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 1223653 : inline_badness ()
138 993594 : : badness (sreal::min ()), uid (0)
139 : {
140 : }
141 3599120 : inline_badness (cgraph_edge *e, sreal b)
142 18009 : : badness (b), uid (e->get_uid ())
143 : {
144 : }
145 891882 : bool operator<= (const inline_badness &other)
146 : {
147 891882 : if (badness != other.badness)
148 891882 : return badness <= other.badness;
149 0 : return uid <= other.uid;
150 : }
151 993594 : bool operator== (const inline_badness &other)
152 : {
153 1842297 : return badness == other.badness && uid == other.uid;
154 : }
155 0 : bool operator!= (const inline_badness &other)
156 : {
157 993594 : return badness != other.badness || uid != other.uid;
158 : }
159 27993138 : bool operator< (const inline_badness &other)
160 : {
161 27993138 : if (badness != other.badness)
162 23236473 : return badness < other.badness;
163 4756665 : return uid < other.uid;
164 : }
165 12867727 : bool operator> (const inline_badness &other)
166 : {
167 12867727 : if (badness != other.badness)
168 10521008 : return badness > other.badness;
169 2346719 : 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 6431814 : caller_growth_limits (struct cgraph_edge *e)
196 : {
197 6431814 : struct cgraph_node *to = e->caller;
198 6431814 : struct cgraph_node *what = e->callee->ultimate_alias_target ();
199 6431814 : int newsize;
200 6431814 : int limit = 0;
201 6431814 : HOST_WIDE_INT stack_size_limit = 0, inlined_stack;
202 6431814 : 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 9139484 : while (true)
213 : {
214 7785649 : ipa_size_summary *size_info = ipa_size_summaries->get (to);
215 7785649 : if (limit < size_info->self_size)
216 : limit = size_info->self_size;
217 7785649 : if (stack_size_limit < size_info->estimated_self_stack_size)
218 : stack_size_limit = size_info->estimated_self_stack_size;
219 7785649 : if (to->inlined_to)
220 1353835 : to = to->callers->caller;
221 : else
222 : break;
223 1353835 : }
224 :
225 6431814 : ipa_fn_summary *what_info = ipa_fn_summaries->get (what);
226 6431814 : ipa_size_summary *what_size_info = ipa_size_summaries->get (what);
227 :
228 6431814 : if (limit < what_size_info->self_size)
229 : limit = what_size_info->self_size;
230 :
231 6431814 : 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 6431814 : newsize = estimate_size_after_inlining (to, e);
236 6431814 : if (newsize >= ipa_size_summaries->get (what)->size
237 6254156 : && newsize > opt_for_fn (to->decl, param_large_function_insns)
238 6684355 : && newsize > limit)
239 : {
240 6046 : e->inline_failed = CIF_LARGE_FUNCTION_GROWTH_LIMIT;
241 6046 : return false;
242 : }
243 :
244 6425768 : 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 2287700 : stack_size_limit += ((gcov_type)stack_size_limit
253 1143850 : * opt_for_fn (to->decl, param_stack_frame_growth)
254 1143850 : / 100);
255 :
256 1143850 : inlined_stack = (ipa_get_stack_frame_offset (to)
257 1143850 : + outer_info->estimated_self_stack_size
258 1143850 : + what_info->estimated_stack_size);
259 : /* Check new stack consumption with stack consumption at the place
260 : stack is used. */
261 1143850 : 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 334973 : && inlined_stack > ipa_fn_summaries->get (to)->estimated_stack_size
267 1465661 : && inlined_stack > opt_for_fn (to->decl, param_large_stack_frame))
268 : {
269 60871 : e->inline_failed = CIF_LARGE_STACK_FRAME_GROWTH_LIMIT;
270 60871 : return false;
271 : }
272 : return true;
273 : }
274 :
275 : /* Dump info about why inlining has failed. */
276 :
277 : static void
278 4966542 : report_inline_failed_reason (struct cgraph_edge *e)
279 : {
280 4966542 : if (dump_enabled_p ())
281 : {
282 2398 : 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 2398 : if ((e->inline_failed == CIF_TARGET_OPTION_MISMATCH
287 2398 : || e->inline_failed == CIF_OPTIMIZATION_MISMATCH)
288 2 : && e->caller->lto_file_data
289 2398 : && 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 2398 : if (e->inline_failed == CIF_TARGET_OPTION_MISMATCH)
297 2 : if (dump_file)
298 0 : cl_target_option_print_diff
299 0 : (dump_file, 2, target_opts_for_fn (e->caller->decl),
300 0 : target_opts_for_fn (e->callee->ultimate_alias_target ()->decl));
301 2398 : if (e->inline_failed == CIF_OPTIMIZATION_MISMATCH)
302 0 : if (dump_file)
303 0 : cl_optimization_print_diff
304 0 : (dump_file, 2, opts_for_fn (e->caller->decl),
305 0 : opts_for_fn (e->callee->ultimate_alias_target ()->decl));
306 : }
307 4966542 : }
308 :
309 : /* Decide whether sanitizer-related attributes allow inlining. */
310 :
311 : static bool
312 9225576 : sanitize_attrs_match_for_inline_p (const_tree caller, const_tree callee)
313 : {
314 9225576 : if (!caller || !callee)
315 : return true;
316 :
317 : /* Follow clang and allow inlining for always_inline functions. */
318 9225576 : if (lookup_attribute ("always_inline", DECL_ATTRIBUTES (callee)))
319 : return true;
320 :
321 8669757 : const sanitize_code codes[] =
322 : {
323 : SANITIZE_ADDRESS,
324 : SANITIZE_THREAD,
325 : SANITIZE_UNDEFINED,
326 : SANITIZE_UNDEFINED_NONDEFAULT,
327 : SANITIZE_POINTER_COMPARE,
328 : SANITIZE_POINTER_SUBTRACT
329 : };
330 :
331 60687259 : for (unsigned i = 0; i < ARRAY_SIZE (codes); i++)
332 104035400 : if (sanitize_flags_p (codes[i], caller)
333 52017700 : != sanitize_flags_p (codes[i], callee))
334 : return false;
335 :
336 8669559 : if (sanitize_coverage_p (caller) != sanitize_coverage_p (callee))
337 : return false;
338 :
339 : return true;
340 : }
341 :
342 : /* Used for flags where it is safe to inline when caller's value is
343 : grater than callee's. */
344 : #define check_maybe_up(flag) \
345 : (opts_for_fn (caller->decl)->x_##flag \
346 : != opts_for_fn (callee->decl)->x_##flag \
347 : && (!always_inline \
348 : || opts_for_fn (caller->decl)->x_##flag \
349 : < opts_for_fn (callee->decl)->x_##flag))
350 : /* Used for flags where it is safe to inline when caller's value is
351 : smaller than callee's. */
352 : #define check_maybe_down(flag) \
353 : (opts_for_fn (caller->decl)->x_##flag \
354 : != opts_for_fn (callee->decl)->x_##flag \
355 : && (!always_inline \
356 : || opts_for_fn (caller->decl)->x_##flag \
357 : > opts_for_fn (callee->decl)->x_##flag))
358 : /* Used for flags where exact match is needed for correctness. */
359 : #define check_match(flag) \
360 : (opts_for_fn (caller->decl)->x_##flag \
361 : != opts_for_fn (callee->decl)->x_##flag)
362 :
363 : /* Decide if we can inline the edge and possibly update
364 : inline_failed reason.
365 : We check whether inlining is possible at all and whether
366 : caller growth limits allow doing so.
367 :
368 : if REPORT is true, output reason to the dump file. */
369 :
370 : static bool
371 13367874 : can_inline_edge_p (struct cgraph_edge *e, bool report,
372 : bool early = false)
373 : {
374 13367874 : gcc_checking_assert (e->inline_failed);
375 :
376 13367874 : if (cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR)
377 : {
378 3651932 : if (report)
379 3619906 : report_inline_failed_reason (e);
380 3651932 : return false;
381 : }
382 :
383 9715942 : bool inlinable = true;
384 9715942 : enum availability avail;
385 8483394 : cgraph_node *caller = (e->caller->inlined_to
386 9715942 : ? e->caller->inlined_to : e->caller);
387 9715942 : cgraph_node *callee = e->callee->ultimate_alias_target (&avail, caller);
388 :
389 9715942 : if (!callee->definition)
390 : {
391 1187 : e->inline_failed = CIF_BODY_NOT_AVAILABLE;
392 1187 : inlinable = false;
393 : }
394 9715942 : if (!early && (!opt_for_fn (callee->decl, optimize)
395 5584044 : || !opt_for_fn (caller->decl, optimize)))
396 : {
397 305 : e->inline_failed = CIF_FUNCTION_NOT_OPTIMIZED;
398 305 : inlinable = false;
399 : }
400 9715637 : else if (callee->calls_comdat_local)
401 : {
402 20382 : e->inline_failed = CIF_USES_COMDAT_LOCAL;
403 20382 : inlinable = false;
404 : }
405 9695255 : else if (avail <= AVAIL_INTERPOSABLE)
406 : {
407 129746 : e->inline_failed = CIF_OVERWRITABLE;
408 129746 : inlinable = false;
409 : }
410 : /* All edges with call_stmt_cannot_inline_p should have inline_failed
411 : initialized to one of FINAL_ERROR reasons. */
412 9565509 : else if (e->call_stmt_cannot_inline_p)
413 0 : gcc_unreachable ();
414 : /* Don't inline if the functions have different EH personalities. */
415 9565509 : else if (DECL_FUNCTION_PERSONALITY (caller->decl)
416 2652716 : && DECL_FUNCTION_PERSONALITY (callee->decl)
417 9565509 : && (DECL_FUNCTION_PERSONALITY (caller->decl)
418 220289 : != DECL_FUNCTION_PERSONALITY (callee->decl)))
419 : {
420 0 : e->inline_failed = CIF_EH_PERSONALITY;
421 0 : inlinable = false;
422 : }
423 : /* TM pure functions should not be inlined into non-TM_pure
424 : functions. */
425 9565509 : else if (is_tm_pure (callee->decl) && !is_tm_pure (caller->decl))
426 : {
427 30 : e->inline_failed = CIF_UNSPECIFIED;
428 30 : inlinable = false;
429 : }
430 : /* Check compatibility of target optimization options. */
431 9565479 : else if (!targetm.target_option.can_inline_p (caller->decl,
432 : callee->decl))
433 : {
434 468 : e->inline_failed = CIF_TARGET_OPTION_MISMATCH;
435 468 : inlinable = false;
436 : }
437 9565011 : else if (ipa_fn_summaries->get (callee) == NULL
438 9565009 : || !ipa_fn_summaries->get (callee)->inlinable)
439 : {
440 339435 : e->inline_failed = CIF_FUNCTION_NOT_INLINABLE;
441 339435 : inlinable = false;
442 : }
443 : /* Don't inline a function with mismatched sanitization attributes. */
444 9225576 : else if (!sanitize_attrs_match_for_inline_p (caller->decl, callee->decl))
445 : {
446 198 : e->inline_failed = CIF_SANITIZE_ATTRIBUTE_MISMATCH;
447 198 : inlinable = false;
448 : }
449 :
450 9715942 : if (inlinable && !strub_inlinable_to_p (callee, caller))
451 : {
452 1101 : e->inline_failed = CIF_UNSPECIFIED;
453 1101 : inlinable = false;
454 : }
455 9715942 : if (inlinable && callee->must_remain_in_tu_body
456 9 : && caller->lto_file_data != callee->lto_file_data)
457 : {
458 9 : e->inline_failed = CIF_MUST_REMAIN_IN_TU;
459 9 : inlinable = false;
460 : }
461 9715942 : if (!inlinable && report)
462 484413 : report_inline_failed_reason (e);
463 : return inlinable;
464 : }
465 :
466 : /* Return inlining_insns_single limit for function N. If HINT or HINT2 is true
467 : scale up the bound. */
468 :
469 : static int
470 8314791 : inline_insns_single (cgraph_node *n, bool hint, bool hint2)
471 : {
472 8314791 : if (hint && hint2)
473 : {
474 2883591 : int64_t spd = opt_for_fn (n->decl, param_inline_heuristics_hint_percent);
475 2883591 : spd = spd * spd;
476 2883591 : if (spd > 1000000)
477 : spd = 1000000;
478 2883591 : return opt_for_fn (n->decl, param_max_inline_insns_single) * spd / 100;
479 : }
480 5431200 : if (hint || hint2)
481 495129 : return opt_for_fn (n->decl, param_max_inline_insns_single)
482 495129 : * opt_for_fn (n->decl, param_inline_heuristics_hint_percent) / 100;
483 4936071 : return opt_for_fn (n->decl, param_max_inline_insns_single);
484 : }
485 :
486 : /* Return inlining_insns_auto limit for function N. If HINT or HINT2 is true
487 : scale up the bound. */
488 :
489 : static int
490 5429174 : inline_insns_auto (cgraph_node *n, bool hint, bool hint2)
491 : {
492 5429174 : int max_inline_insns_auto = opt_for_fn (n->decl, param_max_inline_insns_auto);
493 5429174 : if (hint && hint2)
494 : {
495 1988913 : int64_t spd = opt_for_fn (n->decl, param_inline_heuristics_hint_percent);
496 1988913 : spd = spd * spd;
497 1988913 : if (spd > 1000000)
498 : spd = 1000000;
499 1988913 : return max_inline_insns_auto * spd / 100;
500 : }
501 3440261 : if (hint || hint2)
502 1449314 : return max_inline_insns_auto
503 1449314 : * opt_for_fn (n->decl, param_inline_heuristics_hint_percent) / 100;
504 : return max_inline_insns_auto;
505 : }
506 :
507 : enum can_inline_edge_by_limits_flags
508 : {
509 : /* True if we are early inlining. */
510 : CAN_INLINE_EARLY = 1,
511 : /* Ignore size limits. */
512 : CAN_INLINE_DISREGARD_LIMITS = 2,
513 : /* Force size limits (ignore always_inline). This is used for
514 : recrusive inlining where always_inline may lead to inline bombs
515 : and technically it is non-sential anyway. */
516 : CAN_INLINE_FORCE_LIMITS = 4,
517 : /* Report decision to dump file. */
518 : CAN_INLINE_REPORT = 8,
519 : };
520 :
521 : /* Decide if we can inline the edge and possibly update
522 : inline_failed reason.
523 : We check whether inlining is possible at all and whether
524 : caller growth limits allow doing so. */
525 :
526 : static bool
527 7000376 : can_inline_edge_by_limits_p (struct cgraph_edge *e, int flags)
528 : {
529 7000376 : gcc_checking_assert (e->inline_failed);
530 :
531 7000376 : if (cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR)
532 : {
533 210 : if (flags & CAN_INLINE_REPORT)
534 210 : report_inline_failed_reason (e);
535 210 : return false;
536 : }
537 :
538 7000166 : bool inlinable = true;
539 7000166 : enum availability avail;
540 6290633 : cgraph_node *caller = (e->caller->inlined_to
541 7000166 : ? e->caller->inlined_to : e->caller);
542 7000166 : cgraph_node *callee = e->callee->ultimate_alias_target (&avail, caller);
543 7000166 : tree caller_tree = DECL_FUNCTION_SPECIFIC_OPTIMIZATION (caller->decl);
544 7000166 : tree callee_tree
545 7000166 : = callee ? DECL_FUNCTION_SPECIFIC_OPTIMIZATION (callee->decl) : NULL;
546 : /* Check if caller growth allows the inlining. */
547 7000166 : if (!(flags & CAN_INLINE_DISREGARD_LIMITS)
548 6989886 : && ((flags & CAN_INLINE_FORCE_LIMITS)
549 6955406 : || (!DECL_DISREGARD_INLINE_LIMITS (callee->decl)
550 6397482 : && !lookup_attribute ("flatten",
551 6397482 : DECL_ATTRIBUTES (caller->decl))))
552 13431980 : && !caller_growth_limits (e))
553 : inlinable = false;
554 6933249 : else if (callee->externally_visible
555 5143527 : && !DECL_DISREGARD_INLINE_LIMITS (callee->decl)
556 11634975 : && flag_live_patching == LIVE_PATCHING_INLINE_ONLY_STATIC)
557 : {
558 2 : e->inline_failed = CIF_EXTERN_LIVE_ONLY_STATIC;
559 2 : inlinable = false;
560 : }
561 : /* Don't inline a function with a higher optimization level than the
562 : caller. FIXME: this is really just tip of iceberg of handling
563 : optimization attribute. */
564 6933247 : else if (caller_tree != callee_tree)
565 : {
566 9651 : bool always_inline =
567 9651 : (DECL_DISREGARD_INLINE_LIMITS (callee->decl)
568 12081 : && lookup_attribute ("always_inline",
569 2430 : DECL_ATTRIBUTES (callee->decl)));
570 9651 : ipa_fn_summary *caller_info = ipa_fn_summaries->get (caller);
571 9651 : ipa_fn_summary *callee_info = ipa_fn_summaries->get (callee);
572 :
573 : /* Until GCC 4.9 we did not check the semantics-altering flags
574 : below and inlined across optimization boundaries.
575 : Enabling checks below breaks several packages by refusing
576 : to inline library always_inline functions. See PR65873.
577 : Disable the check for early inlining for now until better solution
578 : is found. */
579 9651 : if (always_inline && (flags & CAN_INLINE_EARLY))
580 : ;
581 : /* There are some options that change IL semantics which means
582 : we cannot inline in these cases for correctness reason.
