Line data Source code
1 : /* Inlining decision heuristics.
2 : Copyright (C) 2003-2026 Free Software Foundation, Inc.
3 : Contributed by Jan Hubicka
4 :
5 : This file is part of GCC.
6 :
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
10 : version.
11 :
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
15 : for more details.
16 :
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/>. */
20 :
21 : /* Inlining decision heuristics
22 :
23 : The implementation of inliner is organized as follows:
24 :
25 : inlining heuristics limits
26 :
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
29 : on).
30 :
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
34 : inlining.
35 :
36 : inlining heuristics
37 :
38 : The inliner itself is split into two passes:
39 :
40 : pass_early_inlining
41 :
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.
45 :
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.
52 :
53 : The pass handle the obvious inlining decisions within the compilation
54 : unit - inlining auto inline functions, inlining for size and
55 : flattening.
56 :
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
60 : optimizers.
61 :
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.
67 :
68 : pass_ipa_inline
69 :
70 : This is the real inliner able to handle inlining with whole program
71 : knowledge. It performs following steps:
72 :
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.
76 :
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.
81 :
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.
87 :
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. */
91 :
92 : #include "config.h"
93 : #include "system.h"
94 : #include "coretypes.h"
95 : #include "backend.h"
96 : #include "target.h"
97 : #include "rtl.h"
98 : #include "tree.h"
99 : #include "gimple.h"
100 : #include "alloc-pool.h"
101 : #include "tree-pass.h"
102 : #include "gimple-ssa.h"
103 : #include "cgraph.h"
104 : #include "lto-streamer.h"
105 : #include "trans-mem.h"
106 : #include "calls.h"
107 : #include "tree-inline.h"
108 : #include "profile.h"
109 : #include "symbol-summary.h"
110 : #include "tree-vrp.h"
111 : #include "sreal.h"
112 : #include "ipa-cp.h"
113 : #include "ipa-prop.h"
114 : #include "ipa-fnsummary.h"
115 : #include "ipa-inline.h"
116 : #include "ipa-utils.h"
117 : #include "auto-profile.h"
118 : #include "builtins.h"
119 : #include "fibonacci_heap.h"
120 : #include "stringpool.h"
121 : #include "attribs.h"
122 : #include "asan.h"
123 : #include "ipa-strub.h"
124 : #include "ipa-modref-tree.h"
125 : #include "ipa-modref.h"
126 :
127 : /* Inliner uses greedy algorithm to inline calls in a priority order.
128 : Badness is used as the key in a Fibonacci heap which roughly corresponds
129 : to negation of benefit to cost ratios.
130 : In case multiple calls has same priority we want to stabilize the outcomes
131 : for which we use ids. */
132 : class inline_badness
133 : {
134 : public:
135 : sreal badness;
136 : int uid;
137 1049713 : inline_badness ()
138 1049713 : : badness (sreal::min ()), uid (0)
139 : {
140 : }
141 3796828 : inline_badness (cgraph_edge *e, sreal b)
142 18205 : : badness (b), uid (e->get_uid ())
143 : {
144 : }
145 931015 : bool operator<= (const inline_badness &other)
146 : {
147 931015 : if (badness != other.badness)
148 931015 : return badness <= other.badness;
149 0 : return uid <= other.uid;
150 : }
151 1049713 : bool operator== (const inline_badness &other)
152 : {
153 1902566 : return badness == other.badness && uid == other.uid;
154 : }
155 0 : bool operator!= (const inline_badness &other)
156 : {
157 1049713 : return badness != other.badness || uid != other.uid;
158 : }
159 29377404 : bool operator< (const inline_badness &other)
160 : {
161 29377404 : if (badness != other.badness)
162 24379352 : return badness < other.badness;
163 4998052 : return uid < other.uid;
164 : }
165 13361671 : bool operator> (const inline_badness &other)
166 : {
167 13361671 : if (badness != other.badness)
168 10979760 : return badness > other.badness;
169 2381911 : return uid > other.uid;
170 : }
171 : };
172 :
173 : typedef fibonacci_heap <inline_badness, cgraph_edge> edge_heap_t;
174 : typedef fibonacci_node <inline_badness, cgraph_edge> edge_heap_node_t;
175 :
176 : /* Statistics we collect about inlining algorithm. */
177 : static int overall_size;
178 : static bool has_nonzero_ipa_profile;
179 : static profile_count spec_rem;
180 :
181 : /* Return false when inlining edge E would lead to violating
182 : limits on function unit growth or stack usage growth.
183 :
184 : The relative function body growth limit is present generally
185 : to avoid problems with non-linear behavior of the compiler.
186 : To allow inlining huge functions into tiny wrapper, the limit
187 : is always based on the bigger of the two functions considered.
188 :
189 : For stack growth limits we always base the growth in stack usage
190 : of the callers. We want to prevent applications from segfaulting
191 : on stack overflow when functions with huge stack frames gets
192 : inlined. */
193 :
194 : static bool
195 6907440 : caller_growth_limits (struct cgraph_edge *e)
196 : {
197 6907440 : struct cgraph_node *to = e->caller;
198 6907440 : struct cgraph_node *what = e->callee->ultimate_alias_target ();
199 6907440 : int newsize;
200 6907440 : int limit = 0;
201 6907440 : HOST_WIDE_INT stack_size_limit = 0, inlined_stack;
202 6907440 : ipa_size_summary *outer_info = ipa_size_summaries->get (to);
203 :
204 : /* Look for function e->caller is inlined to. While doing
205 : so work out the largest function body on the way. As
206 : described above, we want to base our function growth
207 : limits based on that. Not on the self size of the
208 : outer function, not on the self size of inline code
209 : we immediately inline to. This is the most relaxed
210 : interpretation of the rule "do not grow large functions
211 : too much in order to prevent compiler from exploding". */
212 10311000 : while (true)
213 : {
214 8609220 : ipa_size_summary *size_info = ipa_size_summaries->get (to);
215 8609220 : if (limit < size_info->self_size)
216 : limit = size_info->self_size;
217 8609220 : if (stack_size_limit < size_info->estimated_self_stack_size)
218 : stack_size_limit = size_info->estimated_self_stack_size;
219 8609220 : if (to->inlined_to)
220 1701780 : to = to->callers->caller;
221 : else
222 : break;
223 1701780 : }
224 :
225 6907440 : ipa_fn_summary *what_info = ipa_fn_summaries->get (what);
226 6907440 : ipa_size_summary *what_size_info = ipa_size_summaries->get (what);
227 :
228 6907440 : if (limit < what_size_info->self_size)
229 : limit = what_size_info->self_size;
230 :
231 6907440 : limit += limit * opt_for_fn (to->decl, param_large_function_growth) / 100;
232 :
233 : /* Check the size after inlining against the function limits. But allow
234 : the function to shrink if it went over the limits by forced inlining. */
235 6907440 : newsize = estimate_size_after_inlining (to, e);
236 6907440 : if (newsize >= ipa_size_summaries->get (what)->size
237 6725754 : && newsize > opt_for_fn (to->decl, param_large_function_insns)
238 7166683 : && newsize > limit)
239 : {
240 7483 : e->inline_failed = CIF_LARGE_FUNCTION_GROWTH_LIMIT;
241 7483 : return false;
242 : }
243 :
244 6899957 : if (!what_info->estimated_stack_size)
245 : return true;
246 :
247 : /* FIXME: Stack size limit often prevents inlining in Fortran programs
248 : due to large i/o datastructures used by the Fortran front-end.
249 : We ought to ignore this limit when we know that the edge is executed
250 : on every invocation of the caller (i.e. its call statement dominates
251 : exit block). We do not track this information, yet. */
252 2143914 : stack_size_limit += ((gcov_type)stack_size_limit
253 1071957 : * opt_for_fn (to->decl, param_stack_frame_growth)
254 1071957 : / 100);
255 :
256 1071957 : inlined_stack = (ipa_get_stack_frame_offset (to)
257 1071957 : + outer_info->estimated_self_stack_size
258 1071957 : + what_info->estimated_stack_size);
259 : /* Check new stack consumption with stack consumption at the place
260 : stack is used. */
261 1071957 : if (inlined_stack > stack_size_limit
262 : /* If function already has large stack usage from sibling
263 : inline call, we can inline, too.
264 : This bit overoptimistically assume that we are good at stack
265 : packing. */
266 309676 : && inlined_stack > ipa_fn_summaries->get (to)->estimated_stack_size
267 1368576 : && inlined_stack > opt_for_fn (to->decl, param_large_stack_frame))
268 : {
269 63736 : e->inline_failed = CIF_LARGE_STACK_FRAME_GROWTH_LIMIT;
270 63736 : return false;
271 : }
272 : return true;
273 : }
274 :
275 : /* Dump info about why inlining has failed. */
276 :
277 : static void
278 5113698 : report_inline_failed_reason (struct cgraph_edge *e)
279 : {
280 5113698 : if (dump_enabled_p ())
281 : {
282 2386 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
283 : " not inlinable: %C -> %C, %s\n",
284 : e->caller, e->callee,
285 : cgraph_inline_failed_string (e->inline_failed));
286 2386 : if ((e->inline_failed == CIF_TARGET_OPTION_MISMATCH
287 2386 : || e->inline_failed == CIF_OPTIMIZATION_MISMATCH)
288 2 : && e->caller->lto_file_data
289 2386 : && e->callee->ultimate_alias_target ()->lto_file_data)
290 : {
291 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
292 : " LTO objects: %s, %s\n",
293 0 : e->caller->lto_file_data->file_name,
294 0 : e->callee->ultimate_alias_target ()->lto_file_data->file_name);
295 : }
296 2386 : if (e->inline_failed == CIF_TARGET_OPTION_MISMATCH
297 2 : && dump_file)
298 : {
299 0 : struct cl_target_option *opt_caller
300 0 : = target_opts_for_fn (e->caller->decl);
301 0 : struct cl_target_option *opt_callee
302 0 : = target_opts_for_fn (e->callee->ultimate_alias_target ()->decl);
303 0 : if (opt_caller != NULL && opt_callee != NULL)
304 0 : cl_target_option_print_diff (dump_file, 2, opt_caller, opt_callee);
305 : }
306 2386 : if (e->inline_failed == CIF_OPTIMIZATION_MISMATCH)
307 0 : if (dump_file)
308 0 : cl_optimization_print_diff
309 0 : (dump_file, 2, opts_for_fn (e->caller->decl),
310 0 : opts_for_fn (e->callee->ultimate_alias_target ()->decl));
311 : }
312 5113698 : }
313 :
314 : /* Decide whether sanitizer-related attributes allow inlining. */
315 :
316 : static bool
317 10137448 : sanitize_attrs_match_for_inline_p (const_tree caller, const_tree callee)
318 : {
319 10137448 : if (!caller || !callee)
320 : return true;
321 :
322 : /* Follow clang and allow inlining for always_inline functions. */
323 10137448 : if (lookup_attribute ("always_inline", DECL_ATTRIBUTES (callee)))
324 : return true;
325 :
326 9498875 : const sanitize_code codes[] =
327 : {
328 : SANITIZE_ADDRESS,
329 : SANITIZE_THREAD,
330 : SANITIZE_UNDEFINED,
331 : SANITIZE_UNDEFINED_NONDEFAULT,
332 : SANITIZE_POINTER_COMPARE,
333 : SANITIZE_POINTER_SUBTRACT
334 : };
335 :
336 66491133 : for (unsigned i = 0; i < ARRAY_SIZE (codes); i++)
337 113984888 : if (sanitize_flags_p (codes[i], caller)
338 56992444 : != sanitize_flags_p (codes[i], callee))
339 : return false;
340 :
341 9498689 : if (sanitize_coverage_p (caller) != sanitize_coverage_p (callee))
342 0 : return false;
343 :
344 : return true;
345 : }
346 :
347 : /* Used for flags where it is safe to inline when caller's value is
348 : grater than callee's. */
349 : #define check_maybe_up(flag) \
350 : (opts_for_fn (caller->decl)->x_##flag \
351 : != opts_for_fn (callee->decl)->x_##flag \
352 : && (!always_inline \
353 : || opts_for_fn (caller->decl)->x_##flag \
354 : < opts_for_fn (callee->decl)->x_##flag))
355 : /* Used for flags where it is safe to inline when caller's value is
356 : smaller than callee's. */
357 : #define check_maybe_down(flag) \
358 : (opts_for_fn (caller->decl)->x_##flag \
359 : != opts_for_fn (callee->decl)->x_##flag \
360 : && (!always_inline \
361 : || opts_for_fn (caller->decl)->x_##flag \
362 : > opts_for_fn (callee->decl)->x_##flag))
363 : /* Used for flags where exact match is needed for correctness. */
364 : #define check_match(flag) \
365 : (opts_for_fn (caller->decl)->x_##flag \
366 : != opts_for_fn (callee->decl)->x_##flag)
367 :
368 : /* Decide if we can inline the edge and possibly update
369 : inline_failed reason.
370 : We check whether inlining is possible at all and whether
371 : caller growth limits allow doing so.
372 :
373 : if REPORT is true, output reason to the dump file. */
374 :
375 : static bool
376 14374182 : can_inline_edge_p (struct cgraph_edge *e, bool report,
377 : bool early = false)
378 : {
379 14374182 : gcc_checking_assert (e->inline_failed);
380 :
381 14374182 : if (cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR)
382 : {
383 3738368 : if (report)
384 3706598 : report_inline_failed_reason (e);
385 : return false;
386 : }
387 :
388 10635814 : bool inlinable = true;
389 10635814 : enum availability avail;
390 9212790 : cgraph_node *caller = (e->caller->inlined_to
391 10635814 : ? e->caller->inlined_to : e->caller);
392 10635814 : cgraph_node *callee = e->callee->ultimate_alias_target (&avail, caller);
393 :
394 10635814 : if (!callee->definition)
395 : {
396 1435 : e->inline_failed = CIF_BODY_NOT_AVAILABLE;
397 1435 : inlinable = false;
398 : }
399 10635814 : if (!early && (!opt_for_fn (callee->decl, optimize)
400 6225547 : || !opt_for_fn (caller->decl, optimize)))
401 : {
402 364 : e->inline_failed = CIF_FUNCTION_NOT_OPTIMIZED;
403 364 : inlinable = false;
404 : }
405 10635450 : else if (callee->calls_comdat_local)
406 : {
407 21627 : e->inline_failed = CIF_USES_COMDAT_LOCAL;
408 21627 : inlinable = false;
409 : }
410 10613823 : else if (avail <= AVAIL_INTERPOSABLE)
411 : {
412 133418 : e->inline_failed = CIF_OVERWRITABLE;
413 133418 : inlinable = false;
414 : }
415 : /* All edges with call_stmt_cannot_inline_p should have inline_failed
416 : initialized to one of FINAL_ERROR reasons. */
417 10480405 : else if (e->call_stmt_cannot_inline_p)
418 0 : gcc_unreachable ();
419 : /* Don't inline if the functions have different EH personalities. */
420 10480405 : else if (DECL_FUNCTION_PERSONALITY (caller->decl)
421 2980642 : && DECL_FUNCTION_PERSONALITY (callee->decl)
422 10480405 : && (DECL_FUNCTION_PERSONALITY (caller->decl)
423 245460 : != DECL_FUNCTION_PERSONALITY (callee->decl)))
424 : {
425 0 : e->inline_failed = CIF_EH_PERSONALITY;
426 0 : inlinable = false;
427 : }
428 : /* TM pure functions should not be inlined into non-TM_pure
429 : functions. */
430 10480405 : else if (is_tm_pure (callee->decl) && !is_tm_pure (caller->decl))
431 : {
432 30 : e->inline_failed = CIF_UNSPECIFIED;
433 30 : inlinable = false;
434 : }
435 : /* Check compatibility of target optimization options. */
436 10480375 : else if (!targetm.target_option.can_inline_p (caller->decl,
437 : callee->decl))
438 : {
439 477 : e->inline_failed = CIF_TARGET_OPTION_MISMATCH;
440 477 : inlinable = false;
441 : }
442 10479898 : else if (ipa_fn_summaries->get (callee) == NULL
443 10479896 : || !ipa_fn_summaries->get (callee)->inlinable)
444 : {
445 342450 : e->inline_failed = CIF_FUNCTION_NOT_INLINABLE;
446 342450 : inlinable = false;
447 : }
448 : /* Don't inline a function with mismatched sanitization attributes. */
449 10137448 : else if (!sanitize_attrs_match_for_inline_p (caller->decl, callee->decl))
450 : {
451 186 : e->inline_failed = CIF_SANITIZE_ATTRIBUTE_MISMATCH;
452 186 : inlinable = false;
453 : }
454 :
455 10635814 : if (inlinable && !strub_inlinable_to_p (callee, caller))
456 : {
457 1104 : e->inline_failed = CIF_UNSPECIFIED;
458 1104 : inlinable = false;
459 : }
460 10635814 : if (inlinable && callee->must_remain_in_tu_body
461 9 : && caller->lto_file_data != callee->lto_file_data)
462 : {
463 9 : e->inline_failed = CIF_MUST_REMAIN_IN_TU;
464 9 : inlinable = false;
465 : }
466 10635814 : if (!inlinable && report)
467 492263 : report_inline_failed_reason (e);
468 : return inlinable;
469 : }
470 :
471 : /* Return inlining_insns_single limit for function N. If HINT or HINT2 is true
472 : scale up the bound. */
473 :
474 : static int
475 9855428 : inline_insns_single (cgraph_node *n, bool hint, bool hint2)
476 : {
477 9855428 : if (hint && hint2)
478 : {
479 3458461 : int64_t spd = opt_for_fn (n->decl, param_inline_heuristics_hint_percent);
480 3458461 : spd = spd * spd;
481 3458461 : if (spd > 1000000)
482 : spd = 1000000;
483 3458461 : return opt_for_fn (n->decl, param_max_inline_insns_single) * spd / 100;
484 : }
485 6396967 : if (hint || hint2)
486 739144 : return opt_for_fn (n->decl, param_max_inline_insns_single)
487 739144 : * opt_for_fn (n->decl, param_inline_heuristics_hint_percent) / 100;
488 5657823 : return opt_for_fn (n->decl, param_max_inline_insns_single);
489 : }
490 :
491 : /* Return inlining_insns_auto limit for function N. If HINT or HINT2 is true
492 : scale up the bound. */
493 :
494 : static int
495 5579617 : inline_insns_auto (cgraph_node *n, bool hint, bool hint2)
496 : {
497 5579617 : int max_inline_insns_auto = opt_for_fn (n->decl, param_max_inline_insns_auto);
498 5579617 : if (hint && hint2)
499 : {
500 2034609 : int64_t spd = opt_for_fn (n->decl, param_inline_heuristics_hint_percent);
501 2034609 : spd = spd * spd;
502 2034609 : if (spd > 1000000)
503 : spd = 1000000;
504 2034609 : return max_inline_insns_auto * spd / 100;
505 : }
506 3545008 : if (hint || hint2)
507 1488889 : return max_inline_insns_auto
508 1488889 : * opt_for_fn (n->decl, param_inline_heuristics_hint_percent) / 100;
509 : return max_inline_insns_auto;
510 : }
511 :
512 : enum can_inline_edge_by_limits_flags
513 : {
514 : /* True if we are early inlining. */
515 : CAN_INLINE_EARLY = 1,
516 : /* Ignore size limits. */
517 : CAN_INLINE_DISREGARD_LIMITS = 2,
518 : /* Force size limits (ignore always_inline). This is used for
519 : recursive inlining where always_inline may lead to inline bombs
520 : and technically it is non-sential anyway. */
521 : CAN_INLINE_FORCE_LIMITS = 4,
522 : /* Report decision to dump file. */
523 : CAN_INLINE_REPORT = 8,
524 : };
525 :
526 : /* Decide if we can inline the edge and possibly update
527 : inline_failed reason.
