]> git.ipfire.org Git - thirdparty/gcc.git/blob - gcc/ipa-inline.c
8e344001105407369900a23299cbb296a3832496
[thirdparty/gcc.git] / gcc / ipa-inline.c
1 /* Inlining decision heuristics.
2 Copyright (C) 2003-2019 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 can not 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 "params.h"
109 #include "profile.h"
110 #include "symbol-summary.h"
111 #include "tree-vrp.h"
112 #include "ipa-prop.h"
113 #include "ipa-fnsummary.h"
114 #include "ipa-inline.h"
115 #include "ipa-utils.h"
116 #include "sreal.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
124 typedef fibonacci_heap <sreal, cgraph_edge> edge_heap_t;
125 typedef fibonacci_node <sreal, cgraph_edge> edge_heap_node_t;
126
127 /* Statistics we collect about inlining algorithm. */
128 static int overall_size;
129 static profile_count max_count;
130 static profile_count spec_rem;
131
132 /* Return false when inlining edge E would lead to violating
133 limits on function unit growth or stack usage growth.
134
135 The relative function body growth limit is present generally
136 to avoid problems with non-linear behavior of the compiler.
137 To allow inlining huge functions into tiny wrapper, the limit
138 is always based on the bigger of the two functions considered.
139
140 For stack growth limits we always base the growth in stack usage
141 of the callers. We want to prevent applications from segfaulting
142 on stack overflow when functions with huge stack frames gets
143 inlined. */
144
145 static bool
146 caller_growth_limits (struct cgraph_edge *e)
147 {
148 struct cgraph_node *to = e->caller;
149 struct cgraph_node *what = e->callee->ultimate_alias_target ();
150 int newsize;
151 int limit = 0;
152 HOST_WIDE_INT stack_size_limit = 0, inlined_stack;
153 ipa_fn_summary *info, *what_info;
154 ipa_fn_summary *outer_info = ipa_fn_summaries->get (to);
155
156 /* Look for function e->caller is inlined to. While doing
157 so work out the largest function body on the way. As
158 described above, we want to base our function growth
159 limits based on that. Not on the self size of the
160 outer function, not on the self size of inline code
161 we immediately inline to. This is the most relaxed
162 interpretation of the rule "do not grow large functions
163 too much in order to prevent compiler from exploding". */
164 while (true)
165 {
166 info = ipa_fn_summaries->get (to);
167 if (limit < info->self_size)
168 limit = info->self_size;
169 if (stack_size_limit < info->estimated_self_stack_size)
170 stack_size_limit = info->estimated_self_stack_size;
171 if (to->global.inlined_to)
172 to = to->callers->caller;
173 else
174 break;
175 }
176
177 what_info = ipa_fn_summaries->get (what);
178
179 if (limit < what_info->self_size)
180 limit = what_info->self_size;
181
182 limit += limit * PARAM_VALUE (PARAM_LARGE_FUNCTION_GROWTH) / 100;
183
184 /* Check the size after inlining against the function limits. But allow
185 the function to shrink if it went over the limits by forced inlining. */
186 newsize = estimate_size_after_inlining (to, e);
187 if (newsize >= info->size
188 && newsize > PARAM_VALUE (PARAM_LARGE_FUNCTION_INSNS)
189 && newsize > limit)
190 {
191 e->inline_failed = CIF_LARGE_FUNCTION_GROWTH_LIMIT;
192 return false;
193 }
194
195 if (!what_info->estimated_stack_size)
196 return true;
197
198 /* FIXME: Stack size limit often prevents inlining in Fortran programs
199 due to large i/o datastructures used by the Fortran front-end.
200 We ought to ignore this limit when we know that the edge is executed
201 on every invocation of the caller (i.e. its call statement dominates
202 exit block). We do not track this information, yet. */
203 stack_size_limit += ((gcov_type)stack_size_limit
204 * PARAM_VALUE (PARAM_STACK_FRAME_GROWTH) / 100);
205
206 inlined_stack = (outer_info->stack_frame_offset
207 + outer_info->estimated_self_stack_size
208 + what_info->estimated_stack_size);
209 /* Check new stack consumption with stack consumption at the place
210 stack is used. */
211 if (inlined_stack > stack_size_limit
212 /* If function already has large stack usage from sibling
213 inline call, we can inline, too.
214 This bit overoptimistically assume that we are good at stack
215 packing. */
216 && inlined_stack > info->estimated_stack_size
217 && inlined_stack > PARAM_VALUE (PARAM_LARGE_STACK_FRAME))
218 {
219 e->inline_failed = CIF_LARGE_STACK_FRAME_GROWTH_LIMIT;
220 return false;
221 }
222 return true;
223 }
224
225 /* Dump info about why inlining has failed. */
226
227 static void
228 report_inline_failed_reason (struct cgraph_edge *e)
229 {
230 if (dump_enabled_p ())
231 {
232 dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
233 " not inlinable: %C -> %C, %s\n",
234 e->caller, e->callee,
235 cgraph_inline_failed_string (e->inline_failed));
236 if ((e->inline_failed == CIF_TARGET_OPTION_MISMATCH
237 || e->inline_failed == CIF_OPTIMIZATION_MISMATCH)
238 && e->caller->lto_file_data
239 && e->callee->ultimate_alias_target ()->lto_file_data)
240 {
241 dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
242 " LTO objects: %s, %s\n",
243 e->caller->lto_file_data->file_name,
244 e->callee->ultimate_alias_target ()->lto_file_data->file_name);
245 }
246 if (e->inline_failed == CIF_TARGET_OPTION_MISMATCH)
247 if (dump_file)
248 cl_target_option_print_diff
249 (dump_file, 2, target_opts_for_fn (e->caller->decl),
250 target_opts_for_fn (e->callee->ultimate_alias_target ()->decl));
251 if (e->inline_failed == CIF_OPTIMIZATION_MISMATCH)
252 if (dump_file)
253 cl_optimization_print_diff
254 (dump_file, 2, opts_for_fn (e->caller->decl),
255 opts_for_fn (e->callee->ultimate_alias_target ()->decl));
256 }
257 }
258
259 /* Decide whether sanitizer-related attributes allow inlining. */
260
261 static bool
262 sanitize_attrs_match_for_inline_p (const_tree caller, const_tree callee)
263 {
264 if (!caller || !callee)
265 return true;
266
267 return ((sanitize_flags_p (SANITIZE_ADDRESS, caller)
268 == sanitize_flags_p (SANITIZE_ADDRESS, callee))
269 && (sanitize_flags_p (SANITIZE_POINTER_COMPARE, caller)
270 == sanitize_flags_p (SANITIZE_POINTER_COMPARE, callee))
271 && (sanitize_flags_p (SANITIZE_POINTER_SUBTRACT, caller)
272 == sanitize_flags_p (SANITIZE_POINTER_SUBTRACT, callee)));
273 }
274
275 /* Used for flags where it is safe to inline when caller's value is
276 grater than callee's. */
277 #define check_maybe_up(flag) \
278 (opts_for_fn (caller->decl)->x_##flag \
279 != opts_for_fn (callee->decl)->x_##flag \
280 && (!always_inline \
281 || opts_for_fn (caller->decl)->x_##flag \
282 < opts_for_fn (callee->decl)->x_##flag))
283 /* Used for flags where it is safe to inline when caller's value is
284 smaller than callee's. */
285 #define check_maybe_down(flag) \
286 (opts_for_fn (caller->decl)->x_##flag \
287 != opts_for_fn (callee->decl)->x_##flag \
288 && (!always_inline \
289 || opts_for_fn (caller->decl)->x_##flag \
290 > opts_for_fn (callee->decl)->x_##flag))
291 /* Used for flags where exact match is needed for correctness. */
292 #define check_match(flag) \
293 (opts_for_fn (caller->decl)->x_##flag \
294 != opts_for_fn (callee->decl)->x_##flag)
295
296 /* Decide if we can inline the edge and possibly update
297 inline_failed reason.
298 We check whether inlining is possible at all and whether
299 caller growth limits allow doing so.
300
301 if REPORT is true, output reason to the dump file. */
302
303 static bool
304 can_inline_edge_p (struct cgraph_edge *e, bool report,
305 bool early = false)
306 {
307 gcc_checking_assert (e->inline_failed);
308
309 if (cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR)
310 {
311 if (report)
312 report_inline_failed_reason (e);
313 return false;
314 }
315
316 bool inlinable = true;
317 enum availability avail;
318 cgraph_node *caller = e->caller->global.inlined_to
319 ? e->caller->global.inlined_to : e->caller;
320 cgraph_node *callee = e->callee->ultimate_alias_target (&avail, caller);
321
322 if (!callee->definition)
323 {
324 e->inline_failed = CIF_BODY_NOT_AVAILABLE;
325 inlinable = false;
326 }
327 if (!early && (!opt_for_fn (callee->decl, optimize)
328 || !opt_for_fn (caller->decl, optimize)))
329 {
330 e->inline_failed = CIF_FUNCTION_NOT_OPTIMIZED;
331 inlinable = false;
332 }
333 else if (callee->calls_comdat_local)
334 {
335 e->inline_failed = CIF_USES_COMDAT_LOCAL;
336 inlinable = false;
337 }
338 else if (avail <= AVAIL_INTERPOSABLE)
339 {
340 e->inline_failed = CIF_OVERWRITABLE;
341 inlinable = false;
342 }
343 /* All edges with call_stmt_cannot_inline_p should have inline_failed
344 initialized to one of FINAL_ERROR reasons. */
345 else if (e->call_stmt_cannot_inline_p)
346 gcc_unreachable ();
347 /* Don't inline if the functions have different EH personalities. */
348 else if (DECL_FUNCTION_PERSONALITY (caller->decl)
349 && DECL_FUNCTION_PERSONALITY (callee->decl)
350 && (DECL_FUNCTION_PERSONALITY (caller->decl)
351 != DECL_FUNCTION_PERSONALITY (callee->decl)))
352 {
353 e->inline_failed = CIF_EH_PERSONALITY;
354 inlinable = false;
355 }
356 /* TM pure functions should not be inlined into non-TM_pure
357 functions. */
358 else if (is_tm_pure (callee->decl) && !is_tm_pure (caller->decl))
359 {
360 e->inline_failed = CIF_UNSPECIFIED;
361 inlinable = false;
362 }
363 /* Check compatibility of target optimization options. */
364 else if (!targetm.target_option.can_inline_p (caller->decl,
365 callee->decl))
366 {
367 e->inline_failed = CIF_TARGET_OPTION_MISMATCH;
368 inlinable = false;
369 }
370 else if (ipa_fn_summaries->get (callee) == NULL
371 || !ipa_fn_summaries->get (callee)->inlinable)
372 {
373 e->inline_failed = CIF_FUNCTION_NOT_INLINABLE;
374 inlinable = false;
375 }
376 /* Don't inline a function with mismatched sanitization attributes. */
377 else if (!sanitize_attrs_match_for_inline_p (caller->decl, callee->decl))
378 {
379 e->inline_failed = CIF_ATTRIBUTE_MISMATCH;
380 inlinable = false;
381 }
382 else if (callee->externally_visible
383 && flag_live_patching == LIVE_PATCHING_INLINE_ONLY_STATIC)
384 {
385 e->inline_failed = CIF_EXTERN_LIVE_ONLY_STATIC;
386 inlinable = false;
387 }
388 if (!inlinable && report)
389 report_inline_failed_reason (e);
390 return inlinable;
391 }
392
393 /* Decide if we can inline the edge and possibly update
394 inline_failed reason.
395 We check whether inlining is possible at all and whether
396 caller growth limits allow doing so.
397
398 if REPORT is true, output reason to the dump file.
399
400 if DISREGARD_LIMITS is true, ignore size limits. */
401
402 static bool
403 can_inline_edge_by_limits_p (struct cgraph_edge *e, bool report,
404 bool disregard_limits = false, bool early = false)
405 {
406 gcc_checking_assert (e->inline_failed);
407
408 if (cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR)
409 {
410 if (report)
411 report_inline_failed_reason (e);
412 return false;
413 }
414
415 bool inlinable = true;
416 enum availability avail;
417 cgraph_node *caller = e->caller->global.inlined_to
418 ? e->caller->global.inlined_to : e->caller;
419 cgraph_node *callee = e->callee->ultimate_alias_target (&avail, caller);
420 tree caller_tree = DECL_FUNCTION_SPECIFIC_OPTIMIZATION (caller->decl);
421 tree callee_tree
422 = callee ? DECL_FUNCTION_SPECIFIC_OPTIMIZATION (callee->decl) : NULL;
423 /* Check if caller growth allows the inlining. */
424 if (!DECL_DISREGARD_INLINE_LIMITS (callee->decl)
425 && !disregard_limits
426 && !lookup_attribute ("flatten",
427 DECL_ATTRIBUTES (caller->decl))
428 && !caller_growth_limits (e))
429 inlinable = false;
430 /* Don't inline a function with a higher optimization level than the
431 caller. FIXME: this is really just tip of iceberg of handling
432 optimization attribute. */
433 else if (caller_tree != callee_tree)
434 {
435 bool always_inline =
436 (DECL_DISREGARD_INLINE_LIMITS (callee->decl)
437 && lookup_attribute ("always_inline",
438 DECL_ATTRIBUTES (callee->decl)));
439 ipa_fn_summary *caller_info = ipa_fn_summaries->get (caller);
440 ipa_fn_summary *callee_info = ipa_fn_summaries->get (callee);
441
442 /* Until GCC 4.9 we did not check the semantics-altering flags
443 below and inlined across optimization boundaries.
444 Enabling checks below breaks several packages by refusing
445 to inline library always_inline functions. See PR65873.
