1 /* Inlining decision heuristics.
2 Copyright (C) 2003-2014 Free Software Foundation, Inc.
3 Contributed by Jan Hubicka
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
21 /* Analysis used by the inliner and other passes limiting code size growth.
23 We estimate for each function
25 - average function execution time
26 - inlining size benefit (that is how much of function body size
27 and its call sequence is expected to disappear by inlining)
28 - inlining time benefit
31 - call statement size and time
33 inlinie_summary datastructures store above information locally (i.e.
34 parameters of the function itself) and globally (i.e. parameters of
35 the function created by applying all the inline decisions already
36 present in the callgraph).
38 We provide accestor to the inline_summary datastructure and
39 basic logic updating the parameters when inlining is performed.
41 The summaries are context sensitive. Context means
42 1) partial assignment of known constant values of operands
43 2) whether function is inlined into the call or not.
44 It is easy to add more variants. To represent function size and time
45 that depends on context (i.e. it is known to be optimized away when
46 context is known either by inlining or from IP-CP and clonning),
47 we use predicates. Predicates are logical formulas in
48 conjunctive-disjunctive form consisting of clauses. Clauses are bitmaps
49 specifying what conditions must be true. Conditions are simple test
50 of the form described above.
52 In order to make predicate (possibly) true, all of its clauses must
53 be (possibly) true. To make clause (possibly) true, one of conditions
54 it mentions must be (possibly) true. There are fixed bounds on
55 number of clauses and conditions and all the manipulation functions
56 are conservative in positive direction. I.e. we may lose precision
57 by thinking that predicate may be true even when it is not.
59 estimate_edge_size and estimate_edge_growth can be used to query
60 function size/time in the given context. inline_merge_summary merges
61 properties of caller and callee after inlining.
63 Finally pass_inline_parameters is exported. This is used to drive
64 computation of function parameters used by the early inliner. IPA
65 inlined performs analysis via its analyze_function method. */
69 #include "coretypes.h"
72 #include "stor-layout.h"
73 #include "stringpool.h"
74 #include "print-tree.h"
75 #include "tree-inline.h"
76 #include "langhooks.h"
78 #include "diagnostic.h"
79 #include "gimple-pretty-print.h"
81 #include "tree-pass.h"
88 #include "hard-reg-set.h"
91 #include "dominance.h"
94 #include "basic-block.h"
95 #include "tree-ssa-alias.h"
96 #include "internal-fn.h"
97 #include "gimple-expr.h"
100 #include "gimple-iterator.h"
101 #include "gimple-ssa.h"
102 #include "tree-cfg.h"
103 #include "tree-phinodes.h"
104 #include "ssa-iterators.h"
105 #include "tree-ssanames.h"
106 #include "tree-ssa-loop-niter.h"
107 #include "tree-ssa-loop.h"
108 #include "hash-map.h"
109 #include "plugin-api.h"
112 #include "alloc-pool.h"
113 #include "ipa-prop.h"
114 #include "lto-streamer.h"
115 #include "data-streamer.h"
116 #include "tree-streamer.h"
117 #include "ipa-inline.h"
119 #include "tree-scalar-evolution.h"
120 #include "ipa-utils.h"
122 #include "cfgexpand.h"
124 /* Estimate runtime of function can easilly run into huge numbers with many
125 nested loops. Be sure we can compute time * INLINE_SIZE_SCALE * 2 in an
126 integer. For anything larger we use gcov_type. */
127 #define MAX_TIME 500000
129 /* Number of bits in integer, but we really want to be stable across different
131 #define NUM_CONDITIONS 32
133 enum predicate_conditions
135 predicate_false_condition
= 0,
136 predicate_not_inlined_condition
= 1,
137 predicate_first_dynamic_condition
= 2
140 /* Special condition code we use to represent test that operand is compile time
142 #define IS_NOT_CONSTANT ERROR_MARK
143 /* Special condition code we use to represent test that operand is not changed
144 across invocation of the function. When operand IS_NOT_CONSTANT it is always
145 CHANGED, however i.e. loop invariants can be NOT_CHANGED given percentage
146 of executions even when they are not compile time constants. */
147 #define CHANGED IDENTIFIER_NODE
149 /* Holders of ipa cgraph hooks: */
150 static struct cgraph_node_hook_list
*function_insertion_hook_holder
;
151 static struct cgraph_node_hook_list
*node_removal_hook_holder
;
152 static struct cgraph_2node_hook_list
*node_duplication_hook_holder
;
153 static struct cgraph_2edge_hook_list
*edge_duplication_hook_holder
;
154 static struct cgraph_edge_hook_list
*edge_removal_hook_holder
;
155 static void inline_node_removal_hook (struct cgraph_node
*, void *);
156 static void inline_node_duplication_hook (struct cgraph_node
*,
157 struct cgraph_node
*, void *);
158 static void inline_edge_removal_hook (struct cgraph_edge
*, void *);
159 static void inline_edge_duplication_hook (struct cgraph_edge
*,
160 struct cgraph_edge
*, void *);
162 /* VECtor holding inline summaries.
163 In GGC memory because conditions might point to constant trees. */
164 vec
<inline_summary_t
, va_gc
> *inline_summary_vec
;
165 vec
<inline_edge_summary_t
> inline_edge_summary_vec
;
167 /* Cached node/edge growths. */
168 vec
<int> node_growth_cache
;
169 vec
<edge_growth_cache_entry
> edge_growth_cache
;
171 /* Edge predicates goes here. */
172 static alloc_pool edge_predicate_pool
;
174 /* Return true predicate (tautology).
175 We represent it by empty list of clauses. */
177 static inline struct predicate
178 true_predicate (void)
186 /* Return predicate testing single condition number COND. */
188 static inline struct predicate
189 single_cond_predicate (int cond
)
192 p
.clause
[0] = 1 << cond
;
198 /* Return false predicate. First clause require false condition. */
200 static inline struct predicate
201 false_predicate (void)
203 return single_cond_predicate (predicate_false_condition
);
207 /* Return true if P is (true). */
210 true_predicate_p (struct predicate
*p
)
212 return !p
->clause
[0];
216 /* Return true if P is (false). */
219 false_predicate_p (struct predicate
*p
)
221 if (p
->clause
[0] == (1 << predicate_false_condition
))
223 gcc_checking_assert (!p
->clause
[1]
224 && p
->clause
[0] == 1 << predicate_false_condition
);
231 /* Return predicate that is set true when function is not inlined. */
233 static inline struct predicate
234 not_inlined_predicate (void)
236 return single_cond_predicate (predicate_not_inlined_condition
);
239 /* Simple description of whether a memory load or a condition refers to a load
240 from an aggregate and if so, how and where from in the aggregate.
241 Individual fields have the same meaning like fields with the same name in
244 struct agg_position_info
246 HOST_WIDE_INT offset
;
251 /* Add condition to condition list CONDS. AGGPOS describes whether the used
252 oprand is loaded from an aggregate and where in the aggregate it is. It can
253 be NULL, which means this not a load from an aggregate. */
255 static struct predicate
256 add_condition (struct inline_summary
*summary
, int operand_num
,
257 struct agg_position_info
*aggpos
,
258 enum tree_code code
, tree val
)
262 struct condition new_cond
;
263 HOST_WIDE_INT offset
;
264 bool agg_contents
, by_ref
;
268 offset
= aggpos
->offset
;
269 agg_contents
= aggpos
->agg_contents
;
270 by_ref
= aggpos
->by_ref
;
275 agg_contents
= false;
279 gcc_checking_assert (operand_num
>= 0);
280 for (i
= 0; vec_safe_iterate (summary
->conds
, i
, &c
); i
++)
282 if (c
->operand_num
== operand_num
285 && c
->agg_contents
== agg_contents
286 && (!agg_contents
|| (c
->offset
== offset
&& c
->by_ref
== by_ref
)))
287 return single_cond_predicate (i
+ predicate_first_dynamic_condition
);
289 /* Too many conditions. Give up and return constant true. */
290 if (i
== NUM_CONDITIONS
- predicate_first_dynamic_condition
)
291 return true_predicate ();
293 new_cond
.operand_num
= operand_num
;
294 new_cond
.code
= code
;
296 new_cond
.agg_contents
= agg_contents
;
297 new_cond
.by_ref
= by_ref
;
298 new_cond
.offset
= offset
;
299 vec_safe_push (summary
->conds
, new_cond
);
300 return single_cond_predicate (i
+ predicate_first_dynamic_condition
);
304 /* Add clause CLAUSE into the predicate P. */
307 add_clause (conditions conditions
, struct predicate
*p
, clause_t clause
)
311 int insert_here
= -1;
318 /* False clause makes the whole predicate false. Kill the other variants. */
319 if (clause
== (1 << predicate_false_condition
))
321 p
->clause
[0] = (1 << predicate_false_condition
);
325 if (false_predicate_p (p
))
328 /* No one should be silly enough to add false into nontrivial clauses. */
329 gcc_checking_assert (!(clause
& (1 << predicate_false_condition
)));
331 /* Look where to insert the clause. At the same time prune out
332 clauses of P that are implied by the new clause and thus
334 for (i
= 0, i2
= 0; i
<= MAX_CLAUSES
; i
++)
336 p
->clause
[i2
] = p
->clause
[i
];
341 /* If p->clause[i] implies clause, there is nothing to add. */
342 if ((p
->clause
[i
] & clause
) == p
->clause
[i
])
344 /* We had nothing to add, none of clauses should've become
346 gcc_checking_assert (i
== i2
);
350 if (p
->clause
[i
] < clause
&& insert_here
< 0)
353 /* If clause implies p->clause[i], then p->clause[i] becomes redundant.
354 Otherwise the p->clause[i] has to stay. */
355 if ((p
->clause
[i
] & clause
) != clause
)
359 /* Look for clauses that are obviously true. I.e.
360 op0 == 5 || op0 != 5. */
361 for (c1
= predicate_first_dynamic_condition
; c1
< NUM_CONDITIONS
; c1
++)
364 if (!(clause
& (1 << c1
)))
366 cc1
= &(*conditions
)[c1
- predicate_first_dynamic_condition
];
367 /* We have no way to represent !CHANGED and !IS_NOT_CONSTANT
368 and thus there is no point for looking for them. */
369 if (cc1
->code
== CHANGED
|| cc1
->code
== IS_NOT_CONSTANT
)
371 for (c2
= c1
+ 1; c2
< NUM_CONDITIONS
; c2
++)
372 if (clause
& (1 << c2
))
375 &(*conditions
)[c1
- predicate_first_dynamic_condition
];
377 &(*conditions
)[c2
- predicate_first_dynamic_condition
];
378 if (cc1
->operand_num
== cc2
->operand_num
379 && cc1
->val
== cc2
->val
380 && cc2
->code
!= IS_NOT_CONSTANT
381 && cc2
->code
!= CHANGED
382 && cc1
->code
== invert_tree_comparison
384 HONOR_NANS (TYPE_MODE (TREE_TYPE (cc1
->val
)))))
390 /* We run out of variants. Be conservative in positive direction. */
391 if (i2
== MAX_CLAUSES
)
393 /* Keep clauses in decreasing order. This makes equivalence testing easy. */
394 p
->clause
[i2
+ 1] = 0;
395 if (insert_here
>= 0)
396 for (; i2
> insert_here
; i2
--)
397 p
->clause
[i2
] = p
->clause
[i2
- 1];
400 p
->clause
[insert_here
] = clause
;
406 static struct predicate
407 and_predicates (conditions conditions
,
408 struct predicate
*p
, struct predicate
*p2
)
410 struct predicate out
= *p
;
413 /* Avoid busy work. */
414 if (false_predicate_p (p2
) || true_predicate_p (p
))
416 if (false_predicate_p (p
) || true_predicate_p (p2
))
419 /* See how far predicates match. */
420 for (i
= 0; p
->clause
[i
] && p
->clause
[i
] == p2
->clause
[i
]; i
++)
422 gcc_checking_assert (i
< MAX_CLAUSES
);
425 /* Combine the predicates rest. */
426 for (; p2
->clause
[i
]; i
++)
428 gcc_checking_assert (i
< MAX_CLAUSES
);
429 add_clause (conditions
, &out
, p2
->clause
[i
]);
435 /* Return true if predicates are obviously equal. */
438 predicates_equal_p (struct predicate
*p
, struct predicate
*p2
)
441 for (i
= 0; p
->clause
[i
]; i
++)
443 gcc_checking_assert (i
< MAX_CLAUSES
);
444 gcc_checking_assert (p
->clause
[i
] > p
->clause
[i
+ 1]);
445 gcc_checking_assert (!p2
->clause
[i
]
446 || p2
->clause
[i
] > p2
->clause
[i
+ 1]);
447 if (p
->clause
[i
] != p2
->clause
[i
])
450 return !p2
->clause
[i
];
456 static struct predicate
457 or_predicates (conditions conditions
,
458 struct predicate
*p
, struct predicate
*p2
)
460 struct predicate out
= true_predicate ();
463 /* Avoid busy work. */
464 if (false_predicate_p (p2
) || true_predicate_p (p
))
466 if (false_predicate_p (p
) || true_predicate_p (p2
))
468 if (predicates_equal_p (p
, p2
))
471 /* OK, combine the predicates. */
472 for (i
= 0; p
->clause
[i
]; i
++)
473 for (j
= 0; p2
->clause
[j
]; j
++)
475 gcc_checking_assert (i
< MAX_CLAUSES
&& j
< MAX_CLAUSES
);
476 add_clause (conditions
, &out
, p
->clause
[i
] | p2
->clause
[j
]);
482 /* Having partial truth assignment in POSSIBLE_TRUTHS, return false
483 if predicate P is known to be false. */
486 evaluate_predicate (struct predicate
*p
, clause_t possible_truths
)
490 /* True remains true. */
491 if (true_predicate_p (p
))
494 gcc_assert (!(possible_truths
& (1 << predicate_false_condition
)));
496 /* See if we can find clause we can disprove. */
497 for (i
= 0; p
->clause
[i
]; i
++)
499 gcc_checking_assert (i
< MAX_CLAUSES
);
500 if (!(p
->clause
[i
] & possible_truths
))
506 /* Return the probability in range 0...REG_BR_PROB_BASE that the predicated
507 instruction will be recomputed per invocation of the inlined call. */
510 predicate_probability (conditions conds
,
511 struct predicate
*p
, clause_t possible_truths
,
512 vec
<inline_param_summary
> inline_param_summary
)
515 int combined_prob
= REG_BR_PROB_BASE
;
517 /* True remains true. */
518 if (true_predicate_p (p
))
519 return REG_BR_PROB_BASE
;
521 if (false_predicate_p (p
))
524 gcc_assert (!(possible_truths
& (1 << predicate_false_condition
)));
526 /* See if we can find clause we can disprove. */
527 for (i
= 0; p
->clause
[i
]; i
++)
529 gcc_checking_assert (i
< MAX_CLAUSES
);
530 if (!(p
->clause
[i
] & possible_truths
))
536 if (!inline_param_summary
.exists ())
537 return REG_BR_PROB_BASE
;
538 for (i2
= 0; i2
< NUM_CONDITIONS
; i2
++)
539 if ((p
->clause
[i
] & possible_truths
) & (1 << i2
))
541 if (i2
>= predicate_first_dynamic_condition
)
544 &(*conds
)[i2
- predicate_first_dynamic_condition
];
545 if (c
->code
== CHANGED
547 (int) inline_param_summary
.length ()))
550 inline_param_summary
[c
->operand_num
].change_prob
;
551 this_prob
= MAX (this_prob
, iprob
);
554 this_prob
= REG_BR_PROB_BASE
;
557 this_prob
= REG_BR_PROB_BASE
;
559 combined_prob
= MIN (this_prob
, combined_prob
);
564 return combined_prob
;
568 /* Dump conditional COND. */
571 dump_condition (FILE *f
, conditions conditions
, int cond
)
574 if (cond
== predicate_false_condition
)
575 fprintf (f
, "false");
576 else if (cond
== predicate_not_inlined_condition
)
577 fprintf (f
, "not inlined");
580 c
= &(*conditions
)[cond
- predicate_first_dynamic_condition
];
581 fprintf (f
, "op%i", c
->operand_num
);
583 fprintf (f
, "[%soffset: " HOST_WIDE_INT_PRINT_DEC
"]",
584 c
->by_ref
? "ref " : "", c
->offset
);
585 if (c
->code
== IS_NOT_CONSTANT
)
587 fprintf (f
, " not constant");
590 if (c
->code
== CHANGED
)
592 fprintf (f
, " changed");
595 fprintf (f
, " %s ", op_symbol_code (c
->code
));
596 print_generic_expr (f
, c
->val
, 1);
601 /* Dump clause CLAUSE. */
604 dump_clause (FILE *f
, conditions conds
, clause_t clause
)
611 for (i
= 0; i
< NUM_CONDITIONS
; i
++)
612 if (clause
& (1 << i
))
617 dump_condition (f
, conds
, i
);
623 /* Dump predicate PREDICATE. */
626 dump_predicate (FILE *f
, conditions conds
, struct predicate
*pred
)
629 if (true_predicate_p (pred
))
630 dump_clause (f
, conds
, 0);
632 for (i
= 0; pred
->clause
[i
]; i
++)
636 dump_clause (f
, conds
, pred
->clause
[i
]);
642 /* Dump inline hints. */
644 dump_inline_hints (FILE *f
, inline_hints hints
)
648 fprintf (f
, "inline hints:");
649 if (hints
& INLINE_HINT_indirect_call
)
651 hints
&= ~INLINE_HINT_indirect_call
;
652 fprintf (f
, " indirect_call");
654 if (hints
& INLINE_HINT_loop_iterations
)
656 hints
&= ~INLINE_HINT_loop_iterations
;
657 fprintf (f
, " loop_iterations");
659 if (hints
& INLINE_HINT_loop_stride
)
661 hints
&= ~INLINE_HINT_loop_stride
;
662 fprintf (f
, " loop_stride");
664 if (hints
& INLINE_HINT_same_scc
)
666 hints
&= ~INLINE_HINT_same_scc
;
667 fprintf (f
, " same_scc");
669 if (hints
& INLINE_HINT_in_scc
)
671 hints
&= ~INLINE_HINT_in_scc
;
672 fprintf (f
, " in_scc");
674 if (hints
& INLINE_HINT_cross_module
)
676 hints
&= ~INLINE_HINT_cross_module
;
677 fprintf (f
, " cross_module");
679 if (hints
& INLINE_HINT_declared_inline
)
681 hints
&= ~INLINE_HINT_declared_inline
;
682 fprintf (f
, " declared_inline");
684 if (hints
& INLINE_HINT_array_index
)
686 hints
&= ~INLINE_HINT_array_index
;
687 fprintf (f
, " array_index");
689 if (hints
& INLINE_HINT_known_hot
)
691 hints
&= ~INLINE_HINT_known_hot
;
692 fprintf (f
, " known_hot");
698 /* Record SIZE and TIME under condition PRED into the inline summary. */
701 account_size_time (struct inline_summary
*summary
, int size
, int time
,
702 struct predicate
*pred
)
708 if (false_predicate_p (pred
))
711 /* We need to create initial empty unconitional clause, but otherwie
712 we don't need to account empty times and sizes. */
713 if (!size
&& !time
&& summary
->entry
)
716 /* Watch overflow that might result from insane profiles. */
717 if (time
> MAX_TIME
* INLINE_TIME_SCALE
)
718 time
= MAX_TIME
* INLINE_TIME_SCALE
;
719 gcc_assert (time
>= 0);
721 for (i
= 0; vec_safe_iterate (summary
->entry
, i
, &e
); i
++)
722 if (predicates_equal_p (&e
->predicate
, pred
))
731 e
= &(*summary
->entry
)[0];
732 gcc_assert (!e
->predicate
.clause
[0]);
733 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
735 "\t\tReached limit on number of entries, "
736 "ignoring the predicate.");
738 if (dump_file
&& (dump_flags
& TDF_DETAILS
) && (time
|| size
))
741 "\t\tAccounting size:%3.2f, time:%3.2f on %spredicate:",
742 ((double) size
) / INLINE_SIZE_SCALE
,
743 ((double) time
) / INLINE_TIME_SCALE
, found
? "" : "new ");
744 dump_predicate (dump_file
, summary
->conds
, pred
);
748 struct size_time_entry new_entry
;
749 new_entry
.size
= size
;
750 new_entry
.time
= time
;
751 new_entry
.predicate
= *pred
;
752 vec_safe_push (summary
->entry
, new_entry
);
758 if (e
->time
> MAX_TIME
* INLINE_TIME_SCALE
)
759 e
->time
= MAX_TIME
* INLINE_TIME_SCALE
;
763 /* Set predicate for edge E. */
766 edge_set_predicate (struct cgraph_edge
*e
, struct predicate
*predicate
)
768 struct inline_edge_summary
*es
= inline_edge_summary (e
);
770 /* If the edge is determined to be never executed, redirect it
771 to BUILTIN_UNREACHABLE to save inliner from inlining into it. */
772 if (predicate
&& false_predicate_p (predicate
) && e
->callee
)
774 struct cgraph_node
*callee
= !e
->inline_failed
? e
->callee
: NULL
;
776 e
->redirect_callee (cgraph_node::get_create
777 (builtin_decl_implicit (BUILT_IN_UNREACHABLE
)));
778 e
->inline_failed
= CIF_UNREACHABLE
;
780 callee
->remove_symbol_and_inline_clones ();
782 if (predicate
&& !true_predicate_p (predicate
))
785 es
->predicate
= (struct predicate
*) pool_alloc (edge_predicate_pool
);
786 *es
->predicate
= *predicate
;
791 pool_free (edge_predicate_pool
, es
->predicate
);
792 es
->predicate
= NULL
;
796 /* Set predicate for hint *P. */
799 set_hint_predicate (struct predicate
**p
, struct predicate new_predicate
)
801 if (false_predicate_p (&new_predicate
) || true_predicate_p (&new_predicate
))
804 pool_free (edge_predicate_pool
, *p
);
810 *p
= (struct predicate
*) pool_alloc (edge_predicate_pool
);
816 /* KNOWN_VALS is partial mapping of parameters of NODE to constant values.
