]>
Commit | Line | Data |
---|---|---|
2862cf88 | 1 | /* Function splitting pass |
d353bf18 | 2 | Copyright (C) 2010-2015 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" | |
b20a8bb4 | 80 | #include "alias.h" |
81 | #include "symtab.h" | |
82 | #include "options.h" | |
b20a8bb4 | 83 | #include "tree.h" |
84 | #include "fold-const.h" | |
85 | #include "predict.h" | |
94ea8568 | 86 | #include "tm.h" |
87 | #include "hard-reg-set.h" | |
94ea8568 | 88 | #include "function.h" |
89 | #include "dominance.h" | |
90 | #include "cfg.h" | |
91 | #include "cfganal.h" | |
bc61cadb | 92 | #include "basic-block.h" |
93 | #include "tree-ssa-alias.h" | |
94 | #include "internal-fn.h" | |
95 | #include "gimple-expr.h" | |
e795d6e1 | 96 | #include "gimple.h" |
9ed99284 | 97 | #include "stringpool.h" |
d53441c8 | 98 | #include "rtl.h" |
99 | #include "flags.h" | |
d53441c8 | 100 | #include "insn-config.h" |
101 | #include "expmed.h" | |
102 | #include "dojump.h" | |
103 | #include "explow.h" | |
9ed99284 | 104 | #include "calls.h" |
d53441c8 | 105 | #include "emit-rtl.h" |
106 | #include "varasm.h" | |
107 | #include "stmt.h" | |
108 | #include "expr.h" | |
a8783bee | 109 | #include "gimplify.h" |
dcf1a1ec | 110 | #include "gimple-iterator.h" |
e795d6e1 | 111 | #include "gimplify-me.h" |
dcf1a1ec | 112 | #include "gimple-walk.h" |
2862cf88 | 113 | #include "target.h" |
1140c305 | 114 | #include "plugin-api.h" |
115 | #include "ipa-ref.h" | |
116 | #include "cgraph.h" | |
117 | #include "alloc-pool.h" | |
2cc80ac3 | 118 | #include "symbol-summary.h" |
2862cf88 | 119 | #include "ipa-prop.h" |
073c1fd5 | 120 | #include "gimple-ssa.h" |
121 | #include "tree-cfg.h" | |
122 | #include "tree-phinodes.h" | |
123 | #include "ssa-iterators.h" | |
124 | #include "tree-ssanames.h" | |
125 | #include "tree-into-ssa.h" | |
126 | #include "tree-dfa.h" | |
2862cf88 | 127 | #include "tree-pass.h" |
2862cf88 | 128 | #include "diagnostic.h" |
129 | #include "tree-dump.h" | |
130 | #include "tree-inline.h" | |
2862cf88 | 131 | #include "params.h" |
132 | #include "gimple-pretty-print.h" | |
cbd7f5a0 | 133 | #include "ipa-inline.h" |
f6568ea4 | 134 | #include "cfgloop.h" |
058a1b7a | 135 | #include "tree-chkp.h" |
2862cf88 | 136 | |
137 | /* Per basic block info. */ | |
138 | ||
139 | typedef struct | |
140 | { | |
141 | unsigned int size; | |
142 | unsigned int time; | |
9908fe4d | 143 | } split_bb_info; |
2862cf88 | 144 | |
9908fe4d | 145 | static vec<split_bb_info> bb_info_vec; |
2862cf88 | 146 | |
147 | /* Description of split point. */ | |
148 | ||
149 | struct split_point | |
150 | { | |
151 | /* Size of the partitions. */ | |
152 | unsigned int header_time, header_size, split_time, split_size; | |
153 | ||
574c4ab5 | 154 | /* SSA names that need to be passed into spit function. */ |
2862cf88 | 155 | bitmap ssa_names_to_pass; |
156 | ||
157 | /* Basic block where we split (that will become entry point of new function. */ | |
158 | basic_block entry_bb; | |
159 | ||
160 | /* Basic blocks we are splitting away. */ | |
161 | bitmap split_bbs; | |
b04bab7c | 162 | |
163 | /* True when return value is computed on split part and thus it needs | |
164 | to be returned. */ | |
165 | bool split_part_set_retval; | |
2862cf88 | 166 | }; |
167 | ||
168 | /* Best split point found. */ | |
169 | ||
170 | struct split_point best_split_point; | |
171 | ||
aa2edcb9 | 172 | /* Set of basic blocks that are not allowed to dominate a split point. */ |
173 | ||
174 | static bitmap forbidden_dominators; | |
175 | ||
b04bab7c | 176 | static tree find_retval (basic_block return_bb); |
058a1b7a | 177 | static tree find_retbnd (basic_block return_bb); |
b04bab7c | 178 | |
2dc0a33e | 179 | /* Callback for walk_stmt_load_store_addr_ops. If T is non-SSA automatic |
2862cf88 | 180 | variable, check it if it is present in bitmap passed via DATA. */ |
181 | ||
182 | static bool | |
5b26a9e3 | 183 | test_nonssa_use (gimple, tree t, tree, void *data) |
2862cf88 | 184 | { |
185 | t = get_base_address (t); | |
186 | ||
2dc0a33e | 187 | if (!t || is_gimple_reg (t)) |
188 | return false; | |
189 | ||
190 | if (TREE_CODE (t) == PARM_DECL | |
191 | || (TREE_CODE (t) == VAR_DECL | |
2862cf88 | 192 | && auto_var_in_fn_p (t, current_function_decl)) |
2dc0a33e | 193 | || TREE_CODE (t) == RESULT_DECL |
3c659c08 | 194 | /* Normal labels are part of CFG and will be handled gratefuly. |
195 | Forced labels however can be used directly by statements and | |
196 | need to stay in one partition along with their uses. */ | |
197 | || (TREE_CODE (t) == LABEL_DECL | |
198 | && FORCED_LABEL (t))) | |
2862cf88 | 199 | return bitmap_bit_p ((bitmap)data, DECL_UID (t)); |
b04bab7c | 200 | |
2dc0a33e | 201 | /* For DECL_BY_REFERENCE, the return value is actually a pointer. We want |
202 | to pretend that the value pointed to is actual result decl. */ | |
203 | if ((TREE_CODE (t) == MEM_REF || INDIRECT_REF_P (t)) | |
b04bab7c | 204 | && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME |
ec11736b | 205 | && SSA_NAME_VAR (TREE_OPERAND (t, 0)) |
b04bab7c | 206 | && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND (t, 0))) == RESULT_DECL |
207 | && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) | |
2dc0a33e | 208 | return |
209 | bitmap_bit_p ((bitmap)data, | |
210 | DECL_UID (DECL_RESULT (current_function_decl))); | |
211 | ||
2862cf88 | 212 | return false; |
213 | } | |
214 | ||
215 | /* Dump split point CURRENT. */ | |
216 | ||
217 | static void | |
218 | dump_split_point (FILE * file, struct split_point *current) | |
219 | { | |
220 | fprintf (file, | |
1d0b727d | 221 | "Split point at BB %i\n" |
222 | " header time: %i header size: %i\n" | |
223 | " split time: %i split size: %i\n bbs: ", | |
2862cf88 | 224 | current->entry_bb->index, current->header_time, |
225 | current->header_size, current->split_time, current->split_size); | |
226 | dump_bitmap (file, current->split_bbs); | |
227 | fprintf (file, " SSA names to pass: "); | |
228 | dump_bitmap (file, current->ssa_names_to_pass); | |
229 | } | |
230 | ||
2dc0a33e | 231 | /* Look for all BBs in header that might lead to the split part and verify |
232 | that they are not defining any non-SSA var used by the split part. | |
4493dab3 | 233 | Parameters are the same as for consider_split. */ |
234 | ||
235 | static bool | |
236 | verify_non_ssa_vars (struct split_point *current, bitmap non_ssa_vars, | |
237 | basic_block return_bb) | |
238 | { | |
239 | bitmap seen = BITMAP_ALLOC (NULL); | |
1e094109 | 240 | vec<basic_block> worklist = vNULL; |
4493dab3 | 241 | edge e; |
242 | edge_iterator ei; | |
243 | bool ok = true; | |
3c659c08 | 244 | basic_block bb; |
2dc0a33e | 245 | |
4493dab3 | 246 | FOR_EACH_EDGE (e, ei, current->entry_bb->preds) |
34154e27 | 247 | if (e->src != ENTRY_BLOCK_PTR_FOR_FN (cfun) |
4493dab3 | 248 | && !bitmap_bit_p (current->split_bbs, e->src->index)) |
249 | { | |
f1f41a6c | 250 | worklist.safe_push (e->src); |
4493dab3 | 251 | bitmap_set_bit (seen, e->src->index); |
252 | } | |
2dc0a33e | 253 | |
f1f41a6c | 254 | while (!worklist.is_empty ()) |
4493dab3 | 255 | { |
3c659c08 | 256 | bb = worklist.pop (); |
4493dab3 | 257 | FOR_EACH_EDGE (e, ei, bb->preds) |
34154e27 | 258 | if (e->src != ENTRY_BLOCK_PTR_FOR_FN (cfun) |
6ef9bbe0 | 259 | && bitmap_set_bit (seen, e->src->index)) |
4493dab3 | 260 | { |
261 | gcc_checking_assert (!bitmap_bit_p (current->split_bbs, | |
262 | e->src->index)); | |
f1f41a6c | 263 | worklist.safe_push (e->src); |
4493dab3 | 264 | } |
1a91d914 | 265 | for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi); |
266 | gsi_next (&bsi)) | |
4493dab3 | 267 | { |
2dc0a33e | 268 | gimple stmt = gsi_stmt (bsi); |
269 | if (is_gimple_debug (stmt)) | |
4493dab3 | 270 | continue; |
271 | if (walk_stmt_load_store_addr_ops | |
2dc0a33e | 272 | (stmt, non_ssa_vars, test_nonssa_use, test_nonssa_use, |
273 | test_nonssa_use)) | |
4493dab3 | 274 | { |
275 | ok = false; | |
276 | goto done; | |
277 | } | |
1a91d914 | 278 | if (glabel *label_stmt = dyn_cast <glabel *> (stmt)) |
279 | if (test_nonssa_use (stmt, gimple_label_label (label_stmt), | |
280 | NULL_TREE, non_ssa_vars)) | |
281 | { | |
282 | ok = false; | |
283 | goto done; | |
284 | } | |
4493dab3 | 285 | } |
1a91d914 | 286 | for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi); |
287 | gsi_next (&bsi)) | |
4493dab3 | 288 | { |
289 | if (walk_stmt_load_store_addr_ops | |
2dc0a33e | 290 | (gsi_stmt (bsi), non_ssa_vars, test_nonssa_use, test_nonssa_use, |
291 | test_nonssa_use)) | |
4493dab3 | 292 | { |
293 | ok = false; | |
294 | goto done; | |
295 | } | |
296 | } | |
297 | FOR_EACH_EDGE (e, ei, bb->succs) | |
298 | { | |
299 | if (e->dest != return_bb) | |
300 | continue; | |
1a91d914 | 301 | for (gphi_iterator bsi = gsi_start_phis (return_bb); |
302 | !gsi_end_p (bsi); | |
4493dab3 | 303 | gsi_next (&bsi)) |
304 | { | |
1a91d914 | 305 | gphi *stmt = bsi.phi (); |
4493dab3 | 306 | tree op = gimple_phi_arg_def (stmt, e->dest_idx); |
307 | ||
7c782c9b | 308 | if (virtual_operand_p (gimple_phi_result (stmt))) |
4493dab3 | 309 | continue; |
310 | if (TREE_CODE (op) != SSA_NAME | |
5b26a9e3 | 311 | && test_nonssa_use (stmt, op, op, non_ssa_vars)) |
4493dab3 | 312 | { |
313 | ok = false; | |
314 | goto done; | |
315 | } | |
316 | } | |
317 | } | |
318 | } | |
3c659c08 | 319 | |
320 | /* Verify that the rest of function does not define any label | |
321 | used by the split part. */ | |
322 | FOR_EACH_BB_FN (bb, cfun) | |
323 | if (!bitmap_bit_p (current->split_bbs, bb->index) | |
324 | && !bitmap_bit_p (seen, bb->index)) | |
325 | { | |
326 | gimple_stmt_iterator bsi; | |
327 | for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
1a91d914 | 328 | if (glabel *label_stmt = dyn_cast <glabel *> (gsi_stmt (bsi))) |
3c659c08 | 329 | { |
1a91d914 | 330 | if (test_nonssa_use (label_stmt, |
331 | gimple_label_label (label_stmt), | |
