1 /* Inlining decision heuristics.
2 Copyright (C) 2003-2015 Free Software Foundation, Inc.
3 Contributed by Jan Hubicka
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
21 /* Analysis used by the inliner and other passes limiting code size growth.
23 We estimate for each function
25 - average function execution time
26 - inlining size benefit (that is how much of function body size
27 and its call sequence is expected to disappear by inlining)
28 - inlining time benefit
31 - call statement size and time
33 inlinie_summary datastructures store above information locally (i.e.
34 parameters of the function itself) and globally (i.e. parameters of
35 the function created by applying all the inline decisions already
36 present in the callgraph).
38 We provide accestor to the inline_summary datastructure and
39 basic logic updating the parameters when inlining is performed.
41 The summaries are context sensitive. Context means
42 1) partial assignment of known constant values of operands
43 2) whether function is inlined into the call or not.
44 It is easy to add more variants. To represent function size and time
45 that depends on context (i.e. it is known to be optimized away when
46 context is known either by inlining or from IP-CP and clonning),
47 we use predicates. Predicates are logical formulas in
48 conjunctive-disjunctive form consisting of clauses. Clauses are bitmaps
49 specifying what conditions must be true. Conditions are simple test
50 of the form described above.
52 In order to make predicate (possibly) true, all of its clauses must
53 be (possibly) true. To make clause (possibly) true, one of conditions
54 it mentions must be (possibly) true. There are fixed bounds on
55 number of clauses and conditions and all the manipulation functions
56 are conservative in positive direction. I.e. we may lose precision
57 by thinking that predicate may be true even when it is not.
59 estimate_edge_size and estimate_edge_growth can be used to query
60 function size/time in the given context. inline_merge_summary merges
61 properties of caller and callee after inlining.
63 Finally pass_inline_parameters is exported. This is used to drive
64 computation of function parameters used by the early inliner. IPA
65 inlined performs analysis via its analyze_function method. */
69 #include "coretypes.h"
73 #include "hard-reg-set.h"
76 #include "fold-const.h"
77 #include "stor-layout.h"
78 #include "print-tree.h"
79 #include "tree-inline.h"
80 #include "langhooks.h"
82 #include "diagnostic.h"
83 #include "gimple-pretty-print.h"
85 #include "tree-pass.h"
88 #include "internal-fn.h"
89 #include "gimple-iterator.h"
91 #include "tree-ssa-loop-niter.h"
92 #include "tree-ssa-loop.h"
94 #include "alloc-pool.h"
95 #include "symbol-summary.h"
97 #include "tree-streamer.h"
98 #include "ipa-inline.h"
100 #include "tree-scalar-evolution.h"
101 #include "ipa-utils.h"
103 #include "cfgexpand.h"
105 /* Estimate runtime of function can easilly run into huge numbers with many
106 nested loops. Be sure we can compute time * INLINE_SIZE_SCALE * 2 in an
107 integer. For anything larger we use gcov_type. */
108 #define MAX_TIME 500000
110 /* Number of bits in integer, but we really want to be stable across different
112 #define NUM_CONDITIONS 32
114 enum predicate_conditions
116 predicate_false_condition
= 0,
117 predicate_not_inlined_condition
= 1,
118 predicate_first_dynamic_condition
= 2
121 /* Special condition code we use to represent test that operand is compile time
123 #define IS_NOT_CONSTANT ERROR_MARK
124 /* Special condition code we use to represent test that operand is not changed
125 across invocation of the function. When operand IS_NOT_CONSTANT it is always
126 CHANGED, however i.e. loop invariants can be NOT_CHANGED given percentage
127 of executions even when they are not compile time constants. */
128 #define CHANGED IDENTIFIER_NODE
130 /* Holders of ipa cgraph hooks: */
131 static struct cgraph_2edge_hook_list
*edge_duplication_hook_holder
;
132 static struct cgraph_edge_hook_list
*edge_removal_hook_holder
;
133 static void inline_edge_removal_hook (struct cgraph_edge
*, void *);
134 static void inline_edge_duplication_hook (struct cgraph_edge
*,
135 struct cgraph_edge
*, void *);
137 /* VECtor holding inline summaries.
138 In GGC memory because conditions might point to constant trees. */
139 function_summary
<inline_summary
*> *inline_summaries
;
140 vec
<inline_edge_summary_t
> inline_edge_summary_vec
;
142 /* Cached node/edge growths. */
143 vec
<edge_growth_cache_entry
> edge_growth_cache
;
145 /* Edge predicates goes here. */
146 static object_allocator
<predicate
> edge_predicate_pool ("edge predicates");
148 /* Return true predicate (tautology).
149 We represent it by empty list of clauses. */
151 static inline struct predicate
152 true_predicate (void)
160 /* Return predicate testing single condition number COND. */
162 static inline struct predicate
163 single_cond_predicate (int cond
)
166 p
.clause
[0] = 1 << cond
;
172 /* Return false predicate. First clause require false condition. */
174 static inline struct predicate
175 false_predicate (void)
177 return single_cond_predicate (predicate_false_condition
);
181 /* Return true if P is (true). */
184 true_predicate_p (struct predicate
*p
)
186 return !p
->clause
[0];
190 /* Return true if P is (false). */
193 false_predicate_p (struct predicate
*p
)
195 if (p
->clause
[0] == (1 << predicate_false_condition
))
197 gcc_checking_assert (!p
->clause
[1]
198 && p
->clause
[0] == 1 << predicate_false_condition
);
205 /* Return predicate that is set true when function is not inlined. */
207 static inline struct predicate
208 not_inlined_predicate (void)
210 return single_cond_predicate (predicate_not_inlined_condition
);
213 /* Simple description of whether a memory load or a condition refers to a load
214 from an aggregate and if so, how and where from in the aggregate.
215 Individual fields have the same meaning like fields with the same name in
218 struct agg_position_info
220 HOST_WIDE_INT offset
;
225 /* Add condition to condition list CONDS. AGGPOS describes whether the used
226 oprand is loaded from an aggregate and where in the aggregate it is. It can
227 be NULL, which means this not a load from an aggregate. */
229 static struct predicate
230 add_condition (struct inline_summary
*summary
, int operand_num
,
231 struct agg_position_info
*aggpos
,
232 enum tree_code code
, tree val
)
236 struct condition new_cond
;
237 HOST_WIDE_INT offset
;
238 bool agg_contents
, by_ref
;
242 offset
= aggpos
->offset
;
243 agg_contents
= aggpos
->agg_contents
;
244 by_ref
= aggpos
->by_ref
;
249 agg_contents
= false;
253 gcc_checking_assert (operand_num
>= 0);
254 for (i
= 0; vec_safe_iterate (summary
->conds
, i
, &c
); i
++)
256 if (c
->operand_num
== operand_num
259 && c
->agg_contents
== agg_contents
260 && (!agg_contents
|| (c
->offset
== offset
&& c
->by_ref
== by_ref
)))
261 return single_cond_predicate (i
+ predicate_first_dynamic_condition
);
263 /* Too many conditions. Give up and return constant true. */
264 if (i
== NUM_CONDITIONS
- predicate_first_dynamic_condition
)
265 return true_predicate ();
267 new_cond
.operand_num
= operand_num
;
268 new_cond
.code
= code
;
270 new_cond
.agg_contents
= agg_contents
;
271 new_cond
.by_ref
= by_ref
;
272 new_cond
.offset
= offset
;
273 vec_safe_push (summary
->conds
, new_cond
);
274 return single_cond_predicate (i
+ predicate_first_dynamic_condition
);
278 /* Add clause CLAUSE into the predicate P. */
281 add_clause (conditions conditions
, struct predicate
*p
, clause_t clause
)
285 int insert_here
= -1;
292 /* False clause makes the whole predicate false. Kill the other variants. */
293 if (clause
== (1 << predicate_false_condition
))
295 p
->clause
[0] = (1 << predicate_false_condition
);
299 if (false_predicate_p (p
))
302 /* No one should be silly enough to add false into nontrivial clauses. */
303 gcc_checking_assert (!(clause
& (1 << predicate_false_condition
)));
305 /* Look where to insert the clause. At the same time prune out
306 clauses of P that are implied by the new clause and thus
308 for (i
= 0, i2
= 0; i
<= MAX_CLAUSES
; i
++)
310 p
->clause
[i2
] = p
->clause
[i
];
315 /* If p->clause[i] implies clause, there is nothing to add. */
316 if ((p
->clause
[i
] & clause
) == p
->clause
[i
])
318 /* We had nothing to add, none of clauses should've become
320 gcc_checking_assert (i
== i2
);
324 if (p
->clause
[i
] < clause
&& insert_here
< 0)
327 /* If clause implies p->clause[i], then p->clause[i] becomes redundant.
328 Otherwise the p->clause[i] has to stay. */
329 if ((p
->clause
[i
] & clause
) != clause
)
333 /* Look for clauses that are obviously true. I.e.
334 op0 == 5 || op0 != 5. */
335 for (c1
= predicate_first_dynamic_condition
; c1
< NUM_CONDITIONS
; c1
++)
338 if (!(clause
& (1 << c1
)))
340 cc1
= &(*conditions
)[c1
- predicate_first_dynamic_condition
];
341 /* We have no way to represent !CHANGED and !IS_NOT_CONSTANT
342 and thus there is no point for looking for them. */
343 if (cc1
->code
== CHANGED
|| cc1
->code
== IS_NOT_CONSTANT
)
345 for (c2
= c1
+ 1; c2
< NUM_CONDITIONS
; c2
++)
346 if (clause
& (1 << c2
))
349 &(*conditions
)[c1
- predicate_first_dynamic_condition
];
351 &(*conditions
)[c2
- predicate_first_dynamic_condition
];
352 if (cc1
->operand_num
== cc2
->operand_num
353 && cc1
->val
== cc2
->val
354 && cc2
->code
!= IS_NOT_CONSTANT
355 && cc2
->code
!= CHANGED
356 && cc1
->code
== invert_tree_comparison (cc2
->code
,
357 HONOR_NANS (cc1
->val
)))
363 /* We run out of variants. Be conservative in positive direction. */
364 if (i2
== MAX_CLAUSES
)
366 /* Keep clauses in decreasing order. This makes equivalence testing easy. */
367 p
->clause
[i2
+ 1] = 0;
368 if (insert_here
>= 0)
369 for (; i2
> insert_here
; i2
--)
370 p
->clause
[i2
] = p
->clause
[i2
- 1];
373 p
->clause
[insert_here
] = clause
;
379 static struct predicate
380 and_predicates (conditions conditions
,
381 struct predicate
*p
, struct predicate
*p2
)
383 struct predicate out
= *p
;
386 /* Avoid busy work. */
387 if (false_predicate_p (p2
) || true_predicate_p (p
))
389 if (false_predicate_p (p
) || true_predicate_p (p2
))
392 /* See how far predicates match. */
393 for (i
= 0; p
->clause
[i
] && p
->clause
[i
] == p2
->clause
[i
]; i
++)
395 gcc_checking_assert (i
< MAX_CLAUSES
);
398 /* Combine the predicates rest. */
399 for (; p2
->clause
[i
]; i
++)
401 gcc_checking_assert (i
< MAX_CLAUSES
);
402 add_clause (conditions
, &out
, p2
->clause
[i
]);
408 /* Return true if predicates are obviously equal. */
411 predicates_equal_p (struct predicate
*p
, struct predicate
*p2
)
414 for (i
= 0; p
->clause
[i
]; i
++)
416 gcc_checking_assert (i
< MAX_CLAUSES
);
417 gcc_checking_assert (p
->clause
[i
] > p
->clause
[i
+ 1]);
418 gcc_checking_assert (!p2
->clause
[i
]
419 || p2
->clause
[i
] > p2
->clause
[i
+ 1]);
420 if (p
->clause
[i
] != p2
->clause
[i
])
423 return !p2
->clause
[i
];
429 static struct predicate
430 or_predicates (conditions conditions
,
431 struct predicate
*p
, struct predicate
*p2
)
433 struct predicate out
= true_predicate ();
436 /* Avoid busy work. */
437 if (false_predicate_p (p2
) || true_predicate_p (p
))
439 if (false_predicate_p (p
) || true_predicate_p (p2
))
441 if (predicates_equal_p (p
, p2
))
444 /* OK, combine the predicates. */
445 for (i
= 0; p
->clause
[i
]; i
++)
446 for (j
= 0; p2
->clause
[j
]; j
++)
448 gcc_checking_assert (i
< MAX_CLAUSES
&& j
< MAX_CLAUSES
);
449 add_clause (conditions
, &out
, p
->clause
[i
] | p2
->clause
[j
]);
455 /* Having partial truth assignment in POSSIBLE_TRUTHS, return false
456 if predicate P is known to be false. */
459 evaluate_predicate (struct predicate
*p
, clause_t possible_truths
)
463 /* True remains true. */
464 if (true_predicate_p (p
))
467 gcc_assert (!(possible_truths
& (1 << predicate_false_condition
)));
469 /* See if we can find clause we can disprove. */
470 for (i
= 0; p
->clause
[i
]; i
++)
472 gcc_checking_assert (i
< MAX_CLAUSES
);
473 if (!(p
->clause
[i
] & possible_truths
))
479 /* Return the probability in range 0...REG_BR_PROB_BASE that the predicated
480 instruction will be recomputed per invocation of the inlined call. */
483 predicate_probability (conditions conds
,
484 struct predicate
*p
, clause_t possible_truths
,
485 vec
<inline_param_summary
> inline_param_summary
)
488 int combined_prob
= REG_BR_PROB_BASE
;
490 /* True remains true. */
491 if (true_predicate_p (p
))
492 return REG_BR_PROB_BASE
;
494 if (false_predicate_p (p
))
497 gcc_assert (!(possible_truths
& (1 << predicate_false_condition
)));
499 /* See if we can find clause we can disprove. */
500 for (i
= 0; p
->clause
[i
]; i
++)
502 gcc_checking_assert (i
< MAX_CLAUSES
);
503 if (!(p
->clause
[i
] & possible_truths
))
509 if (!inline_param_summary
.exists ())
510 return REG_BR_PROB_BASE
;
511 for (i2
= 0; i2
< NUM_CONDITIONS
; i2
++)
512 if ((p
->clause
[i
] & possible_truths
) & (1 << i2
))
514 if (i2
>= predicate_first_dynamic_condition
)
517 &(*conds
)[i2
- predicate_first_dynamic_condition
];
518 if (c
->code
== CHANGED
520 (int) inline_param_summary
.length ()))
523 inline_param_summary
[c
->operand_num
].change_prob
;
524 this_prob
= MAX (this_prob
, iprob
);
527 this_prob
= REG_BR_PROB_BASE
;
530 this_prob
= REG_BR_PROB_BASE
;
532 combined_prob
= MIN (this_prob
, combined_prob
);
537 return combined_prob
;
541 /* Dump conditional COND. */
544 dump_condition (FILE *f
, conditions conditions
, int cond
)
547 if (cond
== predicate_false_condition
)
548 fprintf (f
, "false");
549 else if (cond
== predicate_not_inlined_condition
)
550 fprintf (f
, "not inlined");
553 c
= &(*conditions
)[cond
- predicate_first_dynamic_condition
];
554 fprintf (f
, "op%i", c
->operand_num
);
556 fprintf (f
, "[%soffset: " HOST_WIDE_INT_PRINT_DEC
"]",
557 c
->by_ref
? "ref " : "", c
->offset
);
558 if (c
->code
== IS_NOT_CONSTANT
)
560 fprintf (f
, " not constant");
563 if (c
->code
== CHANGED
)
565 fprintf (f
, " changed");
568 fprintf (f
, " %s ", op_symbol_code (c
->code
));
569 print_generic_expr (f
, c
->val
, 1);
574 /* Dump clause CLAUSE. */
577 dump_clause (FILE *f
, conditions conds
, clause_t clause
)
584 for (i
= 0; i
< NUM_CONDITIONS
; i
++)
585 if (clause
& (1 << i
))
590 dump_condition (f
, conds
, i
);
596 /* Dump predicate PREDICATE. */
599 dump_predicate (FILE *f
, conditions conds
, struct predicate
*pred
)
602 if (true_predicate_p (pred
))
603 dump_clause (f
, conds
, 0);
605 for (i
= 0; pred
->clause
[i
]; i
++)
609 dump_clause (f
, conds
, pred
->clause
[i
]);
615 /* Dump inline hints. */
617 dump_inline_hints (FILE *f
, inline_hints hints
)
621 fprintf (f
, "inline hints:");
622 if (hints
& INLINE_HINT_indirect_call
)
624 hints
&= ~INLINE_HINT_indirect_call
;
625 fprintf (f
, " indirect_call");
627 if (hints
& INLINE_HINT_loop_iterations
)
629 hints
&= ~INLINE_HINT_loop_iterations
;
630 fprintf (f
, " loop_iterations");
632 if (hints
& INLINE_HINT_loop_stride
)
634 hints
&= ~INLINE_HINT_loop_stride
;
635 fprintf (f
, " loop_stride");
637 if (hints
& INLINE_HINT_same_scc
)
639 hints
&= ~INLINE_HINT_same_scc
;
640 fprintf (f
, " same_scc");
642 if (hints
& INLINE_HINT_in_scc
)
644 hints
&= ~INLINE_HINT_in_scc
;
645 fprintf (f
, " in_scc");
647 if (hints
& INLINE_HINT_cross_module
)
649 hints
&= ~INLINE_HINT_cross_module
;
650 fprintf (f
, " cross_module");
652 if (hints
& INLINE_HINT_declared_inline
)
654 hints
&= ~INLINE_HINT_declared_inline
;
655 fprintf (f
, " declared_inline");
657 if (hints
& INLINE_HINT_array_index
)
659 hints
&= ~INLINE_HINT_array_index
;
660 fprintf (f
, " array_index");
662 if (hints
& INLINE_HINT_known_hot
)
664 hints
&= ~INLINE_HINT_known_hot
;
665 fprintf (f
, " known_hot");
671 /* Record SIZE and TIME under condition PRED into the inline summary. */
674 account_size_time (struct inline_summary
*summary
, int size
, int time
,
675 struct predicate
*pred
)
681 if (false_predicate_p (pred
))
684 /* We need to create initial empty unconitional clause, but otherwie
685 we don't need to account empty times and sizes. */
686 if (!size
&& !time
&& summary
->entry
)
689 /* Watch overflow that might result from insane profiles. */
690 if (time
> MAX_TIME
* INLINE_TIME_SCALE
)
691 time
= MAX_TIME
* INLINE_TIME_SCALE
;
692 gcc_assert (time
>= 0);
694 for (i
= 0; vec_safe_iterate (summary
->entry
, i
, &e
); i
++)
695 if (predicates_equal_p (&e
->predicate
, pred
))
704 e
= &(*summary
->entry
)[0];
705 gcc_assert (!e
->predicate
.clause
[0]);
706 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
708 "\t\tReached limit on number of entries, "
709 "ignoring the predicate.");
711 if (dump_file
&& (dump_flags
& TDF_DETAILS
) && (time
|| size
))
714 "\t\tAccounting size:%3.2f, time:%3.2f on %spredicate:",
715 ((double) size
) / INLINE_SIZE_SCALE
,
716 ((double) time
) / INLINE_TIME_SCALE
, found
? "" : "new ");
717 dump_predicate (dump_file
, summary
->conds
, pred
);
721 struct size_time_entry new_entry
;
722 new_entry
.size
= size
;
723 new_entry
.time
= time
;
724 new_entry
.predicate
= *pred
;
725 vec_safe_push (summary
->entry
, new_entry
);
731 if (e
->time
> MAX_TIME
* INLINE_TIME_SCALE
)
732 e
->time
= MAX_TIME
* INLINE_TIME_SCALE
;
736 /* We proved E to be unreachable, redirect it to __bultin_unreachable. */
738 static struct cgraph_edge
*
739 redirect_to_unreachable (struct cgraph_edge
*e
)
741 struct cgraph_node
*callee
= !e
->inline_failed
? e
->callee
: NULL
;
742 struct cgraph_node
*target
= cgraph_node::get_create
743 (builtin_decl_implicit (BUILT_IN_UNREACHABLE
));
746 e
= e
->resolve_speculation (target
->decl
);
748 e
->make_direct (target
);
750 e
->redirect_callee (target
);
751 struct inline_edge_summary
*es
= inline_edge_summary (e
);
752 e
->inline_failed
= CIF_UNREACHABLE
;
755 es
->call_stmt_size
= 0;
756 es
->call_stmt_time
= 0;
758 callee
->remove_symbol_and_inline_clones ();
762 /* Set predicate for edge E. */
765 edge_set_predicate (struct cgraph_edge
*e
, struct predicate
*predicate
)
767 /* If the edge is determined to be never executed, redirect it
768 to BUILTIN_UNREACHABLE to save inliner from inlining into it. */
769 if (predicate
&& false_predicate_p (predicate
)
770 /* When handling speculative edges, we need to do the redirection
771 just once. Do it always on the direct edge, so we do not
772 attempt to resolve speculation while duplicating the edge. */
773 && (!e
->speculative
|| e
->callee
))
774 e
= redirect_to_unreachable (e
);
776 struct inline_edge_summary
*es
= inline_edge_summary (e
);
777 if (predicate
&& !true_predicate_p (predicate
))
780 es
->predicate
= edge_predicate_pool
.allocate ();
781 *es
->predicate
= *predicate
;
786 edge_predicate_pool
.remove (es
->predicate
);
787 es
->predicate
= NULL
;
791 /* Set predicate for hint *P. */
794 set_hint_predicate (struct predicate
**p
, struct predicate new_predicate
)
796 if (false_predicate_p (&new_predicate
) || true_predicate_p (&new_predicate
))
799 edge_predicate_pool
.remove (*p
);
805 *p
= edge_predicate_pool
.allocate ();
811 /* KNOWN_VALS is partial mapping of parameters of NODE to constant values.
