]>
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 |
1e83bd70 | 187 | && newsize > opt_for_fn (to->decl, param_large_function_insns) |
ca31b95f JH |
188 | && newsize > limit) |
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 | ||
6206a883 JJ |
267 | /* Allow inlining always_inline functions into no_sanitize_address |
268 | functions. */ | |
269 | if (!sanitize_flags_p (SANITIZE_ADDRESS, caller) | |
270 | && lookup_attribute ("always_inline", DECL_ATTRIBUTES (callee))) | |
271 | return true; | |
272 | ||
05abad4c ML |
273 | return ((sanitize_flags_p (SANITIZE_ADDRESS, caller) |
274 | == sanitize_flags_p (SANITIZE_ADDRESS, callee)) | |
275 | && (sanitize_flags_p (SANITIZE_POINTER_COMPARE, caller) | |
276 | == sanitize_flags_p (SANITIZE_POINTER_COMPARE, callee)) | |
277 | && (sanitize_flags_p (SANITIZE_POINTER_SUBTRACT, caller) | |
278 | == sanitize_flags_p (SANITIZE_POINTER_SUBTRACT, callee))); | |
25a07c7e YG |
279 | } |
280 | ||
8e926cb1 JH |
281 | /* Used for flags where it is safe to inline when caller's value is |
282 | grater than callee's. */ | |
283 | #define check_maybe_up(flag) \ | |
284 | (opts_for_fn (caller->decl)->x_##flag \ | |
285 | != opts_for_fn (callee->decl)->x_##flag \ | |
286 | && (!always_inline \ | |
287 | || opts_for_fn (caller->decl)->x_##flag \ | |
288 | < opts_for_fn (callee->decl)->x_##flag)) | |
289 | /* Used for flags where it is safe to inline when caller's value is | |
290 | smaller than callee's. */ | |
291 | #define check_maybe_down(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 exact match is needed for correctness. */ | |
298 | #define check_match(flag) \ | |
299 | (opts_for_fn (caller->decl)->x_##flag \ | |
300 | != opts_for_fn (callee->decl)->x_##flag) | |
301 | ||
9a4841a3 | 302 | /* Decide if we can inline the edge and possibly update |
4c0f7679 JH |
303 | inline_failed reason. |
304 | We check whether inlining is possible at all and whether | |
305 | caller growth limits allow doing so. | |
306 | ||
9a4841a3 | 307 | if REPORT is true, output reason to the dump file. */ |
ca31b95f | 308 | |
61a05df1 | 309 | static bool |
09ce3660 | 310 | can_inline_edge_p (struct cgraph_edge *e, bool report, |
9a4841a3 | 311 | bool early = false) |
ca31b95f | 312 | { |
7ce7e4d4 JH |
313 | gcc_checking_assert (e->inline_failed); |
314 | ||
315 | if (cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR) | |
316 | { | |
317 | if (report) | |
318 | report_inline_failed_reason (e); | |
319 | return false; | |
320 | } | |
321 | ||
4c0f7679 | 322 | bool inlinable = true; |
a5b1779f | 323 | enum availability avail; |
a62bfab5 ML |
324 | cgraph_node *caller = (e->caller->inlined_to |
325 | ? e->caller->inlined_to : e->caller); | |
e6007a27 | 326 | cgraph_node *callee = e->callee->ultimate_alias_target (&avail, caller); |
ea99e0be | 327 | |
7ce7e4d4 | 328 | if (!callee->definition) |
4c0f7679 JH |
329 | { |
330 | e->inline_failed = CIF_BODY_NOT_AVAILABLE; | |
331 | inlinable = false; | |
332 | } | |
8a4a6d2e | 333 | if (!early && (!opt_for_fn (callee->decl, optimize) |
12b8cb2e | 334 | || !opt_for_fn (caller->decl, optimize))) |
29f1e2b1 JH |
335 | { |
336 | e->inline_failed = CIF_FUNCTION_NOT_OPTIMIZED; | |
337 | inlinable = false; | |
338 | } | |
1f26ac87 JM |
339 | else if (callee->calls_comdat_local) |
340 | { | |
341 | e->inline_failed = CIF_USES_COMDAT_LOCAL; | |
342 | inlinable = false; | |
343 | } | |
d52f5295 | 344 | else if (avail <= AVAIL_INTERPOSABLE) |
9de21a23 | 345 | { |
4c0f7679 | 346 | e->inline_failed = CIF_OVERWRITABLE; |
6957a6f6 | 347 | inlinable = false; |
9de21a23 | 348 | } |
1a0bf5e1 JH |
349 | /* All edges with call_stmt_cannot_inline_p should have inline_failed |
350 | initialized to one of FINAL_ERROR reasons. */ | |
89faf322 | 351 | else if (e->call_stmt_cannot_inline_p) |
1a0bf5e1 | 352 | gcc_unreachable (); |
4c0f7679 | 353 | /* Don't inline if the functions have different EH personalities. */ |
bb1e543c | 354 | else if (DECL_FUNCTION_PERSONALITY (caller->decl) |
67348ccc | 355 | && DECL_FUNCTION_PERSONALITY (callee->decl) |
bb1e543c | 356 | && (DECL_FUNCTION_PERSONALITY (caller->decl) |
67348ccc | 357 | != DECL_FUNCTION_PERSONALITY (callee->decl))) |
4c0f7679 JH |
358 | { |
359 | e->inline_failed = CIF_EH_PERSONALITY; | |
360 | inlinable = false; | |
361 | } | |
a7ff6e27 AH |
362 | /* TM pure functions should not be inlined into non-TM_pure |
363 | functions. */ | |
7ce7e4d4 | 364 | else if (is_tm_pure (callee->decl) && !is_tm_pure (caller->decl)) |
0a35513e AH |
365 | { |
366 | e->inline_failed = CIF_UNSPECIFIED; | |
367 | inlinable = false; | |
368 | } | |
09a2806f | 369 | /* Check compatibility of target optimization options. */ |
bb1e543c | 370 | else if (!targetm.target_option.can_inline_p (caller->decl, |
67348ccc | 371 | callee->decl)) |
4c0f7679 JH |
372 | { |
373 | e->inline_failed = CIF_TARGET_OPTION_MISMATCH; | |
374 | inlinable = false; | |
375 | } | |
56f62793 ML |
376 | else if (ipa_fn_summaries->get (callee) == NULL |
377 | || !ipa_fn_summaries->get (callee)->inlinable) | |
5058c037 JH |
378 | { |
379 | e->inline_failed = CIF_FUNCTION_NOT_INLINABLE; | |
380 | inlinable = false; | |
381 | } | |
25a07c7e | 382 | /* Don't inline a function with mismatched sanitization attributes. */ |
bb1e543c | 383 | else if (!sanitize_attrs_match_for_inline_p (caller->decl, callee->decl)) |
25a07c7e YG |
384 | { |
385 | e->inline_failed = CIF_ATTRIBUTE_MISMATCH; | |
386 | inlinable = false; | |
387 | } | |
9a4841a3 JH |
388 | if (!inlinable && report) |
389 | report_inline_failed_reason (e); | |
390 | return inlinable; | |
391 | } | |
392 | ||
2925cad2 JH |
393 | /* Return inlining_insns_single limit for function N. If HINT is true |
394 | scale up the bound. */ | |
562d1e95 JH |
395 | |
396 | static int | |
2925cad2 | 397 | inline_insns_single (cgraph_node *n, bool hint) |
562d1e95 | 398 | { |
1e83bd70 JH |
399 | if (hint) |
400 | return opt_for_fn (n->decl, param_max_inline_insns_single) | |
401 | * opt_for_fn (n->decl, param_inline_heuristics_hint_percent) / 100; | |
402 | return opt_for_fn (n->decl, param_max_inline_insns_single); | |
562d1e95 JH |
403 | } |
404 | ||
2925cad2 JH |
405 | /* Return inlining_insns_auto limit for function N. If HINT is true |
406 | scale up the bound. */ | |
562d1e95 JH |
407 | |
408 | static int | |
2925cad2 | 409 | inline_insns_auto (cgraph_node *n, bool hint) |
562d1e95 | 410 | { |
78a502ca ML |
411 | int max_inline_insns_auto = opt_for_fn (n->decl, param_max_inline_insns_auto); |
412 | if (hint) | |
1e83bd70 JH |
413 | return max_inline_insns_auto |
414 | * opt_for_fn (n->decl, param_inline_heuristics_hint_percent) / 100; | |
78a502ca | 415 | return max_inline_insns_auto; |
562d1e95 JH |
416 | } |
417 | ||
9a4841a3 JH |
418 | /* Decide if we can inline the edge and possibly update |
419 | inline_failed reason. | |
420 | We check whether inlining is possible at all and whether | |
421 | caller growth limits allow doing so. | |
422 | ||
423 | if REPORT is true, output reason to the dump file. | |
424 | ||
425 | if DISREGARD_LIMITS is true, ignore size limits. */ | |
426 | ||
427 | static bool | |
428 | can_inline_edge_by_limits_p (struct cgraph_edge *e, bool report, | |
429 | bool disregard_limits = false, bool early = false) | |
430 | { | |
431 | gcc_checking_assert (e->inline_failed); | |
432 | ||
433 | if (cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR) | |
434 | { | |
435 | if (report) | |
436 | report_inline_failed_reason (e); | |
437 | return false; | |
438 | } | |
439 | ||
440 | bool inlinable = true; | |
441 | enum availability avail; | |
a62bfab5 ML |
442 | cgraph_node *caller = (e->caller->inlined_to |
443 | ? e->caller->inlined_to : e->caller); | |
9a4841a3 JH |
444 | cgraph_node *callee = e->callee->ultimate_alias_target (&avail, caller); |
445 | tree caller_tree = DECL_FUNCTION_SPECIFIC_OPTIMIZATION (caller->decl); | |
446 | tree callee_tree | |
447 | = callee ? DECL_FUNCTION_SPECIFIC_OPTIMIZATION (callee->decl) : NULL; | |
4c0f7679 | 448 | /* Check if caller growth allows the inlining. */ |
9a4841a3 JH |
449 | if (!DECL_DISREGARD_INLINE_LIMITS (callee->decl) |
450 | && !disregard_limits | |
451 | && !lookup_attribute ("flatten", | |
452 | DECL_ATTRIBUTES (caller->decl)) | |
453 | && !caller_growth_limits (e)) | |
4c0f7679 | 454 | inlinable = false; |
bc53dee0 QZ |
455 | else if (callee->externally_visible |
456 | && !DECL_DISREGARD_INLINE_LIMITS (callee->decl) | |
457 | && flag_live_patching == LIVE_PATCHING_INLINE_ONLY_STATIC) | |
458 | { | |
459 | e->inline_failed = CIF_EXTERN_LIVE_ONLY_STATIC; | |
460 | inlinable = false; | |
461 | } | |
4c0f7679 JH |
462 | /* Don't inline a function with a higher optimization level than the |
463 | caller. FIXME: this is really just tip of iceberg of handling | |
464 | optimization attribute. */ | |
465 | else if (caller_tree != callee_tree) | |
9de21a23 | 466 | { |
8e926cb1 JH |
467 | bool always_inline = |
468 | (DECL_DISREGARD_INLINE_LIMITS (callee->decl) | |
469 | && lookup_attribute ("always_inline", | |
470 | DECL_ATTRIBUTES (callee->decl))); | |
56f62793 ML |
471 | ipa_fn_summary *caller_info = ipa_fn_summaries->get (caller); |
472 | ipa_fn_summary *callee_info = ipa_fn_summaries->get (callee); | |
8e926cb1 | 473 | |
553bb257 KT |
474 | /* Until GCC 4.9 we did not check the semantics-altering flags |
475 | below and inlined across optimization boundaries. | |
476 | Enabling checks below breaks several packages by refusing | |
ddb3773a JH |
477 | to inline library always_inline functions. See PR65873. |
478 | Disable the check for early inlining for now until better solution | |
479 | is found. */ | |
480 | if (always_inline && early) | |
481 | ; | |
0c3068e0 RB |
482 | /* There are some options that change IL semantics which means |
483 | we cannot inline in these cases for correctness reason. | |
484 | Not even for always_inline declared functions. */ | |
2bf54d93 | 485 | else if (check_match (flag_wrapv) |
ddb3773a | 486 | || check_match (flag_trapv) |
29f1e2b1 | 487 | || check_match (flag_pcc_struct_return) |
818b88a7 JH |
488 | /* When caller or callee does FP math, be sure FP codegen flags |
489 | compatible. */ | |
490 | || ((caller_info->fp_expressions && callee_info->fp_expressions) | |
491 | && (check_maybe_up (flag_rounding_math) | |
492 | || check_maybe_up (flag_trapping_math) | |
493 | || check_maybe_down (flag_unsafe_math_optimizations) | |
494 | || check_maybe_down (flag_finite_math_only) | |
495 | || check_maybe_up (flag_signaling_nans) | |
496 | || check_maybe_down (flag_cx_limited_range) | |
497 | || check_maybe_up (flag_signed_zeros) | |
498 | || check_maybe_down (flag_associative_math) | |
499 | || check_maybe_down (flag_reciprocal_math) | |
0d2f700f | 500 | || check_maybe_down (flag_fp_int_builtin_inexact) |
818b88a7 JH |
501 | /* Strictly speaking only when the callee contains function |
502 | calls that may end up setting errno. */ | |
503 | || check_maybe_up (flag_errno_math))) | |
ddb3773a JH |
504 | /* We do not want to make code compiled with exceptions to be |
505 | brought into a non-EH function unless we know that the callee | |
506 | does not throw. | |
507 | This is tracked by DECL_FUNCTION_PERSONALITY. */ | |
6a96e917 EB |
508 | || (check_maybe_up (flag_non_call_exceptions) |
509 | && DECL_FUNCTION_PERSONALITY (callee->decl)) | |
ddb3773a JH |
510 | || (check_maybe_up (flag_exceptions) |
511 | && DECL_FUNCTION_PERSONALITY (callee->decl)) | |
956d615d | 512 | /* When devirtualization is disabled for callee, it is not safe |
ddb3773a JH |
513 | to inline it as we possibly mangled the type info. |
514 | Allow early inlining of always inlines. */ | |
515 | || (!early && check_maybe_down (flag_devirtualize))) | |
0c3068e0 RB |
516 | { |
517 | e->inline_failed = CIF_OPTIMIZATION_MISMATCH; | |
518 | inlinable = false; | |
519 | } | |
520 | /* gcc.dg/pr43564.c. Apply user-forced inline even at -O0. */ | |
8e926cb1 | 521 | else if (always_inline) |
bb1e543c | 522 | ; |
0c3068e0 RB |
523 | /* When user added an attribute to the callee honor it. */ |
524 | else if (lookup_attribute ("optimize", DECL_ATTRIBUTES (callee->decl)) | |
525 | && opts_for_fn (caller->decl) != opts_for_fn (callee->decl)) | |
4c0f7679 | 526 | { |
fd811f03 | 527 | e->inline_failed = CIF_OPTIMIZATION_MISMATCH; |
4c0f7679 JH |
528 | inlinable = false; |
529 | } | |
ddb3773a JH |
530 | /* If explicit optimize attribute are not used, the mismatch is caused |
531 | by different command line options used to build different units. | |
532 | Do not care about COMDAT functions - those are intended to be | |
533 | optimized with the optimization flags of module they are used in. | |
534 | Also do not care about mixing up size/speed optimization when | |
535 | DECL_DISREGARD_INLINE_LIMITS is set. */ | |
88636b62 | 536 | else if ((callee->merged_comdat |
ddb3773a JH |
537 | && !lookup_attribute ("optimize", |
538 | DECL_ATTRIBUTES (caller->decl))) | |
539 | || DECL_DISREGARD_INLINE_LIMITS (callee->decl)) | |
540 | ; | |
bb1e543c | 541 | /* If mismatch is caused by merging two LTO units with different |
956d615d | 542 | optimization flags we want to be bit nicer. However never inline |
bb1e543c JH |
543 | if one of functions is not optimized at all. */ |
544 | else if (!opt_for_fn (callee->decl, optimize) | |
545 | || !opt_for_fn (caller->decl, optimize)) | |
546 | { | |
547 | e->inline_failed = CIF_OPTIMIZATION_MISMATCH; | |
548 | inlinable = false; | |
549 | } | |
550 | /* If callee is optimized for size and caller is not, allow inlining if | |
028d4092 ML |
551 | code shrinks or we are in param_max_inline_insns_single limit and |
552 | callee is inline (and thus likely an unified comdat). | |
553 | This will allow caller to run faster. */ | |
bb1e543c JH |
554 | else if (opt_for_fn (callee->decl, optimize_size) |
555 | > opt_for_fn (caller->decl, optimize_size)) | |
556 | { | |
557 | int growth = estimate_edge_growth (e); | |
1e83bd70 | 558 | if (growth > opt_for_fn (caller->decl, param_max_inline_insns_size) |
bb1e543c | 559 | && (!DECL_DECLARED_INLINE_P (callee->decl) |
2925cad2 JH |
560 | && growth >= MAX (inline_insns_single (caller, false), |
561 | inline_insns_auto (caller, false)))) | |
bb1e543c JH |
562 | { |
563 | e->inline_failed = CIF_OPTIMIZATION_MISMATCH; | |
564 | inlinable = false; | |
565 | } | |
566 | } | |
567 | /* If callee is more aggressively optimized for performance than caller, | |
568 | we generally want to inline only cheap (runtime wise) functions. */ | |
569 | else if (opt_for_fn (callee->decl, optimize_size) | |
570 | < opt_for_fn (caller->decl, optimize_size) | |
571 | || (opt_for_fn (callee->decl, optimize) | |
86f46e39 | 572 | > opt_for_fn (caller->decl, optimize))) |
bb1e543c JH |
573 | { |
574 | if (estimate_edge_time (e) | |
56f62793 | 575 | >= 20 + ipa_call_summaries->get (e)->call_stmt_time) |
bb1e543c JH |
576 | { |
577 | e->inline_failed = CIF_OPTIMIZATION_MISMATCH; | |
578 | inlinable = false; | |
579 | } | |
580 | } | |
581 | ||
4c0f7679 JH |
582 | } |
583 | ||
4c0f7679 JH |
584 | if (!inlinable && report) |
585 | report_inline_failed_reason (e); | |
586 | return inlinable; | |
587 | } | |
588 | ||
589 | ||
590 | /* Return true if the edge E is inlinable during early inlining. */ | |
591 | ||
592 | static bool | |
593 | can_early_inline_edge_p (struct cgraph_edge *e) | |
594 | { | |
d52f5295 | 595 | struct cgraph_node *callee = e->callee->ultimate_alias_target (); |
4c0f7679 | 596 | /* Early inliner might get called at WPA stage when IPA pass adds new |
67914693 | 597 | function. In this case we cannot really do any of early inlining |