332 | NULL_TREE, non_ssa_vars)) | |
333 | { | |
334 | ok = false; | |
335 | goto done; | |
336 | } | |
3c659c08 | 337 | } |
1a91d914 | 338 | else |
3c659c08 | 339 | break; |
340 | } | |
341 | ||
4493dab3 | 342 | done: |
343 | BITMAP_FREE (seen); | |
f1f41a6c | 344 | worklist.release (); |
4493dab3 | 345 | return ok; |
346 | } | |
347 | ||
aa2edcb9 | 348 | /* If STMT is a call, check the callee against a list of forbidden |
349 | predicate functions. If a match is found, look for uses of the | |
350 | call result in condition statements that compare against zero. | |
351 | For each such use, find the block targeted by the condition | |
352 | statement for the nonzero result, and set the bit for this block | |
353 | in the forbidden dominators bitmap. The purpose of this is to avoid | |
354 | selecting a split point where we are likely to lose the chance | |
355 | to optimize away an unused function call. */ | |
356 | ||
357 | static void | |
358 | check_forbidden_calls (gimple stmt) | |
359 | { | |
360 | imm_use_iterator use_iter; | |
361 | use_operand_p use_p; | |
362 | tree lhs; | |
363 | ||
364 | /* At the moment, __builtin_constant_p is the only forbidden | |
365 | predicate function call (see PR49642). */ | |
366 | if (!gimple_call_builtin_p (stmt, BUILT_IN_CONSTANT_P)) | |
367 | return; | |
368 | ||
369 | lhs = gimple_call_lhs (stmt); | |
370 | ||
371 | if (!lhs || TREE_CODE (lhs) != SSA_NAME) | |
372 | return; | |
373 | ||
374 | FOR_EACH_IMM_USE_FAST (use_p, use_iter, lhs) | |
375 | { | |
376 | tree op1; | |
377 | basic_block use_bb, forbidden_bb; | |
378 | enum tree_code code; | |
379 | edge true_edge, false_edge; | |
1a91d914 | 380 | gcond *use_stmt; |
aa2edcb9 | 381 | |
1a91d914 | 382 | use_stmt = dyn_cast <gcond *> (USE_STMT (use_p)); |
383 | if (!use_stmt) | |
aa2edcb9 | 384 | continue; |
385 | ||
386 | /* Assuming canonical form for GIMPLE_COND here, with constant | |
387 | in second position. */ | |
388 | op1 = gimple_cond_rhs (use_stmt); | |
389 | code = gimple_cond_code (use_stmt); | |
390 | use_bb = gimple_bb (use_stmt); | |
391 | ||
392 | extract_true_false_edges_from_block (use_bb, &true_edge, &false_edge); | |
393 | ||
394 | /* We're only interested in comparisons that distinguish | |
395 | unambiguously from zero. */ | |
396 | if (!integer_zerop (op1) || code == LE_EXPR || code == GE_EXPR) | |
397 | continue; | |
398 | ||
399 | if (code == EQ_EXPR) | |
400 | forbidden_bb = false_edge->dest; | |
401 | else | |
402 | forbidden_bb = true_edge->dest; | |
403 | ||
404 | bitmap_set_bit (forbidden_dominators, forbidden_bb->index); | |
405 | } | |
406 | } | |
407 | ||
408 | /* If BB is dominated by any block in the forbidden dominators set, | |
409 | return TRUE; else FALSE. */ | |
410 | ||
411 | static bool | |
412 | dominated_by_forbidden (basic_block bb) | |
413 | { | |
414 | unsigned dom_bb; | |
415 | bitmap_iterator bi; | |
416 | ||
417 | EXECUTE_IF_SET_IN_BITMAP (forbidden_dominators, 1, dom_bb, bi) | |
418 | { | |
f5a6b05f | 419 | if (dominated_by_p (CDI_DOMINATORS, bb, |
420 | BASIC_BLOCK_FOR_FN (cfun, dom_bb))) | |
aa2edcb9 | 421 | return true; |
422 | } | |
423 | ||
424 | return false; | |
425 | } | |
426 | ||
058a1b7a | 427 | /* For give split point CURRENT and return block RETURN_BB return 1 |
428 | if ssa name VAL is set by split part and 0 otherwise. */ | |
429 | static bool | |
430 | split_part_set_ssa_name_p (tree val, struct split_point *current, | |
431 | basic_block return_bb) | |
432 | { | |
433 | if (TREE_CODE (val) != SSA_NAME) | |
434 | return false; | |
435 | ||
436 | return (!SSA_NAME_IS_DEFAULT_DEF (val) | |
437 | && (bitmap_bit_p (current->split_bbs, | |
438 | gimple_bb (SSA_NAME_DEF_STMT (val))->index) | |
439 | || gimple_bb (SSA_NAME_DEF_STMT (val)) == return_bb)); | |
440 | } | |
441 | ||
2862cf88 | 442 | /* We found an split_point CURRENT. NON_SSA_VARS is bitmap of all non ssa |
443 | variables used and RETURN_BB is return basic block. | |
444 | See if we can split function here. */ | |
445 | ||
446 | static void | |
447 | consider_split (struct split_point *current, bitmap non_ssa_vars, | |
448 | basic_block return_bb) | |
449 | { | |
450 | tree parm; | |
451 | unsigned int num_args = 0; | |
452 | unsigned int call_overhead; | |
453 | edge e; | |
454 | edge_iterator ei; | |
1a91d914 | 455 | gphi_iterator bsi; |
6a69e813 | 456 | unsigned int i; |
574c4ab5 | 457 | int incoming_freq = 0; |
b04bab7c | 458 | tree retval; |
058a1b7a | 459 | tree retbnd; |
15ca8f90 | 460 | bool back_edge = false; |
6a69e813 | 461 | |
2862cf88 | 462 | if (dump_file && (dump_flags & TDF_DETAILS)) |
463 | dump_split_point (dump_file, current); | |
464 | ||
6a69e813 | 465 | FOR_EACH_EDGE (e, ei, current->entry_bb->preds) |
15ca8f90 | 466 | { |
467 | if (e->flags & EDGE_DFS_BACK) | |
468 | back_edge = true; | |
469 | if (!bitmap_bit_p (current->split_bbs, e->src->index)) | |
470 | incoming_freq += EDGE_FREQUENCY (e); | |
471 | } | |
6a69e813 | 472 | |
2862cf88 | 473 | /* Do not split when we would end up calling function anyway. */ |
574c4ab5 | 474 | if (incoming_freq |
34154e27 | 475 | >= (ENTRY_BLOCK_PTR_FOR_FN (cfun)->frequency |
2862cf88 | 476 | * PARAM_VALUE (PARAM_PARTIAL_INLINING_ENTRY_PROBABILITY) / 100)) |
477 | { | |
15ca8f90 | 478 | /* When profile is guessed, we can not expect it to give us |
479 | realistic estimate on likelyness of function taking the | |
480 | complex path. As a special case, when tail of the function is | |
481 | a loop, enable splitting since inlining code skipping the loop | |
482 | is likely noticeable win. */ | |
483 | if (back_edge | |
f26d8580 | 484 | && profile_status_for_fn (cfun) != PROFILE_READ |
34154e27 | 485 | && incoming_freq < ENTRY_BLOCK_PTR_FOR_FN (cfun)->frequency) |
15ca8f90 | 486 | { |
487 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
488 | fprintf (dump_file, | |
489 | " Split before loop, accepting despite low frequencies %i %i.\n", | |
490 | incoming_freq, | |
34154e27 | 491 | ENTRY_BLOCK_PTR_FOR_FN (cfun)->frequency); |
15ca8f90 | 492 | } |
493 | else | |
494 | { | |
495 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
496 | fprintf (dump_file, | |
497 | " Refused: incoming frequency is too large.\n"); | |
498 | return; | |
499 | } | |
2862cf88 | 500 | } |
501 | ||
502 | if (!current->header_size) | |
503 | { | |
504 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
505 | fprintf (dump_file, " Refused: header empty\n"); | |
2862cf88 | 506 | return; |
507 | } | |
508 | ||
574c4ab5 | 509 | /* Verify that PHI args on entry are either virtual or all their operands |
510 | incoming from header are the same. */ | |
6a69e813 | 511 | for (bsi = gsi_start_phis (current->entry_bb); !gsi_end_p (bsi); gsi_next (&bsi)) |
2862cf88 | 512 | { |
1a91d914 | 513 | gphi *stmt = bsi.phi (); |
6a69e813 | 514 | tree val = NULL; |
515 | ||
7c782c9b | 516 | if (virtual_operand_p (gimple_phi_result (stmt))) |
6a69e813 | 517 | continue; |
518 | for (i = 0; i < gimple_phi_num_args (stmt); i++) | |
519 | { | |
520 | edge e = gimple_phi_arg_edge (stmt, i); | |
521 | if (!bitmap_bit_p (current->split_bbs, e->src->index)) | |
522 | { | |
523 | tree edge_val = gimple_phi_arg_def (stmt, i); | |
524 | if (val && edge_val != val) | |
525 | { | |
526 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
527 | fprintf (dump_file, | |
528 | " Refused: entry BB has PHI with multiple variants\n"); | |
529 | return; | |
530 | } | |
531 | val = edge_val; | |
532 | } | |
533 | } | |
2862cf88 | 534 | } |
535 | ||
536 | ||
537 | /* See what argument we will pass to the split function and compute | |
538 | call overhead. */ | |
539 | call_overhead = eni_size_weights.call_cost; | |
540 | for (parm = DECL_ARGUMENTS (current_function_decl); parm; | |
1767a056 | 541 | parm = DECL_CHAIN (parm)) |
2862cf88 | 542 | { |
543 | if (!is_gimple_reg (parm)) | |
544 | { | |
545 | if (bitmap_bit_p (non_ssa_vars, DECL_UID (parm))) | |
546 | { | |
547 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
548 | fprintf (dump_file, | |
549 | " Refused: need to pass non-ssa param values\n"); | |
550 | return; | |
551 | } | |
552 | } | |
c6dfe037 | 553 | else |
2862cf88 | 554 | { |
c6dfe037 | 555 | tree ddef = ssa_default_def (cfun, parm); |
556 | if (ddef | |
557 | && bitmap_bit_p (current->ssa_names_to_pass, | |
558 | SSA_NAME_VERSION (ddef))) | |
559 | { | |
560 | if (!VOID_TYPE_P (TREE_TYPE (parm))) | |
96c7c99c | 561 | call_overhead += estimate_move_cost (TREE_TYPE (parm), false); |
c6dfe037 | 562 | num_args++; |
563 | } | |
2862cf88 | 564 | } |
565 | } | |
566 | if (!VOID_TYPE_P (TREE_TYPE (current_function_decl))) | |
96c7c99c | 567 | call_overhead += estimate_move_cost (TREE_TYPE (current_function_decl), |
568 | false); | |
2862cf88 | 569 | |
570 | if (current->split_size <= call_overhead) | |
571 | { | |
572 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
573 | fprintf (dump_file, | |
574 | " Refused: split size is smaller than call overhead\n"); | |
575 | return; | |
576 | } | |
577 | if (current->header_size + call_overhead | |
578 | >= (unsigned int)(DECL_DECLARED_INLINE_P (current_function_decl) | |
579 | ? MAX_INLINE_INSNS_SINGLE | |
580 | : MAX_INLINE_INSNS_AUTO)) | |
581 | { | |
582 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
583 | fprintf (dump_file, | |
584 | " Refused: header size is too large for inline candidate\n"); | |
585 | return; | |
586 | } | |
587 | ||
7037b790 | 588 | /* Splitting functions brings the target out of comdat group; this will |
589 | lead to code duplication if the function is reused by other unit. | |
590 | Limit this duplication. This is consistent with limit in tree-sra.c | |
591 | FIXME: with LTO we ought to be able to do better! */ | |
592 | if (DECL_ONE_ONLY (current_function_decl) | |
593 | && current->split_size >= (unsigned int) MAX_INLINE_INSNS_AUTO) | |
594 | { | |
595 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
