index
);
2424 dump_predicate (dump_file
, info
->conds
, &bb_predicate
);
2427 if (parms_info
&& nonconstant_names
.exists ())
2429 struct predicate phi_predicate
;
2430 bool first_phi
= true;
2432 for (bsi
= gsi_start_phis (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
2435 && !phi_result_unknown_predicate (parms_info
, info
, bb
,
2440 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2442 fprintf (dump_file
, " ");
2443 print_gimple_stmt (dump_file
, gsi_stmt (bsi
), 0, 0);
2445 predicate_for_phi_result (info
, gsi_stmt (bsi
), &phi_predicate
,
2450 fix_builtin_expect_stmt
= find_foldable_builtin_expect (bb
);
2452 for (bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
2454 gimple stmt
= gsi_stmt (bsi
);
2455 int this_size
= estimate_num_insns (stmt
, &eni_size_weights
);
2456 int this_time
= estimate_num_insns (stmt
, &eni_time_weights
);
2458 struct predicate will_be_nonconstant
;
2460 /* This relation stmt should be folded after we remove
2461 buildin_expect call. Adjust the cost here. */
2462 if (stmt
== fix_builtin_expect_stmt
)
2468 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2470 fprintf (dump_file
, " ");
2471 print_gimple_stmt (dump_file
, stmt
, 0, 0);
2472 fprintf (dump_file
, "\t\tfreq:%3.2f size:%3i time:%3i\n",
2473 ((double) freq
) / CGRAPH_FREQ_BASE
, this_size
,
2477 if (gimple_assign_load_p (stmt
) && nonconstant_names
.exists ())
2479 struct predicate this_array_index
;
2481 array_index_predicate (info
, nonconstant_names
,
2482 gimple_assign_rhs1 (stmt
));
2483 if (!false_predicate_p (&this_array_index
))
2485 and_predicates (info
->conds
, &array_index
,
2488 if (gimple_store_p (stmt
) && nonconstant_names
.exists ())
2490 struct predicate this_array_index
;
2492 array_index_predicate (info
, nonconstant_names
,
2493 gimple_get_lhs (stmt
));
2494 if (!false_predicate_p (&this_array_index
))
2496 and_predicates (info
->conds
, &array_index
,
2501 if (is_gimple_call (stmt
))
2503 struct cgraph_edge
*edge
= cgraph_edge (node
, stmt
);
2504 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
2506 /* Special case: results of BUILT_IN_CONSTANT_P will be always
2507 resolved as constant. We however don't want to optimize
2508 out the cgraph edges. */
2509 if (nonconstant_names
.exists ()
2510 && gimple_call_builtin_p (stmt
, BUILT_IN_CONSTANT_P
)
2511 && gimple_call_lhs (stmt
)
2512 && TREE_CODE (gimple_call_lhs (stmt
)) == SSA_NAME
)
2514 struct predicate false_p
= false_predicate ();
2515 nonconstant_names
[SSA_NAME_VERSION (gimple_call_lhs (stmt
))]
2518 if (ipa_node_params_vector
.exists ())
2520 int count
= gimple_call_num_args (stmt
);
2524 es
->param
.safe_grow_cleared (count
);
2525 for (i
= 0; i
< count
; i
++)
2527 int prob
= param_change_prob (stmt
, i
);
2528 gcc_assert (prob
>= 0 && prob
<= REG_BR_PROB_BASE
);
2529 es
->param
[i
].change_prob
= prob
;
2533 es
->call_stmt_size
= this_size
;
2534 es
->call_stmt_time
= this_time
;
2535 es
->loop_depth
= bb_loop_depth (bb
);
2536 edge_set_predicate (edge
, &bb_predicate
);
2539 /* TODO: When conditional jump or swithc is known to be constant, but
2540 we did not translate it into the predicates, we really can account
2541 just maximum of the possible paths. */
2544 = will_be_nonconstant_predicate (parms_info
, info
,
2545 stmt
, nonconstant_names
);
2546 if (this_time
|| this_size
)
2552 prob
= eliminated_by_inlining_prob (stmt
);
2553 if (prob
== 1 && dump_file
&& (dump_flags
& TDF_DETAILS
))
2555 "\t\t50%% will be eliminated by inlining\n");
2556 if (prob
== 2 && dump_file
&& (dump_flags
& TDF_DETAILS
))
2557 fprintf (dump_file
, "\t\tWill be eliminated by inlining\n");
2560 p
= and_predicates (info
->conds
, &bb_predicate
,
2561 &will_be_nonconstant
);
2563 p
= true_predicate ();
2565 if (!false_predicate_p (&p
))
2569 if (time
> MAX_TIME
* INLINE_TIME_SCALE
)
2570 time
= MAX_TIME
* INLINE_TIME_SCALE
;
2573 /* We account everything but the calls. Calls have their own
2574 size/time info attached to cgraph edges. This is necessary
2575 in order to make the cost disappear after inlining. */
2576 if (!is_gimple_call (stmt
))
2580 struct predicate ip
= not_inlined_predicate ();
2581 ip
= and_predicates (info
->conds
, &ip
, &p
);
2582 account_size_time (info
, this_size
* prob
,
2583 this_time
* prob
, &ip
);
2586 account_size_time (info
, this_size
* (2 - prob
),
2587 this_time
* (2 - prob
), &p
);
2590 gcc_assert (time
>= 0);
2591 gcc_assert (size
>= 0);
2595 set_hint_predicate (&inline_summary (node
)->array_index
, array_index
);
2596 time
= (time
+ CGRAPH_FREQ_BASE
/ 2) / CGRAPH_FREQ_BASE
;
2597 if (time
> MAX_TIME
)
2601 if (!early
&& nonconstant_names
.exists ())
2604 predicate loop_iterations
= true_predicate ();
2605 predicate loop_stride
= true_predicate ();
2607 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2608 flow_loops_dump (dump_file
, NULL
, 0);
2610 FOR_EACH_LOOP (loop
, 0)
2615 struct tree_niter_desc niter_desc
;
2616 basic_block
*body
= get_loop_body (loop
);
2617 bb_predicate
= *(struct predicate
*) loop
->header
->aux
;
2619 exits
= get_loop_exit_edges (loop
);
2620 FOR_EACH_VEC_ELT (exits
, j
, ex
)
2621 if (number_of_iterations_exit (loop
, ex
, &niter_desc
, false)
2622 && !is_gimple_min_invariant (niter_desc
.niter
))
2624 predicate will_be_nonconstant
2625 = will_be_nonconstant_expr_predicate (parms_info
, info
,
2628 if (!true_predicate_p (&will_be_nonconstant
))
2629 will_be_nonconstant
= and_predicates (info
->conds
,
2631 &will_be_nonconstant
);
2632 if (!true_predicate_p (&will_be_nonconstant
)
2633 && !false_predicate_p (&will_be_nonconstant
))
2634 /* This is slightly inprecise. We may want to represent each
2635 loop with independent predicate. */
2637 and_predicates (info
->conds
, &loop_iterations
,
2638 &will_be_nonconstant
);
2642 for (i
= 0; i
< loop
->num_nodes
; i
++)
2644 gimple_stmt_iterator gsi
;
2645 bb_predicate
= *(struct predicate
*) body
[i
]->aux
;
2646 for (gsi
= gsi_start_bb (body
[i
]); !gsi_end_p (gsi
);
2649 gimple stmt
= gsi_stmt (gsi
);
2654 FOR_EACH_SSA_TREE_OPERAND (use
, stmt
, iter
, SSA_OP_USE
)
2656 predicate will_be_nonconstant
;
2659 (loop
, loop_containing_stmt (stmt
), use
, &iv
, true)
2660 || is_gimple_min_invariant (iv
.step
))
2663 = will_be_nonconstant_expr_predicate (parms_info
, info
,
2666 if (!true_predicate_p (&will_be_nonconstant
))
2668 = and_predicates (info
->conds
,
2670 &will_be_nonconstant
);
2671 if (!true_predicate_p (&will_be_nonconstant
)
2672 && !false_predicate_p (&will_be_nonconstant
))
2673 /* This is slightly inprecise. We may want to represent
2674 each loop with independent predicate. */
2676 and_predicates (info
->conds
, &loop_stride
,
2677 &will_be_nonconstant
);
2683 set_hint_predicate (&inline_summary (node
)->loop_iterations
,
2685 set_hint_predicate (&inline_summary (node
)->loop_stride
, loop_stride
);
2688 FOR_ALL_BB_FN (bb
, my_function
)
2694 pool_free (edge_predicate_pool
, bb
->aux
);
2696 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
2699 pool_free (edge_predicate_pool
, e
->aux
);
2703 inline_summary (node
)->self_time
= time
;
2704 inline_summary (node
)->self_size
= size
;
2705 nonconstant_names
.release ();
2706 if (optimize
&& !early
)
2708 loop_optimizer_finalize ();
2709 free_dominance_info (CDI_DOMINATORS
);
2713 fprintf (dump_file
, "\n");
2714 dump_inline_summary (dump_file
, node
);
2719 /* Compute parameters of functions used by inliner.