446 Disable the check for early inlining for now until better solution
447 is found. */
448 if (always_inline && early)
449 ;
450 /* There are some options that change IL semantics which means
451 we cannot inline in these cases for correctness reason.
452 Not even for always_inline declared functions. */
453 else if (check_match (flag_wrapv)
454 || check_match (flag_trapv)
455 || check_match (flag_pcc_struct_return)
456 /* When caller or callee does FP math, be sure FP codegen flags
457 compatible. */
458 || ((caller_info->fp_expressions && callee_info->fp_expressions)
459 && (check_maybe_up (flag_rounding_math)
460 || check_maybe_up (flag_trapping_math)
461 || check_maybe_down (flag_unsafe_math_optimizations)
462 || check_maybe_down (flag_finite_math_only)
463 || check_maybe_up (flag_signaling_nans)
464 || check_maybe_down (flag_cx_limited_range)
465 || check_maybe_up (flag_signed_zeros)
466 || check_maybe_down (flag_associative_math)
467 || check_maybe_down (flag_reciprocal_math)
468 || check_maybe_down (flag_fp_int_builtin_inexact)
469 /* Strictly speaking only when the callee contains function
470 calls that may end up setting errno. */
471 || check_maybe_up (flag_errno_math)))
472 /* We do not want to make code compiled with exceptions to be
473 brought into a non-EH function unless we know that the callee
474 does not throw.
475 This is tracked by DECL_FUNCTION_PERSONALITY. */
476 || (check_maybe_up (flag_non_call_exceptions)
477 && DECL_FUNCTION_PERSONALITY (callee->decl))
478 || (check_maybe_up (flag_exceptions)
479 && DECL_FUNCTION_PERSONALITY (callee->decl))
480 /* When devirtualization is diabled for callee, it is not safe
481 to inline it as we possibly mangled the type info.
482 Allow early inlining of always inlines. */
483 || (!early && check_maybe_down (flag_devirtualize)))
484 {
485 e->inline_failed = CIF_OPTIMIZATION_MISMATCH;
486 inlinable = false;
487 }
488 /* gcc.dg/pr43564.c. Apply user-forced inline even at -O0. */
489 else if (always_inline)
490 ;
491 /* When user added an attribute to the callee honor it. */
492 else if (lookup_attribute ("optimize", DECL_ATTRIBUTES (callee->decl))
493 && opts_for_fn (caller->decl) != opts_for_fn (callee->decl))
494 {
495 e->inline_failed = CIF_OPTIMIZATION_MISMATCH;
496 inlinable = false;
497 }
498 /* If explicit optimize attribute are not used, the mismatch is caused
499 by different command line options used to build different units.
500 Do not care about COMDAT functions - those are intended to be
501 optimized with the optimization flags of module they are used in.
502 Also do not care about mixing up size/speed optimization when
503 DECL_DISREGARD_INLINE_LIMITS is set. */
504 else if ((callee->merged_comdat
505 && !lookup_attribute ("optimize",
506 DECL_ATTRIBUTES (caller->decl)))
507 || DECL_DISREGARD_INLINE_LIMITS (callee->decl))
508 ;
509 /* If mismatch is caused by merging two LTO units with different
510 optimizationflags we want to be bit nicer. However never inline
511 if one of functions is not optimized at all. */
512 else if (!opt_for_fn (callee->decl, optimize)
513 || !opt_for_fn (caller->decl, optimize))
514 {
515 e->inline_failed = CIF_OPTIMIZATION_MISMATCH;
516 inlinable = false;
517 }
518 /* If callee is optimized for size and caller is not, allow inlining if
519 code shrinks or we are in MAX_INLINE_INSNS_SINGLE limit and callee
520 is inline (and thus likely an unified comdat). This will allow caller
521 to run faster. */
522 else if (opt_for_fn (callee->decl, optimize_size)
523 > opt_for_fn (caller->decl, optimize_size))
524 {
525 int growth = estimate_edge_growth (e);
526 if (growth > PARAM_VALUE (PARAM_MAX_INLINE_INSNS_SIZE)
527 && (!DECL_DECLARED_INLINE_P (callee->decl)
528 && growth >= MAX (MAX_INLINE_INSNS_SINGLE,
529 MAX_INLINE_INSNS_AUTO)))
530 {
531 e->inline_failed = CIF_OPTIMIZATION_MISMATCH;
532 inlinable = false;
533 }
534 }
535 /* If callee is more aggressively optimized for performance than caller,
536 we generally want to inline only cheap (runtime wise) functions. */
537 else if (opt_for_fn (callee->decl, optimize_size)
538 < opt_for_fn (caller->decl, optimize_size)
539 || (opt_for_fn (callee->decl, optimize)
540 > opt_for_fn (caller->decl, optimize)))
541 {
542 if (estimate_edge_time (e)
543 >= 20 + ipa_call_summaries->get (e)->call_stmt_time)
544 {
545 e->inline_failed = CIF_OPTIMIZATION_MISMATCH;
546 inlinable = false;
547 }
548 }
549
550 }
551
552 if (!inlinable && report)
553 report_inline_failed_reason (e);
554 return inlinable;
555 }
556
557
558 /* Return true if the edge E is inlinable during early inlining. */
559
560 static bool
561 can_early_inline_edge_p (struct cgraph_edge *e)
562 {
563 struct cgraph_node *callee = e->callee->ultimate_alias_target ();
564 /* Early inliner might get called at WPA stage when IPA pass adds new
565 function. In this case we can not really do any of early inlining
566 because function bodies are missing. */
567 if (cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR)
568 return false;
569 if (!gimple_has_body_p (callee->decl))
570 {
571 e->inline_failed = CIF_BODY_NOT_AVAILABLE;
572 return false;
573 }
574 /* In early inliner some of callees may not be in SSA form yet
575 (i.e. the callgraph is cyclic and we did not process
576 the callee by early inliner, yet). We don't have CIF code for this
577 case; later we will re-do the decision in the real inliner. */
578 if (!gimple_in_ssa_p (DECL_STRUCT_FUNCTION (e->caller->decl))
579 || !gimple_in_ssa_p (DECL_STRUCT_FUNCTION (callee->decl)))
580 {
581 if (dump_enabled_p ())
582 dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
583 " edge not inlinable: not in SSA form\n");
584 return false;
585 }
586 if (!can_inline_edge_p (e, true, true)
587 || !can_inline_edge_by_limits_p (e, true, false, true))
588 return false;
589 return true;
590 }
591
592
593 /* Return number of calls in N. Ignore cheap builtins. */
594
595 static int
596 num_calls (struct cgraph_node *n)
597 {
598 struct cgraph_edge *e;
599 int num = 0;
600
601 for (e = n->callees; e; e = e->next_callee)
602 if (!is_inexpensive_builtin (e->callee->decl))
603 num++;
604 return num;
605 }
606
607
608 /* Return true if we are interested in inlining small function. */
609
610 static bool
611 want_early_inline_function_p (struct cgraph_edge *e)
612 {
613 bool want_inline = true;
614 struct cgraph_node *callee = e->callee->ultimate_alias_target ();
615
616 if (DECL_DISREGARD_INLINE_LIMITS (callee->decl))
617 ;
618 /* For AutoFDO, we need to make sure that before profile summary, all
619 hot paths' IR look exactly the same as profiled binary. As a result,
620 in einliner, we will disregard size limit and inline those callsites
621 that are:
622 * inlined in the profiled binary, and
623 * the cloned callee has enough samples to be considered "hot". */
624 else if (flag_auto_profile && afdo_callsite_hot_enough_for_early_inline (e))
625 ;
626 else if (!DECL_DECLARED_INLINE_P (callee->decl)
627 && !opt_for_fn (e->caller->decl, flag_inline_small_functions))
628 {
629 e->inline_failed = CIF_FUNCTION_NOT_INLINE_CANDIDATE;
630 report_inline_failed_reason (e);
631 want_inline = false;
632 }
633 else
634 {
635 int growth = estimate_edge_growth (e);
636 int n;
637
638 if (growth <= PARAM_VALUE (PARAM_MAX_INLINE_INSNS_SIZE))
639 ;
640 else if (!e->maybe_hot_p ())
641 {
642 if (dump_enabled_p ())
643 dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
644 " will not early inline: %C->%C, "
645 "call is cold and code would grow by %i\n",
646 e->caller, callee,
647 growth);
648 want_inline = false;
649 }
650 else if (growth > PARAM_VALUE (PARAM_EARLY_INLINING_INSNS))
651 {
652 if (dump_enabled_p ())
653 dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
654 " will not early inline: %C->%C, "
655 "growth %i exceeds --param early-inlining-insns\n",
656 e->caller, callee,
657 growth);
658 want_inline = false;
659 }
660 else if ((n = num_calls (callee)) != 0
661 && growth * (n + 1) > PARAM_VALUE (PARAM_EARLY_INLINING_INSNS))
662 {
663 if (dump_enabled_p ())
664 dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
665 " will not early inline: %C->%C, "
666 "growth %i exceeds --param early-inlining-insns "
667 "divided by number of calls\n",
668 e->caller, callee,
669 growth);
670 want_inline = false;
671 }
672 }
673 return want_inline;
674 }
675
676 /* Compute time of the edge->caller + edge->callee execution when inlining
677 does not happen. */
678
679 inline sreal
680 compute_uninlined_call_time (struct cgraph_edge *edge,
681 sreal uninlined_call_time)
682 {
683 cgraph_node *caller = (edge->caller->global.inlined_to
684 ? edge->caller->global.inlined_to
685 : edge->caller);
686
687 sreal freq = edge->sreal_frequency ();
688 if (freq > 0)
689 uninlined_call_time *= freq;
690 else
691 uninlined_call_time = uninlined_call_time >> 11;
692
693 sreal caller_time = ipa_fn_summaries->get (caller)->time;
694 return uninlined_call_time + caller_time;
695 }
696
697 /* Same as compute_uinlined_call_time but compute time when inlining
698 does happen. */
699
700 inline sreal
701 compute_inlined_call_time (struct cgraph_edge *edge,
702 sreal time)
703 {
704 cgraph_node *caller = (edge->caller->global.inlined_to
705 ? edge->caller->global.inlined_to
706 : edge->caller);
707 sreal caller_time = ipa_fn_summaries->get (caller)->time;
708
709 sreal freq = edge->sreal_frequency ();
710 if (freq > 0)
711 time *= freq;
712 else
713 time = time >> 11;
714
715 /* This calculation should match one in ipa-inline-analysis.c
716 (estimate_edge_size_and_time). */
717 time -= (sreal)ipa_call_summaries->get (edge)->call_stmt_time * freq;
718 time += caller_time;
719 if (time <= 0)
720 time = ((sreal) 1) >> 8;
721 gcc_checking_assert (time >= 0);
722 return time;
723 }
724
725 /* Return true if the speedup for inlining E is bigger than
726 PARAM_MAX_INLINE_MIN_SPEEDUP. */
727
728 static bool
729 big_speedup_p (struct cgraph_edge *e)
730 {
731 sreal unspec_time;
732 sreal spec_time = estimate_edge_time (e, &unspec_time);
733 sreal time = compute_uninlined_call_time (e, unspec_time);
734 sreal inlined_time = compute_inlined_call_time (e, spec_time);
735
736 if ((time - inlined_time) * 100
737 > (sreal) (time * PARAM_VALUE (PARAM_INLINE_MIN_SPEEDUP)))
738 return true;
739 return false;
740 }
741
742 /* Return true if we are interested in inlining small function.
