op-common.h, [...]: Fix comment typos.
[gcc.git] / gcc / tree-ssa-dom.c
1 /* SSA Dominator optimizations for trees
2 Copyright (C) 2001, 2002, 2003, 2004, 2005, 2006
3 Free Software Foundation, Inc.
4 Contributed by Diego Novillo <dnovillo@redhat.com>
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 2, 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 COPYING. If not, write to
20 the Free Software Foundation, 51 Franklin Street, Fifth Floor,
21 Boston, MA 02110-1301, USA. */
22
23 #include "config.h"
24 #include "system.h"
25 #include "coretypes.h"
26 #include "tm.h"
27 #include "tree.h"
28 #include "flags.h"
29 #include "rtl.h"
30 #include "tm_p.h"
31 #include "ggc.h"
32 #include "basic-block.h"
33 #include "cfgloop.h"
34 #include "output.h"
35 #include "expr.h"
36 #include "function.h"
37 #include "diagnostic.h"
38 #include "timevar.h"
39 #include "tree-dump.h"
40 #include "tree-flow.h"
41 #include "domwalk.h"
42 #include "real.h"
43 #include "tree-pass.h"
44 #include "tree-ssa-propagate.h"
45 #include "langhooks.h"
46 #include "params.h"
47
48 /* This file implements optimizations on the dominator tree. */
49
50
51 /* Structure for recording edge equivalences as well as any pending
52 edge redirections during the dominator optimizer.
53
54 Computing and storing the edge equivalences instead of creating
55 them on-demand can save significant amounts of time, particularly
56 for pathological cases involving switch statements.
57
58 These structures live for a single iteration of the dominator
59 optimizer in the edge's AUX field. At the end of an iteration we
60 free each of these structures and update the AUX field to point
61 to any requested redirection target (the code for updating the
62 CFG and SSA graph for edge redirection expects redirection edge
63 targets to be in the AUX field for each edge. */
64
65 struct edge_info
66 {
67 /* If this edge creates a simple equivalence, the LHS and RHS of
68 the equivalence will be stored here. */
69 tree lhs;
70 tree rhs;
71
72 /* Traversing an edge may also indicate one or more particular conditions
73 are true or false. The number of recorded conditions can vary, but
74 can be determined by the condition's code. So we have an array
75 and its maximum index rather than use a varray. */
76 tree *cond_equivalences;
77 unsigned int max_cond_equivalences;
78 };
79
80
81 /* Hash table with expressions made available during the renaming process.
82 When an assignment of the form X_i = EXPR is found, the statement is
83 stored in this table. If the same expression EXPR is later found on the
84 RHS of another statement, it is replaced with X_i (thus performing
85 global redundancy elimination). Similarly as we pass through conditionals
86 we record the conditional itself as having either a true or false value
87 in this table. */
88 static htab_t avail_exprs;
89
90 /* Stack of available expressions in AVAIL_EXPRs. Each block pushes any
91 expressions it enters into the hash table along with a marker entry
92 (null). When we finish processing the block, we pop off entries and
93 remove the expressions from the global hash table until we hit the
94 marker. */
95 static VEC(tree,heap) *avail_exprs_stack;
96
97 /* Stack of statements we need to rescan during finalization for newly
98 exposed variables.
99
100 Statement rescanning must occur after the current block's available
101 expressions are removed from AVAIL_EXPRS. Else we may change the
102 hash code for an expression and be unable to find/remove it from
103 AVAIL_EXPRS. */
104 static VEC(tree,heap) *stmts_to_rescan;
105
106 /* Structure for entries in the expression hash table.
107
108 This requires more memory for the hash table entries, but allows us
109 to avoid creating silly tree nodes and annotations for conditionals,
110 eliminates 2 global hash tables and two block local varrays.
111
112 It also allows us to reduce the number of hash table lookups we
113 have to perform in lookup_avail_expr and finally it allows us to
114 significantly reduce the number of calls into the hashing routine
115 itself. */
116
117 struct expr_hash_elt
118 {
119 /* The value (lhs) of this expression. */
120 tree lhs;
121
122 /* The expression (rhs) we want to record. */
123 tree rhs;
124
125 /* The stmt pointer if this element corresponds to a statement. */
126 tree stmt;
127
128 /* The hash value for RHS/ann. */
129 hashval_t hash;
130 };
131
132 /* Stack of dest,src pairs that need to be restored during finalization.
133
134 A NULL entry is used to mark the end of pairs which need to be
135 restored during finalization of this block. */
136 static VEC(tree,heap) *const_and_copies_stack;
137
138 /* Track whether or not we have changed the control flow graph. */
139 static bool cfg_altered;
140
141 /* Bitmap of blocks that have had EH statements cleaned. We should
142 remove their dead edges eventually. */
143 static bitmap need_eh_cleanup;
144
145 /* Statistics for dominator optimizations. */
146 struct opt_stats_d
147 {
148 long num_stmts;
149 long num_exprs_considered;
150 long num_re;
151 long num_const_prop;
152 long num_copy_prop;
153 };
154
155 static struct opt_stats_d opt_stats;
156
157 struct eq_expr_value
158 {
159 tree src;
160 tree dst;
161 };
162
163 /* Local functions. */
164 static void optimize_stmt (struct dom_walk_data *,
165 basic_block bb,
166 block_stmt_iterator);
167 static tree lookup_avail_expr (tree, bool);
168 static hashval_t avail_expr_hash (const void *);
169 static hashval_t real_avail_expr_hash (const void *);
170 static int avail_expr_eq (const void *, const void *);
171 static void htab_statistics (FILE *, htab_t);
172 static void record_cond (tree, tree);
173 static void record_const_or_copy (tree, tree);
174 static void record_equality (tree, tree);
175 static void record_equivalences_from_phis (basic_block);
176 static void record_equivalences_from_incoming_edge (basic_block);
177 static bool eliminate_redundant_computations (tree);
178 static void record_equivalences_from_stmt (tree, int, stmt_ann_t);
179 static void dom_thread_across_edge (struct dom_walk_data *, edge);
180 static void dom_opt_finalize_block (struct dom_walk_data *, basic_block);
181 static void dom_opt_initialize_block (struct dom_walk_data *, basic_block);
182 static void propagate_to_outgoing_edges (struct dom_walk_data *, basic_block);
183 static void remove_local_expressions_from_table (void);
184 static void restore_vars_to_original_value (void);
185 static edge single_incoming_edge_ignoring_loop_edges (basic_block);
186
187
188 /* Allocate an EDGE_INFO for edge E and attach it to E.
189 Return the new EDGE_INFO structure. */
190
191 static struct edge_info *
192 allocate_edge_info (edge e)
193 {
194 struct edge_info *edge_info;
195
196 edge_info = XCNEW (struct edge_info);
197
198 e->aux = edge_info;
199 return edge_info;
200 }
201
202 /* Free all EDGE_INFO structures associated with edges in the CFG.
203 If a particular edge can be threaded, copy the redirection
204 target from the EDGE_INFO structure into the edge's AUX field
205 as required by code to update the CFG and SSA graph for
206 jump threading. */
207
208 static void
209 free_all_edge_infos (void)
210 {
211 basic_block bb;
212 edge_iterator ei;
213 edge e;
214
215 FOR_EACH_BB (bb)
216 {
217 FOR_EACH_EDGE (e, ei, bb->preds)
218 {
219 struct edge_info *edge_info = (struct edge_info *) e->aux;
220
221 if (edge_info)
222 {
223 if (edge_info->cond_equivalences)
224 free (edge_info->cond_equivalences);
225 free (edge_info);
226 e->aux = NULL;
227 }
228 }
229 }
230 }
231
232 /* Jump threading, redundancy elimination and const/copy propagation.
233
234 This pass may expose new symbols that need to be renamed into SSA. For
235 every new symbol exposed, its corresponding bit will be set in
236 VARS_TO_RENAME. */
237
238 static unsigned int
239 tree_ssa_dominator_optimize (void)
240 {
241 struct dom_walk_data walk_data;
242 unsigned int i;
243 struct loops loops_info;
244
245 memset (&opt_stats, 0, sizeof (opt_stats));
246
247 /* Create our hash tables. */
248 avail_exprs = htab_create (1024, real_avail_expr_hash, avail_expr_eq, free);
249 avail_exprs_stack = VEC_alloc (tree, heap, 20);
250 const_and_copies_stack = VEC_alloc (tree, heap, 20);
251 stmts_to_rescan = VEC_alloc (tree, heap, 20);
252 need_eh_cleanup = BITMAP_ALLOC (NULL);
253
254 /* Setup callbacks for the generic dominator tree walker. */
255 walk_data.walk_stmts_backward = false;
256 walk_data.dom_direction = CDI_DOMINATORS;
257 walk_data.initialize_block_local_data = NULL;
258 walk_data.before_dom_children_before_stmts = dom_opt_initialize_block;
259 walk_data.before_dom_children_walk_stmts = optimize_stmt;
260 walk_data.before_dom_children_after_stmts = propagate_to_outgoing_edges;
261 walk_data.after_dom_children_before_stmts = NULL;
262 walk_data.after_dom_children_walk_stmts = NULL;
263 walk_data.after_dom_children_after_stmts = dom_opt_finalize_block;
264 /* Right now we only attach a dummy COND_EXPR to the global data pointer.
265 When we attach more stuff we'll need to fill this out with a real
266 structure. */
267 walk_data.global_data = NULL;
268 walk_data.block_local_data_size = 0;
269 walk_data.interesting_blocks = NULL;
270
271 /* Now initialize the dominator walker. */
272 init_walk_dominator_tree (&walk_data);
273
274 calculate_dominance_info (CDI_DOMINATORS);
275
276 /* We need to know which edges exit loops so that we can
277 aggressively thread through loop headers to an exit
278 edge. */
279 flow_loops_find (&loops_info);
280 mark_loop_exit_edges (&loops_info);
281 flow_loops_free (&loops_info);
282
283 /* Clean up the CFG so that any forwarder blocks created by loop
284 canonicalization are removed. */
285 cleanup_tree_cfg ();
286 calculate_dominance_info (CDI_DOMINATORS);
287
288 /* We need accurate information regarding back edges in the CFG
289 for jump threading. */
290 mark_dfs_back_edges ();
291
292 /* Recursively walk the dominator tree optimizing statements. */
293 walk_dominator_tree (&walk_data, ENTRY_BLOCK_PTR);
294
295 {
296 block_stmt_iterator bsi;
297 basic_block bb;
298 FOR_EACH_BB (bb)
299 {
300 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
301 update_stmt_if_modified (bsi_stmt (bsi));
302 }
303 }
304
305 /* If we exposed any new variables, go ahead and put them into
306 SSA form now, before we handle jump threading. This simplifies
307 interactions between rewriting of _DECL nodes into SSA form
308 and rewriting SSA_NAME nodes into SSA form after block
309 duplication and CFG manipulation. */
310 update_ssa (TODO_update_ssa);
311
312 free_all_edge_infos ();
313
314 /* Thread jumps, creating duplicate blocks as needed. */
315 cfg_altered |= thread_through_all_blocks ();
316
317 /* Removal of statements may make some EH edges dead. Purge
318 such edges from the CFG as needed. */
319 if (!bitmap_empty_p (need_eh_cleanup))
320 {
321 cfg_altered |= tree_purge_all_dead_eh_edges (need_eh_cleanup);
322 bitmap_zero (need_eh_cleanup);
323 }
324
325 if (cfg_altered)
326 free_dominance_info (CDI_DOMINATORS);
327
328 /* Finally, remove everything except invariants in SSA_NAME_VALUE.