2720 EARLY is true when we compute parameters for the early inliner */
2723 compute_inline_parameters (struct cgraph_node
*node
, bool early
)
2725 HOST_WIDE_INT self_stack_size
;
2726 struct cgraph_edge
*e
;
2727 struct inline_summary
*info
;
2729 gcc_assert (!node
->global
.inlined_to
);
2731 inline_summary_alloc ();
2733 info
= inline_summary (node
);
2734 reset_inline_summary (node
);
2736 /* FIXME: Thunks are inlinable, but tree-inline don't know how to do that.
2737 Once this happen, we will need to more curefully predict call
2739 if (node
->thunk
.thunk_p
)
2741 struct inline_edge_summary
*es
= inline_edge_summary (node
->callees
);
2742 struct predicate t
= true_predicate ();
2744 info
->inlinable
= 0;
2745 node
->callees
->call_stmt_cannot_inline_p
= true;
2746 node
->local
.can_change_signature
= false;
2747 es
->call_stmt_time
= 1;
2748 es
->call_stmt_size
= 1;
2749 account_size_time (info
, 0, 0, &t
);
2753 /* Even is_gimple_min_invariant rely on current_function_decl. */
2754 push_cfun (DECL_STRUCT_FUNCTION (node
->decl
));
2756 /* Estimate the stack size for the function if we're optimizing. */
2757 self_stack_size
= optimize
? estimated_stack_frame_size (node
) : 0;
2758 info
->estimated_self_stack_size
= self_stack_size
;
2759 info
->estimated_stack_size
= self_stack_size
;
2760 info
->stack_frame_offset
= 0;
2762 /* Can this function be inlined at all? */
2763 if (!optimize
&& !lookup_attribute ("always_inline",
2764 DECL_ATTRIBUTES (node
->decl
)))
2765 info
->inlinable
= false;
2767 info
->inlinable
= tree_inlinable_function_p (node
->decl
);
2769 /* Type attributes can use parameter indices to describe them. */
2770 if (TYPE_ATTRIBUTES (TREE_TYPE (node
->decl
)))
2771 node
->local
.can_change_signature
= false;
2774 /* Otherwise, inlinable functions always can change signature. */
2775 if (info
->inlinable
)
2776 node
->local
.can_change_signature
= true;
2779 /* Functions calling builtin_apply can not change signature. */
2780 for (e
= node
->callees
; e
; e
= e
->next_callee
)
2782 tree
cdecl = e
->callee
->decl
;
2783 if (DECL_BUILT_IN (cdecl)
2784 && DECL_BUILT_IN_CLASS (cdecl) == BUILT_IN_NORMAL
2785 && (DECL_FUNCTION_CODE (cdecl) == BUILT_IN_APPLY_ARGS
2786 || DECL_FUNCTION_CODE (cdecl) == BUILT_IN_VA_START
))
2789 node
->local
.can_change_signature
= !e
;
2792 estimate_function_body_sizes (node
, early
);
2794 /* Inlining characteristics are maintained by the cgraph_mark_inline. */
2795 info
->time
= info
->self_time
;
2796 info
->size
= info
->self_size
;
2797 info
->stack_frame_offset
= 0;
2798 info
->estimated_stack_size
= info
->estimated_self_stack_size
;
2799 #ifdef ENABLE_CHECKING
2800 inline_update_overall_summary (node
);
2801 gcc_assert (info
->time
== info
->self_time
&& info
->size
== info
->self_size
);
2808 /* Compute parameters of functions used by inliner using
2809 current_function_decl. */
2812 compute_inline_parameters_for_current (void)
2814 compute_inline_parameters (cgraph_get_node (current_function_decl
), true);
2820 const pass_data pass_data_inline_parameters
=
2822 GIMPLE_PASS
, /* type */
2823 "inline_param", /* name */
2824 OPTGROUP_INLINE
, /* optinfo_flags */
2825 false, /* has_gate */
2826 true, /* has_execute */
2827 TV_INLINE_PARAMETERS
, /* tv_id */
2828 0, /* properties_required */
2829 0, /* properties_provided */
2830 0, /* properties_destroyed */
2831 0, /* todo_flags_start */
2832 0, /* todo_flags_finish */
2835 class pass_inline_parameters
: public gimple_opt_pass
2838 pass_inline_parameters (gcc::context
*ctxt
)
2839 : gimple_opt_pass (pass_data_inline_parameters
, ctxt
)
2842 /* opt_pass methods: */
2843 opt_pass
* clone () { return new pass_inline_parameters (m_ctxt
); }
2844 unsigned int execute () {
2845 return compute_inline_parameters_for_current ();
2848 }; // class pass_inline_parameters
2853 make_pass_inline_parameters (gcc::context
*ctxt
)
2855 return new pass_inline_parameters (ctxt
);
2859 /* Estimate benefit devirtualizing indirect edge IE, provided KNOWN_VALS and
2863 estimate_edge_devirt_benefit (struct cgraph_edge
*ie
,
2864 int *size
, int *time
,
2865 vec
<tree
> known_vals
,
2866 vec
<tree
> known_binfos
,
2867 vec
<ipa_agg_jump_function_p
> known_aggs
)
2870 struct cgraph_node
*callee
;
2871 struct inline_summary
*isummary
;
2873 if (!known_vals
.exists () && !known_binfos
.exists ())
2875 if (!flag_indirect_inlining
)
2878 target
= ipa_get_indirect_edge_target (ie
, known_vals
, known_binfos
,
2883 /* Account for difference in cost between indirect and direct calls. */
2884 *size
-= (eni_size_weights
.indirect_call_cost
- eni_size_weights
.call_cost
);
2885 *time
-= (eni_time_weights
.indirect_call_cost
- eni_time_weights
.call_cost
);
2886 gcc_checking_assert (*time
>= 0);
2887 gcc_checking_assert (*size
>= 0);
2889 callee
= cgraph_get_node (target
);
2890 if (!callee
|| !callee
->definition
)
2892 isummary
= inline_summary (callee
);
2893 return isummary
->inlinable
;
2896 /* Increase SIZE and TIME for size and time needed to handle edge E. */
2899 estimate_edge_size_and_time (struct cgraph_edge
*e
, int *size
, int *time
,
2901 vec
<tree
> known_vals
,
2902 vec
<tree
> known_binfos
,
2903 vec
<ipa_agg_jump_function_p
> known_aggs
,
2904 inline_hints
*hints
)
2906 struct inline_edge_summary
*es
= inline_edge_summary (e
);
2907 int call_size
= es
->call_stmt_size
;
2908 int call_time
= es
->call_stmt_time
;
2910 && estimate_edge_devirt_benefit (e
, &call_size
, &call_time
,
2911 known_vals
, known_binfos
, known_aggs
)
2912 && hints
&& cgraph_maybe_hot_edge_p (e
))
2913 *hints
|= INLINE_HINT_indirect_call
;
2914 *size
+= call_size
* INLINE_SIZE_SCALE
;
2915 *time
+= apply_probability ((gcov_type
) call_time
, prob
)
2916 * e
->frequency
* (INLINE_TIME_SCALE
/ CGRAPH_FREQ_BASE
);
2917 if (*time
> MAX_TIME
* INLINE_TIME_SCALE
)
2918 *time
= MAX_TIME
* INLINE_TIME_SCALE
;
2923 /* Increase SIZE and TIME for size and time needed to handle all calls in NODE.
