Mercurial > hg > CbC > CbC_gcc
annotate gcc/ipa-inline.c @ 120:f93fa5091070
fix conv1.c
author | mir3636 |
---|---|
date | Thu, 08 Mar 2018 14:53:42 +0900 |
parents | 04ced10e8804 |
children | fb3d53c41846 84e7813d76e9 |
rev | line source |
---|---|
0 | 1 /* Inlining decision heuristics. |
111 | 2 Copyright (C) 2003-2017 Free Software Foundation, Inc. |
0 | 3 Contributed by Jan Hubicka |
4 | |
5 This file is part of GCC. | |
6 | |
7 GCC is free software; you can redistribute it and/or modify it under | |
8 the terms of the GNU General Public License as published by the Free | |
9 Software Foundation; either version 3, or (at your option) any later | |
10 version. | |
11 | |
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
15 for more details. | |
16 | |
17 You should have received a copy of the GNU General Public License | |
18 along with GCC; see the file COPYING3. If not see | |
19 <http://www.gnu.org/licenses/>. */ | |
20 | |
21 /* Inlining decision heuristics | |
22 | |
111 | 23 The implementation of inliner is organized as follows: |
0 | 24 |
25 inlining heuristics limits | |
26 | |
111 | 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. | |
0 | 35 |
36 inlining heuristics | |
37 | |
111 | 38 The inliner itself is split into two passes: |
39 | |
40 pass_early_inlining | |
0 | 41 |
111 | 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. | |
0 | 45 |
111 | 46 The strength of the pass is that it is run in topological order |
47 (reverse postorder) on the callgraph. Functions are converted into SSA | |
48 form just before this pass and optimized subsequently. As a result, the | |
49 callees of the function seen by the early inliner was already optimized | |
50 and results of early inlining adds a lot of optimization opportunities | |
51 for the local optimization. | |
0 | 52 |
111 | 53 The pass handle the obvious inlining decisions within the compilation |
54 unit - inlining auto inline functions, inlining for size and | |
55 flattening. | |
0 | 56 |
111 | 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. | |
0 | 61 |
111 | 62 Because of lack of whole unit knowledge, the pass can not really make |
63 good code size/performance tradeoffs. It however does very simple | |
64 speculative inlining allowing code size to grow by | |
65 EARLY_INLINING_INSNS when callee is leaf function. In this case the | |
66 optimizations performed later are very likely to eliminate the cost. | |
0 | 67 |
111 | 68 pass_ipa_inline |
0 | 69 |
111 | 70 This is the real inliner able to handle inlining with whole program |
71 knowledge. It performs following steps: | |
72 | |
73 1) inlining of small functions. This is implemented by greedy | |
74 algorithm ordering all inlinable cgraph edges by their badness and | |
75 inlining them in this order as long as inline limits allows doing so. | |
0 | 76 |
111 | 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. | |
0 | 81 |
111 | 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. | |
0 | 87 |
111 | 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. */ | |
0 | 91 |
92 #include "config.h" | |
93 #include "system.h" | |
94 #include "coretypes.h" | |
111 | 95 #include "backend.h" |
96 #include "target.h" | |
97 #include "rtl.h" | |
0 | 98 #include "tree.h" |
111 | 99 #include "gimple.h" |
100 #include "alloc-pool.h" | |
101 #include "tree-pass.h" | |
102 #include "gimple-ssa.h" | |
103 #include "cgraph.h" | |
104 #include "lto-streamer.h" | |
105 #include "trans-mem.h" | |
106 #include "calls.h" | |
0 | 107 #include "tree-inline.h" |
108 #include "params.h" | |
111 | 109 #include "profile.h" |
110 #include "symbol-summary.h" | |
111 #include "tree-vrp.h" | |
0 | 112 #include "ipa-prop.h" |
111 | 113 #include "ipa-fnsummary.h" |
114 #include "ipa-inline.h" | |
115 #include "ipa-utils.h" | |
116 #include "sreal.h" | |
117 #include "auto-profile.h" | |
118 #include "builtins.h" | |
119 #include "fibonacci_heap.h" | |
120 #include "stringpool.h" | |
121 #include "attribs.h" | |
122 #include "asan.h" | |
0 | 123 |
111 | 124 typedef fibonacci_heap <sreal, cgraph_edge> edge_heap_t; |
125 typedef fibonacci_node <sreal, cgraph_edge> edge_heap_node_t; | |
0 | 126 |
127 /* Statistics we collect about inlining algorithm. */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
128 static int overall_size; |
111 | 129 static profile_count max_count; |
130 static profile_count spec_rem; | |
0 | 131 |
111 | 132 /* Pre-computed constants 1/CGRAPH_FREQ_BASE and 1/100. */ |
133 static sreal cgraph_freq_base_rec, percent_rec; | |
134 | |
135 /* Return false when inlining edge E would lead to violating | |
136 limits on function unit growth or stack usage growth. | |
0 | 137 |
111 | 138 The relative function body growth limit is present generally |
139 to avoid problems with non-linear behavior of the compiler. | |
140 To allow inlining huge functions into tiny wrapper, the limit | |
141 is always based on the bigger of the two functions considered. | |
0 | 142 |
111 | 143 For stack growth limits we always base the growth in stack usage |
144 of the callers. We want to prevent applications from segfaulting | |
145 on stack overflow when functions with huge stack frames gets | |
146 inlined. */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
147 |
111 | 148 static bool |
149 caller_growth_limits (struct cgraph_edge *e) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
150 { |
111 | 151 struct cgraph_node *to = e->caller; |
152 struct cgraph_node *what = e->callee->ultimate_alias_target (); | |
153 int newsize; | |
154 int limit = 0; | |
155 HOST_WIDE_INT stack_size_limit = 0, inlined_stack; | |
156 ipa_fn_summary *info, *what_info, *outer_info = ipa_fn_summaries->get (to); | |
0 | 157 |
111 | 158 /* Look for function e->caller is inlined to. While doing |
159 so work out the largest function body on the way. As | |
160 described above, we want to base our function growth | |
161 limits based on that. Not on the self size of the | |
162 outer function, not on the self size of inline code | |
163 we immediately inline to. This is the most relaxed | |
164 interpretation of the rule "do not grow large functions | |
165 too much in order to prevent compiler from exploding". */ | |
166 while (true) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
167 { |
111 | 168 info = ipa_fn_summaries->get (to); |
169 if (limit < info->self_size) | |
170 limit = info->self_size; | |
171 if (stack_size_limit < info->estimated_self_stack_size) | |
172 stack_size_limit = info->estimated_self_stack_size; | |
173 if (to->global.inlined_to) | |
174 to = to->callers->caller; | |
0 | 175 else |
111 | 176 break; |
0 | 177 } |
178 | |
111 | 179 what_info = ipa_fn_summaries->get (what); |
0 | 180 |
111 | 181 if (limit < what_info->self_size) |
182 limit = what_info->self_size; | |
0 | 183 |
184 limit += limit * PARAM_VALUE (PARAM_LARGE_FUNCTION_GROWTH) / 100; | |
185 | |
186 /* Check the size after inlining against the function limits. But allow | |
187 the function to shrink if it went over the limits by forced inlining. */ | |
111 | 188 newsize = estimate_size_after_inlining (to, e); |
189 if (newsize >= info->size | |
0 | 190 && newsize > PARAM_VALUE (PARAM_LARGE_FUNCTION_INSNS) |
191 && newsize > limit) | |
192 { | |
111 | 193 e->inline_failed = CIF_LARGE_FUNCTION_GROWTH_LIMIT; |
0 | 194 return false; |
195 } | |
196 | |
111 | 197 if (!what_info->estimated_stack_size) |
198 return true; | |
0 | 199 |
111 | 200 /* FIXME: Stack size limit often prevents inlining in Fortran programs |
201 due to large i/o datastructures used by the Fortran front-end. | |
202 We ought to ignore this limit when we know that the edge is executed | |
203 on every invocation of the caller (i.e. its call statement dominates | |
204 exit block). We do not track this information, yet. */ | |
205 stack_size_limit += ((gcov_type)stack_size_limit | |
206 * PARAM_VALUE (PARAM_STACK_FRAME_GROWTH) / 100); | |
0 | 207 |
111 | 208 inlined_stack = (outer_info->stack_frame_offset |
209 + outer_info->estimated_self_stack_size | |
210 + what_info->estimated_stack_size); | |
211 /* Check new stack consumption with stack consumption at the place | |
212 stack is used. */ | |
213 if (inlined_stack > stack_size_limit | |
214 /* If function already has large stack usage from sibling | |
215 inline call, we can inline, too. | |
216 This bit overoptimistically assume that we are good at stack | |
217 packing. */ | |
218 && inlined_stack > info->estimated_stack_size | |
0 | 219 && inlined_stack > PARAM_VALUE (PARAM_LARGE_STACK_FRAME)) |
220 { | |
111 | 221 e->inline_failed = CIF_LARGE_STACK_FRAME_GROWTH_LIMIT; |
0 | 222 return false; |
223 } | |
224 return true; | |
225 } | |
226 | |
111 | 227 /* Dump info about why inlining has failed. */ |
228 | |
229 static void | |
230 report_inline_failed_reason (struct cgraph_edge *e) | |
231 { | |
232 if (dump_file) | |
233 { | |
234 fprintf (dump_file, " not inlinable: %s -> %s, %s\n", | |
235 e->caller->dump_name (), | |
236 e->callee->dump_name (), | |
237 cgraph_inline_failed_string (e->inline_failed)); | |
238 if ((e->inline_failed == CIF_TARGET_OPTION_MISMATCH | |
239 || e->inline_failed == CIF_OPTIMIZATION_MISMATCH) | |
240 && e->caller->lto_file_data | |
241 && e->callee->ultimate_alias_target ()->lto_file_data) | |
242 { | |
243 fprintf (dump_file, " LTO objects: %s, %s\n", | |
244 e->caller->lto_file_data->file_name, | |
245 e->callee->ultimate_alias_target ()->lto_file_data->file_name); | |
246 } | |
247 if (e->inline_failed == CIF_TARGET_OPTION_MISMATCH) | |
248 cl_target_option_print_diff | |
249 (dump_file, 2, target_opts_for_fn (e->caller->decl), | |
250 target_opts_for_fn (e->callee->ultimate_alias_target ()->decl)); | |
251 if (e->inline_failed == CIF_OPTIMIZATION_MISMATCH) | |
252 cl_optimization_print_diff | |
253 (dump_file, 2, opts_for_fn (e->caller->decl), | |
254 opts_for_fn (e->callee->ultimate_alias_target ()->decl)); | |
255 } | |
256 } | |
257 | |
258 /* Decide whether sanitizer-related attributes allow inlining. */ | |
0 | 259 |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
260 static bool |
111 | 261 sanitize_attrs_match_for_inline_p (const_tree caller, const_tree callee) |
0 | 262 { |
111 | 263 if (!caller || !callee) |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
264 return true; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
265 |
111 | 266 return sanitize_flags_p (SANITIZE_ADDRESS, caller) |
267 == sanitize_flags_p (SANITIZE_ADDRESS, callee); | |
268 } | |
269 | |
270 /* Used for flags where it is safe to inline when caller's value is | |
271 grater than callee's. */ | |
272 #define check_maybe_up(flag) \ | |
273 (opts_for_fn (caller->decl)->x_##flag \ | |
274 != opts_for_fn (callee->decl)->x_##flag \ | |
275 && (!always_inline \ | |
276 || opts_for_fn (caller->decl)->x_##flag \ | |
277 < opts_for_fn (callee->decl)->x_##flag)) | |
278 /* Used for flags where it is safe to inline when caller's value is | |
279 smaller than callee's. */ | |
280 #define check_maybe_down(flag) \ | |
281 (opts_for_fn (caller->decl)->x_##flag \ | |
282 != opts_for_fn (callee->decl)->x_##flag \ | |
283 && (!always_inline \ | |
284 || opts_for_fn (caller->decl)->x_##flag \ | |
285 > opts_for_fn (callee->decl)->x_##flag)) | |
286 /* Used for flags where exact match is needed for correctness. */ | |
287 #define check_match(flag) \ | |
288 (opts_for_fn (caller->decl)->x_##flag \ | |
289 != opts_for_fn (callee->decl)->x_##flag) | |
290 | |
291 /* Decide if we can inline the edge and possibly update | |
292 inline_failed reason. | |
293 We check whether inlining is possible at all and whether | |
294 caller growth limits allow doing so. | |
295 | |
296 if REPORT is true, output reason to the dump file. | |
297 | |
298 if DISREGARD_LIMITS is true, ignore size limits.*/ | |
299 | |
300 static bool | |
301 can_inline_edge_p (struct cgraph_edge *e, bool report, | |
302 bool disregard_limits = false, bool early = false) | |
303 { | |
304 gcc_checking_assert (e->inline_failed); | |
305 | |
306 if (cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR) | |
0 | 307 { |
111 | 308 if (report) |
309 report_inline_failed_reason (e); | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
310 return false; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
311 } |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
312 |
111 | 313 bool inlinable = true; |
314 enum availability avail; | |
315 cgraph_node *caller = e->caller->global.inlined_to | |
316 ? e->caller->global.inlined_to : e->caller; | |
317 cgraph_node *callee = e->callee->ultimate_alias_target (&avail, caller); | |
318 tree caller_tree = DECL_FUNCTION_SPECIFIC_OPTIMIZATION (caller->decl); | |
319 tree callee_tree | |
320 = callee ? DECL_FUNCTION_SPECIFIC_OPTIMIZATION (callee->decl) : NULL; | |
0 | 321 |
111 | 322 if (!callee->definition) |
323 { | |
324 e->inline_failed = CIF_BODY_NOT_AVAILABLE; | |
325 inlinable = false; | |
326 } | |
327 if (!early && !opt_for_fn (callee->decl, optimize)) | |
328 { | |
329 e->inline_failed = CIF_FUNCTION_NOT_OPTIMIZED; | |
330 inlinable = false; | |
331 } | |
332 else if (callee->calls_comdat_local) | |
333 { | |
334 e->inline_failed = CIF_USES_COMDAT_LOCAL; | |
335 inlinable = false; | |
336 } | |
337 else if (avail <= AVAIL_INTERPOSABLE) | |
338 { | |
339 e->inline_failed = CIF_OVERWRITABLE; | |
340 inlinable = false; | |
341 } | |
342 /* All edges with call_stmt_cannot_inline_p should have inline_failed | |
343 initialized to one of FINAL_ERROR reasons. */ | |
344 else if (e->call_stmt_cannot_inline_p) | |
345 gcc_unreachable (); | |
346 /* Don't inline if the functions have different EH personalities. */ | |
347 else if (DECL_FUNCTION_PERSONALITY (caller->decl) | |
348 && DECL_FUNCTION_PERSONALITY (callee->decl) | |
349 && (DECL_FUNCTION_PERSONALITY (caller->decl) | |
350 != DECL_FUNCTION_PERSONALITY (callee->decl))) | |
351 { | |
352 e->inline_failed = CIF_EH_PERSONALITY; | |
353 inlinable = false; | |
354 } | |
355 /* TM pure functions should not be inlined into non-TM_pure | |
356 functions. */ | |
357 else if (is_tm_pure (callee->decl) && !is_tm_pure (caller->decl)) | |
358 { | |
359 e->inline_failed = CIF_UNSPECIFIED; | |
360 inlinable = false; | |
361 } | |
362 /* Check compatibility of target optimization options. */ | |
363 else if (!targetm.target_option.can_inline_p (caller->decl, | |
364 callee->decl)) | |
365 { | |
366 e->inline_failed = CIF_TARGET_OPTION_MISMATCH; | |
367 inlinable = false; | |
368 } | |
369 else if (!ipa_fn_summaries->get (callee)->inlinable) | |
370 { | |
371 e->inline_failed = CIF_FUNCTION_NOT_INLINABLE; | |
372 inlinable = false; | |
373 } | |
374 /* Don't inline a function with mismatched sanitization attributes. */ | |