743 When REPORT is true, report reason to dump file. */
744
745 static bool
746 want_inline_small_function_p (struct cgraph_edge *e, bool report)
747 {
748 bool want_inline = true;
749 struct cgraph_node *callee = e->callee->ultimate_alias_target ();
750
751 /* Allow this function to be called before can_inline_edge_p,
752 since it's usually cheaper. */
753 if (cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR)
754 want_inline = false;
755 else if (DECL_DISREGARD_INLINE_LIMITS (callee->decl))
756 ;
757 else if (!DECL_DECLARED_INLINE_P (callee->decl)
758 && !opt_for_fn (e->caller->decl, flag_inline_small_functions))
759 {
760 e->inline_failed = CIF_FUNCTION_NOT_INLINE_CANDIDATE;
761 want_inline = false;
762 }
763 /* Do fast and conservative check if the function can be good
764 inline candidate. At the moment we allow inline hints to
765 promote non-inline functions to inline and we increase
766 MAX_INLINE_INSNS_SINGLE 16-fold for inline functions. */
767 else if ((!DECL_DECLARED_INLINE_P (callee->decl)
768 && (!e->count.ipa ().initialized_p () || !e->maybe_hot_p ()))
769 && ipa_fn_summaries->get (callee)->min_size
770 - ipa_call_summaries->get (e)->call_stmt_size
771 > MAX (MAX_INLINE_INSNS_SINGLE, MAX_INLINE_INSNS_AUTO))
772 {
773 e->inline_failed = CIF_MAX_INLINE_INSNS_AUTO_LIMIT;
774 want_inline = false;
775 }
776 else if ((DECL_DECLARED_INLINE_P (callee->decl)
777 || e->count.ipa ().nonzero_p ())
778 && ipa_fn_summaries->get (callee)->min_size
779 - ipa_call_summaries->get (e)->call_stmt_size
780 > 16 * MAX_INLINE_INSNS_SINGLE)
781 {
782 e->inline_failed = (DECL_DECLARED_INLINE_P (callee->decl)
783 ? CIF_MAX_INLINE_INSNS_SINGLE_LIMIT
784 : CIF_MAX_INLINE_INSNS_AUTO_LIMIT);
785 want_inline = false;
786 }
787 else
788 {
789 int growth = estimate_edge_growth (e);
790 ipa_hints hints = estimate_edge_hints (e);
791 int big_speedup = -1; /* compute this lazily */
792
793 if (growth <= PARAM_VALUE (PARAM_MAX_INLINE_INSNS_SIZE))
794 ;
795 /* Apply MAX_INLINE_INSNS_SINGLE limit. Do not do so when
796 hints suggests that inlining given function is very profitable. */
797 else if (DECL_DECLARED_INLINE_P (callee->decl)
798 && growth >= MAX_INLINE_INSNS_SINGLE
799 && (growth >= MAX_INLINE_INSNS_SINGLE * 16
800 || (!(hints & (INLINE_HINT_indirect_call
801 | INLINE_HINT_known_hot
802 | INLINE_HINT_loop_iterations
803 | INLINE_HINT_array_index
804 | INLINE_HINT_loop_stride))
805 && !(big_speedup = big_speedup_p (e)))))
806 {
807 e->inline_failed = CIF_MAX_INLINE_INSNS_SINGLE_LIMIT;
808 want_inline = false;
809 }
810 else if (!DECL_DECLARED_INLINE_P (callee->decl)
811 && !opt_for_fn (e->caller->decl, flag_inline_functions)
812 && growth >= PARAM_VALUE (PARAM_MAX_INLINE_INSNS_SMALL))
813 {
814 /* growth_likely_positive is expensive, always test it last. */
815 if (growth >= MAX_INLINE_INSNS_SINGLE
816 || growth_likely_positive (callee, growth))
817 {
818 e->inline_failed = CIF_NOT_DECLARED_INLINED;
819 want_inline = false;
820 }
821 }
822 /* Apply MAX_INLINE_INSNS_AUTO limit for functions not declared inline
823 Upgrade it to MAX_INLINE_INSNS_SINGLE when hints suggests that
824 inlining given function is very profitable. */
825 else if (!DECL_DECLARED_INLINE_P (callee->decl)
826 && !(hints & INLINE_HINT_known_hot)
827 && growth >= ((hints & (INLINE_HINT_indirect_call
828 | INLINE_HINT_loop_iterations
829 | INLINE_HINT_array_index
830 | INLINE_HINT_loop_stride))
831 ? MAX (MAX_INLINE_INSNS_AUTO,
832 MAX_INLINE_INSNS_SINGLE)
833 : MAX_INLINE_INSNS_AUTO)
834 && !(big_speedup == -1 ? big_speedup_p (e) : big_speedup))
835 {
836 /* growth_likely_positive is expensive, always test it last. */
837 if (growth >= MAX_INLINE_INSNS_SINGLE
838 || growth_likely_positive (callee, growth))
839 {
840 e->inline_failed = CIF_MAX_INLINE_INSNS_AUTO_LIMIT;
841 want_inline = false;
842 }
843 }
844 /* If call is cold, do not inline when function body would grow. */
845 else if (!e->maybe_hot_p ()
846 && (growth >= MAX_INLINE_INSNS_SINGLE
847 || growth_likely_positive (callee, growth)))
848 {
849 e->inline_failed = CIF_UNLIKELY_CALL;
850 want_inline = false;
851 }
852 }
853 if (!want_inline && report)
854 report_inline_failed_reason (e);
855 return want_inline;
856 }
857
858 /* EDGE is self recursive edge.
859 We hand two cases - when function A is inlining into itself
860 or when function A is being inlined into another inliner copy of function
861 A within function B.
862
863 In first case OUTER_NODE points to the toplevel copy of A, while
864 in the second case OUTER_NODE points to the outermost copy of A in B.
865
866 In both cases we want to be extra selective since
867 inlining the call will just introduce new recursive calls to appear. */
868
869 static bool
870 want_inline_self_recursive_call_p (struct cgraph_edge *edge,
871 struct cgraph_node *outer_node,
872 bool peeling,
873 int depth)
874 {
875 char const *reason = NULL;
876 bool want_inline = true;
877 sreal caller_freq = 1;
878 int max_depth = PARAM_VALUE (PARAM_MAX_INLINE_RECURSIVE_DEPTH_AUTO);
879
880 if (DECL_DECLARED_INLINE_P (edge->caller->decl))
881 max_depth = PARAM_VALUE (PARAM_MAX_INLINE_RECURSIVE_DEPTH);
882
883 if (!edge->maybe_hot_p ())
884 {
885 reason = "recursive call is cold";
886 want_inline = false;
887 }
888 else if (depth > max_depth)
889 {
890 reason = "--param max-inline-recursive-depth exceeded.";
891 want_inline = false;
892 }
893 else if (outer_node->global.inlined_to
894 && (caller_freq = outer_node->callers->sreal_frequency ()) == 0)
895 {
896 reason = "caller frequency is 0";
897 want_inline = false;
898 }
899
900 if (!want_inline)
901 ;
902 /* Inlining of self recursive function into copy of itself within other
903 function is transformation similar to loop peeling.
904
905 Peeling is profitable if we can inline enough copies to make probability
906 of actual call to the self recursive function very small. Be sure that
907 the probability of recursion is small.
908
909 We ensure that the frequency of recursing is at most 1 - (1/max_depth).
910 This way the expected number of recursion is at most max_depth. */
911 else if (peeling)
912 {
913 sreal max_prob = (sreal)1 - ((sreal)1 / (sreal)max_depth);
914 int i;
915 for (i = 1; i < depth; i++)
916 max_prob = max_prob * max_prob;
917 if (edge->sreal_frequency () >= max_prob * caller_freq)
918 {
919 reason = "frequency of recursive call is too large";
920 want_inline = false;
921 }
922 }
923 /* Recursive inlining, i.e. equivalent of unrolling, is profitable if
924 recursion depth is large. We reduce function call overhead and increase
925 chances that things fit in hardware return predictor.
926
927 Recursive inlining might however increase cost of stack frame setup
928 actually slowing down functions whose recursion tree is wide rather than
929 deep.
930
931 Deciding reliably on when to do recursive inlining without profile feedback
932 is tricky. For now we disable recursive inlining when probability of self
933 recursion is low.
934
935 Recursive inlining of self recursive call within loop also results in
936 large loop depths that generally optimize badly. We may want to throttle
937 down inlining in those cases. In particular this seems to happen in one
938 of libstdc++ rb tree methods. */
939 else
940 {
941 if (edge->sreal_frequency () * 100
942 <= caller_freq
943 * PARAM_VALUE (PARAM_MIN_INLINE_RECURSIVE_PROBABILITY))
944 {
945 reason = "frequency of recursive call is too small";
946 want_inline = false;
947 }
948 }
949 if (!want_inline && dump_enabled_p ())
950 dump_printf_loc (MSG_MISSED_OPTIMIZATION, edge->call_stmt,
951 " not inlining recursively: %s\n", reason);
952 return want_inline;
953 }
954
955 /* Return true when NODE has uninlinable caller;
956 set HAS_HOT_CALL if it has hot call.
957 Worker for cgraph_for_node_and_aliases. */
958
959 static bool
960 check_callers (struct cgraph_node *node, void *has_hot_call)
961 {
962 struct cgraph_edge *e;
963 for (e = node->callers; e; e = e->next_caller)
964 {
965 if (!opt_for_fn (e->caller->decl, flag_inline_functions_called_once)
966 || !opt_for_fn (e->caller->decl, optimize))
967 return true;
968 if (!can_inline_edge_p (e, true))
969 return true;
970 if (e->recursive_p ())
971 return true;
972 if (!can_inline_edge_by_limits_p (e, true))
973 return true;
974 if (!(*(bool *)has_hot_call) && e->maybe_hot_p ())
975 *(bool *)has_hot_call = true;
976 }
977 return false;
978 }
979
980 /* If NODE has a caller, return true. */
981
982 static bool
983 has_caller_p (struct cgraph_node *node, void *data ATTRIBUTE_UNUSED)
984 {
985 if (node->callers)
986 return true;
987 return false;
988 }
989
990 /* Decide if inlining NODE would reduce unit size by eliminating
991 the offline copy of function.
992 When COLD is true the cold calls are considered, too. */
993
994 static bool
995 want_inline_function_to_all_callers_p (struct cgraph_node *node, bool cold)
996 {
997 bool has_hot_call = false;
998
999 /* Aliases gets inlined along with the function they alias. */
1000 if (node->alias)
1001 return false;
1002 /* Already inlined? */
1003 if (node->global.inlined_to)
1004 return false;
1005 /* Does it have callers? */
1006 if (!node->call_for_symbol_and_aliases (has_caller_p, NULL, true))
1007 return false;
1008 /* Inlining into all callers would increase size? */
1009 if (estimate_growth (node) > 0)
1010 return false;
1011 /* All inlines must be possible. */
1012 if (node->call_for_symbol_and_aliases (check_callers, &has_hot_call,
1013 true))
1014 return false;
1015 if (!cold && !has_hot_call)
1016 return false;
1017 return true;
1018 }
1019
1020 /* A cost model driving the inlining heuristics in a way so the edges with
1021 smallest badness are inlined first. After each inlining is performed
1022 the costs of all caller edges of nodes affected are recomputed so the
1023 metrics may accurately depend on values such as number of inlinable callers
1024 of the function or function body size. */
1025
1026 static sreal
1027 edge_badness (struct cgraph_edge *edge, bool dump)
1028 {
1029 sreal badness;
1030 int growth;
1031 sreal edge_time, unspec_edge_time;
1032 struct cgraph_node *callee = edge->callee->ultimate_alias_target ();
1033 struct ipa_fn_summary *callee_info = ipa_fn_summaries->get (callee);
1034 ipa_hints hints;
1035 cgraph_node *caller = (edge->caller->global.inlined_to
1036 ? edge->caller->global.inlined_to
1037 : edge->caller);
1038
1039 growth = estimate_edge_growth (edge);
1040 edge_time = estimate_edge_time (edge, &unspec_edge_time);
1041 hints = estimate_edge_hints (edge);
1042 gcc_checking_assert (edge_time >= 0);
1043 /* Check that inlined time is better, but tolerate some roundoff issues.
1044 FIXME: When callee profile drops to 0 we account calls more. This
1045 should be fixed by never doing that. */
1046 gcc_checking_assert ((edge_time * 100
1047 - callee_info->time * 101).to_int () <= 0
1048 || callee->count.ipa ().initialized_p ());
1049 gcc_checking_assert (growth <= callee_info->size);
1050
1051 if (dump)
1052 {
1053 fprintf (dump_file, " Badness calculation for %s -> %s\n",
1054 edge->caller->dump_name (),
1055 edge->callee->dump_name ());
1056 fprintf (dump_file, " size growth %i, time %f unspec %f ",
1057 growth,
1058 edge_time.to_double (),
1059 unspec_edge_time.to_double ());
1060 ipa_dump_hints (dump_file, hints);
1061 if (big_speedup_p (edge))
1062 fprintf (dump_file, " big_speedup");
1063 fprintf (dump_file, "\n");
1064 }
1065
1066 /* Always prefer inlining saving code size. */
1067 if (growth <= 0)
1068 {
1069 badness = (sreal) (-SREAL_MIN_SIG + growth) << (SREAL_MAX_EXP / 256);
1070 if (dump)
1071 fprintf (dump_file, " %f: Growth %d <= 0\n", badness.to_double (),
1072 growth);
1073 }
1074 /* Inlining into EXTERNAL functions is not going to change anything unless
1075 they are themselves inlined. */
1076 else if (DECL_EXTERNAL (caller->decl))
1077 {
1078 if (dump)
1079 fprintf (dump_file, " max: function is external\n");
1080 return sreal::max ();
1081 }
1082 /* When profile is available. Compute badness as:
1083
1084 time_saved * caller_count
1085 goodness = -------------------------------------------------
1086 growth_of_caller * overall_growth * combined_size
1087
1088 badness = - goodness
1089
1090 Again use negative value to make calls with profile appear hotter
1091 then calls without.
1092 */
1093 else if (opt_for_fn (caller->decl, flag_guess_branch_prob)
1094 || caller->count.ipa ().nonzero_p ())
1095 {
1096 sreal numerator, denominator;
1097 int overall_growth;
1098 sreal inlined_time = compute_inlined_call_time (edge, edge_time);
1099
1100 numerator = (compute_uninlined_call_time (edge, unspec_edge_time)
1101 - inlined_time);
1102 if (numerator <= 0)
1103 numerator = ((sreal) 1 >> 8);
1104 if (caller->count.ipa ().nonzero_p ())
1105 numerator *= caller->count.ipa ().to_gcov_type ();
1106 else if (caller->count.ipa ().initialized_p ())
1107 numerator = numerator >> 11;
1108 denominator = growth;
1109
1110 overall_growth = callee_info->growth;
1111
1112 /* Look for inliner wrappers of the form:
1113
1114 inline_caller ()
1115 {
1116 do_fast_job...
1117 if (need_more_work)
1118 noninline_callee ();
1119 }
1120 Withhout panilizing this case, we usually inline noninline_callee
1121 into the inline_caller because overall_growth is small preventing
1122 further inlining of inline_caller.
1123
1124 Penalize only callgraph edges to functions with small overall
1125 growth ...
