]>
Commit | Line | Data |
---|---|---|
ca31b95f | 1 | /* Inlining decision heuristics. |
8d9254fc | 2 | Copyright (C) 2003-2020 Free Software Foundation, Inc. |
ca31b95f JH |
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 | |
9dcd6f09 | 9 | Software Foundation; either version 3, or (at your option) any later |
ca31b95f JH |
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 | |
9dcd6f09 NC |
18 | along with GCC; see the file COPYING3. If not see |
19 | <http://www.gnu.org/licenses/>. */ | |
ca31b95f JH |
20 | |
21 | /* Inlining decision heuristics | |
22 | ||
4c0f7679 | 23 | The implementation of inliner is organized as follows: |
ca31b95f | 24 | |
ca31b95f JH |
25 | inlining heuristics limits |
26 | ||
4c0f7679 JH |
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. | |
ca31b95f JH |
35 | |
36 | inlining heuristics | |
37 | ||
4c0f7679 JH |
38 | The inliner itself is split into two passes: |
39 | ||
40 | pass_early_inlining | |
ca31b95f | 41 | |
4c0f7679 JH |
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. | |
ca31b95f | 45 | |
4c0f7679 JH |
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 | |
09a2806f | 50 | and results of early inlining adds a lot of optimization opportunities |
4c0f7679 | 51 | for the local optimization. |
ca31b95f | 52 | |
09a2806f | 53 | The pass handle the obvious inlining decisions within the compilation |
4c0f7679 JH |
54 | unit - inlining auto inline functions, inlining for size and |
55 | flattening. | |
ca31b95f | 56 | |
4c0f7679 JH |
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. | |
873aa8f5 | 61 | |
67914693 | 62 | Because of lack of whole unit knowledge, the pass cannot really make |
4c0f7679 JH |
63 | good code size/performance tradeoffs. It however does very simple |
64 | speculative inlining allowing code size to grow by | |
09a2806f JH |
65 | EARLY_INLINING_INSNS when callee is leaf function. In this case the |
66 | optimizations performed later are very likely to eliminate the cost. | |
873aa8f5 | 67 | |
4c0f7679 | 68 | pass_ipa_inline |
873aa8f5 | 69 | |
4c0f7679 JH |
70 | This is the real inliner able to handle inlining with whole program |
71 | knowledge. It performs following steps: | |
873aa8f5 | 72 | |
4c0f7679 JH |
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. | |
873aa8f5 | 76 | |
4c0f7679 JH |
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. | |
873aa8f5 | 81 | |
4c0f7679 JH |
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. */ | |
ca31b95f JH |
91 | |
92 | #include "config.h" | |
93 | #include "system.h" | |
94 | #include "coretypes.h" | |
c7131fb2 | 95 | #include "backend.h" |
957060b5 AM |
96 | #include "target.h" |
97 | #include "rtl.h" | |
ca31b95f | 98 | #include "tree.h" |
c7131fb2 | 99 | #include "gimple.h" |
957060b5 AM |
100 | #include "alloc-pool.h" |
101 | #include "tree-pass.h" | |
102 | #include "gimple-ssa.h" | |
103 | #include "cgraph.h" | |
957060b5 | 104 | #include "lto-streamer.h" |
d8a2d370 DN |
105 | #include "trans-mem.h" |
106 | #include "calls.h" | |
ca31b95f | 107 | #include "tree-inline.h" |
59f2e9d8 | 108 | #include "profile.h" |
dd912cb8 | 109 | #include "symbol-summary.h" |
8bc5448f | 110 | #include "tree-vrp.h" |
3e293154 | 111 | #include "ipa-prop.h" |
27d020cf | 112 | #include "ipa-fnsummary.h" |
03dfc36d | 113 | #include "ipa-inline.h" |
af8bca3c | 114 | #include "ipa-utils.h" |
1b08b734 | 115 | #include "sreal.h" |
be3c16c4 | 116 | #include "auto-profile.h" |
9b2b7279 | 117 | #include "builtins.h" |
4a910049 | 118 | #include "fibonacci_heap.h" |
314e6352 ML |
119 | #include "stringpool.h" |
120 | #include "attribs.h" | |
45b2222a | 121 | #include "asan.h" |
4a910049 | 122 | |
f0e1509b ML |
123 | typedef fibonacci_heap <sreal, cgraph_edge> edge_heap_t; |
124 | typedef fibonacci_node <sreal, cgraph_edge> edge_heap_node_t; | |
85057983 | 125 | |
ca31b95f | 126 | /* Statistics we collect about inlining algorithm. */ |
85057983 | 127 | static int overall_size; |
3995f3a2 JH |
128 | static profile_count max_count; |
129 | static profile_count spec_rem; | |
ca31b95f | 130 | |
4c0f7679 JH |
131 | /* Return false when inlining edge E would lead to violating |
132 | limits on function unit growth or stack usage growth. | |
133 | ||
134 | The relative function body growth limit is present generally | |
09a2806f | 135 | to avoid problems with non-linear behavior of the compiler. |
4c0f7679 JH |
136 | To allow inlining huge functions into tiny wrapper, the limit |
137 | is always based on the bigger of the two functions considered. | |
138 | ||
139 | For stack growth limits we always base the growth in stack usage | |
140 | of the callers. We want to prevent applications from segfaulting | |
141 | on stack overflow when functions with huge stack frames gets | |
142 | inlined. */ | |
ca31b95f JH |
143 | |
144 | static bool | |
4c0f7679 | 145 | caller_growth_limits (struct cgraph_edge *e) |
ca31b95f | 146 | { |
d7d1d041 | 147 | struct cgraph_node *to = e->caller; |
d52f5295 | 148 | struct cgraph_node *what = e->callee->ultimate_alias_target (); |
ca31b95f | 149 | int newsize; |
4c0f7679 JH |
150 | int limit = 0; |
151 | HOST_WIDE_INT stack_size_limit = 0, inlined_stack; | |
f658ad30 | 152 | ipa_size_summary *outer_info = ipa_size_summaries->get (to); |
4c0f7679 JH |
153 | |
154 | /* Look for function e->caller is inlined to. While doing | |
155 | so work out the largest function body on the way. As | |
156 | described above, we want to base our function growth | |
157 | limits based on that. Not on the self size of the | |
158 | outer function, not on the self size of inline code | |
159 | we immediately inline to. This is the most relaxed | |
160 | interpretation of the rule "do not grow large functions | |
161 | too much in order to prevent compiler from exploding". */ | |
09dfe187 | 162 | while (true) |
4c0f7679 | 163 | { |
f658ad30 JH |
164 | ipa_size_summary *size_info = ipa_size_summaries->get (to); |
165 | if (limit < size_info->self_size) | |
166 | limit = size_info->self_size; | |
167 | if (stack_size_limit < size_info->estimated_self_stack_size) | |
168 | stack_size_limit = size_info->estimated_self_stack_size; | |
a62bfab5 | 169 | if (to->inlined_to) |
4c0f7679 | 170 | to = to->callers->caller; |
09dfe187 JH |
171 | else |
172 | break; | |
4c0f7679 | 173 | } |
6971d714 | 174 | |
f658ad30 JH |
175 | ipa_fn_summary *what_info = ipa_fn_summaries->get (what); |
176 | ipa_size_summary *what_size_info = ipa_size_summaries->get (what); | |
e7f23018 | 177 | |
f658ad30 JH |
178 | if (limit < what_size_info->self_size) |
179 | limit = what_size_info->self_size; | |
ca31b95f | 180 | |
1e83bd70 | 181 | limit += limit * opt_for_fn (to->decl, param_large_function_growth) / 100; |
ca31b95f | 182 | |
6971d714 RG |
183 | /* Check the size after inlining against the function limits. But allow |
184 | the function to shrink if it went over the limits by forced inlining. */ | |
03dfc36d | 185 | newsize = estimate_size_after_inlining (to, e); |
f658ad30 | 186 | if (newsize >= ipa_size_summaries->get (what)->size |
709d7838 LXH |
187 | && newsize > opt_for_fn (to->decl, param_large_function_insns) |
188 | && newsize > limit) | |
ca31b95f | 189 | { |
4c0f7679 | 190 | e->inline_failed = CIF_LARGE_FUNCTION_GROWTH_LIMIT; |
ca31b95f JH |
191 | return false; |
192 | } | |
ff28a94d | 193 | |
09dfe187 JH |
194 | if (!what_info->estimated_stack_size) |
195 | return true; | |
196 | ||
09a2806f JH |
197 | /* FIXME: Stack size limit often prevents inlining in Fortran programs |
198 | due to large i/o datastructures used by the Fortran front-end. | |
4c0f7679 JH |
199 | We ought to ignore this limit when we know that the edge is executed |
200 | on every invocation of the caller (i.e. its call statement dominates | |
201 | exit block). We do not track this information, yet. */ | |
09dfe187 | 202 | stack_size_limit += ((gcov_type)stack_size_limit |
1e83bd70 JH |
203 | * opt_for_fn (to->decl, param_stack_frame_growth) |
204 | / 100); | |
ff28a94d | 205 | |
f658ad30 | 206 | inlined_stack = (ipa_get_stack_frame_offset (to) |
4c0f7679 | 207 | + outer_info->estimated_self_stack_size |
e7f23018 | 208 | + what_info->estimated_stack_size); |
4c0f7679 JH |
209 | /* Check new stack consumption with stack consumption at the place |
210 | stack is used. */ | |
211 | if (inlined_stack > stack_size_limit | |
09a2806f | 212 | /* If function already has large stack usage from sibling |
4c0f7679 JH |
213 | inline call, we can inline, too. |
214 | This bit overoptimistically assume that we are good at stack | |
215 | packing. */ | |
f658ad30 | 216 | && inlined_stack > ipa_fn_summaries->get (to)->estimated_stack_size |
1e83bd70 | 217 | && inlined_stack > opt_for_fn (to->decl, param_large_stack_frame)) |
ff28a94d | 218 | { |
4c0f7679 | 219 | e->inline_failed = CIF_LARGE_STACK_FRAME_GROWTH_LIMIT; |
ff28a94d JH |
220 | return false; |
221 | } | |
ca31b95f JH |
222 | return true; |
223 | } | |
224 | ||
4c0f7679 JH |
225 | /* Dump info about why inlining has failed. */ |
226 | ||
227 | static void | |
228 | report_inline_failed_reason (struct cgraph_edge *e) | |
229 | { | |
4174a33a | 230 | if (dump_enabled_p ()) |
4c0f7679 | 231 | { |
4174a33a DM |
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)); | |
bb1e543c JH |
236 | if ((e->inline_failed == CIF_TARGET_OPTION_MISMATCH |
237 | || e->inline_failed == CIF_OPTIMIZATION_MISMATCH) | |
238 | && e->caller->lto_file_data | |
1f6f9079 | 239 | && e->callee->ultimate_alias_target ()->lto_file_data) |
bb1e543c | 240 | { |
4174a33a DM |
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); | |
bb1e543c JH |
245 | } |
246 | if (e->inline_failed == CIF_TARGET_OPTION_MISMATCH) | |
9228f64c DM |
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)); | |
bb1e543c | 251 | if (e->inline_failed == CIF_OPTIMIZATION_MISMATCH) |
9228f64c DM |
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)); | |
4c0f7679 JH |
256 | } |
257 | } | |
258 | ||
25a07c7e YG |
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 | { | |
25a07c7e YG |
264 | if (!caller || !callee) |
265 | return true; | |
266 | ||
4089df8e ML |
267 | /* Follow clang and allow inlining for always_inline functions. */ |
268 | if (lookup_attribute ("always_inline", DECL_ATTRIBUTES (callee))) | |
6206a883 JJ |
269 | return true; |
270 | ||
4089df8e ML |
271 | const sanitize_code codes[] = |
272 | { | |
273 | SANITIZE_ADDRESS, | |
274 | SANITIZE_THREAD, | |
275 | SANITIZE_UNDEFINED, | |
276 | SANITIZE_UNDEFINED_NONDEFAULT, | |
277 | SANITIZE_POINTER_COMPARE, | |
278 | SANITIZE_POINTER_SUBTRACT | |
279 | }; | |
280 | ||
281 | for (unsigned i = 0; i < sizeof (codes) / sizeof (codes[0]); i++) | |
282 | if (sanitize_flags_p (codes[i], caller) | |
283 | != sanitize_flags_p (codes[i], callee)) | |
284 | return false; | |
285 | ||
286 | return true; | |
25a07c7e YG |
287 | } |
288 | ||
8e926cb1 JH |
289 | /* Used for flags where it is safe to inline when caller's value is |
290 | grater than callee's. */ | |
291 | #define check_maybe_up(flag) \ | |
292 | (opts_for_fn (caller->decl)->x_##flag \ | |
293 | != opts_for_fn (callee->decl)->x_##flag \ | |
294 | && (!always_inline \ | |
295 | || opts_for_fn (caller->decl)->x_##flag \ | |
296 | < opts_for_fn (callee->decl)->x_##flag)) | |
297 | /* Used for flags where it is safe to inline when caller's value is | |
298 | smaller than callee's. */ | |
299 | #define check_maybe_down(flag) \ | |
300 | (opts_for_fn (caller->decl)->x_##flag \ | |
301 | != opts_for_fn (callee->decl)->x_##flag \ | |
302 | && (!always_inline \ | |
303 | || opts_for_fn (caller->decl)->x_##flag \ | |
304 | > opts_for_fn (callee->decl)->x_##flag)) | |
305 | /* Used for flags where exact match is needed for correctness. */ | |
306 | #define check_match(flag) \ | |
307 | (opts_for_fn (caller->decl)->x_##flag \ | |
308 | != opts_for_fn (callee->decl)->x_##flag) | |
309 | ||
9a4841a3 | 310 | /* Decide if we can inline the edge and possibly update |
4c0f7679 JH |
311 | inline_failed reason. |
312 | We check whether inlining is possible at all and whether | |
313 | caller growth limits allow doing so. | |
314 | ||
9a4841a3 | 315 | if REPORT is true, output reason to the dump file. */ |
ca31b95f | 316 | |
61a05df1 | 317 | static bool |
09ce3660 | 318 | can_inline_edge_p (struct cgraph_edge *e, bool report, |
9a4841a3 | 319 | bool early = false) |
ca31b95f | 320 | { |
7ce7e4d4 JH |
321 | gcc_checking_assert (e->inline_failed); |
322 | ||
323 | if (cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR) | |
324 | { | |
325 | if (report) | |
326 | report_inline_failed_reason (e); | |
327 | return false; | |
328 | } | |
329 | ||
4c0f7679 | 330 | bool inlinable = true; |
a5b1779f | 331 | enum availability avail; |
a62bfab5 ML |
332 | cgraph_node *caller = (e->caller->inlined_to |
333 | ? e->caller->inlined_to : e->caller); | |
e6007a27 | 334 | cgraph_node *callee = e->callee->ultimate_alias_target (&avail, caller); |
ea99e0be | 335 | |
7ce7e4d4 | 336 | if (!callee->definition) |
4c0f7679 JH |
337 | { |
338 | e->inline_failed = CIF_BODY_NOT_AVAILABLE; | |
339 | inlinable = false; | |
340 | } | |
8a4a6d2e | 341 | if (!early && (!opt_for_fn (callee->decl, optimize) |
12b8cb2e | 342 | || !opt_for_fn (caller->decl, optimize))) |
29f1e2b1 JH |
343 | { |
344 | e->inline_failed = CIF_FUNCTION_NOT_OPTIMIZED; | |
345 | inlinable = false; | |
346 | } | |
1f26ac87 JM |
347 | else if (callee->calls_comdat_local) |
348 | { | |
349 | e->inline_failed = CIF_USES_COMDAT_LOCAL; | |
350 | inlinable = false; | |
351 | } | |
d52f5295 | 352 | else if (avail <= AVAIL_INTERPOSABLE) |
9de21a23 | 353 | { |
4c0f7679 | 354 | e->inline_failed = CIF_OVERWRITABLE; |
6957a6f6 | 355 | inlinable = false; |
9de21a23 | 356 | } |
1a0bf5e1 JH |
357 | /* All edges with call_stmt_cannot_inline_p should have inline_failed |
358 | initialized to one of FINAL_ERROR reasons. */ | |
89faf322 | 359 | else if (e->call_stmt_cannot_inline_p) |
1a0bf5e1 | 360 | gcc_unreachable (); |
4c0f7679 | 361 | /* Don't inline if the functions have different EH personalities. */ |
bb1e543c | 362 | else if (DECL_FUNCTION_PERSONALITY (caller->decl) |
67348ccc | 363 | && DECL_FUNCTION_PERSONALITY (callee->decl) |
bb1e543c | 364 | && (DECL_FUNCTION_PERSONALITY (caller->decl) |
67348ccc | 365 | != DECL_FUNCTION_PERSONALITY (callee->decl))) |
4c0f7679 JH |
366 | { |
367 | e->inline_failed = CIF_EH_PERSONALITY; | |
368 | inlinable = false; | |
369 | } | |
a7ff6e27 AH |
370 | /* TM pure functions should not be inlined into non-TM_pure |
371 | functions. */ | |
7ce7e4d4 | 372 | else if (is_tm_pure (callee->decl) && !is_tm_pure (caller->decl)) |
0a35513e AH |
373 | { |
374 | e->inline_failed = CIF_UNSPECIFIED; | |
375 | inlinable = false; | |
376 | } | |
09a2806f | 377 | /* Check compatibility of target optimization options. */ |
bb1e543c | 378 | else if (!targetm.target_option.can_inline_p (caller->decl, |
67348ccc | 379 | callee->decl)) |
4c0f7679 JH |
380 | { |
381 | e->inline_failed = CIF_TARGET_OPTION_MISMATCH; | |
382 | inlinable = false; | |
383 | } | |
56f62793 ML |
384 | else if (ipa_fn_summaries->get (callee) == NULL |
385 | || !ipa_fn_summaries->get (callee)->inlinable) | |
5058c037 JH |
386 | { |
387 | e->inline_failed = CIF_FUNCTION_NOT_INLINABLE; | |
388 | inlinable = false; | |
389 | } | |
25a07c7e | 390 | /* Don't inline a function with mismatched sanitization attributes. */ |
bb1e543c | 391 | else if (!sanitize_attrs_match_for_inline_p (caller->decl, callee->decl)) |
25a07c7e | 392 | { |
4089df8e | 393 | e->inline_failed = CIF_SANITIZE_ATTRIBUTE_MISMATCH; |
25a07c7e YG |
394 | inlinable = false; |
395 | } | |
9a4841a3 JH |
396 | if (!inlinable && report) |
397 | report_inline_failed_reason (e); | |
398 | return inlinable; | |
399 | } | |
400 | ||
2925cad2 JH |
401 | /* Return inlining_insns_single limit for function N. If HINT is true |
402 | scale up the bound. */ | |
562d1e95 JH |
403 | |
404 | static int | |
2925cad2 | 405 | inline_insns_single (cgraph_node *n, bool hint) |
562d1e95 | 406 | { |
1e83bd70 JH |
407 | if (hint) |
408 | return opt_for_fn (n->decl, param_max_inline_insns_single) | |
409 | * opt_for_fn (n->decl, param_inline_heuristics_hint_percent) / 100; | |
410 | return opt_for_fn (n->decl, param_max_inline_insns_single); | |
562d1e95 JH |
411 | } |
412 | ||
2925cad2 JH |
413 | /* Return inlining_insns_auto limit for function N. If HINT is true |
414 | scale up the bound. */ | |
562d1e95 JH |
415 | |
416 | static int | |
2925cad2 | 417 | inline_insns_auto (cgraph_node *n, bool hint) |
562d1e95 | 418 | { |
78a502ca ML |
419 | int max_inline_insns_auto = opt_for_fn (n->decl, param_max_inline_insns_auto); |
420 | if (hint) | |
1e83bd70 JH |
421 | return max_inline_insns_auto |
422 | * opt_for_fn (n->decl, param_inline_heuristics_hint_percent) / 100; | |
78a502ca | 423 | return max_inline_insns_auto; |
562d1e95 JH |
424 | } |
425 | ||
9a4841a3 JH |
426 | /* Decide if we can inline the edge and possibly update |
