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"
110 #include "symbol-summary.h"
111 #include "tree-vrp.h"
112 #include "ipa-prop.h"
113 #include "ipa-fnsummary.h"
114 #include "ipa-inline.h"
115 #include "ipa-utils.h"
117 #include "auto-profile.h"
118 #include "builtins.h"
119 #include "fibonacci_heap.h"
120 #include "stringpool.h"
124 typedef fibonacci_heap
<sreal
, cgraph_edge
> edge_heap_t
;
125 typedef fibonacci_node
<sreal
, cgraph_edge
> edge_heap_node_t
;
127 /* Statistics we collect about inlining algorithm. */
128 static int overall_size
;
129 static profile_count max_count
;
130 static profile_count spec_rem
;
132 /* Return false when inlining edge E would lead to violating
133 limits on function unit growth or stack usage growth.
135 The relative function body growth limit is present generally
136 to avoid problems with non-linear behavior of the compiler.
137 To allow inlining huge functions into tiny wrapper, the limit
138 is always based on the bigger of the two functions considered.
140 For stack growth limits we always base the growth in stack usage
141 of the callers. We want to prevent applications from segfaulting
142 on stack overflow when functions with huge stack frames gets
146 caller_growth_limits (struct cgraph_edge
*e
)
148 struct cgraph_node
*to
= e
->caller
;
149 struct cgraph_node
*what
= e
->callee
->ultimate_alias_target ();
152 HOST_WIDE_INT stack_size_limit
= 0, inlined_stack
;
153 ipa_size_summary
*outer_info
= ipa_size_summaries
->get (to
);
155 /* Look for function e->caller is inlined to. While doing
156 so work out the largest function body on the way. As
157 described above, we want to base our function growth
158 limits based on that. Not on the self size of the
159 outer function, not on the self size of inline code
160 we immediately inline to. This is the most relaxed
161 interpretation of the rule "do not grow large functions
162 too much in order to prevent compiler from exploding". */
165 ipa_size_summary
*size_info
= ipa_size_summaries
->get (to
);
166 if (limit
< size_info
->self_size
)
167 limit
= size_info
->self_size
;
168 if (stack_size_limit
< size_info
->estimated_self_stack_size
)
169 stack_size_limit
= size_info
->estimated_self_stack_size
;
171 to
= to
->callers
->caller
;
176 ipa_fn_summary
*what_info
= ipa_fn_summaries
->get (what
);
177 ipa_size_summary
*what_size_info
= ipa_size_summaries
->get (what
);
179 if (limit
< what_size_info
->self_size
)
180 limit
= what_size_info
->self_size
;
182 limit
+= limit
* param_large_function_growth
/ 100;
184 /* Check the size after inlining against the function limits. But allow
185 the function to shrink if it went over the limits by forced inlining. */
186 newsize
= estimate_size_after_inlining (to
, e
);
187 if (newsize
>= ipa_size_summaries
->get (what
)->size
188 && newsize
> param_large_function_insns
191 e
->inline_failed
= CIF_LARGE_FUNCTION_GROWTH_LIMIT
;
195 if (!what_info
->estimated_stack_size
)
198 /* FIXME: Stack size limit often prevents inlining in Fortran programs
199 due to large i/o datastructures used by the Fortran front-end.
200 We ought to ignore this limit when we know that the edge is executed
201 on every invocation of the caller (i.e. its call statement dominates
202 exit block). We do not track this information, yet. */
203 stack_size_limit
+= ((gcov_type
)stack_size_limit
204 * param_stack_frame_growth
/ 100);
206 inlined_stack
= (ipa_get_stack_frame_offset (to
)
207 + outer_info
->estimated_self_stack_size
208 + what_info
->estimated_stack_size
);
209 /* Check new stack consumption with stack consumption at the place
211 if (inlined_stack
> stack_size_limit
212 /* If function already has large stack usage from sibling
213 inline call, we can inline, too.
214 This bit overoptimistically assume that we are good at stack
216 && inlined_stack
> ipa_fn_summaries
->get (to
)->estimated_stack_size
217 && inlined_stack
> param_large_stack_frame
)
219 e
->inline_failed
= CIF_LARGE_STACK_FRAME_GROWTH_LIMIT
;
225 /* Dump info about why inlining has failed. */
228 report_inline_failed_reason (struct cgraph_edge
*e
)
230 if (dump_enabled_p ())
232 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, e
->call_stmt
,
233 " not inlinable: %C -> %C, %s\n",
234 e
->caller
, e
->callee
,
235 cgraph_inline_failed_string (e
->inline_failed
));
236 if ((e
->inline_failed
== CIF_TARGET_OPTION_MISMATCH
237 || e
->inline_failed
== CIF_OPTIMIZATION_MISMATCH
)
238 && e
->caller
->lto_file_data
239 && e
->callee
->ultimate_alias_target ()->lto_file_data
)
241 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, e
->call_stmt
,
242 " LTO objects: %s, %s\n",
243 e
->caller
->lto_file_data
->file_name
,
244 e
->callee
->ultimate_alias_target ()->lto_file_data
->file_name
);
246 if (e
->inline_failed
== CIF_TARGET_OPTION_MISMATCH
)
248 cl_target_option_print_diff
249 (dump_file
, 2, target_opts_for_fn (e
->caller
->decl
),
250 target_opts_for_fn (e
->callee
->ultimate_alias_target ()->decl
));
251 if (e
->inline_failed
== CIF_OPTIMIZATION_MISMATCH
)
253 cl_optimization_print_diff
254 (dump_file
, 2, opts_for_fn (e
->caller
->decl
),
255 opts_for_fn (e
->callee
->ultimate_alias_target ()->decl
));
259 /* Decide whether sanitizer-related attributes allow inlining. */
262 sanitize_attrs_match_for_inline_p (const_tree caller
, const_tree callee
)
264 if (!caller
|| !callee
)
267 /* Allow inlining always_inline functions into no_sanitize_address
269 if (!sanitize_flags_p (SANITIZE_ADDRESS
, caller
)
270 && lookup_attribute ("always_inline", DECL_ATTRIBUTES (callee
)))
273 return ((sanitize_flags_p (SANITIZE_ADDRESS
, caller
)
274 == sanitize_flags_p (SANITIZE_ADDRESS
, callee
))
275 && (sanitize_flags_p (SANITIZE_POINTER_COMPARE
, caller
)
276 == sanitize_flags_p (SANITIZE_POINTER_COMPARE
, callee
))
277 && (sanitize_flags_p (SANITIZE_POINTER_SUBTRACT
, caller
)
278 == sanitize_flags_p (SANITIZE_POINTER_SUBTRACT
, callee
)));
281 /* Used for flags where it is safe to inline when caller's value is
282 grater than callee's. */
283 #define check_maybe_up(flag) \
284 (opts_for_fn (caller->decl)->x_##flag \
285 != opts_for_fn (callee->decl)->x_##flag \
287 || opts_for_fn (caller->decl)->x_##flag \
288 < opts_for_fn (callee->decl)->x_##flag))
289 /* Used for flags where it is safe to inline when caller's value is
290 smaller than callee's. */
291 #define check_maybe_down(flag) \
292 (opts_for_fn (caller->decl)->x_##flag \
293 != opts_for_fn (callee->decl)->x_##flag \
295 || opts_for_fn (caller->decl)->x_##flag \
296 > opts_for_fn (callee->decl)->x_##flag))
297 /* Used for flags where exact match is needed for correctness. */
298 #define check_match(flag) \
299 (opts_for_fn (caller->decl)->x_##flag \
300 != opts_for_fn (callee->decl)->x_##flag)
302 /* Decide if we can inline the edge and possibly update
303 inline_failed reason.
304 We check whether inlining is possible at all and whether
305 caller growth limits allow doing so.