4c0f7679 | 598 | because function bodies are missing. */ |
1a0bf5e1 JH |
599 | if (cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR) |
600 | return false; | |
67348ccc | 601 | if (!gimple_has_body_p (callee->decl)) |
4c0f7679 JH |
602 | { |
603 | e->inline_failed = CIF_BODY_NOT_AVAILABLE; | |
9de21a23 JC |
604 | return false; |
605 | } | |
4c0f7679 JH |
606 | /* In early inliner some of callees may not be in SSA form yet |
607 | (i.e. the callgraph is cyclic and we did not process | |
608 | the callee by early inliner, yet). We don't have CIF code for this | |
609 | case; later we will re-do the decision in the real inliner. */ | |
67348ccc DM |
610 | if (!gimple_in_ssa_p (DECL_STRUCT_FUNCTION (e->caller->decl)) |
611 | || !gimple_in_ssa_p (DECL_STRUCT_FUNCTION (callee->decl))) | |
f27e50db | 612 | { |
4174a33a DM |
613 | if (dump_enabled_p ()) |
614 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt, | |
615 | " edge not inlinable: not in SSA form\n"); | |
f27e50db JH |
616 | return false; |
617 | } | |
9a4841a3 JH |
618 | if (!can_inline_edge_p (e, true, true) |
619 | || !can_inline_edge_by_limits_p (e, true, false, true)) | |
4c0f7679 JH |
620 | return false; |
621 | return true; | |
622 | } | |
623 | ||
624 | ||
ae6e6a08 | 625 | /* Return number of calls in N. Ignore cheap builtins. */ |
4c0f7679 | 626 | |
ae6e6a08 JH |
627 | static int |
628 | num_calls (struct cgraph_node *n) | |
4c0f7679 JH |
629 | { |
630 | struct cgraph_edge *e; | |
ae6e6a08 JH |
631 | int num = 0; |
632 | ||
4c0f7679 | 633 | for (e = n->callees; e; e = e->next_callee) |
67348ccc | 634 | if (!is_inexpensive_builtin (e->callee->decl)) |
ae6e6a08 JH |
635 | num++; |
636 | return num; | |
4c0f7679 JH |
637 | } |
638 | ||
f27e50db | 639 | |
4c0f7679 | 640 | /* Return true if we are interested in inlining small function. */ |
9de21a23 | 641 | |
4c0f7679 JH |
642 | static bool |
643 | want_early_inline_function_p (struct cgraph_edge *e) | |
644 | { | |
645 | bool want_inline = true; | |
d52f5295 | 646 | struct cgraph_node *callee = e->callee->ultimate_alias_target (); |
4c0f7679 | 647 | |
67348ccc | 648 | if (DECL_DISREGARD_INLINE_LIMITS (callee->decl)) |
4c0f7679 | 649 | ; |
9a1e784a | 650 | /* For AutoFDO, we need to make sure that before profile summary, all |
be3c16c4 DC |
651 | hot paths' IR look exactly the same as profiled binary. As a result, |
652 | in einliner, we will disregard size limit and inline those callsites | |
653 | that are: | |
654 | * inlined in the profiled binary, and | |
655 | * the cloned callee has enough samples to be considered "hot". */ | |
656 | else if (flag_auto_profile && afdo_callsite_hot_enough_for_early_inline (e)) | |
657 | ; | |
67348ccc | 658 | else if (!DECL_DECLARED_INLINE_P (callee->decl) |
2bf86c84 | 659 | && !opt_for_fn (e->caller->decl, flag_inline_small_functions)) |
4c0f7679 JH |
660 | { |
661 | e->inline_failed = CIF_FUNCTION_NOT_INLINE_CANDIDATE; | |
662 | report_inline_failed_reason (e); | |
663 | want_inline = false; | |
664 | } | |
665 | else | |
9de21a23 | 666 | { |
49e26500 JH |
667 | /* First take care of very large functions. */ |
668 | int min_growth = estimate_min_edge_growth (e), growth = 0; | |
ae6e6a08 | 669 | int n; |
1e83bd70 | 670 | int early_inlining_insns = param_early_inlining_insns; |
0b92cf30 | 671 | |
49e26500 JH |
672 | if (min_growth > early_inlining_insns) |
673 | { | |
674 | if (dump_enabled_p ()) | |
675 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt, | |
676 | " will not early inline: %C->%C, " | |
677 | "call is cold and code would grow " | |
678 | "at least by %i\n", | |
679 | e->caller, callee, | |
680 | min_growth); | |
681 | want_inline = false; | |
682 | } | |
683 | else | |
684 | growth = estimate_edge_growth (e); | |
685 | ||
ae6e6a08 | 686 | |
49e26500 | 687 | if (!want_inline || growth <= param_max_inline_insns_size) |
4c0f7679 | 688 | ; |
f256c274 | 689 | else if (!e->maybe_hot_p ()) |
4c0f7679 | 690 | { |
4174a33a DM |
691 | if (dump_enabled_p ()) |
692 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt, | |
693 | " will not early inline: %C->%C, " | |
694 | "call is cold and code would grow by %i\n", | |
695 | e->caller, callee, | |
696 | growth); | |
4c0f7679 JH |
697 | want_inline = false; |
698 | } | |
0b92cf30 | 699 | else if (growth > early_inlining_insns) |
9de21a23 | 700 | { |
4174a33a DM |
701 | if (dump_enabled_p ()) |
702 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt, | |
703 | " will not early inline: %C->%C, " | |
9c28689a JH |
704 | "growth %i exceeds --param early-inlining-insns\n", |
705 | e->caller, callee, growth); | |
4c0f7679 | 706 | want_inline = false; |
9de21a23 | 707 | } |
ae6e6a08 | 708 | else if ((n = num_calls (callee)) != 0 |
0b92cf30 | 709 | && growth * (n + 1) > early_inlining_insns) |
4c0f7679 | 710 | { |
4174a33a DM |
711 | if (dump_enabled_p ()) |
712 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt, | |
713 | " will not early inline: %C->%C, " | |
9c28689a | 714 | "growth %i exceeds --param early-inlining-insns " |
4174a33a | 715 | "divided by number of calls\n", |
9c28689a | 716 | e->caller, callee, growth); |
4c0f7679 JH |
717 | want_inline = false; |
718 | } | |
719 | } | |
720 | return want_inline; | |
721 | } | |
722 | ||
d59171da JH |
723 | /* Compute time of the edge->caller + edge->callee execution when inlining |
724 | does not happen. */ | |
725 | ||
6d4ab5f8 | 726 | inline sreal |
4adaad64 | 727 | compute_uninlined_call_time (struct cgraph_edge *edge, |
b5351388 JH |
728 | sreal uninlined_call_time, |
729 | sreal freq) | |
d59171da | 730 | { |
a62bfab5 ML |
731 | cgraph_node *caller = (edge->caller->inlined_to |
732 | ? edge->caller->inlined_to | |
208e5afa JH |
733 | : edge->caller); |
734 | ||
0009a6c3 | 735 | if (freq > 0) |
41f0e819 | 736 | uninlined_call_time *= freq; |
208e5afa JH |
737 | else |
738 | uninlined_call_time = uninlined_call_time >> 11; | |
739 | ||
56f62793 | 740 | sreal caller_time = ipa_fn_summaries->get (caller)->time; |
d59171da JH |
741 | return uninlined_call_time + caller_time; |
742 | } | |
743 | ||
744 | /* Same as compute_uinlined_call_time but compute time when inlining | |
745 | does happen. */ | |
746 | ||
6d4ab5f8 | 747 | inline sreal |
d59171da | 748 | compute_inlined_call_time (struct cgraph_edge *edge, |
b5351388 JH |
749 | sreal time, |
750 | sreal freq) | |
d59171da | 751 | { |
a62bfab5 ML |
752 | cgraph_node *caller = (edge->caller->inlined_to |
753 | ? edge->caller->inlined_to | |
208e5afa | 754 | : edge->caller); |
56f62793 | 755 | sreal caller_time = ipa_fn_summaries->get (caller)->time; |
208e5afa | 756 | |
0009a6c3 | 757 | if (freq > 0) |
41f0e819 | 758 | time *= freq; |
208e5afa JH |
759 | else |
760 | time = time >> 11; | |
761 | ||
4adaad64 JH |
762 | /* This calculation should match one in ipa-inline-analysis.c |
763 | (estimate_edge_size_and_time). */ | |
56f62793 | 764 | time -= (sreal)ipa_call_summaries->get (edge)->call_stmt_time * freq; |
208e5afa JH |
765 | time += caller_time; |
766 | if (time <= 0) | |
767 | time = ((sreal) 1) >> 8; | |
6d4ab5f8 | 768 | gcc_checking_assert (time >= 0); |
d59171da JH |
769 | return time; |
770 | } | |
771 | ||
956d615d JJ |
772 | /* Determine time saved by inlining EDGE of frequency FREQ |
773 | where callee's runtime w/o inlining is UNINLINED_TYPE | |
ea8dd3b6 JH |
774 | and with inlined is INLINED_TYPE. */ |
775 | ||
776 | inline sreal | |
777 | inlining_speedup (struct cgraph_edge *edge, | |
778 | sreal freq, | |
779 | sreal uninlined_time, | |
780 | sreal inlined_time) | |
781 | { | |
782 | sreal speedup = uninlined_time - inlined_time; | |
783 | /* Handling of call_time should match one in ipa-inline-fnsummary.c | |
784 | (estimate_edge_size_and_time). */ | |
785 | sreal call_time = ipa_call_summaries->get (edge)->call_stmt_time; | |
786 | ||
787 | if (freq > 0) | |
788 | { | |
789 | speedup = (speedup + call_time); | |
790 | if (freq != 1) | |
791 | speedup = speedup * freq; | |
792 | } | |
793 | else if (freq == 0) | |
794 | speedup = speedup >> 11; | |
795 | gcc_checking_assert (speedup >= 0); | |
796 | return speedup; | |
797 | } | |
798 | ||
42f7b0fa JH |
799 | /* Return true if the speedup for inlining E is bigger than |
800 | PARAM_MAX_INLINE_MIN_SPEEDUP. */ | |
801 | ||
802 | static bool | |
803 | big_speedup_p (struct cgraph_edge *e) | |
804 | { | |
4adaad64 JH |
805 | sreal unspec_time; |
806 | sreal spec_time = estimate_edge_time (e, &unspec_time); | |
b5351388 JH |
807 | sreal freq = e->sreal_frequency (); |
808 | sreal time = compute_uninlined_call_time (e, unspec_time, freq); | |
809 | sreal inlined_time = compute_inlined_call_time (e, spec_time, freq); | |
a62bfab5 ML |
810 | cgraph_node *caller = (e->caller->inlined_to |
811 | ? e->caller->inlined_to | |
812 | : e->caller); | |
1e83bd70 | 813 | int limit = opt_for_fn (caller->decl, param_inline_min_speedup); |
208e5afa | 814 | |
562d1e95 | 815 | if ((time - inlined_time) * 100 > time * limit) |
42f7b0fa JH |
816 | return true; |
817 | return false; | |
818 | } | |
819 | ||
4c0f7679 JH |
820 | /* Return true if we are interested in inlining small function. |
821 | When REPORT is true, report reason to dump file. */ | |
822 | ||
823 | static bool | |
824 | want_inline_small_function_p (struct cgraph_edge *e, bool report) | |
825 | { | |
826 | bool want_inline = true; | |
d52f5295 | 827 | struct cgraph_node *callee = e->callee->ultimate_alias_target (); |
1e83bd70 JH |
828 | cgraph_node *to = (e->caller->inlined_to |
829 | ? e->caller->inlined_to : e->caller); | |
4c0f7679 | 830 | |
01b9bf06 RS |
831 | /* Allow this function to be called before can_inline_edge_p, |
832 | since it's usually cheaper. */ | |
833 | if (cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR) | |
834 | want_inline = false; | |
835 | else if (DECL_DISREGARD_INLINE_LIMITS (callee->decl)) | |
4c0f7679 | 836 | ; |
67348ccc | 837 | else if (!DECL_DECLARED_INLINE_P (callee->decl) |
2bf86c84 | 838 | && !opt_for_fn (e->caller->decl, flag_inline_small_functions)) |
4c0f7679 JH |
839 | { |
840 | e->inline_failed = CIF_FUNCTION_NOT_INLINE_CANDIDATE; | |
841 | want_inline = false; | |
9de21a23 | 842 | } |
4cd8957f | 843 | /* Do fast and conservative check if the function can be good |
2925cad2 | 844 | inline candidate. */ |
b6d627e4 | 845 | else if ((!DECL_DECLARED_INLINE_P (callee->decl) |
1bad9c18 | 846 | && (!e->count.ipa ().initialized_p () || !e->maybe_hot_p ())) |
56f62793 ML |
847 | && ipa_fn_summaries->get (callee)->min_size |
848 | - ipa_call_summaries->get (e)->call_stmt_size | |
2925cad2 | 849 | > inline_insns_auto (e->caller, true)) |
4cd8957f | 850 | { |
afeb8875 | 851 | e->inline_failed = CIF_MAX_INLINE_INSNS_AUTO_LIMIT; |
4cd8957f JH |
852 | want_inline = false; |
853 | } | |
3995f3a2 | 854 | else if ((DECL_DECLARED_INLINE_P (callee->decl) |
1bad9c18 | 855 | || e->count.ipa ().nonzero_p ()) |
56f62793 ML |
856 | && ipa_fn_summaries->get (callee)->min_size |
857 | - ipa_call_summaries->get (e)->call_stmt_size | |
2925cad2 | 858 | > inline_insns_single (e->caller, true)) |
4cd8957f | 859 | { |
9c28689a JH |
860 | e->inline_failed = (DECL_DECLARED_INLINE_P (callee->decl) |
861 | ? CIF_MAX_INLINE_INSNS_SINGLE_LIMIT | |
862 | : CIF_MAX_INLINE_INSNS_AUTO_LIMIT); | |
4cd8957f JH |
863 | want_inline = false; |
864 | } | |
ca31b95f | 865 | else |
9de21a23 | 866 | { |
4c0f7679 | 867 | int growth = estimate_edge_growth (e); |
0bceb671 | 868 | ipa_hints hints = estimate_edge_hints (e); |
2925cad2 JH |
869 | bool apply_hints = (hints & (INLINE_HINT_indirect_call |
870 | | INLINE_HINT_known_hot | |
871 | | INLINE_HINT_loop_iterations | |
872 | | INLINE_HINT_loop_stride)); | |
4c0f7679 | 873 | |
1e83bd70 JH |
874 | if (growth <= opt_for_fn (to->decl, |
875 | param_max_inline_insns_size)) | |
4c0f7679 | 876 | ; |
028d4092 | 877 | /* Apply param_max_inline_insns_single limit. Do not do so when |
2925cad2 JH |
878 | hints suggests that inlining given function is very profitable. |
879 | Avoid computation of big_speedup_p when not necessary to change | |
880 | outcome of decision. */ | |
67348ccc | 881 | else if (DECL_DECLARED_INLINE_P (callee->decl) |
2925cad2 JH |
882 | && growth >= inline_insns_single (e->caller, apply_hints) |
883 | && (apply_hints | |
884 | || growth >= inline_insns_single (e->caller, true) | |
885 | || !big_speedup_p (e))) | |
4c0f7679 | 886 | { |
1e83bd70 | 887 | e->inline_failed = CIF_MAX_INLINE_INSNS_SINGLE_LIMIT; |
4c0f7679 JH |
888 | want_inline = false; |
889 | } | |
67348ccc | 890 | else if (!DECL_DECLARED_INLINE_P (callee->decl) |
f256c274 | 891 | && !opt_for_fn (e->caller->decl, flag_inline_functions) |
1e83bd70 JH |
892 | && growth >= opt_for_fn (to->decl, |
893 | param_max_inline_insns_small)) | |
4c0f7679 | 894 | { |
49d9c9d2 | 895 | /* growth_positive_p is expensive, always test it last. */ |
2925cad2 | 896 | if (growth >= inline_insns_single (e->caller, false) |
49d9c9d2 | 897 | || growth_positive_p (callee, e, growth)) |
4cd8957f | 898 | { |
562d1e95 | 899 | e->inline_failed = CIF_NOT_DECLARED_INLINED; |
4cd8957f | 900 | want_inline = false; |
562d1e95 | 901 | } |
4c0f7679 | 902 | } |
028d4092 ML |
903 | /* Apply param_max_inline_insns_auto limit for functions not declared |
904 | inline. Bypass the limit when speedup seems big. */ | |
67348ccc | 905 | else if (!DECL_DECLARED_INLINE_P (callee->decl) |
2925cad2 JH |
906 | && growth >= inline_insns_auto (e->caller, apply_hints) |
907 | && (apply_hints | |
908 | || growth >= inline_insns_auto (e->caller, true) | |
909 | || !big_speedup_p (e))) | |
4c0f7679 | 910 | { |
49d9c9d2 | 911 | /* growth_positive_p is expensive, always test it last. */ |
2925cad2 | 912 | if (growth >= inline_insns_single (e->caller, false) |
49d9c9d2 | 913 | || growth_positive_p (callee, e, growth)) |
4cd8957f | 914 | { |
afeb8875 | 915 | e->inline_failed = CIF_MAX_INLINE_INSNS_AUTO_LIMIT; |
4cd8957f | 916 | want_inline = false; |
562d1e95 | 917 | } |
4c0f7679 | 918 | } |
db22a743 | 919 | /* If call is cold, do not inline when function body would grow. */ |
3dafb85c | 920 | else if (!e->maybe_hot_p () |
2925cad2 | 921 | && (growth >= inline_insns_single (e->caller, false) |
49d9c9d2 | 922 | || growth_positive_p (callee, e, growth))) |
9de21a23 | 923 | { |
562d1e95 | 924 | e->inline_failed = CIF_UNLIKELY_CALL; |
4c0f7679 | 925 | want_inline = false; |
9de21a23 JC |
926 | } |
927 | } | |
4c0f7679 JH |
928 | if (!want_inline && report) |
929 | report_inline_failed_reason (e); | |
930 | return want_inline; | |
931 | } | |
9de21a23 | 932 | |
4c0f7679 | 933 | /* EDGE is self recursive edge. |
1e83bd70 | 934 | We handle two cases - when function A is inlining into itself |
4c0f7679 JH |
935 | or when function A is being inlined into another inliner copy of function |
936 | A within function B. | |
937 | ||
938 | In first case OUTER_NODE points to the toplevel copy of A, while | |
939 | in the second case OUTER_NODE points to the outermost copy of A in B. | |
940 | ||
941 | In both cases we want to be extra selective since | |
942 | inlining the call will just introduce new recursive calls to appear. */ | |
09a2806f | 943 | |
4c0f7679 JH |
944 | static bool |
945 | want_inline_self_recursive_call_p (struct cgraph_edge *edge, | |
946 | struct cgraph_node *outer_node, | |
947 | bool peeling, | |
948 | int depth) | |
949 | { | |
950 | char const *reason = NULL; | |
951 | bool want_inline = true; | |
0009a6c3 | 952 | sreal caller_freq = 1; |
1e83bd70 JH |
953 | int max_depth = opt_for_fn (outer_node->decl, |
954 | param_max_inline_recursive_depth_auto); | |
4c0f7679 | 955 | |
67348ccc | 956 | if (DECL_DECLARED_INLINE_P (edge->caller->decl)) |
