]> git.ipfire.org Git - thirdparty/gcc.git/blame - gcc/ipa-split.c
2011-01-20 Richard Guenther <rguenther@suse.de>
[thirdparty/gcc.git] / gcc / ipa-split.c
CommitLineData
2862cf88 1/* Function splitting pass
e65564ab 2 Copyright (C) 2010, 2011
2862cf88 3 Free Software Foundation, Inc.
4 Contributed by Jan Hubicka <jh@suse.cz>
5
6This file is part of GCC.
7
8GCC is free software; you can redistribute it and/or modify it under
9the terms of the GNU General Public License as published by the Free
10Software Foundation; either version 3, or (at your option) any later
11version.
12
13GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14WARRANTY; without even the implied warranty of MERCHANTABILITY or
15FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16for more details.
17
18You should have received a copy of the GNU General Public License
19along 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,
574c4ab5 49 but not fund.part leads to performance improvement similar as inlining
2862cf88 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
98typedef struct
99{
100 unsigned int size;
101 unsigned int time;
102} bb_info;
103DEF_VEC_O(bb_info);
104DEF_VEC_ALLOC_O(bb_info,heap);
105
106static VEC(bb_info, heap) *bb_info_vec;
107
108/* Description of split point. */
109
110struct split_point
111{
112 /* Size of the partitions. */
113 unsigned int header_time, header_size, split_time, split_size;
114
574c4ab5 115 /* SSA names that need to be passed into spit function. */
2862cf88 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
131struct split_point best_split_point;
132
b04bab7c 133static tree find_retval (basic_block return_bb);
134
2dc0a33e 135/* Callback for walk_stmt_load_store_addr_ops. If T is non-SSA automatic
2862cf88 136 variable, check it if it is present in bitmap passed via DATA. */
137
138static bool
2dc0a33e 139test_nonssa_use (gimple stmt ATTRIBUTE_UNUSED, tree t, void *data)
2862cf88 140{
141 t = get_base_address (t);
142
2dc0a33e 143 if (!t || is_gimple_reg (t))
144 return false;
145
146 if (TREE_CODE (t) == PARM_DECL
147 || (TREE_CODE (t) == VAR_DECL
2862cf88 148 && auto_var_in_fn_p (t, current_function_decl))
2dc0a33e 149 || TREE_CODE (t) == RESULT_DECL
150 || TREE_CODE (t) == LABEL_DECL)
2862cf88 151 return bitmap_bit_p ((bitmap)data, DECL_UID (t));
b04bab7c 152
2dc0a33e 153 /* For DECL_BY_REFERENCE, the return value is actually a pointer. We want
154 to pretend that the value pointed to is actual result decl. */
155 if ((TREE_CODE (t) == MEM_REF || INDIRECT_REF_P (t))
b04bab7c 156 && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME
157 && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND (t, 0))) == RESULT_DECL
158 && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)))
2dc0a33e 159 return
160 bitmap_bit_p ((bitmap)data,
161 DECL_UID (DECL_RESULT (current_function_decl)));
162
2862cf88 163 return false;
164}
165
166/* Dump split point CURRENT. */
167
168static void
169dump_split_point (FILE * file, struct split_point *current)
170{
171 fprintf (file,
172 "Split point at BB %i header time:%i header size: %i"
173 " split time: %i split size: %i\n bbs: ",
174 current->entry_bb->index, current->header_time,
175 current->header_size, current->split_time, current->split_size);
176 dump_bitmap (file, current->split_bbs);
177 fprintf (file, " SSA names to pass: ");
178 dump_bitmap (file, current->ssa_names_to_pass);
179}
180
2dc0a33e 181/* Look for all BBs in header that might lead to the split part and verify
182 that they are not defining any non-SSA var used by the split part.
4493dab3 183 Parameters are the same as for consider_split. */
184
185static bool
186verify_non_ssa_vars (struct split_point *current, bitmap non_ssa_vars,
187 basic_block return_bb)
188{
189 bitmap seen = BITMAP_ALLOC (NULL);
190 VEC (basic_block,heap) *worklist = NULL;
191 edge e;
192 edge_iterator ei;
193 bool ok = true;
2dc0a33e 194
4493dab3 195 FOR_EACH_EDGE (e, ei, current->entry_bb->preds)
196 if (e->src != ENTRY_BLOCK_PTR
197 && !bitmap_bit_p (current->split_bbs, e->src->index))
198 {
199 VEC_safe_push (basic_block, heap, worklist, e->src);
200 bitmap_set_bit (seen, e->src->index);
201 }
2dc0a33e 202
4493dab3 203 while (!VEC_empty (basic_block, worklist))
204 {
205 gimple_stmt_iterator bsi;
206 basic_block bb = VEC_pop (basic_block, worklist);
207
208 FOR_EACH_EDGE (e, ei, bb->preds)
209 if (e->src != ENTRY_BLOCK_PTR
6ef9bbe0 210 && bitmap_set_bit (seen, e->src->index))
4493dab3 211 {
212 gcc_checking_assert (!bitmap_bit_p (current->split_bbs,
213 e->src->index));
214 VEC_safe_push (basic_block, heap, worklist, e->src);
4493dab3 215 }
216 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
217 {
2dc0a33e 218 gimple stmt = gsi_stmt (bsi);
219 if (is_gimple_debug (stmt))
4493dab3 220 continue;
221 if (walk_stmt_load_store_addr_ops
2dc0a33e 222 (stmt, non_ssa_vars, test_nonssa_use, test_nonssa_use,
223 test_nonssa_use))
4493dab3 224 {
225 ok = false;
226 goto done;
227 }
2dc0a33e 228 if (gimple_code (stmt) == GIMPLE_LABEL
229 && test_nonssa_use (stmt, gimple_label_label (stmt),
230 non_ssa_vars))
231 {
232 ok = false;
233 goto done;
234 }
4493dab3 235 }
236 for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi))
237 {
238 if (walk_stmt_load_store_addr_ops
2dc0a33e 239 (gsi_stmt (bsi), non_ssa_vars, test_nonssa_use, test_nonssa_use,
240 test_nonssa_use))
4493dab3 241 {
242 ok = false;
243 goto done;
244 }
245 }
246 FOR_EACH_EDGE (e, ei, bb->succs)
247 {
248 if (e->dest != return_bb)
249 continue;
250 for (bsi = gsi_start_phis (return_bb); !gsi_end_p (bsi);
251 gsi_next (&bsi))
252 {
253 gimple stmt = gsi_stmt (bsi);
254 tree op = gimple_phi_arg_def (stmt, e->dest_idx);
255
256 if (!is_gimple_reg (gimple_phi_result (stmt)))
257 continue;
258 if (TREE_CODE (op) != SSA_NAME
259 && test_nonssa_use (stmt, op, non_ssa_vars))
260 {
261 ok = false;
262 goto done;
263 }
264 }
265 }
266 }
267done:
268 BITMAP_FREE (seen);
269 VEC_free (basic_block, heap, worklist);
270 return ok;
271}
272
2862cf88 273/* We found an split_point CURRENT. NON_SSA_VARS is bitmap of all non ssa
274 variables used and RETURN_BB is return basic block.
