]>
Commit | Line | Data |
---|---|---|
2862cf88 | 1 | /* Function splitting pass |
2 | Copyright (C) 2010 | |
3 | Free Software Foundation, Inc. | |
4 | Contributed by Jan Hubicka <jh@suse.cz> | |
5 | ||
6 | This file is part of GCC. | |
7 | ||
8 | GCC is free software; you can redistribute it and/or modify it under | |
9 | the terms of the GNU General Public License as published by the Free | |
10 | Software Foundation; either version 3, or (at your option) any later | |
11 | version. | |
12 | ||
13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
16 | for more details. | |
17 | ||
18 | You should have received a copy of the GNU General Public License | |
19 | along with GCC; see the file COPYING3. If not see | |
20 | <http://www.gnu.org/licenses/>. */ | |
21 | ||
22 | /* The purpose of this pass is to split function bodies to improve | |
23 | inlining. I.e. for function of the form: | |
24 | ||
25 | func (...) | |
26 | { | |
27 | if (cheap_test) | |
28 | something_small | |
29 | else | |
30 | something_big | |
31 | } | |
32 | ||
33 | Produce: | |
34 | ||
35 | func.part (...) | |
36 | { | |
37 | something_big | |
38 | } | |
39 | ||
40 | func (...) | |
41 | { | |
42 | if (cheap_test) | |
43 | something_small | |
44 | else | |
45 | func.part (...); | |
46 | } | |
47 | ||
48 | When func becomes inlinable and when cheap_test is often true, inlining func, | |
49 | but not fund.part leads to performance imrovement similar as inlining | |
50 | original func while the code size growth is smaller. | |
51 | ||
52 | The pass is organized in three stages: | |
53 | 1) Collect local info about basic block into BB_INFO structure and | |
54 | compute function body estimated size and time. | |
55 | 2) Via DFS walk find all possible basic blocks where we can split | |
56 | and chose best one. | |
57 | 3) If split point is found, split at the specified BB by creating a clone | |
58 | and updating function to call it. | |
59 | ||
60 | The decisions what functions to split are in execute_split_functions | |
61 | and consider_split. | |
62 | ||
63 | There are several possible future improvements for this pass including: | |
64 | ||
65 | 1) Splitting to break up large functions | |
66 | 2) Splitting to reduce stack frame usage | |
67 | 3) Allow split part of function to use values computed in the header part. | |
68 | The values needs to be passed to split function, perhaps via same | |
69 | interface as for nested functions or as argument. | |
70 | 4) Support for simple rematerialization. I.e. when split part use | |
71 | value computed in header from function parameter in very cheap way, we | |
72 | can just recompute it. | |
73 | 5) Support splitting of nested functions. | |
74 | 6) Support non-SSA arguments. | |
75 | 7) There is nothing preventing us from producing multiple parts of single function | |
76 | when needed or splitting also the parts. */ | |
77 | ||
78 | #include "config.h" | |
79 | #include "system.h" | |
80 | #include "coretypes.h" | |
81 | #include "tree.h" | |
82 | #include "target.h" | |
83 | #include "cgraph.h" | |
84 | #include "ipa-prop.h" | |
85 | #include "tree-flow.h" | |
86 | #include "tree-pass.h" | |
87 | #include "flags.h" | |
88 | #include "timevar.h" | |
89 | #include "diagnostic.h" | |
90 | #include "tree-dump.h" | |
91 | #include "tree-inline.h" | |
92 | #include "fibheap.h" | |
93 | #include "params.h" | |
94 | #include "gimple-pretty-print.h" | |
95 | ||
96 | /* Per basic block info. */ | |
97 | ||
98 | typedef struct | |
99 | { | |
100 | unsigned int size; | |
101 | unsigned int time; | |
102 | } bb_info; | |
103 | DEF_VEC_O(bb_info); | |
104 | DEF_VEC_ALLOC_O(bb_info,heap); | |
105 | ||
106 | static VEC(bb_info, heap) *bb_info_vec; | |
107 | ||
108 | /* Description of split point. */ | |
109 | ||
110 | struct split_point | |
111 | { | |
112 | /* Size of the partitions. */ | |
113 | unsigned int header_time, header_size, split_time, split_size; | |
114 | ||
115 | /* SSA names that need to be passed into spit funciton. */ | |
116 | bitmap ssa_names_to_pass; | |
117 | ||
118 | /* Basic block where we split (that will become entry point of new function. */ | |
119 | basic_block entry_bb; | |
120 | ||
121 | /* Basic blocks we are splitting away. */ | |
122 | bitmap split_bbs; | |
b04bab7c | 123 | |
124 | /* True when return value is computed on split part and thus it needs | |
125 | to be returned. */ | |
126 | bool split_part_set_retval; | |
2862cf88 | 127 | }; |
128 | ||
129 | /* Best split point found. */ | |
130 | ||
131 | struct split_point best_split_point; | |
132 | ||
b04bab7c | 133 | static tree find_retval (basic_block return_bb); |
134 | ||
2862cf88 | 135 | /* Callback for walk_stmt_load_store_addr_ops. If T is non-ssa automatic |
136 | variable, check it if it is present in bitmap passed via DATA. */ | |
137 | ||
138 | static bool | |
139 | test_nonssa_use (gimple stmt ATTRIBUTE_UNUSED, tree t, | |
140 | void *data ATTRIBUTE_UNUSED) | |
141 | { | |
142 | t = get_base_address (t); | |
143 | ||
144 | if (t && !is_gimple_reg (t) | |
145 | && ((TREE_CODE (t) == VAR_DECL | |
146 | && auto_var_in_fn_p (t, current_function_decl)) | |
c12a7417 | 147 | || (TREE_CODE (t) == RESULT_DECL) |
2862cf88 | 148 | || (TREE_CODE (t) == PARM_DECL))) |
149 | return bitmap_bit_p ((bitmap)data, DECL_UID (t)); | |
b04bab7c | 150 | |
151 | /* For DECL_BY_REFERENCE, the return value is actually pointer. We want to pretend | |
152 | that the value pointed to is actual result decl. */ | |
153 | if (t && (TREE_CODE (t) == MEM_REF || INDIRECT_REF_P (t)) | |
154 | && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME | |
155 | && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND (t, 0))) == RESULT_DECL | |
156 | && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) | |
157 | return bitmap_bit_p ((bitmap)data, DECL_UID (DECL_RESULT (current_function_decl))); | |
2862cf88 | 158 | return false; |
159 | } | |
160 | ||
161 | /* Dump split point CURRENT. */ | |
162 | ||
163 | static void | |
164 | dump_split_point (FILE * file, struct split_point *current) | |
165 | { | |
166 | fprintf (file, | |
167 | "Split point at BB %i header time:%i header size: %i" | |
168 | " split time: %i split size: %i\n bbs: ", | |
169 | current->entry_bb->index, current->header_time, | |
170 | current->header_size, current->split_time, current->split_size); | |
171 | dump_bitmap (file, current->split_bbs); | |
172 | fprintf (file, " SSA names to pass: "); | |
173 | dump_bitmap (file, current->ssa_names_to_pass); | |
174 | } | |
175 | ||
4493dab3 | 176 | /* Look for all BBs in header that might lead to split part and verify that |