583 : Not even for always_inline declared functions. */
584 7221 : else if (check_match (flag_wrapv)
585 7221 : || check_match (flag_trapv)
586 7221 : || check_match (flag_pcc_struct_return)
587 7221 : || check_maybe_down (optimize_debug)
588 : /* When caller or callee does FP math, be sure FP codegen flags
589 : compatible. */
590 7215 : || ((caller_info->fp_expressions && callee_info->fp_expressions)
591 1273 : && (check_maybe_up (flag_rounding_math)
592 1273 : || check_maybe_up (flag_trapping_math)
593 1271 : || check_maybe_down (flag_unsafe_math_optimizations)
594 1271 : || check_maybe_down (flag_finite_math_only)
595 1270 : || check_maybe_up (flag_signaling_nans)
596 1270 : || check_maybe_up (flag_complex_method)
597 1269 : || check_maybe_up (flag_signed_zeros)
598 1269 : || check_maybe_down (flag_associative_math)
599 1253 : || check_maybe_down (flag_reciprocal_math)
600 1253 : || check_maybe_down (flag_fp_int_builtin_inexact)
601 : /* Strictly speaking only when the callee contains function
602 : calls that may end up setting errno. */
603 1253 : || check_maybe_up (flag_errno_math)))
604 : /* We do not want to make code compiled with exceptions to be
605 : brought into a non-EH function unless we know that the callee
606 : does not throw.
607 : This is tracked by DECL_FUNCTION_PERSONALITY. */
608 7195 : || (check_maybe_up (flag_non_call_exceptions)
609 0 : && DECL_FUNCTION_PERSONALITY (callee->decl))
610 7195 : || (check_maybe_up (flag_exceptions)
611 16 : && DECL_FUNCTION_PERSONALITY (callee->decl))
612 : /* When devirtualization is disabled for callee, it is not safe
613 : to inline it as we possibly mangled the type info.
614 : Allow early inlining of always inlines. */
615 14416 : || (!(flags & CAN_INLINE_EARLY) && check_maybe_down (flag_devirtualize)))
616 : {
617 34 : e->inline_failed = CIF_OPTIMIZATION_MISMATCH;
618 34 : inlinable = false;
619 : }
620 : /* gcc.dg/pr43564.c. Apply user-forced inline even at -O0. */
621 7187 : else if (always_inline)
622 : ;
623 : /* When user added an attribute to the callee honor it. */
624 7187 : else if (lookup_attribute ("optimize", DECL_ATTRIBUTES (callee->decl))
625 7187 : && opts_for_fn (caller->decl) != opts_for_fn (callee->decl))
626 : {
627 2456 : e->inline_failed = CIF_OPTIMIZATION_MISMATCH;
628 2456 : inlinable = false;
629 : }
630 : /* If explicit optimize attribute are not used, the mismatch is caused
631 : by different command line options used to build different units.
632 : Do not care about COMDAT functions - those are intended to be
633 : optimized with the optimization flags of module they are used in.
634 : Also do not care about mixing up size/speed optimization when
635 : DECL_DISREGARD_INLINE_LIMITS is set. */
636 4731 : else if ((callee->merged_comdat
637 0 : && !lookup_attribute ("optimize",
638 0 : DECL_ATTRIBUTES (caller->decl)))
639 4731 : || DECL_DISREGARD_INLINE_LIMITS (callee->decl))
640 : ;
641 : /* If mismatch is caused by merging two LTO units with different
642 : optimization flags we want to be bit nicer. However never inline
643 : if one of functions is not optimized at all. */
644 4731 : else if (!opt_for_fn (callee->decl, optimize)
645 4731 : || !opt_for_fn (caller->decl, optimize))
646 : {
647 0 : e->inline_failed = CIF_OPTIMIZATION_MISMATCH;
648 0 : inlinable = false;
649 : }
650 : /* If callee is optimized for size and caller is not, allow inlining if
651 : code shrinks or we are in param_max_inline_insns_single limit and
652 : callee is inline (and thus likely an unified comdat).
653 : This will allow caller to run faster. */
654 4731 : else if (opt_for_fn (callee->decl, optimize_size)
655 4731 : > opt_for_fn (caller->decl, optimize_size))
656 : {
657 118 : int growth = estimate_edge_growth (e);
658 118 : if (growth > opt_for_fn (caller->decl, param_max_inline_insns_size)
659 118 : && (!DECL_DECLARED_INLINE_P (callee->decl)
660 65 : && growth >= MAX (inline_insns_single (caller, false, false),
661 : inline_insns_auto (caller, false, false))))
662 : {
663 0 : e->inline_failed = CIF_OPTIMIZATION_MISMATCH;
664 0 : inlinable = false;
665 : }
666 : }
667 : /* If callee is more aggressively optimized for performance than caller,
668 : we generally want to inline only cheap (runtime wise) functions. */
669 4613 : else if (opt_for_fn (callee->decl, optimize_size)
670 : < opt_for_fn (caller->decl, optimize_size)
671 4613 : || (opt_for_fn (callee->decl, optimize)
672 : > opt_for_fn (caller->decl, optimize)))
673 : {
674 12675 : if (estimate_edge_time (e)
675 4225 : >= 20 + ipa_call_summaries->get (e)->call_stmt_time)
676 : {
677 1514 : e->inline_failed = CIF_OPTIMIZATION_MISMATCH;
678 1514 : inlinable = false;
679 : }
680 : }
681 :
682 : }
683 :
684 70923 : if (!inlinable && (flags & CAN_INLINE_REPORT))
685 65111 : report_inline_failed_reason (e);
686 : return inlinable;
687 : }
688 :
689 :
690 : /* Return true if the edge E is inlinable during early inlining. */
691 :
692 : static bool
693 4131957 : can_early_inline_edge_p (struct cgraph_edge *e)
694 : {
695 4130394 : cgraph_node *caller = (e->caller->inlined_to
696 4131957 : ? e->caller->inlined_to : e->caller);
697 4131957 : struct cgraph_node *callee = e->callee->ultimate_alias_target ();
698 : /* Early inliner might get called at WPA stage when IPA pass adds new
699 : function. In this case we cannot really do any of early inlining
700 : because function bodies are missing. */
701 4131957 : if (cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR)
702 : return false;
703 4131652 : if (!gimple_has_body_p (callee->decl))
704 : {
705 0 : e->inline_failed = CIF_BODY_NOT_AVAILABLE;
706 0 : return false;
707 : }
708 8263304 : gcc_assert (gimple_in_ssa_p (DECL_STRUCT_FUNCTION (e->caller->decl))
709 : && gimple_in_ssa_p (DECL_STRUCT_FUNCTION (callee->decl)));
710 4131652 : if (coverage_instrumentation_p ()
711 4133266 : && ((lookup_attribute ("no_profile_instrument_function",
712 1614 : DECL_ATTRIBUTES (caller->decl)) == NULL_TREE)
713 1614 : != (lookup_attribute ("no_profile_instrument_function",
714 3228 : DECL_ATTRIBUTES (callee->decl)) == NULL_TREE)))
715 : return false;
716 :
717 4131650 : if (!can_inline_edge_p (e, true, true)
718 4131650 : || !can_inline_edge_by_limits_p (e, CAN_INLINE_EARLY | CAN_INLINE_REPORT))
719 87055 : return false;
720 : /* When inlining regular functions into always-inline functions
721 : during early inlining watch for possible inline cycles. */
722 4044595 : if (DECL_DISREGARD_INLINE_LIMITS (caller->decl)
723 219645 : && lookup_attribute ("always_inline", DECL_ATTRIBUTES (caller->decl))
724 4263661 : && (!DECL_DISREGARD_INLINE_LIMITS (callee->decl)
725 133669 : || !lookup_attribute ("always_inline", DECL_ATTRIBUTES (callee->decl))))
726 : {
727 : /* If there are indirect calls, inlining may produce direct call.
728 : TODO: We may lift this restriction if we avoid errors on formely
729 : indirect calls to always_inline functions. Taking address
730 : of always_inline function is generally bad idea and should
731 : have been declared as undefined, but sadly we allow this. */
732 85398 : if (caller->indirect_calls || e->callee->indirect_calls)
733 : return false;
734 84216 : ipa_fn_summary *callee_info = ipa_fn_summaries->get (callee);
735 84216 : if (callee_info->safe_to_inline_to_always_inline)
736 19882 : return callee_info->safe_to_inline_to_always_inline - 1;
737 141658 : for (cgraph_edge *e2 = callee->callees; e2; e2 = e2->next_callee)
738 : {
739 77357 : struct cgraph_node *callee2 = e2->callee->ultimate_alias_target ();
740 : /* As early inliner runs in RPO order, we will see uninlined
741 : always_inline calls only in the case of cyclic graphs. */
742 77357 : if (DECL_DISREGARD_INLINE_LIMITS (callee2->decl)
743 77357 : || lookup_attribute ("always_inline", DECL_ATTRIBUTES (callee2->decl)))
744 : {
745 0 : callee_info->safe_to_inline_to_always_inline = 1;
746 0 : return false;
747 : }
748 : /* With LTO watch for case where function is later replaced
749 : by always_inline definition.
750 : TODO: We may either stop treating noninlined cross-module always
751 : inlines as errors, or we can extend decl merging to produce
752 : syntacic alias and honor always inline only in units it has
753 : been declared as such. */
754 77357 : if (flag_lto && callee2->externally_visible)
755 : {
756 33 : callee_info->safe_to_inline_to_always_inline = 1;
757 33 : return false;
758 : }
759 : }
760 64301 : callee_info->safe_to_inline_to_always_inline = 2;
761 : }
762 : return true;
763 : }
764 :
765 :
766 : /* Return number of calls in N. Ignore cheap builtins. */
767 :
768 : static int
769 844810 : num_calls (struct cgraph_node *n)
770 : {
771 844810 : struct cgraph_edge *e;
772 844810 : int num = 0;
773 :
774 1706774 : for (e = n->callees; e; e = e->next_callee)
775 861964 : if (!is_inexpensive_builtin (e->callee->decl))
776 770405 : num++;
777 844810 : return num;
778 : }
779 :
780 :
781 : /* Return true if we are interested in inlining small function. */
782 :
783 : static bool
784 3482634 : want_early_inline_function_p (struct cgraph_edge *e)
785 : {
786 3482634 : bool want_inline = true;
787 3482634 : struct cgraph_node *callee = e->callee->ultimate_alias_target ();
788 :
789 3482634 : if (DECL_DISREGARD_INLINE_LIMITS (callee->decl))
790 : ;
791 3482627 : else if (!DECL_DECLARED_INLINE_P (callee->decl)
792 3482627 : && !opt_for_fn (e->caller->decl, flag_inline_small_functions))
793 : {
794 62 : e->inline_failed = CIF_FUNCTION_NOT_INLINE_CANDIDATE;
795 62 : report_inline_failed_reason (e);
796 62 : want_inline = false;
797 : }
798 : else
799 : {
800 : /* First take care of very large functions. */
801 3482565 : int min_growth = estimate_min_edge_growth (e), growth = 0;
802 3482565 : int n;
803 3482565 : int early_inlining_insns = param_early_inlining_insns;
804 :
805 3482565 : if (min_growth > early_inlining_insns)
806 : {
807 407167 : if (dump_enabled_p ())
808 40 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
809 : " will not early inline: %C->%C, "
810 : "call is cold and code would grow "
811 : "at least by %i\n",
812 : e->caller, callee,
813 : min_growth);
814 : want_inline = false;
815 : }
816 : else
817 3075398 : growth = estimate_edge_growth (e);
818 :
819 :
820 3075398 : if (!want_inline || growth <= param_max_inline_insns_size)
821 : ;
822 1232243 : else if (!e->maybe_hot_p ())
823 : {
824 21183 : if (dump_enabled_p ())
825 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
826 : " will not early inline: %C->%C, "
827 : "call is cold and code would grow by %i\n",
828 : e->caller, callee,
829 : growth);
830 : want_inline = false;
831 : }
832 1211060 : else if (growth > early_inlining_insns)
833 : {
834 366250 : if (dump_enabled_p ())
835 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
836 : " will not early inline: %C->%C, "
837 : "growth %i exceeds --param early-inlining-insns\n",
838 : e->caller, callee, growth);
839 : want_inline = false;
840 : }
841 844810 : else if ((n = num_calls (callee)) != 0
842 844810 : && growth * (n + 1) > early_inlining_insns)
843 : {
844 248116 : if (dump_enabled_p ())
845 11 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
846 : " will not early inline: %C->%C, "
847 : "growth %i exceeds --param early-inlining-insns "
848 : "divided by number of calls\n",
849 : e->caller, callee, growth);
850 : want_inline = false;
851 : }
852 : }
853 3482634 : return want_inline;
854 : }
855 :
856 : /* Compute time of the edge->caller + edge->callee execution when inlining
857 : does not happen. */
858 :
859 : inline sreal
860 452373 : compute_uninlined_call_time (struct cgraph_edge *edge,
861 : sreal uninlined_call_time,
862 : sreal freq)
863 : {
864 313859 : cgraph_node *caller = (edge->caller->inlined_to
865 452373 : ? edge->caller->inlined_to
866 : : edge->caller);
867 :
868 452373 : if (freq > 0)
869 438591 : uninlined_call_time *= freq;
870 : else
871 13782 : uninlined_call_time = uninlined_call_time >> 11;
872 :
873 452373 : sreal caller_time = ipa_fn_summaries->get (caller)->time;
874 452373 : return uninlined_call_time + caller_time;
875 : }
876 :
877 : /* Same as compute_uinlined_call_time but compute time when inlining
878 : does happen. */
879 :
880 : inline sreal
881 452373 : compute_inlined_call_time (struct cgraph_edge *edge,
882 : sreal time,
883 : sreal freq)
884 : {
885 313859 : cgraph_node *caller = (edge->caller->inlined_to
886 452373 : ? edge->caller->inlined_to
887 : : edge->caller);
888 452373 : sreal caller_time = ipa_fn_summaries->get (caller)->time;
889 :
890 452373 : if (freq > 0)
891 438591 : time *= freq;
892 : else
893 13782 : time = time >> 11;
894 :
895 : /* This calculation should match one in ipa-inline-analysis.cc
896 : (estimate_edge_size_and_time). */
897 452373 : time -= (sreal)ipa_call_summaries->get (edge)->call_stmt_time * freq;
898 452373 : time += caller_time;
899 452373 : if (time <= 0)
900 0 : time = ((sreal) 1) >> 8;
901 452373 : gcc_checking_assert (time >= 0);
902 452373 : return time;
903 : }
904 :
905 : /* Determine time saved by inlining EDGE of frequency FREQ
906 : where callee's runtime w/o inlining is UNINLINED_TYPE
907 : and with inlined is INLINED_TYPE. */
908 :
909 : inline sreal
910 9697221 : inlining_speedup (struct cgraph_edge *edge,
911 : sreal freq,
912 : sreal uninlined_time,
913 : sreal inlined_time)
914 : {
915 9697221 : sreal speedup = uninlined_time - inlined_time;
916 : /* Handling of call_time should match one in ipa-inline-fnsummary.c
917 : (estimate_edge_size_and_time). */
918 9697221 : sreal call_time = ipa_call_summaries->get (edge)->call_stmt_time;
919 :
920 9697221 : if (freq > 0)
921 : {
922 9661532 : speedup = (speedup + call_time);
923 11960753 : if (freq != 1)
924 7362311 : speedup = speedup * freq;
925 : }
926 35689 : else if (freq == 0)
927 35689 : speedup = speedup >> 11;
928 9697221 : gcc_checking_assert (speedup >= 0);
929 9697221 : return speedup;
930 : }
931 :
932 : /* Return expected speedup of the callee function alone
933 : (i.e. not estimate of call overhead and also no scalling
934 : by call frequency. */
935 :
936 : static sreal
937 3089458 : callee_speedup (struct cgraph_edge *e)
938 : {
939 3089458 : sreal unspec_time;
940 3089458 : sreal spec_time = estimate_edge_time (e, &unspec_time);
941 3089458 : return unspec_time - spec_time;
942 : }
943 :
944 : /* Return true if the speedup for inlining E is bigger than
945 : param_inline_min_speedup. */
946 :
947 : static bool
948 452373 : big_speedup_p (struct cgraph_edge *e)
949 : {
950 452373 : sreal unspec_time;
951 452373 : sreal spec_time = estimate_edge_time (e, &unspec_time);
952 452373 : sreal freq = e->sreal_frequency ();
953 452373 : sreal time = compute_uninlined_call_time (e, unspec_time, freq);
954 452373 : sreal inlined_time = compute_inlined_call_time (e, spec_time, freq);
955 313859 : cgraph_node *caller = (e->caller->inlined_to
956 452373 : ? e->caller->inlined_to
957 : : e->caller);
958 452373 : int limit = opt_for_fn (caller->decl, param_inline_min_speedup);
959 :
960 452373 : if ((time - inlined_time) * 100 > time * limit)
961 : return true;
962 : return false;
963 : }
964 :
965 : /* Return true if we are interested in inlining small function.