329
330 Long term we will be able to let everything in SSA_NAME_VALUE
331 persist. However, for now, we know this is the safe thing to do. */
332 for (i = 0; i < num_ssa_names; i++)
333 {
334 tree name = ssa_name (i);
335 tree value;
336
337 if (!name)
338 continue;
339
340 value = SSA_NAME_VALUE (name);
341 if (value && !is_gimple_min_invariant (value))
342 SSA_NAME_VALUE (name) = NULL;
343 }
344
345 /* Debugging dumps. */
346 if (dump_file && (dump_flags & TDF_STATS))
347 dump_dominator_optimization_stats (dump_file);
348
349 /* Delete our main hashtable. */
350 htab_delete (avail_exprs);
351
352 /* And finalize the dominator walker. */
353 fini_walk_dominator_tree (&walk_data);
354
355 /* Free asserted bitmaps and stacks. */
356 BITMAP_FREE (need_eh_cleanup);
357
358 VEC_free (tree, heap, avail_exprs_stack);
359 VEC_free (tree, heap, const_and_copies_stack);
360 VEC_free (tree, heap, stmts_to_rescan);
361 return 0;
362 }
363
364 static bool
365 gate_dominator (void)
366 {
367 return flag_tree_dom != 0;
368 }
369
370 struct tree_opt_pass pass_dominator =
371 {
372 "dom", /* name */
373 gate_dominator, /* gate */
374 tree_ssa_dominator_optimize, /* execute */
375 NULL, /* sub */
376 NULL, /* next */
377 0, /* static_pass_number */
378 TV_TREE_SSA_DOMINATOR_OPTS, /* tv_id */
379 PROP_cfg | PROP_ssa | PROP_alias, /* properties_required */
380 0, /* properties_provided */
381 PROP_smt_usage, /* properties_destroyed */
382 0, /* todo_flags_start */
383 TODO_dump_func
384 | TODO_update_ssa
385 | TODO_cleanup_cfg
386 | TODO_verify_ssa
387 | TODO_update_smt_usage, /* todo_flags_finish */
388 0 /* letter */
389 };
390
391
392 /* Given a stmt CONDSTMT containing a COND_EXPR, canonicalize the
393 COND_EXPR into a canonical form. */
394
395 static void
396 canonicalize_comparison (tree condstmt)
397 {
398 tree cond = COND_EXPR_COND (condstmt);
399 tree op0;
400 tree op1;
401 enum tree_code code = TREE_CODE (cond);
402
403 if (!COMPARISON_CLASS_P (cond))
404 return;
405
406 op0 = TREE_OPERAND (cond, 0);
407 op1 = TREE_OPERAND (cond, 1);
408
409 /* If it would be profitable to swap the operands, then do so to
410 canonicalize the statement, enabling better optimization.
411
412 By placing canonicalization of such expressions here we
413 transparently keep statements in canonical form, even
414 when the statement is modified. */
415 if (tree_swap_operands_p (op0, op1, false))
416 {
417 /* For relationals we need to swap the operands
418 and change the code. */
419 if (code == LT_EXPR
420 || code == GT_EXPR
421 || code == LE_EXPR
422 || code == GE_EXPR)
423 {
424 TREE_SET_CODE (cond, swap_tree_comparison (code));
425 swap_tree_operands (condstmt,
426 &TREE_OPERAND (cond, 0),
427 &TREE_OPERAND (cond, 1));
428 /* If one operand was in the operand cache, but the other is
429 not, because it is a constant, this is a case that the
430 internal updating code of swap_tree_operands can't handle
431 properly. */
432 if (TREE_CODE_CLASS (TREE_CODE (op0))
433 != TREE_CODE_CLASS (TREE_CODE (op1)))
434 update_stmt (condstmt);
435 }
436 }
437 }
438
439 /* Initialize local stacks for this optimizer and record equivalences
440 upon entry to BB. Equivalences can come from the edge traversed to
441 reach BB or they may come from PHI nodes at the start of BB. */
442
443 static void
444 dom_opt_initialize_block (struct dom_walk_data *walk_data ATTRIBUTE_UNUSED,
445 basic_block bb)
446 {
447 if (dump_file && (dump_flags & TDF_DETAILS))
448 fprintf (dump_file, "\n\nOptimizing block #%d\n\n", bb->index);
449
450 /* Push a marker on the stacks of local information so that we know how
451 far to unwind when we finalize this block. */
452 VEC_safe_push (tree, heap, avail_exprs_stack, NULL_TREE);
453 VEC_safe_push (tree, heap, const_and_copies_stack, NULL_TREE);
454
455 record_equivalences_from_incoming_edge (bb);
456
457 /* PHI nodes can create equivalences too. */
458 record_equivalences_from_phis (bb);
459 }
460
461 /* Given an expression EXPR (a relational expression or a statement),
462 initialize the hash table element pointed to by ELEMENT. */
463
464 static void
465 initialize_hash_element (tree expr, tree lhs, struct expr_hash_elt *element)
466 {
467 /* Hash table elements may be based on conditional expressions or statements.
468
469 For the former case, we have no annotation and we want to hash the
470 conditional expression. In the latter case we have an annotation and
471 we want to record the expression the statement evaluates. */
472 if (COMPARISON_CLASS_P (expr) || TREE_CODE (expr) == TRUTH_NOT_EXPR)
473 {
474 element->stmt = NULL;
475 element->rhs = expr;
476 }
477 else if (TREE_CODE (expr) == COND_EXPR)
478 {
479 element->stmt = expr;
480 element->rhs = COND_EXPR_COND (expr);
481 }
482 else if (TREE_CODE (expr) == SWITCH_EXPR)
483 {
484 element->stmt = expr;
485 element->rhs = SWITCH_COND (expr);
486 }
487 else if (TREE_CODE (expr) == RETURN_EXPR && TREE_OPERAND (expr, 0))
488 {
489 element->stmt = expr;
490 element->rhs = TREE_OPERAND (TREE_OPERAND (expr, 0), 1);
491 }
492 else if (TREE_CODE (expr) == GOTO_EXPR)
493 {
494 element->stmt = expr;
495 element->rhs = GOTO_DESTINATION (expr);
496 }
497 else
498 {
499 element->stmt = expr;
500 element->rhs = TREE_OPERAND (expr, 1);
501 }
502
503 element->lhs = lhs;
504 element->hash = avail_expr_hash (element);
505 }
506
507 /* Remove all the expressions in LOCALS from TABLE, stopping when there are
508 LIMIT entries left in LOCALs. */
509
510 static void
511 remove_local_expressions_from_table (void)
512 {
513 /* Remove all the expressions made available in this block. */
514 while (VEC_length (tree, avail_exprs_stack) > 0)
515 {
516 struct expr_hash_elt element;
517 tree expr = VEC_pop (tree, avail_exprs_stack);
518
519 if (expr == NULL_TREE)
520 break;
521
522 initialize_hash_element (expr, NULL, &element);
523 htab_remove_elt_with_hash (avail_exprs, &element, element.hash);
524 }
525 }
526
527 /* Use the source/dest pairs in CONST_AND_COPIES_STACK to restore
528 CONST_AND_COPIES to its original state, stopping when we hit a
529 NULL marker. */
530
531 static void
532 restore_vars_to_original_value (void)
533 {
534 while (VEC_length (tree, const_and_copies_stack) > 0)
535 {
536 tree prev_value, dest;
537
538 dest = VEC_pop (tree, const_and_copies_stack);
539
540 if (dest == NULL)
541 break;
542
543 prev_value = VEC_pop (tree, const_and_copies_stack);
544 SSA_NAME_VALUE (dest) = prev_value;
545 }
546 }
547
548 /* A trivial wrapper so that we can present the generic jump
549 threading code with a simple API for simplifying statements. */
550 static tree
551 simplify_stmt_for_jump_threading (tree stmt)
552 {
553 return lookup_avail_expr (stmt, false);
554 }
555
556 /* Wrapper for common code to attempt to thread an edge. For example,
557 it handles lazily building the dummy condition and the bookkeeping
558 when jump threading is successful. */
559
560 static void
561 dom_thread_across_edge (struct dom_walk_data *walk_data, edge e)
562 {
563 /* If we don't already have a dummy condition, build it now. */
564 if (! walk_data->global_data)
565 {
566 tree dummy_cond = build2 (NE_EXPR, boolean_type_node,
567 integer_zero_node, integer_zero_node);
568 dummy_cond = build3 (COND_EXPR, void_type_node, dummy_cond, NULL, NULL);
569 walk_data->global_data = dummy_cond;
570 }
571
572 thread_across_edge (walk_data->global_data, e, false,
573 &const_and_copies_stack,
574 simplify_stmt_for_jump_threading);
575 }
576
577 /* We have finished processing the dominator children of BB, perform
578 any finalization actions in preparation for leaving this node in
579 the dominator tree. */
580
581 static void
582 dom_opt_finalize_block (struct dom_walk_data *walk_data, basic_block bb)
583 {
584 tree last;
585
586
587 /* If we have an outgoing edge to a block with multiple incoming and
588 outgoing edges, then we may be able to thread the edge. ie, we
589 may be able to statically determine which of the outgoing edges
590 will be traversed when the incoming edge from BB is traversed. */
591 if (single_succ_p (bb)
592 && (single_succ_edge (bb)->flags & EDGE_ABNORMAL) == 0
593 && potentially_threadable_block (single_succ (bb)))
594 {
595 dom_thread_across_edge (walk_data, single_succ_edge (bb));
596 }
597 else if ((last = last_stmt (bb))
598 && TREE_CODE (last) == COND_EXPR
599 && (COMPARISON_CLASS_P (COND_EXPR_COND (last))
600 || TREE_CODE (COND_EXPR_COND (last)) == SSA_NAME)
601 && EDGE_COUNT (bb->succs) == 2
602 && (EDGE_SUCC (bb, 0)->flags & EDGE_ABNORMAL) == 0
603 && (EDGE_SUCC (bb, 1)->flags & EDGE_ABNORMAL) == 0)
604 {
605 edge true_edge, false_edge;
606
607 extract_true_false_edges_from_block (bb, &true_edge, &false_edge);
608
609 /* Only try to thread the edge if it reaches a target block with
610 more than one predecessor and more than one successor. */
611 if (potentially_threadable_block (true_edge->dest))
612 {
613 struct edge_info *edge_info;
614 unsigned int i;
615
616 /* Push a marker onto the available expression stack so that we
617 unwind any expressions related to the TRUE arm before processing
618 the false arm below. */
619 VEC_safe_push (tree, heap, avail_exprs_stack, NULL_TREE);
620 VEC_safe_push (tree, heap, const_and_copies_stack, NULL_TREE);
621
622 edge_info = (struct edge_info *) true_edge->aux;
623
624 /* If we have info associated with this edge, record it into
625 our equivalency tables. */
626 if (edge_info)
627 {
628 tree *cond_equivalences = edge_info->cond_equivalences;
629 tree lhs = edge_info->lhs;
630 tree rhs = edge_info->rhs;
631
632 /* If we have a simple NAME = VALUE equivalency record it. */
633 if (lhs && TREE_CODE (lhs) == SSA_NAME)
634 record_const_or_copy (lhs, rhs);
635
636 /* If we have 0 = COND or 1 = COND equivalences, record them
637 into our expression hash tables. */
638 if (cond_equivalences)
639 for (i = 0; i < edge_info->max_cond_equivalences; i += 2)
640 {
641 tree expr = cond_equivalences[i];
642 tree value = cond_equivalences[i + 1];
643
644 record_cond (expr, value);
645 }
646 }
647
648 dom_thread_across_edge (walk_data, true_edge);
649
650 /* And restore the various tables to their state before
651 we threaded this edge. */
652 remove_local_expressions_from_table ();
653 }
654
655 /* Similarly for the ELSE arm. */
656 if (potentially_threadable_block (false_edge->dest))
657 {
658 struct edge_info *edge_info;
659 unsigned int i;
660
661 VEC_safe_push (tree, heap, const_and_copies_stack, NULL_TREE);
662 edge_info = (struct edge_info *) false_edge->aux;
663
664 /* If we have info associated with this edge, record it into
665 our equivalency tables. */
666 if (edge_info)
667 {
668 tree *cond_equivalences = edge_info->cond_equivalences;
669 tree lhs = edge_info->lhs;
670 tree rhs = edge_info->rhs;
671
672 /* If we have a simple NAME = VALUE equivalency record it. */
673 if (lhs && TREE_CODE (lhs) == SSA_NAME)
674 record_const_or_copy (lhs, rhs);
675
676 /* If we have 0 = COND or 1 = COND equivalences, record them
677 into our expression hash tables. */
678 if (cond_equivalences)
679 for (i = 0; i < edge_info->max_cond_equivalences; i += 2)
680 {
681 tree expr = cond_equivalences[i];
682 tree value = cond_equivalences[i + 1];
683
684 record_cond (expr, value);
685 }
686 }
687
688 /* Now thread the edge. */
689 dom_thread_across_edge (walk_data, false_edge);
690
691 /* No need to remove local expressions from our tables
692 or restore vars to their original value as that will
693 be done immediately below. */
694 }
695 }
696
697 remove_local_expressions_from_table ();
698 restore_vars_to_original_value ();
699
700 /* If we queued any statements to rescan in this block, then
701 go ahead and rescan them now. */
702 while (VEC_length (tree, stmts_to_rescan) > 0)
703 {
704 tree stmt = VEC_last (tree, stmts_to_rescan);
705 basic_block stmt_bb = bb_for_stmt (stmt);
706
707 if (stmt_bb != bb)
708 break;
709
710 VEC_pop (tree, stmts_to_rescan);
711 mark_new_vars_to_rename (stmt);
712 }
713 }
714
715 /* PHI nodes can create equivalences too.
