1 /* Inlining decision heuristics.
2 Copyright (C) 2003-2013 Free Software Foundation, Inc.
3 Contributed by Jan Hubicka
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
21 /* Inlining decision heuristics
23 The implementation of inliner is organized as follows:
25 inlining heuristics limits
27 can_inline_edge_p allow to check that particular inlining is allowed
28 by the limits specified by user (allowed function growth, growth and so
31 Functions are inlined when it is obvious the result is profitable (such
32 as functions called once or when inlining reduce code size).
33 In addition to that we perform inlining of small functions and recursive
38 The inliner itself is split into two passes:
42 Simple local inlining pass inlining callees into current function.
43 This pass makes no use of whole unit analysis and thus it can do only
44 very simple decisions based on local properties.
46 The strength of the pass is that it is run in topological order
47 (reverse postorder) on the callgraph. Functions are converted into SSA
48 form just before this pass and optimized subsequently. As a result, the
49 callees of the function seen by the early inliner was already optimized
50 and results of early inlining adds a lot of optimization opportunities
51 for the local optimization.
53 The pass handle the obvious inlining decisions within the compilation
54 unit - inlining auto inline functions, inlining for size and
57 main strength of the pass is the ability to eliminate abstraction
58 penalty in C++ code (via combination of inlining and early
59 optimization) and thus improve quality of analysis done by real IPA
62 Because of lack of whole unit knowledge, the pass can not really make
63 good code size/performance tradeoffs. It however does very simple
64 speculative inlining allowing code size to grow by
65 EARLY_INLINING_INSNS when callee is leaf function. In this case the
66 optimizations performed later are very likely to eliminate the cost.
70 This is the real inliner able to handle inlining with whole program
71 knowledge. It performs following steps:
73 1) inlining of small functions. This is implemented by greedy
74 algorithm ordering all inlinable cgraph edges by their badness and
75 inlining them in this order as long as inline limits allows doing so.
77 This heuristics is not very good on inlining recursive calls. Recursive
78 calls can be inlined with results similar to loop unrolling. To do so,
79 special purpose recursive inliner is executed on function when
80 recursive edge is met as viable candidate.
82 2) Unreachable functions are removed from callgraph. Inlining leads
83 to devirtualization and other modification of callgraph so functions
84 may become unreachable during the process. Also functions declared as
85 extern inline or virtual functions are removed, since after inlining
86 we no longer need the offline bodies.
88 3) Functions called once and not exported from the unit are inlined.
89 This should almost always lead to reduction of code size by eliminating
90 the need for offline copy of the function. */
94 #include "coretypes.h"
97 #include "tree-inline.h"
98 #include "langhooks.h"
101 #include "diagnostic.h"
102 #include "gimple-pretty-print.h"
106 #include "tree-pass.h"
107 #include "coverage.h"
110 #include "tree-ssa.h"
111 #include "ipa-prop.h"
114 #include "ipa-inline.h"
115 #include "ipa-utils.h"
118 /* Statistics we collect about inlining algorithm. */
119 static int overall_size
;
120 static gcov_type max_count
;
121 static sreal max_count_real
, max_relbenefit_real
, half_int_min_real
;
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
= cgraph_function_or_thunk_node (e
->callee
, NULL
);
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 xstrdup (cgraph_node_name (e
->caller
)), e
->caller
->symbol
.order
,
224 xstrdup (cgraph_node_name (e
->callee
)), e
->callee
->symbol
.order
,
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.
236 if DISREGARD_LIMITES is true, ignore size limits.*/
239 can_inline_edge_p (struct cgraph_edge
*e
, bool report
,
240 bool disregard_limits
= false)
242 bool inlinable
= true;
243 enum availability avail
;
244 struct cgraph_node
*callee
245 = cgraph_function_or_thunk_node (e
->callee
, &avail
);
246 tree caller_tree
= DECL_FUNCTION_SPECIFIC_OPTIMIZATION (e
->caller
->symbol
.decl
);
248 = callee
? DECL_FUNCTION_SPECIFIC_OPTIMIZATION (callee
->symbol
.decl
) : NULL
;
249 struct function
*caller_cfun
= DECL_STRUCT_FUNCTION (e
->caller
->symbol
.decl
);
250 struct function
*callee_cfun
251 = callee
? DECL_STRUCT_FUNCTION (callee
->symbol
.decl
) : NULL
;
253 if (!caller_cfun
&& e
->caller
->clone_of
)
254 caller_cfun
= DECL_STRUCT_FUNCTION (e
->caller
->clone_of
->symbol
.decl
);
256 if (!callee_cfun
&& callee
&& callee
->clone_of
)
257 callee_cfun
= DECL_STRUCT_FUNCTION (callee
->clone_of
->symbol
.decl
);
259 gcc_assert (e
->inline_failed
);
261 if (!callee
|| !callee
->symbol
.definition
)
263 e
->inline_failed
= CIF_BODY_NOT_AVAILABLE
;
266 else if (!inline_summary (callee
)->inlinable
)
268 e
->inline_failed
= CIF_FUNCTION_NOT_INLINABLE
;
271 else if (avail
<= AVAIL_OVERWRITABLE
)
273 e
->inline_failed
= CIF_OVERWRITABLE
;
276 else if (e
->call_stmt_cannot_inline_p
)
278 e
->inline_failed
= CIF_MISMATCHED_ARGUMENTS
;
281 /* Don't inline if the functions have different EH personalities. */
282 else if (DECL_FUNCTION_PERSONALITY (e
->caller
->symbol
.decl
)
283 && DECL_FUNCTION_PERSONALITY (callee
->symbol
.decl
)
284 && (DECL_FUNCTION_PERSONALITY (e
->caller
->symbol
.decl
)
285 != DECL_FUNCTION_PERSONALITY (callee
->symbol
.decl
)))
287 e
->inline_failed
= CIF_EH_PERSONALITY
;
290 /* TM pure functions should not be inlined into non-TM_pure
292 else if (is_tm_pure (callee
->symbol
.decl
)
293 && !is_tm_pure (e
->caller
->symbol
.decl
))
295 e
->inline_failed
= CIF_UNSPECIFIED
;
298 /* Don't inline if the callee can throw non-call exceptions but the
300 FIXME: this is obviously wrong for LTO where STRUCT_FUNCTION is missing.
301 Move the flag into cgraph node or mirror it in the inline summary. */
302 else if (callee_cfun
&& callee_cfun
->can_throw_non_call_exceptions
303 && !(caller_cfun
&& caller_cfun
->can_throw_non_call_exceptions
))
305 e
->inline_failed
= CIF_NON_CALL_EXCEPTIONS
;
308 /* Check compatibility of target optimization options. */
309 else if (!targetm
.target_option
.can_inline_p (e
->caller
->symbol
.decl
,
310 callee
->symbol
.decl
))
312 e
->inline_failed
= CIF_TARGET_OPTION_MISMATCH
;
315 /* Check if caller growth allows the inlining. */
316 else if (!DECL_DISREGARD_INLINE_LIMITS (callee
->symbol
.decl
)
318 && !lookup_attribute ("flatten",
320 (e
->caller
->global
.inlined_to
321 ? e
->caller
->global
.inlined_to
->symbol
.decl
322 : e
->caller
->symbol
.decl
))
323 && !caller_growth_limits (e
))
325 /* Don't inline a function with a higher optimization level than the
326 caller. FIXME: this is really just tip of iceberg of handling
327 optimization attribute. */
328 else if (caller_tree
!= callee_tree
)
330 struct cl_optimization
*caller_opt
331 = TREE_OPTIMIZATION ((caller_tree
)
333 : optimization_default_node
);
335 struct cl_optimization
*callee_opt
336 = TREE_OPTIMIZATION ((callee_tree
)
338 : optimization_default_node
);
340 if (((caller_opt
->x_optimize
> callee_opt
->x_optimize
)
341 || (caller_opt
->x_optimize_size
!= callee_opt
->x_optimize_size
))
342 /* gcc.dg/pr43564.c. Look at forced inline even in -O0. */
343 && !DECL_DISREGARD_INLINE_LIMITS (e
->callee
->symbol
.decl
))
345 e
->inline_failed
= CIF_OPTIMIZATION_MISMATCH
;
350 if (!inlinable
&& report
)
351 report_inline_failed_reason (e
);
356 /* Return true if the edge E is inlinable during early inlining. */
359 can_early_inline_edge_p (struct cgraph_edge
*e
)
361 struct cgraph_node
*callee
= cgraph_function_or_thunk_node (e
->callee
,
363 /* Early inliner might get called at WPA stage when IPA pass adds new
364 function. In this case we can not really do any of early inlining
365 because function bodies are missing. */
366 if (!gimple_has_body_p (callee