2924 POSSIBLE_TRUTHS, KNOWN_VALS and KNOWN_BINFOS describe context of the call
2928 estimate_calls_size_and_time (struct cgraph_node
*node
, int *size
, int *time
,
2929 inline_hints
*hints
,
2930 clause_t possible_truths
,
2931 vec
<tree
> known_vals
,
2932 vec
<tree
> known_binfos
,
2933 vec
<ipa_agg_jump_function_p
> known_aggs
)
2935 struct cgraph_edge
*e
;
2936 for (e
= node
->callees
; e
; e
= e
->next_callee
)
2938 struct inline_edge_summary
*es
= inline_edge_summary (e
);
2940 || evaluate_predicate (es
->predicate
, possible_truths
))
2942 if (e
->inline_failed
)
2944 /* Predicates of calls shall not use NOT_CHANGED codes,
2945 sowe do not need to compute probabilities. */
2946 estimate_edge_size_and_time (e
, size
, time
, REG_BR_PROB_BASE
,
2947 known_vals
, known_binfos
,
2951 estimate_calls_size_and_time (e
->callee
, size
, time
, hints
,
2953 known_vals
, known_binfos
,
2957 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
2959 struct inline_edge_summary
*es
= inline_edge_summary (e
);
2961 || evaluate_predicate (es
->predicate
, possible_truths
))
2962 estimate_edge_size_and_time (e
, size
, time
, REG_BR_PROB_BASE
,
2963 known_vals
, known_binfos
, known_aggs
,
2969 /* Estimate size and time needed to execute NODE assuming
2970 POSSIBLE_TRUTHS clause, and KNOWN_VALS and KNOWN_BINFOS information
2971 about NODE's arguments. */
2974 estimate_node_size_and_time (struct cgraph_node
*node
,
2975 clause_t possible_truths
,
2976 vec
<tree
> known_vals
,
2977 vec
<tree
> known_binfos
,
2978 vec
<ipa_agg_jump_function_p
> known_aggs
,
2979 int *ret_size
, int *ret_time
,
2980 inline_hints
*ret_hints
,
2981 vec
<inline_param_summary_t
>
2982 inline_param_summary
)
2984 struct inline_summary
*info
= inline_summary (node
);
2988 inline_hints hints
= 0;
2991 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2994 fprintf (dump_file
, " Estimating body: %s/%i\n"
2995 " Known to be false: ", node
->name (),
2998 for (i
= predicate_not_inlined_condition
;
2999 i
< (predicate_first_dynamic_condition
3000 + (int) vec_safe_length (info
->conds
)); i
++)
3001 if (!(possible_truths
& (1 << i
)))
3004 fprintf (dump_file
, ", ");
3006 dump_condition (dump_file
, info
->conds
, i
);
3010 for (i
= 0; vec_safe_iterate (info
->entry
, i
, &e
); i
++)
3011 if (evaluate_predicate (&e
->predicate
, possible_truths
))
3014 gcc_checking_assert (e
->time
>= 0);
3015 gcc_checking_assert (time
>= 0);
3016 if (!inline_param_summary
.exists ())
3020 int prob
= predicate_probability (info
->conds
,
3023 inline_param_summary
);
3024 gcc_checking_assert (prob
>= 0);
3025 gcc_checking_assert (prob
<= REG_BR_PROB_BASE
);
3026 time
+= apply_probability ((gcov_type
) e
->time
, prob
);
3028 if (time
> MAX_TIME
* INLINE_TIME_SCALE
)
3029 time
= MAX_TIME
* INLINE_TIME_SCALE
;
3030 gcc_checking_assert (time
>= 0);
3033 gcc_checking_assert (size
>= 0);
3034 gcc_checking_assert (time
>= 0);
3036 if (info
->loop_iterations
3037 && !evaluate_predicate (info
->loop_iterations
, possible_truths
))
3038 hints
|= INLINE_HINT_loop_iterations
;
3039 if (info
->loop_stride
3040 && !evaluate_predicate (info
->loop_stride
, possible_truths
))
3041 hints
|= INLINE_HINT_loop_stride
;
3042 if (info
->array_index
3043 && !evaluate_predicate (info
->array_index
, possible_truths
))
3044 hints
|= INLINE_HINT_array_index
;
3046 hints
|= INLINE_HINT_in_scc
;
3047 if (DECL_DECLARED_INLINE_P (node
->decl
))
3048 hints
|= INLINE_HINT_declared_inline
;
3050 estimate_calls_size_and_time (node
, &size
, &time
, &hints
, possible_truths
,
3051 known_vals
, known_binfos
, known_aggs
);
3052 gcc_checking_assert (size
>= 0);
3053 gcc_checking_assert (time
>= 0);
3054 time
= RDIV (time
, INLINE_TIME_SCALE
);
3055 size
= RDIV (size
, INLINE_SIZE_SCALE
);
3057 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3058 fprintf (dump_file
, "\n size:%i time:%i\n", (int) size
, (int) time
);
3069 /* Estimate size and time needed to execute callee of EDGE assuming that
3070 parameters known to be constant at caller of EDGE are propagated.
3071 KNOWN_VALS and KNOWN_BINFOS are vectors of assumed known constant values
3072 and types for parameters. */
3075 estimate_ipcp_clone_size_and_time (struct cgraph_node
*node
,
3076 vec
<tree
> known_vals
,
3077 vec
<tree
> known_binfos
,
3078 vec
<ipa_agg_jump_function_p
> known_aggs
,
3079 int *ret_size
, int *ret_time
,
3080 inline_hints
*hints
)
3084 clause
= evaluate_conditions_for_known_args (node
, false, known_vals
,
3086 estimate_node_size_and_time (node
, clause
, known_vals
, known_binfos
,
3087 known_aggs
, ret_size
, ret_time
, hints
, vNULL
);
3090 /* Translate all conditions from callee representation into caller
3091 representation and symbolically evaluate predicate P into new predicate.
