1 /* Inlining decision heuristics.
2 Copyright (C) 2003-2019 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 cannot 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"
100 #include "alloc-pool.h"
101 #include "tree-pass.h"
102 #include "gimple-ssa.h"
104 #include "lto-streamer.h"
105 #include "trans-mem.h"
107 #include "tree-inline.h"
109 #include "symbol-summary.h"
110 #include "tree-vrp.h"
111 #include "ipa-prop.h"
112 #include "ipa-fnsummary.h"
113 #include "ipa-inline.h"
114 #include "ipa-utils.h"
116 #include "auto-profile.h"
117 #include "builtins.h"
118 #include "fibonacci_heap.h"
119 #include "stringpool.h"
123 typedef fibonacci_heap
<sreal
, cgraph_edge
> edge_heap_t
;
124 typedef fibonacci_node
<sreal
, cgraph_edge
> edge_heap_node_t
;
126 /* Statistics we collect about inlining algorithm. */
127 static int overall_size
;
128 static profile_count max_count
;
129 static profile_count spec_rem
;
131 /* Return false when inlining edge E would lead to violating
132 limits on function unit growth or stack usage growth.
134 The relative function body growth limit is present generally
135 to avoid problems with non-linear behavior of the compiler.
136 To allow inlining huge functions into tiny wrapper, the limit
137 is always based on the bigger of the two functions considered.
139 For stack growth limits we always base the growth in stack usage
140 of the callers. We want to prevent applications from segfaulting
141 on stack overflow when functions with huge stack frames gets
145 caller_growth_limits (struct cgraph_edge
*e
)
147 struct cgraph_node
*to
= e
->caller
;
148 struct cgraph_node
*what
= e
->callee
->ultimate_alias_target ();
151 HOST_WIDE_INT stack_size_limit
= 0, inlined_stack
;
152 ipa_size_summary
*outer_info
= ipa_size_summaries
->get (to
);
154 /* Look for function e->caller is inlined to. While doing
155 so work out the largest function body on the way. As
156 described above, we want to base our function growth
157 limits based on that. Not on the self size of the
158 outer function, not on the self size of inline code
159 we immediately inline to. This is the most relaxed
160 interpretation of the rule "do not grow large functions
161 too much in order to prevent compiler from exploding". */
164 ipa_size_summary
*size_info
= ipa_size_summaries
->get (to
);
165 if (limit
< size_info
->self_size
)
166 limit
= size_info
->self_size
;
167 if (stack_size_limit
< size_info
->estimated_self_stack_size
)
168 stack_size_limit
= size_info
->estimated_self_stack_size
;
170 to
= to
->callers
->caller
;
175 ipa_fn_summary
*what_info
= ipa_fn_summaries
->get (what
);
176 ipa_size_summary
*what_size_info
= ipa_size_summaries
->get (what
);
178 if (limit
< what_size_info
->self_size
)
179 limit
= what_size_info
->self_size
;
181 limit
+= limit
* param_large_function_growth
/ 100;
183 /* Check the size after inlining against the function limits. But allow
184 the function to shrink if it went over the limits by forced inlining. */
185 newsize
= estimate_size_after_inlining (to
, e
);
186 if (newsize
>= ipa_size_summaries
->get (what
)->size
187 && newsize
> param_large_function_insns
190 e
->inline_failed
= CIF_LARGE_FUNCTION_GROWTH_LIMIT
;
194 if (!what_info
->estimated_stack_size
)
197 /* FIXME: Stack size limit often prevents inlining in Fortran programs
198 due to large i/o datastructures used by the Fortran front-end.
199 We ought to ignore this limit when we know that the edge is executed
200 on every invocation of the caller (i.e. its call statement dominates
201 exit block). We do not track this information, yet. */
202 stack_size_limit
+= ((gcov_type
)stack_size_limit
203 * param_stack_frame_growth
/ 100);
205 inlined_stack
= (ipa_get_stack_frame_offset (to
)
206 + outer_info
->estimated_self_stack_size
207 + what_info
->estimated_stack_size
);
208 /* Check new stack consumption with stack consumption at the place
210 if (inlined_stack
> stack_size_limit
211 /* If function already has large stack usage from sibling
212 inline call, we can inline, too.
213 This bit overoptimistically assume that we are good at stack
215 && inlined_stack
> ipa_fn_summaries
->get (to
)->estimated_stack_size
216 && inlined_stack
> param_large_stack_frame
)
218 e
->inline_failed
= CIF_LARGE_STACK_FRAME_GROWTH_LIMIT
;
224 /* Dump info about why inlining has failed. */
227 report_inline_failed_reason (struct cgraph_edge
*e
)
229 if (dump_enabled_p ())
231 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, e
->call_stmt
,
232 " not inlinable: %C -> %C, %s\n",
233 e
->caller
, e
->callee
,
234 cgraph_inline_failed_string (e
->inline_failed
));
235 if ((e
->inline_failed
== CIF_TARGET_OPTION_MISMATCH
236 || e
->inline_failed
== CIF_OPTIMIZATION_MISMATCH
)
237 && e
->caller
->lto_file_data
238 && e
->callee
->ultimate_alias_target ()->lto_file_data
)
240 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, e
->call_stmt
,
241 " LTO objects: %s, %s\n",
242 e
->caller
->lto_file_data
->file_name
,
243 e
->callee
->ultimate_alias_target ()->lto_file_data
->file_name
);
245 if (e
->inline_failed
== CIF_TARGET_OPTION_MISMATCH
)
247 cl_target_option_print_diff
248 (dump_file
, 2, target_opts_for_fn (e
->caller
->decl
),
249 target_opts_for_fn (e
->callee
->ultimate_alias_target ()->decl
));
250 if (e
->inline_failed
== CIF_OPTIMIZATION_MISMATCH
)
252 cl_optimization_print_diff
253 (dump_file
, 2, opts_for_fn (e
->caller
->decl
),
254 opts_for_fn (e
->callee
->ultimate_alias_target ()->decl
));
258 /* Decide whether sanitizer-related attributes allow inlining. */
261 sanitize_attrs_match_for_inline_p (const_tree caller
, const_tree callee
)
263 if (!caller
|| !callee
)
266 /* Allow inlining always_inline functions into no_sanitize_address
268 if (!sanitize_flags_p (SANITIZE_ADDRESS
, caller
)
269 && lookup_attribute ("always_inline", DECL_ATTRIBUTES (callee
)))
272 return ((sanitize_flags_p (SANITIZE_ADDRESS
, caller
)
273 == sanitize_flags_p (SANITIZE_ADDRESS
, callee
))
274 && (sanitize_flags_p (SANITIZE_POINTER_COMPARE
, caller
)
275 == sanitize_flags_p (SANITIZE_POINTER_COMPARE
, callee
))
276 && (sanitize_flags_p (SANITIZE_POINTER_SUBTRACT
, caller
)
277 == sanitize_flags_p (SANITIZE_POINTER_SUBTRACT
, callee
)));
280 /* Used for flags where it is safe to inline when caller's value is
281 grater than callee's. */
282 #define check_maybe_up(flag) \
283 (opts_for_fn (caller->decl)->x_##flag \
284 != opts_for_fn (callee->decl)->x_##flag \
286 || opts_for_fn (caller->decl)->x_##flag \
287 < opts_for_fn (callee->decl)->x_##flag))
288 /* Used for flags where it is safe to inline when caller's value is
289 smaller than callee's. */
290 #define check_maybe_down(flag) \
291 (opts_for_fn (caller->decl)->x_##flag \
292 != opts_for_fn (callee->decl)->x_##flag \
294 || opts_for_fn (caller->decl)->x_##flag \
295 > opts_for_fn (callee->decl)->x_##flag))
296 /* Used for flags where exact match is needed for correctness. */
297 #define check_match(flag) \
298 (opts_for_fn (caller->decl)->x_##flag \
299 != opts_for_fn (callee->decl)->x_##flag)
301 /* Decide if we can inline the edge and possibly update
302 inline_failed reason.
303 We check whether inlining is possible at all and whether
304 caller growth limits allow doing so.
306 if REPORT is true, output reason to the dump file. */
309 can_inline_edge_p (struct cgraph_edge
*e
, bool report
,
312 gcc_checking_assert (e
->inline_failed
);
314 if (cgraph_inline_failed_type (e
->inline_failed
) == CIF_FINAL_ERROR
)
317 report_inline_failed_reason (e
);
321 bool inlinable
= true;
322 enum availability avail
;
323 cgraph_node
*caller
= (e
->caller
->inlined_to
324 ? e
->caller
->inlined_to
: e
->caller
);
325 cgraph_node
*callee
= e
->callee
->ultimate_alias_target (&avail
, caller
);
327 if (!callee
->definition
)
329 e
->inline_failed
= CIF_BODY_NOT_AVAILABLE
;
332 if (!early
&& (!opt_for_fn (callee
->decl
, optimize
)
333 || !opt_for_fn (caller
->decl
, optimize
)))
335 e
->inline_failed
= CIF_FUNCTION_NOT_OPTIMIZED
;
338 else if (callee
->calls_comdat_local
)
340 e
->inline_failed
= CIF_USES_COMDAT_LOCAL
;
343 else if (avail
<= AVAIL_INTERPOSABLE
)
345 e
->inline_failed
= CIF_OVERWRITABLE
;
348 /* All edges with call_stmt_cannot_inline_p should have inline_failed
349 initialized to one of FINAL_ERROR reasons. */
350 else if (e
->call_stmt_cannot_inline_p
)
352 /* Don't inline if the functions have different EH personalities. */
353 else if (DECL_FUNCTION_PERSONALITY (caller
->decl
)
354 && DECL_FUNCTION_PERSONALITY (callee
->decl
)
355 && (DECL_FUNCTION_PERSONALITY (caller
->decl
)
356 != DECL_FUNCTION_PERSONALITY (callee
->decl
)))
358 e
->inline_failed
= CIF_EH_PERSONALITY
;
361 /* TM pure functions should not be inlined into non-TM_pure
363 else if (is_tm_pure (callee
->decl
) && !is_tm_pure (caller
->decl
))
365 e
->inline_failed
= CIF_UNSPECIFIED
;
368 /* Check compatibility of target optimization options. */
369 else if (!targetm
.target_option
.can_inline_p (caller
->decl
,
372 e
->inline_failed
= CIF_TARGET_OPTION_MISMATCH
;
375 else if (ipa_fn_summaries
->get (callee
) == NULL
376 || !ipa_fn_summaries
->get (callee
)->inlinable
)
378 e
->inline_failed
= CIF_FUNCTION_NOT_INLINABLE
;
381 /* Don't inline a function with mismatched sanitization attributes. */
382 else if (!sanitize_attrs_match_for_inline_p (caller
->decl
, callee
->decl
))
384 e
->inline_failed
= CIF_ATTRIBUTE_MISMATCH
;
387 if (!inlinable
&& report
)
388 report_inline_failed_reason (e
);
392 /* Return inlining_insns_single limit for function N. If HINT is true
393 scale up the bound. */
396 inline_insns_single (cgraph_node
*n
, bool hint
)
398 if (opt_for_fn (n
->decl
, optimize
) >= 3)
401 return param_max_inline_insns_single
402 * param_inline_heuristics_hint_percent
/ 100;
403 return param_max_inline_insns_single
;
408 return param_max_inline_insns_single_o2
409 * param_inline_heuristics_hint_percent_o2
/ 100;
410 return param_max_inline_insns_single_o2
;
414 /* Return inlining_insns_auto limit for function N. If HINT is true
415 scale up the bound. */
418 inline_insns_auto (cgraph_node
*n
, bool hint
)
420 if (opt_for_fn (n
->decl
, optimize
) >= 3)
423 return param_max_inline_insns_auto
424 * param_inline_heuristics_hint_percent
/ 100;
425 return param_max_inline_insns_auto
;
430 return param_max_inline_insns_auto_o2
431 * param_inline_heuristics_hint_percent_o2
/ 100;
432 return param_max_inline_insns_auto_o2
;
436 /* Decide if we can inline the edge and possibly update
437 inline_failed reason.
438 We check whether inlining is possible at all and whether
439 caller growth limits allow doing so.
441 if REPORT is true, output reason to the dump file.