275 See if we can split function here. */
276
277static void
278consider_split (struct split_point *current, bitmap non_ssa_vars,
279 basic_block return_bb)
280{
281 tree parm;
282 unsigned int num_args = 0;
283 unsigned int call_overhead;
284 edge e;
285 edge_iterator ei;
6a69e813 286 gimple_stmt_iterator bsi;
287 unsigned int i;
574c4ab5 288 int incoming_freq = 0;
b04bab7c 289 tree retval;
6a69e813 290
2862cf88 291 if (dump_file && (dump_flags & TDF_DETAILS))
292 dump_split_point (dump_file, current);
293
6a69e813 294 FOR_EACH_EDGE (e, ei, current->entry_bb->preds)
295 if (!bitmap_bit_p (current->split_bbs, e->src->index))
574c4ab5 296 incoming_freq += EDGE_FREQUENCY (e);
6a69e813 297
2862cf88 298 /* Do not split when we would end up calling function anyway. */
574c4ab5 299 if (incoming_freq
2862cf88 300 >= (ENTRY_BLOCK_PTR->frequency
301 * PARAM_VALUE (PARAM_PARTIAL_INLINING_ENTRY_PROBABILITY) / 100))
302 {
303 if (dump_file && (dump_flags & TDF_DETAILS))
304 fprintf (dump_file,
574c4ab5 305 " Refused: incoming frequency is too large.\n");
2862cf88 306 return;
307 }
308
309 if (!current->header_size)
310 {
311 if (dump_file && (dump_flags & TDF_DETAILS))
312 fprintf (dump_file, " Refused: header empty\n");
2862cf88 313 return;
314 }
315
574c4ab5 316 /* Verify that PHI args on entry are either virtual or all their operands
317 incoming from header are the same. */
6a69e813 318 for (bsi = gsi_start_phis (current->entry_bb); !gsi_end_p (bsi); gsi_next (&bsi))
2862cf88 319 {
6a69e813 320 gimple stmt = gsi_stmt (bsi);
321 tree val = NULL;
322
323 if (!is_gimple_reg (gimple_phi_result (stmt)))
324 continue;
325 for (i = 0; i < gimple_phi_num_args (stmt); i++)
326 {
327 edge e = gimple_phi_arg_edge (stmt, i);
328 if (!bitmap_bit_p (current->split_bbs, e->src->index))
329 {
330 tree edge_val = gimple_phi_arg_def (stmt, i);
331 if (val && edge_val != val)
332 {
333 if (dump_file && (dump_flags & TDF_DETAILS))
334 fprintf (dump_file,
335 " Refused: entry BB has PHI with multiple variants\n");
336 return;
337 }
338 val = edge_val;
339 }
340 }
2862cf88 341 }
342
343
344 /* See what argument we will pass to the split function and compute
345 call overhead. */
346 call_overhead = eni_size_weights.call_cost;
347 for (parm = DECL_ARGUMENTS (current_function_decl); parm;
1767a056 348 parm = DECL_CHAIN (parm))
2862cf88 349 {
350 if (!is_gimple_reg (parm))
351 {
352 if (bitmap_bit_p (non_ssa_vars, DECL_UID (parm)))
353 {
354 if (dump_file && (dump_flags & TDF_DETAILS))
355 fprintf (dump_file,
356 " Refused: need to pass non-ssa param values\n");
357 return;
358 }
359 }
360 else if (gimple_default_def (cfun, parm)
361 && bitmap_bit_p (current->ssa_names_to_pass,
362 SSA_NAME_VERSION (gimple_default_def
363 (cfun, parm))))
364 {
365 if (!VOID_TYPE_P (TREE_TYPE (parm)))
366 call_overhead += estimate_move_cost (TREE_TYPE (parm));
367 num_args++;
368 }
369 }
370 if (!VOID_TYPE_P (TREE_TYPE (current_function_decl)))
371 call_overhead += estimate_move_cost (TREE_TYPE (current_function_decl));
372
373 if (current->split_size <= call_overhead)
374 {
375 if (dump_file && (dump_flags & TDF_DETAILS))
376 fprintf (dump_file,
377 " Refused: split size is smaller than call overhead\n");
378 return;
379 }
380 if (current->header_size + call_overhead
381 >= (unsigned int)(DECL_DECLARED_INLINE_P (current_function_decl)
382 ? MAX_INLINE_INSNS_SINGLE
383 : MAX_INLINE_INSNS_AUTO))
384 {
385 if (dump_file && (dump_flags & TDF_DETAILS))
386 fprintf (dump_file,
387 " Refused: header size is too large for inline candidate\n");
388 return;
389 }
390
391 /* FIXME: we currently can pass only SSA function parameters to the split
fd8d648f 392 arguments. Once parm_adjustment infrastructure is supported by cloning,
2862cf88 393 we can pass more than that. */
394 if (num_args != bitmap_count_bits (current->ssa_names_to_pass))
395 {
6a69e813 396
2862cf88 397 if (dump_file && (dump_flags & TDF_DETAILS))
398 fprintf (dump_file,
399 " Refused: need to pass non-param values\n");
400 return;
401 }
402
403 /* When there are non-ssa vars used in the split region, see if they
404 are used in the header region. If so, reject the split.
405 FIXME: we can use nested function support to access both. */
4493dab3 406 if (!bitmap_empty_p (non_ssa_vars)
407 && !verify_non_ssa_vars (current, non_ssa_vars, return_bb))
2862cf88 408 {
4493dab3 409 if (dump_file && (dump_flags & TDF_DETAILS))
410 fprintf (dump_file,
411 " Refused: split part has non-ssa uses\n");
2862cf88 412 return;
413 }
414 if (dump_file && (dump_flags & TDF_DETAILS))
415 fprintf (dump_file, " Accepted!\n");
416
b04bab7c 417 /* See if retval used by return bb is computed by header or split part.
418 When it is computed by split part, we need to produce return statement
419 in the split part and add code to header to pass it around.
420
421 This is bit tricky to test:
422 1) When there is no return_bb or no return value, we always pass
423 value around.
424 2) Invariants are always computed by caller.