1e83bd70 JH |
957 | max_depth = opt_for_fn (outer_node->decl, |
958 | param_max_inline_recursive_depth); | |
4c0f7679 | 959 | |
3dafb85c | 960 | if (!edge->maybe_hot_p ()) |
4c0f7679 JH |
961 | { |
962 | reason = "recursive call is cold"; | |
963 | want_inline = false; | |
964 | } | |
4c0f7679 JH |
965 | else if (depth > max_depth) |
966 | { | |
967 | reason = "--param max-inline-recursive-depth exceeded."; | |
968 | want_inline = false; | |
969 | } | |
a62bfab5 | 970 | else if (outer_node->inlined_to |
0009a6c3 | 971 | && (caller_freq = outer_node->callers->sreal_frequency ()) == 0) |
bd936951 | 972 | { |
0009a6c3 | 973 | reason = "caller frequency is 0"; |
bd936951 JH |
974 | want_inline = false; |
975 | } | |
976 | ||
4c0f7679 JH |
977 | if (!want_inline) |
978 | ; | |
0009a6c3 JH |
979 | /* Inlining of self recursive function into copy of itself within other |
980 | function is transformation similar to loop peeling. | |
4c0f7679 | 981 | |
09a2806f | 982 | Peeling is profitable if we can inline enough copies to make probability |
4c0f7679 JH |
983 | of actual call to the self recursive function very small. Be sure that |
984 | the probability of recursion is small. | |
985 | ||
09a2806f | 986 | We ensure that the frequency of recursing is at most 1 - (1/max_depth). |
0009a6c3 | 987 | This way the expected number of recursion is at most max_depth. */ |
4c0f7679 JH |
988 | else if (peeling) |
989 | { | |
0009a6c3 | 990 | sreal max_prob = (sreal)1 - ((sreal)1 / (sreal)max_depth); |
4c0f7679 JH |
991 | int i; |
992 | for (i = 1; i < depth; i++) | |
0009a6c3 JH |
993 | max_prob = max_prob * max_prob; |
994 | if (edge->sreal_frequency () >= max_prob * caller_freq) | |
4c0f7679 JH |
995 | { |
996 | reason = "frequency of recursive call is too large"; | |
997 | want_inline = false; | |
998 | } | |
999 | } | |
0009a6c3 JH |
1000 | /* Recursive inlining, i.e. equivalent of unrolling, is profitable if |
1001 | recursion depth is large. We reduce function call overhead and increase | |
1002 | chances that things fit in hardware return predictor. | |
4c0f7679 JH |
1003 | |
1004 | Recursive inlining might however increase cost of stack frame setup | |
1005 | actually slowing down functions whose recursion tree is wide rather than | |
1006 | deep. | |
1007 | ||
09a2806f | 1008 | Deciding reliably on when to do recursive inlining without profile feedback |
4c0f7679 JH |
1009 | is tricky. For now we disable recursive inlining when probability of self |
1010 | recursion is low. | |
1011 | ||
0009a6c3 JH |
1012 | Recursive inlining of self recursive call within loop also results in |
1013 | large loop depths that generally optimize badly. We may want to throttle | |
1014 | down inlining in those cases. In particular this seems to happen in one | |
1015 | of libstdc++ rb tree methods. */ | |
4c0f7679 JH |
1016 | else |
1017 | { | |
0009a6c3 JH |
1018 | if (edge->sreal_frequency () * 100 |
1019 | <= caller_freq | |
1e83bd70 JH |
1020 | * opt_for_fn (outer_node->decl, |
1021 | param_min_inline_recursive_probability)) | |
4c0f7679 JH |
1022 | { |
1023 | reason = "frequency of recursive call is too small"; | |
1024 | want_inline = false; | |
1025 | } | |
1026 | } | |
4174a33a DM |
1027 | if (!want_inline && dump_enabled_p ()) |
1028 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, edge->call_stmt, | |
1029 | " not inlining recursively: %s\n", reason); | |
4c0f7679 | 1030 | return want_inline; |
ca31b95f JH |
1031 | } |
1032 | ||
19ba6aab JH |
1033 | /* Return true when NODE has uninlinable caller; |
1034 | set HAS_HOT_CALL if it has hot call. | |
9aa3f5c5 JH |
1035 | Worker for cgraph_for_node_and_aliases. */ |
1036 | ||
1037 | static bool | |
19ba6aab | 1038 | check_callers (struct cgraph_node *node, void *has_hot_call) |
9aa3f5c5 | 1039 | { |
19ba6aab JH |
1040 | struct cgraph_edge *e; |
1041 | for (e = node->callers; e; e = e->next_caller) | |
1042 | { | |
29f1e2b1 JH |
1043 | if (!opt_for_fn (e->caller->decl, flag_inline_functions_called_once) |
1044 | || !opt_for_fn (e->caller->decl, optimize)) | |
2bf86c84 | 1045 | return true; |
19ba6aab JH |
1046 | if (!can_inline_edge_p (e, true)) |
1047 | return true; | |
1af8bfe5 JH |
1048 | if (e->recursive_p ()) |
1049 | return true; | |
9a4841a3 JH |
1050 | if (!can_inline_edge_by_limits_p (e, true)) |
1051 | return true; | |
3dafb85c | 1052 | if (!(*(bool *)has_hot_call) && e->maybe_hot_p ()) |
19ba6aab JH |
1053 | *(bool *)has_hot_call = true; |
1054 | } | |
1055 | return false; | |
9aa3f5c5 JH |
1056 | } |
1057 | ||
a81b0a3d JH |
1058 | /* If NODE has a caller, return true. */ |
1059 | ||
1060 | static bool | |
1061 | has_caller_p (struct cgraph_node *node, void *data ATTRIBUTE_UNUSED) | |
1062 | { | |
1063 | if (node->callers) | |
1064 | return true; | |
1065 | return false; | |
1066 | } | |
09a2806f | 1067 | |
100411f8 JH |
1068 | /* Decide if inlining NODE would reduce unit size by eliminating |
1069 | the offline copy of function. | |
1070 | When COLD is true the cold calls are considered, too. */ | |
09a2806f JH |
1071 | |
1072 | static bool | |
100411f8 | 1073 | want_inline_function_to_all_callers_p (struct cgraph_node *node, bool cold) |
09a2806f | 1074 | { |
5970b079 EB |
1075 | bool has_hot_call = false; |
1076 | ||
9789b553 JH |
1077 | /* Aliases gets inlined along with the function they alias. */ |
1078 | if (node->alias) | |
5970b079 EB |
1079 | return false; |
1080 | /* Already inlined? */ | |
a62bfab5 | 1081 | if (node->inlined_to) |
5970b079 EB |
1082 | return false; |
1083 | /* Does it have callers? */ | |
1ede94c5 | 1084 | if (!node->call_for_symbol_and_aliases (has_caller_p, NULL, true)) |
5970b079 EB |
1085 | return false; |
1086 | /* Inlining into all callers would increase size? */ | |
49d9c9d2 | 1087 | if (growth_positive_p (node, NULL, INT_MIN) > 0) |
5970b079 EB |
1088 | return false; |
1089 | /* All inlines must be possible. */ | |
1ede94c5 JH |
1090 | if (node->call_for_symbol_and_aliases (check_callers, &has_hot_call, |
1091 | true)) | |
5970b079 EB |
1092 | return false; |
1093 | if (!cold && !has_hot_call) | |
1094 | return false; | |
1095 | return true; | |
09a2806f JH |
1096 | } |
1097 | ||
041cb615 JH |
1098 | /* Return true if WHERE of SIZE is a possible candidate for wrapper heuristics |
1099 | in estimate_edge_badness. */ | |
1100 | ||
1101 | static bool | |
1102 | wrapper_heuristics_may_apply (struct cgraph_node *where, int size) | |
1103 | { | |
1104 | return size < (DECL_DECLARED_INLINE_P (where->decl) | |
1105 | ? inline_insns_single (where, false) | |
1106 | : inline_insns_auto (where, false)); | |
1107 | } | |
1108 | ||
670cd5c5 JH |
1109 | /* A cost model driving the inlining heuristics in a way so the edges with |
1110 | smallest badness are inlined first. After each inlining is performed | |
0fa2e4df | 1111 | the costs of all caller edges of nodes affected are recomputed so the |
670cd5c5 | 1112 | metrics may accurately depend on values such as number of inlinable callers |
45a80bb9 | 1113 | of the function or function body size. */ |
670cd5c5 | 1114 | |
f0e1509b | 1115 | static sreal |
4c0f7679 | 1116 | edge_badness (struct cgraph_edge *edge, bool dump) |
670cd5c5 | 1117 | { |
f0e1509b | 1118 | sreal badness; |
ab38481c | 1119 | int growth; |
4adaad64 | 1120 | sreal edge_time, unspec_edge_time; |
d52f5295 | 1121 | struct cgraph_node *callee = edge->callee->ultimate_alias_target (); |
99b1c316 | 1122 | class ipa_fn_summary *callee_info = ipa_fn_summaries->get (callee); |
0bceb671 | 1123 | ipa_hints hints; |
a62bfab5 ML |
1124 | cgraph_node *caller = (edge->caller->inlined_to |
1125 | ? edge->caller->inlined_to | |
208e5afa | 1126 | : edge->caller); |
1aa14195 | 1127 | |
03dfc36d | 1128 | growth = estimate_edge_growth (edge); |
4adaad64 | 1129 | edge_time = estimate_edge_time (edge, &unspec_edge_time); |
37678631 | 1130 | hints = estimate_edge_hints (edge); |
d59171da | 1131 | gcc_checking_assert (edge_time >= 0); |
1bad9c18 JH |
1132 | /* Check that inlined time is better, but tolerate some roundoff issues. |
1133 | FIXME: When callee profile drops to 0 we account calls more. This | |
1134 | should be fixed by never doing that. */ | |
3f05a4f0 JH |
1135 | gcc_checking_assert ((edge_time * 100 |
1136 | - callee_info->time * 101).to_int () <= 0 | |
1bad9c18 | 1137 | || callee->count.ipa ().initialized_p ()); |
f658ad30 | 1138 | gcc_checking_assert (growth <= ipa_size_summaries->get (callee)->size); |
e89964e3 | 1139 | |
1ce18dc8 JH |
1140 | if (dump) |
1141 | { | |
464d0118 ML |
1142 | fprintf (dump_file, " Badness calculation for %s -> %s\n", |
1143 | edge->caller->dump_name (), | |
1144 | edge->callee->dump_name ()); | |
4adaad64 | 1145 | fprintf (dump_file, " size growth %i, time %f unspec %f ", |
1ce18dc8 | 1146 | growth, |
4adaad64 JH |
1147 | edge_time.to_double (), |
1148 | unspec_edge_time.to_double ()); | |
0bceb671 | 1149 | ipa_dump_hints (dump_file, hints); |
42f7b0fa JH |
1150 | if (big_speedup_p (edge)) |
1151 | fprintf (dump_file, " big_speedup"); | |
37678631 | 1152 | fprintf (dump_file, "\n"); |
1ce18dc8 | 1153 | } |
45a80bb9 JH |
1154 | |
1155 | /* Always prefer inlining saving code size. */ | |
1156 | if (growth <= 0) | |
1ce18dc8 | 1157 | { |
6d4ab5f8 | 1158 | badness = (sreal) (-SREAL_MIN_SIG + growth) << (SREAL_MAX_EXP / 256); |
1ce18dc8 | 1159 | if (dump) |
6d4ab5f8 | 1160 | fprintf (dump_file, " %f: Growth %d <= 0\n", badness.to_double (), |
1ce18dc8 JH |
1161 | growth); |
1162 | } | |
208e5afa JH |
1163 | /* Inlining into EXTERNAL functions is not going to change anything unless |
1164 | they are themselves inlined. */ | |
1165 | else if (DECL_EXTERNAL (caller->decl)) | |
1ce18dc8 | 1166 | { |
1ce18dc8 | 1167 | if (dump) |
208e5afa JH |
1168 | fprintf (dump_file, " max: function is external\n"); |
1169 | return sreal::max (); | |
1ce18dc8 | 1170 | } |
208e5afa | 1171 | /* When profile is available. Compute badness as: |
b4c0a884 | 1172 | |
208e5afa | 1173 | time_saved * caller_count |
133a84ab JH |
1174 | goodness = ------------------------------------------------- |
1175 | growth_of_caller * overall_growth * combined_size | |
d59171da JH |
1176 | |
1177 | badness = - goodness | |
b4c0a884 | 1178 | |
208e5afa JH |
1179 | Again use negative value to make calls with profile appear hotter |
1180 | then calls without. | |
b4c0a884 | 1181 | */ |
3995f3a2 | 1182 | else if (opt_for_fn (caller->decl, flag_guess_branch_prob) |
1bad9c18 | 1183 | || caller->count.ipa ().nonzero_p ()) |
670cd5c5 | 1184 | { |
6d4ab5f8 | 1185 | sreal numerator, denominator; |
41f669d8 | 1186 | int overall_growth; |
b5351388 | 1187 | sreal freq = edge->sreal_frequency (); |
208e5afa | 1188 | |
ea8dd3b6 | 1189 | numerator = inlining_speedup (edge, freq, unspec_edge_time, edge_time); |
0009a6c3 | 1190 | if (numerator <= 0) |
208e5afa | 1191 | numerator = ((sreal) 1 >> 8); |
1bad9c18 JH |
1192 | if (caller->count.ipa ().nonzero_p ()) |
1193 | numerator *= caller->count.ipa ().to_gcov_type (); | |
1194 | else if (caller->count.ipa ().initialized_p ()) | |
208e5afa JH |
1195 | numerator = numerator >> 11; |
1196 | denominator = growth; | |
41f669d8 JH |
1197 | |
1198 | overall_growth = callee_info->growth; | |
1199 | ||
1200 | /* Look for inliner wrappers of the form: | |
1201 | ||
1202 | inline_caller () | |
1203 | { | |
1204 | do_fast_job... | |
1205 | if (need_more_work) | |
1206 | noninline_callee (); | |
1207 | } | |
956d615d | 1208 | Without penalizing this case, we usually inline noninline_callee |
41f669d8 JH |
1209 | into the inline_caller because overall_growth is small preventing |
1210 | further inlining of inline_caller. | |
1211 | ||
1212 | Penalize only callgraph edges to functions with small overall | |
1213 | growth ... | |
1214 | */ | |
1215 | if (growth > overall_growth | |
1216 | /* ... and having only one caller which is not inlined ... */ | |
1217 | && callee_info->single_caller | |
a62bfab5 | 1218 | && !edge->caller->inlined_to |
41f669d8 | 1219 | /* ... and edges executed only conditionally ... */ |
b5351388 | 1220 | && freq < 1 |
41f669d8 JH |
1221 | /* ... consider case where callee is not inline but caller is ... */ |
1222 | && ((!DECL_DECLARED_INLINE_P (edge->callee->decl) | |
1223 | && DECL_DECLARED_INLINE_P (caller->decl)) | |
1224 | /* ... or when early optimizers decided to split and edge | |
1225 | frequency still indicates splitting is a win ... */ | |
1226 | || (callee->split_part && !caller->split_part | |
1e83bd70 JH |
1227 | && freq * 100 |
1228 | < opt_for_fn (caller->decl, | |
1229 | param_partial_inlining_entry_probability) | |
41f669d8 JH |
1230 | /* ... and do not overwrite user specified hints. */ |
1231 | && (!DECL_DECLARED_INLINE_P (edge->callee->decl) | |
1232 | || DECL_DECLARED_INLINE_P (caller->decl))))) | |
1233 | { | |
56f62793 | 1234 | ipa_fn_summary *caller_info = ipa_fn_summaries->get (caller); |
41f669d8 JH |
1235 | int caller_growth = caller_info->growth; |
1236 | ||
1237 | /* Only apply the penalty when caller looks like inline candidate, | |
2925cad2 | 1238 | and it is not called once. */ |
41f669d8 JH |
1239 | if (!caller_info->single_caller && overall_growth < caller_growth |
1240 | && caller_info->inlinable | |
041cb615 JH |
1241 | && wrapper_heuristics_may_apply |
1242 | (caller, ipa_size_summaries->get (caller)->size)) | |
41f669d8 JH |
1243 | { |
1244 | if (dump) | |
1245 | fprintf (dump_file, | |
1246 | " Wrapper penalty. Increasing growth %i to %i\n", | |
1247 | overall_growth, caller_growth); | |
1248 | overall_growth = caller_growth; | |
1249 | } | |
1250 | } | |
1251 | if (overall_growth > 0) | |
1252 | { | |
e4112065 | 1253 | /* Strongly prefer functions with few callers that can be inlined |
41f669d8 JH |
1254 | fully. The square root here leads to smaller binaries at average. |
1255 | Watch however for extreme cases and return to linear function | |
1256 | when growth is large. */ | |
1257 | if (overall_growth < 256) | |
1258 | overall_growth *= overall_growth; | |
1259 | else | |
1260 | overall_growth += 256 * 256 - 256; | |
1261 | denominator *= overall_growth; | |
1262 | } | |
f658ad30 | 1263 | denominator *= ipa_size_summaries->get (caller)->size + growth; |
6d4ab5f8 JH |
1264 | |
1265 | badness = - numerator / denominator; | |
1266 | ||
1ce18dc8 JH |
1267 | if (dump) |
1268 | { | |
1269 | fprintf (dump_file, | |
16998094 JM |
1270 | " %f: guessed profile. frequency %f, count %" PRId64 |
1271 | " caller count %" PRId64 | |
ea8dd3b6 | 1272 | " time saved %f" |
41f669d8 JH |
1273 | " overall growth %i (current) %i (original)" |
1274 | " %i (compensated)\n", | |
1275 | badness.to_double (), | |
b5351388 | 1276 | freq.to_double (), |
1bad9c18 JH |
1277 | edge->count.ipa ().initialized_p () ? edge->count.ipa ().to_gcov_type () : -1, |
1278 | caller->count.ipa ().initialized_p () ? caller->count.ipa ().to_gcov_type () : -1, | |
ea8dd3b6 | 1279 | inlining_speedup (edge, freq, unspec_edge_time, edge_time).to_double (), |
d59171da | 1280 | estimate_growth (callee), |
41f669d8 | 1281 | callee_info->growth, overall_growth); |
1ce18dc8 | 1282 | } |
45a80bb9 JH |
1283 | } |
1284 | /* When function local profile is not available or it does not give | |
956d615d | 1285 | useful information (i.e. frequency is zero), base the cost on |
45a80bb9 JH |
1286 | loop nest and overall size growth, so we optimize for overall number |
1287 | of functions fully inlined in program. */ | |
1288 | else | |
1289 | { | |
56f62793 | 1290 | int nest = MIN (ipa_call_summaries->get (edge)->loop_depth, 8); |
6d4ab5f8 | 1291 | badness = growth; |
670cd5c5 | 1292 | |
45a80bb9 | 1293 | /* Decrease badness if call is nested. */ |