443 if DISREGARD_LIMITS is true, ignore size limits. */
446 can_inline_edge_by_limits_p (struct cgraph_edge
*e
, bool report
,
447 bool disregard_limits
= false, bool early
= false)
449 gcc_checking_assert (e
->inline_failed
);
451 if (cgraph_inline_failed_type (e
->inline_failed
) == CIF_FINAL_ERROR
)
454 report_inline_failed_reason (e
);
458 bool inlinable
= true;
459 enum availability avail
;
460 cgraph_node
*caller
= (e
->caller
->inlined_to
461 ? e
->caller
->inlined_to
: e
->caller
);
462 cgraph_node
*callee
= e
->callee
->ultimate_alias_target (&avail
, caller
);
463 tree caller_tree
= DECL_FUNCTION_SPECIFIC_OPTIMIZATION (caller
->decl
);
465 = callee
? DECL_FUNCTION_SPECIFIC_OPTIMIZATION (callee
->decl
) : NULL
;
466 /* Check if caller growth allows the inlining. */
467 if (!DECL_DISREGARD_INLINE_LIMITS (callee
->decl
)
469 && !lookup_attribute ("flatten",
470 DECL_ATTRIBUTES (caller
->decl
))
471 && !caller_growth_limits (e
))
473 else if (callee
->externally_visible
474 && !DECL_DISREGARD_INLINE_LIMITS (callee
->decl
)
475 && flag_live_patching
== LIVE_PATCHING_INLINE_ONLY_STATIC
)
477 e
->inline_failed
= CIF_EXTERN_LIVE_ONLY_STATIC
;
480 /* Don't inline a function with a higher optimization level than the
481 caller. FIXME: this is really just tip of iceberg of handling
482 optimization attribute. */
483 else if (caller_tree
!= callee_tree
)
486 (DECL_DISREGARD_INLINE_LIMITS (callee
->decl
)
487 && lookup_attribute ("always_inline",
488 DECL_ATTRIBUTES (callee
->decl
)));
489 ipa_fn_summary
*caller_info
= ipa_fn_summaries
->get (caller
);
490 ipa_fn_summary
*callee_info
= ipa_fn_summaries
->get (callee
);
492 /* Until GCC 4.9 we did not check the semantics-altering flags
493 below and inlined across optimization boundaries.
494 Enabling checks below breaks several packages by refusing
495 to inline library always_inline functions. See PR65873.
496 Disable the check for early inlining for now until better solution
498 if (always_inline
&& early
)
500 /* There are some options that change IL semantics which means
501 we cannot inline in these cases for correctness reason.
502 Not even for always_inline declared functions. */
503 else if (check_match (flag_wrapv
)
504 || check_match (flag_trapv
)
505 || check_match (flag_pcc_struct_return
)
506 /* When caller or callee does FP math, be sure FP codegen flags
508 || ((caller_info
->fp_expressions
&& callee_info
->fp_expressions
)
509 && (check_maybe_up (flag_rounding_math
)
510 || check_maybe_up (flag_trapping_math
)
511 || check_maybe_down (flag_unsafe_math_optimizations
)
512 || check_maybe_down (flag_finite_math_only
)
513 || check_maybe_up (flag_signaling_nans
)
514 || check_maybe_down (flag_cx_limited_range
)
515 || check_maybe_up (flag_signed_zeros
)
516 || check_maybe_down (flag_associative_math
)
517 || check_maybe_down (flag_reciprocal_math
)
518 || check_maybe_down (flag_fp_int_builtin_inexact
)
519 /* Strictly speaking only when the callee contains function
520 calls that may end up setting errno. */
521 || check_maybe_up (flag_errno_math
)))
522 /* We do not want to make code compiled with exceptions to be
523 brought into a non-EH function unless we know that the callee
525 This is tracked by DECL_FUNCTION_PERSONALITY. */
526 || (check_maybe_up (flag_non_call_exceptions
)
527 && DECL_FUNCTION_PERSONALITY (callee
->decl
))
528 || (check_maybe_up (flag_exceptions
)
529 && DECL_FUNCTION_PERSONALITY (callee
->decl
))
530 /* When devirtualization is diabled for callee, it is not safe
531 to inline it as we possibly mangled the type info.
532 Allow early inlining of always inlines. */
533 || (!early
&& check_maybe_down (flag_devirtualize
)))
535 e
->inline_failed
= CIF_OPTIMIZATION_MISMATCH
;
538 /* gcc.dg/pr43564.c. Apply user-forced inline even at -O0. */
539 else if (always_inline
)
541 /* When user added an attribute to the callee honor it. */
542 else if (lookup_attribute ("optimize", DECL_ATTRIBUTES (callee
->decl
))
543 && opts_for_fn (caller
->decl
) != opts_for_fn (callee
->decl
))
545 e
->inline_failed
= CIF_OPTIMIZATION_MISMATCH
;
548 /* If explicit optimize attribute are not used, the mismatch is caused
549 by different command line options used to build different units.
550 Do not care about COMDAT functions - those are intended to be
551 optimized with the optimization flags of module they are used in.
552 Also do not care about mixing up size/speed optimization when
553 DECL_DISREGARD_INLINE_LIMITS is set. */
554 else if ((callee
->merged_comdat
555 && !lookup_attribute ("optimize",
556 DECL_ATTRIBUTES (caller
->decl
)))
557 || DECL_DISREGARD_INLINE_LIMITS (callee
->decl
))
559 /* If mismatch is caused by merging two LTO units with different
560 optimizationflags we want to be bit nicer. However never inline
561 if one of functions is not optimized at all. */
562 else if (!opt_for_fn (callee
->decl
, optimize
)
563 || !opt_for_fn (caller
->decl
, optimize
))
565 e
->inline_failed
= CIF_OPTIMIZATION_MISMATCH
;
568 /* If callee is optimized for size and caller is not, allow inlining if
569 code shrinks or we are in param_max_inline_insns_single limit and
570 callee is inline (and thus likely an unified comdat).
571 This will allow caller to run faster. */
572 else if (opt_for_fn (callee
->decl
, optimize_size
)
573 > opt_for_fn (caller
->decl
, optimize_size
))
575 int growth
= estimate_edge_growth (e
);
576 if (growth
> param_max_inline_insns_size
577 && (!DECL_DECLARED_INLINE_P (callee
->decl
)
578 && growth
>= MAX (inline_insns_single (caller
, false),
579 inline_insns_auto (caller
, false))))
581 e
->inline_failed
= CIF_OPTIMIZATION_MISMATCH
;
585 /* If callee is more aggressively optimized for performance than caller,
586 we generally want to inline only cheap (runtime wise) functions. */
587 else if (opt_for_fn (callee
->decl
, optimize_size
)
588 < opt_for_fn (caller
->decl
, optimize_size
)
589 || (opt_for_fn (callee
->decl
, optimize
)
590 > opt_for_fn (caller
->decl
, optimize
)))
592 if (estimate_edge_time (e
)
593 >= 20 + ipa_call_summaries
->get (e
)->call_stmt_time
)
595 e
->inline_failed
= CIF_OPTIMIZATION_MISMATCH
;
602 if (!inlinable
&& report
)
603 report_inline_failed_reason (e
);
608 /* Return true if the edge E is inlinable during early inlining. */
611 can_early_inline_edge_p (struct cgraph_edge
*e
)
613 struct cgraph_node
*callee
= e
->callee
->ultimate_alias_target ();
614 /* Early inliner might get called at WPA stage when IPA pass adds new
615 function. In this case we cannot really do any of early inlining
616 because function bodies are missing. */
617 if (cgraph_inline_failed_type (e
->inline_failed
) == CIF_FINAL_ERROR
)
619 if (!gimple_has_body_p (callee
->decl
))
621 e
->inline_failed
= CIF_BODY_NOT_AVAILABLE
;
624 /* In early inliner some of callees may not be in SSA form yet
625 (i.e. the callgraph is cyclic and we did not process
626 the callee by early inliner, yet). We don't have CIF code for this
627 case; later we will re-do the decision in the real inliner. */
628 if (!gimple_in_ssa_p (DECL_STRUCT_FUNCTION (e
->caller
->decl
))
629 || !gimple_in_ssa_p (DECL_STRUCT_FUNCTION (callee
->decl
)))
631 if (dump_enabled_p ())
632 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, e
->call_stmt
,
633 " edge not inlinable: not in SSA form\n");
636 if (!can_inline_edge_p (e
, true, true)
637 || !can_inline_edge_by_limits_p (e
, true, false, true))
643 /* Return number of calls in N. Ignore cheap builtins. */
646 num_calls (struct cgraph_node
*n
)
648 struct cgraph_edge
*e
;
651 for (e
= n
->callees
; e
; e
= e
->next_callee
)
652 if (!is_inexpensive_builtin (e
->callee
->decl
))
658 /* Return true if we are interested in inlining small function. */
661 want_early_inline_function_p (struct cgraph_edge
*e
)
663 bool want_inline
= true;
664 struct cgraph_node
*callee
= e
->callee
->ultimate_alias_target ();
666 if (DECL_DISREGARD_INLINE_LIMITS (callee
->decl
))
668 /* For AutoFDO, we need to make sure that before profile summary, all
669 hot paths' IR look exactly the same as profiled binary. As a result,
670 in einliner, we will disregard size limit and inline those callsites
672 * inlined in the profiled binary, and
673 * the cloned callee has enough samples to be considered "hot". */
674 else if (flag_auto_profile
&& afdo_callsite_hot_enough_for_early_inline (e
))
676 else if (!DECL_DECLARED_INLINE_P (callee
->decl
)
677 && !opt_for_fn (e
->caller
->decl
, flag_inline_small_functions
))
679 e
->inline_failed
= CIF_FUNCTION_NOT_INLINE_CANDIDATE
;
680 report_inline_failed_reason (e
);
685 int growth
= estimate_edge_growth (e
);
687 int early_inlining_insns
= opt_for_fn (e
->caller
->decl
, optimize
) >= 3
688 ? param_early_inlining_insns
689 : param_early_inlining_insns_o2
;
692 if (growth
<= param_max_inline_insns_size
)
694 else if (!e
->maybe_hot_p ())
696 if (dump_enabled_p ())
697 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, e
->call_stmt
,
698 " will not early inline: %C->%C, "
699 "call is cold and code would grow by %i\n",
704 else if (growth
> early_inlining_insns
)
706 if (dump_enabled_p ())
707 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, e
->call_stmt
,
708 " will not early inline: %C->%C, "
709 "growth %i exceeds --param early-inlining-insns%s\n",
710 e
->caller
, callee
, growth
,
711 opt_for_fn (e
->caller
->decl
, optimize
) >= 3
715 else if ((n
= num_calls (callee
)) != 0
716 && growth
* (n
+ 1) > early_inlining_insns
)
718 if (dump_enabled_p ())
719 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, e
->call_stmt
,
720 " will not early inline: %C->%C, "
721 "growth %i exceeds --param early-inlining-insns%s "
722 "divided by number of calls\n",
723 e
->caller
, callee
, growth
,
724 opt_for_fn (e
->caller
->decl
, optimize
) >= 3
732 /* Compute time of the edge->caller + edge->callee execution when inlining
736 compute_uninlined_call_time (struct cgraph_edge
*edge
,
737 sreal uninlined_call_time
,
740 cgraph_node
*caller
= (edge
->caller
->inlined_to
741 ? edge
->caller
->inlined_to
745 uninlined_call_time
*= freq
;
747 uninlined_call_time
= uninlined_call_time
>> 11;
749 sreal caller_time
= ipa_fn_summaries
->get (caller
)->time
;
750 return uninlined_call_time
+ caller_time
;
753 /* Same as compute_uinlined_call_time but compute time when inlining
757 compute_inlined_call_time (struct cgraph_edge
*edge
,
761 cgraph_node
*caller
= (edge
->caller
->inlined_to
762 ? edge
->caller
->inlined_to
764 sreal caller_time
= ipa_fn_summaries
->get (caller
)->time
;
771 /* This calculation should match one in ipa-inline-analysis.c
772 (estimate_edge_size_and_time). */
773 time
-= (sreal
)ipa_call_summaries
->get (edge
)->call_stmt_time
* freq
;
776 time
= ((sreal
) 1) >> 8;
777 gcc_checking_assert (time
>= 0);
781 /* Return true if the speedup for inlining E is bigger than
782 PARAM_MAX_INLINE_MIN_SPEEDUP. */
785 big_speedup_p (struct cgraph_edge
*e
)
788 sreal spec_time
= estimate_edge_time (e
, &unspec_time
);
789 sreal freq
= e
->sreal_frequency ();
790 sreal time
= compute_uninlined_call_time (e
, unspec_time
, freq
);
791 sreal inlined_time
= compute_inlined_call_time (e
, spec_time
, freq
);
792 cgraph_node
*caller
= (e
->caller
->inlined_to
793 ? e
->caller
->inlined_to
795 int limit
= opt_for_fn (caller
->decl
, optimize
) >= 3
796 ? param_inline_min_speedup
797 : param_inline_min_speedup_o2
;
799 if ((time
- inlined_time
) * 100 > time
* limit
)
804 /* Return true if we are interested in inlining small function.