3093 INFO is inline_summary of function we are adding predicate into, CALLEE_INFO
3094 is summary of function predicate P is from. OPERAND_MAP is array giving
3095 callee formal IDs the caller formal IDs. POSSSIBLE_TRUTHS is clausule of all
3096 callee conditions that may be true in caller context. TOPLEV_PREDICATE is
3097 predicate under which callee is executed. OFFSET_MAP is an array of of
3098 offsets that need to be added to conditions, negative offset means that
3099 conditions relying on values passed by reference have to be discarded
3100 because they might not be preserved (and should be considered offset zero
3101 for other purposes). */
3103 static struct predicate
3104 remap_predicate (struct inline_summary
*info
,
3105 struct inline_summary
*callee_info
,
3106 struct predicate
*p
,
3107 vec
<int> operand_map
,
3108 vec
<int> offset_map
,
3109 clause_t possible_truths
, struct predicate
*toplev_predicate
)
3112 struct predicate out
= true_predicate ();
3114 /* True predicate is easy. */
3115 if (true_predicate_p (p
))
3116 return *toplev_predicate
;
3117 for (i
= 0; p
->clause
[i
]; i
++)
3119 clause_t clause
= p
->clause
[i
];
3121 struct predicate clause_predicate
= false_predicate ();
3123 gcc_assert (i
< MAX_CLAUSES
);
3125 for (cond
= 0; cond
< NUM_CONDITIONS
; cond
++)
3126 /* Do we have condition we can't disprove? */
3127 if (clause
& possible_truths
& (1 << cond
))
3129 struct predicate cond_predicate
;
3130 /* Work out if the condition can translate to predicate in the
3131 inlined function. */
3132 if (cond
>= predicate_first_dynamic_condition
)
3134 struct condition
*c
;
3136 c
= &(*callee_info
->conds
)[cond
3138 predicate_first_dynamic_condition
];
3139 /* See if we can remap condition operand to caller's operand.
3140 Otherwise give up. */
3141 if (!operand_map
.exists ()
3142 || (int) operand_map
.length () <= c
->operand_num
3143 || operand_map
[c
->operand_num
] == -1
3144 /* TODO: For non-aggregate conditions, adding an offset is
3145 basically an arithmetic jump function processing which
3146 we should support in future. */
3147 || ((!c
->agg_contents
|| !c
->by_ref
)
3148 && offset_map
[c
->operand_num
] > 0)
3149 || (c
->agg_contents
&& c
->by_ref
3150 && offset_map
[c
->operand_num
] < 0))
3151 cond_predicate
= true_predicate ();
3154 struct agg_position_info ap
;
3155 HOST_WIDE_INT offset_delta
= offset_map
[c
->operand_num
];
3156 if (offset_delta
< 0)
3158 gcc_checking_assert (!c
->agg_contents
|| !c
->by_ref
);
3161 gcc_assert (!c
->agg_contents
3162 || c
->by_ref
|| offset_delta
== 0);
3163 ap
.offset
= c
->offset
+ offset_delta
;
3164 ap
.agg_contents
= c
->agg_contents
;
3165 ap
.by_ref
= c
->by_ref
;
3166 cond_predicate
= add_condition (info
,
3167 operand_map
[c
->operand_num
],
3168 &ap
, c
->code
, c
->val
);
3171 /* Fixed conditions remains same, construct single
3172 condition predicate. */
3175 cond_predicate
.clause
[0] = 1 << cond
;
3176 cond_predicate
.clause
[1] = 0;
3178 clause_predicate
= or_predicates (info
->conds
, &clause_predicate
,
3181 out
= and_predicates (info
->conds
, &out
, &clause_predicate
);
3183 return and_predicates (info
->conds
, &out
, toplev_predicate
);
3187 /* Update summary information of inline clones after inlining.
3188 Compute peak stack usage. */
3191 inline_update_callee_summaries (struct cgraph_node
*node
, int depth
)
3193 struct cgraph_edge
*e
;
3194 struct inline_summary
*callee_info
= inline_summary (node
);
3195 struct inline_summary
*caller_info
= inline_summary (node
->callers
->caller
);
3198 callee_info
->stack_frame_offset
3199 = caller_info
->stack_frame_offset
3200 + caller_info
->estimated_self_stack_size
;
3201 peak
= callee_info
->stack_frame_offset
3202 + callee_info
->estimated_self_stack_size
;
3203 if (inline_summary (node
->global
.inlined_to
)->estimated_stack_size
< peak
)
3204 inline_summary (node
->global
.inlined_to
)->estimated_stack_size
= peak
;
3205 ipa_propagate_frequency (node
);
3206 for (e
= node
->callees
; e
; e
= e
->next_callee
)
3208 if (!e
->inline_failed
)
3209 inline_update_callee_summaries (e
->callee
, depth
);
3210 inline_edge_summary (e
)->loop_depth
+= depth
;
3212 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
3213 inline_edge_summary (e
)->loop_depth
+= depth
;
3216 /* Update change_prob of EDGE after INLINED_EDGE has been inlined.
3217 When functoin A is inlined in B and A calls C with parameter that
3218 changes with probability PROB1 and C is known to be passthroug
3219 of argument if B that change with probability PROB2, the probability
3220 of change is now PROB1*PROB2. */
3223 remap_edge_change_prob (struct cgraph_edge
*inlined_edge
,
3224 struct cgraph_edge
*edge
)
3226 if (ipa_node_params_vector
.exists ())
3229 struct ipa_edge_args
*args
= IPA_EDGE_REF (edge
);
3230 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3231 struct inline_edge_summary
*inlined_es
3232 = inline_edge_summary (inlined_edge
);
3234 for (i
= 0; i
< ipa_get_cs_argument_count (args
); i
++)
3236 struct ipa_jump_func
*jfunc
= ipa_get_ith_jump_func (args
, i
);
3237 if (jfunc
->type
== IPA_JF_PASS_THROUGH
3238 && (ipa_get_jf_pass_through_formal_id (jfunc
)
3239 < (int) inlined_es
->param
.length ()))
3241 int jf_formal_id
= ipa_get_jf_pass_through_formal_id (jfunc
);
3242 int prob1
= es
->param
[i
].change_prob
;
3243 int prob2
= inlined_es
->param
[jf_formal_id
].change_prob
;
3244 int prob
= combine_probabilities (prob1
, prob2
);
3246 if (prob1
&& prob2
&& !prob
)
3249 es
->param
[i
].change_prob
= prob
;
3255 /* Update edge summaries of NODE after INLINED_EDGE has been inlined.
3257 Remap predicates of callees of NODE. Rest of arguments match
3260 Also update change probabilities. */
3263 remap_edge_summaries (struct cgraph_edge
*inlined_edge
,
3264 struct cgraph_node
*node
,
3265 struct inline_summary
*info
,
3266 struct inline_summary
*callee_info
,
3267 vec
<int> operand_map
,
3268 vec
<int> offset_map
,
3269 clause_t possible_truths
,
3270 struct predicate
*toplev_predicate
)
3272 struct cgraph_edge
*e
;
3273 for (e
= node
->callees
; e
; e
= e
->next_callee
)
3275 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3278 if (e
->inline_failed
)
3280 remap_edge_change_prob (inlined_edge
, e
);
3284 p
= remap_predicate (info
, callee_info
,
3285 es
->predicate
, operand_map
, offset_map
,
3286 possible_truths
, toplev_predicate
);
3287 edge_set_predicate (e
, &p
);
3288 /* TODO: We should remove the edge for code that will be
3289 optimized out, but we need to keep verifiers and tree-inline
3290 happy. Make it cold for now. */
3291 if (false_predicate_p (&p
))
3298 edge_set_predicate (e
, toplev_predicate
);
3301 remap_edge_summaries (inlined_edge
, e
->callee
, info
, callee_info
,
3302 operand_map
, offset_map
, possible_truths
,
3305 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
3307 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3310 remap_edge_change_prob (inlined_edge
, e
);
3313 p
= remap_predicate (info
, callee_info
,
3314 es
->predicate
, operand_map
, offset_map
,
3315 possible_truths
, toplev_predicate
);
3316 edge_set_predicate (e
, &p
);
3317 /* TODO: We should remove the edge for code that will be optimized
3318 out, but we need to keep verifiers and tree-inline happy.