812 KNOWN_AGGS is a vector of aggreggate jump functions for each parameter.
813 Return clause of possible truths. When INLINE_P is true, assume that we are
816 ERROR_MARK means compile time invariant. */
819 evaluate_conditions_for_known_args (struct cgraph_node
*node
,
821 vec
<tree
> known_vals
,
822 vec
<ipa_agg_jump_function_p
>
825 clause_t clause
= inline_p
? 0 : 1 << predicate_not_inlined_condition
;
826 struct inline_summary
*info
= inline_summaries
->get (node
);
830 for (i
= 0; vec_safe_iterate (info
->conds
, i
, &c
); i
++)
835 /* We allow call stmt to have fewer arguments than the callee function
836 (especially for K&R style programs). So bound check here (we assume
837 known_aggs vector, if non-NULL, has the same length as
839 gcc_checking_assert (!known_aggs
.exists ()
840 || (known_vals
.length () == known_aggs
.length ()));
841 if (c
->operand_num
>= (int) known_vals
.length ())
843 clause
|= 1 << (i
+ predicate_first_dynamic_condition
);
849 struct ipa_agg_jump_function
*agg
;
851 if (c
->code
== CHANGED
853 && (known_vals
[c
->operand_num
] == error_mark_node
))
856 if (known_aggs
.exists ())
858 agg
= known_aggs
[c
->operand_num
];
859 val
= ipa_find_agg_cst_for_param (agg
, c
->offset
, c
->by_ref
);
866 val
= known_vals
[c
->operand_num
];
867 if (val
== error_mark_node
&& c
->code
!= CHANGED
)
873 clause
|= 1 << (i
+ predicate_first_dynamic_condition
);
876 if (c
->code
== IS_NOT_CONSTANT
|| c
->code
== CHANGED
)
879 if (operand_equal_p (TYPE_SIZE (TREE_TYPE (c
->val
)),
880 TYPE_SIZE (TREE_TYPE (val
)), 0))
882 val
= fold_unary (VIEW_CONVERT_EXPR
, TREE_TYPE (c
->val
), val
);
885 ? fold_binary_to_constant (c
->code
, boolean_type_node
, val
, c
->val
)
888 if (res
&& integer_zerop (res
))
891 clause
|= 1 << (i
+ predicate_first_dynamic_condition
);
897 /* Work out what conditions might be true at invocation of E. */
900 evaluate_properties_for_edge (struct cgraph_edge
*e
, bool inline_p
,
901 clause_t
*clause_ptr
,
902 vec
<tree
> *known_vals_ptr
,
903 vec
<ipa_polymorphic_call_context
>
905 vec
<ipa_agg_jump_function_p
> *known_aggs_ptr
)
907 struct cgraph_node
*callee
= e
->callee
->ultimate_alias_target ();
908 struct inline_summary
*info
= inline_summaries
->get (callee
);
909 vec
<tree
> known_vals
= vNULL
;
910 vec
<ipa_agg_jump_function_p
> known_aggs
= vNULL
;
913 *clause_ptr
= inline_p
? 0 : 1 << predicate_not_inlined_condition
;
915 known_vals_ptr
->create (0);
916 if (known_contexts_ptr
)
917 known_contexts_ptr
->create (0);
919 if (ipa_node_params_sum
920 && !e
->call_stmt_cannot_inline_p
921 && ((clause_ptr
&& info
->conds
) || known_vals_ptr
|| known_contexts_ptr
))
923 struct ipa_node_params
*parms_info
;
924 struct ipa_edge_args
*args
= IPA_EDGE_REF (e
);
925 struct inline_edge_summary
*es
= inline_edge_summary (e
);
926 int i
, count
= ipa_get_cs_argument_count (args
);
928 if (e
->caller
->global
.inlined_to
)
929 parms_info
= IPA_NODE_REF (e
->caller
->global
.inlined_to
);
931 parms_info
= IPA_NODE_REF (e
->caller
);
933 if (count
&& (info
->conds
|| known_vals_ptr
))
934 known_vals
.safe_grow_cleared (count
);
935 if (count
&& (info
->conds
|| known_aggs_ptr
))
936 known_aggs
.safe_grow_cleared (count
);
937 if (count
&& known_contexts_ptr
)
938 known_contexts_ptr
->safe_grow_cleared (count
);
940 for (i
= 0; i
< count
; i
++)
942 struct ipa_jump_func
*jf
= ipa_get_ith_jump_func (args
, i
);
943 tree cst
= ipa_value_from_jfunc (parms_info
, jf
);
945 if (!cst
&& e
->call_stmt
946 && i
< (int)gimple_call_num_args (e
->call_stmt
))
948 cst
= gimple_call_arg (e
->call_stmt
, i
);
949 if (!is_gimple_min_invariant (cst
))
954 gcc_checking_assert (TREE_CODE (cst
) != TREE_BINFO
);
955 if (known_vals
.exists ())
958 else if (inline_p
&& !es
->param
[i
].change_prob
)
959 known_vals
[i
] = error_mark_node
;
961 if (known_contexts_ptr
)
962 (*known_contexts_ptr
)[i
] = ipa_context_from_jfunc (parms_info
, e
,
964 /* TODO: When IPA-CP starts propagating and merging aggregate jump
965 functions, use its knowledge of the caller too, just like the
966 scalar case above. */
967 known_aggs
[i
] = &jf
->agg
;
970 else if (e
->call_stmt
&& !e
->call_stmt_cannot_inline_p
971 && ((clause_ptr
&& info
->conds
) || known_vals_ptr
))
973 int i
, count
= (int)gimple_call_num_args (e
->call_stmt
);
975 if (count
&& (info
->conds
|| known_vals_ptr
))
976 known_vals
.safe_grow_cleared (count
);
977 for (i
= 0; i
< count
; i
++)
979 tree cst
= gimple_call_arg (e
->call_stmt
, i
);
980 if (!is_gimple_min_invariant (cst
))
988 *clause_ptr
= evaluate_conditions_for_known_args (callee
, inline_p
,
989 known_vals
, known_aggs
);
992 *known_vals_ptr
= known_vals
;
994 known_vals
.release ();
997 *known_aggs_ptr
= known_aggs
;
999 known_aggs
.release ();
1003 /* Allocate the inline summary vector or resize it to cover all cgraph nodes. */
1006 inline_summary_alloc (void)
1008 if (!edge_removal_hook_holder
)
1009 edge_removal_hook_holder
=
1010 symtab
->add_edge_removal_hook (&inline_edge_removal_hook
, NULL
);
1011 if (!edge_duplication_hook_holder
)
1012 edge_duplication_hook_holder
=
1013 symtab
->add_edge_duplication_hook (&inline_edge_duplication_hook
, NULL
);
1015 if (!inline_summaries
)
1016 inline_summaries
= (inline_summary_t
*) inline_summary_t::create_ggc (symtab
);
1018 if (inline_edge_summary_vec
.length () <= (unsigned) symtab
->edges_max_uid
)
1019 inline_edge_summary_vec
.safe_grow_cleared (symtab
->edges_max_uid
+ 1);
1022 /* We are called multiple time for given function; clear
1023 data from previous run so they are not cumulated. */
1026 reset_inline_edge_summary (struct cgraph_edge
*e
)
1028 if (e
->uid
< (int) inline_edge_summary_vec
.length ())
1030 struct inline_edge_summary
*es
= inline_edge_summary (e
);
1032 es
->call_stmt_size
= es
->call_stmt_time
= 0;
1034 edge_predicate_pool
.remove (es
->predicate
);
1035 es
->predicate
= NULL
;
1036 es
->param
.release ();
1040 /* We are called multiple time for given function; clear
1041 data from previous run so they are not cumulated. */
1044 reset_inline_summary (struct cgraph_node
*node
,
1045 inline_summary
*info
)
1047 struct cgraph_edge
*e
;
1049 info
->self_size
= info
->self_time
= 0;
1050 info
->estimated_stack_size
= 0;
1051 info
->estimated_self_stack_size
= 0;
1052 info
->stack_frame_offset
= 0;
1057 if (info
->loop_iterations
)
1059 edge_predicate_pool
.remove (info
->loop_iterations
);
1060 info
->loop_iterations
= NULL
;
1062 if (info
->loop_stride
)
1064 edge_predicate_pool
.remove (info
->loop_stride
);
1065 info
->loop_stride
= NULL
;
1067 if (info
->array_index
)
1069 edge_predicate_pool
.remove (info
->array_index
);
1070 info
->array_index
= NULL
;
1072 vec_free (info
->conds
);
1073 vec_free (info
->entry
);
1074 for (e
= node
->callees
; e
; e
= e
->next_callee
)
1075 reset_inline_edge_summary (e
);
1076 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
1077 reset_inline_edge_summary (e
);
1080 /* Hook that is called by cgraph.c when a node is removed. */
1083 inline_summary_t::remove (cgraph_node
*node
, inline_summary
*info
)
1085 reset_inline_summary (node
, info
);
1088 /* Remap predicate P of former function to be predicate of duplicated function.
1089 POSSIBLE_TRUTHS is clause of possible truths in the duplicated node,
1090 INFO is inline summary of the duplicated node. */
1092 static struct predicate
1093 remap_predicate_after_duplication (struct predicate
*p
,
1094 clause_t possible_truths
,
1095 struct inline_summary
*info
)
1097 struct predicate new_predicate
= true_predicate ();
1099 for (j
= 0; p
->clause
[j
]; j
++)
1100 if (!(possible_truths
& p
->clause
[j
]))
1102 new_predicate
= false_predicate ();
1106 add_clause (info
->conds
, &new_predicate
,
1107 possible_truths
& p
->clause
[j
]);
1108 return new_predicate
;
1111 /* Same as remap_predicate_after_duplication but handle hint predicate *P.
1112 Additionally care about allocating new memory slot for updated predicate
1113 and set it to NULL when it becomes true or false (and thus uninteresting).
1117 remap_hint_predicate_after_duplication (struct predicate
**p
,
1118 clause_t possible_truths
,
1119 struct inline_summary
*info
)
1121 struct predicate new_predicate
;
1126 new_predicate
= remap_predicate_after_duplication (*p
,
1127 possible_truths
, info
);
1128 /* We do not want to free previous predicate; it is used by node origin. */
1130 set_hint_predicate (p
, new_predicate
);
1134 /* Hook that is called by cgraph.c when a node is duplicated. */
1136 inline_summary_t::duplicate (cgraph_node
*src
,
1139 inline_summary
*info
)
1141 inline_summary_alloc ();
1142 memcpy (info
, inline_summaries
->get (src
), sizeof (inline_summary
));
1143 /* TODO: as an optimization, we may avoid copying conditions
1144 that are known to be false or true. */
1145 info
->conds
= vec_safe_copy (info
->conds
);
1147 /* When there are any replacements in the function body, see if we can figure
1148 out that something was optimized out. */
1149 if (ipa_node_params_sum
&& dst
->clone
.tree_map
)
1151 vec
<size_time_entry
, va_gc
> *entry
= info
->entry
;
1152 /* Use SRC parm info since it may not be copied yet. */
1153 struct ipa_node_params
*parms_info
= IPA_NODE_REF (src
);
1154 vec
<tree
> known_vals
= vNULL
;
1155 int count
= ipa_get_param_count (parms_info
);
1157 clause_t possible_truths
;
1158 struct predicate true_pred
= true_predicate ();
1160 int optimized_out_size
= 0;
1161 bool inlined_to_p
= false;
1162 struct cgraph_edge
*edge
, *next
;
1165 known_vals
.safe_grow_cleared (count
);
1166 for (i
= 0; i
< count
; i
++)
1168 struct ipa_replace_map
*r
;
1170 for (j
= 0; vec_safe_iterate (dst
->clone
.tree_map
, j
, &r
); j
++)
1172 if (((!r
->old_tree
&& r
->parm_num
== i
)
1173 || (r
->old_tree
&& r
->old_tree
== ipa_get_param (parms_info
, i
)))
1174 && r
->replace_p
&& !r
->ref_p
)
1176 known_vals
[i
] = r
->new_tree
;
1181 possible_truths
= evaluate_conditions_for_known_args (dst
, false,
1184 known_vals
.release ();
1186 account_size_time (info
, 0, 0, &true_pred
);
1188 /* Remap size_time vectors.
1189 Simplify the predicate by prunning out alternatives that are known
1191 TODO: as on optimization, we can also eliminate conditions known
1193 for (i
= 0; vec_safe_iterate (entry
, i
, &e
); i
++)
1195 struct predicate new_predicate
;
1196 new_predicate
= remap_predicate_after_duplication (&e
->predicate
,
1199 if (false_predicate_p (&new_predicate
))
1200 optimized_out_size
+= e
->size
;
1202 account_size_time (info
, e
->size
, e
->time
, &new_predicate
);
1205 /* Remap edge predicates with the same simplification as above.
1206 Also copy constantness arrays. */
1207 for (edge
= dst
->callees
; edge
; edge
= next
)
1209 struct predicate new_predicate
;
1210 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
1211 next
= edge
->next_callee
;
1213 if (!edge
->inline_failed
)
1214 inlined_to_p
= true;
1217 new_predicate
= remap_predicate_after_duplication (es
->predicate
,
1220 if (false_predicate_p (&new_predicate
)
1221 && !false_predicate_p (es
->predicate
))
1222 optimized_out_size
+= es
->call_stmt_size
* INLINE_SIZE_SCALE
;
1223 edge_set_predicate (edge
, &new_predicate
);
1226 /* Remap indirect edge predicates with the same simplificaiton as above.
