]>
Commit | Line | Data |
---|---|---|
564fe867 | 1 | /* Callgraph clones |
d1e082c2 | 2 | Copyright (C) 2003-2013 Free Software Foundation, Inc. |
564fe867 JH |
3 | Contributed by Jan Hubicka |
4 | ||
5 | This file is part of GCC. | |
6 | ||
7 | GCC is free software; you can redistribute it and/or modify it under | |
8 | the terms of the GNU General Public License as published by the Free | |
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 | /* This module provide facilities for clonning functions. I.e. creating | |
22 | new functions based on existing functions with simple modifications, | |
23 | such as replacement of parameters. | |
24 | ||
25 | To allow whole program optimization without actual presence of function | |
26 | bodies, an additional infrastructure is provided for so-called virtual | |
27 | clones | |
28 | ||
29 | A virtual clone in the callgraph is a function that has no | |
30 | associated body, just a description of how to create its body based | |
31 | on a different function (which itself may be a virtual clone). | |
32 | ||
33 | The description of function modifications includes adjustments to | |
34 | the function's signature (which allows, for example, removing or | |
35 | adding function arguments), substitutions to perform on the | |
36 | function body, and, for inlined functions, a pointer to the | |
37 | function that it will be inlined into. | |
38 | ||
39 | It is also possible to redirect any edge of the callgraph from a | |
40 | function to its virtual clone. This implies updating of the call | |
41 | site to adjust for the new function signature. | |
42 | ||
43 | Most of the transformations performed by inter-procedural | |
44 | optimizations can be represented via virtual clones. For | |
45 | instance, a constant propagation pass can produce a virtual clone | |
46 | of the function which replaces one of its arguments by a | |
47 | constant. The inliner can represent its decisions by producing a | |
48 | clone of a function whose body will be later integrated into | |
49 | a given function. | |
50 | ||
51 | Using virtual clones, the program can be easily updated | |
52 | during the Execute stage, solving most of pass interactions | |
53 | problems that would otherwise occur during Transform. | |
54 | ||
55 | Virtual clones are later materialized in the LTRANS stage and | |
56 | turned into real functions. Passes executed after the virtual | |
57 | clone were introduced also perform their Transform stage | |
58 | on new functions, so for a pass there is no significant | |
59 | difference between operating on a real function or a virtual | |
60 | clone introduced before its Execute stage. | |
61 | ||
62 | Optimization passes then work on virtual clones introduced before | |
63 | their Execute stage as if they were real functions. The | |
64 | only difference is that clones are not visible during the | |
65 | Generate Summary stage. */ | |
66 | ||
67 | #include "config.h" | |
68 | #include "system.h" | |
69 | #include "coretypes.h" | |
70 | #include "tm.h" | |
d8a2d370 | 71 | #include "rtl.h" |
564fe867 | 72 | #include "tree.h" |
d8a2d370 DN |
73 | #include "stringpool.h" |
74 | #include "function.h" | |
75 | #include "emit-rtl.h" | |
8e9055ae | 76 | #include "gimple.h" |
442b4905 AM |
77 | #include "bitmap.h" |
78 | #include "tree-cfg.h" | |
564fe867 JH |
79 | #include "tree-inline.h" |
80 | #include "langhooks.h" | |
81 | #include "pointer-set.h" | |
82 | #include "toplev.h" | |
83 | #include "flags.h" | |
84 | #include "ggc.h" | |
85 | #include "debug.h" | |
86 | #include "target.h" | |
564fe867 | 87 | #include "diagnostic.h" |
564fe867 | 88 | #include "params.h" |
564fe867 JH |
89 | #include "intl.h" |
90 | #include "function.h" | |
91 | #include "ipa-prop.h" | |
564fe867 | 92 | #include "tree-iterator.h" |
564fe867 JH |
93 | #include "tree-dump.h" |
94 | #include "gimple-pretty-print.h" | |
564fe867 | 95 | #include "coverage.h" |
564fe867 JH |
96 | #include "ipa-inline.h" |
97 | #include "ipa-utils.h" | |
98 | #include "lto-streamer.h" | |
99 | #include "except.h" | |
100 | ||
101 | /* Create clone of E in the node N represented by CALL_EXPR the callgraph. */ | |
102 | struct cgraph_edge * | |
103 | cgraph_clone_edge (struct cgraph_edge *e, struct cgraph_node *n, | |
104 | gimple call_stmt, unsigned stmt_uid, gcov_type count_scale, | |
105 | int freq_scale, bool update_original) | |
106 | { | |
107 | struct cgraph_edge *new_edge; | |
8b47039c | 108 | gcov_type count = apply_probability (e->count, count_scale); |
564fe867 JH |
109 | gcov_type freq; |
110 | ||
111 | /* We do not want to ignore loop nest after frequency drops to 0. */ | |
112 | if (!freq_scale) | |
113 | freq_scale = 1; | |
114 | freq = e->frequency * (gcov_type) freq_scale / CGRAPH_FREQ_BASE; | |
115 | if (freq > CGRAPH_FREQ_MAX) | |
116 | freq = CGRAPH_FREQ_MAX; | |
117 | ||
118 | if (e->indirect_unknown_callee) | |
119 | { | |
120 | tree decl; | |
121 | ||
122 | if (call_stmt && (decl = gimple_call_fndecl (call_stmt))) | |
123 | { | |
124 | struct cgraph_node *callee = cgraph_get_node (decl); | |
125 | gcc_checking_assert (callee); | |
126 | new_edge = cgraph_create_edge (n, callee, call_stmt, count, freq); | |