817 KNOWN_AGGS is a vector of aggreggate jump functions for each parameter.
818 Return clause of possible truths. When INLINE_P is true, assume that we are
821 ERROR_MARK means compile time invariant. */
824 evaluate_conditions_for_known_args (struct cgraph_node
*node
,
826 vec
<tree
> known_vals
,
827 vec
<ipa_agg_jump_function_p
>
830 clause_t clause
= inline_p
? 0 : 1 << predicate_not_inlined_condition
;
831 struct inline_summary
*info
= inline_summary (node
);
835 for (i
= 0; vec_safe_iterate (info
->conds
, i
, &c
); i
++)
840 /* We allow call stmt to have fewer arguments than the callee function
841 (especially for K&R style programs). So bound check here (we assume
842 known_aggs vector, if non-NULL, has the same length as
844 gcc_checking_assert (!known_aggs
.exists ()
845 || (known_vals
.length () == known_aggs
.length ()));
846 if (c
->operand_num
>= (int) known_vals
.length ())
848 clause
|= 1 << (i
+ predicate_first_dynamic_condition
);
854 struct ipa_agg_jump_function
*agg
;
856 if (c
->code
== CHANGED
858 && (known_vals
[c
->operand_num
] == error_mark_node
))
861 if (known_aggs
.exists ())
863 agg
= known_aggs
[c
->operand_num
];
864 val
= ipa_find_agg_cst_for_param (agg
, c
->offset
, c
->by_ref
);
871 val
= known_vals
[c
->operand_num
];
872 if (val
== error_mark_node
&& c
->code
!= CHANGED
)
878 clause
|= 1 << (i
+ predicate_first_dynamic_condition
);
881 if (c
->code
== IS_NOT_CONSTANT
|| c
->code
== CHANGED
)
883 res
= fold_binary_to_constant (c
->code
, boolean_type_node
, val
, c
->val
);
884 if (res
&& integer_zerop (res
))
886 clause
|= 1 << (i
+ predicate_first_dynamic_condition
);
892 /* Work out what conditions might be true at invocation of E. */
895 evaluate_properties_for_edge (struct cgraph_edge
*e
, bool inline_p
,
896 clause_t
*clause_ptr
,
897 vec
<tree
> *known_vals_ptr
,
898 vec
<ipa_polymorphic_call_context
>
900 vec
<ipa_agg_jump_function_p
> *known_aggs_ptr
)
902 struct cgraph_node
*callee
= e
->callee
->ultimate_alias_target ();
903 struct inline_summary
*info
= inline_summary (callee
);
904 vec
<tree
> known_vals
= vNULL
;
905 vec
<ipa_agg_jump_function_p
> known_aggs
= vNULL
;
908 *clause_ptr
= inline_p
? 0 : 1 << predicate_not_inlined_condition
;
910 known_vals_ptr
->create (0);
911 if (known_contexts_ptr
)
912 known_contexts_ptr
->create (0);
914 if (ipa_node_params_vector
.exists ()
915 && !e
->call_stmt_cannot_inline_p
916 && ((clause_ptr
&& info
->conds
) || known_vals_ptr
|| known_contexts_ptr
))
918 struct ipa_node_params
*parms_info
;
919 struct ipa_edge_args
*args
= IPA_EDGE_REF (e
);
920 struct inline_edge_summary
*es
= inline_edge_summary (e
);
921 int i
, count
= ipa_get_cs_argument_count (args
);
923 if (e
->caller
->global
.inlined_to
)
924 parms_info
= IPA_NODE_REF (e
->caller
->global
.inlined_to
);
926 parms_info
= IPA_NODE_REF (e
->caller
);
928 if (count
&& (info
->conds
|| known_vals_ptr
))
929 known_vals
.safe_grow_cleared (count
);
930 if (count
&& (info
->conds
|| known_aggs_ptr
))
931 known_aggs
.safe_grow_cleared (count
);
932 if (count
&& known_contexts_ptr
)
933 known_contexts_ptr
->safe_grow_cleared (count
);
935 for (i
= 0; i
< count
; i
++)
937 struct ipa_jump_func
*jf
= ipa_get_ith_jump_func (args
, i
);
938 tree cst
= ipa_value_from_jfunc (parms_info
, jf
);
941 gcc_checking_assert (TREE_CODE (cst
) != TREE_BINFO
);
942 if (known_vals
.exists ())
945 else if (inline_p
&& !es
->param
[i
].change_prob
)
946 known_vals
[i
] = error_mark_node
;
948 if (known_contexts_ptr
)
949 (*known_contexts_ptr
)[i
] = ipa_context_from_jfunc (parms_info
, e
,
951 /* TODO: When IPA-CP starts propagating and merging aggregate jump
952 functions, use its knowledge of the caller too, just like the
953 scalar case above. */
954 known_aggs
[i
] = &jf
->agg
;
959 *clause_ptr
= evaluate_conditions_for_known_args (callee
, inline_p
,
960 known_vals
, known_aggs
);
963 *known_vals_ptr
= known_vals
;
965 known_vals
.release ();
968 *known_aggs_ptr
= known_aggs
;
970 known_aggs
.release ();
974 /* Allocate the inline summary vector or resize it to cover all cgraph nodes. */
977 inline_summary_alloc (void)
979 if (!node_removal_hook_holder
)
980 node_removal_hook_holder
=
981 symtab
->add_cgraph_removal_hook (&inline_node_removal_hook
, NULL
);
982 if (!edge_removal_hook_holder
)
983 edge_removal_hook_holder
=
984 symtab
->add_edge_removal_hook (&inline_edge_removal_hook
, NULL
);
985 if (!node_duplication_hook_holder
)
986 node_duplication_hook_holder
=
987 symtab
->add_cgraph_duplication_hook (&inline_node_duplication_hook
, NULL
);
988 if (!edge_duplication_hook_holder
)
989 edge_duplication_hook_holder
=
990 symtab
->add_edge_duplication_hook (&inline_edge_duplication_hook
, NULL
);
992 if (vec_safe_length (inline_summary_vec
) <= (unsigned) symtab
->cgraph_max_uid
)
993 vec_safe_grow_cleared (inline_summary_vec
, symtab
->cgraph_max_uid
+ 1);
994 if (inline_edge_summary_vec
.length () <= (unsigned) symtab
->edges_max_uid
)
995 inline_edge_summary_vec
.safe_grow_cleared (symtab
->edges_max_uid
+ 1);
996 if (!edge_predicate_pool
)
997 edge_predicate_pool
= create_alloc_pool ("edge predicates",
998 sizeof (struct predicate
), 10);
1001 /* We are called multiple time for given function; clear
1002 data from previous run so they are not cumulated. */
1005 reset_inline_edge_summary (struct cgraph_edge
*e
)
1007 if (e
->uid
< (int) inline_edge_summary_vec
.length ())
1009 struct inline_edge_summary
*es
= inline_edge_summary (e
);
1011 es
->call_stmt_size
= es
->call_stmt_time
= 0;
1013 pool_free (edge_predicate_pool
, es
->predicate
);
1014 es
->predicate
= NULL
;
1015 es
->param
.release ();
1019 /* We are called multiple time for given function; clear
1020 data from previous run so they are not cumulated. */
1023 reset_inline_summary (struct cgraph_node
*node
)
1025 struct inline_summary
*info
= inline_summary (node
);
1026 struct cgraph_edge
*e
;
1028 info
->self_size
= info
->self_time
= 0;
1029 info
->estimated_stack_size
= 0;
1030 info
->estimated_self_stack_size
= 0;
1031 info
->stack_frame_offset
= 0;
1036 if (info
->loop_iterations
)
1038 pool_free (edge_predicate_pool
, info
->loop_iterations
);
1039 info
->loop_iterations
= NULL
;
1041 if (info
->loop_stride
)
1043 pool_free (edge_predicate_pool
, info
->loop_stride
);
1044 info
->loop_stride
= NULL
;
1046 if (info
->array_index
)
1048 pool_free (edge_predicate_pool
, info
->array_index
);
1049 info
->array_index
= NULL
;
1051 vec_free (info
->conds
);
1052 vec_free (info
->entry
);
1053 for (e
= node
->callees
; e
; e
= e
->next_callee
)
1054 reset_inline_edge_summary (e
);
1055 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
1056 reset_inline_edge_summary (e
);
1059 /* Hook that is called by cgraph.c when a node is removed. */
1062 inline_node_removal_hook (struct cgraph_node
*node
,
1063 void *data ATTRIBUTE_UNUSED
)
1065 struct inline_summary
*info
;
1066 if (vec_safe_length (inline_summary_vec
) <= (unsigned) node
->uid
)
1068 info
= inline_summary (node
);
1069 reset_inline_summary (node
);
1070 memset (info
, 0, sizeof (inline_summary_t
));
1073 /* Remap predicate P of former function to be predicate of duplicated function.
1074 POSSIBLE_TRUTHS is clause of possible truths in the duplicated node,
1075 INFO is inline summary of the duplicated node. */
1077 static struct predicate
1078 remap_predicate_after_duplication (struct predicate
*p
,
1079 clause_t possible_truths
,
1080 struct inline_summary
*info
)
1082 struct predicate new_predicate
= true_predicate ();
1084 for (j
= 0; p
->clause
[j
]; j
++)
1085 if (!(possible_truths
& p
->clause
[j
]))
1087 new_predicate
= false_predicate ();
1091 add_clause (info
->conds
, &new_predicate
,
1092 possible_truths
& p
->clause
[j
]);
1093 return new_predicate
;
1096 /* Same as remap_predicate_after_duplication but handle hint predicate *P.
1097 Additionally care about allocating new memory slot for updated predicate
1098 and set it to NULL when it becomes true or false (and thus uninteresting).
1102 remap_hint_predicate_after_duplication (struct predicate
**p
,
1103 clause_t possible_truths
,
1104 struct inline_summary
*info
)
1106 struct predicate new_predicate
;
1111 new_predicate
= remap_predicate_after_duplication (*p
,
1112 possible_truths
, info
);
1113 /* We do not want to free previous predicate; it is used by node origin. */
1115 set_hint_predicate (p
, new_predicate
);
1119 /* Hook that is called by cgraph.c when a node is duplicated. */
1122 inline_node_duplication_hook (struct cgraph_node
*src
,
1123 struct cgraph_node
*dst
,
1124 ATTRIBUTE_UNUSED
void *data
)
1126 struct inline_summary
*info
;
1127 inline_summary_alloc ();
1128 info
= inline_summary (dst
);
1129 memcpy (info
, inline_summary (src
), sizeof (struct inline_summary
));
1130 /* TODO: as an optimization, we may avoid copying conditions
1131 that are known to be false or true. */
1132 info
->conds
= vec_safe_copy (info
->conds
);
1134 /* When there are any replacements in the function body, see if we can figure
1135 out that something was optimized out. */
1136 if (ipa_node_params_vector
.exists () && dst
->clone
.tree_map
)
1138 vec
<size_time_entry
, va_gc
> *entry
= info
->entry
;
1139 /* Use SRC parm info since it may not be copied yet. */
1140 struct ipa_node_params
*parms_info
= IPA_NODE_REF (src
);
1141 vec
<tree
> known_vals
= vNULL
;
1142 int count
= ipa_get_param_count (parms_info
);
1144 clause_t possible_truths
;
1145 struct predicate true_pred
= true_predicate ();
1147 int optimized_out_size
= 0;
1148 bool inlined_to_p
= false;
1149 struct cgraph_edge
*edge
;
1152 known_vals
.safe_grow_cleared (count
);
1153 for (i
= 0; i
< count
; i
++)
1155 struct ipa_replace_map
*r
;
1157 for (j
= 0; vec_safe_iterate (dst
->clone
.tree_map
, j
, &r
); j
++)
1159 if (((!r
->old_tree
&& r
->parm_num
== i
)
1160 || (r
->old_tree
&& r
->old_tree
== ipa_get_param (parms_info
, i
)))
1161 && r
->replace_p
&& !r
->ref_p
)
1163 known_vals
[i
] = r
->new_tree
;
1168 possible_truths
= evaluate_conditions_for_known_args (dst
, false,
1171 known_vals
.release ();
1173 account_size_time (info
, 0, 0, &true_pred
);
1175 /* Remap size_time vectors.
1176 Simplify the predicate by prunning out alternatives that are known
1178 TODO: as on optimization, we can also eliminate conditions known
1180 for (i
= 0; vec_safe_iterate (entry
, i
, &e
); i
++)
1182 struct predicate new_predicate
;
1183 new_predicate
= remap_predicate_after_duplication (&e
->predicate
,
1186 if (false_predicate_p (&new_predicate
))
1187 optimized_out_size
+= e
->size
;
1189 account_size_time (info
, e
->size
, e
->time
, &new_predicate
);
1192 /* Remap edge predicates with the same simplification as above.