1227 Also copy constantness arrays. */
1228 for (edge
= dst
->indirect_calls
; edge
; edge
= next
)
1230 struct predicate new_predicate
;
1231 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
1232 next
= edge
->next_callee
;
1234 gcc_checking_assert (edge
->inline_failed
);
1237 new_predicate
= remap_predicate_after_duplication (es
->predicate
,
1240 if (false_predicate_p (&new_predicate
)
1241 && !false_predicate_p (es
->predicate
))
1242 optimized_out_size
+= es
->call_stmt_size
* INLINE_SIZE_SCALE
;
1243 edge_set_predicate (edge
, &new_predicate
);
1245 remap_hint_predicate_after_duplication (&info
->loop_iterations
,
1246 possible_truths
, info
);
1247 remap_hint_predicate_after_duplication (&info
->loop_stride
,
1248 possible_truths
, info
);
1249 remap_hint_predicate_after_duplication (&info
->array_index
,
1250 possible_truths
, info
);
1252 /* If inliner or someone after inliner will ever start producing
1253 non-trivial clones, we will get trouble with lack of information
1254 about updating self sizes, because size vectors already contains
1255 sizes of the calees. */
1256 gcc_assert (!inlined_to_p
|| !optimized_out_size
);
1260 info
->entry
= vec_safe_copy (info
->entry
);
1261 if (info
->loop_iterations
)
1263 predicate p
= *info
->loop_iterations
;
1264 info
->loop_iterations
= NULL
;
1265 set_hint_predicate (&info
->loop_iterations
, p
);
1267 if (info
->loop_stride
)
1269 predicate p
= *info
->loop_stride
;
1270 info
->loop_stride
= NULL
;
1271 set_hint_predicate (&info
->loop_stride
, p
);
1273 if (info
->array_index
)
1275 predicate p
= *info
->array_index
;
1276 info
->array_index
= NULL
;
1277 set_hint_predicate (&info
->array_index
, p
);
1280 if (!dst
->global
.inlined_to
)
1281 inline_update_overall_summary (dst
);
1285 /* Hook that is called by cgraph.c when a node is duplicated. */
1288 inline_edge_duplication_hook (struct cgraph_edge
*src
,
1289 struct cgraph_edge
*dst
,
1290 ATTRIBUTE_UNUSED
void *data
)
1292 struct inline_edge_summary
*info
;
1293 struct inline_edge_summary
*srcinfo
;
1294 inline_summary_alloc ();
1295 info
= inline_edge_summary (dst
);
1296 srcinfo
= inline_edge_summary (src
);
1297 memcpy (info
, srcinfo
, sizeof (struct inline_edge_summary
));
1298 info
->predicate
= NULL
;
1299 edge_set_predicate (dst
, srcinfo
->predicate
);
1300 info
->param
= srcinfo
->param
.copy ();
1301 if (!dst
->indirect_unknown_callee
&& src
->indirect_unknown_callee
)
1303 info
->call_stmt_size
-= (eni_size_weights
.indirect_call_cost
1304 - eni_size_weights
.call_cost
);
1305 info
->call_stmt_time
-= (eni_time_weights
.indirect_call_cost
1306 - eni_time_weights
.call_cost
);
1311 /* Keep edge cache consistent across edge removal. */
1314 inline_edge_removal_hook (struct cgraph_edge
*edge
,
1315 void *data ATTRIBUTE_UNUSED
)
1317 if (edge_growth_cache
.exists ())
1318 reset_edge_growth_cache (edge
);
1319 reset_inline_edge_summary (edge
);
1323 /* Initialize growth caches. */
1326 initialize_growth_caches (void)
1328 if (symtab
->edges_max_uid
)
1329 edge_growth_cache
.safe_grow_cleared (symtab
->edges_max_uid
);
1333 /* Free growth caches. */
1336 free_growth_caches (void)
1338 edge_growth_cache
.release ();
1342 /* Dump edge summaries associated to NODE and recursively to all clones.
1343 Indent by INDENT. */
1346 dump_inline_edge_summary (FILE *f
, int indent
, struct cgraph_node
*node
,
1347 struct inline_summary
*info
)
1349 struct cgraph_edge
*edge
;
1350 for (edge
= node
->callees
; edge
; edge
= edge
->next_callee
)
1352 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
1353 struct cgraph_node
*callee
= edge
->callee
->ultimate_alias_target ();
1357 "%*s%s/%i %s\n%*s loop depth:%2i freq:%4i size:%2i"
1358 " time: %2i callee size:%2i stack:%2i",
1359 indent
, "", callee
->name (), callee
->order
,
1360 !edge
->inline_failed
1361 ? "inlined" : cgraph_inline_failed_string (edge
-> inline_failed
),
1362 indent
, "", es
->loop_depth
, edge
->frequency
,
1363 es
->call_stmt_size
, es
->call_stmt_time
,
1364 (int) inline_summaries
->get (callee
)->size
/ INLINE_SIZE_SCALE
,
1365 (int) inline_summaries
->get (callee
)->estimated_stack_size
);
1369 fprintf (f
, " predicate: ");
1370 dump_predicate (f
, info
->conds
, es
->predicate
);
1374 if (es
->param
.exists ())
1375 for (i
= 0; i
< (int) es
->param
.length (); i
++)
1377 int prob
= es
->param
[i
].change_prob
;
1380 fprintf (f
, "%*s op%i is compile time invariant\n",
1382 else if (prob
!= REG_BR_PROB_BASE
)
1383 fprintf (f
, "%*s op%i change %f%% of time\n", indent
+ 2, "", i
,
1384 prob
* 100.0 / REG_BR_PROB_BASE
);
1386 if (!edge
->inline_failed
)
1388 fprintf (f
, "%*sStack frame offset %i, callee self size %i,"
1389 " callee size %i\n",
1391 (int) inline_summaries
->get (callee
)->stack_frame_offset
,
1392 (int) inline_summaries
->get (callee
)->estimated_self_stack_size
,
1393 (int) inline_summaries
->get (callee
)->estimated_stack_size
);
1394 dump_inline_edge_summary (f
, indent
+ 2, callee
, info
);
1397 for (edge
= node
->indirect_calls
; edge
; edge
= edge
->next_callee
)
1399 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
1400 fprintf (f
, "%*sindirect call loop depth:%2i freq:%4i size:%2i"
1404 edge
->frequency
, es
->call_stmt_size
, es
->call_stmt_time
);
1407 fprintf (f
, "predicate: ");
1408 dump_predicate (f
, info
->conds
, es
->predicate
);
1417 dump_inline_summary (FILE *f
, struct cgraph_node
*node
)
1419 if (node
->definition
)
1421 struct inline_summary
*s
= inline_summaries
->get (node
);
1424 fprintf (f
, "Inline summary for %s/%i", node
->name (),
1426 if (DECL_DISREGARD_INLINE_LIMITS (node
->decl
))
1427 fprintf (f
, " always_inline");
1429 fprintf (f
, " inlinable");
1430 if (s
->contains_cilk_spawn
)
1431 fprintf (f
, " contains_cilk_spawn");
1432 fprintf (f
, "\n self time: %i\n", s
->self_time
);
1433 fprintf (f
, " global time: %i\n", s
->time
);
1434 fprintf (f
, " self size: %i\n", s
->self_size
);
1435 fprintf (f
, " global size: %i\n", s
->size
);
1436 fprintf (f
, " min size: %i\n", s
->min_size
);
1437 fprintf (f
, " self stack: %i\n",
1438 (int) s
->estimated_self_stack_size
);
1439 fprintf (f
, " global stack: %i\n", (int) s
->estimated_stack_size
);
1441 fprintf (f
, " estimated growth:%i\n", (int) s
->growth
);
1443 fprintf (f
, " In SCC: %i\n", (int) s
->scc_no
);
1444 for (i
= 0; vec_safe_iterate (s
->entry
, i
, &e
); i
++)
1446 fprintf (f
, " size:%f, time:%f, predicate:",
1447 (double) e
->size
/ INLINE_SIZE_SCALE
,
1448 (double) e
->time
/ INLINE_TIME_SCALE
);
1449 dump_predicate (f
, s
->conds
, &e
->predicate
);
1451 if (s
->loop_iterations
)
1453 fprintf (f
, " loop iterations:");
1454 dump_predicate (f
, s
->conds
, s
->loop_iterations
);
1458 fprintf (f
, " loop stride:");
1459 dump_predicate (f
, s
->conds
, s
->loop_stride
);
1463 fprintf (f
, " array index:");
1464 dump_predicate (f
, s
->conds
, s
->array_index
);
1466 fprintf (f
, " calls:\n");
1467 dump_inline_edge_summary (f
, 4, node
, s
);
1473 debug_inline_summary (struct cgraph_node
*node
)
1475 dump_inline_summary (stderr
, node
);
1479 dump_inline_summaries (FILE *f
)
1481 struct cgraph_node
*node
;
1483 FOR_EACH_DEFINED_FUNCTION (node
)
1484 if (!node
->global
.inlined_to
)
1485 dump_inline_summary (f
, node
);
1488 /* Give initial reasons why inlining would fail on EDGE. This gets either
1489 nullified or usually overwritten by more precise reasons later. */
1492 initialize_inline_failed (struct cgraph_edge
*e
)
1494 struct cgraph_node
*callee
= e
->callee
;
1496 if (e
->indirect_unknown_callee
)
1497 e
->inline_failed
= CIF_INDIRECT_UNKNOWN_CALL
;
1498 else if (!callee
->definition
)
1499 e
->inline_failed
= CIF_BODY_NOT_AVAILABLE
;
1500 else if (callee
->local
.redefined_extern_inline
)
1501 e
->inline_failed
= CIF_REDEFINED_EXTERN_INLINE
;
1502 else if (e
->call_stmt_cannot_inline_p
)
1503 e
->inline_failed
= CIF_MISMATCHED_ARGUMENTS
;
1504 else if (cfun
&& fn_contains_cilk_spawn_p (cfun
))
1505 /* We can't inline if the function is spawing a function. */
1506 e
->inline_failed
= CIF_FUNCTION_NOT_INLINABLE
;
1508 e
->inline_failed
= CIF_FUNCTION_NOT_CONSIDERED
;
1511 /* Callback of walk_aliased_vdefs. Flags that it has been invoked to the
1512 boolean variable pointed to by DATA. */
1515 mark_modified (ao_ref
*ao ATTRIBUTE_UNUSED
, tree vdef ATTRIBUTE_UNUSED
,
1518 bool *b
= (bool *) data
;
1523 /* If OP refers to value of function parameter, return the corresponding
1527 unmodified_parm_1 (gimple
*stmt
, tree op
)
1529 /* SSA_NAME referring to parm default def? */
1530 if (TREE_CODE (op
) == SSA_NAME
1531 && SSA_NAME_IS_DEFAULT_DEF (op
)
1532 && TREE_CODE (SSA_NAME_VAR (op
)) == PARM_DECL
)
1533 return SSA_NAME_VAR (op
);
1534 /* Non-SSA parm reference? */
1535 if (TREE_CODE (op
) == PARM_DECL
)
1537 bool modified
= false;
1540 ao_ref_init (&refd
, op
);
1541 walk_aliased_vdefs (&refd
, gimple_vuse (stmt
), mark_modified
, &modified
,
1549 /* If OP refers to value of function parameter, return the corresponding
1550 parameter. Also traverse chains of SSA register assignments. */
1553 unmodified_parm (gimple
*stmt
, tree op
)
1555 tree res
= unmodified_parm_1 (stmt
, op
);
1559 if (TREE_CODE (op
) == SSA_NAME
1560 && !SSA_NAME_IS_DEFAULT_DEF (op
)
1561 && gimple_assign_single_p (SSA_NAME_DEF_STMT (op
)))
1562 return unmodified_parm (SSA_NAME_DEF_STMT (op
),
1563 gimple_assign_rhs1 (SSA_NAME_DEF_STMT (op
)));
1567 /* If OP refers to a value of a function parameter or value loaded from an
1568 aggregate passed to a parameter (either by value or reference), return TRUE
1569 and store the number of the parameter to *INDEX_P and information whether
1570 and how it has been loaded from an aggregate into *AGGPOS. INFO describes
1571 the function parameters, STMT is the statement in which OP is used or
1575 unmodified_parm_or_parm_agg_item (struct ipa_func_body_info
*fbi
,
1576 gimple
*stmt
, tree op
, int *index_p
,
1577 struct agg_position_info
*aggpos
)
1579 tree res
= unmodified_parm_1 (stmt
, op
);
1581 gcc_checking_assert (aggpos
);
1584 *index_p
= ipa_get_param_decl_index (fbi
->info
, res
);
1587 aggpos
->agg_contents
= false;
1588 aggpos
->by_ref
= false;
1592 if (TREE_CODE (op
) == SSA_NAME
)
1594 if (SSA_NAME_IS_DEFAULT_DEF (op
)
1595 || !gimple_assign_single_p (SSA_NAME_DEF_STMT (op
)))
1597 stmt
= SSA_NAME_DEF_STMT (op
);
1598 op
= gimple_assign_rhs1 (stmt
);
1599 if (!REFERENCE_CLASS_P (op
))
1600 return unmodified_parm_or_parm_agg_item (fbi
, stmt
, op
, index_p
,
1604 aggpos
->agg_contents
= true;
1605 return ipa_load_from_parm_agg (fbi
, fbi
->info
->descriptors
,
1606 stmt
, op
, index_p
, &aggpos
->offset
,
1607 NULL
, &aggpos
->by_ref
);
1610 /* See if statement might disappear after inlining.
1611 0 - means not eliminated
1612 1 - half of statements goes away
1613 2 - for sure it is eliminated.
1614 We are not terribly sophisticated, basically looking for simple abstraction
1615 penalty wrappers. */
1618 eliminated_by_inlining_prob (gimple
*stmt
)
1620 enum gimple_code code
= gimple_code (stmt
);
1621 enum tree_code rhs_code
;
1631 if (gimple_num_ops (stmt
) != 2)
1634 rhs_code
= gimple_assign_rhs_code (stmt
);
1636 /* Casts of parameters, loads from parameters passed by reference
1637 and stores to return value or parameters are often free after
1638 inlining dua to SRA and further combining.
1639 Assume that half of statements goes away. */
1640 if (CONVERT_EXPR_CODE_P (rhs_code
)
1641 || rhs_code
== VIEW_CONVERT_EXPR
1642 || rhs_code
== ADDR_EXPR
1643 || gimple_assign_rhs_class (stmt
) == GIMPLE_SINGLE_RHS
)
1645 tree rhs
= gimple_assign_rhs1 (stmt
);
1646 tree lhs
= gimple_assign_lhs (stmt
);
1647 tree inner_rhs
= get_base_address (rhs
);
1648 tree inner_lhs
= get_base_address (lhs
);
1649 bool rhs_free
= false;
1650 bool lhs_free
= false;
1657 /* Reads of parameter are expected to be free. */
1658 if (unmodified_parm (stmt
, inner_rhs
))
1660 /* Match expressions of form &this->field. Those will most likely
1661 combine with something upstream after inlining. */
1662 else if (TREE_CODE (inner_rhs
) == ADDR_EXPR
)
1664 tree op
= get_base_address (TREE_OPERAND (inner_rhs
, 0));
1665 if (TREE_CODE (op
) == PARM_DECL
)
1667 else if (TREE_CODE (op
) == MEM_REF
1668 && unmodified_parm (stmt
, TREE_OPERAND (op
, 0)))
1672 /* When parameter is not SSA register because its address is taken
1673 and it is just copied into one, the statement will be completely
1674 free after inlining (we will copy propagate backward). */
1675 if (rhs_free
&& is_gimple_reg (lhs
))
1678 /* Reads of parameters passed by reference
1679 expected to be free (i.e. optimized out after inlining). */
1680 if (TREE_CODE (inner_rhs
) == MEM_REF
1681 && unmodified_parm (stmt
, TREE_OPERAND (inner_rhs
, 0)))
1684 /* Copying parameter passed by reference into gimple register is
1685 probably also going to copy propagate, but we can't be quite
1687 if (rhs_free
&& is_gimple_reg (lhs
))
1690 /* Writes to parameters, parameters passed by value and return value
1691 (either dirrectly or passed via invisible reference) are free.
1693 TODO: We ought to handle testcase like
1694 struct a {int a,b;};
1696 retrurnsturct (void)
1702 This translate into:
1717 For that we either need to copy ipa-split logic detecting writes
1719 if (TREE_CODE (inner_lhs
) == PARM_DECL
1720 || TREE_CODE (inner_lhs
) == RESULT_DECL
1721 || (TREE_CODE (inner_lhs
) == MEM_REF
1722 && (unmodified_parm (stmt
, TREE_OPERAND (inner_lhs
, 0))
1723 || (TREE_CODE (TREE_OPERAND (inner_lhs
, 0)) == SSA_NAME
1724 && SSA_NAME_VAR (TREE_OPERAND (inner_lhs
, 0))
1725 && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND
1727 0))) == RESULT_DECL
))))
1730 && (is_gimple_reg (rhs
) || is_gimple_min_invariant (rhs
)))
1732 if (lhs_free
&& rhs_free
)
1742 /* If BB ends by a conditional we can turn into predicates, attach corresponding
1743 predicates to the CFG edges. */
1746 set_cond_stmt_execution_predicate (struct ipa_func_body_info
*fbi
,
1747 struct inline_summary
*summary
,
1753 struct agg_position_info aggpos
;
1754 enum tree_code code
, inverted_code
;
1760 last
= last_stmt (bb
);
1761 if (!last
|| gimple_code (last
) != GIMPLE_COND
)
1763 if (!is_gimple_ip_invariant (gimple_cond_rhs (last
)))
1765 op
= gimple_cond_lhs (last
);
1766 /* TODO: handle conditionals like
1769 if (unmodified_parm_or_parm_agg_item (fbi
, last
, op
, &index
, &aggpos
))
1771 code
= gimple_cond_code (last
);
1772 inverted_code
= invert_tree_comparison (code
, HONOR_NANS (op
));
1774 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1776 enum tree_code this_code
= (e
->flags
& EDGE_TRUE_VALUE
1777 ? code
: inverted_code
);
1778 /* invert_tree_comparison will return ERROR_MARK on FP
1779 comparsions that are not EQ/NE instead of returning proper
1780 unordered one. Be sure it is not confused with NON_CONSTANT. */
1781 if (this_code
!= ERROR_MARK
)
1783 struct predicate p
= add_condition (summary
, index
, &aggpos
,
1785 gimple_cond_rhs (last
));
1786 e
->aux
= edge_predicate_pool
.allocate ();
1787 *(struct predicate
*) e
->aux
= p
;
1792 if (TREE_CODE (op
) != SSA_NAME
)
1795 if (builtin_constant_p (op))
1799 Here we can predicate nonconstant_code. We can't
1800 really handle constant_code since we have no predicate
1801 for this and also the constant code is not known to be
1802 optimized away when inliner doen't see operand is constant.