127 | } | |
128 | else | |
129 | { | |
130 | new_edge = cgraph_create_indirect_edge (n, call_stmt, | |
131 | e->indirect_info->ecf_flags, | |
132 | count, freq); | |
133 | *new_edge->indirect_info = *e->indirect_info; | |
134 | } | |
135 | } | |
136 | else | |
137 | { | |
138 | new_edge = cgraph_create_edge (n, e->callee, call_stmt, count, freq); | |
139 | if (e->indirect_info) | |
140 | { | |
141 | new_edge->indirect_info | |
142 | = ggc_alloc_cleared_cgraph_indirect_call_info (); | |
143 | *new_edge->indirect_info = *e->indirect_info; | |
144 | } | |
145 | } | |
146 | ||
147 | new_edge->inline_failed = e->inline_failed; | |
148 | new_edge->indirect_inlining_edge = e->indirect_inlining_edge; | |
149 | new_edge->lto_stmt_uid = stmt_uid; | |
150 | /* Clone flags that depend on call_stmt availability manually. */ | |
151 | new_edge->can_throw_external = e->can_throw_external; | |
152 | new_edge->call_stmt_cannot_inline_p = e->call_stmt_cannot_inline_p; | |
042ae7d2 | 153 | new_edge->speculative = e->speculative; |
564fe867 JH |
154 | if (update_original) |
155 | { | |
156 | e->count -= new_edge->count; | |
157 | if (e->count < 0) | |
158 | e->count = 0; | |
159 | } | |
160 | cgraph_call_edge_duplication_hooks (e, new_edge); | |
161 | return new_edge; | |
162 | } | |
163 | ||
164 | ||
165 | /* Create node representing clone of N executed COUNT times. Decrease | |
166 | the execution counts from original node too. | |
167 | The new clone will have decl set to DECL that may or may not be the same | |
168 | as decl of N. | |
169 | ||
170 | When UPDATE_ORIGINAL is true, the counts are subtracted from the original | |
171 | function's profile to reflect the fact that part of execution is handled | |
172 | by node. | |
173 | When CALL_DUPLICATOIN_HOOK is true, the ipa passes are acknowledged about | |
44a60244 MJ |
174 | the new clone. Otherwise the caller is responsible for doing so later. |
175 | ||
176 | If the new node is being inlined into another one, NEW_INLINED_TO should be | |
177 | the outline function the new one is (even indirectly) inlined to. All hooks | |
178 | will see this in node's global.inlined_to, when invoked. Can be NULL if the | |
179 | node is not inlined. */ | |
564fe867 JH |
180 | |
181 | struct cgraph_node * | |
182 | cgraph_clone_node (struct cgraph_node *n, tree decl, gcov_type count, int freq, | |
183 | bool update_original, | |
9771b263 | 184 | vec<cgraph_edge_p> redirect_callers, |
44a60244 MJ |
185 | bool call_duplication_hook, |
186 | struct cgraph_node *new_inlined_to) | |
564fe867 JH |
187 | { |
188 | struct cgraph_node *new_node = cgraph_create_empty_node (); | |
189 | struct cgraph_edge *e; | |
190 | gcov_type count_scale; | |
191 | unsigned i; | |
192 | ||
67348ccc DM |
193 | new_node->decl = decl; |
194 | symtab_register_node (new_node); | |
564fe867 | 195 | new_node->origin = n->origin; |
67348ccc | 196 | new_node->lto_file_data = n->lto_file_data; |
564fe867 JH |
197 | if (new_node->origin) |
198 | { | |
199 | new_node->next_nested = new_node->origin->nested; | |
200 | new_node->origin->nested = new_node; | |
201 | } | |
67348ccc DM |
202 | new_node->analyzed = n->analyzed; |
203 | new_node->definition = n->definition; | |
564fe867 | 204 | new_node->local = n->local; |
67348ccc | 205 | new_node->externally_visible = false; |
564fe867 JH |
206 | new_node->local.local = true; |
207 | new_node->global = n->global; | |
44a60244 | 208 | new_node->global.inlined_to = new_inlined_to; |
564fe867 JH |
209 | new_node->rtl = n->rtl; |
210 | new_node->count = count; | |
211 | new_node->frequency = n->frequency; | |
212 | new_node->clone = n->clone; | |
9771b263 | 213 | new_node->clone.tree_map = NULL; |
86ce5d2f | 214 | new_node->tp_first_run = n->tp_first_run; |
564fe867 JH |
215 | if (n->count) |
216 | { | |
217 | if (new_node->count > n->count) | |
218 | count_scale = REG_BR_PROB_BASE; | |
219 | else | |
8b47039c | 220 | count_scale = GCOV_COMPUTE_SCALE (new_node->count, n->count); |
564fe867 JH |
221 | } |
222 | else | |
223 | count_scale = 0; | |
224 | if (update_original) | |
225 | { | |
226 | n->count -= count; | |
227 | if (n->count < 0) | |
228 | n->count = 0; | |
229 | } | |
230 | ||
9771b263 | 231 | FOR_EACH_VEC_ELT (redirect_callers, i, e) |
564fe867 JH |
232 | { |
233 | /* Redirect calls to the old version node to point to its new | |
234 | version. */ | |
235 | cgraph_redirect_edge_callee (e, new_node); | |
236 | } | |
237 | ||
238 | ||
239 | for (e = n->callees;e; e=e->next_callee) | |
240 | cgraph_clone_edge (e, new_node, e->call_stmt, e->lto_stmt_uid, | |
241 | count_scale, freq, update_original); | |
242 | ||
243 | for (e = n->indirect_calls; e; e = e->next_callee) | |
244 | cgraph_clone_edge (e, new_node, e->call_stmt, e->lto_stmt_uid, | |
245 | count_scale, freq, update_original); | |
67348ccc | 246 | ipa_clone_references (new_node, &n->ref_list); |
564fe867 JH |
247 | |
248 | new_node->next_sibling_clone = n->clones; | |
249 | if (n->clones) | |
250 | n->clones->prev_sibling_clone = new_node; | |
251 | n->clones = new_node; | |
252 | new_node->clone_of = n; | |
253 | ||
254 | if (call_duplication_hook) | |