425 3) For SSA we need to look if defining statement is in header or split part
426 4) For non-SSA we need to look where the var is computed. */
427 retval = find_retval (return_bb);
428 if (!retval)
429 current->split_part_set_retval = true;
430 else if (is_gimple_min_invariant (retval))
431 current->split_part_set_retval = false;
432 /* Special case is value returned by reference we record as if it was non-ssa
433 set to result_decl. */
434 else if (TREE_CODE (retval) == SSA_NAME
435 && TREE_CODE (SSA_NAME_VAR (retval)) == RESULT_DECL
436 && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)))
437 current->split_part_set_retval
438 = bitmap_bit_p (non_ssa_vars, DECL_UID (SSA_NAME_VAR (retval)));
439 else if (TREE_CODE (retval) == SSA_NAME)
440 current->split_part_set_retval
441 = (!SSA_NAME_IS_DEFAULT_DEF (retval)
442 && (bitmap_bit_p (current->split_bbs,
443 gimple_bb (SSA_NAME_DEF_STMT (retval))->index)
444 || gimple_bb (SSA_NAME_DEF_STMT (retval)) == return_bb));
445 else if (TREE_CODE (retval) == PARM_DECL)
446 current->split_part_set_retval = false;
447 else if (TREE_CODE (retval) == VAR_DECL
448 || TREE_CODE (retval) == RESULT_DECL)
449 current->split_part_set_retval
450 = bitmap_bit_p (non_ssa_vars, DECL_UID (retval));
451 else
452 current->split_part_set_retval = true;
453
2862cf88 454 /* At the moment chose split point with lowest frequency and that leaves
455 out smallest size of header.
456 In future we might re-consider this heuristics. */
457 if (!best_split_point.split_bbs
458 || best_split_point.entry_bb->frequency > current->entry_bb->frequency
459 || (best_split_point.entry_bb->frequency == current->entry_bb->frequency
460 && best_split_point.split_size < current->split_size))
461
462 {
463 if (dump_file && (dump_flags & TDF_DETAILS))
464 fprintf (dump_file, " New best split point!\n");
465 if (best_split_point.ssa_names_to_pass)
466 {
467 BITMAP_FREE (best_split_point.ssa_names_to_pass);
468 BITMAP_FREE (best_split_point.split_bbs);
469 }
470 best_split_point = *current;
471 best_split_point.ssa_names_to_pass = BITMAP_ALLOC (NULL);
472 bitmap_copy (best_split_point.ssa_names_to_pass,
473 current->ssa_names_to_pass);
474 best_split_point.split_bbs = BITMAP_ALLOC (NULL);
475 bitmap_copy (best_split_point.split_bbs, current->split_bbs);
476 }
477}
478
4493dab3 479/* Return basic block containing RETURN statement. We allow basic blocks
480 of the form:
481 <retval> = tmp_var;
482 return <retval>
483 but return_bb can not be more complex than this.
484 If nothing is found, return EXIT_BLOCK_PTR.
485
2862cf88 486 When there are multiple RETURN statement, chose one with return value,
487 since that one is more likely shared by multiple code paths.
4493dab3 488
489 Return BB is special, because for function splitting it is the only
490 basic block that is duplicated in between header and split part of the
491 function.
492
2862cf88 493 TODO: We might support multiple return blocks. */
494
495static basic_block
496find_return_bb (void)
497{
498 edge e;
2862cf88 499 basic_block return_bb = EXIT_BLOCK_PTR;
b234c94d 500 gimple_stmt_iterator bsi;
501 bool found_return = false;
502 tree retval = NULL_TREE;
2862cf88 503
b234c94d 504 if (!single_pred_p (EXIT_BLOCK_PTR))
505 return return_bb;
506
507 e = single_pred_edge (EXIT_BLOCK_PTR);
508 for (bsi = gsi_last_bb (e->src); !gsi_end_p (bsi); gsi_prev (&bsi))
509 {
510 gimple stmt = gsi_stmt (bsi);
511 if (gimple_code (stmt) == GIMPLE_LABEL || is_gimple_debug (stmt))
512 ;
513 else if (gimple_code (stmt) == GIMPLE_ASSIGN
514 && found_return
515 && gimple_assign_single_p (stmt)
516 && (auto_var_in_fn_p (gimple_assign_rhs1 (stmt),
517 current_function_decl)
518 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))
519 && retval == gimple_assign_lhs (stmt))
520 ;
521 else if (gimple_code (stmt) == GIMPLE_RETURN)
522 {
523 found_return = true;
524 retval = gimple_return_retval (stmt);
525 }
526 else
527 break;
528 }
529 if (gsi_end_p (bsi) && found_return)
530 return_bb = e->src;
2862cf88 531
2862cf88 532 return return_bb;
533}
534
574c4ab5 535/* Given return basic block RETURN_BB, see where return value is really
4493dab3 536 stored. */
537static tree
538find_retval (basic_block return_bb)
539{
540 gimple_stmt_iterator bsi;
541 for (bsi = gsi_start_bb (return_bb); !gsi_end_p (bsi); gsi_next (&bsi))
542 if (gimple_code (gsi_stmt (bsi)) == GIMPLE_RETURN)
543 return gimple_return_retval (gsi_stmt (bsi));
544 else if (gimple_code (gsi_stmt (bsi)) == GIMPLE_ASSIGN)
545 return gimple_assign_rhs1 (gsi_stmt (bsi));
546 return NULL;
547}
548
2dc0a33e 549/* Callback for walk_stmt_load_store_addr_ops. If T is non-SSA automatic
550 variable, mark it as used in bitmap passed via DATA.
2862cf88 551 Return true when access to T prevents splitting the function. */
552
553static bool
2dc0a33e 554mark_nonssa_use (gimple stmt ATTRIBUTE_UNUSED, tree t, void *data)
2862cf88 555{
556 t = get_base_address (t);
557
558 if (!t || is_gimple_reg (t))
559 return false;
560
561 /* At present we can't pass non-SSA arguments to split function.
562 FIXME: this can be relaxed by passing references to arguments. */
563 if (TREE_CODE (t) == PARM_DECL)
564 {
565 if (dump_file && (dump_flags & TDF_DETAILS))
2dc0a33e 566 fprintf (dump_file,
567 "Cannot split: use of non-ssa function parameter.\n");
2862cf88 568 return true;
569 }
570
2dc0a33e 571 if ((TREE_CODE (t) == VAR_DECL
572 && auto_var_in_fn_p (t, current_function_decl))
573 || TREE_CODE (t) == RESULT_DECL
574 || TREE_CODE (t) == LABEL_DECL)
2862cf88 575 bitmap_set_bit ((bitmap)data, DECL_UID (t));
b04bab7c 576
2dc0a33e 577 /* For DECL_BY_REFERENCE, the return value is actually a pointer. We want
578 to pretend that the value pointed to is actual result decl. */
579 if ((TREE_CODE (t) == MEM_REF || INDIRECT_REF_P (t))
b04bab7c 580 && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME
581 && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND (t, 0))) == RESULT_DECL
582 && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)))
2dc0a33e 583 return
584 bitmap_bit_p ((bitmap)data,
585 DECL_UID (DECL_RESULT (current_function_decl)));
586
2862cf88 587 return false;
588}
589
590/* Compute local properties of basic block BB we collect when looking for
591 split points. We look for ssa defs and store them in SET_SSA_NAMES,
592 for ssa uses and store them in USED_SSA_NAMES and for any non-SSA automatic
593 vars stored in NON_SSA_VARS.
