* tree-ssa-dom.c (record_edge_info): Fix memory leak.
[official-gcc.git] / gcc / tree-ssa-dom.c
blob3a3bdada4d884eb0a3fc6d19bb61c60b657fd39d
1 /* SSA Dominator optimizations for trees
2 Copyright (C) 2001, 2002, 2003, 2004 Free Software Foundation, Inc.
3 Contributed by Diego Novillo <dnovillo@redhat.com>
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 2, or (at your option)
10 any later version.
12 GCC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING. If not, write to
19 the Free Software Foundation, 59 Temple Place - Suite 330,
20 Boston, MA 02111-1307, USA. */
22 #include "config.h"
23 #include "system.h"
24 #include "coretypes.h"
25 #include "tm.h"
26 #include "tree.h"
27 #include "flags.h"
28 #include "rtl.h"
29 #include "tm_p.h"
30 #include "ggc.h"
31 #include "basic-block.h"
32 #include "output.h"
33 #include "errors.h"
34 #include "expr.h"
35 #include "function.h"
36 #include "diagnostic.h"
37 #include "timevar.h"
38 #include "tree-dump.h"
39 #include "tree-flow.h"
40 #include "domwalk.h"
41 #include "real.h"
42 #include "tree-pass.h"
43 #include "tree-ssa-propagate.h"
44 #include "langhooks.h"
46 /* This file implements optimizations on the dominator tree. */
49 /* Structure for recording edge equivalences as well as any pending
50 edge redirections during the dominator optimizer.
52 Computing and storing the edge equivalences instead of creating
53 them on-demand can save significant amounts of time, particularly
54 for pathological cases involving switch statements.
56 These structures live for a single iteration of the dominator
57 optimizer in the edge's AUX field. At the end of an iteration we
58 free each of these structures and update the AUX field to point
59 to any requested redirection target (the code for updating the
60 CFG and SSA graph for edge redirection expects redirection edge
61 targets to be in the AUX field for each edge. */
63 struct edge_info
65 /* If this edge creates a simple equivalence, the LHS and RHS of
66 the equivalence will be stored here. */
67 tree lhs;
68 tree rhs;
70 /* Traversing an edge may also indicate one or more particular conditions
71 are true or false. The number of recorded conditions can vary, but
72 can be determined by the condition's code. So we have an array
73 and its maximum index rather than use a varray. */
74 tree *cond_equivalences;
75 unsigned int max_cond_equivalences;
77 /* If we can thread this edge this field records the new target. */
78 edge redirection_target;
82 /* Hash table with expressions made available during the renaming process.
83 When an assignment of the form X_i = EXPR is found, the statement is
84 stored in this table. If the same expression EXPR is later found on the
85 RHS of another statement, it is replaced with X_i (thus performing
86 global redundancy elimination). Similarly as we pass through conditionals
87 we record the conditional itself as having either a true or false value
88 in this table. */
89 static htab_t avail_exprs;
91 /* Stack of available expressions in AVAIL_EXPRs. Each block pushes any
92 expressions it enters into the hash table along with a marker entry
93 (null). When we finish processing the block, we pop off entries and
94 remove the expressions from the global hash table until we hit the
95 marker. */
96 static VEC(tree_on_heap) *avail_exprs_stack;
98 /* Stack of trees used to restore the global currdefs to its original
99 state after completing optimization of a block and its dominator children.
101 An SSA_NAME indicates that the current definition of the underlying
102 variable should be set to the given SSA_NAME.
104 A _DECL node indicates that the underlying variable has no current
105 definition.
107 A NULL node is used to mark the last node associated with the
108 current block. */
109 static VEC(tree_on_heap) *block_defs_stack;
111 /* Stack of statements we need to rescan during finalization for newly
112 exposed variables.
114 Statement rescanning must occur after the current block's available
115 expressions are removed from AVAIL_EXPRS. Else we may change the
116 hash code for an expression and be unable to find/remove it from
117 AVAIL_EXPRS. */
118 static VEC(tree_on_heap) *stmts_to_rescan;
120 /* Structure for entries in the expression hash table.
122 This requires more memory for the hash table entries, but allows us
123 to avoid creating silly tree nodes and annotations for conditionals,
124 eliminates 2 global hash tables and two block local varrays.
126 It also allows us to reduce the number of hash table lookups we
127 have to perform in lookup_avail_expr and finally it allows us to
128 significantly reduce the number of calls into the hashing routine
129 itself. */
131 struct expr_hash_elt
133 /* The value (lhs) of this expression. */
134 tree lhs;
136 /* The expression (rhs) we want to record. */
137 tree rhs;
139 /* The annotation if this element corresponds to a statement. */
140 stmt_ann_t ann;
142 /* The hash value for RHS/ann. */
143 hashval_t hash;
146 /* Stack of dest,src pairs that need to be restored during finalization.
148 A NULL entry is used to mark the end of pairs which need to be
149 restored during finalization of this block. */
150 static VEC(tree_on_heap) *const_and_copies_stack;
152 /* Bitmap of SSA_NAMEs known to have a nonzero value, even if we do not
153 know their exact value. */
154 static bitmap nonzero_vars;
156 /* Stack of SSA_NAMEs which need their NONZERO_VARS property cleared
157 when the current block is finalized.
159 A NULL entry is used to mark the end of names needing their
160 entry in NONZERO_VARS cleared during finalization of this block. */
161 static VEC(tree_on_heap) *nonzero_vars_stack;
163 /* Track whether or not we have changed the control flow graph. */
164 static bool cfg_altered;
166 /* Bitmap of blocks that have had EH statements cleaned. We should
167 remove their dead edges eventually. */
168 static bitmap need_eh_cleanup;
170 /* Statistics for dominator optimizations. */
171 struct opt_stats_d
173 long num_stmts;
174 long num_exprs_considered;
175 long num_re;
178 static struct opt_stats_d opt_stats;
180 /* Value range propagation record. Each time we encounter a conditional
181 of the form SSA_NAME COND CONST we create a new vrp_element to record
182 how the condition affects the possible values SSA_NAME may have.
184 Each record contains the condition tested (COND), and the the range of
185 values the variable may legitimately have if COND is true. Note the
186 range of values may be a smaller range than COND specifies if we have
187 recorded other ranges for this variable. Each record also contains the
188 block in which the range was recorded for invalidation purposes.
190 Note that the current known range is computed lazily. This allows us
191 to avoid the overhead of computing ranges which are never queried.
193 When we encounter a conditional, we look for records which constrain
194 the SSA_NAME used in the condition. In some cases those records allow
195 us to determine the condition's result at compile time. In other cases
196 they may allow us to simplify the condition.
198 We also use value ranges to do things like transform signed div/mod
199 operations into unsigned div/mod or to simplify ABS_EXPRs.
201 Simple experiments have shown these optimizations to not be all that
202 useful on switch statements (much to my surprise). So switch statement
203 optimizations are not performed.
205 Note carefully we do not propagate information through each statement
206 in the block. i.e., if we know variable X has a value defined of
207 [0, 25] and we encounter Y = X + 1, we do not track a value range
208 for Y (which would be [1, 26] if we cared). Similarly we do not
209 constrain values as we encounter narrowing typecasts, etc. */
211 struct vrp_element
213 /* The highest and lowest values the variable in COND may contain when
214 COND is true. Note this may not necessarily be the same values
215 tested by COND if the same variable was used in earlier conditionals.
217 Note this is computed lazily and thus can be NULL indicating that
218 the values have not been computed yet. */
219 tree low;
220 tree high;
222 /* The actual conditional we recorded. This is needed since we compute
223 ranges lazily. */
224 tree cond;
226 /* The basic block where this record was created. We use this to determine
227 when to remove records. */
228 basic_block bb;
231 /* A hash table holding value range records (VRP_ELEMENTs) for a given
232 SSA_NAME. We used to use a varray indexed by SSA_NAME_VERSION, but
233 that gets awful wasteful, particularly since the density objects
234 with useful information is very low. */
235 static htab_t vrp_data;
237 /* An entry in the VRP_DATA hash table. We record the variable and a
238 varray of VRP_ELEMENT records associated with that variable. */
239 struct vrp_hash_elt
241 tree var;
242 varray_type records;
245 /* Array of variables which have their values constrained by operations
246 in this basic block. We use this during finalization to know
247 which variables need their VRP data updated. */
249 /* Stack of SSA_NAMEs which had their values constrainted by operations
250 in this basic block. During finalization of this block we use this
251 list to determine which variables need their VRP data updated.
253 A NULL entry marks the end of the SSA_NAMEs associated with this block. */
254 static VEC(tree_on_heap) *vrp_variables_stack;
256 struct eq_expr_value
258 tree src;
259 tree dst;
262 /* Local functions. */
263 static void optimize_stmt (struct dom_walk_data *,
264 basic_block bb,
265 block_stmt_iterator);
266 static tree lookup_avail_expr (tree, bool);
267 static hashval_t vrp_hash (const void *);
268 static int vrp_eq (const void *, const void *);
269 static hashval_t avail_expr_hash (const void *);
270 static hashval_t real_avail_expr_hash (const void *);
271 static int avail_expr_eq (const void *, const void *);
272 static void htab_statistics (FILE *, htab_t);
273 static void record_cond (tree, tree);
274 static void record_const_or_copy (tree, tree);
275 static void record_equality (tree, tree);
276 static tree update_rhs_and_lookup_avail_expr (tree, tree, bool);
277 static tree simplify_rhs_and_lookup_avail_expr (struct dom_walk_data *,
278 tree, int);
279 static tree simplify_cond_and_lookup_avail_expr (tree, stmt_ann_t, int);
280 static tree simplify_switch_and_lookup_avail_expr (tree, int);
281 static tree find_equivalent_equality_comparison (tree);
282 static void record_range (tree, basic_block);
283 static bool extract_range_from_cond (tree, tree *, tree *, int *);
284 static void record_equivalences_from_phis (basic_block);
285 static void record_equivalences_from_incoming_edge (basic_block);
286 static bool eliminate_redundant_computations (struct dom_walk_data *,
287 tree, stmt_ann_t);
288 static void record_equivalences_from_stmt (tree, int, stmt_ann_t);
289 static void thread_across_edge (struct dom_walk_data *, edge);
290 static void dom_opt_finalize_block (struct dom_walk_data *, basic_block);
291 static void dom_opt_initialize_block (struct dom_walk_data *, basic_block);
292 static void propagate_to_outgoing_edges (struct dom_walk_data *, basic_block);
293 static void remove_local_expressions_from_table (void);
294 static void restore_vars_to_original_value (void);
295 static void restore_currdefs_to_original_value (void);
296 static void register_definitions_for_stmt (tree);
297 static edge single_incoming_edge_ignoring_loop_edges (basic_block);
298 static void restore_nonzero_vars_to_original_value (void);
299 static inline bool unsafe_associative_fp_binop (tree);
301 /* Local version of fold that doesn't introduce cruft. */
303 static tree
304 local_fold (tree t)
306 t = fold (t);
308 /* Strip away useless type conversions. Both the NON_LVALUE_EXPR that
309 may have been added by fold, and "useless" type conversions that might
310 now be apparent due to propagation. */
311 STRIP_USELESS_TYPE_CONVERSION (t);
313 return t;
316 /* Allocate an EDGE_INFO for edge E and attach it to E.
317 Return the new EDGE_INFO structure. */
319 static struct edge_info *
320 allocate_edge_info (edge e)
322 struct edge_info *edge_info;
324 edge_info = xcalloc (1, sizeof (struct edge_info));
326 e->aux = edge_info;
327 return edge_info;
330 /* Free all EDGE_INFO structures associated with edges in the CFG.
331 If a particular edge can be threaded, copy the redirection
332 target from the EDGE_INFO structure into the edge's AUX field
333 as required by code to update the CFG and SSA graph for
334 jump threading. */
336 static void
337 free_all_edge_infos (void)
339 basic_block bb;
340 edge_iterator ei;
341 edge e;
343 FOR_EACH_BB (bb)
345 FOR_EACH_EDGE (e, ei, bb->preds)
347 struct edge_info *edge_info = e->aux;
349 if (edge_info)
351 e->aux = edge_info->redirection_target;
352 if (edge_info->cond_equivalences)
353 free (edge_info->cond_equivalences);
354 free (edge_info);
360 /* Jump threading, redundancy elimination and const/copy propagation.
362 This pass may expose new symbols that need to be renamed into SSA. For
363 every new symbol exposed, its corresponding bit will be set in
364 VARS_TO_RENAME. */
366 static void
367 tree_ssa_dominator_optimize (void)
369 struct dom_walk_data walk_data;
370 unsigned int i;
372 memset (&opt_stats, 0, sizeof (opt_stats));
374 for (i = 0; i < num_referenced_vars; i++)
375 var_ann (referenced_var (i))->current_def = NULL;
377 /* Mark loop edges so we avoid threading across loop boundaries.
378 This may result in transforming natural loop into irreducible
379 region. */
380 mark_dfs_back_edges ();
382 /* Create our hash tables. */
383 avail_exprs = htab_create (1024, real_avail_expr_hash, avail_expr_eq, free);
384 vrp_data = htab_create (ceil_log2 (num_ssa_names), vrp_hash, vrp_eq, free);
385 avail_exprs_stack = VEC_alloc (tree_on_heap, 20);
386 block_defs_stack = VEC_alloc (tree_on_heap, 20);
387 const_and_copies_stack = VEC_alloc (tree_on_heap, 20);
388 nonzero_vars_stack = VEC_alloc (tree_on_heap, 20);
389 vrp_variables_stack = VEC_alloc (tree_on_heap, 20);
390 stmts_to_rescan = VEC_alloc (tree_on_heap, 20);
391 nonzero_vars = BITMAP_XMALLOC ();
392 need_eh_cleanup = BITMAP_XMALLOC ();
394 /* Setup callbacks for the generic dominator tree walker. */
395 walk_data.walk_stmts_backward = false;
396 walk_data.dom_direction = CDI_DOMINATORS;
397 walk_data.initialize_block_local_data = NULL;
398 walk_data.before_dom_children_before_stmts = dom_opt_initialize_block;
399 walk_data.before_dom_children_walk_stmts = optimize_stmt;
400 walk_data.before_dom_children_after_stmts = propagate_to_outgoing_edges;
401 walk_data.after_dom_children_before_stmts = NULL;
402 walk_data.after_dom_children_walk_stmts = NULL;
403 walk_data.after_dom_children_after_stmts = dom_opt_finalize_block;
404 /* Right now we only attach a dummy COND_EXPR to the global data pointer.