528 : We check whether inlining is possible at all and whether
529 : caller growth limits allow doing so. */
530 :
531 : static bool
532 7557343 : can_inline_edge_by_limits_p (struct cgraph_edge *e, int flags)
533 : {
534 7557343 : gcc_checking_assert (e->inline_failed);
535 :
536 7557343 : if (cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR)
537 : {
538 210 : if (flags & CAN_INLINE_REPORT)
539 210 : report_inline_failed_reason (e);
540 : return false;
541 : }
542 :
543 7557133 : bool inlinable = true;
544 7557133 : enum availability avail;
545 6712716 : cgraph_node *caller = (e->caller->inlined_to
546 7557133 : ? e->caller->inlined_to : e->caller);
547 7557133 : cgraph_node *callee = e->callee->ultimate_alias_target (&avail, caller);
548 7557133 : tree caller_tree = DECL_FUNCTION_SPECIFIC_OPTIMIZATION (caller->decl);
549 7557133 : tree callee_tree
550 7557133 : = callee ? DECL_FUNCTION_SPECIFIC_OPTIMIZATION (callee->decl) : NULL;
551 : /* Check if caller growth allows the inlining. */
552 7557133 : if (!(flags & CAN_INLINE_DISREGARD_LIMITS)
553 7548778 : && ((flags & CAN_INLINE_FORCE_LIMITS)
554 7513962 : || (!DECL_DISREGARD_INLINE_LIMITS (callee->decl)
555 6872768 : && !lookup_attribute ("flatten",
556 6872768 : DECL_ATTRIBUTES (caller->decl))))
557 14464573 : && !caller_growth_limits (e))
558 : inlinable = false;
559 7485914 : else if (callee->externally_visible
560 4964556 : && !DECL_DISREGARD_INLINE_LIMITS (callee->decl)
561 11946340 : && flag_live_patching == LIVE_PATCHING_INLINE_ONLY_STATIC)
562 : {
563 2 : e->inline_failed = CIF_EXTERN_LIVE_ONLY_STATIC;
564 2 : inlinable = false;
565 : }
566 : /* Don't inline a function with a higher optimization level than the
567 : caller. FIXME: this is really just tip of iceberg of handling
568 : optimization attribute. */
569 7485912 : else if (caller_tree != callee_tree)
570 : {
571 9994 : bool always_inline =
572 9994 : (DECL_DISREGARD_INLINE_LIMITS (callee->decl)
573 12681 : && lookup_attribute ("always_inline",
574 2687 : DECL_ATTRIBUTES (callee->decl)));
575 9994 : ipa_fn_summary *caller_info = ipa_fn_summaries->get (caller);
576 9994 : ipa_fn_summary *callee_info = ipa_fn_summaries->get (callee);
577 :
578 : /* Until GCC 4.9 we did not check the semantics-altering flags
579 : below and inlined across optimization boundaries.
580 : Enabling checks below breaks several packages by refusing
581 : to inline library always_inline functions. See PR65873.
582 : Disable the check for early inlining for now until better solution
583 : is found. */
584 9994 : if (always_inline && (flags & CAN_INLINE_EARLY))
585 : ;
586 : /* There are some options that change IL semantics which means
587 : we cannot inline in these cases for correctness reason.
588 : Not even for always_inline declared functions. */
589 7307 : else if (check_match (flag_wrapv)
590 7307 : || check_match (flag_trapv)
591 7307 : || check_match (flag_pcc_struct_return)
592 7307 : || check_maybe_down (optimize_debug)
593 : /* When caller or callee does FP math, be sure FP codegen flags
594 : compatible. */
595 7301 : || ((caller_info->fp_expressions && callee_info->fp_expressions)
596 1273 : && (check_maybe_up (flag_rounding_math)
597 1273 : || check_maybe_up (flag_trapping_math)
598 1271 : || check_maybe_down (flag_unsafe_math_optimizations)
599 1271 : || check_maybe_down (flag_finite_math_only)
600 1270 : || check_maybe_up (flag_signaling_nans)
601 1270 : || check_maybe_up (flag_complex_method)
602 1269 : || check_maybe_up (flag_signed_zeros)
603 1269 : || check_maybe_down (flag_associative_math)
604 1253 : || check_maybe_down (flag_reciprocal_math)
605 1253 : || check_maybe_down (flag_fp_int_builtin_inexact)
606 : /* Strictly speaking only when the callee contains function
607 : calls that may end up setting errno. */
608 1253 : || check_maybe_up (flag_errno_math)))
609 : /* We do not want to make code compiled with exceptions to be
610 : brought into a non-EH function unless we know that the callee
611 : does not throw.
612 : This is tracked by DECL_FUNCTION_PERSONALITY. */
613 7281 : || (check_maybe_up (flag_non_call_exceptions)
614 0 : && DECL_FUNCTION_PERSONALITY (callee->decl))
615 7281 : || (check_maybe_up (flag_exceptions)
616 16 : && DECL_FUNCTION_PERSONALITY (callee->decl))
617 : /* When devirtualization is disabled for callee, it is not safe
618 : to inline it as we possibly mangled the type info.
619 : Allow early inlining of always inlines. */
620 14588 : || (!(flags & CAN_INLINE_EARLY) && check_maybe_down (flag_devirtualize)))
621 : {
622 34 : e->inline_failed = CIF_OPTIMIZATION_MISMATCH;
623 34 : inlinable = false;
624 : }
625 : /* gcc.dg/pr43564.c. Apply user-forced inline even at -O0. */
626 7273 : else if (always_inline)
627 : ;
628 : /* When user added an attribute to the callee honor it. */
629 7273 : else if (lookup_attribute ("optimize", DECL_ATTRIBUTES (callee->decl))
630 7273 : && opts_for_fn (caller->decl) != opts_for_fn (callee->decl))
631 : {
632 2456 : e->inline_failed = CIF_OPTIMIZATION_MISMATCH;
633 2456 : inlinable = false;
634 : }
635 : /* If explicit optimize attribute are not used, the mismatch is caused
636 : by different command line options used to build different units.
637 : Do not care about COMDAT functions - those are intended to be
638 : optimized with the optimization flags of module they are used in.
639 : Also do not care about mixing up size/speed optimization when
640 : DECL_DISREGARD_INLINE_LIMITS is set. */
641 4817 : else if ((callee->merged_comdat
642 0 : && !lookup_attribute ("optimize",
643 0 : DECL_ATTRIBUTES (caller->decl)))
644 4817 : || DECL_DISREGARD_INLINE_LIMITS (callee->decl))
645 : ;
646 : /* If mismatch is caused by merging two LTO units with different
647 : optimization flags we want to be bit nicer. However never inline
648 : if one of functions is not optimized at all. */
649 4817 : else if (!opt_for_fn (callee->decl, optimize)
650 4817 : || !opt_for_fn (caller->decl, optimize))
651 : {
652 0 : e->inline_failed = CIF_OPTIMIZATION_MISMATCH;
653 0 : inlinable = false;
654 : }
655 : /* If callee is optimized for size and caller is not, allow inlining if
656 : code shrinks or we are in param_max_inline_insns_single limit and
657 : callee is inline (and thus likely an unified comdat).
658 : This will allow caller to run faster. */
659 4817 : else if (opt_for_fn (callee->decl, optimize_size)
660 4817 : > opt_for_fn (caller->decl, optimize_size))
661 : {
662 118 : int growth = estimate_edge_growth (e);
663 118 : if (growth > opt_for_fn (caller->decl, param_max_inline_insns_size)
664 118 : && (!DECL_DECLARED_INLINE_P (callee->decl)
665 65 : && growth >= MAX (inline_insns_single (caller, false, false),
666 : inline_insns_auto (caller, false, false))))
667 : {
668 0 : e->inline_failed = CIF_OPTIMIZATION_MISMATCH;
669 0 : inlinable = false;
670 : }
671 : }
672 : /* If callee is more aggressively optimized for performance than caller,
673 : we generally want to inline only cheap (runtime wise) functions. */
674 4699 : else if (opt_for_fn (callee->decl, optimize_size)
675 : < opt_for_fn (caller->decl, optimize_size)
676 4699 : || (opt_for_fn (callee->decl, optimize)
677 : > opt_for_fn (caller->decl, optimize)))
678 : {
679 12675 : if (estimate_edge_time (e)
680 4225 : >= 20 + ipa_call_summaries->get (e)->call_stmt_time)
681 : {
682 1492 : e->inline_failed = CIF_OPTIMIZATION_MISMATCH;
683 1492 : inlinable = false;
684 : }
685 : }
686 :
687 : }
688 :
689 75203 : if (!inlinable && (flags & CAN_INLINE_REPORT))
690 68726 : report_inline_failed_reason (e);
691 : return inlinable;
692 : }
693 :
694 :
695 : /* Return true if the edge E is inlinable during early inlining. */
696 :
697 : static bool
698 4410297 : can_early_inline_edge_p (struct cgraph_edge *e)
699 : {
700 4408734 : cgraph_node *caller = (e->caller->inlined_to
701 4410297 : ? e->caller->inlined_to : e->caller);
702 4410297 : struct cgraph_node *callee = e->callee->ultimate_alias_target ();
703 : /* Early inliner might get called at WPA stage when IPA pass adds new
704 : function. In this case we cannot really do any of early inlining
705 : because function bodies are missing. */
706 4410297 : if (cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR)
707 : return false;
708 4409965 : if (!gimple_has_body_p (callee->decl))
709 : {
710 0 : e->inline_failed = CIF_BODY_NOT_AVAILABLE;
711 0 : return false;
712 : }
713 8819930 : gcc_assert (gimple_in_ssa_p (DECL_STRUCT_FUNCTION (e->caller->decl))
714 : && gimple_in_ssa_p (DECL_STRUCT_FUNCTION (callee->decl)));
715 4409965 : if (coverage_instrumentation_p ()
716 4411641 : && ((lookup_attribute ("no_profile_instrument_function",
717 1676 : DECL_ATTRIBUTES (caller->decl)) == NULL_TREE)
718 1676 : != (lookup_attribute ("no_profile_instrument_function",
719 3352 : DECL_ATTRIBUTES (callee->decl)) == NULL_TREE)))
720 : return false;
721 :
722 4409963 : if (!can_inline_edge_p (e, true, true)
723 4409963 : || !can_inline_edge_by_limits_p (e, CAN_INLINE_EARLY | CAN_INLINE_REPORT))
724 : return false;
725 : /* When inlining regular functions into always-inline functions
726 : during early inlining watch for possible inline cycles. */
727 4321687 : if (DECL_DISREGARD_INLINE_LIMITS (caller->decl)
728 275131 : && lookup_attribute ("always_inline", DECL_ATTRIBUTES (caller->decl))
729 4595959 : && (!DECL_DISREGARD_INLINE_LIMITS (callee->decl)
730 165402 : || !lookup_attribute ("always_inline", DECL_ATTRIBUTES (callee->decl))))
731 : {
732 : /* If there are indirect calls, inlining may produce direct call.
733 : TODO: We may lift this restriction if we avoid errors on formerly
734 : indirect calls to always_inline functions. Taking address
735 : of always_inline function is generally bad idea and should
736 : have been declared as undefined, but sadly we allow this. */
737 108871 : if (caller->indirect_calls || e->callee->indirect_calls)
738 : return false;
739 107618 : ipa_fn_summary *callee_info = ipa_fn_summaries->get (callee);
740 107618 : if (callee_info->safe_to_inline_to_always_inline)
741 27205 : return callee_info->safe_to_inline_to_always_inline - 1;
742 179801 : for (cgraph_edge *e2 = callee->callees; e2; e2 = e2->next_callee)
743 : {
744 99414 : struct cgraph_node *callee2 = e2->callee->ultimate_alias_target ();
745 : /* As early inliner runs in RPO order, we will see uninlined
746 : always_inline calls only in the case of cyclic graphs. */
747 99414 : if (DECL_DISREGARD_INLINE_LIMITS (callee2->decl)
748 99414 : || lookup_attribute ("always_inline", DECL_ATTRIBUTES (callee2->decl)))
749 : {
750 0 : callee_info->safe_to_inline_to_always_inline = 1;
751 0 : return false;
752 : }
753 : /* With LTO watch for case where function is later replaced
754 : by always_inline definition.
755 : TODO: We may either stop treating noninlined cross-module always
756 : inlines as errors, or we can extend decl merging to produce
757 : syntacic alias and honor always inline only in units it has
758 : been declared as such. */
759 99414 : if (flag_lto && callee2->externally_visible)
760 : {
761 26 : callee_info->safe_to_inline_to_always_inline = 1;
762 26 : return false;
763 : }
764 : }
765 80387 : callee_info->safe_to_inline_to_always_inline = 2;
766 : }
767 : return true;
768 : }
769 :
770 :
771 : /* Return number of calls in N. Ignore cheap builtins. */
772 :
773 : static int
774 891355 : num_calls (struct cgraph_node *n)
775 : {
776 891355 : struct cgraph_edge *e;
777 891355 : int num = 0;
778 :
779 1803842 : for (e = n->callees; e; e = e->next_callee)
780 912487 : if (!is_inexpensive_builtin (e->callee->decl))
781 818948 : num++;
782 891355 : return num;
783 : }
784 :
785 :
786 : /* Return true if we are interested in inlining small function. */
787 :
788 : static bool
789 3676293 : want_early_inline_function_p (struct cgraph_edge *e)
790 : {
791 3676293 : bool want_inline = true;
792 3676293 : struct cgraph_node *callee = e->callee->ultimate_alias_target ();
793 :
794 3676293 : if (DECL_DISREGARD_INLINE_LIMITS (callee->decl))
795 : ;
796 3676281 : else if (!DECL_DECLARED_INLINE_P (callee->decl)
797 3676281 : && !opt_for_fn (e->caller->decl, flag_inline_small_functions))
798 : {
799 62 : e->inline_failed = CIF_FUNCTION_NOT_INLINE_CANDIDATE;
800 62 : report_inline_failed_reason (e);
801 62 : want_inline = false;
802 : }
803 : else
804 : {
805 : /* First take care of very large functions. */
806 3676219 : int min_growth = estimate_min_edge_growth (e), growth = 0;
807 3676219 : int n;
808 3676219 : int early_inlining_insns = param_early_inlining_insns;
809 :
810 3676219 : if (min_growth > early_inlining_insns)
811 : {
812 427014 : if (dump_enabled_p ())
813 40 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
814 : " will not early inline: %C->%C, "
815 : "call is cold and code would grow "
816 : "at least by %i\n",
817 : e->caller, callee,
818 : min_growth);
819 : want_inline = false;
820 : }
821 : else
822 3249205 : growth = estimate_edge_growth (e);
823 :
824 :
825 3249245 : if (!want_inline || growth <= param_max_inline_insns_size)
826 : ;
827 1299864 : else if (!e->maybe_hot_p ())
828 : {
829 21153 : if (dump_enabled_p ())
830 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
831 : " will not early inline: %C->%C, "
832 : "call is cold and code would grow by %i\n",
833 : e->caller, callee,
834 : growth);
835 : want_inline = false;
836 : }
837 1278711 : else if (growth > early_inlining_insns)
838 : {
839 387356 : if (dump_enabled_p ())
840 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
841 : " will not early inline: %C->%C, "
842 : "growth %i exceeds --param early-inlining-insns\n",
843 : e->caller, callee, growth);
844 : want_inline = false;
845 : }
846 891355 : else if ((n = num_calls (callee)) != 0
847 891355 : && growth * (n + 1) > early_inlining_insns)
848 : {
849 263865 : if (dump_enabled_p ())
850 11 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
851 : " will not early inline: %C->%C, "
852 : "growth %i exceeds --param early-inlining-insns "
853 : "divided by number of calls\n",
854 : e->caller, callee, growth);
855 : want_inline = false;
856 : }
857 : }
858 3676293 : return want_inline;
859 : }
860 :
861 : /* Compute time of the edge->caller + edge->callee execution when inlining
862 : does not happen. */
863 :
864 : inline sreal
865 602676 : compute_uninlined_call_time (struct cgraph_edge *edge,
866 : sreal uninlined_call_time,
867 : sreal freq)
868 : {
869 443829 : cgraph_node *caller = (edge->caller->inlined_to
870 602676 : ? edge->caller->inlined_to
871 : : edge->caller);
872 :
873 602676 : if (freq > 0)
874 586370 : uninlined_call_time *= freq;
875 : else
876 16306 : uninlined_call_time = uninlined_call_time >> 11;
877 :
878 602676 : sreal caller_time = ipa_fn_summaries->get (caller)->time;
879 602676 : return uninlined_call_time + caller_time;
880 : }
881 :
882 : /* Same as compute_uinlined_call_time but compute time when inlining
883 : does happen. */
884 :
885 : inline sreal
886 602676 : compute_inlined_call_time (struct cgraph_edge *edge,
887 : sreal time,
888 : sreal freq)
889 : {
890 443829 : cgraph_node *caller = (edge->caller->inlined_to
891 602676 : ? edge->caller->inlined_to
892 : : edge->caller);
893 602676 : sreal caller_time = ipa_fn_summaries->get (caller)->time;
894 :
895 602676 : if (freq > 0)
896 586370 : time *= freq;
897 : else
898 16306 : time = time >> 11;
899 :
900 : /* This calculation should match one in ipa-inline-analysis.cc
901 : (estimate_edge_size_and_time). */
902 602676 : time -= (sreal)ipa_call_summaries->get (edge)->call_stmt_time * freq;
903 602676 : time += caller_time;
904 602676 : if (time <= 0)
905 0 : time = ((sreal) 1) >> 8;
906 602676 : gcc_checking_assert (time >= 0);
907 602676 : return time;
908 : }
909 :
910 : /* Determine time saved by inlining EDGE of frequency FREQ
911 : where callee's runtime w/o inlining is UNINLINED_TYPE
912 : and with inlined is INLINED_TYPE. */
913 :
914 : inline sreal
915 10272207 : inlining_speedup (struct cgraph_edge *edge,
916 : sreal freq,
917 : sreal uninlined_time,
918 : sreal inlined_time)
919 : {
920 10272207 : sreal speedup = uninlined_time - inlined_time;
921 : /* Handling of call_time should match one in ipa-inline-fnsummary.c
922 : (estimate_edge_size_and_time). */
923 10272207 : sreal call_time = ipa_call_summaries->get (edge)->call_stmt_time;
924 :
925 10272207 : if (freq > 0)
926 : {
927 10230997 : speedup = (speedup + call_time);
928 12547717 : if (freq != 1)
929 7914277 : speedup = speedup * freq;
930 : }
931 41210 : else if (freq == 0)
932 41210 : speedup = speedup >> 11;
933 10272207 : gcc_checking_assert (speedup >= 0);
934 10272207 : return speedup;
935 : }
936 :
937 : /* Return expected speedup of the callee function alone
938 : (i.e. not estimate of call overhead and also no scalling
939 : by call frequency. */
940 :
941 : static sreal
942 3416834 : callee_speedup (struct cgraph_edge *e)
943 : {
944 3416834 : sreal unspec_time;
945 3416834 : sreal spec_time = estimate_edge_time (e, &unspec_time);
946 3416834 : return unspec_time - spec_time;
947 : }
948 :
949 : /* Return true if the speedup for inlining E is bigger than
950 : param_inline_min_speedup. */
951 :
952 : static bool
953 602676 : big_speedup_p (struct cgraph_edge *e)
954 : {
955 602676 : sreal unspec_time;
956 602676 : sreal spec_time = estimate_edge_time (e, &unspec_time);
957 602676 : sreal freq = e->sreal_frequency ();
958 602676 : sreal time = compute_uninlined_call_time (e, unspec_time, freq);
959 602676 : sreal inlined_time = compute_inlined_call_time (e, spec_time, freq);
960 443829 : cgraph_node *caller = (e->caller->inlined_to
961 602676 : ? e->caller->inlined_to
962 : : e->caller);
963 602676 : int limit = opt_for_fn (caller->decl, param_inline_min_speedup);
964 :
965 602676 : if ((time - inlined_time) * 100 > time * limit)
966 : return true;
967 : return false;
968 : }
969 :
970 : /* Return true if we are interested in inlining small function.
