]>
Commit | Line | Data |
---|---|---|
2862cf88 | 1 | /* Function splitting pass |
711789cc | 2 | Copyright (C) 2010-2013 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" | |
80 | #include "tree.h" | |
b23fb4cb | 81 | #include "gimple.h" |
2862cf88 | 82 | #include "target.h" |
2862cf88 | 83 | #include "ipa-prop.h" |
073c1fd5 | 84 | #include "gimple-ssa.h" |
85 | #include "tree-cfg.h" | |
86 | #include "tree-phinodes.h" | |
87 | #include "ssa-iterators.h" | |
88 | #include "tree-ssanames.h" | |
89 | #include "tree-into-ssa.h" | |
90 | #include "tree-dfa.h" | |
2862cf88 | 91 | #include "tree-pass.h" |
92 | #include "flags.h" | |
2862cf88 | 93 | #include "diagnostic.h" |
94 | #include "tree-dump.h" | |
95 | #include "tree-inline.h" | |
2862cf88 | 96 | #include "params.h" |
97 | #include "gimple-pretty-print.h" | |
cbd7f5a0 | 98 | #include "ipa-inline.h" |
f6568ea4 | 99 | #include "cfgloop.h" |
2862cf88 | 100 | |
101 | /* Per basic block info. */ | |
102 | ||
103 | typedef struct | |
104 | { | |
105 | unsigned int size; | |
106 | unsigned int time; | |
107 | } bb_info; | |
2862cf88 | 108 | |
f1f41a6c | 109 | static vec<bb_info> bb_info_vec; |
2862cf88 | 110 | |
111 | /* Description of split point. */ | |
112 | ||
113 | struct split_point | |
114 | { | |
115 | /* Size of the partitions. */ | |
116 | unsigned int header_time, header_size, split_time, split_size; | |
117 | ||
574c4ab5 | 118 | /* SSA names that need to be passed into spit function. */ |
2862cf88 | 119 | bitmap ssa_names_to_pass; |
120 | ||
121 | /* Basic block where we split (that will become entry point of new function. */ | |
122 | basic_block entry_bb; | |
123 | ||
124 | /* Basic blocks we are splitting away. */ | |
125 | bitmap split_bbs; | |
b04bab7c | 126 | |
127 | /* True when return value is computed on split part and thus it needs | |
128 | to be returned. */ | |
129 | bool split_part_set_retval; | |
2862cf88 | 130 | }; |
131 | ||
132 | /* Best split point found. */ | |
133 | ||
134 | struct split_point best_split_point; | |
135 | ||
aa2edcb9 | 136 | /* Set of basic blocks that are not allowed to dominate a split point. */ |
137 | ||
138 | static bitmap forbidden_dominators; | |
139 | ||
b04bab7c | 140 | static tree find_retval (basic_block return_bb); |
141 | ||
2dc0a33e | 142 | /* Callback for walk_stmt_load_store_addr_ops. If T is non-SSA automatic |
2862cf88 | 143 | variable, check it if it is present in bitmap passed via DATA. */ |
144 | ||
145 | static bool | |
2dc0a33e | 146 | test_nonssa_use (gimple stmt ATTRIBUTE_UNUSED, tree t, void *data) |
2862cf88 | 147 | { |
148 | t = get_base_address (t); | |
149 | ||
2dc0a33e | 150 | if (!t || is_gimple_reg (t)) |
151 | return false; | |
152 | ||
153 | if (TREE_CODE (t) == PARM_DECL | |
154 | || (TREE_CODE (t) == VAR_DECL | |
2862cf88 | 155 | && auto_var_in_fn_p (t, current_function_decl)) |
2dc0a33e | 156 | || TREE_CODE (t) == RESULT_DECL |
157 | || TREE_CODE (t) == LABEL_DECL) | |
2862cf88 | 158 | return bitmap_bit_p ((bitmap)data, DECL_UID (t)); |
b04bab7c | 159 | |
2dc0a33e | 160 | /* For DECL_BY_REFERENCE, the return value is actually a pointer. We want |
161 | to pretend that the value pointed to is actual result decl. */ | |
162 | if ((TREE_CODE (t) == MEM_REF || INDIRECT_REF_P (t)) | |
b04bab7c | 163 | && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME |
ec11736b | 164 | && SSA_NAME_VAR (TREE_OPERAND (t, 0)) |
b04bab7c | 165 | && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND (t, 0))) == RESULT_DECL |
166 | && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) | |
2dc0a33e | 167 | return |
168 | bitmap_bit_p ((bitmap)data, | |
169 | DECL_UID (DECL_RESULT (current_function_decl))); | |
170 | ||
2862cf88 | 171 | return false; |
172 | } | |
173 | ||
174 | /* Dump split point CURRENT. */ | |
175 | ||
176 | static void | |
177 | dump_split_point (FILE * file, struct split_point *current) | |
178 | { | |
179 | fprintf (file, | |
1d0b727d | 180 | "Split point at BB %i\n" |
181 | " header time: %i header size: %i\n" | |
182 | " split time: %i split size: %i\n bbs: ", | |
2862cf88 | 183 | current->entry_bb->index, current->header_time, |
184 | current->header_size, current->split_time, current->split_size); | |
185 | dump_bitmap (file, current->split_bbs); | |
186 | fprintf (file, " SSA names to pass: "); | |
187 | dump_bitmap (file, current->ssa_names_to_pass); | |
188 | } | |
189 | ||
2dc0a33e | 190 | /* Look for all BBs in header that might lead to the split part and verify |
191 | that they are not defining any non-SSA var used by the split part. | |
4493dab3 | 192 | Parameters are the same as for consider_split. */ |
193 | ||
194 | static bool | |
195 | verify_non_ssa_vars (struct split_point *current, bitmap non_ssa_vars, | |
196 | basic_block return_bb) | |
197 | { | |
198 | bitmap seen = BITMAP_ALLOC (NULL); | |
1e094109 | 199 | vec<basic_block> worklist = vNULL; |
4493dab3 | 200 | edge e; |
201 | edge_iterator ei; | |
202 | bool ok = true; | |
2dc0a33e | 203 | |
4493dab3 | 204 | FOR_EACH_EDGE (e, ei, current->entry_bb->preds) |
205 | if (e->src != ENTRY_BLOCK_PTR | |
206 | && !bitmap_bit_p (current->split_bbs, e->src->index)) | |
207 | { | |
f1f41a6c | 208 | worklist.safe_push (e->src); |
4493dab3 | 209 | bitmap_set_bit (seen, e->src->index); |
210 | } | |
2dc0a33e | 211 | |
f1f41a6c | 212 | while (!worklist.is_empty ()) |
4493dab3 | 213 | { |
214 | gimple_stmt_iterator bsi; | |
f1f41a6c | 215 | basic_block bb = worklist.pop (); |
4493dab3 | 216 | |
217 | FOR_EACH_EDGE (e, ei, bb->preds) | |
218 | if (e->src != ENTRY_BLOCK_PTR | |
6ef9bbe0 | 219 | && bitmap_set_bit (seen, e->src->index)) |
4493dab3 | 220 | { |
221 | gcc_checking_assert (!bitmap_bit_p (current->split_bbs, | |
222 | e->src->index)); | |
f1f41a6c | 223 | worklist.safe_push (e->src); |
4493dab3 | 224 | } |
225 | for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
226 | { | |
2dc0a33e | 227 | gimple stmt = gsi_stmt (bsi); |
228 | if (is_gimple_debug (stmt)) | |
4493dab3 | 229 | continue; |
230 | if (walk_stmt_load_store_addr_ops | |
2dc0a33e | 231 | (stmt, non_ssa_vars, test_nonssa_use, test_nonssa_use, |
232 | test_nonssa_use)) | |
4493dab3 | 233 | { |
234 | ok = false; | |
235 | goto done; | |
236 | } | |
2dc0a33e | 237 | if (gimple_code (stmt) == GIMPLE_LABEL |
238 | && test_nonssa_use (stmt, gimple_label_label (stmt), | |
239 | non_ssa_vars)) | |
240 | { | |
241 | ok = false; | |
242 | goto done; | |
243 | } | |
4493dab3 | 244 | } |
245 | for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
246 | { | |
247 | if (walk_stmt_load_store_addr_ops | |
2dc0a33e | 248 | (gsi_stmt (bsi), non_ssa_vars, test_nonssa_use, test_nonssa_use, |
249 | test_nonssa_use)) | |
4493dab3 | 250 | { |
251 | ok = false; | |
252 | goto done; | |
253 | } | |
254 | } | |
255 | FOR_EACH_EDGE (e, ei, bb->succs) | |
256 | { | |
257 | if (e->dest != return_bb) | |
258 | continue; | |
259 | for (bsi = gsi_start_phis (return_bb); !gsi_end_p (bsi); | |
260 | gsi_next (&bsi)) | |
261 | { | |
262 | gimple stmt = gsi_stmt (bsi); | |
263 | tree op = gimple_phi_arg_def (stmt, e->dest_idx); | |
264 | ||
7c782c9b | 265 | if (virtual_operand_p (gimple_phi_result (stmt))) |
4493dab3 | 266 | continue; |
267 | if (TREE_CODE (op) != SSA_NAME | |
268 | && test_nonssa_use (stmt, op, non_ssa_vars)) | |
269 | { | |
270 | ok = false; | |
271 | goto done; | |
272 | } | |
273 | } | |
274 | } | |
275 | } | |
276 | done: | |
277 | BITMAP_FREE (seen); | |
f1f41a6c | 278 | worklist.release (); |
4493dab3 | 279 | return ok; |
280 | } | |
281 | ||
aa2edcb9 | 282 | /* If STMT is a call, check the callee against a list of forbidden |
283 | predicate functions. If a match is found, look for uses of the | |
284 | call result in condition statements that compare against zero. | |
285 | For each such use, find the block targeted by the condition | |
286 | statement for the nonzero result, and set the bit for this block | |
287 | in the forbidden dominators bitmap. The purpose of this is to avoid | |
288 | selecting a split point where we are likely to lose the chance | |
289 | to optimize away an unused function call. */ | |
290 | ||
291 | static void | |
292 | check_forbidden_calls (gimple stmt) | |
293 | { | |
294 | imm_use_iterator use_iter; | |
295 | use_operand_p use_p; | |
296 | tree lhs; | |
297 | ||
298 | /* At the moment, __builtin_constant_p is the only forbidden | |
299 | predicate function call (see PR49642). */ | |
300 | if (!gimple_call_builtin_p (stmt, BUILT_IN_CONSTANT_P)) | |
