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 1049869 : inline_badness ()
138 1049869 : : badness (sreal::min ()), uid (0)
139 : {
140 : }
141 3606948 : inline_badness (cgraph_edge *e, sreal b)
142 18514 : : badness (b), uid (e->get_uid ())
143 : {
144 : }
145 936841 : bool operator<= (const inline_badness &other)
146 : {
147 936841 : if (badness != other.badness)
148 936841 : return badness <= other.badness;
149 0 : return uid <= other.uid;
150 : }
151 1049869 : bool operator== (const inline_badness &other)
152 : {
153 1911302 : return badness == other.badness && uid == other.uid;
154 : }
155 0 : bool operator!= (const inline_badness &other)
156 : {
157 1049869 : return badness != other.badness || uid != other.uid;
158 : }
159 28019112 : bool operator< (const inline_badness &other)
160 : {
161 28019112 : if (badness != other.badness)
162 23957021 : return badness < other.badness;
163 4062091 : return uid < other.uid;
164 : }
165 12788315 : bool operator> (const inline_badness &other)
166 : {
167 12788315 : if (badness != other.badness)
168 10864831 : return badness > other.badness;
169 1923484 : 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 6941530 : caller_growth_limits (struct cgraph_edge *e)
196 : {
197 6941530 : struct cgraph_node *to = e->caller;
198 6941530 : struct cgraph_node *what = e->callee->ultimate_alias_target ();
199 6941530 : int newsize;
200 6941530 : int limit = 0;
201 6941530 : HOST_WIDE_INT stack_size_limit = 0, inlined_stack;
202 6941530 : 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 10389448 : while (true)
213 : {
214 8665489 : ipa_size_summary *size_info = ipa_size_summaries->get (to);
215 8665489 : if (limit < size_info->self_size)
216 : limit = size_info->self_size;
217 8665489 : if (stack_size_limit < size_info->estimated_self_stack_size)
218 : stack_size_limit = size_info->estimated_self_stack_size;
219 8665489 : if (to->inlined_to)
220 1723959 : to = to->callers->caller;
221 : else
222 : break;
223 1723959 : }
224 :
225 6941530 : ipa_fn_summary *what_info = ipa_fn_summaries->get (what);
226 6941530 : ipa_size_summary *what_size_info = ipa_size_summaries->get (what);
227 :
228 6941530 : if (limit < what_size_info->self_size)
229 : limit = what_size_info->self_size;
230 :
231 6941530 : 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 6941530 : newsize = estimate_size_after_inlining (to, e);
236 6941530 : if (newsize >= ipa_size_summaries->get (what)->size
237 6759670 : && newsize > opt_for_fn (to->decl, param_large_function_insns)
238 7198930 : && newsize > limit)
239 : {
240 5559 : e->inline_failed = CIF_LARGE_FUNCTION_GROWTH_LIMIT;
241 5559 : return false;
242 : }
243 :
244 6935971 : 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 2159824 : stack_size_limit += ((gcov_type)stack_size_limit
253 1079912 : * opt_for_fn (to->decl, param_stack_frame_growth)
254 1079912 : / 100);
255 :
256 1079912 : inlined_stack = (ipa_get_stack_frame_offset (to)
257 1079912 : + outer_info->estimated_self_stack_size
258 1079912 : + what_info->estimated_stack_size);
259 : /* Check new stack consumption with stack consumption at the place
260 : stack is used. */
261 1079912 : 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 313191 : && inlined_stack > ipa_fn_summaries->get (to)->estimated_stack_size
267 1379824 : && inlined_stack > opt_for_fn (to->decl, param_large_stack_frame))
268 : {
269 64628 : e->inline_failed = CIF_LARGE_STACK_FRAME_GROWTH_LIMIT;
270 64628 : return false;
271 : }
272 : return true;
273 : }
274 :
275 : /* Dump info about why inlining has failed. */
276 :
277 : static void
278 5158952 : report_inline_failed_reason (struct cgraph_edge *e)
279 : {
280 5158952 : if (dump_enabled_p ())
281 : {
282 2387 : 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 2387 : if ((e->inline_failed == CIF_TARGET_OPTION_MISMATCH
287 2387 : || e->inline_failed == CIF_OPTIMIZATION_MISMATCH)
288 2 : && e->caller->lto_file_data
289 2387 : && 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 2387 : if (e->inline_failed == CIF_TARGET_OPTION_MISMATCH
297 2 : && dump_file)
298 : {
299 0 : struct cl_target_option *opt_caller
300 0 : = target_opts_for_fn (e->caller->decl);
301 0 : struct cl_target_option *opt_callee
302 0 : = target_opts_for_fn (e->callee->ultimate_alias_target ()->decl);
303 0 : if (opt_caller != NULL && opt_callee != NULL)
304 0 : cl_target_option_print_diff (dump_file, 2, opt_caller, opt_callee);
305 : }
306 2387 : if (e->inline_failed == CIF_OPTIMIZATION_MISMATCH)
307 0 : if (dump_file)
308 0 : cl_optimization_print_diff
309 0 : (dump_file, 2, opts_for_fn (e->caller->decl),
310 0 : opts_for_fn (e->callee->ultimate_alias_target ()->decl));
311 : }
312 5158952 : }
313 :
314 : /* Decide whether sanitizer-related attributes allow inlining. */
315 :
316 : static bool
317 10161699 : sanitize_attrs_match_for_inline_p (const_tree caller, const_tree callee)
318 : {
319 10161699 : if (!caller || !callee)
320 : return true;
321 :
322 : /* Follow clang and allow inlining for always_inline functions. */
323 10161699 : if (lookup_attribute ("always_inline", DECL_ATTRIBUTES (callee)))
324 : return true;
325 :
326 9535635 : const sanitize_code codes[] =
327 : {
328 : SANITIZE_ADDRESS,
329 : SANITIZE_THREAD,
330 : SANITIZE_UNDEFINED,
331 : SANITIZE_UNDEFINED_NONDEFAULT,
332 : SANITIZE_POINTER_COMPARE,
333 : SANITIZE_POINTER_SUBTRACT
334 : };
335 :
336 66748453 : for (unsigned i = 0; i < ARRAY_SIZE (codes); i++)
337 114426008 : if (sanitize_flags_p (codes[i], caller)
338 57213004 : != sanitize_flags_p (codes[i], callee))
339 : return false;
340 :
341 9535449 : if (sanitize_coverage_p (caller) != sanitize_coverage_p (callee))
342 0 : return false;
343 :
344 : return true;
345 : }
346 :
347 : /* Used for flags where it is safe to inline when caller's value is
348 : grater than callee's. */
349 : #define check_maybe_up(flag) \
350 : (opts_for_fn (caller->decl)->x_##flag \
351 : != opts_for_fn (callee->decl)->x_##flag \
352 : && (!always_inline \
353 : || opts_for_fn (caller->decl)->x_##flag \
354 : < opts_for_fn (callee->decl)->x_##flag))
355 : /* Used for flags where it is safe to inline when caller's value is
356 : smaller than callee's. */
357 : #define check_maybe_down(flag) \
358 : (opts_for_fn (caller->decl)->x_##flag \
359 : != opts_for_fn (callee->decl)->x_##flag \
360 : && (!always_inline \
361 : || opts_for_fn (caller->decl)->x_##flag \
362 : > opts_for_fn (callee->decl)->x_##flag))
363 : /* Used for flags where exact match is needed for correctness. */
364 : #define check_match(flag) \
365 : (opts_for_fn (caller->decl)->x_##flag \
366 : != opts_for_fn (callee->decl)->x_##flag)
367 :
368 : /* Decide if we can inline the edge and possibly update
369 : inline_failed reason.
370 : We check whether inlining is possible at all and whether
371 : caller growth limits allow doing so.
372 :
373 : if REPORT is true, output reason to the dump file. */
374 :
375 : static bool
376 14411406 : can_inline_edge_p (struct cgraph_edge *e, bool report,
377 : bool early = false)
378 : {
379 14411406 : gcc_checking_assert (e->inline_failed);
380 :
381 14411406 : if (cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR)
382 : {
383 3754381 : if (report)
384 3722504 : report_inline_failed_reason (e);
385 : return false;
386 : }
387 :
388 10657025 : bool inlinable = true;
389 10657025 : enum availability avail;
390 9225388 : cgraph_node *caller = (e->caller->inlined_to
391 10657025 : ? e->caller->inlined_to : e->caller);
392 10657025 : cgraph_node *callee = e->callee->ultimate_alias_target (&avail, caller);
393 :
394 10657025 : if (!callee->definition)
395 : {
396 950 : e->inline_failed = CIF_BODY_NOT_AVAILABLE;
397 950 : inlinable = false;
398 : }
399 10657025 : if (!early && (!opt_for_fn (callee->decl, optimize)
400 6231198 : || !opt_for_fn (caller->decl, optimize)))
401 : {
402 458 : e->inline_failed = CIF_FUNCTION_NOT_OPTIMIZED;
403 458 : inlinable = false;
404 : }
405 10656567 : else if (callee->calls_comdat_local)
406 : {
407 23489 : e->inline_failed = CIF_USES_COMDAT_LOCAL;
408 23489 : inlinable = false;
409 : }
410 10633078 : else if (avail <= AVAIL_INTERPOSABLE)
411 : {
412 136397 : e->inline_failed = CIF_OVERWRITABLE;
413 136397 : inlinable = false;
414 : }
415 : /* All edges with call_stmt_cannot_inline_p should have inline_failed
416 : initialized to one of FINAL_ERROR reasons. */
417 10496681 : else if (e->call_stmt_cannot_inline_p)
418 0 : gcc_unreachable ();
419 : /* Don't inline if the functions have different EH personalities. */
420 10496681 : else if (DECL_FUNCTION_PERSONALITY (caller->decl)
421 3113791 : && DECL_FUNCTION_PERSONALITY (callee->decl)
422 10496681 : && (DECL_FUNCTION_PERSONALITY (caller->decl)
423 278263 : != DECL_FUNCTION_PERSONALITY (callee->decl)))
424 : {
425 0 : e->inline_failed = CIF_EH_PERSONALITY;
426 0 : inlinable = false;
427 : }
428 : /* TM pure functions should not be inlined into non-TM_pure
429 : functions. */
430 10496681 : else if (is_tm_pure (callee->decl) && !is_tm_pure (caller->decl))
431 : {
432 30 : e->inline_failed = CIF_UNSPECIFIED;
433 30 : inlinable = false;
434 : }
435 : /* Check compatibility of target optimization options. */
436 10496651 : else if (!targetm.target_option.can_inline_p (caller->decl,
437 : callee->decl))
438 : {
439 477 : e->inline_failed = CIF_TARGET_OPTION_MISMATCH;
440 477 : inlinable = false;
441 : }
442 10496174 : else if (ipa_fn_summaries->get (callee) == NULL
443 10496172 : || !ipa_fn_summaries->get (callee)->inlinable)
444 : {
445 334475 : e->inline_failed = CIF_FUNCTION_NOT_INLINABLE;
446 334475 : inlinable = false;
447 : }
448 : /* Don't inline a function with mismatched sanitization attributes. */
449 10161699 : else if (!sanitize_attrs_match_for_inline_p (caller->decl, callee->decl))
450 : {
451 186 : e->inline_failed = CIF_SANITIZE_ATTRIBUTE_MISMATCH;
452 186 : inlinable = false;
453 : }
454 :
455 10657025 : if (inlinable && !strub_inlinable_to_p (callee, caller))
456 : {
457 1104 : e->inline_failed = CIF_UNSPECIFIED;
458 1104 : inlinable = false;
459 : }
460 10657025 : if (inlinable && callee->must_remain_in_tu_body
461 9 : && caller->lto_file_data != callee->lto_file_data)
462 : {
463 9 : e->inline_failed = CIF_MUST_REMAIN_IN_TU;
464 9 : inlinable = false;
465 : }
466 10657025 : if (!inlinable && report)
467 489212 : report_inline_failed_reason (e);
468 : return inlinable;
469 : }
470 :
471 : /* Return inlining_insns_single limit for function N. If HINT or HINT2 is true
472 : scale up the bound. */
473 :
474 : static int
475 9784289 : inline_insns_single (cgraph_node *n, bool hint, bool hint2)
476 : {
477 9784289 : if (hint && hint2)
478 : {
479 3422858 : int64_t spd = opt_for_fn (n->decl, param_inline_heuristics_hint_percent);
480 3422858 : spd = spd * spd;
481 3422858 : if (spd > 1000000)
482 : spd = 1000000;
483 3422858 : return opt_for_fn (n->decl, param_max_inline_insns_single) * spd / 100;
484 : }
485 6361431 : if (hint || hint2)
486 724462 : return opt_for_fn (n->decl, param_max_inline_insns_single)
487 724462 : * opt_for_fn (n->decl, param_inline_heuristics_hint_percent) / 100;
488 5636969 : return opt_for_fn (n->decl, param_max_inline_insns_single);
489 : }
490 :
491 : /* Return inlining_insns_auto limit for function N. If HINT or HINT2 is true
492 : scale up the bound. */
493 :
494 : static int
495 5619700 : inline_insns_auto (cgraph_node *n, bool hint, bool hint2)
496 : {
497 5619700 : int max_inline_insns_auto = opt_for_fn (n->decl, param_max_inline_insns_auto);
498 5619700 : if (hint && hint2)
499 : {
500 2053642 : int64_t spd = opt_for_fn (n->decl, param_inline_heuristics_hint_percent);
501 2053642 : spd = spd * spd;
502 2053642 : if (spd > 1000000)
503 : spd = 1000000;
504 2053642 : return max_inline_insns_auto * spd / 100;
505 : }
506 3566058 : if (hint || hint2)
507 1502032 : return max_inline_insns_auto
508 1502032 : * opt_for_fn (n->decl, param_inline_heuristics_hint_percent) / 100;
509 : return max_inline_insns_auto;
510 : }
511 :
512 : enum can_inline_edge_by_limits_flags
513 : {
514 : /* True if we are early inlining. */
515 : CAN_INLINE_EARLY = 1,
516 : /* Ignore size limits. */
517 : CAN_INLINE_DISREGARD_LIMITS = 2,
518 : /* Force size limits (ignore always_inline). This is used for
519 : recursive inlining where always_inline may lead to inline bombs
520 : and technically it is non-sential anyway. */
521 : CAN_INLINE_FORCE_LIMITS = 4,
522 : /* Report decision to dump file. */
523 : CAN_INLINE_REPORT = 8,
524 : };
525 :
526 : /* Decide if we can inline the edge and possibly update
527 : inline_failed reason.
528 : We check whether inlining is possible at all and whether
529 : caller growth limits allow doing so. */
530 :
531 : static bool
532 7578927 : can_inline_edge_by_limits_p (struct cgraph_edge *e, int flags)
533 : {
534 7578927 : gcc_checking_assert (e->inline_failed);
535 :
536 7578927 : if (cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR)
537 : {
538 210 : if (flags & CAN_INLINE_REPORT)
539 210 : report_inline_failed_reason (e);
540 : return false;
541 : }
542 :
543 7578717 : bool inlinable = true;
544 7578717 : enum availability avail;
545 6730371 : cgraph_node *caller = (e->caller->inlined_to
546 7578717 : ? e->caller->inlined_to : e->caller);
547 7578717 : cgraph_node *callee = e->callee->ultimate_alias_target (&avail, caller);
548 7578717 : tree caller_tree = DECL_FUNCTION_SPECIFIC_OPTIMIZATION (caller->decl);
549 7578717 : tree callee_tree
550 7578717 : = callee ? DECL_FUNCTION_SPECIFIC_OPTIMIZATION (callee->decl) : NULL;
551 : /* Check if caller growth allows the inlining. */
552 7578717 : if (!(flags & CAN_INLINE_DISREGARD_LIMITS)
553 7570268 : && ((flags & CAN_INLINE_FORCE_LIMITS)
554 7534856 : || (!DECL_DISREGARD_INLINE_LIMITS (callee->decl)
555 6906262 : && !lookup_attribute ("flatten",
556 6906262 : DECL_ATTRIBUTES (caller->decl))))
557 14520247 : && !caller_growth_limits (e))
558 : inlinable = false;
559 7508530 : else if (callee->externally_visible
560 4995279 : && !DECL_DISREGARD_INLINE_LIMITS (callee->decl)
561 12004693 : && flag_live_patching == LIVE_PATCHING_INLINE_ONLY_STATIC)
562 : {
563 2 : e->inline_failed = CIF_EXTERN_LIVE_ONLY_STATIC;
564 2 : inlinable = false;
565 : }
566 : /* Don't inline a function with a higher optimization level than the
567 : caller. FIXME: this is really just tip of iceberg of handling
568 : optimization attribute. */
569 7508528 : else if (caller_tree != callee_tree)
570 : {
571 9999 : bool always_inline =
572 9999 : (DECL_DISREGARD_INLINE_LIMITS (callee->decl)
573 12690 : && lookup_attribute ("always_inline",
574 2691 : DECL_ATTRIBUTES (callee->decl)));
575 9999 : ipa_fn_summary *caller_info = ipa_fn_summaries->get (caller);
576 9999 : ipa_fn_summary *callee_info = ipa_fn_summaries->get (callee);
577 :
578 : /* Until GCC 4.9 we did not check the semantics-altering flags
579 : below and inlined across optimization boundaries.
580 : Enabling checks below breaks several packages by refusing
581 : to inline library always_inline functions. See PR65873.
582 : Disable the check for early inlining for now until better solution
583 : is found. */
584 9999 : if (always_inline && (flags & CAN_INLINE_EARLY))
585 : ;
586 : /* There are some options that change IL semantics which means
587 : we cannot inline in these cases for correctness reason.
