1 /* Inlining decision heuristics.
2 Copyright (C) 2003, 2004, 2007, 2008, 2009, 2010, 2011
3 Free Software Foundation, Inc.
4 Contributed by Jan Hubicka
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 3, or (at your option) any later
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
22 /* Inlining decision heuristics
24 The implementation of inliner is organized as follows:
26 inlining heuristics limits
28 can_inline_edge_p allow to check that particular inlining is allowed
29 by the limits specified by user (allowed function growth, growth and so
32 Functions are inlined when it is obvious the result is profitable (such
33 as functions called once or when inlining reduce code size).
34 In addition to that we perform inlining of small functions and recursive
39 The inliner itself is split into two passes:
43 Simple local inlining pass inlining callees into current function.
44 This pass makes no use of whole unit analysis and thus it can do only
45 very simple decisions based on local properties.
47 The strength of the pass is that it is run in topological order
48 (reverse postorder) on the callgraph. Functions are converted into SSA
49 form just before this pass and optimized subsequently. As a result, the
50 callees of the function seen by the early inliner was already optimized
51 and results of early inlining adds a lot of optimization opportunities
52 for the local optimization.
54 The pass handle the obvious inlining decisions within the compilation
55 unit - inlining auto inline functions, inlining for size and
58 main strength of the pass is the ability to eliminate abstraction
59 penalty in C++ code (via combination of inlining and early
60 optimization) and thus improve quality of analysis done by real IPA
63 Because of lack of whole unit knowledge, the pass can not really make
64 good code size/performance tradeoffs. It however does very simple
65 speculative inlining allowing code size to grow by
66 EARLY_INLINING_INSNS when callee is leaf function. In this case the
67 optimizations performed later are very likely to eliminate the cost.
71 This is the real inliner able to handle inlining with whole program
72 knowledge. It performs following steps:
74 1) inlining of small functions. This is implemented by greedy
75 algorithm ordering all inlinable cgraph edges by their badness and
76 inlining them in this order as long as inline limits allows doing so.
78 This heuristics is not very good on inlining recursive calls. Recursive
79 calls can be inlined with results similar to loop unrolling. To do so,
80 special purpose recursive inliner is executed on function when
81 recursive edge is met as viable candidate.
83 2) Unreachable functions are removed from callgraph. Inlining leads
84 to devirtualization and other modification of callgraph so functions
85 may become unreachable during the process. Also functions declared as
86 extern inline or virtual functions are removed, since after inlining
87 we no longer need the offline bodies.
89 3) Functions called once and not exported from the unit are inlined.
90 This should almost always lead to reduction of code size by eliminating
91 the need for offline copy of the function. */
95 #include "coretypes.h"
98 #include "tree-inline.h"
99 #include "langhooks.h"
102 #include "diagnostic.h"
103 #include "gimple-pretty-print.h"
108 #include "tree-pass.h"
109 #include "coverage.h"
112 #include "tree-flow.h"
113 #include "ipa-prop.h"
116 #include "ipa-inline.h"
117 #include "ipa-utils.h"
119 /* Statistics we collect about inlining algorithm. */
120 static int overall_size
;
121 static gcov_type max_count
;
123 /* Return false when inlining edge E would lead to violating
124 limits on function unit growth or stack usage growth.
126 The relative function body growth limit is present generally
127 to avoid problems with non-linear behavior of the compiler.
128 To allow inlining huge functions into tiny wrapper, the limit
129 is always based on the bigger of the two functions considered.
131 For stack growth limits we always base the growth in stack usage
132 of the callers. We want to prevent applications from segfaulting
133 on stack overflow when functions with huge stack frames gets
137 caller_growth_limits (struct cgraph_edge
*e
)
139 struct cgraph_node
*to
= e
->caller
;
140 struct cgraph_node
*what
= e
->callee
;
143 HOST_WIDE_INT stack_size_limit
= 0, inlined_stack
;
144 struct inline_summary
*info
, *what_info
, *outer_info
= inline_summary (to
);
146 /* Look for function e->caller is inlined to. While doing
147 so work out the largest function body on the way. As
148 described above, we want to base our function growth
149 limits based on that. Not on the self size of the
150 outer function, not on the self size of inline code
151 we immediately inline to. This is the most relaxed
152 interpretation of the rule "do not grow large functions
153 too much in order to prevent compiler from exploding". */
156 info
= inline_summary (to
);
157 if (limit
< info
->self_size
)
158 limit
= info
->self_size
;
159 if (stack_size_limit
< info
->estimated_self_stack_size
)
160 stack_size_limit
= info
->estimated_self_stack_size
;
161 if (to
->global
.inlined_to
)
162 to
= to
->callers
->caller
;
167 what_info
= inline_summary (what
);
169 if (limit
< what_info
->self_size
)
170 limit
= what_info
->self_size
;
172 limit
+= limit
* PARAM_VALUE (PARAM_LARGE_FUNCTION_GROWTH
) / 100;
174 /* Check the size after inlining against the function limits. But allow
175 the function to shrink if it went over the limits by forced inlining. */
176 newsize
= estimate_size_after_inlining (to
, e
);
177 if (newsize
>= info
->size
178 && newsize
> PARAM_VALUE (PARAM_LARGE_FUNCTION_INSNS
)
181 e
->inline_failed
= CIF_LARGE_FUNCTION_GROWTH_LIMIT
;
185 if (!what_info
->estimated_stack_size
)
188 /* FIXME: Stack size limit often prevents inlining in Fortran programs
189 due to large i/o datastructures used by the Fortran front-end.
190 We ought to ignore this limit when we know that the edge is executed
191 on every invocation of the caller (i.e. its call statement dominates
192 exit block). We do not track this information, yet. */
193 stack_size_limit
+= ((gcov_type
)stack_size_limit
194 * PARAM_VALUE (PARAM_STACK_FRAME_GROWTH
) / 100);
196 inlined_stack
= (outer_info
->stack_frame_offset
197 + outer_info
->estimated_self_stack_size
198 + what_info
->estimated_stack_size
);
199 /* Check new stack consumption with stack consumption at the place
201 if (inlined_stack
> stack_size_limit
202 /* If function already has large stack usage from sibling
203 inline call, we can inline, too.
204 This bit overoptimistically assume that we are good at stack
206 && inlined_stack
> info
->estimated_stack_size
207 && inlined_stack
> PARAM_VALUE (PARAM_LARGE_STACK_FRAME
))
209 e
->inline_failed
= CIF_LARGE_STACK_FRAME_GROWTH_LIMIT
;
215 /* Dump info about why inlining has failed. */
218 report_inline_failed_reason (struct cgraph_edge
*e
)
222 fprintf (dump_file
, " not inlinable: %s/%i -> %s/%i, %s\n",
223 cgraph_node_name (e
->caller
), e
->caller
->uid
,
224 cgraph_node_name (e
->callee
), e
->callee
->uid
,
225 cgraph_inline_failed_string (e
->inline_failed
));
229 /* Decide if we can inline the edge and possibly update
230 inline_failed reason.
231 We check whether inlining is possible at all and whether
232 caller growth limits allow doing so.