594
595 When BB has edge to RETURN_BB, collect uses in RETURN_BB too.
596
597 Return false when BB contains something that prevents it from being put into
598 split function. */
599
600static bool
601visit_bb (basic_block bb, basic_block return_bb,
602 bitmap set_ssa_names, bitmap used_ssa_names,
603 bitmap non_ssa_vars)
604{
605 gimple_stmt_iterator bsi;
606 edge e;
607 edge_iterator ei;
608 bool can_split = true;
609
610 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
611 {
612 gimple stmt = gsi_stmt (bsi);
613 tree op;
614 ssa_op_iter iter;
615 tree decl;
616
617 if (is_gimple_debug (stmt))
618 continue;
619
620 /* FIXME: We can split regions containing EH. We can not however
621 split RESX, EH_DISPATCH and EH_POINTER referring to same region
622 into different partitions. This would require tracking of
623 EH regions and checking in consider_split_point if they
624 are not used elsewhere. */
625 if (gimple_code (stmt) == GIMPLE_RESX
626 && stmt_can_throw_external (stmt))
627 {
628 if (dump_file && (dump_flags & TDF_DETAILS))
2dc0a33e 629 fprintf (dump_file, "Cannot split: external resx.\n");
2862cf88 630 can_split = false;
631 }
632 if (gimple_code (stmt) == GIMPLE_EH_DISPATCH)
633 {
634 if (dump_file && (dump_flags & TDF_DETAILS))
2dc0a33e 635 fprintf (dump_file, "Cannot split: eh dispatch.\n");
2862cf88 636 can_split = false;
637 }
638
639 /* Check builtins that prevent splitting. */
640 if (gimple_code (stmt) == GIMPLE_CALL
641 && (decl = gimple_call_fndecl (stmt)) != NULL_TREE
642 && DECL_BUILT_IN (decl)
643 && DECL_BUILT_IN_CLASS (decl) == BUILT_IN_NORMAL)
644 switch (DECL_FUNCTION_CODE (decl))
645 {
646 /* FIXME: once we will allow passing non-parm values to split part,
647 we need to be sure to handle correct builtin_stack_save and
648 builtin_stack_restore. At the moment we are safe; there is no
649 way to store builtin_stack_save result in non-SSA variable
650 since all calls to those are compiler generated. */
651 case BUILT_IN_APPLY:
652 case BUILT_IN_VA_START:
653 if (dump_file && (dump_flags & TDF_DETAILS))
2dc0a33e 654 fprintf (dump_file,
655 "Cannot split: builtin_apply and va_start.\n");
2862cf88 656 can_split = false;
657 break;
658 case BUILT_IN_EH_POINTER:
659 if (dump_file && (dump_flags & TDF_DETAILS))
2dc0a33e 660 fprintf (dump_file, "Cannot split: builtin_eh_pointer.\n");
2862cf88 661 can_split = false;
662 break;
663 default:
664 break;
665 }
666
667 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
668 bitmap_set_bit (set_ssa_names, SSA_NAME_VERSION (op));
669 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
670 bitmap_set_bit (used_ssa_names, SSA_NAME_VERSION (op));
671 can_split &= !walk_stmt_load_store_addr_ops (stmt, non_ssa_vars,
672 mark_nonssa_use,
673 mark_nonssa_use,
674 mark_nonssa_use);
675 }
676 for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi))
677 {
678 gimple stmt = gsi_stmt (bsi);
6a69e813 679 unsigned int i;
2862cf88 680
681 if (is_gimple_debug (stmt))
682 continue;
683 if (!is_gimple_reg (gimple_phi_result (stmt)))
684 continue;
6a69e813 685 bitmap_set_bit (set_ssa_names,
686 SSA_NAME_VERSION (gimple_phi_result (stmt)));
687 for (i = 0; i < gimple_phi_num_args (stmt); i++)
688 {
689 tree op = gimple_phi_arg_def (stmt, i);
690 if (TREE_CODE (op) == SSA_NAME)
691 bitmap_set_bit (used_ssa_names, SSA_NAME_VERSION (op));
692 }
2862cf88 693 can_split &= !walk_stmt_load_store_addr_ops (stmt, non_ssa_vars,
694 mark_nonssa_use,
695 mark_nonssa_use,
696 mark_nonssa_use);
697 }
574c4ab5 698 /* Record also uses coming from PHI operand in return BB. */
2862cf88 699 FOR_EACH_EDGE (e, ei, bb->succs)
700 if (e->dest == return_bb)
701 {
2862cf88 702 for (bsi = gsi_start_phis (return_bb); !gsi_end_p (bsi); gsi_next (&bsi))
703 {
704 gimple stmt = gsi_stmt (bsi);
705 tree op = gimple_phi_arg_def (stmt, e->dest_idx);
706
707 if (is_gimple_debug (stmt))
708 continue;
709 if (!is_gimple_reg (gimple_phi_result (stmt)))
710 continue;
2862cf88 711 if (TREE_CODE (op) == SSA_NAME)
712 bitmap_set_bit (used_ssa_names, SSA_NAME_VERSION (op));
713 else
714 can_split &= !mark_nonssa_use (stmt, op, non_ssa_vars);
715 }
2862cf88 716 }
717 return can_split;
718}
719
720/* Stack entry for recursive DFS walk in find_split_point. */
721
722typedef struct
723{
724 /* Basic block we are examining. */
725 basic_block bb;
726
727 /* SSA names set and used by the BB and all BBs reachable
728 from it via DFS walk. */
729 bitmap set_ssa_names, used_ssa_names;
730 bitmap non_ssa_vars;
731
732 /* All BBS visited from this BB via DFS walk. */
733 bitmap bbs_visited;
734
735 /* Last examined edge in DFS walk. Since we walk unoriented graph,
574c4ab5 736 the value is up to sum of incoming and outgoing edges of BB. */
2862cf88 737 unsigned int edge_num;
738
739 /* Stack entry index of earliest BB reachable from current BB
574c4ab5 740 or any BB visited later in DFS walk. */
2862cf88 741 int earliest;
742
743 /* Overall time and size of all BBs reached from this BB in DFS walk. */
744 int overall_time, overall_size;
745
746 /* When false we can not split on this BB. */
747 bool can_split;
748} stack_entry;
749DEF_VEC_O(stack_entry);
750DEF_VEC_ALLOC_O(stack_entry,heap);
751
752
753/* Find all articulations and call consider_split on them.
754 OVERALL_TIME and OVERALL_SIZE is time and size of the function.
755
756 We perform basic algorithm for finding an articulation in a graph
757 created from CFG by considering it to be an unoriented graph.
758
759 The articulation is discovered via DFS walk. We collect earliest
760 basic block on stack that is reachable via backward edge. Articulation
761 is any basic block such that there is no backward edge bypassing it.