971 : When REPORT is true, report reason to dump file. */
972 :
973 : static bool
974 5551403 : want_inline_small_function_p (struct cgraph_edge *e, bool report)
975 : {
976 5551403 : bool want_inline = true;
977 5551403 : struct cgraph_node *callee = e->callee->ultimate_alias_target ();
978 4209539 : cgraph_node *to = (e->caller->inlined_to
979 5551403 : ? e->caller->inlined_to : e->caller);
980 :
981 : /* Allow this function to be called before can_inline_edge_p,
982 : since it's usually cheaper. */
983 5551403 : if (cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR)
984 : want_inline = false;
985 5551403 : else if (DECL_DISREGARD_INLINE_LIMITS (callee->decl))
986 : return true;
987 5545428 : else if (!DECL_DECLARED_INLINE_P (callee->decl)
988 5545428 : && !opt_for_fn (e->caller->decl, flag_inline_small_functions))
989 : {
990 52382 : e->inline_failed = CIF_FUNCTION_NOT_INLINE_CANDIDATE;
991 52382 : want_inline = false;
992 : }
993 :
994 : /* Early return before lookup of summaries. */
995 52382 : if (!want_inline)
996 : {
997 52382 : if (report)
998 48118 : report_inline_failed_reason (e);
999 : return false;
1000 : }
1001 :
1002 5493046 : ipa_fn_summary *callee_info = ipa_fn_summaries->get (callee);
1003 5493046 : ipa_call_summary *call_info = ipa_call_summaries->get (e);
1004 :
1005 : /* Do fast and conservative check if the function can be good
1006 : inline candidate. */
1007 5493046 : if ((!DECL_DECLARED_INLINE_P (callee->decl)
1008 2034856 : && (!e->count.ipa ().initialized_p ()
1009 42418 : || !e->maybe_hot_p (callee_info->time)))
1010 7527655 : && callee_info->min_size - call_info->call_stmt_size
1011 2034609 : > inline_insns_auto (e->caller, true, true))
1012 : {
1013 107 : e->inline_failed = CIF_MAX_INLINE_INSNS_AUTO_LIMIT;
1014 107 : want_inline = false;
1015 : }
1016 5492939 : else if ((DECL_DECLARED_INLINE_P (callee->decl)
1017 2034749 : || e->count.ipa ().nonzero_p ())
1018 8951400 : && callee_info->min_size - call_info->call_stmt_size
1019 3458461 : > inline_insns_single (e->caller, true, true))
1020 : {
1021 0 : e->inline_failed = (DECL_DECLARED_INLINE_P (callee->decl)
1022 0 : ? CIF_MAX_INLINE_INSNS_SINGLE_LIMIT
1023 : : CIF_MAX_INLINE_INSNS_AUTO_LIMIT);
1024 0 : want_inline = false;
1025 : }
1026 : else
1027 : {
1028 5492939 : int growth = estimate_edge_growth (e);
1029 5492939 : ipa_hints hints = estimate_edge_hints (e);
1030 : /* We have two independent groups of hints. If one matches in each
1031 : of groups the limits are inreased. If both groups matches, limit
1032 : is increased even more. */
1033 5492939 : bool apply_hints = (hints & (INLINE_HINT_indirect_call
1034 : | INLINE_HINT_known_hot
1035 : | INLINE_HINT_loop_iterations
1036 : | INLINE_HINT_loop_stride));
1037 5492939 : bool apply_hints2 = (hints & INLINE_HINT_builtin_constant_p);
1038 :
1039 5492939 : if (growth <= opt_for_fn (to->decl,
1040 : param_max_inline_insns_size))
1041 : ;
1042 : /* Apply param_max_inline_insns_single limit. Do not do so when
1043 : hints suggests that inlining given function is very profitable.
1044 : Avoid computation of big_speedup_p when not necessary to change
1045 : outcome of decision. */
1046 5354926 : else if (DECL_DECLARED_INLINE_P (callee->decl)
1047 3396607 : && growth >= inline_insns_single (e->caller, apply_hints,
1048 : apply_hints2)
1049 6076378 : && (apply_hints || apply_hints2
1050 719809 : || growth >= inline_insns_single (e->caller, true,
1051 : apply_hints2)
1052 382634 : || !big_speedup_p (e)))
1053 : {
1054 720141 : e->inline_failed = CIF_MAX_INLINE_INSNS_SINGLE_LIMIT;
1055 720141 : want_inline = false;
1056 : }
1057 4634785 : else if (!DECL_DECLARED_INLINE_P (callee->decl)
1058 1958319 : && !opt_for_fn (e->caller->decl, flag_inline_functions)
1059 4640136 : && growth >= opt_for_fn (to->decl,
1060 : param_max_inline_insns_small))
1061 : {
1062 : /* growth_positive_p is expensive, always test it last. */
1063 5351 : if (growth >= inline_insns_single (e->caller, false, false)
1064 5351 : || growth_positive_p (callee, e, growth))
1065 : {
1066 4964 : e->inline_failed = CIF_NOT_DECLARED_INLINED;
1067 4964 : want_inline = false;
1068 : }
1069 : }
1070 : /* Apply param_max_inline_insns_auto limit for functions not declared
1071 : inline. Bypass the limit when speedup seems big. */
1072 4629434 : else if (!DECL_DECLARED_INLINE_P (callee->decl)
1073 1952968 : && growth >= inline_insns_auto (e->caller, apply_hints,
1074 : apply_hints2)
1075 5853072 : && (apply_hints || apply_hints2
1076 1208098 : || growth >= inline_insns_auto (e->caller, true,
1077 : apply_hints2)
1078 219857 : || !big_speedup_p (e)))
1079 : {
1080 : /* growth_positive_p is expensive, always test it last. */
1081 1212600 : if (growth >= inline_insns_single (e->caller, false, false)
1082 1212600 : || growth_positive_p (callee, e, growth))
1083 : {
1084 1128822 : e->inline_failed = CIF_MAX_INLINE_INSNS_AUTO_LIMIT;
1085 1128822 : want_inline = false;
1086 : }
1087 : }
1088 : /* If call is cold, do not inline when function body would grow. */
1089 3416834 : else if (!e->maybe_hot_p (callee_speedup (e))
1090 3416834 : && (growth >= inline_insns_single (e->caller, false, false)
1091 773690 : || growth_positive_p (callee, e, growth)))
1092 : {
1093 694675 : e->inline_failed = CIF_UNLIKELY_CALL;
1094 694675 : want_inline = false;
1095 : }
1096 : }
1097 5493046 : if (!want_inline && report)
1098 597046 : report_inline_failed_reason (e);
1099 : return want_inline;
1100 : }
1101 :
1102 : /* EDGE is self recursive edge.
1103 : We handle two cases - when function A is inlining into itself
1104 : or when function A is being inlined into another inliner copy of function
1105 : A within function B.
1106 :
1107 : In first case OUTER_NODE points to the toplevel copy of A, while
1108 : in the second case OUTER_NODE points to the outermost copy of A in B.
1109 :
1110 : In both cases we want to be extra selective since
1111 : inlining the call will just introduce new recursive calls to appear. */
1112 :
1113 : static bool
1114 19832 : want_inline_self_recursive_call_p (struct cgraph_edge *edge,
1115 : struct cgraph_node *outer_node,
1116 : bool peeling,
1117 : int depth)
1118 : {
1119 19832 : char const *reason = NULL;
1120 19832 : bool want_inline = true;
1121 19832 : sreal caller_freq = 1;
1122 19832 : int max_depth = opt_for_fn (outer_node->decl,
1123 : param_max_inline_recursive_depth_auto);
1124 :
1125 19832 : if (DECL_DECLARED_INLINE_P (edge->caller->decl))
1126 2930 : max_depth = opt_for_fn (outer_node->decl,
1127 : param_max_inline_recursive_depth);
1128 :
1129 19832 : if (!edge->maybe_hot_p ())
1130 : {
1131 : reason = "recursive call is cold";
1132 : want_inline = false;
1133 : }
1134 19346 : else if (depth > max_depth)
1135 : {
1136 : reason = "--param max-inline-recursive-depth exceeded.";
1137 : want_inline = false;
1138 : }
1139 17144 : else if (outer_node->inlined_to
1140 21595 : && (caller_freq = outer_node->callers->sreal_frequency ()) == 0)
1141 : {
1142 : reason = "caller frequency is 0";
1143 : want_inline = false;
1144 : }
1145 :
1146 17144 : if (!want_inline)
1147 : ;
1148 : /* Inlining of self recursive function into copy of itself within other
1149 : function is transformation similar to loop peeling.
1150 :
1151 : Peeling is profitable if we can inline enough copies to make probability
1152 : of actual call to the self recursive function very small. Be sure that
1153 : the probability of recursion is small.
1154 :
1155 : We ensure that the frequency of recursing is at most 1 - (1/max_depth).
1156 : This way the expected number of recursion is at most max_depth. */
1157 17144 : else if (peeling)
1158 : {
1159 4451 : sreal max_prob = (sreal)1 - ((sreal)1 / (sreal)max_depth);
1160 4451 : int i;
1161 9865 : for (i = 1; i < depth; i++)
1162 5414 : max_prob = max_prob * max_prob;
1163 4451 : if (edge->sreal_frequency () >= max_prob * caller_freq)
1164 : {
1165 1593 : reason = "frequency of recursive call is too large";
1166 1593 : want_inline = false;
1167 : }
1168 : }
1169 : /* Recursive inlining, i.e. equivalent of unrolling, is profitable if
1170 : recursion depth is large. We reduce function call overhead and increase
1171 : chances that things fit in hardware return predictor.
1172 :
1173 : Recursive inlining might however increase cost of stack frame setup
1174 : actually slowing down functions whose recursion tree is wide rather than
1175 : deep.
1176 :
1177 : Deciding reliably on when to do recursive inlining without profile feedback
1178 : is tricky. For now we disable recursive inlining when probability of self
1179 : recursion is low.
1180 :
1181 : Recursive inlining of self recursive call within loop also results in
1182 : large loop depths that generally optimize badly. We may want to throttle
1183 : down inlining in those cases. In particular this seems to happen in one
1184 : of libstdc++ rb tree methods. */
1185 : else
1186 : {
1187 12693 : if (edge->sreal_frequency () * 100
1188 12693 : <= caller_freq
1189 25386 : * opt_for_fn (outer_node->decl,
1190 : param_min_inline_recursive_probability))
1191 : {
1192 800 : reason = "frequency of recursive call is too small";
1193 800 : want_inline = false;
1194 : }
1195 : }
1196 19832 : if (!can_inline_edge_by_limits_p (edge, CAN_INLINE_FORCE_LIMITS | CAN_INLINE_REPORT))
1197 : {
1198 : reason = "inline limits exceeded for always_inline function";
1199 : want_inline = false;
1200 : }
1201 19832 : if (!want_inline && dump_enabled_p ())
1202 7 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, edge->call_stmt,
1203 : " not inlining recursively: %s\n", reason);
1204 19832 : return want_inline;
1205 : }
1206 :
1207 : /* Return true when NODE has uninlinable caller;
1208 : set HAS_HOT_CALL if it has hot call.
1209 : Worker for cgraph_for_node_and_aliases. */
1210 :
1211 : static bool
1212 76803 : check_callers (struct cgraph_node *node, void *has_hot_call)
1213 : {
1214 76803 : struct cgraph_edge *e;
1215 133769 : for (e = node->callers; e; e = e->next_caller)
1216 : {
1217 84902 : if (!opt_for_fn (e->caller->decl, flag_inline_functions_called_once)
1218 84902 : || !opt_for_fn (e->caller->decl, optimize))
1219 : return true;
1220 84902 : if (!can_inline_edge_p (e, true))
1221 : return true;
1222 84875 : if (e->recursive_p ())
1223 : return true;
1224 84875 : if (!can_inline_edge_by_limits_p (e, CAN_INLINE_REPORT))
1225 : return true;
1226 : /* Inlining large functions to large loop depth is often harmful because
1227 : of register pressure it implies. */
1228 57038 : if ((int)ipa_call_summaries->get (e)->loop_depth
1229 57038 : > param_inline_functions_called_once_loop_depth)
1230 : return true;
1231 : /* Do not produce gigantic functions. */
1232 104229 : if (estimate_size_after_inlining (e->caller->inlined_to ?
1233 : e->caller->inlined_to : e->caller, e)
1234 57038 : > param_inline_functions_called_once_insns)
1235 : return true;
1236 56966 : if (!(*(bool *)has_hot_call) && e->maybe_hot_p ())
1237 12119 : *(bool *)has_hot_call = true;
1238 : }
1239 : return false;
1240 : }
1241 :
1242 : /* If NODE has a caller, return true. */
1243 :
1244 : static bool
1245 2204822 : has_caller_p (struct cgraph_node *node, void *data ATTRIBUTE_UNUSED)
1246 : {
1247 2204822 : if (node->callers)
1248 693414 : return true;
1249 : return false;
1250 : }
1251 :
1252 : /* Decide if inlining NODE would reduce unit size by eliminating
1253 : the offline copy of function.
1254 : When COLD is true the cold calls are considered, too. */
1255 :
1256 : static bool
1257 5132201 : want_inline_function_to_all_callers_p (struct cgraph_node *node, bool cold)
1258 : {
1259 5132201 : bool has_hot_call = false;
1260 :
1261 : /* Aliases gets inlined along with the function they alias. */
1262 5132201 : if (node->alias)
1263 : return false;
1264 : /* Already inlined? */
1265 5050920 : if (node->inlined_to)
1266 : return false;
1267 : /* Does it have callers? */
1268 2152435 : if (!node->call_for_symbol_and_aliases (has_caller_p, NULL, true))
1269 : return false;
1270 : /* Inlining into all callers would increase size? */
1271 693414 : if (growth_positive_p (node, NULL, INT_MIN) > 0)
1272 : return false;
1273 : /* All inlines must be possible. */
1274 71816 : if (node->call_for_symbol_and_aliases (check_callers, &has_hot_call,
1275 : true))
1276 : return false;
1277 43880 : if (!cold && !has_hot_call)
1278 : return false;
1279 : return true;
1280 : }
1281 :
1282 : /* Return true if WHERE of SIZE is a possible candidate for wrapper heuristics
1283 : in estimate_edge_badness. */
1284 :
1285 : static bool
1286 672501 : wrapper_heuristics_may_apply (struct cgraph_node *where, int size)
1287 : {
1288 672501 : return size < (DECL_DECLARED_INLINE_P (where->decl)
1289 672501 : ? inline_insns_single (where, false, false)
1290 383877 : : inline_insns_auto (where, false, false));
1291 : }
1292 :
1293 : /* A cost model driving the inlining heuristics in a way so the edges with
1294 : smallest badness are inlined first. After each inlining is performed
1295 : the costs of all caller edges of nodes affected are recomputed so the
1296 : metrics may accurately depend on values such as number of inlinable callers
1297 : of the function or function body size. */
1298 :
1299 : static sreal
1300 10846476 : edge_badness (struct cgraph_edge *edge, bool dump)
1301 : {
1302 10846476 : sreal badness;
1303 10846476 : int growth;
1304 10846476 : sreal edge_time, unspec_edge_time;
1305 10846476 : struct cgraph_node *callee = edge->callee->ultimate_alias_target ();
1306 10846476 : class ipa_fn_summary *callee_info = ipa_fn_summaries->get (callee);
1307 10846476 : ipa_hints hints;
1308 8413745 : cgraph_node *caller = (edge->caller->inlined_to
1309 10846476 : ? edge->caller->inlined_to
1310 : : edge->caller);
1311 :
1312 10846476 : growth = estimate_edge_growth (edge);
1313 10846476 : edge_time = estimate_edge_time (edge, &unspec_edge_time);
1314 10846476 : hints = estimate_edge_hints (edge);
1315 10846476 : gcc_checking_assert (edge_time >= 0);
1316 : /* Check that inlined time is better, but tolerate some roundoff issues.