805 When REPORT is true, report reason to dump file. */
808 want_inline_small_function_p (struct cgraph_edge
*e
, bool report
)
810 bool want_inline
= true;
811 struct cgraph_node
*callee
= e
->callee
->ultimate_alias_target ();
813 /* Allow this function to be called before can_inline_edge_p,
814 since it's usually cheaper. */
815 if (cgraph_inline_failed_type (e
->inline_failed
) == CIF_FINAL_ERROR
)
817 else if (DECL_DISREGARD_INLINE_LIMITS (callee
->decl
))
819 else if (!DECL_DECLARED_INLINE_P (callee
->decl
)
820 && !opt_for_fn (e
->caller
->decl
, flag_inline_small_functions
))
822 e
->inline_failed
= CIF_FUNCTION_NOT_INLINE_CANDIDATE
;
825 /* Do fast and conservative check if the function can be good
827 else if ((!DECL_DECLARED_INLINE_P (callee
->decl
)
828 && (!e
->count
.ipa ().initialized_p () || !e
->maybe_hot_p ()))
829 && ipa_fn_summaries
->get (callee
)->min_size
830 - ipa_call_summaries
->get (e
)->call_stmt_size
831 > inline_insns_auto (e
->caller
, true))
833 if (opt_for_fn (e
->caller
->decl
, optimize
) >= 3)
834 e
->inline_failed
= CIF_MAX_INLINE_INSNS_AUTO_LIMIT
;
836 e
->inline_failed
= CIF_MAX_INLINE_INSNS_AUTO_O2_LIMIT
;
839 else if ((DECL_DECLARED_INLINE_P (callee
->decl
)
840 || e
->count
.ipa ().nonzero_p ())
841 && ipa_fn_summaries
->get (callee
)->min_size
842 - ipa_call_summaries
->get (e
)->call_stmt_size
843 > inline_insns_single (e
->caller
, true))
845 if (opt_for_fn (e
->caller
->decl
, optimize
) >= 3)
846 e
->inline_failed
= (DECL_DECLARED_INLINE_P (callee
->decl
)
847 ? CIF_MAX_INLINE_INSNS_SINGLE_LIMIT
848 : CIF_MAX_INLINE_INSNS_AUTO_LIMIT
);
850 e
->inline_failed
= (DECL_DECLARED_INLINE_P (callee
->decl
)
851 ? CIF_MAX_INLINE_INSNS_SINGLE_O2_LIMIT
852 : CIF_MAX_INLINE_INSNS_AUTO_O2_LIMIT
);
857 int growth
= estimate_edge_growth (e
);
858 ipa_hints hints
= estimate_edge_hints (e
);
859 bool apply_hints
= (hints
& (INLINE_HINT_indirect_call
860 | INLINE_HINT_known_hot
861 | INLINE_HINT_loop_iterations
862 | INLINE_HINT_loop_stride
));
864 if (growth
<= param_max_inline_insns_size
)
866 /* Apply param_max_inline_insns_single limit. Do not do so when
867 hints suggests that inlining given function is very profitable.
868 Avoid computation of big_speedup_p when not necessary to change
869 outcome of decision. */
870 else if (DECL_DECLARED_INLINE_P (callee
->decl
)
871 && growth
>= inline_insns_single (e
->caller
, apply_hints
)
873 || growth
>= inline_insns_single (e
->caller
, true)
874 || !big_speedup_p (e
)))
876 if (opt_for_fn (e
->caller
->decl
, optimize
) >= 3)
877 e
->inline_failed
= CIF_MAX_INLINE_INSNS_SINGLE_LIMIT
;
879 e
->inline_failed
= CIF_MAX_INLINE_INSNS_SINGLE_O2_LIMIT
;
882 else if (!DECL_DECLARED_INLINE_P (callee
->decl
)
883 && !opt_for_fn (e
->caller
->decl
, flag_inline_functions
)
884 && growth
>= param_max_inline_insns_small
)
886 /* growth_positive_p is expensive, always test it last. */
887 if (growth
>= inline_insns_single (e
->caller
, false)
888 || growth_positive_p (callee
, e
, growth
))
890 e
->inline_failed
= CIF_NOT_DECLARED_INLINED
;
894 /* Apply param_max_inline_insns_auto limit for functions not declared
895 inline. Bypass the limit when speedup seems big. */
896 else if (!DECL_DECLARED_INLINE_P (callee
->decl
)
897 && growth
>= inline_insns_auto (e
->caller
, apply_hints
)
899 || growth
>= inline_insns_auto (e
->caller
, true)
900 || !big_speedup_p (e
)))
902 /* growth_positive_p is expensive, always test it last. */
903 if (growth
>= inline_insns_single (e
->caller
, false)
904 || growth_positive_p (callee
, e
, growth
))
906 if (opt_for_fn (e
->caller
->decl
, optimize
) >= 3)
907 e
->inline_failed
= CIF_MAX_INLINE_INSNS_AUTO_LIMIT
;
909 e
->inline_failed
= CIF_MAX_INLINE_INSNS_AUTO_O2_LIMIT
;
913 /* If call is cold, do not inline when function body would grow. */
914 else if (!e
->maybe_hot_p ()
915 && (growth
>= inline_insns_single (e
->caller
, false)
916 || growth_positive_p (callee
, e
, growth
)))
918 e
->inline_failed
= CIF_UNLIKELY_CALL
;
922 if (!want_inline
&& report
)
923 report_inline_failed_reason (e
);
927 /* EDGE is self recursive edge.
928 We hand two cases - when function A is inlining into itself
929 or when function A is being inlined into another inliner copy of function
932 In first case OUTER_NODE points to the toplevel copy of A, while
933 in the second case OUTER_NODE points to the outermost copy of A in B.
935 In both cases we want to be extra selective since
936 inlining the call will just introduce new recursive calls to appear. */
939 want_inline_self_recursive_call_p (struct cgraph_edge
*edge
,
940 struct cgraph_node
*outer_node
,
944 char const *reason
= NULL
;
945 bool want_inline
= true;
946 sreal caller_freq
= 1;
947 int max_depth
= param_max_inline_recursive_depth_auto
;
949 if (DECL_DECLARED_INLINE_P (edge
->caller
->decl
))
950 max_depth
= param_max_inline_recursive_depth
;
952 if (!edge
->maybe_hot_p ())
954 reason
= "recursive call is cold";
957 else if (depth
> max_depth
)
959 reason
= "--param max-inline-recursive-depth exceeded.";
962 else if (outer_node
->inlined_to
963 && (caller_freq
= outer_node
->callers
->sreal_frequency ()) == 0)
965 reason
= "caller frequency is 0";
971 /* Inlining of self recursive function into copy of itself within other
972 function is transformation similar to loop peeling.
974 Peeling is profitable if we can inline enough copies to make probability
975 of actual call to the self recursive function very small. Be sure that
976 the probability of recursion is small.
978 We ensure that the frequency of recursing is at most 1 - (1/max_depth).
979 This way the expected number of recursion is at most max_depth. */
982 sreal max_prob
= (sreal
)1 - ((sreal
)1 / (sreal
)max_depth
);
984 for (i
= 1; i
< depth
; i
++)
985 max_prob
= max_prob
* max_prob
;
986 if (edge
->sreal_frequency () >= max_prob
* caller_freq
)
988 reason
= "frequency of recursive call is too large";
992 /* Recursive inlining, i.e. equivalent of unrolling, is profitable if
993 recursion depth is large. We reduce function call overhead and increase
994 chances that things fit in hardware return predictor.
996 Recursive inlining might however increase cost of stack frame setup
997 actually slowing down functions whose recursion tree is wide rather than
1000 Deciding reliably on when to do recursive inlining without profile feedback
1001 is tricky. For now we disable recursive inlining when probability of self
1004 Recursive inlining of self recursive call within loop also results in
1005 large loop depths that generally optimize badly. We may want to throttle
1006 down inlining in those cases. In particular this seems to happen in one
1007 of libstdc++ rb tree methods. */
1010 if (edge
->sreal_frequency () * 100
1012 * param_min_inline_recursive_probability
)
1014 reason
= "frequency of recursive call is too small";
1015 want_inline
= false;
1018 if (!want_inline
&& dump_enabled_p ())
1019 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, edge
->call_stmt
,
1020 " not inlining recursively: %s\n", reason
);
1024 /* Return true when NODE has uninlinable caller;
1025 set HAS_HOT_CALL if it has hot call.
1026 Worker for cgraph_for_node_and_aliases. */
1029 check_callers (struct cgraph_node
*node
, void *has_hot_call
)
1031 struct cgraph_edge
*e
;
1032 for (e
= node
->callers
; e
; e
= e
->next_caller
)
1034 if (!opt_for_fn (e
->caller
->decl
, flag_inline_functions_called_once
)
1035 || !opt_for_fn (e
->caller
->decl
, optimize
))
1037 if (!can_inline_edge_p (e
, true))
1039 if (e
->recursive_p ())
1041 if (!can_inline_edge_by_limits_p (e
, true))
1043 if (!(*(bool *)has_hot_call
) && e
->maybe_hot_p ())
1044 *(bool *)has_hot_call
= true;
1049 /* If NODE has a caller, return true. */
1052 has_caller_p (struct cgraph_node
*node
, void *data ATTRIBUTE_UNUSED
)
1059 /* Decide if inlining NODE would reduce unit size by eliminating
1060 the offline copy of function.
1061 When COLD is true the cold calls are considered, too. */
1064 want_inline_function_to_all_callers_p (struct cgraph_node
*node
, bool cold
)
1066 bool has_hot_call
= false;
1068 /* Aliases gets inlined along with the function they alias. */
1071 /* Already inlined? */
1072 if (node
->inlined_to
)
1074 /* Does it have callers? */
1075 if (!node
->call_for_symbol_and_aliases (has_caller_p
, NULL
, true))
1077 /* Inlining into all callers would increase size? */
1078 if (growth_positive_p (node
, NULL
, INT_MIN
) > 0)
1080 /* All inlines must be possible. */
1081 if (node
->call_for_symbol_and_aliases (check_callers
, &has_hot_call
,
1084 if (!cold
&& !has_hot_call
)
1089 /* A cost model driving the inlining heuristics in a way so the edges with
1090 smallest badness are inlined first. After each inlining is performed
1091 the costs of all caller edges of nodes affected are recomputed so the
1092 metrics may accurately depend on values such as number of inlinable callers
1093 of the function or function body size. */
1096 edge_badness (struct cgraph_edge
*edge
, bool dump
)
1100 sreal edge_time
, unspec_edge_time
;
1101 struct cgraph_node
*callee
= edge
->callee
->ultimate_alias_target ();
1102 class ipa_fn_summary
*callee_info
= ipa_fn_summaries
->get (callee
);
1104 cgraph_node
*caller
= (edge
->caller
->inlined_to
1105 ? edge
->caller
->inlined_to
1108 growth
= estimate_edge_growth (edge
);
1109 edge_time
= estimate_edge_time (edge
, &unspec_edge_time
);
1110 hints
= estimate_edge_hints (edge
);
1111 gcc_checking_assert (edge_time
>= 0);
1112 /* Check that inlined time is better, but tolerate some roundoff issues.