762 To reduce stack usage we maintain heap allocated stack in STACK vector.
763 AUX pointer of BB is set to index it appears in the stack or -1 once
764 it is visited and popped off the stack.
765
766 The algorithm finds articulation after visiting the whole component
767 reachable by it. This makes it convenient to collect information about
768 the component used by consider_split. */
769
770static void
771find_split_points (int overall_time, int overall_size)
772{
773 stack_entry first;
774 VEC(stack_entry, heap) *stack = NULL;
775 basic_block bb;
776 basic_block return_bb = find_return_bb ();
777 struct split_point current;
778
779 current.header_time = overall_time;
780 current.header_size = overall_size;
781 current.split_time = 0;
782 current.split_size = 0;
783 current.ssa_names_to_pass = BITMAP_ALLOC (NULL);
784
785 first.bb = ENTRY_BLOCK_PTR;
786 first.edge_num = 0;
787 first.overall_time = 0;
788 first.overall_size = 0;
789 first.earliest = INT_MAX;
790 first.set_ssa_names = 0;
791 first.used_ssa_names = 0;
792 first.bbs_visited = 0;
793 VEC_safe_push (stack_entry, heap, stack, &first);
794 ENTRY_BLOCK_PTR->aux = (void *)(intptr_t)-1;
795
796 while (!VEC_empty (stack_entry, stack))
797 {
798 stack_entry *entry = VEC_last (stack_entry, stack);
799
800 /* We are walking an acyclic graph, so edge_num counts
801 succ and pred edges together. However when considering
802 articulation, we want to have processed everything reachable
803 from articulation but nothing that reaches into it. */
804 if (entry->edge_num == EDGE_COUNT (entry->bb->succs)
805 && entry->bb != ENTRY_BLOCK_PTR)
806 {
807 int pos = VEC_length (stack_entry, stack);
808 entry->can_split &= visit_bb (entry->bb, return_bb,
809 entry->set_ssa_names,
810 entry->used_ssa_names,
811 entry->non_ssa_vars);
812 if (pos <= entry->earliest && !entry->can_split
813 && dump_file && (dump_flags & TDF_DETAILS))
814 fprintf (dump_file,
815 "found articulation at bb %i but can not split\n",
816 entry->bb->index);
817 if (pos <= entry->earliest && entry->can_split)
818 {
819 if (dump_file && (dump_flags & TDF_DETAILS))
820 fprintf (dump_file, "found articulation at bb %i\n",
821 entry->bb->index);
822 current.entry_bb = entry->bb;
823 current.ssa_names_to_pass = BITMAP_ALLOC (NULL);
824 bitmap_and_compl (current.ssa_names_to_pass,
825 entry->used_ssa_names, entry->set_ssa_names);
826 current.header_time = overall_time - entry->overall_time;
827 current.header_size = overall_size - entry->overall_size;
828 current.split_time = entry->overall_time;
829 current.split_size = entry->overall_size;
830 current.split_bbs = entry->bbs_visited;
831 consider_split (&current, entry->non_ssa_vars, return_bb);
832 BITMAP_FREE (current.ssa_names_to_pass);
833 }
834 }
835 /* Do actual DFS walk. */
836 if (entry->edge_num
837 < (EDGE_COUNT (entry->bb->succs)
838 + EDGE_COUNT (entry->bb->preds)))
839 {
840 edge e;
841 basic_block dest;
842 if (entry->edge_num < EDGE_COUNT (entry->bb->succs))
843 {
844 e = EDGE_SUCC (entry->bb, entry->edge_num);
845 dest = e->dest;
846 }
847 else
848 {
849 e = EDGE_PRED (entry->bb, entry->edge_num
850 - EDGE_COUNT (entry->bb->succs));
851 dest = e->src;
852 }
853
854 entry->edge_num++;
855
856 /* New BB to visit, push it to the stack. */
857 if (dest != return_bb && dest != EXIT_BLOCK_PTR
858 && !dest->aux)
859 {
860 stack_entry new_entry;
861
862 new_entry.bb = dest;
863 new_entry.edge_num = 0;
864 new_entry.overall_time
865 = VEC_index (bb_info, bb_info_vec, dest->index)->time;
866 new_entry.overall_size
867 = VEC_index (bb_info, bb_info_vec, dest->index)->size;
868 new_entry.earliest = INT_MAX;
869 new_entry.set_ssa_names = BITMAP_ALLOC (NULL);
870 new_entry.used_ssa_names = BITMAP_ALLOC (NULL);
871 new_entry.bbs_visited = BITMAP_ALLOC (NULL);
872 new_entry.non_ssa_vars = BITMAP_ALLOC (NULL);
873 new_entry.can_split = true;
874 bitmap_set_bit (new_entry.bbs_visited, dest->index);
875 VEC_safe_push (stack_entry, heap, stack, &new_entry);
876 dest->aux = (void *)(intptr_t)VEC_length (stack_entry, stack);
877 }
878 /* Back edge found, record the earliest point. */
879 else if ((intptr_t)dest->aux > 0
880 && (intptr_t)dest->aux < entry->earliest)
881 entry->earliest = (intptr_t)dest->aux;
882 }
574c4ab5 883 /* We are done with examining the edges. Pop off the value from stack
884 and merge stuff we accumulate during the walk. */
2862cf88 885 else if (entry->bb != ENTRY_BLOCK_PTR)
886 {
887 stack_entry *prev = VEC_index (stack_entry, stack,
888 VEC_length (stack_entry, stack) - 2);
889
890 entry->bb->aux = (void *)(intptr_t)-1;
891 prev->can_split &= entry->can_split;
892 if (prev->set_ssa_names)
893 {
894 bitmap_ior_into (prev->set_ssa_names, entry->set_ssa_names);
895 bitmap_ior_into (prev->used_ssa_names, entry->used_ssa_names);
896 bitmap_ior_into (prev->bbs_visited, entry->bbs_visited);
897 bitmap_ior_into (prev->non_ssa_vars, entry->non_ssa_vars);
898 }
899 if (prev->earliest > entry->earliest)
900 prev->earliest = entry->earliest;
901 prev->overall_time += entry->overall_time;
902 prev->overall_size += entry->overall_size;
903 BITMAP_FREE (entry->set_ssa_names);
904 BITMAP_FREE (entry->used_ssa_names);
905 BITMAP_FREE (entry->bbs_visited);
906 BITMAP_FREE (entry->non_ssa_vars);
907 VEC_pop (stack_entry, stack);
908 }
909 else
910 VEC_pop (stack_entry, stack);
911 }
912 ENTRY_BLOCK_PTR->aux = NULL;
913 FOR_EACH_BB (bb)
914 bb->aux = NULL;
84b53675 915 VEC_free (stack_entry, heap, stack);
2862cf88 916 BITMAP_FREE (current.ssa_names_to_pass);
917}
918
919/* Split function at SPLIT_POINT. */
920
921static void
922split_function (struct split_point *split_point)
923{
924 VEC (tree, heap) *args_to_pass = NULL;
925 bitmap args_to_skip = BITMAP_ALLOC (NULL);
926 tree parm;
927 int num = 0;
928 struct cgraph_node *node;
929 basic_block return_bb = find_return_bb ();
930 basic_block call_bb;
931 gimple_stmt_iterator gsi;
932 gimple call;
933 edge e;
934 edge_iterator ei;
935 tree retval = NULL, real_retval = NULL;
936 bool split_part_return_p = false;
937 gimple last_stmt = NULL;
e65564ab 938 bool conv_needed = false;
939 unsigned int i;
940 tree arg;
2862cf88 941
942 if (dump_file)
943 {
944 fprintf (dump_file, "\n\nSplitting function at:\n");
945 dump_split_point (dump_file, split_point);
946 }
947
948 /* Collect the parameters of new function and args_to_skip bitmap. */
