]>
Commit | Line | Data |
---|---|---|
cd6bca02 | 1 | /* Callgraph based interprocedural optimizations. |
cfaf579d | 2 | Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009 |
e3a37aef | 3 | Free Software Foundation, Inc. |
ae01b312 | 4 | Contributed by Jan Hubicka |
5 | ||
6 | This file is part of GCC. | |
7 | ||
8 | GCC is free software; you can redistribute it and/or modify it under | |
9 | the terms of the GNU General Public License as published by the Free | |
8c4c00c1 | 10 | Software Foundation; either version 3, or (at your option) any later |
ae01b312 | 11 | version. |
12 | ||
13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
16 | for more details. | |
17 | ||
18 | You should have received a copy of the GNU General Public License | |
8c4c00c1 | 19 | along with GCC; see the file COPYING3. If not see |
20 | <http://www.gnu.org/licenses/>. */ | |
ae01b312 | 21 | |
b0cdf642 | 22 | /* This module implements main driver of compilation process as well as |
cd6bca02 | 23 | few basic interprocedural optimizers. |
b0cdf642 | 24 | |
25 | The main scope of this file is to act as an interface in between | |
26 | tree based frontends and the backend (and middle end) | |
27 | ||
28 | The front-end is supposed to use following functionality: | |
29 | ||
30 | - cgraph_finalize_function | |
31 | ||
32 | This function is called once front-end has parsed whole body of function | |
33 | and it is certain that the function body nor the declaration will change. | |
34 | ||
b326746d | 35 | (There is one exception needed for implementing GCC extern inline |
36 | function.) | |
b0cdf642 | 37 | |
1d416bd7 | 38 | - varpool_finalize_variable |
b0cdf642 | 39 | |
7bd28bba | 40 | This function has same behavior as the above but is used for static |
b0cdf642 | 41 | variables. |
42 | ||
43 | - cgraph_finalize_compilation_unit | |
44 | ||
b326746d | 45 | This function is called once (source level) compilation unit is finalized |
46 | and it will no longer change. | |
b0cdf642 | 47 | |
6329636b | 48 | In the the call-graph construction and local function |
b0cdf642 | 49 | analysis takes place here. Bodies of unreachable functions are released |
50 | to conserve memory usage. | |
51 | ||
b326746d | 52 | The function can be called multiple times when multiple source level |
53 | compilation units are combined (such as in C frontend) | |
b0cdf642 | 54 | |
55 | - cgraph_optimize | |
56 | ||
57 | In this unit-at-a-time compilation the intra procedural analysis takes | |
58 | place here. In particular the static functions whose address is never | |
59 | taken are marked as local. Backend can then use this information to | |
60 | modify calling conventions, do better inlining or similar optimizations. | |
61 | ||
b0cdf642 | 62 | - cgraph_mark_needed_node |
1d416bd7 | 63 | - varpool_mark_needed_node |
b0cdf642 | 64 | |
b326746d | 65 | When function or variable is referenced by some hidden way the call-graph |
66 | data structure must be updated accordingly by this function. | |
67 | There should be little need to call this function and all the references | |
68 | should be made explicit to cgraph code. At present these functions are | |
ccd2f3d1 | 69 | used by C++ frontend to explicitly mark the keyed methods. |
b0cdf642 | 70 | |
71 | - analyze_expr callback | |
72 | ||
73 | This function is responsible for lowering tree nodes not understood by | |
74 | generic code into understandable ones or alternatively marking | |
75 | callgraph and varpool nodes referenced by the as needed. | |
76 | ||
77 | ??? On the tree-ssa genericizing should take place here and we will avoid | |
78 | need for these hooks (replacing them by genericizing hook) | |
79 | ||
6329636b | 80 | Analyzing of all functions is deferred |
b0cdf642 | 81 | to cgraph_finalize_compilation_unit and expansion into cgraph_optimize. |
82 | ||
83 | In cgraph_finalize_compilation_unit the reachable functions are | |
84 | analyzed. During analysis the call-graph edges from reachable | |
85 | functions are constructed and their destinations are marked as | |
86 | reachable. References to functions and variables are discovered too | |
87 | and variables found to be needed output to the assembly file. Via | |
88 | mark_referenced call in assemble_variable functions referenced by | |
89 | static variables are noticed too. | |
90 | ||
ca67a72b | 91 | The intra-procedural information is produced and its existence |
b0cdf642 | 92 | indicated by global_info_ready. Once this flag is set it is impossible |
93 | to change function from !reachable to reachable and thus | |
94 | assemble_variable no longer call mark_referenced. | |
95 | ||
96 | Finally the call-graph is topologically sorted and all reachable functions | |
97 | that has not been completely inlined or are not external are output. | |
98 | ||
99 | ??? It is possible that reference to function or variable is optimized | |
100 | out. We can not deal with this nicely because topological order is not | |
101 | suitable for it. For tree-ssa we may consider another pass doing | |
102 | optimization and re-discovering reachable functions. | |
103 | ||
104 | ??? Reorganize code so variables are output very last and only if they | |
105 | really has been referenced by produced code, so we catch more cases | |
6329636b | 106 | where reference has been optimized out. */ |
121f3051 | 107 | |
acc70efa | 108 | |
ae01b312 | 109 | #include "config.h" |
110 | #include "system.h" | |
111 | #include "coretypes.h" | |
112 | #include "tm.h" | |
113 | #include "tree.h" | |
b5530559 | 114 | #include "rtl.h" |
acc70efa | 115 | #include "tree-flow.h" |
ae01b312 | 116 | #include "tree-inline.h" |
117 | #include "langhooks.h" | |
c6224531 | 118 | #include "pointer-set.h" |
ae01b312 | 119 | #include "toplev.h" |
120 | #include "flags.h" | |
121 | #include "ggc.h" | |
122 | #include "debug.h" | |
123 | #include "target.h" | |
124 | #include "cgraph.h" | |
80a85d8a | 125 | #include "diagnostic.h" |
f79b6507 | 126 | #include "timevar.h" |
d7c6d889 | 127 | #include "params.h" |
128 | #include "fibheap.h" | |
611e5405 | 129 | #include "intl.h" |
b69eb0ff | 130 | #include "function.h" |
b5d36404 | 131 | #include "ipa-prop.h" |
75a70cf9 | 132 | #include "gimple.h" |
133 | #include "tree-iterator.h" | |
f1e2a033 | 134 | #include "tree-pass.h" |
c1dcd13c | 135 | #include "output.h" |
9ed5b1f5 | 136 | #include "coverage.h" |
d7c6d889 | 137 | |
a6868229 | 138 | static void cgraph_expand_all_functions (void); |
d9d9733a | 139 | static void cgraph_mark_functions_to_output (void); |
140 | static void cgraph_expand_function (struct cgraph_node *); | |
f788fff2 | 141 | static void cgraph_output_pending_asms (void); |
25bb88de | 142 | |
121f3051 | 143 | static FILE *cgraph_dump_file; |
144 | ||
2c56f72e | 145 | /* A vector of FUNCTION_DECLs declared as static constructors. */ |
146 | static GTY (()) VEC(tree, gc) *static_ctors; | |
147 | /* A vector of FUNCTION_DECLs declared as static destructors. */ | |
148 | static GTY (()) VEC(tree, gc) *static_dtors; | |
a861fe52 | 149 | |
150 | /* When target does not have ctors and dtors, we call all constructor | |
310d2511 | 151 | and destructor by special initialization/destruction function |
a861fe52 | 152 | recognized by collect2. |
153 | ||
154 | When we are going to build this function, collect all constructors and | |
155 | destructors and turn them into normal functions. */ | |
156 | ||
157 | static void | |
158 | record_cdtor_fn (tree fndecl) | |
159 | { | |
2de29097 | 160 | struct cgraph_node *node; |
161 | if (targetm.have_ctors_dtors | |
162 | || (!DECL_STATIC_CONSTRUCTOR (fndecl) | |
163 | && !DECL_STATIC_DESTRUCTOR (fndecl))) | |
a861fe52 | 164 | return; |
165 | ||
166 | if (DECL_STATIC_CONSTRUCTOR (fndecl)) | |
167 | { | |
2c56f72e | 168 | VEC_safe_push (tree, gc, static_ctors, fndecl); |
a861fe52 | 169 | DECL_STATIC_CONSTRUCTOR (fndecl) = 0; |
a861fe52 | 170 | } |
171 | if (DECL_STATIC_DESTRUCTOR (fndecl)) | |
172 | { | |
2c56f72e | 173 | VEC_safe_push (tree, gc, static_dtors, fndecl); |
a861fe52 | 174 | DECL_STATIC_DESTRUCTOR (fndecl) = 0; |
a861fe52 | 175 | } |
2de29097 | 176 | node = cgraph_node (fndecl); |
177 | node->local.disregard_inline_limits = 1; | |
178 | cgraph_mark_reachable_node (node); | |
a861fe52 | 179 | } |
180 | ||
2c56f72e | 181 | /* Define global constructors/destructor functions for the CDTORS, of |
182 | which they are LEN. The CDTORS are sorted by initialization | |
183 | priority. If CTOR_P is true, these are constructors; otherwise, | |
184 | they are destructors. */ | |
185 | ||
a861fe52 | 186 | static void |
2c56f72e | 187 | build_cdtor (bool ctor_p, tree *cdtors, size_t len) |
a861fe52 | 188 | { |
2c56f72e | 189 | size_t i; |
a861fe52 | 190 | |
2c56f72e | 191 | i = 0; |
192 | while (i < len) | |
193 | { | |
194 | tree body; | |
195 | tree fn; | |
196 | priority_type priority; | |
197 | ||
198 | priority = 0; | |
199 | body = NULL_TREE; | |
200 | /* Find the next batch of constructors/destructors with the same | |
201 | initialization priority. */ | |
202 | do | |
203 | { | |
204 | priority_type p; | |
205 | fn = cdtors[i]; | |
206 | p = ctor_p ? DECL_INIT_PRIORITY (fn) : DECL_FINI_PRIORITY (fn); | |
