]>
Commit | Line | Data |
---|---|---|
3e485f62 | 1 | /* Function splitting pass |
cbe34bb5 | 2 | Copyright (C) 2010-2017 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" | |
27d020cf | 105 | #include "ipa-fnsummary.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; | |
c2e87766 | 114 | sreal 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. */ | |
c2e87766 JH |
124 | sreal header_time, split_time; |
125 | unsigned int header_size, split_size; | |
3e485f62 | 126 | |
ed7656f6 | 127 | /* SSA names that need to be passed into spit function. */ |
3e485f62 JH |
128 | bitmap ssa_names_to_pass; |
129 | ||
130 | /* Basic block where we split (that will become entry point of new function. */ | |
131 | basic_block entry_bb; | |
132 | ||
ed10d09b JH |
133 | /* Count for entering the split part. |
134 | This is not count of the entry_bb because it may be in loop. */ | |
135 | profile_count count; | |
136 | ||
3e485f62 JH |
137 | /* Basic blocks we are splitting away. */ |
138 | bitmap split_bbs; | |
241a2b9e JH |
139 | |
140 | /* True when return value is computed on split part and thus it needs | |
141 | to be returned. */ | |
142 | bool split_part_set_retval; | |
3e485f62 JH |
143 | }; |
144 | ||
145 | /* Best split point found. */ | |
146 | ||
147 | struct split_point best_split_point; | |
148 | ||
b2e25729 BS |
149 | /* Set of basic blocks that are not allowed to dominate a split point. */ |
150 | ||
151 | static bitmap forbidden_dominators; | |
152 | ||
241a2b9e | 153 | static tree find_retval (basic_block return_bb); |
d5e254e1 | 154 | static tree find_retbnd (basic_block return_bb); |
241a2b9e | 155 | |
1802378d | 156 | /* Callback for walk_stmt_load_store_addr_ops. If T is non-SSA automatic |
3e485f62 JH |
157 | variable, check it if it is present in bitmap passed via DATA. */ |
158 | ||
159 | static bool | |
355fe088 | 160 | test_nonssa_use (gimple *, tree t, tree, void *data) |
3e485f62 JH |
161 | { |
162 | t = get_base_address (t); | |
163 | ||
1802378d EB |
164 | if (!t || is_gimple_reg (t)) |
165 | return false; | |
166 | ||
167 | if (TREE_CODE (t) == PARM_DECL | |
8813a647 | 168 | || (VAR_P (t) |
3e485f62 | 169 | && auto_var_in_fn_p (t, current_function_decl)) |
1802378d | 170 | || TREE_CODE (t) == RESULT_DECL |
412c4af7 JH |
171 | /* Normal labels are part of CFG and will be handled gratefuly. |
172 | Forced labels however can be used directly by statements and | |
173 | need to stay in one partition along with their uses. */ | |
174 | || (TREE_CODE (t) == LABEL_DECL | |
175 | && FORCED_LABEL (t))) | |
3e485f62 | 176 | return bitmap_bit_p ((bitmap)data, DECL_UID (t)); |
241a2b9e | 177 | |
1802378d EB |
178 | /* For DECL_BY_REFERENCE, the return value is actually a pointer. We want |
179 | to pretend that the value pointed to is actual result decl. */ | |
180 | if ((TREE_CODE (t) == MEM_REF || INDIRECT_REF_P (t)) | |
241a2b9e | 181 | && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME |
70b5e7dc | 182 | && SSA_NAME_VAR (TREE_OPERAND (t, 0)) |
241a2b9e JH |
183 | && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND (t, 0))) == RESULT_DECL |
184 | && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) | |
1802378d EB |
185 | return |
186 | bitmap_bit_p ((bitmap)data, | |
187 | DECL_UID (DECL_RESULT (current_function_decl))); | |
188 | ||
3e485f62 JH |
189 | return false; |
190 | } | |
191 | ||
192 | /* Dump split point CURRENT. */ | |
193 | ||
194 | static void | |
195 | dump_split_point (FILE * file, struct split_point *current) | |
196 | { | |
197 | fprintf (file, | |
cfef45c8 | 198 | "Split point at BB %i\n" |
c2e87766 JH |
199 | " header time: %f header size: %i\n" |
200 | " split time: %f split size: %i\n bbs: ", | |
201 | current->entry_bb->index, current->header_time.to_double (), | |
202 | current->header_size, current->split_time.to_double (), | |
203 | current->split_size); | |
3e485f62 JH |
204 | dump_bitmap (file, current->split_bbs); |
205 | fprintf (file, " SSA names to pass: "); | |
206 | dump_bitmap (file, current->ssa_names_to_pass); | |
207 | } | |
208 | ||
1802378d EB |
209 | /* Look for all BBs in header that might lead to the split part and verify |
210 | that they are not defining any non-SSA var used by the split part. | |
2094f1fc JH |
211 | Parameters are the same as for consider_split. */ |
212 | ||
213 | static bool | |
214 | verify_non_ssa_vars (struct split_point *current, bitmap non_ssa_vars, | |
215 | basic_block return_bb) | |
216 | { | |
217 | bitmap seen = BITMAP_ALLOC (NULL); | |
6e1aa848 | 218 | vec<basic_block> worklist = vNULL; |
2094f1fc JH |
219 | edge e; |
220 | edge_iterator ei; | |
221 | bool ok = true; | |
412c4af7 | 222 | basic_block bb; |
1802378d | 223 | |
2094f1fc | 224 | FOR_EACH_EDGE (e, ei, current->entry_bb->preds) |
fefa31b5 | 225 | if (e->src != ENTRY_BLOCK_PTR_FOR_FN (cfun) |
2094f1fc JH |
226 | && !bitmap_bit_p (current->split_bbs, e->src->index)) |
227 | { | |
9771b263 | 228 | worklist.safe_push (e->src); |
2094f1fc JH |
229 | bitmap_set_bit (seen, e->src->index); |
230 | } | |
1802378d | 231 | |
9771b263 | 232 | while (!worklist.is_empty ()) |
2094f1fc | 233 | { |
412c4af7 | 234 | bb = worklist.pop (); |
2094f1fc | 235 | FOR_EACH_EDGE (e, ei, bb->preds) |
fefa31b5 | 236 | if (e->src != ENTRY_BLOCK_PTR_FOR_FN (cfun) |
fcaa4ca4 | 237 | && bitmap_set_bit (seen, e->src->index)) |
2094f1fc JH |
238 | { |
239 | gcc_checking_assert (!bitmap_bit_p (current->split_bbs, | |
240 | e->src->index)); | |
9771b263 | 241 | worklist.safe_push (e->src); |
2094f1fc | 242 | } |
538dd0b7 DM |
243 | for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi); |
244 | gsi_next (&bsi)) | |
2094f1fc | 245 | { |
355fe088 | 246 | gimple *stmt = gsi_stmt (bsi); |
1802378d | 247 | if (is_gimple_debug (stmt)) |
2094f1fc JH |
248 | continue; |
249 | if (walk_stmt_load_store_addr_ops | |
1802378d EB |
250 | (stmt, non_ssa_vars, test_nonssa_use, test_nonssa_use, |
251 | test_nonssa_use)) | |
2094f1fc JH |
252 | { |
253 | ok = false; | |
254 | goto done; | |
255 | } | |
538dd0b7 DM |
256 | if (glabel *label_stmt = dyn_cast <glabel *> (stmt)) |
257 | if (test_nonssa_use (stmt, gimple_label_label (label_stmt), | |
258 | NULL_TREE, non_ssa_vars)) | |
259 | { | |
260 | ok = false; | |
261 | goto done; | |
262 | } | |
2094f1fc | 263 | } |
538dd0b7 DM |
264 | for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi); |
265 | gsi_next (&bsi)) | |
2094f1fc JH |
266 | { |
267 | if (walk_stmt_load_store_addr_ops | |
1802378d EB |
268 | (gsi_stmt (bsi), non_ssa_vars, test_nonssa_use, test_nonssa_use, |
269 | test_nonssa_use)) | |
2094f1fc JH |
270 | { |
271 | ok = false; | |
272 | goto done; | |
273 | } | |
274 | } | |
275 | FOR_EACH_EDGE (e, ei, bb->succs) | |
276 | { | |
277 | if (e->dest != return_bb) | |
278 | continue; | |
538dd0b7 DM |
279 | for (gphi_iterator bsi = gsi_start_phis (return_bb); |
280 | !gsi_end_p (bsi); | |
2094f1fc JH |
281 | gsi_next (&bsi)) |
282 | { | |
538dd0b7 | 283 | gphi *stmt = bsi.phi (); |
2094f1fc JH |
284 | tree op = gimple_phi_arg_def (stmt, e->dest_idx); |
285 | ||
ea057359 | 286 | if (virtual_operand_p (gimple_phi_result (stmt))) |
2094f1fc JH |
287 | continue; |
288 | if (TREE_CODE (op) != SSA_NAME | |
9f1363cd | 289 | && test_nonssa_use (stmt, op, op, non_ssa_vars)) |
2094f1fc JH |
290 | { |
291 | ok = false; | |
292 | goto done; | |
293 | } | |
294 | } | |
295 | } | |
296 | } | |
412c4af7 JH |
297 | |
298 | /* Verify that the rest of function does not define any label | |
299 | used by the split part. */ | |
300 | FOR_EACH_BB_FN (bb, cfun) | |
301 | if (!bitmap_bit_p (current->split_bbs, bb->index) | |
302 | && !bitmap_bit_p (seen, bb->index)) | |
303 | { | |
304 | gimple_stmt_iterator bsi; | |
305 | for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
538dd0b7 | 306 | if (glabel *label_stmt = dyn_cast <glabel *> (gsi_stmt (bsi))) |
412c4af7 | 307 | { |
538dd0b7 DM |
308 | if (test_nonssa_use (label_stmt, |
309 | gimple_label_label (label_stmt), | |
310 | NULL_TREE, non_ssa_vars)) | |
311 | { | |
312 | ok = false; | |
313 | goto done; | |
314 | } | |
412c4af7 | 315 | } |
538dd0b7 | 316 | else |
412c4af7 JH |
317 | break; |
318 | } | |
319 | ||
2094f1fc JH |
320 | done: |
321 | BITMAP_FREE (seen); | |
9771b263 | 322 | worklist.release (); |
2094f1fc JH |
323 | return ok; |
324 | } | |
325 | ||
b2e25729 BS |
326 | /* If STMT is a call, check the callee against a list of forbidden |
327 | predicate functions. If a match is found, look for uses of the | |
328 | call result in condition statements that compare against zero. | |
329 | For each such use, find the block targeted by the condition | |
330 | statement for the nonzero result, and set the bit for this block | |
331 | in the forbidden dominators bitmap. The purpose of this is to avoid | |
332 | selecting a split point where we are likely to lose the chance | |
333 | to optimize away an unused function call. */ | |
334 | ||
335 | static void | |
355fe088 | 336 | check_forbidden_calls (gimple *stmt) |
b2e25729 BS |
337 | { |
338 | imm_use_iterator use_iter; | |
339 | use_operand_p use_p; | |
340 | tree lhs; | |
341 | ||
342 | /* At the moment, __builtin_constant_p is the only forbidden | |
343 | predicate function call (see PR49642). */ | |
344 | if (!gimple_call_builtin_p (stmt, BUILT_IN_CONSTANT_P)) | |
345 | return; | |
346 | ||
347 | lhs = gimple_call_lhs (stmt); | |
348 | ||
349 | if (!lhs || TREE_CODE (lhs) != SSA_NAME) | |
350 | return; | |
351 | ||
352 | FOR_EACH_IMM_USE_FAST (use_p, use_iter, lhs) | |
353 | { | |
354 | tree op1; | |
355 | basic_block use_bb, forbidden_bb; | |
356 | enum tree_code code; | |
357 | edge true_edge, false_edge; | |
538dd0b7 | 358 | gcond *use_stmt; |
b2e25729 | 359 | |
538dd0b7 DM |
360 | use_stmt = dyn_cast <gcond *> (USE_STMT (use_p)); |
361 | if (!use_stmt) | |
b2e25729 BS |
362 | continue; |
363 | ||
364 | /* Assuming canonical form for GIMPLE_COND here, with constant | |
365 | in second position. */ | |
366 | op1 = gimple_cond_rhs (use_stmt); | |