375 else if (!sanitize_attrs_match_for_inline_p (caller->decl, callee->decl)) | |
0 | 376 { |
111 | 377 e->inline_failed = CIF_ATTRIBUTE_MISMATCH; |
378 inlinable = false; | |
379 } | |
380 /* Check if caller growth allows the inlining. */ | |
381 else if (!DECL_DISREGARD_INLINE_LIMITS (callee->decl) | |
382 && !disregard_limits | |
383 && !lookup_attribute ("flatten", | |
384 DECL_ATTRIBUTES (caller->decl)) | |
385 && !caller_growth_limits (e)) | |
386 inlinable = false; | |
387 /* Don't inline a function with a higher optimization level than the | |
388 caller. FIXME: this is really just tip of iceberg of handling | |
389 optimization attribute. */ | |
390 else if (caller_tree != callee_tree) | |
391 { | |
392 bool always_inline = | |
393 (DECL_DISREGARD_INLINE_LIMITS (callee->decl) | |
394 && lookup_attribute ("always_inline", | |
395 DECL_ATTRIBUTES (callee->decl))); | |
396 ipa_fn_summary *caller_info = ipa_fn_summaries->get (caller); | |
397 ipa_fn_summary *callee_info = ipa_fn_summaries->get (callee); | |
398 | |
399 /* Until GCC 4.9 we did not check the semantics alterning flags | |
400 bellow and inline across optimization boundry. | |
401 Enabling checks bellow breaks several packages by refusing | |
402 to inline library always_inline functions. See PR65873. | |
403 Disable the check for early inlining for now until better solution | |
404 is found. */ | |
405 if (always_inline && early) | |
406 ; | |
407 /* There are some options that change IL semantics which means | |
408 we cannot inline in these cases for correctness reason. | |
409 Not even for always_inline declared functions. */ | |
410 else if (check_match (flag_wrapv) | |
411 || check_match (flag_trapv) | |
412 || check_match (flag_pcc_struct_return) | |
413 /* When caller or callee does FP math, be sure FP codegen flags | |
414 compatible. */ | |
415 || ((caller_info->fp_expressions && callee_info->fp_expressions) | |
416 && (check_maybe_up (flag_rounding_math) | |
417 || check_maybe_up (flag_trapping_math) | |
418 || check_maybe_down (flag_unsafe_math_optimizations) | |
419 || check_maybe_down (flag_finite_math_only) | |
420 || check_maybe_up (flag_signaling_nans) | |
421 || check_maybe_down (flag_cx_limited_range) | |
422 || check_maybe_up (flag_signed_zeros) | |
423 || check_maybe_down (flag_associative_math) | |
424 || check_maybe_down (flag_reciprocal_math) | |
425 || check_maybe_down (flag_fp_int_builtin_inexact) | |
426 /* Strictly speaking only when the callee contains function | |
427 calls that may end up setting errno. */ | |
428 || check_maybe_up (flag_errno_math))) | |
429 /* We do not want to make code compiled with exceptions to be | |
430 brought into a non-EH function unless we know that the callee | |
431 does not throw. | |
432 This is tracked by DECL_FUNCTION_PERSONALITY. */ | |
433 || (check_maybe_up (flag_non_call_exceptions) | |
434 && DECL_FUNCTION_PERSONALITY (callee->decl)) | |
435 || (check_maybe_up (flag_exceptions) | |
436 && DECL_FUNCTION_PERSONALITY (callee->decl)) | |
437 /* When devirtualization is diabled for callee, it is not safe | |
438 to inline it as we possibly mangled the type info. | |
439 Allow early inlining of always inlines. */ | |
440 || (!early && check_maybe_down (flag_devirtualize))) | |
0 | 441 { |
111 | 442 e->inline_failed = CIF_OPTIMIZATION_MISMATCH; |
443 inlinable = false; | |
444 } | |
445 /* gcc.dg/pr43564.c. Apply user-forced inline even at -O0. */ | |
446 else if (always_inline) | |
447 ; | |
448 /* When user added an attribute to the callee honor it. */ | |
449 else if (lookup_attribute ("optimize", DECL_ATTRIBUTES (callee->decl)) | |
450 && opts_for_fn (caller->decl) != opts_for_fn (callee->decl)) | |
451 { | |
452 e->inline_failed = CIF_OPTIMIZATION_MISMATCH; | |
453 inlinable = false; | |
454 } | |
455 /* If explicit optimize attribute are not used, the mismatch is caused | |
456 by different command line options used to build different units. | |
457 Do not care about COMDAT functions - those are intended to be | |
458 optimized with the optimization flags of module they are used in. | |
459 Also do not care about mixing up size/speed optimization when | |
460 DECL_DISREGARD_INLINE_LIMITS is set. */ | |
461 else if ((callee->merged_comdat | |
462 && !lookup_attribute ("optimize", | |
463 DECL_ATTRIBUTES (caller->decl))) | |
464 || DECL_DISREGARD_INLINE_LIMITS (callee->decl)) | |
465 ; | |
466 /* If mismatch is caused by merging two LTO units with different | |
467 optimizationflags we want to be bit nicer. However never inline | |
468 if one of functions is not optimized at all. */ | |
469 else if (!opt_for_fn (callee->decl, optimize) | |
470 || !opt_for_fn (caller->decl, optimize)) | |
471 { | |
472 e->inline_failed = CIF_OPTIMIZATION_MISMATCH; | |
473 inlinable = false; | |
474 } | |
475 /* If callee is optimized for size and caller is not, allow inlining if | |
476 code shrinks or we are in MAX_INLINE_INSNS_SINGLE limit and callee | |
477 is inline (and thus likely an unified comdat). This will allow caller | |
478 to run faster. */ | |
479 else if (opt_for_fn (callee->decl, optimize_size) | |
480 > opt_for_fn (caller->decl, optimize_size)) | |
481 { | |
482 int growth = estimate_edge_growth (e); | |
483 if (growth > 0 | |
484 && (!DECL_DECLARED_INLINE_P (callee->decl) | |
485 && growth >= MAX (MAX_INLINE_INSNS_SINGLE, | |
486 MAX_INLINE_INSNS_AUTO))) | |
487 { | |
488 e->inline_failed = CIF_OPTIMIZATION_MISMATCH; | |
489 inlinable = false; | |
490 } | |
491 } | |
492 /* If callee is more aggressively optimized for performance than caller, | |
493 we generally want to inline only cheap (runtime wise) functions. */ | |
494 else if (opt_for_fn (callee->decl, optimize_size) | |
495 < opt_for_fn (caller->decl, optimize_size) | |
496 || (opt_for_fn (callee->decl, optimize) | |
497 > opt_for_fn (caller->decl, optimize))) | |
498 { | |
499 if (estimate_edge_time (e) | |
500 >= 20 + ipa_call_summaries->get (e)->call_stmt_time) | |
501 { | |
502 e->inline_failed = CIF_OPTIMIZATION_MISMATCH; | |
503 inlinable = false; | |
504 } | |
0 | 505 } |
111 | 506 |
507 } | |
508 | |
509 if (!inlinable && report) | |
510 report_inline_failed_reason (e); | |
511 return inlinable; | |
512 } | |
513 | |
514 | |
515 /* Return true if the edge E is inlinable during early inlining. */ | |
516 | |
517 static bool | |
518 can_early_inline_edge_p (struct cgraph_edge *e) | |
519 { | |
520 struct cgraph_node *callee = e->callee->ultimate_alias_target (); | |
521 /* Early inliner might get called at WPA stage when IPA pass adds new | |
522 function. In this case we can not really do any of early inlining | |
523 because function bodies are missing. */ | |
524 if (cgraph_inline_failed_type (e->inline_failed) == CIF_FINAL_ERROR) | |
525 return false; | |
526 if (!gimple_has_body_p (callee->decl)) | |
527 { | |
528 e->inline_failed = CIF_BODY_NOT_AVAILABLE; | |
529 return false; | |
530 } | |
531 /* In early inliner some of callees may not be in SSA form yet | |
532 (i.e. the callgraph is cyclic and we did not process | |
533 the callee by early inliner, yet). We don't have CIF code for this | |
534 case; later we will re-do the decision in the real inliner. */ | |
535 if (!gimple_in_ssa_p (DECL_STRUCT_FUNCTION (e->caller->decl)) | |
536 || !gimple_in_ssa_p (DECL_STRUCT_FUNCTION (callee->decl))) | |
537 { | |
538 if (dump_file) | |
539 fprintf (dump_file, " edge not inlinable: not in SSA form\n"); | |
540 return false; | |
541 } | |
542 if (!can_inline_edge_p (e, true, false, true)) | |
543 return false; | |
544 return true; | |
545 } | |
546 | |
547 | |
548 /* Return number of calls in N. Ignore cheap builtins. */ | |
549 | |
550 static int | |
551 num_calls (struct cgraph_node *n) | |
552 { | |
553 struct cgraph_edge *e; | |
554 int num = 0; | |
555 | |
556 for (e = n->callees; e; e = e->next_callee) | |
557 if (!is_inexpensive_builtin (e->callee->decl)) | |
558 num++; | |
559 return num; | |
560 } | |
561 | |
562 | |
563 /* Return true if we are interested in inlining small function. */ | |
564 | |
565 static bool | |
566 want_early_inline_function_p (struct cgraph_edge *e) | |
567 { | |
568 bool want_inline = true; | |
569 struct cgraph_node *callee = e->callee->ultimate_alias_target (); | |
570 | |
571 if (DECL_DISREGARD_INLINE_LIMITS (callee->decl)) | |
572 ; | |
573 /* For AutoFDO, we need to make sure that before profile summary, all | |
574 hot paths' IR look exactly the same as profiled binary. As a result, | |
575 in einliner, we will disregard size limit and inline those callsites | |
576 that are: | |
577 * inlined in the profiled binary, and | |
578 * the cloned callee has enough samples to be considered "hot". */ | |
579 else if (flag_auto_profile && afdo_callsite_hot_enough_for_early_inline (e)) | |
580 ; | |
581 else if (!DECL_DECLARED_INLINE_P (callee->decl) | |
582 && !opt_for_fn (e->caller->decl, flag_inline_small_functions)) | |
583 { | |
584 e->inline_failed = CIF_FUNCTION_NOT_INLINE_CANDIDATE; | |
585 report_inline_failed_reason (e); | |
586 want_inline = false; | |
0 | 587 } |
588 else | |
589 { | |
111 | 590 int growth = estimate_edge_growth (e); |
591 int n; | |
592 | |
593 if (growth <= 0) | |
594 ; | |
595 else if (!e->maybe_hot_p () | |
596 && growth > 0) | |
597 { | |
598 if (dump_file) | |
599 fprintf (dump_file, " will not early inline: %s->%s, " | |
600 "call is cold and code would grow by %i\n", | |
601 e->caller->dump_name (), | |
602 callee->dump_name (), | |
603 growth); | |
604 want_inline = false; | |
605 } | |
606 else if (growth > PARAM_VALUE (PARAM_EARLY_INLINING_INSNS)) | |
0 | 607 { |
111 | 608 if (dump_file) |
609 fprintf (dump_file, " will not early inline: %s->%s, " | |
610 "growth %i exceeds --param early-inlining-insns\n", | |
611 e->caller->dump_name (), | |
612 callee->dump_name (), | |
613 growth); | |
614 want_inline = false; | |
615 } | |
616 else if ((n = num_calls (callee)) != 0 | |
617 && growth * (n + 1) > PARAM_VALUE (PARAM_EARLY_INLINING_INSNS)) | |
618 { | |
619 if (dump_file) | |
620 fprintf (dump_file, " will not early inline: %s->%s, " | |
621 "growth %i exceeds --param early-inlining-insns " | |
622 "divided by number of calls\n", | |
623 e->caller->dump_name (), | |
624 callee->dump_name (), | |
625 growth); | |
626 want_inline = false; | |
0 | 627 } |
628 } | |
111 | 629 return want_inline; |
630 } | |
0 | 631 |
111 | 632 /* Compute time of the edge->caller + edge->callee execution when inlining |
633 does not happen. */ | |
634 | |
635 inline sreal | |
636 compute_uninlined_call_time (struct cgraph_edge *edge, | |
637 sreal uninlined_call_time) | |
638 { | |
639 cgraph_node *caller = (edge->caller->global.inlined_to | |
640 ? edge->caller->global.inlined_to | |
641 : edge->caller); | |
642 | |
643 if (edge->count > profile_count::zero () | |
644 && caller->count > profile_count::zero ()) | |
645 uninlined_call_time *= (sreal)edge->count.to_gcov_type () | |
646 / caller->count.to_gcov_type (); | |
647 if (edge->frequency) | |
648 uninlined_call_time *= cgraph_freq_base_rec * edge->frequency; | |
649 else | |
650 uninlined_call_time = uninlined_call_time >> 11; | |
651 | |
652 sreal caller_time = ipa_fn_summaries->get (caller)->time; | |
653 return uninlined_call_time + caller_time; | |
654 } | |
655 | |
656 /* Same as compute_uinlined_call_time but compute time when inlining | |
657 does happen. */ | |
658 | |
659 inline sreal | |
660 compute_inlined_call_time (struct cgraph_edge *edge, | |
661 sreal time) | |
662 { | |
663 cgraph_node *caller = (edge->caller->global.inlined_to | |
664 ? edge->caller->global.inlined_to | |
665 : edge->caller); | |
666 sreal caller_time = ipa_fn_summaries->get (caller)->time; | |
667 | |
668 if (edge->count > profile_count::zero () | |
669 && caller->count > profile_count::zero ()) | |
670 time *= (sreal)edge->count.to_gcov_type () / caller->count.to_gcov_type (); | |
671 if (edge->frequency) | |
672 time *= cgraph_freq_base_rec * edge->frequency; | |
673 else | |
674 time = time >> 11; | |
675 | |
676 /* This calculation should match one in ipa-inline-analysis.c | |
677 (estimate_edge_size_and_time). */ | |
678 time -= (sreal) edge->frequency | |
679 * ipa_call_summaries->get (edge)->call_stmt_time / CGRAPH_FREQ_BASE; | |
680 time += caller_time; | |
681 if (time <= 0) | |
682 time = ((sreal) 1) >> 8; | |
683 gcc_checking_assert (time >= 0); | |
684 return time; | |
685 } | |
686 | |
687 /* Return true if the speedup for inlining E is bigger than | |
688 PARAM_MAX_INLINE_MIN_SPEEDUP. */ | |
689 | |
690 static bool | |
691 big_speedup_p (struct cgraph_edge *e) | |
692 { | |
693 sreal unspec_time; | |
694 sreal spec_time = estimate_edge_time (e, &unspec_time); | |
695 sreal time = compute_uninlined_call_time (e, unspec_time); | |
696 sreal inlined_time = compute_inlined_call_time (e, spec_time); | |
697 | |
698 if (time - inlined_time | |
699 > (sreal) (time * PARAM_VALUE (PARAM_INLINE_MIN_SPEEDUP)) | |
700 * percent_rec) | |
701 return true; | |
702 return false; | |
0 | 703 } |
704 | |
111 | 705 /* Return true if we are interested in inlining small function. |
706 When REPORT is true, report reason to dump file. */ | |
707 | |
708 static bool | |
709 want_inline_small_function_p (struct cgraph_edge *e, bool report) | |
710 { | |
711 bool want_inline = true; | |
712 struct cgraph_node *callee = e->callee->ultimate_alias_target (); | |
713 | |
714 if (DECL_DISREGARD_INLINE_LIMITS (callee->decl)) | |
715 ; | |
716 else if (!DECL_DECLARED_INLINE_P (callee->decl) | |
717 && !opt_for_fn (e->caller->decl, flag_inline_small_functions)) | |
718 { | |
719 e->inline_failed = CIF_FUNCTION_NOT_INLINE_CANDIDATE; | |
720 want_inline = false; | |
721 } | |
722 /* Do fast and conservative check if the function can be good | |
723 inline candidate. At the moment we allow inline hints to | |
724 promote non-inline functions to inline and we increase | |
725 MAX_INLINE_INSNS_SINGLE 16-fold for inline functions. */ | |
726 else if ((!DECL_DECLARED_INLINE_P (callee->decl) | |
727 && (!e->count.initialized_p () || !e->maybe_hot_p ())) | |
728 && ipa_fn_summaries->get (callee)->min_size | |
729 - ipa_call_summaries->get (e)->call_stmt_size | |
730 > MAX (MAX_INLINE_INSNS_SINGLE, MAX_INLINE_INSNS_AUTO)) | |
731 { | |
732 e->inline_failed = CIF_MAX_INLINE_INSNS_AUTO_LIMIT; | |
733 want_inline = false; | |
734 } | |
735 else if ((DECL_DECLARED_INLINE_P (callee->decl) | |
736 || e->count > profile_count::zero ()) | |
737 && ipa_fn_summaries->get (callee)->min_size | |
738 - ipa_call_summaries->get (e)->call_stmt_size | |
739 > 16 * MAX_INLINE_INSNS_SINGLE) | |
740 { | |
741 e->inline_failed = (DECL_DECLARED_INLINE_P (callee->decl) | |
742 ? CIF_MAX_INLINE_INSNS_SINGLE_LIMIT | |