966 : When REPORT is true, report reason to dump file. */
967 :
968 : static bool
969 4929764 : want_inline_small_function_p (struct cgraph_edge *e, bool report)
970 : {
971 4929764 : bool want_inline = true;
972 4929764 : struct cgraph_node *callee = e->callee->ultimate_alias_target ();
973 3772810 : cgraph_node *to = (e->caller->inlined_to
974 4929764 : ? e->caller->inlined_to : e->caller);
975 :
976 : /* Allow this function to be called before can_inline_edge_p,
977 : since it's usually cheaper. */
978 4929764 : if (cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR)
979 : want_inline = false;
980 4929764 : else if (DECL_DISREGARD_INLINE_LIMITS (callee->decl))
981 : return true;
982 4923945 : else if (!DECL_DECLARED_INLINE_P (callee->decl)
983 4923945 : && !opt_for_fn (e->caller->decl, flag_inline_small_functions))
984 : {
985 51465 : e->inline_failed = CIF_FUNCTION_NOT_INLINE_CANDIDATE;
986 51465 : want_inline = false;
987 : }
988 :
989 : /* Early return before lookup of summaries. */
990 51465 : if (!want_inline)
991 : {
992 51465 : if (report)
993 47441 : report_inline_failed_reason (e);
994 51465 : return false;
995 : }
996 :
997 4872480 : ipa_fn_summary *callee_info = ipa_fn_summaries->get (callee);
998 4872480 : ipa_call_summary *call_info = ipa_call_summaries->get (e);
999 :
1000 : /* Do fast and conservative check if the function can be good
1001 : inline candidate. */
1002 4872480 : if ((!DECL_DECLARED_INLINE_P (callee->decl)
1003 1989143 : && (!e->count.ipa ().initialized_p ()
1004 39720 : || !e->maybe_hot_p (callee_info->time)))
1005 6861393 : && callee_info->min_size - call_info->call_stmt_size
1006 1988913 : > inline_insns_auto (e->caller, true, true))
1007 : {
1008 2293 : e->inline_failed = CIF_MAX_INLINE_INSNS_AUTO_LIMIT;
1009 2293 : want_inline = false;
1010 : }
1011 4870187 : else if ((DECL_DECLARED_INLINE_P (callee->decl)
1012 1986850 : || e->count.ipa ().nonzero_p ())
1013 7753778 : && callee_info->min_size - call_info->call_stmt_size
1014 2883591 : > inline_insns_single (e->caller, true, true))
1015 : {
1016 0 : e->inline_failed = (DECL_DECLARED_INLINE_P (callee->decl)
1017 0 : ? CIF_MAX_INLINE_INSNS_SINGLE_LIMIT
1018 : : CIF_MAX_INLINE_INSNS_AUTO_LIMIT);
1019 0 : want_inline = false;
1020 : }
1021 : else
1022 : {
1023 4870187 : int growth = estimate_edge_growth (e);
1024 4870187 : ipa_hints hints = estimate_edge_hints (e);
1025 : /* We have two independent groups of hints. If one matches in each
1026 : of groups the limits are inreased. If both groups matches, limit
1027 : is increased even more. */
1028 4870187 : bool apply_hints = (hints & (INLINE_HINT_indirect_call
1029 : | INLINE_HINT_known_hot
1030 : | INLINE_HINT_loop_iterations
1031 : | INLINE_HINT_loop_stride));
1032 4870187 : bool apply_hints2 = (hints & INLINE_HINT_builtin_constant_p);
1033 :
1034 4870187 : if (growth <= opt_for_fn (to->decl,
1035 : param_max_inline_insns_size))
1036 : ;
1037 : /* Apply param_max_inline_insns_single limit. Do not do so when
1038 : hints suggests that inlining given function is very profitable.
1039 : Avoid computation of big_speedup_p when not necessary to change
1040 : outcome of decision. */
1041 4744462 : else if (DECL_DECLARED_INLINE_P (callee->decl)
1042 2832758 : && growth >= inline_insns_single (e->caller, apply_hints,
1043 : apply_hints2)
1044 5221683 : && (apply_hints || apply_hints2
1045 474665 : || growth >= inline_insns_single (e->caller, true,
1046 : apply_hints2)
1047 241554 : || !big_speedup_p (e)))
1048 : {
1049 475956 : e->inline_failed = CIF_MAX_INLINE_INSNS_SINGLE_LIMIT;
1050 475956 : want_inline = false;
1051 : }
1052 4268506 : else if (!DECL_DECLARED_INLINE_P (callee->decl)
1053 1911704 : && !opt_for_fn (e->caller->decl, flag_inline_functions)
1054 4273866 : && growth >= opt_for_fn (to->decl,
1055 : param_max_inline_insns_small))
1056 : {
1057 : /* growth_positive_p is expensive, always test it last. */
1058 5360 : if (growth >= inline_insns_single (e->caller, false, false)
1059 5360 : || growth_positive_p (callee, e, growth))
1060 : {
1061 4973 : e->inline_failed = CIF_NOT_DECLARED_INLINED;
1062 4973 : want_inline = false;
1063 : }
1064 : }
1065 : /* Apply param_max_inline_insns_auto limit for functions not declared
1066 : inline. Bypass the limit when speedup seems big. */
1067 4263146 : else if (!DECL_DECLARED_INLINE_P (callee->decl)
1068 1906344 : && growth >= inline_insns_auto (e->caller, apply_hints,
1069 : apply_hints2)
1070 5447809 : && (apply_hints || apply_hints2
1071 1170293 : || growth >= inline_insns_auto (e->caller, true,
1072 : apply_hints2)
1073 210634 : || !big_speedup_p (e)))
1074 : {
1075 : /* growth_positive_p is expensive, always test it last. */
1076 1173688 : if (growth >= inline_insns_single (e->caller, false, false)
1077 1173688 : || growth_positive_p (callee, e, growth))
1078 : {
1079 1093632 : e->inline_failed = CIF_MAX_INLINE_INSNS_AUTO_LIMIT;
1080 1093632 : want_inline = false;
1081 : }
1082 : }
1083 : /* If call is cold, do not inline when function body would grow. */
1084 3089458 : else if (!e->maybe_hot_p (callee_speedup (e))
1085 3089458 : && (growth >= inline_insns_single (e->caller, false, false)
1086 688918 : || growth_positive_p (callee, e, growth)))
1087 : {
1088 617582 : e->inline_failed = CIF_UNLIKELY_CALL;
1089 617582 : want_inline = false;
1090 : }
1091 : }
1092 4872480 : if (!want_inline && report)
1093 574393 : report_inline_failed_reason (e);
1094 : return want_inline;
1095 : }
1096 :
1097 : /* EDGE is self recursive edge.
1098 : We handle two cases - when function A is inlining into itself
1099 : or when function A is being inlined into another inliner copy of function
1100 : A within function B.
1101 :
1102 : In first case OUTER_NODE points to the toplevel copy of A, while
1103 : in the second case OUTER_NODE points to the outermost copy of A in B.
1104 :
1105 : In both cases we want to be extra selective since
1106 : inlining the call will just introduce new recursive calls to appear. */
1107 :
1108 : static bool
1109 19673 : want_inline_self_recursive_call_p (struct cgraph_edge *edge,
1110 : struct cgraph_node *outer_node,
1111 : bool peeling,
1112 : int depth)
1113 : {
1114 19673 : char const *reason = NULL;
1115 19673 : bool want_inline = true;
1116 19673 : sreal caller_freq = 1;
1117 19673 : int max_depth = opt_for_fn (outer_node->decl,
1118 : param_max_inline_recursive_depth_auto);
1119 :
1120 19673 : if (DECL_DECLARED_INLINE_P (edge->caller->decl))
1121 2885 : max_depth = opt_for_fn (outer_node->decl,
1122 : param_max_inline_recursive_depth);
1123 :
1124 19673 : if (!edge->maybe_hot_p ())
1125 : {
1126 : reason = "recursive call is cold";
1127 : want_inline = false;
1128 : }
1129 19198 : else if (depth > max_depth)
1130 : {
1131 : reason = "--param max-inline-recursive-depth exceeded.";
1132 : want_inline = false;
1133 : }
1134 17019 : else if (outer_node->inlined_to
1135 21494 : && (caller_freq = outer_node->callers->sreal_frequency ()) == 0)
1136 : {
1137 0 : reason = "caller frequency is 0";
1138 0 : want_inline = false;
1139 : }
1140 :
1141 0 : if (!want_inline)
1142 : ;
1143 : /* Inlining of self recursive function into copy of itself within other
1144 : function is transformation similar to loop peeling.
1145 :
1146 : Peeling is profitable if we can inline enough copies to make probability
1147 : of actual call to the self recursive function very small. Be sure that
1148 : the probability of recursion is small.
1149 :
1150 : We ensure that the frequency of recursing is at most 1 - (1/max_depth).
1151 : This way the expected number of recursion is at most max_depth. */
1152 17019 : else if (peeling)
1153 : {
1154 4475 : sreal max_prob = (sreal)1 - ((sreal)1 / (sreal)max_depth);
1155 4475 : int i;
1156 9928 : for (i = 1; i < depth; i++)
1157 5453 : max_prob = max_prob * max_prob;
1158 4475 : if (edge->sreal_frequency () >= max_prob * caller_freq)
1159 : {
1160 1559 : reason = "frequency of recursive call is too large";
1161 1559 : want_inline = false;
1162 : }
1163 : }
1164 : /* Recursive inlining, i.e. equivalent of unrolling, is profitable if
1165 : recursion depth is large. We reduce function call overhead and increase
1166 : chances that things fit in hardware return predictor.
1167 :
1168 : Recursive inlining might however increase cost of stack frame setup
1169 : actually slowing down functions whose recursion tree is wide rather than
1170 : deep.
1171 :
1172 : Deciding reliably on when to do recursive inlining without profile feedback
1173 : is tricky. For now we disable recursive inlining when probability of self
1174 : recursion is low.
1175 :
1176 : Recursive inlining of self recursive call within loop also results in
1177 : large loop depths that generally optimize badly. We may want to throttle
1178 : down inlining in those cases. In particular this seems to happen in one
1179 : of libstdc++ rb tree methods. */
1180 : else
1181 : {
1182 12544 : if (edge->sreal_frequency () * 100
1183 12544 : <= caller_freq
1184 25088 : * opt_for_fn (outer_node->decl,
1185 : param_min_inline_recursive_probability))
1186 : {
1187 794 : reason = "frequency of recursive call is too small";
1188 794 : want_inline = false;
1189 : }
1190 : }
1191 19673 : if (!can_inline_edge_by_limits_p (edge, CAN_INLINE_FORCE_LIMITS | CAN_INLINE_REPORT))
1192 : {
1193 : reason = "inline limits exceeded for always_inline function";
1194 : want_inline = false;
1195 : }
1196 19673 : if (!want_inline && dump_enabled_p ())
1197 7 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, edge->call_stmt,
1198 : " not inlining recursively: %s\n", reason);
1199 19673 : return want_inline;
1200 : }
1201 :
1202 : /* Return true when NODE has uninlinable caller;
1203 : set HAS_HOT_CALL if it has hot call.
1204 : Worker for cgraph_for_node_and_aliases. */
1205 :
1206 : static bool
1207 73655 : check_callers (struct cgraph_node *node, void *has_hot_call)
1208 : {
1209 73655 : struct cgraph_edge *e;
1210 129572 : for (e = node->callers; e; e = e->next_caller)
1211 : {
1212 82521 : if (!opt_for_fn (e->caller->decl, flag_inline_functions_called_once)
1213 82521 : || !opt_for_fn (e->caller->decl, optimize))
1214 : return true;
1215 82521 : if (!can_inline_edge_p (e, true))
1216 : return true;
1217 82494 : if (e->recursive_p ())
1218 : return true;
1219 82494 : if (!can_inline_edge_by_limits_p (e, CAN_INLINE_REPORT))
1220 : return true;
1221 : /* Inlining large functions to large loop depth is often harmful because
1222 : of register pressure it implies. */
1223 55969 : if ((int)ipa_call_summaries->get (e)->loop_depth
1224 55969 : > param_inline_functions_called_once_loop_depth)
1225 : return true;
1226 : /* Do not produce gigantic functions. */
1227 102643 : if (estimate_size_after_inlining (e->caller->inlined_to ?
1228 : e->caller->inlined_to : e->caller, e)
1229 55969 : > param_inline_functions_called_once_insns)
1230 : return true;
1231 55917 : if (!(*(bool *)has_hot_call) && e->maybe_hot_p ())
1232 10702 : *(bool *)has_hot_call = true;
1233 : }
1234 : return false;
1235 : }
1236 :
1237 : /* If NODE has a caller, return true. */
1238 :
1239 : static bool
1240 2159110 : has_caller_p (struct cgraph_node *node, void *data ATTRIBUTE_UNUSED)
1241 : {
1242 2159110 : if (node->callers)
1243 676963 : return true;
1244 : return false;
1245 : }
1246 :
1247 : /* Decide if inlining NODE would reduce unit size by eliminating
1248 : the offline copy of function.
1249 : When COLD is true the cold calls are considered, too. */
1250 :
1251 : static bool
1252 4782128 : want_inline_function_to_all_callers_p (struct cgraph_node *node, bool cold)
1253 : {
1254 4782128 : bool has_hot_call = false;
1255 :
1256 : /* Aliases gets inlined along with the function they alias. */
1257 4782128 : if (node->alias)
1258 : return false;
1259 : /* Already inlined? */
1260 4704819 : if (node->inlined_to)
1261 : return false;
1262 : /* Does it have callers? */
1263 2109389 : if (!node->call_for_symbol_and_aliases (has_caller_p, NULL, true))
1264 : return false;
1265 : /* Inlining into all callers would increase size? */
1266 676963 : if (growth_positive_p (node, NULL, INT_MIN) > 0)
1267 : return false;
1268 : /* All inlines must be possible. */
1269 68969 : if (node->call_for_symbol_and_aliases (check_callers, &has_hot_call,
1270 : true))
1271 : return false;
1272 42365 : if (!cold && !has_hot_call)
1273 : return false;
1274 : return true;
1275 : }
1276 :
1277 : /* Return true if WHERE of SIZE is a possible candidate for wrapper heuristics
1278 : in estimate_edge_badness. */
1279 :
1280 : static bool
1281 619177 : wrapper_heuristics_may_apply (struct cgraph_node *where, int size)
1282 : {
1283 619177 : return size < (DECL_DECLARED_INLINE_P (where->decl)
1284 619177 : ? inline_insns_single (where, false, false)
1285 363559 : : inline_insns_auto (where, false, false));
1286 : }
1287 :
1288 : /* A cost model driving the inlining heuristics in a way so the edges with
1289 : smallest badness are inlined first. After each inlining is performed
1290 : the costs of all caller edges of nodes affected are recomputed so the
1291 : metrics may accurately depend on values such as number of inlinable callers
1292 : of the function or function body size. */
1293 :
1294 : static sreal
1295 10227748 : edge_badness (struct cgraph_edge *edge, bool dump)
1296 : {
1297 10227748 : sreal badness;
1298 10227748 : int growth;
1299 10227748 : sreal edge_time, unspec_edge_time;
1300 10227748 : struct cgraph_node *callee = edge->callee->ultimate_alias_target ();
1301 10227748 : class ipa_fn_summary *callee_info = ipa_fn_summaries->get (callee);
1302 10227748 : ipa_hints hints;
1303 8180002 : cgraph_node *caller = (edge->caller->inlined_to
1304 10227748 : ? edge->caller->inlined_to
1305 : : edge->caller);
1306 :
1307 10227748 : growth = estimate_edge_growth (edge);
1308 10227748 : edge_time = estimate_edge_time (edge, &unspec_edge_time);
1309 10227748 : hints = estimate_edge_hints (edge);
1310 10227748 : gcc_checking_assert (edge_time >= 0);
1311 : /* Check that inlined time is better, but tolerate some roundoff issues.
1312 : FIXME: When callee profile drops to 0 we account calls more. This
1313 : should be fixed by never doing that. */
1314 10227748 : gcc_checking_assert ((edge_time * 100
1315 : - callee_info->time * 101).to_int () <= 0
1316 : || callee->count.ipa ().initialized_p ());
1317 10227748 : gcc_checking_assert (growth <= ipa_size_summaries->get (callee)->size);
1318 :
1319 10227748 : if (dump)
1320 : {
1321 185 : fprintf (dump_file, " Badness calculation for %s -> %s\n",
1322 185 : edge->caller->dump_name (),
1323 185 : edge->callee->dump_name ());
1324 185 : fprintf (dump_file, " size growth %i, time %f unspec %f ",
1325 : growth,
1326 : edge_time.to_double (),
1327 : unspec_edge_time.to_double ());
1328 185 : ipa_dump_hints (dump_file, hints);
1329 185 : if (big_speedup_p (edge))
1330 151 : fprintf (dump_file, " big_speedup");
1331 185 : fprintf (dump_file, "\n");
1332 : }
1333 :
1334 : /* Always prefer inlining saving code size. */
1335 10227748 : if (growth <= 0)
1336 : {
1337 504834 : badness = (sreal) (-SREAL_MIN_SIG + growth) << (SREAL_MAX_EXP / 256);
1338 504834 : if (dump)
1339 108 : fprintf (dump_file, " %f: Growth %d <= 0\n", badness.to_double (),
1340 : growth);
1341 : }
1342 : /* Inlining into EXTERNAL functions is not going to change anything unless
1343 : they are themselves inlined. */
1344 9722914 : else if (DECL_EXTERNAL (caller->decl))
1345 : {
1346 24448 : if (dump)
1347 0 : fprintf (dump_file, " max: function is external\n");
1348 24448 : return sreal::max ();
1349 : }
1350 : /* When profile is available. Compute badness as:
1351 :
1352 : time_saved * caller_count
1353 : goodness = -------------------------------------------------
1354 : growth_of_caller * overall_growth * combined_size
1355 :
1356 : badness = - goodness
1357 :
1358 : Again use negative value to make calls with profile appear hotter
1359 : then calls without.
1360 : */
1361 9698466 : else if (opt_for_fn (caller->decl, flag_guess_branch_prob)
1362 9698466 : || caller->count.ipa ().nonzero_p ())
1363 : {
1364 9697144 : sreal numerator, denominator;
1365 9697144 : int overall_growth;
1366 9697144 : sreal freq = edge->sreal_frequency ();
1367 :
1368 9697144 : numerator = inlining_speedup (edge, freq, unspec_edge_time, edge_time);
1369 9697144 : if (numerator <= 0)
1370 4839 : numerator = ((sreal) 1 >> 8);
1371 9697144 : if (caller->count.ipa ().nonzero_p ())
1372 179 : numerator *= caller->count.ipa ().to_gcov_type ();
1373 9696965 : else if (caller->count.ipa ().initialized_p ())
1374 757 : numerator = numerator >> 11;
1375 9697144 : denominator = growth;
1376 :
1377 9697144 : overall_growth = callee_info->growth;
1378 :
1379 : /* Look for inliner wrappers of the form:
1380 :
1381 : inline_caller ()
1382 : {
1383 : do_fast_job...
1384 : if (need_more_work)
1385 : noninline_callee ();
1386 : }
1387 : Without penalizing this case, we usually inline noninline_callee
1388 : into the inline_caller because overall_growth is small preventing
1389 : further inlining of inline_caller.
1390 :
1391 : Penalize only callgraph edges to functions with small overall
1392 : growth ...