367 | code = gimple_cond_code (use_stmt); | |
368 | use_bb = gimple_bb (use_stmt); | |
369 | ||
370 | extract_true_false_edges_from_block (use_bb, &true_edge, &false_edge); | |
371 | ||
372 | /* We're only interested in comparisons that distinguish | |
373 | unambiguously from zero. */ | |
374 | if (!integer_zerop (op1) || code == LE_EXPR || code == GE_EXPR) | |
375 | continue; | |
376 | ||
377 | if (code == EQ_EXPR) | |
378 | forbidden_bb = false_edge->dest; | |
379 | else | |
380 | forbidden_bb = true_edge->dest; | |
381 | ||
382 | bitmap_set_bit (forbidden_dominators, forbidden_bb->index); | |
383 | } | |
384 | } | |
385 | ||
386 | /* If BB is dominated by any block in the forbidden dominators set, | |
387 | return TRUE; else FALSE. */ | |
388 | ||
389 | static bool | |
390 | dominated_by_forbidden (basic_block bb) | |
391 | { | |
392 | unsigned dom_bb; | |
393 | bitmap_iterator bi; | |
394 | ||
395 | EXECUTE_IF_SET_IN_BITMAP (forbidden_dominators, 1, dom_bb, bi) | |
396 | { | |
06e28de2 DM |
397 | if (dominated_by_p (CDI_DOMINATORS, bb, |
398 | BASIC_BLOCK_FOR_FN (cfun, dom_bb))) | |
b2e25729 BS |
399 | return true; |
400 | } | |
401 | ||
402 | return false; | |
403 | } | |
404 | ||
d5e254e1 IE |
405 | /* For give split point CURRENT and return block RETURN_BB return 1 |
406 | if ssa name VAL is set by split part and 0 otherwise. */ | |
407 | static bool | |
408 | split_part_set_ssa_name_p (tree val, struct split_point *current, | |
409 | basic_block return_bb) | |
410 | { | |
411 | if (TREE_CODE (val) != SSA_NAME) | |
412 | return false; | |
413 | ||
414 | return (!SSA_NAME_IS_DEFAULT_DEF (val) | |
415 | && (bitmap_bit_p (current->split_bbs, | |
416 | gimple_bb (SSA_NAME_DEF_STMT (val))->index) | |
417 | || gimple_bb (SSA_NAME_DEF_STMT (val)) == return_bb)); | |
418 | } | |
419 | ||
3e485f62 JH |
420 | /* We found an split_point CURRENT. NON_SSA_VARS is bitmap of all non ssa |
421 | variables used and RETURN_BB is return basic block. | |
422 | See if we can split function here. */ | |
423 | ||
424 | static void | |
425 | consider_split (struct split_point *current, bitmap non_ssa_vars, | |
426 | basic_block return_bb) | |
427 | { | |
428 | tree parm; | |
429 | unsigned int num_args = 0; | |
430 | unsigned int call_overhead; | |
431 | edge e; | |
432 | edge_iterator ei; | |
538dd0b7 | 433 | gphi_iterator bsi; |
8b3057b3 | 434 | unsigned int i; |
241a2b9e | 435 | tree retval; |
d5e254e1 | 436 | tree retbnd; |
e70670cf | 437 | bool back_edge = false; |
8b3057b3 | 438 | |
3e485f62 JH |
439 | if (dump_file && (dump_flags & TDF_DETAILS)) |
440 | dump_split_point (dump_file, current); | |
441 | ||
ed10d09b | 442 | current->count = profile_count::zero (); |
8b3057b3 | 443 | FOR_EACH_EDGE (e, ei, current->entry_bb->preds) |
e70670cf JH |
444 | { |
445 | if (e->flags & EDGE_DFS_BACK) | |
446 | back_edge = true; | |
447 | if (!bitmap_bit_p (current->split_bbs, e->src->index)) | |
ed10d09b | 448 | current->count += e->count (); |
e70670cf | 449 | } |
8b3057b3 | 450 | |
ed10d09b JH |
451 | /* Do not split when we would end up calling function anyway. |
452 | Compares are three state, use !(...<...) to also give up when outcome | |
453 | is unknown. */ | |
454 | if (!(current->count | |
455 | < (ENTRY_BLOCK_PTR_FOR_FN (cfun)->count.apply_scale | |
456 | (PARAM_VALUE (PARAM_PARTIAL_INLINING_ENTRY_PROBABILITY), 100)))) | |
3e485f62 | 457 | { |
e70670cf JH |
458 | /* When profile is guessed, we can not expect it to give us |
459 | realistic estimate on likelyness of function taking the | |
460 | complex path. As a special case, when tail of the function is | |
461 | a loop, enable splitting since inlining code skipping the loop | |
462 | is likely noticeable win. */ | |
463 | if (back_edge | |
0a6a6ac9 | 464 | && profile_status_for_fn (cfun) != PROFILE_READ |
ed10d09b JH |
465 | && current->count |
466 | < ENTRY_BLOCK_PTR_FOR_FN (cfun)->count) | |
e70670cf JH |
467 | { |
468 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
ed10d09b JH |
469 | { |
470 | fprintf (dump_file, | |
471 | " Split before loop, accepting despite low counts"); | |
472 | current->count.dump (dump_file); | |
473 | fprintf (dump_file, " "); | |
474 | ENTRY_BLOCK_PTR_FOR_FN (cfun)->count.dump (dump_file); | |
475 | } | |
e70670cf JH |
476 | } |
477 | else | |
478 | { | |
479 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
480 | fprintf (dump_file, | |
481 | " Refused: incoming frequency is too large.\n"); | |
482 | return; | |
483 | } | |
3e485f62 JH |
484 | } |
485 | ||
486 | if (!current->header_size) | |
487 | { | |
488 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
489 | fprintf (dump_file, " Refused: header empty\n"); | |
3e485f62 JH |
490 | return; |
491 | } | |
492 | ||
ed7656f6 JJ |
493 | /* Verify that PHI args on entry are either virtual or all their operands |
494 | incoming from header are the same. */ | |
8b3057b3 | 495 | for (bsi = gsi_start_phis (current->entry_bb); !gsi_end_p (bsi); gsi_next (&bsi)) |
3e485f62 | 496 | { |
538dd0b7 | 497 | gphi *stmt = bsi.phi (); |
8b3057b3 JH |
498 | tree val = NULL; |
499 | ||
ea057359 | 500 | if (virtual_operand_p (gimple_phi_result (stmt))) |
8b3057b3 JH |
501 | continue; |
502 | for (i = 0; i < gimple_phi_num_args (stmt); i++) | |
503 | { | |
504 | edge e = gimple_phi_arg_edge (stmt, i); | |
505 | if (!bitmap_bit_p (current->split_bbs, e->src->index)) | |
506 | { | |
507 | tree edge_val = gimple_phi_arg_def (stmt, i); | |
508 | if (val && edge_val != val) | |
509 | { | |
510 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
511 | fprintf (dump_file, | |
512 | " Refused: entry BB has PHI with multiple variants\n"); | |
513 | return; | |
514 | } | |
515 | val = edge_val; | |
516 | } | |
517 | } | |
3e485f62 JH |
518 | } |
519 | ||
520 | ||
521 | /* See what argument we will pass to the split function and compute | |
522 | call overhead. */ | |
523 | call_overhead = eni_size_weights.call_cost; | |
524 | for (parm = DECL_ARGUMENTS (current_function_decl); parm; | |
910ad8de | 525 | parm = DECL_CHAIN (parm)) |
3e485f62 JH |
526 | { |
527 | if (!is_gimple_reg (parm)) | |
528 | { | |
529 | if (bitmap_bit_p (non_ssa_vars, DECL_UID (parm))) | |
530 | { | |
531 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
532 | fprintf (dump_file, | |
533 | " Refused: need to pass non-ssa param values\n"); | |
534 | return; | |
535 | } | |
536 | } | |
32244553 | 537 | else |
3e485f62 | 538 | { |
32244553 RG |
539 | tree ddef = ssa_default_def (cfun, parm); |
540 | if (ddef | |
541 | && bitmap_bit_p (current->ssa_names_to_pass, | |
542 | SSA_NAME_VERSION (ddef))) | |
543 | { | |
544 | if (!VOID_TYPE_P (TREE_TYPE (parm))) | |
b4c9af96 | 545 | call_overhead += estimate_move_cost (TREE_TYPE (parm), false); |
32244553 RG |
546 | num_args++; |
547 | } | |
3e485f62 JH |
548 | } |
549 | } | |
550 | if (!VOID_TYPE_P (TREE_TYPE (current_function_decl))) | |
b4c9af96 RB |
551 | call_overhead += estimate_move_cost (TREE_TYPE (current_function_decl), |
552 | false); | |
3e485f62 JH |
553 | |
554 | if (current->split_size <= call_overhead) | |
555 | { | |
556 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
557 | fprintf (dump_file, | |
558 | " Refused: split size is smaller than call overhead\n"); | |
559 | return; | |
560 | } | |
561 | if (current->header_size + call_overhead | |
562 | >= (unsigned int)(DECL_DECLARED_INLINE_P (current_function_decl) | |
563 | ? MAX_INLINE_INSNS_SINGLE | |
564 | : MAX_INLINE_INSNS_AUTO)) | |
565 | { | |
566 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
567 | fprintf (dump_file, | |
568 | " Refused: header size is too large for inline candidate\n"); | |
569 | return; | |
570 | } | |
571 | ||
7a161d5b JH |
572 | /* Splitting functions brings the target out of comdat group; this will |
573 | lead to code duplication if the function is reused by other unit. | |
574 | Limit this duplication. This is consistent with limit in tree-sra.c | |
575 | FIXME: with LTO we ought to be able to do better! */ | |
576 | if (DECL_ONE_ONLY (current_function_decl) | |
577 | && current->split_size >= (unsigned int) MAX_INLINE_INSNS_AUTO) | |
578 | { | |
579 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
580 | fprintf (dump_file, | |
581 | " Refused: function is COMDAT and tail is too large\n"); | |
582 | return; | |
583 | } | |
584 | /* For comdat functions also reject very small tails; those will likely get | |
585 | inlined back and we do not want to risk the duplication overhead. | |
586 | FIXME: with LTO we ought to be able to do better! */ | |
587 | if (DECL_ONE_ONLY (current_function_decl) | |
588 | && current->split_size | |
589 | <= (unsigned int) PARAM_VALUE (PARAM_EARLY_INLINING_INSNS) / 2) | |
590 | { | |
591 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
592 | fprintf (dump_file, | |
593 | " Refused: function is COMDAT and tail is too small\n"); | |
594 | return; | |
595 | } | |
596 | ||
3e485f62 | 597 | /* FIXME: we currently can pass only SSA function parameters to the split |
d402c33d | 598 | arguments. Once parm_adjustment infrastructure is supported by cloning, |
3e485f62 JH |
599 | we can pass more than that. */ |
600 | if (num_args != bitmap_count_bits (current->ssa_names_to_pass)) | |
601 | { | |
8b3057b3 | 602 | |
3e485f62 JH |
603 | if (dump_file && (dump_flags & TDF_DETAILS)) |
604 | fprintf (dump_file, | |
605 | " Refused: need to pass non-param values\n"); | |
606 | return; | |
607 | } | |
608 | ||
609 | /* When there are non-ssa vars used in the split region, see if they | |
610 | are used in the header region. If so, reject the split. | |
611 | FIXME: we can use nested function support to access both. */ | |
2094f1fc JH |
612 | if (!bitmap_empty_p (non_ssa_vars) |
613 | && !verify_non_ssa_vars (current, non_ssa_vars, return_bb)) | |
3e485f62 | 614 | { |
2094f1fc JH |
615 | if (dump_file && (dump_flags & TDF_DETAILS)) |
616 | fprintf (dump_file, | |
617 | " Refused: split part has non-ssa uses\n"); | |
3e485f62 JH |
618 | return; |
619 | } | |
b2e25729 BS |