596 | fprintf (dump_file, | |
597 | " Refused: function is COMDAT and tail is too large\n"); | |
598 | return; | |
599 | } | |
600 | /* For comdat functions also reject very small tails; those will likely get | |
601 | inlined back and we do not want to risk the duplication overhead. | |
602 | FIXME: with LTO we ought to be able to do better! */ | |
603 | if (DECL_ONE_ONLY (current_function_decl) | |
604 | && current->split_size | |
605 | <= (unsigned int) PARAM_VALUE (PARAM_EARLY_INLINING_INSNS) / 2) | |
606 | { | |
607 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
608 | fprintf (dump_file, | |
609 | " Refused: function is COMDAT and tail is too small\n"); | |
610 | return; | |
611 | } | |
612 | ||
2862cf88 | 613 | /* FIXME: we currently can pass only SSA function parameters to the split |
fd8d648f | 614 | arguments. Once parm_adjustment infrastructure is supported by cloning, |
2862cf88 | 615 | we can pass more than that. */ |
616 | if (num_args != bitmap_count_bits (current->ssa_names_to_pass)) | |
617 | { | |
6a69e813 | 618 | |
2862cf88 | 619 | if (dump_file && (dump_flags & TDF_DETAILS)) |
620 | fprintf (dump_file, | |
621 | " Refused: need to pass non-param values\n"); | |
622 | return; | |
623 | } | |
624 | ||
625 | /* When there are non-ssa vars used in the split region, see if they | |
626 | are used in the header region. If so, reject the split. | |
627 | FIXME: we can use nested function support to access both. */ | |
4493dab3 | 628 | if (!bitmap_empty_p (non_ssa_vars) |
629 | && !verify_non_ssa_vars (current, non_ssa_vars, return_bb)) | |
2862cf88 | 630 | { |
4493dab3 | 631 | if (dump_file && (dump_flags & TDF_DETAILS)) |
632 | fprintf (dump_file, | |
633 | " Refused: split part has non-ssa uses\n"); | |
2862cf88 | 634 | return; |
635 | } | |
aa2edcb9 | 636 | |
637 | /* If the split point is dominated by a forbidden block, reject | |
638 | the split. */ | |
639 | if (!bitmap_empty_p (forbidden_dominators) | |
640 | && dominated_by_forbidden (current->entry_bb)) | |
641 | { | |
642 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
643 | fprintf (dump_file, | |
644 | " Refused: split point dominated by forbidden block\n"); | |
645 | return; | |
646 | } | |
647 | ||
b04bab7c | 648 | /* See if retval used by return bb is computed by header or split part. |
649 | When it is computed by split part, we need to produce return statement | |
650 | in the split part and add code to header to pass it around. | |
651 | ||
652 | This is bit tricky to test: | |
653 | 1) When there is no return_bb or no return value, we always pass | |
654 | value around. | |
655 | 2) Invariants are always computed by caller. | |
656 | 3) For SSA we need to look if defining statement is in header or split part | |
657 | 4) For non-SSA we need to look where the var is computed. */ | |
658 | retval = find_retval (return_bb); | |
659 | if (!retval) | |
660 | current->split_part_set_retval = true; | |
661 | else if (is_gimple_min_invariant (retval)) | |
662 | current->split_part_set_retval = false; | |
663 | /* Special case is value returned by reference we record as if it was non-ssa | |
664 | set to result_decl. */ | |
665 | else if (TREE_CODE (retval) == SSA_NAME | |
ec11736b | 666 | && SSA_NAME_VAR (retval) |
b04bab7c | 667 | && TREE_CODE (SSA_NAME_VAR (retval)) == RESULT_DECL |
668 | && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) | |
669 | current->split_part_set_retval | |
670 | = bitmap_bit_p (non_ssa_vars, DECL_UID (SSA_NAME_VAR (retval))); | |
671 | else if (TREE_CODE (retval) == SSA_NAME) | |
672 | current->split_part_set_retval | |
058a1b7a | 673 | = split_part_set_ssa_name_p (retval, current, return_bb); |
b04bab7c | 674 | else if (TREE_CODE (retval) == PARM_DECL) |
675 | current->split_part_set_retval = false; | |
676 | else if (TREE_CODE (retval) == VAR_DECL | |
677 | || TREE_CODE (retval) == RESULT_DECL) | |
678 | current->split_part_set_retval | |
679 | = bitmap_bit_p (non_ssa_vars, DECL_UID (retval)); | |
680 | else | |
681 | current->split_part_set_retval = true; | |
682 | ||
058a1b7a | 683 | /* See if retbnd used by return bb is computed by header or split part. */ |
684 | retbnd = find_retbnd (return_bb); | |
685 | if (retbnd) | |
686 | { | |
687 | bool split_part_set_retbnd | |
688 | = split_part_set_ssa_name_p (retbnd, current, return_bb); | |
689 | ||
690 | /* If we have both return value and bounds then keep their definitions | |
691 | in a single function. We use SSA names to link returned bounds and | |
692 | value and therefore do not handle cases when result is passed by | |
693 | reference (which should not be our case anyway since bounds are | |
694 | returned for pointers only). */ | |
695 | if ((DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)) | |
696 | && current->split_part_set_retval) | |
697 | || split_part_set_retbnd != current->split_part_set_retval) | |
698 | { | |
699 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
700 | fprintf (dump_file, | |
701 | " Refused: split point splits return value and bounds\n"); | |
702 | return; | |
703 | } | |
704 | } | |
705 | ||
9aa9c1e6 | 706 | /* split_function fixes up at most one PHI non-virtual PHI node in return_bb, |
707 | for the return value. If there are other PHIs, give up. */ | |
34154e27 | 708 | if (return_bb != EXIT_BLOCK_PTR_FOR_FN (cfun)) |
9aa9c1e6 | 709 | { |
1a91d914 | 710 | gphi_iterator psi; |
9aa9c1e6 | 711 | |
712 | for (psi = gsi_start_phis (return_bb); !gsi_end_p (psi); gsi_next (&psi)) | |
1a91d914 | 713 | if (!virtual_operand_p (gimple_phi_result (psi.phi ())) |
9aa9c1e6 | 714 | && !(retval |
715 | && current->split_part_set_retval | |
716 | && TREE_CODE (retval) == SSA_NAME | |
717 | && !DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)) | |
1a91d914 | 718 | && SSA_NAME_DEF_STMT (retval) == psi.phi ())) |
9aa9c1e6 | 719 | { |
720 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
721 | fprintf (dump_file, | |
722 | " Refused: return bb has extra PHIs\n"); | |
723 | return; | |
724 | } | |
725 | } | |
726 | ||
727 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
728 | fprintf (dump_file, " Accepted!\n"); | |
729 | ||
2862cf88 | 730 | /* At the moment chose split point with lowest frequency and that leaves |
731 | out smallest size of header. | |
732 | In future we might re-consider this heuristics. */ | |
733 | if (!best_split_point.split_bbs | |
734 | || best_split_point.entry_bb->frequency > current->entry_bb->frequency | |
735 | || (best_split_point.entry_bb->frequency == current->entry_bb->frequency | |
736 | && best_split_point.split_size < current->split_size)) | |
737 | ||
738 | { | |
739 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
740 | fprintf (dump_file, " New best split point!\n"); | |
741 | if (best_split_point.ssa_names_to_pass) | |
742 | { | |
743 | BITMAP_FREE (best_split_point.ssa_names_to_pass); | |
744 | BITMAP_FREE (best_split_point.split_bbs); | |
745 | } | |
746 | best_split_point = *current; | |
747 | best_split_point.ssa_names_to_pass = BITMAP_ALLOC (NULL); | |
748 | bitmap_copy (best_split_point.ssa_names_to_pass, | |
749 | current->ssa_names_to_pass); | |
750 | best_split_point.split_bbs = BITMAP_ALLOC (NULL); | |
751 | bitmap_copy (best_split_point.split_bbs, current->split_bbs); | |
752 | } | |
753 | } | |
754 | ||
4493dab3 | 755 | /* Return basic block containing RETURN statement. We allow basic blocks |
756 | of the form: | |
757 | <retval> = tmp_var; | |
758 | return <retval> | |
a63cfc4d | 759 | but return_bb can not be more complex than this (except for |
760 | -fsanitize=thread we allow TSAN_FUNC_EXIT () internal call in there). | |
efee62d1 | 761 | If nothing is found, return the exit block. |
4493dab3 | 762 | |
2862cf88 | 763 | When there are multiple RETURN statement, chose one with return value, |
764 | since that one is more likely shared by multiple code paths. | |
4493dab3 | 765 | |
766 | Return BB is special, because for function splitting it is the only | |
767 | basic block that is duplicated in between header and split part of the | |
768 | function. | |
769 | ||
2862cf88 | 770 | TODO: We might support multiple return blocks. */ |
771 | ||
772 | static basic_block | |
773 | find_return_bb (void) | |
774 | { | |
775 | edge e; | |
34154e27 | 776 | basic_block return_bb = EXIT_BLOCK_PTR_FOR_FN (cfun); |
b234c94d | 777 | gimple_stmt_iterator bsi; |
778 | bool found_return = false; | |
779 | tree retval = NULL_TREE; | |
2862cf88 | 780 | |
34154e27 | 781 | if (!single_pred_p (EXIT_BLOCK_PTR_FOR_FN (cfun))) |
b234c94d | 782 | return return_bb; |
783 | ||
34154e27 | 784 | e = single_pred_edge (EXIT_BLOCK_PTR_FOR_FN (cfun)); |
b234c94d | 785 | for (bsi = gsi_last_bb (e->src); !gsi_end_p (bsi); gsi_prev (&bsi)) |
786 | { | |
787 | gimple stmt = gsi_stmt (bsi); | |
1a3c6423 | 788 | if (gimple_code (stmt) == GIMPLE_LABEL |
789 | || is_gimple_debug (stmt) | |
790 | || gimple_clobber_p (stmt)) | |
b234c94d | 791 | ; |
792 | else if (gimple_code (stmt) == GIMPLE_ASSIGN | |
793 | && found_return | |
794 | && gimple_assign_single_p (stmt) | |
795 | && (auto_var_in_fn_p (gimple_assign_rhs1 (stmt), | |
796 | current_function_decl) | |
797 | || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))) | |
798 | && retval == gimple_assign_lhs (stmt)) | |
799 | ; | |
1a91d914 | 800 | else if (greturn *return_stmt = dyn_cast <greturn *> (stmt)) |
b234c94d | 801 | { |
802 | found_return = true; | |
1a91d914 | 803 | retval = gimple_return_retval (return_stmt); |
b234c94d | 804 | } |
a63cfc4d | 805 | /* For -fsanitize=thread, allow also TSAN_FUNC_EXIT () in the return |
806 | bb. */ | |
807 | else if ((flag_sanitize & SANITIZE_THREAD) | |
808 | && is_gimple_call (stmt) | |
809 | && gimple_call_internal_p (stmt) | |
810 | && gimple_call_internal_fn (stmt) == IFN_TSAN_FUNC_EXIT) | |
811 | ; | |
b234c94d | 812 | else |
813 | break; | |
814 | } | |
815 | if (gsi_end_p (bsi) && found_return) | |
