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"
107 #include "tree-pass.h"
108 #include "coverage.h"
111 #include "tree-flow.h"
112 #include "ipa-prop.h"
115 #include "ipa-inline.h"
116 #include "ipa-utils.h"
118 /* Statistics we collect about inlining algorithm. */
119 static int overall_size
;
120 static gcov_type max_count
;
122 /* Return false when inlining edge E would lead to violating
123 limits on function unit growth or stack usage growth.
125 The relative function body growth limit is present generally
126 to avoid problems with non-linear behavior of the compiler.
127 To allow inlining huge functions into tiny wrapper, the limit
128 is always based on the bigger of the two functions considered.
130 For stack growth limits we always base the growth in stack usage
131 of the callers. We want to prevent applications from segfaulting
132 on stack overflow when functions with huge stack frames gets
136 caller_growth_limits (struct cgraph_edge
*e
)
138 struct cgraph_node
*to
= e
->caller
;
139 struct cgraph_node
*what
= cgraph_function_or_thunk_node (e
->callee
, NULL
);
142 HOST_WIDE_INT stack_size_limit
= 0, inlined_stack
;
143 struct inline_summary
*info
, *what_info
, *outer_info
= inline_summary (to
);
145 /* Look for function e->caller is inlined to. While doing
146 so work out the largest function body on the way. As
147 described above, we want to base our function growth
148 limits based on that. Not on the self size of the
149 outer function, not on the self size of inline code
150 we immediately inline to. This is the most relaxed
151 interpretation of the rule "do not grow large functions
152 too much in order to prevent compiler from exploding". */
155 info
= inline_summary (to
);
156 if (limit
< info
->self_size
)
157 limit
= info
->self_size
;
158 if (stack_size_limit
< info
->estimated_self_stack_size
)
159 stack_size_limit
= info
->estimated_self_stack_size
;
160 if (to
->global
.inlined_to
)
161 to
= to
->callers
->caller
;
166 what_info
= inline_summary (what
);
168 if (limit
< what_info
->self_size
)
169 limit
= what_info
->self_size
;
171 limit
+= limit
* PARAM_VALUE (PARAM_LARGE_FUNCTION_GROWTH
) / 100;
173 /* Check the size after inlining against the function limits. But allow
174 the function to shrink if it went over the limits by forced inlining. */
175 newsize
= estimate_size_after_inlining (to
, e
);
176 if (newsize
>= info
->size
177 && newsize
> PARAM_VALUE (PARAM_LARGE_FUNCTION_INSNS
)
180 e
->inline_failed
= CIF_LARGE_FUNCTION_GROWTH_LIMIT
;
184 if (!what_info
->estimated_stack_size
)
187 /* FIXME: Stack size limit often prevents inlining in Fortran programs
188 due to large i/o datastructures used by the Fortran front-end.
189 We ought to ignore this limit when we know that the edge is executed
190 on every invocation of the caller (i.e. its call statement dominates
191 exit block). We do not track this information, yet. */
192 stack_size_limit
+= ((gcov_type
)stack_size_limit
193 * PARAM_VALUE (PARAM_STACK_FRAME_GROWTH
) / 100);
195 inlined_stack
= (outer_info
->stack_frame_offset
196 + outer_info
->estimated_self_stack_size
197 + what_info
->estimated_stack_size
);
198 /* Check new stack consumption with stack consumption at the place
200 if (inlined_stack
> stack_size_limit
201 /* If function already has large stack usage from sibling
202 inline call, we can inline, too.
203 This bit overoptimistically assume that we are good at stack
205 && inlined_stack
> info
->estimated_stack_size
206 && inlined_stack
> PARAM_VALUE (PARAM_LARGE_STACK_FRAME
))
208 e
->inline_failed
= CIF_LARGE_STACK_FRAME_GROWTH_LIMIT
;
214 /* Dump info about why inlining has failed. */
217 report_inline_failed_reason (struct cgraph_edge
*e
)
221 fprintf (dump_file
, " not inlinable: %s/%i -> %s/%i, %s\n",
222 xstrdup (cgraph_node_name (e
->caller
)), e
->caller
->uid
,
223 xstrdup (cgraph_node_name (e
->callee
)), e
->callee
->uid
,
224 cgraph_inline_failed_string (e
->inline_failed
));
228 /* Decide if we can inline the edge and possibly update
229 inline_failed reason.
230 We check whether inlining is possible at all and whether
231 caller growth limits allow doing so.
233 if REPORT is true, output reason to the dump file. */
236 can_inline_edge_p (struct cgraph_edge
*e
, bool report
)
238 bool inlinable
= true;
239 enum availability avail
;
240 struct cgraph_node
*callee
241 = cgraph_function_or_thunk_node (e
->callee
, &avail
);
242 tree caller_tree
= DECL_FUNCTION_SPECIFIC_OPTIMIZATION (e
->caller
->symbol
.decl
);
244 = callee
? DECL_FUNCTION_SPECIFIC_OPTIMIZATION (callee
->symbol
.decl
) : NULL
;
245 struct function
*caller_cfun
= DECL_STRUCT_FUNCTION (e
->caller
->symbol
.decl
);
246 struct function
*callee_cfun
247 = callee
? DECL_STRUCT_FUNCTION (callee
->symbol
.decl
) : NULL
;
249 if (!caller_cfun
&& e
->caller
->clone_of
)
250 caller_cfun
= DECL_STRUCT_FUNCTION (e
->caller
->clone_of
->symbol
.decl
);
252 if (!callee_cfun
&& callee
&& callee
->clone_of
)
253 callee_cfun
= DECL_STRUCT_FUNCTION (callee
->clone_of
->symbol
.decl
);
255 gcc_assert (e
->inline_failed
);
257 if (!callee
|| !callee
->analyzed
)
259 e
->inline_failed
= CIF_BODY_NOT_AVAILABLE
;
262 else if (!inline_summary (callee
)->inlinable
)
264 e
->inline_failed
= CIF_FUNCTION_NOT_INLINABLE
;
267 else if (avail
<= AVAIL_OVERWRITABLE
)
269 e
->inline_failed
= CIF_OVERWRITABLE
;
272 else if (e
->call_stmt_cannot_inline_p
)
274 e
->inline_failed
= CIF_MISMATCHED_ARGUMENTS
;
277 /* Don't inline if the functions have different EH personalities. */
278 else if (DECL_FUNCTION_PERSONALITY (e
->caller
->symbol
.decl
)
279 && DECL_FUNCTION_PERSONALITY (callee
->symbol
.decl
)
280 && (DECL_FUNCTION_PERSONALITY (e
->caller
->symbol
.decl
)
281 != DECL_FUNCTION_PERSONALITY (callee
->symbol
.decl
)))
283 e
->inline_failed
= CIF_EH_PERSONALITY
;
286 /* TM pure functions should not be inlined into non-TM_pure
288 else if (is_tm_pure (callee
->symbol
.decl
)
289 && !is_tm_pure (e
->caller
->symbol
.decl
))
291 e
->inline_failed
= CIF_UNSPECIFIED
;
