]>
Commit | Line | Data |
---|---|---|
1 | /* Callgraph based intraprocedural optimizations. | |
2 | Copyright (C) 2003, 2004, 2005 Free Software Foundation, Inc. | |
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 2, 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 COPYING. If not, write to the Free | |
19 | Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA | |
20 | 02110-1301, USA. */ | |
21 | ||
22 | /* This module implements main driver of compilation process as well as | |
23 | few basic intraprocedural optimizers. | |
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 | ||
35 | (There is one exception needed for implementing GCC extern inline function.) | |
36 | ||
37 | - cgraph_varpool_finalize_variable | |
38 | ||
39 | This function has same behavior as the above but is used for static | |
40 | variables. | |
41 | ||
42 | - cgraph_finalize_compilation_unit | |
43 | ||
44 | This function is called once compilation unit is finalized and it will | |
45 | no longer change. | |
46 | ||
47 | In the unit-at-a-time the call-graph construction and local function | |
48 | analysis takes place here. Bodies of unreachable functions are released | |
49 | to conserve memory usage. | |
50 | ||
51 | ??? The compilation unit in this point of view should be compilation | |
52 | unit as defined by the language - for instance C frontend allows multiple | |
53 | compilation units to be parsed at once and it should call function each | |
54 | time parsing is done so we save memory. | |
55 | ||
56 | - cgraph_optimize | |
57 | ||
58 | In this unit-at-a-time compilation the intra procedural analysis takes | |
59 | place here. In particular the static functions whose address is never | |
60 | taken are marked as local. Backend can then use this information to | |
61 | modify calling conventions, do better inlining or similar optimizations. | |
62 | ||
63 | - cgraph_assemble_pending_functions | |
64 | - cgraph_varpool_assemble_pending_variables | |
65 | ||
66 | In non-unit-at-a-time mode these functions can be used to force compilation | |
67 | of functions or variables that are known to be needed at given stage | |
68 | of compilation | |
69 | ||
70 | - cgraph_mark_needed_node | |
71 | - cgraph_varpool_mark_needed_node | |
72 | ||
73 | When function or variable is referenced by some hidden way (for instance | |
74 | via assembly code and marked by attribute "used"), the call-graph data structure | |
75 | must be updated accordingly by this function. | |
76 | ||
77 | - analyze_expr callback | |
78 | ||
79 | This function is responsible for lowering tree nodes not understood by | |
80 | generic code into understandable ones or alternatively marking | |
81 | callgraph and varpool nodes referenced by the as needed. | |
82 | ||
83 | ??? On the tree-ssa genericizing should take place here and we will avoid | |
84 | need for these hooks (replacing them by genericizing hook) | |
85 | ||
86 | - expand_function callback | |
87 | ||
88 | This function is used to expand function and pass it into RTL back-end. | |
89 | Front-end should not make any assumptions about when this function can be | |
90 | called. In particular cgraph_assemble_pending_functions, | |
91 | cgraph_varpool_assemble_pending_variables, cgraph_finalize_function, | |
92 | cgraph_varpool_finalize_function, cgraph_optimize can cause arbitrarily | |
93 | previously finalized functions to be expanded. | |
94 | ||
95 | We implement two compilation modes. | |
96 | ||
97 | - unit-at-a-time: In this mode analyzing of all functions is deferred | |
98 | to cgraph_finalize_compilation_unit and expansion into cgraph_optimize. | |
99 | ||
100 | In cgraph_finalize_compilation_unit the reachable functions are | |
101 | analyzed. During analysis the call-graph edges from reachable | |
102 | functions are constructed and their destinations are marked as | |
103 | reachable. References to functions and variables are discovered too | |
104 | and variables found to be needed output to the assembly file. Via | |
105 | mark_referenced call in assemble_variable functions referenced by | |
106 | static variables are noticed too. | |
107 | ||
108 | The intra-procedural information is produced and its existence | |
109 | indicated by global_info_ready. Once this flag is set it is impossible | |
110 | to change function from !reachable to reachable and thus | |
111 | assemble_variable no longer call mark_referenced. | |
112 | ||
113 | Finally the call-graph is topologically sorted and all reachable functions | |
114 | that has not been completely inlined or are not external are output. | |
115 | ||
116 | ??? It is possible that reference to function or variable is optimized | |
117 | out. We can not deal with this nicely because topological order is not | |
118 | suitable for it. For tree-ssa we may consider another pass doing | |
119 | optimization and re-discovering reachable functions. | |
120 | ||
121 | ??? Reorganize code so variables are output very last and only if they | |
122 | really has been referenced by produced code, so we catch more cases | |
123 | where reference has been optimized out. | |
124 | ||
125 | - non-unit-at-a-time | |
126 | ||
127 | All functions are variables are output as early as possible to conserve | |
128 | memory consumption. This may or may not result in less memory used but | |
129 | it is still needed for some legacy code that rely on particular ordering | |
130 | of things output from the compiler. | |
131 | ||
132 | Varpool data structures are not used and variables are output directly. | |
133 | ||
134 | Functions are output early using call of | |
135 | cgraph_assemble_pending_function from cgraph_finalize_function. The | |
136 | decision on whether function is needed is made more conservative so | |
137 | uninlininable static functions are needed too. During the call-graph | |
138 | construction the edge destinations are not marked as reachable and it | |
139 | is completely relied upn assemble_variable to mark them. */ | |
140 | ||
141 | ||
142 | #include "config.h" | |
143 | #include "system.h" | |
144 | #include "coretypes.h" | |
145 | #include "tm.h" | |
146 | #include "tree.h" | |
147 | #include "rtl.h" | |
148 | #include "tree-flow.h" | |
149 | #include "tree-inline.h" | |
150 | #include "langhooks.h" | |
151 | #include "pointer-set.h" | |
152 | #include "toplev.h" | |
153 | #include "flags.h" | |
154 | #include "ggc.h" | |
155 | #include "debug.h" | |
156 | #include "target.h" | |
157 | #include "cgraph.h" | |
158 | #include "diagnostic.h" | |
159 | #include "timevar.h" | |
160 | #include "params.h" | |
161 | #include "fibheap.h" | |
162 | #include "c-common.h" | |
163 | #include "intl.h" | |
164 | #include "function.h" | |
165 | #include "ipa-prop.h" | |
166 | #include "tree-gimple.h" | |
167 | #include "tree-pass.h" | |
168 | #include "output.h" | |
169 | ||
170 | static void cgraph_expand_all_functions (void); | |
171 | static void cgraph_mark_functions_to_output (void); | |
172 | static void cgraph_expand_function (struct cgraph_node *); | |
173 | static tree record_reference (tree *, int *, void *); | |
174 | ||
175 | /* Records tree nodes seen in record_reference. Simply using | |
176 | walk_tree_without_duplicates doesn't guarantee each node is visited | |
177 | once because it gets a new htab upon each recursive call from | |
178 | record_reference itself. */ | |
179 | static struct pointer_set_t *visited_nodes; | |
180 | ||
181 | static FILE *cgraph_dump_file; | |
182 | ||
183 | /* Determine if function DECL is needed. That is, visible to something | |
184 | either outside this translation unit, something magic in the system | |
185 | configury, or (if not doing unit-at-a-time) to something we havn't | |
186 | seen yet. */ | |
187 | ||
188 | static bool | |
189 | decide_is_function_needed (struct cgraph_node *node, tree decl) | |
190 | { | |
191 | tree origin; | |
192 | if (MAIN_NAME_P (DECL_NAME (decl)) | |
193 | && TREE_PUBLIC (decl)) | |