427 | inline_failed reason. | |
428 | We check whether inlining is possible at all and whether | |
429 | caller growth limits allow doing so. | |
430 | ||
431 | if REPORT is true, output reason to the dump file. | |
432 | ||
433 | if DISREGARD_LIMITS is true, ignore size limits. */ | |
434 | ||
435 | static bool | |
436 | can_inline_edge_by_limits_p (struct cgraph_edge *e, bool report, | |
437 | bool disregard_limits = false, bool early = false) | |
438 | { | |
439 | gcc_checking_assert (e->inline_failed); | |
440 | ||
441 | if (cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR) | |
442 | { | |
443 | if (report) | |
444 | report_inline_failed_reason (e); | |
445 | return false; | |
446 | } | |
447 | ||
448 | bool inlinable = true; | |
449 | enum availability avail; | |
a62bfab5 ML |
450 | cgraph_node *caller = (e->caller->inlined_to |
451 | ? e->caller->inlined_to : e->caller); | |
9a4841a3 JH |
452 | cgraph_node *callee = e->callee->ultimate_alias_target (&avail, caller); |
453 | tree caller_tree = DECL_FUNCTION_SPECIFIC_OPTIMIZATION (caller->decl); | |
454 | tree callee_tree | |
455 | = callee ? DECL_FUNCTION_SPECIFIC_OPTIMIZATION (callee->decl) : NULL; | |
4c0f7679 | 456 | /* Check if caller growth allows the inlining. */ |
9a4841a3 JH |
457 | if (!DECL_DISREGARD_INLINE_LIMITS (callee->decl) |
458 | && !disregard_limits | |
459 | && !lookup_attribute ("flatten", | |
460 | DECL_ATTRIBUTES (caller->decl)) | |
461 | && !caller_growth_limits (e)) | |
4c0f7679 | 462 | inlinable = false; |
bc53dee0 QZ |
463 | else if (callee->externally_visible |
464 | && !DECL_DISREGARD_INLINE_LIMITS (callee->decl) | |
465 | && flag_live_patching == LIVE_PATCHING_INLINE_ONLY_STATIC) | |
466 | { | |
467 | e->inline_failed = CIF_EXTERN_LIVE_ONLY_STATIC; | |
468 | inlinable = false; | |
469 | } | |
4c0f7679 JH |
470 | /* Don't inline a function with a higher optimization level than the |
471 | caller. FIXME: this is really just tip of iceberg of handling | |
472 | optimization attribute. */ | |
473 | else if (caller_tree != callee_tree) | |
9de21a23 | 474 | { |
8e926cb1 JH |
475 | bool always_inline = |
476 | (DECL_DISREGARD_INLINE_LIMITS (callee->decl) | |
477 | && lookup_attribute ("always_inline", | |
478 | DECL_ATTRIBUTES (callee->decl))); | |
56f62793 ML |
479 | ipa_fn_summary *caller_info = ipa_fn_summaries->get (caller); |
480 | ipa_fn_summary *callee_info = ipa_fn_summaries->get (callee); | |
8e926cb1 | 481 | |
553bb257 KT |
482 | /* Until GCC 4.9 we did not check the semantics-altering flags |
483 | below and inlined across optimization boundaries. | |
484 | Enabling checks below breaks several packages by refusing | |
ddb3773a JH |
485 | to inline library always_inline functions. See PR65873. |
486 | Disable the check for early inlining for now until better solution | |
487 | is found. */ | |
488 | if (always_inline && early) | |
489 | ; | |
0c3068e0 RB |
490 | /* There are some options that change IL semantics which means |
491 | we cannot inline in these cases for correctness reason. | |
492 | Not even for always_inline declared functions. */ | |
2bf54d93 | 493 | else if (check_match (flag_wrapv) |
ddb3773a | 494 | || check_match (flag_trapv) |
29f1e2b1 | 495 | || check_match (flag_pcc_struct_return) |
2523d721 | 496 | || check_maybe_down (optimize_debug) |
818b88a7 JH |
497 | /* When caller or callee does FP math, be sure FP codegen flags |
498 | compatible. */ | |
499 | || ((caller_info->fp_expressions && callee_info->fp_expressions) | |
500 | && (check_maybe_up (flag_rounding_math) | |
501 | || check_maybe_up (flag_trapping_math) | |
502 | || check_maybe_down (flag_unsafe_math_optimizations) | |
503 | || check_maybe_down (flag_finite_math_only) | |
504 | || check_maybe_up (flag_signaling_nans) | |
505 | || check_maybe_down (flag_cx_limited_range) | |
506 | || check_maybe_up (flag_signed_zeros) | |
507 | || check_maybe_down (flag_associative_math) | |
508 | || check_maybe_down (flag_reciprocal_math) | |
0d2f700f | 509 | || check_maybe_down (flag_fp_int_builtin_inexact) |
818b88a7 JH |
510 | /* Strictly speaking only when the callee contains function |
511 | calls that may end up setting errno. */ | |
512 | || check_maybe_up (flag_errno_math))) | |
ddb3773a JH |
513 | /* We do not want to make code compiled with exceptions to be |
514 | brought into a non-EH function unless we know that the callee | |
515 | does not throw. | |
516 | This is tracked by DECL_FUNCTION_PERSONALITY. */ | |
6a96e917 EB |
517 | || (check_maybe_up (flag_non_call_exceptions) |
518 | && DECL_FUNCTION_PERSONALITY (callee->decl)) | |
ddb3773a JH |
519 | || (check_maybe_up (flag_exceptions) |
520 | && DECL_FUNCTION_PERSONALITY (callee->decl)) | |
956d615d | 521 | /* When devirtualization is disabled for callee, it is not safe |
ddb3773a JH |
522 | to inline it as we possibly mangled the type info. |
523 | Allow early inlining of always inlines. */ | |
524 | || (!early && check_maybe_down (flag_devirtualize))) | |
0c3068e0 RB |
525 | { |
526 | e->inline_failed = CIF_OPTIMIZATION_MISMATCH; | |
527 | inlinable = false; | |
528 | } | |
529 | /* gcc.dg/pr43564.c. Apply user-forced inline even at -O0. */ | |
8e926cb1 | 530 | else if (always_inline) |
bb1e543c | 531 | ; |
0c3068e0 RB |
532 | /* When user added an attribute to the callee honor it. */ |
533 | else if (lookup_attribute ("optimize", DECL_ATTRIBUTES (callee->decl)) | |
534 | && opts_for_fn (caller->decl) != opts_for_fn (callee->decl)) | |
4c0f7679 | 535 | { |
fd811f03 | 536 | e->inline_failed = CIF_OPTIMIZATION_MISMATCH; |
4c0f7679 JH |
537 | inlinable = false; |
538 | } | |
ddb3773a JH |
539 | /* If explicit optimize attribute are not used, the mismatch is caused |
540 | by different command line options used to build different units. | |
541 | Do not care about COMDAT functions - those are intended to be | |
542 | optimized with the optimization flags of module they are used in. | |
543 | Also do not care about mixing up size/speed optimization when | |
544 | DECL_DISREGARD_INLINE_LIMITS is set. */ | |
88636b62 | 545 | else if ((callee->merged_comdat |
ddb3773a JH |
546 | && !lookup_attribute ("optimize", |
547 | DECL_ATTRIBUTES (caller->decl))) | |
548 | || DECL_DISREGARD_INLINE_LIMITS (callee->decl)) | |
549 | ; | |
bb1e543c | 550 | /* If mismatch is caused by merging two LTO units with different |
956d615d | 551 | optimization flags we want to be bit nicer. However never inline |
bb1e543c JH |
552 | if one of functions is not optimized at all. */ |
553 | else if (!opt_for_fn (callee->decl, optimize) | |
554 | || !opt_for_fn (caller->decl, optimize)) | |
555 | { | |
556 | e->inline_failed = CIF_OPTIMIZATION_MISMATCH; | |
557 | inlinable = false; | |
558 | } | |
559 | /* If callee is optimized for size and caller is not, allow inlining if | |
028d4092 ML |
560 | code shrinks or we are in param_max_inline_insns_single limit and |
561 | callee is inline (and thus likely an unified comdat). | |
562 | This will allow caller to run faster. */ | |
bb1e543c JH |
563 | else if (opt_for_fn (callee->decl, optimize_size) |
564 | > opt_for_fn (caller->decl, optimize_size)) | |
565 | { | |
566 | int growth = estimate_edge_growth (e); | |
1e83bd70 | 567 | if (growth > opt_for_fn (caller->decl, param_max_inline_insns_size) |
bb1e543c | 568 | && (!DECL_DECLARED_INLINE_P (callee->decl) |
2925cad2 JH |
569 | && growth >= MAX (inline_insns_single (caller, false), |
570 | inline_insns_auto (caller, false)))) | |
bb1e543c JH |
571 | { |
572 | e->inline_failed = CIF_OPTIMIZATION_MISMATCH; | |
573 | inlinable = false; | |
574 | } | |
575 | } | |
576 | /* If callee is more aggressively optimized for performance than caller, | |
577 | we generally want to inline only cheap (runtime wise) functions. */ | |
578 | else if (opt_for_fn (callee->decl, optimize_size) | |
579 | < opt_for_fn (caller->decl, optimize_size) | |
580 | || (opt_for_fn (callee->decl, optimize) | |
86f46e39 | 581 | > opt_for_fn (caller->decl, optimize))) |
bb1e543c JH |
582 | { |
583 | if (estimate_edge_time (e) | |
56f62793 | 584 | >= 20 + ipa_call_summaries->get (e)->call_stmt_time) |
bb1e543c JH |
585 | { |
586 | e->inline_failed = CIF_OPTIMIZATION_MISMATCH; | |
587 | inlinable = false; | |
588 | } | |
589 | } | |
590 | ||
4c0f7679 JH |
591 | } |
592 | ||
4c0f7679 JH |
593 | if (!inlinable && report) |
594 | report_inline_failed_reason (e); | |
595 | return inlinable; | |
596 | } | |
597 | ||
598 | ||
599 | /* Return true if the edge E is inlinable during early inlining. */ | |
600 | ||
601 | static bool | |
602 | can_early_inline_edge_p (struct cgraph_edge *e) | |
603 | { | |
d52f5295 | 604 | struct cgraph_node *callee = e->callee->ultimate_alias_target (); |
4c0f7679 | 605 | /* Early inliner might get called at WPA stage when IPA pass adds new |
67914693 | 606 | function. In this case we cannot really do any of early inlining |
4c0f7679 | 607 | because function bodies are missing. */ |
1a0bf5e1 JH |
608 | if (cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR) |
609 | return false; | |
67348ccc | 610 | if (!gimple_has_body_p (callee->decl)) |
4c0f7679 JH |
611 | { |
612 | e->inline_failed = CIF_BODY_NOT_AVAILABLE; | |
9de21a23 JC |
613 | return false; |
614 | } | |
4c0f7679 JH |
615 | /* In early inliner some of callees may not be in SSA form yet |
616 | (i.e. the callgraph is cyclic and we did not process | |
617 | the callee by early inliner, yet). We don't have CIF code for this | |
618 | case; later we will re-do the decision in the real inliner. */ | |
67348ccc DM |
619 | if (!gimple_in_ssa_p (DECL_STRUCT_FUNCTION (e->caller->decl)) |
620 | || !gimple_in_ssa_p (DECL_STRUCT_FUNCTION (callee->decl))) | |
f27e50db | 621 | { |
4174a33a DM |
622 | if (dump_enabled_p ()) |
623 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt, | |
624 | " edge not inlinable: not in SSA form\n"); | |
f27e50db JH |
625 | return false; |
626 | } | |
9a4841a3 JH |
627 | if (!can_inline_edge_p (e, true, true) |
628 | || !can_inline_edge_by_limits_p (e, true, false, true)) | |
4c0f7679 JH |
629 | return false; |
630 | return true; | |
631 | } | |
632 | ||
633 | ||
ae6e6a08 | 634 | /* Return number of calls in N. Ignore cheap builtins. */ |
4c0f7679 | 635 | |
ae6e6a08 JH |
636 | static int |
637 | num_calls (struct cgraph_node *n) | |
4c0f7679 JH |
638 | { |
639 | struct cgraph_edge *e; | |
ae6e6a08 JH |
640 | int num = 0; |
641 | ||
4c0f7679 | 642 | for (e = n->callees; e; e = e->next_callee) |
67348ccc | 643 | if (!is_inexpensive_builtin (e->callee->decl)) |
ae6e6a08 JH |
644 | num++; |
645 | return num; | |
4c0f7679 JH |
646 | } |
647 | ||
f27e50db | 648 | |
4c0f7679 | 649 | /* Return true if we are interested in inlining small function. */ |
9de21a23 | 650 | |
4c0f7679 JH |
651 | static bool |
652 | want_early_inline_function_p (struct cgraph_edge *e) | |
653 | { | |
654 | bool want_inline = true; | |
d52f5295 | 655 | struct cgraph_node *callee = e->callee->ultimate_alias_target (); |
4c0f7679 | 656 | |
67348ccc | 657 | if (DECL_DISREGARD_INLINE_LIMITS (callee->decl)) |
4c0f7679 | 658 | ; |
9a1e784a | 659 | /* For AutoFDO, we need to make sure that before profile summary, all |
be3c16c4 DC |
660 | hot paths' IR look exactly the same as profiled binary. As a result, |
661 | in einliner, we will disregard size limit and inline those callsites | |
662 | that are: | |
663 | * inlined in the profiled binary, and | |
664 | * the cloned callee has enough samples to be considered "hot". */ | |
665 | else if (flag_auto_profile && afdo_callsite_hot_enough_for_early_inline (e)) | |
666 | ; | |
67348ccc | 667 | else if (!DECL_DECLARED_INLINE_P (callee->decl) |
2bf86c84 | 668 | && !opt_for_fn (e->caller->decl, flag_inline_small_functions)) |
4c0f7679 JH |
669 | { |
670 | e->inline_failed = CIF_FUNCTION_NOT_INLINE_CANDIDATE; | |
671 | report_inline_failed_reason (e); | |
672 | want_inline = false; | |
673 | } | |
674 | else | |
9de21a23 | 675 | { |
49e26500 JH |
676 | /* First take care of very large functions. */ |
677 | int min_growth = estimate_min_edge_growth (e), growth = 0; | |
ae6e6a08 | 678 | int n; |
1e83bd70 | 679 | int early_inlining_insns = param_early_inlining_insns; |
0b92cf30 | 680 | |
49e26500 JH |
681 | if (min_growth > early_inlining_insns) |
682 | { | |
683 | if (dump_enabled_p ()) | |
684 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt, | |
685 | " will not early inline: %C->%C, " | |
686 | "call is cold and code would grow " | |
687 | "at least by %i\n", | |
688 | e->caller, callee, | |
689 | min_growth); | |
690 | want_inline = false; | |
691 | } | |
692 | else | |
693 | growth = estimate_edge_growth (e); | |
694 | ||
ae6e6a08 | 695 | |
49e26500 | 696 | if (!want_inline || growth <= param_max_inline_insns_size) |
4c0f7679 | 697 | ; |
f256c274 | 698 | else if (!e->maybe_hot_p ()) |
4c0f7679 | 699 | { |
4174a33a DM |
700 | if (dump_enabled_p ()) |
701 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt, | |
702 | " will not early inline: %C->%C, " | |
703 | "call is cold and code would grow by %i\n", | |
704 | e->caller, callee, | |
705 | growth); | |
4c0f7679 JH |
706 | want_inline = false; |
707 | } | |
0b92cf30 | 708 | else if (growth > early_inlining_insns) |
9de21a23 | 709 | { |
4174a33a DM |
710 | if (dump_enabled_p ()) |
711 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt, | |
712 | " will not early inline: %C->%C, " | |
9c28689a JH |
713 | "growth %i exceeds --param early-inlining-insns\n", |
714 | e->caller, callee, growth); | |
4c0f7679 | 715 | want_inline = false; |
9de21a23 | 716 | } |
ae6e6a08 | 717 | else if ((n = num_calls (callee)) != 0 |
0b92cf30 | 718 | && growth * (n + 1) > early_inlining_insns) |
4c0f7679 | 719 | { |
4174a33a DM |
720 | if (dump_enabled_p ()) |
721 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt, | |
722 | " will not early inline: %C->%C, " | |
9c28689a | 723 | "growth %i exceeds --param early-inlining-insns " |
4174a33a | 724 | "divided by number of calls\n", |
9c28689a | 725 | e->caller, callee, growth); |
4c0f7679 JH |
726 | want_inline = false; |
727 | } | |
728 | } | |
729 | return want_inline; | |
730 | } | |
731 | ||
d59171da JH |
732 | /* Compute time of the edge->caller + edge->callee execution when inlining |
733 | does not happen. */ | |
734 | ||
6d4ab5f8 | 735 | inline sreal |
4adaad64 | 736 | compute_uninlined_call_time (struct cgraph_edge *edge, |
b5351388 JH |
737 | sreal uninlined_call_time, |
738 | sreal freq) | |
d59171da | 739 | { |
a62bfab5 ML |
740 | cgraph_node *caller = (edge->caller->inlined_to |
741 | ? edge->caller->inlined_to | |
208e5afa JH |
742 | : edge->caller); |
743 | ||
0009a6c3 | 744 | if (freq > 0) |
41f0e819 | 745 | uninlined_call_time *= freq; |
208e5afa JH |
746 | else |
747 | uninlined_call_time = uninlined_call_time >> 11; | |
748 | ||
56f62793 | 749 | sreal caller_time = ipa_fn_summaries->get (caller)->time; |
d59171da JH |
750 | return uninlined_call_time + caller_time; |
751 | } | |
752 | ||
753 | /* Same as compute_uinlined_call_time but compute time when inlining | |
754 | does happen. */ | |
755 | ||
6d4ab5f8 | 756 | inline sreal |
d59171da | 757 | compute_inlined_call_time (struct cgraph_edge *edge, |
b5351388 JH |
758 | sreal time, |
759 | sreal freq) | |
d59171da | 760 | { |
a62bfab5 ML |
761 | cgraph_node *caller = (edge->caller->inlined_to |
762 | ? edge->caller->inlined_to | |
208e5afa | 763 | : edge->caller); |
56f62793 | 764 | sreal caller_time = ipa_fn_summaries->get (caller)->time; |
208e5afa | 765 | |
0009a6c3 | 766 | if (freq > 0) |
41f0e819 | 767 | time *= freq; |
208e5afa JH |
768 | else |
769 | time = time >> 11; | |
770 | ||
4adaad64 JH |
771 | /* This calculation should match one in ipa-inline-analysis.c |
772 | (estimate_edge_size_and_time). */ | |
56f62793 | 773 | time -= (sreal)ipa_call_summaries->get (edge)->call_stmt_time * freq; |
208e5afa JH |
774 | time += caller_time; |
775 | if (time <= 0) | |
776 | time = ((sreal) 1) >> 8; | |
6d4ab5f8 | 777 | gcc_checking_assert (time >= 0); |
d59171da JH |
778 | return time; |
779 | } | |
780 | ||
956d615d JJ |
781 | /* Determine time saved by inlining EDGE of frequency FREQ |
782 | where callee's runtime w/o inlining is UNINLINED_TYPE | |
ea8dd3b6 JH |
783 | and with inlined is INLINED_TYPE. */ |
784 | ||
785 | inline sreal | |
786 | inlining_speedup (struct cgraph_edge *edge, | |
787 | sreal freq, | |
788 | sreal uninlined_time, | |
789 | sreal inlined_time) | |
790 | { | |
791 | sreal speedup = uninlined_time - inlined_time; | |
792 | /* Handling of call_time should match one in ipa-inline-fnsummary.c | |
793 | (estimate_edge_size_and_time). */ | |
794 | sreal call_time = ipa_call_summaries->get (edge)->call_stmt_time; | |
795 | ||
796 | if (freq > 0) | |
797 | { | |
798 | speedup = (speedup + call_time); | |
799 | if (freq != 1) | |
800 | speedup = speedup * freq; | |
801 | } | |
802 | else if (freq == 0) | |
803 | speedup = speedup >> 11; | |