588 : Not even for always_inline declared functions. */
589 7308 : else if (check_match (flag_wrapv)
590 7308 : || check_match (flag_wrapv_pointer)
591 7307 : || check_match (flag_trapv)
592 7307 : || check_match (flag_pcc_struct_return)
593 7307 : || check_maybe_down (optimize_debug)
594 : /* When caller or callee does FP math, be sure FP codegen flags
595 : compatible. */
596 7301 : || ((caller_info->fp_expressions && callee_info->fp_expressions)
597 1273 : && (check_maybe_up (flag_rounding_math)
598 1273 : || check_maybe_up (flag_trapping_math)
599 1271 : || check_maybe_down (flag_unsafe_math_optimizations)
600 1271 : || check_maybe_down (flag_finite_math_only)
601 1270 : || check_maybe_up (flag_signaling_nans)
602 1270 : || check_maybe_up (flag_complex_method)
603 1269 : || check_maybe_up (flag_signed_zeros)
604 1269 : || check_maybe_down (flag_associative_math)
605 1253 : || check_maybe_down (flag_reciprocal_math)
606 1253 : || check_maybe_down (flag_fp_int_builtin_inexact)
607 : /* Strictly speaking only when the callee contains function
608 : calls that may end up setting errno. */
609 1253 : || check_maybe_up (flag_errno_math)))
610 : /* We do not want to make code compiled with exceptions to be
611 : brought into a non-EH function unless we know that the callee
612 : does not throw.
613 : This is tracked by DECL_FUNCTION_PERSONALITY. */
614 7281 : || (check_maybe_up (flag_non_call_exceptions)
615 0 : && DECL_FUNCTION_PERSONALITY (callee->decl))
616 7281 : || (check_maybe_up (flag_exceptions)
617 16 : && DECL_FUNCTION_PERSONALITY (callee->decl))
618 : /* When devirtualization is disabled for callee, it is not safe
619 : to inline it as we possibly mangled the type info.
620 : Allow early inlining of always inlines. */
621 14589 : || (!(flags & CAN_INLINE_EARLY) && check_maybe_down (flag_devirtualize)))
622 : {
623 35 : e->inline_failed = CIF_OPTIMIZATION_MISMATCH;
624 35 : inlinable = false;
625 : }
626 : /* gcc.dg/pr43564.c. Apply user-forced inline even at -O0. */
627 7273 : else if (always_inline)
628 : ;
629 : /* When user added an attribute to the callee honor it. */
630 7273 : else if (lookup_attribute ("optimize", DECL_ATTRIBUTES (callee->decl))
631 7273 : && opts_for_fn (caller->decl) != opts_for_fn (callee->decl))
632 : {
633 2456 : e->inline_failed = CIF_OPTIMIZATION_MISMATCH;
634 2456 : inlinable = false;
635 : }
636 : /* If explicit optimize attribute are not used, the mismatch is caused
637 : by different command line options used to build different units.
638 : Do not care about COMDAT functions - those are intended to be
639 : optimized with the optimization flags of module they are used in.
640 : Also do not care about mixing up size/speed optimization when
641 : DECL_DISREGARD_INLINE_LIMITS is set. */
642 4817 : else if ((callee->merged_comdat
643 0 : && !lookup_attribute ("optimize",
644 0 : DECL_ATTRIBUTES (caller->decl)))
645 4817 : || DECL_DISREGARD_INLINE_LIMITS (callee->decl))
646 : ;
647 : /* If mismatch is caused by merging two LTO units with different
648 : optimization flags we want to be bit nicer. However never inline
649 : if one of functions is not optimized at all. */
650 4817 : else if (!opt_for_fn (callee->decl, optimize)
651 4817 : || !opt_for_fn (caller->decl, optimize))
652 : {
653 0 : e->inline_failed = CIF_OPTIMIZATION_MISMATCH;
654 0 : inlinable = false;
655 : }
656 : /* If callee is optimized for size and caller is not, allow inlining if
657 : code shrinks or we are in param_max_inline_insns_single limit and
658 : callee is inline (and thus likely an unified comdat).
659 : This will allow caller to run faster. */
660 4817 : else if (opt_for_fn (callee->decl, optimize_size)
661 4817 : > opt_for_fn (caller->decl, optimize_size))
662 : {
663 118 : int growth = estimate_edge_growth (e);
664 118 : if (growth > opt_for_fn (caller->decl, param_max_inline_insns_size)
665 118 : && (!DECL_DECLARED_INLINE_P (callee->decl)
666 65 : && growth >= MAX (inline_insns_single (caller, false, false),
667 : inline_insns_auto (caller, false, false))))
668 : {
669 0 : e->inline_failed = CIF_OPTIMIZATION_MISMATCH;
670 0 : inlinable = false;
671 : }
672 : }
673 : /* If callee is more aggressively optimized for performance than caller,
674 : we generally want to inline only cheap (runtime wise) functions. */
675 4699 : else if (opt_for_fn (callee->decl, optimize_size)
676 : < opt_for_fn (caller->decl, optimize_size)
677 4699 : || (opt_for_fn (callee->decl, optimize)
678 : > opt_for_fn (caller->decl, optimize)))
679 : {
680 12675 : if (estimate_edge_time (e)
681 4225 : >= 20 + ipa_call_summaries->get (e)->call_stmt_time)
682 : {
683 1492 : e->inline_failed = CIF_OPTIMIZATION_MISMATCH;
684 1492 : inlinable = false;
685 : }
686 : }
687 :
688 : }
689 :
690 74172 : if (!inlinable && (flags & CAN_INLINE_REPORT))
691 67477 : report_inline_failed_reason (e);
692 : return inlinable;
693 : }
694 :
695 :
696 : /* Return true if the edge E is inlinable during early inlining. */
697 :
698 : static bool
699 4425763 : can_early_inline_edge_p (struct cgraph_edge *e)
700 : {
701 4424200 : cgraph_node *caller = (e->caller->inlined_to
702 4425763 : ? e->caller->inlined_to : e->caller);
703 4425763 : struct cgraph_node *callee = e->callee->ultimate_alias_target ();
704 : /* Early inliner might get called at WPA stage when IPA pass adds new
705 : function. In this case we cannot really do any of early inlining
706 : because function bodies are missing. */
707 4425763 : if (cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR)
708 : return false;
709 4425431 : if (!gimple_has_body_p (callee->decl))
710 : {
711 0 : e->inline_failed = CIF_BODY_NOT_AVAILABLE;
712 0 : return false;
713 : }
714 8850862 : gcc_assert (gimple_in_ssa_p (DECL_STRUCT_FUNCTION (e->caller->decl))
715 : && gimple_in_ssa_p (DECL_STRUCT_FUNCTION (callee->decl)));
716 4425431 : if (coverage_instrumentation_p ()
717 4427107 : && ((lookup_attribute ("no_profile_instrument_function",
718 1676 : DECL_ATTRIBUTES (caller->decl)) == NULL_TREE)
719 1676 : != (lookup_attribute ("no_profile_instrument_function",
720 3352 : DECL_ATTRIBUTES (callee->decl)) == NULL_TREE)))
721 : return false;
722 :
723 4425429 : if (!can_inline_edge_p (e, true, true)
724 4425429 : || !can_inline_edge_by_limits_p (e, CAN_INLINE_EARLY | CAN_INLINE_REPORT))
725 : return false;
726 : /* When inlining regular functions into always-inline functions
727 : during early inlining watch for possible inline cycles. */
728 4336019 : if (DECL_DISREGARD_INLINE_LIMITS (caller->decl)
729 276872 : && lookup_attribute ("always_inline", DECL_ATTRIBUTES (caller->decl))
730 4612015 : && (!DECL_DISREGARD_INLINE_LIMITS (callee->decl)
731 164109 : || !lookup_attribute ("always_inline", DECL_ATTRIBUTES (callee->decl))))
732 : {
733 : /* If there are indirect calls, inlining may produce direct call.
734 : TODO: We may lift this restriction if we avoid errors on formerly
735 : indirect calls to always_inline functions. Taking address
736 : of always_inline function is generally bad idea and should
737 : have been declared as undefined, but sadly we allow this. */
738 111888 : if (caller->indirect_calls || e->callee->indirect_calls)
739 : return false;
740 110326 : ipa_fn_summary *callee_info = ipa_fn_summaries->get (callee);
741 110326 : if (callee_info->safe_to_inline_to_always_inline)
742 31230 : return callee_info->safe_to_inline_to_always_inline - 1;
743 174235 : for (cgraph_edge *e2 = callee->callees; e2; e2 = e2->next_callee)
744 : {
745 95165 : struct cgraph_node *callee2 = e2->callee->ultimate_alias_target ();
746 : /* As early inliner runs in RPO order, we will see uninlined
747 : always_inline calls only in the case of cyclic graphs. */
748 95165 : if (DECL_DISREGARD_INLINE_LIMITS (callee2->decl)
749 95165 : || lookup_attribute ("always_inline", DECL_ATTRIBUTES (callee2->decl)))
750 : {
751 0 : callee_info->safe_to_inline_to_always_inline = 1;
752 0 : return false;
753 : }
754 : /* With LTO watch for case where function is later replaced
755 : by always_inline definition.
756 : TODO: We may either stop treating noninlined cross-module always
757 : inlines as errors, or we can extend decl merging to produce
758 : syntacic alias and honor always inline only in units it has
759 : been declared as such. */
760 95165 : if (flag_lto && callee2->externally_visible)
761 : {
762 26 : callee_info->safe_to_inline_to_always_inline = 1;
763 26 : return false;
764 : }
765 : }
766 79070 : callee_info->safe_to_inline_to_always_inline = 2;
767 : }
768 : return true;
769 : }
770 :
771 :
772 : /* Return number of calls in N. Ignore cheap builtins. */
773 :
774 : static int
775 914205 : num_calls (struct cgraph_node *n)
776 : {
777 914205 : struct cgraph_edge *e;
778 914205 : int num = 0;
779 :
780 1849898 : for (e = n->callees; e; e = e->next_callee)
781 935693 : if (!is_inexpensive_builtin (e->callee->decl))
782 843084 : num++;
783 914205 : return num;
784 : }
785 :
786 :
787 : /* Return true if we are interested in inlining small function. */
788 :
789 : static bool
790 3702183 : want_early_inline_function_p (struct cgraph_edge *e)
791 : {
792 3702183 : bool want_inline = true;
793 3702183 : struct cgraph_node *callee = e->callee->ultimate_alias_target ();
794 :
795 3702183 : if (DECL_DISREGARD_INLINE_LIMITS (callee->decl))
796 : ;
797 3702171 : else if (!DECL_DECLARED_INLINE_P (callee->decl)
798 3702171 : && !opt_for_fn (e->caller->decl, flag_inline_small_functions))
799 : {
800 62 : e->inline_failed = CIF_FUNCTION_NOT_INLINE_CANDIDATE;
801 62 : report_inline_failed_reason (e);
802 62 : want_inline = false;
803 : }
804 : else
805 : {
806 : /* First take care of very large functions. */
807 3702109 : int min_growth = estimate_min_edge_growth (e), growth = 0;
808 3702109 : int n;
809 3702109 : int early_inlining_insns = param_early_inlining_insns;
810 :
811 3702109 : if (min_growth > early_inlining_insns)
812 : {
813 436886 : if (dump_enabled_p ())
814 40 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
815 : " will not early inline: %C->%C, "
816 : "call is cold and code would grow "
817 : "at least by %i\n",
818 : e->caller, callee,
819 : min_growth);
820 : want_inline = false;
821 : }
822 : else
823 3265223 : growth = estimate_edge_growth (e);
824 :
825 :
826 3265263 : if (!want_inline || growth <= param_max_inline_insns_size)
827 : ;
828 1321399 : else if (!e->maybe_hot_p ())
829 : {
830 21402 : if (dump_enabled_p ())
831 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
832 : " will not early inline: %C->%C, "
833 : "call is cold and code would grow by %i\n",
834 : e->caller, callee,
835 : growth);
836 : want_inline = false;
837 : }
838 1299997 : else if (growth > early_inlining_insns)
839 : {
840 385792 : if (dump_enabled_p ())
841 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
842 : " will not early inline: %C->%C, "
843 : "growth %i exceeds --param early-inlining-insns\n",
844 : e->caller, callee, growth);
845 : want_inline = false;
846 : }
847 914205 : else if ((n = num_calls (callee)) != 0
848 914205 : && growth * (n + 1) > early_inlining_insns)
849 : {
850 271124 : if (dump_enabled_p ())
851 11 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
852 : " will not early inline: %C->%C, "
853 : "growth %i exceeds --param early-inlining-insns "
854 : "divided by number of calls\n",
855 : e->caller, callee, growth);
856 : want_inline = false;
857 : }
858 : }
859 3702183 : return want_inline;
860 : }
861 :
862 : /* Compute time of the edge->caller + edge->callee execution when inlining
863 : does not happen. */
864 :
865 : inline sreal
866 587504 : compute_uninlined_call_time (struct cgraph_edge *edge,
867 : sreal uninlined_call_time,
868 : sreal freq)
869 : {
870 428376 : cgraph_node *caller = (edge->caller->inlined_to
871 587504 : ? edge->caller->inlined_to
872 : : edge->caller);
873 :
874 587504 : if (freq > 0)
875 570199 : uninlined_call_time *= freq;
876 : else
877 17305 : uninlined_call_time = uninlined_call_time >> 11;
878 :
879 587504 : sreal caller_time = ipa_fn_summaries->get (caller)->time;
880 587504 : return uninlined_call_time + caller_time;
881 : }
882 :
883 : /* Same as compute_uinlined_call_time but compute time when inlining
884 : does happen. */
885 :
886 : inline sreal
887 587504 : compute_inlined_call_time (struct cgraph_edge *edge,
888 : sreal time,
889 : sreal freq)
890 : {
891 428376 : cgraph_node *caller = (edge->caller->inlined_to
892 587504 : ? edge->caller->inlined_to
893 : : edge->caller);
894 587504 : sreal caller_time = ipa_fn_summaries->get (caller)->time;
895 :
896 587504 : if (freq > 0)
897 570199 : time *= freq;
898 : else
899 17305 : time = time >> 11;
900 :
901 : /* This calculation should match one in ipa-inline-analysis.cc
902 : (estimate_edge_size_and_time). */
903 587504 : time -= (sreal)ipa_call_summaries->get (edge)->call_stmt_time * freq;
904 587504 : time += caller_time;
905 587504 : if (time <= 0)
906 0 : time = ((sreal) 1) >> 8;
907 587504 : gcc_checking_assert (time >= 0);
908 587504 : return time;
909 : }
910 :
911 : /* Determine time saved by inlining EDGE of frequency FREQ
912 : where callee's runtime w/o inlining is UNINLINED_TYPE
913 : and with inlined is INLINED_TYPE. */
914 :
915 : inline sreal
916 9645053 : inlining_speedup (struct cgraph_edge *edge,
917 : sreal freq,
918 : sreal uninlined_time,
919 : sreal inlined_time)
920 : {
921 9645053 : sreal speedup = uninlined_time - inlined_time;
922 : /* Handling of call_time should match one in ipa-inline-fnsummary.c
923 : (estimate_edge_size_and_time). */
924 9645053 : sreal call_time = ipa_call_summaries->get (edge)->call_stmt_time;
925 :
926 9645053 : if (freq > 0)
927 : {
928 9606652 : speedup = (speedup + call_time);
929 12016864 : if (freq != 1)
930 7196440 : speedup = speedup * freq;
931 : }
932 38401 : else if (freq == 0)
933 38401 : speedup = speedup >> 11;
934 9645053 : gcc_checking_assert (speedup >= 0);
935 9645053 : return speedup;
936 : }
937 :
938 : /* Return expected speedup of the callee function alone
939 : (i.e. not estimate of call overhead and also no scalling
940 : by call frequency. */
941 :
942 : static sreal
943 3401688 : callee_speedup (struct cgraph_edge *e)
944 : {
945 3401688 : sreal unspec_time;
946 3401688 : sreal spec_time = estimate_edge_time (e, &unspec_time);
947 3401688 : return unspec_time - spec_time;
948 : }
949 :
950 : /* Return true if the speedup for inlining E is bigger than
951 : param_inline_min_speedup. */
952 :
953 : static bool
954 587504 : big_speedup_p (struct cgraph_edge *e)
955 : {
956 587504 : sreal unspec_time;
957 587504 : sreal spec_time = estimate_edge_time (e, &unspec_time);
958 587504 : sreal freq = e->sreal_frequency ();
959 587504 : sreal time = compute_uninlined_call_time (e, unspec_time, freq);
960 587504 : sreal inlined_time = compute_inlined_call_time (e, spec_time, freq);
961 428376 : cgraph_node *caller = (e->caller->inlined_to
962 587504 : ? e->caller->inlined_to
963 : : e->caller);
964 587504 : int limit = opt_for_fn (caller->decl, param_inline_min_speedup);
965 :
966 587504 : if ((time - inlined_time) * 100 > time * limit)
967 : return true;
968 : return false;
969 : }
970 :
971 : /* Return true if we are interested in inlining small function.