1193 Also copy constantness arrays. */
1194 for (edge
= dst
->callees
; edge
; edge
= edge
->next_callee
)
1196 struct predicate new_predicate
;
1197 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
1199 if (!edge
->inline_failed
)
1200 inlined_to_p
= true;
1203 new_predicate
= remap_predicate_after_duplication (es
->predicate
,
1206 if (false_predicate_p (&new_predicate
)
1207 && !false_predicate_p (es
->predicate
))
1209 optimized_out_size
+= es
->call_stmt_size
* INLINE_SIZE_SCALE
;
1210 edge
->frequency
= 0;
1212 edge_set_predicate (edge
, &new_predicate
);
1215 /* Remap indirect edge predicates with the same simplificaiton as above.
1216 Also copy constantness arrays. */
1217 for (edge
= dst
->indirect_calls
; edge
; edge
= edge
->next_callee
)
1219 struct predicate new_predicate
;
1220 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
1222 gcc_checking_assert (edge
->inline_failed
);
1225 new_predicate
= remap_predicate_after_duplication (es
->predicate
,
1228 if (false_predicate_p (&new_predicate
)
1229 && !false_predicate_p (es
->predicate
))
1231 optimized_out_size
+= es
->call_stmt_size
* INLINE_SIZE_SCALE
;
1232 edge
->frequency
= 0;
1234 edge_set_predicate (edge
, &new_predicate
);
1236 remap_hint_predicate_after_duplication (&info
->loop_iterations
,
1237 possible_truths
, info
);
1238 remap_hint_predicate_after_duplication (&info
->loop_stride
,
1239 possible_truths
, info
);
1240 remap_hint_predicate_after_duplication (&info
->array_index
,
1241 possible_truths
, info
);
1243 /* If inliner or someone after inliner will ever start producing
1244 non-trivial clones, we will get trouble with lack of information
1245 about updating self sizes, because size vectors already contains
1246 sizes of the calees. */
1247 gcc_assert (!inlined_to_p
|| !optimized_out_size
);
1251 info
->entry
= vec_safe_copy (info
->entry
);
1252 if (info
->loop_iterations
)
1254 predicate p
= *info
->loop_iterations
;
1255 info
->loop_iterations
= NULL
;
1256 set_hint_predicate (&info
->loop_iterations
, p
);
1258 if (info
->loop_stride
)
1260 predicate p
= *info
->loop_stride
;
1261 info
->loop_stride
= NULL
;
1262 set_hint_predicate (&info
->loop_stride
, p
);
1264 if (info
->array_index
)
1266 predicate p
= *info
->array_index
;
1267 info
->array_index
= NULL
;
1268 set_hint_predicate (&info
->array_index
, p
);
1271 inline_update_overall_summary (dst
);
1275 /* Hook that is called by cgraph.c when a node is duplicated. */
1278 inline_edge_duplication_hook (struct cgraph_edge
*src
,
1279 struct cgraph_edge
*dst
,
1280 ATTRIBUTE_UNUSED
void *data
)
1282 struct inline_edge_summary
*info
;
1283 struct inline_edge_summary
*srcinfo
;
1284 inline_summary_alloc ();
1285 info
= inline_edge_summary (dst
);
1286 srcinfo
= inline_edge_summary (src
);
1287 memcpy (info
, srcinfo
, sizeof (struct inline_edge_summary
));
1288 info
->predicate
= NULL
;
1289 edge_set_predicate (dst
, srcinfo
->predicate
);
1290 info
->param
= srcinfo
->param
.copy ();
1294 /* Keep edge cache consistent across edge removal. */
1297 inline_edge_removal_hook (struct cgraph_edge
*edge
,
1298 void *data ATTRIBUTE_UNUSED
)
1300 if (edge_growth_cache
.exists ())
1301 reset_edge_growth_cache (edge
);
1302 reset_inline_edge_summary (edge
);
1306 /* Initialize growth caches. */
1309 initialize_growth_caches (void)
1311 if (symtab
->edges_max_uid
)
1312 edge_growth_cache
.safe_grow_cleared (symtab
->edges_max_uid
);
1313 if (symtab
->cgraph_max_uid
)
1314 node_growth_cache
.safe_grow_cleared (symtab
->cgraph_max_uid
);
1318 /* Free growth caches. */
1321 free_growth_caches (void)
1323 edge_growth_cache
.release ();
1324 node_growth_cache
.release ();
1328 /* Dump edge summaries associated to NODE and recursively to all clones.
1329 Indent by INDENT. */
1332 dump_inline_edge_summary (FILE *f
, int indent
, struct cgraph_node
*node
,
1333 struct inline_summary
*info
)
1335 struct cgraph_edge
*edge
;
1336 for (edge
= node
->callees
; edge
; edge
= edge
->next_callee
)
1338 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
1339 struct cgraph_node
*callee
= edge
->callee
->ultimate_alias_target ();
1343 "%*s%s/%i %s\n%*s loop depth:%2i freq:%4i size:%2i"
1344 " time: %2i callee size:%2i stack:%2i",
1345 indent
, "", callee
->name (), callee
->order
,
1346 !edge
->inline_failed
1347 ? "inlined" : cgraph_inline_failed_string (edge
-> inline_failed
),
1348 indent
, "", es
->loop_depth
, edge
->frequency
,
1349 es
->call_stmt_size
, es
->call_stmt_time
,
1350 (int) inline_summary (callee
)->size
/ INLINE_SIZE_SCALE
,
1351 (int) inline_summary (callee
)->estimated_stack_size
);
1355 fprintf (f
, " predicate: ");
1356 dump_predicate (f
, info
->conds
, es
->predicate
);
1360 if (es
->param
.exists ())
1361 for (i
= 0; i
< (int) es
->param
.length (); i
++)
1363 int prob
= es
->param
[i
].change_prob
;
1366 fprintf (f
, "%*s op%i is compile time invariant\n",
1368 else if (prob
!= REG_BR_PROB_BASE
)
1369 fprintf (f
, "%*s op%i change %f%% of time\n", indent
+ 2, "", i
,
1370 prob
* 100.0 / REG_BR_PROB_BASE
);
1372 if (!edge
->inline_failed
)
1374 fprintf (f
, "%*sStack frame offset %i, callee self size %i,"
1375 " callee size %i\n",
1377 (int) inline_summary (callee
)->stack_frame_offset
,
1378 (int) inline_summary (callee
)->estimated_self_stack_size
,
1379 (int) inline_summary (callee
)->estimated_stack_size
);
1380 dump_inline_edge_summary (f
, indent
+ 2, callee
, info
);
1383 for (edge
= node
->indirect_calls
; edge
; edge
= edge
->next_callee
)
1385 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
1386 fprintf (f
, "%*sindirect call loop depth:%2i freq:%4i size:%2i"
1390 edge
->frequency
, es
->call_stmt_size
, es
->call_stmt_time
);
1393 fprintf (f
, "predicate: ");
1394 dump_predicate (f
, info
->conds
, es
->predicate
);
1403 dump_inline_summary (FILE *f
, struct cgraph_node
*node
)
1405 if (node
->definition
)
1407 struct inline_summary
*s
= inline_summary (node
);
1410 fprintf (f
, "Inline summary for %s/%i", node
->name (),
1412 if (DECL_DISREGARD_INLINE_LIMITS (node
->decl
))
1413 fprintf (f
, " always_inline");
1415 fprintf (f
, " inlinable");
1416 fprintf (f
, "\n self time: %i\n", s
->self_time
);
1417 fprintf (f
, " global time: %i\n", s
->time
);
1418 fprintf (f
, " self size: %i\n", s
->self_size
);
1419 fprintf (f
, " global size: %i\n", s
->size
);
1420 fprintf (f
, " min size: %i\n", s
->min_size
);
1421 fprintf (f
, " self stack: %i\n",
1422 (int) s
->estimated_self_stack_size
);
1423 fprintf (f
, " global stack: %i\n", (int) s
->estimated_stack_size
);
1425 fprintf (f
, " estimated growth:%i\n", (int) s
->growth
);
1427 fprintf (f
, " In SCC: %i\n", (int) s
->scc_no
);
1428 for (i
= 0; vec_safe_iterate (s
->entry
, i
, &e
); i
++)
1430 fprintf (f
, " size:%f, time:%f, predicate:",
1431 (double) e
->size
/ INLINE_SIZE_SCALE
,
1432 (double) e
->time
/ INLINE_TIME_SCALE
);
1433 dump_predicate (f
, s
->conds
, &e
->predicate
);
1435 if (s
->loop_iterations
)
1437 fprintf (f
, " loop iterations:");
1438 dump_predicate (f
, s
->conds
, s
->loop_iterations
);
1442 fprintf (f
, " loop stride:");
1443 dump_predicate (f
, s
->conds
, s
->loop_stride
);
1447 fprintf (f
, " array index:");
1448 dump_predicate (f
, s
->conds
, s
->array_index
);
1450 fprintf (f
, " calls:\n");
1451 dump_inline_edge_summary (f
, 4, node
, s
);
1457 debug_inline_summary (struct cgraph_node
*node
)
1459 dump_inline_summary (stderr
, node
);
1463 dump_inline_summaries (FILE *f
)
1465 struct cgraph_node
*node
;
1467 FOR_EACH_DEFINED_FUNCTION (node
)
1468 if (!node
->global
.inlined_to
)
1469 dump_inline_summary (f
, node
);
1472 /* Give initial reasons why inlining would fail on EDGE. This gets either
1473 nullified or usually overwritten by more precise reasons later. */
1476 initialize_inline_failed (struct cgraph_edge
*e
)
1478 struct cgraph_node
*callee
= e
->callee
;
1480 if (e
->indirect_unknown_callee
)
1481 e
->inline_failed
= CIF_INDIRECT_UNKNOWN_CALL
;
1482 else if (!callee
->definition
)
1483 e
->inline_failed
= CIF_BODY_NOT_AVAILABLE
;
1484 else if (callee
->local
.redefined_extern_inline
)
1485 e
->inline_failed
= CIF_REDEFINED_EXTERN_INLINE
;
1486 else if (e
->call_stmt_cannot_inline_p
)
1487 e
->inline_failed
= CIF_MISMATCHED_ARGUMENTS
;
1488 else if (cfun
&& fn_contains_cilk_spawn_p (cfun
))
1489 /* We can't inline if the function is spawing a function. */
1490 e
->inline_failed
= CIF_FUNCTION_NOT_INLINABLE
;
1492 e
->inline_failed
= CIF_FUNCTION_NOT_CONSIDERED
;
1495 /* Callback of walk_aliased_vdefs. Flags that it has been invoked to the
1496 boolean variable pointed to by DATA. */
1499 mark_modified (ao_ref
*ao ATTRIBUTE_UNUSED
, tree vdef ATTRIBUTE_UNUSED
,
1502 bool *b
= (bool *) data
;
1507 /* If OP refers to value of function parameter, return the corresponding
1511 unmodified_parm_1 (gimple stmt
, tree op
)
1513 /* SSA_NAME referring to parm default def? */
1514 if (TREE_CODE (op
) == SSA_NAME
1515 && SSA_NAME_IS_DEFAULT_DEF (op
)
1516 && TREE_CODE (SSA_NAME_VAR (op
)) == PARM_DECL
)
1517 return SSA_NAME_VAR (op
);
1518 /* Non-SSA parm reference? */
1519 if (TREE_CODE (op
) == PARM_DECL
)
1521 bool modified
= false;
1524 ao_ref_init (&refd
, op
);
1525 walk_aliased_vdefs (&refd
, gimple_vuse (stmt
), mark_modified
, &modified
,
1533 /* If OP refers to value of function parameter, return the corresponding
1534 parameter. Also traverse chains of SSA register assignments. */
1537 unmodified_parm (gimple stmt
, tree op
)
1539 tree res
= unmodified_parm_1 (stmt
, op
);
1543 if (TREE_CODE (op
) == SSA_NAME
1544 && !SSA_NAME_IS_DEFAULT_DEF (op
)
1545 && gimple_assign_single_p (SSA_NAME_DEF_STMT (op
)))
1546 return unmodified_parm (SSA_NAME_DEF_STMT (op
),
1547 gimple_assign_rhs1 (SSA_NAME_DEF_STMT (op
)));
1551 /* If OP refers to a value of a function parameter or value loaded from an
1552 aggregate passed to a parameter (either by value or reference), return TRUE
1553 and store the number of the parameter to *INDEX_P and information whether
1554 and how it has been loaded from an aggregate into *AGGPOS. INFO describes
1555 the function parameters, STMT is the statement in which OP is used or
1559 unmodified_parm_or_parm_agg_item (struct ipa_node_params
*info
,
1560 gimple stmt
, tree op
, int *index_p
,
1561 struct agg_position_info
*aggpos
)
1563 tree res
= unmodified_parm_1 (stmt
, op
);
1565 gcc_checking_assert (aggpos
);
1568 *index_p
= ipa_get_param_decl_index (info
, res
);
1571 aggpos
->agg_contents
= false;
1572 aggpos
->by_ref
= false;
1576 if (TREE_CODE (op
) == SSA_NAME
)
1578 if (SSA_NAME_IS_DEFAULT_DEF (op
)
1579 || !gimple_assign_single_p (SSA_NAME_DEF_STMT (op
)))
1581 stmt
= SSA_NAME_DEF_STMT (op
);
1582 op
= gimple_assign_rhs1 (stmt
);
1583 if (!REFERENCE_CLASS_P (op
))
1584 return unmodified_parm_or_parm_agg_item (info
, stmt
, op
, index_p
,
1588 aggpos
->agg_contents
= true;
1589 return ipa_load_from_parm_agg (info
, stmt
, op
, index_p
, &aggpos
->offset
,
1593 /* See if statement might disappear after inlining.
1594 0 - means not eliminated
1595 1 - half of statements goes away
1596 2 - for sure it is eliminated.
1597 We are not terribly sophisticated, basically looking for simple abstraction
1598 penalty wrappers. */
1601 eliminated_by_inlining_prob (gimple stmt
)
1603 enum gimple_code code
= gimple_code (stmt
);
1604 enum tree_code rhs_code
;
1614 if (gimple_num_ops (stmt
) != 2)
1617 rhs_code
= gimple_assign_rhs_code (stmt
);
1619 /* Casts of parameters, loads from parameters passed by reference
1620 and stores to return value or parameters are often free after
1621 inlining dua to SRA and further combining.
1622 Assume that half of statements goes away. */
1623 if (CONVERT_EXPR_CODE_P (rhs_code
)
1624 || rhs_code
== VIEW_CONVERT_EXPR
1625 || rhs_code
== ADDR_EXPR
1626 || gimple_assign_rhs_class (stmt
) == GIMPLE_SINGLE_RHS
)
1628 tree rhs
= gimple_assign_rhs1 (stmt
);
1629 tree lhs
= gimple_assign_lhs (stmt
);
1630 tree inner_rhs
= get_base_address (rhs
);
1631 tree inner_lhs
= get_base_address (lhs
);
1632 bool rhs_free
= false;
1633 bool lhs_free
= false;
1640 /* Reads of parameter are expected to be free. */
1641 if (unmodified_parm (stmt
, inner_rhs
))
1643 /* Match expressions of form &this->field. Those will most likely
1644 combine with something upstream after inlining. */
1645 else if (TREE_CODE (inner_rhs
) == ADDR_EXPR
)
1647 tree op
= get_base_address (TREE_OPERAND (inner_rhs
, 0));
1648 if (TREE_CODE (op
) == PARM_DECL
)
1650 else if (TREE_CODE (op
) == MEM_REF
1651 && unmodified_parm (stmt
, TREE_OPERAND (op
, 0)))
1655 /* When parameter is not SSA register because its address is taken
1656 and it is just copied into one, the statement will be completely
1657 free after inlining (we will copy propagate backward). */
1658 if (rhs_free
&& is_gimple_reg (lhs
))
1661 /* Reads of parameters passed by reference
1662 expected to be free (i.e. optimized out after inlining). */
1663 if (TREE_CODE (inner_rhs
) == MEM_REF
1664 && unmodified_parm (stmt
, TREE_OPERAND (inner_rhs
, 0)))
1667 /* Copying parameter passed by reference into gimple register is
1668 probably also going to copy propagate, but we can't be quite
1670 if (rhs_free
&& is_gimple_reg (lhs
))
1673 /* Writes to parameters, parameters passed by value and return value
1674 (either dirrectly or passed via invisible reference) are free.