194 | { | |
195 | node->local.externally_visible = true; | |
196 | return true; | |
197 | } | |
198 | ||
199 | /* If the user told us it is used, then it must be so. */ | |
200 | if (node->local.externally_visible | |
201 | || lookup_attribute ("used", DECL_ATTRIBUTES (decl))) | |
202 | return true; | |
203 | ||
204 | /* ??? If the assembler name is set by hand, it is possible to assemble | |
205 | the name later after finalizing the function and the fact is noticed | |
206 | in assemble_name then. This is arguably a bug. */ | |
207 | if (DECL_ASSEMBLER_NAME_SET_P (decl) | |
208 | && TREE_SYMBOL_REFERENCED (DECL_ASSEMBLER_NAME (decl))) | |
209 | return true; | |
210 | ||
211 | /* If we decided it was needed before, but at the time we didn't have | |
212 | the body of the function available, then it's still needed. We have | |
213 | to go back and re-check its dependencies now. */ | |
214 | if (node->needed) | |
215 | return true; | |
216 | ||
217 | /* Externally visible functions must be output. The exception is | |
218 | COMDAT functions that must be output only when they are needed. */ | |
219 | if ((TREE_PUBLIC (decl) && !flag_whole_program) | |
220 | && !DECL_COMDAT (decl) && !DECL_EXTERNAL (decl)) | |
221 | return true; | |
222 | ||
223 | /* Constructors and destructors are reachable from the runtime by | |
224 | some mechanism. */ | |
225 | if (DECL_STATIC_CONSTRUCTOR (decl) || DECL_STATIC_DESTRUCTOR (decl)) | |
226 | return true; | |
227 | ||
228 | if (flag_unit_at_a_time) | |
229 | return false; | |
230 | ||
231 | /* If not doing unit at a time, then we'll only defer this function | |
232 | if its marked for inlining. Otherwise we want to emit it now. */ | |
233 | ||
234 | /* "extern inline" functions are never output locally. */ | |
235 | if (DECL_EXTERNAL (decl)) | |
236 | return false; | |
237 | /* Nested functions of extern inline function shall not be emit unless | |
238 | we inlined the origin. */ | |
239 | for (origin = decl_function_context (decl); origin; | |
240 | origin = decl_function_context (origin)) | |
241 | if (DECL_EXTERNAL (origin)) | |
242 | return false; | |
243 | /* We want to emit COMDAT functions only when absolutely necessary. */ | |
244 | if (DECL_COMDAT (decl)) | |
245 | return false; | |
246 | if (!DECL_INLINE (decl) | |
247 | || (!node->local.disregard_inline_limits | |
248 | /* When declared inline, defer even the uninlinable functions. | |
249 | This allows them to be eliminated when unused. */ | |
250 | && !DECL_DECLARED_INLINE_P (decl) | |
251 | && (!node->local.inlinable || !cgraph_default_inline_p (node, NULL)))) | |
252 | return true; | |
253 | ||
254 | return false; | |
255 | } | |
256 | ||
257 | /* Walk the decls we marked as necessary and see if they reference new | |
258 | variables or functions and add them into the worklists. */ | |
259 | static bool | |
260 | cgraph_varpool_analyze_pending_decls (void) | |
261 | { | |
262 | bool changed = false; | |
263 | timevar_push (TV_CGRAPH); | |
264 | ||
265 | while (cgraph_varpool_first_unanalyzed_node) | |
266 | { | |
267 | tree decl = cgraph_varpool_first_unanalyzed_node->decl; | |
268 | ||
269 | cgraph_varpool_first_unanalyzed_node->analyzed = true; | |
270 | ||
271 | cgraph_varpool_first_unanalyzed_node = cgraph_varpool_first_unanalyzed_node->next_needed; | |
272 | ||
273 | if (DECL_INITIAL (decl)) | |
274 | { | |
275 | visited_nodes = pointer_set_create (); | |
276 | walk_tree (&DECL_INITIAL (decl), record_reference, NULL, visited_nodes); | |
277 | pointer_set_destroy (visited_nodes); | |
278 | visited_nodes = NULL; | |
279 | } | |
280 | changed = true; | |
281 | } | |
282 | timevar_pop (TV_CGRAPH); | |
283 | return changed; | |
284 | } | |
285 | ||
286 | /* Optimization of function bodies might've rendered some variables as | |
287 | unnecessary so we want to avoid these from being compiled. | |
288 | ||
289 | This is done by pruning the queue and keeping only the variables that | |
290 | really appear needed (ie they are either externally visible or referenced | |
291 | by compiled function). Re-doing the reachability analysis on variables | |
292 | brings back the remaining variables referenced by these. */ | |
293 | static void | |
294 | cgraph_varpool_remove_unreferenced_decls (void) | |
295 | { | |
296 | struct cgraph_varpool_node *next, *node = cgraph_varpool_nodes_queue; | |
297 | ||
298 | cgraph_varpool_reset_queue (); | |
299 | ||
300 | if (errorcount || sorrycount) | |
301 | return; | |
302 | ||
303 | while (node) | |
304 | { | |
305 | tree decl = node->decl; | |
306 | next = node->next_needed; | |
307 | node->needed = 0; | |
308 | ||
309 | if (node->finalized | |
310 | && ((DECL_ASSEMBLER_NAME_SET_P (decl) | |
311 | && TREE_SYMBOL_REFERENCED (DECL_ASSEMBLER_NAME (decl))) | |
312 | || node->force_output | |
313 | || decide_is_variable_needed (node, decl) | |
314 | /* ??? Cgraph does not yet rule the world with an iron hand, | |
315 | and does not control the emission of debug information. | |
316 | After a variable has its DECL_RTL set, we must assume that | |
317 | it may be referenced by the debug information, and we can | |
318 | no longer elide it. */ | |
319 | || DECL_RTL_SET_P (decl))) | |
320 | cgraph_varpool_mark_needed_node (node); | |
321 | ||
322 | node = next; | |
323 | } | |
324 | /* Make sure we mark alias targets as used targets. */ | |
325 | finish_aliases_1 (); | |
326 | cgraph_varpool_analyze_pending_decls (); | |
327 | } | |
328 | ||
329 | ||
330 | /* When not doing unit-at-a-time, output all functions enqueued. | |
331 | Return true when such a functions were found. */ | |
332 | ||
333 | bool | |
334 | cgraph_assemble_pending_functions (void) | |
335 | { | |
336 | bool output = false; | |
337 | ||
338 | if (flag_unit_at_a_time) | |
339 | return false; | |
340 | ||
341 | while (cgraph_nodes_queue) | |
342 | { | |
343 | struct cgraph_node *n = cgraph_nodes_queue; | |
344 | ||
345 | cgraph_nodes_queue = cgraph_nodes_queue->next_needed; | |
346 | n->next_needed = NULL; | |
347 | if (!n->global.inlined_to | |
348 | && !n->alias | |
349 | && !DECL_EXTERNAL (n->decl)) | |
350 | { | |
351 | cgraph_expand_function (n); | |
352 | output = true; | |
353 | } | |
354 | } | |
355 | ||
356 | /* Process CGRAPH_EXPAND_QUEUE, these are functions created during | |
357 | the expansion process. Note that this queue may grow as its | |
358 | being processed, as the new functions may generate new ones. */ | |
359 | while (cgraph_expand_queue) | |
360 | { | |
361 | struct cgraph_node *n = cgraph_expand_queue; | |
362 | cgraph_expand_queue = cgraph_expand_queue->next_needed; | |
363 | n->next_needed = NULL; | |
364 | cgraph_finalize_function (n->decl, false); | |
365 | output = true; | |
366 | } | |
367 | ||
368 | return output; | |
369 | } | |
370 | ||
371 | ||
372 | /* As an GCC extension we allow redefinition of the function. The | |
373 | semantics when both copies of bodies differ is not well defined. | |
374 | We replace the old body with new body so in unit at a time mode | |
375 | we always use new body, while in normal mode we may end up with | |
376 | old body inlined into some functions and new body expanded and | |
377 | inlined in others. | |
378 | ||
379 | ??? It may make more sense to use one body for inlining and other | |
380 | body for expanding the function but this is difficult to do. */ | |
381 | ||
382 | static void | |
383 | cgraph_reset_node (struct cgraph_node *node) | |
384 | { | |
385 | /* If node->output is set, then this is a unit-at-a-time compilation | |
386 | and we have already begun whole-unit analysis. This is *not* | |
387 | testing for whether we've already emitted the function. That | |
388 | case can be sort-of legitimately seen with real function | |
389 | redefinition errors. I would argue that the front end should | |
390 | never present us with such a case, but don't enforce that for now. */ | |
391 | gcc_assert (!node->output); | |
392 | ||