1317 : FIXME: When callee profile drops to 0 we account calls more. This
1318 : should be fixed by never doing that. */
1319 10846476 : gcc_checking_assert ((edge_time * 100
1320 : - callee_info->time * 101).to_int () <= 0
1321 : || callee->count.ipa ().initialized_p ());
1322 10846476 : gcc_checking_assert (growth <= ipa_size_summaries->get (callee)->size);
1323 :
1324 10846476 : if (dump)
1325 : {
1326 185 : fprintf (dump_file, " Badness calculation for %s -> %s\n",
1327 185 : edge->caller->dump_name (),
1328 185 : edge->callee->dump_name ());
1329 185 : fprintf (dump_file, " size growth %i, time %f unspec %f ",
1330 : growth,
1331 : edge_time.to_double (),
1332 : unspec_edge_time.to_double ());
1333 185 : ipa_dump_hints (dump_file, hints);
1334 185 : if (big_speedup_p (edge))
1335 151 : fprintf (dump_file, " big_speedup");
1336 185 : fprintf (dump_file, "\n");
1337 : }
1338 :
1339 : /* Always prefer inlining saving code size. */
1340 10846476 : if (growth <= 0)
1341 : {
1342 525084 : badness = (sreal) (-SREAL_MIN_SIG + growth) << (SREAL_MAX_EXP / 256);
1343 525084 : if (dump)
1344 108 : fprintf (dump_file, " %f: Growth %d <= 0\n", badness.to_double (),
1345 : growth);
1346 : }
1347 : /* Inlining into EXTERNAL functions is not going to change anything unless
1348 : they are themselves inlined. */
1349 10321392 : else if (DECL_EXTERNAL (caller->decl))
1350 : {
1351 47910 : if (dump)
1352 0 : fprintf (dump_file, " max: function is external\n");
1353 47910 : return sreal::max ();
1354 : }
1355 : /* When profile is available. Compute badness as:
1356 :
1357 : time_saved * caller_count
1358 : goodness = -------------------------------------------------
1359 : growth_of_caller * overall_growth * combined_size
1360 :
1361 : badness = - goodness
1362 :
1363 : Again use negative value to make calls with profile appear hotter
1364 : then calls without.
1365 : */
1366 10273482 : else if (opt_for_fn (caller->decl, flag_guess_branch_prob)
1367 10273482 : || caller->count.ipa ().nonzero_p ())
1368 : {
1369 10272130 : sreal numerator, denominator;
1370 10272130 : int overall_growth;
1371 10272130 : sreal freq = edge->sreal_frequency ();
1372 :
1373 10272130 : numerator = inlining_speedup (edge, freq, unspec_edge_time, edge_time);
1374 10272130 : if (numerator <= 0)
1375 8798 : numerator = ((sreal) 1 >> 8);
1376 10272130 : if (caller->count.ipa ().nonzero_p ())
1377 200 : numerator *= caller->count.ipa ().to_gcov_type ();
1378 10271930 : else if (caller->count.ipa ().initialized_p ())
1379 918 : numerator = numerator >> 11;
1380 10272130 : denominator = growth;
1381 :
1382 10272130 : overall_growth = callee_info->growth;
1383 :
1384 : /* Look for inliner wrappers of the form:
1385 :
1386 : inline_caller ()
1387 : {
1388 : do_fast_job...
1389 : if (need_more_work)
1390 : noninline_callee ();
1391 : }
1392 : Without penalizing this case, we usually inline noninline_callee
1393 : into the inline_caller because overall_growth is small preventing
1394 : further inlining of inline_caller.
1395 :
1396 : Penalize only callgraph edges to functions with small overall
1397 : growth ...
1398 : */
1399 10272130 : if (growth > overall_growth
1400 : /* ... and having only one caller which is not inlined ... */
1401 2169053 : && callee_info->single_caller
1402 1213743 : && !edge->caller->inlined_to
1403 : /* ... and edges executed only conditionally ... */
1404 840230 : && freq < 1
1405 : /* ... consider case where callee is not inline but caller is ... */
1406 10632344 : && ((!DECL_DECLARED_INLINE_P (edge->callee->decl)
1407 99314 : && DECL_DECLARED_INLINE_P (caller->decl))
1408 : /* ... or when early optimizers decided to split and edge
1409 : frequency still indicates splitting is a win ... */
1410 351930 : || (callee->split_part && !caller->split_part
1411 95777 : && freq * 100
1412 10463272 : < opt_for_fn (caller->decl,
1413 : param_partial_inlining_entry_probability)
1414 : /* ... and do not overwrite user specified hints. */
1415 95365 : && (!DECL_DECLARED_INLINE_P (edge->callee->decl)
1416 70768 : || DECL_DECLARED_INLINE_P (caller->decl)))))
1417 : {
1418 102371 : ipa_fn_summary *caller_info = ipa_fn_summaries->get (caller);
1419 102371 : int caller_growth = caller_info->growth;
1420 :
1421 : /* Only apply the penalty when caller looks like inline candidate,
1422 : and it is not called once. */
1423 54225 : if (!caller_info->single_caller && overall_growth < caller_growth
1424 51864 : && caller_info->inlinable
1425 154187 : && wrapper_heuristics_may_apply
1426 51816 : (caller, ipa_size_summaries->get (caller)->size))
1427 : {
1428 41672 : if (dump)
1429 1 : fprintf (dump_file,
1430 : " Wrapper penalty. Increasing growth %i to %i\n",
1431 : overall_growth, caller_growth);
1432 : overall_growth = caller_growth;
1433 : }
1434 : }
1435 10272130 : if (overall_growth > 0)
1436 : {
1437 : /* Strongly prefer functions with few callers that can be inlined
1438 : fully. The square root here leads to smaller binaries at average.
1439 : Watch however for extreme cases and return to linear function
1440 : when growth is large. */
1441 8827285 : if (overall_growth < 256)
1442 4534854 : overall_growth *= overall_growth;
1443 : else
1444 4292431 : overall_growth += 256 * 256 - 256;
1445 8827285 : denominator *= overall_growth;
1446 : }
1447 10272130 : denominator *= ipa_size_summaries->get (caller)->size + growth;
1448 :
1449 10272130 : badness = - numerator / denominator;
1450 :
1451 10272130 : if (dump)
1452 : {
1453 308 : fprintf (dump_file,
1454 : " %f: guessed profile. frequency %f, count %" PRId64
1455 : " caller count %" PRId64
1456 : " time saved %f"
1457 : " overall growth %i (current) %i (original)"
1458 : " %i (compensated)\n",
1459 : badness.to_double (),
1460 : freq.to_double (),
1461 77 : edge->count.ipa ().initialized_p ()
1462 0 : ? edge->count.ipa ().to_gcov_type () : -1,
1463 77 : caller->count.ipa ().initialized_p ()
1464 0 : ? caller->count.ipa ().to_gcov_type () : -1,
1465 154 : inlining_speedup (edge, freq, unspec_edge_time,
1466 : edge_time).to_double (),
1467 : estimate_growth (callee),
1468 : callee_info->growth, overall_growth);
1469 : }
1470 : }
1471 : /* When function local profile is not available or it does not give
1472 : useful information (i.e. frequency is zero), base the cost on
1473 : loop nest and overall size growth, so we optimize for overall number
1474 : of functions fully inlined in program. */
1475 : else
1476 : {
1477 1352 : int nest = MIN (ipa_call_summaries->get (edge)->loop_depth, 8);
1478 1352 : badness = growth;
1479 :
1480 : /* Decrease badness if call is nested. */
1481 1352 : if (badness > 0)
1482 1352 : badness = badness >> nest;
1483 : else
1484 0 : badness = badness << nest;
1485 1352 : if (dump)
1486 0 : fprintf (dump_file, " %f: no profile. nest %i\n",
1487 : badness.to_double (), nest);
1488 : }
1489 10798566 : gcc_checking_assert (badness != 0);
1490 :
1491 10798566 : if (edge->recursive_p ())
1492 18613 : badness = badness.shift (badness > 0 ? 4 : -4);
1493 10798566 : if ((hints & (INLINE_HINT_indirect_call
1494 : | INLINE_HINT_loop_iterations
1495 : | INLINE_HINT_loop_stride))
1496 9977271 : || callee_info->growth <= 0)
1497 4887705 : badness = badness.shift (badness > 0 ? -2 : 2);
1498 10798566 : if (hints & INLINE_HINT_builtin_constant_p)
1499 23734 : badness = badness.shift (badness > 0 ? -4 : 4);
1500 10798566 : if (hints & (INLINE_HINT_same_scc))
1501 70492 : badness = badness.shift (badness > 0 ? 3 : -3);
1502 10763319 : else if (hints & (INLINE_HINT_in_scc))
1503 174652 : badness = badness.shift (badness > 0 ? 2 : -2);
1504 10675993 : else if (hints & (INLINE_HINT_cross_module))
1505 3898 : badness = badness.shift (badness > 0 ? 1 : -1);
1506 10798566 : if (DECL_DISREGARD_INLINE_LIMITS (callee->decl))
1507 17896 : badness = badness.shift (badness > 0 ? -4 : 4);
1508 10789618 : else if ((hints & INLINE_HINT_declared_inline))
1509 16630461 : badness = badness.shift (badness > 0 ? -3 : 3);
1510 10798566 : if (dump)
1511 185 : fprintf (dump_file, " Adjusted by hints %f\n", badness.to_double ());
1512 10798566 : return badness;
1513 : }
1514 :
1515 : /* Recompute badness of EDGE and update its key in HEAP if needed. */
1516 : static inline void
1517 5811501 : update_edge_key (edge_heap_t *heap, struct cgraph_edge *edge)
1518 : {
1519 5811501 : sreal badness = edge_badness (edge, false);
1520 5811501 : if (edge->aux)
1521 : {
1522 4654018 : edge_heap_node_t *n = (edge_heap_node_t *) edge->aux;
1523 4654018 : gcc_checking_assert (n->get_data () == edge);
1524 :
1525 : /* fibonacci_heap::replace_key does busy updating of the
1526 : heap that is unnecessarily expensive.
1527 : We do lazy increases: after extracting minimum if the key
1528 : turns out to be out of date, it is re-inserted into heap
1529 : with correct value. */
1530 4654018 : if (badness < n->get_key ().badness)
1531 : {
1532 931015 : if (dump_file && (dump_flags & TDF_DETAILS))
1533 : {
1534 90 : fprintf (dump_file,
1535 : " decreasing badness %s -> %s, %f to %f\n",
1536 45 : edge->caller->dump_name (),
1537 45 : edge->callee->dump_name (),
1538 90 : n->get_key ().badness.to_double (),
1539 : badness.to_double ());
1540 : }
1541 931015 : inline_badness b (edge, badness);
1542 931015 : heap->decrease_key (n, b);
1543 : }
1544 : }
1545 : else
1546 : {
1547 1157483 : if (dump_file && (dump_flags & TDF_DETAILS))
1548 : {
1549 338 : fprintf (dump_file,
1550 : " enqueuing call %s -> %s, badness %f\n",
1551 169 : edge->caller->dump_name (),
1552 169 : edge->callee->dump_name (),
1553 : badness.to_double ());
1554 : }
1555 1157483 : inline_badness b (edge, badness);
1556 1157483 : edge->aux = heap->insert (b, edge);
1557 : }
1558 5811501 : }
1559 :
1560 :
1561 : /* NODE was inlined.
1562 : All caller edges needs to be reset because
1563 : size estimates change. Similarly callees needs reset
1564 : because better context may be known. */
1565 :
1566 : static void
1567 1009160 : reset_edge_caches (struct cgraph_node *node)
1568 : {
1569 1009160 : struct cgraph_edge *edge;
1570 1009160 : struct cgraph_edge *e = node->callees;
1571 1009160 : struct cgraph_node *where = node;
1572 1009160 : struct ipa_ref *ref;
1573 :
1574 1009160 : if (where->inlined_to)
1575 941996 : where = where->inlined_to;
1576 :
1577 1009160 : reset_node_cache (where);
1578 :
1579 1009160 : if (edge_growth_cache != NULL)
1580 3534474 : for (edge = where->callers; edge; edge = edge->next_caller)
1581 2527654 : if (edge->inline_failed)
1582 2527654 : edge_growth_cache->remove (edge);
1583 :
1584 1071193 : FOR_EACH_ALIAS (where, ref)
1585 124066 : reset_edge_caches (dyn_cast <cgraph_node *> (ref->referring));
1586 :
1587 1009160 : if (!e)
1588 : return;
1589 :
1590 3283157 : while (true)
1591 3283157 : if (!e->inline_failed && e->callee->callees)
1592 : e = e->callee->callees;
1593 : else
1594 : {
1595 2602032 : if (edge_growth_cache != NULL && e->inline_failed)
1596 2350021 : edge_growth_cache->remove (e);
1597 2602032 : if (e->next_callee)
1598 : e = e->next_callee;
1599 : else
1600 : {
1601 1528321 : do
1602 : {
1603 1528321 : if (e->caller == node)
1604 : return;
1605 681125 : e = e->caller->callers;
1606 : }
1607 681125 : while (!e->next_callee);
1608 : e = e->next_callee;
1609 : }
1610 : }
1611 : }
1612 :
1613 : /* Recompute HEAP nodes for each of caller of NODE.
1614 : UPDATED_NODES track nodes we already visited, to avoid redundant work.
1615 : When CHECK_INLINABLITY_FOR is set, re-check for specified edge that
1616 : it is inlinable. Otherwise check all edges. */
1617 :
1618 : static void
1619 1006515 : update_caller_keys (edge_heap_t *heap, struct cgraph_node *node,
1620 : bitmap updated_nodes,
1621 : struct cgraph_edge *check_inlinablity_for)
1622 : {
1623 1006515 : struct cgraph_edge *edge;
1624 1006515 : struct ipa_ref *ref;
1625 :
1626 1006515 : if ((!node->alias && !ipa_fn_summaries->get (node)->inlinable)
1627 994282 : || node->inlined_to)
1628 12233 : return;
1629 994282 : if (!bitmap_set_bit (updated_nodes, node->get_summary_id ()))
1630 : return;
1631 :
1632 1055815 : FOR_EACH_ALIAS (node, ref)
1633 : {
1634 61533 : struct cgraph_node *alias = dyn_cast <cgraph_node *> (ref->referring);
1635 61533 : update_caller_keys (heap, alias, updated_nodes, check_inlinablity_for);
1636 : }
1637 :
1638 3415566 : for (edge = node->callers; edge; edge = edge->next_caller)
1639 2421284 : if (edge->inline_failed)
1640 : {
1641 2421284 : if (!check_inlinablity_for
1642 2421284 : || check_inlinablity_for == edge)
1643 : {
1644 2421284 : if (can_inline_edge_p (edge, false)
1645 2387428 : && want_inline_small_function_p (edge, false)
1646 3069284 : && can_inline_edge_by_limits_p (edge, 0))
1647 642284 : update_edge_key (heap, edge);
1648 1779000 : else if (edge->aux)
1649 : {
1650 107562 : report_inline_failed_reason (edge);
1651 107562 : heap->delete_node ((edge_heap_node_t *) edge->aux);
1652 107562 : edge->aux = NULL;
1653 : }
1654 : }
1655 0 : else if (edge->aux)
1656 0 : update_edge_key (heap, edge);
1657 : }
1658 : }
1659 :
1660 : /* Recompute HEAP nodes for each uninlined call in NODE
1661 : If UPDATE_SINCE is non-NULL check if edges called within that function
1662 : are inlinable (typically UPDATE_SINCE is the inline clone we introduced
1663 : where all edges have new context).