234 if REPORT is true, output reason to the dump file. */
237 can_inline_edge_p (struct cgraph_edge
*e
, bool report
)
239 bool inlinable
= true;
240 tree caller_tree
= DECL_FUNCTION_SPECIFIC_OPTIMIZATION (e
->caller
->decl
);
241 tree callee_tree
= DECL_FUNCTION_SPECIFIC_OPTIMIZATION (e
->callee
->decl
);
243 gcc_assert (e
->inline_failed
);
245 if (!e
->callee
->analyzed
)
247 e
->inline_failed
= CIF_BODY_NOT_AVAILABLE
;
250 else if (!inline_summary (e
->callee
)->inlinable
)
252 e
->inline_failed
= CIF_FUNCTION_NOT_INLINABLE
;
255 else if (cgraph_function_body_availability (e
->callee
) <= AVAIL_OVERWRITABLE
)
257 e
->inline_failed
= CIF_OVERWRITABLE
;
260 else if (e
->call_stmt_cannot_inline_p
)
262 e
->inline_failed
= CIF_MISMATCHED_ARGUMENTS
;
265 /* Don't inline if the functions have different EH personalities. */
266 else if (DECL_FUNCTION_PERSONALITY (e
->caller
->decl
)
267 && DECL_FUNCTION_PERSONALITY (e
->callee
->decl
)
268 && (DECL_FUNCTION_PERSONALITY (e
->caller
->decl
)
269 != DECL_FUNCTION_PERSONALITY (e
->callee
->decl
)))
271 e
->inline_failed
= CIF_EH_PERSONALITY
;
274 /* Don't inline if the callee can throw non-call exceptions but the
276 FIXME: this is obviously wrong for LTO where STRUCT_FUNCTION is missing.
277 Move the flag into cgraph node or mirror it in the inline summary. */
278 else if (DECL_STRUCT_FUNCTION (e
->callee
->decl
)
279 && DECL_STRUCT_FUNCTION
280 (e
->callee
->decl
)->can_throw_non_call_exceptions
281 && !(DECL_STRUCT_FUNCTION (e
->caller
->decl
)
282 && DECL_STRUCT_FUNCTION
283 (e
->caller
->decl
)->can_throw_non_call_exceptions
))
285 e
->inline_failed
= CIF_NON_CALL_EXCEPTIONS
;
288 /* Check compatibility of target optimization options. */
289 else if (!targetm
.target_option
.can_inline_p (e
->caller
->decl
,
292 e
->inline_failed
= CIF_TARGET_OPTION_MISMATCH
;
295 /* Check if caller growth allows the inlining. */
296 else if (!DECL_DISREGARD_INLINE_LIMITS (e
->callee
->decl
)
297 && !lookup_attribute ("flatten",
299 (e
->caller
->global
.inlined_to
300 ? e
->caller
->global
.inlined_to
->decl
302 && !caller_growth_limits (e
))
304 /* Don't inline a function with a higher optimization level than the
305 caller. FIXME: this is really just tip of iceberg of handling
306 optimization attribute. */
307 else if (caller_tree
!= callee_tree
)
309 struct cl_optimization
*caller_opt
310 = TREE_OPTIMIZATION ((caller_tree
)
312 : optimization_default_node
);
314 struct cl_optimization
*callee_opt
315 = TREE_OPTIMIZATION ((callee_tree
)
317 : optimization_default_node
);
319 if ((caller_opt
->x_optimize
> callee_opt
->x_optimize
)
320 || (caller_opt
->x_optimize_size
!= callee_opt
->x_optimize_size
))
322 e
->inline_failed
= CIF_TARGET_OPTIMIZATION_MISMATCH
;
327 /* Be sure that the cannot_inline_p flag is up to date. */
328 gcc_checking_assert (!e
->call_stmt
329 || (gimple_call_cannot_inline_p (e
->call_stmt
)
330 == e
->call_stmt_cannot_inline_p
)
331 /* In -flto-partition=none mode we really keep things out of
332 sync because call_stmt_cannot_inline_p is set at cgraph
333 merging when function bodies are not there yet. */
334 || (in_lto_p
&& !gimple_call_cannot_inline_p (e
->call_stmt
)));
335 if (!inlinable
&& report
)
336 report_inline_failed_reason (e
);
341 /* Return true if the edge E is inlinable during early inlining. */
344 can_early_inline_edge_p (struct cgraph_edge
*e
)
346 /* Early inliner might get called at WPA stage when IPA pass adds new
347 function. In this case we can not really do any of early inlining
348 because function bodies are missing. */
349 if (!gimple_has_body_p (e
->callee
->decl
))
351 e
->inline_failed
= CIF_BODY_NOT_AVAILABLE
;
354 /* In early inliner some of callees may not be in SSA form yet
355 (i.e. the callgraph is cyclic and we did not process
356 the callee by early inliner, yet). We don't have CIF code for this
357 case; later we will re-do the decision in the real inliner. */
358 if (!gimple_in_ssa_p (DECL_STRUCT_FUNCTION (e
->caller
->decl
))
359 || !gimple_in_ssa_p (DECL_STRUCT_FUNCTION (e
->callee
->decl
)))
362 fprintf (dump_file
, " edge not inlinable: not in SSA form\n");
365 if (!can_inline_edge_p (e
, true))
371 /* Return true when N is leaf function. Accept cheap builtins
372 in leaf functions. */
375 leaf_node_p (struct cgraph_node
*n
)
377 struct cgraph_edge
*e
;
378 for (e
= n
->callees
; e
; e
= e
->next_callee
)
379 if (!is_inexpensive_builtin (e
->callee
->decl
))
385 /* Return true if we are interested in inlining small function. */
388 want_early_inline_function_p (struct cgraph_edge
*e
)
390 bool want_inline
= true;
392 if (DECL_DISREGARD_INLINE_LIMITS (e
->callee
->decl
))
394 else if (!DECL_DECLARED_INLINE_P (e
->callee
->decl
)
395 && !flag_inline_small_functions
)
397 e
->inline_failed
= CIF_FUNCTION_NOT_INLINE_CANDIDATE
;
398 report_inline_failed_reason (e
);
403 int growth
= estimate_edge_growth (e
);
406 else if (!cgraph_maybe_hot_edge_p (e
)
410 fprintf (dump_file
, " will not early inline: %s/%i->%s/%i, "
411 "call is cold and code would grow by %i\n",
412 cgraph_node_name (e
->caller
), e
->caller
->uid
,
413 cgraph_node_name (e
->callee
), e
->callee
->uid
,
417 else if (!leaf_node_p (e
->callee
)
421 fprintf (dump_file
, " will not early inline: %s/%i->%s/%i, "
422 "callee is not leaf and code would grow by %i\n",
423 cgraph_node_name (e
->caller
), e
->caller
->uid
,
424 cgraph_node_name (e
->callee
), e
->callee
->uid
,
428 else if (growth
> PARAM_VALUE (PARAM_EARLY_INLINING_INSNS
))
431 fprintf (dump_file
, " will not early inline: %s/%i->%s/%i, "
432 "growth %i exceeds --param early-inlining-insns\n",
433 cgraph_node_name (e
->caller
), e
->caller
->uid
,
434 cgraph_node_name (e
->callee
), e
->callee
->uid
,
442 /* Return true if we are interested in inlining small function.