716
717 Ignoring any alternatives which are the same as the result, if
718 all the alternatives are equal, then the PHI node creates an
719 equivalence. */
720
721 static void
722 record_equivalences_from_phis (basic_block bb)
723 {
724 tree phi;
725
726 for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi))
727 {
728 tree lhs = PHI_RESULT (phi);
729 tree rhs = NULL;
730 int i;
731
732 for (i = 0; i < PHI_NUM_ARGS (phi); i++)
733 {
734 tree t = PHI_ARG_DEF (phi, i);
735
736 /* Ignore alternatives which are the same as our LHS. Since
737 LHS is a PHI_RESULT, it is known to be a SSA_NAME, so we
738 can simply compare pointers. */
739 if (lhs == t)
740 continue;
741
742 /* If we have not processed an alternative yet, then set
743 RHS to this alternative. */
744 if (rhs == NULL)
745 rhs = t;
746 /* If we have processed an alternative (stored in RHS), then
747 see if it is equal to this one. If it isn't, then stop
748 the search. */
749 else if (! operand_equal_for_phi_arg_p (rhs, t))
750 break;
751 }
752
753 /* If we had no interesting alternatives, then all the RHS alternatives
754 must have been the same as LHS. */
755 if (!rhs)
756 rhs = lhs;
757
758 /* If we managed to iterate through each PHI alternative without
759 breaking out of the loop, then we have a PHI which may create
760 a useful equivalence. We do not need to record unwind data for
761 this, since this is a true assignment and not an equivalence
762 inferred from a comparison. All uses of this ssa name are dominated
763 by this assignment, so unwinding just costs time and space. */
764 if (i == PHI_NUM_ARGS (phi)
765 && may_propagate_copy (lhs, rhs))
766 SSA_NAME_VALUE (lhs) = rhs;
767 }
768 }
769
770 /* Ignoring loop backedges, if BB has precisely one incoming edge then
771 return that edge. Otherwise return NULL. */
772 static edge
773 single_incoming_edge_ignoring_loop_edges (basic_block bb)
774 {
775 edge retval = NULL;
776 edge e;
777 edge_iterator ei;
778
779 FOR_EACH_EDGE (e, ei, bb->preds)
780 {
781 /* A loop back edge can be identified by the destination of
782 the edge dominating the source of the edge. */
783 if (dominated_by_p (CDI_DOMINATORS, e->src, e->dest))
784 continue;
785
786 /* If we have already seen a non-loop edge, then we must have
787 multiple incoming non-loop edges and thus we return NULL. */
788 if (retval)
789 return NULL;
790
791 /* This is the first non-loop incoming edge we have found. Record
792 it. */
793 retval = e;
794 }
795
796 return retval;
797 }
798
799 /* Record any equivalences created by the incoming edge to BB. If BB
800 has more than one incoming edge, then no equivalence is created. */
801
802 static void
803 record_equivalences_from_incoming_edge (basic_block bb)
804 {
805 edge e;
806 basic_block parent;
807 struct edge_info *edge_info;
808
809 /* If our parent block ended with a control statement, then we may be
810 able to record some equivalences based on which outgoing edge from
811 the parent was followed. */
812 parent = get_immediate_dominator (CDI_DOMINATORS, bb);
813
814 e = single_incoming_edge_ignoring_loop_edges (bb);
815
816 /* If we had a single incoming edge from our parent block, then enter
817 any data associated with the edge into our tables. */
818 if (e && e->src == parent)
819 {
820 unsigned int i;
821
822 edge_info = (struct edge_info *) e->aux;
823
824 if (edge_info)
825 {
826 tree lhs = edge_info->lhs;
827 tree rhs = edge_info->rhs;
828 tree *cond_equivalences = edge_info->cond_equivalences;
829
830 if (lhs)
831 record_equality (lhs, rhs);
832
833 if (cond_equivalences)
834 {
835 for (i = 0; i < edge_info->max_cond_equivalences; i += 2)
836 {
837 tree expr = cond_equivalences[i];
838 tree value = cond_equivalences[i + 1];
839
840 record_cond (expr, value);
841 }
842 }
843 }
844 }
845 }
846
847 /* Dump SSA statistics on FILE. */
848
849 void
850 dump_dominator_optimization_stats (FILE *file)
851 {
852 long n_exprs;
853
854 fprintf (file, "Total number of statements: %6ld\n\n",
855 opt_stats.num_stmts);
856 fprintf (file, "Exprs considered for dominator optimizations: %6ld\n",
857 opt_stats.num_exprs_considered);
858
859 n_exprs = opt_stats.num_exprs_considered;
860 if (n_exprs == 0)
861 n_exprs = 1;
862
863 fprintf (file, " Redundant expressions eliminated: %6ld (%.0f%%)\n",
864 opt_stats.num_re, PERCENT (opt_stats.num_re,
865 n_exprs));
866 fprintf (file, " Constants propagated: %6ld\n",
867 opt_stats.num_const_prop);
868 fprintf (file, " Copies propagated: %6ld\n",
869 opt_stats.num_copy_prop);
870
871 fprintf (file, "\nHash table statistics:\n");
872
873 fprintf (file, " avail_exprs: ");
874 htab_statistics (file, avail_exprs);
875 }
876
877
878 /* Dump SSA statistics on stderr. */
879
880 void
881 debug_dominator_optimization_stats (void)
882 {
883 dump_dominator_optimization_stats (stderr);
884 }
885
886
887 /* Dump statistics for the hash table HTAB. */
888
889 static void
890 htab_statistics (FILE *file, htab_t htab)
891 {
892 fprintf (file, "size %ld, %ld elements, %f collision/search ratio\n",
893 (long) htab_size (htab),
894 (long) htab_elements (htab),
895 htab_collisions (htab));
896 }
897
898 /* Enter a statement into the true/false expression hash table indicating
899 that the condition COND has the value VALUE. */
900
901 static void
902 record_cond (tree cond, tree value)
903 {
904 struct expr_hash_elt *element = XCNEW (struct expr_hash_elt);
905 void **slot;
906
907 initialize_hash_element (cond, value, element);
908
909 slot = htab_find_slot_with_hash (avail_exprs, (void *)element,
910 element->hash, INSERT);
911 if (*slot == NULL)
912 {
913 *slot = (void *) element;
914 VEC_safe_push (tree, heap, avail_exprs_stack, cond);
915 }
916 else
917 free (element);
918 }
919
920 /* Build a new conditional using NEW_CODE, OP0 and OP1 and store
921 the new conditional into *p, then store a boolean_true_node
922 into *(p + 1). */
923
924 static void
925 build_and_record_new_cond (enum tree_code new_code, tree op0, tree op1, tree *p)
926 {
927 *p = build2 (new_code, boolean_type_node, op0, op1);
928 p++;
929 *p = boolean_true_node;
930 }
931
932 /* Record that COND is true and INVERTED is false into the edge information
933 structure. Also record that any conditions dominated by COND are true
934 as well.
935
936 For example, if a < b is true, then a <= b must also be true. */
937
938 static void
939 record_conditions (struct edge_info *edge_info, tree cond, tree inverted)
940 {
941 tree op0, op1;
942
943 if (!COMPARISON_CLASS_P (cond))
944 return;
945
946 op0 = TREE_OPERAND (cond, 0);
947 op1 = TREE_OPERAND (cond, 1);
948
949 switch (TREE_CODE (cond))
950 {
951 case LT_EXPR:
952 case GT_EXPR:
953 edge_info->max_cond_equivalences = 12;
954 edge_info->cond_equivalences = XNEWVEC (tree, 12);
955 build_and_record_new_cond ((TREE_CODE (cond) == LT_EXPR
956 ? LE_EXPR : GE_EXPR),
957 op0, op1, &edge_info->cond_equivalences[4]);
958 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
959 &edge_info->cond_equivalences[6]);
960 build_and_record_new_cond (NE_EXPR, op0, op1,
961 &edge_info->cond_equivalences[8]);
962 build_and_record_new_cond (LTGT_EXPR, op0, op1,
963 &edge_info->cond_equivalences[10]);
964 break;
965
966 case GE_EXPR:
967 case LE_EXPR:
968 edge_info->max_cond_equivalences = 6;
969 edge_info->cond_equivalences = XNEWVEC (tree, 6);
970 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
971 &edge_info->cond_equivalences[4]);
972 break;
973
974 case EQ_EXPR:
975 edge_info->max_cond_equivalences = 10;
976 edge_info->cond_equivalences = XNEWVEC (tree, 10);
977 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
978 &edge_info->cond_equivalences[4]);
979 build_and_record_new_cond (LE_EXPR, op0, op1,
980 &edge_info->cond_equivalences[6]);
981 build_and_record_new_cond (GE_EXPR, op0, op1,
982 &edge_info->cond_equivalences[8]);
983 break;
984
985 case UNORDERED_EXPR:
986 edge_info->max_cond_equivalences = 16;
987 edge_info->cond_equivalences = XNEWVEC (tree, 16);
988 build_and_record_new_cond (NE_EXPR, op0, op1,
989 &edge_info->cond_equivalences[4]);
990 build_and_record_new_cond (UNLE_EXPR, op0, op1,
991 &edge_info->cond_equivalences[6]);
992 build_and_record_new_cond (UNGE_EXPR, op0, op1,
993 &edge_info->cond_equivalences[8]);
994 build_and_record_new_cond (UNEQ_EXPR, op0, op1,
995 &edge_info->cond_equivalences[10]);
996 build_and_record_new_cond (UNLT_EXPR, op0, op1,
997 &edge_info->cond_equivalences[12]);
998 build_and_record_new_cond (UNGT_EXPR, op0, op1,
999 &edge_info->cond_equivalences[14]);
1000 break;
1001
1002 case UNLT_EXPR:
1003 case UNGT_EXPR:
1004 edge_info->max_cond_equivalences = 8;
1005 edge_info->cond_equivalences = XNEWVEC (tree, 8);
1006 build_and_record_new_cond ((TREE_CODE (cond) == UNLT_EXPR
1007 ? UNLE_EXPR : UNGE_EXPR),
1008 op0, op1, &edge_info->cond_equivalences[4]);
1009 build_and_record_new_cond (NE_EXPR, op0, op1,
1010 &edge_info->cond_equivalences[6]);
1011 break;
1012
1013 case UNEQ_EXPR:
1014 edge_info->max_cond_equivalences = 8;
1015 edge_info->cond_equivalences = XNEWVEC (tree, 8);
1016 build_and_record_new_cond (UNLE_EXPR, op0, op1,
1017 &edge_info->cond_equivalences[4]);
1018 build_and_record_new_cond (UNGE_EXPR, op0, op1,
1019 &edge_info->cond_equivalences[6]);
1020 break;
1021
1022 case LTGT_EXPR:
1023 edge_info->max_cond_equivalences = 8;
1024 edge_info->cond_equivalences = XNEWVEC (tree, 8);
1025 build_and_record_new_cond (NE_EXPR, op0, op1,
1026 &edge_info->cond_equivalences[4]);
1027 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1028 &edge_info->cond_equivalences[6]);
1029 break;
1030
1031 default:
1032 edge_info->max_cond_equivalences = 4;
1033 edge_info->cond_equivalences = XNEWVEC (tree, 4);
1034 break;
1035 }
1036
1037 /* Now store the original true and false conditions into the first
1038 two slots. */
1039 edge_info->cond_equivalences[0] = cond;
1040 edge_info->cond_equivalences[1] = boolean_true_node;
1041 edge_info->cond_equivalences[2] = inverted;
1042 edge_info->cond_equivalences[3] = boolean_false_node;
1043 }
1044
1045 /* A helper function for record_const_or_copy and record_equality.