->symbol
.decl
))
368 e
->inline_failed
= CIF_BODY_NOT_AVAILABLE
;
371 /* In early inliner some of callees may not be in SSA form yet
372 (i.e. the callgraph is cyclic and we did not process
373 the callee by early inliner, yet). We don't have CIF code for this
374 case; later we will re-do the decision in the real inliner. */
375 if (!gimple_in_ssa_p (DECL_STRUCT_FUNCTION (e
->caller
->symbol
.decl
))
376 || !gimple_in_ssa_p (DECL_STRUCT_FUNCTION (callee
->symbol
.decl
)))
379 fprintf (dump_file
, " edge not inlinable: not in SSA form\n");
382 if (!can_inline_edge_p (e
, true))
388 /* Return number of calls in N. Ignore cheap builtins. */
391 num_calls (struct cgraph_node
*n
)
393 struct cgraph_edge
*e
;
396 for (e
= n
->callees
; e
; e
= e
->next_callee
)
397 if (!is_inexpensive_builtin (e
->callee
->symbol
.decl
))
403 /* Return true if we are interested in inlining small function. */
406 want_early_inline_function_p (struct cgraph_edge
*e
)
408 bool want_inline
= true;
409 struct cgraph_node
*callee
= cgraph_function_or_thunk_node (e
->callee
, NULL
);
411 if (DECL_DISREGARD_INLINE_LIMITS (callee
->symbol
.decl
))
413 else if (!DECL_DECLARED_INLINE_P (callee
->symbol
.decl
)
414 && !flag_inline_small_functions
)
416 e
->inline_failed
= CIF_FUNCTION_NOT_INLINE_CANDIDATE
;
417 report_inline_failed_reason (e
);
422 int growth
= estimate_edge_growth (e
);
427 else if (!cgraph_maybe_hot_edge_p (e
)
431 fprintf (dump_file
, " will not early inline: %s/%i->%s/%i, "
432 "call is cold and code would grow by %i\n",
433 xstrdup (cgraph_node_name (e
->caller
)),
434 e
->caller
->symbol
.order
,
435 xstrdup (cgraph_node_name (callee
)), callee
->symbol
.order
,
439 else if (growth
> PARAM_VALUE (PARAM_EARLY_INLINING_INSNS
))
442 fprintf (dump_file
, " will not early inline: %s/%i->%s/%i, "
443 "growth %i exceeds --param early-inlining-insns\n",
444 xstrdup (cgraph_node_name (e
->caller
)),
445 e
->caller
->symbol
.order
,
446 xstrdup (cgraph_node_name (callee
)), callee
->symbol
.order
,
450 else if ((n
= num_calls (callee
)) != 0
451 && growth
* (n
+ 1) > PARAM_VALUE (PARAM_EARLY_INLINING_INSNS
))
454 fprintf (dump_file
, " will not early inline: %s/%i->%s/%i, "
455 "growth %i exceeds --param early-inlining-insns "
456 "divided by number of calls\n",
457 xstrdup (cgraph_node_name (e
->caller
)),
458 e
->caller
->symbol
.order
,
459 xstrdup (cgraph_node_name (callee
)), callee
->symbol
.order
,
467 /* Compute time of the edge->caller + edge->callee execution when inlining
471 compute_uninlined_call_time (struct inline_summary
*callee_info
,
472 struct cgraph_edge
*edge
)
474 gcov_type uninlined_call_time
=
475 RDIV ((gcov_type
)callee_info
->time
* MAX (edge
->frequency
, 1),
477 gcov_type caller_time
= inline_summary (edge
->caller
->global
.inlined_to
478 ? edge
->caller
->global
.inlined_to
479 : edge
->caller
)->time
;
480 return uninlined_call_time
+ caller_time
;
483 /* Same as compute_uinlined_call_time but compute time when inlining
487 compute_inlined_call_time (struct cgraph_edge
*edge
,
490 gcov_type caller_time
= inline_summary (edge
->caller
->global
.inlined_to
491 ? edge
->caller
->global
.inlined_to
492 : edge
->caller
)->time
;
493 gcov_type time
= (caller_time
494 + RDIV (((gcov_type
) edge_time
495 - inline_edge_summary (edge
)->call_stmt_time
)
496 * MAX (edge
->frequency
, 1), CGRAPH_FREQ_BASE
));
497 /* Possible one roundoff error, but watch for overflows. */
498 gcc_checking_assert (time
>= INT_MIN
/ 2);
504 /* Return true if the speedup for inlining E is bigger than
505 PARAM_MAX_INLINE_MIN_SPEEDUP. */
508 big_speedup_p (struct cgraph_edge
*e
)
510 gcov_type time
= compute_uninlined_call_time (inline_summary (e
->callee
),
512 gcov_type inlined_time
= compute_inlined_call_time (e
,
513 estimate_edge_time (e
));
514 if (time
- inlined_time
515 > RDIV (time
* PARAM_VALUE (PARAM_INLINE_MIN_SPEEDUP
), 100))
520 /* Return true if we are interested in inlining small function.
521 When REPORT is true, report reason to dump file. */
524 want_inline_small_function_p (struct cgraph_edge
*e
, bool report
)
526 bool want_inline
= true;
527 struct cgraph_node
*callee
= cgraph_function_or_thunk_node (e
->callee
, NULL
);
529 if (DECL_DISREGARD_INLINE_LIMITS (callee
->symbol
.decl
))
531 else if (!DECL_DECLARED_INLINE_P (callee
->symbol
.decl
)
532 && !flag_inline_small_functions
)
534 e
->inline_failed
= CIF_FUNCTION_NOT_INLINE_CANDIDATE
;
539 int growth
= estimate_edge_growth (e
);
540 inline_hints hints
= estimate_edge_hints (e
);
541 bool big_speedup
= big_speedup_p (e
);
545 /* Apply MAX_INLINE_INSNS_SINGLE limit. Do not do so when
546 hints suggests that inlining given function is very profitable. */
547 else if (DECL_DECLARED_INLINE_P (callee
->symbol
.decl
)
548 && growth
>= MAX_INLINE_INSNS_SINGLE
550 && !(hints
& (INLINE_HINT_indirect_call
551 | INLINE_HINT_loop_iterations
552 | INLINE_HINT_array_index
553 | INLINE_HINT_loop_stride
)))
555 e
->inline_failed
= CIF_MAX_INLINE_INSNS_SINGLE_LIMIT
;
558 /* Before giving up based on fact that caller size will grow, allow
559 functions that are called few times and eliminating the offline
560 copy will lead to overall code size reduction.
561 Not all of these will be handled by subsequent inlining of functions
562 called once: in particular weak functions are not handled or funcitons
563 that inline to multiple calls but a lot of bodies is optimized out.
564 Finally we want to inline earlier to allow inlining of callbacks.
566 This is slightly wrong on aggressive side: it is entirely possible
567 that function is called many times with a context where inlining
568 reduces code size and few times with a context where inlining increase
569 code size. Resoluting growth estimate will be negative even if it
570 would make more sense to keep offline copy and do not inline into the
571 call sites that makes the code size grow.
573 When badness orders the calls in a way that code reducing calls come
574 first, this situation is not a problem at all: after inlining all
575 "good" calls, we will realize that keeping the function around is
577 else if (growth
<= MAX_INLINE_INSNS_SINGLE
578 /* Unlike for functions called once, we play unsafe with
579 COMDATs. We can allow that since we know functions
580 in consideration are small (and thus risk is small) and
581 moreover grow estimates already accounts that COMDAT
582 functions may or may not disappear when eliminated from
583 current unit. With good probability making aggressive
584 choice in all units is going to make overall program
587 Consequently we ask cgraph_can_remove_if_no_direct_calls_p
589 cgraph_will_be_removed_from_program_if_no_direct_calls */
590 && !DECL_EXTERNAL (callee
->symbol
.decl
)
591 && cgraph_can_remove_if_no_direct_calls_p (callee
)
592 && estimate_growth (callee
) <= 0)
594 else if (!DECL_DECLARED_INLINE_P (callee
->symbol
.decl
)
595 && !flag_inline_functions
)
597 e
->inline_failed
= CIF_NOT_DECLARED_INLINED
;
600 /* Apply MAX_INLINE_INSNS_AUTO limit for functions not declared inline
601 Upgrade it to MAX_INLINE_INSNS_SINGLE when hints suggests that
602 inlining given function is very profitable. */
603 else if (!DECL_DECLARED_INLINE_P (callee
->symbol
.decl
)
605 && growth
>= ((hints
& (INLINE_HINT_indirect_call
606 | INLINE_HINT_loop_iterations
607 | INLINE_HINT_array_index
608 | INLINE_HINT_loop_stride
))
609 ? MAX (MAX_INLINE_INSNS_AUTO
,
610 MAX_INLINE_INSNS_SINGLE
)
611 : MAX_INLINE_INSNS_AUTO
))
613 e
->inline_failed
= CIF_MAX_INLINE_INSNS_AUTO_LIMIT
;
616 /* If call is cold, do not inline when function body would grow. */
617 else if (!cgraph_maybe_hot_edge_p (e
))
619 e
->inline_failed
= CIF_UNLIKELY_CALL
;
623 if (!want_inline
&& report
)
624 report_inline_failed_reason (e
);
628 /* EDGE is self recursive edge.
629 We hand two cases - when function A is inlining into itself
630 or when function A is being inlined into another inliner copy of function
633 In first case OUTER_NODE points to the toplevel copy of A, while
634 in the second case OUTER_NODE points to the outermost copy of A in B.