207 | if (!body) | |
208 | priority = p; | |
209 | else if (p != priority) | |
210 | break; | |
211 | append_to_statement_list (build_function_call_expr (fn, 0), | |
212 | &body); | |
213 | ++i; | |
214 | } | |
215 | while (i < len); | |
216 | gcc_assert (body != NULL_TREE); | |
217 | /* Generate a function to call all the function of like | |
218 | priority. */ | |
219 | cgraph_build_static_cdtor (ctor_p ? 'I' : 'D', body, priority); | |
220 | } | |
221 | } | |
222 | ||
223 | /* Comparison function for qsort. P1 and P2 are actually of type | |
224 | "tree *" and point to static constructors. DECL_INIT_PRIORITY is | |
225 | used to determine the sort order. */ | |
a861fe52 | 226 | |
2c56f72e | 227 | static int |
228 | compare_ctor (const void *p1, const void *p2) | |
229 | { | |
230 | tree f1; | |
231 | tree f2; | |
232 | int priority1; | |
233 | int priority2; | |
234 | ||
235 | f1 = *(const tree *)p1; | |
236 | f2 = *(const tree *)p2; | |
237 | priority1 = DECL_INIT_PRIORITY (f1); | |
238 | priority2 = DECL_INIT_PRIORITY (f2); | |
239 | ||
240 | if (priority1 < priority2) | |
241 | return -1; | |
242 | else if (priority1 > priority2) | |
243 | return 1; | |
244 | else | |
245 | /* Ensure a stable sort. */ | |
246 | return (const tree *)p1 - (const tree *)p2; | |
247 | } | |
248 | ||
249 | /* Comparison function for qsort. P1 and P2 are actually of type | |
250 | "tree *" and point to static destructors. DECL_FINI_PRIORITY is | |
251 | used to determine the sort order. */ | |
a861fe52 | 252 | |
2c56f72e | 253 | static int |
254 | compare_dtor (const void *p1, const void *p2) | |
255 | { | |
256 | tree f1; | |
257 | tree f2; | |
258 | int priority1; | |
259 | int priority2; | |
260 | ||
261 | f1 = *(const tree *)p1; | |
262 | f2 = *(const tree *)p2; | |
263 | priority1 = DECL_FINI_PRIORITY (f1); | |
264 | priority2 = DECL_FINI_PRIORITY (f2); | |
265 | ||
266 | if (priority1 < priority2) | |
267 | return -1; | |
268 | else if (priority1 > priority2) | |
269 | return 1; | |
270 | else | |
271 | /* Ensure a stable sort. */ | |
272 | return (const tree *)p1 - (const tree *)p2; | |
a861fe52 | 273 | } |
274 | ||
275 | /* Generate functions to call static constructors and destructors | |
276 | for targets that do not support .ctors/.dtors sections. These | |
277 | functions have magic names which are detected by collect2. */ | |
278 | ||
279 | static void | |
280 | cgraph_build_cdtor_fns (void) | |
281 | { | |
2c56f72e | 282 | if (!VEC_empty (tree, static_ctors)) |
a861fe52 | 283 | { |
2c56f72e | 284 | gcc_assert (!targetm.have_ctors_dtors); |
285 | qsort (VEC_address (tree, static_ctors), | |
286 | VEC_length (tree, static_ctors), | |
287 | sizeof (tree), | |
288 | compare_ctor); | |
289 | build_cdtor (/*ctor_p=*/true, | |
290 | VEC_address (tree, static_ctors), | |
291 | VEC_length (tree, static_ctors)); | |
292 | VEC_truncate (tree, static_ctors, 0); | |
a861fe52 | 293 | } |
2c56f72e | 294 | |
295 | if (!VEC_empty (tree, static_dtors)) | |
a861fe52 | 296 | { |
2c56f72e | 297 | gcc_assert (!targetm.have_ctors_dtors); |
298 | qsort (VEC_address (tree, static_dtors), | |
299 | VEC_length (tree, static_dtors), | |
300 | sizeof (tree), | |
301 | compare_dtor); | |
302 | build_cdtor (/*ctor_p=*/false, | |
303 | VEC_address (tree, static_dtors), | |
304 | VEC_length (tree, static_dtors)); | |
305 | VEC_truncate (tree, static_dtors, 0); | |
a861fe52 | 306 | } |
307 | } | |
308 | ||
2c0b522d | 309 | /* Determine if function DECL is needed. That is, visible to something |
310 | either outside this translation unit, something magic in the system | |
6329636b | 311 | configury. */ |
2c0b522d | 312 | |
313 | static bool | |
314 | decide_is_function_needed (struct cgraph_node *node, tree decl) | |
315 | { | |
62eec3b4 | 316 | if (MAIN_NAME_P (DECL_NAME (decl)) |
317 | && TREE_PUBLIC (decl)) | |
318 | { | |
319 | node->local.externally_visible = true; | |
320 | return true; | |
321 | } | |
4ee9c684 | 322 | |
3f82b628 | 323 | /* If the user told us it is used, then it must be so. */ |
05806473 | 324 | if (node->local.externally_visible) |
325 | return true; | |
326 | ||
3f82b628 | 327 | /* ??? If the assembler name is set by hand, it is possible to assemble |
328 | the name later after finalizing the function and the fact is noticed | |
329 | in assemble_name then. This is arguably a bug. */ | |
330 | if (DECL_ASSEMBLER_NAME_SET_P (decl) | |
331 | && TREE_SYMBOL_REFERENCED (DECL_ASSEMBLER_NAME (decl))) | |
332 | return true; | |
333 | ||
55680bef | 334 | /* With -fkeep-inline-functions we are keeping all inline functions except |
335 | for extern inline ones. */ | |
336 | if (flag_keep_inline_functions | |
337 | && DECL_DECLARED_INLINE_P (decl) | |
316ef6d8 | 338 | && !DECL_EXTERNAL (decl) |
339 | && !lookup_attribute ("always_inline", DECL_ATTRIBUTES (decl))) | |
55680bef | 340 | return true; |
341 | ||
2c0b522d | 342 | /* If we decided it was needed before, but at the time we didn't have |
343 | the body of the function available, then it's still needed. We have | |
344 | to go back and re-check its dependencies now. */ | |
345 | if (node->needed) | |
346 | return true; | |
347 | ||
348 | /* Externally visible functions must be output. The exception is | |
a0c938f0 | 349 | COMDAT functions that must be output only when they are needed. |
8baa9d15 | 350 | |
351 | When not optimizing, also output the static functions. (see | |
95da6220 | 352 | PR24561), but don't do so for always_inline functions, functions |
d3d410e1 | 353 | declared inline and nested functions. These was optimized out |
354 | in the original implementation and it is unclear whether we want | |
554f2707 | 355 | to change the behavior here. */ |
bba7ddf8 | 356 | if (((TREE_PUBLIC (decl) |
d3d410e1 | 357 | || (!optimize && !node->local.disregard_inline_limits |
358 | && !DECL_DECLARED_INLINE_P (decl) | |
359 | && !node->origin)) | |
bba7ddf8 | 360 | && !flag_whole_program) |
62eec3b4 | 361 | && !DECL_COMDAT (decl) && !DECL_EXTERNAL (decl)) |
2c0b522d | 362 | return true; |
363 | ||
364 | /* Constructors and destructors are reachable from the runtime by | |
365 | some mechanism. */ | |
366 | if (DECL_STATIC_CONSTRUCTOR (decl) || DECL_STATIC_DESTRUCTOR (decl)) | |
367 | return true; | |
368 | ||
2c0b522d | 369 | return false; |
370 | } | |
371 | ||
bdc40eb8 | 372 | /* Process CGRAPH_NEW_FUNCTIONS and perform actions necessary to add these |
523c1122 | 373 | functions into callgraph in a way so they look like ordinary reachable |
374 | functions inserted into callgraph already at construction time. */ | |
375 | ||
376 | bool | |
377 | cgraph_process_new_functions (void) | |
378 | { | |
379 | bool output = false; | |
380 | tree fndecl; | |
381 | struct cgraph_node *node; | |
382 | ||
383 | /* Note that this queue may grow as its being processed, as the new | |
384 | functions may generate new ones. */ | |
385 | while (cgraph_new_nodes) | |
386 | { | |
387 | node = cgraph_new_nodes; | |
388 | fndecl = node->decl; | |
389 | cgraph_new_nodes = cgraph_new_nodes->next_needed; | |
390 | switch (cgraph_state) | |
391 | { | |
392 | case CGRAPH_STATE_CONSTRUCTION: | |
393 | /* At construction time we just need to finalize function and move | |
394 | it into reachable functions list. */ | |
395 | ||
396 | node->next_needed = NULL; | |
397 | cgraph_finalize_function (fndecl, false); | |
398 | cgraph_mark_reachable_node (node); | |
399 | output = true; | |
400 | break; | |
401 | ||
402 | case CGRAPH_STATE_IPA: | |
f517b36e | 403 | case CGRAPH_STATE_IPA_SSA: |
523c1122 | 404 | /* When IPA optimization already started, do all essential |
405 | transformations that has been already performed on the whole | |
406 | cgraph but not on this function. */ | |
407 | ||
75a70cf9 | 408 | gimple_register_cfg_hooks (); |
523c1122 | 409 | if (!node->analyzed) |
410 | cgraph_analyze_function (node); | |
411 | push_cfun (DECL_STRUCT_FUNCTION (fndecl)); | |
412 | current_function_decl = fndecl; | |
9c1bff7a | 413 | compute_inline_parameters (node); |
f517b36e | 414 | if ((cgraph_state == CGRAPH_STATE_IPA_SSA |
415 | && !gimple_in_ssa_p (DECL_STRUCT_FUNCTION (fndecl))) | |
416 | /* When not optimizing, be sure we run early local passes anyway | |
417 | to expand OMP. */ | |
418 | || !optimize) | |
20099e35 | 419 | execute_pass_list (pass_early_local_passes.pass.sub); |
523c1122 | 420 | free_dominance_info (CDI_POST_DOMINATORS); |
421 | free_dominance_info (CDI_DOMINATORS); | |
422 | pop_cfun (); | |
423 | current_function_decl = NULL; | |
424 | break; | |
425 | ||
426 | case CGRAPH_STATE_EXPANSION: | |
427 | /* Functions created during expansion shall be compiled | |
428 | directly. */ | |
09fc9532 | 429 | node->process = 0; |
523c1122 | 430 | cgraph_expand_function (node); |
431 | break; | |
432 | ||
433 | default: | |
434 | gcc_unreachable (); | |
435 | break; | |
436 | } | |
50828ed8 | 437 | cgraph_call_function_insertion_hooks (node); |
523c1122 | 438 | } |
439 | return output; | |
440 | } | |
441 | ||
9b8fb23a | 442 | /* As an GCC extension we allow redefinition of the function. The |
443 | semantics when both copies of bodies differ is not well defined. | |
