1 /* Code sinking for trees
2 Copyright (C) 2001-2022 Free Software Foundation, Inc.
3 Contributed by Daniel Berlin <dan@dberlin.org>
4
5 This file is part of GCC.
6
7 GCC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 3, or (at your option)
10 any later version.
11
12 GCC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
16
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
20
21 #include "config.h"
22 #include "system.h"
23 #include "coretypes.h"
24 #include "backend.h"
25 #include "tree.h"
26 #include "gimple.h"
27 #include "cfghooks.h"
28 #include "tree-pass.h"
29 #include "ssa.h"
30 #include "gimple-pretty-print.h"
31 #include "fold-const.h"
32 #include "stor-layout.h"
33 #include "cfganal.h"
34 #include "gimple-iterator.h"
35 #include "tree-cfg.h"
36 #include "cfgloop.h"
37 #include "tree-eh.h"
38
39 /* TODO:
40 1. Sinking store only using scalar promotion (IE without moving the RHS):
41
42 *q = p;
43 p = p + 1;
44 if (something)
45 *q = <not p>;
46 else
47 y = *q;
48
49
50 should become
51 sinktemp = p;
52 p = p + 1;
53 if (something)
54 *q = <not p>;
55 else
56 {
57 *q = sinktemp;
58 y = *q
59 }
60 Store copy propagation will take care of the store elimination above.
61
62
63 2. Sinking using Partial Dead Code Elimination. */
64
65
66 static struct
67 {
68 /* The number of statements sunk down the flowgraph by code sinking. */
69 int sunk;
70
71 /* The number of stores commoned and sunk down by store commoning. */
72 int commoned;
73 } sink_stats;
74
75
76 /* Given a PHI, and one of its arguments (DEF), find the edge for
77 that argument and return it. If the argument occurs twice in the PHI node,
78 we return NULL. */
79
80 static basic_block
find_bb_for_arg(gphi * phi,tree def)81 find_bb_for_arg (gphi *phi, tree def)
82 {
83 size_t i;
84 bool foundone = false;
85 basic_block result = NULL;
86 for (i = 0; i < gimple_phi_num_args (phi); i++)
87 if (PHI_ARG_DEF (phi, i) == def)
88 {
89 if (foundone)
90 return NULL;
91 foundone = true;
92 result = gimple_phi_arg_edge (phi, i)->src;
93 }
94 return result;
95 }
96
97 /* When the first immediate use is in a statement, then return true if all
98 immediate uses in IMM are in the same statement.
99 We could also do the case where the first immediate use is in a phi node,
100 and all the other uses are in phis in the same basic block, but this
101 requires some expensive checking later (you have to make sure no def/vdef
102 in the statement occurs for multiple edges in the various phi nodes it's
103 used in, so that you only have one place you can sink it to. */
104
105 static bool
all_immediate_uses_same_place(def_operand_p def_p)106 all_immediate_uses_same_place (def_operand_p def_p)
107 {
108 tree var = DEF_FROM_PTR (def_p);
109 imm_use_iterator imm_iter;
110 use_operand_p use_p;
111
112 gimple *firstuse = NULL;
113 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, var)
114 {
115 if (is_gimple_debug (USE_STMT (use_p)))
116 continue;
117 if (firstuse == NULL)
118 firstuse = USE_STMT (use_p);
119 else
120 if (firstuse != USE_STMT (use_p))
121 return false;
122 }
123
124 return true;
125 }
126
127 /* Find the nearest common dominator of all of the immediate uses in IMM. */
128
129 static basic_block
nearest_common_dominator_of_uses(def_operand_p def_p,bool * debug_stmts)130 nearest_common_dominator_of_uses (def_operand_p def_p, bool *debug_stmts)
131 {
132 tree var = DEF_FROM_PTR (def_p);
133 auto_bitmap blocks;
134 basic_block commondom;
135 unsigned int j;
136 bitmap_iterator bi;
137 imm_use_iterator imm_iter;
138 use_operand_p use_p;
139
140 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, var)
