]>
Commit | Line | Data |
---|---|---|
2862cf88 | 1 | /* Function splitting pass |
aad93da1 | 2 | Copyright (C) 2010-2017 Free Software Foundation, Inc. |
2862cf88 | 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, | |
574c4ab5 | 48 | but not fund.part leads to performance improvement similar as inlining |
2862cf88 | 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" | |
9ef16211 | 80 | #include "backend.h" |
7c29e30e | 81 | #include "rtl.h" |
b20a8bb4 | 82 | #include "tree.h" |
9ef16211 | 83 | #include "gimple.h" |
7c29e30e | 84 | #include "cfghooks.h" |
85 | #include "alloc-pool.h" | |
86 | #include "tree-pass.h" | |
9ef16211 | 87 | #include "ssa.h" |
7c29e30e | 88 | #include "cgraph.h" |
89 | #include "diagnostic.h" | |
b20a8bb4 | 90 | #include "fold-const.h" |
94ea8568 | 91 | #include "cfganal.h" |
9ed99284 | 92 | #include "calls.h" |
a8783bee | 93 | #include "gimplify.h" |
dcf1a1ec | 94 | #include "gimple-iterator.h" |
e795d6e1 | 95 | #include "gimplify-me.h" |
dcf1a1ec | 96 | #include "gimple-walk.h" |
2cc80ac3 | 97 | #include "symbol-summary.h" |
2862cf88 | 98 | #include "ipa-prop.h" |
073c1fd5 | 99 | #include "tree-cfg.h" |
073c1fd5 | 100 | #include "tree-into-ssa.h" |
101 | #include "tree-dfa.h" | |
2862cf88 | 102 | #include "tree-inline.h" |
2862cf88 | 103 | #include "params.h" |
104 | #include "gimple-pretty-print.h" | |
b9a58fc5 | 105 | #include "ipa-fnsummary.h" |
f6568ea4 | 106 | #include "cfgloop.h" |
058a1b7a | 107 | #include "tree-chkp.h" |
2862cf88 | 108 | |
109 | /* Per basic block info. */ | |
110 | ||
df8eb490 | 111 | struct split_bb_info |
2862cf88 | 112 | { |
113 | unsigned int size; | |
96824d35 | 114 | sreal time; |
df8eb490 | 115 | }; |
2862cf88 | 116 | |
9908fe4d | 117 | static vec<split_bb_info> bb_info_vec; |
2862cf88 | 118 | |
119 | /* Description of split point. */ | |
120 | ||
121 | struct split_point | |
122 | { | |
123 | /* Size of the partitions. */ | |
96824d35 | 124 | sreal header_time, split_time; |
125 | unsigned int header_size, split_size; | |
2862cf88 | 126 | |
574c4ab5 | 127 | /* SSA names that need to be passed into spit function. */ |
2862cf88 | 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 | ||
bfb98797 | 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 | ||
2862cf88 | 137 | /* Basic blocks we are splitting away. */ |
138 | bitmap split_bbs; | |
b04bab7c | 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; | |
2862cf88 | 143 | }; |
144 | ||
145 | /* Best split point found. */ | |
146 | ||
147 | struct split_point best_split_point; | |
148 | ||
aa2edcb9 | 149 | /* Set of basic blocks that are not allowed to dominate a split point. */ |
150 | ||
151 | static bitmap forbidden_dominators; | |
152 | ||
b04bab7c | 153 | static tree find_retval (basic_block return_bb); |
058a1b7a | 154 | static tree find_retbnd (basic_block return_bb); |
b04bab7c | 155 | |
2dc0a33e | 156 | /* Callback for walk_stmt_load_store_addr_ops. If T is non-SSA automatic |
2862cf88 | 157 | variable, check it if it is present in bitmap passed via DATA. */ |
158 | ||
159 | static bool | |
42acab1c | 160 | test_nonssa_use (gimple *, tree t, tree, void *data) |
2862cf88 | 161 | { |
162 | t = get_base_address (t); | |
163 | ||
2dc0a33e | 164 | if (!t || is_gimple_reg (t)) |
165 | return false; | |
166 | ||
167 | if (TREE_CODE (t) == PARM_DECL | |
53e9c5c4 | 168 | || (VAR_P (t) |
2862cf88 | 169 | && auto_var_in_fn_p (t, current_function_decl)) |
2dc0a33e | 170 | || TREE_CODE (t) == RESULT_DECL |
3c659c08 | 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))) | |
2862cf88 | 176 | return bitmap_bit_p ((bitmap)data, DECL_UID (t)); |
b04bab7c | 177 | |
2dc0a33e | 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)) | |
b04bab7c | 181 | && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME |
ec11736b | 182 | && SSA_NAME_VAR (TREE_OPERAND (t, 0)) |
b04bab7c | 183 | && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND (t, 0))) == RESULT_DECL |
184 | && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) | |
2dc0a33e | 185 | return |
186 | bitmap_bit_p ((bitmap)data, | |
187 | DECL_UID (DECL_RESULT (current_function_decl))); | |
188 | ||
2862cf88 | 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, | |
1d0b727d | 198 | "Split point at BB %i\n" |
96824d35 | 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); | |
2862cf88 | 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 | ||
2dc0a33e | 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. | |
4493dab3 | 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); | |
1e094109 | 218 | vec<basic_block> worklist = vNULL; |
4493dab3 | 219 | edge e; |
220 | edge_iterator ei; | |
221 | bool ok = true; | |
3c659c08 | 222 | basic_block bb; |
2dc0a33e | 223 | |
4493dab3 | 224 | FOR_EACH_EDGE (e, ei, current->entry_bb->preds) |
34154e27 | 225 | if (e->src != ENTRY_BLOCK_PTR_FOR_FN (cfun) |
4493dab3 | 226 | && !bitmap_bit_p (current->split_bbs, e->src->index)) |
227 | { | |
f1f41a6c | 228 | worklist.safe_push (e->src); |
4493dab3 | 229 | bitmap_set_bit (seen, e->src->index); |
230 | } | |
2dc0a33e | 231 | |
f1f41a6c | 232 | while (!worklist.is_empty ()) |
4493dab3 | 233 | { |
3c659c08 | 234 | bb = worklist.pop (); |
4493dab3 | 235 | FOR_EACH_EDGE (e, ei, bb->preds) |
34154e27 | 236 | if (e->src != ENTRY_BLOCK_PTR_FOR_FN (cfun) |
6ef9bbe0 | 237 | && bitmap_set_bit (seen, e->src->index)) |
4493dab3 | 238 | { |
239 | gcc_checking_assert (!bitmap_bit_p (current->split_bbs, | |
240 | e->src->index)); | |
f1f41a6c | 241 | worklist.safe_push (e->src); |
4493dab3 | 242 | } |
1a91d914 | 243 | for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi); |
244 | gsi_next (&bsi)) | |
4493dab3 | 245 | { |
42acab1c | 246 | gimple *stmt = gsi_stmt (bsi); |
2dc0a33e | 247 | if (is_gimple_debug (stmt)) |
4493dab3 | 248 | continue; |
249 | if (walk_stmt_load_store_addr_ops | |
2dc0a33e | 250 | (stmt, non_ssa_vars, test_nonssa_use, test_nonssa_use, |
251 | test_nonssa_use)) | |
4493dab3 | 252 | { |
253 | ok = false; | |
254 | goto done; | |
255 | } | |
1a91d914 | 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 | } | |
4493dab3 | 263 | } |
1a91d914 | 264 | for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi); |
265 | gsi_next (&bsi)) | |
4493dab3 | 266 | { |
267 | if (walk_stmt_load_store_addr_ops | |
2dc0a33e | 268 | (gsi_stmt (bsi), non_ssa_vars, test_nonssa_use, test_nonssa_use, |
269 | test_nonssa_use)) | |
4493dab3 | 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; | |
1a91d914 | 279 | for (gphi_iterator bsi = gsi_start_phis (return_bb); |
280 | !gsi_end_p (bsi); | |
4493dab3 | 281 | gsi_next (&bsi)) |
282 | { | |
1a91d914 | 283 | gphi *stmt = bsi.phi (); |