636 In both cases we want to be extra selective since
637 inlining the call will just introduce new recursive calls to appear. */
640 want_inline_self_recursive_call_p (struct cgraph_edge
*edge
,
641 struct cgraph_node
*outer_node
,
645 char const *reason
= NULL
;
646 bool want_inline
= true;
647 int caller_freq
= CGRAPH_FREQ_BASE
;
648 int max_depth
= PARAM_VALUE (PARAM_MAX_INLINE_RECURSIVE_DEPTH_AUTO
);
650 if (DECL_DECLARED_INLINE_P (edge
->caller
->symbol
.decl
))
651 max_depth
= PARAM_VALUE (PARAM_MAX_INLINE_RECURSIVE_DEPTH
);
653 if (!cgraph_maybe_hot_edge_p (edge
))
655 reason
= "recursive call is cold";
658 else if (max_count
&& !outer_node
->count
)
660 reason
= "not executed in profile";
663 else if (depth
> max_depth
)
665 reason
= "--param max-inline-recursive-depth exceeded.";
669 if (outer_node
->global
.inlined_to
)
670 caller_freq
= outer_node
->callers
->frequency
;
674 /* Inlining of self recursive function into copy of itself within other function
675 is transformation similar to loop peeling.
677 Peeling is profitable if we can inline enough copies to make probability
678 of actual call to the self recursive function very small. Be sure that
679 the probability of recursion is small.
681 We ensure that the frequency of recursing is at most 1 - (1/max_depth).
682 This way the expected number of recision is at most max_depth. */
685 int max_prob
= CGRAPH_FREQ_BASE
- ((CGRAPH_FREQ_BASE
+ max_depth
- 1)
688 for (i
= 1; i
< depth
; i
++)
689 max_prob
= max_prob
* max_prob
/ CGRAPH_FREQ_BASE
;
691 && (edge
->count
* CGRAPH_FREQ_BASE
/ outer_node
->count
694 reason
= "profile of recursive call is too large";
698 && (edge
->frequency
* CGRAPH_FREQ_BASE
/ caller_freq
701 reason
= "frequency of recursive call is too large";
705 /* Recursive inlining, i.e. equivalent of unrolling, is profitable if recursion
706 depth is large. We reduce function call overhead and increase chances that
707 things fit in hardware return predictor.
709 Recursive inlining might however increase cost of stack frame setup
710 actually slowing down functions whose recursion tree is wide rather than
713 Deciding reliably on when to do recursive inlining without profile feedback
714 is tricky. For now we disable recursive inlining when probability of self
717 Recursive inlining of self recursive call within loop also results in large loop
718 depths that generally optimize badly. We may want to throttle down inlining
719 in those cases. In particular this seems to happen in one of libstdc++ rb tree
724 && (edge
->count
* 100 / outer_node
->count
725 <= PARAM_VALUE (PARAM_MIN_INLINE_RECURSIVE_PROBABILITY
)))
727 reason
= "profile of recursive call is too small";
731 && (edge
->frequency
* 100 / caller_freq
732 <= PARAM_VALUE (PARAM_MIN_INLINE_RECURSIVE_PROBABILITY
)))
734 reason
= "frequency of recursive call is too small";
738 if (!want_inline
&& dump_file
)
739 fprintf (dump_file
, " not inlining recursively: %s\n", reason
);
743 /* Return true when NODE has caller other than EDGE.
744 Worker for cgraph_for_node_and_aliases. */
747 check_caller_edge (struct cgraph_node
*node
, void *edge
)
749 return (node
->callers
750 && node
->callers
!= edge
);
753 /* If NODE has a caller, return true. */
756 has_caller_p (struct cgraph_node
*node
, void *data ATTRIBUTE_UNUSED
)
763 /* Decide if inlining NODE would reduce unit size by eliminating
764 the offline copy of function.
765 When COLD is true the cold calls are considered, too. */
768 want_inline_function_to_all_callers_p (struct cgraph_node
*node
, bool cold
)
770 struct cgraph_node
*function
= cgraph_function_or_thunk_node (node
, NULL
);
771 struct cgraph_edge
*e
;
772 bool has_hot_call
= false;
774 /* Does it have callers? */
775 if (!cgraph_for_node_and_aliases (node
, has_caller_p
, NULL
, true))
777 /* Already inlined? */
778 if (function
->global
.inlined_to
)
780 if (cgraph_function_or_thunk_node (node
, NULL
) != node
)
782 /* Inlining into all callers would increase size? */
783 if (estimate_growth (node
) > 0)
785 /* Maybe other aliases has more direct calls. */
786 if (cgraph_for_node_and_aliases (node
, check_caller_edge
, node
->callers
, true))
788 /* All inlines must be possible. */
789 for (e
= node
->callers
; e
; e
= e
->next_caller
)
791 if (!can_inline_edge_p (e
, true))
793 if (!has_hot_call
&& cgraph_maybe_hot_edge_p (e
))
797 if (!cold
&& !has_hot_call
)
802 #define RELATIVE_TIME_BENEFIT_RANGE (INT_MAX / 64)
804 /* Return relative time improvement for inlining EDGE in range
805 1...RELATIVE_TIME_BENEFIT_RANGE */
808 relative_time_benefit (struct inline_summary
*callee_info
,
809 struct cgraph_edge
*edge
,
812 gcov_type relbenefit
;
813 gcov_type uninlined_call_time
= compute_uninlined_call_time (callee_info
, edge
);
814 gcov_type inlined_call_time
= compute_inlined_call_time (edge
, edge_time
);
816 /* Inlining into extern inline function is not a win. */
817 if (DECL_EXTERNAL (edge
->caller
->global
.inlined_to
818 ? edge
->caller
->global
.inlined_to
->symbol
.decl
819 : edge
->caller
->symbol
.decl
))
822 /* Watch overflows. */
823 gcc_checking_assert (uninlined_call_time
>= 0);
824 gcc_checking_assert (inlined_call_time
>= 0);
825 gcc_checking_assert (uninlined_call_time
>= inlined_call_time
);
827 /* Compute relative time benefit, i.e. how much the call becomes faster.
828 ??? perhaps computing how much the caller+calle together become faster
829 would lead to more realistic results. */
830 if (!uninlined_call_time
)
831 uninlined_call_time
= 1;
833 RDIV (((gcov_type
)uninlined_call_time
- inlined_call_time
) * RELATIVE_TIME_BENEFIT_RANGE
,
834 uninlined_call_time
);
835 relbenefit
= MIN (relbenefit
, RELATIVE_TIME_BENEFIT_RANGE
);
836 gcc_checking_assert (relbenefit
>= 0);
837 relbenefit
= MAX (relbenefit
, 1);
842 /* A cost model driving the inlining heuristics in a way so the edges with
843 smallest badness are inlined first. After each inlining is performed
844 the costs of all caller edges of nodes affected are recomputed so the
845 metrics may accurately depend on values such as number of inlinable callers
846 of the function or function body size. */
849 edge_badness (struct cgraph_edge
*edge
, bool dump
)
852 int growth
, edge_time
;
853 struct cgraph_node
*callee
= cgraph_function_or_thunk_node (edge
->callee
,
855 struct inline_summary
*callee_info
= inline_summary (callee
);
858 if (DECL_DISREGARD_INLINE_LIMITS (callee
->symbol
.decl
))
861 growth
= estimate_edge_growth (edge
);
862 edge_time
= estimate_edge_time (edge
);
863 hints
= estimate_edge_hints (edge
);
864 gcc_checking_assert (edge_time
>= 0);
865 gcc_checking_assert (edge_time
<= callee_info
->time
);
866 gcc_checking_assert (growth
<= callee_info
->size
);
870 fprintf (dump_file
, " Badness calculation for %s/%i -> %s/%i\n",
871 xstrdup (cgraph_node_name (edge
->caller
)),
872 edge
->caller
->symbol
.order
,
873 xstrdup (cgraph_node_name (callee
)),
874 edge
->callee
->symbol
.order
);
875 fprintf (dump_file
, " size growth %i, time %i ",
878 dump_inline_hints (dump_file
, hints
);
879 if (big_speedup_p (edge
))
880 fprintf (dump_file
, " big_speedup");
881 fprintf (dump_file
, "\n");
884 /* Always prefer inlining saving code size. */
887 badness
= INT_MIN
/ 2 + growth
;
889 fprintf (dump_file
, " %i: Growth %i <= 0\n", (int) badness
,
893 /* When profiling is available, compute badness as:
895 relative_edge_count * relative_time_benefit
896 goodness = -------------------------------------------
900 The fraction is upside down, because on edge counts and time beneits
901 the bounds are known. Edge growth is essentially unlimited. */
905 sreal tmp
, relbenefit_real
, growth_real
;
906 int relbenefit
= relative_time_benefit (callee_info
, edge
, edge_time
);
908 sreal_init(&relbenefit_real
, relbenefit
, 0);
909 sreal_init(&growth_real
, growth
, 0);
911 /* relative_edge_count. */
912 sreal_init (&tmp
, edge
->count
, 0);
913 sreal_div (&tmp
, &tmp
, &max_count_real
);
915 /* relative_time_benefit. */
916 sreal_mul (&tmp
, &tmp
, &relbenefit_real
);
917 sreal_div (&tmp
, &tmp
, &max_relbenefit_real
);
919 /* growth_f_caller. */
920 sreal_mul (&tmp
, &tmp
, &half_int_min_real
);
921 sreal_div (&tmp
, &tmp
, &growth_real
);
923 badness
= -1 * sreal_to_int (&tmp
);
925 /* Be sure that insanity of the profile won't lead to increasing counts
926 in the scalling and thus to overflow in the computation above. */
927 gcc_assert (max_count
>= edge
->count
);
931 " %i (relative %f): profile info. Relative count %f"
932 " * Relative benefit %f\n",
933 (int) badness
, (double) badness
/ INT_MIN
,
934 (double) edge
->count
/ max_count
,
935 relbenefit
* 100.0 / RELATIVE_TIME_BENEFIT_RANGE
);
939 /* When function local profile is available. Compute badness as:
941 relative_time_benefit
942 goodness = ---------------------------------
943 growth_of_caller * overall_growth
947 compensated by the inline hints.
