]>
Commit | Line | Data |
---|---|---|
3e485f62 | 1 | /* Function splitting pass |
5624e564 | 2 | Copyright (C) 2010-2015 Free Software Foundation, Inc. |
3e485f62 JH |
3 | Contributed by Jan Hubicka <jh@suse.cz> |
4 | ||
5 | This file is part of GCC. | |
6 | ||
7 | GCC is free software; you can redistribute it and/or modify it under | |
8 | the terms of the GNU General Public License as published by the Free | |
9 | Software Foundation; either version 3, or (at your option) any later | |
10 | version. | |
11 | ||
12 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
13 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
14 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
15 | for more details. | |
16 | ||
17 | You should have received a copy of the GNU General Public License | |
18 | along with GCC; see the file COPYING3. If not see | |
19 | <http://www.gnu.org/licenses/>. */ | |
20 | ||
21 | /* The purpose of this pass is to split function bodies to improve | |
22 | inlining. I.e. for function of the form: | |
23 | ||
24 | func (...) | |
25 | { | |
26 | if (cheap_test) | |
27 | something_small | |
28 | else | |
29 | something_big | |
30 | } | |
31 | ||
32 | Produce: | |
33 | ||
34 | func.part (...) | |
35 | { | |
36 | something_big | |
37 | } | |
38 | ||
39 | func (...) | |
40 | { | |
41 | if (cheap_test) | |
42 | something_small | |
43 | else | |
44 | func.part (...); | |
45 | } | |
46 | ||
47 | When func becomes inlinable and when cheap_test is often true, inlining func, | |
ed7656f6 | 48 | but not fund.part leads to performance improvement similar as inlining |
3e485f62 JH |
49 | original func while the code size growth is smaller. |
50 | ||
51 | The pass is organized in three stages: | |
52 | 1) Collect local info about basic block into BB_INFO structure and | |
53 | compute function body estimated size and time. | |
54 | 2) Via DFS walk find all possible basic blocks where we can split | |
55 | and chose best one. | |
56 | 3) If split point is found, split at the specified BB by creating a clone | |
57 | and updating function to call it. | |
58 | ||
59 | The decisions what functions to split are in execute_split_functions | |
60 | and consider_split. | |
61 | ||
62 | There are several possible future improvements for this pass including: | |
63 | ||
64 | 1) Splitting to break up large functions | |
65 | 2) Splitting to reduce stack frame usage | |
66 | 3) Allow split part of function to use values computed in the header part. | |
67 | The values needs to be passed to split function, perhaps via same | |
68 | interface as for nested functions or as argument. | |
69 | 4) Support for simple rematerialization. I.e. when split part use | |
70 | value computed in header from function parameter in very cheap way, we | |
71 | can just recompute it. | |
72 | 5) Support splitting of nested functions. | |
73 | 6) Support non-SSA arguments. | |
74 | 7) There is nothing preventing us from producing multiple parts of single function | |
75 | when needed or splitting also the parts. */ | |
76 | ||
77 | #include "config.h" | |
78 | #include "system.h" | |
79 | #include "coretypes.h" | |
c7131fb2 | 80 | #include "backend.h" |
957060b5 | 81 | #include "rtl.h" |
40e23961 | 82 | #include "tree.h" |
c7131fb2 | 83 | #include "gimple.h" |
957060b5 AM |
84 | #include "cfghooks.h" |
85 | #include "alloc-pool.h" | |
86 | #include "tree-pass.h" | |
c7131fb2 | 87 | #include "ssa.h" |
957060b5 AM |
88 | #include "cgraph.h" |
89 | #include "diagnostic.h" | |
40e23961 | 90 | #include "fold-const.h" |
60393bbc | 91 | #include "cfganal.h" |
d8a2d370 | 92 | #include "calls.h" |
45b0be94 | 93 | #include "gimplify.h" |
5be5c238 | 94 | #include "gimple-iterator.h" |
18f429e2 | 95 | #include "gimplify-me.h" |
5be5c238 | 96 | #include "gimple-walk.h" |
dd912cb8 | 97 | #include "symbol-summary.h" |
3e485f62 | 98 | #include "ipa-prop.h" |
442b4905 | 99 | #include "tree-cfg.h" |
442b4905 AM |
100 | #include "tree-into-ssa.h" |
101 | #include "tree-dfa.h" | |
3e485f62 | 102 | #include "tree-inline.h" |
3e485f62 JH |
103 | #include "params.h" |
104 | #include "gimple-pretty-print.h" | |
e7f23018 | 105 | #include "ipa-inline.h" |
a9e0d843 | 106 | #include "cfgloop.h" |
d5e254e1 | 107 | #include "tree-chkp.h" |
3e485f62 JH |
108 | |
109 | /* Per basic block info. */ | |
110 | ||
50686850 | 111 | struct split_bb_info |
3e485f62 JH |
112 | { |
113 | unsigned int size; | |
114 | unsigned int time; | |
50686850 | 115 | }; |
3e485f62 | 116 | |
11478306 | 117 | static vec<split_bb_info> bb_info_vec; |
3e485f62 JH |
118 | |
119 | /* Description of split point. */ | |
120 | ||
121 | struct split_point | |
122 | { | |
123 | /* Size of the partitions. */ | |
124 | unsigned int header_time, header_size, split_time, split_size; | |
125 | ||
ed7656f6 | 126 | /* SSA names that need to be passed into spit function. */ |
3e485f62 JH |
127 | bitmap ssa_names_to_pass; |
128 | ||
129 | /* Basic block where we split (that will become entry point of new function. */ | |
130 | basic_block entry_bb; | |
131 | ||
132 | /* Basic blocks we are splitting away. */ | |
133 | bitmap split_bbs; | |
241a2b9e JH |
134 | |
135 | /* True when return value is computed on split part and thus it needs | |
136 | to be returned. */ | |
137 | bool split_part_set_retval; | |
3e485f62 JH |
138 | }; |
139 | ||
140 | /* Best split point found. */ | |
141 | ||
142 | struct split_point best_split_point; | |
143 | ||
b2e25729 BS |
144 | /* Set of basic blocks that are not allowed to dominate a split point. */ |
145 | ||
146 | static bitmap forbidden_dominators; | |
147 | ||
241a2b9e | 148 | static tree find_retval (basic_block return_bb); |
d5e254e1 | 149 | static tree find_retbnd (basic_block return_bb); |
241a2b9e | 150 | |
1802378d | 151 | /* Callback for walk_stmt_load_store_addr_ops. If T is non-SSA automatic |
3e485f62 JH |
152 | variable, check it if it is present in bitmap passed via DATA. */ |
153 | ||
154 | static bool | |
355fe088 | 155 | test_nonssa_use (gimple *, tree t, tree, void *data) |
3e485f62 JH |
156 | { |
157 | t = get_base_address (t); | |
158 | ||
1802378d EB |
159 | if (!t || is_gimple_reg (t)) |
160 | return false; | |
161 | ||
162 | if (TREE_CODE (t) == PARM_DECL | |
163 | || (TREE_CODE (t) == VAR_DECL | |
3e485f62 | 164 | && auto_var_in_fn_p (t, current_function_decl)) |
1802378d | 165 | || TREE_CODE (t) == RESULT_DECL |
412c4af7 JH |
166 | /* Normal labels are part of CFG and will be handled gratefuly. |
167 | Forced labels however can be used directly by statements and | |
168 | need to stay in one partition along with their uses. */ | |
169 | || (TREE_CODE (t) == LABEL_DECL | |
170 | && FORCED_LABEL (t))) | |
3e485f62 | 171 | return bitmap_bit_p ((bitmap)data, DECL_UID (t)); |
241a2b9e | 172 | |
1802378d EB |
173 | /* For DECL_BY_REFERENCE, the return value is actually a pointer. We want |
174 | to pretend that the value pointed to is actual result decl. */ | |
175 | if ((TREE_CODE (t) == MEM_REF || INDIRECT_REF_P (t)) | |
241a2b9e | 176 | && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME |
70b5e7dc | 177 | && SSA_NAME_VAR (TREE_OPERAND (t, 0)) |
241a2b9e JH |
178 | && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND (t, 0))) == RESULT_DECL |
179 | && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) | |
1802378d EB |
180 | return |
181 | bitmap_bit_p ((bitmap)data, | |
182 | DECL_UID (DECL_RESULT (current_function_decl))); | |
183 | ||
3e485f62 JH |
184 | return false; |
185 | } | |
186 | ||
187 | /* Dump split point CURRENT. */ | |
188 | ||
189 | static void | |
190 | dump_split_point (FILE * file, struct split_point *current) | |
191 | { | |
192 | fprintf (file, | |
cfef45c8 RG |
193 | "Split point at BB %i\n" |
194 | " header time: %i header size: %i\n" | |
195 | " split time: %i split size: %i\n bbs: ", | |
3e485f62 JH |
196 | current->entry_bb->index, current->header_time, |
197 | current->header_size, current->split_time, current->split_size); | |
198 | dump_bitmap (file, current->split_bbs); | |
199 | fprintf (file, " SSA names to pass: "); | |
200 | dump_bitmap (file, current->ssa_names_to_pass); | |
201 | } | |
202 | ||
1802378d EB |
203 | /* Look for all BBs in header that might lead to the split part and verify |
204 | that they are not defining any non-SSA var used by the split part. | |
2094f1fc JH |
205 | Parameters are the same as for consider_split. */ |
206 | ||
207 | static bool | |
208 | verify_non_ssa_vars (struct split_point *current, bitmap non_ssa_vars, | |
209 | basic_block return_bb) | |
210 | { | |
211 | bitmap seen = BITMAP_ALLOC (NULL); | |
6e1aa848 | 212 | vec<basic_block> worklist = vNULL; |
2094f1fc JH |
213 | edge e; |
214 | edge_iterator ei; | |
215 | bool ok = true; | |
412c4af7 | 216 | basic_block bb; |
1802378d | 217 | |
2094f1fc | 218 | FOR_EACH_EDGE (e, ei, current->entry_bb->preds) |
fefa31b5 | 219 | if (e->src != ENTRY_BLOCK_PTR_FOR_FN (cfun) |
2094f1fc JH |
220 | && !bitmap_bit_p (current->split_bbs, e->src->index)) |
221 | { | |
9771b263 | 222 | worklist.safe_push (e->src); |
2094f1fc JH |
223 | bitmap_set_bit (seen, e->src->index); |
224 | } | |
1802378d | 225 | |
9771b263 | 226 | while (!worklist.is_empty ()) |
2094f1fc | 227 | { |
412c4af7 | 228 | bb = worklist.pop (); |
2094f1fc | 229 | FOR_EACH_EDGE (e, ei, bb->preds) |
fefa31b5 | 230 | if (e->src != ENTRY_BLOCK_PTR_FOR_FN (cfun) |
fcaa4ca4 | 231 | && bitmap_set_bit (seen, e->src->index)) |
2094f1fc JH |
232 | { |
233 | gcc_checking_assert (!bitmap_bit_p (current->split_bbs, | |
234 | e->src->index)); | |
9771b263 | 235 | worklist.safe_push (e->src); |
2094f1fc | 236 | } |
538dd0b7 DM |
237 | for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi); |
238 | gsi_next (&bsi)) | |
2094f1fc | 239 | { |
355fe088 | 240 | gimple *stmt = gsi_stmt (bsi); |
1802378d | 241 | if (is_gimple_debug (stmt)) |
2094f1fc JH |
242 | continue; |
243 | if (walk_stmt_load_store_addr_ops | |
1802378d EB |
244 | (stmt, non_ssa_vars, test_nonssa_use, test_nonssa_use, |
245 | test_nonssa_use)) | |
2094f1fc JH |
246 | { |
247 | ok = false; | |
248 | goto done; | |
249 | } | |
538dd0b7 DM |
250 | if (glabel *label_stmt = dyn_cast <glabel *> (stmt)) |
251 | if (test_nonssa_use (stmt, gimple_label_label (label_stmt), | |