1113 FIXME: When callee profile drops to 0 we account calls more. This
1114 should be fixed by never doing that. */
1115 gcc_checking_assert ((edge_time
* 100
1116 - callee_info
->time
* 101).to_int () <= 0
1117 || callee
->count
.ipa ().initialized_p ());
1118 gcc_checking_assert (growth
<= ipa_size_summaries
->get (callee
)->size
);
1122 fprintf (dump_file
, " Badness calculation for %s -> %s\n",
1123 edge
->caller
->dump_name (),
1124 edge
->callee
->dump_name ());
1125 fprintf (dump_file
, " size growth %i, time %f unspec %f ",
1127 edge_time
.to_double (),
1128 unspec_edge_time
.to_double ());
1129 ipa_dump_hints (dump_file
, hints
);
1130 if (big_speedup_p (edge
))
1131 fprintf (dump_file
, " big_speedup");
1132 fprintf (dump_file
, "\n");
1135 /* Always prefer inlining saving code size. */
1138 badness
= (sreal
) (-SREAL_MIN_SIG
+ growth
) << (SREAL_MAX_EXP
/ 256);
1140 fprintf (dump_file
, " %f: Growth %d <= 0\n", badness
.to_double (),
1143 /* Inlining into EXTERNAL functions is not going to change anything unless
1144 they are themselves inlined. */
1145 else if (DECL_EXTERNAL (caller
->decl
))
1148 fprintf (dump_file
, " max: function is external\n");
1149 return sreal::max ();
1151 /* When profile is available. Compute badness as:
1153 time_saved * caller_count
1154 goodness = -------------------------------------------------
1155 growth_of_caller * overall_growth * combined_size
1157 badness = - goodness
1159 Again use negative value to make calls with profile appear hotter
1162 else if (opt_for_fn (caller
->decl
, flag_guess_branch_prob
)
1163 || caller
->count
.ipa ().nonzero_p ())
1165 sreal numerator
, denominator
;
1167 sreal freq
= edge
->sreal_frequency ();
1168 sreal inlined_time
= compute_inlined_call_time (edge
, edge_time
, freq
);
1170 numerator
= (compute_uninlined_call_time (edge
, unspec_edge_time
, freq
)
1173 numerator
= ((sreal
) 1 >> 8);
1174 if (caller
->count
.ipa ().nonzero_p ())
1175 numerator
*= caller
->count
.ipa ().to_gcov_type ();
1176 else if (caller
->count
.ipa ().initialized_p ())
1177 numerator
= numerator
>> 11;
1178 denominator
= growth
;
1180 overall_growth
= callee_info
->growth
;
1182 /* Look for inliner wrappers of the form:
1188 noninline_callee ();
1190 Withhout penalizing this case, we usually inline noninline_callee
1191 into the inline_caller because overall_growth is small preventing
1192 further inlining of inline_caller.
1194 Penalize only callgraph edges to functions with small overall
1197 if (growth
> overall_growth
1198 /* ... and having only one caller which is not inlined ... */
1199 && callee_info
->single_caller
1200 && !edge
->caller
->inlined_to
1201 /* ... and edges executed only conditionally ... */
1203 /* ... consider case where callee is not inline but caller is ... */
1204 && ((!DECL_DECLARED_INLINE_P (edge
->callee
->decl
)
1205 && DECL_DECLARED_INLINE_P (caller
->decl
))
1206 /* ... or when early optimizers decided to split and edge
1207 frequency still indicates splitting is a win ... */
1208 || (callee
->split_part
&& !caller
->split_part
1209 && freq
* 100 < param_partial_inlining_entry_probability
1210 /* ... and do not overwrite user specified hints. */
1211 && (!DECL_DECLARED_INLINE_P (edge
->callee
->decl
)
1212 || DECL_DECLARED_INLINE_P (caller
->decl
)))))
1214 ipa_fn_summary
*caller_info
= ipa_fn_summaries
->get (caller
);
1215 int caller_growth
= caller_info
->growth
;
1217 /* Only apply the penalty when caller looks like inline candidate,
1218 and it is not called once. */
1219 if (!caller_info
->single_caller
&& overall_growth
< caller_growth
1220 && caller_info
->inlinable
1221 && ipa_size_summaries
->get (caller
)->size
1222 < (DECL_DECLARED_INLINE_P (caller
->decl
)
1223 ? inline_insns_single (caller
, false)
1224 : inline_insns_auto (caller
, false)))
1228 " Wrapper penalty. Increasing growth %i to %i\n",
1229 overall_growth
, caller_growth
);
1230 overall_growth
= caller_growth
;
1233 if (overall_growth
> 0)
1235 /* Strongly preffer functions with few callers that can be inlined
1236 fully. The square root here leads to smaller binaries at average.
1237 Watch however for extreme cases and return to linear function
1238 when growth is large. */
1239 if (overall_growth
< 256)
1240 overall_growth
*= overall_growth
;
1242 overall_growth
+= 256 * 256 - 256;
1243 denominator
*= overall_growth
;
1245 denominator
*= ipa_size_summaries
->get (caller
)->size
+ growth
;
1247 badness
= - numerator
/ denominator
;
1252 " %f: guessed profile. frequency %f, count %" PRId64
1253 " caller count %" PRId64
1254 " time w/o inlining %f, time with inlining %f"
1255 " overall growth %i (current) %i (original)"
1256 " %i (compensated)\n",
1257 badness
.to_double (),
1259 edge
->count
.ipa ().initialized_p () ? edge
->count
.ipa ().to_gcov_type () : -1,
1260 caller
->count
.ipa ().initialized_p () ? caller
->count
.ipa ().to_gcov_type () : -1,
1261 compute_uninlined_call_time (edge
,
1262 unspec_edge_time
, freq
).to_double (),
1263 inlined_time
.to_double (),
1264 estimate_growth (callee
),
1265 callee_info
->growth
, overall_growth
);
1268 /* When function local profile is not available or it does not give
1269 useful information (ie frequency is zero), base the cost on
1270 loop nest and overall size growth, so we optimize for overall number
1271 of functions fully inlined in program. */
1274 int nest
= MIN (ipa_call_summaries
->get (edge
)->loop_depth
, 8);
1277 /* Decrease badness if call is nested. */
1279 badness
= badness
>> nest
;
1281 badness
= badness
<< nest
;
1283 fprintf (dump_file
, " %f: no profile. nest %i\n",
1284 badness
.to_double (), nest
);
1286 gcc_checking_assert (badness
!= 0);
1288 if (edge
->recursive_p ())
1289 badness
= badness
.shift (badness
> 0 ? 4 : -4);
1290 if ((hints
& (INLINE_HINT_indirect_call
1291 | INLINE_HINT_loop_iterations
1292 | INLINE_HINT_loop_stride
))
1293 || callee_info
->growth
<= 0)
1294 badness
= badness
.shift (badness
> 0 ? -2 : 2);
1295 if (hints
& (INLINE_HINT_same_scc
))
1296 badness
= badness
.shift (badness
> 0 ? 3 : -3);
1297 else if (hints
& (INLINE_HINT_in_scc
))
1298 badness
= badness
.shift (badness
> 0 ? 2 : -2);
1299 else if (hints
& (INLINE_HINT_cross_module
))
1300 badness
= badness
.shift (badness
> 0 ? 1 : -1);
1301 if (DECL_DISREGARD_INLINE_LIMITS (callee
->decl
))
1302 badness
= badness
.shift (badness
> 0 ? -4 : 4);
1303 else if ((hints
& INLINE_HINT_declared_inline
))
1304 badness
= badness
.shift (badness
> 0 ? -3 : 3);
1306 fprintf (dump_file
, " Adjusted by hints %f\n", badness
.to_double ());
1310 /* Recompute badness of EDGE and update its key in HEAP if needed. */
1312 update_edge_key (edge_heap_t
*heap
, struct cgraph_edge
*edge
)
1314 sreal badness
= edge_badness (edge
, false);
1317 edge_heap_node_t
*n
= (edge_heap_node_t
*) edge
->aux
;
1318 gcc_checking_assert (n
->get_data () == edge
);
1320 /* fibonacci_heap::replace_key does busy updating of the
1321 heap that is unnecesarily expensive.
1322 We do lazy increases: after extracting minimum if the key
1323 turns out to be out of date, it is re-inserted into heap
1324 with correct value. */
1325 if (badness
< n
->get_key ())
1327 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1330 " decreasing badness %s -> %s, %f to %f\n",
1331 edge
->caller
->dump_name (),
1332 edge
->callee
->dump_name (),
1333 n
->get_key ().to_double (),
1334 badness
.to_double ());
1336 heap
->decrease_key (n
, badness
);
1341 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1344 " enqueuing call %s -> %s, badness %f\n",
1345 edge
->caller
->dump_name (),
1346 edge
->callee
->dump_name (),
1347 badness
.to_double ());
1349 edge
->aux
= heap
->insert (badness
, edge
);
1354 /* NODE was inlined.
1355 All caller edges needs to be resetted because
1356 size estimates change. Similarly callees needs reset
1357 because better context may be known. */
1360 reset_edge_caches (struct cgraph_node
*node
)
1362 struct cgraph_edge
*edge
;
1363 struct cgraph_edge
*e
= node
->callees
;
1364 struct cgraph_node
*where
= node
;
1365 struct ipa_ref
*ref
;
1367 if (where
->inlined_to
)
1368 where
= where
->inlined_to
;
1370 reset_node_cache (where
);
1372 if (edge_growth_cache
!= NULL
)
1373 for (edge
= where
->callers
; edge
; edge
= edge
->next_caller
)
1374 if (edge
->inline_failed
)
1375 edge_growth_cache
->remove (edge
);
1377 FOR_EACH_ALIAS (where
, ref
)
1378 reset_edge_caches (dyn_cast
<cgraph_node
*> (ref
->referring
));
1384 if (!e
->inline_failed
&& e
->callee
->callees
)
1385 e
= e
->callee
->callees
;
1388 if (edge_growth_cache
!= NULL
&& e
->inline_failed
)
1389 edge_growth_cache
->remove (e
);
1396 if (e
->caller
== node
)
1398 e
= e
->caller
->callers
;
1400 while (!e
->next_callee
);
1406 /* Recompute HEAP nodes for each of caller of NODE.
1407 UPDATED_NODES track nodes we already visited, to avoid redundant work.
1408 When CHECK_INLINABLITY_FOR is set, re-check for specified edge that
1409 it is inlinable. Otherwise check all edges. */
1412 update_caller_keys (edge_heap_t
*heap
, struct cgraph_node
*node
,
1413 bitmap updated_nodes
,
1414 struct cgraph_edge
*check_inlinablity_for
)
1416 struct cgraph_edge
*edge
;
1417 struct ipa_ref
*ref
;
1419 if ((!node
->alias
&& !ipa_fn_summaries
->get (node
)->inlinable
)
1420 || node
->inlined_to
)
1422 if (!bitmap_set_bit (updated_nodes
, node
->get_uid ()))
1425 FOR_EACH_ALIAS (node
, ref
)
1427 struct cgraph_node
*alias
= dyn_cast
<cgraph_node
*> (ref
->referring
);
1428 update_caller_keys (heap
, alias
, updated_nodes
, check_inlinablity_for
);
1431 for (edge
= node
->callers
; edge
; edge
= edge
->next_caller
)
1432 if (edge
->inline_failed
)
1434 if (!check_inlinablity_for
1435 || check_inlinablity_for
== edge
)
1437 if (can_inline_edge_p (edge
, false)
1438 && want_inline_small_function_p (edge
, false)
1439 && can_inline_edge_by_limits_p (edge
, false))
1440 update_edge_key (heap
, edge
);
1443 report_inline_failed_reason (edge
);
1444 heap
->delete_node ((edge_heap_node_t
*) edge
->aux
);
1449 update_edge_key (heap
, edge
);
1453 /* Recompute HEAP nodes for each uninlined call in NODE.