620 | |
621 | /* If the split point is dominated by a forbidden block, reject | |
622 | the split. */ | |
623 | if (!bitmap_empty_p (forbidden_dominators) | |
624 | && dominated_by_forbidden (current->entry_bb)) | |
625 | { | |
626 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
627 | fprintf (dump_file, | |
628 | " Refused: split point dominated by forbidden block\n"); | |
629 | return; | |
630 | } | |
631 | ||
241a2b9e JH |
632 | /* See if retval used by return bb is computed by header or split part. |
633 | When it is computed by split part, we need to produce return statement | |
634 | in the split part and add code to header to pass it around. | |
635 | ||
636 | This is bit tricky to test: | |
637 | 1) When there is no return_bb or no return value, we always pass | |
638 | value around. | |
639 | 2) Invariants are always computed by caller. | |
640 | 3) For SSA we need to look if defining statement is in header or split part | |
641 | 4) For non-SSA we need to look where the var is computed. */ | |
642 | retval = find_retval (return_bb); | |
643 | if (!retval) | |
06ff7000 JJ |
644 | { |
645 | /* If there is a return_bb with no return value in function returning | |
646 | value by reference, also make the split part return void, otherwise | |
647 | we expansion would try to create a non-POD temporary, which is | |
648 | invalid. */ | |
649 | if (return_bb != EXIT_BLOCK_PTR_FOR_FN (cfun) | |
650 | && DECL_RESULT (current_function_decl) | |
651 | && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) | |
652 | current->split_part_set_retval = false; | |
653 | else | |
654 | current->split_part_set_retval = true; | |
655 | } | |
241a2b9e JH |
656 | else if (is_gimple_min_invariant (retval)) |
657 | current->split_part_set_retval = false; | |
658 | /* Special case is value returned by reference we record as if it was non-ssa | |
659 | set to result_decl. */ | |
660 | else if (TREE_CODE (retval) == SSA_NAME | |
70b5e7dc | 661 | && SSA_NAME_VAR (retval) |
241a2b9e JH |
662 | && TREE_CODE (SSA_NAME_VAR (retval)) == RESULT_DECL |
663 | && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) | |
664 | current->split_part_set_retval | |
665 | = bitmap_bit_p (non_ssa_vars, DECL_UID (SSA_NAME_VAR (retval))); | |
666 | else if (TREE_CODE (retval) == SSA_NAME) | |
667 | current->split_part_set_retval | |
d5e254e1 | 668 | = split_part_set_ssa_name_p (retval, current, return_bb); |
241a2b9e JH |
669 | else if (TREE_CODE (retval) == PARM_DECL) |
670 | current->split_part_set_retval = false; | |
8813a647 | 671 | else if (VAR_P (retval) |
241a2b9e JH |
672 | || TREE_CODE (retval) == RESULT_DECL) |
673 | current->split_part_set_retval | |
674 | = bitmap_bit_p (non_ssa_vars, DECL_UID (retval)); | |
675 | else | |
676 | current->split_part_set_retval = true; | |
677 | ||
d5e254e1 IE |
678 | /* See if retbnd used by return bb is computed by header or split part. */ |
679 | retbnd = find_retbnd (return_bb); | |
680 | if (retbnd) | |
681 | { | |
682 | bool split_part_set_retbnd | |
683 | = split_part_set_ssa_name_p (retbnd, current, return_bb); | |
684 | ||
685 | /* If we have both return value and bounds then keep their definitions | |
686 | in a single function. We use SSA names to link returned bounds and | |
687 | value and therefore do not handle cases when result is passed by | |
688 | reference (which should not be our case anyway since bounds are | |
689 | returned for pointers only). */ | |
690 | if ((DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)) | |
691 | && current->split_part_set_retval) | |
692 | || split_part_set_retbnd != current->split_part_set_retval) | |
693 | { | |
694 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
695 | fprintf (dump_file, | |
696 | " Refused: split point splits return value and bounds\n"); | |
697 | return; | |
698 | } | |
699 | } | |
700 | ||
28fc44f3 JJ |
701 | /* split_function fixes up at most one PHI non-virtual PHI node in return_bb, |
702 | for the return value. If there are other PHIs, give up. */ | |
fefa31b5 | 703 | if (return_bb != EXIT_BLOCK_PTR_FOR_FN (cfun)) |
28fc44f3 | 704 | { |
538dd0b7 | 705 | gphi_iterator psi; |
28fc44f3 JJ |
706 | |
707 | for (psi = gsi_start_phis (return_bb); !gsi_end_p (psi); gsi_next (&psi)) | |
538dd0b7 | 708 | if (!virtual_operand_p (gimple_phi_result (psi.phi ())) |
28fc44f3 JJ |
709 | && !(retval |
710 | && current->split_part_set_retval | |
711 | && TREE_CODE (retval) == SSA_NAME | |
712 | && !DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)) | |
538dd0b7 | 713 | && SSA_NAME_DEF_STMT (retval) == psi.phi ())) |
28fc44f3 JJ |
714 | { |
715 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
716 | fprintf (dump_file, | |
717 | " Refused: return bb has extra PHIs\n"); | |
718 | return; | |
719 | } | |
720 | } | |
721 | ||
722 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
723 | fprintf (dump_file, " Accepted!\n"); | |
724 | ||
ed10d09b | 725 | /* At the moment chose split point with lowest count and that leaves |
3e485f62 JH |
726 | out smallest size of header. |
727 | In future we might re-consider this heuristics. */ | |
728 | if (!best_split_point.split_bbs | |
ed10d09b JH |
729 | || best_split_point.count |
730 | > current->count | |
731 | || (best_split_point.count == current->count | |
3e485f62 JH |
732 | && best_split_point.split_size < current->split_size)) |
733 | ||
734 | { | |
735 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
736 | fprintf (dump_file, " New best split point!\n"); | |
737 | if (best_split_point.ssa_names_to_pass) | |
738 | { | |
739 | BITMAP_FREE (best_split_point.ssa_names_to_pass); | |
740 | BITMAP_FREE (best_split_point.split_bbs); | |
741 | } | |
742 | best_split_point = *current; | |
743 | best_split_point.ssa_names_to_pass = BITMAP_ALLOC (NULL); | |
744 | bitmap_copy (best_split_point.ssa_names_to_pass, | |
745 | current->ssa_names_to_pass); | |
746 | best_split_point.split_bbs = BITMAP_ALLOC (NULL); | |
747 | bitmap_copy (best_split_point.split_bbs, current->split_bbs); | |
748 | } | |
749 | } | |
750 | ||
2094f1fc JH |
751 | /* Return basic block containing RETURN statement. We allow basic blocks |
752 | of the form: | |
753 | <retval> = tmp_var; | |
754 | return <retval> | |
bfd71482 JJ |
755 | but return_bb can not be more complex than this (except for |
756 | -fsanitize=thread we allow TSAN_FUNC_EXIT () internal call in there). | |
6626665f | 757 | If nothing is found, return the exit block. |
2094f1fc | 758 | |
3e485f62 JH |
759 | When there are multiple RETURN statement, chose one with return value, |
760 | since that one is more likely shared by multiple code paths. | |
2094f1fc JH |
761 | |
762 | Return BB is special, because for function splitting it is the only | |
763 | basic block that is duplicated in between header and split part of the | |
764 | function. | |
765 | ||
3e485f62 JH |
766 | TODO: We might support multiple return blocks. */ |
767 | ||
768 | static basic_block | |
769 | find_return_bb (void) | |
770 | { | |
771 | edge e; | |
fefa31b5 | 772 | basic_block return_bb = EXIT_BLOCK_PTR_FOR_FN (cfun); |
68457901 JJ |
773 | gimple_stmt_iterator bsi; |
774 | bool found_return = false; | |
775 | tree retval = NULL_TREE; | |
3e485f62 | 776 | |
fefa31b5 | 777 | if (!single_pred_p (EXIT_BLOCK_PTR_FOR_FN (cfun))) |
68457901 JJ |
778 | return return_bb; |
779 | ||
fefa31b5 | 780 | e = single_pred_edge (EXIT_BLOCK_PTR_FOR_FN (cfun)); |
68457901 JJ |
781 | for (bsi = gsi_last_bb (e->src); !gsi_end_p (bsi); gsi_prev (&bsi)) |
782 | { | |
355fe088 | 783 | gimple *stmt = gsi_stmt (bsi); |
a348dc7f JJ |
784 | if (gimple_code (stmt) == GIMPLE_LABEL |
785 | || is_gimple_debug (stmt) | |
786 | || gimple_clobber_p (stmt)) | |
68457901 JJ |
787 | ; |
788 | else if (gimple_code (stmt) == GIMPLE_ASSIGN | |
789 | && found_return | |
790 | && gimple_assign_single_p (stmt) | |
791 | && (auto_var_in_fn_p (gimple_assign_rhs1 (stmt), | |
792 | current_function_decl) | |
793 | || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))) | |
794 | && retval == gimple_assign_lhs (stmt)) | |
795 | ; | |
538dd0b7 | 796 | else if (greturn *return_stmt = dyn_cast <greturn *> (stmt)) |
68457901 JJ |
797 | { |
798 | found_return = true; | |
538dd0b7 | 799 | retval = gimple_return_retval (return_stmt); |
68457901 | 800 | } |
bfd71482 JJ |
801 | /* For -fsanitize=thread, allow also TSAN_FUNC_EXIT () in the return |
802 | bb. */ | |
803 | else if ((flag_sanitize & SANITIZE_THREAD) | |
8e4284d0 | 804 | && gimple_call_internal_p (stmt, IFN_TSAN_FUNC_EXIT)) |
bfd71482 | 805 | ; |
68457901 JJ |
806 | else |
807 | break; | |
808 | } | |
809 | if (gsi_end_p (bsi) && found_return) | |
810 | return_bb = e->src; | |
3e485f62 | 811 | |
3e485f62 JH |
812 | return return_bb; |
813 | } | |
814 | ||
ed7656f6 | 815 | /* Given return basic block RETURN_BB, see where return value is really |
2094f1fc JH |
816 | stored. */ |
817 | static tree | |
818 | find_retval (basic_block return_bb) | |
819 | { | |
820 | gimple_stmt_iterator bsi; | |
821 | for (bsi = gsi_start_bb (return_bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
538dd0b7 DM |
822 | if (greturn *return_stmt = dyn_cast <greturn *> (gsi_stmt (bsi))) |
823 | return gimple_return_retval (return_stmt); | |
a348dc7f JJ |
824 | else if (gimple_code (gsi_stmt (bsi)) == GIMPLE_ASSIGN |
825 | && !gimple_clobber_p (gsi_stmt (bsi))) | |
2094f1fc JH |
826 | return gimple_assign_rhs1 (gsi_stmt (bsi)); |
827 | return NULL; | |
828 | } | |
829 | ||
d5e254e1 IE |
830 | /* Given return basic block RETURN_BB, see where return bounds are really |
831 | stored. */ | |
832 | static tree | |
833 | find_retbnd (basic_block return_bb) | |
834 | { | |
835 | gimple_stmt_iterator bsi; | |
836 | for (bsi = gsi_last_bb (return_bb); !gsi_end_p (bsi); gsi_prev (&bsi)) | |
837 | if (gimple_code (gsi_stmt (bsi)) == GIMPLE_RETURN) | |
838 | return gimple_return_retbnd (gsi_stmt (bsi)); | |
839 | return NULL; | |
840 | } | |
841 | ||
1802378d EB |
