]>
Commit | Line | Data |
---|---|---|
c02159a7 | 1 | /* Callgraph clones |
fbd26352 | 2 | Copyright (C) 2003-2019 Free Software Foundation, Inc. |
c02159a7 | 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" | |
9ef16211 | 70 | #include "backend.h" |
7c29e30e | 71 | #include "target.h" |
72 | #include "rtl.h" | |
9ef16211 | 73 | #include "tree.h" |
74 | #include "gimple.h" | |
b20a8bb4 | 75 | #include "stringpool.h" |
7c29e30e | 76 | #include "cgraph.h" |
7c29e30e | 77 | #include "lto-streamer.h" |
bc61cadb | 78 | #include "tree-eh.h" |
073c1fd5 | 79 | #include "tree-cfg.h" |
c02159a7 | 80 | #include "tree-inline.h" |
930c75f4 | 81 | #include "dumpfile.h" |
c02159a7 | 82 | #include "gimple-pretty-print.h" |
c02159a7 | 83 | |
35ee1c66 | 84 | /* Create clone of edge in the node N represented by CALL_EXPR |
85 | the callgraph. */ | |
86 | ||
87 | cgraph_edge * | |
1a91d914 | 88 | cgraph_edge::clone (cgraph_node *n, gcall *call_stmt, unsigned stmt_uid, |
f7f3f5f6 | 89 | profile_count num, profile_count den, |
151b9ff5 | 90 | bool update_original) |
c02159a7 | 91 | { |
35ee1c66 | 92 | cgraph_edge *new_edge; |
151b9ff5 | 93 | profile_count::adjust_for_ipa_scaling (&num, &den); |
ed0831a9 | 94 | profile_count prof_count = count.apply_scale (num, den); |
c02159a7 | 95 | |
35ee1c66 | 96 | if (indirect_unknown_callee) |
c02159a7 | 97 | { |
98 | tree decl; | |
99 | ||
9ccaed86 | 100 | if (call_stmt && (decl = gimple_call_fndecl (call_stmt)) |
101 | /* When the call is speculative, we need to resolve it | |
102 | via cgraph_resolve_speculation and not here. */ | |
35ee1c66 | 103 | && !speculative) |
c02159a7 | 104 | { |
35ee1c66 | 105 | cgraph_node *callee = cgraph_node::get (decl); |
c02159a7 | 106 | gcc_checking_assert (callee); |
ed0831a9 | 107 | new_edge = n->create_edge (callee, call_stmt, prof_count); |
c02159a7 | 108 | } |
109 | else | |
110 | { | |
415d1b9a | 111 | new_edge = n->create_indirect_edge (call_stmt, |
35ee1c66 | 112 | indirect_info->ecf_flags, |
ed0831a9 | 113 | prof_count, false); |
35ee1c66 | 114 | *new_edge->indirect_info = *indirect_info; |
c02159a7 | 115 | } |
116 | } | |
117 | else | |
118 | { | |
ed0831a9 | 119 | new_edge = n->create_edge (callee, call_stmt, prof_count); |
35ee1c66 | 120 | if (indirect_info) |
c02159a7 | 121 | { |
122 | new_edge->indirect_info | |
25a27413 | 123 | = ggc_cleared_alloc<cgraph_indirect_call_info> (); |
35ee1c66 | 124 | *new_edge->indirect_info = *indirect_info; |
c02159a7 | 125 | } |
126 | } | |
127 | ||
35ee1c66 | 128 | new_edge->inline_failed = inline_failed; |
129 | new_edge->indirect_inlining_edge = indirect_inlining_edge; | |
c02159a7 | 130 | new_edge->lto_stmt_uid = stmt_uid; |
131 | /* Clone flags that depend on call_stmt availability manually. */ | |
35ee1c66 | 132 | new_edge->can_throw_external = can_throw_external; |
133 | new_edge->call_stmt_cannot_inline_p = call_stmt_cannot_inline_p; | |
134 | new_edge->speculative = speculative; | |
007a6c27 | 135 | new_edge->in_polymorphic_cdtor = in_polymorphic_cdtor; |
151b9ff5 | 136 | |
137 | /* Update IPA profile. Local profiles need no updating in original. */ | |
ed0831a9 | 138 | if (update_original) |
139 | count = count.combine_with_ipa_count (count.ipa () | |
140 | - new_edge->count.ipa ()); | |
35ee1c66 | 141 | symtab->call_edge_duplication_hooks (this, new_edge); |
c02159a7 | 142 | return new_edge; |
143 | } | |
144 | ||
98d52bcd | 145 | /* Build variant of function type ORIG_TYPE skipping ARGS_TO_SKIP and the |
146 | return value if SKIP_RETURN is true. */ | |
147 | ||
61896fab | 148 | tree |
149 | cgraph_build_function_type_skip_args (tree orig_type, bitmap args_to_skip, | |
150 | bool skip_return) | |
98d52bcd | 151 | { |
152 | tree new_type = NULL; | |
40e1cddf | 153 | tree args, new_args = NULL; |
98d52bcd | 154 | tree new_reversed; |
155 | int i = 0; | |
156 | ||
157 | for (args = TYPE_ARG_TYPES (orig_type); args && args != void_list_node; | |
158 | args = TREE_CHAIN (args), i++) | |
159 | if (!args_to_skip || !bitmap_bit_p (args_to_skip, i)) | |
160 | new_args = tree_cons (NULL_TREE, TREE_VALUE (args), new_args); | |
161 | ||
162 | new_reversed = nreverse (new_args); | |
163 | if (args) | |
164 | { | |
165 | if (new_reversed) | |
166 | TREE_CHAIN (new_args) = void_list_node; | |
167 | else | |
168 | new_reversed = void_list_node; | |
169 | } | |
170 | ||
171 | /* Use copy_node to preserve as much as possible from original type | |
172 | (debug info, attribute lists etc.) | |
173 | Exception is METHOD_TYPEs must have THIS argument. | |
174 | When we are asked to remove it, we need to build new FUNCTION_TYPE | |
175 | instead. */ | |
176 | if (TREE_CODE (orig_type) != METHOD_TYPE | |
177 | || !args_to_skip | |
178 | || !bitmap_bit_p (args_to_skip, 0)) | |
179 | { | |
180 | new_type = build_distinct_type_copy (orig_type); | |
181 | TYPE_ARG_TYPES (new_type) = new_reversed; | |
182 | } | |
183 | else | |
184 | { | |
185 | new_type | |
186 | = build_distinct_type_copy (build_function_type (TREE_TYPE (orig_type), | |
187 | new_reversed)); | |
188 | TYPE_CONTEXT (new_type) = TYPE_CONTEXT (orig_type); | |
189 | } | |
190 | ||
191 | if (skip_return) | |
192 | TREE_TYPE (new_type) = void_type_node; | |
193 | ||
98d52bcd | 194 | return new_type; |
195 | } | |
196 | ||
197 | /* Build variant of function decl ORIG_DECL skipping ARGS_TO_SKIP and the | |
198 | return value if SKIP_RETURN is true. | |
199 | ||
200 | Arguments from DECL_ARGUMENTS list can't be removed now, since they are | |
201 | linked by TREE_CHAIN directly. The caller is responsible for eliminating | |
202 | them when they are being duplicated (i.e. copy_arguments_for_versioning). */ | |
203 | ||
204 | static tree | |
205 | build_function_decl_skip_args (tree orig_decl, bitmap args_to_skip, | |
206 | bool skip_return) | |
207 | { | |
208 | tree new_decl = copy_node (orig_decl); | |
209 | tree new_type; | |
210 | ||
211 | new_type = TREE_TYPE (orig_decl); | |
212 | if (prototype_p (new_type) | |
213 | || (skip_return && !VOID_TYPE_P (TREE_TYPE (new_type)))) | |
214 | new_type | |
61896fab | 215 | = cgraph_build_function_type_skip_args (new_type, args_to_skip, |
216 | skip_return); | |