177 | they are not defining any of SSA vars used by split part. | |
178 | Parameters are the same as for consider_split. */ | |
179 | ||
180 | static bool | |
181 | verify_non_ssa_vars (struct split_point *current, bitmap non_ssa_vars, | |
182 | basic_block return_bb) | |
183 | { | |
184 | bitmap seen = BITMAP_ALLOC (NULL); | |
185 | VEC (basic_block,heap) *worklist = NULL; | |
186 | edge e; | |
187 | edge_iterator ei; | |
188 | bool ok = true; | |
189 | ||
190 | FOR_EACH_EDGE (e, ei, current->entry_bb->preds) | |
191 | if (e->src != ENTRY_BLOCK_PTR | |
192 | && !bitmap_bit_p (current->split_bbs, e->src->index)) | |
193 | { | |
194 | VEC_safe_push (basic_block, heap, worklist, e->src); | |
195 | bitmap_set_bit (seen, e->src->index); | |
196 | } | |
197 | ||
198 | while (!VEC_empty (basic_block, worklist)) | |
199 | { | |
200 | gimple_stmt_iterator bsi; | |
201 | basic_block bb = VEC_pop (basic_block, worklist); | |
202 | ||
203 | FOR_EACH_EDGE (e, ei, bb->preds) | |
204 | if (e->src != ENTRY_BLOCK_PTR | |
6ef9bbe0 | 205 | && bitmap_set_bit (seen, e->src->index)) |
4493dab3 | 206 | { |
207 | gcc_checking_assert (!bitmap_bit_p (current->split_bbs, | |
208 | e->src->index)); | |
209 | VEC_safe_push (basic_block, heap, worklist, e->src); | |
4493dab3 | 210 | } |
211 | for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
212 | { | |
213 | if (is_gimple_debug (gsi_stmt (bsi))) | |
214 | continue; | |
215 | if (walk_stmt_load_store_addr_ops | |
216 | (gsi_stmt (bsi), non_ssa_vars, test_nonssa_use, | |
217 | test_nonssa_use, test_nonssa_use)) | |
218 | { | |
219 | ok = false; | |
220 | goto done; | |
221 | } | |
222 | } | |
223 | for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
224 | { | |
225 | if (walk_stmt_load_store_addr_ops | |
226 | (gsi_stmt (bsi), non_ssa_vars, test_nonssa_use, | |
227 | test_nonssa_use, test_nonssa_use)) | |
228 | { | |
229 | ok = false; | |
230 | goto done; | |
231 | } | |
232 | } | |
233 | FOR_EACH_EDGE (e, ei, bb->succs) | |
234 | { | |
235 | if (e->dest != return_bb) | |
236 | continue; | |
237 | for (bsi = gsi_start_phis (return_bb); !gsi_end_p (bsi); | |
238 | gsi_next (&bsi)) | |
239 | { | |
240 | gimple stmt = gsi_stmt (bsi); | |
241 | tree op = gimple_phi_arg_def (stmt, e->dest_idx); | |
242 | ||
243 | if (!is_gimple_reg (gimple_phi_result (stmt))) | |
244 | continue; | |
245 | if (TREE_CODE (op) != SSA_NAME | |
246 | && test_nonssa_use (stmt, op, non_ssa_vars)) | |
247 | { | |
248 | ok = false; | |
249 | goto done; | |
250 | } | |
251 | } | |
252 | } | |
253 | } | |
254 | done: | |
255 | BITMAP_FREE (seen); | |
256 | VEC_free (basic_block, heap, worklist); | |
257 | return ok; | |
258 | } | |
259 | ||
2862cf88 | 260 | /* We found an split_point CURRENT. NON_SSA_VARS is bitmap of all non ssa |
261 | variables used and RETURN_BB is return basic block. | |
262 | See if we can split function here. */ | |
263 | ||
264 | static void | |
265 | consider_split (struct split_point *current, bitmap non_ssa_vars, | |
266 | basic_block return_bb) | |
267 | { | |
268 | tree parm; | |
269 | unsigned int num_args = 0; | |
270 | unsigned int call_overhead; | |
271 | edge e; | |
272 | edge_iterator ei; | |
6a69e813 | 273 | gimple_stmt_iterator bsi; |
274 | unsigned int i; | |
275 | int incomming_freq = 0; | |
b04bab7c | 276 | tree retval; |
6a69e813 | 277 | |
2862cf88 | 278 | if (dump_file && (dump_flags & TDF_DETAILS)) |
279 | dump_split_point (dump_file, current); | |
280 | ||
6a69e813 | 281 | FOR_EACH_EDGE (e, ei, current->entry_bb->preds) |
282 | if (!bitmap_bit_p (current->split_bbs, e->src->index)) | |
283 | incomming_freq += EDGE_FREQUENCY (e); | |
284 | ||
2862cf88 | 285 | /* Do not split when we would end up calling function anyway. */ |
6a69e813 | 286 | if (incomming_freq |
2862cf88 | 287 | >= (ENTRY_BLOCK_PTR->frequency |
288 | * PARAM_VALUE (PARAM_PARTIAL_INLINING_ENTRY_PROBABILITY) / 100)) | |
289 | { | |
290 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
291 | fprintf (dump_file, | |
6a69e813 | 292 | " Refused: incomming frequency is too large.\n"); |
2862cf88 | 293 | return; |
294 | } | |
295 | ||
296 | if (!current->header_size) | |
297 | { | |
298 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
299 | fprintf (dump_file, " Refused: header empty\n"); | |
300 | gcc_unreachable (); | |
301 | return; | |
302 | } | |
303 | ||
6a69e813 | 304 | /* Verify that PHI args on entry are either virutal or all their operands |
305 | incomming from header are the same. */ | |
306 | for (bsi = gsi_start_phis (current->entry_bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
2862cf88 | 307 | { |
6a69e813 | 308 | gimple stmt = gsi_stmt (bsi); |
309 | tree val = NULL; | |
310 | ||
311 | if (!is_gimple_reg (gimple_phi_result (stmt))) | |
312 | continue; | |
313 | for (i = 0; i < gimple_phi_num_args (stmt); i++) | |
314 | { | |
315 | edge e = gimple_phi_arg_edge (stmt, i); | |
316 | if (!bitmap_bit_p (current->split_bbs, e->src->index)) | |
317 | { | |
318 | tree edge_val = gimple_phi_arg_def (stmt, i); | |
319 | if (val && edge_val != val) | |
320 | { | |
321 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
322 | fprintf (dump_file, | |
323 | " Refused: entry BB has PHI with multiple variants\n"); | |
324 | return; | |
325 | } | |
326 | val = edge_val; | |
327 | } | |
328 | } | |
2862cf88 | 329 | } |
330 | ||
331 | ||
332 | /* See what argument we will pass to the split function and compute | |
333 | call overhead. */ | |
334 | call_overhead = eni_size_weights.call_cost; | |
335 | for (parm = DECL_ARGUMENTS (current_function_decl); parm; | |
1767a056 | 336 | parm = DECL_CHAIN (parm)) |
2862cf88 | 337 | { |
338 | if (!is_gimple_reg (parm)) | |
339 | { | |
340 | if (bitmap_bit_p (non_ssa_vars, DECL_UID (parm))) | |
341 | { | |
342 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
343 | fprintf (dump_file, | |
344 | " Refused: need to pass non-ssa param values\n"); | |
345 | return; | |
346 | } | |
347 | } | |
348 | else if (gimple_default_def (cfun, parm) | |
349 | && bitmap_bit_p (current->ssa_names_to_pass, | |
350 | SSA_NAME_VERSION (gimple_default_def | |
351 | (cfun, parm)))) | |
352 | { | |
353 | if (!VOID_TYPE_P (TREE_TYPE (parm))) | |
354 | call_overhead += estimate_move_cost (TREE_TYPE (parm)); | |
355 | num_args++; | |
356 | } | |
357 | } | |
358 | if (!VOID_TYPE_P (TREE_TYPE (current_function_decl))) | |
359 | call_overhead += estimate_move_cost (TREE_TYPE (current_function_decl)); | |
360 | ||
361 | if (current->split_size <= call_overhead) | |
362 | { | |
363 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