816 | return_bb = e->src; | |
2862cf88 | 817 | |
2862cf88 | 818 | return return_bb; |
819 | } | |
820 | ||
574c4ab5 | 821 | /* Given return basic block RETURN_BB, see where return value is really |
4493dab3 | 822 | stored. */ |
823 | static tree | |
824 | find_retval (basic_block return_bb) | |
825 | { | |
826 | gimple_stmt_iterator bsi; | |
827 | for (bsi = gsi_start_bb (return_bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
1a91d914 | 828 | if (greturn *return_stmt = dyn_cast <greturn *> (gsi_stmt (bsi))) |
829 | return gimple_return_retval (return_stmt); | |
1a3c6423 | 830 | else if (gimple_code (gsi_stmt (bsi)) == GIMPLE_ASSIGN |
831 | && !gimple_clobber_p (gsi_stmt (bsi))) | |
4493dab3 | 832 | return gimple_assign_rhs1 (gsi_stmt (bsi)); |
833 | return NULL; | |
834 | } | |
835 | ||
058a1b7a | 836 | /* Given return basic block RETURN_BB, see where return bounds are really |
837 | stored. */ | |
838 | static tree | |
839 | find_retbnd (basic_block return_bb) | |
840 | { | |
841 | gimple_stmt_iterator bsi; | |
842 | for (bsi = gsi_last_bb (return_bb); !gsi_end_p (bsi); gsi_prev (&bsi)) | |
843 | if (gimple_code (gsi_stmt (bsi)) == GIMPLE_RETURN) | |
844 | return gimple_return_retbnd (gsi_stmt (bsi)); | |
845 | return NULL; | |
846 | } | |
847 | ||
2dc0a33e | 848 | /* Callback for walk_stmt_load_store_addr_ops. If T is non-SSA automatic |
849 | variable, mark it as used in bitmap passed via DATA. | |
2862cf88 | 850 | Return true when access to T prevents splitting the function. */ |
851 | ||
852 | static bool | |
5b26a9e3 | 853 | mark_nonssa_use (gimple, tree t, tree, void *data) |
2862cf88 | 854 | { |
855 | t = get_base_address (t); | |
856 | ||
857 | if (!t || is_gimple_reg (t)) | |
858 | return false; | |
859 | ||
860 | /* At present we can't pass non-SSA arguments to split function. | |
861 | FIXME: this can be relaxed by passing references to arguments. */ | |
862 | if (TREE_CODE (t) == PARM_DECL) | |
863 | { | |
864 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2dc0a33e | 865 | fprintf (dump_file, |
866 | "Cannot split: use of non-ssa function parameter.\n"); | |
2862cf88 | 867 | return true; |
868 | } | |
869 | ||
2dc0a33e | 870 | if ((TREE_CODE (t) == VAR_DECL |
871 | && auto_var_in_fn_p (t, current_function_decl)) | |
872 | || TREE_CODE (t) == RESULT_DECL | |
3c659c08 | 873 | || (TREE_CODE (t) == LABEL_DECL |
874 | && FORCED_LABEL (t))) | |
2862cf88 | 875 | bitmap_set_bit ((bitmap)data, DECL_UID (t)); |
b04bab7c | 876 | |
2dc0a33e | 877 | /* For DECL_BY_REFERENCE, the return value is actually a pointer. We want |
878 | to pretend that the value pointed to is actual result decl. */ | |
879 | if ((TREE_CODE (t) == MEM_REF || INDIRECT_REF_P (t)) | |
b04bab7c | 880 | && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME |
ec11736b | 881 | && SSA_NAME_VAR (TREE_OPERAND (t, 0)) |
b04bab7c | 882 | && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND (t, 0))) == RESULT_DECL |
883 | && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) | |
2dc0a33e | 884 | return |
885 | bitmap_bit_p ((bitmap)data, | |
886 | DECL_UID (DECL_RESULT (current_function_decl))); | |
887 | ||
2862cf88 | 888 | return false; |
889 | } | |
890 | ||
891 | /* Compute local properties of basic block BB we collect when looking for | |
892 | split points. We look for ssa defs and store them in SET_SSA_NAMES, | |
893 | for ssa uses and store them in USED_SSA_NAMES and for any non-SSA automatic | |
894 | vars stored in NON_SSA_VARS. | |
895 | ||
896 | When BB has edge to RETURN_BB, collect uses in RETURN_BB too. | |
897 | ||
898 | Return false when BB contains something that prevents it from being put into | |
899 | split function. */ | |
900 | ||
901 | static bool | |
902 | visit_bb (basic_block bb, basic_block return_bb, | |
903 | bitmap set_ssa_names, bitmap used_ssa_names, | |
904 | bitmap non_ssa_vars) | |
905 | { | |
2862cf88 | 906 | edge e; |
907 | edge_iterator ei; | |
908 | bool can_split = true; | |
909 | ||
1a91d914 | 910 | for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi); |
911 | gsi_next (&bsi)) | |
2862cf88 | 912 | { |
913 | gimple stmt = gsi_stmt (bsi); | |
914 | tree op; | |
915 | ssa_op_iter iter; | |
916 | tree decl; | |
917 | ||
918 | if (is_gimple_debug (stmt)) | |
919 | continue; | |
920 | ||
1a3c6423 | 921 | if (gimple_clobber_p (stmt)) |
922 | continue; | |
923 | ||
2862cf88 | 924 | /* FIXME: We can split regions containing EH. We can not however |
925 | split RESX, EH_DISPATCH and EH_POINTER referring to same region | |
926 | into different partitions. This would require tracking of | |
927 | EH regions and checking in consider_split_point if they | |
928 | are not used elsewhere. */ | |
892c141e | 929 | if (gimple_code (stmt) == GIMPLE_RESX) |
2862cf88 | 930 | { |
931 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
892c141e | 932 | fprintf (dump_file, "Cannot split: resx.\n"); |
2862cf88 | 933 | can_split = false; |
934 | } | |
935 | if (gimple_code (stmt) == GIMPLE_EH_DISPATCH) | |
936 | { | |
937 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2dc0a33e | 938 | fprintf (dump_file, "Cannot split: eh dispatch.\n"); |
2862cf88 | 939 | can_split = false; |
940 | } | |
941 | ||
942 | /* Check builtins that prevent splitting. */ | |
943 | if (gimple_code (stmt) == GIMPLE_CALL | |
944 | && (decl = gimple_call_fndecl (stmt)) != NULL_TREE | |
945 | && DECL_BUILT_IN (decl) | |
946 | && DECL_BUILT_IN_CLASS (decl) == BUILT_IN_NORMAL) | |
947 | switch (DECL_FUNCTION_CODE (decl)) | |
948 | { | |
949 | /* FIXME: once we will allow passing non-parm values to split part, | |
950 | we need to be sure to handle correct builtin_stack_save and | |
951 | builtin_stack_restore. At the moment we are safe; there is no | |
952 | way to store builtin_stack_save result in non-SSA variable | |
953 | since all calls to those are compiler generated. */ | |
954 | case BUILT_IN_APPLY: | |
3c97c75d | 955 | case BUILT_IN_APPLY_ARGS: |
2862cf88 | 956 | case BUILT_IN_VA_START: |
957 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2dc0a33e | 958 | fprintf (dump_file, |
959 | "Cannot split: builtin_apply and va_start.\n"); | |
2862cf88 | 960 | can_split = false; |
961 | break; | |
962 | case BUILT_IN_EH_POINTER: | |
963 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2dc0a33e | 964 | fprintf (dump_file, "Cannot split: builtin_eh_pointer.\n"); |
2862cf88 | 965 | can_split = false; |
966 | break; | |
967 | default: | |
968 | break; | |
969 | } | |
970 | ||
971 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF) | |
972 | bitmap_set_bit (set_ssa_names, SSA_NAME_VERSION (op)); | |
973 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE) | |
974 | bitmap_set_bit (used_ssa_names, SSA_NAME_VERSION (op)); | |
975 | can_split &= !walk_stmt_load_store_addr_ops (stmt, non_ssa_vars, | |
976 | mark_nonssa_use, | |
977 | mark_nonssa_use, | |
978 | mark_nonssa_use); | |
979 | } | |
1a91d914 | 980 | for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi); |
981 | gsi_next (&bsi)) | |
2862cf88 | 982 | { |
1a91d914 | 983 | gphi *stmt = bsi.phi (); |
6a69e813 | 984 | unsigned int i; |
2862cf88 | 985 | |
7c782c9b | 986 | if (virtual_operand_p (gimple_phi_result (stmt))) |
2862cf88 | 987 | continue; |
6a69e813 | 988 | bitmap_set_bit (set_ssa_names, |
989 | SSA_NAME_VERSION (gimple_phi_result (stmt))); | |
990 | for (i = 0; i < gimple_phi_num_args (stmt); i++) | |
991 | { | |
992 | tree op = gimple_phi_arg_def (stmt, i); | |
993 | if (TREE_CODE (op) == SSA_NAME) | |
994 | bitmap_set_bit (used_ssa_names, SSA_NAME_VERSION (op)); | |
995 | } | |
2862cf88 | 996 | can_split &= !walk_stmt_load_store_addr_ops (stmt, non_ssa_vars, |
997 | mark_nonssa_use, | |
998 | mark_nonssa_use, | |
999 | mark_nonssa_use); | |
1000 | } | |
574c4ab5 | 1001 | /* Record also uses coming from PHI operand in return BB. */ |
2862cf88 | 1002 | FOR_EACH_EDGE (e, ei, bb->succs) |
1003 | if (e->dest == return_bb) | |
1004 | { | |
1a91d914 | 1005 | for (gphi_iterator bsi = gsi_start_phis (return_bb); |
1006 | !gsi_end_p (bsi); | |
1007 | gsi_next (&bsi)) | |
2862cf88 | 1008 | { |
1a91d914 | 1009 | gphi *stmt = bsi.phi (); |
2862cf88 | 1010 | tree op = gimple_phi_arg_def (stmt, e->dest_idx); |
1011 | ||
7c782c9b | 1012 | if (virtual_operand_p (gimple_phi_result (stmt))) |
2862cf88 | 1013 | continue; |
2862cf88 | 1014 | if (TREE_CODE (op) == SSA_NAME) |
1015 | bitmap_set_bit (used_ssa_names, SSA_NAME_VERSION (op)); | |
1016 | else | |
5b26a9e3 | 1017 | can_split &= !mark_nonssa_use (stmt, op, op, non_ssa_vars); |
2862cf88 | 1018 | } |
2862cf88 | 1019 | } |
1020 | return can_split; | |
1021 | } | |
1022 | ||
1023 | /* Stack entry for recursive DFS walk in find_split_point. */ | |
1024 | ||
1025 | typedef struct | |
1026 | { | |
1027 | /* Basic block we are examining. */ | |
1028 | basic_block bb; | |
1029 | ||
1030 | /* SSA names set and used by the BB and all BBs reachable | |
1031 | from it via DFS walk. */ | |
1032 | bitmap set_ssa_names, used_ssa_names; | |
1033 | bitmap non_ssa_vars; | |
1034 | ||
1035 | /* All BBS visited from this BB via DFS walk. */ | |
1036 | bitmap bbs_visited; | |
1037 | ||
1038 | /* Last examined edge in DFS walk. Since we walk unoriented graph, | |
574c4ab5 | 1039 | the value is up to sum of incoming and outgoing edges of BB. */ |
2862cf88 | 1040 | unsigned int edge_num; |
1041 | ||
1042 | /* Stack entry index of earliest BB reachable from current BB | |
574c4ab5 | 1043 | or any BB visited later in DFS walk. */ |
2862cf88 | 1044 | int earliest; |
1045 | ||
1046 | /* Overall time and size of all BBs reached from this BB in DFS walk. */ | |
1047 | int overall_time, overall_size; | |
1048 | ||
1049 | /* When false we can not split on this BB. */ | |
1050 | bool can_split; | |
1051 | } stack_entry; | |
2862cf88 | 1052 | |
1053 | ||
1054 | /* Find all articulations and call consider_split on them. | |