255 | cgraph_call_node_duplication_hooks (n, new_node); | |
256 | return new_node; | |
257 | } | |
258 | ||
440a5082 | 259 | /* Return a new assembler name for a clone of DECL with SUFFIX. */ |
564fe867 JH |
260 | |
261 | static GTY(()) unsigned int clone_fn_id_num; | |
262 | ||
263 | tree | |
264 | clone_function_name (tree decl, const char *suffix) | |
265 | { | |
266 | tree name = DECL_ASSEMBLER_NAME (decl); | |
267 | size_t len = IDENTIFIER_LENGTH (name); | |
268 | char *tmp_name, *prefix; | |
269 | ||
270 | prefix = XALLOCAVEC (char, len + strlen (suffix) + 2); | |
271 | memcpy (prefix, IDENTIFIER_POINTER (name), len); | |
272 | strcpy (prefix + len + 1, suffix); | |
273 | #ifndef NO_DOT_IN_LABEL | |
274 | prefix[len] = '.'; | |
275 | #elif !defined NO_DOLLAR_IN_LABEL | |
276 | prefix[len] = '$'; | |
277 | #else | |
278 | prefix[len] = '_'; | |
279 | #endif | |
280 | ASM_FORMAT_PRIVATE_NAME (tmp_name, prefix, clone_fn_id_num++); | |
281 | return get_identifier (tmp_name); | |
282 | } | |
283 | ||
862d0b35 DN |
284 | /* Build variant of function type ORIG_TYPE skipping ARGS_TO_SKIP and the |
285 | return value if SKIP_RETURN is true. */ | |
286 | ||
287 | static tree | |
288 | build_function_type_skip_args (tree orig_type, bitmap args_to_skip, | |
289 | bool skip_return) | |
290 | { | |
291 | tree new_type = NULL; | |
292 | tree args, new_args = NULL, t; | |
293 | tree new_reversed; | |
294 | int i = 0; | |
295 | ||
296 | for (args = TYPE_ARG_TYPES (orig_type); args && args != void_list_node; | |
297 | args = TREE_CHAIN (args), i++) | |
298 | if (!args_to_skip || !bitmap_bit_p (args_to_skip, i)) | |
299 | new_args = tree_cons (NULL_TREE, TREE_VALUE (args), new_args); | |
300 | ||
301 | new_reversed = nreverse (new_args); | |
302 | if (args) | |
303 | { | |
304 | if (new_reversed) | |
305 | TREE_CHAIN (new_args) = void_list_node; | |
306 | else | |
307 | new_reversed = void_list_node; | |
308 | } | |
309 | ||
310 | /* Use copy_node to preserve as much as possible from original type | |
311 | (debug info, attribute lists etc.) | |
312 | Exception is METHOD_TYPEs must have THIS argument. | |
313 | When we are asked to remove it, we need to build new FUNCTION_TYPE | |
314 | instead. */ | |
315 | if (TREE_CODE (orig_type) != METHOD_TYPE | |
316 | || !args_to_skip | |
317 | || !bitmap_bit_p (args_to_skip, 0)) | |
318 | { | |
319 | new_type = build_distinct_type_copy (orig_type); | |
320 | TYPE_ARG_TYPES (new_type) = new_reversed; | |
321 | } | |
322 | else | |
323 | { | |
324 | new_type | |
325 | = build_distinct_type_copy (build_function_type (TREE_TYPE (orig_type), | |
326 | new_reversed)); | |
327 | TYPE_CONTEXT (new_type) = TYPE_CONTEXT (orig_type); | |
328 | } | |
329 | ||
330 | if (skip_return) | |
331 | TREE_TYPE (new_type) = void_type_node; | |
332 | ||
333 | /* This is a new type, not a copy of an old type. Need to reassociate | |
334 | variants. We can handle everything except the main variant lazily. */ | |
335 | t = TYPE_MAIN_VARIANT (orig_type); | |
336 | if (t != orig_type) | |
337 | { | |
338 | t = build_function_type_skip_args (t, args_to_skip, skip_return); | |
339 | TYPE_MAIN_VARIANT (new_type) = t; | |
340 | TYPE_NEXT_VARIANT (new_type) = TYPE_NEXT_VARIANT (t); | |
341 | TYPE_NEXT_VARIANT (t) = new_type; | |
342 | } | |
343 | else | |
344 | { | |
345 | TYPE_MAIN_VARIANT (new_type) = new_type; | |
346 | TYPE_NEXT_VARIANT (new_type) = NULL; | |
347 | } | |
348 | ||
349 | return new_type; | |
350 | } | |
351 | ||
352 | /* Build variant of function decl ORIG_DECL skipping ARGS_TO_SKIP and the | |
353 | return value if SKIP_RETURN is true. | |
354 | ||
355 | Arguments from DECL_ARGUMENTS list can't be removed now, since they are | |
356 | linked by TREE_CHAIN directly. The caller is responsible for eliminating | |
357 | them when they are being duplicated (i.e. copy_arguments_for_versioning). */ | |
358 | ||
359 | static tree | |
360 | build_function_decl_skip_args (tree orig_decl, bitmap args_to_skip, | |
361 | bool skip_return) | |
362 | { | |
363 | tree new_decl = copy_node (orig_decl); | |
364 | tree new_type; | |
365 | ||
366 | new_type = TREE_TYPE (orig_decl); | |
367 | if (prototype_p (new_type) | |
368 | || (skip_return && !VOID_TYPE_P (TREE_TYPE (new_type)))) | |
369 | new_type | |
370 | = build_function_type_skip_args (new_type, args_to_skip, skip_return); | |
371 | TREE_TYPE (new_decl) = new_type; | |
372 | ||
373 | /* For declarations setting DECL_VINDEX (i.e. methods) | |
374 | we expect first argument to be THIS pointer. */ | |
375 | if (args_to_skip && bitmap_bit_p (args_to_skip, 0)) | |
376 | DECL_VINDEX (new_decl) = NULL_TREE; | |
377 | ||
378 | /* When signature changes, we need to clear builtin info. */ | |
379 | if (DECL_BUILT_IN (new_decl) | |
380 | && args_to_skip | |
381 | && !bitmap_empty_p (args_to_skip)) | |
382 | { | |
383 | DECL_BUILT_IN_CLASS (new_decl) = NOT_BUILT_IN; | |
384 | DECL_FUNCTION_CODE (new_decl) = (enum built_in_function) 0; | |
385 | } | |
386 | return new_decl; | |
387 | } | |
388 | ||
564fe867 JH |
389 | /* Create callgraph node clone with new declaration. The actual body will |
390 | be copied later at compilation stage. | |