294 /* Don't inline if the callee can throw non-call exceptions but the
296 FIXME: this is obviously wrong for LTO where STRUCT_FUNCTION is missing.
297 Move the flag into cgraph node or mirror it in the inline summary. */
298 else if (callee_cfun
&& callee_cfun
->can_throw_non_call_exceptions
299 && !(caller_cfun
&& caller_cfun
->can_throw_non_call_exceptions
))
301 e
->inline_failed
= CIF_NON_CALL_EXCEPTIONS
;
304 /* Check compatibility of target optimization options. */
305 else if (!targetm
.target_option
.can_inline_p (e
->caller
->symbol
.decl
,
306 callee
->symbol
.decl
))
308 e
->inline_failed
= CIF_TARGET_OPTION_MISMATCH
;
311 /* Check if caller growth allows the inlining. */
312 else if (!DECL_DISREGARD_INLINE_LIMITS (callee
->symbol
.decl
)
313 && !lookup_attribute ("flatten",
315 (e
->caller
->global
.inlined_to
316 ? e
->caller
->global
.inlined_to
->symbol
.decl
317 : e
->caller
->symbol
.decl
))
318 && !caller_growth_limits (e
))
320 /* Don't inline a function with a higher optimization level than the
321 caller. FIXME: this is really just tip of iceberg of handling
322 optimization attribute. */
323 else if (caller_tree
!= callee_tree
)
325 struct cl_optimization
*caller_opt
326 = TREE_OPTIMIZATION ((caller_tree
)
328 : optimization_default_node
);
330 struct cl_optimization
*callee_opt
331 = TREE_OPTIMIZATION ((callee_tree
)
333 : optimization_default_node
);
335 if (((caller_opt
->x_optimize
> callee_opt
->x_optimize
)
336 || (caller_opt
->x_optimize_size
!= callee_opt
->x_optimize_size
))
337 /* gcc.dg/pr43564.c. Look at forced inline even in -O0. */
338 && !DECL_DISREGARD_INLINE_LIMITS (e
->callee
->symbol
.decl
))
340 e
->inline_failed
= CIF_OPTIMIZATION_MISMATCH
;
345 if (!inlinable
&& report
)
346 report_inline_failed_reason (e
);
351 /* Return true if the edge E is inlinable during early inlining. */
354 can_early_inline_edge_p (struct cgraph_edge
*e
)
356 struct cgraph_node
*callee
= cgraph_function_or_thunk_node (e
->callee
,
358 /* Early inliner might get called at WPA stage when IPA pass adds new
359 function. In this case we can not really do any of early inlining
360 because function bodies are missing. */
361 if (!gimple_has_body_p (callee
->symbol
.decl
))
363 e
->inline_failed
= CIF_BODY_NOT_AVAILABLE
;
366 /* In early inliner some of callees may not be in SSA form yet
367 (i.e. the callgraph is cyclic and we did not process
368 the callee by early inliner, yet). We don't have CIF code for this
369 case; later we will re-do the decision in the real inliner. */
370 if (!gimple_in_ssa_p (DECL_STRUCT_FUNCTION (e
->caller
->symbol
.decl
))
371 || !gimple_in_ssa_p (DECL_STRUCT_FUNCTION (callee
->symbol
.decl
)))
374 fprintf (dump_file
, " edge not inlinable: not in SSA form\n");
377 if (!can_inline_edge_p (e
, true))
383 /* Return true when N is leaf function. Accept cheap builtins
384 in leaf functions. */
387 leaf_node_p (struct cgraph_node
*n
)
389 struct cgraph_edge
*e
;
390 for (e
= n
->callees
; e
; e
= e
->next_callee
)
391 if (!is_inexpensive_builtin (e
->callee
->symbol
.decl
))
397 /* Return true if we are interested in inlining small function. */
400 want_early_inline_function_p (struct cgraph_edge
*e
)
402 bool want_inline
= true;
403 struct cgraph_node
*callee
= cgraph_function_or_thunk_node (e
->callee
, NULL
);
405 if (DECL_DISREGARD_INLINE_LIMITS (callee
->symbol
.decl
))
407 else if (!DECL_DECLARED_INLINE_P (callee
->symbol
.decl
)
408 && !flag_inline_small_functions
)
410 e
->inline_failed
= CIF_FUNCTION_NOT_INLINE_CANDIDATE
;
411 report_inline_failed_reason (e
);
416 int growth
= estimate_edge_growth (e
);
419 else if (!cgraph_maybe_hot_edge_p (e
)
423 fprintf (dump_file
, " will not early inline: %s/%i->%s/%i, "
424 "call is cold and code would grow by %i\n",
425 xstrdup (cgraph_node_name (e
->caller
)), e
->caller
->uid
,
426 xstrdup (cgraph_node_name (callee
)), callee
->uid
,
430 else if (!leaf_node_p (callee
)
434 fprintf (dump_file
, " will not early inline: %s/%i->%s/%i, "
435 "callee is not leaf and code would grow by %i\n",
436 xstrdup (cgraph_node_name (e
->caller
)), e
->caller
->uid
,
437 xstrdup (cgraph_node_name (callee
)), callee
->uid
,
441 else if (growth
> PARAM_VALUE (PARAM_EARLY_INLINING_INSNS
))
444 fprintf (dump_file
, " will not early inline: %s/%i->%s/%i, "
445 "growth %i exceeds --param early-inlining-insns\n",
446 xstrdup (cgraph_node_name (e
->caller
)), e
->caller
->uid
,
447 xstrdup (cgraph_node_name (callee
)), callee
->uid
,
455 /* Return true if we are interested in inlining small function.
456 When REPORT is true, report reason to dump file. */
459 want_inline_small_function_p (struct cgraph_edge
*e
, bool report
)
461 bool want_inline
= true;
462 struct cgraph_node
*callee
= cgraph_function_or_thunk_node (e
->callee
, NULL
);
464 if (DECL_DISREGARD_INLINE_LIMITS (callee
->symbol
.decl
))
466 else if (!DECL_DECLARED_INLINE_P (callee
->symbol
.decl
)
467 && !flag_inline_small_functions
)
469 e
->inline_failed
= CIF_FUNCTION_NOT_INLINE_CANDIDATE
;
474 int growth
= estimate_edge_growth (e
);
478 else if (DECL_DECLARED_INLINE_P (callee
->symbol
.decl
)
479 && growth
>= MAX_INLINE_INSNS_SINGLE
)
481 e
->inline_failed
= CIF_MAX_INLINE_INSNS_SINGLE_LIMIT
;
484 /* Before giving up based on fact that caller size will grow, allow
485 functions that are called few times and eliminating the offline
486 copy will lead to overall code size reduction.
487 Not all of these will be handled by subsequent inlining of functions
488 called once: in particular weak functions are not handled or funcitons
489 that inline to multiple calls but a lot of bodies is optimized out.
490 Finally we want to inline earlier to allow inlining of callbacks.
492 This is slightly wrong on aggressive side: it is entirely possible
493 that function is called many times with a context where inlining
494 reduces code size and few times with a context where inlining increase
495 code size. Resoluting growth estimate will be negative even if it
496 would make more sense to keep offline copy and do not inline into the
497 call sites that makes the code size grow.
499 When badness orders the calls in a way that code reducing calls come
500 first, this situation is not a problem at all: after inlining all
501 "good" calls, we will realize that keeping the function around is
503 else if (growth
<= MAX_INLINE_INSNS_SINGLE
504 /* Unlike for functions called once, we play unsafe with
505 COMDATs. We can allow that since we know functions
506 in consideration are small (and thus risk is small) and
507 moreover grow estimates already accounts that COMDAT
508 functions may or may not disappear when eliminated from
509 current unit. With good probability making aggressive
510 choice in all units is going to make overall program
513 Consequently we ask cgraph_can_remove_if_no_direct_calls_p
515 cgraph_will_be_removed_from_program_if_no_direct_calls */
516 && !DECL_EXTERNAL (callee
->symbol
.decl
)
517 && cgraph_can_remove_if_no_direct_calls_p (callee
)
518 && estimate_growth (callee
) <= 0)
520 else if (!DECL_DECLARED_INLINE_P (callee
->symbol
.decl
)
521 && !flag_inline_functions
)
523 e
->inline_failed
= CIF_NOT_DECLARED_INLINED
;
526 else if (!DECL_DECLARED_INLINE_P (callee
->symbol
.decl
)
527 && growth
>= MAX_INLINE_INSNS_AUTO
)
529 e
->inline_failed
= CIF_MAX_INLINE_INSNS_AUTO_LIMIT
;
532 /* If call is cold, do not inline when function body would grow. */
533 else if (!cgraph_maybe_hot_edge_p (e
))
535 e
->inline_failed
= CIF_UNLIKELY_CALL
;
539 if (!want_inline
&& report
)
540 report_inline_failed_reason (e
);
544 /* EDGE is self recursive edge.