1055 | OVERALL_TIME and OVERALL_SIZE is time and size of the function. | |
1056 | ||
1057 | We perform basic algorithm for finding an articulation in a graph | |
1058 | created from CFG by considering it to be an unoriented graph. | |
1059 | ||
1060 | The articulation is discovered via DFS walk. We collect earliest | |
1061 | basic block on stack that is reachable via backward edge. Articulation | |
1062 | is any basic block such that there is no backward edge bypassing it. | |
1063 | To reduce stack usage we maintain heap allocated stack in STACK vector. | |
1064 | AUX pointer of BB is set to index it appears in the stack or -1 once | |
1065 | it is visited and popped off the stack. | |
1066 | ||
1067 | The algorithm finds articulation after visiting the whole component | |
1068 | reachable by it. This makes it convenient to collect information about | |
1069 | the component used by consider_split. */ | |
1070 | ||
1071 | static void | |
a63cfc4d | 1072 | find_split_points (basic_block return_bb, int overall_time, int overall_size) |
2862cf88 | 1073 | { |
1074 | stack_entry first; | |
1e094109 | 1075 | vec<stack_entry> stack = vNULL; |
2862cf88 | 1076 | basic_block bb; |
2862cf88 | 1077 | struct split_point current; |
1078 | ||
1079 | current.header_time = overall_time; | |
1080 | current.header_size = overall_size; | |
1081 | current.split_time = 0; | |
1082 | current.split_size = 0; | |
1083 | current.ssa_names_to_pass = BITMAP_ALLOC (NULL); | |
1084 | ||
34154e27 | 1085 | first.bb = ENTRY_BLOCK_PTR_FOR_FN (cfun); |
2862cf88 | 1086 | first.edge_num = 0; |
1087 | first.overall_time = 0; | |
1088 | first.overall_size = 0; | |
1089 | first.earliest = INT_MAX; | |
1090 | first.set_ssa_names = 0; | |
1091 | first.used_ssa_names = 0; | |
60141df0 | 1092 | first.non_ssa_vars = 0; |
2862cf88 | 1093 | first.bbs_visited = 0; |
60141df0 | 1094 | first.can_split = false; |
f1f41a6c | 1095 | stack.safe_push (first); |
34154e27 | 1096 | ENTRY_BLOCK_PTR_FOR_FN (cfun)->aux = (void *)(intptr_t)-1; |
2862cf88 | 1097 | |
f1f41a6c | 1098 | while (!stack.is_empty ()) |
2862cf88 | 1099 | { |
f1f41a6c | 1100 | stack_entry *entry = &stack.last (); |
2862cf88 | 1101 | |
1102 | /* We are walking an acyclic graph, so edge_num counts | |
1103 | succ and pred edges together. However when considering | |
1104 | articulation, we want to have processed everything reachable | |
1105 | from articulation but nothing that reaches into it. */ | |
1106 | if (entry->edge_num == EDGE_COUNT (entry->bb->succs) | |
34154e27 | 1107 | && entry->bb != ENTRY_BLOCK_PTR_FOR_FN (cfun)) |
2862cf88 | 1108 | { |
f1f41a6c | 1109 | int pos = stack.length (); |
2862cf88 | 1110 | entry->can_split &= visit_bb (entry->bb, return_bb, |
1111 | entry->set_ssa_names, | |
1112 | entry->used_ssa_names, | |
1113 | entry->non_ssa_vars); | |
1114 | if (pos <= entry->earliest && !entry->can_split | |
1115 | && dump_file && (dump_flags & TDF_DETAILS)) | |
1116 | fprintf (dump_file, | |
1117 | "found articulation at bb %i but can not split\n", | |
1118 | entry->bb->index); | |
1119 | if (pos <= entry->earliest && entry->can_split) | |
1120 | { | |
1121 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1122 | fprintf (dump_file, "found articulation at bb %i\n", | |
1123 | entry->bb->index); | |
1124 | current.entry_bb = entry->bb; | |
1125 | current.ssa_names_to_pass = BITMAP_ALLOC (NULL); | |
1126 | bitmap_and_compl (current.ssa_names_to_pass, | |
1127 | entry->used_ssa_names, entry->set_ssa_names); | |
1128 | current.header_time = overall_time - entry->overall_time; | |
1129 | current.header_size = overall_size - entry->overall_size; | |
1130 | current.split_time = entry->overall_time; | |
1131 | current.split_size = entry->overall_size; | |
1132 | current.split_bbs = entry->bbs_visited; | |
1133 | consider_split (¤t, entry->non_ssa_vars, return_bb); | |
1134 | BITMAP_FREE (current.ssa_names_to_pass); | |
1135 | } | |
1136 | } | |
1137 | /* Do actual DFS walk. */ | |
1138 | if (entry->edge_num | |
1139 | < (EDGE_COUNT (entry->bb->succs) | |
1140 | + EDGE_COUNT (entry->bb->preds))) | |
1141 | { | |
1142 | edge e; | |
1143 | basic_block dest; | |
1144 | if (entry->edge_num < EDGE_COUNT (entry->bb->succs)) | |
1145 | { | |
1146 | e = EDGE_SUCC (entry->bb, entry->edge_num); | |
1147 | dest = e->dest; | |
1148 | } | |
1149 | else | |
1150 | { | |
1151 | e = EDGE_PRED (entry->bb, entry->edge_num | |
1152 | - EDGE_COUNT (entry->bb->succs)); | |
1153 | dest = e->src; | |
1154 | } | |
1155 | ||
1156 | entry->edge_num++; | |
1157 | ||
1158 | /* New BB to visit, push it to the stack. */ | |
34154e27 | 1159 | if (dest != return_bb && dest != EXIT_BLOCK_PTR_FOR_FN (cfun) |
2862cf88 | 1160 | && !dest->aux) |
1161 | { | |
1162 | stack_entry new_entry; | |
1163 | ||
1164 | new_entry.bb = dest; | |
1165 | new_entry.edge_num = 0; | |
1166 | new_entry.overall_time | |
f1f41a6c | 1167 | = bb_info_vec[dest->index].time; |
2862cf88 | 1168 | new_entry.overall_size |
f1f41a6c | 1169 | = bb_info_vec[dest->index].size; |
2862cf88 | 1170 | new_entry.earliest = INT_MAX; |
1171 | new_entry.set_ssa_names = BITMAP_ALLOC (NULL); | |
1172 | new_entry.used_ssa_names = BITMAP_ALLOC (NULL); | |
1173 | new_entry.bbs_visited = BITMAP_ALLOC (NULL); | |
1174 | new_entry.non_ssa_vars = BITMAP_ALLOC (NULL); | |
1175 | new_entry.can_split = true; | |
1176 | bitmap_set_bit (new_entry.bbs_visited, dest->index); | |
f1f41a6c | 1177 | stack.safe_push (new_entry); |
1178 | dest->aux = (void *)(intptr_t)stack.length (); | |
2862cf88 | 1179 | } |
1180 | /* Back edge found, record the earliest point. */ | |
1181 | else if ((intptr_t)dest->aux > 0 | |
1182 | && (intptr_t)dest->aux < entry->earliest) | |
1183 | entry->earliest = (intptr_t)dest->aux; | |
1184 | } | |
574c4ab5 | 1185 | /* We are done with examining the edges. Pop off the value from stack |
1186 | and merge stuff we accumulate during the walk. */ | |
34154e27 | 1187 | else if (entry->bb != ENTRY_BLOCK_PTR_FOR_FN (cfun)) |
2862cf88 | 1188 | { |
f1f41a6c | 1189 | stack_entry *prev = &stack[stack.length () - 2]; |
2862cf88 | 1190 | |
1191 | entry->bb->aux = (void *)(intptr_t)-1; | |
1192 | prev->can_split &= entry->can_split; | |
1193 | if (prev->set_ssa_names) | |
1194 | { | |
1195 | bitmap_ior_into (prev->set_ssa_names, entry->set_ssa_names); | |
1196 | bitmap_ior_into (prev->used_ssa_names, entry->used_ssa_names); | |
1197 | bitmap_ior_into (prev->bbs_visited, entry->bbs_visited); | |
1198 | bitmap_ior_into (prev->non_ssa_vars, entry->non_ssa_vars); | |
1199 | } | |
1200 | if (prev->earliest > entry->earliest) | |
1201 | prev->earliest = entry->earliest; | |
1202 | prev->overall_time += entry->overall_time; | |
1203 | prev->overall_size += entry->overall_size; | |
1204 | BITMAP_FREE (entry->set_ssa_names); | |
1205 | BITMAP_FREE (entry->used_ssa_names); | |
1206 | BITMAP_FREE (entry->bbs_visited); | |
1207 | BITMAP_FREE (entry->non_ssa_vars); | |
f1f41a6c | 1208 | stack.pop (); |
2862cf88 | 1209 | } |
1210 | else | |
f1f41a6c | 1211 | stack.pop (); |
2862cf88 | 1212 | } |
34154e27 | 1213 | ENTRY_BLOCK_PTR_FOR_FN (cfun)->aux = NULL; |
fc00614f | 1214 | FOR_EACH_BB_FN (bb, cfun) |
2862cf88 | 1215 | bb->aux = NULL; |
f1f41a6c | 1216 | stack.release (); |
2862cf88 | 1217 | BITMAP_FREE (current.ssa_names_to_pass); |
1218 | } | |
1219 | ||
1220 | /* Split function at SPLIT_POINT. */ | |
1221 | ||
1222 | static void | |
a63cfc4d | 1223 | split_function (basic_block return_bb, struct split_point *split_point, |
1224 | bool add_tsan_func_exit) | |
2862cf88 | 1225 | { |
1e094109 | 1226 | vec<tree> args_to_pass = vNULL; |
982ffd8d | 1227 | bitmap args_to_skip; |
2862cf88 | 1228 | tree parm; |
1229 | int num = 0; | |
415d1b9a | 1230 | cgraph_node *node, *cur_node = cgraph_node::get (current_function_decl); |
2862cf88 | 1231 | basic_block call_bb; |
a63cfc4d | 1232 | gcall *call, *tsan_func_exit_call = NULL; |
2862cf88 | 1233 | edge e; |
1234 | edge_iterator ei; | |
058a1b7a | 1235 | tree retval = NULL, real_retval = NULL, retbnd = NULL; |
2862cf88 | 1236 | bool split_part_return_p = false; |
058a1b7a | 1237 | bool with_bounds = chkp_function_instrumented_p (current_function_decl); |
2862cf88 | 1238 | gimple last_stmt = NULL; |
e65564ab | 1239 | unsigned int i; |
c6dfe037 | 1240 | tree arg, ddef; |
f1f41a6c | 1241 | vec<tree, va_gc> **debug_args = NULL; |
2862cf88 | 1242 | |
1243 | if (dump_file) | |
1244 | { | |
1245 | fprintf (dump_file, "\n\nSplitting function at:\n"); | |
1246 | dump_split_point (dump_file, split_point); | |
1247 | } | |
1248 | ||
982ffd8d | 1249 | if (cur_node->local.can_change_signature) |
1250 | args_to_skip = BITMAP_ALLOC (NULL); | |
1251 | else | |
1252 | args_to_skip = NULL; | |
1253 | ||
2862cf88 | 1254 | /* Collect the parameters of new function and args_to_skip bitmap. */ |
1255 | for (parm = DECL_ARGUMENTS (current_function_decl); | |
1767a056 | 1256 | parm; parm = DECL_CHAIN (parm), num++) |
982ffd8d | 1257 | if (args_to_skip |
1258 | && (!is_gimple_reg (parm) | |
c6dfe037 | 1259 | || (ddef = ssa_default_def (cfun, parm)) == NULL_TREE |
982ffd8d | 1260 | || !bitmap_bit_p (split_point->ssa_names_to_pass, |
c6dfe037 | 1261 | SSA_NAME_VERSION (ddef)))) |
2862cf88 | 1262 | bitmap_set_bit (args_to_skip, num); |
1263 | else | |
e65564ab | 1264 | { |
17c5f86d | 1265 | /* This parm might not have been used up to now, but is going to be |
1266 | used, hence register it. */ | |
17c5f86d | 1267 | if (is_gimple_reg (parm)) |
c6dfe037 | 1268 | arg = get_or_create_ssa_default_def (cfun, parm); |
17c5f86d | 1269 | else |
1270 | arg = parm; | |
982ffd8d | 1271 | |
0ed25048 | 1272 | if (!useless_type_conversion_p (DECL_ARG_TYPE (parm), TREE_TYPE (arg))) |
1273 | arg = fold_convert (DECL_ARG_TYPE (parm), arg); | |
f1f41a6c | 1274 | args_to_pass.safe_push (arg); |
e65564ab | 1275 | } |
2862cf88 | 1276 | |
1277 | /* See if the split function will return. */ | |