b8698a0f | 1294 | if (badness > 0) |
f0e1509b | 1295 | badness = badness >> nest; |
45a80bb9 | 1296 | else |
6d4ab5f8 | 1297 | badness = badness << nest; |
1ce18dc8 | 1298 | if (dump) |
41f669d8 JH |
1299 | fprintf (dump_file, " %f: no profile. nest %i\n", |
1300 | badness.to_double (), nest); | |
670cd5c5 | 1301 | } |
6d4ab5f8 | 1302 | gcc_checking_assert (badness != 0); |
1ce18dc8 | 1303 | |
3dafb85c | 1304 | if (edge->recursive_p ()) |
6d4ab5f8 JH |
1305 | badness = badness.shift (badness > 0 ? 4 : -4); |
1306 | if ((hints & (INLINE_HINT_indirect_call | |
1307 | | INLINE_HINT_loop_iterations | |
6d4ab5f8 JH |
1308 | | INLINE_HINT_loop_stride)) |
1309 | || callee_info->growth <= 0) | |
1310 | badness = badness.shift (badness > 0 ? -2 : 2); | |
1311 | if (hints & (INLINE_HINT_same_scc)) | |
1312 | badness = badness.shift (badness > 0 ? 3 : -3); | |
1313 | else if (hints & (INLINE_HINT_in_scc)) | |
1314 | badness = badness.shift (badness > 0 ? 2 : -2); | |
1315 | else if (hints & (INLINE_HINT_cross_module)) | |
1316 | badness = badness.shift (badness > 0 ? 1 : -1); | |
208e5afa JH |
1317 | if (DECL_DISREGARD_INLINE_LIMITS (callee->decl)) |
1318 | badness = badness.shift (badness > 0 ? -4 : 4); | |
1319 | else if ((hints & INLINE_HINT_declared_inline)) | |
6d4ab5f8 JH |
1320 | badness = badness.shift (badness > 0 ? -3 : 3); |
1321 | if (dump) | |
1322 | fprintf (dump_file, " Adjusted by hints %f\n", badness.to_double ()); | |
1323 | return badness; | |
670cd5c5 JH |
1324 | } |
1325 | ||
9b8051b4 | 1326 | /* Recompute badness of EDGE and update its key in HEAP if needed. */ |
4c0f7679 | 1327 | static inline void |
4a910049 | 1328 | update_edge_key (edge_heap_t *heap, struct cgraph_edge *edge) |
9b8051b4 | 1329 | { |
f0e1509b | 1330 | sreal badness = edge_badness (edge, false); |
9b8051b4 JH |
1331 | if (edge->aux) |
1332 | { | |
4a910049 ML |
1333 | edge_heap_node_t *n = (edge_heap_node_t *) edge->aux; |
1334 | gcc_checking_assert (n->get_data () == edge); | |
9b8051b4 | 1335 | |
6d4ab5f8 | 1336 | /* fibonacci_heap::replace_key does busy updating of the |
956d615d | 1337 | heap that is unnecessarily expensive. |
6d4ab5f8 JH |
1338 | We do lazy increases: after extracting minimum if the key |
1339 | turns out to be out of date, it is re-inserted into heap | |
1340 | with correct value. */ | |
4a910049 | 1341 | if (badness < n->get_key ()) |
9b8051b4 | 1342 | { |
4c0f7679 JH |
1343 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1344 | { | |
1345 | fprintf (dump_file, | |
464d0118 ML |
1346 | " decreasing badness %s -> %s, %f to %f\n", |
1347 | edge->caller->dump_name (), | |
1348 | edge->callee->dump_name (), | |
6d4ab5f8 JH |
1349 | n->get_key ().to_double (), |
1350 | badness.to_double ()); | |
4c0f7679 | 1351 | } |
4a910049 | 1352 | heap->decrease_key (n, badness); |
9b8051b4 JH |
1353 | } |
1354 | } | |
1355 | else | |
4c0f7679 JH |
1356 | { |
1357 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1358 | { | |
1359 | fprintf (dump_file, | |
464d0118 ML |
1360 | " enqueuing call %s -> %s, badness %f\n", |
1361 | edge->caller->dump_name (), | |
1362 | edge->callee->dump_name (), | |
6d4ab5f8 | 1363 | badness.to_double ()); |
4c0f7679 | 1364 | } |
4a910049 | 1365 | edge->aux = heap->insert (badness, edge); |
4c0f7679 | 1366 | } |
9b8051b4 JH |
1367 | } |
1368 | ||
40fda55b JH |
1369 | |
1370 | /* NODE was inlined. | |
956d615d | 1371 | All caller edges needs to be reset because |
40fda55b JH |
1372 | size estimates change. Similarly callees needs reset |
1373 | because better context may be known. */ | |
1374 | ||
1375 | static void | |
1376 | reset_edge_caches (struct cgraph_node *node) | |
1377 | { | |
1378 | struct cgraph_edge *edge; | |
1379 | struct cgraph_edge *e = node->callees; | |
1380 | struct cgraph_node *where = node; | |
e55637b7 | 1381 | struct ipa_ref *ref; |
40fda55b | 1382 | |
a62bfab5 ML |
1383 | if (where->inlined_to) |
1384 | where = where->inlined_to; | |
40fda55b | 1385 | |
ac6f2e59 JH |
1386 | reset_node_cache (where); |
1387 | ||
9fb50ad8 ML |
1388 | if (edge_growth_cache != NULL) |
1389 | for (edge = where->callers; edge; edge = edge->next_caller) | |
1390 | if (edge->inline_failed) | |
1391 | edge_growth_cache->remove (edge); | |
e55637b7 ML |
1392 | |
1393 | FOR_EACH_ALIAS (where, ref) | |
1394 | reset_edge_caches (dyn_cast <cgraph_node *> (ref->referring)); | |
40fda55b JH |
1395 | |
1396 | if (!e) | |
1397 | return; | |
1398 | ||
1399 | while (true) | |
1400 | if (!e->inline_failed && e->callee->callees) | |
1401 | e = e->callee->callees; | |
1402 | else | |
1403 | { | |
9fb50ad8 ML |
1404 | if (edge_growth_cache != NULL && e->inline_failed) |
1405 | edge_growth_cache->remove (e); | |
40fda55b JH |
1406 | if (e->next_callee) |
1407 | e = e->next_callee; | |
1408 | else | |
1409 | { | |
1410 | do | |
1411 | { | |
1412 | if (e->caller == node) | |
1413 | return; | |
1414 | e = e->caller->callers; | |
1415 | } | |
1416 | while (!e->next_callee); | |
1417 | e = e->next_callee; | |
1418 | } | |
1419 | } | |
1420 | } | |
1421 | ||
1422 | /* Recompute HEAP nodes for each of caller of NODE. | |
1423 | UPDATED_NODES track nodes we already visited, to avoid redundant work. | |
1424 | When CHECK_INLINABLITY_FOR is set, re-check for specified edge that | |
1425 | it is inlinable. Otherwise check all edges. */ | |
670cd5c5 JH |
1426 | |
1427 | static void | |
4a910049 | 1428 | update_caller_keys (edge_heap_t *heap, struct cgraph_node *node, |
40fda55b JH |
1429 | bitmap updated_nodes, |
1430 | struct cgraph_edge *check_inlinablity_for) | |
670cd5c5 JH |
1431 | { |
1432 | struct cgraph_edge *edge; | |
e55637b7 | 1433 | struct ipa_ref *ref; |
670cd5c5 | 1434 | |
56f62793 | 1435 | if ((!node->alias && !ipa_fn_summaries->get (node)->inlinable) |
a62bfab5 | 1436 | || node->inlined_to) |
670cd5c5 | 1437 | return; |
4325656f | 1438 | if (!bitmap_set_bit (updated_nodes, node->get_uid ())) |
670cd5c5 | 1439 | return; |
670cd5c5 | 1440 | |
e55637b7 ML |
1441 | FOR_EACH_ALIAS (node, ref) |
1442 | { | |
1443 | struct cgraph_node *alias = dyn_cast <cgraph_node *> (ref->referring); | |
1444 | update_caller_keys (heap, alias, updated_nodes, check_inlinablity_for); | |
1445 | } | |
39e2db00 | 1446 | |
cdc029b9 | 1447 | for (edge = node->callers; edge; edge = edge->next_caller) |
4c0f7679 JH |
1448 | if (edge->inline_failed) |
1449 | { | |
40fda55b JH |
1450 | if (!check_inlinablity_for |
1451 | || check_inlinablity_for == edge) | |
f10d1a74 | 1452 | { |
9a4841a3 JH |
1453 | if (can_inline_edge_p (edge, false) |
1454 | && want_inline_small_function_p (edge, false) | |
1455 | && can_inline_edge_by_limits_p (edge, false)) | |
40fda55b JH |
1456 | update_edge_key (heap, edge); |
1457 | else if (edge->aux) | |
1458 | { | |
1459 | report_inline_failed_reason (edge); | |
4a910049 | 1460 | heap->delete_node ((edge_heap_node_t *) edge->aux); |
40fda55b JH |
1461 | edge->aux = NULL; |
1462 | } | |
f10d1a74 | 1463 | } |
40fda55b JH |
1464 | else if (edge->aux) |
1465 | update_edge_key (heap, edge); | |
4c0f7679 | 1466 | } |
9b8051b4 JH |
1467 | } |
1468 | ||
7c6f2fb9 JH |
1469 | /* Recompute HEAP nodes for each uninlined call in NODE |
1470 | If UPDATE_SINCE is non-NULL check if edges called within that function | |
1471 | are inlinable (typically UPDATE_SINCE is the inline clone we introduced | |
1472 | where all edges have new context). | |
1473 | ||
9b8051b4 JH |
1474 | This is used when we know that edge badnesses are going only to increase |
1475 | (we introduced new call site) and thus all we need is to insert newly | |
1476 | created edges into heap. */ | |
1477 | ||
1478 | static void | |
4a910049 | 1479 | update_callee_keys (edge_heap_t *heap, struct cgraph_node *node, |
7c6f2fb9 | 1480 | struct cgraph_node *update_since, |
9b8051b4 JH |
1481 | bitmap updated_nodes) |
1482 | { | |
1483 | struct cgraph_edge *e = node->callees; | |
7c6f2fb9 | 1484 | bool check_inlinability = update_since == node; |
09a2806f | 1485 | |
9b8051b4 JH |
1486 | if (!e) |
1487 | return; | |
1488 | while (true) | |
1489 | if (!e->inline_failed && e->callee->callees) | |
7c6f2fb9 JH |
1490 | { |
1491 | if (e->callee == update_since) | |
1492 | check_inlinability = true; | |
1493 | e = e->callee->callees; | |
1494 | } | |
9b8051b4 | 1495 | else |
670cd5c5 | 1496 | { |
a5b1779f JH |
1497 | enum availability avail; |
1498 | struct cgraph_node *callee; | |
7c6f2fb9 JH |
1499 | if (!check_inlinability) |
1500 | { | |
1501 | if (e->aux | |
1502 | && !bitmap_bit_p (updated_nodes, | |
1503 | e->callee->ultimate_alias_target | |
1504 | (&avail, e->caller)->get_uid ())) | |
1505 | update_edge_key (heap, e); | |
1506 | } | |
58696ce5 | 1507 | /* We do not reset callee growth cache here. Since we added a new call, |
956d615d | 1508 | growth should have just increased and consequently badness metric |
58696ce5 | 1509 | don't need updating. */ |
7c6f2fb9 JH |
1510 | else if (e->inline_failed |
1511 | && (callee = e->callee->ultimate_alias_target (&avail, | |
1512 | e->caller)) | |
1513 | && avail >= AVAIL_AVAILABLE | |
1514 | && ipa_fn_summaries->get (callee) != NULL | |
1515 | && ipa_fn_summaries->get (callee)->inlinable | |
1516 | && !bitmap_bit_p (updated_nodes, callee->get_uid ())) | |
670cd5c5 | 1517 | { |
9a4841a3 JH |
1518 | if (can_inline_edge_p (e, false) |
1519 | && want_inline_small_function_p (e, false) | |
1520 | && can_inline_edge_by_limits_p (e, false)) | |
7c6f2fb9 JH |
1521 | { |
1522 | gcc_checking_assert (check_inlinability || can_inline_edge_p (e, false)); | |
1523 | gcc_checking_assert (check_inlinability || e->aux); | |
1524 | update_edge_key (heap, e); | |
1525 | } | |
40fda55b JH |
1526 | else if (e->aux) |
1527 | { | |
1528 | report_inline_failed_reason (e); | |
4a910049 | 1529 | heap->delete_node ((edge_heap_node_t *) e->aux); |
40fda55b JH |
1530 | e->aux = NULL; |
1531 | } | |
9b8051b4 | 1532 | } |
7c6f2fb9 JH |
1533 | /* In case we redirected to unreachable node we only need to remove the |
1534 | fibheap entry. */ | |
1535 | else if (e->aux) | |
1536 | { | |
1537 | heap->delete_node ((edge_heap_node_t *) e->aux); | |
1538 | e->aux = NULL; | |
1539 | } | |
9b8051b4 JH |
1540 | if (e->next_callee) |
1541 | e = e->next_callee; | |
1542 | else | |
1543 | { | |
1544 | do | |
1ce18dc8 | 1545 | { |
9b8051b4 JH |
1546 | if (e->caller == node) |
1547 | return; | |
7c6f2fb9 JH |
1548 | if (e->caller == update_since) |
1549 | check_inlinability = false; | |
9b8051b4 | 1550 | e = e->caller->callers; |
1ce18dc8 | 1551 | } |
9b8051b4 JH |
1552 | while (!e->next_callee); |
1553 | e = e->next_callee; | |
670cd5c5 | 1554 | } |
670cd5c5 JH |
1555 | } |
1556 | } | |
1557 | ||
670cd5c5 | 1558 | /* Enqueue all recursive calls from NODE into priority queue depending on |
0fa2e4df | 1559 | how likely we want to recursively inline the call. */ |
670cd5c5 | 1560 | |
ca31b95f JH |
1561 | static void |
1562 | lookup_recursive_calls (struct cgraph_node *node, struct cgraph_node *where, | |
4a910049 | 1563 | edge_heap_t *heap) |
ca31b95f JH |
1564 | { |
1565 | struct cgraph_edge *e; | |
a5b1779f JH |
1566 | enum availability avail; |
1567 | ||
ca31b95f | 1568 | for (e = where->callees; e; e = e->next_callee) |
a5b1779f | 1569 | if (e->callee == node |
e6007a27 | 1570 | || (e->callee->ultimate_alias_target (&avail, e->caller) == node |
d52f5295 | 1571 | && avail > AVAIL_INTERPOSABLE)) |
0009a6c3 | 1572 | heap->insert (-e->sreal_frequency (), e); |
ca31b95f JH |
1573 | for (e = where->callees; e; e = e->next_callee) |
1574 | if (!e->inline_failed) | |
670cd5c5 | 1575 | lookup_recursive_calls (node, e->callee, heap); |
ca31b95f JH |
1576 | } |
1577 | ||
1578 | /* Decide on recursive inlining: in the case function has recursive calls, | |
3e293154 | 1579 | inline until body size reaches given argument. If any new indirect edges |
e56f5f3e JJ |
1580 | are discovered in the process, add them to *NEW_EDGES, unless NEW_EDGES |
1581 | is NULL. */ | |
670cd5c5 JH |
1582 | |
1583 | static bool | |
4c0f7679 | 1584 | recursive_inlining (struct cgraph_edge *edge, |
d52f5295 | 1585 | vec<cgraph_edge *> *new_edges) |
ca31b95f | 1586 | { |
1e83bd70 JH |
1587 | cgraph_node *to = (edge->caller->inlined_to |
1588 | ? edge->caller->inlined_to : edge->caller); | |
1589 | int limit = opt_for_fn (to->decl, | |
1590 | param_max_inline_insns_recursive_auto); | |
d1704358 | 1591 | edge_heap_t heap (sreal::min ()); |
d7d1d041 | 1592 | struct cgraph_node *node; |
ca31b95f | 1593 | struct cgraph_edge *e; |
4c0f7679 | 1594 | struct cgraph_node *master_clone = NULL, *next; |
ca31b95f JH |
1595 | int depth = 0; |
1596 | int n = 0; | |
1597 | ||
d7d1d041 | 1598 | node = edge->caller; |
a62bfab5 ML |
1599 | if (node->inlined_to) |
1600 | node = node->inlined_to; | |
d7d1d041 | 1601 | |
67348ccc | 1602 | if (DECL_DECLARED_INLINE_P (node->decl)) |
1e83bd70 | 1603 | limit = opt_for_fn (to->decl, param_max_inline_insns_recursive); |
ca31b95f JH |
1604 | |
1605 | /* Make sure that function is small enough to be considered for inlining. */ | |
4c0f7679 | 1606 | if (estimate_size_after_inlining (node, edge) >= limit) |
670cd5c5 | 1607 | return false; |
4a910049 ML |
1608 | lookup_recursive_calls (node, node, &heap); |
1609 | if (heap.empty ()) | |
1610 | return false; | |
ca31b95f JH |
1611 | |
1612 | if (dump_file) | |
b8698a0f | 1613 | fprintf (dump_file, |
670cd5c5 | 1614 | " Performing recursive inlining on %s\n", |
fec39fa6 | 1615 | node->name ()); |
ca31b95f | 1616 | |
ca31b95f | 1617 | /* Do the inlining and update list of recursive call during process. */ |
4a910049 | 1618 | while (!heap.empty ()) |
ca31b95f | 1619 | { |
4a910049 | 1620 | struct cgraph_edge *curr = heap.extract_min (); |
6ced940d | 1621 | struct cgraph_node *cnode, *dest = curr->callee; |
d7d1d041 | 1622 | |
9a4841a3 | 1623 | if (!can_inline_edge_p (curr, true) |
4eb50396 | 1624 | || !can_inline_edge_by_limits_p (curr, true)) |
4c0f7679 JH |
1625 | continue; |
1626 | ||
6ced940d JH |
1627 | /* MASTER_CLONE is produced in the case we already started modified |
1628 | the function. Be sure to redirect edge to the original body before | |
1629 | estimating growths otherwise we will be seeing growths after inlining | |
1630 | the already modified body. */ | |
1631 | if (master_clone) | |
1632 | { | |
3dafb85c | 1633 | curr->redirect_callee (master_clone); |
9fb50ad8 ML |
1634 | if (edge_growth_cache != NULL) |
1635 | edge_growth_cache->remove (curr); | |
6ced940d JH |
1636 | } |
1637 | ||
1638 | if (estimate_size_after_inlining (node, curr) > limit) | |
1639 | { | |
3dafb85c | 1640 | curr->redirect_callee (dest); |
9fb50ad8 ML |
1641 | if (edge_growth_cache != NULL) |
1642 | edge_growth_cache->remove (curr); | |
6ced940d JH |
1643 | break; |
1644 | } | |
1645 | ||
c5a4444c JH |
1646 | depth = 1; |
1647 | for (cnode = curr->caller; | |
a62bfab5 | 1648 | cnode->inlined_to; cnode = cnode->callers->caller) |
67348ccc | 1649 | if (node->decl |
d52f5295 | 1650 | == curr->callee->ultimate_alias_target ()->decl) |
f791d333 | 1651 | depth++; |
c5a4444c | 1652 | |
4c0f7679 | 1653 | if (!want_inline_self_recursive_call_p (curr, node, false, depth)) |
6ced940d | 1654 | { |
3dafb85c | 1655 | curr->redirect_callee (dest); |
9fb50ad8 ML |
1656 | if (edge_growth_cache != NULL) |
1657 | edge_growth_cache->remove (curr); | |
6ced940d JH |
1658 | continue; |
1659 | } | |
ca31b95f | 1660 | |
670cd5c5 | 1661 | if (dump_file) |
c5a4444c | 1662 | { |
b8698a0f | 1663 | fprintf (dump_file, |
c5a4444c | 1664 | " Inlining call of depth %i", depth); |
ae94bb0e | 1665 | if (node->count.nonzero_p () && curr->count.initialized_p ()) |
c5a4444c JH |
1666 | { |
