Daily bump.
[official-gcc.git] / gcc / tree-ssa-sink.c
blob0d9029bc66e9686c1a212b6fda24b3f264352df9
1 /* Code sinking for trees
2 Copyright (C) 2001-2013 Free Software Foundation, Inc.
3 Contributed by Daniel Berlin <dan@dberlin.org>
5 This file is part of GCC.
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.
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.
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/>. */
21 #include "config.h"
22 #include "system.h"
23 #include "coretypes.h"
24 #include "tm.h"
25 #include "tree.h"
26 #include "basic-block.h"
27 #include "gimple-pretty-print.h"
28 #include "tree-inline.h"
29 #include "tree-flow.h"
30 #include "gimple.h"
31 #include "hashtab.h"
32 #include "tree-iterator.h"
33 #include "alloc-pool.h"
34 #include "tree-pass.h"
35 #include "flags.h"
36 #include "bitmap.h"
37 #include "cfgloop.h"
38 #include "params.h"
40 /* TODO:
41 1. Sinking store only using scalar promotion (IE without moving the RHS):
43 *q = p;
44 p = p + 1;
45 if (something)
46 *q = <not p>;
47 else
48 y = *q;
51 should become
52 sinktemp = p;
53 p = p + 1;
54 if (something)
55 *q = <not p>;
56 else
58 *q = sinktemp;
59 y = *q
61 Store copy propagation will take care of the store elimination above.
64 2. Sinking using Partial Dead Code Elimination. */
67 static struct
69 /* The number of statements sunk down the flowgraph by code sinking. */
70 int sunk;
72 } sink_stats;
75 /* Given a PHI, and one of its arguments (DEF), find the edge for
76 that argument and return it. If the argument occurs twice in the PHI node,
77 we return NULL. */
79 static basic_block
80 find_bb_for_arg (gimple phi, tree def)
82 size_t i;
83 bool foundone = false;
84 basic_block result = NULL;
85 for (i = 0; i < gimple_phi_num_args (phi); i++)
86 if (PHI_ARG_DEF (phi, i) == def)
88 if (foundone)
89 return NULL;
90 foundone = true;
91 result = gimple_phi_arg_edge (phi, i)->src;
93 return result;
96 /* When the first immediate use is in a statement, then return true if all
97 immediate uses in IMM are in the same statement.
98 We could also do the case where the first immediate use is in a phi node,
99 and all the other uses are in phis in the same basic block, but this
100 requires some expensive checking later (you have to make sure no def/vdef
101 in the statement occurs for multiple edges in the various phi nodes it's
102 used in, so that you only have one place you can sink it to. */
104 static bool
105 all_immediate_uses_same_place (gimple stmt)
107 gimple firstuse = NULL;
108 ssa_op_iter op_iter;
109 imm_use_iterator imm_iter;
110 use_operand_p use_p;
111 tree var;
113 FOR_EACH_SSA_TREE_OPERAND (var, stmt, op_iter, SSA_OP_ALL_DEFS)
115 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, var)
117 if (is_gimple_debug (USE_STMT (use_p)))
118 continue;
119 if (firstuse == NULL)
120 firstuse = USE_STMT (use_p);
121 else
122 if (firstuse != USE_STMT (use_p))
123 return false;
127 return true;
130 /* Find the nearest common dominator of all of the immediate uses in IMM. */
132 static basic_block
133 nearest_common_dominator_of_uses (gimple stmt, bool *debug_stmts)
135 bitmap blocks = BITMAP_ALLOC (NULL);
136 basic_block commondom;
137 unsigned int j;
138 bitmap_iterator bi;
139 ssa_op_iter op_iter;
140 imm_use_iterator imm_iter;
141 use_operand_p use_p;
142 tree var;
144 bitmap_clear (blocks);
145 FOR_EACH_SSA_TREE_OPERAND (var, stmt, op_iter, SSA_OP_ALL_DEFS)
147 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, var)
149 gimple usestmt = USE_STMT (use_p);
150 basic_block useblock;
152 if (gimple_code (usestmt) == GIMPLE_PHI)
154 int idx = PHI_ARG_INDEX_FROM_USE (use_p);
156 useblock = gimple_phi_arg_edge (usestmt, idx)->src;
158 else if (is_gimple_debug (usestmt))
160 *debug_stmts = true;
161 continue;
163 else
165 useblock = gimple_bb (usestmt);
168 /* Short circuit. Nothing dominates the entry block. */
169 if (useblock == ENTRY_BLOCK_PTR)
171 BITMAP_FREE (blocks);
172 return NULL;
174 bitmap_set_bit (blocks, useblock->index);
177 commondom = BASIC_BLOCK (bitmap_first_set_bit (blocks));
178 EXECUTE_IF_SET_IN_BITMAP (blocks, 0, j, bi)
179 commondom = nearest_common_dominator (CDI_DOMINATORS, commondom,
180 BASIC_BLOCK (j));
181 BITMAP_FREE (blocks);
182 return commondom;
185 /* Given EARLY_BB and LATE_BB, two blocks in a path through the dominator
186 tree, return the best basic block between them (inclusive) to place
187 statements.