1046 Do the work of recording the value and undo info. */
1047
1048 static void
1049 record_const_or_copy_1 (tree x, tree y, tree prev_x)
1050 {
1051 SSA_NAME_VALUE (x) = y;
1052
1053 VEC_reserve (tree, heap, const_and_copies_stack, 2);
1054 VEC_quick_push (tree, const_and_copies_stack, prev_x);
1055 VEC_quick_push (tree, const_and_copies_stack, x);
1056 }
1057
1058
1059 /* Return the loop depth of the basic block of the defining statement of X.
1060 This number should not be treated as absolutely correct because the loop
1061 information may not be completely up-to-date when dom runs. However, it
1062 will be relatively correct, and as more passes are taught to keep loop info
1063 up to date, the result will become more and more accurate. */
1064
1065 int
1066 loop_depth_of_name (tree x)
1067 {
1068 tree defstmt;
1069 basic_block defbb;
1070
1071 /* If it's not an SSA_NAME, we have no clue where the definition is. */
1072 if (TREE_CODE (x) != SSA_NAME)
1073 return 0;
1074
1075 /* Otherwise return the loop depth of the defining statement's bb.
1076 Note that there may not actually be a bb for this statement, if the
1077 ssa_name is live on entry. */
1078 defstmt = SSA_NAME_DEF_STMT (x);
1079 defbb = bb_for_stmt (defstmt);
1080 if (!defbb)
1081 return 0;
1082
1083 return defbb->loop_depth;
1084 }
1085
1086
1087 /* Record that X is equal to Y in const_and_copies. Record undo
1088 information in the block-local vector. */
1089
1090 static void
1091 record_const_or_copy (tree x, tree y)
1092 {
1093 tree prev_x = SSA_NAME_VALUE (x);
1094
1095 if (TREE_CODE (y) == SSA_NAME)
1096 {
1097 tree tmp = SSA_NAME_VALUE (y);
1098 if (tmp)
1099 y = tmp;
1100 }
1101
1102 record_const_or_copy_1 (x, y, prev_x);
1103 }
1104
1105 /* Similarly, but assume that X and Y are the two operands of an EQ_EXPR.
1106 This constrains the cases in which we may treat this as assignment. */
1107
1108 static void
1109 record_equality (tree x, tree y)
1110 {
1111 tree prev_x = NULL, prev_y = NULL;
1112
1113 if (TREE_CODE (x) == SSA_NAME)
1114 prev_x = SSA_NAME_VALUE (x);
1115 if (TREE_CODE (y) == SSA_NAME)
1116 prev_y = SSA_NAME_VALUE (y);
1117
1118 /* If one of the previous values is invariant, or invariant in more loops
1119 (by depth), then use that.
1120 Otherwise it doesn't matter which value we choose, just so
1121 long as we canonicalize on one value. */
1122 if (TREE_INVARIANT (y))
1123 ;
1124 else if (TREE_INVARIANT (x) || (loop_depth_of_name (x) <= loop_depth_of_name (y)))
1125 prev_x = x, x = y, y = prev_x, prev_x = prev_y;
1126 else if (prev_x && TREE_INVARIANT (prev_x))
1127 x = y, y = prev_x, prev_x = prev_y;
1128 else if (prev_y && TREE_CODE (prev_y) != VALUE_HANDLE)
1129 y = prev_y;
1130
1131 /* After the swapping, we must have one SSA_NAME. */
1132 if (TREE_CODE (x) != SSA_NAME)
1133 return;
1134
1135 /* For IEEE, -0.0 == 0.0, so we don't necessarily know the sign of a
1136 variable compared against zero. If we're honoring signed zeros,
1137 then we cannot record this value unless we know that the value is
1138 nonzero. */
1139 if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (x)))
1140 && (TREE_CODE (y) != REAL_CST
1141 || REAL_VALUES_EQUAL (dconst0, TREE_REAL_CST (y))))
1142 return;
1143
1144 record_const_or_copy_1 (x, y, prev_x);
1145 }
1146
1147 /* Returns true when STMT is a simple iv increment. It detects the
1148 following situation:
1149
1150 i_1 = phi (..., i_2)
1151 i_2 = i_1 +/- ... */
1152
1153 static bool
1154 simple_iv_increment_p (tree stmt)
1155 {
1156 tree lhs, rhs, preinc, phi;
1157 unsigned i;
1158
1159 if (TREE_CODE (stmt) != MODIFY_EXPR)
1160 return false;
1161
1162 lhs = TREE_OPERAND (stmt, 0);
1163 if (TREE_CODE (lhs) != SSA_NAME)
1164 return false;
1165
1166 rhs = TREE_OPERAND (stmt, 1);
1167
1168 if (TREE_CODE (rhs) != PLUS_EXPR
1169 && TREE_CODE (rhs) != MINUS_EXPR)
1170 return false;
1171
1172 preinc = TREE_OPERAND (rhs, 0);
1173 if (TREE_CODE (preinc) != SSA_NAME)
1174 return false;
1175
1176 phi = SSA_NAME_DEF_STMT (preinc);
1177 if (TREE_CODE (phi) != PHI_NODE)
1178 return false;
1179
1180 for (i = 0; i < (unsigned) PHI_NUM_ARGS (phi); i++)
1181 if (PHI_ARG_DEF (phi, i) == lhs)
1182 return true;
1183
1184 return false;
1185 }
1186
1187 /* CONST_AND_COPIES is a table which maps an SSA_NAME to the current
1188 known value for that SSA_NAME (or NULL if no value is known).
1189
1190 Propagate values from CONST_AND_COPIES into the PHI nodes of the
1191 successors of BB. */
1192
1193 static void
1194 cprop_into_successor_phis (basic_block bb)
1195 {
1196 edge e;
1197 edge_iterator ei;
1198
1199 FOR_EACH_EDGE (e, ei, bb->succs)
1200 {
1201 tree phi;
1202 int indx;
1203
1204 /* If this is an abnormal edge, then we do not want to copy propagate
1205 into the PHI alternative associated with this edge. */
1206 if (e->flags & EDGE_ABNORMAL)
1207 continue;
1208
1209 phi = phi_nodes (e->dest);
1210 if (! phi)
1211 continue;
1212
1213 indx = e->dest_idx;
1214 for ( ; phi; phi = PHI_CHAIN (phi))
1215 {
1216 tree new;
1217 use_operand_p orig_p;
1218 tree orig;
1219
1220 /* The alternative may be associated with a constant, so verify
1221 it is an SSA_NAME before doing anything with it. */
1222 orig_p = PHI_ARG_DEF_PTR (phi, indx);
1223 orig = USE_FROM_PTR (orig_p);
1224 if (TREE_CODE (orig) != SSA_NAME)
1225 continue;
1226
1227 /* If we have *ORIG_P in our constant/copy table, then replace
1228 ORIG_P with its value in our constant/copy table. */
1229 new = SSA_NAME_VALUE (orig);
1230 if (new
1231 && new != orig
1232 && (TREE_CODE (new) == SSA_NAME
1233 || is_gimple_min_invariant (new))
1234 && may_propagate_copy (orig, new))
1235 propagate_value (orig_p, new);
1236 }
1237 }
1238 }
1239
1240 /* We have finished optimizing BB, record any information implied by
1241 taking a specific outgoing edge from BB. */
1242
1243 static void
1244 record_edge_info (basic_block bb)
1245 {
1246 block_stmt_iterator bsi = bsi_last (bb);
1247 struct edge_info *edge_info;
1248
1249 if (! bsi_end_p (bsi))
1250 {
1251 tree stmt = bsi_stmt (bsi);
1252
1253 if (stmt && TREE_CODE (stmt) == SWITCH_EXPR)
1254 {
1255 tree cond = SWITCH_COND (stmt);
1256
1257 if (TREE_CODE (cond) == SSA_NAME)
1258 {
1259 tree labels = SWITCH_LABELS (stmt);
1260 int i, n_labels = TREE_VEC_LENGTH (labels);
1261 tree *info = XCNEWVEC (tree, last_basic_block);
1262 edge e;
1263 edge_iterator ei;
1264
1265 for (i = 0; i < n_labels; i++)
1266 {
1267 tree label = TREE_VEC_ELT (labels, i);
1268 basic_block target_bb = label_to_block (CASE_LABEL (label));
1269
1270 if (CASE_HIGH (label)
1271 || !CASE_LOW (label)
1272 || info[target_bb->index])
1273 info[target_bb->index] = error_mark_node;
1274 else
1275 info[target_bb->index] = label;
1276 }
1277
1278 FOR_EACH_EDGE (e, ei, bb->succs)
1279 {
1280 basic_block target_bb = e->dest;
1281 tree node = info[target_bb->index];
1282
1283 if (node != NULL && node != error_mark_node)
1284 {
1285 tree x = fold_convert (TREE_TYPE (cond), CASE_LOW (node));
1286 edge_info = allocate_edge_info (e);
1287 edge_info->lhs = cond;
1288 edge_info->rhs = x;
1289 }
1290 }
1291 free (info);
1292 }
1293 }
1294
1295 /* A COND_EXPR may create equivalences too. */
1296 if (stmt && TREE_CODE (stmt) == COND_EXPR)
1297 {
1298 tree cond = COND_EXPR_COND (stmt);
1299 edge true_edge;
1300 edge false_edge;
1301
1302 extract_true_false_edges_from_block (bb, &true_edge, &false_edge);
1303
1304 /* If the conditional is a single variable 'X', record 'X = 1'
1305 for the true edge and 'X = 0' on the false edge. */
1306 if (SSA_VAR_P (cond))
1307 {
1308 struct edge_info *edge_info;
1309
1310 edge_info = allocate_edge_info (true_edge);
1311 edge_info->lhs = cond;
1312 edge_info->rhs = constant_boolean_node (1, TREE_TYPE (cond));
1313
1314 edge_info = allocate_edge_info (false_edge);
1315 edge_info->lhs = cond;
1316 edge_info->rhs = constant_boolean_node (0, TREE_TYPE (cond));
1317 }
1318 /* Equality tests may create one or two equivalences. */
1319 else if (COMPARISON_CLASS_P (cond))
1320 {
1321 tree op0 = TREE_OPERAND (cond, 0);
1322 tree op1 = TREE_OPERAND (cond, 1);
1323
1324 /* Special case comparing booleans against a constant as we
1325 know the value of OP0 on both arms of the branch. i.e., we
1326 can record an equivalence for OP0 rather than COND. */
1327 if ((TREE_CODE (cond) == EQ_EXPR || TREE_CODE (cond) == NE_EXPR)
1328 && TREE_CODE (op0) == SSA_NAME
1329 && TREE_CODE (TREE_TYPE (op0)) == BOOLEAN_TYPE
1330 && is_gimple_min_invariant (op1))
1331 {
1332 if (TREE_CODE (cond) == EQ_EXPR)
1333 {
1334 edge_info = allocate_edge_info (true_edge);
1335 edge_info->lhs = op0;
1336 edge_info->rhs = (integer_zerop (op1)
1337 ? boolean_false_node
1338 : boolean_true_node);
1339
1340 edge_info = allocate_edge_info (false_edge);
1341 edge_info->lhs = op0;
1342 edge_info->rhs = (integer_zerop (op1)
1343 ? boolean_true_node
1344 : boolean_false_node);
1345 }
1346 else
1347 {
1348 edge_info = allocate_edge_info (true_edge);
1349 edge_info->lhs = op0;
1350 edge_info->rhs = (integer_zerop (op1)
1351 ? boolean_true_node
1352 : boolean_false_node);
1353
1354 edge_info = allocate_edge_info (false_edge);
1355 edge_info->lhs = op0;
1356 edge_info->rhs = (integer_zerop (op1)
1357 ? boolean_false_node
1358 : boolean_true_node);
1359 }
1360 }
1361
1362 else if (is_gimple_min_invariant (op0)
1363 && (TREE_CODE (op1) == SSA_NAME
1364 || is_gimple_min_invariant (op1)))
1365 {
1366 tree inverted = invert_truthvalue (cond);
1367 struct edge_info *edge_info;
1368
1369 edge_info = allocate_edge_info (true_edge);
1370 record_conditions (edge_info, cond, inverted);
1371
1372 if (TREE_CODE (cond) == EQ_EXPR)
1373 {
1374 edge_info->lhs = op1;
1375 edge_info->rhs = op0;
1376 }
1377
1378 edge_info = allocate_edge_info (false_edge);
1379 record_conditions (edge_info, inverted, cond);
1380
1381 if (TREE_CODE (cond) == NE_EXPR)
1382 {
1383 edge_info->lhs = op1;
1384 edge_info->rhs = op0;
1385 }
1386 }
1387
1388 else if (TREE_CODE (op0) == SSA_NAME
1389 && (is_gimple_min_invariant (op1)
1390 || TREE_CODE (op1) == SSA_NAME))
1391 {
1392 tree inverted = invert_truthvalue (cond);
1393 struct edge_info *edge_info;
1394
1395 edge_info = allocate_edge_info (true_edge);
1396 record_conditions (edge_info, cond, inverted);
1397
1398 if (TREE_CODE (cond) == EQ_EXPR)
1399 {
1400 edge_info->lhs = op0;
1401 edge_info->rhs = op1;
1402 }
1403
1404 edge_info = allocate_edge_info (false_edge);
1405 record_conditions (edge_info, inverted, cond);
1406
1407 if (TREE_CODE (cond) == NE_EXPR)
1408 {
1409 edge_info->lhs = op0;
1410 edge_info->rhs = op1;
1411 }
1412 }
1413 }
1414
1415 /* ??? TRUTH_NOT_EXPR can create an equivalence too. */
1416 }
1417 }
1418 }
1419
1420 /* Propagate information from BB to its outgoing edges.