443 When REPORT is true, report reason to dump file. */
446 want_inline_small_function_p (struct cgraph_edge
*e
, bool report
)
448 bool want_inline
= true;
450 if (DECL_DISREGARD_INLINE_LIMITS (e
->callee
->decl
))
452 else if (!DECL_DECLARED_INLINE_P (e
->callee
->decl
)
453 && !flag_inline_small_functions
)
455 e
->inline_failed
= CIF_FUNCTION_NOT_INLINE_CANDIDATE
;
460 int growth
= estimate_edge_growth (e
);
464 else if (DECL_DECLARED_INLINE_P (e
->callee
->decl
)
465 && growth
>= MAX_INLINE_INSNS_SINGLE
)
467 e
->inline_failed
= CIF_MAX_INLINE_INSNS_SINGLE_LIMIT
;
470 else if (!DECL_DECLARED_INLINE_P (e
->callee
->decl
)
471 && !flag_inline_functions
)
473 e
->inline_failed
= CIF_NOT_DECLARED_INLINED
;
476 else if (!DECL_DECLARED_INLINE_P (e
->callee
->decl
)
477 && growth
>= MAX_INLINE_INSNS_AUTO
)
479 e
->inline_failed
= CIF_MAX_INLINE_INSNS_AUTO_LIMIT
;
482 /* If call is cold, do not inline when function body would grow.
483 Still inline when the overall unit size will shrink because the offline
484 copy of function being eliminated.
486 This is slightly wrong on aggressive side: it is entirely possible
487 that function is called many times with a context where inlining
488 reduces code size and few times with a context where inlining increase
489 code size. Resoluting growth estimate will be negative even if it
490 would make more sense to keep offline copy and do not inline into the
491 call sites that makes the code size grow.
493 When badness orders the calls in a way that code reducing calls come
494 first, this situation is not a problem at all: after inlining all
495 "good" calls, we will realize that keeping the function around is
497 else if (!cgraph_maybe_hot_edge_p (e
)
498 && (DECL_EXTERNAL (e
->callee
->decl
)
500 /* Unlike for functions called once, we play unsafe with
501 COMDATs. We can allow that since we know functions
502 in consideration are small (and thus risk is small) and
503 moreover grow estimates already accounts that COMDAT
504 functions may or may not disappear when eliminated from
505 current unit. With good probability making aggressive
506 choice in all units is going to make overall program
509 Consequently we ask cgraph_can_remove_if_no_direct_calls_p
511 cgraph_will_be_removed_from_program_if_no_direct_calls */
513 || !cgraph_can_remove_if_no_direct_calls_p (e
->callee
)
514 || estimate_growth (e
->callee
) > 0))
516 e
->inline_failed
= CIF_UNLIKELY_CALL
;
520 if (!want_inline
&& report
)
521 report_inline_failed_reason (e
);
525 /* EDGE is self recursive edge.
526 We hand two cases - when function A is inlining into itself
527 or when function A is being inlined into another inliner copy of function
530 In first case OUTER_NODE points to the toplevel copy of A, while
531 in the second case OUTER_NODE points to the outermost copy of A in B.
533 In both cases we want to be extra selective since
534 inlining the call will just introduce new recursive calls to appear. */
537 want_inline_self_recursive_call_p (struct cgraph_edge
*edge
,
538 struct cgraph_node
*outer_node
,
542 char const *reason
= NULL
;
543 bool want_inline
= true;
544 int caller_freq
= CGRAPH_FREQ_BASE
;
545 int max_depth
= PARAM_VALUE (PARAM_MAX_INLINE_RECURSIVE_DEPTH_AUTO
);
547 if (DECL_DECLARED_INLINE_P (edge
->callee
->decl
))
548 max_depth
= PARAM_VALUE (PARAM_MAX_INLINE_RECURSIVE_DEPTH
);
550 if (!cgraph_maybe_hot_edge_p (edge
))
552 reason
= "recursive call is cold";
555 else if (max_count
&& !outer_node
->count
)
557 reason
= "not executed in profile";
560 else if (depth
> max_depth
)
562 reason
= "--param max-inline-recursive-depth exceeded.";
566 if (outer_node
->global
.inlined_to
)
567 caller_freq
= outer_node
->callers
->frequency
;
571 /* Inlining of self recursive function into copy of itself within other function
572 is transformation similar to loop peeling.
574 Peeling is profitable if we can inline enough copies to make probability
575 of actual call to the self recursive function very small. Be sure that
576 the probability of recursion is small.
578 We ensure that the frequency of recursing is at most 1 - (1/max_depth).
579 This way the expected number of recision is at most max_depth. */
582 int max_prob
= CGRAPH_FREQ_BASE
- ((CGRAPH_FREQ_BASE
+ max_depth
- 1)
585 for (i
= 1; i
< depth
; i
++)
586 max_prob
= max_prob
* max_prob
/ CGRAPH_FREQ_BASE
;
588 && (edge
->count
* CGRAPH_FREQ_BASE
/ outer_node
->count
591 reason
= "profile of recursive call is too large";
595 && (edge
->frequency
* CGRAPH_FREQ_BASE
/ caller_freq
598 reason
= "frequency of recursive call is too large";
602 /* Recursive inlining, i.e. equivalent of unrolling, is profitable if recursion
603 depth is large. We reduce function call overhead and increase chances that
604 things fit in hardware return predictor.
606 Recursive inlining might however increase cost of stack frame setup
607 actually slowing down functions whose recursion tree is wide rather than
610 Deciding reliably on when to do recursive inlining without profile feedback
611 is tricky. For now we disable recursive inlining when probability of self
614 Recursive inlining of self recursive call within loop also results in large loop
615 depths that generally optimize badly. We may want to throttle down inlining
616 in those cases. In particular this seems to happen in one of libstdc++ rb tree
621 && (edge
->count
* 100 / outer_node
->count
622 <= PARAM_VALUE (PARAM_MIN_INLINE_RECURSIVE_PROBABILITY
)))
624 reason
= "profile of recursive call is too small";
628 && (edge
->frequency
* 100 / caller_freq
629 <= PARAM_VALUE (PARAM_MIN_INLINE_RECURSIVE_PROBABILITY
)))
631 reason
= "frequency of recursive call is too small";
635 if (!want_inline
&& dump_file
)
636 fprintf (dump_file
, " not inlining recursively: %s\n", reason
);
641 /* Decide if NODE is called once inlining it would eliminate need
642 for the offline copy of function. */
645 want_inline_function_called_once_p (struct cgraph_node
*node
)
647 /* Already inlined? */
648 if (node
->global
.inlined_to
)
650 /* Zero or more then one callers? */
652 || node
->callers
->next_caller
)
654 /* Recursive call makes no sense to inline. */
655 if (node
->callers
->caller
== node
)
657 /* External functions are not really in the unit, so inlining
658 them when called once would just increase the program size. */
659 if (DECL_EXTERNAL (node
->decl
))
661 /* Offline body must be optimized out. */
662 if (!cgraph_will_be_removed_from_program_if_no_direct_calls (node
))
664 if (!can_inline_edge_p (node
->callers
, true))
669 /* A cost model driving the inlining heuristics in a way so the edges with
670 smallest badness are inlined first. After each inlining is performed
671 the costs of all caller edges of nodes affected are recomputed so the
672 metrics may accurately depend on values such as number of inlinable callers
673 of the function or function body size. */
676 edge_badness (struct cgraph_edge
*edge
, bool dump
)
679 int growth
, time_growth
;
680 struct inline_summary
*callee_info
= inline_summary (edge
->callee
);
682 if (DECL_DISREGARD_INLINE_LIMITS (edge
->callee
->decl
))
685 growth
= estimate_edge_growth (edge
);
686 time_growth
= estimate_edge_time (edge
);
690 fprintf (dump_file
, " Badness calculation for %s -> %s\n",
691 cgraph_node_name (edge
->caller
),
692 cgraph_node_name (edge
->callee
));
693 fprintf (dump_file
, " growth size %i, time %i\n",
698 /* Always prefer inlining saving code size. */
701 badness
= INT_MIN
- growth
;
703 fprintf (dump_file
, " %i: Growth %i < 0\n", (int) badness
,
707 /* When profiling is available, base priorities -(#calls / growth).