364 | fprintf (dump_file, | |
365 | " Refused: split size is smaller than call overhead\n"); | |
366 | return; | |
367 | } | |
368 | if (current->header_size + call_overhead | |
369 | >= (unsigned int)(DECL_DECLARED_INLINE_P (current_function_decl) | |
370 | ? MAX_INLINE_INSNS_SINGLE | |
371 | : MAX_INLINE_INSNS_AUTO)) | |
372 | { | |
373 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
374 | fprintf (dump_file, | |
375 | " Refused: header size is too large for inline candidate\n"); | |
376 | return; | |
377 | } | |
378 | ||
379 | /* FIXME: we currently can pass only SSA function parameters to the split | |
fd8d648f | 380 | arguments. Once parm_adjustment infrastructure is supported by cloning, |
2862cf88 | 381 | we can pass more than that. */ |
382 | if (num_args != bitmap_count_bits (current->ssa_names_to_pass)) | |
383 | { | |
6a69e813 | 384 | |
2862cf88 | 385 | if (dump_file && (dump_flags & TDF_DETAILS)) |
386 | fprintf (dump_file, | |
387 | " Refused: need to pass non-param values\n"); | |
388 | return; | |
389 | } | |
390 | ||
391 | /* When there are non-ssa vars used in the split region, see if they | |
392 | are used in the header region. If so, reject the split. | |
393 | FIXME: we can use nested function support to access both. */ | |
4493dab3 | 394 | if (!bitmap_empty_p (non_ssa_vars) |
395 | && !verify_non_ssa_vars (current, non_ssa_vars, return_bb)) | |
2862cf88 | 396 | { |
4493dab3 | 397 | if (dump_file && (dump_flags & TDF_DETAILS)) |
398 | fprintf (dump_file, | |
399 | " Refused: split part has non-ssa uses\n"); | |
2862cf88 | 400 | return; |
401 | } | |
402 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
403 | fprintf (dump_file, " Accepted!\n"); | |
404 | ||
b04bab7c | 405 | /* See if retval used by return bb is computed by header or split part. |
406 | When it is computed by split part, we need to produce return statement | |
407 | in the split part and add code to header to pass it around. | |
408 | ||
409 | This is bit tricky to test: | |
410 | 1) When there is no return_bb or no return value, we always pass | |
411 | value around. | |
412 | 2) Invariants are always computed by caller. | |
413 | 3) For SSA we need to look if defining statement is in header or split part | |
414 | 4) For non-SSA we need to look where the var is computed. */ | |
415 | retval = find_retval (return_bb); | |
416 | if (!retval) | |
417 | current->split_part_set_retval = true; | |
418 | else if (is_gimple_min_invariant (retval)) | |
419 | current->split_part_set_retval = false; | |
420 | /* Special case is value returned by reference we record as if it was non-ssa | |
421 | set to result_decl. */ | |
422 | else if (TREE_CODE (retval) == SSA_NAME | |
423 | && TREE_CODE (SSA_NAME_VAR (retval)) == RESULT_DECL | |
424 | && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) | |
425 | current->split_part_set_retval | |
426 | = bitmap_bit_p (non_ssa_vars, DECL_UID (SSA_NAME_VAR (retval))); | |
427 | else if (TREE_CODE (retval) == SSA_NAME) | |
428 | current->split_part_set_retval | |
429 | = (!SSA_NAME_IS_DEFAULT_DEF (retval) | |
430 | && (bitmap_bit_p (current->split_bbs, | |
431 | gimple_bb (SSA_NAME_DEF_STMT (retval))->index) | |
432 | || gimple_bb (SSA_NAME_DEF_STMT (retval)) == return_bb)); | |
433 | else if (TREE_CODE (retval) == PARM_DECL) | |
434 | current->split_part_set_retval = false; | |
435 | else if (TREE_CODE (retval) == VAR_DECL | |
436 | || TREE_CODE (retval) == RESULT_DECL) | |
437 | current->split_part_set_retval | |
438 | = bitmap_bit_p (non_ssa_vars, DECL_UID (retval)); | |
439 | else | |
440 | current->split_part_set_retval = true; | |
441 | ||
2862cf88 | 442 | /* At the moment chose split point with lowest frequency and that leaves |
443 | out smallest size of header. | |
444 | In future we might re-consider this heuristics. */ | |
445 | if (!best_split_point.split_bbs | |
446 | || best_split_point.entry_bb->frequency > current->entry_bb->frequency | |
447 | || (best_split_point.entry_bb->frequency == current->entry_bb->frequency | |
448 | && best_split_point.split_size < current->split_size)) | |
449 | ||
450 | { | |
451 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
452 | fprintf (dump_file, " New best split point!\n"); | |
453 | if (best_split_point.ssa_names_to_pass) | |
454 | { | |
455 | BITMAP_FREE (best_split_point.ssa_names_to_pass); | |
456 | BITMAP_FREE (best_split_point.split_bbs); | |
457 | } | |
458 | best_split_point = *current; | |
459 | best_split_point.ssa_names_to_pass = BITMAP_ALLOC (NULL); | |
460 | bitmap_copy (best_split_point.ssa_names_to_pass, | |
461 | current->ssa_names_to_pass); | |
462 | best_split_point.split_bbs = BITMAP_ALLOC (NULL); | |
463 | bitmap_copy (best_split_point.split_bbs, current->split_bbs); | |
464 | } | |
465 | } | |
466 | ||
4493dab3 | 467 | /* Return basic block containing RETURN statement. We allow basic blocks |
468 | of the form: | |
469 | <retval> = tmp_var; | |
470 | return <retval> | |
471 | but return_bb can not be more complex than this. | |
472 | If nothing is found, return EXIT_BLOCK_PTR. | |
473 | ||
2862cf88 | 474 | When there are multiple RETURN statement, chose one with return value, |
475 | since that one is more likely shared by multiple code paths. | |
4493dab3 | 476 | |
477 | Return BB is special, because for function splitting it is the only | |
478 | basic block that is duplicated in between header and split part of the | |
479 | function. | |
480 | ||
2862cf88 | 481 | TODO: We might support multiple return blocks. */ |
482 | ||
483 | static basic_block | |
484 | find_return_bb (void) | |
485 | { | |
486 | edge e; | |
487 | edge_iterator ei; | |
488 | basic_block return_bb = EXIT_BLOCK_PTR; | |
489 | ||
490 | if (EDGE_COUNT (EXIT_BLOCK_PTR->preds) == 1) | |
491 | FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR->preds) | |
492 | { | |
493 | gimple_stmt_iterator bsi; | |
494 | bool found_return = false; | |
495 | tree retval = NULL_TREE; | |
496 | ||
4493dab3 | 497 | for (bsi = gsi_last_bb (e->src); !gsi_end_p (bsi); gsi_prev (&bsi)) |
498 | { | |
499 | gimple stmt = gsi_stmt (bsi); | |
500 | if (gimple_code (stmt) == GIMPLE_LABEL | |
501 | || is_gimple_debug (stmt)) | |
502 | ; | |
503 | else if (gimple_code (stmt) == GIMPLE_ASSIGN | |
504 | && found_return | |
505 | && gimple_assign_single_p (stmt) | |
506 | && (auto_var_in_fn_p (gimple_assign_rhs1 (stmt), | |
507 | current_function_decl) | |
508 | || is_gimple_min_invariant | |
509 | (gimple_assign_rhs1 (stmt))) | |
510 | && retval == gimple_assign_lhs (stmt)) | |
511 | ; | |
512 | else if (gimple_code (stmt) == GIMPLE_RETURN) | |
513 | { | |