391 | ||
392 | TODO: after merging in ipa-sra use function call notes instead of args_to_skip | |
393 | bitmap interface. | |
394 | */ | |
395 | struct cgraph_node * | |
396 | cgraph_create_virtual_clone (struct cgraph_node *old_node, | |
9771b263 DN |
397 | vec<cgraph_edge_p> redirect_callers, |
398 | vec<ipa_replace_map_p, va_gc> *tree_map, | |
564fe867 JH |
399 | bitmap args_to_skip, |
400 | const char * suffix) | |
401 | { | |
67348ccc | 402 | tree old_decl = old_node->decl; |
564fe867 JH |
403 | struct cgraph_node *new_node = NULL; |
404 | tree new_decl; | |
440a5082 | 405 | size_t len, i; |
564fe867 | 406 | struct ipa_replace_map *map; |
440a5082 | 407 | char *name; |
564fe867 | 408 | |
a2e2a668 | 409 | if (!in_lto_p) |
564fe867 JH |
410 | gcc_checking_assert (tree_versionable_function_p (old_decl)); |
411 | ||
412 | gcc_assert (old_node->local.can_change_signature || !args_to_skip); | |
413 | ||
414 | /* Make a new FUNCTION_DECL tree node */ | |
415 | if (!args_to_skip) | |
416 | new_decl = copy_node (old_decl); | |
417 | else | |
418 | new_decl = build_function_decl_skip_args (old_decl, args_to_skip, false); | |
49bde175 JH |
419 | |
420 | /* These pointers represent function body and will be populated only when clone | |
421 | is materialized. */ | |
422 | gcc_assert (new_decl != old_decl); | |
564fe867 | 423 | DECL_STRUCT_FUNCTION (new_decl) = NULL; |
49bde175 JH |
424 | DECL_ARGUMENTS (new_decl) = NULL; |
425 | DECL_INITIAL (new_decl) = NULL; | |
426 | DECL_RESULT (new_decl) = NULL; | |
427 | /* We can not do DECL_RESULT (new_decl) = NULL; here because of LTO partitioning | |
428 | sometimes storing only clone decl instead of original. */ | |
564fe867 JH |
429 | |
430 | /* Generate a new name for the new version. */ | |
440a5082 EB |
431 | len = IDENTIFIER_LENGTH (DECL_NAME (old_decl)); |
432 | name = XALLOCAVEC (char, len + strlen (suffix) + 2); | |
433 | memcpy (name, IDENTIFIER_POINTER (DECL_NAME (old_decl)), len); | |
434 | strcpy (name + len + 1, suffix); | |
435 | name[len] = '.'; | |
436 | DECL_NAME (new_decl) = get_identifier (name); | |
437 | SET_DECL_ASSEMBLER_NAME (new_decl, clone_function_name (old_decl, suffix)); | |
564fe867 JH |
438 | SET_DECL_RTL (new_decl, NULL); |
439 | ||
440 | new_node = cgraph_clone_node (old_node, new_decl, old_node->count, | |
441 | CGRAPH_FREQ_BASE, false, | |
44a60244 | 442 | redirect_callers, false, NULL); |
564fe867 JH |
443 | /* Update the properties. |
444 | Make clone visible only within this translation unit. Make sure | |
445 | that is not weak also. | |
446 | ??? We cannot use COMDAT linkage because there is no | |
447 | ABI support for this. */ | |
67348ccc | 448 | DECL_EXTERNAL (new_node->decl) = 0; |
564fe867 | 449 | if (DECL_ONE_ONLY (old_decl)) |
67348ccc DM |
450 | DECL_SECTION_NAME (new_node->decl) = NULL; |
451 | DECL_COMDAT_GROUP (new_node->decl) = 0; | |
452 | TREE_PUBLIC (new_node->decl) = 0; | |
453 | DECL_COMDAT (new_node->decl) = 0; | |
454 | DECL_WEAK (new_node->decl) = 0; | |
455 | DECL_VIRTUAL_P (new_node->decl) = 0; | |
456 | DECL_STATIC_CONSTRUCTOR (new_node->decl) = 0; | |
457 | DECL_STATIC_DESTRUCTOR (new_node->decl) = 0; | |
564fe867 JH |
458 | new_node->clone.tree_map = tree_map; |
459 | new_node->clone.args_to_skip = args_to_skip; | |
702d8703 JH |
460 | |
461 | /* Clones of global symbols or symbols with unique names are unique. */ | |
462 | if ((TREE_PUBLIC (old_decl) | |
463 | && !DECL_EXTERNAL (old_decl) | |
464 | && !DECL_WEAK (old_decl) | |
465 | && !DECL_COMDAT (old_decl)) | |
466 | || in_lto_p) | |
67348ccc | 467 | new_node->unique_name = true; |
9771b263 | 468 | FOR_EACH_VEC_SAFE_ELT (tree_map, i, map) |
67348ccc | 469 | ipa_maybe_record_reference (new_node, map->new_tree, |
79ee9826 | 470 | IPA_REF_ADDR, NULL); |
564fe867 JH |
471 | if (!args_to_skip) |
472 | new_node->clone.combined_args_to_skip = old_node->clone.combined_args_to_skip; | |
473 | else if (old_node->clone.combined_args_to_skip) | |
474 | { | |
475 | int newi = 0, oldi = 0; | |
476 | tree arg; | |
477 | bitmap new_args_to_skip = BITMAP_GGC_ALLOC (); | |
478 | struct cgraph_node *orig_node; | |
479 | for (orig_node = old_node; orig_node->clone_of; orig_node = orig_node->clone_of) | |
480 | ; | |
67348ccc | 481 | for (arg = DECL_ARGUMENTS (orig_node->decl); |
564fe867 JH |
482 | arg; arg = DECL_CHAIN (arg), oldi++) |
483 | { | |
484 | if (bitmap_bit_p (old_node->clone.combined_args_to_skip, oldi)) | |
485 | { | |
486 | bitmap_set_bit (new_args_to_skip, oldi); | |
487 | continue; | |
488 | } | |
489 | if (bitmap_bit_p (args_to_skip, newi)) | |
490 | bitmap_set_bit (new_args_to_skip, oldi); | |
491 | newi++; | |
492 | } | |
493 | new_node->clone.combined_args_to_skip = new_args_to_skip; | |
494 | } | |
495 | else | |
496 | new_node->clone.combined_args_to_skip = args_to_skip; | |
67348ccc | 497 | new_node->externally_visible = 0; |
564fe867 JH |
498 | new_node->local.local = 1; |
499 | new_node->lowered = true; | |
500 | ||
501 | cgraph_call_node_duplication_hooks (old_node, new_node); | |
502 | ||
503 | ||
504 | return new_node; | |
505 | } | |
506 | ||
507 | /* NODE is being removed from symbol table; see if its entry can be replaced by | |