708 So we optimize for overall number of "executed" inlined calls. */
712 benefitperc
= (((gcov_type
)callee_info
->time
713 * edge
->frequency
/ CGRAPH_FREQ_BASE
- time_growth
) * 100
714 / (callee_info
->time
+ 1) + 1);
715 benefitperc
= MIN (benefitperc
, 100);
716 benefitperc
= MAX (benefitperc
, 0);
719 ((double) edge
->count
* INT_MIN
/ max_count
/ 100) *
720 benefitperc
) / growth
;
722 /* Be sure that insanity of the profile won't lead to increasing counts
723 in the scalling and thus to overflow in the computation above. */
724 gcc_assert (max_count
>= edge
->count
);
728 " %i (relative %f): profile info. Relative count %f"
729 " * Relative benefit %f\n",
730 (int) badness
, (double) badness
/ INT_MIN
,
731 (double) edge
->count
/ max_count
,
732 (double) benefitperc
);
736 /* When function local profile is available, base priorities on
737 growth / frequency, so we optimize for overall frequency of inlined
738 calls. This is not too accurate since while the call might be frequent
739 within function, the function itself is infrequent.
741 Other objective to optimize for is number of different calls inlined.
742 We add the estimated growth after inlining all functions to bias the
743 priorities slightly in this direction (so fewer times called functions
744 of the same size gets priority). */
745 else if (flag_guess_branch_prob
)
747 int div
= edge
->frequency
* 100 / CGRAPH_FREQ_BASE
+ 1;
750 badness
= growth
* 10000;
751 benefitperc
= (((gcov_type
)callee_info
->time
752 * edge
->frequency
/ CGRAPH_FREQ_BASE
- time_growth
) * 100
753 / (callee_info
->time
+ 1) + 1);
754 benefitperc
= MIN (benefitperc
, 100);
755 benefitperc
= MAX (benefitperc
, 0);
758 /* Decrease badness if call is nested. */
759 /* Compress the range so we don't overflow. */
761 div
= 10000 + ceil_log2 (div
) - 8;
766 growth_for_all
= estimate_growth (edge
->callee
);
767 badness
+= growth_for_all
;
768 if (badness
> INT_MAX
)
773 " %i: guessed profile. frequency %i, overall growth %i,"
774 " benefit %i%%, divisor %i\n",
775 (int) badness
, edge
->frequency
, growth_for_all
,
779 /* When function local profile is not available or it does not give
780 useful information (ie frequency is zero), base the cost on
781 loop nest and overall size growth, so we optimize for overall number
782 of functions fully inlined in program. */
785 int nest
= MIN (inline_edge_summary (edge
)->loop_depth
, 8);
786 badness
= estimate_growth (edge
->callee
) * 256;
788 /* Decrease badness if call is nested. */
796 fprintf (dump_file
, " %i: no profile. nest %i\n", (int) badness
,
800 /* Ensure that we did not overflow in all the fixed point math above. */
801 gcc_assert (badness
>= INT_MIN
);
802 gcc_assert (badness
<= INT_MAX
- 1);
803 /* Make recursive inlining happen always after other inlining is done. */
804 if (cgraph_edge_recursive_p (edge
))
810 /* Recompute badness of EDGE and update its key in HEAP if needed. */
812 update_edge_key (fibheap_t heap
, struct cgraph_edge
*edge
)
814 int badness
= edge_badness (edge
, false);
817 fibnode_t n
= (fibnode_t
) edge
->aux
;
818 gcc_checking_assert (n
->data
== edge
);
820 /* fibheap_replace_key only decrease the keys.
821 When we increase the key we do not update heap
822 and instead re-insert the element once it becomes
823 a minimum of heap. */
824 if (badness
< n
->key
)
826 fibheap_replace_key (heap
, n
, badness
);
827 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
830 " decreasing badness %s/%i -> %s/%i, %i to %i\n",
831 cgraph_node_name (edge
->caller
), edge
->caller
->uid
,
832 cgraph_node_name (edge
->callee
), edge
->callee
->uid
,
836 gcc_checking_assert (n
->key
== badness
);
841 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
844 " enqueuing call %s/%i -> %s/%i, badness %i\n",
845 cgraph_node_name (edge
->caller
), edge
->caller
->uid
,
846 cgraph_node_name (edge
->callee
), edge
->callee
->uid
,
849 edge
->aux
= fibheap_insert (heap
, badness
, edge
);
855 All caller edges needs to be resetted because
856 size estimates change. Similarly callees needs reset
857 because better context may be known. */
860 reset_edge_caches (struct cgraph_node
*node
)
862 struct cgraph_edge
*edge
;
863 struct cgraph_edge
*e
= node
->callees
;
864 struct cgraph_node
*where
= node
;
866 if (where
->global
.inlined_to
)
867 where
= where
->global
.inlined_to
;
869 /* WHERE body size has changed, the cached growth is invalid. */
870 reset_node_growth_cache (where
);
872 for (edge
= where
->callers
; edge
; edge
= edge
->next_caller
)
873 if (edge
->inline_failed
)
874 reset_edge_growth_cache (edge
);
880 if (!e
->inline_failed
&& e
->callee
->callees
)
881 e
= e
->callee
->callees
;
884 if (e
->inline_failed
)
885 reset_edge_growth_cache (e
);
892 if (e
->caller
== node
)
894 e
= e
->caller
->callers
;
896 while (!e
->next_callee
);
902 /* Recompute HEAP nodes for each of caller of NODE.
903 UPDATED_NODES track nodes we already visited, to avoid redundant work.
904 When CHECK_INLINABLITY_FOR is set, re-check for specified edge that
905 it is inlinable. Otherwise check all edges. */
908 update_caller_keys (fibheap_t heap
, struct cgraph_node
*node
,
909 bitmap updated_nodes
,
910 struct cgraph_edge
*check_inlinablity_for
)
912 struct cgraph_edge
*edge
;
914 if (!inline_summary (node
)->inlinable
915 || cgraph_function_body_availability (node
) <= AVAIL_OVERWRITABLE
916 || node
->global
.inlined_to
)
918 if (!bitmap_set_bit (updated_nodes
, node
->uid
))
921 for (edge
= node
->callers
; edge
; edge
= edge
->next_caller
)
922 if (edge
->inline_failed
)
924 if (!check_inlinablity_for
925 || check_inlinablity_for
== edge
)
927 if (can_inline_edge_p (edge
, false)
928 && want_inline_small_function_p (edge
, false))
929 update_edge_key (heap
, edge
);
932 report_inline_failed_reason (edge
);
933 fibheap_delete_node (heap
, (fibnode_t
) edge
->aux
);
938 update_edge_key (heap
, edge
);
942 /* Recompute HEAP nodes for each uninlined call in NODE.
943 This is used when we know that edge badnesses are going only to increase
944 (we introduced new call site) and thus all we need is to insert newly
945 created edges into heap. */
948 update_callee_keys (fibheap_t heap
, struct cgraph_node
*node
,
949 bitmap updated_nodes
)
951 struct cgraph_edge
*e
= node
->callees
;
956 if (!e
->inline_failed
&& e
->callee
->callees
)
957 e
= e
->callee
->callees
;
960 /* We inlined and thus callees might have different number of calls.