393 | /* Reset our data structures so we can analyze the function again. */ | |
394 | memset (&node->local, 0, sizeof (node->local)); | |
395 | memset (&node->global, 0, sizeof (node->global)); | |
396 | memset (&node->rtl, 0, sizeof (node->rtl)); | |
397 | node->analyzed = false; | |
398 | node->local.redefined_extern_inline = true; | |
399 | node->local.finalized = false; | |
400 | ||
401 | if (!flag_unit_at_a_time) | |
402 | { | |
403 | struct cgraph_node *n; | |
404 | ||
405 | for (n = cgraph_nodes; n; n = n->next) | |
406 | if (n->global.inlined_to == node) | |
407 | cgraph_remove_node (n); | |
408 | } | |
409 | ||
410 | cgraph_node_remove_callees (node); | |
411 | ||
412 | /* We may need to re-queue the node for assembling in case | |
413 | we already proceeded it and ignored as not needed. */ | |
414 | if (node->reachable && !flag_unit_at_a_time) | |
415 | { | |
416 | struct cgraph_node *n; | |
417 | ||
418 | for (n = cgraph_nodes_queue; n; n = n->next_needed) | |
419 | if (n == node) | |
420 | break; | |
421 | if (!n) | |
422 | node->reachable = 0; | |
423 | } | |
424 | } | |
425 | ||
426 | static void | |
427 | cgraph_lower_function (struct cgraph_node *node) | |
428 | { | |
429 | if (node->lowered) | |
430 | return; | |
431 | tree_lowering_passes (node->decl); | |
432 | node->lowered = true; | |
433 | } | |
434 | ||
435 | /* DECL has been parsed. Take it, queue it, compile it at the whim of the | |
436 | logic in effect. If NESTED is true, then our caller cannot stand to have | |
437 | the garbage collector run at the moment. We would need to either create | |
438 | a new GC context, or just not compile right now. */ | |
439 | ||
440 | void | |
441 | cgraph_finalize_function (tree decl, bool nested) | |
442 | { | |
443 | struct cgraph_node *node = cgraph_node (decl); | |
444 | ||
445 | if (node->local.finalized) | |
446 | cgraph_reset_node (node); | |
447 | ||
448 | notice_global_symbol (decl); | |
449 | node->decl = decl; | |
450 | node->local.finalized = true; | |
451 | node->lowered = DECL_STRUCT_FUNCTION (decl)->cfg != NULL; | |
452 | if (node->nested) | |
453 | lower_nested_functions (decl); | |
454 | gcc_assert (!node->nested); | |
455 | ||
456 | /* If not unit at a time, then we need to create the call graph | |
457 | now, so that called functions can be queued and emitted now. */ | |
458 | if (!flag_unit_at_a_time) | |
459 | { | |
460 | cgraph_analyze_function (node); | |
461 | cgraph_decide_inlining_incrementally (node, false); | |
462 | } | |
463 | ||
464 | if (decide_is_function_needed (node, decl)) | |
465 | cgraph_mark_needed_node (node); | |
466 | ||
467 | /* Since we reclaim unreachable nodes at the end of every language | |
468 | level unit, we need to be conservative about possible entry points | |
469 | there. */ | |
470 | if ((TREE_PUBLIC (decl) && !DECL_COMDAT (decl) && !DECL_EXTERNAL (decl))) | |
471 | cgraph_mark_reachable_node (node); | |
472 | ||
473 | /* If not unit at a time, go ahead and emit everything we've found | |
474 | to be reachable at this time. */ | |
475 | if (!nested) | |
476 | { | |
477 | if (!cgraph_assemble_pending_functions ()) | |
478 | ggc_collect (); | |
479 | } | |
480 | ||
481 | /* If we've not yet emitted decl, tell the debug info about it. */ | |
482 | if (!TREE_ASM_WRITTEN (decl)) | |
483 | (*debug_hooks->deferred_inline_function) (decl); | |
484 | ||
485 | /* Possibly warn about unused parameters. */ | |
486 | if (warn_unused_parameter) | |
487 | do_warn_unused_parameter (decl); | |
488 | } | |
489 | ||
490 | /* Walk tree and record all calls. Called via walk_tree. */ | |
491 | static tree | |
492 | record_reference (tree *tp, int *walk_subtrees, void *data) | |
493 | { | |
494 | tree t = *tp; | |
495 | ||
496 | switch (TREE_CODE (t)) | |
497 | { | |
498 | case VAR_DECL: | |
499 | /* ??? Really, we should mark this decl as *potentially* referenced | |
500 | by this function and re-examine whether the decl is actually used | |
501 | after rtl has been generated. */ | |
502 | if (TREE_STATIC (t) || DECL_EXTERNAL (t)) | |
503 | { | |
504 | cgraph_varpool_mark_needed_node (cgraph_varpool_node (t)); | |
505 | if (lang_hooks.callgraph.analyze_expr) | |
506 | return lang_hooks.callgraph.analyze_expr (tp, walk_subtrees, | |
507 | data); | |
508 | } | |
509 | break; | |
510 | ||
511 | case FDESC_EXPR: | |
512 | case ADDR_EXPR: | |
513 | if (flag_unit_at_a_time) | |
514 | { | |
515 | /* Record dereferences to the functions. This makes the | |
516 | functions reachable unconditionally. */ | |
517 | tree decl = TREE_OPERAND (*tp, 0); | |
518 | if (TREE_CODE (decl) == FUNCTION_DECL) | |
519 | cgraph_mark_needed_node (cgraph_node (decl)); | |
520 | } | |
521 | break; | |
522 | ||
523 | default: | |
524 | /* Save some cycles by not walking types and declaration as we | |
525 | won't find anything useful there anyway. */ | |
526 | if (IS_TYPE_OR_DECL_P (*tp)) | |
527 | { | |
528 | *walk_subtrees = 0; | |
529 | break; | |
530 | } | |
531 | ||
532 | if ((unsigned int) TREE_CODE (t) >= LAST_AND_UNUSED_TREE_CODE) | |
533 | return lang_hooks.callgraph.analyze_expr (tp, walk_subtrees, data); | |
534 | break; | |
535 | } | |
536 | ||
537 | return NULL; | |
538 | } | |
539 | ||
540 | /* Create cgraph edges for function calls inside BODY from NODE. */ | |
541 | ||
542 | static void | |
543 | cgraph_create_edges (struct cgraph_node *node, tree body) | |
544 | { | |
545 | basic_block bb; | |
546 | ||
547 | struct function *this_cfun = DECL_STRUCT_FUNCTION (body); | |
548 | block_stmt_iterator bsi; | |
549 | tree step; | |
550 | visited_nodes = pointer_set_create (); | |
551 | ||
552 | /* Reach the trees by walking over the CFG, and note the | |
553 | enclosing basic-blocks in the call edges. */ | |
554 | FOR_EACH_BB_FN (bb, this_cfun) | |
555 | for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi)) | |
556 | { | |
557 | tree stmt = bsi_stmt (bsi); | |
558 | tree call = get_call_expr_in (stmt); | |
559 | tree decl; | |
560 | ||
561 | if (call && (decl = get_callee_fndecl (call))) | |
562 | { | |
563 | cgraph_create_edge (node, cgraph_node (decl), stmt, | |
564 | bb->count, | |
565 | bb->loop_depth); | |
566 | walk_tree (&TREE_OPERAND (call, 1), | |
567 | record_reference, node, visited_nodes); | |
568 | if (TREE_CODE (stmt) == MODIFY_EXPR) | |
569 | walk_tree (&TREE_OPERAND (stmt, 0), | |
570 | record_reference, node, visited_nodes); | |
571 | } | |
572 | else | |
573 | walk_tree (bsi_stmt_ptr (bsi), record_reference, node, visited_nodes); | |
574 | } | |
575 | ||
576 | /* Look for initializers of constant variables and private statics. */ | |
577 | for (step = DECL_STRUCT_FUNCTION (body)->unexpanded_var_list; | |
578 | step; | |
579 | step = TREE_CHAIN (step)) | |
580 | { | |
581 | tree decl = TREE_VALUE (step); | |
582 | if (TREE_CODE (decl) == VAR_DECL | |
583 | && (TREE_STATIC (decl) && !DECL_EXTERNAL (decl)) | |
584 | && flag_unit_at_a_time) | |
585 | cgraph_varpool_finalize_decl (decl); | |
586 | else if (TREE_CODE (decl) == VAR_DECL && DECL_INITIAL (decl)) | |
587 | walk_tree (&DECL_INITIAL (decl), record_reference, node, visited_nodes); | |
588 | } | |
589 | ||
590 | pointer_set_destroy (visited_nodes); | |
591 | visited_nodes = NULL; | |
592 | } | |
593 | ||
594 | /* Give initial reasons why inlining would fail. Those gets | |
595 | either NULLified or usually overwritten by more precise reason | |
596 | later. */ | |
597 | static void | |
598 | initialize_inline_failed (struct cgraph_node *node) | |
599 | { | |
600 | struct cgraph_edge *e; | |
601 | ||
602 | for (e = node->callers; e; e = e->next_caller) | |
603 | { | |
604 | gcc_assert (!e->callee->global.inlined_to); | |
605 | gcc_assert (e->inline_failed); | |
606 | if (node->local.redefined_extern_inline) | |
607 | e->inline_failed = N_("redefined extern inline functions are not " | |
608 | "considered for inlining"); | |
609 | else if (!node->local.inlinable) | |
610 | e->inline_failed = N_("function not inlinable"); | |
611 | else | |