4493dab3 | 284 | tree op = gimple_phi_arg_def (stmt, e->dest_idx); |
285 | ||
7c782c9b | 286 | if (virtual_operand_p (gimple_phi_result (stmt))) |
4493dab3 | 287 | continue; |
288 | if (TREE_CODE (op) != SSA_NAME | |
5b26a9e3 | 289 | && test_nonssa_use (stmt, op, op, non_ssa_vars)) |
4493dab3 | 290 | { |
291 | ok = false; | |
292 | goto done; | |
293 | } | |
294 | } | |
295 | } | |
296 | } | |
3c659c08 | 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)) | |
1a91d914 | 306 | if (glabel *label_stmt = dyn_cast <glabel *> (gsi_stmt (bsi))) |
3c659c08 | 307 | { |
1a91d914 | 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 | } | |
3c659c08 | 315 | } |
1a91d914 | 316 | else |
3c659c08 | 317 | break; |
318 | } | |
319 | ||
4493dab3 | 320 | done: |
321 | BITMAP_FREE (seen); | |
f1f41a6c | 322 | worklist.release (); |
4493dab3 | 323 | return ok; |
324 | } | |
325 | ||
aa2edcb9 | 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 | |
42acab1c | 336 | check_forbidden_calls (gimple *stmt) |
aa2edcb9 | 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; | |
1a91d914 | 358 | gcond *use_stmt; |
aa2edcb9 | 359 | |
1a91d914 | 360 | use_stmt = dyn_cast <gcond *> (USE_STMT (use_p)); |
361 | if (!use_stmt) | |
aa2edcb9 | 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 | { | |
f5a6b05f | 397 | if (dominated_by_p (CDI_DOMINATORS, bb, |
398 | BASIC_BLOCK_FOR_FN (cfun, dom_bb))) | |
aa2edcb9 | 399 | return true; |
400 | } | |
401 | ||
402 | return false; | |
403 | } | |
404 | ||
058a1b7a | 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 | ||
2862cf88 | 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; | |
1a91d914 | 433 | gphi_iterator bsi; |
6a69e813 | 434 | unsigned int i; |
b04bab7c | 435 | tree retval; |
058a1b7a | 436 | tree retbnd; |
15ca8f90 | 437 | bool back_edge = false; |
6a69e813 | 438 | |
2862cf88 | 439 | if (dump_file && (dump_flags & TDF_DETAILS)) |
440 | dump_split_point (dump_file, current); | |
441 | ||
bfb98797 | 442 | current->count = profile_count::zero (); |
6a69e813 | 443 | FOR_EACH_EDGE (e, ei, current->entry_bb->preds) |
15ca8f90 | 444 | { |
445 | if (e->flags & EDGE_DFS_BACK) | |
446 | back_edge = true; | |
447 | if (!bitmap_bit_p (current->split_bbs, e->src->index)) | |
bfb98797 | 448 | current->count += e->count (); |
15ca8f90 | 449 | } |
6a69e813 | 450 | |
bfb98797 | 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)))) | |
2862cf88 | 457 | { |
15ca8f90 | 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 | |
f26d8580 | 464 | && profile_status_for_fn (cfun) != PROFILE_READ |
bfb98797 | 465 | && current->count |
466 | < ENTRY_BLOCK_PTR_FOR_FN (cfun)->count) | |
15ca8f90 | 467 | { |
468 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
bfb98797 | 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 | } | |
15ca8f90 | 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 | } | |
2862cf88 | 484 | } |
485 | ||
486 | if (!current->header_size) | |
487 | { | |
488 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
489 | fprintf (dump_file, " Refused: header empty\n"); | |
2862cf88 | 490 | return; |
491 | } | |
492 | ||
574c4ab5 | 493 | /* Verify that PHI args on entry are either virtual or all their operands |
494 | incoming from header are the same. */ | |
6a69e813 | 495 | for (bsi = gsi_start_phis (current->entry_bb); !gsi_end_p (bsi); gsi_next (&bsi)) |
2862cf88 | 496 | { |
1a91d914 | 497 | gphi *stmt = bsi.phi (); |
6a69e813 | 498 | tree val = NULL; |
499 | ||
7c782c9b | 500 | if (virtual_operand_p (gimple_phi_result (stmt))) |
6a69e813 | 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 | } | |
2862cf88 | 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; | |
1767a056 | 525 | parm = DECL_CHAIN (parm)) |
2862cf88 | 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 | } | |
c6dfe037 | 537 | else |
2862cf88 | 538 | { |
c6dfe037 | 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))) | |
96c7c99c | 545 | call_overhead += estimate_move_cost (TREE_TYPE (parm), false); |
c6dfe037 | 546 | num_args++; |
547 | } | |
2862cf88 | 548 | } |
549 | } | |
550 | if (!VOID_TYPE_P (TREE_TYPE (current_function_decl))) | |
96c7c99c | 551 | call_overhead += estimate_move_cost (TREE_TYPE (current_function_decl), |
552 | false); | |
2862cf88 | 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 | ||
7037b790 | 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 | ||
2862cf88 | 597 | /* FIXME: we currently can pass only SSA function parameters to the split |
fd8d648f | 598 | arguments. Once parm_adjustment infrastructure is supported by cloning, |
2862cf88 | 599 | we can pass more than that. */ |
600 | if (num_args != bitmap_count_bits (current->ssa_names_to_pass)) | |
601 | { | |
6a69e813 | 602 | |
2862cf88 | 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. */ | |
4493dab3 | 612 | if (!bitmap_empty_p (non_ssa_vars) |
613 | && !verify_non_ssa_vars (current, non_ssa_vars, return_bb)) | |
2862cf88 | 614 | { |
4493dab3 | 615 | if (dump_file && (dump_flags & TDF_DETAILS)) |
616 | fprintf (dump_file, | |
617 | " Refused: split part has non-ssa uses\n"); | |
2862cf88 | 618 | return; |
619 | } | |
aa2edcb9 | 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 | ||
b04bab7c | 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) | |
9181a4ec | 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 | } | |
b04bab7c | 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 | |
ec11736b | 661 | && SSA_NAME_VAR (retval) |
b04bab7c | 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 | |
058a1b7a | 668 | = split_part_set_ssa_name_p (retval, current, return_bb); |
b04bab7c | 669 | else if (TREE_CODE (retval) == PARM_DECL) |
670 | current->split_part_set_retval = false; | |
53e9c5c4 | 671 | else if (VAR_P (retval) |
b04bab7c | 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 | ||
058a1b7a | 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 | ||
9aa9c1e6 | 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. */ | |
34154e27 | 703 | if (return_bb != EXIT_BLOCK_PTR_FOR_FN (cfun)) |
9aa9c1e6 | 704 | { |
1a91d914 | 705 | gphi_iterator psi; |
9aa9c1e6 | 706 | |
707 | for (psi = gsi_start_phis (return_bb); !gsi_end_p (psi); gsi_next (&psi)) | |
1a91d914 | 708 | if (!virtual_operand_p (gimple_phi_result (psi.phi ())) |
9aa9c1e6 | 709 | && !(retval |
710 | && current->split_part_set_retval | |
711 | && TREE_CODE (retval) == SSA_NAME | |
712 | && !DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)) | |
1a91d914 | 713 | && SSA_NAME_DEF_STMT (retval) == psi.phi ())) |
9aa9c1e6 | 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 | ||
bfb98797 | 725 | /* At the moment chose split point with lowest count and that leaves |
2862cf88 | 726 | out smallest size of header. |