1126 */
1127 if (growth > overall_growth
1128 /* ... and having only one caller which is not inlined ... */
1129 && callee_info->single_caller
1130 && !edge->caller->global.inlined_to
1131 /* ... and edges executed only conditionally ... */
1132 && edge->sreal_frequency () < 1
1133 /* ... consider case where callee is not inline but caller is ... */
1134 && ((!DECL_DECLARED_INLINE_P (edge->callee->decl)
1135 && DECL_DECLARED_INLINE_P (caller->decl))
1136 /* ... or when early optimizers decided to split and edge
1137 frequency still indicates splitting is a win ... */
1138 || (callee->split_part && !caller->split_part
1139 && edge->sreal_frequency () * 100
1140 < PARAM_VALUE
1141 (PARAM_PARTIAL_INLINING_ENTRY_PROBABILITY)
1142 /* ... and do not overwrite user specified hints. */
1143 && (!DECL_DECLARED_INLINE_P (edge->callee->decl)
1144 || DECL_DECLARED_INLINE_P (caller->decl)))))
1145 {
1146 ipa_fn_summary *caller_info = ipa_fn_summaries->get (caller);
1147 int caller_growth = caller_info->growth;
1148
1149 /* Only apply the penalty when caller looks like inline candidate,
1150 and it is not called once and. */
1151 if (!caller_info->single_caller && overall_growth < caller_growth
1152 && caller_info->inlinable
1153 && caller_info->size
1154 < (DECL_DECLARED_INLINE_P (caller->decl)
1155 ? MAX_INLINE_INSNS_SINGLE : MAX_INLINE_INSNS_AUTO))
1156 {
1157 if (dump)
1158 fprintf (dump_file,
1159 " Wrapper penalty. Increasing growth %i to %i\n",
1160 overall_growth, caller_growth);
1161 overall_growth = caller_growth;
1162 }
1163 }
1164 if (overall_growth > 0)
1165 {
1166 /* Strongly preffer functions with few callers that can be inlined
1167 fully. The square root here leads to smaller binaries at average.
1168 Watch however for extreme cases and return to linear function
1169 when growth is large. */
1170 if (overall_growth < 256)
1171 overall_growth *= overall_growth;
1172 else
1173 overall_growth += 256 * 256 - 256;
1174 denominator *= overall_growth;
1175 }
1176 denominator *= ipa_fn_summaries->get (caller)->self_size + growth;
1177
1178 badness = - numerator / denominator;
1179
1180 if (dump)
1181 {
1182 fprintf (dump_file,
1183 " %f: guessed profile. frequency %f, count %" PRId64
1184 " caller count %" PRId64
1185 " time w/o inlining %f, time with inlining %f"
1186 " overall growth %i (current) %i (original)"
1187 " %i (compensated)\n",
1188 badness.to_double (),
1189 edge->sreal_frequency ().to_double (),
1190 edge->count.ipa ().initialized_p () ? edge->count.ipa ().to_gcov_type () : -1,
1191 caller->count.ipa ().initialized_p () ? caller->count.ipa ().to_gcov_type () : -1,
1192 compute_uninlined_call_time (edge,
1193 unspec_edge_time).to_double (),
1194 inlined_time.to_double (),
1195 estimate_growth (callee),
1196 callee_info->growth, overall_growth);
1197 }
1198 }
1199 /* When function local profile is not available or it does not give
1200 useful information (ie frequency is zero), base the cost on
1201 loop nest and overall size growth, so we optimize for overall number
1202 of functions fully inlined in program. */
1203 else
1204 {
1205 int nest = MIN (ipa_call_summaries->get (edge)->loop_depth, 8);
1206 badness = growth;
1207
1208 /* Decrease badness if call is nested. */
1209 if (badness > 0)
1210 badness = badness >> nest;
1211 else
1212 badness = badness << nest;
1213 if (dump)
1214 fprintf (dump_file, " %f: no profile. nest %i\n",
1215 badness.to_double (), nest);
1216 }
1217 gcc_checking_assert (badness != 0);
1218
1219 if (edge->recursive_p ())
1220 badness = badness.shift (badness > 0 ? 4 : -4);
1221 if ((hints & (INLINE_HINT_indirect_call
1222 | INLINE_HINT_loop_iterations
1223 | INLINE_HINT_array_index
1224 | INLINE_HINT_loop_stride))
1225 || callee_info->growth <= 0)
1226 badness = badness.shift (badness > 0 ? -2 : 2);
1227 if (hints & (INLINE_HINT_same_scc))
1228 badness = badness.shift (badness > 0 ? 3 : -3);
1229 else if (hints & (INLINE_HINT_in_scc))
1230 badness = badness.shift (badness > 0 ? 2 : -2);
1231 else if (hints & (INLINE_HINT_cross_module))
1232 badness = badness.shift (badness > 0 ? 1 : -1);
1233 if (DECL_DISREGARD_INLINE_LIMITS (callee->decl))
1234 badness = badness.shift (badness > 0 ? -4 : 4);
1235 else if ((hints & INLINE_HINT_declared_inline))
1236 badness = badness.shift (badness > 0 ? -3 : 3);
1237 if (dump)
1238 fprintf (dump_file, " Adjusted by hints %f\n", badness.to_double ());
1239 return badness;
1240 }
1241
1242 /* Recompute badness of EDGE and update its key in HEAP if needed. */
1243 static inline void
1244 update_edge_key (edge_heap_t *heap, struct cgraph_edge *edge)
1245 {
1246 sreal badness = edge_badness (edge, false);
1247 if (edge->aux)
1248 {
1249 edge_heap_node_t *n = (edge_heap_node_t *) edge->aux;
1250 gcc_checking_assert (n->get_data () == edge);
1251
1252 /* fibonacci_heap::replace_key does busy updating of the
1253 heap that is unnecesarily expensive.
1254 We do lazy increases: after extracting minimum if the key
1255 turns out to be out of date, it is re-inserted into heap
1256 with correct value. */
1257 if (badness < n->get_key ())
1258 {
1259 if (dump_file && (dump_flags & TDF_DETAILS))
1260 {
1261 fprintf (dump_file,
1262 " decreasing badness %s -> %s, %f to %f\n",
1263 edge->caller->dump_name (),
1264 edge->callee->dump_name (),
1265 n->get_key ().to_double (),
1266 badness.to_double ());
1267 }
1268 heap->decrease_key (n, badness);
1269 }
1270 }
1271 else
1272 {
1273 if (dump_file && (dump_flags & TDF_DETAILS))
1274 {
1275 fprintf (dump_file,
1276 " enqueuing call %s -> %s, badness %f\n",
1277 edge->caller->dump_name (),
1278 edge->callee->dump_name (),
1279 badness.to_double ());
1280 }
1281 edge->aux = heap->insert (badness, edge);
1282 }
1283 }
1284
1285
1286 /* NODE was inlined.
1287 All caller edges needs to be resetted because
1288 size estimates change. Similarly callees needs reset
1289 because better context may be known. */
1290
1291 static void
1292 reset_edge_caches (struct cgraph_node *node)
1293 {
1294 struct cgraph_edge *edge;
1295 struct cgraph_edge *e = node->callees;
1296 struct cgraph_node *where = node;
1297 struct ipa_ref *ref;
1298
1299 if (where->global.inlined_to)
1300 where = where->global.inlined_to;
1301
1302 if (edge_growth_cache != NULL)
1303 for (edge = where->callers; edge; edge = edge->next_caller)
1304 if (edge->inline_failed)
1305 edge_growth_cache->remove (edge);
1306
1307 FOR_EACH_ALIAS (where, ref)
1308 reset_edge_caches (dyn_cast <cgraph_node *> (ref->referring));
1309
1310 if (!e)
1311 return;
1312
1313 while (true)
1314 if (!e->inline_failed && e->callee->callees)
1315 e = e->callee->callees;
1316 else
1317 {
1318 if (edge_growth_cache != NULL && e->inline_failed)
1319 edge_growth_cache->remove (e);
1320 if (e->next_callee)
1321 e = e->next_callee;
1322 else
1323 {
1324 do
1325 {
1326 if (e->caller == node)
1327 return;
1328 e = e->caller->callers;
1329 }
1330 while (!e->next_callee);
1331 e = e->next_callee;
1332 }
1333 }
1334 }
1335
1336 /* Recompute HEAP nodes for each of caller of NODE.
1337 UPDATED_NODES track nodes we already visited, to avoid redundant work.
1338 When CHECK_INLINABLITY_FOR is set, re-check for specified edge that
1339 it is inlinable. Otherwise check all edges. */
1340
1341 static void
1342 update_caller_keys (edge_heap_t *heap, struct cgraph_node *node,
1343 bitmap updated_nodes,
1344 struct cgraph_edge *check_inlinablity_for)
1345 {
1346 struct cgraph_edge *edge;
1347 struct ipa_ref *ref;
1348
1349 if ((!node->alias && !ipa_fn_summaries->get (node)->inlinable)
1350 || node->global.inlined_to)
1351 return;
1352 if (!bitmap_set_bit (updated_nodes, node->get_uid ()))
1353 return;
1354
1355 FOR_EACH_ALIAS (node, ref)
1356 {
1357 struct cgraph_node *alias = dyn_cast <cgraph_node *> (ref->referring);
1358 update_caller_keys (heap, alias, updated_nodes, check_inlinablity_for);
1359 }
1360
1361 for (edge = node->callers; edge; edge = edge->next_caller)
1362 if (edge->inline_failed)
1363 {
1364 if (!check_inlinablity_for
1365 || check_inlinablity_for == edge)
1366 {
1367 if (can_inline_edge_p (edge, false)
1368 && want_inline_small_function_p (edge, false)
1369 && can_inline_edge_by_limits_p (edge, false))
1370 update_edge_key (heap, edge);
1371 else if (edge->aux)
1372 {
1373 report_inline_failed_reason (edge);
1374 heap->delete_node ((edge_heap_node_t *) edge->aux);
1375 edge->aux = NULL;
1376 }
1377 }
1378 else if (edge->aux)
1379 update_edge_key (heap, edge);
1380 }
1381 }
1382
1383 /* Recompute HEAP nodes for each uninlined call in NODE.
1384 This is used when we know that edge badnesses are going only to increase
1385 (we introduced new call site) and thus all we need is to insert newly
1386 created edges into heap. */
1387
1388 static void
1389 update_callee_keys (edge_heap_t *heap, struct cgraph_node *node,
1390 bitmap updated_nodes)
1391 {
1392 struct cgraph_edge *e = node->callees;
1393
1394 if (!e)
1395 return;
1396 while (true)
1397 if (!e->inline_failed && e->callee->callees)
1398 e = e->callee->callees;
1399 else
1400 {
1401 enum availability avail;
1402 struct cgraph_node *callee;
1403 /* We do not reset callee growth cache here. Since we added a new call,
1404 growth chould have just increased and consequentely badness metric
1405 don't need updating. */
1406 if (e->inline_failed
1407 && (callee = e->callee->ultimate_alias_target (&avail, e->caller))
1408 && ipa_fn_summaries->get (callee) != NULL
1409 && ipa_fn_summaries->get (callee)->inlinable
1410 && avail >= AVAIL_AVAILABLE
1411 && !bitmap_bit_p (updated_nodes, callee->get_uid ()))
1412 {
1413 if (can_inline_edge_p (e, false)
1414 && want_inline_small_function_p (e, false)
1415 && can_inline_edge_by_limits_p (e, false))
1416 update_edge_key (heap, e);
1417 else if (e->aux)
1418 {
1419 report_inline_failed_reason (e);
1420 heap->delete_node ((edge_heap_node_t *) e->aux);
1421 e->aux = NULL;
1422 }
1423 }
1424 if (e->next_callee)
1425 e = e->next_callee;
1426 else
1427 {
1428 do
1429 {
1430 if (e->caller == node)
1431 return;
1432 e = e->caller->callers;
1433 }
1434 while (!e->next_callee);
1435 e = e->next_callee;
1436 }
1437 }
1438 }
1439
1440 /* Enqueue all recursive calls from NODE into priority queue depending on
1441 how likely we want to recursively inline the call. */
1442
1443 static void
1444 lookup_recursive_calls (struct cgraph_node *node, struct cgraph_node *where,
1445 edge_heap_t *heap)
1446 {
1447 struct cgraph_edge *e;
1448 enum availability avail;
1449
1450 for (e = where->callees; e; e = e->next_callee)
1451 if (e->callee == node
1452 || (e->callee->ultimate_alias_target (&avail, e->caller) == node
1453 && avail > AVAIL_INTERPOSABLE))
1454 heap->insert (-e->sreal_frequency (), e);
1455 for (e = where->callees; e; e = e->next_callee)
1456 if (!e->inline_failed)
1457 lookup_recursive_calls (node, e->callee, heap);
1458 }
1459
1460 /* Decide on recursive inlining: in the case function has recursive calls,
1461 inline until body size reaches given argument. If any new indirect edges
1462 are discovered in the process, add them to *NEW_EDGES, unless NEW_EDGES
1463 is NULL. */
1464
1465 static bool
1466 recursive_inlining (struct cgraph_edge *edge,
1467 vec<cgraph_edge *> *new_edges)
1468 {
1469 int limit = PARAM_VALUE (PARAM_MAX_INLINE_INSNS_RECURSIVE_AUTO);
1470 edge_heap_t heap (sreal::min ());
1471 struct cgraph_node *node;
1472 struct cgraph_edge *e;
1473 struct cgraph_node *master_clone = NULL, *next;
1474 int depth = 0;
1475 int n = 0;
1476
1477 node = edge->caller;
1478 if (node->global.inlined_to)
1479 node = node->global.inlined_to;
1480
1481 if (DECL_DECLARED_INLINE_P (node->decl))
1482 limit = PARAM_VALUE (PARAM_MAX_INLINE_INSNS_RECURSIVE);
1483
1484 /* Make sure that function is small enough to be considered for inlining. */
1485 if (estimate_size_after_inlining (node, edge) >= limit)
1486 return false;
1487 lookup_recursive_calls (node, node, &heap);
1488 if (heap.empty ())
1489 return false;
1490
1491 if (dump_file)
1492 fprintf (dump_file,
1493 " Performing recursive inlining on %s\n",
1494 node->name ());
1495
1496 /* Do the inlining and update list of recursive call during process. */
1497 while (!heap.empty ())
1498 {
1499 struct cgraph_edge *curr = heap.extract_min ();
1500 struct cgraph_node *cnode, *dest = curr->callee;
1501
1502 if (!can_inline_edge_p (curr, true)
1503 || can_inline_edge_by_limits_p (curr, true))
1504 continue;
1505