1454 This is used when we know that edge badnesses are going only to increase
1455 (we introduced new call site) and thus all we need is to insert newly
1456 created edges into heap. */
1459 update_callee_keys (edge_heap_t
*heap
, struct cgraph_node
*node
,
1460 bitmap updated_nodes
)
1462 struct cgraph_edge
*e
= node
->callees
;
1467 if (!e
->inline_failed
&& e
->callee
->callees
)
1468 e
= e
->callee
->callees
;
1471 enum availability avail
;
1472 struct cgraph_node
*callee
;
1473 /* We do not reset callee growth cache here. Since we added a new call,
1474 growth chould have just increased and consequentely badness metric
1475 don't need updating. */
1476 if (e
->inline_failed
1477 && (callee
= e
->callee
->ultimate_alias_target (&avail
, e
->caller
))
1478 && ipa_fn_summaries
->get (callee
) != NULL
1479 && ipa_fn_summaries
->get (callee
)->inlinable
1480 && avail
>= AVAIL_AVAILABLE
1481 && !bitmap_bit_p (updated_nodes
, callee
->get_uid ()))
1483 if (can_inline_edge_p (e
, false)
1484 && want_inline_small_function_p (e
, false)
1485 && can_inline_edge_by_limits_p (e
, false))
1486 update_edge_key (heap
, e
);
1489 report_inline_failed_reason (e
);
1490 heap
->delete_node ((edge_heap_node_t
*) e
->aux
);
1500 if (e
->caller
== node
)
1502 e
= e
->caller
->callers
;
1504 while (!e
->next_callee
);
1510 /* Enqueue all recursive calls from NODE into priority queue depending on
1511 how likely we want to recursively inline the call. */
1514 lookup_recursive_calls (struct cgraph_node
*node
, struct cgraph_node
*where
,
1517 struct cgraph_edge
*e
;
1518 enum availability avail
;
1520 for (e
= where
->callees
; e
; e
= e
->next_callee
)
1521 if (e
->callee
== node
1522 || (e
->callee
->ultimate_alias_target (&avail
, e
->caller
) == node
1523 && avail
> AVAIL_INTERPOSABLE
))
1524 heap
->insert (-e
->sreal_frequency (), e
);
1525 for (e
= where
->callees
; e
; e
= e
->next_callee
)
1526 if (!e
->inline_failed
)
1527 lookup_recursive_calls (node
, e
->callee
, heap
);
1530 /* Decide on recursive inlining: in the case function has recursive calls,
1531 inline until body size reaches given argument. If any new indirect edges
1532 are discovered in the process, add them to *NEW_EDGES, unless NEW_EDGES
1536 recursive_inlining (struct cgraph_edge
*edge
,
1537 vec
<cgraph_edge
*> *new_edges
)
1539 int limit
= param_max_inline_insns_recursive_auto
;
1540 edge_heap_t
heap (sreal::min ());
1541 struct cgraph_node
*node
;
1542 struct cgraph_edge
*e
;
1543 struct cgraph_node
*master_clone
= NULL
, *next
;
1547 node
= edge
->caller
;
1548 if (node
->inlined_to
)
1549 node
= node
->inlined_to
;
1551 if (DECL_DECLARED_INLINE_P (node
->decl
))
1552 limit
= param_max_inline_insns_recursive
;
1554 /* Make sure that function is small enough to be considered for inlining. */
1555 if (estimate_size_after_inlining (node
, edge
) >= limit
)
1557 lookup_recursive_calls (node
, node
, &heap
);
1563 " Performing recursive inlining on %s\n",
1566 /* Do the inlining and update list of recursive call during process. */
1567 while (!heap
.empty ())
1569 struct cgraph_edge
*curr
= heap
.extract_min ();
1570 struct cgraph_node
*cnode
, *dest
= curr
->callee
;
1572 if (!can_inline_edge_p (curr
, true)
1573 || !can_inline_edge_by_limits_p (curr
, true))
1576 /* MASTER_CLONE is produced in the case we already started modified
1577 the function. Be sure to redirect edge to the original body before
1578 estimating growths otherwise we will be seeing growths after inlining
1579 the already modified body. */
1582 curr
->redirect_callee (master_clone
);
1583 if (edge_growth_cache
!= NULL
)
1584 edge_growth_cache
->remove (curr
);
1587 if (estimate_size_after_inlining (node
, curr
) > limit
)
1589 curr
->redirect_callee (dest
);
1590 if (edge_growth_cache
!= NULL
)
1591 edge_growth_cache
->remove (curr
);
1596 for (cnode
= curr
->caller
;
1597 cnode
->inlined_to
; cnode
= cnode
->callers
->caller
)
1599 == curr
->callee
->ultimate_alias_target ()->decl
)
1602 if (!want_inline_self_recursive_call_p (curr
, node
, false, depth
))
1604 curr
->redirect_callee (dest
);
1605 if (edge_growth_cache
!= NULL
)
1606 edge_growth_cache
->remove (curr
);
1613 " Inlining call of depth %i", depth
);
1614 if (node
->count
.nonzero_p () && curr
->count
.initialized_p ())
1616 fprintf (dump_file
, " called approx. %.2f times per call",
1617 (double)curr
->count
.to_gcov_type ()
1618 / node
->count
.to_gcov_type ());
1620 fprintf (dump_file
, "\n");
1624 /* We need original clone to copy around. */
1625 master_clone
= node
->create_clone (node
->decl
, node
->count
,
1626 false, vNULL
, true, NULL
, NULL
);
1627 for (e
= master_clone
->callees
; e
; e
= e
->next_callee
)
1628 if (!e
->inline_failed
)
1629 clone_inlined_nodes (e
, true, false, NULL
);
1630 curr
->redirect_callee (master_clone
);
1631 if (edge_growth_cache
!= NULL
)
1632 edge_growth_cache
->remove (curr
);
1635 inline_call (curr
, false, new_edges
, &overall_size
, true);
1636 reset_node_cache (node
);
1637 lookup_recursive_calls (node
, curr
->callee
, &heap
);
1641 if (!heap
.empty () && dump_file
)
1642 fprintf (dump_file
, " Recursive inlining growth limit met.\n");
1647 if (dump_enabled_p ())
1648 dump_printf_loc (MSG_NOTE
, edge
->call_stmt
,
1649 "\n Inlined %i times, "
1650 "body grown from size %i to %i, time %f to %f\n", n
,
1651 ipa_size_summaries
->get (master_clone
)->size
,
1652 ipa_size_summaries
->get (node
)->size
,
1653 ipa_fn_summaries
->get (master_clone
)->time
.to_double (),
1654 ipa_fn_summaries
->get (node
)->time
.to_double ());
1656 /* Remove master clone we used for inlining. We rely that clones inlined
1657 into master clone gets queued just before master clone so we don't
1659 for (node
= symtab
->first_function (); node
!= master_clone
;
1662 next
= symtab
->next_function (node
);
1663 if (node
->inlined_to
== master_clone
)
1666 master_clone
->remove ();
1671 /* Given whole compilation unit estimate of INSNS, compute how large we can
1672 allow the unit to grow. */
1675 compute_max_insns (int insns
)
1677 int max_insns
= insns
;
1678 if (max_insns
< param_large_unit_insns
)
1679 max_insns
= param_large_unit_insns
;
1681 return ((int64_t) max_insns
1682 * (100 + param_inline_unit_growth
) / 100);
1686 /* Compute badness of all edges in NEW_EDGES and add them to the HEAP. */
1689 add_new_edges_to_heap (edge_heap_t
*heap
, vec
<cgraph_edge
*> new_edges
)
1691 while (new_edges
.length () > 0)
1693 struct cgraph_edge
*edge
= new_edges
.pop ();
1695 gcc_assert (!edge
->aux
);
1696 gcc_assert (edge
->callee
);
1697 if (edge
->inline_failed
1698 && can_inline_edge_p (edge
, true)
1699 && want_inline_small_function_p (edge
, true)
1700 && can_inline_edge_by_limits_p (edge
, true))
1701 edge
->aux
= heap
->insert (edge_badness (edge
, false), edge
);
1705 /* Remove EDGE from the fibheap. */
1708 heap_edge_removal_hook (struct cgraph_edge
*e
, void *data
)
1712 ((edge_heap_t
*)data
)->delete_node ((edge_heap_node_t
*)e
->aux
);
1717 /* Return true if speculation of edge E seems useful.
1718 If ANTICIPATE_INLINING is true, be conservative and hope that E
1722 speculation_useful_p (struct cgraph_edge
*e
, bool anticipate_inlining
)
1724 /* If we have already decided to inline the edge, it seems useful. */
1725 if (!e
->inline_failed
)
1728 enum availability avail
;
1729 struct cgraph_node
*target
= e
->callee
->ultimate_alias_target (&avail
,
1731 struct cgraph_edge
*direct
, *indirect
;
1732 struct ipa_ref
*ref
;
1734 gcc_assert (e
->speculative
&& !e
->indirect_unknown_callee
);
1736 if (!e
->maybe_hot_p ())
1739 /* See if IP optimizations found something potentially useful about the
1740 function. For now we look only for CONST/PURE flags. Almost everything
1741 else we propagate is useless. */
1742 if (avail
>= AVAIL_AVAILABLE
)
1744 int ecf_flags
= flags_from_decl_or_type (target
->decl
);
1745 if (ecf_flags
& ECF_CONST
)
1747 e
->speculative_call_info (direct
, indirect
, ref
);
1748 if (!(indirect
->indirect_info
->ecf_flags
& ECF_CONST
))
1751 else if (ecf_flags
& ECF_PURE
)
1753 e
->speculative_call_info (direct
, indirect
, ref
);
1754 if (!(indirect
->indirect_info
->ecf_flags
& ECF_PURE
))
1758 /* If we did not managed to inline the function nor redirect
1759 to an ipa-cp clone (that are seen by having local flag set),
1760 it is probably pointless to inline it unless hardware is missing
1761 indirect call predictor. */
1762 if (!anticipate_inlining
&& !target
->local
)
1764 /* For overwritable targets there is not much to do. */
1765 if (!can_inline_edge_p (e
, false)
1766 || !can_inline_edge_by_limits_p (e
, false, true))
1768 /* OK, speculation seems interesting. */
1772 /* We know that EDGE is not going to be inlined.
1773 See if we can remove speculation. */
1776 resolve_noninline_speculation (edge_heap_t
*edge_heap
, struct cgraph_edge
*edge
)
1778 if (edge
->speculative
&& !speculation_useful_p (edge
, false))
1780 struct cgraph_node
*node
= edge
->caller
;
1781 struct cgraph_node
*where
= node
->inlined_to
1782 ? node
->inlined_to
: node
;
1783 auto_bitmap updated_nodes
;
1785 if (edge
->count
.ipa ().initialized_p ())
1786 spec_rem
+= edge
->count
.ipa ();
1787 edge
->resolve_speculation ();
1788 reset_edge_caches (where
);
1789 ipa_update_overall_fn_summary (where
);
1790 update_caller_keys (edge_heap
, where
,
1791 updated_nodes
, NULL
);
1792 update_callee_keys (edge_heap
, where
,
1797 /* Return true if NODE should be accounted for overall size estimate.
1798 Skip all nodes optimized for size so we can measure the growth of hot
1799 part of program no matter of the padding. */
1802 inline_account_function_p (struct cgraph_node
*node
)
1804 return (!DECL_EXTERNAL (node
->decl
)
1805 && !opt_for_fn (node
->decl
, optimize_size
)
1806 && node
->frequency
!= NODE_FREQUENCY_UNLIKELY_EXECUTED
);
1809 /* Count number of callers of NODE and store it into DATA (that
1810 points to int. Worker for cgraph_for_node_and_aliases. */
1813 sum_callers (struct cgraph_node
*node
, void *data
)
1815 struct cgraph_edge
*e
;
1816 int *num_calls
= (int *)data
;
1818 for (e
= node
->callers
; e
; e
= e
->next_caller
)
1823 /* We only propagate across edges with non-interposable callee. */
1826 ignore_edge_p (struct cgraph_edge
*e
)
1828 enum availability avail
;
1829 e
->callee
->function_or_virtual_thunk_symbol (&avail
, e
->caller
);
1830 return (avail
<= AVAIL_INTERPOSABLE
);
1833 /* We use greedy algorithm for inlining of small functions:
1834 All inline candidates are put into prioritized heap ordered in
1837 The inlining of small functions is bounded by unit growth parameters. */
1840 inline_small_functions (void)
1842 struct cgraph_node
*node
;
1843 struct cgraph_edge
*edge
;
1844 edge_heap_t
edge_heap (sreal::min ());
1845 auto_bitmap updated_nodes
;
1846 int min_size
, max_size
;
1847 auto_vec
<cgraph_edge
*> new_indirect_edges
;
1848 int initial_size
= 0;
1849 struct cgraph_node
**order
= XCNEWVEC (cgraph_node
*, symtab
->cgraph_count
);
1850 struct cgraph_edge_hook_list
*edge_removal_hook_holder
;
1851 new_indirect_edges
.create (8);
1853 edge_removal_hook_holder
1854 = symtab
->add_edge_removal_hook (&heap_edge_removal_hook
, &edge_heap
);
1856 /* Compute overall unit size and other global parameters used by badness
1859 max_count
= profile_count::uninitialized ();
1860 ipa_reduced_postorder (order
, true, ignore_edge_p
);
1863 FOR_EACH_DEFINED_FUNCTION (node
)
1864 if (!node
->inlined_to
)
1866 if (!node
->alias
&& node
->analyzed
1867 && (node
->has_gimple_body_p () || node
->thunk
.thunk_p
)
1868 && opt_for_fn (node
->decl
, optimize
))
1870 class ipa_fn_summary
*info
= ipa_fn_summaries
->get (node
);
1871 struct ipa_dfs_info
*dfs
= (struct ipa_dfs_info
*) node
->aux
;
1873 /* Do not account external functions, they will be optimized out
1874 if not inlined. Also only count the non-cold portion of program. */
1875 if (inline_account_function_p (node
))
1876 initial_size
+= ipa_size_summaries
->get (node
)->size
;
1877 info
->growth
= estimate_growth (node
);
1880 node
->call_for_symbol_and_aliases (sum_callers
, &num_calls
,
1883 info
->single_caller
= true;
1884 if (dfs
&& dfs
->next_cycle
)
1886 struct cgraph_node
*n2
;
1887 int id
= dfs
->scc_no
+ 1;
1889 n2
= ((struct ipa_dfs_info
*) n2
->aux
)->next_cycle
)
1890 if (opt_for_fn (n2
->decl
, optimize
))
1892 ipa_fn_summary
*info2
= ipa_fn_summaries
->get
1893 (n2
->inlined_to
? n2
->inlined_to
: n2
);
1901 for (edge
= node
->callers
; edge
; edge
= edge
->next_caller
)
1902 max_count
= max_count
.max (edge
->count
.ipa ());
1904 ipa_free_postorder_info ();
1905 initialize_growth_caches ();
1909 "\nDeciding on inlining of small functions. Starting with size %i.\n",
1912 overall_size
= initial_size
;
1913 max_size
= compute_max_insns (overall_size
);
1914 min_size
= overall_size
;
1916 /* Populate the heap with all edges we might inline. */
1918 FOR_EACH_DEFINED_FUNCTION (node
)
1920 bool update
= false;
1921 struct cgraph_edge
*next
= NULL
;
1922 bool has_speculative
= false;
1924 if (!opt_for_fn (node
->decl
, optimize
))
1928 fprintf (dump_file
, "Enqueueing calls in %s.\n", node
->dump_name ());
1930 for (edge
= node
->callees
; edge
; edge
= next
)
1932 next
= edge
->next_callee
;
1933 if (edge
->inline_failed
1935 && can_inline_edge_p (edge
, true)
1936 && want_inline_small_function_p (edge
, true)
1937 && can_inline_edge_by_limits_p (edge
, true)
1938 && edge
->inline_failed
)
1940 gcc_assert (!edge
->aux
);
1941 update_edge_key (&edge_heap
, edge
);
1943 if (edge
->speculative
)
1944 has_speculative
= true;
1946 if (has_speculative
)
1947 for (edge
= node
->callees
; edge
; edge
= next
)
1948 if (edge
->speculative
&& !speculation_useful_p (edge
,
1951 edge
->resolve_speculation ();
1956 struct cgraph_node
*where
= node
->inlined_to
1957 ? node
->inlined_to
: node
;
1958 ipa_update_overall_fn_summary (where
);
1959 reset_edge_caches (where
);
1960 update_caller_keys (&edge_heap
, where
,
1961 updated_nodes
, NULL
);
1962 update_callee_keys (&edge_heap
, where
,
1964 bitmap_clear (updated_nodes
);
1968 gcc_assert (in_lto_p
1970 || (profile_info
&& flag_branch_probabilities
));
1972 while (!edge_heap
.empty ())
1974 int old_size
= overall_size
;
1975 struct cgraph_node
*where
, *callee
;
1976 sreal badness
= edge_heap
.min_key ();
1977 sreal current_badness
;
1980 edge
= edge_heap
.extract_min ();
1981 gcc_assert (edge
->aux
);
1983 if (!edge
->inline_failed
|| !edge
->callee
->analyzed
)
1986 /* Be sure that caches are maintained consistent.