1667 | fprintf (dump_file, " called approx. %.2f times per call", | |
3995f3a2 JH |
1668 | (double)curr->count.to_gcov_type () |
1669 | / node->count.to_gcov_type ()); | |
c5a4444c JH |
1670 | } |
1671 | fprintf (dump_file, "\n"); | |
1672 | } | |
4c0f7679 JH |
1673 | if (!master_clone) |
1674 | { | |
1675 | /* We need original clone to copy around. */ | |
d52f5295 | 1676 | master_clone = node->create_clone (node->decl, node->count, |
1bad9c18 | 1677 | false, vNULL, true, NULL, NULL); |
4c0f7679 JH |
1678 | for (e = master_clone->callees; e; e = e->next_callee) |
1679 | if (!e->inline_failed) | |
1bad9c18 | 1680 | clone_inlined_nodes (e, true, false, NULL); |
3dafb85c | 1681 | curr->redirect_callee (master_clone); |
9fb50ad8 ML |
1682 | if (edge_growth_cache != NULL) |
1683 | edge_growth_cache->remove (curr); | |
4c0f7679 JH |
1684 | } |
1685 | ||
c170d40f | 1686 | inline_call (curr, false, new_edges, &overall_size, true); |
b914768c | 1687 | reset_node_cache (node); |
4a910049 | 1688 | lookup_recursive_calls (node, curr->callee, &heap); |
ca31b95f JH |
1689 | n++; |
1690 | } | |
4c0f7679 | 1691 | |
4a910049 | 1692 | if (!heap.empty () && dump_file) |
c5a4444c | 1693 | fprintf (dump_file, " Recursive inlining growth limit met.\n"); |
4c0f7679 JH |
1694 | |
1695 | if (!master_clone) | |
1696 | return false; | |
1697 | ||
4174a33a DM |
1698 | if (dump_enabled_p ()) |
1699 | dump_printf_loc (MSG_NOTE, edge->call_stmt, | |
1700 | "\n Inlined %i times, " | |
1701 | "body grown from size %i to %i, time %f to %f\n", n, | |
f658ad30 JH |
1702 | ipa_size_summaries->get (master_clone)->size, |
1703 | ipa_size_summaries->get (node)->size, | |
4174a33a DM |
1704 | ipa_fn_summaries->get (master_clone)->time.to_double (), |
1705 | ipa_fn_summaries->get (node)->time.to_double ()); | |
ca31b95f JH |
1706 | |
1707 | /* Remove master clone we used for inlining. We rely that clones inlined | |
1708 | into master clone gets queued just before master clone so we don't | |
1709 | need recursion. */ | |
3dafb85c | 1710 | for (node = symtab->first_function (); node != master_clone; |
96fc428c JH |
1711 | node = next) |
1712 | { | |
3dafb85c | 1713 | next = symtab->next_function (node); |
a62bfab5 | 1714 | if (node->inlined_to == master_clone) |
d52f5295 | 1715 | node->remove (); |
96fc428c | 1716 | } |
d52f5295 | 1717 | master_clone->remove (); |
4c0f7679 | 1718 | return true; |
ca31b95f JH |
1719 | } |
1720 | ||
09a2806f | 1721 | |
88512ba0 | 1722 | /* Given whole compilation unit estimate of INSNS, compute how large we can |
b7c27d51 | 1723 | allow the unit to grow. */ |
09a2806f | 1724 | |
8fcfc44f | 1725 | static int64_t |
1e83bd70 | 1726 | compute_max_insns (cgraph_node *node, int insns) |
b7c27d51 JH |
1727 | { |
1728 | int max_insns = insns; | |
1e83bd70 JH |
1729 | if (max_insns < opt_for_fn (node->decl, param_large_unit_insns)) |
1730 | max_insns = opt_for_fn (node->decl, param_large_unit_insns); | |
b7c27d51 | 1731 | |
a9243bfc | 1732 | return ((int64_t) max_insns |
1e83bd70 | 1733 | * (100 + opt_for_fn (node->decl, param_inline_unit_growth)) / 100); |
b7c27d51 JH |
1734 | } |
1735 | ||
09a2806f | 1736 | |
3e293154 | 1737 | /* Compute badness of all edges in NEW_EDGES and add them to the HEAP. */ |
09a2806f | 1738 | |
3e293154 | 1739 | static void |
4a910049 | 1740 | add_new_edges_to_heap (edge_heap_t *heap, vec<cgraph_edge *> new_edges) |
3e293154 | 1741 | { |
9771b263 | 1742 | while (new_edges.length () > 0) |
3e293154 | 1743 | { |
9771b263 | 1744 | struct cgraph_edge *edge = new_edges.pop (); |
3e293154 MJ |
1745 | |
1746 | gcc_assert (!edge->aux); | |
4517b378 | 1747 | gcc_assert (edge->callee); |
a5b1779f | 1748 | if (edge->inline_failed |
4c0f7679 | 1749 | && can_inline_edge_p (edge, true) |
9a4841a3 JH |
1750 | && want_inline_small_function_p (edge, true) |
1751 | && can_inline_edge_by_limits_p (edge, true)) | |
4a910049 | 1752 | edge->aux = heap->insert (edge_badness (edge, false), edge); |
3e293154 MJ |
1753 | } |
1754 | } | |
1755 | ||
042ae7d2 JH |
1756 | /* Remove EDGE from the fibheap. */ |
1757 | ||
1758 | static void | |
1759 | heap_edge_removal_hook (struct cgraph_edge *e, void *data) | |
1760 | { | |
1761 | if (e->aux) | |
1762 | { | |
4a910049 | 1763 | ((edge_heap_t *)data)->delete_node ((edge_heap_node_t *)e->aux); |
042ae7d2 JH |
1764 | e->aux = NULL; |
1765 | } | |
1766 | } | |
3e293154 | 1767 | |
09ce3660 JH |
1768 | /* Return true if speculation of edge E seems useful. |
1769 | If ANTICIPATE_INLINING is true, be conservative and hope that E | |
1770 | may get inlined. */ | |
1771 | ||
1772 | bool | |
1773 | speculation_useful_p (struct cgraph_edge *e, bool anticipate_inlining) | |
1774 | { | |
4517b378 MJ |
1775 | /* If we have already decided to inline the edge, it seems useful. */ |
1776 | if (!e->inline_failed) | |
1777 | return true; | |
1778 | ||
09ce3660 | 1779 | enum availability avail; |
e6007a27 JH |
1780 | struct cgraph_node *target = e->callee->ultimate_alias_target (&avail, |
1781 | e->caller); | |
09ce3660 JH |
1782 | struct cgraph_edge *direct, *indirect; |
1783 | struct ipa_ref *ref; | |
1784 | ||
1785 | gcc_assert (e->speculative && !e->indirect_unknown_callee); | |
1786 | ||
3dafb85c | 1787 | if (!e->maybe_hot_p ()) |
09ce3660 JH |
1788 | return false; |
1789 | ||
1790 | /* See if IP optimizations found something potentially useful about the | |
1791 | function. For now we look only for CONST/PURE flags. Almost everything | |
1792 | else we propagate is useless. */ | |
1793 | if (avail >= AVAIL_AVAILABLE) | |
1794 | { | |
67348ccc | 1795 | int ecf_flags = flags_from_decl_or_type (target->decl); |
09ce3660 JH |
1796 | if (ecf_flags & ECF_CONST) |
1797 | { | |
3dafb85c | 1798 | e->speculative_call_info (direct, indirect, ref); |
09ce3660 JH |
1799 | if (!(indirect->indirect_info->ecf_flags & ECF_CONST)) |
1800 | return true; | |
1801 | } | |
1802 | else if (ecf_flags & ECF_PURE) | |
1803 | { | |
3dafb85c | 1804 | e->speculative_call_info (direct, indirect, ref); |
09ce3660 JH |
1805 | if (!(indirect->indirect_info->ecf_flags & ECF_PURE)) |
1806 | return true; | |
1807 | } | |
1808 | } | |
1809 | /* If we did not managed to inline the function nor redirect | |
1810 | to an ipa-cp clone (that are seen by having local flag set), | |
1811 | it is probably pointless to inline it unless hardware is missing | |
1812 | indirect call predictor. */ | |
87f94429 | 1813 | if (!anticipate_inlining && !target->local) |
09ce3660 JH |
1814 | return false; |
1815 | /* For overwritable targets there is not much to do. */ | |
4517b378 MJ |
1816 | if (!can_inline_edge_p (e, false) |
1817 | || !can_inline_edge_by_limits_p (e, false, true)) | |
09ce3660 JH |
1818 | return false; |
1819 | /* OK, speculation seems interesting. */ | |
1820 | return true; | |
1821 | } | |
1822 | ||
1823 | /* We know that EDGE is not going to be inlined. | |
1824 | See if we can remove speculation. */ | |
1825 | ||
1826 | static void | |
4a910049 | 1827 | resolve_noninline_speculation (edge_heap_t *edge_heap, struct cgraph_edge *edge) |
09ce3660 JH |
1828 | { |
1829 | if (edge->speculative && !speculation_useful_p (edge, false)) | |
1830 | { | |
1831 | struct cgraph_node *node = edge->caller; | |
a62bfab5 ML |
1832 | struct cgraph_node *where = node->inlined_to |
1833 | ? node->inlined_to : node; | |
0e3de1d4 | 1834 | auto_bitmap updated_nodes; |
09ce3660 | 1835 | |
1bad9c18 JH |
1836 | if (edge->count.ipa ().initialized_p ()) |
1837 | spec_rem += edge->count.ipa (); | |
3dafb85c | 1838 | edge->resolve_speculation (); |
09ce3660 | 1839 | reset_edge_caches (where); |
0bceb671 | 1840 | ipa_update_overall_fn_summary (where); |
09ce3660 JH |
1841 | update_caller_keys (edge_heap, where, |
1842 | updated_nodes, NULL); | |
7c6f2fb9 | 1843 | update_callee_keys (edge_heap, where, NULL, |
d0b66480 | 1844 | updated_nodes); |
09ce3660 JH |
1845 | } |
1846 | } | |
1847 | ||
bb1e543c JH |
1848 | /* Return true if NODE should be accounted for overall size estimate. |
1849 | Skip all nodes optimized for size so we can measure the growth of hot | |
1850 | part of program no matter of the padding. */ | |
1851 | ||
1852 | bool | |
1853 | inline_account_function_p (struct cgraph_node *node) | |
1854 | { | |
1855 | return (!DECL_EXTERNAL (node->decl) | |
1856 | && !opt_for_fn (node->decl, optimize_size) | |
1857 | && node->frequency != NODE_FREQUENCY_UNLIKELY_EXECUTED); | |
1858 | } | |
1859 | ||
41f669d8 JH |
1860 | /* Count number of callers of NODE and store it into DATA (that |
1861 | points to int. Worker for cgraph_for_node_and_aliases. */ | |
1862 | ||
1863 | static bool | |
1864 | sum_callers (struct cgraph_node *node, void *data) | |
1865 | { | |
1866 | struct cgraph_edge *e; | |
1867 | int *num_calls = (int *)data; | |
1868 | ||
1869 | for (e = node->callers; e; e = e->next_caller) | |
1870 | (*num_calls)++; | |
1871 | return false; | |
1872 | } | |
1873 | ||
97e59627 ML |
1874 | /* We only propagate across edges with non-interposable callee. */ |
1875 | ||
1876 | inline bool | |
1877 | ignore_edge_p (struct cgraph_edge *e) | |
1878 | { | |
1879 | enum availability avail; | |
1880 | e->callee->function_or_virtual_thunk_symbol (&avail, e->caller); | |
1881 | return (avail <= AVAIL_INTERPOSABLE); | |
1882 | } | |
1883 | ||
ca31b95f | 1884 | /* We use greedy algorithm for inlining of small functions: |
09a2806f JH |
1885 | All inline candidates are put into prioritized heap ordered in |
1886 | increasing badness. | |
ca31b95f | 1887 | |
09a2806f | 1888 | The inlining of small functions is bounded by unit growth parameters. */ |
ca31b95f JH |
1889 | |
1890 | static void | |
4c0f7679 | 1891 | inline_small_functions (void) |
ca31b95f JH |
1892 | { |
1893 | struct cgraph_node *node; | |
670cd5c5 | 1894 | struct cgraph_edge *edge; |
f0e1509b | 1895 | edge_heap_t edge_heap (sreal::min ()); |
0e3de1d4 | 1896 | auto_bitmap updated_nodes; |
1e83bd70 | 1897 | int min_size; |
d52f5295 | 1898 | auto_vec<cgraph_edge *> new_indirect_edges; |
09a2806f | 1899 | int initial_size = 0; |
3dafb85c | 1900 | struct cgraph_node **order = XCNEWVEC (cgraph_node *, symtab->cgraph_count); |
042ae7d2 | 1901 | struct cgraph_edge_hook_list *edge_removal_hook_holder; |
2bf86c84 | 1902 | new_indirect_edges.create (8); |
670cd5c5 | 1903 | |
042ae7d2 | 1904 | edge_removal_hook_holder |
4a910049 | 1905 | = symtab->add_edge_removal_hook (&heap_edge_removal_hook, &edge_heap); |
042ae7d2 | 1906 | |
1a3118e9 JH |
1907 | /* Compute overall unit size and other global parameters used by badness |
1908 | metrics. */ | |
ca31b95f | 1909 | |
3995f3a2 | 1910 | max_count = profile_count::uninitialized (); |
97e59627 | 1911 | ipa_reduced_postorder (order, true, ignore_edge_p); |
68cc8feb | 1912 | free (order); |
1a3118e9 | 1913 | |
c47d0034 | 1914 | FOR_EACH_DEFINED_FUNCTION (node) |
a62bfab5 | 1915 | if (!node->inlined_to) |
e7f23018 | 1916 | { |
bb1e543c | 1917 | if (!node->alias && node->analyzed |
29f1e2b1 JH |
1918 | && (node->has_gimple_body_p () || node->thunk.thunk_p) |
1919 | && opt_for_fn (node->decl, optimize)) | |
a5b1779f | 1920 | { |
99b1c316 | 1921 | class ipa_fn_summary *info = ipa_fn_summaries->get (node); |
67348ccc | 1922 | struct ipa_dfs_info *dfs = (struct ipa_dfs_info *) node->aux; |
ca31b95f | 1923 | |
5e750dc6 JH |
1924 | /* Do not account external functions, they will be optimized out |
1925 | if not inlined. Also only count the non-cold portion of program. */ | |
bb1e543c | 1926 | if (inline_account_function_p (node)) |
f658ad30 | 1927 | initial_size += ipa_size_summaries->get (node)->size; |
d59171da | 1928 | info->growth = estimate_growth (node); |
41f669d8 JH |
1929 | |
1930 | int num_calls = 0; | |
1931 | node->call_for_symbol_and_aliases (sum_callers, &num_calls, | |
1932 | true); | |
1933 | if (num_calls == 1) | |
1934 | info->single_caller = true; | |
bf3f6510 JH |
1935 | if (dfs && dfs->next_cycle) |
1936 | { | |
1937 | struct cgraph_node *n2; | |
1938 | int id = dfs->scc_no + 1; | |
1939 | for (n2 = node; n2; | |
1a03b929 | 1940 | n2 = ((struct ipa_dfs_info *) n2->aux)->next_cycle) |
29f1e2b1 JH |
1941 | if (opt_for_fn (n2->decl, optimize)) |
1942 | { | |
f658ad30 | 1943 | ipa_fn_summary *info2 = ipa_fn_summaries->get |
a62bfab5 | 1944 | (n2->inlined_to ? n2->inlined_to : n2); |
29f1e2b1 JH |
1945 | if (info2->scc_no) |
1946 | break; | |
1947 | info2->scc_no = id; | |
1948 | } | |
bf3f6510 | 1949 | } |
a5b1779f | 1950 | } |
09a2806f | 1951 | |
e7f23018 | 1952 | for (edge = node->callers; edge; edge = edge->next_caller) |
1bad9c18 | 1953 | max_count = max_count.max (edge->count.ipa ()); |
e7f23018 | 1954 | } |
b48ccf0d | 1955 | ipa_free_postorder_info (); |
ac6f2e59 | 1956 | initialize_growth_caches (); |
b48ccf0d JH |
1957 | |
1958 | if (dump_file) | |
1959 | fprintf (dump_file, | |
1960 | "\nDeciding on inlining of small functions. Starting with size %i.\n", | |
1961 | initial_size); | |
b7c27d51 | 1962 | |
8a8dccb2 | 1963 | overall_size = initial_size; |
85057983 | 1964 | min_size = overall_size; |
1a3118e9 | 1965 | |
a99be3c9 | 1966 | /* Populate the heap with all edges we might inline. */ |
1a3118e9 | 1967 | |
c47d0034 | 1968 | FOR_EACH_DEFINED_FUNCTION (node) |
09ce3660 JH |
1969 | { |
1970 | bool update = false; | |
bcda57c1 | 1971 | struct cgraph_edge *next = NULL; |
c1eed5a1 | 1972 | bool has_speculative = false; |
1a3118e9 | 1973 | |
29f1e2b1 JH |
1974 | if (!opt_for_fn (node->decl, optimize)) |
1975 | continue; | |
1976 | ||
09ce3660 | 1977 | if (dump_file) |
464d0118 | 1978 | fprintf (dump_file, "Enqueueing calls in %s.\n", node->dump_name ()); |
09ce3660 | 1979 | |
8fcfc44f | 1980 | for (edge = node->callees; edge; edge = edge->next_callee) |
09ce3660 | 1981 | { |
1a3118e9 | 1982 | if (edge->inline_failed |
09ce3660 | 1983 | && !edge->aux |
1a3118e9 JH |
1984 | && can_inline_edge_p (edge, true) |
1985 | && want_inline_small_function_p (edge, true) | |
9a4841a3 | 1986 | && can_inline_edge_by_limits_p (edge, true) |
1a3118e9 JH |
1987 | && edge->inline_failed) |
1988 | { | |
1989 | gcc_assert (!edge->aux); | |
4a910049 | 1990 | update_edge_key (&edge_heap, edge); |
1a3118e9 | 1991 | } |
c1eed5a1 JH |
1992 | if (edge->speculative) |
1993 | has_speculative = true; | |
1994 | } | |
1995 | if (has_speculative) | |
1996 | for (edge = node->callees; edge; edge = next) | |
2aae99f7 | 1997 | { |
2e98ac86 | 1998 | next = edge->next_callee; |
2aae99f7 XHL |
1999 | if (edge->speculative |
2000 | && !speculation_useful_p (edge, edge->aux != NULL)) | |
2001 | { | |
2002 | edge->resolve_speculation (); | |
2003 | update = true; | |
2004 | } | |
2aae99f7 | 2005 | } |
09ce3660 JH |
2006 | if (update) |
2007 | { | |
a62bfab5 ML |
2008 | struct cgraph_node *where = node->inlined_to |
2009 | ? node->inlined_to : node; | |
0bceb671 | 2010 | ipa_update_overall_fn_summary (where); |
09ce3660 | 2011 | reset_edge_caches (where); |
4a910049 | 2012 | update_caller_keys (&edge_heap, where, |
09ce3660 | 2013 | updated_nodes, NULL); |
7c6f2fb9 | 2014 | update_callee_keys (&edge_heap, where, NULL, |
2001028a | 2015 | updated_nodes); |
09ce3660 JH |
2016 | bitmap_clear (updated_nodes); |
2017 | } | |
2018 | } | |
1a3118e9 | 2019 | |
09a2806f | 2020 | gcc_assert (in_lto_p |
3995f3a2 | 2021 | || !(max_count > 0) |
09a2806f | 2022 | || (profile_info && flag_branch_probabilities)); |
b7c27d51 | 2023 | |
4a910049 | 2024 | while (!edge_heap.empty ()) |
ca31b95f | 2025 | { |
85057983 | 2026 | int old_size = overall_size; |
1ce18dc8 | 2027 | struct cgraph_node *where, *callee; |
f0e1509b ML |
2028 | sreal badness = edge_heap.min_key (); |
2029 | sreal current_badness; | |
1ce18dc8 | 2030 | int growth; |
670cd5c5 | 2031 | |