1506 /* MASTER_CLONE is produced in the case we already started modified
1507 the function. Be sure to redirect edge to the original body before
1508 estimating growths otherwise we will be seeing growths after inlining
1509 the already modified body. */
1510 if (master_clone)
1511 {
1512 curr->redirect_callee (master_clone);
1513 if (edge_growth_cache != NULL)
1514 edge_growth_cache->remove (curr);
1515 }
1516
1517 if (estimate_size_after_inlining (node, curr) > limit)
1518 {
1519 curr->redirect_callee (dest);
1520 if (edge_growth_cache != NULL)
1521 edge_growth_cache->remove (curr);
1522 break;
1523 }
1524
1525 depth = 1;
1526 for (cnode = curr->caller;
1527 cnode->global.inlined_to; cnode = cnode->callers->caller)
1528 if (node->decl
1529 == curr->callee->ultimate_alias_target ()->decl)
1530 depth++;
1531
1532 if (!want_inline_self_recursive_call_p (curr, node, false, depth))
1533 {
1534 curr->redirect_callee (dest);
1535 if (edge_growth_cache != NULL)
1536 edge_growth_cache->remove (curr);
1537 continue;
1538 }
1539
1540 if (dump_file)
1541 {
1542 fprintf (dump_file,
1543 " Inlining call of depth %i", depth);
1544 if (node->count.nonzero_p ())
1545 {
1546 fprintf (dump_file, " called approx. %.2f times per call",
1547 (double)curr->count.to_gcov_type ()
1548 / node->count.to_gcov_type ());
1549 }
1550 fprintf (dump_file, "\n");
1551 }
1552 if (!master_clone)
1553 {
1554 /* We need original clone to copy around. */
1555 master_clone = node->create_clone (node->decl, node->count,
1556 false, vNULL, true, NULL, NULL);
1557 for (e = master_clone->callees; e; e = e->next_callee)
1558 if (!e->inline_failed)
1559 clone_inlined_nodes (e, true, false, NULL);
1560 curr->redirect_callee (master_clone);
1561 if (edge_growth_cache != NULL)
1562 edge_growth_cache->remove (curr);
1563 }
1564
1565 inline_call (curr, false, new_edges, &overall_size, true);
1566 lookup_recursive_calls (node, curr->callee, &heap);
1567 n++;
1568 }
1569
1570 if (!heap.empty () && dump_file)
1571 fprintf (dump_file, " Recursive inlining growth limit met.\n");
1572
1573 if (!master_clone)
1574 return false;
1575
1576 if (dump_enabled_p ())
1577 dump_printf_loc (MSG_NOTE, edge->call_stmt,
1578 "\n Inlined %i times, "
1579 "body grown from size %i to %i, time %f to %f\n", n,
1580 ipa_fn_summaries->get (master_clone)->size,
1581 ipa_fn_summaries->get (node)->size,
1582 ipa_fn_summaries->get (master_clone)->time.to_double (),
1583 ipa_fn_summaries->get (node)->time.to_double ());
1584
1585 /* Remove master clone we used for inlining. We rely that clones inlined
1586 into master clone gets queued just before master clone so we don't
1587 need recursion. */
1588 for (node = symtab->first_function (); node != master_clone;
1589 node = next)
1590 {
1591 next = symtab->next_function (node);
1592 if (node->global.inlined_to == master_clone)
1593 node->remove ();
1594 }
1595 master_clone->remove ();
1596 return true;
1597 }
1598
1599
1600 /* Given whole compilation unit estimate of INSNS, compute how large we can
1601 allow the unit to grow. */
1602
1603 static int
1604 compute_max_insns (int insns)
1605 {
1606 int max_insns = insns;
1607 if (max_insns < PARAM_VALUE (PARAM_LARGE_UNIT_INSNS))
1608 max_insns = PARAM_VALUE (PARAM_LARGE_UNIT_INSNS);
1609
1610 return ((int64_t) max_insns
1611 * (100 + PARAM_VALUE (PARAM_INLINE_UNIT_GROWTH)) / 100);
1612 }
1613
1614
1615 /* Compute badness of all edges in NEW_EDGES and add them to the HEAP. */
1616
1617 static void
1618 add_new_edges_to_heap (edge_heap_t *heap, vec<cgraph_edge *> new_edges)
1619 {
1620 while (new_edges.length () > 0)
1621 {
1622 struct cgraph_edge *edge = new_edges.pop ();
1623
1624 gcc_assert (!edge->aux);
1625 if (edge->inline_failed
1626 && can_inline_edge_p (edge, true)
1627 && want_inline_small_function_p (edge, true)
1628 && can_inline_edge_by_limits_p (edge, true))
1629 edge->aux = heap->insert (edge_badness (edge, false), edge);
1630 }
1631 }
1632
1633 /* Remove EDGE from the fibheap. */
1634
1635 static void
1636 heap_edge_removal_hook (struct cgraph_edge *e, void *data)
1637 {
1638 if (e->aux)
1639 {
1640 ((edge_heap_t *)data)->delete_node ((edge_heap_node_t *)e->aux);
1641 e->aux = NULL;
1642 }
1643 }
1644
1645 /* Return true if speculation of edge E seems useful.
1646 If ANTICIPATE_INLINING is true, be conservative and hope that E
1647 may get inlined. */
1648
1649 bool
1650 speculation_useful_p (struct cgraph_edge *e, bool anticipate_inlining)
1651 {
1652 enum availability avail;
1653 struct cgraph_node *target = e->callee->ultimate_alias_target (&avail,
1654 e->caller);
1655 struct cgraph_edge *direct, *indirect;
1656 struct ipa_ref *ref;
1657
1658 gcc_assert (e->speculative && !e->indirect_unknown_callee);
1659
1660 if (!e->maybe_hot_p ())
1661 return false;
1662
1663 /* See if IP optimizations found something potentially useful about the
1664 function. For now we look only for CONST/PURE flags. Almost everything
1665 else we propagate is useless. */
1666 if (avail >= AVAIL_AVAILABLE)
1667 {
1668 int ecf_flags = flags_from_decl_or_type (target->decl);
1669 if (ecf_flags & ECF_CONST)
1670 {
1671 e->speculative_call_info (direct, indirect, ref);
1672 if (!(indirect->indirect_info->ecf_flags & ECF_CONST))
1673 return true;
1674 }
1675 else if (ecf_flags & ECF_PURE)
1676 {
1677 e->speculative_call_info (direct, indirect, ref);
1678 if (!(indirect->indirect_info->ecf_flags & ECF_PURE))
1679 return true;
1680 }
1681 }
1682 /* If we did not managed to inline the function nor redirect
1683 to an ipa-cp clone (that are seen by having local flag set),
1684 it is probably pointless to inline it unless hardware is missing
1685 indirect call predictor. */
1686 if (!anticipate_inlining && e->inline_failed && !target->local.local)
1687 return false;
1688 /* For overwritable targets there is not much to do. */
1689 if (e->inline_failed
1690 && (!can_inline_edge_p (e, false)
1691 || !can_inline_edge_by_limits_p (e, false, true)))
1692 return false;
1693 /* OK, speculation seems interesting. */
1694 return true;
1695 }
1696
1697 /* We know that EDGE is not going to be inlined.
1698 See if we can remove speculation. */
1699
1700 static void
1701 resolve_noninline_speculation (edge_heap_t *edge_heap, struct cgraph_edge *edge)
1702 {
1703 if (edge->speculative && !speculation_useful_p (edge, false))
1704 {
1705 struct cgraph_node *node = edge->caller;
1706 struct cgraph_node *where = node->global.inlined_to
1707 ? node->global.inlined_to : node;
1708 auto_bitmap updated_nodes;
1709
1710 if (edge->count.ipa ().initialized_p ())
1711 spec_rem += edge->count.ipa ();
1712 edge->resolve_speculation ();
1713 reset_edge_caches (where);
1714 ipa_update_overall_fn_summary (where);
1715 update_caller_keys (edge_heap, where,
1716 updated_nodes, NULL);
1717 update_callee_keys (edge_heap, where,
1718 updated_nodes);
1719 }
1720 }
1721
1722 /* Return true if NODE should be accounted for overall size estimate.
1723 Skip all nodes optimized for size so we can measure the growth of hot
1724 part of program no matter of the padding. */
1725
1726 bool
1727 inline_account_function_p (struct cgraph_node *node)
1728 {
1729 return (!DECL_EXTERNAL (node->decl)
1730 && !opt_for_fn (node->decl, optimize_size)
1731 && node->frequency != NODE_FREQUENCY_UNLIKELY_EXECUTED);
1732 }
1733
1734 /* Count number of callers of NODE and store it into DATA (that
1735 points to int. Worker for cgraph_for_node_and_aliases. */
1736
1737 static bool
1738 sum_callers (struct cgraph_node *node, void *data)
1739 {
1740 struct cgraph_edge *e;
1741 int *num_calls = (int *)data;
1742
1743 for (e = node->callers; e; e = e->next_caller)
1744 (*num_calls)++;
1745 return false;
1746 }
1747
1748 /* We use greedy algorithm for inlining of small functions:
1749 All inline candidates are put into prioritized heap ordered in
1750 increasing badness.
1751
1752 The inlining of small functions is bounded by unit growth parameters. */
1753
1754 static void
1755 inline_small_functions (void)
1756 {
1757 struct cgraph_node *node;
1758 struct cgraph_edge *edge;
1759 edge_heap_t edge_heap (sreal::min ());
1760 auto_bitmap updated_nodes;
1761 int min_size, max_size;
1762 auto_vec<cgraph_edge *> new_indirect_edges;
1763 int initial_size = 0;
1764 struct cgraph_node **order = XCNEWVEC (cgraph_node *, symtab->cgraph_count);
1765 struct cgraph_edge_hook_list *edge_removal_hook_holder;
1766 new_indirect_edges.create (8);
1767
1768 edge_removal_hook_holder
1769 = symtab->add_edge_removal_hook (&heap_edge_removal_hook, &edge_heap);
1770
1771 /* Compute overall unit size and other global parameters used by badness
1772 metrics. */
1773
1774 max_count = profile_count::uninitialized ();
1775 ipa_reduced_postorder (order, true, true, NULL);
1776 free (order);
1777
1778 FOR_EACH_DEFINED_FUNCTION (node)
1779 if (!node->global.inlined_to)
1780 {
1781 if (!node->alias && node->analyzed
1782 && (node->has_gimple_body_p () || node->thunk.thunk_p)
1783 && opt_for_fn (node->decl, optimize))
1784 {
1785 struct ipa_fn_summary *info = ipa_fn_summaries->get (node);
1786 struct ipa_dfs_info *dfs = (struct ipa_dfs_info *) node->aux;
1787
1788 /* Do not account external functions, they will be optimized out
1789 if not inlined. Also only count the non-cold portion of program. */
1790 if (inline_account_function_p (node))
1791 initial_size += info->size;
1792 info->growth = estimate_growth (node);
1793
1794 int num_calls = 0;
1795 node->call_for_symbol_and_aliases (sum_callers, &num_calls,
1796 true);
1797 if (num_calls == 1)
1798 info->single_caller = true;
1799 if (dfs && dfs->next_cycle)
1800 {
1801 struct cgraph_node *n2;
1802 int id = dfs->scc_no + 1;
1803 for (n2 = node; n2;
1804 n2 = ((struct ipa_dfs_info *) n2->aux)->next_cycle)
1805 if (opt_for_fn (n2->decl, optimize))
1806 {
1807 ipa_fn_summary *info2 = ipa_fn_summaries->get (n2);
1808 if (info2->scc_no)
1809 break;
1810 info2->scc_no = id;
1811 }
1812 }
1813 }
1814
1815 for (edge = node->callers; edge; edge = edge->next_caller)
1816 max_count = max_count.max (edge->count.ipa ());
1817 }
1818 ipa_free_postorder_info ();
1819 edge_growth_cache
1820 = new call_summary<edge_growth_cache_entry *> (symtab, false);
1821
1822 if (dump_file)
1823 fprintf (dump_file,
1824 "\nDeciding on inlining of small functions. Starting with size %i.\n",
1825 initial_size);
1826
1827 overall_size = initial_size;
1828 max_size = compute_max_insns (overall_size);
1829 min_size = overall_size;
1830
1831 /* Populate the heap with all edges we might inline. */
1832
1833 FOR_EACH_DEFINED_FUNCTION (node)
1834 {
1835 bool update = false;
1836 struct cgraph_edge *next = NULL;
1837 bool has_speculative = false;
1838
1839 if (!opt_for_fn (node->decl, optimize))
1840 continue;
1841
1842 if (dump_file)
1843 fprintf (dump_file, "Enqueueing calls in %s.\n", node->dump_name ());
1844
1845 for (edge = node->callees; edge; edge = next)
1846 {
1847 next = edge->next_callee;
1848 if (edge->inline_failed
1849 && !edge->aux
1850 && can_inline_edge_p (edge, true)
1851 && want_inline_small_function_p (edge, true)
1852 && can_inline_edge_by_limits_p (edge, true)
1853 && edge->inline_failed)
1854 {
1855 gcc_assert (!edge->aux);
1856 update_edge_key (&edge_heap, edge);
1857 }
1858 if (edge->speculative)
1859 has_speculative = true;
1860 }
1861 if (has_speculative)
1862 for (edge = node->callees; edge; edge = next)
1863 if (edge->speculative && !speculation_useful_p (edge,
1864 edge->aux != NULL))
1865 {
1866 edge->resolve_speculation ();
1867 update = true;
1868 }
1869 if (update)
1870 {
1871 struct cgraph_node *where = node->global.inlined_to
1872 ? node->global.inlined_to : node;
1873 ipa_update_overall_fn_summary (where);
1874 reset_edge_caches (where);
1875 update_caller_keys (&edge_heap, where,
1876 updated_nodes, NULL);
1877 update_callee_keys (&edge_heap, where,
1878 updated_nodes);
1879 bitmap_clear (updated_nodes);
1880 }
1881 }
1882
1883 gcc_assert (in_lto_p
1884 || !(max_count > 0)
1885 || (profile_info && flag_branch_probabilities));
1886
1887 while (!edge_heap.empty ())
1888 {
1889 int old_size = overall_size;
1890 struct cgraph_node *where, *callee;
1891 sreal badness = edge_heap.min_key ();
1892 sreal current_badness;
1893 int growth;
1894
1895 edge = edge_heap.extract_min ();
1896 gcc_assert (edge->aux);
1897 edge->aux = NULL;
1898 if (!edge->inline_failed || !edge->callee->analyzed)
1899 continue;
1900
1901 #if CHECKING_P
1902 /* Be sure that caches are maintained consistent.