405 When we attach more stuff we'll need to fill this out with a real
406 structure. */
407 walk_data.global_data = NULL;
408 walk_data.block_local_data_size = 0;
410 /* Now initialize the dominator walker. */
411 init_walk_dominator_tree (&walk_data);
413 calculate_dominance_info (CDI_DOMINATORS);
415 /* If we prove certain blocks are unreachable, then we want to
416 repeat the dominator optimization process as PHI nodes may
417 have turned into copies which allows better propagation of
418 values. So we repeat until we do not identify any new unreachable
419 blocks. */
422 /* Optimize the dominator tree. */
423 cfg_altered = false;
425 /* Recursively walk the dominator tree optimizing statements. */
426 walk_dominator_tree (&walk_data, ENTRY_BLOCK_PTR);
428 /* If we exposed any new variables, go ahead and put them into
429 SSA form now, before we handle jump threading. This simplifies
430 interactions between rewriting of _DECL nodes into SSA form
431 and rewriting SSA_NAME nodes into SSA form after block
432 duplication and CFG manipulation. */
433 if (!bitmap_empty_p (vars_to_rename))
435 rewrite_into_ssa (false);
436 bitmap_clear (vars_to_rename);
439 free_all_edge_infos ();
441 /* Thread jumps, creating duplicate blocks as needed. */
442 cfg_altered = thread_through_all_blocks ();
444 /* Removal of statements may make some EH edges dead. Purge
445 such edges from the CFG as needed. */
446 if (!bitmap_empty_p (need_eh_cleanup))
448 cfg_altered |= tree_purge_all_dead_eh_edges (need_eh_cleanup);
449 bitmap_zero (need_eh_cleanup);
452 free_dominance_info (CDI_DOMINATORS);
453 cfg_altered = cleanup_tree_cfg ();
454 calculate_dominance_info (CDI_DOMINATORS);
456 rewrite_ssa_into_ssa ();
458 /* Reinitialize the various tables. */
459 bitmap_clear (nonzero_vars);
460 htab_empty (avail_exprs);
461 htab_empty (vrp_data);
463 for (i = 0; i < num_referenced_vars; i++)
464 var_ann (referenced_var (i))->current_def = NULL;
466 while (cfg_altered);
468 /* Debugging dumps. */
469 if (dump_file && (dump_flags & TDF_STATS))
470 dump_dominator_optimization_stats (dump_file);
472 /* We emptied the hash table earlier, now delete it completely. */
473 htab_delete (avail_exprs);
474 htab_delete (vrp_data);
476 /* It is not necessary to clear CURRDEFS, REDIRECTION_EDGES, VRP_DATA,
477 CONST_AND_COPIES, and NONZERO_VARS as they all get cleared at the bottom
478 of the do-while loop above. */
480 /* And finalize the dominator walker. */
481 fini_walk_dominator_tree (&walk_data);
483 /* Free nonzero_vars. */
484 BITMAP_XFREE (nonzero_vars);
485 BITMAP_XFREE (need_eh_cleanup);
487 /* Finally, remove everything except invariants in SSA_NAME_VALUE.
489 Long term we will be able to let everything in SSA_NAME_VALUE
490 persist. However, for now, we know this is the safe thing to
491 do. */
492 for (i = 0; i < num_ssa_names; i++)
494 tree name = ssa_name (i);
495 tree value;
497 if (!name)
498 continue;
500 value = SSA_NAME_VALUE (name);
501 if (value && !is_gimple_min_invariant (value))
502 SSA_NAME_VALUE (name) = NULL;
505 VEC_free (tree_on_heap, block_defs_stack);
506 VEC_free (tree_on_heap, avail_exprs_stack);
507 VEC_free (tree_on_heap, const_and_copies_stack);
508 VEC_free (tree_on_heap, nonzero_vars_stack);
509 VEC_free (tree_on_heap, vrp_variables_stack);
510 VEC_free (tree_on_heap, stmts_to_rescan);
513 static bool
514 gate_dominator (void)
516 return flag_tree_dom != 0;
519 struct tree_opt_pass pass_dominator =
521 "dom", /* name */
522 gate_dominator, /* gate */
523 tree_ssa_dominator_optimize, /* execute */
524 NULL, /* sub */
525 NULL, /* next */
526 0, /* static_pass_number */
527 TV_TREE_SSA_DOMINATOR_OPTS, /* tv_id */
528 PROP_cfg | PROP_ssa | PROP_alias, /* properties_required */
529 0, /* properties_provided */
530 0, /* properties_destroyed */
531 0, /* todo_flags_start */
532 TODO_dump_func | TODO_rename_vars
533 | TODO_verify_ssa, /* todo_flags_finish */
534 0 /* letter */
538 /* We are exiting BB, see if the target block begins with a conditional
539 jump which has a known value when reached via BB. */
541 static void
542 thread_across_edge (struct dom_walk_data *walk_data, edge e)
544 block_stmt_iterator bsi;
545 tree stmt = NULL;
546 tree phi;
548 /* Each PHI creates a temporary equivalence, record them. */
549 for (phi = phi_nodes (e->dest); phi; phi = PHI_CHAIN (phi))
551 tree src = PHI_ARG_DEF_FROM_EDGE (phi, e);
552 tree dst = PHI_RESULT (phi);
553 record_const_or_copy (dst, src);
554 register_new_def (dst, &block_defs_stack);
557 for (bsi = bsi_start (e->dest); ! bsi_end_p (bsi); bsi_next (&bsi))
559 tree lhs, cached_lhs;
561 stmt = bsi_stmt (bsi);
563 /* Ignore empty statements and labels. */
564 if (IS_EMPTY_STMT (stmt) || TREE_CODE (stmt) == LABEL_EXPR)
565 continue;
567 /* If this is not a MODIFY_EXPR which sets an SSA_NAME to a new
568 value, then stop our search here. Ideally when we stop a
569 search we stop on a COND_EXPR or SWITCH_EXPR. */
570 if (TREE_CODE (stmt) != MODIFY_EXPR
571 || TREE_CODE (TREE_OPERAND (stmt, 0)) != SSA_NAME)
572 break;
574 /* At this point we have a statement which assigns an RHS to an
575 SSA_VAR on the LHS. We want to prove that the RHS is already
576 available and that its value is held in the current definition
577 of the LHS -- meaning that this assignment is a NOP when
578 reached via edge E. */
579 if (TREE_CODE (TREE_OPERAND (stmt, 1)) == SSA_NAME)
580 cached_lhs = TREE_OPERAND (stmt, 1);
581 else
582 cached_lhs = lookup_avail_expr (stmt, false);
584 lhs = TREE_OPERAND (stmt, 0);
586 /* This can happen if we thread around to the start of a loop. */
587 if (lhs == cached_lhs)
588 break;
590 /* If we did not find RHS in the hash table, then try again after
591 temporarily const/copy propagating the operands. */
592 if (!cached_lhs)
594 /* Copy the operands. */
595 stmt_ann_t ann = stmt_ann (stmt);
596 use_optype uses = USE_OPS (ann);
597 vuse_optype vuses = VUSE_OPS (ann);
598 tree *uses_copy = xcalloc (NUM_USES (uses), sizeof (tree));
599 tree *vuses_copy = xcalloc (NUM_VUSES (vuses), sizeof (tree));
600 unsigned int i;
602 /* Make a copy of the uses into USES_COPY, then cprop into
603 the use operands. */
604 for (i = 0; i < NUM_USES (uses); i++)
606 tree tmp = NULL;
608 uses_copy[i] = USE_OP (uses, i);
609 if (TREE_CODE (USE_OP (uses, i)) == SSA_NAME)
610 tmp = SSA_NAME_VALUE (USE_OP (uses, i));
611 if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
612 SET_USE_OP (uses, i, tmp);
615 /* Similarly for virtual uses. */
616 for (i = 0; i < NUM_VUSES (vuses); i++)
618 tree tmp = NULL;
620 vuses_copy[i] = VUSE_OP (vuses, i);
621 if (TREE_CODE (VUSE_OP (vuses, i)) == SSA_NAME)
622 tmp = SSA_NAME_VALUE (VUSE_OP (vuses, i));
623 if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
624 SET_VUSE_OP (vuses, i, tmp);
627 /* Try to lookup the new expression. */
628 cached_lhs = lookup_avail_expr (stmt, false);
630 /* Restore the statement's original uses/defs. */
631 for (i = 0; i < NUM_USES (uses); i++)
632 SET_USE_OP (uses, i, uses_copy[i]);
634 for (i = 0; i < NUM_VUSES (vuses); i++)
635 SET_VUSE_OP (vuses, i, vuses_copy[i]);
637 free (uses_copy);
638 free (vuses_copy);
640 /* If we still did not find the expression in the hash table,
641 then we can not ignore this statement. */
642 if (! cached_lhs)
643 break;
646 /* If the expression in the hash table was not assigned to an
647 SSA_NAME, then we can not ignore this statement. */
648 if (TREE_CODE (cached_lhs) != SSA_NAME)
649 break;
651 /* If we have different underlying variables, then we can not
652 ignore this statement. */
653 if (SSA_NAME_VAR (cached_lhs) != SSA_NAME_VAR (lhs))
654 break;
656 /* If CACHED_LHS does not represent the current value of the underlying
657 variable in CACHED_LHS/LHS, then we can not ignore this statement. */
658 if (var_ann (SSA_NAME_VAR (lhs))->current_def != cached_lhs)
659 break;
661 /* If we got here, then we can ignore this statement and continue
662 walking through the statements in the block looking for a threadable
663 COND_EXPR.
665 We want to record an equivalence lhs = cache_lhs so that if
666 the result of this statement is used later we can copy propagate
667 suitably. */
668 record_const_or_copy (lhs, cached_lhs);
669 register_new_def (lhs, &block_defs_stack);
672 /* If we stopped at a COND_EXPR or SWITCH_EXPR, then see if we know which
673 arm will be taken. */
674 if (stmt
675 && (TREE_CODE (stmt) == COND_EXPR
676 || TREE_CODE (stmt) == SWITCH_EXPR))
678 tree cond, cached_lhs;
679 edge e1;
680 edge_iterator ei;
682 /* Do not forward entry edges into the loop. In the case loop
683 has multiple entry edges we may end up in constructing irreducible
684 region.
685 ??? We may consider forwarding the edges in the case all incoming
686 edges forward to the same destination block. */
687 if (!e->flags & EDGE_DFS_BACK)
689 FOR_EACH_EDGE (e1, ei, e->dest->preds)
690 if (e1->flags & EDGE_DFS_BACK)
691 break;
692 if (e1)
693 return;
696 /* Now temporarily cprop the operands and try to find the resulting
697 expression in the hash tables. */
698 if (TREE_CODE (stmt) == COND_EXPR)
699 cond = COND_EXPR_COND (stmt);
700 else
701 cond = SWITCH_COND (stmt);
703 if (COMPARISON_CLASS_P (cond))
705 tree dummy_cond, op0, op1;
706 enum tree_code cond_code;
708 op0 = TREE_OPERAND (cond, 0);
709 op1 = TREE_OPERAND (cond, 1);
710 cond_code = TREE_CODE (cond);
712 /* Get the current value of both operands. */
713 if (TREE_CODE (op0) == SSA_NAME)
715 tree tmp = SSA_NAME_VALUE (op0);
716 if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
717 op0 = tmp;
720 if (TREE_CODE (op1) == SSA_NAME)
722 tree tmp = SSA_NAME_VALUE (op1);
723 if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
724 op1 = tmp;
727 /* Stuff the operator and operands into our dummy conditional
728 expression, creating the dummy conditional if necessary. */
729 dummy_cond = walk_data->global_data;
730 if (! dummy_cond)
732 dummy_cond = build (cond_code, boolean_type_node, op0, op1);
733 dummy_cond = build (COND_EXPR, void_type_node,
734 dummy_cond, NULL, NULL);
735 walk_data->global_data = dummy_cond;
737 else
739 TREE_SET_CODE (COND_EXPR_COND (dummy_cond), cond_code);
740 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 0) = op0;
741 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 1) = op1;
744 /* If the conditional folds to an invariant, then we are done,
745 otherwise look it up in the hash tables. */
746 cached_lhs = local_fold (COND_EXPR_COND (dummy_cond));
747 if (! is_gimple_min_invariant (cached_lhs))
749 cached_lhs = lookup_avail_expr (dummy_cond, false);
750 if (!cached_lhs || ! is_gimple_min_invariant (cached_lhs))
751 cached_lhs = simplify_cond_and_lookup_avail_expr (dummy_cond,
752 NULL,
753 false);
756 /* We can have conditionals which just test the state of a
757 variable rather than use a relational operator. These are
758 simpler to handle. */
759 else if (TREE_CODE (cond) == SSA_NAME)
761 cached_lhs = cond;
762 cached_lhs = SSA_NAME_VALUE (cached_lhs);
763 if (cached_lhs && ! is_gimple_min_invariant (cached_lhs))
764 cached_lhs = 0;
766 else
767 cached_lhs = lookup_avail_expr (stmt, false);
769 if (cached_lhs)
771 edge taken_edge = find_taken_edge (e->dest, cached_lhs);
772 basic_block dest = (taken_edge ? taken_edge->dest : NULL);
774 if (dest == e->dest)
775 return;
777 /* If we have a known destination for the conditional, then
778 we can perform this optimization, which saves at least one
779 conditional jump each time it applies since we get to
780 bypass the conditional at our original destination. */
781 if (dest)
783 struct edge_info *edge_info;
785 update_bb_profile_for_threading (e->dest, EDGE_FREQUENCY (e),
786 e->count, taken_edge);
787 if (e->aux)
788 edge_info = e->aux;
789 else
790 edge_info = allocate_edge_info (e);
791 edge_info->redirection_target = taken_edge;
792 bb_ann (e->dest)->incoming_edge_threaded = true;
799 /* Initialize local stacks for this optimizer and record equivalences
800 upon entry to BB. Equivalences can come from the edge traversed to
801 reach BB or they may come from PHI nodes at the start of BB. */
803 static void
804 dom_opt_initialize_block (struct dom_walk_data *walk_data ATTRIBUTE_UNUSED,
805 basic_block bb)
807 if (dump_file && (dump_flags & TDF_DETAILS))
808 fprintf (dump_file, "\n\nOptimizing block #%d\n\n", bb->index);
810 /* Push a marker on the stacks of local information so that we know how
811 far to unwind when we finalize this block. */
812 VEC_safe_push (tree_on_heap, avail_exprs_stack, NULL_TREE);
813 VEC_safe_push (tree_on_heap, block_defs_stack, NULL_TREE);
814 VEC_safe_push (tree_on_heap, const_and_copies_stack, NULL_TREE);
815 VEC_safe_push (tree_on_heap, nonzero_vars_stack, NULL_TREE);
816 VEC_safe_push (tree_on_heap, vrp_variables_stack, NULL_TREE);
818 record_equivalences_from_incoming_edge (bb);
820 /* PHI nodes can create equivalences too. */
821 record_equivalences_from_phis (bb);
824 /* Given an expression EXPR (a relational expression or a statement),
825 initialize the hash table element pointed by by ELEMENT. */
827 static void
828 initialize_hash_element (tree expr, tree lhs, struct expr_hash_elt *element)
830 /* Hash table elements may be based on conditional expressions or statements.