1393 : */
1394 9697144 : if (growth > overall_growth
1395 : /* ... and having only one caller which is not inlined ... */
1396 2097544 : && callee_info->single_caller
1397 1167901 : && !edge->caller->inlined_to
1398 : /* ... and edges executed only conditionally ... */
1399 815669 : && freq < 1
1400 : /* ... consider case where callee is not inline but caller is ... */
1401 10065342 : && ((!DECL_DECLARED_INLINE_P (edge->callee->decl)
1402 94672 : && DECL_DECLARED_INLINE_P (caller->decl))
1403 : /* ... or when early optimizers decided to split and edge
1404 : frequency still indicates splitting is a win ... */
1405 360778 : || (callee->split_part && !caller->split_part
1406 89897 : && freq * 100
1407 9780876 : < opt_for_fn (caller->decl,
1408 : param_partial_inlining_entry_probability)
1409 : /* ... and do not overwrite user specified hints. */
1410 89432 : && (!DECL_DECLARED_INLINE_P (edge->callee->decl)
1411 66385 : || DECL_DECLARED_INLINE_P (caller->decl)))))
1412 : {
1413 95597 : ipa_fn_summary *caller_info = ipa_fn_summaries->get (caller);
1414 95597 : int caller_growth = caller_info->growth;
1415 :
1416 : /* Only apply the penalty when caller looks like inline candidate,
1417 : and it is not called once. */
1418 52749 : if (!caller_info->single_caller && overall_growth < caller_growth
1419 50521 : && caller_info->inlinable
1420 146064 : && wrapper_heuristics_may_apply
1421 50467 : (caller, ipa_size_summaries->get (caller)->size))
1422 : {
1423 40821 : if (dump)
1424 1 : fprintf (dump_file,
1425 : " Wrapper penalty. Increasing growth %i to %i\n",
1426 : overall_growth, caller_growth);
1427 : overall_growth = caller_growth;
1428 : }
1429 : }
1430 9697144 : if (overall_growth > 0)
1431 : {
1432 : /* Strongly prefer functions with few callers that can be inlined
1433 : fully. The square root here leads to smaller binaries at average.
1434 : Watch however for extreme cases and return to linear function
1435 : when growth is large. */
1436 8323658 : if (overall_growth < 256)
1437 4319235 : overall_growth *= overall_growth;
1438 : else
1439 4004423 : overall_growth += 256 * 256 - 256;
1440 8323658 : denominator *= overall_growth;
1441 : }
1442 9697144 : denominator *= ipa_size_summaries->get (caller)->size + growth;
1443 :
1444 9697144 : badness = - numerator / denominator;
1445 :
1446 9697144 : if (dump)
1447 : {
1448 308 : fprintf (dump_file,
1449 : " %f: guessed profile. frequency %f, count %" PRId64
1450 : " caller count %" PRId64
1451 : " time saved %f"
1452 : " overall growth %i (current) %i (original)"
1453 : " %i (compensated)\n",
1454 : badness.to_double (),
1455 : freq.to_double (),
1456 77 : edge->count.ipa ().initialized_p ()
1457 0 : ? edge->count.ipa ().to_gcov_type () : -1,
1458 77 : caller->count.ipa ().initialized_p ()
1459 0 : ? caller->count.ipa ().to_gcov_type () : -1,
1460 154 : inlining_speedup (edge, freq, unspec_edge_time,
1461 : edge_time).to_double (),
1462 : estimate_growth (callee),
1463 : callee_info->growth, overall_growth);
1464 : }
1465 : }
1466 : /* When function local profile is not available or it does not give
1467 : useful information (i.e. frequency is zero), base the cost on
1468 : loop nest and overall size growth, so we optimize for overall number
1469 : of functions fully inlined in program. */
1470 : else
1471 : {
1472 1322 : int nest = MIN (ipa_call_summaries->get (edge)->loop_depth, 8);
1473 1322 : badness = growth;
1474 :
1475 : /* Decrease badness if call is nested. */
1476 1322 : if (badness > 0)
1477 1322 : badness = badness >> nest;
1478 : else
1479 0 : badness = badness << nest;
1480 1322 : if (dump)
1481 0 : fprintf (dump_file, " %f: no profile. nest %i\n",
1482 : badness.to_double (), nest);
1483 : }
1484 10203300 : gcc_checking_assert (badness != 0);
1485 :
1486 10203300 : if (edge->recursive_p ())
1487 18729 : badness = badness.shift (badness > 0 ? 4 : -4);
1488 10203300 : if ((hints & (INLINE_HINT_indirect_call
1489 : | INLINE_HINT_loop_iterations
1490 : | INLINE_HINT_loop_stride))
1491 9382948 : || callee_info->growth <= 0)
1492 4709440 : badness = badness.shift (badness > 0 ? -2 : 2);
1493 10203300 : if (hints & INLINE_HINT_builtin_constant_p)
1494 20986 : badness = badness.shift (badness > 0 ? -4 : 4);
1495 10203300 : if (hints & (INLINE_HINT_same_scc))
1496 60344 : badness = badness.shift (badness > 0 ? 3 : -3);
1497 10173127 : else if (hints & (INLINE_HINT_in_scc))
1498 113944 : badness = badness.shift (badness > 0 ? 2 : -2);
1499 10116155 : else if (hints & (INLINE_HINT_cross_module))
1500 3582 : badness = badness.shift (badness > 0 ? 1 : -1);
1501 10203300 : if (DECL_DISREGARD_INLINE_LIMITS (callee->decl))
1502 17448 : badness = badness.shift (badness > 0 ? -4 : 4);
1503 10194576 : else if ((hints & INLINE_HINT_declared_inline))
1504 15492071 : badness = badness.shift (badness > 0 ? -3 : 3);
1505 10203300 : if (dump)
1506 185 : fprintf (dump_file, " Adjusted by hints %f\n", badness.to_double ());
1507 10203300 : return badness;
1508 : }
1509 :
1510 : /* Recompute badness of EDGE and update its key in HEAP if needed. */
1511 : static inline void
1512 5220578 : update_edge_key (edge_heap_t *heap, struct cgraph_edge *edge)
1513 : {
1514 5220578 : sreal badness = edge_badness (edge, false);
1515 5220578 : if (edge->aux)
1516 : {
1517 4170963 : edge_heap_node_t *n = (edge_heap_node_t *) edge->aux;
1518 4170963 : gcc_checking_assert (n->get_data () == edge);
1519 :
1520 : /* fibonacci_heap::replace_key does busy updating of the
1521 : heap that is unnecessarily expensive.
1522 : We do lazy increases: after extracting minimum if the key
1523 : turns out to be out of date, it is re-inserted into heap
1524 : with correct value. */
1525 4170963 : if (badness < n->get_key ().badness)
1526 : {
1527 891882 : if (dump_file && (dump_flags & TDF_DETAILS))
1528 : {
1529 90 : fprintf (dump_file,
1530 : " decreasing badness %s -> %s, %f to %f\n",
1531 45 : edge->caller->dump_name (),
1532 45 : edge->callee->dump_name (),
1533 90 : n->get_key ().badness.to_double (),
1534 : badness.to_double ());
1535 : }
1536 891882 : inline_badness b (edge, badness);
1537 891882 : heap->decrease_key (n, b);
1538 : }
1539 : }
1540 : else
1541 : {
1542 1049615 : if (dump_file && (dump_flags & TDF_DETAILS))
1543 : {
1544 338 : fprintf (dump_file,
1545 : " enqueuing call %s -> %s, badness %f\n",
1546 169 : edge->caller->dump_name (),
1547 169 : edge->callee->dump_name (),
1548 : badness.to_double ());
1549 : }
1550 1049615 : inline_badness b (edge, badness);
1551 1049615 : edge->aux = heap->insert (b, edge);
1552 : }
1553 5220578 : }
1554 :
1555 :
1556 : /* NODE was inlined.
1557 : All caller edges needs to be reset because
1558 : size estimates change. Similarly callees needs reset
1559 : because better context may be known. */
1560 :
1561 : static void
1562 923057 : reset_edge_caches (struct cgraph_node *node)
1563 : {
1564 923057 : struct cgraph_edge *edge;
1565 923057 : struct cgraph_edge *e = node->callees;
1566 923057 : struct cgraph_node *where = node;
1567 923057 : struct ipa_ref *ref;
1568 :
1569 923057 : if (where->inlined_to)
1570 861621 : where = where->inlined_to;
1571 :
1572 923057 : reset_node_cache (where);
1573 :
1574 923057 : if (edge_growth_cache != NULL)
1575 3063434 : for (edge = where->callers; edge; edge = edge->next_caller)
1576 2141858 : if (edge->inline_failed)
1577 2141858 : edge_growth_cache->remove (edge);
1578 :
1579 979464 : FOR_EACH_ALIAS (where, ref)
1580 112814 : reset_edge_caches (dyn_cast <cgraph_node *> (ref->referring));
1581 :
1582 923057 : if (!e)
1583 : return;
1584 :
1585 2946087 : while (true)
1586 2946087 : if (!e->inline_failed && e->callee->callees)
1587 : e = e->callee->callees;
1588 : else
1589 : {
1590 2396143 : if (edge_growth_cache != NULL && e->inline_failed)
1591 2196114 : edge_growth_cache->remove (e);
1592 2396143 : if (e->next_callee)
1593 : e = e->next_callee;
1594 : else
1595 : {
1596 1317106 : do
1597 : {
1598 1317106 : if (e->caller == node)
1599 : return;
1600 549944 : e = e->caller->callers;
1601 : }
1602 549944 : while (!e->next_callee);
1603 : e = e->next_callee;
1604 : }
1605 : }
1606 : }
1607 :
1608 : /* Recompute HEAP nodes for each of caller of NODE.
1609 : UPDATED_NODES track nodes we already visited, to avoid redundant work.
1610 : When CHECK_INLINABLITY_FOR is set, re-check for specified edge that
1611 : it is inlinable. Otherwise check all edges. */
1612 :
1613 : static void
1614 921270 : update_caller_keys (edge_heap_t *heap, struct cgraph_node *node,
1615 : bitmap updated_nodes,
1616 : struct cgraph_edge *check_inlinablity_for)
1617 : {
1618 921270 : struct cgraph_edge *edge;
1619 921270 : struct ipa_ref *ref;
1620 :
1621 921270 : if ((!node->alias && !ipa_fn_summaries->get (node)->inlinable)
1622 908723 : || node->inlined_to)
1623 12547 : return;
1624 908723 : if (!bitmap_set_bit (updated_nodes, node->get_summary_id ()))
1625 : return;
1626 :
1627 964629 : FOR_EACH_ALIAS (node, ref)
1628 : {
1629 55906 : struct cgraph_node *alias = dyn_cast <cgraph_node *> (ref->referring);
1630 55906 : update_caller_keys (heap, alias, updated_nodes, check_inlinablity_for);
1631 : }
1632 :
1633 2941192 : for (edge = node->callers; edge; edge = edge->next_caller)
1634 2032469 : if (edge->inline_failed)
1635 : {
1636 2032469 : if (!check_inlinablity_for
1637 2032469 : || check_inlinablity_for == edge)
1638 : {
1639 2032469 : if (can_inline_edge_p (edge, false)
1640 1998355 : && want_inline_small_function_p (edge, false)
1641 2622209 : && can_inline_edge_by_limits_p (edge, 0))
1642 584585 : update_edge_key (heap, edge);
1643 1447884 : else if (edge->aux)
1644 : {
1645 91937 : report_inline_failed_reason (edge);
1646 91937 : heap->delete_node ((edge_heap_node_t *) edge->aux);
1647 91937 : edge->aux = NULL;
1648 : }
1649 : }
1650 0 : else if (edge->aux)
1651 0 : update_edge_key (heap, edge);
1652 : }
1653 : }
1654 :
1655 : /* Recompute HEAP nodes for each uninlined call in NODE
1656 : If UPDATE_SINCE is non-NULL check if edges called within that function
1657 : are inlinable (typically UPDATE_SINCE is the inline clone we introduced
1658 : where all edges have new context).
1659 :
1660 : This is used when we know that edge badnesses are going only to increase
1661 : (we introduced new call site) and thus all we need is to insert newly
1662 : created edges into heap. */
1663 :
1664 : static void
1665 865421 : update_callee_keys (edge_heap_t *heap, struct cgraph_node *node,
1666 : struct cgraph_node *update_since,
1667 : bitmap updated_nodes)
1668 : {
1669 865421 : struct cgraph_edge *e = node->callees;
1670 865421 : bool check_inlinability = update_since == node;
1671 :
1672 865421 : if (!e)
1673 : return;
1674 25631198 : while (true)
1675 25631198 : if (!e->inline_failed && e->callee->callees)
1676 : {
1677 3985607 : if (e->callee == update_since)
1678 398478 : check_inlinability = true;
1679 : e = e->callee->callees;
1680 : }
1681 : else
1682 : {
1683 21645591 : enum availability avail;
1684 21645591 : struct cgraph_node *callee;
1685 21645591 : if (!check_inlinability)
1686 : {
1687 19428709 : if (e->aux
1688 22844960 : && !bitmap_bit_p (updated_nodes,
1689 3416251 : e->callee->ultimate_alias_target
1690 3416251 : (&avail, e->caller)->get_summary_id ()))
1691 3416244 : update_edge_key (heap, e);
1692 : }
1693 : /* We do not reset callee growth cache here. Since we added a new call,
1694 : growth should have just increased and consequently badness metric
1695 : don't need updating. */
1696 2216882 : else if (e->inline_failed
1697 2114410 : && (callee = e->callee->ultimate_alias_target (&avail,
1698 2114410 : e->caller))
1699 2114410 : && avail >= AVAIL_AVAILABLE
1700 536340 : && ipa_fn_summaries->get (callee) != NULL
1701 536335 : && ipa_fn_summaries->get (callee)->inlinable
1702 2738956 : && !bitmap_bit_p (updated_nodes, callee->get_summary_id ()))
1703 : {
1704 522074 : if (can_inline_edge_p (e, false)
1705 516907 : && want_inline_small_function_p (e, false)
1706 823529 : && can_inline_edge_by_limits_p (e, 0))
1707 : {
1708 300798 : gcc_checking_assert (check_inlinability || can_inline_edge_p (e, false));
1709 300798 : gcc_checking_assert (check_inlinability || e->aux);
1710 300798 : update_edge_key (heap, e);
1711 : }
1712 221276 : else if (e->aux)
1713 : {
1714 6649 : report_inline_failed_reason (e);
1715 6649 : heap->delete_node ((edge_heap_node_t *) e->aux);
1716 6649 : e->aux = NULL;
1717 : }
1718 : }
1719 : /* In case we redirected to unreachable node we only need to remove the
1720 : fibheap entry. */
1721 1694808 : else if (e->aux)
1722 : {
1723 3028 : heap->delete_node ((edge_heap_node_t *) e->aux);
1724 3028 : e->aux = NULL;
1725 : }
1726 21645591 : if (e->next_callee)
1727 : e = e->next_callee;
1728 : else
1729 : {
1730 4815998 : do
1731 : {
1732 4815998 : if (e->caller == node)
1733 830391 : return;
1734 3985607 : if (e->caller == update_since)
1735 398478 : check_inlinability = false;
1736 3985607 : e = e->caller->callers;
1737 : }
1738 3985607 : while (!e->next_callee);
1739 : e = e->next_callee;
1740 : }
1741 : }
1742 : }
1743 :
1744 : /* Enqueue all recursive calls from NODE into priority queue depending on
1745 : how likely we want to recursively inline the call. */
1746 :
1747 : static void
1748 21289 : lookup_recursive_calls (struct cgraph_node *node, struct cgraph_node *where,
1749 : edge_heap_t *heap)
1750 : {
1751 21289 : struct cgraph_edge *e;
1752 21289 : enum availability avail;
1753 :
1754 59805 : for (e = where->callees; e; e = e->next_callee)
1755 38516 : if (e->callee == node
1756 38516 : || (e->callee->ultimate_alias_target (&avail, e->caller) == node
1757 1142 : && avail > AVAIL_INTERPOSABLE))
1758 : {
1759 16048 : inline_badness b (e, -e->sreal_frequency ());
1760 16048 : heap->insert (b, e);
1761 : }
1762 59805 : for (e = where->callees; e; e = e->next_callee)
1763 38516 : if (!e->inline_failed)
1764 7578 : lookup_recursive_calls (node, e->callee, heap);
1765 21289 : }
1766 :
1767 : /* Decide on recursive inlining: in the case function has recursive calls,
1768 : inline until body size reaches given argument. If any new indirect edges
1769 : are discovered in the process, add them to *NEW_EDGES, unless NEW_EDGES
1770 : is NULL. */
1771 :
1772 : static bool
1773 1961 : recursive_inlining (struct cgraph_edge *edge,
1774 : vec<cgraph_edge *> *new_edges)
1775 : {
1776 1402 : cgraph_node *to = (edge->caller->inlined_to
1777 1961 : ? edge->caller->inlined_to : edge->caller);
1778 1961 : int limit = opt_for_fn (to->decl,
1779 : param_max_inline_insns_recursive_auto);
1780 1961 : inline_badness b (edge, sreal::min ());
1781 1961 : edge_heap_t heap (b);
1782 1961 : struct cgraph_node *node;
1783 1961 : struct cgraph_edge *e;
1784 1961 : struct cgraph_node *master_clone = NULL, *next;
1785 1961 : int depth = 0;
1786 1961 : int n = 0;
1787 :
1788 1961 : node = edge->caller;
1789 1961 : if (node->inlined_to)
1790 559 : node = node->inlined_to;
1791 :
1792 1961 : if (DECL_DECLARED_INLINE_P (node->decl))
1793 417 : limit = opt_for_fn (to->decl, param_max_inline_insns_recursive);
1794 :
1795 : /* Make sure that function is small enough to be considered for inlining. */
1796 1961 : if (estimate_size_after_inlining (node, edge) >= limit)
1797 : return false;
1798 1961 : lookup_recursive_calls (node, node, &heap);
1799 1961 : if (heap.empty ())
1800 : return false;
1801 :
1802 1961 : if (dump_file)
1803 4 : fprintf (dump_file,
1804 : " Performing recursive inlining on %s\n", node->dump_name ());
1805 :
1806 : /* Do the inlining and update list of recursive call during process. */
1807 16745 : while (!heap.empty ())
1808 : {
1809 14807 : struct cgraph_edge *curr = heap.extract_min ();
1810 14807 : struct cgraph_node *cnode, *dest = curr->callee;
1811 :
1812 14807 : if (!can_inline_edge_p (curr, true)
1813 14807 : || !can_inline_edge_by_limits_p (curr, CAN_INLINE_REPORT | CAN_INLINE_FORCE_LIMITS))
1814 0 : continue;
1815 :
1816 : /* MASTER_CLONE is produced in the case we already started modified
1817 : the function. Be sure to redirect edge to the original body before