612 | e->inline_failed = N_("function not considered for inlining"); | |
613 | } | |
614 | } | |
615 | ||
616 | /* Rebuild call edges from current function after a passes not aware | |
617 | of cgraph updating. */ | |
618 | static void | |
619 | rebuild_cgraph_edges (void) | |
620 | { | |
621 | basic_block bb; | |
622 | struct cgraph_node *node = cgraph_node (current_function_decl); | |
623 | block_stmt_iterator bsi; | |
624 | ||
625 | cgraph_node_remove_callees (node); | |
626 | ||
627 | node->count = ENTRY_BLOCK_PTR->count; | |
628 | ||
629 | FOR_EACH_BB (bb) | |
630 | for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi)) | |
631 | { | |
632 | tree stmt = bsi_stmt (bsi); | |
633 | tree call = get_call_expr_in (stmt); | |
634 | tree decl; | |
635 | ||
636 | if (call && (decl = get_callee_fndecl (call))) | |
637 | cgraph_create_edge (node, cgraph_node (decl), stmt, | |
638 | bb->count, | |
639 | bb->loop_depth); | |
640 | } | |
641 | initialize_inline_failed (node); | |
642 | gcc_assert (!node->global.inlined_to); | |
643 | } | |
644 | ||
645 | struct tree_opt_pass pass_rebuild_cgraph_edges = | |
646 | { | |
647 | NULL, /* name */ | |
648 | NULL, /* gate */ | |
649 | rebuild_cgraph_edges, /* execute */ | |
650 | NULL, /* sub */ | |
651 | NULL, /* next */ | |
652 | 0, /* static_pass_number */ | |
653 | 0, /* tv_id */ | |
654 | PROP_cfg, /* properties_required */ | |
655 | 0, /* properties_provided */ | |
656 | 0, /* properties_destroyed */ | |
657 | 0, /* todo_flags_start */ | |
658 | 0, /* todo_flags_finish */ | |
659 | 0 /* letter */ | |
660 | }; | |
661 | ||
662 | /* Verify cgraph nodes of given cgraph node. */ | |
663 | void | |
664 | verify_cgraph_node (struct cgraph_node *node) | |
665 | { | |
666 | struct cgraph_edge *e; | |
667 | struct cgraph_node *main_clone; | |
668 | struct function *this_cfun = DECL_STRUCT_FUNCTION (node->decl); | |
669 | basic_block this_block; | |
670 | block_stmt_iterator bsi; | |
671 | bool error_found = false; | |
672 | ||
673 | timevar_push (TV_CGRAPH_VERIFY); | |
674 | for (e = node->callees; e; e = e->next_callee) | |
675 | if (e->aux) | |
676 | { | |
677 | error ("aux field set for edge %s->%s", | |
678 | cgraph_node_name (e->caller), cgraph_node_name (e->callee)); | |
679 | error_found = true; | |
680 | } | |
681 | if (node->count < 0) | |
682 | { | |
683 | error ("Execution count is negative"); | |
684 | error_found = true; | |
685 | } | |
686 | for (e = node->callers; e; e = e->next_caller) | |
687 | { | |
688 | if (e->count < 0) | |
689 | { | |
690 | error ("caller edge count is negative"); | |
691 | error_found = true; | |
692 | } | |
693 | if (!e->inline_failed) | |
694 | { | |
695 | if (node->global.inlined_to | |
696 | != (e->caller->global.inlined_to | |
697 | ? e->caller->global.inlined_to : e->caller)) | |
698 | { | |
699 | error ("inlined_to pointer is wrong"); | |
700 | error_found = true; | |
701 | } | |
702 | if (node->callers->next_caller) | |
703 | { | |
704 | error ("multiple inline callers"); | |
705 | error_found = true; | |
706 | } | |
707 | } | |
708 | else | |
709 | if (node->global.inlined_to) | |
710 | { | |
711 | error ("inlined_to pointer set for noninline callers"); | |
712 | error_found = true; | |
713 | } | |
714 | } | |
715 | if (!node->callers && node->global.inlined_to) | |
716 | { | |
717 | error ("inlined_to pointer is set but no predecesors found"); | |
718 | error_found = true; | |
719 | } | |
720 | if (node->global.inlined_to == node) | |
721 | { | |
722 | error ("inlined_to pointer refers to itself"); | |
723 | error_found = true; | |
724 | } | |
725 | ||
726 | for (main_clone = cgraph_node (node->decl); main_clone; | |
727 | main_clone = main_clone->next_clone) | |
728 | if (main_clone == node) | |
729 | break; | |
730 | if (!node) | |
731 | { | |
732 | error ("node not found in DECL_ASSEMBLER_NAME hash"); | |
733 | error_found = true; | |
734 | } | |
735 | ||
736 | if (node->analyzed | |
737 | && DECL_SAVED_TREE (node->decl) && !TREE_ASM_WRITTEN (node->decl) | |
738 | && (!DECL_EXTERNAL (node->decl) || node->global.inlined_to)) | |
739 | { | |
740 | if (this_cfun->cfg) | |
741 | { | |
742 | /* The nodes we're interested in are never shared, so walk | |
743 | the tree ignoring duplicates. */ | |
744 | visited_nodes = pointer_set_create (); | |
745 | /* Reach the trees by walking over the CFG, and note the | |
746 | enclosing basic-blocks in the call edges. */ | |
747 | FOR_EACH_BB_FN (this_block, this_cfun) | |
748 | for (bsi = bsi_start (this_block); !bsi_end_p (bsi); bsi_next (&bsi)) | |
749 | { | |
750 | tree stmt = bsi_stmt (bsi); | |
751 | tree call = get_call_expr_in (stmt); | |
752 | tree decl; | |
753 | if (call && (decl = get_callee_fndecl (call))) | |
754 | { | |
755 | struct cgraph_edge *e = cgraph_edge (node, stmt); | |
756 | if (e) | |
757 | { | |
758 | if (e->aux) | |
759 | { | |
760 | error ("shared call_stmt:"); | |
761 | debug_generic_stmt (stmt); | |
762 | error_found = true; | |
763 | } | |
764 | if (e->callee->decl != cgraph_node (decl)->decl | |
765 | && e->inline_failed) | |
766 | { | |
767 | error ("edge points to wrong declaration:"); | |
768 | debug_tree (e->callee->decl); | |
769 | fprintf (stderr," Instead of:"); | |
770 | debug_tree (decl); | |
771 | } | |
772 | e->aux = (void *)1; | |
773 | } | |
774 | else | |
775 | { | |
776 | error ("missing callgraph edge for call stmt:"); | |
777 | debug_generic_stmt (stmt); | |
778 | error_found = true; | |
779 | } | |
780 | } | |
781 | } | |
782 | pointer_set_destroy (visited_nodes); | |
783 | visited_nodes = NULL; | |
784 | } | |
785 | else | |
786 | /* No CFG available?! */ | |
787 | gcc_unreachable (); | |
788 | ||
789 | for (e = node->callees; e; e = e->next_callee) | |
790 | { | |
791 | if (!e->aux) | |
792 | { | |
793 | error ("edge %s->%s has no corresponding call_stmt", | |
794 | cgraph_node_name (e->caller), | |
795 | cgraph_node_name (e->callee)); | |
796 | debug_generic_stmt (e->call_stmt); | |
797 | error_found = true; | |
798 | } | |
799 | e->aux = 0; | |
800 | } | |
801 | } | |
802 | if (error_found) | |
803 | { | |
804 | dump_cgraph_node (stderr, node); | |
805 | internal_error ("verify_cgraph_node failed"); | |
806 | } | |
807 | timevar_pop (TV_CGRAPH_VERIFY); | |
808 | } | |
809 | ||
810 | /* Verify whole cgraph structure. */ | |
811 | void | |
812 | verify_cgraph (void) | |
813 | { | |
814 | struct cgraph_node *node; | |
815 | ||
816 | if (sorrycount || errorcount) | |
817 | return; | |
818 | ||
819 | for (node = cgraph_nodes; node; node = node->next) | |
820 | verify_cgraph_node (node); | |
821 | } | |
822 | ||
823 | /* Output one variable, if necessary. Return whether we output it. */ | |
824 | static bool | |
825 | cgraph_varpool_assemble_decl (struct cgraph_varpool_node *node) | |
826 | { | |
827 | tree decl = node->decl; | |
828 | ||
829 | if (!TREE_ASM_WRITTEN (decl) | |
830 | && !node->alias | |
831 | && !DECL_EXTERNAL (decl) | |
832 | && (TREE_CODE (decl) != VAR_DECL || !DECL_HAS_VALUE_EXPR_P (decl))) | |
833 | { | |
834 | assemble_variable (decl, 0, 1, 0); | |
835 | /* Local static variables are never seen by check_global_declarations | |
836 | so we need to output debug info by hand. */ | |
837 | if (DECL_CONTEXT (decl) | |
838 | && (TREE_CODE (DECL_CONTEXT (decl)) == BLOCK | |
839 | || TREE_CODE (DECL_CONTEXT (decl)) == FUNCTION_DECL) | |
840 | && errorcount == 0 && sorrycount == 0) | |
841 | { | |
842 | timevar_push (TV_SYMOUT); | |
843 | (*debug_hooks->global_decl) (decl); | |
844 | timevar_pop (TV_SYMOUT); | |
845 | } | |
846 | return true; | |
847 | } | |
848 | ||
849 | return false; | |
850 | } | |
851 | ||
852 | /* Output all variables enqueued to be assembled. */ | |
853 | bool | |
854 | cgraph_varpool_assemble_pending_decls (void) | |
855 | { | |
856 | bool changed = false; | |
857 | ||
858 | if (errorcount || sorrycount) | |
859 | return false; | |
860 | ||
861 | /* EH might mark decls as needed during expansion. This should be safe since | |