514 | found_return = true; | |
515 | retval = gimple_return_retval (stmt); | |
516 | } | |
517 | else | |
518 | break; | |
519 | } | |
2862cf88 | 520 | if (gsi_end_p (bsi) && found_return) |
521 | { | |
522 | if (retval) | |
523 | return e->src; | |
524 | else | |
525 | return_bb = e->src; | |
526 | } | |
527 | } | |
528 | return return_bb; | |
529 | } | |
530 | ||
4493dab3 | 531 | /* Given return basicblock RETURN_BB, see where return value is really |
532 | stored. */ | |
533 | static tree | |
534 | find_retval (basic_block return_bb) | |
535 | { | |
536 | gimple_stmt_iterator bsi; | |
537 | for (bsi = gsi_start_bb (return_bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
538 | if (gimple_code (gsi_stmt (bsi)) == GIMPLE_RETURN) | |
539 | return gimple_return_retval (gsi_stmt (bsi)); | |
540 | else if (gimple_code (gsi_stmt (bsi)) == GIMPLE_ASSIGN) | |
541 | return gimple_assign_rhs1 (gsi_stmt (bsi)); | |
542 | return NULL; | |
543 | } | |
544 | ||
2862cf88 | 545 | /* Callback for walk_stmt_load_store_addr_ops. If T is non-ssa automatic |
546 | variable, mark it as used in bitmap passed via DATA. | |
547 | Return true when access to T prevents splitting the function. */ | |
548 | ||
549 | static bool | |
550 | mark_nonssa_use (gimple stmt ATTRIBUTE_UNUSED, tree t, | |
551 | void *data ATTRIBUTE_UNUSED) | |
552 | { | |
553 | t = get_base_address (t); | |
554 | ||
555 | if (!t || is_gimple_reg (t)) | |
556 | return false; | |
557 | ||
558 | /* At present we can't pass non-SSA arguments to split function. | |
559 | FIXME: this can be relaxed by passing references to arguments. */ | |
560 | if (TREE_CODE (t) == PARM_DECL) | |
561 | { | |
562 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
563 | fprintf (dump_file, "Can not split use of non-ssa function parameter.\n"); | |
564 | return true; | |
565 | } | |
566 | ||
c12a7417 | 567 | if ((TREE_CODE (t) == VAR_DECL && auto_var_in_fn_p (t, current_function_decl)) |
568 | || (TREE_CODE (t) == RESULT_DECL)) | |
2862cf88 | 569 | bitmap_set_bit ((bitmap)data, DECL_UID (t)); |
b04bab7c | 570 | |
571 | /* For DECL_BY_REFERENCE, the return value is actually pointer. We want to pretend | |
572 | that the value pointed to is actual result decl. */ | |
573 | if (t && (TREE_CODE (t) == MEM_REF || INDIRECT_REF_P (t)) | |
574 | && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME | |
575 | && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND (t, 0))) == RESULT_DECL | |
576 | && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) | |
577 | return bitmap_bit_p ((bitmap)data, DECL_UID (DECL_RESULT (current_function_decl))); | |
2862cf88 | 578 | return false; |
579 | } | |
580 | ||
581 | /* Compute local properties of basic block BB we collect when looking for | |
582 | split points. We look for ssa defs and store them in SET_SSA_NAMES, | |
583 | for ssa uses and store them in USED_SSA_NAMES and for any non-SSA automatic | |
584 | vars stored in NON_SSA_VARS. | |
585 | ||
586 | When BB has edge to RETURN_BB, collect uses in RETURN_BB too. | |
587 | ||
588 | Return false when BB contains something that prevents it from being put into | |
589 | split function. */ | |
590 | ||
591 | static bool | |
592 | visit_bb (basic_block bb, basic_block return_bb, | |
593 | bitmap set_ssa_names, bitmap used_ssa_names, | |
594 | bitmap non_ssa_vars) | |
595 | { | |
596 | gimple_stmt_iterator bsi; | |
597 | edge e; | |
598 | edge_iterator ei; | |
599 | bool can_split = true; | |
600 | ||
601 | for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
602 | { | |
603 | gimple stmt = gsi_stmt (bsi); | |
604 | tree op; | |
605 | ssa_op_iter iter; | |
606 | tree decl; | |
607 | ||
608 | if (is_gimple_debug (stmt)) | |
609 | continue; | |
610 | ||
611 | /* FIXME: We can split regions containing EH. We can not however | |
612 | split RESX, EH_DISPATCH and EH_POINTER referring to same region | |
613 | into different partitions. This would require tracking of | |
614 | EH regions and checking in consider_split_point if they | |
615 | are not used elsewhere. */ | |
616 | if (gimple_code (stmt) == GIMPLE_RESX | |
617 | && stmt_can_throw_external (stmt)) | |
618 | { | |
619 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
620 | fprintf (dump_file, "Can not split external resx.\n"); | |
621 | can_split = false; | |
622 | } | |
623 | if (gimple_code (stmt) == GIMPLE_EH_DISPATCH) | |
624 | { | |
625 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
626 | fprintf (dump_file, "Can not split eh dispatch.\n"); | |
627 | can_split = false; | |
628 | } | |
629 | ||
630 | /* Check builtins that prevent splitting. */ | |
631 | if (gimple_code (stmt) == GIMPLE_CALL | |
632 | && (decl = gimple_call_fndecl (stmt)) != NULL_TREE | |
633 | && DECL_BUILT_IN (decl) | |
634 | && DECL_BUILT_IN_CLASS (decl) == BUILT_IN_NORMAL) | |
635 | switch (DECL_FUNCTION_CODE (decl)) | |
636 | { | |
637 | /* FIXME: once we will allow passing non-parm values to split part, | |
638 | we need to be sure to handle correct builtin_stack_save and | |
639 | builtin_stack_restore. At the moment we are safe; there is no | |
640 | way to store builtin_stack_save result in non-SSA variable | |
641 | since all calls to those are compiler generated. */ | |
642 | case BUILT_IN_APPLY: | |
643 | case BUILT_IN_VA_START: | |
644 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
645 | fprintf (dump_file, "Can not split builtin_apply and va_start.\n"); | |
646 | can_split = false; | |
647 | break; | |
648 | case BUILT_IN_EH_POINTER: | |
649 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
650 | fprintf (dump_file, "Can not split builtin_eh_pointer.\n"); | |
651 | can_split = false; | |
652 | break; | |
653 | default: | |
654 | break; | |
655 | } | |
656 | ||
657 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF) | |
658 | bitmap_set_bit (set_ssa_names, SSA_NAME_VERSION (op)); | |
659 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE) | |
660 | bitmap_set_bit (used_ssa_names, SSA_NAME_VERSION (op)); | |
661 | can_split &= !walk_stmt_load_store_addr_ops (stmt, non_ssa_vars, | |
662 | mark_nonssa_use, | |
663 | mark_nonssa_use, | |
664 | mark_nonssa_use); | |
665 | } | |
666 | for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
667 | { | |
668 | gimple stmt = gsi_stmt (bsi); | |
6a69e813 | 669 | unsigned int i; |
2862cf88 | 670 | |
671 | if (is_gimple_debug (stmt)) | |
672 | continue; | |
673 | if (!is_gimple_reg (gimple_phi_result (stmt))) | |
674 | continue; | |
6a69e813 | 675 | bitmap_set_bit (set_ssa_names, |
676 | SSA_NAME_VERSION (gimple_phi_result (stmt))); | |
677 | for (i = 0; i < gimple_phi_num_args (stmt); i++) | |