1818 : estimating growths otherwise we will be seeing growths after inlining
1819 : the already modified body. */
1820 14807 : if (master_clone)
1821 : {
1822 12740 : curr->redirect_callee (master_clone);
1823 12740 : if (edge_growth_cache != NULL)
1824 12740 : edge_growth_cache->remove (curr);
1825 : }
1826 :
1827 14807 : if (estimate_size_after_inlining (node, curr) > limit)
1828 : {
1829 23 : curr->redirect_callee (dest);
1830 23 : if (edge_growth_cache != NULL)
1831 23 : edge_growth_cache->remove (curr);
1832 : break;
1833 : }
1834 :
1835 14784 : depth = 1;
1836 14784 : for (cnode = curr->caller;
1837 77024 : cnode->inlined_to; cnode = cnode->callers->caller)
1838 124480 : if (node->decl
1839 62240 : == curr->callee->ultimate_alias_target ()->decl)
1840 62240 : depth++;
1841 :
1842 14784 : if (!want_inline_self_recursive_call_p (curr, node, false, depth))
1843 : {
1844 3034 : curr->redirect_callee (dest);
1845 3034 : if (edge_growth_cache != NULL)
1846 3034 : edge_growth_cache->remove (curr);
1847 3034 : continue;
1848 : }
1849 :
1850 11750 : if (dump_file)
1851 : {
1852 14 : fprintf (dump_file,
1853 : " Inlining call of depth %i", depth);
1854 28 : if (node->count.nonzero_p () && curr->count.initialized_p ())
1855 : {
1856 2 : fprintf (dump_file, " called approx. %.2f times per call",
1857 2 : (double)curr->count.to_gcov_type ()
1858 2 : / node->count.to_gcov_type ());
1859 : }
1860 14 : fprintf (dump_file, "\n");
1861 : }
1862 11750 : if (!master_clone)
1863 : {
1864 : /* We need original clone to copy around. */
1865 1609 : master_clone = node->create_clone (node->decl, node->count,
1866 1609 : false, vNULL, true, NULL, NULL, NULL);
1867 4587 : for (e = master_clone->callees; e; e = e->next_callee)
1868 2978 : if (!e->inline_failed)
1869 488 : clone_inlined_nodes (e, true, true, false, NULL);
1870 1609 : curr->redirect_callee (master_clone);
1871 1609 : if (edge_growth_cache != NULL)
1872 1609 : edge_growth_cache->remove (curr);
1873 : }
1874 :
1875 11750 : inline_call (curr, false, new_edges, &overall_size, true);
1876 11750 : reset_node_cache (node);
1877 11750 : lookup_recursive_calls (node, curr->callee, &heap);
1878 11750 : n++;
1879 : }
1880 :
1881 1961 : if (!heap.empty () && dump_file)
1882 0 : fprintf (dump_file, " Recursive inlining growth limit met.\n");
1883 :
1884 1961 : if (!master_clone)
1885 : return false;
1886 :
1887 1609 : if (dump_enabled_p ())
1888 4 : dump_printf_loc (MSG_NOTE, edge->call_stmt,
1889 : "\n Inlined %i times, "
1890 : "body grown from size %i to %i, time %f to %f\n", n,
1891 4 : ipa_size_summaries->get (master_clone)->size,
1892 4 : ipa_size_summaries->get (node)->size,
1893 4 : ipa_fn_summaries->get (master_clone)->time.to_double (),
1894 4 : ipa_fn_summaries->get (node)->time.to_double ());
1895 :
1896 : /* Remove master clone we used for inlining. We rely that clones inlined
1897 : into master clone gets queued just before master clone so we don't
1898 : need recursion. */
1899 18672 : for (node = symtab->first_function (); node != master_clone;
1900 : node = next)
1901 : {
1902 17063 : next = symtab->next_function (node);
1903 17063 : if (node->inlined_to == master_clone)
1904 865 : node->remove ();
1905 : }
1906 1609 : master_clone->remove ();
1907 1609 : return true;
1908 1961 : }
1909 :
1910 :
1911 : /* Given whole compilation unit estimate of INSNS, compute how large we can
1912 : allow the unit to grow. */
1913 :
1914 : static int64_t
1915 944136 : compute_max_insns (cgraph_node *node, int insns)
1916 : {
1917 944136 : int max_insns = insns;
1918 944136 : if (max_insns < opt_for_fn (node->decl, param_large_unit_insns))
1919 : max_insns = opt_for_fn (node->decl, param_large_unit_insns);
1920 :
1921 944136 : return ((int64_t) max_insns
1922 944136 : * (100 + opt_for_fn (node->decl, param_inline_unit_growth)) / 100);
1923 : }
1924 :
1925 :
1926 : /* Compute badness of all edges in NEW_EDGES and add them to the HEAP. */
1927 :
1928 : static void
1929 863205 : add_new_edges_to_heap (edge_heap_t *heap, vec<cgraph_edge *> &new_edges)
1930 : {
1931 1729015 : while (new_edges.length () > 0)
1932 : {
1933 2605 : struct cgraph_edge *edge = new_edges.pop ();
1934 :
1935 2605 : gcc_assert (!edge->aux);
1936 2605 : gcc_assert (edge->callee);
1937 2605 : if (edge->inline_failed
1938 2605 : && can_inline_edge_p (edge, true)
1939 1148 : && want_inline_small_function_p (edge, true)
1940 3423 : && can_inline_edge_by_limits_p (edge, CAN_INLINE_REPORT))
1941 : {
1942 818 : inline_badness b (edge, edge_badness (edge, false));
1943 818 : edge->aux = heap->insert (b, edge);
1944 : }
1945 : }
1946 863205 : }
1947 :
1948 : /* Remove EDGE from the fibheap. */
1949 :
1950 : static void
1951 9014 : heap_edge_removal_hook (struct cgraph_edge *e, void *data)
1952 : {
1953 9014 : if (e->aux)
1954 : {
1955 98 : ((edge_heap_t *)data)->delete_node ((edge_heap_node_t *)e->aux);
1956 98 : e->aux = NULL;
1957 : }
1958 9014 : }
1959 :
1960 : /* Return true if speculation of edge E seems useful.
1961 : If ANTICIPATE_INLINING is true, be conservative and hope that E
1962 : may get inlined. */
1963 :
1964 : bool
1965 118594 : speculation_useful_p (struct cgraph_edge *e, bool anticipate_inlining)
1966 : {
1967 : /* If we have already decided to inline the edge, it seems useful.
1968 : Also if ipa-cp or other pass worked hard enough to produce a clone,
1969 : we already decided this is a good idea. */
1970 118594 : if (!e->inline_failed
1971 34229 : || e->callee->clone_of)
1972 : return true;
1973 :
1974 32181 : enum availability avail;
1975 32181 : struct cgraph_node *target = e->callee->ultimate_alias_target (&avail,
1976 : e->callee);
1977 :
1978 32181 : gcc_assert (e->speculative && !e->indirect_unknown_callee);
1979 :
1980 : /* Even if call statement is not hot, we can still have useful speculation
1981 : in cases where a lot of time is spent is callee.
1982 : Do not check maybe_hot_p. */
1983 36663 : if (!e->count.nonzero_p ())
1984 : return false;
1985 :
1986 : /* See if IP optimizations found something potentially useful about the
1987 : function. Do this only if the call seems hot since this is about
1988 : optimizing the code surrounding call site rahter than improving
1989 : callee. */
1990 32142 : if (avail >= AVAIL_AVAILABLE && e->maybe_hot_p ())
1991 : {
1992 31176 : int ecf_flags = flags_from_decl_or_type (target->decl);
1993 31176 : if (ecf_flags & ECF_CONST)
1994 : {
1995 469 : if (!(e->speculative_call_indirect_edge ()->indirect_info
1996 469 : ->ecf_flags & ECF_CONST))
1997 : return true;
1998 : }
1999 30707 : else if (ecf_flags & ECF_PURE)
2000 : {
2001 2775 : if (!(e->speculative_call_indirect_edge ()->indirect_info
2002 2775 : ->ecf_flags & ECF_PURE))
2003 : return true;
2004 : }
2005 27932 : else if (get_modref_function_summary (target))
2006 : return true;
2007 : }
2008 : /* If we did not managed to inline the function nor redirect
2009 : to an ipa-cp clone (that are seen by having local flag set),
2010 : it is probably pointless to inline it unless hardware is missing
2011 : indirect call predictor.
2012 :
2013 : At this point we know we will not dispatch into faster version of
2014 : callee, so if call itself is not hot, we definitely can give up
2015 : speculating. */
2016 14761 : if (!anticipate_inlining && (!target->local || !e->maybe_hot_p ()))
2017 4475 : return false;
2018 : /* For overwritable targets there is not much to do. */
2019 10286 : if (!can_inline_edge_p (e, false)
2020 10286 : || !can_inline_edge_by_limits_p (e, CAN_INLINE_DISREGARD_LIMITS))
2021 6 : return false;
2022 : /* OK, speculation seems interesting. */
2023 : return true;
2024 : }
2025 :
2026 : /* We know that EDGE is not going to be inlined.
2027 : See if we can remove speculation. */
2028 :
2029 : static void
2030 85404 : resolve_noninline_speculation (edge_heap_t *edge_heap, struct cgraph_edge *edge)
2031 : {
2032 85404 : if (edge->speculative && !speculation_useful_p (edge, false))
2033 : {
2034 1759 : struct cgraph_node *node = edge->caller;
2035 1753 : struct cgraph_node *where = node->inlined_to
2036 1759 : ? node->inlined_to : node;
2037 1759 : auto_bitmap updated_nodes;
2038 :
2039 1759 : if (edge->count.ipa ().initialized_p ())
2040 0 : spec_rem += edge->count.ipa ();
2041 1759 : cgraph_edge::resolve_speculation (edge);
2042 1759 : reset_edge_caches (where);
2043 1759 : ipa_update_overall_fn_summary (where);
2044 1759 : update_caller_keys (edge_heap, where,
2045 : updated_nodes, NULL);
2046 1759 : update_callee_keys (edge_heap, where, NULL,
2047 : updated_nodes);
2048 1759 : }
2049 85404 : }
2050 :
2051 : /* Return true if NODE should be accounted for overall size estimate.
2052 : Skip all nodes optimized for size so we can measure the growth of hot
2053 : part of program no matter of the padding. */
2054 :
2055 : bool
2056 3639514 : inline_account_function_p (struct cgraph_node *node)
2057 : {
2058 3639514 : return (!DECL_EXTERNAL (node->decl)
2059 3531060 : && !opt_for_fn (node->decl, optimize_size)
2060 7077527 : && node->frequency != NODE_FREQUENCY_UNLIKELY_EXECUTED);
2061 : }
2062 :
2063 : /* Count number of callers of NODE and store it into DATA (that
2064 : points to int. Worker for cgraph_for_node_and_aliases. */
2065 :
2066 : static bool
2067 1459853 : sum_callers (struct cgraph_node *node, void *data)
2068 : {
2069 1459853 : struct cgraph_edge *e;
2070 1459853 : int *num_calls = (int *)data;
2071 :
2072 3468394 : for (e = node->callers; e; e = e->next_caller)
2073 2008541 : (*num_calls)++;
2074 1459853 : return false;
2075 : }
2076 :
2077 : /* We only propagate across edges with non-interposable callee. */
2078 :
2079 : inline bool
2080 6978631 : ignore_edge_p (struct cgraph_edge *e)
2081 : {
2082 6978631 : enum availability avail;
2083 6978631 : e->callee->function_or_virtual_thunk_symbol (&avail, e->caller);
2084 6978631 : return (avail <= AVAIL_INTERPOSABLE);
2085 : }
2086 :
2087 : /* We use greedy algorithm for inlining of small functions:
2088 : All inline candidates are put into prioritized heap ordered in
2089 : increasing badness.
2090 :
2091 : The inlining of small functions is bounded by unit growth parameters. */
2092 :
2093 : static void
2094 230059 : inline_small_functions (void)
2095 : {
2096 230059 : struct cgraph_node *node;
2097 230059 : struct cgraph_edge *edge;
2098 230059 : inline_badness b;
2099 230059 : edge_heap_t edge_heap (b);
2100 230059 : auto_bitmap updated_nodes;
2101 230059 : int min_size;
2102 230059 : auto_vec<cgraph_edge *> new_indirect_edges;
2103 230059 : int initial_size = 0;
2104 230059 : struct cgraph_node **order = XCNEWVEC (cgraph_node *, symtab->cgraph_count);
2105 230059 : struct cgraph_edge_hook_list *edge_removal_hook_holder;
2106 230059 : new_indirect_edges.create (8);
2107 :
2108 230059 : edge_removal_hook_holder
2109 230059 : = symtab->add_edge_removal_hook (&heap_edge_removal_hook, &edge_heap);
2110 :
2111 : /* Compute overall unit size and other global parameters used by badness
2112 : metrics. */
2113 :
2114 230059 : has_nonzero_ipa_profile = false;
2115 230059 : ipa_reduced_postorder (order, true, ignore_edge_p);
2116 230059 : free (order);
2117 :
2118 2113732 : FOR_EACH_DEFINED_FUNCTION (node)
2119 1883673 : if (!node->inlined_to)
2120 : {
2121 1784654 : if (!node->alias && node->analyzed
2122 1784654 : && (node->has_gimple_body_p () || node->thunk)
2123 3668296 : && opt_for_fn (node->decl, optimize))
2124 : {
2125 1357337 : class ipa_fn_summary *info = ipa_fn_summaries->get (node);
2126 1357337 : struct ipa_dfs_info *dfs = (struct ipa_dfs_info *) node->aux;
2127 :
2128 : /* Do not account external functions, they will be optimized out
2129 : if not inlined. Also only count the non-cold portion of program. */
2130 1357337 : if (inline_account_function_p (node))
2131 1260957 : initial_size += ipa_size_summaries->get (node)->size;
2132 1357337 : info->growth = estimate_growth (node);
2133 :
2134 1357337 : int num_calls = 0;
2135 1357337 : node->call_for_symbol_and_aliases (sum_callers, &num_calls,
2136 : true);
2137 1357337 : if (num_calls == 1)
2138 450269 : info->single_caller = true;
2139 1357337 : if (dfs && dfs->next_cycle)
2140 : {
2141 4585 : struct cgraph_node *n2;
2142 4585 : int id = dfs->scc_no + 1;
2143 10473 : for (n2 = node; n2;
2144 5888 : n2 = ((struct ipa_dfs_info *) n2->aux)->next_cycle)
2145 9171 : if (opt_for_fn (n2->decl, optimize))
2146 : {
2147 9163 : ipa_fn_summary *info2 = ipa_fn_summaries->get
2148 9163 : (n2->inlined_to ? n2->inlined_to : n2);
2149 9163 : if (info2->scc_no)
2150 : break;
2151 5880 : info2->scc_no = id;
2152 : }
2153 : }
2154 : }
2155 :
2156 4273605 : for (edge = node->callers; edge; edge = edge->next_caller)
2157 2389963 : if (edge->count.ipa ().initialized_p ()
2158 2509228 : && edge->count.ipa ().nonzero_p ())
2159 400 : has_nonzero_ipa_profile = true;
2160 : }
2161 230059 : ipa_free_postorder_info ();
2162 230059 : initialize_growth_caches ();
2163 :
2164 230059 : if (dump_file)
2165 178 : fprintf (dump_file,
2166 : "\nDeciding on inlining of small functions. Starting with size %i.\n",
2167 : initial_size);
2168 :
2169 230059 : overall_size = initial_size;
2170 230059 : min_size = overall_size;
2171 :
2172 : /* Populate the heap with all edges we might inline. */
2173 :
2174 2113732 : FOR_EACH_DEFINED_FUNCTION (node)
2175 : {
2176 1883673 : bool update = false;
2177 1883673 : struct cgraph_edge *next = NULL;
2178 1883673 : bool has_speculative = false;
2179 :
2180 1883673 : if (!opt_for_fn (node->decl, optimize)
2181 : /* With -Og we do not want to perform IPA inlining of small
2182 : functions since there are no scalar cleanups after it
2183 : that would realize the anticipated win. All abstraction
2184 : is removed during early inlining. */
2185 1883673 : || opt_for_fn (node->decl, optimize_debug))
2186 455086 : continue;
2187 :
2188 1428587 : if (dump_file)
2189 807 : fprintf (dump_file, "Enqueueing calls in %s.\n", node->dump_name ());
2190 :
2191 7016693 : for (edge = node->callees; edge; edge = edge->next_callee)
2192 : {
2193 5588106 : if (edge->inline_failed
2194 5588075 : && !edge->aux
2195 5587952 : && can_inline_edge_p (edge, true)
2196 1545638 : && want_inline_small_function_p (edge, true)
2197 926295 : && can_inline_edge_by_limits_p (edge, CAN_INLINE_REPORT)
2198 6507057 : && edge->inline_failed)
2199 : {
2200 918951 : gcc_assert (!edge->aux);
2201 918951 : update_edge_key (&edge_heap, edge);
2202 : }
2203 5588106 : if (edge->speculative)
2204 17395 : has_speculative = true;
2205 : }
2206 1428587 : if (has_speculative)
2207 51910 : for (edge = node->callees; edge; edge = next)
2208 : {
2209 43911 : next = edge->next_callee;
2210 43911 : if (edge->speculative
2211 43911 : && !speculation_useful_p (edge, edge->aux != NULL))
2212 : {
2213 449 : cgraph_edge::resolve_speculation (edge);
2214 449 : update = true;
2215 : }
2216 : }
2217 7999 : if (update)
2218 : {
2219 0 : struct cgraph_node *where = node->inlined_to
2220 375 : ? node->inlined_to : node;
2221 375 : ipa_update_overall_fn_summary (where);
2222 375 : reset_edge_caches (where);
2223 375 : update_caller_keys (&edge_heap, where,
2224 : updated_nodes, NULL);
2225 375 : update_callee_keys (&edge_heap, where, NULL,
2226 : updated_nodes);
2227 375 : bitmap_clear (updated_nodes);
2228 : }
2229 : }
2230 :
2231 230059 : gcc_assert (in_lto_p
2232 : || !has_nonzero_ipa_profile
2233 : || flag_auto_profile
2234 : || (profile_info && flag_branch_probabilities));
2235 :
2236 2817576 : while (!edge_heap.empty ())
2237 : {
2238 2587517 : int old_size = overall_size;
2239 2587517 : struct cgraph_node *where, *callee;
2240 2587517 : sreal badness = edge_heap.min_key ().badness;
2241 2587517 : sreal current_badness;
2242 2587517 : int growth;
2243 :
2244 2587517 : edge = edge_heap.extract_min ();
2245 2587517 : gcc_assert (edge->aux);
2246 2587517 : edge->aux = NULL;
2247 2587517 : if (!edge->inline_failed || !edge->callee->analyzed)
2248 1724287 : continue;
2249 :
2250 : /* Be sure that caches are maintained consistent.