189 We want the most control dependent block in the shallowest loop nest.
191 If the resulting block is in a shallower loop nest, then use it. Else
192 only use the resulting block if it has significantly lower execution
193 frequency than EARLY_BB to avoid gratutious statement movement. We
194 consider statements with VOPS more desirable to move.
196 This pass would obviously benefit from PDO as it utilizes block
197 frequencies. It would also benefit from recomputing frequencies
198 if profile data is not available since frequencies often get out
199 of sync with reality. */
201 static basic_block
202 select_best_block (basic_block early_bb,
203 basic_block late_bb,
204 gimple stmt)
206 basic_block best_bb = late_bb;
207 basic_block temp_bb = late_bb;
208 int threshold;
210 while (temp_bb != early_bb)
212 /* If we've moved into a lower loop nest, then that becomes
213 our best block. */
214 if (bb_loop_depth (temp_bb) < bb_loop_depth (best_bb))
215 best_bb = temp_bb;
217 /* Walk up the dominator tree, hopefully we'll find a shallower
218 loop nest. */
219 temp_bb = get_immediate_dominator (CDI_DOMINATORS, temp_bb);
222 /* If we found a shallower loop nest, then we always consider that
223 a win. This will always give us the most control dependent block
224 within that loop nest. */
225 if (bb_loop_depth (best_bb) < bb_loop_depth (early_bb))
226 return best_bb;
228 /* Get the sinking threshold. If the statement to be moved has memory
229 operands, then increase the threshold by 7% as those are even more
230 profitable to avoid, clamping at 100%. */
231 threshold = PARAM_VALUE (PARAM_SINK_FREQUENCY_THRESHOLD);
232 if (gimple_vuse (stmt) || gimple_vdef (stmt))
234 threshold += 7;
235 if (threshold > 100)
236 threshold = 100;
239 /* If BEST_BB is at the same nesting level, then require it to have
240 significantly lower execution frequency to avoid gratutious movement. */
241 if (bb_loop_depth (best_bb) == bb_loop_depth (early_bb)
242 && best_bb->frequency < (early_bb->frequency * threshold / 100.0))
243 return best_bb;
245 /* No better block found, so return EARLY_BB, which happens to be the
246 statement's original block. */
247 return early_bb;
250 /* Given a statement (STMT) and the basic block it is currently in (FROMBB),
251 determine the location to sink the statement to, if any.
252 Returns true if there is such location; in that case, TOGSI points to the
253 statement before that STMT should be moved. */
255 static bool
256 statement_sink_location (gimple stmt, basic_block frombb,
257 gimple_stmt_iterator *togsi)
259 gimple use;
260 use_operand_p one_use = NULL_USE_OPERAND_P;
261 basic_block sinkbb;
262 use_operand_p use_p;
263 def_operand_p def_p;
264 ssa_op_iter iter;
265 imm_use_iterator imm_iter;
267 /* We only can sink assignments. */
268 if (!is_gimple_assign (stmt))
269 return false;
271 /* We only can sink stmts with a single definition. */
272 def_p = single_ssa_def_operand (stmt, SSA_OP_ALL_DEFS);
273 if (def_p == NULL_DEF_OPERAND_P)
274 return false;
276 /* Return if there are no immediate uses of this stmt. */
277 if (has_zero_uses (DEF_FROM_PTR (def_p)))
278 return false;
280 /* There are a few classes of things we can't or don't move, some because we
281 don't have code to handle it, some because it's not profitable and some
282 because it's not legal.
284 We can't sink things that may be global stores, at least not without
285 calculating a lot more information, because we may cause it to no longer
286 be seen by an external routine that needs it depending on where it gets
287 moved to.
289 We don't want to sink loads from memory.
291 We can't sink statements that end basic blocks without splitting the
292 incoming edge for the sink location to place it there.