301 | return; | |
302 | ||
303 | lhs = gimple_call_lhs (stmt); | |
304 | ||
305 | if (!lhs || TREE_CODE (lhs) != SSA_NAME) | |
306 | return; | |
307 | ||
308 | FOR_EACH_IMM_USE_FAST (use_p, use_iter, lhs) | |
309 | { | |
310 | tree op1; | |
311 | basic_block use_bb, forbidden_bb; | |
312 | enum tree_code code; | |
313 | edge true_edge, false_edge; | |
314 | gimple use_stmt = USE_STMT (use_p); | |
315 | ||
316 | if (gimple_code (use_stmt) != GIMPLE_COND) | |
317 | continue; | |
318 | ||
319 | /* Assuming canonical form for GIMPLE_COND here, with constant | |
320 | in second position. */ | |
321 | op1 = gimple_cond_rhs (use_stmt); | |
322 | code = gimple_cond_code (use_stmt); | |
323 | use_bb = gimple_bb (use_stmt); | |
324 | ||
325 | extract_true_false_edges_from_block (use_bb, &true_edge, &false_edge); | |
326 | ||
327 | /* We're only interested in comparisons that distinguish | |
328 | unambiguously from zero. */ | |
329 | if (!integer_zerop (op1) || code == LE_EXPR || code == GE_EXPR) | |
330 | continue; | |
331 | ||
332 | if (code == EQ_EXPR) | |
333 | forbidden_bb = false_edge->dest; | |
334 | else | |
335 | forbidden_bb = true_edge->dest; | |
336 | ||
337 | bitmap_set_bit (forbidden_dominators, forbidden_bb->index); | |
338 | } | |
339 | } | |
340 | ||
341 | /* If BB is dominated by any block in the forbidden dominators set, | |
342 | return TRUE; else FALSE. */ | |
343 | ||
344 | static bool | |
345 | dominated_by_forbidden (basic_block bb) | |
346 | { | |
347 | unsigned dom_bb; | |
348 | bitmap_iterator bi; | |
349 | ||
350 | EXECUTE_IF_SET_IN_BITMAP (forbidden_dominators, 1, dom_bb, bi) | |
351 | { | |
352 | if (dominated_by_p (CDI_DOMINATORS, bb, BASIC_BLOCK (dom_bb))) | |
353 | return true; | |
354 | } | |
355 | ||
356 | return false; | |
357 | } | |
358 | ||
2862cf88 | 359 | /* We found an split_point CURRENT. NON_SSA_VARS is bitmap of all non ssa |
360 | variables used and RETURN_BB is return basic block. | |
361 | See if we can split function here. */ | |
362 | ||
363 | static void | |
364 | consider_split (struct split_point *current, bitmap non_ssa_vars, | |
365 | basic_block return_bb) | |
366 | { | |
367 | tree parm; | |
368 | unsigned int num_args = 0; | |
369 | unsigned int call_overhead; | |
370 | edge e; | |
371 | edge_iterator ei; | |
6a69e813 | 372 | gimple_stmt_iterator bsi; |
373 | unsigned int i; | |
574c4ab5 | 374 | int incoming_freq = 0; |
b04bab7c | 375 | tree retval; |
15ca8f90 | 376 | bool back_edge = false; |
6a69e813 | 377 | |
2862cf88 | 378 | if (dump_file && (dump_flags & TDF_DETAILS)) |
379 | dump_split_point (dump_file, current); | |
380 | ||
6a69e813 | 381 | FOR_EACH_EDGE (e, ei, current->entry_bb->preds) |
15ca8f90 | 382 | { |
383 | if (e->flags & EDGE_DFS_BACK) | |
384 | back_edge = true; | |
385 | if (!bitmap_bit_p (current->split_bbs, e->src->index)) | |
386 | incoming_freq += EDGE_FREQUENCY (e); | |
387 | } | |
6a69e813 | 388 | |
2862cf88 | 389 | /* Do not split when we would end up calling function anyway. */ |
574c4ab5 | 390 | if (incoming_freq |
2862cf88 | 391 | >= (ENTRY_BLOCK_PTR->frequency |
392 | * PARAM_VALUE (PARAM_PARTIAL_INLINING_ENTRY_PROBABILITY) / 100)) | |
393 | { | |
15ca8f90 | 394 | /* When profile is guessed, we can not expect it to give us |
395 | realistic estimate on likelyness of function taking the | |
396 | complex path. As a special case, when tail of the function is | |
397 | a loop, enable splitting since inlining code skipping the loop | |
398 | is likely noticeable win. */ | |
399 | if (back_edge | |
400 | && profile_status != PROFILE_READ | |
401 | && incoming_freq < ENTRY_BLOCK_PTR->frequency) | |
402 | { | |
403 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
404 | fprintf (dump_file, | |
405 | " Split before loop, accepting despite low frequencies %i %i.\n", | |
406 | incoming_freq, | |
407 | ENTRY_BLOCK_PTR->frequency); | |
408 | } | |
409 | else | |
410 | { | |
411 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
412 | fprintf (dump_file, | |
413 | " Refused: incoming frequency is too large.\n"); | |
414 | return; | |
415 | } | |
2862cf88 | 416 | } |
417 | ||
418 | if (!current->header_size) | |
419 | { | |
420 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
421 | fprintf (dump_file, " Refused: header empty\n"); | |
2862cf88 | 422 | return; |
423 | } | |
424 | ||
574c4ab5 | 425 | /* Verify that PHI args on entry are either virtual or all their operands |
426 | incoming from header are the same. */ | |
6a69e813 | 427 | for (bsi = gsi_start_phis (current->entry_bb); !gsi_end_p (bsi); gsi_next (&bsi)) |
2862cf88 | 428 | { |
6a69e813 | 429 | gimple stmt = gsi_stmt (bsi); |
430 | tree val = NULL; | |
431 | ||
7c782c9b | 432 | if (virtual_operand_p (gimple_phi_result (stmt))) |
6a69e813 | 433 | continue; |
434 | for (i = 0; i < gimple_phi_num_args (stmt); i++) | |
435 | { | |
436 | edge e = gimple_phi_arg_edge (stmt, i); | |
437 | if (!bitmap_bit_p (current->split_bbs, e->src->index)) | |
438 | { | |
439 | tree edge_val = gimple_phi_arg_def (stmt, i); | |
440 | if (val && edge_val != val) | |
441 | { | |
442 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
443 | fprintf (dump_file, | |
444 | " Refused: entry BB has PHI with multiple variants\n"); | |
445 | return; | |
446 | } | |
447 | val = edge_val; | |
448 | } | |
449 | } | |
2862cf88 | 450 | } |
451 | ||
452 | ||
453 | /* See what argument we will pass to the split function and compute | |
454 | call overhead. */ | |
455 | call_overhead = eni_size_weights.call_cost; | |
456 | for (parm = DECL_ARGUMENTS (current_function_decl); parm; | |
1767a056 | 457 | parm = DECL_CHAIN (parm)) |
2862cf88 | 458 | { |
459 | if (!is_gimple_reg (parm)) | |
460 | { | |
461 | if (bitmap_bit_p (non_ssa_vars, DECL_UID (parm))) | |
462 | { | |
463 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
464 | fprintf (dump_file, | |
465 | " Refused: need to pass non-ssa param values\n"); | |
466 | return; | |
467 | } | |
468 | } | |
c6dfe037 | 469 | else |
2862cf88 | 470 | { |
c6dfe037 | 471 | tree ddef = ssa_default_def (cfun, parm); |
472 | if (ddef | |
473 | && bitmap_bit_p (current->ssa_names_to_pass, | |
474 | SSA_NAME_VERSION (ddef))) | |
475 | { | |
476 | if (!VOID_TYPE_P (TREE_TYPE (parm))) | |
477 | call_overhead += estimate_move_cost (TREE_TYPE (parm)); | |
478 | num_args++; | |
479 | } | |
2862cf88 | 480 | } |
481 | } | |
482 | if (!VOID_TYPE_P (TREE_TYPE (current_function_decl))) | |
483 | call_overhead += estimate_move_cost (TREE_TYPE (current_function_decl)); | |
484 | ||
485 | if (current->split_size <= call_overhead) | |
486 | { | |
487 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
488 | fprintf (dump_file, | |
489 | " Refused: split size is smaller than call overhead\n"); | |
490 | return; | |
491 | } | |
492 | if (current->header_size + call_overhead | |
493 | >= (unsigned int)(DECL_DECLARED_INLINE_P (current_function_decl) | |
494 | ? MAX_INLINE_INSNS_SINGLE | |
495 | : MAX_INLINE_INSNS_AUTO)) | |
496 | { | |
497 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
498 | fprintf (dump_file, | |
499 | " Refused: header size is too large for inline candidate\n"); | |
500 | return; | |
501 | } | |
502 | ||
503 | /* FIXME: we currently can pass only SSA function parameters to the split | |
fd8d648f | 504 | arguments. Once parm_adjustment infrastructure is supported by cloning, |
2862cf88 | 505 | we can pass more than that. */ |
506 | if (num_args != bitmap_count_bits (current->ssa_names_to_pass)) | |
507 | { | |
6a69e813 | 508 | |
2862cf88 | 509 | if (dump_file && (dump_flags & TDF_DETAILS)) |
510 | fprintf (dump_file, | |
511 | " Refused: need to pass non-param values\n"); | |
512 | return; | |
513 | } | |
514 | ||
515 | /* When there are non-ssa vars used in the split region, see if they | |
516 | are used in the header region. If so, reject the split. | |
517 | FIXME: we can use nested function support to access both. */ | |
4493dab3 | 518 | if (!bitmap_empty_p (non_ssa_vars) |
519 | && !verify_non_ssa_vars (current, non_ssa_vars, return_bb)) | |
2862cf88 | 520 | { |
4493dab3 | 521 | if (dump_file && (dump_flags & TDF_DETAILS)) |
522 | fprintf (dump_file, | |
523 | " Refused: split part has non-ssa uses\n"); | |
2862cf88 | 524 | return; |
525 | } | |
aa2edcb9 | 526 | |
527 | /* If the split point is dominated by a forbidden block, reject | |
528 | the split. */ | |
529 | if (!bitmap_empty_p (forbidden_dominators) | |
530 | && dominated_by_forbidden (current->entry_bb)) | |
531 | { | |