4a910049 | 2032 | edge = edge_heap.extract_min (); |
1ce18dc8 JH |
2033 | gcc_assert (edge->aux); |
2034 | edge->aux = NULL; | |
9de6f6c3 | 2035 | if (!edge->inline_failed || !edge->callee->analyzed) |
1ce18dc8 | 2036 | continue; |
cdc029b9 | 2037 | |
1bad9c18 JH |
2038 | /* Be sure that caches are maintained consistent. |
2039 | This check is affected by scaling roundoff errors when compiling for | |
2040 | IPA this we skip it in that case. */ | |
b914768c | 2041 | if (flag_checking && !edge->callee->count.ipa_p () |
517048ce | 2042 | && (!max_count.initialized_p () || !max_count.nonzero_p ())) |
1bad9c18 JH |
2043 | { |
2044 | sreal cached_badness = edge_badness (edge, false); | |
2045 | ||
2046 | int old_size_est = estimate_edge_size (edge); | |
2047 | sreal old_time_est = estimate_edge_time (edge); | |
2048 | int old_hints_est = estimate_edge_hints (edge); | |
2049 | ||
9fb50ad8 ML |
2050 | if (edge_growth_cache != NULL) |
2051 | edge_growth_cache->remove (edge); | |
b914768c JH |
2052 | reset_node_cache (edge->caller->inlined_to |
2053 | ? edge->caller->inlined_to | |
2054 | : edge->caller); | |
1bad9c18 JH |
2055 | gcc_assert (old_size_est == estimate_edge_size (edge)); |
2056 | gcc_assert (old_time_est == estimate_edge_time (edge)); | |
2057 | /* FIXME: | |
2058 | ||
2059 | gcc_assert (old_hints_est == estimate_edge_hints (edge)); | |
2060 | ||
2061 | fails with profile feedback because some hints depends on | |
2062 | maybe_hot_edge_p predicate and because callee gets inlined to other | |
2063 | calls, the edge may become cold. | |
2064 | This ought to be fixed by computing relative probabilities | |
2065 | for given invocation but that will be better done once whole | |
2066 | code is converted to sreals. Disable for now and revert to "wrong" | |
2067 | value so enable/disable checking paths agree. */ | |
9fb50ad8 | 2068 | edge_growth_cache->get (edge)->hints = old_hints_est + 1; |
1bad9c18 JH |
2069 | |
2070 | /* When updating the edge costs, we only decrease badness in the keys. | |
956d615d | 2071 | Increases of badness are handled lazily; when we see key with out |
1bad9c18 JH |
2072 | of date value on it, we re-insert it now. */ |
2073 | current_badness = edge_badness (edge, false); | |
2074 | gcc_assert (cached_badness == current_badness); | |
2075 | gcc_assert (current_badness >= badness); | |
2076 | } | |
f4118c87 JH |
2077 | else |
2078 | current_badness = edge_badness (edge, false); | |
cdc029b9 JH |
2079 | if (current_badness != badness) |
2080 | { | |
d75de25b | 2081 | if (edge_heap.min () && current_badness > edge_heap.min_key ()) |
6d4ab5f8 JH |
2082 | { |
2083 | edge->aux = edge_heap.insert (current_badness, edge); | |
2084 | continue; | |
2085 | } | |
2086 | else | |
2087 | badness = current_badness; | |
cdc029b9 | 2088 | } |
4c0f7679 | 2089 | |
9a4841a3 JH |
2090 | if (!can_inline_edge_p (edge, true) |
2091 | || !can_inline_edge_by_limits_p (edge, true)) | |
09ce3660 | 2092 | { |
4a910049 | 2093 | resolve_noninline_speculation (&edge_heap, edge); |
09ce3660 JH |
2094 | continue; |
2095 | } | |
cdc029b9 | 2096 | |
d52f5295 | 2097 | callee = edge->callee->ultimate_alias_target (); |
03dfc36d | 2098 | growth = estimate_edge_growth (edge); |
ca31b95f | 2099 | if (dump_file) |
ca31b95f | 2100 | { |
b8698a0f | 2101 | fprintf (dump_file, |
464d0118 ML |
2102 | "\nConsidering %s with %i size\n", |
2103 | callee->dump_name (), | |
f658ad30 | 2104 | ipa_size_summaries->get (callee)->size); |
b8698a0f | 2105 | fprintf (dump_file, |
464d0118 | 2106 | " to be inlined into %s in %s:%i\n" |
6d4ab5f8 | 2107 | " Estimated badness is %f, frequency %.2f.\n", |
464d0118 | 2108 | edge->caller->dump_name (), |
9e145afd | 2109 | edge->call_stmt |
355fe088 | 2110 | && (LOCATION_LOCUS (gimple_location ((const gimple *) |
1b7706c8 JJ |
2111 | edge->call_stmt)) |
2112 | > BUILTINS_LOCATION) | |
355fe088 | 2113 | ? gimple_filename ((const gimple *) edge->call_stmt) |
9e145afd N |
2114 | : "unknown", |
2115 | edge->call_stmt | |
355fe088 | 2116 | ? gimple_lineno ((const gimple *) edge->call_stmt) |
9e145afd | 2117 | : -1, |
6d4ab5f8 | 2118 | badness.to_double (), |
0cea1d34 | 2119 | edge->sreal_frequency ().to_double ()); |
1bad9c18 | 2120 | if (edge->count.ipa ().initialized_p ()) |
3995f3a2 JH |
2121 | { |
2122 | fprintf (dump_file, " Called "); | |
1bad9c18 | 2123 | edge->count.ipa ().dump (dump_file); |
517048ce | 2124 | fprintf (dump_file, " times\n"); |
3995f3a2 | 2125 | } |
1ce18dc8 | 2126 | if (dump_flags & TDF_DETAILS) |
4c0f7679 | 2127 | edge_badness (edge, true); |
ca31b95f JH |
2128 | } |
2129 | ||
1e83bd70 JH |
2130 | where = edge->caller; |
2131 | ||
2132 | if (overall_size + growth > compute_max_insns (where, min_size) | |
67348ccc | 2133 | && !DECL_DISREGARD_INLINE_LIMITS (callee->decl)) |
670cd5c5 | 2134 | { |
4c0f7679 JH |
2135 | edge->inline_failed = CIF_INLINE_UNIT_GROWTH_LIMIT; |
2136 | report_inline_failed_reason (edge); | |
4a910049 | 2137 | resolve_noninline_speculation (&edge_heap, edge); |
670cd5c5 JH |
2138 | continue; |
2139 | } | |
4c0f7679 JH |
2140 | |
2141 | if (!want_inline_small_function_p (edge, true)) | |
09ce3660 | 2142 | { |
4a910049 | 2143 | resolve_noninline_speculation (&edge_heap, edge); |
09ce3660 JH |
2144 | continue; |
2145 | } | |
09a2806f | 2146 | |
7c6f2fb9 JH |
2147 | profile_count old_count = callee->count; |
2148 | ||
7ba03e5e JL |
2149 | /* Heuristics for inlining small functions work poorly for |
2150 | recursive calls where we do effects similar to loop unrolling. | |
2151 | When inlining such edge seems profitable, leave decision on | |
09a2806f | 2152 | specific inliner. */ |
3dafb85c | 2153 | if (edge->recursive_p ()) |
670cd5c5 | 2154 | { |
a62bfab5 ML |
2155 | if (where->inlined_to) |
2156 | where = where->inlined_to; | |
4c0f7679 | 2157 | if (!recursive_inlining (edge, |
2bf86c84 JH |
2158 | opt_for_fn (edge->caller->decl, |
2159 | flag_indirect_inlining) | |
4c0f7679 | 2160 | ? &new_indirect_edges : NULL)) |
d7d1d041 RG |
2161 | { |
2162 | edge->inline_failed = CIF_RECURSIVE_INLINING; | |
4a910049 | 2163 | resolve_noninline_speculation (&edge_heap, edge); |
d7d1d041 RG |
2164 | continue; |
2165 | } | |
40fda55b | 2166 | reset_edge_caches (where); |
09a2806f JH |
2167 | /* Recursive inliner inlines all recursive calls of the function |
2168 | at once. Consequently we need to update all callee keys. */ | |
2bf86c84 | 2169 | if (opt_for_fn (edge->caller->decl, flag_indirect_inlining)) |
4a910049 | 2170 | add_new_edges_to_heap (&edge_heap, new_indirect_edges); |
7c6f2fb9 | 2171 | update_callee_keys (&edge_heap, where, where, updated_nodes); |
09ce3660 | 2172 | bitmap_clear (updated_nodes); |
670cd5c5 JH |
2173 | } |
2174 | else | |
2175 | { | |
4c0f7679 JH |
2176 | struct cgraph_node *outer_node = NULL; |
2177 | int depth = 0; | |
2178 | ||
7ba03e5e JL |
2179 | /* Consider the case where self recursive function A is inlined |
2180 | into B. This is desired optimization in some cases, since it | |
2181 | leads to effect similar of loop peeling and we might completely | |
2182 | optimize out the recursive call. However we must be extra | |
2183 | selective. */ | |
4c0f7679 JH |
2184 | |
2185 | where = edge->caller; | |
a62bfab5 | 2186 | while (where->inlined_to) |
670cd5c5 | 2187 | { |
67348ccc | 2188 | if (where->decl == callee->decl) |
4c0f7679 JH |
2189 | outer_node = where, depth++; |
2190 | where = where->callers->caller; | |
2191 | } | |
2192 | if (outer_node | |
2193 | && !want_inline_self_recursive_call_p (edge, outer_node, | |
2194 | true, depth)) | |
2195 | { | |
2196 | edge->inline_failed | |
67348ccc | 2197 | = (DECL_DISREGARD_INLINE_LIMITS (edge->callee->decl) |
4c0f7679 | 2198 | ? CIF_RECURSIVE_INLINING : CIF_UNSPECIFIED); |
4a910049 | 2199 | resolve_noninline_speculation (&edge_heap, edge); |
670cd5c5 JH |
2200 | continue; |
2201 | } | |
4c0f7679 JH |
2202 | else if (depth && dump_file) |
2203 | fprintf (dump_file, " Peeling recursion with depth %i\n", depth); | |
2204 | ||
a62bfab5 | 2205 | gcc_checking_assert (!callee->inlined_to); |
041cb615 JH |
2206 | |
2207 | int old_size = ipa_size_summaries->get (where)->size; | |
2208 | sreal old_time = ipa_fn_summaries->get (where)->time; | |
2209 | ||
c170d40f | 2210 | inline_call (edge, true, &new_indirect_edges, &overall_size, true); |
1f6f9079 | 2211 | reset_edge_caches (edge->callee); |
8d890d37 | 2212 | add_new_edges_to_heap (&edge_heap, new_indirect_edges); |
40fda55b | 2213 | |
041cb615 | 2214 | /* If caller's size and time increased we do not need to update |
956d615d | 2215 | all edges because badness is not going to decrease. */ |
041cb615 JH |
2216 | if (old_size <= ipa_size_summaries->get (where)->size |
2217 | && old_time <= ipa_fn_summaries->get (where)->time | |
2218 | /* Wrapper penalty may be non-monotonous in this respect. | |
2219 | Fortunately it only affects small functions. */ | |
2220 | && !wrapper_heuristics_may_apply (where, old_size)) | |
7c6f2fb9 JH |
2221 | update_callee_keys (&edge_heap, edge->callee, edge->callee, |
2222 | updated_nodes); | |
041cb615 | 2223 | else |
7c6f2fb9 JH |
2224 | update_callee_keys (&edge_heap, where, |
2225 | edge->callee, | |
2226 | updated_nodes); | |
670cd5c5 JH |
2227 | } |
2228 | where = edge->caller; | |
a62bfab5 ML |
2229 | if (where->inlined_to) |
2230 | where = where->inlined_to; | |
670cd5c5 JH |
2231 | |
2232 | /* Our profitability metric can depend on local properties | |
2233 | such as number of inlinable calls and size of the function body. | |
2234 | After inlining these properties might change for the function we | |
2235 | inlined into (since it's body size changed) and for the functions | |
2236 | called by function we inlined (since number of it inlinable callers | |
2237 | might change). */ | |
4a910049 | 2238 | update_caller_keys (&edge_heap, where, updated_nodes, NULL); |
208e5afa JH |
2239 | /* Offline copy count has possibly changed, recompute if profile is |
2240 | available. */ | |
7c6f2fb9 JH |
2241 | struct cgraph_node *n |
2242 | = cgraph_node::get (edge->callee->decl)->ultimate_alias_target (); | |
2243 | if (n != edge->callee && n->analyzed && !(n->count == old_count) | |
2244 | && n->count.ipa_p ()) | |
2245 | update_callee_keys (&edge_heap, n, NULL, updated_nodes); | |
670cd5c5 | 2246 | bitmap_clear (updated_nodes); |
ca31b95f | 2247 | |
4174a33a | 2248 | if (dump_enabled_p ()) |
50fe876d | 2249 | { |
f658ad30 | 2250 | ipa_fn_summary *s = ipa_fn_summaries->get (where); |
4174a33a DM |
2251 | |
2252 | /* dump_printf can't handle %+i. */ | |
2253 | char buf_net_change[100]; | |
2254 | snprintf (buf_net_change, sizeof buf_net_change, "%+i", | |
2255 | overall_size - old_size); | |
2256 | ||
2257 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, edge->call_stmt, | |
2258 | " Inlined %C into %C which now has time %f and " | |
b74d8dc4 | 2259 | "size %i, net change of %s%s.\n", |
4174a33a | 2260 | edge->callee, edge->caller, |
f658ad30 JH |
2261 | s->time.to_double (), |
2262 | ipa_size_summaries->get (edge->caller)->size, | |
b74d8dc4 JH |
2263 | buf_net_change, |
2264 | cross_module_call_p (edge) ? " (cross module)":""); | |
50fe876d | 2265 | } |
85057983 | 2266 | if (min_size > overall_size) |
b7c27d51 | 2267 | { |
85057983 | 2268 | min_size = overall_size; |
b7c27d51 JH |
2269 | |
2270 | if (dump_file) | |
85057983 | 2271 | fprintf (dump_file, "New minimal size reached: %i\n", min_size); |
b7c27d51 | 2272 | } |
ca31b95f | 2273 | } |
3e293154 | 2274 | |
632b4f8e | 2275 | free_growth_caches (); |
4174a33a DM |
2276 | if (dump_enabled_p ()) |
2277 | dump_printf (MSG_NOTE, | |
2278 | "Unit growth for small function inlining: %i->%i (%i%%)\n", | |
2279 | initial_size, overall_size, | |
2280 | initial_size ? overall_size * 100 / (initial_size) - 100: 0); | |
3dafb85c | 2281 | symtab->remove_edge_removal_hook (edge_removal_hook_holder); |
ca31b95f JH |
2282 | } |
2283 | ||
09a2806f JH |
2284 | /* Flatten NODE. Performed both during early inlining and |
2285 | at IPA inlining time. */ | |
af961c7f RG |
2286 | |
2287 | static void | |
f6e809c8 | 2288 | flatten_function (struct cgraph_node *node, bool early, bool update) |
af961c7f RG |
2289 | { |
2290 | struct cgraph_edge *e; | |
2291 | ||
2292 | /* We shouldn't be called recursively when we are being processed. */ | |
67348ccc | 2293 | gcc_assert (node->aux == NULL); |
af961c7f | 2294 | |
67348ccc | 2295 | node->aux = (void *) node; |
af961c7f RG |
2296 | |
2297 | for (e = node->callees; e; e = e->next_callee) | |
2298 | { | |
2299 | struct cgraph_node *orig_callee; | |
d52f5295 | 2300 | struct cgraph_node *callee = e->callee->ultimate_alias_target (); |
af961c7f | 2301 | |
af961c7f | 2302 | /* We've hit cycle? It is time to give up. */ |
67348ccc | 2303 | if (callee->aux) |
af961c7f | 2304 | { |
4174a33a DM |
2305 | if (dump_enabled_p ()) |
2306 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt, | |
2307 | "Not inlining %C into %C to avoid cycle.\n", | |
2308 | callee, e->caller); | |
a99670f9 JH |
2309 | if (cgraph_inline_failed_type (e->inline_failed) != CIF_FINAL_ERROR) |
2310 | e->inline_failed = CIF_RECURSIVE_INLINING; | |
af961c7f RG |
2311 | continue; |
2312 | } | |
2313 | ||
2314 | /* When the edge is already inlined, we just need to recurse into | |
2315 | it in order to fully flatten the leaves. */ | |
2316 | if (!e->inline_failed) | |
2317 | { | |
f6e809c8 | 2318 | flatten_function (callee, early, false); |
af961c7f RG |
2319 | continue; |
2320 | } | |
2321 | ||
4c0f7679 JH |
2322 | /* Flatten attribute needs to be processed during late inlining. For |
2323 | extra code quality we however do flattening during early optimization, | |
2324 | too. */ | |
632b4f8e | 2325 | if (!early |
4c0f7679 | 2326 | ? !can_inline_edge_p (e, true) |
9a4841a3 | 2327 | && !can_inline_edge_by_limits_p (e, true) |
4c0f7679 JH |
2328 | : !can_early_inline_edge_p (e)) |
2329 | continue; | |
2330 | ||
3dafb85c | 2331 | if (e->recursive_p ()) |
af961c7f | 2332 | { |
4174a33a DM |
2333 | if (dump_enabled_p ()) |
2334 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt, | |
2335 | "Not inlining: recursive call.\n"); | |
af961c7f RG |
2336 | continue; |
2337 | } | |
2338 | ||
67348ccc DM |
2339 | if (gimple_in_ssa_p (DECL_STRUCT_FUNCTION (node->decl)) |
2340 | != gimple_in_ssa_p (DECL_STRUCT_FUNCTION (callee->decl))) | |
59e0c6b7 | 2341 | { |
4174a33a DM |
2342 | if (dump_enabled_p ()) |
2343 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt, | |
2344 | "Not inlining: SSA form does not match.\n"); | |
59e0c6b7 RG |
2345 | continue; |
2346 | } | |
2347 | ||
af961c7f RG |
2348 | /* Inline the edge and flatten the inline clone. Avoid |
2349 | recursing through the original node if the node was cloned. */ | |
4174a33a DM |
2350 | if (dump_enabled_p ()) |
2351 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, e->call_stmt, | |
2352 | " Inlining %C into %C.\n", | |
2353 | callee, e->caller); | |
a5b1779f | 2354 | orig_callee = callee; |
c170d40f | 2355 | inline_call (e, true, NULL, NULL, false); |
af961c7f | 2356 | if (e->callee != orig_callee) |
67348ccc | 2357 | orig_callee->aux = (void *) node; |
f6e809c8 | 2358 | flatten_function (e->callee, early, false); |
af961c7f | 2359 | if (e->callee != orig_callee) |
67348ccc | 2360 | orig_callee->aux = NULL; |
af961c7f RG |
2361 | } |
2362 | ||
67348ccc | 2363 | node->aux = NULL; |
7237f93e JH |
2364 | cgraph_node *where = node->inlined_to ? node->inlined_to : node; |
2365 | if (update && opt_for_fn (where->decl, optimize)) | |
2366 | ipa_update_overall_fn_summary (where); | |
af961c7f RG |
2367 | } |
2368 | ||
a81b0a3d JH |