832 For the former case, we have no annotation and we want to hash the
833 conditional expression. In the latter case we have an annotation and
834 we want to record the expression the statement evaluates. */
835 if (COMPARISON_CLASS_P (expr) || TREE_CODE (expr) == TRUTH_NOT_EXPR)
837 element->ann = NULL;
838 element->rhs = expr;
840 else if (TREE_CODE (expr) == COND_EXPR)
842 element->ann = stmt_ann (expr);
843 element->rhs = COND_EXPR_COND (expr);
845 else if (TREE_CODE (expr) == SWITCH_EXPR)
847 element->ann = stmt_ann (expr);
848 element->rhs = SWITCH_COND (expr);
850 else if (TREE_CODE (expr) == RETURN_EXPR && TREE_OPERAND (expr, 0))
852 element->ann = stmt_ann (expr);
853 element->rhs = TREE_OPERAND (TREE_OPERAND (expr, 0), 1);
855 else
857 element->ann = stmt_ann (expr);
858 element->rhs = TREE_OPERAND (expr, 1);
861 element->lhs = lhs;
862 element->hash = avail_expr_hash (element);
865 /* Remove all the expressions in LOCALS from TABLE, stopping when there are
866 LIMIT entries left in LOCALs. */
868 static void
869 remove_local_expressions_from_table (void)
871 /* Remove all the expressions made available in this block. */
872 while (VEC_length (tree_on_heap, avail_exprs_stack) > 0)
874 struct expr_hash_elt element;
875 tree expr = VEC_pop (tree_on_heap, avail_exprs_stack);
877 if (expr == NULL_TREE)
878 break;
880 initialize_hash_element (expr, NULL, &element);
881 htab_remove_elt_with_hash (avail_exprs, &element, element.hash);
885 /* Use the SSA_NAMES in LOCALS to restore TABLE to its original
886 state, stopping when there are LIMIT entries left in LOCALs. */
888 static void
889 restore_nonzero_vars_to_original_value (void)
891 while (VEC_length (tree_on_heap, nonzero_vars_stack) > 0)
893 tree name = VEC_pop (tree_on_heap, nonzero_vars_stack);
895 if (name == NULL)
896 break;
898 bitmap_clear_bit (nonzero_vars, SSA_NAME_VERSION (name));
902 /* Use the source/dest pairs in CONST_AND_COPIES_STACK to restore
903 CONST_AND_COPIES to its original state, stopping when we hit a
904 NULL marker. */
906 static void
907 restore_vars_to_original_value (void)
909 while (VEC_length (tree_on_heap, const_and_copies_stack) > 0)
911 tree prev_value, dest;
913 dest = VEC_pop (tree_on_heap, const_and_copies_stack);
915 if (dest == NULL)
916 break;
918 prev_value = VEC_pop (tree_on_heap, const_and_copies_stack);
919 SSA_NAME_VALUE (dest) = prev_value;
923 /* Similar to restore_vars_to_original_value, except that it restores
924 CURRDEFS to its original value. */
925 static void
926 restore_currdefs_to_original_value (void)
928 /* Restore CURRDEFS to its original state. */
929 while (VEC_length (tree_on_heap, block_defs_stack) > 0)
931 tree tmp = VEC_pop (tree_on_heap, block_defs_stack);
932 tree saved_def, var;
934 if (tmp == NULL_TREE)
935 break;
937 /* If we recorded an SSA_NAME, then make the SSA_NAME the current
938 definition of its underlying variable. If we recorded anything
939 else, it must have been an _DECL node and its current reaching
940 definition must have been NULL. */
941 if (TREE_CODE (tmp) == SSA_NAME)
943 saved_def = tmp;
944 var = SSA_NAME_VAR (saved_def);
946 else
948 saved_def = NULL;
949 var = tmp;
952 var_ann (var)->current_def = saved_def;
956 /* We have finished processing the dominator children of BB, perform
957 any finalization actions in preparation for leaving this node in
958 the dominator tree. */
960 static void
961 dom_opt_finalize_block (struct dom_walk_data *walk_data, basic_block bb)
963 tree last;
965 /* If we are at a leaf node in the dominator tree, see if we can thread
966 the edge from BB through its successor.
968 Do this before we remove entries from our equivalence tables. */
969 if (EDGE_COUNT (bb->succs) == 1
970 && (EDGE_SUCC (bb, 0)->flags & EDGE_ABNORMAL) == 0
971 && (get_immediate_dominator (CDI_DOMINATORS, EDGE_SUCC (bb, 0)->dest) != bb
972 || phi_nodes (EDGE_SUCC (bb, 0)->dest)))
975 thread_across_edge (walk_data, EDGE_SUCC (bb, 0));
977 else if ((last = last_stmt (bb))
978 && TREE_CODE (last) == COND_EXPR
979 && (COMPARISON_CLASS_P (COND_EXPR_COND (last))
980 || TREE_CODE (COND_EXPR_COND (last)) == SSA_NAME)
981 && EDGE_COUNT (bb->succs) == 2
982 && (EDGE_SUCC (bb, 0)->flags & EDGE_ABNORMAL) == 0
983 && (EDGE_SUCC (bb, 1)->flags & EDGE_ABNORMAL) == 0)
985 edge true_edge, false_edge;
987 extract_true_false_edges_from_block (bb, &true_edge, &false_edge);
989 /* If the THEN arm is the end of a dominator tree or has PHI nodes,
990 then try to thread through its edge. */
991 if (get_immediate_dominator (CDI_DOMINATORS, true_edge->dest) != bb
992 || phi_nodes (true_edge->dest))
994 struct edge_info *edge_info;
995 unsigned int i;
997 /* Push a marker onto the available expression stack so that we
998 unwind any expressions related to the TRUE arm before processing
999 the false arm below. */
1000 VEC_safe_push (tree_on_heap, avail_exprs_stack, NULL_TREE);
1001 VEC_safe_push (tree_on_heap, block_defs_stack, NULL_TREE);
1002 VEC_safe_push (tree_on_heap, const_and_copies_stack, NULL_TREE);
1004 edge_info = true_edge->aux;
1006 /* If we have info associated with this edge, record it into
1007 our equivalency tables. */
1008 if (edge_info)
1010 tree *cond_equivalences = edge_info->cond_equivalences;
1011 tree lhs = edge_info->lhs;
1012 tree rhs = edge_info->rhs;
1014 /* If we have a simple NAME = VALUE equivalency record it.
1015 Until the jump threading selection code improves, only
1016 do this if both the name and value are SSA_NAMEs with
1017 the same underlying variable to avoid missing threading
1018 opportunities. */
1019 if (lhs
1020 && TREE_CODE (COND_EXPR_COND (last)) == SSA_NAME
1021 && TREE_CODE (edge_info->rhs) == SSA_NAME
1022 && SSA_NAME_VAR (lhs) == SSA_NAME_VAR (rhs))
1023 record_const_or_copy (lhs, rhs);
1025 /* If we have 0 = COND or 1 = COND equivalences, record them
1026 into our expression hash tables. */
1027 if (cond_equivalences)
1028 for (i = 0; i < edge_info->max_cond_equivalences; i += 2)
1030 tree expr = cond_equivalences[i];
1031 tree value = cond_equivalences[i + 1];
1033 record_cond (expr, value);
1037 /* Now thread the edge. */
1038 thread_across_edge (walk_data, true_edge);
1040 /* And restore the various tables to their state before
1041 we threaded this edge. */
1042 remove_local_expressions_from_table ();
1043 restore_vars_to_original_value ();
1044 restore_currdefs_to_original_value ();
1047 /* Similarly for the ELSE arm. */
1048 if (get_immediate_dominator (CDI_DOMINATORS, false_edge->dest) != bb
1049 || phi_nodes (false_edge->dest))
1051 struct edge_info *edge_info;
1052 unsigned int i;
1054 edge_info = false_edge->aux;
1056 /* If we have info associated with this edge, record it into
1057 our equivalency tables. */
1058 if (edge_info)
1060 tree *cond_equivalences = edge_info->cond_equivalences;
1061 tree lhs = edge_info->lhs;
1062 tree rhs = edge_info->rhs;
1064 /* If we have a simple NAME = VALUE equivalency record it.
1065 Until the jump threading selection code improves, only
1066 do this if both the name and value are SSA_NAMEs with
1067 the same underlying variable to avoid missing threading
1068 opportunities. */
1069 if (lhs
1070 && TREE_CODE (COND_EXPR_COND (last)) == SSA_NAME)
1071 record_const_or_copy (lhs, rhs);
1073 /* If we have 0 = COND or 1 = COND equivalences, record them
1074 into our expression hash tables. */
1075 if (cond_equivalences)
1076 for (i = 0; i < edge_info->max_cond_equivalences; i += 2)
1078 tree expr = cond_equivalences[i];
1079 tree value = cond_equivalences[i + 1];
1081 record_cond (expr, value);
1085 thread_across_edge (walk_data, false_edge);
1087 /* No need to remove local expressions from our tables
1088 or restore vars to their original value as that will
1089 be done immediately below. */
1093 remove_local_expressions_from_table ();
1094 restore_nonzero_vars_to_original_value ();
1095 restore_vars_to_original_value ();
1096 restore_currdefs_to_original_value ();
1098 /* Remove VRP records associated with this basic block. They are no
1099 longer valid.
1101 To be efficient, we note which variables have had their values
1102 constrained in this block. So walk over each variable in the
1103 VRP_VARIABLEs array. */
1104 while (VEC_length (tree_on_heap, vrp_variables_stack) > 0)
1106 tree var = VEC_pop (tree_on_heap, vrp_variables_stack);
1107 struct vrp_hash_elt vrp_hash_elt, *vrp_hash_elt_p;
1108 void **slot;
1110 /* Each variable has a stack of value range records. We want to
1111 invalidate those associated with our basic block. So we walk
1112 the array backwards popping off records associated with our
1113 block. Once we hit a record not associated with our block
1114 we are done. */
1115 varray_type var_vrp_records;
1117 if (var == NULL)
1118 break;
1120 vrp_hash_elt.var = var;
1121 vrp_hash_elt.records = NULL;
1123 slot = htab_find_slot (vrp_data, &vrp_hash_elt, NO_INSERT);
1125 vrp_hash_elt_p = (struct vrp_hash_elt *) *slot;
1126 var_vrp_records = vrp_hash_elt_p->records;
1128 while (VARRAY_ACTIVE_SIZE (var_vrp_records) > 0)
1130 struct vrp_element *element
1131 = (struct vrp_element *)VARRAY_TOP_GENERIC_PTR (var_vrp_records);
1133 if (element->bb != bb)
1134 break;
1136 VARRAY_POP (var_vrp_records);
1140 /* If we queued any statements to rescan in this block, then
1141 go ahead and rescan them now. */
1142 while (VEC_length (tree_on_heap, stmts_to_rescan) > 0)
1144 tree stmt = VEC_last (tree_on_heap, stmts_to_rescan);
1145 basic_block stmt_bb = bb_for_stmt (stmt);
1147 if (stmt_bb != bb)
1148 break;
1150 VEC_pop (tree_on_heap, stmts_to_rescan);
1151 mark_new_vars_to_rename (stmt, vars_to_rename);
1155 /* PHI nodes can create equivalences too.
1157 Ignoring any alternatives which are the same as the result, if
1158 all the alternatives are equal, then the PHI node creates an
1159 equivalence.
1161 Additionally, if all the PHI alternatives are known to have a nonzero
1162 value, then the result of this PHI is known to have a nonzero value,
1163 even if we do not know its exact value. */
1165 static void
1166 record_equivalences_from_phis (basic_block bb)
1168 tree phi;
1170 for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi))
1172 tree lhs = PHI_RESULT (phi);
1173 tree rhs = NULL;
1174 int i;
1176 for (i = 0; i < PHI_NUM_ARGS (phi); i++)
1178 tree t = PHI_ARG_DEF (phi, i);
1180 /* Ignore alternatives which are the same as our LHS. Since
1181 LHS is a PHI_RESULT, it is known to be a SSA_NAME, so we
1182 can simply compare pointers. */
1183 if (lhs == t)
1184 continue;
1186 /* If we have not processed an alternative yet, then set
1187 RHS to this alternative. */
1188 if (rhs == NULL)
1189 rhs = t;
1190 /* If we have processed an alternative (stored in RHS), then
1191 see if it is equal to this one. If it isn't, then stop
1192 the search. */
1193 else if (! operand_equal_for_phi_arg_p (rhs, t))
1194 break;
1197 /* If we had no interesting alternatives, then all the RHS alternatives
1198 must have been the same as LHS. */
1199 if (!rhs)
1200 rhs = lhs;
1202 /* If we managed to iterate through each PHI alternative without
1203 breaking out of the loop, then we have a PHI which may create
1204 a useful equivalence. We do not need to record unwind data for
1205 this, since this is a true assignment and not an equivalence
1206 inferred from a comparison. All uses of this ssa name are dominated
1207 by this assignment, so unwinding just costs time and space. */
1208 if (i == PHI_NUM_ARGS (phi)
1209 && may_propagate_copy (lhs, rhs))
1210 SSA_NAME_VALUE (lhs) = rhs;
1212 /* Now see if we know anything about the nonzero property for the
1213 result of this PHI. */
1214 for (i = 0; i < PHI_NUM_ARGS (phi); i++)
1216 if (!PHI_ARG_NONZERO (phi, i))
1217 break;
1220 if (i == PHI_NUM_ARGS (phi))
1221 bitmap_set_bit (nonzero_vars, SSA_NAME_VERSION (PHI_RESULT (phi)));
1223 register_new_def (lhs, &block_defs_stack);
1227 /* Ignoring loop backedges, if BB has precisely one incoming edge then
1228 return that edge. Otherwise return NULL. */
1229 static edge
1230 single_incoming_edge_ignoring_loop_edges (basic_block bb)
1232 edge retval = NULL;
1233 edge e;
1234 edge_iterator ei;
1236 FOR_EACH_EDGE (e, ei, bb->preds)
1238 /* A loop back edge can be identified by the destination of
1239 the edge dominating the source of the edge. */
1240 if (dominated_by_p (CDI_DOMINATORS, e->src, e->dest))
1241 continue;
1243 /* If we have already seen a non-loop edge, then we must have
1244 multiple incoming non-loop edges and thus we return NULL. */
1245 if (retval)
1246 return NULL;
1248 /* This is the first non-loop incoming edge we have found. Record
1249 it. */
1250 retval = e;
1253 return retval;
1256 /* Record any equivalences created by the incoming edge to BB. If BB
1257 has more than one incoming edge, then no equivalence is created. */
1259 static void
1260 record_equivalences_from_incoming_edge (basic_block bb)
1262 edge e;
1263 basic_block parent;
1264 struct edge_info *edge_info;
1266 /* If our parent block ended with a control statment, then we may be
1267 able to record some equivalences based on which outgoing edge from
1268 the parent was followed. */
1269 parent = get_immediate_dominator (CDI_DOMINATORS, bb);
1271 e = single_incoming_edge_ignoring_loop_edges (bb);
1273 /* If we had a single incoming edge from our parent block, then enter
1274 any data associated with the edge into our tables. */
1275 if (e && e->src == parent)
1277 unsigned int i;
1279 edge_info = e->aux;
1281 if (edge_info)
1283 tree lhs = edge_info->lhs;
1284 tree rhs = edge_info->rhs;
1285 tree *cond_equivalences = edge_info->cond_equivalences;
1287 if (lhs)
1288 record_equality (lhs, rhs);
1290 if (cond_equivalences)
1292 bool recorded_range = false;
1293 for (i = 0; i < edge_info->max_cond_equivalences; i += 2)
1295 tree expr = cond_equivalences[i];
1296 tree value = cond_equivalences[i + 1];
1298 record_cond (expr, value);
1300 /* For the first true equivalence, record range
1301 information. We only do this for the first
1302 true equivalence as it should dominate any
1303 later true equivalences. */
1304 if (! recorded_range
1305 && COMPARISON_CLASS_P (expr)
1306 && value == boolean_true_node
1307 && TREE_CONSTANT (TREE_OPERAND (expr, 1)))
1309 record_range (expr, bb);
1310 recorded_range = true;
1318 /* Dump SSA statistics on FILE. */
1320 void
1321 dump_dominator_optimization_stats (FILE *file)
1323 long n_exprs;
1325 fprintf (file, "Total number of statements: %6ld\n\n",
1326 opt_stats.num_stmts);
1327 fprintf (file, "Exprs considered for dominator optimizations: %6ld\n",
1328 opt_stats.num_exprs_considered);
1330 n_exprs = opt_stats.num_exprs_considered;
1331 if (n_exprs == 0)
1332 n_exprs = 1;
1334 fprintf (file, " Redundant expressions eliminated: %6ld (%.0f%%)\n",
1335 opt_stats.num_re, PERCENT (opt_stats.num_re,
1336 n_exprs));
1338 fprintf (file, "\nHash table statistics:\n");
1340 fprintf (file, " avail_exprs: ");
1341 htab_statistics (file, avail_exprs);
1345 /* Dump SSA statistics on stderr. */
1347 void
1348 debug_dominator_optimization_stats (void)
1350 dump_dominator_optimization_stats (stderr);
1354 /* Dump statistics for the hash table HTAB. */
1356 static void
1357 htab_statistics (FILE *file, htab_t htab)
1359 fprintf (file, "size %ld, %ld elements, %f collision/search ratio\n",
1360 (long) htab_size (htab),
1361 (long) htab_elements (htab),
1362 htab_collisions (htab));
1365 /* Record the fact that VAR has a nonzero value, though we may not know
1366 its exact value. Note that if VAR is already known to have a nonzero
1367 value, then we do nothing. */
1369 static void
1370 record_var_is_nonzero (tree var)
1372 int indx = SSA_NAME_VERSION (var);
1374 if (bitmap_bit_p (nonzero_vars, indx))
1375 return;
1377 /* Mark it in the global table. */
1378 bitmap_set_bit (nonzero_vars, indx);
1380 /* Record this SSA_NAME so that we can reset the global table
1381 when we leave this block. */
1382 VEC_safe_push (tree_on_heap, nonzero_vars_stack, var);
1385 /* Enter a statement into the true/false expression hash table indicating
1386 that the condition COND has the value VALUE. */
1388 static void
1389 record_cond (tree cond, tree value)
1391 struct expr_hash_elt *element = xmalloc (sizeof (struct expr_hash_elt));
1392 void **slot;
1394 initialize_hash_element (cond, value, element);
1396 slot = htab_find_slot_with_hash (avail_exprs, (void *)element,
1397 element->hash, true);
1398 if (*slot == NULL)
1400 *slot = (void *) element;
1401 VEC_safe_push (tree_on_heap, avail_exprs_stack, cond);
1403 else
1404 free (element);
1407 /* Build a new conditional using NEW_CODE, OP0 and OP1 and store
1408 the new conditional into *p, then store a boolean_true_node
1409 into the the *(p + 1). */
1411 static void
1412 build_and_record_new_cond (enum tree_code new_code, tree op0, tree op1, tree *p)
1414 *p = build2 (new_code, boolean_type_node, op0, op1);
1415 p++;
1416 *p = boolean_true_node;
1419 /* Record that COND is true and INVERTED is false into the edge information
1420 structure. Also record that any conditions dominated by COND are true
1421 as well.