532 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
533 | fprintf (dump_file, | |
534 | " Refused: split point dominated by forbidden block\n"); | |
535 | return; | |
536 | } | |
537 | ||
b04bab7c | 538 | /* See if retval used by return bb is computed by header or split part. |
539 | When it is computed by split part, we need to produce return statement | |
540 | in the split part and add code to header to pass it around. | |
541 | ||
542 | This is bit tricky to test: | |
543 | 1) When there is no return_bb or no return value, we always pass | |
544 | value around. | |
545 | 2) Invariants are always computed by caller. | |
546 | 3) For SSA we need to look if defining statement is in header or split part | |
547 | 4) For non-SSA we need to look where the var is computed. */ | |
548 | retval = find_retval (return_bb); | |
549 | if (!retval) | |
550 | current->split_part_set_retval = true; | |
551 | else if (is_gimple_min_invariant (retval)) | |
552 | current->split_part_set_retval = false; | |
553 | /* Special case is value returned by reference we record as if it was non-ssa | |
554 | set to result_decl. */ | |
555 | else if (TREE_CODE (retval) == SSA_NAME | |
ec11736b | 556 | && SSA_NAME_VAR (retval) |
b04bab7c | 557 | && TREE_CODE (SSA_NAME_VAR (retval)) == RESULT_DECL |
558 | && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) | |
559 | current->split_part_set_retval | |
560 | = bitmap_bit_p (non_ssa_vars, DECL_UID (SSA_NAME_VAR (retval))); | |
561 | else if (TREE_CODE (retval) == SSA_NAME) | |
562 | current->split_part_set_retval | |
563 | = (!SSA_NAME_IS_DEFAULT_DEF (retval) | |
564 | && (bitmap_bit_p (current->split_bbs, | |
565 | gimple_bb (SSA_NAME_DEF_STMT (retval))->index) | |
566 | || gimple_bb (SSA_NAME_DEF_STMT (retval)) == return_bb)); | |
567 | else if (TREE_CODE (retval) == PARM_DECL) | |
568 | current->split_part_set_retval = false; | |
569 | else if (TREE_CODE (retval) == VAR_DECL | |
570 | || TREE_CODE (retval) == RESULT_DECL) | |
571 | current->split_part_set_retval | |
572 | = bitmap_bit_p (non_ssa_vars, DECL_UID (retval)); | |
573 | else | |
574 | current->split_part_set_retval = true; | |
575 | ||
9aa9c1e6 | 576 | /* split_function fixes up at most one PHI non-virtual PHI node in return_bb, |
577 | for the return value. If there are other PHIs, give up. */ | |
578 | if (return_bb != EXIT_BLOCK_PTR) | |
579 | { | |
580 | gimple_stmt_iterator psi; | |
581 | ||
582 | for (psi = gsi_start_phis (return_bb); !gsi_end_p (psi); gsi_next (&psi)) | |
7c782c9b | 583 | if (!virtual_operand_p (gimple_phi_result (gsi_stmt (psi))) |
9aa9c1e6 | 584 | && !(retval |
585 | && current->split_part_set_retval | |
586 | && TREE_CODE (retval) == SSA_NAME | |
587 | && !DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)) | |
588 | && SSA_NAME_DEF_STMT (retval) == gsi_stmt (psi))) | |
589 | { | |
590 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
591 | fprintf (dump_file, | |
592 | " Refused: return bb has extra PHIs\n"); | |
593 | return; | |
594 | } | |
595 | } | |
596 | ||
597 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
598 | fprintf (dump_file, " Accepted!\n"); | |
599 | ||
2862cf88 | 600 | /* At the moment chose split point with lowest frequency and that leaves |
601 | out smallest size of header. | |
602 | In future we might re-consider this heuristics. */ | |
603 | if (!best_split_point.split_bbs | |
604 | || best_split_point.entry_bb->frequency > current->entry_bb->frequency | |
605 | || (best_split_point.entry_bb->frequency == current->entry_bb->frequency | |
606 | && best_split_point.split_size < current->split_size)) | |
607 | ||
608 | { | |
609 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
610 | fprintf (dump_file, " New best split point!\n"); | |
611 | if (best_split_point.ssa_names_to_pass) | |
612 | { | |
613 | BITMAP_FREE (best_split_point.ssa_names_to_pass); | |
614 | BITMAP_FREE (best_split_point.split_bbs); | |
615 | } | |
616 | best_split_point = *current; | |
617 | best_split_point.ssa_names_to_pass = BITMAP_ALLOC (NULL); | |
618 | bitmap_copy (best_split_point.ssa_names_to_pass, | |
619 | current->ssa_names_to_pass); | |
620 | best_split_point.split_bbs = BITMAP_ALLOC (NULL); | |
621 | bitmap_copy (best_split_point.split_bbs, current->split_bbs); | |
622 | } | |
623 | } | |
624 | ||
4493dab3 | 625 | /* Return basic block containing RETURN statement. We allow basic blocks |
626 | of the form: | |
627 | <retval> = tmp_var; | |
628 | return <retval> | |
629 | but return_bb can not be more complex than this. | |
630 | If nothing is found, return EXIT_BLOCK_PTR. | |
631 | ||
2862cf88 | 632 | When there are multiple RETURN statement, chose one with return value, |
633 | since that one is more likely shared by multiple code paths. | |
4493dab3 | 634 | |
635 | Return BB is special, because for function splitting it is the only | |
636 | basic block that is duplicated in between header and split part of the | |
637 | function. | |
638 | ||
2862cf88 | 639 | TODO: We might support multiple return blocks. */ |
640 | ||
641 | static basic_block | |
642 | find_return_bb (void) | |
643 | { | |
644 | edge e; | |
2862cf88 | 645 | basic_block return_bb = EXIT_BLOCK_PTR; |
b234c94d | 646 | gimple_stmt_iterator bsi; |
647 | bool found_return = false; | |
648 | tree retval = NULL_TREE; | |
2862cf88 | 649 | |
b234c94d | 650 | if (!single_pred_p (EXIT_BLOCK_PTR)) |
651 | return return_bb; | |
652 | ||
653 | e = single_pred_edge (EXIT_BLOCK_PTR); | |
654 | for (bsi = gsi_last_bb (e->src); !gsi_end_p (bsi); gsi_prev (&bsi)) | |
655 | { | |
656 | gimple stmt = gsi_stmt (bsi); | |
1a3c6423 | 657 | if (gimple_code (stmt) == GIMPLE_LABEL |
658 | || is_gimple_debug (stmt) | |
659 | || gimple_clobber_p (stmt)) | |
b234c94d | 660 | ; |
661 | else if (gimple_code (stmt) == GIMPLE_ASSIGN | |
662 | && found_return | |
663 | && gimple_assign_single_p (stmt) | |
664 | && (auto_var_in_fn_p (gimple_assign_rhs1 (stmt), | |
665 | current_function_decl) | |
666 | || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))) | |
667 | && retval == gimple_assign_lhs (stmt)) | |
668 | ; | |
669 | else if (gimple_code (stmt) == GIMPLE_RETURN) | |
670 | { | |
671 | found_return = true; | |
672 | retval = gimple_return_retval (stmt); | |
673 | } | |
674 | else | |
675 | break; | |
676 | } | |
677 | if (gsi_end_p (bsi) && found_return) | |
678 | return_bb = e->src; | |
2862cf88 | 679 | |
2862cf88 | 680 | return return_bb; |
681 | } | |
682 | ||
574c4ab5 | 683 | /* Given return basic block RETURN_BB, see where return value is really |
4493dab3 | 684 | stored. */ |
685 | static tree | |
686 | find_retval (basic_block return_bb) | |
687 | { | |
688 | gimple_stmt_iterator bsi; | |
689 | for (bsi = gsi_start_bb (return_bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
690 | if (gimple_code (gsi_stmt (bsi)) == GIMPLE_RETURN) | |
691 | return gimple_return_retval (gsi_stmt (bsi)); | |
1a3c6423 | 692 | else if (gimple_code (gsi_stmt (bsi)) == GIMPLE_ASSIGN |
693 | && !gimple_clobber_p (gsi_stmt (bsi))) | |
4493dab3 | 694 | return gimple_assign_rhs1 (gsi_stmt (bsi)); |
695 | return NULL; | |
696 | } | |
697 | ||
2dc0a33e | 698 | /* Callback for walk_stmt_load_store_addr_ops. If T is non-SSA automatic |
699 | variable, mark it as used in bitmap passed via DATA. | |
2862cf88 | 700 | Return true when access to T prevents splitting the function. */ |
701 | ||
702 | static bool | |
2dc0a33e | 703 | mark_nonssa_use (gimple stmt ATTRIBUTE_UNUSED, tree t, void *data) |
2862cf88 | 704 | { |
705 | t = get_base_address (t); | |
706 | ||
707 | if (!t || is_gimple_reg (t)) | |
708 | return false; | |
709 | ||
710 | /* At present we can't pass non-SSA arguments to split function. | |
711 | FIXME: this can be relaxed by passing references to arguments. */ | |
712 | if (TREE_CODE (t) == PARM_DECL) | |
713 | { | |
714 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2dc0a33e | 715 | fprintf (dump_file, |
716 | "Cannot split: use of non-ssa function parameter.\n"); | |
2862cf88 | 717 | return true; |
718 | } | |
719 | ||
2dc0a33e | 720 | if ((TREE_CODE (t) == VAR_DECL |
721 | && auto_var_in_fn_p (t, current_function_decl)) | |
722 | || TREE_CODE (t) == RESULT_DECL | |
723 | || TREE_CODE (t) == LABEL_DECL) | |
2862cf88 | 724 | bitmap_set_bit ((bitmap)data, DECL_UID (t)); |
b04bab7c | 725 | |
2dc0a33e | 726 | /* For DECL_BY_REFERENCE, the return value is actually a pointer. We want |
727 | to pretend that the value pointed to is actual result decl. */ | |
