0
|
1 /* Code sinking for trees
|
|
2 Copyright (C) 2001, 2002, 2003, 2004, 2007, 2008, 2009
|
|
3 Free Software Foundation, Inc.
|
|
4 Contributed by Daniel Berlin <dan@dberlin.org>
|
|
5
|
|
6 This file is part of GCC.
|
|
7
|
|
8 GCC is free software; you can redistribute it and/or modify
|
|
9 it under the terms of the GNU General Public License as published by
|
|
10 the Free Software Foundation; either version 3, or (at your option)
|
|
11 any later version.
|
|
12
|
|
13 GCC is distributed in the hope that it will be useful,
|
|
14 but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
15 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
|
16 GNU General Public License 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 #include "config.h"
|
|
23 #include "system.h"
|
|
24 #include "coretypes.h"
|
|
25 #include "tm.h"
|
|
26 #include "ggc.h"
|
|
27 #include "tree.h"
|
|
28 #include "basic-block.h"
|
|
29 #include "diagnostic.h"
|
|
30 #include "tree-inline.h"
|
|
31 #include "tree-flow.h"
|
|
32 #include "gimple.h"
|
|
33 #include "tree-dump.h"
|
|
34 #include "timevar.h"
|
|
35 #include "fibheap.h"
|
|
36 #include "hashtab.h"
|
|
37 #include "tree-iterator.h"
|
|
38 #include "real.h"
|
|
39 #include "alloc-pool.h"
|
|
40 #include "tree-pass.h"
|
|
41 #include "flags.h"
|
|
42 #include "bitmap.h"
|
|
43 #include "langhooks.h"
|
|
44 #include "cfgloop.h"
|
|
45
|
|
46 /* TODO:
|
|
47 1. Sinking store only using scalar promotion (IE without moving the RHS):
|
|
48
|
|
49 *q = p;
|
|
50 p = p + 1;
|
|
51 if (something)
|
|
52 *q = <not p>;
|
|
53 else
|
|
54 y = *q;
|
|
55
|
|
56
|
|
57 should become
|
|
58 sinktemp = p;
|
|
59 p = p + 1;
|
|
60 if (something)
|
|
61 *q = <not p>;
|
|
62 else
|
|
63 {
|
|
64 *q = sinktemp;
|
|
65 y = *q
|
|
66 }
|
|
67 Store copy propagation will take care of the store elimination above.
|
|
68
|
|
69
|
|
70 2. Sinking using Partial Dead Code Elimination. */
|
|
71
|
|
72
|
|
73 static struct
|
|
74 {
|
|
75 /* The number of statements sunk down the flowgraph by code sinking. */
|
|
76 int sunk;
|
|
77
|
|
78 } sink_stats;
|
|
79
|
|
80
|
|
81 /* Given a PHI, and one of its arguments (DEF), find the edge for
|
|
82 that argument and return it. If the argument occurs twice in the PHI node,
|
|
83 we return NULL. */
|
|
84
|
|
85 static basic_block
|
|
86 find_bb_for_arg (gimple phi, tree def)
|
|
87 {
|
|
88 size_t i;
|
|
89 bool foundone = false;
|
|
90 basic_block result = NULL;
|
|
91 for (i = 0; i < gimple_phi_num_args (phi); i++)
|
|
92 if (PHI_ARG_DEF (phi, i) == def)
|
|
93 {
|
|
94 if (foundone)
|
|
95 return NULL;
|
|
96 foundone = true;
|
|
97 result = gimple_phi_arg_edge (phi, i)->src;
|
|
98 }
|
|
99 return result;
|
|
100 }
|
|
101
|
|
102 /* When the first immediate use is in a statement, then return true if all
|
|
103 immediate uses in IMM are in the same statement.
