PR target/19236
[official-gcc.git] / gcc / tree-ssa-dom.c
blob67446b1432f39e20da57cfee9b21412108f5951a
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);
554 /* If the desired argument is not the same as this PHI's result
555 and it is set by a PHI in this block, then we can not thread
556 through this block. */
557 if (src != dst
558 && TREE_CODE (src) == SSA_NAME
559 && TREE_CODE (SSA_NAME_DEF_STMT (src)) == PHI_NODE
560 && bb_for_stmt (SSA_NAME_DEF_STMT (src)) == e->dest)
561 return;
563 record_const_or_copy (dst, src);
564 register_new_def (dst, &block_defs_stack);
567 for (bsi = bsi_start (e->dest); ! bsi_end_p (bsi); bsi_next (&bsi))
569 tree lhs, cached_lhs;
571 stmt = bsi_stmt (bsi);
573 /* Ignore empty statements and labels. */
574 if (IS_EMPTY_STMT (stmt) || TREE_CODE (stmt) == LABEL_EXPR)
575 continue;
577 /* If this is not a MODIFY_EXPR which sets an SSA_NAME to a new
578 value, then stop our search here. Ideally when we stop a
579 search we stop on a COND_EXPR or SWITCH_EXPR. */
580 if (TREE_CODE (stmt) != MODIFY_EXPR
581 || TREE_CODE (TREE_OPERAND (stmt, 0)) != SSA_NAME)
582 break;
584 /* At this point we have a statement which assigns an RHS to an
585 SSA_VAR on the LHS. We want to prove that the RHS is already
586 available and that its value is held in the current definition
587 of the LHS -- meaning that this assignment is a NOP when
588 reached via edge E. */
589 if (TREE_CODE (TREE_OPERAND (stmt, 1)) == SSA_NAME)
590 cached_lhs = TREE_OPERAND (stmt, 1);
591 else
592 cached_lhs = lookup_avail_expr (stmt, false);
594 lhs = TREE_OPERAND (stmt, 0);
596 /* This can happen if we thread around to the start of a loop. */
597 if (lhs == cached_lhs)
598 break;
600 /* If we did not find RHS in the hash table, then try again after
601 temporarily const/copy propagating the operands. */
602 if (!cached_lhs)
604 /* Copy the operands. */
605 stmt_ann_t ann = stmt_ann (stmt);
606 use_optype uses = USE_OPS (ann);
607 vuse_optype vuses = VUSE_OPS (ann);
608 tree *uses_copy = xcalloc (NUM_USES (uses), sizeof (tree));
609 tree *vuses_copy = xcalloc (NUM_VUSES (vuses), sizeof (tree));
610 unsigned int i;
612 /* Make a copy of the uses into USES_COPY, then cprop into
613 the use operands. */
614 for (i = 0; i < NUM_USES (uses); i++)
616 tree tmp = NULL;
618 uses_copy[i] = USE_OP (uses, i);
619 if (TREE_CODE (USE_OP (uses, i)) == SSA_NAME)
620 tmp = SSA_NAME_VALUE (USE_OP (uses, i));
621 if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
622 SET_USE_OP (uses, i, tmp);
625 /* Similarly for virtual uses. */
626 for (i = 0; i < NUM_VUSES (vuses); i++)
628 tree tmp = NULL;
630 vuses_copy[i] = VUSE_OP (vuses, i);
631 if (TREE_CODE (VUSE_OP (vuses, i)) == SSA_NAME)
632 tmp = SSA_NAME_VALUE (VUSE_OP (vuses, i));
633 if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
634 SET_VUSE_OP (vuses, i, tmp);
637 /* Try to lookup the new expression. */
638 cached_lhs = lookup_avail_expr (stmt, false);
640 /* Restore the statement's original uses/defs. */
641 for (i = 0; i < NUM_USES (uses); i++)
642 SET_USE_OP (uses, i, uses_copy[i]);
644 for (i = 0; i < NUM_VUSES (vuses); i++)
645 SET_VUSE_OP (vuses, i, vuses_copy[i]);
647 free (uses_copy);
648 free (vuses_copy);
650 /* If we still did not find the expression in the hash table,
651 then we can not ignore this statement. */
652 if (! cached_lhs)
653 break;
656 /* If the expression in the hash table was not assigned to an
657 SSA_NAME, then we can not ignore this statement. */
658 if (TREE_CODE (cached_lhs) != SSA_NAME)
659 break;
661 /* If we have different underlying variables, then we can not
662 ignore this statement. */
663 if (SSA_NAME_VAR (cached_lhs) != SSA_NAME_VAR (lhs))
664 break;
666 /* If CACHED_LHS does not represent the current value of the underlying
667 variable in CACHED_LHS/LHS, then we can not ignore this statement. */
668 if (var_ann (SSA_NAME_VAR (lhs))->current_def != cached_lhs)
669 break;
671 /* If we got here, then we can ignore this statement and continue
672 walking through the statements in the block looking for a threadable
673 COND_EXPR.
675 We want to record an equivalence lhs = cache_lhs so that if
676 the result of this statement is used later we can copy propagate
677 suitably. */
678 record_const_or_copy (lhs, cached_lhs);
679 register_new_def (lhs, &block_defs_stack);
682 /* If we stopped at a COND_EXPR or SWITCH_EXPR, then see if we know which
683 arm will be taken. */
684 if (stmt
685 && (TREE_CODE (stmt) == COND_EXPR
686 || TREE_CODE (stmt) == SWITCH_EXPR))
688 tree cond, cached_lhs;
690 /* Now temporarily cprop the operands and try to find the resulting
691 expression in the hash tables. */
692 if (TREE_CODE (stmt) == COND_EXPR)
693 cond = COND_EXPR_COND (stmt);
694 else
695 cond = SWITCH_COND (stmt);
697 if (COMPARISON_CLASS_P (cond))
699 tree dummy_cond, op0, op1;
700 enum tree_code cond_code;
702 op0 = TREE_OPERAND (cond, 0);
703 op1 = TREE_OPERAND (cond, 1);
704 cond_code = TREE_CODE (cond);
706 /* Get the current value of both operands. */
707 if (TREE_CODE (op0) == SSA_NAME)
709 tree tmp = SSA_NAME_VALUE (op0);
710 if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
711 op0 = tmp;
714 if (TREE_CODE (op1) == SSA_NAME)
716 tree tmp = SSA_NAME_VALUE (op1);
717 if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
718 op1 = tmp;
721 /* Stuff the operator and operands into our dummy conditional
722 expression, creating the dummy conditional if necessary. */
723 dummy_cond = walk_data->global_data;
724 if (! dummy_cond)
726 dummy_cond = build (cond_code, boolean_type_node, op0, op1);
727 dummy_cond = build (COND_EXPR, void_type_node,
728 dummy_cond, NULL, NULL);
729 walk_data->global_data = dummy_cond;
731 else
733 TREE_SET_CODE (COND_EXPR_COND (dummy_cond), cond_code);
734 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 0) = op0;
735 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 1) = op1;
738 /* If the conditional folds to an invariant, then we are done,
739 otherwise look it up in the hash tables. */
740 cached_lhs = local_fold (COND_EXPR_COND (dummy_cond));
741 if (! is_gimple_min_invariant (cached_lhs))
743 cached_lhs = lookup_avail_expr (dummy_cond, false);
744 if (!cached_lhs || ! is_gimple_min_invariant (cached_lhs))
745 cached_lhs = simplify_cond_and_lookup_avail_expr (dummy_cond,
746 NULL,
747 false);
750 /* We can have conditionals which just test the state of a
751 variable rather than use a relational operator. These are
752 simpler to handle. */
753 else if (TREE_CODE (cond) == SSA_NAME)
755 cached_lhs = cond;
756 cached_lhs = SSA_NAME_VALUE (cached_lhs);
757 if (cached_lhs && ! is_gimple_min_invariant (cached_lhs))
758 cached_lhs = 0;
760 else
761 cached_lhs = lookup_avail_expr (stmt, false);
763 if (cached_lhs)
765 edge taken_edge = find_taken_edge (e->dest, cached_lhs);
766 basic_block dest = (taken_edge ? taken_edge->dest : NULL);
768 if (dest == e->dest)
769 return;
771 /* If we have a known destination for the conditional, then
772 we can perform this optimization, which saves at least one
773 conditional jump each time it applies since we get to
774 bypass the conditional at our original destination. */
775 if (dest)
777 struct edge_info *edge_info;
779 update_bb_profile_for_threading (e->dest, EDGE_FREQUENCY (e),
780 e->count, taken_edge);
781 if (e->aux)
782 edge_info = e->aux;
783 else
784 edge_info = allocate_edge_info (e);
785 edge_info->redirection_target = taken_edge;
786 bb_ann (e->dest)->incoming_edge_threaded = true;
793 /* Initialize local stacks for this optimizer and record equivalences
794 upon entry to BB. Equivalences can come from the edge traversed to
795 reach BB or they may come from PHI nodes at the start of BB. */
797 static void
798 dom_opt_initialize_block (struct dom_walk_data *walk_data ATTRIBUTE_UNUSED,
799 basic_block bb)
801 if (dump_file && (dump_flags & TDF_DETAILS))
802 fprintf (dump_file, "\n\nOptimizing block #%d\n\n", bb->index);
804 /* Push a marker on the stacks of local information so that we know how
805 far to unwind when we finalize this block. */
806 VEC_safe_push (tree_on_heap, avail_exprs_stack, NULL_TREE);
807 VEC_safe_push (tree_on_heap, block_defs_stack, NULL_TREE);
808 VEC_safe_push (tree_on_heap, const_and_copies_stack, NULL_TREE);
809 VEC_safe_push (tree_on_heap, nonzero_vars_stack, NULL_TREE);
810 VEC_safe_push (tree_on_heap, vrp_variables_stack, NULL_TREE);
812 record_equivalences_from_incoming_edge (bb);
814 /* PHI nodes can create equivalences too. */
815 record_equivalences_from_phis (bb);
818 /* Given an expression EXPR (a relational expression or a statement),
819 initialize the hash table element pointed by by ELEMENT. */
821 static void
822 initialize_hash_element (tree expr, tree lhs, struct expr_hash_elt *element)
824 /* Hash table elements may be based on conditional expressions or statements.