949 else if (flag_guess_branch_prob
)
951 badness
= (relative_time_benefit (callee_info
, edge
, edge_time
)
952 * (INT_MIN
/ 16 / RELATIVE_TIME_BENEFIT_RANGE
));
953 badness
/= (MIN (65536/2, growth
) * MIN (65536/2, MAX (1, callee_info
->growth
)));
954 gcc_checking_assert (badness
<=0 && badness
>= INT_MIN
/ 16);
955 if ((hints
& (INLINE_HINT_indirect_call
956 | INLINE_HINT_loop_iterations
957 | INLINE_HINT_array_index
958 | INLINE_HINT_loop_stride
))
959 || callee_info
->growth
<= 0)
961 if (hints
& (INLINE_HINT_same_scc
))
963 else if (hints
& (INLINE_HINT_in_scc
))
965 else if (hints
& (INLINE_HINT_cross_module
))
967 gcc_checking_assert (badness
<= 0 && badness
>= INT_MIN
/ 2);
968 if ((hints
& INLINE_HINT_declared_inline
) && badness
>= INT_MIN
/ 32)
973 " %i: guessed profile. frequency %f,"
974 " benefit %f%%, time w/o inlining %i, time w inlining %i"
975 " overall growth %i (current) %i (original)\n",
976 (int) badness
, (double)edge
->frequency
/ CGRAPH_FREQ_BASE
,
977 relative_time_benefit (callee_info
, edge
, edge_time
) * 100.0
978 / RELATIVE_TIME_BENEFIT_RANGE
,
979 (int)compute_uninlined_call_time (callee_info
, edge
),
980 (int)compute_inlined_call_time (edge
, edge_time
),
981 estimate_growth (callee
),
982 callee_info
->growth
);
985 /* When function local profile is not available or it does not give
986 useful information (ie frequency is zero), base the cost on
987 loop nest and overall size growth, so we optimize for overall number
988 of functions fully inlined in program. */
991 int nest
= MIN (inline_edge_summary (edge
)->loop_depth
, 8);
992 badness
= growth
* 256;
994 /* Decrease badness if call is nested. */
1002 fprintf (dump_file
, " %i: no profile. nest %i\n", (int) badness
,
1006 /* Ensure that we did not overflow in all the fixed point math above. */
1007 gcc_assert (badness
>= INT_MIN
);
1008 gcc_assert (badness
<= INT_MAX
- 1);
1009 /* Make recursive inlining happen always after other inlining is done. */
1010 if (cgraph_edge_recursive_p (edge
))
1016 /* Recompute badness of EDGE and update its key in HEAP if needed. */
1018 update_edge_key (fibheap_t heap
, struct cgraph_edge
*edge
)
1020 int badness
= edge_badness (edge
, false);
1023 fibnode_t n
= (fibnode_t
) edge
->aux
;
1024 gcc_checking_assert (n
->data
== edge
);
1026 /* fibheap_replace_key only decrease the keys.
1027 When we increase the key we do not update heap
1028 and instead re-insert the element once it becomes
1029 a minimum of heap. */
1030 if (badness
< n
->key
)
1032 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1035 " decreasing badness %s/%i -> %s/%i, %i to %i\n",
1036 xstrdup (cgraph_node_name (edge
->caller
)),
1037 edge
->caller
->symbol
.order
,
1038 xstrdup (cgraph_node_name (edge
->callee
)),
1039 edge
->callee
->symbol
.order
,
1043 fibheap_replace_key (heap
, n
, badness
);
1044 gcc_checking_assert (n
->key
== badness
);
1049 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1052 " enqueuing call %s/%i -> %s/%i, badness %i\n",
1053 xstrdup (cgraph_node_name (edge
->caller
)),
1054 edge
->caller
->symbol
.order
,
1055 xstrdup (cgraph_node_name (edge
->callee
)),
1056 edge
->callee
->symbol
.order
,
1059 edge
->aux
= fibheap_insert (heap
, badness
, edge
);
1064 /* NODE was inlined.
1065 All caller edges needs to be resetted because
1066 size estimates change. Similarly callees needs reset
1067 because better context may be known. */
1070 reset_edge_caches (struct cgraph_node
*node
)
1072 struct cgraph_edge
*edge
;
1073 struct cgraph_edge
*e
= node
->callees
;
1074 struct cgraph_node
*where
= node
;
1076 struct ipa_ref
*ref
;
1078 if (where
->global
.inlined_to
)
1079 where
= where
->global
.inlined_to
;
1081 /* WHERE body size has changed, the cached growth is invalid. */
1082 reset_node_growth_cache (where
);
1084 for (edge
= where
->callers
; edge
; edge
= edge
->next_caller
)
1085 if (edge
->inline_failed
)
1086 reset_edge_growth_cache (edge
);
1087 for (i
= 0; ipa_ref_list_referring_iterate (&where
->symbol
.ref_list
,
1089 if (ref
->use
== IPA_REF_ALIAS
)
1090 reset_edge_caches (ipa_ref_referring_node (ref
));
1096 if (!e
->inline_failed
&& e
->callee
->callees
)
1097 e
= e
->callee
->callees
;
1100 if (e
->inline_failed
)
1101 reset_edge_growth_cache (e
);
1108 if (e
->caller
== node
)
1110 e
= e
->caller
->callers
;
1112 while (!e
->next_callee
);
1118 /* Recompute HEAP nodes for each of caller of NODE.
1119 UPDATED_NODES track nodes we already visited, to avoid redundant work.
1120 When CHECK_INLINABLITY_FOR is set, re-check for specified edge that
1121 it is inlinable. Otherwise check all edges. */
1124 update_caller_keys (fibheap_t heap
, struct cgraph_node
*node
,
1125 bitmap updated_nodes
,
1126 struct cgraph_edge
*check_inlinablity_for
)
1128 struct cgraph_edge
*edge
;
1130 struct ipa_ref
*ref
;
1132 if ((!node
->symbol
.alias
&& !inline_summary (node
)->inlinable
)
1133 || node
->global
.inlined_to
)
1135 if (!bitmap_set_bit (updated_nodes
, node
->uid
))
1138 for (i
= 0; ipa_ref_list_referring_iterate (&node
->symbol
.ref_list
,
1140 if (ref
->use
== IPA_REF_ALIAS
)
1142 struct cgraph_node
*alias
= ipa_ref_referring_node (ref
);
1143 update_caller_keys (heap
, alias
, updated_nodes
, check_inlinablity_for
);
1146 for (edge
= node
->callers
; edge
; edge
= edge
->next_caller
)
1147 if (edge
->inline_failed
)
1149 if (!check_inlinablity_for
1150 || check_inlinablity_for
== edge
)
1152 if (can_inline_edge_p (edge
, false)
1153 && want_inline_small_function_p (edge
, false))
1154 update_edge_key (heap
, edge
);
1157 report_inline_failed_reason (edge
);
1158 fibheap_delete_node (heap
, (fibnode_t
) edge
->aux
);
1163 update_edge_key (heap
, edge
);
1167 /* Recompute HEAP nodes for each uninlined call in NODE.