1676 TODO: We ought to handle testcase like
1677 struct a {int a,b;};
1679 retrurnsturct (void)
1685 This translate into:
1700 For that we either need to copy ipa-split logic detecting writes
1702 if (TREE_CODE (inner_lhs
) == PARM_DECL
1703 || TREE_CODE (inner_lhs
) == RESULT_DECL
1704 || (TREE_CODE (inner_lhs
) == MEM_REF
1705 && (unmodified_parm (stmt
, TREE_OPERAND (inner_lhs
, 0))
1706 || (TREE_CODE (TREE_OPERAND (inner_lhs
, 0)) == SSA_NAME
1707 && SSA_NAME_VAR (TREE_OPERAND (inner_lhs
, 0))
1708 && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND
1710 0))) == RESULT_DECL
))))
1713 && (is_gimple_reg (rhs
) || is_gimple_min_invariant (rhs
)))
1715 if (lhs_free
&& rhs_free
)
1725 /* If BB ends by a conditional we can turn into predicates, attach corresponding
1726 predicates to the CFG edges. */
1729 set_cond_stmt_execution_predicate (struct ipa_node_params
*info
,
1730 struct inline_summary
*summary
,
1736 struct agg_position_info aggpos
;
1737 enum tree_code code
, inverted_code
;
1743 last
= last_stmt (bb
);
1744 if (!last
|| gimple_code (last
) != GIMPLE_COND
)
1746 if (!is_gimple_ip_invariant (gimple_cond_rhs (last
)))
1748 op
= gimple_cond_lhs (last
);
1749 /* TODO: handle conditionals like
1752 if (unmodified_parm_or_parm_agg_item (info
, last
, op
, &index
, &aggpos
))
1754 code
= gimple_cond_code (last
);
1756 = invert_tree_comparison (code
,
1757 HONOR_NANS (TYPE_MODE (TREE_TYPE (op
))));
1759 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1761 enum tree_code this_code
= (e
->flags
& EDGE_TRUE_VALUE
1762 ? code
: inverted_code
);
1763 /* invert_tree_comparison will return ERROR_MARK on FP
1764 comparsions that are not EQ/NE instead of returning proper
1765 unordered one. Be sure it is not confused with NON_CONSTANT. */
1766 if (this_code
!= ERROR_MARK
)
1768 struct predicate p
= add_condition (summary
, index
, &aggpos
,
1770 gimple_cond_rhs (last
));
1771 e
->aux
= pool_alloc (edge_predicate_pool
);
1772 *(struct predicate
*) e
->aux
= p
;
1777 if (TREE_CODE (op
) != SSA_NAME
)
1780 if (builtin_constant_p (op))
1784 Here we can predicate nonconstant_code. We can't
1785 really handle constant_code since we have no predicate
1786 for this and also the constant code is not known to be
1787 optimized away when inliner doen't see operand is constant.
1788 Other optimizers might think otherwise. */
1789 if (gimple_cond_code (last
) != NE_EXPR
1790 || !integer_zerop (gimple_cond_rhs (last
)))
1792 set_stmt
= SSA_NAME_DEF_STMT (op
);
1793 if (!gimple_call_builtin_p (set_stmt
, BUILT_IN_CONSTANT_P
)
1794 || gimple_call_num_args (set_stmt
) != 1)
1796 op2
= gimple_call_arg (set_stmt
, 0);
1797 if (!unmodified_parm_or_parm_agg_item
1798 (info
, set_stmt
, op2
, &index
, &aggpos
))
1800 FOR_EACH_EDGE (e
, ei
, bb
->succs
) if (e
->flags
& EDGE_FALSE_VALUE
)
1802 struct predicate p
= add_condition (summary
, index
, &aggpos
,
1803 IS_NOT_CONSTANT
, NULL_TREE
);
1804 e
->aux
= pool_alloc (edge_predicate_pool
);
1805 *(struct predicate
*) e
->aux
= p
;
1810 /* If BB ends by a switch we can turn into predicates, attach corresponding
1811 predicates to the CFG edges. */
1814 set_switch_stmt_execution_predicate (struct ipa_node_params
*info
,
1815 struct inline_summary
*summary
,
1821 struct agg_position_info aggpos
;
1827 lastg
= last_stmt (bb
);
1828 if (!lastg
|| gimple_code (lastg
) != GIMPLE_SWITCH
)
1830 gswitch
*last
= as_a
<gswitch
*> (lastg
);
1831 op
= gimple_switch_index (last
);
1832 if (!unmodified_parm_or_parm_agg_item (info
, last
, op
, &index
, &aggpos
))
1835 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1837 e
->aux
= pool_alloc (edge_predicate_pool
);
1838 *(struct predicate
*) e
->aux
= false_predicate ();
1840 n
= gimple_switch_num_labels (last
);
1841 for (case_idx
= 0; case_idx
< n
; ++case_idx
)
1843 tree cl
= gimple_switch_label (last
, case_idx
);
1847 e
= find_edge (bb
, label_to_block (CASE_LABEL (cl
)));
1848 min
= CASE_LOW (cl
);
1849 max
= CASE_HIGH (cl
);
1851 /* For default we might want to construct predicate that none
1852 of cases is met, but it is bit hard to do not having negations
1853 of conditionals handy. */
1855 p
= true_predicate ();
1857 p
= add_condition (summary
, index
, &aggpos
, EQ_EXPR
, min
);
1860 struct predicate p1
, p2
;
1861 p1
= add_condition (summary
, index
, &aggpos
, GE_EXPR
, min
);
1862 p2
= add_condition (summary
, index
, &aggpos
, LE_EXPR
, max
);
1863 p
= and_predicates (summary
->conds
, &p1
, &p2
);
1865 *(struct predicate
*) e
->aux
1866 = or_predicates (summary
->conds
, &p
, (struct predicate
*) e
->aux
);
1871 /* For each BB in NODE attach to its AUX pointer predicate under
1872 which it is executable. */
1875 compute_bb_predicates (struct cgraph_node
*node
,
1876 struct ipa_node_params
*parms_info
,
1877 struct inline_summary
*summary
)
1879 struct function
*my_function
= DECL_STRUCT_FUNCTION (node
->decl
);
1883 FOR_EACH_BB_FN (bb
, my_function
)
1885 set_cond_stmt_execution_predicate (parms_info
, summary
, bb
);
1886 set_switch_stmt_execution_predicate (parms_info
, summary
, bb
);
1889 /* Entry block is always executable. */
1890 ENTRY_BLOCK_PTR_FOR_FN (my_function
)->aux
1891 = pool_alloc (edge_predicate_pool
);
1892 *(struct predicate
*) ENTRY_BLOCK_PTR_FOR_FN (my_function
)->aux
1893 = true_predicate ();
1895 /* A simple dataflow propagation of predicates forward in the CFG.
1896 TODO: work in reverse postorder. */
1900 FOR_EACH_BB_FN (bb
, my_function
)
1902 struct predicate p
= false_predicate ();
1905 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
1909 struct predicate this_bb_predicate
1910 = *(struct predicate
*) e
->src
->aux
;
1913 = and_predicates (summary
->conds
, &this_bb_predicate
,
1914 (struct predicate
*) e
->aux
);
1915 p
= or_predicates (summary
->conds
, &p
, &this_bb_predicate
);
1916 if (true_predicate_p (&p
))
1920 if (false_predicate_p (&p
))
1921 gcc_assert (!bb
->aux
);
1927 bb
->aux
= pool_alloc (edge_predicate_pool
);
1928 *((struct predicate
*) bb
->aux
) = p
;
1930 else if (!predicates_equal_p (&p
, (struct predicate
*) bb
->aux
))
1932 /* This OR operation is needed to ensure monotonous data flow
1933 in the case we hit the limit on number of clauses and the
1934 and/or operations above give approximate answers. */
1935 p
= or_predicates (summary
->conds
, &p
, (struct predicate
*)bb
->aux
);
1936 if (!predicates_equal_p (&p
, (struct predicate
*) bb
->aux
))
1939 *((struct predicate
*) bb
->aux
) = p
;
1948 /* We keep info about constantness of SSA names. */
1950 typedef struct predicate predicate_t
;
1951 /* Return predicate specifying when the STMT might have result that is not
1952 a compile time constant. */
1954 static struct predicate
1955 will_be_nonconstant_expr_predicate (struct ipa_node_params
*info
,
1956 struct inline_summary
*summary
,
1958 vec
<predicate_t
> nonconstant_names
)
1963 while (UNARY_CLASS_P (expr
))
1964 expr
= TREE_OPERAND (expr
, 0);
1966 parm
= unmodified_parm (NULL
, expr
);
1967 if (parm
&& (index
= ipa_get_param_decl_index (info
, parm
)) >= 0)
1968 return add_condition (summary
, index
, NULL
, CHANGED
, NULL_TREE
);
1969 if (is_gimple_min_invariant (expr
))
1970 return false_predicate ();
1971 if (TREE_CODE (expr
) == SSA_NAME
)
1972 return nonconstant_names
[SSA_NAME_VERSION (expr
)];
1973 if (BINARY_CLASS_P (expr
) || COMPARISON_CLASS_P (expr
))
1975 struct predicate p1
= will_be_nonconstant_expr_predicate
1976 (info
, summary
, TREE_OPERAND (expr
, 0),
1978 struct predicate p2
;
1979 if (true_predicate_p (&p1
))
1981 p2
= will_be_nonconstant_expr_predicate (info
, summary
,
1982 TREE_OPERAND (expr
, 1),
1984 return or_predicates (summary
->conds
, &p1
, &p2
);
1986 else if (TREE_CODE (expr
) == COND_EXPR
)
1988 struct predicate p1
= will_be_nonconstant_expr_predicate
1989 (info
, summary
, TREE_OPERAND (expr
, 0),
1991 struct predicate p2
;
1992 if (true_predicate_p (&p1
))
1994 p2
= will_be_nonconstant_expr_predicate (info
, summary
,
1995 TREE_OPERAND (expr
, 1),
1997 if (true_predicate_p (&p2
))
1999 p1
= or_predicates (summary
->conds
, &p1
, &p2
);
2000 p2
= will_be_nonconstant_expr_predicate (info
, summary
,
2001 TREE_OPERAND (expr
, 2),
2003 return or_predicates (summary
->conds
, &p1
, &p2
);
2010 return false_predicate ();
2014 /* Return predicate specifying when the STMT might have result that is not
2015 a compile time constant. */
2017 static struct predicate
2018 will_be_nonconstant_predicate (struct ipa_node_params
*info
,
2019 struct inline_summary
*summary
,
2021 vec
<predicate_t
> nonconstant_names
)
2023 struct predicate p
= true_predicate ();
2026 struct predicate op_non_const
;
2029 struct agg_position_info aggpos
;
2031 /* What statments might be optimized away
2032 when their arguments are constant
2033 TODO: also trivial builtins.
2034 builtin_constant_p is already handled later. */
2035 if (gimple_code (stmt
) != GIMPLE_ASSIGN
2036 && gimple_code (stmt
) != GIMPLE_COND
2037 && gimple_code (stmt
) != GIMPLE_SWITCH
)
2040 /* Stores will stay anyway. */
2041 if (gimple_store_p (stmt
))
2044 is_load
= gimple_assign_load_p (stmt
);
2046 /* Loads can be optimized when the value is known. */
2050 gcc_assert (gimple_assign_single_p (stmt
));
2051 op
= gimple_assign_rhs1 (stmt
);
2052 if (!unmodified_parm_or_parm_agg_item (info
, stmt
, op
, &base_index
,
2059 /* See if we understand all operands before we start
2060 adding conditionals. */
2061 FOR_EACH_SSA_TREE_OPERAND (use
, stmt
, iter
, SSA_OP_USE
)
2063 tree parm
= unmodified_parm (stmt
, use
);
2064 /* For arguments we can build a condition. */
2065 if (parm
&& ipa_get_param_decl_index (info
, parm
) >= 0)
2067 if (TREE_CODE (use
) != SSA_NAME
)
2069 /* If we know when operand is constant,
2070 we still can say something useful. */
2071 if (!true_predicate_p (&nonconstant_names
[SSA_NAME_VERSION (use
)]))
2078 add_condition (summary
, base_index
, &aggpos
, CHANGED
, NULL
);
2080 op_non_const
= false_predicate ();
2081 FOR_EACH_SSA_TREE_OPERAND (use
, stmt
, iter
, SSA_OP_USE
)
2083 tree parm
= unmodified_parm (stmt
, use
);
2086 if (parm
&& (index
= ipa_get_param_decl_index (info
, parm
)) >= 0)
2088 if (index
!= base_index
)
2089 p
= add_condition (summary
, index
, NULL
, CHANGED
, NULL_TREE
);
2094 p
= nonconstant_names
[SSA_NAME_VERSION (use
)];
2095 op_non_const
= or_predicates (summary
->conds
, &p
, &op_non_const
);
2097 if (gimple_code (stmt
) == GIMPLE_ASSIGN
2098 && TREE_CODE (gimple_assign_lhs (stmt
)) == SSA_NAME
)
2099 nonconstant_names
[SSA_NAME_VERSION (gimple_assign_lhs (stmt
))]
2101 return op_non_const
;
2104 struct record_modified_bb_info
2110 /* Callback of walk_aliased_vdefs. Records basic blocks where the value may be
2111 set except for info->stmt. */
2114 record_modified (ao_ref
*ao ATTRIBUTE_UNUSED
, tree vdef
, void *data
)
2116 struct record_modified_bb_info
*info
=
2117 (struct record_modified_bb_info
*) data
;
2118 if (SSA_NAME_DEF_STMT (vdef
) == info
->stmt
)
2120 bitmap_set_bit (info
->bb_set
,
2121 SSA_NAME_IS_DEFAULT_DEF (vdef
)
2122 ? ENTRY_BLOCK_PTR_FOR_FN (cfun
)->index
2123 : gimple_bb (SSA_NAME_DEF_STMT (vdef
))->index
);
2127 /* Return probability (based on REG_BR_PROB_BASE) that I-th parameter of STMT
2128 will change since last invocation of STMT.
2130 Value 0 is reserved for compile time invariants.
2131 For common parameters it is REG_BR_PROB_BASE. For loop invariants it
2132 ought to be REG_BR_PROB_BASE / estimated_iters. */
2135 param_change_prob (gimple stmt
, int i
)
2137 tree op
= gimple_call_arg (stmt
, i
);
2138 basic_block bb
= gimple_bb (stmt
);
2141 /* Global invariants neve change. */
2142 if (is_gimple_min_invariant (op
))
2144 /* We would have to do non-trivial analysis to really work out what
2145 is the probability of value to change (i.e. when init statement
2146 is in a sibling loop of the call).
2148 We do an conservative estimate: when call is executed N times more often
2149 than the statement defining value, we take the frequency 1/N. */
2150 if (TREE_CODE (op
) == SSA_NAME
)
2155 return REG_BR_PROB_BASE
;
2157 if (SSA_NAME_IS_DEFAULT_DEF (op
))
2158 init_freq
= ENTRY_BLOCK_PTR_FOR_FN (cfun
)->frequency
;
2160 init_freq
= gimple_bb (SSA_NAME_DEF_STMT (op
))->frequency
;
2164 if (init_freq
< bb
->frequency
)
2165 return MAX (GCOV_COMPUTE_SCALE (init_freq
, bb
->frequency
), 1);
2167 return REG_BR_PROB_BASE
;
2170 base
= get_base_address (op
);
2175 struct record_modified_bb_info info
;
2178 tree init
= ctor_for_folding (base
);
2180 if (init
!= error_mark_node
)
2183 return REG_BR_PROB_BASE
;
2184 ao_ref_init (&refd
, op
);
2186 info
.bb_set
= BITMAP_ALLOC (NULL
);
2187 walk_aliased_vdefs (&refd
, gimple_vuse (stmt
), record_modified
, &info
,
2189 if (bitmap_bit_p (info
.bb_set
, bb
->index
))
2191 BITMAP_FREE (info
.bb_set
);
2192 return REG_BR_PROB_BASE
;
2195 /* Assume that every memory is initialized at entry.
2196 TODO: Can we easilly determine if value is always defined
2197 and thus we may skip entry block? */
2198 if (ENTRY_BLOCK_PTR_FOR_FN (cfun
)->frequency
)
2199 max
= ENTRY_BLOCK_PTR_FOR_FN (cfun
)->frequency
;
2203 EXECUTE_IF_SET_IN_BITMAP (info
.bb_set
, 0, index
, bi
)
2204 max
= MIN (max
, BASIC_BLOCK_FOR_FN (cfun
, index
)->frequency
);
2206 BITMAP_FREE (info
.bb_set
);
2207 if (max
< bb
->frequency
)
2208 return MAX (GCOV_COMPUTE_SCALE (max
, bb
->frequency
), 1);
2210 return REG_BR_PROB_BASE
;
2212 return REG_BR_PROB_BASE
;
2215 /* Find whether a basic block BB is the final block of a (half) diamond CFG
2216 sub-graph and if the predicate the condition depends on is known. If so,
2217 return true and store the pointer the predicate in *P. */
2220 phi_result_unknown_predicate (struct ipa_node_params
*info
,
2221 struct inline_summary
*summary
, basic_block bb
,
2222 struct predicate
*p
,
2223 vec
<predicate_t
> nonconstant_names
)
2227 basic_block first_bb
= NULL
;
2230 if (single_pred_p (bb
))
2232 *p
= false_predicate ();
2236 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
2238 if (single_succ_p (e
->src
))
2240 if (!single_pred_p (e
->src
))
2243 first_bb
= single_pred (e
->src
);
2244 else if (single_pred (e
->src
) != first_bb
)
2251 else if (e
->src
!= first_bb
)
2259 stmt
= last_stmt (first_bb
);
2261 || gimple_code (stmt
) != GIMPLE_COND
2262 || !is_gimple_ip_invariant (gimple_cond_rhs (stmt
)))
2265 *p
= will_be_nonconstant_expr_predicate (info
, summary
,
2266 gimple_cond_lhs (stmt
),
2268 if (true_predicate_p (p
))
2274 /* Given a PHI statement in a function described by inline properties SUMMARY
2275 and *P being the predicate describing whether the selected PHI argument is
2276 known, store a predicate for the result of the PHI statement into
2277 NONCONSTANT_NAMES, if possible. */
2280 predicate_for_phi_result (struct inline_summary
*summary
, gphi
*phi
,
2281 struct predicate
*p
,
2282 vec
<predicate_t
> nonconstant_names
)
2286 for (i
= 0; i
< gimple_phi_num_args (phi
); i
++)
2288 tree arg
= gimple_phi_arg (phi
, i
)->def
;
2289 if (!is_gimple_min_invariant (arg
))
2291 gcc_assert (TREE_CODE (arg
) == SSA_NAME
);
2292 *p
= or_predicates (summary
->conds
, p
,
2293 &nonconstant_names
[SSA_NAME_VERSION (arg
)]);
2294 if (true_predicate_p (p
))
2299 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2301 fprintf (dump_file
, "\t\tphi predicate: ");
2302 dump_predicate (dump_file
, summary
->conds
, p
);
2304 nonconstant_names
[SSA_NAME_VERSION (gimple_phi_result (phi
))] = *p
;
2307 /* Return predicate specifying when array index in access OP becomes non-constant. */
2309 static struct predicate
2310 array_index_predicate (struct inline_summary
*info
,
2311 vec
< predicate_t
> nonconstant_names
, tree op
)
2313 struct predicate p
= false_predicate ();
2314 while (handled_component_p (op
))
2316 if (TREE_CODE (op
) == ARRAY_REF
|| TREE_CODE (op
) == ARRAY_RANGE_REF
)
2318 if (TREE_CODE (TREE_OPERAND (op
, 1)) == SSA_NAME
)
2319 p
= or_predicates (info
->conds
, &p
,
2320 &nonconstant_names
[SSA_NAME_VERSION
2321 (TREE_OPERAND (op
, 1))]);
2323 op
= TREE_OPERAND (op
, 0);
2328 /* For a typical usage of __builtin_expect (a<b, 1), we
2329 may introduce an extra relation stmt:
2330 With the builtin, we have
2333 t3 = __builtin_expect (t2, 1);
2336 Without the builtin, we have
2339 This affects the size/time estimation and may have
2340 an impact on the earlier inlining.