972 : When REPORT is true, report reason to dump file. */
973 :
974 : static bool
975 5535260 : want_inline_small_function_p (struct cgraph_edge *e, bool report)
976 : {
977 5535260 : bool want_inline = true;
978 5535260 : struct cgraph_node *callee = e->callee->ultimate_alias_target ();
979 4190785 : cgraph_node *to = (e->caller->inlined_to
980 5535260 : ? e->caller->inlined_to : e->caller);
981 :
982 : /* Allow this function to be called before can_inline_edge_p,
983 : since it's usually cheaper. */
984 5535260 : if (cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR)
985 : want_inline = false;
986 5535260 : else if (DECL_DISREGARD_INLINE_LIMITS (callee->decl))
987 : return true;
988 5529285 : else if (!DECL_DECLARED_INLINE_P (callee->decl)
989 5529285 : && !opt_for_fn (e->caller->decl, flag_inline_small_functions))
990 : {
991 52809 : e->inline_failed = CIF_FUNCTION_NOT_INLINE_CANDIDATE;
992 52809 : want_inline = false;
993 : }
994 :
995 : /* Early return before lookup of summaries. */
996 52809 : if (!want_inline)
997 : {
998 52809 : if (report)
999 48546 : report_inline_failed_reason (e);
1000 : return false;
1001 : }
1002 :
1003 5476476 : ipa_fn_summary *callee_info = ipa_fn_summaries->get (callee);
1004 5476476 : ipa_call_summary *call_info = ipa_call_summaries->get (e);
1005 :
1006 : /* Do fast and conservative check if the function can be good
1007 : inline candidate. */
1008 5476476 : if ((!DECL_DECLARED_INLINE_P (callee->decl)
1009 2053927 : && (!e->count.ipa ().initialized_p ()
1010 51106 : || !e->maybe_hot_p (callee_info->time)))
1011 7530118 : && callee_info->min_size - call_info->call_stmt_size
1012 2053642 : > inline_insns_auto (e->caller, true, true))
1013 : {
1014 107 : e->inline_failed = CIF_MAX_INLINE_INSNS_AUTO_LIMIT;
1015 107 : want_inline = false;
1016 : }
1017 5476369 : else if ((DECL_DECLARED_INLINE_P (callee->decl)
1018 2053820 : || e->count.ipa ().nonzero_p ())
1019 8899227 : && callee_info->min_size - call_info->call_stmt_size
1020 3422858 : > inline_insns_single (e->caller, true, true))
1021 : {
1022 0 : e->inline_failed = (DECL_DECLARED_INLINE_P (callee->decl)
1023 0 : ? CIF_MAX_INLINE_INSNS_SINGLE_LIMIT
1024 : : CIF_MAX_INLINE_INSNS_AUTO_LIMIT);
1025 0 : want_inline = false;
1026 : }
1027 : else
1028 : {
1029 5476369 : int growth = estimate_edge_growth (e);
1030 5476369 : ipa_hints hints = estimate_edge_hints (e);
1031 : /* We have two independent groups of hints. If one matches in each
1032 : of groups the limits are inreased. If both groups matches, limit
1033 : is increased even more. */
1034 5476369 : bool apply_hints = (hints & (INLINE_HINT_indirect_call
1035 : | INLINE_HINT_known_hot
1036 : | INLINE_HINT_loop_iterations
1037 : | INLINE_HINT_loop_stride));
1038 5476369 : bool apply_hints2 = (hints & INLINE_HINT_builtin_constant_p);
1039 :
1040 5476369 : if (growth <= opt_for_fn (to->decl,
1041 : param_max_inline_insns_size))
1042 : ;
1043 : /* Apply param_max_inline_insns_single limit. Do not do so when
1044 : hints suggests that inlining given function is very profitable.
1045 : Avoid computation of big_speedup_p when not necessary to change
1046 : outcome of decision. */
1047 5337549 : else if (DECL_DECLARED_INLINE_P (callee->decl)
1048 3360734 : && growth >= inline_insns_single (e->caller, apply_hints,
1049 : apply_hints2)
1050 6044317 : && (apply_hints || apply_hints2
1051 705545 : || growth >= inline_insns_single (e->caller, true,
1052 : apply_hints2)
1053 365288 : || !big_speedup_p (e)))
1054 : {
1055 705417 : e->inline_failed = CIF_MAX_INLINE_INSNS_SINGLE_LIMIT;
1056 705417 : want_inline = false;
1057 : }
1058 4632132 : else if (!DECL_DECLARED_INLINE_P (callee->decl)
1059 1976815 : && !opt_for_fn (e->caller->decl, flag_inline_functions)
1060 4637483 : && growth >= opt_for_fn (to->decl,
1061 : param_max_inline_insns_small))
1062 : {
1063 : /* growth_positive_p is expensive, always test it last. */
1064 5351 : if (growth >= inline_insns_single (e->caller, false, false)
1065 5351 : || growth_positive_p (callee, e, growth))
1066 : {
1067 4964 : e->inline_failed = CIF_NOT_DECLARED_INLINED;
1068 4964 : want_inline = false;
1069 : }
1070 : }
1071 : /* Apply param_max_inline_insns_auto limit for functions not declared
1072 : inline. Bypass the limit when speedup seems big. */
1073 4626781 : else if (!DECL_DECLARED_INLINE_P (callee->decl)
1074 1971464 : && growth >= inline_insns_auto (e->caller, apply_hints,
1075 : apply_hints2)
1076 5863006 : && (apply_hints || apply_hints2
1077 1220613 : || growth >= inline_insns_auto (e->caller, true,
1078 : apply_hints2)
1079 222031 : || !big_speedup_p (e)))
1080 : {
1081 : /* growth_positive_p is expensive, always test it last. */
1082 1225093 : if (growth >= inline_insns_single (e->caller, false, false)
1083 1225093 : || growth_positive_p (callee, e, growth))
1084 : {
1085 1142500 : e->inline_failed = CIF_MAX_INLINE_INSNS_AUTO_LIMIT;
1086 1142500 : want_inline = false;
1087 : }
1088 : }
1089 : /* If call is cold, do not inline when function body would grow. */
1090 3401688 : else if (!e->maybe_hot_p (callee_speedup (e))
1091 3401688 : && (growth >= inline_insns_single (e->caller, false, false)
1092 777147 : || growth_positive_p (callee, e, growth)))
1093 : {
1094 698376 : e->inline_failed = CIF_UNLIKELY_CALL;
1095 698376 : want_inline = false;
1096 : }
1097 : }
1098 5476476 : if (!want_inline && report)
1099 609828 : report_inline_failed_reason (e);
1100 : return want_inline;
1101 : }
1102 :
1103 : /* EDGE is self recursive edge.
1104 : We handle two cases - when function A is inlining into itself
1105 : or when function A is being inlined into another inliner copy of function
1106 : A within function B.
1107 :
1108 : In first case OUTER_NODE points to the toplevel copy of A, while
1109 : in the second case OUTER_NODE points to the outermost copy of A in B.
1110 :
1111 : In both cases we want to be extra selective since
1112 : inlining the call will just introduce new recursive calls to appear. */
1113 :
1114 : static bool
1115 20160 : want_inline_self_recursive_call_p (struct cgraph_edge *edge,
1116 : struct cgraph_node *outer_node,
1117 : bool peeling,
1118 : int depth)
1119 : {
1120 20160 : char const *reason = NULL;
1121 20160 : bool want_inline = true;
1122 20160 : sreal caller_freq = 1;
1123 20160 : int max_depth = opt_for_fn (outer_node->decl,
1124 : param_max_inline_recursive_depth_auto);
1125 :
1126 20160 : if (DECL_DECLARED_INLINE_P (edge->caller->decl))
1127 3043 : max_depth = opt_for_fn (outer_node->decl,
1128 : param_max_inline_recursive_depth);
1129 :
1130 20160 : if (!edge->maybe_hot_p ())
1131 : {
1132 : reason = "recursive call is cold";
1133 : want_inline = false;
1134 : }
1135 19676 : else if (depth > max_depth)
1136 : {
1137 : reason = "--param max-inline-recursive-depth exceeded.";
1138 : want_inline = false;
1139 : }
1140 17437 : else if (outer_node->inlined_to
1141 21948 : && (caller_freq = outer_node->callers->sreal_frequency ()) == 0)
1142 : {
1143 : reason = "caller frequency is 0";
1144 : want_inline = false;
1145 : }
1146 :
1147 17437 : if (!want_inline)
1148 : ;
1149 : /* Inlining of self recursive function into copy of itself within other
1150 : function is transformation similar to loop peeling.
1151 :
1152 : Peeling is profitable if we can inline enough copies to make probability
1153 : of actual call to the self recursive function very small. Be sure that
1154 : the probability of recursion is small.
1155 :
1156 : We ensure that the frequency of recursing is at most 1 - (1/max_depth).
1157 : This way the expected number of recursion is at most max_depth. */
1158 17437 : else if (peeling)
1159 : {
1160 4511 : sreal max_prob = (sreal)1 - ((sreal)1 / (sreal)max_depth);
1161 4511 : int i;
1162 10001 : for (i = 1; i < depth; i++)
1163 5490 : max_prob = max_prob * max_prob;
1164 4511 : if (edge->sreal_frequency () >= max_prob * caller_freq)
1165 : {
1166 1615 : reason = "frequency of recursive call is too large";
1167 1615 : want_inline = false;
1168 : }
1169 : }
1170 : /* Recursive inlining, i.e. equivalent of unrolling, is profitable if
1171 : recursion depth is large. We reduce function call overhead and increase
1172 : chances that things fit in hardware return predictor.
1173 :
1174 : Recursive inlining might however increase cost of stack frame setup
1175 : actually slowing down functions whose recursion tree is wide rather than
1176 : deep.
1177 :
1178 : Deciding reliably on when to do recursive inlining without profile feedback
1179 : is tricky. For now we disable recursive inlining when probability of self
1180 : recursion is low.
1181 :
1182 : Recursive inlining of self recursive call within loop also results in
1183 : large loop depths that generally optimize badly. We may want to throttle
1184 : down inlining in those cases. In particular this seems to happen in one
1185 : of libstdc++ rb tree methods. */
1186 : else
1187 : {
1188 12926 : if (edge->sreal_frequency () * 100
1189 12926 : <= caller_freq
1190 25852 : * opt_for_fn (outer_node->decl,
1191 : param_min_inline_recursive_probability))
1192 : {
1193 806 : reason = "frequency of recursive call is too small";
1194 806 : want_inline = false;
1195 : }
1196 : }
1197 20160 : if (!can_inline_edge_by_limits_p (edge, CAN_INLINE_FORCE_LIMITS | CAN_INLINE_REPORT))
1198 : {
1199 : reason = "inline limits exceeded for always_inline function";
1200 : want_inline = false;
1201 : }
1202 20160 : if (!want_inline && dump_enabled_p ())
1203 7 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, edge->call_stmt,
1204 : " not inlining recursively: %s\n", reason);
1205 20160 : return want_inline;
1206 : }
1207 :
1208 : /* Return true when NODE has uninlinable caller;
1209 : set HAS_HOT_CALL if it has hot call.
1210 : Worker for cgraph_for_node_and_aliases. */
1211 :
1212 : static bool
1213 77976 : check_callers (struct cgraph_node *node, void *has_hot_call)
1214 : {
1215 77976 : struct cgraph_edge *e;
1216 135672 : for (e = node->callers; e; e = e->next_caller)
1217 : {
1218 86138 : if (!opt_for_fn (e->caller->decl, flag_inline_functions_called_once)
1219 86138 : || !opt_for_fn (e->caller->decl, optimize))
1220 : return true;
1221 86138 : if (!can_inline_edge_p (e, true))
1222 : return true;
1223 86099 : if (e->recursive_p ())
1224 : return true;
1225 86099 : if (!can_inline_edge_by_limits_p (e, CAN_INLINE_REPORT))
1226 : return true;
1227 : /* Inlining large functions to large loop depth is often harmful because
1228 : of register pressure it implies. */
1229 57768 : if ((int)ipa_call_summaries->get (e)->loop_depth
1230 57768 : > param_inline_functions_called_once_loop_depth)
1231 : return true;
1232 : /* Do not produce gigantic functions. */
1233 105754 : if (estimate_size_after_inlining (e->caller->inlined_to ?
1234 : e->caller->inlined_to : e->caller, e)
1235 57768 : > param_inline_functions_called_once_insns)
1236 : return true;
1237 57696 : if (!(*(bool *)has_hot_call) && e->maybe_hot_p ())
1238 12020 : *(bool *)has_hot_call = true;
1239 : }
1240 : return false;
1241 : }
1242 :
1243 : /* If NODE has a caller, return true. */
1244 :
1245 : static bool
1246 2204388 : has_caller_p (struct cgraph_node *node, void *data ATTRIBUTE_UNUSED)
1247 : {
1248 2204388 : if (node->callers)
1249 691577 : return true;
1250 : return false;
1251 : }
1252 :
1253 : /* Decide if inlining NODE would reduce unit size by eliminating
1254 : the offline copy of function.
1255 : When COLD is true the cold calls are considered, too. */
1256 :
1257 : static bool
1258 5052771 : want_inline_function_to_all_callers_p (struct cgraph_node *node, bool cold)
1259 : {
1260 5052771 : bool has_hot_call = false;
1261 :
1262 : /* Aliases gets inlined along with the function they alias. */
1263 5052771 : if (node->alias)
1264 : return false;
1265 : /* Already inlined? */
1266 4972971 : if (node->inlined_to)
1267 : return false;
1268 : /* Does it have callers? */
1269 2154086 : if (!node->call_for_symbol_and_aliases (has_caller_p, NULL, true))
1270 : return false;
1271 : /* Inlining into all callers would increase size? */
1272 691577 : if (growth_positive_p (node, NULL, INT_MIN) > 0)
1273 : return false;
1274 : /* All inlines must be possible. */
1275 72915 : if (node->call_for_symbol_and_aliases (check_callers, &has_hot_call,
1276 : true))
1277 : return false;
1278 44473 : if (!cold && !has_hot_call)
1279 : return false;
1280 : return true;
1281 : }
1282 :
1283 : /* Return true if WHERE of SIZE is a possible candidate for wrapper heuristics
1284 : in estimate_edge_badness. */
1285 :
1286 : static bool
1287 661359 : wrapper_heuristics_may_apply (struct cgraph_node *where, int size)
1288 : {
1289 661359 : return size < (DECL_DECLARED_INLINE_P (where->decl)
1290 661359 : ? inline_insns_single (where, false, false)
1291 373916 : : inline_insns_auto (where, false, false));
1292 : }
1293 :
1294 : /* A cost model driving the inlining heuristics in a way so the edges with
1295 : smallest badness are inlined first. After each inlining is performed
1296 : the costs of all caller edges of nodes affected are recomputed so the
1297 : metrics may accurately depend on values such as number of inlinable callers
1298 : of the function or function body size. */
1299 :
1300 : static sreal
1301 10219620 : edge_badness (struct cgraph_edge *edge, bool dump)
1302 : {
1303 10219620 : sreal badness;
1304 10219620 : int growth;
1305 10219620 : sreal edge_time, unspec_edge_time;
1306 10219620 : struct cgraph_node *callee = edge->callee->ultimate_alias_target ();
1307 10219620 : class ipa_fn_summary *callee_info = ipa_fn_summaries->get (callee);
1308 10219620 : ipa_hints hints;
1309 7765988 : cgraph_node *caller = (edge->caller->inlined_to
1310 10219620 : ? edge->caller->inlined_to
1311 : : edge->caller);
1312 :
1313 10219620 : growth = estimate_edge_growth (edge);
1314 10219620 : edge_time = estimate_edge_time (edge, &unspec_edge_time);
1315 10219620 : hints = estimate_edge_hints (edge);
1316 10219620 : gcc_checking_assert (edge_time >= 0);
1317 : /* Check that inlined time is better, but tolerate some roundoff issues.
1318 : FIXME: When callee profile drops to 0 we account calls more. This
1319 : should be fixed by never doing that. */
1320 10219620 : gcc_checking_assert ((edge_time * 100
1321 : - callee_info->time * 101).to_int () <= 0
1322 : || callee->count.ipa ().initialized_p ());
1323 10219620 : gcc_checking_assert (growth <= ipa_size_summaries->get (callee)->size);
1324 :
1325 10219620 : if (dump)
1326 : {
1327 185 : fprintf (dump_file, " Badness calculation for %s -> %s\n",
1328 185 : edge->caller->dump_name (),
1329 185 : edge->callee->dump_name ());
1330 185 : fprintf (dump_file, " size growth %i, time %f unspec %f ",
1331 : growth,
1332 : edge_time.to_double (),
1333 : unspec_edge_time.to_double ());
1334 185 : ipa_dump_hints (dump_file, hints);
1335 185 : if (big_speedup_p (edge))
1336 151 : fprintf (dump_file, " big_speedup");
1337 185 : fprintf (dump_file, "\n");
1338 : }
1339 :
1340 : /* Always prefer inlining saving code size. */
1341 10219620 : if (growth <= 0)
1342 : {
1343 527306 : badness = (sreal) (-SREAL_MIN_SIG + growth) << (SREAL_MAX_EXP / 256);
1344 527306 : if (dump)
1345 108 : fprintf (dump_file, " %f: Growth %d <= 0\n", badness.to_double (),
1346 : growth);
1347 : }
1348 : /* Inlining into EXTERNAL functions is not going to change anything unless
1349 : they are themselves inlined. */
1350 9692314 : else if (DECL_EXTERNAL (caller->decl))
1351 : {
1352 45982 : if (dump)
1353 0 : fprintf (dump_file, " max: function is external\n");
1354 45982 : return sreal::max ();
1355 : }
1356 : /* When profile is available. Compute badness as:
1357 :
1358 : time_saved * caller_count
1359 : goodness = -------------------------------------------------
1360 : growth_of_caller * overall_growth * combined_size
1361 :
1362 : badness = - goodness
1363 :
1364 : Again use negative value to make calls with profile appear hotter
1365 : then calls without.
1366 : */
1367 9646332 : else if (opt_for_fn (caller->decl, flag_guess_branch_prob)
1368 9646332 : || caller->count.ipa ().nonzero_p ())
1369 : {
1370 9644976 : sreal numerator, denominator;
1371 9644976 : int overall_growth;
1372 9644976 : sreal freq = edge->sreal_frequency ();
1373 :
1374 9644976 : numerator = inlining_speedup (edge, freq, unspec_edge_time, edge_time);
1375 9644976 : if (numerator <= 0)
1376 8855 : numerator = ((sreal) 1 >> 8);
1377 9644976 : if (caller->count.ipa ().nonzero_p ())
1378 204 : numerator *= caller->count.ipa ().to_gcov_type ();
1379 9644772 : else if (caller->count.ipa ().initialized_p ())
1380 926 : numerator = numerator >> 11;
1381 9644976 : denominator = growth;
1382 :
1383 9644976 : overall_growth = callee_info->growth;
1384 :
1385 : /* Look for inliner wrappers of the form:
1386 :
1387 : inline_caller ()
1388 : {
1389 : do_fast_job...
1390 : if (need_more_work)
1391 : noninline_callee ();
1392 : }
1393 : Without penalizing this case, we usually inline noninline_callee
1394 : into the inline_caller because overall_growth is small preventing
1395 : further inlining of inline_caller.