728 | if ((TREE_CODE (t) == MEM_REF || INDIRECT_REF_P (t)) | |
b04bab7c | 729 | && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME |
ec11736b | 730 | && SSA_NAME_VAR (TREE_OPERAND (t, 0)) |
b04bab7c | 731 | && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND (t, 0))) == RESULT_DECL |
732 | && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) | |
2dc0a33e | 733 | return |
734 | bitmap_bit_p ((bitmap)data, | |
735 | DECL_UID (DECL_RESULT (current_function_decl))); | |
736 | ||
2862cf88 | 737 | return false; |
738 | } | |
739 | ||
740 | /* Compute local properties of basic block BB we collect when looking for | |
741 | split points. We look for ssa defs and store them in SET_SSA_NAMES, | |
742 | for ssa uses and store them in USED_SSA_NAMES and for any non-SSA automatic | |
743 | vars stored in NON_SSA_VARS. | |
744 | ||
745 | When BB has edge to RETURN_BB, collect uses in RETURN_BB too. | |
746 | ||
747 | Return false when BB contains something that prevents it from being put into | |
748 | split function. */ | |
749 | ||
750 | static bool | |
751 | visit_bb (basic_block bb, basic_block return_bb, | |
752 | bitmap set_ssa_names, bitmap used_ssa_names, | |
753 | bitmap non_ssa_vars) | |
754 | { | |
755 | gimple_stmt_iterator bsi; | |
756 | edge e; | |
757 | edge_iterator ei; | |
758 | bool can_split = true; | |
759 | ||
760 | for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
761 | { | |
762 | gimple stmt = gsi_stmt (bsi); | |
763 | tree op; | |
764 | ssa_op_iter iter; | |
765 | tree decl; | |
766 | ||
767 | if (is_gimple_debug (stmt)) | |
768 | continue; | |
769 | ||
1a3c6423 | 770 | if (gimple_clobber_p (stmt)) |
771 | continue; | |
772 | ||
2862cf88 | 773 | /* FIXME: We can split regions containing EH. We can not however |
774 | split RESX, EH_DISPATCH and EH_POINTER referring to same region | |
775 | into different partitions. This would require tracking of | |
776 | EH regions and checking in consider_split_point if they | |
777 | are not used elsewhere. */ | |
892c141e | 778 | if (gimple_code (stmt) == GIMPLE_RESX) |
2862cf88 | 779 | { |
780 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
892c141e | 781 | fprintf (dump_file, "Cannot split: resx.\n"); |
2862cf88 | 782 | can_split = false; |
783 | } | |
784 | if (gimple_code (stmt) == GIMPLE_EH_DISPATCH) | |
785 | { | |
786 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2dc0a33e | 787 | fprintf (dump_file, "Cannot split: eh dispatch.\n"); |
2862cf88 | 788 | can_split = false; |
789 | } | |
790 | ||
791 | /* Check builtins that prevent splitting. */ | |
792 | if (gimple_code (stmt) == GIMPLE_CALL | |
793 | && (decl = gimple_call_fndecl (stmt)) != NULL_TREE | |
794 | && DECL_BUILT_IN (decl) | |
795 | && DECL_BUILT_IN_CLASS (decl) == BUILT_IN_NORMAL) | |
796 | switch (DECL_FUNCTION_CODE (decl)) | |
797 | { | |
798 | /* FIXME: once we will allow passing non-parm values to split part, | |
799 | we need to be sure to handle correct builtin_stack_save and | |
800 | builtin_stack_restore. At the moment we are safe; there is no | |
801 | way to store builtin_stack_save result in non-SSA variable | |
802 | since all calls to those are compiler generated. */ | |
803 | case BUILT_IN_APPLY: | |
3c97c75d | 804 | case BUILT_IN_APPLY_ARGS: |
2862cf88 | 805 | case BUILT_IN_VA_START: |
806 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2dc0a33e | 807 | fprintf (dump_file, |
808 | "Cannot split: builtin_apply and va_start.\n"); | |
2862cf88 | 809 | can_split = false; |
810 | break; | |
811 | case BUILT_IN_EH_POINTER: | |
812 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2dc0a33e | 813 | fprintf (dump_file, "Cannot split: builtin_eh_pointer.\n"); |
2862cf88 | 814 | can_split = false; |
815 | break; | |
816 | default: | |
817 | break; | |
818 | } | |
819 | ||
820 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF) | |
821 | bitmap_set_bit (set_ssa_names, SSA_NAME_VERSION (op)); | |
822 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE) | |
823 | bitmap_set_bit (used_ssa_names, SSA_NAME_VERSION (op)); | |
824 | can_split &= !walk_stmt_load_store_addr_ops (stmt, non_ssa_vars, | |
825 | mark_nonssa_use, | |
826 | mark_nonssa_use, | |
827 | mark_nonssa_use); | |
828 | } | |
829 | for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
830 | { | |
831 | gimple stmt = gsi_stmt (bsi); | |
6a69e813 | 832 | unsigned int i; |
2862cf88 | 833 | |
7c782c9b | 834 | if (virtual_operand_p (gimple_phi_result (stmt))) |
2862cf88 | 835 | continue; |
6a69e813 | 836 | bitmap_set_bit (set_ssa_names, |
837 | SSA_NAME_VERSION (gimple_phi_result (stmt))); | |
838 | for (i = 0; i < gimple_phi_num_args (stmt); i++) | |
839 | { | |
840 | tree op = gimple_phi_arg_def (stmt, i); | |
841 | if (TREE_CODE (op) == SSA_NAME) | |
842 | bitmap_set_bit (used_ssa_names, SSA_NAME_VERSION (op)); | |
843 | } | |
2862cf88 | 844 | can_split &= !walk_stmt_load_store_addr_ops (stmt, non_ssa_vars, |
845 | mark_nonssa_use, | |
846 | mark_nonssa_use, | |
847 | mark_nonssa_use); | |
848 | } | |
574c4ab5 | 849 | /* Record also uses coming from PHI operand in return BB. */ |
2862cf88 | 850 | FOR_EACH_EDGE (e, ei, bb->succs) |
851 | if (e->dest == return_bb) | |
852 | { | |
2862cf88 | 853 | for (bsi = gsi_start_phis (return_bb); !gsi_end_p (bsi); gsi_next (&bsi)) |
854 | { | |
855 | gimple stmt = gsi_stmt (bsi); | |
856 | tree op = gimple_phi_arg_def (stmt, e->dest_idx); | |
857 | ||
7c782c9b | 858 | if (virtual_operand_p (gimple_phi_result (stmt))) |
2862cf88 | 859 | continue; |
2862cf88 | 860 | if (TREE_CODE (op) == SSA_NAME) |
861 | bitmap_set_bit (used_ssa_names, SSA_NAME_VERSION (op)); | |
862 | else | |
863 | can_split &= !mark_nonssa_use (stmt, op, non_ssa_vars); | |
864 | } | |
2862cf88 | 865 | } |
866 | return can_split; | |
867 | } | |
868 | ||
869 | /* Stack entry for recursive DFS walk in find_split_point. */ | |
870 | ||
871 | typedef struct | |
872 | { | |
873 | /* Basic block we are examining. */ | |
874 | basic_block bb; | |
875 | ||
876 | /* SSA names set and used by the BB and all BBs reachable | |
877 | from it via DFS walk. */ | |
878 | bitmap set_ssa_names, used_ssa_names; | |
879 | bitmap non_ssa_vars; | |
880 | ||
881 | /* All BBS visited from this BB via DFS walk. */ | |
882 | bitmap bbs_visited; | |
883 | ||
884 | /* Last examined edge in DFS walk. Since we walk unoriented graph, | |
574c4ab5 | 885 | the value is up to sum of incoming and outgoing edges of BB. */ |
2862cf88 | 886 | unsigned int edge_num; |
887 | ||
888 | /* Stack entry index of earliest BB reachable from current BB | |
574c4ab5 | 889 | or any BB visited later in DFS walk. */ |
2862cf88 | 890 | int earliest; |
891 | ||
892 | /* Overall time and size of all BBs reached from this BB in DFS walk. */ | |
893 | int overall_time, overall_size; | |
894 | ||
895 | /* When false we can not split on this BB. */ | |
896 | bool can_split; | |
897 | } stack_entry; | |
2862cf88 | 898 | |
899 | ||
900 | /* Find all articulations and call consider_split on them. | |
901 | OVERALL_TIME and OVERALL_SIZE is time and size of the function. | |
902 | ||
903 | We perform basic algorithm for finding an articulation in a graph | |
904 | created from CFG by considering it to be an unoriented graph. | |
905 | ||
906 | The articulation is discovered via DFS walk. We collect earliest | |
907 | basic block on stack that is reachable via backward edge. Articulation | |
908 | is any basic block such that there is no backward edge bypassing it. | |
909 | To reduce stack usage we maintain heap allocated stack in STACK vector. | |
910 | AUX pointer of BB is set to index it appears in the stack or -1 once | |
911 | it is visited and popped off the stack. | |
912 | ||
913 | The algorithm finds articulation after visiting the whole component | |
914 | reachable by it. This makes it convenient to collect information about | |
915 | the component used by consider_split. */ | |
916 | ||
917 | static void | |
918 | find_split_points (int overall_time, int overall_size) | |
919 | { | |
920 | stack_entry first; | |
1e094109 | 921 | vec<stack_entry> stack = vNULL; |
2862cf88 | 922 | basic_block bb; |
923 | basic_block return_bb = find_return_bb (); | |
924 | struct split_point current; | |
925 | ||
926 | current.header_time = overall_time; | |
927 | current.header_size = overall_size; | |
928 | current.split_time = 0; | |
929 | current.split_size = 0; | |
930 | current.ssa_names_to_pass = BITMAP_ALLOC (NULL); | |
931 | ||
932 | first.bb = ENTRY_BLOCK_PTR; | |
933 | first.edge_num = 0; | |