2369 | /* Inline NODE to all callers. Worker for cgraph_for_node_and_aliases. |
2370 | DATA points to number of calls originally found so we avoid infinite | |
2371 | recursion. */ | |
2372 | ||
2373 | static bool | |
bddead15 RB |
2374 | inline_to_all_callers_1 (struct cgraph_node *node, void *data, |
2375 | hash_set<cgraph_node *> *callers) | |
a81b0a3d JH |
2376 | { |
2377 | int *num_calls = (int *)data; | |
1bbb87c4 JH |
2378 | bool callee_removed = false; |
2379 | ||
a62bfab5 | 2380 | while (node->callers && !node->inlined_to) |
a81b0a3d JH |
2381 | { |
2382 | struct cgraph_node *caller = node->callers->caller; | |
2383 | ||
1af8bfe5 | 2384 | if (!can_inline_edge_p (node->callers, true) |
9a4841a3 | 2385 | || !can_inline_edge_by_limits_p (node->callers, true) |
1af8bfe5 JH |
2386 | || node->callers->recursive_p ()) |
2387 | { | |
2388 | if (dump_file) | |
2389 | fprintf (dump_file, "Uninlinable call found; giving up.\n"); | |
2390 | *num_calls = 0; | |
2391 | return false; | |
2392 | } | |
2393 | ||
a81b0a3d JH |
2394 | if (dump_file) |
2395 | { | |
5e626cd9 | 2396 | cgraph_node *ultimate = node->ultimate_alias_target (); |
a81b0a3d JH |
2397 | fprintf (dump_file, |
2398 | "\nInlining %s size %i.\n", | |
5e626cd9 | 2399 | ultimate->name (), |
f658ad30 | 2400 | ipa_size_summaries->get (ultimate)->size); |
a81b0a3d JH |
2401 | fprintf (dump_file, |
2402 | " Called once from %s %i insns.\n", | |
fec39fa6 | 2403 | node->callers->caller->name (), |
f658ad30 | 2404 | ipa_size_summaries->get (node->callers->caller)->size); |
a81b0a3d JH |
2405 | } |
2406 | ||
bddead15 RB |
2407 | /* Remember which callers we inlined to, delaying updating the |
2408 | overall summary. */ | |
2409 | callers->add (node->callers->caller); | |
2410 | inline_call (node->callers, true, NULL, NULL, false, &callee_removed); | |
a81b0a3d JH |
2411 | if (dump_file) |
2412 | fprintf (dump_file, | |
2413 | " Inlined into %s which now has %i size\n", | |
fec39fa6 | 2414 | caller->name (), |
f658ad30 | 2415 | ipa_size_summaries->get (caller)->size); |
a81b0a3d JH |
2416 | if (!(*num_calls)--) |
2417 | { | |
2418 | if (dump_file) | |
2419 | fprintf (dump_file, "New calls found; giving up.\n"); | |
1bbb87c4 | 2420 | return callee_removed; |
a81b0a3d | 2421 | } |
1bbb87c4 JH |
2422 | if (callee_removed) |
2423 | return true; | |
a81b0a3d JH |
2424 | } |
2425 | return false; | |
2426 | } | |
2427 | ||
bddead15 RB |
2428 | /* Wrapper around inline_to_all_callers_1 doing delayed overall summary |
2429 | update. */ | |
2430 | ||
2431 | static bool | |
2432 | inline_to_all_callers (struct cgraph_node *node, void *data) | |
2433 | { | |
2434 | hash_set<cgraph_node *> callers; | |
2435 | bool res = inline_to_all_callers_1 (node, data, &callers); | |
2436 | /* Perform the delayed update of the overall summary of all callers | |
2437 | processed. This avoids quadratic behavior in the cases where | |
2438 | we have a lot of calls to the same function. */ | |
2439 | for (hash_set<cgraph_node *>::iterator i = callers.begin (); | |
2440 | i != callers.end (); ++i) | |
7237f93e | 2441 | ipa_update_overall_fn_summary ((*i)->inlined_to ? (*i)->inlined_to : *i); |
bddead15 RB |
2442 | return res; |
2443 | } | |
2444 | ||
e86a910f JH |
2445 | /* Output overall time estimate. */ |
2446 | static void | |
2447 | dump_overall_stats (void) | |
2448 | { | |
ab38481c | 2449 | sreal sum_weighted = 0, sum = 0; |
e86a910f JH |
2450 | struct cgraph_node *node; |
2451 | ||
2452 | FOR_EACH_DEFINED_FUNCTION (node) | |
a62bfab5 | 2453 | if (!node->inlined_to |
e86a910f JH |
2454 | && !node->alias) |
2455 | { | |
56f62793 ML |
2456 | ipa_fn_summary *s = ipa_fn_summaries->get (node); |
2457 | if (s != NULL) | |
2458 | { | |
2459 | sum += s->time; | |
2460 | if (node->count.ipa ().initialized_p ()) | |
2461 | sum_weighted += s->time * node->count.ipa ().to_gcov_type (); | |
2462 | } | |
e86a910f JH |
2463 | } |
2464 | fprintf (dump_file, "Overall time estimate: " | |
ab38481c JH |
2465 | "%f weighted by profile: " |
2466 | "%f\n", sum.to_double (), sum_weighted.to_double ()); | |
e86a910f JH |
2467 | } |
2468 | ||
2469 | /* Output some useful stats about inlining. */ | |
2470 | ||
2471 | static void | |
2472 | dump_inline_stats (void) | |
2473 | { | |
a9243bfc RB |
2474 | int64_t inlined_cnt = 0, inlined_indir_cnt = 0; |
2475 | int64_t inlined_virt_cnt = 0, inlined_virt_indir_cnt = 0; | |
2476 | int64_t noninlined_cnt = 0, noninlined_indir_cnt = 0; | |
2477 | int64_t noninlined_virt_cnt = 0, noninlined_virt_indir_cnt = 0; | |
2478 | int64_t inlined_speculative = 0, inlined_speculative_ply = 0; | |
2479 | int64_t indirect_poly_cnt = 0, indirect_cnt = 0; | |
0009a6c3 JH |
2480 | int64_t reason[CIF_N_REASONS][2]; |
2481 | sreal reason_freq[CIF_N_REASONS]; | |
e86a910f JH |
2482 | int i; |
2483 | struct cgraph_node *node; | |
2484 | ||
2485 | memset (reason, 0, sizeof (reason)); | |
0009a6c3 JH |
2486 | for (i=0; i < CIF_N_REASONS; i++) |
2487 | reason_freq[i] = 0; | |
e86a910f JH |
2488 | FOR_EACH_DEFINED_FUNCTION (node) |
2489 | { | |
2490 | struct cgraph_edge *e; | |
2491 | for (e = node->callees; e; e = e->next_callee) | |
2492 | { | |
2493 | if (e->inline_failed) | |
2494 | { | |
1bad9c18 JH |
2495 | if (e->count.ipa ().initialized_p ()) |
2496 | reason[(int) e->inline_failed][0] += e->count.ipa ().to_gcov_type (); | |
0009a6c3 JH |
2497 | reason_freq[(int) e->inline_failed] += e->sreal_frequency (); |
2498 | reason[(int) e->inline_failed][1] ++; | |
3995f3a2 | 2499 | if (DECL_VIRTUAL_P (e->callee->decl) |
1bad9c18 | 2500 | && e->count.ipa ().initialized_p ()) |
e86a910f JH |
2501 | { |
2502 | if (e->indirect_inlining_edge) | |
1bad9c18 | 2503 | noninlined_virt_indir_cnt += e->count.ipa ().to_gcov_type (); |
e86a910f | 2504 | else |
1bad9c18 | 2505 | noninlined_virt_cnt += e->count.ipa ().to_gcov_type (); |
e86a910f | 2506 | } |
1bad9c18 | 2507 | else if (e->count.ipa ().initialized_p ()) |
e86a910f JH |
2508 | { |
2509 | if (e->indirect_inlining_edge) | |
1bad9c18 | 2510 | noninlined_indir_cnt += e->count.ipa ().to_gcov_type (); |
e86a910f | 2511 | else |
1bad9c18 | 2512 | noninlined_cnt += e->count.ipa ().to_gcov_type (); |
e86a910f JH |
2513 | } |
2514 | } | |
1bad9c18 | 2515 | else if (e->count.ipa ().initialized_p ()) |
e86a910f JH |
2516 | { |
2517 | if (e->speculative) | |
2518 | { | |
2519 | if (DECL_VIRTUAL_P (e->callee->decl)) | |
1bad9c18 | 2520 | inlined_speculative_ply += e->count.ipa ().to_gcov_type (); |
e86a910f | 2521 | else |
1bad9c18 | 2522 | inlined_speculative += e->count.ipa ().to_gcov_type (); |
e86a910f JH |
2523 | } |
2524 | else if (DECL_VIRTUAL_P (e->callee->decl)) | |
2525 | { | |
2526 | if (e->indirect_inlining_edge) | |
1bad9c18 | 2527 | inlined_virt_indir_cnt += e->count.ipa ().to_gcov_type (); |
e86a910f | 2528 | else |
1bad9c18 | 2529 | inlined_virt_cnt += e->count.ipa ().to_gcov_type (); |
e86a910f JH |
2530 | } |
2531 | else | |
2532 | { | |
2533 | if (e->indirect_inlining_edge) | |
1bad9c18 | 2534 | inlined_indir_cnt += e->count.ipa ().to_gcov_type (); |
e86a910f | 2535 | else |
1bad9c18 | 2536 | inlined_cnt += e->count.ipa ().to_gcov_type (); |
e86a910f JH |
2537 | } |
2538 | } | |
2539 | } | |
2540 | for (e = node->indirect_calls; e; e = e->next_callee) | |
3995f3a2 | 2541 | if (e->indirect_info->polymorphic |
1bad9c18 JH |
2542 | & e->count.ipa ().initialized_p ()) |
2543 | indirect_poly_cnt += e->count.ipa ().to_gcov_type (); | |
2544 | else if (e->count.ipa ().initialized_p ()) | |
2545 | indirect_cnt += e->count.ipa ().to_gcov_type (); | |
e86a910f | 2546 | } |
3995f3a2 | 2547 | if (max_count.initialized_p ()) |
e86a910f JH |
2548 | { |
2549 | fprintf (dump_file, | |
16998094 JM |
2550 | "Inlined %" PRId64 " + speculative " |
2551 | "%" PRId64 " + speculative polymorphic " | |
2552 | "%" PRId64 " + previously indirect " | |
2553 | "%" PRId64 " + virtual " | |
2554 | "%" PRId64 " + virtual and previously indirect " | |
2555 | "%" PRId64 "\n" "Not inlined " | |
2556 | "%" PRId64 " + previously indirect " | |
2557 | "%" PRId64 " + virtual " | |
2558 | "%" PRId64 " + virtual and previously indirect " | |
956d615d | 2559 | "%" PRId64 " + still indirect " |
16998094 JM |
2560 | "%" PRId64 " + still indirect polymorphic " |
2561 | "%" PRId64 "\n", inlined_cnt, | |
e86a910f JH |
2562 | inlined_speculative, inlined_speculative_ply, |
2563 | inlined_indir_cnt, inlined_virt_cnt, inlined_virt_indir_cnt, | |
2564 | noninlined_cnt, noninlined_indir_cnt, noninlined_virt_cnt, | |
2565 | noninlined_virt_indir_cnt, indirect_cnt, indirect_poly_cnt); | |
3995f3a2 JH |
2566 | fprintf (dump_file, "Removed speculations "); |
2567 | spec_rem.dump (dump_file); | |
2568 | fprintf (dump_file, "\n"); | |
e86a910f JH |
2569 | } |
2570 | dump_overall_stats (); | |
2571 | fprintf (dump_file, "\nWhy inlining failed?\n"); | |
2572 | for (i = 0; i < CIF_N_REASONS; i++) | |
0009a6c3 JH |
2573 | if (reason[i][1]) |
2574 | fprintf (dump_file, "%-50s: %8i calls, %8f freq, %" PRId64" count\n", | |
e86a910f | 2575 | cgraph_inline_failed_string ((cgraph_inline_failed_t) i), |
0009a6c3 | 2576 | (int) reason[i][1], reason_freq[i].to_double (), reason[i][0]); |
e86a910f JH |
2577 | } |
2578 | ||
d6ea70a0 JJ |
2579 | /* Called when node is removed. */ |
2580 | ||
2581 | static void | |
2582 | flatten_remove_node_hook (struct cgraph_node *node, void *data) | |
2583 | { | |
2584 | if (lookup_attribute ("flatten", DECL_ATTRIBUTES (node->decl)) == NULL) | |
2585 | return; | |
2586 | ||
2587 | hash_set<struct cgraph_node *> *removed | |
2588 | = (hash_set<struct cgraph_node *> *) data; | |
2589 | removed->add (node); | |
2590 | } | |
2591 | ||
ca31b95f JH |
2592 | /* Decide on the inlining. We do so in the topological order to avoid |
2593 | expenses on updating data structures. */ | |
2594 | ||
c2924966 | 2595 | static unsigned int |
4c0f7679 | 2596 | ipa_inline (void) |
ca31b95f JH |
2597 | { |
2598 | struct cgraph_node *node; | |
2599 | int nnodes; | |
b591a8b7 | 2600 | struct cgraph_node **order; |
d6ea70a0 | 2601 | int i, j; |
09ce3660 | 2602 | int cold; |
8a41354f JH |
2603 | bool remove_functions = false; |
2604 | ||
3dafb85c | 2605 | order = XCNEWVEC (struct cgraph_node *, symtab->cgraph_count); |
b591a8b7 | 2606 | |
10a5dd5d | 2607 | if (dump_file) |
0bceb671 | 2608 | ipa_dump_fn_summaries (dump_file); |
670cd5c5 | 2609 | |
af8bca3c | 2610 | nnodes = ipa_reverse_postorder (order); |
e7a74006 | 2611 | spec_rem = profile_count::zero (); |
ca31b95f | 2612 | |
65c70e6b | 2613 | FOR_EACH_FUNCTION (node) |
7ce7e4d4 JH |
2614 | { |
2615 | node->aux = 0; | |
2616 | ||
2617 | /* Recompute the default reasons for inlining because they may have | |
2618 | changed during merging. */ | |
2619 | if (in_lto_p) | |
2620 | { | |
2621 | for (cgraph_edge *e = node->callees; e; e = e->next_callee) | |
2622 | { | |
2623 | gcc_assert (e->inline_failed); | |
2624 | initialize_inline_failed (e); | |
2625 | } | |
2626 | for (cgraph_edge *e = node->indirect_calls; e; e = e->next_callee) | |
2627 | initialize_inline_failed (e); | |
2628 | } | |
2629 | } | |
ca31b95f JH |
2630 | |
2631 | if (dump_file) | |
af961c7f | 2632 | fprintf (dump_file, "\nFlattening functions:\n"); |
ca31b95f | 2633 | |
d6ea70a0 JJ |
2634 | /* First shrink order array, so that it only contains nodes with |
2635 | flatten attribute. */ | |
2636 | for (i = nnodes - 1, j = i; i >= 0; i--) | |
2637 | { | |
2638 | node = order[i]; | |
2895b172 JH |
2639 | if (node->definition |
2640 | && lookup_attribute ("flatten", | |
2641 | DECL_ATTRIBUTES (node->decl)) != NULL) | |
d6ea70a0 JJ |
2642 | order[j--] = order[i]; |
2643 | } | |
2644 | ||
2645 | /* After the above loop, order[j + 1] ... order[nnodes - 1] contain | |
2646 | nodes with flatten attribute. If there is more than one such | |
2647 | node, we need to register a node removal hook, as flatten_function | |
2648 | could remove other nodes with flatten attribute. See PR82801. */ | |
2649 | struct cgraph_node_hook_list *node_removal_hook_holder = NULL; | |
2650 | hash_set<struct cgraph_node *> *flatten_removed_nodes = NULL; | |
2651 | if (j < nnodes - 2) | |
2652 | { | |
2653 | flatten_removed_nodes = new hash_set<struct cgraph_node *>; | |
2654 | node_removal_hook_holder | |
2655 | = symtab->add_cgraph_removal_hook (&flatten_remove_node_hook, | |
2656 | flatten_removed_nodes); | |
2657 | } | |
2658 | ||
af961c7f RG |
2659 | /* In the first pass handle functions to be flattened. Do this with |
2660 | a priority so none of our later choices will make this impossible. */ | |
d6ea70a0 | 2661 | for (i = nnodes - 1; i > j; i--) |
ca31b95f | 2662 | { |
af961c7f | 2663 | node = order[i]; |
d6ea70a0 JJ |
2664 | if (flatten_removed_nodes |
2665 | && flatten_removed_nodes->contains (node)) | |
2666 | continue; | |
af961c7f | 2667 | |
09a2806f | 2668 | /* Handle nodes to be flattened. |
af961c7f RG |
2669 | Ideally when processing callees we stop inlining at the |
2670 | entry of cycles, possibly cloning that entry point and | |
2671 | try to flatten itself turning it into a self-recursive | |
2672 | function. */ | |
d6ea70a0 JJ |
2673 | if (dump_file) |
2674 | fprintf (dump_file, "Flattening %s\n", node->name ()); | |
f6e809c8 | 2675 | flatten_function (node, false, true); |
d6ea70a0 JJ |
2676 | } |
2677 | ||
2678 | if (j < nnodes - 2) | |
2679 | { | |
2680 | symtab->remove_cgraph_removal_hook (node_removal_hook_holder); | |
2681 | delete flatten_removed_nodes; | |
ca31b95f | 2682 | } |
d6ea70a0 JJ |
2683 | free (order); |
2684 | ||
e86a910f JH |
2685 | if (dump_file) |
2686 | dump_overall_stats (); | |
ca31b95f | 2687 | |
4c0f7679 | 2688 | inline_small_functions (); |
e70670cf | 2689 | |
17e0fc92 JH |
2690 | gcc_assert (symtab->state == IPA_SSA); |
2691 | symtab->state = IPA_SSA_AFTER_INLINING; | |
2692 | /* Do first after-inlining removal. We want to remove all "stale" extern | |
2693 | inline functions and virtual functions so we really know what is called | |
2694 | once. */ | |
2695 | symtab->remove_unreachable_nodes (dump_file); | |
ca31b95f | 2696 | |
100411f8 JH |
2697 | /* Inline functions with a property that after inlining into all callers the |
2698 | code size will shrink because the out-of-line copy is eliminated. | |
2699 | We do this regardless on the callee size as long as function growth limits | |
2700 | are met. */ | |
09ce3660 JH |
2701 | if (dump_file) |
2702 | fprintf (dump_file, | |
17e0fc92 JH |
2703 | "\nDeciding on functions to be inlined into all callers and " |
2704 | "removing useless speculations:\n"); | |
09ce3660 JH |
2705 | |
2706 | /* Inlining one function called once has good chance of preventing | |
2707 | inlining other function into the same callee. Ideally we should | |
2708 | work in priority order, but probably inlining hot functions first | |
2709 | is good cut without the extra pain of maintaining the queue. | |
2710 | ||
2711 | ??? this is not really fitting the bill perfectly: inlining function | |
2712 | into callee often leads to better optimization of callee due to | |
2713 | increased context for optimization. | |
2714 | For example if main() function calls a function that outputs help | |
956d615d | 2715 | and then function that does the main optimization, we should inline |
09ce3660 JH |
2716 | the second with priority even if both calls are cold by themselves. |
2717 | ||
2718 | We probably want to implement new predicate replacing our use of | |
2719 | maybe_hot_edge interpreted as maybe_hot_edge || callee is known | |
2720 | to be hot. */ | |
2721 | for (cold = 0; cold <= 1; cold ++) | |
355866de | 2722 | { |
09ce3660 | 2723 | FOR_EACH_DEFINED_FUNCTION (node) |
ca31b95f | 2724 | { |