1396 :
1397 : Penalize only callgraph edges to functions with small overall
1398 : growth ...
1399 : */
1400 9644976 : if (growth > overall_growth
1401 : /* ... and having only one caller which is not inlined ... */
1402 2163175 : && callee_info->single_caller
1403 1208554 : && !edge->caller->inlined_to
1404 : /* ... and edges executed only conditionally ... */
1405 831125 : && freq < 1
1406 : /* ... consider case where callee is not inline but caller is ... */
1407 9999022 : && ((!DECL_DECLARED_INLINE_P (edge->callee->decl)
1408 93946 : && DECL_DECLARED_INLINE_P (caller->decl))
1409 : /* ... or when early optimizers decided to split and edge
1410 : frequency still indicates splitting is a win ... */
1411 345588 : || (callee->split_part && !caller->split_part
1412 95474 : && freq * 100
1413 9835509 : < opt_for_fn (caller->decl,
1414 : param_partial_inlining_entry_probability)
1415 : /* ... and do not overwrite user specified hints. */
1416 95059 : && (!DECL_DECLARED_INLINE_P (edge->callee->decl)
1417 70721 : || DECL_DECLARED_INLINE_P (caller->decl)))))
1418 : {
1419 102079 : ipa_fn_summary *caller_info = ipa_fn_summaries->get (caller);
1420 102079 : int caller_growth = caller_info->growth;
1421 :
1422 : /* Only apply the penalty when caller looks like inline candidate,
1423 : and it is not called once. */
1424 54810 : if (!caller_info->single_caller && overall_growth < caller_growth
1425 52349 : && caller_info->inlinable
1426 154380 : && wrapper_heuristics_may_apply
1427 52301 : (caller, ipa_size_summaries->get (caller)->size))
1428 : {
1429 41966 : if (dump)
1430 1 : fprintf (dump_file,
1431 : " Wrapper penalty. Increasing growth %i to %i\n",
1432 : overall_growth, caller_growth);
1433 : overall_growth = caller_growth;
1434 : }
1435 : }
1436 9644976 : if (overall_growth > 0)
1437 : {
1438 : /* Strongly prefer functions with few callers that can be inlined
1439 : fully. The square root here leads to smaller binaries at average.
1440 : Watch however for extreme cases and return to linear function
1441 : when growth is large. */
1442 8215703 : if (overall_growth < 256)
1443 4519631 : overall_growth *= overall_growth;
1444 : else
1445 3696072 : overall_growth += 256 * 256 - 256;
1446 8215703 : denominator *= overall_growth;
1447 : }
1448 9644976 : denominator *= ipa_size_summaries->get (caller)->size + growth;
1449 :
1450 9644976 : badness = - numerator / denominator;
1451 :
1452 9644976 : if (dump)
1453 : {
1454 308 : fprintf (dump_file,
1455 : " %f: guessed profile. frequency %f, count %" PRId64
1456 : " caller count %" PRId64
1457 : " time saved %f"
1458 : " overall growth %i (current) %i (original)"
1459 : " %i (compensated)\n",
1460 : badness.to_double (),
1461 : freq.to_double (),
1462 77 : edge->count.ipa ().initialized_p ()
1463 0 : ? edge->count.ipa ().to_gcov_type () : -1,
1464 77 : caller->count.ipa ().initialized_p ()
1465 0 : ? caller->count.ipa ().to_gcov_type () : -1,
1466 154 : inlining_speedup (edge, freq, unspec_edge_time,
1467 : edge_time).to_double (),
1468 : estimate_growth (callee),
1469 : callee_info->growth, overall_growth);
1470 : }
1471 : }
1472 : /* When function local profile is not available or it does not give
1473 : useful information (i.e. frequency is zero), base the cost on
1474 : loop nest and overall size growth, so we optimize for overall number
1475 : of functions fully inlined in program. */
1476 : else
1477 : {
1478 1356 : int nest = MIN (ipa_call_summaries->get (edge)->loop_depth, 8);
1479 1356 : badness = growth;
1480 :
1481 : /* Decrease badness if call is nested. */
1482 1356 : if (badness > 0)
1483 1356 : badness = badness >> nest;
1484 : else
1485 0 : badness = badness << nest;
1486 1356 : if (dump)
1487 0 : fprintf (dump_file, " %f: no profile. nest %i\n",
1488 : badness.to_double (), nest);
1489 : }
1490 10173638 : gcc_checking_assert (badness != 0);
1491 :
1492 10173638 : if (edge->recursive_p ())
1493 19089 : badness = badness.shift (badness > 0 ? 4 : -4);
1494 10173638 : if ((hints & (INLINE_HINT_indirect_call
1495 : | INLINE_HINT_loop_iterations
1496 : | INLINE_HINT_loop_stride))
1497 9351047 : || callee_info->growth <= 0)
1498 4865609 : badness = badness.shift (badness > 0 ? -2 : 2);
1499 10173638 : if (hints & INLINE_HINT_builtin_constant_p)
1500 24228 : badness = badness.shift (badness > 0 ? -4 : 4);
1501 10173638 : if (hints & (INLINE_HINT_same_scc))
1502 72340 : badness = badness.shift (badness > 0 ? 3 : -3);
1503 10137467 : else if (hints & (INLINE_HINT_in_scc))
1504 172894 : badness = badness.shift (badness > 0 ? 2 : -2);
1505 10051020 : else if (hints & (INLINE_HINT_cross_module))
1506 3970 : badness = badness.shift (badness > 0 ? 1 : -1);
1507 10173638 : if (DECL_DISREGARD_INLINE_LIMITS (callee->decl))
1508 17896 : badness = badness.shift (badness > 0 ? -4 : 4);
1509 10164690 : else if ((hints & INLINE_HINT_declared_inline))
1510 15356561 : badness = badness.shift (badness > 0 ? -3 : 3);
1511 10173638 : if (dump)
1512 185 : fprintf (dump_file, " Adjusted by hints %f\n", badness.to_double ());
1513 10173638 : return badness;
1514 : }
1515 :
1516 : /* Recompute badness of EDGE and update its key in HEAP if needed. */
1517 : static inline void
1518 5467990 : update_edge_key (edge_heap_t *heap, struct cgraph_edge *edge)
1519 : {
1520 5467990 : sreal badness = edge_badness (edge, false);
1521 5467990 : if (edge->aux)
1522 : {
1523 4306813 : edge_heap_node_t *n = (edge_heap_node_t *) edge->aux;
1524 4306813 : gcc_checking_assert (n->get_data () == edge);
1525 :
1526 : /* fibonacci_heap::replace_key does busy updating of the
1527 : heap that is unnecessarily expensive.
1528 : We do lazy increases: after extracting minimum if the key
1529 : turns out to be out of date, it is re-inserted into heap
1530 : with correct value. */
1531 4306813 : if (badness < n->get_key ().badness)
1532 : {
1533 936841 : if (dump_file && (dump_flags & TDF_DETAILS))
1534 : {
1535 90 : fprintf (dump_file,
1536 : " decreasing badness %s -> %s, %f to %f\n",
1537 45 : edge->caller->dump_name (),
1538 45 : edge->callee->dump_name (),
1539 90 : n->get_key ().badness.to_double (),
1540 : badness.to_double ());
1541 : }
1542 936841 : inline_badness b (edge, badness);
1543 936841 : heap->decrease_key (n, b);
1544 : }
1545 : }
1546 : else
1547 : {
1548 1161177 : if (dump_file && (dump_flags & TDF_DETAILS))
1549 : {
1550 338 : fprintf (dump_file,
1551 : " enqueuing call %s -> %s, badness %f\n",
1552 169 : edge->caller->dump_name (),
1553 169 : edge->callee->dump_name (),
1554 : badness.to_double ());
1555 : }
1556 1161177 : inline_badness b (edge, badness);
1557 1161177 : edge->aux = heap->insert (b, edge);
1558 : }
1559 5467990 : }
1560 :
1561 :
1562 : /* NODE was inlined.
1563 : All caller edges needs to be reset because
1564 : size estimates change. Similarly callees needs reset
1565 : because better context may be known. */
1566 :
1567 : static void
1568 992716 : reset_edge_caches (struct cgraph_node *node)
1569 : {
1570 992716 : struct cgraph_edge *edge;
1571 992716 : struct cgraph_edge *e = node->callees;
1572 992716 : struct cgraph_node *where = node;
1573 992716 : struct ipa_ref *ref;
1574 :
1575 992716 : if (where->inlined_to)
1576 927212 : where = where->inlined_to;
1577 :
1578 992716 : reset_node_cache (where);
1579 :
1580 992716 : if (edge_growth_cache != NULL)
1581 3500630 : for (edge = where->callers; edge; edge = edge->next_caller)
1582 2510172 : if (edge->inline_failed)
1583 2510172 : edge_growth_cache->remove (edge);
1584 :
1585 1053094 : FOR_EACH_ALIAS (where, ref)
1586 120756 : reset_edge_caches (dyn_cast <cgraph_node *> (ref->referring));
1587 :
1588 992716 : if (!e)
1589 : return;
1590 :
1591 3192001 : while (true)
1592 3192001 : if (!e->inline_failed && e->callee->callees)
1593 : e = e->callee->callees;
1594 : else
1595 : {
1596 2545622 : if (edge_growth_cache != NULL && e->inline_failed)
1597 2300119 : edge_growth_cache->remove (e);
1598 2545622 : if (e->next_callee)
1599 : e = e->next_callee;
1600 : else
1601 : {
1602 1477366 : do
1603 : {
1604 1477366 : if (e->caller == node)
1605 : return;
1606 646379 : e = e->caller->callers;
1607 : }
1608 646379 : while (!e->next_callee);
1609 : e = e->next_callee;
1610 : }
1611 : }
1612 : }
1613 :
1614 : /* Recompute HEAP nodes for each of caller of NODE.
1615 : UPDATED_NODES track nodes we already visited, to avoid redundant work.
1616 : When CHECK_INLINABLITY_FOR is set, re-check for specified edge that
1617 : it is inlinable. Otherwise check all edges. */
1618 :
1619 : static void
1620 990152 : update_caller_keys (edge_heap_t *heap, struct cgraph_node *node,
1621 : bitmap updated_nodes,
1622 : struct cgraph_edge *check_inlinablity_for)
1623 : {
1624 990152 : struct cgraph_edge *edge;
1625 990152 : struct ipa_ref *ref;
1626 :
1627 990152 : if ((!node->alias && !ipa_fn_summaries->get (node)->inlinable)
1628 977817 : || node->inlined_to)
1629 12335 : return;
1630 977817 : if (!bitmap_set_bit (updated_nodes, node->get_summary_id ()))
1631 : return;
1632 :
1633 1037694 : FOR_EACH_ALIAS (node, ref)
1634 : {
1635 59877 : struct cgraph_node *alias = dyn_cast <cgraph_node *> (ref->referring);
1636 59877 : update_caller_keys (heap, alias, updated_nodes, check_inlinablity_for);
1637 : }
1638 :
1639 3379885 : for (edge = node->callers; edge; edge = edge->next_caller)
1640 2402068 : if (edge->inline_failed)
1641 : {
1642 2402068 : if (!check_inlinablity_for
1643 2402068 : || check_inlinablity_for == edge)
1644 : {
1645 2402068 : if (can_inline_edge_p (edge, false)
1646 2368142 : && want_inline_small_function_p (edge, false)
1647 3043815 : && can_inline_edge_by_limits_p (edge, 0))
1648 635875 : update_edge_key (heap, edge);
1649 1766193 : else if (edge->aux)
1650 : {
1651 102479 : report_inline_failed_reason (edge);
1652 102479 : heap->delete_node ((edge_heap_node_t *) edge->aux);
1653 102479 : edge->aux = NULL;
1654 : }
1655 : }
1656 0 : else if (edge->aux)
1657 0 : update_edge_key (heap, edge);
1658 : }
1659 : }
1660 :
1661 : /* Recompute HEAP nodes for each uninlined call in NODE
1662 : If UPDATE_SINCE is non-NULL check if edges called within that function
1663 : are inlinable (typically UPDATE_SINCE is the inline clone we introduced
1664 : where all edges have new context).
1665 :
1666 : This is used when we know that edge badnesses are going only to increase
1667 : (we introduced new call site) and thus all we need is to insert newly
1668 : created edges into heap. */
1669 :
1670 : static void
1671 930344 : update_callee_keys (edge_heap_t *heap, struct cgraph_node *node,
1672 : struct cgraph_node *update_since,
1673 : bitmap updated_nodes)
1674 : {
1675 930344 : struct cgraph_edge *e = node->callees;
1676 930344 : bool check_inlinability = update_since == node;
1677 :
1678 930344 : if (!e)
1679 : return;
1680 29267849 : while (true)
1681 29267849 : if (!e->inline_failed && e->callee->callees)
1682 : {
1683 5349330 : if (e->callee == update_since)
1684 433667 : check_inlinability = true;
1685 28370780 : e = e->callee->callees;
1686 : }
1687 : else
1688 : {
1689 23918519 : enum availability avail;
1690 23918519 : struct cgraph_node *callee;
1691 23918519 : if (!check_inlinability)
1692 : {
1693 21560975 : if (e->aux
1694 25044486 : && !bitmap_bit_p (updated_nodes,
1695 3483511 : e->callee->ultimate_alias_target
1696 3483511 : (&avail, e->caller)->get_summary_id ()))
1697 3483505 : update_edge_key (heap, e);
1698 : }
1699 : /* We do not reset callee growth cache here. Since we added a new call,
1700 : growth should have just increased and consequently badness metric
1701 : don't need updating. */
1702 2357544 : else if (e->inline_failed
1703 2257838 : && (callee = e->callee->ultimate_alias_target (&avail,
1704 2257838 : e->caller))
1705 2257838 : && avail >= AVAIL_AVAILABLE
1706 612006 : && ipa_fn_summaries->get (callee) != NULL
1707 611990 : && ipa_fn_summaries->get (callee)->inlinable
1708 2955240 : && !bitmap_bit_p (updated_nodes, callee->get_summary_id ()))
1709 : {
1710 597696 : if (can_inline_edge_p (e, false)
1711 592338 : && want_inline_small_function_p (e, false)
1712 970630 : && can_inline_edge_by_limits_p (e, 0))
1713 : {
1714 372111 : gcc_checking_assert (check_inlinability || can_inline_edge_p (e, false));
1715 : gcc_checking_assert (check_inlinability || e->aux);
1716 372111 : update_edge_key (heap, e);
1717 : }
1718 225585 : else if (e->aux)
1719 : {
1720 6997 : report_inline_failed_reason (e);
1721 6997 : heap->delete_node ((edge_heap_node_t *) e->aux);
1722 6997 : e->aux = NULL;
1723 : }
1724 : }
1725 : /* In case we redirected to unreachable node we only need to remove the
1726 : fibheap entry. */
1727 1759848 : else if (e->aux)
1728 : {
1729 3515 : heap->delete_node ((edge_heap_node_t *) e->aux);
1730 3515 : e->aux = NULL;
1731 : }
1732 23918519 : if (e->next_callee)
1733 : e = e->next_callee;
1734 : else
1735 : {
1736 6246399 : do
1737 : {
1738 6246399 : if (e->caller == node)
1739 897069 : return;
1740 5349330 : if (e->caller == update_since)
1741 433667 : check_inlinability = false;
1742 5349330 : e = e->caller->callers;
1743 : }
1744 5349330 : while (!e->next_callee);
1745 : e = e->next_callee;
1746 : }
1747 : }
1748 : }
1749 :
1750 : /* Enqueue all recursive calls from NODE into priority queue depending on
1751 : how likely we want to recursively inline the call. */
1752 :
1753 : static void
1754 22514 : lookup_recursive_calls (struct cgraph_node *node, struct cgraph_node *where,
1755 : edge_heap_t *heap)
1756 : {
1757 22514 : struct cgraph_edge *e;
1758 22514 : enum availability avail;
1759 :
1760 62660 : for (e = where->callees; e; e = e->next_callee)
1761 40146 : if (e->callee == node
1762 40146 : || (e->callee->ultimate_alias_target (&avail, e->caller) == node
1763 1193 : && avail > AVAIL_INTERPOSABLE))
1764 : {
1765 16473 : inline_badness b (e, -e->sreal_frequency ());
1766 16473 : heap->insert (b, e);
1767 : }
1768 62660 : for (e = where->callees; e; e = e->next_callee)
1769 40146 : if (!e->inline_failed)
1770 8353 : lookup_recursive_calls (node, e->callee, heap);
1771 22514 : }
1772 :
1773 : /* Decide on recursive inlining: in the case function has recursive calls,
1774 : inline until body size reaches given argument. If any new indirect edges
1775 : are discovered in the process, add them to *NEW_EDGES, unless NEW_EDGES
1776 : is NULL. */
1777 :
1778 : static bool
1779 2041 : recursive_inlining (struct cgraph_edge *edge,
1780 : vec<cgraph_edge *> *new_edges)
1781 : {
1782 1462 : cgraph_node *to = (edge->caller->inlined_to
1783 2041 : ? edge->caller->inlined_to : edge->caller);
1784 2041 : int limit = opt_for_fn (to->decl,
1785 : param_max_inline_insns_recursive_auto);
1786 2041 : inline_badness b (edge, sreal::min ());
1787 2041 : edge_heap_t heap (b);
1788 2041 : struct cgraph_node *node;
1789 2041 : struct cgraph_edge *e;
1790 2041 : struct cgraph_node *master_clone = NULL, *next;
1791 2041 : int depth = 0;
1792 2041 : int n = 0;
1793 :
1794 2041 : node = edge->caller;
1795 2041 : if (node->inlined_to)
1796 579 : node = node->inlined_to;
1797 :
1798 2041 : if (DECL_DECLARED_INLINE_P (node->decl))
1799 446 : limit = opt_for_fn (to->decl, param_max_inline_insns_recursive);
1800 :
1801 : /* Make sure that function is small enough to be considered for inlining. */
1802 2041 : if (estimate_size_after_inlining (node, edge) >= limit)
1803 : return false;
1804 2041 : lookup_recursive_calls (node, node, &heap);
1805 2041 : if (heap.empty ())
1806 : return false;
1807 :
1808 2041 : if (dump_file)
1809 4 : fprintf (dump_file,
1810 : " Performing recursive inlining on %s\n", node->dump_name ());
1811 :
1812 : /* Do the inlining and update list of recursive call during process. */
1813 17265 : while (!heap.empty ())
1814 : {
1815 15252 : struct cgraph_edge *curr = heap.extract_min ();
1816 15252 : struct cgraph_node *cnode, *dest = curr->callee;
1817 :
1818 15252 : if (!can_inline_edge_p (curr, true)
1819 15252 : || !can_inline_edge_by_limits_p (curr, CAN_INLINE_REPORT | CAN_INLINE_FORCE_LIMITS))
1820 0 : continue;
1821 :
1822 : /* MASTER_CLONE is produced in the case we already started modified