98d52bcd | 217 | TREE_TYPE (new_decl) = new_type; |
218 | ||
219 | /* For declarations setting DECL_VINDEX (i.e. methods) | |
220 | we expect first argument to be THIS pointer. */ | |
221 | if (args_to_skip && bitmap_bit_p (args_to_skip, 0)) | |
222 | DECL_VINDEX (new_decl) = NULL_TREE; | |
223 | ||
224 | /* When signature changes, we need to clear builtin info. */ | |
a0e9bfbb | 225 | if (fndecl_built_in_p (new_decl) |
98d52bcd | 226 | && args_to_skip |
227 | && !bitmap_empty_p (args_to_skip)) | |
228 | { | |
229 | DECL_BUILT_IN_CLASS (new_decl) = NOT_BUILT_IN; | |
230 | DECL_FUNCTION_CODE (new_decl) = (enum built_in_function) 0; | |
231 | } | |
232 | /* The FE might have information and assumptions about the other | |
233 | arguments. */ | |
234 | DECL_LANG_SPECIFIC (new_decl) = NULL; | |
235 | return new_decl; | |
236 | } | |
237 | ||
238 | /* Set flags of NEW_NODE and its decl. NEW_NODE is a newly created private | |
239 | clone or its thunk. */ | |
240 | ||
241 | static void | |
242 | set_new_clone_decl_and_node_flags (cgraph_node *new_node) | |
243 | { | |
244 | DECL_EXTERNAL (new_node->decl) = 0; | |
98d52bcd | 245 | TREE_PUBLIC (new_node->decl) = 0; |
246 | DECL_COMDAT (new_node->decl) = 0; | |
247 | DECL_WEAK (new_node->decl) = 0; | |
248 | DECL_VIRTUAL_P (new_node->decl) = 0; | |
249 | DECL_STATIC_CONSTRUCTOR (new_node->decl) = 0; | |
250 | DECL_STATIC_DESTRUCTOR (new_node->decl) = 0; | |
251 | ||
252 | new_node->externally_visible = 0; | |
253 | new_node->local.local = 1; | |
254 | new_node->lowered = true; | |
255 | } | |
256 | ||
257 | /* Duplicate thunk THUNK if necessary but make it to refer to NODE. | |
258 | ARGS_TO_SKIP, if non-NULL, determines which parameters should be omitted. | |
259 | Function can return NODE if no thunk is necessary, which can happen when | |
260 | thunk is this_adjusting but we are removing this parameter. */ | |
261 | ||
262 | static cgraph_node * | |
2e1475da | 263 | duplicate_thunk_for_node (cgraph_node *thunk, cgraph_node *node) |
98d52bcd | 264 | { |
265 | cgraph_node *new_thunk, *thunk_of; | |
415d1b9a | 266 | thunk_of = thunk->callees->callee->ultimate_alias_target (); |
98d52bcd | 267 | |
268 | if (thunk_of->thunk.thunk_p) | |
2e1475da | 269 | node = duplicate_thunk_for_node (thunk_of, node); |
98d52bcd | 270 | |
a3a6d499 | 271 | if (!DECL_ARGUMENTS (thunk->decl)) |
4560777c | 272 | thunk->get_untransformed_body (); |
a3a6d499 | 273 | |
35ee1c66 | 274 | cgraph_edge *cs; |
98d52bcd | 275 | for (cs = node->callers; cs; cs = cs->next_caller) |
276 | if (cs->caller->thunk.thunk_p | |
98d52bcd | 277 | && cs->caller->thunk.fixed_offset == thunk->thunk.fixed_offset |
4c7db812 | 278 | && cs->caller->thunk.virtual_value == thunk->thunk.virtual_value |
279 | && cs->caller->thunk.indirect_offset == thunk->thunk.indirect_offset | |
280 | && cs->caller->thunk.this_adjusting == thunk->thunk.this_adjusting | |
281 | && cs->caller->thunk.virtual_offset_p == thunk->thunk.virtual_offset_p) | |
98d52bcd | 282 | return cs->caller; |
283 | ||
284 | tree new_decl; | |
2e1475da | 285 | if (!node->clone.args_to_skip) |
98d52bcd | 286 | new_decl = copy_node (thunk->decl); |
287 | else | |
288 | { | |
289 | /* We do not need to duplicate this_adjusting thunks if we have removed | |
290 | this. */ | |
291 | if (thunk->thunk.this_adjusting | |
2e1475da | 292 | && bitmap_bit_p (node->clone.args_to_skip, 0)) |
98d52bcd | 293 | return node; |
294 | ||
2e1475da | 295 | new_decl = build_function_decl_skip_args (thunk->decl, |
296 | node->clone.args_to_skip, | |
98d52bcd | 297 | false); |
298 | } | |
d779c7b1 | 299 | |
300 | tree *link = &DECL_ARGUMENTS (new_decl); | |
301 | int i = 0; | |
302 | for (tree pd = DECL_ARGUMENTS (thunk->decl); pd; pd = DECL_CHAIN (pd), i++) | |
303 | { | |
304 | if (!node->clone.args_to_skip | |
305 | || !bitmap_bit_p (node->clone.args_to_skip, i)) | |
306 | { | |
307 | tree nd = copy_node (pd); | |
308 | DECL_CONTEXT (nd) = new_decl; | |
309 | *link = nd; | |
310 | link = &DECL_CHAIN (nd); | |
311 | } | |
312 | } | |
313 | *link = NULL_TREE; | |
314 | ||
98d52bcd | 315 | gcc_checking_assert (!DECL_STRUCT_FUNCTION (new_decl)); |
316 | gcc_checking_assert (!DECL_INITIAL (new_decl)); | |
317 | gcc_checking_assert (!DECL_RESULT (new_decl)); | |
318 | gcc_checking_assert (!DECL_RTL_SET_P (new_decl)); | |
319 | ||
87943388 | 320 | DECL_NAME (new_decl) = clone_function_name_numbered (thunk->decl, |
321 | "artificial_thunk"); | |
98d52bcd | 322 | SET_DECL_ASSEMBLER_NAME (new_decl, DECL_NAME (new_decl)); |
98d52bcd | 323 | |
e926b156 | 324 | /* We need to force DECL_IGNORED_P because the new thunk is created after |
325 | early debug was run. */ | |
326 | DECL_IGNORED_P (new_decl) = 1; | |
327 | ||
415d1b9a | 328 | new_thunk = cgraph_node::create (new_decl); |
98d52bcd | 329 | set_new_clone_decl_and_node_flags (new_thunk); |
330 | new_thunk->definition = true; | |
2ebf191d | 331 | new_thunk->local.can_change_signature = node->local.can_change_signature; |
98d52bcd | 332 | new_thunk->thunk = thunk->thunk; |
333 | new_thunk->unique_name = in_lto_p; | |
334 | new_thunk->former_clone_of = thunk->decl; | |
2e1475da | 335 | new_thunk->clone.args_to_skip = node->clone.args_to_skip; |
336 | new_thunk->clone.combined_args_to_skip = node->clone.combined_args_to_skip; | |
98d52bcd | 337 | |
151b9ff5 | 338 | cgraph_edge *e = new_thunk->create_edge (node, NULL, new_thunk->count); |
35ee1c66 | 339 | symtab->call_edge_duplication_hooks (thunk->callees, e); |
35ee1c66 | 340 | symtab->call_cgraph_duplication_hooks (thunk, new_thunk); |
98d52bcd | 341 | return new_thunk; |
342 | } | |
343 | ||
344 | /* If E does not lead to a thunk, simply redirect it to N. Otherwise create | |
345 | one or more equivalent thunks for N and redirect E to the first in the | |
fe3c9f1e | 346 | chain. Note that it is then necessary to call |
347 | n->expand_all_artificial_thunks once all callers are redirected. */ | |
98d52bcd | 348 | |
349 | void | |
fe3c9f1e | 350 | cgraph_edge::redirect_callee_duplicating_thunks (cgraph_node *n) |
98d52bcd | 351 | { |
fe3c9f1e | 352 | cgraph_node *orig_to = callee->ultimate_alias_target (); |
98d52bcd | 353 | if (orig_to->thunk.thunk_p) |