727 | In future we might re-consider this heuristics. */ | |
728 | if (!best_split_point.split_bbs | |
bfb98797 | 729 | || best_split_point.count |
730 | > current->count | |
731 | || (best_split_point.count == current->count | |
2862cf88 | 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 | ||
4493dab3 | 751 | /* Return basic block containing RETURN statement. We allow basic blocks |
752 | of the form: | |
753 | <retval> = tmp_var; | |
754 | return <retval> | |
a63cfc4d | 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). | |
efee62d1 | 757 | If nothing is found, return the exit block. |
4493dab3 | 758 | |
2862cf88 | 759 | When there are multiple RETURN statement, chose one with return value, |
760 | since that one is more likely shared by multiple code paths. | |
4493dab3 | 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 | ||
2862cf88 | 766 | TODO: We might support multiple return blocks. */ |
767 | ||
768 | static basic_block | |
769 | find_return_bb (void) | |
770 | { | |
771 | edge e; | |
34154e27 | 772 | basic_block return_bb = EXIT_BLOCK_PTR_FOR_FN (cfun); |
b234c94d | 773 | gimple_stmt_iterator bsi; |
774 | bool found_return = false; | |
775 | tree retval = NULL_TREE; | |
2862cf88 | 776 | |
34154e27 | 777 | if (!single_pred_p (EXIT_BLOCK_PTR_FOR_FN (cfun))) |
b234c94d | 778 | return return_bb; |
779 | ||
34154e27 | 780 | e = single_pred_edge (EXIT_BLOCK_PTR_FOR_FN (cfun)); |
b234c94d | 781 | for (bsi = gsi_last_bb (e->src); !gsi_end_p (bsi); gsi_prev (&bsi)) |
782 | { | |
42acab1c | 783 | gimple *stmt = gsi_stmt (bsi); |
1a3c6423 | 784 | if (gimple_code (stmt) == GIMPLE_LABEL |
785 | || is_gimple_debug (stmt) | |
786 | || gimple_clobber_p (stmt)) | |
b234c94d | 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 | ; | |
1a91d914 | 796 | else if (greturn *return_stmt = dyn_cast <greturn *> (stmt)) |
b234c94d | 797 | { |
798 | found_return = true; | |
1a91d914 | 799 | retval = gimple_return_retval (return_stmt); |
b234c94d | 800 | } |
a63cfc4d | 801 | /* For -fsanitize=thread, allow also TSAN_FUNC_EXIT () in the return |
802 | bb. */ | |
803 | else if ((flag_sanitize & SANITIZE_THREAD) | |
7408cd7d | 804 | && gimple_call_internal_p (stmt, IFN_TSAN_FUNC_EXIT)) |
a63cfc4d | 805 | ; |
b234c94d | 806 | else |
807 | break; | |
808 | } | |
809 | if (gsi_end_p (bsi) && found_return) | |
810 | return_bb = e->src; | |
2862cf88 | 811 | |
2862cf88 | 812 | return return_bb; |
813 | } | |
814 | ||
574c4ab5 | 815 | /* Given return basic block RETURN_BB, see where return value is really |
4493dab3 | 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)) | |
1a91d914 | 822 | if (greturn *return_stmt = dyn_cast <greturn *> (gsi_stmt (bsi))) |
823 | return gimple_return_retval (return_stmt); | |
1a3c6423 | 824 | else if (gimple_code (gsi_stmt (bsi)) == GIMPLE_ASSIGN |
825 | && !gimple_clobber_p (gsi_stmt (bsi))) | |
4493dab3 | 826 | return gimple_assign_rhs1 (gsi_stmt (bsi)); |
827 | return NULL; | |
828 | } | |
829 | ||
058a1b7a | 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 | ||
2dc0a33e | 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. | |
2862cf88 | 844 | Return true when access to T prevents splitting the function. */ |
845 | ||
846 | static bool | |
42acab1c | 847 | mark_nonssa_use (gimple *, tree t, tree, void *data) |
2862cf88 | 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)) | |
2dc0a33e | 859 | fprintf (dump_file, |
860 | "Cannot split: use of non-ssa function parameter.\n"); | |
2862cf88 | 861 | return true; |
862 | } | |
863 | ||
53e9c5c4 | 864 | if ((VAR_P (t) && auto_var_in_fn_p (t, current_function_decl)) |
2dc0a33e | 865 | || TREE_CODE (t) == RESULT_DECL |
53e9c5c4 | 866 | || (TREE_CODE (t) == LABEL_DECL && FORCED_LABEL (t))) |
2862cf88 | 867 | bitmap_set_bit ((bitmap)data, DECL_UID (t)); |
b04bab7c | 868 | |
2dc0a33e | 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)) | |
b04bab7c | 872 | && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME |
ec11736b | 873 | && SSA_NAME_VAR (TREE_OPERAND (t, 0)) |
b04bab7c | 874 | && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND (t, 0))) == RESULT_DECL |
875 | && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) | |
2dc0a33e | 876 | return |
877 | bitmap_bit_p ((bitmap)data, | |
878 | DECL_UID (DECL_RESULT (current_function_decl))); | |
879 | ||
2862cf88 | 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 | { | |
2862cf88 | 898 | edge e; |
899 | edge_iterator ei; | |
900 | bool can_split = true; | |
901 | ||
1a91d914 | 902 | for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi); |
903 | gsi_next (&bsi)) | |
2862cf88 | 904 | { |
42acab1c | 905 | gimple *stmt = gsi_stmt (bsi); |
2862cf88 | 906 | tree op; |
907 | ssa_op_iter iter; | |
908 | tree decl; | |
909 | ||
910 | if (is_gimple_debug (stmt)) | |
911 | continue; | |
912 | ||
1a3c6423 | 913 | if (gimple_clobber_p (stmt)) |
914 | continue; | |
915 | ||
2862cf88 | 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. */ | |
892c141e | 921 | if (gimple_code (stmt) == GIMPLE_RESX) |
2862cf88 | 922 | { |
923 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
892c141e | 924 | fprintf (dump_file, "Cannot split: resx.\n"); |
2862cf88 | 925 | can_split = false; |
926 | } | |
927 | if (gimple_code (stmt) == GIMPLE_EH_DISPATCH) | |
928 | { | |
929 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2dc0a33e | 930 | fprintf (dump_file, "Cannot split: eh dispatch.\n"); |
2862cf88 | 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: | |
3c97c75d | 947 | case BUILT_IN_APPLY_ARGS: |
2862cf88 | 948 | case BUILT_IN_VA_START: |
949 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2dc0a33e | 950 | fprintf (dump_file, |
951 | "Cannot split: builtin_apply and va_start.\n"); | |
2862cf88 | 952 | can_split = false; |
953 | break; | |
954 | case BUILT_IN_EH_POINTER: | |
955 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2dc0a33e | 956 | fprintf (dump_file, "Cannot split: builtin_eh_pointer.\n"); |
2862cf88 | 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 | } | |
1a91d914 | 972 | for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi); |
973 | gsi_next (&bsi)) | |
2862cf88 | 974 | { |
1a91d914 | 975 | gphi *stmt = bsi.phi (); |
6a69e813 | 976 | unsigned int i; |
2862cf88 | 977 | |
7c782c9b | 978 | if (virtual_operand_p (gimple_phi_result (stmt))) |
2862cf88 | 979 | continue; |
6a69e813 | 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 | } | |
2862cf88 | 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 | } | |
574c4ab5 | 993 | /* Record also uses coming from PHI operand in return BB. */ |
2862cf88 | 994 | FOR_EACH_EDGE (e, ei, bb->succs) |
995 | if (e->dest == return_bb) | |