1423 For example, if a < b is true, then a <= b must also be true. */
1425 static void
1426 record_conditions (struct edge_info *edge_info, tree cond, tree inverted)
1428 tree op0, op1;
1430 if (!COMPARISON_CLASS_P (cond))
1431 return;
1433 op0 = TREE_OPERAND (cond, 0);
1434 op1 = TREE_OPERAND (cond, 1);
1436 switch (TREE_CODE (cond))
1438 case LT_EXPR:
1439 case GT_EXPR:
1440 edge_info->max_cond_equivalences = 12;
1441 edge_info->cond_equivalences = xmalloc (12 * sizeof (tree));
1442 build_and_record_new_cond ((TREE_CODE (cond) == LT_EXPR
1443 ? LE_EXPR : GE_EXPR),
1444 op0, op1, &edge_info->cond_equivalences[4]);
1445 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1446 &edge_info->cond_equivalences[6]);
1447 build_and_record_new_cond (NE_EXPR, op0, op1,
1448 &edge_info->cond_equivalences[8]);
1449 build_and_record_new_cond (LTGT_EXPR, op0, op1,
1450 &edge_info->cond_equivalences[10]);
1451 break;
1453 case GE_EXPR:
1454 case LE_EXPR:
1455 edge_info->max_cond_equivalences = 6;
1456 edge_info->cond_equivalences = xmalloc (6 * sizeof (tree));
1457 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1458 &edge_info->cond_equivalences[4]);
1459 break;
1461 case EQ_EXPR:
1462 edge_info->max_cond_equivalences = 10;
1463 edge_info->cond_equivalences = xmalloc (10 * sizeof (tree));
1464 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1465 &edge_info->cond_equivalences[4]);
1466 build_and_record_new_cond (LE_EXPR, op0, op1,
1467 &edge_info->cond_equivalences[6]);
1468 build_and_record_new_cond (GE_EXPR, op0, op1,
1469 &edge_info->cond_equivalences[8]);
1470 break;
1472 case UNORDERED_EXPR:
1473 edge_info->max_cond_equivalences = 16;
1474 edge_info->cond_equivalences = xmalloc (16 * sizeof (tree));
1475 build_and_record_new_cond (NE_EXPR, op0, op1,
1476 &edge_info->cond_equivalences[4]);
1477 build_and_record_new_cond (UNLE_EXPR, op0, op1,
1478 &edge_info->cond_equivalences[6]);
1479 build_and_record_new_cond (UNGE_EXPR, op0, op1,
1480 &edge_info->cond_equivalences[8]);
1481 build_and_record_new_cond (UNEQ_EXPR, op0, op1,
1482 &edge_info->cond_equivalences[10]);
1483 build_and_record_new_cond (UNLT_EXPR, op0, op1,
1484 &edge_info->cond_equivalences[12]);
1485 build_and_record_new_cond (UNGT_EXPR, op0, op1,
1486 &edge_info->cond_equivalences[14]);
1487 break;
1489 case UNLT_EXPR:
1490 case UNGT_EXPR:
1491 edge_info->max_cond_equivalences = 8;
1492 edge_info->cond_equivalences = xmalloc (8 * sizeof (tree));
1493 build_and_record_new_cond ((TREE_CODE (cond) == UNLT_EXPR
1494 ? UNLE_EXPR : UNGE_EXPR),
1495 op0, op1, &edge_info->cond_equivalences[4]);
1496 build_and_record_new_cond (NE_EXPR, op0, op1,
1497 &edge_info->cond_equivalences[6]);
1498 break;
1500 case UNEQ_EXPR:
1501 edge_info->max_cond_equivalences = 8;
1502 edge_info->cond_equivalences = xmalloc (8 * sizeof (tree));
1503 build_and_record_new_cond (UNLE_EXPR, op0, op1,
1504 &edge_info->cond_equivalences[4]);
1505 build_and_record_new_cond (UNGE_EXPR, op0, op1,
1506 &edge_info->cond_equivalences[6]);
1507 break;
1509 case LTGT_EXPR:
1510 edge_info->max_cond_equivalences = 8;
1511 edge_info->cond_equivalences = xmalloc (8 * sizeof (tree));
1512 build_and_record_new_cond (NE_EXPR, op0, op1,
1513 &edge_info->cond_equivalences[4]);
1514 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1515 &edge_info->cond_equivalences[6]);
1516 break;
1518 default:
1519 edge_info->max_cond_equivalences = 4;
1520 edge_info->cond_equivalences = xmalloc (4 * sizeof (tree));
1521 break;
1524 /* Now store the original true and false conditions into the first
1525 two slots. */
1526 edge_info->cond_equivalences[0] = cond;
1527 edge_info->cond_equivalences[1] = boolean_true_node;
1528 edge_info->cond_equivalences[2] = inverted;
1529 edge_info->cond_equivalences[3] = boolean_false_node;
1532 /* A helper function for record_const_or_copy and record_equality.
1533 Do the work of recording the value and undo info. */
1535 static void
1536 record_const_or_copy_1 (tree x, tree y, tree prev_x)
1538 SSA_NAME_VALUE (x) = y;
1540 VEC_safe_push (tree_on_heap, const_and_copies_stack, prev_x);
1541 VEC_safe_push (tree_on_heap, const_and_copies_stack, x);
1545 /* Return the loop depth of the basic block of the defining statement of X.
1546 This number should not be treated as absolutely correct because the loop
1547 information may not be completely up-to-date when dom runs. However, it
1548 will be relatively correct, and as more passes are taught to keep loop info
1549 up to date, the result will become more and more accurate. */
1551 static int
1552 loop_depth_of_name (tree x)
1554 tree defstmt;
1555 basic_block defbb;
1557 /* If it's not an SSA_NAME, we have no clue where the definition is. */
1558 if (TREE_CODE (x) != SSA_NAME)
1559 return 0;
1561 /* Otherwise return the loop depth of the defining statement's bb.
1562 Note that there may not actually be a bb for this statement, if the
1563 ssa_name is live on entry. */
1564 defstmt = SSA_NAME_DEF_STMT (x);
1565 defbb = bb_for_stmt (defstmt);
1566 if (!defbb)
1567 return 0;
1569 return defbb->loop_depth;
1573 /* Record that X is equal to Y in const_and_copies. Record undo
1574 information in the block-local vector. */
1576 static void
1577 record_const_or_copy (tree x, tree y)
1579 tree prev_x = SSA_NAME_VALUE (x);
1581 if (TREE_CODE (y) == SSA_NAME)
1583 tree tmp = SSA_NAME_VALUE (y);
1584 if (tmp)
1585 y = tmp;
1588 record_const_or_copy_1 (x, y, prev_x);
1591 /* Similarly, but assume that X and Y are the two operands of an EQ_EXPR.
1592 This constrains the cases in which we may treat this as assignment. */
1594 static void
1595 record_equality (tree x, tree y)
1597 tree prev_x = NULL, prev_y = NULL;
1599 if (TREE_CODE (x) == SSA_NAME)
1600 prev_x = SSA_NAME_VALUE (x);
1601 if (TREE_CODE (y) == SSA_NAME)
1602 prev_y = SSA_NAME_VALUE (y);
1604 /* If one of the previous values is invariant, or invariant in more loops
1605 (by depth), then use that.
1606 Otherwise it doesn't matter which value we choose, just so
1607 long as we canonicalize on one value. */
1608 if (TREE_INVARIANT (y))
1610 else if (TREE_INVARIANT (x) || (loop_depth_of_name (x) <= loop_depth_of_name (y)))
1611 prev_x = x, x = y, y = prev_x, prev_x = prev_y;
1612 else if (prev_x && TREE_INVARIANT (prev_x))
1613 x = y, y = prev_x, prev_x = prev_y;
1614 else if (prev_y && TREE_CODE (prev_y) != VALUE_HANDLE)
1615 y = prev_y;
1617 /* After the swapping, we must have one SSA_NAME. */
1618 if (TREE_CODE (x) != SSA_NAME)
1619 return;
1621 /* For IEEE, -0.0 == 0.0, so we don't necessarily know the sign of a
1622 variable compared against zero. If we're honoring signed zeros,
1623 then we cannot record this value unless we know that the value is
1624 nonzero. */
1625 if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (x)))
1626 && (TREE_CODE (y) != REAL_CST
1627 || REAL_VALUES_EQUAL (dconst0, TREE_REAL_CST (y))))
1628 return;
1630 record_const_or_copy_1 (x, y, prev_x);
1633 /* Return true, if it is ok to do folding of an associative expression.
1634 EXP is the tree for the associative expression. */
1636 static inline bool
1637 unsafe_associative_fp_binop (tree exp)
1639 enum tree_code code = TREE_CODE (exp);
1640 return !(!flag_unsafe_math_optimizations
1641 && (code == MULT_EXPR || code == PLUS_EXPR
1642 || code == MINUS_EXPR)
1643 && FLOAT_TYPE_P (TREE_TYPE (exp)));
1646 /* STMT is a MODIFY_EXPR for which we were unable to find RHS in the
1647 hash tables. Try to simplify the RHS using whatever equivalences
1648 we may have recorded.
1650 If we are able to simplify the RHS, then lookup the simplified form in
1651 the hash table and return the result. Otherwise return NULL. */
1653 static tree
1654 simplify_rhs_and_lookup_avail_expr (struct dom_walk_data *walk_data,
1655 tree stmt, int insert)
1657 tree rhs = TREE_OPERAND (stmt, 1);
1658 enum tree_code rhs_code = TREE_CODE (rhs);
1659 tree result = NULL;
1661 /* If we have lhs = ~x, look and see if we earlier had x = ~y.
1662 In which case we can change this statement to be lhs = y.
1663 Which can then be copy propagated.
1665 Similarly for negation. */
1666 if ((rhs_code == BIT_NOT_EXPR || rhs_code == NEGATE_EXPR)
1667 && TREE_CODE (TREE_OPERAND (rhs, 0)) == SSA_NAME)
1669 /* Get the definition statement for our RHS. */
1670 tree rhs_def_stmt = SSA_NAME_DEF_STMT (TREE_OPERAND (rhs, 0));
1672 /* See if the RHS_DEF_STMT has the same form as our statement. */
1673 if (TREE_CODE (rhs_def_stmt) == MODIFY_EXPR
1674 && TREE_CODE (TREE_OPERAND (rhs_def_stmt, 1)) == rhs_code)
1676 tree rhs_def_operand;
1678 rhs_def_operand = TREE_OPERAND (TREE_OPERAND (rhs_def_stmt, 1), 0);
1680 /* Verify that RHS_DEF_OPERAND is a suitable SSA variable. */
1681 if (TREE_CODE (rhs_def_operand) == SSA_NAME
1682 && ! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (rhs_def_operand))
1683 result = update_rhs_and_lookup_avail_expr (stmt,
1684 rhs_def_operand,
1685 insert);
1689 /* If we have z = (x OP C1), see if we earlier had x = y OP C2.