2e1475da | 354 | n = duplicate_thunk_for_node (orig_to, n); |
98d52bcd | 355 | |
fe3c9f1e | 356 | redirect_callee (n); |
357 | } | |
358 | ||
359 | /* Call expand_thunk on all callers that are thunks and if analyze those nodes | |
360 | that were expanded. */ | |
361 | ||
362 | void | |
363 | cgraph_node::expand_all_artificial_thunks () | |
364 | { | |
365 | cgraph_edge *e; | |
366 | for (e = callers; e;) | |
367 | if (e->caller->thunk.thunk_p) | |
368 | { | |
369 | cgraph_node *thunk = e->caller; | |
370 | ||
371 | e = e->next_caller; | |
372 | if (thunk->expand_thunk (false, false)) | |
373 | { | |
374 | thunk->thunk.thunk_p = false; | |
375 | thunk->analyze (); | |
376 | } | |
377 | thunk->expand_all_artificial_thunks (); | |
378 | } | |
379 | else | |
380 | e = e->next_caller; | |
98d52bcd | 381 | } |
c02159a7 | 382 | |
9dc70d59 | 383 | void |
384 | dump_callgraph_transformation (const cgraph_node *original, | |
385 | const cgraph_node *clone, | |
386 | const char *suffix) | |
387 | { | |
388 | if (symtab->ipa_clones_dump_file) | |
389 | { | |
390 | fprintf (symtab->ipa_clones_dump_file, | |
391 | "Callgraph clone;%s;%d;%s;%d;%d;%s;%d;%s;%d;%d;%s\n", | |
392 | original->asm_name (), original->order, | |
393 | DECL_SOURCE_FILE (original->decl), | |
394 | DECL_SOURCE_LINE (original->decl), | |
395 | DECL_SOURCE_COLUMN (original->decl), clone->asm_name (), | |
396 | clone->order, DECL_SOURCE_FILE (clone->decl), | |
397 | DECL_SOURCE_LINE (clone->decl), DECL_SOURCE_COLUMN (clone->decl), | |
398 | suffix); | |
399 | ||
400 | symtab->cloned_nodes.add (original); | |
401 | symtab->cloned_nodes.add (clone); | |
402 | } | |
403 | } | |
404 | ||
c02159a7 | 405 | /* Create node representing clone of N executed COUNT times. Decrease |
406 | the execution counts from original node too. | |
407 | The new clone will have decl set to DECL that may or may not be the same | |
408 | as decl of N. | |
409 | ||
410 | When UPDATE_ORIGINAL is true, the counts are subtracted from the original | |
411 | function's profile to reflect the fact that part of execution is handled | |
412 | by node. | |
413 | When CALL_DUPLICATOIN_HOOK is true, the ipa passes are acknowledged about | |
48f42a9a | 414 | the new clone. Otherwise the caller is responsible for doing so later. |
415 | ||
416 | If the new node is being inlined into another one, NEW_INLINED_TO should be | |
417 | the outline function the new one is (even indirectly) inlined to. All hooks | |
418 | will see this in node's global.inlined_to, when invoked. Can be NULL if the | |
419 | node is not inlined. */ | |
c02159a7 | 420 | |
415d1b9a | 421 | cgraph_node * |
151b9ff5 | 422 | cgraph_node::create_clone (tree new_decl, profile_count prof_count, |
415d1b9a | 423 | bool update_original, |
424 | vec<cgraph_edge *> redirect_callers, | |
425 | bool call_duplication_hook, | |
35ee1c66 | 426 | cgraph_node *new_inlined_to, |
9dc70d59 | 427 | bitmap args_to_skip, const char *suffix) |
c02159a7 | 428 | { |
35ee1c66 | 429 | cgraph_node *new_node = symtab->create_empty (); |
430 | cgraph_edge *e; | |
c02159a7 | 431 | unsigned i; |
151b9ff5 | 432 | profile_count old_count = count; |
c02159a7 | 433 | |
9dc70d59 | 434 | if (new_inlined_to) |
435 | dump_callgraph_transformation (this, new_inlined_to, "inlining to"); | |
436 | ||
6251b3d9 | 437 | /* When inlining we scale precisely to prof_count, when cloning we can |
438 | preserve local profile. */ | |
439 | if (!new_inlined_to) | |
440 | prof_count = count.combine_with_ipa_count (prof_count); | |
db9cef39 | 441 | new_node->count = prof_count; |
151b9ff5 | 442 | |
443 | /* Update IPA profile. Local profiles need no updating in original. */ | |
ed0831a9 | 444 | if (update_original) |
445 | count = count.combine_with_ipa_count (count.ipa () - prof_count.ipa ()); | |
c28f0245 | 446 | new_node->decl = new_decl; |
415d1b9a | 447 | new_node->register_symbol (); |
448 | new_node->origin = origin; | |
449 | new_node->lto_file_data = lto_file_data; | |
c02159a7 | 450 | if (new_node->origin) |
451 | { | |
452 | new_node->next_nested = new_node->origin->nested; | |
453 | new_node->origin->nested = new_node; | |
454 | } | |
415d1b9a | 455 | new_node->analyzed = analyzed; |
456 | new_node->definition = definition; | |
457 | new_node->local = local; | |
02774f2d | 458 | new_node->externally_visible = false; |
6b722052 | 459 | new_node->no_reorder = no_reorder; |
c02159a7 | 460 | new_node->local.local = true; |
415d1b9a | 461 | new_node->global = global; |
48f42a9a | 462 | new_node->global.inlined_to = new_inlined_to; |
415d1b9a | 463 | new_node->rtl = rtl; |
415d1b9a | 464 | new_node->frequency = frequency; |
465 | new_node->tp_first_run = tp_first_run; | |
f852c239 | 466 | new_node->tm_clone = tm_clone; |
ce7711df | 467 | new_node->icf_merged = icf_merged; |
7d38b7bc | 468 | new_node->merged_comdat = merged_comdat; |
76e435b5 | 469 | new_node->thunk = thunk; |
2e1475da | 470 | |
471 | new_node->clone.tree_map = NULL; | |
472 | new_node->clone.args_to_skip = args_to_skip; | |
c28f0245 | 473 | new_node->split_part = split_part; |
2e1475da | 474 | if (!args_to_skip) |
415d1b9a | 475 | new_node->clone.combined_args_to_skip = clone.combined_args_to_skip; |
476 | else if (clone.combined_args_to_skip) | |
2e1475da | 477 | { |
478 | new_node->clone.combined_args_to_skip = BITMAP_GGC_ALLOC (); | |
479 | bitmap_ior (new_node->clone.combined_args_to_skip, | |
415d1b9a | 480 | clone.combined_args_to_skip, args_to_skip); |
2e1475da | 481 | } |
482 | else | |
483 | new_node->clone.combined_args_to_skip = args_to_skip; | |
484 | ||
f1f41a6c | 485 | FOR_EACH_VEC_ELT (redirect_callers, i, e) |
c02159a7 | 486 | { |
487 | /* Redirect calls to the old version node to point to its new | |
5a7ad253 | 488 | version. The only exception is when the edge was proved to |
489 | be unreachable during the clonning procedure. */ | |
490 | if (!e->callee | |
a0e9bfbb | 491 | || !fndecl_built_in_p (e->callee->decl, BUILT_IN_UNREACHABLE)) |
fe3c9f1e | 492 | e->redirect_callee_duplicating_thunks (new_node); |
c02159a7 | 493 | } |
fe3c9f1e | 494 | new_node->expand_all_artificial_thunks (); |
c02159a7 | 495 | |