294 We can't sink statements that have volatile operands.
296 We don't want to sink dead code, so anything with 0 immediate uses is not
297 sunk.
299 Don't sink BLKmode assignments if current function has any local explicit
300 register variables, as BLKmode assignments may involve memcpy or memset
301 calls or, on some targets, inline expansion thereof that sometimes need
302 to use specific hard registers.
305 if (stmt_ends_bb_p (stmt)
306 || gimple_has_side_effects (stmt)
307 || gimple_has_volatile_ops (stmt)
308 || (gimple_vuse (stmt) && !gimple_vdef (stmt))
309 || (cfun->has_local_explicit_reg_vars
310 && TYPE_MODE (TREE_TYPE (gimple_assign_lhs (stmt))) == BLKmode))
311 return false;
313 if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (DEF_FROM_PTR (def_p)))
314 return false;
316 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_ALL_USES)
318 tree use = USE_FROM_PTR (use_p);
319 if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (use))
320 return false;
323 use = NULL;
325 /* If stmt is a store the one and only use needs to be the VOP
326 merging PHI node. */
327 if (gimple_vdef (stmt))
329 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p))
331 gimple use_stmt = USE_STMT (use_p);
333 /* A killing definition is not a use. */
334 if (gimple_assign_single_p (use_stmt)
335 && gimple_vdef (use_stmt)
336 && operand_equal_p (gimple_assign_lhs (stmt),
337 gimple_assign_lhs (use_stmt), 0))
338 continue;
340 if (gimple_code (use_stmt) != GIMPLE_PHI)
341 return false;
343 if (use
344 && use != use_stmt)
345 return false;
347 use = use_stmt;
349 if (!use)
350 return false;
352 /* If all the immediate uses are not in the same place, find the nearest
353 common dominator of all the immediate uses. For PHI nodes, we have to
354 find the nearest common dominator of all of the predecessor blocks, since
355 that is where insertion would have to take place. */
356 else if (!all_immediate_uses_same_place (stmt))
358 bool debug_stmts = false;
359 basic_block commondom = nearest_common_dominator_of_uses (stmt,
360 &debug_stmts);
362 if (commondom == frombb)
363 return false;
365 /* Our common dominator has to be dominated by frombb in order to be a
366 trivially safe place to put this statement, since it has multiple
367 uses. */
368 if (!dominated_by_p (CDI_DOMINATORS, commondom, frombb))
369 return false;
371 commondom = select_best_block (frombb, commondom, stmt);
373 if (commondom == frombb)
374 return false;
376 *togsi = gsi_after_labels (commondom);
378 return true;
380 else
382 FOR_EACH_IMM_USE_FAST (one_use, imm_iter, DEF_FROM_PTR (def_p))
384 if (is_gimple_debug (USE_STMT (one_use)))
385 continue;
386 break;
388 use = USE_STMT (one_use);
390 if (gimple_code (use) != GIMPLE_PHI)
392 sinkbb = gimple_bb (use);
393 sinkbb = select_best_block (frombb, gimple_bb (use), stmt);
395 if (sinkbb == frombb)
396 return false;
398 *togsi = gsi_for_stmt (use);
400 return true;
404 sinkbb = find_bb_for_arg (use, DEF_FROM_PTR (def_p));
406 /* This can happen if there are multiple uses in a PHI. */
407 if (!sinkbb)
408 return false;
410 sinkbb = select_best_block (frombb, sinkbb, stmt);
411 if (!sinkbb || sinkbb == frombb)
412 return false;
414 /* If the latch block is empty, don't make it non-empty by sinking
415 something into it. */
416 if (sinkbb == frombb->loop_father->latch
417 && empty_block_p (sinkbb))
418 return false;
420 *togsi = gsi_after_labels (sinkbb);
422 return true;
425 /* Perform code sinking on BB */
427 static void
428 sink_code_in_bb (basic_block bb)
430 basic_block son;
431 gimple_stmt_iterator gsi;
432 edge_iterator ei;
433 edge e;
434 bool last = true;
436 /* If this block doesn't dominate anything, there can't be any place to sink