1168 This is used when we know that edge badnesses are going only to increase
1169 (we introduced new call site) and thus all we need is to insert newly
1170 created edges into heap. */
1173 update_callee_keys (fibheap_t heap
, struct cgraph_node
*node
,
1174 bitmap updated_nodes
)
1176 struct cgraph_edge
*e
= node
->callees
;
1181 if (!e
->inline_failed
&& e
->callee
->callees
)
1182 e
= e
->callee
->callees
;
1185 enum availability avail
;
1186 struct cgraph_node
*callee
;
1187 /* We do not reset callee growth cache here. Since we added a new call,
1188 growth chould have just increased and consequentely badness metric
1189 don't need updating. */
1190 if (e
->inline_failed
1191 && (callee
= cgraph_function_or_thunk_node (e
->callee
, &avail
))
1192 && inline_summary (callee
)->inlinable
1193 && avail
>= AVAIL_AVAILABLE
1194 && !bitmap_bit_p (updated_nodes
, callee
->uid
))
1196 if (can_inline_edge_p (e
, false)
1197 && want_inline_small_function_p (e
, false))
1198 update_edge_key (heap
, e
);
1201 report_inline_failed_reason (e
);
1202 fibheap_delete_node (heap
, (fibnode_t
) e
->aux
);
1212 if (e
->caller
== node
)
1214 e
= e
->caller
->callers
;
1216 while (!e
->next_callee
);
1222 /* Enqueue all recursive calls from NODE into priority queue depending on
1223 how likely we want to recursively inline the call. */
1226 lookup_recursive_calls (struct cgraph_node
*node
, struct cgraph_node
*where
,
1229 struct cgraph_edge
*e
;
1230 enum availability avail
;
1232 for (e
= where
->callees
; e
; e
= e
->next_callee
)
1233 if (e
->callee
== node
1234 || (cgraph_function_or_thunk_node (e
->callee
, &avail
) == node
1235 && avail
> AVAIL_OVERWRITABLE
))
1237 /* When profile feedback is available, prioritize by expected number
1239 fibheap_insert (heap
,
1240 !max_count
? -e
->frequency
1241 : -(e
->count
/ ((max_count
+ (1<<24) - 1) / (1<<24))),
1244 for (e
= where
->callees
; e
; e
= e
->next_callee
)
1245 if (!e
->inline_failed
)
1246 lookup_recursive_calls (node
, e
->callee
, heap
);
1249 /* Decide on recursive inlining: in the case function has recursive calls,
1250 inline until body size reaches given argument. If any new indirect edges
1251 are discovered in the process, add them to *NEW_EDGES, unless NEW_EDGES
1255 recursive_inlining (struct cgraph_edge
*edge
,
1256 vec
<cgraph_edge_p
> *new_edges
)
1258 int limit
= PARAM_VALUE (PARAM_MAX_INLINE_INSNS_RECURSIVE_AUTO
);
1260 struct cgraph_node
*node
;
1261 struct cgraph_edge
*e
;
1262 struct cgraph_node
*master_clone
= NULL
, *next
;
1266 node
= edge
->caller
;
1267 if (node
->global
.inlined_to
)
1268 node
= node
->global
.inlined_to
;
1270 if (DECL_DECLARED_INLINE_P (node
->symbol
.decl
))
1271 limit
= PARAM_VALUE (PARAM_MAX_INLINE_INSNS_RECURSIVE
);
1273 /* Make sure that function is small enough to be considered for inlining. */
1274 if (estimate_size_after_inlining (node
, edge
) >= limit
)
1276 heap
= fibheap_new ();
1277 lookup_recursive_calls (node
, node
, heap
);
1278 if (fibheap_empty (heap
))
1280 fibheap_delete (heap
);
1286 " Performing recursive inlining on %s\n",
1287 cgraph_node_name (node
));
1289 /* Do the inlining and update list of recursive call during process. */
1290 while (!fibheap_empty (heap
))
1292 struct cgraph_edge
*curr
1293 = (struct cgraph_edge
*) fibheap_extract_min (heap
);
1294 struct cgraph_node
*cnode
, *dest
= curr
->callee
;
1296 if (!can_inline_edge_p (curr
, true))
1299 /* MASTER_CLONE is produced in the case we already started modified
1300 the function. Be sure to redirect edge to the original body before
1301 estimating growths otherwise we will be seeing growths after inlining
1302 the already modified body. */
1305 cgraph_redirect_edge_callee (curr
, master_clone
);
1306 reset_edge_growth_cache (curr
);
1309 if (estimate_size_after_inlining (node
, curr
) > limit
)
1311 cgraph_redirect_edge_callee (curr
, dest
);
1312 reset_edge_growth_cache (curr
);
1317 for (cnode
= curr
->caller
;
1318 cnode
->global
.inlined_to
; cnode
= cnode
->callers
->caller
)
1319 if (node
->symbol
.decl
1320 == cgraph_function_or_thunk_node (curr
->callee
, NULL
)->symbol
.decl
)
1323 if (!want_inline_self_recursive_call_p (curr
, node
, false, depth
))
1325 cgraph_redirect_edge_callee (curr
, dest
);
1326 reset_edge_growth_cache (curr
);
1333 " Inlining call of depth %i", depth
);
1336 fprintf (dump_file
, " called approx. %.2f times per call",
1337 (double)curr
->count
/ node
->count
);
1339 fprintf (dump_file
, "\n");
1343 /* We need original clone to copy around. */
1344 master_clone
= cgraph_clone_node (node
, node
->symbol
.decl
,
1345 node
->count
, CGRAPH_FREQ_BASE
,
1346 false, vNULL
, true, NULL
);
1347 for (e
= master_clone
->callees
; e
; e
= e
->next_callee
)
1348 if (!e
->inline_failed
)
1349 clone_inlined_nodes (e
, true, false, NULL
);
1350 cgraph_redirect_edge_callee (curr
, master_clone
);
1351 reset_edge_growth_cache (curr
);
1354 inline_call (curr
, false, new_edges
, &overall_size
, true);
1355 lookup_recursive_calls (node
, curr
->callee
, heap
);
1359 if (!fibheap_empty (heap
) && dump_file
)
1360 fprintf (dump_file
, " Recursive inlining growth limit met.\n");
1361 fibheap_delete (heap
);
1368 "\n Inlined %i times, "
1369 "body grown from size %i to %i, time %i to %i\n", n
,
1370 inline_summary (master_clone
)->size
, inline_summary (node
)->size
,
1371 inline_summary (master_clone
)->time
, inline_summary (node
)->time
);
1373 /* Remove master clone we used for inlining. We rely that clones inlined
1374 into master clone gets queued just before master clone so we don't
1376 for (node
= cgraph_first_function (); node
!= master_clone
;
1379 next
= cgraph_next_function (node
);
1380 if (node
->global
.inlined_to
== master_clone
)
1381 cgraph_remove_node (node
);
1383 cgraph_remove_node (master_clone
);
1388 /* Given whole compilation unit estimate of INSNS, compute how large we can
1389 allow the unit to grow. */
1392 compute_max_insns (int insns
)
1394 int max_insns
= insns
;
1395 if (max_insns
< PARAM_VALUE (PARAM_LARGE_UNIT_INSNS
))
1396 max_insns
= PARAM_VALUE (PARAM_LARGE_UNIT_INSNS
);
1398 return ((HOST_WIDEST_INT
) max_insns
1399 * (100 + PARAM_VALUE (PARAM_INLINE_UNIT_GROWTH
)) / 100);
1403 /* Compute badness of all edges in NEW_EDGES and add them to the HEAP. */
1406 add_new_edges_to_heap (fibheap_t heap
, vec
<cgraph_edge_p
> new_edges
)
1408 while (new_edges
.length () > 0)
1410 struct cgraph_edge
*edge
= new_edges
.pop ();
1412 gcc_assert (!edge
->aux
);
1413 if (edge
->inline_failed
1414 && can_inline_edge_p (edge
, true)
1415 && want_inline_small_function_p (edge
, true))
1416 edge
->aux
= fibheap_insert (heap
, edge_badness (edge
, false), edge
);
1420 /* Remove EDGE from the fibheap. */
1423 heap_edge_removal_hook (struct cgraph_edge
*e
, void *data
)
1426 reset_node_growth_cache (e
->callee
);
1429 fibheap_delete_node ((fibheap_t
)data
, (fibnode_t
)e
->aux
);
1434 /* Return true if speculation of edge E seems useful.
1435 If ANTICIPATE_INLINING is true, be conservative and hope that E
1439 speculation_useful_p (struct cgraph_edge
*e
, bool anticipate_inlining
)
1441 enum availability avail
;
1442 struct cgraph_node
*target
= cgraph_function_or_thunk_node (e
->callee
, &avail
);
1443 struct cgraph_edge
*direct
, *indirect
;
1444 struct ipa_ref
*ref
;
1446 gcc_assert (e
->speculative
&& !e
->indirect_unknown_callee
);
1448 if (!cgraph_maybe_hot_edge_p (e
))
1451 /* See if IP optimizations found something potentially useful about the
1452 function. For now we look only for CONST/PURE flags. Almost everything
1453 else we propagate is useless. */
1454 if (avail
>= AVAIL_AVAILABLE
)
1456 int ecf_flags
= flags_from_decl_or_type (target
->symbol
.decl
);
1457 if (ecf_flags
& ECF_CONST
)
1459 cgraph_speculative_call_info (e
, direct
, indirect
, ref
);
1460 if (!(indirect
->indirect_info
->ecf_flags
& ECF_CONST
))
1463 else if (ecf_flags
& ECF_PURE
)
1465 cgraph_speculative_call_info (e
, direct
, indirect
, ref
);
1466 if (!(indirect
->indirect_info
->ecf_flags
& ECF_PURE
))
1470 /* If we did not managed to inline the function nor redirect
1471 to an ipa-cp clone (that are seen by having local flag set),
1472 it is probably pointless to inline it unless hardware is missing
1473 indirect call predictor. */
1474 if (!anticipate_inlining
&& e
->inline_failed
&& !target
->local
.local
)
1476 /* For overwritable targets there is not much to do. */
1477 if (e
->inline_failed
&& !can_inline_edge_p (e
, false, true))
1479 /* OK, speculation seems interesting. */
1483 /* We know that EDGE is not going to be inlined.