961 Reset their caches */
962 reset_node_growth_cache (e
->callee
);
964 && inline_summary (e
->callee
)->inlinable
965 && cgraph_function_body_availability (e
->callee
) >= AVAIL_AVAILABLE
966 && !bitmap_bit_p (updated_nodes
, e
->callee
->uid
))
968 if (can_inline_edge_p (e
, false)
969 && want_inline_small_function_p (e
, false))
970 update_edge_key (heap
, e
);
973 report_inline_failed_reason (e
);
974 fibheap_delete_node (heap
, (fibnode_t
) e
->aux
);
984 if (e
->caller
== node
)
986 e
= e
->caller
->callers
;
988 while (!e
->next_callee
);
994 /* Recompute heap nodes for each of caller edges of each of callees.
995 Walk recursively into all inline clones. */
998 update_all_callee_keys (fibheap_t heap
, struct cgraph_node
*node
,
999 bitmap updated_nodes
)
1001 struct cgraph_edge
*e
= node
->callees
;
1006 if (!e
->inline_failed
&& e
->callee
->callees
)
1007 e
= e
->callee
->callees
;
1010 /* We inlined and thus callees might have different number of calls.
1011 Reset their caches */
1012 reset_node_growth_cache (e
->callee
);
1013 if (e
->inline_failed
)
1014 update_caller_keys (heap
, e
->callee
, updated_nodes
, e
);
1021 if (e
->caller
== node
)
1023 e
= e
->caller
->callers
;
1025 while (!e
->next_callee
);
1031 /* Enqueue all recursive calls from NODE into priority queue depending on
1032 how likely we want to recursively inline the call. */
1035 lookup_recursive_calls (struct cgraph_node
*node
, struct cgraph_node
*where
,
1038 struct cgraph_edge
*e
;
1039 for (e
= where
->callees
; e
; e
= e
->next_callee
)
1040 if (e
->callee
== node
)
1042 /* When profile feedback is available, prioritize by expected number
1044 fibheap_insert (heap
,
1045 !max_count
? -e
->frequency
1046 : -(e
->count
/ ((max_count
+ (1<<24) - 1) / (1<<24))),
1049 for (e
= where
->callees
; e
; e
= e
->next_callee
)
1050 if (!e
->inline_failed
)
1051 lookup_recursive_calls (node
, e
->callee
, heap
);
1054 /* Decide on recursive inlining: in the case function has recursive calls,
1055 inline until body size reaches given argument. If any new indirect edges
1056 are discovered in the process, add them to *NEW_EDGES, unless NEW_EDGES
1060 recursive_inlining (struct cgraph_edge
*edge
,
1061 VEC (cgraph_edge_p
, heap
) **new_edges
)
1063 int limit
= PARAM_VALUE (PARAM_MAX_INLINE_INSNS_RECURSIVE_AUTO
);
1065 struct cgraph_node
*node
;
1066 struct cgraph_edge
*e
;
1067 struct cgraph_node
*master_clone
= NULL
, *next
;
1071 node
= edge
->caller
;
1072 if (node
->global
.inlined_to
)
1073 node
= node
->global
.inlined_to
;
1075 if (DECL_DECLARED_INLINE_P (node
->decl
))
1076 limit
= PARAM_VALUE (PARAM_MAX_INLINE_INSNS_RECURSIVE
);
1078 /* Make sure that function is small enough to be considered for inlining. */
1079 if (estimate_size_after_inlining (node
, edge
) >= limit
)
1081 heap
= fibheap_new ();
1082 lookup_recursive_calls (node
, node
, heap
);
1083 if (fibheap_empty (heap
))
1085 fibheap_delete (heap
);
1091 " Performing recursive inlining on %s\n",
1092 cgraph_node_name (node
));
1094 /* Do the inlining and update list of recursive call during process. */
1095 while (!fibheap_empty (heap
))
1097 struct cgraph_edge
*curr
1098 = (struct cgraph_edge
*) fibheap_extract_min (heap
);
1099 struct cgraph_node
*cnode
;
1101 if (estimate_size_after_inlining (node
, curr
) > limit
)
1104 if (!can_inline_edge_p (curr
, true))
1108 for (cnode
= curr
->caller
;
1109 cnode
->global
.inlined_to
; cnode
= cnode
->callers
->caller
)
1110 if (node
->decl
== curr
->callee
->decl
)
1113 if (!want_inline_self_recursive_call_p (curr
, node
, false, depth
))
1119 " Inlining call of depth %i", depth
);
1122 fprintf (dump_file
, " called approx. %.2f times per call",
1123 (double)curr
->count
/ node
->count
);
1125 fprintf (dump_file
, "\n");
1129 /* We need original clone to copy around. */
1130 master_clone
= cgraph_clone_node (node
, node
->decl
,
1131 node
->count
, CGRAPH_FREQ_BASE
,
1133 for (e
= master_clone
->callees
; e
; e
= e
->next_callee
)
1134 if (!e
->inline_failed
)
1135 clone_inlined_nodes (e
, true, false, NULL
);
1138 cgraph_redirect_edge_callee (curr
, master_clone
);
1139 inline_call (curr
, false, new_edges
, &overall_size
);
1140 lookup_recursive_calls (node
, curr
->callee
, heap
);
1144 if (!fibheap_empty (heap
) && dump_file
)
1145 fprintf (dump_file
, " Recursive inlining growth limit met.\n");
1146 fibheap_delete (heap
);
1153 "\n Inlined %i times, "
1154 "body grown from size %i to %i, time %i to %i\n", n
,
1155 inline_summary (master_clone
)->size
, inline_summary (node
)->size
,
1156 inline_summary (master_clone
)->time
, inline_summary (node
)->time
);
1158 /* Remove master clone we used for inlining. We rely that clones inlined
1159 into master clone gets queued just before master clone so we don't
1161 for (node
= cgraph_nodes
; node
!= master_clone
;
1165 if (node
->global
.inlined_to
== master_clone
)
1166 cgraph_remove_node (node
);
1168 cgraph_remove_node (master_clone
);
1173 /* Given whole compilation unit estimate of INSNS, compute how large we can
1174 allow the unit to grow. */
1177 compute_max_insns (int insns
)
1179 int max_insns
= insns
;
1180 if (max_insns
< PARAM_VALUE (PARAM_LARGE_UNIT_INSNS
))
1181 max_insns
= PARAM_VALUE (PARAM_LARGE_UNIT_INSNS
);
1183 return ((HOST_WIDEST_INT
) max_insns
1184 * (100 + PARAM_VALUE (PARAM_INLINE_UNIT_GROWTH
)) / 100);
1188 /* Compute badness of all edges in NEW_EDGES and add them to the HEAP. */
1191 add_new_edges_to_heap (fibheap_t heap
, VEC (cgraph_edge_p
, heap
) *new_edges
)
1193 while (VEC_length (cgraph_edge_p
, new_edges
) > 0)
1195 struct cgraph_edge
*edge
= VEC_pop (cgraph_edge_p
, new_edges
);
1197 gcc_assert (!edge
->aux
);
1198 if (inline_summary (edge
->callee
)->inlinable
1199 && edge
->inline_failed
1200 && can_inline_edge_p (edge
, true)
1201 && want_inline_small_function_p (edge
, true))
1202 edge
->aux
= fibheap_insert (heap
, edge_badness (edge
, false), edge
);
1207 /* We use greedy algorithm for inlining of small functions:
1208 All inline candidates are put into prioritized heap ordered in
1211 The inlining of small functions is bounded by unit growth parameters. */
1214 inline_small_functions (void)