862 | we don't create references to new function, but it should not be used | |
863 | elsewhere. */ | |
864 | cgraph_varpool_analyze_pending_decls (); | |
865 | ||
866 | while (cgraph_varpool_nodes_queue) | |
867 | { | |
868 | struct cgraph_varpool_node *node = cgraph_varpool_nodes_queue; | |
869 | ||
870 | cgraph_varpool_nodes_queue = cgraph_varpool_nodes_queue->next_needed; | |
871 | if (cgraph_varpool_assemble_decl (node)) | |
872 | changed = true; | |
873 | node->next_needed = NULL; | |
874 | } | |
875 | return changed; | |
876 | } | |
877 | ||
878 | /* Output all asm statements we have stored up to be output. */ | |
879 | ||
880 | static void | |
881 | cgraph_output_pending_asms (void) | |
882 | { | |
883 | struct cgraph_asm_node *can; | |
884 | ||
885 | if (errorcount || sorrycount) | |
886 | return; | |
887 | ||
888 | for (can = cgraph_asm_nodes; can; can = can->next) | |
889 | assemble_asm (can->asm_str); | |
890 | cgraph_asm_nodes = NULL; | |
891 | } | |
892 | ||
893 | /* Analyze the function scheduled to be output. */ | |
894 | void | |
895 | cgraph_analyze_function (struct cgraph_node *node) | |
896 | { | |
897 | tree decl = node->decl; | |
898 | ||
899 | current_function_decl = decl; | |
900 | push_cfun (DECL_STRUCT_FUNCTION (decl)); | |
901 | cgraph_lower_function (node); | |
902 | ||
903 | /* First kill forward declaration so reverse inlining works properly. */ | |
904 | cgraph_create_edges (node, decl); | |
905 | ||
906 | node->local.inlinable = tree_inlinable_function_p (decl); | |
907 | node->local.self_insns = estimate_num_insns (decl); | |
908 | if (node->local.inlinable) | |
909 | node->local.disregard_inline_limits | |
910 | = lang_hooks.tree_inlining.disregard_inline_limits (decl); | |
911 | initialize_inline_failed (node); | |
912 | if (flag_really_no_inline && !node->local.disregard_inline_limits) | |
913 | node->local.inlinable = 0; | |
914 | /* Inlining characteristics are maintained by the cgraph_mark_inline. */ | |
915 | node->global.insns = node->local.self_insns; | |
916 | ||
917 | node->analyzed = true; | |
918 | pop_cfun (); | |
919 | current_function_decl = NULL; | |
920 | } | |
921 | ||
922 | /* Analyze the whole compilation unit once it is parsed completely. */ | |
923 | ||
924 | void | |
925 | cgraph_finalize_compilation_unit (void) | |
926 | { | |
927 | struct cgraph_node *node; | |
928 | /* Keep track of already processed nodes when called multiple times for | |
929 | intermodule optimization. */ | |
930 | static struct cgraph_node *first_analyzed; | |
931 | ||
932 | finish_aliases_1 (); | |
933 | ||
934 | if (!flag_unit_at_a_time) | |
935 | { | |
936 | cgraph_output_pending_asms (); | |
937 | cgraph_assemble_pending_functions (); | |
938 | return; | |
939 | } | |
940 | ||
941 | if (!quiet_flag) | |
942 | { | |
943 | fprintf (stderr, "\nAnalyzing compilation unit"); | |
944 | fflush (stderr); | |
945 | } | |
946 | ||
947 | timevar_push (TV_CGRAPH); | |
948 | cgraph_varpool_analyze_pending_decls (); | |
949 | if (cgraph_dump_file) | |
950 | { | |
951 | fprintf (cgraph_dump_file, "Initial entry points:"); | |
952 | for (node = cgraph_nodes; node != first_analyzed; node = node->next) | |
953 | if (node->needed && DECL_SAVED_TREE (node->decl)) | |
954 | fprintf (cgraph_dump_file, " %s", cgraph_node_name (node)); | |
955 | fprintf (cgraph_dump_file, "\n"); | |
956 | } | |
957 | ||
958 | /* Propagate reachability flag and lower representation of all reachable | |
959 | functions. In the future, lowering will introduce new functions and | |
960 | new entry points on the way (by template instantiation and virtual | |
961 | method table generation for instance). */ | |
962 | while (cgraph_nodes_queue) | |
963 | { | |
964 | struct cgraph_edge *edge; | |
965 | tree decl = cgraph_nodes_queue->decl; | |
966 | ||
967 | node = cgraph_nodes_queue; | |
968 | cgraph_nodes_queue = cgraph_nodes_queue->next_needed; | |
969 | node->next_needed = NULL; | |
970 | ||
971 | /* ??? It is possible to create extern inline function and later using | |
972 | weak alias attribute to kill its body. See | |
973 | gcc.c-torture/compile/20011119-1.c */ | |
974 | if (!DECL_SAVED_TREE (decl)) | |
975 | { | |
976 | cgraph_reset_node (node); | |
977 | continue; | |
978 | } | |
979 | ||
980 | gcc_assert (!node->analyzed && node->reachable); | |
981 | gcc_assert (DECL_SAVED_TREE (decl)); | |
982 | ||
983 | cgraph_analyze_function (node); | |
984 | ||
985 | for (edge = node->callees; edge; edge = edge->next_callee) | |
986 | if (!edge->callee->reachable) | |
987 | cgraph_mark_reachable_node (edge->callee); | |
988 | ||
989 | cgraph_varpool_analyze_pending_decls (); | |
990 | } | |
991 | ||
992 | /* Collect entry points to the unit. */ | |
993 | ||
994 | if (cgraph_dump_file) | |
995 | { | |
996 | fprintf (cgraph_dump_file, "Unit entry points:"); | |
997 | for (node = cgraph_nodes; node != first_analyzed; node = node->next) | |
998 | if (node->needed && DECL_SAVED_TREE (node->decl)) | |
999 | fprintf (cgraph_dump_file, " %s", cgraph_node_name (node)); | |
1000 | fprintf (cgraph_dump_file, "\n\nInitial "); | |
1001 | dump_cgraph (cgraph_dump_file); | |
1002 | } | |
1003 | ||
1004 | if (cgraph_dump_file) | |
1005 | fprintf (cgraph_dump_file, "\nReclaiming functions:"); | |
1006 | ||
1007 | for (node = cgraph_nodes; node != first_analyzed; node = node->next) | |
1008 | { | |
1009 | tree decl = node->decl; | |
1010 | ||
1011 | if (node->local.finalized && !DECL_SAVED_TREE (decl)) | |
1012 | cgraph_reset_node (node); | |
1013 | ||
1014 | if (!node->reachable && DECL_SAVED_TREE (decl)) | |
1015 | { | |
1016 | if (cgraph_dump_file) | |
1017 | fprintf (cgraph_dump_file, " %s", cgraph_node_name (node)); | |
1018 | cgraph_remove_node (node); | |
1019 | continue; | |
1020 | } | |
1021 | else | |
1022 | node->next_needed = NULL; | |
1023 | gcc_assert (!node->local.finalized || DECL_SAVED_TREE (decl)); | |
1024 | gcc_assert (node->analyzed == node->local.finalized); | |
1025 | } | |
1026 | if (cgraph_dump_file) | |
1027 | { | |
1028 | fprintf (cgraph_dump_file, "\n\nReclaimed "); | |
1029 | dump_cgraph (cgraph_dump_file); | |
1030 | } | |
1031 | first_analyzed = cgraph_nodes; | |
1032 | ggc_collect (); | |
1033 | timevar_pop (TV_CGRAPH); | |
1034 | } | |
1035 | /* Figure out what functions we want to assemble. */ | |
1036 | ||
1037 | static void | |
1038 | cgraph_mark_functions_to_output (void) | |
1039 | { | |
1040 | struct cgraph_node *node; | |
1041 | ||
1042 | for (node = cgraph_nodes; node; node = node->next) | |
1043 | { | |
1044 | tree decl = node->decl; | |
1045 | struct cgraph_edge *e; | |
1046 | ||
1047 | gcc_assert (!node->output); | |
1048 | ||
1049 | for (e = node->callers; e; e = e->next_caller) | |
1050 | if (e->inline_failed) | |
1051 | break; | |
1052 | ||
1053 | /* We need to output all local functions that are used and not | |
1054 | always inlined, as well as those that are reachable from | |
1055 | outside the current compilation unit. */ | |
1056 | if (DECL_SAVED_TREE (decl) | |
1057 | && !node->global.inlined_to | |
1058 | && (node->needed | |
1059 | || (e && node->reachable)) | |
1060 | && !TREE_ASM_WRITTEN (decl) | |
1061 | && !DECL_EXTERNAL (decl)) | |
1062 | node->output = 1; | |
1063 | else | |
1064 | { | |
1065 | /* We should've reclaimed all functions that are not needed. */ | |
1066 | #ifdef ENABLE_CHECKING | |
1067 | if (!node->global.inlined_to && DECL_SAVED_TREE (decl) | |
1068 | && !DECL_EXTERNAL (decl)) | |
1069 | { | |
1070 | dump_cgraph_node (stderr, node); | |
1071 | internal_error ("failed to reclaim unneeded function"); | |
1072 | } | |
1073 | #endif | |
1074 | gcc_assert (node->global.inlined_to || !DECL_SAVED_TREE (decl) | |
1075 | || DECL_EXTERNAL (decl)); | |
1076 | ||
1077 | } | |
1078 | ||
1079 | } | |
1080 | } | |
1081 | ||
1082 | /* Expand function specified by NODE. */ | |
1083 | ||
1084 | static void | |
1085 | cgraph_expand_function (struct cgraph_node *node) | |
1086 | { | |
1087 | tree decl = node->decl; | |