743 : CIF_MAX_INLINE_INSNS_AUTO_LIMIT); | |
744 want_inline = false; | |
745 } | |
746 else | |
747 { | |
748 int growth = estimate_edge_growth (e); | |
749 ipa_hints hints = estimate_edge_hints (e); | |
750 bool big_speedup = big_speedup_p (e); | |
0 | 751 |
111 | 752 if (growth <= 0) |
753 ; | |
754 /* Apply MAX_INLINE_INSNS_SINGLE limit. Do not do so when | |
755 hints suggests that inlining given function is very profitable. */ | |
756 else if (DECL_DECLARED_INLINE_P (callee->decl) | |
757 && growth >= MAX_INLINE_INSNS_SINGLE | |
758 && ((!big_speedup | |
759 && !(hints & (INLINE_HINT_indirect_call | |
760 | INLINE_HINT_known_hot | |
761 | INLINE_HINT_loop_iterations | |
762 | INLINE_HINT_array_index | |
763 | INLINE_HINT_loop_stride))) | |
764 || growth >= MAX_INLINE_INSNS_SINGLE * 16)) | |
765 { | |
766 e->inline_failed = CIF_MAX_INLINE_INSNS_SINGLE_LIMIT; | |
767 want_inline = false; | |
768 } | |
769 else if (!DECL_DECLARED_INLINE_P (callee->decl) | |
770 && !opt_for_fn (e->caller->decl, flag_inline_functions)) | |
771 { | |
772 /* growth_likely_positive is expensive, always test it last. */ | |
773 if (growth >= MAX_INLINE_INSNS_SINGLE | |
774 || growth_likely_positive (callee, growth)) | |
775 { | |
776 e->inline_failed = CIF_NOT_DECLARED_INLINED; | |
777 want_inline = false; | |
778 } | |
779 } | |
780 /* Apply MAX_INLINE_INSNS_AUTO limit for functions not declared inline | |
781 Upgrade it to MAX_INLINE_INSNS_SINGLE when hints suggests that | |
782 inlining given function is very profitable. */ | |
783 else if (!DECL_DECLARED_INLINE_P (callee->decl) | |
784 && !big_speedup | |
785 && !(hints & INLINE_HINT_known_hot) | |
786 && growth >= ((hints & (INLINE_HINT_indirect_call | |
787 | INLINE_HINT_loop_iterations | |
788 | INLINE_HINT_array_index | |
789 | INLINE_HINT_loop_stride)) | |
790 ? MAX (MAX_INLINE_INSNS_AUTO, | |
791 MAX_INLINE_INSNS_SINGLE) | |
792 : MAX_INLINE_INSNS_AUTO)) | |
793 { | |
794 /* growth_likely_positive is expensive, always test it last. */ | |
795 if (growth >= MAX_INLINE_INSNS_SINGLE | |
796 || growth_likely_positive (callee, growth)) | |
797 { | |
798 e->inline_failed = CIF_MAX_INLINE_INSNS_AUTO_LIMIT; | |
799 want_inline = false; | |
800 } | |
801 } | |
802 /* If call is cold, do not inline when function body would grow. */ | |
803 else if (!e->maybe_hot_p () | |
804 && (growth >= MAX_INLINE_INSNS_SINGLE | |
805 || growth_likely_positive (callee, growth))) | |
806 { | |
807 e->inline_failed = CIF_UNLIKELY_CALL; | |
808 want_inline = false; | |
809 } | |
810 } | |
811 if (!want_inline && report) | |
812 report_inline_failed_reason (e); | |
813 return want_inline; | |
814 } | |
815 | |
816 /* EDGE is self recursive edge. | |
817 We hand two cases - when function A is inlining into itself | |
818 or when function A is being inlined into another inliner copy of function | |
819 A within function B. | |
820 | |
821 In first case OUTER_NODE points to the toplevel copy of A, while | |
822 in the second case OUTER_NODE points to the outermost copy of A in B. | |
823 | |
824 In both cases we want to be extra selective since | |
825 inlining the call will just introduce new recursive calls to appear. */ | |
826 | |
827 static bool | |
828 want_inline_self_recursive_call_p (struct cgraph_edge *edge, | |
829 struct cgraph_node *outer_node, | |
830 bool peeling, | |
831 int depth) | |
0 | 832 { |
111 | 833 char const *reason = NULL; |
834 bool want_inline = true; | |
835 int caller_freq = CGRAPH_FREQ_BASE; | |
836 int max_depth = PARAM_VALUE (PARAM_MAX_INLINE_RECURSIVE_DEPTH_AUTO); | |
837 | |
838 if (DECL_DECLARED_INLINE_P (edge->caller->decl)) | |
839 max_depth = PARAM_VALUE (PARAM_MAX_INLINE_RECURSIVE_DEPTH); | |
840 | |
841 if (!edge->maybe_hot_p ()) | |
842 { | |
843 reason = "recursive call is cold"; | |
844 want_inline = false; | |
845 } | |
846 else if (outer_node->count == profile_count::zero ()) | |
847 { | |
848 reason = "not executed in profile"; | |
849 want_inline = false; | |
850 } | |
851 else if (depth > max_depth) | |
852 { | |
853 reason = "--param max-inline-recursive-depth exceeded."; | |
854 want_inline = false; | |
855 } | |
856 | |
857 if (outer_node->global.inlined_to) | |
858 caller_freq = outer_node->callers->frequency; | |
859 | |
860 if (!caller_freq) | |
861 { | |
862 reason = "function is inlined and unlikely"; | |
863 want_inline = false; | |
864 } | |
865 | |
866 if (!want_inline) | |
867 ; | |
868 /* Inlining of self recursive function into copy of itself within other function | |
869 is transformation similar to loop peeling. | |
870 | |
871 Peeling is profitable if we can inline enough copies to make probability | |
872 of actual call to the self recursive function very small. Be sure that | |
873 the probability of recursion is small. | |
874 | |
875 We ensure that the frequency of recursing is at most 1 - (1/max_depth). | |
876 This way the expected number of recision is at most max_depth. */ | |
877 else if (peeling) | |
878 { | |
879 int max_prob = CGRAPH_FREQ_BASE - ((CGRAPH_FREQ_BASE + max_depth - 1) | |
880 / max_depth); | |
881 int i; | |
882 for (i = 1; i < depth; i++) | |
883 max_prob = max_prob * max_prob / CGRAPH_FREQ_BASE; | |
884 if (max_count > profile_count::zero () && edge->count > profile_count::zero () | |
885 && (edge->count.to_gcov_type () * CGRAPH_FREQ_BASE | |
886 / outer_node->count.to_gcov_type () | |
887 >= max_prob)) | |
888 { | |
889 reason = "profile of recursive call is too large"; | |
890 want_inline = false; | |
891 } | |
892 if (max_count == profile_count::zero () | |
893 && (edge->frequency * CGRAPH_FREQ_BASE / caller_freq | |
894 >= max_prob)) | |
895 { | |
896 reason = "frequency of recursive call is too large"; | |
897 want_inline = false; | |
898 } | |
899 } | |
900 /* Recursive inlining, i.e. equivalent of unrolling, is profitable if recursion | |
901 depth is large. We reduce function call overhead and increase chances that | |
902 things fit in hardware return predictor. | |
903 | |
904 Recursive inlining might however increase cost of stack frame setup | |
905 actually slowing down functions whose recursion tree is wide rather than | |
906 deep. | |
907 | |
908 Deciding reliably on when to do recursive inlining without profile feedback | |
909 is tricky. For now we disable recursive inlining when probability of self | |
910 recursion is low. | |
911 | |
912 Recursive inlining of self recursive call within loop also results in large loop | |
913 depths that generally optimize badly. We may want to throttle down inlining | |
914 in those cases. In particular this seems to happen in one of libstdc++ rb tree | |
915 methods. */ | |
0 | 916 else |
111 | 917 { |
918 if (max_count > profile_count::zero () && edge->count.initialized_p () | |
919 && (edge->count.to_gcov_type () * 100 | |
920 / outer_node->count.to_gcov_type () | |
921 <= PARAM_VALUE (PARAM_MIN_INLINE_RECURSIVE_PROBABILITY))) | |
922 { | |
923 reason = "profile of recursive call is too small"; | |
924 want_inline = false; | |
925 } | |
926 else if ((max_count == profile_count::zero () | |
927 || !edge->count.initialized_p ()) | |
928 && (edge->frequency * 100 / caller_freq | |
929 <= PARAM_VALUE (PARAM_MIN_INLINE_RECURSIVE_PROBABILITY))) | |
930 { | |
931 reason = "frequency of recursive call is too small"; | |
932 want_inline = false; | |
933 } | |
934 } | |
935 if (!want_inline && dump_file) | |
936 fprintf (dump_file, " not inlining recursively: %s\n", reason); | |
937 return want_inline; | |
938 } | |
939 | |
940 /* Return true when NODE has uninlinable caller; | |
941 set HAS_HOT_CALL if it has hot call. | |
942 Worker for cgraph_for_node_and_aliases. */ | |
943 | |
944 static bool | |
945 check_callers (struct cgraph_node *node, void *has_hot_call) | |
946 { | |
947 struct cgraph_edge *e; | |
948 for (e = node->callers; e; e = e->next_caller) | |
949 { | |
950 if (!opt_for_fn (e->caller->decl, flag_inline_functions_called_once) | |
951 || !opt_for_fn (e->caller->decl, optimize)) | |
952 return true; | |
953 if (!can_inline_edge_p (e, true)) | |
954 return true; | |
955 if (e->recursive_p ()) | |
956 return true; | |
957 if (!(*(bool *)has_hot_call) && e->maybe_hot_p ()) | |
958 *(bool *)has_hot_call = true; | |
959 } | |
960 return false; | |
961 } | |
962 | |
963 /* If NODE has a caller, return true. */ | |
964 | |
965 static bool | |
966 has_caller_p (struct cgraph_node *node, void *data ATTRIBUTE_UNUSED) | |
967 { | |
968 if (node->callers) | |
969 return true; | |
970 return false; | |
971 } | |
972 | |
973 /* Decide if inlining NODE would reduce unit size by eliminating | |
974 the offline copy of function. | |
975 When COLD is true the cold calls are considered, too. */ | |
976 | |
977 static bool | |
978 want_inline_function_to_all_callers_p (struct cgraph_node *node, bool cold) | |
979 { | |
980 bool has_hot_call = false; | |
981 | |
982 /* Aliases gets inlined along with the function they alias. */ | |
983 if (node->alias) | |
984 return false; | |
985 /* Already inlined? */ | |
986 if (node->global.inlined_to) | |
987 return false; | |
988 /* Does it have callers? */ | |
989 if (!node->call_for_symbol_and_aliases (has_caller_p, NULL, true)) | |
990 return false; | |
991 /* Inlining into all callers would increase size? */ | |
992 if (estimate_growth (node) > 0) | |
993 return false; | |
994 /* All inlines must be possible. */ | |
995 if (node->call_for_symbol_and_aliases (check_callers, &has_hot_call, | |
996 true)) | |
997 return false; | |
998 if (!cold && !has_hot_call) | |
999 return false; | |
1000 return true; | |
0 | 1001 } |
1002 | |
1003 /* A cost model driving the inlining heuristics in a way so the edges with | |
1004 smallest badness are inlined first. After each inlining is performed | |
1005 the costs of all caller edges of nodes affected are recomputed so the | |
1006 metrics may accurately depend on values such as number of inlinable callers | |
1007 of the function or function body size. */ | |
1008 | |
111 | 1009 static sreal |
1010 edge_badness (struct cgraph_edge *edge, bool dump) | |
0 | 1011 { |
111 | 1012 sreal badness; |
1013 int growth; | |
1014 sreal edge_time, unspec_edge_time; | |
1015 struct cgraph_node *callee = edge->callee->ultimate_alias_target (); | |
1016 struct ipa_fn_summary *callee_info = ipa_fn_summaries->get (callee); | |
1017 ipa_hints hints; | |
1018 cgraph_node *caller = (edge->caller->global.inlined_to | |
1019 ? edge->caller->global.inlined_to | |
1020 : edge->caller); | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1021 |
111 | 1022 growth = estimate_edge_growth (edge); |
1023 edge_time = estimate_edge_time (edge, &unspec_edge_time); | |
1024 hints = estimate_edge_hints (edge); | |
1025 gcc_checking_assert (edge_time >= 0); | |
1026 /* Check that inlined time is better, but tolerate some roundoff issues. */ | |
1027 gcc_checking_assert ((edge_time - callee_info->time).to_int () <= 0); | |
1028 gcc_checking_assert (growth <= callee_info->size); | |
0 | 1029 |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1030 if (dump) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1031 { |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1032 fprintf (dump_file, " Badness calculation for %s -> %s\n", |
111 | 1033 edge->caller->dump_name (), |
1034 edge->callee->dump_name ()); | |
1035 fprintf (dump_file, " size growth %i, time %f unspec %f ", | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1036 growth, |
111 | 1037 edge_time.to_double (), |
1038 unspec_edge_time.to_double ()); | |
1039 ipa_dump_hints (dump_file, hints); | |
1040 if (big_speedup_p (edge)) | |
1041 fprintf (dump_file, " big_speedup"); | |
1042 fprintf (dump_file, "\n"); | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1043 } |
0 | 1044 |
1045 /* Always prefer inlining saving code size. */ | |
1046 if (growth <= 0) | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1047 { |
111 | 1048 badness = (sreal) (-SREAL_MIN_SIG + growth) << (SREAL_MAX_EXP / 256); |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1049 if (dump) |
111 | 1050 fprintf (dump_file, " %f: Growth %d <= 0\n", badness.to_double (), |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1051 growth); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1052 } |
111 | 1053 /* Inlining into EXTERNAL functions is not going to change anything unless |
1054 they are themselves inlined. */ | |
1055 else if (DECL_EXTERNAL (caller->decl)) | |
1056 { | |
1057 if (dump) | |
1058 fprintf (dump_file, " max: function is external\n"); | |
1059 return sreal::max (); | |
1060 } | |
1061 /* When profile is available. Compute badness as: | |
1062 | |
1063 time_saved * caller_count | |
1064 goodness = ------------------------------------------------- | |
1065 growth_of_caller * overall_growth * combined_size | |
0 | 1066 |
111 | 1067 badness = - goodness |
1068 | |
1069 Again use negative value to make calls with profile appear hotter | |
1070 then calls without. | |
1071 */ | |
1072 else if (opt_for_fn (caller->decl, flag_guess_branch_prob) | |
1073 || caller->count > profile_count::zero ()) | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1074 { |
111 | 1075 sreal numerator, denominator; |
1076 int overall_growth; | |
1077 sreal inlined_time = compute_inlined_call_time (edge, edge_time); | |
1078 | |
1079 numerator = (compute_uninlined_call_time (edge, unspec_edge_time) | |
1080 - inlined_time); | |
1081 if (numerator == 0) | |
1082 numerator = ((sreal) 1 >> 8); | |
1083 if (caller->count > profile_count::zero ()) | |
1084 numerator *= caller->count.to_gcov_type (); | |
1085 else if (caller->count.initialized_p ()) | |
1086 numerator = numerator >> 11; | |
1087 denominator = growth; | |
1088 | |
1089 overall_growth = callee_info->growth; | |
1090 | |
1091 /* Look for inliner wrappers of the form: | |
1092 | |
1093 inline_caller () | |
1094 { | |
1095 do_fast_job... | |
1096 if (need_more_work) | |
1097 noninline_callee (); | |
1098 } | |
1099 Withhout panilizing this case, we usually inline noninline_callee | |
1100 into the inline_caller because overall_growth is small preventing | |
1101 further inlining of inline_caller. | |
1102 | |
1103 Penalize only callgraph edges to functions with small overall | |
1104 growth ... | |
1105 */ | |
1106 if (growth > overall_growth | |
1107 /* ... and having only one caller which is not inlined ... */ | |
1108 && callee_info->single_caller | |
1109 && !edge->caller->global.inlined_to | |
1110 /* ... and edges executed only conditionally ... */ | |
1111 && edge->frequency < CGRAPH_FREQ_BASE | |
1112 /* ... consider case where callee is not inline but caller is ... */ | |
1113 && ((!DECL_DECLARED_INLINE_P (edge->callee->decl) | |