1216 struct cgraph_node
*node
;
1217 struct cgraph_edge
*edge
;
1218 fibheap_t heap
= fibheap_new ();
1219 bitmap updated_nodes
= BITMAP_ALLOC (NULL
);
1220 int min_size
, max_size
;
1221 VEC (cgraph_edge_p
, heap
) *new_indirect_edges
= NULL
;
1222 int initial_size
= 0;
1224 if (flag_indirect_inlining
)
1225 new_indirect_edges
= VEC_alloc (cgraph_edge_p
, heap
, 8);
1229 "\nDeciding on inlining of small functions. Starting with size %i.\n",
1232 /* Compute overall unit size and other global parameters used by badness
1236 initialize_growth_caches ();
1238 FOR_EACH_DEFINED_FUNCTION (node
)
1239 if (!node
->global
.inlined_to
)
1241 struct inline_summary
*info
= inline_summary (node
);
1243 if (!DECL_EXTERNAL (node
->decl
))
1244 initial_size
+= info
->size
;
1246 for (edge
= node
->callers
; edge
; edge
= edge
->next_caller
)
1247 if (max_count
< edge
->count
)
1248 max_count
= edge
->count
;
1251 overall_size
= initial_size
;
1252 max_size
= compute_max_insns (overall_size
);
1253 min_size
= overall_size
;
1255 /* Populate the heeap with all edges we might inline. */
1257 FOR_EACH_DEFINED_FUNCTION (node
)
1258 if (!node
->global
.inlined_to
)
1261 fprintf (dump_file
, "Enqueueing calls of %s/%i.\n",
1262 cgraph_node_name (node
), node
->uid
);
1264 for (edge
= node
->callers
; edge
; edge
= edge
->next_caller
)
1265 if (edge
->inline_failed
1266 && can_inline_edge_p (edge
, true)
1267 && want_inline_small_function_p (edge
, true)
1268 && edge
->inline_failed
)
1270 gcc_assert (!edge
->aux
);
1271 update_edge_key (heap
, edge
);
1275 gcc_assert (in_lto_p
1277 || (profile_info
&& flag_branch_probabilities
));
1279 while (!fibheap_empty (heap
))
1281 int old_size
= overall_size
;
1282 struct cgraph_node
*where
, *callee
;
1283 int badness
= fibheap_min_key (heap
);
1284 int current_badness
;
1287 edge
= (struct cgraph_edge
*) fibheap_extract_min (heap
);
1288 gcc_assert (edge
->aux
);
1290 if (!edge
->inline_failed
)
1293 /* Be sure that caches are maintained consistent. */
1294 #ifdef ENABLE_CHECKING
1295 reset_edge_growth_cache (edge
);
1296 reset_node_growth_cache (edge
->callee
);
1299 /* When updating the edge costs, we only decrease badness in the keys.
1300 Increases of badness are handled lazilly; when we see key with out
1301 of date value on it, we re-insert it now. */
1302 current_badness
= edge_badness (edge
, false);
1303 gcc_assert (current_badness
>= badness
);
1304 if (current_badness
!= badness
)
1306 edge
->aux
= fibheap_insert (heap
, current_badness
, edge
);
1310 if (!can_inline_edge_p (edge
, true))
1313 callee
= edge
->callee
;
1314 growth
= estimate_edge_growth (edge
);
1318 "\nConsidering %s with %i size\n",
1319 cgraph_node_name (edge
->callee
),
1320 inline_summary (edge
->callee
)->size
);
1322 " to be inlined into %s in %s:%i\n"
1323 " Estimated growth after inlined into all is %+i insns.\n"
1324 " Estimated badness is %i, frequency %.2f.\n",
1325 cgraph_node_name (edge
->caller
),
1326 flag_wpa
? "unknown"
1327 : gimple_filename ((const_gimple
) edge
->call_stmt
),
1329 : gimple_lineno ((const_gimple
) edge
->call_stmt
),
1330 estimate_growth (edge
->callee
),
1332 edge
->frequency
/ (double)CGRAPH_FREQ_BASE
);
1334 fprintf (dump_file
," Called "HOST_WIDEST_INT_PRINT_DEC
"x\n",
1336 if (dump_flags
& TDF_DETAILS
)
1337 edge_badness (edge
, true);
1340 if (overall_size
+ growth
> max_size
1341 && !DECL_DISREGARD_INLINE_LIMITS (edge
->callee
->decl
))
1343 edge
->inline_failed
= CIF_INLINE_UNIT_GROWTH_LIMIT
;
1344 report_inline_failed_reason (edge
);
1348 if (!want_inline_small_function_p (edge
, true))
1351 /* Heuristics for inlining small functions works poorly for
1352 recursive calls where we do efect similar to loop unrolling.
1353 When inliing such edge seems profitable, leave decision on
1354 specific inliner. */
1355 if (cgraph_edge_recursive_p (edge
))
1357 where
= edge
->caller
;
1358 if (where
->global
.inlined_to
)
1359 where
= where
->global
.inlined_to
;
1360 if (!recursive_inlining (edge
,
1361 flag_indirect_inlining
1362 ? &new_indirect_edges
: NULL
))
1364 edge
->inline_failed
= CIF_RECURSIVE_INLINING
;
1367 reset_edge_caches (where
);
1368 /* Recursive inliner inlines all recursive calls of the function
1369 at once. Consequently we need to update all callee keys. */
1370 if (flag_indirect_inlining
)
1371 add_new_edges_to_heap (heap
, new_indirect_edges
);
1372 update_all_callee_keys (heap
, where
, updated_nodes
);
1376 struct cgraph_node
*callee
;
1377 struct cgraph_node
*outer_node
= NULL
;
1380 /* Consider the case where self recursive function A is inlined into B.
1381 This is desired optimization in some cases, since it leads to effect
1382 similar of loop peeling and we might completely optimize out the
1383 recursive call. However we must be extra selective. */
1385 where
= edge
->caller
;
1386 while (where
->global
.inlined_to
)
1388 if (where
->decl
== edge
->callee
->decl
)
1389 outer_node
= where
, depth
++;
1390 where
= where
->callers
->caller
;
1393 && !want_inline_self_recursive_call_p (edge
, outer_node
,
1397 = (DECL_DISREGARD_INLINE_LIMITS (edge
->callee
->decl
)
1398 ? CIF_RECURSIVE_INLINING
: CIF_UNSPECIFIED
);
1401 else if (depth
&& dump_file
)
1402 fprintf (dump_file
, " Peeling recursion with depth %i\n", depth
);
1404 callee
= edge
->callee
;
1405 gcc_checking_assert (!callee
->global
.inlined_to
);
1406 inline_call (edge
, true, &new_indirect_edges
, &overall_size
);
1407 if (flag_indirect_inlining
)
1408 add_new_edges_to_heap (heap
, new_indirect_edges
);
1410 reset_edge_caches (edge
->callee
);
1411 reset_node_growth_cache (callee
);
1413 /* We inlined last offline copy to the body. This might lead
1414 to callees of function having fewer call sites and thus they
1415 may need updating. */
1416 if (callee
->global
.inlined_to
)
1417 update_all_callee_keys (heap
, callee
, updated_nodes
);
1419 update_callee_keys (heap
, edge
->callee
, updated_nodes
);
1421 where
= edge
->caller
;
1422 if (where
->global
.inlined_to
)
1423 where
= where
->global
.inlined_to
;
1425 /* Our profitability metric can depend on local properties
1426 such as number of inlinable calls and size of the function body.