545 We hand two cases - when function A is inlining into itself
546 or when function A is being inlined into another inliner copy of function
549 In first case OUTER_NODE points to the toplevel copy of A, while
550 in the second case OUTER_NODE points to the outermost copy of A in B.
552 In both cases we want to be extra selective since
553 inlining the call will just introduce new recursive calls to appear. */
556 want_inline_self_recursive_call_p (struct cgraph_edge
*edge
,
557 struct cgraph_node
*outer_node
,
561 char const *reason
= NULL
;
562 bool want_inline
= true;
563 int caller_freq
= CGRAPH_FREQ_BASE
;
564 int max_depth
= PARAM_VALUE (PARAM_MAX_INLINE_RECURSIVE_DEPTH_AUTO
);
566 if (DECL_DECLARED_INLINE_P (edge
->caller
->symbol
.decl
))
567 max_depth
= PARAM_VALUE (PARAM_MAX_INLINE_RECURSIVE_DEPTH
);
569 if (!cgraph_maybe_hot_edge_p (edge
))
571 reason
= "recursive call is cold";
574 else if (max_count
&& !outer_node
->count
)
576 reason
= "not executed in profile";
579 else if (depth
> max_depth
)
581 reason
= "--param max-inline-recursive-depth exceeded.";
585 if (outer_node
->global
.inlined_to
)
586 caller_freq
= outer_node
->callers
->frequency
;
590 /* Inlining of self recursive function into copy of itself within other function
591 is transformation similar to loop peeling.
593 Peeling is profitable if we can inline enough copies to make probability
594 of actual call to the self recursive function very small. Be sure that
595 the probability of recursion is small.
597 We ensure that the frequency of recursing is at most 1 - (1/max_depth).
598 This way the expected number of recision is at most max_depth. */
601 int max_prob
= CGRAPH_FREQ_BASE
- ((CGRAPH_FREQ_BASE
+ max_depth
- 1)
604 for (i
= 1; i
< depth
; i
++)
605 max_prob
= max_prob
* max_prob
/ CGRAPH_FREQ_BASE
;
607 && (edge
->count
* CGRAPH_FREQ_BASE
/ outer_node
->count
610 reason
= "profile of recursive call is too large";
614 && (edge
->frequency
* CGRAPH_FREQ_BASE
/ caller_freq
617 reason
= "frequency of recursive call is too large";
621 /* Recursive inlining, i.e. equivalent of unrolling, is profitable if recursion
622 depth is large. We reduce function call overhead and increase chances that
623 things fit in hardware return predictor.
625 Recursive inlining might however increase cost of stack frame setup
626 actually slowing down functions whose recursion tree is wide rather than
629 Deciding reliably on when to do recursive inlining without profile feedback
630 is tricky. For now we disable recursive inlining when probability of self
633 Recursive inlining of self recursive call within loop also results in large loop
634 depths that generally optimize badly. We may want to throttle down inlining
635 in those cases. In particular this seems to happen in one of libstdc++ rb tree
640 && (edge
->count
* 100 / outer_node
->count
641 <= PARAM_VALUE (PARAM_MIN_INLINE_RECURSIVE_PROBABILITY
)))
643 reason
= "profile of recursive call is too small";
647 && (edge
->frequency
* 100 / caller_freq
648 <= PARAM_VALUE (PARAM_MIN_INLINE_RECURSIVE_PROBABILITY
)))
650 reason
= "frequency of recursive call is too small";
654 if (!want_inline
&& dump_file
)
655 fprintf (dump_file
, " not inlining recursively: %s\n", reason
);
659 /* Return true when NODE has caller other than EDGE.
660 Worker for cgraph_for_node_and_aliases. */
663 check_caller_edge (struct cgraph_node
*node
, void *edge
)
665 return (node
->callers
666 && node
->callers
!= edge
);
670 /* Decide if NODE is called once inlining it would eliminate need
671 for the offline copy of function. */
674 want_inline_function_called_once_p (struct cgraph_node
*node
)
676 struct cgraph_node
*function
= cgraph_function_or_thunk_node (node
, NULL
);
677 /* Already inlined? */
678 if (function
->global
.inlined_to
)
680 /* Zero or more then one callers? */
682 || node
->callers
->next_caller
)
684 /* Maybe other aliases has more direct calls. */
685 if (cgraph_for_node_and_aliases (node
, check_caller_edge
, node
->callers
, true))
687 /* Recursive call makes no sense to inline. */
688 if (cgraph_edge_recursive_p (node
->callers
))
690 /* External functions are not really in the unit, so inlining
691 them when called once would just increase the program size. */
692 if (DECL_EXTERNAL (function
->symbol
.decl
))
694 /* Offline body must be optimized out. */
695 if (!cgraph_will_be_removed_from_program_if_no_direct_calls (function
))
697 if (!can_inline_edge_p (node
->callers
, true))
703 /* Return relative time improvement for inlining EDGE in range
707 relative_time_benefit (struct inline_summary
*callee_info
,
708 struct cgraph_edge
*edge
,
712 gcov_type uninlined_call_time
;
714 uninlined_call_time
=
717 + inline_edge_summary (edge
)->call_stmt_time
) * edge
->frequency
718 + CGRAPH_FREQ_BASE
/ 2) / CGRAPH_FREQ_BASE
;
719 /* Compute relative time benefit, i.e. how much the call becomes faster.
720 ??? perhaps computing how much the caller+calle together become faster
721 would lead to more realistic results. */
722 if (!uninlined_call_time
)
723 uninlined_call_time
= 1;
725 (uninlined_call_time
- time_growth
) * 256 / (uninlined_call_time
);
726 relbenefit
= MIN (relbenefit
, 512);
727 relbenefit
= MAX (relbenefit
, 1);
732 /* A cost model driving the inlining heuristics in a way so the edges with
733 smallest badness are inlined first. After each inlining is performed
734 the costs of all caller edges of nodes affected are recomputed so the
735 metrics may accurately depend on values such as number of inlinable callers
736 of the function or function body size. */
739 edge_badness (struct cgraph_edge
*edge
, bool dump
)
742 int growth
, time_growth
;
743 struct cgraph_node
*callee
= cgraph_function_or_thunk_node (edge
->callee
,
745 struct inline_summary
*callee_info
= inline_summary (callee
);
747 if (DECL_DISREGARD_INLINE_LIMITS (callee
->symbol
.decl
))
750 growth
= estimate_edge_growth (edge
);
751 time_growth
= estimate_edge_time (edge
);
755 fprintf (dump_file
, " Badness calculation for %s -> %s\n",
756 xstrdup (cgraph_node_name (edge
->caller
)),
757 xstrdup (cgraph_node_name (callee
)));
758 fprintf (dump_file
, " size growth %i, time growth %i\n",
763 /* Always prefer inlining saving code size. */
766 badness
= INT_MIN
/ 2 + growth
;
768 fprintf (dump_file
, " %i: Growth %i <= 0\n", (int) badness
,
772 /* When profiling is available, compute badness as:
774 relative_edge_count * relative_time_benefit
775 goodness = -------------------------------------------
779 The fraction is upside down, because on edge counts and time beneits
780 the bounds are known. Edge growth is essentially unlimited. */
784 int relbenefit
= relative_time_benefit (callee_info
, edge
, time_growth
);
787 ((double) edge
->count
* INT_MIN
/ 2 / max_count
/ 512) *
788 relative_time_benefit (callee_info
, edge
, time_growth
)) / growth
;
790 /* Be sure that insanity of the profile won't lead to increasing counts
791 in the scalling and thus to overflow in the computation above. */
792 gcc_assert (max_count
>= edge
->count
);
796 " %i (relative %f): profile info. Relative count %f"
797 " * Relative benefit %f\n",
798 (int) badness
, (double) badness
/ INT_MIN
,
799 (double) edge
->count
/ max_count
,
800 relbenefit
* 100 / 256.0);
804 /* When function local profile is available. Compute badness as:
808 badness = -------------------------------------- + growth_for-all
809 relative_time_benefit * edge_frequency
812 else if (flag_guess_branch_prob
)
814 int div
= edge
->frequency
* (1<<10) / CGRAPH_FREQ_MAX
;
817 gcc_checking_assert (edge
->frequency
<= CGRAPH_FREQ_MAX
);
818 div
*= relative_time_benefit (callee_info
, edge
, time_growth
);
820 /* frequency is normalized in range 1...2^10.