1823 : the function. Be sure to redirect edge to the original body before
1824 : estimating growths otherwise we will be seeing growths after inlining
1825 : the already modified body. */
1826 15252 : if (master_clone)
1827 : {
1828 13105 : curr->redirect_callee (master_clone);
1829 13105 : if (edge_growth_cache != NULL)
1830 13105 : edge_growth_cache->remove (curr);
1831 : }
1832 :
1833 15252 : if (estimate_size_after_inlining (node, curr) > limit)
1834 : {
1835 28 : curr->redirect_callee (dest);
1836 28 : if (edge_growth_cache != NULL)
1837 28 : edge_growth_cache->remove (curr);
1838 : break;
1839 : }
1840 :
1841 15224 : depth = 1;
1842 15224 : for (cnode = curr->caller;
1843 79145 : cnode->inlined_to; cnode = cnode->callers->caller)
1844 127842 : if (node->decl
1845 63921 : == curr->callee->ultimate_alias_target ()->decl)
1846 63921 : depth++;
1847 :
1848 15224 : if (!want_inline_self_recursive_call_p (curr, node, false, depth))
1849 : {
1850 3104 : curr->redirect_callee (dest);
1851 3104 : if (edge_growth_cache != NULL)
1852 3104 : edge_growth_cache->remove (curr);
1853 3104 : continue;
1854 : }
1855 :
1856 12120 : if (dump_file)
1857 : {
1858 14 : fprintf (dump_file,
1859 : " Inlining call of depth %i", depth);
1860 28 : if (node->count.nonzero_p () && curr->count.initialized_p ())
1861 : {
1862 2 : fprintf (dump_file, " called approx. %.2f times per call",
1863 2 : (double)curr->count.to_gcov_type ()
1864 2 : / node->count.to_gcov_type ());
1865 : }
1866 14 : fprintf (dump_file, "\n");
1867 : }
1868 12120 : if (!master_clone)
1869 : {
1870 : /* We need original clone to copy around. */
1871 1670 : master_clone = node->create_clone (node->decl, node->count,
1872 1670 : false, vNULL, true, NULL, NULL, NULL);
1873 4782 : for (e = master_clone->callees; e; e = e->next_callee)
1874 3112 : if (!e->inline_failed)
1875 511 : clone_inlined_nodes (e, true, true, false, NULL);
1876 1670 : curr->redirect_callee (master_clone);
1877 1670 : if (edge_growth_cache != NULL)
1878 1670 : edge_growth_cache->remove (curr);
1879 : }
1880 :
1881 12120 : inline_call (curr, false, new_edges, &overall_size, true);
1882 12120 : reset_node_cache (node);
1883 12120 : lookup_recursive_calls (node, curr->callee, &heap);
1884 12120 : n++;
1885 : }
1886 :
1887 2041 : if (!heap.empty () && dump_file)
1888 0 : fprintf (dump_file, " Recursive inlining growth limit met.\n");
1889 :
1890 2041 : if (!master_clone)
1891 : return false;
1892 :
1893 1670 : if (dump_enabled_p ())
1894 4 : dump_printf_loc (MSG_NOTE, edge->call_stmt,
1895 : "\n Inlined %i times, "
1896 : "body grown from size %i to %i, time %f to %f\n", n,
1897 4 : ipa_size_summaries->get (master_clone)->size,
1898 4 : ipa_size_summaries->get (node)->size,
1899 4 : ipa_fn_summaries->get (master_clone)->time.to_double (),
1900 4 : ipa_fn_summaries->get (node)->time.to_double ());
1901 :
1902 : /* Remove master clone we used for inlining. We rely that clones inlined
1903 : into master clone gets queued just before master clone so we don't
1904 : need recursion. */
1905 19670 : for (node = symtab->first_function (); node != master_clone;
1906 18000 : node = next)
1907 : {
1908 18000 : next = symtab->next_function (node);
1909 18000 : if (node->inlined_to == master_clone)
1910 974 : node->remove ();
1911 : }
1912 1670 : master_clone->remove ();
1913 1670 : return true;
1914 2041 : }
1915 :
1916 :
1917 : /* Given whole compilation unit estimate of INSNS, compute how large we can
1918 : allow the unit to grow. */
1919 :
1920 : static int64_t
1921 1045296 : compute_max_insns (cgraph_node *node, int insns)
1922 : {
1923 1045296 : int max_insns = insns;
1924 1045296 : if (max_insns < opt_for_fn (node->decl, param_large_unit_insns))
1925 : max_insns = opt_for_fn (node->decl, param_large_unit_insns);
1926 :
1927 1045296 : return ((int64_t) max_insns
1928 1045296 : * (100 + opt_for_fn (node->decl, param_inline_unit_growth)) / 100);
1929 : }
1930 :
1931 :
1932 : /* Compute badness of all edges in NEW_EDGES and add them to the HEAP. */
1933 :
1934 : static void
1935 928857 : add_new_edges_to_heap (edge_heap_t *heap, vec<cgraph_edge *> &new_edges)
1936 : {
1937 931626 : while (new_edges.length () > 0)
1938 : {
1939 2769 : struct cgraph_edge *edge = new_edges.pop ();
1940 :
1941 2769 : gcc_assert (!edge->aux);
1942 2769 : gcc_assert (edge->callee);
1943 2769 : if (edge->inline_failed
1944 2769 : && can_inline_edge_p (edge, true)
1945 1294 : && want_inline_small_function_p (edge, true)
1946 3691 : && can_inline_edge_by_limits_p (edge, CAN_INLINE_REPORT))
1947 : {
1948 922 : inline_badness b (edge, edge_badness (edge, false));
1949 922 : edge->aux = heap->insert (b, edge);
1950 : }
1951 : }
1952 928857 : }
1953 :
1954 : /* Remove EDGE from the fibheap. */
1955 :
1956 : static void
1957 8055 : heap_edge_removal_hook (struct cgraph_edge *e, void *data)
1958 : {
1959 8055 : if (e->aux)
1960 : {
1961 37 : ((edge_heap_t *)data)->delete_node ((edge_heap_node_t *)e->aux);
1962 37 : e->aux = NULL;
1963 : }
1964 8055 : }
1965 :
1966 : /* Return true if speculation of edge E seems useful.
1967 : If ANTICIPATE_INLINING is true, be conservative and hope that E
1968 : may get inlined. */
1969 :
1970 : bool
1971 123024 : speculation_useful_p (struct cgraph_edge *e, bool anticipate_inlining)
1972 : {
1973 : /* If we have already decided to inline the edge, it seems useful.
1974 : Also if ipa-cp or other pass worked hard enough to produce a clone,
1975 : we already decided this is a good idea. */
1976 123024 : if (!e->inline_failed
1977 35359 : || e->callee->clone_of)
1978 : return true;
1979 :
1980 34606 : enum availability avail;
1981 34606 : struct cgraph_node *target = e->callee->ultimate_alias_target (&avail,
1982 : e->callee);
1983 :
1984 34606 : gcc_assert (e->speculative && !e->indirect_unknown_callee);
1985 :
1986 : /* Even if call statement is not hot, we can still have useful speculation
1987 : in cases where a lot of time is spent is callee.
1988 : Do not check maybe_hot_p. */
1989 34606 : if (!e->count.nonzero_p ())
1990 : return false;
1991 :
1992 : /* See if IP optimizations found something potentially useful about the
1993 : function. Do this only if the call seems hot since this is about
1994 : optimizing the code surrounding call site rahter than improving
1995 : callee. */
1996 34564 : if (avail >= AVAIL_AVAILABLE && e->maybe_hot_p ())
1997 : {
1998 33618 : int ecf_flags = flags_from_decl_or_type (target->decl);
1999 33618 : if (ecf_flags & ECF_CONST)
2000 : {
2001 488 : if (!(e->speculative_call_indirect_edge ()->indirect_info
2002 488 : ->ecf_flags & ECF_CONST))
2003 : return true;
2004 : }
2005 33130 : else if (ecf_flags & ECF_PURE)
2006 : {
2007 4188 : if (!(e->speculative_call_indirect_edge ()->indirect_info
2008 4188 : ->ecf_flags & ECF_PURE))
2009 : return true;
2010 : }
2011 28942 : else if (get_modref_function_summary (target))
2012 : return true;
2013 : }
2014 : /* If we did not managed to inline the function nor redirect
2015 : to an ipa-cp clone (that are seen by having local flag set),
2016 : it is probably pointless to inline it unless hardware is missing
2017 : indirect call predictor.
2018 :
2019 : At this point we know we will not dispatch into faster version of
2020 : callee, so if call itself is not hot, we definitely can give up
2021 : speculating. */
2022 12452 : if (!anticipate_inlining && (!target->local || !e->maybe_hot_p ()))
2023 : return false;
2024 : /* For overwritable targets there is not much to do. */
2025 8455 : if (!can_inline_edge_p (e, false)
2026 8455 : || !can_inline_edge_by_limits_p (e, CAN_INLINE_DISREGARD_LIMITS))
2027 6 : return false;
2028 : /* OK, speculation seems interesting. */
2029 : return true;
2030 : }
2031 :
2032 : /* We know that EDGE is not going to be inlined.
2033 : See if we can remove speculation. */
2034 :
2035 : static void
2036 120104 : resolve_noninline_speculation (edge_heap_t *edge_heap, struct cgraph_edge *edge)
2037 : {
2038 120104 : if (edge->speculative && !speculation_useful_p (edge, false))
2039 : {
2040 981 : struct cgraph_node *node = edge->caller;
2041 976 : struct cgraph_node *where = node->inlined_to
2042 981 : ? node->inlined_to : node;
2043 981 : auto_bitmap updated_nodes;
2044 :
2045 981 : if (edge->count.ipa ().initialized_p ())
2046 0 : spec_rem += edge->count.ipa ();
2047 981 : cgraph_edge::resolve_speculation (edge);
2048 981 : reset_edge_caches (where);
2049 981 : ipa_update_overall_fn_summary (where);
2050 981 : update_caller_keys (edge_heap, where,
2051 : updated_nodes, NULL);
2052 981 : update_callee_keys (edge_heap, where, NULL,
2053 : updated_nodes);
2054 981 : }
2055 120104 : }
2056 :
2057 : /* Return true if NODE should be accounted for overall size estimate.
2058 : Skip all nodes optimized for size so we can measure the growth of hot
2059 : part of program no matter of the padding. */
2060 :
2061 : bool
2062 3839011 : inline_account_function_p (struct cgraph_node *node)
2063 : {
2064 3839011 : return (!DECL_EXTERNAL (node->decl)
2065 3645285 : && !opt_for_fn (node->decl, optimize_size)
2066 7387905 : && node->frequency != NODE_FREQUENCY_UNLIKELY_EXECUTED);
2067 : }
2068 :
2069 : /* Count number of callers of NODE and store it into DATA (that
2070 : points to int. Worker for cgraph_for_node_and_aliases. */
2071 :
2072 : static bool
2073 1504212 : sum_callers (struct cgraph_node *node, void *data)
2074 : {
2075 1504212 : struct cgraph_edge *e;
2076 1504212 : int *num_calls = (int *)data;
2077 :
2078 3611945 : for (e = node->callers; e; e = e->next_caller)
2079 2107733 : (*num_calls)++;
2080 1504212 : return false;
2081 : }
2082 :
2083 : /* We only propagate across edges with non-interposable callee. */
2084 :
2085 : inline bool
2086 7232082 : ignore_edge_p (struct cgraph_edge *e)
2087 : {
2088 7232082 : enum availability avail;
2089 7232082 : e->callee->function_or_virtual_thunk_symbol (&avail, e->caller);
2090 7232082 : return (avail <= AVAIL_INTERPOSABLE);
2091 : }
2092 :
2093 : /* We use greedy algorithm for inlining of small functions:
2094 : All inline candidates are put into prioritized heap ordered in
2095 : increasing badness.
2096 :
2097 : The inlining of small functions is bounded by unit growth parameters. */
2098 :
2099 : static void
2100 239354 : inline_small_functions (void)
2101 : {
2102 239354 : struct cgraph_node *node;
2103 239354 : struct cgraph_edge *edge;
2104 239354 : inline_badness b;
2105 239354 : edge_heap_t edge_heap (b);
2106 239354 : auto_bitmap updated_nodes;
2107 239354 : int min_size;
2108 239354 : auto_vec<cgraph_edge *> new_indirect_edges;
2109 239354 : int initial_size = 0;
2110 239354 : struct cgraph_node **order = XCNEWVEC (cgraph_node *, symtab->cgraph_count);
2111 239354 : struct cgraph_edge_hook_list *edge_removal_hook_holder;
2112 239354 : new_indirect_edges.create (8);
2113 :
2114 239354 : edge_removal_hook_holder
2115 239354 : = symtab->add_edge_removal_hook (&heap_edge_removal_hook, &edge_heap);
2116 :
2117 : /* Compute overall unit size and other global parameters used by badness
2118 : metrics. */
2119 :
2120 239354 : has_nonzero_ipa_profile = false;
2121 239354 : ipa_reduced_postorder (order, true, ignore_edge_p);
2122 239354 : free (order);
2123 :
2124 2191618 : FOR_EACH_DEFINED_FUNCTION (node)
2125 1952264 : if (!node->inlined_to)
2126 : {
2127 1849433 : if (!node->alias && node->analyzed
2128 1849433 : && (node->has_gimple_body_p () || node->thunk)
2129 3801663 : && opt_for_fn (node->decl, optimize))
2130 : {
2131 1397451 : class ipa_fn_summary *info = ipa_fn_summaries->get (node);
2132 1397451 : struct ipa_dfs_info *dfs = (struct ipa_dfs_info *) node->aux;
2133 :
2134 : /* Do not account external functions, they will be optimized out
2135 : if not inlined. Also only count the non-cold portion of program. */
2136 1397451 : if (inline_account_function_p (node))
2137 1289786 : initial_size += ipa_size_summaries->get (node)->size;
2138 1397451 : info->growth = estimate_growth (node);
2139 :
2140 1397451 : int num_calls = 0;
2141 1397451 : node->call_for_symbol_and_aliases (sum_callers, &num_calls,
2142 : true);
2143 1397451 : if (num_calls == 1)
2144 473625 : info->single_caller = true;
2145 1397451 : if (dfs && dfs->next_cycle)
2146 : {
2147 5250 : struct cgraph_node *n2;
2148 5250 : int id = dfs->scc_no + 1;
2149 11924 : for (n2 = node; n2;
2150 6674 : n2 = ((struct ipa_dfs_info *) n2->aux)->next_cycle)
2151 10501 : if (opt_for_fn (n2->decl, optimize))
2152 : {
2153 10493 : ipa_fn_summary *info2 = ipa_fn_summaries->get
2154 10493 : (n2->inlined_to ? n2->inlined_to : n2);
2155 10493 : if (info2->scc_no)
2156 : break;
2157 6666 : info2->scc_no = id;
2158 : }
2159 : }
2160 : }
2161 :
2162 4463197 : for (edge = node->callers; edge; edge = edge->next_caller)
2163 2510967 : if (edge->count.ipa ().initialized_p ()
2164 2646332 : && edge->count.ipa ().nonzero_p ())
2165 439 : has_nonzero_ipa_profile = true;
2166 : }
2167 239354 : ipa_free_postorder_info ();
2168 239354 : initialize_growth_caches ();
2169 :
2170 239354 : if (dump_file)
2171 178 : fprintf (dump_file,
2172 : "\nDeciding on inlining of small functions. Starting with size %i.\n",
2173 : initial_size);
2174 :
2175 239354 : overall_size = initial_size;
2176 239354 : min_size = overall_size;
2177 :
2178 : /* Populate the heap with all edges we might inline. */
2179 :
2180 2191618 : FOR_EACH_DEFINED_FUNCTION (node)
2181 : {
2182 1952264 : bool update = false;
2183 1952264 : struct cgraph_edge *next = NULL;
2184 1952264 : bool has_speculative = false;
2185 :
2186 1952264 : if (!opt_for_fn (node->decl, optimize)
2187 : /* With -Og we do not want to perform IPA inlining of small
2188 : functions since there are no scalar cleanups after it
2189 : that would realize the anticipated win. All abstraction
2190 : is removed during early inlining. */
2191 1952264 : || opt_for_fn (node->decl, optimize_debug))
2192 481027 : continue;
2193 :
2194 1471237 : if (dump_file)
2195 803 : fprintf (dump_file, "Enqueueing calls in %s.\n", node->dump_name ());
2196 :
2197 7259583 : for (edge = node->callees; edge; edge = edge->next_callee)
2198 : {
2199 5788346 : if (edge->inline_failed
2200 5788312 : && !edge->aux
2201 5788172 : && can_inline_edge_p (edge, true)
2202 1639827 : && want_inline_small_function_p (edge, true)
2203 984191 : && can_inline_edge_by_limits_p (edge, CAN_INLINE_REPORT)
2204 6764845 : && edge->inline_failed)
2205 : {
2206 976499 : gcc_assert (!edge->aux);
2207 976499 : update_edge_key (&edge_heap, edge);
2208 : }
2209 5788346 : if (edge->speculative)
2210 15907 : has_speculative = true;
2211 : }
2212 1471237 : if (has_speculative)
2213 64794 : for (edge = node->callees; edge; edge = next)
2214 : {
2215 56633 : next = edge->next_callee;
2216 56633 : if (edge->speculative
2217 56633 : && !speculation_useful_p (edge, edge->aux != NULL))
2218 : {
2219 486 : cgraph_edge::resolve_speculation (edge);
2220 486 : update = true;
2221 : }
2222 : }
2223 8161 : if (update)
2224 : {
2225 0 : struct cgraph_node *where = node->inlined_to
2226 412 : ? node->inlined_to : node;
2227 412 : ipa_update_overall_fn_summary (where);
2228 412 : reset_edge_caches (where);
2229 412 : update_caller_keys (&edge_heap, where,
2230 : updated_nodes, NULL);
2231 412 : update_callee_keys (&edge_heap, where, NULL,
2232 : updated_nodes);
2233 412 : bitmap_clear (updated_nodes);
2234 : }
2235 : }
2236 :
2237 239354 : gcc_assert (in_lto_p
2238 : || !has_nonzero_ipa_profile
2239 : || flag_auto_profile
2240 : || (profile_info && flag_branch_probabilities));
2241 :
2242 2777919 : while (!edge_heap.empty ())
2243 : {
2244 2538565 : int old_size = overall_size;
2245 2538565 : struct cgraph_node *where, *callee;
2246 2538565 : sreal badness = edge_heap.min_key ().badness;
2247 2538565 : sreal current_badness;
2248 2538565 : int growth;
2249 :
2250 2538565 : edge = edge_heap.extract_min ();
2251 2538565 : gcc_assert (edge->aux);
2252 2538565 : edge->aux = NULL;
2253 2538565 : if (!edge->inline_failed || !edge->callee->analyzed)
2254 1609683 : continue;
2255 :
2256 : /* Be sure that caches are maintained consistent.