1903 This check is affected by scaling roundoff errors when compiling for
1904 IPA this we skip it in that case. */
1905 if (!edge->callee->count.ipa_p ()
1906 && (!max_count.initialized_p () || !max_count.nonzero_p ()))
1907 {
1908 sreal cached_badness = edge_badness (edge, false);
1909
1910 int old_size_est = estimate_edge_size (edge);
1911 sreal old_time_est = estimate_edge_time (edge);
1912 int old_hints_est = estimate_edge_hints (edge);
1913
1914 if (edge_growth_cache != NULL)
1915 edge_growth_cache->remove (edge);
1916 gcc_assert (old_size_est == estimate_edge_size (edge));
1917 gcc_assert (old_time_est == estimate_edge_time (edge));
1918 /* FIXME:
1919
1920 gcc_assert (old_hints_est == estimate_edge_hints (edge));
1921
1922 fails with profile feedback because some hints depends on
1923 maybe_hot_edge_p predicate and because callee gets inlined to other
1924 calls, the edge may become cold.
1925 This ought to be fixed by computing relative probabilities
1926 for given invocation but that will be better done once whole
1927 code is converted to sreals. Disable for now and revert to "wrong"
1928 value so enable/disable checking paths agree. */
1929 edge_growth_cache->get (edge)->hints = old_hints_est + 1;
1930
1931 /* When updating the edge costs, we only decrease badness in the keys.
1932 Increases of badness are handled lazilly; when we see key with out
1933 of date value on it, we re-insert it now. */
1934 current_badness = edge_badness (edge, false);
1935 gcc_assert (cached_badness == current_badness);
1936 gcc_assert (current_badness >= badness);
1937 }
1938 else
1939 current_badness = edge_badness (edge, false);
1940 #else
1941 current_badness = edge_badness (edge, false);
1942 #endif
1943 if (current_badness != badness)
1944 {
1945 if (edge_heap.min () && current_badness > edge_heap.min_key ())
1946 {
1947 edge->aux = edge_heap.insert (current_badness, edge);
1948 continue;
1949 }
1950 else
1951 badness = current_badness;
1952 }
1953
1954 if (!can_inline_edge_p (edge, true)
1955 || !can_inline_edge_by_limits_p (edge, true))
1956 {
1957 resolve_noninline_speculation (&edge_heap, edge);
1958 continue;
1959 }
1960
1961 callee = edge->callee->ultimate_alias_target ();
1962 growth = estimate_edge_growth (edge);
1963 if (dump_file)
1964 {
1965 fprintf (dump_file,
1966 "\nConsidering %s with %i size\n",
1967 callee->dump_name (),
1968 ipa_fn_summaries->get (callee)->size);
1969 fprintf (dump_file,
1970 " to be inlined into %s in %s:%i\n"
1971 " Estimated badness is %f, frequency %.2f.\n",
1972 edge->caller->dump_name (),
1973 edge->call_stmt
1974 && (LOCATION_LOCUS (gimple_location ((const gimple *)
1975 edge->call_stmt))
1976 > BUILTINS_LOCATION)
1977 ? gimple_filename ((const gimple *) edge->call_stmt)
1978 : "unknown",
1979 edge->call_stmt
1980 ? gimple_lineno ((const gimple *) edge->call_stmt)
1981 : -1,
1982 badness.to_double (),
1983 edge->sreal_frequency ().to_double ());
1984 if (edge->count.ipa ().initialized_p ())
1985 {
1986 fprintf (dump_file, " Called ");
1987 edge->count.ipa ().dump (dump_file);
1988 fprintf (dump_file, " times\n");
1989 }
1990 if (dump_flags & TDF_DETAILS)
1991 edge_badness (edge, true);
1992 }
1993
1994 if (overall_size + growth > max_size
1995 && !DECL_DISREGARD_INLINE_LIMITS (callee->decl))
1996 {
1997 edge->inline_failed = CIF_INLINE_UNIT_GROWTH_LIMIT;
1998 report_inline_failed_reason (edge);
1999 resolve_noninline_speculation (&edge_heap, edge);
2000 continue;
2001 }
2002
2003 if (!want_inline_small_function_p (edge, true))
2004 {
2005 resolve_noninline_speculation (&edge_heap, edge);
2006 continue;
2007 }
2008
2009 /* Heuristics for inlining small functions work poorly for
2010 recursive calls where we do effects similar to loop unrolling.
2011 When inlining such edge seems profitable, leave decision on
2012 specific inliner. */
2013 if (edge->recursive_p ())
2014 {
2015 where = edge->caller;
2016 if (where->global.inlined_to)
2017 where = where->global.inlined_to;
2018 if (!recursive_inlining (edge,
2019 opt_for_fn (edge->caller->decl,
2020 flag_indirect_inlining)
2021 ? &new_indirect_edges : NULL))
2022 {
2023 edge->inline_failed = CIF_RECURSIVE_INLINING;
2024 resolve_noninline_speculation (&edge_heap, edge);
2025 continue;
2026 }
2027 reset_edge_caches (where);
2028 /* Recursive inliner inlines all recursive calls of the function
2029 at once. Consequently we need to update all callee keys. */
2030 if (opt_for_fn (edge->caller->decl, flag_indirect_inlining))
2031 add_new_edges_to_heap (&edge_heap, new_indirect_edges);
2032 update_callee_keys (&edge_heap, where, updated_nodes);
2033 bitmap_clear (updated_nodes);
2034 }
2035 else
2036 {
2037 struct cgraph_node *outer_node = NULL;
2038 int depth = 0;
2039
2040 /* Consider the case where self recursive function A is inlined
2041 into B. This is desired optimization in some cases, since it
2042 leads to effect similar of loop peeling and we might completely
2043 optimize out the recursive call. However we must be extra
2044 selective. */
2045
2046 where = edge->caller;
2047 while (where->global.inlined_to)
2048 {
2049 if (where->decl == callee->decl)
2050 outer_node = where, depth++;
2051 where = where->callers->caller;
2052 }
2053 if (outer_node
2054 && !want_inline_self_recursive_call_p (edge, outer_node,
2055 true, depth))
2056 {
2057 edge->inline_failed
2058 = (DECL_DISREGARD_INLINE_LIMITS (edge->callee->decl)
2059 ? CIF_RECURSIVE_INLINING : CIF_UNSPECIFIED);
2060 resolve_noninline_speculation (&edge_heap, edge);
2061 continue;
2062 }
2063 else if (depth && dump_file)
2064 fprintf (dump_file, " Peeling recursion with depth %i\n", depth);
2065
2066 gcc_checking_assert (!callee->global.inlined_to);
2067 inline_call (edge, true, &new_indirect_edges, &overall_size, true);
2068 add_new_edges_to_heap (&edge_heap, new_indirect_edges);
2069
2070 reset_edge_caches (edge->callee);
2071
2072 update_callee_keys (&edge_heap, where, updated_nodes);
2073 }
2074 where = edge->caller;
2075 if (where->global.inlined_to)
2076 where = where->global.inlined_to;
2077
2078 /* Our profitability metric can depend on local properties
2079 such as number of inlinable calls and size of the function body.
2080 After inlining these properties might change for the function we
2081 inlined into (since it's body size changed) and for the functions
2082 called by function we inlined (since number of it inlinable callers
2083 might change). */
2084 update_caller_keys (&edge_heap, where, updated_nodes, NULL);
2085 /* Offline copy count has possibly changed, recompute if profile is
2086 available. */
2087 struct cgraph_node *n = cgraph_node::get (edge->callee->decl);
2088 if (n != edge->callee && n->analyzed && n->count.ipa ().initialized_p ())
2089 update_callee_keys (&edge_heap, n, updated_nodes);
2090 bitmap_clear (updated_nodes);
2091
2092 if (dump_enabled_p ())
2093 {
2094 ipa_fn_summary *s = ipa_fn_summaries->get (edge->caller);
2095
2096 /* dump_printf can't handle %+i. */
2097 char buf_net_change[100];
2098 snprintf (buf_net_change, sizeof buf_net_change, "%+i",
2099 overall_size - old_size);
2100
2101 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, edge->call_stmt,
2102 " Inlined %C into %C which now has time %f and "
2103 "size %i, net change of %s.\n",
2104 edge->callee, edge->caller,
2105 s->time.to_double (), s->size, buf_net_change);
2106 }
2107 if (min_size > overall_size)
2108 {
2109 min_size = overall_size;
2110 max_size = compute_max_insns (min_size);
2111
2112 if (dump_file)
2113 fprintf (dump_file, "New minimal size reached: %i\n", min_size);
2114 }
2115 }
2116
2117 free_growth_caches ();
2118 if (dump_enabled_p ())
2119 dump_printf (MSG_NOTE,
2120 "Unit growth for small function inlining: %i->%i (%i%%)\n",
2121 initial_size, overall_size,
2122 initial_size ? overall_size * 100 / (initial_size) - 100: 0);
2123 symtab->remove_edge_removal_hook (edge_removal_hook_holder);
2124 }
2125
2126 /* Flatten NODE. Performed both during early inlining and
2127 at IPA inlining time. */
2128
2129 static void
2130 flatten_function (struct cgraph_node *node, bool early)
2131 {
2132 struct cgraph_edge *e;
2133
2134 /* We shouldn't be called recursively when we are being processed. */
2135 gcc_assert (node->aux == NULL);
2136
2137 node->aux = (void *) node;
2138
2139 for (e = node->callees; e; e = e->next_callee)
2140 {
2141 struct cgraph_node *orig_callee;
2142 struct cgraph_node *callee = e->callee->ultimate_alias_target ();
2143
2144 /* We've hit cycle? It is time to give up. */
2145 if (callee->aux)
2146 {
2147 if (dump_enabled_p ())
2148 dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
2149 "Not inlining %C into %C to avoid cycle.\n",
2150 callee, e->caller);
2151 if (cgraph_inline_failed_type (e->inline_failed) != CIF_FINAL_ERROR)
2152 e->inline_failed = CIF_RECURSIVE_INLINING;
2153 continue;
2154 }
2155
2156 /* When the edge is already inlined, we just need to recurse into
2157 it in order to fully flatten the leaves. */
2158 if (!e->inline_failed)
2159 {
2160 flatten_function (callee, early);
2161 continue;
2162 }
2163
2164 /* Flatten attribute needs to be processed during late inlining. For
2165 extra code quality we however do flattening during early optimization,
2166 too. */
2167 if (!early
2168 ? !can_inline_edge_p (e, true)
2169 && !can_inline_edge_by_limits_p (e, true)
2170 : !can_early_inline_edge_p (e))
2171 continue;
2172
2173 if (e->recursive_p ())
2174 {
2175 if (dump_enabled_p ())
2176 dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
2177 "Not inlining: recursive call.\n");
2178 continue;
2179 }
2180
2181 if (gimple_in_ssa_p (DECL_STRUCT_FUNCTION (node->decl))
2182 != gimple_in_ssa_p (DECL_STRUCT_FUNCTION (callee->decl)))
2183 {
2184 if (dump_enabled_p ())
2185 dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
2186 "Not inlining: SSA form does not match.\n");
2187 continue;
2188 }
2189
2190 /* Inline the edge and flatten the inline clone. Avoid
2191 recursing through the original node if the node was cloned. */
2192 if (dump_enabled_p ())
2193 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, e->call_stmt,
2194 " Inlining %C into %C.\n",
2195 callee, e->caller);
2196 orig_callee = callee;
2197 inline_call (e, true, NULL, NULL, false);
2198 if (e->callee != orig_callee)
2199 orig_callee->aux = (void *) node;
2200 flatten_function (e->callee, early);
2201 if (e->callee != orig_callee)
2202 orig_callee->aux = NULL;
2203 }
2204
2205 node->aux = NULL;
2206 if (!node->global.inlined_to)
2207 ipa_update_overall_fn_summary (node);
2208 }
2209
2210 /* Inline NODE to all callers. Worker for cgraph_for_node_and_aliases.