1664 :
1665 : This is used when we know that edge badnesses are going only to increase
1666 : (we introduced new call site) and thus all we need is to insert newly
1667 : created edges into heap. */
1668 :
1669 : static void
1670 945055 : update_callee_keys (edge_heap_t *heap, struct cgraph_node *node,
1671 : struct cgraph_node *update_since,
1672 : bitmap updated_nodes)
1673 : {
1674 945055 : struct cgraph_edge *e = node->callees;
1675 945055 : bool check_inlinability = update_since == node;
1676 :
1677 945055 : if (!e)
1678 : return;
1679 31002344 : while (true)
1680 31002344 : if (!e->inline_failed && e->callee->callees)
1681 : {
1682 5745734 : if (e->callee == update_since)
1683 439915 : check_inlinability = true;
1684 30090129 : e = e->callee->callees;
1685 : }
1686 : else
1687 : {
1688 25256610 : enum availability avail;
1689 25256610 : struct cgraph_node *callee;
1690 25256610 : if (!check_inlinability)
1691 : {
1692 22846667 : if (e->aux
1693 26668886 : && !bitmap_bit_p (updated_nodes,
1694 3822219 : e->callee->ultimate_alias_target
1695 3822219 : (&avail, e->caller)->get_summary_id ()))
1696 3822212 : update_edge_key (heap, e);
1697 : }
1698 : /* We do not reset callee growth cache here. Since we added a new call,
1699 : growth should have just increased and consequently badness metric
1700 : don't need updating. */
1701 2409943 : else if (e->inline_failed
1702 2309917 : && (callee = e->callee->ultimate_alias_target (&avail,
1703 2309917 : e->caller))
1704 2309917 : && avail >= AVAIL_AVAILABLE
1705 609084 : && ipa_fn_summaries->get (callee) != NULL
1706 609068 : && ipa_fn_summaries->get (callee)->inlinable
1707 3004747 : && !bitmap_bit_p (updated_nodes, callee->get_summary_id ()))
1708 : {
1709 594804 : if (can_inline_edge_p (e, false)
1710 589494 : && want_inline_small_function_p (e, false)
1711 967799 : && can_inline_edge_by_limits_p (e, 0))
1712 : {
1713 372234 : gcc_checking_assert (check_inlinability || can_inline_edge_p (e, false));
1714 : gcc_checking_assert (check_inlinability || e->aux);
1715 372234 : update_edge_key (heap, e);
1716 : }
1717 222570 : else if (e->aux)
1718 : {
1719 7478 : report_inline_failed_reason (e);
1720 7478 : heap->delete_node ((edge_heap_node_t *) e->aux);
1721 7478 : e->aux = NULL;
1722 : }
1723 : }
1724 : /* In case we redirected to unreachable node we only need to remove the
1725 : fibheap entry. */
1726 1815139 : else if (e->aux)
1727 : {
1728 3620 : heap->delete_node ((edge_heap_node_t *) e->aux);
1729 3620 : e->aux = NULL;
1730 : }
1731 25256610 : if (e->next_callee)
1732 : e = e->next_callee;
1733 : else
1734 : {
1735 6657949 : do
1736 : {
1737 6657949 : if (e->caller == node)
1738 912215 : return;
1739 5745734 : if (e->caller == update_since)
1740 439915 : check_inlinability = false;
1741 5745734 : e = e->caller->callers;
1742 : }
1743 5745734 : while (!e->next_callee);
1744 : e = e->next_callee;
1745 : }
1746 : }
1747 : }
1748 :
1749 : /* Enqueue all recursive calls from NODE into priority queue depending on
1750 : how likely we want to recursively inline the call. */
1751 :
1752 : static void
1753 21979 : lookup_recursive_calls (struct cgraph_node *node, struct cgraph_node *where,
1754 : edge_heap_t *heap)
1755 : {
1756 21979 : struct cgraph_edge *e;
1757 21979 : enum availability avail;
1758 :
1759 61379 : for (e = where->callees; e; e = e->next_callee)
1760 39400 : if (e->callee == node
1761 39400 : || (e->callee->ultimate_alias_target (&avail, e->caller) == node
1762 1193 : && avail > AVAIL_INTERPOSABLE))
1763 : {
1764 16205 : inline_badness b (e, -e->sreal_frequency ());
1765 16205 : heap->insert (b, e);
1766 : }
1767 61379 : for (e = where->callees; e; e = e->next_callee)
1768 39400 : if (!e->inline_failed)
1769 8086 : lookup_recursive_calls (node, e->callee, heap);
1770 21979 : }
1771 :
1772 : /* Decide on recursive inlining: in the case function has recursive calls,
1773 : inline until body size reaches given argument. If any new indirect edges
1774 : are discovered in the process, add them to *NEW_EDGES, unless NEW_EDGES
1775 : is NULL. */
1776 :
1777 : static bool
1778 2000 : recursive_inlining (struct cgraph_edge *edge,
1779 : vec<cgraph_edge *> *new_edges)
1780 : {
1781 1432 : cgraph_node *to = (edge->caller->inlined_to
1782 2000 : ? edge->caller->inlined_to : edge->caller);
1783 2000 : int limit = opt_for_fn (to->decl,
1784 : param_max_inline_insns_recursive_auto);
1785 2000 : inline_badness b (edge, sreal::min ());
1786 2000 : edge_heap_t heap (b);
1787 2000 : struct cgraph_node *node;
1788 2000 : struct cgraph_edge *e;
1789 2000 : struct cgraph_node *master_clone = NULL, *next;
1790 2000 : int depth = 0;
1791 2000 : int n = 0;
1792 :
1793 2000 : node = edge->caller;
1794 2000 : if (node->inlined_to)
1795 568 : node = node->inlined_to;
1796 :
1797 2000 : if (DECL_DECLARED_INLINE_P (node->decl))
1798 425 : limit = opt_for_fn (to->decl, param_max_inline_insns_recursive);
1799 :
1800 : /* Make sure that function is small enough to be considered for inlining. */
1801 2000 : if (estimate_size_after_inlining (node, edge) >= limit)
1802 : return false;
1803 2000 : lookup_recursive_calls (node, node, &heap);
1804 2000 : if (heap.empty ())
1805 : return false;
1806 :
1807 2000 : if (dump_file)
1808 4 : fprintf (dump_file,
1809 : " Performing recursive inlining on %s\n", node->dump_name ());
1810 :
1811 : /* Do the inlining and update list of recursive call during process. */
1812 16956 : while (!heap.empty ())
1813 : {
1814 14984 : struct cgraph_edge *curr = heap.extract_min ();
1815 14984 : struct cgraph_node *cnode, *dest = curr->callee;
1816 :
1817 14984 : if (!can_inline_edge_p (curr, true)
1818 14984 : || !can_inline_edge_by_limits_p (curr, CAN_INLINE_REPORT | CAN_INLINE_FORCE_LIMITS))
1819 0 : continue;
1820 :
1821 : /* MASTER_CLONE is produced in the case we already started modified
1822 : the function. Be sure to redirect edge to the original body before
1823 : estimating growths otherwise we will be seeing growths after inlining
1824 : the already modified body. */
1825 14984 : if (master_clone)
1826 : {
1827 12878 : curr->redirect_callee (master_clone);
1828 12878 : if (edge_growth_cache != NULL)
1829 12878 : edge_growth_cache->remove (curr);
1830 : }
1831 :
1832 14984 : if (estimate_size_after_inlining (node, curr) > limit)
1833 : {
1834 28 : curr->redirect_callee (dest);
1835 28 : if (edge_growth_cache != NULL)
1836 28 : edge_growth_cache->remove (curr);
1837 : break;
1838 : }
1839 :
1840 14956 : depth = 1;
1841 14956 : for (cnode = curr->caller;
1842 77783 : cnode->inlined_to; cnode = cnode->callers->caller)
1843 125654 : if (node->decl
1844 62827 : == curr->callee->ultimate_alias_target ()->decl)
1845 62827 : depth++;
1846 :
1847 14956 : if (!want_inline_self_recursive_call_p (curr, node, false, depth))
1848 : {
1849 3063 : curr->redirect_callee (dest);
1850 3063 : if (edge_growth_cache != NULL)
1851 3063 : edge_growth_cache->remove (curr);
1852 3063 : continue;
1853 : }
1854 :
1855 11893 : if (dump_file)
1856 : {
1857 14 : fprintf (dump_file,
1858 : " Inlining call of depth %i", depth);
1859 28 : if (node->count.nonzero_p () && curr->count.initialized_p ())
1860 : {
1861 2 : fprintf (dump_file, " called approx. %.2f times per call",
1862 2 : (double)curr->count.to_gcov_type ()
1863 2 : / node->count.to_gcov_type ());
1864 : }
1865 14 : fprintf (dump_file, "\n");
1866 : }
1867 11893 : if (!master_clone)
1868 : {
1869 : /* We need original clone to copy around. */
1870 1639 : master_clone = node->create_clone (node->decl, node->count,
1871 1639 : false, vNULL, true, NULL, NULL, NULL);
1872 4691 : for (e = master_clone->callees; e; e = e->next_callee)
1873 3052 : if (!e->inline_failed)
1874 496 : clone_inlined_nodes (e, true, true, false, NULL);
1875 1639 : curr->redirect_callee (master_clone);
1876 1639 : if (edge_growth_cache != NULL)
1877 1639 : edge_growth_cache->remove (curr);
1878 : }
1879 :
1880 11893 : inline_call (curr, false, new_edges, &overall_size, true);
1881 11893 : reset_node_cache (node);
1882 11893 : lookup_recursive_calls (node, curr->callee, &heap);
1883 11893 : n++;
1884 : }
1885 :
1886 2000 : if (!heap.empty () && dump_file)
1887 0 : fprintf (dump_file, " Recursive inlining growth limit met.\n");
1888 :
1889 2000 : if (!master_clone)
1890 : return false;
1891 :
1892 1639 : if (dump_enabled_p ())
1893 4 : dump_printf_loc (MSG_NOTE, edge->call_stmt,
1894 : "\n Inlined %i times, "
1895 : "body grown from size %i to %i, time %f to %f\n", n,
1896 4 : ipa_size_summaries->get (master_clone)->size,
1897 4 : ipa_size_summaries->get (node)->size,
1898 4 : ipa_fn_summaries->get (master_clone)->time.to_double (),
1899 4 : ipa_fn_summaries->get (node)->time.to_double ());
1900 :
1901 : /* Remove master clone we used for inlining. We rely that clones inlined
1902 : into master clone gets queued just before master clone so we don't
1903 : need recursion. */
1904 19305 : for (node = symtab->first_function (); node != master_clone;
1905 17666 : node = next)
1906 : {
1907 17666 : next = symtab->next_function (node);
1908 17666 : if (node->inlined_to == master_clone)
1909 955 : node->remove ();
1910 : }
1911 1639 : master_clone->remove ();
1912 1639 : return true;
1913 2000 : }
1914 :
1915 :
1916 : /* Given whole compilation unit estimate of INSNS, compute how large we can
1917 : allow the unit to grow. */
1918 :
1919 : static int64_t
1920 1033928 : compute_max_insns (cgraph_node *node, int insns)
1921 : {
1922 1033928 : int max_insns = insns;
1923 1033928 : if (max_insns < opt_for_fn (node->decl, param_large_unit_insns))
1924 : max_insns = opt_for_fn (node->decl, param_large_unit_insns);
1925 :
1926 1033928 : return ((int64_t) max_insns
1927 1033928 : * (100 + opt_for_fn (node->decl, param_inline_unit_growth)) / 100);
1928 : }
1929 :
1930 :
1931 : /* Compute badness of all edges in NEW_EDGES and add them to the HEAP. */
1932 :
1933 : static void
1934 943610 : add_new_edges_to_heap (edge_heap_t *heap, vec<cgraph_edge *> &new_edges)
1935 : {
1936 946371 : while (new_edges.length () > 0)
1937 : {
1938 2761 : struct cgraph_edge *edge = new_edges.pop ();
1939 :
1940 2761 : gcc_assert (!edge->aux);
1941 2761 : gcc_assert (edge->callee);
1942 2761 : if (edge->inline_failed
1943 2761 : && can_inline_edge_p (edge, true)
1944 1292 : && want_inline_small_function_p (edge, true)
1945 3681 : && can_inline_edge_by_limits_p (edge, CAN_INLINE_REPORT))
1946 : {
1947 920 : inline_badness b (edge, edge_badness (edge, false));
1948 920 : edge->aux = heap->insert (b, edge);
1949 : }
1950 : }
1951 943610 : }
1952 :
1953 : /* Remove EDGE from the fibheap. */
1954 :
1955 : static void
1956 7898 : heap_edge_removal_hook (struct cgraph_edge *e, void *data)
1957 : {
1958 7898 : if (e->aux)
1959 : {
1960 38 : ((edge_heap_t *)data)->delete_node ((edge_heap_node_t *)e->aux);
1961 38 : e->aux = NULL;
1962 : }
1963 7898 : }
1964 :
1965 : /* Return true if speculation of edge E seems useful.
1966 : If ANTICIPATE_INLINING is true, be conservative and hope that E
1967 : may get inlined. */
1968 :
1969 : bool
1970 127611 : speculation_useful_p (struct cgraph_edge *e, bool anticipate_inlining)
1971 : {
1972 : /* If we have already decided to inline the edge, it seems useful.
1973 : Also if ipa-cp or other pass worked hard enough to produce a clone,
1974 : we already decided this is a good idea. */
1975 127611 : if (!e->inline_failed
1976 35035 : || e->callee->clone_of)
1977 : return true;
1978 :
1979 34262 : enum availability avail;
1980 34262 : struct cgraph_node *target = e->callee->ultimate_alias_target (&avail,
1981 : e->callee);
1982 :
1983 34262 : gcc_assert (e->speculative && !e->indirect_unknown_callee);
1984 :
1985 : /* Even if call statement is not hot, we can still have useful speculation
1986 : in cases where a lot of time is spent is callee.
1987 : Do not check maybe_hot_p. */
1988 34262 : if (!e->count.nonzero_p ())
1989 : return false;
1990 :
1991 : /* See if IP optimizations found something potentially useful about the
1992 : function. Do this only if the call seems hot since this is about
1993 : optimizing the code surrounding call site rahter than improving
1994 : callee. */
1995 34220 : if (avail >= AVAIL_AVAILABLE && e->maybe_hot_p ())
1996 : {
1997 33232 : int ecf_flags = flags_from_decl_or_type (target->decl);
1998 33232 : if (ecf_flags & ECF_CONST)
1999 : {
2000 596 : if (!(e->speculative_call_indirect_edge ()->indirect_info
2001 596 : ->ecf_flags & ECF_CONST))
2002 : return true;
2003 : }
2004 32636 : else if (ecf_flags & ECF_PURE)
2005 : {
2006 3951 : if (!(e->speculative_call_indirect_edge ()->indirect_info
2007 3951 : ->ecf_flags & ECF_PURE))
2008 : return true;
2009 : }
2010 28685 : else if (get_modref_function_summary (target))
2011 : return true;
2012 : }
2013 : /* If we did not managed to inline the function nor redirect
2014 : to an ipa-cp clone (that are seen by having local flag set),
2015 : it is probably pointless to inline it unless hardware is missing
2016 : indirect call predictor.
2017 :
2018 : At this point we know we will not dispatch into faster version of
2019 : callee, so if call itself is not hot, we definitely can give up
2020 : speculating. */
2021 12372 : if (!anticipate_inlining && (!target->local || !e->maybe_hot_p ()))
2022 : return false;
2023 : /* For overwritable targets there is not much to do. */
2024 8361 : if (!can_inline_edge_p (e, false)
2025 8361 : || !can_inline_edge_by_limits_p (e, CAN_INLINE_DISREGARD_LIMITS))
2026 6 : return false;
2027 : /* OK, speculation seems interesting. */
2028 : return true;
2029 : }
2030 :
2031 : /* We know that EDGE is not going to be inlined.
2032 : See if we can remove speculation. */
2033 :
2034 : static void
2035 95985 : resolve_noninline_speculation (edge_heap_t *edge_heap, struct cgraph_edge *edge)
2036 : {
2037 95985 : if (edge->speculative && !speculation_useful_p (edge, false))
2038 : {
2039 935 : struct cgraph_node *node = edge->caller;
2040 930 : struct cgraph_node *where = node->inlined_to
2041 935 : ? node->inlined_to : node;
2042 935 : auto_bitmap updated_nodes;
2043 :
2044 935 : if (edge->count.ipa ().initialized_p ())
2045 0 : spec_rem += edge->count.ipa ();
2046 935 : cgraph_edge::resolve_speculation (edge);
2047 935 : reset_edge_caches (where);
2048 935 : ipa_update_overall_fn_summary (where);
2049 935 : update_caller_keys (edge_heap, where,
2050 : updated_nodes, NULL);
2051 935 : update_callee_keys (edge_heap, where, NULL,
2052 : updated_nodes);
2053 935 : }
2054 95985 : }
2055 :
2056 : /* Return true if NODE should be accounted for overall size estimate.
2057 : Skip all nodes optimized for size so we can measure the growth of hot
2058 : part of program no matter of the padding. */
2059 :
2060 : bool
2061 3872165 : inline_account_function_p (struct cgraph_node *node)
2062 : {
2063 3872165 : return (!DECL_EXTERNAL (node->decl)
2064 3669068 : && !opt_for_fn (node->decl, optimize_size)
2065 7446287 : && node->frequency != NODE_FREQUENCY_UNLIKELY_EXECUTED);
2066 : }
2067 :
2068 : /* Count number of callers of NODE and store it into DATA (that
2069 : points to int. Worker for cgraph_for_node_and_aliases. */
2070 :
2071 : static bool
2072 1507541 : sum_callers (struct cgraph_node *node, void *data)
2073 : {
2074 1507541 : struct cgraph_edge *e;
2075 1507541 : int *num_calls = (int *)data;
2076 :
2077 3603488 : for (e = node->callers; e; e = e->next_caller)
2078 2095947 : (*num_calls)++;
2079 1507541 : return false;
2080 : }
2081 :
2082 : /* We only propagate across edges with non-interposable callee. */
2083 :
2084 : inline bool
2085 7192665 : ignore_edge_p (struct cgraph_edge *e)
2086 : {
2087 7192665 : enum availability avail;
2088 7192665 : e->callee->function_or_virtual_thunk_symbol (&avail, e->caller);
2089 7192665 : return (avail <= AVAIL_INTERPOSABLE);
2090 : }
2091 :
2092 : /* We use greedy algorithm for inlining of small functions:
2093 : All inline candidates are put into prioritized heap ordered in
2094 : increasing badness.