3319 Make it cold for now. */
3320 if (false_predicate_p (&p
))
3327 edge_set_predicate (e
, toplev_predicate
);
3331 /* Same as remap_predicate, but set result into hint *HINT. */
3334 remap_hint_predicate (struct inline_summary
*info
,
3335 struct inline_summary
*callee_info
,
3336 struct predicate
**hint
,
3337 vec
<int> operand_map
,
3338 vec
<int> offset_map
,
3339 clause_t possible_truths
,
3340 struct predicate
*toplev_predicate
)
3346 p
= remap_predicate (info
, callee_info
,
3348 operand_map
, offset_map
,
3349 possible_truths
, toplev_predicate
);
3350 if (!false_predicate_p (&p
) && !true_predicate_p (&p
))
3353 set_hint_predicate (hint
, p
);
3355 **hint
= and_predicates (info
->conds
, *hint
, &p
);
3359 /* We inlined EDGE. Update summary of the function we inlined into. */
3362 inline_merge_summary (struct cgraph_edge
*edge
)
3364 struct inline_summary
*callee_info
= inline_summary (edge
->callee
);
3365 struct cgraph_node
*to
= (edge
->caller
->global
.inlined_to
3366 ? edge
->caller
->global
.inlined_to
: edge
->caller
);
3367 struct inline_summary
*info
= inline_summary (to
);
3368 clause_t clause
= 0; /* not_inline is known to be false. */
3370 vec
<int> operand_map
= vNULL
;
3371 vec
<int> offset_map
= vNULL
;
3373 struct predicate toplev_predicate
;
3374 struct predicate true_p
= true_predicate ();
3375 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3378 toplev_predicate
= *es
->predicate
;
3380 toplev_predicate
= true_predicate ();
3382 if (ipa_node_params_vector
.exists () && callee_info
->conds
)
3384 struct ipa_edge_args
*args
= IPA_EDGE_REF (edge
);
3385 int count
= ipa_get_cs_argument_count (args
);
3388 evaluate_properties_for_edge (edge
, true, &clause
, NULL
, NULL
, NULL
);
3391 operand_map
.safe_grow_cleared (count
);
3392 offset_map
.safe_grow_cleared (count
);
3394 for (i
= 0; i
< count
; i
++)
3396 struct ipa_jump_func
*jfunc
= ipa_get_ith_jump_func (args
, i
);
3399 /* TODO: handle non-NOPs when merging. */
3400 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
3402 if (ipa_get_jf_pass_through_operation (jfunc
) == NOP_EXPR
)
3403 map
= ipa_get_jf_pass_through_formal_id (jfunc
);
3404 if (!ipa_get_jf_pass_through_agg_preserved (jfunc
))
3407 else if (jfunc
->type
== IPA_JF_ANCESTOR
)
3409 HOST_WIDE_INT offset
= ipa_get_jf_ancestor_offset (jfunc
);
3410 if (offset
>= 0 && offset
< INT_MAX
)
3412 map
= ipa_get_jf_ancestor_formal_id (jfunc
);
3413 if (!ipa_get_jf_ancestor_agg_preserved (jfunc
))
3415 offset_map
[i
] = offset
;
3418 operand_map
[i
] = map
;
3419 gcc_assert (map
< ipa_get_param_count (IPA_NODE_REF (to
)));
3422 for (i
= 0; vec_safe_iterate (callee_info
->entry
, i
, &e
); i
++)
3424 struct predicate p
= remap_predicate (info
, callee_info
,
3425 &e
->predicate
, operand_map
,
3428 if (!false_predicate_p (&p
))
3430 gcov_type add_time
= ((gcov_type
) e
->time
* edge
->frequency
3431 + CGRAPH_FREQ_BASE
/ 2) / CGRAPH_FREQ_BASE
;
3432 int prob
= predicate_probability (callee_info
->conds
,
3435 add_time
= apply_probability ((gcov_type
) add_time
, prob
);
3436 if (add_time
> MAX_TIME
* INLINE_TIME_SCALE
)
3437 add_time
= MAX_TIME
* INLINE_TIME_SCALE
;
3438 if (prob
!= REG_BR_PROB_BASE
3439 && dump_file
&& (dump_flags
& TDF_DETAILS
))
3441 fprintf (dump_file
, "\t\tScaling time by probability:%f\n",
3442 (double) prob
/ REG_BR_PROB_BASE
);
3444 account_size_time (info
, e
->size
, add_time
, &p
);
3447 remap_edge_summaries (edge
, edge
->callee
, info
, callee_info
, operand_map
,
3448 offset_map
, clause
, &toplev_predicate
);
3449 remap_hint_predicate (info
, callee_info
,
3450 &callee_info
->loop_iterations
,
3451 operand_map
, offset_map
, clause
, &toplev_predicate
);
3452 remap_hint_predicate (info
, callee_info
,
3453 &callee_info
->loop_stride
,
3454 operand_map
, offset_map
, clause
, &toplev_predicate
);
3455 remap_hint_predicate (info
, callee_info
,
3456 &callee_info
->array_index
,
3457 operand_map
, offset_map
, clause
, &toplev_predicate
);
3459 inline_update_callee_summaries (edge
->callee
,
3460 inline_edge_summary (edge
)->loop_depth
);
3462 /* We do not maintain predicates of inlined edges, free it. */
3463 edge_set_predicate (edge
, &true_p
);
3464 /* Similarly remove param summaries. */
3465 es
->param
.release ();
3466 operand_map
.release ();
3467 offset_map
.release ();
3470 /* For performance reasons inline_merge_summary is not updating overall size
3471 and time. Recompute it. */
3474 inline_update_overall_summary (struct cgraph_node
*node
)
3476 struct inline_summary
*info
= inline_summary (node
);
3482 for (i
= 0; vec_safe_iterate (info
->entry
, i
, &e
); i
++)
3484 info
->size
+= e
->size
, info
->time
+= e
->time
;
3485 if (info
->time
> MAX_TIME
* INLINE_TIME_SCALE
)
3486 info
->time
= MAX_TIME
* INLINE_TIME_SCALE
;
3488 estimate_calls_size_and_time (node
, &info
->size
, &info
->time
, NULL
,
3489 ~(clause_t
) (1 << predicate_false_condition
),
3490 vNULL
, vNULL
, vNULL
);
3491 info
->time
= (info
->time
+ INLINE_TIME_SCALE
/ 2) / INLINE_TIME_SCALE
;
3492 info
->size
= (info
->size
+ INLINE_SIZE_SCALE
/ 2) / INLINE_SIZE_SCALE
;
3495 /* Return hints derrived from EDGE. */
3497 simple_edge_hints (struct cgraph_edge
*edge
)
3500 struct cgraph_node
*to
= (edge
->caller
->global
.inlined_to
3501 ? edge
->caller
->global
.inlined_to
: edge
->caller
);
3502 if (inline_summary (to
)->scc_no
3503 && inline_summary (to
)->scc_no
== inline_summary (edge
->callee
)->scc_no
3504 && !cgraph_edge_recursive_p (edge
))
3505 hints
|= INLINE_HINT_same_scc
;
3507 if (to
->lto_file_data
&& edge
->callee
->lto_file_data
3508 && to
->lto_file_data
!= edge
->callee
->lto_file_data
)
3509 hints
|= INLINE_HINT_cross_module
;
3514 /* Estimate the time cost for the caller when inlining EDGE.