444 | We replace the old body with new body so in unit at a time mode | |
445 | we always use new body, while in normal mode we may end up with | |
446 | old body inlined into some functions and new body expanded and | |
447 | inlined in others. | |
448 | ||
449 | ??? It may make more sense to use one body for inlining and other | |
450 | body for expanding the function but this is difficult to do. */ | |
451 | ||
452 | static void | |
453 | cgraph_reset_node (struct cgraph_node *node) | |
454 | { | |
09fc9532 | 455 | /* If node->process is set, then we have already begun whole-unit analysis. |
6329636b | 456 | This is *not* testing for whether we've already emitted the function. |
457 | That case can be sort-of legitimately seen with real function redefinition | |
458 | errors. I would argue that the front end should never present us with | |
459 | such a case, but don't enforce that for now. */ | |
09fc9532 | 460 | gcc_assert (!node->process); |
9b8fb23a | 461 | |
462 | /* Reset our data structures so we can analyze the function again. */ | |
463 | memset (&node->local, 0, sizeof (node->local)); | |
464 | memset (&node->global, 0, sizeof (node->global)); | |
465 | memset (&node->rtl, 0, sizeof (node->rtl)); | |
466 | node->analyzed = false; | |
467 | node->local.redefined_extern_inline = true; | |
468 | node->local.finalized = false; | |
469 | ||
9b8fb23a | 470 | cgraph_node_remove_callees (node); |
471 | ||
472 | /* We may need to re-queue the node for assembling in case | |
46beef9a | 473 | we already proceeded it and ignored as not needed or got |
474 | a re-declaration in IMA mode. */ | |
475 | if (node->reachable) | |
9b8fb23a | 476 | { |
477 | struct cgraph_node *n; | |
478 | ||
479 | for (n = cgraph_nodes_queue; n; n = n->next_needed) | |
480 | if (n == node) | |
481 | break; | |
482 | if (!n) | |
483 | node->reachable = 0; | |
484 | } | |
485 | } | |
c08871a9 | 486 | |
1e8e9920 | 487 | static void |
488 | cgraph_lower_function (struct cgraph_node *node) | |
489 | { | |
490 | if (node->lowered) | |
491 | return; | |
492 | tree_lowering_passes (node->decl); | |
493 | node->lowered = true; | |
494 | } | |
495 | ||
28df663b | 496 | /* DECL has been parsed. Take it, queue it, compile it at the whim of the |
497 | logic in effect. If NESTED is true, then our caller cannot stand to have | |
498 | the garbage collector run at the moment. We would need to either create | |
499 | a new GC context, or just not compile right now. */ | |
ae01b312 | 500 | |
501 | void | |
28df663b | 502 | cgraph_finalize_function (tree decl, bool nested) |
ae01b312 | 503 | { |
504 | struct cgraph_node *node = cgraph_node (decl); | |
505 | ||
c08871a9 | 506 | if (node->local.finalized) |
9b8fb23a | 507 | cgraph_reset_node (node); |
28df663b | 508 | |
167b550b | 509 | node->pid = cgraph_max_pid ++; |
c08871a9 | 510 | notice_global_symbol (decl); |
79bb87b4 | 511 | node->local.finalized = true; |
e27482aa | 512 | node->lowered = DECL_STRUCT_FUNCTION (decl)->cfg != NULL; |
22c748a1 | 513 | node->finalized_by_frontend = true; |
a861fe52 | 514 | record_cdtor_fn (node->decl); |
9d95b2b0 | 515 | if (node->nested) |
516 | lower_nested_functions (decl); | |
517 | gcc_assert (!node->nested); | |
ae01b312 | 518 | |
2c0b522d | 519 | if (decide_is_function_needed (node, decl)) |
520 | cgraph_mark_needed_node (node); | |
521 | ||
ecda6e51 | 522 | /* Since we reclaim unreachable nodes at the end of every language |
3f82b628 | 523 | level unit, we need to be conservative about possible entry points |
524 | there. */ | |
62eec3b4 | 525 | if ((TREE_PUBLIC (decl) && !DECL_COMDAT (decl) && !DECL_EXTERNAL (decl))) |
3f82b628 | 526 | cgraph_mark_reachable_node (node); |
527 | ||
2c0b522d | 528 | /* If we've not yet emitted decl, tell the debug info about it. */ |
28df663b | 529 | if (!TREE_ASM_WRITTEN (decl)) |
2c0b522d | 530 | (*debug_hooks->deferred_inline_function) (decl); |
4e8871a0 | 531 | |
b69eb0ff | 532 | /* Possibly warn about unused parameters. */ |
533 | if (warn_unused_parameter) | |
534 | do_warn_unused_parameter (decl); | |
6329636b | 535 | |
536 | if (!nested) | |
537 | ggc_collect (); | |
ae01b312 | 538 | } |
539 | ||
0da03d11 | 540 | /* C99 extern inline keywords allow changing of declaration after function |
541 | has been finalized. We need to re-decide if we want to mark the function as | |
542 | needed then. */ | |
543 | ||
544 | void | |
545 | cgraph_mark_if_needed (tree decl) | |
546 | { | |
547 | struct cgraph_node *node = cgraph_node (decl); | |
548 | if (node->local.finalized && decide_is_function_needed (node, decl)) | |
549 | cgraph_mark_needed_node (node); | |
550 | } | |
551 | ||
ccf4ab6b | 552 | /* Return TRUE if NODE2 is equivalent to NODE or its clone. */ |
553 | static bool | |
554 | clone_of_p (struct cgraph_node *node, struct cgraph_node *node2) | |
555 | { | |
556 | while (node != node2 && node2) | |
557 | node2 = node2->clone_of; | |
558 | return node2 != NULL; | |
559 | } | |
560 | ||
b0cdf642 | 561 | /* Verify cgraph nodes of given cgraph node. */ |
562 | void | |
563 | verify_cgraph_node (struct cgraph_node *node) | |
564 | { | |
565 | struct cgraph_edge *e; | |
e27482aa | 566 | struct function *this_cfun = DECL_STRUCT_FUNCTION (node->decl); |
117ef3d7 | 567 | struct function *saved_cfun = cfun; |
e27482aa | 568 | basic_block this_block; |
75a70cf9 | 569 | gimple_stmt_iterator gsi; |
9bfec7c2 | 570 | bool error_found = false; |
b0cdf642 | 571 | |
bd09cd3e | 572 | if (errorcount || sorrycount) |
573 | return; | |
574 | ||
b0cdf642 | 575 | timevar_push (TV_CGRAPH_VERIFY); |
117ef3d7 | 576 | /* debug_generic_stmt needs correct cfun */ |
577 | set_cfun (this_cfun); | |
b0cdf642 | 578 | for (e = node->callees; e; e = e->next_callee) |
579 | if (e->aux) | |
580 | { | |
0a81f5a0 | 581 | error ("aux field set for edge %s->%s", |
abd3e6b5 | 582 | identifier_to_locale (cgraph_node_name (e->caller)), |
583 | identifier_to_locale (cgraph_node_name (e->callee))); | |
b0cdf642 | 584 | error_found = true; |
585 | } | |
a2cb9b3b | 586 | if (node->count < 0) |
587 | { | |
588 | error ("Execution count is negative"); | |
589 | error_found = true; | |
590 | } | |
b0cdf642 | 591 | for (e = node->callers; e; e = e->next_caller) |
592 | { | |
a2cb9b3b | 593 | if (e->count < 0) |
594 | { | |
595 | error ("caller edge count is negative"); | |
596 | error_found = true; | |
597 | } | |
4ae20857 | 598 | if (e->frequency < 0) |
599 | { | |
600 | error ("caller edge frequency is negative"); | |
601 | error_found = true; | |
602 | } | |
603 | if (e->frequency > CGRAPH_FREQ_MAX) | |
604 | { | |
605 | error ("caller edge frequency is too large"); | |
606 | error_found = true; | |
607 | } | |
b0cdf642 | 608 | if (!e->inline_failed) |
609 | { | |
610 | if (node->global.inlined_to | |
611 | != (e->caller->global.inlined_to | |
612 | ? e->caller->global.inlined_to : e->caller)) | |
613 | { | |
0a81f5a0 | 614 | error ("inlined_to pointer is wrong"); |
b0cdf642 | 615 | error_found = true; |
616 | } | |
617 | if (node->callers->next_caller) | |
618 | { | |
0a81f5a0 | 619 | error ("multiple inline callers"); |
b0cdf642 | 620 | error_found = true; |
621 | } | |
622 | } | |
623 | else | |
624 | if (node->global.inlined_to) | |
625 | { | |
0a81f5a0 | 626 | error ("inlined_to pointer set for noninline callers"); |
b0cdf642 | 627 | error_found = true; |
628 | } | |
629 | } | |
630 | if (!node->callers && node->global.inlined_to) | |
631 | { | |
5cd75817 | 632 | error ("inlined_to pointer is set but no predecessors found"); |
b0cdf642 | 633 | error_found = true; |
634 | } | |
635 | if (node->global.inlined_to == node) | |
636 | { | |
0a81f5a0 | 637 | error ("inlined_to pointer refers to itself"); |
b0cdf642 | 638 | error_found = true; |
639 | } | |
640 | ||
0f6439b9 | 641 | if (!cgraph_node (node->decl)) |
b0cdf642 | 642 | { |
0f6439b9 | 643 | error ("node not found in cgraph_hash"); |
b0cdf642 | 644 | error_found = true; |
645 | } | |
a0c938f0 | 646 | |
ccf4ab6b | 647 | if (node->clone_of) |
648 | { | |
649 | struct cgraph_node *n; | |
650 | for (n = node->clone_of->clones; n; n = n->next_sibling_clone) | |
651 | if (n == node) | |
652 | break; | |
653 | if (!n) | |
654 | { | |
655 | error ("node has wrong clone_of"); | |
656 | error_found = true; | |
657 | } | |
658 | } | |
659 | if (node->clones) | |
660 | { | |
661 | struct cgraph_node *n; | |
662 | for (n = node->clones; n; n = n->next_sibling_clone) | |
663 | if (n->clone_of != node) | |
664 | break; | |
665 | if (n) | |
666 | { | |
667 | error ("node has wrong clone list"); | |
668 | error_found = true; | |
669 | } | |
670 | } | |
671 | if ((node->prev_sibling_clone || node->next_sibling_clone) && !node->clone_of) | |
672 | { | |
673 | error ("node is in clone list but it is not clone"); | |
674 | error_found = true; | |
675 | } | |
676 | if (!node->prev_sibling_clone && node->clone_of && node->clone_of->clones != node) | |
677 | { | |
678 | error ("node has wrong prev_clone pointer"); | |
679 | error_found = true; | |
680 | } | |
681 | if (node->prev_sibling_clone && node->prev_sibling_clone->next_sibling_clone != node) | |
682 | { | |