1803 Other optimizers might think otherwise. */
1804 if (gimple_cond_code (last
) != NE_EXPR
1805 || !integer_zerop (gimple_cond_rhs (last
)))
1807 set_stmt
= SSA_NAME_DEF_STMT (op
);
1808 if (!gimple_call_builtin_p (set_stmt
, BUILT_IN_CONSTANT_P
)
1809 || gimple_call_num_args (set_stmt
) != 1)
1811 op2
= gimple_call_arg (set_stmt
, 0);
1812 if (!unmodified_parm_or_parm_agg_item (fbi
, set_stmt
, op2
, &index
, &aggpos
))
1814 FOR_EACH_EDGE (e
, ei
, bb
->succs
) if (e
->flags
& EDGE_FALSE_VALUE
)
1816 struct predicate p
= add_condition (summary
, index
, &aggpos
,
1817 IS_NOT_CONSTANT
, NULL_TREE
);
1818 e
->aux
= edge_predicate_pool
.allocate ();
1819 *(struct predicate
*) e
->aux
= p
;
1824 /* If BB ends by a switch we can turn into predicates, attach corresponding
1825 predicates to the CFG edges. */
1828 set_switch_stmt_execution_predicate (struct ipa_func_body_info
*fbi
,
1829 struct inline_summary
*summary
,
1835 struct agg_position_info aggpos
;
1841 lastg
= last_stmt (bb
);
1842 if (!lastg
|| gimple_code (lastg
) != GIMPLE_SWITCH
)
1844 gswitch
*last
= as_a
<gswitch
*> (lastg
);
1845 op
= gimple_switch_index (last
);
1846 if (!unmodified_parm_or_parm_agg_item (fbi
, last
, op
, &index
, &aggpos
))
1849 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1851 e
->aux
= edge_predicate_pool
.allocate ();
1852 *(struct predicate
*) e
->aux
= false_predicate ();
1854 n
= gimple_switch_num_labels (last
);
1855 for (case_idx
= 0; case_idx
< n
; ++case_idx
)
1857 tree cl
= gimple_switch_label (last
, case_idx
);
1861 e
= find_edge (bb
, label_to_block (CASE_LABEL (cl
)));
1862 min
= CASE_LOW (cl
);
1863 max
= CASE_HIGH (cl
);
1865 /* For default we might want to construct predicate that none
1866 of cases is met, but it is bit hard to do not having negations
1867 of conditionals handy. */
1869 p
= true_predicate ();
1871 p
= add_condition (summary
, index
, &aggpos
, EQ_EXPR
, min
);
1874 struct predicate p1
, p2
;
1875 p1
= add_condition (summary
, index
, &aggpos
, GE_EXPR
, min
);
1876 p2
= add_condition (summary
, index
, &aggpos
, LE_EXPR
, max
);
1877 p
= and_predicates (summary
->conds
, &p1
, &p2
);
1879 *(struct predicate
*) e
->aux
1880 = or_predicates (summary
->conds
, &p
, (struct predicate
*) e
->aux
);
1885 /* For each BB in NODE attach to its AUX pointer predicate under
1886 which it is executable. */
1889 compute_bb_predicates (struct ipa_func_body_info
*fbi
,
1890 struct cgraph_node
*node
,
1891 struct inline_summary
*summary
)
1893 struct function
*my_function
= DECL_STRUCT_FUNCTION (node
->decl
);
1897 FOR_EACH_BB_FN (bb
, my_function
)
1899 set_cond_stmt_execution_predicate (fbi
, summary
, bb
);
1900 set_switch_stmt_execution_predicate (fbi
, summary
, bb
);
1903 /* Entry block is always executable. */
1904 ENTRY_BLOCK_PTR_FOR_FN (my_function
)->aux
1905 = edge_predicate_pool
.allocate ();
1906 *(struct predicate
*) ENTRY_BLOCK_PTR_FOR_FN (my_function
)->aux
1907 = true_predicate ();
1909 /* A simple dataflow propagation of predicates forward in the CFG.
1910 TODO: work in reverse postorder. */
1914 FOR_EACH_BB_FN (bb
, my_function
)
1916 struct predicate p
= false_predicate ();
1919 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
1923 struct predicate this_bb_predicate
1924 = *(struct predicate
*) e
->src
->aux
;
1927 = and_predicates (summary
->conds
, &this_bb_predicate
,
1928 (struct predicate
*) e
->aux
);
1929 p
= or_predicates (summary
->conds
, &p
, &this_bb_predicate
);
1930 if (true_predicate_p (&p
))
1934 if (false_predicate_p (&p
))
1935 gcc_assert (!bb
->aux
);
1941 bb
->aux
= edge_predicate_pool
.allocate ();
1942 *((struct predicate
*) bb
->aux
) = p
;
1944 else if (!predicates_equal_p (&p
, (struct predicate
*) bb
->aux
))
1946 /* This OR operation is needed to ensure monotonous data flow
1947 in the case we hit the limit on number of clauses and the
1948 and/or operations above give approximate answers. */
1949 p
= or_predicates (summary
->conds
, &p
, (struct predicate
*)bb
->aux
);
1950 if (!predicates_equal_p (&p
, (struct predicate
*) bb
->aux
))
1953 *((struct predicate
*) bb
->aux
) = p
;
1962 /* We keep info about constantness of SSA names. */
1964 typedef struct predicate predicate_t
;
1965 /* Return predicate specifying when the STMT might have result that is not
1966 a compile time constant. */
1968 static struct predicate
1969 will_be_nonconstant_expr_predicate (struct ipa_node_params
*info
,
1970 struct inline_summary
*summary
,
1972 vec
<predicate_t
> nonconstant_names
)
1977 while (UNARY_CLASS_P (expr
))
1978 expr
= TREE_OPERAND (expr
, 0);
1980 parm
= unmodified_parm (NULL
, expr
);
1981 if (parm
&& (index
= ipa_get_param_decl_index (info
, parm
)) >= 0)
1982 return add_condition (summary
, index
, NULL
, CHANGED
, NULL_TREE
);
1983 if (is_gimple_min_invariant (expr
))
1984 return false_predicate ();
1985 if (TREE_CODE (expr
) == SSA_NAME
)
1986 return nonconstant_names
[SSA_NAME_VERSION (expr
)];
1987 if (BINARY_CLASS_P (expr
) || COMPARISON_CLASS_P (expr
))
1989 struct predicate p1
= will_be_nonconstant_expr_predicate
1990 (info
, summary
, TREE_OPERAND (expr
, 0),
1992 struct predicate p2
;
1993 if (true_predicate_p (&p1
))
1995 p2
= will_be_nonconstant_expr_predicate (info
, summary
,
1996 TREE_OPERAND (expr
, 1),
1998 return or_predicates (summary
->conds
, &p1
, &p2
);
2000 else if (TREE_CODE (expr
) == COND_EXPR
)
2002 struct predicate p1
= will_be_nonconstant_expr_predicate
2003 (info
, summary
, TREE_OPERAND (expr
, 0),
2005 struct predicate p2
;
2006 if (true_predicate_p (&p1
))
2008 p2
= will_be_nonconstant_expr_predicate (info
, summary
,
2009 TREE_OPERAND (expr
, 1),
2011 if (true_predicate_p (&p2
))
2013 p1
= or_predicates (summary
->conds
, &p1
, &p2
);
2014 p2
= will_be_nonconstant_expr_predicate (info
, summary
,
2015 TREE_OPERAND (expr
, 2),
2017 return or_predicates (summary
->conds
, &p1
, &p2
);
2024 return false_predicate ();
2028 /* Return predicate specifying when the STMT might have result that is not
2029 a compile time constant. */
2031 static struct predicate
2032 will_be_nonconstant_predicate (struct ipa_func_body_info
*fbi
,
2033 struct inline_summary
*summary
,
2035 vec
<predicate_t
> nonconstant_names
)
2037 struct predicate p
= true_predicate ();
2040 struct predicate op_non_const
;
2043 struct agg_position_info aggpos
;
2045 /* What statments might be optimized away
2046 when their arguments are constant. */
2047 if (gimple_code (stmt
) != GIMPLE_ASSIGN
2048 && gimple_code (stmt
) != GIMPLE_COND
2049 && gimple_code (stmt
) != GIMPLE_SWITCH
2050 && (gimple_code (stmt
) != GIMPLE_CALL
2051 || !(gimple_call_flags (stmt
) & ECF_CONST
)))
2054 /* Stores will stay anyway. */
2055 if (gimple_store_p (stmt
))
2058 is_load
= gimple_assign_load_p (stmt
);
2060 /* Loads can be optimized when the value is known. */
2064 gcc_assert (gimple_assign_single_p (stmt
));
2065 op
= gimple_assign_rhs1 (stmt
);
2066 if (!unmodified_parm_or_parm_agg_item (fbi
, stmt
, op
, &base_index
,
2073 /* See if we understand all operands before we start
2074 adding conditionals. */
2075 FOR_EACH_SSA_TREE_OPERAND (use
, stmt
, iter
, SSA_OP_USE
)
2077 tree parm
= unmodified_parm (stmt
, use
);
2078 /* For arguments we can build a condition. */
2079 if (parm
&& ipa_get_param_decl_index (fbi
->info
, parm
) >= 0)
2081 if (TREE_CODE (use
) != SSA_NAME
)
2083 /* If we know when operand is constant,
2084 we still can say something useful. */
2085 if (!true_predicate_p (&nonconstant_names
[SSA_NAME_VERSION (use
)]))
2092 add_condition (summary
, base_index
, &aggpos
, CHANGED
, NULL
);
2094 op_non_const
= false_predicate ();
2095 FOR_EACH_SSA_TREE_OPERAND (use
, stmt
, iter
, SSA_OP_USE
)
2097 tree parm
= unmodified_parm (stmt
, use
);
2100 if (parm
&& (index
= ipa_get_param_decl_index (fbi
->info
, parm
)) >= 0)
2102 if (index
!= base_index
)
2103 p
= add_condition (summary
, index
, NULL
, CHANGED
, NULL_TREE
);
2108 p
= nonconstant_names
[SSA_NAME_VERSION (use
)];
2109 op_non_const
= or_predicates (summary
->conds
, &p
, &op_non_const
);
2111 if ((gimple_code (stmt
) == GIMPLE_ASSIGN
|| gimple_code (stmt
) == GIMPLE_CALL
)
2112 && gimple_op (stmt
, 0)
2113 && TREE_CODE (gimple_op (stmt
, 0)) == SSA_NAME
)
2114 nonconstant_names
[SSA_NAME_VERSION (gimple_op (stmt
, 0))]
2116 return op_non_const
;
2119 struct record_modified_bb_info
2125 /* Callback of walk_aliased_vdefs. Records basic blocks where the value may be
2126 set except for info->stmt. */
2129 record_modified (ao_ref
*ao ATTRIBUTE_UNUSED
, tree vdef
, void *data
)
2131 struct record_modified_bb_info
*info
=
2132 (struct record_modified_bb_info
*) data
;
2133 if (SSA_NAME_DEF_STMT (vdef
) == info
->stmt
)
2135 bitmap_set_bit (info
->bb_set
,
2136 SSA_NAME_IS_DEFAULT_DEF (vdef
)
2137 ? ENTRY_BLOCK_PTR_FOR_FN (cfun
)->index
2138 : gimple_bb (SSA_NAME_DEF_STMT (vdef
))->index
);
2142 /* Return probability (based on REG_BR_PROB_BASE) that I-th parameter of STMT
2143 will change since last invocation of STMT.
2145 Value 0 is reserved for compile time invariants.
2146 For common parameters it is REG_BR_PROB_BASE. For loop invariants it
2147 ought to be REG_BR_PROB_BASE / estimated_iters. */
2150 param_change_prob (gimple
*stmt
, int i
)
2152 tree op
= gimple_call_arg (stmt
, i
);
2153 basic_block bb
= gimple_bb (stmt
);
2156 /* Global invariants neve change. */
2157 if (is_gimple_min_invariant (op
))
2159 /* We would have to do non-trivial analysis to really work out what
2160 is the probability of value to change (i.e. when init statement
2161 is in a sibling loop of the call).
2163 We do an conservative estimate: when call is executed N times more often
2164 than the statement defining value, we take the frequency 1/N. */
2165 if (TREE_CODE (op
) == SSA_NAME
)
2170 return REG_BR_PROB_BASE
;
2172 if (SSA_NAME_IS_DEFAULT_DEF (op
))
2173 init_freq
= ENTRY_BLOCK_PTR_FOR_FN (cfun
)->frequency
;
2175 init_freq
= gimple_bb (SSA_NAME_DEF_STMT (op
))->frequency
;
2179 if (init_freq
< bb
->frequency
)
2180 return MAX (GCOV_COMPUTE_SCALE (init_freq
, bb
->frequency
), 1);
2182 return REG_BR_PROB_BASE
;
2185 base
= get_base_address (op
);
2190 struct record_modified_bb_info info
;
2193 tree init
= ctor_for_folding (base
);
2195 if (init
!= error_mark_node
)
2198 return REG_BR_PROB_BASE
;
2199 ao_ref_init (&refd
, op
);
2201 info
.bb_set
= BITMAP_ALLOC (NULL
);
2202 walk_aliased_vdefs (&refd
, gimple_vuse (stmt
), record_modified
, &info
,
2204 if (bitmap_bit_p (info
.bb_set
, bb
->index
))
2206 BITMAP_FREE (info
.bb_set
);
2207 return REG_BR_PROB_BASE
;
2210 /* Assume that every memory is initialized at entry.
2211 TODO: Can we easilly determine if value is always defined
2212 and thus we may skip entry block? */
2213 if (ENTRY_BLOCK_PTR_FOR_FN (cfun
)->frequency
)
2214 max
= ENTRY_BLOCK_PTR_FOR_FN (cfun
)->frequency
;
2218 EXECUTE_IF_SET_IN_BITMAP (info
.bb_set
, 0, index
, bi
)
2219 max
= MIN (max
, BASIC_BLOCK_FOR_FN (cfun
, index
)->frequency
);
2221 BITMAP_FREE (info
.bb_set
);
2222 if (max
< bb
->frequency
)
2223 return MAX (GCOV_COMPUTE_SCALE (max
, bb
->frequency
), 1);
2225 return REG_BR_PROB_BASE
;
2227 return REG_BR_PROB_BASE
;
2230 /* Find whether a basic block BB is the final block of a (half) diamond CFG
2231 sub-graph and if the predicate the condition depends on is known. If so,
2232 return true and store the pointer the predicate in *P. */
2235 phi_result_unknown_predicate (struct ipa_node_params
*info
,
2236 inline_summary
*summary
, basic_block bb
,
2237 struct predicate
*p
,
2238 vec
<predicate_t
> nonconstant_names
)
2242 basic_block first_bb
= NULL
;
2245 if (single_pred_p (bb
))
2247 *p
= false_predicate ();
2251 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
2253 if (single_succ_p (e
->src
))
2255 if (!single_pred_p (e
->src
))
2258 first_bb
= single_pred (e
->src
);
2259 else if (single_pred (e
->src
) != first_bb
)
2266 else if (e
->src
!= first_bb
)
2274 stmt
= last_stmt (first_bb
);
2276 || gimple_code (stmt
) != GIMPLE_COND
2277 || !is_gimple_ip_invariant (gimple_cond_rhs (stmt
)))
2280 *p
= will_be_nonconstant_expr_predicate (info
, summary
,
2281 gimple_cond_lhs (stmt
),
2283 if (true_predicate_p (p
))
2289 /* Given a PHI statement in a function described by inline properties SUMMARY
2290 and *P being the predicate describing whether the selected PHI argument is
2291 known, store a predicate for the result of the PHI statement into
2292 NONCONSTANT_NAMES, if possible. */
2295 predicate_for_phi_result (struct inline_summary
*summary
, gphi
*phi
,
2296 struct predicate
*p
,
2297 vec
<predicate_t
> nonconstant_names
)
2301 for (i
= 0; i
< gimple_phi_num_args (phi
); i
++)
2303 tree arg
= gimple_phi_arg (phi
, i
)->def
;
2304 if (!is_gimple_min_invariant (arg
))
2306 gcc_assert (TREE_CODE (arg
) == SSA_NAME
);
2307 *p
= or_predicates (summary
->conds
, p
,
2308 &nonconstant_names
[SSA_NAME_VERSION (arg
)]);
2309 if (true_predicate_p (p
))
2314 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2316 fprintf (dump_file
, "\t\tphi predicate: ");
2317 dump_predicate (dump_file
, summary
->conds
, p
);
2319 nonconstant_names
[SSA_NAME_VERSION (gimple_phi_result (phi
))] = *p
;
2322 /* Return predicate specifying when array index in access OP becomes non-constant. */
2324 static struct predicate
2325 array_index_predicate (inline_summary
*info
,
2326 vec
< predicate_t
> nonconstant_names
, tree op
)
2328 struct predicate p
= false_predicate ();
2329 while (handled_component_p (op
))
2331 if (TREE_CODE (op
) == ARRAY_REF
|| TREE_CODE (op
) == ARRAY_RANGE_REF
)
2333 if (TREE_CODE (TREE_OPERAND (op
, 1)) == SSA_NAME
)
2334 p
= or_predicates (info
->conds
, &p
,
2335 &nonconstant_names
[SSA_NAME_VERSION
2336 (TREE_OPERAND (op
, 1))]);
2338 op
= TREE_OPERAND (op
, 0);
2343 /* For a typical usage of __builtin_expect (a<b, 1), we
2344 may introduce an extra relation stmt:
2345 With the builtin, we have
2348 t3 = __builtin_expect (t2, 1);
2351 Without the builtin, we have
2354 This affects the size/time estimation and may have
2355 an impact on the earlier inlining.