141 {
142 gimple *usestmt = USE_STMT (use_p);
143 basic_block useblock;
144
145 if (gphi *phi = dyn_cast <gphi *> (usestmt))
146 {
147 int idx = PHI_ARG_INDEX_FROM_USE (use_p);
148
149 useblock = gimple_phi_arg_edge (phi, idx)->src;
150 }
151 else if (is_gimple_debug (usestmt))
152 {
153 *debug_stmts = true;
154 continue;
155 }
156 else
157 {
158 useblock = gimple_bb (usestmt);
159 }
160
161 /* Short circuit. Nothing dominates the entry block. */
162 if (useblock == ENTRY_BLOCK_PTR_FOR_FN (cfun))
163 return NULL;
164
165 bitmap_set_bit (blocks, useblock->index);
166 }
167 commondom = BASIC_BLOCK_FOR_FN (cfun, bitmap_first_set_bit (blocks));
168 EXECUTE_IF_SET_IN_BITMAP (blocks, 0, j, bi)
169 commondom = nearest_common_dominator (CDI_DOMINATORS, commondom,
170 BASIC_BLOCK_FOR_FN (cfun, j));
171 return commondom;
172 }
173
174 /* Given EARLY_BB and LATE_BB, two blocks in a path through the dominator
175 tree, return the best basic block between them (inclusive) to place
176 statements.
177
178 We want the most control dependent block in the shallowest loop nest.
179
180 If the resulting block is in a shallower loop nest, then use it. Else
181 only use the resulting block if it has significantly lower execution
182 frequency than EARLY_BB to avoid gratuitous statement movement. We
183 consider statements with VOPS more desirable to move.
184
185 This pass would obviously benefit from PDO as it utilizes block
186 frequencies. It would also benefit from recomputing frequencies
187 if profile data is not available since frequencies often get out
188 of sync with reality. */
189
190 static basic_block
select_best_block(basic_block early_bb,basic_block late_bb,gimple * stmt)191 select_best_block (basic_block early_bb,
192 basic_block late_bb,
193 gimple *stmt)
194 {
195 basic_block best_bb = late_bb;
196 basic_block temp_bb = late_bb;
197 int threshold;
198
199 while (temp_bb != early_bb)
200 {
201 /* If we've moved into a lower loop nest, then that becomes
202 our best block. */
203 if (bb_loop_depth (temp_bb) < bb_loop_depth (best_bb))
204 best_bb = temp_bb;
205
206 /* Walk up the dominator tree, hopefully we'll find a shallower
207 loop nest. */
208 temp_bb = get_immediate_dominator (CDI_DOMINATORS, temp_bb);
209 }
210
211 /* If we found a shallower loop nest, then we always consider that
212 a win. This will always give us the most control dependent block
213 within that loop nest. */
214 if (bb_loop_depth (best_bb) < bb_loop_depth (early_bb))
215 return best_bb;
216
217 /* Get the sinking threshold. If the statement to be moved has memory
218 operands, then increase the threshold by 7% as those are even more
219 profitable to avoid, clamping at 100%. */
220 threshold = param_sink_frequency_threshold;
221 if (gimple_vuse (stmt) || gimple_vdef (stmt))
222 {
223 threshold += 7;
224 if (threshold > 100)
225 threshold = 100;
226 }
227
228 /* If BEST_BB is at the same nesting level, then require it to have
229 significantly lower execution frequency to avoid gratuitous movement. */
230 if (bb_loop_depth (best_bb) == bb_loop_depth (early_bb)
231 /* If result of comparsion is unknown, prefer EARLY_BB.
232 Thus use !(...>=..) rather than (...<...) */
233 && !(best_bb->count.apply_scale (100, 1)
234 >= early_bb->count.apply_scale (threshold, 1)))
235 return best_bb;
236
237 /* No better block found, so return EARLY_BB, which happens to be the
238 statement's original block. */
239 return early_bb;
240 }
241
242 /* Given a statement (STMT) and the basic block it is currently in (FROMBB),
243 determine the location to sink the statement to, if any.