821 relbenefit in range 1...2^9
822 DIV should be in range 1....2^19. */
823 gcc_checking_assert (div
>= 1 && div
<= (1<<19));
825 /* Result must be integer in range 0...INT_MAX.
826 Set the base of fixed point calculation so we don't lose much of
827 precision for small bandesses (those are interesting) yet we don't
828 overflow for growths that are still in interesting range.
830 Fixed point arithmetic with point at 8th bit. */
831 badness
= ((gcov_type
)growth
) * (1<<(19+8));
832 badness
= (badness
+ div
/ 2) / div
;
834 /* Overall growth of inlining all calls of function matters: we want to
835 inline so offline copy of function is no longer needed.
837 Additionally functions that can be fully inlined without much of
838 effort are better inline candidates than functions that can be fully
839 inlined only after noticeable overall unit growths. The latter
840 are better in a sense compressing of code size by factoring out common
841 code into separate function shared by multiple code paths.
843 We might mix the valud into the fraction by taking into account
844 relative growth of the unit, but for now just add the number
845 into resulting fraction. */
846 if (badness
> INT_MAX
/ 2)
848 badness
= INT_MAX
/ 2;
850 fprintf (dump_file
, "Badness overflow\n");
855 " %i: guessed profile. frequency %f,"
856 " benefit %f%%, divisor %i\n",
857 (int) badness
, (double)edge
->frequency
/ CGRAPH_FREQ_BASE
,
858 relative_time_benefit (callee_info
, edge
, time_growth
) * 100 / 256.0, div
);
861 /* When function local profile is not available or it does not give
862 useful information (ie frequency is zero), base the cost on
863 loop nest and overall size growth, so we optimize for overall number
864 of functions fully inlined in program. */
867 int nest
= MIN (inline_edge_summary (edge
)->loop_depth
, 8);
868 badness
= growth
* 256;
870 /* Decrease badness if call is nested. */
878 fprintf (dump_file
, " %i: no profile. nest %i\n", (int) badness
,
882 /* Ensure that we did not overflow in all the fixed point math above. */
883 gcc_assert (badness
>= INT_MIN
);
884 gcc_assert (badness
<= INT_MAX
- 1);
885 /* Make recursive inlining happen always after other inlining is done. */
886 if (cgraph_edge_recursive_p (edge
))
892 /* Recompute badness of EDGE and update its key in HEAP if needed. */
894 update_edge_key (fibheap_t heap
, struct cgraph_edge
*edge
)
896 int badness
= edge_badness (edge
, false);
899 fibnode_t n
= (fibnode_t
) edge
->aux
;
900 gcc_checking_assert (n
->data
== edge
);
902 /* fibheap_replace_key only decrease the keys.
903 When we increase the key we do not update heap
904 and instead re-insert the element once it becomes
905 a minimum of heap. */
906 if (badness
< n
->key
)
908 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
911 " decreasing badness %s/%i -> %s/%i, %i to %i\n",
912 xstrdup (cgraph_node_name (edge
->caller
)),
914 xstrdup (cgraph_node_name (edge
->callee
)),
919 fibheap_replace_key (heap
, n
, badness
);
920 gcc_checking_assert (n
->key
== badness
);
925 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
928 " enqueuing call %s/%i -> %s/%i, badness %i\n",
929 xstrdup (cgraph_node_name (edge
->caller
)),
931 xstrdup (cgraph_node_name (edge
->callee
)),
935 edge
->aux
= fibheap_insert (heap
, badness
, edge
);
941 All caller edges needs to be resetted because
942 size estimates change. Similarly callees needs reset
943 because better context may be known. */
946 reset_edge_caches (struct cgraph_node
*node
)
948 struct cgraph_edge
*edge
;
949 struct cgraph_edge
*e
= node
->callees
;
950 struct cgraph_node
*where
= node
;
954 if (where
->global
.inlined_to
)
955 where
= where
->global
.inlined_to
;
957 /* WHERE body size has changed, the cached growth is invalid. */
958 reset_node_growth_cache (where
);
960 for (edge
= where
->callers
; edge
; edge
= edge
->next_caller
)
961 if (edge
->inline_failed
)
962 reset_edge_growth_cache (edge
);
963 for (i
= 0; ipa_ref_list_referring_iterate (&where
->symbol
.ref_list
,
965 if (ref
->use
== IPA_REF_ALIAS
)
966 reset_edge_caches (ipa_ref_referring_node (ref
));
972 if (!e
->inline_failed
&& e
->callee
->callees
)
973 e
= e
->callee
->callees
;
976 if (e
->inline_failed
)
977 reset_edge_growth_cache (e
);
984 if (e
->caller
== node
)
986 e
= e
->caller
->callers
;
988 while (!e
->next_callee
);
994 /* Recompute HEAP nodes for each of caller of NODE.
995 UPDATED_NODES track nodes we already visited, to avoid redundant work.
996 When CHECK_INLINABLITY_FOR is set, re-check for specified edge that
997 it is inlinable. Otherwise check all edges. */
1000 update_caller_keys (fibheap_t heap
, struct cgraph_node
*node
,
1001 bitmap updated_nodes
,
1002 struct cgraph_edge
*check_inlinablity_for
)
1004 struct cgraph_edge
*edge
;
1006 struct ipa_ref
*ref
;
1008 if ((!node
->alias
&& !inline_summary (node
)->inlinable
)
1009 || cgraph_function_body_availability (node
) <= AVAIL_OVERWRITABLE
1010 || node
->global
.inlined_to
)
1012 if (!bitmap_set_bit (updated_nodes
, node
->uid
))
1015 for (i
= 0; ipa_ref_list_referring_iterate (&node
->symbol
.ref_list
,
1017 if (ref
->use
== IPA_REF_ALIAS
)
1019 struct cgraph_node
*alias
= ipa_ref_referring_node (ref
);
1020 update_caller_keys (heap
, alias
, updated_nodes
, check_inlinablity_for
);
1023 for (edge
= node
->callers
; edge
; edge
= edge
->next_caller
)
1024 if (edge
->inline_failed
)
1026 if (!check_inlinablity_for
1027 || check_inlinablity_for
== edge
)
1029 if (can_inline_edge_p (edge
, false)
1030 && want_inline_small_function_p (edge
, false))
1031 update_edge_key (heap
, edge
);
1034 report_inline_failed_reason (edge
);
1035 fibheap_delete_node (heap
, (fibnode_t
) edge
->aux
);
1040 update_edge_key (heap
, edge
);
1044 /* Recompute HEAP nodes for each uninlined call in NODE.