1987 This check is affected by scaling roundoff errors when compiling for
1988 IPA this we skip it in that case. */
1989 if (flag_checking
&& !edge
->callee
->count
.ipa_p ()
1990 && (!max_count
.initialized_p () || !max_count
.nonzero_p ()))
1992 sreal cached_badness
= edge_badness (edge
, false);
1994 int old_size_est
= estimate_edge_size (edge
);
1995 sreal old_time_est
= estimate_edge_time (edge
);
1996 int old_hints_est
= estimate_edge_hints (edge
);
1998 if (edge_growth_cache
!= NULL
)
1999 edge_growth_cache
->remove (edge
);
2000 reset_node_cache (edge
->caller
->inlined_to
2001 ? edge
->caller
->inlined_to
2003 gcc_assert (old_size_est
== estimate_edge_size (edge
));
2004 gcc_assert (old_time_est
== estimate_edge_time (edge
));
2007 gcc_assert (old_hints_est == estimate_edge_hints (edge));
2009 fails with profile feedback because some hints depends on
2010 maybe_hot_edge_p predicate and because callee gets inlined to other
2011 calls, the edge may become cold.
2012 This ought to be fixed by computing relative probabilities
2013 for given invocation but that will be better done once whole
2014 code is converted to sreals. Disable for now and revert to "wrong"
2015 value so enable/disable checking paths agree. */
2016 edge_growth_cache
->get (edge
)->hints
= old_hints_est
+ 1;
2018 /* When updating the edge costs, we only decrease badness in the keys.
2019 Increases of badness are handled lazilly; when we see key with out
2020 of date value on it, we re-insert it now. */
2021 current_badness
= edge_badness (edge
, false);
2022 gcc_assert (cached_badness
== current_badness
);
2023 gcc_assert (current_badness
>= badness
);
2026 current_badness
= edge_badness (edge
, false);
2027 if (current_badness
!= badness
)
2029 if (edge_heap
.min () && current_badness
> edge_heap
.min_key ())
2031 edge
->aux
= edge_heap
.insert (current_badness
, edge
);
2035 badness
= current_badness
;
2038 if (!can_inline_edge_p (edge
, true)
2039 || !can_inline_edge_by_limits_p (edge
, true))
2041 resolve_noninline_speculation (&edge_heap
, edge
);
2045 callee
= edge
->callee
->ultimate_alias_target ();
2046 growth
= estimate_edge_growth (edge
);
2050 "\nConsidering %s with %i size\n",
2051 callee
->dump_name (),
2052 ipa_size_summaries
->get (callee
)->size
);
2054 " to be inlined into %s in %s:%i\n"
2055 " Estimated badness is %f, frequency %.2f.\n",
2056 edge
->caller
->dump_name (),
2058 && (LOCATION_LOCUS (gimple_location ((const gimple
*)
2060 > BUILTINS_LOCATION
)
2061 ? gimple_filename ((const gimple
*) edge
->call_stmt
)
2064 ? gimple_lineno ((const gimple
*) edge
->call_stmt
)
2066 badness
.to_double (),
2067 edge
->sreal_frequency ().to_double ());
2068 if (edge
->count
.ipa ().initialized_p ())
2070 fprintf (dump_file
, " Called ");
2071 edge
->count
.ipa ().dump (dump_file
);
2072 fprintf (dump_file
, " times\n");
2074 if (dump_flags
& TDF_DETAILS
)
2075 edge_badness (edge
, true);
2078 if (overall_size
+ growth
> max_size
2079 && !DECL_DISREGARD_INLINE_LIMITS (callee
->decl
))
2081 edge
->inline_failed
= CIF_INLINE_UNIT_GROWTH_LIMIT
;
2082 report_inline_failed_reason (edge
);
2083 resolve_noninline_speculation (&edge_heap
, edge
);
2087 if (!want_inline_small_function_p (edge
, true))
2089 resolve_noninline_speculation (&edge_heap
, edge
);
2093 /* Heuristics for inlining small functions work poorly for
2094 recursive calls where we do effects similar to loop unrolling.
2095 When inlining such edge seems profitable, leave decision on
2096 specific inliner. */
2097 if (edge
->recursive_p ())
2099 where
= edge
->caller
;
2100 if (where
->inlined_to
)
2101 where
= where
->inlined_to
;
2102 if (!recursive_inlining (edge
,
2103 opt_for_fn (edge
->caller
->decl
,
2104 flag_indirect_inlining
)
2105 ? &new_indirect_edges
: NULL
))
2107 edge
->inline_failed
= CIF_RECURSIVE_INLINING
;
2108 resolve_noninline_speculation (&edge_heap
, edge
);
2111 reset_edge_caches (where
);
2112 /* Recursive inliner inlines all recursive calls of the function
2113 at once. Consequently we need to update all callee keys. */
2114 if (opt_for_fn (edge
->caller
->decl
, flag_indirect_inlining
))
2115 add_new_edges_to_heap (&edge_heap
, new_indirect_edges
);
2116 update_callee_keys (&edge_heap
, where
, updated_nodes
);
2117 bitmap_clear (updated_nodes
);
2121 struct cgraph_node
*outer_node
= NULL
;
2124 /* Consider the case where self recursive function A is inlined
2125 into B. This is desired optimization in some cases, since it
2126 leads to effect similar of loop peeling and we might completely
2127 optimize out the recursive call. However we must be extra
2130 where
= edge
->caller
;
2131 while (where
->inlined_to
)
2133 if (where
->decl
== callee
->decl
)
2134 outer_node
= where
, depth
++;
2135 where
= where
->callers
->caller
;
2138 && !want_inline_self_recursive_call_p (edge
, outer_node
,
2142 = (DECL_DISREGARD_INLINE_LIMITS (edge
->callee
->decl
)
2143 ? CIF_RECURSIVE_INLINING
: CIF_UNSPECIFIED
);
2144 resolve_noninline_speculation (&edge_heap
, edge
);
2147 else if (depth
&& dump_file
)
2148 fprintf (dump_file
, " Peeling recursion with depth %i\n", depth
);
2150 gcc_checking_assert (!callee
->inlined_to
);
2151 inline_call (edge
, true, &new_indirect_edges
, &overall_size
, true);
2152 add_new_edges_to_heap (&edge_heap
, new_indirect_edges
);
2154 reset_edge_caches (edge
->callee
);
2156 update_callee_keys (&edge_heap
, where
, updated_nodes
);
2158 where
= edge
->caller
;
2159 if (where
->inlined_to
)
2160 where
= where
->inlined_to
;
2162 /* Our profitability metric can depend on local properties
2163 such as number of inlinable calls and size of the function body.
2164 After inlining these properties might change for the function we
2165 inlined into (since it's body size changed) and for the functions
2166 called by function we inlined (since number of it inlinable callers
2168 update_caller_keys (&edge_heap
, where
, updated_nodes
, NULL
);
2169 /* Offline copy count has possibly changed, recompute if profile is
2171 struct cgraph_node
*n
= cgraph_node::get (edge
->callee
->decl
);
2172 if (n
!= edge
->callee
&& n
->analyzed
&& n
->count
.ipa ().initialized_p ())
2173 update_callee_keys (&edge_heap
, n
, updated_nodes
);
2174 bitmap_clear (updated_nodes
);
2176 if (dump_enabled_p ())
2178 ipa_fn_summary
*s
= ipa_fn_summaries
->get (where
);
2180 /* dump_printf can't handle %+i. */
2181 char buf_net_change
[100];
2182 snprintf (buf_net_change
, sizeof buf_net_change
, "%+i",
2183 overall_size
- old_size
);
2185 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS
, edge
->call_stmt
,
2186 " Inlined %C into %C which now has time %f and "
2187 "size %i, net change of %s.\n",
2188 edge
->callee
, edge
->caller
,
2189 s
->time
.to_double (),
2190 ipa_size_summaries
->get (edge
->caller
)->size
,
2193 if (min_size
> overall_size
)
2195 min_size
= overall_size
;
2196 max_size
= compute_max_insns (min_size
);
2199 fprintf (dump_file
, "New minimal size reached: %i\n", min_size
);
2203 free_growth_caches ();
2204 if (dump_enabled_p ())
2205 dump_printf (MSG_NOTE
,
2206 "Unit growth for small function inlining: %i->%i (%i%%)\n",
2207 initial_size
, overall_size
,
2208 initial_size
? overall_size
* 100 / (initial_size
) - 100: 0);
2209 symtab
->remove_edge_removal_hook (edge_removal_hook_holder
);
2212 /* Flatten NODE. Performed both during early inlining and
2213 at IPA inlining time. */
2216 flatten_function (struct cgraph_node
*node
, bool early
, bool update
)
2218 struct cgraph_edge
*e
;
2220 /* We shouldn't be called recursively when we are being processed. */
2221 gcc_assert (node
->aux
== NULL
);
2223 node
->aux
= (void *) node
;
2225 for (e
= node
->callees
; e
; e
= e
->next_callee
)
2227 struct cgraph_node
*orig_callee
;
2228 struct cgraph_node
*callee
= e
->callee
->ultimate_alias_target ();
2230 /* We've hit cycle? It is time to give up. */
2233 if (dump_enabled_p ())
2234 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, e
->call_stmt
,
2235 "Not inlining %C into %C to avoid cycle.\n",
2237 if (cgraph_inline_failed_type (e
->inline_failed
) != CIF_FINAL_ERROR
)
2238 e
->inline_failed
= CIF_RECURSIVE_INLINING
;
2242 /* When the edge is already inlined, we just need to recurse into
2243 it in order to fully flatten the leaves. */
2244 if (!e
->inline_failed
)
2246 flatten_function (callee
, early
, false);
2250 /* Flatten attribute needs to be processed during late inlining. For
2251 extra code quality we however do flattening during early optimization,
2254 ? !can_inline_edge_p (e
, true)
2255 && !can_inline_edge_by_limits_p (e
, true)
2256 : !can_early_inline_edge_p (e
))
2259 if (e
->recursive_p ())
2261 if (dump_enabled_p ())
2262 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, e
->call_stmt
,
2263 "Not inlining: recursive call.\n");
2267 if (gimple_in_ssa_p (DECL_STRUCT_FUNCTION (node
->decl
))
2268 != gimple_in_ssa_p (DECL_STRUCT_FUNCTION (callee
->decl
)))
2270 if (dump_enabled_p ())
2271 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, e
->call_stmt
,
2272 "Not inlining: SSA form does not match.\n");
2276 /* Inline the edge and flatten the inline clone. Avoid
2277 recursing through the original node if the node was cloned. */
2278 if (dump_enabled_p ())
2279 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS
, e
->call_stmt
,
2280 " Inlining %C into %C.\n",
2282 orig_callee
= callee
;
2283 inline_call (e
, true, NULL
, NULL
, false);
2284 if (e
->callee
!= orig_callee
)
2285 orig_callee
->aux
= (void *) node
;
2286 flatten_function (e
->callee
, early
, false);
2287 if (e
->callee
!= orig_callee
)
2288 orig_callee
->aux
= NULL
;
2292 cgraph_node
*where
= node
->inlined_to
? node
->inlined_to
: node
;
2293 if (update
&& opt_for_fn (where
->decl
, optimize
))
2294 ipa_update_overall_fn_summary (where
);
2297 /* Inline NODE to all callers. Worker for cgraph_for_node_and_aliases.