842 | /* Callback for walk_stmt_load_store_addr_ops. If T is non-SSA automatic |
843 | variable, mark it as used in bitmap passed via DATA. | |
3e485f62 JH |
844 | Return true when access to T prevents splitting the function. */ |
845 | ||
846 | static bool | |
355fe088 | 847 | mark_nonssa_use (gimple *, tree t, tree, void *data) |
3e485f62 JH |
848 | { |
849 | t = get_base_address (t); | |
850 | ||
851 | if (!t || is_gimple_reg (t)) | |
852 | return false; | |
853 | ||
854 | /* At present we can't pass non-SSA arguments to split function. | |
855 | FIXME: this can be relaxed by passing references to arguments. */ | |
856 | if (TREE_CODE (t) == PARM_DECL) | |
857 | { | |
858 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1802378d EB |
859 | fprintf (dump_file, |
860 | "Cannot split: use of non-ssa function parameter.\n"); | |
3e485f62 JH |
861 | return true; |
862 | } | |
863 | ||
8813a647 | 864 | if ((VAR_P (t) && auto_var_in_fn_p (t, current_function_decl)) |
1802378d | 865 | || TREE_CODE (t) == RESULT_DECL |
8813a647 | 866 | || (TREE_CODE (t) == LABEL_DECL && FORCED_LABEL (t))) |
3e485f62 | 867 | bitmap_set_bit ((bitmap)data, DECL_UID (t)); |
241a2b9e | 868 | |
1802378d EB |
869 | /* For DECL_BY_REFERENCE, the return value is actually a pointer. We want |
870 | to pretend that the value pointed to is actual result decl. */ | |
871 | if ((TREE_CODE (t) == MEM_REF || INDIRECT_REF_P (t)) | |
241a2b9e | 872 | && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME |
70b5e7dc | 873 | && SSA_NAME_VAR (TREE_OPERAND (t, 0)) |
241a2b9e JH |
874 | && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND (t, 0))) == RESULT_DECL |
875 | && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) | |
1802378d EB |
876 | return |
877 | bitmap_bit_p ((bitmap)data, | |
878 | DECL_UID (DECL_RESULT (current_function_decl))); | |
879 | ||
3e485f62 JH |
880 | return false; |
881 | } | |
882 | ||
883 | /* Compute local properties of basic block BB we collect when looking for | |
884 | split points. We look for ssa defs and store them in SET_SSA_NAMES, | |
885 | for ssa uses and store them in USED_SSA_NAMES and for any non-SSA automatic | |
886 | vars stored in NON_SSA_VARS. | |
887 | ||
888 | When BB has edge to RETURN_BB, collect uses in RETURN_BB too. | |
889 | ||
890 | Return false when BB contains something that prevents it from being put into | |
891 | split function. */ | |
892 | ||
893 | static bool | |
894 | visit_bb (basic_block bb, basic_block return_bb, | |
895 | bitmap set_ssa_names, bitmap used_ssa_names, | |
896 | bitmap non_ssa_vars) | |
897 | { | |
3e485f62 JH |
898 | edge e; |
899 | edge_iterator ei; | |
900 | bool can_split = true; | |
901 | ||
538dd0b7 DM |
902 | for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi); |
903 | gsi_next (&bsi)) | |
3e485f62 | 904 | { |
355fe088 | 905 | gimple *stmt = gsi_stmt (bsi); |
3e485f62 JH |
906 | tree op; |
907 | ssa_op_iter iter; | |
908 | tree decl; | |
909 | ||
910 | if (is_gimple_debug (stmt)) | |
911 | continue; | |
912 | ||
a348dc7f JJ |
913 | if (gimple_clobber_p (stmt)) |
914 | continue; | |
915 | ||
3e485f62 JH |
916 | /* FIXME: We can split regions containing EH. We can not however |
917 | split RESX, EH_DISPATCH and EH_POINTER referring to same region | |
918 | into different partitions. This would require tracking of | |
919 | EH regions and checking in consider_split_point if they | |
920 | are not used elsewhere. */ | |
1da7d8c0 | 921 | if (gimple_code (stmt) == GIMPLE_RESX) |
3e485f62 JH |
922 | { |
923 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1da7d8c0 | 924 | fprintf (dump_file, "Cannot split: resx.\n"); |
3e485f62 JH |
925 | can_split = false; |
926 | } | |
927 | if (gimple_code (stmt) == GIMPLE_EH_DISPATCH) | |
928 | { | |
929 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1802378d | 930 | fprintf (dump_file, "Cannot split: eh dispatch.\n"); |
3e485f62 JH |
931 | can_split = false; |
932 | } | |
933 | ||
934 | /* Check builtins that prevent splitting. */ | |
935 | if (gimple_code (stmt) == GIMPLE_CALL | |
936 | && (decl = gimple_call_fndecl (stmt)) != NULL_TREE | |
937 | && DECL_BUILT_IN (decl) | |
938 | && DECL_BUILT_IN_CLASS (decl) == BUILT_IN_NORMAL) | |
939 | switch (DECL_FUNCTION_CODE (decl)) | |
940 | { | |
941 | /* FIXME: once we will allow passing non-parm values to split part, | |
942 | we need to be sure to handle correct builtin_stack_save and | |
943 | builtin_stack_restore. At the moment we are safe; there is no | |
944 | way to store builtin_stack_save result in non-SSA variable | |
945 | since all calls to those are compiler generated. */ | |
946 | case BUILT_IN_APPLY: | |
61e03ffc | 947 | case BUILT_IN_APPLY_ARGS: |
3e485f62 JH |
948 | case BUILT_IN_VA_START: |
949 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1802378d EB |
950 | fprintf (dump_file, |
951 | "Cannot split: builtin_apply and va_start.\n"); | |
3e485f62 JH |
952 | can_split = false; |
953 | break; | |
954 | case BUILT_IN_EH_POINTER: | |
955 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1802378d | 956 | fprintf (dump_file, "Cannot split: builtin_eh_pointer.\n"); |
3e485f62 JH |
957 | can_split = false; |
958 | break; | |
959 | default: | |
960 | break; | |
961 | } | |
962 | ||
963 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF) | |
964 | bitmap_set_bit (set_ssa_names, SSA_NAME_VERSION (op)); | |
965 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE) | |
966 | bitmap_set_bit (used_ssa_names, SSA_NAME_VERSION (op)); | |
967 | can_split &= !walk_stmt_load_store_addr_ops (stmt, non_ssa_vars, | |
968 | mark_nonssa_use, | |
969 | mark_nonssa_use, | |
970 | mark_nonssa_use); | |
971 | } | |
538dd0b7 DM |
972 | for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi); |
973 | gsi_next (&bsi)) | |
3e485f62 | 974 | { |
538dd0b7 | 975 | gphi *stmt = bsi.phi (); |
8b3057b3 | 976 | unsigned int i; |
3e485f62 | 977 | |
ea057359 | 978 | if (virtual_operand_p (gimple_phi_result (stmt))) |
3e485f62 | 979 | continue; |
8b3057b3 JH |
980 | bitmap_set_bit (set_ssa_names, |
981 | SSA_NAME_VERSION (gimple_phi_result (stmt))); | |
982 | for (i = 0; i < gimple_phi_num_args (stmt); i++) | |
983 | { | |
984 | tree op = gimple_phi_arg_def (stmt, i); | |
985 | if (TREE_CODE (op) == SSA_NAME) | |
986 | bitmap_set_bit (used_ssa_names, SSA_NAME_VERSION (op)); | |
987 | } | |
3e485f62 JH |
988 | can_split &= !walk_stmt_load_store_addr_ops (stmt, non_ssa_vars, |
989 | mark_nonssa_use, | |
990 | mark_nonssa_use, | |
991 | mark_nonssa_use); | |
992 | } | |
ed7656f6 | 993 | /* Record also uses coming from PHI operand in return BB. */ |
3e485f62 JH |
994 | FOR_EACH_EDGE (e, ei, bb->succs) |
995 | if (e->dest == return_bb) | |
996 | { | |
538dd0b7 DM |
997 | for (gphi_iterator bsi = gsi_start_phis (return_bb); |
998 | !gsi_end_p (bsi); | |
999 | gsi_next (&bsi)) | |
3e485f62 | 1000 | { |
538dd0b7 | 1001 | gphi *stmt = bsi.phi (); |
3e485f62 JH |
1002 | tree op = gimple_phi_arg_def (stmt, e->dest_idx); |
1003 | ||
ea057359 | 1004 | if (virtual_operand_p (gimple_phi_result (stmt))) |
3e485f62 | 1005 | continue; |
3e485f62 JH |
1006 | if (TREE_CODE (op) == SSA_NAME) |
1007 | bitmap_set_bit (used_ssa_names, SSA_NAME_VERSION (op)); | |
1008 | else | |
9f1363cd | 1009 | can_split &= !mark_nonssa_use (stmt, op, op, non_ssa_vars); |
3e485f62 | 1010 | } |
3e485f62 JH |
1011 | } |
1012 | return can_split; | |
1013 | } | |
1014 | ||
1015 | /* Stack entry for recursive DFS walk in find_split_point. */ | |
1016 | ||
50686850 | 1017 | struct stack_entry |
3e485f62 JH |
1018 | { |
1019 | /* Basic block we are examining. */ | |
1020 | basic_block bb; | |
1021 | ||
1022 | /* SSA names set and used by the BB and all BBs reachable | |
1023 | from it via DFS walk. */ | |
1024 | bitmap set_ssa_names, used_ssa_names; | |
1025 | bitmap non_ssa_vars; | |
1026 | ||
1027 | /* All BBS visited from this BB via DFS walk. */ | |
1028 | bitmap bbs_visited; | |
1029 | ||
1030 | /* Last examined edge in DFS walk. Since we walk unoriented graph, | |
ed7656f6 | 1031 | the value is up to sum of incoming and outgoing edges of BB. */ |
3e485f62 JH |
1032 | unsigned int edge_num; |
1033 | ||
1034 | /* Stack entry index of earliest BB reachable from current BB | |
ed7656f6 | 1035 | or any BB visited later in DFS walk. */ |
3e485f62 JH |
1036 | int earliest; |
1037 | ||
1038 | /* Overall time and size of all BBs reached from this BB in DFS walk. */ | |
c2e87766 JH |
1039 | sreal overall_time; |
1040 | int overall_size; | |
3e485f62 JH |
1041 | |
1042 | /* When false we can not split on this BB. */ | |
1043 | bool can_split; | |
50686850 | 1044 | }; |
3e485f62 JH |
1045 | |
1046 | ||
1047 | /* Find all articulations and call consider_split on them. | |
1048 | OVERALL_TIME and OVERALL_SIZE is time and size of the function. | |
1049 | ||
1050 | We perform basic algorithm for finding an articulation in a graph | |
1051 | created from CFG by considering it to be an unoriented graph. | |
1052 | ||
1053 | The articulation is discovered via DFS walk. We collect earliest | |
1054 | basic block on stack that is reachable via backward edge. Articulation | |
1055 | is any basic block such that there is no backward edge bypassing it. | |
1056 | To reduce stack usage we maintain heap allocated stack in STACK vector. | |
1057 | AUX pointer of BB is set to index it appears in the stack or -1 once | |
1058 | it is visited and popped off the stack. | |
1059 | ||
1060 | The algorithm finds articulation after visiting the whole component | |
1061 | reachable by it. This makes it convenient to collect information about | |
1062 | the component used by consider_split. */ | |
1063 | ||
1064 | static void | |
c2e87766 | 1065 | find_split_points (basic_block return_bb, sreal overall_time, int overall_size) |
3e485f62 JH |
1066 | { |
1067 | stack_entry first; | |
6e1aa848 | 1068 | vec<stack_entry> stack = vNULL; |
3e485f62 | 1069 | basic_block bb; |
3e485f62 JH |
1070 | struct split_point current; |
1071 | ||
1072 | current.header_time = overall_time; | |