1114 && DECL_DECLARED_INLINE_P (caller->decl)) | |
1115 /* ... or when early optimizers decided to split and edge | |
1116 frequency still indicates splitting is a win ... */ | |
1117 || (callee->split_part && !caller->split_part | |
1118 && edge->frequency | |
1119 < CGRAPH_FREQ_BASE | |
1120 * PARAM_VALUE | |
1121 (PARAM_PARTIAL_INLINING_ENTRY_PROBABILITY) / 100 | |
1122 /* ... and do not overwrite user specified hints. */ | |
1123 && (!DECL_DECLARED_INLINE_P (edge->callee->decl) | |
1124 || DECL_DECLARED_INLINE_P (caller->decl))))) | |
1125 { | |
1126 struct ipa_fn_summary *caller_info = ipa_fn_summaries->get (caller); | |
1127 int caller_growth = caller_info->growth; | |
1128 | |
1129 /* Only apply the penalty when caller looks like inline candidate, | |
1130 and it is not called once and. */ | |
1131 if (!caller_info->single_caller && overall_growth < caller_growth | |
1132 && caller_info->inlinable | |
1133 && caller_info->size | |
1134 < (DECL_DECLARED_INLINE_P (caller->decl) | |
1135 ? MAX_INLINE_INSNS_SINGLE : MAX_INLINE_INSNS_AUTO)) | |
1136 { | |
1137 if (dump) | |
1138 fprintf (dump_file, | |
1139 " Wrapper penalty. Increasing growth %i to %i\n", | |
1140 overall_growth, caller_growth); | |
1141 overall_growth = caller_growth; | |
1142 } | |
1143 } | |
1144 if (overall_growth > 0) | |
1145 { | |
1146 /* Strongly preffer functions with few callers that can be inlined | |
1147 fully. The square root here leads to smaller binaries at average. | |
1148 Watch however for extreme cases and return to linear function | |
1149 when growth is large. */ | |
1150 if (overall_growth < 256) | |
1151 overall_growth *= overall_growth; | |
1152 else | |
1153 overall_growth += 256 * 256 - 256; | |
1154 denominator *= overall_growth; | |
1155 } | |
1156 denominator *= inlined_time; | |
1157 | |
1158 badness = - numerator / denominator; | |
1159 | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1160 if (dump) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1161 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1162 fprintf (dump_file, |
111 | 1163 " %f: guessed profile. frequency %f, count %" PRId64 |
1164 " caller count %" PRId64 | |
1165 " time w/o inlining %f, time with inlining %f" | |
1166 " overall growth %i (current) %i (original)" | |
1167 " %i (compensated)\n", | |
1168 badness.to_double (), | |
1169 (double)edge->frequency / CGRAPH_FREQ_BASE, | |
1170 edge->count.initialized_p () ? edge->count.to_gcov_type () : -1, | |
1171 caller->count.initialized_p () ? caller->count.to_gcov_type () : -1, | |
1172 compute_uninlined_call_time (edge, | |
1173 unspec_edge_time).to_double (), | |
1174 compute_inlined_call_time (edge, edge_time).to_double (), | |
1175 estimate_growth (callee), | |
1176 callee_info->growth, overall_growth); | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1177 } |
0 | 1178 } |
1179 /* When function local profile is not available or it does not give | |
1180 useful information (ie frequency is zero), base the cost on | |
1181 loop nest and overall size growth, so we optimize for overall number | |
1182 of functions fully inlined in program. */ | |
1183 else | |
1184 { | |
111 | 1185 int nest = MIN (ipa_call_summaries->get (edge)->loop_depth, 8); |
1186 badness = growth; | |
0 | 1187 |
1188 /* Decrease badness if call is nested. */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1189 if (badness > 0) |
111 | 1190 badness = badness >> nest; |
0 | 1191 else |
111 | 1192 badness = badness << nest; |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1193 if (dump) |
111 | 1194 fprintf (dump_file, " %f: no profile. nest %i\n", |
1195 badness.to_double (), nest); | |
0 | 1196 } |
111 | 1197 gcc_checking_assert (badness != 0); |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1198 |
111 | 1199 if (edge->recursive_p ()) |
1200 badness = badness.shift (badness > 0 ? 4 : -4); | |
1201 if ((hints & (INLINE_HINT_indirect_call | |
1202 | INLINE_HINT_loop_iterations | |
1203 | INLINE_HINT_array_index | |
1204 | INLINE_HINT_loop_stride)) | |
1205 || callee_info->growth <= 0) | |
1206 badness = badness.shift (badness > 0 ? -2 : 2); | |
1207 if (hints & (INLINE_HINT_same_scc)) | |
1208 badness = badness.shift (badness > 0 ? 3 : -3); | |
1209 else if (hints & (INLINE_HINT_in_scc)) | |
1210 badness = badness.shift (badness > 0 ? 2 : -2); | |
1211 else if (hints & (INLINE_HINT_cross_module)) | |
1212 badness = badness.shift (badness > 0 ? 1 : -1); | |
1213 if (DECL_DISREGARD_INLINE_LIMITS (callee->decl)) | |
1214 badness = badness.shift (badness > 0 ? -4 : 4); | |
1215 else if ((hints & INLINE_HINT_declared_inline)) | |
1216 badness = badness.shift (badness > 0 ? -3 : 3); | |
1217 if (dump) | |
1218 fprintf (dump_file, " Adjusted by hints %f\n", badness.to_double ()); | |
1219 return badness; | |
0 | 1220 } |
1221 | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1222 /* Recompute badness of EDGE and update its key in HEAP if needed. */ |
111 | 1223 static inline void |
1224 update_edge_key (edge_heap_t *heap, struct cgraph_edge *edge) | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1225 { |
111 | 1226 sreal badness = edge_badness (edge, false); |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1227 if (edge->aux) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1228 { |
111 | 1229 edge_heap_node_t *n = (edge_heap_node_t *) edge->aux; |
1230 gcc_checking_assert (n->get_data () == edge); | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1231 |
111 | 1232 /* fibonacci_heap::replace_key does busy updating of the |
1233 heap that is unnecesarily expensive. | |
1234 We do lazy increases: after extracting minimum if the key | |
1235 turns out to be out of date, it is re-inserted into heap | |
1236 with correct value. */ | |
1237 if (badness < n->get_key ()) | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1238 { |
111 | 1239 if (dump_file && (dump_flags & TDF_DETAILS)) |
1240 { | |
1241 fprintf (dump_file, | |
1242 " decreasing badness %s -> %s, %f to %f\n", | |
1243 edge->caller->dump_name (), | |
1244 edge->callee->dump_name (), | |
1245 n->get_key ().to_double (), | |
1246 badness.to_double ()); | |
1247 } | |
1248 heap->decrease_key (n, badness); | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1249 } |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1250 } |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1251 else |
111 | 1252 { |
1253 if (dump_file && (dump_flags & TDF_DETAILS)) | |
1254 { | |
1255 fprintf (dump_file, | |
1256 " enqueuing call %s -> %s, badness %f\n", | |
1257 edge->caller->dump_name (), | |
1258 edge->callee->dump_name (), | |
1259 badness.to_double ()); | |
1260 } | |
1261 edge->aux = heap->insert (badness, edge); | |
1262 } | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1263 } |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1264 |
111 | 1265 |
1266 /* NODE was inlined. | |
1267 All caller edges needs to be resetted because | |
1268 size estimates change. Similarly callees needs reset | |
1269 because better context may be known. */ | |
0 | 1270 |
1271 static void | |
111 | 1272 reset_edge_caches (struct cgraph_node *node) |
0 | 1273 { |
1274 struct cgraph_edge *edge; | |
111 | 1275 struct cgraph_edge *e = node->callees; |
1276 struct cgraph_node *where = node; | |
1277 struct ipa_ref *ref; | |
0 | 1278 |
111 | 1279 if (where->global.inlined_to) |
1280 where = where->global.inlined_to; | |
0 | 1281 |
111 | 1282 for (edge = where->callers; edge; edge = edge->next_caller) |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1283 if (edge->inline_failed) |
111 | 1284 reset_edge_growth_cache (edge); |
0 | 1285 |
111 | 1286 FOR_EACH_ALIAS (where, ref) |
1287 reset_edge_caches (dyn_cast <cgraph_node *> (ref->referring)); | |
0 | 1288 |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1289 if (!e) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1290 return; |
111 | 1291 |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1292 while (true) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1293 if (!e->inline_failed && e->callee->callees) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1294 e = e->callee->callees; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1295 else |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1296 { |
111 | 1297 if (e->inline_failed) |
1298 reset_edge_growth_cache (e); | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1299 if (e->next_callee) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1300 e = e->next_callee; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1301 else |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1302 { |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1303 do |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1304 { |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1305 if (e->caller == node) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1306 return; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1307 e = e->caller->callers; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1308 } |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1309 while (!e->next_callee); |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1310 e = e->next_callee; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1311 } |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1312 } |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1313 } |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1314 |
111 | 1315 /* Recompute HEAP nodes for each of caller of NODE. |
1316 UPDATED_NODES track nodes we already visited, to avoid redundant work. | |
1317 When CHECK_INLINABLITY_FOR is set, re-check for specified edge that | |
1318 it is inlinable. Otherwise check all edges. */ | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1319 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1320 static void |
111 | 1321 update_caller_keys (edge_heap_t *heap, struct cgraph_node *node, |
1322 bitmap updated_nodes, | |
1323 struct cgraph_edge *check_inlinablity_for) | |
1324 { | |
1325 struct cgraph_edge *edge; | |
1326 struct ipa_ref *ref; | |
1327 | |
1328 if ((!node->alias && !ipa_fn_summaries->get (node)->inlinable) | |
1329 || node->global.inlined_to) | |
1330 return; | |
1331 if (!bitmap_set_bit (updated_nodes, node->uid)) | |
1332 return; | |
1333 | |
1334 FOR_EACH_ALIAS (node, ref) | |
1335 { | |
1336 struct cgraph_node *alias = dyn_cast <cgraph_node *> (ref->referring); | |
1337 update_caller_keys (heap, alias, updated_nodes, check_inlinablity_for); | |
1338 } | |
1339 | |
1340 for (edge = node->callers; edge; edge = edge->next_caller) | |
1341 if (edge->inline_failed) | |
1342 { | |
1343 if (!check_inlinablity_for | |
1344 || check_inlinablity_for == edge) | |
1345 { | |
1346 if (can_inline_edge_p (edge, false) | |
1347 && want_inline_small_function_p (edge, false)) | |
1348 update_edge_key (heap, edge); | |
1349 else if (edge->aux) | |
1350 { | |
1351 report_inline_failed_reason (edge); | |
1352 heap->delete_node ((edge_heap_node_t *) edge->aux); | |
1353 edge->aux = NULL; | |
1354 } | |
1355 } | |
1356 else if (edge->aux) | |
1357 update_edge_key (heap, edge); | |
1358 } | |
1359 } | |
1360 | |
1361 /* Recompute HEAP nodes for each uninlined call in NODE. | |
1362 This is used when we know that edge badnesses are going only to increase | |
1363 (we introduced new call site) and thus all we need is to insert newly | |
1364 created edges into heap. */ | |
1365 | |
1366 static void | |
1367 update_callee_keys (edge_heap_t *heap, struct cgraph_node *node, | |
1368 bitmap updated_nodes) | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1369 { |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1370 struct cgraph_edge *e = node->callees; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1371 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1372 if (!e) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1373 return; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1374 while (true) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1375 if (!e->inline_failed && e->callee->callees) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1376 e = e->callee->callees; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1377 else |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1378 { |
111 | 1379 enum availability avail; |
1380 struct cgraph_node *callee; | |
1381 /* We do not reset callee growth cache here. Since we added a new call, | |
1382 growth chould have just increased and consequentely badness metric | |
1383 don't need updating. */ | |
1384 if (e->inline_failed | |
1385 && (callee = e->callee->ultimate_alias_target (&avail, e->caller)) | |
1386 && ipa_fn_summaries->get (callee)->inlinable | |
1387 && avail >= AVAIL_AVAILABLE | |
1388 && !bitmap_bit_p (updated_nodes, callee->uid)) | |
1389 { | |
1390 if (can_inline_edge_p (e, false) | |
1391 && want_inline_small_function_p (e, false)) | |
1392 update_edge_key (heap, e); | |
1393 else if (e->aux) | |
1394 { | |
1395 report_inline_failed_reason (e); | |
1396 heap->delete_node ((edge_heap_node_t *) e->aux); | |
1397 e->aux = NULL; | |
1398 } | |
1399 } | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1400 if (e->next_callee) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1401 e = e->next_callee; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1402 else |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1403 { |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1404 do |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1405 { |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1406 if (e->caller == node) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1407 return; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1408 e = e->caller->callers; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1409 } |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1410 while (!e->next_callee); |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1411 e = e->next_callee; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1412 } |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1413 } |
0 | 1414 } |
1415 | |
1416 /* Enqueue all recursive calls from NODE into priority queue depending on | |
1417 how likely we want to recursively inline the call. */ | |
1418 | |
1419 static void | |
1420 lookup_recursive_calls (struct cgraph_node *node, struct cgraph_node *where, | |
111 | 1421 edge_heap_t *heap) |
0 | 1422 { |
1423 struct cgraph_edge *e; | |
111 | 1424 enum availability avail; |
1425 | |
0 | 1426 for (e = where->callees; e; e = e->next_callee) |
111 | 1427 if (e->callee == node |
1428 || (e->callee->ultimate_alias_target (&avail, e->caller) == node | |
1429 && avail > AVAIL_INTERPOSABLE)) | |
0 | 1430 { |