2251 : This check is affected by scaling roundoff errors when compiling for
2252 : IPA this we skip it in that case. */
2253 2587427 : if (flag_checking && !edge->callee->count.ipa_p ()
2254 5006167 : && !has_nonzero_ipa_profile)
2255 : {
2256 2418737 : sreal cached_badness = edge_badness (edge, false);
2257 :
2258 2418737 : int old_size_est = estimate_edge_size (edge);
2259 2418737 : sreal old_time_est = estimate_edge_time (edge);
2260 2418737 : int old_hints_est = estimate_edge_hints (edge);
2261 :
2262 2418737 : if (edge_growth_cache != NULL)
2263 2418737 : edge_growth_cache->remove (edge);
2264 4325661 : reset_node_cache (edge->caller->inlined_to
2265 : ? edge->caller->inlined_to
2266 : : edge->caller);
2267 2418737 : gcc_assert (old_size_est == estimate_edge_size (edge));
2268 2418737 : gcc_assert (old_time_est == estimate_edge_time (edge));
2269 : /* FIXME:
2270 :
2271 : gcc_assert (old_hints_est == estimate_edge_hints (edge));
2272 :
2273 : fails with profile feedback because some hints depends on
2274 : maybe_hot_edge_p predicate and because callee gets inlined to other
2275 : calls, the edge may become cold.
2276 : This ought to be fixed by computing relative probabilities
2277 : for given invocation but that will be better done once whole
2278 : code is converted to sreals. Disable for now and revert to "wrong"
2279 : value so enable/disable checking paths agree. */
2280 2418737 : edge_growth_cache->get (edge)->hints = old_hints_est + 1;
2281 :
2282 : /* When updating the edge costs, we only decrease badness in the keys.
2283 : Increases of badness are handled lazily; when we see key with out
2284 : of date value on it, we re-insert it now. */
2285 2418737 : current_badness = edge_badness (edge, false);
2286 2418737 : gcc_assert (cached_badness == current_badness);
2287 2418737 : gcc_assert (current_badness >= badness);
2288 : }
2289 : else
2290 168693 : current_badness = edge_badness (edge, false);
2291 2587430 : if (current_badness != badness)
2292 : {
2293 1800187 : if (edge_heap.min () && current_badness > edge_heap.min_key ().badness)
2294 : {
2295 1638796 : inline_badness b (edge, current_badness);
2296 1638796 : edge->aux = edge_heap.insert (b, edge);
2297 1638796 : continue;
2298 1638796 : }
2299 : else
2300 161391 : badness = current_badness;
2301 : }
2302 :
2303 948634 : if (!can_inline_edge_p (edge, true)
2304 948634 : || !can_inline_edge_by_limits_p (edge, CAN_INLINE_REPORT))
2305 : {
2306 4498 : resolve_noninline_speculation (&edge_heap, edge);
2307 4498 : continue;
2308 : }
2309 :
2310 944136 : callee = edge->callee->ultimate_alias_target ();
2311 944136 : growth = estimate_edge_growth (edge);
2312 944136 : if (dump_file)
2313 : {
2314 528 : fprintf (dump_file,
2315 : "\nConsidering %s with %i size\n",
2316 : callee->dump_name (),
2317 528 : ipa_size_summaries->get (callee)->size);
2318 1056 : fprintf (dump_file,
2319 : " to be inlined into %s in %s:%i\n"
2320 : " Estimated badness is %f, frequency %.2f.\n",
2321 528 : edge->caller->dump_name (),
2322 528 : edge->call_stmt
2323 500 : && (LOCATION_LOCUS (gimple_location ((const gimple *)
2324 : edge->call_stmt))
2325 : > BUILTINS_LOCATION)
2326 491 : ? gimple_filename ((const gimple *) edge->call_stmt)
2327 : : "unknown",
2328 528 : edge->call_stmt
2329 500 : ? gimple_lineno ((const gimple *) edge->call_stmt)
2330 : : -1,
2331 : badness.to_double (),
2332 528 : edge->sreal_frequency ().to_double ());
2333 528 : if (edge->count.ipa ().initialized_p ())
2334 : {
2335 0 : fprintf (dump_file, " Called ");
2336 0 : edge->count.ipa ().dump (dump_file);
2337 0 : fprintf (dump_file, " times\n");
2338 : }
2339 528 : if (dump_flags & TDF_DETAILS)
2340 185 : edge_badness (edge, true);
2341 : }
2342 :
2343 944136 : where = edge->caller;
2344 :
2345 944136 : if (overall_size + growth > compute_max_insns (where, min_size)
2346 944136 : && !DECL_DISREGARD_INLINE_LIMITS (callee->decl))
2347 : {
2348 76420 : edge->inline_failed = CIF_INLINE_UNIT_GROWTH_LIMIT;
2349 76420 : report_inline_failed_reason (edge);
2350 76420 : resolve_noninline_speculation (&edge_heap, edge);
2351 76420 : continue;
2352 : }
2353 :
2354 867716 : if (!want_inline_small_function_p (edge, true))
2355 : {
2356 2161 : resolve_noninline_speculation (&edge_heap, edge);
2357 2161 : continue;
2358 : }
2359 :
2360 865555 : profile_count old_count = callee->count;
2361 :
2362 : /* Heuristics for inlining small functions work poorly for
2363 : recursive calls where we do effects similar to loop unrolling.
2364 : When inlining such edge seems profitable, leave decision on
2365 : specific inliner. */
2366 865555 : if (edge->recursive_p ())
2367 : {
2368 1961 : if (where->inlined_to)
2369 559 : where = where->inlined_to;
2370 :
2371 : /* Disable always_inline on self recursive functions.
2372 : This prevents some inlining bombs such as one in PR113291
2373 : from exploding.
2374 : It is not enough to stop inlining in self recursive always_inlines
2375 : since they may grow large enough so always inlining them even
2376 : with recursin depth 0 is too much.
2377 :
2378 : All sane uses of always_inline should be handled during
2379 : early optimizations. */
2380 1961 : DECL_DISREGARD_INLINE_LIMITS (where->decl) = false;
2381 :
2382 1961 : if (!recursive_inlining (edge,
2383 1961 : opt_for_fn (edge->caller->decl,
2384 : flag_indirect_inlining)
2385 : ? &new_indirect_edges : NULL))
2386 : {
2387 352 : edge->inline_failed = CIF_RECURSIVE_INLINING;
2388 352 : resolve_noninline_speculation (&edge_heap, edge);
2389 352 : continue;
2390 : }
2391 1609 : reset_edge_caches (where);
2392 : /* Recursive inliner inlines all recursive calls of the function
2393 : at once. Consequently we need to update all callee keys. */
2394 1609 : if (opt_for_fn (edge->caller->decl, flag_indirect_inlining))
2395 1584 : add_new_edges_to_heap (&edge_heap, new_indirect_edges);
2396 1609 : update_callee_keys (&edge_heap, where, where, updated_nodes);
2397 1609 : bitmap_clear (updated_nodes);
2398 : }
2399 : else
2400 : {
2401 863594 : struct cgraph_node *outer_node = NULL;
2402 863594 : int depth = 0;
2403 :
2404 : /* Consider the case where self recursive function A is inlined
2405 : into B. This is desired optimization in some cases, since it
2406 : leads to effect similar of loop peeling and we might completely
2407 : optimize out the recursive call. However we must be extra
2408 : selective. */
2409 :
2410 863594 : where = edge->caller;
2411 1281799 : while (where->inlined_to)
2412 : {
2413 418205 : if (where->decl == callee->decl)
2414 11121 : outer_node = where, depth++;
2415 418205 : where = where->callers->caller;
2416 : }
2417 865567 : if (outer_node
2418 863594 : && !want_inline_self_recursive_call_p (edge, outer_node,
2419 : true, depth))
2420 : {
2421 1973 : edge->inline_failed
2422 1973 : = (DECL_DISREGARD_INLINE_LIMITS (edge->callee->decl)
2423 1973 : ? CIF_RECURSIVE_INLINING : CIF_UNSPECIFIED);
2424 1973 : resolve_noninline_speculation (&edge_heap, edge);
2425 1973 : continue;
2426 : }
2427 861621 : else if (depth && dump_file)
2428 6 : fprintf (dump_file, " Peeling recursion with depth %i\n", depth);
2429 :
2430 861621 : gcc_checking_assert (!callee->inlined_to);
2431 :
2432 861621 : int old_size = ipa_size_summaries->get (where)->size;
2433 861621 : sreal old_time = ipa_fn_summaries->get (where)->time;
2434 :
2435 861621 : inline_call (edge, true, &new_indirect_edges, &overall_size, true);
2436 861621 : reset_edge_caches (edge->callee);
2437 861621 : add_new_edges_to_heap (&edge_heap, new_indirect_edges);
2438 :
2439 : /* If caller's size and time increased we do not need to update
2440 : all edges because badness is not going to decrease. */
2441 861621 : if (old_size <= ipa_size_summaries->get (where)->size
2442 810329 : && old_time <= ipa_fn_summaries->get (where)->time
2443 : /* Wrapper penalty may be non-monotonous in this respect.
2444 : Fortunately it only affects small functions. */
2445 1430331 : && !wrapper_heuristics_may_apply (where, old_size))
2446 398657 : update_callee_keys (&edge_heap, edge->callee, edge->callee,
2447 : updated_nodes);
2448 : else
2449 462964 : update_callee_keys (&edge_heap, where,
2450 : edge->callee,
2451 : updated_nodes);
2452 : }
2453 863230 : where = edge->caller;
2454 863230 : if (where->inlined_to)
2455 218389 : where = where->inlined_to;
2456 :
2457 : /* Our profitability metric can depend on local properties
2458 : such as number of inlinable calls and size of the function body.
2459 : After inlining these properties might change for the function we
2460 : inlined into (since it's body size changed) and for the functions
2461 : called by function we inlined (since number of it inlinable callers
2462 : might change). */
2463 863230 : update_caller_keys (&edge_heap, where, updated_nodes, NULL);
2464 : /* Offline copy count has possibly changed, recompute if profile is
2465 : available. */
2466 863230 : struct cgraph_node *n
2467 863230 : = cgraph_node::get (edge->callee->decl)->ultimate_alias_target ();
2468 581268 : if (n != edge->callee && n->analyzed && !(n->count == old_count)
2469 863287 : && n->count.ipa_p ())
2470 57 : update_callee_keys (&edge_heap, n, NULL, updated_nodes);
2471 863230 : bitmap_clear (updated_nodes);
2472 :
2473 863230 : if (dump_enabled_p ())
2474 : {
2475 537 : ipa_fn_summary *s = ipa_fn_summaries->get (where);
2476 :
2477 : /* dump_printf can't handle %+i. */
2478 537 : char buf_net_change[100];
2479 537 : snprintf (buf_net_change, sizeof buf_net_change, "%+i",
2480 : overall_size - old_size);
2481 :
2482 1074 : dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, edge->call_stmt,
2483 : " Inlined %C into %C which now has time %f and "
2484 : "size %i, net change of %s%s.\n",
2485 : edge->callee, edge->caller,
2486 : s->time.to_double (),
2487 537 : ipa_size_summaries->get (edge->caller)->size,
2488 : buf_net_change,
2489 537 : cross_module_call_p (edge)
2490 : ? " (cross module)" : "");
2491 : }
2492 863230 : if (min_size > overall_size)
2493 : {
2494 215309 : min_size = overall_size;
2495 :
2496 215309 : if (dump_file)
2497 401 : fprintf (dump_file, "New minimal size reached: %i\n", min_size);
2498 : }
2499 : }
2500 :
2501 230059 : free_growth_caches ();
2502 230059 : if (dump_enabled_p ())
2503 436 : dump_printf (MSG_NOTE,
2504 : "Unit growth for small function inlining: %i->%i (%i%%)\n",
2505 : initial_size, overall_size,
2506 194 : initial_size ? overall_size * 100 / (initial_size) - 100 : 0);
2507 230059 : symtab->remove_edge_removal_hook (edge_removal_hook_holder);
2508 230059 : }
2509 :
2510 : /* Flatten NODE. Performed both during early inlining and
2511 : at IPA inlining time. */
2512 :
2513 : static void
2514 713 : flatten_function (struct cgraph_node *node, bool early, bool update)
2515 : {
2516 713 : struct cgraph_edge *e;
2517 :
2518 : /* We shouldn't be called recursively when we are being processed. */
2519 713 : gcc_assert (node->aux == NULL);
2520 :
2521 713 : node->aux = (void *) node;
2522 :
2523 1644 : for (e = node->callees; e; e = e->next_callee)
2524 : {
2525 931 : struct cgraph_node *orig_callee;
2526 931 : struct cgraph_node *callee = e->callee->ultimate_alias_target ();
2527 :
2528 : /* We've hit cycle? It is time to give up. */
2529 931 : if (callee->aux)
2530 : {
2531 15 : if (dump_enabled_p ())
2532 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
2533 : "Not inlining %C into %C to avoid cycle.\n",
2534 : callee, e->caller);
2535 15 : if (cgraph_inline_failed_type (e->inline_failed) != CIF_FINAL_ERROR)
2536 15 : e->inline_failed = CIF_RECURSIVE_INLINING;
2537 15 : continue;
2538 : }
2539 :
2540 : /* When the edge is already inlined, we just need to recurse into
2541 : it in order to fully flatten the leaves. */
2542 916 : if (!e->inline_failed)
2543 : {
2544 350 : flatten_function (callee, early, false);
2545 350 : continue;
2546 : }
2547 :
2548 : /* Flatten attribute needs to be processed during late inlining. For
2549 : extra code quality we however do flattening during early optimization,
2550 : too. */
2551 321 : if (!early
2552 566 : ? !can_inline_edge_p (e, true)
2553 245 : && !can_inline_edge_by_limits_p (e, CAN_INLINE_REPORT)
2554 321 : : !can_early_inline_edge_p (e))
2555 419 : continue;
2556 :
2557 147 : if (e->recursive_p ())
2558 : {
2559 0 : if (dump_enabled_p ())
2560 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
2561 : "Not inlining: recursive call.\n");
2562 0 : continue;
2563 : }
2564 :
2565 147 : if (gimple_in_ssa_p (DECL_STRUCT_FUNCTION (node->decl))
2566 294 : != gimple_in_ssa_p (DECL_STRUCT_FUNCTION (callee->decl)))
2567 : {
2568 4 : if (dump_enabled_p ())
2569 4 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
2570 : "Not inlining: SSA form does not match.\n");
2571 4 : continue;
2572 : }
2573 :
2574 : /* Inline the edge and flatten the inline clone. Avoid
2575 : recursing through the original node if the node was cloned. */
2576 143 : if (dump_enabled_p ())
2577 3 : dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, e->call_stmt,
2578 : " Inlining %C into %C.\n",
2579 : callee, e->caller);
2580 143 : orig_callee = callee;
2581 143 : inline_call (e, true, NULL, NULL, false);
2582 143 : if (e->callee != orig_callee)
2583 106 : orig_callee->aux = (void *) node;
2584 143 : flatten_function (e->callee, early, false);
2585 143 : if (e->callee != orig_callee)
2586 106 : orig_callee->aux = NULL;
2587 : }
2588 :
2589 713 : node->aux = NULL;
2590 713 : cgraph_node *where = node->inlined_to ? node->inlined_to : node;
2591 713 : if (update && opt_for_fn (where->decl, optimize))
2592 209 : ipa_update_overall_fn_summary (where);
2593 713 : }
2594 :
2595 : /* Inline NODE to all callers. Worker for cgraph_for_node_and_aliases.