307 if REPORT is true, output reason to the dump file. */
310 can_inline_edge_p (struct cgraph_edge
*e
, bool report
,
313 gcc_checking_assert (e
->inline_failed
);
315 if (cgraph_inline_failed_type (e
->inline_failed
) == CIF_FINAL_ERROR
)
318 report_inline_failed_reason (e
);
322 bool inlinable
= true;
323 enum availability avail
;
324 cgraph_node
*caller
= (e
->caller
->inlined_to
325 ? e
->caller
->inlined_to
: e
->caller
);
326 cgraph_node
*callee
= e
->callee
->ultimate_alias_target (&avail
, caller
);
328 if (!callee
->definition
)
330 e
->inline_failed
= CIF_BODY_NOT_AVAILABLE
;
333 if (!early
&& (!opt_for_fn (callee
->decl
, optimize
)
334 || !opt_for_fn (caller
->decl
, optimize
)))
336 e
->inline_failed
= CIF_FUNCTION_NOT_OPTIMIZED
;
339 else if (callee
->calls_comdat_local
)
341 e
->inline_failed
= CIF_USES_COMDAT_LOCAL
;
344 else if (avail
<= AVAIL_INTERPOSABLE
)
346 e
->inline_failed
= CIF_OVERWRITABLE
;
349 /* All edges with call_stmt_cannot_inline_p should have inline_failed
350 initialized to one of FINAL_ERROR reasons. */
351 else if (e
->call_stmt_cannot_inline_p
)
353 /* Don't inline if the functions have different EH personalities. */
354 else if (DECL_FUNCTION_PERSONALITY (caller
->decl
)
355 && DECL_FUNCTION_PERSONALITY (callee
->decl
)
356 && (DECL_FUNCTION_PERSONALITY (caller
->decl
)
357 != DECL_FUNCTION_PERSONALITY (callee
->decl
)))
359 e
->inline_failed
= CIF_EH_PERSONALITY
;
362 /* TM pure functions should not be inlined into non-TM_pure
364 else if (is_tm_pure (callee
->decl
) && !is_tm_pure (caller
->decl
))
366 e
->inline_failed
= CIF_UNSPECIFIED
;
369 /* Check compatibility of target optimization options. */
370 else if (!targetm
.target_option
.can_inline_p (caller
->decl
,
373 e
->inline_failed
= CIF_TARGET_OPTION_MISMATCH
;
376 else if (ipa_fn_summaries
->get (callee
) == NULL
377 || !ipa_fn_summaries
->get (callee
)->inlinable
)
379 e
->inline_failed
= CIF_FUNCTION_NOT_INLINABLE
;
382 /* Don't inline a function with mismatched sanitization attributes. */
383 else if (!sanitize_attrs_match_for_inline_p (caller
->decl
, callee
->decl
))
385 e
->inline_failed
= CIF_ATTRIBUTE_MISMATCH
;
388 if (!inlinable
&& report
)
389 report_inline_failed_reason (e
);
393 /* Return inlining_insns_single limit for function N. If HINT is true
394 scale up the bound. */
397 inline_insns_single (cgraph_node
*n
, bool hint
)
399 if (opt_for_fn (n
->decl
, optimize
) >= 3)
402 return param_max_inline_insns_single
403 * param_inline_heuristics_hint_percent
/ 100;
404 return param_max_inline_insns_single
;
409 return param_max_inline_insns_single_o2
410 * param_inline_heuristics_hint_percent_o2
/ 100;
411 return param_max_inline_insns_single_o2
;
415 /* Return inlining_insns_auto limit for function N. If HINT is true
416 scale up the bound. */
419 inline_insns_auto (cgraph_node
*n
, bool hint
)
421 if (opt_for_fn (n
->decl
, optimize
) >= 3)
424 return param_max_inline_insns_auto
425 * param_inline_heuristics_hint_percent
/ 100;
426 return param_max_inline_insns_auto
;
431 return param_max_inline_insns_auto_o2
432 * param_inline_heuristics_hint_percent_o2
/ 100;
433 return param_max_inline_insns_auto_o2
;
437 /* Decide if we can inline the edge and possibly update
438 inline_failed reason.
439 We check whether inlining is possible at all and whether
440 caller growth limits allow doing so.
442 if REPORT is true, output reason to the dump file.
444 if DISREGARD_LIMITS is true, ignore size limits. */
447 can_inline_edge_by_limits_p (struct cgraph_edge
*e
, bool report
,
448 bool disregard_limits
= false, bool early
= false)
450 gcc_checking_assert (e
->inline_failed
);
452 if (cgraph_inline_failed_type (e
->inline_failed
) == CIF_FINAL_ERROR
)
455 report_inline_failed_reason (e
);
459 bool inlinable
= true;
460 enum availability avail
;
461 cgraph_node
*caller
= (e
->caller
->inlined_to
462 ? e
->caller
->inlined_to
: e
->caller
);
463 cgraph_node
*callee
= e
->callee
->ultimate_alias_target (&avail
, caller
);
464 tree caller_tree
= DECL_FUNCTION_SPECIFIC_OPTIMIZATION (caller
->decl
);
466 = callee
? DECL_FUNCTION_SPECIFIC_OPTIMIZATION (callee
->decl
) : NULL
;
467 /* Check if caller growth allows the inlining. */
468 if (!DECL_DISREGARD_INLINE_LIMITS (callee
->decl
)
470 && !lookup_attribute ("flatten",
471 DECL_ATTRIBUTES (caller
->decl
))
472 && !caller_growth_limits (e
))
474 else if (callee
->externally_visible
475 && !DECL_DISREGARD_INLINE_LIMITS (callee
->decl
)
476 && flag_live_patching
== LIVE_PATCHING_INLINE_ONLY_STATIC
)
478 e
->inline_failed
= CIF_EXTERN_LIVE_ONLY_STATIC
;
481 /* Don't inline a function with a higher optimization level than the
482 caller. FIXME: this is really just tip of iceberg of handling
483 optimization attribute. */
484 else if (caller_tree
!= callee_tree
)
487 (DECL_DISREGARD_INLINE_LIMITS (callee
->decl
)
488 && lookup_attribute ("always_inline",
489 DECL_ATTRIBUTES (callee
->decl
)));
490 ipa_fn_summary
*caller_info
= ipa_fn_summaries
->get (caller
);
491 ipa_fn_summary
*callee_info
= ipa_fn_summaries
->get (callee
);
493 /* Until GCC 4.9 we did not check the semantics-altering flags
494 below and inlined across optimization boundaries.
495 Enabling checks below breaks several packages by refusing
496 to inline library always_inline functions. See PR65873.
497 Disable the check for early inlining for now until better solution
499 if (always_inline
&& early
)
501 /* There are some options that change IL semantics which means
502 we cannot inline in these cases for correctness reason.
503 Not even for always_inline declared functions. */
504 else if (check_match (flag_wrapv
)
505 || check_match (flag_trapv
)
506 || check_match (flag_pcc_struct_return
)
507 /* When caller or callee does FP math, be sure FP codegen flags
509 || ((caller_info
->fp_expressions
&& callee_info
->fp_expressions
)
510 && (check_maybe_up (flag_rounding_math
)
511 || check_maybe_up (flag_trapping_math
)
512 || check_maybe_down (flag_unsafe_math_optimizations
)
513 || check_maybe_down (flag_finite_math_only
)
514 || check_maybe_up (flag_signaling_nans
)
515 || check_maybe_down (flag_cx_limited_range
)
516 || check_maybe_up (flag_signed_zeros
)
517 || check_maybe_down (flag_associative_math
)
518 || check_maybe_down (flag_reciprocal_math
)
519 || check_maybe_down (flag_fp_int_builtin_inexact
)
520 /* Strictly speaking only when the callee contains function
521 calls that may end up setting errno. */
522 || check_maybe_up (flag_errno_math
)))
523 /* We do not want to make code compiled with exceptions to be
524 brought into a non-EH function unless we know that the callee
526 This is tracked by DECL_FUNCTION_PERSONALITY. */
527 || (check_maybe_up (flag_non_call_exceptions
)
528 && DECL_FUNCTION_PERSONALITY (callee
->decl
))
529 || (check_maybe_up (flag_exceptions
)
530 && DECL_FUNCTION_PERSONALITY (callee
->decl
))
531 /* When devirtualization is diabled for callee, it is not safe
532 to inline it as we possibly mangled the type info.
533 Allow early inlining of always inlines. */
534 || (!early
&& check_maybe_down (flag_devirtualize
)))
536 e
->inline_failed
= CIF_OPTIMIZATION_MISMATCH
;
539 /* gcc.dg/pr43564.c. Apply user-forced inline even at -O0. */
540 else if (always_inline
)
542 /* When user added an attribute to the callee honor it. */
543 else if (lookup_attribute ("optimize", DECL_ATTRIBUTES (callee
->decl
))
544 && opts_for_fn (caller
->decl
) != opts_for_fn (callee
->decl
))
546 e
->inline_failed
= CIF_OPTIMIZATION_MISMATCH
;
549 /* If explicit optimize attribute are not used, the mismatch is caused
550 by different command line options used to build different units.
551 Do not care about COMDAT functions - those are intended to be
552 optimized with the optimization flags of module they are used in.
553 Also do not care about mixing up size/speed optimization when
554 DECL_DISREGARD_INLINE_LIMITS is set. */
555 else if ((callee
->merged_comdat
556 && !lookup_attribute ("optimize",
557 DECL_ATTRIBUTES (caller
->decl
)))
558 || DECL_DISREGARD_INLINE_LIMITS (callee
->decl
))
560 /* If mismatch is caused by merging two LTO units with different
561 optimizationflags we want to be bit nicer. However never inline
562 if one of functions is not optimized at all. */
563 else if (!opt_for_fn (callee
->decl
, optimize
)
564 || !opt_for_fn (caller
->decl
, optimize
))
566 e
->inline_failed
= CIF_OPTIMIZATION_MISMATCH
;
569 /* If callee is optimized for size and caller is not, allow inlining if
570 code shrinks or we are in param_max_inline_insns_single limit and
571 callee is inline (and thus likely an unified comdat).