1690 If OP is associative, create and fold (y OP C2) OP C1 which
1691 should result in (y OP C3), use that as the RHS for the
1692 assignment. Add minus to this, as we handle it specially below. */
1693 if ((associative_tree_code (rhs_code) || rhs_code == MINUS_EXPR)
1694 && TREE_CODE (TREE_OPERAND (rhs, 0)) == SSA_NAME
1695 && is_gimple_min_invariant (TREE_OPERAND (rhs, 1)))
1697 tree rhs_def_stmt = SSA_NAME_DEF_STMT (TREE_OPERAND (rhs, 0));
1699 /* See if the RHS_DEF_STMT has the same form as our statement. */
1700 if (TREE_CODE (rhs_def_stmt) == MODIFY_EXPR)
1702 tree rhs_def_rhs = TREE_OPERAND (rhs_def_stmt, 1);
1703 enum tree_code rhs_def_code = TREE_CODE (rhs_def_rhs);
1705 if ((rhs_code == rhs_def_code && unsafe_associative_fp_binop (rhs))
1706 || (rhs_code == PLUS_EXPR && rhs_def_code == MINUS_EXPR)
1707 || (rhs_code == MINUS_EXPR && rhs_def_code == PLUS_EXPR))
1709 tree def_stmt_op0 = TREE_OPERAND (rhs_def_rhs, 0);
1710 tree def_stmt_op1 = TREE_OPERAND (rhs_def_rhs, 1);
1712 if (TREE_CODE (def_stmt_op0) == SSA_NAME
1713 && ! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def_stmt_op0)
1714 && is_gimple_min_invariant (def_stmt_op1))
1716 tree outer_const = TREE_OPERAND (rhs, 1);
1717 tree type = TREE_TYPE (TREE_OPERAND (stmt, 0));
1718 tree t;
1720 /* If we care about correct floating point results, then
1721 don't fold x + c1 - c2. Note that we need to take both
1722 the codes and the signs to figure this out. */
1723 if (FLOAT_TYPE_P (type)
1724 && !flag_unsafe_math_optimizations
1725 && (rhs_def_code == PLUS_EXPR
1726 || rhs_def_code == MINUS_EXPR))
1728 bool neg = false;
1730 neg ^= (rhs_code == MINUS_EXPR);
1731 neg ^= (rhs_def_code == MINUS_EXPR);
1732 neg ^= real_isneg (TREE_REAL_CST_PTR (outer_const));
1733 neg ^= real_isneg (TREE_REAL_CST_PTR (def_stmt_op1));
1735 if (neg)
1736 goto dont_fold_assoc;
1739 /* Ho hum. So fold will only operate on the outermost
1740 thingy that we give it, so we have to build the new
1741 expression in two pieces. This requires that we handle
1742 combinations of plus and minus. */
1743 if (rhs_def_code != rhs_code)
1745 if (rhs_def_code == MINUS_EXPR)
1746 t = build (MINUS_EXPR, type, outer_const, def_stmt_op1);
1747 else
1748 t = build (MINUS_EXPR, type, def_stmt_op1, outer_const);
1749 rhs_code = PLUS_EXPR;
1751 else if (rhs_def_code == MINUS_EXPR)
1752 t = build (PLUS_EXPR, type, def_stmt_op1, outer_const);
1753 else
1754 t = build (rhs_def_code, type, def_stmt_op1, outer_const);
1755 t = local_fold (t);
1756 t = build (rhs_code, type, def_stmt_op0, t);
1757 t = local_fold (t);
1759 /* If the result is a suitable looking gimple expression,
1760 then use it instead of the original for STMT. */
1761 if (TREE_CODE (t) == SSA_NAME
1762 || (UNARY_CLASS_P (t)
1763 && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME)
1764 || ((BINARY_CLASS_P (t) || COMPARISON_CLASS_P (t))
1765 && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME
1766 && is_gimple_val (TREE_OPERAND (t, 1))))
1767 result = update_rhs_and_lookup_avail_expr (stmt, t, insert);
1771 dont_fold_assoc:;
1774 /* Transform TRUNC_DIV_EXPR and TRUNC_MOD_EXPR into RSHIFT_EXPR
1775 and BIT_AND_EXPR respectively if the first operand is greater
1776 than zero and the second operand is an exact power of two. */
1777 if ((rhs_code == TRUNC_DIV_EXPR || rhs_code == TRUNC_MOD_EXPR)
1778 && INTEGRAL_TYPE_P (TREE_TYPE (TREE_OPERAND (rhs, 0)))
1779 && integer_pow2p (TREE_OPERAND (rhs, 1)))
1781 tree val;
1782 tree op = TREE_OPERAND (rhs, 0);
1784 if (TYPE_UNSIGNED (TREE_TYPE (op)))
1786 val = integer_one_node;
1788 else
1790 tree dummy_cond = walk_data->global_data;
1792 if (! dummy_cond)
1794 dummy_cond = build (GT_EXPR, boolean_type_node,
1795 op, integer_zero_node);
1796 dummy_cond = build (COND_EXPR, void_type_node,
1797 dummy_cond, NULL, NULL);
1798 walk_data->global_data = dummy_cond;
1800 else
1802 TREE_SET_CODE (COND_EXPR_COND (dummy_cond), GT_EXPR);
1803 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 0) = op;
1804 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 1)
1805 = integer_zero_node;
1807 val = simplify_cond_and_lookup_avail_expr (dummy_cond, NULL, false);
1810 if (val && integer_onep (val))
1812 tree t;
1813 tree op0 = TREE_OPERAND (rhs, 0);
1814 tree op1 = TREE_OPERAND (rhs, 1);
1816 if (rhs_code == TRUNC_DIV_EXPR)
1817 t = build (RSHIFT_EXPR, TREE_TYPE (op0), op0,
1818 build_int_cst (NULL_TREE, tree_log2 (op1)));
1819 else
1820 t = build (BIT_AND_EXPR, TREE_TYPE (op0), op0,
1821 local_fold (build (MINUS_EXPR, TREE_TYPE (op1),
1822 op1, integer_one_node)));
1824 result = update_rhs_and_lookup_avail_expr (stmt, t, insert);
1828 /* Transform ABS (X) into X or -X as appropriate. */
1829 if (rhs_code == ABS_EXPR
1830 && INTEGRAL_TYPE_P (TREE_TYPE (TREE_OPERAND (rhs, 0))))
1832 tree val;
1833 tree op = TREE_OPERAND (rhs, 0);
1834 tree type = TREE_TYPE (op);
1836 if (TYPE_UNSIGNED (type))
1838 val = integer_zero_node;
1840 else
1842 tree dummy_cond = walk_data->global_data;
1844 if (! dummy_cond)
1846 dummy_cond = build (LE_EXPR, boolean_type_node,
1847 op, integer_zero_node);
1848 dummy_cond = build (COND_EXPR, void_type_node,
1849 dummy_cond, NULL, NULL);
1850 walk_data->global_data = dummy_cond;
1852 else
1854 TREE_SET_CODE (COND_EXPR_COND (dummy_cond), LE_EXPR);
1855 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 0) = op;
1856 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 1)
1857 = build_int_cst (type, 0);
1859 val = simplify_cond_and_lookup_avail_expr (dummy_cond, NULL, false);
1861 if (!val)
1863 TREE_SET_CODE (COND_EXPR_COND (dummy_cond), GE_EXPR);
1864 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 0) = op;
1865 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 1)
1866 = build_int_cst (type, 0);
1868 val = simplify_cond_and_lookup_avail_expr (dummy_cond,
1869 NULL, false);
1871 if (val)
1873 if (integer_zerop (val))
1874 val = integer_one_node;
1875 else if (integer_onep (val))
1876 val = integer_zero_node;
1881 if (val
1882 && (integer_onep (val) || integer_zerop (val)))
1884 tree t;
1886 if (integer_onep (val))
1887 t = build1 (NEGATE_EXPR, TREE_TYPE (op), op);
1888 else
1889 t = op;
1891 result = update_rhs_and_lookup_avail_expr (stmt, t, insert);
1895 /* Optimize *"foo" into 'f'. This is done here rather than
1896 in fold to avoid problems with stuff like &*"foo". */
1897 if (TREE_CODE (rhs) == INDIRECT_REF || TREE_CODE (rhs) == ARRAY_REF)
1899 tree t = fold_read_from_constant_string (rhs);
1901 if (t)
1902 result = update_rhs_and_lookup_avail_expr (stmt, t, insert);
1905 return result;
1908 /* COND is a condition of the form:
1910 x == const or x != const
1912 Look back to x's defining statement and see if x is defined as
1914 x = (type) y;
1916 If const is unchanged if we convert it to type, then we can build
1917 the equivalent expression:
1920 y == const or y != const
1922 Which may allow further optimizations.
1924 Return the equivalent comparison or NULL if no such equivalent comparison
1925 was found. */
1927 static tree
1928 find_equivalent_equality_comparison (tree cond)
1930 tree op0 = TREE_OPERAND (cond, 0);
1931 tree op1 = TREE_OPERAND (cond, 1);
1932 tree def_stmt = SSA_NAME_DEF_STMT (op0);
1934 /* OP0 might have been a parameter, so first make sure it
1935 was defined by a MODIFY_EXPR. */
1936 if (def_stmt && TREE_CODE (def_stmt) == MODIFY_EXPR)
1938 tree def_rhs = TREE_OPERAND (def_stmt, 1);
1940 /* Now make sure the RHS of the MODIFY_EXPR is a typecast. */
1941 if ((TREE_CODE (def_rhs) == NOP_EXPR
1942 || TREE_CODE (def_rhs) == CONVERT_EXPR)
1943 && TREE_CODE (TREE_OPERAND (def_rhs, 0)) == SSA_NAME)
1945 tree def_rhs_inner = TREE_OPERAND (def_rhs, 0);
1946 tree def_rhs_inner_type = TREE_TYPE (def_rhs_inner);
1947 tree new;
1949 if (TYPE_PRECISION (def_rhs_inner_type)
1950 > TYPE_PRECISION (TREE_TYPE (def_rhs)))
1951 return NULL;
1953 /* What we want to prove is that if we convert OP1 to
1954 the type of the object inside the NOP_EXPR that the
1955 result is still equivalent to SRC.
1957 If that is true, the build and return new equivalent
1958 condition which uses the source of the typecast and the
1959 new constant (which has only changed its type). */
1960 new = build1 (TREE_CODE (def_rhs), def_rhs_inner_type, op1);
1961 new = local_fold (new);
1962 if (is_gimple_val (new) && tree_int_cst_equal (new, op1))
1963 return build (TREE_CODE (cond), TREE_TYPE (cond),
1964 def_rhs_inner, new);
1967 return NULL;
1970 /* STMT is a COND_EXPR for which we could not trivially determine its
1971 result. This routine attempts to find equivalent forms of the
1972 condition which we may be able to optimize better. It also
1973 uses simple value range propagation to optimize conditionals. */
1975 static tree
1976 simplify_cond_and_lookup_avail_expr (tree stmt,
1977 stmt_ann_t ann,
1978 int insert)
1980 tree cond = COND_EXPR_COND (stmt);
1982 if (COMPARISON_CLASS_P (cond))
1984 tree op0 = TREE_OPERAND (cond, 0);
1985 tree op1 = TREE_OPERAND (cond, 1);
1987 if (TREE_CODE (op0) == SSA_NAME && is_gimple_min_invariant (op1))
1989 int limit;
1990 tree low, high, cond_low, cond_high;
1991 int lowequal, highequal, swapped, no_overlap, subset, cond_inverted;
1992 varray_type vrp_records;
1993 struct vrp_element *element;
1994 struct vrp_hash_elt vrp_hash_elt, *vrp_hash_elt_p;
1995 void **slot;
1997 /* First see if we have test of an SSA_NAME against a constant
1998 where the SSA_NAME is defined by an earlier typecast which
1999 is irrelevant when performing tests against the given
2000 constant. */
2001 if (TREE_CODE (cond) == EQ_EXPR || TREE_CODE (cond) == NE_EXPR)
2003 tree new_cond = find_equivalent_equality_comparison (cond);
2005 if (new_cond)
2007 /* Update the statement to use the new equivalent
2008 condition. */
2009 COND_EXPR_COND (stmt) = new_cond;
2011 /* If this is not a real stmt, ann will be NULL and we
2012 avoid processing the operands. */
2013 if (ann)
2014 modify_stmt (stmt);
2016 /* Lookup the condition and return its known value if it
2017 exists. */
2018 new_cond = lookup_avail_expr (stmt, insert);
2019 if (new_cond)
2020 return new_cond;
2022 /* The operands have changed, so update op0 and op1. */
2023 op0 = TREE_OPERAND (cond, 0);
2024 op1 = TREE_OPERAND (cond, 1);
2028 /* Consult the value range records for this variable (if they exist)
2029 to see if we can eliminate or simplify this conditional.
2031 Note two tests are necessary to determine no records exist.
2032 First we have to see if the virtual array exists, if it
2033 exists, then we have to check its active size.
2035 Also note the vast majority of conditionals are not testing
2036 a variable which has had its range constrained by an earlier
2037 conditional. So this filter avoids a lot of unnecessary work. */
2038 vrp_hash_elt.var = op0;
2039 vrp_hash_elt.records = NULL;
2040 slot = htab_find_slot (vrp_data, &vrp_hash_elt, NO_INSERT);
2041 if (slot == NULL)
2042 return NULL;
2044 vrp_hash_elt_p = (struct vrp_hash_elt *) *slot;
2045 vrp_records = vrp_hash_elt_p->records;
2046 if (vrp_records == NULL)
2047 return NULL;
2049 limit = VARRAY_ACTIVE_SIZE (vrp_records);
2051 /* If we have no value range records for this variable, or we are
2052 unable to extract a range for this condition, then there is
2053 nothing to do. */
2054 if (limit == 0
2055 || ! extract_range_from_cond (cond, &cond_high,
2056 &cond_low, &cond_inverted))
2057 return NULL;
2059 /* We really want to avoid unnecessary computations of range
2060 info. So all ranges are computed lazily; this avoids a
2061 lot of unnecessary work. i.e., we record the conditional,
2062 but do not process how it constrains the variable's
2063 potential values until we know that processing the condition
2064 could be helpful.
2066 However, we do not want to have to walk a potentially long
2067 list of ranges, nor do we want to compute a variable's
2068 range more than once for a given path.
2070 Luckily, each time we encounter a conditional that can not
2071 be otherwise optimized we will end up here and we will
2072 compute the necessary range information for the variable
2073 used in this condition.
2075 Thus you can conclude that there will never be more than one
2076 conditional associated with a variable which has not been
2077 processed. So we never need to merge more than one new
2078 conditional into the current range.
2080 These properties also help us avoid unnecessary work. */
2081 element
2082 = (struct vrp_element *)VARRAY_GENERIC_PTR (vrp_records, limit - 1);
2084 if (element->high && element->low)
2086 /* The last element has been processed, so there is no range
2087 merging to do, we can simply use the high/low values
2088 recorded in the last element. */
2089 low = element->low;
2090 high = element->high;
2092 else
2094 tree tmp_high, tmp_low;
2095 int dummy;
2097 /* The last element has not been processed. Process it now. */
2098 extract_range_from_cond (element->cond, &tmp_high,
2099 &tmp_low, &dummy);
2101 /* If this is the only element, then no merging is necessary,
2102 the high/low values from extract_range_from_cond are all
2103 we need. */
2104 if (limit == 1)
2106 low = tmp_low;
2107 high = tmp_high;
2109 else
2111 /* Get the high/low value from the previous element. */
2112 struct vrp_element *prev
2113 = (struct vrp_element *)VARRAY_GENERIC_PTR (vrp_records,
2114 limit - 2);
2115 low = prev->low;
2116 high = prev->high;
2118 /* Merge in this element's range with the range from the
2119 previous element.
2121 The low value for the merged range is the maximum of
2122 the previous low value and the low value of this record.
2124 Similarly the high value for the merged range is the
2125 minimum of the previous high value and the high value of
2126 this record. */
2127 low = (tree_int_cst_compare (low, tmp_low) == 1
2128 ? low : tmp_low);
2129 high = (tree_int_cst_compare (high, tmp_high) == -1
2130 ? high : tmp_high);
2133 /* And record the computed range. */
2134 element->low = low;
2135 element->high = high;
2139 /* After we have constrained this variable's potential values,
2140 we try to determine the result of the given conditional.
2142 To simplify later tests, first determine if the current
2143 low value is the same low value as the conditional.