1045 This is used when we know that edge badnesses are going only to increase
1046 (we introduced new call site) and thus all we need is to insert newly
1047 created edges into heap. */
1050 update_callee_keys (fibheap_t heap
, struct cgraph_node
*node
,
1051 bitmap updated_nodes
)
1053 struct cgraph_edge
*e
= node
->callees
;
1058 if (!e
->inline_failed
&& e
->callee
->callees
)
1059 e
= e
->callee
->callees
;
1062 enum availability avail
;
1063 struct cgraph_node
*callee
;
1064 /* We do not reset callee growth cache here. Since we added a new call,
1065 growth chould have just increased and consequentely badness metric
1066 don't need updating. */
1067 if (e
->inline_failed
1068 && (callee
= cgraph_function_or_thunk_node (e
->callee
, &avail
))
1069 && inline_summary (callee
)->inlinable
1070 && cgraph_function_body_availability (callee
) >= AVAIL_AVAILABLE
1071 && !bitmap_bit_p (updated_nodes
, callee
->uid
))
1073 if (can_inline_edge_p (e
, false)
1074 && want_inline_small_function_p (e
, false))
1075 update_edge_key (heap
, e
);
1078 report_inline_failed_reason (e
);
1079 fibheap_delete_node (heap
, (fibnode_t
) e
->aux
);
1089 if (e
->caller
== node
)
1091 e
= e
->caller
->callers
;
1093 while (!e
->next_callee
);
1099 /* Enqueue all recursive calls from NODE into priority queue depending on
1100 how likely we want to recursively inline the call. */
1103 lookup_recursive_calls (struct cgraph_node
*node
, struct cgraph_node
*where
,
1106 struct cgraph_edge
*e
;
1107 enum availability avail
;
1109 for (e
= where
->callees
; e
; e
= e
->next_callee
)
1110 if (e
->callee
== node
1111 || (cgraph_function_or_thunk_node (e
->callee
, &avail
) == node
1112 && avail
> AVAIL_OVERWRITABLE
))
1114 /* When profile feedback is available, prioritize by expected number
1116 fibheap_insert (heap
,
1117 !max_count
? -e
->frequency
1118 : -(e
->count
/ ((max_count
+ (1<<24) - 1) / (1<<24))),
1121 for (e
= where
->callees
; e
; e
= e
->next_callee
)
1122 if (!e
->inline_failed
)
1123 lookup_recursive_calls (node
, e
->callee
, heap
);
1126 /* Decide on recursive inlining: in the case function has recursive calls,
1127 inline until body size reaches given argument. If any new indirect edges
1128 are discovered in the process, add them to *NEW_EDGES, unless NEW_EDGES
1132 recursive_inlining (struct cgraph_edge
*edge
,
1133 VEC (cgraph_edge_p
, heap
) **new_edges
)
1135 int limit
= PARAM_VALUE (PARAM_MAX_INLINE_INSNS_RECURSIVE_AUTO
);
1137 struct cgraph_node
*node
;
1138 struct cgraph_edge
*e
;
1139 struct cgraph_node
*master_clone
= NULL
, *next
;
1143 node
= edge
->caller
;
1144 if (node
->global
.inlined_to
)
1145 node
= node
->global
.inlined_to
;
1147 if (DECL_DECLARED_INLINE_P (node
->symbol
.decl
))
1148 limit
= PARAM_VALUE (PARAM_MAX_INLINE_INSNS_RECURSIVE
);
1150 /* Make sure that function is small enough to be considered for inlining. */
1151 if (estimate_size_after_inlining (node
, edge
) >= limit
)
1153 heap
= fibheap_new ();
1154 lookup_recursive_calls (node
, node
, heap
);
1155 if (fibheap_empty (heap
))
1157 fibheap_delete (heap
);
1163 " Performing recursive inlining on %s\n",
1164 cgraph_node_name (node
));
1166 /* Do the inlining and update list of recursive call during process. */
1167 while (!fibheap_empty (heap
))
1169 struct cgraph_edge
*curr
1170 = (struct cgraph_edge
*) fibheap_extract_min (heap
);
1171 struct cgraph_node
*cnode
;
1173 if (estimate_size_after_inlining (node
, curr
) > limit
)
1176 if (!can_inline_edge_p (curr
, true))
1180 for (cnode
= curr
->caller
;
1181 cnode
->global
.inlined_to
; cnode
= cnode
->callers
->caller
)
1182 if (node
->symbol
.decl
1183 == cgraph_function_or_thunk_node (curr
->callee
, NULL
)->symbol
.decl
)
1186 if (!want_inline_self_recursive_call_p (curr
, node
, false, depth
))
1192 " Inlining call of depth %i", depth
);
1195 fprintf (dump_file
, " called approx. %.2f times per call",
1196 (double)curr
->count
/ node
->count
);
1198 fprintf (dump_file
, "\n");
1202 /* We need original clone to copy around. */
1203 master_clone
= cgraph_clone_node (node
, node
->symbol
.decl
,
1204 node
->count
, CGRAPH_FREQ_BASE
,
1206 for (e
= master_clone
->callees
; e
; e
= e
->next_callee
)
1207 if (!e
->inline_failed
)
1208 clone_inlined_nodes (e
, true, false, NULL
);
1211 cgraph_redirect_edge_callee (curr
, master_clone
);
1212 inline_call (curr
, false, new_edges
, &overall_size
);
1213 lookup_recursive_calls (node
, curr
->callee
, heap
);
1217 if (!fibheap_empty (heap
) && dump_file
)
1218 fprintf (dump_file
, " Recursive inlining growth limit met.\n");
1219 fibheap_delete (heap
);
1226 "\n Inlined %i times, "
1227 "body grown from size %i to %i, time %i to %i\n", n
,
1228 inline_summary (master_clone
)->size
, inline_summary (node
)->size
,
1229 inline_summary (master_clone
)->time
, inline_summary (node
)->time
);
1231 /* Remove master clone we used for inlining. We rely that clones inlined
1232 into master clone gets queued just before master clone so we don't
1234 for (node
= cgraph_first_function (); node
!= master_clone
;
1237 next
= cgraph_next_function (node
);
1238 if (node
->global
.inlined_to
== master_clone
)
1239 cgraph_remove_node (node
);
1241 cgraph_remove_node (master_clone
);
1246 /* Given whole compilation unit estimate of INSNS, compute how large we can
1247 allow the unit to grow. */
1250 compute_max_insns (int insns
)
1252 int max_insns
= insns
;
1253 if (max_insns
< PARAM_VALUE (PARAM_LARGE_UNIT_INSNS
))
1254 max_insns
= PARAM_VALUE (PARAM_LARGE_UNIT_INSNS
);
1256 return ((HOST_WIDEST_INT
) max_insns
1257 * (100 + PARAM_VALUE (PARAM_INLINE_UNIT_GROWTH
)) / 100);
1261 /* Compute badness of all edges in NEW_EDGES and add them to the HEAP. */
1264 add_new_edges_to_heap (fibheap_t heap
, VEC (cgraph_edge_p
, heap
) *new_edges
)
1266 while (VEC_length (cgraph_edge_p
, new_edges
) > 0)
1268 struct cgraph_edge
*edge
= VEC_pop (cgraph_edge_p
, new_edges
);
1270 gcc_assert (!edge
->aux
);
1271 if (edge
->inline_failed
1272 && can_inline_edge_p (edge
, true)
1273 && want_inline_small_function_p (edge
, true))
1274 edge
->aux
= fibheap_insert (heap
, edge_badness (edge
, false), edge
);
1279 /* We use greedy algorithm for inlining of small functions:
1280 All inline candidates are put into prioritized heap ordered in