683 | error ("double linked list of clones corrupted"); | |
684 | error_found = true; | |
685 | } | |
686 | ||
687 | if (node->analyzed && gimple_has_body_p (node->decl) | |
75a70cf9 | 688 | && !TREE_ASM_WRITTEN (node->decl) |
b0cdf642 | 689 | && (!DECL_EXTERNAL (node->decl) || node->global.inlined_to)) |
690 | { | |
e27482aa | 691 | if (this_cfun->cfg) |
692 | { | |
693 | /* The nodes we're interested in are never shared, so walk | |
694 | the tree ignoring duplicates. */ | |
e7c352d1 | 695 | struct pointer_set_t *visited_nodes = pointer_set_create (); |
e27482aa | 696 | /* Reach the trees by walking over the CFG, and note the |
697 | enclosing basic-blocks in the call edges. */ | |
698 | FOR_EACH_BB_FN (this_block, this_cfun) | |
75a70cf9 | 699 | for (gsi = gsi_start_bb (this_block); |
700 | !gsi_end_p (gsi); | |
701 | gsi_next (&gsi)) | |
9bfec7c2 | 702 | { |
75a70cf9 | 703 | gimple stmt = gsi_stmt (gsi); |
9bfec7c2 | 704 | tree decl; |
75a70cf9 | 705 | if (is_gimple_call (stmt) && (decl = gimple_call_fndecl (stmt))) |
9bfec7c2 | 706 | { |
707 | struct cgraph_edge *e = cgraph_edge (node, stmt); | |
708 | if (e) | |
709 | { | |
710 | if (e->aux) | |
711 | { | |
0a81f5a0 | 712 | error ("shared call_stmt:"); |
75a70cf9 | 713 | debug_gimple_stmt (stmt); |
9bfec7c2 | 714 | error_found = true; |
715 | } | |
ccf4ab6b | 716 | if (!clone_of_p (cgraph_node (decl), e->callee) |
717 | && !e->callee->global.inlined_to) | |
9bfec7c2 | 718 | { |
0a81f5a0 | 719 | error ("edge points to wrong declaration:"); |
9bfec7c2 | 720 | debug_tree (e->callee->decl); |
721 | fprintf (stderr," Instead of:"); | |
722 | debug_tree (decl); | |
723 | } | |
724 | e->aux = (void *)1; | |
725 | } | |
726 | else | |
727 | { | |
0a81f5a0 | 728 | error ("missing callgraph edge for call stmt:"); |
75a70cf9 | 729 | debug_gimple_stmt (stmt); |
9bfec7c2 | 730 | error_found = true; |
731 | } | |
732 | } | |
733 | } | |
e27482aa | 734 | pointer_set_destroy (visited_nodes); |
e27482aa | 735 | } |
736 | else | |
737 | /* No CFG available?! */ | |
738 | gcc_unreachable (); | |
739 | ||
b0cdf642 | 740 | for (e = node->callees; e; e = e->next_callee) |
741 | { | |
f8daee9b | 742 | if (!e->aux && !e->indirect_call) |
b0cdf642 | 743 | { |
0a81f5a0 | 744 | error ("edge %s->%s has no corresponding call_stmt", |
abd3e6b5 | 745 | identifier_to_locale (cgraph_node_name (e->caller)), |
746 | identifier_to_locale (cgraph_node_name (e->callee))); | |
75a70cf9 | 747 | debug_gimple_stmt (e->call_stmt); |
b0cdf642 | 748 | error_found = true; |
749 | } | |
750 | e->aux = 0; | |
751 | } | |
752 | } | |
753 | if (error_found) | |
754 | { | |
755 | dump_cgraph_node (stderr, node); | |
0a81f5a0 | 756 | internal_error ("verify_cgraph_node failed"); |
b0cdf642 | 757 | } |
117ef3d7 | 758 | set_cfun (saved_cfun); |
b0cdf642 | 759 | timevar_pop (TV_CGRAPH_VERIFY); |
760 | } | |
761 | ||
762 | /* Verify whole cgraph structure. */ | |
763 | void | |
764 | verify_cgraph (void) | |
765 | { | |
766 | struct cgraph_node *node; | |
767 | ||
8ec2a798 | 768 | if (sorrycount || errorcount) |
769 | return; | |
770 | ||
b0cdf642 | 771 | for (node = cgraph_nodes; node; node = node->next) |
772 | verify_cgraph_node (node); | |
773 | } | |
774 | ||
56af936e | 775 | /* Output all asm statements we have stored up to be output. */ |
776 | ||
777 | static void | |
778 | cgraph_output_pending_asms (void) | |
779 | { | |
780 | struct cgraph_asm_node *can; | |
781 | ||
782 | if (errorcount || sorrycount) | |
783 | return; | |
784 | ||
785 | for (can = cgraph_asm_nodes; can; can = can->next) | |
786 | assemble_asm (can->asm_str); | |
787 | cgraph_asm_nodes = NULL; | |
788 | } | |
789 | ||
0785e435 | 790 | /* Analyze the function scheduled to be output. */ |
1e8e9920 | 791 | void |
0785e435 | 792 | cgraph_analyze_function (struct cgraph_node *node) |
793 | { | |
794 | tree decl = node->decl; | |
795 | ||
ec1e35b2 | 796 | current_function_decl = decl; |
e27482aa | 797 | push_cfun (DECL_STRUCT_FUNCTION (decl)); |
798 | cgraph_lower_function (node); | |
6e8d6e86 | 799 | node->analyzed = true; |
0785e435 | 800 | |
e27482aa | 801 | pop_cfun (); |
c08871a9 | 802 | current_function_decl = NULL; |
0785e435 | 803 | } |
804 | ||
05806473 | 805 | /* Look for externally_visible and used attributes and mark cgraph nodes |
806 | accordingly. | |
807 | ||
808 | We cannot mark the nodes at the point the attributes are processed (in | |
809 | handle_*_attribute) because the copy of the declarations available at that | |
810 | point may not be canonical. For example, in: | |
811 | ||
812 | void f(); | |
813 | void f() __attribute__((used)); | |
814 | ||
815 | the declaration we see in handle_used_attribute will be the second | |
816 | declaration -- but the front end will subsequently merge that declaration | |
817 | with the original declaration and discard the second declaration. | |
818 | ||
819 | Furthermore, we can't mark these nodes in cgraph_finalize_function because: | |
820 | ||
821 | void f() {} | |
822 | void f() __attribute__((externally_visible)); | |
823 | ||
824 | is valid. | |
825 | ||
826 | So, we walk the nodes at the end of the translation unit, applying the | |
827 | attributes at that point. */ | |
828 | ||
829 | static void | |
830 | process_function_and_variable_attributes (struct cgraph_node *first, | |
1d416bd7 | 831 | struct varpool_node *first_var) |
05806473 | 832 | { |
833 | struct cgraph_node *node; | |
1d416bd7 | 834 | struct varpool_node *vnode; |
05806473 | 835 | |
836 | for (node = cgraph_nodes; node != first; node = node->next) | |
837 | { | |
838 | tree decl = node->decl; | |
839 | if (lookup_attribute ("used", DECL_ATTRIBUTES (decl))) | |
840 | { | |
841 | mark_decl_referenced (decl); | |
842 | if (node->local.finalized) | |
843 | cgraph_mark_needed_node (node); | |
844 | } | |
845 | if (lookup_attribute ("externally_visible", DECL_ATTRIBUTES (decl))) | |
846 | { | |
ba12ea31 | 847 | if (! TREE_PUBLIC (node->decl)) |
848 | warning (OPT_Wattributes, | |
849 | "%J%<externally_visible%> attribute have effect only on public objects", | |
850 | node->decl); | |
851 | else | |
852 | { | |
853 | if (node->local.finalized) | |
854 | cgraph_mark_needed_node (node); | |
855 | node->local.externally_visible = true; | |
856 | } | |
05806473 | 857 | } |
858 | } | |
1d416bd7 | 859 | for (vnode = varpool_nodes; vnode != first_var; vnode = vnode->next) |
05806473 | 860 | { |
861 | tree decl = vnode->decl; | |
862 | if (lookup_attribute ("used", DECL_ATTRIBUTES (decl))) | |
863 | { | |
864 | mark_decl_referenced (decl); | |
865 | if (vnode->finalized) | |
1d416bd7 | 866 | varpool_mark_needed_node (vnode); |
05806473 | 867 | } |
868 | if (lookup_attribute ("externally_visible", DECL_ATTRIBUTES (decl))) | |
869 | { | |
ba12ea31 | 870 | if (! TREE_PUBLIC (vnode->decl)) |
871 | warning (OPT_Wattributes, | |
872 | "%J%<externally_visible%> attribute have effect only on public objects", | |
873 | vnode->decl); | |
874 | else | |
875 | { | |
876 | if (vnode->finalized) | |
1d416bd7 | 877 | varpool_mark_needed_node (vnode); |
ba12ea31 | 878 | vnode->externally_visible = true; |
879 | } | |
05806473 | 880 | } |
881 | } | |
882 | } | |
883 | ||
aeeb194b | 884 | /* Process CGRAPH_NODES_NEEDED queue, analyze each function (and transitively |
885 | each reachable functions) and build cgraph. | |
886 | The function can be called multiple times after inserting new nodes | |
0d424440 | 887 | into beginning of queue. Just the new part of queue is re-scanned then. */ |
ae01b312 | 888 | |
aeeb194b | 889 | static void |
890 | cgraph_analyze_functions (void) | |
ae01b312 | 891 | { |
c1dcd13c | 892 | /* Keep track of already processed nodes when called multiple times for |
06b27565 | 893 | intermodule optimization. */ |
c1dcd13c | 894 | static struct cgraph_node *first_analyzed; |
c17d0de1 | 895 | struct cgraph_node *first_processed = first_analyzed; |
1d416bd7 | 896 | static struct varpool_node *first_analyzed_var; |
aeeb194b | 897 | struct cgraph_node *node, *next; |
ae01b312 | 898 | |
c17d0de1 | 899 | process_function_and_variable_attributes (first_processed, |
900 | first_analyzed_var); | |
901 | first_processed = cgraph_nodes; | |
1d416bd7 | 902 | first_analyzed_var = varpool_nodes; |
903 | varpool_analyze_pending_decls (); | |
f79b6507 | 904 | if (cgraph_dump_file) |
ae01b312 | 905 | { |
e4200070 | 906 | fprintf (cgraph_dump_file, "Initial entry points:"); |
c1dcd13c | 907 | for (node = cgraph_nodes; node != first_analyzed; node = node->next) |
1a1a827a | 908 | if (node->needed) |
f79b6507 | 909 | fprintf (cgraph_dump_file, " %s", cgraph_node_name (node)); |
910 | fprintf (cgraph_dump_file, "\n"); | |
ae01b312 | 911 | } |
aeeb194b | 912 | cgraph_process_new_functions (); |
ae01b312 | 913 | |
e6d2b2d8 | 914 | /* Propagate reachability flag and lower representation of all reachable |
915 | functions. In the future, lowering will introduce new functions and | |
916 | new entry points on the way (by template instantiation and virtual | |
917 | method table generation for instance). */ | |
3d7bfc56 | 918 | while (cgraph_nodes_queue) |
ae01b312 | 919 | { |
0785e435 | 920 | struct cgraph_edge *edge; |