934 | first.overall_time = 0; | |
935 | first.overall_size = 0; | |
936 | first.earliest = INT_MAX; | |
937 | first.set_ssa_names = 0; | |
938 | first.used_ssa_names = 0; | |
939 | first.bbs_visited = 0; | |
f1f41a6c | 940 | stack.safe_push (first); |
2862cf88 | 941 | ENTRY_BLOCK_PTR->aux = (void *)(intptr_t)-1; |
942 | ||
f1f41a6c | 943 | while (!stack.is_empty ()) |
2862cf88 | 944 | { |
f1f41a6c | 945 | stack_entry *entry = &stack.last (); |
2862cf88 | 946 | |
947 | /* We are walking an acyclic graph, so edge_num counts | |
948 | succ and pred edges together. However when considering | |
949 | articulation, we want to have processed everything reachable | |
950 | from articulation but nothing that reaches into it. */ | |
951 | if (entry->edge_num == EDGE_COUNT (entry->bb->succs) | |
952 | && entry->bb != ENTRY_BLOCK_PTR) | |
953 | { | |
f1f41a6c | 954 | int pos = stack.length (); |
2862cf88 | 955 | entry->can_split &= visit_bb (entry->bb, return_bb, |
956 | entry->set_ssa_names, | |
957 | entry->used_ssa_names, | |
958 | entry->non_ssa_vars); | |
959 | if (pos <= entry->earliest && !entry->can_split | |
960 | && dump_file && (dump_flags & TDF_DETAILS)) | |
961 | fprintf (dump_file, | |
962 | "found articulation at bb %i but can not split\n", | |
963 | entry->bb->index); | |
964 | if (pos <= entry->earliest && entry->can_split) | |
965 | { | |
966 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
967 | fprintf (dump_file, "found articulation at bb %i\n", | |
968 | entry->bb->index); | |
969 | current.entry_bb = entry->bb; | |
970 | current.ssa_names_to_pass = BITMAP_ALLOC (NULL); | |
971 | bitmap_and_compl (current.ssa_names_to_pass, | |
972 | entry->used_ssa_names, entry->set_ssa_names); | |
973 | current.header_time = overall_time - entry->overall_time; | |
974 | current.header_size = overall_size - entry->overall_size; | |
975 | current.split_time = entry->overall_time; | |
976 | current.split_size = entry->overall_size; | |
977 | current.split_bbs = entry->bbs_visited; | |
978 | consider_split (¤t, entry->non_ssa_vars, return_bb); | |
979 | BITMAP_FREE (current.ssa_names_to_pass); | |
980 | } | |
981 | } | |
982 | /* Do actual DFS walk. */ | |
983 | if (entry->edge_num | |
984 | < (EDGE_COUNT (entry->bb->succs) | |
985 | + EDGE_COUNT (entry->bb->preds))) | |
986 | { | |
987 | edge e; | |
988 | basic_block dest; | |
989 | if (entry->edge_num < EDGE_COUNT (entry->bb->succs)) | |
990 | { | |
991 | e = EDGE_SUCC (entry->bb, entry->edge_num); | |
992 | dest = e->dest; | |
993 | } | |
994 | else | |
995 | { | |
996 | e = EDGE_PRED (entry->bb, entry->edge_num | |
997 | - EDGE_COUNT (entry->bb->succs)); | |
998 | dest = e->src; | |
999 | } | |
1000 | ||
1001 | entry->edge_num++; | |
1002 | ||
1003 | /* New BB to visit, push it to the stack. */ | |
1004 | if (dest != return_bb && dest != EXIT_BLOCK_PTR | |
1005 | && !dest->aux) | |
1006 | { | |
1007 | stack_entry new_entry; | |
1008 | ||
1009 | new_entry.bb = dest; | |
1010 | new_entry.edge_num = 0; | |
1011 | new_entry.overall_time | |
f1f41a6c | 1012 | = bb_info_vec[dest->index].time; |
2862cf88 | 1013 | new_entry.overall_size |
f1f41a6c | 1014 | = bb_info_vec[dest->index].size; |
2862cf88 | 1015 | new_entry.earliest = INT_MAX; |
1016 | new_entry.set_ssa_names = BITMAP_ALLOC (NULL); | |
1017 | new_entry.used_ssa_names = BITMAP_ALLOC (NULL); | |
1018 | new_entry.bbs_visited = BITMAP_ALLOC (NULL); | |
1019 | new_entry.non_ssa_vars = BITMAP_ALLOC (NULL); | |
1020 | new_entry.can_split = true; | |
1021 | bitmap_set_bit (new_entry.bbs_visited, dest->index); | |
f1f41a6c | 1022 | stack.safe_push (new_entry); |
1023 | dest->aux = (void *)(intptr_t)stack.length (); | |
2862cf88 | 1024 | } |
1025 | /* Back edge found, record the earliest point. */ | |
1026 | else if ((intptr_t)dest->aux > 0 | |
1027 | && (intptr_t)dest->aux < entry->earliest) | |
1028 | entry->earliest = (intptr_t)dest->aux; | |
1029 | } | |
574c4ab5 | 1030 | /* We are done with examining the edges. Pop off the value from stack |
1031 | and merge stuff we accumulate during the walk. */ | |
2862cf88 | 1032 | else if (entry->bb != ENTRY_BLOCK_PTR) |
1033 | { | |
f1f41a6c | 1034 | stack_entry *prev = &stack[stack.length () - 2]; |
2862cf88 | 1035 | |
1036 | entry->bb->aux = (void *)(intptr_t)-1; | |
1037 | prev->can_split &= entry->can_split; | |
1038 | if (prev->set_ssa_names) | |
1039 | { | |
1040 | bitmap_ior_into (prev->set_ssa_names, entry->set_ssa_names); | |
1041 | bitmap_ior_into (prev->used_ssa_names, entry->used_ssa_names); | |
1042 | bitmap_ior_into (prev->bbs_visited, entry->bbs_visited); | |
1043 | bitmap_ior_into (prev->non_ssa_vars, entry->non_ssa_vars); | |
1044 | } | |
1045 | if (prev->earliest > entry->earliest) | |
1046 | prev->earliest = entry->earliest; | |
1047 | prev->overall_time += entry->overall_time; | |
1048 | prev->overall_size += entry->overall_size; | |
1049 | BITMAP_FREE (entry->set_ssa_names); | |
1050 | BITMAP_FREE (entry->used_ssa_names); | |
1051 | BITMAP_FREE (entry->bbs_visited); | |
1052 | BITMAP_FREE (entry->non_ssa_vars); | |
f1f41a6c | 1053 | stack.pop (); |
2862cf88 | 1054 | } |
1055 | else | |
f1f41a6c | 1056 | stack.pop (); |
2862cf88 | 1057 | } |
1058 | ENTRY_BLOCK_PTR->aux = NULL; | |
1059 | FOR_EACH_BB (bb) | |
1060 | bb->aux = NULL; | |
f1f41a6c | 1061 | stack.release (); |
2862cf88 | 1062 | BITMAP_FREE (current.ssa_names_to_pass); |
1063 | } | |
1064 | ||
1065 | /* Split function at SPLIT_POINT. */ | |
1066 | ||
1067 | static void | |
1068 | split_function (struct split_point *split_point) | |
1069 | { | |
1e094109 | 1070 | vec<tree> args_to_pass = vNULL; |
982ffd8d | 1071 | bitmap args_to_skip; |
2862cf88 | 1072 | tree parm; |
1073 | int num = 0; | |
982ffd8d | 1074 | struct cgraph_node *node, *cur_node = cgraph_get_node (current_function_decl); |
2862cf88 | 1075 | basic_block return_bb = find_return_bb (); |
1076 | basic_block call_bb; | |
1077 | gimple_stmt_iterator gsi; | |
1078 | gimple call; | |
1079 | edge e; | |
1080 | edge_iterator ei; | |
1081 | tree retval = NULL, real_retval = NULL; | |
1082 | bool split_part_return_p = false; | |
1083 | gimple last_stmt = NULL; | |
e65564ab | 1084 | unsigned int i; |
c6dfe037 | 1085 | tree arg, ddef; |
f1f41a6c | 1086 | vec<tree, va_gc> **debug_args = NULL; |
2862cf88 | 1087 | |
1088 | if (dump_file) | |
1089 | { | |
1090 | fprintf (dump_file, "\n\nSplitting function at:\n"); | |
1091 | dump_split_point (dump_file, split_point); | |
1092 | } | |
1093 | ||
982ffd8d | 1094 | if (cur_node->local.can_change_signature) |
1095 | args_to_skip = BITMAP_ALLOC (NULL); | |
1096 | else | |
1097 | args_to_skip = NULL; | |
1098 | ||
2862cf88 | 1099 | /* Collect the parameters of new function and args_to_skip bitmap. */ |
1100 | for (parm = DECL_ARGUMENTS (current_function_decl); | |
1767a056 | 1101 | parm; parm = DECL_CHAIN (parm), num++) |
982ffd8d | 1102 | if (args_to_skip |
1103 | && (!is_gimple_reg (parm) | |
c6dfe037 | 1104 | || (ddef = ssa_default_def (cfun, parm)) == NULL_TREE |
982ffd8d | 1105 | || !bitmap_bit_p (split_point->ssa_names_to_pass, |
c6dfe037 | 1106 | SSA_NAME_VERSION (ddef)))) |
2862cf88 | 1107 | bitmap_set_bit (args_to_skip, num); |
1108 | else | |
e65564ab | 1109 | { |
17c5f86d | 1110 | /* This parm might not have been used up to now, but is going to be |
1111 | used, hence register it. */ | |
17c5f86d | 1112 | if (is_gimple_reg (parm)) |
c6dfe037 | 1113 | arg = get_or_create_ssa_default_def (cfun, parm); |
17c5f86d | 1114 | else |
1115 | arg = parm; | |
982ffd8d | 1116 | |
0ed25048 | 1117 | if (!useless_type_conversion_p (DECL_ARG_TYPE (parm), TREE_TYPE (arg))) |
1118 | arg = fold_convert (DECL_ARG_TYPE (parm), arg); | |
f1f41a6c | 1119 | args_to_pass.safe_push (arg); |
e65564ab | 1120 | } |
2862cf88 | 1121 | |
1122 | /* See if the split function will return. */ | |
1123 | FOR_EACH_EDGE (e, ei, return_bb->preds) | |
1124 | if (bitmap_bit_p (split_point->split_bbs, e->src->index)) | |
1125 | break; | |
1126 | if (e) | |
1127 | split_part_return_p = true; | |
1128 | ||
b04bab7c | 1129 | /* Add return block to what will become the split function. |
1130 | We do not return; no return block is needed. */ | |
1131 | if (!split_part_return_p) | |
1132 | ; | |
1133 | /* We have no return block, so nothing is needed. */ | |
1134 | else if (return_bb == EXIT_BLOCK_PTR) | |
1135 | ; | |
1136 | /* When we do not want to return value, we need to construct | |
1137 | new return block with empty return statement. | |