826 For the former case, we have no annotation and we want to hash the
827 conditional expression. In the latter case we have an annotation and
828 we want to record the expression the statement evaluates. */
829 if (COMPARISON_CLASS_P (expr) || TREE_CODE (expr) == TRUTH_NOT_EXPR)
831 element->ann = NULL;
832 element->rhs = expr;
834 else if (TREE_CODE (expr) == COND_EXPR)
836 element->ann = stmt_ann (expr);
837 element->rhs = COND_EXPR_COND (expr);
839 else if (TREE_CODE (expr) == SWITCH_EXPR)
841 element->ann = stmt_ann (expr);
842 element->rhs = SWITCH_COND (expr);
844 else if (TREE_CODE (expr) == RETURN_EXPR && TREE_OPERAND (expr, 0))
846 element->ann = stmt_ann (expr);
847 element->rhs = TREE_OPERAND (TREE_OPERAND (expr, 0), 1);
849 else
851 element->ann = stmt_ann (expr);
852 element->rhs = TREE_OPERAND (expr, 1);
855 element->lhs = lhs;
856 element->hash = avail_expr_hash (element);
859 /* Remove all the expressions in LOCALS from TABLE, stopping when there are
860 LIMIT entries left in LOCALs. */
862 static void
863 remove_local_expressions_from_table (void)
865 /* Remove all the expressions made available in this block. */
866 while (VEC_length (tree_on_heap, avail_exprs_stack) > 0)
868 struct expr_hash_elt element;
869 tree expr = VEC_pop (tree_on_heap, avail_exprs_stack);
871 if (expr == NULL_TREE)
872 break;
874 initialize_hash_element (expr, NULL, &element);
875 htab_remove_elt_with_hash (avail_exprs, &element, element.hash);
879 /* Use the SSA_NAMES in LOCALS to restore TABLE to its original
880 state, stopping when there are LIMIT entries left in LOCALs. */
882 static void
883 restore_nonzero_vars_to_original_value (void)
885 while (VEC_length (tree_on_heap, nonzero_vars_stack) > 0)
887 tree name = VEC_pop (tree_on_heap, nonzero_vars_stack);
889 if (name == NULL)
890 break;
892 bitmap_clear_bit (nonzero_vars, SSA_NAME_VERSION (name));
896 /* Use the source/dest pairs in CONST_AND_COPIES_STACK to restore
897 CONST_AND_COPIES to its original state, stopping when we hit a
898 NULL marker. */
900 static void
901 restore_vars_to_original_value (void)
903 while (VEC_length (tree_on_heap, const_and_copies_stack) > 0)
905 tree prev_value, dest;
907 dest = VEC_pop (tree_on_heap, const_and_copies_stack);
909 if (dest == NULL)
910 break;
912 prev_value = VEC_pop (tree_on_heap, const_and_copies_stack);
913 SSA_NAME_VALUE (dest) = prev_value;
917 /* Similar to restore_vars_to_original_value, except that it restores
918 CURRDEFS to its original value. */
919 static void
920 restore_currdefs_to_original_value (void)
922 /* Restore CURRDEFS to its original state. */
923 while (VEC_length (tree_on_heap, block_defs_stack) > 0)
925 tree tmp = VEC_pop (tree_on_heap, block_defs_stack);
926 tree saved_def, var;
928 if (tmp == NULL_TREE)
929 break;
931 /* If we recorded an SSA_NAME, then make the SSA_NAME the current
932 definition of its underlying variable. If we recorded anything
933 else, it must have been an _DECL node and its current reaching
934 definition must have been NULL. */
935 if (TREE_CODE (tmp) == SSA_NAME)
937 saved_def = tmp;
938 var = SSA_NAME_VAR (saved_def);
940 else
942 saved_def = NULL;
943 var = tmp;
946 var_ann (var)->current_def = saved_def;
950 /* We have finished processing the dominator children of BB, perform
951 any finalization actions in preparation for leaving this node in
952 the dominator tree. */
954 static void
955 dom_opt_finalize_block (struct dom_walk_data *walk_data, basic_block bb)
957 tree last;
959 /* If we are at a leaf node in the dominator tree, see if we can thread
960 the edge from BB through its successor.
962 Do this before we remove entries from our equivalence tables. */
963 if (EDGE_COUNT (bb->succs) == 1
964 && (EDGE_SUCC (bb, 0)->flags & EDGE_ABNORMAL) == 0
965 && (get_immediate_dominator (CDI_DOMINATORS, EDGE_SUCC (bb, 0)->dest) != bb
966 || phi_nodes (EDGE_SUCC (bb, 0)->dest)))
969 thread_across_edge (walk_data, EDGE_SUCC (bb, 0));
971 else if ((last = last_stmt (bb))
972 && TREE_CODE (last) == COND_EXPR
973 && (COMPARISON_CLASS_P (COND_EXPR_COND (last))
974 || TREE_CODE (COND_EXPR_COND (last)) == SSA_NAME)
975 && EDGE_COUNT (bb->succs) == 2
976 && (EDGE_SUCC (bb, 0)->flags & EDGE_ABNORMAL) == 0
977 && (EDGE_SUCC (bb, 1)->flags & EDGE_ABNORMAL) == 0)
979 edge true_edge, false_edge;
981 extract_true_false_edges_from_block (bb, &true_edge, &false_edge);
983 /* If the THEN arm is the end of a dominator tree or has PHI nodes,
984 then try to thread through its edge. */
985 if (get_immediate_dominator (CDI_DOMINATORS, true_edge->dest) != bb
986 || phi_nodes (true_edge->dest))
988 struct edge_info *edge_info;
989 unsigned int i;
991 /* Push a marker onto the available expression stack so that we
992 unwind any expressions related to the TRUE arm before processing
993 the false arm below. */
994 VEC_safe_push (tree_on_heap, avail_exprs_stack, NULL_TREE);
995 VEC_safe_push (tree_on_heap, block_defs_stack, NULL_TREE);
996 VEC_safe_push (tree_on_heap, const_and_copies_stack, NULL_TREE);
998 edge_info = true_edge->aux;
1000 /* If we have info associated with this edge, record it into
1001 our equivalency tables. */
1002 if (edge_info)
1004 tree *cond_equivalences = edge_info->cond_equivalences;
1005 tree lhs = edge_info->lhs;
1006 tree rhs = edge_info->rhs;
1008 /* If we have a simple NAME = VALUE equivalency record it.
1009 Until the jump threading selection code improves, only
1010 do this if both the name and value are SSA_NAMEs with
1011 the same underlying variable to avoid missing threading
1012 opportunities. */
1013 if (lhs
1014 && TREE_CODE (COND_EXPR_COND (last)) == SSA_NAME
1015 && TREE_CODE (edge_info->rhs) == SSA_NAME
1016 && SSA_NAME_VAR (lhs) == SSA_NAME_VAR (rhs))
1017 record_const_or_copy (lhs, rhs);
1019 /* If we have 0 = COND or 1 = COND equivalences, record them
1020 into our expression hash tables. */
1021 if (cond_equivalences)
1022 for (i = 0; i < edge_info->max_cond_equivalences; i += 2)
1024 tree expr = cond_equivalences[i];
1025 tree value = cond_equivalences[i + 1];
1027 record_cond (expr, value);
1031 /* Now thread the edge. */
1032 thread_across_edge (walk_data, true_edge);
1034 /* And restore the various tables to their state before
1035 we threaded this edge. */
1036 remove_local_expressions_from_table ();
1037 restore_vars_to_original_value ();
1038 restore_currdefs_to_original_value ();
1041 /* Similarly for the ELSE arm. */
1042 if (get_immediate_dominator (CDI_DOMINATORS, false_edge->dest) != bb
1043 || phi_nodes (false_edge->dest))
1045 struct edge_info *edge_info;
1046 unsigned int i;
1048 edge_info = false_edge->aux;
1050 /* If we have info associated with this edge, record it into
1051 our equivalency tables. */
1052 if (edge_info)
1054 tree *cond_equivalences = edge_info->cond_equivalences;
1055 tree lhs = edge_info->lhs;
1056 tree rhs = edge_info->rhs;
1058 /* If we have a simple NAME = VALUE equivalency record it.