244 Returns true if there is such location; in that case, TOGSI points to the
245 statement before that STMT should be moved. */
246
247 static bool
statement_sink_location(gimple * stmt,basic_block frombb,gimple_stmt_iterator * togsi,bool * zero_uses_p)248 statement_sink_location (gimple *stmt, basic_block frombb,
249 gimple_stmt_iterator *togsi, bool *zero_uses_p)
250 {
251 gimple *use;
252 use_operand_p one_use = NULL_USE_OPERAND_P;
253 basic_block sinkbb;
254 use_operand_p use_p;
255 def_operand_p def_p;
256 ssa_op_iter iter;
257 imm_use_iterator imm_iter;
258
259 *zero_uses_p = false;
260
261 /* We only can sink assignments and non-looping const/pure calls. */
262 int cf;
263 if (!is_gimple_assign (stmt)
264 && (!is_gimple_call (stmt)
265 || !((cf = gimple_call_flags (stmt)) & (ECF_CONST|ECF_PURE))
266 || (cf & ECF_LOOPING_CONST_OR_PURE)))
267 return false;
268
269 /* We only can sink stmts with a single definition. */
270 def_p = single_ssa_def_operand (stmt, SSA_OP_ALL_DEFS);
271 if (def_p == NULL_DEF_OPERAND_P)
272 return false;
273
274 /* There are a few classes of things we can't or don't move, some because we
275 don't have code to handle it, some because it's not profitable and some
276 because it's not legal.
277
278 We can't sink things that may be global stores, at least not without
279 calculating a lot more information, because we may cause it to no longer
280 be seen by an external routine that needs it depending on where it gets
281 moved to.
282
283 We can't sink statements that end basic blocks without splitting the
284 incoming edge for the sink location to place it there.
285
286 We can't sink statements that have volatile operands.
287
288 We don't want to sink dead code, so anything with 0 immediate uses is not
289 sunk.
290
291 Don't sink BLKmode assignments if current function has any local explicit
292 register variables, as BLKmode assignments may involve memcpy or memset
293 calls or, on some targets, inline expansion thereof that sometimes need
294 to use specific hard registers.
295
296 */
297 if (stmt_ends_bb_p (stmt)
298 || gimple_has_side_effects (stmt)
299 || (cfun->has_local_explicit_reg_vars
300 && TYPE_MODE (TREE_TYPE (gimple_get_lhs (stmt))) == BLKmode))
301 return false;
302
303 /* Return if there are no immediate uses of this stmt. */
304 if (has_zero_uses (DEF_FROM_PTR (def_p)))
305 {
306 *zero_uses_p = true;
307 return false;
308 }
309
310 if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (DEF_FROM_PTR (def_p)))
311 return false;
312
313 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_ALL_USES)
314 {
315 tree use = USE_FROM_PTR (use_p);
316 if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (use))
317 return false;
318 }
319
320 use = NULL;
321
322 /* If stmt is a store the one and only use needs to be the VOP
323 merging PHI node. */
324 if (virtual_operand_p (DEF_FROM_PTR (def_p)))
325 {
326 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p))
327 {
328 gimple *use_stmt = USE_STMT (use_p);
329
330 /* A killing definition is not a use. */
331 if ((gimple_has_lhs (use_stmt)
332 && operand_equal_p (gimple_get_lhs (stmt),
333 gimple_get_lhs (use_stmt), 0))
334 || stmt_kills_ref_p (use_stmt, gimple_get_lhs (stmt)))
335 {
336 /* If use_stmt is or might be a nop assignment then USE_STMT
337 acts as a use as well as definition. */
338 if (stmt != use_stmt
339 && ref_maybe_used_by_stmt_p (use_stmt,
340 gimple_get_lhs (stmt)))
341 return false;
342 continue;
343 }
344
345 if (gimple_code (use_stmt) != GIMPLE_PHI)
346 return false;
347
348 if (use
349 && use != use_stmt)
350 return false;
351
352 use = use_stmt;
353 }
354 if (!use)
355 return false;
356 }
357 /* If all the immediate uses are not in the same place, find the nearest
358 common dominator of all the immediate uses. For PHI nodes, we have to
359 find the nearest common dominator of all of the predecessor blocks, since
360 that is where insertion would have to take place. */
361 else if (gimple_vuse (stmt)
362 || !all_immediate_uses_same_place (def_p))
363 {
364 bool debug_stmts = false;
365 basic_block commondom = nearest_common_dominator_of_uses (def_p,
366 &debug_stmts);
367
368 if (commondom == frombb)
369 return false;
370
371 /* If this is a load then do not sink past any stores.