2596 : DATA points to number of calls originally found so we avoid infinite
2597 : recursion. */
2598 :
2599 : static bool
2600 29723 : inline_to_all_callers_1 (struct cgraph_node *node, void *data,
2601 : hash_set<cgraph_node *> *callers)
2602 : {
2603 29723 : int *num_calls = (int *)data;
2604 29723 : bool callee_removed = false;
2605 :
2606 63191 : while (node->callers && !node->inlined_to)
2607 : {
2608 34631 : struct cgraph_node *caller = node->callers->caller;
2609 :
2610 34631 : if (!can_inline_edge_p (node->callers, true)
2611 34631 : || !can_inline_edge_by_limits_p (node->callers, CAN_INLINE_REPORT)
2612 69262 : || node->callers->recursive_p ())
2613 : {
2614 0 : if (dump_file)
2615 0 : fprintf (dump_file, "Uninlinable call found; giving up.\n");
2616 0 : *num_calls = 0;
2617 0 : return false;
2618 : }
2619 :
2620 34631 : if (dump_file)
2621 : {
2622 3 : cgraph_node *ultimate = node->ultimate_alias_target ();
2623 3 : fprintf (dump_file,
2624 : "\nInlining %s size %i.\n",
2625 : ultimate->dump_name (),
2626 3 : ipa_size_summaries->get (ultimate)->size);
2627 3 : fprintf (dump_file,
2628 : " Called once from %s %i insns.\n",
2629 : node->callers->caller->dump_name (),
2630 3 : ipa_size_summaries->get (node->callers->caller)->size);
2631 : }
2632 :
2633 : /* Remember which callers we inlined to, delaying updating the
2634 : overall summary. */
2635 34631 : callers->add (node->callers->caller);
2636 34631 : inline_call (node->callers, true, NULL, NULL, false, &callee_removed);
2637 34631 : if (dump_file)
2638 3 : fprintf (dump_file,
2639 : " Inlined into %s which now has %i size\n",
2640 : caller->dump_name (),
2641 3 : ipa_size_summaries->get (caller)->size);
2642 34631 : if (!(*num_calls)--)
2643 : {
2644 0 : if (dump_file)
2645 0 : fprintf (dump_file, "New calls found; giving up.\n");
2646 0 : return callee_removed;
2647 : }
2648 34631 : if (callee_removed)
2649 : return true;
2650 : }
2651 : return false;
2652 : }
2653 :
2654 : /* Wrapper around inline_to_all_callers_1 doing delayed overall summary
2655 : update. */
2656 :
2657 : static bool
2658 29723 : inline_to_all_callers (struct cgraph_node *node, void *data)
2659 : {
2660 29723 : hash_set<cgraph_node *> callers;
2661 29723 : bool res = inline_to_all_callers_1 (node, data, &callers);
2662 : /* Perform the delayed update of the overall summary of all callers
2663 : processed. This avoids quadratic behavior in the cases where
2664 : we have a lot of calls to the same function. */
2665 58469 : for (hash_set<cgraph_node *>::iterator i = callers.begin ();
2666 87215 : i != callers.end (); ++i)
2667 28746 : ipa_update_overall_fn_summary ((*i)->inlined_to ? (*i)->inlined_to : *i);
2668 29723 : return res;
2669 29723 : }
2670 :
2671 : /* Output overall time estimate. */
2672 : static void
2673 356 : dump_overall_stats (void)
2674 : {
2675 356 : sreal sum_weighted = 0, sum = 0;
2676 356 : struct cgraph_node *node;
2677 :
2678 2296 : FOR_EACH_DEFINED_FUNCTION (node)
2679 1940 : if (!node->inlined_to
2680 1383 : && !node->alias)
2681 : {
2682 1284 : ipa_fn_summary *s = ipa_fn_summaries->get (node);
2683 1284 : if (s != NULL)
2684 : {
2685 1180 : sum += s->time;
2686 1180 : if (node->count.ipa ().initialized_p ())
2687 14 : sum_weighted += s->time * node->count.ipa ().to_gcov_type ();
2688 : }
2689 : }
2690 356 : fprintf (dump_file, "Overall time estimate: "
2691 : "%f weighted by profile: "
2692 : "%f\n", sum.to_double (), sum_weighted.to_double ());
2693 356 : }
2694 :
2695 : /* Output some useful stats about inlining. */
2696 :
2697 : static void
2698 178 : dump_inline_stats (void)
2699 : {
2700 178 : int64_t inlined_cnt = 0, inlined_indir_cnt = 0;
2701 178 : int64_t inlined_virt_cnt = 0, inlined_virt_indir_cnt = 0;
2702 178 : int64_t noninlined_cnt = 0, noninlined_indir_cnt = 0;
2703 178 : int64_t noninlined_virt_cnt = 0, noninlined_virt_indir_cnt = 0;
2704 178 : int64_t inlined_speculative = 0, inlined_speculative_ply = 0;
2705 178 : int64_t indirect_poly_cnt = 0, indirect_cnt = 0;
2706 178 : int64_t reason[CIF_N_REASONS][2];
2707 5874 : sreal reason_freq[CIF_N_REASONS];
2708 178 : int i;
2709 178 : struct cgraph_node *node;
2710 :
2711 178 : memset (reason, 0, sizeof (reason));
2712 5874 : for (i=0; i < CIF_N_REASONS; i++)
2713 5696 : reason_freq[i] = 0;
2714 1246 : FOR_EACH_DEFINED_FUNCTION (node)
2715 : {
2716 1068 : struct cgraph_edge *e;
2717 5962 : for (e = node->callees; e; e = e->next_callee)
2718 : {
2719 4894 : if (e->inline_failed)
2720 : {
2721 4340 : if (e->count.ipa ().initialized_p ())
2722 2611 : reason[(int) e->inline_failed][0] += e->count.ipa ().to_gcov_type ();
2723 4340 : reason_freq[(int) e->inline_failed] += e->sreal_frequency ();
2724 4340 : reason[(int) e->inline_failed][1] ++;
2725 4340 : if (DECL_VIRTUAL_P (e->callee->decl)
2726 4340 : && e->count.ipa ().initialized_p ())
2727 : {
2728 0 : if (e->indirect_inlining_edge)
2729 0 : noninlined_virt_indir_cnt += e->count.ipa ().to_gcov_type ();
2730 : else
2731 0 : noninlined_virt_cnt += e->count.ipa ().to_gcov_type ();
2732 : }
2733 4340 : else if (e->count.ipa ().initialized_p ())
2734 : {
2735 2611 : if (e->indirect_inlining_edge)
2736 0 : noninlined_indir_cnt += e->count.ipa ().to_gcov_type ();
2737 : else
2738 2611 : noninlined_cnt += e->count.ipa ().to_gcov_type ();
2739 : }
2740 : }
2741 554 : else if (e->count.ipa ().initialized_p ())
2742 : {
2743 0 : if (e->speculative)
2744 : {
2745 0 : if (DECL_VIRTUAL_P (e->callee->decl))
2746 0 : inlined_speculative_ply += e->count.ipa ().to_gcov_type ();
2747 : else
2748 0 : inlined_speculative += e->count.ipa ().to_gcov_type ();
2749 : }
2750 0 : else if (DECL_VIRTUAL_P (e->callee->decl))
2751 : {
2752 0 : if (e->indirect_inlining_edge)
2753 0 : inlined_virt_indir_cnt += e->count.ipa ().to_gcov_type ();
2754 : else
2755 0 : inlined_virt_cnt += e->count.ipa ().to_gcov_type ();
2756 : }
2757 : else
2758 : {
2759 0 : if (e->indirect_inlining_edge)
2760 0 : inlined_indir_cnt += e->count.ipa ().to_gcov_type ();
2761 : else
2762 0 : inlined_cnt += e->count.ipa ().to_gcov_type ();
2763 : }
2764 : }
2765 : }
2766 1167 : for (e = node->indirect_calls; e; e = e->next_callee)
2767 198 : if (is_a <cgraph_polymorphic_indirect_info *> (e->indirect_info)
2768 99 : & e->count.ipa ().initialized_p ())
2769 0 : indirect_poly_cnt += e->count.ipa ().to_gcov_type ();
2770 99 : else if (e->count.ipa ().initialized_p ())
2771 0 : indirect_cnt += e->count.ipa ().to_gcov_type ();
2772 : }
2773 178 : if (has_nonzero_ipa_profile)
2774 : {
2775 0 : fprintf (dump_file,
2776 : "Inlined %" PRId64 " + speculative "
2777 : "%" PRId64 " + speculative polymorphic "
2778 : "%" PRId64 " + previously indirect "
2779 : "%" PRId64 " + virtual "
2780 : "%" PRId64 " + virtual and previously indirect "
2781 : "%" PRId64 "\n" "Not inlined "
2782 : "%" PRId64 " + previously indirect "
2783 : "%" PRId64 " + virtual "
2784 : "%" PRId64 " + virtual and previously indirect "
2785 : "%" PRId64 " + still indirect "
2786 : "%" PRId64 " + still indirect polymorphic "
2787 : "%" PRId64 "\n", inlined_cnt,
2788 : inlined_speculative, inlined_speculative_ply,
2789 : inlined_indir_cnt, inlined_virt_cnt, inlined_virt_indir_cnt,
2790 : noninlined_cnt, noninlined_indir_cnt, noninlined_virt_cnt,
2791 : noninlined_virt_indir_cnt, indirect_cnt, indirect_poly_cnt);
2792 0 : fprintf (dump_file, "Removed speculations ");
2793 0 : spec_rem.dump (dump_file);
2794 0 : fprintf (dump_file, "\n");
2795 : }
2796 178 : dump_overall_stats ();
2797 178 : fprintf (dump_file, "\nWhy inlining failed?\n");
2798 6052 : for (i = 0; i < CIF_N_REASONS; i++)
2799 5696 : if (reason[i][1])
2800 149 : fprintf (dump_file, "%-50s: %8i calls, %8f freq, %" PRId64" count\n",
2801 : cgraph_inline_failed_string ((cgraph_inline_failed_t) i),
2802 : (int) reason[i][1], reason_freq[i].to_double (), reason[i][0]);
2803 178 : }
2804 :
2805 : /* Called when node is removed. */
2806 :
2807 : static void
2808 0 : flatten_remove_node_hook (struct cgraph_node *node, void *data)
2809 : {
2810 0 : if (lookup_attribute ("flatten", DECL_ATTRIBUTES (node->decl)) == NULL)
2811 : return;
2812 :
2813 0 : hash_set<struct cgraph_node *> *removed
2814 : = (hash_set<struct cgraph_node *> *) data;
2815 0 : removed->add (node);
2816 : }
2817 :
2818 : /* Decide on the inlining. We do so in the topological order to avoid
2819 : expenses on updating data structures. */
2820 :
2821 : static unsigned int
2822 230059 : ipa_inline (void)
2823 : {
2824 230059 : struct cgraph_node *node;
2825 230059 : int nnodes;
2826 230059 : struct cgraph_node **order;
2827 230059 : int i, j;
2828 230059 : int cold;
2829 230059 : bool remove_functions = false;
2830 :
2831 230059 : order = XCNEWVEC (struct cgraph_node *, symtab->cgraph_count);
2832 :
2833 230059 : if (dump_file)
2834 178 : ipa_dump_fn_summaries (dump_file);
2835 :
2836 230059 : nnodes = ipa_reverse_postorder (order);
2837 230059 : spec_rem = profile_count::zero ();
2838 :
2839 3881171 : FOR_EACH_FUNCTION (node)
2840 : {
2841 3651112 : node->aux = 0;
2842 :
2843 : /* Recompute the default reasons for inlining because they may have
2844 : changed during merging. */
2845 3651112 : if (in_lto_p)
2846 : {
2847 435886 : for (cgraph_edge *e = node->callees; e; e = e->next_callee)
2848 : {
2849 330484 : gcc_assert (e->inline_failed);
2850 330484 : initialize_inline_failed (e);
2851 : }
2852 106664 : for (cgraph_edge *e = node->indirect_calls; e; e = e->next_callee)
2853 1262 : initialize_inline_failed (e);
2854 : }
2855 : }
2856 :
2857 230059 : if (dump_file)
2858 178 : fprintf (dump_file, "\nFlattening functions:\n");
2859 :
2860 : /* First shrink order array, so that it only contains nodes with
2861 : flatten attribute. */
2862 3881171 : for (i = nnodes - 1, j = i; i >= 0; i--)
2863 : {
2864 3651112 : node = order[i];
2865 3651112 : if (node->definition
2866 : /* Do not try to flatten aliases. These may happen for example when
2867 : creating local aliases. */
2868 1883655 : && !node->alias
2869 5435779 : && lookup_attribute ("flatten",
2870 1784667 : DECL_ATTRIBUTES (node->decl)) != NULL)
2871 85 : order[j--] = order[i];
2872 : }
2873 :
2874 : /* After the above loop, order[j + 1] ... order[nnodes - 1] contain
2875 : nodes with flatten attribute. If there is more than one such
2876 : node, we need to register a node removal hook, as flatten_function
2877 : could remove other nodes with flatten attribute. See PR82801. */
2878 230059 : struct cgraph_node_hook_list *node_removal_hook_holder = NULL;
2879 230059 : hash_set<struct cgraph_node *> *flatten_removed_nodes = NULL;
2880 230059 : if (j < nnodes - 2)
2881 : {
2882 15 : flatten_removed_nodes = new hash_set<struct cgraph_node *>;
2883 15 : node_removal_hook_holder
2884 15 : = symtab->add_cgraph_removal_hook (&flatten_remove_node_hook,
2885 : flatten_removed_nodes);
2886 : }
2887 :
2888 : /* In the first pass handle functions to be flattened. Do this with
2889 : a priority so none of our later choices will make this impossible. */
2890 230144 : for (i = nnodes - 1; i > j; i--)
2891 : {
2892 85 : node = order[i];
2893 85 : if (flatten_removed_nodes
2894 85 : && flatten_removed_nodes->contains (node))
2895 0 : continue;
2896 :
2897 : /* Handle nodes to be flattened.
2898 : Ideally when processing callees we stop inlining at the
2899 : entry of cycles, possibly cloning that entry point and
2900 : try to flatten itself turning it into a self-recursive
2901 : function. */
2902 85 : if (dump_file)
2903 4 : fprintf (dump_file, "Flattening %s\n", node->dump_name ());
2904 85 : flatten_function (node, false, true);
2905 : }
2906 :
2907 230059 : if (j < nnodes - 2)
2908 : {
2909 15 : symtab->remove_cgraph_removal_hook (node_removal_hook_holder);
2910 30 : delete flatten_removed_nodes;
2911 : }
2912 230059 : free (order);
2913 :
2914 230059 : if (dump_file)
2915 178 : dump_overall_stats ();
2916 :
2917 230059 : inline_small_functions ();
2918 :
2919 230059 : gcc_assert (symtab->state == IPA_SSA);
2920 230059 : symtab->state = IPA_SSA_AFTER_INLINING;
2921 : /* Do first after-inlining removal. We want to remove all "stale" extern
2922 : inline functions and virtual functions so we really know what is called
2923 : once. */
2924 230059 : symtab->remove_unreachable_nodes (dump_file);
2925 :
2926 : /* Inline functions with a property that after inlining into all callers the
2927 : code size will shrink because the out-of-line copy is eliminated.
2928 : We do this regardless on the callee size as long as function growth limits
2929 : are met. */
2930 230059 : if (dump_file)
2931 178 : fprintf (dump_file,
2932 : "\nDeciding on functions to be inlined into all callers and "
2933 : "removing useless speculations:\n");
2934 :
2935 : /* Inlining one function called once has good chance of preventing
2936 : inlining other function into the same callee. Ideally we should
2937 : work in priority order, but probably inlining hot functions first
2938 : is good cut without the extra pain of maintaining the queue.
2939 :
2940 : ??? this is not really fitting the bill perfectly: inlining function
2941 : into callee often leads to better optimization of callee due to
2942 : increased context for optimization.
2943 : For example if main() function calls a function that outputs help
2944 : and then function that does the main optimization, we should inline
2945 : the second with priority even if both calls are cold by themselves.