804 | gcc_checking_assert (speedup >= 0); | |
805 | return speedup; | |
806 | } | |
807 | ||
42f7b0fa JH |
808 | /* Return true if the speedup for inlining E is bigger than |
809 | PARAM_MAX_INLINE_MIN_SPEEDUP. */ | |
810 | ||
811 | static bool | |
812 | big_speedup_p (struct cgraph_edge *e) | |
813 | { | |
4adaad64 JH |
814 | sreal unspec_time; |
815 | sreal spec_time = estimate_edge_time (e, &unspec_time); | |
b5351388 JH |
816 | sreal freq = e->sreal_frequency (); |
817 | sreal time = compute_uninlined_call_time (e, unspec_time, freq); | |
818 | sreal inlined_time = compute_inlined_call_time (e, spec_time, freq); | |
a62bfab5 ML |
819 | cgraph_node *caller = (e->caller->inlined_to |
820 | ? e->caller->inlined_to | |
821 | : e->caller); | |
1e83bd70 | 822 | int limit = opt_for_fn (caller->decl, param_inline_min_speedup); |
208e5afa | 823 | |
562d1e95 | 824 | if ((time - inlined_time) * 100 > time * limit) |
42f7b0fa JH |
825 | return true; |
826 | return false; | |
827 | } | |
828 | ||
4c0f7679 JH |
829 | /* Return true if we are interested in inlining small function. |
830 | When REPORT is true, report reason to dump file. */ | |
831 | ||
832 | static bool | |
833 | want_inline_small_function_p (struct cgraph_edge *e, bool report) | |
834 | { | |
835 | bool want_inline = true; | |
d52f5295 | 836 | struct cgraph_node *callee = e->callee->ultimate_alias_target (); |
1e83bd70 JH |
837 | cgraph_node *to = (e->caller->inlined_to |
838 | ? e->caller->inlined_to : e->caller); | |
4c0f7679 | 839 | |
01b9bf06 RS |
840 | /* Allow this function to be called before can_inline_edge_p, |
841 | since it's usually cheaper. */ | |
842 | if (cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR) | |
843 | want_inline = false; | |
844 | else if (DECL_DISREGARD_INLINE_LIMITS (callee->decl)) | |
4c0f7679 | 845 | ; |
67348ccc | 846 | else if (!DECL_DECLARED_INLINE_P (callee->decl) |
2bf86c84 | 847 | && !opt_for_fn (e->caller->decl, flag_inline_small_functions)) |
4c0f7679 JH |
848 | { |
849 | e->inline_failed = CIF_FUNCTION_NOT_INLINE_CANDIDATE; | |
850 | want_inline = false; | |
9de21a23 | 851 | } |
4cd8957f | 852 | /* Do fast and conservative check if the function can be good |
2925cad2 | 853 | inline candidate. */ |
b6d627e4 | 854 | else if ((!DECL_DECLARED_INLINE_P (callee->decl) |
1bad9c18 | 855 | && (!e->count.ipa ().initialized_p () || !e->maybe_hot_p ())) |
56f62793 ML |
856 | && ipa_fn_summaries->get (callee)->min_size |
857 | - ipa_call_summaries->get (e)->call_stmt_size | |
2925cad2 | 858 | > inline_insns_auto (e->caller, true)) |
4cd8957f | 859 | { |
afeb8875 | 860 | e->inline_failed = CIF_MAX_INLINE_INSNS_AUTO_LIMIT; |
4cd8957f JH |
861 | want_inline = false; |
862 | } | |
3995f3a2 | 863 | else if ((DECL_DECLARED_INLINE_P (callee->decl) |
1bad9c18 | 864 | || e->count.ipa ().nonzero_p ()) |
56f62793 ML |
865 | && ipa_fn_summaries->get (callee)->min_size |
866 | - ipa_call_summaries->get (e)->call_stmt_size | |
2925cad2 | 867 | > inline_insns_single (e->caller, true)) |
4cd8957f | 868 | { |
9c28689a JH |
869 | e->inline_failed = (DECL_DECLARED_INLINE_P (callee->decl) |
870 | ? CIF_MAX_INLINE_INSNS_SINGLE_LIMIT | |
871 | : CIF_MAX_INLINE_INSNS_AUTO_LIMIT); | |
4cd8957f JH |
872 | want_inline = false; |
873 | } | |
ca31b95f | 874 | else |
9de21a23 | 875 | { |
4c0f7679 | 876 | int growth = estimate_edge_growth (e); |
0bceb671 | 877 | ipa_hints hints = estimate_edge_hints (e); |
2925cad2 JH |
878 | bool apply_hints = (hints & (INLINE_HINT_indirect_call |
879 | | INLINE_HINT_known_hot | |
880 | | INLINE_HINT_loop_iterations | |
881 | | INLINE_HINT_loop_stride)); | |
4c0f7679 | 882 | |
1e83bd70 JH |
883 | if (growth <= opt_for_fn (to->decl, |
884 | param_max_inline_insns_size)) | |
4c0f7679 | 885 | ; |
028d4092 | 886 | /* Apply param_max_inline_insns_single limit. Do not do so when |
2925cad2 JH |
887 | hints suggests that inlining given function is very profitable. |
888 | Avoid computation of big_speedup_p when not necessary to change | |
889 | outcome of decision. */ | |
67348ccc | 890 | else if (DECL_DECLARED_INLINE_P (callee->decl) |
2925cad2 JH |
891 | && growth >= inline_insns_single (e->caller, apply_hints) |
892 | && (apply_hints | |
893 | || growth >= inline_insns_single (e->caller, true) | |
894 | || !big_speedup_p (e))) | |
4c0f7679 | 895 | { |
1e83bd70 | 896 | e->inline_failed = CIF_MAX_INLINE_INSNS_SINGLE_LIMIT; |
4c0f7679 JH |
897 | want_inline = false; |
898 | } | |
67348ccc | 899 | else if (!DECL_DECLARED_INLINE_P (callee->decl) |
f256c274 | 900 | && !opt_for_fn (e->caller->decl, flag_inline_functions) |
1e83bd70 JH |
901 | && growth >= opt_for_fn (to->decl, |
902 | param_max_inline_insns_small)) | |
4c0f7679 | 903 | { |
49d9c9d2 | 904 | /* growth_positive_p is expensive, always test it last. */ |
2925cad2 | 905 | if (growth >= inline_insns_single (e->caller, false) |
49d9c9d2 | 906 | || growth_positive_p (callee, e, growth)) |
4cd8957f | 907 | { |
562d1e95 | 908 | e->inline_failed = CIF_NOT_DECLARED_INLINED; |
4cd8957f | 909 | want_inline = false; |
562d1e95 | 910 | } |
4c0f7679 | 911 | } |
028d4092 ML |
912 | /* Apply param_max_inline_insns_auto limit for functions not declared |
913 | inline. Bypass the limit when speedup seems big. */ | |
67348ccc | 914 | else if (!DECL_DECLARED_INLINE_P (callee->decl) |
2925cad2 JH |
915 | && growth >= inline_insns_auto (e->caller, apply_hints) |
916 | && (apply_hints | |
917 | || growth >= inline_insns_auto (e->caller, true) | |
918 | || !big_speedup_p (e))) | |
4c0f7679 | 919 | { |
49d9c9d2 | 920 | /* growth_positive_p is expensive, always test it last. */ |
2925cad2 | 921 | if (growth >= inline_insns_single (e->caller, false) |
49d9c9d2 | 922 | || growth_positive_p (callee, e, growth)) |
4cd8957f | 923 | { |
afeb8875 | 924 | e->inline_failed = CIF_MAX_INLINE_INSNS_AUTO_LIMIT; |
4cd8957f | 925 | want_inline = false; |
562d1e95 | 926 | } |
4c0f7679 | 927 | } |
db22a743 | 928 | /* If call is cold, do not inline when function body would grow. */ |
3dafb85c | 929 | else if (!e->maybe_hot_p () |
2925cad2 | 930 | && (growth >= inline_insns_single (e->caller, false) |
49d9c9d2 | 931 | || growth_positive_p (callee, e, growth))) |
9de21a23 | 932 | { |
562d1e95 | 933 | e->inline_failed = CIF_UNLIKELY_CALL; |
4c0f7679 | 934 | want_inline = false; |
9de21a23 JC |
935 | } |
936 | } | |
4c0f7679 JH |
937 | if (!want_inline && report) |
938 | report_inline_failed_reason (e); | |
939 | return want_inline; | |
940 | } | |
9de21a23 | 941 | |
4c0f7679 | 942 | /* EDGE is self recursive edge. |
1e83bd70 | 943 | We handle two cases - when function A is inlining into itself |
4c0f7679 JH |
944 | or when function A is being inlined into another inliner copy of function |
945 | A within function B. | |
946 | ||
947 | In first case OUTER_NODE points to the toplevel copy of A, while | |
948 | in the second case OUTER_NODE points to the outermost copy of A in B. | |
949 | ||
950 | In both cases we want to be extra selective since | |
951 | inlining the call will just introduce new recursive calls to appear. */ | |
09a2806f | 952 | |
4c0f7679 JH |
953 | static bool |
954 | want_inline_self_recursive_call_p (struct cgraph_edge *edge, | |
955 | struct cgraph_node *outer_node, | |
956 | bool peeling, | |
957 | int depth) | |
958 | { | |
959 | char const *reason = NULL; | |
960 | bool want_inline = true; | |
0009a6c3 | 961 | sreal caller_freq = 1; |
1e83bd70 JH |
962 | int max_depth = opt_for_fn (outer_node->decl, |
963 | param_max_inline_recursive_depth_auto); | |
4c0f7679 | 964 | |
67348ccc | 965 | if (DECL_DECLARED_INLINE_P (edge->caller->decl)) |
1e83bd70 JH |
966 | max_depth = opt_for_fn (outer_node->decl, |
967 | param_max_inline_recursive_depth); | |
4c0f7679 | 968 | |
3dafb85c | 969 | if (!edge->maybe_hot_p ()) |
4c0f7679 JH |
970 | { |
971 | reason = "recursive call is cold"; | |
972 | want_inline = false; | |
973 | } | |
4c0f7679 JH |
974 | else if (depth > max_depth) |
975 | { | |
976 | reason = "--param max-inline-recursive-depth exceeded."; | |
977 | want_inline = false; | |
978 | } | |
a62bfab5 | 979 | else if (outer_node->inlined_to |
0009a6c3 | 980 | && (caller_freq = outer_node->callers->sreal_frequency ()) == 0) |
bd936951 | 981 | { |
0009a6c3 | 982 | reason = "caller frequency is 0"; |
bd936951 JH |
983 | want_inline = false; |
984 | } | |
985 | ||
4c0f7679 JH |
986 | if (!want_inline) |
987 | ; | |
0009a6c3 JH |
988 | /* Inlining of self recursive function into copy of itself within other |
989 | function is transformation similar to loop peeling. | |
4c0f7679 | 990 | |
09a2806f | 991 | Peeling is profitable if we can inline enough copies to make probability |
4c0f7679 JH |
992 | of actual call to the self recursive function very small. Be sure that |
993 | the probability of recursion is small. | |
994 | ||
09a2806f | 995 | We ensure that the frequency of recursing is at most 1 - (1/max_depth). |
0009a6c3 | 996 | This way the expected number of recursion is at most max_depth. */ |
4c0f7679 JH |
997 | else if (peeling) |
998 | { | |
0009a6c3 | 999 | sreal max_prob = (sreal)1 - ((sreal)1 / (sreal)max_depth); |
4c0f7679 JH |
1000 | int i; |
1001 | for (i = 1; i < depth; i++) | |
0009a6c3 JH |
1002 | max_prob = max_prob * max_prob; |
1003 | if (edge->sreal_frequency () >= max_prob * caller_freq) | |
4c0f7679 JH |
1004 | { |
1005 | reason = "frequency of recursive call is too large"; | |
1006 | want_inline = false; | |
1007 | } | |
1008 | } | |
0009a6c3 JH |
1009 | /* Recursive inlining, i.e. equivalent of unrolling, is profitable if |
1010 | recursion depth is large. We reduce function call overhead and increase | |
1011 | chances that things fit in hardware return predictor. | |
4c0f7679 JH |
1012 | |
1013 | Recursive inlining might however increase cost of stack frame setup | |
1014 | actually slowing down functions whose recursion tree is wide rather than | |
1015 | deep. | |
1016 | ||
09a2806f | 1017 | Deciding reliably on when to do recursive inlining without profile feedback |
4c0f7679 JH |
1018 | is tricky. For now we disable recursive inlining when probability of self |
1019 | recursion is low. | |
1020 | ||
0009a6c3 JH |
1021 | Recursive inlining of self recursive call within loop also results in |
1022 | large loop depths that generally optimize badly. We may want to throttle | |
1023 | down inlining in those cases. In particular this seems to happen in one | |
1024 | of libstdc++ rb tree methods. */ | |
4c0f7679 JH |
1025 | else |
1026 | { | |
0009a6c3 JH |
1027 | if (edge->sreal_frequency () * 100 |
1028 | <= caller_freq | |
1e83bd70 JH |
1029 | * opt_for_fn (outer_node->decl, |
1030 | param_min_inline_recursive_probability)) | |
4c0f7679 JH |
1031 | { |
1032 | reason = "frequency of recursive call is too small"; | |
1033 | want_inline = false; | |
1034 | } | |
1035 | } | |
4174a33a DM |
1036 | if (!want_inline && dump_enabled_p ()) |
1037 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, edge->call_stmt, | |
1038 | " not inlining recursively: %s\n", reason); | |
4c0f7679 | 1039 | return want_inline; |
ca31b95f JH |
1040 | } |
1041 | ||
19ba6aab JH |
1042 | /* Return true when NODE has uninlinable caller; |
1043 | set HAS_HOT_CALL if it has hot call. | |
9aa3f5c5 JH |
1044 | Worker for cgraph_for_node_and_aliases. */ |
1045 | ||
1046 | static bool | |
19ba6aab | 1047 | check_callers (struct cgraph_node *node, void *has_hot_call) |
9aa3f5c5 | 1048 | { |
19ba6aab JH |
1049 | struct cgraph_edge *e; |
1050 | for (e = node->callers; e; e = e->next_caller) | |
1051 | { | |
29f1e2b1 JH |
1052 | if (!opt_for_fn (e->caller->decl, flag_inline_functions_called_once) |
1053 | || !opt_for_fn (e->caller->decl, optimize)) | |
2bf86c84 | 1054 | return true; |
19ba6aab JH |
1055 | if (!can_inline_edge_p (e, true)) |
1056 | return true; | |
1af8bfe5 JH |
1057 | if (e->recursive_p ()) |
1058 | return true; | |
9a4841a3 JH |
1059 | if (!can_inline_edge_by_limits_p (e, true)) |
1060 | return true; | |
3dafb85c | 1061 | if (!(*(bool *)has_hot_call) && e->maybe_hot_p ()) |
19ba6aab JH |
1062 | *(bool *)has_hot_call = true; |
1063 | } | |
1064 | return false; | |
9aa3f5c5 JH |
1065 | } |
1066 | ||
a81b0a3d JH |
1067 | /* If NODE has a caller, return true. */ |
1068 | ||
1069 | static bool | |
1070 | has_caller_p (struct cgraph_node *node, void *data ATTRIBUTE_UNUSED) | |
1071 | { | |
1072 | if (node->callers) | |
1073 | return true; | |
1074 | return false; | |
1075 | } | |
09a2806f | 1076 | |
100411f8 JH |
1077 | /* Decide if inlining NODE would reduce unit size by eliminating |
1078 | the offline copy of function. | |
1079 | When COLD is true the cold calls are considered, too. */ | |
09a2806f JH |
1080 | |
1081 | static bool | |
100411f8 | 1082 | want_inline_function_to_all_callers_p (struct cgraph_node *node, bool cold) |
09a2806f | 1083 | { |
5970b079 EB |
1084 | bool has_hot_call = false; |
1085 | ||
9789b553 JH |
1086 | /* Aliases gets inlined along with the function they alias. */ |
1087 | if (node->alias) | |
5970b079 EB |
1088 | return false; |
1089 | /* Already inlined? */ | |
a62bfab5 | 1090 | if (node->inlined_to) |
5970b079 EB |
1091 | return false; |
1092 | /* Does it have callers? */ | |
1ede94c5 | 1093 | if (!node->call_for_symbol_and_aliases (has_caller_p, NULL, true)) |
5970b079 EB |
1094 | return false; |
1095 | /* Inlining into all callers would increase size? */ | |
49d9c9d2 | 1096 | if (growth_positive_p (node, NULL, INT_MIN) > 0) |
5970b079 EB |
1097 | return false; |
1098 | /* All inlines must be possible. */ | |
1ede94c5 JH |
1099 | if (node->call_for_symbol_and_aliases (check_callers, &has_hot_call, |
1100 | true)) | |
5970b079 EB |
1101 | return false; |
1102 | if (!cold && !has_hot_call) | |
1103 | return false; | |
1104 | return true; | |
09a2806f JH |
1105 | } |
1106 | ||
041cb615 JH |
1107 | /* Return true if WHERE of SIZE is a possible candidate for wrapper heuristics |
1108 | in estimate_edge_badness. */ | |
1109 | ||
1110 | static bool | |
1111 | wrapper_heuristics_may_apply (struct cgraph_node *where, int size) | |
1112 | { | |
1113 | return size < (DECL_DECLARED_INLINE_P (where->decl) | |
1114 | ? inline_insns_single (where, false) | |
1115 | : inline_insns_auto (where, false)); | |
1116 | } | |
1117 | ||
670cd5c5 JH |
1118 | /* A cost model driving the inlining heuristics in a way so the edges with |
1119 | smallest badness are inlined first. After each inlining is performed | |
0fa2e4df | 1120 | the costs of all caller edges of nodes affected are recomputed so the |
670cd5c5 | 1121 | metrics may accurately depend on values such as number of inlinable callers |
45a80bb9 | 1122 | of the function or function body size. */ |
670cd5c5 | 1123 | |
f0e1509b | 1124 | static sreal |
4c0f7679 | 1125 | edge_badness (struct cgraph_edge *edge, bool dump) |
670cd5c5 | 1126 | { |
f0e1509b | 1127 | sreal badness; |
ab38481c | 1128 | int growth; |
4adaad64 | 1129 | sreal edge_time, unspec_edge_time; |
d52f5295 | 1130 | struct cgraph_node *callee = edge->callee->ultimate_alias_target (); |
99b1c316 | 1131 | class ipa_fn_summary *callee_info = ipa_fn_summaries->get (callee); |
0bceb671 | 1132 | ipa_hints hints; |
a62bfab5 ML |
1133 | cgraph_node *caller = (edge->caller->inlined_to |
1134 | ? edge->caller->inlined_to | |
208e5afa | 1135 | : edge->caller); |
1aa14195 | 1136 | |
03dfc36d | 1137 | growth = estimate_edge_growth (edge); |
4adaad64 | 1138 | edge_time = estimate_edge_time (edge, &unspec_edge_time); |
37678631 | 1139 | hints = estimate_edge_hints (edge); |
d59171da | 1140 | gcc_checking_assert (edge_time >= 0); |
1bad9c18 JH |
1141 | /* Check that inlined time is better, but tolerate some roundoff issues. |
1142 | FIXME: When callee profile drops to 0 we account calls more. This | |
1143 | should be fixed by never doing that. */ | |
3f05a4f0 JH |
1144 | gcc_checking_assert ((edge_time * 100 |
1145 | - callee_info->time * 101).to_int () <= 0 | |
1bad9c18 | 1146 | || callee->count.ipa ().initialized_p ()); |
f658ad30 | 1147 | gcc_checking_assert (growth <= ipa_size_summaries->get (callee)->size); |
e89964e3 | 1148 | |
1ce18dc8 JH |
1149 | if (dump) |
1150 | { | |
464d0118 ML |
1151 | fprintf (dump_file, " Badness calculation for %s -> %s\n", |
1152 | edge->caller->dump_name (), | |
1153 | edge->callee->dump_name ()); | |
4adaad64 | 1154 | fprintf (dump_file, " size growth %i, time %f unspec %f ", |
1ce18dc8 | 1155 | growth, |
4adaad64 JH |
1156 | edge_time.to_double (), |
1157 | unspec_edge_time.to_double ()); | |
0bceb671 | 1158 | ipa_dump_hints (dump_file, hints); |
42f7b0fa JH |
1159 | if (big_speedup_p (edge)) |
1160 | fprintf (dump_file, " big_speedup"); | |
37678631 | 1161 | fprintf (dump_file, "\n"); |
1ce18dc8 | 1162 | } |
45a80bb9 JH |
1163 | |
1164 | /* Always prefer inlining saving code size. */ | |
1165 | if (growth <= 0) | |
1ce18dc8 | 1166 | { |
6d4ab5f8 | 1167 | badness = (sreal) (-SREAL_MIN_SIG + growth) << (SREAL_MAX_EXP / 256); |