1283 The inlining of small functions is bounded by unit growth parameters. */
1286 inline_small_functions (void)
1288 struct cgraph_node
*node
;
1289 struct cgraph_edge
*edge
;
1290 fibheap_t heap
= fibheap_new ();
1291 bitmap updated_nodes
= BITMAP_ALLOC (NULL
);
1292 int min_size
, max_size
;
1293 VEC (cgraph_edge_p
, heap
) *new_indirect_edges
= NULL
;
1294 int initial_size
= 0;
1296 if (flag_indirect_inlining
)
1297 new_indirect_edges
= VEC_alloc (cgraph_edge_p
, heap
, 8);
1301 "\nDeciding on inlining of small functions. Starting with size %i.\n",
1304 /* Compute overall unit size and other global parameters used by badness
1308 initialize_growth_caches ();
1310 FOR_EACH_DEFINED_FUNCTION (node
)
1311 if (!node
->global
.inlined_to
)
1313 if (cgraph_function_with_gimple_body_p (node
)
1314 || node
->thunk
.thunk_p
)
1316 struct inline_summary
*info
= inline_summary (node
);
1318 if (!DECL_EXTERNAL (node
->symbol
.decl
))
1319 initial_size
+= info
->size
;
1322 for (edge
= node
->callers
; edge
; edge
= edge
->next_caller
)
1323 if (max_count
< edge
->count
)
1324 max_count
= edge
->count
;
1327 overall_size
= initial_size
;
1328 max_size
= compute_max_insns (overall_size
);
1329 min_size
= overall_size
;
1331 /* Populate the heeap with all edges we might inline. */
1333 FOR_EACH_DEFINED_FUNCTION (node
)
1334 if (!node
->global
.inlined_to
)
1337 fprintf (dump_file
, "Enqueueing calls of %s/%i.\n",
1338 cgraph_node_name (node
), node
->uid
);
1340 for (edge
= node
->callers
; edge
; edge
= edge
->next_caller
)
1341 if (edge
->inline_failed
1342 && can_inline_edge_p (edge
, true)
1343 && want_inline_small_function_p (edge
, true)
1344 && edge
->inline_failed
)
1346 gcc_assert (!edge
->aux
);
1347 update_edge_key (heap
, edge
);
1351 gcc_assert (in_lto_p
1353 || (profile_info
&& flag_branch_probabilities
));
1355 while (!fibheap_empty (heap
))
1357 int old_size
= overall_size
;
1358 struct cgraph_node
*where
, *callee
;
1359 int badness
= fibheap_min_key (heap
);
1360 int current_badness
;
1364 edge
= (struct cgraph_edge
*) fibheap_extract_min (heap
);
1365 gcc_assert (edge
->aux
);
1367 if (!edge
->inline_failed
)
1370 /* Be sure that caches are maintained consistent.
1371 We can not make this ENABLE_CHECKING only because it cause different
1372 updates of the fibheap queue. */
1373 cached_badness
= edge_badness (edge
, false);
1374 reset_edge_growth_cache (edge
);
1375 reset_node_growth_cache (edge
->callee
);
1377 /* When updating the edge costs, we only decrease badness in the keys.
1378 Increases of badness are handled lazilly; when we see key with out
1379 of date value on it, we re-insert it now. */
1380 current_badness
= edge_badness (edge
, false);
1381 gcc_assert (cached_badness
== current_badness
);
1382 gcc_assert (current_badness
>= badness
);
1383 if (current_badness
!= badness
)
1385 edge
->aux
= fibheap_insert (heap
, current_badness
, edge
);
1389 if (!can_inline_edge_p (edge
, true))
1392 callee
= cgraph_function_or_thunk_node (edge
->callee
, NULL
);
1393 growth
= estimate_edge_growth (edge
);
1397 "\nConsidering %s with %i size\n",
1398 cgraph_node_name (callee
),
1399 inline_summary (callee
)->size
);
1401 " to be inlined into %s in %s:%i\n"
1402 " Estimated growth after inlined into all is %+i insns.\n"
1403 " Estimated badness is %i, frequency %.2f.\n",
1404 cgraph_node_name (edge
->caller
),
1405 flag_wpa
? "unknown"
1406 : gimple_filename ((const_gimple
) edge
->call_stmt
),
1408 : gimple_lineno ((const_gimple
) edge
->call_stmt
),
1409 estimate_growth (callee
),
1411 edge
->frequency
/ (double)CGRAPH_FREQ_BASE
);
1413 fprintf (dump_file
," Called "HOST_WIDEST_INT_PRINT_DEC
"x\n",
1415 if (dump_flags
& TDF_DETAILS
)
1416 edge_badness (edge
, true);
1419 if (overall_size
+ growth
> max_size
1420 && !DECL_DISREGARD_INLINE_LIMITS (callee
->symbol
.decl
))
1422 edge
->inline_failed
= CIF_INLINE_UNIT_GROWTH_LIMIT
;
1423 report_inline_failed_reason (edge
);
1427 if (!want_inline_small_function_p (edge
, true))
1430 /* Heuristics for inlining small functions works poorly for
1431 recursive calls where we do efect similar to loop unrolling.
1432 When inliing such edge seems profitable, leave decision on
1433 specific inliner. */
1434 if (cgraph_edge_recursive_p (edge
))
1436 where
= edge
->caller
;
1437 if (where
->global
.inlined_to
)
1438 where
= where
->global
.inlined_to
;
1439 if (!recursive_inlining (edge
,
1440 flag_indirect_inlining
1441 ? &new_indirect_edges
: NULL
))
1443 edge
->inline_failed
= CIF_RECURSIVE_INLINING
;
1446 reset_edge_caches (where
);
1447 /* Recursive inliner inlines all recursive calls of the function
1448 at once. Consequently we need to update all callee keys. */
1449 if (flag_indirect_inlining
)
1450 add_new_edges_to_heap (heap
, new_indirect_edges
);
1451 update_callee_keys (heap
, where
, updated_nodes
);
1455 struct cgraph_node
*outer_node
= NULL
;
1458 /* Consider the case where self recursive function A is inlined into B.
1459 This is desired optimization in some cases, since it leads to effect
1460 similar of loop peeling and we might completely optimize out the
1461 recursive call. However we must be extra selective. */
1463 where
= edge
->caller
;
1464 while (where
->global
.inlined_to
)
1466 if (where
->symbol
.decl
== callee
->symbol
.decl
)
1467 outer_node
= where
, depth
++;
1468 where
= where
->callers
->caller
;
1471 && !want_inline_self_recursive_call_p (edge
, outer_node
,
1475 = (DECL_DISREGARD_INLINE_LIMITS (edge
->callee
->symbol
.decl
)
1476 ? CIF_RECURSIVE_INLINING
: CIF_UNSPECIFIED
);
1479 else if (depth
&& dump_file
)
1480 fprintf (dump_file
, " Peeling recursion with depth %i\n", depth
);
1482 gcc_checking_assert (!callee
->global
.inlined_to
);
1483 inline_call (edge
, true, &new_indirect_edges
, &overall_size
);
1484 if (flag_indirect_inlining
)
1485 add_new_edges_to_heap (heap
, new_indirect_edges
);
1487 reset_edge_caches (edge
->callee
);
1488 reset_node_growth_cache (callee
);
1490 update_callee_keys (heap
, edge
->callee
, updated_nodes
);
1492 where
= edge
->caller
;
1493 if (where
->global
.inlined_to
)
1494 where
= where
->global
.inlined_to
;
1496 /* Our profitability metric can depend on local properties
1497 such as number of inlinable calls and size of the function body.