|
|
104 We could also do the case where the first immediate use is in a phi node,
|
|
105 and all the other uses are in phis in the same basic block, but this
|
|
106 requires some expensive checking later (you have to make sure no def/vdef
|
|
107 in the statement occurs for multiple edges in the various phi nodes it's
|
|
108 used in, so that you only have one place you can sink it to. */
|
|
109
|
|
110 static bool
|
|
111 all_immediate_uses_same_place (gimple stmt)
|
|
112 {
|
|
113 gimple firstuse = NULL;
|
|
114 ssa_op_iter op_iter;
|
|
115 imm_use_iterator imm_iter;
|
|
116 use_operand_p use_p;
|
|
117 tree var;
|
|
118
|
|
119 FOR_EACH_SSA_TREE_OPERAND (var, stmt, op_iter, SSA_OP_ALL_DEFS)
|
|
120 {
|
|
121 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, var)
|
|
122 {
|
|
123 if (firstuse == NULL)
|
|
124 firstuse = USE_STMT (use_p);
|
|
125 else
|
|
126 if (firstuse != USE_STMT (use_p))
|
|
127 return false;
|
|
128 }
|
|
129 }
|
|
130
|
|
131 return true;
|
|
132 }
|
|
133
|
|
134 /* Some global stores don't necessarily have VDEF's of global variables,
|
|
135 but we still must avoid moving them around. */
|
|
136
|
|
137 bool
|
|
138 is_hidden_global_store (gimple stmt)
|
|
139 {
|
|
140 /* Check virtual definitions. If we get here, the only virtual
|
|
141 definitions we should see are those generated by assignment or call
|
|
142 statements. */
|
|
143 if (!ZERO_SSA_OPERANDS (stmt, SSA_OP_VIRTUAL_DEFS))
|
|
144 {
|
|
145 tree lhs;
|
|
146
|
|
147 gcc_assert (is_gimple_assign (stmt) || is_gimple_call (stmt));
|
|
148
|
|
149 /* Note that we must not check the individual virtual operands
|
|
150 here. In particular, if this is an aliased store, we could
|
|
151 end up with something like the following (SSA notation
|
|
152 redacted for brevity):
|
|
153
|
|
154 foo (int *p, int i)
|
|
155 {
|
|
156 int x;
|
|
157 p_1 = (i_2 > 3) ? &x : p;
|
|
158
|
|
159 # x_4 = VDEF <x_3>
|
|
160 *p_1 = 5;
|
|
161
|
|
162 return 2;
|
|
163 }
|
|
164
|
|
165 Notice that the store to '*p_1' should be preserved, if we
|
|
166 were to check the virtual definitions in that store, we would
|
|
167 not mark it needed. This is because 'x' is not a global
|
|
168 variable.
|
|
169
|
|
170 Therefore, we check the base address of the LHS. If the
|
|
171 address is a pointer, we check if its name tag or symbol tag is
|
|
172 a global variable. Otherwise, we check if the base variable
|
|
173 is a global. */
|
|
174 lhs = gimple_get_lhs (stmt);
|
|
175
|
|
176 if (REFERENCE_CLASS_P (lhs))
|
|
177 lhs = get_base_address (lhs);
|
|
178
|
|
179 if (lhs == NULL_TREE)
|
|
180 {
|
|
181 /* If LHS is NULL, it means that we couldn't get the base
|
|
182 address of the reference. In which case, we should not
|
|
183 move this store. */
|
|
184 return true;
|
|
185 }
|
|
186 else if (DECL_P (lhs))
|
|
187 {
|
|
188 /* If the store is to a global symbol, we need to keep it. */
|
|
189 if (is_global_var (lhs))
|
|
190 return true;
|
|
191
|
|
192 }
|
|
193 else if (INDIRECT_REF_P (lhs))
|
|
194 return may_point_to_global_var (TREE_OPERAND (lhs, 0));
|
|
195 else
|
|
196 gcc_unreachable ();
|
|
197 }
|
|
198
|
|
199 return false;
|
|
200 }
|
|
201
|
|
202 /* Find the nearest common dominator of all of the immediate uses in IMM. */