572 This will allow caller to run faster. */
573 else if (opt_for_fn (callee
->decl
, optimize_size
)
574 > opt_for_fn (caller
->decl
, optimize_size
))
576 int growth
= estimate_edge_growth (e
);
577 if (growth
> param_max_inline_insns_size
578 && (!DECL_DECLARED_INLINE_P (callee
->decl
)
579 && growth
>= MAX (inline_insns_single (caller
, false),
580 inline_insns_auto (caller
, false))))
582 e
->inline_failed
= CIF_OPTIMIZATION_MISMATCH
;
586 /* If callee is more aggressively optimized for performance than caller,
587 we generally want to inline only cheap (runtime wise) functions. */
588 else if (opt_for_fn (callee
->decl
, optimize_size
)
589 < opt_for_fn (caller
->decl
, optimize_size
)
590 || (opt_for_fn (callee
->decl
, optimize
)
591 > opt_for_fn (caller
->decl
, optimize
)))
593 if (estimate_edge_time (e
)
594 >= 20 + ipa_call_summaries
->get (e
)->call_stmt_time
)
596 e
->inline_failed
= CIF_OPTIMIZATION_MISMATCH
;
603 if (!inlinable
&& report
)
604 report_inline_failed_reason (e
);
609 /* Return true if the edge E is inlinable during early inlining. */
612 can_early_inline_edge_p (struct cgraph_edge
*e
)
614 struct cgraph_node
*callee
= e
->callee
->ultimate_alias_target ();
615 /* Early inliner might get called at WPA stage when IPA pass adds new
616 function. In this case we cannot really do any of early inlining
617 because function bodies are missing. */
618 if (cgraph_inline_failed_type (e
->inline_failed
) == CIF_FINAL_ERROR
)
620 if (!gimple_has_body_p (callee
->decl
))
622 e
->inline_failed
= CIF_BODY_NOT_AVAILABLE
;
625 /* In early inliner some of callees may not be in SSA form yet
626 (i.e. the callgraph is cyclic and we did not process
627 the callee by early inliner, yet). We don't have CIF code for this
628 case; later we will re-do the decision in the real inliner. */
629 if (!gimple_in_ssa_p (DECL_STRUCT_FUNCTION (e
->caller
->decl
))
630 || !gimple_in_ssa_p (DECL_STRUCT_FUNCTION (callee
->decl
)))
632 if (dump_enabled_p ())
633 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, e
->call_stmt
,
634 " edge not inlinable: not in SSA form\n");
637 if (!can_inline_edge_p (e
, true, true)
638 || !can_inline_edge_by_limits_p (e
, true, false, true))
644 /* Return number of calls in N. Ignore cheap builtins. */
647 num_calls (struct cgraph_node
*n
)
649 struct cgraph_edge
*e
;
652 for (e
= n
->callees
; e
; e
= e
->next_callee
)
653 if (!is_inexpensive_builtin (e
->callee
->decl
))
659 /* Return true if we are interested in inlining small function. */
662 want_early_inline_function_p (struct cgraph_edge
*e
)
664 bool want_inline
= true;
665 struct cgraph_node
*callee
= e
->callee
->ultimate_alias_target ();
667 if (DECL_DISREGARD_INLINE_LIMITS (callee
->decl
))
669 /* For AutoFDO, we need to make sure that before profile summary, all
670 hot paths' IR look exactly the same as profiled binary. As a result,
671 in einliner, we will disregard size limit and inline those callsites
673 * inlined in the profiled binary, and
674 * the cloned callee has enough samples to be considered "hot". */
675 else if (flag_auto_profile
&& afdo_callsite_hot_enough_for_early_inline (e
))
677 else if (!DECL_DECLARED_INLINE_P (callee
->decl
)
678 && !opt_for_fn (e
->caller
->decl
, flag_inline_small_functions
))
680 e
->inline_failed
= CIF_FUNCTION_NOT_INLINE_CANDIDATE
;
681 report_inline_failed_reason (e
);
686 int growth
= estimate_edge_growth (e
);
688 int early_inlining_insns
= opt_for_fn (e
->caller
->decl
, optimize
) >= 3
689 ? param_early_inlining_insns
690 : param_early_inlining_insns_o2
;
693 if (growth
<= param_max_inline_insns_size
)
695 else if (!e
->maybe_hot_p ())
697 if (dump_enabled_p ())
698 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, e
->call_stmt
,
699 " will not early inline: %C->%C, "
700 "call is cold and code would grow by %i\n",
705 else if (growth
> early_inlining_insns
)
707 if (dump_enabled_p ())
708 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, e
->call_stmt
,
709 " will not early inline: %C->%C, "
710 "growth %i exceeds --param early-inlining-insns%s\n",
711 e
->caller
, callee
, growth
,
712 opt_for_fn (e
->caller
->decl
, optimize
) >= 3
716 else if ((n
= num_calls (callee
)) != 0
717 && growth
* (n
+ 1) > early_inlining_insns
)
719 if (dump_enabled_p ())
720 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, e
->call_stmt
,
721 " will not early inline: %C->%C, "
722 "growth %i exceeds --param early-inlining-insns%s "
723 "divided by number of calls\n",
724 e
->caller
, callee
, growth
,
725 opt_for_fn (e
->caller
->decl
, optimize
) >= 3
733 /* Compute time of the edge->caller + edge->callee execution when inlining
737 compute_uninlined_call_time (struct cgraph_edge
*edge
,
738 sreal uninlined_call_time
,
741 cgraph_node
*caller
= (edge
->caller
->inlined_to
742 ? edge
->caller
->inlined_to
746 uninlined_call_time
*= freq
;
748 uninlined_call_time
= uninlined_call_time
>> 11;
750 sreal caller_time
= ipa_fn_summaries
->get (caller
)->time
;
751 return uninlined_call_time
+ caller_time
;
754 /* Same as compute_uinlined_call_time but compute time when inlining
758 compute_inlined_call_time (struct cgraph_edge
*edge
,
762 cgraph_node
*caller
= (edge
->caller
->inlined_to
763 ? edge
->caller
->inlined_to
765 sreal caller_time
= ipa_fn_summaries
->get (caller
)->time
;
772 /* This calculation should match one in ipa-inline-analysis.c
773 (estimate_edge_size_and_time). */
774 time
-= (sreal
)ipa_call_summaries
->get (edge
)->call_stmt_time
* freq
;
777 time
= ((sreal
) 1) >> 8;
778 gcc_checking_assert (time
>= 0);
782 /* Return true if the speedup for inlining E is bigger than
783 PARAM_MAX_INLINE_MIN_SPEEDUP. */
786 big_speedup_p (struct cgraph_edge
*e
)
789 sreal spec_time
= estimate_edge_time (e
, &unspec_time
);
790 sreal freq
= e
->sreal_frequency ();
791 sreal time
= compute_uninlined_call_time (e
, unspec_time
, freq
);
792 sreal inlined_time
= compute_inlined_call_time (e
, spec_time
, freq
);
793 cgraph_node
*caller
= (e
->caller
->inlined_to
794 ? e
->caller
->inlined_to
796 int limit
= opt_for_fn (caller
->decl
, optimize
) >= 3
797 ? param_inline_min_speedup
798 : param_inline_min_speedup_o2
;
800 if ((time
- inlined_time
) * 100 > time
* limit
)
805 /* Return true if we are interested in inlining small function.
806 When REPORT is true, report reason to dump file. */
809 want_inline_small_function_p (struct cgraph_edge
*e
, bool report
)
811 bool want_inline
= true;
812 struct cgraph_node
*callee
= e
->callee
->ultimate_alias_target ();
814 /* Allow this function to be called before can_inline_edge_p,
815 since it's usually cheaper. */
816 if (cgraph_inline_failed_type (e
->inline_failed
) == CIF_FINAL_ERROR
)
818 else if (DECL_DISREGARD_INLINE_LIMITS (callee
->decl
))
820 else if (!DECL_DECLARED_INLINE_P (callee
->decl
)
821 && !opt_for_fn (e
->caller
->decl
, flag_inline_small_functions
))
823 e
->inline_failed
= CIF_FUNCTION_NOT_INLINE_CANDIDATE
;
826 /* Do fast and conservative check if the function can be good
828 else if ((!DECL_DECLARED_INLINE_P (callee
->decl
)
829 && (!e
->count
.ipa ().initialized_p () || !e
->maybe_hot_p ()))
830 && ipa_fn_summaries
->get (callee
)->min_size
831 - ipa_call_summaries
->get (e
)->call_stmt_size
832 > inline_insns_auto (e
->caller
, true))
834 if (opt_for_fn (e
->caller
->decl
, optimize
) >= 3)
835 e
->inline_failed
= CIF_MAX_INLINE_INSNS_AUTO_LIMIT
;
837 e
->inline_failed
= CIF_MAX_INLINE_INSNS_AUTO_O2_LIMIT
;
840 else if ((DECL_DECLARED_INLINE_P (callee
->decl
)
841 || e
->count
.ipa ().nonzero_p ())
842 && ipa_fn_summaries
->get (callee
)->min_size
843 - ipa_call_summaries
->get (e
)->call_stmt_size
844 > inline_insns_single (e
->caller
, true))
846 if (opt_for_fn (e
->caller
->decl
, optimize
) >= 3)
847 e
->inline_failed
= (DECL_DECLARED_INLINE_P (callee
->decl
)
848 ? CIF_MAX_INLINE_INSNS_SINGLE_LIMIT
849 : CIF_MAX_INLINE_INSNS_AUTO_LIMIT
);
851 e
->inline_failed
= (DECL_DECLARED_INLINE_P (callee
->decl
)
852 ? CIF_MAX_INLINE_INSNS_SINGLE_O2_LIMIT
853 : CIF_MAX_INLINE_INSNS_AUTO_O2_LIMIT
);
858 int growth
= estimate_edge_growth (e
);
859 ipa_hints hints
= estimate_edge_hints (e
);
860 bool apply_hints
= (hints
& (INLINE_HINT_indirect_call
861 | INLINE_HINT_known_hot
862 | INLINE_HINT_loop_iterations
863 | INLINE_HINT_loop_stride
));
865 if (growth
<= param_max_inline_insns_size
)
867 /* Apply param_max_inline_insns_single limit. Do not do so when
868 hints suggests that inlining given function is very profitable.