252 | NULL_TREE, non_ssa_vars)) | |
253 | { | |
254 | ok = false; | |
255 | goto done; | |
256 | } | |
2094f1fc | 257 | } |
538dd0b7 DM |
258 | for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi); |
259 | gsi_next (&bsi)) | |
2094f1fc JH |
260 | { |
261 | if (walk_stmt_load_store_addr_ops | |
1802378d EB |
262 | (gsi_stmt (bsi), non_ssa_vars, test_nonssa_use, test_nonssa_use, |
263 | test_nonssa_use)) | |
2094f1fc JH |
264 | { |
265 | ok = false; | |
266 | goto done; | |
267 | } | |
268 | } | |
269 | FOR_EACH_EDGE (e, ei, bb->succs) | |
270 | { | |
271 | if (e->dest != return_bb) | |
272 | continue; | |
538dd0b7 DM |
273 | for (gphi_iterator bsi = gsi_start_phis (return_bb); |
274 | !gsi_end_p (bsi); | |
2094f1fc JH |
275 | gsi_next (&bsi)) |
276 | { | |
538dd0b7 | 277 | gphi *stmt = bsi.phi (); |
2094f1fc JH |
278 | tree op = gimple_phi_arg_def (stmt, e->dest_idx); |
279 | ||
ea057359 | 280 | if (virtual_operand_p (gimple_phi_result (stmt))) |
2094f1fc JH |
281 | continue; |
282 | if (TREE_CODE (op) != SSA_NAME | |
9f1363cd | 283 | && test_nonssa_use (stmt, op, op, non_ssa_vars)) |
2094f1fc JH |
284 | { |
285 | ok = false; | |
286 | goto done; | |
287 | } | |
288 | } | |
289 | } | |
290 | } | |
412c4af7 JH |
291 | |
292 | /* Verify that the rest of function does not define any label | |
293 | used by the split part. */ | |
294 | FOR_EACH_BB_FN (bb, cfun) | |
295 | if (!bitmap_bit_p (current->split_bbs, bb->index) | |
296 | && !bitmap_bit_p (seen, bb->index)) | |
297 | { | |
298 | gimple_stmt_iterator bsi; | |
299 | for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
538dd0b7 | 300 | if (glabel *label_stmt = dyn_cast <glabel *> (gsi_stmt (bsi))) |
412c4af7 | 301 | { |
538dd0b7 DM |
302 | if (test_nonssa_use (label_stmt, |
303 | gimple_label_label (label_stmt), | |
304 | NULL_TREE, non_ssa_vars)) | |
305 | { | |
306 | ok = false; | |
307 | goto done; | |
308 | } | |
412c4af7 | 309 | } |
538dd0b7 | 310 | else |
412c4af7 JH |
311 | break; |
312 | } | |
313 | ||
2094f1fc JH |
314 | done: |
315 | BITMAP_FREE (seen); | |
9771b263 | 316 | worklist.release (); |
2094f1fc JH |
317 | return ok; |
318 | } | |
319 | ||
b2e25729 BS |
320 | /* If STMT is a call, check the callee against a list of forbidden |
321 | predicate functions. If a match is found, look for uses of the | |
322 | call result in condition statements that compare against zero. | |
323 | For each such use, find the block targeted by the condition | |
324 | statement for the nonzero result, and set the bit for this block | |
325 | in the forbidden dominators bitmap. The purpose of this is to avoid | |
326 | selecting a split point where we are likely to lose the chance | |
327 | to optimize away an unused function call. */ | |
328 | ||
329 | static void | |
355fe088 | 330 | check_forbidden_calls (gimple *stmt) |
b2e25729 BS |
331 | { |
332 | imm_use_iterator use_iter; | |
333 | use_operand_p use_p; | |
334 | tree lhs; | |
335 | ||
336 | /* At the moment, __builtin_constant_p is the only forbidden | |
337 | predicate function call (see PR49642). */ | |
338 | if (!gimple_call_builtin_p (stmt, BUILT_IN_CONSTANT_P)) | |
339 | return; | |
340 | ||
341 | lhs = gimple_call_lhs (stmt); | |
342 | ||
343 | if (!lhs || TREE_CODE (lhs) != SSA_NAME) | |
344 | return; | |
345 | ||
346 | FOR_EACH_IMM_USE_FAST (use_p, use_iter, lhs) | |
347 | { | |
348 | tree op1; | |
349 | basic_block use_bb, forbidden_bb; | |
350 | enum tree_code code; | |
351 | edge true_edge, false_edge; | |
538dd0b7 | 352 | gcond *use_stmt; |
b2e25729 | 353 | |
538dd0b7 DM |
354 | use_stmt = dyn_cast <gcond *> (USE_STMT (use_p)); |
355 | if (!use_stmt) | |
b2e25729 BS |
356 | continue; |
357 | ||
358 | /* Assuming canonical form for GIMPLE_COND here, with constant | |
359 | in second position. */ | |
360 | op1 = gimple_cond_rhs (use_stmt); | |
361 | code = gimple_cond_code (use_stmt); | |
362 | use_bb = gimple_bb (use_stmt); | |
363 | ||
364 | extract_true_false_edges_from_block (use_bb, &true_edge, &false_edge); | |
365 | ||
366 | /* We're only interested in comparisons that distinguish | |
367 | unambiguously from zero. */ | |
368 | if (!integer_zerop (op1) || code == LE_EXPR || code == GE_EXPR) | |
369 | continue; | |
370 | ||
371 | if (code == EQ_EXPR) | |
372 | forbidden_bb = false_edge->dest; | |
373 | else | |
374 | forbidden_bb = true_edge->dest; | |
375 | ||
376 | bitmap_set_bit (forbidden_dominators, forbidden_bb->index); | |
377 | } | |
378 | } | |
379 | ||
380 | /* If BB is dominated by any block in the forbidden dominators set, | |
381 | return TRUE; else FALSE. */ | |
382 | ||
383 | static bool | |
384 | dominated_by_forbidden (basic_block bb) | |
385 | { | |
386 | unsigned dom_bb; | |
387 | bitmap_iterator bi; | |
388 | ||
389 | EXECUTE_IF_SET_IN_BITMAP (forbidden_dominators, 1, dom_bb, bi) | |
390 | { | |
06e28de2 DM |
391 | if (dominated_by_p (CDI_DOMINATORS, bb, |
392 | BASIC_BLOCK_FOR_FN (cfun, dom_bb))) | |
b2e25729 BS |
393 | return true; |
394 | } | |
395 | ||
396 | return false; | |
397 | } | |
398 | ||
d5e254e1 IE |
399 | /* For give split point CURRENT and return block RETURN_BB return 1 |
400 | if ssa name VAL is set by split part and 0 otherwise. */ | |
401 | static bool | |
402 | split_part_set_ssa_name_p (tree val, struct split_point *current, | |
403 | basic_block return_bb) | |
404 | { | |
405 | if (TREE_CODE (val) != SSA_NAME) | |
406 | return false; | |
407 | ||
408 | return (!SSA_NAME_IS_DEFAULT_DEF (val) | |
409 | && (bitmap_bit_p (current->split_bbs, | |
410 | gimple_bb (SSA_NAME_DEF_STMT (val))->index) | |
411 | || gimple_bb (SSA_NAME_DEF_STMT (val)) == return_bb)); | |
412 | } | |
413 | ||
3e485f62 JH |
414 | /* We found an split_point CURRENT. NON_SSA_VARS is bitmap of all non ssa |
415 | variables used and RETURN_BB is return basic block. | |
416 | See if we can split function here. */ | |
417 | ||
418 | static void | |
419 | consider_split (struct split_point *current, bitmap non_ssa_vars, | |
420 | basic_block return_bb) | |
421 | { | |
422 | tree parm; | |
423 | unsigned int num_args = 0; | |
424 | unsigned int call_overhead; | |
425 | edge e; | |
426 | edge_iterator ei; | |
538dd0b7 | 427 | gphi_iterator bsi; |
8b3057b3 | 428 | unsigned int i; |
ed7656f6 | 429 | int incoming_freq = 0; |
241a2b9e | 430 | tree retval; |
d5e254e1 | 431 | tree retbnd; |
e70670cf | 432 | bool back_edge = false; |
8b3057b3 | 433 | |
3e485f62 JH |
434 | if (dump_file && (dump_flags & TDF_DETAILS)) |
435 | dump_split_point (dump_file, current); | |
436 | ||
8b3057b3 | 437 | FOR_EACH_EDGE (e, ei, current->entry_bb->preds) |
e70670cf JH |
438 | { |
439 | if (e->flags & EDGE_DFS_BACK) | |
440 | back_edge = true; | |
441 | if (!bitmap_bit_p (current->split_bbs, e->src->index)) | |
442 | incoming_freq += EDGE_FREQUENCY (e); | |
443 | } | |
8b3057b3 | 444 | |
3e485f62 | 445 | /* Do not split when we would end up calling function anyway. */ |
ed7656f6 | 446 | if (incoming_freq |
fefa31b5 | 447 | >= (ENTRY_BLOCK_PTR_FOR_FN (cfun)->frequency |
3e485f62 JH |
448 | * PARAM_VALUE (PARAM_PARTIAL_INLINING_ENTRY_PROBABILITY) / 100)) |
449 | { | |
e70670cf JH |
450 | /* When profile is guessed, we can not expect it to give us |
451 | realistic estimate on likelyness of function taking the | |
452 | complex path. As a special case, when tail of the function is | |
453 | a loop, enable splitting since inlining code skipping the loop | |
454 | is likely noticeable win. */ | |
455 | if (back_edge | |
0a6a6ac9 | 456 | && profile_status_for_fn (cfun) != PROFILE_READ |
fefa31b5 | 457 | && incoming_freq < ENTRY_BLOCK_PTR_FOR_FN (cfun)->frequency) |
e70670cf JH |
458 | { |
459 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
460 | fprintf (dump_file, | |
461 | " Split before loop, accepting despite low frequencies %i %i.\n", | |
462 | incoming_freq, | |
fefa31b5 | 463 | ENTRY_BLOCK_PTR_FOR_FN (cfun)->frequency); |
e70670cf JH |
464 | } |
465 | else | |
466 | { | |
467 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
468 | fprintf (dump_file, | |
469 | " Refused: incoming frequency is too large.\n"); | |
470 | return; | |
471 | } | |
3e485f62 JH |
472 | } |
473 | ||
474 | if (!current->header_size) | |
475 | { | |
476 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
477 | fprintf (dump_file, " Refused: header empty\n"); | |
3e485f62 JH |
478 | return; |
479 | } | |
480 | ||
ed7656f6 JJ |
481 | /* Verify that PHI args on entry are either virtual or all their operands |
482 | incoming from header are the same. */ | |
8b3057b3 | 483 | for (bsi = gsi_start_phis (current->entry_bb); !gsi_end_p (bsi); gsi_next (&bsi)) |
3e485f62 | 484 | { |
538dd0b7 | 485 | gphi *stmt = bsi.phi (); |
8b3057b3 JH |
486 | tree val = NULL; |
487 | ||
ea057359 | 488 | if (virtual_operand_p (gimple_phi_result (stmt))) |
8b3057b3 JH |
489 | continue; |
490 | for (i = 0; i < gimple_phi_num_args (stmt); i++) | |
491 | { | |
492 | edge e = gimple_phi_arg_edge (stmt, i); | |
493 | if (!bitmap_bit_p (current->split_bbs, e->src->index)) | |
494 | { | |
495 | tree edge_val = gimple_phi_arg_def (stmt, i); | |
496 | if (val && edge_val != val) | |
497 | { | |
498 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
499 | fprintf (dump_file, | |
500 | " Refused: entry BB has PHI with multiple variants\n"); | |
501 | return; | |
502 | } | |
503 | val = edge_val; | |
504 | } | |
505 | } | |
3e485f62 JH |
506 | } |
507 | ||
508 | ||
509 | /* See what argument we will pass to the split function and compute | |
510 | call overhead. */ | |
511 | call_overhead = eni_size_weights.call_cost; | |
512 | for (parm = DECL_ARGUMENTS (current_function_decl); parm; | |
910ad8de | 513 | parm = DECL_CHAIN (parm)) |
3e485f62 JH |
514 | { |
515 | if (!is_gimple_reg (parm)) | |
516 | { | |
517 | if (bitmap_bit_p (non_ssa_vars, DECL_UID (parm))) | |
518 | { | |
519 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
520 | fprintf (dump_file, | |
521 | " Refused: need to pass non-ssa param values\n"); | |
522 | return; | |
523 | } | |
524 | } | |
32244553 | 525 | else |
3e485f62 | 526 | { |
32244553 RG |
527 | tree ddef = ssa_default_def (cfun, parm); |
528 | if (ddef | |
529 | && bitmap_bit_p (current->ssa_names_to_pass, | |
530 | SSA_NAME_VERSION (ddef))) | |
531 | { | |
532 | if (!VOID_TYPE_P (TREE_TYPE (parm))) | |