2144 Similarly for the current high value and the high value
2145 for the conditional. */
2146 lowequal = tree_int_cst_equal (low, cond_low);
2147 highequal = tree_int_cst_equal (high, cond_high);
2149 if (lowequal && highequal)
2150 return (cond_inverted ? boolean_false_node : boolean_true_node);
2152 /* To simplify the overlap/subset tests below we may want
2153 to swap the two ranges so that the larger of the two
2154 ranges occurs "first". */
2155 swapped = 0;
2156 if (tree_int_cst_compare (low, cond_low) == 1
2157 || (lowequal
2158 && tree_int_cst_compare (cond_high, high) == 1))
2160 tree temp;
2162 swapped = 1;
2163 temp = low;
2164 low = cond_low;
2165 cond_low = temp;
2166 temp = high;
2167 high = cond_high;
2168 cond_high = temp;
2171 /* Now determine if there is no overlap in the ranges
2172 or if the second range is a subset of the first range. */
2173 no_overlap = tree_int_cst_lt (high, cond_low);
2174 subset = tree_int_cst_compare (cond_high, high) != 1;
2176 /* If there was no overlap in the ranges, then this conditional
2177 always has a false value (unless we had to invert this
2178 conditional, in which case it always has a true value). */
2179 if (no_overlap)
2180 return (cond_inverted ? boolean_true_node : boolean_false_node);
2182 /* If the current range is a subset of the condition's range,
2183 then this conditional always has a true value (unless we
2184 had to invert this conditional, in which case it always
2185 has a true value). */
2186 if (subset && swapped)
2187 return (cond_inverted ? boolean_false_node : boolean_true_node);
2189 /* We were unable to determine the result of the conditional.
2190 However, we may be able to simplify the conditional. First
2191 merge the ranges in the same manner as range merging above. */
2192 low = tree_int_cst_compare (low, cond_low) == 1 ? low : cond_low;
2193 high = tree_int_cst_compare (high, cond_high) == -1 ? high : cond_high;
2195 /* If the range has converged to a single point, then turn this
2196 into an equality comparison. */
2197 if (TREE_CODE (cond) != EQ_EXPR
2198 && TREE_CODE (cond) != NE_EXPR
2199 && tree_int_cst_equal (low, high))
2201 TREE_SET_CODE (cond, EQ_EXPR);
2202 TREE_OPERAND (cond, 1) = high;
2206 return 0;
2209 /* STMT is a SWITCH_EXPR for which we could not trivially determine its
2210 result. This routine attempts to find equivalent forms of the
2211 condition which we may be able to optimize better. */
2213 static tree
2214 simplify_switch_and_lookup_avail_expr (tree stmt, int insert)
2216 tree cond = SWITCH_COND (stmt);
2217 tree def, to, ti;
2219 /* The optimization that we really care about is removing unnecessary
2220 casts. That will let us do much better in propagating the inferred
2221 constant at the switch target. */
2222 if (TREE_CODE (cond) == SSA_NAME)
2224 def = SSA_NAME_DEF_STMT (cond);
2225 if (TREE_CODE (def) == MODIFY_EXPR)
2227 def = TREE_OPERAND (def, 1);
2228 if (TREE_CODE (def) == NOP_EXPR)
2230 int need_precision;
2231 bool fail;
2233 def = TREE_OPERAND (def, 0);
2235 #ifdef ENABLE_CHECKING
2236 /* ??? Why was Jeff testing this? We are gimple... */
2237 gcc_assert (is_gimple_val (def));
2238 #endif
2240 to = TREE_TYPE (cond);
2241 ti = TREE_TYPE (def);
2243 /* If we have an extension that preserves value, then we
2244 can copy the source value into the switch. */
2246 need_precision = TYPE_PRECISION (ti);
2247 fail = false;
2248 if (TYPE_UNSIGNED (to) && !TYPE_UNSIGNED (ti))
2249 fail = true;
2250 else if (!TYPE_UNSIGNED (to) && TYPE_UNSIGNED (ti))
2251 need_precision += 1;
2252 if (TYPE_PRECISION (to) < need_precision)
2253 fail = true;
2255 if (!fail)
2257 SWITCH_COND (stmt) = def;
2258 modify_stmt (stmt);
2260 return lookup_avail_expr (stmt, insert);
2266 return 0;
2270 /* CONST_AND_COPIES is a table which maps an SSA_NAME to the current
2271 known value for that SSA_NAME (or NULL if no value is known).
2273 NONZERO_VARS is the set SSA_NAMES known to have a nonzero value,
2274 even if we don't know their precise value.
2276 Propagate values from CONST_AND_COPIES and NONZERO_VARS into the PHI
2277 nodes of the successors of BB. */
2279 static void
2280 cprop_into_successor_phis (basic_block bb, bitmap nonzero_vars)
2282 edge e;
2283 edge_iterator ei;
2285 /* This can get rather expensive if the implementation is naive in
2286 how it finds the phi alternative associated with a particular edge. */
2287 FOR_EACH_EDGE (e, ei, bb->succs)
2289 tree phi;
2290 int indx;
2292 /* If this is an abnormal edge, then we do not want to copy propagate
2293 into the PHI alternative associated with this edge. */
2294 if (e->flags & EDGE_ABNORMAL)
2295 continue;
2297 phi = phi_nodes (e->dest);
2298 if (! phi)
2299 continue;
2301 indx = e->dest_idx;
2302 for ( ; phi; phi = PHI_CHAIN (phi))
2304 tree new;
2305 use_operand_p orig_p;
2306 tree orig;
2308 /* The alternative may be associated with a constant, so verify
2309 it is an SSA_NAME before doing anything with it. */
2310 orig_p = PHI_ARG_DEF_PTR (phi, indx);
2311 orig = USE_FROM_PTR (orig_p);
2312 if (TREE_CODE (orig) != SSA_NAME)
2313 continue;
2315 /* If the alternative is known to have a nonzero value, record
2316 that fact in the PHI node itself for future use. */
2317 if (bitmap_bit_p (nonzero_vars, SSA_NAME_VERSION (orig)))
2318 PHI_ARG_NONZERO (phi, indx) = true;
2320 /* If we have *ORIG_P in our constant/copy table, then replace
2321 ORIG_P with its value in our constant/copy table. */
2322 new = SSA_NAME_VALUE (orig);
2323 if (new
2324 && (TREE_CODE (new) == SSA_NAME
2325 || is_gimple_min_invariant (new))
2326 && may_propagate_copy (orig, new))
2328 propagate_value (orig_p, new);
2334 /* We have finished optimizing BB, record any information implied by
2335 taking a specific outgoing edge from BB. */
2337 static void
2338 record_edge_info (basic_block bb)
2340 block_stmt_iterator bsi = bsi_last (bb);
2341 struct edge_info *edge_info;
2343 if (! bsi_end_p (bsi))
2345 tree stmt = bsi_stmt (bsi);
2347 if (stmt && TREE_CODE (stmt) == SWITCH_EXPR)
2349 tree cond = SWITCH_COND (stmt);
2351 if (TREE_CODE (cond) == SSA_NAME)
2353 tree labels = SWITCH_LABELS (stmt);
2354 int i, n_labels = TREE_VEC_LENGTH (labels);
2355 tree *info = xcalloc (n_basic_blocks, sizeof (tree));
2356 edge e;
2357 edge_iterator ei;
2359 for (i = 0; i < n_labels; i++)
2361 tree label = TREE_VEC_ELT (labels, i);
2362 basic_block target_bb = label_to_block (CASE_LABEL (label));
2364 if (CASE_HIGH (label)
2365 || !CASE_LOW (label)
2366 || info[target_bb->index])
2367 info[target_bb->index] = error_mark_node;
2368 else
2369 info[target_bb->index] = label;
2372 FOR_EACH_EDGE (e, ei, bb->succs)
2374 basic_block target_bb = e->dest;
2375 tree node = info[target_bb->index];
2377 if (node != NULL && node != error_mark_node)
2379 tree x = fold_convert (TREE_TYPE (cond), CASE_LOW (node));
2380 edge_info = allocate_edge_info (e);
2381 edge_info->lhs = cond;
2382 edge_info->rhs = x;
2385 free (info);
2389 /* A COND_EXPR may create equivalences too. */
2390 if (stmt && TREE_CODE (stmt) == COND_EXPR)
2392 tree cond = COND_EXPR_COND (stmt);
2393 edge true_edge;
2394 edge false_edge;
2396 extract_true_false_edges_from_block (bb, &true_edge, &false_edge);
2398 /* If the conditional is a single variable 'X', record 'X = 1'
2399 for the true edge and 'X = 0' on the false edge. */
2400 if (SSA_VAR_P (cond))
2402 struct edge_info *edge_info;
2404 edge_info = allocate_edge_info (true_edge);
2405 edge_info->lhs = cond;
2406 edge_info->rhs = constant_boolean_node (1, TREE_TYPE (cond));
2408 edge_info = allocate_edge_info (false_edge);
2409 edge_info->lhs = cond;
2410 edge_info->rhs = constant_boolean_node (0, TREE_TYPE (cond));
2412 /* Equality tests may create one or two equivalences. */
2413 else if (COMPARISON_CLASS_P (cond))
2415 tree op0 = TREE_OPERAND (cond, 0);
2416 tree op1 = TREE_OPERAND (cond, 1);
2418 /* Special case comparing booleans against a constant as we
2419 know the value of OP0 on both arms of the branch. i.e., we
2420 can record an equivalence for OP0 rather than COND. */
2421 if ((TREE_CODE (cond) == EQ_EXPR || TREE_CODE (cond) == NE_EXPR)
2422 && TREE_CODE (op0) == SSA_NAME
2423 && TREE_CODE (TREE_TYPE (op0)) == BOOLEAN_TYPE
2424 && is_gimple_min_invariant (op1))
2426 if (TREE_CODE (cond) == EQ_EXPR)
2428 edge_info = allocate_edge_info (true_edge);
2429 edge_info->lhs = op0;
2430 edge_info->rhs = (integer_zerop (op1)
2431 ? boolean_false_node
2432 : boolean_true_node);
2434 edge_info = allocate_edge_info (false_edge);
2435 edge_info->lhs = op0;
2436 edge_info->rhs = (integer_zerop (op1)
2437 ? boolean_true_node
2438 : boolean_false_node);
2440 else
2442 edge_info = allocate_edge_info (true_edge);
2443 edge_info->lhs = op0;
2444 edge_info->rhs = (integer_zerop (op1)
2445 ? boolean_true_node
2446 : boolean_false_node);
2448 edge_info = allocate_edge_info (false_edge);
2449 edge_info->lhs = op0;
2450 edge_info->rhs = (integer_zerop (op1)
2451 ? boolean_false_node
2452 : boolean_true_node);
2456 else if (is_gimple_min_invariant (op0)
2457 && (TREE_CODE (op1) == SSA_NAME
2458 || is_gimple_min_invariant (op1)))
2460 tree inverted = invert_truthvalue (cond);
2461 struct edge_info *edge_info;
2463 edge_info = allocate_edge_info (true_edge);
2464 record_conditions (edge_info, cond, inverted);
2466 if (TREE_CODE (cond) == EQ_EXPR)
2468 edge_info->lhs = op1;
2469 edge_info->rhs = op0;
2472 edge_info = allocate_edge_info (false_edge);
2473 record_conditions (edge_info, inverted, cond);
2475 if (TREE_CODE (cond) == NE_EXPR)
2477 edge_info->lhs = op1;
2478 edge_info->rhs = op0;
2482 else if (TREE_CODE (op0) == SSA_NAME
2483 && (is_gimple_min_invariant (op1)
2484 || TREE_CODE (op1) == SSA_NAME))
2486 tree inverted = invert_truthvalue (cond);
2487 struct edge_info *edge_info;
2489 edge_info = allocate_edge_info (true_edge);
2490 record_conditions (edge_info, cond, inverted);
2492 if (TREE_CODE (cond) == EQ_EXPR)
2494 edge_info->lhs = op0;
2495 edge_info->rhs = op1;
2498 edge_info = allocate_edge_info (false_edge);
2499 record_conditions (edge_info, inverted, cond);
2501 if (TREE_CODE (cond) == NE_EXPR)
2503 edge_info->lhs = op0;
2504 edge_info->rhs = op1;
2509 /* ??? TRUTH_NOT_EXPR can create an equivalence too. */
2514 /* Propagate information from BB to its outgoing edges.
2516 This can include equivalency information implied by control statements
2517 at the end of BB and const/copy propagation into PHIs in BB's
2518 successor blocks. */
2520 static void
2521 propagate_to_outgoing_edges (struct dom_walk_data *walk_data ATTRIBUTE_UNUSED,
2522 basic_block bb)
2525 record_edge_info (bb);
2526 cprop_into_successor_phis (bb, nonzero_vars);
2529 /* Search for redundant computations in STMT. If any are found, then
2530 replace them with the variable holding the result of the computation.
2532 If safe, record this expression into the available expression hash
2533 table. */
2535 static bool
2536 eliminate_redundant_computations (struct dom_walk_data *walk_data,
2537 tree stmt, stmt_ann_t ann)
2539 v_may_def_optype v_may_defs = V_MAY_DEF_OPS (ann);
2540 tree *expr_p, def = NULL_TREE;
2541 bool insert = true;
2542 tree cached_lhs;
2543 bool retval = false;
2545 if (TREE_CODE (stmt) == MODIFY_EXPR)
2546 def = TREE_OPERAND (stmt, 0);
2548 /* Certain expressions on the RHS can be optimized away, but can not
2549 themselves be entered into the hash tables. */
2550 if (ann->makes_aliased_stores
2551 || ! def
2552 || TREE_CODE (def) != SSA_NAME
2553 || SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def)
2554 || NUM_V_MAY_DEFS (v_may_defs) != 0)
2555 insert = false;
2557 /* Check if the expression has been computed before. */
2558 cached_lhs = lookup_avail_expr (stmt, insert);
2560 /* If this is an assignment and the RHS was not in the hash table,
2561 then try to simplify the RHS and lookup the new RHS in the
2562 hash table. */
2563 if (! cached_lhs && TREE_CODE (stmt) == MODIFY_EXPR)
2564 cached_lhs = simplify_rhs_and_lookup_avail_expr (walk_data, stmt, insert);
2565 /* Similarly if this is a COND_EXPR and we did not find its
2566 expression in the hash table, simplify the condition and
2567 try again. */
2568 else if (! cached_lhs && TREE_CODE (stmt) == COND_EXPR)
2569 cached_lhs = simplify_cond_and_lookup_avail_expr (stmt, ann, insert);
2570 /* Similarly for a SWITCH_EXPR. */
2571 else if (!cached_lhs && TREE_CODE (stmt) == SWITCH_EXPR)
2572 cached_lhs = simplify_switch_and_lookup_avail_expr (stmt, insert);
2574 opt_stats.num_exprs_considered++;
2576 /* Get a pointer to the expression we are trying to optimize. */
2577 if (TREE_CODE (stmt) == COND_EXPR)
2578 expr_p = &COND_EXPR_COND (stmt);
2579 else if (TREE_CODE (stmt) == SWITCH_EXPR)
2580 expr_p = &SWITCH_COND (stmt);
2581 else if (TREE_CODE (stmt) == RETURN_EXPR && TREE_OPERAND (stmt, 0))
2582 expr_p = &TREE_OPERAND (TREE_OPERAND (stmt, 0), 1);
2583 else
2584 expr_p = &TREE_OPERAND (stmt, 1);
2586 /* It is safe to ignore types here since we have already done
2587 type checking in the hashing and equality routines. In fact
2588 type checking here merely gets in the way of constant
2589 propagation. Also, make sure that it is safe to propagate
2590 CACHED_LHS into *EXPR_P. */
2591 if (cached_lhs
2592 && (TREE_CODE (cached_lhs) != SSA_NAME
2593 || may_propagate_copy (*expr_p, cached_lhs)))
2595 if (dump_file && (dump_flags & TDF_DETAILS))
2597 fprintf (dump_file, " Replaced redundant expr '");
2598 print_generic_expr (dump_file, *expr_p, dump_flags);
2599 fprintf (dump_file, "' with '");
2600 print_generic_expr (dump_file, cached_lhs, dump_flags);
2601 fprintf (dump_file, "'\n");
2604 opt_stats.num_re++;
2606 #if defined ENABLE_CHECKING
2607 gcc_assert (TREE_CODE (cached_lhs) == SSA_NAME
2608 || is_gimple_min_invariant (cached_lhs));
2609 #endif
2611 if (TREE_CODE (cached_lhs) == ADDR_EXPR
2612 || (POINTER_TYPE_P (TREE_TYPE (*expr_p))
2613 && is_gimple_min_invariant (cached_lhs)))
2614 retval = true;
2616 propagate_tree_value (expr_p, cached_lhs);
2617 modify_stmt (stmt);
2619 return retval;
2622 /* STMT, a MODIFY_EXPR, may create certain equivalences, in either
2623 the available expressions table or the const_and_copies table.