678 | { | |
679 | tree op = gimple_phi_arg_def (stmt, i); | |
680 | if (TREE_CODE (op) == SSA_NAME) | |
681 | bitmap_set_bit (used_ssa_names, SSA_NAME_VERSION (op)); | |
682 | } | |
2862cf88 | 683 | can_split &= !walk_stmt_load_store_addr_ops (stmt, non_ssa_vars, |
684 | mark_nonssa_use, | |
685 | mark_nonssa_use, | |
686 | mark_nonssa_use); | |
687 | } | |
688 | /* Record also uses comming from PHI operand in return BB. */ | |
689 | FOR_EACH_EDGE (e, ei, bb->succs) | |
690 | if (e->dest == return_bb) | |
691 | { | |
2862cf88 | 692 | for (bsi = gsi_start_phis (return_bb); !gsi_end_p (bsi); gsi_next (&bsi)) |
693 | { | |
694 | gimple stmt = gsi_stmt (bsi); | |
695 | tree op = gimple_phi_arg_def (stmt, e->dest_idx); | |
696 | ||
697 | if (is_gimple_debug (stmt)) | |
698 | continue; | |
699 | if (!is_gimple_reg (gimple_phi_result (stmt))) | |
700 | continue; | |
2862cf88 | 701 | if (TREE_CODE (op) == SSA_NAME) |
702 | bitmap_set_bit (used_ssa_names, SSA_NAME_VERSION (op)); | |
703 | else | |
704 | can_split &= !mark_nonssa_use (stmt, op, non_ssa_vars); | |
705 | } | |
2862cf88 | 706 | } |
707 | return can_split; | |
708 | } | |
709 | ||
710 | /* Stack entry for recursive DFS walk in find_split_point. */ | |
711 | ||
712 | typedef struct | |
713 | { | |
714 | /* Basic block we are examining. */ | |
715 | basic_block bb; | |
716 | ||
717 | /* SSA names set and used by the BB and all BBs reachable | |
718 | from it via DFS walk. */ | |
719 | bitmap set_ssa_names, used_ssa_names; | |
720 | bitmap non_ssa_vars; | |
721 | ||
722 | /* All BBS visited from this BB via DFS walk. */ | |
723 | bitmap bbs_visited; | |
724 | ||
725 | /* Last examined edge in DFS walk. Since we walk unoriented graph, | |
726 | the value is up to sum of incomming and outgoing edges of BB. */ | |
727 | unsigned int edge_num; | |
728 | ||
729 | /* Stack entry index of earliest BB reachable from current BB | |
730 | or any BB visited later in DFS valk. */ | |
731 | int earliest; | |
732 | ||
733 | /* Overall time and size of all BBs reached from this BB in DFS walk. */ | |
734 | int overall_time, overall_size; | |
735 | ||
736 | /* When false we can not split on this BB. */ | |
737 | bool can_split; | |
738 | } stack_entry; | |
739 | DEF_VEC_O(stack_entry); | |
740 | DEF_VEC_ALLOC_O(stack_entry,heap); | |
741 | ||
742 | ||
743 | /* Find all articulations and call consider_split on them. | |
744 | OVERALL_TIME and OVERALL_SIZE is time and size of the function. | |
745 | ||
746 | We perform basic algorithm for finding an articulation in a graph | |
747 | created from CFG by considering it to be an unoriented graph. | |
748 | ||
749 | The articulation is discovered via DFS walk. We collect earliest | |
750 | basic block on stack that is reachable via backward edge. Articulation | |
751 | is any basic block such that there is no backward edge bypassing it. | |
752 | To reduce stack usage we maintain heap allocated stack in STACK vector. | |
753 | AUX pointer of BB is set to index it appears in the stack or -1 once | |
754 | it is visited and popped off the stack. | |
755 | ||
756 | The algorithm finds articulation after visiting the whole component | |
757 | reachable by it. This makes it convenient to collect information about | |
758 | the component used by consider_split. */ | |
759 | ||
760 | static void | |
761 | find_split_points (int overall_time, int overall_size) | |
762 | { | |
763 | stack_entry first; | |
764 | VEC(stack_entry, heap) *stack = NULL; | |
765 | basic_block bb; | |
766 | basic_block return_bb = find_return_bb (); | |
767 | struct split_point current; | |
768 | ||
769 | current.header_time = overall_time; | |
770 | current.header_size = overall_size; | |
771 | current.split_time = 0; | |
772 | current.split_size = 0; | |
773 | current.ssa_names_to_pass = BITMAP_ALLOC (NULL); | |
774 | ||
775 | first.bb = ENTRY_BLOCK_PTR; | |
776 | first.edge_num = 0; | |
777 | first.overall_time = 0; | |
778 | first.overall_size = 0; | |
779 | first.earliest = INT_MAX; | |
780 | first.set_ssa_names = 0; | |
781 | first.used_ssa_names = 0; | |
782 | first.bbs_visited = 0; | |
783 | VEC_safe_push (stack_entry, heap, stack, &first); | |
784 | ENTRY_BLOCK_PTR->aux = (void *)(intptr_t)-1; | |
785 | ||
786 | while (!VEC_empty (stack_entry, stack)) | |
787 | { | |
788 | stack_entry *entry = VEC_last (stack_entry, stack); | |
789 | ||
790 | /* We are walking an acyclic graph, so edge_num counts | |
791 | succ and pred edges together. However when considering | |
792 | articulation, we want to have processed everything reachable | |
793 | from articulation but nothing that reaches into it. */ | |
794 | if (entry->edge_num == EDGE_COUNT (entry->bb->succs) | |
795 | && entry->bb != ENTRY_BLOCK_PTR) | |
796 | { | |
797 | int pos = VEC_length (stack_entry, stack); | |
798 | entry->can_split &= visit_bb (entry->bb, return_bb, | |
799 | entry->set_ssa_names, | |
800 | entry->used_ssa_names, | |
801 | entry->non_ssa_vars); | |
802 | if (pos <= entry->earliest && !entry->can_split | |
803 | && dump_file && (dump_flags & TDF_DETAILS)) | |
804 | fprintf (dump_file, | |
805 | "found articulation at bb %i but can not split\n", | |
806 | entry->bb->index); | |
807 | if (pos <= entry->earliest && entry->can_split) | |
808 | { | |
809 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
810 | fprintf (dump_file, "found articulation at bb %i\n", | |
811 | entry->bb->index); | |
812 | current.entry_bb = entry->bb; | |
813 | current.ssa_names_to_pass = BITMAP_ALLOC (NULL); | |
814 | bitmap_and_compl (current.ssa_names_to_pass, | |
815 | entry->used_ssa_names, entry->set_ssa_names); | |
816 | current.header_time = overall_time - entry->overall_time; | |
817 | current.header_size = overall_size - entry->overall_size; | |
818 | current.split_time = entry->overall_time; | |
819 | current.split_size = entry->overall_size; | |
820 | current.split_bbs = entry->bbs_visited; | |
821 | consider_split (¤t, entry->non_ssa_vars, return_bb); | |
822 | BITMAP_FREE (current.ssa_names_to_pass); | |
823 | } | |
824 | } | |
825 | /* Do actual DFS walk. */ | |
826 | if (entry->edge_num | |
827 | < (EDGE_COUNT (entry->bb->succs) | |
828 | + EDGE_COUNT (entry->bb->preds))) | |
829 | { | |
830 | edge e; | |
831 | basic_block dest; | |
832 | if (entry->edge_num < EDGE_COUNT (entry->bb->succs)) | |
833 | { | |
834 | e = EDGE_SUCC (entry->bb, entry->edge_num); | |
835 | dest = e->dest; | |
836 | } | |
837 | else | |
838 | { | |
839 | e = EDGE_PRED (entry->bb, entry->edge_num | |
840 | - EDGE_COUNT (entry->bb->succs)); | |
841 | dest = e->src; | |
842 | } | |
843 | ||