996 | { | |
1a91d914 | 997 | for (gphi_iterator bsi = gsi_start_phis (return_bb); |
998 | !gsi_end_p (bsi); | |
999 | gsi_next (&bsi)) | |
2862cf88 | 1000 | { |
1a91d914 | 1001 | gphi *stmt = bsi.phi (); |
2862cf88 | 1002 | tree op = gimple_phi_arg_def (stmt, e->dest_idx); |
1003 | ||
7c782c9b | 1004 | if (virtual_operand_p (gimple_phi_result (stmt))) |
2862cf88 | 1005 | continue; |
2862cf88 | 1006 | if (TREE_CODE (op) == SSA_NAME) |
1007 | bitmap_set_bit (used_ssa_names, SSA_NAME_VERSION (op)); | |
1008 | else | |
5b26a9e3 | 1009 | can_split &= !mark_nonssa_use (stmt, op, op, non_ssa_vars); |
2862cf88 | 1010 | } |
2862cf88 | 1011 | } |
1012 | return can_split; | |
1013 | } | |
1014 | ||
1015 | /* Stack entry for recursive DFS walk in find_split_point. */ | |
1016 | ||
df8eb490 | 1017 | struct stack_entry |
2862cf88 | 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, | |
574c4ab5 | 1031 | the value is up to sum of incoming and outgoing edges of BB. */ |
2862cf88 | 1032 | unsigned int edge_num; |
1033 | ||
1034 | /* Stack entry index of earliest BB reachable from current BB | |
574c4ab5 | 1035 | or any BB visited later in DFS walk. */ |
2862cf88 | 1036 | int earliest; |
1037 | ||
1038 | /* Overall time and size of all BBs reached from this BB in DFS walk. */ | |
96824d35 | 1039 | sreal overall_time; |
1040 | int overall_size; | |
2862cf88 | 1041 | |
1042 | /* When false we can not split on this BB. */ | |
1043 | bool can_split; | |
df8eb490 | 1044 | }; |
2862cf88 | 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 | |
96824d35 | 1065 | find_split_points (basic_block return_bb, sreal overall_time, int overall_size) |
2862cf88 | 1066 | { |
1067 | stack_entry first; | |
1e094109 | 1068 | vec<stack_entry> stack = vNULL; |
2862cf88 | 1069 | basic_block bb; |
2862cf88 | 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 | ||
34154e27 | 1078 | first.bb = ENTRY_BLOCK_PTR_FOR_FN (cfun); |
2862cf88 | 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; | |
60141df0 | 1085 | first.non_ssa_vars = 0; |
2862cf88 | 1086 | first.bbs_visited = 0; |
60141df0 | 1087 | first.can_split = false; |
f1f41a6c | 1088 | stack.safe_push (first); |
34154e27 | 1089 | ENTRY_BLOCK_PTR_FOR_FN (cfun)->aux = (void *)(intptr_t)-1; |
2862cf88 | 1090 | |
f1f41a6c | 1091 | while (!stack.is_empty ()) |
2862cf88 | 1092 | { |
f1f41a6c | 1093 | stack_entry *entry = &stack.last (); |
2862cf88 | 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) | |
34154e27 | 1100 | && entry->bb != ENTRY_BLOCK_PTR_FOR_FN (cfun)) |
2862cf88 | 1101 | { |
f1f41a6c | 1102 | int pos = stack.length (); |
2862cf88 | 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. */ | |
34154e27 | 1152 | if (dest != return_bb && dest != EXIT_BLOCK_PTR_FOR_FN (cfun) |
2862cf88 | 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 | |
f1f41a6c | 1160 | = bb_info_vec[dest->index].time; |
2862cf88 | 1161 | new_entry.overall_size |
f1f41a6c | 1162 | = bb_info_vec[dest->index].size; |
2862cf88 | 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); | |
f1f41a6c | 1170 | stack.safe_push (new_entry); |
1171 | dest->aux = (void *)(intptr_t)stack.length (); | |
2862cf88 | 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 | } | |
574c4ab5 | 1178 | /* We are done with examining the edges. Pop off the value from stack |
1179 | and merge stuff we accumulate during the walk. */ | |
34154e27 | 1180 | else if (entry->bb != ENTRY_BLOCK_PTR_FOR_FN (cfun)) |
2862cf88 | 1181 | { |
f1f41a6c | 1182 | stack_entry *prev = &stack[stack.length () - 2]; |
2862cf88 | 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); | |
f1f41a6c | 1201 | stack.pop (); |
2862cf88 | 1202 | } |
1203 | else | |
f1f41a6c | 1204 | stack.pop (); |
2862cf88 | 1205 | } |
34154e27 | 1206 | ENTRY_BLOCK_PTR_FOR_FN (cfun)->aux = NULL; |
fc00614f | 1207 | FOR_EACH_BB_FN (bb, cfun) |
2862cf88 | 1208 | bb->aux = NULL; |
f1f41a6c | 1209 | stack.release (); |
2862cf88 | 1210 | BITMAP_FREE (current.ssa_names_to_pass); |
1211 | } | |
1212 | ||
1213 | /* Split function at SPLIT_POINT. */ | |
1214 | ||
1215 | static void | |
a63cfc4d | 1216 | split_function (basic_block return_bb, struct split_point *split_point, |
1217 | bool add_tsan_func_exit) | |
2862cf88 | 1218 | { |
1e094109 | 1219 | vec<tree> args_to_pass = vNULL; |
982ffd8d | 1220 | bitmap args_to_skip; |
2862cf88 | 1221 | tree parm; |
1222 | int num = 0; | |
415d1b9a | 1223 | cgraph_node *node, *cur_node = cgraph_node::get (current_function_decl); |
2862cf88 | 1224 | basic_block call_bb; |
a63cfc4d | 1225 | gcall *call, *tsan_func_exit_call = NULL; |
2862cf88 | 1226 | edge e; |
1227 | edge_iterator ei; | |
058a1b7a | 1228 | tree retval = NULL, real_retval = NULL, retbnd = NULL; |
058a1b7a | 1229 | bool with_bounds = chkp_function_instrumented_p (current_function_decl); |
42acab1c | 1230 | gimple *last_stmt = NULL; |
e65564ab | 1231 | unsigned int i; |
c6dfe037 | 1232 | tree arg, ddef; |
2862cf88 | 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 | ||
982ffd8d | 1240 | if (cur_node->local.can_change_signature) |
1241 | args_to_skip = BITMAP_ALLOC (NULL); | |
1242 | else | |
1243 | args_to_skip = NULL; | |
1244 | ||
2862cf88 | 1245 | /* Collect the parameters of new function and args_to_skip bitmap. */ |
1246 | for (parm = DECL_ARGUMENTS (current_function_decl); | |
1767a056 | 1247 | parm; parm = DECL_CHAIN (parm), num++) |
982ffd8d | 1248 | if (args_to_skip |
1249 | && (!is_gimple_reg (parm) | |
c6dfe037 | 1250 | || (ddef = ssa_default_def (cfun, parm)) == NULL_TREE |
982ffd8d | 1251 | || !bitmap_bit_p (split_point->ssa_names_to_pass, |
c6dfe037 | 1252 | SSA_NAME_VERSION (ddef)))) |
2862cf88 | 1253 | bitmap_set_bit (args_to_skip, num); |
1254 | else | |
e65564ab | 1255 | { |
17c5f86d | 1256 | /* This parm might not have been used up to now, but is going to be |
1257 | used, hence register it. */ | |
17c5f86d | 1258 | if (is_gimple_reg (parm)) |
c6dfe037 | 1259 | arg = get_or_create_ssa_default_def (cfun, parm); |
17c5f86d | 1260 | else |
1261 | arg = parm; | |
982ffd8d | 1262 | |
0ed25048 | 1263 | if (!useless_type_conversion_p (DECL_ARG_TYPE (parm), TREE_TYPE (arg))) |
1264 | arg = fold_convert (DECL_ARG_TYPE (parm), arg); | |
f1f41a6c | 1265 | args_to_pass.safe_push (arg); |
e65564ab | 1266 | } |
2862cf88 | 1267 | |
34947478 | 1268 | /* See if the split function will return. */ |
a206617d | 1269 | bool split_part_return_p = false; |
2862cf88 | 1270 | FOR_EACH_EDGE (e, ei, return_bb->preds) |
a206617d | 1271 | { |
1272 | if (bitmap_bit_p (split_point->split_bbs, e->src->index)) | |
1273 | split_part_return_p = true; | |
a206617d | 1274 | } |
2862cf88 | 1275 | |
b04bab7c | 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. */ | |