1088 | ||
1089 | /* We ought to not compile any inline clones. */ | |
1090 | gcc_assert (!node->global.inlined_to); | |
1091 | ||
1092 | if (flag_unit_at_a_time) | |
1093 | announce_function (decl); | |
1094 | ||
1095 | cgraph_lower_function (node); | |
1096 | ||
1097 | /* Generate RTL for the body of DECL. */ | |
1098 | lang_hooks.callgraph.expand_function (decl); | |
1099 | ||
1100 | /* Make sure that BE didn't give up on compiling. */ | |
1101 | /* ??? Can happen with nested function of extern inline. */ | |
1102 | gcc_assert (TREE_ASM_WRITTEN (node->decl)); | |
1103 | ||
1104 | current_function_decl = NULL; | |
1105 | if (!cgraph_preserve_function_body_p (node->decl)) | |
1106 | { | |
1107 | DECL_SAVED_TREE (node->decl) = NULL; | |
1108 | DECL_STRUCT_FUNCTION (node->decl) = NULL; | |
1109 | DECL_INITIAL (node->decl) = error_mark_node; | |
1110 | /* Eliminate all call edges. This is important so the call_expr no longer | |
1111 | points to the dead function body. */ | |
1112 | cgraph_node_remove_callees (node); | |
1113 | } | |
1114 | ||
1115 | cgraph_function_flags_ready = true; | |
1116 | } | |
1117 | ||
1118 | /* Return true when CALLER_DECL should be inlined into CALLEE_DECL. */ | |
1119 | ||
1120 | bool | |
1121 | cgraph_inline_p (struct cgraph_edge *e, const char **reason) | |
1122 | { | |
1123 | *reason = e->inline_failed; | |
1124 | return !e->inline_failed; | |
1125 | } | |
1126 | ||
1127 | ||
1128 | ||
1129 | /* Expand all functions that must be output. | |
1130 | ||
1131 | Attempt to topologically sort the nodes so function is output when | |
1132 | all called functions are already assembled to allow data to be | |
1133 | propagated across the callgraph. Use a stack to get smaller distance | |
1134 | between a function and its callees (later we may choose to use a more | |
1135 | sophisticated algorithm for function reordering; we will likely want | |
1136 | to use subsections to make the output functions appear in top-down | |
1137 | order). */ | |
1138 | ||
1139 | static void | |
1140 | cgraph_expand_all_functions (void) | |
1141 | { | |
1142 | struct cgraph_node *node; | |
1143 | struct cgraph_node **order = | |
1144 | xcalloc (cgraph_n_nodes, sizeof (struct cgraph_node *)); | |
1145 | int order_pos = 0, new_order_pos = 0; | |
1146 | int i; | |
1147 | ||
1148 | order_pos = cgraph_postorder (order); | |
1149 | gcc_assert (order_pos == cgraph_n_nodes); | |
1150 | ||
1151 | /* Garbage collector may remove inline clones we eliminate during | |
1152 | optimization. So we must be sure to not reference them. */ | |
1153 | for (i = 0; i < order_pos; i++) | |
1154 | if (order[i]->output) | |
1155 | order[new_order_pos++] = order[i]; | |
1156 | ||
1157 | for (i = new_order_pos - 1; i >= 0; i--) | |
1158 | { | |
1159 | node = order[i]; | |
1160 | if (node->output) | |
1161 | { | |
1162 | gcc_assert (node->reachable); | |
1163 | node->output = 0; | |
1164 | cgraph_expand_function (node); | |
1165 | } | |
1166 | } | |
1167 | ||
1168 | free (order); | |
1169 | ||
1170 | /* Process CGRAPH_EXPAND_QUEUE, these are functions created during | |
1171 | the expansion process. Note that this queue may grow as its | |
1172 | being processed, as the new functions may generate new ones. */ | |
1173 | while (cgraph_expand_queue) | |
1174 | { | |
1175 | node = cgraph_expand_queue; | |
1176 | cgraph_expand_queue = cgraph_expand_queue->next_needed; | |
1177 | node->next_needed = NULL; | |
1178 | node->output = 0; | |
1179 | node->lowered = DECL_STRUCT_FUNCTION (node->decl)->cfg != NULL; | |
1180 | cgraph_expand_function (node); | |
1181 | } | |
1182 | } | |
1183 | ||
1184 | /* This is used to sort the node types by the cgraph order number. */ | |
1185 | ||
1186 | struct cgraph_order_sort | |
1187 | { | |
1188 | enum { ORDER_UNDEFINED = 0, ORDER_FUNCTION, ORDER_VAR, ORDER_ASM } kind; | |
1189 | union | |
1190 | { | |
1191 | struct cgraph_node *f; | |
1192 | struct cgraph_varpool_node *v; | |
1193 | struct cgraph_asm_node *a; | |
1194 | } u; | |
1195 | }; | |
1196 | ||
1197 | /* Output all functions, variables, and asm statements in the order | |
1198 | according to their order fields, which is the order in which they | |
1199 | appeared in the file. This implements -fno-toplevel-reorder. In | |
1200 | this mode we may output functions and variables which don't really | |
1201 | need to be output. */ | |
1202 | ||
1203 | static void | |
1204 | cgraph_output_in_order (void) | |
1205 | { | |
1206 | int max; | |
1207 | size_t size; | |
1208 | struct cgraph_order_sort *nodes; | |
1209 | int i; | |
1210 | struct cgraph_node *pf; | |
1211 | struct cgraph_varpool_node *pv; | |
1212 | struct cgraph_asm_node *pa; | |
1213 | ||
1214 | max = cgraph_order; | |
1215 | size = max * sizeof (struct cgraph_order_sort); | |
1216 | nodes = (struct cgraph_order_sort *) alloca (size); | |
1217 | memset (nodes, 0, size); | |
1218 | ||
1219 | cgraph_varpool_analyze_pending_decls (); | |
1220 | ||
1221 | for (pf = cgraph_nodes; pf; pf = pf->next) | |
1222 | { | |
1223 | if (pf->output) | |
1224 | { | |
1225 | i = pf->order; | |
1226 | gcc_assert (nodes[i].kind == ORDER_UNDEFINED); | |
1227 | nodes[i].kind = ORDER_FUNCTION; | |
1228 | nodes[i].u.f = pf; | |
1229 | } | |
1230 | } | |
1231 | ||
1232 | for (pv = cgraph_varpool_nodes_queue; pv; pv = pv->next_needed) | |
1233 | { | |
1234 | i = pv->order; | |
1235 | gcc_assert (nodes[i].kind == ORDER_UNDEFINED); | |
1236 | nodes[i].kind = ORDER_VAR; | |
1237 | nodes[i].u.v = pv; | |
1238 | } | |
1239 | ||
1240 | for (pa = cgraph_asm_nodes; pa; pa = pa->next) | |
1241 | { | |
1242 | i = pa->order; | |
1243 | gcc_assert (nodes[i].kind == ORDER_UNDEFINED); | |
1244 | nodes[i].kind = ORDER_ASM; | |
1245 | nodes[i].u.a = pa; | |
1246 | } | |
1247 | cgraph_asm_nodes = NULL; | |
1248 | ||
1249 | for (i = 0; i < max; ++i) | |
1250 | { | |
1251 | switch (nodes[i].kind) | |
1252 | { | |
1253 | case ORDER_FUNCTION: | |
1254 | nodes[i].u.f->output = 0; | |
1255 | cgraph_expand_function (nodes[i].u.f); | |
1256 | break; | |
1257 | ||
1258 | case ORDER_VAR: | |
1259 | cgraph_varpool_assemble_decl (nodes[i].u.v); | |
1260 | break; | |
1261 | ||
1262 | case ORDER_ASM: | |
1263 | assemble_asm (nodes[i].u.a->asm_str); | |
1264 | break; | |
1265 | ||
1266 | case ORDER_UNDEFINED: | |
1267 | break; | |
1268 | ||
1269 | default: | |
1270 | gcc_unreachable (); | |
1271 | } | |
1272 | } | |
1273 | } | |
1274 | ||
1275 | /* Mark visibility of all functions. | |
1276 | ||
1277 | A local function is one whose calls can occur only in the current | |
1278 | compilation unit and all its calls are explicit, so we can change | |
1279 | its calling convention. We simply mark all static functions whose | |
1280 | address is not taken as local. | |
1281 | ||
1282 | We also change the TREE_PUBLIC flag of all declarations that are public | |
1283 | in language point of view but we want to overwrite this default | |
1284 | via visibilities for the backend point of view. */ | |
1285 | ||
1286 | static void | |
1287 | cgraph_function_and_variable_visibility (void) | |
1288 | { | |
1289 | struct cgraph_node *node; | |
1290 | struct cgraph_varpool_node *vnode; | |
1291 | ||
1292 | for (node = cgraph_nodes; node; node = node->next) | |
1293 | { | |
1294 | if (node->reachable | |
1295 | && (DECL_COMDAT (node->decl) | |
1296 | || (!flag_whole_program | |
1297 | && TREE_PUBLIC (node->decl) && !DECL_EXTERNAL (node->decl)))) | |
1298 | node->local.externally_visible = true; | |
1299 | if (!node->local.externally_visible && node->analyzed | |
1300 | && !DECL_EXTERNAL (node->decl)) | |
1301 | { | |
1302 | gcc_assert (flag_whole_program || !TREE_PUBLIC (node->decl)); | |
1303 | TREE_PUBLIC (node->decl) = 0; | |
1304 | } | |
1305 | node->local.local = (!node->needed | |
1306 | && node->analyzed | |
1307 | && !DECL_EXTERNAL (node->decl) | |
1308 | && !node->local.externally_visible); | |
1309 | } | |
1310 | for (vnode = cgraph_varpool_nodes_queue; vnode; vnode = vnode->next_needed) | |