437 the statements to. */
438 if (first_dom_son (CDI_DOMINATORS, bb) == NULL)
439 goto earlyout;
441 /* We can't move things across abnormal edges, so don't try. */
442 FOR_EACH_EDGE (e, ei, bb->succs)
443 if (e->flags & EDGE_ABNORMAL)
444 goto earlyout;
446 for (gsi = gsi_last_bb (bb); !gsi_end_p (gsi);)
448 gimple stmt = gsi_stmt (gsi);
449 gimple_stmt_iterator togsi;
451 if (!statement_sink_location (stmt, bb, &togsi))
453 if (!gsi_end_p (gsi))
454 gsi_prev (&gsi);
455 last = false;
456 continue;
458 if (dump_file)
460 fprintf (dump_file, "Sinking ");
461 print_gimple_stmt (dump_file, stmt, 0, TDF_VOPS);
462 fprintf (dump_file, " from bb %d to bb %d\n",
463 bb->index, (gsi_bb (togsi))->index);
466 /* Update virtual operands of statements in the path we
467 do not sink to. */
468 if (gimple_vdef (stmt))
470 imm_use_iterator iter;
471 use_operand_p use_p;
472 gimple vuse_stmt;
474 FOR_EACH_IMM_USE_STMT (vuse_stmt, iter, gimple_vdef (stmt))
475 if (gimple_code (vuse_stmt) != GIMPLE_PHI)
476 FOR_EACH_IMM_USE_ON_STMT (use_p, iter)
477 SET_USE (use_p, gimple_vuse (stmt));
480 /* If this is the end of the basic block, we need to insert at the end
481 of the basic block. */
482 if (gsi_end_p (togsi))
483 gsi_move_to_bb_end (&gsi, gsi_bb (togsi));
484 else
485 gsi_move_before (&gsi, &togsi);
487 sink_stats.sunk++;
489 /* If we've just removed the last statement of the BB, the
490 gsi_end_p() test below would fail, but gsi_prev() would have
491 succeeded, and we want it to succeed. So we keep track of
492 whether we're at the last statement and pick up the new last
493 statement. */
494 if (last)
496 gsi = gsi_last_bb (bb);
497 continue;
500 last = false;
501 if (!gsi_end_p (gsi))
502 gsi_prev (&gsi);
505 earlyout:
506 for (son = first_dom_son (CDI_POST_DOMINATORS, bb);
507 son;
508 son = next_dom_son (CDI_POST_DOMINATORS, son))
510 sink_code_in_bb (son);
514 /* Perform code sinking.
515 This moves code down the flowgraph when we know it would be
516 profitable to do so, or it wouldn't increase the number of
517 executions of the statement.
519 IE given
521 a_1 = b + c;
522 if (<something>)
525 else
527 foo (&b, &c);
528 a_5 = b + c;
530 a_6 = PHI (a_5, a_1);
531 USE a_6.
533 we'll transform this into:
535 if (<something>)
537 a_1 = b + c;
539 else
541 foo (&b, &c);
542 a_5 = b + c;
544 a_6 = PHI (a_5, a_1);
545 USE a_6.
547 Note that this reduces the number of computations of a = b + c to 1
548 when we take the else edge, instead of 2.
550 static void
551 execute_sink_code (void)
553 loop_optimizer_init (LOOPS_NORMAL);
555 connect_infinite_loops_to_exit ();
556 memset (&sink_stats, 0, sizeof (sink_stats));
557 calculate_dominance_info (CDI_DOMINATORS);
558 calculate_dominance_info (CDI_POST_DOMINATORS);
559 sink_code_in_bb (EXIT_BLOCK_PTR);
560 statistics_counter_event (cfun, "Sunk statements", sink_stats.sunk);
561 free_dominance_info (CDI_POST_DOMINATORS);
562 remove_fake_exit_edges ();
563 loop_optimizer_finalize ();
566 /* Gate and execute functions for PRE. */
568 static unsigned int
569 do_sink (void)
571 execute_sink_code ();
572 return 0;
575 static bool
576 gate_sink (void)
578 return flag_tree_sink != 0;
581 struct gimple_opt_pass pass_sink_code =
584 GIMPLE_PASS,
585 "sink", /* name */
586 OPTGROUP_NONE, /* optinfo_flags */
587 gate_sink, /* gate */
588 do_sink, /* execute */
589 NULL, /* sub */
590 NULL, /* next */
591 0, /* static_pass_number */
592 TV_TREE_SINK, /* tv_id */
593 PROP_no_crit_edges | PROP_cfg
594 | PROP_ssa, /* properties_required */
595 0, /* properties_provided */
596 0, /* properties_destroyed */
597 0, /* todo_flags_start */
598 TODO_update_ssa
599 | TODO_verify_ssa
600 | TODO_verify_flow /* todo_flags_finish */