34154e27 | 1281 | else if (return_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)) |
b04bab7c | 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. */ | |
34947478 | 1288 | else if ((retval = find_retval (return_bb)) |
1289 | && !split_point->split_part_set_retval) | |
b04bab7c | 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); | |
25bd54a7 | 1295 | new_return_bb->count = profile_count::zero (); |
b04bab7c | 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 | { | |
205ce1aa | 1302 | new_return_bb->count += e->count (); |
b04bab7c | 1303 | redirect_edge_and_branch (e, new_return_bb); |
1304 | redirected = true; | |
1305 | break; | |
1306 | } | |
1307 | } | |
720cfc43 | 1308 | e = make_single_succ_edge (new_return_bb, EXIT_BLOCK_PTR_FOR_FN (cfun), 0); |
b3083327 | 1309 | add_bb_to_loop (new_return_bb, current_loops->tree_root); |
b04bab7c | 1310 | bitmap_set_bit (split_point->split_bbs, new_return_bb->index); |
bca33fb8 | 1311 | retbnd = find_retbnd (return_bb); |
c13d6d8f | 1312 | } |
1313 | /* When we pass around the value, use existing return block. */ | |
1314 | else | |
34947478 | 1315 | { |
1316 | bitmap_set_bit (split_point->split_bbs, return_bb->index); | |
1317 | retbnd = find_retbnd (return_bb); | |
1318 | } | |
a206617d | 1319 | |
c13d6d8f | 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 | |
1d0b727d | 1322 | tree-inline is not able to compensate for. |
c13d6d8f | 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. */ | |
34154e27 | 1326 | if (return_bb != EXIT_BLOCK_PTR_FOR_FN (cfun)) |
c13d6d8f | 1327 | { |
1d0b727d | 1328 | bool phi_p = false; |
1a91d914 | 1329 | for (gphi_iterator gsi = gsi_start_phis (return_bb); |
1330 | !gsi_end_p (gsi);) | |
b04bab7c | 1331 | { |
1a91d914 | 1332 | gphi *stmt = gsi.phi (); |
7c782c9b | 1333 | if (!virtual_operand_p (gimple_phi_result (stmt))) |
c13d6d8f | 1334 | { |
1335 | gsi_next (&gsi); | |
1336 | continue; | |
1337 | } | |
904ad591 | 1338 | mark_virtual_phi_result_for_renaming (stmt); |
1339 | remove_phi_node (&gsi, true); | |
1d0b727d | 1340 | phi_p = true; |
b04bab7c | 1341 | } |
1d0b727d | 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) | |
1a91d914 | 1351 | for (gimple_stmt_iterator gsi = gsi_start_bb (return_bb); |
1352 | !gsi_end_p (gsi); | |
1353 | gsi_next (&gsi)) | |
1d0b727d | 1354 | { |
42acab1c | 1355 | gimple *stmt = gsi_stmt (gsi); |
1d0b727d | 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 | } | |
b04bab7c | 1364 | } |
2862cf88 | 1365 | |
1366 | /* Now create the actual clone. */ | |
35ee1c66 | 1367 | cgraph_edge::rebuild_edges (); |
415d1b9a | 1368 | node = cur_node->create_version_clone_with_body |
e5d1ea31 | 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"); | |
af48f0b1 | 1372 | |
71e37927 | 1373 | node->split_part = true; |
1374 | ||
236594e6 | 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 | ||
af48f0b1 | 1384 | /* Let's take a time profile for splitted function. */ |
1385 | node->tp_first_run = cur_node->tp_first_run + 1; | |
1386 | ||
fd8d648f | 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. */ | |
02774f2d | 1390 | if (DECL_BUILT_IN (node->decl)) |
fd8d648f | 1391 | { |
02774f2d | 1392 | DECL_BUILT_IN_CLASS (node->decl) = NOT_BUILT_IN; |
1393 | DECL_FUNCTION_CODE (node->decl) = (enum built_in_function) 0; | |
fd8d648f | 1394 | } |
058a1b7a | 1395 | |
34947478 | 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 | ||
058a1b7a | 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 | ||
ea03e543 | 1439 | /* If the original function is declared inline, there is no point in issuing |
1440 | a warning for the non-inlinable part. */ | |
02774f2d | 1441 | DECL_NO_INLINE_WARNING_P (node->decl) = 1; |
415d1b9a | 1442 | cur_node->remove_callees (); |
51ce5652 | 1443 | cur_node->remove_all_references (); |
2862cf88 | 1444 | if (!split_part_return_p) |
02774f2d | 1445 | TREE_THIS_VOLATILE (node->decl) = 1; |
2862cf88 | 1446 | if (dump_file) |
02774f2d | 1447 | dump_function_to_file (node->decl, dump_file, dump_flags); |
2862cf88 | 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; | |
1a91d914 | 1452 | for (gimple_stmt_iterator gsi = gsi_start_bb (call_bb); !gsi_end_p (gsi);) |
2862cf88 | 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; | |
bfb98797 | 1460 | call_bb->count = split_point->count; |
2862cf88 | 1461 | e = split_block (split_point->entry_bb, last_stmt); |
1462 | remove_edge (e); | |
1463 | ||
1464 | /* Produce the call statement. */ | |
1a91d914 | 1465 | gimple_stmt_iterator gsi = gsi_last_bb (call_bb); |
f1f41a6c | 1466 | FOR_EACH_VEC_ELT (args_to_pass, i, arg) |
0ed25048 | 1467 | if (!is_gimple_val (arg)) |
1468 | { | |
1469 | arg = force_gimple_operand_gsi (&gsi, arg, true, NULL_TREE, | |
ad676199 | 1470 | false, GSI_CONTINUE_LINKING); |
f1f41a6c | 1471 | args_to_pass[i] = arg; |
0ed25048 | 1472 | } |
02774f2d | 1473 | call = gimple_build_call_vec (node->decl, args_to_pass); |
058a1b7a | 1474 | gimple_call_set_with_bounds (call, with_bounds); |
2862cf88 | 1475 | gimple_set_block (call, DECL_INITIAL (current_function_decl)); |
f1f41a6c | 1476 | args_to_pass.release (); |
2862cf88 | 1477 | |
eee0ea7c | 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) | |
eee0ea7c | 1485 | { |
6ec840b0 | 1486 | vec<tree, va_gc> **debug_args = NULL; |
1487 | unsigned i = 0, len = 0; | |
c64f38bf | 1488 | if (MAY_HAVE_DEBUG_BIND_STMTS) |
eee0ea7c | 1489 | { |
6ec840b0 | 1490 | debug_args = decl_debug_args_lookup (node->decl); |
1491 | if (debug_args) | |
1492 | len = vec_safe_length (*debug_args); | |
eee0ea7c | 1493 | } |
6ec840b0 | 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 | ||
c64f38bf | 1501 | /* This needs to be done even without |
1502 | MAY_HAVE_DEBUG_BIND_STMTS, otherwise if it didn't exist | |
1503 | before, we'd end up with different SSA_NAME_VERSIONs | |
1504 | between -g and -g0. */ | |
6ec840b0 | 1505 | arg = get_or_create_ssa_default_def (cfun, parm); |
c64f38bf | 1506 | if (!MAY_HAVE_DEBUG_BIND_STMTS || debug_args == NULL) |
6ec840b0 | 1507 | continue; |
1508 | ||
1509 | while (i < len && (**debug_args)[i] != DECL_ORIGIN (parm)) | |
1510 | i += 2; | |
1511 | if (i >= len) | |
1512 | continue; | |
1513 | ddecl = (**debug_args)[i + 1]; | |
1514 | def_temp | |
1515 | = gimple_build_debug_bind (ddecl, unshare_expr (arg), call); | |
1516 | gsi_insert_after (&gsi, def_temp, GSI_NEW_STMT); | |
1517 | } | |
eee0ea7c | 1518 | } |
1519 | ||
a8005893 | 1520 | /* We avoid address being taken on any variable used by split part, |
1521 | so return slot optimization is always possible. Moreover this is | |
1522 | required to make DECL_BY_REFERENCE work. */ | |