1138 | FIXME: Once we are able to change return type, we should change function | |
1139 | to return void instead of just outputting function with undefined return | |
1140 | value. For structures this affects quality of codegen. */ | |
1141 | else if (!split_point->split_part_set_retval | |
1142 | && find_retval (return_bb)) | |
1143 | { | |
1144 | bool redirected = true; | |
1145 | basic_block new_return_bb = create_basic_block (NULL, 0, return_bb); | |
1146 | gimple_stmt_iterator gsi = gsi_start_bb (new_return_bb); | |
1147 | gsi_insert_after (&gsi, gimple_build_return (NULL), GSI_NEW_STMT); | |
1148 | while (redirected) | |
1149 | { | |
1150 | redirected = false; | |
1151 | FOR_EACH_EDGE (e, ei, return_bb->preds) | |
1152 | if (bitmap_bit_p (split_point->split_bbs, e->src->index)) | |
1153 | { | |
1154 | new_return_bb->count += e->count; | |
1155 | new_return_bb->frequency += EDGE_FREQUENCY (e); | |
1156 | redirect_edge_and_branch (e, new_return_bb); | |
1157 | redirected = true; | |
1158 | break; | |
1159 | } | |
1160 | } | |
1161 | e = make_edge (new_return_bb, EXIT_BLOCK_PTR, 0); | |
1162 | e->probability = REG_BR_PROB_BASE; | |
1163 | e->count = new_return_bb->count; | |
f6568ea4 | 1164 | if (current_loops) |
1165 | add_bb_to_loop (new_return_bb, current_loops->tree_root); | |
b04bab7c | 1166 | bitmap_set_bit (split_point->split_bbs, new_return_bb->index); |
c13d6d8f | 1167 | } |
1168 | /* When we pass around the value, use existing return block. */ | |
1169 | else | |
1170 | bitmap_set_bit (split_point->split_bbs, return_bb->index); | |
1171 | ||
1172 | /* If RETURN_BB has virtual operand PHIs, they must be removed and the | |
1173 | virtual operand marked for renaming as we change the CFG in a way that | |
1d0b727d | 1174 | tree-inline is not able to compensate for. |
c13d6d8f | 1175 | |
1176 | Note this can happen whether or not we have a return value. If we have | |
1177 | a return value, then RETURN_BB may have PHIs for real operands too. */ | |
1178 | if (return_bb != EXIT_BLOCK_PTR) | |
1179 | { | |
1d0b727d | 1180 | bool phi_p = false; |
b04bab7c | 1181 | for (gsi = gsi_start_phis (return_bb); !gsi_end_p (gsi);) |
1182 | { | |
1183 | gimple stmt = gsi_stmt (gsi); | |
7c782c9b | 1184 | if (!virtual_operand_p (gimple_phi_result (stmt))) |
c13d6d8f | 1185 | { |
1186 | gsi_next (&gsi); | |
1187 | continue; | |
1188 | } | |
904ad591 | 1189 | mark_virtual_phi_result_for_renaming (stmt); |
1190 | remove_phi_node (&gsi, true); | |
1d0b727d | 1191 | phi_p = true; |
b04bab7c | 1192 | } |
1d0b727d | 1193 | /* In reality we have to rename the reaching definition of the |
1194 | virtual operand at return_bb as we will eventually release it | |
1195 | when we remove the code region we outlined. | |
1196 | So we have to rename all immediate virtual uses of that region | |
1197 | if we didn't see a PHI definition yet. */ | |
1198 | /* ??? In real reality we want to set the reaching vdef of the | |
1199 | entry of the SESE region as the vuse of the call and the reaching | |
1200 | vdef of the exit of the SESE region as the vdef of the call. */ | |
1201 | if (!phi_p) | |
1202 | for (gsi = gsi_start_bb (return_bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
1203 | { | |
1204 | gimple stmt = gsi_stmt (gsi); | |
1205 | if (gimple_vuse (stmt)) | |
1206 | { | |
1207 | gimple_set_vuse (stmt, NULL_TREE); | |
1208 | update_stmt (stmt); | |
1209 | } | |
1210 | if (gimple_vdef (stmt)) | |
1211 | break; | |
1212 | } | |
b04bab7c | 1213 | } |
2862cf88 | 1214 | |
1215 | /* Now create the actual clone. */ | |
1216 | rebuild_cgraph_edges (); | |
1e094109 | 1217 | node = cgraph_function_versioning (cur_node, vNULL, |
f1f41a6c | 1218 | NULL, |
1219 | args_to_skip, | |
7a3ec978 | 1220 | !split_part_return_p, |
2862cf88 | 1221 | split_point->split_bbs, |
4493dab3 | 1222 | split_point->entry_bb, "part"); |
fd8d648f | 1223 | /* For usual cloning it is enough to clear builtin only when signature |
1224 | changes. For partial inlining we however can not expect the part | |
1225 | of builtin implementation to have same semantic as the whole. */ | |
7d0d0ce1 | 1226 | if (DECL_BUILT_IN (node->symbol.decl)) |
fd8d648f | 1227 | { |
7d0d0ce1 | 1228 | DECL_BUILT_IN_CLASS (node->symbol.decl) = NOT_BUILT_IN; |
1229 | DECL_FUNCTION_CODE (node->symbol.decl) = (enum built_in_function) 0; | |
fd8d648f | 1230 | } |
ea03e543 | 1231 | /* If the original function is declared inline, there is no point in issuing |
1232 | a warning for the non-inlinable part. */ | |
1233 | DECL_NO_INLINE_WARNING_P (node->symbol.decl) = 1; | |
982ffd8d | 1234 | cgraph_node_remove_callees (cur_node); |
9df17e79 | 1235 | ipa_remove_all_references (&cur_node->symbol.ref_list); |
2862cf88 | 1236 | if (!split_part_return_p) |
7d0d0ce1 | 1237 | TREE_THIS_VOLATILE (node->symbol.decl) = 1; |
2862cf88 | 1238 | if (dump_file) |
7d0d0ce1 | 1239 | dump_function_to_file (node->symbol.decl, dump_file, dump_flags); |
2862cf88 | 1240 | |
1241 | /* Create the basic block we place call into. It is the entry basic block | |
1242 | split after last label. */ | |
1243 | call_bb = split_point->entry_bb; | |
1244 | for (gsi = gsi_start_bb (call_bb); !gsi_end_p (gsi);) | |
1245 | if (gimple_code (gsi_stmt (gsi)) == GIMPLE_LABEL) | |
1246 | { | |
1247 | last_stmt = gsi_stmt (gsi); | |
1248 | gsi_next (&gsi); | |
1249 | } | |
1250 | else | |
1251 | break; | |
1252 | e = split_block (split_point->entry_bb, last_stmt); | |
1253 | remove_edge (e); | |
1254 | ||
1255 | /* Produce the call statement. */ | |
1256 | gsi = gsi_last_bb (call_bb); | |
f1f41a6c | 1257 | FOR_EACH_VEC_ELT (args_to_pass, i, arg) |
0ed25048 | 1258 | if (!is_gimple_val (arg)) |
1259 | { | |
1260 | arg = force_gimple_operand_gsi (&gsi, arg, true, NULL_TREE, | |
ad676199 | 1261 | false, GSI_CONTINUE_LINKING); |
f1f41a6c | 1262 | args_to_pass[i] = arg; |
0ed25048 | 1263 | } |
7d0d0ce1 | 1264 | call = gimple_build_call_vec (node->symbol.decl, args_to_pass); |
2862cf88 | 1265 | gimple_set_block (call, DECL_INITIAL (current_function_decl)); |
f1f41a6c | 1266 | args_to_pass.release (); |
2862cf88 | 1267 | |
eee0ea7c | 1268 | /* For optimized away parameters, add on the caller side |
1269 | before the call | |
1270 | DEBUG D#X => parm_Y(D) | |
1271 | stmts and associate D#X with parm in decl_debug_args_lookup | |
1272 | vector to say for debug info that if parameter parm had been passed, | |
1273 | it would have value parm_Y(D). */ | |
1274 | if (args_to_skip) | |
1275 | for (parm = DECL_ARGUMENTS (current_function_decl), num = 0; | |
1276 | parm; parm = DECL_CHAIN (parm), num++) | |
1277 | if (bitmap_bit_p (args_to_skip, num) | |
1278 | && is_gimple_reg (parm)) | |
1279 | { | |
1280 | tree ddecl; | |
1281 | gimple def_temp; | |
1282 | ||
1283 | /* This needs to be done even without MAY_HAVE_DEBUG_STMTS, | |
1284 | otherwise if it didn't exist before, we'd end up with | |
1285 | different SSA_NAME_VERSIONs between -g and -g0. */ | |
1286 | arg = get_or_create_ssa_default_def (cfun, parm); | |
1287 | if (!MAY_HAVE_DEBUG_STMTS) | |
1288 | continue; | |
1289 | ||
1290 | if (debug_args == NULL) | |
1291 | debug_args = decl_debug_args_insert (node->symbol.decl); | |
1292 | ddecl = make_node (DEBUG_EXPR_DECL); | |
1293 | DECL_ARTIFICIAL (ddecl) = 1; | |
1294 | TREE_TYPE (ddecl) = TREE_TYPE (parm); | |
1295 | DECL_MODE (ddecl) = DECL_MODE (parm); | |
f1f41a6c | 1296 | vec_safe_push (*debug_args, DECL_ORIGIN (parm)); |
1297 | vec_safe_push (*debug_args, ddecl); | |
eee0ea7c | 1298 | def_temp = gimple_build_debug_bind (ddecl, unshare_expr (arg), |
1299 | call); | |
1300 | gsi_insert_after (&gsi, def_temp, GSI_NEW_STMT); | |
1301 | } | |
1302 | /* And on the callee side, add | |
1303 | DEBUG D#Y s=> parm | |
1304 | DEBUG var => D#Y | |
1305 | stmts to the first bb where var is a VAR_DECL created for the | |
1306 | optimized away parameter in DECL_INITIAL block. This hints | |
1307 | in the debug info that var (whole DECL_ORIGIN is the parm PARM_DECL) | |
1308 | is optimized away, but could be looked up at the call site | |
1309 | as value of D#X there. */ | |
1310 | if (debug_args != NULL) | |
1311 | { | |
1312 | unsigned int i; | |
1313 | tree var, vexpr; | |
1314 | gimple_stmt_iterator cgsi; | |
1315 | gimple def_temp; | |
1316 | ||
1317 | push_cfun (DECL_STRUCT_FUNCTION (node->symbol.decl)); | |
1318 | var = BLOCK_VARS (DECL_INITIAL (node->symbol.decl)); | |
f1f41a6c | 1319 | i = vec_safe_length (*debug_args); |
eee0ea7c | 1320 | cgsi = gsi_after_labels (single_succ (ENTRY_BLOCK_PTR)); |