869 Avoid computation of big_speedup_p when not necessary to change
870 outcome of decision. */
871 else if (DECL_DECLARED_INLINE_P (callee
->decl
)
872 && growth
>= inline_insns_single (e
->caller
, apply_hints
)
874 || growth
>= inline_insns_single (e
->caller
, true)
875 || !big_speedup_p (e
)))
877 if (opt_for_fn (e
->caller
->decl
, optimize
) >= 3)
878 e
->inline_failed
= CIF_MAX_INLINE_INSNS_SINGLE_LIMIT
;
880 e
->inline_failed
= CIF_MAX_INLINE_INSNS_SINGLE_O2_LIMIT
;
883 else if (!DECL_DECLARED_INLINE_P (callee
->decl
)
884 && !opt_for_fn (e
->caller
->decl
, flag_inline_functions
)
885 && growth
>= param_max_inline_insns_small
)
887 /* growth_positive_p is expensive, always test it last. */
888 if (growth
>= inline_insns_single (e
->caller
, false)
889 || growth_positive_p (callee
, e
, growth
))
891 e
->inline_failed
= CIF_NOT_DECLARED_INLINED
;
895 /* Apply param_max_inline_insns_auto limit for functions not declared
896 inline. Bypass the limit when speedup seems big. */
897 else if (!DECL_DECLARED_INLINE_P (callee
->decl
)
898 && growth
>= inline_insns_auto (e
->caller
, apply_hints
)
900 || growth
>= inline_insns_auto (e
->caller
, true)
901 || !big_speedup_p (e
)))
903 /* growth_positive_p is expensive, always test it last. */
904 if (growth
>= inline_insns_single (e
->caller
, false)
905 || growth_positive_p (callee
, e
, growth
))
907 if (opt_for_fn (e
->caller
->decl
, optimize
) >= 3)
908 e
->inline_failed
= CIF_MAX_INLINE_INSNS_AUTO_LIMIT
;
910 e
->inline_failed
= CIF_MAX_INLINE_INSNS_AUTO_O2_LIMIT
;
914 /* If call is cold, do not inline when function body would grow. */
915 else if (!e
->maybe_hot_p ()
916 && (growth
>= inline_insns_single (e
->caller
, false)
917 || growth_positive_p (callee
, e
, growth
)))
919 e
->inline_failed
= CIF_UNLIKELY_CALL
;
923 if (!want_inline
&& report
)
924 report_inline_failed_reason (e
);
928 /* EDGE is self recursive edge.
929 We hand two cases - when function A is inlining into itself
930 or when function A is being inlined into another inliner copy of function
933 In first case OUTER_NODE points to the toplevel copy of A, while
934 in the second case OUTER_NODE points to the outermost copy of A in B.
936 In both cases we want to be extra selective since
937 inlining the call will just introduce new recursive calls to appear. */
940 want_inline_self_recursive_call_p (struct cgraph_edge
*edge
,
941 struct cgraph_node
*outer_node
,
945 char const *reason
= NULL
;
946 bool want_inline
= true;
947 sreal caller_freq
= 1;
948 int max_depth
= param_max_inline_recursive_depth_auto
;
950 if (DECL_DECLARED_INLINE_P (edge
->caller
->decl
))
951 max_depth
= param_max_inline_recursive_depth
;
953 if (!edge
->maybe_hot_p ())
955 reason
= "recursive call is cold";
958 else if (depth
> max_depth
)
960 reason
= "--param max-inline-recursive-depth exceeded.";
963 else if (outer_node
->inlined_to
964 && (caller_freq
= outer_node
->callers
->sreal_frequency ()) == 0)
966 reason
= "caller frequency is 0";
972 /* Inlining of self recursive function into copy of itself within other
973 function is transformation similar to loop peeling.
975 Peeling is profitable if we can inline enough copies to make probability
976 of actual call to the self recursive function very small. Be sure that
977 the probability of recursion is small.
979 We ensure that the frequency of recursing is at most 1 - (1/max_depth).
980 This way the expected number of recursion is at most max_depth. */
983 sreal max_prob
= (sreal
)1 - ((sreal
)1 / (sreal
)max_depth
);
985 for (i
= 1; i
< depth
; i
++)
986 max_prob
= max_prob
* max_prob
;
987 if (edge
->sreal_frequency () >= max_prob
* caller_freq
)
989 reason
= "frequency of recursive call is too large";
993 /* Recursive inlining, i.e. equivalent of unrolling, is profitable if
994 recursion depth is large. We reduce function call overhead and increase
995 chances that things fit in hardware return predictor.
997 Recursive inlining might however increase cost of stack frame setup
998 actually slowing down functions whose recursion tree is wide rather than
1001 Deciding reliably on when to do recursive inlining without profile feedback
1002 is tricky. For now we disable recursive inlining when probability of self
1005 Recursive inlining of self recursive call within loop also results in
1006 large loop depths that generally optimize badly. We may want to throttle
1007 down inlining in those cases. In particular this seems to happen in one
1008 of libstdc++ rb tree methods. */
1011 if (edge
->sreal_frequency () * 100
1013 * param_min_inline_recursive_probability
)
1015 reason
= "frequency of recursive call is too small";
1016 want_inline
= false;
1019 if (!want_inline
&& dump_enabled_p ())
1020 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, edge
->call_stmt
,
1021 " not inlining recursively: %s\n", reason
);
1025 /* Return true when NODE has uninlinable caller;
1026 set HAS_HOT_CALL if it has hot call.
1027 Worker for cgraph_for_node_and_aliases. */
1030 check_callers (struct cgraph_node
*node
, void *has_hot_call
)
1032 struct cgraph_edge
*e
;
1033 for (e
= node
->callers
; e
; e
= e
->next_caller
)
1035 if (!opt_for_fn (e
->caller
->decl
, flag_inline_functions_called_once
)
1036 || !opt_for_fn (e
->caller
->decl
, optimize
))
1038 if (!can_inline_edge_p (e
, true))
1040 if (e
->recursive_p ())
1042 if (!can_inline_edge_by_limits_p (e
, true))
1044 if (!(*(bool *)has_hot_call
) && e
->maybe_hot_p ())
1045 *(bool *)has_hot_call
= true;
1050 /* If NODE has a caller, return true. */
1053 has_caller_p (struct cgraph_node
*node
, void *data ATTRIBUTE_UNUSED
)
1060 /* Decide if inlining NODE would reduce unit size by eliminating
1061 the offline copy of function.
1062 When COLD is true the cold calls are considered, too. */
1065 want_inline_function_to_all_callers_p (struct cgraph_node
*node
, bool cold
)
1067 bool has_hot_call
= false;
1069 /* Aliases gets inlined along with the function they alias. */
1072 /* Already inlined? */
1073 if (node
->inlined_to
)
1075 /* Does it have callers? */
1076 if (!node
->call_for_symbol_and_aliases (has_caller_p
, NULL
, true))
1078 /* Inlining into all callers would increase size? */
1079 if (growth_positive_p (node
, NULL
, INT_MIN
) > 0)
1081 /* All inlines must be possible. */
1082 if (node
->call_for_symbol_and_aliases (check_callers
, &has_hot_call
,
1085 if (!cold
&& !has_hot_call
)
1090 /* A cost model driving the inlining heuristics in a way so the edges with
1091 smallest badness are inlined first. After each inlining is performed
1092 the costs of all caller edges of nodes affected are recomputed so the
1093 metrics may accurately depend on values such as number of inlinable callers
1094 of the function or function body size. */
1097 edge_badness (struct cgraph_edge
*edge
, bool dump
)
1101 sreal edge_time
, unspec_edge_time
;
1102 struct cgraph_node
*callee
= edge
->callee
->ultimate_alias_target ();
1103 class ipa_fn_summary
*callee_info
= ipa_fn_summaries
->get (callee
);
1105 cgraph_node
*caller
= (edge
->caller
->inlined_to
1106 ? edge
->caller
->inlined_to
1109 growth
= estimate_edge_growth (edge
);
1110 edge_time
= estimate_edge_time (edge
, &unspec_edge_time
);
1111 hints
= estimate_edge_hints (edge
);
1112 gcc_checking_assert (edge_time
>= 0);
1113 /* Check that inlined time is better, but tolerate some roundoff issues.