372 Look for virtual definitions in the path from frombb to the sink
373 location computed from the real uses and if found, adjust
374 that it a common dominator. */
375 if (gimple_vuse (stmt))
376 {
377 /* Do not sink loads from hard registers. */
378 if (gimple_assign_single_p (stmt)
379 && TREE_CODE (gimple_assign_rhs1 (stmt)) == VAR_DECL
380 && DECL_HARD_REGISTER (gimple_assign_rhs1 (stmt)))
381 return false;
382
383 imm_use_iterator imm_iter;
384 use_operand_p use_p;
385 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, gimple_vuse (stmt))
386 {
387 gimple *use_stmt = USE_STMT (use_p);
388 basic_block bb = gimple_bb (use_stmt);
389 /* For PHI nodes the block we know sth about is the incoming block
390 with the use. */
391 if (gimple_code (use_stmt) == GIMPLE_PHI)
392 {
393 /* If the PHI defines the virtual operand, ignore it. */
394 if (gimple_phi_result (use_stmt) == gimple_vuse (stmt))
395 continue;
396 /* In case the PHI node post-dominates the current insert
397 location we can disregard it. But make sure it is not
398 dominating it as well as can happen in a CFG cycle. */
399 if (commondom != bb
400 && !dominated_by_p (CDI_DOMINATORS, commondom, bb)
401 && dominated_by_p (CDI_POST_DOMINATORS, commondom, bb)
402 /* If the blocks are possibly within the same irreducible
403 cycle the above check breaks down. */
404 && !((bb->flags & commondom->flags & BB_IRREDUCIBLE_LOOP)
405 && bb->loop_father == commondom->loop_father)
406 && !((commondom->flags & BB_IRREDUCIBLE_LOOP)
407 && flow_loop_nested_p (commondom->loop_father,
408 bb->loop_father))
409 && !((bb->flags & BB_IRREDUCIBLE_LOOP)
410 && flow_loop_nested_p (bb->loop_father,
411 commondom->loop_father)))
412 continue;
413 bb = EDGE_PRED (bb, PHI_ARG_INDEX_FROM_USE (use_p))->src;
414 }
415 else if (!gimple_vdef (use_stmt))
416 continue;
417 /* If the use is not dominated by the path entry it is not on
418 the path. */
419 if (!dominated_by_p (CDI_DOMINATORS, bb, frombb))
420 continue;
421 /* There is no easy way to disregard defs not on the path from
422 frombb to commondom so just consider them all. */
423 commondom = nearest_common_dominator (CDI_DOMINATORS,
424 bb, commondom);
425 if (commondom == frombb)
426 return false;
427 }
428 }
429
430 /* Our common dominator has to be dominated by frombb in order to be a
431 trivially safe place to put this statement, since it has multiple
432 uses. */
433 if (!dominated_by_p (CDI_DOMINATORS, commondom, frombb))
434 return false;
435
436 commondom = select_best_block (frombb, commondom, stmt);
437
438 if (commondom == frombb)
439 return false;
440
441 *togsi = gsi_after_labels (commondom);
442
443 return true;
444 }
445 else
446 {
447 FOR_EACH_IMM_USE_FAST (one_use, imm_iter, DEF_FROM_PTR (def_p))
448 {
449 if (is_gimple_debug (USE_STMT (one_use)))
450 continue;
451 break;
452 }
453 use = USE_STMT (one_use);
454
455 if (gimple_code (use) != GIMPLE_PHI)
456 {
457 sinkbb = select_best_block (frombb, gimple_bb (use), stmt);
458
459 if (sinkbb == frombb)
460 return false;
461
462 if (sinkbb == gimple_bb (use))
463 *togsi = gsi_for_stmt (use);