1059 Until the jump threading selection code improves, only
1060 do this if both the name and value are SSA_NAMEs with
1061 the same underlying variable to avoid missing threading
1062 opportunities. */
1063 if (lhs
1064 && TREE_CODE (COND_EXPR_COND (last)) == SSA_NAME)
1065 record_const_or_copy (lhs, rhs);
1067 /* If we have 0 = COND or 1 = COND equivalences, record them
1068 into our expression hash tables. */
1069 if (cond_equivalences)
1070 for (i = 0; i < edge_info->max_cond_equivalences; i += 2)
1072 tree expr = cond_equivalences[i];
1073 tree value = cond_equivalences[i + 1];
1075 record_cond (expr, value);
1079 thread_across_edge (walk_data, false_edge);
1081 /* No need to remove local expressions from our tables
1082 or restore vars to their original value as that will
1083 be done immediately below. */
1087 remove_local_expressions_from_table ();
1088 restore_nonzero_vars_to_original_value ();
1089 restore_vars_to_original_value ();
1090 restore_currdefs_to_original_value ();
1092 /* Remove VRP records associated with this basic block. They are no
1093 longer valid.
1095 To be efficient, we note which variables have had their values
1096 constrained in this block. So walk over each variable in the
1097 VRP_VARIABLEs array. */
1098 while (VEC_length (tree_on_heap, vrp_variables_stack) > 0)
1100 tree var = VEC_pop (tree_on_heap, vrp_variables_stack);
1101 struct vrp_hash_elt vrp_hash_elt, *vrp_hash_elt_p;
1102 void **slot;
1104 /* Each variable has a stack of value range records. We want to
1105 invalidate those associated with our basic block. So we walk
1106 the array backwards popping off records associated with our
1107 block. Once we hit a record not associated with our block
1108 we are done. */
1109 varray_type var_vrp_records;
1111 if (var == NULL)
1112 break;
1114 vrp_hash_elt.var = var;
1115 vrp_hash_elt.records = NULL;
1117 slot = htab_find_slot (vrp_data, &vrp_hash_elt, NO_INSERT);
1119 vrp_hash_elt_p = (struct vrp_hash_elt *) *slot;
1120 var_vrp_records = vrp_hash_elt_p->records;
1122 while (VARRAY_ACTIVE_SIZE (var_vrp_records) > 0)
1124 struct vrp_element *element
1125 = (struct vrp_element *)VARRAY_TOP_GENERIC_PTR (var_vrp_records);
1127 if (element->bb != bb)
1128 break;
1130 VARRAY_POP (var_vrp_records);
1134 /* If we queued any statements to rescan in this block, then
1135 go ahead and rescan them now. */
1136 while (VEC_length (tree_on_heap, stmts_to_rescan) > 0)
1138 tree stmt = VEC_last (tree_on_heap, stmts_to_rescan);
1139 basic_block stmt_bb = bb_for_stmt (stmt);
1141 if (stmt_bb != bb)
1142 break;
1144 VEC_pop (tree_on_heap, stmts_to_rescan);
1145 mark_new_vars_to_rename (stmt, vars_to_rename);
1149 /* PHI nodes can create equivalences too.
1151 Ignoring any alternatives which are the same as the result, if
1152 all the alternatives are equal, then the PHI node creates an
1153 equivalence.
1155 Additionally, if all the PHI alternatives are known to have a nonzero
1156 value, then the result of this PHI is known to have a nonzero value,
1157 even if we do not know its exact value. */
1159 static void
1160 record_equivalences_from_phis (basic_block bb)
1162 tree phi;
1164 for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi))
1166 tree lhs = PHI_RESULT (phi);
1167 tree rhs = NULL;
1168 int i;
1170 for (i = 0; i < PHI_NUM_ARGS (phi); i++)
1172 tree t = PHI_ARG_DEF (phi, i);
1174 /* Ignore alternatives which are the same as our LHS. Since
1175 LHS is a PHI_RESULT, it is known to be a SSA_NAME, so we
1176 can simply compare pointers. */
1177 if (lhs == t)
1178 continue;
1180 /* If we have not processed an alternative yet, then set
1181 RHS to this alternative. */
1182 if (rhs == NULL)
1183 rhs = t;
1184 /* If we have processed an alternative (stored in RHS), then
1185 see if it is equal to this one. If it isn't, then stop
1186 the search. */
1187 else if (! operand_equal_for_phi_arg_p (rhs, t))
1188 break;
1191 /* If we had no interesting alternatives, then all the RHS alternatives
1192 must have been the same as LHS. */
1193 if (!rhs)
1194 rhs = lhs;
1196 /* If we managed to iterate through each PHI alternative without
1197 breaking out of the loop, then we have a PHI which may create
1198 a useful equivalence. We do not need to record unwind data for
1199 this, since this is a true assignment and not an equivalence
1200 inferred from a comparison. All uses of this ssa name are dominated
1201 by this assignment, so unwinding just costs time and space. */
1202 if (i == PHI_NUM_ARGS (phi)
1203 && may_propagate_copy (lhs, rhs))
1204 SSA_NAME_VALUE (lhs) = rhs;
1206 /* Now see if we know anything about the nonzero property for the
1207 result of this PHI. */
1208 for (i = 0; i < PHI_NUM_ARGS (phi); i++)
1210 if (!PHI_ARG_NONZERO (phi, i))
1211 break;
1214 if (i == PHI_NUM_ARGS (phi))
1215 bitmap_set_bit (nonzero_vars, SSA_NAME_VERSION (PHI_RESULT (phi)));
1217 register_new_def (lhs, &block_defs_stack);
1221 /* Ignoring loop backedges, if BB has precisely one incoming edge then
1222 return that edge. Otherwise return NULL. */
1223 static edge
1224 single_incoming_edge_ignoring_loop_edges (basic_block bb)
1226 edge retval = NULL;
1227 edge e;
1228 edge_iterator ei;
1230 FOR_EACH_EDGE (e, ei, bb->preds)
1232 /* A loop back edge can be identified by the destination of
1233 the edge dominating the source of the edge. */
1234 if (dominated_by_p (CDI_DOMINATORS, e->src, e->dest))
1235 continue;
1237 /* If we have already seen a non-loop edge, then we must have
1238 multiple incoming non-loop edges and thus we return NULL. */
1239 if (retval)
1240 return NULL;
1242 /* This is the first non-loop incoming edge we have found. Record
1243 it. */
1244 retval = e;
1247 return retval;
1250 /* Record any equivalences created by the incoming edge to BB. If BB
1251 has more than one incoming edge, then no equivalence is created. */
1253 static void
1254 record_equivalences_from_incoming_edge (basic_block bb)
1256 edge e;
1257 basic_block parent;
1258 struct edge_info *edge_info;
1260 /* If our parent block ended with a control statment, then we may be
1261 able to record some equivalences based on which outgoing edge from
1262 the parent was followed. */
1263 parent = get_immediate_dominator (CDI_DOMINATORS, bb);
1265 e = single_incoming_edge_ignoring_loop_edges (bb);
1267 /* If we had a single incoming edge from our parent block, then enter
1268 any data associated with the edge into our tables. */
1269 if (e && e->src == parent)
1271 unsigned int i;
1273 edge_info = e->aux;
1275 if (edge_info)
1277 tree lhs = edge_info->lhs;
1278 tree rhs = edge_info->rhs;
1279 tree *cond_equivalences = edge_info->cond_equivalences;
1281 if (lhs)
1282 record_equality (lhs, rhs);
1284 if (cond_equivalences)
1286 bool recorded_range = false;
1287 for (i = 0; i < edge_info->max_cond_equivalences; i += 2)
1289 tree expr = cond_equivalences[i];
1290 tree value = cond_equivalences[i + 1];
1292 record_cond (expr, value);
1294 /* For the first true equivalence, record range
1295 information. We only do this for the first
1296 true equivalence as it should dominate any
1297 later true equivalences. */
1298 if (! recorded_range
1299 && COMPARISON_CLASS_P (expr)
1300 && value == boolean_true_node
1301 && TREE_CONSTANT (TREE_OPERAND (expr, 1)))
1303 record_range (expr, bb);
1304 recorded_range = true;
1312 /* Dump SSA statistics on FILE. */
1314 void
1315 dump_dominator_optimization_stats (FILE *file)
1317 long n_exprs;
1319 fprintf (file, "Total number of statements: %6ld\n\n",
1320 opt_stats.num_stmts);
1321 fprintf (file, "Exprs considered for dominator optimizations: %6ld\n",
1322 opt_stats.num_exprs_considered);
1324 n_exprs = opt_stats.num_exprs_considered;
1325 if (n_exprs == 0)
1326 n_exprs = 1;
1328 fprintf (file, " Redundant expressions eliminated: %6ld (%.0f%%)\n",
1329 opt_stats.num_re, PERCENT (opt_stats.num_re,
1330 n_exprs));
1332 fprintf (file, "\nHash table statistics:\n");
1334 fprintf (file, " avail_exprs: ");
1335 htab_statistics (file, avail_exprs);
1339 /* Dump SSA statistics on stderr. */
1341 void
1342 debug_dominator_optimization_stats (void)
1344 dump_dominator_optimization_stats (stderr);
1348 /* Dump statistics for the hash table HTAB. */
1350 static void
1351 htab_statistics (FILE *file, htab_t htab)
1353 fprintf (file, "size %ld, %ld elements, %f collision/search ratio\n",
1354 (long) htab_size (htab),
1355 (long) htab_elements (htab),
1356 htab_collisions (htab));
1359 /* Record the fact that VAR has a nonzero value, though we may not know
1360 its exact value. Note that if VAR is already known to have a nonzero
1361 value, then we do nothing. */
1363 static void
1364 record_var_is_nonzero (tree var)
1366 int indx = SSA_NAME_VERSION (var);
1368 if (bitmap_bit_p (nonzero_vars, indx))
1369 return;
1371 /* Mark it in the global table. */
1372 bitmap_set_bit (nonzero_vars, indx);
1374 /* Record this SSA_NAME so that we can reset the global table
1375 when we leave this block. */
1376 VEC_safe_push (tree_on_heap, nonzero_vars_stack, var);
1379 /* Enter a statement into the true/false expression hash table indicating
1380 that the condition COND has the value VALUE. */
1382 static void
1383 record_cond (tree cond, tree value)
1385 struct expr_hash_elt *element = xmalloc (sizeof (struct expr_hash_elt));
1386 void **slot;
1388 initialize_hash_element (cond, value, element);
1390 slot = htab_find_slot_with_hash (avail_exprs, (void *)element,
1391 element->hash, true);
1392 if (*slot == NULL)
1394 *slot = (void *) element;
1395 VEC_safe_push (tree_on_heap, avail_exprs_stack, cond);
1397 else
1398 free (element);
1401 /* Build a new conditional using NEW_CODE, OP0 and OP1 and store
1402 the new conditional into *p, then store a boolean_true_node
1403 into the the *(p + 1). */
1405 static void
1406 build_and_record_new_cond (enum tree_code new_code, tree op0, tree op1, tree *p)
1408 *p = build2 (new_code, boolean_type_node, op0, op1);
1409 p++;
1410 *p = boolean_true_node;
1413 /* Record that COND is true and INVERTED is false into the edge information
1414 structure. Also record that any conditions dominated by COND are true
1415 as well.