2211 DATA points to number of calls originally found so we avoid infinite
2212 recursion. */
2213
2214 static bool
2215 inline_to_all_callers_1 (struct cgraph_node *node, void *data,
2216 hash_set<cgraph_node *> *callers)
2217 {
2218 int *num_calls = (int *)data;
2219 bool callee_removed = false;
2220
2221 while (node->callers && !node->global.inlined_to)
2222 {
2223 struct cgraph_node *caller = node->callers->caller;
2224
2225 if (!can_inline_edge_p (node->callers, true)
2226 || !can_inline_edge_by_limits_p (node->callers, true)
2227 || node->callers->recursive_p ())
2228 {
2229 if (dump_file)
2230 fprintf (dump_file, "Uninlinable call found; giving up.\n");
2231 *num_calls = 0;
2232 return false;
2233 }
2234
2235 if (dump_file)
2236 {
2237 cgraph_node *ultimate = node->ultimate_alias_target ();
2238 fprintf (dump_file,
2239 "\nInlining %s size %i.\n",
2240 ultimate->name (),
2241 ipa_fn_summaries->get (ultimate)->size);
2242 fprintf (dump_file,
2243 " Called once from %s %i insns.\n",
2244 node->callers->caller->name (),
2245 ipa_fn_summaries->get (node->callers->caller)->size);
2246 }
2247
2248 /* Remember which callers we inlined to, delaying updating the
2249 overall summary. */
2250 callers->add (node->callers->caller);
2251 inline_call (node->callers, true, NULL, NULL, false, &callee_removed);
2252 if (dump_file)
2253 fprintf (dump_file,
2254 " Inlined into %s which now has %i size\n",
2255 caller->name (),
2256 ipa_fn_summaries->get (caller)->size);
2257 if (!(*num_calls)--)
2258 {
2259 if (dump_file)
2260 fprintf (dump_file, "New calls found; giving up.\n");
2261 return callee_removed;
2262 }
2263 if (callee_removed)
2264 return true;
2265 }
2266 return false;
2267 }
2268
2269 /* Wrapper around inline_to_all_callers_1 doing delayed overall summary
2270 update. */
2271
2272 static bool
2273 inline_to_all_callers (struct cgraph_node *node, void *data)
2274 {
2275 hash_set<cgraph_node *> callers;
2276 bool res = inline_to_all_callers_1 (node, data, &callers);
2277 /* Perform the delayed update of the overall summary of all callers
2278 processed. This avoids quadratic behavior in the cases where
2279 we have a lot of calls to the same function. */
2280 for (hash_set<cgraph_node *>::iterator i = callers.begin ();
2281 i != callers.end (); ++i)
2282 ipa_update_overall_fn_summary (*i);
2283 return res;
2284 }
2285
2286 /* Output overall time estimate. */
2287 static void
2288 dump_overall_stats (void)
2289 {
2290 sreal sum_weighted = 0, sum = 0;
2291 struct cgraph_node *node;
2292
2293 FOR_EACH_DEFINED_FUNCTION (node)
2294 if (!node->global.inlined_to
2295 && !node->alias)
2296 {
2297 ipa_fn_summary *s = ipa_fn_summaries->get (node);
2298 if (s != NULL)
2299 {
2300 sum += s->time;
2301 if (node->count.ipa ().initialized_p ())
2302 sum_weighted += s->time * node->count.ipa ().to_gcov_type ();
2303 }
2304 }
2305 fprintf (dump_file, "Overall time estimate: "
2306 "%f weighted by profile: "
2307 "%f\n", sum.to_double (), sum_weighted.to_double ());
2308 }
2309
2310 /* Output some useful stats about inlining. */
2311
2312 static void
2313 dump_inline_stats (void)
2314 {
2315 int64_t inlined_cnt = 0, inlined_indir_cnt = 0;
2316 int64_t inlined_virt_cnt = 0, inlined_virt_indir_cnt = 0;
2317 int64_t noninlined_cnt = 0, noninlined_indir_cnt = 0;
2318 int64_t noninlined_virt_cnt = 0, noninlined_virt_indir_cnt = 0;
2319 int64_t inlined_speculative = 0, inlined_speculative_ply = 0;
2320 int64_t indirect_poly_cnt = 0, indirect_cnt = 0;
2321 int64_t reason[CIF_N_REASONS][2];
2322 sreal reason_freq[CIF_N_REASONS];
2323 int i;
2324 struct cgraph_node *node;
2325
2326 memset (reason, 0, sizeof (reason));
2327 for (i=0; i < CIF_N_REASONS; i++)
2328 reason_freq[i] = 0;
2329 FOR_EACH_DEFINED_FUNCTION (node)
2330 {
2331 struct cgraph_edge *e;
2332 for (e = node->callees; e; e = e->next_callee)
2333 {
2334 if (e->inline_failed)
2335 {
2336 if (e->count.ipa ().initialized_p ())
2337 reason[(int) e->inline_failed][0] += e->count.ipa ().to_gcov_type ();
2338 reason_freq[(int) e->inline_failed] += e->sreal_frequency ();
2339 reason[(int) e->inline_failed][1] ++;
2340 if (DECL_VIRTUAL_P (e->callee->decl)
2341 && e->count.ipa ().initialized_p ())
2342 {
2343 if (e->indirect_inlining_edge)
2344 noninlined_virt_indir_cnt += e->count.ipa ().to_gcov_type ();
2345 else
2346 noninlined_virt_cnt += e->count.ipa ().to_gcov_type ();
2347 }
2348 else if (e->count.ipa ().initialized_p ())
2349 {
2350 if (e->indirect_inlining_edge)
2351 noninlined_indir_cnt += e->count.ipa ().to_gcov_type ();
2352 else
2353 noninlined_cnt += e->count.ipa ().to_gcov_type ();
2354 }
2355 }
2356 else if (e->count.ipa ().initialized_p ())
2357 {
2358 if (e->speculative)
2359 {
2360 if (DECL_VIRTUAL_P (e->callee->decl))
2361 inlined_speculative_ply += e->count.ipa ().to_gcov_type ();
2362 else
2363 inlined_speculative += e->count.ipa ().to_gcov_type ();
2364 }
2365 else if (DECL_VIRTUAL_P (e->callee->decl))
2366 {
2367 if (e->indirect_inlining_edge)
2368 inlined_virt_indir_cnt += e->count.ipa ().to_gcov_type ();
2369 else
2370 inlined_virt_cnt += e->count.ipa ().to_gcov_type ();
2371 }
2372 else
2373 {
2374 if (e->indirect_inlining_edge)
2375 inlined_indir_cnt += e->count.ipa ().to_gcov_type ();
2376 else
2377 inlined_cnt += e->count.ipa ().to_gcov_type ();
2378 }
2379 }
2380 }
2381 for (e = node->indirect_calls; e; e = e->next_callee)
2382 if (e->indirect_info->polymorphic
2383 & e->count.ipa ().initialized_p ())
2384 indirect_poly_cnt += e->count.ipa ().to_gcov_type ();
2385 else if (e->count.ipa ().initialized_p ())
2386 indirect_cnt += e->count.ipa ().to_gcov_type ();
2387 }
2388 if (max_count.initialized_p ())
2389 {
2390 fprintf (dump_file,
2391 "Inlined %" PRId64 " + speculative "
2392 "%" PRId64 " + speculative polymorphic "
2393 "%" PRId64 " + previously indirect "
2394 "%" PRId64 " + virtual "
2395 "%" PRId64 " + virtual and previously indirect "
2396 "%" PRId64 "\n" "Not inlined "
2397 "%" PRId64 " + previously indirect "
2398 "%" PRId64 " + virtual "
2399 "%" PRId64 " + virtual and previously indirect "
2400 "%" PRId64 " + stil indirect "
2401 "%" PRId64 " + still indirect polymorphic "
2402 "%" PRId64 "\n", inlined_cnt,
2403 inlined_speculative, inlined_speculative_ply,
2404 inlined_indir_cnt, inlined_virt_cnt, inlined_virt_indir_cnt,
2405 noninlined_cnt, noninlined_indir_cnt, noninlined_virt_cnt,
2406 noninlined_virt_indir_cnt, indirect_cnt, indirect_poly_cnt);
2407 fprintf (dump_file, "Removed speculations ");
2408 spec_rem.dump (dump_file);
2409 fprintf (dump_file, "\n");
2410 }
2411 dump_overall_stats ();
2412 fprintf (dump_file, "\nWhy inlining failed?\n");
2413 for (i = 0; i < CIF_N_REASONS; i++)
2414 if (reason[i][1])
2415 fprintf (dump_file, "%-50s: %8i calls, %8f freq, %" PRId64" count\n",
2416 cgraph_inline_failed_string ((cgraph_inline_failed_t) i),
2417 (int) reason[i][1], reason_freq[i].to_double (), reason[i][0]);
2418 }
2419
2420 /* Called when node is removed. */
2421
2422 static void
2423 flatten_remove_node_hook (struct cgraph_node *node, void *data)
2424 {
2425 if (lookup_attribute ("flatten", DECL_ATTRIBUTES (node->decl)) == NULL)
2426 return;
2427
2428 hash_set<struct cgraph_node *> *removed
2429 = (hash_set<struct cgraph_node *> *) data;
2430 removed->add (node);
2431 }
2432
2433 /* Decide on the inlining. We do so in the topological order to avoid
2434 expenses on updating data structures. */
2435
2436 static unsigned int
2437 ipa_inline (void)
2438 {
2439 struct cgraph_node *node;
2440 int nnodes;
2441 struct cgraph_node **order;
2442 int i, j;
2443 int cold;
2444 bool remove_functions = false;
2445
2446 order = XCNEWVEC (struct cgraph_node *, symtab->cgraph_count);
2447
2448 if (dump_file)
2449 ipa_dump_fn_summaries (dump_file);
2450
2451 nnodes = ipa_reverse_postorder (order);
2452 spec_rem = profile_count::zero ();
2453
2454 FOR_EACH_FUNCTION (node)
2455 {
2456 node->aux = 0;
2457
2458 /* Recompute the default reasons for inlining because they may have
2459 changed during merging. */
2460 if (in_lto_p)
2461 {
2462 for (cgraph_edge *e = node->callees; e; e = e->next_callee)
2463 {
2464 gcc_assert (e->inline_failed);
2465 initialize_inline_failed (e);
2466 }
2467 for (cgraph_edge *e = node->indirect_calls; e; e = e->next_callee)
2468 initialize_inline_failed (e);
2469 }
2470 }
2471
2472 if (dump_file)
2473 fprintf (dump_file, "\nFlattening functions:\n");
2474
2475 /* First shrink order array, so that it only contains nodes with
2476 flatten attribute. */
2477 for (i = nnodes - 1, j = i; i >= 0; i--)
2478 {
2479 node = order[i];
2480 if (lookup_attribute ("flatten",
2481 DECL_ATTRIBUTES (node->decl)) != NULL)
2482 order[j--] = order[i];
2483 }
2484
2485 /* After the above loop, order[j + 1] ... order[nnodes - 1] contain
2486 nodes with flatten attribute. If there is more than one such
2487 node, we need to register a node removal hook, as flatten_function
2488 could remove other nodes with flatten attribute. See PR82801. */
2489 struct cgraph_node_hook_list *node_removal_hook_holder = NULL;
2490 hash_set<struct cgraph_node *> *flatten_removed_nodes = NULL;
2491 if (j < nnodes - 2)
2492 {
2493 flatten_removed_nodes = new hash_set<struct cgraph_node *>;
2494 node_removal_hook_holder
2495 = symtab->add_cgraph_removal_hook (&flatten_remove_node_hook,
2496 flatten_removed_nodes);
2497 }
2498
2499 /* In the first pass handle functions to be flattened. Do this with
2500 a priority so none of our later choices will make this impossible. */
2501 for (i = nnodes - 1; i > j; i--)
2502 {
2503 node = order[i];
2504 if (flatten_removed_nodes
2505 && flatten_removed_nodes->contains (node))
2506 continue;
2507
2508 /* Handle nodes to be flattened.
2509 Ideally when processing callees we stop inlining at the
2510 entry of cycles, possibly cloning that entry point and
2511 try to flatten itself turning it into a self-recursive
2512 function. */
2513 if (dump_file)
2514 fprintf (dump_file, "Flattening %s\n", node->name ());
2515 flatten_function (node, false);
2516 }
2517
2518 if (j < nnodes - 2)
2519 {
2520 symtab->remove_cgraph_removal_hook (node_removal_hook_holder);
2521 delete flatten_removed_nodes;
2522 }
2523 free (order);
2524
2525 if (dump_file)
2526 dump_overall_stats ();
2527
2528 inline_small_functions ();
2529
2530 gcc_assert (symtab->state == IPA_SSA);
2531 symtab->state = IPA_SSA_AFTER_INLINING;
2532 /* Do first after-inlining removal. We want to remove all "stale" extern
2533 inline functions and virtual functions so we really know what is called
2534 once. */
2535 symtab->remove_unreachable_nodes (dump_file);
2536
2537 /* Inline functions with a property that after inlining into all callers the
2538 code size will shrink because the out-of-line copy is eliminated.
2539 We do this regardless on the callee size as long as function growth limits
2540 are met. */
2541 if (dump_file)
2542 fprintf (dump_file,
2543 "\nDeciding on functions to be inlined into all callers and "
2544 "removing useless speculations:\n");
2545
2546 /* Inlining one function called once has good chance of preventing
2547 inlining other function into the same callee. Ideally we should
2548 work in priority order, but probably inlining hot functions first
2549 is good cut without the extra pain of maintaining the queue.
2550
2551 ??? this is not really fitting the bill perfectly: inlining function
2552 into callee often leads to better optimization of callee due to
2553 increased context for optimization.
2554 For example if main() function calls a function that outputs help
2555 and then function that does the main optmization, we should inline
2556 the second with priority even if both calls are cold by themselves.