2341 Here find this pattern and fix it up later. */
2344 find_foldable_builtin_expect (basic_block bb
)
2346 gimple_stmt_iterator bsi
;
2348 for (bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
2350 gimple stmt
= gsi_stmt (bsi
);
2351 if (gimple_call_builtin_p (stmt
, BUILT_IN_EXPECT
)
2352 || (is_gimple_call (stmt
)
2353 && gimple_call_internal_p (stmt
)
2354 && gimple_call_internal_fn (stmt
) == IFN_BUILTIN_EXPECT
))
2356 tree var
= gimple_call_lhs (stmt
);
2357 tree arg
= gimple_call_arg (stmt
, 0);
2358 use_operand_p use_p
;
2365 gcc_assert (TREE_CODE (var
) == SSA_NAME
);
2367 while (TREE_CODE (arg
) == SSA_NAME
)
2369 gimple stmt_tmp
= SSA_NAME_DEF_STMT (arg
);
2370 if (!is_gimple_assign (stmt_tmp
))
2372 switch (gimple_assign_rhs_code (stmt_tmp
))
2391 arg
= gimple_assign_rhs1 (stmt_tmp
);
2394 if (match
&& single_imm_use (var
, &use_p
, &use_stmt
)
2395 && gimple_code (use_stmt
) == GIMPLE_COND
)
2402 /* Return true when the basic blocks contains only clobbers followed by RESX.
2403 Such BBs are kept around to make removal of dead stores possible with
2404 presence of EH and will be optimized out by optimize_clobbers later in the
2407 NEED_EH is used to recurse in case the clobber has non-EH predecestors
2408 that can be clobber only, too.. When it is false, the RESX is not necessary
2409 on the end of basic block. */
2412 clobber_only_eh_bb_p (basic_block bb
, bool need_eh
= true)
2414 gimple_stmt_iterator gsi
= gsi_last_bb (bb
);
2420 if (gsi_end_p (gsi
))
2422 if (gimple_code (gsi_stmt (gsi
)) != GIMPLE_RESX
)
2426 else if (!single_succ_p (bb
))
2429 for (; !gsi_end_p (gsi
); gsi_prev (&gsi
))
2431 gimple stmt
= gsi_stmt (gsi
);
2432 if (is_gimple_debug (stmt
))
2434 if (gimple_clobber_p (stmt
))
2436 if (gimple_code (stmt
) == GIMPLE_LABEL
)
2441 /* See if all predecestors are either throws or clobber only BBs. */
2442 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
2443 if (!(e
->flags
& EDGE_EH
)
2444 && !clobber_only_eh_bb_p (e
->src
, false))
2450 /* Compute function body size parameters for NODE.
2451 When EARLY is true, we compute only simple summaries without
2452 non-trivial predicates to drive the early inliner. */
2455 estimate_function_body_sizes (struct cgraph_node
*node
, bool early
)
2458 /* Estimate static overhead for function prologue/epilogue and alignment. */
2460 /* Benefits are scaled by probability of elimination that is in range
2463 struct function
*my_function
= DECL_STRUCT_FUNCTION (node
->decl
);
2465 struct inline_summary
*info
= inline_summary (node
);
2466 struct predicate bb_predicate
;
2467 struct ipa_node_params
*parms_info
= NULL
;
2468 vec
<predicate_t
> nonconstant_names
= vNULL
;
2471 predicate array_index
= true_predicate ();
2472 gimple fix_builtin_expect_stmt
;
2477 if (optimize
&& !early
)
2479 calculate_dominance_info (CDI_DOMINATORS
);
2480 loop_optimizer_init (LOOPS_NORMAL
| LOOPS_HAVE_RECORDED_EXITS
);
2482 if (ipa_node_params_vector
.exists ())
2484 parms_info
= IPA_NODE_REF (node
);
2485 nonconstant_names
.safe_grow_cleared
2486 (SSANAMES (my_function
)->length ());
2491 fprintf (dump_file
, "\nAnalyzing function body size: %s\n",
2494 /* When we run into maximal number of entries, we assign everything to the
2495 constant truth case. Be sure to have it in list. */
2496 bb_predicate
= true_predicate ();
2497 account_size_time (info
, 0, 0, &bb_predicate
);
2499 bb_predicate
= not_inlined_predicate ();
2500 account_size_time (info
, 2 * INLINE_SIZE_SCALE
, 0, &bb_predicate
);
2502 gcc_assert (my_function
&& my_function
->cfg
);
2504 compute_bb_predicates (node
, parms_info
, info
);
2505 gcc_assert (cfun
== my_function
);
2506 order
= XNEWVEC (int, n_basic_blocks_for_fn (cfun
));
2507 nblocks
= pre_and_rev_post_order_compute (NULL
, order
, false);
2508 for (n
= 0; n
< nblocks
; n
++)
2510 bb
= BASIC_BLOCK_FOR_FN (cfun
, order
[n
]);
2511 freq
= compute_call_stmt_bb_frequency (node
->decl
, bb
);
2512 if (clobber_only_eh_bb_p (bb
))
2514 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2515 fprintf (dump_file
, "\n Ignoring BB %i;"
2516 " it will be optimized away by cleanup_clobbers\n",
2521 /* TODO: Obviously predicates can be propagated down across CFG. */
2525 bb_predicate
= *(struct predicate
*) bb
->aux
;
2527 bb_predicate
= false_predicate ();
2530 bb_predicate
= true_predicate ();
2532 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2534 fprintf (dump_file
, "\n BB %i predicate:", bb
->index
);
2535 dump_predicate (dump_file
, info
->conds
, &bb_predicate
);
2538 if (parms_info
&& nonconstant_names
.exists ())
2540 struct predicate phi_predicate
;
2541 bool first_phi
= true;
2543 for (gphi_iterator bsi
= gsi_start_phis (bb
); !gsi_end_p (bsi
);
2547 && !phi_result_unknown_predicate (parms_info
, info
, bb
,
2552 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2554 fprintf (dump_file
, " ");
2555 print_gimple_stmt (dump_file
, gsi_stmt (bsi
), 0, 0);
2557 predicate_for_phi_result (info
, bsi
.phi (), &phi_predicate
,
2562 fix_builtin_expect_stmt
= find_foldable_builtin_expect (bb
);
2564 for (gimple_stmt_iterator bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
);
2567 gimple stmt
= gsi_stmt (bsi
);
2568 int this_size
= estimate_num_insns (stmt
, &eni_size_weights
);
2569 int this_time
= estimate_num_insns (stmt
, &eni_time_weights
);
2571 struct predicate will_be_nonconstant
;
2573 /* This relation stmt should be folded after we remove
2574 buildin_expect call. Adjust the cost here. */
2575 if (stmt
== fix_builtin_expect_stmt
)
2581 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2583 fprintf (dump_file
, " ");
2584 print_gimple_stmt (dump_file
, stmt
, 0, 0);
2585 fprintf (dump_file
, "\t\tfreq:%3.2f size:%3i time:%3i\n",
2586 ((double) freq
) / CGRAPH_FREQ_BASE
, this_size
,
2590 if (gimple_assign_load_p (stmt
) && nonconstant_names
.exists ())
2592 struct predicate this_array_index
;
2594 array_index_predicate (info
, nonconstant_names
,
2595 gimple_assign_rhs1 (stmt
));
2596 if (!false_predicate_p (&this_array_index
))
2598 and_predicates (info
->conds
, &array_index
,
2601 if (gimple_store_p (stmt
) && nonconstant_names
.exists ())
2603 struct predicate this_array_index
;
2605 array_index_predicate (info
, nonconstant_names
,
2606 gimple_get_lhs (stmt
));
2607 if (!false_predicate_p (&this_array_index
))
2609 and_predicates (info
->conds
, &array_index
,
2614 if (is_gimple_call (stmt
)
2615 && !gimple_call_internal_p (stmt
))
2617 struct cgraph_edge
*edge
= node
->get_edge (stmt
);
2618 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
2620 /* Special case: results of BUILT_IN_CONSTANT_P will be always
2621 resolved as constant. We however don't want to optimize
2622 out the cgraph edges. */
2623 if (nonconstant_names
.exists ()
2624 && gimple_call_builtin_p (stmt
, BUILT_IN_CONSTANT_P
)
2625 && gimple_call_lhs (stmt
)
2626 && TREE_CODE (gimple_call_lhs (stmt
)) == SSA_NAME
)
2628 struct predicate false_p
= false_predicate ();
2629 nonconstant_names
[SSA_NAME_VERSION (gimple_call_lhs (stmt
))]
2632 if (ipa_node_params_vector
.exists ())
2634 int count
= gimple_call_num_args (stmt
);
2638 es
->param
.safe_grow_cleared (count
);
2639 for (i
= 0; i
< count
; i
++)
2641 int prob
= param_change_prob (stmt
, i
);
2642 gcc_assert (prob
>= 0 && prob
<= REG_BR_PROB_BASE
);
2643 es
->param
[i
].change_prob
= prob
;
2647 es
->call_stmt_size
= this_size
;
2648 es
->call_stmt_time
= this_time
;
2649 es
->loop_depth
= bb_loop_depth (bb
);
2650 edge_set_predicate (edge
, &bb_predicate
);
2653 /* TODO: When conditional jump or swithc is known to be constant, but
2654 we did not translate it into the predicates, we really can account
2655 just maximum of the possible paths. */
2658 = will_be_nonconstant_predicate (parms_info
, info
,
2659 stmt
, nonconstant_names
);
2660 if (this_time
|| this_size
)
2666 prob
= eliminated_by_inlining_prob (stmt
);
2667 if (prob
== 1 && dump_file
&& (dump_flags
& TDF_DETAILS
))
2669 "\t\t50%% will be eliminated by inlining\n");
2670 if (prob
== 2 && dump_file
&& (dump_flags
& TDF_DETAILS
))
2671 fprintf (dump_file
, "\t\tWill be eliminated by inlining\n");
2674 p
= and_predicates (info
->conds
, &bb_predicate
,
2675 &will_be_nonconstant
);
2677 p
= true_predicate ();
2679 if (!false_predicate_p (&p
))
2683 if (time
> MAX_TIME
* INLINE_TIME_SCALE
)
2684 time
= MAX_TIME
* INLINE_TIME_SCALE
;
2687 /* We account everything but the calls. Calls have their own
2688 size/time info attached to cgraph edges. This is necessary
2689 in order to make the cost disappear after inlining. */
2690 if (!is_gimple_call (stmt
))
2694 struct predicate ip
= not_inlined_predicate ();
2695 ip
= and_predicates (info
->conds
, &ip
, &p
);
2696 account_size_time (info
, this_size
* prob
,
2697 this_time
* prob
, &ip
);
2700 account_size_time (info
, this_size
* (2 - prob
),
2701 this_time
* (2 - prob
), &p
);
2704 gcc_assert (time
>= 0);
2705 gcc_assert (size
>= 0);
2709 set_hint_predicate (&inline_summary (node
)->array_index
, array_index
);
2710 time
= (time
+ CGRAPH_FREQ_BASE
/ 2) / CGRAPH_FREQ_BASE
;
2711 if (time
> MAX_TIME
)
2715 if (!early
&& nonconstant_names
.exists ())
2718 predicate loop_iterations
= true_predicate ();
2719 predicate loop_stride
= true_predicate ();
2721 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2722 flow_loops_dump (dump_file
, NULL
, 0);
2724 FOR_EACH_LOOP (loop
, 0)
2729 struct tree_niter_desc niter_desc
;
2730 basic_block
*body
= get_loop_body (loop
);
2731 bb_predicate
= *(struct predicate
*) loop
->header
->aux
;
2733 exits
= get_loop_exit_edges (loop
);
2734 FOR_EACH_VEC_ELT (exits
, j
, ex
)
2735 if (number_of_iterations_exit (loop
, ex
, &niter_desc
, false)
2736 && !is_gimple_min_invariant (niter_desc
.niter
))
2738 predicate will_be_nonconstant
2739 = will_be_nonconstant_expr_predicate (parms_info
, info
,
2742 if (!true_predicate_p (&will_be_nonconstant
))
2743 will_be_nonconstant
= and_predicates (info
->conds
,
2745 &will_be_nonconstant
);
2746 if (!true_predicate_p (&will_be_nonconstant
)
2747 && !false_predicate_p (&will_be_nonconstant
))
2748 /* This is slightly inprecise. We may want to represent each
2749 loop with independent predicate. */
2751 and_predicates (info
->conds
, &loop_iterations
,
2752 &will_be_nonconstant
);
2756 for (i
= 0; i
< loop
->num_nodes
; i
++)
2758 gimple_stmt_iterator gsi
;
2759 bb_predicate
= *(struct predicate
*) body
[i
]->aux
;
2760 for (gsi
= gsi_start_bb (body
[i
]); !gsi_end_p (gsi
);
2763 gimple stmt
= gsi_stmt (gsi
);
2768 FOR_EACH_SSA_TREE_OPERAND (use
, stmt
, iter
, SSA_OP_USE
)
2770 predicate will_be_nonconstant
;
2773 (loop
, loop_containing_stmt (stmt
), use
, &iv
, true)
2774 || is_gimple_min_invariant (iv
.step
))
2777 = will_be_nonconstant_expr_predicate (parms_info
, info
,
2780 if (!true_predicate_p (&will_be_nonconstant
))
2782 = and_predicates (info
->conds
,
2784 &will_be_nonconstant
);
2785 if (!true_predicate_p (&will_be_nonconstant
)
2786 && !false_predicate_p (&will_be_nonconstant
))
2787 /* This is slightly inprecise. We may want to represent
2788 each loop with independent predicate. */
2790 and_predicates (info
->conds
, &loop_stride
,
2791 &will_be_nonconstant
);
2797 set_hint_predicate (&inline_summary (node
)->loop_iterations
,
2799 set_hint_predicate (&inline_summary (node
)->loop_stride
, loop_stride
);
2802 FOR_ALL_BB_FN (bb
, my_function
)
2808 pool_free (edge_predicate_pool
, bb
->aux
);
2810 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
2813 pool_free (edge_predicate_pool
, e
->aux
);
2817 inline_summary (node
)->self_time
= time
;
2818 inline_summary (node
)->self_size
= size
;
2819 nonconstant_names
.release ();
2820 if (optimize
&& !early
)
2822 loop_optimizer_finalize ();
2823 free_dominance_info (CDI_DOMINATORS
);
2827 fprintf (dump_file
, "\n");
2828 dump_inline_summary (dump_file
, node
);
2833 /* Compute parameters of functions used by inliner.
2834 EARLY is true when we compute parameters for the early inliner */
2837 compute_inline_parameters (struct cgraph_node
*node
, bool early
)
2839 HOST_WIDE_INT self_stack_size
;
2840 struct cgraph_edge
*e
;
2841 struct inline_summary
*info
;
2843 gcc_assert (!node
->global
.inlined_to
);
2845 inline_summary_alloc ();
2847 info
= inline_summary (node
);
2848 reset_inline_summary (node
);
2850 /* FIXME: Thunks are inlinable, but tree-inline don't know how to do that.