1431 /* When profile feedback is available, prioritize by expected number | |
111 | 1432 of calls. */ |
1433 heap->insert (!(max_count > 0) || !e->count.initialized_p () ? -e->frequency | |
1434 : -(e->count.to_gcov_type () | |
1435 / ((max_count.to_gcov_type () + (1<<24) - 1) | |
1436 / (1<<24))), | |
1437 e); | |
0 | 1438 } |
1439 for (e = where->callees; e; e = e->next_callee) | |
1440 if (!e->inline_failed) | |
1441 lookup_recursive_calls (node, e->callee, heap); | |
1442 } | |
1443 | |
1444 /* Decide on recursive inlining: in the case function has recursive calls, | |
1445 inline until body size reaches given argument. If any new indirect edges | |
1446 are discovered in the process, add them to *NEW_EDGES, unless NEW_EDGES | |
1447 is NULL. */ | |
1448 | |
1449 static bool | |
111 | 1450 recursive_inlining (struct cgraph_edge *edge, |
1451 vec<cgraph_edge *> *new_edges) | |
0 | 1452 { |
1453 int limit = PARAM_VALUE (PARAM_MAX_INLINE_INSNS_RECURSIVE_AUTO); | |
111 | 1454 edge_heap_t heap (sreal::min ()); |
1455 struct cgraph_node *node; | |
0 | 1456 struct cgraph_edge *e; |
111 | 1457 struct cgraph_node *master_clone = NULL, *next; |
0 | 1458 int depth = 0; |
1459 int n = 0; | |
1460 | |
111 | 1461 node = edge->caller; |
1462 if (node->global.inlined_to) | |
1463 node = node->global.inlined_to; | |
0 | 1464 |
1465 if (DECL_DECLARED_INLINE_P (node->decl)) | |
111 | 1466 limit = PARAM_VALUE (PARAM_MAX_INLINE_INSNS_RECURSIVE); |
0 | 1467 |
1468 /* Make sure that function is small enough to be considered for inlining. */ | |
111 | 1469 if (estimate_size_after_inlining (node, edge) >= limit) |
0 | 1470 return false; |
111 | 1471 lookup_recursive_calls (node, node, &heap); |
1472 if (heap.empty ()) | |
1473 return false; | |
0 | 1474 |
1475 if (dump_file) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1476 fprintf (dump_file, |
0 | 1477 " Performing recursive inlining on %s\n", |
111 | 1478 node->name ()); |
0 | 1479 |
1480 /* Do the inlining and update list of recursive call during process. */ | |
111 | 1481 while (!heap.empty ()) |
0 | 1482 { |
111 | 1483 struct cgraph_edge *curr = heap.extract_min (); |
1484 struct cgraph_node *cnode, *dest = curr->callee; | |
1485 | |
1486 if (!can_inline_edge_p (curr, true)) | |
1487 continue; | |
1488 | |
1489 /* MASTER_CLONE is produced in the case we already started modified | |
1490 the function. Be sure to redirect edge to the original body before | |
1491 estimating growths otherwise we will be seeing growths after inlining | |
1492 the already modified body. */ | |
1493 if (master_clone) | |
1494 { | |
1495 curr->redirect_callee (master_clone); | |
1496 reset_edge_growth_cache (curr); | |
1497 } | |
1498 | |
1499 if (estimate_size_after_inlining (node, curr) > limit) | |
1500 { | |
1501 curr->redirect_callee (dest); | |
1502 reset_edge_growth_cache (curr); | |
1503 break; | |
1504 } | |
0 | 1505 |
1506 depth = 1; | |
1507 for (cnode = curr->caller; | |
1508 cnode->global.inlined_to; cnode = cnode->callers->caller) | |
111 | 1509 if (node->decl |
1510 == curr->callee->ultimate_alias_target ()->decl) | |
1511 depth++; | |
0 | 1512 |
111 | 1513 if (!want_inline_self_recursive_call_p (curr, node, false, depth)) |
0 | 1514 { |
111 | 1515 curr->redirect_callee (dest); |
1516 reset_edge_growth_cache (curr); | |
1517 continue; | |
0 | 1518 } |
1519 | |
1520 if (dump_file) | |
1521 { | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1522 fprintf (dump_file, |
0 | 1523 " Inlining call of depth %i", depth); |
111 | 1524 if (node->count > profile_count::zero ()) |
0 | 1525 { |
1526 fprintf (dump_file, " called approx. %.2f times per call", | |
111 | 1527 (double)curr->count.to_gcov_type () |
1528 / node->count.to_gcov_type ()); | |
0 | 1529 } |
1530 fprintf (dump_file, "\n"); | |
1531 } | |
111 | 1532 if (!master_clone) |
1533 { | |
1534 /* We need original clone to copy around. */ | |
1535 master_clone = node->create_clone (node->decl, node->count, | |
1536 CGRAPH_FREQ_BASE, false, vNULL, | |
1537 true, NULL, NULL); | |
1538 for (e = master_clone->callees; e; e = e->next_callee) | |
1539 if (!e->inline_failed) | |
1540 clone_inlined_nodes (e, true, false, NULL, CGRAPH_FREQ_BASE); | |
1541 curr->redirect_callee (master_clone); | |
1542 reset_edge_growth_cache (curr); | |
1543 } | |
1544 | |
1545 inline_call (curr, false, new_edges, &overall_size, true); | |
1546 lookup_recursive_calls (node, curr->callee, &heap); | |
0 | 1547 n++; |
1548 } | |
111 | 1549 |
1550 if (!heap.empty () && dump_file) | |
0 | 1551 fprintf (dump_file, " Recursive inlining growth limit met.\n"); |
1552 | |
111 | 1553 if (!master_clone) |
1554 return false; | |
1555 | |
0 | 1556 if (dump_file) |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1557 fprintf (dump_file, |
111 | 1558 "\n Inlined %i times, " |
1559 "body grown from size %i to %i, time %f to %f\n", n, | |
1560 ipa_fn_summaries->get (master_clone)->size, | |
1561 ipa_fn_summaries->get (node)->size, | |
1562 ipa_fn_summaries->get (master_clone)->time.to_double (), | |
1563 ipa_fn_summaries->get (node)->time.to_double ()); | |
0 | 1564 |
1565 /* Remove master clone we used for inlining. We rely that clones inlined | |
1566 into master clone gets queued just before master clone so we don't | |
1567 need recursion. */ | |
111 | 1568 for (node = symtab->first_function (); node != master_clone; |
0 | 1569 node = next) |
1570 { | |
111 | 1571 next = symtab->next_function (node); |
0 | 1572 if (node->global.inlined_to == master_clone) |
111 | 1573 node->remove (); |
0 | 1574 } |
111 | 1575 master_clone->remove (); |
1576 return true; | |
0 | 1577 } |
1578 | |
1579 | |
1580 /* Given whole compilation unit estimate of INSNS, compute how large we can | |
1581 allow the unit to grow. */ | |
111 | 1582 |
0 | 1583 static int |
1584 compute_max_insns (int insns) | |
1585 { | |
1586 int max_insns = insns; | |
1587 if (max_insns < PARAM_VALUE (PARAM_LARGE_UNIT_INSNS)) | |
1588 max_insns = PARAM_VALUE (PARAM_LARGE_UNIT_INSNS); | |
1589 | |
111 | 1590 return ((int64_t) max_insns |
0 | 1591 * (100 + PARAM_VALUE (PARAM_INLINE_UNIT_GROWTH)) / 100); |
1592 } | |
1593 | |
111 | 1594 |
0 | 1595 /* Compute badness of all edges in NEW_EDGES and add them to the HEAP. */ |
111 | 1596 |
0 | 1597 static void |
111 | 1598 add_new_edges_to_heap (edge_heap_t *heap, vec<cgraph_edge *> new_edges) |
0 | 1599 { |
111 | 1600 while (new_edges.length () > 0) |
0 | 1601 { |
111 | 1602 struct cgraph_edge *edge = new_edges.pop (); |
0 | 1603 |
1604 gcc_assert (!edge->aux); | |
111 | 1605 if (edge->inline_failed |
1606 && can_inline_edge_p (edge, true) | |
1607 && want_inline_small_function_p (edge, true)) | |
1608 edge->aux = heap->insert (edge_badness (edge, false), edge); | |
1609 } | |
1610 } | |
1611 | |
1612 /* Remove EDGE from the fibheap. */ | |
1613 | |
1614 static void | |
1615 heap_edge_removal_hook (struct cgraph_edge *e, void *data) | |
1616 { | |
1617 if (e->aux) | |
1618 { | |
1619 ((edge_heap_t *)data)->delete_node ((edge_heap_node_t *)e->aux); | |
1620 e->aux = NULL; | |
0 | 1621 } |
1622 } | |
1623 | |
111 | 1624 /* Return true if speculation of edge E seems useful. |
1625 If ANTICIPATE_INLINING is true, be conservative and hope that E | |
1626 may get inlined. */ | |
1627 | |
1628 bool | |
1629 speculation_useful_p (struct cgraph_edge *e, bool anticipate_inlining) | |
1630 { | |
1631 enum availability avail; | |
1632 struct cgraph_node *target = e->callee->ultimate_alias_target (&avail, | |
1633 e->caller); | |
1634 struct cgraph_edge *direct, *indirect; | |
1635 struct ipa_ref *ref; | |
1636 | |
1637 gcc_assert (e->speculative && !e->indirect_unknown_callee); | |
1638 | |
1639 if (!e->maybe_hot_p ()) | |
1640 return false; | |
1641 | |
1642 /* See if IP optimizations found something potentially useful about the | |
1643 function. For now we look only for CONST/PURE flags. Almost everything | |
1644 else we propagate is useless. */ | |
1645 if (avail >= AVAIL_AVAILABLE) | |
1646 { | |
1647 int ecf_flags = flags_from_decl_or_type (target->decl); | |
1648 if (ecf_flags & ECF_CONST) | |
1649 { | |
1650 e->speculative_call_info (direct, indirect, ref); | |
1651 if (!(indirect->indirect_info->ecf_flags & ECF_CONST)) | |
1652 return true; | |
1653 } | |
1654 else if (ecf_flags & ECF_PURE) | |
1655 { | |
1656 e->speculative_call_info (direct, indirect, ref); | |
1657 if (!(indirect->indirect_info->ecf_flags & ECF_PURE)) | |
1658 return true; | |
1659 } | |
1660 } | |
1661 /* If we did not managed to inline the function nor redirect | |
1662 to an ipa-cp clone (that are seen by having local flag set), | |
1663 it is probably pointless to inline it unless hardware is missing | |
1664 indirect call predictor. */ | |
1665 if (!anticipate_inlining && e->inline_failed && !target->local.local) | |
1666 return false; | |
1667 /* For overwritable targets there is not much to do. */ | |
1668 if (e->inline_failed && !can_inline_edge_p (e, false, true)) | |
1669 return false; | |
1670 /* OK, speculation seems interesting. */ | |
1671 return true; | |
1672 } | |
1673 | |
1674 /* We know that EDGE is not going to be inlined. | |
1675 See if we can remove speculation. */ | |
1676 | |
1677 static void | |
1678 resolve_noninline_speculation (edge_heap_t *edge_heap, struct cgraph_edge *edge) | |
1679 { | |
1680 if (edge->speculative && !speculation_useful_p (edge, false)) | |
1681 { | |
1682 struct cgraph_node *node = edge->caller; | |
1683 struct cgraph_node *where = node->global.inlined_to | |
1684 ? node->global.inlined_to : node; | |
1685 auto_bitmap updated_nodes; | |
1686 | |
1687 spec_rem += edge->count; | |
1688 edge->resolve_speculation (); | |
1689 reset_edge_caches (where); | |
1690 ipa_update_overall_fn_summary (where); | |
1691 update_caller_keys (edge_heap, where, | |
1692 updated_nodes, NULL); | |
1693 update_callee_keys (edge_heap, where, | |
1694 updated_nodes); | |
1695 } | |
1696 } | |
1697 | |
1698 /* Return true if NODE should be accounted for overall size estimate. | |
1699 Skip all nodes optimized for size so we can measure the growth of hot | |
1700 part of program no matter of the padding. */ | |
1701 | |
1702 bool | |
1703 inline_account_function_p (struct cgraph_node *node) | |
1704 { | |
1705 return (!DECL_EXTERNAL (node->decl) | |
1706 && !opt_for_fn (node->decl, optimize_size) | |
1707 && node->frequency != NODE_FREQUENCY_UNLIKELY_EXECUTED); | |
1708 } | |
1709 | |
1710 /* Count number of callers of NODE and store it into DATA (that | |
1711 points to int. Worker for cgraph_for_node_and_aliases. */ | |
1712 | |
1713 static bool | |
1714 sum_callers (struct cgraph_node *node, void *data) | |
1715 { | |
1716 struct cgraph_edge *e; | |
1717 int *num_calls = (int *)data; | |
1718 | |
1719 for (e = node->callers; e; e = e->next_caller) | |
1720 (*num_calls)++; | |
1721 return false; | |
1722 } | |
0 | 1723 |
1724 /* We use greedy algorithm for inlining of small functions: | |
111 | 1725 All inline candidates are put into prioritized heap ordered in |
1726 increasing badness. | |
0 | 1727 |
111 | 1728 The inlining of small functions is bounded by unit growth parameters. */ |
0 | 1729 |
1730 static void | |
111 | 1731 inline_small_functions (void) |
0 | 1732 { |
1733 struct cgraph_node *node; | |
1734 struct cgraph_edge *edge; | |
111 | 1735 edge_heap_t edge_heap (sreal::min ()); |
1736 auto_bitmap updated_nodes; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1737 int min_size, max_size; |
111 | 1738 auto_vec<cgraph_edge *> new_indirect_edges; |
1739 int initial_size = 0; | |
1740 struct cgraph_node **order = XCNEWVEC (cgraph_node *, symtab->cgraph_count); | |
1741 struct cgraph_edge_hook_list *edge_removal_hook_holder; | |
1742 new_indirect_edges.create (8); | |
1743 | |
1744 edge_removal_hook_holder | |
1745 = symtab->add_edge_removal_hook (&heap_edge_removal_hook, &edge_heap); | |
1746 | |
1747 /* Compute overall unit size and other global parameters used by badness | |
1748 metrics. */ | |
1749 | |
1750 max_count = profile_count::uninitialized (); | |
1751 ipa_reduced_postorder (order, true, true, NULL); | |
1752 free (order); | |
1753 | |
1754 FOR_EACH_DEFINED_FUNCTION (node) | |
1755 if (!node->global.inlined_to) | |
1756 { | |
1757 if (!node->alias && node->analyzed | |
1758 && (node->has_gimple_body_p () || node->thunk.thunk_p) | |
1759 && opt_for_fn (node->decl, optimize)) | |
1760 { | |
1761 struct ipa_fn_summary *info = ipa_fn_summaries->get (node); | |
1762 struct ipa_dfs_info *dfs = (struct ipa_dfs_info *) node->aux; | |
0 | 1763 |
111 | 1764 /* Do not account external functions, they will be optimized out |
1765 if not inlined. Also only count the non-cold portion of program. */ | |
1766 if (inline_account_function_p (node)) | |
1767 initial_size += info->size; | |
1768 info->growth = estimate_growth (node); | |
1769 | |
1770 int num_calls = 0; | |
1771 node->call_for_symbol_and_aliases (sum_callers, &num_calls, | |
1772 true); | |
1773 if (num_calls == 1) | |
1774 info->single_caller = true; | |
1775 if (dfs && dfs->next_cycle) | |
1776 { | |
1777 struct cgraph_node *n2; | |
1778 int id = dfs->scc_no + 1; | |
1779 for (n2 = node; n2; | |
1780 n2 = ((struct ipa_dfs_info *) node->aux)->next_cycle) | |
1781 if (opt_for_fn (n2->decl, optimize)) | |
1782 { | |
1783 struct ipa_fn_summary *info2 = ipa_fn_summaries->get (n2); | |
1784 if (info2->scc_no) | |
1785 break; | |
1786 info2->scc_no = id; | |
1787 } | |
1788 } | |
1789 } | |
1790 | |
1791 for (edge = node->callers; edge; edge = edge->next_caller) | |
1792 if (!(max_count >= edge->count)) | |
1793 max_count = edge->count; | |
1794 } | |
1795 ipa_free_postorder_info (); | |
1796 initialize_growth_caches (); | |
0 | 1797 |
1798 if (dump_file) | |
111 | 1799 fprintf (dump_file, |
1800 "\nDeciding on inlining of small functions. Starting with size %i.\n", | |
1801 initial_size); | |
0 | 1802 |
111 | 1803 overall_size = initial_size; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1804 max_size = compute_max_insns (overall_size); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1805 min_size = overall_size; |
0 | 1806 |
111 | 1807 /* Populate the heap with all edges we might inline. */ |
1808 | |
1809 FOR_EACH_DEFINED_FUNCTION (node) | |
1810 { | |
1811 bool update = false; | |
1812 struct cgraph_edge *next = NULL; | |
1813 bool has_speculative = false; | |
1814 | |
1815 if (!opt_for_fn (node->decl, optimize)) | |
1816 continue; | |
1817 | |
1818 if (dump_file) | |
1819 fprintf (dump_file, "Enqueueing calls in %s.\n", node->dump_name ()); | |
1820 | |
1821 for (edge = node->callees; edge; edge = next) | |
1822 { | |
1823 next = edge->next_callee; | |
1824 if (edge->inline_failed | |
1825 && !edge->aux | |
1826 && can_inline_edge_p (edge, true) | |