508 | other inline clone. */ | |
509 | struct cgraph_node * | |
510 | cgraph_find_replacement_node (struct cgraph_node *node) | |
511 | { | |
512 | struct cgraph_node *next_inline_clone, *replacement; | |
513 | ||
514 | for (next_inline_clone = node->clones; | |
515 | next_inline_clone | |
67348ccc | 516 | && next_inline_clone->decl != node->decl; |
564fe867 JH |
517 | next_inline_clone = next_inline_clone->next_sibling_clone) |
518 | ; | |
519 | ||
520 | /* If there is inline clone of the node being removed, we need | |
521 | to put it into the position of removed node and reorganize all | |
522 | other clones to be based on it. */ | |
523 | if (next_inline_clone) | |
524 | { | |
525 | struct cgraph_node *n; | |
526 | struct cgraph_node *new_clones; | |
527 | ||
528 | replacement = next_inline_clone; | |
529 | ||
530 | /* Unlink inline clone from the list of clones of removed node. */ | |
531 | if (next_inline_clone->next_sibling_clone) | |
532 | next_inline_clone->next_sibling_clone->prev_sibling_clone | |
533 | = next_inline_clone->prev_sibling_clone; | |
534 | if (next_inline_clone->prev_sibling_clone) | |
535 | { | |
536 | gcc_assert (node->clones != next_inline_clone); | |
537 | next_inline_clone->prev_sibling_clone->next_sibling_clone | |
538 | = next_inline_clone->next_sibling_clone; | |
539 | } | |
540 | else | |
541 | { | |
542 | gcc_assert (node->clones == next_inline_clone); | |
543 | node->clones = next_inline_clone->next_sibling_clone; | |
544 | } | |
545 | ||
546 | new_clones = node->clones; | |
547 | node->clones = NULL; | |
548 | ||
549 | /* Copy clone info. */ | |
550 | next_inline_clone->clone = node->clone; | |
551 | ||
552 | /* Now place it into clone tree at same level at NODE. */ | |
553 | next_inline_clone->clone_of = node->clone_of; | |
554 | next_inline_clone->prev_sibling_clone = NULL; | |
555 | next_inline_clone->next_sibling_clone = NULL; | |
556 | if (node->clone_of) | |
557 | { | |
558 | if (node->clone_of->clones) | |
559 | node->clone_of->clones->prev_sibling_clone = next_inline_clone; | |
560 | next_inline_clone->next_sibling_clone = node->clone_of->clones; | |
561 | node->clone_of->clones = next_inline_clone; | |
562 | } | |
563 | ||
564 | /* Merge the clone list. */ | |
565 | if (new_clones) | |
566 | { | |
567 | if (!next_inline_clone->clones) | |
568 | next_inline_clone->clones = new_clones; | |
569 | else | |
570 | { | |
571 | n = next_inline_clone->clones; | |
572 | while (n->next_sibling_clone) | |
573 | n = n->next_sibling_clone; | |
574 | n->next_sibling_clone = new_clones; | |
575 | new_clones->prev_sibling_clone = n; | |
576 | } | |
577 | } | |
578 | ||
579 | /* Update clone_of pointers. */ | |
580 | n = new_clones; | |
581 | while (n) | |
582 | { | |
583 | n->clone_of = next_inline_clone; | |
584 | n = n->next_sibling_clone; | |
585 | } | |
586 | return replacement; | |
587 | } | |
588 | else | |
589 | return NULL; | |
590 | } | |
591 | ||
592 | /* Like cgraph_set_call_stmt but walk the clone tree and update all | |
042ae7d2 JH |
593 | clones sharing the same function body. |
594 | When WHOLE_SPECULATIVE_EDGES is true, all three components of | |
595 | speculative edge gets updated. Otherwise we update only direct | |
596 | call. */ | |
564fe867 JH |
597 | |
598 | void | |
599 | cgraph_set_call_stmt_including_clones (struct cgraph_node *orig, | |
042ae7d2 JH |
600 | gimple old_stmt, gimple new_stmt, |
601 | bool update_speculative) | |
564fe867 JH |
602 | { |
603 | struct cgraph_node *node; | |
604 | struct cgraph_edge *edge = cgraph_edge (orig, old_stmt); | |
605 | ||
606 | if (edge) | |
042ae7d2 | 607 | cgraph_set_call_stmt (edge, new_stmt, update_speculative); |
564fe867 JH |
608 | |
609 | node = orig->clones; | |
610 | if (node) | |
611 | while (node != orig) | |
612 | { | |
613 | struct cgraph_edge *edge = cgraph_edge (node, old_stmt); | |
614 | if (edge) | |
042ae7d2 JH |
615 | { |
616 | cgraph_set_call_stmt (edge, new_stmt, update_speculative); | |
617 | /* If UPDATE_SPECULATIVE is false, it means that we are turning | |
618 | speculative call into a real code sequence. Update the | |
619 | callgraph edges. */ | |
620 | if (edge->speculative && !update_speculative) | |
621 | { | |
622 | struct cgraph_edge *direct, *indirect; | |
623 | struct ipa_ref *ref; | |
624 | ||
625 | gcc_assert (!edge->indirect_unknown_callee); | |
626 | cgraph_speculative_call_info (edge, direct, indirect, ref); | |
627 | direct->speculative = false; | |
628 | indirect->speculative = false; | |
629 | ref->speculative = false; | |
630 | } | |
631 | } | |
564fe867 JH |
632 | if (node->clones) |
633 | node = node->clones; | |
634 | else if (node->next_sibling_clone) | |
635 | node = node->next_sibling_clone; | |
636 | else | |
637 | { | |
638 | while (node != orig && !node->next_sibling_clone) | |
639 | node = node->clone_of; | |
640 | if (node != orig) | |
641 | node = node->next_sibling_clone; | |
642 | } | |
643 | } | |
644 | } | |
645 | ||
646 | /* Like cgraph_create_edge walk the clone tree and update all clones sharing | |
647 | same function body. If clones already have edge for OLD_STMT; only | |