b4c9af96 | 533 | call_overhead += estimate_move_cost (TREE_TYPE (parm), false); |
32244553 RG |
534 | num_args++; |
535 | } | |
3e485f62 JH |
536 | } |
537 | } | |
538 | if (!VOID_TYPE_P (TREE_TYPE (current_function_decl))) | |
b4c9af96 RB |
539 | call_overhead += estimate_move_cost (TREE_TYPE (current_function_decl), |
540 | false); | |
3e485f62 JH |
541 | |
542 | if (current->split_size <= call_overhead) | |
543 | { | |
544 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
545 | fprintf (dump_file, | |
546 | " Refused: split size is smaller than call overhead\n"); | |
547 | return; | |
548 | } | |
549 | if (current->header_size + call_overhead | |
550 | >= (unsigned int)(DECL_DECLARED_INLINE_P (current_function_decl) | |
551 | ? MAX_INLINE_INSNS_SINGLE | |
552 | : MAX_INLINE_INSNS_AUTO)) | |
553 | { | |
554 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
555 | fprintf (dump_file, | |
556 | " Refused: header size is too large for inline candidate\n"); | |
557 | return; | |
558 | } | |
559 | ||
7a161d5b JH |
560 | /* Splitting functions brings the target out of comdat group; this will |
561 | lead to code duplication if the function is reused by other unit. | |
562 | Limit this duplication. This is consistent with limit in tree-sra.c | |
563 | FIXME: with LTO we ought to be able to do better! */ | |
564 | if (DECL_ONE_ONLY (current_function_decl) | |
565 | && current->split_size >= (unsigned int) MAX_INLINE_INSNS_AUTO) | |
566 | { | |
567 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
568 | fprintf (dump_file, | |
569 | " Refused: function is COMDAT and tail is too large\n"); | |
570 | return; | |
571 | } | |
572 | /* For comdat functions also reject very small tails; those will likely get | |
573 | inlined back and we do not want to risk the duplication overhead. | |
574 | FIXME: with LTO we ought to be able to do better! */ | |
575 | if (DECL_ONE_ONLY (current_function_decl) | |
576 | && current->split_size | |
577 | <= (unsigned int) PARAM_VALUE (PARAM_EARLY_INLINING_INSNS) / 2) | |
578 | { | |
579 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
580 | fprintf (dump_file, | |
581 | " Refused: function is COMDAT and tail is too small\n"); | |
582 | return; | |
583 | } | |
584 | ||
3e485f62 | 585 | /* FIXME: we currently can pass only SSA function parameters to the split |
d402c33d | 586 | arguments. Once parm_adjustment infrastructure is supported by cloning, |
3e485f62 JH |
587 | we can pass more than that. */ |
588 | if (num_args != bitmap_count_bits (current->ssa_names_to_pass)) | |
589 | { | |
8b3057b3 | 590 | |
3e485f62 JH |
591 | if (dump_file && (dump_flags & TDF_DETAILS)) |
592 | fprintf (dump_file, | |
593 | " Refused: need to pass non-param values\n"); | |
594 | return; | |
595 | } | |
596 | ||
597 | /* When there are non-ssa vars used in the split region, see if they | |
598 | are used in the header region. If so, reject the split. | |
599 | FIXME: we can use nested function support to access both. */ | |
2094f1fc JH |
600 | if (!bitmap_empty_p (non_ssa_vars) |
601 | && !verify_non_ssa_vars (current, non_ssa_vars, return_bb)) | |
3e485f62 | 602 | { |
2094f1fc JH |
603 | if (dump_file && (dump_flags & TDF_DETAILS)) |
604 | fprintf (dump_file, | |
605 | " Refused: split part has non-ssa uses\n"); | |
3e485f62 JH |
606 | return; |
607 | } | |
b2e25729 BS |
608 | |
609 | /* If the split point is dominated by a forbidden block, reject | |
610 | the split. */ | |
611 | if (!bitmap_empty_p (forbidden_dominators) | |
612 | && dominated_by_forbidden (current->entry_bb)) | |
613 | { | |
614 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
615 | fprintf (dump_file, | |
616 | " Refused: split point dominated by forbidden block\n"); | |
617 | return; | |
618 | } | |
619 | ||
241a2b9e JH |
620 | /* See if retval used by return bb is computed by header or split part. |
621 | When it is computed by split part, we need to produce return statement | |
622 | in the split part and add code to header to pass it around. | |
623 | ||
624 | This is bit tricky to test: | |
625 | 1) When there is no return_bb or no return value, we always pass | |
626 | value around. | |
627 | 2) Invariants are always computed by caller. | |
628 | 3) For SSA we need to look if defining statement is in header or split part | |
629 | 4) For non-SSA we need to look where the var is computed. */ | |
630 | retval = find_retval (return_bb); | |
631 | if (!retval) | |
632 | current->split_part_set_retval = true; | |
633 | else if (is_gimple_min_invariant (retval)) | |
634 | current->split_part_set_retval = false; | |
635 | /* Special case is value returned by reference we record as if it was non-ssa | |
636 | set to result_decl. */ | |
637 | else if (TREE_CODE (retval) == SSA_NAME | |
70b5e7dc | 638 | && SSA_NAME_VAR (retval) |
241a2b9e JH |
639 | && TREE_CODE (SSA_NAME_VAR (retval)) == RESULT_DECL |
640 | && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) | |
641 | current->split_part_set_retval | |
642 | = bitmap_bit_p (non_ssa_vars, DECL_UID (SSA_NAME_VAR (retval))); | |
643 | else if (TREE_CODE (retval) == SSA_NAME) | |
644 | current->split_part_set_retval | |
d5e254e1 | 645 | = split_part_set_ssa_name_p (retval, current, return_bb); |
241a2b9e JH |
646 | else if (TREE_CODE (retval) == PARM_DECL) |
647 | current->split_part_set_retval = false; | |
648 | else if (TREE_CODE (retval) == VAR_DECL | |
649 | || TREE_CODE (retval) == RESULT_DECL) | |
650 | current->split_part_set_retval | |
651 | = bitmap_bit_p (non_ssa_vars, DECL_UID (retval)); | |
652 | else | |
653 | current->split_part_set_retval = true; | |
654 | ||
d5e254e1 IE |
655 | /* See if retbnd used by return bb is computed by header or split part. */ |
656 | retbnd = find_retbnd (return_bb); | |
657 | if (retbnd) | |
658 | { | |
659 | bool split_part_set_retbnd | |
660 | = split_part_set_ssa_name_p (retbnd, current, return_bb); | |
661 | ||
662 | /* If we have both return value and bounds then keep their definitions | |
663 | in a single function. We use SSA names to link returned bounds and | |
664 | value and therefore do not handle cases when result is passed by | |
665 | reference (which should not be our case anyway since bounds are | |
666 | returned for pointers only). */ | |
667 | if ((DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)) | |
668 | && current->split_part_set_retval) | |
669 | || split_part_set_retbnd != current->split_part_set_retval) | |
670 | { | |
671 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
672 | fprintf (dump_file, | |
673 | " Refused: split point splits return value and bounds\n"); | |
674 | return; | |
675 | } | |
676 | } | |
677 | ||
28fc44f3 JJ |
678 | /* split_function fixes up at most one PHI non-virtual PHI node in return_bb, |
679 | for the return value. If there are other PHIs, give up. */ | |
fefa31b5 | 680 | if (return_bb != EXIT_BLOCK_PTR_FOR_FN (cfun)) |
28fc44f3 | 681 | { |
538dd0b7 | 682 | gphi_iterator psi; |
28fc44f3 JJ |
683 | |
684 | for (psi = gsi_start_phis (return_bb); !gsi_end_p (psi); gsi_next (&psi)) | |
538dd0b7 | 685 | if (!virtual_operand_p (gimple_phi_result (psi.phi ())) |
28fc44f3 JJ |
686 | && !(retval |
687 | && current->split_part_set_retval | |
688 | && TREE_CODE (retval) == SSA_NAME | |
689 | && !DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)) | |
538dd0b7 | 690 | && SSA_NAME_DEF_STMT (retval) == psi.phi ())) |
28fc44f3 JJ |
691 | { |
692 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
693 | fprintf (dump_file, | |
694 | " Refused: return bb has extra PHIs\n"); | |
695 | return; | |
696 | } | |
697 | } | |
698 | ||
699 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
700 | fprintf (dump_file, " Accepted!\n"); | |
701 | ||
3e485f62 JH |
702 | /* At the moment chose split point with lowest frequency and that leaves |
703 | out smallest size of header. | |
704 | In future we might re-consider this heuristics. */ | |
705 | if (!best_split_point.split_bbs | |
706 | || best_split_point.entry_bb->frequency > current->entry_bb->frequency | |
707 | || (best_split_point.entry_bb->frequency == current->entry_bb->frequency | |
708 | && best_split_point.split_size < current->split_size)) | |
709 | ||
710 | { | |
711 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
712 | fprintf (dump_file, " New best split point!\n"); | |
713 | if (best_split_point.ssa_names_to_pass) | |
714 | { | |
715 | BITMAP_FREE (best_split_point.ssa_names_to_pass); | |
716 | BITMAP_FREE (best_split_point.split_bbs); | |
717 | } | |
718 | best_split_point = *current; | |
719 | best_split_point.ssa_names_to_pass = BITMAP_ALLOC (NULL); | |
720 | bitmap_copy (best_split_point.ssa_names_to_pass, | |
721 | current->ssa_names_to_pass); | |
722 | best_split_point.split_bbs = BITMAP_ALLOC (NULL); | |
723 | bitmap_copy (best_split_point.split_bbs, current->split_bbs); | |
724 | } | |
725 | } | |
726 | ||
2094f1fc JH |
727 | /* Return basic block containing RETURN statement. We allow basic blocks |
728 | of the form: | |
729 | <retval> = tmp_var; | |
730 | return <retval> | |
bfd71482 JJ |
731 | but return_bb can not be more complex than this (except for |
732 | -fsanitize=thread we allow TSAN_FUNC_EXIT () internal call in there). | |
6626665f | 733 | If nothing is found, return the exit block. |
2094f1fc | 734 | |
3e485f62 JH |
735 | When there are multiple RETURN statement, chose one with return value, |
736 | since that one is more likely shared by multiple code paths. | |
2094f1fc JH |
737 | |
738 | Return BB is special, because for function splitting it is the only | |
739 | basic block that is duplicated in between header and split part of the | |
740 | function. | |
741 | ||
3e485f62 JH |
742 | TODO: We might support multiple return blocks. */ |
743 | ||
744 | static basic_block | |
745 | find_return_bb (void) | |
746 | { | |
747 | edge e; | |
fefa31b5 | 748 | basic_block return_bb = EXIT_BLOCK_PTR_FOR_FN (cfun); |
68457901 JJ |
749 | gimple_stmt_iterator bsi; |
750 | bool found_return = false; | |
751 | tree retval = NULL_TREE; | |
3e485f62 | 752 | |
fefa31b5 | 753 | if (!single_pred_p (EXIT_BLOCK_PTR_FOR_FN (cfun))) |
68457901 JJ |
754 | return return_bb; |
755 | ||
fefa31b5 | 756 | e = single_pred_edge (EXIT_BLOCK_PTR_FOR_FN (cfun)); |
68457901 JJ |