1114 FIXME: When callee profile drops to 0 we account calls more. This
1115 should be fixed by never doing that. */
1116 gcc_checking_assert ((edge_time
* 100
1117 - callee_info
->time
* 101).to_int () <= 0
1118 || callee
->count
.ipa ().initialized_p ());
1119 gcc_checking_assert (growth
<= ipa_size_summaries
->get (callee
)->size
);
1123 fprintf (dump_file
, " Badness calculation for %s -> %s\n",
1124 edge
->caller
->dump_name (),
1125 edge
->callee
->dump_name ());
1126 fprintf (dump_file
, " size growth %i, time %f unspec %f ",
1128 edge_time
.to_double (),
1129 unspec_edge_time
.to_double ());
1130 ipa_dump_hints (dump_file
, hints
);
1131 if (big_speedup_p (edge
))
1132 fprintf (dump_file
, " big_speedup");
1133 fprintf (dump_file
, "\n");
1136 /* Always prefer inlining saving code size. */
1139 badness
= (sreal
) (-SREAL_MIN_SIG
+ growth
) << (SREAL_MAX_EXP
/ 256);
1141 fprintf (dump_file
, " %f: Growth %d <= 0\n", badness
.to_double (),
1144 /* Inlining into EXTERNAL functions is not going to change anything unless
1145 they are themselves inlined. */
1146 else if (DECL_EXTERNAL (caller
->decl
))
1149 fprintf (dump_file
, " max: function is external\n");
1150 return sreal::max ();
1152 /* When profile is available. Compute badness as:
1154 time_saved * caller_count
1155 goodness = -------------------------------------------------
1156 growth_of_caller * overall_growth * combined_size
1158 badness = - goodness
1160 Again use negative value to make calls with profile appear hotter
1163 else if (opt_for_fn (caller
->decl
, flag_guess_branch_prob
)
1164 || caller
->count
.ipa ().nonzero_p ())
1166 sreal numerator
, denominator
;
1168 sreal freq
= edge
->sreal_frequency ();
1169 sreal inlined_time
= compute_inlined_call_time (edge
, edge_time
, freq
);
1171 numerator
= (compute_uninlined_call_time (edge
, unspec_edge_time
, freq
)
1174 numerator
= ((sreal
) 1 >> 8);
1175 if (caller
->count
.ipa ().nonzero_p ())
1176 numerator
*= caller
->count
.ipa ().to_gcov_type ();
1177 else if (caller
->count
.ipa ().initialized_p ())
1178 numerator
= numerator
>> 11;
1179 denominator
= growth
;
1181 overall_growth
= callee_info
->growth
;
1183 /* Look for inliner wrappers of the form:
1189 noninline_callee ();
1191 Withhout penalizing this case, we usually inline noninline_callee
1192 into the inline_caller because overall_growth is small preventing
1193 further inlining of inline_caller.
1195 Penalize only callgraph edges to functions with small overall
1198 if (growth
> overall_growth
1199 /* ... and having only one caller which is not inlined ... */
1200 && callee_info
->single_caller
1201 && !edge
->caller
->inlined_to
1202 /* ... and edges executed only conditionally ... */
1204 /* ... consider case where callee is not inline but caller is ... */
1205 && ((!DECL_DECLARED_INLINE_P (edge
->callee
->decl
)
1206 && DECL_DECLARED_INLINE_P (caller
->decl
))
1207 /* ... or when early optimizers decided to split and edge
1208 frequency still indicates splitting is a win ... */
1209 || (callee
->split_part
&& !caller
->split_part
1210 && freq
* 100 < param_partial_inlining_entry_probability
1211 /* ... and do not overwrite user specified hints. */
1212 && (!DECL_DECLARED_INLINE_P (edge
->callee
->decl
)
1213 || DECL_DECLARED_INLINE_P (caller
->decl
)))))
1215 ipa_fn_summary
*caller_info
= ipa_fn_summaries
->get (caller
);
1216 int caller_growth
= caller_info
->growth
;
1218 /* Only apply the penalty when caller looks like inline candidate,
1219 and it is not called once. */
1220 if (!caller_info
->single_caller
&& overall_growth
< caller_growth
1221 && caller_info
->inlinable
1222 && ipa_size_summaries
->get (caller
)->size
1223 < (DECL_DECLARED_INLINE_P (caller
->decl
)
1224 ? inline_insns_single (caller
, false)
1225 : inline_insns_auto (caller
, false)))
1229 " Wrapper penalty. Increasing growth %i to %i\n",
1230 overall_growth
, caller_growth
);
1231 overall_growth
= caller_growth
;
1234 if (overall_growth
> 0)
1236 /* Strongly preffer functions with few callers that can be inlined
1237 fully. The square root here leads to smaller binaries at average.
1238 Watch however for extreme cases and return to linear function
1239 when growth is large. */
1240 if (overall_growth
< 256)
1241 overall_growth
*= overall_growth
;
1243 overall_growth
+= 256 * 256 - 256;
1244 denominator
*= overall_growth
;
1246 denominator
*= ipa_size_summaries
->get (caller
)->size
+ growth
;
1248 badness
= - numerator
/ denominator
;
1253 " %f: guessed profile. frequency %f, count %" PRId64
1254 " caller count %" PRId64
1255 " time w/o inlining %f, time with inlining %f"
1256 " overall growth %i (current) %i (original)"
1257 " %i (compensated)\n",
1258 badness
.to_double (),
1260 edge
->count
.ipa ().initialized_p () ? edge
->count
.ipa ().to_gcov_type () : -1,
1261 caller
->count
.ipa ().initialized_p () ? caller
->count
.ipa ().to_gcov_type () : -1,
1262 compute_uninlined_call_time (edge
,
1263 unspec_edge_time
, freq
).to_double (),
1264 inlined_time
.to_double (),
1265 estimate_growth (callee
),
1266 callee_info
->growth
, overall_growth
);
1269 /* When function local profile is not available or it does not give
1270 useful information (ie frequency is zero), base the cost on
1271 loop nest and overall size growth, so we optimize for overall number
1272 of functions fully inlined in program. */
1275 int nest
= MIN (ipa_call_summaries
->get (edge
)->loop_depth
, 8);
1278 /* Decrease badness if call is nested. */
1280 badness
= badness
>> nest
;
1282 badness
= badness
<< nest
;
1284 fprintf (dump_file
, " %f: no profile. nest %i\n",
1285 badness
.to_double (), nest
);
1287 gcc_checking_assert (badness
!= 0);
1289 if (edge
->recursive_p ())
1290 badness
= badness
.shift (badness
> 0 ? 4 : -4);
1291 if ((hints
& (INLINE_HINT_indirect_call
1292 | INLINE_HINT_loop_iterations
1293 | INLINE_HINT_loop_stride
))
1294 || callee_info
->growth
<= 0)
1295 badness
= badness
.shift (badness
> 0 ? -2 : 2);
1296 if (hints
& (INLINE_HINT_same_scc
))
1297 badness
= badness
.shift (badness
> 0 ? 3 : -3);
1298 else if (hints
& (INLINE_HINT_in_scc
))
1299 badness
= badness
.shift (badness
> 0 ? 2 : -2);
1300 else if (hints
& (INLINE_HINT_cross_module
))
1301 badness
= badness
.shift (badness
> 0 ? 1 : -1);
1302 if (DECL_DISREGARD_INLINE_LIMITS (callee
->decl
))
1303 badness
= badness
.shift (badness
> 0 ? -4 : 4);
1304 else if ((hints
& INLINE_HINT_declared_inline
))
1305 badness
= badness
.shift (badness
> 0 ? -3 : 3);
1307 fprintf (dump_file
, " Adjusted by hints %f\n", badness
.to_double ());
1311 /* Recompute badness of EDGE and update its key in HEAP if needed. */
1313 update_edge_key (edge_heap_t
*heap
, struct cgraph_edge
*edge
)
1315 sreal badness
= edge_badness (edge
, false);
1318 edge_heap_node_t
*n
= (edge_heap_node_t
*) edge
->aux
;
1319 gcc_checking_assert (n
->get_data () == edge
);
1321 /* fibonacci_heap::replace_key does busy updating of the
1322 heap that is unnecesarily expensive.
1323 We do lazy increases: after extracting minimum if the key
1324 turns out to be out of date, it is re-inserted into heap
1325 with correct value. */
1326 if (badness
< n
->get_key ())
1328 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1331 " decreasing badness %s -> %s, %f to %f\n",
1332 edge
->caller
->dump_name (),
1333 edge
->callee
->dump_name (),
1334 n
->get_key ().to_double (),
1335 badness
.to_double ());
1337 heap
->decrease_key (n
, badness
);
1342 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1345 " enqueuing call %s -> %s, badness %f\n",
1346 edge
->caller
->dump_name (),
1347 edge
->callee
->dump_name (),
1348 badness
.to_double ());
1350 edge
->aux
= heap
->insert (badness
, edge
);
1355 /* NODE was inlined.
1356 All caller edges needs to be resetted because
1357 size estimates change. Similarly callees needs reset
1358 because better context may be known. */
1361 reset_edge_caches (struct cgraph_node
*node
)
1363 struct cgraph_edge
*edge
;
1364 struct cgraph_edge
*e
= node
->callees
;
1365 struct cgraph_node
*where
= node
;
1366 struct ipa_ref
*ref
;
1368 if (where
->inlined_to
)
1369 where
= where
->inlined_to
;
1371 reset_node_cache (where
);
1373 if (edge_growth_cache
!= NULL
)
1374 for (edge
= where
->callers
; edge
; edge
= edge
->next_caller
)
1375 if (edge
->inline_failed
)
1376 edge_growth_cache
->remove (edge
);
1378 FOR_EACH_ALIAS (where
, ref
)
1379 reset_edge_caches (dyn_cast
<cgraph_node
*> (ref
->referring
));
1385 if (!e
->inline_failed
&& e
->callee
->callees
)
1386 e
= e
->callee
->callees
;
1389 if (edge_growth_cache
!= NULL
&& e
->inline_failed
)
1390 edge_growth_cache
->remove (e
);
1397 if (e
->caller
== node
)
1399 e
= e
->caller
->callers
;
1401 while (!e
->next_callee
);
1407 /* Recompute HEAP nodes for each of caller of NODE.