1ce18dc8 | 1168 | if (dump) |
6d4ab5f8 | 1169 | fprintf (dump_file, " %f: Growth %d <= 0\n", badness.to_double (), |
1ce18dc8 JH |
1170 | growth); |
1171 | } | |
208e5afa JH |
1172 | /* Inlining into EXTERNAL functions is not going to change anything unless |
1173 | they are themselves inlined. */ | |
1174 | else if (DECL_EXTERNAL (caller->decl)) | |
1ce18dc8 | 1175 | { |
1ce18dc8 | 1176 | if (dump) |
208e5afa JH |
1177 | fprintf (dump_file, " max: function is external\n"); |
1178 | return sreal::max (); | |
1ce18dc8 | 1179 | } |
208e5afa | 1180 | /* When profile is available. Compute badness as: |
b4c0a884 | 1181 | |
208e5afa | 1182 | time_saved * caller_count |
133a84ab JH |
1183 | goodness = ------------------------------------------------- |
1184 | growth_of_caller * overall_growth * combined_size | |
d59171da JH |
1185 | |
1186 | badness = - goodness | |
b4c0a884 | 1187 | |
208e5afa JH |
1188 | Again use negative value to make calls with profile appear hotter |
1189 | then calls without. | |
b4c0a884 | 1190 | */ |
3995f3a2 | 1191 | else if (opt_for_fn (caller->decl, flag_guess_branch_prob) |
1bad9c18 | 1192 | || caller->count.ipa ().nonzero_p ()) |
670cd5c5 | 1193 | { |
6d4ab5f8 | 1194 | sreal numerator, denominator; |
41f669d8 | 1195 | int overall_growth; |
b5351388 | 1196 | sreal freq = edge->sreal_frequency (); |
208e5afa | 1197 | |
ea8dd3b6 | 1198 | numerator = inlining_speedup (edge, freq, unspec_edge_time, edge_time); |
0009a6c3 | 1199 | if (numerator <= 0) |
208e5afa | 1200 | numerator = ((sreal) 1 >> 8); |
1bad9c18 JH |
1201 | if (caller->count.ipa ().nonzero_p ()) |
1202 | numerator *= caller->count.ipa ().to_gcov_type (); | |
1203 | else if (caller->count.ipa ().initialized_p ()) | |
208e5afa JH |
1204 | numerator = numerator >> 11; |
1205 | denominator = growth; | |
41f669d8 JH |
1206 | |
1207 | overall_growth = callee_info->growth; | |
1208 | ||
1209 | /* Look for inliner wrappers of the form: | |
1210 | ||
1211 | inline_caller () | |
1212 | { | |
1213 | do_fast_job... | |
1214 | if (need_more_work) | |
1215 | noninline_callee (); | |
1216 | } | |
956d615d | 1217 | Without penalizing this case, we usually inline noninline_callee |
41f669d8 JH |
1218 | into the inline_caller because overall_growth is small preventing |
1219 | further inlining of inline_caller. | |
1220 | ||
1221 | Penalize only callgraph edges to functions with small overall | |
1222 | growth ... | |
1223 | */ | |
1224 | if (growth > overall_growth | |
1225 | /* ... and having only one caller which is not inlined ... */ | |
1226 | && callee_info->single_caller | |
a62bfab5 | 1227 | && !edge->caller->inlined_to |
41f669d8 | 1228 | /* ... and edges executed only conditionally ... */ |
b5351388 | 1229 | && freq < 1 |
41f669d8 JH |
1230 | /* ... consider case where callee is not inline but caller is ... */ |
1231 | && ((!DECL_DECLARED_INLINE_P (edge->callee->decl) | |
1232 | && DECL_DECLARED_INLINE_P (caller->decl)) | |
1233 | /* ... or when early optimizers decided to split and edge | |
1234 | frequency still indicates splitting is a win ... */ | |
1235 | || (callee->split_part && !caller->split_part | |
1e83bd70 JH |
1236 | && freq * 100 |
1237 | < opt_for_fn (caller->decl, | |
1238 | param_partial_inlining_entry_probability) | |
41f669d8 JH |
1239 | /* ... and do not overwrite user specified hints. */ |
1240 | && (!DECL_DECLARED_INLINE_P (edge->callee->decl) | |
1241 | || DECL_DECLARED_INLINE_P (caller->decl))))) | |
1242 | { | |
56f62793 | 1243 | ipa_fn_summary *caller_info = ipa_fn_summaries->get (caller); |
41f669d8 JH |
1244 | int caller_growth = caller_info->growth; |
1245 | ||
1246 | /* Only apply the penalty when caller looks like inline candidate, | |
2925cad2 | 1247 | and it is not called once. */ |
41f669d8 JH |
1248 | if (!caller_info->single_caller && overall_growth < caller_growth |
1249 | && caller_info->inlinable | |
041cb615 JH |
1250 | && wrapper_heuristics_may_apply |
1251 | (caller, ipa_size_summaries->get (caller)->size)) | |
41f669d8 JH |
1252 | { |
1253 | if (dump) | |
1254 | fprintf (dump_file, | |
1255 | " Wrapper penalty. Increasing growth %i to %i\n", | |
1256 | overall_growth, caller_growth); | |
1257 | overall_growth = caller_growth; | |
1258 | } | |
1259 | } | |
1260 | if (overall_growth > 0) | |
1261 | { | |
e4112065 | 1262 | /* Strongly prefer functions with few callers that can be inlined |
41f669d8 JH |
1263 | fully. The square root here leads to smaller binaries at average. |
1264 | Watch however for extreme cases and return to linear function | |
1265 | when growth is large. */ | |
1266 | if (overall_growth < 256) | |
1267 | overall_growth *= overall_growth; | |
1268 | else | |
1269 | overall_growth += 256 * 256 - 256; | |
1270 | denominator *= overall_growth; | |
1271 | } | |
f658ad30 | 1272 | denominator *= ipa_size_summaries->get (caller)->size + growth; |
6d4ab5f8 JH |
1273 | |
1274 | badness = - numerator / denominator; | |
1275 | ||
1ce18dc8 JH |
1276 | if (dump) |
1277 | { | |
1278 | fprintf (dump_file, | |
16998094 JM |
1279 | " %f: guessed profile. frequency %f, count %" PRId64 |
1280 | " caller count %" PRId64 | |
ea8dd3b6 | 1281 | " time saved %f" |
41f669d8 JH |
1282 | " overall growth %i (current) %i (original)" |
1283 | " %i (compensated)\n", | |
1284 | badness.to_double (), | |
b5351388 | 1285 | freq.to_double (), |
1bad9c18 JH |
1286 | edge->count.ipa ().initialized_p () ? edge->count.ipa ().to_gcov_type () : -1, |
1287 | caller->count.ipa ().initialized_p () ? caller->count.ipa ().to_gcov_type () : -1, | |
ea8dd3b6 | 1288 | inlining_speedup (edge, freq, unspec_edge_time, edge_time).to_double (), |
d59171da | 1289 | estimate_growth (callee), |
41f669d8 | 1290 | callee_info->growth, overall_growth); |
1ce18dc8 | 1291 | } |
45a80bb9 JH |
1292 | } |
1293 | /* When function local profile is not available or it does not give | |
956d615d | 1294 | useful information (i.e. frequency is zero), base the cost on |
45a80bb9 JH |
1295 | loop nest and overall size growth, so we optimize for overall number |
1296 | of functions fully inlined in program. */ | |
1297 | else | |
1298 | { | |
56f62793 | 1299 | int nest = MIN (ipa_call_summaries->get (edge)->loop_depth, 8); |
6d4ab5f8 | 1300 | badness = growth; |
670cd5c5 | 1301 | |
45a80bb9 | 1302 | /* Decrease badness if call is nested. */ |
b8698a0f | 1303 | if (badness > 0) |
f0e1509b | 1304 | badness = badness >> nest; |
45a80bb9 | 1305 | else |
6d4ab5f8 | 1306 | badness = badness << nest; |
1ce18dc8 | 1307 | if (dump) |
41f669d8 JH |
1308 | fprintf (dump_file, " %f: no profile. nest %i\n", |
1309 | badness.to_double (), nest); | |
670cd5c5 | 1310 | } |
6d4ab5f8 | 1311 | gcc_checking_assert (badness != 0); |
1ce18dc8 | 1312 | |
3dafb85c | 1313 | if (edge->recursive_p ()) |
6d4ab5f8 JH |
1314 | badness = badness.shift (badness > 0 ? 4 : -4); |
1315 | if ((hints & (INLINE_HINT_indirect_call | |
1316 | | INLINE_HINT_loop_iterations | |
6d4ab5f8 JH |
1317 | | INLINE_HINT_loop_stride)) |
1318 | || callee_info->growth <= 0) | |
1319 | badness = badness.shift (badness > 0 ? -2 : 2); | |
1320 | if (hints & (INLINE_HINT_same_scc)) | |
1321 | badness = badness.shift (badness > 0 ? 3 : -3); | |
1322 | else if (hints & (INLINE_HINT_in_scc)) | |
1323 | badness = badness.shift (badness > 0 ? 2 : -2); | |
1324 | else if (hints & (INLINE_HINT_cross_module)) | |
1325 | badness = badness.shift (badness > 0 ? 1 : -1); | |
208e5afa JH |
1326 | if (DECL_DISREGARD_INLINE_LIMITS (callee->decl)) |
1327 | badness = badness.shift (badness > 0 ? -4 : 4); | |
1328 | else if ((hints & INLINE_HINT_declared_inline)) | |
6d4ab5f8 JH |
1329 | badness = badness.shift (badness > 0 ? -3 : 3); |
1330 | if (dump) | |
1331 | fprintf (dump_file, " Adjusted by hints %f\n", badness.to_double ()); | |
1332 | return badness; | |
670cd5c5 JH |
1333 | } |
1334 | ||
9b8051b4 | 1335 | /* Recompute badness of EDGE and update its key in HEAP if needed. */ |
4c0f7679 | 1336 | static inline void |
4a910049 | 1337 | update_edge_key (edge_heap_t *heap, struct cgraph_edge *edge) |
9b8051b4 | 1338 | { |
f0e1509b | 1339 | sreal badness = edge_badness (edge, false); |
9b8051b4 JH |
1340 | if (edge->aux) |
1341 | { | |
4a910049 ML |
1342 | edge_heap_node_t *n = (edge_heap_node_t *) edge->aux; |
1343 | gcc_checking_assert (n->get_data () == edge); | |
9b8051b4 | 1344 | |
6d4ab5f8 | 1345 | /* fibonacci_heap::replace_key does busy updating of the |
956d615d | 1346 | heap that is unnecessarily expensive. |
6d4ab5f8 JH |
1347 | We do lazy increases: after extracting minimum if the key |
1348 | turns out to be out of date, it is re-inserted into heap | |
1349 | with correct value. */ | |
4a910049 | 1350 | if (badness < n->get_key ()) |
9b8051b4 | 1351 | { |
4c0f7679 JH |
1352 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1353 | { | |
1354 | fprintf (dump_file, | |
464d0118 ML |
1355 | " decreasing badness %s -> %s, %f to %f\n", |
1356 | edge->caller->dump_name (), | |
1357 | edge->callee->dump_name (), | |
6d4ab5f8 JH |
1358 | n->get_key ().to_double (), |
1359 | badness.to_double ()); | |
4c0f7679 | 1360 | } |
4a910049 | 1361 | heap->decrease_key (n, badness); |
9b8051b4 JH |
1362 | } |
1363 | } | |
1364 | else | |
4c0f7679 JH |
1365 | { |
1366 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1367 | { | |
1368 | fprintf (dump_file, | |
464d0118 ML |
1369 | " enqueuing call %s -> %s, badness %f\n", |
1370 | edge->caller->dump_name (), | |
1371 | edge->callee->dump_name (), | |
6d4ab5f8 | 1372 | badness.to_double ()); |
4c0f7679 | 1373 | } |
4a910049 | 1374 | edge->aux = heap->insert (badness, edge); |
4c0f7679 | 1375 | } |
9b8051b4 JH |
1376 | } |
1377 | ||
40fda55b JH |
1378 | |
1379 | /* NODE was inlined. | |
956d615d | 1380 | All caller edges needs to be reset because |
40fda55b JH |
1381 | size estimates change. Similarly callees needs reset |
1382 | because better context may be known. */ | |
1383 | ||
1384 | static void | |
1385 | reset_edge_caches (struct cgraph_node *node) | |
1386 | { | |
1387 | struct cgraph_edge *edge; | |
1388 | struct cgraph_edge *e = node->callees; | |
1389 | struct cgraph_node *where = node; | |
e55637b7 | 1390 | struct ipa_ref *ref; |
40fda55b | 1391 | |
a62bfab5 ML |
1392 | if (where->inlined_to) |
1393 | where = where->inlined_to; | |
40fda55b | 1394 | |
ac6f2e59 JH |
1395 | reset_node_cache (where); |
1396 | ||
9fb50ad8 ML |
1397 | if (edge_growth_cache != NULL) |
1398 | for (edge = where->callers; edge; edge = edge->next_caller) | |
1399 | if (edge->inline_failed) | |
1400 | edge_growth_cache->remove (edge); | |
e55637b7 ML |
1401 | |
1402 | FOR_EACH_ALIAS (where, ref) | |
1403 | reset_edge_caches (dyn_cast <cgraph_node *> (ref->referring)); | |
40fda55b JH |
1404 | |
1405 | if (!e) | |
1406 | return; | |
1407 | ||
1408 | while (true) | |
1409 | if (!e->inline_failed && e->callee->callees) | |
1410 | e = e->callee->callees; | |
1411 | else | |
1412 | { | |
9fb50ad8 ML |
1413 | if (edge_growth_cache != NULL && e->inline_failed) |
1414 | edge_growth_cache->remove (e); | |
40fda55b JH |
1415 | if (e->next_callee) |
1416 | e = e->next_callee; | |
1417 | else | |
1418 | { | |
1419 | do | |
1420 | { | |
1421 | if (e->caller == node) | |
1422 | return; | |
1423 | e = e->caller->callers; | |
1424 | } | |
1425 | while (!e->next_callee); | |
1426 | e = e->next_callee; | |
1427 | } | |
1428 | } | |
1429 | } | |
1430 | ||
1431 | /* Recompute HEAP nodes for each of caller of NODE. | |
1432 | UPDATED_NODES track nodes we already visited, to avoid redundant work. | |
1433 | When CHECK_INLINABLITY_FOR is set, re-check for specified edge that | |
1434 | it is inlinable. Otherwise check all edges. */ | |
670cd5c5 JH |
1435 | |
1436 | static void | |
4a910049 | 1437 | update_caller_keys (edge_heap_t *heap, struct cgraph_node *node, |
40fda55b JH |
1438 | bitmap updated_nodes, |
1439 | struct cgraph_edge *check_inlinablity_for) | |
670cd5c5 JH |
1440 | { |
1441 | struct cgraph_edge *edge; | |
e55637b7 | 1442 | struct ipa_ref *ref; |
670cd5c5 | 1443 | |
56f62793 | 1444 | if ((!node->alias && !ipa_fn_summaries->get (node)->inlinable) |
a62bfab5 | 1445 | || node->inlined_to) |
670cd5c5 | 1446 | return; |
4325656f | 1447 | if (!bitmap_set_bit (updated_nodes, node->get_uid ())) |
670cd5c5 | 1448 | return; |
670cd5c5 | 1449 | |
e55637b7 ML |
1450 | FOR_EACH_ALIAS (node, ref) |
1451 | { | |
1452 | struct cgraph_node *alias = dyn_cast <cgraph_node *> (ref->referring); | |
1453 | update_caller_keys (heap, alias, updated_nodes, check_inlinablity_for); | |
1454 | } | |
39e2db00 | 1455 | |
cdc029b9 | 1456 | for (edge = node->callers; edge; edge = edge->next_caller) |
4c0f7679 JH |
1457 | if (edge->inline_failed) |
1458 | { | |
40fda55b JH |
1459 | if (!check_inlinablity_for |
1460 | || check_inlinablity_for == edge) | |
f10d1a74 | 1461 | { |
9a4841a3 JH |
1462 | if (can_inline_edge_p (edge, false) |
1463 | && want_inline_small_function_p (edge, false) | |
1464 | && can_inline_edge_by_limits_p (edge, false)) | |
40fda55b JH |
1465 | update_edge_key (heap, edge); |
1466 | else if (edge->aux) | |
1467 | { | |
1468 | report_inline_failed_reason (edge); | |
4a910049 | 1469 | heap->delete_node ((edge_heap_node_t *) edge->aux); |
40fda55b JH |
1470 | edge->aux = NULL; |
1471 | } | |
f10d1a74 | 1472 | } |
40fda55b JH |
1473 | else if (edge->aux) |
1474 | update_edge_key (heap, edge); | |
4c0f7679 | 1475 | } |
9b8051b4 JH |
1476 | } |
1477 | ||
7c6f2fb9 JH |
1478 | /* Recompute HEAP nodes for each uninlined call in NODE |
1479 | If UPDATE_SINCE is non-NULL check if edges called within that function | |
1480 | are inlinable (typically UPDATE_SINCE is the inline clone we introduced | |
1481 | where all edges have new context). | |
1482 | ||
9b8051b4 JH |
1483 | This is used when we know that edge badnesses are going only to increase |
1484 | (we introduced new call site) and thus all we need is to insert newly | |
1485 | created edges into heap. */ | |
1486 | ||
1487 | static void | |
4a910049 | 1488 | update_callee_keys (edge_heap_t *heap, struct cgraph_node *node, |
7c6f2fb9 | 1489 | struct cgraph_node *update_since, |
9b8051b4 JH |
1490 | bitmap updated_nodes) |
1491 | { | |
1492 | struct cgraph_edge *e = node->callees; | |
7c6f2fb9 | 1493 | bool check_inlinability = update_since == node; |
09a2806f | 1494 | |
9b8051b4 JH |
1495 | if (!e) |
1496 | return; | |
1497 | while (true) | |
1498 | if (!e->inline_failed && e->callee->callees) | |
7c6f2fb9 JH |
1499 | { |
1500 | if (e->callee == update_since) | |
1501 | check_inlinability = true; | |
1502 | e = e->callee->callees; | |
1503 | } | |
9b8051b4 | 1504 | else |
670cd5c5 | 1505 | { |
a5b1779f JH |
1506 | enum availability avail; |
1507 | struct cgraph_node *callee; | |
7c6f2fb9 JH |
1508 | if (!check_inlinability) |
1509 | { | |
1510 | if (e->aux | |
1511 | && !bitmap_bit_p (updated_nodes, | |
1512 | e->callee->ultimate_alias_target | |
1513 | (&avail, e->caller)->get_uid ())) | |
1514 | update_edge_key (heap, e); | |
1515 | } | |
58696ce5 | 1516 | /* We do not reset callee growth cache here. Since we added a new call, |
956d615d | 1517 | growth should have just increased and consequently badness metric |
58696ce5 | 1518 | don't need updating. */ |
7c6f2fb9 JH |
1519 | else if (e->inline_failed |
1520 | && (callee = e->callee->ultimate_alias_target (&avail, | |
1521 | e->caller)) | |
1522 | && avail >= AVAIL_AVAILABLE | |
1523 | && ipa_fn_summaries->get (callee) != NULL | |
1524 | && ipa_fn_summaries->get (callee)->inlinable | |
1525 | && !bitmap_bit_p (updated_nodes, callee->get_uid ())) | |
670cd5c5 | 1526 | { |
9a4841a3 JH |
1527 | if (can_inline_edge_p (e, false) |
1528 | && want_inline_small_function_p (e, false) | |
1529 | && can_inline_edge_by_limits_p (e, false)) | |
7c6f2fb9 JH |
1530 | { |
1531 | gcc_checking_assert (check_inlinability || can_inline_edge_p (e, false)); | |
1532 | gcc_checking_assert (check_inlinability || e->aux); | |
1533 | update_edge_key (heap, e); | |
1534 | } | |
40fda55b JH |
1535 | else if (e->aux) |
1536 | { | |
1537 | report_inline_failed_reason (e); | |
4a910049 | 1538 | heap->delete_node ((edge_heap_node_t *) e->aux); |
40fda55b JH |
1539 | e->aux = NULL; |
1540 | } | |
9b8051b4 | 1541 | } |
7c6f2fb9 JH |
1542 | /* In case we redirected to unreachable node we only need to remove the |
1543 | fibheap entry. */ | |
1544 | else if (e->aux) | |
1545 | { | |
1546 | heap->delete_node ((edge_heap_node_t *) e->aux); | |
1547 | e->aux = NULL; | |
1548 | } | |
9b8051b4 JH |
1549 | if (e->next_callee) |
1550 | e = e->next_callee; | |
1551 | else | |
1552 | { | |
1553 | do | |
1ce18dc8 | 1554 | { |
9b8051b4 JH |
1555 | if (e->caller == node) |
1556 | return; | |
7c6f2fb9 JH |
1557 | if (e->caller == update_since) |
1558 | check_inlinability = false; | |
9b8051b4 | 1559 | e = e->caller->callers; |
1ce18dc8 | 1560 | } |
9b8051b4 JH |
1561 | while (!e->next_callee); |
1562 | e = e->next_callee; | |
670cd5c5 | 1563 | } |
670cd5c5 JH |