1311 | { | |
1312 | if (vnode->needed | |
1313 | && !flag_whole_program | |
1314 | && (DECL_COMDAT (vnode->decl) || TREE_PUBLIC (vnode->decl))) | |
1315 | vnode->externally_visible = 1; | |
1316 | if (!vnode->externally_visible) | |
1317 | { | |
1318 | gcc_assert (flag_whole_program || !TREE_PUBLIC (vnode->decl)); | |
1319 | TREE_PUBLIC (vnode->decl) = 0; | |
1320 | } | |
1321 | gcc_assert (TREE_STATIC (vnode->decl)); | |
1322 | } | |
1323 | ||
1324 | /* Because we have to be conservative on the boundaries of source | |
1325 | level units, it is possible that we marked some functions in | |
1326 | reachable just because they might be used later via external | |
1327 | linkage, but after making them local they are really unreachable | |
1328 | now. */ | |
1329 | cgraph_remove_unreachable_nodes (true, cgraph_dump_file); | |
1330 | ||
1331 | if (cgraph_dump_file) | |
1332 | { | |
1333 | fprintf (cgraph_dump_file, "\nMarking local functions:"); | |
1334 | for (node = cgraph_nodes; node; node = node->next) | |
1335 | if (node->local.local) | |
1336 | fprintf (cgraph_dump_file, " %s", cgraph_node_name (node)); | |
1337 | fprintf (cgraph_dump_file, "\n\n"); | |
1338 | fprintf (cgraph_dump_file, "\nMarking externally visible functions:"); | |
1339 | for (node = cgraph_nodes; node; node = node->next) | |
1340 | if (node->local.externally_visible) | |
1341 | fprintf (cgraph_dump_file, " %s", cgraph_node_name (node)); | |
1342 | fprintf (cgraph_dump_file, "\n\n"); | |
1343 | } | |
1344 | cgraph_function_flags_ready = true; | |
1345 | } | |
1346 | ||
1347 | /* Return true when function body of DECL still needs to be kept around | |
1348 | for later re-use. */ | |
1349 | bool | |
1350 | cgraph_preserve_function_body_p (tree decl) | |
1351 | { | |
1352 | struct cgraph_node *node; | |
1353 | if (!cgraph_global_info_ready) | |
1354 | return (DECL_INLINE (decl) && !flag_really_no_inline); | |
1355 | /* Look if there is any clone around. */ | |
1356 | for (node = cgraph_node (decl); node; node = node->next_clone) | |
1357 | if (node->global.inlined_to) | |
1358 | return true; | |
1359 | return false; | |
1360 | } | |
1361 | ||
1362 | static void | |
1363 | ipa_passes (void) | |
1364 | { | |
1365 | cfun = NULL; | |
1366 | tree_register_cfg_hooks (); | |
1367 | bitmap_obstack_initialize (NULL); | |
1368 | execute_ipa_pass_list (all_ipa_passes); | |
1369 | bitmap_obstack_release (NULL); | |
1370 | } | |
1371 | ||
1372 | /* Perform simple optimizations based on callgraph. */ | |
1373 | ||
1374 | void | |
1375 | cgraph_optimize (void) | |
1376 | { | |
1377 | #ifdef ENABLE_CHECKING | |
1378 | verify_cgraph (); | |
1379 | #endif | |
1380 | if (!flag_unit_at_a_time) | |
1381 | { | |
1382 | cgraph_output_pending_asms (); | |
1383 | cgraph_varpool_assemble_pending_decls (); | |
1384 | return; | |
1385 | } | |
1386 | ||
1387 | process_pending_assemble_externals (); | |
1388 | ||
1389 | /* Frontend may output common variables after the unit has been finalized. | |
1390 | It is safe to deal with them here as they are always zero initialized. */ | |
1391 | cgraph_varpool_analyze_pending_decls (); | |
1392 | ||
1393 | timevar_push (TV_CGRAPHOPT); | |
1394 | if (!quiet_flag) | |
1395 | fprintf (stderr, "Performing intraprocedural optimizations\n"); | |
1396 | ||
1397 | cgraph_function_and_variable_visibility (); | |
1398 | if (cgraph_dump_file) | |
1399 | { | |
1400 | fprintf (cgraph_dump_file, "Marked "); | |
1401 | dump_cgraph (cgraph_dump_file); | |
1402 | } | |
1403 | ipa_passes (); | |
1404 | /* This pass remove bodies of extern inline functions we never inlined. | |
1405 | Do this later so other IPA passes see what is really going on. */ | |
1406 | cgraph_remove_unreachable_nodes (false, dump_file); | |
1407 | cgraph_global_info_ready = true; | |
1408 | if (cgraph_dump_file) | |
1409 | { | |
1410 | fprintf (cgraph_dump_file, "Optimized "); | |
1411 | dump_cgraph (cgraph_dump_file); | |
1412 | dump_varpool (cgraph_dump_file); | |
1413 | } | |
1414 | timevar_pop (TV_CGRAPHOPT); | |
1415 | ||
1416 | /* Output everything. */ | |
1417 | if (!quiet_flag) | |
1418 | fprintf (stderr, "Assembling functions:\n"); | |
1419 | #ifdef ENABLE_CHECKING | |
1420 | verify_cgraph (); | |
1421 | #endif | |
1422 | ||
1423 | cgraph_mark_functions_to_output (); | |
1424 | ||
1425 | if (!flag_toplevel_reorder) | |
1426 | cgraph_output_in_order (); | |
1427 | else | |
1428 | { | |
1429 | cgraph_output_pending_asms (); | |
1430 | ||
1431 | cgraph_expand_all_functions (); | |
1432 | cgraph_varpool_remove_unreferenced_decls (); | |
1433 | ||
1434 | cgraph_varpool_assemble_pending_decls (); | |
1435 | } | |
1436 | ||
1437 | if (cgraph_dump_file) | |
1438 | { | |
1439 | fprintf (cgraph_dump_file, "\nFinal "); | |
1440 | dump_cgraph (cgraph_dump_file); | |
1441 | } | |
1442 | #ifdef ENABLE_CHECKING | |
1443 | verify_cgraph (); | |
1444 | /* Double check that all inline clones are gone and that all | |
1445 | function bodies have been released from memory. */ | |
1446 | if (flag_unit_at_a_time | |
1447 | && !dump_enabled_p (TDI_tree_all) | |
1448 | && !(sorrycount || errorcount)) | |
1449 | { | |
1450 | struct cgraph_node *node; | |
1451 | bool error_found = false; | |
1452 | ||
1453 | for (node = cgraph_nodes; node; node = node->next) | |
1454 | if (node->analyzed | |
1455 | && (node->global.inlined_to | |
1456 | || DECL_SAVED_TREE (node->decl))) | |
1457 | { | |
1458 | error_found = true; | |
1459 | dump_cgraph_node (stderr, node); | |
1460 | } | |
1461 | if (error_found) | |
1462 | internal_error ("nodes with no released memory found"); | |
1463 | } | |
1464 | #endif | |
1465 | } | |
1466 | ||
1467 | /* Generate and emit a static constructor or destructor. WHICH must be | |
1468 | one of 'I' or 'D'. BODY should be a STATEMENT_LIST containing | |
1469 | GENERIC statements. */ | |
1470 | ||
1471 | void | |
1472 | cgraph_build_static_cdtor (char which, tree body, int priority) | |
1473 | { | |
1474 | static int counter = 0; | |
1475 | char which_buf[16]; | |
1476 | tree decl, name, resdecl; | |
1477 | ||
1478 | sprintf (which_buf, "%c_%d", which, counter++); | |
1479 | name = get_file_function_name_long (which_buf); | |
1480 | ||
1481 | decl = build_decl (FUNCTION_DECL, name, | |
1482 | build_function_type (void_type_node, void_list_node)); | |
1483 | current_function_decl = decl; | |
1484 | ||
1485 | resdecl = build_decl (RESULT_DECL, NULL_TREE, void_type_node); | |
1486 | DECL_ARTIFICIAL (resdecl) = 1; | |
1487 | DECL_IGNORED_P (resdecl) = 1; | |
1488 | DECL_RESULT (decl) = resdecl; | |
1489 | ||
1490 | allocate_struct_function (decl); | |
1491 | ||
1492 | TREE_STATIC (decl) = 1; | |
1493 | TREE_USED (decl) = 1; | |
1494 | DECL_ARTIFICIAL (decl) = 1; | |
1495 | DECL_IGNORED_P (decl) = 1; | |
1496 | DECL_NO_INSTRUMENT_FUNCTION_ENTRY_EXIT (decl) = 1; | |
1497 | DECL_SAVED_TREE (decl) = body; | |
1498 | TREE_PUBLIC (decl) = ! targetm.have_ctors_dtors; | |
1499 | DECL_UNINLINABLE (decl) = 1; | |
1500 | ||
1501 | DECL_INITIAL (decl) = make_node (BLOCK); | |
1502 | TREE_USED (DECL_INITIAL (decl)) = 1; | |
1503 | ||
1504 | DECL_SOURCE_LOCATION (decl) = input_location; | |
1505 | cfun->function_end_locus = input_location; | |
1506 | ||
1507 | switch (which) | |
1508 | { | |
1509 | case 'I': | |
1510 | DECL_STATIC_CONSTRUCTOR (decl) = 1; | |
1511 | break; | |
1512 | case 'D': | |
1513 | DECL_STATIC_DESTRUCTOR (decl) = 1; | |
1514 | break; | |
1515 | default: | |
1516 | gcc_unreachable (); | |
1517 | } | |
1518 | ||
1519 | gimplify_function_tree (decl); | |
1520 | ||
1521 | /* ??? We will get called LATE in the compilation process. */ | |
1522 | if (cgraph_global_info_ready) | |
1523 | { | |
1524 | tree_lowering_passes (decl); | |
1525 | tree_rest_of_compilation (decl); | |
1526 | } | |
1527 | else | |
1528 | cgraph_finalize_function (decl, 0); | |
1529 | ||
1530 | if (targetm.have_ctors_dtors) | |
1531 | { | |
1532 | void (*fn) (rtx, int); | |
1533 | ||