1427 After inlining these properties might change for the function we
1428 inlined into (since it's body size changed) and for the functions
1429 called by function we inlined (since number of it inlinable callers
1431 update_caller_keys (heap
, where
, updated_nodes
, NULL
);
1433 /* We removed one call of the function we just inlined. If offline
1434 copy is still needed, be sure to update the keys. */
1435 if (callee
!= where
&& !callee
->global
.inlined_to
)
1436 update_caller_keys (heap
, callee
, updated_nodes
, NULL
);
1437 bitmap_clear (updated_nodes
);
1442 " Inlined into %s which now has time %i and size %i,"
1443 "net change of %+i.\n",
1444 cgraph_node_name (edge
->caller
),
1445 inline_summary (edge
->caller
)->time
,
1446 inline_summary (edge
->caller
)->size
,
1447 overall_size
- old_size
);
1449 if (min_size
> overall_size
)
1451 min_size
= overall_size
;
1452 max_size
= compute_max_insns (min_size
);
1455 fprintf (dump_file
, "New minimal size reached: %i\n", min_size
);
1459 free_growth_caches ();
1460 if (new_indirect_edges
)
1461 VEC_free (cgraph_edge_p
, heap
, new_indirect_edges
);
1462 fibheap_delete (heap
);
1465 "Unit growth for small function inlining: %i->%i (%i%%)\n",
1466 initial_size
, overall_size
,
1467 initial_size
? overall_size
* 100 / (initial_size
) - 100: 0);
1468 BITMAP_FREE (updated_nodes
);
1471 /* Flatten NODE. Performed both during early inlining and
1472 at IPA inlining time. */
1475 flatten_function (struct cgraph_node
*node
, bool early
)
1477 struct cgraph_edge
*e
;
1479 /* We shouldn't be called recursively when we are being processed. */
1480 gcc_assert (node
->aux
== NULL
);
1482 node
->aux
= (void *) node
;
1484 for (e
= node
->callees
; e
; e
= e
->next_callee
)
1486 struct cgraph_node
*orig_callee
;
1488 /* We've hit cycle? It is time to give up. */
1493 "Not inlining %s into %s to avoid cycle.\n",
1494 cgraph_node_name (e
->callee
),
1495 cgraph_node_name (e
->caller
));
1496 e
->inline_failed
= CIF_RECURSIVE_INLINING
;
1500 /* When the edge is already inlined, we just need to recurse into
1501 it in order to fully flatten the leaves. */
1502 if (!e
->inline_failed
)
1504 flatten_function (e
->callee
, early
);
1508 /* Flatten attribute needs to be processed during late inlining. For
1509 extra code quality we however do flattening during early optimization,
1512 ? !can_inline_edge_p (e
, true)
1513 : !can_early_inline_edge_p (e
))
1516 if (cgraph_edge_recursive_p (e
))
1519 fprintf (dump_file
, "Not inlining: recursive call.\n");
1523 if (gimple_in_ssa_p (DECL_STRUCT_FUNCTION (node
->decl
))
1524 != gimple_in_ssa_p (DECL_STRUCT_FUNCTION (e
->callee
->decl
)))
1527 fprintf (dump_file
, "Not inlining: SSA form does not match.\n");
1531 /* Inline the edge and flatten the inline clone. Avoid
1532 recursing through the original node if the node was cloned. */
1534 fprintf (dump_file
, " Inlining %s into %s.\n",
1535 cgraph_node_name (e
->callee
),
1536 cgraph_node_name (e
->caller
));
1537 orig_callee
= e
->callee
;
1538 inline_call (e
, true, NULL
, NULL
);
1539 if (e
->callee
!= orig_callee
)
1540 orig_callee
->aux
= (void *) node
;
1541 flatten_function (e
->callee
, early
);
1542 if (e
->callee
!= orig_callee
)
1543 orig_callee
->aux
= NULL
;
1549 /* Decide on the inlining. We do so in the topological order to avoid
1550 expenses on updating data structures. */
1555 struct cgraph_node
*node
;
1557 struct cgraph_node
**order
=
1558 XCNEWVEC (struct cgraph_node
*, cgraph_n_nodes
);
1561 if (in_lto_p
&& flag_indirect_inlining
)
1562 ipa_update_after_lto_read ();
1563 if (flag_indirect_inlining
)
1564 ipa_create_all_structures_for_iinln ();
1567 dump_inline_summaries (dump_file
);
1569 nnodes
= ipa_reverse_postorder (order
);
1571 for (node
= cgraph_nodes
; node
; node
= node
->next
)
1575 fprintf (dump_file
, "\nFlattening functions:\n");
1577 /* In the first pass handle functions to be flattened. Do this with
1578 a priority so none of our later choices will make this impossible. */
1579 for (i
= nnodes
- 1; i
>= 0; i
--)
1583 /* Handle nodes to be flattened.
1584 Ideally when processing callees we stop inlining at the
1585 entry of cycles, possibly cloning that entry point and
1586 try to flatten itself turning it into a self-recursive
1588 if (lookup_attribute ("flatten",
1589 DECL_ATTRIBUTES (node
->decl
)) != NULL
)
1593 "Flattening %s\n", cgraph_node_name (node
));
1594 flatten_function (node
, false);
1598 inline_small_functions ();
1599 cgraph_remove_unreachable_nodes (true, dump_file
);
1602 /* We already perform some inlining of functions called once during
1603 inlining small functions above. After unreachable nodes are removed,
1604 we still might do a quick check that nothing new is found. */
1605 if (flag_inline_functions_called_once
)
1609 fprintf (dump_file
, "\nDeciding on functions called once:\n");
1611 /* Inlining one function called once has good chance of preventing
1612 inlining other function into the same callee. Ideally we should
1613 work in priority order, but probably inlining hot functions first
1614 is good cut without the extra pain of maintaining the queue.
1616 ??? this is not really fitting the bill perfectly: inlining function
1617 into callee often leads to better optimization of callee due to
1618 increased context for optimization.
1619 For example if main() function calls a function that outputs help
1620 and then function that does the main optmization, we should inline
1621 the second with priority even if both calls are cold by themselves.