3515 Only to be called via estimate_edge_time, that handles the
3518 When caching, also update the cache entry. Compute both time and
3519 size, since we always need both metrics eventually. */
3522 do_estimate_edge_time (struct cgraph_edge
*edge
)
3527 struct cgraph_node
*callee
;
3529 vec
<tree
> known_vals
;
3530 vec
<tree
> known_binfos
;
3531 vec
<ipa_agg_jump_function_p
> known_aggs
;
3532 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3534 callee
= cgraph_function_or_thunk_node (edge
->callee
, NULL
);
3536 gcc_checking_assert (edge
->inline_failed
);
3537 evaluate_properties_for_edge (edge
, true,
3538 &clause
, &known_vals
, &known_binfos
,
3540 estimate_node_size_and_time (callee
, clause
, known_vals
, known_binfos
,
3541 known_aggs
, &size
, &time
, &hints
, es
->param
);
3542 known_vals
.release ();
3543 known_binfos
.release ();
3544 known_aggs
.release ();
3545 gcc_checking_assert (size
>= 0);
3546 gcc_checking_assert (time
>= 0);
3548 /* When caching, update the cache entry. */
3549 if (edge_growth_cache
.exists ())
3551 if ((int) edge_growth_cache
.length () <= edge
->uid
)
3552 edge_growth_cache
.safe_grow_cleared (cgraph_edge_max_uid
);
3553 edge_growth_cache
[edge
->uid
].time
= time
+ (time
>= 0);
3555 edge_growth_cache
[edge
->uid
].size
= size
+ (size
>= 0);
3556 hints
|= simple_edge_hints (edge
);
3557 edge_growth_cache
[edge
->uid
].hints
= hints
+ 1;
3563 /* Return estimated callee growth after inlining EDGE.
3564 Only to be called via estimate_edge_size. */
3567 do_estimate_edge_size (struct cgraph_edge
*edge
)
3570 struct cgraph_node
*callee
;
3572 vec
<tree
> known_vals
;
3573 vec
<tree
> known_binfos
;
3574 vec
<ipa_agg_jump_function_p
> known_aggs
;
3576 /* When we do caching, use do_estimate_edge_time to populate the entry. */
3578 if (edge_growth_cache
.exists ())
3580 do_estimate_edge_time (edge
);
3581 size
= edge_growth_cache
[edge
->uid
].size
;
3582 gcc_checking_assert (size
);
3583 return size
- (size
> 0);
3586 callee
= cgraph_function_or_thunk_node (edge
->callee
, NULL
);
3588 /* Early inliner runs without caching, go ahead and do the dirty work. */
3589 gcc_checking_assert (edge
->inline_failed
);
3590 evaluate_properties_for_edge (edge
, true,
3591 &clause
, &known_vals
, &known_binfos
,
3593 estimate_node_size_and_time (callee
, clause
, known_vals
, known_binfos
,
3594 known_aggs
, &size
, NULL
, NULL
, vNULL
);
3595 known_vals
.release ();
3596 known_binfos
.release ();
3597 known_aggs
.release ();
3602 /* Estimate the growth of the caller when inlining EDGE.
3603 Only to be called via estimate_edge_size. */
3606 do_estimate_edge_hints (struct cgraph_edge
*edge
)
3609 struct cgraph_node
*callee
;
3611 vec
<tree
> known_vals
;
3612 vec
<tree
> known_binfos
;
3613 vec
<ipa_agg_jump_function_p
> known_aggs
;
3615 /* When we do caching, use do_estimate_edge_time to populate the entry. */
3617 if (edge_growth_cache
.exists ())
3619 do_estimate_edge_time (edge
);
3620 hints
= edge_growth_cache
[edge
->uid
].hints
;
3621 gcc_checking_assert (hints
);
3625 callee
= cgraph_function_or_thunk_node (edge
->callee
, NULL
);
3627 /* Early inliner runs without caching, go ahead and do the dirty work. */
3628 gcc_checking_assert (edge
->inline_failed
);
3629 evaluate_properties_for_edge (edge
, true,
3630 &clause
, &known_vals
, &known_binfos
,
3632 estimate_node_size_and_time (callee
, clause
, known_vals
, known_binfos
,
3633 known_aggs
, NULL
, NULL
, &hints
, vNULL
);
3634 known_vals
.release ();
3635 known_binfos
.release ();
3636 known_aggs
.release ();
3637 hints
|= simple_edge_hints (edge
);
3642 /* Estimate self time of the function NODE after inlining EDGE. */
3645 estimate_time_after_inlining (struct cgraph_node
*node
,
3646 struct cgraph_edge
*edge
)
3648 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3649 if (!es
->predicate
|| !false_predicate_p (es
->predicate
))
3652 inline_summary (node
)->time
+ estimate_edge_time (edge
);
3655 if (time
> MAX_TIME
)
3659 return inline_summary (node
)->time
;
3663 /* Estimate the size of NODE after inlining EDGE which should be an
3664 edge to either NODE or a call inlined into NODE. */
3667 estimate_size_after_inlining (struct cgraph_node
*node
,
3668 struct cgraph_edge
*edge
)
3670 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3671 if (!es
->predicate
|| !false_predicate_p (es
->predicate
))
3673 int size
= inline_summary (node
)->size
+ estimate_edge_growth (edge
);
3674 gcc_assert (size
>= 0);
3677 return inline_summary (node
)->size
;
3683 struct cgraph_node
*node
;
3684 bool self_recursive
;
3689 /* Worker for do_estimate_growth. Collect growth for all callers. */
3692 do_estimate_growth_1 (struct cgraph_node
*node
, void *data
)
3694 struct cgraph_edge
*e
;
3695 struct growth_data
*d
= (struct growth_data
*) data
;
3697 for (e
= node
->callers
; e
; e
= e
->next_caller
)
3699 gcc_checking_assert (e
->inline_failed
);
3701 if (e
->caller
== d
->node
3702 || (e
->caller
->global
.inlined_to
3703 && e
->caller
->global
.inlined_to
== d
->node
))
3704 d
->self_recursive
= true;
3705 d
->growth
+= estimate_edge_growth (e
);
3711 /* Estimate the growth caused by inlining NODE into all callees. */
3714 do_estimate_growth (struct cgraph_node
*node
)
3716 struct growth_data d
= { node
, 0, false };
3717 struct inline_summary
*info
= inline_summary (node
);
3719 cgraph_for_node_and_aliases (node
, do_estimate_growth_1
, &d
, true);
3721 /* For self recursive functions the growth estimation really should be
3722 infinity. We don't want to return very large values because the growth
3723 plays various roles in badness computation fractions. Be sure to not
3724 return zero or negative growths. */
3725 if (d
.self_recursive
)
3726 d
.growth
= d
.growth
< info
->size
? info
->size
: d
.growth
;
3727 else if (DECL_EXTERNAL (node
->decl
))
3731 if (cgraph_will_be_removed_from_program_if_no_direct_calls (node
))
3732 d
.growth
-= info
->size
;
3733 /* COMDAT functions are very often not shared across multiple units
3734 since they come from various template instantiations.