1564 | } |
1565 | } | |
1566 | ||
670cd5c5 | 1567 | /* Enqueue all recursive calls from NODE into priority queue depending on |
0fa2e4df | 1568 | how likely we want to recursively inline the call. */ |
670cd5c5 | 1569 | |
ca31b95f JH |
1570 | static void |
1571 | lookup_recursive_calls (struct cgraph_node *node, struct cgraph_node *where, | |
4a910049 | 1572 | edge_heap_t *heap) |
ca31b95f JH |
1573 | { |
1574 | struct cgraph_edge *e; | |
a5b1779f JH |
1575 | enum availability avail; |
1576 | ||
ca31b95f | 1577 | for (e = where->callees; e; e = e->next_callee) |
a5b1779f | 1578 | if (e->callee == node |
e6007a27 | 1579 | || (e->callee->ultimate_alias_target (&avail, e->caller) == node |
d52f5295 | 1580 | && avail > AVAIL_INTERPOSABLE)) |
0009a6c3 | 1581 | heap->insert (-e->sreal_frequency (), e); |
ca31b95f JH |
1582 | for (e = where->callees; e; e = e->next_callee) |
1583 | if (!e->inline_failed) | |
670cd5c5 | 1584 | lookup_recursive_calls (node, e->callee, heap); |
ca31b95f JH |
1585 | } |
1586 | ||
1587 | /* Decide on recursive inlining: in the case function has recursive calls, | |
3e293154 | 1588 | inline until body size reaches given argument. If any new indirect edges |
e56f5f3e JJ |
1589 | are discovered in the process, add them to *NEW_EDGES, unless NEW_EDGES |
1590 | is NULL. */ | |
670cd5c5 JH |
1591 | |
1592 | static bool | |
4c0f7679 | 1593 | recursive_inlining (struct cgraph_edge *edge, |
d52f5295 | 1594 | vec<cgraph_edge *> *new_edges) |
ca31b95f | 1595 | { |
1e83bd70 JH |
1596 | cgraph_node *to = (edge->caller->inlined_to |
1597 | ? edge->caller->inlined_to : edge->caller); | |
1598 | int limit = opt_for_fn (to->decl, | |
1599 | param_max_inline_insns_recursive_auto); | |
d1704358 | 1600 | edge_heap_t heap (sreal::min ()); |
d7d1d041 | 1601 | struct cgraph_node *node; |
ca31b95f | 1602 | struct cgraph_edge *e; |
4c0f7679 | 1603 | struct cgraph_node *master_clone = NULL, *next; |
ca31b95f JH |
1604 | int depth = 0; |
1605 | int n = 0; | |
1606 | ||
d7d1d041 | 1607 | node = edge->caller; |
a62bfab5 ML |
1608 | if (node->inlined_to) |
1609 | node = node->inlined_to; | |
d7d1d041 | 1610 | |
67348ccc | 1611 | if (DECL_DECLARED_INLINE_P (node->decl)) |
1e83bd70 | 1612 | limit = opt_for_fn (to->decl, param_max_inline_insns_recursive); |
ca31b95f JH |
1613 | |
1614 | /* Make sure that function is small enough to be considered for inlining. */ | |
4c0f7679 | 1615 | if (estimate_size_after_inlining (node, edge) >= limit) |
670cd5c5 | 1616 | return false; |
4a910049 ML |
1617 | lookup_recursive_calls (node, node, &heap); |
1618 | if (heap.empty ()) | |
1619 | return false; | |
ca31b95f JH |
1620 | |
1621 | if (dump_file) | |
b8698a0f | 1622 | fprintf (dump_file, |
3629ff8a | 1623 | " Performing recursive inlining on %s\n", node->dump_name ()); |
ca31b95f | 1624 | |
ca31b95f | 1625 | /* Do the inlining and update list of recursive call during process. */ |
4a910049 | 1626 | while (!heap.empty ()) |
ca31b95f | 1627 | { |
4a910049 | 1628 | struct cgraph_edge *curr = heap.extract_min (); |
6ced940d | 1629 | struct cgraph_node *cnode, *dest = curr->callee; |
d7d1d041 | 1630 | |
9a4841a3 | 1631 | if (!can_inline_edge_p (curr, true) |
4eb50396 | 1632 | || !can_inline_edge_by_limits_p (curr, true)) |
4c0f7679 JH |
1633 | continue; |
1634 | ||
6ced940d JH |
1635 | /* MASTER_CLONE is produced in the case we already started modified |
1636 | the function. Be sure to redirect edge to the original body before | |
1637 | estimating growths otherwise we will be seeing growths after inlining | |
1638 | the already modified body. */ | |
1639 | if (master_clone) | |
1640 | { | |
3dafb85c | 1641 | curr->redirect_callee (master_clone); |
9fb50ad8 ML |
1642 | if (edge_growth_cache != NULL) |
1643 | edge_growth_cache->remove (curr); | |
6ced940d JH |
1644 | } |
1645 | ||
1646 | if (estimate_size_after_inlining (node, curr) > limit) | |
1647 | { | |
3dafb85c | 1648 | curr->redirect_callee (dest); |
9fb50ad8 ML |
1649 | if (edge_growth_cache != NULL) |
1650 | edge_growth_cache->remove (curr); | |
6ced940d JH |
1651 | break; |
1652 | } | |
1653 | ||
c5a4444c JH |
1654 | depth = 1; |
1655 | for (cnode = curr->caller; | |
a62bfab5 | 1656 | cnode->inlined_to; cnode = cnode->callers->caller) |
67348ccc | 1657 | if (node->decl |
d52f5295 | 1658 | == curr->callee->ultimate_alias_target ()->decl) |
f791d333 | 1659 | depth++; |
c5a4444c | 1660 | |
4c0f7679 | 1661 | if (!want_inline_self_recursive_call_p (curr, node, false, depth)) |
6ced940d | 1662 | { |
3dafb85c | 1663 | curr->redirect_callee (dest); |
9fb50ad8 ML |
1664 | if (edge_growth_cache != NULL) |
1665 | edge_growth_cache->remove (curr); | |
6ced940d JH |
1666 | continue; |
1667 | } | |
ca31b95f | 1668 | |
670cd5c5 | 1669 | if (dump_file) |
c5a4444c | 1670 | { |
b8698a0f | 1671 | fprintf (dump_file, |
c5a4444c | 1672 | " Inlining call of depth %i", depth); |
ae94bb0e | 1673 | if (node->count.nonzero_p () && curr->count.initialized_p ()) |
c5a4444c JH |
1674 | { |
1675 | fprintf (dump_file, " called approx. %.2f times per call", | |
3995f3a2 JH |
1676 | (double)curr->count.to_gcov_type () |
1677 | / node->count.to_gcov_type ()); | |
c5a4444c JH |
1678 | } |
1679 | fprintf (dump_file, "\n"); | |
1680 | } | |
4c0f7679 JH |
1681 | if (!master_clone) |
1682 | { | |
1683 | /* We need original clone to copy around. */ | |
d52f5295 | 1684 | master_clone = node->create_clone (node->decl, node->count, |
1bad9c18 | 1685 | false, vNULL, true, NULL, NULL); |
4c0f7679 JH |
1686 | for (e = master_clone->callees; e; e = e->next_callee) |
1687 | if (!e->inline_failed) | |
1bad9c18 | 1688 | clone_inlined_nodes (e, true, false, NULL); |
3dafb85c | 1689 | curr->redirect_callee (master_clone); |
9fb50ad8 ML |
1690 | if (edge_growth_cache != NULL) |
1691 | edge_growth_cache->remove (curr); | |
4c0f7679 JH |
1692 | } |
1693 | ||
c170d40f | 1694 | inline_call (curr, false, new_edges, &overall_size, true); |
b914768c | 1695 | reset_node_cache (node); |
4a910049 | 1696 | lookup_recursive_calls (node, curr->callee, &heap); |
ca31b95f JH |
1697 | n++; |
1698 | } | |
4c0f7679 | 1699 | |
4a910049 | 1700 | if (!heap.empty () && dump_file) |
c5a4444c | 1701 | fprintf (dump_file, " Recursive inlining growth limit met.\n"); |
4c0f7679 JH |
1702 | |
1703 | if (!master_clone) | |
1704 | return false; | |
1705 | ||
4174a33a DM |
1706 | if (dump_enabled_p ()) |
1707 | dump_printf_loc (MSG_NOTE, edge->call_stmt, | |
1708 | "\n Inlined %i times, " | |
1709 | "body grown from size %i to %i, time %f to %f\n", n, | |
f658ad30 JH |
1710 | ipa_size_summaries->get (master_clone)->size, |
1711 | ipa_size_summaries->get (node)->size, | |
4174a33a DM |
1712 | ipa_fn_summaries->get (master_clone)->time.to_double (), |
1713 | ipa_fn_summaries->get (node)->time.to_double ()); | |
ca31b95f JH |
1714 | |
1715 | /* Remove master clone we used for inlining. We rely that clones inlined | |
1716 | into master clone gets queued just before master clone so we don't | |
1717 | need recursion. */ | |
3dafb85c | 1718 | for (node = symtab->first_function (); node != master_clone; |
96fc428c JH |
1719 | node = next) |
1720 | { | |
3dafb85c | 1721 | next = symtab->next_function (node); |
a62bfab5 | 1722 | if (node->inlined_to == master_clone) |
d52f5295 | 1723 | node->remove (); |
96fc428c | 1724 | } |
d52f5295 | 1725 | master_clone->remove (); |
4c0f7679 | 1726 | return true; |
ca31b95f JH |
1727 | } |
1728 | ||
09a2806f | 1729 | |
88512ba0 | 1730 | /* Given whole compilation unit estimate of INSNS, compute how large we can |
b7c27d51 | 1731 | allow the unit to grow. */ |
09a2806f | 1732 | |
8fcfc44f | 1733 | static int64_t |
1e83bd70 | 1734 | compute_max_insns (cgraph_node *node, int insns) |
b7c27d51 JH |
1735 | { |
1736 | int max_insns = insns; | |
1e83bd70 JH |
1737 | if (max_insns < opt_for_fn (node->decl, param_large_unit_insns)) |
1738 | max_insns = opt_for_fn (node->decl, param_large_unit_insns); | |
b7c27d51 | 1739 | |
a9243bfc | 1740 | return ((int64_t) max_insns |
1e83bd70 | 1741 | * (100 + opt_for_fn (node->decl, param_inline_unit_growth)) / 100); |
b7c27d51 JH |
1742 | } |
1743 | ||
09a2806f | 1744 | |
3e293154 | 1745 | /* Compute badness of all edges in NEW_EDGES and add them to the HEAP. */ |
09a2806f | 1746 | |
3e293154 | 1747 | static void |
4a910049 | 1748 | add_new_edges_to_heap (edge_heap_t *heap, vec<cgraph_edge *> new_edges) |
3e293154 | 1749 | { |
9771b263 | 1750 | while (new_edges.length () > 0) |
3e293154 | 1751 | { |
9771b263 | 1752 | struct cgraph_edge *edge = new_edges.pop (); |
3e293154 MJ |
1753 | |
1754 | gcc_assert (!edge->aux); | |
4517b378 | 1755 | gcc_assert (edge->callee); |
a5b1779f | 1756 | if (edge->inline_failed |
4c0f7679 | 1757 | && can_inline_edge_p (edge, true) |
9a4841a3 JH |
1758 | && want_inline_small_function_p (edge, true) |
1759 | && can_inline_edge_by_limits_p (edge, true)) | |
4a910049 | 1760 | edge->aux = heap->insert (edge_badness (edge, false), edge); |
3e293154 MJ |
1761 | } |
1762 | } | |
1763 | ||
042ae7d2 JH |
1764 | /* Remove EDGE from the fibheap. */ |
1765 | ||
1766 | static void | |
1767 | heap_edge_removal_hook (struct cgraph_edge *e, void *data) | |
1768 | { | |
1769 | if (e->aux) | |
1770 | { | |
4a910049 | 1771 | ((edge_heap_t *)data)->delete_node ((edge_heap_node_t *)e->aux); |
042ae7d2 JH |
1772 | e->aux = NULL; |
1773 | } | |
1774 | } | |
3e293154 | 1775 | |
09ce3660 JH |
1776 | /* Return true if speculation of edge E seems useful. |
1777 | If ANTICIPATE_INLINING is true, be conservative and hope that E | |
1778 | may get inlined. */ | |
1779 | ||
1780 | bool | |
1781 | speculation_useful_p (struct cgraph_edge *e, bool anticipate_inlining) | |
1782 | { | |
4517b378 MJ |
1783 | /* If we have already decided to inline the edge, it seems useful. */ |
1784 | if (!e->inline_failed) | |
1785 | return true; | |
1786 | ||
09ce3660 | 1787 | enum availability avail; |
e6007a27 JH |
1788 | struct cgraph_node *target = e->callee->ultimate_alias_target (&avail, |
1789 | e->caller); | |
09ce3660 JH |
1790 | |
1791 | gcc_assert (e->speculative && !e->indirect_unknown_callee); | |
1792 | ||
3dafb85c | 1793 | if (!e->maybe_hot_p ()) |
09ce3660 JH |
1794 | return false; |
1795 | ||
1796 | /* See if IP optimizations found something potentially useful about the | |
1797 | function. For now we look only for CONST/PURE flags. Almost everything | |
1798 | else we propagate is useless. */ | |
1799 | if (avail >= AVAIL_AVAILABLE) | |
1800 | { | |
67348ccc | 1801 | int ecf_flags = flags_from_decl_or_type (target->decl); |
09ce3660 JH |
1802 | if (ecf_flags & ECF_CONST) |
1803 | { | |
845bb366 JH |
1804 | if (!(e->speculative_call_indirect_edge ()->indirect_info |
1805 | ->ecf_flags & ECF_CONST)) | |
09ce3660 JH |
1806 | return true; |
1807 | } | |
1808 | else if (ecf_flags & ECF_PURE) | |
1809 | { | |
845bb366 JH |
1810 | if (!(e->speculative_call_indirect_edge ()->indirect_info |
1811 | ->ecf_flags & ECF_PURE)) | |
09ce3660 JH |
1812 | return true; |
1813 | } | |
1814 | } | |
1815 | /* If we did not managed to inline the function nor redirect | |
1816 | to an ipa-cp clone (that are seen by having local flag set), | |
1817 | it is probably pointless to inline it unless hardware is missing | |
1818 | indirect call predictor. */ | |
87f94429 | 1819 | if (!anticipate_inlining && !target->local) |
09ce3660 JH |
1820 | return false; |
1821 | /* For overwritable targets there is not much to do. */ | |
4517b378 MJ |
1822 | if (!can_inline_edge_p (e, false) |
1823 | || !can_inline_edge_by_limits_p (e, false, true)) | |
09ce3660 JH |
1824 | return false; |
1825 | /* OK, speculation seems interesting. */ | |
1826 | return true; | |
1827 | } | |
1828 | ||
1829 | /* We know that EDGE is not going to be inlined. | |
1830 | See if we can remove speculation. */ | |
1831 | ||
1832 | static void | |
4a910049 | 1833 | resolve_noninline_speculation (edge_heap_t *edge_heap, struct cgraph_edge *edge) |
09ce3660 JH |
1834 | { |
1835 | if (edge->speculative && !speculation_useful_p (edge, false)) | |
1836 | { | |
1837 | struct cgraph_node *node = edge->caller; | |
a62bfab5 ML |
1838 | struct cgraph_node *where = node->inlined_to |
1839 | ? node->inlined_to : node; | |
0e3de1d4 | 1840 | auto_bitmap updated_nodes; |
09ce3660 | 1841 | |
1bad9c18 JH |
1842 | if (edge->count.ipa ().initialized_p ()) |
1843 | spec_rem += edge->count.ipa (); | |
27c5a177 | 1844 | cgraph_edge::resolve_speculation (edge); |
09ce3660 | 1845 | reset_edge_caches (where); |
0bceb671 | 1846 | ipa_update_overall_fn_summary (where); |
09ce3660 JH |
1847 | update_caller_keys (edge_heap, where, |
1848 | updated_nodes, NULL); | |
7c6f2fb9 | 1849 | update_callee_keys (edge_heap, where, NULL, |
d0b66480 | 1850 | updated_nodes); |
09ce3660 JH |
1851 | } |
1852 | } | |
1853 | ||
bb1e543c JH |
1854 | /* Return true if NODE should be accounted for overall size estimate. |
1855 | Skip all nodes optimized for size so we can measure the growth of hot | |
1856 | part of program no matter of the padding. */ | |
1857 | ||
1858 | bool | |
1859 | inline_account_function_p (struct cgraph_node *node) | |
1860 | { | |
1861 | return (!DECL_EXTERNAL (node->decl) | |
1862 | && !opt_for_fn (node->decl, optimize_size) | |
1863 | && node->frequency != NODE_FREQUENCY_UNLIKELY_EXECUTED); | |
1864 | } | |
1865 | ||
41f669d8 JH |
1866 | /* Count number of callers of NODE and store it into DATA (that |
1867 | points to int. Worker for cgraph_for_node_and_aliases. */ | |
1868 | ||
1869 | static bool | |
1870 | sum_callers (struct cgraph_node *node, void *data) | |
1871 | { | |
1872 | struct cgraph_edge *e; | |
1873 | int *num_calls = (int *)data; | |
1874 | ||
1875 | for (e = node->callers; e; e = e->next_caller) | |
1876 | (*num_calls)++; | |
1877 | return false; | |
1878 | } | |
1879 | ||
97e59627 ML |
1880 | /* We only propagate across edges with non-interposable callee. */ |
1881 | ||
1882 | inline bool | |
1883 | ignore_edge_p (struct cgraph_edge *e) | |
1884 | { | |
1885 | enum availability avail; | |
1886 | e->callee->function_or_virtual_thunk_symbol (&avail, e->caller); | |
1887 | return (avail <= AVAIL_INTERPOSABLE); | |
1888 | } | |
1889 | ||
ca31b95f | 1890 | /* We use greedy algorithm for inlining of small functions: |
09a2806f JH |
1891 | All inline candidates are put into prioritized heap ordered in |
1892 | increasing badness. | |
ca31b95f | 1893 | |
09a2806f | 1894 | The inlining of small functions is bounded by unit growth parameters. */ |
ca31b95f JH |
1895 | |
1896 | static void | |
4c0f7679 | 1897 | inline_small_functions (void) |
ca31b95f JH |
1898 | { |
1899 | struct cgraph_node *node; | |
670cd5c5 | 1900 | struct cgraph_edge *edge; |
f0e1509b | 1901 | edge_heap_t edge_heap (sreal::min ()); |
0e3de1d4 | 1902 | auto_bitmap updated_nodes; |
1e83bd70 | 1903 | int min_size; |
d52f5295 | 1904 | auto_vec<cgraph_edge *> new_indirect_edges; |
09a2806f | 1905 | int initial_size = 0; |
3dafb85c | 1906 | struct cgraph_node **order = XCNEWVEC (cgraph_node *, symtab->cgraph_count); |
042ae7d2 | 1907 | struct cgraph_edge_hook_list *edge_removal_hook_holder; |
2bf86c84 | 1908 | new_indirect_edges.create (8); |
670cd5c5 | 1909 | |
042ae7d2 | 1910 | edge_removal_hook_holder |
4a910049 | 1911 | = symtab->add_edge_removal_hook (&heap_edge_removal_hook, &edge_heap); |
042ae7d2 | 1912 | |
1a3118e9 JH |
1913 | /* Compute overall unit size and other global parameters used by badness |
1914 | metrics. */ | |
ca31b95f | 1915 | |
3995f3a2 | 1916 | max_count = profile_count::uninitialized (); |
97e59627 | 1917 | ipa_reduced_postorder (order, true, ignore_edge_p); |
68cc8feb | 1918 | free (order); |
1a3118e9 | 1919 | |
c47d0034 | 1920 | FOR_EACH_DEFINED_FUNCTION (node) |
a62bfab5 | 1921 | if (!node->inlined_to) |
e7f23018 | 1922 | { |
bb1e543c | 1923 | if (!node->alias && node->analyzed |
29f1e2b1 JH |
1924 | && (node->has_gimple_body_p () || node->thunk.thunk_p) |
1925 | && opt_for_fn (node->decl, optimize)) | |
a5b1779f | 1926 | { |
99b1c316 | 1927 | class ipa_fn_summary *info = ipa_fn_summaries->get (node); |
67348ccc | 1928 | struct ipa_dfs_info *dfs = (struct ipa_dfs_info *) node->aux; |
ca31b95f | 1929 | |
5e750dc6 JH |
1930 | /* Do not account external functions, they will be optimized out |
1931 | if not inlined. Also only count the non-cold portion of program. */ | |
bb1e543c | 1932 | if (inline_account_function_p (node)) |
f658ad30 | 1933 | initial_size += ipa_size_summaries->get (node)->size; |
d59171da | 1934 | info->growth = estimate_growth (node); |
41f669d8 JH |
1935 | |
1936 | int num_calls = 0; | |
1937 | node->call_for_symbol_and_aliases (sum_callers, &num_calls, | |
1938 | true); | |
1939 | if (num_calls == 1) | |
1940 | info->single_caller = true; | |
bf3f6510 JH |
1941 | if (dfs && dfs->next_cycle) |
1942 | { | |
1943 | struct cgraph_node *n2; | |
1944 | int id = dfs->scc_no + 1; | |
1945 | for (n2 = node; n2; | |
1a03b929 | 1946 | n2 = ((struct ipa_dfs_info *) n2->aux)->next_cycle) |