2624 Detect and record those equivalences. */
2626 static void
2627 record_equivalences_from_stmt (tree stmt,
2628 int may_optimize_p,
2629 stmt_ann_t ann)
2631 tree lhs = TREE_OPERAND (stmt, 0);
2632 enum tree_code lhs_code = TREE_CODE (lhs);
2633 int i;
2635 if (lhs_code == SSA_NAME)
2637 tree rhs = TREE_OPERAND (stmt, 1);
2639 /* Strip away any useless type conversions. */
2640 STRIP_USELESS_TYPE_CONVERSION (rhs);
2642 /* If the RHS of the assignment is a constant or another variable that
2643 may be propagated, register it in the CONST_AND_COPIES table. We
2644 do not need to record unwind data for this, since this is a true
2645 assignment and not an equivalence inferred from a comparison. All
2646 uses of this ssa name are dominated by this assignment, so unwinding
2647 just costs time and space. */
2648 if (may_optimize_p
2649 && (TREE_CODE (rhs) == SSA_NAME
2650 || is_gimple_min_invariant (rhs)))
2651 SSA_NAME_VALUE (lhs) = rhs;
2653 /* alloca never returns zero and the address of a non-weak symbol
2654 is never zero. NOP_EXPRs and CONVERT_EXPRs can be completely
2655 stripped as they do not affect this equivalence. */
2656 while (TREE_CODE (rhs) == NOP_EXPR
2657 || TREE_CODE (rhs) == CONVERT_EXPR)
2658 rhs = TREE_OPERAND (rhs, 0);
2660 if (alloca_call_p (rhs)
2661 || (TREE_CODE (rhs) == ADDR_EXPR
2662 && DECL_P (TREE_OPERAND (rhs, 0))
2663 && ! DECL_WEAK (TREE_OPERAND (rhs, 0))))
2664 record_var_is_nonzero (lhs);
2666 /* IOR of any value with a nonzero value will result in a nonzero
2667 value. Even if we do not know the exact result recording that
2668 the result is nonzero is worth the effort. */
2669 if (TREE_CODE (rhs) == BIT_IOR_EXPR
2670 && integer_nonzerop (TREE_OPERAND (rhs, 1)))
2671 record_var_is_nonzero (lhs);
2674 /* Look at both sides for pointer dereferences. If we find one, then
2675 the pointer must be nonnull and we can enter that equivalence into
2676 the hash tables. */
2677 if (flag_delete_null_pointer_checks)
2678 for (i = 0; i < 2; i++)
2680 tree t = TREE_OPERAND (stmt, i);
2682 /* Strip away any COMPONENT_REFs. */
2683 while (TREE_CODE (t) == COMPONENT_REF)
2684 t = TREE_OPERAND (t, 0);
2686 /* Now see if this is a pointer dereference. */
2687 if (INDIRECT_REF_P (t))
2689 tree op = TREE_OPERAND (t, 0);
2691 /* If the pointer is a SSA variable, then enter new
2692 equivalences into the hash table. */
2693 while (TREE_CODE (op) == SSA_NAME)
2695 tree def = SSA_NAME_DEF_STMT (op);
2697 record_var_is_nonzero (op);
2699 /* And walk up the USE-DEF chains noting other SSA_NAMEs
2700 which are known to have a nonzero value. */
2701 if (def
2702 && TREE_CODE (def) == MODIFY_EXPR
2703 && TREE_CODE (TREE_OPERAND (def, 1)) == NOP_EXPR)
2704 op = TREE_OPERAND (TREE_OPERAND (def, 1), 0);
2705 else
2706 break;
2711 /* A memory store, even an aliased store, creates a useful
2712 equivalence. By exchanging the LHS and RHS, creating suitable
2713 vops and recording the result in the available expression table,
2714 we may be able to expose more redundant loads. */
2715 if (!ann->has_volatile_ops
2716 && (TREE_CODE (TREE_OPERAND (stmt, 1)) == SSA_NAME
2717 || is_gimple_min_invariant (TREE_OPERAND (stmt, 1)))
2718 && !is_gimple_reg (lhs))
2720 tree rhs = TREE_OPERAND (stmt, 1);
2721 tree new;
2723 /* FIXME: If the LHS of the assignment is a bitfield and the RHS
2724 is a constant, we need to adjust the constant to fit into the
2725 type of the LHS. If the LHS is a bitfield and the RHS is not
2726 a constant, then we can not record any equivalences for this
2727 statement since we would need to represent the widening or
2728 narrowing of RHS. This fixes gcc.c-torture/execute/921016-1.c
2729 and should not be necessary if GCC represented bitfields
2730 properly. */
2731 if (lhs_code == COMPONENT_REF
2732 && DECL_BIT_FIELD (TREE_OPERAND (lhs, 1)))
2734 if (TREE_CONSTANT (rhs))
2735 rhs = widen_bitfield (rhs, TREE_OPERAND (lhs, 1), lhs);
2736 else
2737 rhs = NULL;
2739 /* If the value overflowed, then we can not use this equivalence. */
2740 if (rhs && ! is_gimple_min_invariant (rhs))
2741 rhs = NULL;
2744 if (rhs)
2746 /* Build a new statement with the RHS and LHS exchanged. */
2747 new = build (MODIFY_EXPR, TREE_TYPE (stmt), rhs, lhs);
2749 create_ssa_artficial_load_stmt (&(ann->operands), new);
2751 /* Finally enter the statement into the available expression
2752 table. */
2753 lookup_avail_expr (new, true);
2758 /* Replace *OP_P in STMT with any known equivalent value for *OP_P from
2759 CONST_AND_COPIES. */
2761 static bool
2762 cprop_operand (tree stmt, use_operand_p op_p)
2764 bool may_have_exposed_new_symbols = false;
2765 tree val;
2766 tree op = USE_FROM_PTR (op_p);
2768 /* If the operand has a known constant value or it is known to be a
2769 copy of some other variable, use the value or copy stored in
2770 CONST_AND_COPIES. */
2771 val = SSA_NAME_VALUE (op);
2772 if (val && TREE_CODE (val) != VALUE_HANDLE)
2774 tree op_type, val_type;
2776 /* Do not change the base variable in the virtual operand
2777 tables. That would make it impossible to reconstruct
2778 the renamed virtual operand if we later modify this
2779 statement. Also only allow the new value to be an SSA_NAME
2780 for propagation into virtual operands. */
2781 if (!is_gimple_reg (op)
2782 && (get_virtual_var (val) != get_virtual_var (op)
2783 || TREE_CODE (val) != SSA_NAME))
2784 return false;
2786 /* Do not replace hard register operands in asm statements. */
2787 if (TREE_CODE (stmt) == ASM_EXPR
2788 && !may_propagate_copy_into_asm (op))
2789 return false;
2791 /* Get the toplevel type of each operand. */
2792 op_type = TREE_TYPE (op);
2793 val_type = TREE_TYPE (val);
2795 /* While both types are pointers, get the type of the object
2796 pointed to. */
2797 while (POINTER_TYPE_P (op_type) && POINTER_TYPE_P (val_type))
2799 op_type = TREE_TYPE (op_type);
2800 val_type = TREE_TYPE (val_type);
2803 /* Make sure underlying types match before propagating a constant by
2804 converting the constant to the proper type. Note that convert may
2805 return a non-gimple expression, in which case we ignore this
2806 propagation opportunity. */
2807 if (TREE_CODE (val) != SSA_NAME)
2809 if (!lang_hooks.types_compatible_p (op_type, val_type))
2811 val = fold_convert (TREE_TYPE (op), val);
2812 if (!is_gimple_min_invariant (val))
2813 return false;
2817 /* Certain operands are not allowed to be copy propagated due
2818 to their interaction with exception handling and some GCC
2819 extensions. */
2820 else if (!may_propagate_copy (op, val))
2821 return false;
2823 /* Dump details. */
2824 if (dump_file && (dump_flags & TDF_DETAILS))
2826 fprintf (dump_file, " Replaced '");
2827 print_generic_expr (dump_file, op, dump_flags);
2828 fprintf (dump_file, "' with %s '",
2829 (TREE_CODE (val) != SSA_NAME ? "constant" : "variable"));
2830 print_generic_expr (dump_file, val, dump_flags);
2831 fprintf (dump_file, "'\n");
2834 /* If VAL is an ADDR_EXPR or a constant of pointer type, note
2835 that we may have exposed a new symbol for SSA renaming. */
2836 if (TREE_CODE (val) == ADDR_EXPR
2837 || (POINTER_TYPE_P (TREE_TYPE (op))
2838 && is_gimple_min_invariant (val)))
2839 may_have_exposed_new_symbols = true;
2841 propagate_value (op_p, val);
2843 /* And note that we modified this statement. This is now
2844 safe, even if we changed virtual operands since we will
2845 rescan the statement and rewrite its operands again. */
2846 modify_stmt (stmt);
2848 return may_have_exposed_new_symbols;
2851 /* CONST_AND_COPIES is a table which maps an SSA_NAME to the current
2852 known value for that SSA_NAME (or NULL if no value is known).
2854 Propagate values from CONST_AND_COPIES into the uses, vuses and
2855 v_may_def_ops of STMT. */
2857 static bool
2858 cprop_into_stmt (tree stmt)
2860 bool may_have_exposed_new_symbols = false;
2861 use_operand_p op_p;
2862 ssa_op_iter iter;
2863 tree rhs;
2865 FOR_EACH_SSA_USE_OPERAND (op_p, stmt, iter, SSA_OP_ALL_USES)
2867 if (TREE_CODE (USE_FROM_PTR (op_p)) == SSA_NAME)
2868 may_have_exposed_new_symbols |= cprop_operand (stmt, op_p);
2871 if (may_have_exposed_new_symbols)
2873 rhs = get_rhs (stmt);
2874 if (rhs && TREE_CODE (rhs) == ADDR_EXPR)
2875 recompute_tree_invarant_for_addr_expr (rhs);
2878 return may_have_exposed_new_symbols;
2882 /* Optimize the statement pointed by iterator SI.
2884 We try to perform some simplistic global redundancy elimination and
2885 constant propagation:
2887 1- To detect global redundancy, we keep track of expressions that have
2888 been computed in this block and its dominators. If we find that the
2889 same expression is computed more than once, we eliminate repeated
2890 computations by using the target of the first one.
2892 2- Constant values and copy assignments. This is used to do very
2893 simplistic constant and copy propagation. When a constant or copy
2894 assignment is found, we map the value on the RHS of the assignment to
2895 the variable in the LHS in the CONST_AND_COPIES table. */
2897 static void
2898 optimize_stmt (struct dom_walk_data *walk_data, basic_block bb,
2899 block_stmt_iterator si)
2901 stmt_ann_t ann;
2902 tree stmt;
2903 bool may_optimize_p;
2904 bool may_have_exposed_new_symbols = false;
2906 stmt = bsi_stmt (si);
2908 get_stmt_operands (stmt);
2909 ann = stmt_ann (stmt);
2910 opt_stats.num_stmts++;
2911 may_have_exposed_new_symbols = false;
2913 if (dump_file && (dump_flags & TDF_DETAILS))
2915 fprintf (dump_file, "Optimizing statement ");
2916 print_generic_stmt (dump_file, stmt, TDF_SLIM);
2919 /* Const/copy propagate into USES, VUSES and the RHS of V_MAY_DEFs. */
2920 may_have_exposed_new_symbols = cprop_into_stmt (stmt);
2922 /* If the statement has been modified with constant replacements,
2923 fold its RHS before checking for redundant computations. */
2924 if (ann->modified)
2926 /* Try to fold the statement making sure that STMT is kept
2927 up to date. */
2928 if (fold_stmt (bsi_stmt_ptr (si)))
2930 stmt = bsi_stmt (si);
2931 ann = stmt_ann (stmt);
2933 if (dump_file && (dump_flags & TDF_DETAILS))
2935 fprintf (dump_file, " Folded to: ");
2936 print_generic_stmt (dump_file, stmt, TDF_SLIM);
2940 /* Constant/copy propagation above may change the set of
2941 virtual operands associated with this statement. Folding
2942 may remove the need for some virtual operands.
2944 Indicate we will need to rescan and rewrite the statement. */
2945 may_have_exposed_new_symbols = true;
2948 /* Check for redundant computations. Do this optimization only
2949 for assignments that have no volatile ops and conditionals. */
2950 may_optimize_p = (!ann->has_volatile_ops
2951 && ((TREE_CODE (stmt) == RETURN_EXPR
2952 && TREE_OPERAND (stmt, 0)
2953 && TREE_CODE (TREE_OPERAND (stmt, 0)) == MODIFY_EXPR
2954 && ! (TREE_SIDE_EFFECTS
2955 (TREE_OPERAND (TREE_OPERAND (stmt, 0), 1))))
2956 || (TREE_CODE (stmt) == MODIFY_EXPR
2957 && ! TREE_SIDE_EFFECTS (TREE_OPERAND (stmt, 1)))
2958 || TREE_CODE (stmt) == COND_EXPR
2959 || TREE_CODE (stmt) == SWITCH_EXPR));
2961 if (may_optimize_p)
2962 may_have_exposed_new_symbols
2963 |= eliminate_redundant_computations (walk_data, stmt, ann);
2965 /* Record any additional equivalences created by this statement. */
2966 if (TREE_CODE (stmt) == MODIFY_EXPR)
2967 record_equivalences_from_stmt (stmt,
2968 may_optimize_p,
2969 ann);
2971 register_definitions_for_stmt (stmt);
2973 /* If STMT is a COND_EXPR and it was modified, then we may know
2974 where it goes. If that is the case, then mark the CFG as altered.
2976 This will cause us to later call remove_unreachable_blocks and
2977 cleanup_tree_cfg when it is safe to do so. It is not safe to
2978 clean things up here since removal of edges and such can trigger
2979 the removal of PHI nodes, which in turn can release SSA_NAMEs to
2980 the manager.
2982 That's all fine and good, except that once SSA_NAMEs are released
2983 to the manager, we must not call create_ssa_name until all references
2984 to released SSA_NAMEs have been eliminated.
2986 All references to the deleted SSA_NAMEs can not be eliminated until
2987 we remove unreachable blocks.