949 for (parm = DECL_ARGUMENTS (current_function_decl);
1767a056 950 parm; parm = DECL_CHAIN (parm), num++)
2862cf88 951 if (!is_gimple_reg (parm)
952 || !gimple_default_def (cfun, parm)
953 || !bitmap_bit_p (split_point->ssa_names_to_pass,
954 SSA_NAME_VERSION (gimple_default_def (cfun, parm))))
955 bitmap_set_bit (args_to_skip, num);
956 else
e65564ab 957 {
958 arg = gimple_default_def (cfun, parm);
959 if (TYPE_MAIN_VARIANT (DECL_ARG_TYPE (parm))
960 != TYPE_MAIN_VARIANT (TREE_TYPE (arg)))
961 {
962 conv_needed = true;
963 arg = fold_convert (DECL_ARG_TYPE (parm), arg);
964 }
965 VEC_safe_push (tree, heap, args_to_pass, arg);
966 }
2862cf88 967
968 /* See if the split function will return. */
969 FOR_EACH_EDGE (e, ei, return_bb->preds)
970 if (bitmap_bit_p (split_point->split_bbs, e->src->index))
971 break;
972 if (e)
973 split_part_return_p = true;
974
b04bab7c 975 /* Add return block to what will become the split function.
976 We do not return; no return block is needed. */
977 if (!split_part_return_p)
978 ;
979 /* We have no return block, so nothing is needed. */
980 else if (return_bb == EXIT_BLOCK_PTR)
981 ;
982 /* When we do not want to return value, we need to construct
983 new return block with empty return statement.
984 FIXME: Once we are able to change return type, we should change function
985 to return void instead of just outputting function with undefined return
986 value. For structures this affects quality of codegen. */
987 else if (!split_point->split_part_set_retval
988 && find_retval (return_bb))
989 {
990 bool redirected = true;
991 basic_block new_return_bb = create_basic_block (NULL, 0, return_bb);
992 gimple_stmt_iterator gsi = gsi_start_bb (new_return_bb);
993 gsi_insert_after (&gsi, gimple_build_return (NULL), GSI_NEW_STMT);
994 while (redirected)
995 {
996 redirected = false;
997 FOR_EACH_EDGE (e, ei, return_bb->preds)
998 if (bitmap_bit_p (split_point->split_bbs, e->src->index))
999 {
1000 new_return_bb->count += e->count;
1001 new_return_bb->frequency += EDGE_FREQUENCY (e);
1002 redirect_edge_and_branch (e, new_return_bb);
1003 redirected = true;
1004 break;
1005 }
1006 }
1007 e = make_edge (new_return_bb, EXIT_BLOCK_PTR, 0);
1008 e->probability = REG_BR_PROB_BASE;
1009 e->count = new_return_bb->count;
1010 bitmap_set_bit (split_point->split_bbs, new_return_bb->index);
c13d6d8f 1011 }
1012 /* When we pass around the value, use existing return block. */
1013 else
1014 bitmap_set_bit (split_point->split_bbs, return_bb->index);
1015
1016 /* If RETURN_BB has virtual operand PHIs, they must be removed and the
1017 virtual operand marked for renaming as we change the CFG in a way that
1018 tree-inline is not able to compensate for.
1019
1020 Note this can happen whether or not we have a return value. If we have
1021 a return value, then RETURN_BB may have PHIs for real operands too. */
1022 if (return_bb != EXIT_BLOCK_PTR)
1023 {
b04bab7c 1024 for (gsi = gsi_start_phis (return_bb); !gsi_end_p (gsi);)
1025 {
1026 gimple stmt = gsi_stmt (gsi);
c13d6d8f 1027 if (is_gimple_reg (gimple_phi_result (stmt)))
1028 {
1029 gsi_next (&gsi);
1030 continue;
1031 }
904ad591 1032 mark_virtual_phi_result_for_renaming (stmt);
1033 remove_phi_node (&gsi, true);
b04bab7c 1034 }
1035 }
2862cf88 1036
1037 /* Now create the actual clone. */
1038 rebuild_cgraph_edges ();
1039 node = cgraph_function_versioning (cgraph_node (current_function_decl),
1040 NULL, NULL,
1041 args_to_skip,
1042 split_point->split_bbs,
4493dab3 1043 split_point->entry_bb, "part");
fd8d648f 1044 /* For usual cloning it is enough to clear builtin only when signature
1045 changes. For partial inlining we however can not expect the part
1046 of builtin implementation to have same semantic as the whole. */
1047 if (DECL_BUILT_IN (node->decl))
1048 {
1049 DECL_BUILT_IN_CLASS (node->decl) = NOT_BUILT_IN;
1050 DECL_FUNCTION_CODE (node->decl) = (enum built_in_function) 0;
1051 }
2862cf88 1052 cgraph_node_remove_callees (cgraph_node (current_function_decl));
1053 if (!split_part_return_p)
1054 TREE_THIS_VOLATILE (node->decl) = 1;
1055 if (dump_file)
1056 dump_function_to_file (node->decl, dump_file, dump_flags);
1057
1058 /* Create the basic block we place call into. It is the entry basic block
1059 split after last label. */
1060 call_bb = split_point->entry_bb;
1061 for (gsi = gsi_start_bb (call_bb); !gsi_end_p (gsi);)
1062 if (gimple_code (gsi_stmt (gsi)) == GIMPLE_LABEL)
1063 {
1064 last_stmt = gsi_stmt (gsi);
1065 gsi_next (&gsi);
1066 }
1067 else
1068 break;
1069 e = split_block (split_point->entry_bb, last_stmt);
1070 remove_edge (e);
1071
1072 /* Produce the call statement. */
1073 gsi = gsi_last_bb (call_bb);
e65564ab 1074 if (conv_needed)
1075 FOR_EACH_VEC_ELT (tree, args_to_pass, i, arg)
1076 if (!is_gimple_val (arg))
1077 {
1078 arg = force_gimple_operand_gsi (&gsi, arg, true, NULL_TREE,
1079 false, GSI_NEW_STMT);
1080 VEC_replace (tree, args_to_pass, i, arg);
1081 }
2862cf88 1082 call = gimple_build_call_vec (node->decl, args_to_pass);
1083 gimple_set_block (call, DECL_INITIAL (current_function_decl));
1084
a8005893 1085 /* We avoid address being taken on any variable used by split part,
1086 so return slot optimization is always possible. Moreover this is
1087 required to make DECL_BY_REFERENCE work. */
1088 if (aggregate_value_p (DECL_RESULT (current_function_decl),
1089 TREE_TYPE (current_function_decl)))
1090 gimple_call_set_return_slot_opt (call, true);
1091
2862cf88 1092 /* Update return value. This is bit tricky. When we do not return,
1093 do nothing. When we return we might need to update return_bb
1094 or produce a new return statement. */
1095 if (!split_part_return_p)
1096 gsi_insert_after (&gsi, call, GSI_NEW_STMT);
1097 else
1098 {
1099 e = make_edge (call_bb, return_bb,
1100 return_bb == EXIT_BLOCK_PTR ? 0 : EDGE_FALLTHRU);
1101 e->count = call_bb->count;
1102 e->probability = REG_BR_PROB_BASE;
524a0531 1103
1104 /* If there is return basic block, see what value we need to store
1105 return value into and put call just before it. */
2862cf88 1106 if (return_bb != EXIT_BLOCK_PTR)
1107 {
4493dab3 1108 real_retval = retval = find_retval (return_bb);
524a0531 1109
b04bab7c 1110 if (real_retval && split_point->split_part_set_retval)
2862cf88 1111 {
1112 gimple_stmt_iterator psi;
1113
524a0531 1114 /* See if we need new SSA_NAME for the result.