844 | entry->edge_num++; | |
845 | ||
846 | /* New BB to visit, push it to the stack. */ | |
847 | if (dest != return_bb && dest != EXIT_BLOCK_PTR | |
848 | && !dest->aux) | |
849 | { | |
850 | stack_entry new_entry; | |
851 | ||
852 | new_entry.bb = dest; | |
853 | new_entry.edge_num = 0; | |
854 | new_entry.overall_time | |
855 | = VEC_index (bb_info, bb_info_vec, dest->index)->time; | |
856 | new_entry.overall_size | |
857 | = VEC_index (bb_info, bb_info_vec, dest->index)->size; | |
858 | new_entry.earliest = INT_MAX; | |
859 | new_entry.set_ssa_names = BITMAP_ALLOC (NULL); | |
860 | new_entry.used_ssa_names = BITMAP_ALLOC (NULL); | |
861 | new_entry.bbs_visited = BITMAP_ALLOC (NULL); | |
862 | new_entry.non_ssa_vars = BITMAP_ALLOC (NULL); | |
863 | new_entry.can_split = true; | |
864 | bitmap_set_bit (new_entry.bbs_visited, dest->index); | |
865 | VEC_safe_push (stack_entry, heap, stack, &new_entry); | |
866 | dest->aux = (void *)(intptr_t)VEC_length (stack_entry, stack); | |
867 | } | |
868 | /* Back edge found, record the earliest point. */ | |
869 | else if ((intptr_t)dest->aux > 0 | |
870 | && (intptr_t)dest->aux < entry->earliest) | |
871 | entry->earliest = (intptr_t)dest->aux; | |
872 | } | |
873 | /* We are done with examing the edges. pop off the value from stack and | |
874 | merge stuff we cummulate during the walk. */ | |
875 | else if (entry->bb != ENTRY_BLOCK_PTR) | |
876 | { | |
877 | stack_entry *prev = VEC_index (stack_entry, stack, | |
878 | VEC_length (stack_entry, stack) - 2); | |
879 | ||
880 | entry->bb->aux = (void *)(intptr_t)-1; | |
881 | prev->can_split &= entry->can_split; | |
882 | if (prev->set_ssa_names) | |
883 | { | |
884 | bitmap_ior_into (prev->set_ssa_names, entry->set_ssa_names); | |
885 | bitmap_ior_into (prev->used_ssa_names, entry->used_ssa_names); | |
886 | bitmap_ior_into (prev->bbs_visited, entry->bbs_visited); | |
887 | bitmap_ior_into (prev->non_ssa_vars, entry->non_ssa_vars); | |
888 | } | |
889 | if (prev->earliest > entry->earliest) | |
890 | prev->earliest = entry->earliest; | |
891 | prev->overall_time += entry->overall_time; | |
892 | prev->overall_size += entry->overall_size; | |
893 | BITMAP_FREE (entry->set_ssa_names); | |
894 | BITMAP_FREE (entry->used_ssa_names); | |
895 | BITMAP_FREE (entry->bbs_visited); | |
896 | BITMAP_FREE (entry->non_ssa_vars); | |
897 | VEC_pop (stack_entry, stack); | |
898 | } | |
899 | else | |
900 | VEC_pop (stack_entry, stack); | |
901 | } | |
902 | ENTRY_BLOCK_PTR->aux = NULL; | |
903 | FOR_EACH_BB (bb) | |
904 | bb->aux = NULL; | |
84b53675 | 905 | VEC_free (stack_entry, heap, stack); |
2862cf88 | 906 | BITMAP_FREE (current.ssa_names_to_pass); |
907 | } | |
908 | ||
909 | /* Split function at SPLIT_POINT. */ | |
910 | ||
911 | static void | |
912 | split_function (struct split_point *split_point) | |
913 | { | |
914 | VEC (tree, heap) *args_to_pass = NULL; | |
915 | bitmap args_to_skip = BITMAP_ALLOC (NULL); | |
916 | tree parm; | |
917 | int num = 0; | |
918 | struct cgraph_node *node; | |
919 | basic_block return_bb = find_return_bb (); | |
920 | basic_block call_bb; | |
921 | gimple_stmt_iterator gsi; | |
922 | gimple call; | |
923 | edge e; | |
924 | edge_iterator ei; | |
925 | tree retval = NULL, real_retval = NULL; | |
926 | bool split_part_return_p = false; | |
927 | gimple last_stmt = NULL; | |
928 | ||
929 | if (dump_file) | |
930 | { | |
931 | fprintf (dump_file, "\n\nSplitting function at:\n"); | |
932 | dump_split_point (dump_file, split_point); | |
933 | } | |
934 | ||
935 | /* Collect the parameters of new function and args_to_skip bitmap. */ | |
936 | for (parm = DECL_ARGUMENTS (current_function_decl); | |
1767a056 | 937 | parm; parm = DECL_CHAIN (parm), num++) |
2862cf88 | 938 | if (!is_gimple_reg (parm) |
939 | || !gimple_default_def (cfun, parm) | |
940 | || !bitmap_bit_p (split_point->ssa_names_to_pass, | |
941 | SSA_NAME_VERSION (gimple_default_def (cfun, parm)))) | |
942 | bitmap_set_bit (args_to_skip, num); | |
943 | else | |
944 | VEC_safe_push (tree, heap, args_to_pass, gimple_default_def (cfun, parm)); | |
945 | ||
946 | /* See if the split function will return. */ | |
947 | FOR_EACH_EDGE (e, ei, return_bb->preds) | |
948 | if (bitmap_bit_p (split_point->split_bbs, e->src->index)) | |
949 | break; | |
950 | if (e) | |
951 | split_part_return_p = true; | |
952 | ||
b04bab7c | 953 | /* Add return block to what will become the split function. |
954 | We do not return; no return block is needed. */ | |
955 | if (!split_part_return_p) | |
956 | ; | |
957 | /* We have no return block, so nothing is needed. */ | |
958 | else if (return_bb == EXIT_BLOCK_PTR) | |
959 | ; | |
960 | /* When we do not want to return value, we need to construct | |
961 | new return block with empty return statement. | |
962 | FIXME: Once we are able to change return type, we should change function | |
963 | to return void instead of just outputting function with undefined return | |
964 | value. For structures this affects quality of codegen. */ | |
965 | else if (!split_point->split_part_set_retval | |
966 | && find_retval (return_bb)) | |
967 | { | |
968 | bool redirected = true; | |
969 | basic_block new_return_bb = create_basic_block (NULL, 0, return_bb); | |
970 | gimple_stmt_iterator gsi = gsi_start_bb (new_return_bb); | |
971 | gsi_insert_after (&gsi, gimple_build_return (NULL), GSI_NEW_STMT); | |
972 | while (redirected) | |
973 | { | |
974 | redirected = false; | |
975 | FOR_EACH_EDGE (e, ei, return_bb->preds) | |
976 | if (bitmap_bit_p (split_point->split_bbs, e->src->index)) | |
977 | { | |
978 | new_return_bb->count += e->count; | |
979 | new_return_bb->frequency += EDGE_FREQUENCY (e); | |
980 | redirect_edge_and_branch (e, new_return_bb); | |
981 | redirected = true; | |
982 | break; | |
983 | } | |
984 | } | |
985 | e = make_edge (new_return_bb, EXIT_BLOCK_PTR, 0); | |
986 | e->probability = REG_BR_PROB_BASE; | |
987 | e->count = new_return_bb->count; | |
988 | bitmap_set_bit (split_point->split_bbs, new_return_bb->index); | |
989 | /* We change CFG in a way tree-inline is not able to compensate on while | |
990 | updating PHIs. There are only virtuals in return_bb, so recompute | |
991 | them. */ | |
992 | for (gsi = gsi_start_phis (return_bb); !gsi_end_p (gsi);) | |
993 | { | |
994 | gimple stmt = gsi_stmt (gsi); | |
995 | gcc_assert (!is_gimple_reg (gimple_phi_result (stmt))); | |
996 | mark_sym_for_renaming (SSA_NAME_VAR (PHI_RESULT (stmt))); | |
997 | gsi_remove (&gsi, false); | |
998 | } | |
999 | } | |
1000 | /* When we pass aorund the value, use existing return block. */ | |
1001 | else | |