2851 Once this happen, we will need to more curefully predict call
2853 if (node
->thunk
.thunk_p
)
2855 struct inline_edge_summary
*es
= inline_edge_summary (node
->callees
);
2856 struct predicate t
= true_predicate ();
2858 info
->inlinable
= 0;
2859 node
->callees
->call_stmt_cannot_inline_p
= true;
2860 node
->local
.can_change_signature
= false;
2861 es
->call_stmt_time
= 1;
2862 es
->call_stmt_size
= 1;
2863 account_size_time (info
, 0, 0, &t
);
2867 /* Even is_gimple_min_invariant rely on current_function_decl. */
2868 push_cfun (DECL_STRUCT_FUNCTION (node
->decl
));
2870 /* Estimate the stack size for the function if we're optimizing. */
2871 self_stack_size
= optimize
? estimated_stack_frame_size (node
) : 0;
2872 info
->estimated_self_stack_size
= self_stack_size
;
2873 info
->estimated_stack_size
= self_stack_size
;
2874 info
->stack_frame_offset
= 0;
2876 /* Can this function be inlined at all? */
2877 if (!optimize
&& !lookup_attribute ("always_inline",
2878 DECL_ATTRIBUTES (node
->decl
)))
2879 info
->inlinable
= false;
2881 info
->inlinable
= tree_inlinable_function_p (node
->decl
);
2883 /* Type attributes can use parameter indices to describe them. */
2884 if (TYPE_ATTRIBUTES (TREE_TYPE (node
->decl
)))
2885 node
->local
.can_change_signature
= false;
2888 /* Otherwise, inlinable functions always can change signature. */
2889 if (info
->inlinable
)
2890 node
->local
.can_change_signature
= true;
2893 /* Functions calling builtin_apply can not change signature. */
2894 for (e
= node
->callees
; e
; e
= e
->next_callee
)
2896 tree
cdecl = e
->callee
->decl
;
2897 if (DECL_BUILT_IN (cdecl)
2898 && DECL_BUILT_IN_CLASS (cdecl) == BUILT_IN_NORMAL
2899 && (DECL_FUNCTION_CODE (cdecl) == BUILT_IN_APPLY_ARGS
2900 || DECL_FUNCTION_CODE (cdecl) == BUILT_IN_VA_START
))
2903 node
->local
.can_change_signature
= !e
;
2906 estimate_function_body_sizes (node
, early
);
2908 for (e
= node
->callees
; e
; e
= e
->next_callee
)
2909 if (e
->callee
->comdat_local_p ())
2911 node
->calls_comdat_local
= (e
!= NULL
);
2913 /* Inlining characteristics are maintained by the cgraph_mark_inline. */
2914 info
->time
= info
->self_time
;
2915 info
->size
= info
->self_size
;
2916 info
->stack_frame_offset
= 0;
2917 info
->estimated_stack_size
= info
->estimated_self_stack_size
;
2918 #ifdef ENABLE_CHECKING
2919 inline_update_overall_summary (node
);
2920 gcc_assert (info
->time
== info
->self_time
&& info
->size
== info
->self_size
);
2927 /* Compute parameters of functions used by inliner using
2928 current_function_decl. */
2931 compute_inline_parameters_for_current (void)
2933 compute_inline_parameters (cgraph_node::get (current_function_decl
), true);
2939 const pass_data pass_data_inline_parameters
=
2941 GIMPLE_PASS
, /* type */
2942 "inline_param", /* name */
2943 OPTGROUP_INLINE
, /* optinfo_flags */
2944 TV_INLINE_PARAMETERS
, /* tv_id */
2945 0, /* properties_required */
2946 0, /* properties_provided */
2947 0, /* properties_destroyed */
2948 0, /* todo_flags_start */
2949 0, /* todo_flags_finish */
2952 class pass_inline_parameters
: public gimple_opt_pass
2955 pass_inline_parameters (gcc::context
*ctxt
)
2956 : gimple_opt_pass (pass_data_inline_parameters
, ctxt
)
2959 /* opt_pass methods: */
2960 opt_pass
* clone () { return new pass_inline_parameters (m_ctxt
); }
2961 virtual unsigned int execute (function
*)
2963 return compute_inline_parameters_for_current ();
2966 }; // class pass_inline_parameters
2971 make_pass_inline_parameters (gcc::context
*ctxt
)
2973 return new pass_inline_parameters (ctxt
);
2977 /* Estimate benefit devirtualizing indirect edge IE, provided KNOWN_VALS,
2978 KNOWN_CONTEXTS and KNOWN_AGGS. */
2981 estimate_edge_devirt_benefit (struct cgraph_edge
*ie
,
2982 int *size
, int *time
,
2983 vec
<tree
> known_vals
,
2984 vec
<ipa_polymorphic_call_context
> known_contexts
,
2985 vec
<ipa_agg_jump_function_p
> known_aggs
)
2988 struct cgraph_node
*callee
;
2989 struct inline_summary
*isummary
;
2990 enum availability avail
;
2993 if (!known_vals
.exists () && !known_contexts
.exists ())
2995 if (!flag_indirect_inlining
)
2998 target
= ipa_get_indirect_edge_target (ie
, known_vals
, known_contexts
,
2999 known_aggs
, &speculative
);
3000 if (!target
|| speculative
)
3003 /* Account for difference in cost between indirect and direct calls. */
3004 *size
-= (eni_size_weights
.indirect_call_cost
- eni_size_weights
.call_cost
);
3005 *time
-= (eni_time_weights
.indirect_call_cost
- eni_time_weights
.call_cost
);
3006 gcc_checking_assert (*time
>= 0);
3007 gcc_checking_assert (*size
>= 0);
3009 callee
= cgraph_node::get (target
);
3010 if (!callee
|| !callee
->definition
)
3012 callee
= callee
->function_symbol (&avail
);
3013 if (avail
< AVAIL_AVAILABLE
)
3015 isummary
= inline_summary (callee
);
3016 return isummary
->inlinable
;
3019 /* Increase SIZE, MIN_SIZE (if non-NULL) and TIME for size and time needed to
3020 handle edge E with probability PROB.
3021 Set HINTS if edge may be devirtualized.
3022 KNOWN_VALS, KNOWN_AGGS and KNOWN_CONTEXTS describe context of the call
3026 estimate_edge_size_and_time (struct cgraph_edge
*e
, int *size
, int *min_size
,
3029 vec
<tree
> known_vals
,
3030 vec
<ipa_polymorphic_call_context
> known_contexts
,
3031 vec
<ipa_agg_jump_function_p
> known_aggs
,
3032 inline_hints
*hints
)
3034 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3035 int call_size
= es
->call_stmt_size
;
3036 int call_time
= es
->call_stmt_time
;
3039 && estimate_edge_devirt_benefit (e
, &call_size
, &call_time
,
3040 known_vals
, known_contexts
, known_aggs
)
3041 && hints
&& e
->maybe_hot_p ())
3042 *hints
|= INLINE_HINT_indirect_call
;
3043 cur_size
= call_size
* INLINE_SIZE_SCALE
;
3046 *min_size
+= cur_size
;
3047 *time
+= apply_probability ((gcov_type
) call_time
, prob
)
3048 * e
->frequency
* (INLINE_TIME_SCALE
/ CGRAPH_FREQ_BASE
);
3049 if (*time
> MAX_TIME
* INLINE_TIME_SCALE
)
3050 *time
= MAX_TIME
* INLINE_TIME_SCALE
;
3055 /* Increase SIZE, MIN_SIZE and TIME for size and time needed to handle all
3056 calls in NODE. POSSIBLE_TRUTHS, KNOWN_VALS, KNOWN_AGGS and KNOWN_CONTEXTS
3057 describe context of the call site. */
3060 estimate_calls_size_and_time (struct cgraph_node
*node
, int *size
,
3061 int *min_size
, int *time
,
3062 inline_hints
*hints
,
3063 clause_t possible_truths
,
3064 vec
<tree
> known_vals
,
3065 vec
<ipa_polymorphic_call_context
> known_contexts
,
3066 vec
<ipa_agg_jump_function_p
> known_aggs
)
3068 struct cgraph_edge
*e
;
3069 for (e
= node
->callees
; e
; e
= e
->next_callee
)
3071 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3073 || evaluate_predicate (es
->predicate
, possible_truths
))
3075 if (e
->inline_failed
)
3077 /* Predicates of calls shall not use NOT_CHANGED codes,
3078 sowe do not need to compute probabilities. */
3079 estimate_edge_size_and_time (e
, size
,
3080 es
->predicate
? NULL
: min_size
,
3081 time
, REG_BR_PROB_BASE
,
3082 known_vals
, known_contexts
,
3086 estimate_calls_size_and_time (e
->callee
, size
, min_size
, time
,
3089 known_vals
, known_contexts
,
3093 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
3095 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3097 || evaluate_predicate (es
->predicate
, possible_truths
))
3098 estimate_edge_size_and_time (e
, size
,
3099 es
->predicate
? NULL
: min_size
,
3100 time
, REG_BR_PROB_BASE
,
3101 known_vals
, known_contexts
, known_aggs
,
3107 /* Estimate size and time needed to execute NODE assuming
3108 POSSIBLE_TRUTHS clause, and KNOWN_VALS, KNOWN_AGGS and KNOWN_CONTEXTS
3109 information about NODE's arguments. If non-NULL use also probability
3110 information present in INLINE_PARAM_SUMMARY vector.
3111 Additionally detemine hints determined by the context. Finally compute
3112 minimal size needed for the call that is independent on the call context and
3113 can be used for fast estimates. Return the values in RET_SIZE,
3114 RET_MIN_SIZE, RET_TIME and RET_HINTS. */
3117 estimate_node_size_and_time (struct cgraph_node
*node
,
3118 clause_t possible_truths
,
3119 vec
<tree
> known_vals
,
3120 vec
<ipa_polymorphic_call_context
> known_contexts
,
3121 vec
<ipa_agg_jump_function_p
> known_aggs
,
3122 int *ret_size
, int *ret_min_size
, int *ret_time
,
3123 inline_hints
*ret_hints
,
3124 vec
<inline_param_summary
>
3125 inline_param_summary
)
3127 struct inline_summary
*info
= inline_summary (node
);
3132 inline_hints hints
= 0;
3135 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3138 fprintf (dump_file
, " Estimating body: %s/%i\n"
3139 " Known to be false: ", node
->name (),
3142 for (i
= predicate_not_inlined_condition
;
3143 i
< (predicate_first_dynamic_condition
3144 + (int) vec_safe_length (info
->conds
)); i
++)
3145 if (!(possible_truths
& (1 << i
)))
3148 fprintf (dump_file
, ", ");
3150 dump_condition (dump_file
, info
->conds
, i
);
3154 for (i
= 0; vec_safe_iterate (info
->entry
, i
, &e
); i
++)
3155 if (evaluate_predicate (&e
->predicate
, possible_truths
))
3158 gcc_checking_assert (e
->time
>= 0);
3159 gcc_checking_assert (time
>= 0);
3160 if (!inline_param_summary
.exists ())
3164 int prob
= predicate_probability (info
->conds
,
3167 inline_param_summary
);
3168 gcc_checking_assert (prob
>= 0);
3169 gcc_checking_assert (prob
<= REG_BR_PROB_BASE
);
3170 time
+= apply_probability ((gcov_type
) e
->time
, prob
);
3172 if (time
> MAX_TIME
* INLINE_TIME_SCALE
)
3173 time
= MAX_TIME
* INLINE_TIME_SCALE
;
3174 gcc_checking_assert (time
>= 0);
3177 gcc_checking_assert (true_predicate_p (&(*info
->entry
)[0].predicate
));
3178 min_size
= (*info
->entry
)[0].size
;
3179 gcc_checking_assert (size
>= 0);
3180 gcc_checking_assert (time
>= 0);
3182 if (info
->loop_iterations
3183 && !evaluate_predicate (info
->loop_iterations
, possible_truths
))
3184 hints
|= INLINE_HINT_loop_iterations
;
3185 if (info
->loop_stride
3186 && !evaluate_predicate (info
->loop_stride
, possible_truths
))
3187 hints
|= INLINE_HINT_loop_stride
;
3188 if (info
->array_index
3189 && !evaluate_predicate (info
->array_index
, possible_truths
))
3190 hints
|= INLINE_HINT_array_index
;
3192 hints
|= INLINE_HINT_in_scc
;
3193 if (DECL_DECLARED_INLINE_P (node
->decl
))
3194 hints
|= INLINE_HINT_declared_inline
;
3196 estimate_calls_size_and_time (node
, &size
, &min_size
, &time
, &hints
, possible_truths
,
3197 known_vals
, known_contexts
, known_aggs
);
3198 gcc_checking_assert (size
>= 0);
3199 gcc_checking_assert (time
>= 0);
3200 time
= RDIV (time
, INLINE_TIME_SCALE
);
3201 size
= RDIV (size
, INLINE_SIZE_SCALE
);
3202 min_size
= RDIV (min_size
, INLINE_SIZE_SCALE
);
3204 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3205 fprintf (dump_file
, "\n size:%i time:%i\n", (int) size
, (int) time
);
3211 *ret_min_size
= min_size
;
3218 /* Estimate size and time needed to execute callee of EDGE assuming that
3219 parameters known to be constant at caller of EDGE are propagated.
3220 KNOWN_VALS and KNOWN_CONTEXTS are vectors of assumed known constant values
3221 and types for parameters. */
3224 estimate_ipcp_clone_size_and_time (struct cgraph_node
*node
,
3225 vec
<tree
> known_vals
,
3226 vec
<ipa_polymorphic_call_context
>
3228 vec
<ipa_agg_jump_function_p
> known_aggs
,
3229 int *ret_size
, int *ret_time
,
3230 inline_hints
*hints
)
3234 clause
= evaluate_conditions_for_known_args (node
, false, known_vals
,
3236 estimate_node_size_and_time (node
, clause
, known_vals
, known_contexts
,
3237 known_aggs
, ret_size
, NULL
, ret_time
, hints
, vNULL
);
3240 /* Translate all conditions from callee representation into caller
3241 representation and symbolically evaluate predicate P into new predicate.
3243 INFO is inline_summary of function we are adding predicate into, CALLEE_INFO
3244 is summary of function predicate P is from. OPERAND_MAP is array giving
3245 callee formal IDs the caller formal IDs. POSSSIBLE_TRUTHS is clausule of all
3246 callee conditions that may be true in caller context. TOPLEV_PREDICATE is
3247 predicate under which callee is executed. OFFSET_MAP is an array of of
3248 offsets that need to be added to conditions, negative offset means that
3249 conditions relying on values passed by reference have to be discarded
3250 because they might not be preserved (and should be considered offset zero
3251 for other purposes). */
3253 static struct predicate
3254 remap_predicate (struct inline_summary
*info
,
3255 struct inline_summary
*callee_info
,
3256 struct predicate
*p
,
3257 vec
<int> operand_map
,
3258 vec
<int> offset_map
,
3259 clause_t possible_truths
, struct predicate
*toplev_predicate
)
3262 struct predicate out
= true_predicate ();
3264 /* True predicate is easy. */
3265 if (true_predicate_p (p
))
3266 return *toplev_predicate
;
3267 for (i
= 0; p
->clause
[i
]; i
++)
3269 clause_t clause
= p
->clause
[i
];
3271 struct predicate clause_predicate
= false_predicate ();
3273 gcc_assert (i
< MAX_CLAUSES
);
3275 for (cond
= 0; cond
< NUM_CONDITIONS
; cond
++)
3276 /* Do we have condition we can't disprove? */
3277 if (clause
& possible_truths
& (1 << cond
))
3279 struct predicate cond_predicate
;
3280 /* Work out if the condition can translate to predicate in the
3281 inlined function. */
3282 if (cond
>= predicate_first_dynamic_condition
)
3284 struct condition
*c
;
3286 c
= &(*callee_info
->conds
)[cond
3288 predicate_first_dynamic_condition
];
3289 /* See if we can remap condition operand to caller's operand.
3290 Otherwise give up. */
3291 if (!operand_map
.exists ()
3292 || (int) operand_map
.length () <= c
->operand_num
3293 || operand_map
[c
->operand_num
] == -1
3294 /* TODO: For non-aggregate conditions, adding an offset is
3295 basically an arithmetic jump function processing which
3296 we should support in future. */
3297 || ((!c
->agg_contents
|| !c
->by_ref
)
3298 && offset_map
[c
->operand_num
] > 0)
3299 || (c
->agg_contents
&& c
->by_ref
3300 && offset_map
[c
->operand_num
] < 0))
3301 cond_predicate
= true_predicate ();
3304 struct agg_position_info ap
;
3305 HOST_WIDE_INT offset_delta
= offset_map
[c
->operand_num
];
3306 if (offset_delta
< 0)
3308 gcc_checking_assert (!c
->agg_contents
|| !c
->by_ref
);
3311 gcc_assert (!c
->agg_contents
3312 || c
->by_ref
|| offset_delta
== 0);
3313 ap
.offset
= c
->offset
+ offset_delta
;
3314 ap
.agg_contents
= c
->agg_contents
;
3315 ap
.by_ref
= c
->by_ref
;
3316 cond_predicate
= add_condition (info
,
3317 operand_map
[c
->operand_num
],
3318 &ap
, c
->code
, c
->val
);
3321 /* Fixed conditions remains same, construct single
3322 condition predicate. */
3325 cond_predicate
.clause
[0] = 1 << cond
;
3326 cond_predicate
.clause
[1] = 0;
3328 clause_predicate
= or_predicates (info
->conds
, &clause_predicate
,
3331 out
= and_predicates (info
->conds
, &out
, &clause_predicate
);
3333 return and_predicates (info
->conds
, &out
, toplev_predicate
);
3337 /* Update summary information of inline clones after inlining.
3338 Compute peak stack usage. */
3341 inline_update_callee_summaries (struct cgraph_node
*node
, int depth
)
3343 struct cgraph_edge
*e
;
3344 struct inline_summary
*callee_info
= inline_summary (node
);
3345 struct inline_summary
*caller_info
= inline_summary (node
->callers
->caller
);
3348 callee_info
->stack_frame_offset
3349 = caller_info
->stack_frame_offset
3350 + caller_info
->estimated_self_stack_size
;
3351 peak
= callee_info
->stack_frame_offset
3352 + callee_info
->estimated_self_stack_size
;
3353 if (inline_summary (node
->global
.inlined_to
)->estimated_stack_size
< peak
)
3354 inline_summary (node
->global
.inlined_to
)->estimated_stack_size
= peak
;
3355 ipa_propagate_frequency (node
);
3356 for (e
= node
->callees
; e
; e
= e
->next_callee
)
3358 if (!e
->inline_failed
)
3359 inline_update_callee_summaries (e
->callee
, depth
);
3360 inline_edge_summary (e
)->loop_depth
+= depth
;
3362 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
3363 inline_edge_summary (e
)->loop_depth
+= depth
;
3366 /* Update change_prob of EDGE after INLINED_EDGE has been inlined.