1827 && want_inline_small_function_p (edge, true) | |
1828 && edge->inline_failed) | |
1829 { | |
1830 gcc_assert (!edge->aux); | |
1831 update_edge_key (&edge_heap, edge); | |
1832 } | |
1833 if (edge->speculative) | |
1834 has_speculative = true; | |
1835 } | |
1836 if (has_speculative) | |
1837 for (edge = node->callees; edge; edge = next) | |
1838 if (edge->speculative && !speculation_useful_p (edge, | |
1839 edge->aux != NULL)) | |
1840 { | |
1841 edge->resolve_speculation (); | |
1842 update = true; | |
1843 } | |
1844 if (update) | |
1845 { | |
1846 struct cgraph_node *where = node->global.inlined_to | |
1847 ? node->global.inlined_to : node; | |
1848 ipa_update_overall_fn_summary (where); | |
1849 reset_edge_caches (where); | |
1850 update_caller_keys (&edge_heap, where, | |
1851 updated_nodes, NULL); | |
1852 update_callee_keys (&edge_heap, where, | |
1853 updated_nodes); | |
1854 bitmap_clear (updated_nodes); | |
1855 } | |
1856 } | |
1857 | |
1858 gcc_assert (in_lto_p | |
1859 || !(max_count > 0) | |
1860 || (profile_info && flag_branch_probabilities)); | |
1861 | |
1862 while (!edge_heap.empty ()) | |
0 | 1863 { |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1864 int old_size = overall_size; |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1865 struct cgraph_node *where, *callee; |
111 | 1866 sreal badness = edge_heap.min_key (); |
1867 sreal current_badness; | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1868 int growth; |
0 | 1869 |
111 | 1870 edge = edge_heap.extract_min (); |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1871 gcc_assert (edge->aux); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1872 edge->aux = NULL; |
111 | 1873 if (!edge->inline_failed || !edge->callee->analyzed) |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1874 continue; |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1875 |
111 | 1876 #if CHECKING_P |
1877 /* Be sure that caches are maintained consistent. */ | |
1878 sreal cached_badness = edge_badness (edge, false); | |
1879 | |
1880 int old_size_est = estimate_edge_size (edge); | |
1881 sreal old_time_est = estimate_edge_time (edge); | |
1882 int old_hints_est = estimate_edge_hints (edge); | |
1883 | |
1884 reset_edge_growth_cache (edge); | |
1885 gcc_assert (old_size_est == estimate_edge_size (edge)); | |
1886 gcc_assert (old_time_est == estimate_edge_time (edge)); | |
1887 /* FIXME: | |
1888 | |
1889 gcc_assert (old_hints_est == estimate_edge_hints (edge)); | |
1890 | |
1891 fails with profile feedback because some hints depends on | |
1892 maybe_hot_edge_p predicate and because callee gets inlined to other | |
1893 calls, the edge may become cold. | |
1894 This ought to be fixed by computing relative probabilities | |
1895 for given invocation but that will be better done once whole | |
1896 code is converted to sreals. Disable for now and revert to "wrong" | |
1897 value so enable/disable checking paths agree. */ | |
1898 edge_growth_cache[edge->uid].hints = old_hints_est + 1; | |
1899 | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1900 /* When updating the edge costs, we only decrease badness in the keys. |
111 | 1901 Increases of badness are handled lazilly; when we see key with out |
1902 of date value on it, we re-insert it now. */ | |
1903 current_badness = edge_badness (edge, false); | |
1904 gcc_assert (cached_badness == current_badness); | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1905 gcc_assert (current_badness >= badness); |
111 | 1906 #else |
1907 current_badness = edge_badness (edge, false); | |
1908 #endif | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1909 if (current_badness != badness) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1910 { |
111 | 1911 if (edge_heap.min () && current_badness > edge_heap.min_key ()) |
1912 { | |
1913 edge->aux = edge_heap.insert (current_badness, edge); | |
1914 continue; | |
1915 } | |
1916 else | |
1917 badness = current_badness; | |
1918 } | |
1919 | |
1920 if (!can_inline_edge_p (edge, true)) | |
1921 { | |
1922 resolve_noninline_speculation (&edge_heap, edge); | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1923 continue; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1924 } |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1925 |
111 | 1926 callee = edge->callee->ultimate_alias_target (); |
1927 growth = estimate_edge_growth (edge); | |
0 | 1928 if (dump_file) |
1929 { | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1930 fprintf (dump_file, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1931 "\nConsidering %s with %i size\n", |
111 | 1932 callee->dump_name (), |
1933 ipa_fn_summaries->get (callee)->size); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1934 fprintf (dump_file, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1935 " to be inlined into %s in %s:%i\n" |
111 | 1936 " Estimated badness is %f, frequency %.2f.\n", |
1937 edge->caller->dump_name (), | |
1938 edge->call_stmt | |
1939 && (LOCATION_LOCUS (gimple_location ((const gimple *) | |
1940 edge->call_stmt)) | |
1941 > BUILTINS_LOCATION) | |
1942 ? gimple_filename ((const gimple *) edge->call_stmt) | |
1943 : "unknown", | |
1944 edge->call_stmt | |
1945 ? gimple_lineno ((const gimple *) edge->call_stmt) | |
1946 : -1, | |
1947 badness.to_double (), | |
0 | 1948 edge->frequency / (double)CGRAPH_FREQ_BASE); |
111 | 1949 if (edge->count.initialized_p ()) |
1950 { | |
1951 fprintf (dump_file, " Called "); | |
1952 edge->count.dump (dump_file); | |
1953 fprintf (dump_file, "times\n"); | |
1954 } | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1955 if (dump_flags & TDF_DETAILS) |
111 | 1956 edge_badness (edge, true); |
0 | 1957 } |
1958 | |
111 | 1959 if (overall_size + growth > max_size |
1960 && !DECL_DISREGARD_INLINE_LIMITS (callee->decl)) | |
0 | 1961 { |
111 | 1962 edge->inline_failed = CIF_INLINE_UNIT_GROWTH_LIMIT; |
1963 report_inline_failed_reason (edge); | |
1964 resolve_noninline_speculation (&edge_heap, edge); | |
1965 continue; | |
0 | 1966 } |
1967 | |
111 | 1968 if (!want_inline_small_function_p (edge, true)) |
0 | 1969 { |
111 | 1970 resolve_noninline_speculation (&edge_heap, edge); |
0 | 1971 continue; |
1972 } | |
111 | 1973 |
1974 /* Heuristics for inlining small functions work poorly for | |
1975 recursive calls where we do effects similar to loop unrolling. | |
1976 When inlining such edge seems profitable, leave decision on | |
1977 specific inliner. */ | |
1978 if (edge->recursive_p ()) | |
0 | 1979 { |
1980 where = edge->caller; | |
1981 if (where->global.inlined_to) | |
1982 where = where->global.inlined_to; | |
111 | 1983 if (!recursive_inlining (edge, |
1984 opt_for_fn (edge->caller->decl, | |
1985 flag_indirect_inlining) | |
1986 ? &new_indirect_edges : NULL)) | |
1987 { | |
1988 edge->inline_failed = CIF_RECURSIVE_INLINING; | |
1989 resolve_noninline_speculation (&edge_heap, edge); | |
1990 continue; | |
1991 } | |
1992 reset_edge_caches (where); | |
1993 /* Recursive inliner inlines all recursive calls of the function | |
1994 at once. Consequently we need to update all callee keys. */ | |
1995 if (opt_for_fn (edge->caller->decl, flag_indirect_inlining)) | |
1996 add_new_edges_to_heap (&edge_heap, new_indirect_edges); | |
1997 update_callee_keys (&edge_heap, where, updated_nodes); | |
1998 bitmap_clear (updated_nodes); | |
0 | 1999 } |
2000 else | |
2001 { | |
111 | 2002 struct cgraph_node *outer_node = NULL; |
2003 int depth = 0; | |
2004 | |
2005 /* Consider the case where self recursive function A is inlined | |
2006 into B. This is desired optimization in some cases, since it | |
2007 leads to effect similar of loop peeling and we might completely | |
2008 optimize out the recursive call. However we must be extra | |
2009 selective. */ | |
2010 | |
2011 where = edge->caller; | |
2012 while (where->global.inlined_to) | |
0 | 2013 { |
111 | 2014 if (where->decl == callee->decl) |
2015 outer_node = where, depth++; | |
2016 where = where->callers->caller; | |
2017 } | |
2018 if (outer_node | |
2019 && !want_inline_self_recursive_call_p (edge, outer_node, | |
2020 true, depth)) | |
2021 { | |
2022 edge->inline_failed | |
2023 = (DECL_DISREGARD_INLINE_LIMITS (edge->callee->decl) | |
2024 ? CIF_RECURSIVE_INLINING : CIF_UNSPECIFIED); | |
2025 resolve_noninline_speculation (&edge_heap, edge); | |
0 | 2026 continue; |
2027 } | |
111 | 2028 else if (depth && dump_file) |
2029 fprintf (dump_file, " Peeling recursion with depth %i\n", depth); | |
0 | 2030 |
111 | 2031 gcc_checking_assert (!callee->global.inlined_to); |
2032 inline_call (edge, true, &new_indirect_edges, &overall_size, true); | |
2033 add_new_edges_to_heap (&edge_heap, new_indirect_edges); | |
2034 | |
2035 reset_edge_caches (edge->callee); | |
2036 | |
2037 update_callee_keys (&edge_heap, where, updated_nodes); | |
0 | 2038 } |
2039 where = edge->caller; | |
2040 if (where->global.inlined_to) | |
2041 where = where->global.inlined_to; | |
2042 | |
2043 /* Our profitability metric can depend on local properties | |
2044 such as number of inlinable calls and size of the function body. | |
2045 After inlining these properties might change for the function we | |
2046 inlined into (since it's body size changed) and for the functions | |
2047 called by function we inlined (since number of it inlinable callers | |
2048 might change). */ | |
111 | 2049 update_caller_keys (&edge_heap, where, updated_nodes, NULL); |
2050 /* Offline copy count has possibly changed, recompute if profile is | |
2051 available. */ | |
2052 if (max_count > profile_count::zero ()) | |
2053 { | |
2054 struct cgraph_node *n = cgraph_node::get (edge->callee->decl); | |
2055 if (n != edge->callee && n->analyzed) | |
2056 update_callee_keys (&edge_heap, n, updated_nodes); | |
2057 } | |
0 | 2058 bitmap_clear (updated_nodes); |
2059 | |
2060 if (dump_file) | |
2061 { | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2062 fprintf (dump_file, |
111 | 2063 " Inlined %s into %s which now has time %f and size %i, " |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2064 "net change of %+i.\n", |
111 | 2065 xstrdup_for_dump (edge->callee->name ()), |
2066 xstrdup_for_dump (edge->caller->name ()), | |
2067 ipa_fn_summaries->get (edge->caller)->time.to_double (), | |
2068 ipa_fn_summaries->get (edge->caller)->size, | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2069 overall_size - old_size); |
0 | 2070 } |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2071 if (min_size > overall_size) |
0 | 2072 { |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2073 min_size = overall_size; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2074 max_size = compute_max_insns (min_size); |
0 | 2075 |
2076 if (dump_file) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2077 fprintf (dump_file, "New minimal size reached: %i\n", min_size); |
0 | 2078 } |
2079 } | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2080 |
111 | 2081 free_growth_caches (); |
2082 if (dump_file) | |
2083 fprintf (dump_file, | |
2084 "Unit growth for small function inlining: %i->%i (%i%%)\n", | |
2085 initial_size, overall_size, | |
2086 initial_size ? overall_size * 100 / (initial_size) - 100: 0); | |
2087 symtab->remove_edge_removal_hook (edge_removal_hook_holder); | |
0 | 2088 } |
2089 | |
111 | 2090 /* Flatten NODE. Performed both during early inlining and |
2091 at IPA inlining time. */ | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2092 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2093 static void |
111 | 2094 flatten_function (struct cgraph_node *node, bool early) |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2095 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2096 struct cgraph_edge *e; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2097 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2098 /* We shouldn't be called recursively when we are being processed. */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2099 gcc_assert (node->aux == NULL); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2100 |
111 | 2101 node->aux = (void *) node; |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2102 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2103 for (e = node->callees; e; e = e->next_callee) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2104 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2105 struct cgraph_node *orig_callee; |
111 | 2106 struct cgraph_node *callee = e->callee->ultimate_alias_target (); |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2107 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2108 /* We've hit cycle? It is time to give up. */ |
111 | 2109 if (callee->aux) |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2110 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2111 if (dump_file) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2112 fprintf (dump_file, |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2113 "Not inlining %s into %s to avoid cycle.\n", |
111 | 2114 xstrdup_for_dump (callee->name ()), |
2115 xstrdup_for_dump (e->caller->name ())); | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2116 e->inline_failed = CIF_RECURSIVE_INLINING; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2117 continue; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2118 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2119 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2120 /* When the edge is already inlined, we just need to recurse into |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2121 it in order to fully flatten the leaves. */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2122 if (!e->inline_failed) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2123 { |
111 | 2124 flatten_function (callee, early); |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2125 continue; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2126 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2127 |
111 | 2128 /* Flatten attribute needs to be processed during late inlining. For |
2129 extra code quality we however do flattening during early optimization, | |
2130 too. */ | |
2131 if (!early | |
2132 ? !can_inline_edge_p (e, true) | |
2133 : !can_early_inline_edge_p (e)) | |
2134 continue; | |
2135 | |
2136 if (e->recursive_p ()) | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2137 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2138 if (dump_file) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2139 fprintf (dump_file, "Not inlining: recursive call.\n"); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2140 continue; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2141 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2142 |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