415d1b9a | 496 | for (e = callees;e; e=e->next_callee) |
151b9ff5 | 497 | e->clone (new_node, e->call_stmt, e->lto_stmt_uid, new_node->count, old_count, |
498 | update_original); | |
c02159a7 | 499 | |
415d1b9a | 500 | for (e = indirect_calls; e; e = e->next_callee) |
35ee1c66 | 501 | e->clone (new_node, e->call_stmt, e->lto_stmt_uid, |
151b9ff5 | 502 | new_node->count, old_count, update_original); |
415d1b9a | 503 | new_node->clone_references (this); |
c02159a7 | 504 | |
415d1b9a | 505 | new_node->next_sibling_clone = clones; |
506 | if (clones) | |
507 | clones->prev_sibling_clone = new_node; | |
508 | clones = new_node; | |
509 | new_node->clone_of = this; | |
c02159a7 | 510 | |
511 | if (call_duplication_hook) | |
35ee1c66 | 512 | symtab->call_cgraph_duplication_hooks (this, new_node); |
9dc70d59 | 513 | |
514 | if (!new_inlined_to) | |
515 | dump_callgraph_transformation (this, new_node, suffix); | |
516 | ||
c02159a7 | 517 | return new_node; |
518 | } | |
519 | ||
194b1f6b | 520 | static GTY(()) hash_map<const char *, unsigned> *clone_fn_ids; |
c02159a7 | 521 | |
87943388 | 522 | /* Return a new assembler name for a clone of decl named NAME. Apart |
523 | from the string SUFFIX, the new name will end with a unique (for | |
524 | each NAME) unspecified number. If clone numbering is not needed | |
525 | then the two argument clone_function_name should be used instead. | |
526 | Should not be called directly except for by | |
527 | lto-partition.c:privatize_symbol_name_1. */ | |
a73afd90 | 528 | |
c02159a7 | 529 | tree |
87943388 | 530 | clone_function_name_numbered (const char *name, const char *suffix) |
531 | { | |
194b1f6b | 532 | /* Initialize the function->counter mapping the first time it's |
533 | needed. */ | |
534 | if (!clone_fn_ids) | |
535 | clone_fn_ids = hash_map<const char *, unsigned int>::create_ggc (64); | |
536 | unsigned int &suffix_counter = clone_fn_ids->get_or_insert ( | |
537 | IDENTIFIER_POINTER (get_identifier (name))); | |
538 | return clone_function_name (name, suffix, suffix_counter++); | |
87943388 | 539 | } |
540 | ||
541 | /* Return a new assembler name for a clone of DECL. Apart from string | |
542 | SUFFIX, the new name will end with a unique (for each DECL | |
543 | assembler name) unspecified number. If clone numbering is not | |
544 | needed then the two argument clone_function_name should be used | |
545 | instead. */ | |
546 | ||
547 | tree | |
548 | clone_function_name_numbered (tree decl, const char *suffix) | |
549 | { | |
550 | tree name = DECL_ASSEMBLER_NAME (decl); | |
551 | return clone_function_name_numbered (IDENTIFIER_POINTER (name), | |
552 | suffix); | |
553 | } | |
554 | ||
555 | /* Return a new assembler name for a clone of decl named NAME. Apart | |
556 | from the string SUFFIX, the new name will end with the specified | |
557 | NUMBER. If clone numbering is not needed then the two argument | |
558 | clone_function_name should be used instead. */ | |
559 | ||
560 | tree | |
561 | clone_function_name (const char *name, const char *suffix, | |
562 | unsigned long number) | |
c02159a7 | 563 | { |
a73afd90 | 564 | size_t len = strlen (name); |
c02159a7 | 565 | char *tmp_name, *prefix; |
566 | ||
567 | prefix = XALLOCAVEC (char, len + strlen (suffix) + 2); | |
a73afd90 | 568 | memcpy (prefix, name, len); |
c02159a7 | 569 | strcpy (prefix + len + 1, suffix); |
11075c28 | 570 | prefix[len] = symbol_table::symbol_suffix_separator (); |
87943388 | 571 | ASM_FORMAT_PRIVATE_NAME (tmp_name, prefix, number); |
c02159a7 | 572 | return get_identifier (tmp_name); |
573 | } | |
574 | ||
9175d3dd | 575 | /* Return a new assembler name for a clone of DECL. Apart from the |
576 | string SUFFIX, the new name will end with the specified NUMBER. If | |
577 | clone numbering is not needed then the two argument | |
578 | clone_function_name should be used instead. */ | |
579 | ||
580 | tree | |
581 | clone_function_name (tree decl, const char *suffix, | |
582 | unsigned long number) | |
583 | { | |
584 | return clone_function_name ( | |
585 | IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (decl)), suffix, number); | |
586 | } | |
587 | ||
87943388 | 588 | /* Return a new assembler name ending with the string SUFFIX for a |
589 | clone of DECL. */ | |
a73afd90 | 590 | |
591 | tree | |
592 | clone_function_name (tree decl, const char *suffix) | |
593 | { | |
87943388 | 594 | tree identifier = DECL_ASSEMBLER_NAME (decl); |
595 | /* For consistency this needs to behave the same way as | |
596 | ASM_FORMAT_PRIVATE_NAME does, but without the final number | |
597 | suffix. */ | |
598 | char *separator = XALLOCAVEC (char, 2); | |
599 | separator[0] = symbol_table::symbol_suffix_separator (); | |
600 | separator[1] = 0; | |
601 | #if defined (NO_DOT_IN_LABEL) && defined (NO_DOLLAR_IN_LABEL) | |
602 | const char *prefix = "__"; | |
603 | #else | |
604 | const char *prefix = ""; | |
605 | #endif | |
606 | char *result = ACONCAT ((prefix, | |
607 | IDENTIFIER_POINTER (identifier), | |
608 | separator, | |
609 | suffix, | |
610 | (char*)0)); | |
611 | return get_identifier (result); | |
a73afd90 | 612 | } |
613 | ||
614 | ||
9175d3dd | 615 | /* Create callgraph node clone with new declaration. The actual body will be |
616 | copied later at compilation stage. The name of the new clone will be | |
617 | constructed from the name of the original node, SUFFIX and NUM_SUFFIX. | |
c02159a7 | 618 | |
619 | TODO: after merging in ipa-sra use function call notes instead of args_to_skip | |
620 | bitmap interface. | |
621 | */ | |
35ee1c66 | 622 | cgraph_node * |
415d1b9a | 623 | cgraph_node::create_virtual_clone (vec<cgraph_edge *> redirect_callers, |
624 | vec<ipa_replace_map *, va_gc> *tree_map, | |
9175d3dd | 625 | bitmap args_to_skip, const char * suffix, |
626 | unsigned num_suffix) | |
c02159a7 | 627 | { |
415d1b9a | 628 | tree old_decl = decl; |
35ee1c66 | 629 | cgraph_node *new_node = NULL; |
c02159a7 | 630 | tree new_decl; |
cac1ceba | 631 | size_t len, i; |
35ee1c66 | 632 | ipa_replace_map *map; |
cac1ceba | 633 | char *name; |
c02159a7 | 634 | |
b8681788 | 635 | gcc_checking_assert (local.versionable); |
415d1b9a | 636 | gcc_assert (local.can_change_signature || !args_to_skip); |