648 | update the edge same way as cgraph_set_call_stmt_including_clones does. | |
649 | ||
650 | TODO: COUNT and LOOP_DEPTH should be properly distributed based on relative | |
651 | frequencies of the clones. */ | |
652 | ||
653 | void | |
654 | cgraph_create_edge_including_clones (struct cgraph_node *orig, | |
655 | struct cgraph_node *callee, | |
656 | gimple old_stmt, | |
657 | gimple stmt, gcov_type count, | |
658 | int freq, | |
659 | cgraph_inline_failed_t reason) | |
660 | { | |
661 | struct cgraph_node *node; | |
662 | struct cgraph_edge *edge; | |
663 | ||
664 | if (!cgraph_edge (orig, stmt)) | |
665 | { | |
666 | edge = cgraph_create_edge (orig, callee, stmt, count, freq); | |
667 | edge->inline_failed = reason; | |
668 | } | |
669 | ||
670 | node = orig->clones; | |
671 | if (node) | |
672 | while (node != orig) | |
673 | { | |
674 | struct cgraph_edge *edge = cgraph_edge (node, old_stmt); | |
675 | ||
676 | /* It is possible that clones already contain the edge while | |
677 | master didn't. Either we promoted indirect call into direct | |
678 | call in the clone or we are processing clones of unreachable | |
679 | master where edges has been removed. */ | |
680 | if (edge) | |
681 | cgraph_set_call_stmt (edge, stmt); | |
682 | else if (!cgraph_edge (node, stmt)) | |
683 | { | |
684 | edge = cgraph_create_edge (node, callee, stmt, count, | |
685 | freq); | |
686 | edge->inline_failed = reason; | |
687 | } | |
688 | ||
689 | if (node->clones) | |
690 | node = node->clones; | |
691 | else if (node->next_sibling_clone) | |
692 | node = node->next_sibling_clone; | |
693 | else | |
694 | { | |
695 | while (node != orig && !node->next_sibling_clone) | |
696 | node = node->clone_of; | |
697 | if (node != orig) | |
698 | node = node->next_sibling_clone; | |
699 | } | |
700 | } | |
701 | } | |
702 | ||
703 | /* Remove the node from cgraph and all inline clones inlined into it. | |
704 | Skip however removal of FORBIDDEN_NODE and return true if it needs to be | |
705 | removed. This allows to call the function from outer loop walking clone | |
706 | tree. */ | |
707 | ||
708 | bool | |
709 | cgraph_remove_node_and_inline_clones (struct cgraph_node *node, struct cgraph_node *forbidden_node) | |
710 | { | |
711 | struct cgraph_edge *e, *next; | |
712 | bool found = false; | |
713 | ||
714 | if (node == forbidden_node) | |
39f9719e JH |
715 | { |
716 | cgraph_remove_edge (node->callers); | |
717 | return true; | |
718 | } | |
564fe867 JH |
719 | for (e = node->callees; e; e = next) |
720 | { | |
721 | next = e->next_callee; | |
722 | if (!e->inline_failed) | |
723 | found |= cgraph_remove_node_and_inline_clones (e->callee, forbidden_node); | |
724 | } | |
725 | cgraph_remove_node (node); | |
726 | return found; | |
727 | } | |
728 | ||
729 | /* The edges representing the callers of the NEW_VERSION node were | |
730 | fixed by cgraph_function_versioning (), now the call_expr in their | |
731 | respective tree code should be updated to call the NEW_VERSION. */ | |
732 | ||
733 | static void | |
734 | update_call_expr (struct cgraph_node *new_version) | |
735 | { | |
736 | struct cgraph_edge *e; | |
737 | ||
738 | gcc_assert (new_version); | |
739 | ||
740 | /* Update the call expr on the edges to call the new version. */ | |
741 | for (e = new_version->callers; e; e = e->next_caller) | |
742 | { | |
67348ccc DM |
743 | struct function *inner_function = DECL_STRUCT_FUNCTION (e->caller->decl); |
744 | gimple_call_set_fndecl (e->call_stmt, new_version->decl); | |
564fe867 JH |
745 | maybe_clean_eh_stmt_fn (inner_function, e->call_stmt); |
746 | } | |
747 | } | |
748 | ||
749 | ||
750 | /* Create a new cgraph node which is the new version of | |
751 | OLD_VERSION node. REDIRECT_CALLERS holds the callers | |
752 | edges which should be redirected to point to | |
753 | NEW_VERSION. ALL the callees edges of OLD_VERSION | |
754 | are cloned to the new version node. Return the new | |
755 | version node. | |
756 | ||
757 | If non-NULL BLOCK_TO_COPY determine what basic blocks | |
758 | was copied to prevent duplications of calls that are dead | |
759 | in the clone. */ | |
760 | ||
761 | struct cgraph_node * | |
762 | cgraph_copy_node_for_versioning (struct cgraph_node *old_version, | |
763 | tree new_decl, | |
9771b263 | 764 | vec<cgraph_edge_p> redirect_callers, |
564fe867 JH |
765 | bitmap bbs_to_copy) |
766 | { | |
767 | struct cgraph_node *new_version; | |
768 | struct cgraph_edge *e; | |
769 | unsigned i; | |
770 | ||
771 | gcc_assert (old_version); | |
772 | ||
773 | new_version = cgraph_create_node (new_decl); | |
774 | ||
67348ccc DM |
775 | new_version->analyzed = old_version->analyzed; |
776 | new_version->definition = old_version->definition; | |
564fe867 | 777 | new_version->local = old_version->local; |
67348ccc DM |
778 | new_version->externally_visible = false; |
779 | new_version->local.local = new_version->definition; | |
564fe867 JH |
780 | new_version->global = old_version->global; |
781 | new_version->rtl = old_version->rtl; | |
782 | new_version->count = old_version->count; | |
783 | ||
784 | for (e = old_version->callees; e; e=e->next_callee) | |
785 | if (!bbs_to_copy | |
786 | || bitmap_bit_p (bbs_to_copy, gimple_bb (e->call_stmt)->index)) | |