3d7bfc56 | 921 | tree decl = cgraph_nodes_queue->decl; |
922 | ||
923 | node = cgraph_nodes_queue; | |
d87976fb | 924 | cgraph_nodes_queue = cgraph_nodes_queue->next_needed; |
b0cdf642 | 925 | node->next_needed = NULL; |
ae01b312 | 926 | |
638531ad | 927 | /* ??? It is possible to create extern inline function and later using |
bbd5cba2 | 928 | weak alias attribute to kill its body. See |
638531ad | 929 | gcc.c-torture/compile/20011119-1.c */ |
75a70cf9 | 930 | if (!DECL_STRUCT_FUNCTION (decl)) |
9b8fb23a | 931 | { |
932 | cgraph_reset_node (node); | |
933 | continue; | |
934 | } | |
638531ad | 935 | |
cc636d56 | 936 | gcc_assert (!node->analyzed && node->reachable); |
75a70cf9 | 937 | gcc_assert (gimple_body (decl)); |
ae01b312 | 938 | |
0785e435 | 939 | cgraph_analyze_function (node); |
2c0b522d | 940 | |
ae01b312 | 941 | for (edge = node->callees; edge; edge = edge->next_callee) |
0785e435 | 942 | if (!edge->callee->reachable) |
2c0b522d | 943 | cgraph_mark_reachable_node (edge->callee); |
944 | ||
d544ceff | 945 | /* If decl is a clone of an abstract function, mark that abstract |
946 | function so that we don't release its body. The DECL_INITIAL() of that | |
947 | abstract function declaration will be later needed to output debug info. */ | |
948 | if (DECL_ABSTRACT_ORIGIN (decl)) | |
949 | { | |
950 | struct cgraph_node *origin_node = cgraph_node (DECL_ABSTRACT_ORIGIN (decl)); | |
951 | origin_node->abstract_and_needed = true; | |
952 | } | |
953 | ||
c17d0de1 | 954 | /* We finalize local static variables during constructing callgraph |
955 | edges. Process their attributes too. */ | |
956 | process_function_and_variable_attributes (first_processed, | |
957 | first_analyzed_var); | |
958 | first_processed = cgraph_nodes; | |
1d416bd7 | 959 | first_analyzed_var = varpool_nodes; |
960 | varpool_analyze_pending_decls (); | |
aeeb194b | 961 | cgraph_process_new_functions (); |
ae01b312 | 962 | } |
2c0b522d | 963 | |
aa5e06c7 | 964 | /* Collect entry points to the unit. */ |
f79b6507 | 965 | if (cgraph_dump_file) |
3d7bfc56 | 966 | { |
e4200070 | 967 | fprintf (cgraph_dump_file, "Unit entry points:"); |
c1dcd13c | 968 | for (node = cgraph_nodes; node != first_analyzed; node = node->next) |
1a1a827a | 969 | if (node->needed) |
f79b6507 | 970 | fprintf (cgraph_dump_file, " %s", cgraph_node_name (node)); |
e4200070 | 971 | fprintf (cgraph_dump_file, "\n\nInitial "); |
0785e435 | 972 | dump_cgraph (cgraph_dump_file); |
3d7bfc56 | 973 | } |
e6d2b2d8 | 974 | |
f79b6507 | 975 | if (cgraph_dump_file) |
976 | fprintf (cgraph_dump_file, "\nReclaiming functions:"); | |
ae01b312 | 977 | |
f4ec5ce1 | 978 | for (node = cgraph_nodes; node != first_analyzed; node = next) |
ae01b312 | 979 | { |
980 | tree decl = node->decl; | |
f4ec5ce1 | 981 | next = node->next; |
ae01b312 | 982 | |
1a1a827a | 983 | if (node->local.finalized && !gimple_has_body_p (decl)) |
a0c938f0 | 984 | cgraph_reset_node (node); |
9b8fb23a | 985 | |
1a1a827a | 986 | if (!node->reachable && gimple_has_body_p (decl)) |
ae01b312 | 987 | { |
f79b6507 | 988 | if (cgraph_dump_file) |
989 | fprintf (cgraph_dump_file, " %s", cgraph_node_name (node)); | |
b0cdf642 | 990 | cgraph_remove_node (node); |
9b8fb23a | 991 | continue; |
ae01b312 | 992 | } |
bc5cab3b | 993 | else |
994 | node->next_needed = NULL; | |
1a1a827a | 995 | gcc_assert (!node->local.finalized || gimple_has_body_p (decl)); |
9b8fb23a | 996 | gcc_assert (node->analyzed == node->local.finalized); |
ae01b312 | 997 | } |
f79b6507 | 998 | if (cgraph_dump_file) |
e4200070 | 999 | { |
1000 | fprintf (cgraph_dump_file, "\n\nReclaimed "); | |
1001 | dump_cgraph (cgraph_dump_file); | |
1002 | } | |
c1dcd13c | 1003 | first_analyzed = cgraph_nodes; |
ae01b312 | 1004 | ggc_collect (); |
aeeb194b | 1005 | } |
1006 | ||
1007 | /* Analyze the whole compilation unit once it is parsed completely. */ | |
1008 | ||
1009 | void | |
1010 | cgraph_finalize_compilation_unit (void) | |
1011 | { | |
1012 | if (errorcount || sorrycount) | |
1013 | return; | |
1014 | ||
1015 | finish_aliases_1 (); | |
1016 | ||
aeeb194b | 1017 | if (!quiet_flag) |
1018 | { | |
1019 | fprintf (stderr, "\nAnalyzing compilation unit\n"); | |
1020 | fflush (stderr); | |
1021 | } | |
1022 | ||
1023 | timevar_push (TV_CGRAPH); | |
1024 | cgraph_analyze_functions (); | |
f79b6507 | 1025 | timevar_pop (TV_CGRAPH); |
ae01b312 | 1026 | } |
9ed5b1f5 | 1027 | |
1028 | ||
ae01b312 | 1029 | /* Figure out what functions we want to assemble. */ |
1030 | ||
1031 | static void | |
d9d9733a | 1032 | cgraph_mark_functions_to_output (void) |
ae01b312 | 1033 | { |
1034 | struct cgraph_node *node; | |
1035 | ||
ae01b312 | 1036 | for (node = cgraph_nodes; node; node = node->next) |
1037 | { | |
1038 | tree decl = node->decl; | |
d7c6d889 | 1039 | struct cgraph_edge *e; |
a0c938f0 | 1040 | |
09fc9532 | 1041 | gcc_assert (!node->process); |
d7c6d889 | 1042 | |
1043 | for (e = node->callers; e; e = e->next_caller) | |
611e5405 | 1044 | if (e->inline_failed) |
d7c6d889 | 1045 | break; |
ae01b312 | 1046 | |
e6d2b2d8 | 1047 | /* We need to output all local functions that are used and not |
1048 | always inlined, as well as those that are reachable from | |
1049 | outside the current compilation unit. */ | |
1a1a827a | 1050 | if (node->analyzed |
b0cdf642 | 1051 | && !node->global.inlined_to |
ae01b312 | 1052 | && (node->needed |
d7c6d889 | 1053 | || (e && node->reachable)) |
4ee9c684 | 1054 | && !TREE_ASM_WRITTEN (decl) |
ae01b312 | 1055 | && !DECL_EXTERNAL (decl)) |
09fc9532 | 1056 | node->process = 1; |
cc636d56 | 1057 | else |
9cee7c3f | 1058 | { |
1059 | /* We should've reclaimed all functions that are not needed. */ | |
1060 | #ifdef ENABLE_CHECKING | |
75a70cf9 | 1061 | if (!node->global.inlined_to |
1a1a827a | 1062 | && gimple_has_body_p (decl) |
9cee7c3f | 1063 | && !DECL_EXTERNAL (decl)) |
1064 | { | |
1065 | dump_cgraph_node (stderr, node); | |
1066 | internal_error ("failed to reclaim unneeded function"); | |
1067 | } | |
1068 | #endif | |
75a70cf9 | 1069 | gcc_assert (node->global.inlined_to |
1a1a827a | 1070 | || !gimple_has_body_p (decl) |
9cee7c3f | 1071 | || DECL_EXTERNAL (decl)); |
1072 | ||
1073 | } | |
a0c938f0 | 1074 | |
961e3b13 | 1075 | } |
1076 | } | |
1077 | ||
ae01b312 | 1078 | /* Expand function specified by NODE. */ |
e6d2b2d8 | 1079 | |
ae01b312 | 1080 | static void |
d9d9733a | 1081 | cgraph_expand_function (struct cgraph_node *node) |
ae01b312 | 1082 | { |
1083 | tree decl = node->decl; | |
1084 | ||
b0cdf642 | 1085 | /* We ought to not compile any inline clones. */ |
cc636d56 | 1086 | gcc_assert (!node->global.inlined_to); |
b0cdf642 | 1087 | |
6329636b | 1088 | announce_function (decl); |
09fc9532 | 1089 | node->process = 0; |
961e3b13 | 1090 | |
e7c352d1 | 1091 | gcc_assert (node->lowered); |
f8deefc1 | 1092 | |
794da2bb | 1093 | /* Generate RTL for the body of DECL. */ |
22c748a1 | 1094 | if (lang_hooks.callgraph.emit_associated_thunks |
1095 | && node->finalized_by_frontend) | |
84e10000 | 1096 | lang_hooks.callgraph.emit_associated_thunks (decl); |
1097 | tree_rest_of_compilation (decl); | |
961e3b13 | 1098 | |
4ee9c684 | 1099 | /* Make sure that BE didn't give up on compiling. */ |
c04e3894 | 1100 | gcc_assert (TREE_ASM_WRITTEN (decl)); |
ae01b312 | 1101 | current_function_decl = NULL; |
1a1a827a | 1102 | gcc_assert (!cgraph_preserve_function_body_p (decl)); |
1103 | cgraph_release_function_body (node); | |
1104 | /* Eliminate all call edges. This is important so the GIMPLE_CALL no longer | |
1105 | points to the dead function body. */ | |
1106 | cgraph_node_remove_callees (node); | |
e1be32b8 | 1107 | |
1108 | cgraph_function_flags_ready = true; | |
ae01b312 | 1109 | } |
1110 | ||
b0cdf642 | 1111 | /* Return true when CALLER_DECL should be inlined into CALLEE_DECL. */ |
d7c6d889 | 1112 | |
1113 | bool | |
326a9581 | 1114 | cgraph_inline_p (struct cgraph_edge *e, cgraph_inline_failed_t *reason) |
d7c6d889 | 1115 | { |
b0cdf642 | 1116 | *reason = e->inline_failed; |
1117 | return !e->inline_failed; | |
d7c6d889 | 1118 | } |
b0cdf642 | 1119 | |
acc70efa | 1120 | |
acc70efa | 1121 | |
d9d9733a | 1122 | /* Expand all functions that must be output. |
1123 | ||
d7c6d889 | 1124 | Attempt to topologically sort the nodes so function is output when |
1125 | all called functions are already assembled to allow data to be | |
91c82c20 | 1126 | propagated across the callgraph. Use a stack to get smaller distance |
3927afe0 | 1127 | between a function and its callees (later we may choose to use a more |
d7c6d889 | 1128 | sophisticated algorithm for function reordering; we will likely want |
1129 | to use subsections to make the output functions appear in top-down | |
1130 | order). */ | |
1131 | ||
1132 | static void | |
a6868229 | 1133 | cgraph_expand_all_functions (void) |
d7c6d889 | 1134 | { |
1135 | struct cgraph_node *node; | |
4c36ffe6 | 1136 | struct cgraph_node **order = XCNEWVEC (struct cgraph_node *, cgraph_n_nodes); |
c04e3894 | 1137 | int order_pos, new_order_pos = 0; |
d7c6d889 | 1138 | int i; |
1139 | ||
d7c6d889 | 1140 | order_pos = cgraph_postorder (order); |
cc636d56 | 1141 | gcc_assert (order_pos == cgraph_n_nodes); |