1278 | FOR_EACH_EDGE (e, ei, return_bb->preds) | |
1279 | if (bitmap_bit_p (split_point->split_bbs, e->src->index)) | |
1280 | break; | |
1281 | if (e) | |
1282 | split_part_return_p = true; | |
1283 | ||
b04bab7c | 1284 | /* Add return block to what will become the split function. |
1285 | We do not return; no return block is needed. */ | |
1286 | if (!split_part_return_p) | |
1287 | ; | |
1288 | /* We have no return block, so nothing is needed. */ | |
34154e27 | 1289 | else if (return_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)) |
b04bab7c | 1290 | ; |
1291 | /* When we do not want to return value, we need to construct | |
1292 | new return block with empty return statement. | |
1293 | FIXME: Once we are able to change return type, we should change function | |
1294 | to return void instead of just outputting function with undefined return | |
1295 | value. For structures this affects quality of codegen. */ | |
1296 | else if (!split_point->split_part_set_retval | |
1297 | && find_retval (return_bb)) | |
1298 | { | |
1299 | bool redirected = true; | |
1300 | basic_block new_return_bb = create_basic_block (NULL, 0, return_bb); | |
1301 | gimple_stmt_iterator gsi = gsi_start_bb (new_return_bb); | |
1302 | gsi_insert_after (&gsi, gimple_build_return (NULL), GSI_NEW_STMT); | |
1303 | while (redirected) | |
1304 | { | |
1305 | redirected = false; | |
1306 | FOR_EACH_EDGE (e, ei, return_bb->preds) | |
1307 | if (bitmap_bit_p (split_point->split_bbs, e->src->index)) | |
1308 | { | |
1309 | new_return_bb->count += e->count; | |
1310 | new_return_bb->frequency += EDGE_FREQUENCY (e); | |
1311 | redirect_edge_and_branch (e, new_return_bb); | |
1312 | redirected = true; | |
1313 | break; | |
1314 | } | |
1315 | } | |
34154e27 | 1316 | e = make_edge (new_return_bb, EXIT_BLOCK_PTR_FOR_FN (cfun), 0); |
b04bab7c | 1317 | e->probability = REG_BR_PROB_BASE; |
1318 | e->count = new_return_bb->count; | |
b3083327 | 1319 | add_bb_to_loop (new_return_bb, current_loops->tree_root); |
b04bab7c | 1320 | bitmap_set_bit (split_point->split_bbs, new_return_bb->index); |
c13d6d8f | 1321 | } |
1322 | /* When we pass around the value, use existing return block. */ | |
1323 | else | |
1324 | bitmap_set_bit (split_point->split_bbs, return_bb->index); | |
1325 | ||
1326 | /* If RETURN_BB has virtual operand PHIs, they must be removed and the | |
1327 | virtual operand marked for renaming as we change the CFG in a way that | |
1d0b727d | 1328 | tree-inline is not able to compensate for. |
c13d6d8f | 1329 | |
1330 | Note this can happen whether or not we have a return value. If we have | |
1331 | a return value, then RETURN_BB may have PHIs for real operands too. */ | |
34154e27 | 1332 | if (return_bb != EXIT_BLOCK_PTR_FOR_FN (cfun)) |
c13d6d8f | 1333 | { |
1d0b727d | 1334 | bool phi_p = false; |
1a91d914 | 1335 | for (gphi_iterator gsi = gsi_start_phis (return_bb); |
1336 | !gsi_end_p (gsi);) | |
b04bab7c | 1337 | { |
1a91d914 | 1338 | gphi *stmt = gsi.phi (); |
7c782c9b | 1339 | if (!virtual_operand_p (gimple_phi_result (stmt))) |
c13d6d8f | 1340 | { |
1341 | gsi_next (&gsi); | |
1342 | continue; | |
1343 | } | |
904ad591 | 1344 | mark_virtual_phi_result_for_renaming (stmt); |
1345 | remove_phi_node (&gsi, true); | |
1d0b727d | 1346 | phi_p = true; |
b04bab7c | 1347 | } |
1d0b727d | 1348 | /* In reality we have to rename the reaching definition of the |
1349 | virtual operand at return_bb as we will eventually release it | |
1350 | when we remove the code region we outlined. | |
1351 | So we have to rename all immediate virtual uses of that region | |
1352 | if we didn't see a PHI definition yet. */ | |
1353 | /* ??? In real reality we want to set the reaching vdef of the | |
1354 | entry of the SESE region as the vuse of the call and the reaching | |
1355 | vdef of the exit of the SESE region as the vdef of the call. */ | |
1356 | if (!phi_p) | |
1a91d914 | 1357 | for (gimple_stmt_iterator gsi = gsi_start_bb (return_bb); |
1358 | !gsi_end_p (gsi); | |
1359 | gsi_next (&gsi)) | |
1d0b727d | 1360 | { |
1361 | gimple stmt = gsi_stmt (gsi); | |
1362 | if (gimple_vuse (stmt)) | |
1363 | { | |
1364 | gimple_set_vuse (stmt, NULL_TREE); | |
1365 | update_stmt (stmt); | |
1366 | } | |
1367 | if (gimple_vdef (stmt)) | |
1368 | break; | |
1369 | } | |
b04bab7c | 1370 | } |
2862cf88 | 1371 | |
1372 | /* Now create the actual clone. */ | |
35ee1c66 | 1373 | cgraph_edge::rebuild_edges (); |
415d1b9a | 1374 | node = cur_node->create_version_clone_with_body |
1375 | (vNULL, NULL, args_to_skip, !split_part_return_p, split_point->split_bbs, | |
1376 | split_point->entry_bb, "part"); | |
af48f0b1 | 1377 | |
71e37927 | 1378 | node->split_part = true; |
1379 | ||
af48f0b1 | 1380 | /* Let's take a time profile for splitted function. */ |
1381 | node->tp_first_run = cur_node->tp_first_run + 1; | |
1382 | ||
fd8d648f | 1383 | /* For usual cloning it is enough to clear builtin only when signature |
1384 | changes. For partial inlining we however can not expect the part | |
1385 | of builtin implementation to have same semantic as the whole. */ | |
02774f2d | 1386 | if (DECL_BUILT_IN (node->decl)) |
fd8d648f | 1387 | { |
02774f2d | 1388 | DECL_BUILT_IN_CLASS (node->decl) = NOT_BUILT_IN; |
1389 | DECL_FUNCTION_CODE (node->decl) = (enum built_in_function) 0; | |
fd8d648f | 1390 | } |
058a1b7a | 1391 | |
1392 | /* If the original function is instrumented then it's | |
1393 | part is also instrumented. */ | |
1394 | if (with_bounds) | |
1395 | chkp_function_mark_instrumented (node->decl); | |
1396 | ||
ea03e543 | 1397 | /* If the original function is declared inline, there is no point in issuing |
1398 | a warning for the non-inlinable part. */ | |
02774f2d | 1399 | DECL_NO_INLINE_WARNING_P (node->decl) = 1; |
415d1b9a | 1400 | cur_node->remove_callees (); |
51ce5652 | 1401 | cur_node->remove_all_references (); |
2862cf88 | 1402 | if (!split_part_return_p) |
02774f2d | 1403 | TREE_THIS_VOLATILE (node->decl) = 1; |
2862cf88 | 1404 | if (dump_file) |
02774f2d | 1405 | dump_function_to_file (node->decl, dump_file, dump_flags); |
2862cf88 | 1406 | |
1407 | /* Create the basic block we place call into. It is the entry basic block | |
1408 | split after last label. */ | |
1409 | call_bb = split_point->entry_bb; | |
1a91d914 | 1410 | for (gimple_stmt_iterator gsi = gsi_start_bb (call_bb); !gsi_end_p (gsi);) |
2862cf88 | 1411 | if (gimple_code (gsi_stmt (gsi)) == GIMPLE_LABEL) |
1412 | { | |
1413 | last_stmt = gsi_stmt (gsi); | |
1414 | gsi_next (&gsi); | |
1415 | } | |
1416 | else | |
1417 | break; | |
1418 | e = split_block (split_point->entry_bb, last_stmt); | |
1419 | remove_edge (e); | |
1420 | ||
1421 | /* Produce the call statement. */ | |
1a91d914 | 1422 | gimple_stmt_iterator gsi = gsi_last_bb (call_bb); |
f1f41a6c | 1423 | FOR_EACH_VEC_ELT (args_to_pass, i, arg) |
0ed25048 | 1424 | if (!is_gimple_val (arg)) |
1425 | { | |
1426 | arg = force_gimple_operand_gsi (&gsi, arg, true, NULL_TREE, | |
ad676199 | 1427 | false, GSI_CONTINUE_LINKING); |
f1f41a6c | 1428 | args_to_pass[i] = arg; |
0ed25048 | 1429 | } |
02774f2d | 1430 | call = gimple_build_call_vec (node->decl, args_to_pass); |
058a1b7a | 1431 | gimple_call_set_with_bounds (call, with_bounds); |
2862cf88 | 1432 | gimple_set_block (call, DECL_INITIAL (current_function_decl)); |
f1f41a6c | 1433 | args_to_pass.release (); |
2862cf88 | 1434 | |
eee0ea7c | 1435 | /* For optimized away parameters, add on the caller side |
1436 | before the call | |
1437 | DEBUG D#X => parm_Y(D) | |
1438 | stmts and associate D#X with parm in decl_debug_args_lookup | |
1439 | vector to say for debug info that if parameter parm had been passed, | |
1440 | it would have value parm_Y(D). */ | |
1441 | if (args_to_skip) | |
1442 | for (parm = DECL_ARGUMENTS (current_function_decl), num = 0; | |
1443 | parm; parm = DECL_CHAIN (parm), num++) | |
1444 | if (bitmap_bit_p (args_to_skip, num) | |
1445 | && is_gimple_reg (parm)) | |
1446 | { | |
1447 | tree ddecl; | |
1448 | gimple def_temp; | |
1449 | ||
1450 | /* This needs to be done even without MAY_HAVE_DEBUG_STMTS, | |
1451 | otherwise if it didn't exist before, we'd end up with | |
1452 | different SSA_NAME_VERSIONs between -g and -g0. */ | |
1453 | arg = get_or_create_ssa_default_def (cfun, parm); | |
1454 | if (!MAY_HAVE_DEBUG_STMTS) | |
1455 | continue; | |
1456 | ||
1457 | if (debug_args == NULL) | |
02774f2d | 1458 | debug_args = decl_debug_args_insert (node->decl); |
eee0ea7c | 1459 | ddecl = make_node (DEBUG_EXPR_DECL); |
1460 | DECL_ARTIFICIAL (ddecl) = 1; | |
1461 | TREE_TYPE (ddecl) = TREE_TYPE (parm); | |
1462 | DECL_MODE (ddecl) = DECL_MODE (parm); | |
f1f41a6c | 1463 | vec_safe_push (*debug_args, DECL_ORIGIN (parm)); |
1464 | vec_safe_push (*debug_args, ddecl); | |
eee0ea7c | 1465 | def_temp = gimple_build_debug_bind (ddecl, unshare_expr (arg), |
1466 | call); | |
1467 | gsi_insert_after (&gsi, def_temp, GSI_NEW_STMT); | |
1468 | } | |
1469 | /* And on the callee side, add | |
1470 | DEBUG D#Y s=> parm | |
1471 | DEBUG var => D#Y | |
1472 | stmts to the first bb where var is a VAR_DECL created for the | |
1473 | optimized away parameter in DECL_INITIAL block. This hints | |
1474 | in the debug info that var (whole DECL_ORIGIN is the parm PARM_DECL) | |
1475 | is optimized away, but could be looked up at the call site | |
1476 | as value of D#X there. */ | |
1477 | if (debug_args != NULL) | |
1478 | { | |
1479 | unsigned int i; | |
1480 | tree var, vexpr; | |
1481 | gimple_stmt_iterator cgsi; | |
1482 | gimple def_temp; | |
1483 | ||
02774f2d | 1484 | push_cfun (DECL_STRUCT_FUNCTION (node->decl)); |
1485 | var = BLOCK_VARS (DECL_INITIAL (node->decl)); | |
f1f41a6c | 1486 | i = vec_safe_length (*debug_args); |
34154e27 | 1487 | cgsi = gsi_after_labels (single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun))); |
eee0ea7c | 1488 | do |
1489 | { | |