1484 See if we can remove speculation. */
1487 resolve_noninline_speculation (fibheap_t edge_heap
, struct cgraph_edge
*edge
)
1489 if (edge
->speculative
&& !speculation_useful_p (edge
, false))
1491 struct cgraph_node
*node
= edge
->caller
;
1492 struct cgraph_node
*where
= node
->global
.inlined_to
1493 ? node
->global
.inlined_to
: node
;
1494 bitmap updated_nodes
= BITMAP_ALLOC (NULL
);
1496 cgraph_resolve_speculation (edge
, NULL
);
1497 reset_edge_caches (where
);
1498 inline_update_overall_summary (where
);
1499 update_caller_keys (edge_heap
, where
,
1500 updated_nodes
, NULL
);
1501 update_callee_keys (edge_heap
, where
,
1503 BITMAP_FREE (updated_nodes
);
1507 /* We use greedy algorithm for inlining of small functions:
1508 All inline candidates are put into prioritized heap ordered in
1511 The inlining of small functions is bounded by unit growth parameters. */
1514 inline_small_functions (void)
1516 struct cgraph_node
*node
;
1517 struct cgraph_edge
*edge
;
1518 fibheap_t edge_heap
= fibheap_new ();
1519 bitmap updated_nodes
= BITMAP_ALLOC (NULL
);
1520 int min_size
, max_size
;
1521 vec
<cgraph_edge_p
> new_indirect_edges
= vNULL
;
1522 int initial_size
= 0;
1523 struct cgraph_node
**order
= XCNEWVEC (struct cgraph_node
*, cgraph_n_nodes
);
1524 struct cgraph_edge_hook_list
*edge_removal_hook_holder
;
1526 if (flag_indirect_inlining
)
1527 new_indirect_edges
.create (8);
1529 edge_removal_hook_holder
1530 = cgraph_add_edge_removal_hook (&heap_edge_removal_hook
, edge_heap
);
1532 /* Compute overall unit size and other global parameters used by badness
1536 ipa_reduced_postorder (order
, true, true, NULL
);
1539 FOR_EACH_DEFINED_FUNCTION (node
)
1540 if (!node
->global
.inlined_to
)
1542 if (cgraph_function_with_gimple_body_p (node
)
1543 || node
->thunk
.thunk_p
)
1545 struct inline_summary
*info
= inline_summary (node
);
1546 struct ipa_dfs_info
*dfs
= (struct ipa_dfs_info
*) node
->symbol
.aux
;
1548 if (!DECL_EXTERNAL (node
->symbol
.decl
))
1549 initial_size
+= info
->size
;
1550 info
->growth
= estimate_growth (node
);
1551 if (dfs
&& dfs
->next_cycle
)
1553 struct cgraph_node
*n2
;
1554 int id
= dfs
->scc_no
+ 1;
1556 n2
= ((struct ipa_dfs_info
*) node
->symbol
.aux
)->next_cycle
)
1558 struct inline_summary
*info2
= inline_summary (n2
);
1566 for (edge
= node
->callers
; edge
; edge
= edge
->next_caller
)
1567 if (max_count
< edge
->count
)
1568 max_count
= edge
->count
;
1570 sreal_init (&max_count_real
, max_count
, 0);
1571 sreal_init (&max_relbenefit_real
, RELATIVE_TIME_BENEFIT_RANGE
, 0);
1572 sreal_init (&half_int_min_real
, INT_MAX
/ 2, 0);
1573 ipa_free_postorder_info ();
1574 initialize_growth_caches ();
1578 "\nDeciding on inlining of small functions. Starting with size %i.\n",
1581 overall_size
= initial_size
;
1582 max_size
= compute_max_insns (overall_size
);
1583 min_size
= overall_size
;
1585 /* Populate the heeap with all edges we might inline. */
1587 FOR_EACH_DEFINED_FUNCTION (node
)
1589 bool update
= false;
1590 struct cgraph_edge
*next
;
1593 fprintf (dump_file
, "Enqueueing calls in %s/%i.\n",
1594 cgraph_node_name (node
), node
->symbol
.order
);
1596 for (edge
= node
->callees
; edge
; edge
= next
)
1598 next
= edge
->next_callee
;
1599 if (edge
->inline_failed
1601 && can_inline_edge_p (edge
, true)
1602 && want_inline_small_function_p (edge
, true)
1603 && edge
->inline_failed
)
1605 gcc_assert (!edge
->aux
);
1606 update_edge_key (edge_heap
, edge
);
1608 if (edge
->speculative
&& !speculation_useful_p (edge
, edge
->aux
!= NULL
))
1610 cgraph_resolve_speculation (edge
, NULL
);
1616 struct cgraph_node
*where
= node
->global
.inlined_to
1617 ? node
->global
.inlined_to
: node
;
1618 inline_update_overall_summary (where
);
1619 reset_node_growth_cache (where
);
1620 reset_edge_caches (where
);
1621 update_caller_keys (edge_heap
, where
,
1622 updated_nodes
, NULL
);
1623 bitmap_clear (updated_nodes
);
1627 gcc_assert (in_lto_p
1629 || (profile_info
&& flag_branch_probabilities
));
1631 while (!fibheap_empty (edge_heap
))
1633 int old_size
= overall_size
;
1634 struct cgraph_node
*where
, *callee
;
1635 int badness
= fibheap_min_key (edge_heap
);
1636 int current_badness
;
1640 edge
= (struct cgraph_edge
*) fibheap_extract_min (edge_heap
);
1641 gcc_assert (edge
->aux
);
1643 if (!edge
->inline_failed
)
1646 /* Be sure that caches are maintained consistent.
1647 We can not make this ENABLE_CHECKING only because it cause different
1648 updates of the fibheap queue. */
1649 cached_badness
= edge_badness (edge
, false);
1650 reset_edge_growth_cache (edge
);
1651 reset_node_growth_cache (edge
->callee
);
1653 /* When updating the edge costs, we only decrease badness in the keys.
1654 Increases of badness are handled lazilly; when we see key with out
1655 of date value on it, we re-insert it now. */
1656 current_badness
= edge_badness (edge
, false);
1657 gcc_assert (cached_badness
== current_badness
);
1658 gcc_assert (current_badness
>= badness
);
1659 if (current_badness
!= badness
)
1661 edge
->aux
= fibheap_insert (edge_heap
, current_badness
, edge
);
1665 if (!can_inline_edge_p (edge
, true))
1667 resolve_noninline_speculation (edge_heap
, edge
);
1671 callee
= cgraph_function_or_thunk_node (edge
->callee
, NULL
);
1672 growth
= estimate_edge_growth (edge
);
1676 "\nConsidering %s/%i with %i size\n",
1677 cgraph_node_name (callee
), callee
->symbol
.order
,
1678 inline_summary (callee
)->size
);
1680 " to be inlined into %s/%i in %s:%i\n"
1681 " Estimated growth after inlined into all is %+i insns.\n"
1682 " Estimated badness is %i, frequency %.2f.\n",
1683 cgraph_node_name (edge
->caller
), edge
->caller
->symbol
.order
,
1684 flag_wpa
? "unknown"
1685 : gimple_filename ((const_gimple
) edge
->call_stmt
),
1687 : gimple_lineno ((const_gimple
) edge
->call_stmt
),
1688 estimate_growth (callee
),
1690 edge
->frequency
/ (double)CGRAPH_FREQ_BASE
);
1692 fprintf (dump_file
," Called "HOST_WIDEST_INT_PRINT_DEC
"x\n",
1694 if (dump_flags
& TDF_DETAILS
)
1695 edge_badness (edge
, true);
1698 if (overall_size
+ growth
> max_size
1699 && !DECL_DISREGARD_INLINE_LIMITS (callee
->symbol
.decl
))
1701 edge
->inline_failed
= CIF_INLINE_UNIT_GROWTH_LIMIT
;
1702 report_inline_failed_reason (edge
);
1703 resolve_noninline_speculation (edge_heap
, edge
);
1707 if (!want_inline_small_function_p (edge
, true))
1709 resolve_noninline_speculation (edge_heap
, edge
);
1713 /* Heuristics for inlining small functions works poorly for
1714 recursive calls where we do efect similar to loop unrolling.
1715 When inliing such edge seems profitable, leave decision on
1716 specific inliner. */
1717 if (cgraph_edge_recursive_p (edge
))
1719 where
= edge
->caller
;
1720 if (where
->global
.inlined_to
)
1721 where
= where
->global
.inlined_to
;
1722 if (!recursive_inlining (edge
,
1723 flag_indirect_inlining
1724 ? &new_indirect_edges
: NULL
))
1726 edge
->inline_failed
= CIF_RECURSIVE_INLINING
;
1727 resolve_noninline_speculation (edge_heap
, edge
);
1730 reset_edge_caches (where
);
1731 /* Recursive inliner inlines all recursive calls of the function
1732 at once. Consequently we need to update all callee keys. */
1733 if (flag_indirect_inlining
)
1734 add_new_edges_to_heap (edge_heap
, new_indirect_edges
);
1735 update_callee_keys (edge_heap
, where
, updated_nodes
);
1736 bitmap_clear (updated_nodes
);
1740 struct cgraph_node
*outer_node
= NULL
;
1743 /* Consider the case where self recursive function A is inlined into B.
1744 This is desired optimization in some cases, since it leads to effect
1745 similar of loop peeling and we might completely optimize out the
1746 recursive call. However we must be extra selective. */
1748 where
= edge
->caller
;
1749 while (where
->global
.inlined_to
)
1751 if (where
->symbol
.decl
== callee
->symbol
.decl
)
1752 outer_node
= where
, depth
++;
1753 where
= where
->callers
->caller
;
1756 && !want_inline_self_recursive_call_p (edge
, outer_node
,
1760 = (DECL_DISREGARD_INLINE_LIMITS (edge
->callee
->symbol
.decl
)
1761 ? CIF_RECURSIVE_INLINING
: CIF_UNSPECIFIED
);
1762 resolve_noninline_speculation (edge_heap
, edge
);
1765 else if (depth
&& dump_file
)
1766 fprintf (dump_file
, " Peeling recursion with depth %i\n", depth
);
1768 gcc_checking_assert (!callee
->global
.inlined_to
);
1769 inline_call (edge
, true, &new_indirect_edges
, &overall_size
, true);
1770 if (flag_indirect_inlining
)
1771 add_new_edges_to_heap (edge_heap
, new_indirect_edges
);
1773 reset_edge_caches (edge
->callee
);
1774 reset_node_growth_cache (callee
);
1776 update_callee_keys (edge_heap
, where
, updated_nodes
);
1778 where
= edge
->caller
;
1779 if (where
->global
.inlined_to
)
1780 where
= where
->global
.inlined_to
;
1782 /* Our profitability metric can depend on local properties
1783 such as number of inlinable calls and size of the function body.