d7c6d889 | 1142 | |
7bd28bba | 1143 | /* Garbage collector may remove inline clones we eliminate during |
b0cdf642 | 1144 | optimization. So we must be sure to not reference them. */ |
1145 | for (i = 0; i < order_pos; i++) | |
09fc9532 | 1146 | if (order[i]->process) |
b0cdf642 | 1147 | order[new_order_pos++] = order[i]; |
1148 | ||
1149 | for (i = new_order_pos - 1; i >= 0; i--) | |
d7c6d889 | 1150 | { |
1151 | node = order[i]; | |
09fc9532 | 1152 | if (node->process) |
d7c6d889 | 1153 | { |
cc636d56 | 1154 | gcc_assert (node->reachable); |
09fc9532 | 1155 | node->process = 0; |
d7c6d889 | 1156 | cgraph_expand_function (node); |
1157 | } | |
1158 | } | |
523c1122 | 1159 | cgraph_process_new_functions (); |
773c5ba7 | 1160 | |
d7c6d889 | 1161 | free (order); |
773c5ba7 | 1162 | |
d7c6d889 | 1163 | } |
1164 | ||
56af936e | 1165 | /* This is used to sort the node types by the cgraph order number. */ |
1166 | ||
0b09525f | 1167 | enum cgraph_order_sort_kind |
1168 | { | |
1169 | ORDER_UNDEFINED = 0, | |
1170 | ORDER_FUNCTION, | |
1171 | ORDER_VAR, | |
1172 | ORDER_ASM | |
1173 | }; | |
1174 | ||
56af936e | 1175 | struct cgraph_order_sort |
1176 | { | |
0b09525f | 1177 | enum cgraph_order_sort_kind kind; |
56af936e | 1178 | union |
1179 | { | |
1180 | struct cgraph_node *f; | |
1d416bd7 | 1181 | struct varpool_node *v; |
56af936e | 1182 | struct cgraph_asm_node *a; |
1183 | } u; | |
1184 | }; | |
1185 | ||
1186 | /* Output all functions, variables, and asm statements in the order | |
1187 | according to their order fields, which is the order in which they | |
1188 | appeared in the file. This implements -fno-toplevel-reorder. In | |
1189 | this mode we may output functions and variables which don't really | |
1190 | need to be output. */ | |
1191 | ||
1192 | static void | |
1193 | cgraph_output_in_order (void) | |
1194 | { | |
1195 | int max; | |
1196 | size_t size; | |
1197 | struct cgraph_order_sort *nodes; | |
1198 | int i; | |
1199 | struct cgraph_node *pf; | |
1d416bd7 | 1200 | struct varpool_node *pv; |
56af936e | 1201 | struct cgraph_asm_node *pa; |
1202 | ||
1203 | max = cgraph_order; | |
1204 | size = max * sizeof (struct cgraph_order_sort); | |
1205 | nodes = (struct cgraph_order_sort *) alloca (size); | |
1206 | memset (nodes, 0, size); | |
1207 | ||
1d416bd7 | 1208 | varpool_analyze_pending_decls (); |
56af936e | 1209 | |
1210 | for (pf = cgraph_nodes; pf; pf = pf->next) | |
1211 | { | |
09fc9532 | 1212 | if (pf->process) |
56af936e | 1213 | { |
1214 | i = pf->order; | |
1215 | gcc_assert (nodes[i].kind == ORDER_UNDEFINED); | |
1216 | nodes[i].kind = ORDER_FUNCTION; | |
1217 | nodes[i].u.f = pf; | |
1218 | } | |
1219 | } | |
1220 | ||
1d416bd7 | 1221 | for (pv = varpool_nodes_queue; pv; pv = pv->next_needed) |
56af936e | 1222 | { |
1223 | i = pv->order; | |
1224 | gcc_assert (nodes[i].kind == ORDER_UNDEFINED); | |
1225 | nodes[i].kind = ORDER_VAR; | |
1226 | nodes[i].u.v = pv; | |
1227 | } | |
1228 | ||
1229 | for (pa = cgraph_asm_nodes; pa; pa = pa->next) | |
1230 | { | |
1231 | i = pa->order; | |
1232 | gcc_assert (nodes[i].kind == ORDER_UNDEFINED); | |
1233 | nodes[i].kind = ORDER_ASM; | |
1234 | nodes[i].u.a = pa; | |
1235 | } | |
56af936e | 1236 | |
304e5318 | 1237 | /* In toplevel reorder mode we output all statics; mark them as needed. */ |
1238 | for (i = 0; i < max; ++i) | |
1239 | { | |
1240 | if (nodes[i].kind == ORDER_VAR) | |
1241 | { | |
1242 | varpool_mark_needed_node (nodes[i].u.v); | |
1243 | } | |
1244 | } | |
1245 | varpool_empty_needed_queue (); | |
1246 | ||
56af936e | 1247 | for (i = 0; i < max; ++i) |
1248 | { | |
1249 | switch (nodes[i].kind) | |
1250 | { | |
1251 | case ORDER_FUNCTION: | |
09fc9532 | 1252 | nodes[i].u.f->process = 0; |
56af936e | 1253 | cgraph_expand_function (nodes[i].u.f); |
1254 | break; | |
1255 | ||
1256 | case ORDER_VAR: | |
1d416bd7 | 1257 | varpool_assemble_decl (nodes[i].u.v); |
56af936e | 1258 | break; |
1259 | ||
1260 | case ORDER_ASM: | |
1261 | assemble_asm (nodes[i].u.a->asm_str); | |
1262 | break; | |
1263 | ||
1264 | case ORDER_UNDEFINED: | |
1265 | break; | |
1266 | ||
1267 | default: | |
1268 | gcc_unreachable (); | |
1269 | } | |
1270 | } | |
4b4ea2db | 1271 | |
1272 | cgraph_asm_nodes = NULL; | |
56af936e | 1273 | } |
1274 | ||
b0cdf642 | 1275 | /* Return true when function body of DECL still needs to be kept around |
1276 | for later re-use. */ | |
1277 | bool | |
1278 | cgraph_preserve_function_body_p (tree decl) | |
1279 | { | |
1280 | struct cgraph_node *node; | |
8d8c4c8d | 1281 | |
1282 | gcc_assert (cgraph_global_info_ready); | |
b0cdf642 | 1283 | /* Look if there is any clone around. */ |
ccf4ab6b | 1284 | node = cgraph_node (decl); |
1285 | if (node->clones) | |
1286 | return true; | |
b0cdf642 | 1287 | return false; |
1288 | } | |
1289 | ||
77fce4cd | 1290 | static void |
1291 | ipa_passes (void) | |
1292 | { | |
87d4aa85 | 1293 | set_cfun (NULL); |
4b14adf9 | 1294 | current_function_decl = NULL; |
75a70cf9 | 1295 | gimple_register_cfg_hooks (); |
77fce4cd | 1296 | bitmap_obstack_initialize (NULL); |
1297 | execute_ipa_pass_list (all_ipa_passes); | |
9ed5b1f5 | 1298 | |
1299 | /* Generate coverage variables and constructors. */ | |
1300 | coverage_finish (); | |
1301 | ||
1302 | /* Process new functions added. */ | |
1303 | set_cfun (NULL); | |
1304 | current_function_decl = NULL; | |
1305 | cgraph_process_new_functions (); | |
1306 | ||
77fce4cd | 1307 | bitmap_obstack_release (NULL); |
1308 | } | |
1309 | ||
ae01b312 | 1310 | /* Perform simple optimizations based on callgraph. */ |
1311 | ||
1312 | void | |
d9d9733a | 1313 | cgraph_optimize (void) |
ae01b312 | 1314 | { |
cb2b5570 | 1315 | if (errorcount || sorrycount) |
1316 | return; | |
1317 | ||
b0cdf642 | 1318 | #ifdef ENABLE_CHECKING |
1319 | verify_cgraph (); | |
1320 | #endif | |
a861fe52 | 1321 | |
1322 | /* Call functions declared with the "constructor" or "destructor" | |
1323 | attribute. */ | |
1324 | cgraph_build_cdtor_fns (); | |
e9f08e82 | 1325 | |
c1dcd13c | 1326 | /* Frontend may output common variables after the unit has been finalized. |
1327 | It is safe to deal with them here as they are always zero initialized. */ | |
1d416bd7 | 1328 | varpool_analyze_pending_decls (); |
aeeb194b | 1329 | cgraph_analyze_functions (); |
e9f08e82 | 1330 | |
f79b6507 | 1331 | timevar_push (TV_CGRAPHOPT); |
51949610 | 1332 | if (pre_ipa_mem_report) |
1333 | { | |
1334 | fprintf (stderr, "Memory consumption before IPA\n"); | |
1335 | dump_memory_report (false); | |
1336 | } | |
d7c6d889 | 1337 | if (!quiet_flag) |
cd6bca02 | 1338 | fprintf (stderr, "Performing interprocedural optimizations\n"); |
523c1122 | 1339 | cgraph_state = CGRAPH_STATE_IPA; |
c04e3894 | 1340 | |
be4d0974 | 1341 | /* Don't run the IPA passes if there was any error or sorry messages. */ |
1342 | if (errorcount == 0 && sorrycount == 0) | |
1343 | ipa_passes (); | |
1344 | ||
e1be32b8 | 1345 | /* This pass remove bodies of extern inline functions we never inlined. |
1346 | Do this later so other IPA passes see what is really going on. */ | |
1347 | cgraph_remove_unreachable_nodes (false, dump_file); | |
80a85d8a | 1348 | cgraph_global_info_ready = true; |
f79b6507 | 1349 | if (cgraph_dump_file) |
1350 | { | |
e4200070 | 1351 | fprintf (cgraph_dump_file, "Optimized "); |
f79b6507 | 1352 | dump_cgraph (cgraph_dump_file); |
c1dcd13c | 1353 | dump_varpool (cgraph_dump_file); |
f79b6507 | 1354 | } |
51949610 | 1355 | if (post_ipa_mem_report) |
1356 | { | |
defa2fa6 | 1357 | fprintf (stderr, "Memory consumption after IPA\n"); |
51949610 | 1358 | dump_memory_report (false); |
1359 | } | |
f79b6507 | 1360 | timevar_pop (TV_CGRAPHOPT); |
ae01b312 | 1361 | |
d7c6d889 | 1362 | /* Output everything. */ |
e4200070 | 1363 | if (!quiet_flag) |
1364 | fprintf (stderr, "Assembling functions:\n"); | |
b0cdf642 | 1365 | #ifdef ENABLE_CHECKING |
1366 | verify_cgraph (); | |
1367 | #endif | |
56af936e | 1368 | |
ccf4ab6b | 1369 | cgraph_materialize_all_clones (); |
acc70efa | 1370 | cgraph_mark_functions_to_output (); |
c1dcd13c | 1371 | |
523c1122 | 1372 | cgraph_state = CGRAPH_STATE_EXPANSION; |
56af936e | 1373 | if (!flag_toplevel_reorder) |
1374 | cgraph_output_in_order (); | |
1375 | else | |
1376 | { | |
1377 | cgraph_output_pending_asms (); | |
1378 | ||
1379 | cgraph_expand_all_functions (); | |
1d416bd7 | 1380 | varpool_remove_unreferenced_decls (); |
56af936e | 1381 | |
1d416bd7 | 1382 | varpool_assemble_pending_decls (); |
56af936e | 1383 | } |
523c1122 | 1384 | cgraph_process_new_functions (); |
1385 | cgraph_state = CGRAPH_STATE_FINISHED; | |
c1dcd13c | 1386 | |
f79b6507 | 1387 | if (cgraph_dump_file) |
1388 | { | |
e4200070 | 1389 | fprintf (cgraph_dump_file, "\nFinal "); |
f79b6507 | 1390 | dump_cgraph (cgraph_dump_file); |
1391 | } | |
b0cdf642 | 1392 | #ifdef ENABLE_CHECKING |
1393 | verify_cgraph (); | |
4ee9c684 | 1394 | /* Double check that all inline clones are gone and that all |
1395 | function bodies have been released from memory. */ | |
6329636b | 1396 | if (!(sorrycount || errorcount)) |
4ee9c684 | 1397 | { |
1398 | struct cgraph_node *node; | |
1399 | bool error_found = false; | |
1400 | ||