1421
1422 This can include equivalency information implied by control statements
1423 at the end of BB and const/copy propagation into PHIs in BB's
1424 successor blocks. */
1425
1426 static void
1427 propagate_to_outgoing_edges (struct dom_walk_data *walk_data ATTRIBUTE_UNUSED,
1428 basic_block bb)
1429 {
1430 record_edge_info (bb);
1431 cprop_into_successor_phis (bb);
1432 }
1433
1434 /* Search for redundant computations in STMT. If any are found, then
1435 replace them with the variable holding the result of the computation.
1436
1437 If safe, record this expression into the available expression hash
1438 table. */
1439
1440 static bool
1441 eliminate_redundant_computations (tree stmt)
1442 {
1443 tree *expr_p, def = NULL_TREE;
1444 bool insert = true;
1445 tree cached_lhs;
1446 bool retval = false;
1447 bool modify_expr_p = false;
1448
1449 if (TREE_CODE (stmt) == MODIFY_EXPR)
1450 def = TREE_OPERAND (stmt, 0);
1451
1452 /* Certain expressions on the RHS can be optimized away, but can not
1453 themselves be entered into the hash tables. */
1454 if (! def
1455 || TREE_CODE (def) != SSA_NAME
1456 || SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def)
1457 || !ZERO_SSA_OPERANDS (stmt, SSA_OP_VMAYDEF)
1458 /* Do not record equivalences for increments of ivs. This would create
1459 overlapping live ranges for a very questionable gain. */
1460 || simple_iv_increment_p (stmt))
1461 insert = false;
1462
1463 /* Check if the expression has been computed before. */
1464 cached_lhs = lookup_avail_expr (stmt, insert);
1465
1466 opt_stats.num_exprs_considered++;
1467
1468 /* Get a pointer to the expression we are trying to optimize. */
1469 if (TREE_CODE (stmt) == COND_EXPR)
1470 expr_p = &COND_EXPR_COND (stmt);
1471 else if (TREE_CODE (stmt) == SWITCH_EXPR)
1472 expr_p = &SWITCH_COND (stmt);
1473 else if (TREE_CODE (stmt) == RETURN_EXPR && TREE_OPERAND (stmt, 0))
1474 {
1475 expr_p = &TREE_OPERAND (TREE_OPERAND (stmt, 0), 1);
1476 modify_expr_p = true;
1477 }
1478 else
1479 {
1480 expr_p = &TREE_OPERAND (stmt, 1);
1481 modify_expr_p = true;
1482 }
1483
1484 /* It is safe to ignore types here since we have already done
1485 type checking in the hashing and equality routines. In fact
1486 type checking here merely gets in the way of constant
1487 propagation. Also, make sure that it is safe to propagate
1488 CACHED_LHS into *EXPR_P. */
1489 if (cached_lhs
1490 && ((TREE_CODE (cached_lhs) != SSA_NAME
1491 && (modify_expr_p
1492 || tree_ssa_useless_type_conversion_1 (TREE_TYPE (*expr_p),
1493 TREE_TYPE (cached_lhs))))
1494 || may_propagate_copy (*expr_p, cached_lhs)))
1495 {
1496 if (dump_file && (dump_flags & TDF_DETAILS))
1497 {
1498 fprintf (dump_file, " Replaced redundant expr '");
1499 print_generic_expr (dump_file, *expr_p, dump_flags);
1500 fprintf (dump_file, "' with '");
1501 print_generic_expr (dump_file, cached_lhs, dump_flags);
1502 fprintf (dump_file, "'\n");
1503 }
1504
1505 opt_stats.num_re++;
1506
1507 #if defined ENABLE_CHECKING
1508 gcc_assert (TREE_CODE (cached_lhs) == SSA_NAME
1509 || is_gimple_min_invariant (cached_lhs));
1510 #endif
1511
1512 if (TREE_CODE (cached_lhs) == ADDR_EXPR
1513 || (POINTER_TYPE_P (TREE_TYPE (*expr_p))
1514 && is_gimple_min_invariant (cached_lhs)))
1515 retval = true;
1516
1517 if (modify_expr_p
1518 && !tree_ssa_useless_type_conversion_1 (TREE_TYPE (*expr_p),
1519 TREE_TYPE (cached_lhs)))
1520 cached_lhs = fold_convert (TREE_TYPE (*expr_p), cached_lhs);
1521
1522 propagate_tree_value (expr_p, cached_lhs);
1523 mark_stmt_modified (stmt);
1524 }
1525 return retval;
1526 }
1527
1528 /* STMT, a MODIFY_EXPR, may create certain equivalences, in either
1529 the available expressions table or the const_and_copies table.
1530 Detect and record those equivalences. */
1531
1532 static void
1533 record_equivalences_from_stmt (tree stmt,
1534 int may_optimize_p,
1535 stmt_ann_t ann)
1536 {
1537 tree lhs = TREE_OPERAND (stmt, 0);
1538 enum tree_code lhs_code = TREE_CODE (lhs);
1539
1540 if (lhs_code == SSA_NAME)
1541 {
1542 tree rhs = TREE_OPERAND (stmt, 1);
1543
1544 /* Strip away any useless type conversions. */
1545 STRIP_USELESS_TYPE_CONVERSION (rhs);
1546
1547 /* If the RHS of the assignment is a constant or another variable that
1548 may be propagated, register it in the CONST_AND_COPIES table. We
1549 do not need to record unwind data for this, since this is a true
1550 assignment and not an equivalence inferred from a comparison. All
1551 uses of this ssa name are dominated by this assignment, so unwinding
1552 just costs time and space. */
1553 if (may_optimize_p
1554 && (TREE_CODE (rhs) == SSA_NAME
1555 || is_gimple_min_invariant (rhs)))
1556 SSA_NAME_VALUE (lhs) = rhs;
1557 }
1558
1559 /* A memory store, even an aliased store, creates a useful
1560 equivalence. By exchanging the LHS and RHS, creating suitable
1561 vops and recording the result in the available expression table,
1562 we may be able to expose more redundant loads. */
1563 if (!ann->has_volatile_ops
1564 && (TREE_CODE (TREE_OPERAND (stmt, 1)) == SSA_NAME
1565 || is_gimple_min_invariant (TREE_OPERAND (stmt, 1)))
1566 && !is_gimple_reg (lhs))
1567 {
1568 tree rhs = TREE_OPERAND (stmt, 1);
1569 tree new;
1570
1571 /* FIXME: If the LHS of the assignment is a bitfield and the RHS
1572 is a constant, we need to adjust the constant to fit into the
1573 type of the LHS. If the LHS is a bitfield and the RHS is not
1574 a constant, then we can not record any equivalences for this
1575 statement since we would need to represent the widening or
1576 narrowing of RHS. This fixes gcc.c-torture/execute/921016-1.c
1577 and should not be necessary if GCC represented bitfields
1578 properly. */
1579 if (lhs_code == COMPONENT_REF
1580 && DECL_BIT_FIELD (TREE_OPERAND (lhs, 1)))
1581 {
1582 if (TREE_CONSTANT (rhs))
1583 rhs = widen_bitfield (rhs, TREE_OPERAND (lhs, 1), lhs);
1584 else
1585 rhs = NULL;
1586
1587 /* If the value overflowed, then we can not use this equivalence. */
1588 if (rhs && ! is_gimple_min_invariant (rhs))
1589 rhs = NULL;
1590 }
1591
1592 if (rhs)
1593 {
1594 /* Build a new statement with the RHS and LHS exchanged. */
1595 new = build2 (MODIFY_EXPR, TREE_TYPE (stmt), rhs, lhs);
1596
1597 create_ssa_artficial_load_stmt (new, stmt);
1598
1599 /* Finally enter the statement into the available expression
1600 table. */
1601 lookup_avail_expr (new, true);
1602 }
1603 }
1604 }
1605
1606 /* Replace *OP_P in STMT with any known equivalent value for *OP_P from
1607 CONST_AND_COPIES. */
1608
1609 static bool
1610 cprop_operand (tree stmt, use_operand_p op_p)
1611 {
1612 bool may_have_exposed_new_symbols = false;
1613 tree val;
1614 tree op = USE_FROM_PTR (op_p);
1615
1616 /* If the operand has a known constant value or it is known to be a
1617 copy of some other variable, use the value or copy stored in
1618 CONST_AND_COPIES. */
1619 val = SSA_NAME_VALUE (op);
1620 if (val && val != op && TREE_CODE (val) != VALUE_HANDLE)
1621 {
1622 tree op_type, val_type;
1623
1624 /* Do not change the base variable in the virtual operand
1625 tables. That would make it impossible to reconstruct
1626 the renamed virtual operand if we later modify this
1627 statement. Also only allow the new value to be an SSA_NAME
1628 for propagation into virtual operands. */
1629 if (!is_gimple_reg (op)
1630 && (TREE_CODE (val) != SSA_NAME
1631 || is_gimple_reg (val)
1632 || get_virtual_var (val) != get_virtual_var (op)))
1633 return false;
1634
1635 /* Do not replace hard register operands in asm statements. */
1636 if (TREE_CODE (stmt) == ASM_EXPR
1637 && !may_propagate_copy_into_asm (op))
1638 return false;
1639
1640 /* Get the toplevel type of each operand. */
1641 op_type = TREE_TYPE (op);
1642 val_type = TREE_TYPE (val);
1643
1644 /* While both types are pointers, get the type of the object
1645 pointed to. */
1646 while (POINTER_TYPE_P (op_type) && POINTER_TYPE_P (val_type))
1647 {
1648 op_type = TREE_TYPE (op_type);
1649 val_type = TREE_TYPE (val_type);
1650 }
1651
1652 /* Make sure underlying types match before propagating a constant by
1653 converting the constant to the proper type. Note that convert may
1654 return a non-gimple expression, in which case we ignore this
1655 propagation opportunity. */
1656 if (TREE_CODE (val) != SSA_NAME)
1657 {
1658 if (!lang_hooks.types_compatible_p (op_type, val_type))
1659 {
1660 val = fold_convert (TREE_TYPE (op), val);
1661 if (!is_gimple_min_invariant (val))
1662 return false;
1663 }
1664 }
1665
1666 /* Certain operands are not allowed to be copy propagated due
1667 to their interaction with exception handling and some GCC
1668 extensions. */
1669 else if (!may_propagate_copy (op, val))
1670 return false;
1671
1672 /* Do not propagate copies if the propagated value is at a deeper loop
1673 depth than the propagatee. Otherwise, this may move loop variant
1674 variables outside of their loops and prevent coalescing
1675 opportunities. If the value was loop invariant, it will be hoisted
1676 by LICM and exposed for copy propagation. */
1677 if (loop_depth_of_name (val) > loop_depth_of_name (op))
1678 return false;
1679
1680 /* Dump details. */
1681 if (dump_file && (dump_flags & TDF_DETAILS))
1682 {
1683 fprintf (dump_file, " Replaced '");
1684 print_generic_expr (dump_file, op, dump_flags);
1685 fprintf (dump_file, "' with %s '",
1686 (TREE_CODE (val) != SSA_NAME ? "constant" : "variable"));
1687 print_generic_expr (dump_file, val, dump_flags);
1688 fprintf (dump_file, "'\n");
1689 }
1690
1691 /* If VAL is an ADDR_EXPR or a constant of pointer type, note
1692 that we may have exposed a new symbol for SSA renaming. */
1693 if (TREE_CODE (val) == ADDR_EXPR
1694 || (POINTER_TYPE_P (TREE_TYPE (op))
1695 && is_gimple_min_invariant (val)))
1696 may_have_exposed_new_symbols = true;
1697
1698 if (TREE_CODE (val) != SSA_NAME)
1699 opt_stats.num_const_prop++;
1700 else
1701 opt_stats.num_copy_prop++;
1702
1703 propagate_value (op_p, val);
1704
1705 /* And note that we modified this statement. This is now
1706 safe, even if we changed virtual operands since we will
1707 rescan the statement and rewrite its operands again. */
1708 mark_stmt_modified (stmt);
1709 }
1710 return may_have_exposed_new_symbols;
1711 }
1712
1713 /* CONST_AND_COPIES is a table which maps an SSA_NAME to the current
1714 known value for that SSA_NAME (or NULL if no value is known).