1784 After inlining these properties might change for the function we
1785 inlined into (since it's body size changed) and for the functions
1786 called by function we inlined (since number of it inlinable callers
1788 update_caller_keys (edge_heap
, where
, updated_nodes
, NULL
);
1789 bitmap_clear (updated_nodes
);
1794 " Inlined into %s which now has time %i and size %i,"
1795 "net change of %+i.\n",
1796 cgraph_node_name (edge
->caller
),
1797 inline_summary (edge
->caller
)->time
,
1798 inline_summary (edge
->caller
)->size
,
1799 overall_size
- old_size
);
1801 if (min_size
> overall_size
)
1803 min_size
= overall_size
;
1804 max_size
= compute_max_insns (min_size
);
1807 fprintf (dump_file
, "New minimal size reached: %i\n", min_size
);
1811 free_growth_caches ();
1812 new_indirect_edges
.release ();
1813 fibheap_delete (edge_heap
);
1816 "Unit growth for small function inlining: %i->%i (%i%%)\n",
1817 initial_size
, overall_size
,
1818 initial_size
? overall_size
* 100 / (initial_size
) - 100: 0);
1819 BITMAP_FREE (updated_nodes
);
1820 cgraph_remove_edge_removal_hook (edge_removal_hook_holder
);
1823 /* Flatten NODE. Performed both during early inlining and
1824 at IPA inlining time. */
1827 flatten_function (struct cgraph_node
*node
, bool early
)
1829 struct cgraph_edge
*e
;
1831 /* We shouldn't be called recursively when we are being processed. */
1832 gcc_assert (node
->symbol
.aux
== NULL
);
1834 node
->symbol
.aux
= (void *) node
;
1836 for (e
= node
->callees
; e
; e
= e
->next_callee
)
1838 struct cgraph_node
*orig_callee
;
1839 struct cgraph_node
*callee
= cgraph_function_or_thunk_node (e
->callee
, NULL
);
1841 /* We've hit cycle? It is time to give up. */
1842 if (callee
->symbol
.aux
)
1846 "Not inlining %s into %s to avoid cycle.\n",
1847 xstrdup (cgraph_node_name (callee
)),
1848 xstrdup (cgraph_node_name (e
->caller
)));
1849 e
->inline_failed
= CIF_RECURSIVE_INLINING
;
1853 /* When the edge is already inlined, we just need to recurse into
1854 it in order to fully flatten the leaves. */
1855 if (!e
->inline_failed
)
1857 flatten_function (callee
, early
);
1861 /* Flatten attribute needs to be processed during late inlining. For
1862 extra code quality we however do flattening during early optimization,
1865 ? !can_inline_edge_p (e
, true)
1866 : !can_early_inline_edge_p (e
))
1869 if (cgraph_edge_recursive_p (e
))
1872 fprintf (dump_file
, "Not inlining: recursive call.\n");
1876 if (gimple_in_ssa_p (DECL_STRUCT_FUNCTION (node
->symbol
.decl
))
1877 != gimple_in_ssa_p (DECL_STRUCT_FUNCTION (callee
->symbol
.decl
)))
1880 fprintf (dump_file
, "Not inlining: SSA form does not match.\n");
1884 /* Inline the edge and flatten the inline clone. Avoid
1885 recursing through the original node if the node was cloned. */
1887 fprintf (dump_file
, " Inlining %s into %s.\n",
1888 xstrdup (cgraph_node_name (callee
)),
1889 xstrdup (cgraph_node_name (e
->caller
)));
1890 orig_callee
= callee
;
1891 inline_call (e
, true, NULL
, NULL
, false);
1892 if (e
->callee
!= orig_callee
)
1893 orig_callee
->symbol
.aux
= (void *) node
;
1894 flatten_function (e
->callee
, early
);
1895 if (e
->callee
!= orig_callee
)
1896 orig_callee
->symbol
.aux
= NULL
;
1899 node
->symbol
.aux
= NULL
;
1900 if (!node
->global
.inlined_to
)
1901 inline_update_overall_summary (node
);
1904 /* Count number of callers of NODE and store it into DATA (that
1905 points to int. Worker for cgraph_for_node_and_aliases. */
1908 sum_callers (struct cgraph_node
*node
, void *data
)
1910 struct cgraph_edge
*e
;
1911 int *num_calls
= (int *)data
;
1913 for (e
= node
->callers
; e
; e
= e
->next_caller
)
1918 /* Inline NODE to all callers. Worker for cgraph_for_node_and_aliases.
1919 DATA points to number of calls originally found so we avoid infinite
1923 inline_to_all_callers (struct cgraph_node
*node
, void *data
)
1925 int *num_calls
= (int *)data
;
1926 while (node
->callers
&& !node
->global
.inlined_to
)
1928 struct cgraph_node
*caller
= node
->callers
->caller
;
1933 "\nInlining %s size %i.\n",
1934 cgraph_node_name (node
),
1935 inline_summary (node
)->size
);
1937 " Called once from %s %i insns.\n",
1938 cgraph_node_name (node
->callers
->caller
),
1939 inline_summary (node
->callers
->caller
)->size
);
1942 inline_call (node
->callers
, true, NULL
, NULL
, true);
1945 " Inlined into %s which now has %i size\n",
1946 cgraph_node_name (caller
),
1947 inline_summary (caller
)->size
);
1948 if (!(*num_calls
)--)
1951 fprintf (dump_file
, "New calls found; giving up.\n");
1958 /* Decide on the inlining. We do so in the topological order to avoid
1959 expenses on updating data structures. */
1964 struct cgraph_node
*node
;
1966 struct cgraph_node
**order
=
1967 XCNEWVEC (struct cgraph_node
*, cgraph_n_nodes
);
1970 bool remove_functions
= false;
1975 if (in_lto_p
&& optimize
)
1976 ipa_update_after_lto_read ();
1979 dump_inline_summaries (dump_file
);
1981 nnodes
= ipa_reverse_postorder (order
);
1983 FOR_EACH_FUNCTION (node
)
1984 node
->symbol
.aux
= 0;
1987 fprintf (dump_file
, "\nFlattening functions:\n");
1989 /* In the first pass handle functions to be flattened. Do this with
1990 a priority so none of our later choices will make this impossible. */
1991 for (i
= nnodes
- 1; i
>= 0; i
--)
1995 /* Handle nodes to be flattened.
1996 Ideally when processing callees we stop inlining at the
1997 entry of cycles, possibly cloning that entry point and
1998 try to flatten itself turning it into a self-recursive
2000 if (lookup_attribute ("flatten",
2001 DECL_ATTRIBUTES (node
->symbol
.decl
)) != NULL
)
2005 "Flattening %s\n", cgraph_node_name (node
));
2006 flatten_function (node
, false);
2010 inline_small_functions ();
2012 /* Do first after-inlining removal. We want to remove all "stale" extern inline
2013 functions and virtual functions so we really know what is called once. */
2014 symtab_remove_unreachable_nodes (false, dump_file
);
2017 /* Inline functions with a property that after inlining into all callers the
2018 code size will shrink because the out-of-line copy is eliminated.
2019 We do this regardless on the callee size as long as function growth limits
2023 "\nDeciding on functions to be inlined into all callers and removing useless speculations:\n");
2025 /* Inlining one function called once has good chance of preventing
2026 inlining other function into the same callee. Ideally we should
2027 work in priority order, but probably inlining hot functions first
2028 is good cut without the extra pain of maintaining the queue.
2030 ??? this is not really fitting the bill perfectly: inlining function
2031 into callee often leads to better optimization of callee due to
2032 increased context for optimization.
2033 For example if main() function calls a function that outputs help
2034 and then function that does the main optmization, we should inline
2035 the second with priority even if both calls are cold by themselves.