1417 For example, if a < b is true, then a <= b must also be true. */
1419 static void
1420 record_conditions (struct edge_info *edge_info, tree cond, tree inverted)
1422 tree op0, op1;
1424 if (!COMPARISON_CLASS_P (cond))
1425 return;
1427 op0 = TREE_OPERAND (cond, 0);
1428 op1 = TREE_OPERAND (cond, 1);
1430 switch (TREE_CODE (cond))
1432 case LT_EXPR:
1433 case GT_EXPR:
1434 edge_info->max_cond_equivalences = 12;
1435 edge_info->cond_equivalences = xmalloc (12 * sizeof (tree));
1436 build_and_record_new_cond ((TREE_CODE (cond) == LT_EXPR
1437 ? LE_EXPR : GE_EXPR),
1438 op0, op1, &edge_info->cond_equivalences[4]);
1439 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1440 &edge_info->cond_equivalences[6]);
1441 build_and_record_new_cond (NE_EXPR, op0, op1,
1442 &edge_info->cond_equivalences[8]);
1443 build_and_record_new_cond (LTGT_EXPR, op0, op1,
1444 &edge_info->cond_equivalences[10]);
1445 break;
1447 case GE_EXPR:
1448 case LE_EXPR:
1449 edge_info->max_cond_equivalences = 6;
1450 edge_info->cond_equivalences = xmalloc (6 * sizeof (tree));
1451 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1452 &edge_info->cond_equivalences[4]);
1453 break;
1455 case EQ_EXPR:
1456 edge_info->max_cond_equivalences = 10;
1457 edge_info->cond_equivalences = xmalloc (10 * sizeof (tree));
1458 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1459 &edge_info->cond_equivalences[4]);
1460 build_and_record_new_cond (LE_EXPR, op0, op1,
1461 &edge_info->cond_equivalences[6]);
1462 build_and_record_new_cond (GE_EXPR, op0, op1,
1463 &edge_info->cond_equivalences[8]);
1464 break;
1466 case UNORDERED_EXPR:
1467 edge_info->max_cond_equivalences = 16;
1468 edge_info->cond_equivalences = xmalloc (16 * sizeof (tree));
1469 build_and_record_new_cond (NE_EXPR, op0, op1,
1470 &edge_info->cond_equivalences[4]);
1471 build_and_record_new_cond (UNLE_EXPR, op0, op1,
1472 &edge_info->cond_equivalences[6]);
1473 build_and_record_new_cond (UNGE_EXPR, op0, op1,
1474 &edge_info->cond_equivalences[8]);
1475 build_and_record_new_cond (UNEQ_EXPR, op0, op1,
1476 &edge_info->cond_equivalences[10]);
1477 build_and_record_new_cond (UNLT_EXPR, op0, op1,
1478 &edge_info->cond_equivalences[12]);
1479 build_and_record_new_cond (UNGT_EXPR, op0, op1,
1480 &edge_info->cond_equivalences[14]);
1481 break;
1483 case UNLT_EXPR:
1484 case UNGT_EXPR:
1485 edge_info->max_cond_equivalences = 8;
1486 edge_info->cond_equivalences = xmalloc (8 * sizeof (tree));
1487 build_and_record_new_cond ((TREE_CODE (cond) == UNLT_EXPR
1488 ? UNLE_EXPR : UNGE_EXPR),
1489 op0, op1, &edge_info->cond_equivalences[4]);
1490 build_and_record_new_cond (NE_EXPR, op0, op1,
1491 &edge_info->cond_equivalences[6]);
1492 break;
1494 case UNEQ_EXPR:
1495 edge_info->max_cond_equivalences = 8;
1496 edge_info->cond_equivalences = xmalloc (8 * sizeof (tree));
1497 build_and_record_new_cond (UNLE_EXPR, op0, op1,
1498 &edge_info->cond_equivalences[4]);
1499 build_and_record_new_cond (UNGE_EXPR, op0, op1,
1500 &edge_info->cond_equivalences[6]);
1501 break;
1503 case LTGT_EXPR:
1504 edge_info->max_cond_equivalences = 8;
1505 edge_info->cond_equivalences = xmalloc (8 * sizeof (tree));
1506 build_and_record_new_cond (NE_EXPR, op0, op1,
1507 &edge_info->cond_equivalences[4]);
1508 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1509 &edge_info->cond_equivalences[6]);
1510 break;
1512 default:
1513 edge_info->max_cond_equivalences = 4;
1514 edge_info->cond_equivalences = xmalloc (4 * sizeof (tree));
1515 break;
1518 /* Now store the original true and false conditions into the first
1519 two slots. */
1520 edge_info->cond_equivalences[0] = cond;
1521 edge_info->cond_equivalences[1] = boolean_true_node;
1522 edge_info->cond_equivalences[2] = inverted;
1523 edge_info->cond_equivalences[3] = boolean_false_node;
1526 /* A helper function for record_const_or_copy and record_equality.
1527 Do the work of recording the value and undo info. */
1529 static void
1530 record_const_or_copy_1 (tree x, tree y, tree prev_x)
1532 SSA_NAME_VALUE (x) = y;
1534 VEC_safe_push (tree_on_heap, const_and_copies_stack, prev_x);
1535 VEC_safe_push (tree_on_heap, const_and_copies_stack, x);
1539 /* Return the loop depth of the basic block of the defining statement of X.
1540 This number should not be treated as absolutely correct because the loop
1541 information may not be completely up-to-date when dom runs. However, it
1542 will be relatively correct, and as more passes are taught to keep loop info
1543 up to date, the result will become more and more accurate. */
1545 static int
1546 loop_depth_of_name (tree x)
1548 tree defstmt;
1549 basic_block defbb;
1551 /* If it's not an SSA_NAME, we have no clue where the definition is. */
1552 if (TREE_CODE (x) != SSA_NAME)
1553 return 0;
1555 /* Otherwise return the loop depth of the defining statement's bb.
1556 Note that there may not actually be a bb for this statement, if the
1557 ssa_name is live on entry. */
1558 defstmt = SSA_NAME_DEF_STMT (x);
1559 defbb = bb_for_stmt (defstmt);
1560 if (!defbb)
1561 return 0;
1563 return defbb->loop_depth;
1567 /* Record that X is equal to Y in const_and_copies. Record undo
1568 information in the block-local vector. */
1570 static void
1571 record_const_or_copy (tree x, tree y)
1573 tree prev_x = SSA_NAME_VALUE (x);
1575 if (TREE_CODE (y) == SSA_NAME)
1577 tree tmp = SSA_NAME_VALUE (y);
1578 if (tmp)
1579 y = tmp;
1582 record_const_or_copy_1 (x, y, prev_x);
1585 /* Similarly, but assume that X and Y are the two operands of an EQ_EXPR.