2298 DATA points to number of calls originally found so we avoid infinite
2302 inline_to_all_callers_1 (struct cgraph_node
*node
, void *data
,
2303 hash_set
<cgraph_node
*> *callers
)
2305 int *num_calls
= (int *)data
;
2306 bool callee_removed
= false;
2308 while (node
->callers
&& !node
->inlined_to
)
2310 struct cgraph_node
*caller
= node
->callers
->caller
;
2312 if (!can_inline_edge_p (node
->callers
, true)
2313 || !can_inline_edge_by_limits_p (node
->callers
, true)
2314 || node
->callers
->recursive_p ())
2317 fprintf (dump_file
, "Uninlinable call found; giving up.\n");
2324 cgraph_node
*ultimate
= node
->ultimate_alias_target ();
2326 "\nInlining %s size %i.\n",
2328 ipa_size_summaries
->get (ultimate
)->size
);
2330 " Called once from %s %i insns.\n",
2331 node
->callers
->caller
->name (),
2332 ipa_size_summaries
->get (node
->callers
->caller
)->size
);
2335 /* Remember which callers we inlined to, delaying updating the
2337 callers
->add (node
->callers
->caller
);
2338 inline_call (node
->callers
, true, NULL
, NULL
, false, &callee_removed
);
2341 " Inlined into %s which now has %i size\n",
2343 ipa_size_summaries
->get (caller
)->size
);
2344 if (!(*num_calls
)--)
2347 fprintf (dump_file
, "New calls found; giving up.\n");
2348 return callee_removed
;
2356 /* Wrapper around inline_to_all_callers_1 doing delayed overall summary
2360 inline_to_all_callers (struct cgraph_node
*node
, void *data
)
2362 hash_set
<cgraph_node
*> callers
;
2363 bool res
= inline_to_all_callers_1 (node
, data
, &callers
);
2364 /* Perform the delayed update of the overall summary of all callers
2365 processed. This avoids quadratic behavior in the cases where
2366 we have a lot of calls to the same function. */
2367 for (hash_set
<cgraph_node
*>::iterator i
= callers
.begin ();
2368 i
!= callers
.end (); ++i
)
2369 ipa_update_overall_fn_summary ((*i
)->inlined_to
? (*i
)->inlined_to
: *i
);
2373 /* Output overall time estimate. */
2375 dump_overall_stats (void)
2377 sreal sum_weighted
= 0, sum
= 0;
2378 struct cgraph_node
*node
;
2380 FOR_EACH_DEFINED_FUNCTION (node
)
2381 if (!node
->inlined_to
2384 ipa_fn_summary
*s
= ipa_fn_summaries
->get (node
);
2388 if (node
->count
.ipa ().initialized_p ())
2389 sum_weighted
+= s
->time
* node
->count
.ipa ().to_gcov_type ();
2392 fprintf (dump_file
, "Overall time estimate: "
2393 "%f weighted by profile: "
2394 "%f\n", sum
.to_double (), sum_weighted
.to_double ());
2397 /* Output some useful stats about inlining. */
2400 dump_inline_stats (void)
2402 int64_t inlined_cnt
= 0, inlined_indir_cnt
= 0;
2403 int64_t inlined_virt_cnt
= 0, inlined_virt_indir_cnt
= 0;
2404 int64_t noninlined_cnt
= 0, noninlined_indir_cnt
= 0;
2405 int64_t noninlined_virt_cnt
= 0, noninlined_virt_indir_cnt
= 0;
2406 int64_t inlined_speculative
= 0, inlined_speculative_ply
= 0;
2407 int64_t indirect_poly_cnt
= 0, indirect_cnt
= 0;
2408 int64_t reason
[CIF_N_REASONS
][2];
2409 sreal reason_freq
[CIF_N_REASONS
];
2411 struct cgraph_node
*node
;
2413 memset (reason
, 0, sizeof (reason
));
2414 for (i
=0; i
< CIF_N_REASONS
; i
++)
2416 FOR_EACH_DEFINED_FUNCTION (node
)
2418 struct cgraph_edge
*e
;
2419 for (e
= node
->callees
; e
; e
= e
->next_callee
)
2421 if (e
->inline_failed
)
2423 if (e
->count
.ipa ().initialized_p ())
2424 reason
[(int) e
->inline_failed
][0] += e
->count
.ipa ().to_gcov_type ();
2425 reason_freq
[(int) e
->inline_failed
] += e
->sreal_frequency ();
2426 reason
[(int) e
->inline_failed
][1] ++;
2427 if (DECL_VIRTUAL_P (e
->callee
->decl
)
2428 && e
->count
.ipa ().initialized_p ())
2430 if (e
->indirect_inlining_edge
)
2431 noninlined_virt_indir_cnt
+= e
->count
.ipa ().to_gcov_type ();
2433 noninlined_virt_cnt
+= e
->count
.ipa ().to_gcov_type ();
2435 else if (e
->count
.ipa ().initialized_p ())
2437 if (e
->indirect_inlining_edge
)
2438 noninlined_indir_cnt
+= e
->count
.ipa ().to_gcov_type ();
2440 noninlined_cnt
+= e
->count
.ipa ().to_gcov_type ();
2443 else if (e
->count
.ipa ().initialized_p ())
2447 if (DECL_VIRTUAL_P (e
->callee
->decl
))
2448 inlined_speculative_ply
+= e
->count
.ipa ().to_gcov_type ();
2450 inlined_speculative
+= e
->count
.ipa ().to_gcov_type ();
2452 else if (DECL_VIRTUAL_P (e
->callee
->decl
))
2454 if (e
->indirect_inlining_edge
)
2455 inlined_virt_indir_cnt
+= e
->count
.ipa ().to_gcov_type ();
2457 inlined_virt_cnt
+= e
->count
.ipa ().to_gcov_type ();
2461 if (e
->indirect_inlining_edge
)
2462 inlined_indir_cnt
+= e
->count
.ipa ().to_gcov_type ();
2464 inlined_cnt
+= e
->count
.ipa ().to_gcov_type ();
2468 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
2469 if (e
->indirect_info
->polymorphic
2470 & e
->count
.ipa ().initialized_p ())
2471 indirect_poly_cnt
+= e
->count
.ipa ().to_gcov_type ();
2472 else if (e
->count
.ipa ().initialized_p ())
2473 indirect_cnt
+= e
->count
.ipa ().to_gcov_type ();
2475 if (max_count
.initialized_p ())
2478 "Inlined %" PRId64
" + speculative "
2479 "%" PRId64
" + speculative polymorphic "
2480 "%" PRId64
" + previously indirect "
2481 "%" PRId64
" + virtual "
2482 "%" PRId64
" + virtual and previously indirect "
2483 "%" PRId64
"\n" "Not inlined "
2484 "%" PRId64
" + previously indirect "
2485 "%" PRId64
" + virtual "
2486 "%" PRId64
" + virtual and previously indirect "
2487 "%" PRId64
" + stil indirect "
2488 "%" PRId64
" + still indirect polymorphic "
2489 "%" PRId64
"\n", inlined_cnt
,
2490 inlined_speculative
, inlined_speculative_ply
,
2491 inlined_indir_cnt
, inlined_virt_cnt
, inlined_virt_indir_cnt
,
2492 noninlined_cnt
, noninlined_indir_cnt
, noninlined_virt_cnt
,
2493 noninlined_virt_indir_cnt
, indirect_cnt
, indirect_poly_cnt
);
2494 fprintf (dump_file
, "Removed speculations ");
2495 spec_rem
.dump (dump_file
);
2496 fprintf (dump_file
, "\n");
2498 dump_overall_stats ();
2499 fprintf (dump_file
, "\nWhy inlining failed?\n");
2500 for (i
= 0; i
< CIF_N_REASONS
; i
++)
2502 fprintf (dump_file
, "%-50s: %8i calls, %8f freq, %" PRId64
" count\n",
2503 cgraph_inline_failed_string ((cgraph_inline_failed_t
) i
),
2504 (int) reason
[i
][1], reason_freq
[i
].to_double (), reason
[i
][0]);
2507 /* Called when node is removed. */
2510 flatten_remove_node_hook (struct cgraph_node
*node
, void *data
)
2512 if (lookup_attribute ("flatten", DECL_ATTRIBUTES (node
->decl
)) == NULL
)
2515 hash_set
<struct cgraph_node
*> *removed
2516 = (hash_set
<struct cgraph_node
*> *) data
;
2517 removed
->add (node
);
2520 /* Decide on the inlining. We do so in the topological order to avoid
2521 expenses on updating data structures. */
2526 struct cgraph_node
*node
;
2528 struct cgraph_node
**order
;
2531 bool remove_functions
= false;
2533 order
= XCNEWVEC (struct cgraph_node
*, symtab
->cgraph_count
);
2536 ipa_dump_fn_summaries (dump_file
);
2538 nnodes
= ipa_reverse_postorder (order
);
2539 spec_rem
= profile_count::zero ();
2541 FOR_EACH_FUNCTION (node
)
2545 /* Recompute the default reasons for inlining because they may have
2546 changed during merging. */
2549 for (cgraph_edge
*e
= node
->callees
; e
; e
= e
->next_callee
)
2551 gcc_assert (e
->inline_failed
);
2552 initialize_inline_failed (e
);
2554 for (cgraph_edge
*e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
2555 initialize_inline_failed (e
);
2560 fprintf (dump_file
, "\nFlattening functions:\n");
2562 /* First shrink order array, so that it only contains nodes with
2563 flatten attribute. */
2564 for (i
= nnodes
- 1, j
= i
; i
>= 0; i
--)
2567 if (node
->definition
2568 && lookup_attribute ("flatten",
2569 DECL_ATTRIBUTES (node
->decl
)) != NULL
)
2570 order
[j
--] = order
[i
];
2573 /* After the above loop, order[j + 1] ... order[nnodes - 1] contain
2574 nodes with flatten attribute. If there is more than one such
2575 node, we need to register a node removal hook, as flatten_function
2576 could remove other nodes with flatten attribute. See PR82801. */
2577 struct cgraph_node_hook_list
*node_removal_hook_holder
= NULL
;
2578 hash_set
<struct cgraph_node
*> *flatten_removed_nodes
= NULL
;
2581 flatten_removed_nodes
= new hash_set
<struct cgraph_node
*>;
2582 node_removal_hook_holder
2583 = symtab
->add_cgraph_removal_hook (&flatten_remove_node_hook
,
2584 flatten_removed_nodes
);
2587 /* In the first pass handle functions to be flattened. Do this with
2588 a priority so none of our later choices will make this impossible. */
2589 for (i
= nnodes
- 1; i
> j
; i
--)
2592 if (flatten_removed_nodes
2593 && flatten_removed_nodes
->contains (node
))
2596 /* Handle nodes to be flattened.
2597 Ideally when processing callees we stop inlining at the
2598 entry of cycles, possibly cloning that entry point and
2599 try to flatten itself turning it into a self-recursive
2602 fprintf (dump_file
, "Flattening %s\n", node
->name ());
2603 flatten_function (node
, false, true);
2608 symtab
->remove_cgraph_removal_hook (node_removal_hook_holder
);
2609 delete flatten_removed_nodes
;
2614 dump_overall_stats ();
2616 inline_small_functions ();
2618 gcc_assert (symtab
->state
== IPA_SSA
);
2619 symtab
->state
= IPA_SSA_AFTER_INLINING
;
2620 /* Do first after-inlining removal. We want to remove all "stale" extern
2621 inline functions and virtual functions so we really know what is called
2623 symtab
->remove_unreachable_nodes (dump_file
);
2625 /* Inline functions with a property that after inlining into all callers the
2626 code size will shrink because the out-of-line copy is eliminated.