1623 We probably want to implement new predicate replacing our use of
1624 maybe_hot_edge interpreted as maybe_hot_edge || callee is known
1626 for (cold
= 0; cold
<= 1; cold
++)
1628 for (node
= cgraph_nodes
; node
; node
= node
->next
)
1630 if (want_inline_function_called_once_p (node
)
1632 || cgraph_maybe_hot_edge_p (node
->callers
)))
1634 struct cgraph_node
*caller
= node
->callers
->caller
;
1639 "\nInlining %s size %i.\n",
1640 cgraph_node_name (node
), inline_summary (node
)->size
);
1642 " Called once from %s %i insns.\n",
1643 cgraph_node_name (node
->callers
->caller
),
1644 inline_summary (node
->callers
->caller
)->size
);
1647 inline_call (node
->callers
, true, NULL
, NULL
);
1650 " Inlined into %s which now has %i size\n",
1651 cgraph_node_name (caller
),
1652 inline_summary (caller
)->size
);
1658 /* Free ipa-prop structures if they are no longer needed. */
1659 if (flag_indirect_inlining
)
1660 ipa_free_all_structures_after_iinln ();
1664 "\nInlined %i calls, eliminated %i functions\n\n",
1665 ncalls_inlined
, nfunctions_inlined
);
1668 dump_inline_summaries (dump_file
);
1669 /* In WPA we use inline summaries for partitioning process. */
1671 inline_free_summary ();
1675 /* Inline always-inline function calls in NODE. */
1678 inline_always_inline_functions (struct cgraph_node
*node
)
1680 struct cgraph_edge
*e
;
1681 bool inlined
= false;
1683 for (e
= node
->callees
; e
; e
= e
->next_callee
)
1685 if (!DECL_DISREGARD_INLINE_LIMITS (e
->callee
->decl
))
1688 if (cgraph_edge_recursive_p (e
))
1691 fprintf (dump_file
, " Not inlining recursive call to %s.\n",
1692 cgraph_node_name (e
->callee
));
1693 e
->inline_failed
= CIF_RECURSIVE_INLINING
;
1697 if (!can_early_inline_edge_p (e
))
1701 fprintf (dump_file
, " Inlining %s into %s (always_inline).\n",
1702 cgraph_node_name (e
->callee
),
1703 cgraph_node_name (e
->caller
));
1704 inline_call (e
, true, NULL
, NULL
);
1711 /* Decide on the inlining. We do so in the topological order to avoid
1712 expenses on updating data structures. */
1715 early_inline_small_functions (struct cgraph_node
*node
)
1717 struct cgraph_edge
*e
;
1718 bool inlined
= false;
1720 for (e
= node
->callees
; e
; e
= e
->next_callee
)
1722 if (!inline_summary (e
->callee
)->inlinable
1723 || !e
->inline_failed
)
1726 /* Do not consider functions not declared inline. */
1727 if (!DECL_DECLARED_INLINE_P (e
->callee
->decl
)
1728 && !flag_inline_small_functions
1729 && !flag_inline_functions
)
1733 fprintf (dump_file
, "Considering inline candidate %s.\n",
1734 cgraph_node_name (e
->callee
));
1736 if (!can_early_inline_edge_p (e
))
1739 if (cgraph_edge_recursive_p (e
))
1742 fprintf (dump_file
, " Not inlining: recursive call.\n");
1746 if (!want_early_inline_function_p (e
))
1750 fprintf (dump_file
, " Inlining %s into %s.\n",
1751 cgraph_node_name (e
->callee
),
1752 cgraph_node_name (e
->caller
));
1753 inline_call (e
, true, NULL
, NULL
);
1760 /* Do inlining of small functions. Doing so early helps profiling and other
1761 passes to be somewhat more effective and avoids some code duplication in
1762 later real inlining pass for testcases with very many function calls. */
1764 early_inliner (void)
1766 struct cgraph_node
*node
= cgraph_get_node (current_function_decl
);
1767 struct cgraph_edge
*edge
;
1768 unsigned int todo
= 0;
1770 bool inlined
= false;
1775 /* Do nothing if datastructures for ipa-inliner are already computed. This
1776 happens when some pass decides to construct new function and
1777 cgraph_add_new_function calls lowering passes and early optimization on
1778 it. This may confuse ourself when early inliner decide to inline call to
1779 function clone, because function clones don't have parameter list in
1780 ipa-prop matching their signature. */
1781 if (ipa_node_params_vector
)
1784 #ifdef ENABLE_CHECKING
1785 verify_cgraph_node (node
);
1788 /* Even when not optimizing or not inlining inline always-inline
1790 inlined
= inline_always_inline_functions (node
);
1794 || !flag_early_inlining
1795 /* Never inline regular functions into always-inline functions
1796 during incremental inlining. This sucks as functions calling
1797 always inline functions will get less optimized, but at the
1798 same time inlining of functions calling always inline
1799 function into an always inline function might introduce
1800 cycles of edges to be always inlined in the callgraph.
1802 We might want to be smarter and just avoid this type of inlining. */
1803 || DECL_DISREGARD_INLINE_LIMITS (node
->decl
))
1805 else if (lookup_attribute ("flatten",
1806 DECL_ATTRIBUTES (node
->decl
)) != NULL
)
1808 /* When the function is marked to be flattened, recursively inline
1812 "Flattening %s\n", cgraph_node_name (node
));
1813 flatten_function (node
, true);
1818 /* We iterate incremental inlining to get trivial cases of indirect
1820 while (iterations
< PARAM_VALUE (PARAM_EARLY_INLINER_MAX_ITERATIONS
)
1821 && early_inline_small_functions (node
))
1823 timevar_push (TV_INTEGRATION
);
1824 todo
|= optimize_inline_calls (current_function_decl
);
1826 /* Technically we ought to recompute inline parameters so the new
1827 iteration of early inliner works as expected. We however have
1828 values approximately right and thus we only need to update edge
1829 info that might be cleared out for newly discovered edges. */
1830 for (edge
= node
->callees
; edge
; edge
= edge
->next_callee
)
1832 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
1834 = estimate_num_insns (edge
->call_stmt
, &eni_size_weights
);
1836 = estimate_num_insns (edge
->call_stmt
, &eni_time_weights
);
1838 timevar_pop (TV_INTEGRATION
);
1843 fprintf (dump_file
, "Iterations: %i\n", iterations
);
1848 timevar_push (TV_INTEGRATION
);
1849 todo
|= optimize_inline_calls (current_function_decl
);
1850 timevar_pop (TV_INTEGRATION
);
1853 cfun
->always_inline_functions_inlined
= true;
1858 struct gimple_opt_pass pass_early_inline
=
1862 "einline", /* name */
1864 early_inliner
, /* execute */
1867 0, /* static_pass_number */
1868 TV_INLINE_HEURISTICS
, /* tv_id */
1869 PROP_ssa
, /* properties_required */
1870 0, /* properties_provided */
1871 0, /* properties_destroyed */
1872 0, /* todo_flags_start */
1873 TODO_dump_func
/* todo_flags_finish */
1878 /* When to run IPA inlining. Inlining of always-inline functions
1879 happens during early inlining. */
1882 gate_ipa_inline (void)
1884 /* ??? We'd like to skip this if not optimizing or not inlining as
1885 all always-inline functions have been processed by early
1886 inlining already. But this at least breaks EH with C++ as
1887 we need to unconditionally run fixup_cfg even at -O0.
1888 So leave it on unconditionally for now. */
1892 struct ipa_opt_pass_d pass_ipa_inline
=
1896 "inline", /* name */
1897 gate_ipa_inline
, /* gate */
1898 ipa_inline
, /* execute */
1901 0, /* static_pass_number */
1902 TV_INLINE_HEURISTICS
, /* tv_id */
1903 0, /* properties_required */
1904 0, /* properties_provided */
1905 0, /* properties_destroyed */
1906 TODO_remove_functions
, /* todo_flags_finish */
1907 TODO_dump_cgraph
| TODO_dump_func
1908 | TODO_remove_functions
| TODO_ggc_collect
/* todo_flags_finish */
1910 inline_generate_summary
, /* generate_summary */
1911 inline_write_summary
, /* write_summary */
1912 inline_read_summary
, /* read_summary */
1913 NULL
, /* write_optimization_summary */
1914 NULL
, /* read_optimization_summary */
1915 NULL
, /* stmt_fixup */
1917 inline_transform
, /* function_transform */
1918 NULL
, /* variable_transform */