29f1e2b1 JH |
1947 | if (opt_for_fn (n2->decl, optimize)) |
1948 | { | |
f658ad30 | 1949 | ipa_fn_summary *info2 = ipa_fn_summaries->get |
a62bfab5 | 1950 | (n2->inlined_to ? n2->inlined_to : n2); |
29f1e2b1 JH |
1951 | if (info2->scc_no) |
1952 | break; | |
1953 | info2->scc_no = id; | |
1954 | } | |
bf3f6510 | 1955 | } |
a5b1779f | 1956 | } |
09a2806f | 1957 | |
e7f23018 | 1958 | for (edge = node->callers; edge; edge = edge->next_caller) |
1bad9c18 | 1959 | max_count = max_count.max (edge->count.ipa ()); |
e7f23018 | 1960 | } |
b48ccf0d | 1961 | ipa_free_postorder_info (); |
ac6f2e59 | 1962 | initialize_growth_caches (); |
b48ccf0d JH |
1963 | |
1964 | if (dump_file) | |
1965 | fprintf (dump_file, | |
1966 | "\nDeciding on inlining of small functions. Starting with size %i.\n", | |
1967 | initial_size); | |
b7c27d51 | 1968 | |
8a8dccb2 | 1969 | overall_size = initial_size; |
85057983 | 1970 | min_size = overall_size; |
1a3118e9 | 1971 | |
a99be3c9 | 1972 | /* Populate the heap with all edges we might inline. */ |
1a3118e9 | 1973 | |
c47d0034 | 1974 | FOR_EACH_DEFINED_FUNCTION (node) |
09ce3660 JH |
1975 | { |
1976 | bool update = false; | |
bcda57c1 | 1977 | struct cgraph_edge *next = NULL; |
c1eed5a1 | 1978 | bool has_speculative = false; |
1a3118e9 | 1979 | |
29f1e2b1 JH |
1980 | if (!opt_for_fn (node->decl, optimize)) |
1981 | continue; | |
1982 | ||
09ce3660 | 1983 | if (dump_file) |
464d0118 | 1984 | fprintf (dump_file, "Enqueueing calls in %s.\n", node->dump_name ()); |
09ce3660 | 1985 | |
8fcfc44f | 1986 | for (edge = node->callees; edge; edge = edge->next_callee) |
09ce3660 | 1987 | { |
1a3118e9 | 1988 | if (edge->inline_failed |
09ce3660 | 1989 | && !edge->aux |
1a3118e9 JH |
1990 | && can_inline_edge_p (edge, true) |
1991 | && want_inline_small_function_p (edge, true) | |
9a4841a3 | 1992 | && can_inline_edge_by_limits_p (edge, true) |
1a3118e9 JH |
1993 | && edge->inline_failed) |
1994 | { | |
1995 | gcc_assert (!edge->aux); | |
4a910049 | 1996 | update_edge_key (&edge_heap, edge); |
1a3118e9 | 1997 | } |
c1eed5a1 JH |
1998 | if (edge->speculative) |
1999 | has_speculative = true; | |
2000 | } | |
2001 | if (has_speculative) | |
2002 | for (edge = node->callees; edge; edge = next) | |
2aae99f7 | 2003 | { |
2e98ac86 | 2004 | next = edge->next_callee; |
2aae99f7 XHL |
2005 | if (edge->speculative |
2006 | && !speculation_useful_p (edge, edge->aux != NULL)) | |
2007 | { | |
27c5a177 | 2008 | cgraph_edge::resolve_speculation (edge); |
2aae99f7 XHL |
2009 | update = true; |
2010 | } | |
2aae99f7 | 2011 | } |
09ce3660 JH |
2012 | if (update) |
2013 | { | |
a62bfab5 ML |
2014 | struct cgraph_node *where = node->inlined_to |
2015 | ? node->inlined_to : node; | |
0bceb671 | 2016 | ipa_update_overall_fn_summary (where); |
09ce3660 | 2017 | reset_edge_caches (where); |
4a910049 | 2018 | update_caller_keys (&edge_heap, where, |
09ce3660 | 2019 | updated_nodes, NULL); |
7c6f2fb9 | 2020 | update_callee_keys (&edge_heap, where, NULL, |
2001028a | 2021 | updated_nodes); |
09ce3660 JH |
2022 | bitmap_clear (updated_nodes); |
2023 | } | |
2024 | } | |
1a3118e9 | 2025 | |
09a2806f | 2026 | gcc_assert (in_lto_p |
3995f3a2 | 2027 | || !(max_count > 0) |
09a2806f | 2028 | || (profile_info && flag_branch_probabilities)); |
b7c27d51 | 2029 | |
4a910049 | 2030 | while (!edge_heap.empty ()) |
ca31b95f | 2031 | { |
85057983 | 2032 | int old_size = overall_size; |
1ce18dc8 | 2033 | struct cgraph_node *where, *callee; |
f0e1509b ML |
2034 | sreal badness = edge_heap.min_key (); |
2035 | sreal current_badness; | |
1ce18dc8 | 2036 | int growth; |
670cd5c5 | 2037 | |
4a910049 | 2038 | edge = edge_heap.extract_min (); |
1ce18dc8 JH |
2039 | gcc_assert (edge->aux); |
2040 | edge->aux = NULL; | |
9de6f6c3 | 2041 | if (!edge->inline_failed || !edge->callee->analyzed) |
1ce18dc8 | 2042 | continue; |
cdc029b9 | 2043 | |
1bad9c18 JH |
2044 | /* Be sure that caches are maintained consistent. |
2045 | This check is affected by scaling roundoff errors when compiling for | |
2046 | IPA this we skip it in that case. */ | |
b914768c | 2047 | if (flag_checking && !edge->callee->count.ipa_p () |
517048ce | 2048 | && (!max_count.initialized_p () || !max_count.nonzero_p ())) |
1bad9c18 JH |
2049 | { |
2050 | sreal cached_badness = edge_badness (edge, false); | |
2051 | ||
2052 | int old_size_est = estimate_edge_size (edge); | |
2053 | sreal old_time_est = estimate_edge_time (edge); | |
2054 | int old_hints_est = estimate_edge_hints (edge); | |
2055 | ||
9fb50ad8 ML |
2056 | if (edge_growth_cache != NULL) |
2057 | edge_growth_cache->remove (edge); | |
b914768c JH |
2058 | reset_node_cache (edge->caller->inlined_to |
2059 | ? edge->caller->inlined_to | |
2060 | : edge->caller); | |
1bad9c18 JH |
2061 | gcc_assert (old_size_est == estimate_edge_size (edge)); |
2062 | gcc_assert (old_time_est == estimate_edge_time (edge)); | |
2063 | /* FIXME: | |
2064 | ||
2065 | gcc_assert (old_hints_est == estimate_edge_hints (edge)); | |
2066 | ||
2067 | fails with profile feedback because some hints depends on | |
2068 | maybe_hot_edge_p predicate and because callee gets inlined to other | |
2069 | calls, the edge may become cold. | |
2070 | This ought to be fixed by computing relative probabilities | |
2071 | for given invocation but that will be better done once whole | |
2072 | code is converted to sreals. Disable for now and revert to "wrong" | |
2073 | value so enable/disable checking paths agree. */ | |
9fb50ad8 | 2074 | edge_growth_cache->get (edge)->hints = old_hints_est + 1; |
1bad9c18 JH |
2075 | |
2076 | /* When updating the edge costs, we only decrease badness in the keys. | |
956d615d | 2077 | Increases of badness are handled lazily; when we see key with out |
1bad9c18 JH |
2078 | of date value on it, we re-insert it now. */ |
2079 | current_badness = edge_badness (edge, false); | |
2080 | gcc_assert (cached_badness == current_badness); | |
2081 | gcc_assert (current_badness >= badness); | |
2082 | } | |
f4118c87 JH |
2083 | else |
2084 | current_badness = edge_badness (edge, false); | |
cdc029b9 JH |
2085 | if (current_badness != badness) |
2086 | { | |
d75de25b | 2087 | if (edge_heap.min () && current_badness > edge_heap.min_key ()) |
6d4ab5f8 JH |
2088 | { |
2089 | edge->aux = edge_heap.insert (current_badness, edge); | |
2090 | continue; | |
2091 | } | |
2092 | else | |
2093 | badness = current_badness; | |
cdc029b9 | 2094 | } |
4c0f7679 | 2095 | |
9a4841a3 JH |
2096 | if (!can_inline_edge_p (edge, true) |
2097 | || !can_inline_edge_by_limits_p (edge, true)) | |
09ce3660 | 2098 | { |
4a910049 | 2099 | resolve_noninline_speculation (&edge_heap, edge); |
09ce3660 JH |
2100 | continue; |
2101 | } | |
cdc029b9 | 2102 | |
d52f5295 | 2103 | callee = edge->callee->ultimate_alias_target (); |
03dfc36d | 2104 | growth = estimate_edge_growth (edge); |
ca31b95f | 2105 | if (dump_file) |
ca31b95f | 2106 | { |
b8698a0f | 2107 | fprintf (dump_file, |
464d0118 ML |
2108 | "\nConsidering %s with %i size\n", |
2109 | callee->dump_name (), | |
f658ad30 | 2110 | ipa_size_summaries->get (callee)->size); |
b8698a0f | 2111 | fprintf (dump_file, |
464d0118 | 2112 | " to be inlined into %s in %s:%i\n" |
6d4ab5f8 | 2113 | " Estimated badness is %f, frequency %.2f.\n", |
464d0118 | 2114 | edge->caller->dump_name (), |
9e145afd | 2115 | edge->call_stmt |
355fe088 | 2116 | && (LOCATION_LOCUS (gimple_location ((const gimple *) |
1b7706c8 JJ |
2117 | edge->call_stmt)) |
2118 | > BUILTINS_LOCATION) | |
355fe088 | 2119 | ? gimple_filename ((const gimple *) edge->call_stmt) |
9e145afd N |
2120 | : "unknown", |
2121 | edge->call_stmt | |
355fe088 | 2122 | ? gimple_lineno ((const gimple *) edge->call_stmt) |
9e145afd | 2123 | : -1, |
6d4ab5f8 | 2124 | badness.to_double (), |
0cea1d34 | 2125 | edge->sreal_frequency ().to_double ()); |
1bad9c18 | 2126 | if (edge->count.ipa ().initialized_p ()) |
3995f3a2 JH |
2127 | { |
2128 | fprintf (dump_file, " Called "); | |
1bad9c18 | 2129 | edge->count.ipa ().dump (dump_file); |
517048ce | 2130 | fprintf (dump_file, " times\n"); |
3995f3a2 | 2131 | } |
1ce18dc8 | 2132 | if (dump_flags & TDF_DETAILS) |
4c0f7679 | 2133 | edge_badness (edge, true); |
ca31b95f JH |
2134 | } |
2135 | ||
1e83bd70 JH |
2136 | where = edge->caller; |
2137 | ||
2138 | if (overall_size + growth > compute_max_insns (where, min_size) | |
67348ccc | 2139 | && !DECL_DISREGARD_INLINE_LIMITS (callee->decl)) |
670cd5c5 | 2140 | { |
4c0f7679 JH |
2141 | edge->inline_failed = CIF_INLINE_UNIT_GROWTH_LIMIT; |
2142 | report_inline_failed_reason (edge); | |
4a910049 | 2143 | resolve_noninline_speculation (&edge_heap, edge); |
670cd5c5 JH |
2144 | continue; |
2145 | } | |
4c0f7679 JH |
2146 | |
2147 | if (!want_inline_small_function_p (edge, true)) | |
09ce3660 | 2148 | { |
4a910049 | 2149 | resolve_noninline_speculation (&edge_heap, edge); |
09ce3660 JH |
2150 | continue; |
2151 | } | |
09a2806f | 2152 | |
7c6f2fb9 JH |
2153 | profile_count old_count = callee->count; |
2154 | ||
7ba03e5e JL |
2155 | /* Heuristics for inlining small functions work poorly for |
2156 | recursive calls where we do effects similar to loop unrolling. | |
2157 | When inlining such edge seems profitable, leave decision on | |
09a2806f | 2158 | specific inliner. */ |
3dafb85c | 2159 | if (edge->recursive_p ()) |
670cd5c5 | 2160 | { |
a62bfab5 ML |
2161 | if (where->inlined_to) |
2162 | where = where->inlined_to; | |
4c0f7679 | 2163 | if (!recursive_inlining (edge, |
2bf86c84 JH |
2164 | opt_for_fn (edge->caller->decl, |
2165 | flag_indirect_inlining) | |
4c0f7679 | 2166 | ? &new_indirect_edges : NULL)) |
d7d1d041 RG |
2167 | { |
2168 | edge->inline_failed = CIF_RECURSIVE_INLINING; | |
4a910049 | 2169 | resolve_noninline_speculation (&edge_heap, edge); |
d7d1d041 RG |
2170 | continue; |
2171 | } | |
40fda55b | 2172 | reset_edge_caches (where); |
09a2806f JH |
2173 | /* Recursive inliner inlines all recursive calls of the function |
2174 | at once. Consequently we need to update all callee keys. */ | |
2bf86c84 | 2175 | if (opt_for_fn (edge->caller->decl, flag_indirect_inlining)) |
4a910049 | 2176 | add_new_edges_to_heap (&edge_heap, new_indirect_edges); |
7c6f2fb9 | 2177 | update_callee_keys (&edge_heap, where, where, updated_nodes); |
09ce3660 | 2178 | bitmap_clear (updated_nodes); |
670cd5c5 JH |
2179 | } |
2180 | else | |
2181 | { | |
4c0f7679 JH |
2182 | struct cgraph_node *outer_node = NULL; |
2183 | int depth = 0; | |
2184 | ||
7ba03e5e JL |
2185 | /* Consider the case where self recursive function A is inlined |
2186 | into B. This is desired optimization in some cases, since it | |
2187 | leads to effect similar of loop peeling and we might completely | |
2188 | optimize out the recursive call. However we must be extra | |
2189 | selective. */ | |
4c0f7679 JH |
2190 | |
2191 | where = edge->caller; | |
a62bfab5 | 2192 | while (where->inlined_to) |
670cd5c5 | 2193 | { |
67348ccc | 2194 | if (where->decl == callee->decl) |
4c0f7679 JH |
2195 | outer_node = where, depth++; |
2196 | where = where->callers->caller; | |
2197 | } | |
2198 | if (outer_node | |
2199 | && !want_inline_self_recursive_call_p (edge, outer_node, | |
2200 | true, depth)) | |
2201 | { | |
2202 | edge->inline_failed | |
67348ccc | 2203 | = (DECL_DISREGARD_INLINE_LIMITS (edge->callee->decl) |
4c0f7679 | 2204 | ? CIF_RECURSIVE_INLINING : CIF_UNSPECIFIED); |
4a910049 | 2205 | resolve_noninline_speculation (&edge_heap, edge); |
670cd5c5 JH |
2206 | continue; |
2207 | } | |
4c0f7679 JH |
2208 | else if (depth && dump_file) |
2209 | fprintf (dump_file, " Peeling recursion with depth %i\n", depth); | |
2210 | ||
a62bfab5 | 2211 | gcc_checking_assert (!callee->inlined_to); |
041cb615 JH |
2212 | |
2213 | int old_size = ipa_size_summaries->get (where)->size; | |
2214 | sreal old_time = ipa_fn_summaries->get (where)->time; | |
2215 | ||
c170d40f | 2216 | inline_call (edge, true, &new_indirect_edges, &overall_size, true); |
1f6f9079 | 2217 | reset_edge_caches (edge->callee); |
8d890d37 | 2218 | add_new_edges_to_heap (&edge_heap, new_indirect_edges); |
40fda55b | 2219 | |
041cb615 | 2220 | /* If caller's size and time increased we do not need to update |
956d615d | 2221 | all edges because badness is not going to decrease. */ |
041cb615 JH |
2222 | if (old_size <= ipa_size_summaries->get (where)->size |
2223 | && old_time <= ipa_fn_summaries->get (where)->time | |
2224 | /* Wrapper penalty may be non-monotonous in this respect. | |
2225 | Fortunately it only affects small functions. */ | |
2226 | && !wrapper_heuristics_may_apply (where, old_size)) | |
7c6f2fb9 JH |
2227 | update_callee_keys (&edge_heap, edge->callee, edge->callee, |
2228 | updated_nodes); | |
041cb615 | 2229 | else |
7c6f2fb9 JH |
2230 | update_callee_keys (&edge_heap, where, |
2231 | edge->callee, | |
2232 | updated_nodes); | |
670cd5c5 JH |
2233 | } |
2234 | where = edge->caller; | |
a62bfab5 ML |
2235 | if (where->inlined_to) |
2236 | where = where->inlined_to; | |
670cd5c5 JH |
2237 | |
2238 | /* Our profitability metric can depend on local properties | |
2239 | such as number of inlinable calls and size of the function body. | |
2240 | After inlining these properties might change for the function we | |
2241 | inlined into (since it's body size changed) and for the functions | |
2242 | called by function we inlined (since number of it inlinable callers | |
2243 | might change). */ | |
4a910049 | 2244 | update_caller_keys (&edge_heap, where, updated_nodes, NULL); |
208e5afa JH |
2245 | /* Offline copy count has possibly changed, recompute if profile is |
2246 | available. */ | |
7c6f2fb9 JH |
2247 | struct cgraph_node *n |
2248 | = cgraph_node::get (edge->callee->decl)->ultimate_alias_target (); | |
2249 | if (n != edge->callee && n->analyzed && !(n->count == old_count) | |
2250 | && n->count.ipa_p ()) | |
2251 | update_callee_keys (&edge_heap, n, NULL, updated_nodes); | |
670cd5c5 | 2252 | bitmap_clear (updated_nodes); |
ca31b95f | 2253 | |
4174a33a | 2254 | if (dump_enabled_p ()) |
50fe876d | 2255 | { |
f658ad30 | 2256 | ipa_fn_summary *s = ipa_fn_summaries->get (where); |
4174a33a DM |
2257 | |
2258 | /* dump_printf can't handle %+i. */ | |
2259 | char buf_net_change[100]; | |
2260 | snprintf (buf_net_change, sizeof buf_net_change, "%+i", | |
2261 | overall_size - old_size); | |
2262 | ||
2263 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, edge->call_stmt, | |
2264 | " Inlined %C into %C which now has time %f and " | |
b74d8dc4 | 2265 | "size %i, net change of %s%s.\n", |
4174a33a | 2266 | edge->callee, edge->caller, |
f658ad30 JH |
2267 | s->time.to_double (), |
2268 | ipa_size_summaries->get (edge->caller)->size, | |
b74d8dc4 JH |
2269 | buf_net_change, |
2270 | cross_module_call_p (edge) ? " (cross module)":""); | |
50fe876d | 2271 | } |
85057983 | 2272 | if (min_size > overall_size) |
b7c27d51 | 2273 | { |
85057983 | 2274 | min_size = overall_size; |
b7c27d51 JH |
2275 | |
2276 | if (dump_file) | |
85057983 | 2277 | fprintf (dump_file, "New minimal size reached: %i\n", min_size); |
b7c27d51 | 2278 | } |
ca31b95f | 2279 | } |
3e293154 | 2280 | |
632b4f8e | 2281 | free_growth_caches (); |
4174a33a DM |
2282 | if (dump_enabled_p ()) |
2283 | dump_printf (MSG_NOTE, | |
2284 | "Unit growth for small function inlining: %i->%i (%i%%)\n", | |
2285 | initial_size, overall_size, | |
2286 | initial_size ? overall_size * 100 / (initial_size) - 100: 0); | |
3dafb85c | 2287 | symtab->remove_edge_removal_hook (edge_removal_hook_holder); |
ca31b95f JH |
2288 | } |
2289 | ||
09a2806f JH |
2290 | /* Flatten NODE. Performed both during early inlining and |
2291 | at IPA inlining time. */ | |
af961c7f RG |
2292 | |
2293 | static void | |
f6e809c8 | 2294 | flatten_function (struct cgraph_node *node, bool early, bool update) |
af961c7f RG |
2295 | { |
2296 | struct cgraph_edge *e; | |
2297 | ||
2298 | /* We shouldn't be called recursively when we are being processed. */ | |
67348ccc | 2299 | gcc_assert (node->aux == NULL); |
af961c7f | 2300 | |
67348ccc | 2301 | node->aux = (void *) node; |
af961c7f RG |
2302 | |
2303 | for (e = node->callees; e; e = e->next_callee) | |
2304 | { | |
2305 | struct cgraph_node *orig_callee; | |
d52f5295 | 2306 | struct cgraph_node *callee = e->callee->ultimate_alias_target (); |
af961c7f | 2307 | |
af961c7f | 2308 | /* We've hit cycle? It is time to give up. */ |
67348ccc | 2309 | if (callee->aux) |
af961c7f | 2310 | { |
4174a33a DM |
2311 | if (dump_enabled_p ()) |
2312 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt, | |
2313 | "Not inlining %C into %C to avoid cycle.\n", | |
2314 | callee, e->caller); | |
a99670f9 JH |
2315 | if (cgraph_inline_failed_type (e->inline_failed) != CIF_FINAL_ERROR) |
2316 | e->inline_failed = CIF_RECURSIVE_INLINING; | |
af961c7f RG |
2317 | continue; |
2318 | } | |
2319 | ||
2320 | /* When the edge is already inlined, we just need to recurse into | |
2321 | it in order to fully flatten the leaves. */ | |