787 | cgraph_clone_edge (e, new_version, e->call_stmt, | |
788 | e->lto_stmt_uid, REG_BR_PROB_BASE, | |
789 | CGRAPH_FREQ_BASE, | |
790 | true); | |
791 | for (e = old_version->indirect_calls; e; e=e->next_callee) | |
792 | if (!bbs_to_copy | |
793 | || bitmap_bit_p (bbs_to_copy, gimple_bb (e->call_stmt)->index)) | |
794 | cgraph_clone_edge (e, new_version, e->call_stmt, | |
795 | e->lto_stmt_uid, REG_BR_PROB_BASE, | |
796 | CGRAPH_FREQ_BASE, | |
797 | true); | |
9771b263 | 798 | FOR_EACH_VEC_ELT (redirect_callers, i, e) |
564fe867 JH |
799 | { |
800 | /* Redirect calls to the old version node to point to its new | |
801 | version. */ | |
802 | cgraph_redirect_edge_callee (e, new_version); | |
803 | } | |
804 | ||
805 | cgraph_call_node_duplication_hooks (old_version, new_version); | |
806 | ||
807 | return new_version; | |
808 | } | |
809 | ||
810 | /* Perform function versioning. | |
811 | Function versioning includes copying of the tree and | |
812 | a callgraph update (creating a new cgraph node and updating | |
813 | its callees and callers). | |
814 | ||
815 | REDIRECT_CALLERS varray includes the edges to be redirected | |
816 | to the new version. | |
817 | ||
818 | TREE_MAP is a mapping of tree nodes we want to replace with | |
819 | new ones (according to results of prior analysis). | |
820 | OLD_VERSION_NODE is the node that is versioned. | |
821 | ||
822 | If non-NULL ARGS_TO_SKIP determine function parameters to remove | |
823 | from new version. | |
824 | If SKIP_RETURN is true, the new version will return void. | |
825 | If non-NULL BLOCK_TO_COPY determine what basic blocks to copy. | |
826 | If non_NULL NEW_ENTRY determine new entry BB of the clone. | |
827 | ||
828 | Return the new version's cgraph node. */ | |
829 | ||
830 | struct cgraph_node * | |
831 | cgraph_function_versioning (struct cgraph_node *old_version_node, | |
9771b263 DN |
832 | vec<cgraph_edge_p> redirect_callers, |
833 | vec<ipa_replace_map_p, va_gc> *tree_map, | |
564fe867 JH |
834 | bitmap args_to_skip, |
835 | bool skip_return, | |
836 | bitmap bbs_to_copy, | |
837 | basic_block new_entry_block, | |
838 | const char *clone_name) | |
839 | { | |
67348ccc | 840 | tree old_decl = old_version_node->decl; |
564fe867 JH |
841 | struct cgraph_node *new_version_node = NULL; |
842 | tree new_decl; | |
843 | ||
844 | if (!tree_versionable_function_p (old_decl)) | |
845 | return NULL; | |
846 | ||
847 | gcc_assert (old_version_node->local.can_change_signature || !args_to_skip); | |
848 | ||
849 | /* Make a new FUNCTION_DECL tree node for the new version. */ | |
850 | if (!args_to_skip && !skip_return) | |
851 | new_decl = copy_node (old_decl); | |
852 | else | |
853 | new_decl | |
854 | = build_function_decl_skip_args (old_decl, args_to_skip, skip_return); | |
855 | ||
856 | /* Generate a new name for the new version. */ | |
857 | DECL_NAME (new_decl) = clone_function_name (old_decl, clone_name); | |
858 | SET_DECL_ASSEMBLER_NAME (new_decl, DECL_NAME (new_decl)); | |
859 | SET_DECL_RTL (new_decl, NULL); | |
860 | ||
861 | /* When the old decl was a con-/destructor make sure the clone isn't. */ | |
c3284718 RS |
862 | DECL_STATIC_CONSTRUCTOR (new_decl) = 0; |
863 | DECL_STATIC_DESTRUCTOR (new_decl) = 0; | |
564fe867 JH |
864 | |
865 | /* Create the new version's call-graph node. | |
866 | and update the edges of the new node. */ | |
867 | new_version_node = | |
868 | cgraph_copy_node_for_versioning (old_version_node, new_decl, | |
869 | redirect_callers, bbs_to_copy); | |
870 | ||
871 | /* Copy the OLD_VERSION_NODE function tree to the new version. */ | |
872 | tree_function_versioning (old_decl, new_decl, tree_map, false, args_to_skip, | |
873 | skip_return, bbs_to_copy, new_entry_block); | |
874 | ||
875 | /* Update the new version's properties. | |
876 | Make The new version visible only within this translation unit. Make sure | |
877 | that is not weak also. | |
878 | ??? We cannot use COMDAT linkage because there is no | |
879 | ABI support for this. */ | |
67348ccc DM |
880 | symtab_make_decl_local (new_version_node->decl); |
881 | DECL_VIRTUAL_P (new_version_node->decl) = 0; | |
882 | new_version_node->externally_visible = 0; | |
564fe867 JH |
883 | new_version_node->local.local = 1; |
884 | new_version_node->lowered = true; | |
702d8703 JH |
885 | /* Clones of global symbols or symbols with unique names are unique. */ |
886 | if ((TREE_PUBLIC (old_decl) | |
887 | && !DECL_EXTERNAL (old_decl) | |
888 | && !DECL_WEAK (old_decl) | |
889 | && !DECL_COMDAT (old_decl)) | |
890 | || in_lto_p) | |
67348ccc | 891 | new_version_node->unique_name = true; |
564fe867 JH |
892 | |
893 | /* Update the call_expr on the edges to call the new version node. */ | |
894 | update_call_expr (new_version_node); | |
895 | ||
896 | cgraph_call_function_insertion_hooks (new_version_node); | |
897 | return new_version_node; | |
898 | } | |
899 | ||
900 | /* Given virtual clone, turn it into actual clone. */ | |
901 | ||
902 | static void | |
903 | cgraph_materialize_clone (struct cgraph_node *node) | |
904 | { | |
905 | bitmap_obstack_initialize (NULL); | |
67348ccc | 906 | node->former_clone_of = node->clone_of->decl; |