464 else
465 *togsi = gsi_after_labels (sinkbb);
466
467 return true;
468 }
469 }
470
471 sinkbb = find_bb_for_arg (as_a <gphi *> (use), DEF_FROM_PTR (def_p));
472
473 /* This can happen if there are multiple uses in a PHI. */
474 if (!sinkbb)
475 return false;
476
477 sinkbb = select_best_block (frombb, sinkbb, stmt);
478 if (!sinkbb || sinkbb == frombb)
479 return false;
480
481 /* If the latch block is empty, don't make it non-empty by sinking
482 something into it. */
483 if (sinkbb == frombb->loop_father->latch
484 && empty_block_p (sinkbb))
485 return false;
486
487 *togsi = gsi_after_labels (sinkbb);
488
489 return true;
490 }
491
492 /* Very simplistic code to sink common stores from the predecessor through
493 our virtual PHI. We do this before sinking stmts from BB as it might
494 expose sinking opportunities of the merged stores.
495 Once we have partial dead code elimination through sth like SSU-PRE this
496 should be moved there. */
497
498 static unsigned
sink_common_stores_to_bb(basic_block bb)499 sink_common_stores_to_bb (basic_block bb)
500 {
501 unsigned todo = 0;
502 gphi *phi;
503
504 if (EDGE_COUNT (bb->preds) > 1
505 && (phi = get_virtual_phi (bb)))
506 {
507 /* Repeat until no more common stores are found. */
508 while (1)
509 {
510 gimple *first_store = NULL;
511 auto_vec <tree, 5> vdefs;
512 gimple_stmt_iterator gsi;
513
514 /* Search for common stores defined by all virtual PHI args.
515 ??? Common stores not present in all predecessors could
516 be handled by inserting a forwarder to sink to. Generally
517 this involves deciding which stores to do this for if
518 multiple common stores are present for different sets of
519 predecessors. See PR11832 for an interesting case. */
520 for (unsigned i = 0; i < gimple_phi_num_args (phi); ++i)
521 {
522 tree arg = gimple_phi_arg_def (phi, i);
523 gimple *def = SSA_NAME_DEF_STMT (arg);
524 if (! is_gimple_assign (def)
525 || stmt_can_throw_internal (cfun, def)
526 || (gimple_phi_arg_edge (phi, i)->flags & EDGE_ABNORMAL))
527 {
528 /* ??? We could handle some cascading with the def being
529 another PHI. We'd have to insert multiple PHIs for
530 the rhs then though (if they are not all equal). */
531 first_store = NULL;
532 break;
533 }
534 /* ??? Do not try to do anything fancy with aliasing, thus
535 do not sink across non-aliased loads (or even stores,
536 so different store order will make the sinking fail). */
537 bool all_uses_on_phi = true;
538 imm_use_iterator iter;
539 use_operand_p use_p;
540 FOR_EACH_IMM_USE_FAST (use_p, iter, arg)
541 if (USE_STMT (use_p) != phi)
542 {
543 all_uses_on_phi = false;
544 break;
545 }
546 if (! all_uses_on_phi)
547 {
548 first_store = NULL;
549 break;
550 }
551 /* Check all stores are to the same LHS. */
552 if (! first_store)
553 first_store = def;
554 /* ??? We could handle differing SSA uses in the LHS by inserting
555 PHIs for them. */
556 else if (! operand_equal_p (gimple_assign_lhs (first_store),
557 gimple_assign_lhs (def), 0)
558 || (gimple_clobber_p (first_store)
559 != gimple_clobber_p (def)))
560 {
561 first_store = NULL;
562 break;