2356 Here find this pattern and fix it up later. */
2359 find_foldable_builtin_expect (basic_block bb
)
2361 gimple_stmt_iterator bsi
;
2363 for (bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
2365 gimple
*stmt
= gsi_stmt (bsi
);
2366 if (gimple_call_builtin_p (stmt
, BUILT_IN_EXPECT
)
2367 || (is_gimple_call (stmt
)
2368 && gimple_call_internal_p (stmt
)
2369 && gimple_call_internal_fn (stmt
) == IFN_BUILTIN_EXPECT
))
2371 tree var
= gimple_call_lhs (stmt
);
2372 tree arg
= gimple_call_arg (stmt
, 0);
2373 use_operand_p use_p
;
2380 gcc_assert (TREE_CODE (var
) == SSA_NAME
);
2382 while (TREE_CODE (arg
) == SSA_NAME
)
2384 gimple
*stmt_tmp
= SSA_NAME_DEF_STMT (arg
);
2385 if (!is_gimple_assign (stmt_tmp
))
2387 switch (gimple_assign_rhs_code (stmt_tmp
))
2406 arg
= gimple_assign_rhs1 (stmt_tmp
);
2409 if (match
&& single_imm_use (var
, &use_p
, &use_stmt
)
2410 && gimple_code (use_stmt
) == GIMPLE_COND
)
2417 /* Return true when the basic blocks contains only clobbers followed by RESX.
2418 Such BBs are kept around to make removal of dead stores possible with
2419 presence of EH and will be optimized out by optimize_clobbers later in the
2422 NEED_EH is used to recurse in case the clobber has non-EH predecestors
2423 that can be clobber only, too.. When it is false, the RESX is not necessary
2424 on the end of basic block. */
2427 clobber_only_eh_bb_p (basic_block bb
, bool need_eh
= true)
2429 gimple_stmt_iterator gsi
= gsi_last_bb (bb
);
2435 if (gsi_end_p (gsi
))
2437 if (gimple_code (gsi_stmt (gsi
)) != GIMPLE_RESX
)
2441 else if (!single_succ_p (bb
))
2444 for (; !gsi_end_p (gsi
); gsi_prev (&gsi
))
2446 gimple
*stmt
= gsi_stmt (gsi
);
2447 if (is_gimple_debug (stmt
))
2449 if (gimple_clobber_p (stmt
))
2451 if (gimple_code (stmt
) == GIMPLE_LABEL
)
2456 /* See if all predecestors are either throws or clobber only BBs. */
2457 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
2458 if (!(e
->flags
& EDGE_EH
)
2459 && !clobber_only_eh_bb_p (e
->src
, false))
2465 /* Compute function body size parameters for NODE.
2466 When EARLY is true, we compute only simple summaries without
2467 non-trivial predicates to drive the early inliner. */
2470 estimate_function_body_sizes (struct cgraph_node
*node
, bool early
)
2473 /* Estimate static overhead for function prologue/epilogue and alignment. */
2475 /* Benefits are scaled by probability of elimination that is in range
2478 struct function
*my_function
= DECL_STRUCT_FUNCTION (node
->decl
);
2480 struct inline_summary
*info
= inline_summaries
->get (node
);
2481 struct predicate bb_predicate
;
2482 struct ipa_func_body_info fbi
;
2483 vec
<predicate_t
> nonconstant_names
= vNULL
;
2486 predicate array_index
= true_predicate ();
2487 gimple
*fix_builtin_expect_stmt
;
2489 gcc_assert (my_function
&& my_function
->cfg
);
2490 gcc_assert (cfun
== my_function
);
2492 memset(&fbi
, 0, sizeof(fbi
));
2496 /* When optimizing and analyzing for IPA inliner, initialize loop optimizer
2497 so we can produce proper inline hints.
2499 When optimizing and analyzing for early inliner, initialize node params
2500 so we can produce correct BB predicates. */
2502 if (opt_for_fn (node
->decl
, optimize
))
2504 calculate_dominance_info (CDI_DOMINATORS
);
2506 loop_optimizer_init (LOOPS_NORMAL
| LOOPS_HAVE_RECORDED_EXITS
);
2509 ipa_check_create_node_params ();
2510 ipa_initialize_node_params (node
);
2513 if (ipa_node_params_sum
)
2516 fbi
.info
= IPA_NODE_REF (node
);
2517 fbi
.bb_infos
= vNULL
;
2518 fbi
.bb_infos
.safe_grow_cleared (last_basic_block_for_fn (cfun
));
2519 fbi
.param_count
= count_formal_params(node
->decl
);
2520 nonconstant_names
.safe_grow_cleared
2521 (SSANAMES (my_function
)->length ());
2526 fprintf (dump_file
, "\nAnalyzing function body size: %s\n",
2529 /* When we run into maximal number of entries, we assign everything to the
2530 constant truth case. Be sure to have it in list. */
2531 bb_predicate
= true_predicate ();
2532 account_size_time (info
, 0, 0, &bb_predicate
);
2534 bb_predicate
= not_inlined_predicate ();
2535 account_size_time (info
, 2 * INLINE_SIZE_SCALE
, 0, &bb_predicate
);
2538 compute_bb_predicates (&fbi
, node
, info
);
2539 order
= XNEWVEC (int, n_basic_blocks_for_fn (cfun
));
2540 nblocks
= pre_and_rev_post_order_compute (NULL
, order
, false);
2541 for (n
= 0; n
< nblocks
; n
++)
2543 bb
= BASIC_BLOCK_FOR_FN (cfun
, order
[n
]);
2544 freq
= compute_call_stmt_bb_frequency (node
->decl
, bb
);
2545 if (clobber_only_eh_bb_p (bb
))
2547 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2548 fprintf (dump_file
, "\n Ignoring BB %i;"
2549 " it will be optimized away by cleanup_clobbers\n",
2554 /* TODO: Obviously predicates can be propagated down across CFG. */
2558 bb_predicate
= *(struct predicate
*) bb
->aux
;
2560 bb_predicate
= false_predicate ();
2563 bb_predicate
= true_predicate ();
2565 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2567 fprintf (dump_file
, "\n BB %i predicate:", bb
->index
);
2568 dump_predicate (dump_file
, info
->conds
, &bb_predicate
);
2571 if (fbi
.info
&& nonconstant_names
.exists ())
2573 struct predicate phi_predicate
;
2574 bool first_phi
= true;
2576 for (gphi_iterator bsi
= gsi_start_phis (bb
); !gsi_end_p (bsi
);
2580 && !phi_result_unknown_predicate (fbi
.info
, info
, bb
,
2585 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2587 fprintf (dump_file
, " ");
2588 print_gimple_stmt (dump_file
, gsi_stmt (bsi
), 0, 0);
2590 predicate_for_phi_result (info
, bsi
.phi (), &phi_predicate
,
2595 fix_builtin_expect_stmt
= find_foldable_builtin_expect (bb
);
2597 for (gimple_stmt_iterator bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
);
2600 gimple
*stmt
= gsi_stmt (bsi
);
2601 int this_size
= estimate_num_insns (stmt
, &eni_size_weights
);
2602 int this_time
= estimate_num_insns (stmt
, &eni_time_weights
);
2604 struct predicate will_be_nonconstant
;
2606 /* This relation stmt should be folded after we remove
2607 buildin_expect call. Adjust the cost here. */
2608 if (stmt
== fix_builtin_expect_stmt
)
2614 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2616 fprintf (dump_file
, " ");
2617 print_gimple_stmt (dump_file
, stmt
, 0, 0);
2618 fprintf (dump_file
, "\t\tfreq:%3.2f size:%3i time:%3i\n",
2619 ((double) freq
) / CGRAPH_FREQ_BASE
, this_size
,
2623 if (gimple_assign_load_p (stmt
) && nonconstant_names
.exists ())
2625 struct predicate this_array_index
;
2627 array_index_predicate (info
, nonconstant_names
,
2628 gimple_assign_rhs1 (stmt
));
2629 if (!false_predicate_p (&this_array_index
))
2631 and_predicates (info
->conds
, &array_index
,
2634 if (gimple_store_p (stmt
) && nonconstant_names
.exists ())
2636 struct predicate this_array_index
;
2638 array_index_predicate (info
, nonconstant_names
,
2639 gimple_get_lhs (stmt
));
2640 if (!false_predicate_p (&this_array_index
))
2642 and_predicates (info
->conds
, &array_index
,
2647 if (is_gimple_call (stmt
)
2648 && !gimple_call_internal_p (stmt
))
2650 struct cgraph_edge
*edge
= node
->get_edge (stmt
);
2651 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
2653 /* Special case: results of BUILT_IN_CONSTANT_P will be always
2654 resolved as constant. We however don't want to optimize
2655 out the cgraph edges. */
2656 if (nonconstant_names
.exists ()
2657 && gimple_call_builtin_p (stmt
, BUILT_IN_CONSTANT_P
)
2658 && gimple_call_lhs (stmt
)
2659 && TREE_CODE (gimple_call_lhs (stmt
)) == SSA_NAME
)
2661 struct predicate false_p
= false_predicate ();
2662 nonconstant_names
[SSA_NAME_VERSION (gimple_call_lhs (stmt
))]
2665 if (ipa_node_params_sum
)
2667 int count
= gimple_call_num_args (stmt
);
2671 es
->param
.safe_grow_cleared (count
);
2672 for (i
= 0; i
< count
; i
++)
2674 int prob
= param_change_prob (stmt
, i
);
2675 gcc_assert (prob
>= 0 && prob
<= REG_BR_PROB_BASE
);
2676 es
->param
[i
].change_prob
= prob
;
2680 es
->call_stmt_size
= this_size
;
2681 es
->call_stmt_time
= this_time
;
2682 es
->loop_depth
= bb_loop_depth (bb
);
2683 edge_set_predicate (edge
, &bb_predicate
);
2686 /* TODO: When conditional jump or swithc is known to be constant, but
2687 we did not translate it into the predicates, we really can account
2688 just maximum of the possible paths. */
2691 = will_be_nonconstant_predicate (&fbi
, info
,
2692 stmt
, nonconstant_names
);
2693 if (this_time
|| this_size
)
2699 prob
= eliminated_by_inlining_prob (stmt
);
2700 if (prob
== 1 && dump_file
&& (dump_flags
& TDF_DETAILS
))
2702 "\t\t50%% will be eliminated by inlining\n");
2703 if (prob
== 2 && dump_file
&& (dump_flags
& TDF_DETAILS
))
2704 fprintf (dump_file
, "\t\tWill be eliminated by inlining\n");
2707 p
= and_predicates (info
->conds
, &bb_predicate
,
2708 &will_be_nonconstant
);
2710 p
= true_predicate ();
2712 if (!false_predicate_p (&p
)
2713 || (is_gimple_call (stmt
)
2714 && !false_predicate_p (&bb_predicate
)))
2718 if (time
> MAX_TIME
* INLINE_TIME_SCALE
)
2719 time
= MAX_TIME
* INLINE_TIME_SCALE
;
2722 /* We account everything but the calls. Calls have their own
2723 size/time info attached to cgraph edges. This is necessary
2724 in order to make the cost disappear after inlining. */
2725 if (!is_gimple_call (stmt
))
2729 struct predicate ip
= not_inlined_predicate ();
2730 ip
= and_predicates (info
->conds
, &ip
, &p
);
2731 account_size_time (info
, this_size
* prob
,
2732 this_time
* prob
, &ip
);
2735 account_size_time (info
, this_size
* (2 - prob
),
2736 this_time
* (2 - prob
), &p
);
2739 gcc_assert (time
>= 0);
2740 gcc_assert (size
>= 0);
2744 set_hint_predicate (&inline_summaries
->get (node
)->array_index
, array_index
);
2745 time
= (time
+ CGRAPH_FREQ_BASE
/ 2) / CGRAPH_FREQ_BASE
;
2746 if (time
> MAX_TIME
)
2750 if (nonconstant_names
.exists () && !early
)
2753 predicate loop_iterations
= true_predicate ();
2754 predicate loop_stride
= true_predicate ();
2756 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2757 flow_loops_dump (dump_file
, NULL
, 0);
2759 FOR_EACH_LOOP (loop
, 0)
2764 struct tree_niter_desc niter_desc
;
2765 bb_predicate
= *(struct predicate
*) loop
->header
->aux
;
2767 exits
= get_loop_exit_edges (loop
);
2768 FOR_EACH_VEC_ELT (exits
, j
, ex
)
2769 if (number_of_iterations_exit (loop
, ex
, &niter_desc
, false)
2770 && !is_gimple_min_invariant (niter_desc
.niter
))
2772 predicate will_be_nonconstant
2773 = will_be_nonconstant_expr_predicate (fbi
.info
, info
,
2776 if (!true_predicate_p (&will_be_nonconstant
))
2777 will_be_nonconstant
= and_predicates (info
->conds
,
2779 &will_be_nonconstant
);
2780 if (!true_predicate_p (&will_be_nonconstant
)
2781 && !false_predicate_p (&will_be_nonconstant
))
2782 /* This is slightly inprecise. We may want to represent each
2783 loop with independent predicate. */
2785 and_predicates (info
->conds
, &loop_iterations
,
2786 &will_be_nonconstant
);
2790 for (gphi_iterator gsi
= gsi_start_phis (loop
->header
);
2791 !gsi_end_p (gsi
); gsi_next (&gsi
))
2793 gphi
*phi
= gsi
.phi ();
2794 tree use
= gimple_phi_result (phi
);
2796 predicate will_be_nonconstant
;
2797 if (virtual_operand_p (use
)
2798 || !simple_iv (loop
, loop
, use
, &iv
, true)
2799 || is_gimple_min_invariant (iv
.step
))
2802 = will_be_nonconstant_expr_predicate (fbi
.info
, info
,
2805 if (!true_predicate_p (&will_be_nonconstant
))
2806 will_be_nonconstant
= and_predicates (info
->conds
,
2808 &will_be_nonconstant
);
2809 if (!true_predicate_p (&will_be_nonconstant
)
2810 && !false_predicate_p (&will_be_nonconstant
))
2811 /* This is slightly inprecise. We may want to represent
2812 each loop with independent predicate. */
2813 loop_stride
= and_predicates (info
->conds
, &loop_stride
,
2814 &will_be_nonconstant
);
2817 set_hint_predicate (&inline_summaries
->get (node
)->loop_iterations
,
2819 set_hint_predicate (&inline_summaries
->get (node
)->loop_stride
, loop_stride
);
2822 FOR_ALL_BB_FN (bb
, my_function
)
2828 edge_predicate_pool
.remove ((predicate
*)bb
->aux
);
2830 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
2833 edge_predicate_pool
.remove ((predicate
*) e
->aux
);
2837 inline_summaries
->get (node
)->self_time
= time
;
2838 inline_summaries
->get (node
)->self_size
= size
;
2839 nonconstant_names
.release ();
2840 if (opt_for_fn (node
->decl
, optimize
))
2843 loop_optimizer_finalize ();
2844 else if (!ipa_edge_args_vector
)
2845 ipa_free_all_node_params ();
2846 free_dominance_info (CDI_DOMINATORS
);
2850 fprintf (dump_file
, "\n");
2851 dump_inline_summary (dump_file
, node
);
2856 /* Compute parameters of functions used by inliner.
2857 EARLY is true when we compute parameters for the early inliner */
2860 compute_inline_parameters (struct cgraph_node
*node
, bool early
)
2862 HOST_WIDE_INT self_stack_size
;
2863 struct cgraph_edge
*e
;
2864 struct inline_summary
*info
;
2866 gcc_assert (!node
->global
.inlined_to
);
2868 inline_summary_alloc ();
2870 info
= inline_summaries
->get (node
);
2871 reset_inline_summary (node
, info
);
2873 /* FIXME: Thunks are inlinable, but tree-inline don't know how to do that.