757 | for (bsi = gsi_last_bb (e->src); !gsi_end_p (bsi); gsi_prev (&bsi)) |
758 | { | |
355fe088 | 759 | gimple *stmt = gsi_stmt (bsi); |
a348dc7f JJ |
760 | if (gimple_code (stmt) == GIMPLE_LABEL |
761 | || is_gimple_debug (stmt) | |
762 | || gimple_clobber_p (stmt)) | |
68457901 JJ |
763 | ; |
764 | else if (gimple_code (stmt) == GIMPLE_ASSIGN | |
765 | && found_return | |
766 | && gimple_assign_single_p (stmt) | |
767 | && (auto_var_in_fn_p (gimple_assign_rhs1 (stmt), | |
768 | current_function_decl) | |
769 | || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))) | |
770 | && retval == gimple_assign_lhs (stmt)) | |
771 | ; | |
538dd0b7 | 772 | else if (greturn *return_stmt = dyn_cast <greturn *> (stmt)) |
68457901 JJ |
773 | { |
774 | found_return = true; | |
538dd0b7 | 775 | retval = gimple_return_retval (return_stmt); |
68457901 | 776 | } |
bfd71482 JJ |
777 | /* For -fsanitize=thread, allow also TSAN_FUNC_EXIT () in the return |
778 | bb. */ | |
779 | else if ((flag_sanitize & SANITIZE_THREAD) | |
780 | && is_gimple_call (stmt) | |
781 | && gimple_call_internal_p (stmt) | |
782 | && gimple_call_internal_fn (stmt) == IFN_TSAN_FUNC_EXIT) | |
783 | ; | |
68457901 JJ |
784 | else |
785 | break; | |
786 | } | |
787 | if (gsi_end_p (bsi) && found_return) | |
788 | return_bb = e->src; | |
3e485f62 | 789 | |
3e485f62 JH |
790 | return return_bb; |
791 | } | |
792 | ||
ed7656f6 | 793 | /* Given return basic block RETURN_BB, see where return value is really |
2094f1fc JH |
794 | stored. */ |
795 | static tree | |
796 | find_retval (basic_block return_bb) | |
797 | { | |
798 | gimple_stmt_iterator bsi; | |
799 | for (bsi = gsi_start_bb (return_bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
538dd0b7 DM |
800 | if (greturn *return_stmt = dyn_cast <greturn *> (gsi_stmt (bsi))) |
801 | return gimple_return_retval (return_stmt); | |
a348dc7f JJ |
802 | else if (gimple_code (gsi_stmt (bsi)) == GIMPLE_ASSIGN |
803 | && !gimple_clobber_p (gsi_stmt (bsi))) | |
2094f1fc JH |
804 | return gimple_assign_rhs1 (gsi_stmt (bsi)); |
805 | return NULL; | |
806 | } | |
807 | ||
d5e254e1 IE |
808 | /* Given return basic block RETURN_BB, see where return bounds are really |
809 | stored. */ | |
810 | static tree | |
811 | find_retbnd (basic_block return_bb) | |
812 | { | |
813 | gimple_stmt_iterator bsi; | |
814 | for (bsi = gsi_last_bb (return_bb); !gsi_end_p (bsi); gsi_prev (&bsi)) | |
815 | if (gimple_code (gsi_stmt (bsi)) == GIMPLE_RETURN) | |
816 | return gimple_return_retbnd (gsi_stmt (bsi)); | |
817 | return NULL; | |
818 | } | |
819 | ||
1802378d EB |
820 | /* Callback for walk_stmt_load_store_addr_ops. If T is non-SSA automatic |
821 | variable, mark it as used in bitmap passed via DATA. | |
3e485f62 JH |
822 | Return true when access to T prevents splitting the function. */ |
823 | ||
824 | static bool | |
355fe088 | 825 | mark_nonssa_use (gimple *, tree t, tree, void *data) |
3e485f62 JH |
826 | { |
827 | t = get_base_address (t); | |
828 | ||
829 | if (!t || is_gimple_reg (t)) | |
830 | return false; | |
831 | ||
832 | /* At present we can't pass non-SSA arguments to split function. | |
833 | FIXME: this can be relaxed by passing references to arguments. */ | |
834 | if (TREE_CODE (t) == PARM_DECL) | |
835 | { | |
836 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1802378d EB |
837 | fprintf (dump_file, |
838 | "Cannot split: use of non-ssa function parameter.\n"); | |
3e485f62 JH |
839 | return true; |
840 | } | |
841 | ||
1802378d EB |
842 | if ((TREE_CODE (t) == VAR_DECL |
843 | && auto_var_in_fn_p (t, current_function_decl)) | |
844 | || TREE_CODE (t) == RESULT_DECL | |
412c4af7 JH |
845 | || (TREE_CODE (t) == LABEL_DECL |
846 | && FORCED_LABEL (t))) | |
3e485f62 | 847 | bitmap_set_bit ((bitmap)data, DECL_UID (t)); |
241a2b9e | 848 | |
1802378d EB |
849 | /* For DECL_BY_REFERENCE, the return value is actually a pointer. We want |
850 | to pretend that the value pointed to is actual result decl. */ | |
851 | if ((TREE_CODE (t) == MEM_REF || INDIRECT_REF_P (t)) | |
241a2b9e | 852 | && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME |
70b5e7dc | 853 | && SSA_NAME_VAR (TREE_OPERAND (t, 0)) |
241a2b9e JH |
854 | && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND (t, 0))) == RESULT_DECL |
855 | && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) | |
1802378d EB |
856 | return |
857 | bitmap_bit_p ((bitmap)data, | |
858 | DECL_UID (DECL_RESULT (current_function_decl))); | |
859 | ||
3e485f62 JH |
860 | return false; |
861 | } | |
862 | ||
863 | /* Compute local properties of basic block BB we collect when looking for | |
864 | split points. We look for ssa defs and store them in SET_SSA_NAMES, | |
865 | for ssa uses and store them in USED_SSA_NAMES and for any non-SSA automatic | |
866 | vars stored in NON_SSA_VARS. | |
867 | ||
868 | When BB has edge to RETURN_BB, collect uses in RETURN_BB too. | |
869 | ||
870 | Return false when BB contains something that prevents it from being put into | |
871 | split function. */ | |
872 | ||
873 | static bool | |
874 | visit_bb (basic_block bb, basic_block return_bb, | |
875 | bitmap set_ssa_names, bitmap used_ssa_names, | |
876 | bitmap non_ssa_vars) | |
877 | { | |
3e485f62 JH |
878 | edge e; |
879 | edge_iterator ei; | |
880 | bool can_split = true; | |
881 | ||
538dd0b7 DM |
882 | for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi); |
883 | gsi_next (&bsi)) | |
3e485f62 | 884 | { |
355fe088 | 885 | gimple *stmt = gsi_stmt (bsi); |
3e485f62 JH |
886 | tree op; |
887 | ssa_op_iter iter; | |
888 | tree decl; | |
889 | ||
890 | if (is_gimple_debug (stmt)) | |
891 | continue; | |
892 | ||
a348dc7f JJ |
893 | if (gimple_clobber_p (stmt)) |
894 | continue; | |
895 | ||
3e485f62 JH |
896 | /* FIXME: We can split regions containing EH. We can not however |
897 | split RESX, EH_DISPATCH and EH_POINTER referring to same region | |
898 | into different partitions. This would require tracking of | |
899 | EH regions and checking in consider_split_point if they | |
900 | are not used elsewhere. */ | |
1da7d8c0 | 901 | if (gimple_code (stmt) == GIMPLE_RESX) |
3e485f62 JH |
902 | { |
903 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1da7d8c0 | 904 | fprintf (dump_file, "Cannot split: resx.\n"); |
3e485f62 JH |
905 | can_split = false; |
906 | } | |
907 | if (gimple_code (stmt) == GIMPLE_EH_DISPATCH) | |
908 | { | |
909 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1802378d | 910 | fprintf (dump_file, "Cannot split: eh dispatch.\n"); |
3e485f62 JH |
911 | can_split = false; |
912 | } | |
913 | ||
914 | /* Check builtins that prevent splitting. */ | |
915 | if (gimple_code (stmt) == GIMPLE_CALL | |
916 | && (decl = gimple_call_fndecl (stmt)) != NULL_TREE | |
917 | && DECL_BUILT_IN (decl) | |
918 | && DECL_BUILT_IN_CLASS (decl) == BUILT_IN_NORMAL) | |
919 | switch (DECL_FUNCTION_CODE (decl)) | |
920 | { | |
921 | /* FIXME: once we will allow passing non-parm values to split part, | |
922 | we need to be sure to handle correct builtin_stack_save and | |
923 | builtin_stack_restore. At the moment we are safe; there is no | |
924 | way to store builtin_stack_save result in non-SSA variable | |
925 | since all calls to those are compiler generated. */ | |
926 | case BUILT_IN_APPLY: | |
61e03ffc | 927 | case BUILT_IN_APPLY_ARGS: |
3e485f62 JH |
928 | case BUILT_IN_VA_START: |
929 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1802378d EB |
930 | fprintf (dump_file, |
931 | "Cannot split: builtin_apply and va_start.\n"); | |
3e485f62 JH |
932 | can_split = false; |
933 | break; | |
934 | case BUILT_IN_EH_POINTER: | |
935 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1802378d | 936 | fprintf (dump_file, "Cannot split: builtin_eh_pointer.\n"); |
3e485f62 JH |
937 | can_split = false; |
938 | break; | |
939 | default: | |
940 | break; | |
941 | } | |
942 | ||
943 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF) | |
944 | bitmap_set_bit (set_ssa_names, SSA_NAME_VERSION (op)); | |
945 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE) | |
946 | bitmap_set_bit (used_ssa_names, SSA_NAME_VERSION (op)); | |
947 | can_split &= !walk_stmt_load_store_addr_ops (stmt, non_ssa_vars, | |
948 | mark_nonssa_use, | |
949 | mark_nonssa_use, | |
950 | mark_nonssa_use); | |
951 | } | |
538dd0b7 DM |
952 | for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi); |
953 | gsi_next (&bsi)) | |
3e485f62 | 954 | { |
538dd0b7 | 955 | gphi *stmt = bsi.phi (); |
8b3057b3 | 956 | unsigned int i; |
3e485f62 | 957 | |
ea057359 | 958 | if (virtual_operand_p (gimple_phi_result (stmt))) |
3e485f62 | 959 | continue; |
8b3057b3 JH |
960 | bitmap_set_bit (set_ssa_names, |
961 | SSA_NAME_VERSION (gimple_phi_result (stmt))); | |
962 | for (i = 0; i < gimple_phi_num_args (stmt); i++) | |
963 | { | |
964 | tree op = gimple_phi_arg_def (stmt, i); | |
965 | if (TREE_CODE (op) == SSA_NAME) | |
966 | bitmap_set_bit (used_ssa_names, SSA_NAME_VERSION (op)); | |
967 | } | |
3e485f62 JH |
968 | can_split &= !walk_stmt_load_store_addr_ops (stmt, non_ssa_vars, |
969 | mark_nonssa_use, | |
970 | mark_nonssa_use, | |
971 | mark_nonssa_use); | |
972 | } | |
ed7656f6 | 973 | /* Record also uses coming from PHI operand in return BB. */ |
3e485f62 JH |
974 | FOR_EACH_EDGE (e, ei, bb->succs) |
975 | if (e->dest == return_bb) | |
976 | { | |
538dd0b7 DM |
977 | for (gphi_iterator bsi = gsi_start_phis (return_bb); |
978 | !gsi_end_p (bsi); | |
979 | gsi_next (&bsi)) | |
3e485f62 | 980 | { |
538dd0b7 | 981 | gphi *stmt = bsi.phi (); |
3e485f62 JH |
982 | tree op = gimple_phi_arg_def (stmt, e->dest_idx); |
983 | ||
ea057359 | 984 | if (virtual_operand_p (gimple_phi_result (stmt))) |
3e485f62 | 985 | continue; |
3e485f62 JH |
986 | if (TREE_CODE (op) == SSA_NAME) |
987 | bitmap_set_bit (used_ssa_names, SSA_NAME_VERSION (op)); | |
988 | else | |
9f1363cd | 989 | can_split &= !mark_nonssa_use (stmt, op, op, non_ssa_vars); |
3e485f62 | 990 | } |
3e485f62 JH |
991 | } |
992 | return can_split; | |
993 | } | |
994 | ||
995 | /* Stack entry for recursive DFS walk in find_split_point. */ | |
996 | ||
50686850 | 997 | struct stack_entry |
3e485f62 JH |
998 | { |
999 | /* Basic block we are examining. */ | |
1000 | basic_block bb; | |
1001 | ||
1002 | /* SSA names set and used by the BB and all BBs reachable | |