1321 | do | |
1322 | { | |
1323 | i -= 2; | |
1324 | while (var != NULL_TREE | |
f1f41a6c | 1325 | && DECL_ABSTRACT_ORIGIN (var) != (**debug_args)[i]) |
eee0ea7c | 1326 | var = TREE_CHAIN (var); |
1327 | if (var == NULL_TREE) | |
1328 | break; | |
1329 | vexpr = make_node (DEBUG_EXPR_DECL); | |
f1f41a6c | 1330 | parm = (**debug_args)[i]; |
eee0ea7c | 1331 | DECL_ARTIFICIAL (vexpr) = 1; |
1332 | TREE_TYPE (vexpr) = TREE_TYPE (parm); | |
1333 | DECL_MODE (vexpr) = DECL_MODE (parm); | |
1334 | def_temp = gimple_build_debug_source_bind (vexpr, parm, | |
1335 | NULL); | |
1336 | gsi_insert_before (&cgsi, def_temp, GSI_SAME_STMT); | |
1337 | def_temp = gimple_build_debug_bind (var, vexpr, NULL); | |
1338 | gsi_insert_before (&cgsi, def_temp, GSI_SAME_STMT); | |
1339 | } | |
1340 | while (i); | |
1341 | pop_cfun (); | |
1342 | } | |
1343 | ||
a8005893 | 1344 | /* We avoid address being taken on any variable used by split part, |
1345 | so return slot optimization is always possible. Moreover this is | |
1346 | required to make DECL_BY_REFERENCE work. */ | |
1347 | if (aggregate_value_p (DECL_RESULT (current_function_decl), | |
8ad1eb69 | 1348 | TREE_TYPE (current_function_decl)) |
1349 | && (!is_gimple_reg_type (TREE_TYPE (DECL_RESULT (current_function_decl))) | |
1350 | || DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)))) | |
a8005893 | 1351 | gimple_call_set_return_slot_opt (call, true); |
1352 | ||
2862cf88 | 1353 | /* Update return value. This is bit tricky. When we do not return, |
1354 | do nothing. When we return we might need to update return_bb | |
1355 | or produce a new return statement. */ | |
1356 | if (!split_part_return_p) | |
1357 | gsi_insert_after (&gsi, call, GSI_NEW_STMT); | |
1358 | else | |
1359 | { | |
1360 | e = make_edge (call_bb, return_bb, | |
1361 | return_bb == EXIT_BLOCK_PTR ? 0 : EDGE_FALLTHRU); | |
1362 | e->count = call_bb->count; | |
1363 | e->probability = REG_BR_PROB_BASE; | |
524a0531 | 1364 | |
1365 | /* If there is return basic block, see what value we need to store | |
1366 | return value into and put call just before it. */ | |
2862cf88 | 1367 | if (return_bb != EXIT_BLOCK_PTR) |
1368 | { | |
4493dab3 | 1369 | real_retval = retval = find_retval (return_bb); |
524a0531 | 1370 | |
b04bab7c | 1371 | if (real_retval && split_point->split_part_set_retval) |
2862cf88 | 1372 | { |
1373 | gimple_stmt_iterator psi; | |
1374 | ||
524a0531 | 1375 | /* See if we need new SSA_NAME for the result. |
1376 | When DECL_BY_REFERENCE is true, retval is actually pointer to | |
1377 | return value and it is constant in whole function. */ | |
1378 | if (TREE_CODE (retval) == SSA_NAME | |
1379 | && !DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) | |
2862cf88 | 1380 | { |
874117c8 | 1381 | retval = copy_ssa_name (retval, call); |
524a0531 | 1382 | |
1383 | /* See if there is PHI defining return value. */ | |
1384 | for (psi = gsi_start_phis (return_bb); | |
1385 | !gsi_end_p (psi); gsi_next (&psi)) | |
7c782c9b | 1386 | if (!virtual_operand_p (gimple_phi_result (gsi_stmt (psi)))) |
524a0531 | 1387 | break; |
1388 | ||
1389 | /* When there is PHI, just update its value. */ | |
2862cf88 | 1390 | if (TREE_CODE (retval) == SSA_NAME |
1391 | && !gsi_end_p (psi)) | |
60d535d2 | 1392 | add_phi_arg (gsi_stmt (psi), retval, e, UNKNOWN_LOCATION); |
524a0531 | 1393 | /* Otherwise update the return BB itself. |
1394 | find_return_bb allows at most one assignment to return value, | |
1395 | so update first statement. */ | |
1396 | else | |
2862cf88 | 1397 | { |
4493dab3 | 1398 | gimple_stmt_iterator bsi; |
1399 | for (bsi = gsi_start_bb (return_bb); !gsi_end_p (bsi); | |
1400 | gsi_next (&bsi)) | |
1401 | if (gimple_code (gsi_stmt (bsi)) == GIMPLE_RETURN) | |
1402 | { | |
1403 | gimple_return_set_retval (gsi_stmt (bsi), retval); | |
1404 | break; | |
1405 | } | |
0f41c0cf | 1406 | else if (gimple_code (gsi_stmt (bsi)) == GIMPLE_ASSIGN |
1407 | && !gimple_clobber_p (gsi_stmt (bsi))) | |
4493dab3 | 1408 | { |
1409 | gimple_assign_set_rhs1 (gsi_stmt (bsi), retval); | |
1410 | break; | |
1411 | } | |
1412 | update_stmt (gsi_stmt (bsi)); | |
2862cf88 | 1413 | } |
1414 | } | |
a8005893 | 1415 | if (DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) |
9ece8f57 | 1416 | { |
1417 | gimple_call_set_lhs (call, build_simple_mem_ref (retval)); | |
1418 | gsi_insert_after (&gsi, call, GSI_NEW_STMT); | |
1419 | } | |
a8005893 | 1420 | else |
9ece8f57 | 1421 | { |
1422 | tree restype; | |
1423 | restype = TREE_TYPE (DECL_RESULT (current_function_decl)); | |
1424 | gsi_insert_after (&gsi, call, GSI_NEW_STMT); | |
1425 | if (!useless_type_conversion_p (TREE_TYPE (retval), restype)) | |
1426 | { | |
1427 | gimple cpy; | |
1428 | tree tem = create_tmp_reg (restype, NULL); | |
1429 | tem = make_ssa_name (tem, call); | |
1430 | cpy = gimple_build_assign_with_ops (NOP_EXPR, retval, | |
1431 | tem, NULL_TREE); | |
1432 | gsi_insert_after (&gsi, cpy, GSI_NEW_STMT); | |
1433 | retval = tem; | |
1434 | } | |
1435 | gimple_call_set_lhs (call, retval); | |
1436 | update_stmt (call); | |
1437 | } | |
2862cf88 | 1438 | } |
9ece8f57 | 1439 | else |
1440 | gsi_insert_after (&gsi, call, GSI_NEW_STMT); | |
2862cf88 | 1441 | } |
524a0531 | 1442 | /* We don't use return block (there is either no return in function or |
1443 | multiple of them). So create new basic block with return statement. | |
1444 | */ | |
2862cf88 | 1445 | else |
1446 | { | |
1447 | gimple ret; | |
b04bab7c | 1448 | if (split_point->split_part_set_retval |
1449 | && !VOID_TYPE_P (TREE_TYPE (TREE_TYPE (current_function_decl)))) | |
2862cf88 | 1450 | { |
1b268210 | 1451 | retval = DECL_RESULT (current_function_decl); |
9eb0b8ee | 1452 | |
1453 | /* We use temporary register to hold value when aggregate_value_p | |
1454 | is false. Similarly for DECL_BY_REFERENCE we must avoid extra | |
1455 | copy. */ | |
1456 | if (!aggregate_value_p (retval, TREE_TYPE (current_function_decl)) | |
1457 | && !DECL_BY_REFERENCE (retval)) | |
1458 | retval = create_tmp_reg (TREE_TYPE (retval), NULL); | |
2862cf88 | 1459 | if (is_gimple_reg (retval)) |
524a0531 | 1460 | { |
1461 | /* When returning by reference, there is only one SSA name | |
1462 | assigned to RESULT_DECL (that is pointer to return value). | |
1463 | Look it up or create new one if it is missing. */ | |
1464 | if (DECL_BY_REFERENCE (retval)) | |
c6dfe037 | 1465 | retval = get_or_create_ssa_default_def (cfun, retval); |
524a0531 | 1466 | /* Otherwise produce new SSA name for return value. */ |
1467 | else | |
1468 | retval = make_ssa_name (retval, call); | |
1469 | } | |
a8005893 | 1470 | if (DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) |
1471 | gimple_call_set_lhs (call, build_simple_mem_ref (retval)); | |
1472 | else | |
1473 | gimple_call_set_lhs (call, retval); | |
2862cf88 | 1474 | } |
1475 | gsi_insert_after (&gsi, call, GSI_NEW_STMT); | |
1476 | ret = gimple_build_return (retval); | |
1477 | gsi_insert_after (&gsi, ret, GSI_NEW_STMT); | |
1478 | } | |
1479 | } | |
1480 | free_dominance_info (CDI_DOMINATORS); | |
1481 | free_dominance_info (CDI_POST_DOMINATORS); | |
a41f2a28 | 1482 | compute_inline_parameters (node, true); |
2862cf88 | 1483 | } |
1484 | ||
1485 | /* Execute function splitting pass. */ | |
1486 | ||
1487 | static unsigned int | |
1488 | execute_split_functions (void) | |
1489 | { | |
1490 | gimple_stmt_iterator bsi; | |
1491 | basic_block bb; | |
1492 | int overall_time = 0, overall_size = 0; | |
1493 | int todo = 0; | |
fd6a3c41 | 1494 | struct cgraph_node *node = cgraph_get_node (current_function_decl); |
2862cf88 | 1495 | |
f3941af5 | 1496 | if (flags_from_decl_or_type (current_function_decl) |
1497 | & (ECF_NORETURN|ECF_MALLOC)) | |
2862cf88 | 1498 | { |
1499 | if (dump_file) | |
f3941af5 | 1500 | fprintf (dump_file, "Not splitting: noreturn/malloc function.\n"); |
2862cf88 | 1501 | return 0; |
1502 | } | |
1503 | if (MAIN_NAME_P (DECL_NAME (current_function_decl))) | |
1504 | { | |
1505 | if (dump_file) | |
1506 | fprintf (dump_file, "Not splitting: main function.\n"); | |
1507 | return 0; | |
1508 | } | |
1509 | /* This can be relaxed; function might become inlinable after splitting | |
1510 | away the uninlinable part. */ | |
f1f41a6c | 1511 | if (inline_edge_summary_vec.exists () |
1512 | && !inline_summary (node)->inlinable) | |
2862cf88 | 1513 | { |
1514 | if (dump_file) | |
1515 | fprintf (dump_file, "Not splitting: not inlinable.\n"); | |
1516 | return 0; | |
1517 | } | |