2946 :
2947 : We probably want to implement new predicate replacing our use of
2948 : maybe_hot_edge interpreted as maybe_hot_edge || callee is known
2949 : to be hot. */
2950 690177 : for (cold = 0; cold <= 1; cold ++)
2951 : {
2952 6152330 : FOR_EACH_DEFINED_FUNCTION (node)
2953 : {
2954 5692212 : struct cgraph_edge *edge, *next;
2955 5692212 : bool update=false;
2956 :
2957 5692212 : if (!opt_for_fn (node->decl, optimize)
2958 5692212 : || !opt_for_fn (node->decl, flag_inline_functions_called_once))
2959 910084 : continue;
2960 :
2961 19409465 : for (edge = node->callees; edge; edge = next)
2962 : {
2963 14627337 : next = edge->next_callee;
2964 14627337 : if (edge->speculative && !speculation_useful_p (edge, false))
2965 : {
2966 2267 : if (edge->count.ipa ().initialized_p ())
2967 0 : spec_rem += edge->count.ipa ();
2968 2267 : cgraph_edge::resolve_speculation (edge);
2969 2267 : update = true;
2970 2267 : remove_functions = true;
2971 : }
2972 : }
2973 4782128 : if (update)
2974 : {
2975 120 : struct cgraph_node *where = node->inlined_to
2976 1286 : ? node->inlined_to : node;
2977 1286 : reset_edge_caches (where);
2978 1286 : ipa_update_overall_fn_summary (where);
2979 : }
2980 4782128 : if (want_inline_function_to_all_callers_p (node, cold))
2981 : {
2982 26164 : int num_calls = 0;
2983 26164 : node->call_for_symbol_and_aliases (sum_callers, &num_calls,
2984 : true);
2985 26164 : while (node->call_for_symbol_and_aliases
2986 27327 : (inline_to_all_callers, &num_calls, true))
2987 : ;
2988 26164 : remove_functions = true;
2989 : }
2990 : }
2991 : }
2992 :
2993 230059 : if (dump_enabled_p ())
2994 242 : dump_printf (MSG_NOTE,
2995 : "\nInlined %i calls, eliminated %i functions\n\n",
2996 : ncalls_inlined, nfunctions_inlined);
2997 230059 : if (dump_file)
2998 178 : dump_inline_stats ();
2999 :
3000 230059 : if (dump_file)
3001 178 : ipa_dump_fn_summaries (dump_file);
3002 230059 : return remove_functions ? TODO_remove_functions : 0;
3003 : }
3004 :
3005 : /* Inline always-inline function calls in NODE
3006 : (which itself is possibly inline). */
3007 :
3008 : static bool
3009 3400410 : inline_always_inline_functions (struct cgraph_node *node)
3010 : {
3011 3400410 : struct cgraph_edge *e;
3012 3400410 : bool inlined = false;
3013 :
3014 13218181 : for (e = node->callees; e; e = e->next_callee)
3015 : {
3016 9817771 : struct cgraph_node *callee = e->callee->ultimate_alias_target ();
3017 9817771 : gcc_checking_assert (!callee->aux || callee->aux == (void *)(size_t)1);
3018 9817771 : if (!DECL_DISREGARD_INLINE_LIMITS (callee->decl)
3019 : /* Watch for self-recursive cycles. */
3020 9817771 : || callee->aux)
3021 9265535 : continue;
3022 :
3023 552236 : if (e->recursive_p ())
3024 : {
3025 6 : if (dump_enabled_p ())
3026 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
3027 : " Not inlining recursive call to %C.\n",
3028 : e->callee);
3029 6 : e->inline_failed = CIF_RECURSIVE_INLINING;
3030 6 : continue;
3031 : }
3032 552230 : if (callee->definition
3033 552135 : && !ipa_fn_summaries->get (callee))
3034 221 : compute_fn_summary (callee, true);
3035 :
3036 552230 : if (!can_early_inline_edge_p (e))
3037 : {
3038 : /* Set inlined to true if the callee is marked "always_inline" but
3039 : is not inlinable. This will allow flagging an error later in
3040 : expand_call_inline in tree-inline.cc. */
3041 137 : if (lookup_attribute ("always_inline",
3042 137 : DECL_ATTRIBUTES (callee->decl)) != NULL)
3043 28 : inlined = true;
3044 137 : continue;
3045 : }
3046 :
3047 552093 : if (dump_enabled_p ())
3048 18 : dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, e->call_stmt,
3049 : " Inlining %C into %C (always_inline).\n",
3050 : e->callee, e->caller);
3051 552093 : inline_call (e, true, NULL, NULL, false);
3052 552093 : callee->aux = (void *)(size_t)1;
3053 : /* Inline recursively to handle the case where always_inline function was
3054 : not optimized yet since it is a part of a cycle in callgraph. */
3055 552093 : inline_always_inline_functions (e->callee);
3056 552093 : callee->aux = NULL;
3057 552093 : inlined = true;
3058 : }
3059 3400410 : return inlined;
3060 : }
3061 :
3062 : /* Decide on the inlining. We do so in the topological order to avoid
3063 : expenses on updating data structures. */
3064 :
3065 : static bool
3066 2379346 : early_inline_small_functions (struct cgraph_node *node)
3067 : {
3068 2379346 : struct cgraph_edge *e;
3069 2379346 : bool inlined = false;
3070 :
3071 10199812 : for (e = node->callees; e; e = e->next_callee)
3072 : {
3073 7820466 : struct cgraph_node *callee = e->callee->ultimate_alias_target ();
3074 :
3075 : /* We can encounter not-yet-analyzed function during
3076 : early inlining on callgraphs with strongly
3077 : connected components. */
3078 7820466 : ipa_fn_summary *s = ipa_fn_summaries->get (callee);
3079 7820466 : if (s == NULL || !s->inlinable || !e->inline_failed)
3080 4188893 : continue;
3081 :
3082 : /* Do not consider functions not declared inline. */
3083 3631573 : if (!DECL_DECLARED_INLINE_P (callee->decl)
3084 859981 : && !opt_for_fn (node->decl, flag_inline_small_functions)
3085 3683805 : && !opt_for_fn (node->decl, flag_inline_functions))
3086 52167 : continue;
3087 :
3088 3579406 : if (dump_enabled_p ())
3089 155 : dump_printf_loc (MSG_NOTE, e->call_stmt,
3090 : "Considering inline candidate %C.\n",
3091 : callee);
3092 :
3093 3579406 : if (!can_early_inline_edge_p (e))
3094 88255 : continue;
3095 :
3096 3491151 : if (e->recursive_p ())
3097 : {
3098 8517 : if (dump_enabled_p ())
3099 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
3100 : " Not inlining: recursive call.\n");
3101 8517 : continue;
3102 : }
3103 :
3104 3482634 : if (!want_early_inline_function_p (e))
3105 1042778 : continue;
3106 :
3107 2439856 : if (dump_enabled_p ())
3108 102 : dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, e->call_stmt,
3109 : " Inlining %C into %C.\n",
3110 : callee, e->caller);
3111 2439856 : inline_call (e, true, NULL, NULL, false);
3112 2439856 : inlined = true;
3113 : }
3114 :
3115 2379346 : if (inlined)
3116 824752 : ipa_update_overall_fn_summary (node);
3117 :
3118 2379346 : return inlined;
3119 : }
3120 :
3121 : /* With auto-fdo inline all functions that was inlined in the train run
3122 : and inlining seems useful. That is there are enough samples in the callee
3123 : function.
3124 :
3125 : Unlike early inlining, we inline recursively. Profile data is also used
3126 : to produce speculative calls which we then inline. In the case some
3127 : speculatin was introduced, set SPECULATIVE_CALLS. */
3128 :
3129 : static bool
3130 2380871 : inline_functions_by_afdo (struct cgraph_node *node, bool *speculative_calls)
3131 : {
3132 2380871 : if (!flag_auto_profile || !flag_auto_profile_inlining)
3133 : return false;
3134 0 : struct cgraph_edge *e;
3135 0 : bool inlined = false;
3136 :
3137 0 : *speculative_calls |= afdo_vpt_for_early_inline (node);
3138 :
3139 0 : cgraph_edge *next;
3140 0 : for (e = node->callees; e; e = next)
3141 : {
3142 0 : next = e->next_callee;
3143 :
3144 0 : if (!e->inline_failed)
3145 : {
3146 0 : inlined |= inline_functions_by_afdo (e->callee, speculative_calls);
3147 0 : continue;
3148 : }
3149 0 : if (!afdo_callsite_hot_enough_for_early_inline (e))
3150 : {
3151 : /* If we do not want to inline, remove the speculation. */
3152 0 : if (e->speculative)
3153 0 : cgraph_edge::resolve_speculation (e);
3154 0 : continue;
3155 : }
3156 :
3157 0 : struct cgraph_node *callee = e->callee->ultimate_alias_target ();
3158 0 : if (callee->definition
3159 0 : && !ipa_fn_summaries->get (callee))
3160 0 : compute_fn_summary (callee, true);
3161 :
3162 0 : if (!can_early_inline_edge_p (e))
3163 : {
3164 0 : if (dump_enabled_p ())
3165 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
3166 : "Not inlining %C -> %C using auto-profile, %s.",
3167 : e->caller, e->callee,
3168 : cgraph_inline_failed_string (e->inline_failed));
3169 : /* If we do not want to inline, remove the speculation. */
3170 0 : if (e->speculative)
3171 0 : cgraph_edge::resolve_speculation (e);
3172 0 : continue;
3173 : }
3174 : /* We can handle recursive inlining by first producing
3175 : inline clone. */
3176 0 : if (e->recursive_p ())
3177 : {
3178 0 : if (dump_enabled_p ())
3179 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
3180 : "Not inlining %C recursively"
3181 : " using auto-profile.\n",
3182 : e->callee);
3183 : /* If we do not want to inline, remove the speculation. */
3184 0 : if (e->speculative)
3185 0 : cgraph_edge::resolve_speculation (e);
3186 0 : continue;
3187 : }
3188 :
3189 0 : if (dump_enabled_p ())
3190 : {
3191 0 : if (e->caller->inlined_to)
3192 0 : dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, e->call_stmt,
3193 : "Inlining using auto-profile %C into %C "
3194 : "which is transitively inlined to %C.\n",
3195 : callee, e->caller, e->caller->inlined_to);
3196 : else
3197 0 : dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, e->call_stmt,
3198 : "Inlining using auto-profile %C into %C.\n",
3199 : callee, e->caller);
3200 : }
3201 0 : if (e->speculative)
3202 0 : remove_afdo_speculative_target (e);
3203 0 : inline_call (e, true, NULL, NULL, false);
3204 0 : inlined |= inline_functions_by_afdo (e->callee, speculative_calls);
3205 0 : inlined = true;
3206 : }
3207 :
3208 0 : if (inlined && !node->inlined_to)
3209 0 : ipa_update_overall_fn_summary (node);
3210 :
3211 : return inlined;
3212 : }
3213 :
3214 : unsigned int
3215 2848331 : early_inliner (function *fun)
3216 : {
3217 2848331 : struct cgraph_node *node = cgraph_node::get (current_function_decl);
3218 2848331 : struct cgraph_edge *edge;
3219 2848331 : unsigned int todo = 0;
3220 2848331 : int iterations = 0;
3221 2848331 : bool inlined = false;
3222 :
3223 2848331 : if (seen_error ())
3224 : return 0;
3225 :
3226 : /* Do nothing if datastructures for ipa-inliner are already computed. This
3227 : happens when some pass decides to construct new function and
3228 : cgraph_add_new_function calls lowering passes and early optimization on
3229 : it. This may confuse ourself when early inliner decide to inline call to
3230 : function clone, because function clones don't have parameter list in
3231 : ipa-prop matching their signature. */
3232 2848325 : if (ipa_node_params_sum)
3233 : return 0;
3234 :
3235 2848317 : if (flag_checking)
3236 2848287 : node->verify ();
3237 2848317 : node->remove_all_references ();
3238 :
3239 : /* Even when not optimizing or not inlining inline always-inline
3240 : functions. */
3241 2848317 : inlined = inline_always_inline_functions (node);
3242 :
3243 2848317 : if (!optimize
3244 2412294 : || flag_no_inline
3245 2380912 : || !flag_early_inlining)
3246 : ;
3247 2379387 : else if (lookup_attribute ("flatten",
3248 2379387 : DECL_ATTRIBUTES (node->decl)) != NULL)
3249 : {
3250 : /* When the function is marked to be flattened, recursively inline
3251 : all calls in it. */
3252 135 : if (dump_enabled_p ())
3253 0 : dump_printf (MSG_OPTIMIZED_LOCATIONS,
3254 : "Flattening %C\n", node);
3255 135 : flatten_function (node, true, true);
3256 135 : inlined = true;
3257 : }
3258 : else
3259 : {
3260 : /* If some always_inline functions was inlined, apply the changes.
3261 : This way we will not account always inline into growth limits and
3262 : moreover we will inline calls from always inlines that we skipped
3263 : previously because of conditional in can_early_inline_edge_p
3264 : which prevents some inlining to always_inline. */
3265 2379252 : if (inlined)
3266 : {
3267 290981 : timevar_push (TV_INTEGRATION);
3268 290981 : todo |= optimize_inline_calls (current_function_decl);
3269 : /* optimize_inline_calls call above might have introduced new
3270 : statements that don't have inline parameters computed. */
3271 1456179 : for (edge = node->callees; edge; edge = edge->next_callee)
3272 : {
3273 : /* We can enounter not-yet-analyzed function during
3274 : early inlining on callgraphs with strongly
3275 : connected components. */
3276 1165198 : ipa_call_summary *es = ipa_call_summaries->get_create (edge);
3277 1165198 : es->call_stmt_size
3278 1165198 : = estimate_num_insns (edge->call_stmt, &eni_size_weights);
3279 1165198 : es->call_stmt_time
3280 1165198 : = estimate_num_insns (edge->call_stmt, &eni_time_weights);
3281 : }
3282 290981 : ipa_update_overall_fn_summary (node);
3283 290981 : inlined = false;
3284 290981 : timevar_pop (TV_INTEGRATION);
3285 : }
3286 : /* We iterate incremental inlining to get trivial cases of indirect
3287 : inlining. */
3288 3204004 : while (iterations < opt_for_fn (node->decl,
3289 : param_early_inliner_max_iterations))
3290 : {
3291 2379346 : bool inlined = early_inline_small_functions (node);
3292 2379346 : bool speculative_calls = false;
3293 2379346 : inlined |= inline_functions_by_afdo (node, &speculative_calls);
3294 2379346 : if (!inlined)
3295 : break;
3296 824752 : timevar_push (TV_INTEGRATION);
3297 824752 : if (speculative_calls)
3298 : {
3299 0 : cgraph_edge *next;
3300 0 : for (cgraph_edge *e = node->callees; e; e = next)
3301 : {
3302 0 : next = e->next_callee;
3303 0 : cgraph_edge::redirect_call_stmt_to_callee (e);
3304 : }
3305 : }
3306 824752 : todo |= optimize_inline_calls (current_function_decl);
3307 :
3308 : /* Technically we ought to recompute inline parameters so the new
3309 : iteration of early inliner works as expected. We however have
3310 : values approximately right and thus we only need to update edge
3311 : info that might be cleared out for newly discovered edges. */
3312 3406874 : for (edge = node->callees; edge; edge = edge->next_callee)
3313 : {
3314 : /* We have no summary for new bound store calls yet. */
3315 2582122 : ipa_call_summary *es = ipa_call_summaries->get_create (edge);
3316 2582122 : es->call_stmt_size
3317 2582122 : = estimate_num_insns (edge->call_stmt, &eni_size_weights);
3318 2582122 : es->call_stmt_time
3319 2582122 : = estimate_num_insns (edge->call_stmt, &eni_time_weights);
3320 : }
3321 824752 : if (iterations < opt_for_fn (node->decl,
3322 824752 : param_early_inliner_max_iterations) - 1)
3323 94 : ipa_update_overall_fn_summary (node);
3324 824752 : timevar_pop (TV_INTEGRATION);
3325 824752 : iterations++;
3326 824752 : inlined = false;
3327 : }
3328 2379252 : if (dump_file)
3329 203 : fprintf (dump_file, "Iterations: %i\n", iterations);
3330 : }
3331 :
3332 : /* do AFDO inlining in case it was not done as part of early inlining. */
3333 2848317 : if (optimize
3334 2412294 : && !flag_no_inline
3335 2380912 : && !flag_early_inlining
3336 1525 : && flag_auto_profile_inlining)
3337 : {
3338 1525 : bool speculative_calls = false;
3339 1525 : inlined |= inline_functions_by_afdo (node, &speculative_calls);
3340 1525 : if (speculative_calls)
3341 : {
3342 0 : cgraph_edge *next;
3343 0 : for (cgraph_edge *e = node->callees; e; e = next)
3344 : {
3345 0 : next = e->next_callee;
3346 0 : cgraph_edge::redirect_call_stmt_to_callee (e);
3347 : }
3348 : }
3349 : }
3350 :
3351 2848317 : if (inlined)
3352 : {
3353 23182 : timevar_push (TV_INTEGRATION);
3354 23182 : todo |= optimize_inline_calls (current_function_decl);
3355 23182 : timevar_pop (TV_INTEGRATION);
3356 : }
3357 :
3358 2848317 : fun->always_inline_functions_inlined = true;
3359 :
3360 2848317 : return todo;
3361 : }
3362 :
3363 : /* Do inlining of small functions. Doing so early helps profiling and other
3364 : passes to be somewhat more effective and avoids some code duplication in
3365 : later real inlining pass for testcases with very many function calls. */
3366 :
3367 : namespace {
3368 :
3369 : const pass_data pass_data_early_inline =
3370 : {
3371 : GIMPLE_PASS, /* type */
3372 : "einline", /* name */
3373 : OPTGROUP_INLINE, /* optinfo_flags */
3374 : TV_EARLY_INLINING, /* tv_id */
3375 : PROP_ssa, /* properties_required */
3376 : 0, /* properties_provided */
3377 : 0, /* properties_destroyed */
3378 : 0, /* todo_flags_start */
3379 : 0, /* todo_flags_finish */
3380 : };
3381 :
3382 : class pass_early_inline : public gimple_opt_pass
3383 : {
3384 : public:
3385 285722 : pass_early_inline (gcc::context *ctxt)
3386 571444 : : gimple_opt_pass (pass_data_early_inline, ctxt)
3387 : {}
3388 :
3389 : /* opt_pass methods: */
3390 : unsigned int execute (function *) final override;
3391 :
3392 : }; // class pass_early_inline
3393 :
3394 : unsigned int
3395 2848331 : pass_early_inline::execute (function *fun)
3396 : {
3397 2848331 : return early_inliner (fun);
3398 : }
3399 :
3400 : } // anon namespace
3401 :
3402 : gimple_opt_pass *
3403 285722 : make_pass_early_inline (gcc::context *ctxt)
3404 : {
3405 285722 : return new pass_early_inline (ctxt);
3406 : }
3407 :
3408 : namespace {
3409 :
3410 : const pass_data pass_data_ipa_inline =
3411 : {
3412 : IPA_PASS, /* type */
3413 : "inline", /* name */
3414 : OPTGROUP_INLINE, /* optinfo_flags */
3415 : TV_IPA_INLINING, /* tv_id */
3416 : 0, /* properties_required */
3417 : 0, /* properties_provided */
3418 : 0, /* properties_destroyed */
3419 : 0, /* todo_flags_start */
3420 : ( TODO_dump_symtab ), /* todo_flags_finish */
3421 : };
3422 :
3423 : class pass_ipa_inline : public ipa_opt_pass_d
3424 : {
3425 : public:
3426 285722 : pass_ipa_inline (gcc::context *ctxt)
3427 : : ipa_opt_pass_d (pass_data_ipa_inline, ctxt,
3428 : NULL, /* generate_summary */
3429 : NULL, /* write_summary */
3430 : NULL, /* read_summary */
3431 : NULL, /* write_optimization_summary */
3432 : NULL, /* read_optimization_summary */
3433 : NULL, /* stmt_fixup */
3434 : 0, /* function_transform_todo_flags_start */
3435 : inline_transform, /* function_transform */
3436 285722 : NULL) /* variable_transform */
3437 285722 : {}
3438 :
3439 : /* opt_pass methods: */
3440 230059 : unsigned int execute (function *) final override { return ipa_inline (); }
3441 :
3442 : }; // class pass_ipa_inline
3443 :
3444 : } // anon namespace
3445 :
3446 : ipa_opt_pass_d *
3447 285722 : make_pass_ipa_inline (gcc::context *ctxt)
3448 : {
3449 285722 : return new pass_ipa_inline (ctxt);
3450 : }
|