1003 | from it via DFS walk. */ | |
1004 | bitmap set_ssa_names, used_ssa_names; | |
1005 | bitmap non_ssa_vars; | |
1006 | ||
1007 | /* All BBS visited from this BB via DFS walk. */ | |
1008 | bitmap bbs_visited; | |
1009 | ||
1010 | /* Last examined edge in DFS walk. Since we walk unoriented graph, | |
ed7656f6 | 1011 | the value is up to sum of incoming and outgoing edges of BB. */ |
3e485f62 JH |
1012 | unsigned int edge_num; |
1013 | ||
1014 | /* Stack entry index of earliest BB reachable from current BB | |
ed7656f6 | 1015 | or any BB visited later in DFS walk. */ |
3e485f62 JH |
1016 | int earliest; |
1017 | ||
1018 | /* Overall time and size of all BBs reached from this BB in DFS walk. */ | |
1019 | int overall_time, overall_size; | |
1020 | ||
1021 | /* When false we can not split on this BB. */ | |
1022 | bool can_split; | |
50686850 | 1023 | }; |
3e485f62 JH |
1024 | |
1025 | ||
1026 | /* Find all articulations and call consider_split on them. | |
1027 | OVERALL_TIME and OVERALL_SIZE is time and size of the function. | |
1028 | ||
1029 | We perform basic algorithm for finding an articulation in a graph | |
1030 | created from CFG by considering it to be an unoriented graph. | |
1031 | ||
1032 | The articulation is discovered via DFS walk. We collect earliest | |
1033 | basic block on stack that is reachable via backward edge. Articulation | |
1034 | is any basic block such that there is no backward edge bypassing it. | |
1035 | To reduce stack usage we maintain heap allocated stack in STACK vector. | |
1036 | AUX pointer of BB is set to index it appears in the stack or -1 once | |
1037 | it is visited and popped off the stack. | |
1038 | ||
1039 | The algorithm finds articulation after visiting the whole component | |
1040 | reachable by it. This makes it convenient to collect information about | |
1041 | the component used by consider_split. */ | |
1042 | ||
1043 | static void | |
bfd71482 | 1044 | find_split_points (basic_block return_bb, int overall_time, int overall_size) |
3e485f62 JH |
1045 | { |
1046 | stack_entry first; | |
6e1aa848 | 1047 | vec<stack_entry> stack = vNULL; |
3e485f62 | 1048 | basic_block bb; |
3e485f62 JH |
1049 | struct split_point current; |
1050 | ||
1051 | current.header_time = overall_time; | |
1052 | current.header_size = overall_size; | |
1053 | current.split_time = 0; | |
1054 | current.split_size = 0; | |
1055 | current.ssa_names_to_pass = BITMAP_ALLOC (NULL); | |
1056 | ||
fefa31b5 | 1057 | first.bb = ENTRY_BLOCK_PTR_FOR_FN (cfun); |
3e485f62 JH |
1058 | first.edge_num = 0; |
1059 | first.overall_time = 0; | |
1060 | first.overall_size = 0; | |
1061 | first.earliest = INT_MAX; | |
1062 | first.set_ssa_names = 0; | |
1063 | first.used_ssa_names = 0; | |
e15eb172 | 1064 | first.non_ssa_vars = 0; |
3e485f62 | 1065 | first.bbs_visited = 0; |
e15eb172 | 1066 | first.can_split = false; |
9771b263 | 1067 | stack.safe_push (first); |
fefa31b5 | 1068 | ENTRY_BLOCK_PTR_FOR_FN (cfun)->aux = (void *)(intptr_t)-1; |
3e485f62 | 1069 | |
9771b263 | 1070 | while (!stack.is_empty ()) |
3e485f62 | 1071 | { |
9771b263 | 1072 | stack_entry *entry = &stack.last (); |
3e485f62 JH |
1073 | |
1074 | /* We are walking an acyclic graph, so edge_num counts | |
1075 | succ and pred edges together. However when considering | |
1076 | articulation, we want to have processed everything reachable | |
1077 | from articulation but nothing that reaches into it. */ | |
1078 | if (entry->edge_num == EDGE_COUNT (entry->bb->succs) | |
fefa31b5 | 1079 | && entry->bb != ENTRY_BLOCK_PTR_FOR_FN (cfun)) |
3e485f62 | 1080 | { |
9771b263 | 1081 | int pos = stack.length (); |
3e485f62 JH |
1082 | entry->can_split &= visit_bb (entry->bb, return_bb, |
1083 | entry->set_ssa_names, | |
1084 | entry->used_ssa_names, | |
1085 | entry->non_ssa_vars); | |
1086 | if (pos <= entry->earliest && !entry->can_split | |
1087 | && dump_file && (dump_flags & TDF_DETAILS)) | |
1088 | fprintf (dump_file, | |
1089 | "found articulation at bb %i but can not split\n", | |
1090 | entry->bb->index); | |
1091 | if (pos <= entry->earliest && entry->can_split) | |
1092 | { | |
1093 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1094 | fprintf (dump_file, "found articulation at bb %i\n", | |
1095 | entry->bb->index); | |
1096 | current.entry_bb = entry->bb; | |
1097 | current.ssa_names_to_pass = BITMAP_ALLOC (NULL); | |
1098 | bitmap_and_compl (current.ssa_names_to_pass, | |
1099 | entry->used_ssa_names, entry->set_ssa_names); | |
1100 | current.header_time = overall_time - entry->overall_time; | |
1101 | current.header_size = overall_size - entry->overall_size; | |
1102 | current.split_time = entry->overall_time; | |
1103 | current.split_size = entry->overall_size; | |
1104 | current.split_bbs = entry->bbs_visited; | |
1105 | consider_split (¤t, entry->non_ssa_vars, return_bb); | |
1106 | BITMAP_FREE (current.ssa_names_to_pass); | |
1107 | } | |
1108 | } | |
1109 | /* Do actual DFS walk. */ | |
1110 | if (entry->edge_num | |
1111 | < (EDGE_COUNT (entry->bb->succs) | |
1112 | + EDGE_COUNT (entry->bb->preds))) | |
1113 | { | |
1114 | edge e; | |
1115 | basic_block dest; | |
1116 | if (entry->edge_num < EDGE_COUNT (entry->bb->succs)) | |
1117 | { | |
1118 | e = EDGE_SUCC (entry->bb, entry->edge_num); | |
1119 | dest = e->dest; | |
1120 | } | |
1121 | else | |
1122 | { | |
1123 | e = EDGE_PRED (entry->bb, entry->edge_num | |
1124 | - EDGE_COUNT (entry->bb->succs)); | |
1125 | dest = e->src; | |
1126 | } | |
1127 | ||
1128 | entry->edge_num++; | |
1129 | ||
1130 | /* New BB to visit, push it to the stack. */ | |
fefa31b5 | 1131 | if (dest != return_bb && dest != EXIT_BLOCK_PTR_FOR_FN (cfun) |
3e485f62 JH |
1132 | && !dest->aux) |
1133 | { | |
1134 | stack_entry new_entry; | |
1135 | ||
1136 | new_entry.bb = dest; | |
1137 | new_entry.edge_num = 0; | |
1138 | new_entry.overall_time | |
9771b263 | 1139 | = bb_info_vec[dest->index].time; |
3e485f62 | 1140 | new_entry.overall_size |
9771b263 | 1141 | = bb_info_vec[dest->index].size; |
3e485f62 JH |
1142 | new_entry.earliest = INT_MAX; |
1143 | new_entry.set_ssa_names = BITMAP_ALLOC (NULL); | |
1144 | new_entry.used_ssa_names = BITMAP_ALLOC (NULL); | |
1145 | new_entry.bbs_visited = BITMAP_ALLOC (NULL); | |
1146 | new_entry.non_ssa_vars = BITMAP_ALLOC (NULL); | |
1147 | new_entry.can_split = true; | |
1148 | bitmap_set_bit (new_entry.bbs_visited, dest->index); | |
9771b263 DN |
1149 | stack.safe_push (new_entry); |
1150 | dest->aux = (void *)(intptr_t)stack.length (); | |
3e485f62 JH |
1151 | } |
1152 | /* Back edge found, record the earliest point. */ | |
1153 | else if ((intptr_t)dest->aux > 0 | |
1154 | && (intptr_t)dest->aux < entry->earliest) | |
1155 | entry->earliest = (intptr_t)dest->aux; | |
1156 | } | |
ed7656f6 JJ |
1157 | /* We are done with examining the edges. Pop off the value from stack |
1158 | and merge stuff we accumulate during the walk. */ | |
fefa31b5 | 1159 | else if (entry->bb != ENTRY_BLOCK_PTR_FOR_FN (cfun)) |
3e485f62 | 1160 | { |
9771b263 | 1161 | stack_entry *prev = &stack[stack.length () - 2]; |
3e485f62 JH |
1162 | |
1163 | entry->bb->aux = (void *)(intptr_t)-1; | |
1164 | prev->can_split &= entry->can_split; | |
1165 | if (prev->set_ssa_names) | |
1166 | { | |
1167 | bitmap_ior_into (prev->set_ssa_names, entry->set_ssa_names); | |
1168 | bitmap_ior_into (prev->used_ssa_names, entry->used_ssa_names); | |
1169 | bitmap_ior_into (prev->bbs_visited, entry->bbs_visited); | |
1170 | bitmap_ior_into (prev->non_ssa_vars, entry->non_ssa_vars); | |
1171 | } | |
1172 | if (prev->earliest > entry->earliest) | |
1173 | prev->earliest = entry->earliest; | |
1174 | prev->overall_time += entry->overall_time; | |
1175 | prev->overall_size += entry->overall_size; | |
1176 | BITMAP_FREE (entry->set_ssa_names); | |
1177 | BITMAP_FREE (entry->used_ssa_names); | |
1178 | BITMAP_FREE (entry->bbs_visited); | |
1179 | BITMAP_FREE (entry->non_ssa_vars); | |
9771b263 | 1180 | stack.pop (); |
3e485f62 JH |
1181 | } |
1182 | else | |
9771b263 | 1183 | stack.pop (); |
3e485f62 | 1184 | } |
fefa31b5 | 1185 | ENTRY_BLOCK_PTR_FOR_FN (cfun)->aux = NULL; |
11cd3bed | 1186 | FOR_EACH_BB_FN (bb, cfun) |
3e485f62 | 1187 | bb->aux = NULL; |
9771b263 | 1188 | stack.release (); |
3e485f62 JH |
1189 | BITMAP_FREE (current.ssa_names_to_pass); |
1190 | } | |
1191 | ||
1192 | /* Split function at SPLIT_POINT. */ | |
1193 | ||
1194 | static void | |
bfd71482 JJ |
1195 | split_function (basic_block return_bb, struct split_point *split_point, |
1196 | bool add_tsan_func_exit) | |
3e485f62 | 1197 | { |
6e1aa848 | 1198 | vec<tree> args_to_pass = vNULL; |
201176d3 | 1199 | bitmap args_to_skip; |
3e485f62 JH |
1200 | tree parm; |
1201 | int num = 0; | |
d52f5295 | 1202 | cgraph_node *node, *cur_node = cgraph_node::get (current_function_decl); |
3e485f62 | 1203 | basic_block call_bb; |
bfd71482 | 1204 | gcall *call, *tsan_func_exit_call = NULL; |
3e485f62 JH |
1205 | edge e; |
1206 | edge_iterator ei; | |
d5e254e1 | 1207 | tree retval = NULL, real_retval = NULL, retbnd = NULL; |
3e485f62 | 1208 | bool split_part_return_p = false; |
d5e254e1 | 1209 | bool with_bounds = chkp_function_instrumented_p (current_function_decl); |
355fe088 | 1210 | gimple *last_stmt = NULL; |
371556ee | 1211 | unsigned int i; |
32244553 | 1212 | tree arg, ddef; |
9771b263 | 1213 | vec<tree, va_gc> **debug_args = NULL; |
3e485f62 JH |
1214 | |
1215 | if (dump_file) | |
1216 | { | |
1217 | fprintf (dump_file, "\n\nSplitting function at:\n"); | |
1218 | dump_split_point (dump_file, split_point); | |
1219 | } | |
1220 | ||
201176d3 MJ |
1221 | if (cur_node->local.can_change_signature) |
1222 | args_to_skip = BITMAP_ALLOC (NULL); | |
1223 | else | |
1224 | args_to_skip = NULL; | |
1225 | ||
3e485f62 JH |
1226 | /* Collect the parameters of new function and args_to_skip bitmap. */ |
1227 | for (parm = DECL_ARGUMENTS (current_function_decl); | |
910ad8de | 1228 | parm; parm = DECL_CHAIN (parm), num++) |
201176d3 MJ |
1229 | if (args_to_skip |
1230 | && (!is_gimple_reg (parm) | |
32244553 | 1231 | || (ddef = ssa_default_def (cfun, parm)) == NULL_TREE |
201176d3 | 1232 | || !bitmap_bit_p (split_point->ssa_names_to_pass, |
32244553 | 1233 | SSA_NAME_VERSION (ddef)))) |
3e485f62 JH |