1490 | i -= 2; | |
1491 | while (var != NULL_TREE | |
f1f41a6c | 1492 | && DECL_ABSTRACT_ORIGIN (var) != (**debug_args)[i]) |
eee0ea7c | 1493 | var = TREE_CHAIN (var); |
1494 | if (var == NULL_TREE) | |
1495 | break; | |
1496 | vexpr = make_node (DEBUG_EXPR_DECL); | |
f1f41a6c | 1497 | parm = (**debug_args)[i]; |
eee0ea7c | 1498 | DECL_ARTIFICIAL (vexpr) = 1; |
1499 | TREE_TYPE (vexpr) = TREE_TYPE (parm); | |
1500 | DECL_MODE (vexpr) = DECL_MODE (parm); | |
1501 | def_temp = gimple_build_debug_source_bind (vexpr, parm, | |
1502 | NULL); | |
1503 | gsi_insert_before (&cgsi, def_temp, GSI_SAME_STMT); | |
1504 | def_temp = gimple_build_debug_bind (var, vexpr, NULL); | |
1505 | gsi_insert_before (&cgsi, def_temp, GSI_SAME_STMT); | |
1506 | } | |
1507 | while (i); | |
1508 | pop_cfun (); | |
1509 | } | |
1510 | ||
a8005893 | 1511 | /* We avoid address being taken on any variable used by split part, |
1512 | so return slot optimization is always possible. Moreover this is | |
1513 | required to make DECL_BY_REFERENCE work. */ | |
1514 | if (aggregate_value_p (DECL_RESULT (current_function_decl), | |
8ad1eb69 | 1515 | TREE_TYPE (current_function_decl)) |
1516 | && (!is_gimple_reg_type (TREE_TYPE (DECL_RESULT (current_function_decl))) | |
1517 | || DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)))) | |
a8005893 | 1518 | gimple_call_set_return_slot_opt (call, true); |
1519 | ||
a63cfc4d | 1520 | if (add_tsan_func_exit) |
1521 | tsan_func_exit_call = gimple_build_call_internal (IFN_TSAN_FUNC_EXIT, 0); | |
1522 | ||
2862cf88 | 1523 | /* Update return value. This is bit tricky. When we do not return, |
1524 | do nothing. When we return we might need to update return_bb | |
1525 | or produce a new return statement. */ | |
1526 | if (!split_part_return_p) | |
a63cfc4d | 1527 | { |
1528 | gsi_insert_after (&gsi, call, GSI_NEW_STMT); | |
1529 | if (tsan_func_exit_call) | |
1530 | gsi_insert_after (&gsi, tsan_func_exit_call, GSI_NEW_STMT); | |
1531 | } | |
2862cf88 | 1532 | else |
1533 | { | |
1534 | e = make_edge (call_bb, return_bb, | |
34154e27 | 1535 | return_bb == EXIT_BLOCK_PTR_FOR_FN (cfun) |
1536 | ? 0 : EDGE_FALLTHRU); | |
2862cf88 | 1537 | e->count = call_bb->count; |
1538 | e->probability = REG_BR_PROB_BASE; | |
524a0531 | 1539 | |
1540 | /* If there is return basic block, see what value we need to store | |
1541 | return value into and put call just before it. */ | |
34154e27 | 1542 | if (return_bb != EXIT_BLOCK_PTR_FOR_FN (cfun)) |
2862cf88 | 1543 | { |
4493dab3 | 1544 | real_retval = retval = find_retval (return_bb); |
058a1b7a | 1545 | retbnd = find_retbnd (return_bb); |
524a0531 | 1546 | |
b04bab7c | 1547 | if (real_retval && split_point->split_part_set_retval) |
2862cf88 | 1548 | { |
1a91d914 | 1549 | gphi_iterator psi; |
2862cf88 | 1550 | |
524a0531 | 1551 | /* See if we need new SSA_NAME for the result. |
1552 | When DECL_BY_REFERENCE is true, retval is actually pointer to | |
1553 | return value and it is constant in whole function. */ | |
1554 | if (TREE_CODE (retval) == SSA_NAME | |
1555 | && !DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) | |
2862cf88 | 1556 | { |
874117c8 | 1557 | retval = copy_ssa_name (retval, call); |
524a0531 | 1558 | |
1559 | /* See if there is PHI defining return value. */ | |
1560 | for (psi = gsi_start_phis (return_bb); | |
1561 | !gsi_end_p (psi); gsi_next (&psi)) | |
1a91d914 | 1562 | if (!virtual_operand_p (gimple_phi_result (psi.phi ()))) |
524a0531 | 1563 | break; |
1564 | ||
1565 | /* When there is PHI, just update its value. */ | |
2862cf88 | 1566 | if (TREE_CODE (retval) == SSA_NAME |
1567 | && !gsi_end_p (psi)) | |
1a91d914 | 1568 | add_phi_arg (psi.phi (), retval, e, UNKNOWN_LOCATION); |
524a0531 | 1569 | /* Otherwise update the return BB itself. |
1570 | find_return_bb allows at most one assignment to return value, | |
1571 | so update first statement. */ | |
1572 | else | |
2862cf88 | 1573 | { |
4493dab3 | 1574 | gimple_stmt_iterator bsi; |
1575 | for (bsi = gsi_start_bb (return_bb); !gsi_end_p (bsi); | |
1576 | gsi_next (&bsi)) | |
1a91d914 | 1577 | if (greturn *return_stmt |
1578 | = dyn_cast <greturn *> (gsi_stmt (bsi))) | |
4493dab3 | 1579 | { |
1a91d914 | 1580 | gimple_return_set_retval (return_stmt, retval); |
4493dab3 | 1581 | break; |
1582 | } | |
0f41c0cf | 1583 | else if (gimple_code (gsi_stmt (bsi)) == GIMPLE_ASSIGN |
1584 | && !gimple_clobber_p (gsi_stmt (bsi))) | |
4493dab3 | 1585 | { |
1586 | gimple_assign_set_rhs1 (gsi_stmt (bsi), retval); | |
1587 | break; | |
1588 | } | |
1589 | update_stmt (gsi_stmt (bsi)); | |
2862cf88 | 1590 | } |
058a1b7a | 1591 | |
1592 | /* Replace retbnd with new one. */ | |
1593 | if (retbnd) | |
1594 | { | |
1595 | gimple_stmt_iterator bsi; | |
1596 | for (bsi = gsi_last_bb (return_bb); !gsi_end_p (bsi); | |
1597 | gsi_prev (&bsi)) | |
1598 | if (gimple_code (gsi_stmt (bsi)) == GIMPLE_RETURN) | |
1599 | { | |
1600 | retbnd = copy_ssa_name (retbnd, call); | |
1601 | gimple_return_set_retbnd (gsi_stmt (bsi), retbnd); | |
1602 | update_stmt (gsi_stmt (bsi)); | |
1603 | break; | |
1604 | } | |
1605 | } | |
2862cf88 | 1606 | } |
a8005893 | 1607 | if (DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) |
9ece8f57 | 1608 | { |
1609 | gimple_call_set_lhs (call, build_simple_mem_ref (retval)); | |
1610 | gsi_insert_after (&gsi, call, GSI_NEW_STMT); | |
1611 | } | |
a8005893 | 1612 | else |
9ece8f57 | 1613 | { |
1614 | tree restype; | |
1615 | restype = TREE_TYPE (DECL_RESULT (current_function_decl)); | |
1616 | gsi_insert_after (&gsi, call, GSI_NEW_STMT); | |
1617 | if (!useless_type_conversion_p (TREE_TYPE (retval), restype)) | |
1618 | { | |
1619 | gimple cpy; | |
f9e245b2 | 1620 | tree tem = create_tmp_reg (restype); |
9ece8f57 | 1621 | tem = make_ssa_name (tem, call); |
e9cf809e | 1622 | cpy = gimple_build_assign (retval, NOP_EXPR, tem); |
9ece8f57 | 1623 | gsi_insert_after (&gsi, cpy, GSI_NEW_STMT); |
1624 | retval = tem; | |
1625 | } | |
058a1b7a | 1626 | /* Build bndret call to obtain returned bounds. */ |
1627 | if (retbnd) | |
17c596cc | 1628 | chkp_insert_retbnd_call (retbnd, retval, &gsi); |
9ece8f57 | 1629 | gimple_call_set_lhs (call, retval); |
1630 | update_stmt (call); | |
1631 | } | |
2862cf88 | 1632 | } |
9ece8f57 | 1633 | else |
1634 | gsi_insert_after (&gsi, call, GSI_NEW_STMT); | |
a63cfc4d | 1635 | if (tsan_func_exit_call) |
1636 | gsi_insert_after (&gsi, tsan_func_exit_call, GSI_NEW_STMT); | |
2862cf88 | 1637 | } |
524a0531 | 1638 | /* We don't use return block (there is either no return in function or |
1639 | multiple of them). So create new basic block with return statement. | |
1640 | */ | |
2862cf88 | 1641 | else |
1642 | { | |
1a91d914 | 1643 | greturn *ret; |
b04bab7c | 1644 | if (split_point->split_part_set_retval |
1645 | && !VOID_TYPE_P (TREE_TYPE (TREE_TYPE (current_function_decl)))) | |
2862cf88 | 1646 | { |
1b268210 | 1647 | retval = DECL_RESULT (current_function_decl); |
9eb0b8ee | 1648 | |
058a1b7a | 1649 | if (chkp_function_instrumented_p (current_function_decl) |
1650 | && BOUNDED_P (retval)) | |
f9e245b2 | 1651 | retbnd = create_tmp_reg (pointer_bounds_type_node); |
058a1b7a | 1652 | |
9eb0b8ee | 1653 | /* We use temporary register to hold value when aggregate_value_p |
1654 | is false. Similarly for DECL_BY_REFERENCE we must avoid extra | |
1655 | copy. */ | |
1656 | if (!aggregate_value_p (retval, TREE_TYPE (current_function_decl)) | |
1657 | && !DECL_BY_REFERENCE (retval)) | |
f9e245b2 | 1658 | retval = create_tmp_reg (TREE_TYPE (retval)); |
2862cf88 | 1659 | if (is_gimple_reg (retval)) |
524a0531 | 1660 | { |
1661 | /* When returning by reference, there is only one SSA name | |
1662 | assigned to RESULT_DECL (that is pointer to return value). | |
1663 | Look it up or create new one if it is missing. */ | |
1664 | if (DECL_BY_REFERENCE (retval)) | |
c6dfe037 | 1665 | retval = get_or_create_ssa_default_def (cfun, retval); |
524a0531 | 1666 | /* Otherwise produce new SSA name for return value. */ |
1667 | else | |
1668 | retval = make_ssa_name (retval, call); | |
1669 | } | |
a8005893 | 1670 | if (DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) |
1671 | gimple_call_set_lhs (call, build_simple_mem_ref (retval)); | |
1672 | else | |
1673 | gimple_call_set_lhs (call, retval); | |
2862cf88 | 1674 | } |
1675 | gsi_insert_after (&gsi, call, GSI_NEW_STMT); | |
058a1b7a | 1676 | /* Build bndret call to obtain returned bounds. */ |
1677 | if (retbnd) | |
17c596cc | 1678 | chkp_insert_retbnd_call (retbnd, retval, &gsi); |
a63cfc4d | 1679 | if (tsan_func_exit_call) |
1680 | gsi_insert_after (&gsi, tsan_func_exit_call, GSI_NEW_STMT); | |
2862cf88 | 1681 | ret = gimple_build_return (retval); |
1682 | gsi_insert_after (&gsi, ret, GSI_NEW_STMT); | |
1683 | } | |
1684 | } | |
1685 | free_dominance_info (CDI_DOMINATORS); | |
1686 | free_dominance_info (CDI_POST_DOMINATORS); | |
a41f2a28 | 1687 | compute_inline_parameters (node, true); |
2862cf88 | 1688 | } |
1689 | ||
1690 | /* Execute function splitting pass. */ | |
1691 | ||
1692 | static unsigned int | |
1693 | execute_split_functions (void) | |
1694 | { | |
1695 | gimple_stmt_iterator bsi; | |
1696 | basic_block bb; | |
1697 | int overall_time = 0, overall_size = 0; | |
1698 | int todo = 0; | |
415d1b9a | 1699 | struct cgraph_node *node = cgraph_node::get (current_function_decl); |
2862cf88 | 1700 | |
f3941af5 | 1701 | if (flags_from_decl_or_type (current_function_decl) |
1702 | & (ECF_NORETURN|ECF_MALLOC)) | |
2862cf88 | 1703 | { |
1704 | if (dump_file) | |
f3941af5 | 1705 | fprintf (dump_file, "Not splitting: noreturn/malloc function.\n"); |
2862cf88 | 1706 | return 0; |
1707 | } | |
1708 | if (MAIN_NAME_P (DECL_NAME (current_function_decl))) | |
1709 | { | |
1710 | if (dump_file) | |
1711 | fprintf (dump_file, "Not splitting: main function.\n"); | |
1712 | return 0; | |
1713 | } | |
1714 | /* This can be relaxed; function might become inlinable after splitting | |