|
|
203
|
|
204 static basic_block
|
|
205 nearest_common_dominator_of_uses (gimple stmt)
|
|
206 {
|
|
207 bitmap blocks = BITMAP_ALLOC (NULL);
|
|
208 basic_block commondom;
|
|
209 unsigned int j;
|
|
210 bitmap_iterator bi;
|
|
211 ssa_op_iter op_iter;
|
|
212 imm_use_iterator imm_iter;
|
|
213 use_operand_p use_p;
|
|
214 tree var;
|
|
215
|
|
216 bitmap_clear (blocks);
|
|
217 FOR_EACH_SSA_TREE_OPERAND (var, stmt, op_iter, SSA_OP_ALL_DEFS)
|
|
218 {
|
|
219 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, var)
|
|
220 {
|
|
221 gimple usestmt = USE_STMT (use_p);
|
|
222 basic_block useblock;
|
|
223
|
|
224 if (gimple_code (usestmt) == GIMPLE_PHI)
|
|
225 {
|
|
226 int idx = PHI_ARG_INDEX_FROM_USE (use_p);
|
|
227
|
|
228 useblock = gimple_phi_arg_edge (usestmt, idx)->src;
|
|
229 }
|
|
230 else
|
|
231 {
|
|
232 useblock = gimple_bb (usestmt);
|
|
233 }
|
|
234
|
|
235 /* Short circuit. Nothing dominates the entry block. */
|
|
236 if (useblock == ENTRY_BLOCK_PTR)
|
|
237 {
|
|
238 BITMAP_FREE (blocks);
|
|
239 return NULL;
|
|
240 }
|
|
241 bitmap_set_bit (blocks, useblock->index);
|
|
242 }
|
|
243 }
|
|
244 commondom = BASIC_BLOCK (bitmap_first_set_bit (blocks));
|
|
245 EXECUTE_IF_SET_IN_BITMAP (blocks, 0, j, bi)
|
|
246 commondom = nearest_common_dominator (CDI_DOMINATORS, commondom,
|
|
247 BASIC_BLOCK (j));
|
|
248 BITMAP_FREE (blocks);
|
|
249 return commondom;
|
|
250 }
|
|
251
|
|
252 /* Given a statement (STMT) and the basic block it is currently in (FROMBB),
|
|
253 determine the location to sink the statement to, if any.
|
|
254 Returns true if there is such location; in that case, TOGSI points to the
|
|
255 statement before that STMT should be moved. */
|
|
256
|
|
257 static bool
|
|
258 statement_sink_location (gimple stmt, basic_block frombb,
|
|
259 gimple_stmt_iterator *togsi)
|
|
260 {
|
|
261 gimple use;
|
|
262 tree def;
|
|
263 use_operand_p one_use = NULL_USE_OPERAND_P;
|
|
264 basic_block sinkbb;
|
|
265 use_operand_p use_p;
|
|
266 def_operand_p def_p;
|
|
267 ssa_op_iter iter;
|
|
268 imm_use_iterator imm_iter;
|
|
269 enum tree_code code;
|
|
270
|
|
271 FOR_EACH_SSA_TREE_OPERAND (def, stmt, iter, SSA_OP_ALL_DEFS)
|
|
272 {
|
|
273 FOR_EACH_IMM_USE_FAST (one_use, imm_iter, def)
|
|
274 {
|
|
275 break;
|
|
276 }
|
|
277 if (one_use != NULL_USE_OPERAND_P)
|
|
278 break;
|
|
279 }
|
|
280
|
|
281 /* Return if there are no immediate uses of this stmt. */
|
|
282 if (one_use == NULL_USE_OPERAND_P)
|
|
283 return false;
|
|
284
|
|
285 if (gimple_code (stmt) != GIMPLE_ASSIGN)
|
|
286 return false;
|
|
287
|
|
288 /* There are a few classes of things we can't or don't move, some because we
|
|
289 don't have code to handle it, some because it's not profitable and some
|
|
290 because it's not legal.
|
|
291
|
|
292 We can't sink things that may be global stores, at least not without
|
|
293 calculating a lot more information, because we may cause it to no longer
|
|
294 be seen by an external routine that needs it depending on where it gets
|
|
295 moved to.
|
|
296
|
|
297 We don't want to sink loads from memory.
|
|
298
|
|
299 We can't sink statements that end basic blocks without splitting the
|
|
300 incoming edge for the sink location to place it there.