c02159a7 | 637 | |
638 | /* Make a new FUNCTION_DECL tree node */ | |
639 | if (!args_to_skip) | |
640 | new_decl = copy_node (old_decl); | |
641 | else | |
642 | new_decl = build_function_decl_skip_args (old_decl, args_to_skip, false); | |
79e830ee | 643 | |
644 | /* These pointers represent function body and will be populated only when clone | |
645 | is materialized. */ | |
646 | gcc_assert (new_decl != old_decl); | |
c02159a7 | 647 | DECL_STRUCT_FUNCTION (new_decl) = NULL; |
79e830ee | 648 | DECL_ARGUMENTS (new_decl) = NULL; |
649 | DECL_INITIAL (new_decl) = NULL; | |
650 | DECL_RESULT (new_decl) = NULL; | |
f4d3c071 | 651 | /* We cannot do DECL_RESULT (new_decl) = NULL; here because of LTO partitioning |
79e830ee | 652 | sometimes storing only clone decl instead of original. */ |
c02159a7 | 653 | |
654 | /* Generate a new name for the new version. */ | |
cac1ceba | 655 | len = IDENTIFIER_LENGTH (DECL_NAME (old_decl)); |
656 | name = XALLOCAVEC (char, len + strlen (suffix) + 2); | |
657 | memcpy (name, IDENTIFIER_POINTER (DECL_NAME (old_decl)), len); | |
658 | strcpy (name + len + 1, suffix); | |
659 | name[len] = '.'; | |
660 | DECL_NAME (new_decl) = get_identifier (name); | |
9175d3dd | 661 | SET_DECL_ASSEMBLER_NAME (new_decl, |
662 | clone_function_name (old_decl, suffix, num_suffix)); | |
c02159a7 | 663 | SET_DECL_RTL (new_decl, NULL); |
664 | ||
151b9ff5 | 665 | new_node = create_clone (new_decl, count, false, |
9dc70d59 | 666 | redirect_callers, false, NULL, args_to_skip, suffix); |
415d1b9a | 667 | |
c02159a7 | 668 | /* Update the properties. |
669 | Make clone visible only within this translation unit. Make sure | |
670 | that is not weak also. | |
671 | ??? We cannot use COMDAT linkage because there is no | |
672 | ABI support for this. */ | |
98d52bcd | 673 | set_new_clone_decl_and_node_flags (new_node); |
c02159a7 | 674 | new_node->clone.tree_map = tree_map; |
033ab5d3 | 675 | if (!implicit_section) |
022a2340 | 676 | new_node->set_section (get_section ()); |
5c368d8a | 677 | |
678 | /* Clones of global symbols or symbols with unique names are unique. */ | |
679 | if ((TREE_PUBLIC (old_decl) | |
680 | && !DECL_EXTERNAL (old_decl) | |
681 | && !DECL_WEAK (old_decl) | |
682 | && !DECL_COMDAT (old_decl)) | |
683 | || in_lto_p) | |
02774f2d | 684 | new_node->unique_name = true; |
f1f41a6c | 685 | FOR_EACH_VEC_SAFE_ELT (tree_map, i, map) |
7dab0c61 | 686 | new_node->maybe_create_reference (map->new_tree, NULL); |
2e1475da | 687 | |
415d1b9a | 688 | if (ipa_transforms_to_apply.exists ()) |
2cf4c3b3 | 689 | new_node->ipa_transforms_to_apply |
415d1b9a | 690 | = ipa_transforms_to_apply.copy (); |
c02159a7 | 691 | |
35ee1c66 | 692 | symtab->call_cgraph_duplication_hooks (this, new_node); |
c02159a7 | 693 | |
694 | return new_node; | |
695 | } | |
696 | ||
415d1b9a | 697 | /* callgraph node being removed from symbol table; see if its entry can be |
698 | replaced by other inline clone. */ | |
699 | cgraph_node * | |
700 | cgraph_node::find_replacement (void) | |
c02159a7 | 701 | { |
35ee1c66 | 702 | cgraph_node *next_inline_clone, *replacement; |
c02159a7 | 703 | |
415d1b9a | 704 | for (next_inline_clone = clones; |
c02159a7 | 705 | next_inline_clone |
415d1b9a | 706 | && next_inline_clone->decl != decl; |
c02159a7 | 707 | next_inline_clone = next_inline_clone->next_sibling_clone) |
708 | ; | |
709 | ||
710 | /* If there is inline clone of the node being removed, we need | |
711 | to put it into the position of removed node and reorganize all | |
712 | other clones to be based on it. */ | |
713 | if (next_inline_clone) | |
714 | { | |
35ee1c66 | 715 | cgraph_node *n; |
716 | cgraph_node *new_clones; | |
c02159a7 | 717 | |
718 | replacement = next_inline_clone; | |
719 | ||
720 | /* Unlink inline clone from the list of clones of removed node. */ | |
721 | if (next_inline_clone->next_sibling_clone) | |
722 | next_inline_clone->next_sibling_clone->prev_sibling_clone | |
723 | = next_inline_clone->prev_sibling_clone; | |
724 | if (next_inline_clone->prev_sibling_clone) | |
725 | { | |
415d1b9a | 726 | gcc_assert (clones != next_inline_clone); |
c02159a7 | 727 | next_inline_clone->prev_sibling_clone->next_sibling_clone |
728 | = next_inline_clone->next_sibling_clone; | |
729 | } | |
730 | else | |
731 | { | |
415d1b9a | 732 | gcc_assert (clones == next_inline_clone); |
733 | clones = next_inline_clone->next_sibling_clone; | |
c02159a7 | 734 | } |
735 | ||
415d1b9a | 736 | new_clones = clones; |
737 | clones = NULL; | |
c02159a7 | 738 | |
739 | /* Copy clone info. */ | |
415d1b9a | 740 | next_inline_clone->clone = clone; |
c02159a7 | 741 | |
742 | /* Now place it into clone tree at same level at NODE. */ | |
415d1b9a | 743 | next_inline_clone->clone_of = clone_of; |
c02159a7 | 744 | next_inline_clone->prev_sibling_clone = NULL; |
745 | next_inline_clone->next_sibling_clone = NULL; | |
415d1b9a | 746 | if (clone_of) |
c02159a7 | 747 | { |
415d1b9a | 748 | if (clone_of->clones) |
749 | clone_of->clones->prev_sibling_clone = next_inline_clone; | |
750 | next_inline_clone->next_sibling_clone = clone_of->clones; | |
751 | clone_of->clones = next_inline_clone; | |
c02159a7 | 752 | } |
753 | ||
754 | /* Merge the clone list. */ | |
755 | if (new_clones) | |
756 | { | |
757 | if (!next_inline_clone->clones) | |
758 | next_inline_clone->clones = new_clones; | |
759 | else | |
760 | { | |
761 | n = next_inline_clone->clones; | |
762 | while (n->next_sibling_clone) | |
415d1b9a | 763 | n = n->next_sibling_clone; |
c02159a7 | 764 | n->next_sibling_clone = new_clones; |
765 | new_clones->prev_sibling_clone = n; | |
766 | } | |
767 | } | |
768 | ||
769 | /* Update clone_of pointers. */ | |
770 | n = new_clones; | |
771 | while (n) | |
772 | { | |
773 | n->clone_of = next_inline_clone; | |
774 | n = n->next_sibling_clone; | |
775 | } | |
776 | return replacement; | |
777 | } | |
778 | else | |
779 | return NULL; | |
780 | } | |
781 | ||
782 | /* Like cgraph_set_call_stmt but walk the clone tree and update all | |
4d044066 | 783 | clones sharing the same function body. |
784 | When WHOLE_SPECULATIVE_EDGES is true, all three components of | |