1408 UPDATED_NODES track nodes we already visited, to avoid redundant work.
1409 When CHECK_INLINABLITY_FOR is set, re-check for specified edge that
1410 it is inlinable. Otherwise check all edges. */
1413 update_caller_keys (edge_heap_t
*heap
, struct cgraph_node
*node
,
1414 bitmap updated_nodes
,
1415 struct cgraph_edge
*check_inlinablity_for
)
1417 struct cgraph_edge
*edge
;
1418 struct ipa_ref
*ref
;
1420 if ((!node
->alias
&& !ipa_fn_summaries
->get (node
)->inlinable
)
1421 || node
->inlined_to
)
1423 if (!bitmap_set_bit (updated_nodes
, node
->get_uid ()))
1426 FOR_EACH_ALIAS (node
, ref
)
1428 struct cgraph_node
*alias
= dyn_cast
<cgraph_node
*> (ref
->referring
);
1429 update_caller_keys (heap
, alias
, updated_nodes
, check_inlinablity_for
);
1432 for (edge
= node
->callers
; edge
; edge
= edge
->next_caller
)
1433 if (edge
->inline_failed
)
1435 if (!check_inlinablity_for
1436 || check_inlinablity_for
== edge
)
1438 if (can_inline_edge_p (edge
, false)
1439 && want_inline_small_function_p (edge
, false)
1440 && can_inline_edge_by_limits_p (edge
, false))
1441 update_edge_key (heap
, edge
);
1444 report_inline_failed_reason (edge
);
1445 heap
->delete_node ((edge_heap_node_t
*) edge
->aux
);
1450 update_edge_key (heap
, edge
);
1454 /* Recompute HEAP nodes for each uninlined call in NODE.
1455 This is used when we know that edge badnesses are going only to increase
1456 (we introduced new call site) and thus all we need is to insert newly
1457 created edges into heap. */
1460 update_callee_keys (edge_heap_t
*heap
, struct cgraph_node
*node
,
1461 bitmap updated_nodes
)
1463 struct cgraph_edge
*e
= node
->callees
;
1468 if (!e
->inline_failed
&& e
->callee
->callees
)
1469 e
= e
->callee
->callees
;
1472 enum availability avail
;
1473 struct cgraph_node
*callee
;
1474 /* We do not reset callee growth cache here. Since we added a new call,
1475 growth chould have just increased and consequentely badness metric
1476 don't need updating. */
1477 if (e
->inline_failed
1478 && (callee
= e
->callee
->ultimate_alias_target (&avail
, e
->caller
))
1479 && ipa_fn_summaries
->get (callee
) != NULL
1480 && ipa_fn_summaries
->get (callee
)->inlinable
1481 && avail
>= AVAIL_AVAILABLE
1482 && !bitmap_bit_p (updated_nodes
, callee
->get_uid ()))
1484 if (can_inline_edge_p (e
, false)
1485 && want_inline_small_function_p (e
, false)
1486 && can_inline_edge_by_limits_p (e
, false))
1487 update_edge_key (heap
, e
);
1490 report_inline_failed_reason (e
);
1491 heap
->delete_node ((edge_heap_node_t
*) e
->aux
);
1501 if (e
->caller
== node
)
1503 e
= e
->caller
->callers
;
1505 while (!e
->next_callee
);
1511 /* Enqueue all recursive calls from NODE into priority queue depending on
1512 how likely we want to recursively inline the call. */
1515 lookup_recursive_calls (struct cgraph_node
*node
, struct cgraph_node
*where
,
1518 struct cgraph_edge
*e
;
1519 enum availability avail
;
1521 for (e
= where
->callees
; e
; e
= e
->next_callee
)
1522 if (e
->callee
== node
1523 || (e
->callee
->ultimate_alias_target (&avail
, e
->caller
) == node
1524 && avail
> AVAIL_INTERPOSABLE
))
1525 heap
->insert (-e
->sreal_frequency (), e
);
1526 for (e
= where
->callees
; e
; e
= e
->next_callee
)
1527 if (!e
->inline_failed
)
1528 lookup_recursive_calls (node
, e
->callee
, heap
);
1531 /* Decide on recursive inlining: in the case function has recursive calls,
1532 inline until body size reaches given argument. If any new indirect edges
1533 are discovered in the process, add them to *NEW_EDGES, unless NEW_EDGES
1537 recursive_inlining (struct cgraph_edge
*edge
,
1538 vec
<cgraph_edge
*> *new_edges
)
1540 int limit
= param_max_inline_insns_recursive_auto
;
1541 edge_heap_t
heap (sreal::min ());
1542 struct cgraph_node
*node
;
1543 struct cgraph_edge
*e
;
1544 struct cgraph_node
*master_clone
= NULL
, *next
;
1548 node
= edge
->caller
;
1549 if (node
->inlined_to
)
1550 node
= node
->inlined_to
;
1552 if (DECL_DECLARED_INLINE_P (node
->decl
))
1553 limit
= param_max_inline_insns_recursive
;
1555 /* Make sure that function is small enough to be considered for inlining. */
1556 if (estimate_size_after_inlining (node
, edge
) >= limit
)
1558 lookup_recursive_calls (node
, node
, &heap
);
1564 " Performing recursive inlining on %s\n",
1567 /* Do the inlining and update list of recursive call during process. */
1568 while (!heap
.empty ())
1570 struct cgraph_edge
*curr
= heap
.extract_min ();
1571 struct cgraph_node
*cnode
, *dest
= curr
->callee
;
1573 if (!can_inline_edge_p (curr
, true)
1574 || !can_inline_edge_by_limits_p (curr
, true))
1577 /* MASTER_CLONE is produced in the case we already started modified
1578 the function. Be sure to redirect edge to the original body before
1579 estimating growths otherwise we will be seeing growths after inlining
1580 the already modified body. */
1583 curr
->redirect_callee (master_clone
);
1584 if (edge_growth_cache
!= NULL
)
1585 edge_growth_cache
->remove (curr
);
1588 if (estimate_size_after_inlining (node
, curr
) > limit
)
1590 curr
->redirect_callee (dest
);
1591 if (edge_growth_cache
!= NULL
)
1592 edge_growth_cache
->remove (curr
);
1597 for (cnode
= curr
->caller
;
1598 cnode
->inlined_to
; cnode
= cnode
->callers
->caller
)
1600 == curr
->callee
->ultimate_alias_target ()->decl
)
1603 if (!want_inline_self_recursive_call_p (curr
, node
, false, depth
))
1605 curr
->redirect_callee (dest
);
1606 if (edge_growth_cache
!= NULL
)
1607 edge_growth_cache
->remove (curr
);
1614 " Inlining call of depth %i", depth
);
1615 if (node
->count
.nonzero_p () && curr
->count
.initialized_p ())
1617 fprintf (dump_file
, " called approx. %.2f times per call",
1618 (double)curr
->count
.to_gcov_type ()
1619 / node
->count
.to_gcov_type ());
1621 fprintf (dump_file
, "\n");
1625 /* We need original clone to copy around. */
1626 master_clone
= node
->create_clone (node
->decl
, node
->count
,
1627 false, vNULL
, true, NULL
, NULL
);
1628 for (e
= master_clone
->callees
; e
; e
= e
->next_callee
)
1629 if (!e
->inline_failed
)
1630 clone_inlined_nodes (e
, true, false, NULL
);
1631 curr
->redirect_callee (master_clone
);
1632 if (edge_growth_cache
!= NULL
)
1633 edge_growth_cache
->remove (curr
);
1636 inline_call (curr
, false, new_edges
, &overall_size
, true);
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
)
1987 /* Be sure that caches are maintained consistent.
1988 This check is affected by scaling roundoff errors when compiling for
1989 IPA this we skip it in that case. */
1990 if (!edge
->callee
->count
.ipa_p ()
1991 && (!max_count
.initialized_p () || !max_count
.nonzero_p ()))
1993 sreal cached_badness
= edge_badness (edge
, false);
1995 int old_size_est
= estimate_edge_size (edge
);
1996 sreal old_time_est
= estimate_edge_time (edge
);
1997 int old_hints_est
= estimate_edge_hints (edge
);
1999 if (edge_growth_cache
!= NULL
)
2000 edge_growth_cache
->remove (edge
);
2001 gcc_assert (old_size_est
== estimate_edge_size (edge
));
2002 gcc_assert (old_time_est
== estimate_edge_time (edge
));
2005 gcc_assert (old_hints_est == estimate_edge_hints (edge));
2007 fails with profile feedback because some hints depends on
2008 maybe_hot_edge_p predicate and because callee gets inlined to other
2009 calls, the edge may become cold.
2010 This ought to be fixed by computing relative probabilities
2011 for given invocation but that will be better done once whole
2012 code is converted to sreals. Disable for now and revert to "wrong"
2013 value so enable/disable checking paths agree. */
2014 edge_growth_cache
->get (edge
)->hints
= old_hints_est
+ 1;
2016 /* When updating the edge costs, we only decrease badness in the keys.