|
|
301
|
|
302 We can't sink statements that have volatile operands.
|
|
303
|
|
304 We don't want to sink dead code, so anything with 0 immediate uses is not
|
|
305 sunk.
|
|
306
|
|
307 Don't sink BLKmode assignments if current function has any local explicit
|
|
308 register variables, as BLKmode assignments may involve memcpy or memset
|
|
309 calls or, on some targets, inline expansion thereof that sometimes need
|
|
310 to use specific hard registers.
|
|
311
|
|
312 */
|
|
313 code = gimple_assign_rhs_code (stmt);
|
|
314 if (stmt_ends_bb_p (stmt)
|
|
315 || gimple_has_side_effects (stmt)
|
|
316 || code == EXC_PTR_EXPR
|
|
317 || code == FILTER_EXPR
|
|
318 || is_hidden_global_store (stmt)
|
|
319 || gimple_has_volatile_ops (stmt)
|
|
320 || !ZERO_SSA_OPERANDS (stmt, SSA_OP_VUSE)
|
|
321 || (cfun->has_local_explicit_reg_vars
|
|
322 && TYPE_MODE (TREE_TYPE (gimple_assign_lhs (stmt))) == BLKmode))
|
|
323 return false;
|
|
324
|
|
325 FOR_EACH_SSA_DEF_OPERAND (def_p, stmt, iter, SSA_OP_ALL_DEFS)
|
|
326 {
|
|
327 tree def = DEF_FROM_PTR (def_p);
|
|
328 if (is_global_var (SSA_NAME_VAR (def))
|
|
329 || SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def))
|
|
330 return false;
|
|
331 }
|
|
332
|
|
333 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_ALL_USES)
|
|
334 {
|
|
335 tree use = USE_FROM_PTR (use_p);
|
|
336 if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (use))
|
|
337 return false;
|
|
338 }
|
|
339
|
|
340 /* If all the immediate uses are not in the same place, find the nearest
|
|
341 common dominator of all the immediate uses. For PHI nodes, we have to
|
|
342 find the nearest common dominator of all of the predecessor blocks, since
|
|
343 that is where insertion would have to take place. */
|
|
344 if (!all_immediate_uses_same_place (stmt))
|
|
345 {
|
|
346 basic_block commondom = nearest_common_dominator_of_uses (stmt);
|
|
347
|
|
348 if (commondom == frombb)
|
|
349 return false;
|
|
350
|
|
351 /* Our common dominator has to be dominated by frombb in order to be a
|
|
352 trivially safe place to put this statement, since it has multiple
|
|
353 uses. */
|
|
354 if (!dominated_by_p (CDI_DOMINATORS, commondom, frombb))
|
|
355 return false;
|
|
356
|
|
357 /* It doesn't make sense to move to a dominator that post-dominates
|
|
358 frombb, because it means we've just moved it into a path that always
|
|
359 executes if frombb executes, instead of reducing the number of
|
|
360 executions . */
|
|
361 if (dominated_by_p (CDI_POST_DOMINATORS, frombb, commondom))
|
|
362 {
|
|
363 if (dump_file && (dump_flags & TDF_DETAILS))
|
|
364 fprintf (dump_file, "Not moving store, common dominator post-dominates from block.\n");
|
|
365 return false;
|
|
366 }
|
|
367
|
|
368 if (commondom == frombb || commondom->loop_depth > frombb->loop_depth)
|
|
369 return false;
|
|
370 if (dump_file && (dump_flags & TDF_DETAILS))
|
|
371 {
|
|
372 fprintf (dump_file, "Common dominator of all uses is %d\n",
|
|
373 commondom->index);
|
|
374 }
|
|
375 *togsi = gsi_after_labels (commondom);
|
|
376 return true;
|
|
377 }
|
|
378
|
|
379 use = USE_STMT (one_use);
|
|
380 if (gimple_code (use) != GIMPLE_PHI)
|
|
381 {
|
|
382 sinkbb = gimple_bb (use);
|
|
383 if (sinkbb == frombb || sinkbb->loop_depth > frombb->loop_depth