1401 | for (node = cgraph_nodes; node; node = node->next) | |
1402 | if (node->analyzed | |
1403 | && (node->global.inlined_to | |
1a1a827a | 1404 | || gimple_has_body_p (node->decl))) |
4ee9c684 | 1405 | { |
1406 | error_found = true; | |
1407 | dump_cgraph_node (stderr, node); | |
a0c938f0 | 1408 | } |
4ee9c684 | 1409 | if (error_found) |
c04e3894 | 1410 | internal_error ("nodes with unreleased memory found"); |
4ee9c684 | 1411 | } |
b0cdf642 | 1412 | #endif |
ae01b312 | 1413 | } |
2c56f72e | 1414 | /* Generate and emit a static constructor or destructor. WHICH must |
1415 | be one of 'I' (for a constructor) or 'D' (for a destructor). BODY | |
1416 | is a STATEMENT_LIST containing GENERIC statements. PRIORITY is the | |
f0b5f617 | 1417 | initialization priority for this constructor or destructor. */ |
b5530559 | 1418 | |
1419 | void | |
c5344b58 | 1420 | cgraph_build_static_cdtor (char which, tree body, int priority) |
b5530559 | 1421 | { |
1422 | static int counter = 0; | |
1423 | char which_buf[16]; | |
540edea7 | 1424 | tree decl, name, resdecl; |
b5530559 | 1425 | |
2c56f72e | 1426 | /* The priority is encoded in the constructor or destructor name. |
1427 | collect2 will sort the names and arrange that they are called at | |
1428 | program startup. */ | |
1429 | sprintf (which_buf, "%c_%.5d_%d", which, priority, counter++); | |
db85cc4f | 1430 | name = get_file_function_name (which_buf); |
b5530559 | 1431 | |
e60a6f7b | 1432 | decl = build_decl (input_location, FUNCTION_DECL, name, |
b5530559 | 1433 | build_function_type (void_type_node, void_list_node)); |
1434 | current_function_decl = decl; | |
1435 | ||
e60a6f7b | 1436 | resdecl = build_decl (input_location, |
1437 | RESULT_DECL, NULL_TREE, void_type_node); | |
540edea7 | 1438 | DECL_ARTIFICIAL (resdecl) = 1; |
540edea7 | 1439 | DECL_RESULT (decl) = resdecl; |
8e5b4ed6 | 1440 | DECL_CONTEXT (resdecl) = decl; |
540edea7 | 1441 | |
80f2ef47 | 1442 | allocate_struct_function (decl, false); |
b5530559 | 1443 | |
1444 | TREE_STATIC (decl) = 1; | |
1445 | TREE_USED (decl) = 1; | |
1446 | DECL_ARTIFICIAL (decl) = 1; | |
b5530559 | 1447 | DECL_NO_INSTRUMENT_FUNCTION_ENTRY_EXIT (decl) = 1; |
1448 | DECL_SAVED_TREE (decl) = body; | |
1449 | TREE_PUBLIC (decl) = ! targetm.have_ctors_dtors; | |
1450 | DECL_UNINLINABLE (decl) = 1; | |
1451 | ||
1452 | DECL_INITIAL (decl) = make_node (BLOCK); | |
1453 | TREE_USED (DECL_INITIAL (decl)) = 1; | |
1454 | ||
1455 | DECL_SOURCE_LOCATION (decl) = input_location; | |
1456 | cfun->function_end_locus = input_location; | |
1457 | ||
cc636d56 | 1458 | switch (which) |
1459 | { | |
1460 | case 'I': | |
1461 | DECL_STATIC_CONSTRUCTOR (decl) = 1; | |
64c2e9b0 | 1462 | decl_init_priority_insert (decl, priority); |
cc636d56 | 1463 | break; |
1464 | case 'D': | |
1465 | DECL_STATIC_DESTRUCTOR (decl) = 1; | |
64c2e9b0 | 1466 | decl_fini_priority_insert (decl, priority); |
cc636d56 | 1467 | break; |
1468 | default: | |
1469 | gcc_unreachable (); | |
1470 | } | |
b5530559 | 1471 | |
1472 | gimplify_function_tree (decl); | |
1473 | ||
523c1122 | 1474 | cgraph_add_new_function (decl, false); |
1475 | cgraph_mark_needed_node (cgraph_node (decl)); | |
e3a37aef | 1476 | set_cfun (NULL); |
b5530559 | 1477 | } |
121f3051 | 1478 | |
1479 | void | |
1480 | init_cgraph (void) | |
1481 | { | |
1482 | cgraph_dump_file = dump_begin (TDI_cgraph, NULL); | |
1483 | } | |
b5d36404 | 1484 | |
a0c938f0 | 1485 | /* The edges representing the callers of the NEW_VERSION node were |
b5d36404 | 1486 | fixed by cgraph_function_versioning (), now the call_expr in their |
1487 | respective tree code should be updated to call the NEW_VERSION. */ | |
1488 | ||
1489 | static void | |
1490 | update_call_expr (struct cgraph_node *new_version) | |
1491 | { | |
1492 | struct cgraph_edge *e; | |
1493 | ||
1494 | gcc_assert (new_version); | |
75a70cf9 | 1495 | |
1496 | /* Update the call expr on the edges to call the new version. */ | |
b5d36404 | 1497 | for (e = new_version->callers; e; e = e->next_caller) |
e03a95e7 | 1498 | { |
1499 | struct function *inner_function = DECL_STRUCT_FUNCTION (e->caller->decl); | |
1500 | gimple_call_set_fndecl (e->call_stmt, new_version->decl); | |
1501 | /* Update EH information too, just in case. */ | |
1502 | if (!stmt_could_throw_p (e->call_stmt) | |
1503 | && lookup_stmt_eh_region_fn (inner_function, e->call_stmt)) | |
1504 | remove_stmt_from_eh_region_fn (inner_function, e->call_stmt); | |
1505 | } | |
b5d36404 | 1506 | } |
1507 | ||
1508 | ||
1509 | /* Create a new cgraph node which is the new version of | |
1510 | OLD_VERSION node. REDIRECT_CALLERS holds the callers | |
1511 | edges which should be redirected to point to | |
1512 | NEW_VERSION. ALL the callees edges of OLD_VERSION | |
1513 | are cloned to the new version node. Return the new | |
1514 | version node. */ | |
1515 | ||
1516 | static struct cgraph_node * | |
1517 | cgraph_copy_node_for_versioning (struct cgraph_node *old_version, | |
4460a647 | 1518 | tree new_decl, |
1519 | VEC(cgraph_edge_p,heap) *redirect_callers) | |
b5d36404 | 1520 | { |
1521 | struct cgraph_node *new_version; | |
1522 | struct cgraph_edge *e, *new_e; | |
1523 | struct cgraph_edge *next_callee; | |
1524 | unsigned i; | |
1525 | ||
1526 | gcc_assert (old_version); | |
a0c938f0 | 1527 | |
b5d36404 | 1528 | new_version = cgraph_node (new_decl); |
1529 | ||
1530 | new_version->analyzed = true; | |
1531 | new_version->local = old_version->local; | |
1532 | new_version->global = old_version->global; | |
1533 | new_version->rtl = new_version->rtl; | |
1534 | new_version->reachable = true; | |
1535 | new_version->count = old_version->count; | |
1536 | ||
1537 | /* Clone the old node callees. Recursive calls are | |
1538 | also cloned. */ | |
1539 | for (e = old_version->callees;e; e=e->next_callee) | |
1540 | { | |
4ae20857 | 1541 | new_e = cgraph_clone_edge (e, new_version, e->call_stmt, 0, e->frequency, |
1542 | e->loop_nest, true); | |
b5d36404 | 1543 | new_e->count = e->count; |
1544 | } | |
1545 | /* Fix recursive calls. | |
1546 | If OLD_VERSION has a recursive call after the | |
1547 | previous edge cloning, the new version will have an edge | |
1548 | pointing to the old version, which is wrong; | |
1549 | Redirect it to point to the new version. */ | |
1550 | for (e = new_version->callees ; e; e = next_callee) | |
1551 | { | |
1552 | next_callee = e->next_callee; | |
1553 | if (e->callee == old_version) | |
1554 | cgraph_redirect_edge_callee (e, new_version); | |
a0c938f0 | 1555 | |
b5d36404 | 1556 | if (!next_callee) |
1557 | break; | |
1558 | } | |
4460a647 | 1559 | for (i = 0; VEC_iterate (cgraph_edge_p, redirect_callers, i, e); i++) |
1560 | { | |
1561 | /* Redirect calls to the old version node to point to its new | |
1562 | version. */ | |
1563 | cgraph_redirect_edge_callee (e, new_version); | |
1564 | } | |
b5d36404 | 1565 | |
1566 | return new_version; | |
1567 | } | |
1568 | ||
1569 | /* Perform function versioning. | |
a0c938f0 | 1570 | Function versioning includes copying of the tree and |
b5d36404 | 1571 | a callgraph update (creating a new cgraph node and updating |
1572 | its callees and callers). | |
1573 | ||
1574 | REDIRECT_CALLERS varray includes the edges to be redirected | |
1575 | to the new version. | |
1576 | ||
1577 | TREE_MAP is a mapping of tree nodes we want to replace with | |
1578 | new ones (according to results of prior analysis). | |
1579 | OLD_VERSION_NODE is the node that is versioned. | |
5afe38fe | 1580 | It returns the new version's cgraph node. |
1581 | ARGS_TO_SKIP lists arguments to be omitted from functions | |
1582 | */ | |
b5d36404 | 1583 | |
1584 | struct cgraph_node * | |
1585 | cgraph_function_versioning (struct cgraph_node *old_version_node, | |
4460a647 | 1586 | VEC(cgraph_edge_p,heap) *redirect_callers, |
ccf4ab6b | 1587 | VEC (ipa_replace_map_p,gc)* tree_map, |
5afe38fe | 1588 | bitmap args_to_skip) |
b5d36404 | 1589 | { |
1590 | tree old_decl = old_version_node->decl; | |
1591 | struct cgraph_node *new_version_node = NULL; | |
1592 | tree new_decl; | |
1593 | ||
1594 | if (!tree_versionable_function_p (old_decl)) | |
1595 | return NULL; | |
1596 | ||
1597 | /* Make a new FUNCTION_DECL tree node for the | |
1598 | new version. */ | |
5afe38fe | 1599 | if (!args_to_skip) |
1600 | new_decl = copy_node (old_decl); | |
1601 | else | |
1602 | new_decl = build_function_decl_skip_args (old_decl, args_to_skip); | |
b5d36404 | 1603 | |
1604 | /* Create the new version's call-graph node. | |
1605 | and update the edges of the new node. */ | |
1606 | new_version_node = | |
1607 | cgraph_copy_node_for_versioning (old_version_node, new_decl, | |
1608 | redirect_callers); | |
1609 | ||
1610 | /* Copy the OLD_VERSION_NODE function tree to the new version. */ | |
5afe38fe | 1611 | tree_function_versioning (old_decl, new_decl, tree_map, false, args_to_skip); |
b5d36404 | 1612 | |
a0c938f0 | 1613 | /* Update the new version's properties. |
e03a95e7 | 1614 | Make The new version visible only within this translation unit. Make sure |
1615 | that is not weak also. | |
a0c938f0 | 1616 | ??? We cannot use COMDAT linkage because there is no |