1073 | current.header_size = overall_size; | |
1074 | current.split_time = 0; | |
1075 | current.split_size = 0; | |
1076 | current.ssa_names_to_pass = BITMAP_ALLOC (NULL); | |
1077 | ||
fefa31b5 | 1078 | first.bb = ENTRY_BLOCK_PTR_FOR_FN (cfun); |
3e485f62 JH |
1079 | first.edge_num = 0; |
1080 | first.overall_time = 0; | |
1081 | first.overall_size = 0; | |
1082 | first.earliest = INT_MAX; | |
1083 | first.set_ssa_names = 0; | |
1084 | first.used_ssa_names = 0; | |
e15eb172 | 1085 | first.non_ssa_vars = 0; |
3e485f62 | 1086 | first.bbs_visited = 0; |
e15eb172 | 1087 | first.can_split = false; |
9771b263 | 1088 | stack.safe_push (first); |
fefa31b5 | 1089 | ENTRY_BLOCK_PTR_FOR_FN (cfun)->aux = (void *)(intptr_t)-1; |
3e485f62 | 1090 | |
9771b263 | 1091 | while (!stack.is_empty ()) |
3e485f62 | 1092 | { |
9771b263 | 1093 | stack_entry *entry = &stack.last (); |
3e485f62 JH |
1094 | |
1095 | /* We are walking an acyclic graph, so edge_num counts | |
1096 | succ and pred edges together. However when considering | |
1097 | articulation, we want to have processed everything reachable | |
1098 | from articulation but nothing that reaches into it. */ | |
1099 | if (entry->edge_num == EDGE_COUNT (entry->bb->succs) | |
fefa31b5 | 1100 | && entry->bb != ENTRY_BLOCK_PTR_FOR_FN (cfun)) |
3e485f62 | 1101 | { |
9771b263 | 1102 | int pos = stack.length (); |
3e485f62 JH |
1103 | entry->can_split &= visit_bb (entry->bb, return_bb, |
1104 | entry->set_ssa_names, | |
1105 | entry->used_ssa_names, | |
1106 | entry->non_ssa_vars); | |
1107 | if (pos <= entry->earliest && !entry->can_split | |
1108 | && dump_file && (dump_flags & TDF_DETAILS)) | |
1109 | fprintf (dump_file, | |
1110 | "found articulation at bb %i but can not split\n", | |
1111 | entry->bb->index); | |
1112 | if (pos <= entry->earliest && entry->can_split) | |
1113 | { | |
1114 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1115 | fprintf (dump_file, "found articulation at bb %i\n", | |
1116 | entry->bb->index); | |
1117 | current.entry_bb = entry->bb; | |
1118 | current.ssa_names_to_pass = BITMAP_ALLOC (NULL); | |
1119 | bitmap_and_compl (current.ssa_names_to_pass, | |
1120 | entry->used_ssa_names, entry->set_ssa_names); | |
1121 | current.header_time = overall_time - entry->overall_time; | |
1122 | current.header_size = overall_size - entry->overall_size; | |
1123 | current.split_time = entry->overall_time; | |
1124 | current.split_size = entry->overall_size; | |
1125 | current.split_bbs = entry->bbs_visited; | |
1126 | consider_split (¤t, entry->non_ssa_vars, return_bb); | |
1127 | BITMAP_FREE (current.ssa_names_to_pass); | |
1128 | } | |
1129 | } | |
1130 | /* Do actual DFS walk. */ | |
1131 | if (entry->edge_num | |
1132 | < (EDGE_COUNT (entry->bb->succs) | |
1133 | + EDGE_COUNT (entry->bb->preds))) | |
1134 | { | |
1135 | edge e; | |
1136 | basic_block dest; | |
1137 | if (entry->edge_num < EDGE_COUNT (entry->bb->succs)) | |
1138 | { | |
1139 | e = EDGE_SUCC (entry->bb, entry->edge_num); | |
1140 | dest = e->dest; | |
1141 | } | |
1142 | else | |
1143 | { | |
1144 | e = EDGE_PRED (entry->bb, entry->edge_num | |
1145 | - EDGE_COUNT (entry->bb->succs)); | |
1146 | dest = e->src; | |
1147 | } | |
1148 | ||
1149 | entry->edge_num++; | |
1150 | ||
1151 | /* New BB to visit, push it to the stack. */ | |
fefa31b5 | 1152 | if (dest != return_bb && dest != EXIT_BLOCK_PTR_FOR_FN (cfun) |
3e485f62 JH |
1153 | && !dest->aux) |
1154 | { | |
1155 | stack_entry new_entry; | |
1156 | ||
1157 | new_entry.bb = dest; | |
1158 | new_entry.edge_num = 0; | |
1159 | new_entry.overall_time | |
9771b263 | 1160 | = bb_info_vec[dest->index].time; |
3e485f62 | 1161 | new_entry.overall_size |
9771b263 | 1162 | = bb_info_vec[dest->index].size; |
3e485f62 JH |
1163 | new_entry.earliest = INT_MAX; |
1164 | new_entry.set_ssa_names = BITMAP_ALLOC (NULL); | |
1165 | new_entry.used_ssa_names = BITMAP_ALLOC (NULL); | |
1166 | new_entry.bbs_visited = BITMAP_ALLOC (NULL); | |
1167 | new_entry.non_ssa_vars = BITMAP_ALLOC (NULL); | |
1168 | new_entry.can_split = true; | |
1169 | bitmap_set_bit (new_entry.bbs_visited, dest->index); | |
9771b263 DN |
1170 | stack.safe_push (new_entry); |
1171 | dest->aux = (void *)(intptr_t)stack.length (); | |
3e485f62 JH |
1172 | } |
1173 | /* Back edge found, record the earliest point. */ | |
1174 | else if ((intptr_t)dest->aux > 0 | |
1175 | && (intptr_t)dest->aux < entry->earliest) | |
1176 | entry->earliest = (intptr_t)dest->aux; | |
1177 | } | |
ed7656f6 JJ |
1178 | /* We are done with examining the edges. Pop off the value from stack |
1179 | and merge stuff we accumulate during the walk. */ | |
fefa31b5 | 1180 | else if (entry->bb != ENTRY_BLOCK_PTR_FOR_FN (cfun)) |
3e485f62 | 1181 | { |
9771b263 | 1182 | stack_entry *prev = &stack[stack.length () - 2]; |
3e485f62 JH |
1183 | |
1184 | entry->bb->aux = (void *)(intptr_t)-1; | |
1185 | prev->can_split &= entry->can_split; | |
1186 | if (prev->set_ssa_names) | |
1187 | { | |
1188 | bitmap_ior_into (prev->set_ssa_names, entry->set_ssa_names); | |
1189 | bitmap_ior_into (prev->used_ssa_names, entry->used_ssa_names); | |
1190 | bitmap_ior_into (prev->bbs_visited, entry->bbs_visited); | |
1191 | bitmap_ior_into (prev->non_ssa_vars, entry->non_ssa_vars); | |
1192 | } | |
1193 | if (prev->earliest > entry->earliest) | |
1194 | prev->earliest = entry->earliest; | |
1195 | prev->overall_time += entry->overall_time; | |
1196 | prev->overall_size += entry->overall_size; | |
1197 | BITMAP_FREE (entry->set_ssa_names); | |
1198 | BITMAP_FREE (entry->used_ssa_names); | |
1199 | BITMAP_FREE (entry->bbs_visited); | |
1200 | BITMAP_FREE (entry->non_ssa_vars); | |
9771b263 | 1201 | stack.pop (); |
3e485f62 JH |
1202 | } |
1203 | else | |
9771b263 | 1204 | stack.pop (); |
3e485f62 | 1205 | } |
fefa31b5 | 1206 | ENTRY_BLOCK_PTR_FOR_FN (cfun)->aux = NULL; |
11cd3bed | 1207 | FOR_EACH_BB_FN (bb, cfun) |
3e485f62 | 1208 | bb->aux = NULL; |
9771b263 | 1209 | stack.release (); |
3e485f62 JH |
1210 | BITMAP_FREE (current.ssa_names_to_pass); |
1211 | } | |
1212 | ||
1213 | /* Split function at SPLIT_POINT. */ | |
1214 | ||
1215 | static void | |
bfd71482 JJ |
1216 | split_function (basic_block return_bb, struct split_point *split_point, |
1217 | bool add_tsan_func_exit) | |
3e485f62 | 1218 | { |
6e1aa848 | 1219 | vec<tree> args_to_pass = vNULL; |
201176d3 | 1220 | bitmap args_to_skip; |
3e485f62 JH |
1221 | tree parm; |
1222 | int num = 0; | |
d52f5295 | 1223 | cgraph_node *node, *cur_node = cgraph_node::get (current_function_decl); |
3e485f62 | 1224 | basic_block call_bb; |
bfd71482 | 1225 | gcall *call, *tsan_func_exit_call = NULL; |
3e485f62 JH |
1226 | edge e; |
1227 | edge_iterator ei; | |
d5e254e1 | 1228 | tree retval = NULL, real_retval = NULL, retbnd = NULL; |
d5e254e1 | 1229 | bool with_bounds = chkp_function_instrumented_p (current_function_decl); |
355fe088 | 1230 | gimple *last_stmt = NULL; |
371556ee | 1231 | unsigned int i; |
32244553 | 1232 | tree arg, ddef; |
3e485f62 JH |
1233 | |
1234 | if (dump_file) | |
1235 | { | |
1236 | fprintf (dump_file, "\n\nSplitting function at:\n"); | |
1237 | dump_split_point (dump_file, split_point); | |
1238 | } | |
1239 | ||
201176d3 MJ |
1240 | if (cur_node->local.can_change_signature) |
1241 | args_to_skip = BITMAP_ALLOC (NULL); | |
1242 | else | |
1243 | args_to_skip = NULL; | |
1244 | ||
3e485f62 JH |
1245 | /* Collect the parameters of new function and args_to_skip bitmap. */ |
1246 | for (parm = DECL_ARGUMENTS (current_function_decl); | |
910ad8de | 1247 | parm; parm = DECL_CHAIN (parm), num++) |
201176d3 MJ |
1248 | if (args_to_skip |
1249 | && (!is_gimple_reg (parm) | |
32244553 | 1250 | || (ddef = ssa_default_def (cfun, parm)) == NULL_TREE |
201176d3 | 1251 | || !bitmap_bit_p (split_point->ssa_names_to_pass, |
32244553 | 1252 | SSA_NAME_VERSION (ddef)))) |
3e485f62 JH |
1253 | bitmap_set_bit (args_to_skip, num); |
1254 | else | |
371556ee | 1255 | { |
86814190 MJ |
1256 | /* This parm might not have been used up to now, but is going to be |
1257 | used, hence register it. */ | |
86814190 | 1258 | if (is_gimple_reg (parm)) |
32244553 | 1259 | arg = get_or_create_ssa_default_def (cfun, parm); |
86814190 MJ |
1260 | else |
1261 | arg = parm; | |
201176d3 | 1262 | |
2b3c0885 RG |
1263 | if (!useless_type_conversion_p (DECL_ARG_TYPE (parm), TREE_TYPE (arg))) |
1264 | arg = fold_convert (DECL_ARG_TYPE (parm), arg); | |
9771b263 | 1265 | args_to_pass.safe_push (arg); |
371556ee | 1266 | } |
3e485f62 | 1267 | |
b69539cb | 1268 | /* See if the split function will return. */ |
ece95567 | 1269 | bool split_part_return_p = false; |
3e485f62 | 1270 | FOR_EACH_EDGE (e, ei, return_bb->preds) |
ece95567 RB |
1271 | { |
1272 | if (bitmap_bit_p (split_point->split_bbs, e->src->index)) | |
1273 | split_part_return_p = true; | |
ece95567 | 1274 | } |
3e485f62 | 1275 | |
241a2b9e JH |
1276 | /* Add return block to what will become the split function. |
1277 | We do not return; no return block is needed. */ | |
1278 | if (!split_part_return_p) | |
1279 | ; | |
1280 | /* We have no return block, so nothing is needed. */ | |
fefa31b5 | 1281 | else if (return_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)) |
241a2b9e JH |
1282 | ; |
1283 | /* When we do not want to return value, we need to construct | |
1284 | new return block with empty return statement. | |
1285 | FIXME: Once we are able to change return type, we should change function | |
1286 | to return void instead of just outputting function with undefined return | |
1287 | value. For structures this affects quality of codegen. */ | |
b69539cb JJ |
1288 | else if ((retval = find_retval (return_bb)) |
1289 | && !split_point->split_part_set_retval) | |
241a2b9e JH |
1290 | { |
1291 | bool redirected = true; | |
1292 | basic_block new_return_bb = create_basic_block (NULL, 0, return_bb); | |
1293 | gimple_stmt_iterator gsi = gsi_start_bb (new_return_bb); | |
1294 | gsi_insert_after (&gsi, gimple_build_return (NULL), GSI_NEW_STMT); | |