2322 | if (!e->inline_failed) | |
2323 | { | |
f6e809c8 | 2324 | flatten_function (callee, early, false); |
af961c7f RG |
2325 | continue; |
2326 | } | |
2327 | ||
4c0f7679 JH |
2328 | /* Flatten attribute needs to be processed during late inlining. For |
2329 | extra code quality we however do flattening during early optimization, | |
2330 | too. */ | |
632b4f8e | 2331 | if (!early |
4c0f7679 | 2332 | ? !can_inline_edge_p (e, true) |
9a4841a3 | 2333 | && !can_inline_edge_by_limits_p (e, true) |
4c0f7679 JH |
2334 | : !can_early_inline_edge_p (e)) |
2335 | continue; | |
2336 | ||
3dafb85c | 2337 | if (e->recursive_p ()) |
af961c7f | 2338 | { |
4174a33a DM |
2339 | if (dump_enabled_p ()) |
2340 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt, | |
2341 | "Not inlining: recursive call.\n"); | |
af961c7f RG |
2342 | continue; |
2343 | } | |
2344 | ||
67348ccc DM |
2345 | if (gimple_in_ssa_p (DECL_STRUCT_FUNCTION (node->decl)) |
2346 | != gimple_in_ssa_p (DECL_STRUCT_FUNCTION (callee->decl))) | |
59e0c6b7 | 2347 | { |
4174a33a DM |
2348 | if (dump_enabled_p ()) |
2349 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt, | |
2350 | "Not inlining: SSA form does not match.\n"); | |
59e0c6b7 RG |
2351 | continue; |
2352 | } | |
2353 | ||
af961c7f RG |
2354 | /* Inline the edge and flatten the inline clone. Avoid |
2355 | recursing through the original node if the node was cloned. */ | |
4174a33a DM |
2356 | if (dump_enabled_p ()) |
2357 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, e->call_stmt, | |
2358 | " Inlining %C into %C.\n", | |
2359 | callee, e->caller); | |
a5b1779f | 2360 | orig_callee = callee; |
c170d40f | 2361 | inline_call (e, true, NULL, NULL, false); |
af961c7f | 2362 | if (e->callee != orig_callee) |
67348ccc | 2363 | orig_callee->aux = (void *) node; |
f6e809c8 | 2364 | flatten_function (e->callee, early, false); |
af961c7f | 2365 | if (e->callee != orig_callee) |
67348ccc | 2366 | orig_callee->aux = NULL; |
af961c7f RG |
2367 | } |
2368 | ||
67348ccc | 2369 | node->aux = NULL; |
7237f93e JH |
2370 | cgraph_node *where = node->inlined_to ? node->inlined_to : node; |
2371 | if (update && opt_for_fn (where->decl, optimize)) | |
2372 | ipa_update_overall_fn_summary (where); | |
af961c7f RG |
2373 | } |
2374 | ||
a81b0a3d JH |
2375 | /* Inline NODE to all callers. Worker for cgraph_for_node_and_aliases. |
2376 | DATA points to number of calls originally found so we avoid infinite | |
2377 | recursion. */ | |
2378 | ||
2379 | static bool | |
bddead15 RB |
2380 | inline_to_all_callers_1 (struct cgraph_node *node, void *data, |
2381 | hash_set<cgraph_node *> *callers) | |
a81b0a3d JH |
2382 | { |
2383 | int *num_calls = (int *)data; | |
1bbb87c4 JH |
2384 | bool callee_removed = false; |
2385 | ||
a62bfab5 | 2386 | while (node->callers && !node->inlined_to) |
a81b0a3d JH |
2387 | { |
2388 | struct cgraph_node *caller = node->callers->caller; | |
2389 | ||
1af8bfe5 | 2390 | if (!can_inline_edge_p (node->callers, true) |
9a4841a3 | 2391 | || !can_inline_edge_by_limits_p (node->callers, true) |
1af8bfe5 JH |
2392 | || node->callers->recursive_p ()) |
2393 | { | |
2394 | if (dump_file) | |
2395 | fprintf (dump_file, "Uninlinable call found; giving up.\n"); | |
2396 | *num_calls = 0; | |
2397 | return false; | |
2398 | } | |
2399 | ||
a81b0a3d JH |
2400 | if (dump_file) |
2401 | { | |
5e626cd9 | 2402 | cgraph_node *ultimate = node->ultimate_alias_target (); |
a81b0a3d JH |
2403 | fprintf (dump_file, |
2404 | "\nInlining %s size %i.\n", | |
3629ff8a | 2405 | ultimate->dump_name (), |
f658ad30 | 2406 | ipa_size_summaries->get (ultimate)->size); |
a81b0a3d JH |
2407 | fprintf (dump_file, |
2408 | " Called once from %s %i insns.\n", | |
3629ff8a | 2409 | node->callers->caller->dump_name (), |
f658ad30 | 2410 | ipa_size_summaries->get (node->callers->caller)->size); |
a81b0a3d JH |
2411 | } |
2412 | ||
bddead15 RB |
2413 | /* Remember which callers we inlined to, delaying updating the |
2414 | overall summary. */ | |
2415 | callers->add (node->callers->caller); | |
2416 | inline_call (node->callers, true, NULL, NULL, false, &callee_removed); | |
a81b0a3d JH |
2417 | if (dump_file) |
2418 | fprintf (dump_file, | |
2419 | " Inlined into %s which now has %i size\n", | |
3629ff8a | 2420 | caller->dump_name (), |
f658ad30 | 2421 | ipa_size_summaries->get (caller)->size); |
a81b0a3d JH |
2422 | if (!(*num_calls)--) |
2423 | { | |
2424 | if (dump_file) | |
2425 | fprintf (dump_file, "New calls found; giving up.\n"); | |
1bbb87c4 | 2426 | return callee_removed; |
a81b0a3d | 2427 | } |
1bbb87c4 JH |
2428 | if (callee_removed) |
2429 | return true; | |
a81b0a3d JH |
2430 | } |
2431 | return false; | |
2432 | } | |
2433 | ||
bddead15 RB |
2434 | /* Wrapper around inline_to_all_callers_1 doing delayed overall summary |
2435 | update. */ | |
2436 | ||
2437 | static bool | |
2438 | inline_to_all_callers (struct cgraph_node *node, void *data) | |
2439 | { | |
2440 | hash_set<cgraph_node *> callers; | |
2441 | bool res = inline_to_all_callers_1 (node, data, &callers); | |
2442 | /* Perform the delayed update of the overall summary of all callers | |
2443 | processed. This avoids quadratic behavior in the cases where | |
2444 | we have a lot of calls to the same function. */ | |
2445 | for (hash_set<cgraph_node *>::iterator i = callers.begin (); | |
2446 | i != callers.end (); ++i) | |
7237f93e | 2447 | ipa_update_overall_fn_summary ((*i)->inlined_to ? (*i)->inlined_to : *i); |
bddead15 RB |
2448 | return res; |
2449 | } | |
2450 | ||
e86a910f JH |
2451 | /* Output overall time estimate. */ |
2452 | static void | |
2453 | dump_overall_stats (void) | |
2454 | { | |
ab38481c | 2455 | sreal sum_weighted = 0, sum = 0; |
e86a910f JH |
2456 | struct cgraph_node *node; |
2457 | ||
2458 | FOR_EACH_DEFINED_FUNCTION (node) | |
a62bfab5 | 2459 | if (!node->inlined_to |
e86a910f JH |
2460 | && !node->alias) |
2461 | { | |
56f62793 ML |
2462 | ipa_fn_summary *s = ipa_fn_summaries->get (node); |
2463 | if (s != NULL) | |
2464 | { | |
2465 | sum += s->time; | |
2466 | if (node->count.ipa ().initialized_p ()) | |
2467 | sum_weighted += s->time * node->count.ipa ().to_gcov_type (); | |
2468 | } | |
e86a910f JH |
2469 | } |
2470 | fprintf (dump_file, "Overall time estimate: " | |
ab38481c JH |
2471 | "%f weighted by profile: " |
2472 | "%f\n", sum.to_double (), sum_weighted.to_double ()); | |
e86a910f JH |
2473 | } |
2474 | ||
2475 | /* Output some useful stats about inlining. */ | |
2476 | ||
2477 | static void | |
2478 | dump_inline_stats (void) | |
2479 | { | |
a9243bfc RB |
2480 | int64_t inlined_cnt = 0, inlined_indir_cnt = 0; |
2481 | int64_t inlined_virt_cnt = 0, inlined_virt_indir_cnt = 0; | |
2482 | int64_t noninlined_cnt = 0, noninlined_indir_cnt = 0; | |
2483 | int64_t noninlined_virt_cnt = 0, noninlined_virt_indir_cnt = 0; | |
2484 | int64_t inlined_speculative = 0, inlined_speculative_ply = 0; | |
2485 | int64_t indirect_poly_cnt = 0, indirect_cnt = 0; | |
0009a6c3 JH |
2486 | int64_t reason[CIF_N_REASONS][2]; |
2487 | sreal reason_freq[CIF_N_REASONS]; | |
e86a910f JH |
2488 | int i; |
2489 | struct cgraph_node *node; | |
2490 | ||
2491 | memset (reason, 0, sizeof (reason)); | |
0009a6c3 JH |
2492 | for (i=0; i < CIF_N_REASONS; i++) |
2493 | reason_freq[i] = 0; | |
e86a910f JH |
2494 | FOR_EACH_DEFINED_FUNCTION (node) |
2495 | { | |
2496 | struct cgraph_edge *e; | |
2497 | for (e = node->callees; e; e = e->next_callee) | |
2498 | { | |
2499 | if (e->inline_failed) | |
2500 | { | |
1bad9c18 JH |
2501 | if (e->count.ipa ().initialized_p ()) |
2502 | reason[(int) e->inline_failed][0] += e->count.ipa ().to_gcov_type (); | |
0009a6c3 JH |
2503 | reason_freq[(int) e->inline_failed] += e->sreal_frequency (); |
2504 | reason[(int) e->inline_failed][1] ++; | |
3995f3a2 | 2505 | if (DECL_VIRTUAL_P (e->callee->decl) |
1bad9c18 | 2506 | && e->count.ipa ().initialized_p ()) |
e86a910f JH |
2507 | { |
2508 | if (e->indirect_inlining_edge) | |
1bad9c18 | 2509 | noninlined_virt_indir_cnt += e->count.ipa ().to_gcov_type (); |
e86a910f | 2510 | else |
1bad9c18 | 2511 | noninlined_virt_cnt += e->count.ipa ().to_gcov_type (); |
e86a910f | 2512 | } |
1bad9c18 | 2513 | else if (e->count.ipa ().initialized_p ()) |
e86a910f JH |
2514 | { |
2515 | if (e->indirect_inlining_edge) | |
1bad9c18 | 2516 | noninlined_indir_cnt += e->count.ipa ().to_gcov_type (); |
e86a910f | 2517 | else |
1bad9c18 | 2518 | noninlined_cnt += e->count.ipa ().to_gcov_type (); |
e86a910f JH |
2519 | } |
2520 | } | |
1bad9c18 | 2521 | else if (e->count.ipa ().initialized_p ()) |
e86a910f JH |
2522 | { |
2523 | if (e->speculative) | |
2524 | { | |
2525 | if (DECL_VIRTUAL_P (e->callee->decl)) | |
1bad9c18 | 2526 | inlined_speculative_ply += e->count.ipa ().to_gcov_type (); |
e86a910f | 2527 | else |
1bad9c18 | 2528 | inlined_speculative += e->count.ipa ().to_gcov_type (); |
e86a910f JH |
2529 | } |
2530 | else if (DECL_VIRTUAL_P (e->callee->decl)) | |
2531 | { | |
2532 | if (e->indirect_inlining_edge) | |
1bad9c18 | 2533 | inlined_virt_indir_cnt += e->count.ipa ().to_gcov_type (); |
e86a910f | 2534 | else |
1bad9c18 | 2535 | inlined_virt_cnt += e->count.ipa ().to_gcov_type (); |
e86a910f JH |
2536 | } |
2537 | else | |
2538 | { | |
2539 | if (e->indirect_inlining_edge) | |
1bad9c18 | 2540 | inlined_indir_cnt += e->count.ipa ().to_gcov_type (); |
e86a910f | 2541 | else |
1bad9c18 | 2542 | inlined_cnt += e->count.ipa ().to_gcov_type (); |
e86a910f JH |
2543 | } |
2544 | } | |
2545 | } | |
2546 | for (e = node->indirect_calls; e; e = e->next_callee) | |
3995f3a2 | 2547 | if (e->indirect_info->polymorphic |
1bad9c18 JH |
2548 | & e->count.ipa ().initialized_p ()) |
2549 | indirect_poly_cnt += e->count.ipa ().to_gcov_type (); | |
2550 | else if (e->count.ipa ().initialized_p ()) | |
2551 | indirect_cnt += e->count.ipa ().to_gcov_type (); | |
e86a910f | 2552 | } |
3995f3a2 | 2553 | if (max_count.initialized_p ()) |
e86a910f JH |
2554 | { |
2555 | fprintf (dump_file, | |
16998094 JM |
2556 | "Inlined %" PRId64 " + speculative " |
2557 | "%" PRId64 " + speculative polymorphic " | |
2558 | "%" PRId64 " + previously indirect " | |
2559 | "%" PRId64 " + virtual " | |
2560 | "%" PRId64 " + virtual and previously indirect " | |
2561 | "%" PRId64 "\n" "Not inlined " | |
2562 | "%" PRId64 " + previously indirect " | |
2563 | "%" PRId64 " + virtual " | |
2564 | "%" PRId64 " + virtual and previously indirect " | |
956d615d | 2565 | "%" PRId64 " + still indirect " |
16998094 JM |
2566 | "%" PRId64 " + still indirect polymorphic " |
2567 | "%" PRId64 "\n", inlined_cnt, | |
e86a910f JH |
2568 | inlined_speculative, inlined_speculative_ply, |
2569 | inlined_indir_cnt, inlined_virt_cnt, inlined_virt_indir_cnt, | |
2570 | noninlined_cnt, noninlined_indir_cnt, noninlined_virt_cnt, | |
2571 | noninlined_virt_indir_cnt, indirect_cnt, indirect_poly_cnt); | |
3995f3a2 JH |
2572 | fprintf (dump_file, "Removed speculations "); |
2573 | spec_rem.dump (dump_file); | |
2574 | fprintf (dump_file, "\n"); | |
e86a910f JH |
2575 | } |
2576 | dump_overall_stats (); | |
2577 | fprintf (dump_file, "\nWhy inlining failed?\n"); | |
2578 | for (i = 0; i < CIF_N_REASONS; i++) | |
0009a6c3 JH |
2579 | if (reason[i][1]) |
2580 | fprintf (dump_file, "%-50s: %8i calls, %8f freq, %" PRId64" count\n", | |
e86a910f | 2581 | cgraph_inline_failed_string ((cgraph_inline_failed_t) i), |
0009a6c3 | 2582 | (int) reason[i][1], reason_freq[i].to_double (), reason[i][0]); |
e86a910f JH |
2583 | } |
2584 | ||
d6ea70a0 JJ |
2585 | /* Called when node is removed. */ |
2586 | ||
2587 | static void | |
2588 | flatten_remove_node_hook (struct cgraph_node *node, void *data) | |
2589 | { | |
2590 | if (lookup_attribute ("flatten", DECL_ATTRIBUTES (node->decl)) == NULL) | |
2591 | return; | |
2592 | ||
2593 | hash_set<struct cgraph_node *> *removed | |
2594 | = (hash_set<struct cgraph_node *> *) data; | |
2595 | removed->add (node); | |
2596 | } | |
2597 | ||
ca31b95f JH |
2598 | /* Decide on the inlining. We do so in the topological order to avoid |
2599 | expenses on updating data structures. */ | |
2600 | ||
c2924966 | 2601 | static unsigned int |
4c0f7679 | 2602 | ipa_inline (void) |
ca31b95f JH |
2603 | { |
2604 | struct cgraph_node *node; | |
2605 | int nnodes; | |
b591a8b7 | 2606 | struct cgraph_node **order; |
d6ea70a0 | 2607 | int i, j; |
09ce3660 | 2608 | int cold; |
8a41354f JH |
2609 | bool remove_functions = false; |
2610 | ||
3dafb85c | 2611 | order = XCNEWVEC (struct cgraph_node *, symtab->cgraph_count); |
b591a8b7 | 2612 | |
10a5dd5d | 2613 | if (dump_file) |
0bceb671 | 2614 | ipa_dump_fn_summaries (dump_file); |
670cd5c5 | 2615 | |
af8bca3c | 2616 | nnodes = ipa_reverse_postorder (order); |
e7a74006 | 2617 | spec_rem = profile_count::zero (); |
ca31b95f | 2618 | |
65c70e6b | 2619 | FOR_EACH_FUNCTION (node) |
7ce7e4d4 JH |
2620 | { |
2621 | node->aux = 0; | |
2622 | ||
2623 | /* Recompute the default reasons for inlining because they may have | |
2624 | changed during merging. */ | |
2625 | if (in_lto_p) | |
2626 | { | |
2627 | for (cgraph_edge *e = node->callees; e; e = e->next_callee) | |
2628 | { | |
2629 | gcc_assert (e->inline_failed); | |
2630 | initialize_inline_failed (e); | |
2631 | } | |
2632 | for (cgraph_edge *e = node->indirect_calls; e; e = e->next_callee) | |
2633 | initialize_inline_failed (e); | |
2634 | } | |
2635 | } | |
ca31b95f JH |
2636 | |
2637 | if (dump_file) | |
af961c7f | 2638 | fprintf (dump_file, "\nFlattening functions:\n"); |
ca31b95f | 2639 | |
d6ea70a0 JJ |
2640 | /* First shrink order array, so that it only contains nodes with |
2641 | flatten attribute. */ | |
2642 | for (i = nnodes - 1, j = i; i >= 0; i--) | |
2643 | { | |
2644 | node = order[i]; | |
2895b172 | 2645 | if (node->definition |
f22712bd JH |
2646 | /* Do not try to flatten aliases. These may happen for example when |
2647 | creating local aliases. */ | |
2648 | && !node->alias | |
2895b172 JH |
2649 | && lookup_attribute ("flatten", |
2650 | DECL_ATTRIBUTES (node->decl)) != NULL) | |
d6ea70a0 JJ |
2651 | order[j--] = order[i]; |
2652 | } | |
2653 | ||
2654 | /* After the above loop, order[j + 1] ... order[nnodes - 1] contain | |
2655 | nodes with flatten attribute. If there is more than one such | |
2656 | node, we need to register a node removal hook, as flatten_function | |
2657 | could remove other nodes with flatten attribute. See PR82801. */ | |
2658 | struct cgraph_node_hook_list *node_removal_hook_holder = NULL; | |
2659 | hash_set<struct cgraph_node *> *flatten_removed_nodes = NULL; | |
2660 | if (j < nnodes - 2) | |
2661 | { | |
2662 | flatten_removed_nodes = new hash_set<struct cgraph_node *>; | |
2663 | node_removal_hook_holder | |
2664 | = symtab->add_cgraph_removal_hook (&flatten_remove_node_hook, | |
2665 | flatten_removed_nodes); | |
2666 | } | |
2667 | ||
af961c7f RG |
2668 | /* In the first pass handle functions to be flattened. Do this with |
2669 | a priority so none of our later choices will make this impossible. */ | |
d6ea70a0 | 2670 | for (i = nnodes - 1; i > j; i--) |
ca31b95f | 2671 | { |
af961c7f | 2672 | node = order[i]; |
d6ea70a0 JJ |
2673 | if (flatten_removed_nodes |
2674 | && flatten_removed_nodes->contains (node)) | |
2675 | continue; | |
af961c7f | 2676 | |
09a2806f | 2677 | /* Handle nodes to be flattened. |
af961c7f RG |
2678 | Ideally when processing callees we stop inlining at the |
2679 | entry of cycles, possibly cloning that entry point and | |
2680 | try to flatten itself turning it into a self-recursive | |
2681 | function. */ | |
d6ea70a0 | 2682 | if (dump_file) |
3629ff8a | 2683 | fprintf (dump_file, "Flattening %s\n", node->dump_name ()); |
f6e809c8 | 2684 | flatten_function (node, false, true); |
d6ea70a0 JJ |
2685 | } |
2686 | ||
2687 | if (j < nnodes - 2) | |
2688 | { | |
2689 | symtab->remove_cgraph_removal_hook (node_removal_hook_holder); | |
2690 | delete flatten_removed_nodes; | |
ca31b95f | 2691 | } |
d6ea70a0 JJ |
2692 | free (order); |
2693 | ||
e86a910f JH |
2694 | if (dump_file) |
2695 | dump_overall_stats (); | |
ca31b95f | 2696 | |
4c0f7679 | 2697 | inline_small_functions (); |
e70670cf | 2698 | |
17e0fc92 JH |
2699 | gcc_assert (symtab->state == IPA_SSA); |
2700 | symtab->state = IPA_SSA_AFTER_INLINING; | |
2701 | /* Do first after-inlining removal. We want to remove all "stale" extern | |
2702 | inline functions and virtual functions so we really know what is called | |
2703 | once. */ | |
2704 | symtab->remove_unreachable_nodes (dump_file); | |
ca31b95f | 2705 | |
100411f8 JH |
2706 | /* Inline functions with a property that after inlining into all callers the |
2707 | code size will shrink because the out-of-line copy is eliminated. | |
2708 | We do this regardless on the callee size as long as function growth limits | |
2709 | are met. */ | |
09ce3660 JH |
2710 | if (dump_file) |
2711 | fprintf (dump_file, | |
17e0fc92 JH |