1715
1716 Propagate values from CONST_AND_COPIES into the uses, vuses and
1717 v_may_def_ops of STMT. */
1718
1719 static bool
1720 cprop_into_stmt (tree stmt)
1721 {
1722 bool may_have_exposed_new_symbols = false;
1723 use_operand_p op_p;
1724 ssa_op_iter iter;
1725
1726 FOR_EACH_SSA_USE_OPERAND (op_p, stmt, iter, SSA_OP_ALL_USES)
1727 {
1728 if (TREE_CODE (USE_FROM_PTR (op_p)) == SSA_NAME)
1729 may_have_exposed_new_symbols |= cprop_operand (stmt, op_p);
1730 }
1731
1732 return may_have_exposed_new_symbols;
1733 }
1734
1735
1736 /* Optimize the statement pointed to by iterator SI.
1737
1738 We try to perform some simplistic global redundancy elimination and
1739 constant propagation:
1740
1741 1- To detect global redundancy, we keep track of expressions that have
1742 been computed in this block and its dominators. If we find that the
1743 same expression is computed more than once, we eliminate repeated
1744 computations by using the target of the first one.
1745
1746 2- Constant values and copy assignments. This is used to do very
1747 simplistic constant and copy propagation. When a constant or copy
1748 assignment is found, we map the value on the RHS of the assignment to
1749 the variable in the LHS in the CONST_AND_COPIES table. */
1750
1751 static void
1752 optimize_stmt (struct dom_walk_data *walk_data ATTRIBUTE_UNUSED,
1753 basic_block bb, block_stmt_iterator si)
1754 {
1755 stmt_ann_t ann;
1756 tree stmt, old_stmt;
1757 bool may_optimize_p;
1758 bool may_have_exposed_new_symbols = false;
1759
1760 old_stmt = stmt = bsi_stmt (si);
1761
1762 if (TREE_CODE (stmt) == COND_EXPR)
1763 canonicalize_comparison (stmt);
1764
1765 update_stmt_if_modified (stmt);
1766 ann = stmt_ann (stmt);
1767 opt_stats.num_stmts++;
1768 may_have_exposed_new_symbols = false;
1769
1770 if (dump_file && (dump_flags & TDF_DETAILS))
1771 {
1772 fprintf (dump_file, "Optimizing statement ");
1773 print_generic_stmt (dump_file, stmt, TDF_SLIM);
1774 }
1775
1776 /* Const/copy propagate into USES, VUSES and the RHS of V_MAY_DEFs. */
1777 may_have_exposed_new_symbols = cprop_into_stmt (stmt);
1778
1779 /* If the statement has been modified with constant replacements,
1780 fold its RHS before checking for redundant computations. */
1781 if (ann->modified)
1782 {
1783 tree rhs;
1784
1785 /* Try to fold the statement making sure that STMT is kept
1786 up to date. */
1787 if (fold_stmt (bsi_stmt_ptr (si)))
1788 {
1789 stmt = bsi_stmt (si);
1790 ann = stmt_ann (stmt);
1791
1792 if (dump_file && (dump_flags & TDF_DETAILS))
1793 {
1794 fprintf (dump_file, " Folded to: ");
1795 print_generic_stmt (dump_file, stmt, TDF_SLIM);
1796 }
1797 }
1798
1799 rhs = get_rhs (stmt);
1800 if (rhs && TREE_CODE (rhs) == ADDR_EXPR)
1801 recompute_tree_invariant_for_addr_expr (rhs);
1802
1803 /* Constant/copy propagation above may change the set of
1804 virtual operands associated with this statement. Folding
1805 may remove the need for some virtual operands.
1806
1807 Indicate we will need to rescan and rewrite the statement. */
1808 may_have_exposed_new_symbols = true;
1809 }
1810
1811 /* Check for redundant computations. Do this optimization only
1812 for assignments that have no volatile ops and conditionals. */
1813 may_optimize_p = (!ann->has_volatile_ops
1814 && ((TREE_CODE (stmt) == RETURN_EXPR
1815 && TREE_OPERAND (stmt, 0)
1816 && TREE_CODE (TREE_OPERAND (stmt, 0)) == MODIFY_EXPR
1817 && ! (TREE_SIDE_EFFECTS
1818 (TREE_OPERAND (TREE_OPERAND (stmt, 0), 1))))
1819 || (TREE_CODE (stmt) == MODIFY_EXPR
1820 && ! TREE_SIDE_EFFECTS (TREE_OPERAND (stmt, 1)))
1821 || TREE_CODE (stmt) == COND_EXPR
1822 || TREE_CODE (stmt) == SWITCH_EXPR));
1823
1824 if (may_optimize_p)
1825 may_have_exposed_new_symbols |= eliminate_redundant_computations (stmt);
1826
1827 /* Record any additional equivalences created by this statement. */
1828 if (TREE_CODE (stmt) == MODIFY_EXPR)
1829 record_equivalences_from_stmt (stmt,
1830 may_optimize_p,
1831 ann);
1832
1833 /* If STMT is a COND_EXPR and it was modified, then we may know
1834 where it goes. If that is the case, then mark the CFG as altered.
1835
1836 This will cause us to later call remove_unreachable_blocks and
1837 cleanup_tree_cfg when it is safe to do so. It is not safe to
1838 clean things up here since removal of edges and such can trigger
1839 the removal of PHI nodes, which in turn can release SSA_NAMEs to
1840 the manager.
1841
1842 That's all fine and good, except that once SSA_NAMEs are released
1843 to the manager, we must not call create_ssa_name until all references
1844 to released SSA_NAMEs have been eliminated.
1845
1846 All references to the deleted SSA_NAMEs can not be eliminated until
1847 we remove unreachable blocks.
1848
1849 We can not remove unreachable blocks until after we have completed
1850 any queued jump threading.
1851
1852 We can not complete any queued jump threads until we have taken
1853 appropriate variables out of SSA form. Taking variables out of
1854 SSA form can call create_ssa_name and thus we lose.
1855
1856 Ultimately I suspect we're going to need to change the interface
1857 into the SSA_NAME manager. */
1858
1859 if (ann->modified)
1860 {
1861 tree val = NULL;
1862
1863 if (TREE_CODE (stmt) == COND_EXPR)
1864 val = COND_EXPR_COND (stmt);
1865 else if (TREE_CODE (stmt) == SWITCH_EXPR)
1866 val = SWITCH_COND (stmt);
1867
1868 if (val && TREE_CODE (val) == INTEGER_CST && find_taken_edge (bb, val))
1869 cfg_altered = true;
1870
1871 /* If we simplified a statement in such a way as to be shown that it
1872 cannot trap, update the eh information and the cfg to match. */
1873 if (maybe_clean_or_replace_eh_stmt (old_stmt, stmt))
1874 {
1875 bitmap_set_bit (need_eh_cleanup, bb->index);
1876 if (dump_file && (dump_flags & TDF_DETAILS))
1877 fprintf (dump_file, " Flagged to clear EH edges.\n");
1878 }
1879 }
1880
1881 if (may_have_exposed_new_symbols)
1882 VEC_safe_push (tree, heap, stmts_to_rescan, bsi_stmt (si));
1883 }
1884
1885 /* Search for an existing instance of STMT in the AVAIL_EXPRS table. If
1886 found, return its LHS. Otherwise insert STMT in the table and return
1887 NULL_TREE.
1888
1889 Also, when an expression is first inserted in the AVAIL_EXPRS table, it
1890 is also added to the stack pointed to by BLOCK_AVAIL_EXPRS_P, so that they
1891 can be removed when we finish processing this block and its children.
1892
1893 NOTE: This function assumes that STMT is a MODIFY_EXPR node that
1894 contains no CALL_EXPR on its RHS and makes no volatile nor
1895 aliased references. */
1896
1897 static tree
1898 lookup_avail_expr (tree stmt, bool insert)
1899 {
1900 void **slot;
1901 tree lhs;
1902 tree temp;
1903 struct expr_hash_elt *element = XNEW (struct expr_hash_elt);
1904
1905 lhs = TREE_CODE (stmt) == MODIFY_EXPR ? TREE_OPERAND (stmt, 0) : NULL;
1906
1907 initialize_hash_element (stmt, lhs, element);
1908
1909 /* Don't bother remembering constant assignments and copy operations.