2874 Once this happen, we will need to more curefully predict call
2876 if (node
->thunk
.thunk_p
)
2878 struct inline_edge_summary
*es
= inline_edge_summary (node
->callees
);
2879 struct predicate t
= true_predicate ();
2881 info
->inlinable
= 0;
2882 node
->callees
->call_stmt_cannot_inline_p
= true;
2883 node
->local
.can_change_signature
= false;
2884 es
->call_stmt_time
= 1;
2885 es
->call_stmt_size
= 1;
2886 account_size_time (info
, 0, 0, &t
);
2890 /* Even is_gimple_min_invariant rely on current_function_decl. */
2891 push_cfun (DECL_STRUCT_FUNCTION (node
->decl
));
2893 /* Estimate the stack size for the function if we're optimizing. */
2894 self_stack_size
= optimize
? estimated_stack_frame_size (node
) : 0;
2895 info
->estimated_self_stack_size
= self_stack_size
;
2896 info
->estimated_stack_size
= self_stack_size
;
2897 info
->stack_frame_offset
= 0;
2899 /* Can this function be inlined at all? */
2900 if (!opt_for_fn (node
->decl
, optimize
)
2901 && !lookup_attribute ("always_inline",
2902 DECL_ATTRIBUTES (node
->decl
)))
2903 info
->inlinable
= false;
2905 info
->inlinable
= tree_inlinable_function_p (node
->decl
);
2907 info
->contains_cilk_spawn
= fn_contains_cilk_spawn_p (cfun
);
2909 /* Type attributes can use parameter indices to describe them. */
2910 if (TYPE_ATTRIBUTES (TREE_TYPE (node
->decl
)))
2911 node
->local
.can_change_signature
= false;
2914 /* Otherwise, inlinable functions always can change signature. */
2915 if (info
->inlinable
)
2916 node
->local
.can_change_signature
= true;
2919 /* Functions calling builtin_apply can not change signature. */
2920 for (e
= node
->callees
; e
; e
= e
->next_callee
)
2922 tree
cdecl = e
->callee
->decl
;
2923 if (DECL_BUILT_IN (cdecl)
2924 && DECL_BUILT_IN_CLASS (cdecl) == BUILT_IN_NORMAL
2925 && (DECL_FUNCTION_CODE (cdecl) == BUILT_IN_APPLY_ARGS
2926 || DECL_FUNCTION_CODE (cdecl) == BUILT_IN_VA_START
))
2929 node
->local
.can_change_signature
= !e
;
2932 estimate_function_body_sizes (node
, early
);
2934 for (e
= node
->callees
; e
; e
= e
->next_callee
)
2935 if (e
->callee
->comdat_local_p ())
2937 node
->calls_comdat_local
= (e
!= NULL
);
2939 /* Inlining characteristics are maintained by the cgraph_mark_inline. */
2940 info
->time
= info
->self_time
;
2941 info
->size
= info
->self_size
;
2942 info
->stack_frame_offset
= 0;
2943 info
->estimated_stack_size
= info
->estimated_self_stack_size
;
2944 #ifdef ENABLE_CHECKING
2945 inline_update_overall_summary (node
);
2946 gcc_assert (info
->time
== info
->self_time
&& info
->size
== info
->self_size
);
2953 /* Compute parameters of functions used by inliner using
2954 current_function_decl. */
2957 compute_inline_parameters_for_current (void)
2959 compute_inline_parameters (cgraph_node::get (current_function_decl
), true);
2965 const pass_data pass_data_inline_parameters
=
2967 GIMPLE_PASS
, /* type */
2968 "inline_param", /* name */
2969 OPTGROUP_INLINE
, /* optinfo_flags */
2970 TV_INLINE_PARAMETERS
, /* tv_id */
2971 0, /* properties_required */
2972 0, /* properties_provided */
2973 0, /* properties_destroyed */
2974 0, /* todo_flags_start */
2975 0, /* todo_flags_finish */
2978 class pass_inline_parameters
: public gimple_opt_pass
2981 pass_inline_parameters (gcc::context
*ctxt
)
2982 : gimple_opt_pass (pass_data_inline_parameters
, ctxt
)
2985 /* opt_pass methods: */
2986 opt_pass
* clone () { return new pass_inline_parameters (m_ctxt
); }
2987 virtual unsigned int execute (function
*)
2989 return compute_inline_parameters_for_current ();
2992 }; // class pass_inline_parameters
2997 make_pass_inline_parameters (gcc::context
*ctxt
)
2999 return new pass_inline_parameters (ctxt
);
3003 /* Estimate benefit devirtualizing indirect edge IE, provided KNOWN_VALS,
3004 KNOWN_CONTEXTS and KNOWN_AGGS. */
3007 estimate_edge_devirt_benefit (struct cgraph_edge
*ie
,
3008 int *size
, int *time
,
3009 vec
<tree
> known_vals
,
3010 vec
<ipa_polymorphic_call_context
> known_contexts
,
3011 vec
<ipa_agg_jump_function_p
> known_aggs
)
3014 struct cgraph_node
*callee
;
3015 struct inline_summary
*isummary
;
3016 enum availability avail
;
3019 if (!known_vals
.exists () && !known_contexts
.exists ())
3021 if (!opt_for_fn (ie
->caller
->decl
, flag_indirect_inlining
))
3024 target
= ipa_get_indirect_edge_target (ie
, known_vals
, known_contexts
,
3025 known_aggs
, &speculative
);
3026 if (!target
|| speculative
)
3029 /* Account for difference in cost between indirect and direct calls. */
3030 *size
-= (eni_size_weights
.indirect_call_cost
- eni_size_weights
.call_cost
);
3031 *time
-= (eni_time_weights
.indirect_call_cost
- eni_time_weights
.call_cost
);
3032 gcc_checking_assert (*time
>= 0);
3033 gcc_checking_assert (*size
>= 0);
3035 callee
= cgraph_node::get (target
);
3036 if (!callee
|| !callee
->definition
)
3038 callee
= callee
->function_symbol (&avail
);
3039 if (avail
< AVAIL_AVAILABLE
)
3041 isummary
= inline_summaries
->get (callee
);
3042 return isummary
->inlinable
;
3045 /* Increase SIZE, MIN_SIZE (if non-NULL) and TIME for size and time needed to
3046 handle edge E with probability PROB.
3047 Set HINTS if edge may be devirtualized.
3048 KNOWN_VALS, KNOWN_AGGS and KNOWN_CONTEXTS describe context of the call
3052 estimate_edge_size_and_time (struct cgraph_edge
*e
, int *size
, int *min_size
,
3055 vec
<tree
> known_vals
,
3056 vec
<ipa_polymorphic_call_context
> known_contexts
,
3057 vec
<ipa_agg_jump_function_p
> known_aggs
,
3058 inline_hints
*hints
)
3060 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3061 int call_size
= es
->call_stmt_size
;
3062 int call_time
= es
->call_stmt_time
;
3065 && estimate_edge_devirt_benefit (e
, &call_size
, &call_time
,
3066 known_vals
, known_contexts
, known_aggs
)
3067 && hints
&& e
->maybe_hot_p ())
3068 *hints
|= INLINE_HINT_indirect_call
;
3069 cur_size
= call_size
* INLINE_SIZE_SCALE
;
3072 *min_size
+= cur_size
;
3073 *time
+= apply_probability ((gcov_type
) call_time
, prob
)
3074 * e
->frequency
* (INLINE_TIME_SCALE
/ CGRAPH_FREQ_BASE
);
3075 if (*time
> MAX_TIME
* INLINE_TIME_SCALE
)
3076 *time
= MAX_TIME
* INLINE_TIME_SCALE
;
3081 /* Increase SIZE, MIN_SIZE and TIME for size and time needed to handle all
3082 calls in NODE. POSSIBLE_TRUTHS, KNOWN_VALS, KNOWN_AGGS and KNOWN_CONTEXTS
3083 describe context of the call site. */
3086 estimate_calls_size_and_time (struct cgraph_node
*node
, int *size
,
3087 int *min_size
, int *time
,
3088 inline_hints
*hints
,
3089 clause_t possible_truths
,
3090 vec
<tree
> known_vals
,
3091 vec
<ipa_polymorphic_call_context
> known_contexts
,
3092 vec
<ipa_agg_jump_function_p
> known_aggs
)
3094 struct cgraph_edge
*e
;
3095 for (e
= node
->callees
; e
; e
= e
->next_callee
)
3097 if (inline_edge_summary_vec
.length () <= (unsigned) e
->uid
)
3100 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3102 /* Do not care about zero sized builtins. */
3103 if (e
->inline_failed
&& !es
->call_stmt_size
)
3105 gcc_checking_assert (!es
->call_stmt_time
);
3109 || evaluate_predicate (es
->predicate
, possible_truths
))
3111 if (e
->inline_failed
)
3113 /* Predicates of calls shall not use NOT_CHANGED codes,
3114 sowe do not need to compute probabilities. */
3115 estimate_edge_size_and_time (e
, size
,
3116 es
->predicate
? NULL
: min_size
,
3117 time
, REG_BR_PROB_BASE
,
3118 known_vals
, known_contexts
,
3122 estimate_calls_size_and_time (e
->callee
, size
, min_size
, time
,
3125 known_vals
, known_contexts
,
3129 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
3131 if (inline_edge_summary_vec
.length () <= (unsigned) e
->uid
)
3134 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3136 || evaluate_predicate (es
->predicate
, possible_truths
))
3137 estimate_edge_size_and_time (e
, size
,
3138 es
->predicate
? NULL
: min_size
,
3139 time
, REG_BR_PROB_BASE
,
3140 known_vals
, known_contexts
, known_aggs
,
3146 /* Estimate size and time needed to execute NODE assuming
3147 POSSIBLE_TRUTHS clause, and KNOWN_VALS, KNOWN_AGGS and KNOWN_CONTEXTS
3148 information about NODE's arguments. If non-NULL use also probability
3149 information present in INLINE_PARAM_SUMMARY vector.
3150 Additionally detemine hints determined by the context. Finally compute
3151 minimal size needed for the call that is independent on the call context and
3152 can be used for fast estimates. Return the values in RET_SIZE,
3153 RET_MIN_SIZE, RET_TIME and RET_HINTS. */
3156 estimate_node_size_and_time (struct cgraph_node
*node
,
3157 clause_t possible_truths
,
3158 vec
<tree
> known_vals
,
3159 vec
<ipa_polymorphic_call_context
> known_contexts
,
3160 vec
<ipa_agg_jump_function_p
> known_aggs
,
3161 int *ret_size
, int *ret_min_size
, int *ret_time
,
3162 inline_hints
*ret_hints
,
3163 vec
<inline_param_summary
>
3164 inline_param_summary
)
3166 struct inline_summary
*info
= inline_summaries
->get (node
);
3171 inline_hints hints
= 0;
3174 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3177 fprintf (dump_file
, " Estimating body: %s/%i\n"
3178 " Known to be false: ", node
->name (),
3181 for (i
= predicate_not_inlined_condition
;
3182 i
< (predicate_first_dynamic_condition
3183 + (int) vec_safe_length (info
->conds
)); i
++)
3184 if (!(possible_truths
& (1 << i
)))
3187 fprintf (dump_file
, ", ");
3189 dump_condition (dump_file
, info
->conds
, i
);
3193 for (i
= 0; vec_safe_iterate (info
->entry
, i
, &e
); i
++)
3194 if (evaluate_predicate (&e
->predicate
, possible_truths
))
3197 gcc_checking_assert (e
->time
>= 0);
3198 gcc_checking_assert (time
>= 0);
3199 if (!inline_param_summary
.exists ())
3203 int prob
= predicate_probability (info
->conds
,
3206 inline_param_summary
);
3207 gcc_checking_assert (prob
>= 0);
3208 gcc_checking_assert (prob
<= REG_BR_PROB_BASE
);
3209 time
+= apply_probability ((gcov_type
) e
->time
, prob
);
3211 if (time
> MAX_TIME
* INLINE_TIME_SCALE
)
3212 time
= MAX_TIME
* INLINE_TIME_SCALE
;
3213 gcc_checking_assert (time
>= 0);
3216 gcc_checking_assert (true_predicate_p (&(*info
->entry
)[0].predicate
));
3217 min_size
= (*info
->entry
)[0].size
;
3218 gcc_checking_assert (size
>= 0);
3219 gcc_checking_assert (time
>= 0);
3221 if (info
->loop_iterations
3222 && !evaluate_predicate (info
->loop_iterations
, possible_truths
))
3223 hints
|= INLINE_HINT_loop_iterations
;
3224 if (info
->loop_stride
3225 && !evaluate_predicate (info
->loop_stride
, possible_truths
))
3226 hints
|= INLINE_HINT_loop_stride
;
3227 if (info
->array_index
3228 && !evaluate_predicate (info
->array_index
, possible_truths
))
3229 hints
|= INLINE_HINT_array_index
;
3231 hints
|= INLINE_HINT_in_scc
;
3232 if (DECL_DECLARED_INLINE_P (node
->decl
))
3233 hints
|= INLINE_HINT_declared_inline
;
3235 estimate_calls_size_and_time (node
, &size
, &min_size
, &time
, &hints
, possible_truths
,
3236 known_vals
, known_contexts
, known_aggs
);
3237 gcc_checking_assert (size
>= 0);
3238 gcc_checking_assert (time
>= 0);
3239 time
= RDIV (time
, INLINE_TIME_SCALE
);
3240 size
= RDIV (size
, INLINE_SIZE_SCALE
);
3241 min_size
= RDIV (min_size
, INLINE_SIZE_SCALE
);
3243 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3244 fprintf (dump_file
, "\n size:%i time:%i\n", (int) size
, (int) time
);
3250 *ret_min_size
= min_size
;
3257 /* Estimate size and time needed to execute callee of EDGE assuming that
3258 parameters known to be constant at caller of EDGE are propagated.
3259 KNOWN_VALS and KNOWN_CONTEXTS are vectors of assumed known constant values
3260 and types for parameters. */
3263 estimate_ipcp_clone_size_and_time (struct cgraph_node
*node
,
3264 vec
<tree
> known_vals
,
3265 vec
<ipa_polymorphic_call_context
>
3267 vec
<ipa_agg_jump_function_p
> known_aggs
,
3268 int *ret_size
, int *ret_time
,
3269 inline_hints
*hints
)
3273 clause
= evaluate_conditions_for_known_args (node
, false, known_vals
,
3275 estimate_node_size_and_time (node
, clause
, known_vals
, known_contexts
,
3276 known_aggs
, ret_size
, NULL
, ret_time
, hints
, vNULL
);
3279 /* Translate all conditions from callee representation into caller
3280 representation and symbolically evaluate predicate P into new predicate.
3282 INFO is inline_summary of function we are adding predicate into, CALLEE_INFO
3283 is summary of function predicate P is from. OPERAND_MAP is array giving
3284 callee formal IDs the caller formal IDs. POSSSIBLE_TRUTHS is clausule of all
3285 callee conditions that may be true in caller context. TOPLEV_PREDICATE is
3286 predicate under which callee is executed. OFFSET_MAP is an array of of
3287 offsets that need to be added to conditions, negative offset means that
3288 conditions relying on values passed by reference have to be discarded
3289 because they might not be preserved (and should be considered offset zero
3290 for other purposes). */
3292 static struct predicate
3293 remap_predicate (struct inline_summary
*info
,
3294 struct inline_summary
*callee_info
,
3295 struct predicate
*p
,
3296 vec
<int> operand_map
,
3297 vec
<int> offset_map
,
3298 clause_t possible_truths
, struct predicate
*toplev_predicate
)
3301 struct predicate out
= true_predicate ();
3303 /* True predicate is easy. */
3304 if (true_predicate_p (p
))
3305 return *toplev_predicate
;
3306 for (i
= 0; p
->clause
[i
]; i
++)
3308 clause_t clause
= p
->clause
[i
];
3310 struct predicate clause_predicate
= false_predicate ();
3312 gcc_assert (i
< MAX_CLAUSES
);
3314 for (cond
= 0; cond
< NUM_CONDITIONS
; cond
++)
3315 /* Do we have condition we can't disprove? */
3316 if (clause
& possible_truths
& (1 << cond
))
3318 struct predicate cond_predicate
;
3319 /* Work out if the condition can translate to predicate in the
3320 inlined function. */
3321 if (cond
>= predicate_first_dynamic_condition
)
3323 struct condition
*c
;
3325 c
= &(*callee_info
->conds
)[cond
3327 predicate_first_dynamic_condition
];
3328 /* See if we can remap condition operand to caller's operand.
3329 Otherwise give up. */
3330 if (!operand_map
.exists ()
3331 || (int) operand_map
.length () <= c
->operand_num
3332 || operand_map
[c
->operand_num
] == -1
3333 /* TODO: For non-aggregate conditions, adding an offset is
3334 basically an arithmetic jump function processing which
3335 we should support in future. */
3336 || ((!c
->agg_contents
|| !c
->by_ref
)
3337 && offset_map
[c
->operand_num
] > 0)
3338 || (c
->agg_contents
&& c
->by_ref
3339 && offset_map
[c
->operand_num
] < 0))
3340 cond_predicate
= true_predicate ();
3343 struct agg_position_info ap
;
3344 HOST_WIDE_INT offset_delta
= offset_map
[c
->operand_num
];
3345 if (offset_delta
< 0)
3347 gcc_checking_assert (!c
->agg_contents
|| !c
->by_ref
);
3350 gcc_assert (!c
->agg_contents
3351 || c
->by_ref
|| offset_delta
== 0);
3352 ap
.offset
= c
->offset
+ offset_delta
;
3353 ap
.agg_contents
= c
->agg_contents
;
3354 ap
.by_ref
= c
->by_ref
;
3355 cond_predicate
= add_condition (info
,
3356 operand_map
[c
->operand_num
],
3357 &ap
, c
->code
, c
->val
);
3360 /* Fixed conditions remains same, construct single
3361 condition predicate. */
3364 cond_predicate
.clause
[0] = 1 << cond
;
3365 cond_predicate
.clause
[1] = 0;
3367 clause_predicate
= or_predicates (info
->conds
, &clause_predicate
,
3370 out
= and_predicates (info
->conds
, &out
, &clause_predicate
);
3372 return and_predicates (info
->conds
, &out
, toplev_predicate
);
3376 /* Update summary information of inline clones after inlining.