3367 When functoin A is inlined in B and A calls C with parameter that
3368 changes with probability PROB1 and C is known to be passthroug
3369 of argument if B that change with probability PROB2, the probability
3370 of change is now PROB1*PROB2. */
3373 remap_edge_change_prob (struct cgraph_edge
*inlined_edge
,
3374 struct cgraph_edge
*edge
)
3376 if (ipa_node_params_vector
.exists ())
3379 struct ipa_edge_args
*args
= IPA_EDGE_REF (edge
);
3380 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3381 struct inline_edge_summary
*inlined_es
3382 = inline_edge_summary (inlined_edge
);
3384 for (i
= 0; i
< ipa_get_cs_argument_count (args
); i
++)
3386 struct ipa_jump_func
*jfunc
= ipa_get_ith_jump_func (args
, i
);
3387 if (jfunc
->type
== IPA_JF_PASS_THROUGH
3388 && (ipa_get_jf_pass_through_formal_id (jfunc
)
3389 < (int) inlined_es
->param
.length ()))
3391 int jf_formal_id
= ipa_get_jf_pass_through_formal_id (jfunc
);
3392 int prob1
= es
->param
[i
].change_prob
;
3393 int prob2
= inlined_es
->param
[jf_formal_id
].change_prob
;
3394 int prob
= combine_probabilities (prob1
, prob2
);
3396 if (prob1
&& prob2
&& !prob
)
3399 es
->param
[i
].change_prob
= prob
;
3405 /* Update edge summaries of NODE after INLINED_EDGE has been inlined.
3407 Remap predicates of callees of NODE. Rest of arguments match
3410 Also update change probabilities. */
3413 remap_edge_summaries (struct cgraph_edge
*inlined_edge
,
3414 struct cgraph_node
*node
,
3415 struct inline_summary
*info
,
3416 struct inline_summary
*callee_info
,
3417 vec
<int> operand_map
,
3418 vec
<int> offset_map
,
3419 clause_t possible_truths
,
3420 struct predicate
*toplev_predicate
)
3422 struct cgraph_edge
*e
;
3423 for (e
= node
->callees
; e
; e
= e
->next_callee
)
3425 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3428 if (e
->inline_failed
)
3430 remap_edge_change_prob (inlined_edge
, e
);
3434 p
= remap_predicate (info
, callee_info
,
3435 es
->predicate
, operand_map
, offset_map
,
3436 possible_truths
, toplev_predicate
);
3437 edge_set_predicate (e
, &p
);
3438 /* TODO: We should remove the edge for code that will be
3439 optimized out, but we need to keep verifiers and tree-inline
3440 happy. Make it cold for now. */
3441 if (false_predicate_p (&p
))
3448 edge_set_predicate (e
, toplev_predicate
);
3451 remap_edge_summaries (inlined_edge
, e
->callee
, info
, callee_info
,
3452 operand_map
, offset_map
, possible_truths
,
3455 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
3457 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3460 remap_edge_change_prob (inlined_edge
, e
);
3463 p
= remap_predicate (info
, callee_info
,
3464 es
->predicate
, operand_map
, offset_map
,
3465 possible_truths
, toplev_predicate
);
3466 edge_set_predicate (e
, &p
);
3467 /* TODO: We should remove the edge for code that will be optimized
3468 out, but we need to keep verifiers and tree-inline happy.
3469 Make it cold for now. */
3470 if (false_predicate_p (&p
))
3477 edge_set_predicate (e
, toplev_predicate
);
3481 /* Same as remap_predicate, but set result into hint *HINT. */
3484 remap_hint_predicate (struct inline_summary
*info
,
3485 struct inline_summary
*callee_info
,
3486 struct predicate
**hint
,
3487 vec
<int> operand_map
,
3488 vec
<int> offset_map
,
3489 clause_t possible_truths
,
3490 struct predicate
*toplev_predicate
)
3496 p
= remap_predicate (info
, callee_info
,
3498 operand_map
, offset_map
,
3499 possible_truths
, toplev_predicate
);
3500 if (!false_predicate_p (&p
) && !true_predicate_p (&p
))
3503 set_hint_predicate (hint
, p
);
3505 **hint
= and_predicates (info
->conds
, *hint
, &p
);
3509 /* We inlined EDGE. Update summary of the function we inlined into. */
3512 inline_merge_summary (struct cgraph_edge
*edge
)
3514 struct inline_summary
*callee_info
= inline_summary (edge
->callee
);
3515 struct cgraph_node
*to
= (edge
->caller
->global
.inlined_to
3516 ? edge
->caller
->global
.inlined_to
: edge
->caller
);
3517 struct inline_summary
*info
= inline_summary (to
);
3518 clause_t clause
= 0; /* not_inline is known to be false. */
3520 vec
<int> operand_map
= vNULL
;
3521 vec
<int> offset_map
= vNULL
;
3523 struct predicate toplev_predicate
;
3524 struct predicate true_p
= true_predicate ();
3525 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3528 toplev_predicate
= *es
->predicate
;
3530 toplev_predicate
= true_predicate ();
3532 if (ipa_node_params_vector
.exists () && callee_info
->conds
)
3534 struct ipa_edge_args
*args
= IPA_EDGE_REF (edge
);
3535 int count
= ipa_get_cs_argument_count (args
);
3538 evaluate_properties_for_edge (edge
, true, &clause
, NULL
, NULL
, NULL
);
3541 operand_map
.safe_grow_cleared (count
);
3542 offset_map
.safe_grow_cleared (count
);
3544 for (i
= 0; i
< count
; i
++)
3546 struct ipa_jump_func
*jfunc
= ipa_get_ith_jump_func (args
, i
);
3549 /* TODO: handle non-NOPs when merging. */
3550 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
3552 if (ipa_get_jf_pass_through_operation (jfunc
) == NOP_EXPR
)
3553 map
= ipa_get_jf_pass_through_formal_id (jfunc
);
3554 if (!ipa_get_jf_pass_through_agg_preserved (jfunc
))
3557 else if (jfunc
->type
== IPA_JF_ANCESTOR
)
3559 HOST_WIDE_INT offset
= ipa_get_jf_ancestor_offset (jfunc
);
3560 if (offset
>= 0 && offset
< INT_MAX
)
3562 map
= ipa_get_jf_ancestor_formal_id (jfunc
);
3563 if (!ipa_get_jf_ancestor_agg_preserved (jfunc
))
3565 offset_map
[i
] = offset
;
3568 operand_map
[i
] = map
;
3569 gcc_assert (map
< ipa_get_param_count (IPA_NODE_REF (to
)));
3572 for (i
= 0; vec_safe_iterate (callee_info
->entry
, i
, &e
); i
++)
3574 struct predicate p
= remap_predicate (info
, callee_info
,
3575 &e
->predicate
, operand_map
,
3578 if (!false_predicate_p (&p
))
3580 gcov_type add_time
= ((gcov_type
) e
->time
* edge
->frequency
3581 + CGRAPH_FREQ_BASE
/ 2) / CGRAPH_FREQ_BASE
;
3582 int prob
= predicate_probability (callee_info
->conds
,
3585 add_time
= apply_probability ((gcov_type
) add_time
, prob
);
3586 if (add_time
> MAX_TIME
* INLINE_TIME_SCALE
)
3587 add_time
= MAX_TIME
* INLINE_TIME_SCALE
;
3588 if (prob
!= REG_BR_PROB_BASE
3589 && dump_file
&& (dump_flags
& TDF_DETAILS
))
3591 fprintf (dump_file
, "\t\tScaling time by probability:%f\n",
3592 (double) prob
/ REG_BR_PROB_BASE
);
3594 account_size_time (info
, e
->size
, add_time
, &p
);
3597 remap_edge_summaries (edge
, edge
->callee
, info
, callee_info
, operand_map
,
3598 offset_map
, clause
, &toplev_predicate
);
3599 remap_hint_predicate (info
, callee_info
,
3600 &callee_info
->loop_iterations
,
3601 operand_map
, offset_map
, clause
, &toplev_predicate
);
3602 remap_hint_predicate (info
, callee_info
,
3603 &callee_info
->loop_stride
,
3604 operand_map
, offset_map
, clause
, &toplev_predicate
);
3605 remap_hint_predicate (info
, callee_info
,
3606 &callee_info
->array_index
,
3607 operand_map
, offset_map
, clause
, &toplev_predicate
);
3609 inline_update_callee_summaries (edge
->callee
,
3610 inline_edge_summary (edge
)->loop_depth
);
3612 /* We do not maintain predicates of inlined edges, free it. */
3613 edge_set_predicate (edge
, &true_p
);
3614 /* Similarly remove param summaries. */
3615 es
->param
.release ();
3616 operand_map
.release ();
3617 offset_map
.release ();
3620 /* For performance reasons inline_merge_summary is not updating overall size
3621 and time. Recompute it. */
3624 inline_update_overall_summary (struct cgraph_node
*node
)
3626 struct inline_summary
*info
= inline_summary (node
);
3632 for (i
= 0; vec_safe_iterate (info
->entry
, i
, &e
); i
++)
3634 info
->size
+= e
->size
, info
->time
+= e
->time
;
3635 if (info
->time
> MAX_TIME
* INLINE_TIME_SCALE
)
3636 info
->time
= MAX_TIME
* INLINE_TIME_SCALE
;
3638 estimate_calls_size_and_time (node
, &info
->size
, &info
->min_size
,
3640 ~(clause_t
) (1 << predicate_false_condition
),
3641 vNULL
, vNULL
, vNULL
);
3642 info
->time
= (info
->time
+ INLINE_TIME_SCALE
/ 2) / INLINE_TIME_SCALE
;
3643 info
->size
= (info
->size
+ INLINE_SIZE_SCALE
/ 2) / INLINE_SIZE_SCALE
;
3646 /* Return hints derrived from EDGE. */
3648 simple_edge_hints (struct cgraph_edge
*edge
)
3651 struct cgraph_node
*to
= (edge
->caller
->global
.inlined_to
3652 ? edge
->caller
->global
.inlined_to
: edge
->caller
);
3653 if (inline_summary (to
)->scc_no
3654 && inline_summary (to
)->scc_no
== inline_summary (edge
->callee
)->scc_no
3655 && !edge
->recursive_p ())
3656 hints
|= INLINE_HINT_same_scc
;
3658 if (to
->lto_file_data
&& edge
->callee
->lto_file_data
3659 && to
->lto_file_data
!= edge
->callee
->lto_file_data
)
3660 hints
|= INLINE_HINT_cross_module
;
3665 /* Estimate the time cost for the caller when inlining EDGE.
3666 Only to be called via estimate_edge_time, that handles the
3669 When caching, also update the cache entry. Compute both time and
3670 size, since we always need both metrics eventually. */
3673 do_estimate_edge_time (struct cgraph_edge
*edge
)
3678 struct cgraph_node
*callee
;
3680 vec
<tree
> known_vals
;
3681 vec
<ipa_polymorphic_call_context
> known_contexts
;
3682 vec
<ipa_agg_jump_function_p
> known_aggs
;
3683 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3686 callee
= edge
->callee
->ultimate_alias_target ();
3688 gcc_checking_assert (edge
->inline_failed
);
3689 evaluate_properties_for_edge (edge
, true,
3690 &clause
, &known_vals
, &known_contexts
,
3692 estimate_node_size_and_time (callee
, clause
, known_vals
, known_contexts
,
3693 known_aggs
, &size
, &min_size
, &time
, &hints
, es
->param
);
3695 /* When we have profile feedback, we can quite safely identify hot
3696 edges and for those we disable size limits. Don't do that when
3697 probability that caller will call the callee is low however, since it
3698 may hurt optimization of the caller's hot path. */
3699 if (edge
->count
&& edge
->maybe_hot_p ()
3701 > (edge
->caller
->global
.inlined_to
3702 ? edge
->caller
->global
.inlined_to
->count
: edge
->caller
->count
)))
3703 hints
|= INLINE_HINT_known_hot
;
3705 known_vals
.release ();
3706 known_contexts
.release ();
3707 known_aggs
.release ();
3708 gcc_checking_assert (size
>= 0);
3709 gcc_checking_assert (time
>= 0);
3711 /* When caching, update the cache entry. */
3712 if (edge_growth_cache
.exists ())
3714 inline_summary (edge
->callee
)->min_size
= min_size
;
3715 if ((int) edge_growth_cache
.length () <= edge
->uid
)
3716 edge_growth_cache
.safe_grow_cleared (symtab
->edges_max_uid
);
3717 edge_growth_cache
[edge
->uid
].time
= time
+ (time
>= 0);
3719 edge_growth_cache
[edge
->uid
].size
= size
+ (size
>= 0);
3720 hints
|= simple_edge_hints (edge
);
3721 edge_growth_cache
[edge
->uid
].hints
= hints
+ 1;
3727 /* Return estimated callee growth after inlining EDGE.
3728 Only to be called via estimate_edge_size. */
3731 do_estimate_edge_size (struct cgraph_edge
*edge
)
3734 struct cgraph_node
*callee
;
3736 vec
<tree
> known_vals
;
3737 vec
<ipa_polymorphic_call_context
> known_contexts
;
3738 vec
<ipa_agg_jump_function_p
> known_aggs
;
3740 /* When we do caching, use do_estimate_edge_time to populate the entry. */
3742 if (edge_growth_cache
.exists ())
3744 do_estimate_edge_time (edge
);
3745 size
= edge_growth_cache
[edge
->uid
].size
;
3746 gcc_checking_assert (size
);
3747 return size
- (size
> 0);
3750 callee
= edge
->callee
->ultimate_alias_target ();
3752 /* Early inliner runs without caching, go ahead and do the dirty work. */
3753 gcc_checking_assert (edge
->inline_failed
);
3754 evaluate_properties_for_edge (edge
, true,
3755 &clause
, &known_vals
, &known_contexts
,
3757 estimate_node_size_and_time (callee
, clause
, known_vals
, known_contexts
,
3758 known_aggs
, &size
, NULL
, NULL
, NULL
, vNULL
);
3759 known_vals
.release ();
3760 known_contexts
.release ();
3761 known_aggs
.release ();
3766 /* Estimate the growth of the caller when inlining EDGE.
3767 Only to be called via estimate_edge_size. */
3770 do_estimate_edge_hints (struct cgraph_edge
*edge
)
3773 struct cgraph_node
*callee
;
3775 vec
<tree
> known_vals
;
3776 vec
<ipa_polymorphic_call_context
> known_contexts
;
3777 vec
<ipa_agg_jump_function_p
> known_aggs
;
3779 /* When we do caching, use do_estimate_edge_time to populate the entry. */
3781 if (edge_growth_cache
.exists ())
3783 do_estimate_edge_time (edge
);
3784 hints
= edge_growth_cache
[edge
->uid
].hints
;
3785 gcc_checking_assert (hints
);
3789 callee
= edge
->callee
->ultimate_alias_target ();
3791 /* Early inliner runs without caching, go ahead and do the dirty work. */
3792 gcc_checking_assert (edge
->inline_failed
);
3793 evaluate_properties_for_edge (edge
, true,
3794 &clause
, &known_vals
, &known_contexts
,
3796 estimate_node_size_and_time (callee
, clause
, known_vals
, known_contexts
,
3797 known_aggs
, NULL
, NULL
, NULL
, &hints
, vNULL
);
3798 known_vals
.release ();
3799 known_contexts
.release ();
3800 known_aggs
.release ();
3801 hints
|= simple_edge_hints (edge
);
3806 /* Estimate self time of the function NODE after inlining EDGE. */
3809 estimate_time_after_inlining (struct cgraph_node
*node
,
3810 struct cgraph_edge
*edge
)
3812 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3813 if (!es
->predicate
|| !false_predicate_p (es
->predicate
))
3816 inline_summary (node
)->time
+ estimate_edge_time (edge
);
3819 if (time
> MAX_TIME
)
3823 return inline_summary (node
)->time
;
3827 /* Estimate the size of NODE after inlining EDGE which should be an
3828 edge to either NODE or a call inlined into NODE. */
3831 estimate_size_after_inlining (struct cgraph_node
*node
,
3832 struct cgraph_edge
*edge
)
3834 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3835 if (!es
->predicate
|| !false_predicate_p (es
->predicate
))
3837 int size
= inline_summary (node
)->size
+ estimate_edge_growth (edge
);
3838 gcc_assert (size
>= 0);
3841 return inline_summary (node
)->size
;
3847 struct cgraph_node
*node
;
3848 bool self_recursive
;
3853 /* Worker for do_estimate_growth. Collect growth for all callers. */
3856 do_estimate_growth_1 (struct cgraph_node
*node
, void *data
)
3858 struct cgraph_edge
*e
;
3859 struct growth_data
*d
= (struct growth_data
*) data
;
3861 for (e
= node
->callers
; e
; e
= e
->next_caller
)
3863 gcc_checking_assert (e
->inline_failed
);
3865 if (e
->caller
== d
->node
3866 || (e
->caller
->global
.inlined_to
3867 && e
->caller
->global
.inlined_to
== d
->node
))
3868 d
->self_recursive
= true;
3869 d
->growth
+= estimate_edge_growth (e
);
3875 /* Estimate the growth caused by inlining NODE into all callees. */
3878 do_estimate_growth (struct cgraph_node
*node
)
3880 struct growth_data d
= { node
, 0, false };
3881 struct inline_summary
*info
= inline_summary (node
);
3883 node
->call_for_symbol_thunks_and_aliases (do_estimate_growth_1
, &d
, true);
3885 /* For self recursive functions the growth estimation really should be
3886 infinity. We don't want to return very large values because the growth
3887 plays various roles in badness computation fractions. Be sure to not
3888 return zero or negative growths. */
3889 if (d
.self_recursive
)
3890 d
.growth
= d
.growth
< info
->size
? info
->size
: d
.growth
;
3891 else if (DECL_EXTERNAL (node
->decl
))
3895 if (node
->will_be_removed_from_program_if_no_direct_calls_p ())
3896 d
.growth
-= info
->size
;
3897 /* COMDAT functions are very often not shared across multiple units
3898 since they come from various template instantiations.