|
|
384 || sinkbb->loop_father != frombb->loop_father)
|
|
385 return false;
|
|
386
|
|
387 *togsi = gsi_for_stmt (use);
|
|
388 return true;
|
|
389 }
|
|
390
|
|
391 /* Note that at this point, all uses must be in the same statement, so it
|
|
392 doesn't matter which def op we choose, pick the first one. */
|
|
393 FOR_EACH_SSA_TREE_OPERAND (def, stmt, iter, SSA_OP_ALL_DEFS)
|
|
394 break;
|
|
395
|
|
396 sinkbb = find_bb_for_arg (use, def);
|
|
397 if (!sinkbb)
|
|
398 return false;
|
|
399
|
|
400 /* This will happen when you have
|
|
401 a_3 = PHI <a_13, a_26>
|
|
402
|
|
403 a_26 = VDEF <a_3>
|
|
404
|
|
405 If the use is a phi, and is in the same bb as the def,
|
|
406 we can't sink it. */
|
|
407
|
|
408 if (gimple_bb (use) == frombb)
|
|
409 return false;
|
|
410 if (sinkbb == frombb || sinkbb->loop_depth > frombb->loop_depth
|
|
411 || sinkbb->loop_father != frombb->loop_father)
|
|
412 return false;
|
|
413
|
|
414 *togsi = gsi_after_labels (sinkbb);
|
|
415
|
|
416 return true;
|
|
417 }
|
|
418
|
|
419 /* Perform code sinking on BB */
|
|
420
|
|
421 static void
|
|
422 sink_code_in_bb (basic_block bb)
|
|
423 {
|
|
424 basic_block son;
|
|
425 gimple_stmt_iterator gsi;
|
|
426 edge_iterator ei;
|
|
427 edge e;
|
|
428 bool last = true;
|
|
429
|
|
430 /* If this block doesn't dominate anything, there can't be any place to sink
|
|
431 the statements to. */
|
|
432 if (first_dom_son (CDI_DOMINATORS, bb) == NULL)
|
|
433 goto earlyout;
|
|
434
|
|
435 /* We can't move things across abnormal edges, so don't try. */
|
|
436 FOR_EACH_EDGE (e, ei, bb->succs)
|
|
437 if (e->flags & EDGE_ABNORMAL)
|
|
438 goto earlyout;
|
|
439
|
|
440 for (gsi = gsi_last_bb (bb); !gsi_end_p (gsi);)
|
|
441 {
|
|
442 gimple stmt = gsi_stmt (gsi);
|
|
443 gimple_stmt_iterator togsi;
|
|
444
|
|
445 if (!statement_sink_location (stmt, bb, &togsi))
|
|
446 {
|
|
447 if (!gsi_end_p (gsi))
|
|
448 gsi_prev (&gsi);
|
|
449 last = false;
|
|
450 continue;
|
|
451 }
|
|
452 if (dump_file)
|
|
453 {
|
|
454 fprintf (dump_file, "Sinking ");
|
|
455 print_gimple_stmt (dump_file, stmt, 0, TDF_VOPS);
|
|
456 fprintf (dump_file, " from bb %d to bb %d\n",
|
|
457 bb->index, (gsi_bb (togsi))->index);
|
|
458 }
|
|
459
|
|
460 /* If this is the end of the basic block, we need to insert at the end
|
|
461 of the basic block. */
|
|
462 if (gsi_end_p (togsi))
|
|
463 gsi_move_to_bb_end (&gsi, gsi_bb (togsi));
|
|
464 else
|
|
465 gsi_move_before (&gsi, &togsi);
|
|
466
|
|
467 sink_stats.sunk++;
|
|
468
|
|
469 /* If we've just removed the last statement of the BB, the
|
|
470 gsi_end_p() test below would fail, but gsi_prev() would have
|
|
471 succeeded, and we want it to succeed. So we keep track of
|
|
472 whether we're at the last statement and pick up the new last
|
|
473 statement. */
|
|
474 if (last)
|
|
475 {
|
|
476 gsi = gsi_last_bb (bb);
|
|
477 continue;
|
|
478 }
|
|
479
|
|
480 last = false;
|
|
481 if (!gsi_end_p (gsi))
|
|
482 gsi_prev (&gsi);
|
|
483
|
|
484 }
|
|
485 earlyout:
|
|
486 for (son = first_dom_son (CDI_POST_DOMINATORS, bb);
|
|
487 son;
|
|
488 son = next_dom_son (CDI_POST_DOMINATORS, son))
|
|
489 {
|
|
490 sink_code_in_bb (son);
|
|
491 }
|
|
492 }
|
|
493
|
|
494 /* Perform code sinking.