1234 | bitmap_set_bit (args_to_skip, num); |
1235 | else | |
371556ee | 1236 | { |
86814190 MJ |
1237 | /* This parm might not have been used up to now, but is going to be |
1238 | used, hence register it. */ | |
86814190 | 1239 | if (is_gimple_reg (parm)) |
32244553 | 1240 | arg = get_or_create_ssa_default_def (cfun, parm); |
86814190 MJ |
1241 | else |
1242 | arg = parm; | |
201176d3 | 1243 | |
2b3c0885 RG |
1244 | if (!useless_type_conversion_p (DECL_ARG_TYPE (parm), TREE_TYPE (arg))) |
1245 | arg = fold_convert (DECL_ARG_TYPE (parm), arg); | |
9771b263 | 1246 | args_to_pass.safe_push (arg); |
371556ee | 1247 | } |
3e485f62 JH |
1248 | |
1249 | /* See if the split function will return. */ | |
1250 | FOR_EACH_EDGE (e, ei, return_bb->preds) | |
1251 | if (bitmap_bit_p (split_point->split_bbs, e->src->index)) | |
1252 | break; | |
1253 | if (e) | |
1254 | split_part_return_p = true; | |
1255 | ||
241a2b9e JH |
1256 | /* Add return block to what will become the split function. |
1257 | We do not return; no return block is needed. */ | |
1258 | if (!split_part_return_p) | |
1259 | ; | |
1260 | /* We have no return block, so nothing is needed. */ | |
fefa31b5 | 1261 | else if (return_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)) |
241a2b9e JH |
1262 | ; |
1263 | /* When we do not want to return value, we need to construct | |
1264 | new return block with empty return statement. | |
1265 | FIXME: Once we are able to change return type, we should change function | |
1266 | to return void instead of just outputting function with undefined return | |
1267 | value. For structures this affects quality of codegen. */ | |
1268 | else if (!split_point->split_part_set_retval | |
1269 | && find_retval (return_bb)) | |
1270 | { | |
1271 | bool redirected = true; | |
1272 | basic_block new_return_bb = create_basic_block (NULL, 0, return_bb); | |
1273 | gimple_stmt_iterator gsi = gsi_start_bb (new_return_bb); | |
1274 | gsi_insert_after (&gsi, gimple_build_return (NULL), GSI_NEW_STMT); | |
1275 | while (redirected) | |
1276 | { | |
1277 | redirected = false; | |
1278 | FOR_EACH_EDGE (e, ei, return_bb->preds) | |
1279 | if (bitmap_bit_p (split_point->split_bbs, e->src->index)) | |
1280 | { | |
1281 | new_return_bb->count += e->count; | |
1282 | new_return_bb->frequency += EDGE_FREQUENCY (e); | |
1283 | redirect_edge_and_branch (e, new_return_bb); | |
1284 | redirected = true; | |
1285 | break; | |
1286 | } | |
1287 | } | |
fefa31b5 | 1288 | e = make_edge (new_return_bb, EXIT_BLOCK_PTR_FOR_FN (cfun), 0); |
241a2b9e JH |
1289 | e->probability = REG_BR_PROB_BASE; |
1290 | e->count = new_return_bb->count; | |
726338f4 | 1291 | add_bb_to_loop (new_return_bb, current_loops->tree_root); |
241a2b9e | 1292 | bitmap_set_bit (split_point->split_bbs, new_return_bb->index); |
2e5e346d JL |
1293 | } |
1294 | /* When we pass around the value, use existing return block. */ | |
1295 | else | |
1296 | bitmap_set_bit (split_point->split_bbs, return_bb->index); | |
1297 | ||
1298 | /* If RETURN_BB has virtual operand PHIs, they must be removed and the | |
1299 | virtual operand marked for renaming as we change the CFG in a way that | |
cfef45c8 | 1300 | tree-inline is not able to compensate for. |
2e5e346d JL |
1301 | |
1302 | Note this can happen whether or not we have a return value. If we have | |
1303 | a return value, then RETURN_BB may have PHIs for real operands too. */ | |
fefa31b5 | 1304 | if (return_bb != EXIT_BLOCK_PTR_FOR_FN (cfun)) |
2e5e346d | 1305 | { |
cfef45c8 | 1306 | bool phi_p = false; |
538dd0b7 DM |
1307 | for (gphi_iterator gsi = gsi_start_phis (return_bb); |
1308 | !gsi_end_p (gsi);) | |
241a2b9e | 1309 | { |
538dd0b7 | 1310 | gphi *stmt = gsi.phi (); |
ea057359 | 1311 | if (!virtual_operand_p (gimple_phi_result (stmt))) |
2e5e346d JL |
1312 | { |
1313 | gsi_next (&gsi); | |
1314 | continue; | |
1315 | } | |
6b8c9df8 RG |
1316 | mark_virtual_phi_result_for_renaming (stmt); |
1317 | remove_phi_node (&gsi, true); | |
cfef45c8 | 1318 | phi_p = true; |
241a2b9e | 1319 | } |
cfef45c8 RG |
1320 | /* In reality we have to rename the reaching definition of the |
1321 | virtual operand at return_bb as we will eventually release it | |
1322 | when we remove the code region we outlined. | |
1323 | So we have to rename all immediate virtual uses of that region | |
1324 | if we didn't see a PHI definition yet. */ | |
1325 | /* ??? In real reality we want to set the reaching vdef of the | |
1326 | entry of the SESE region as the vuse of the call and the reaching | |
1327 | vdef of the exit of the SESE region as the vdef of the call. */ | |
1328 | if (!phi_p) | |
538dd0b7 DM |
1329 | for (gimple_stmt_iterator gsi = gsi_start_bb (return_bb); |
1330 | !gsi_end_p (gsi); | |
1331 | gsi_next (&gsi)) | |
cfef45c8 | 1332 | { |
355fe088 | 1333 | gimple *stmt = gsi_stmt (gsi); |
cfef45c8 RG |
1334 | if (gimple_vuse (stmt)) |
1335 | { | |
1336 | gimple_set_vuse (stmt, NULL_TREE); | |
1337 | update_stmt (stmt); | |
1338 | } | |
1339 | if (gimple_vdef (stmt)) | |
1340 | break; | |
1341 | } | |
241a2b9e | 1342 | } |
3e485f62 JH |
1343 | |
1344 | /* Now create the actual clone. */ | |
3dafb85c | 1345 | cgraph_edge::rebuild_edges (); |
d52f5295 ML |
1346 | node = cur_node->create_version_clone_with_body |
1347 | (vNULL, NULL, args_to_skip, !split_part_return_p, split_point->split_bbs, | |
1348 | split_point->entry_bb, "part"); | |
9cec31f4 | 1349 | |
41f669d8 JH |
1350 | node->split_part = true; |
1351 | ||
9cec31f4 ML |
1352 | /* Let's take a time profile for splitted function. */ |
1353 | node->tp_first_run = cur_node->tp_first_run + 1; | |
1354 | ||
d402c33d JH |
1355 | /* For usual cloning it is enough to clear builtin only when signature |
1356 | changes. For partial inlining we however can not expect the part | |
1357 | of builtin implementation to have same semantic as the whole. */ | |
67348ccc | 1358 | if (DECL_BUILT_IN (node->decl)) |
d402c33d | 1359 | { |
67348ccc DM |
1360 | DECL_BUILT_IN_CLASS (node->decl) = NOT_BUILT_IN; |
1361 | DECL_FUNCTION_CODE (node->decl) = (enum built_in_function) 0; | |
d402c33d | 1362 | } |
d5e254e1 IE |
1363 | |
1364 | /* If the original function is instrumented then it's | |
1365 | part is also instrumented. */ | |
1366 | if (with_bounds) | |
1367 | chkp_function_mark_instrumented (node->decl); | |
1368 | ||
9a6af450 EB |
1369 | /* If the original function is declared inline, there is no point in issuing |
1370 | a warning for the non-inlinable part. */ | |
67348ccc | 1371 | DECL_NO_INLINE_WARNING_P (node->decl) = 1; |
d52f5295 | 1372 | cur_node->remove_callees (); |
d122681a | 1373 | cur_node->remove_all_references (); |
3e485f62 | 1374 | if (!split_part_return_p) |
67348ccc | 1375 | TREE_THIS_VOLATILE (node->decl) = 1; |
3e485f62 | 1376 | if (dump_file) |
67348ccc | 1377 | dump_function_to_file (node->decl, dump_file, dump_flags); |
3e485f62 JH |
1378 | |
1379 | /* Create the basic block we place call into. It is the entry basic block | |
1380 | split after last label. */ | |
1381 | call_bb = split_point->entry_bb; | |
538dd0b7 | 1382 | for (gimple_stmt_iterator gsi = gsi_start_bb (call_bb); !gsi_end_p (gsi);) |
3e485f62 JH |
1383 | if (gimple_code (gsi_stmt (gsi)) == GIMPLE_LABEL) |
1384 | { | |
1385 | last_stmt = gsi_stmt (gsi); | |
1386 | gsi_next (&gsi); | |
1387 | } | |
1388 | else | |
1389 | break; | |
1390 | e = split_block (split_point->entry_bb, last_stmt); | |
1391 | remove_edge (e); | |
1392 | ||
1393 | /* Produce the call statement. */ | |
538dd0b7 | 1394 | gimple_stmt_iterator gsi = gsi_last_bb (call_bb); |
9771b263 | 1395 | FOR_EACH_VEC_ELT (args_to_pass, i, arg) |
2b3c0885 RG |
1396 | if (!is_gimple_val (arg)) |
1397 | { | |
1398 | arg = force_gimple_operand_gsi (&gsi, arg, true, NULL_TREE, | |
f6e52e91 | 1399 | false, GSI_CONTINUE_LINKING); |
9771b263 | 1400 | args_to_pass[i] = arg; |
2b3c0885 | 1401 | } |
67348ccc | 1402 | call = gimple_build_call_vec (node->decl, args_to_pass); |
d5e254e1 | 1403 | gimple_call_set_with_bounds (call, with_bounds); |
3e485f62 | 1404 | gimple_set_block (call, DECL_INITIAL (current_function_decl)); |
9771b263 | 1405 | args_to_pass.release (); |
3e485f62 | 1406 | |
878eef4a JJ |
1407 | /* For optimized away parameters, add on the caller side |
1408 | before the call | |
1409 | DEBUG D#X => parm_Y(D) | |
1410 | stmts and associate D#X with parm in decl_debug_args_lookup | |
1411 | vector to say for debug info that if parameter parm had been passed, | |
1412 | it would have value parm_Y(D). */ | |
1413 | if (args_to_skip) | |
1414 | for (parm = DECL_ARGUMENTS (current_function_decl), num = 0; | |
1415 | parm; parm = DECL_CHAIN (parm), num++) | |
1416 | if (bitmap_bit_p (args_to_skip, num) | |
1417 | && is_gimple_reg (parm)) | |
1418 | { | |
1419 | tree ddecl; | |
355fe088 | 1420 | gimple *def_temp; |
878eef4a JJ |
1421 | |
1422 | /* This needs to be done even without MAY_HAVE_DEBUG_STMTS, | |
1423 | otherwise if it didn't exist before, we'd end up with | |
1424 | different SSA_NAME_VERSIONs between -g and -g0. */ | |
1425 | arg = get_or_create_ssa_default_def (cfun, parm); | |
1426 | if (!MAY_HAVE_DEBUG_STMTS) | |
1427 | continue; | |
1428 | ||
1429 | if (debug_args == NULL) | |
67348ccc | 1430 | debug_args = decl_debug_args_insert (node->decl); |
878eef4a JJ |
1431 | ddecl = make_node (DEBUG_EXPR_DECL); |
1432 | DECL_ARTIFICIAL (ddecl) = 1; | |
1433 | TREE_TYPE (ddecl) = TREE_TYPE (parm); | |
1434 | DECL_MODE (ddecl) = DECL_MODE (parm); | |
9771b263 DN |
1435 | vec_safe_push (*debug_args, DECL_ORIGIN (parm)); |
1436 | vec_safe_push (*debug_args, ddecl); | |
878eef4a JJ |
1437 | def_temp = gimple_build_debug_bind (ddecl, unshare_expr (arg), |
1438 | call); | |
1439 | gsi_insert_after (&gsi, def_temp, GSI_NEW_STMT); | |
1440 | } | |
1441 | /* And on the callee side, add | |
1442 | DEBUG D#Y s=> parm | |
1443 | DEBUG var => D#Y | |
1444 | stmts to the first bb where var is a VAR_DECL created for the | |
1445 | optimized away parameter in DECL_INITIAL block. This hints | |
1446 | in the debug info that var (whole DECL_ORIGIN is the parm PARM_DECL) | |
1447 | is optimized away, but could be looked up at the call site | |
1448 | as value of D#X there. */ | |
1449 | if (debug_args != NULL) | |
1450 | { | |
1451 | unsigned int i; | |