2257 : This check is affected by scaling roundoff errors when compiling for
2258 : IPA this we skip it in that case. */
2259 2538477 : if (flag_checking && !edge->callee->count.ipa_p ()
2260 4750523 : && !has_nonzero_ipa_profile)
2261 : {
2262 2212043 : sreal cached_badness = edge_badness (edge, false);
2263 :
2264 2212043 : int old_size_est = estimate_edge_size (edge);
2265 2212043 : sreal old_time_est = estimate_edge_time (edge);
2266 2212043 : int old_hints_est = estimate_edge_hints (edge);
2267 :
2268 2212043 : if (edge_growth_cache != NULL)
2269 2212043 : edge_growth_cache->remove (edge);
2270 3806881 : reset_node_cache (edge->caller->inlined_to
2271 : ? edge->caller->inlined_to
2272 : : edge->caller);
2273 2212043 : gcc_assert (old_size_est == estimate_edge_size (edge));
2274 2212043 : gcc_assert (old_time_est == estimate_edge_time (edge));
2275 : /* FIXME:
2276 :
2277 : gcc_assert (old_hints_est == estimate_edge_hints (edge));
2278 :
2279 : fails with profile feedback because some hints depends on
2280 : maybe_hot_edge_p predicate and because callee gets inlined to other
2281 : calls, the edge may become cold.
2282 : This ought to be fixed by computing relative probabilities
2283 : for given invocation but that will be better done once whole
2284 : code is converted to sreals. Disable for now and revert to "wrong"
2285 : value so enable/disable checking paths agree. */
2286 2212043 : edge_growth_cache->get (edge)->hints = old_hints_est + 1;
2287 :
2288 : /* When updating the edge costs, we only decrease badness in the keys.
2289 : Increases of badness are handled lazily; when we see key with out
2290 : of date value on it, we re-insert it now. */
2291 2212043 : current_badness = edge_badness (edge, false);
2292 2212043 : gcc_assert (cached_badness == current_badness);
2293 2212043 : gcc_assert (current_badness >= badness);
2294 : }
2295 : else
2296 326437 : current_badness = edge_badness (edge, false);
2297 2538480 : if (current_badness != badness)
2298 : {
2299 1671494 : if (edge_heap.min () && current_badness > edge_heap.min_key ().badness)
2300 : {
2301 1489494 : inline_badness b (edge, current_badness);
2302 1489494 : edge->aux = edge_heap.insert (b, edge);
2303 1489494 : continue;
2304 1489494 : }
2305 : else
2306 182000 : badness = current_badness;
2307 : }
2308 :
2309 1048986 : if (!can_inline_edge_p (edge, true)
2310 1048986 : || !can_inline_edge_by_limits_p (edge, CAN_INLINE_REPORT))
2311 : {
2312 3690 : resolve_noninline_speculation (&edge_heap, edge);
2313 3690 : continue;
2314 : }
2315 :
2316 1045296 : callee = edge->callee->ultimate_alias_target ();
2317 1045296 : growth = estimate_edge_growth (edge);
2318 1045296 : if (dump_file)
2319 : {
2320 525 : fprintf (dump_file,
2321 : "\nConsidering %s with %i size\n",
2322 : callee->dump_name (),
2323 525 : ipa_size_summaries->get (callee)->size);
2324 1050 : fprintf (dump_file,
2325 : " to be inlined into %s in %s:%i\n"
2326 : " Estimated badness is %f, frequency %.2f.\n",
2327 525 : edge->caller->dump_name (),
2328 525 : edge->call_stmt
2329 497 : && (LOCATION_LOCUS (gimple_location ((const gimple *)
2330 : edge->call_stmt))
2331 : > BUILTINS_LOCATION)
2332 488 : ? gimple_filename ((const gimple *) edge->call_stmt)
2333 : : "unknown",
2334 525 : edge->call_stmt
2335 497 : ? gimple_lineno ((const gimple *) edge->call_stmt)
2336 : : -1,
2337 : badness.to_double (),
2338 525 : edge->sreal_frequency ().to_double ());
2339 525 : if (edge->count.ipa ().initialized_p ())
2340 : {
2341 0 : fprintf (dump_file, " Called ");
2342 0 : edge->count.ipa ().dump (dump_file);
2343 0 : fprintf (dump_file, " times\n");
2344 : }
2345 525 : if (dump_flags & TDF_DETAILS)
2346 185 : edge_badness (edge, true);
2347 : }
2348 :
2349 1045296 : where = edge->caller;
2350 :
2351 1045296 : if (overall_size + growth > compute_max_insns (where, min_size)
2352 1045296 : && !DECL_DISREGARD_INLINE_LIMITS (callee->decl))
2353 : {
2354 111637 : edge->inline_failed = CIF_INLINE_UNIT_GROWTH_LIMIT;
2355 111637 : report_inline_failed_reason (edge);
2356 111637 : resolve_noninline_speculation (&edge_heap, edge);
2357 111637 : continue;
2358 : }
2359 :
2360 933659 : if (!want_inline_small_function_p (edge, true))
2361 : {
2362 2366 : resolve_noninline_speculation (&edge_heap, edge);
2363 2366 : continue;
2364 : }
2365 :
2366 931293 : profile_count old_count = callee->count;
2367 :
2368 : /* Heuristics for inlining small functions work poorly for
2369 : recursive calls where we do effects similar to loop unrolling.
2370 : When inlining such edge seems profitable, leave decision on
2371 : specific inliner. */
2372 931293 : if (edge->recursive_p ())
2373 : {
2374 2041 : if (where->inlined_to)
2375 579 : where = where->inlined_to;
2376 :
2377 : /* Disable always_inline on self recursive functions.
2378 : This prevents some inlining bombs such as one in PR113291
2379 : from exploding.
2380 : It is not enough to stop inlining in self recursive always_inlines
2381 : since they may grow large enough so always inlining them even
2382 : with recursin depth 0 is too much.
2383 :
2384 : All sane uses of always_inline should be handled during
2385 : early optimizations. */
2386 2041 : DECL_DISREGARD_INLINE_LIMITS (where->decl) = false;
2387 :
2388 2041 : if (!recursive_inlining (edge,
2389 2041 : opt_for_fn (edge->caller->decl,
2390 : flag_indirect_inlining)
2391 : ? &new_indirect_edges : NULL))
2392 : {
2393 371 : edge->inline_failed = CIF_RECURSIVE_INLINING;
2394 371 : resolve_noninline_speculation (&edge_heap, edge);
2395 371 : continue;
2396 : }
2397 1670 : reset_edge_caches (where);
2398 : /* Recursive inliner inlines all recursive calls of the function
2399 : at once. Consequently we need to update all callee keys. */
2400 1670 : if (opt_for_fn (edge->caller->decl, flag_indirect_inlining))
2401 1645 : add_new_edges_to_heap (&edge_heap, new_indirect_edges);
2402 1670 : update_callee_keys (&edge_heap, where, where, updated_nodes);
2403 1670 : bitmap_clear (updated_nodes);
2404 : }
2405 : else
2406 : {
2407 929252 : struct cgraph_node *outer_node = NULL;
2408 929252 : int depth = 0;
2409 :
2410 : /* Consider the case where self recursive function A is inlined
2411 : into B. This is desired optimization in some cases, since it
2412 : leads to effect similar of loop peeling and we might completely
2413 : optimize out the recursive call. However we must be extra
2414 : selective. */
2415 :
2416 929252 : where = edge->caller;
2417 1492116 : while (where->inlined_to)
2418 : {
2419 562864 : if (where->decl == callee->decl)
2420 11268 : outer_node = where, depth++;
2421 562864 : where = where->callers->caller;
2422 : }
2423 931292 : if (outer_node
2424 929252 : && !want_inline_self_recursive_call_p (edge, outer_node,
2425 : true, depth))
2426 : {
2427 2040 : edge->inline_failed
2428 2040 : = (DECL_DISREGARD_INLINE_LIMITS (edge->callee->decl)
2429 2040 : ? CIF_RECURSIVE_INLINING : CIF_UNSPECIFIED);
2430 2040 : resolve_noninline_speculation (&edge_heap, edge);
2431 2040 : continue;
2432 : }
2433 927212 : else if (depth && dump_file)
2434 6 : fprintf (dump_file, " Peeling recursion with depth %i\n", depth);
2435 :
2436 927212 : gcc_checking_assert (!callee->inlined_to);
2437 :
2438 927212 : int old_size = ipa_size_summaries->get (where)->size;
2439 927212 : sreal old_time = ipa_fn_summaries->get (where)->time;
2440 :
2441 927212 : inline_call (edge, true, &new_indirect_edges, &overall_size, true);
2442 927212 : reset_edge_caches (edge->callee);
2443 927212 : add_new_edges_to_heap (&edge_heap, new_indirect_edges);
2444 :
2445 : /* If caller's size and time increased we do not need to update
2446 : all edges because badness is not going to decrease. */
2447 927212 : if (old_size <= ipa_size_summaries->get (where)->size
2448 871553 : && old_time <= ipa_fn_summaries->get (where)->time
2449 : /* Wrapper penalty may be non-monotonous in this respect.
2450 : Fortunately it only affects small functions. */
2451 1536270 : && !wrapper_heuristics_may_apply (where, old_size))
2452 425428 : update_callee_keys (&edge_heap, edge->callee, edge->callee,
2453 : updated_nodes);
2454 : else
2455 501784 : update_callee_keys (&edge_heap, where,
2456 : edge->callee,
2457 : updated_nodes);
2458 : }
2459 928882 : where = edge->caller;
2460 928882 : if (where->inlined_to)
2461 271957 : where = where->inlined_to;
2462 :
2463 : /* Our profitability metric can depend on local properties
2464 : such as number of inlinable calls and size of the function body.
2465 : After inlining these properties might change for the function we
2466 : inlined into (since it's body size changed) and for the functions
2467 : called by function we inlined (since number of it inlinable callers
2468 : might change). */
2469 928882 : update_caller_keys (&edge_heap, where, updated_nodes, NULL);
2470 : /* Offline copy count has possibly changed, recompute if profile is
2471 : available. */
2472 928882 : struct cgraph_node *n
2473 928882 : = cgraph_node::get (edge->callee->decl)->ultimate_alias_target ();
2474 632736 : if (n != edge->callee && n->analyzed && !(n->count == old_count)
2475 928951 : && n->count.ipa_p ())
2476 69 : update_callee_keys (&edge_heap, n, NULL, updated_nodes);
2477 928882 : bitmap_clear (updated_nodes);
2478 :
2479 928882 : if (dump_enabled_p ())
2480 : {
2481 534 : ipa_fn_summary *s = ipa_fn_summaries->get (where);
2482 :
2483 : /* dump_printf can't handle %+i. */
2484 534 : char buf_net_change[100];
2485 534 : snprintf (buf_net_change, sizeof buf_net_change, "%+i",
2486 : overall_size - old_size);
2487 :
2488 1068 : dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, edge->call_stmt,
2489 : " Inlined %C into %C which now has time %f and "
2490 : "size %i, net change of %s%s.\n",
2491 : edge->callee, edge->caller,
2492 : s->time.to_double (),
2493 534 : ipa_size_summaries->get (edge->caller)->size,
2494 : buf_net_change,
2495 534 : cross_module_call_p (edge)
2496 : ? " (cross module)" : "");
2497 : }
2498 928882 : if (min_size > overall_size)
2499 : {
2500 225233 : min_size = overall_size;
2501 :
2502 225233 : if (dump_file)
2503 398 : fprintf (dump_file, "New minimal size reached: %i\n", min_size);
2504 : }
2505 : }
2506 :
2507 239354 : free_growth_caches ();
2508 239354 : if (dump_enabled_p ())
2509 434 : dump_printf (MSG_NOTE,
2510 : "Unit growth for small function inlining: %i->%i (%i%%)\n",
2511 : initial_size, overall_size,
2512 193 : initial_size ? overall_size * 100 / (initial_size) - 100 : 0);
2513 239354 : symtab->remove_edge_removal_hook (edge_removal_hook_holder);
2514 239354 : }
2515 :
2516 : /* Flatten NODE. Performed both during early inlining and
2517 : at IPA inlining time. */
2518 :
2519 : static void
2520 716 : flatten_function (struct cgraph_node *node, bool early, bool update)
2521 : {
2522 716 : struct cgraph_edge *e;
2523 :
2524 : /* We shouldn't be called recursively when we are being processed. */
2525 716 : gcc_assert (node->aux == NULL);
2526 :
2527 716 : node->aux = (void *) node;
2528 :
2529 1648 : for (e = node->callees; e; e = e->next_callee)
2530 : {
2531 932 : struct cgraph_node *orig_callee;
2532 932 : struct cgraph_node *callee = e->callee->ultimate_alias_target ();
2533 :
2534 : /* We've hit cycle? It is time to give up. */
2535 932 : if (callee->aux)
2536 : {
2537 15 : if (dump_enabled_p ())
2538 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
2539 : "Not inlining %C into %C to avoid cycle.\n",
2540 : callee, e->caller);
2541 15 : if (cgraph_inline_failed_type (e->inline_failed) != CIF_FINAL_ERROR)
2542 15 : e->inline_failed = CIF_RECURSIVE_INLINING;
2543 15 : continue;
2544 : }
2545 :
2546 : /* When the edge is already inlined, we just need to recurse into
2547 : it in order to fully flatten the leaves. */
2548 917 : if (!e->inline_failed)
2549 : {
2550 350 : flatten_function (callee, early, false);
2551 350 : continue;
2552 : }
2553 :
2554 : /* Flatten attribute needs to be processed during late inlining. For
2555 : extra code quality we however do flattening during early optimization,
2556 : too. */
2557 321 : if (!early
2558 567 : ? !can_inline_edge_p (e, true)
2559 246 : && !can_inline_edge_by_limits_p (e, CAN_INLINE_REPORT)
2560 321 : : !can_early_inline_edge_p (e))
2561 419 : continue;
2562 :
2563 148 : if (e->recursive_p ())
2564 : {
2565 0 : if (dump_enabled_p ())
2566 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
2567 : "Not inlining: recursive call.\n");
2568 0 : continue;
2569 : }
2570 :
2571 148 : if (gimple_in_ssa_p (DECL_STRUCT_FUNCTION (node->decl))
2572 296 : != gimple_in_ssa_p (DECL_STRUCT_FUNCTION (callee->decl)))
2573 : {
2574 2 : if (dump_enabled_p ())
2575 2 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
2576 : "Not inlining: SSA form does not match.\n");
2577 2 : continue;
2578 : }
2579 :
2580 : /* Inline the edge and flatten the inline clone. Avoid
2581 : recursing through the original node if the node was cloned. */
2582 146 : if (dump_enabled_p ())
2583 6 : dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, e->call_stmt,
2584 : " Inlining %C into %C.\n",
2585 : callee, e->caller);
2586 146 : orig_callee = callee;
2587 146 : inline_call (e, true, NULL, NULL, false);
2588 146 : if (e->callee != orig_callee)
2589 106 : orig_callee->aux = (void *) node;
2590 146 : flatten_function (e->callee, early, false);
2591 146 : if (e->callee != orig_callee)
2592 106 : orig_callee->aux = NULL;
2593 : }
2594 :
2595 716 : node->aux = NULL;
2596 716 : cgraph_node *where = node->inlined_to ? node->inlined_to : node;
2597 716 : if (update && opt_for_fn (where->decl, optimize))
2598 209 : ipa_update_overall_fn_summary (where);
2599 716 : }
2600 :
2601 : /* Inline NODE to all callers. Worker for cgraph_for_node_and_aliases.