1498 After inlining these properties might change for the function we
1499 inlined into (since it's body size changed) and for the functions
1500 called by function we inlined (since number of it inlinable callers
1502 update_caller_keys (heap
, where
, updated_nodes
, NULL
);
1503 bitmap_clear (updated_nodes
);
1508 " Inlined into %s which now has time %i and size %i,"
1509 "net change of %+i.\n",
1510 cgraph_node_name (edge
->caller
),
1511 inline_summary (edge
->caller
)->time
,
1512 inline_summary (edge
->caller
)->size
,
1513 overall_size
- old_size
);
1515 if (min_size
> overall_size
)
1517 min_size
= overall_size
;
1518 max_size
= compute_max_insns (min_size
);
1521 fprintf (dump_file
, "New minimal size reached: %i\n", min_size
);
1525 free_growth_caches ();
1526 if (new_indirect_edges
)
1527 VEC_free (cgraph_edge_p
, heap
, new_indirect_edges
);
1528 fibheap_delete (heap
);
1531 "Unit growth for small function inlining: %i->%i (%i%%)\n",
1532 initial_size
, overall_size
,
1533 initial_size
? overall_size
* 100 / (initial_size
) - 100: 0);
1534 BITMAP_FREE (updated_nodes
);
1537 /* Flatten NODE. Performed both during early inlining and
1538 at IPA inlining time. */
1541 flatten_function (struct cgraph_node
*node
, bool early
)
1543 struct cgraph_edge
*e
;
1545 /* We shouldn't be called recursively when we are being processed. */
1546 gcc_assert (node
->symbol
.aux
== NULL
);
1548 node
->symbol
.aux
= (void *) node
;
1550 for (e
= node
->callees
; e
; e
= e
->next_callee
)
1552 struct cgraph_node
*orig_callee
;
1553 struct cgraph_node
*callee
= cgraph_function_or_thunk_node (e
->callee
, NULL
);
1555 /* We've hit cycle? It is time to give up. */
1556 if (callee
->symbol
.aux
)
1560 "Not inlining %s into %s to avoid cycle.\n",
1561 xstrdup (cgraph_node_name (callee
)),
1562 xstrdup (cgraph_node_name (e
->caller
)));
1563 e
->inline_failed
= CIF_RECURSIVE_INLINING
;
1567 /* When the edge is already inlined, we just need to recurse into
1568 it in order to fully flatten the leaves. */
1569 if (!e
->inline_failed
)
1571 flatten_function (callee
, early
);
1575 /* Flatten attribute needs to be processed during late inlining. For
1576 extra code quality we however do flattening during early optimization,
1579 ? !can_inline_edge_p (e
, true)
1580 : !can_early_inline_edge_p (e
))
1583 if (cgraph_edge_recursive_p (e
))
1586 fprintf (dump_file
, "Not inlining: recursive call.\n");
1590 if (gimple_in_ssa_p (DECL_STRUCT_FUNCTION (node
->symbol
.decl
))
1591 != gimple_in_ssa_p (DECL_STRUCT_FUNCTION (callee
->symbol
.decl
)))
1594 fprintf (dump_file
, "Not inlining: SSA form does not match.\n");
1598 /* Inline the edge and flatten the inline clone. Avoid
1599 recursing through the original node if the node was cloned. */
1601 fprintf (dump_file
, " Inlining %s into %s.\n",
1602 xstrdup (cgraph_node_name (callee
)),
1603 xstrdup (cgraph_node_name (e
->caller
)));
1604 orig_callee
= callee
;
1605 inline_call (e
, true, NULL
, NULL
);
1606 if (e
->callee
!= orig_callee
)
1607 orig_callee
->symbol
.aux
= (void *) node
;
1608 flatten_function (e
->callee
, early
);
1609 if (e
->callee
!= orig_callee
)
1610 orig_callee
->symbol
.aux
= NULL
;
1613 node
->symbol
.aux
= NULL
;
1616 /* Decide on the inlining. We do so in the topological order to avoid
1617 expenses on updating data structures. */
1622 struct cgraph_node
*node
;
1624 struct cgraph_node
**order
=
1625 XCNEWVEC (struct cgraph_node
*, cgraph_n_nodes
);
1628 if (in_lto_p
&& optimize
)
1629 ipa_update_after_lto_read ();
1632 dump_inline_summaries (dump_file
);
1634 nnodes
= ipa_reverse_postorder (order
);
1636 FOR_EACH_FUNCTION (node
)
1637 node
->symbol
.aux
= 0;
1640 fprintf (dump_file
, "\nFlattening functions:\n");
1642 /* In the first pass handle functions to be flattened. Do this with
1643 a priority so none of our later choices will make this impossible. */
1644 for (i
= nnodes
- 1; i
>= 0; i
--)
1648 /* Handle nodes to be flattened.
1649 Ideally when processing callees we stop inlining at the
1650 entry of cycles, possibly cloning that entry point and
1651 try to flatten itself turning it into a self-recursive
1653 if (lookup_attribute ("flatten",
1654 DECL_ATTRIBUTES (node
->symbol
.decl
)) != NULL
)
1658 "Flattening %s\n", cgraph_node_name (node
));
1659 flatten_function (node
, false);
1663 inline_small_functions ();
1664 symtab_remove_unreachable_nodes (true, dump_file
);
1667 /* We already perform some inlining of functions called once during
1668 inlining small functions above. After unreachable nodes are removed,
1669 we still might do a quick check that nothing new is found. */
1670 if (flag_inline_functions_called_once
)
1674 fprintf (dump_file
, "\nDeciding on functions called once:\n");
1676 /* Inlining one function called once has good chance of preventing
1677 inlining other function into the same callee. Ideally we should
1678 work in priority order, but probably inlining hot functions first
1679 is good cut without the extra pain of maintaining the queue.
1681 ??? this is not really fitting the bill perfectly: inlining function
1682 into callee often leads to better optimization of callee due to
1683 increased context for optimization.
1684 For example if main() function calls a function that outputs help
1685 and then function that does the main optmization, we should inline
1686 the second with priority even if both calls are cold by themselves.