09ce3660 JH |
2725 | struct cgraph_edge *edge, *next; |
2726 | bool update=false; | |
2727 | ||
29f1e2b1 JH |
2728 | if (!opt_for_fn (node->decl, optimize) |
2729 | || !opt_for_fn (node->decl, flag_inline_functions_called_once)) | |
2730 | continue; | |
2731 | ||
09ce3660 | 2732 | for (edge = node->callees; edge; edge = next) |
ca31b95f | 2733 | { |
09ce3660 JH |
2734 | next = edge->next_callee; |
2735 | if (edge->speculative && !speculation_useful_p (edge, false)) | |
e3c7b49c | 2736 | { |
1bad9c18 JH |
2737 | if (edge->count.ipa ().initialized_p ()) |
2738 | spec_rem += edge->count.ipa (); | |
3dafb85c | 2739 | edge->resolve_speculation (); |
09ce3660 | 2740 | update = true; |
8a41354f | 2741 | remove_functions = true; |
09ce3660 JH |
2742 | } |
2743 | } | |
2744 | if (update) | |
2745 | { | |
a62bfab5 ML |
2746 | struct cgraph_node *where = node->inlined_to |
2747 | ? node->inlined_to : node; | |
09ce3660 | 2748 | reset_edge_caches (where); |
0bceb671 | 2749 | ipa_update_overall_fn_summary (where); |
09ce3660 | 2750 | } |
2bf86c84 | 2751 | if (want_inline_function_to_all_callers_p (node, cold)) |
09ce3660 JH |
2752 | { |
2753 | int num_calls = 0; | |
1ede94c5 JH |
2754 | node->call_for_symbol_and_aliases (sum_callers, &num_calls, |
2755 | true); | |
2756 | while (node->call_for_symbol_and_aliases | |
17e0fc92 | 2757 | (inline_to_all_callers, &num_calls, true)) |
1bbb87c4 | 2758 | ; |
f019b607 | 2759 | remove_functions = true; |
ca31b95f JH |
2760 | } |
2761 | } | |
2762 | } | |
2763 | ||
f8e2a1ed | 2764 | /* Free ipa-prop structures if they are no longer needed. */ |
29f1e2b1 | 2765 | ipa_free_all_structures_after_iinln (); |
f8e2a1ed | 2766 | |
4174a33a DM |
2767 | if (dump_enabled_p ()) |
2768 | dump_printf (MSG_NOTE, | |
2769 | "\nInlined %i calls, eliminated %i functions\n\n", | |
2770 | ncalls_inlined, nfunctions_inlined); | |
ca31b95f | 2771 | if (dump_file) |
4174a33a | 2772 | dump_inline_stats (); |
09a2806f | 2773 | |
898b8927 | 2774 | if (dump_file) |
0bceb671 | 2775 | ipa_dump_fn_summaries (dump_file); |
8a41354f | 2776 | return remove_functions ? TODO_remove_functions : 0; |
ca31b95f JH |
2777 | } |
2778 | ||
275b4baa RG |
2779 | /* Inline always-inline function calls in NODE. */ |
2780 | ||
2781 | static bool | |
4c0f7679 | 2782 | inline_always_inline_functions (struct cgraph_node *node) |
275b4baa RG |
2783 | { |
2784 | struct cgraph_edge *e; | |
2785 | bool inlined = false; | |
2786 | ||
2787 | for (e = node->callees; e; e = e->next_callee) | |
2788 | { | |
d52f5295 | 2789 | struct cgraph_node *callee = e->callee->ultimate_alias_target (); |
67348ccc | 2790 | if (!DECL_DISREGARD_INLINE_LIMITS (callee->decl)) |
275b4baa RG |
2791 | continue; |
2792 | ||
3dafb85c | 2793 | if (e->recursive_p ()) |
275b4baa | 2794 | { |
4174a33a DM |
2795 | if (dump_enabled_p ()) |
2796 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt, | |
2797 | " Not inlining recursive call to %C.\n", | |
2798 | e->callee); | |
275b4baa RG |
2799 | e->inline_failed = CIF_RECURSIVE_INLINING; |
2800 | continue; | |
2801 | } | |
2802 | ||
4c0f7679 | 2803 | if (!can_early_inline_edge_p (e)) |
bef8491a ST |
2804 | { |
2805 | /* Set inlined to true if the callee is marked "always_inline" but | |
2806 | is not inlinable. This will allow flagging an error later in | |
2807 | expand_call_inline in tree-inline.c. */ | |
2808 | if (lookup_attribute ("always_inline", | |
67348ccc | 2809 | DECL_ATTRIBUTES (callee->decl)) != NULL) |
bef8491a ST |
2810 | inlined = true; |
2811 | continue; | |
2812 | } | |
275b4baa | 2813 | |
4174a33a DM |
2814 | if (dump_enabled_p ()) |
2815 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, e->call_stmt, | |
2816 | " Inlining %C into %C (always_inline).\n", | |
2817 | e->callee, e->caller); | |
c170d40f | 2818 | inline_call (e, true, NULL, NULL, false); |
275b4baa RG |
2819 | inlined = true; |
2820 | } | |
c170d40f | 2821 | if (inlined) |
0bceb671 | 2822 | ipa_update_overall_fn_summary (node); |
275b4baa RG |
2823 | |
2824 | return inlined; | |
2825 | } | |
2826 | ||
ca31b95f | 2827 | /* Decide on the inlining. We do so in the topological order to avoid |
af961c7f | 2828 | expenses on updating data structures. */ |
ca31b95f | 2829 | |
7fa49e7b | 2830 | static bool |
4c0f7679 | 2831 | early_inline_small_functions (struct cgraph_node *node) |
ca31b95f JH |
2832 | { |
2833 | struct cgraph_edge *e; | |
d63db217 | 2834 | bool inlined = false; |
7fa49e7b | 2835 | |
275b4baa | 2836 | for (e = node->callees; e; e = e->next_callee) |
c3056c2d | 2837 | { |
d52f5295 | 2838 | struct cgraph_node *callee = e->callee->ultimate_alias_target (); |
56f62793 | 2839 | |
956d615d | 2840 | /* We can encounter not-yet-analyzed function during |
56f62793 ML |
2841 | early inlining on callgraphs with strongly |
2842 | connected components. */ | |
2843 | ipa_fn_summary *s = ipa_fn_summaries->get (callee); | |
2844 | if (s == NULL || !s->inlinable || !e->inline_failed) | |
275b4baa RG |
2845 | continue; |
2846 | ||
2847 | /* Do not consider functions not declared inline. */ | |
67348ccc | 2848 | if (!DECL_DECLARED_INLINE_P (callee->decl) |
2bf86c84 JH |
2849 | && !opt_for_fn (node->decl, flag_inline_small_functions) |
2850 | && !opt_for_fn (node->decl, flag_inline_functions)) | |
275b4baa RG |
2851 | continue; |
2852 | ||
4174a33a DM |
2853 | if (dump_enabled_p ()) |
2854 | dump_printf_loc (MSG_NOTE, e->call_stmt, | |
2855 | "Considering inline candidate %C.\n", | |
2856 | callee); | |
ca31b95f | 2857 | |
4c0f7679 JH |
2858 | if (!can_early_inline_edge_p (e)) |
2859 | continue; | |
2860 | ||
3dafb85c | 2861 | if (e->recursive_p ()) |
275b4baa | 2862 | { |
4174a33a DM |
2863 | if (dump_enabled_p ()) |
2864 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, e->call_stmt, | |
2865 | " Not inlining: recursive call.\n"); | |
22ad64b6 | 2866 | continue; |
275b4baa | 2867 | } |
af961c7f | 2868 | |
4c0f7679 | 2869 | if (!want_early_inline_function_p (e)) |
275b4baa | 2870 | continue; |
ca31b95f | 2871 | |
4174a33a DM |
2872 | if (dump_enabled_p ()) |
2873 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, e->call_stmt, | |
2874 | " Inlining %C into %C.\n", | |
2875 | callee, e->caller); | |
bddead15 | 2876 | inline_call (e, true, NULL, NULL, false); |
4c0f7679 | 2877 | inlined = true; |
38bc76da | 2878 | } |
275b4baa | 2879 | |
bddead15 | 2880 | if (inlined) |
0bceb671 | 2881 | ipa_update_overall_fn_summary (node); |
bddead15 | 2882 | |
7fa49e7b | 2883 | return inlined; |
ca31b95f JH |
2884 | } |
2885 | ||
be55bfe6 | 2886 | unsigned int |
be3c16c4 | 2887 | early_inliner (function *fun) |
d63db217 | 2888 | { |
d52f5295 | 2889 | struct cgraph_node *node = cgraph_node::get (current_function_decl); |
10a5dd5d | 2890 | struct cgraph_edge *edge; |
7fa49e7b | 2891 | unsigned int todo = 0; |
796bda22 | 2892 | int iterations = 0; |
275b4baa | 2893 | bool inlined = false; |
d63db217 | 2894 | |
1da2ed5f | 2895 | if (seen_error ()) |
c2924966 | 2896 | return 0; |
af961c7f | 2897 | |
ecb62563 JH |
2898 | /* Do nothing if datastructures for ipa-inliner are already computed. This |
2899 | happens when some pass decides to construct new function and | |
2900 | cgraph_add_new_function calls lowering passes and early optimization on | |
2901 | it. This may confuse ourself when early inliner decide to inline call to | |
2902 | function clone, because function clones don't have parameter list in | |
2903 | ipa-prop matching their signature. */ | |
dd912cb8 | 2904 | if (ipa_node_params_sum) |
ecb62563 JH |
2905 | return 0; |
2906 | ||
b2b29377 MM |
2907 | if (flag_checking) |
2908 | node->verify (); | |
d122681a | 2909 | node->remove_all_references (); |
275b4baa RG |
2910 | |
2911 | /* Even when not optimizing or not inlining inline always-inline | |
2912 | functions. */ | |
4c0f7679 | 2913 | inlined = inline_always_inline_functions (node); |
275b4baa | 2914 | |
af961c7f RG |
2915 | if (!optimize |
2916 | || flag_no_inline | |
4c0f7679 JH |
2917 | || !flag_early_inlining |
2918 | /* Never inline regular functions into always-inline functions | |
2919 | during incremental inlining. This sucks as functions calling | |
2920 | always inline functions will get less optimized, but at the | |
2921 | same time inlining of functions calling always inline | |
09a2806f | 2922 | function into an always inline function might introduce |
4c0f7679 JH |
2923 | cycles of edges to be always inlined in the callgraph. |
2924 | ||
2925 | We might want to be smarter and just avoid this type of inlining. */ | |
d67bce7c JH |
2926 | || (DECL_DISREGARD_INLINE_LIMITS (node->decl) |
2927 | && lookup_attribute ("always_inline", | |
2928 | DECL_ATTRIBUTES (node->decl)))) | |
275b4baa RG |
2929 | ; |
2930 | else if (lookup_attribute ("flatten", | |
67348ccc | 2931 | DECL_ATTRIBUTES (node->decl)) != NULL) |
7fa49e7b | 2932 | { |
275b4baa RG |
2933 | /* When the function is marked to be flattened, recursively inline |
2934 | all calls in it. */ | |
4174a33a DM |
2935 | if (dump_enabled_p ()) |
2936 | dump_printf (MSG_OPTIMIZED_LOCATIONS, | |
2937 | "Flattening %C\n", node); | |
f6e809c8 | 2938 | flatten_function (node, true, true); |
275b4baa | 2939 | inlined = true; |
7fa49e7b | 2940 | } |
af961c7f RG |
2941 | else |
2942 | { | |
d67bce7c JH |
2943 | /* If some always_inline functions was inlined, apply the changes. |
2944 | This way we will not account always inline into growth limits and | |
2945 | moreover we will inline calls from always inlines that we skipped | |
56aae4b7 | 2946 | previously because of conditional above. */ |
d67bce7c JH |
2947 | if (inlined) |
2948 | { | |
2949 | timevar_push (TV_INTEGRATION); | |
2950 | todo |= optimize_inline_calls (current_function_decl); | |
1cf06f1e MP |
2951 | /* optimize_inline_calls call above might have introduced new |
2952 | statements that don't have inline parameters computed. */ | |
2953 | for (edge = node->callees; edge; edge = edge->next_callee) | |
2954 | { | |
56f62793 ML |
2955 | /* We can enounter not-yet-analyzed function during |
2956 | early inlining on callgraphs with strongly | |
2957 | connected components. */ | |
99353fcf | 2958 | ipa_call_summary *es = ipa_call_summaries->get_create (edge); |
263e19c7 JH |
2959 | es->call_stmt_size |
2960 | = estimate_num_insns (edge->call_stmt, &eni_size_weights); | |
2961 | es->call_stmt_time | |
2962 | = estimate_num_insns (edge->call_stmt, &eni_time_weights); | |
1cf06f1e | 2963 | } |
0bceb671 | 2964 | ipa_update_overall_fn_summary (node); |
d67bce7c JH |
2965 | inlined = false; |
2966 | timevar_pop (TV_INTEGRATION); | |
2967 | } | |
af961c7f RG |
2968 | /* We iterate incremental inlining to get trivial cases of indirect |
2969 | inlining. */ | |
028d4092 | 2970 | while (iterations < param_early_inliner_max_iterations |
4c0f7679 | 2971 | && early_inline_small_functions (node)) |
af961c7f RG |
2972 | { |
2973 | timevar_push (TV_INTEGRATION); | |
2974 | todo |= optimize_inline_calls (current_function_decl); | |
4c0f7679 JH |
2975 | |
2976 | /* Technically we ought to recompute inline parameters so the new | |
2977 | iteration of early inliner works as expected. We however have | |
2978 | values approximately right and thus we only need to update edge | |
2979 | info that might be cleared out for newly discovered edges. */ | |
2980 | for (edge = node->callees; edge; edge = edge->next_callee) | |
2981 | { | |
d5e254e1 | 2982 | /* We have no summary for new bound store calls yet. */ |
b412559e ML |
2983 | ipa_call_summary *es = ipa_call_summaries->get_create (edge); |
2984 | es->call_stmt_size | |
2985 | = estimate_num_insns (edge->call_stmt, &eni_size_weights); | |
2986 | es->call_stmt_time | |
2987 | = estimate_num_insns (edge->call_stmt, &eni_time_weights); | |
4c0f7679 | 2988 | } |
028d4092 | 2989 | if (iterations < param_early_inliner_max_iterations - 1) |
0bceb671 | 2990 | ipa_update_overall_fn_summary (node); |
af961c7f | 2991 | timevar_pop (TV_INTEGRATION); |
275b4baa RG |
2992 | iterations++; |
2993 | inlined = false; | |
af961c7f RG |
2994 | } |
2995 | if (dump_file) | |
2996 | fprintf (dump_file, "Iterations: %i\n", iterations); | |
2997 | } | |
2998 | ||
275b4baa RG |
2999 | if (inlined) |
3000 | { | |
3001 | timevar_push (TV_INTEGRATION); | |
3002 | todo |= optimize_inline_calls (current_function_decl); | |
3003 | timevar_pop (TV_INTEGRATION); | |
3004 | } | |
3005 | ||
be55bfe6 | 3006 | fun->always_inline_functions_inlined = true; |
d63db217 | 3007 | |
af961c7f | 3008 | return todo; |
d63db217 JH |
3009 | } |
3010 | ||
be3c16c4 DC |
3011 | /* Do inlining of small functions. Doing so early helps profiling and other |
3012 | passes to be somewhat more effective and avoids some code duplication in | |
3013 | later real inlining pass for testcases with very many function calls. */ | |
3014 | ||
3015 | namespace { | |
3016 | ||
3017 | const pass_data pass_data_early_inline = | |
3018 | { | |
3019 | GIMPLE_PASS, /* type */ | |
3020 | "einline", /* name */ | |
3021 | OPTGROUP_INLINE, /* optinfo_flags */ | |
3022 | TV_EARLY_INLINING, /* tv_id */ | |
3023 | PROP_ssa, /* properties_required */ | |
3024 | 0, /* properties_provided */ | |
3025 | 0, /* properties_destroyed */ | |
3026 | 0, /* todo_flags_start */ | |
3027 | 0, /* todo_flags_finish */ | |
3028 | }; | |
3029 | ||
3030 | class pass_early_inline : public gimple_opt_pass | |
3031 | { | |
3032 | public: | |
3033 | pass_early_inline (gcc::context *ctxt) | |
3034 | : gimple_opt_pass (pass_data_early_inline, ctxt) | |
3035 | {} | |
3036 | ||
3037 | /* opt_pass methods: */ | |
3038 | virtual unsigned int execute (function *); | |
3039 | ||
3040 | }; // class pass_early_inline | |
3041 | ||
3042 | unsigned int | |
3043 | pass_early_inline::execute (function *fun) | |
3044 | { | |
3045 | return early_inliner (fun); | |
3046 | } | |
3047 | ||
27a4cd48 DM |
3048 | } // anon namespace |
3049 | ||
3050 | gimple_opt_pass * | |
3051 | make_pass_early_inline (gcc::context *ctxt) | |
3052 | { | |
3053 | return new pass_early_inline (ctxt); | |
3054 | } | |
3055 | ||
27a4cd48 DM |
3056 | namespace { |
3057 | ||
3058 | const pass_data pass_data_ipa_inline = | |
873aa8f5 | 3059 | { |
27a4cd48 DM |
3060 | IPA_PASS, /* type */ |
3061 | "inline", /* name */ | |
3062 | OPTGROUP_INLINE, /* optinfo_flags */ | |
27a4cd48 DM |
3063 | TV_IPA_INLINING, /* tv_id */ |
3064 | 0, /* properties_required */ | |
3065 | 0, /* properties_provided */ | |
3066 | 0, /* properties_destroyed */ | |
8605403e | 3067 | 0, /* todo_flags_start */ |
8a41354f | 3068 | ( TODO_dump_symtab ), /* todo_flags_finish */ |
ca31b95f | 3069 | }; |
27a4cd48 DM |
3070 | |
3071 | class pass_ipa_inline : public ipa_opt_pass_d | |
3072 | { | |
3073 | public: | |
c3284718 RS |
3074 | pass_ipa_inline (gcc::context *ctxt) |
3075 | : ipa_opt_pass_d (pass_data_ipa_inline, ctxt, | |
d2db2e6b JH |
3076 | NULL, /* generate_summary */ |
3077 | NULL, /* write_summary */ | |
3078 | NULL, /* read_summary */ | |
c3284718 RS |
3079 | NULL, /* write_optimization_summary */ |
3080 | NULL, /* read_optimization_summary */ | |
3081 | NULL, /* stmt_fixup */ | |
3082 | 0, /* function_transform_todo_flags_start */ | |
3083 | inline_transform, /* function_transform */ | |
3084 | NULL) /* variable_transform */ | |
27a4cd48 DM |
3085 | {} |
3086 | ||
3087 | /* opt_pass methods: */ | |
be55bfe6 | 3088 | virtual unsigned int execute (function *) { return ipa_inline (); } |
27a4cd48 DM |
3089 | |
3090 | }; // class pass_ipa_inline | |
3091 | ||
3092 | } // anon namespace | |
3093 | ||
3094 | ipa_opt_pass_d * | |
3095 | make_pass_ipa_inline (gcc::context *ctxt) | |
3096 | { | |
3097 | return new pass_ipa_inline (ctxt); | |
3098 | } |