1452 | tree var, vexpr; | |
1453 | gimple_stmt_iterator cgsi; | |
355fe088 | 1454 | gimple *def_temp; |
878eef4a | 1455 | |
67348ccc DM |
1456 | push_cfun (DECL_STRUCT_FUNCTION (node->decl)); |
1457 | var = BLOCK_VARS (DECL_INITIAL (node->decl)); | |
9771b263 | 1458 | i = vec_safe_length (*debug_args); |
fefa31b5 | 1459 | cgsi = gsi_after_labels (single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun))); |
878eef4a JJ |
1460 | do |
1461 | { | |
1462 | i -= 2; | |
1463 | while (var != NULL_TREE | |
9771b263 | 1464 | && DECL_ABSTRACT_ORIGIN (var) != (**debug_args)[i]) |
878eef4a JJ |
1465 | var = TREE_CHAIN (var); |
1466 | if (var == NULL_TREE) | |
1467 | break; | |
1468 | vexpr = make_node (DEBUG_EXPR_DECL); | |
9771b263 | 1469 | parm = (**debug_args)[i]; |
878eef4a JJ |
1470 | DECL_ARTIFICIAL (vexpr) = 1; |
1471 | TREE_TYPE (vexpr) = TREE_TYPE (parm); | |
1472 | DECL_MODE (vexpr) = DECL_MODE (parm); | |
1473 | def_temp = gimple_build_debug_source_bind (vexpr, parm, | |
1474 | NULL); | |
1475 | gsi_insert_before (&cgsi, def_temp, GSI_SAME_STMT); | |
1476 | def_temp = gimple_build_debug_bind (var, vexpr, NULL); | |
1477 | gsi_insert_before (&cgsi, def_temp, GSI_SAME_STMT); | |
1478 | } | |
1479 | while (i); | |
1480 | pop_cfun (); | |
1481 | } | |
1482 | ||
556e9ba0 JH |
1483 | /* We avoid address being taken on any variable used by split part, |
1484 | so return slot optimization is always possible. Moreover this is | |
1485 | required to make DECL_BY_REFERENCE work. */ | |
1486 | if (aggregate_value_p (DECL_RESULT (current_function_decl), | |
22110e6c EB |
1487 | TREE_TYPE (current_function_decl)) |
1488 | && (!is_gimple_reg_type (TREE_TYPE (DECL_RESULT (current_function_decl))) | |
1489 | || DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)))) | |
556e9ba0 JH |
1490 | gimple_call_set_return_slot_opt (call, true); |
1491 | ||
bfd71482 JJ |
1492 | if (add_tsan_func_exit) |
1493 | tsan_func_exit_call = gimple_build_call_internal (IFN_TSAN_FUNC_EXIT, 0); | |
1494 | ||
3e485f62 JH |
1495 | /* Update return value. This is bit tricky. When we do not return, |
1496 | do nothing. When we return we might need to update return_bb | |
1497 | or produce a new return statement. */ | |
1498 | if (!split_part_return_p) | |
bfd71482 JJ |
1499 | { |
1500 | gsi_insert_after (&gsi, call, GSI_NEW_STMT); | |
1501 | if (tsan_func_exit_call) | |
1502 | gsi_insert_after (&gsi, tsan_func_exit_call, GSI_NEW_STMT); | |
1503 | } | |
3e485f62 JH |
1504 | else |
1505 | { | |
1506 | e = make_edge (call_bb, return_bb, | |
fefa31b5 DM |
1507 | return_bb == EXIT_BLOCK_PTR_FOR_FN (cfun) |
1508 | ? 0 : EDGE_FALLTHRU); | |
3e485f62 JH |
1509 | e->count = call_bb->count; |
1510 | e->probability = REG_BR_PROB_BASE; | |
6938f93f JH |
1511 | |
1512 | /* If there is return basic block, see what value we need to store | |
1513 | return value into and put call just before it. */ | |
fefa31b5 | 1514 | if (return_bb != EXIT_BLOCK_PTR_FOR_FN (cfun)) |
3e485f62 | 1515 | { |
2094f1fc | 1516 | real_retval = retval = find_retval (return_bb); |
d5e254e1 | 1517 | retbnd = find_retbnd (return_bb); |
6938f93f | 1518 | |
241a2b9e | 1519 | if (real_retval && split_point->split_part_set_retval) |
3e485f62 | 1520 | { |
538dd0b7 | 1521 | gphi_iterator psi; |
3e485f62 | 1522 | |
6938f93f JH |
1523 | /* See if we need new SSA_NAME for the result. |
1524 | When DECL_BY_REFERENCE is true, retval is actually pointer to | |
1525 | return value and it is constant in whole function. */ | |
1526 | if (TREE_CODE (retval) == SSA_NAME | |
1527 | && !DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) | |
3e485f62 | 1528 | { |
070ecdfd | 1529 | retval = copy_ssa_name (retval, call); |
6938f93f JH |
1530 | |
1531 | /* See if there is PHI defining return value. */ | |
1532 | for (psi = gsi_start_phis (return_bb); | |
1533 | !gsi_end_p (psi); gsi_next (&psi)) | |
538dd0b7 | 1534 | if (!virtual_operand_p (gimple_phi_result (psi.phi ()))) |
6938f93f JH |
1535 | break; |
1536 | ||
1537 | /* When there is PHI, just update its value. */ | |
3e485f62 JH |
1538 | if (TREE_CODE (retval) == SSA_NAME |
1539 | && !gsi_end_p (psi)) | |
538dd0b7 | 1540 | add_phi_arg (psi.phi (), retval, e, UNKNOWN_LOCATION); |
6938f93f JH |
1541 | /* Otherwise update the return BB itself. |
1542 | find_return_bb allows at most one assignment to return value, | |
1543 | so update first statement. */ | |
1544 | else | |
3e485f62 | 1545 | { |
2094f1fc JH |
1546 | gimple_stmt_iterator bsi; |
1547 | for (bsi = gsi_start_bb (return_bb); !gsi_end_p (bsi); | |
1548 | gsi_next (&bsi)) | |
538dd0b7 DM |
1549 | if (greturn *return_stmt |
1550 | = dyn_cast <greturn *> (gsi_stmt (bsi))) | |
2094f1fc | 1551 | { |
538dd0b7 | 1552 | gimple_return_set_retval (return_stmt, retval); |
2094f1fc JH |
1553 | break; |
1554 | } | |
2e216592 JJ |
1555 | else if (gimple_code (gsi_stmt (bsi)) == GIMPLE_ASSIGN |
1556 | && !gimple_clobber_p (gsi_stmt (bsi))) | |
2094f1fc JH |
1557 | { |
1558 | gimple_assign_set_rhs1 (gsi_stmt (bsi), retval); | |
1559 | break; | |
1560 | } | |
1561 | update_stmt (gsi_stmt (bsi)); | |
3e485f62 | 1562 | } |
d5e254e1 IE |
1563 | |
1564 | /* Replace retbnd with new one. */ | |
1565 | if (retbnd) | |
1566 | { | |
1567 | gimple_stmt_iterator bsi; | |
1568 | for (bsi = gsi_last_bb (return_bb); !gsi_end_p (bsi); | |
1569 | gsi_prev (&bsi)) | |
1570 | if (gimple_code (gsi_stmt (bsi)) == GIMPLE_RETURN) | |
1571 | { | |
1572 | retbnd = copy_ssa_name (retbnd, call); | |
1573 | gimple_return_set_retbnd (gsi_stmt (bsi), retbnd); | |
1574 | update_stmt (gsi_stmt (bsi)); | |
1575 | break; | |
1576 | } | |
1577 | } | |
3e485f62 | 1578 | } |
556e9ba0 | 1579 | if (DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) |
42b05b6e RG |
1580 | { |
1581 | gimple_call_set_lhs (call, build_simple_mem_ref (retval)); | |
1582 | gsi_insert_after (&gsi, call, GSI_NEW_STMT); | |
1583 | } | |
556e9ba0 | 1584 | else |
42b05b6e RG |
1585 | { |
1586 | tree restype; | |
1587 | restype = TREE_TYPE (DECL_RESULT (current_function_decl)); | |
1588 | gsi_insert_after (&gsi, call, GSI_NEW_STMT); | |
1589 | if (!useless_type_conversion_p (TREE_TYPE (retval), restype)) | |
1590 | { | |
355fe088 | 1591 | gimple *cpy; |
b731b390 | 1592 | tree tem = create_tmp_reg (restype); |
42b05b6e | 1593 | tem = make_ssa_name (tem, call); |
0d0e4a03 | 1594 | cpy = gimple_build_assign (retval, NOP_EXPR, tem); |
42b05b6e RG |
1595 | gsi_insert_after (&gsi, cpy, GSI_NEW_STMT); |
1596 | retval = tem; | |
1597 | } | |
d5e254e1 IE |
1598 | /* Build bndret call to obtain returned bounds. */ |
1599 | if (retbnd) | |
4f5a53cc | 1600 | chkp_insert_retbnd_call (retbnd, retval, &gsi); |
42b05b6e RG |
1601 | gimple_call_set_lhs (call, retval); |
1602 | update_stmt (call); | |
1603 | } | |
3e485f62 | 1604 | } |
42b05b6e RG |
1605 | else |
1606 | gsi_insert_after (&gsi, call, GSI_NEW_STMT); | |
bfd71482 JJ |
1607 | if (tsan_func_exit_call) |
1608 | gsi_insert_after (&gsi, tsan_func_exit_call, GSI_NEW_STMT); | |
3e485f62 | 1609 | } |
6938f93f JH |
1610 | /* We don't use return block (there is either no return in function or |
1611 | multiple of them). So create new basic block with return statement. | |
1612 | */ | |
3e485f62 JH |
1613 | else |
1614 | { | |
538dd0b7 | 1615 | greturn *ret; |
241a2b9e JH |
1616 | if (split_point->split_part_set_retval |
1617 | && !VOID_TYPE_P (TREE_TYPE (TREE_TYPE (current_function_decl)))) | |
3e485f62 | 1618 | { |
4021f4a1 | 1619 | retval = DECL_RESULT (current_function_decl); |
8a9c1ae6 | 1620 | |
d5e254e1 IE |
1621 | if (chkp_function_instrumented_p (current_function_decl) |
1622 | && BOUNDED_P (retval)) | |
b731b390 | 1623 | retbnd = create_tmp_reg (pointer_bounds_type_node); |
d5e254e1 | 1624 | |
8a9c1ae6 JH |
1625 | /* We use temporary register to hold value when aggregate_value_p |
1626 | is false. Similarly for DECL_BY_REFERENCE we must avoid extra | |
1627 | copy. */ | |
1628 | if (!aggregate_value_p (retval, TREE_TYPE (current_function_decl)) | |
1629 | && !DECL_BY_REFERENCE (retval)) | |
b731b390 | 1630 | retval = create_tmp_reg (TREE_TYPE (retval)); |
3e485f62 | 1631 | if (is_gimple_reg (retval)) |
6938f93f JH |
1632 | { |
1633 | /* When returning by reference, there is only one SSA name | |
1634 | assigned to RESULT_DECL (that is pointer to return value). | |
1635 | Look it up or create new one if it is missing. */ | |
1636 | if (DECL_BY_REFERENCE (retval)) | |
32244553 | 1637 | retval = get_or_create_ssa_default_def (cfun, retval); |
6938f93f JH |
1638 | /* Otherwise produce new SSA name for return value. */ |
1639 | else | |
1640 | retval = make_ssa_name (retval, call); | |
1641 | } | |
556e9ba0 JH |
1642 | if (DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) |
1643 | gimple_call_set_lhs (call, build_simple_mem_ref (retval)); | |
1644 | else | |
1645 | gimple_call_set_lhs (call, retval); | |
3e485f62 JH |
1646 | } |
1647 | gsi_insert_after (&gsi, call, GSI_NEW_STMT); | |
d5e254e1 IE |
1648 | /* Build bndret call to obtain returned bounds. */ |
1649 | if (retbnd) | |
4f5a53cc | 1650 | chkp_insert_retbnd_call (retbnd, retval, &gsi); |
bfd71482 JJ |
1651 | if (tsan_func_exit_call) |
1652 | gsi_insert_after (&gsi, tsan_func_exit_call, GSI_NEW_STMT); | |
3e485f62 JH |
1653 | ret = gimple_build_return (retval); |
1654 | gsi_insert_after (&gsi, ret, GSI_NEW_STMT); | |
1655 | } | |
1656 | } | |
1657 | free_dominance_info (CDI_DOMINATORS); | |
1658 | free_dominance_info (CDI_POST_DOMINATORS); | |
632b4f8e | 1659 | compute_inline_parameters (node, true); |
3e485f62 JH |
1660 | } |
1661 | ||
1662 | /* Execute function splitting pass. */ | |
1663 | ||
1664 | static unsigned int | |
1665 | execute_split_functions (void) | |
1666 | { | |
1667 | gimple_stmt_iterator bsi; | |
1668 | basic_block bb; | |
1669 | int overall_time = 0, overall_size = 0; | |
1670 | int todo = 0; | |
d52f5295 | 1671 | struct cgraph_node *node = cgraph_node::get (current_function_decl); |
3e485f62 | 1672 | |
b2d2adc6 RG |
1673 | if (flags_from_decl_or_type (current_function_decl) |
1674 | & (ECF_NORETURN|ECF_MALLOC)) | |
3e485f62 JH |
1675 | { |
1676 | if (dump_file) | |
b2d2adc6 | 1677 | fprintf (dump_file, "Not splitting: noreturn/malloc function.\n"); |
3e485f62 JH |
1678 | return 0; |
1679 | } | |