1523 | if (aggregate_value_p (DECL_RESULT (current_function_decl), | |
8ad1eb69 | 1524 | TREE_TYPE (current_function_decl)) |
1525 | && (!is_gimple_reg_type (TREE_TYPE (DECL_RESULT (current_function_decl))) | |
1526 | || DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)))) | |
a8005893 | 1527 | gimple_call_set_return_slot_opt (call, true); |
1528 | ||
a63cfc4d | 1529 | if (add_tsan_func_exit) |
1530 | tsan_func_exit_call = gimple_build_call_internal (IFN_TSAN_FUNC_EXIT, 0); | |
1531 | ||
2862cf88 | 1532 | /* Update return value. This is bit tricky. When we do not return, |
1533 | do nothing. When we return we might need to update return_bb | |
1534 | or produce a new return statement. */ | |
1535 | if (!split_part_return_p) | |
a63cfc4d | 1536 | { |
1537 | gsi_insert_after (&gsi, call, GSI_NEW_STMT); | |
1538 | if (tsan_func_exit_call) | |
1539 | gsi_insert_after (&gsi, tsan_func_exit_call, GSI_NEW_STMT); | |
1540 | } | |
2862cf88 | 1541 | else |
1542 | { | |
720cfc43 | 1543 | e = make_single_succ_edge (call_bb, return_bb, |
1544 | return_bb == EXIT_BLOCK_PTR_FOR_FN (cfun) | |
1545 | ? 0 : EDGE_FALLTHRU); | |
524a0531 | 1546 | |
1547 | /* If there is return basic block, see what value we need to store | |
1548 | return value into and put call just before it. */ | |
34154e27 | 1549 | if (return_bb != EXIT_BLOCK_PTR_FOR_FN (cfun)) |
2862cf88 | 1550 | { |
34947478 | 1551 | real_retval = retval; |
b04bab7c | 1552 | if (real_retval && split_point->split_part_set_retval) |
2862cf88 | 1553 | { |
1a91d914 | 1554 | gphi_iterator psi; |
2862cf88 | 1555 | |
524a0531 | 1556 | /* See if we need new SSA_NAME for the result. |
1557 | When DECL_BY_REFERENCE is true, retval is actually pointer to | |
1558 | return value and it is constant in whole function. */ | |
1559 | if (TREE_CODE (retval) == SSA_NAME | |
1560 | && !DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) | |
2862cf88 | 1561 | { |
874117c8 | 1562 | retval = copy_ssa_name (retval, call); |
524a0531 | 1563 | |
1564 | /* See if there is PHI defining return value. */ | |
1565 | for (psi = gsi_start_phis (return_bb); | |
1566 | !gsi_end_p (psi); gsi_next (&psi)) | |
1a91d914 | 1567 | if (!virtual_operand_p (gimple_phi_result (psi.phi ()))) |
524a0531 | 1568 | break; |
1569 | ||
1570 | /* When there is PHI, just update its value. */ | |
2862cf88 | 1571 | if (TREE_CODE (retval) == SSA_NAME |
1572 | && !gsi_end_p (psi)) | |
1a91d914 | 1573 | add_phi_arg (psi.phi (), retval, e, UNKNOWN_LOCATION); |
524a0531 | 1574 | /* Otherwise update the return BB itself. |
1575 | find_return_bb allows at most one assignment to return value, | |
1576 | so update first statement. */ | |
1577 | else | |
2862cf88 | 1578 | { |
4493dab3 | 1579 | gimple_stmt_iterator bsi; |
1580 | for (bsi = gsi_start_bb (return_bb); !gsi_end_p (bsi); | |
1581 | gsi_next (&bsi)) | |
1a91d914 | 1582 | if (greturn *return_stmt |
1583 | = dyn_cast <greturn *> (gsi_stmt (bsi))) | |
4493dab3 | 1584 | { |
1a91d914 | 1585 | gimple_return_set_retval (return_stmt, retval); |
4493dab3 | 1586 | break; |
1587 | } | |
0f41c0cf | 1588 | else if (gimple_code (gsi_stmt (bsi)) == GIMPLE_ASSIGN |
1589 | && !gimple_clobber_p (gsi_stmt (bsi))) | |
4493dab3 | 1590 | { |
1591 | gimple_assign_set_rhs1 (gsi_stmt (bsi), retval); | |
1592 | break; | |
1593 | } | |
1594 | update_stmt (gsi_stmt (bsi)); | |
34947478 | 1595 | /* Also adjust clobbers and debug stmts in return_bb. */ |
1596 | for (bsi = gsi_start_bb (return_bb); !gsi_end_p (bsi); | |
1597 | gsi_next (&bsi)) | |
1598 | { | |
1599 | gimple *stmt = gsi_stmt (bsi); | |
1600 | if (gimple_clobber_p (stmt) | |
1601 | || is_gimple_debug (stmt)) | |
1602 | { | |
1603 | ssa_op_iter iter; | |
1604 | use_operand_p use_p; | |
1605 | bool update = false; | |
1606 | FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, | |
1607 | SSA_OP_USE) | |
1608 | if (USE_FROM_PTR (use_p) == real_retval) | |
1609 | { | |
1610 | SET_USE (use_p, retval); | |
1611 | update = true; | |
1612 | } | |
1613 | if (update) | |
1614 | update_stmt (stmt); | |
1615 | } | |
1616 | } | |
2862cf88 | 1617 | } |
058a1b7a | 1618 | |
1619 | /* Replace retbnd with new one. */ | |
1620 | if (retbnd) | |
1621 | { | |
1622 | gimple_stmt_iterator bsi; | |
1623 | for (bsi = gsi_last_bb (return_bb); !gsi_end_p (bsi); | |
1624 | gsi_prev (&bsi)) | |
1625 | if (gimple_code (gsi_stmt (bsi)) == GIMPLE_RETURN) | |
1626 | { | |
1627 | retbnd = copy_ssa_name (retbnd, call); | |
1628 | gimple_return_set_retbnd (gsi_stmt (bsi), retbnd); | |
1629 | update_stmt (gsi_stmt (bsi)); | |
1630 | break; | |
1631 | } | |
1632 | } | |
2862cf88 | 1633 | } |
a8005893 | 1634 | if (DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) |
9ece8f57 | 1635 | { |
1636 | gimple_call_set_lhs (call, build_simple_mem_ref (retval)); | |
1637 | gsi_insert_after (&gsi, call, GSI_NEW_STMT); | |
1638 | } | |
a8005893 | 1639 | else |
9ece8f57 | 1640 | { |
1641 | tree restype; | |
1642 | restype = TREE_TYPE (DECL_RESULT (current_function_decl)); | |
1643 | gsi_insert_after (&gsi, call, GSI_NEW_STMT); | |
1644 | if (!useless_type_conversion_p (TREE_TYPE (retval), restype)) | |
1645 | { | |
42acab1c | 1646 | gimple *cpy; |
f9e245b2 | 1647 | tree tem = create_tmp_reg (restype); |
9ece8f57 | 1648 | tem = make_ssa_name (tem, call); |
e9cf809e | 1649 | cpy = gimple_build_assign (retval, NOP_EXPR, tem); |
9ece8f57 | 1650 | gsi_insert_after (&gsi, cpy, GSI_NEW_STMT); |
1651 | retval = tem; | |
1652 | } | |
058a1b7a | 1653 | /* Build bndret call to obtain returned bounds. */ |
1654 | if (retbnd) | |
17c596cc | 1655 | chkp_insert_retbnd_call (retbnd, retval, &gsi); |
9ece8f57 | 1656 | gimple_call_set_lhs (call, retval); |
1657 | update_stmt (call); | |
1658 | } | |
2862cf88 | 1659 | } |
9ece8f57 | 1660 | else |
1661 | gsi_insert_after (&gsi, call, GSI_NEW_STMT); | |
a63cfc4d | 1662 | if (tsan_func_exit_call) |
1663 | gsi_insert_after (&gsi, tsan_func_exit_call, GSI_NEW_STMT); | |
2862cf88 | 1664 | } |
524a0531 | 1665 | /* We don't use return block (there is either no return in function or |
1666 | multiple of them). So create new basic block with return statement. | |
1667 | */ | |
2862cf88 | 1668 | else |
1669 | { | |
1a91d914 | 1670 | greturn *ret; |
b04bab7c | 1671 | if (split_point->split_part_set_retval |
1672 | && !VOID_TYPE_P (TREE_TYPE (TREE_TYPE (current_function_decl)))) | |
2862cf88 | 1673 | { |
1b268210 | 1674 | retval = DECL_RESULT (current_function_decl); |
9eb0b8ee | 1675 | |
058a1b7a | 1676 | if (chkp_function_instrumented_p (current_function_decl) |
1677 | && BOUNDED_P (retval)) | |
f9e245b2 | 1678 | retbnd = create_tmp_reg (pointer_bounds_type_node); |
058a1b7a | 1679 | |
9eb0b8ee | 1680 | /* We use temporary register to hold value when aggregate_value_p |