2143 if (gimple_in_ssa_p (DECL_STRUCT_FUNCTION (node->decl)) |
111 | 2144 != gimple_in_ssa_p (DECL_STRUCT_FUNCTION (callee->decl))) |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
2145 { |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
2146 if (dump_file) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
2147 fprintf (dump_file, "Not inlining: SSA form does not match.\n"); |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
2148 continue; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
2149 } |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
2150 |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2151 /* Inline the edge and flatten the inline clone. Avoid |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2152 recursing through the original node if the node was cloned. */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2153 if (dump_file) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2154 fprintf (dump_file, " Inlining %s into %s.\n", |
111 | 2155 xstrdup_for_dump (callee->name ()), |
2156 xstrdup_for_dump (e->caller->name ())); | |
2157 orig_callee = callee; | |
2158 inline_call (e, true, NULL, NULL, false); | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2159 if (e->callee != orig_callee) |
111 | 2160 orig_callee->aux = (void *) node; |
2161 flatten_function (e->callee, early); | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2162 if (e->callee != orig_callee) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2163 orig_callee->aux = NULL; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2164 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2165 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2166 node->aux = NULL; |
111 | 2167 if (!node->global.inlined_to) |
2168 ipa_update_overall_fn_summary (node); | |
2169 } | |
2170 | |
2171 /* Inline NODE to all callers. Worker for cgraph_for_node_and_aliases. | |
2172 DATA points to number of calls originally found so we avoid infinite | |
2173 recursion. */ | |
2174 | |
2175 static bool | |
2176 inline_to_all_callers_1 (struct cgraph_node *node, void *data, | |
2177 hash_set<cgraph_node *> *callers) | |
2178 { | |
2179 int *num_calls = (int *)data; | |
2180 bool callee_removed = false; | |
2181 | |
2182 while (node->callers && !node->global.inlined_to) | |
2183 { | |
2184 struct cgraph_node *caller = node->callers->caller; | |
2185 | |
2186 if (!can_inline_edge_p (node->callers, true) | |
2187 || node->callers->recursive_p ()) | |
2188 { | |
2189 if (dump_file) | |
2190 fprintf (dump_file, "Uninlinable call found; giving up.\n"); | |
2191 *num_calls = 0; | |
2192 return false; | |
2193 } | |
2194 | |
2195 if (dump_file) | |
2196 { | |
2197 fprintf (dump_file, | |
2198 "\nInlining %s size %i.\n", | |
2199 node->name (), | |
2200 ipa_fn_summaries->get (node)->size); | |
2201 fprintf (dump_file, | |
2202 " Called once from %s %i insns.\n", | |
2203 node->callers->caller->name (), | |
2204 ipa_fn_summaries->get (node->callers->caller)->size); | |
2205 } | |
2206 | |
2207 /* Remember which callers we inlined to, delaying updating the | |
2208 overall summary. */ | |
2209 callers->add (node->callers->caller); | |
2210 inline_call (node->callers, true, NULL, NULL, false, &callee_removed); | |
2211 if (dump_file) | |
2212 fprintf (dump_file, | |
2213 " Inlined into %s which now has %i size\n", | |
2214 caller->name (), | |
2215 ipa_fn_summaries->get (caller)->size); | |
2216 if (!(*num_calls)--) | |
2217 { | |
2218 if (dump_file) | |
2219 fprintf (dump_file, "New calls found; giving up.\n"); | |
2220 return callee_removed; | |
2221 } | |
2222 if (callee_removed) | |
2223 return true; | |
2224 } | |
2225 return false; | |
2226 } | |
2227 | |
2228 /* Wrapper around inline_to_all_callers_1 doing delayed overall summary | |
2229 update. */ | |
2230 | |
2231 static bool | |
2232 inline_to_all_callers (struct cgraph_node *node, void *data) | |
2233 { | |
2234 hash_set<cgraph_node *> callers; | |
2235 bool res = inline_to_all_callers_1 (node, data, &callers); | |
2236 /* Perform the delayed update of the overall summary of all callers | |
2237 processed. This avoids quadratic behavior in the cases where | |
2238 we have a lot of calls to the same function. */ | |
2239 for (hash_set<cgraph_node *>::iterator i = callers.begin (); | |
2240 i != callers.end (); ++i) | |
2241 ipa_update_overall_fn_summary (*i); | |
2242 return res; | |
2243 } | |
2244 | |
2245 /* Output overall time estimate. */ | |
2246 static void | |
2247 dump_overall_stats (void) | |
2248 { | |
2249 sreal sum_weighted = 0, sum = 0; | |
2250 struct cgraph_node *node; | |
2251 | |
2252 FOR_EACH_DEFINED_FUNCTION (node) | |
2253 if (!node->global.inlined_to | |
2254 && !node->alias) | |
2255 { | |
2256 sreal time = ipa_fn_summaries->get (node)->time; | |
2257 sum += time; | |
2258 if (node->count.initialized_p ()) | |
2259 sum_weighted += time * node->count.to_gcov_type (); | |
2260 } | |
2261 fprintf (dump_file, "Overall time estimate: " | |
2262 "%f weighted by profile: " | |
2263 "%f\n", sum.to_double (), sum_weighted.to_double ()); | |
2264 } | |
2265 | |
2266 /* Output some useful stats about inlining. */ | |
2267 | |
2268 static void | |
2269 dump_inline_stats (void) | |
2270 { | |
2271 int64_t inlined_cnt = 0, inlined_indir_cnt = 0; | |
2272 int64_t inlined_virt_cnt = 0, inlined_virt_indir_cnt = 0; | |
2273 int64_t noninlined_cnt = 0, noninlined_indir_cnt = 0; | |
2274 int64_t noninlined_virt_cnt = 0, noninlined_virt_indir_cnt = 0; | |
2275 int64_t inlined_speculative = 0, inlined_speculative_ply = 0; | |
2276 int64_t indirect_poly_cnt = 0, indirect_cnt = 0; | |
2277 int64_t reason[CIF_N_REASONS][3]; | |
2278 int i; | |
2279 struct cgraph_node *node; | |
2280 | |
2281 memset (reason, 0, sizeof (reason)); | |
2282 FOR_EACH_DEFINED_FUNCTION (node) | |
2283 { | |
2284 struct cgraph_edge *e; | |
2285 for (e = node->callees; e; e = e->next_callee) | |
2286 { | |
2287 if (e->inline_failed) | |
2288 { | |
2289 if (e->count.initialized_p ()) | |
2290 reason[(int) e->inline_failed][0] += e->count.to_gcov_type (); | |
2291 reason[(int) e->inline_failed][1] += e->frequency; | |
2292 reason[(int) e->inline_failed][2] ++; | |
2293 if (DECL_VIRTUAL_P (e->callee->decl) | |
2294 && e->count.initialized_p ()) | |
2295 { | |
2296 if (e->indirect_inlining_edge) | |
2297 noninlined_virt_indir_cnt += e->count.to_gcov_type (); | |
2298 else | |
2299 noninlined_virt_cnt += e->count.to_gcov_type (); | |
2300 } | |
2301 else if (e->count.initialized_p ()) | |
2302 { | |
2303 if (e->indirect_inlining_edge) | |
2304 noninlined_indir_cnt += e->count.to_gcov_type (); | |
2305 else | |
2306 noninlined_cnt += e->count.to_gcov_type (); | |
2307 } | |
2308 } | |
2309 else if (e->count.initialized_p ()) | |
2310 { | |
2311 if (e->speculative) | |
2312 { | |
2313 if (DECL_VIRTUAL_P (e->callee->decl)) | |
2314 inlined_speculative_ply += e->count.to_gcov_type (); | |
2315 else | |
2316 inlined_speculative += e->count.to_gcov_type (); | |
2317 } | |
2318 else if (DECL_VIRTUAL_P (e->callee->decl)) | |
2319 { | |
2320 if (e->indirect_inlining_edge) | |
2321 inlined_virt_indir_cnt += e->count.to_gcov_type (); | |
2322 else | |
2323 inlined_virt_cnt += e->count.to_gcov_type (); | |
2324 } | |
2325 else | |
2326 { | |
2327 if (e->indirect_inlining_edge) | |
2328 inlined_indir_cnt += e->count.to_gcov_type (); | |
2329 else | |
2330 inlined_cnt += e->count.to_gcov_type (); | |
2331 } | |
2332 } | |
2333 } | |
2334 for (e = node->indirect_calls; e; e = e->next_callee) | |
2335 if (e->indirect_info->polymorphic | |
2336 & e->count.initialized_p ()) | |
2337 indirect_poly_cnt += e->count.to_gcov_type (); | |
2338 else if (e->count.initialized_p ()) | |
2339 indirect_cnt += e->count.to_gcov_type (); | |
2340 } | |
2341 if (max_count.initialized_p ()) | |
2342 { | |
2343 fprintf (dump_file, | |
2344 "Inlined %" PRId64 " + speculative " | |
2345 "%" PRId64 " + speculative polymorphic " | |
2346 "%" PRId64 " + previously indirect " | |
2347 "%" PRId64 " + virtual " | |
2348 "%" PRId64 " + virtual and previously indirect " | |
2349 "%" PRId64 "\n" "Not inlined " | |
2350 "%" PRId64 " + previously indirect " | |
2351 "%" PRId64 " + virtual " | |
2352 "%" PRId64 " + virtual and previously indirect " | |
2353 "%" PRId64 " + stil indirect " | |
2354 "%" PRId64 " + still indirect polymorphic " | |
2355 "%" PRId64 "\n", inlined_cnt, | |
2356 inlined_speculative, inlined_speculative_ply, | |
2357 inlined_indir_cnt, inlined_virt_cnt, inlined_virt_indir_cnt, | |
2358 noninlined_cnt, noninlined_indir_cnt, noninlined_virt_cnt, | |
2359 noninlined_virt_indir_cnt, indirect_cnt, indirect_poly_cnt); | |
2360 fprintf (dump_file, "Removed speculations "); | |
2361 spec_rem.dump (dump_file); | |
2362 fprintf (dump_file, "\n"); | |
2363 } | |
2364 dump_overall_stats (); | |
2365 fprintf (dump_file, "\nWhy inlining failed?\n"); | |
2366 for (i = 0; i < CIF_N_REASONS; i++) | |
2367 if (reason[i][2]) | |
2368 fprintf (dump_file, "%-50s: %8i calls, %8i freq, %" PRId64" count\n", | |
2369 cgraph_inline_failed_string ((cgraph_inline_failed_t) i), | |
2370 (int) reason[i][2], (int) reason[i][1], reason[i][0]); | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2371 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2372 |
0 | 2373 /* Decide on the inlining. We do so in the topological order to avoid |
2374 expenses on updating data structures. */ | |
2375 | |
2376 static unsigned int | |
111 | 2377 ipa_inline (void) |
0 | 2378 { |
2379 struct cgraph_node *node; | |
2380 int nnodes; | |
111 | 2381 struct cgraph_node **order; |
0 | 2382 int i; |
111 | 2383 int cold; |
2384 bool remove_functions = false; | |
0 | 2385 |
111 | 2386 cgraph_freq_base_rec = (sreal) 1 / (sreal) CGRAPH_FREQ_BASE; |
2387 percent_rec = (sreal) 1 / (sreal) 100; | |
0 | 2388 |
111 | 2389 order = XCNEWVEC (struct cgraph_node *, symtab->cgraph_count); |
0 | 2390 |
2391 if (dump_file) | |
111 | 2392 ipa_dump_fn_summaries (dump_file); |
2393 | |
2394 nnodes = ipa_reverse_postorder (order); | |
2395 | |
2396 FOR_EACH_FUNCTION (node) | |
2397 { | |
2398 node->aux = 0; | |
0 | 2399 |
111 | 2400 /* Recompute the default reasons for inlining because they may have |
2401 changed during merging. */ | |
2402 if (in_lto_p) | |
2403 { | |
2404 for (cgraph_edge *e = node->callees; e; e = e->next_callee) | |
2405 { | |
2406 gcc_assert (e->inline_failed); | |
2407 initialize_inline_failed (e); | |
2408 } | |
2409 for (cgraph_edge *e = node->indirect_calls; e; e = e->next_callee) | |
2410 initialize_inline_failed (e); | |
2411 } | |
2412 } | |
0 | 2413 |
2414 if (dump_file) | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2415 fprintf (dump_file, "\nFlattening functions:\n"); |
0 | 2416 |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2417 /* In the first pass handle functions to be flattened. Do this with |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2418 a priority so none of our later choices will make this impossible. */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2419 for (i = nnodes - 1; i >= 0; i--) |
0 | 2420 { |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2421 node = order[i]; |
0 | 2422 |
111 | 2423 /* Handle nodes to be flattened. |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2424 Ideally when processing callees we stop inlining at the |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2425 entry of cycles, possibly cloning that entry point and |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2426 try to flatten itself turning it into a self-recursive |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2427 function. */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2428 if (lookup_attribute ("flatten", |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2429 DECL_ATTRIBUTES (node->decl)) != NULL) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2430 { |
0 | 2431 if (dump_file) |
2432 fprintf (dump_file, | |
111 | 2433 "Flattening %s\n", node->name ()); |
2434 flatten_function (node, false); | |
0 | 2435 } |
2436 } | |
111 | 2437 if (dump_file) |
2438 dump_overall_stats (); | |
0 | 2439 |
111 | 2440 inline_small_functions (); |
0 | 2441 |
111 | 2442 gcc_assert (symtab->state == IPA_SSA); |
2443 symtab->state = IPA_SSA_AFTER_INLINING; | |
2444 /* Do first after-inlining removal. We want to remove all "stale" extern | |
2445 inline functions and virtual functions so we really know what is called | |
2446 once. */ | |
2447 symtab->remove_unreachable_nodes (dump_file); | |
2448 free (order); | |
0 | 2449 |
111 | 2450 /* Inline functions with a property that after inlining into all callers the |
2451 code size will shrink because the out-of-line copy is eliminated. | |
2452 We do this regardless on the callee size as long as function growth limits | |
2453 are met. */ | |
2454 if (dump_file) | |
2455 fprintf (dump_file, | |
2456 "\nDeciding on functions to be inlined into all callers and " | |
2457 "removing useless speculations:\n"); | |
0 | 2458 |
111 | 2459 /* Inlining one function called once has good chance of preventing |
2460 inlining other function into the same callee. Ideally we should | |
2461 work in priority order, but probably inlining hot functions first | |
2462 is good cut without the extra pain of maintaining the queue. | |
2463 | |
2464 ??? this is not really fitting the bill perfectly: inlining function | |
2465 into callee often leads to better optimization of callee due to | |
2466 increased context for optimization. | |
2467 For example if main() function calls a function that outputs help | |
2468 and then function that does the main optmization, we should inline | |
2469 the second with priority even if both calls are cold by themselves. | |
2470 | |
2471 We probably want to implement new predicate replacing our use of | |
2472 maybe_hot_edge interpreted as maybe_hot_edge || callee is known | |
2473 to be hot. */ | |
2474 for (cold = 0; cold <= 1; cold ++) | |
2475 { | |
2476 FOR_EACH_DEFINED_FUNCTION (node) | |
2477 { | |
2478 struct cgraph_edge *edge, *next; | |
2479 bool update=false; | |
2480 | |
2481 if (!opt_for_fn (node->decl, optimize) | |
2482 || !opt_for_fn (node->decl, flag_inline_functions_called_once)) | |
2483 continue; | |
2484 | |
2485 for (edge = node->callees; edge; edge = next) | |
0 | 2486 { |
111 | 2487 next = edge->next_callee; |
2488 if (edge->speculative && !speculation_useful_p (edge, false)) | |
0 | 2489 { |
111 | 2490 edge->resolve_speculation (); |
2491 spec_rem += edge->count; | |
2492 update = true; | |
2493 remove_functions = true; | |
0 | 2494 } |
111 | 2495 } |
2496 if (update) | |
2497 { | |
2498 struct cgraph_node *where = node->global.inlined_to | |