b5d36404 | 1617 | ABI support for this. */ |
1618 | DECL_EXTERNAL (new_version_node->decl) = 0; | |
1619 | DECL_ONE_ONLY (new_version_node->decl) = 0; | |
1620 | TREE_PUBLIC (new_version_node->decl) = 0; | |
1621 | DECL_COMDAT (new_version_node->decl) = 0; | |
e03a95e7 | 1622 | DECL_WEAK (new_version_node->decl) = 0; |
f014e39d | 1623 | DECL_VIRTUAL_P (new_version_node->decl) = 0; |
b5d36404 | 1624 | new_version_node->local.externally_visible = 0; |
1625 | new_version_node->local.local = 1; | |
1626 | new_version_node->lowered = true; | |
f014e39d | 1627 | |
e03a95e7 | 1628 | /* Update the call_expr on the edges to call the new version node. */ |
1629 | update_call_expr (new_version_node); | |
1630 | ||
50828ed8 | 1631 | cgraph_call_function_insertion_hooks (new_version_node); |
b5d36404 | 1632 | return new_version_node; |
1633 | } | |
469679ab | 1634 | |
1635 | /* Produce separate function body for inline clones so the offline copy can be | |
1636 | modified without affecting them. */ | |
1637 | struct cgraph_node * | |
1638 | save_inline_function_body (struct cgraph_node *node) | |
1639 | { | |
ccf4ab6b | 1640 | struct cgraph_node *first_clone, *n; |
469679ab | 1641 | |
1642 | gcc_assert (node == cgraph_node (node->decl)); | |
1643 | ||
1644 | cgraph_lower_function (node); | |
1645 | ||
ccf4ab6b | 1646 | first_clone = node->clones; |
469679ab | 1647 | |
1648 | first_clone->decl = copy_node (node->decl); | |
469679ab | 1649 | cgraph_insert_node_to_hashtable (first_clone); |
1650 | gcc_assert (first_clone == cgraph_node (first_clone->decl)); | |
ccf4ab6b | 1651 | if (first_clone->next_sibling_clone) |
1652 | { | |
1653 | for (n = first_clone->next_sibling_clone; n->next_sibling_clone; n = n->next_sibling_clone) | |
1654 | n->clone_of = first_clone; | |
1655 | n->clone_of = first_clone; | |
1656 | n->next_sibling_clone = first_clone->clones; | |
1657 | if (first_clone->clones) | |
1658 | first_clone->clones->prev_sibling_clone = n; | |
1659 | first_clone->clones = first_clone->next_sibling_clone; | |
1660 | first_clone->next_sibling_clone->prev_sibling_clone = NULL; | |
1661 | first_clone->next_sibling_clone = NULL; | |
1662 | gcc_assert (!first_clone->prev_sibling_clone); | |
1663 | } | |
1664 | first_clone->clone_of = NULL; | |
1665 | node->clones = NULL; | |
1666 | ||
1667 | if (first_clone->clones) | |
1668 | for (n = first_clone->clones; n != first_clone;) | |
1669 | { | |
1670 | gcc_assert (n->decl == node->decl); | |
1671 | n->decl = first_clone->decl; | |
1672 | if (n->clones) | |
1673 | n = n->clones; | |
1674 | else if (n->next_sibling_clone) | |
1675 | n = n->next_sibling_clone; | |
1676 | else | |
1677 | { | |
1678 | while (n != first_clone && !n->next_sibling_clone) | |
1679 | n = n->clone_of; | |
1680 | if (n != first_clone) | |
1681 | n = n->next_sibling_clone; | |
1682 | } | |
1683 | } | |
469679ab | 1684 | |
1685 | /* Copy the OLD_VERSION_NODE function tree to the new version. */ | |
5afe38fe | 1686 | tree_function_versioning (node->decl, first_clone->decl, NULL, true, NULL); |
469679ab | 1687 | |
1688 | DECL_EXTERNAL (first_clone->decl) = 0; | |
1689 | DECL_ONE_ONLY (first_clone->decl) = 0; | |
1690 | TREE_PUBLIC (first_clone->decl) = 0; | |
1691 | DECL_COMDAT (first_clone->decl) = 0; | |
7fe9b425 | 1692 | VEC_free (ipa_opt_pass, heap, |
1693 | DECL_STRUCT_FUNCTION (first_clone->decl)->ipa_transforms_to_apply); | |
1694 | DECL_STRUCT_FUNCTION (first_clone->decl)->ipa_transforms_to_apply = NULL; | |
469679ab | 1695 | |
469679ab | 1696 | #ifdef ENABLE_CHECKING |
1697 | verify_cgraph_node (first_clone); | |
1698 | #endif | |
1699 | return first_clone; | |
1700 | } | |
a861fe52 | 1701 | |
ccf4ab6b | 1702 | /* Given virtual clone, turn it into actual clone. */ |
1703 | static void | |
1704 | cgraph_materialize_clone (struct cgraph_node *node) | |
1705 | { | |
1706 | bitmap_obstack_initialize (NULL); | |
1707 | /* Copy the OLD_VERSION_NODE function tree to the new version. */ | |
1708 | tree_function_versioning (node->clone_of->decl, node->decl, | |
1709 | node->clone.tree_map, true, | |
1710 | node->clone.args_to_skip); | |
1711 | ||
1712 | /* Function is no longer clone. */ | |
1713 | if (node->next_sibling_clone) | |
1714 | node->next_sibling_clone->prev_sibling_clone = node->prev_sibling_clone; | |
1715 | if (node->prev_sibling_clone) | |
1716 | node->prev_sibling_clone->next_sibling_clone = node->next_sibling_clone; | |
1717 | else | |
1718 | node->clone_of->clones = node->next_sibling_clone; | |
1719 | node->next_sibling_clone = NULL; | |
1720 | node->prev_sibling_clone = NULL; | |
1721 | node->clone_of = NULL; | |
1722 | bitmap_obstack_release (NULL); | |
1723 | } | |
1724 | ||
1725 | /* Once all functions from compilation unit are in memory, produce all clones | |
1726 | and update all calls. | |
1727 | We might also do this on demand if we don't want to bring all functions to | |
1728 | memory prior compilation, but current WHOPR implementation does that and it is | |
1729 | is bit easier to keep everything right in this order. */ | |
1730 | void | |
1731 | cgraph_materialize_all_clones (void) | |
1732 | { | |
1733 | struct cgraph_node *node; | |
1734 | bool stabilized = false; | |
1735 | ||
1736 | if (cgraph_dump_file) | |
1737 | fprintf (cgraph_dump_file, "Materializing clones\n"); | |
1738 | #ifdef ENABLE_CHECKING | |
1739 | verify_cgraph (); | |
1740 | #endif | |
1741 | ||
1742 | /* We can also do topological order, but number of iterations should be | |
1743 | bounded by number of IPA passes since single IPA pass is probably not | |
1744 | going to create clones of clones it created itself. */ | |
1745 | while (!stabilized) | |
1746 | { | |
1747 | stabilized = true; | |
1748 | for (node = cgraph_nodes; node; node = node->next) | |
1749 | { | |
1750 | if (node->clone_of && node->decl != node->clone_of->decl | |
1751 | && !gimple_has_body_p (node->decl)) | |
1752 | { | |
1753 | if (gimple_has_body_p (node->clone_of->decl)) | |
1754 | { | |
1755 | if (cgraph_dump_file) | |
1756 | fprintf (cgraph_dump_file, " clonning %s to %s", | |
1757 | cgraph_node_name (node->clone_of), | |
1758 | cgraph_node_name (node)); | |
1759 | cgraph_materialize_clone (node); | |
1760 | } | |
1761 | else | |
1762 | stabilized = false; | |
1763 | } | |
1764 | } | |
1765 | } | |
1766 | if (cgraph_dump_file) | |
1767 | fprintf (cgraph_dump_file, "Updating call sites\n"); | |
1768 | for (node = cgraph_nodes; node; node = node->next) | |
1769 | if (node->analyzed && gimple_has_body_p (node->decl) | |
1770 | && (!node->clone_of || node->clone_of->decl != node->decl)) | |
1771 | { | |
1772 | struct cgraph_edge *e; | |
1773 | ||
1774 | current_function_decl = node->decl; | |
1775 | push_cfun (DECL_STRUCT_FUNCTION (node->decl)); | |
1776 | for (e = node->callees; e; e = e->next_callee) | |
1777 | { | |
1778 | tree decl = gimple_call_fndecl (e->call_stmt); | |
947781ac | 1779 | /* When function gets inlined, indirect inlining might've invented |
1780 | new edge for orginally indirect stmt. Since we are not | |
1781 | preserving clones in the original form, we must not update here | |
1782 | since other inline clones don't need to contain call to the same | |
1783 | call. Inliner will do the substitution for us later. */ | |
1784 | if (decl && decl != e->callee->decl) | |
ccf4ab6b | 1785 | { |
1786 | gimple new_stmt; | |
1787 | gimple_stmt_iterator gsi; | |
1788 | ||
1789 | if (cgraph_dump_file) | |
1790 | { | |
1791 | fprintf (cgraph_dump_file, "updating call of %s in %s:", | |
1792 | cgraph_node_name (node), | |
1793 | cgraph_node_name (e->callee)); | |
1794 | print_gimple_stmt (cgraph_dump_file, e->call_stmt, 0, dump_flags); | |
1795 | } | |
1796 | ||
1797 | if (e->callee->clone.args_to_skip) | |
1798 | new_stmt = gimple_call_copy_skip_args (e->call_stmt, | |
1799 | e->callee->clone.args_to_skip); | |
1800 | else | |
1801 | new_stmt = e->call_stmt; | |
1802 | if (gimple_vdef (new_stmt) | |
1803 | && TREE_CODE (gimple_vdef (new_stmt)) == SSA_NAME) | |
1804 | SSA_NAME_DEF_STMT (gimple_vdef (new_stmt)) = new_stmt; | |
1805 | gimple_call_set_fndecl (new_stmt, e->callee->decl); | |
1806 | ||
1807 | gsi = gsi_for_stmt (e->call_stmt); | |
1808 | gsi_replace (&gsi, new_stmt, true); | |
1809 | ||
1810 | /* Update EH information too, just in case. */ | |
1811 | if (!stmt_could_throw_p (new_stmt) | |
1812 | && lookup_stmt_eh_region (new_stmt)) | |
1813 | remove_stmt_from_eh_region (new_stmt); | |
1814 | ||
1815 | cgraph_set_call_stmt_including_clones (node, e->call_stmt, new_stmt); | |
1816 | ||
1817 | if (cgraph_dump_file) | |
1818 | { | |
1819 | fprintf (cgraph_dump_file, " updated to:"); | |
1820 | print_gimple_stmt (cgraph_dump_file, e->call_stmt, 0, dump_flags); | |
1821 | } | |
1822 | } | |
1823 | } | |
1824 | pop_cfun (); | |
1825 | current_function_decl = NULL; | |
1826 | #ifdef ENABLE_CHECKING | |
1827 | verify_cgraph_node (node); | |
1828 | #endif | |
1829 | } | |
947781ac | 1830 | #ifdef ENABLE_CHECKING |
1831 | verify_cgraph (); | |
1832 | #endif | |
ccf4ab6b | 1833 | cgraph_remove_unreachable_nodes (false, cgraph_dump_file); |
1834 | } | |
1835 | ||
a861fe52 | 1836 | #include "gt-cgraphunit.h" |