1681 | is false. Similarly for DECL_BY_REFERENCE we must avoid extra | |
1682 | copy. */ | |
1683 | if (!aggregate_value_p (retval, TREE_TYPE (current_function_decl)) | |
1684 | && !DECL_BY_REFERENCE (retval)) | |
f9e245b2 | 1685 | retval = create_tmp_reg (TREE_TYPE (retval)); |
2862cf88 | 1686 | if (is_gimple_reg (retval)) |
524a0531 | 1687 | { |
1688 | /* When returning by reference, there is only one SSA name | |
1689 | assigned to RESULT_DECL (that is pointer to return value). | |
1690 | Look it up or create new one if it is missing. */ | |
1691 | if (DECL_BY_REFERENCE (retval)) | |
c6dfe037 | 1692 | retval = get_or_create_ssa_default_def (cfun, retval); |
524a0531 | 1693 | /* Otherwise produce new SSA name for return value. */ |
1694 | else | |
1695 | retval = make_ssa_name (retval, call); | |
1696 | } | |
a8005893 | 1697 | if (DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) |
1698 | gimple_call_set_lhs (call, build_simple_mem_ref (retval)); | |
1699 | else | |
1700 | gimple_call_set_lhs (call, retval); | |
394a86f3 | 1701 | gsi_insert_after (&gsi, call, GSI_NEW_STMT); |
1702 | } | |
1703 | else | |
1704 | { | |
1705 | gsi_insert_after (&gsi, call, GSI_NEW_STMT); | |
1706 | if (retval | |
1707 | && is_gimple_reg_type (TREE_TYPE (retval)) | |
1708 | && !is_gimple_val (retval)) | |
1709 | { | |
1710 | gassign *g | |
1711 | = gimple_build_assign (make_ssa_name (TREE_TYPE (retval)), | |
1712 | retval); | |
1713 | retval = gimple_assign_lhs (g); | |
1714 | gsi_insert_after (&gsi, g, GSI_NEW_STMT); | |
1715 | } | |
2862cf88 | 1716 | } |
058a1b7a | 1717 | /* Build bndret call to obtain returned bounds. */ |
1718 | if (retbnd) | |
17c596cc | 1719 | chkp_insert_retbnd_call (retbnd, retval, &gsi); |
a63cfc4d | 1720 | if (tsan_func_exit_call) |
1721 | gsi_insert_after (&gsi, tsan_func_exit_call, GSI_NEW_STMT); | |
2862cf88 | 1722 | ret = gimple_build_return (retval); |
1723 | gsi_insert_after (&gsi, ret, GSI_NEW_STMT); | |
1724 | } | |
1725 | } | |
1726 | free_dominance_info (CDI_DOMINATORS); | |
1727 | free_dominance_info (CDI_POST_DOMINATORS); | |
1297cbcd | 1728 | compute_fn_summary (node, true); |
2862cf88 | 1729 | } |
1730 | ||
1731 | /* Execute function splitting pass. */ | |
1732 | ||
1733 | static unsigned int | |
1734 | execute_split_functions (void) | |
1735 | { | |
1736 | gimple_stmt_iterator bsi; | |
1737 | basic_block bb; | |
96824d35 | 1738 | sreal overall_time = 0; |
1739 | int overall_size = 0; | |
2862cf88 | 1740 | int todo = 0; |
415d1b9a | 1741 | struct cgraph_node *node = cgraph_node::get (current_function_decl); |
2862cf88 | 1742 | |
f3941af5 | 1743 | if (flags_from_decl_or_type (current_function_decl) |
1744 | & (ECF_NORETURN|ECF_MALLOC)) | |
2862cf88 | 1745 | { |
1746 | if (dump_file) | |
f3941af5 | 1747 | fprintf (dump_file, "Not splitting: noreturn/malloc function.\n"); |
2862cf88 | 1748 | return 0; |
1749 | } | |
1750 | if (MAIN_NAME_P (DECL_NAME (current_function_decl))) | |
1751 | { | |
1752 | if (dump_file) | |
1753 | fprintf (dump_file, "Not splitting: main function.\n"); | |
1754 | return 0; | |
1755 | } | |
ed0831a9 | 1756 | if (node->frequency == NODE_FREQUENCY_UNLIKELY_EXECUTED) |
1757 | { | |
1758 | if (dump_file) | |
1759 | fprintf (dump_file, "Not splitting: function is unlikely executed.\n"); | |
1760 | return 0; | |
1761 | } | |
2862cf88 | 1762 | /* This can be relaxed; function might become inlinable after splitting |
1763 | away the uninlinable part. */ | |
1297cbcd | 1764 | if (ipa_fn_summaries |
1765 | && !ipa_fn_summaries->get (node)->inlinable) | |
2862cf88 | 1766 | { |
1767 | if (dump_file) | |
1768 | fprintf (dump_file, "Not splitting: not inlinable.\n"); | |
1769 | return 0; | |
1770 | } | |
02774f2d | 1771 | if (DECL_DISREGARD_INLINE_LIMITS (node->decl)) |
2862cf88 | 1772 | { |
1773 | if (dump_file) | |
574c4ab5 | 1774 | fprintf (dump_file, "Not splitting: disregarding inline limits.\n"); |
2862cf88 | 1775 | return 0; |
1776 | } | |
1777 | /* This can be relaxed; most of versioning tests actually prevents | |
1778 | a duplication. */ | |
1779 | if (!tree_versionable_function_p (current_function_decl)) | |
1780 | { | |
1781 | if (dump_file) | |
1782 | fprintf (dump_file, "Not splitting: not versionable.\n"); | |
1783 | return 0; | |
1784 | } | |
1785 | /* FIXME: we could support this. */ | |
1786 | if (DECL_STRUCT_FUNCTION (current_function_decl)->static_chain_decl) | |
1787 | { | |
1788 | if (dump_file) | |
1789 | fprintf (dump_file, "Not splitting: nested function.\n"); | |
1790 | return 0; | |
1791 | } | |
2862cf88 | 1792 | |
1793 | /* See if it makes sense to try to split. | |
1794 | It makes sense to split if we inline, that is if we have direct calls to | |
1795 | handle or direct calls are possibly going to appear as result of indirect | |
a15d5ede | 1796 | inlining or LTO. Also handle -fprofile-generate as LTO to allow non-LTO |
1797 | training for LTO -fprofile-use build. | |
1798 | ||
2862cf88 | 1799 | Note that we are not completely conservative about disqualifying functions |
1800 | called once. It is possible that the caller is called more then once and | |
1801 | then inlining would still benefit. */ | |
8745ccbf | 1802 | if ((!node->callers |
1803 | /* Local functions called once will be completely inlined most of time. */ | |
1804 | || (!node->callers->next_caller && node->local.local)) | |
02774f2d | 1805 | && !node->address_taken |
c670165e | 1806 | && !node->has_aliases_p () |
02774f2d | 1807 | && (!flag_lto || !node->externally_visible)) |
2862cf88 | 1808 | { |
1809 | if (dump_file) | |
1810 | fprintf (dump_file, "Not splitting: not called directly " | |
1811 | "or called once.\n"); | |
1812 | return 0; | |
1813 | } | |
1814 | ||
1815 | /* FIXME: We can actually split if splitting reduces call overhead. */ | |
1816 | if (!flag_inline_small_functions | |
1817 | && !DECL_DECLARED_INLINE_P (current_function_decl)) | |
1818 | { | |
1819 | if (dump_file) | |
1820 | fprintf (dump_file, "Not splitting: not autoinlining and function" | |
1821 | " is not inline.\n"); | |
1822 | return 0; | |
1823 | } | |
1824 | ||
15ca8f90 | 1825 | /* We enforce splitting after loop headers when profile info is not |
1826 | available. */ | |
f26d8580 | 1827 | if (profile_status_for_fn (cfun) != PROFILE_READ) |
15ca8f90 | 1828 | mark_dfs_back_edges (); |
1829 | ||
aa2edcb9 | 1830 | /* Initialize bitmap to track forbidden calls. */ |
1831 | forbidden_dominators = BITMAP_ALLOC (NULL); | |
1832 | calculate_dominance_info (CDI_DOMINATORS); | |
1833 | ||
2862cf88 | 1834 | /* Compute local info about basic blocks and determine function size/time. */ |
fe672ac0 | 1835 | bb_info_vec.safe_grow_cleared (last_basic_block_for_fn (cfun) + 1); |