|
|
495 This moves code down the flowgraph when we know it would be
|
|
496 profitable to do so, or it wouldn't increase the number of
|
|
497 executions of the statement.
|
|
498
|
|
499 IE given
|
|
500
|
|
501 a_1 = b + c;
|
|
502 if (<something>)
|
|
503 {
|
|
504 }
|
|
505 else
|
|
506 {
|
|
507 foo (&b, &c);
|
|
508 a_5 = b + c;
|
|
509 }
|
|
510 a_6 = PHI (a_5, a_1);
|
|
511 USE a_6.
|
|
512
|
|
513 we'll transform this into:
|
|
514
|
|
515 if (<something>)
|
|
516 {
|
|
517 a_1 = b + c;
|
|
518 }
|
|
519 else
|
|
520 {
|
|
521 foo (&b, &c);
|
|
522 a_5 = b + c;
|
|
523 }
|
|
524 a_6 = PHI (a_5, a_1);
|
|
525 USE a_6.
|
|
526
|
|
527 Note that this reduces the number of computations of a = b + c to 1
|
|
528 when we take the else edge, instead of 2.
|
|
529 */
|
|
530 static void
|
|
531 execute_sink_code (void)
|
|
532 {
|
|
533 loop_optimizer_init (LOOPS_NORMAL);
|
|
534
|
|
535 connect_infinite_loops_to_exit ();
|
|
536 memset (&sink_stats, 0, sizeof (sink_stats));
|
|
537 calculate_dominance_info (CDI_DOMINATORS);
|
|
538 calculate_dominance_info (CDI_POST_DOMINATORS);
|
|
539 sink_code_in_bb (EXIT_BLOCK_PTR);
|
|
540 statistics_counter_event (cfun, "Sunk statements", sink_stats.sunk);
|
|
541 free_dominance_info (CDI_POST_DOMINATORS);
|
|
542 remove_fake_exit_edges ();
|
|
543 loop_optimizer_finalize ();
|
|
544 }
|
|
545
|
|
546 /* Gate and execute functions for PRE. */
|
|
547
|
|
548 static unsigned int
|
|
549 do_sink (void)
|
|
550 {
|
|
551 execute_sink_code ();
|
|
552 return 0;
|
|
553 }
|
|
554
|
|
555 static bool
|
|
556 gate_sink (void)
|
|
557 {
|
|
558 return flag_tree_sink != 0;
|
|
559 }
|
|
560
|
|
561 struct gimple_opt_pass pass_sink_code =
|
|
562 {
|
|
563 {
|
|
564 GIMPLE_PASS,
|
|
565 "sink", /* name */
|
|
566 gate_sink, /* gate */
|
|
567 do_sink, /* execute */
|
|
568 NULL, /* sub */
|
|
569 NULL, /* next */
|
|
570 0, /* static_pass_number */
|
|
571 TV_TREE_SINK, /* tv_id */
|
|
572 PROP_no_crit_edges | PROP_cfg
|
|
573 | PROP_ssa | PROP_alias, /* properties_required */
|
|
574 0, /* properties_provided */
|
|
575 0, /* properties_destroyed */
|
|
576 0, /* todo_flags_start */
|
|
577 TODO_update_ssa
|
|
578 | TODO_dump_func
|
|
579 | TODO_ggc_collect
|
|
580 | TODO_verify_ssa /* todo_flags_finish */
|
|
581 }
|
|
582 };
|