1586 This constrains the cases in which we may treat this as assignment. */
1588 static void
1589 record_equality (tree x, tree y)
1591 tree prev_x = NULL, prev_y = NULL;
1593 if (TREE_CODE (x) == SSA_NAME)
1594 prev_x = SSA_NAME_VALUE (x);
1595 if (TREE_CODE (y) == SSA_NAME)
1596 prev_y = SSA_NAME_VALUE (y);
1598 /* If one of the previous values is invariant, or invariant in more loops
1599 (by depth), then use that.
1600 Otherwise it doesn't matter which value we choose, just so
1601 long as we canonicalize on one value. */
1602 if (TREE_INVARIANT (y))
1604 else if (TREE_INVARIANT (x) || (loop_depth_of_name (x) <= loop_depth_of_name (y)))
1605 prev_x = x, x = y, y = prev_x, prev_x = prev_y;
1606 else if (prev_x && TREE_INVARIANT (prev_x))
1607 x = y, y = prev_x, prev_x = prev_y;
1608 else if (prev_y && TREE_CODE (prev_y) != VALUE_HANDLE)
1609 y = prev_y;
1611 /* After the swapping, we must have one SSA_NAME. */
1612 if (TREE_CODE (x) != SSA_NAME)
1613 return;
1615 /* For IEEE, -0.0 == 0.0, so we don't necessarily know the sign of a
1616 variable compared against zero. If we're honoring signed zeros,
1617 then we cannot record this value unless we know that the value is
1618 nonzero. */
1619 if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (x)))
1620 && (TREE_CODE (y) != REAL_CST
1621 || REAL_VALUES_EQUAL (dconst0, TREE_REAL_CST (y))))
1622 return;
1624 record_const_or_copy_1 (x, y, prev_x);
1627 /* Return true, if it is ok to do folding of an associative expression.
1628 EXP is the tree for the associative expression. */
1630 static inline bool
1631 unsafe_associative_fp_binop (tree exp)
1633 enum tree_code code = TREE_CODE (exp);
1634 return !(!flag_unsafe_math_optimizations
1635 && (code == MULT_EXPR || code == PLUS_EXPR
1636 || code == MINUS_EXPR)
1637 && FLOAT_TYPE_P (TREE_TYPE (exp)));
1640 /* STMT is a MODIFY_EXPR for which we were unable to find RHS in the
1641 hash tables. Try to simplify the RHS using whatever equivalences
1642 we may have recorded.
1644 If we are able to simplify the RHS, then lookup the simplified form in
1645 the hash table and return the result. Otherwise return NULL. */
1647 static tree
1648 simplify_rhs_and_lookup_avail_expr (struct dom_walk_data *walk_data,
1649 tree stmt, int insert)
1651 tree rhs = TREE_OPERAND (stmt, 1);
1652 enum tree_code rhs_code = TREE_CODE (rhs);
1653 tree result = NULL;
1655 /* If we have lhs = ~x, look and see if we earlier had x = ~y.
1656 In which case we can change this statement to be lhs = y.
1657 Which can then be copy propagated.
1659 Similarly for negation. */
1660 if ((rhs_code == BIT_NOT_EXPR || rhs_code == NEGATE_EXPR)
1661 && TREE_CODE (TREE_OPERAND (rhs, 0)) == SSA_NAME)
1663 /* Get the definition statement for our RHS. */
1664 tree rhs_def_stmt = SSA_NAME_DEF_STMT (TREE_OPERAND (rhs, 0));
1666 /* See if the RHS_DEF_STMT has the same form as our statement. */
1667 if (TREE_CODE (rhs_def_stmt) == MODIFY_EXPR
1668 && TREE_CODE (TREE_OPERAND (rhs_def_stmt, 1)) == rhs_code)
1670 tree rhs_def_operand;
1672 rhs_def_operand = TREE_OPERAND (TREE_OPERAND (rhs_def_stmt, 1), 0);
1674 /* Verify that RHS_DEF_OPERAND is a suitable SSA variable. */
1675 if (TREE_CODE (rhs_def_operand) == SSA_NAME
1676 && ! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (rhs_def_operand))
1677 result = update_rhs_and_lookup_avail_expr (stmt,
1678 rhs_def_operand,
1679 insert);
1683 /* If we have z = (x OP C1), see if we earlier had x = y OP C2.
1684 If OP is associative, create and fold (y OP C2) OP C1 which
1685 should result in (y OP C3), use that as the RHS for the
1686 assignment. Add minus to this, as we handle it specially below. */
1687 if ((associative_tree_code (rhs_code) || rhs_code == MINUS_EXPR)
1688 && TREE_CODE (TREE_OPERAND (rhs, 0)) == SSA_NAME
1689 && is_gimple_min_invariant (TREE_OPERAND (rhs, 1)))
1691 tree rhs_def_stmt = SSA_NAME_DEF_STMT (TREE_OPERAND (rhs, 0));
1693 /* See if the RHS_DEF_STMT has the same form as our statement. */
1694 if (TREE_CODE (rhs_def_stmt) == MODIFY_EXPR)
1696 tree rhs_def_rhs = TREE_OPERAND (rhs_def_stmt, 1);
1697 enum tree_code rhs_def_code = TREE_CODE (rhs_def_rhs);
1699 if ((rhs_code == rhs_def_code && unsafe_associative_fp_binop (rhs))
1700 || (rhs_code == PLUS_EXPR && rhs_def_code == MINUS_EXPR)
1701 || (rhs_code == MINUS_EXPR && rhs_def_code == PLUS_EXPR))
1703 tree def_stmt_op0 = TREE_OPERAND (rhs_def_rhs, 0);
1704 tree def_stmt_op1 = TREE_OPERAND (rhs_def_rhs, 1);
1706 if (TREE_CODE (def_stmt_op0) == SSA_NAME
1707 && ! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def_stmt_op0)
1708 && is_gimple_min_invariant (def_stmt_op1))
1710 tree outer_const = TREE_OPERAND (rhs, 1);
1711 tree type = TREE_TYPE (TREE_OPERAND (stmt, 0));
1712 tree t;
1714 /* If we care about correct floating point results, then
1715 don't fold x + c1 - c2. Note that we need to take both
1716 the codes and the signs to figure this out. */
1717 if (FLOAT_TYPE_P (type)
1718 && !flag_unsafe_math_optimizations
1719 && (rhs_def_code == PLUS_EXPR
1720 || rhs_def_code == MINUS_EXPR))
1722 bool neg = false;
1724 neg ^= (rhs_code == MINUS_EXPR);
1725 neg ^= (rhs_def_code == MINUS_EXPR);
1726 neg ^= real_isneg (TREE_REAL_CST_PTR (outer_const));
1727 neg ^= real_isneg (TREE_REAL_CST_PTR (def_stmt_op1));
1729 if (neg)
1730 goto dont_fold_assoc;
1733 /* Ho hum. So fold will only operate on the outermost
1734 thingy that we give it, so we have to build the new
1735 expression in two pieces. This requires that we handle
1736 combinations of plus and minus. */
1737 if (rhs_def_code != rhs_code)
1739 if (rhs_def_code == MINUS_EXPR)
1740 t = build (MINUS_EXPR, type, outer_const, def_stmt_op1);
1741 else
1742 t = build (MINUS_EXPR, type, def_stmt_op1, outer_const);
1743 rhs_code = PLUS_EXPR;
1745 else if (rhs_def_code == MINUS_EXPR)
1746 t = build (PLUS_EXPR, type, def_stmt_op1, outer_const);
1747 else
1748 t = build (rhs_def_code, type, def_stmt_op1, outer_const);
1749 t = local_fold (t);
1750 t = build (rhs_code, type, def_stmt_op0, t);
1751 t = local_fold (t);
1753 /* If the result is a suitable looking gimple expression,
1754 then use it instead of the original for STMT. */
1755 if (TREE_CODE (t) == SSA_NAME
1756 || (UNARY_CLASS_P (t)
1757 && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME)
1758 || ((BINARY_CLASS_P (t) || COMPARISON_CLASS_P (t))
1759 && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME
1760 && is_gimple_val (TREE_OPERAND (t, 1))))
1761 result = update_rhs_and_lookup_avail_expr (stmt, t, insert);
1765 dont_fold_assoc:;
1768 /* Transform TRUNC_DIV_EXPR and TRUNC_MOD_EXPR into RSHIFT_EXPR
1769 and BIT_AND_EXPR respectively if the first operand is greater
1770 than zero and the second operand is an exact power of two. */
1771 if ((rhs_code == TRUNC_DIV_EXPR || rhs_code == TRUNC_MOD_EXPR)
1772 && INTEGRAL_TYPE_P (TREE_TYPE (TREE_OPERAND (rhs, 0)))
1773 && integer_pow2p (TREE_OPERAND (rhs, 1)))
1775 tree val;
1776 tree op = TREE_OPERAND (rhs, 0);
1778 if (TYPE_UNSIGNED (TREE_TYPE (op)))
1780 val = integer_one_node;
1782 else
1784 tree dummy_cond = walk_data->global_data;
1786 if (! dummy_cond)
1788 dummy_cond = build (GT_EXPR, boolean_type_node,
1789 op, integer_zero_node);
1790 dummy_cond = build (COND_EXPR, void_type_node,
1791 dummy_cond, NULL, NULL);
1792 walk_data->global_data = dummy_cond;
1794 else
1796 TREE_SET_CODE (COND_EXPR_COND (dummy_cond), GT_EXPR);
1797 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 0) = op;
1798 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 1)
1799 = integer_zero_node;
1801 val = simplify_cond_and_lookup_avail_expr (dummy_cond, NULL, false);