96824d35 | 1836 | best_split_point.split_bbs = NULL; |
a63cfc4d | 1837 | basic_block return_bb = find_return_bb (); |
1838 | int tsan_exit_found = -1; | |
fc00614f | 1839 | FOR_EACH_BB_FN (bb, cfun) |
2862cf88 | 1840 | { |
96824d35 | 1841 | sreal time = 0; |
2862cf88 | 1842 | int size = 0; |
96824d35 | 1843 | sreal freq = bb->count.to_sreal_scale |
1844 | (ENTRY_BLOCK_PTR_FOR_FN (cfun)->count); | |
2862cf88 | 1845 | |
1846 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1847 | fprintf (dump_file, "Basic block %i\n", bb->index); | |
1848 | ||
1849 | for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
1850 | { | |
96824d35 | 1851 | sreal this_time; |
1852 | int this_size; | |
42acab1c | 1853 | gimple *stmt = gsi_stmt (bsi); |
2862cf88 | 1854 | |
1855 | this_size = estimate_num_insns (stmt, &eni_size_weights); | |
96824d35 | 1856 | this_time = (sreal)estimate_num_insns (stmt, &eni_time_weights) |
1857 | * freq; | |
2862cf88 | 1858 | size += this_size; |
1859 | time += this_time; | |
aa2edcb9 | 1860 | check_forbidden_calls (stmt); |
2862cf88 | 1861 | |
1862 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1863 | { | |
96824d35 | 1864 | fprintf (dump_file, " freq:%4.2f size:%3i time:%4.2f ", |
1865 | freq.to_double (), this_size, this_time.to_double ()); | |
1ffa4346 | 1866 | print_gimple_stmt (dump_file, stmt, 0); |
2862cf88 | 1867 | } |
a63cfc4d | 1868 | |
1869 | if ((flag_sanitize & SANITIZE_THREAD) | |
7408cd7d | 1870 | && gimple_call_internal_p (stmt, IFN_TSAN_FUNC_EXIT)) |
a63cfc4d | 1871 | { |
1872 | /* We handle TSAN_FUNC_EXIT for splitting either in the | |
1873 | return_bb, or in its immediate predecessors. */ | |
1874 | if ((bb != return_bb && !find_edge (bb, return_bb)) | |
1875 | || (tsan_exit_found != -1 | |
1876 | && tsan_exit_found != (bb != return_bb))) | |
1877 | { | |
1878 | if (dump_file) | |
1879 | fprintf (dump_file, "Not splitting: TSAN_FUNC_EXIT" | |
1880 | " in unexpected basic block.\n"); | |
1881 | BITMAP_FREE (forbidden_dominators); | |
1882 | bb_info_vec.release (); | |
1883 | return 0; | |
1884 | } | |
1885 | tsan_exit_found = bb != return_bb; | |
1886 | } | |
2862cf88 | 1887 | } |
1888 | overall_time += time; | |
1889 | overall_size += size; | |
f1f41a6c | 1890 | bb_info_vec[bb->index].time = time; |
1891 | bb_info_vec[bb->index].size = size; | |
2862cf88 | 1892 | } |
a63cfc4d | 1893 | find_split_points (return_bb, overall_time, overall_size); |
2862cf88 | 1894 | if (best_split_point.split_bbs) |
1895 | { | |
a63cfc4d | 1896 | split_function (return_bb, &best_split_point, tsan_exit_found == 1); |
2862cf88 | 1897 | BITMAP_FREE (best_split_point.ssa_names_to_pass); |
1898 | BITMAP_FREE (best_split_point.split_bbs); | |
1899 | todo = TODO_update_ssa | TODO_cleanup_cfg; | |
1900 | } | |
aa2edcb9 | 1901 | BITMAP_FREE (forbidden_dominators); |
f1f41a6c | 1902 | bb_info_vec.release (); |
2862cf88 | 1903 | return todo; |
1904 | } | |
1905 | ||
cbe8bda8 | 1906 | namespace { |
1907 | ||
1908 | const pass_data pass_data_split_functions = | |
2862cf88 | 1909 | { |
cbe8bda8 | 1910 | GIMPLE_PASS, /* type */ |
1911 | "fnsplit", /* name */ | |
1912 | OPTGROUP_NONE, /* optinfo_flags */ | |
cbe8bda8 | 1913 | TV_IPA_FNSPLIT, /* tv_id */ |
1914 | PROP_cfg, /* properties_required */ | |
1915 | 0, /* properties_provided */ | |
1916 | 0, /* properties_destroyed */ | |
1917 | 0, /* todo_flags_start */ | |
8b88439e | 1918 | 0, /* todo_flags_finish */ |
2862cf88 | 1919 | }; |
a15d5ede | 1920 | |
cbe8bda8 | 1921 | class pass_split_functions : public gimple_opt_pass |
1922 | { | |
1923 | public: | |
9af5ce0c | 1924 | pass_split_functions (gcc::context *ctxt) |
1925 | : gimple_opt_pass (pass_data_split_functions, ctxt) | |
cbe8bda8 | 1926 | {} |
1927 | ||
1928 | /* opt_pass methods: */ | |
31315c24 | 1929 | virtual bool gate (function *); |
65b0537f | 1930 | virtual unsigned int execute (function *) |
1931 | { | |
1932 | return execute_split_functions (); | |
1933 | } | |
cbe8bda8 | 1934 | |
1935 | }; // class pass_split_functions | |
1936 | ||
31315c24 | 1937 | bool |
1938 | pass_split_functions::gate (function *) | |
1939 | { | |
1940 | /* When doing profile feedback, we want to execute the pass after profiling | |
1941 | is read. So disable one in early optimization. */ | |
1942 | return (flag_partial_inlining | |
1943 | && !profile_arc_flag && !flag_branch_probabilities); | |
1944 | } | |
1945 | ||
cbe8bda8 | 1946 | } // anon namespace |
1947 | ||
1948 | gimple_opt_pass * | |
1949 | make_pass_split_functions (gcc::context *ctxt) | |
1950 | { | |
1951 | return new pass_split_functions (ctxt); | |
1952 | } | |
1953 | ||
a15d5ede | 1954 | /* Execute function splitting pass. */ |
1955 | ||
1956 | static unsigned int | |
1957 | execute_feedback_split_functions (void) | |
1958 | { | |
1959 | unsigned int retval = execute_split_functions (); | |
1960 | if (retval) | |
1961 | retval |= TODO_rebuild_cgraph_edges; | |
1962 | return retval; | |
1963 | } | |
1964 | ||
cbe8bda8 | 1965 | namespace { |
1966 | ||
1967 | const pass_data pass_data_feedback_split_functions = | |
a15d5ede | 1968 | { |
cbe8bda8 | 1969 | GIMPLE_PASS, /* type */ |
1970 | "feedback_fnsplit", /* name */ | |
1971 | OPTGROUP_NONE, /* optinfo_flags */ | |
cbe8bda8 | 1972 | TV_IPA_FNSPLIT, /* tv_id */ |
1973 | PROP_cfg, /* properties_required */ | |
1974 | 0, /* properties_provided */ | |
1975 | 0, /* properties_destroyed */ | |
1976 | 0, /* todo_flags_start */ | |
8b88439e | 1977 | 0, /* todo_flags_finish */ |
a15d5ede | 1978 | }; |
cbe8bda8 | 1979 | |
1980 | class pass_feedback_split_functions : public gimple_opt_pass | |
1981 | { | |
1982 | public: | |
9af5ce0c | 1983 | pass_feedback_split_functions (gcc::context *ctxt) |
1984 | : gimple_opt_pass (pass_data_feedback_split_functions, ctxt) | |
cbe8bda8 | 1985 | {} |
1986 | ||
1987 | /* opt_pass methods: */ | |
31315c24 | 1988 | virtual bool gate (function *); |
65b0537f | 1989 | virtual unsigned int execute (function *) |
1990 | { | |
1991 | return execute_feedback_split_functions (); | |
1992 | } | |
cbe8bda8 | 1993 | |
1994 | }; // class pass_feedback_split_functions | |
1995 | ||
31315c24 | 1996 | bool |
1997 | pass_feedback_split_functions::gate (function *) | |
1998 | { | |
1999 | /* We don't need to split when profiling at all, we are producing | |
2000 | lousy code anyway. */ | |
2001 | return (flag_partial_inlining | |
2002 | && flag_branch_probabilities); | |
2003 | } | |
2004 | ||
cbe8bda8 | 2005 | } // anon namespace |
2006 | ||
2007 | gimple_opt_pass * | |
2008 | make_pass_feedback_split_functions (gcc::context *ctxt) | |
2009 | { | |
2010 | return new pass_feedback_split_functions (ctxt); | |
2011 | } |