ce6ec234 | 1295 | new_return_bb->count = profile_count::zero (); |
241a2b9e JH |
1296 | while (redirected) |
1297 | { | |
1298 | redirected = false; | |
1299 | FOR_EACH_EDGE (e, ei, return_bb->preds) | |
1300 | if (bitmap_bit_p (split_point->split_bbs, e->src->index)) | |
1301 | { | |
e7a74006 | 1302 | new_return_bb->count += e->count (); |
241a2b9e JH |
1303 | redirect_edge_and_branch (e, new_return_bb); |
1304 | redirected = true; | |
1305 | break; | |
1306 | } | |
1307 | } | |
357067f2 | 1308 | e = make_single_succ_edge (new_return_bb, EXIT_BLOCK_PTR_FOR_FN (cfun), 0); |
726338f4 | 1309 | add_bb_to_loop (new_return_bb, current_loops->tree_root); |
241a2b9e | 1310 | bitmap_set_bit (split_point->split_bbs, new_return_bb->index); |
d4f4e71c | 1311 | retbnd = find_retbnd (return_bb); |
2e5e346d JL |
1312 | } |
1313 | /* When we pass around the value, use existing return block. */ | |
1314 | else | |
b69539cb JJ |
1315 | { |
1316 | bitmap_set_bit (split_point->split_bbs, return_bb->index); | |
1317 | retbnd = find_retbnd (return_bb); | |
1318 | } | |
ece95567 | 1319 | |
2e5e346d JL |
1320 | /* If RETURN_BB has virtual operand PHIs, they must be removed and the |
1321 | virtual operand marked for renaming as we change the CFG in a way that | |
cfef45c8 | 1322 | tree-inline is not able to compensate for. |
2e5e346d JL |
1323 | |
1324 | Note this can happen whether or not we have a return value. If we have | |
1325 | a return value, then RETURN_BB may have PHIs for real operands too. */ | |
fefa31b5 | 1326 | if (return_bb != EXIT_BLOCK_PTR_FOR_FN (cfun)) |
2e5e346d | 1327 | { |
cfef45c8 | 1328 | bool phi_p = false; |
538dd0b7 DM |
1329 | for (gphi_iterator gsi = gsi_start_phis (return_bb); |
1330 | !gsi_end_p (gsi);) | |
241a2b9e | 1331 | { |
538dd0b7 | 1332 | gphi *stmt = gsi.phi (); |
ea057359 | 1333 | if (!virtual_operand_p (gimple_phi_result (stmt))) |
2e5e346d JL |
1334 | { |
1335 | gsi_next (&gsi); | |
1336 | continue; | |
1337 | } | |
6b8c9df8 RG |
1338 | mark_virtual_phi_result_for_renaming (stmt); |
1339 | remove_phi_node (&gsi, true); | |
cfef45c8 | 1340 | phi_p = true; |
241a2b9e | 1341 | } |
cfef45c8 RG |
1342 | /* In reality we have to rename the reaching definition of the |
1343 | virtual operand at return_bb as we will eventually release it | |
1344 | when we remove the code region we outlined. | |
1345 | So we have to rename all immediate virtual uses of that region | |
1346 | if we didn't see a PHI definition yet. */ | |
1347 | /* ??? In real reality we want to set the reaching vdef of the | |
1348 | entry of the SESE region as the vuse of the call and the reaching | |
1349 | vdef of the exit of the SESE region as the vdef of the call. */ | |
1350 | if (!phi_p) | |
538dd0b7 DM |
1351 | for (gimple_stmt_iterator gsi = gsi_start_bb (return_bb); |
1352 | !gsi_end_p (gsi); | |
1353 | gsi_next (&gsi)) | |
cfef45c8 | 1354 | { |
355fe088 | 1355 | gimple *stmt = gsi_stmt (gsi); |
cfef45c8 RG |
1356 | if (gimple_vuse (stmt)) |
1357 | { | |
1358 | gimple_set_vuse (stmt, NULL_TREE); | |
1359 | update_stmt (stmt); | |
1360 | } | |
1361 | if (gimple_vdef (stmt)) | |
1362 | break; | |
1363 | } | |
241a2b9e | 1364 | } |
3e485f62 JH |
1365 | |
1366 | /* Now create the actual clone. */ | |
3dafb85c | 1367 | cgraph_edge::rebuild_edges (); |
d52f5295 | 1368 | node = cur_node->create_version_clone_with_body |
e199dd0a JJ |
1369 | (vNULL, NULL, args_to_skip, |
1370 | !split_part_return_p || !split_point->split_part_set_retval, | |
1371 | split_point->split_bbs, split_point->entry_bb, "part"); | |
9cec31f4 | 1372 | |
41f669d8 JH |
1373 | node->split_part = true; |
1374 | ||
c6cf6ef7 ML |
1375 | if (cur_node->same_comdat_group) |
1376 | { | |
1377 | /* TODO: call is versionable if we make sure that all | |
1378 | callers are inside of a comdat group. */ | |
1379 | cur_node->calls_comdat_local = 1; | |
1380 | node->add_to_same_comdat_group (cur_node); | |
1381 | } | |
1382 | ||
1383 | ||
9cec31f4 ML |
1384 | /* Let's take a time profile for splitted function. */ |
1385 | node->tp_first_run = cur_node->tp_first_run + 1; | |
1386 | ||
d402c33d JH |
1387 | /* For usual cloning it is enough to clear builtin only when signature |
1388 | changes. For partial inlining we however can not expect the part | |
1389 | of builtin implementation to have same semantic as the whole. */ | |
67348ccc | 1390 | if (DECL_BUILT_IN (node->decl)) |
d402c33d | 1391 | { |
67348ccc DM |
1392 | DECL_BUILT_IN_CLASS (node->decl) = NOT_BUILT_IN; |
1393 | DECL_FUNCTION_CODE (node->decl) = (enum built_in_function) 0; | |
d402c33d | 1394 | } |
d5e254e1 | 1395 | |
b69539cb JJ |
1396 | /* If return_bb contains any clobbers that refer to SSA_NAMEs |
1397 | set in the split part, remove them. Also reset debug stmts that | |
1398 | refer to SSA_NAMEs set in the split part. */ | |
1399 | if (return_bb != EXIT_BLOCK_PTR_FOR_FN (cfun)) | |
1400 | { | |
1401 | gimple_stmt_iterator gsi = gsi_start_bb (return_bb); | |
1402 | while (!gsi_end_p (gsi)) | |
1403 | { | |
1404 | tree op; | |
1405 | ssa_op_iter iter; | |
1406 | gimple *stmt = gsi_stmt (gsi); | |
1407 | bool remove = false; | |
1408 | if (gimple_clobber_p (stmt) || is_gimple_debug (stmt)) | |
1409 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE) | |
1410 | { | |
1411 | basic_block bb = gimple_bb (SSA_NAME_DEF_STMT (op)); | |
1412 | if (op != retval | |
1413 | && bb | |
1414 | && bb != return_bb | |
1415 | && bitmap_bit_p (split_point->split_bbs, bb->index)) | |
1416 | { | |
1417 | if (is_gimple_debug (stmt)) | |
1418 | { | |
1419 | gimple_debug_bind_reset_value (stmt); | |
1420 | update_stmt (stmt); | |
1421 | } | |
1422 | else | |
1423 | remove = true; | |
1424 | break; | |
1425 | } | |
1426 | } | |
1427 | if (remove) | |
1428 | gsi_remove (&gsi, true); | |
1429 | else | |
1430 | gsi_next (&gsi); | |
1431 | } | |
1432 | } | |
1433 | ||
d5e254e1 IE |
1434 | /* If the original function is instrumented then it's |
1435 | part is also instrumented. */ | |
1436 | if (with_bounds) | |
1437 | chkp_function_mark_instrumented (node->decl); | |
1438 | ||
9a6af450 EB |
1439 | /* If the original function is declared inline, there is no point in issuing |
1440 | a warning for the non-inlinable part. */ | |
67348ccc | 1441 | DECL_NO_INLINE_WARNING_P (node->decl) = 1; |
d52f5295 | 1442 | cur_node->remove_callees (); |
d122681a | 1443 | cur_node->remove_all_references (); |
3e485f62 | 1444 | if (!split_part_return_p) |
67348ccc | 1445 | TREE_THIS_VOLATILE (node->decl) = 1; |
3e485f62 | 1446 | if (dump_file) |
67348ccc | 1447 | dump_function_to_file (node->decl, dump_file, dump_flags); |
3e485f62 JH |
1448 | |
1449 | /* Create the basic block we place call into. It is the entry basic block | |
1450 | split after last label. */ | |
1451 | call_bb = split_point->entry_bb; | |
538dd0b7 | 1452 | for (gimple_stmt_iterator gsi = gsi_start_bb (call_bb); !gsi_end_p (gsi);) |
3e485f62 JH |
1453 | if (gimple_code (gsi_stmt (gsi)) == GIMPLE_LABEL) |
1454 | { | |
1455 | last_stmt = gsi_stmt (gsi); | |
1456 | gsi_next (&gsi); | |
1457 | } | |
1458 | else | |
1459 | break; | |
ed10d09b | 1460 | call_bb->count = split_point->count; |
3e485f62 JH |
1461 | e = split_block (split_point->entry_bb, last_stmt); |
1462 | remove_edge (e); | |
1463 | ||
1464 | /* Produce the call statement. */ | |
538dd0b7 | 1465 | gimple_stmt_iterator gsi = gsi_last_bb (call_bb); |
9771b263 | 1466 | FOR_EACH_VEC_ELT (args_to_pass, i, arg) |
2b3c0885 RG |
1467 | if (!is_gimple_val (arg)) |
1468 | { | |
1469 | arg = force_gimple_operand_gsi (&gsi, arg, true, NULL_TREE, | |
f6e52e91 | 1470 | false, GSI_CONTINUE_LINKING); |
9771b263 | 1471 | args_to_pass[i] = arg; |
2b3c0885 | 1472 | } |
67348ccc | 1473 | call = gimple_build_call_vec (node->decl, args_to_pass); |
d5e254e1 | 1474 | gimple_call_set_with_bounds (call, with_bounds); |
3e485f62 | 1475 | gimple_set_block (call, DECL_INITIAL (current_function_decl)); |
9771b263 | 1476 | args_to_pass.release (); |
3e485f62 | 1477 | |
878eef4a JJ |
1478 | /* For optimized away parameters, add on the caller side |
1479 | before the call | |
1480 | DEBUG D#X => parm_Y(D) | |
1481 | stmts and associate D#X with parm in decl_debug_args_lookup | |
1482 | vector to say for debug info that if parameter parm had been passed, | |
1483 | it would have value parm_Y(D). */ | |
1484 | if (args_to_skip) | |
878eef4a | 1485 | { |
a49de7a4 JJ |
1486 | vec<tree, va_gc> **debug_args = NULL; |
1487 | unsigned i = 0, len = 0; | |
1488 | if (MAY_HAVE_DEBUG_STMTS) | |
878eef4a | 1489 | { |
a49de7a4 JJ |
1490 | debug_args = decl_debug_args_lookup (node->decl); |
1491 | if (debug_args) | |
1492 | len = vec_safe_length (*debug_args); | |
878eef4a | 1493 | } |
a49de7a4 JJ |
1494 | for (parm = DECL_ARGUMENTS (current_function_decl), num = 0; |
1495 | parm; parm = DECL_CHAIN (parm), num++) | |
1496 | if (bitmap_bit_p (args_to_skip, num) && is_gimple_reg (parm)) | |
1497 | { | |
1498 | tree ddecl; | |
1499 | gimple *def_temp; | |
1500 | ||
1501 | /* This needs to be done even without MAY_HAVE_DEBUG_STMTS, | |
1502 | otherwise if it didn't exist before, we'd end up with | |
1503 | different SSA_NAME_VERSIONs between -g and -g0. */ | |
1504 | arg = get_or_create_ssa_default_def (cfun, parm); | |
1505 | if (!MAY_HAVE_DEBUG_STMTS || debug_args == NULL) | |
1506 | continue; | |
1507 | ||
1508 | while (i < len && (**debug_args)[i] != DECL_ORIGIN (parm)) | |
1509 | i += 2; | |
1510 | if (i >= len) | |
1511 | continue; | |
1512 | ddecl = (**debug_args)[i + 1]; | |
1513 | def_temp | |
1514 | = gimple_build_debug_bind (ddecl, unshare_expr (arg), call); | |
1515 | gsi_insert_after (&gsi, def_temp, GSI_NEW_STMT); | |
1516 | } | |
878eef4a JJ |
1517 | } |
1518 | ||
556e9ba0 JH |
1519 | /* We avoid address being taken on any variable used by split part, |
1520 | so return slot optimization is always possible. Moreover this is | |
1521 | required to make DECL_BY_REFERENCE work. */ | |
1522 | if (aggregate_value_p (DECL_RESULT (current_function_decl), | |
22110e6c EB |
1523 | TREE_TYPE (current_function_decl)) |