2862cf88 | 1002 | bitmap_set_bit (split_point->split_bbs, return_bb->index); |
1003 | ||
1004 | /* Now create the actual clone. */ | |
1005 | rebuild_cgraph_edges (); | |
1006 | node = cgraph_function_versioning (cgraph_node (current_function_decl), | |
1007 | NULL, NULL, | |
1008 | args_to_skip, | |
1009 | split_point->split_bbs, | |
4493dab3 | 1010 | split_point->entry_bb, "part"); |
fd8d648f | 1011 | /* For usual cloning it is enough to clear builtin only when signature |
1012 | changes. For partial inlining we however can not expect the part | |
1013 | of builtin implementation to have same semantic as the whole. */ | |
1014 | if (DECL_BUILT_IN (node->decl)) | |
1015 | { | |
1016 | DECL_BUILT_IN_CLASS (node->decl) = NOT_BUILT_IN; | |
1017 | DECL_FUNCTION_CODE (node->decl) = (enum built_in_function) 0; | |
1018 | } | |
2862cf88 | 1019 | cgraph_node_remove_callees (cgraph_node (current_function_decl)); |
1020 | if (!split_part_return_p) | |
1021 | TREE_THIS_VOLATILE (node->decl) = 1; | |
1022 | if (dump_file) | |
1023 | dump_function_to_file (node->decl, dump_file, dump_flags); | |
1024 | ||
1025 | /* Create the basic block we place call into. It is the entry basic block | |
1026 | split after last label. */ | |
1027 | call_bb = split_point->entry_bb; | |
1028 | for (gsi = gsi_start_bb (call_bb); !gsi_end_p (gsi);) | |
1029 | if (gimple_code (gsi_stmt (gsi)) == GIMPLE_LABEL) | |
1030 | { | |
1031 | last_stmt = gsi_stmt (gsi); | |
1032 | gsi_next (&gsi); | |
1033 | } | |
1034 | else | |
1035 | break; | |
1036 | e = split_block (split_point->entry_bb, last_stmt); | |
1037 | remove_edge (e); | |
1038 | ||
1039 | /* Produce the call statement. */ | |
1040 | gsi = gsi_last_bb (call_bb); | |
1041 | call = gimple_build_call_vec (node->decl, args_to_pass); | |
1042 | gimple_set_block (call, DECL_INITIAL (current_function_decl)); | |
1043 | ||
a8005893 | 1044 | /* We avoid address being taken on any variable used by split part, |
1045 | so return slot optimization is always possible. Moreover this is | |
1046 | required to make DECL_BY_REFERENCE work. */ | |
1047 | if (aggregate_value_p (DECL_RESULT (current_function_decl), | |
1048 | TREE_TYPE (current_function_decl))) | |
1049 | gimple_call_set_return_slot_opt (call, true); | |
1050 | ||
2862cf88 | 1051 | /* Update return value. This is bit tricky. When we do not return, |
1052 | do nothing. When we return we might need to update return_bb | |
1053 | or produce a new return statement. */ | |
1054 | if (!split_part_return_p) | |
1055 | gsi_insert_after (&gsi, call, GSI_NEW_STMT); | |
1056 | else | |
1057 | { | |
1058 | e = make_edge (call_bb, return_bb, | |
1059 | return_bb == EXIT_BLOCK_PTR ? 0 : EDGE_FALLTHRU); | |
1060 | e->count = call_bb->count; | |
1061 | e->probability = REG_BR_PROB_BASE; | |
524a0531 | 1062 | |
1063 | /* If there is return basic block, see what value we need to store | |
1064 | return value into and put call just before it. */ | |
2862cf88 | 1065 | if (return_bb != EXIT_BLOCK_PTR) |
1066 | { | |
4493dab3 | 1067 | real_retval = retval = find_retval (return_bb); |
524a0531 | 1068 | |
b04bab7c | 1069 | if (real_retval && split_point->split_part_set_retval) |
2862cf88 | 1070 | { |
1071 | gimple_stmt_iterator psi; | |
1072 | ||
524a0531 | 1073 | /* See if we need new SSA_NAME for the result. |
1074 | When DECL_BY_REFERENCE is true, retval is actually pointer to | |
1075 | return value and it is constant in whole function. */ | |
1076 | if (TREE_CODE (retval) == SSA_NAME | |
1077 | && !DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) | |
2862cf88 | 1078 | { |
1079 | retval = make_ssa_name (SSA_NAME_VAR (retval), call); | |
524a0531 | 1080 | |
1081 | /* See if there is PHI defining return value. */ | |
1082 | for (psi = gsi_start_phis (return_bb); | |
1083 | !gsi_end_p (psi); gsi_next (&psi)) | |
1084 | if (is_gimple_reg (gimple_phi_result (gsi_stmt (psi)))) | |
1085 | break; | |
1086 | ||
1087 | /* When there is PHI, just update its value. */ | |
2862cf88 | 1088 | if (TREE_CODE (retval) == SSA_NAME |
1089 | && !gsi_end_p (psi)) | |
1090 | add_phi_arg (gsi_stmt (psi), retval, e, UNKNOWN_LOCATION); | |
524a0531 | 1091 | /* Otherwise update the return BB itself. |
1092 | find_return_bb allows at most one assignment to return value, | |
1093 | so update first statement. */ | |
1094 | else | |
2862cf88 | 1095 | { |
4493dab3 | 1096 | gimple_stmt_iterator bsi; |
1097 | for (bsi = gsi_start_bb (return_bb); !gsi_end_p (bsi); | |
1098 | gsi_next (&bsi)) | |
1099 | if (gimple_code (gsi_stmt (bsi)) == GIMPLE_RETURN) | |
1100 | { | |
1101 | gimple_return_set_retval (gsi_stmt (bsi), retval); | |
1102 | break; | |
1103 | } | |
1104 | else if (gimple_code (gsi_stmt (bsi)) == GIMPLE_ASSIGN) | |
1105 | { | |
1106 | gimple_assign_set_rhs1 (gsi_stmt (bsi), retval); | |
1107 | break; | |
1108 | } | |
1109 | update_stmt (gsi_stmt (bsi)); | |
2862cf88 | 1110 | } |
1111 | } | |
a8005893 | 1112 | if (DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) |
1113 | gimple_call_set_lhs (call, build_simple_mem_ref (retval)); | |
1114 | else | |
1115 | gimple_call_set_lhs (call, retval); | |
2862cf88 | 1116 | } |
1117 | gsi_insert_after (&gsi, call, GSI_NEW_STMT); | |
1118 | } | |
524a0531 | 1119 | /* We don't use return block (there is either no return in function or |
1120 | multiple of them). So create new basic block with return statement. | |
1121 | */ | |
2862cf88 | 1122 | else |
1123 | { | |
1124 | gimple ret; | |
b04bab7c | 1125 | if (split_point->split_part_set_retval |
1126 | && !VOID_TYPE_P (TREE_TYPE (TREE_TYPE (current_function_decl)))) | |
2862cf88 | 1127 | { |
1b268210 | 1128 | retval = DECL_RESULT (current_function_decl); |
9eb0b8ee | 1129 | |
1130 | /* We use temporary register to hold value when aggregate_value_p | |
1131 | is false. Similarly for DECL_BY_REFERENCE we must avoid extra | |
1132 | copy. */ | |
1133 | if (!aggregate_value_p (retval, TREE_TYPE (current_function_decl)) | |
1134 | && !DECL_BY_REFERENCE (retval)) | |
1135 | retval = create_tmp_reg (TREE_TYPE (retval), NULL); | |
2862cf88 | 1136 | if (is_gimple_reg (retval)) |
524a0531 | 1137 | { |
1138 | /* When returning by reference, there is only one SSA name | |
1139 | assigned to RESULT_DECL (that is pointer to return value). | |
1140 | Look it up or create new one if it is missing. */ | |
1141 | if (DECL_BY_REFERENCE (retval)) | |
1142 | { | |
1143 | tree retval_name; | |
1144 | if ((retval_name = gimple_default_def (cfun, retval)) | |
1145 | != NULL) | |
1146 | retval = retval_name; | |
1147 | else | |
1148 | { | |
1149 | retval_name = make_ssa_name (retval, | |