2712 | "\nDeciding on functions to be inlined into all callers and " |
2713 | "removing useless speculations:\n"); | |
09ce3660 JH |
2714 | |
2715 | /* Inlining one function called once has good chance of preventing | |
2716 | inlining other function into the same callee. Ideally we should | |
2717 | work in priority order, but probably inlining hot functions first | |
2718 | is good cut without the extra pain of maintaining the queue. | |
2719 | ||
2720 | ??? this is not really fitting the bill perfectly: inlining function | |
2721 | into callee often leads to better optimization of callee due to | |
2722 | increased context for optimization. | |
2723 | For example if main() function calls a function that outputs help | |
956d615d | 2724 | and then function that does the main optimization, we should inline |
09ce3660 JH |
2725 | the second with priority even if both calls are cold by themselves. |
2726 | ||
2727 | We probably want to implement new predicate replacing our use of | |
2728 | maybe_hot_edge interpreted as maybe_hot_edge || callee is known | |
2729 | to be hot. */ | |
2730 | for (cold = 0; cold <= 1; cold ++) | |
355866de | 2731 | { |
09ce3660 | 2732 | FOR_EACH_DEFINED_FUNCTION (node) |
ca31b95f | 2733 | { |
09ce3660 JH |
2734 | struct cgraph_edge *edge, *next; |
2735 | bool update=false; | |
2736 | ||
29f1e2b1 JH |
2737 | if (!opt_for_fn (node->decl, optimize) |
2738 | || !opt_for_fn (node->decl, flag_inline_functions_called_once)) | |
2739 | continue; | |
2740 | ||
09ce3660 | 2741 | for (edge = node->callees; edge; edge = next) |
ca31b95f | 2742 | { |
09ce3660 JH |
2743 | next = edge->next_callee; |
2744 | if (edge->speculative && !speculation_useful_p (edge, false)) | |
e3c7b49c | 2745 | { |
1bad9c18 JH |
2746 | if (edge->count.ipa ().initialized_p ()) |
2747 | spec_rem += edge->count.ipa (); | |
27c5a177 | 2748 | cgraph_edge::resolve_speculation (edge); |
09ce3660 | 2749 | update = true; |
8a41354f | 2750 | remove_functions = true; |
09ce3660 JH |
2751 | } |
2752 | } | |
2753 | if (update) | |
2754 | { | |
a62bfab5 ML |
2755 | struct cgraph_node *where = node->inlined_to |
2756 | ? node->inlined_to : node; | |
09ce3660 | 2757 | reset_edge_caches (where); |
0bceb671 | 2758 | ipa_update_overall_fn_summary (where); |
09ce3660 | 2759 | } |
2bf86c84 | 2760 | if (want_inline_function_to_all_callers_p (node, cold)) |
09ce3660 JH |
2761 | { |
2762 | int num_calls = 0; | |
1ede94c5 JH |
2763 | node->call_for_symbol_and_aliases (sum_callers, &num_calls, |
2764 | true); | |
2765 | while (node->call_for_symbol_and_aliases | |
17e0fc92 | 2766 | (inline_to_all_callers, &num_calls, true)) |
1bbb87c4 | 2767 | ; |
f019b607 | 2768 | remove_functions = true; |
ca31b95f JH |
2769 | } |
2770 | } | |
2771 | } | |
2772 | ||
f8e2a1ed | 2773 | /* Free ipa-prop structures if they are no longer needed. */ |
29f1e2b1 | 2774 | ipa_free_all_structures_after_iinln (); |
f8e2a1ed | 2775 | |
4174a33a DM |
2776 | if (dump_enabled_p ()) |
2777 | dump_printf (MSG_NOTE, | |
2778 | "\nInlined %i calls, eliminated %i functions\n\n", | |
2779 | ncalls_inlined, nfunctions_inlined); | |
ca31b95f | 2780 | if (dump_file) |
4174a33a | 2781 | dump_inline_stats (); |
09a2806f | 2782 | |
898b8927 | 2783 | if (dump_file) |
0bceb671 | 2784 | ipa_dump_fn_summaries (dump_file); |
8a41354f | 2785 | return remove_functions ? TODO_remove_functions : 0; |
ca31b95f JH |
2786 | } |
2787 | ||
275b4baa RG |
2788 | /* Inline always-inline function calls in NODE. */ |
2789 | ||
2790 | static bool | |
4c0f7679 | 2791 | inline_always_inline_functions (struct cgraph_node *node) |
275b4baa RG |
2792 | { |
2793 | struct cgraph_edge *e; | |
2794 | bool inlined = false; | |
2795 | ||
2796 | for (e = node->callees; e; e = e->next_callee) | |
2797 | { | |
d52f5295 | 2798 | struct cgraph_node *callee = e->callee->ultimate_alias_target (); |
67348ccc | 2799 | if (!DECL_DISREGARD_INLINE_LIMITS (callee->decl)) |
275b4baa RG |
2800 | continue; |
2801 | ||
3dafb85c | 2802 | if (e->recursive_p ()) |
275b4baa | 2803 | { |
4174a33a DM |
2804 | if (dump_enabled_p ()) |
2805 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt, | |
2806 | " Not inlining recursive call to %C.\n", | |
2807 | e->callee); | |
275b4baa RG |
2808 | e->inline_failed = CIF_RECURSIVE_INLINING; |
2809 | continue; | |
2810 | } | |
2811 | ||
4c0f7679 | 2812 | if (!can_early_inline_edge_p (e)) |
bef8491a ST |
2813 | { |
2814 | /* Set inlined to true if the callee is marked "always_inline" but | |
2815 | is not inlinable. This will allow flagging an error later in | |
2816 | expand_call_inline in tree-inline.c. */ | |
2817 | if (lookup_attribute ("always_inline", | |
67348ccc | 2818 | DECL_ATTRIBUTES (callee->decl)) != NULL) |
bef8491a ST |
2819 | inlined = true; |
2820 | continue; | |
2821 | } | |
275b4baa | 2822 | |
4174a33a DM |
2823 | if (dump_enabled_p ()) |
2824 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, e->call_stmt, | |
2825 | " Inlining %C into %C (always_inline).\n", | |
2826 | e->callee, e->caller); | |
c170d40f | 2827 | inline_call (e, true, NULL, NULL, false); |
275b4baa RG |
2828 | inlined = true; |
2829 | } | |
c170d40f | 2830 | if (inlined) |
0bceb671 | 2831 | ipa_update_overall_fn_summary (node); |
275b4baa RG |
2832 | |
2833 | return inlined; | |
2834 | } | |
2835 | ||
ca31b95f | 2836 | /* Decide on the inlining. We do so in the topological order to avoid |
af961c7f | 2837 | expenses on updating data structures. */ |
ca31b95f | 2838 | |
7fa49e7b | 2839 | static bool |
4c0f7679 | 2840 | early_inline_small_functions (struct cgraph_node *node) |
ca31b95f JH |
2841 | { |
2842 | struct cgraph_edge *e; | |
d63db217 | 2843 | bool inlined = false; |
7fa49e7b | 2844 | |
275b4baa | 2845 | for (e = node->callees; e; e = e->next_callee) |
c3056c2d | 2846 | { |
d52f5295 | 2847 | struct cgraph_node *callee = e->callee->ultimate_alias_target (); |
56f62793 | 2848 | |
956d615d | 2849 | /* We can encounter not-yet-analyzed function during |
56f62793 ML |
2850 | early inlining on callgraphs with strongly |
2851 | connected components. */ | |
2852 | ipa_fn_summary *s = ipa_fn_summaries->get (callee); | |
2853 | if (s == NULL || !s->inlinable || !e->inline_failed) | |
275b4baa RG |
2854 | continue; |
2855 | ||
2856 | /* Do not consider functions not declared inline. */ | |
67348ccc | 2857 | if (!DECL_DECLARED_INLINE_P (callee->decl) |
2bf86c84 JH |
2858 | && !opt_for_fn (node->decl, flag_inline_small_functions) |
2859 | && !opt_for_fn (node->decl, flag_inline_functions)) | |
275b4baa RG |
2860 | continue; |
2861 | ||
4174a33a DM |
2862 | if (dump_enabled_p ()) |
2863 | dump_printf_loc (MSG_NOTE, e->call_stmt, | |
2864 | "Considering inline candidate %C.\n", | |
2865 | callee); | |
ca31b95f | 2866 | |
4c0f7679 JH |
2867 | if (!can_early_inline_edge_p (e)) |
2868 | continue; | |
2869 | ||
3dafb85c | 2870 | if (e->recursive_p ()) |
275b4baa | 2871 | { |
4174a33a DM |
2872 | if (dump_enabled_p ()) |
2873 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt, | |
2874 | " Not inlining: recursive call.\n"); | |
22ad64b6 | 2875 | continue; |
275b4baa | 2876 | } |
af961c7f | 2877 | |
4c0f7679 | 2878 | if (!want_early_inline_function_p (e)) |
275b4baa | 2879 | continue; |
ca31b95f | 2880 | |
4174a33a DM |
2881 | if (dump_enabled_p ()) |
2882 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, e->call_stmt, | |
2883 | " Inlining %C into %C.\n", | |
2884 | callee, e->caller); | |
bddead15 | 2885 | inline_call (e, true, NULL, NULL, false); |
4c0f7679 | 2886 | inlined = true; |
38bc76da | 2887 | } |
275b4baa | 2888 | |
bddead15 | 2889 | if (inlined) |
0bceb671 | 2890 | ipa_update_overall_fn_summary (node); |
bddead15 | 2891 | |
7fa49e7b | 2892 | return inlined; |
ca31b95f JH |
2893 | } |
2894 | ||
be55bfe6 | 2895 | unsigned int |
be3c16c4 | 2896 | early_inliner (function *fun) |
d63db217 | 2897 | { |
d52f5295 | 2898 | struct cgraph_node *node = cgraph_node::get (current_function_decl); |
10a5dd5d | 2899 | struct cgraph_edge *edge; |
7fa49e7b | 2900 | unsigned int todo = 0; |
796bda22 | 2901 | int iterations = 0; |
275b4baa | 2902 | bool inlined = false; |
d63db217 | 2903 | |
1da2ed5f | 2904 | if (seen_error ()) |
c2924966 | 2905 | return 0; |
af961c7f | 2906 | |
ecb62563 JH |
2907 | /* Do nothing if datastructures for ipa-inliner are already computed. This |
2908 | happens when some pass decides to construct new function and | |
2909 | cgraph_add_new_function calls lowering passes and early optimization on | |
2910 | it. This may confuse ourself when early inliner decide to inline call to | |
2911 | function clone, because function clones don't have parameter list in | |
2912 | ipa-prop matching their signature. */ | |
dd912cb8 | 2913 | if (ipa_node_params_sum) |
ecb62563 JH |
2914 | return 0; |
2915 | ||
b2b29377 MM |
2916 | if (flag_checking) |
2917 | node->verify (); | |
d122681a | 2918 | node->remove_all_references (); |
275b4baa RG |
2919 | |
2920 | /* Even when not optimizing or not inlining inline always-inline | |
2921 | functions. */ | |
4c0f7679 | 2922 | inlined = inline_always_inline_functions (node); |
275b4baa | 2923 | |
af961c7f RG |
2924 | if (!optimize |
2925 | || flag_no_inline | |
4c0f7679 JH |
2926 | || !flag_early_inlining |
2927 | /* Never inline regular functions into always-inline functions | |
2928 | during incremental inlining. This sucks as functions calling | |
2929 | always inline functions will get less optimized, but at the | |
2930 | same time inlining of functions calling always inline | |
09a2806f | 2931 | function into an always inline function might introduce |
4c0f7679 JH |
2932 | cycles of edges to be always inlined in the callgraph. |
2933 | ||
2934 | We might want to be smarter and just avoid this type of inlining. */ | |
d67bce7c JH |
2935 | || (DECL_DISREGARD_INLINE_LIMITS (node->decl) |
2936 | && lookup_attribute ("always_inline", | |
2937 | DECL_ATTRIBUTES (node->decl)))) | |
275b4baa RG |
2938 | ; |
2939 | else if (lookup_attribute ("flatten", | |
67348ccc | 2940 | DECL_ATTRIBUTES (node->decl)) != NULL) |
7fa49e7b | 2941 | { |
275b4baa RG |
2942 | /* When the function is marked to be flattened, recursively inline |
2943 | all calls in it. */ | |
4174a33a DM |
2944 | if (dump_enabled_p ()) |
2945 | dump_printf (MSG_OPTIMIZED_LOCATIONS, | |
2946 | "Flattening %C\n", node); | |
f6e809c8 | 2947 | flatten_function (node, true, true); |
275b4baa | 2948 | inlined = true; |
7fa49e7b | 2949 | } |
af961c7f RG |
2950 | else |
2951 | { | |
d67bce7c JH |
2952 | /* If some always_inline functions was inlined, apply the changes. |
2953 | This way we will not account always inline into growth limits and | |
2954 | moreover we will inline calls from always inlines that we skipped | |
56aae4b7 | 2955 | previously because of conditional above. */ |
d67bce7c JH |
2956 | if (inlined) |
2957 | { | |
2958 | timevar_push (TV_INTEGRATION); | |
2959 | todo |= optimize_inline_calls (current_function_decl); | |
1cf06f1e MP |
2960 | /* optimize_inline_calls call above might have introduced new |
2961 | statements that don't have inline parameters computed. */ | |
2962 | for (edge = node->callees; edge; edge = edge->next_callee) | |
2963 | { | |
56f62793 ML |
2964 | /* We can enounter not-yet-analyzed function during |
2965 | early inlining on callgraphs with strongly | |
2966 | connected components. */ | |
99353fcf | 2967 | ipa_call_summary *es = ipa_call_summaries->get_create (edge); |
263e19c7 JH |
2968 | es->call_stmt_size |
2969 | = estimate_num_insns (edge->call_stmt, &eni_size_weights); | |
2970 | es->call_stmt_time | |
2971 | = estimate_num_insns (edge->call_stmt, &eni_time_weights); | |
1cf06f1e | 2972 | } |
0bceb671 | 2973 | ipa_update_overall_fn_summary (node); |
d67bce7c JH |
2974 | inlined = false; |
2975 | timevar_pop (TV_INTEGRATION); | |
2976 | } | |
af961c7f RG |
2977 | /* We iterate incremental inlining to get trivial cases of indirect |
2978 | inlining. */ | |
fdfd7f53 ML |
2979 | while (iterations < opt_for_fn (node->decl, |
2980 | param_early_inliner_max_iterations) | |
4c0f7679 | 2981 | && early_inline_small_functions (node)) |
af961c7f RG |
2982 | { |
2983 | timevar_push (TV_INTEGRATION); | |
2984 | todo |= optimize_inline_calls (current_function_decl); | |
4c0f7679 JH |
2985 | |
2986 | /* Technically we ought to recompute inline parameters so the new | |
2987 | iteration of early inliner works as expected. We however have | |
2988 | values approximately right and thus we only need to update edge | |
2989 | info that might be cleared out for newly discovered edges. */ | |
2990 | for (edge = node->callees; edge; edge = edge->next_callee) | |
2991 | { | |
d5e254e1 | 2992 | /* We have no summary for new bound store calls yet. */ |
b412559e ML |
2993 | ipa_call_summary *es = ipa_call_summaries->get_create (edge); |
2994 | es->call_stmt_size | |
2995 | = estimate_num_insns (edge->call_stmt, &eni_size_weights); | |
2996 | es->call_stmt_time | |
2997 | = estimate_num_insns (edge->call_stmt, &eni_time_weights); | |
4c0f7679 | 2998 | } |
fdfd7f53 ML |
2999 | if (iterations < opt_for_fn (node->decl, |
3000 | param_early_inliner_max_iterations) - 1) | |
0bceb671 | 3001 | ipa_update_overall_fn_summary (node); |
af961c7f | 3002 | timevar_pop (TV_INTEGRATION); |
275b4baa RG |
3003 | iterations++; |
3004 | inlined = false; | |
af961c7f RG |
3005 | } |
3006 | if (dump_file) | |
3007 | fprintf (dump_file, "Iterations: %i\n", iterations); | |
3008 | } | |
3009 | ||
275b4baa RG |
3010 | if (inlined) |
3011 | { | |
3012 | timevar_push (TV_INTEGRATION); | |
3013 | todo |= optimize_inline_calls (current_function_decl); | |
3014 | timevar_pop (TV_INTEGRATION); | |
3015 | } | |
3016 | ||
be55bfe6 | 3017 | fun->always_inline_functions_inlined = true; |
d63db217 | 3018 | |
af961c7f | 3019 | return todo; |
d63db217 JH |
3020 | } |
3021 | ||
be3c16c4 DC |
3022 | /* Do inlining of small functions. Doing so early helps profiling and other |
3023 | passes to be somewhat more effective and avoids some code duplication in | |
3024 | later real inlining pass for testcases with very many function calls. */ | |
3025 | ||
3026 | namespace { | |
3027 | ||
3028 | const pass_data pass_data_early_inline = | |
3029 | { | |
3030 | GIMPLE_PASS, /* type */ | |
3031 | "einline", /* name */ | |
3032 | OPTGROUP_INLINE, /* optinfo_flags */ | |
3033 | TV_EARLY_INLINING, /* tv_id */ | |
3034 | PROP_ssa, /* properties_required */ | |
3035 | 0, /* properties_provided */ | |
3036 | 0, /* properties_destroyed */ | |
3037 | 0, /* todo_flags_start */ | |
3038 | 0, /* todo_flags_finish */ | |
3039 | }; | |
3040 | ||
3041 | class pass_early_inline : public gimple_opt_pass | |
3042 | { | |
3043 | public: | |
3044 | pass_early_inline (gcc::context *ctxt) | |
3045 | : gimple_opt_pass (pass_data_early_inline, ctxt) | |
3046 | {} | |
3047 | ||
3048 | /* opt_pass methods: */ | |
3049 | virtual unsigned int execute (function *); | |
3050 | ||
3051 | }; // class pass_early_inline | |
3052 | ||
3053 | unsigned int | |
3054 | pass_early_inline::execute (function *fun) | |
3055 | { | |
3056 | return early_inliner (fun); | |
3057 | } | |
3058 | ||
27a4cd48 DM |
3059 | } // anon namespace |
3060 | ||
3061 | gimple_opt_pass * | |
3062 | make_pass_early_inline (gcc::context *ctxt) | |
3063 | { | |
3064 | return new pass_early_inline (ctxt); | |
3065 | } | |
3066 | ||
27a4cd48 DM |
3067 | namespace { |
3068 | ||
3069 | const pass_data pass_data_ipa_inline = | |
873aa8f5 | 3070 | { |
27a4cd48 DM |
3071 | IPA_PASS, /* type */ |
3072 | "inline", /* name */ | |
3073 | OPTGROUP_INLINE, /* optinfo_flags */ | |
27a4cd48 DM |
3074 | TV_IPA_INLINING, /* tv_id */ |
3075 | 0, /* properties_required */ | |
3076 | 0, /* properties_provided */ | |
3077 | 0, /* properties_destroyed */ | |
8605403e | 3078 | 0, /* todo_flags_start */ |
8a41354f | 3079 | ( TODO_dump_symtab ), /* todo_flags_finish */ |
ca31b95f | 3080 | }; |
27a4cd48 DM |
3081 | |
3082 | class pass_ipa_inline : public ipa_opt_pass_d | |
3083 | { | |
3084 | public: | |
c3284718 RS |
3085 | pass_ipa_inline (gcc::context *ctxt) |
3086 | : ipa_opt_pass_d (pass_data_ipa_inline, ctxt, | |
d2db2e6b JH |
3087 | NULL, /* generate_summary */ |
3088 | NULL, /* write_summary */ | |
3089 | NULL, /* read_summary */ | |
c3284718 RS |
3090 | NULL, /* write_optimization_summary */ |
3091 | NULL, /* read_optimization_summary */ | |
3092 | NULL, /* stmt_fixup */ | |
3093 | 0, /* function_transform_todo_flags_start */ | |
3094 | inline_transform, /* function_transform */ | |
3095 | NULL) /* variable_transform */ | |
27a4cd48 DM |
3096 | {} |
3097 | ||
3098 | /* opt_pass methods: */ | |
be55bfe6 | 3099 | virtual unsigned int execute (function *) { return ipa_inline (); } |
27a4cd48 DM |
3100 | |
3101 | }; // class pass_ipa_inline | |
3102 | ||
3103 | } // anon namespace | |
3104 | ||
3105 | ipa_opt_pass_d * | |
3106 | make_pass_ipa_inline (gcc::context *ctxt) | |
3107 | { | |
3108 | return new pass_ipa_inline (ctxt); | |
3109 | } |