2095 :
2096 : The inlining of small functions is bounded by unit growth parameters. */
2097 :
2098 : static void
2099 237280 : inline_small_functions (void)
2100 : {
2101 237280 : struct cgraph_node *node;
2102 237280 : struct cgraph_edge *edge;
2103 237280 : inline_badness b;
2104 237280 : edge_heap_t edge_heap (b);
2105 237280 : auto_bitmap updated_nodes;
2106 237280 : int min_size;
2107 237280 : auto_vec<cgraph_edge *> new_indirect_edges;
2108 237280 : int initial_size = 0;
2109 237280 : struct cgraph_node **order = XCNEWVEC (cgraph_node *, symtab->cgraph_count);
2110 237280 : struct cgraph_edge_hook_list *edge_removal_hook_holder;
2111 237280 : new_indirect_edges.create (8);
2112 :
2113 237280 : edge_removal_hook_holder
2114 237280 : = symtab->add_edge_removal_hook (&heap_edge_removal_hook, &edge_heap);
2115 :
2116 : /* Compute overall unit size and other global parameters used by badness
2117 : metrics. */
2118 :
2119 237280 : has_nonzero_ipa_profile = false;
2120 237280 : ipa_reduced_postorder (order, true, ignore_edge_p);
2121 237280 : free (order);
2122 :
2123 2187840 : FOR_EACH_DEFINED_FUNCTION (node)
2124 1950560 : if (!node->inlined_to)
2125 : {
2126 1846562 : if (!node->alias && node->analyzed
2127 1846562 : && (node->has_gimple_body_p () || node->thunk)
2128 3797088 : && opt_for_fn (node->decl, optimize))
2129 : {
2130 1399769 : class ipa_fn_summary *info = ipa_fn_summaries->get (node);
2131 1399769 : struct ipa_dfs_info *dfs = (struct ipa_dfs_info *) node->aux;
2132 :
2133 : /* Do not account external functions, they will be optimized out
2134 : if not inlined. Also only count the non-cold portion of program. */
2135 1399769 : if (inline_account_function_p (node))
2136 1292036 : initial_size += ipa_size_summaries->get (node)->size;
2137 1399769 : info->growth = estimate_growth (node);
2138 :
2139 1399769 : int num_calls = 0;
2140 1399769 : node->call_for_symbol_and_aliases (sum_callers, &num_calls,
2141 : true);
2142 1399769 : if (num_calls == 1)
2143 476240 : info->single_caller = true;
2144 1399769 : if (dfs && dfs->next_cycle)
2145 : {
2146 5075 : struct cgraph_node *n2;
2147 5075 : int id = dfs->scc_no + 1;
2148 11543 : for (n2 = node; n2;
2149 6468 : n2 = ((struct ipa_dfs_info *) n2->aux)->next_cycle)
2150 10151 : if (opt_for_fn (n2->decl, optimize))
2151 : {
2152 10143 : ipa_fn_summary *info2 = ipa_fn_summaries->get
2153 10143 : (n2->inlined_to ? n2->inlined_to : n2);
2154 10143 : if (info2->scc_no)
2155 : break;
2156 6460 : info2->scc_no = id;
2157 : }
2158 : }
2159 : }
2160 :
2161 4446195 : for (edge = node->callers; edge; edge = edge->next_caller)
2162 2495669 : if (edge->count.ipa ().initialized_p ()
2163 2626413 : && edge->count.ipa ().nonzero_p ())
2164 420 : has_nonzero_ipa_profile = true;
2165 : }
2166 237280 : ipa_free_postorder_info ();
2167 237280 : initialize_growth_caches ();
2168 :
2169 237280 : if (dump_file)
2170 178 : fprintf (dump_file,
2171 : "\nDeciding on inlining of small functions. Starting with size %i.\n",
2172 : initial_size);
2173 :
2174 237280 : overall_size = initial_size;
2175 237280 : min_size = overall_size;
2176 :
2177 : /* Populate the heap with all edges we might inline. */
2178 :
2179 2187840 : FOR_EACH_DEFINED_FUNCTION (node)
2180 : {
2181 1950560 : bool update = false;
2182 1950560 : struct cgraph_edge *next = NULL;
2183 1950560 : bool has_speculative = false;
2184 :
2185 1950560 : if (!opt_for_fn (node->decl, optimize)
2186 : /* With -Og we do not want to perform IPA inlining of small
2187 : functions since there are no scalar cleanups after it
2188 : that would realize the anticipated win. All abstraction
2189 : is removed during early inlining. */
2190 1950560 : || opt_for_fn (node->decl, optimize_debug))
2191 475744 : continue;
2192 :
2193 1474816 : if (dump_file)
2194 803 : fprintf (dump_file, "Enqueueing calls in %s.\n", node->dump_name ());
2195 :
2196 7236349 : for (edge = node->callees; edge; edge = edge->next_callee)
2197 : {
2198 5761533 : if (edge->inline_failed
2199 5761499 : && !edge->aux
2200 5761359 : && can_inline_edge_p (edge, true)
2201 1624896 : && want_inline_small_function_p (edge, true)
2202 982383 : && can_inline_edge_by_limits_p (edge, CAN_INLINE_REPORT)
2203 6736304 : && edge->inline_failed)
2204 : {
2205 974771 : gcc_assert (!edge->aux);
2206 974771 : update_edge_key (&edge_heap, edge);
2207 : }
2208 5761533 : if (edge->speculative)
2209 15405 : has_speculative = true;
2210 : }
2211 1474816 : if (has_speculative)
2212 50497 : for (edge = node->callees; edge; edge = next)
2213 : {
2214 42473 : next = edge->next_callee;
2215 42473 : if (edge->speculative
2216 42473 : && !speculation_useful_p (edge, edge->aux != NULL))
2217 : {
2218 486 : cgraph_edge::resolve_speculation (edge);
2219 486 : update = true;
2220 : }
2221 : }
2222 8024 : if (update)
2223 : {
2224 0 : struct cgraph_node *where = node->inlined_to
2225 412 : ? node->inlined_to : node;
2226 412 : ipa_update_overall_fn_summary (where);
2227 412 : reset_edge_caches (where);
2228 412 : update_caller_keys (&edge_heap, where,
2229 : updated_nodes, NULL);
2230 412 : update_callee_keys (&edge_heap, where, NULL,
2231 : updated_nodes);
2232 412 : bitmap_clear (updated_nodes);
2233 : }
2234 : }
2235 :
2236 237280 : gcc_assert (in_lto_p
2237 : || !has_nonzero_ipa_profile
2238 : || flag_auto_profile
2239 : || (profile_info && flag_branch_probabilities));
2240 :
2241 2966190 : while (!edge_heap.empty ())
2242 : {
2243 2728910 : int old_size = overall_size;
2244 2728910 : struct cgraph_node *where, *callee;
2245 2728910 : sreal badness = edge_heap.min_key ().badness;
2246 2728910 : sreal current_badness;
2247 2728910 : int growth;
2248 :
2249 2728910 : edge = edge_heap.extract_min ();
2250 2728910 : gcc_assert (edge->aux);
2251 2728910 : edge->aux = NULL;
2252 2728910 : if (!edge->inline_failed || !edge->callee->analyzed)
2253 1785275 : continue;
2254 :
2255 : /* Be sure that caches are maintained consistent.
2256 : This check is affected by scaling roundoff errors when compiling for
2257 : IPA this we skip it in that case. */
2258 2728822 : if (flag_checking && !edge->callee->count.ipa_p ()
2259 5033870 : && !has_nonzero_ipa_profile)
2260 : {
2261 2305045 : sreal cached_badness = edge_badness (edge, false);
2262 :
2263 2305045 : int old_size_est = estimate_edge_size (edge);
2264 2305045 : sreal old_time_est = estimate_edge_time (edge);
2265 2305045 : int old_hints_est = estimate_edge_hints (edge);
2266 :
2267 2305045 : if (edge_growth_cache != NULL)
2268 2305045 : edge_growth_cache->remove (edge);
2269 3998373 : reset_node_cache (edge->caller->inlined_to
2270 : ? edge->caller->inlined_to
2271 : : edge->caller);
2272 2305045 : gcc_assert (old_size_est == estimate_edge_size (edge));
2273 2305045 : gcc_assert (old_time_est == estimate_edge_time (edge));
2274 : /* FIXME:
2275 :
2276 : gcc_assert (old_hints_est == estimate_edge_hints (edge));
2277 :
2278 : fails with profile feedback because some hints depends on
2279 : maybe_hot_edge_p predicate and because callee gets inlined to other
2280 : calls, the edge may become cold.
2281 : This ought to be fixed by computing relative probabilities
2282 : for given invocation but that will be better done once whole
2283 : code is converted to sreals. Disable for now and revert to "wrong"
2284 : value so enable/disable checking paths agree. */
2285 2305045 : edge_growth_cache->get (edge)->hints = old_hints_est + 1;
2286 :
2287 : /* When updating the edge costs, we only decrease badness in the keys.
2288 : Increases of badness are handled lazily; when we see key with out
2289 : of date value on it, we re-insert it now. */
2290 2305045 : current_badness = edge_badness (edge, false);
2291 2305045 : gcc_assert (cached_badness == current_badness);
2292 2305045 : gcc_assert (current_badness >= badness);
2293 : }
2294 : else
2295 423780 : current_badness = edge_badness (edge, false);
2296 2728825 : if (current_badness != badness)
2297 : {
2298 1880828 : if (edge_heap.min () && current_badness > edge_heap.min_key ().badness)
2299 : {
2300 1689205 : inline_badness b (edge, current_badness);
2301 1689205 : edge->aux = edge_heap.insert (b, edge);
2302 1689205 : continue;
2303 1689205 : }
2304 : else
2305 191623 : badness = current_badness;
2306 : }
2307 :
2308 1039620 : if (!can_inline_edge_p (edge, true)
2309 1039620 : || !can_inline_edge_by_limits_p (edge, CAN_INLINE_REPORT))
2310 : {
2311 5692 : resolve_noninline_speculation (&edge_heap, edge);
2312 5692 : continue;
2313 : }
2314 :
2315 1033928 : callee = edge->callee->ultimate_alias_target ();
2316 1033928 : growth = estimate_edge_growth (edge);
2317 1033928 : if (dump_file)
2318 : {
2319 525 : fprintf (dump_file,
2320 : "\nConsidering %s with %i size\n",
2321 : callee->dump_name (),
2322 525 : ipa_size_summaries->get (callee)->size);
2323 1050 : fprintf (dump_file,
2324 : " to be inlined into %s in %s:%i\n"
2325 : " Estimated badness is %f, frequency %.2f.\n",
2326 525 : edge->caller->dump_name (),
2327 525 : edge->call_stmt
2328 497 : && (LOCATION_LOCUS (gimple_location ((const gimple *)
2329 : edge->call_stmt))
2330 : > BUILTINS_LOCATION)
2331 488 : ? gimple_filename ((const gimple *) edge->call_stmt)
2332 : : "unknown",
2333 525 : edge->call_stmt
2334 497 : ? gimple_lineno ((const gimple *) edge->call_stmt)
2335 : : -1,
2336 : badness.to_double (),
2337 525 : edge->sreal_frequency ().to_double ());
2338 525 : if (edge->count.ipa ().initialized_p ())
2339 : {
2340 0 : fprintf (dump_file, " Called ");
2341 0 : edge->count.ipa ().dump (dump_file);
2342 0 : fprintf (dump_file, " times\n");
2343 : }
2344 525 : if (dump_flags & TDF_DETAILS)
2345 185 : edge_badness (edge, true);
2346 : }
2347 :
2348 1033928 : where = edge->caller;
2349 :
2350 1033928 : if (overall_size + growth > compute_max_insns (where, min_size)
2351 1033928 : && !DECL_DISREGARD_INLINE_LIMITS (callee->decl))
2352 : {
2353 85635 : edge->inline_failed = CIF_INLINE_UNIT_GROWTH_LIMIT;
2354 85635 : report_inline_failed_reason (edge);
2355 85635 : resolve_noninline_speculation (&edge_heap, edge);
2356 85635 : continue;
2357 : }
2358 :
2359 948293 : if (!want_inline_small_function_p (edge, true))
2360 : {
2361 2279 : resolve_noninline_speculation (&edge_heap, edge);
2362 2279 : continue;
2363 : }
2364 :
2365 946014 : profile_count old_count = callee->count;
2366 :
2367 : /* Heuristics for inlining small functions work poorly for
2368 : recursive calls where we do effects similar to loop unrolling.
2369 : When inlining such edge seems profitable, leave decision on
2370 : specific inliner. */
2371 946014 : if (edge->recursive_p ())
2372 : {
2373 2000 : if (where->inlined_to)
2374 568 : where = where->inlined_to;
2375 :
2376 : /* Disable always_inline on self recursive functions.
2377 : This prevents some inlining bombs such as one in PR113291
2378 : from exploding.
2379 : It is not enough to stop inlining in self recursive always_inlines
2380 : since they may grow large enough so always inlining them even
2381 : with recursin depth 0 is too much.
2382 :
2383 : All sane uses of always_inline should be handled during
2384 : early optimizations. */
2385 2000 : DECL_DISREGARD_INLINE_LIMITS (where->decl) = false;
2386 :
2387 2000 : if (!recursive_inlining (edge,
2388 2000 : opt_for_fn (edge->caller->decl,
2389 : flag_indirect_inlining)
2390 : ? &new_indirect_edges : NULL))
2391 : {
2392 361 : edge->inline_failed = CIF_RECURSIVE_INLINING;
2393 361 : resolve_noninline_speculation (&edge_heap, edge);
2394 361 : continue;
2395 : }
2396 1639 : reset_edge_caches (where);
2397 : /* Recursive inliner inlines all recursive calls of the function
2398 : at once. Consequently we need to update all callee keys. */
2399 1639 : if (opt_for_fn (edge->caller->decl, flag_indirect_inlining))
2400 1614 : add_new_edges_to_heap (&edge_heap, new_indirect_edges);
2401 1639 : update_callee_keys (&edge_heap, where, where, updated_nodes);
2402 1639 : bitmap_clear (updated_nodes);
2403 : }
2404 : else
2405 : {
2406 944014 : struct cgraph_node *outer_node = NULL;
2407 944014 : int depth = 0;
2408 :
2409 : /* Consider the case where self recursive function A is inlined
2410 : into B. This is desired optimization in some cases, since it
2411 : leads to effect similar of loop peeling and we might completely
2412 : optimize out the recursive call. However we must be extra
2413 : selective. */
2414 :
2415 944014 : where = edge->caller;
2416 1510995 : while (where->inlined_to)
2417 : {
2418 566981 : if (where->decl == callee->decl)
2419 11132 : outer_node = where, depth++;
2420 566981 : where = where->callers->caller;
2421 : }
2422 946032 : if (outer_node
2423 944014 : && !want_inline_self_recursive_call_p (edge, outer_node,
2424 : true, depth))
2425 : {
2426 2018 : edge->inline_failed
2427 2018 : = (DECL_DISREGARD_INLINE_LIMITS (edge->callee->decl)
2428 2018 : ? CIF_RECURSIVE_INLINING : CIF_UNSPECIFIED);
2429 2018 : resolve_noninline_speculation (&edge_heap, edge);
2430 2018 : continue;
2431 : }
2432 941996 : else if (depth && dump_file)
2433 6 : fprintf (dump_file, " Peeling recursion with depth %i\n", depth);
2434 :
2435 941996 : gcc_checking_assert (!callee->inlined_to);
2436 :
2437 941996 : int old_size = ipa_size_summaries->get (where)->size;
2438 941996 : sreal old_time = ipa_fn_summaries->get (where)->time;
2439 :
2440 941996 : inline_call (edge, true, &new_indirect_edges, &overall_size, true);
2441 941996 : reset_edge_caches (edge->callee);
2442 941996 : add_new_edges_to_heap (&edge_heap, new_indirect_edges);
2443 :
2444 : /* If caller's size and time increased we do not need to update
2445 : all edges because badness is not going to decrease. */
2446 941996 : if (old_size <= ipa_size_summaries->get (where)->size
2447 886507 : && old_time <= ipa_fn_summaries->get (where)->time
2448 : /* Wrapper penalty may be non-monotonous in this respect.
2449 : Fortunately it only affects small functions. */
2450 1562681 : && !wrapper_heuristics_may_apply (where, old_size))
2451 434950 : update_callee_keys (&edge_heap, edge->callee, edge->callee,
2452 : updated_nodes);
2453 : else
2454 507046 : update_callee_keys (&edge_heap, where,
2455 : edge->callee,
2456 : updated_nodes);
2457 : }
2458 943635 : where = edge->caller;
2459 943635 : if (where->inlined_to)
2460 276205 : where = where->inlined_to;
2461 :
2462 : /* Our profitability metric can depend on local properties
2463 : such as number of inlinable calls and size of the function body.
2464 : After inlining these properties might change for the function we
2465 : inlined into (since it's body size changed) and for the functions
2466 : called by function we inlined (since number of it inlinable callers
2467 : might change). */
2468 943635 : update_caller_keys (&edge_heap, where, updated_nodes, NULL);
2469 : /* Offline copy count has possibly changed, recompute if profile is
2470 : available. */
2471 943635 : struct cgraph_node *n
2472 943635 : = cgraph_node::get (edge->callee->decl)->ultimate_alias_target ();
2473 644657 : if (n != edge->callee && n->analyzed && !(n->count == old_count)
2474 943708 : && n->count.ipa_p ())
2475 73 : update_callee_keys (&edge_heap, n, NULL, updated_nodes);
2476 943635 : bitmap_clear (updated_nodes);
2477 :
2478 943635 : if (dump_enabled_p ())
2479 : {
2480 534 : ipa_fn_summary *s = ipa_fn_summaries->get (where);
2481 :
2482 : /* dump_printf can't handle %+i. */
2483 534 : char buf_net_change[100];
2484 534 : snprintf (buf_net_change, sizeof buf_net_change, "%+i",
2485 : overall_size - old_size);
2486 :
2487 1068 : dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, edge->call_stmt,
2488 : " Inlined %C into %C which now has time %f and "
2489 : "size %i, net change of %s%s.\n",
2490 : edge->callee, edge->caller,
2491 : s->time.to_double (),
2492 534 : ipa_size_summaries->get (edge->caller)->size,
2493 : buf_net_change,
2494 534 : cross_module_call_p (edge)
2495 : ? " (cross module)" : "");
2496 : }
2497 943635 : if (min_size > overall_size)
2498 : {
2499 226456 : min_size = overall_size;
2500 :
2501 226456 : if (dump_file)
2502 398 : fprintf (dump_file, "New minimal size reached: %i\n", min_size);
2503 : }
2504 : }
2505 :
2506 237280 : free_growth_caches ();
2507 237280 : if (dump_enabled_p ())
2508 434 : dump_printf (MSG_NOTE,
2509 : "Unit growth for small function inlining: %i->%i (%i%%)\n",
2510 : initial_size, overall_size,
2511 193 : initial_size ? overall_size * 100 / (initial_size) - 100 : 0);
2512 237280 : symtab->remove_edge_removal_hook (edge_removal_hook_holder);
2513 237280 : }
2514 :
2515 : /* Flatten NODE. Performed both during early inlining and
2516 : at IPA inlining time. */
2517 :
2518 : static void
2519 716 : flatten_function (struct cgraph_node *node, bool early, bool update)
2520 : {
2521 716 : struct cgraph_edge *e;
2522 :
2523 : /* We shouldn't be called recursively when we are being processed. */
2524 716 : gcc_assert (node->aux == NULL);
2525 :
2526 716 : node->aux = (void *) node;
2527 :
2528 1648 : for (e = node->callees; e; e = e->next_callee)
2529 : {
2530 932 : struct cgraph_node *orig_callee;
2531 932 : struct cgraph_node *callee = e->callee->ultimate_alias_target ();
2532 :
2533 : /* We've hit cycle? It is time to give up. */
2534 932 : if (callee->aux)
2535 : {
2536 15 : if (dump_enabled_p ())
2537 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
2538 : "Not inlining %C into %C to avoid cycle.\n",
2539 : callee, e->caller);
2540 15 : if (cgraph_inline_failed_type (e->inline_failed) != CIF_FINAL_ERROR)
2541 15 : e->inline_failed = CIF_RECURSIVE_INLINING;
2542 15 : continue;
2543 : }
2544 :
2545 : /* When the edge is already inlined, we just need to recurse into
2546 : it in order to fully flatten the leaves. */
2547 917 : if (!e->inline_failed)
2548 : {
2549 350 : flatten_function (callee, early, false);
2550 350 : continue;
2551 : }
2552 :
2553 : /* Flatten attribute needs to be processed during late inlining. For
2554 : extra code quality we however do flattening during early optimization,
2555 : too. */
2556 321 : if (!early
2557 567 : ? !can_inline_edge_p (e, true)
2558 246 : && !can_inline_edge_by_limits_p (e, CAN_INLINE_REPORT)
2559 321 : : !can_early_inline_edge_p (e))
2560 419 : continue;
2561 :
2562 148 : if (e->recursive_p ())
2563 : {
2564 0 : if (dump_enabled_p ())
2565 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
2566 : "Not inlining: recursive call.\n");
2567 0 : continue;
2568 : }
2569 :
2570 148 : if (gimple_in_ssa_p (DECL_STRUCT_FUNCTION (node->decl))
2571 296 : != gimple_in_ssa_p (DECL_STRUCT_FUNCTION (callee->decl)))
2572 : {
2573 2 : if (dump_enabled_p ())
2574 2 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
2575 : "Not inlining: SSA form does not match.\n");
2576 2 : continue;
2577 : }
2578 :
2579 : /* Inline the edge and flatten the inline clone. Avoid
2580 : recursing through the original node if the node was cloned. */
2581 146 : if (dump_enabled_p ())
2582 6 : dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, e->call_stmt,
2583 : " Inlining %C into %C.\n",
2584 : callee, e->caller);
2585 146 : orig_callee = callee;
2586 146 : inline_call (e, true, NULL, NULL, false);
2587 146 : if (e->callee != orig_callee)
2588 106 : orig_callee->aux = (void *) node;
2589 146 : flatten_function (e->callee, early, false);
2590 146 : if (e->callee != orig_callee)
2591 106 : orig_callee->aux = NULL;
2592 : }
2593 :
2594 716 : node->aux = NULL;
2595 716 : cgraph_node *where = node->inlined_to ? node->inlined_to : node;
2596 716 : if (update && opt_for_fn (where->decl, optimize))
2597 209 : ipa_update_overall_fn_summary (where);
2598 716 : }
2599 :
2600 : /* Inline NODE to all callers. Worker for cgraph_for_node_and_aliases.