1804 if (val && integer_onep (val))
1806 tree t;
1807 tree op0 = TREE_OPERAND (rhs, 0);
1808 tree op1 = TREE_OPERAND (rhs, 1);
1810 if (rhs_code == TRUNC_DIV_EXPR)
1811 t = build (RSHIFT_EXPR, TREE_TYPE (op0), op0,
1812 build_int_cst (NULL_TREE, tree_log2 (op1)));
1813 else
1814 t = build (BIT_AND_EXPR, TREE_TYPE (op0), op0,
1815 local_fold (build (MINUS_EXPR, TREE_TYPE (op1),
1816 op1, integer_one_node)));
1818 result = update_rhs_and_lookup_avail_expr (stmt, t, insert);
1822 /* Transform ABS (X) into X or -X as appropriate. */
1823 if (rhs_code == ABS_EXPR
1824 && INTEGRAL_TYPE_P (TREE_TYPE (TREE_OPERAND (rhs, 0))))
1826 tree val;
1827 tree op = TREE_OPERAND (rhs, 0);
1828 tree type = TREE_TYPE (op);
1830 if (TYPE_UNSIGNED (type))
1832 val = integer_zero_node;
1834 else
1836 tree dummy_cond = walk_data->global_data;
1838 if (! dummy_cond)
1840 dummy_cond = build (LE_EXPR, boolean_type_node,
1841 op, integer_zero_node);
1842 dummy_cond = build (COND_EXPR, void_type_node,
1843 dummy_cond, NULL, NULL);
1844 walk_data->global_data = dummy_cond;
1846 else
1848 TREE_SET_CODE (COND_EXPR_COND (dummy_cond), LE_EXPR);
1849 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 0) = op;
1850 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 1)
1851 = build_int_cst (type, 0);
1853 val = simplify_cond_and_lookup_avail_expr (dummy_cond, NULL, false);
1855 if (!val)
1857 TREE_SET_CODE (COND_EXPR_COND (dummy_cond), GE_EXPR);
1858 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 0) = op;
1859 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 1)
1860 = build_int_cst (type, 0);
1862 val = simplify_cond_and_lookup_avail_expr (dummy_cond,
1863 NULL, false);
1865 if (val)
1867 if (integer_zerop (val))
1868 val = integer_one_node;
1869 else if (integer_onep (val))
1870 val = integer_zero_node;
1875 if (val
1876 && (integer_onep (val) || integer_zerop (val)))
1878 tree t;
1880 if (integer_onep (val))
1881 t = build1 (NEGATE_EXPR, TREE_TYPE (op), op);
1882 else
1883 t = op;
1885 result = update_rhs_and_lookup_avail_expr (stmt, t, insert);
1889 /* Optimize *"foo" into 'f'. This is done here rather than
1890 in fold to avoid problems with stuff like &*"foo". */
1891 if (TREE_CODE (rhs) == INDIRECT_REF || TREE_CODE (rhs) == ARRAY_REF)
1893 tree t = fold_read_from_constant_string (rhs);
1895 if (t)
1896 result = update_rhs_and_lookup_avail_expr (stmt, t, insert);
1899 return result;
1902 /* COND is a condition of the form:
1904 x == const or x != const
1906 Look back to x's defining statement and see if x is defined as
1908 x = (type) y;
1910 If const is unchanged if we convert it to type, then we can build
1911 the equivalent expression:
1914 y == const or y != const
1916 Which may allow further optimizations.
1918 Return the equivalent comparison or NULL if no such equivalent comparison
1919 was found. */
1921 static tree
1922 find_equivalent_equality_comparison (tree cond)
1924 tree op0 = TREE_OPERAND (cond, 0);
1925 tree op1 = TREE_OPERAND (cond, 1);
1926 tree def_stmt = SSA_NAME_DEF_STMT (op0);
1928 /* OP0 might have been a parameter, so first make sure it
1929 was defined by a MODIFY_EXPR. */
1930 if (def_stmt && TREE_CODE (def_stmt) == MODIFY_EXPR)
1932 tree def_rhs = TREE_OPERAND (def_stmt, 1);
1934 /* Now make sure the RHS of the MODIFY_EXPR is a typecast. */
1935 if ((TREE_CODE (def_rhs) == NOP_EXPR
1936 || TREE_CODE (def_rhs) == CONVERT_EXPR)
1937 && TREE_CODE (TREE_OPERAND (def_rhs, 0)) == SSA_NAME)
1939 tree def_rhs_inner = TREE_OPERAND (def_rhs, 0);
1940 tree def_rhs_inner_type = TREE_TYPE (def_rhs_inner);
1941 tree new;
1943 if (TYPE_PRECISION (def_rhs_inner_type)
1944 > TYPE_PRECISION (TREE_TYPE (def_rhs)))
1945 return NULL;
1947 /* What we want to prove is that if we convert OP1 to
1948 the type of the object inside the NOP_EXPR that the
1949 result is still equivalent to SRC.
1951 If that is true, the build and return new equivalent
1952 condition which uses the source of the typecast and the
1953 new constant (which has only changed its type). */
1954 new = build1 (TREE_CODE (def_rhs), def_rhs_inner_type, op1);
1955 new = local_fold (new);
1956 if (is_gimple_val (new) && tree_int_cst_equal (new, op1))
1957 return build (TREE_CODE (cond), TREE_TYPE (cond),
1958 def_rhs_inner, new);
1961 return NULL;
1964 /* STMT is a COND_EXPR for which we could not trivially determine its
1965 result. This routine attempts to find equivalent forms of the
1966 condition which we may be able to optimize better. It also
1967 uses simple value range propagation to optimize conditionals. */
1969 static tree
1970 simplify_cond_and_lookup_avail_expr (tree stmt,
1971 stmt_ann_t ann,
1972 int insert)
1974 tree cond = COND_EXPR_COND (stmt);
1976 if (COMPARISON_CLASS_P (cond))
1978 tree op0 = TREE_OPERAND (cond, 0);
1979 tree op1 = TREE_OPERAND (cond, 1);
1981 if (TREE_CODE (op0) == SSA_NAME && is_gimple_min_invariant (op1))
1983 int limit;
1984 tree low, high, cond_low, cond_high;
1985 int lowequal, highequal, swapped, no_overlap, subset, cond_inverted;
1986 varray_type vrp_records;
1987 struct vrp_element *element;
1988 struct vrp_hash_elt vrp_hash_elt, *vrp_hash_elt_p;
1989 void **slot;
1991 /* First see if we have test of an SSA_NAME against a constant
1992 where the SSA_NAME is defined by an earlier typecast which
1993 is irrelevant when performing tests against the given
1994 constant. */
1995 if (TREE_CODE (cond) == EQ_EXPR || TREE_CODE (cond) == NE_EXPR)
1997 tree new_cond = find_equivalent_equality_comparison (cond);
1999 if (new_cond)
2001 /* Update the statement to use the new equivalent
2002 condition. */
2003 COND_EXPR_COND (stmt) = new_cond;
2005 /* If this is not a real stmt, ann will be NULL and we
2006 avoid processing the operands. */
2007 if (ann)
2008 modify_stmt (stmt);
2010 /* Lookup the condition and return its known value if it
2011 exists. */
2012 new_cond = lookup_avail_expr (stmt, insert);
2013 if (new_cond)
2014 return new_cond;
2016 /* The operands have changed, so update op0 and op1. */
2017 op0 = TREE_OPERAND (cond, 0);
2018 op1 = TREE_OPERAND (cond, 1);
2022 /* Consult the value range records for this variable (if they exist)
2023 to see if we can eliminate or simplify this conditional.
2025 Note two tests are necessary to determine no records exist.
2026 First we have to see if the virtual array exists, if it
2027 exists, then we have to check its active size.
2029 Also note the vast majority of conditionals are not testing
2030 a variable which has had its range constrained by an earlier
2031 conditional. So this filter avoids a lot of unnecessary work. */
2032 vrp_hash_elt.var = op0;
2033 vrp_hash_elt.records = NULL;
2034 slot = htab_find_slot (vrp_data, &vrp_hash_elt, NO_INSERT);
2035 if (slot == NULL)
2036 return NULL;
2038 vrp_hash_elt_p = (struct vrp_hash_elt *) *slot;
2039 vrp_records = vrp_hash_elt_p->records;
2040 if (vrp_records == NULL)
2041 return NULL;
2043 limit = VARRAY_ACTIVE_SIZE (vrp_records);
2045 /* If we have no value range records for this variable, or we are
2046 unable to extract a range for this condition, then there is
2047 nothing to do. */
2048 if (limit == 0
2049 || ! extract_range_from_cond (cond, &cond_high,
2050 &cond_low, &cond_inverted))
2051 return NULL;
2053 /* We really want to avoid unnecessary computations of range
2054 info. So all ranges are computed lazily; this avoids a
2055 lot of unnecessary work. i.e., we record the conditional,
2056 but do not process how it constrains the variable's
2057 potential values until we know that processing the condition
2058 could be helpful.