3735 Take this into account. */
3736 else if (DECL_COMDAT (node
->decl
)
3737 && cgraph_can_remove_if_no_direct_calls_p (node
))
3738 d
.growth
-= (info
->size
3739 * (100 - PARAM_VALUE (PARAM_COMDAT_SHARING_PROBABILITY
))
3743 if (node_growth_cache
.exists ())
3745 if ((int) node_growth_cache
.length () <= node
->uid
)
3746 node_growth_cache
.safe_grow_cleared (cgraph_max_uid
);
3747 node_growth_cache
[node
->uid
] = d
.growth
+ (d
.growth
>= 0);
3753 /* This function performs intraprocedural analysis in NODE that is required to
3754 inline indirect calls. */
3757 inline_indirect_intraprocedural_analysis (struct cgraph_node
*node
)
3759 ipa_analyze_node (node
);
3760 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3762 ipa_print_node_params (dump_file
, node
);
3763 ipa_print_node_jump_functions (dump_file
, node
);
3768 /* Note function body size. */
3771 inline_analyze_function (struct cgraph_node
*node
)
3773 push_cfun (DECL_STRUCT_FUNCTION (node
->decl
));
3776 fprintf (dump_file
, "\nAnalyzing function: %s/%u\n",
3777 node
->name (), node
->order
);
3778 if (optimize
&& !node
->thunk
.thunk_p
)
3779 inline_indirect_intraprocedural_analysis (node
);
3780 compute_inline_parameters (node
, false);
3783 struct cgraph_edge
*e
;
3784 for (e
= node
->callees
; e
; e
= e
->next_callee
)
3786 if (e
->inline_failed
== CIF_FUNCTION_NOT_CONSIDERED
)
3787 e
->inline_failed
= CIF_FUNCTION_NOT_OPTIMIZED
;
3788 e
->call_stmt_cannot_inline_p
= true;
3790 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
3792 if (e
->inline_failed
== CIF_FUNCTION_NOT_CONSIDERED
)
3793 e
->inline_failed
= CIF_FUNCTION_NOT_OPTIMIZED
;
3794 e
->call_stmt_cannot_inline_p
= true;
3802 /* Called when new function is inserted to callgraph late. */
3805 add_new_function (struct cgraph_node
*node
, void *data ATTRIBUTE_UNUSED
)
3807 inline_analyze_function (node
);
3811 /* Note function body size. */
3814 inline_generate_summary (void)
3816 struct cgraph_node
*node
;
3818 /* When not optimizing, do not bother to analyze. Inlining is still done
3819 because edge redirection needs to happen there. */
3820 if (!optimize
&& !flag_lto
&& !flag_wpa
)
3823 function_insertion_hook_holder
=
3824 cgraph_add_function_insertion_hook (&add_new_function
, NULL
);
3826 ipa_register_cgraph_hooks ();
3827 inline_free_summary ();
3829 FOR_EACH_DEFINED_FUNCTION (node
)
3831 inline_analyze_function (node
);
3835 /* Read predicate from IB. */
3837 static struct predicate
3838 read_predicate (struct lto_input_block
*ib
)
3840 struct predicate out
;
3846 gcc_assert (k
<= MAX_CLAUSES
);
3847 clause
= out
.clause
[k
++] = streamer_read_uhwi (ib
);
3851 /* Zero-initialize the remaining clauses in OUT. */
3852 while (k
<= MAX_CLAUSES
)
3853 out
.clause
[k
++] = 0;
3859 /* Write inline summary for edge E to OB. */
3862 read_inline_edge_summary (struct lto_input_block
*ib
, struct cgraph_edge
*e
)
3864 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3868 es
->call_stmt_size
= streamer_read_uhwi (ib
);
3869 es
->call_stmt_time
= streamer_read_uhwi (ib
);
3870 es
->loop_depth
= streamer_read_uhwi (ib
);
3871 p
= read_predicate (ib
);
3872 edge_set_predicate (e
, &p
);
3873 length
= streamer_read_uhwi (ib
);
3876 es
->param
.safe_grow_cleared (length
);
3877 for (i
= 0; i
< length
; i
++)
3878 es
->param
[i
].change_prob
= streamer_read_uhwi (ib
);
3883 /* Stream in inline summaries from the section. */
3886 inline_read_section (struct lto_file_decl_data
*file_data
, const char *data
,
3889 const struct lto_function_header
*header
=
3890 (const struct lto_function_header
*) data
;
3891 const int cfg_offset
= sizeof (struct lto_function_header
);
3892 const int main_offset
= cfg_offset
+ header
->cfg_size
;
3893 const int string_offset
= main_offset
+ header
->main_size
;
3894 struct data_in
*data_in
;
3895 struct lto_input_block ib
;
3896 unsigned int i
, count2
, j
;
3897 unsigned int f_count
;
3899 LTO_INIT_INPUT_BLOCK (ib
, (const char *) data
+ main_offset
, 0,
3903 lto_data_in_create (file_data
, (const char *) data
+ string_offset
,
3904 header
->string_size
, vNULL
);
3905 f_count
= streamer_read_uhwi (&ib
);
3906 for (i
= 0; i
< f_count
; i
++)
3909 struct cgraph_node
*node
;
3910 struct inline_summary
*info
;
3911 lto_symtab_encoder_t encoder
;
3912 struct bitpack_d bp
;
3913 struct cgraph_edge
*e
;
3916 index
= streamer_read_uhwi (&ib
);
3917 encoder
= file_data
->symtab_node_encoder
;
3918 node
= cgraph (lto_symtab_encoder_deref (encoder
, index
));
3919 info
= inline_summary (node
);
3921 info
->estimated_stack_size
3922 = info
->estimated_self_stack_size
= streamer_read_uhwi (&ib
);
3923 info
->size
= info
->self_size
= streamer_read_uhwi (&ib
);
3924 info
->time
= info
->self_time
= streamer_read_uhwi (&ib
);
3926 bp
= streamer_read_bitpack (&ib
);
3927 info
->inlinable
= bp_unpack_value (&bp
, 1);
3929 count2
= streamer_read_uhwi (&ib
);
3930 gcc_assert (!info
->conds
);
3931 for (j
= 0; j
< count2
; j
++)
3934 c
.operand_num
= streamer_read_uhwi (&ib
);
3935 c
.code
= (enum tree_code
) streamer_read_uhwi (&ib
);
3936 c
.val
= stream_read_tree (&ib
, data_in
);
3937 bp
= streamer_read_bitpack (&ib
);
3938 c
.agg_contents
= bp_unpack_value (&bp
, 1);
3939 c
.by_ref
= bp_unpack_value (&bp
, 1);
3941 c
.offset
= streamer_read_uhwi (&ib
);
3942 vec_safe_push (info
->conds
, c
);
3944 count2
= streamer_read_uhwi (&ib
);
3945 gcc_assert (!info
->entry
);
3946 for (j
= 0; j
< count2
; j
++)
3948 struct size_time_entry e
;
3950 e
.size
= streamer_read_uhwi (&ib
);
3951 e
.time
= streamer_read_uhwi (&ib
);
3952 e
.predicate
= read_predicate (&ib
);
3954 vec_safe_push (info
->entry
, e
);
3957 p
= read_predicate (&ib
);
3958 set_hint_predicate (&info
->loop_iterations
, p
);
3959 p
= read_predicate (&ib
);
3960 set_hint_predicate (&info
->loop_stride
, p
);
3961 p
= read_predicate (&ib
);
3962 set_hint_predicate (&info
->array_index
, p
);
3963 for (e
= node
->callees
; e
; e
= e
->next_callee
)
3964 read_inline_edge_summary (&ib
, e
);