1115 When DECL_BY_REFERENCE is true, retval is actually pointer to
1116 return value and it is constant in whole function. */
1117 if (TREE_CODE (retval) == SSA_NAME
1118 && !DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)))
2862cf88 1119 {
1120 retval = make_ssa_name (SSA_NAME_VAR (retval), call);
524a0531 1121
1122 /* See if there is PHI defining return value. */
1123 for (psi = gsi_start_phis (return_bb);
1124 !gsi_end_p (psi); gsi_next (&psi))
1125 if (is_gimple_reg (gimple_phi_result (gsi_stmt (psi))))
1126 break;
1127
1128 /* When there is PHI, just update its value. */
2862cf88 1129 if (TREE_CODE (retval) == SSA_NAME
1130 && !gsi_end_p (psi))
1131 add_phi_arg (gsi_stmt (psi), retval, e, UNKNOWN_LOCATION);
524a0531 1132 /* Otherwise update the return BB itself.
1133 find_return_bb allows at most one assignment to return value,
1134 so update first statement. */
1135 else
2862cf88 1136 {
4493dab3 1137 gimple_stmt_iterator bsi;
1138 for (bsi = gsi_start_bb (return_bb); !gsi_end_p (bsi);
1139 gsi_next (&bsi))
1140 if (gimple_code (gsi_stmt (bsi)) == GIMPLE_RETURN)
1141 {
1142 gimple_return_set_retval (gsi_stmt (bsi), retval);
1143 break;
1144 }
1145 else if (gimple_code (gsi_stmt (bsi)) == GIMPLE_ASSIGN)
1146 {
1147 gimple_assign_set_rhs1 (gsi_stmt (bsi), retval);
1148 break;
1149 }
1150 update_stmt (gsi_stmt (bsi));
2862cf88 1151 }
1152 }
a8005893 1153 if (DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)))
1154 gimple_call_set_lhs (call, build_simple_mem_ref (retval));
1155 else
1156 gimple_call_set_lhs (call, retval);
2862cf88 1157 }
1158 gsi_insert_after (&gsi, call, GSI_NEW_STMT);
1159 }
524a0531 1160 /* We don't use return block (there is either no return in function or
1161 multiple of them). So create new basic block with return statement.
1162 */
2862cf88 1163 else
1164 {
1165 gimple ret;
b04bab7c 1166 if (split_point->split_part_set_retval
1167 && !VOID_TYPE_P (TREE_TYPE (TREE_TYPE (current_function_decl))))
2862cf88 1168 {
1b268210 1169 retval = DECL_RESULT (current_function_decl);
9eb0b8ee 1170
1171 /* We use temporary register to hold value when aggregate_value_p
1172 is false. Similarly for DECL_BY_REFERENCE we must avoid extra
1173 copy. */
1174 if (!aggregate_value_p (retval, TREE_TYPE (current_function_decl))
1175 && !DECL_BY_REFERENCE (retval))
1176 retval = create_tmp_reg (TREE_TYPE (retval), NULL);
2862cf88 1177 if (is_gimple_reg (retval))
524a0531 1178 {
1179 /* When returning by reference, there is only one SSA name
1180 assigned to RESULT_DECL (that is pointer to return value).
1181 Look it up or create new one if it is missing. */
1182 if (DECL_BY_REFERENCE (retval))
1183 {
1184 tree retval_name;
1185 if ((retval_name = gimple_default_def (cfun, retval))
1186 != NULL)
1187 retval = retval_name;
1188 else
1189 {
1190 retval_name = make_ssa_name (retval,
1191 gimple_build_nop ());
1192 set_default_def (retval, retval_name);
1193 retval = retval_name;
1194 }
1195 }
1196 /* Otherwise produce new SSA name for return value. */
1197 else
1198 retval = make_ssa_name (retval, call);
1199 }
a8005893 1200 if (DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)))
1201 gimple_call_set_lhs (call, build_simple_mem_ref (retval));
1202 else
1203 gimple_call_set_lhs (call, retval);
2862cf88 1204 }
1205 gsi_insert_after (&gsi, call, GSI_NEW_STMT);
1206 ret = gimple_build_return (retval);
1207 gsi_insert_after (&gsi, ret, GSI_NEW_STMT);
1208 }
1209 }
1210 free_dominance_info (CDI_DOMINATORS);
1211 free_dominance_info (CDI_POST_DOMINATORS);
1212 compute_inline_parameters (node);
1213}
1214
1215/* Execute function splitting pass. */
1216
1217static unsigned int
1218execute_split_functions (void)
1219{
1220 gimple_stmt_iterator bsi;
1221 basic_block bb;
1222 int overall_time = 0, overall_size = 0;
1223 int todo = 0;
1224 struct cgraph_node *node = cgraph_node (current_function_decl);
1225
1226 if (flags_from_decl_or_type (current_function_decl) & ECF_NORETURN)
1227 {
1228 if (dump_file)
1229 fprintf (dump_file, "Not splitting: noreturn function.\n");
1230 return 0;
1231 }
1232 if (MAIN_NAME_P (DECL_NAME (current_function_decl)))
1233 {
1234 if (dump_file)
1235 fprintf (dump_file, "Not splitting: main function.\n");
1236 return 0;
1237 }
1238 /* This can be relaxed; function might become inlinable after splitting
1239 away the uninlinable part. */
1240 if (!node->local.inlinable)
1241 {
1242 if (dump_file)
1243 fprintf (dump_file, "Not splitting: not inlinable.\n");
1244 return 0;
1245 }
1246 if (node->local.disregard_inline_limits)
1247 {
1248 if (dump_file)
574c4ab5 1249 fprintf (dump_file, "Not splitting: disregarding inline limits.\n");
2862cf88 1250 return 0;
1251 }
1252 /* This can be relaxed; most of versioning tests actually prevents
1253 a duplication. */
1254 if (!tree_versionable_function_p (current_function_decl))
1255 {
1256 if (dump_file)
1257 fprintf (dump_file, "Not splitting: not versionable.\n");
1258 return 0;
1259 }
1260 /* FIXME: we could support this. */
1261 if (DECL_STRUCT_FUNCTION (current_function_decl)->static_chain_decl)
1262 {
1263 if (dump_file)
1264 fprintf (dump_file, "Not splitting: nested function.\n");
1265 return 0;
1266 }
2862cf88 1267
1268 /* See if it makes sense to try to split.