563 }
564 vdefs.safe_push (arg);
565 }
566 if (! first_store)
567 break;
568
569 /* Check if we need a PHI node to merge the stored values. */
570 bool allsame = true;
571 if (!gimple_clobber_p (first_store))
572 for (unsigned i = 1; i < vdefs.length (); ++i)
573 {
574 gimple *def = SSA_NAME_DEF_STMT (vdefs[i]);
575 if (! operand_equal_p (gimple_assign_rhs1 (first_store),
576 gimple_assign_rhs1 (def), 0))
577 {
578 allsame = false;
579 break;
580 }
581 }
582
583 /* We cannot handle aggregate values if we need to merge them. */
584 tree type = TREE_TYPE (gimple_assign_lhs (first_store));
585 if (! allsame
586 && ! is_gimple_reg_type (type))
587 break;
588
589 if (dump_enabled_p ())
590 {
591 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS,
592 first_store,
593 "sinking common stores %sto ",
594 allsame ? "with same value " : "");
595 dump_generic_expr (MSG_OPTIMIZED_LOCATIONS, TDF_SLIM,
596 gimple_assign_lhs (first_store));
597 dump_printf (MSG_OPTIMIZED_LOCATIONS, "\n");
598 }
599
600 /* Insert a PHI to merge differing stored values if necessary.
601 Note that in general inserting PHIs isn't a very good idea as
602 it makes the job of coalescing and register allocation harder.
603 Even common SSA uses on the rhs/lhs might extend their lifetime
604 across multiple edges by this code motion which makes
605 register allocation harder. */
606 tree from;
607 if (! allsame)
608 {
609 from = make_ssa_name (type);
610 gphi *newphi = create_phi_node (from, bb);
611 for (unsigned i = 0; i < vdefs.length (); ++i)
612 {
613 gimple *def = SSA_NAME_DEF_STMT (vdefs[i]);
614 add_phi_arg (newphi, gimple_assign_rhs1 (def),
615 EDGE_PRED (bb, i), UNKNOWN_LOCATION);
616 }
617 }
618 else
619 from = gimple_assign_rhs1 (first_store);
620
621 /* Remove all stores. */
622 for (unsigned i = 0; i < vdefs.length (); ++i)
623 TREE_VISITED (vdefs[i]) = 1;
624 for (unsigned i = 0; i < vdefs.length (); ++i)
625 /* If we have more than one use of a VDEF on the PHI make sure
626 we remove the defining stmt only once. */
627 if (TREE_VISITED (vdefs[i]))
628 {
629 TREE_VISITED (vdefs[i]) = 0;
630 gimple *def = SSA_NAME_DEF_STMT (vdefs[i]);
631 gsi = gsi_for_stmt (def);
632 unlink_stmt_vdef (def);
633 gsi_remove (&gsi, true);
634 release_defs (def);
635 }
636
637 /* Insert the first store at the beginning of the merge BB. */
638 gimple_set_vdef (first_store, gimple_phi_result (phi));
639 SSA_NAME_DEF_STMT (gimple_vdef (first_store)) = first_store;
640 gimple_phi_set_result (phi, make_ssa_name (gimple_vop (cfun)));
641 gimple_set_vuse (first_store, gimple_phi_result (phi));
642 gimple_assign_set_rhs1 (first_store, from);
643 /* ??? Should we reset first_stores location? */
644 gsi = gsi_after_labels (bb);
645 gsi_insert_before (&gsi, first_store, GSI_SAME_STMT);
646 sink_stats.commoned++;
647
648 todo |= TODO_cleanup_cfg;
649 }
650
651 /* We could now have empty predecessors that we could remove,
652 forming a proper CFG for further sinking. Note that even
653 CFG cleanup doesn't do this fully at the moment and it
654 doesn't preserve post-dominators in the process either.