2602 : DATA points to number of calls originally found so we avoid infinite
2603 : recursion. */
2604 :
2605 : static bool
2606 31807 : inline_to_all_callers_1 (struct cgraph_node *node, void *data,
2607 : hash_set<cgraph_node *> *callers)
2608 : {
2609 31807 : int *num_calls = (int *)data;
2610 31807 : bool callee_removed = false;
2611 :
2612 66665 : while (node->callers && !node->inlined_to)
2613 : {
2614 36195 : struct cgraph_node *caller = node->callers->caller;
2615 :
2616 36195 : if (!can_inline_edge_p (node->callers, true)
2617 36195 : || !can_inline_edge_by_limits_p (node->callers, CAN_INLINE_REPORT)
2618 72389 : || node->callers->recursive_p ())
2619 : {
2620 1 : if (dump_file)
2621 0 : fprintf (dump_file, "Uninlinable call found; giving up.\n");
2622 1 : *num_calls = 0;
2623 1 : return false;
2624 : }
2625 :
2626 36194 : if (dump_file)
2627 : {
2628 2 : cgraph_node *ultimate = node->ultimate_alias_target ();
2629 2 : fprintf (dump_file,
2630 : "\nInlining %s size %i.\n",
2631 : ultimate->dump_name (),
2632 2 : ipa_size_summaries->get (ultimate)->size);
2633 2 : fprintf (dump_file,
2634 : " Called once from %s %i insns.\n",
2635 : node->callers->caller->dump_name (),
2636 2 : ipa_size_summaries->get (node->callers->caller)->size);
2637 : }
2638 :
2639 : /* Remember which callers we inlined to, delaying updating the
2640 : overall summary. */
2641 36194 : callers->add (node->callers->caller);
2642 36194 : inline_call (node->callers, true, NULL, NULL, false, &callee_removed);
2643 36194 : if (dump_file)
2644 2 : fprintf (dump_file,
2645 : " Inlined into %s which now has %i size\n",
2646 : caller->dump_name (),
2647 2 : ipa_size_summaries->get (caller)->size);
2648 36194 : if (!(*num_calls)--)
2649 : {
2650 0 : if (dump_file)
2651 0 : fprintf (dump_file, "New calls found; giving up.\n");
2652 0 : return callee_removed;
2653 : }
2654 36194 : if (callee_removed)
2655 : return true;
2656 : }
2657 : return false;
2658 : }
2659 :
2660 : /* Wrapper around inline_to_all_callers_1 doing delayed overall summary
2661 : update. */
2662 :
2663 : static bool
2664 31807 : inline_to_all_callers (struct cgraph_node *node, void *data)
2665 : {
2666 31807 : hash_set<cgraph_node *> callers;
2667 31807 : bool res = inline_to_all_callers_1 (node, data, &callers);
2668 : /* Perform the delayed update of the overall summary of all callers
2669 : processed. This avoids quadratic behavior in the cases where
2670 : we have a lot of calls to the same function. */
2671 62165 : for (hash_set<cgraph_node *>::iterator i = callers.begin ();
2672 62165 : i != callers.end (); ++i)
2673 30358 : ipa_update_overall_fn_summary ((*i)->inlined_to ? (*i)->inlined_to : *i);
2674 31807 : return res;
2675 31807 : }
2676 :
2677 : /* Output overall time estimate. */
2678 : static void
2679 356 : dump_overall_stats (void)
2680 : {
2681 356 : sreal sum_weighted = 0, sum = 0;
2682 356 : struct cgraph_node *node;
2683 :
2684 2288 : FOR_EACH_DEFINED_FUNCTION (node)
2685 1932 : if (!node->inlined_to
2686 1373 : && !node->alias)
2687 : {
2688 1265 : ipa_fn_summary *s = ipa_fn_summaries->get (node);
2689 1265 : if (s != NULL)
2690 : {
2691 1161 : sum += s->time;
2692 1161 : if (node->count.ipa ().initialized_p ())
2693 14 : sum_weighted += s->time * node->count.ipa ().to_gcov_type ();
2694 : }
2695 : }
2696 356 : fprintf (dump_file, "Overall time estimate: "
2697 : "%f weighted by profile: "
2698 : "%f\n", sum.to_double (), sum_weighted.to_double ());
2699 356 : }
2700 :
2701 : /* Output some useful stats about inlining. */
2702 :
2703 : static void
2704 178 : dump_inline_stats (void)
2705 : {
2706 178 : int64_t inlined_cnt = 0, inlined_indir_cnt = 0;
2707 178 : int64_t inlined_virt_cnt = 0, inlined_virt_indir_cnt = 0;
2708 178 : int64_t noninlined_cnt = 0, noninlined_indir_cnt = 0;
2709 178 : int64_t noninlined_virt_cnt = 0, noninlined_virt_indir_cnt = 0;
2710 178 : int64_t inlined_speculative = 0, inlined_speculative_ply = 0;
2711 178 : int64_t indirect_poly_cnt = 0, indirect_cnt = 0;
2712 178 : int64_t reason[CIF_N_REASONS][2];
2713 5874 : sreal reason_freq[CIF_N_REASONS];
2714 178 : int i;
2715 178 : struct cgraph_node *node;
2716 :
2717 178 : memset (reason, 0, sizeof (reason));
2718 5874 : for (i=0; i < CIF_N_REASONS; i++)
2719 5696 : reason_freq[i] = 0;
2720 1242 : FOR_EACH_DEFINED_FUNCTION (node)
2721 : {
2722 1064 : struct cgraph_edge *e;
2723 5947 : for (e = node->callees; e; e = e->next_callee)
2724 : {
2725 4883 : if (e->inline_failed)
2726 : {
2727 4330 : if (e->count.ipa ().initialized_p ())
2728 2611 : reason[(int) e->inline_failed][0] += e->count.ipa ().to_gcov_type ();
2729 4330 : reason_freq[(int) e->inline_failed] += e->sreal_frequency ();
2730 4330 : reason[(int) e->inline_failed][1] ++;
2731 4330 : if (DECL_VIRTUAL_P (e->callee->decl)
2732 4330 : && e->count.ipa ().initialized_p ())
2733 : {
2734 0 : if (e->indirect_inlining_edge)
2735 0 : noninlined_virt_indir_cnt += e->count.ipa ().to_gcov_type ();
2736 : else
2737 0 : noninlined_virt_cnt += e->count.ipa ().to_gcov_type ();
2738 : }
2739 4330 : else if (e->count.ipa ().initialized_p ())
2740 : {
2741 2611 : if (e->indirect_inlining_edge)
2742 0 : noninlined_indir_cnt += e->count.ipa ().to_gcov_type ();
2743 : else
2744 2611 : noninlined_cnt += e->count.ipa ().to_gcov_type ();
2745 : }
2746 : }
2747 553 : else if (e->count.ipa ().initialized_p ())
2748 : {
2749 0 : if (e->speculative)
2750 : {
2751 0 : if (DECL_VIRTUAL_P (e->callee->decl))
2752 0 : inlined_speculative_ply += e->count.ipa ().to_gcov_type ();
2753 : else
2754 0 : inlined_speculative += e->count.ipa ().to_gcov_type ();
2755 : }
2756 0 : else if (DECL_VIRTUAL_P (e->callee->decl))
2757 : {
2758 0 : if (e->indirect_inlining_edge)
2759 0 : inlined_virt_indir_cnt += e->count.ipa ().to_gcov_type ();
2760 : else
2761 0 : inlined_virt_cnt += e->count.ipa ().to_gcov_type ();
2762 : }
2763 : else
2764 : {
2765 0 : if (e->indirect_inlining_edge)
2766 0 : inlined_indir_cnt += e->count.ipa ().to_gcov_type ();
2767 : else
2768 0 : inlined_cnt += e->count.ipa ().to_gcov_type ();
2769 : }
2770 : }
2771 : }
2772 1163 : for (e = node->indirect_calls; e; e = e->next_callee)
2773 198 : if (is_a <cgraph_polymorphic_indirect_info *> (e->indirect_info)
2774 99 : & e->count.ipa ().initialized_p ())
2775 0 : indirect_poly_cnt += e->count.ipa ().to_gcov_type ();
2776 99 : else if (e->count.ipa ().initialized_p ())
2777 0 : indirect_cnt += e->count.ipa ().to_gcov_type ();
2778 : }
2779 178 : if (has_nonzero_ipa_profile)
2780 : {
2781 0 : fprintf (dump_file,
2782 : "Inlined %" PRId64 " + speculative "
2783 : "%" PRId64 " + speculative polymorphic "
2784 : "%" PRId64 " + previously indirect "
2785 : "%" PRId64 " + virtual "
2786 : "%" PRId64 " + virtual and previously indirect "
2787 : "%" PRId64 "\n" "Not inlined "
2788 : "%" PRId64 " + previously indirect "
2789 : "%" PRId64 " + virtual "
2790 : "%" PRId64 " + virtual and previously indirect "
2791 : "%" PRId64 " + still indirect "
2792 : "%" PRId64 " + still indirect polymorphic "
2793 : "%" PRId64 "\n", inlined_cnt,
2794 : inlined_speculative, inlined_speculative_ply,
2795 : inlined_indir_cnt, inlined_virt_cnt, inlined_virt_indir_cnt,
2796 : noninlined_cnt, noninlined_indir_cnt, noninlined_virt_cnt,
2797 : noninlined_virt_indir_cnt, indirect_cnt, indirect_poly_cnt);
2798 0 : fprintf (dump_file, "Removed speculations ");
2799 0 : spec_rem.dump (dump_file);
2800 0 : fprintf (dump_file, "\n");
2801 : }
2802 178 : dump_overall_stats ();
2803 178 : fprintf (dump_file, "\nWhy inlining failed?\n");
2804 6052 : for (i = 0; i < CIF_N_REASONS; i++)
2805 5696 : if (reason[i][1])
2806 146 : fprintf (dump_file, "%-50s: %8i calls, %8f freq, %" PRId64" count\n",
2807 : cgraph_inline_failed_string ((cgraph_inline_failed_t) i),
2808 : (int) reason[i][1], reason_freq[i].to_double (), reason[i][0]);
2809 178 : }
2810 :
2811 : /* Called when node is removed. */
2812 :
2813 : static void
2814 0 : flatten_remove_node_hook (struct cgraph_node *node, void *data)
2815 : {
2816 0 : if (lookup_attribute ("flatten", DECL_ATTRIBUTES (node->decl)) == NULL)
2817 : return;
2818 :
2819 0 : hash_set<struct cgraph_node *> *removed
2820 : = (hash_set<struct cgraph_node *> *) data;
2821 0 : removed->add (node);
2822 : }
2823 :
2824 : /* Decide on the inlining. We do so in the topological order to avoid
2825 : expenses on updating data structures. */
2826 :
2827 : static unsigned int
2828 239354 : ipa_inline (void)
2829 : {
2830 239354 : struct cgraph_node *node;
2831 239354 : int nnodes;
2832 239354 : struct cgraph_node **order;
2833 239354 : int i, j;
2834 239354 : int cold;
2835 239354 : bool remove_functions = false;
2836 :
2837 239354 : order = XCNEWVEC (struct cgraph_node *, symtab->cgraph_count);
2838 :
2839 239354 : if (dump_file)
2840 178 : ipa_dump_fn_summaries (dump_file);
2841 :
2842 239354 : nnodes = ipa_reverse_postorder (order);
2843 239354 : spec_rem = profile_count::zero ();
2844 :
2845 3997654 : FOR_EACH_FUNCTION (node)
2846 : {
2847 3758300 : node->aux = 0;
2848 :
2849 : /* Recompute the default reasons for inlining because they may have
2850 : changed during merging. */
2851 3758300 : if (in_lto_p)
2852 : {
2853 443154 : for (cgraph_edge *e = node->callees; e; e = e->next_callee)
2854 : {
2855 334298 : gcc_assert (e->inline_failed);
2856 334298 : initialize_inline_failed (e);
2857 : }
2858 110148 : for (cgraph_edge *e = node->indirect_calls; e; e = e->next_callee)
2859 1292 : initialize_inline_failed (e);
2860 : }
2861 : }
2862 :
2863 239354 : if (dump_file)
2864 178 : fprintf (dump_file, "\nFlattening functions:\n");
2865 :
2866 : /* First shrink order array, so that it only contains nodes with
2867 : flatten attribute. */
2868 3997654 : for (i = nnodes - 1, j = i; i >= 0; i--)
2869 : {
2870 3758300 : node = order[i];
2871 3758300 : if (node->definition
2872 : /* Do not try to flatten aliases. These may happen for example when
2873 : creating local aliases. */
2874 1952246 : && !node->alias
2875 5607749 : && lookup_attribute ("flatten",
2876 1849449 : DECL_ATTRIBUTES (node->decl)) != NULL)
2877 85 : order[j--] = order[i];
2878 : }
2879 :
2880 : /* After the above loop, order[j + 1] ... order[nnodes - 1] contain
2881 : nodes with flatten attribute. If there is more than one such
2882 : node, we need to register a node removal hook, as flatten_function
2883 : could remove other nodes with flatten attribute. See PR82801. */
2884 239354 : struct cgraph_node_hook_list *node_removal_hook_holder = NULL;
2885 239354 : hash_set<struct cgraph_node *> *flatten_removed_nodes = NULL;
2886 239354 : if (j < nnodes - 2)
2887 : {
2888 15 : flatten_removed_nodes = new hash_set<struct cgraph_node *>;
2889 15 : node_removal_hook_holder
2890 15 : = symtab->add_cgraph_removal_hook (&flatten_remove_node_hook,
2891 : flatten_removed_nodes);
2892 : }
2893 :
2894 : /* In the first pass handle functions to be flattened. Do this with
2895 : a priority so none of our later choices will make this impossible. */
2896 239439 : for (i = nnodes - 1; i > j; i--)
2897 : {
2898 85 : node = order[i];
2899 85 : if (flatten_removed_nodes
2900 85 : && flatten_removed_nodes->contains (node))
2901 0 : continue;
2902 :
2903 : /* Handle nodes to be flattened.
2904 : Ideally when processing callees we stop inlining at the
2905 : entry of cycles, possibly cloning that entry point and
2906 : try to flatten itself turning it into a self-recursive
2907 : function. */
2908 85 : if (dump_file)
2909 4 : fprintf (dump_file, "Flattening %s\n", node->dump_name ());
2910 85 : flatten_function (node, false, true);
2911 : }
2912 :
2913 239354 : if (j < nnodes - 2)
2914 : {
2915 15 : symtab->remove_cgraph_removal_hook (node_removal_hook_holder);
2916 30 : delete flatten_removed_nodes;
2917 : }
2918 239354 : free (order);
2919 :
2920 239354 : if (dump_file)
2921 178 : dump_overall_stats ();
2922 :
2923 239354 : inline_small_functions ();
2924 :
2925 239354 : gcc_assert (symtab->state == IPA_SSA);
2926 239354 : symtab->state = IPA_SSA_AFTER_INLINING;
2927 : /* Do first after-inlining removal. We want to remove all "stale" extern
2928 : inline functions and virtual functions so we really know what is called
2929 : once. */
2930 239354 : symtab->remove_unreachable_nodes (dump_file);
2931 :
2932 : /* Inline functions with a property that after inlining into all callers the
2933 : code size will shrink because the out-of-line copy is eliminated.
2934 : We do this regardless on the callee size as long as function growth limits
2935 : are met. */
2936 239354 : if (dump_file)
2937 178 : fprintf (dump_file,
2938 : "\nDeciding on functions to be inlined into all callers and "
2939 : "removing useless speculations:\n");
2940 :
2941 : /* Inlining one function called once has good chance of preventing
2942 : inlining other function into the same callee. Ideally we should
2943 : work in priority order, but probably inlining hot functions first
2944 : is good cut without the extra pain of maintaining the queue.
2945 :
2946 : ??? this is not really fitting the bill perfectly: inlining function
2947 : into callee often leads to better optimization of callee due to
2948 : increased context for optimization.
2949 : For example if main() function calls a function that outputs help
2950 : and then function that does the main optimization, we should inline
2951 : the second with priority even if both calls are cold by themselves.