2060 However, we do not want to have to walk a potentially long
2061 list of ranges, nor do we want to compute a variable's
2062 range more than once for a given path.
2064 Luckily, each time we encounter a conditional that can not
2065 be otherwise optimized we will end up here and we will
2066 compute the necessary range information for the variable
2067 used in this condition.
2069 Thus you can conclude that there will never be more than one
2070 conditional associated with a variable which has not been
2071 processed. So we never need to merge more than one new
2072 conditional into the current range.
2074 These properties also help us avoid unnecessary work. */
2075 element
2076 = (struct vrp_element *)VARRAY_GENERIC_PTR (vrp_records, limit - 1);
2078 if (element->high && element->low)
2080 /* The last element has been processed, so there is no range
2081 merging to do, we can simply use the high/low values
2082 recorded in the last element. */
2083 low = element->low;
2084 high = element->high;
2086 else
2088 tree tmp_high, tmp_low;
2089 int dummy;
2091 /* The last element has not been processed. Process it now. */
2092 extract_range_from_cond (element->cond, &tmp_high,
2093 &tmp_low, &dummy);
2095 /* If this is the only element, then no merging is necessary,
2096 the high/low values from extract_range_from_cond are all
2097 we need. */
2098 if (limit == 1)
2100 low = tmp_low;
2101 high = tmp_high;
2103 else
2105 /* Get the high/low value from the previous element. */
2106 struct vrp_element *prev
2107 = (struct vrp_element *)VARRAY_GENERIC_PTR (vrp_records,
2108 limit - 2);
2109 low = prev->low;
2110 high = prev->high;
2112 /* Merge in this element's range with the range from the
2113 previous element.
2115 The low value for the merged range is the maximum of
2116 the previous low value and the low value of this record.
2118 Similarly the high value for the merged range is the
2119 minimum of the previous high value and the high value of
2120 this record. */
2121 low = (tree_int_cst_compare (low, tmp_low) == 1
2122 ? low : tmp_low);
2123 high = (tree_int_cst_compare (high, tmp_high) == -1
2124 ? high : tmp_high);
2127 /* And record the computed range. */
2128 element->low = low;
2129 element->high = high;
2133 /* After we have constrained this variable's potential values,
2134 we try to determine the result of the given conditional.
2136 To simplify later tests, first determine if the current
2137 low value is the same low value as the conditional.
2138 Similarly for the current high value and the high value
2139 for the conditional. */
2140 lowequal = tree_int_cst_equal (low, cond_low);
2141 highequal = tree_int_cst_equal (high, cond_high);
2143 if (lowequal && highequal)
2144 return (cond_inverted ? boolean_false_node : boolean_true_node);
2146 /* To simplify the overlap/subset tests below we may want
2147 to swap the two ranges so that the larger of the two
2148 ranges occurs "first". */
2149 swapped = 0;
2150 if (tree_int_cst_compare (low, cond_low) == 1
2151 || (lowequal
2152 && tree_int_cst_compare (cond_high, high) == 1))
2154 tree temp;
2156 swapped = 1;
2157 temp = low;
2158 low = cond_low;
2159 cond_low = temp;
2160 temp = high;
2161 high = cond_high;
2162 cond_high = temp;
2165 /* Now determine if there is no overlap in the ranges
2166 or if the second range is a subset of the first range. */
2167 no_overlap = tree_int_cst_lt (high, cond_low);
2168 subset = tree_int_cst_compare (cond_high, high) != 1;
2170 /* If there was no overlap in the ranges, then this conditional
2171 always has a false value (unless we had to invert this
2172 conditional, in which case it always has a true value). */
2173 if (no_overlap)
2174 return (cond_inverted ? boolean_true_node : boolean_false_node);
2176 /* If the current range is a subset of the condition's range,
2177 then this conditional always has a true value (unless we
2178 had to invert this conditional, in which case it always
2179 has a true value). */
2180 if (subset && swapped)
2181 return (cond_inverted ? boolean_false_node : boolean_true_node);
2183 /* We were unable to determine the result of the conditional.
2184 However, we may be able to simplify the conditional. First
2185 merge the ranges in the same manner as range merging above. */
2186 low = tree_int_cst_compare (low, cond_low) == 1 ? low : cond_low;
2187 high = tree_int_cst_compare (high, cond_high) == -1 ? high : cond_high;
2189 /* If the range has converged to a single point, then turn this
2190 into an equality comparison. */
2191 if (TREE_CODE (cond) != EQ_EXPR
2192 && TREE_CODE (cond) != NE_EXPR
2193 && tree_int_cst_equal (low, high))
2195 TREE_SET_CODE (cond, EQ_EXPR);
2196 TREE_OPERAND (cond, 1) = high;
2200 return 0;
2203 /* STMT is a SWITCH_EXPR for which we could not trivially determine its
2204 result. This routine attempts to find equivalent forms of the
2205 condition which we may be able to optimize better. */
2207 static tree
2208 simplify_switch_and_lookup_avail_expr (tree stmt, int insert)
2210 tree cond = SWITCH_COND (stmt);
2211 tree def, to, ti;
2213 /* The optimization that we really care about is removing unnecessary
2214 casts. That will let us do much better in propagating the inferred
2215 constant at the switch target. */
2216 if (TREE_CODE (cond) == SSA_NAME)
2218 def = SSA_NAME_DEF_STMT (cond);
2219 if (TREE_CODE (def) == MODIFY_EXPR)
2221 def = TREE_OPERAND (def, 1);
2222 if (TREE_CODE (def) == NOP_EXPR)
2224 int need_precision;
2225 bool fail;
2227 def = TREE_OPERAND (def, 0);
2229 #ifdef ENABLE_CHECKING
2230 /* ??? Why was Jeff testing this? We are gimple... */
2231 gcc_assert (is_gimple_val (def));
2232 #endif
2234 to = TREE_TYPE (cond);
2235 ti = TREE_TYPE (def);
2237 /* If we have an extension that preserves value, then we
2238 can copy the source value into the switch. */
2240 need_precision = TYPE_PRECISION (ti);
2241 fail = false;
2242 if (TYPE_UNSIGNED (to) && !TYPE_UNSIGNED (ti))
2243 fail = true;
2244 else if (!TYPE_UNSIGNED (to) && TYPE_UNSIGNED (ti))
2245 need_precision += 1;
2246 if (TYPE_PRECISION (to) < need_precision)
2247 fail = true;
2249 if (!fail)
2251 SWITCH_COND (stmt) = def;
2252 modify_stmt (stmt);
2254 return lookup_avail_expr (stmt, insert);
2260 return 0;
2264 /* CONST_AND_COPIES is a table which maps an SSA_NAME to the current
2265 known value for that SSA_NAME (or NULL if no value is known).
2267 NONZERO_VARS is the set SSA_NAMES known to have a nonzero value,
2268 even if we don't know their precise value.