2627 We do this regardless on the callee size as long as function growth limits
2631 "\nDeciding on functions to be inlined into all callers and "
2632 "removing useless speculations:\n");
2634 /* Inlining one function called once has good chance of preventing
2635 inlining other function into the same callee. Ideally we should
2636 work in priority order, but probably inlining hot functions first
2637 is good cut without the extra pain of maintaining the queue.
2639 ??? this is not really fitting the bill perfectly: inlining function
2640 into callee often leads to better optimization of callee due to
2641 increased context for optimization.
2642 For example if main() function calls a function that outputs help
2643 and then function that does the main optmization, we should inline
2644 the second with priority even if both calls are cold by themselves.
2646 We probably want to implement new predicate replacing our use of
2647 maybe_hot_edge interpreted as maybe_hot_edge || callee is known
2649 for (cold
= 0; cold
<= 1; cold
++)
2651 FOR_EACH_DEFINED_FUNCTION (node
)
2653 struct cgraph_edge
*edge
, *next
;
2656 if (!opt_for_fn (node
->decl
, optimize
)
2657 || !opt_for_fn (node
->decl
, flag_inline_functions_called_once
))
2660 for (edge
= node
->callees
; edge
; edge
= next
)
2662 next
= edge
->next_callee
;
2663 if (edge
->speculative
&& !speculation_useful_p (edge
, false))
2665 if (edge
->count
.ipa ().initialized_p ())
2666 spec_rem
+= edge
->count
.ipa ();
2667 edge
->resolve_speculation ();
2669 remove_functions
= true;
2674 struct cgraph_node
*where
= node
->inlined_to
2675 ? node
->inlined_to
: node
;
2676 reset_edge_caches (where
);
2677 ipa_update_overall_fn_summary (where
);
2679 if (want_inline_function_to_all_callers_p (node
, cold
))
2682 node
->call_for_symbol_and_aliases (sum_callers
, &num_calls
,
2684 while (node
->call_for_symbol_and_aliases
2685 (inline_to_all_callers
, &num_calls
, true))
2687 remove_functions
= true;
2692 /* Free ipa-prop structures if they are no longer needed. */
2693 ipa_free_all_structures_after_iinln ();
2695 if (dump_enabled_p ())
2696 dump_printf (MSG_NOTE
,
2697 "\nInlined %i calls, eliminated %i functions\n\n",
2698 ncalls_inlined
, nfunctions_inlined
);
2700 dump_inline_stats ();
2703 ipa_dump_fn_summaries (dump_file
);
2704 return remove_functions
? TODO_remove_functions
: 0;
2707 /* Inline always-inline function calls in NODE. */
2710 inline_always_inline_functions (struct cgraph_node
*node
)
2712 struct cgraph_edge
*e
;
2713 bool inlined
= false;
2715 for (e
= node
->callees
; e
; e
= e
->next_callee
)
2717 struct cgraph_node
*callee
= e
->callee
->ultimate_alias_target ();
2718 if (!DECL_DISREGARD_INLINE_LIMITS (callee
->decl
))
2721 if (e
->recursive_p ())
2723 if (dump_enabled_p ())
2724 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, e
->call_stmt
,
2725 " Not inlining recursive call to %C.\n",
2727 e
->inline_failed
= CIF_RECURSIVE_INLINING
;
2731 if (!can_early_inline_edge_p (e
))
2733 /* Set inlined to true if the callee is marked "always_inline" but
2734 is not inlinable. This will allow flagging an error later in
2735 expand_call_inline in tree-inline.c. */
2736 if (lookup_attribute ("always_inline",
2737 DECL_ATTRIBUTES (callee
->decl
)) != NULL
)
2742 if (dump_enabled_p ())
2743 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS
, e
->call_stmt
,
2744 " Inlining %C into %C (always_inline).\n",
2745 e
->callee
, e
->caller
);
2746 inline_call (e
, true, NULL
, NULL
, false);
2750 ipa_update_overall_fn_summary (node
);
2755 /* Decide on the inlining. We do so in the topological order to avoid
2756 expenses on updating data structures. */
2759 early_inline_small_functions (struct cgraph_node
*node
)
2761 struct cgraph_edge
*e
;
2762 bool inlined
= false;
2764 for (e
= node
->callees
; e
; e
= e
->next_callee
)
2766 struct cgraph_node
*callee
= e
->callee
->ultimate_alias_target ();
2768 /* We can enounter not-yet-analyzed function during
2769 early inlining on callgraphs with strongly
2770 connected components. */
2771 ipa_fn_summary
*s
= ipa_fn_summaries
->get (callee
);
2772 if (s
== NULL
|| !s
->inlinable
|| !e
->inline_failed
)
2775 /* Do not consider functions not declared inline. */
2776 if (!DECL_DECLARED_INLINE_P (callee
->decl
)
2777 && !opt_for_fn (node
->decl
, flag_inline_small_functions
)
2778 && !opt_for_fn (node
->decl
, flag_inline_functions
))
2781 if (dump_enabled_p ())
2782 dump_printf_loc (MSG_NOTE
, e
->call_stmt
,
2783 "Considering inline candidate %C.\n",
2786 if (!can_early_inline_edge_p (e
))
2789 if (e
->recursive_p ())
2791 if (dump_enabled_p ())
2792 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, e
->call_stmt
,
2793 " Not inlining: recursive call.\n");
2797 if (!want_early_inline_function_p (e
))
2800 if (dump_enabled_p ())
2801 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS
, e
->call_stmt
,
2802 " Inlining %C into %C.\n",
2804 inline_call (e
, true, NULL
, NULL
, false);
2809 ipa_update_overall_fn_summary (node
);
2815 early_inliner (function
*fun
)
2817 struct cgraph_node
*node
= cgraph_node::get (current_function_decl
);
2818 struct cgraph_edge
*edge
;
2819 unsigned int todo
= 0;
2821 bool inlined
= false;
2826 /* Do nothing if datastructures for ipa-inliner are already computed. This
2827 happens when some pass decides to construct new function and
2828 cgraph_add_new_function calls lowering passes and early optimization on
2829 it. This may confuse ourself when early inliner decide to inline call to
2830 function clone, because function clones don't have parameter list in
2831 ipa-prop matching their signature. */
2832 if (ipa_node_params_sum
)
2837 node
->remove_all_references ();
2839 /* Even when not optimizing or not inlining inline always-inline
2841 inlined
= inline_always_inline_functions (node
);
2845 || !flag_early_inlining
2846 /* Never inline regular functions into always-inline functions
2847 during incremental inlining. This sucks as functions calling
2848 always inline functions will get less optimized, but at the
2849 same time inlining of functions calling always inline
2850 function into an always inline function might introduce
2851 cycles of edges to be always inlined in the callgraph.
2853 We might want to be smarter and just avoid this type of inlining. */
2854 || (DECL_DISREGARD_INLINE_LIMITS (node
->decl
)
2855 && lookup_attribute ("always_inline",
2856 DECL_ATTRIBUTES (node
->decl
))))
2858 else if (lookup_attribute ("flatten",
2859 DECL_ATTRIBUTES (node
->decl
)) != NULL
)
2861 /* When the function is marked to be flattened, recursively inline
2863 if (dump_enabled_p ())
2864 dump_printf (MSG_OPTIMIZED_LOCATIONS
,
2865 "Flattening %C\n", node
);
2866 flatten_function (node
, true, true);
2871 /* If some always_inline functions was inlined, apply the changes.
2872 This way we will not account always inline into growth limits and
2873 moreover we will inline calls from always inlines that we skipped
2874 previously because of conditional above. */
2877 timevar_push (TV_INTEGRATION
);
2878 todo
|= optimize_inline_calls (current_function_decl
);
2879 /* optimize_inline_calls call above might have introduced new
2880 statements that don't have inline parameters computed. */
2881 for (edge
= node
->callees
; edge
; edge
= edge
->next_callee
)
2883 /* We can enounter not-yet-analyzed function during
2884 early inlining on callgraphs with strongly
2885 connected components. */
2886 ipa_call_summary
*es
= ipa_call_summaries
->get_create (edge
);
2888 = estimate_num_insns (edge
->call_stmt
, &eni_size_weights
);
2890 = estimate_num_insns (edge
->call_stmt
, &eni_time_weights
);
2892 ipa_update_overall_fn_summary (node
);
2894 timevar_pop (TV_INTEGRATION
);
2896 /* We iterate incremental inlining to get trivial cases of indirect
2898 while (iterations
< param_early_inliner_max_iterations
2899 && early_inline_small_functions (node
))
2901 timevar_push (TV_INTEGRATION
);
2902 todo
|= optimize_inline_calls (current_function_decl
);
2904 /* Technically we ought to recompute inline parameters so the new
2905 iteration of early inliner works as expected. We however have
2906 values approximately right and thus we only need to update edge
2907 info that might be cleared out for newly discovered edges. */
2908 for (edge
= node
->callees
; edge
; edge
= edge
->next_callee
)
2910 /* We have no summary for new bound store calls yet. */
2911 ipa_call_summary
*es
= ipa_call_summaries
->get_create (edge
);
2913 = estimate_num_insns (edge
->call_stmt
, &eni_size_weights
);
2915 = estimate_num_insns (edge
->call_stmt
, &eni_time_weights
);
2917 if (iterations
< param_early_inliner_max_iterations
- 1)
2918 ipa_update_overall_fn_summary (node
);
2919 timevar_pop (TV_INTEGRATION
);
2924 fprintf (dump_file
, "Iterations: %i\n", iterations
);
2929 timevar_push (TV_INTEGRATION
);
2930 todo
|= optimize_inline_calls (current_function_decl
);
2931 timevar_pop (TV_INTEGRATION
);
2934 fun
->always_inline_functions_inlined
= true;
2939 /* Do inlining of small functions. Doing so early helps profiling and other
2940 passes to be somewhat more effective and avoids some code duplication in
2941 later real inlining pass for testcases with very many function calls. */
2945 const pass_data pass_data_early_inline
=
2947 GIMPLE_PASS
, /* type */
2948 "einline", /* name */
2949 OPTGROUP_INLINE
, /* optinfo_flags */
2950 TV_EARLY_INLINING
, /* tv_id */
2951 PROP_ssa
, /* properties_required */
2952 0, /* properties_provided */
2953 0, /* properties_destroyed */
2954 0, /* todo_flags_start */
2955 0, /* todo_flags_finish */
2958 class pass_early_inline
: public gimple_opt_pass
2961 pass_early_inline (gcc::context
*ctxt
)
2962 : gimple_opt_pass (pass_data_early_inline
, ctxt
)
2965 /* opt_pass methods: */
2966 virtual unsigned int execute (function
*);
2968 }; // class pass_early_inline
2971 pass_early_inline::execute (function
*fun
)
2973 return early_inliner (fun
);
2979 make_pass_early_inline (gcc::context
*ctxt
)
2981 return new pass_early_inline (ctxt
);
2986 const pass_data pass_data_ipa_inline
=
2988 IPA_PASS
, /* type */
2989 "inline", /* name */
2990 OPTGROUP_INLINE
, /* optinfo_flags */
2991 TV_IPA_INLINING
, /* tv_id */
2992 0, /* properties_required */
2993 0, /* properties_provided */
2994 0, /* properties_destroyed */
2995 0, /* todo_flags_start */
2996 ( TODO_dump_symtab
), /* todo_flags_finish */
2999 class pass_ipa_inline
: public ipa_opt_pass_d
3002 pass_ipa_inline (gcc::context
*ctxt
)
3003 : ipa_opt_pass_d (pass_data_ipa_inline
, ctxt
,
3004 NULL
, /* generate_summary */
3005 NULL
, /* write_summary */
3006 NULL
, /* read_summary */
3007 NULL
, /* write_optimization_summary */
3008 NULL
, /* read_optimization_summary */
3009 NULL
, /* stmt_fixup */
3010 0, /* function_transform_todo_flags_start */
3011 inline_transform
, /* function_transform */
3012 NULL
) /* variable_transform */
3015 /* opt_pass methods: */
3016 virtual unsigned int execute (function
*) { return ipa_inline (); }
3018 }; // class pass_ipa_inline
3023 make_pass_ipa_inline (gcc::context
*ctxt
)
3025 return new pass_ipa_inline (ctxt
);