1269 It makes sense to split if we inline, that is if we have direct calls to
1270 handle or direct calls are possibly going to appear as result of indirect
a15d5ede 1271 inlining or LTO. Also handle -fprofile-generate as LTO to allow non-LTO
1272 training for LTO -fprofile-use build.
1273
2862cf88 1274 Note that we are not completely conservative about disqualifying functions
1275 called once. It is possible that the caller is called more then once and
1276 then inlining would still benefit. */
1277 if ((!node->callers || !node->callers->next_caller)
1278 && !node->address_taken
cbcf2791 1279 && (!flag_lto || !node->local.externally_visible))
2862cf88 1280 {
1281 if (dump_file)
1282 fprintf (dump_file, "Not splitting: not called directly "
1283 "or called once.\n");
1284 return 0;
1285 }
1286
1287 /* FIXME: We can actually split if splitting reduces call overhead. */
1288 if (!flag_inline_small_functions
1289 && !DECL_DECLARED_INLINE_P (current_function_decl))
1290 {
1291 if (dump_file)
1292 fprintf (dump_file, "Not splitting: not autoinlining and function"
1293 " is not inline.\n");
1294 return 0;
1295 }
1296
1297 /* Compute local info about basic blocks and determine function size/time. */
1298 VEC_safe_grow_cleared (bb_info, heap, bb_info_vec, last_basic_block + 1);
1299 memset (&best_split_point, 0, sizeof (best_split_point));
1300 FOR_EACH_BB (bb)
1301 {
1302 int time = 0;
1303 int size = 0;
1304 int freq = compute_call_stmt_bb_frequency (current_function_decl, bb);
1305
1306 if (dump_file && (dump_flags & TDF_DETAILS))
1307 fprintf (dump_file, "Basic block %i\n", bb->index);
1308
1309 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
1310 {
1311 int this_time, this_size;
1312 gimple stmt = gsi_stmt (bsi);
1313
1314 this_size = estimate_num_insns (stmt, &eni_size_weights);
1315 this_time = estimate_num_insns (stmt, &eni_time_weights) * freq;
1316 size += this_size;
1317 time += this_time;
1318
1319 if (dump_file && (dump_flags & TDF_DETAILS))
1320 {
1321 fprintf (dump_file, " freq:%6i size:%3i time:%3i ",
1322 freq, this_size, this_time);
1323 print_gimple_stmt (dump_file, stmt, 0, 0);
1324 }
1325 }
1326 overall_time += time;
1327 overall_size += size;
1328 VEC_index (bb_info, bb_info_vec, bb->index)->time = time;
1329 VEC_index (bb_info, bb_info_vec, bb->index)->size = size;
1330 }
1331 find_split_points (overall_time, overall_size);
1332 if (best_split_point.split_bbs)
1333 {
1334 split_function (&best_split_point);
1335 BITMAP_FREE (best_split_point.ssa_names_to_pass);
1336 BITMAP_FREE (best_split_point.split_bbs);
1337 todo = TODO_update_ssa | TODO_cleanup_cfg;
1338 }
1339 VEC_free (bb_info, heap, bb_info_vec);
1340 bb_info_vec = NULL;
1341 return todo;
1342}
1343
a15d5ede 1344/* Gate function splitting pass. When doing profile feedback, we want
1345 to execute the pass after profiling is read. So disable one in
1346 early optimization. */
1347
2862cf88 1348static bool
1349gate_split_functions (void)
1350{
a15d5ede 1351 return (flag_partial_inlining
1352 && !profile_arc_flag && !flag_branch_probabilities);
2862cf88 1353}
1354
1355struct gimple_opt_pass pass_split_functions =
1356{
1357 {
1358 GIMPLE_PASS,
1359 "fnsplit", /* name */
1360 gate_split_functions, /* gate */
1361 execute_split_functions, /* execute */
1362 NULL, /* sub */
1363 NULL, /* next */
1364 0, /* static_pass_number */
1365 TV_IPA_FNSPLIT, /* tv_id */
1366 PROP_cfg, /* properties_required */
1367 0, /* properties_provided */
1368 0, /* properties_destroyed */
1369 0, /* todo_flags_start */
1370 TODO_dump_func /* todo_flags_finish */
1371 }
1372};
a15d5ede 1373
1374/* Gate feedback driven function splitting pass.
1375 We don't need to split when profiling at all, we are producing
1376 lousy code anyway. */
1377
1378static bool
1379gate_feedback_split_functions (void)
1380{
1381 return (flag_partial_inlining
1382 && flag_branch_probabilities);
1383}
1384
1385/* Execute function splitting pass. */
1386
1387static unsigned int
1388execute_feedback_split_functions (void)
1389{
1390 unsigned int retval = execute_split_functions ();
1391 if (retval)
1392 retval |= TODO_rebuild_cgraph_edges;
1393 return retval;
1394}
1395
1396struct gimple_opt_pass pass_feedback_split_functions =
1397{
1398 {
1399 GIMPLE_PASS,
1400 "feedback_fnsplit", /* name */
1401 gate_feedback_split_functions, /* gate */
1402 execute_feedback_split_functions, /* execute */
1403 NULL, /* sub */
1404 NULL, /* next */
1405 0, /* static_pass_number */
1406 TV_IPA_FNSPLIT, /* tv_id */
1407 PROP_cfg, /* properties_required */
1408 0, /* properties_provided */
1409 0, /* properties_destroyed */
1410 0, /* todo_flags_start */
1411 TODO_dump_func /* todo_flags_finish */
1412 }
1413};