1715 | away the uninlinable part. */ | |
f1f41a6c | 1716 | if (inline_edge_summary_vec.exists () |
b4bae7a0 | 1717 | && !inline_summaries->get (node)->inlinable) |
2862cf88 | 1718 | { |
1719 | if (dump_file) | |
1720 | fprintf (dump_file, "Not splitting: not inlinable.\n"); | |
1721 | return 0; | |
1722 | } | |
02774f2d | 1723 | if (DECL_DISREGARD_INLINE_LIMITS (node->decl)) |
2862cf88 | 1724 | { |
1725 | if (dump_file) | |
574c4ab5 | 1726 | fprintf (dump_file, "Not splitting: disregarding inline limits.\n"); |
2862cf88 | 1727 | return 0; |
1728 | } | |
1729 | /* This can be relaxed; most of versioning tests actually prevents | |
1730 | a duplication. */ | |
1731 | if (!tree_versionable_function_p (current_function_decl)) | |
1732 | { | |
1733 | if (dump_file) | |
1734 | fprintf (dump_file, "Not splitting: not versionable.\n"); | |
1735 | return 0; | |
1736 | } | |
1737 | /* FIXME: we could support this. */ | |
1738 | if (DECL_STRUCT_FUNCTION (current_function_decl)->static_chain_decl) | |
1739 | { | |
1740 | if (dump_file) | |
1741 | fprintf (dump_file, "Not splitting: nested function.\n"); | |
1742 | return 0; | |
1743 | } | |
2862cf88 | 1744 | |
1745 | /* See if it makes sense to try to split. | |
1746 | It makes sense to split if we inline, that is if we have direct calls to | |
1747 | handle or direct calls are possibly going to appear as result of indirect | |
a15d5ede | 1748 | inlining or LTO. Also handle -fprofile-generate as LTO to allow non-LTO |
1749 | training for LTO -fprofile-use build. | |
1750 | ||
2862cf88 | 1751 | Note that we are not completely conservative about disqualifying functions |
1752 | called once. It is possible that the caller is called more then once and | |
1753 | then inlining would still benefit. */ | |
8745ccbf | 1754 | if ((!node->callers |
1755 | /* Local functions called once will be completely inlined most of time. */ | |
1756 | || (!node->callers->next_caller && node->local.local)) | |
02774f2d | 1757 | && !node->address_taken |
c670165e | 1758 | && !node->has_aliases_p () |
02774f2d | 1759 | && (!flag_lto || !node->externally_visible)) |
2862cf88 | 1760 | { |
1761 | if (dump_file) | |
1762 | fprintf (dump_file, "Not splitting: not called directly " | |
1763 | "or called once.\n"); | |
1764 | return 0; | |
1765 | } | |
1766 | ||
1767 | /* FIXME: We can actually split if splitting reduces call overhead. */ | |
1768 | if (!flag_inline_small_functions | |
1769 | && !DECL_DECLARED_INLINE_P (current_function_decl)) | |
1770 | { | |
1771 | if (dump_file) | |
1772 | fprintf (dump_file, "Not splitting: not autoinlining and function" | |
1773 | " is not inline.\n"); | |
1774 | return 0; | |
1775 | } | |
1776 | ||
15ca8f90 | 1777 | /* We enforce splitting after loop headers when profile info is not |
1778 | available. */ | |
f26d8580 | 1779 | if (profile_status_for_fn (cfun) != PROFILE_READ) |
15ca8f90 | 1780 | mark_dfs_back_edges (); |
1781 | ||
aa2edcb9 | 1782 | /* Initialize bitmap to track forbidden calls. */ |
1783 | forbidden_dominators = BITMAP_ALLOC (NULL); | |
1784 | calculate_dominance_info (CDI_DOMINATORS); | |
1785 | ||
2862cf88 | 1786 | /* Compute local info about basic blocks and determine function size/time. */ |
fe672ac0 | 1787 | bb_info_vec.safe_grow_cleared (last_basic_block_for_fn (cfun) + 1); |
2862cf88 | 1788 | memset (&best_split_point, 0, sizeof (best_split_point)); |
a63cfc4d | 1789 | basic_block return_bb = find_return_bb (); |
1790 | int tsan_exit_found = -1; | |
fc00614f | 1791 | FOR_EACH_BB_FN (bb, cfun) |
2862cf88 | 1792 | { |
1793 | int time = 0; | |
1794 | int size = 0; | |
1795 | int freq = compute_call_stmt_bb_frequency (current_function_decl, bb); | |
1796 | ||
1797 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1798 | fprintf (dump_file, "Basic block %i\n", bb->index); | |
1799 | ||
1800 | for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
1801 | { | |
1802 | int this_time, this_size; | |
1803 | gimple stmt = gsi_stmt (bsi); | |
1804 | ||
1805 | this_size = estimate_num_insns (stmt, &eni_size_weights); | |
1806 | this_time = estimate_num_insns (stmt, &eni_time_weights) * freq; | |
1807 | size += this_size; | |
1808 | time += this_time; | |
aa2edcb9 | 1809 | check_forbidden_calls (stmt); |
2862cf88 | 1810 | |
1811 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1812 | { | |
1813 | fprintf (dump_file, " freq:%6i size:%3i time:%3i ", | |
1814 | freq, this_size, this_time); | |
1815 | print_gimple_stmt (dump_file, stmt, 0, 0); | |
1816 | } | |
a63cfc4d | 1817 | |
1818 | if ((flag_sanitize & SANITIZE_THREAD) | |
1819 | && is_gimple_call (stmt) | |
1820 | && gimple_call_internal_p (stmt) | |
1821 | && gimple_call_internal_fn (stmt) == IFN_TSAN_FUNC_EXIT) | |
1822 | { | |
1823 | /* We handle TSAN_FUNC_EXIT for splitting either in the | |
1824 | return_bb, or in its immediate predecessors. */ | |
1825 | if ((bb != return_bb && !find_edge (bb, return_bb)) | |
1826 | || (tsan_exit_found != -1 | |
1827 | && tsan_exit_found != (bb != return_bb))) | |
1828 | { | |
1829 | if (dump_file) | |
1830 | fprintf (dump_file, "Not splitting: TSAN_FUNC_EXIT" | |
1831 | " in unexpected basic block.\n"); | |
1832 | BITMAP_FREE (forbidden_dominators); | |
1833 | bb_info_vec.release (); | |
1834 | return 0; | |
1835 | } | |
1836 | tsan_exit_found = bb != return_bb; | |
1837 | } | |
2862cf88 | 1838 | } |
1839 | overall_time += time; | |
1840 | overall_size += size; | |
f1f41a6c | 1841 | bb_info_vec[bb->index].time = time; |
1842 | bb_info_vec[bb->index].size = size; | |
2862cf88 | 1843 | } |
a63cfc4d | 1844 | find_split_points (return_bb, overall_time, overall_size); |
2862cf88 | 1845 | if (best_split_point.split_bbs) |
1846 | { | |
a63cfc4d | 1847 | split_function (return_bb, &best_split_point, tsan_exit_found == 1); |
2862cf88 | 1848 | BITMAP_FREE (best_split_point.ssa_names_to_pass); |
1849 | BITMAP_FREE (best_split_point.split_bbs); | |
1850 | todo = TODO_update_ssa | TODO_cleanup_cfg; | |
1851 | } | |
aa2edcb9 | 1852 | BITMAP_FREE (forbidden_dominators); |
f1f41a6c | 1853 | bb_info_vec.release (); |
2862cf88 | 1854 | return todo; |
1855 | } | |
1856 | ||
cbe8bda8 | 1857 | namespace { |
1858 | ||
1859 | const pass_data pass_data_split_functions = | |
2862cf88 | 1860 | { |
cbe8bda8 | 1861 | GIMPLE_PASS, /* type */ |
1862 | "fnsplit", /* name */ | |
1863 | OPTGROUP_NONE, /* optinfo_flags */ | |
cbe8bda8 | 1864 | TV_IPA_FNSPLIT, /* tv_id */ |
1865 | PROP_cfg, /* properties_required */ | |
1866 | 0, /* properties_provided */ | |
1867 | 0, /* properties_destroyed */ | |
1868 | 0, /* todo_flags_start */ | |
8b88439e | 1869 | 0, /* todo_flags_finish */ |
2862cf88 | 1870 | }; |
a15d5ede | 1871 | |
cbe8bda8 | 1872 | class pass_split_functions : public gimple_opt_pass |
1873 | { | |
1874 | public: | |
9af5ce0c | 1875 | pass_split_functions (gcc::context *ctxt) |
1876 | : gimple_opt_pass (pass_data_split_functions, ctxt) | |
cbe8bda8 | 1877 | {} |
1878 | ||
1879 | /* opt_pass methods: */ | |
31315c24 | 1880 | virtual bool gate (function *); |
65b0537f | 1881 | virtual unsigned int execute (function *) |
1882 | { | |
1883 | return execute_split_functions (); | |
1884 | } | |
cbe8bda8 | 1885 | |
1886 | }; // class pass_split_functions | |
1887 | ||
31315c24 | 1888 | bool |
1889 | pass_split_functions::gate (function *) | |
1890 | { | |
1891 | /* When doing profile feedback, we want to execute the pass after profiling | |
1892 | is read. So disable one in early optimization. */ | |
1893 | return (flag_partial_inlining | |
1894 | && !profile_arc_flag && !flag_branch_probabilities); | |
1895 | } | |
1896 | ||
cbe8bda8 | 1897 | } // anon namespace |
1898 | ||
1899 | gimple_opt_pass * | |
1900 | make_pass_split_functions (gcc::context *ctxt) | |
1901 | { | |
1902 | return new pass_split_functions (ctxt); | |
1903 | } | |
1904 | ||
a15d5ede | 1905 | /* Execute function splitting pass. */ |
1906 | ||
1907 | static unsigned int | |
1908 | execute_feedback_split_functions (void) | |
1909 | { | |
1910 | unsigned int retval = execute_split_functions (); | |
1911 | if (retval) | |
1912 | retval |= TODO_rebuild_cgraph_edges; | |
1913 | return retval; | |
1914 | } | |
1915 | ||
cbe8bda8 | 1916 | namespace { |
1917 | ||
1918 | const pass_data pass_data_feedback_split_functions = | |
a15d5ede | 1919 | { |
cbe8bda8 | 1920 | GIMPLE_PASS, /* type */ |
1921 | "feedback_fnsplit", /* name */ | |
1922 | OPTGROUP_NONE, /* optinfo_flags */ | |
cbe8bda8 | 1923 | TV_IPA_FNSPLIT, /* tv_id */ |
1924 | PROP_cfg, /* properties_required */ | |
1925 | 0, /* properties_provided */ | |
1926 | 0, /* properties_destroyed */ | |
1927 | 0, /* todo_flags_start */ | |
8b88439e | 1928 | 0, /* todo_flags_finish */ |
a15d5ede | 1929 | }; |
cbe8bda8 | 1930 | |
1931 | class pass_feedback_split_functions : public gimple_opt_pass | |
1932 | { | |
1933 | public: | |
9af5ce0c | 1934 | pass_feedback_split_functions (gcc::context *ctxt) |
1935 | : gimple_opt_pass (pass_data_feedback_split_functions, ctxt) | |
cbe8bda8 | 1936 | {} |
1937 | ||
1938 | /* opt_pass methods: */ | |
31315c24 | 1939 | virtual bool gate (function *); |
65b0537f | 1940 | virtual unsigned int execute (function *) |
1941 | { | |
1942 | return execute_feedback_split_functions (); | |
1943 | } | |
cbe8bda8 | 1944 | |
1945 | }; // class pass_feedback_split_functions | |
1946 | ||
31315c24 | 1947 | bool |
1948 | pass_feedback_split_functions::gate (function *) | |
1949 | { | |
1950 | /* We don't need to split when profiling at all, we are producing | |
1951 | lousy code anyway. */ | |
1952 | return (flag_partial_inlining | |
1953 | && flag_branch_probabilities); | |
1954 | } | |
1955 | ||
cbe8bda8 | 1956 | } // anon namespace |
1957 | ||
1958 | gimple_opt_pass * | |
1959 | make_pass_feedback_split_functions (gcc::context *ctxt) | |
1960 | { | |
1961 | return new pass_feedback_split_functions (ctxt); | |
1962 | } |