564fe867 JH |
907 | if (node->clone_of->former_clone_of) |
908 | node->former_clone_of = node->clone_of->former_clone_of; | |
909 | /* Copy the OLD_VERSION_NODE function tree to the new version. */ | |
67348ccc | 910 | tree_function_versioning (node->clone_of->decl, node->decl, |
564fe867 JH |
911 | node->clone.tree_map, true, |
912 | node->clone.args_to_skip, false, | |
913 | NULL, NULL); | |
914 | if (cgraph_dump_file) | |
915 | { | |
67348ccc DM |
916 | dump_function_to_file (node->clone_of->decl, cgraph_dump_file, dump_flags); |
917 | dump_function_to_file (node->decl, cgraph_dump_file, dump_flags); | |
564fe867 JH |
918 | } |
919 | ||
920 | /* Function is no longer clone. */ | |
921 | if (node->next_sibling_clone) | |
922 | node->next_sibling_clone->prev_sibling_clone = node->prev_sibling_clone; | |
923 | if (node->prev_sibling_clone) | |
924 | node->prev_sibling_clone->next_sibling_clone = node->next_sibling_clone; | |
925 | else | |
926 | node->clone_of->clones = node->next_sibling_clone; | |
927 | node->next_sibling_clone = NULL; | |
928 | node->prev_sibling_clone = NULL; | |
67348ccc | 929 | if (!node->clone_of->analyzed && !node->clone_of->clones) |
564fe867 JH |
930 | { |
931 | cgraph_release_function_body (node->clone_of); | |
932 | cgraph_node_remove_callees (node->clone_of); | |
67348ccc | 933 | ipa_remove_all_references (&node->clone_of->ref_list); |
564fe867 JH |
934 | } |
935 | node->clone_of = NULL; | |
936 | bitmap_obstack_release (NULL); | |
937 | } | |
938 | ||
939 | /* Once all functions from compilation unit are in memory, produce all clones | |
940 | and update all calls. We might also do this on demand if we don't want to | |
941 | bring all functions to memory prior compilation, but current WHOPR | |
942 | implementation does that and it is is bit easier to keep everything right in | |
943 | this order. */ | |
944 | ||
945 | void | |
946 | cgraph_materialize_all_clones (void) | |
947 | { | |
948 | struct cgraph_node *node; | |
949 | bool stabilized = false; | |
042ae7d2 | 950 | |
564fe867 JH |
951 | |
952 | if (cgraph_dump_file) | |
953 | fprintf (cgraph_dump_file, "Materializing clones\n"); | |
954 | #ifdef ENABLE_CHECKING | |
955 | verify_cgraph (); | |
956 | #endif | |
957 | ||
958 | /* We can also do topological order, but number of iterations should be | |
959 | bounded by number of IPA passes since single IPA pass is probably not | |
960 | going to create clones of clones it created itself. */ | |
961 | while (!stabilized) | |
962 | { | |
963 | stabilized = true; | |
964 | FOR_EACH_FUNCTION (node) | |
965 | { | |
67348ccc DM |
966 | if (node->clone_of && node->decl != node->clone_of->decl |
967 | && !gimple_has_body_p (node->decl)) | |
564fe867 | 968 | { |
a2e2a668 JH |
969 | if (!node->clone_of->clone_of) |
970 | cgraph_get_body (node->clone_of); | |
67348ccc | 971 | if (gimple_has_body_p (node->clone_of->decl)) |
564fe867 JH |
972 | { |
973 | if (cgraph_dump_file) | |
974 | { | |
975 | fprintf (cgraph_dump_file, "cloning %s to %s\n", | |
fec39fa6 TS |
976 | xstrdup (node->clone_of->name ()), |
977 | xstrdup (node->name ())); | |
564fe867 JH |
978 | if (node->clone.tree_map) |
979 | { | |
980 | unsigned int i; | |
981 | fprintf (cgraph_dump_file, " replace map: "); | |
9771b263 DN |
982 | for (i = 0; |
983 | i < vec_safe_length (node->clone.tree_map); | |
984 | i++) | |
564fe867 JH |
985 | { |
986 | struct ipa_replace_map *replace_info; | |
9771b263 | 987 | replace_info = (*node->clone.tree_map)[i]; |
564fe867 JH |
988 | print_generic_expr (cgraph_dump_file, replace_info->old_tree, 0); |
989 | fprintf (cgraph_dump_file, " -> "); | |
990 | print_generic_expr (cgraph_dump_file, replace_info->new_tree, 0); | |
991 | fprintf (cgraph_dump_file, "%s%s;", | |
992 | replace_info->replace_p ? "(replace)":"", | |
993 | replace_info->ref_p ? "(ref)":""); | |
994 | } | |
995 | fprintf (cgraph_dump_file, "\n"); | |
996 | } | |
997 | if (node->clone.args_to_skip) | |
998 | { | |
999 | fprintf (cgraph_dump_file, " args_to_skip: "); | |
1000 | dump_bitmap (cgraph_dump_file, node->clone.args_to_skip); | |
1001 | } | |
1002 | if (node->clone.args_to_skip) | |
1003 | { | |
1004 | fprintf (cgraph_dump_file, " combined_args_to_skip:"); | |
1005 | dump_bitmap (cgraph_dump_file, node->clone.combined_args_to_skip); | |
1006 | } | |
1007 | } | |
1008 | cgraph_materialize_clone (node); | |
1009 | stabilized = false; | |
1010 | } | |
1011 | } | |
1012 | } | |
1013 | } | |
1014 | FOR_EACH_FUNCTION (node) | |
67348ccc | 1015 | if (!node->analyzed && node->callees) |
71cafea9 JH |
1016 | { |
1017 | cgraph_node_remove_callees (node); | |
67348ccc | 1018 | ipa_remove_all_references (&node->ref_list); |
71cafea9 JH |
1019 | } |
1020 | else | |
67348ccc | 1021 | ipa_clear_stmts_in_references (node); |
564fe867 JH |
1022 | if (cgraph_dump_file) |
1023 | fprintf (cgraph_dump_file, "Materialization Call site updates done.\n"); | |
1024 | #ifdef ENABLE_CHECKING | |
1025 | verify_cgraph (); | |
1026 | #endif | |
04142cc3 | 1027 | symtab_remove_unreachable_nodes (false, cgraph_dump_file); |
564fe867 JH |
1028 | } |
1029 | ||
1030 | #include "gt-cgraphclones.h" |