2017 Increases of badness are handled lazilly; when we see key with out
2018 of date value on it, we re-insert it now. */
2019 current_badness
= edge_badness (edge
, false);
2020 gcc_assert (cached_badness
== current_badness
);
2021 gcc_assert (current_badness
>= badness
);
2024 current_badness
= edge_badness (edge
, false);
2026 current_badness
= edge_badness (edge
, false);
2028 if (current_badness
!= badness
)
2030 if (edge_heap
.min () && current_badness
> edge_heap
.min_key ())
2032 edge
->aux
= edge_heap
.insert (current_badness
, edge
);
2036 badness
= current_badness
;
2039 if (!can_inline_edge_p (edge
, true)
2040 || !can_inline_edge_by_limits_p (edge
, true))
2042 resolve_noninline_speculation (&edge_heap
, edge
);
2046 callee
= edge
->callee
->ultimate_alias_target ();
2047 growth
= estimate_edge_growth (edge
);
2051 "\nConsidering %s with %i size\n",
2052 callee
->dump_name (),
2053 ipa_size_summaries
->get (callee
)->size
);
2055 " to be inlined into %s in %s:%i\n"
2056 " Estimated badness is %f, frequency %.2f.\n",
2057 edge
->caller
->dump_name (),
2059 && (LOCATION_LOCUS (gimple_location ((const gimple
*)
2061 > BUILTINS_LOCATION
)
2062 ? gimple_filename ((const gimple
*) edge
->call_stmt
)
2065 ? gimple_lineno ((const gimple
*) edge
->call_stmt
)
2067 badness
.to_double (),
2068 edge
->sreal_frequency ().to_double ());
2069 if (edge
->count
.ipa ().initialized_p ())
2071 fprintf (dump_file
, " Called ");
2072 edge
->count
.ipa ().dump (dump_file
);
2073 fprintf (dump_file
, " times\n");
2075 if (dump_flags
& TDF_DETAILS
)
2076 edge_badness (edge
, true);
2079 if (overall_size
+ growth
> max_size
2080 && !DECL_DISREGARD_INLINE_LIMITS (callee
->decl
))
2082 edge
->inline_failed
= CIF_INLINE_UNIT_GROWTH_LIMIT
;
2083 report_inline_failed_reason (edge
);
2084 resolve_noninline_speculation (&edge_heap
, edge
);
2088 if (!want_inline_small_function_p (edge
, true))
2090 resolve_noninline_speculation (&edge_heap
, edge
);
2094 /* Heuristics for inlining small functions work poorly for
2095 recursive calls where we do effects similar to loop unrolling.
2096 When inlining such edge seems profitable, leave decision on
2097 specific inliner. */
2098 if (edge
->recursive_p ())
2100 where
= edge
->caller
;
2101 if (where
->inlined_to
)
2102 where
= where
->inlined_to
;
2103 if (!recursive_inlining (edge
,
2104 opt_for_fn (edge
->caller
->decl
,
2105 flag_indirect_inlining
)
2106 ? &new_indirect_edges
: NULL
))
2108 edge
->inline_failed
= CIF_RECURSIVE_INLINING
;
2109 resolve_noninline_speculation (&edge_heap
, edge
);
2112 reset_edge_caches (where
);
2113 /* Recursive inliner inlines all recursive calls of the function
2114 at once. Consequently we need to update all callee keys. */
2115 if (opt_for_fn (edge
->caller
->decl
, flag_indirect_inlining
))
2116 add_new_edges_to_heap (&edge_heap
, new_indirect_edges
);
2117 update_callee_keys (&edge_heap
, where
, updated_nodes
);
2118 bitmap_clear (updated_nodes
);
2122 struct cgraph_node
*outer_node
= NULL
;
2125 /* Consider the case where self recursive function A is inlined
2126 into B. This is desired optimization in some cases, since it
2127 leads to effect similar of loop peeling and we might completely
2128 optimize out the recursive call. However we must be extra
2131 where
= edge
->caller
;
2132 while (where
->inlined_to
)
2134 if (where
->decl
== callee
->decl
)
2135 outer_node
= where
, depth
++;
2136 where
= where
->callers
->caller
;
2139 && !want_inline_self_recursive_call_p (edge
, outer_node
,
2143 = (DECL_DISREGARD_INLINE_LIMITS (edge
->callee
->decl
)
2144 ? CIF_RECURSIVE_INLINING
: CIF_UNSPECIFIED
);
2145 resolve_noninline_speculation (&edge_heap
, edge
);
2148 else if (depth
&& dump_file
)
2149 fprintf (dump_file
, " Peeling recursion with depth %i\n", depth
);
2151 gcc_checking_assert (!callee
->inlined_to
);
2152 inline_call (edge
, true, &new_indirect_edges
, &overall_size
, true);
2153 add_new_edges_to_heap (&edge_heap
, new_indirect_edges
);
2155 reset_edge_caches (edge
->callee
);
2157 update_callee_keys (&edge_heap
, where
, updated_nodes
);
2159 where
= edge
->caller
;
2160 if (where
->inlined_to
)
2161 where
= where
->inlined_to
;
2163 /* Our profitability metric can depend on local properties
2164 such as number of inlinable calls and size of the function body.
2165 After inlining these properties might change for the function we
2166 inlined into (since it's body size changed) and for the functions
2167 called by function we inlined (since number of it inlinable callers
2169 update_caller_keys (&edge_heap
, where
, updated_nodes
, NULL
);
2170 /* Offline copy count has possibly changed, recompute if profile is
2172 struct cgraph_node
*n
= cgraph_node::get (edge
->callee
->decl
);
2173 if (n
!= edge
->callee
&& n
->analyzed
&& n
->count
.ipa ().initialized_p ())
2174 update_callee_keys (&edge_heap
, n
, updated_nodes
);
2175 bitmap_clear (updated_nodes
);
2177 if (dump_enabled_p ())
2179 ipa_fn_summary
*s
= ipa_fn_summaries
->get (where
);
2181 /* dump_printf can't handle %+i. */
2182 char buf_net_change
[100];
2183 snprintf (buf_net_change
, sizeof buf_net_change
, "%+i",
2184 overall_size
- old_size
);
2186 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS
, edge
->call_stmt
,
2187 " Inlined %C into %C which now has time %f and "
2188 "size %i, net change of %s.\n",
2189 edge
->callee
, edge
->caller
,
2190 s
->time
.to_double (),
2191 ipa_size_summaries
->get (edge
->caller
)->size
,
2194 if (min_size
> overall_size
)
2196 min_size
= overall_size
;
2197 max_size
= compute_max_insns (min_size
);
2200 fprintf (dump_file
, "New minimal size reached: %i\n", min_size
);
2204 free_growth_caches ();
2205 if (dump_enabled_p ())
2206 dump_printf (MSG_NOTE
,
2207 "Unit growth for small function inlining: %i->%i (%i%%)\n",
2208 initial_size
, overall_size
,
2209 initial_size
? overall_size
* 100 / (initial_size
) - 100: 0);
2210 symtab
->remove_edge_removal_hook (edge_removal_hook_holder
);
2213 /* Flatten NODE. Performed both during early inlining and
2214 at IPA inlining time. */
2217 flatten_function (struct cgraph_node
*node
, bool early
, bool update
)
2219 struct cgraph_edge
*e
;
2221 /* We shouldn't be called recursively when we are being processed. */
2222 gcc_assert (node
->aux
== NULL
);
2224 node
->aux
= (void *) node
;
2226 for (e
= node
->callees
; e
; e
= e
->next_callee
)
2228 struct cgraph_node
*orig_callee
;
2229 struct cgraph_node
*callee
= e
->callee
->ultimate_alias_target ();
2231 /* We've hit cycle? It is time to give up. */
2234 if (dump_enabled_p ())
2235 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, e
->call_stmt
,
2236 "Not inlining %C into %C to avoid cycle.\n",
2238 if (cgraph_inline_failed_type (e
->inline_failed
) != CIF_FINAL_ERROR
)
2239 e
->inline_failed
= CIF_RECURSIVE_INLINING
;
2243 /* When the edge is already inlined, we just need to recurse into
2244 it in order to fully flatten the leaves. */
2245 if (!e
->inline_failed
)
2247 flatten_function (callee
, early
, false);
2251 /* Flatten attribute needs to be processed during late inlining. For
2252 extra code quality we however do flattening during early optimization,
2255 ? !can_inline_edge_p (e
, true)
2256 && !can_inline_edge_by_limits_p (e
, true)
2257 : !can_early_inline_edge_p (e
))
2260 if (e
->recursive_p ())
2262 if (dump_enabled_p ())
2263 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, e
->call_stmt
,
2264 "Not inlining: recursive call.\n");
2268 if (gimple_in_ssa_p (DECL_STRUCT_FUNCTION (node
->decl
))
2269 != gimple_in_ssa_p (DECL_STRUCT_FUNCTION (callee
->decl
)))
2271 if (dump_enabled_p ())
2272 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, e
->call_stmt
,
2273 "Not inlining: SSA form does not match.\n");
2277 /* Inline the edge and flatten the inline clone. Avoid
2278 recursing through the original node if the node was cloned. */
2279 if (dump_enabled_p ())
2280 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS
, e
->call_stmt
,
2281 " Inlining %C into %C.\n",
2283 orig_callee
= callee
;
2284 inline_call (e
, true, NULL
, NULL
, false);
2285 if (e
->callee
!= orig_callee
)
2286 orig_callee
->aux
= (void *) node
;
2287 flatten_function (e
->callee
, early
, false);
2288 if (e
->callee
!= orig_callee
)
2289 orig_callee
->aux
= NULL
;
2293 cgraph_node
*where
= node
->inlined_to
? node
->inlined_to
: node
;
2294 if (update
&& opt_for_fn (where
->decl
, optimize
))
2295 ipa_update_overall_fn_summary (where
);
2298 /* Inline NODE to all callers. Worker for cgraph_for_node_and_aliases.