1150 | gimple_build_nop ()); | |
1151 | set_default_def (retval, retval_name); | |
1152 | retval = retval_name; | |
1153 | } | |
1154 | } | |
1155 | /* Otherwise produce new SSA name for return value. */ | |
1156 | else | |
1157 | retval = make_ssa_name (retval, call); | |
1158 | } | |
a8005893 | 1159 | if (DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) |
1160 | gimple_call_set_lhs (call, build_simple_mem_ref (retval)); | |
1161 | else | |
1162 | gimple_call_set_lhs (call, retval); | |
2862cf88 | 1163 | } |
1164 | gsi_insert_after (&gsi, call, GSI_NEW_STMT); | |
1165 | ret = gimple_build_return (retval); | |
1166 | gsi_insert_after (&gsi, ret, GSI_NEW_STMT); | |
1167 | } | |
1168 | } | |
1169 | free_dominance_info (CDI_DOMINATORS); | |
1170 | free_dominance_info (CDI_POST_DOMINATORS); | |
1171 | compute_inline_parameters (node); | |
1172 | } | |
1173 | ||
1174 | /* Execute function splitting pass. */ | |
1175 | ||
1176 | static unsigned int | |
1177 | execute_split_functions (void) | |
1178 | { | |
1179 | gimple_stmt_iterator bsi; | |
1180 | basic_block bb; | |
1181 | int overall_time = 0, overall_size = 0; | |
1182 | int todo = 0; | |
1183 | struct cgraph_node *node = cgraph_node (current_function_decl); | |
1184 | ||
1185 | if (flags_from_decl_or_type (current_function_decl) & ECF_NORETURN) | |
1186 | { | |
1187 | if (dump_file) | |
1188 | fprintf (dump_file, "Not splitting: noreturn function.\n"); | |
1189 | return 0; | |
1190 | } | |
1191 | if (MAIN_NAME_P (DECL_NAME (current_function_decl))) | |
1192 | { | |
1193 | if (dump_file) | |
1194 | fprintf (dump_file, "Not splitting: main function.\n"); | |
1195 | return 0; | |
1196 | } | |
1197 | /* This can be relaxed; function might become inlinable after splitting | |
1198 | away the uninlinable part. */ | |
1199 | if (!node->local.inlinable) | |
1200 | { | |
1201 | if (dump_file) | |
1202 | fprintf (dump_file, "Not splitting: not inlinable.\n"); | |
1203 | return 0; | |
1204 | } | |
1205 | if (node->local.disregard_inline_limits) | |
1206 | { | |
1207 | if (dump_file) | |
1208 | fprintf (dump_file, "Not splitting: disregading inline limits.\n"); | |
1209 | return 0; | |
1210 | } | |
1211 | /* This can be relaxed; most of versioning tests actually prevents | |
1212 | a duplication. */ | |
1213 | if (!tree_versionable_function_p (current_function_decl)) | |
1214 | { | |
1215 | if (dump_file) | |
1216 | fprintf (dump_file, "Not splitting: not versionable.\n"); | |
1217 | return 0; | |
1218 | } | |
1219 | /* FIXME: we could support this. */ | |
1220 | if (DECL_STRUCT_FUNCTION (current_function_decl)->static_chain_decl) | |
1221 | { | |
1222 | if (dump_file) | |
1223 | fprintf (dump_file, "Not splitting: nested function.\n"); | |
1224 | return 0; | |
1225 | } | |
2862cf88 | 1226 | |
1227 | /* See if it makes sense to try to split. | |
1228 | It makes sense to split if we inline, that is if we have direct calls to | |
1229 | handle or direct calls are possibly going to appear as result of indirect | |
1230 | inlining or LTO. | |
1231 | Note that we are not completely conservative about disqualifying functions | |
1232 | called once. It is possible that the caller is called more then once and | |
1233 | then inlining would still benefit. */ | |
1234 | if ((!node->callers || !node->callers->next_caller) | |
1235 | && !node->address_taken | |
1236 | && ((!flag_lto && !flag_whopr) || !node->local.externally_visible)) | |
1237 | { | |
1238 | if (dump_file) | |
1239 | fprintf (dump_file, "Not splitting: not called directly " | |
1240 | "or called once.\n"); | |
1241 | return 0; | |
1242 | } | |
1243 | ||
1244 | /* FIXME: We can actually split if splitting reduces call overhead. */ | |
1245 | if (!flag_inline_small_functions | |
1246 | && !DECL_DECLARED_INLINE_P (current_function_decl)) | |
1247 | { | |
1248 | if (dump_file) | |
1249 | fprintf (dump_file, "Not splitting: not autoinlining and function" | |
1250 | " is not inline.\n"); | |
1251 | return 0; | |
1252 | } | |
1253 | ||
1254 | /* Compute local info about basic blocks and determine function size/time. */ | |
1255 | VEC_safe_grow_cleared (bb_info, heap, bb_info_vec, last_basic_block + 1); | |
1256 | memset (&best_split_point, 0, sizeof (best_split_point)); | |
1257 | FOR_EACH_BB (bb) | |
1258 | { | |
1259 | int time = 0; | |
1260 | int size = 0; | |
1261 | int freq = compute_call_stmt_bb_frequency (current_function_decl, bb); | |
1262 | ||
1263 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1264 | fprintf (dump_file, "Basic block %i\n", bb->index); | |
1265 | ||
1266 | for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
1267 | { | |
1268 | int this_time, this_size; | |
1269 | gimple stmt = gsi_stmt (bsi); | |
1270 | ||
1271 | this_size = estimate_num_insns (stmt, &eni_size_weights); | |
1272 | this_time = estimate_num_insns (stmt, &eni_time_weights) * freq; | |
1273 | size += this_size; | |
1274 | time += this_time; | |
1275 | ||
1276 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1277 | { | |
1278 | fprintf (dump_file, " freq:%6i size:%3i time:%3i ", | |
1279 | freq, this_size, this_time); | |
1280 | print_gimple_stmt (dump_file, stmt, 0, 0); | |
1281 | } | |
1282 | } | |
1283 | overall_time += time; | |
1284 | overall_size += size; | |
1285 | VEC_index (bb_info, bb_info_vec, bb->index)->time = time; | |
1286 | VEC_index (bb_info, bb_info_vec, bb->index)->size = size; | |
1287 | } | |
1288 | find_split_points (overall_time, overall_size); | |
1289 | if (best_split_point.split_bbs) | |
1290 | { | |
1291 | split_function (&best_split_point); | |
1292 | BITMAP_FREE (best_split_point.ssa_names_to_pass); | |
1293 | BITMAP_FREE (best_split_point.split_bbs); | |
1294 | todo = TODO_update_ssa | TODO_cleanup_cfg; | |
1295 | } | |
1296 | VEC_free (bb_info, heap, bb_info_vec); | |
1297 | bb_info_vec = NULL; | |
1298 | return todo; | |
1299 | } | |
1300 | ||
1301 | static bool | |
1302 | gate_split_functions (void) | |
1303 | { | |
1304 | return flag_partial_inlining; | |
1305 | } | |
1306 | ||
1307 | struct gimple_opt_pass pass_split_functions = | |
1308 | { | |
1309 | { | |
1310 | GIMPLE_PASS, | |
1311 | "fnsplit", /* name */ | |
1312 | gate_split_functions, /* gate */ | |
1313 | execute_split_functions, /* execute */ | |
1314 | NULL, /* sub */ | |
1315 | NULL, /* next */ | |
1316 | 0, /* static_pass_number */ | |
1317 | TV_IPA_FNSPLIT, /* tv_id */ | |
1318 | PROP_cfg, /* properties_required */ | |
1319 | 0, /* properties_provided */ | |
1320 | 0, /* properties_destroyed */ | |
1321 | 0, /* todo_flags_start */ | |
1322 | TODO_dump_func /* todo_flags_finish */ | |
1323 | } | |
1324 | }; |