785 | speculative edge gets updated. Otherwise we update only direct | |
786 | call. */ | |
c02159a7 | 787 | |
788 | void | |
42acab1c | 789 | cgraph_node::set_call_stmt_including_clones (gimple *old_stmt, |
1a91d914 | 790 | gcall *new_stmt, |
415d1b9a | 791 | bool update_speculative) |
c02159a7 | 792 | { |
35ee1c66 | 793 | cgraph_node *node; |
794 | cgraph_edge *edge = get_edge (old_stmt); | |
c02159a7 | 795 | |
796 | if (edge) | |
35ee1c66 | 797 | edge->set_call_stmt (new_stmt, update_speculative); |
c02159a7 | 798 | |
415d1b9a | 799 | node = clones; |
c02159a7 | 800 | if (node) |
415d1b9a | 801 | while (node != this) |
c02159a7 | 802 | { |
35ee1c66 | 803 | cgraph_edge *edge = node->get_edge (old_stmt); |
c02159a7 | 804 | if (edge) |
4d044066 | 805 | { |
35ee1c66 | 806 | edge->set_call_stmt (new_stmt, update_speculative); |
4d044066 | 807 | /* If UPDATE_SPECULATIVE is false, it means that we are turning |
808 | speculative call into a real code sequence. Update the | |
809 | callgraph edges. */ | |
810 | if (edge->speculative && !update_speculative) | |
811 | { | |
35ee1c66 | 812 | cgraph_edge *direct, *indirect; |
813 | ipa_ref *ref; | |
4d044066 | 814 | |
815 | gcc_assert (!edge->indirect_unknown_callee); | |
35ee1c66 | 816 | edge->speculative_call_info (direct, indirect, ref); |
4d044066 | 817 | direct->speculative = false; |
818 | indirect->speculative = false; | |
819 | ref->speculative = false; | |
820 | } | |
821 | } | |
c02159a7 | 822 | if (node->clones) |
823 | node = node->clones; | |
824 | else if (node->next_sibling_clone) | |
825 | node = node->next_sibling_clone; | |
826 | else | |
827 | { | |
415d1b9a | 828 | while (node != this && !node->next_sibling_clone) |
c02159a7 | 829 | node = node->clone_of; |
415d1b9a | 830 | if (node != this) |
c02159a7 | 831 | node = node->next_sibling_clone; |
832 | } | |
833 | } | |
834 | } | |
835 | ||
836 | /* Like cgraph_create_edge walk the clone tree and update all clones sharing | |
837 | same function body. If clones already have edge for OLD_STMT; only | |
838 | update the edge same way as cgraph_set_call_stmt_including_clones does. | |
839 | ||
840 | TODO: COUNT and LOOP_DEPTH should be properly distributed based on relative | |
841 | frequencies of the clones. */ | |
842 | ||
843 | void | |
35ee1c66 | 844 | cgraph_node::create_edge_including_clones (cgraph_node *callee, |
42acab1c | 845 | gimple *old_stmt, gcall *stmt, |
db9cef39 | 846 | profile_count count, |
415d1b9a | 847 | cgraph_inline_failed_t reason) |
c02159a7 | 848 | { |
35ee1c66 | 849 | cgraph_node *node; |
850 | cgraph_edge *edge; | |
c02159a7 | 851 | |
415d1b9a | 852 | if (!get_edge (stmt)) |
c02159a7 | 853 | { |
151b9ff5 | 854 | edge = create_edge (callee, stmt, count); |
c02159a7 | 855 | edge->inline_failed = reason; |
856 | } | |
857 | ||
415d1b9a | 858 | node = clones; |
c02159a7 | 859 | if (node) |
415d1b9a | 860 | while (node != this) |
6a353dda | 861 | /* Thunk clones do not get updated while copying inline function body. */ |
862 | if (!node->thunk.thunk_p) | |
863 | { | |
864 | cgraph_edge *edge = node->get_edge (old_stmt); | |
865 | ||
866 | /* It is possible that clones already contain the edge while | |
867 | master didn't. Either we promoted indirect call into direct | |
868 | call in the clone or we are processing clones of unreachable | |
869 | master where edges has been removed. */ | |
870 | if (edge) | |
871 | edge->set_call_stmt (stmt); | |
872 | else if (! node->get_edge (stmt)) | |
873 | { | |
151b9ff5 | 874 | edge = node->create_edge (callee, stmt, count); |
6a353dda | 875 | edge->inline_failed = reason; |
876 | } | |
c02159a7 | 877 | |
6a353dda | 878 | if (node->clones) |
879 | node = node->clones; | |
880 | else if (node->next_sibling_clone) | |
881 | node = node->next_sibling_clone; | |
882 | else | |
883 | { | |
884 | while (node != this && !node->next_sibling_clone) | |
885 | node = node->clone_of; | |
886 | if (node != this) | |
887 | node = node->next_sibling_clone; | |
888 | } | |
889 | } | |
c02159a7 | 890 | } |
891 | ||
892 | /* Remove the node from cgraph and all inline clones inlined into it. | |
893 | Skip however removal of FORBIDDEN_NODE and return true if it needs to be | |
894 | removed. This allows to call the function from outer loop walking clone | |
895 | tree. */ | |
896 | ||
897 | bool | |
415d1b9a | 898 | cgraph_node::remove_symbol_and_inline_clones (cgraph_node *forbidden_node) |
c02159a7 | 899 | { |
35ee1c66 | 900 | cgraph_edge *e, *next; |
c02159a7 | 901 | bool found = false; |
902 | ||
415d1b9a | 903 | if (this == forbidden_node) |
8afc926b | 904 | { |
35ee1c66 | 905 | callers->remove (); |
8afc926b | 906 | return true; |
907 | } | |
415d1b9a | 908 | for (e = callees; e; e = next) |
c02159a7 | 909 | { |
910 | next = e->next_callee; | |
911 | if (!e->inline_failed) | |
415d1b9a | 912 | found |= e->callee->remove_symbol_and_inline_clones (forbidden_node); |
c02159a7 | 913 | } |
415d1b9a | 914 | remove (); |
c02159a7 | 915 | return found; |
916 | } | |
917 | ||
918 | /* The edges representing the callers of the NEW_VERSION node were | |
919 | fixed by cgraph_function_versioning (), now the call_expr in their | |
920 | respective tree code should be updated to call the NEW_VERSION. */ | |
921 | ||
922 | static void | |
35ee1c66 | 923 | update_call_expr (cgraph_node *new_version) |
c02159a7 | 924 | { |
35ee1c66 | 925 | cgraph_edge *e; |
c02159a7 | 926 | |
927 | gcc_assert (new_version); | |
928 | ||
929 | /* Update the call expr on the edges to call the new version. */ | |
930 | for (e = new_version->callers; e; e = e->next_caller) | |
931 | { | |
35ee1c66 | 932 | function *inner_function = DECL_STRUCT_FUNCTION (e->caller->decl); |
02774f2d | 933 | gimple_call_set_fndecl (e->call_stmt, new_version->decl); |
c02159a7 | 934 | maybe_clean_eh_stmt_fn (inner_function, e->call_stmt); |
935 | } | |
936 | } | |
937 | ||
938 | ||
939 | /* Create a new cgraph node which is the new version of | |
415d1b9a | 940 | callgraph node. REDIRECT_CALLERS holds the callers |
c02159a7 | 941 | edges which should be redirected to point to |
415d1b9a | 942 | NEW_VERSION. ALL the callees edges of the node |