1910 Constants and copy operations are handled by the constant/copy propagator
1911 in optimize_stmt. */
1912 if (TREE_CODE (element->rhs) == SSA_NAME
1913 || is_gimple_min_invariant (element->rhs))
1914 {
1915 free (element);
1916 return NULL_TREE;
1917 }
1918
1919 /* Finally try to find the expression in the main expression hash table. */
1920 slot = htab_find_slot_with_hash (avail_exprs, element, element->hash,
1921 (insert ? INSERT : NO_INSERT));
1922 if (slot == NULL)
1923 {
1924 free (element);
1925 return NULL_TREE;
1926 }
1927
1928 if (*slot == NULL)
1929 {
1930 *slot = (void *) element;
1931 VEC_safe_push (tree, heap, avail_exprs_stack,
1932 stmt ? stmt : element->rhs);
1933 return NULL_TREE;
1934 }
1935
1936 /* Extract the LHS of the assignment so that it can be used as the current
1937 definition of another variable. */
1938 lhs = ((struct expr_hash_elt *)*slot)->lhs;
1939
1940 /* See if the LHS appears in the CONST_AND_COPIES table. If it does, then
1941 use the value from the const_and_copies table. */
1942 if (TREE_CODE (lhs) == SSA_NAME)
1943 {
1944 temp = SSA_NAME_VALUE (lhs);
1945 if (temp && TREE_CODE (temp) != VALUE_HANDLE)
1946 lhs = temp;
1947 }
1948
1949 free (element);
1950 return lhs;
1951 }
1952
1953 /* Hashing and equality functions for AVAIL_EXPRS. The table stores
1954 MODIFY_EXPR statements. We compute a value number for expressions using
1955 the code of the expression and the SSA numbers of its operands. */
1956
1957 static hashval_t
1958 avail_expr_hash (const void *p)
1959 {
1960 tree stmt = ((struct expr_hash_elt *)p)->stmt;
1961 tree rhs = ((struct expr_hash_elt *)p)->rhs;
1962 tree vuse;
1963 ssa_op_iter iter;
1964 hashval_t val = 0;
1965
1966 /* iterative_hash_expr knows how to deal with any expression and
1967 deals with commutative operators as well, so just use it instead
1968 of duplicating such complexities here. */
1969 val = iterative_hash_expr (rhs, val);
1970
1971 /* If the hash table entry is not associated with a statement, then we
1972 can just hash the expression and not worry about virtual operands
1973 and such. */
1974 if (!stmt || !stmt_ann (stmt))
1975 return val;
1976
1977 /* Add the SSA version numbers of every vuse operand. This is important
1978 because compound variables like arrays are not renamed in the
1979 operands. Rather, the rename is done on the virtual variable
1980 representing all the elements of the array. */
1981 FOR_EACH_SSA_TREE_OPERAND (vuse, stmt, iter, SSA_OP_VUSE)
1982 val = iterative_hash_expr (vuse, val);
1983
1984 return val;
1985 }
1986
1987 static hashval_t
1988 real_avail_expr_hash (const void *p)
1989 {
1990 return ((const struct expr_hash_elt *)p)->hash;
1991 }
1992
1993 static int
1994 avail_expr_eq (const void *p1, const void *p2)
1995 {
1996 tree stmt1 = ((struct expr_hash_elt *)p1)->stmt;
1997 tree rhs1 = ((struct expr_hash_elt *)p1)->rhs;
1998 tree stmt2 = ((struct expr_hash_elt *)p2)->stmt;
1999 tree rhs2 = ((struct expr_hash_elt *)p2)->rhs;
2000
2001 /* If they are the same physical expression, return true. */
2002 if (rhs1 == rhs2 && stmt1 == stmt2)
2003 return true;
2004
2005 /* If their codes are not equal, then quit now. */
2006 if (TREE_CODE (rhs1) != TREE_CODE (rhs2))
2007 return false;
2008
2009 /* In case of a collision, both RHS have to be identical and have the
2010 same VUSE operands. */
2011 if ((TREE_TYPE (rhs1) == TREE_TYPE (rhs2)
2012 || lang_hooks.types_compatible_p (TREE_TYPE (rhs1), TREE_TYPE (rhs2)))
2013 && operand_equal_p (rhs1, rhs2, OEP_PURE_SAME))
2014 {
2015 bool ret = compare_ssa_operands_equal (stmt1, stmt2, SSA_OP_VUSE);
2016 gcc_assert (!ret || ((struct expr_hash_elt *)p1)->hash
2017 == ((struct expr_hash_elt *)p2)->hash);
2018 return ret;
2019 }
2020
2021 return false;
2022 }
2023
2024 /* PHI-ONLY copy and constant propagation. This pass is meant to clean
2025 up degenerate PHIs created by or exposed by jump threading. */
2026
2027 /* Given PHI, return its RHS if the PHI is a degenerate, otherwise return
2028 NULL. */
2029
2030 static tree
2031 degenerate_phi_result (tree phi)
2032 {
2033 tree lhs = PHI_RESULT (phi);
2034 tree val = NULL;
2035 int i;
2036
2037 /* Ignoring arguments which are the same as LHS, if all the remaining
2038 arguments are the same, then the PHI is a degenerate and has the
2039 value of that common argument. */
2040 for (i = 0; i < PHI_NUM_ARGS (phi); i++)
2041 {
2042 tree arg = PHI_ARG_DEF (phi, i);
2043
2044 if (arg == lhs)
2045 continue;
2046 else if (!val)
2047 val = arg;
2048 else if (!operand_equal_p (arg, val, 0))
2049 break;
2050 }
2051 return (i == PHI_NUM_ARGS (phi) ? val : NULL);
2052 }
2053
2054 /* Given a tree node T, which is either a PHI_NODE or MODIFY_EXPR,
2055 remove it from the IL. */
2056
2057 static void
2058 remove_stmt_or_phi (tree t)
2059 {
2060 if (TREE_CODE (t) == PHI_NODE)
2061 remove_phi_node (t, NULL);
2062 else
2063 {
2064 block_stmt_iterator bsi = bsi_for_stmt (t);
2065 bsi_remove (&bsi, true);
2066 }
2067 }
2068
2069 /* Given a tree node T, which is either a PHI_NODE or MODIFY_EXPR,
2070 return the "rhs" of the node, in the case of a non-degenerate
2071 PHI, NULL is returned. */
2072
2073 static tree
2074 get_rhs_or_phi_arg (tree t)
2075 {
2076 if (TREE_CODE (t) == PHI_NODE)
2077 return degenerate_phi_result (t);
2078 else if (TREE_CODE (t) == MODIFY_EXPR)
2079 return TREE_OPERAND (t, 1);
2080 gcc_unreachable ();
2081 }
2082
2083
2084 /* Given a tree node T, which is either a PHI_NODE or a MODIFY_EXPR,
2085 return the "lhs" of the node. */
2086
2087 static tree
2088 get_lhs_or_phi_result (tree t)
2089 {
2090 if (TREE_CODE (t) == PHI_NODE)
2091 return PHI_RESULT (t);
2092 else if (TREE_CODE (t) == MODIFY_EXPR)
2093 return TREE_OPERAND (t, 0);
2094 gcc_unreachable ();
2095 }
2096
2097 /* Propagate RHS into all uses of LHS (when possible).
2098
2099 RHS and LHS are derived from STMT, which is passed in solely so
2100 that we can remove it if propagation is successful.
2101
2102 When propagating into a PHI node or into a statement which turns
2103 into a trivial copy or constant initialization, set the
2104 appropriate bit in INTERESTING_NAMEs so that we will visit those
2105 nodes as well in an effort to pick up secondary optimization
2106 opportunities. */
2107
2108 static void
2109 propagate_rhs_into_lhs (tree stmt, tree lhs, tree rhs, bitmap interesting_names)
2110 {
2111 /* First verify that propagation is valid and isn't going to move a
2112 loop variant variable outside its loop. */
2113 if (! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs)
2114 && (TREE_CODE (rhs) != SSA_NAME
2115 || ! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (rhs))
2116 && may_propagate_copy (lhs, rhs)
2117 && loop_depth_of_name (lhs) >= loop_depth_of_name (rhs))
2118 {
2119 use_operand_p use_p;
2120 imm_use_iterator iter;
2121 bool all = true;
2122
2123 /* Dump details. */
2124 if (dump_file && (dump_flags & TDF_DETAILS))
2125 {
2126 fprintf (dump_file, " Replacing '");
2127 print_generic_expr (dump_file, lhs, dump_flags);
2128 fprintf (dump_file, "' with %s '",
2129 (TREE_CODE (rhs) != SSA_NAME ? "constant" : "variable"));
2130 print_generic_expr (dump_file, rhs, dump_flags);
2131 fprintf (dump_file, "'\n");
2132 }
2133
2134 /* Walk over every use of LHS and try to replace the use with RHS.
2135 At this point the only reason why such a propagation would not
2136 be successful would be if the use occurs in an ASM_EXPR. */
2137 repeat:
2138 FOR_EACH_IMM_USE_SAFE (use_p, iter, lhs)
2139 {
2140 tree use_stmt = USE_STMT (use_p);
2141
2142 /* It's not always safe to propagate into an ASM_EXPR. */
2143 if (TREE_CODE (use_stmt) == ASM_EXPR
2144 && ! may_propagate_copy_into_asm (lhs))
2145 {
2146 all = false;
2147 continue;
2148 }
2149
2150 /* Dump details. */
2151 if (dump_file && (dump_flags & TDF_DETAILS))
2152 {
2153 fprintf (dump_file, " Original statement:");
2154 print_generic_expr (dump_file, use_stmt, dump_flags);
2155 fprintf (dump_file, "\n");
2156 }
2157
2158 /* Propagate the RHS into this use of the LHS. */
2159 propagate_value (use_p, rhs);
2160
2161 /* Special cases to avoid useless calls into the folding
2162 routines, operand scanning, etc.
2163
2164 First, propagation into a PHI may cause the PHI to become
2165 a degenerate, so mark the PHI as interesting. No other
2166 actions are necessary.