7d0d0ce1 | 1518 | if (DECL_DISREGARD_INLINE_LIMITS (node->symbol.decl)) |
2862cf88 | 1519 | { |
1520 | if (dump_file) | |
574c4ab5 | 1521 | fprintf (dump_file, "Not splitting: disregarding inline limits.\n"); |
2862cf88 | 1522 | return 0; |
1523 | } | |
1524 | /* This can be relaxed; most of versioning tests actually prevents | |
1525 | a duplication. */ | |
1526 | if (!tree_versionable_function_p (current_function_decl)) | |
1527 | { | |
1528 | if (dump_file) | |
1529 | fprintf (dump_file, "Not splitting: not versionable.\n"); | |
1530 | return 0; | |
1531 | } | |
1532 | /* FIXME: we could support this. */ | |
1533 | if (DECL_STRUCT_FUNCTION (current_function_decl)->static_chain_decl) | |
1534 | { | |
1535 | if (dump_file) | |
1536 | fprintf (dump_file, "Not splitting: nested function.\n"); | |
1537 | return 0; | |
1538 | } | |
2862cf88 | 1539 | |
1540 | /* See if it makes sense to try to split. | |
1541 | It makes sense to split if we inline, that is if we have direct calls to | |
1542 | handle or direct calls are possibly going to appear as result of indirect | |
a15d5ede | 1543 | inlining or LTO. Also handle -fprofile-generate as LTO to allow non-LTO |
1544 | training for LTO -fprofile-use build. | |
1545 | ||
2862cf88 | 1546 | Note that we are not completely conservative about disqualifying functions |
1547 | called once. It is possible that the caller is called more then once and | |
1548 | then inlining would still benefit. */ | |
8745ccbf | 1549 | if ((!node->callers |
1550 | /* Local functions called once will be completely inlined most of time. */ | |
1551 | || (!node->callers->next_caller && node->local.local)) | |
7d0d0ce1 | 1552 | && !node->symbol.address_taken |
1553 | && (!flag_lto || !node->symbol.externally_visible)) | |
2862cf88 | 1554 | { |
1555 | if (dump_file) | |
1556 | fprintf (dump_file, "Not splitting: not called directly " | |
1557 | "or called once.\n"); | |
1558 | return 0; | |
1559 | } | |
1560 | ||
1561 | /* FIXME: We can actually split if splitting reduces call overhead. */ | |
1562 | if (!flag_inline_small_functions | |
1563 | && !DECL_DECLARED_INLINE_P (current_function_decl)) | |
1564 | { | |
1565 | if (dump_file) | |
1566 | fprintf (dump_file, "Not splitting: not autoinlining and function" | |
1567 | " is not inline.\n"); | |
1568 | return 0; | |
1569 | } | |
1570 | ||
15ca8f90 | 1571 | /* We enforce splitting after loop headers when profile info is not |
1572 | available. */ | |
1573 | if (profile_status != PROFILE_READ) | |
1574 | mark_dfs_back_edges (); | |
1575 | ||
aa2edcb9 | 1576 | /* Initialize bitmap to track forbidden calls. */ |
1577 | forbidden_dominators = BITMAP_ALLOC (NULL); | |
1578 | calculate_dominance_info (CDI_DOMINATORS); | |
1579 | ||
2862cf88 | 1580 | /* Compute local info about basic blocks and determine function size/time. */ |
f1f41a6c | 1581 | bb_info_vec.safe_grow_cleared (last_basic_block + 1); |
2862cf88 | 1582 | memset (&best_split_point, 0, sizeof (best_split_point)); |
1583 | FOR_EACH_BB (bb) | |
1584 | { | |
1585 | int time = 0; | |
1586 | int size = 0; | |
1587 | int freq = compute_call_stmt_bb_frequency (current_function_decl, bb); | |
1588 | ||
1589 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1590 | fprintf (dump_file, "Basic block %i\n", bb->index); | |
1591 | ||
1592 | for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
1593 | { | |
1594 | int this_time, this_size; | |
1595 | gimple stmt = gsi_stmt (bsi); | |
1596 | ||
1597 | this_size = estimate_num_insns (stmt, &eni_size_weights); | |
1598 | this_time = estimate_num_insns (stmt, &eni_time_weights) * freq; | |
1599 | size += this_size; | |
1600 | time += this_time; | |
aa2edcb9 | 1601 | check_forbidden_calls (stmt); |
2862cf88 | 1602 | |
1603 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1604 | { | |
1605 | fprintf (dump_file, " freq:%6i size:%3i time:%3i ", | |
1606 | freq, this_size, this_time); | |
1607 | print_gimple_stmt (dump_file, stmt, 0, 0); | |
1608 | } | |
1609 | } | |
1610 | overall_time += time; | |
1611 | overall_size += size; | |
f1f41a6c | 1612 | bb_info_vec[bb->index].time = time; |
1613 | bb_info_vec[bb->index].size = size; | |
2862cf88 | 1614 | } |
1615 | find_split_points (overall_time, overall_size); | |
1616 | if (best_split_point.split_bbs) | |
1617 | { | |
1618 | split_function (&best_split_point); | |
1619 | BITMAP_FREE (best_split_point.ssa_names_to_pass); | |
1620 | BITMAP_FREE (best_split_point.split_bbs); | |
1621 | todo = TODO_update_ssa | TODO_cleanup_cfg; | |
1622 | } | |
aa2edcb9 | 1623 | BITMAP_FREE (forbidden_dominators); |
f1f41a6c | 1624 | bb_info_vec.release (); |
2862cf88 | 1625 | return todo; |
1626 | } | |
1627 | ||
a15d5ede | 1628 | /* Gate function splitting pass. When doing profile feedback, we want |
1629 | to execute the pass after profiling is read. So disable one in | |
1630 | early optimization. */ | |
1631 | ||
2862cf88 | 1632 | static bool |
1633 | gate_split_functions (void) | |
1634 | { | |
a15d5ede | 1635 | return (flag_partial_inlining |
1636 | && !profile_arc_flag && !flag_branch_probabilities); | |
2862cf88 | 1637 | } |
1638 | ||
cbe8bda8 | 1639 | namespace { |
1640 | ||
1641 | const pass_data pass_data_split_functions = | |
2862cf88 | 1642 | { |
cbe8bda8 | 1643 | GIMPLE_PASS, /* type */ |
1644 | "fnsplit", /* name */ | |
1645 | OPTGROUP_NONE, /* optinfo_flags */ | |
1646 | true, /* has_gate */ | |
1647 | true, /* has_execute */ | |
1648 | TV_IPA_FNSPLIT, /* tv_id */ | |
1649 | PROP_cfg, /* properties_required */ | |
1650 | 0, /* properties_provided */ | |
1651 | 0, /* properties_destroyed */ | |
1652 | 0, /* todo_flags_start */ | |
1653 | TODO_verify_all, /* todo_flags_finish */ | |
2862cf88 | 1654 | }; |
a15d5ede | 1655 | |
cbe8bda8 | 1656 | class pass_split_functions : public gimple_opt_pass |
1657 | { | |
1658 | public: | |
9af5ce0c | 1659 | pass_split_functions (gcc::context *ctxt) |
1660 | : gimple_opt_pass (pass_data_split_functions, ctxt) | |
cbe8bda8 | 1661 | {} |
1662 | ||
1663 | /* opt_pass methods: */ | |
1664 | bool gate () { return gate_split_functions (); } | |
1665 | unsigned int execute () { return execute_split_functions (); } | |
1666 | ||
1667 | }; // class pass_split_functions | |
1668 | ||
1669 | } // anon namespace | |
1670 | ||
1671 | gimple_opt_pass * | |
1672 | make_pass_split_functions (gcc::context *ctxt) | |
1673 | { | |
1674 | return new pass_split_functions (ctxt); | |
1675 | } | |
1676 | ||
a15d5ede | 1677 | /* Gate feedback driven function splitting pass. |
1678 | We don't need to split when profiling at all, we are producing | |
1679 | lousy code anyway. */ | |
1680 | ||
1681 | static bool | |
1682 | gate_feedback_split_functions (void) | |
1683 | { | |
1684 | return (flag_partial_inlining | |
1685 | && flag_branch_probabilities); | |
1686 | } | |
1687 | ||
1688 | /* Execute function splitting pass. */ | |
1689 | ||
1690 | static unsigned int | |
1691 | execute_feedback_split_functions (void) | |
1692 | { | |
1693 | unsigned int retval = execute_split_functions (); | |
1694 | if (retval) | |
1695 | retval |= TODO_rebuild_cgraph_edges; | |
1696 | return retval; | |
1697 | } | |
1698 | ||
cbe8bda8 | 1699 | namespace { |
1700 | ||
1701 | const pass_data pass_data_feedback_split_functions = | |
a15d5ede | 1702 | { |
cbe8bda8 | 1703 | GIMPLE_PASS, /* type */ |
1704 | "feedback_fnsplit", /* name */ | |
1705 | OPTGROUP_NONE, /* optinfo_flags */ | |
1706 | true, /* has_gate */ | |
1707 | true, /* has_execute */ | |
1708 | TV_IPA_FNSPLIT, /* tv_id */ | |
1709 | PROP_cfg, /* properties_required */ | |
1710 | 0, /* properties_provided */ | |
1711 | 0, /* properties_destroyed */ | |
1712 | 0, /* todo_flags_start */ | |
1713 | TODO_verify_all, /* todo_flags_finish */ | |
a15d5ede | 1714 | }; |
cbe8bda8 | 1715 | |
1716 | class pass_feedback_split_functions : public gimple_opt_pass | |
1717 | { | |
1718 | public: | |
9af5ce0c | 1719 | pass_feedback_split_functions (gcc::context *ctxt) |
1720 | : gimple_opt_pass (pass_data_feedback_split_functions, ctxt) | |
cbe8bda8 | 1721 | {} |
1722 | ||
1723 | /* opt_pass methods: */ | |
1724 | bool gate () { return gate_feedback_split_functions (); } | |
1725 | unsigned int execute () { return execute_feedback_split_functions (); } | |
1726 | ||
1727 | }; // class pass_feedback_split_functions | |
1728 | ||
1729 | } // anon namespace | |
1730 | ||
1731 | gimple_opt_pass * | |
1732 | make_pass_feedback_split_functions (gcc::context *ctxt) | |
1733 | { | |
1734 | return new pass_feedback_split_functions (ctxt); | |
1735 | } |