2499 ? node->global.inlined_to : node; | |
2500 reset_edge_caches (where); | |
2501 ipa_update_overall_fn_summary (where); | |
2502 } | |
2503 if (want_inline_function_to_all_callers_p (node, cold)) | |
2504 { | |
2505 int num_calls = 0; | |
2506 node->call_for_symbol_and_aliases (sum_callers, &num_calls, | |
2507 true); | |
2508 while (node->call_for_symbol_and_aliases | |
2509 (inline_to_all_callers, &num_calls, true)) | |
2510 ; | |
2511 remove_functions = true; | |
0 | 2512 } |
2513 } | |
2514 } | |
2515 | |
2516 /* Free ipa-prop structures if they are no longer needed. */ | |
111 | 2517 ipa_free_all_structures_after_iinln (); |
2518 | |
2519 if (dump_file) | |
2520 { | |
2521 fprintf (dump_file, | |
2522 "\nInlined %i calls, eliminated %i functions\n\n", | |
2523 ncalls_inlined, nfunctions_inlined); | |
2524 dump_inline_stats (); | |
2525 } | |
0 | 2526 |
2527 if (dump_file) | |
111 | 2528 ipa_dump_fn_summaries (dump_file); |
2529 /* In WPA we use inline summaries for partitioning process. */ | |
2530 if (!flag_wpa) | |
2531 ipa_free_fn_summary (); | |
2532 return remove_functions ? TODO_remove_functions : 0; | |
0 | 2533 } |
2534 | |
111 | 2535 /* Inline always-inline function calls in NODE. */ |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
2536 |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2537 static bool |
111 | 2538 inline_always_inline_functions (struct cgraph_node *node) |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2539 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2540 struct cgraph_edge *e; |
111 | 2541 bool inlined = false; |
2542 | |
2543 for (e = node->callees; e; e = e->next_callee) | |
2544 { | |
2545 struct cgraph_node *callee = e->callee->ultimate_alias_target (); | |
2546 if (!DECL_DISREGARD_INLINE_LIMITS (callee->decl)) | |
2547 continue; | |
2548 | |
2549 if (e->recursive_p ()) | |
2550 { | |
2551 if (dump_file) | |
2552 fprintf (dump_file, " Not inlining recursive call to %s.\n", | |
2553 e->callee->name ()); | |
2554 e->inline_failed = CIF_RECURSIVE_INLINING; | |
2555 continue; | |
2556 } | |
2557 | |
2558 if (!can_early_inline_edge_p (e)) | |
2559 { | |
2560 /* Set inlined to true if the callee is marked "always_inline" but | |
2561 is not inlinable. This will allow flagging an error later in | |
2562 expand_call_inline in tree-inline.c. */ | |
2563 if (lookup_attribute ("always_inline", | |
2564 DECL_ATTRIBUTES (callee->decl)) != NULL) | |
2565 inlined = true; | |
2566 continue; | |
2567 } | |
2568 | |
2569 if (dump_file) | |
2570 fprintf (dump_file, " Inlining %s into %s (always_inline).\n", | |
2571 xstrdup_for_dump (e->callee->name ()), | |
2572 xstrdup_for_dump (e->caller->name ())); | |
2573 inline_call (e, true, NULL, NULL, false); | |
2574 inlined = true; | |
2575 } | |
2576 if (inlined) | |
2577 ipa_update_overall_fn_summary (node); | |
2578 | |
2579 return inlined; | |
0 | 2580 } |
2581 | |
2582 /* Decide on the inlining. We do so in the topological order to avoid | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2583 expenses on updating data structures. */ |
0 | 2584 |
2585 static bool | |
111 | 2586 early_inline_small_functions (struct cgraph_node *node) |
0 | 2587 { |
2588 struct cgraph_edge *e; | |
2589 bool inlined = false; | |
111 | 2590 |
2591 for (e = node->callees; e; e = e->next_callee) | |
2592 { | |
2593 struct cgraph_node *callee = e->callee->ultimate_alias_target (); | |
2594 if (!ipa_fn_summaries->get (callee)->inlinable | |
2595 || !e->inline_failed) | |
2596 continue; | |
0 | 2597 |
111 | 2598 /* Do not consider functions not declared inline. */ |
2599 if (!DECL_DECLARED_INLINE_P (callee->decl) | |
2600 && !opt_for_fn (node->decl, flag_inline_small_functions) | |
2601 && !opt_for_fn (node->decl, flag_inline_functions)) | |
2602 continue; | |
2603 | |
2604 if (dump_file) | |
2605 fprintf (dump_file, "Considering inline candidate %s.\n", | |
2606 callee->name ()); | |
0 | 2607 |
111 | 2608 if (!can_early_inline_edge_p (e)) |
2609 continue; | |
2610 | |
2611 if (e->recursive_p ()) | |
2612 { | |
2613 if (dump_file) | |
2614 fprintf (dump_file, " Not inlining: recursive call.\n"); | |
2615 continue; | |
2616 } | |
2617 | |
2618 if (!want_early_inline_function_p (e)) | |
2619 continue; | |
2620 | |
0 | 2621 if (dump_file) |
111 | 2622 fprintf (dump_file, " Inlining %s into %s.\n", |
2623 xstrdup_for_dump (callee->name ()), | |
2624 xstrdup_for_dump (e->caller->name ())); | |
2625 inline_call (e, true, NULL, NULL, false); | |
2626 inlined = true; | |
0 | 2627 } |
2628 | |
111 | 2629 if (inlined) |
2630 ipa_update_overall_fn_summary (node); | |
0 | 2631 |
2632 return inlined; | |
2633 } | |
2634 | |
111 | 2635 unsigned int |
2636 early_inliner (function *fun) | |
0 | 2637 { |
111 | 2638 struct cgraph_node *node = cgraph_node::get (current_function_decl); |
2639 struct cgraph_edge *edge; | |
0 | 2640 unsigned int todo = 0; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2641 int iterations = 0; |
111 | 2642 bool inlined = false; |
0 | 2643 |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
2644 if (seen_error ()) |
0 | 2645 return 0; |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2646 |
111 | 2647 /* Do nothing if datastructures for ipa-inliner are already computed. This |
2648 happens when some pass decides to construct new function and | |
2649 cgraph_add_new_function calls lowering passes and early optimization on | |
2650 it. This may confuse ourself when early inliner decide to inline call to | |
2651 function clone, because function clones don't have parameter list in | |
2652 ipa-prop matching their signature. */ | |
2653 if (ipa_node_params_sum) | |
2654 return 0; | |
2655 | |
2656 if (flag_checking) | |
2657 node->verify (); | |
2658 node->remove_all_references (); | |
2659 | |
2660 /* Rebuild this reference because it dosn't depend on | |
2661 function's body and it's required to pass cgraph_node | |
2662 verification. */ | |
2663 if (node->instrumented_version | |
2664 && !node->instrumentation_clone) | |
2665 node->create_reference (node->instrumented_version, IPA_REF_CHKP, NULL); | |
2666 | |
2667 /* Even when not optimizing or not inlining inline always-inline | |
2668 functions. */ | |
2669 inlined = inline_always_inline_functions (node); | |
2670 | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2671 if (!optimize |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2672 || flag_no_inline |
111 | 2673 || !flag_early_inlining |
2674 /* Never inline regular functions into always-inline functions | |
2675 during incremental inlining. This sucks as functions calling | |
2676 always inline functions will get less optimized, but at the | |
2677 same time inlining of functions calling always inline | |
2678 function into an always inline function might introduce | |
2679 cycles of edges to be always inlined in the callgraph. | |
2680 | |
2681 We might want to be smarter and just avoid this type of inlining. */ | |
2682 || (DECL_DISREGARD_INLINE_LIMITS (node->decl) | |
2683 && lookup_attribute ("always_inline", | |
2684 DECL_ATTRIBUTES (node->decl)))) | |
2685 ; | |
2686 else if (lookup_attribute ("flatten", | |
2687 DECL_ATTRIBUTES (node->decl)) != NULL) | |
0 | 2688 { |
111 | 2689 /* When the function is marked to be flattened, recursively inline |
2690 all calls in it. */ | |
2691 if (dump_file) | |
2692 fprintf (dump_file, | |
2693 "Flattening %s\n", node->name ()); | |
2694 flatten_function (node, true); | |
2695 inlined = true; | |
0 | 2696 } |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2697 else |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2698 { |
111 | 2699 /* If some always_inline functions was inlined, apply the changes. |
2700 This way we will not account always inline into growth limits and | |
2701 moreover we will inline calls from always inlines that we skipped | |
2702 previously because of conditional above. */ | |
2703 if (inlined) | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2704 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2705 timevar_push (TV_INTEGRATION); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2706 todo |= optimize_inline_calls (current_function_decl); |
111 | 2707 /* optimize_inline_calls call above might have introduced new |
2708 statements that don't have inline parameters computed. */ | |
2709 for (edge = node->callees; edge; edge = edge->next_callee) | |
2710 { | |
2711 struct ipa_call_summary *es = ipa_call_summaries->get (edge); | |
2712 es->call_stmt_size | |
2713 = estimate_num_insns (edge->call_stmt, &eni_size_weights); | |
2714 es->call_stmt_time | |
2715 = estimate_num_insns (edge->call_stmt, &eni_time_weights); | |
2716 } | |
2717 ipa_update_overall_fn_summary (node); | |
2718 inlined = false; | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2719 timevar_pop (TV_INTEGRATION); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2720 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2721 /* We iterate incremental inlining to get trivial cases of indirect |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2722 inlining. */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2723 while (iterations < PARAM_VALUE (PARAM_EARLY_INLINER_MAX_ITERATIONS) |
111 | 2724 && early_inline_small_functions (node)) |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2725 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2726 timevar_push (TV_INTEGRATION); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2727 todo |= optimize_inline_calls (current_function_decl); |
111 | 2728 |
2729 /* Technically we ought to recompute inline parameters so the new | |
2730 iteration of early inliner works as expected. We however have | |
2731 values approximately right and thus we only need to update edge | |
2732 info that might be cleared out for newly discovered edges. */ | |
2733 for (edge = node->callees; edge; edge = edge->next_callee) | |
2734 { | |
2735 /* We have no summary for new bound store calls yet. */ | |
2736 struct ipa_call_summary *es = ipa_call_summaries->get (edge); | |
2737 es->call_stmt_size | |
2738 = estimate_num_insns (edge->call_stmt, &eni_size_weights); | |
2739 es->call_stmt_time | |
2740 = estimate_num_insns (edge->call_stmt, &eni_time_weights); | |
2741 | |
2742 if (edge->callee->decl | |
2743 && !gimple_check_call_matching_types ( | |
2744 edge->call_stmt, edge->callee->decl, false)) | |
2745 { | |
2746 edge->inline_failed = CIF_MISMATCHED_ARGUMENTS; | |
2747 edge->call_stmt_cannot_inline_p = true; | |
2748 } | |
2749 } | |
2750 if (iterations < PARAM_VALUE (PARAM_EARLY_INLINER_MAX_ITERATIONS) - 1) | |
2751 ipa_update_overall_fn_summary (node); | |
2752 timevar_pop (TV_INTEGRATION); | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2753 iterations++; |
111 | 2754 inlined = false; |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2755 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2756 if (dump_file) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2757 fprintf (dump_file, "Iterations: %i\n", iterations); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2758 } |
0 | 2759 |
111 | 2760 if (inlined) |
2761 { | |
2762 timevar_push (TV_INTEGRATION); | |
2763 todo |= optimize_inline_calls (current_function_decl); | |
2764 timevar_pop (TV_INTEGRATION); | |
2765 } | |
2766 | |
2767 fun->always_inline_functions_inlined = true; | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2768 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2769 return todo; |
0 | 2770 } |
2771 | |
111 | 2772 /* Do inlining of small functions. Doing so early helps profiling and other |
2773 passes to be somewhat more effective and avoids some code duplication in | |
2774 later real inlining pass for testcases with very many function calls. */ | |
2775 | |
2776 namespace { | |
2777 | |
2778 const pass_data pass_data_early_inline = | |
0 | 2779 { |
111 | 2780 GIMPLE_PASS, /* type */ |
2781 "einline", /* name */ | |
2782 OPTGROUP_INLINE, /* optinfo_flags */ | |
2783 TV_EARLY_INLINING, /* tv_id */ | |
2784 PROP_ssa, /* properties_required */ | |
2785 0, /* properties_provided */ | |
2786 0, /* properties_destroyed */ | |
2787 0, /* todo_flags_start */ | |
2788 0, /* todo_flags_finish */ | |
0 | 2789 }; |
2790 | |
111 | 2791 class pass_early_inline : public gimple_opt_pass |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2792 { |
111 | 2793 public: |
2794 pass_early_inline (gcc::context *ctxt) | |
2795 : gimple_opt_pass (pass_data_early_inline, ctxt) | |
2796 {} | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2797 |
111 | 2798 /* opt_pass methods: */ |
2799 virtual unsigned int execute (function *); | |
2800 | |
2801 }; // class pass_early_inline | |
2802 | |
2803 unsigned int | |
2804 pass_early_inline::execute (function *fun) | |
2805 { | |
2806 return early_inliner (fun); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2807 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2808 |
111 | 2809 } // anon namespace |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
2810 |
111 | 2811 gimple_opt_pass * |
2812 make_pass_early_inline (gcc::context *ctxt) | |
0 | 2813 { |
111 | 2814 return new pass_early_inline (ctxt); |
0 | 2815 } |
2816 | |
111 | 2817 namespace { |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2818 |
111 | 2819 const pass_data pass_data_ipa_inline = |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2820 { |
111 | 2821 IPA_PASS, /* type */ |
2822 "inline", /* name */ | |
2823 OPTGROUP_INLINE, /* optinfo_flags */ | |
2824 TV_IPA_INLINING, /* tv_id */ | |
2825 0, /* properties_required */ | |
2826 0, /* properties_provided */ | |
2827 0, /* properties_destroyed */ | |
2828 0, /* todo_flags_start */ | |
2829 ( TODO_dump_symtab ), /* todo_flags_finish */ | |
0 | 2830 }; |
2831 | |
111 | 2832 class pass_ipa_inline : public ipa_opt_pass_d |
2833 { | |
2834 public: | |
2835 pass_ipa_inline (gcc::context *ctxt) | |
2836 : ipa_opt_pass_d (pass_data_ipa_inline, ctxt, | |
2837 NULL, /* generate_summary */ | |
2838 NULL, /* write_summary */ | |
2839 NULL, /* read_summary */ | |
2840 NULL, /* write_optimization_summary */ | |
2841 NULL, /* read_optimization_summary */ | |
2842 NULL, /* stmt_fixup */ | |
2843 0, /* function_transform_todo_flags_start */ | |
2844 inline_transform, /* function_transform */ | |
2845 NULL) /* variable_transform */ | |
2846 {} | |
0 | 2847 |
111 | 2848 /* opt_pass methods: */ |
2849 virtual unsigned int execute (function *) { return ipa_inline (); } | |
2850 | |
2851 }; // class pass_ipa_inline | |
2852 | |
2853 } // anon namespace | |
2854 | |
2855 ipa_opt_pass_d * | |
2856 make_pass_ipa_inline (gcc::context *ctxt) | |
2857 { | |
2858 return new pass_ipa_inline (ctxt); | |
2859 } |