2989 We can not remove unreachable blocks until after we have completed
2990 any queued jump threading.
2992 We can not complete any queued jump threads until we have taken
2993 appropriate variables out of SSA form. Taking variables out of
2994 SSA form can call create_ssa_name and thus we lose.
2996 Ultimately I suspect we're going to need to change the interface
2997 into the SSA_NAME manager. */
2999 if (ann->modified)
3001 tree val = NULL;
3003 if (TREE_CODE (stmt) == COND_EXPR)
3004 val = COND_EXPR_COND (stmt);
3005 else if (TREE_CODE (stmt) == SWITCH_EXPR)
3006 val = SWITCH_COND (stmt);
3008 if (val && TREE_CODE (val) == INTEGER_CST && find_taken_edge (bb, val))
3009 cfg_altered = true;
3011 /* If we simplified a statement in such a way as to be shown that it
3012 cannot trap, update the eh information and the cfg to match. */
3013 if (maybe_clean_eh_stmt (stmt))
3015 bitmap_set_bit (need_eh_cleanup, bb->index);
3016 if (dump_file && (dump_flags & TDF_DETAILS))
3017 fprintf (dump_file, " Flagged to clear EH edges.\n");
3021 if (may_have_exposed_new_symbols)
3022 VEC_safe_push (tree_on_heap, stmts_to_rescan, bsi_stmt (si));
3025 /* Replace the RHS of STMT with NEW_RHS. If RHS can be found in the
3026 available expression hashtable, then return the LHS from the hash
3027 table.
3029 If INSERT is true, then we also update the available expression
3030 hash table to account for the changes made to STMT. */
3032 static tree
3033 update_rhs_and_lookup_avail_expr (tree stmt, tree new_rhs, bool insert)
3035 tree cached_lhs = NULL;
3037 /* Remove the old entry from the hash table. */
3038 if (insert)
3040 struct expr_hash_elt element;
3042 initialize_hash_element (stmt, NULL, &element);
3043 htab_remove_elt_with_hash (avail_exprs, &element, element.hash);
3046 /* Now update the RHS of the assignment. */
3047 TREE_OPERAND (stmt, 1) = new_rhs;
3049 /* Now lookup the updated statement in the hash table. */
3050 cached_lhs = lookup_avail_expr (stmt, insert);
3052 /* We have now called lookup_avail_expr twice with two different
3053 versions of this same statement, once in optimize_stmt, once here.
3055 We know the call in optimize_stmt did not find an existing entry
3056 in the hash table, so a new entry was created. At the same time
3057 this statement was pushed onto the AVAIL_EXPRS_STACK vector.
3059 If this call failed to find an existing entry on the hash table,
3060 then the new version of this statement was entered into the
3061 hash table. And this statement was pushed onto BLOCK_AVAIL_EXPR
3062 for the second time. So there are two copies on BLOCK_AVAIL_EXPRs
3064 If this call succeeded, we still have one copy of this statement
3065 on the BLOCK_AVAIL_EXPRs vector.
3067 For both cases, we need to pop the most recent entry off the
3068 BLOCK_AVAIL_EXPRs vector. For the case where we never found this
3069 statement in the hash tables, that will leave precisely one
3070 copy of this statement on BLOCK_AVAIL_EXPRs. For the case where
3071 we found a copy of this statement in the second hash table lookup
3072 we want _no_ copies of this statement in BLOCK_AVAIL_EXPRs. */
3073 if (insert)
3074 VEC_pop (tree_on_heap, avail_exprs_stack);
3076 /* And make sure we record the fact that we modified this
3077 statement. */
3078 modify_stmt (stmt);
3080 return cached_lhs;
3083 /* Search for an existing instance of STMT in the AVAIL_EXPRS table. If
3084 found, return its LHS. Otherwise insert STMT in the table and return
3085 NULL_TREE.
3087 Also, when an expression is first inserted in the AVAIL_EXPRS table, it
3088 is also added to the stack pointed by BLOCK_AVAIL_EXPRS_P, so that they
3089 can be removed when we finish processing this block and its children.
3091 NOTE: This function assumes that STMT is a MODIFY_EXPR node that
3092 contains no CALL_EXPR on its RHS and makes no volatile nor
3093 aliased references. */
3095 static tree
3096 lookup_avail_expr (tree stmt, bool insert)
3098 void **slot;
3099 tree lhs;
3100 tree temp;
3101 struct expr_hash_elt *element = xcalloc (sizeof (struct expr_hash_elt), 1);
3103 lhs = TREE_CODE (stmt) == MODIFY_EXPR ? TREE_OPERAND (stmt, 0) : NULL;
3105 initialize_hash_element (stmt, lhs, element);
3107 /* Don't bother remembering constant assignments and copy operations.
3108 Constants and copy operations are handled by the constant/copy propagator
3109 in optimize_stmt. */
3110 if (TREE_CODE (element->rhs) == SSA_NAME
3111 || is_gimple_min_invariant (element->rhs))
3113 free (element);
3114 return NULL_TREE;
3117 /* If this is an equality test against zero, see if we have recorded a
3118 nonzero value for the variable in question. */
3119 if ((TREE_CODE (element->rhs) == EQ_EXPR
3120 || TREE_CODE (element->rhs) == NE_EXPR)
3121 && TREE_CODE (TREE_OPERAND (element->rhs, 0)) == SSA_NAME
3122 && integer_zerop (TREE_OPERAND (element->rhs, 1)))
3124 int indx = SSA_NAME_VERSION (TREE_OPERAND (element->rhs, 0));
3126 if (bitmap_bit_p (nonzero_vars, indx))
3128 tree t = element->rhs;
3129 free (element);
3131 if (TREE_CODE (t) == EQ_EXPR)
3132 return boolean_false_node;
3133 else
3134 return boolean_true_node;
3138 /* Finally try to find the expression in the main expression hash table. */
3139 slot = htab_find_slot_with_hash (avail_exprs, element, element->hash,
3140 (insert ? INSERT : NO_INSERT));
3141 if (slot == NULL)
3143 free (element);
3144 return NULL_TREE;
3147 if (*slot == NULL)
3149 *slot = (void *) element;
3150 VEC_safe_push (tree_on_heap, avail_exprs_stack,
3151 stmt ? stmt : element->rhs);
3152 return NULL_TREE;
3155 /* Extract the LHS of the assignment so that it can be used as the current
3156 definition of another variable. */
3157 lhs = ((struct expr_hash_elt *)*slot)->lhs;
3159 /* See if the LHS appears in the CONST_AND_COPIES table. If it does, then
3160 use the value from the const_and_copies table. */
3161 if (TREE_CODE (lhs) == SSA_NAME)
3163 temp = SSA_NAME_VALUE (lhs);
3164 if (temp && TREE_CODE (temp) != VALUE_HANDLE)
3165 lhs = temp;
3168 free (element);
3169 return lhs;
3172 /* Given a condition COND, record into HI_P, LO_P and INVERTED_P the
3173 range of values that result in the conditional having a true value.
3175 Return true if we are successful in extracting a range from COND and
3176 false if we are unsuccessful. */
3178 static bool
3179 extract_range_from_cond (tree cond, tree *hi_p, tree *lo_p, int *inverted_p)
3181 tree op1 = TREE_OPERAND (cond, 1);
3182 tree high, low, type;
3183 int inverted;
3185 /* Experiments have shown that it's rarely, if ever useful to
3186 record ranges for enumerations. Presumably this is due to
3187 the fact that they're rarely used directly. They are typically
3188 cast into an integer type and used that way. */
3189 if (TREE_CODE (TREE_TYPE (op1)) != INTEGER_TYPE)
3190 return 0;
3192 type = TREE_TYPE (op1);
3194 switch (TREE_CODE (cond))
3196 case EQ_EXPR:
3197 high = low = op1;
3198 inverted = 0;
3199 break;
3201 case NE_EXPR:
3202 high = low = op1;
3203 inverted = 1;
3204 break;
3206 case GE_EXPR:
3207 low = op1;
3208 high = TYPE_MAX_VALUE (type);
3209 inverted = 0;
3210 break;
3212 case GT_EXPR:
3213 low = int_const_binop (PLUS_EXPR, op1, integer_one_node, 1);
3214 high = TYPE_MAX_VALUE (type);
3215 inverted = 0;
3216 break;
3218 case LE_EXPR:
3219 high = op1;
3220 low = TYPE_MIN_VALUE (type);
3221 inverted = 0;
3222 break;
3224 case LT_EXPR:
3225 high = int_const_binop (MINUS_EXPR, op1, integer_one_node, 1);
3226 low = TYPE_MIN_VALUE (type);
3227 inverted = 0;
3228 break;
3230 default:
3231 return 0;
3234 *hi_p = high;
3235 *lo_p = low;
3236 *inverted_p = inverted;
3237 return 1;
3240 /* Record a range created by COND for basic block BB. */
3242 static void
3243 record_range (tree cond, basic_block bb)
3245 enum tree_code code = TREE_CODE (cond);
3247 /* We explicitly ignore NE_EXPRs and all the unordered comparisons.
3248 They rarely allow for meaningful range optimizations and significantly
3249 complicate the implementation. */
3250 if ((code == LT_EXPR || code == LE_EXPR || code == GT_EXPR
3251 || code == GE_EXPR || code == EQ_EXPR)
3252 && TREE_CODE (TREE_TYPE (TREE_OPERAND (cond, 1))) == INTEGER_TYPE)
3254 struct vrp_hash_elt *vrp_hash_elt;
3255 struct vrp_element *element;
3256 varray_type *vrp_records_p;
3257 void **slot;
3260 vrp_hash_elt = xmalloc (sizeof (struct vrp_hash_elt));
3261 vrp_hash_elt->var = TREE_OPERAND (cond, 0);
3262 vrp_hash_elt->records = NULL;
3263 slot = htab_find_slot (vrp_data, vrp_hash_elt, INSERT);
3265 if (*slot == NULL)
3266 *slot = (void *) vrp_hash_elt;
3267 else
3268 free (vrp_hash_elt);
3270 vrp_hash_elt = (struct vrp_hash_elt *) *slot;
3271 vrp_records_p = &vrp_hash_elt->records;
3273 element = ggc_alloc (sizeof (struct vrp_element));
3274 element->low = NULL;
3275 element->high = NULL;
3276 element->cond = cond;
3277 element->bb = bb;
3279 if (*vrp_records_p == NULL)
3280 VARRAY_GENERIC_PTR_INIT (*vrp_records_p, 2, "vrp records");
3282 VARRAY_PUSH_GENERIC_PTR (*vrp_records_p, element);
3283 VEC_safe_push (tree_on_heap, vrp_variables_stack, TREE_OPERAND (cond, 0));
3287 /* Hashing and equality functions for VRP_DATA.
3289 Since this hash table is addressed by SSA_NAMEs, we can hash on
3290 their version number and equality can be determined with a
3291 pointer comparison. */
3293 static hashval_t
3294 vrp_hash (const void *p)
3296 tree var = ((struct vrp_hash_elt *)p)->var;
3298 return SSA_NAME_VERSION (var);
3301 static int
3302 vrp_eq (const void *p1, const void *p2)
3304 tree var1 = ((struct vrp_hash_elt *)p1)->var;
3305 tree var2 = ((struct vrp_hash_elt *)p2)->var;
3307 return var1 == var2;
3310 /* Hashing and equality functions for AVAIL_EXPRS. The table stores
3311 MODIFY_EXPR statements. We compute a value number for expressions using
3312 the code of the expression and the SSA numbers of its operands. */
3314 static hashval_t
3315 avail_expr_hash (const void *p)
3317 stmt_ann_t ann = ((struct expr_hash_elt *)p)->ann;
3318 tree rhs = ((struct expr_hash_elt *)p)->rhs;
3319 hashval_t val = 0;
3320 size_t i;
3321 vuse_optype vuses;
3323 /* iterative_hash_expr knows how to deal with any expression and
3324 deals with commutative operators as well, so just use it instead
3325 of duplicating such complexities here. */
3326 val = iterative_hash_expr (rhs, val);
3328 /* If the hash table entry is not associated with a statement, then we
3329 can just hash the expression and not worry about virtual operands
3330 and such. */
3331 if (!ann)
3332 return val;
3334 /* Add the SSA version numbers of every vuse operand. This is important
3335 because compound variables like arrays are not renamed in the
3336 operands. Rather, the rename is done on the virtual variable
3337 representing all the elements of the array. */
3338 vuses = VUSE_OPS (ann);
3339 for (i = 0; i < NUM_VUSES (vuses); i++)
3340 val = iterative_hash_expr (VUSE_OP (vuses, i), val);
3342 return val;
3345 static hashval_t
3346 real_avail_expr_hash (const void *p)
3348 return ((const struct expr_hash_elt *)p)->hash;
3351 static int
3352 avail_expr_eq (const void *p1, const void *p2)
3354 stmt_ann_t ann1 = ((struct expr_hash_elt *)p1)->ann;
3355 tree rhs1 = ((struct expr_hash_elt *)p1)->rhs;
3356 stmt_ann_t ann2 = ((struct expr_hash_elt *)p2)->ann;
3357 tree rhs2 = ((struct expr_hash_elt *)p2)->rhs;
3359 /* If they are the same physical expression, return true. */
3360 if (rhs1 == rhs2 && ann1 == ann2)
3361 return true;
3363 /* If their codes are not equal, then quit now. */
3364 if (TREE_CODE (rhs1) != TREE_CODE (rhs2))
3365 return false;
3367 /* In case of a collision, both RHS have to be identical and have the
3368 same VUSE operands. */
3369 if ((TREE_TYPE (rhs1) == TREE_TYPE (rhs2)
3370 || lang_hooks.types_compatible_p (TREE_TYPE (rhs1), TREE_TYPE (rhs2)))
3371 && operand_equal_p (rhs1, rhs2, OEP_PURE_SAME))
3373 vuse_optype ops1 = NULL;
3374 vuse_optype ops2 = NULL;
3375 size_t num_ops1 = 0;
3376 size_t num_ops2 = 0;
3377 size_t i;
3379 if (ann1)
3381 ops1 = VUSE_OPS (ann1);
3382 num_ops1 = NUM_VUSES (ops1);
3385 if (ann2)
3387 ops2 = VUSE_OPS (ann2);
3388 num_ops2 = NUM_VUSES (ops2);
3391 /* If the number of virtual uses is different, then we consider
3392 them not equal. */
3393 if (num_ops1 != num_ops2)
3394 return false;
3396 for (i = 0; i < num_ops1; i++)
3397 if (VUSE_OP (ops1, i) != VUSE_OP (ops2, i))
3398 return false;
3400 gcc_assert (((struct expr_hash_elt *)p1)->hash
3401 == ((struct expr_hash_elt *)p2)->hash);
3402 return true;
3405 return false;
3408 /* Given STMT and a pointer to the block local definitions BLOCK_DEFS_P,
3409 register register all objects set by this statement into BLOCK_DEFS_P
3410 and CURRDEFS. */
3412 static void
3413 register_definitions_for_stmt (tree stmt)
3415 tree def;
3416 ssa_op_iter iter;
3418 FOR_EACH_SSA_TREE_OPERAND (def, stmt, iter, SSA_OP_ALL_DEFS)
3421 /* FIXME: We shouldn't be registering new defs if the variable
3422 doesn't need to be renamed. */
3423 register_new_def (def, &block_defs_stack);