2557
2558 We probably want to implement new predicate replacing our use of
2559 maybe_hot_edge interpreted as maybe_hot_edge || callee is known
2560 to be hot. */
2561 for (cold = 0; cold <= 1; cold ++)
2562 {
2563 FOR_EACH_DEFINED_FUNCTION (node)
2564 {
2565 struct cgraph_edge *edge, *next;
2566 bool update=false;
2567
2568 if (!opt_for_fn (node->decl, optimize)
2569 || !opt_for_fn (node->decl, flag_inline_functions_called_once))
2570 continue;
2571
2572 for (edge = node->callees; edge; edge = next)
2573 {
2574 next = edge->next_callee;
2575 if (edge->speculative && !speculation_useful_p (edge, false))
2576 {
2577 if (edge->count.ipa ().initialized_p ())
2578 spec_rem += edge->count.ipa ();
2579 edge->resolve_speculation ();
2580 update = true;
2581 remove_functions = true;
2582 }
2583 }
2584 if (update)
2585 {
2586 struct cgraph_node *where = node->global.inlined_to
2587 ? node->global.inlined_to : node;
2588 reset_edge_caches (where);
2589 ipa_update_overall_fn_summary (where);
2590 }
2591 if (want_inline_function_to_all_callers_p (node, cold))
2592 {
2593 int num_calls = 0;
2594 node->call_for_symbol_and_aliases (sum_callers, &num_calls,
2595 true);
2596 while (node->call_for_symbol_and_aliases
2597 (inline_to_all_callers, &num_calls, true))
2598 ;
2599 remove_functions = true;
2600 }
2601 }
2602 }
2603
2604 /* Free ipa-prop structures if they are no longer needed. */
2605 ipa_free_all_structures_after_iinln ();
2606
2607 if (dump_enabled_p ())
2608 dump_printf (MSG_NOTE,
2609 "\nInlined %i calls, eliminated %i functions\n\n",
2610 ncalls_inlined, nfunctions_inlined);
2611 if (dump_file)
2612 dump_inline_stats ();
2613
2614 if (dump_file)
2615 ipa_dump_fn_summaries (dump_file);
2616 return remove_functions ? TODO_remove_functions : 0;
2617 }
2618
2619 /* Inline always-inline function calls in NODE. */
2620
2621 static bool
2622 inline_always_inline_functions (struct cgraph_node *node)
2623 {
2624 struct cgraph_edge *e;
2625 bool inlined = false;
2626
2627 for (e = node->callees; e; e = e->next_callee)
2628 {
2629 struct cgraph_node *callee = e->callee->ultimate_alias_target ();
2630 if (!DECL_DISREGARD_INLINE_LIMITS (callee->decl))
2631 continue;
2632
2633 if (e->recursive_p ())
2634 {
2635 if (dump_enabled_p ())
2636 dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
2637 " Not inlining recursive call to %C.\n",
2638 e->callee);
2639 e->inline_failed = CIF_RECURSIVE_INLINING;
2640 continue;
2641 }
2642
2643 if (!can_early_inline_edge_p (e))
2644 {
2645 /* Set inlined to true if the callee is marked "always_inline" but
2646 is not inlinable. This will allow flagging an error later in
2647 expand_call_inline in tree-inline.c. */
2648 if (lookup_attribute ("always_inline",
2649 DECL_ATTRIBUTES (callee->decl)) != NULL)
2650 inlined = true;
2651 continue;
2652 }
2653
2654 if (dump_enabled_p ())
2655 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, e->call_stmt,
2656 " Inlining %C into %C (always_inline).\n",
2657 e->callee, e->caller);
2658 inline_call (e, true, NULL, NULL, false);
2659 inlined = true;
2660 }
2661 if (inlined)
2662 ipa_update_overall_fn_summary (node);
2663
2664 return inlined;
2665 }
2666
2667 /* Decide on the inlining. We do so in the topological order to avoid
2668 expenses on updating data structures. */
2669
2670 static bool
2671 early_inline_small_functions (struct cgraph_node *node)
2672 {
2673 struct cgraph_edge *e;
2674 bool inlined = false;
2675
2676 for (e = node->callees; e; e = e->next_callee)
2677 {
2678 struct cgraph_node *callee = e->callee->ultimate_alias_target ();
2679
2680 /* We can enounter not-yet-analyzed function during
2681 early inlining on callgraphs with strongly
2682 connected components. */
2683 ipa_fn_summary *s = ipa_fn_summaries->get (callee);
2684 if (s == NULL || !s->inlinable || !e->inline_failed)
2685 continue;
2686
2687 /* Do not consider functions not declared inline. */
2688 if (!DECL_DECLARED_INLINE_P (callee->decl)
2689 && !opt_for_fn (node->decl, flag_inline_small_functions)
2690 && !opt_for_fn (node->decl, flag_inline_functions))
2691 continue;
2692
2693 if (dump_enabled_p ())
2694 dump_printf_loc (MSG_NOTE, e->call_stmt,
2695 "Considering inline candidate %C.\n",
2696 callee);
2697
2698 if (!can_early_inline_edge_p (e))
2699 continue;
2700
2701 if (e->recursive_p ())
2702 {
2703 if (dump_enabled_p ())
2704 dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt,
2705 " Not inlining: recursive call.\n");
2706 continue;
2707 }
2708
2709 if (!want_early_inline_function_p (e))
2710 continue;
2711
2712 if (dump_enabled_p ())
2713 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, e->call_stmt,
2714 " Inlining %C into %C.\n",
2715 callee, e->caller);
2716 inline_call (e, true, NULL, NULL, false);
2717 inlined = true;
2718 }
2719
2720 if (inlined)
2721 ipa_update_overall_fn_summary (node);
2722
2723 return inlined;
2724 }
2725
2726 unsigned int
2727 early_inliner (function *fun)
2728 {
2729 struct cgraph_node *node = cgraph_node::get (current_function_decl);
2730 struct cgraph_edge *edge;
2731 unsigned int todo = 0;
2732 int iterations = 0;
2733 bool inlined = false;
2734
2735 if (seen_error ())
2736 return 0;
2737
2738 /* Do nothing if datastructures for ipa-inliner are already computed. This
2739 happens when some pass decides to construct new function and
2740 cgraph_add_new_function calls lowering passes and early optimization on
2741 it. This may confuse ourself when early inliner decide to inline call to
2742 function clone, because function clones don't have parameter list in
2743 ipa-prop matching their signature. */
2744 if (ipa_node_params_sum)
2745 return 0;
2746
2747 if (flag_checking)
2748 node->verify ();
2749 node->remove_all_references ();
2750
2751 /* Even when not optimizing or not inlining inline always-inline
2752 functions. */
2753 inlined = inline_always_inline_functions (node);
2754
2755 if (!optimize
2756 || flag_no_inline
2757 || !flag_early_inlining
2758 /* Never inline regular functions into always-inline functions
2759 during incremental inlining. This sucks as functions calling
2760 always inline functions will get less optimized, but at the
2761 same time inlining of functions calling always inline
2762 function into an always inline function might introduce
2763 cycles of edges to be always inlined in the callgraph.
2764
2765 We might want to be smarter and just avoid this type of inlining. */
2766 || (DECL_DISREGARD_INLINE_LIMITS (node->decl)
2767 && lookup_attribute ("always_inline",
2768 DECL_ATTRIBUTES (node->decl))))
2769 ;
2770 else if (lookup_attribute ("flatten",
2771 DECL_ATTRIBUTES (node->decl)) != NULL)
2772 {
2773 /* When the function is marked to be flattened, recursively inline
2774 all calls in it. */
2775 if (dump_enabled_p ())
2776 dump_printf (MSG_OPTIMIZED_LOCATIONS,
2777 "Flattening %C\n", node);
2778 flatten_function (node, true);
2779 inlined = true;
2780 }
2781 else
2782 {
2783 /* If some always_inline functions was inlined, apply the changes.
2784 This way we will not account always inline into growth limits and
2785 moreover we will inline calls from always inlines that we skipped
2786 previously because of conditional above. */
2787 if (inlined)
2788 {
2789 timevar_push (TV_INTEGRATION);
2790 todo |= optimize_inline_calls (current_function_decl);
2791 /* optimize_inline_calls call above might have introduced new
2792 statements that don't have inline parameters computed. */
2793 for (edge = node->callees; edge; edge = edge->next_callee)
2794 {
2795 /* We can enounter not-yet-analyzed function during
2796 early inlining on callgraphs with strongly
2797 connected components. */
2798 ipa_call_summary *es = ipa_call_summaries->get_create (edge);
2799 es->call_stmt_size
2800 = estimate_num_insns (edge->call_stmt, &eni_size_weights);
2801 es->call_stmt_time
2802 = estimate_num_insns (edge->call_stmt, &eni_time_weights);
2803 }
2804 ipa_update_overall_fn_summary (node);
2805 inlined = false;
2806 timevar_pop (TV_INTEGRATION);
2807 }
2808 /* We iterate incremental inlining to get trivial cases of indirect
2809 inlining. */
2810 while (iterations < PARAM_VALUE (PARAM_EARLY_INLINER_MAX_ITERATIONS)
2811 && early_inline_small_functions (node))
2812 {
2813 timevar_push (TV_INTEGRATION);
2814 todo |= optimize_inline_calls (current_function_decl);
2815
2816 /* Technically we ought to recompute inline parameters so the new
2817 iteration of early inliner works as expected. We however have
2818 values approximately right and thus we only need to update edge
2819 info that might be cleared out for newly discovered edges. */
2820 for (edge = node->callees; edge; edge = edge->next_callee)
2821 {
2822 /* We have no summary for new bound store calls yet. */
2823 ipa_call_summary *es = ipa_call_summaries->get_create (edge);
2824 es->call_stmt_size
2825 = estimate_num_insns (edge->call_stmt, &eni_size_weights);
2826 es->call_stmt_time
2827 = estimate_num_insns (edge->call_stmt, &eni_time_weights);
2828
2829 if (edge->callee->decl
2830 && !gimple_check_call_matching_types (
2831 edge->call_stmt, edge->callee->decl, false))
2832 {
2833 edge->inline_failed = CIF_MISMATCHED_ARGUMENTS;
2834 edge->call_stmt_cannot_inline_p = true;
2835 }
2836 }
2837 if (iterations < PARAM_VALUE (PARAM_EARLY_INLINER_MAX_ITERATIONS) - 1)
2838 ipa_update_overall_fn_summary (node);
2839 timevar_pop (TV_INTEGRATION);
2840 iterations++;
2841 inlined = false;
2842 }
2843 if (dump_file)
2844 fprintf (dump_file, "Iterations: %i\n", iterations);
2845 }
2846
2847 if (inlined)
2848 {
2849 timevar_push (TV_INTEGRATION);
2850 todo |= optimize_inline_calls (current_function_decl);
2851 timevar_pop (TV_INTEGRATION);
2852 }
2853
2854 fun->always_inline_functions_inlined = true;
2855
2856 return todo;
2857 }
2858
2859 /* Do inlining of small functions. Doing so early helps profiling and other
2860 passes to be somewhat more effective and avoids some code duplication in
2861 later real inlining pass for testcases with very many function calls. */
2862
2863 namespace {
2864
2865 const pass_data pass_data_early_inline =
2866 {
2867 GIMPLE_PASS, /* type */
2868 "einline", /* name */
2869 OPTGROUP_INLINE, /* optinfo_flags */
2870 TV_EARLY_INLINING, /* tv_id */
2871 PROP_ssa, /* properties_required */
2872 0, /* properties_provided */
2873 0, /* properties_destroyed */
2874 0, /* todo_flags_start */
2875 0, /* todo_flags_finish */
2876 };
2877
2878 class pass_early_inline : public gimple_opt_pass
2879 {
2880 public:
2881 pass_early_inline (gcc::context *ctxt)
2882 : gimple_opt_pass (pass_data_early_inline, ctxt)
2883 {}
2884
2885 /* opt_pass methods: */
2886 virtual unsigned int execute (function *);
2887
2888 }; // class pass_early_inline
2889
2890 unsigned int
2891 pass_early_inline::execute (function *fun)
2892 {
2893 return early_inliner (fun);
2894 }
2895
2896 } // anon namespace
2897
2898 gimple_opt_pass *
2899 make_pass_early_inline (gcc::context *ctxt)
2900 {
2901 return new pass_early_inline (ctxt);
2902 }
2903
2904 namespace {
2905
2906 const pass_data pass_data_ipa_inline =
2907 {
2908 IPA_PASS, /* type */
2909 "inline", /* name */
2910 OPTGROUP_INLINE, /* optinfo_flags */
2911 TV_IPA_INLINING, /* tv_id */
2912 0, /* properties_required */
2913 0, /* properties_provided */
2914 0, /* properties_destroyed */
2915 0, /* todo_flags_start */
2916 ( TODO_dump_symtab ), /* todo_flags_finish */
2917 };
2918
2919 class pass_ipa_inline : public ipa_opt_pass_d
2920 {
2921 public:
2922 pass_ipa_inline (gcc::context *ctxt)
2923 : ipa_opt_pass_d (pass_data_ipa_inline, ctxt,
2924 NULL, /* generate_summary */
2925 NULL, /* write_summary */
2926 NULL, /* read_summary */
2927 NULL, /* write_optimization_summary */
2928 NULL, /* read_optimization_summary */
2929 NULL, /* stmt_fixup */
2930 0, /* function_transform_todo_flags_start */
2931 inline_transform, /* function_transform */
2932 NULL) /* variable_transform */
2933 {}
2934
2935 /* opt_pass methods: */
2936 virtual unsigned int execute (function *) { return ipa_inline (); }
2937
2938 }; // class pass_ipa_inline
2939
2940 } // anon namespace
2941
2942 ipa_opt_pass_d *
2943 make_pass_ipa_inline (gcc::context *ctxt)
2944 {
2945 return new pass_ipa_inline (ctxt);
2946 }