3377 Compute peak stack usage. */
3380 inline_update_callee_summaries (struct cgraph_node
*node
, int depth
)
3382 struct cgraph_edge
*e
;
3383 struct inline_summary
*callee_info
= inline_summaries
->get (node
);
3384 struct inline_summary
*caller_info
= inline_summaries
->get (node
->callers
->caller
);
3387 callee_info
->stack_frame_offset
3388 = caller_info
->stack_frame_offset
3389 + caller_info
->estimated_self_stack_size
;
3390 peak
= callee_info
->stack_frame_offset
3391 + callee_info
->estimated_self_stack_size
;
3392 if (inline_summaries
->get (node
->global
.inlined_to
)->estimated_stack_size
< peak
)
3393 inline_summaries
->get (node
->global
.inlined_to
)->estimated_stack_size
= peak
;
3394 ipa_propagate_frequency (node
);
3395 for (e
= node
->callees
; e
; e
= e
->next_callee
)
3397 if (!e
->inline_failed
)
3398 inline_update_callee_summaries (e
->callee
, depth
);
3399 inline_edge_summary (e
)->loop_depth
+= depth
;
3401 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
3402 inline_edge_summary (e
)->loop_depth
+= depth
;
3405 /* Update change_prob of EDGE after INLINED_EDGE has been inlined.
3406 When functoin A is inlined in B and A calls C with parameter that
3407 changes with probability PROB1 and C is known to be passthroug
3408 of argument if B that change with probability PROB2, the probability
3409 of change is now PROB1*PROB2. */
3412 remap_edge_change_prob (struct cgraph_edge
*inlined_edge
,
3413 struct cgraph_edge
*edge
)
3415 if (ipa_node_params_sum
)
3418 struct ipa_edge_args
*args
= IPA_EDGE_REF (edge
);
3419 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3420 struct inline_edge_summary
*inlined_es
3421 = inline_edge_summary (inlined_edge
);
3423 for (i
= 0; i
< ipa_get_cs_argument_count (args
); i
++)
3425 struct ipa_jump_func
*jfunc
= ipa_get_ith_jump_func (args
, i
);
3426 if (jfunc
->type
== IPA_JF_PASS_THROUGH
3427 && (ipa_get_jf_pass_through_formal_id (jfunc
)
3428 < (int) inlined_es
->param
.length ()))
3430 int jf_formal_id
= ipa_get_jf_pass_through_formal_id (jfunc
);
3431 int prob1
= es
->param
[i
].change_prob
;
3432 int prob2
= inlined_es
->param
[jf_formal_id
].change_prob
;
3433 int prob
= combine_probabilities (prob1
, prob2
);
3435 if (prob1
&& prob2
&& !prob
)
3438 es
->param
[i
].change_prob
= prob
;
3444 /* Update edge summaries of NODE after INLINED_EDGE has been inlined.
3446 Remap predicates of callees of NODE. Rest of arguments match
3449 Also update change probabilities. */
3452 remap_edge_summaries (struct cgraph_edge
*inlined_edge
,
3453 struct cgraph_node
*node
,
3454 struct inline_summary
*info
,
3455 struct inline_summary
*callee_info
,
3456 vec
<int> operand_map
,
3457 vec
<int> offset_map
,
3458 clause_t possible_truths
,
3459 struct predicate
*toplev_predicate
)
3461 struct cgraph_edge
*e
, *next
;
3462 for (e
= node
->callees
; e
; e
= next
)
3464 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3466 next
= e
->next_callee
;
3468 if (e
->inline_failed
)
3470 remap_edge_change_prob (inlined_edge
, e
);
3474 p
= remap_predicate (info
, callee_info
,
3475 es
->predicate
, operand_map
, offset_map
,
3476 possible_truths
, toplev_predicate
);
3477 edge_set_predicate (e
, &p
);
3480 edge_set_predicate (e
, toplev_predicate
);
3483 remap_edge_summaries (inlined_edge
, e
->callee
, info
, callee_info
,
3484 operand_map
, offset_map
, possible_truths
,
3487 for (e
= node
->indirect_calls
; e
; e
= next
)
3489 struct inline_edge_summary
*es
= inline_edge_summary (e
);
3491 next
= e
->next_callee
;
3493 remap_edge_change_prob (inlined_edge
, e
);
3496 p
= remap_predicate (info
, callee_info
,
3497 es
->predicate
, operand_map
, offset_map
,
3498 possible_truths
, toplev_predicate
);
3499 edge_set_predicate (e
, &p
);
3502 edge_set_predicate (e
, toplev_predicate
);
3506 /* Same as remap_predicate, but set result into hint *HINT. */
3509 remap_hint_predicate (struct inline_summary
*info
,
3510 struct inline_summary
*callee_info
,
3511 struct predicate
**hint
,
3512 vec
<int> operand_map
,
3513 vec
<int> offset_map
,
3514 clause_t possible_truths
,
3515 struct predicate
*toplev_predicate
)
3521 p
= remap_predicate (info
, callee_info
,
3523 operand_map
, offset_map
,
3524 possible_truths
, toplev_predicate
);
3525 if (!false_predicate_p (&p
) && !true_predicate_p (&p
))
3528 set_hint_predicate (hint
, p
);
3530 **hint
= and_predicates (info
->conds
, *hint
, &p
);
3534 /* We inlined EDGE. Update summary of the function we inlined into. */
3537 inline_merge_summary (struct cgraph_edge
*edge
)
3539 struct inline_summary
*callee_info
= inline_summaries
->get (edge
->callee
);
3540 struct cgraph_node
*to
= (edge
->caller
->global
.inlined_to
3541 ? edge
->caller
->global
.inlined_to
: edge
->caller
);
3542 struct inline_summary
*info
= inline_summaries
->get (to
);
3543 clause_t clause
= 0; /* not_inline is known to be false. */
3545 vec
<int> operand_map
= vNULL
;
3546 vec
<int> offset_map
= vNULL
;
3548 struct predicate toplev_predicate
;
3549 struct predicate true_p
= true_predicate ();
3550 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3553 toplev_predicate
= *es
->predicate
;
3555 toplev_predicate
= true_predicate ();
3557 if (callee_info
->conds
)
3558 evaluate_properties_for_edge (edge
, true, &clause
, NULL
, NULL
, NULL
);
3559 if (ipa_node_params_sum
&& callee_info
->conds
)
3561 struct ipa_edge_args
*args
= IPA_EDGE_REF (edge
);
3562 int count
= ipa_get_cs_argument_count (args
);
3567 operand_map
.safe_grow_cleared (count
);
3568 offset_map
.safe_grow_cleared (count
);
3570 for (i
= 0; i
< count
; i
++)
3572 struct ipa_jump_func
*jfunc
= ipa_get_ith_jump_func (args
, i
);
3575 /* TODO: handle non-NOPs when merging. */
3576 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
3578 if (ipa_get_jf_pass_through_operation (jfunc
) == NOP_EXPR
)
3579 map
= ipa_get_jf_pass_through_formal_id (jfunc
);
3580 if (!ipa_get_jf_pass_through_agg_preserved (jfunc
))
3583 else if (jfunc
->type
== IPA_JF_ANCESTOR
)
3585 HOST_WIDE_INT offset
= ipa_get_jf_ancestor_offset (jfunc
);
3586 if (offset
>= 0 && offset
< INT_MAX
)
3588 map
= ipa_get_jf_ancestor_formal_id (jfunc
);
3589 if (!ipa_get_jf_ancestor_agg_preserved (jfunc
))
3591 offset_map
[i
] = offset
;
3594 operand_map
[i
] = map
;
3595 gcc_assert (map
< ipa_get_param_count (IPA_NODE_REF (to
)));
3598 for (i
= 0; vec_safe_iterate (callee_info
->entry
, i
, &e
); i
++)
3600 struct predicate p
= remap_predicate (info
, callee_info
,
3601 &e
->predicate
, operand_map
,
3604 if (!false_predicate_p (&p
))
3606 gcov_type add_time
= ((gcov_type
) e
->time
* edge
->frequency
3607 + CGRAPH_FREQ_BASE
/ 2) / CGRAPH_FREQ_BASE
;
3608 int prob
= predicate_probability (callee_info
->conds
,
3611 add_time
= apply_probability ((gcov_type
) add_time
, prob
);
3612 if (add_time
> MAX_TIME
* INLINE_TIME_SCALE
)
3613 add_time
= MAX_TIME
* INLINE_TIME_SCALE
;
3614 if (prob
!= REG_BR_PROB_BASE
3615 && dump_file
&& (dump_flags
& TDF_DETAILS
))
3617 fprintf (dump_file
, "\t\tScaling time by probability:%f\n",
3618 (double) prob
/ REG_BR_PROB_BASE
);
3620 account_size_time (info
, e
->size
, add_time
, &p
);
3623 remap_edge_summaries (edge
, edge
->callee
, info
, callee_info
, operand_map
,
3624 offset_map
, clause
, &toplev_predicate
);
3625 remap_hint_predicate (info
, callee_info
,
3626 &callee_info
->loop_iterations
,
3627 operand_map
, offset_map
, clause
, &toplev_predicate
);
3628 remap_hint_predicate (info
, callee_info
,
3629 &callee_info
->loop_stride
,
3630 operand_map
, offset_map
, clause
, &toplev_predicate
);
3631 remap_hint_predicate (info
, callee_info
,
3632 &callee_info
->array_index
,
3633 operand_map
, offset_map
, clause
, &toplev_predicate
);
3635 inline_update_callee_summaries (edge
->callee
,
3636 inline_edge_summary (edge
)->loop_depth
);
3638 /* We do not maintain predicates of inlined edges, free it. */
3639 edge_set_predicate (edge
, &true_p
);
3640 /* Similarly remove param summaries. */
3641 es
->param
.release ();
3642 operand_map
.release ();
3643 offset_map
.release ();
3646 /* For performance reasons inline_merge_summary is not updating overall size
3647 and time. Recompute it. */
3650 inline_update_overall_summary (struct cgraph_node
*node
)
3652 struct inline_summary
*info
= inline_summaries
->get (node
);
3658 for (i
= 0; vec_safe_iterate (info
->entry
, i
, &e
); i
++)
3660 info
->size
+= e
->size
, info
->time
+= e
->time
;
3661 if (info
->time
> MAX_TIME
* INLINE_TIME_SCALE
)
3662 info
->time
= MAX_TIME
* INLINE_TIME_SCALE
;
3664 estimate_calls_size_and_time (node
, &info
->size
, &info
->min_size
,
3666 ~(clause_t
) (1 << predicate_false_condition
),
3667 vNULL
, vNULL
, vNULL
);
3668 info
->time
= (info
->time
+ INLINE_TIME_SCALE
/ 2) / INLINE_TIME_SCALE
;
3669 info
->size
= (info
->size
+ INLINE_SIZE_SCALE
/ 2) / INLINE_SIZE_SCALE
;
3672 /* Return hints derrived from EDGE. */
3674 simple_edge_hints (struct cgraph_edge
*edge
)
3677 struct cgraph_node
*to
= (edge
->caller
->global
.inlined_to
3678 ? edge
->caller
->global
.inlined_to
: edge
->caller
);
3679 struct cgraph_node
*callee
= edge
->callee
->ultimate_alias_target ();
3680 if (inline_summaries
->get (to
)->scc_no
3681 && inline_summaries
->get (to
)->scc_no
3682 == inline_summaries
->get (callee
)->scc_no
3683 && !edge
->recursive_p ())
3684 hints
|= INLINE_HINT_same_scc
;
3686 if (callee
->lto_file_data
&& edge
->caller
->lto_file_data
3687 && edge
->caller
->lto_file_data
!= callee
->lto_file_data
3689 hints
|= INLINE_HINT_cross_module
;
3694 /* Estimate the time cost for the caller when inlining EDGE.
3695 Only to be called via estimate_edge_time, that handles the
3698 When caching, also update the cache entry. Compute both time and
3699 size, since we always need both metrics eventually. */
3702 do_estimate_edge_time (struct cgraph_edge
*edge
)
3707 struct cgraph_node
*callee
;
3709 vec
<tree
> known_vals
;
3710 vec
<ipa_polymorphic_call_context
> known_contexts
;
3711 vec
<ipa_agg_jump_function_p
> known_aggs
;
3712 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3715 callee
= edge
->callee
->ultimate_alias_target ();
3717 gcc_checking_assert (edge
->inline_failed
);
3718 evaluate_properties_for_edge (edge
, true,
3719 &clause
, &known_vals
, &known_contexts
,
3721 estimate_node_size_and_time (callee
, clause
, known_vals
, known_contexts
,
3722 known_aggs
, &size
, &min_size
, &time
, &hints
, es
->param
);
3724 /* When we have profile feedback, we can quite safely identify hot
3725 edges and for those we disable size limits. Don't do that when
3726 probability that caller will call the callee is low however, since it
3727 may hurt optimization of the caller's hot path. */
3728 if (edge
->count
&& edge
->maybe_hot_p ()
3730 > (edge
->caller
->global
.inlined_to
3731 ? edge
->caller
->global
.inlined_to
->count
: edge
->caller
->count
)))
3732 hints
|= INLINE_HINT_known_hot
;
3734 known_vals
.release ();
3735 known_contexts
.release ();
3736 known_aggs
.release ();
3737 gcc_checking_assert (size
>= 0);
3738 gcc_checking_assert (time
>= 0);
3740 /* When caching, update the cache entry. */
3741 if (edge_growth_cache
.exists ())
3743 inline_summaries
->get (edge
->callee
)->min_size
= min_size
;
3744 if ((int) edge_growth_cache
.length () <= edge
->uid
)
3745 edge_growth_cache
.safe_grow_cleared (symtab
->edges_max_uid
);
3746 edge_growth_cache
[edge
->uid
].time
= time
+ (time
>= 0);
3748 edge_growth_cache
[edge
->uid
].size
= size
+ (size
>= 0);
3749 hints
|= simple_edge_hints (edge
);
3750 edge_growth_cache
[edge
->uid
].hints
= hints
+ 1;
3756 /* Return estimated callee growth after inlining EDGE.
3757 Only to be called via estimate_edge_size. */
3760 do_estimate_edge_size (struct cgraph_edge
*edge
)
3763 struct cgraph_node
*callee
;
3765 vec
<tree
> known_vals
;
3766 vec
<ipa_polymorphic_call_context
> known_contexts
;
3767 vec
<ipa_agg_jump_function_p
> known_aggs
;
3769 /* When we do caching, use do_estimate_edge_time to populate the entry. */
3771 if (edge_growth_cache
.exists ())
3773 do_estimate_edge_time (edge
);
3774 size
= edge_growth_cache
[edge
->uid
].size
;
3775 gcc_checking_assert (size
);
3776 return size
- (size
> 0);
3779 callee
= edge
->callee
->ultimate_alias_target ();
3781 /* Early inliner runs without caching, go ahead and do the dirty work. */
3782 gcc_checking_assert (edge
->inline_failed
);
3783 evaluate_properties_for_edge (edge
, true,
3784 &clause
, &known_vals
, &known_contexts
,
3786 estimate_node_size_and_time (callee
, clause
, known_vals
, known_contexts
,
3787 known_aggs
, &size
, NULL
, NULL
, NULL
, vNULL
);
3788 known_vals
.release ();
3789 known_contexts
.release ();
3790 known_aggs
.release ();
3795 /* Estimate the growth of the caller when inlining EDGE.
3796 Only to be called via estimate_edge_size. */
3799 do_estimate_edge_hints (struct cgraph_edge
*edge
)
3802 struct cgraph_node
*callee
;
3804 vec
<tree
> known_vals
;
3805 vec
<ipa_polymorphic_call_context
> known_contexts
;
3806 vec
<ipa_agg_jump_function_p
> known_aggs
;
3808 /* When we do caching, use do_estimate_edge_time to populate the entry. */
3810 if (edge_growth_cache
.exists ())
3812 do_estimate_edge_time (edge
);
3813 hints
= edge_growth_cache
[edge
->uid
].hints
;
3814 gcc_checking_assert (hints
);
3818 callee
= edge
->callee
->ultimate_alias_target ();
3820 /* Early inliner runs without caching, go ahead and do the dirty work. */
3821 gcc_checking_assert (edge
->inline_failed
);
3822 evaluate_properties_for_edge (edge
, true,
3823 &clause
, &known_vals
, &known_contexts
,
3825 estimate_node_size_and_time (callee
, clause
, known_vals
, known_contexts
,
3826 known_aggs
, NULL
, NULL
, NULL
, &hints
, vNULL
);
3827 known_vals
.release ();
3828 known_contexts
.release ();
3829 known_aggs
.release ();
3830 hints
|= simple_edge_hints (edge
);
3835 /* Estimate self time of the function NODE after inlining EDGE. */
3838 estimate_time_after_inlining (struct cgraph_node
*node
,
3839 struct cgraph_edge
*edge
)
3841 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3842 if (!es
->predicate
|| !false_predicate_p (es
->predicate
))
3845 inline_summaries
->get (node
)->time
+ estimate_edge_time (edge
);
3848 if (time
> MAX_TIME
)
3852 return inline_summaries
->get (node
)->time
;
3856 /* Estimate the size of NODE after inlining EDGE which should be an
3857 edge to either NODE or a call inlined into NODE. */
3860 estimate_size_after_inlining (struct cgraph_node
*node
,
3861 struct cgraph_edge
*edge
)
3863 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
3864 if (!es
->predicate
|| !false_predicate_p (es
->predicate
))
3866 int size
= inline_summaries
->get (node
)->size
+ estimate_edge_growth (edge
);
3867 gcc_assert (size
>= 0);
3870 return inline_summaries
->get (node
)->size
;
3876 struct cgraph_node
*node
;
3877 bool self_recursive
;
3883 /* Worker for do_estimate_growth. Collect growth for all callers. */
3886 do_estimate_growth_1 (struct cgraph_node
*node
, void *data
)
3888 struct cgraph_edge
*e
;
3889 struct growth_data
*d
= (struct growth_data
*) data
;
3891 for (e
= node
->callers
; e
; e
= e
->next_caller
)
3893 gcc_checking_assert (e
->inline_failed
);
3895 if (cgraph_inline_failed_type (e
->inline_failed
) == CIF_FINAL_ERROR
)
3897 d
->uninlinable
= true;
3901 if (e
->recursive_p ())
3903 d
->self_recursive
= true;
3906 d
->growth
+= estimate_edge_growth (e
);
3912 /* Estimate the growth caused by inlining NODE into all callees. */
3915 estimate_growth (struct cgraph_node
*node
)
3917 struct growth_data d
= { node
, false, false, 0 };
3918 struct inline_summary
*info
= inline_summaries
->get (node
);
3920 node
->call_for_symbol_and_aliases (do_estimate_growth_1
, &d
, true);
3922 /* For self recursive functions the growth estimation really should be
3923 infinity. We don't want to return very large values because the growth
3924 plays various roles in badness computation fractions. Be sure to not
3925 return zero or negative growths. */
3926 if (d
.self_recursive
)
3927 d
.growth
= d
.growth
< info
->size
? info
->size
: d
.growth
;
3928 else if (DECL_EXTERNAL (node
->decl
) || d
.uninlinable
)
3932 if (node
->will_be_removed_from_program_if_no_direct_calls_p ())
3933 d
.growth
-= info
->size
;
3934 /* COMDAT functions are very often not shared across multiple units
3935 since they come from various template instantiations.