3965 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
3966 read_inline_edge_summary (&ib
, e
);
3969 lto_free_section_data (file_data
, LTO_section_inline_summary
, NULL
, data
,
3971 lto_data_in_delete (data_in
);
3975 /* Read inline summary. Jump functions are shared among ipa-cp
3976 and inliner, so when ipa-cp is active, we don't need to write them
3980 inline_read_summary (void)
3982 struct lto_file_decl_data
**file_data_vec
= lto_get_file_decl_data ();
3983 struct lto_file_decl_data
*file_data
;
3986 inline_summary_alloc ();
3988 while ((file_data
= file_data_vec
[j
++]))
3991 const char *data
= lto_get_section_data (file_data
,
3992 LTO_section_inline_summary
,
3995 inline_read_section (file_data
, data
, len
);
3997 /* Fatal error here. We do not want to support compiling ltrans units
3998 with different version of compiler or different flags than the WPA
3999 unit, so this should never happen. */
4000 fatal_error ("ipa inline summary is missing in input file");
4004 ipa_register_cgraph_hooks ();
4006 ipa_prop_read_jump_functions ();
4008 function_insertion_hook_holder
=
4009 cgraph_add_function_insertion_hook (&add_new_function
, NULL
);
4013 /* Write predicate P to OB. */
4016 write_predicate (struct output_block
*ob
, struct predicate
*p
)
4020 for (j
= 0; p
->clause
[j
]; j
++)
4022 gcc_assert (j
< MAX_CLAUSES
);
4023 streamer_write_uhwi (ob
, p
->clause
[j
]);
4025 streamer_write_uhwi (ob
, 0);
4029 /* Write inline summary for edge E to OB. */
4032 write_inline_edge_summary (struct output_block
*ob
, struct cgraph_edge
*e
)
4034 struct inline_edge_summary
*es
= inline_edge_summary (e
);
4037 streamer_write_uhwi (ob
, es
->call_stmt_size
);
4038 streamer_write_uhwi (ob
, es
->call_stmt_time
);
4039 streamer_write_uhwi (ob
, es
->loop_depth
);
4040 write_predicate (ob
, es
->predicate
);
4041 streamer_write_uhwi (ob
, es
->param
.length ());
4042 for (i
= 0; i
< (int) es
->param
.length (); i
++)
4043 streamer_write_uhwi (ob
, es
->param
[i
].change_prob
);
4047 /* Write inline summary for node in SET.
4048 Jump functions are shared among ipa-cp and inliner, so when ipa-cp is
4049 active, we don't need to write them twice. */
4052 inline_write_summary (void)
4054 struct cgraph_node
*node
;
4055 struct output_block
*ob
= create_output_block (LTO_section_inline_summary
);
4056 lto_symtab_encoder_t encoder
= ob
->decl_state
->symtab_node_encoder
;
4057 unsigned int count
= 0;
4060 for (i
= 0; i
< lto_symtab_encoder_size (encoder
); i
++)
4062 symtab_node
*snode
= lto_symtab_encoder_deref (encoder
, i
);
4063 cgraph_node
*cnode
= dyn_cast
<cgraph_node
> (snode
);
4064 if (cnode
&& cnode
->definition
&& !cnode
->alias
)
4067 streamer_write_uhwi (ob
, count
);
4069 for (i
= 0; i
< lto_symtab_encoder_size (encoder
); i
++)
4071 symtab_node
*snode
= lto_symtab_encoder_deref (encoder
, i
);
4072 cgraph_node
*cnode
= dyn_cast
<cgraph_node
> (snode
);
4073 if (cnode
&& (node
= cnode
)->definition
&& !node
->alias
)
4075 struct inline_summary
*info
= inline_summary (node
);
4076 struct bitpack_d bp
;
4077 struct cgraph_edge
*edge
;
4080 struct condition
*c
;
4082 streamer_write_uhwi (ob
,
4083 lto_symtab_encoder_encode (encoder
,
4086 streamer_write_hwi (ob
, info
->estimated_self_stack_size
);
4087 streamer_write_hwi (ob
, info
->self_size
);
4088 streamer_write_hwi (ob
, info
->self_time
);
4089 bp
= bitpack_create (ob
->main_stream
);
4090 bp_pack_value (&bp
, info
->inlinable
, 1);
4091 streamer_write_bitpack (&bp
);
4092 streamer_write_uhwi (ob
, vec_safe_length (info
->conds
));
4093 for (i
= 0; vec_safe_iterate (info
->conds
, i
, &c
); i
++)
4095 streamer_write_uhwi (ob
, c
->operand_num
);
4096 streamer_write_uhwi (ob
, c
->code
);
4097 stream_write_tree (ob
, c
->val
, true);
4098 bp
= bitpack_create (ob
->main_stream
);
4099 bp_pack_value (&bp
, c
->agg_contents
, 1);
4100 bp_pack_value (&bp
, c
->by_ref
, 1);
4101 streamer_write_bitpack (&bp
);
4102 if (c
->agg_contents
)
4103 streamer_write_uhwi (ob
, c
->offset
);
4105 streamer_write_uhwi (ob
, vec_safe_length (info
->entry
));
4106 for (i
= 0; vec_safe_iterate (info
->entry
, i
, &e
); i
++)
4108 streamer_write_uhwi (ob
, e
->size
);
4109 streamer_write_uhwi (ob
, e
->time
);
4110 write_predicate (ob
, &e
->predicate
);
4112 write_predicate (ob
, info
->loop_iterations
);
4113 write_predicate (ob
, info
->loop_stride
);
4114 write_predicate (ob
, info
->array_index
);
4115 for (edge
= node
->callees
; edge
; edge
= edge
->next_callee
)
4116 write_inline_edge_summary (ob
, edge
);
4117 for (edge
= node
->indirect_calls
; edge
; edge
= edge
->next_callee
)
4118 write_inline_edge_summary (ob
, edge
);
4121 streamer_write_char_stream (ob
->main_stream
, 0);
4122 produce_asm (ob
, NULL
);
4123 destroy_output_block (ob
);
4125 if (optimize
&& !flag_ipa_cp
)
4126 ipa_prop_write_jump_functions ();
4130 /* Release inline summary. */
4133 inline_free_summary (void)
4135 struct cgraph_node
*node
;
4136 if (!inline_edge_summary_vec
.exists ())
4138 FOR_EACH_DEFINED_FUNCTION (node
)
4139 reset_inline_summary (node
);
4140 if (function_insertion_hook_holder
)
4141 cgraph_remove_function_insertion_hook (function_insertion_hook_holder
);
4142 function_insertion_hook_holder
= NULL
;
4143 if (node_removal_hook_holder
)
4144 cgraph_remove_node_removal_hook (node_removal_hook_holder
);
4145 node_removal_hook_holder
= NULL
;
4146 if (edge_removal_hook_holder
)
4147 cgraph_remove_edge_removal_hook (edge_removal_hook_holder
);
4148 edge_removal_hook_holder
= NULL
;
4149 if (node_duplication_hook_holder
)
4150 cgraph_remove_node_duplication_hook (node_duplication_hook_holder
);
4151 node_duplication_hook_holder
= NULL
;
4152 if (edge_duplication_hook_holder
)
4153 cgraph_remove_edge_duplication_hook (edge_duplication_hook_holder
);
4154 edge_duplication_hook_holder
= NULL
;
4155 vec_free (inline_summary_vec
);
4156 inline_edge_summary_vec
.release ();
4157 if (edge_predicate_pool
)
4158 free_alloc_pool (edge_predicate_pool
);
4159 edge_predicate_pool
= 0;