1680 | if (MAIN_NAME_P (DECL_NAME (current_function_decl))) | |
1681 | { | |
1682 | if (dump_file) | |
1683 | fprintf (dump_file, "Not splitting: main function.\n"); | |
1684 | return 0; | |
1685 | } | |
1686 | /* This can be relaxed; function might become inlinable after splitting | |
1687 | away the uninlinable part. */ | |
9771b263 | 1688 | if (inline_edge_summary_vec.exists () |
9a1e784a | 1689 | && !inline_summaries->get (node)->inlinable) |
3e485f62 JH |
1690 | { |
1691 | if (dump_file) | |
1692 | fprintf (dump_file, "Not splitting: not inlinable.\n"); | |
1693 | return 0; | |
1694 | } | |
67348ccc | 1695 | if (DECL_DISREGARD_INLINE_LIMITS (node->decl)) |
3e485f62 JH |
1696 | { |
1697 | if (dump_file) | |
ed7656f6 | 1698 | fprintf (dump_file, "Not splitting: disregarding inline limits.\n"); |
3e485f62 JH |
1699 | return 0; |
1700 | } | |
1701 | /* This can be relaxed; most of versioning tests actually prevents | |
1702 | a duplication. */ | |
1703 | if (!tree_versionable_function_p (current_function_decl)) | |
1704 | { | |
1705 | if (dump_file) | |
1706 | fprintf (dump_file, "Not splitting: not versionable.\n"); | |
1707 | return 0; | |
1708 | } | |
1709 | /* FIXME: we could support this. */ | |
1710 | if (DECL_STRUCT_FUNCTION (current_function_decl)->static_chain_decl) | |
1711 | { | |
1712 | if (dump_file) | |
1713 | fprintf (dump_file, "Not splitting: nested function.\n"); | |
1714 | return 0; | |
1715 | } | |
3e485f62 JH |
1716 | |
1717 | /* See if it makes sense to try to split. | |
1718 | It makes sense to split if we inline, that is if we have direct calls to | |
1719 | handle or direct calls are possibly going to appear as result of indirect | |
cf9712cc JH |
1720 | inlining or LTO. Also handle -fprofile-generate as LTO to allow non-LTO |
1721 | training for LTO -fprofile-use build. | |
1722 | ||
3e485f62 JH |
1723 | Note that we are not completely conservative about disqualifying functions |
1724 | called once. It is possible that the caller is called more then once and | |
1725 | then inlining would still benefit. */ | |
c91061e6 JH |
1726 | if ((!node->callers |
1727 | /* Local functions called once will be completely inlined most of time. */ | |
1728 | || (!node->callers->next_caller && node->local.local)) | |
67348ccc | 1729 | && !node->address_taken |
42685f72 | 1730 | && !node->has_aliases_p () |
67348ccc | 1731 | && (!flag_lto || !node->externally_visible)) |
3e485f62 JH |
1732 | { |
1733 | if (dump_file) | |
1734 | fprintf (dump_file, "Not splitting: not called directly " | |
1735 | "or called once.\n"); | |
1736 | return 0; | |
1737 | } | |
1738 | ||
1739 | /* FIXME: We can actually split if splitting reduces call overhead. */ | |
1740 | if (!flag_inline_small_functions | |
1741 | && !DECL_DECLARED_INLINE_P (current_function_decl)) | |
1742 | { | |
1743 | if (dump_file) | |
1744 | fprintf (dump_file, "Not splitting: not autoinlining and function" | |
1745 | " is not inline.\n"); | |
1746 | return 0; | |
1747 | } | |
1748 | ||
e70670cf JH |
1749 | /* We enforce splitting after loop headers when profile info is not |
1750 | available. */ | |
0a6a6ac9 | 1751 | if (profile_status_for_fn (cfun) != PROFILE_READ) |
e70670cf JH |
1752 | mark_dfs_back_edges (); |
1753 | ||
b2e25729 BS |
1754 | /* Initialize bitmap to track forbidden calls. */ |
1755 | forbidden_dominators = BITMAP_ALLOC (NULL); | |
1756 | calculate_dominance_info (CDI_DOMINATORS); | |
1757 | ||
3e485f62 | 1758 | /* Compute local info about basic blocks and determine function size/time. */ |
8b1c6fd7 | 1759 | bb_info_vec.safe_grow_cleared (last_basic_block_for_fn (cfun) + 1); |
3e485f62 | 1760 | memset (&best_split_point, 0, sizeof (best_split_point)); |
bfd71482 JJ |
1761 | basic_block return_bb = find_return_bb (); |
1762 | int tsan_exit_found = -1; | |
11cd3bed | 1763 | FOR_EACH_BB_FN (bb, cfun) |
3e485f62 JH |
1764 | { |
1765 | int time = 0; | |
1766 | int size = 0; | |
1767 | int freq = compute_call_stmt_bb_frequency (current_function_decl, bb); | |
1768 | ||
1769 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1770 | fprintf (dump_file, "Basic block %i\n", bb->index); | |
1771 | ||
1772 | for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
1773 | { | |
1774 | int this_time, this_size; | |
355fe088 | 1775 | gimple *stmt = gsi_stmt (bsi); |
3e485f62 JH |
1776 | |
1777 | this_size = estimate_num_insns (stmt, &eni_size_weights); | |
1778 | this_time = estimate_num_insns (stmt, &eni_time_weights) * freq; | |
1779 | size += this_size; | |
1780 | time += this_time; | |
b2e25729 | 1781 | check_forbidden_calls (stmt); |
3e485f62 JH |
1782 | |
1783 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1784 | { | |
1785 | fprintf (dump_file, " freq:%6i size:%3i time:%3i ", | |
1786 | freq, this_size, this_time); | |
1787 | print_gimple_stmt (dump_file, stmt, 0, 0); | |
1788 | } | |
bfd71482 JJ |
1789 | |
1790 | if ((flag_sanitize & SANITIZE_THREAD) | |
1791 | && is_gimple_call (stmt) | |
1792 | && gimple_call_internal_p (stmt) | |
1793 | && gimple_call_internal_fn (stmt) == IFN_TSAN_FUNC_EXIT) | |
1794 | { | |
1795 | /* We handle TSAN_FUNC_EXIT for splitting either in the | |
1796 | return_bb, or in its immediate predecessors. */ | |
1797 | if ((bb != return_bb && !find_edge (bb, return_bb)) | |
1798 | || (tsan_exit_found != -1 | |
1799 | && tsan_exit_found != (bb != return_bb))) | |
1800 | { | |
1801 | if (dump_file) | |
1802 | fprintf (dump_file, "Not splitting: TSAN_FUNC_EXIT" | |
1803 | " in unexpected basic block.\n"); | |
1804 | BITMAP_FREE (forbidden_dominators); | |
1805 | bb_info_vec.release (); | |
1806 | return 0; | |
1807 | } | |
1808 | tsan_exit_found = bb != return_bb; | |
1809 | } | |
3e485f62 JH |
1810 | } |
1811 | overall_time += time; | |
1812 | overall_size += size; | |
9771b263 DN |
1813 | bb_info_vec[bb->index].time = time; |
1814 | bb_info_vec[bb->index].size = size; | |
3e485f62 | 1815 | } |
bfd71482 | 1816 | find_split_points (return_bb, overall_time, overall_size); |
3e485f62 JH |
1817 | if (best_split_point.split_bbs) |
1818 | { | |
bfd71482 | 1819 | split_function (return_bb, &best_split_point, tsan_exit_found == 1); |
3e485f62 JH |
1820 | BITMAP_FREE (best_split_point.ssa_names_to_pass); |
1821 | BITMAP_FREE (best_split_point.split_bbs); | |
1822 | todo = TODO_update_ssa | TODO_cleanup_cfg; | |
1823 | } | |
b2e25729 | 1824 | BITMAP_FREE (forbidden_dominators); |
9771b263 | 1825 | bb_info_vec.release (); |
3e485f62 JH |
1826 | return todo; |
1827 | } | |
1828 | ||
27a4cd48 DM |
1829 | namespace { |
1830 | ||
1831 | const pass_data pass_data_split_functions = | |
3e485f62 | 1832 | { |
27a4cd48 DM |
1833 | GIMPLE_PASS, /* type */ |
1834 | "fnsplit", /* name */ | |
1835 | OPTGROUP_NONE, /* optinfo_flags */ | |
27a4cd48 DM |
1836 | TV_IPA_FNSPLIT, /* tv_id */ |
1837 | PROP_cfg, /* properties_required */ | |
1838 | 0, /* properties_provided */ | |
1839 | 0, /* properties_destroyed */ | |
1840 | 0, /* todo_flags_start */ | |
3bea341f | 1841 | 0, /* todo_flags_finish */ |
3e485f62 | 1842 | }; |
cf9712cc | 1843 | |
27a4cd48 DM |
1844 | class pass_split_functions : public gimple_opt_pass |
1845 | { | |
1846 | public: | |
c3284718 RS |
1847 | pass_split_functions (gcc::context *ctxt) |
1848 | : gimple_opt_pass (pass_data_split_functions, ctxt) | |
27a4cd48 DM |
1849 | {} |
1850 | ||
1851 | /* opt_pass methods: */ | |
1a3d085c | 1852 | virtual bool gate (function *); |
be55bfe6 TS |
1853 | virtual unsigned int execute (function *) |
1854 | { | |
1855 | return execute_split_functions (); | |
1856 | } | |
27a4cd48 DM |
1857 | |
1858 | }; // class pass_split_functions | |
1859 | ||
1a3d085c TS |
1860 | bool |
1861 | pass_split_functions::gate (function *) | |
1862 | { | |
1863 | /* When doing profile feedback, we want to execute the pass after profiling | |
1864 | is read. So disable one in early optimization. */ | |
1865 | return (flag_partial_inlining | |
1866 | && !profile_arc_flag && !flag_branch_probabilities); | |
1867 | } | |
1868 | ||
27a4cd48 DM |
1869 | } // anon namespace |
1870 | ||
1871 | gimple_opt_pass * | |
1872 | make_pass_split_functions (gcc::context *ctxt) | |
1873 | { | |
1874 | return new pass_split_functions (ctxt); | |
1875 | } | |
1876 | ||
cf9712cc JH |
1877 | /* Execute function splitting pass. */ |
1878 | ||
1879 | static unsigned int | |
1880 | execute_feedback_split_functions (void) | |
1881 | { | |
1882 | unsigned int retval = execute_split_functions (); | |
1883 | if (retval) | |
1884 | retval |= TODO_rebuild_cgraph_edges; | |
1885 | return retval; | |
1886 | } | |
1887 | ||
27a4cd48 DM |
1888 | namespace { |
1889 | ||
1890 | const pass_data pass_data_feedback_split_functions = | |
cf9712cc | 1891 | { |
27a4cd48 DM |
1892 | GIMPLE_PASS, /* type */ |
1893 | "feedback_fnsplit", /* name */ | |
1894 | OPTGROUP_NONE, /* optinfo_flags */ | |
27a4cd48 DM |
1895 | TV_IPA_FNSPLIT, /* tv_id */ |
1896 | PROP_cfg, /* properties_required */ | |
1897 | 0, /* properties_provided */ | |
1898 | 0, /* properties_destroyed */ | |
1899 | 0, /* todo_flags_start */ | |
3bea341f | 1900 | 0, /* todo_flags_finish */ |
cf9712cc | 1901 | }; |
27a4cd48 DM |
1902 | |
1903 | class pass_feedback_split_functions : public gimple_opt_pass | |
1904 | { | |
1905 | public: | |
c3284718 RS |
1906 | pass_feedback_split_functions (gcc::context *ctxt) |
1907 | : gimple_opt_pass (pass_data_feedback_split_functions, ctxt) | |
27a4cd48 DM |
1908 | {} |
1909 | ||
1910 | /* opt_pass methods: */ | |
1a3d085c | 1911 | virtual bool gate (function *); |
be55bfe6 TS |
1912 | virtual unsigned int execute (function *) |
1913 | { | |
1914 | return execute_feedback_split_functions (); | |
1915 | } | |
27a4cd48 DM |
1916 | |
1917 | }; // class pass_feedback_split_functions | |
1918 | ||
1a3d085c TS |
1919 | bool |
1920 | pass_feedback_split_functions::gate (function *) | |
1921 | { | |
1922 | /* We don't need to split when profiling at all, we are producing | |
1923 | lousy code anyway. */ | |
1924 | return (flag_partial_inlining | |
1925 | && flag_branch_probabilities); | |
1926 | } | |
1927 | ||
27a4cd48 DM |
1928 | } // anon namespace |
1929 | ||
1930 | gimple_opt_pass * | |
1931 | make_pass_feedback_split_functions (gcc::context *ctxt) | |
1932 | { | |
1933 | return new pass_feedback_split_functions (ctxt); | |
1934 | } |