2167
2168 Second, if we're propagating a virtual operand and the
2169 propagation does not change the underlying _DECL node for
2170 the virtual operand, then no further actions are necessary. */
2171 if (TREE_CODE (use_stmt) == PHI_NODE
2172 || (! is_gimple_reg (lhs)
2173 && TREE_CODE (rhs) == SSA_NAME
2174 && SSA_NAME_VAR (lhs) == SSA_NAME_VAR (rhs)))
2175 {
2176 /* Dump details. */
2177 if (dump_file && (dump_flags & TDF_DETAILS))
2178 {
2179 fprintf (dump_file, " Updated statement:");
2180 print_generic_expr (dump_file, use_stmt, dump_flags);
2181 fprintf (dump_file, "\n");
2182 }
2183
2184 /* Propagation into a PHI may expose new degenerate PHIs,
2185 so mark the result of the PHI as interesting. */
2186 if (TREE_CODE (use_stmt) == PHI_NODE)
2187 {
2188 tree result = get_lhs_or_phi_result (use_stmt);
2189 bitmap_set_bit (interesting_names, SSA_NAME_VERSION (result));
2190 }
2191 continue;
2192 }
2193
2194 /* From this point onward we are propagating into a
2195 real statement. Folding may (or may not) be possible,
2196 we may expose new operands, expose dead EH edges,
2197 etc. */
2198 fold_stmt_inplace (use_stmt);
2199
2200 /* Sometimes propagation can expose new operands to the
2201 renamer. Note this will call update_stmt at the
2202 appropriate time. */
2203 mark_new_vars_to_rename (use_stmt);
2204
2205 /* Dump details. */
2206 if (dump_file && (dump_flags & TDF_DETAILS))
2207 {
2208 fprintf (dump_file, " Updated statement:");
2209 print_generic_expr (dump_file, use_stmt, dump_flags);
2210 fprintf (dump_file, "\n");
2211 }
2212
2213 /* If we replaced a variable index with a constant, then
2214 we would need to update the invariant flag for ADDR_EXPRs. */
2215 if (TREE_CODE (use_stmt) == MODIFY_EXPR
2216 && TREE_CODE (TREE_OPERAND (use_stmt, 1)) == ADDR_EXPR)
2217 recompute_tree_invariant_for_addr_expr (TREE_OPERAND (use_stmt, 1));
2218
2219 /* If we cleaned up EH information from the statement,
2220 mark its containing block as needing EH cleanups. */
2221 if (maybe_clean_or_replace_eh_stmt (use_stmt, use_stmt))
2222 {
2223 bitmap_set_bit (need_eh_cleanup, bb_for_stmt (use_stmt)->index);
2224 if (dump_file && (dump_flags & TDF_DETAILS))
2225 fprintf (dump_file, " Flagged to clear EH edges.\n");
2226 }
2227
2228 /* Propagation may expose new trivial copy/constant propagation
2229 opportunities. */
2230 if (TREE_CODE (use_stmt) == MODIFY_EXPR
2231 && TREE_CODE (TREE_OPERAND (use_stmt, 0)) == SSA_NAME
2232 && (TREE_CODE (TREE_OPERAND (use_stmt, 1)) == SSA_NAME
2233 || is_gimple_min_invariant (TREE_OPERAND (use_stmt, 1))))
2234 {
2235 tree result = get_lhs_or_phi_result (use_stmt);
2236 bitmap_set_bit (interesting_names, SSA_NAME_VERSION (result));
2237 }
2238
2239 /* Propagation into these nodes may make certain edges in
2240 the CFG unexecutable. We want to identify them as PHI nodes
2241 at the destination of those unexecutable edges may become
2242 degenerates. */
2243 else if (TREE_CODE (use_stmt) == COND_EXPR
2244 || TREE_CODE (use_stmt) == SWITCH_EXPR
2245 || TREE_CODE (use_stmt) == GOTO_EXPR)
2246 {
2247 tree val;
2248
2249 if (TREE_CODE (use_stmt) == COND_EXPR)
2250 val = COND_EXPR_COND (use_stmt);
2251 else if (TREE_CODE (use_stmt) == SWITCH_EXPR)
2252 val = SWITCH_COND (use_stmt);
2253 else
2254 val = GOTO_DESTINATION (use_stmt);
2255
2256 if (is_gimple_min_invariant (val))
2257 {
2258 basic_block bb = bb_for_stmt (use_stmt);
2259 edge te = find_taken_edge (bb, val);
2260 edge_iterator ei;
2261 edge e;
2262 block_stmt_iterator bsi;
2263
2264 /* Remove all outgoing edges except TE. */
2265 for (ei = ei_start (bb->succs); (e = ei_safe_edge (ei));)
2266 {
2267 if (e != te)
2268 {
2269 tree phi;
2270
2271 /* Mark all the PHI nodes at the destination of
2272 the unexecutable edge as interesting. */
2273 for (phi = phi_nodes (e->dest);
2274 phi;
2275 phi = PHI_CHAIN (phi))
2276 {
2277 tree result = PHI_RESULT (phi);
2278 int version = SSA_NAME_VERSION (result);
2279
2280 bitmap_set_bit (interesting_names, version);
2281 }
2282
2283 te->probability += e->probability;
2284
2285 te->count += e->count;
2286 remove_edge (e);
2287 cfg_altered = 1;
2288 }
2289 else
2290 ei_next (&ei);
2291 }
2292
2293 bsi = bsi_last (bb_for_stmt (use_stmt));
2294 bsi_remove (&bsi, true);
2295
2296 /* And fixup the flags on the single remaining edge. */
2297 te->flags &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE);
2298 te->flags &= ~EDGE_ABNORMAL;
2299 te->flags |= EDGE_FALLTHRU;
2300 if (te->probability > REG_BR_PROB_BASE)
2301 te->probability = REG_BR_PROB_BASE;
2302 }
2303 }
2304 }
2305
2306 /* Due to a bug in the immediate use iterator code, we can
2307 miss visiting uses in some cases when there is more than
2308 one use in a statement. Missing a use can cause a multitude
2309 of problems if we expected to eliminate all uses and remove
2310 the defining statement.
2311
2312 Until Andrew can fix the iterator, this hack will detect
2313 the cases which cause us problems. Namely if ALL is set
2314 and we still have some immediate uses, then we must have
2315 skipped one or more in the loop above. So just re-execute
2316 the loop.
2317
2318 The maximum number of times we can re-execute the loop is
2319 bounded by the maximum number of times a given SSA_NAME
2320 appears in a single statement. */
2321 if (all && !has_zero_uses (lhs))
2322 goto repeat;
2323
2324 /* If we were able to propagate away all uses of LHS, then
2325 we can remove STMT. */
2326 if (all)
2327 remove_stmt_or_phi (stmt);
2328 }
2329 }
2330
2331 /* T is either a PHI node (potentially a degenerate PHI node) or
2332 a statement that is a trivial copy or constant initialization.
2333
2334 Attempt to eliminate T by propagating its RHS into all uses of
2335 its LHS. This may in turn set new bits in INTERESTING_NAMES
2336 for nodes we want to revisit later.
2337
2338 All exit paths should clear INTERESTING_NAMES for the result
2339 of T. */
2340
2341 static void
2342 eliminate_const_or_copy (tree t, bitmap interesting_names)
2343 {
2344 tree lhs = get_lhs_or_phi_result (t);
2345 tree rhs;
2346 int version = SSA_NAME_VERSION (lhs);
2347
2348 /* If the LHS of this statement or PHI has no uses, then we can
2349 just eliminate it. This can occur if, for example, the PHI
2350 was created by block duplication due to threading and its only
2351 use was in the conditional at the end of the block which was
2352 deleted. */
2353 if (has_zero_uses (lhs))
2354 {
2355 bitmap_clear_bit (interesting_names, version);
2356 remove_stmt_or_phi (t);
2357 return;
2358 }
2359
2360 /* Get the RHS of the assignment or PHI node if the PHI is a
2361 degenerate. */
2362 rhs = get_rhs_or_phi_arg (t);
2363 if (!rhs)
2364 {
2365 bitmap_clear_bit (interesting_names, version);
2366 return;
2367 }
2368
2369 propagate_rhs_into_lhs (t, lhs, rhs, interesting_names);
2370
2371 /* Note that T may well have been deleted by now, so do
2372 not access it, instead use the saved version # to clear
2373 T's entry in the worklist. */
2374 bitmap_clear_bit (interesting_names, version);
2375 }
2376
2377 /* The first phase in degenerate PHI elimination.
2378
2379 Eliminate the degenerate PHIs in BB, then recurse on the
2380 dominator children of BB. */
2381
2382 static void
2383 eliminate_degenerate_phis_1 (basic_block bb, bitmap interesting_names)
2384 {
2385 tree phi, next;
2386 basic_block son;
2387
2388 for (phi = phi_nodes (bb); phi; phi = next)
2389 {
2390 next = PHI_CHAIN (phi);
2391 eliminate_const_or_copy (phi, interesting_names);
2392 }
2393
2394 /* Recurse into the dominator children of BB. */
2395 for (son = first_dom_son (CDI_DOMINATORS, bb);
2396 son;
2397 son = next_dom_son (CDI_DOMINATORS, son))
2398 eliminate_degenerate_phis_1 (son, interesting_names);
2399 }
2400
2401
2402 /* A very simple pass to eliminate degenerate PHI nodes from the
2403 IL. This is meant to be fast enough to be able to be run several
2404 times in the optimization pipeline.
2405
2406 Certain optimizations, particularly those which duplicate blocks
2407 or remove edges from the CFG can create or expose PHIs which are
2408 trivial copies or constant initializations.
2409
2410 While we could pick up these optimizations in DOM or with the
2411 combination of copy-prop and CCP, those solutions are far too
2412 heavy-weight for our needs.
2413
2414 This implementation has two phases so that we can efficiently
2415 eliminate the first order degenerate PHIs and second order
2416 degenerate PHIs.
2417
2418 The first phase performs a dominator walk to identify and eliminate
2419 the vast majority of the degenerate PHIs. When a degenerate PHI
2420 is identified and eliminated any affected statements or PHIs
2421 are put on a worklist.
2422
2423 The second phase eliminates degenerate PHIs and trivial copies
2424 or constant initializations using the worklist. This is how we
2425 pick up the secondary optimization opportunities with minimal
2426 cost. */
2427
2428 static unsigned int
2429 eliminate_degenerate_phis (void)
2430 {
2431 bitmap interesting_names;
2432
2433 /* Bitmap of blocks which need EH information updated. We can not
2434 update it on-the-fly as doing so invalidates the dominator tree. */
2435 need_eh_cleanup = BITMAP_ALLOC (NULL);
2436
2437 /* INTERESTING_NAMES is effectively our worklist, indexed by
2438 SSA_NAME_VERSION.
2439
2440 A set bit indicates that the statement or PHI node which
2441 defines the SSA_NAME should be (re)examined to determine if
2442 it has become a degenerate PHI or trivial const/copy propagation
2443 opportunity.
2444
2445 Experiments have show we generally get better compilation
2446 time behavior with bitmaps rather than sbitmaps. */
2447 interesting_names = BITMAP_ALLOC (NULL);
2448
2449 /* First phase. Eliminate degenerate PHIs via a dominator
2450 walk of the CFG.
2451
2452 Experiments have indicated that we generally get better
2453 compile-time behavior by visiting blocks in the first
2454 phase in dominator order. Presumably this is because walking
2455 in dominator order leaves fewer PHIs for later examination
2456 by the worklist phase. */
2457 calculate_dominance_info (CDI_DOMINATORS);
2458 eliminate_degenerate_phis_1 (ENTRY_BLOCK_PTR, interesting_names);
2459
2460 /* Second phase. Eliminate second order degenerate PHIs as well
2461 as trivial copies or constant initializations identified by
2462 the first phase or this phase. Basically we keep iterating
2463 until our set of INTERESTING_NAMEs is empty. */
2464 while (!bitmap_empty_p (interesting_names))
2465 {
2466 unsigned int i;
2467 bitmap_iterator bi;
2468
2469 EXECUTE_IF_SET_IN_BITMAP (interesting_names, 0, i, bi)
2470 {
2471 tree name = ssa_name (i);
2472
2473 /* Ignore SSA_NAMEs that have been released because
2474 their defining statement was deleted (unreachable). */
2475 if (name)
2476 eliminate_const_or_copy (SSA_NAME_DEF_STMT (ssa_name (i)),
2477 interesting_names);
2478 }
2479 }
2480
2481 /* Propagation of const and copies may make some EH edges dead. Purge
2482 such edges from the CFG as needed. */
2483 if (!bitmap_empty_p (need_eh_cleanup))
2484 {
2485 cfg_altered |= tree_purge_all_dead_eh_edges (need_eh_cleanup);
2486 BITMAP_FREE (need_eh_cleanup);
2487 }
2488
2489 BITMAP_FREE (interesting_names);
2490 if (cfg_altered)
2491 free_dominance_info (CDI_DOMINATORS);
2492 return 0;
2493 }
2494
2495 struct tree_opt_pass pass_phi_only_cprop =
2496 {
2497 "phicprop", /* name */
2498 gate_dominator, /* gate */
2499 eliminate_degenerate_phis, /* execute */
2500 NULL, /* sub */
2501 NULL, /* next */
2502 0, /* static_pass_number */
2503 TV_TREE_PHI_CPROP, /* tv_id */
2504 PROP_cfg | PROP_ssa | PROP_alias, /* properties_required */
2505 0, /* properties_provided */
2506 PROP_smt_usage, /* properties_destroyed */
2507 0, /* todo_flags_start */
2508 TODO_cleanup_cfg | TODO_dump_func
2509 | TODO_ggc_collect | TODO_verify_ssa
2510 | TODO_verify_stmts | TODO_update_smt_usage
2511 | TODO_update_ssa, /* todo_flags_finish */
2512 0 /* letter */
2513 };