1534 | if (which == 'I') | |
1535 | fn = targetm.asm_out.constructor; | |
1536 | else | |
1537 | fn = targetm.asm_out.destructor; | |
1538 | fn (XEXP (DECL_RTL (decl), 0), priority); | |
1539 | } | |
1540 | } | |
1541 | ||
1542 | void | |
1543 | init_cgraph (void) | |
1544 | { | |
1545 | cgraph_dump_file = dump_begin (TDI_cgraph, NULL); | |
1546 | } | |
1547 | ||
1548 | /* The edges representing the callers of the NEW_VERSION node were | |
1549 | fixed by cgraph_function_versioning (), now the call_expr in their | |
1550 | respective tree code should be updated to call the NEW_VERSION. */ | |
1551 | ||
1552 | static void | |
1553 | update_call_expr (struct cgraph_node *new_version) | |
1554 | { | |
1555 | struct cgraph_edge *e; | |
1556 | ||
1557 | gcc_assert (new_version); | |
1558 | for (e = new_version->callers; e; e = e->next_caller) | |
1559 | /* Update the call expr on the edges | |
1560 | to call the new version. */ | |
1561 | TREE_OPERAND (TREE_OPERAND (get_call_expr_in (e->call_stmt), 0), 0) = new_version->decl; | |
1562 | } | |
1563 | ||
1564 | ||
1565 | /* Create a new cgraph node which is the new version of | |
1566 | OLD_VERSION node. REDIRECT_CALLERS holds the callers | |
1567 | edges which should be redirected to point to | |
1568 | NEW_VERSION. ALL the callees edges of OLD_VERSION | |
1569 | are cloned to the new version node. Return the new | |
1570 | version node. */ | |
1571 | ||
1572 | static struct cgraph_node * | |
1573 | cgraph_copy_node_for_versioning (struct cgraph_node *old_version, | |
1574 | tree new_decl, varray_type redirect_callers) | |
1575 | { | |
1576 | struct cgraph_node *new_version; | |
1577 | struct cgraph_edge *e, *new_e; | |
1578 | struct cgraph_edge *next_callee; | |
1579 | unsigned i; | |
1580 | ||
1581 | gcc_assert (old_version); | |
1582 | ||
1583 | new_version = cgraph_node (new_decl); | |
1584 | ||
1585 | new_version->analyzed = true; | |
1586 | new_version->local = old_version->local; | |
1587 | new_version->global = old_version->global; | |
1588 | new_version->rtl = new_version->rtl; | |
1589 | new_version->reachable = true; | |
1590 | new_version->count = old_version->count; | |
1591 | ||
1592 | /* Clone the old node callees. Recursive calls are | |
1593 | also cloned. */ | |
1594 | for (e = old_version->callees;e; e=e->next_callee) | |
1595 | { | |
1596 | new_e = cgraph_clone_edge (e, new_version, e->call_stmt, 0, e->loop_nest, true); | |
1597 | new_e->count = e->count; | |
1598 | } | |
1599 | /* Fix recursive calls. | |
1600 | If OLD_VERSION has a recursive call after the | |
1601 | previous edge cloning, the new version will have an edge | |
1602 | pointing to the old version, which is wrong; | |
1603 | Redirect it to point to the new version. */ | |
1604 | for (e = new_version->callees ; e; e = next_callee) | |
1605 | { | |
1606 | next_callee = e->next_callee; | |
1607 | if (e->callee == old_version) | |
1608 | cgraph_redirect_edge_callee (e, new_version); | |
1609 | ||
1610 | if (!next_callee) | |
1611 | break; | |
1612 | } | |
1613 | if (redirect_callers) | |
1614 | for (i = 0; i < VARRAY_ACTIVE_SIZE (redirect_callers); i++) | |
1615 | { | |
1616 | e = VARRAY_GENERIC_PTR (redirect_callers, i); | |
1617 | /* Redirect calls to the old version node | |
1618 | to point to it's new version. */ | |
1619 | cgraph_redirect_edge_callee (e, new_version); | |
1620 | } | |
1621 | ||
1622 | return new_version; | |
1623 | } | |
1624 | ||
1625 | /* Perform function versioning. | |
1626 | Function versioning includes copying of the tree and | |
1627 | a callgraph update (creating a new cgraph node and updating | |
1628 | its callees and callers). | |
1629 | ||
1630 | REDIRECT_CALLERS varray includes the edges to be redirected | |
1631 | to the new version. | |
1632 | ||
1633 | TREE_MAP is a mapping of tree nodes we want to replace with | |
1634 | new ones (according to results of prior analysis). | |
1635 | OLD_VERSION_NODE is the node that is versioned. | |
1636 | It returns the new version's cgraph node. */ | |
1637 | ||
1638 | struct cgraph_node * | |
1639 | cgraph_function_versioning (struct cgraph_node *old_version_node, | |
1640 | varray_type redirect_callers, | |
1641 | varray_type tree_map) | |
1642 | { | |
1643 | tree old_decl = old_version_node->decl; | |
1644 | struct cgraph_node *new_version_node = NULL; | |
1645 | tree new_decl; | |
1646 | ||
1647 | if (!tree_versionable_function_p (old_decl)) | |
1648 | return NULL; | |
1649 | ||
1650 | /* Make a new FUNCTION_DECL tree node for the | |
1651 | new version. */ | |
1652 | new_decl = copy_node (old_decl); | |
1653 | ||
1654 | /* Create the new version's call-graph node. | |
1655 | and update the edges of the new node. */ | |
1656 | new_version_node = | |
1657 | cgraph_copy_node_for_versioning (old_version_node, new_decl, | |
1658 | redirect_callers); | |
1659 | ||
1660 | /* Copy the OLD_VERSION_NODE function tree to the new version. */ | |
1661 | tree_function_versioning (old_decl, new_decl, tree_map, false); | |
1662 | /* Update the call_expr on the edges to call the new version node. */ | |
1663 | update_call_expr (new_version_node); | |
1664 | ||
1665 | /* Update the new version's properties. | |
1666 | Make The new version visible only within this translation unit. | |
1667 | ??? We cannot use COMDAT linkage because there is no | |
1668 | ABI support for this. */ | |
1669 | DECL_EXTERNAL (new_version_node->decl) = 0; | |
1670 | DECL_ONE_ONLY (new_version_node->decl) = 0; | |
1671 | TREE_PUBLIC (new_version_node->decl) = 0; | |
1672 | DECL_COMDAT (new_version_node->decl) = 0; | |
1673 | new_version_node->local.externally_visible = 0; | |
1674 | new_version_node->local.local = 1; | |
1675 | new_version_node->lowered = true; | |
1676 | return new_version_node; | |
1677 | } | |
1678 | ||
1679 | /* Produce separate function body for inline clones so the offline copy can be | |
1680 | modified without affecting them. */ | |
1681 | struct cgraph_node * | |
1682 | save_inline_function_body (struct cgraph_node *node) | |
1683 | { | |
1684 | struct cgraph_node *first_clone; | |
1685 | ||
1686 | gcc_assert (node == cgraph_node (node->decl)); | |
1687 | ||
1688 | cgraph_lower_function (node); | |
1689 | ||
1690 | /* In non-unit-at-a-time we construct full fledged clone we never output to | |
1691 | assembly file. This clone is pointed out by inline_decl of orginal function | |
1692 | and inlining infrastructure knows how to deal with this. */ | |
1693 | if (!flag_unit_at_a_time) | |
1694 | { | |
1695 | struct cgraph_edge *e; | |
1696 | ||
1697 | first_clone = cgraph_clone_node (node, node->count, 0, false); | |
1698 | first_clone->needed = 0; | |
1699 | first_clone->reachable = 1; | |
1700 | /* Recursively clone all bodies. */ | |
1701 | for (e = first_clone->callees; e; e = e->next_callee) | |
1702 | if (!e->inline_failed) | |
1703 | cgraph_clone_inlined_nodes (e, true, false); | |
1704 | } | |
1705 | else | |
1706 | first_clone = node->next_clone; | |
1707 | ||
1708 | first_clone->decl = copy_node (node->decl); | |
1709 | node->next_clone = NULL; | |
1710 | if (!flag_unit_at_a_time) | |
1711 | node->inline_decl = first_clone->decl; | |
1712 | first_clone->prev_clone = NULL; | |
1713 | cgraph_insert_node_to_hashtable (first_clone); | |
1714 | gcc_assert (first_clone == cgraph_node (first_clone->decl)); | |
1715 | ||
1716 | /* Copy the OLD_VERSION_NODE function tree to the new version. */ | |
1717 | tree_function_versioning (node->decl, first_clone->decl, NULL, true); | |
1718 | ||
1719 | DECL_EXTERNAL (first_clone->decl) = 0; | |
1720 | DECL_ONE_ONLY (first_clone->decl) = 0; | |
1721 | TREE_PUBLIC (first_clone->decl) = 0; | |
1722 | DECL_COMDAT (first_clone->decl) = 0; | |
1723 | ||
1724 | for (node = first_clone->next_clone; node; node = node->next_clone) | |
1725 | node->decl = first_clone->decl; | |
1726 | #ifdef ENABLE_CHECKING | |
1727 | verify_cgraph_node (first_clone); | |
1728 | #endif | |
1729 | return first_clone; | |
1730 | } | |
1731 |