3936 Take this into account. */
3937 else if (DECL_COMDAT (node
->decl
)
3938 && node
->can_remove_if_no_direct_calls_p ())
3939 d
.growth
-= (info
->size
3940 * (100 - PARAM_VALUE (PARAM_COMDAT_SHARING_PROBABILITY
))
3947 /* Verify if there are fewer than MAX_CALLERS. */
3950 check_callers (cgraph_node
*node
, int *max_callers
)
3954 if (!node
->can_remove_if_no_direct_calls_and_refs_p ())
3957 for (cgraph_edge
*e
= node
->callers
; e
; e
= e
->next_caller
)
3961 || cgraph_inline_failed_type (e
->inline_failed
) == CIF_FINAL_ERROR
)
3965 FOR_EACH_ALIAS (node
, ref
)
3966 if (check_callers (dyn_cast
<cgraph_node
*> (ref
->referring
), max_callers
))
3973 /* Make cheap estimation if growth of NODE is likely positive knowing
3974 EDGE_GROWTH of one particular edge.
3975 We assume that most of other edges will have similar growth
3976 and skip computation if there are too many callers. */
3979 growth_likely_positive (struct cgraph_node
*node
,
3983 struct cgraph_edge
*e
;
3984 gcc_checking_assert (edge_growth
> 0);
3986 /* First quickly check if NODE is removable at all. */
3987 if (DECL_EXTERNAL (node
->decl
))
3989 if (!node
->can_remove_if_no_direct_calls_and_refs_p ()
3990 || node
->address_taken
)
3993 max_callers
= inline_summaries
->get (node
)->size
* 4 / edge_growth
+ 2;
3995 for (e
= node
->callers
; e
; e
= e
->next_caller
)
3999 || cgraph_inline_failed_type (e
->inline_failed
) == CIF_FINAL_ERROR
)
4004 FOR_EACH_ALIAS (node
, ref
)
4005 if (check_callers (dyn_cast
<cgraph_node
*> (ref
->referring
), &max_callers
))
4008 /* Unlike for functions called once, we play unsafe with
4009 COMDATs. We can allow that since we know functions
4010 in consideration are small (and thus risk is small) and
4011 moreover grow estimates already accounts that COMDAT
4012 functions may or may not disappear when eliminated from
4013 current unit. With good probability making aggressive
4014 choice in all units is going to make overall program
4016 if (DECL_COMDAT (node
->decl
))
4018 if (!node
->can_remove_if_no_direct_calls_p ())
4021 else if (!node
->will_be_removed_from_program_if_no_direct_calls_p ())
4024 return estimate_growth (node
) > 0;
4028 /* This function performs intraprocedural analysis in NODE that is required to
4029 inline indirect calls. */
4032 inline_indirect_intraprocedural_analysis (struct cgraph_node
*node
)
4034 ipa_analyze_node (node
);
4035 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4037 ipa_print_node_params (dump_file
, node
);
4038 ipa_print_node_jump_functions (dump_file
, node
);
4043 /* Note function body size. */
4046 inline_analyze_function (struct cgraph_node
*node
)
4048 push_cfun (DECL_STRUCT_FUNCTION (node
->decl
));
4051 fprintf (dump_file
, "\nAnalyzing function: %s/%u\n",
4052 node
->name (), node
->order
);
4053 if (opt_for_fn (node
->decl
, optimize
) && !node
->thunk
.thunk_p
)
4054 inline_indirect_intraprocedural_analysis (node
);
4055 compute_inline_parameters (node
, false);
4058 struct cgraph_edge
*e
;
4059 for (e
= node
->callees
; e
; e
= e
->next_callee
)
4061 if (e
->inline_failed
== CIF_FUNCTION_NOT_CONSIDERED
)
4062 e
->inline_failed
= CIF_FUNCTION_NOT_OPTIMIZED
;
4063 e
->call_stmt_cannot_inline_p
= true;
4065 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
4067 if (e
->inline_failed
== CIF_FUNCTION_NOT_CONSIDERED
)
4068 e
->inline_failed
= CIF_FUNCTION_NOT_OPTIMIZED
;
4069 e
->call_stmt_cannot_inline_p
= true;
4077 /* Called when new function is inserted to callgraph late. */
4080 inline_summary_t::insert (struct cgraph_node
*node
, inline_summary
*)
4082 inline_analyze_function (node
);
4085 /* Note function body size. */
4088 inline_generate_summary (void)
4090 struct cgraph_node
*node
;
4092 /* When not optimizing, do not bother to analyze. Inlining is still done
4093 because edge redirection needs to happen there. */
4094 if (!optimize
&& !flag_generate_lto
&& !flag_generate_offload
&& !flag_wpa
)
4097 if (!inline_summaries
)
4098 inline_summaries
= (inline_summary_t
*) inline_summary_t::create_ggc (symtab
);
4100 inline_summaries
->enable_insertion_hook ();
4102 ipa_register_cgraph_hooks ();
4103 inline_free_summary ();
4105 FOR_EACH_DEFINED_FUNCTION (node
)
4107 inline_analyze_function (node
);
4111 /* Read predicate from IB. */
4113 static struct predicate
4114 read_predicate (struct lto_input_block
*ib
)
4116 struct predicate out
;
4122 gcc_assert (k
<= MAX_CLAUSES
);
4123 clause
= out
.clause
[k
++] = streamer_read_uhwi (ib
);
4127 /* Zero-initialize the remaining clauses in OUT. */
4128 while (k
<= MAX_CLAUSES
)
4129 out
.clause
[k
++] = 0;
4135 /* Write inline summary for edge E to OB. */
4138 read_inline_edge_summary (struct lto_input_block
*ib
, struct cgraph_edge
*e
)
4140 struct inline_edge_summary
*es
= inline_edge_summary (e
);
4144 es
->call_stmt_size
= streamer_read_uhwi (ib
);
4145 es
->call_stmt_time
= streamer_read_uhwi (ib
);
4146 es
->loop_depth
= streamer_read_uhwi (ib
);
4147 p
= read_predicate (ib
);
4148 edge_set_predicate (e
, &p
);
4149 length
= streamer_read_uhwi (ib
);
4152 es
->param
.safe_grow_cleared (length
);
4153 for (i
= 0; i
< length
; i
++)
4154 es
->param
[i
].change_prob
= streamer_read_uhwi (ib
);
4159 /* Stream in inline summaries from the section. */
4162 inline_read_section (struct lto_file_decl_data
*file_data
, const char *data
,
4165 const struct lto_function_header
*header
=
4166 (const struct lto_function_header
*) data
;
4167 const int cfg_offset
= sizeof (struct lto_function_header
);
4168 const int main_offset
= cfg_offset
+ header
->cfg_size
;
4169 const int string_offset
= main_offset
+ header
->main_size
;
4170 struct data_in
*data_in
;
4171 unsigned int i
, count2
, j
;
4172 unsigned int f_count
;
4174 lto_input_block
ib ((const char *) data
+ main_offset
, header
->main_size
,
4175 file_data
->mode_table
);
4178 lto_data_in_create (file_data
, (const char *) data
+ string_offset
,
4179 header
->string_size
, vNULL
);
4180 f_count
= streamer_read_uhwi (&ib
);
4181 for (i
= 0; i
< f_count
; i
++)
4184 struct cgraph_node
*node
;
4185 struct inline_summary
*info
;
4186 lto_symtab_encoder_t encoder
;
4187 struct bitpack_d bp
;
4188 struct cgraph_edge
*e
;
4191 index
= streamer_read_uhwi (&ib
);
4192 encoder
= file_data
->symtab_node_encoder
;
4193 node
= dyn_cast
<cgraph_node
*> (lto_symtab_encoder_deref (encoder
,
4195 info
= inline_summaries
->get (node
);
4197 info
->estimated_stack_size
4198 = info
->estimated_self_stack_size
= streamer_read_uhwi (&ib
);
4199 info
->size
= info
->self_size
= streamer_read_uhwi (&ib
);
4200 info
->time
= info
->self_time
= streamer_read_uhwi (&ib
);
4202 bp
= streamer_read_bitpack (&ib
);
4203 info
->inlinable
= bp_unpack_value (&bp
, 1);
4204 info
->contains_cilk_spawn
= bp_unpack_value (&bp
, 1);
4206 count2
= streamer_read_uhwi (&ib
);
4207 gcc_assert (!info
->conds
);
4208 for (j
= 0; j
< count2
; j
++)
4211 c
.operand_num
= streamer_read_uhwi (&ib
);
4212 c
.code
= (enum tree_code
) streamer_read_uhwi (&ib
);
4213 c
.val
= stream_read_tree (&ib
, data_in
);
4214 bp
= streamer_read_bitpack (&ib
);
4215 c
.agg_contents
= bp_unpack_value (&bp
, 1);
4216 c
.by_ref
= bp_unpack_value (&bp
, 1);
4218 c
.offset
= streamer_read_uhwi (&ib
);
4219 vec_safe_push (info
->conds
, c
);
4221 count2
= streamer_read_uhwi (&ib
);
4222 gcc_assert (!info
->entry
);
4223 for (j
= 0; j
< count2
; j
++)
4225 struct size_time_entry e
;
4227 e
.size
= streamer_read_uhwi (&ib
);
4228 e
.time
= streamer_read_uhwi (&ib
);
4229 e
.predicate
= read_predicate (&ib
);
4231 vec_safe_push (info
->entry
, e
);
4234 p
= read_predicate (&ib
);
4235 set_hint_predicate (&info
->loop_iterations
, p
);
4236 p
= read_predicate (&ib
);
4237 set_hint_predicate (&info
->loop_stride
, p
);
4238 p
= read_predicate (&ib
);
4239 set_hint_predicate (&info
->array_index
, p
);
4240 for (e
= node
->callees
; e
; e
= e
->next_callee
)
4241 read_inline_edge_summary (&ib
, e
);
4242 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
4243 read_inline_edge_summary (&ib
, e
);
4246 lto_free_section_data (file_data
, LTO_section_inline_summary
, NULL
, data
,
4248 lto_data_in_delete (data_in
);
4252 /* Read inline summary. Jump functions are shared among ipa-cp
4253 and inliner, so when ipa-cp is active, we don't need to write them
4257 inline_read_summary (void)
4259 struct lto_file_decl_data
**file_data_vec
= lto_get_file_decl_data ();
4260 struct lto_file_decl_data
*file_data
;
4263 inline_summary_alloc ();
4265 while ((file_data
= file_data_vec
[j
++]))
4268 const char *data
= lto_get_section_data (file_data
,
4269 LTO_section_inline_summary
,
4272 inline_read_section (file_data
, data
, len
);
4274 /* Fatal error here. We do not want to support compiling ltrans units
4275 with different version of compiler or different flags than the WPA
4276 unit, so this should never happen. */
4277 fatal_error (input_location
,
4278 "ipa inline summary is missing in input file");
4282 ipa_register_cgraph_hooks ();
4284 ipa_prop_read_jump_functions ();
4287 gcc_assert (inline_summaries
);
4288 inline_summaries
->enable_insertion_hook ();
4292 /* Write predicate P to OB. */
4295 write_predicate (struct output_block
*ob
, struct predicate
*p
)
4299 for (j
= 0; p
->clause
[j
]; j
++)
4301 gcc_assert (j
< MAX_CLAUSES
);
4302 streamer_write_uhwi (ob
, p
->clause
[j
]);
4304 streamer_write_uhwi (ob
, 0);
4308 /* Write inline summary for edge E to OB. */
4311 write_inline_edge_summary (struct output_block
*ob
, struct cgraph_edge
*e
)
4313 struct inline_edge_summary
*es
= inline_edge_summary (e
);
4316 streamer_write_uhwi (ob
, es
->call_stmt_size
);
4317 streamer_write_uhwi (ob
, es
->call_stmt_time
);
4318 streamer_write_uhwi (ob
, es
->loop_depth
);
4319 write_predicate (ob
, es
->predicate
);
4320 streamer_write_uhwi (ob
, es
->param
.length ());
4321 for (i
= 0; i
< (int) es
->param
.length (); i
++)
4322 streamer_write_uhwi (ob
, es
->param
[i
].change_prob
);
4326 /* Write inline summary for node in SET.
4327 Jump functions are shared among ipa-cp and inliner, so when ipa-cp is
4328 active, we don't need to write them twice. */
4331 inline_write_summary (void)
4333 struct cgraph_node
*node
;
4334 struct output_block
*ob
= create_output_block (LTO_section_inline_summary
);
4335 lto_symtab_encoder_t encoder
= ob
->decl_state
->symtab_node_encoder
;
4336 unsigned int count
= 0;
4339 for (i
= 0; i
< lto_symtab_encoder_size (encoder
); i
++)
4341 symtab_node
*snode
= lto_symtab_encoder_deref (encoder
, i
);
4342 cgraph_node
*cnode
= dyn_cast
<cgraph_node
*> (snode
);
4343 if (cnode
&& cnode
->definition
&& !cnode
->alias
)
4346 streamer_write_uhwi (ob
, count
);
4348 for (i
= 0; i
< lto_symtab_encoder_size (encoder
); i
++)
4350 symtab_node
*snode
= lto_symtab_encoder_deref (encoder
, i
);
4351 cgraph_node
*cnode
= dyn_cast
<cgraph_node
*> (snode
);
4352 if (cnode
&& (node
= cnode
)->definition
&& !node
->alias
)
4354 struct inline_summary
*info
= inline_summaries
->get (node
);
4355 struct bitpack_d bp
;
4356 struct cgraph_edge
*edge
;
4359 struct condition
*c
;
4361 streamer_write_uhwi (ob
,
4362 lto_symtab_encoder_encode (encoder
,
4365 streamer_write_hwi (ob
, info
->estimated_self_stack_size
);
4366 streamer_write_hwi (ob
, info
->self_size
);
4367 streamer_write_hwi (ob
, info
->self_time
);
4368 bp
= bitpack_create (ob
->main_stream
);
4369 bp_pack_value (&bp
, info
->inlinable
, 1);
4370 bp_pack_value (&bp
, info
->contains_cilk_spawn
, 1);
4371 streamer_write_bitpack (&bp
);
4372 streamer_write_uhwi (ob
, vec_safe_length (info
->conds
));
4373 for (i
= 0; vec_safe_iterate (info
->conds
, i
, &c
); i
++)
4375 streamer_write_uhwi (ob
, c
->operand_num
);
4376 streamer_write_uhwi (ob
, c
->code
);
4377 stream_write_tree (ob
, c
->val
, true);
4378 bp
= bitpack_create (ob
->main_stream
);
4379 bp_pack_value (&bp
, c
->agg_contents
, 1);
4380 bp_pack_value (&bp
, c
->by_ref
, 1);
4381 streamer_write_bitpack (&bp
);
4382 if (c
->agg_contents
)
4383 streamer_write_uhwi (ob
, c
->offset
);
4385 streamer_write_uhwi (ob
, vec_safe_length (info
->entry
));
4386 for (i
= 0; vec_safe_iterate (info
->entry
, i
, &e
); i
++)
4388 streamer_write_uhwi (ob
, e
->size
);
4389 streamer_write_uhwi (ob
, e
->time
);
4390 write_predicate (ob
, &e
->predicate
);
4392 write_predicate (ob
, info
->loop_iterations
);
4393 write_predicate (ob
, info
->loop_stride
);
4394 write_predicate (ob
, info
->array_index
);
4395 for (edge
= node
->callees
; edge
; edge
= edge
->next_callee
)
4396 write_inline_edge_summary (ob
, edge
);
4397 for (edge
= node
->indirect_calls
; edge
; edge
= edge
->next_callee
)
4398 write_inline_edge_summary (ob
, edge
);
4401 streamer_write_char_stream (ob
->main_stream
, 0);
4402 produce_asm (ob
, NULL
);
4403 destroy_output_block (ob
);
4405 if (optimize
&& !flag_ipa_cp
)
4406 ipa_prop_write_jump_functions ();
4410 /* Release inline summary. */
4413 inline_free_summary (void)
4415 struct cgraph_node
*node
;
4416 if (edge_removal_hook_holder
)
4417 symtab
->remove_edge_removal_hook (edge_removal_hook_holder
);
4418 edge_removal_hook_holder
= NULL
;
4419 if (edge_duplication_hook_holder
)
4420 symtab
->remove_edge_duplication_hook (edge_duplication_hook_holder
);
4421 edge_duplication_hook_holder
= NULL
;
4422 if (!inline_edge_summary_vec
.exists ())
4424 FOR_EACH_DEFINED_FUNCTION (node
)
4426 reset_inline_summary (node
, inline_summaries
->get (node
));
4427 inline_summaries
->release ();
4428 inline_summaries
= NULL
;
4429 inline_edge_summary_vec
.release ();
4430 edge_predicate_pool
.release ();