1524 | && (!is_gimple_reg_type (TREE_TYPE (DECL_RESULT (current_function_decl))) | |
1525 | || DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)))) | |
556e9ba0 JH |
1526 | gimple_call_set_return_slot_opt (call, true); |
1527 | ||
bfd71482 JJ |
1528 | if (add_tsan_func_exit) |
1529 | tsan_func_exit_call = gimple_build_call_internal (IFN_TSAN_FUNC_EXIT, 0); | |
1530 | ||
3e485f62 JH |
1531 | /* Update return value. This is bit tricky. When we do not return, |
1532 | do nothing. When we return we might need to update return_bb | |
1533 | or produce a new return statement. */ | |
1534 | if (!split_part_return_p) | |
bfd71482 JJ |
1535 | { |
1536 | gsi_insert_after (&gsi, call, GSI_NEW_STMT); | |
1537 | if (tsan_func_exit_call) | |
1538 | gsi_insert_after (&gsi, tsan_func_exit_call, GSI_NEW_STMT); | |
1539 | } | |
3e485f62 JH |
1540 | else |
1541 | { | |
357067f2 JH |
1542 | e = make_single_succ_edge (call_bb, return_bb, |
1543 | return_bb == EXIT_BLOCK_PTR_FOR_FN (cfun) | |
1544 | ? 0 : EDGE_FALLTHRU); | |
6938f93f JH |
1545 | |
1546 | /* If there is return basic block, see what value we need to store | |
1547 | return value into and put call just before it. */ | |
fefa31b5 | 1548 | if (return_bb != EXIT_BLOCK_PTR_FOR_FN (cfun)) |
3e485f62 | 1549 | { |
b69539cb | 1550 | real_retval = retval; |
241a2b9e | 1551 | if (real_retval && split_point->split_part_set_retval) |
3e485f62 | 1552 | { |
538dd0b7 | 1553 | gphi_iterator psi; |
3e485f62 | 1554 | |
6938f93f JH |
1555 | /* See if we need new SSA_NAME for the result. |
1556 | When DECL_BY_REFERENCE is true, retval is actually pointer to | |
1557 | return value and it is constant in whole function. */ | |
1558 | if (TREE_CODE (retval) == SSA_NAME | |
1559 | && !DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) | |
3e485f62 | 1560 | { |
070ecdfd | 1561 | retval = copy_ssa_name (retval, call); |
6938f93f JH |
1562 | |
1563 | /* See if there is PHI defining return value. */ | |
1564 | for (psi = gsi_start_phis (return_bb); | |
1565 | !gsi_end_p (psi); gsi_next (&psi)) | |
538dd0b7 | 1566 | if (!virtual_operand_p (gimple_phi_result (psi.phi ()))) |
6938f93f JH |
1567 | break; |
1568 | ||
1569 | /* When there is PHI, just update its value. */ | |
3e485f62 JH |
1570 | if (TREE_CODE (retval) == SSA_NAME |
1571 | && !gsi_end_p (psi)) | |
538dd0b7 | 1572 | add_phi_arg (psi.phi (), retval, e, UNKNOWN_LOCATION); |
6938f93f JH |
1573 | /* Otherwise update the return BB itself. |
1574 | find_return_bb allows at most one assignment to return value, | |
1575 | so update first statement. */ | |
1576 | else | |
3e485f62 | 1577 | { |
2094f1fc JH |
1578 | gimple_stmt_iterator bsi; |
1579 | for (bsi = gsi_start_bb (return_bb); !gsi_end_p (bsi); | |
1580 | gsi_next (&bsi)) | |
538dd0b7 DM |
1581 | if (greturn *return_stmt |
1582 | = dyn_cast <greturn *> (gsi_stmt (bsi))) | |
2094f1fc | 1583 | { |
538dd0b7 | 1584 | gimple_return_set_retval (return_stmt, retval); |
2094f1fc JH |
1585 | break; |
1586 | } | |
2e216592 JJ |
1587 | else if (gimple_code (gsi_stmt (bsi)) == GIMPLE_ASSIGN |
1588 | && !gimple_clobber_p (gsi_stmt (bsi))) | |
2094f1fc JH |
1589 | { |
1590 | gimple_assign_set_rhs1 (gsi_stmt (bsi), retval); | |
1591 | break; | |
1592 | } | |
1593 | update_stmt (gsi_stmt (bsi)); | |
b69539cb JJ |
1594 | /* Also adjust clobbers and debug stmts in return_bb. */ |
1595 | for (bsi = gsi_start_bb (return_bb); !gsi_end_p (bsi); | |
1596 | gsi_next (&bsi)) | |
1597 | { | |
1598 | gimple *stmt = gsi_stmt (bsi); | |
1599 | if (gimple_clobber_p (stmt) | |
1600 | || is_gimple_debug (stmt)) | |
1601 | { | |
1602 | ssa_op_iter iter; | |
1603 | use_operand_p use_p; | |
1604 | bool update = false; | |
1605 | FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, | |
1606 | SSA_OP_USE) | |
1607 | if (USE_FROM_PTR (use_p) == real_retval) | |
1608 | { | |
1609 | SET_USE (use_p, retval); | |
1610 | update = true; | |
1611 | } | |
1612 | if (update) | |
1613 | update_stmt (stmt); | |
1614 | } | |
1615 | } | |
3e485f62 | 1616 | } |
d5e254e1 IE |
1617 | |
1618 | /* Replace retbnd with new one. */ | |
1619 | if (retbnd) | |
1620 | { | |
1621 | gimple_stmt_iterator bsi; | |
1622 | for (bsi = gsi_last_bb (return_bb); !gsi_end_p (bsi); | |
1623 | gsi_prev (&bsi)) | |
1624 | if (gimple_code (gsi_stmt (bsi)) == GIMPLE_RETURN) | |
1625 | { | |
1626 | retbnd = copy_ssa_name (retbnd, call); | |
1627 | gimple_return_set_retbnd (gsi_stmt (bsi), retbnd); | |
1628 | update_stmt (gsi_stmt (bsi)); | |
1629 | break; | |
1630 | } | |
1631 | } | |
3e485f62 | 1632 | } |
556e9ba0 | 1633 | if (DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) |
42b05b6e RG |
1634 | { |
1635 | gimple_call_set_lhs (call, build_simple_mem_ref (retval)); | |
1636 | gsi_insert_after (&gsi, call, GSI_NEW_STMT); | |
1637 | } | |
556e9ba0 | 1638 | else |
42b05b6e RG |
1639 | { |
1640 | tree restype; | |
1641 | restype = TREE_TYPE (DECL_RESULT (current_function_decl)); | |
1642 | gsi_insert_after (&gsi, call, GSI_NEW_STMT); | |
1643 | if (!useless_type_conversion_p (TREE_TYPE (retval), restype)) | |
1644 | { | |
355fe088 | 1645 | gimple *cpy; |
b731b390 | 1646 | tree tem = create_tmp_reg (restype); |
42b05b6e | 1647 | tem = make_ssa_name (tem, call); |
0d0e4a03 | 1648 | cpy = gimple_build_assign (retval, NOP_EXPR, tem); |
42b05b6e RG |
1649 | gsi_insert_after (&gsi, cpy, GSI_NEW_STMT); |
1650 | retval = tem; | |
1651 | } | |
d5e254e1 IE |
1652 | /* Build bndret call to obtain returned bounds. */ |
1653 | if (retbnd) | |
4f5a53cc | 1654 | chkp_insert_retbnd_call (retbnd, retval, &gsi); |
42b05b6e RG |
1655 | gimple_call_set_lhs (call, retval); |
1656 | update_stmt (call); | |
1657 | } | |
3e485f62 | 1658 | } |
42b05b6e RG |
1659 | else |
1660 | gsi_insert_after (&gsi, call, GSI_NEW_STMT); | |
bfd71482 JJ |
1661 | if (tsan_func_exit_call) |
1662 | gsi_insert_after (&gsi, tsan_func_exit_call, GSI_NEW_STMT); | |
3e485f62 | 1663 | } |
6938f93f JH |
1664 | /* We don't use return block (there is either no return in function or |
1665 | multiple of them). So create new basic block with return statement. | |
1666 | */ | |
3e485f62 JH |
1667 | else |
1668 | { | |
538dd0b7 | 1669 | greturn *ret; |
241a2b9e JH |
1670 | if (split_point->split_part_set_retval |
1671 | && !VOID_TYPE_P (TREE_TYPE (TREE_TYPE (current_function_decl)))) | |
3e485f62 | 1672 | { |
4021f4a1 | 1673 | retval = DECL_RESULT (current_function_decl); |
8a9c1ae6 | 1674 | |
d5e254e1 IE |
1675 | if (chkp_function_instrumented_p (current_function_decl) |
1676 | && BOUNDED_P (retval)) | |
b731b390 | 1677 | retbnd = create_tmp_reg (pointer_bounds_type_node); |
d5e254e1 | 1678 | |
8a9c1ae6 JH |
1679 | /* We use temporary register to hold value when aggregate_value_p |
1680 | is false. Similarly for DECL_BY_REFERENCE we must avoid extra | |
1681 | copy. */ | |
1682 | if (!aggregate_value_p (retval, TREE_TYPE (current_function_decl)) | |
1683 | && !DECL_BY_REFERENCE (retval)) | |
b731b390 | 1684 | retval = create_tmp_reg (TREE_TYPE (retval)); |
3e485f62 | 1685 | if (is_gimple_reg (retval)) |
6938f93f JH |
1686 | { |
1687 | /* When returning by reference, there is only one SSA name | |
1688 | assigned to RESULT_DECL (that is pointer to return value). | |
1689 | Look it up or create new one if it is missing. */ | |
1690 | if (DECL_BY_REFERENCE (retval)) | |
32244553 | 1691 | retval = get_or_create_ssa_default_def (cfun, retval); |
6938f93f JH |
1692 | /* Otherwise produce new SSA name for return value. */ |
1693 | else | |
1694 | retval = make_ssa_name (retval, call); | |
1695 | } | |
556e9ba0 JH |
1696 | if (DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) |
1697 | gimple_call_set_lhs (call, build_simple_mem_ref (retval)); | |
1698 | else | |
1699 | gimple_call_set_lhs (call, retval); | |
a162f3af JJ |
1700 | gsi_insert_after (&gsi, call, GSI_NEW_STMT); |
1701 | } | |
1702 | else | |
1703 | { | |
1704 | gsi_insert_after (&gsi, call, GSI_NEW_STMT); | |
1705 | if (retval | |
1706 | && is_gimple_reg_type (TREE_TYPE (retval)) | |
1707 | && !is_gimple_val (retval)) | |
1708 | { | |
1709 | gassign *g | |
1710 | = gimple_build_assign (make_ssa_name (TREE_TYPE (retval)), | |
1711 | retval); | |
1712 | retval = gimple_assign_lhs (g); | |
1713 | gsi_insert_after (&gsi, g, GSI_NEW_STMT); | |
1714 | } | |
3e485f62 | 1715 | } |
d5e254e1 IE |
1716 | /* Build bndret call to obtain returned bounds. */ |
1717 | if (retbnd) | |
4f5a53cc | 1718 | chkp_insert_retbnd_call (retbnd, retval, &gsi); |
bfd71482 JJ |
1719 | if (tsan_func_exit_call) |
1720 | gsi_insert_after (&gsi, tsan_func_exit_call, GSI_NEW_STMT); | |
3e485f62 JH |
1721 | ret = gimple_build_return (retval); |
1722 | gsi_insert_after (&gsi, ret, GSI_NEW_STMT); | |
1723 | } | |
1724 | } | |
1725 | free_dominance_info (CDI_DOMINATORS); | |
1726 | free_dominance_info (CDI_POST_DOMINATORS); | |
0bceb671 | 1727 | compute_fn_summary (node, true); |
3e485f62 JH |
1728 | } |
1729 | ||
1730 | /* Execute function splitting pass. */ | |
1731 | ||
1732 | static unsigned int | |
1733 | execute_split_functions (void) | |
1734 | { | |
1735 | gimple_stmt_iterator bsi; | |
1736 | basic_block bb; | |
c2e87766 JH |
1737 | sreal overall_time = 0; |
1738 | int overall_size = 0; | |
3e485f62 | 1739 | int todo = 0; |
d52f5295 | 1740 | struct cgraph_node *node = cgraph_node::get (current_function_decl); |
3e485f62 | 1741 | |
b2d2adc6 RG |
1742 | if (flags_from_decl_or_type (current_function_decl) |
1743 | & (ECF_NORETURN|ECF_MALLOC)) | |
3e485f62 JH |
1744 | { |
1745 | if (dump_file) | |
b2d2adc6 | 1746 | fprintf (dump_file, "Not splitting: noreturn/malloc function.\n"); |
3e485f62 JH |
1747 | return 0; |
1748 | } | |
1749 | if (MAIN_NAME_P (DECL_NAME (current_function_decl))) | |
1750 | { | |
1751 | if (dump_file) | |
1752 | fprintf (dump_file, "Not splitting: main function.\n"); | |
1753 | return 0; | |
1754 | } | |
1755 | /* This can be relaxed; function might become inlinable after splitting | |
1756 | away the uninlinable part. */ | |
0bceb671 JH |
1757 | if (ipa_fn_summaries |