2601 : DATA points to number of calls originally found so we avoid infinite
2602 : recursion. */
2603 :
2604 : static bool
2605 31567 : inline_to_all_callers_1 (struct cgraph_node *node, void *data,
2606 : hash_set<cgraph_node *> *callers)
2607 : {
2608 31567 : int *num_calls = (int *)data;
2609 31567 : bool callee_removed = false;
2610 :
2611 66123 : while (node->callers && !node->inlined_to)
2612 : {
2613 35898 : struct cgraph_node *caller = node->callers->caller;
2614 :
2615 35898 : if (!can_inline_edge_p (node->callers, true)
2616 35898 : || !can_inline_edge_by_limits_p (node->callers, CAN_INLINE_REPORT)
2617 71795 : || node->callers->recursive_p ())
2618 : {
2619 1 : if (dump_file)
2620 0 : fprintf (dump_file, "Uninlinable call found; giving up.\n");
2621 1 : *num_calls = 0;
2622 1 : return false;
2623 : }
2624 :
2625 35897 : if (dump_file)
2626 : {
2627 2 : cgraph_node *ultimate = node->ultimate_alias_target ();
2628 2 : fprintf (dump_file,
2629 : "\nInlining %s size %i.\n",
2630 : ultimate->dump_name (),
2631 2 : ipa_size_summaries->get (ultimate)->size);
2632 2 : fprintf (dump_file,
2633 : " Called once from %s %i insns.\n",
2634 : node->callers->caller->dump_name (),
2635 2 : ipa_size_summaries->get (node->callers->caller)->size);
2636 : }
2637 :
2638 : /* Remember which callers we inlined to, delaying updating the
2639 : overall summary. */
2640 35897 : callers->add (node->callers->caller);
2641 35897 : inline_call (node->callers, true, NULL, NULL, false, &callee_removed);
2642 35897 : if (dump_file)
2643 2 : fprintf (dump_file,
2644 : " Inlined into %s which now has %i size\n",
2645 : caller->dump_name (),
2646 2 : ipa_size_summaries->get (caller)->size);
2647 35897 : if (!(*num_calls)--)
2648 : {
2649 0 : if (dump_file)
2650 0 : fprintf (dump_file, "New calls found; giving up.\n");
2651 0 : return callee_removed;
2652 : }
2653 35897 : if (callee_removed)
2654 : return true;
2655 : }
2656 : return false;
2657 : }
2658 :
2659 : /* Wrapper around inline_to_all_callers_1 doing delayed overall summary
2660 : update. */
2661 :
2662 : static bool
2663 31567 : inline_to_all_callers (struct cgraph_node *node, void *data)
2664 : {
2665 31567 : hash_set<cgraph_node *> callers;
2666 31567 : bool res = inline_to_all_callers_1 (node, data, &callers);
2667 : /* Perform the delayed update of the overall summary of all callers
2668 : processed. This avoids quadratic behavior in the cases where
2669 : we have a lot of calls to the same function. */
2670 61649 : for (hash_set<cgraph_node *>::iterator i = callers.begin ();
2671 61649 : i != callers.end (); ++i)
2672 30082 : ipa_update_overall_fn_summary ((*i)->inlined_to ? (*i)->inlined_to : *i);
2673 31567 : return res;
2674 31567 : }
2675 :
2676 : /* Output overall time estimate. */
2677 : static void
2678 356 : dump_overall_stats (void)
2679 : {
2680 356 : sreal sum_weighted = 0, sum = 0;
2681 356 : struct cgraph_node *node;
2682 :
2683 2288 : FOR_EACH_DEFINED_FUNCTION (node)
2684 1932 : if (!node->inlined_to
2685 1373 : && !node->alias)
2686 : {
2687 1265 : ipa_fn_summary *s = ipa_fn_summaries->get (node);
2688 1265 : if (s != NULL)
2689 : {
2690 1161 : sum += s->time;
2691 1161 : if (node->count.ipa ().initialized_p ())
2692 14 : sum_weighted += s->time * node->count.ipa ().to_gcov_type ();
2693 : }
2694 : }
2695 356 : fprintf (dump_file, "Overall time estimate: "
2696 : "%f weighted by profile: "
2697 : "%f\n", sum.to_double (), sum_weighted.to_double ());
2698 356 : }
2699 :
2700 : /* Output some useful stats about inlining. */
2701 :
2702 : static void
2703 178 : dump_inline_stats (void)
2704 : {
2705 178 : int64_t inlined_cnt = 0, inlined_indir_cnt = 0;
2706 178 : int64_t inlined_virt_cnt = 0, inlined_virt_indir_cnt = 0;
2707 178 : int64_t noninlined_cnt = 0, noninlined_indir_cnt = 0;
2708 178 : int64_t noninlined_virt_cnt = 0, noninlined_virt_indir_cnt = 0;
2709 178 : int64_t inlined_speculative = 0, inlined_speculative_ply = 0;
2710 178 : int64_t indirect_poly_cnt = 0, indirect_cnt = 0;
2711 178 : int64_t reason[CIF_N_REASONS][2];
2712 5874 : sreal reason_freq[CIF_N_REASONS];
2713 178 : int i;
2714 178 : struct cgraph_node *node;
2715 :
2716 178 : memset (reason, 0, sizeof (reason));
2717 5874 : for (i=0; i < CIF_N_REASONS; i++)
2718 5696 : reason_freq[i] = 0;
2719 1242 : FOR_EACH_DEFINED_FUNCTION (node)
2720 : {
2721 1064 : struct cgraph_edge *e;
2722 5947 : for (e = node->callees; e; e = e->next_callee)
2723 : {
2724 4883 : if (e->inline_failed)
2725 : {
2726 4330 : if (e->count.ipa ().initialized_p ())
2727 2611 : reason[(int) e->inline_failed][0] += e->count.ipa ().to_gcov_type ();
2728 4330 : reason_freq[(int) e->inline_failed] += e->sreal_frequency ();
2729 4330 : reason[(int) e->inline_failed][1] ++;
2730 4330 : if (DECL_VIRTUAL_P (e->callee->decl)
2731 4330 : && e->count.ipa ().initialized_p ())
2732 : {
2733 0 : if (e->indirect_inlining_edge)
2734 0 : noninlined_virt_indir_cnt += e->count.ipa ().to_gcov_type ();
2735 : else
2736 0 : noninlined_virt_cnt += e->count.ipa ().to_gcov_type ();
2737 : }
2738 4330 : else if (e->count.ipa ().initialized_p ())
2739 : {
2740 2611 : if (e->indirect_inlining_edge)
2741 0 : noninlined_indir_cnt += e->count.ipa ().to_gcov_type ();
2742 : else
2743 2611 : noninlined_cnt += e->count.ipa ().to_gcov_type ();
2744 : }
2745 : }
2746 553 : else if (e->count.ipa ().initialized_p ())
2747 : {
2748 0 : if (e->speculative)
2749 : {
2750 0 : if (DECL_VIRTUAL_P (e->callee->decl))
2751 0 : inlined_speculative_ply += e->count.ipa ().to_gcov_type ();
2752 : else
2753 0 : inlined_speculative += e->count.ipa ().to_gcov_type ();
2754 : }
2755 0 : else if (DECL_VIRTUAL_P (e->callee->decl))
2756 : {
2757 0 : if (e->indirect_inlining_edge)
2758 0 : inlined_virt_indir_cnt += e->count.ipa ().to_gcov_type ();
2759 : else
2760 0 : inlined_virt_cnt += e->count.ipa ().to_gcov_type ();
2761 : }
2762 : else
2763 : {
2764 0 : if (e->indirect_inlining_edge)
2765 0 : inlined_indir_cnt += e->count.ipa ().to_gcov_type ();
2766 : else
2767 0 : inlined_cnt += e->count.ipa ().to_gcov_type ();
2768 : }
2769 : }
2770 : }
2771 1163 : for (e = node->indirect_calls; e; e = e->next_callee)
2772 198 : if (is_a <cgraph_polymorphic_indirect_info *> (e->indirect_info)
2773 99 : & e->count.ipa ().initialized_p ())
2774 0 : indirect_poly_cnt += e->count.ipa ().to_gcov_type ();
2775 99 : else if (e->count.ipa ().initialized_p ())
2776 0 : indirect_cnt += e->count.ipa ().to_gcov_type ();
2777 : }
2778 178 : if (has_nonzero_ipa_profile)
2779 : {
2780 0 : fprintf (dump_file,
2781 : "Inlined %" PRId64 " + speculative "
2782 : "%" PRId64 " + speculative polymorphic "
2783 : "%" PRId64 " + previously indirect "
2784 : "%" PRId64 " + virtual "
2785 : "%" PRId64 " + virtual and previously indirect "
2786 : "%" PRId64 "\n" "Not inlined "
2787 : "%" PRId64 " + previously indirect "
2788 : "%" PRId64 " + virtual "
2789 : "%" PRId64 " + virtual and previously indirect "
2790 : "%" PRId64 " + still indirect "
2791 : "%" PRId64 " + still indirect polymorphic "
2792 : "%" PRId64 "\n", inlined_cnt,
2793 : inlined_speculative, inlined_speculative_ply,
2794 : inlined_indir_cnt, inlined_virt_cnt, inlined_virt_indir_cnt,
2795 : noninlined_cnt, noninlined_indir_cnt, noninlined_virt_cnt,
2796 : noninlined_virt_indir_cnt, indirect_cnt, indirect_poly_cnt);
2797 0 : fprintf (dump_file, "Removed speculations ");
2798 0 : spec_rem.dump (dump_file);
2799 0 : fprintf (dump_file, "\n");
2800 : }
2801 178 : dump_overall_stats ();
2802 178 : fprintf (dump_file, "\nWhy inlining failed?\n");
2803 6052 : for (i = 0; i < CIF_N_REASONS; i++)
2804 5696 : if (reason[i][1])
2805 146 : fprintf (dump_file, "%-50s: %8i calls, %8f freq, %" PRId64" count\n",
2806 : cgraph_inline_failed_string ((cgraph_inline_failed_t) i),
2807 : (int) reason[i][1], reason_freq[i].to_double (), reason[i][0]);
2808 178 : }
2809 :
2810 : /* Called when node is removed. */
2811 :
2812 : static void
2813 0 : flatten_remove_node_hook (struct cgraph_node *node, void *data)
2814 : {
2815 0 : if (lookup_attribute ("flatten", DECL_ATTRIBUTES (node->decl)) == NULL)
2816 : return;
2817 :
2818 0 : hash_set<struct cgraph_node *> *removed
2819 : = (hash_set<struct cgraph_node *> *) data;
2820 0 : removed->add (node);
2821 : }
2822 :
2823 : /* Decide on the inlining. We do so in the topological order to avoid
2824 : expenses on updating data structures. */
2825 :
2826 : static unsigned int
2827 237280 : ipa_inline (void)
2828 : {
2829 237280 : struct cgraph_node *node;
2830 237280 : int nnodes;
2831 237280 : struct cgraph_node **order;
2832 237280 : int i, j;
2833 237280 : int cold;
2834 237280 : bool remove_functions = false;
2835 :
2836 237280 : order = XCNEWVEC (struct cgraph_node *, symtab->cgraph_count);
2837 :
2838 237280 : if (dump_file)
2839 178 : ipa_dump_fn_summaries (dump_file);
2840 :
2841 237280 : nnodes = ipa_reverse_postorder (order);
2842 237280 : spec_rem = profile_count::zero ();
2843 :
2844 3997254 : FOR_EACH_FUNCTION (node)
2845 : {
2846 3759974 : node->aux = 0;
2847 :
2848 : /* Recompute the default reasons for inlining because they may have
2849 : changed during merging. */
2850 3759974 : if (in_lto_p)
2851 : {
2852 439327 : for (cgraph_edge *e = node->callees; e; e = e->next_callee)
2853 : {
2854 332118 : gcc_assert (e->inline_failed);
2855 332118 : initialize_inline_failed (e);
2856 : }
2857 108495 : for (cgraph_edge *e = node->indirect_calls; e; e = e->next_callee)
2858 1286 : initialize_inline_failed (e);
2859 : }
2860 : }
2861 :
2862 237280 : if (dump_file)
2863 178 : fprintf (dump_file, "\nFlattening functions:\n");
2864 :
2865 : /* First shrink order array, so that it only contains nodes with
2866 : flatten attribute. */
2867 3997254 : for (i = nnodes - 1, j = i; i >= 0; i--)
2868 : {
2869 3759974 : node = order[i];
2870 3759974 : if (node->definition
2871 : /* Do not try to flatten aliases. These may happen for example when
2872 : creating local aliases. */
2873 1950542 : && !node->alias
2874 5606552 : && lookup_attribute ("flatten",
2875 1846578 : DECL_ATTRIBUTES (node->decl)) != NULL)
2876 85 : order[j--] = order[i];
2877 : }
2878 :
2879 : /* After the above loop, order[j + 1] ... order[nnodes - 1] contain
2880 : nodes with flatten attribute. If there is more than one such
2881 : node, we need to register a node removal hook, as flatten_function
2882 : could remove other nodes with flatten attribute. See PR82801. */
2883 237280 : struct cgraph_node_hook_list *node_removal_hook_holder = NULL;
2884 237280 : hash_set<struct cgraph_node *> *flatten_removed_nodes = NULL;
2885 237280 : if (j < nnodes - 2)
2886 : {
2887 15 : flatten_removed_nodes = new hash_set<struct cgraph_node *>;
2888 15 : node_removal_hook_holder
2889 15 : = symtab->add_cgraph_removal_hook (&flatten_remove_node_hook,
2890 : flatten_removed_nodes);
2891 : }
2892 :
2893 : /* In the first pass handle functions to be flattened. Do this with
2894 : a priority so none of our later choices will make this impossible. */
2895 237365 : for (i = nnodes - 1; i > j; i--)
2896 : {
2897 85 : node = order[i];
2898 85 : if (flatten_removed_nodes
2899 85 : && flatten_removed_nodes->contains (node))
2900 0 : continue;
2901 :
2902 : /* Handle nodes to be flattened.
2903 : Ideally when processing callees we stop inlining at the
2904 : entry of cycles, possibly cloning that entry point and
2905 : try to flatten itself turning it into a self-recursive
2906 : function. */
2907 85 : if (dump_file)
2908 4 : fprintf (dump_file, "Flattening %s\n", node->dump_name ());
2909 85 : flatten_function (node, false, true);
2910 : }
2911 :
2912 237280 : if (j < nnodes - 2)
2913 : {
2914 15 : symtab->remove_cgraph_removal_hook (node_removal_hook_holder);
2915 30 : delete flatten_removed_nodes;
2916 : }
2917 237280 : free (order);
2918 :
2919 237280 : if (dump_file)
2920 178 : dump_overall_stats ();
2921 :
2922 237280 : inline_small_functions ();
2923 :
2924 237280 : gcc_assert (symtab->state == IPA_SSA);
2925 237280 : symtab->state = IPA_SSA_AFTER_INLINING;
2926 : /* Do first after-inlining removal. We want to remove all "stale" extern
2927 : inline functions and virtual functions so we really know what is called
2928 : once. */
2929 237280 : symtab->remove_unreachable_nodes (dump_file);
2930 :
2931 : /* Inline functions with a property that after inlining into all callers the
2932 : code size will shrink because the out-of-line copy is eliminated.
2933 : We do this regardless on the callee size as long as function growth limits
2934 : are met. */
2935 237280 : if (dump_file)
2936 178 : fprintf (dump_file,
2937 : "\nDeciding on functions to be inlined into all callers and "
2938 : "removing useless speculations:\n");
2939 :
2940 : /* Inlining one function called once has good chance of preventing
2941 : inlining other function into the same callee. Ideally we should
2942 : work in priority order, but probably inlining hot functions first
2943 : is good cut without the extra pain of maintaining the queue.
2944 :
2945 : ??? this is not really fitting the bill perfectly: inlining function
2946 : into callee often leads to better optimization of callee due to
2947 : increased context for optimization.
2948 : For example if main() function calls a function that outputs help
2949 : and then function that does the main optimization, we should inline
2950 : the second with priority even if both calls are cold by themselves.