655 The mergephi pass might do it though. gcc.dg/tree-ssa/ssa-sink-13.c
656 shows this nicely if you disable tail merging or (same effect)
657 make the stored values unequal. */
658 }
659
660 return todo;
661 }
662
663 /* Perform code sinking on BB */
664
665 static unsigned
sink_code_in_bb(basic_block bb)666 sink_code_in_bb (basic_block bb)
667 {
668 basic_block son;
669 gimple_stmt_iterator gsi;
670 edge_iterator ei;
671 edge e;
672 bool last = true;
673 unsigned todo = 0;
674
675 /* Sink common stores from the predecessor through our virtual PHI. */
676 todo |= sink_common_stores_to_bb (bb);
677
678 /* If this block doesn't dominate anything, there can't be any place to sink
679 the statements to. */
680 if (first_dom_son (CDI_DOMINATORS, bb) == NULL)
681 goto earlyout;
682
683 /* We can't move things across abnormal edges, so don't try. */
684 FOR_EACH_EDGE (e, ei, bb->succs)
685 if (e->flags & EDGE_ABNORMAL)
686 goto earlyout;
687
688 for (gsi = gsi_last_bb (bb); !gsi_end_p (gsi);)
689 {
690 gimple *stmt = gsi_stmt (gsi);
691 gimple_stmt_iterator togsi;
692 bool zero_uses_p;
693
694 if (!statement_sink_location (stmt, bb, &togsi, &zero_uses_p))
695 {
696 gimple_stmt_iterator saved = gsi;
697 if (!gsi_end_p (gsi))
698 gsi_prev (&gsi);
699 /* If we face a dead stmt remove it as it possibly blocks
700 sinking of uses. */
701 if (zero_uses_p
702 && !gimple_vdef (stmt)
703 && (cfun->can_delete_dead_exceptions
704 || !stmt_could_throw_p (cfun, stmt)))
705 {
706 gsi_remove (&saved, true);
707 release_defs (stmt);
708 }
709 else
710 last = false;
711 continue;
712 }
713 if (dump_file)
714 {
715 fprintf (dump_file, "Sinking ");
716 print_gimple_stmt (dump_file, stmt, 0, TDF_VOPS);
717 fprintf (dump_file, " from bb %d to bb %d\n",
718 bb->index, (gsi_bb (togsi))->index);
719 }
720
721 /* Update virtual operands of statements in the path we
722 do not sink to. */
723 if (gimple_vdef (stmt))
724 {
725 imm_use_iterator iter;
726 use_operand_p use_p;
727 gimple *vuse_stmt;
728
729 FOR_EACH_IMM_USE_STMT (vuse_stmt, iter, gimple_vdef (stmt))
730 if (gimple_code (vuse_stmt) != GIMPLE_PHI)
731 FOR_EACH_IMM_USE_ON_STMT (use_p, iter)
732 SET_USE (use_p, gimple_vuse (stmt));
733 }
734
735 /* If this is the end of the basic block, we need to insert at the end
736 of the basic block. */
737 if (gsi_end_p (togsi))
738 gsi_move_to_bb_end (&gsi, gsi_bb (togsi));
739 else
740 gsi_move_before (&gsi, &togsi);
741
742 sink_stats.sunk++;
743
744 /* If we've just removed the last statement of the BB, the
745 gsi_end_p() test below would fail, but gsi_prev() would have
746 succeeded, and we want it to succeed. So we keep track of
747 whether we're at the last statement and pick up the new last
748 statement. */
749 if (last)
750 {
751 gsi = gsi_last_bb (bb);
752 continue;
753 }
754
755 last = false;
756 if (!gsi_end_p (gsi))
757 gsi_prev (&gsi);
758
759 }
760 earlyout:
761 for (son = first_dom_son (CDI_POST_DOMINATORS, bb);
762 son;
763 son = next_dom_son (CDI_POST_DOMINATORS, son))
764 {
765 todo |= sink_code_in_bb (son);
766 }
767
768 return todo;
769 }
770
771 /* Perform code sinking.