c02159a7 | 943 | are cloned to the new version node. Return the new |
944 | version node. | |
945 | ||
946 | If non-NULL BLOCK_TO_COPY determine what basic blocks | |
947 | was copied to prevent duplications of calls that are dead | |
948 | in the clone. */ | |
949 | ||
415d1b9a | 950 | cgraph_node * |
951 | cgraph_node::create_version_clone (tree new_decl, | |
952 | vec<cgraph_edge *> redirect_callers, | |
9dc70d59 | 953 | bitmap bbs_to_copy, |
954 | const char *suffix) | |
c02159a7 | 955 | { |
35ee1c66 | 956 | cgraph_node *new_version; |
957 | cgraph_edge *e; | |
c02159a7 | 958 | unsigned i; |
959 | ||
415d1b9a | 960 | new_version = cgraph_node::create (new_decl); |
c02159a7 | 961 | |
415d1b9a | 962 | new_version->analyzed = analyzed; |
963 | new_version->definition = definition; | |
964 | new_version->local = local; | |
02774f2d | 965 | new_version->externally_visible = false; |
6b722052 | 966 | new_version->no_reorder = no_reorder; |
02774f2d | 967 | new_version->local.local = new_version->definition; |
415d1b9a | 968 | new_version->global = global; |
969 | new_version->rtl = rtl; | |
970 | new_version->count = count; | |
c02159a7 | 971 | |
415d1b9a | 972 | for (e = callees; e; e=e->next_callee) |
c02159a7 | 973 | if (!bbs_to_copy |
974 | || bitmap_bit_p (bbs_to_copy, gimple_bb (e->call_stmt)->index)) | |
35ee1c66 | 975 | e->clone (new_version, e->call_stmt, |
f7f3f5f6 | 976 | e->lto_stmt_uid, count, count, |
35ee1c66 | 977 | true); |
415d1b9a | 978 | for (e = indirect_calls; e; e=e->next_callee) |
c02159a7 | 979 | if (!bbs_to_copy |
980 | || bitmap_bit_p (bbs_to_copy, gimple_bb (e->call_stmt)->index)) | |
35ee1c66 | 981 | e->clone (new_version, e->call_stmt, |
f7f3f5f6 | 982 | e->lto_stmt_uid, count, count, |
35ee1c66 | 983 | true); |
f1f41a6c | 984 | FOR_EACH_VEC_ELT (redirect_callers, i, e) |
c02159a7 | 985 | { |
986 | /* Redirect calls to the old version node to point to its new | |
987 | version. */ | |
35ee1c66 | 988 | e->redirect_callee (new_version); |
c02159a7 | 989 | } |
990 | ||
35ee1c66 | 991 | symtab->call_cgraph_duplication_hooks (this, new_version); |
c02159a7 | 992 | |
9dc70d59 | 993 | dump_callgraph_transformation (this, new_version, suffix); |
994 | ||
c02159a7 | 995 | return new_version; |
996 | } | |
997 | ||
998 | /* Perform function versioning. | |
999 | Function versioning includes copying of the tree and | |
1000 | a callgraph update (creating a new cgraph node and updating | |
1001 | its callees and callers). | |
1002 | ||
1003 | REDIRECT_CALLERS varray includes the edges to be redirected | |
1004 | to the new version. | |
1005 | ||
1006 | TREE_MAP is a mapping of tree nodes we want to replace with | |
1007 | new ones (according to results of prior analysis). | |
c02159a7 | 1008 | |
1009 | If non-NULL ARGS_TO_SKIP determine function parameters to remove | |
1010 | from new version. | |
1011 | If SKIP_RETURN is true, the new version will return void. | |
1012 | If non-NULL BLOCK_TO_COPY determine what basic blocks to copy. | |
1013 | If non_NULL NEW_ENTRY determine new entry BB of the clone. | |
1014 | ||
1015 | Return the new version's cgraph node. */ | |
1016 | ||
415d1b9a | 1017 | cgraph_node * |
1018 | cgraph_node::create_version_clone_with_body | |
1019 | (vec<cgraph_edge *> redirect_callers, | |
1020 | vec<ipa_replace_map *, va_gc> *tree_map, bitmap args_to_skip, | |
1021 | bool skip_return, bitmap bbs_to_copy, basic_block new_entry_block, | |
9dc70d59 | 1022 | const char *suffix) |
c02159a7 | 1023 | { |
415d1b9a | 1024 | tree old_decl = decl; |
35ee1c66 | 1025 | cgraph_node *new_version_node = NULL; |
c02159a7 | 1026 | tree new_decl; |
1027 | ||
1028 | if (!tree_versionable_function_p (old_decl)) | |
1029 | return NULL; | |
1030 | ||
415d1b9a | 1031 | gcc_assert (local.can_change_signature || !args_to_skip); |
c02159a7 | 1032 | |
1033 | /* Make a new FUNCTION_DECL tree node for the new version. */ | |
1034 | if (!args_to_skip && !skip_return) | |
1035 | new_decl = copy_node (old_decl); | |
1036 | else | |
1037 | new_decl | |
1038 | = build_function_decl_skip_args (old_decl, args_to_skip, skip_return); | |
1039 | ||
1040 | /* Generate a new name for the new version. */ | |
87943388 | 1041 | DECL_NAME (new_decl) = clone_function_name_numbered (old_decl, suffix); |
c02159a7 | 1042 | SET_DECL_ASSEMBLER_NAME (new_decl, DECL_NAME (new_decl)); |
1043 | SET_DECL_RTL (new_decl, NULL); | |
1044 | ||
4728cefc | 1045 | DECL_VIRTUAL_P (new_decl) = 0; |
1046 | ||
c02159a7 | 1047 | /* When the old decl was a con-/destructor make sure the clone isn't. */ |
9af5ce0c | 1048 | DECL_STATIC_CONSTRUCTOR (new_decl) = 0; |
1049 | DECL_STATIC_DESTRUCTOR (new_decl) = 0; | |
c02159a7 | 1050 | |
1051 | /* Create the new version's call-graph node. | |
1052 | and update the edges of the new node. */ | |
415d1b9a | 1053 | new_version_node = create_version_clone (new_decl, redirect_callers, |
9dc70d59 | 1054 | bbs_to_copy, suffix); |
c02159a7 | 1055 | |
415d1b9a | 1056 | if (ipa_transforms_to_apply.exists ()) |
9be0994b | 1057 | new_version_node->ipa_transforms_to_apply |
415d1b9a | 1058 | = ipa_transforms_to_apply.copy (); |
c02159a7 | 1059 | /* Copy the OLD_VERSION_NODE function tree to the new version. */ |
1060 | tree_function_versioning (old_decl, new_decl, tree_map, false, args_to_skip, | |
1061 | skip_return, bbs_to_copy, new_entry_block); | |
1062 | ||
1063 | /* Update the new version's properties. | |
1064 | Make The new version visible only within this translation unit. Make sure | |
1065 | that is not weak also. | |
1066 | ??? We cannot use COMDAT linkage because there is no | |
1067 | ABI support for this. */ | |
415d1b9a | 1068 | new_version_node->make_decl_local (); |
02774f2d | 1069 | DECL_VIRTUAL_P (new_version_node->decl) = 0; |
1070 | new_version_node->externally_visible = 0; | |
c02159a7 | 1071 | new_version_node->local.local = 1; |
1072 | new_version_node->lowered = true; | |
033ab5d3 | 1073 | if (!implicit_section) |
1074 | new_version_node->set_section (get_section ()); | |
5c368d8a | 1075 | /* Clones of global symbols or symbols with unique names are unique. */ |
1076 | if ((TREE_PUBLIC (old_decl) | |
1077 | && !DECL_EXTERNAL (old_decl) | |
1078 | && !DECL_WEAK (old_decl) | |