1688 We probably want to implement new predicate replacing our use of
1689 maybe_hot_edge interpreted as maybe_hot_edge || callee is known
1691 for (cold
= 0; cold
<= 1; cold
++)
1693 FOR_EACH_DEFINED_FUNCTION (node
)
1695 if (want_inline_function_called_once_p (node
)
1697 || cgraph_maybe_hot_edge_p (node
->callers
)))
1699 struct cgraph_node
*caller
= node
->callers
->caller
;
1704 "\nInlining %s size %i.\n",
1705 cgraph_node_name (node
),
1706 inline_summary (node
)->size
);
1708 " Called once from %s %i insns.\n",
1709 cgraph_node_name (node
->callers
->caller
),
1710 inline_summary (node
->callers
->caller
)->size
);
1713 inline_call (node
->callers
, true, NULL
, NULL
);
1716 " Inlined into %s which now has %i size\n",
1717 cgraph_node_name (caller
),
1718 inline_summary (caller
)->size
);
1724 /* Free ipa-prop structures if they are no longer needed. */
1726 ipa_free_all_structures_after_iinln ();
1730 "\nInlined %i calls, eliminated %i functions\n\n",
1731 ncalls_inlined
, nfunctions_inlined
);
1734 dump_inline_summaries (dump_file
);
1735 /* In WPA we use inline summaries for partitioning process. */
1737 inline_free_summary ();
1741 /* Inline always-inline function calls in NODE. */
1744 inline_always_inline_functions (struct cgraph_node
*node
)
1746 struct cgraph_edge
*e
;
1747 bool inlined
= false;
1749 for (e
= node
->callees
; e
; e
= e
->next_callee
)
1751 struct cgraph_node
*callee
= cgraph_function_or_thunk_node (e
->callee
, NULL
);
1752 if (!DECL_DISREGARD_INLINE_LIMITS (callee
->symbol
.decl
))
1755 if (cgraph_edge_recursive_p (e
))
1758 fprintf (dump_file
, " Not inlining recursive call to %s.\n",
1759 cgraph_node_name (e
->callee
));
1760 e
->inline_failed
= CIF_RECURSIVE_INLINING
;
1764 if (!can_early_inline_edge_p (e
))
1768 fprintf (dump_file
, " Inlining %s into %s (always_inline).\n",
1769 xstrdup (cgraph_node_name (e
->callee
)),
1770 xstrdup (cgraph_node_name (e
->caller
)));
1771 inline_call (e
, true, NULL
, NULL
);
1778 /* Decide on the inlining. We do so in the topological order to avoid
1779 expenses on updating data structures. */
1782 early_inline_small_functions (struct cgraph_node
*node
)
1784 struct cgraph_edge
*e
;
1785 bool inlined
= false;
1787 for (e
= node
->callees
; e
; e
= e
->next_callee
)
1789 struct cgraph_node
*callee
= cgraph_function_or_thunk_node (e
->callee
, NULL
);
1790 if (!inline_summary (callee
)->inlinable
1791 || !e
->inline_failed
)
1794 /* Do not consider functions not declared inline. */
1795 if (!DECL_DECLARED_INLINE_P (callee
->symbol
.decl
)
1796 && !flag_inline_small_functions
1797 && !flag_inline_functions
)
1801 fprintf (dump_file
, "Considering inline candidate %s.\n",
1802 cgraph_node_name (callee
));
1804 if (!can_early_inline_edge_p (e
))
1807 if (cgraph_edge_recursive_p (e
))
1810 fprintf (dump_file
, " Not inlining: recursive call.\n");
1814 if (!want_early_inline_function_p (e
))
1818 fprintf (dump_file
, " Inlining %s into %s.\n",
1819 xstrdup (cgraph_node_name (callee
)),
1820 xstrdup (cgraph_node_name (e
->caller
)));
1821 inline_call (e
, true, NULL
, NULL
);
1828 /* Do inlining of small functions. Doing so early helps profiling and other
1829 passes to be somewhat more effective and avoids some code duplication in
1830 later real inlining pass for testcases with very many function calls. */
1832 early_inliner (void)
1834 struct cgraph_node
*node
= cgraph_get_node (current_function_decl
);
1835 struct cgraph_edge
*edge
;
1836 unsigned int todo
= 0;
1838 bool inlined
= false;
1843 /* Do nothing if datastructures for ipa-inliner are already computed. This
1844 happens when some pass decides to construct new function and
1845 cgraph_add_new_function calls lowering passes and early optimization on
1846 it. This may confuse ourself when early inliner decide to inline call to
1847 function clone, because function clones don't have parameter list in
1848 ipa-prop matching their signature. */
1849 if (ipa_node_params_vector
)
1852 #ifdef ENABLE_CHECKING
1853 verify_cgraph_node (node
);
1856 /* Even when not optimizing or not inlining inline always-inline
1858 inlined
= inline_always_inline_functions (node
);
1862 || !flag_early_inlining
1863 /* Never inline regular functions into always-inline functions
1864 during incremental inlining. This sucks as functions calling
1865 always inline functions will get less optimized, but at the
1866 same time inlining of functions calling always inline
1867 function into an always inline function might introduce
1868 cycles of edges to be always inlined in the callgraph.
1870 We might want to be smarter and just avoid this type of inlining. */
1871 || DECL_DISREGARD_INLINE_LIMITS (node
->symbol
.decl
))
1873 else if (lookup_attribute ("flatten",
1874 DECL_ATTRIBUTES (node
->symbol
.decl
)) != NULL
)
1876 /* When the function is marked to be flattened, recursively inline
1880 "Flattening %s\n", cgraph_node_name (node
));
1881 flatten_function (node
, true);
1886 /* We iterate incremental inlining to get trivial cases of indirect
1888 while (iterations
< PARAM_VALUE (PARAM_EARLY_INLINER_MAX_ITERATIONS
)
1889 && early_inline_small_functions (node
))
1891 timevar_push (TV_INTEGRATION
);
1892 todo
|= optimize_inline_calls (current_function_decl
);
1894 /* Technically we ought to recompute inline parameters so the new
1895 iteration of early inliner works as expected. We however have
1896 values approximately right and thus we only need to update edge
1897 info that might be cleared out for newly discovered edges. */
1898 for (edge
= node
->callees
; edge
; edge
= edge
->next_callee
)
1900 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
1902 = estimate_num_insns (edge
->call_stmt
, &eni_size_weights
);
1904 = estimate_num_insns (edge
->call_stmt
, &eni_time_weights
);
1905 if (edge
->callee
->symbol
.decl
1906 && !gimple_check_call_matching_types (edge
->call_stmt
,
1907 edge
->callee
->symbol
.decl
))
1908 edge
->call_stmt_cannot_inline_p
= true;
1910 timevar_pop (TV_INTEGRATION
);
1915 fprintf (dump_file
, "Iterations: %i\n", iterations
);
1920 timevar_push (TV_INTEGRATION
);
1921 todo
|= optimize_inline_calls (current_function_decl
);
1922 timevar_pop (TV_INTEGRATION
);
1925 cfun
->always_inline_functions_inlined
= true;
1930 struct gimple_opt_pass pass_early_inline
=
1934 "einline", /* name */
1936 early_inliner
, /* execute */
1939 0, /* static_pass_number */
1940 TV_INLINE_HEURISTICS
, /* tv_id */
1941 PROP_ssa
, /* properties_required */
1942 0, /* properties_provided */
1943 0, /* properties_destroyed */
1944 0, /* todo_flags_start */
1945 0 /* todo_flags_finish */
1950 /* When to run IPA inlining. Inlining of always-inline functions
1951 happens during early inlining.
1953 Enable inlining unconditoinally at -flto. We need size estimates to
1954 drive partitioning. */
1957 gate_ipa_inline (void)
1959 return optimize
|| flag_lto
|| flag_wpa
;
1962 struct ipa_opt_pass_d pass_ipa_inline
=
1966 "inline", /* name */
1967 gate_ipa_inline
, /* gate */
1968 ipa_inline
, /* execute */
1971 0, /* static_pass_number */
1972 TV_INLINE_HEURISTICS
, /* tv_id */
1973 0, /* properties_required */
1974 0, /* properties_provided */
1975 0, /* properties_destroyed */
1976 TODO_remove_functions
, /* todo_flags_finish */
1978 | TODO_remove_functions
| TODO_ggc_collect
/* todo_flags_finish */
1980 inline_generate_summary
, /* generate_summary */
1981 inline_write_summary
, /* write_summary */
1982 inline_read_summary
, /* read_summary */
1983 NULL
, /* write_optimization_summary */
1984 NULL
, /* read_optimization_summary */
1985 NULL
, /* stmt_fixup */
1987 inline_transform
, /* function_transform */
1988 NULL
, /* variable_transform */