3899 Take this into account. */
3900 else if (DECL_COMDAT (node
->decl
)
3901 && node
->can_remove_if_no_direct_calls_p ())
3902 d
.growth
-= (info
->size
3903 * (100 - PARAM_VALUE (PARAM_COMDAT_SHARING_PROBABILITY
))
3907 if (node_growth_cache
.exists ())
3909 if ((int) node_growth_cache
.length () <= node
->uid
)
3910 node_growth_cache
.safe_grow_cleared (symtab
->cgraph_max_uid
);
3911 node_growth_cache
[node
->uid
] = d
.growth
+ (d
.growth
>= 0);
3917 /* Make cheap estimation if growth of NODE is likely positive knowing
3918 EDGE_GROWTH of one particular edge.
3919 We assume that most of other edges will have similar growth
3920 and skip computation if there are too many callers. */
3923 growth_likely_positive (struct cgraph_node
*node
, int edge_growth ATTRIBUTE_UNUSED
)
3927 struct cgraph_edge
*e
;
3928 gcc_checking_assert (edge_growth
> 0);
3930 /* Unlike for functions called once, we play unsafe with
3931 COMDATs. We can allow that since we know functions
3932 in consideration are small (and thus risk is small) and
3933 moreover grow estimates already accounts that COMDAT
3934 functions may or may not disappear when eliminated from
3935 current unit. With good probability making aggressive
3936 choice in all units is going to make overall program
3939 Consequently we ask cgraph_can_remove_if_no_direct_calls_p
3941 cgraph_will_be_removed_from_program_if_no_direct_calls */
3942 if (DECL_EXTERNAL (node
->decl
)
3943 || !node
->can_remove_if_no_direct_calls_p ())
3946 /* If there is cached value, just go ahead. */
3947 if ((int)node_growth_cache
.length () > node
->uid
3948 && (ret
= node_growth_cache
[node
->uid
]))
3950 if (!node
->will_be_removed_from_program_if_no_direct_calls_p ()
3951 && (!DECL_COMDAT (node
->decl
)
3952 || !node
->can_remove_if_no_direct_calls_p ()))
3954 max_callers
= inline_summary (node
)->size
* 4 / edge_growth
+ 2;
3956 for (e
= node
->callers
; e
; e
= e
->next_caller
)
3962 return estimate_growth (node
) > 0;
3966 /* This function performs intraprocedural analysis in NODE that is required to
3967 inline indirect calls. */
3970 inline_indirect_intraprocedural_analysis (struct cgraph_node
*node
)
3972 ipa_analyze_node (node
);
3973 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3975 ipa_print_node_params (dump_file
, node
);
3976 ipa_print_node_jump_functions (dump_file
, node
);
3981 /* Note function body size. */
3984 inline_analyze_function (struct cgraph_node
*node
)
3986 push_cfun (DECL_STRUCT_FUNCTION (node
->decl
));
3989 fprintf (dump_file
, "\nAnalyzing function: %s/%u\n",
3990 node
->name (), node
->order
);
3991 if (optimize
&& !node
->thunk
.thunk_p
)
3992 inline_indirect_intraprocedural_analysis (node
);
3993 compute_inline_parameters (node
, false);
3996 struct cgraph_edge
*e
;
3997 for (e
= node
->callees
; e
; e
= e
->next_callee
)
3999 if (e
->inline_failed
== CIF_FUNCTION_NOT_CONSIDERED
)
4000 e
->inline_failed
= CIF_FUNCTION_NOT_OPTIMIZED
;
4001 e
->call_stmt_cannot_inline_p
= true;
4003 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
4005 if (e
->inline_failed
== CIF_FUNCTION_NOT_CONSIDERED
)
4006 e
->inline_failed
= CIF_FUNCTION_NOT_OPTIMIZED
;
4007 e
->call_stmt_cannot_inline_p
= true;
4015 /* Called when new function is inserted to callgraph late. */
4018 add_new_function (struct cgraph_node
*node
, void *data ATTRIBUTE_UNUSED
)
4020 inline_analyze_function (node
);
4024 /* Note function body size. */
4027 inline_generate_summary (void)
4029 struct cgraph_node
*node
;
4031 /* When not optimizing, do not bother to analyze. Inlining is still done
4032 because edge redirection needs to happen there. */
4033 if (!optimize
&& !flag_generate_lto
&& !flag_wpa
)
4036 function_insertion_hook_holder
=
4037 symtab
->add_cgraph_insertion_hook (&add_new_function
, NULL
);
4039 ipa_register_cgraph_hooks ();
4040 inline_free_summary ();
4042 FOR_EACH_DEFINED_FUNCTION (node
)
4044 inline_analyze_function (node
);
4048 /* Read predicate from IB. */
4050 static struct predicate
4051 read_predicate (struct lto_input_block
*ib
)
4053 struct predicate out
;
4059 gcc_assert (k
<= MAX_CLAUSES
);
4060 clause
= out
.clause
[k
++] = streamer_read_uhwi (ib
);
4064 /* Zero-initialize the remaining clauses in OUT. */
4065 while (k
<= MAX_CLAUSES
)
4066 out
.clause
[k
++] = 0;
4072 /* Write inline summary for edge E to OB. */
4075 read_inline_edge_summary (struct lto_input_block
*ib
, struct cgraph_edge
*e
)
4077 struct inline_edge_summary
*es
= inline_edge_summary (e
);
4081 es
->call_stmt_size
= streamer_read_uhwi (ib
);
4082 es
->call_stmt_time
= streamer_read_uhwi (ib
);
4083 es
->loop_depth
= streamer_read_uhwi (ib
);
4084 p
= read_predicate (ib
);
4085 edge_set_predicate (e
, &p
);
4086 length
= streamer_read_uhwi (ib
);
4089 es
->param
.safe_grow_cleared (length
);
4090 for (i
= 0; i
< length
; i
++)
4091 es
->param
[i
].change_prob
= streamer_read_uhwi (ib
);
4096 /* Stream in inline summaries from the section. */
4099 inline_read_section (struct lto_file_decl_data
*file_data
, const char *data
,
4102 const struct lto_function_header
*header
=
4103 (const struct lto_function_header
*) data
;
4104 const int cfg_offset
= sizeof (struct lto_function_header
);
4105 const int main_offset
= cfg_offset
+ header
->cfg_size
;
4106 const int string_offset
= main_offset
+ header
->main_size
;
4107 struct data_in
*data_in
;
4108 unsigned int i
, count2
, j
;
4109 unsigned int f_count
;
4111 lto_input_block
ib ((const char *) data
+ main_offset
, header
->main_size
);
4114 lto_data_in_create (file_data
, (const char *) data
+ string_offset
,
4115 header
->string_size
, vNULL
);
4116 f_count
= streamer_read_uhwi (&ib
);
4117 for (i
= 0; i
< f_count
; i
++)
4120 struct cgraph_node
*node
;
4121 struct inline_summary
*info
;
4122 lto_symtab_encoder_t encoder
;
4123 struct bitpack_d bp
;
4124 struct cgraph_edge
*e
;
4127 index
= streamer_read_uhwi (&ib
);
4128 encoder
= file_data
->symtab_node_encoder
;
4129 node
= dyn_cast
<cgraph_node
*> (lto_symtab_encoder_deref (encoder
,
4131 info
= inline_summary (node
);
4133 info
->estimated_stack_size
4134 = info
->estimated_self_stack_size
= streamer_read_uhwi (&ib
);
4135 info
->size
= info
->self_size
= streamer_read_uhwi (&ib
);
4136 info
->time
= info
->self_time
= streamer_read_uhwi (&ib
);
4138 bp
= streamer_read_bitpack (&ib
);
4139 info
->inlinable
= bp_unpack_value (&bp
, 1);
4141 count2
= streamer_read_uhwi (&ib
);
4142 gcc_assert (!info
->conds
);
4143 for (j
= 0; j
< count2
; j
++)
4146 c
.operand_num
= streamer_read_uhwi (&ib
);
4147 c
.code
= (enum tree_code
) streamer_read_uhwi (&ib
);
4148 c
.val
= stream_read_tree (&ib
, data_in
);
4149 bp
= streamer_read_bitpack (&ib
);
4150 c
.agg_contents
= bp_unpack_value (&bp
, 1);
4151 c
.by_ref
= bp_unpack_value (&bp
, 1);
4153 c
.offset
= streamer_read_uhwi (&ib
);
4154 vec_safe_push (info
->conds
, c
);
4156 count2
= streamer_read_uhwi (&ib
);
4157 gcc_assert (!info
->entry
);
4158 for (j
= 0; j
< count2
; j
++)
4160 struct size_time_entry e
;
4162 e
.size
= streamer_read_uhwi (&ib
);
4163 e
.time
= streamer_read_uhwi (&ib
);
4164 e
.predicate
= read_predicate (&ib
);
4166 vec_safe_push (info
->entry
, e
);
4169 p
= read_predicate (&ib
);
4170 set_hint_predicate (&info
->loop_iterations
, p
);
4171 p
= read_predicate (&ib
);
4172 set_hint_predicate (&info
->loop_stride
, p
);
4173 p
= read_predicate (&ib
);
4174 set_hint_predicate (&info
->array_index
, p
);
4175 for (e
= node
->callees
; e
; e
= e
->next_callee
)
4176 read_inline_edge_summary (&ib
, e
);
4177 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
4178 read_inline_edge_summary (&ib
, e
);
4181 lto_free_section_data (file_data
, LTO_section_inline_summary
, NULL
, data
,
4183 lto_data_in_delete (data_in
);
4187 /* Read inline summary. Jump functions are shared among ipa-cp
4188 and inliner, so when ipa-cp is active, we don't need to write them
4192 inline_read_summary (void)
4194 struct lto_file_decl_data
**file_data_vec
= lto_get_file_decl_data ();
4195 struct lto_file_decl_data
*file_data
;
4198 inline_summary_alloc ();
4200 while ((file_data
= file_data_vec
[j
++]))
4203 const char *data
= lto_get_section_data (file_data
,
4204 LTO_section_inline_summary
,
4207 inline_read_section (file_data
, data
, len
);
4209 /* Fatal error here. We do not want to support compiling ltrans units
4210 with different version of compiler or different flags than the WPA
4211 unit, so this should never happen. */
4212 fatal_error ("ipa inline summary is missing in input file");
4216 ipa_register_cgraph_hooks ();
4218 ipa_prop_read_jump_functions ();
4220 function_insertion_hook_holder
=
4221 symtab
->add_cgraph_insertion_hook (&add_new_function
, NULL
);
4225 /* Write predicate P to OB. */
4228 write_predicate (struct output_block
*ob
, struct predicate
*p
)
4232 for (j
= 0; p
->clause
[j
]; j
++)
4234 gcc_assert (j
< MAX_CLAUSES
);
4235 streamer_write_uhwi (ob
, p
->clause
[j
]);
4237 streamer_write_uhwi (ob
, 0);
4241 /* Write inline summary for edge E to OB. */
4244 write_inline_edge_summary (struct output_block
*ob
, struct cgraph_edge
*e
)
4246 struct inline_edge_summary
*es
= inline_edge_summary (e
);
4249 streamer_write_uhwi (ob
, es
->call_stmt_size
);
4250 streamer_write_uhwi (ob
, es
->call_stmt_time
);
4251 streamer_write_uhwi (ob
, es
->loop_depth
);
4252 write_predicate (ob
, es
->predicate
);
4253 streamer_write_uhwi (ob
, es
->param
.length ());
4254 for (i
= 0; i
< (int) es
->param
.length (); i
++)
4255 streamer_write_uhwi (ob
, es
->param
[i
].change_prob
);
4259 /* Write inline summary for node in SET.
4260 Jump functions are shared among ipa-cp and inliner, so when ipa-cp is
4261 active, we don't need to write them twice. */
4264 inline_write_summary (void)
4266 struct cgraph_node
*node
;
4267 struct output_block
*ob
= create_output_block (LTO_section_inline_summary
);
4268 lto_symtab_encoder_t encoder
= ob
->decl_state
->symtab_node_encoder
;
4269 unsigned int count
= 0;
4272 for (i
= 0; i
< lto_symtab_encoder_size (encoder
); i
++)
4274 symtab_node
*snode
= lto_symtab_encoder_deref (encoder
, i
);
4275 cgraph_node
*cnode
= dyn_cast
<cgraph_node
*> (snode
);
4276 if (cnode
&& cnode
->definition
&& !cnode
->alias
)
4279 streamer_write_uhwi (ob
, count
);
4281 for (i
= 0; i
< lto_symtab_encoder_size (encoder
); i
++)
4283 symtab_node
*snode
= lto_symtab_encoder_deref (encoder
, i
);
4284 cgraph_node
*cnode
= dyn_cast
<cgraph_node
*> (snode
);
4285 if (cnode
&& (node
= cnode
)->definition
&& !node
->alias
)
4287 struct inline_summary
*info
= inline_summary (node
);
4288 struct bitpack_d bp
;
4289 struct cgraph_edge
*edge
;
4292 struct condition
*c
;
4294 streamer_write_uhwi (ob
,
4295 lto_symtab_encoder_encode (encoder
,
4298 streamer_write_hwi (ob
, info
->estimated_self_stack_size
);
4299 streamer_write_hwi (ob
, info
->self_size
);
4300 streamer_write_hwi (ob
, info
->self_time
);
4301 bp
= bitpack_create (ob
->main_stream
);
4302 bp_pack_value (&bp
, info
->inlinable
, 1);
4303 streamer_write_bitpack (&bp
);
4304 streamer_write_uhwi (ob
, vec_safe_length (info
->conds
));
4305 for (i
= 0; vec_safe_iterate (info
->conds
, i
, &c
); i
++)
4307 streamer_write_uhwi (ob
, c
->operand_num
);
4308 streamer_write_uhwi (ob
, c
->code
);
4309 stream_write_tree (ob
, c
->val
, true);
4310 bp
= bitpack_create (ob
->main_stream
);
4311 bp_pack_value (&bp
, c
->agg_contents
, 1);
4312 bp_pack_value (&bp
, c
->by_ref
, 1);
4313 streamer_write_bitpack (&bp
);
4314 if (c
->agg_contents
)
4315 streamer_write_uhwi (ob
, c
->offset
);
4317 streamer_write_uhwi (ob
, vec_safe_length (info
->entry
));
4318 for (i
= 0; vec_safe_iterate (info
->entry
, i
, &e
); i
++)
4320 streamer_write_uhwi (ob
, e
->size
);
4321 streamer_write_uhwi (ob
, e
->time
);
4322 write_predicate (ob
, &e
->predicate
);
4324 write_predicate (ob
, info
->loop_iterations
);
4325 write_predicate (ob
, info
->loop_stride
);
4326 write_predicate (ob
, info
->array_index
);
4327 for (edge
= node
->callees
; edge
; edge
= edge
->next_callee
)
4328 write_inline_edge_summary (ob
, edge
);
4329 for (edge
= node
->indirect_calls
; edge
; edge
= edge
->next_callee
)
4330 write_inline_edge_summary (ob
, edge
);
4333 streamer_write_char_stream (ob
->main_stream
, 0);
4334 produce_asm (ob
, NULL
);
4335 destroy_output_block (ob
);
4337 if (optimize
&& !flag_ipa_cp
)
4338 ipa_prop_write_jump_functions ();
4342 /* Release inline summary. */
4345 inline_free_summary (void)
4347 struct cgraph_node
*node
;
4348 if (function_insertion_hook_holder
)
4349 symtab
->remove_cgraph_insertion_hook (function_insertion_hook_holder
);
4350 function_insertion_hook_holder
= NULL
;
4351 if (node_removal_hook_holder
)
4352 symtab
->remove_cgraph_removal_hook (node_removal_hook_holder
);
4353 node_removal_hook_holder
= NULL
;
4354 if (edge_removal_hook_holder
)
4355 symtab
->remove_edge_removal_hook (edge_removal_hook_holder
);
4356 edge_removal_hook_holder
= NULL
;
4357 if (node_duplication_hook_holder
)
4358 symtab
->remove_cgraph_duplication_hook (node_duplication_hook_holder
);
4359 node_duplication_hook_holder
= NULL
;
4360 if (edge_duplication_hook_holder
)
4361 symtab
->remove_edge_duplication_hook (edge_duplication_hook_holder
);
4362 edge_duplication_hook_holder
= NULL
;
4363 if (!inline_edge_summary_vec
.exists ())
4365 FOR_EACH_DEFINED_FUNCTION (node
)
4367 reset_inline_summary (node
);
4368 vec_free (inline_summary_vec
);
4369 inline_edge_summary_vec
.release ();
4370 if (edge_predicate_pool
)
4371 free_alloc_pool (edge_predicate_pool
);
4372 edge_predicate_pool
= 0;