772 This moves code down the flowgraph when we know it would be
773 profitable to do so, or it wouldn't increase the number of
774 executions of the statement.
775
776 IE given
777
778 a_1 = b + c;
779 if (<something>)
780 {
781 }
782 else
783 {
784 foo (&b, &c);
785 a_5 = b + c;
786 }
787 a_6 = PHI (a_5, a_1);
788 USE a_6.
789
790 we'll transform this into:
791
792 if (<something>)
793 {
794 a_1 = b + c;
795 }
796 else
797 {
798 foo (&b, &c);
799 a_5 = b + c;
800 }
801 a_6 = PHI (a_5, a_1);
802 USE a_6.
803
804 Note that this reduces the number of computations of a = b + c to 1
805 when we take the else edge, instead of 2.
806 */
807 namespace {
808
809 const pass_data pass_data_sink_code =
810 {
811 GIMPLE_PASS, /* type */
812 "sink", /* name */
813 OPTGROUP_NONE, /* optinfo_flags */
814 TV_TREE_SINK, /* tv_id */
815 /* PROP_no_crit_edges is ensured by running split_edges_for_insertion in
816 pass_data_sink_code::execute (). */
817 ( PROP_cfg | PROP_ssa ), /* properties_required */
818 0, /* properties_provided */
819 0, /* properties_destroyed */
820 0, /* todo_flags_start */
821 TODO_update_ssa, /* todo_flags_finish */
822 };
823
824 class pass_sink_code : public gimple_opt_pass
825 {
826 public:
pass_sink_code(gcc::context * ctxt)827 pass_sink_code (gcc::context *ctxt)
828 : gimple_opt_pass (pass_data_sink_code, ctxt), unsplit_edges (false)
829 {}
830
831 /* opt_pass methods: */
gate(function *)832 virtual bool gate (function *) { return flag_tree_sink != 0; }
833 virtual unsigned int execute (function *);
clone(void)834 opt_pass *clone (void) { return new pass_sink_code (m_ctxt); }
set_pass_param(unsigned n,bool param)835 void set_pass_param (unsigned n, bool param)
836 {
837 gcc_assert (n == 0);
838 unsplit_edges = param;
839 }
840
841 private:
842 bool unsplit_edges;
843 }; // class pass_sink_code
844
845 unsigned int
execute(function * fun)846 pass_sink_code::execute (function *fun)
847 {
848 loop_optimizer_init (LOOPS_NORMAL);
849 split_edges_for_insertion ();
850 /* Arrange for the critical edge splitting to be undone if requested. */
851 unsigned todo = unsplit_edges ? TODO_cleanup_cfg : 0;
852 connect_infinite_loops_to_exit ();
853 memset (&sink_stats, 0, sizeof (sink_stats));
854 calculate_dominance_info (CDI_DOMINATORS);
855 calculate_dominance_info (CDI_POST_DOMINATORS);
856 todo |= sink_code_in_bb (EXIT_BLOCK_PTR_FOR_FN (fun));
857 statistics_counter_event (fun, "Sunk statements", sink_stats.sunk);
858 statistics_counter_event (fun, "Commoned stores", sink_stats.commoned);
859 free_dominance_info (CDI_POST_DOMINATORS);
860 remove_fake_exit_edges ();
861 loop_optimizer_finalize ();
862
863 return todo;
864 }
865
866 } // anon namespace
867
868 gimple_opt_pass *
make_pass_sink_code(gcc::context * ctxt)869 make_pass_sink_code (gcc::context *ctxt)
870 {
871 return new pass_sink_code (ctxt);
872 }
873