1758 | && !ipa_fn_summaries->get (node)->inlinable) | |
3e485f62 JH |
1759 | { |
1760 | if (dump_file) | |
1761 | fprintf (dump_file, "Not splitting: not inlinable.\n"); | |
1762 | return 0; | |
1763 | } | |
67348ccc | 1764 | if (DECL_DISREGARD_INLINE_LIMITS (node->decl)) |
3e485f62 JH |
1765 | { |
1766 | if (dump_file) | |
ed7656f6 | 1767 | fprintf (dump_file, "Not splitting: disregarding inline limits.\n"); |
3e485f62 JH |
1768 | return 0; |
1769 | } | |
1770 | /* This can be relaxed; most of versioning tests actually prevents | |
1771 | a duplication. */ | |
1772 | if (!tree_versionable_function_p (current_function_decl)) | |
1773 | { | |
1774 | if (dump_file) | |
1775 | fprintf (dump_file, "Not splitting: not versionable.\n"); | |
1776 | return 0; | |
1777 | } | |
1778 | /* FIXME: we could support this. */ | |
1779 | if (DECL_STRUCT_FUNCTION (current_function_decl)->static_chain_decl) | |
1780 | { | |
1781 | if (dump_file) | |
1782 | fprintf (dump_file, "Not splitting: nested function.\n"); | |
1783 | return 0; | |
1784 | } | |
3e485f62 JH |
1785 | |
1786 | /* See if it makes sense to try to split. | |
1787 | It makes sense to split if we inline, that is if we have direct calls to | |
1788 | handle or direct calls are possibly going to appear as result of indirect | |
cf9712cc JH |
1789 | inlining or LTO. Also handle -fprofile-generate as LTO to allow non-LTO |
1790 | training for LTO -fprofile-use build. | |
1791 | ||
3e485f62 JH |
1792 | Note that we are not completely conservative about disqualifying functions |
1793 | called once. It is possible that the caller is called more then once and | |
1794 | then inlining would still benefit. */ | |
c91061e6 JH |
1795 | if ((!node->callers |
1796 | /* Local functions called once will be completely inlined most of time. */ | |
1797 | || (!node->callers->next_caller && node->local.local)) | |
67348ccc | 1798 | && !node->address_taken |
42685f72 | 1799 | && !node->has_aliases_p () |
67348ccc | 1800 | && (!flag_lto || !node->externally_visible)) |
3e485f62 JH |
1801 | { |
1802 | if (dump_file) | |
1803 | fprintf (dump_file, "Not splitting: not called directly " | |
1804 | "or called once.\n"); | |
1805 | return 0; | |
1806 | } | |
1807 | ||
1808 | /* FIXME: We can actually split if splitting reduces call overhead. */ | |
1809 | if (!flag_inline_small_functions | |
1810 | && !DECL_DECLARED_INLINE_P (current_function_decl)) | |
1811 | { | |
1812 | if (dump_file) | |
1813 | fprintf (dump_file, "Not splitting: not autoinlining and function" | |
1814 | " is not inline.\n"); | |
1815 | return 0; | |
1816 | } | |
1817 | ||
e70670cf JH |
1818 | /* We enforce splitting after loop headers when profile info is not |
1819 | available. */ | |
0a6a6ac9 | 1820 | if (profile_status_for_fn (cfun) != PROFILE_READ) |
e70670cf JH |
1821 | mark_dfs_back_edges (); |
1822 | ||
b2e25729 BS |
1823 | /* Initialize bitmap to track forbidden calls. */ |
1824 | forbidden_dominators = BITMAP_ALLOC (NULL); | |
1825 | calculate_dominance_info (CDI_DOMINATORS); | |
1826 | ||
3e485f62 | 1827 | /* Compute local info about basic blocks and determine function size/time. */ |
8b1c6fd7 | 1828 | bb_info_vec.safe_grow_cleared (last_basic_block_for_fn (cfun) + 1); |
c2e87766 | 1829 | best_split_point.split_bbs = NULL; |
bfd71482 JJ |
1830 | basic_block return_bb = find_return_bb (); |
1831 | int tsan_exit_found = -1; | |
11cd3bed | 1832 | FOR_EACH_BB_FN (bb, cfun) |
3e485f62 | 1833 | { |
c2e87766 | 1834 | sreal time = 0; |
3e485f62 | 1835 | int size = 0; |
c2e87766 JH |
1836 | sreal freq = bb->count.to_sreal_scale |
1837 | (ENTRY_BLOCK_PTR_FOR_FN (cfun)->count); | |
3e485f62 JH |
1838 | |
1839 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1840 | fprintf (dump_file, "Basic block %i\n", bb->index); | |
1841 | ||
1842 | for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
1843 | { | |
c2e87766 JH |
1844 | sreal this_time; |
1845 | int this_size; | |
355fe088 | 1846 | gimple *stmt = gsi_stmt (bsi); |
3e485f62 JH |
1847 | |
1848 | this_size = estimate_num_insns (stmt, &eni_size_weights); | |
c2e87766 JH |
1849 | this_time = (sreal)estimate_num_insns (stmt, &eni_time_weights) |
1850 | * freq; | |
3e485f62 JH |
1851 | size += this_size; |
1852 | time += this_time; | |
b2e25729 | 1853 | check_forbidden_calls (stmt); |
3e485f62 JH |
1854 | |
1855 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1856 | { | |
c2e87766 JH |
1857 | fprintf (dump_file, " freq:%4.2f size:%3i time:%4.2f ", |
1858 | freq.to_double (), this_size, this_time.to_double ()); | |
ef6cb4c7 | 1859 | print_gimple_stmt (dump_file, stmt, 0); |
3e485f62 | 1860 | } |
bfd71482 JJ |
1861 | |
1862 | if ((flag_sanitize & SANITIZE_THREAD) | |
8e4284d0 | 1863 | && gimple_call_internal_p (stmt, IFN_TSAN_FUNC_EXIT)) |
bfd71482 JJ |
1864 | { |
1865 | /* We handle TSAN_FUNC_EXIT for splitting either in the | |
1866 | return_bb, or in its immediate predecessors. */ | |
1867 | if ((bb != return_bb && !find_edge (bb, return_bb)) | |
1868 | || (tsan_exit_found != -1 | |
1869 | && tsan_exit_found != (bb != return_bb))) | |
1870 | { | |
1871 | if (dump_file) | |
1872 | fprintf (dump_file, "Not splitting: TSAN_FUNC_EXIT" | |
1873 | " in unexpected basic block.\n"); | |
1874 | BITMAP_FREE (forbidden_dominators); | |
1875 | bb_info_vec.release (); | |
1876 | return 0; | |
1877 | } | |
1878 | tsan_exit_found = bb != return_bb; | |
1879 | } | |
3e485f62 JH |
1880 | } |
1881 | overall_time += time; | |
1882 | overall_size += size; | |
9771b263 DN |
1883 | bb_info_vec[bb->index].time = time; |
1884 | bb_info_vec[bb->index].size = size; | |
3e485f62 | 1885 | } |
bfd71482 | 1886 | find_split_points (return_bb, overall_time, overall_size); |
3e485f62 JH |
1887 | if (best_split_point.split_bbs) |
1888 | { | |
bfd71482 | 1889 | split_function (return_bb, &best_split_point, tsan_exit_found == 1); |
3e485f62 JH |
1890 | BITMAP_FREE (best_split_point.ssa_names_to_pass); |
1891 | BITMAP_FREE (best_split_point.split_bbs); | |
1892 | todo = TODO_update_ssa | TODO_cleanup_cfg; | |
1893 | } | |
b2e25729 | 1894 | BITMAP_FREE (forbidden_dominators); |
9771b263 | 1895 | bb_info_vec.release (); |
3e485f62 JH |
1896 | return todo; |
1897 | } | |
1898 | ||
27a4cd48 DM |
1899 | namespace { |
1900 | ||
1901 | const pass_data pass_data_split_functions = | |
3e485f62 | 1902 | { |
27a4cd48 DM |
1903 | GIMPLE_PASS, /* type */ |
1904 | "fnsplit", /* name */ | |
1905 | OPTGROUP_NONE, /* optinfo_flags */ | |
27a4cd48 DM |
1906 | TV_IPA_FNSPLIT, /* tv_id */ |
1907 | PROP_cfg, /* properties_required */ | |
1908 | 0, /* properties_provided */ | |
1909 | 0, /* properties_destroyed */ | |
1910 | 0, /* todo_flags_start */ | |
3bea341f | 1911 | 0, /* todo_flags_finish */ |
3e485f62 | 1912 | }; |
cf9712cc | 1913 | |
27a4cd48 DM |
1914 | class pass_split_functions : public gimple_opt_pass |
1915 | { | |
1916 | public: | |
c3284718 RS |
1917 | pass_split_functions (gcc::context *ctxt) |
1918 | : gimple_opt_pass (pass_data_split_functions, ctxt) | |
27a4cd48 DM |
1919 | {} |
1920 | ||
1921 | /* opt_pass methods: */ | |
1a3d085c | 1922 | virtual bool gate (function *); |
be55bfe6 TS |
1923 | virtual unsigned int execute (function *) |
1924 | { | |
1925 | return execute_split_functions (); | |
1926 | } | |
27a4cd48 DM |
1927 | |
1928 | }; // class pass_split_functions | |
1929 | ||
1a3d085c TS |
1930 | bool |
1931 | pass_split_functions::gate (function *) | |
1932 | { | |
1933 | /* When doing profile feedback, we want to execute the pass after profiling | |
1934 | is read. So disable one in early optimization. */ | |
1935 | return (flag_partial_inlining | |
1936 | && !profile_arc_flag && !flag_branch_probabilities); | |
1937 | } | |
1938 | ||
27a4cd48 DM |
1939 | } // anon namespace |
1940 | ||
1941 | gimple_opt_pass * | |
1942 | make_pass_split_functions (gcc::context *ctxt) | |
1943 | { | |
1944 | return new pass_split_functions (ctxt); | |
1945 | } | |
1946 | ||
cf9712cc JH |
1947 | /* Execute function splitting pass. */ |
1948 | ||
1949 | static unsigned int | |
1950 | execute_feedback_split_functions (void) | |
1951 | { | |
1952 | unsigned int retval = execute_split_functions (); | |
1953 | if (retval) | |
1954 | retval |= TODO_rebuild_cgraph_edges; | |
1955 | return retval; | |
1956 | } | |
1957 | ||
27a4cd48 DM |
1958 | namespace { |
1959 | ||
1960 | const pass_data pass_data_feedback_split_functions = | |
cf9712cc | 1961 | { |
27a4cd48 DM |
1962 | GIMPLE_PASS, /* type */ |
1963 | "feedback_fnsplit", /* name */ | |
1964 | OPTGROUP_NONE, /* optinfo_flags */ | |
27a4cd48 DM |
1965 | TV_IPA_FNSPLIT, /* tv_id */ |
1966 | PROP_cfg, /* properties_required */ | |
1967 | 0, /* properties_provided */ | |
1968 | 0, /* properties_destroyed */ | |
1969 | 0, /* todo_flags_start */ | |
3bea341f | 1970 | 0, /* todo_flags_finish */ |
cf9712cc | 1971 | }; |
27a4cd48 DM |
1972 | |
1973 | class pass_feedback_split_functions : public gimple_opt_pass | |
1974 | { | |
1975 | public: | |
c3284718 RS |
1976 | pass_feedback_split_functions (gcc::context *ctxt) |
1977 | : gimple_opt_pass (pass_data_feedback_split_functions, ctxt) | |
27a4cd48 DM |
1978 | {} |
1979 | ||
1980 | /* opt_pass methods: */ | |
1a3d085c | 1981 | virtual bool gate (function *); |
be55bfe6 TS |
1982 | virtual unsigned int execute (function *) |
1983 | { | |
1984 | return execute_feedback_split_functions (); | |
1985 | } | |
27a4cd48 DM |
1986 | |
1987 | }; // class pass_feedback_split_functions | |
1988 | ||
1a3d085c TS |
1989 | bool |
1990 | pass_feedback_split_functions::gate (function *) | |
1991 | { | |
1992 | /* We don't need to split when profiling at all, we are producing | |
1993 | lousy code anyway. */ | |
1994 | return (flag_partial_inlining | |
1995 | && flag_branch_probabilities); | |
1996 | } | |
1997 | ||
27a4cd48 DM |
1998 | } // anon namespace |
1999 | ||
2000 | gimple_opt_pass * | |
2001 | make_pass_feedback_split_functions (gcc::context *ctxt) | |
2002 | { | |
2003 | return new pass_feedback_split_functions (ctxt); | |
2004 | } |