1079 | && !DECL_COMDAT (old_decl)) | |
1080 | || in_lto_p) | |
02774f2d | 1081 | new_version_node->unique_name = true; |
c02159a7 | 1082 | |
1083 | /* Update the call_expr on the edges to call the new version node. */ | |
1084 | update_call_expr (new_version_node); | |
1085 | ||
b136b820 | 1086 | symtab->call_cgraph_insertion_hooks (new_version_node); |
c02159a7 | 1087 | return new_version_node; |
1088 | } | |
1089 | ||
1090 | /* Given virtual clone, turn it into actual clone. */ | |
1091 | ||
1092 | static void | |
35ee1c66 | 1093 | cgraph_materialize_clone (cgraph_node *node) |
c02159a7 | 1094 | { |
1095 | bitmap_obstack_initialize (NULL); | |
02774f2d | 1096 | node->former_clone_of = node->clone_of->decl; |
c02159a7 | 1097 | if (node->clone_of->former_clone_of) |
1098 | node->former_clone_of = node->clone_of->former_clone_of; | |
1099 | /* Copy the OLD_VERSION_NODE function tree to the new version. */ | |
02774f2d | 1100 | tree_function_versioning (node->clone_of->decl, node->decl, |
c02159a7 | 1101 | node->clone.tree_map, true, |
1102 | node->clone.args_to_skip, false, | |
1103 | NULL, NULL); | |
35ee1c66 | 1104 | if (symtab->dump_file) |
c02159a7 | 1105 | { |
35ee1c66 | 1106 | dump_function_to_file (node->clone_of->decl, symtab->dump_file, |
1107 | dump_flags); | |
1108 | dump_function_to_file (node->decl, symtab->dump_file, dump_flags); | |
c02159a7 | 1109 | } |
1110 | ||
1111 | /* Function is no longer clone. */ | |
1112 | if (node->next_sibling_clone) | |
1113 | node->next_sibling_clone->prev_sibling_clone = node->prev_sibling_clone; | |
1114 | if (node->prev_sibling_clone) | |
1115 | node->prev_sibling_clone->next_sibling_clone = node->next_sibling_clone; | |
1116 | else | |
1117 | node->clone_of->clones = node->next_sibling_clone; | |
1118 | node->next_sibling_clone = NULL; | |
1119 | node->prev_sibling_clone = NULL; | |
02774f2d | 1120 | if (!node->clone_of->analyzed && !node->clone_of->clones) |
c02159a7 | 1121 | { |
415d1b9a | 1122 | node->clone_of->release_body (); |
1123 | node->clone_of->remove_callees (); | |
51ce5652 | 1124 | node->clone_of->remove_all_references (); |
c02159a7 | 1125 | } |
1126 | node->clone_of = NULL; | |
1127 | bitmap_obstack_release (NULL); | |
1128 | } | |
1129 | ||
1130 | /* Once all functions from compilation unit are in memory, produce all clones | |
1131 | and update all calls. We might also do this on demand if we don't want to | |
1132 | bring all functions to memory prior compilation, but current WHOPR | |
47ae02b7 | 1133 | implementation does that and it is a bit easier to keep everything right in |
c02159a7 | 1134 | this order. */ |
1135 | ||
1136 | void | |
35ee1c66 | 1137 | symbol_table::materialize_all_clones (void) |
c02159a7 | 1138 | { |
35ee1c66 | 1139 | cgraph_node *node; |
c02159a7 | 1140 | bool stabilized = false; |
4d044066 | 1141 | |
c02159a7 | 1142 | |
35ee1c66 | 1143 | if (symtab->dump_file) |
1144 | fprintf (symtab->dump_file, "Materializing clones\n"); | |
382ecba7 | 1145 | |
1146 | cgraph_node::checking_verify_cgraph_nodes (); | |
c02159a7 | 1147 | |
1148 | /* We can also do topological order, but number of iterations should be | |
1149 | bounded by number of IPA passes since single IPA pass is probably not | |
1150 | going to create clones of clones it created itself. */ | |
1151 | while (!stabilized) | |
1152 | { | |
1153 | stabilized = true; | |
1154 | FOR_EACH_FUNCTION (node) | |
1155 | { | |
02774f2d | 1156 | if (node->clone_of && node->decl != node->clone_of->decl |
1157 | && !gimple_has_body_p (node->decl)) | |
c02159a7 | 1158 | { |
eaad46f2 | 1159 | if (!node->clone_of->clone_of) |
4560777c | 1160 | node->clone_of->get_untransformed_body (); |
02774f2d | 1161 | if (gimple_has_body_p (node->clone_of->decl)) |
c02159a7 | 1162 | { |
35ee1c66 | 1163 | if (symtab->dump_file) |
c02159a7 | 1164 | { |
35ee1c66 | 1165 | fprintf (symtab->dump_file, "cloning %s to %s\n", |
5ae49d3e | 1166 | xstrdup_for_dump (node->clone_of->name ()), |
1167 | xstrdup_for_dump (node->name ())); | |
c02159a7 | 1168 | if (node->clone.tree_map) |
1169 | { | |
1170 | unsigned int i; | |
35ee1c66 | 1171 | fprintf (symtab->dump_file, " replace map: "); |
f1f41a6c | 1172 | for (i = 0; |
1173 | i < vec_safe_length (node->clone.tree_map); | |
1174 | i++) | |
c02159a7 | 1175 | { |
35ee1c66 | 1176 | ipa_replace_map *replace_info; |
f1f41a6c | 1177 | replace_info = (*node->clone.tree_map)[i]; |
1ffa4346 | 1178 | print_generic_expr (symtab->dump_file, |
1179 | replace_info->old_tree); | |
35ee1c66 | 1180 | fprintf (symtab->dump_file, " -> "); |
1ffa4346 | 1181 | print_generic_expr (symtab->dump_file, |
1182 | replace_info->new_tree); | |
35ee1c66 | 1183 | fprintf (symtab->dump_file, "%s%s;", |
c02159a7 | 1184 | replace_info->replace_p ? "(replace)":"", |
1185 | replace_info->ref_p ? "(ref)":""); | |
1186 | } | |
35ee1c66 | 1187 | fprintf (symtab->dump_file, "\n"); |
c02159a7 | 1188 | } |
1189 | if (node->clone.args_to_skip) | |
1190 | { | |
35ee1c66 | 1191 | fprintf (symtab->dump_file, " args_to_skip: "); |
1192 | dump_bitmap (symtab->dump_file, | |
1193 | node->clone.args_to_skip); | |
c02159a7 | 1194 | } |
1195 | if (node->clone.args_to_skip) | |
1196 | { | |
35ee1c66 | 1197 | fprintf (symtab->dump_file, " combined_args_to_skip:"); |
1198 | dump_bitmap (symtab->dump_file, node->clone.combined_args_to_skip); | |
c02159a7 | 1199 | } |
1200 | } | |
1201 | cgraph_materialize_clone (node); | |
1202 | stabilized = false; | |
1203 | } | |
1204 | } | |
1205 | } | |
1206 | } | |
1207 | FOR_EACH_FUNCTION (node) | |
02774f2d | 1208 | if (!node->analyzed && node->callees) |
9df17e79 | 1209 | { |
415d1b9a | 1210 | node->remove_callees (); |
51ce5652 | 1211 | node->remove_all_references (); |
9df17e79 | 1212 | } |
1213 | else | |
51ce5652 | 1214 | node->clear_stmts_in_references (); |
35ee1c66 | 1215 | if (symtab->dump_file) |
1216 | fprintf (symtab->dump_file, "Materialization Call site updates done.\n"); | |
382ecba7 | 1217 | |
1218 | cgraph_node::checking_verify_cgraph_nodes (); | |
1219 | ||
366970c6 | 1220 | symtab->remove_unreachable_nodes (symtab->dump_file); |
c02159a7 | 1221 | } |
1222 | ||
1223 | #include "gt-cgraphclones.h" |