2951 :
2952 : We probably want to implement new predicate replacing our use of
2953 : maybe_hot_edge interpreted as maybe_hot_edge || callee is known
2954 : to be hot. */
2955 711840 : for (cold = 0; cold <= 1; cold ++)
2956 : {
2957 6558125 : FOR_EACH_DEFINED_FUNCTION (node)
2958 : {
2959 6083565 : struct cgraph_edge *edge, *next;
2960 6083565 : bool update=false;
2961 :
2962 6083565 : if (!opt_for_fn (node->decl, optimize)
2963 6083565 : || !opt_for_fn (node->decl, flag_inline_functions_called_once))
2964 951364 : continue;
2965 :
2966 20532195 : for (edge = node->callees; edge; edge = next)
2967 : {
2968 15399994 : next = edge->next_callee;
2969 15399994 : if (edge->speculative && !speculation_useful_p (edge, false))
2970 : {
2971 2590 : if (edge->count.ipa ().initialized_p ())
2972 0 : spec_rem += edge->count.ipa ();
2973 2590 : cgraph_edge::resolve_speculation (edge);
2974 2590 : update = true;
2975 2590 : remove_functions = true;
2976 : }
2977 : }
2978 5132201 : if (update)
2979 : {
2980 158 : struct cgraph_node *where = node->inlined_to
2981 2145 : ? node->inlined_to : node;
2982 2145 : reset_edge_caches (where);
2983 2145 : ipa_update_overall_fn_summary (where);
2984 : }
2985 5132201 : if (want_inline_function_to_all_callers_p (node, cold))
2986 : {
2987 27613 : int num_calls = 0;
2988 27613 : node->call_for_symbol_and_aliases (sum_callers, &num_calls,
2989 : true);
2990 27613 : while (node->call_for_symbol_and_aliases
2991 28954 : (inline_to_all_callers, &num_calls, true))
2992 : ;
2993 27613 : remove_functions = true;
2994 : }
2995 : }
2996 : }
2997 :
2998 237280 : if (dump_enabled_p ())
2999 241 : dump_printf (MSG_NOTE,
3000 : "\nInlined %i calls, eliminated %i functions\n\n",
3001 : ncalls_inlined, nfunctions_inlined);
3002 237280 : if (dump_file)
3003 178 : dump_inline_stats ();
3004 :
3005 237280 : if (dump_file)
3006 178 : ipa_dump_fn_summaries (dump_file);
3007 237280 : return remove_functions ? TODO_remove_functions : 0;
3008 : }
3009 :
3010 : /* Inline always-inline function calls in NODE
3011 : (which itself is possibly inline). */
3012 :
3013 : static bool
3014 3636456 : inline_always_inline_functions (struct cgraph_node *node)
3015 : {
3016 3636456 : struct cgraph_edge *e;
3017 3636456 : bool inlined = false;
3018 :
3019 13939926 : for (e = node->callees; e; e = e->next_callee)
3020 : {
3021 10303470 : struct cgraph_node *callee = e->callee->ultimate_alias_target ();
3022 10303470 : gcc_checking_assert (!callee->aux || callee->aux == (void *)(size_t)1);
3023 10303470 : if (!DECL_DISREGARD_INLINE_LIMITS (callee->decl)
3024 : /* Watch for self-recursive cycles. */
3025 10303470 : || callee->aux)
3026 9668091 : continue;
3027 :
3028 635379 : if (e->recursive_p ())
3029 : {
3030 6 : if (dump_enabled_p ())
3031 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
3032 : " Not inlining recursive call to %C.\n",
3033 : e->callee);
3034 6 : e->inline_failed = CIF_RECURSIVE_INLINING;
3035 6 : continue;
3036 : }
3037 635373 : if (callee->definition
3038 635251 : && !ipa_fn_summaries->get (callee))
3039 221 : compute_fn_summary (callee, true);
3040 :
3041 635373 : if (!can_early_inline_edge_p (e))
3042 : {
3043 : /* Set inlined to true if the callee is marked "always_inline" but
3044 : is not inlinable. This will allow flagging an error later in
3045 : expand_call_inline in tree-inline.cc. */
3046 171 : if (lookup_attribute ("always_inline",
3047 171 : DECL_ATTRIBUTES (callee->decl)) != NULL)
3048 28 : inlined = true;
3049 171 : continue;
3050 : }
3051 :
3052 635202 : if (dump_enabled_p ())
3053 18 : dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, e->call_stmt,
3054 : " Inlining %C into %C (always_inline).\n",
3055 : e->callee, e->caller);
3056 635202 : inline_call (e, true, NULL, NULL, false);
3057 635202 : callee->aux = (void *)(size_t)1;
3058 : /* Inline recursively to handle the case where always_inline function was
3059 : not optimized yet since it is a part of a cycle in callgraph. */
3060 635202 : inline_always_inline_functions (e->callee);
3061 635202 : callee->aux = NULL;
3062 635202 : inlined = true;
3063 : }
3064 3636456 : return inlined;
3065 : }
3066 :
3067 : /* Decide on the inlining. We do so in the topological order to avoid
3068 : expenses on updating data structures. */
3069 :
3070 : static bool
3071 2507313 : early_inline_small_functions (struct cgraph_node *node)
3072 : {
3073 2507313 : struct cgraph_edge *e;
3074 2507313 : bool inlined = false;
3075 :
3076 10680366 : for (e = node->callees; e; e = e->next_callee)
3077 : {
3078 8173053 : struct cgraph_node *callee = e->callee->ultimate_alias_target ();
3079 :
3080 : /* We can encounter not-yet-analyzed function during
3081 : early inlining on callgraphs with strongly
3082 : connected components. */
3083 8173053 : ipa_fn_summary *s = ipa_fn_summaries->get (callee);
3084 8173053 : if (s == NULL || !s->inlinable || !e->inline_failed)
3085 4345513 : continue;
3086 :
3087 : /* Do not consider functions not declared inline. */
3088 3827540 : if (!DECL_DECLARED_INLINE_P (callee->decl)
3089 869775 : && !opt_for_fn (node->decl, flag_inline_small_functions)
3090 3880542 : && !opt_for_fn (node->decl, flag_inline_functions))
3091 52937 : continue;
3092 :
3093 3774603 : if (dump_enabled_p ())
3094 158 : dump_printf_loc (MSG_NOTE, e->call_stmt,
3095 : "Considering inline candidate %C.\n",
3096 : callee);
3097 :
3098 3774603 : if (!can_early_inline_edge_p (e))
3099 89533 : continue;
3100 :
3101 3685070 : if (e->recursive_p ())
3102 : {
3103 8777 : if (dump_enabled_p ())
3104 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
3105 : " Not inlining: recursive call.\n");
3106 8777 : continue;
3107 : }
3108 :
3109 3676293 : if (!want_early_inline_function_p (e))
3110 1099450 : continue;
3111 :
3112 2576843 : if (dump_enabled_p ())
3113 105 : dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, e->call_stmt,
3114 : " Inlining %C into %C.\n",
3115 : callee, e->caller);
3116 2576843 : inline_call (e, true, NULL, NULL, false);
3117 2576843 : inlined = true;
3118 : }
3119 :
3120 2507313 : if (inlined)
3121 877197 : ipa_update_overall_fn_summary (node);
3122 :
3123 2507313 : return inlined;
3124 : }
3125 :
3126 : /* With auto-fdo inline all functions that was inlined in the train run
3127 : and inlining seems useful. That is there are enough samples in the callee
3128 : function.
3129 :
3130 : Unlike early inlining, we inline recursively. Profile data is also used
3131 : to produce speculative calls which we then inline. In the case some
3132 : speculatin was introduced, set SPECULATIVE_CALLS. */
3133 :
3134 : static bool
3135 2508846 : inline_functions_by_afdo (struct cgraph_node *node, bool *speculative_calls)
3136 : {
3137 2508846 : if (!flag_auto_profile || !flag_auto_profile_inlining)
3138 : return false;
3139 0 : struct cgraph_edge *e;
3140 0 : bool inlined = false;
3141 :
3142 0 : *speculative_calls |= afdo_vpt_for_early_inline (node);
3143 :
3144 0 : cgraph_edge *next;
3145 0 : for (e = node->callees; e; e = next)
3146 : {
3147 0 : next = e->next_callee;
3148 :
3149 0 : if (!e->inline_failed)
3150 : {
3151 0 : inlined |= inline_functions_by_afdo (e->callee, speculative_calls);
3152 0 : continue;
3153 : }
3154 0 : if (!afdo_callsite_hot_enough_for_early_inline (e))
3155 : {
3156 : /* If we do not want to inline, remove the speculation. */
3157 0 : if (e->speculative)
3158 0 : cgraph_edge::resolve_speculation (e);
3159 0 : continue;
3160 : }
3161 :
3162 0 : struct cgraph_node *callee = e->callee->ultimate_alias_target ();
3163 0 : if (callee->definition
3164 0 : && !ipa_fn_summaries->get (callee))
3165 0 : compute_fn_summary (callee, true);
3166 :
3167 0 : if (!can_early_inline_edge_p (e))
3168 : {
3169 0 : if (dump_enabled_p ())
3170 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
3171 : "Not inlining %C -> %C using auto-profile, %s.",
3172 : e->caller, e->callee,
3173 : cgraph_inline_failed_string (e->inline_failed));
3174 : /* If we do not want to inline, remove the speculation. */
3175 0 : if (e->speculative)
3176 0 : cgraph_edge::resolve_speculation (e);
3177 0 : continue;
3178 : }
3179 : /* We can handle recursive inlining by first producing
3180 : inline clone. */
3181 0 : if (e->recursive_p ())
3182 : {
3183 0 : if (dump_enabled_p ())
3184 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
3185 : "Not inlining %C recursively"
3186 : " using auto-profile.\n",
3187 : e->callee);
3188 : /* If we do not want to inline, remove the speculation. */
3189 0 : if (e->speculative)
3190 0 : cgraph_edge::resolve_speculation (e);
3191 0 : continue;
3192 : }
3193 :
3194 0 : if (dump_enabled_p ())
3195 : {
3196 0 : if (e->caller->inlined_to)
3197 0 : dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, e->call_stmt,
3198 : "Inlining using auto-profile %C into %C "
3199 : "which is transitively inlined to %C.\n",
3200 : callee, e->caller, e->caller->inlined_to);
3201 : else
3202 0 : dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, e->call_stmt,
3203 : "Inlining using auto-profile %C into %C.\n",
3204 : callee, e->caller);
3205 : }
3206 0 : if (e->speculative)
3207 0 : remove_afdo_speculative_target (e);
3208 0 : inline_call (e, true, NULL, NULL, false);
3209 0 : inlined |= inline_functions_by_afdo (e->callee, speculative_calls);
3210 0 : inlined = true;
3211 : }
3212 :
3213 0 : if (inlined && !node->inlined_to)
3214 0 : ipa_update_overall_fn_summary (node);
3215 :
3216 : return inlined;
3217 : }
3218 :
3219 : unsigned int
3220 3001268 : early_inliner (function *fun)
3221 : {
3222 3001268 : struct cgraph_node *node = cgraph_node::get (current_function_decl);
3223 3001268 : struct cgraph_edge *edge;
3224 3001268 : unsigned int todo = 0;
3225 3001268 : int iterations = 0;
3226 3001268 : bool inlined = false;
3227 :
3228 3001268 : if (seen_error ())
3229 : return 0;
3230 :
3231 : /* Do nothing if datastructures for ipa-inliner are already computed. This
3232 : happens when some pass decides to construct new function and
3233 : cgraph_add_new_function calls lowering passes and early optimization on
3234 : it. This may confuse ourself when early inliner decide to inline call to
3235 : function clone, because function clones don't have parameter list in
3236 : ipa-prop matching their signature. */
3237 3001262 : if (ipa_node_params_sum)
3238 : return 0;
3239 :
3240 3001254 : if (flag_checking)
3241 3001224 : node->verify ();
3242 3001254 : node->remove_all_references ();
3243 :
3244 : /* Even when not optimizing or not inlining inline always-inline
3245 : functions. */
3246 3001254 : inlined = inline_always_inline_functions (node);
3247 :
3248 3001254 : if (!optimize
3249 2544722 : || flag_no_inline
3250 2508884 : || !flag_early_inlining)
3251 : ;
3252 2507351 : else if (lookup_attribute ("flatten",
3253 2507351 : DECL_ATTRIBUTES (node->decl)) != NULL)
3254 : {
3255 : /* When the function is marked to be flattened, recursively inline
3256 : all calls in it. */
3257 135 : if (dump_enabled_p ())
3258 0 : dump_printf (MSG_OPTIMIZED_LOCATIONS,
3259 : "Flattening %C\n", node);
3260 135 : flatten_function (node, true, true);
3261 135 : inlined = true;
3262 : }
3263 : else
3264 : {
3265 : /* If some always_inline functions was inlined, apply the changes.
3266 : This way we will not account always inline into growth limits and
3267 : moreover we will inline calls from always inlines that we skipped
3268 : previously because of conditional in can_early_inline_edge_p
3269 : which prevents some inlining to always_inline. */
3270 2507216 : if (inlined)
3271 : {
3272 325515 : timevar_push (TV_INTEGRATION);
3273 325515 : todo |= optimize_inline_calls (current_function_decl);
3274 : /* optimize_inline_calls call above might have introduced new
3275 : statements that don't have inline parameters computed. */
3276 1653599 : for (edge = node->callees; edge; edge = edge->next_callee)
3277 : {
3278 : /* We can encounter not-yet-analyzed function during
3279 : early inlining on callgraphs with strongly
3280 : connected components. */
3281 1328084 : ipa_call_summary *es = ipa_call_summaries->get_create (edge);
3282 1328084 : es->call_stmt_size
3283 1328084 : = estimate_num_insns (edge->call_stmt, &eni_size_weights);
3284 1328084 : es->call_stmt_time
3285 1328084 : = estimate_num_insns (edge->call_stmt, &eni_time_weights);
3286 : }
3287 325515 : ipa_update_overall_fn_summary (node);
3288 325515 : inlined = false;
3289 325515 : timevar_pop (TV_INTEGRATION);
3290 : }
3291 : /* We iterate incremental inlining to get trivial cases of indirect
3292 : inlining. */
3293 3384413 : while (iterations < opt_for_fn (node->decl,
3294 : param_early_inliner_max_iterations))
3295 : {
3296 2507313 : bool inlined = early_inline_small_functions (node);
3297 2507313 : bool speculative_calls = false;
3298 2507313 : inlined |= inline_functions_by_afdo (node, &speculative_calls);
3299 2507313 : if (!inlined)
3300 : break;
3301 877197 : timevar_push (TV_INTEGRATION);
3302 877197 : if (speculative_calls)
3303 : {
3304 0 : cgraph_edge *next;
3305 0 : for (cgraph_edge *e = node->callees; e; e = next)
3306 : {
3307 0 : next = e->next_callee;
3308 0 : cgraph_edge::redirect_call_stmt_to_callee (e);
3309 : }
3310 : }
3311 877197 : todo |= optimize_inline_calls (current_function_decl);
3312 :
3313 : /* Technically we ought to recompute inline parameters so the new
3314 : iteration of early inliner works as expected. We however have
3315 : values approximately right and thus we only need to update edge
3316 : info that might be cleared out for newly discovered edges. */
3317 3599433 : for (edge = node->callees; edge; edge = edge->next_callee)
3318 : {
3319 : /* We have no summary for new bound store calls yet. */
3320 2722236 : ipa_call_summary *es = ipa_call_summaries->get_create (edge);
3321 2722236 : es->call_stmt_size
3322 2722236 : = estimate_num_insns (edge->call_stmt, &eni_size_weights);
3323 2722236 : es->call_stmt_time
3324 2722236 : = estimate_num_insns (edge->call_stmt, &eni_time_weights);
3325 : }
3326 877197 : if (iterations < opt_for_fn (node->decl,
3327 877197 : param_early_inliner_max_iterations) - 1)
3328 97 : ipa_update_overall_fn_summary (node);
3329 877197 : timevar_pop (TV_INTEGRATION);
3330 877197 : iterations++;
3331 877197 : inlined = false;
3332 : }
3333 2507216 : if (dump_file)
3334 206 : fprintf (dump_file, "Iterations: %i\n", iterations);
3335 : }
3336 :
3337 : /* do AFDO inlining in case it was not done as part of early inlining. */
3338 3001254 : if (optimize
3339 2544722 : && !flag_no_inline
3340 2508884 : && !flag_early_inlining
3341 1533 : && flag_auto_profile_inlining)
3342 : {
3343 1533 : bool speculative_calls = false;
3344 1533 : inlined |= inline_functions_by_afdo (node, &speculative_calls);
3345 1533 : if (speculative_calls)
3346 : {
3347 0 : cgraph_edge *next;
3348 0 : for (cgraph_edge *e = node->callees; e; e = next)
3349 : {
3350 0 : next = e->next_callee;
3351 0 : cgraph_edge::redirect_call_stmt_to_callee (e);
3352 : }
3353 : }
3354 : }
3355 :
3356 3001254 : if (inlined)
3357 : {
3358 25476 : timevar_push (TV_INTEGRATION);
3359 25476 : todo |= optimize_inline_calls (current_function_decl);
3360 25476 : timevar_pop (TV_INTEGRATION);
3361 : }
3362 :
3363 3001254 : fun->always_inline_functions_inlined = true;
3364 :
3365 3001254 : return todo;
3366 : }
3367 :
3368 : /* Do inlining of small functions. Doing so early helps profiling and other
3369 : passes to be somewhat more effective and avoids some code duplication in
3370 : later real inlining pass for testcases with very many function calls. */
3371 :
3372 : namespace {
3373 :
3374 : const pass_data pass_data_early_inline =
3375 : {
3376 : GIMPLE_PASS, /* type */
3377 : "einline", /* name */
3378 : OPTGROUP_INLINE, /* optinfo_flags */
3379 : TV_EARLY_INLINING, /* tv_id */
3380 : PROP_ssa, /* properties_required */
3381 : 0, /* properties_provided */
3382 : 0, /* properties_destroyed */
3383 : 0, /* todo_flags_start */
3384 : 0, /* todo_flags_finish */
3385 : };
3386 :
3387 : class pass_early_inline : public gimple_opt_pass
3388 : {
3389 : public:
3390 294587 : pass_early_inline (gcc::context *ctxt)
3391 589174 : : gimple_opt_pass (pass_data_early_inline, ctxt)
3392 : {}
3393 :
3394 : /* opt_pass methods: */
3395 : unsigned int execute (function *) final override;
3396 :
3397 : }; // class pass_early_inline
3398 :
3399 : unsigned int
3400 3001268 : pass_early_inline::execute (function *fun)
3401 : {
3402 3001268 : return early_inliner (fun);
3403 : }
3404 :
3405 : } // anon namespace
3406 :
3407 : gimple_opt_pass *
3408 294587 : make_pass_early_inline (gcc::context *ctxt)
3409 : {
3410 294587 : return new pass_early_inline (ctxt);
3411 : }
3412 :
3413 : namespace {
3414 :
3415 : const pass_data pass_data_ipa_inline =
3416 : {
3417 : IPA_PASS, /* type */
3418 : "inline", /* name */
3419 : OPTGROUP_INLINE, /* optinfo_flags */
3420 : TV_IPA_INLINING, /* tv_id */
3421 : 0, /* properties_required */
3422 : 0, /* properties_provided */
3423 : 0, /* properties_destroyed */
3424 : 0, /* todo_flags_start */
3425 : ( TODO_dump_symtab ), /* todo_flags_finish */
3426 : };
3427 :
3428 : class pass_ipa_inline : public ipa_opt_pass_d
3429 : {
3430 : public:
3431 294587 : pass_ipa_inline (gcc::context *ctxt)
3432 : : ipa_opt_pass_d (pass_data_ipa_inline, ctxt,
3433 : NULL, /* generate_summary */
3434 : NULL, /* write_summary */
3435 : NULL, /* read_summary */
3436 : NULL, /* write_optimization_summary */
3437 : NULL, /* read_optimization_summary */
3438 : NULL, /* stmt_fixup */
3439 : 0, /* function_transform_todo_flags_start */
3440 : inline_transform, /* function_transform */
3441 294587 : NULL) /* variable_transform */
3442 294587 : {}
3443 :
3444 : /* opt_pass methods: */
3445 237280 : unsigned int execute (function *) final override { return ipa_inline (); }
3446 :
3447 : }; // class pass_ipa_inline
3448 :
3449 : } // anon namespace
3450 :
3451 : ipa_opt_pass_d *
3452 294587 : make_pass_ipa_inline (gcc::context *ctxt)
3453 : {
3454 294587 : return new pass_ipa_inline (ctxt);
3455 : }
|