2952 :
2953 : We probably want to implement new predicate replacing our use of
2954 : maybe_hot_edge interpreted as maybe_hot_edge || callee is known
2955 : to be hot. */
2956 718062 : for (cold = 0; cold <= 1; cold ++)
2957 : {
2958 6493409 : FOR_EACH_DEFINED_FUNCTION (node)
2959 : {
2960 6014701 : struct cgraph_edge *edge, *next;
2961 6014701 : bool update=false;
2962 :
2963 6014701 : if (!opt_for_fn (node->decl, optimize)
2964 6014701 : || !opt_for_fn (node->decl, flag_inline_functions_called_once))
2965 961930 : continue;
2966 :
2967 20373224 : for (edge = node->callees; edge; edge = next)
2968 : {
2969 15320453 : next = edge->next_callee;
2970 15320453 : if (edge->speculative && !speculation_useful_p (edge, false))
2971 : {
2972 2530 : if (edge->count.ipa ().initialized_p ())
2973 0 : spec_rem += edge->count.ipa ();
2974 2530 : cgraph_edge::resolve_speculation (edge);
2975 2530 : update = true;
2976 2530 : remove_functions = true;
2977 : }
2978 : }
2979 5052771 : if (update)
2980 : {
2981 152 : struct cgraph_node *where = node->inlined_to
2982 2063 : ? node->inlined_to : node;
2983 2063 : reset_edge_caches (where);
2984 2063 : ipa_update_overall_fn_summary (where);
2985 : }
2986 5052771 : if (want_inline_function_to_all_callers_p (node, cold))
2987 : {
2988 27856 : int num_calls = 0;
2989 27856 : node->call_for_symbol_and_aliases (sum_callers, &num_calls,
2990 : true);
2991 27856 : while (node->call_for_symbol_and_aliases
2992 29192 : (inline_to_all_callers, &num_calls, true))
2993 : ;
2994 27856 : remove_functions = true;
2995 : }
2996 : }
2997 : }
2998 :
2999 239354 : if (dump_enabled_p ())
3000 241 : dump_printf (MSG_NOTE,
3001 : "\nInlined %i calls, eliminated %i functions\n\n",
3002 : ncalls_inlined, nfunctions_inlined);
3003 239354 : if (dump_file)
3004 178 : dump_inline_stats ();
3005 :
3006 239354 : if (dump_file)
3007 178 : ipa_dump_fn_summaries (dump_file);
3008 239354 : return remove_functions ? TODO_remove_functions : 0;
3009 : }
3010 :
3011 : /* Inline always-inline function calls in NODE
3012 : (which itself is possibly inline). */
3013 :
3014 : static bool
3015 3608823 : inline_always_inline_functions (struct cgraph_node *node)
3016 : {
3017 3608823 : struct cgraph_edge *e;
3018 3608823 : bool inlined = false;
3019 :
3020 13955580 : for (e = node->callees; e; e = e->next_callee)
3021 : {
3022 10346757 : struct cgraph_node *callee = e->callee->ultimate_alias_target ();
3023 10346757 : gcc_checking_assert (!callee->aux || callee->aux == (void *)(size_t)1);
3024 10346757 : if (!DECL_DISREGARD_INLINE_LIMITS (callee->decl)
3025 : /* Watch for self-recursive cycles. */
3026 10346757 : || callee->aux)
3027 9723977 : continue;
3028 :
3029 622780 : if (e->recursive_p ())
3030 : {
3031 6 : if (dump_enabled_p ())
3032 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
3033 : " Not inlining recursive call to %C.\n",
3034 : e->callee);
3035 6 : e->inline_failed = CIF_RECURSIVE_INLINING;
3036 6 : continue;
3037 : }
3038 622774 : if (callee->definition
3039 622652 : && !ipa_fn_summaries->get (callee))
3040 221 : compute_fn_summary (callee, true);
3041 :
3042 622774 : if (!can_early_inline_edge_p (e))
3043 : {
3044 : /* Set inlined to true if the callee is marked "always_inline" but
3045 : is not inlinable. This will allow flagging an error later in
3046 : expand_call_inline in tree-inline.cc. */
3047 172 : if (lookup_attribute ("always_inline",
3048 172 : DECL_ATTRIBUTES (callee->decl)) != NULL)
3049 28 : inlined = true;
3050 172 : continue;
3051 : }
3052 :
3053 622602 : if (dump_enabled_p ())
3054 18 : dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, e->call_stmt,
3055 : " Inlining %C into %C (always_inline).\n",
3056 : e->callee, e->caller);
3057 622602 : inline_call (e, true, NULL, NULL, false);
3058 622602 : callee->aux = (void *)(size_t)1;
3059 : /* Inline recursively to handle the case where always_inline function was
3060 : not optimized yet since it is a part of a cycle in callgraph. */
3061 622602 : inline_always_inline_functions (e->callee);
3062 622602 : callee->aux = NULL;
3063 622602 : inlined = true;
3064 : }
3065 3608823 : return inlined;
3066 : }
3067 :
3068 : /* Decide on the inlining. We do so in the topological order to avoid
3069 : expenses on updating data structures. */
3070 :
3071 : static bool
3072 2499159 : early_inline_small_functions (struct cgraph_node *node)
3073 : {
3074 2499159 : struct cgraph_edge *e;
3075 2499159 : bool inlined = false;
3076 :
3077 10733348 : for (e = node->callees; e; e = e->next_callee)
3078 : {
3079 8234189 : struct cgraph_node *callee = e->callee->ultimate_alias_target ();
3080 :
3081 : /* We can encounter not-yet-analyzed function during
3082 : early inlining on callgraphs with strongly
3083 : connected components. */
3084 8234189 : ipa_fn_summary *s = ipa_fn_summaries->get (callee);
3085 8234189 : if (s == NULL || !s->inlinable || !e->inline_failed)
3086 4378155 : continue;
3087 :
3088 : /* Do not consider functions not declared inline. */
3089 3856034 : if (!DECL_DECLARED_INLINE_P (callee->decl)
3090 882923 : && !opt_for_fn (node->decl, flag_inline_small_functions)
3091 3909465 : && !opt_for_fn (node->decl, flag_inline_functions))
3092 53366 : continue;
3093 :
3094 3802668 : if (dump_enabled_p ())
3095 158 : dump_printf_loc (MSG_NOTE, e->call_stmt,
3096 : "Considering inline candidate %C.\n",
3097 : callee);
3098 :
3099 3802668 : if (!can_early_inline_edge_p (e))
3100 90975 : continue;
3101 :
3102 3711693 : if (e->recursive_p ())
3103 : {
3104 9510 : if (dump_enabled_p ())
3105 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
3106 : " Not inlining: recursive call.\n");
3107 9510 : continue;
3108 : }
3109 :
3110 3702183 : if (!want_early_inline_function_p (e))
3111 1115266 : continue;
3112 :
3113 2586917 : if (dump_enabled_p ())
3114 105 : dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, e->call_stmt,
3115 : " Inlining %C into %C.\n",
3116 : callee, e->caller);
3117 2586917 : inline_call (e, true, NULL, NULL, false);
3118 2586917 : inlined = true;
3119 : }
3120 :
3121 2499159 : if (inlined)
3122 874444 : ipa_update_overall_fn_summary (node);
3123 :
3124 2499159 : return inlined;
3125 : }
3126 :
3127 : /* With auto-fdo inline all functions that was inlined in the train run
3128 : and inlining seems useful. That is there are enough samples in the callee
3129 : function.
3130 :
3131 : Unlike early inlining, we inline recursively. Profile data is also used
3132 : to produce speculative calls which we then inline. In the case some
3133 : speculatin was introduced, set SPECULATIVE_CALLS. */
3134 :
3135 : static bool
3136 2500708 : inline_functions_by_afdo (struct cgraph_node *node, bool *speculative_calls)
3137 : {
3138 2500708 : if (!flag_auto_profile || !flag_auto_profile_inlining)
3139 : return false;
3140 0 : struct cgraph_edge *e;
3141 0 : bool inlined = false;
3142 :
3143 0 : *speculative_calls |= afdo_vpt_for_early_inline (node);
3144 :
3145 0 : cgraph_edge *next;
3146 0 : for (e = node->callees; e; e = next)
3147 : {
3148 0 : next = e->next_callee;
3149 :
3150 0 : if (!e->inline_failed)
3151 : {
3152 0 : inlined |= inline_functions_by_afdo (e->callee, speculative_calls);
3153 0 : continue;
3154 : }
3155 0 : if (!afdo_callsite_hot_enough_for_early_inline (e))
3156 : {
3157 : /* If we do not want to inline, remove the speculation. */
3158 0 : if (e->speculative)
3159 0 : cgraph_edge::resolve_speculation (e);
3160 0 : continue;
3161 : }
3162 :
3163 0 : struct cgraph_node *callee = e->callee->ultimate_alias_target ();
3164 0 : if (callee->definition
3165 0 : && !ipa_fn_summaries->get (callee))
3166 0 : compute_fn_summary (callee, true);
3167 :
3168 0 : if (!can_early_inline_edge_p (e))
3169 : {
3170 0 : if (dump_enabled_p ())
3171 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
3172 : "Not inlining %C -> %C using auto-profile, %s.",
3173 : e->caller, e->callee,
3174 : cgraph_inline_failed_string (e->inline_failed));
3175 : /* If we do not want to inline, remove the speculation. */
3176 0 : if (e->speculative)
3177 0 : cgraph_edge::resolve_speculation (e);
3178 0 : continue;
3179 : }
3180 : /* We can handle recursive inlining by first producing
3181 : inline clone. */
3182 0 : if (e->recursive_p ())
3183 : {
3184 0 : if (dump_enabled_p ())
3185 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
3186 : "Not inlining %C recursively"
3187 : " using auto-profile.\n",
3188 : e->callee);
3189 : /* If we do not want to inline, remove the speculation. */
3190 0 : if (e->speculative)
3191 0 : cgraph_edge::resolve_speculation (e);
3192 0 : continue;
3193 : }
3194 :
3195 0 : if (dump_enabled_p ())
3196 : {
3197 0 : if (e->caller->inlined_to)
3198 0 : dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, e->call_stmt,
3199 : "Inlining using auto-profile %C into %C "
3200 : "which is transitively inlined to %C.\n",
3201 : callee, e->caller, e->caller->inlined_to);
3202 : else
3203 0 : dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, e->call_stmt,
3204 : "Inlining using auto-profile %C into %C.\n",
3205 : callee, e->caller);
3206 : }
3207 0 : if (e->speculative)
3208 0 : remove_afdo_speculative_target (e);
3209 0 : inline_call (e, true, NULL, NULL, false);
3210 0 : inlined |= inline_functions_by_afdo (e->callee, speculative_calls);
3211 0 : inlined = true;
3212 : }
3213 :
3214 0 : if (inlined && !node->inlined_to)
3215 0 : ipa_update_overall_fn_summary (node);
3216 :
3217 : return inlined;
3218 : }
3219 :
3220 : unsigned int
3221 2986235 : early_inliner (function *fun)
3222 : {
3223 2986235 : struct cgraph_node *node = cgraph_node::get (current_function_decl);
3224 2986235 : struct cgraph_edge *edge;
3225 2986235 : unsigned int todo = 0;
3226 2986235 : int iterations = 0;
3227 2986235 : bool inlined = false;
3228 :
3229 2986235 : if (seen_error ())
3230 : return 0;
3231 :
3232 : /* Do nothing if datastructures for ipa-inliner are already computed. This
3233 : happens when some pass decides to construct new function and
3234 : cgraph_add_new_function calls lowering passes and early optimization on
3235 : it. This may confuse ourself when early inliner decide to inline call to
3236 : function clone, because function clones don't have parameter list in
3237 : ipa-prop matching their signature. */
3238 2986229 : if (ipa_node_params_sum)
3239 : return 0;
3240 :
3241 2986221 : if (flag_checking)
3242 2986191 : node->verify ();
3243 2986221 : node->remove_all_references ();
3244 :
3245 : /* Even when not optimizing or not inlining inline always-inline
3246 : functions. */
3247 2986221 : inlined = inline_always_inline_functions (node);
3248 :
3249 2986221 : if (!optimize
3250 2524504 : || flag_no_inline
3251 2500746 : || !flag_early_inlining)
3252 : ;
3253 2499197 : else if (lookup_attribute ("flatten",
3254 2499197 : DECL_ATTRIBUTES (node->decl)) != NULL)
3255 : {
3256 : /* When the function is marked to be flattened, recursively inline
3257 : all calls in it. */
3258 135 : if (dump_enabled_p ())
3259 0 : dump_printf (MSG_OPTIMIZED_LOCATIONS,
3260 : "Flattening %C\n", node);
3261 135 : flatten_function (node, true, true);
3262 135 : inlined = true;
3263 : }
3264 : else
3265 : {
3266 : /* If some always_inline functions was inlined, apply the changes.
3267 : This way we will not account always inline into growth limits and
3268 : moreover we will inline calls from always inlines that we skipped
3269 : previously because of conditional in can_early_inline_edge_p
3270 : which prevents some inlining to always_inline. */
3271 2499062 : if (inlined)
3272 : {
3273 321454 : timevar_push (TV_INTEGRATION);
3274 321454 : todo |= optimize_inline_calls (current_function_decl);
3275 : /* optimize_inline_calls call above might have introduced new
3276 : statements that don't have inline parameters computed. */
3277 1671593 : for (edge = node->callees; edge; edge = edge->next_callee)
3278 : {
3279 : /* We can encounter not-yet-analyzed function during
3280 : early inlining on callgraphs with strongly
3281 : connected components. */
3282 1350139 : ipa_call_summary *es = ipa_call_summaries->get_create (edge);
3283 1350139 : es->call_stmt_size
3284 1350139 : = estimate_num_insns (edge->call_stmt, &eni_size_weights);
3285 1350139 : es->call_stmt_time
3286 1350139 : = estimate_num_insns (edge->call_stmt, &eni_time_weights);
3287 : }
3288 321454 : ipa_update_overall_fn_summary (node);
3289 321454 : inlined = false;
3290 321454 : timevar_pop (TV_INTEGRATION);
3291 : }
3292 : /* We iterate incremental inlining to get trivial cases of indirect
3293 : inlining. */
3294 3373506 : while (iterations < opt_for_fn (node->decl,
3295 : param_early_inliner_max_iterations))
3296 : {
3297 2499159 : bool inlined = early_inline_small_functions (node);
3298 2499159 : bool speculative_calls = false;
3299 2499159 : inlined |= inline_functions_by_afdo (node, &speculative_calls);
3300 2499159 : if (!inlined)
3301 : break;
3302 874444 : timevar_push (TV_INTEGRATION);
3303 874444 : if (speculative_calls)
3304 : {
3305 0 : cgraph_edge *next;
3306 0 : for (cgraph_edge *e = node->callees; e; e = next)
3307 : {
3308 0 : next = e->next_callee;
3309 0 : cgraph_edge::redirect_call_stmt_to_callee (e);
3310 : }
3311 : }
3312 874444 : todo |= optimize_inline_calls (current_function_decl);
3313 :
3314 : /* Technically we ought to recompute inline parameters so the new
3315 : iteration of early inliner works as expected. We however have
3316 : values approximately right and thus we only need to update edge
3317 : info that might be cleared out for newly discovered edges. */
3318 3638516 : for (edge = node->callees; edge; edge = edge->next_callee)
3319 : {
3320 : /* We have no summary for new bound store calls yet. */
3321 2764072 : ipa_call_summary *es = ipa_call_summaries->get_create (edge);
3322 2764072 : es->call_stmt_size
3323 2764072 : = estimate_num_insns (edge->call_stmt, &eni_size_weights);
3324 2764072 : es->call_stmt_time
3325 2764072 : = estimate_num_insns (edge->call_stmt, &eni_time_weights);
3326 : }
3327 874444 : if (iterations < opt_for_fn (node->decl,
3328 874444 : param_early_inliner_max_iterations) - 1)
3329 97 : ipa_update_overall_fn_summary (node);
3330 874444 : timevar_pop (TV_INTEGRATION);
3331 874444 : iterations++;
3332 874444 : inlined = false;
3333 : }
3334 2499062 : if (dump_file)
3335 206 : fprintf (dump_file, "Iterations: %i\n", iterations);
3336 : }
3337 :
3338 : /* do AFDO inlining in case it was not done as part of early inlining. */
3339 2986221 : if (optimize
3340 2524504 : && !flag_no_inline
3341 2500746 : && !flag_early_inlining
3342 1549 : && flag_auto_profile_inlining)
3343 : {
3344 1549 : bool speculative_calls = false;
3345 1549 : inlined |= inline_functions_by_afdo (node, &speculative_calls);
3346 1549 : if (speculative_calls)
3347 : {
3348 0 : cgraph_edge *next;
3349 0 : for (cgraph_edge *e = node->callees; e; e = next)
3350 : {
3351 0 : next = e->next_callee;
3352 0 : cgraph_edge::redirect_call_stmt_to_callee (e);
3353 : }
3354 : }
3355 : }
3356 :
3357 2986221 : if (inlined)
3358 : {
3359 23683 : timevar_push (TV_INTEGRATION);
3360 23683 : todo |= optimize_inline_calls (current_function_decl);
3361 23683 : timevar_pop (TV_INTEGRATION);
3362 : }
3363 :
3364 2986221 : fun->always_inline_functions_inlined = true;
3365 :
3366 2986221 : return todo;
3367 : }
3368 :
3369 : /* Do inlining of small functions. Doing so early helps profiling and other
3370 : passes to be somewhat more effective and avoids some code duplication in
3371 : later real inlining pass for testcases with very many function calls. */
3372 :
3373 : namespace {
3374 :
3375 : const pass_data pass_data_early_inline =
3376 : {
3377 : GIMPLE_PASS, /* type */
3378 : "einline", /* name */
3379 : OPTGROUP_INLINE, /* optinfo_flags */
3380 : TV_EARLY_INLINING, /* tv_id */
3381 : PROP_ssa, /* properties_required */
3382 : 0, /* properties_provided */
3383 : 0, /* properties_destroyed */
3384 : 0, /* todo_flags_start */
3385 : 0, /* todo_flags_finish */
3386 : };
3387 :
3388 : class pass_early_inline : public gimple_opt_pass
3389 : {
3390 : public:
3391 297180 : pass_early_inline (gcc::context *ctxt)
3392 594360 : : gimple_opt_pass (pass_data_early_inline, ctxt)
3393 : {}
3394 :
3395 : /* opt_pass methods: */
3396 : unsigned int execute (function *) final override;
3397 :
3398 : }; // class pass_early_inline
3399 :
3400 : unsigned int
3401 2986235 : pass_early_inline::execute (function *fun)
3402 : {
3403 2986235 : return early_inliner (fun);
3404 : }
3405 :
3406 : } // anon namespace
3407 :
3408 : gimple_opt_pass *
3409 297180 : make_pass_early_inline (gcc::context *ctxt)
3410 : {
3411 297180 : return new pass_early_inline (ctxt);
3412 : }
3413 :
3414 : namespace {
3415 :
3416 : const pass_data pass_data_ipa_inline =
3417 : {
3418 : IPA_PASS, /* type */
3419 : "inline", /* name */
3420 : OPTGROUP_INLINE, /* optinfo_flags */
3421 : TV_IPA_INLINING, /* tv_id */
3422 : 0, /* properties_required */
3423 : 0, /* properties_provided */
3424 : 0, /* properties_destroyed */
3425 : 0, /* todo_flags_start */
3426 : ( TODO_dump_symtab ), /* todo_flags_finish */
3427 : };
3428 :
3429 : class pass_ipa_inline : public ipa_opt_pass_d
3430 : {
3431 : public:
3432 297180 : pass_ipa_inline (gcc::context *ctxt)
3433 : : ipa_opt_pass_d (pass_data_ipa_inline, ctxt,
3434 : NULL, /* generate_summary */
3435 : NULL, /* write_summary */
3436 : NULL, /* read_summary */
3437 : NULL, /* write_optimization_summary */
3438 : NULL, /* read_optimization_summary */
3439 : NULL, /* stmt_fixup */
3440 : 0, /* function_transform_todo_flags_start */
3441 : inline_transform, /* function_transform */
3442 297180 : NULL) /* variable_transform */
3443 297180 : {}
3444 :
3445 : /* opt_pass methods: */
3446 239354 : unsigned int execute (function *) final override { return ipa_inline (); }
3447 :
3448 : }; // class pass_ipa_inline
3449 :
3450 : } // anon namespace
3451 :
3452 : ipa_opt_pass_d *
3453 297180 : make_pass_ipa_inline (gcc::context *ctxt)
3454 : {
3455 297180 : return new pass_ipa_inline (ctxt);
3456 : }
|