2299 DATA points to number of calls originally found so we avoid infinite
2303 inline_to_all_callers_1 (struct cgraph_node
*node
, void *data
,
2304 hash_set
<cgraph_node
*> *callers
)
2306 int *num_calls
= (int *)data
;
2307 bool callee_removed
= false;
2309 while (node
->callers
&& !node
->inlined_to
)
2311 struct cgraph_node
*caller
= node
->callers
->caller
;
2313 if (!can_inline_edge_p (node
->callers
, true)
2314 || !can_inline_edge_by_limits_p (node
->callers
, true)
2315 || node
->callers
->recursive_p ())
2318 fprintf (dump_file
, "Uninlinable call found; giving up.\n");
2325 cgraph_node
*ultimate
= node
->ultimate_alias_target ();
2327 "\nInlining %s size %i.\n",
2329 ipa_size_summaries
->get (ultimate
)->size
);
2331 " Called once from %s %i insns.\n",
2332 node
->callers
->caller
->name (),
2333 ipa_size_summaries
->get (node
->callers
->caller
)->size
);
2336 /* Remember which callers we inlined to, delaying updating the
2338 callers
->add (node
->callers
->caller
);
2339 inline_call (node
->callers
, true, NULL
, NULL
, false, &callee_removed
);
2342 " Inlined into %s which now has %i size\n",
2344 ipa_size_summaries
->get (caller
)->size
);
2345 if (!(*num_calls
)--)
2348 fprintf (dump_file
, "New calls found; giving up.\n");
2349 return callee_removed
;
2357 /* Wrapper around inline_to_all_callers_1 doing delayed overall summary
2361 inline_to_all_callers (struct cgraph_node
*node
, void *data
)
2363 hash_set
<cgraph_node
*> callers
;
2364 bool res
= inline_to_all_callers_1 (node
, data
, &callers
);
2365 /* Perform the delayed update of the overall summary of all callers
2366 processed. This avoids quadratic behavior in the cases where
2367 we have a lot of calls to the same function. */
2368 for (hash_set
<cgraph_node
*>::iterator i
= callers
.begin ();
2369 i
!= callers
.end (); ++i
)
2370 ipa_update_overall_fn_summary ((*i
)->inlined_to
? (*i
)->inlined_to
: *i
);
2374 /* Output overall time estimate. */
2376 dump_overall_stats (void)
2378 sreal sum_weighted
= 0, sum
= 0;
2379 struct cgraph_node
*node
;
2381 FOR_EACH_DEFINED_FUNCTION (node
)
2382 if (!node
->inlined_to
2385 ipa_fn_summary
*s
= ipa_fn_summaries
->get (node
);
2389 if (node
->count
.ipa ().initialized_p ())
2390 sum_weighted
+= s
->time
* node
->count
.ipa ().to_gcov_type ();
2393 fprintf (dump_file
, "Overall time estimate: "
2394 "%f weighted by profile: "
2395 "%f\n", sum
.to_double (), sum_weighted
.to_double ());
2398 /* Output some useful stats about inlining. */
2401 dump_inline_stats (void)
2403 int64_t inlined_cnt
= 0, inlined_indir_cnt
= 0;
2404 int64_t inlined_virt_cnt
= 0, inlined_virt_indir_cnt
= 0;
2405 int64_t noninlined_cnt
= 0, noninlined_indir_cnt
= 0;
2406 int64_t noninlined_virt_cnt
= 0, noninlined_virt_indir_cnt
= 0;
2407 int64_t inlined_speculative
= 0, inlined_speculative_ply
= 0;
2408 int64_t indirect_poly_cnt
= 0, indirect_cnt
= 0;
2409 int64_t reason
[CIF_N_REASONS
][2];
2410 sreal reason_freq
[CIF_N_REASONS
];
2412 struct cgraph_node
*node
;
2414 memset (reason
, 0, sizeof (reason
));
2415 for (i
=0; i
< CIF_N_REASONS
; i
++)
2417 FOR_EACH_DEFINED_FUNCTION (node
)
2419 struct cgraph_edge
*e
;
2420 for (e
= node
->callees
; e
; e
= e
->next_callee
)
2422 if (e
->inline_failed
)
2424 if (e
->count
.ipa ().initialized_p ())
2425 reason
[(int) e
->inline_failed
][0] += e
->count
.ipa ().to_gcov_type ();
2426 reason_freq
[(int) e
->inline_failed
] += e
->sreal_frequency ();
2427 reason
[(int) e
->inline_failed
][1] ++;
2428 if (DECL_VIRTUAL_P (e
->callee
->decl
)
2429 && e
->count
.ipa ().initialized_p ())
2431 if (e
->indirect_inlining_edge
)
2432 noninlined_virt_indir_cnt
+= e
->count
.ipa ().to_gcov_type ();
2434 noninlined_virt_cnt
+= e
->count
.ipa ().to_gcov_type ();
2436 else if (e
->count
.ipa ().initialized_p ())
2438 if (e
->indirect_inlining_edge
)
2439 noninlined_indir_cnt
+= e
->count
.ipa ().to_gcov_type ();
2441 noninlined_cnt
+= e
->count
.ipa ().to_gcov_type ();
2444 else if (e
->count
.ipa ().initialized_p ())
2448 if (DECL_VIRTUAL_P (e
->callee
->decl
))
2449 inlined_speculative_ply
+= e
->count
.ipa ().to_gcov_type ();
2451 inlined_speculative
+= e
->count
.ipa ().to_gcov_type ();
2453 else if (DECL_VIRTUAL_P (e
->callee
->decl
))
2455 if (e
->indirect_inlining_edge
)
2456 inlined_virt_indir_cnt
+= e
->count
.ipa ().to_gcov_type ();
2458 inlined_virt_cnt
+= e
->count
.ipa ().to_gcov_type ();
2462 if (e
->indirect_inlining_edge
)
2463 inlined_indir_cnt
+= e
->count
.ipa ().to_gcov_type ();
2465 inlined_cnt
+= e
->count
.ipa ().to_gcov_type ();
2469 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
2470 if (e
->indirect_info
->polymorphic
2471 & e
->count
.ipa ().initialized_p ())
2472 indirect_poly_cnt
+= e
->count
.ipa ().to_gcov_type ();
2473 else if (e
->count
.ipa ().initialized_p ())
2474 indirect_cnt
+= e
->count
.ipa ().to_gcov_type ();
2476 if (max_count
.initialized_p ())
2479 "Inlined %" PRId64
" + speculative "
2480 "%" PRId64
" + speculative polymorphic "
2481 "%" PRId64
" + previously indirect "
2482 "%" PRId64
" + virtual "
2483 "%" PRId64
" + virtual and previously indirect "
2484 "%" PRId64
"\n" "Not inlined "
2485 "%" PRId64
" + previously indirect "
2486 "%" PRId64
" + virtual "
2487 "%" PRId64
" + virtual and previously indirect "
2488 "%" PRId64
" + stil indirect "
2489 "%" PRId64
" + still indirect polymorphic "
2490 "%" PRId64
"\n", inlined_cnt
,
2491 inlined_speculative
, inlined_speculative_ply
,
2492 inlined_indir_cnt
, inlined_virt_cnt
, inlined_virt_indir_cnt
,
2493 noninlined_cnt
, noninlined_indir_cnt
, noninlined_virt_cnt
,
2494 noninlined_virt_indir_cnt
, indirect_cnt
, indirect_poly_cnt
);
2495 fprintf (dump_file
, "Removed speculations ");
2496 spec_rem
.dump (dump_file
);
2497 fprintf (dump_file
, "\n");
2499 dump_overall_stats ();
2500 fprintf (dump_file
, "\nWhy inlining failed?\n");
2501 for (i
= 0; i
< CIF_N_REASONS
; i
++)
2503 fprintf (dump_file
, "%-50s: %8i calls, %8f freq, %" PRId64
" count\n",
2504 cgraph_inline_failed_string ((cgraph_inline_failed_t
) i
),
2505 (int) reason
[i
][1], reason_freq
[i
].to_double (), reason
[i
][0]);
2508 /* Called when node is removed. */
2511 flatten_remove_node_hook (struct cgraph_node
*node
, void *data
)
2513 if (lookup_attribute ("flatten", DECL_ATTRIBUTES (node
->decl
)) == NULL
)
2516 hash_set
<struct cgraph_node
*> *removed
2517 = (hash_set
<struct cgraph_node
*> *) data
;
2518 removed
->add (node
);
2521 /* Decide on the inlining. We do so in the topological order to avoid
2522 expenses on updating data structures. */
2527 struct cgraph_node
*node
;
2529 struct cgraph_node
**order
;
2532 bool remove_functions
= false;
2534 order
= XCNEWVEC (struct cgraph_node
*, symtab
->cgraph_count
);
2537 ipa_dump_fn_summaries (dump_file
);
2539 nnodes
= ipa_reverse_postorder (order
);
2540 spec_rem
= profile_count::zero ();
2542 FOR_EACH_FUNCTION (node
)
2546 /* Recompute the default reasons for inlining because they may have
2547 changed during merging. */
2550 for (cgraph_edge
*e
= node
->callees
; e
; e
= e
->next_callee
)
2552 gcc_assert (e
->inline_failed
);
2553 initialize_inline_failed (e
);
2555 for (cgraph_edge
*e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
2556 initialize_inline_failed (e
);
2561 fprintf (dump_file
, "\nFlattening functions:\n");
2563 /* First shrink order array, so that it only contains nodes with
2564 flatten attribute. */
2565 for (i
= nnodes
- 1, j
= i
; i
>= 0; i
--)
2568 if (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
);