2037 We probably want to implement new predicate replacing our use of
2038 maybe_hot_edge interpreted as maybe_hot_edge || callee is known
2040 for (cold
= 0; cold
<= 1; cold
++)
2042 FOR_EACH_DEFINED_FUNCTION (node
)
2044 struct cgraph_edge
*edge
, *next
;
2047 for (edge
= node
->callees
; edge
; edge
= next
)
2049 next
= edge
->next_callee
;
2050 if (edge
->speculative
&& !speculation_useful_p (edge
, false))
2052 cgraph_resolve_speculation (edge
, NULL
);
2054 remove_functions
= true;
2059 struct cgraph_node
*where
= node
->global
.inlined_to
2060 ? node
->global
.inlined_to
: node
;
2061 reset_node_growth_cache (where
);
2062 reset_edge_caches (where
);
2063 inline_update_overall_summary (where
);
2065 if (flag_inline_functions_called_once
2066 && want_inline_function_to_all_callers_p (node
, cold
))
2069 cgraph_for_node_and_aliases (node
, sum_callers
,
2071 cgraph_for_node_and_aliases (node
, inline_to_all_callers
,
2073 remove_functions
= true;
2078 /* Free ipa-prop structures if they are no longer needed. */
2080 ipa_free_all_structures_after_iinln ();
2084 "\nInlined %i calls, eliminated %i functions\n\n",
2085 ncalls_inlined
, nfunctions_inlined
);
2088 dump_inline_summaries (dump_file
);
2089 /* In WPA we use inline summaries for partitioning process. */
2091 inline_free_summary ();
2092 return remove_functions
? TODO_remove_functions
: 0;
2095 /* Inline always-inline function calls in NODE. */
2098 inline_always_inline_functions (struct cgraph_node
*node
)
2100 struct cgraph_edge
*e
;
2101 bool inlined
= false;
2103 for (e
= node
->callees
; e
; e
= e
->next_callee
)
2105 struct cgraph_node
*callee
= cgraph_function_or_thunk_node (e
->callee
, NULL
);
2106 if (!DECL_DISREGARD_INLINE_LIMITS (callee
->symbol
.decl
))
2109 if (cgraph_edge_recursive_p (e
))
2112 fprintf (dump_file
, " Not inlining recursive call to %s.\n",
2113 cgraph_node_name (e
->callee
));
2114 e
->inline_failed
= CIF_RECURSIVE_INLINING
;
2118 if (!can_early_inline_edge_p (e
))
2120 /* Set inlined to true if the callee is marked "always_inline" but
2121 is not inlinable. This will allow flagging an error later in
2122 expand_call_inline in tree-inline.c. */
2123 if (lookup_attribute ("always_inline",
2124 DECL_ATTRIBUTES (callee
->symbol
.decl
)) != NULL
)
2130 fprintf (dump_file
, " Inlining %s into %s (always_inline).\n",
2131 xstrdup (cgraph_node_name (e
->callee
)),
2132 xstrdup (cgraph_node_name (e
->caller
)));
2133 inline_call (e
, true, NULL
, NULL
, false);
2137 inline_update_overall_summary (node
);
2142 /* Decide on the inlining. We do so in the topological order to avoid
2143 expenses on updating data structures. */
2146 early_inline_small_functions (struct cgraph_node
*node
)
2148 struct cgraph_edge
*e
;
2149 bool inlined
= false;
2151 for (e
= node
->callees
; e
; e
= e
->next_callee
)
2153 struct cgraph_node
*callee
= cgraph_function_or_thunk_node (e
->callee
, NULL
);
2154 if (!inline_summary (callee
)->inlinable
2155 || !e
->inline_failed
)
2158 /* Do not consider functions not declared inline. */
2159 if (!DECL_DECLARED_INLINE_P (callee
->symbol
.decl
)
2160 && !flag_inline_small_functions
2161 && !flag_inline_functions
)
2165 fprintf (dump_file
, "Considering inline candidate %s.\n",
2166 cgraph_node_name (callee
));
2168 if (!can_early_inline_edge_p (e
))
2171 if (cgraph_edge_recursive_p (e
))
2174 fprintf (dump_file
, " Not inlining: recursive call.\n");
2178 if (!want_early_inline_function_p (e
))
2182 fprintf (dump_file
, " Inlining %s into %s.\n",
2183 xstrdup (cgraph_node_name (callee
)),
2184 xstrdup (cgraph_node_name (e
->caller
)));
2185 inline_call (e
, true, NULL
, NULL
, true);
2192 /* Do inlining of small functions. Doing so early helps profiling and other
2193 passes to be somewhat more effective and avoids some code duplication in
2194 later real inlining pass for testcases with very many function calls. */
2196 early_inliner (void)
2198 struct cgraph_node
*node
= cgraph_get_node (current_function_decl
);
2199 struct cgraph_edge
*edge
;
2200 unsigned int todo
= 0;
2202 bool inlined
= false;
2207 /* Do nothing if datastructures for ipa-inliner are already computed. This
2208 happens when some pass decides to construct new function and
2209 cgraph_add_new_function calls lowering passes and early optimization on
2210 it. This may confuse ourself when early inliner decide to inline call to
2211 function clone, because function clones don't have parameter list in
2212 ipa-prop matching their signature. */
2213 if (ipa_node_params_vector
.exists ())
2216 #ifdef ENABLE_CHECKING
2217 verify_cgraph_node (node
);
2219 ipa_remove_all_references (&node
->symbol
.ref_list
);
2221 /* Even when not optimizing or not inlining inline always-inline
2223 inlined
= inline_always_inline_functions (node
);
2227 || !flag_early_inlining
2228 /* Never inline regular functions into always-inline functions
2229 during incremental inlining. This sucks as functions calling
2230 always inline functions will get less optimized, but at the
2231 same time inlining of functions calling always inline
2232 function into an always inline function might introduce
2233 cycles of edges to be always inlined in the callgraph.
2235 We might want to be smarter and just avoid this type of inlining. */
2236 || DECL_DISREGARD_INLINE_LIMITS (node
->symbol
.decl
))
2238 else if (lookup_attribute ("flatten",
2239 DECL_ATTRIBUTES (node
->symbol
.decl
)) != NULL
)
2241 /* When the function is marked to be flattened, recursively inline
2245 "Flattening %s\n", cgraph_node_name (node
));
2246 flatten_function (node
, true);
2251 /* We iterate incremental inlining to get trivial cases of indirect
2253 while (iterations
< PARAM_VALUE (PARAM_EARLY_INLINER_MAX_ITERATIONS
)
2254 && early_inline_small_functions (node
))
2256 timevar_push (TV_INTEGRATION
);
2257 todo
|= optimize_inline_calls (current_function_decl
);
2259 /* Technically we ought to recompute inline parameters so the new
2260 iteration of early inliner works as expected. We however have
2261 values approximately right and thus we only need to update edge
2262 info that might be cleared out for newly discovered edges. */
2263 for (edge
= node
->callees
; edge
; edge
= edge
->next_callee
)
2265 struct inline_edge_summary
*es
= inline_edge_summary (edge
);
2267 = estimate_num_insns (edge
->call_stmt
, &eni_size_weights
);
2269 = estimate_num_insns (edge
->call_stmt
, &eni_time_weights
);
2270 if (edge
->callee
->symbol
.decl
2271 && !gimple_check_call_matching_types (
2272 edge
->call_stmt
, edge
->callee
->symbol
.decl
, false))
2273 edge
->call_stmt_cannot_inline_p
= true;
2275 timevar_pop (TV_INTEGRATION
);
2280 fprintf (dump_file
, "Iterations: %i\n", iterations
);
2285 timevar_push (TV_INTEGRATION
);
2286 todo
|= optimize_inline_calls (current_function_decl
);
2287 timevar_pop (TV_INTEGRATION
);
2290 cfun
->always_inline_functions_inlined
= true;
2297 const pass_data pass_data_early_inline
=
2299 GIMPLE_PASS
, /* type */
2300 "einline", /* name */
2301 OPTGROUP_INLINE
, /* optinfo_flags */
2302 false, /* has_gate */
2303 true, /* has_execute */
2304 TV_EARLY_INLINING
, /* tv_id */
2305 PROP_ssa
, /* properties_required */
2306 0, /* properties_provided */
2307 0, /* properties_destroyed */
2308 0, /* todo_flags_start */
2309 0, /* todo_flags_finish */
2312 class pass_early_inline
: public gimple_opt_pass
2315 pass_early_inline(gcc::context
*ctxt
)
2316 : gimple_opt_pass(pass_data_early_inline
, ctxt
)
2319 /* opt_pass methods: */
2320 unsigned int execute () { return early_inliner (); }
2322 }; // class pass_early_inline
2327 make_pass_early_inline (gcc::context
*ctxt
)
2329 return new pass_early_inline (ctxt
);
2333 /* When to run IPA inlining. Inlining of always-inline functions
2334 happens during early inlining.
2336 Enable inlining unconditoinally, because callgraph redirection
2340 gate_ipa_inline (void)
2347 const pass_data pass_data_ipa_inline
=
2349 IPA_PASS
, /* type */
2350 "inline", /* name */
2351 OPTGROUP_INLINE
, /* optinfo_flags */
2352 true, /* has_gate */
2353 true, /* has_execute */
2354 TV_IPA_INLINING
, /* tv_id */
2355 0, /* properties_required */
2356 0, /* properties_provided */
2357 0, /* properties_destroyed */
2358 TODO_remove_functions
, /* todo_flags_start */
2359 ( TODO_dump_symtab
), /* todo_flags_finish */
2362 class pass_ipa_inline
: public ipa_opt_pass_d
2365 pass_ipa_inline(gcc::context
*ctxt
)
2366 : ipa_opt_pass_d(pass_data_ipa_inline
, ctxt
,
2367 inline_generate_summary
, /* generate_summary */
2368 inline_write_summary
, /* write_summary */
2369 inline_read_summary
, /* read_summary */
2370 NULL
, /* write_optimization_summary */
2371 NULL
, /* read_optimization_summary */
2372 NULL
, /* stmt_fixup */
2373 0, /* function_transform_todo_flags_start */
2374 inline_transform
, /* function_transform */
2375 NULL
) /* variable_transform */
2378 /* opt_pass methods: */
2379 bool gate () { return gate_ipa_inline (); }
2380 unsigned int execute () { return ipa_inline (); }
2382 }; // class pass_ipa_inline
2387 make_pass_ipa_inline (gcc::context
*ctxt
)
2389 return new pass_ipa_inline (ctxt
);