2270 Propagate values from CONST_AND_COPIES and NONZERO_VARS into the PHI
2271 nodes of the successors of BB. */
2273 static void
2274 cprop_into_successor_phis (basic_block bb, bitmap nonzero_vars)
2276 edge e;
2277 edge_iterator ei;
2279 /* This can get rather expensive if the implementation is naive in
2280 how it finds the phi alternative associated with a particular edge. */
2281 FOR_EACH_EDGE (e, ei, bb->succs)
2283 tree phi;
2284 int indx;
2286 /* If this is an abnormal edge, then we do not want to copy propagate
2287 into the PHI alternative associated with this edge. */
2288 if (e->flags & EDGE_ABNORMAL)
2289 continue;
2291 phi = phi_nodes (e->dest);
2292 if (! phi)
2293 continue;
2295 indx = e->dest_idx;
2296 for ( ; phi; phi = PHI_CHAIN (phi))
2298 tree new;
2299 use_operand_p orig_p;
2300 tree orig;
2302 /* The alternative may be associated with a constant, so verify
2303 it is an SSA_NAME before doing anything with it. */
2304 orig_p = PHI_ARG_DEF_PTR (phi, indx);
2305 orig = USE_FROM_PTR (orig_p);
2306 if (TREE_CODE (orig) != SSA_NAME)
2307 continue;
2309 /* If the alternative is known to have a nonzero value, record
2310 that fact in the PHI node itself for future use. */
2311 if (bitmap_bit_p (nonzero_vars, SSA_NAME_VERSION (orig)))
2312 PHI_ARG_NONZERO (phi, indx) = true;
2314 /* If we have *ORIG_P in our constant/copy table, then replace
2315 ORIG_P with its value in our constant/copy table. */
2316 new = SSA_NAME_VALUE (orig);
2317 if (new
2318 && (TREE_CODE (new) == SSA_NAME
2319 || is_gimple_min_invariant (new))
2320 && may_propagate_copy (orig, new))
2322 propagate_value (orig_p, new);
2328 /* We have finished optimizing BB, record any information implied by
2329 taking a specific outgoing edge from BB. */
2331 static void
2332 record_edge_info (basic_block bb)
2334 block_stmt_iterator bsi = bsi_last (bb);
2335 struct edge_info *edge_info;
2337 if (! bsi_end_p (bsi))
2339 tree stmt = bsi_stmt (bsi);
2341 if (stmt && TREE_CODE (stmt) == SWITCH_EXPR)
2343 tree cond = SWITCH_COND (stmt);
2345 if (TREE_CODE (cond) == SSA_NAME)
2347 tree labels = SWITCH_LABELS (stmt);
2348 int i, n_labels = TREE_VEC_LENGTH (labels);
2349 tree *info = xcalloc (n_basic_blocks, sizeof (tree));
2350 edge e;
2351 edge_iterator ei;
2353 for (i = 0; i < n_labels; i++)
2355 tree label = TREE_VEC_ELT (labels, i);
2356 basic_block target_bb = label_to_block (CASE_LABEL (label));
2358 if (CASE_HIGH (label)
2359 || !CASE_LOW (label)
2360 || info[target_bb->index])
2361 info[target_bb->index] = error_mark_node;
2362 else
2363 info[target_bb->index] = label;
2366 FOR_EACH_EDGE (e, ei, bb->succs)
2368 basic_block target_bb = e->dest;
2369 tree node = info[target_bb->index];
2371 if (node != NULL && node != error_mark_node)
2373 tree x = fold_convert (TREE_TYPE (cond), CASE_LOW (node));
2374 edge_info = allocate_edge_info (e);
2375 edge_info->lhs = cond;
2376 edge_info->rhs = x;
2379 free (info);
2383 /* A COND_EXPR may create equivalences too. */
2384 if (stmt && TREE_CODE (stmt) == COND_EXPR)
2386 tree cond = COND_EXPR_COND (stmt);
2387 edge true_edge;
2388 edge false_edge;
2390 extract_true_false_edges_from_block (bb, &true_edge, &false_edge);
2392 /* If the conditional is a single variable 'X', record 'X = 1'
2393 for the true edge and 'X = 0' on the false edge. */
2394 if (SSA_VAR_P (cond))
2396 struct edge_info *edge_info;
2398 edge_info = allocate_edge_info (true_edge);
2399 edge_info->lhs = cond;
2400 edge_info->rhs = constant_boolean_node (1, TREE_TYPE (cond));
2402 edge_info = allocate_edge_info (false_edge);
2403 edge_info->lhs = cond;
2404 edge_info->rhs = constant_boolean_node (0, TREE_TYPE (cond));
2406 /* Equality tests may create one or two equivalences. */
2407 else if (COMPARISON_CLASS_P (cond))
2409 tree op0 = TREE_OPERAND (cond, 0);
2410 tree op1 = TREE_OPERAND (cond, 1);
2412 /* Special case comparing booleans against a constant as we
2413 know the value of OP0 on both arms of the branch. i.e., we
2414 can record an equivalence for OP0 rather than COND. */
2415 if ((TREE_CODE (cond) == EQ_EXPR || TREE_CODE (cond) == NE_EXPR)
2416 && TREE_CODE (op0) == SSA_NAME
2417 && TREE_CODE (TREE_TYPE (op0)) == BOOLEAN_TYPE
2418 && is_gimple_min_invariant (op1))
2420 if (TREE_CODE (cond) == EQ_EXPR)
2422 edge_info = allocate_edge_info (true_edge);
2423 edge_info->lhs = op0;
2424 edge_info->rhs = (integer_zerop (op1)
2425 ? boolean_false_node
2426 : boolean_true_node);
2428 edge_info = allocate_edge_info (false_edge);
2429 edge_info->lhs = op0;
2430 edge_info->rhs = (integer_zerop (op1)
2431 ? boolean_true_node
2432 : boolean_false_node);
2434 else
2436 edge_info = allocate_edge_info (true_edge);
2437 edge_info->lhs = op0;
2438 edge_info->rhs = (integer_zerop (op1)
2439 ? boolean_true_node
2440 : boolean_false_node);
2442 edge_info = allocate_edge_info (false_edge);
2443 edge_info->lhs = op0;
2444 edge_info->rhs = (integer_zerop (op1)
2445 ? boolean_false_node
2446 : boolean_true_node);
2450 else if (is_gimple_min_invariant (op0)
2451 && (TREE_CODE (op1) == SSA_NAME
2452 || is_gimple_min_invariant (op1)))
2454 tree inverted = invert_truthvalue (cond);
2455 struct edge_info *edge_info;
2457 edge_info = allocate_edge_info (true_edge);
2458 record_conditions (edge_info, cond, inverted);
2460 if (TREE_CODE (cond) == EQ_EXPR)
2462 edge_info->lhs = op1;
2463 edge_info->rhs = op0;
2466 edge_info = allocate_edge_info (false_edge);
2467 record_conditions (edge_info, inverted, cond);
2469 if (TREE_CODE (cond) == NE_EXPR)
2471 edge_info->lhs = op1;
2472 edge_info->rhs = op0;
2476 else if (TREE_CODE (op0) == SSA_NAME
2477 && (is_gimple_min_invariant (op1)
2478 || TREE_CODE (op1) == SSA_NAME))
2480 tree inverted = invert_truthvalue (cond);
2481 struct edge_info *edge_info;
2483 edge_info = allocate_edge_info (true_edge);
2484 record_conditions (edge_info, cond, inverted);
2486 if (TREE_CODE (cond) == EQ_EXPR)
2488 edge_info->lhs = op0;
2489 edge_info->rhs = op1;
2492 edge_info = allocate_edge_info (false_edge);
2493 record_conditions (edge_info, inverted, cond);
2495 if (TREE_CODE (cond) == NE_EXPR)
2497 edge_info->lhs = op0;
2498 edge_info->rhs = op1;
2503 /* ??? TRUTH_NOT_EXPR can create an equivalence too. */
2508 /* Propagate information from BB to its outgoing edges.
2510 This can include equivalency information implied by control statements
2511 at the end of BB and const/copy propagation into PHIs in BB's
2512 successor blocks. */
2514 static void
2515 propagate_to_outgoing_edges (struct dom_walk_data *walk_data ATTRIBUTE_UNUSED,
2516 basic_block bb)
2519 record_edge_info (bb);
2520 cprop_into_successor_phis (bb, nonzero_vars);
2523 /* Search for redundant computations in STMT. If any are found, then
2524 replace them with the variable holding the result of the computation.
2526 If safe, record this expression into the available expression hash
2527 table. */
2529 static bool
2530 eliminate_redundant_computations (struct dom_walk_data *walk_data,
2531 tree stmt, stmt_ann_t ann)
2533 v_may_def_optype v_may_defs = V_MAY_DEF_OPS (ann);
2534 tree *expr_p, def = NULL_TREE;
2535 bool insert = true;
2536 tree cached_lhs;
2537 bool retval = false;
2539 if (TREE_CODE (stmt) == MODIFY_EXPR)
2540 def = TREE_OPERAND (stmt, 0);
2542 /* Certain expressions on the RHS can be optimized away, but can not
2543 themselves be entered into the hash tables. */
2544 if (ann->makes_aliased_stores
2545 || ! def
2546 || TREE_CODE (def) != SSA_NAME
2547 || SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def)
2548 || NUM_V_MAY_DEFS (v_may_defs) != 0)
2549 insert = false;
2551 /* Check if the expression has been computed before. */
2552 cached_lhs = lookup_avail_expr (stmt, insert);
2554 /* If this is an assignment and the RHS was not in the hash table,
2555 then try to simplify the RHS and lookup the new RHS in the
2556 hash table. */
2557 if (! cached_lhs && TREE_CODE (stmt) == MODIFY_EXPR)
2558 cached_lhs = simplify_rhs_and_lookup_avail_expr (walk_data, stmt, insert);
2559 /* Similarly if this is a COND_EXPR and we did not find its
2560 expression in the hash table, simplify the condition and
2561 try again. */
2562 else if (! cached_lhs && TREE_CODE (stmt) == COND_EXPR)
2563 cached_lhs = simplify_cond_and_lookup_avail_expr (stmt, ann, insert);
2564 /* Similarly for a SWITCH_EXPR. */
2565 else if (!cached_lhs && TREE_CODE (stmt) == SWITCH_EXPR)
2566 cached_lhs = simplify_switch_and_lookup_avail_expr (stmt, insert);
2568 opt_stats.num_exprs_considered++;
2570 /* Get a pointer to the expression we are trying to optimize. */
2571 if (TREE_CODE (stmt) == COND_EXPR)
2572 expr_p = &COND_EXPR_COND (stmt);
2573 else if (TREE_CODE (stmt) == SWITCH_EXPR)
2574 expr_p = &SWITCH_COND (stmt);
2575 else if (TREE_CODE (stmt) == RETURN_EXPR && TREE_OPERAND (stmt, 0))
2576 expr_p = &TREE_OPERAND (TREE_OPERAND (stmt, 0), 1);
2577 else
2578 expr_p = &TREE_OPERAND (stmt, 1);
2580 /* It is safe to ignore types here since we have already done
2581 type checking in the hashing and equality routines. In fact
2582 type checking here merely gets in the way of constant
2583 propagation. Also, make sure that it is safe to propagate
2584 CACHED_LHS into *EXPR_P. */
2585 if (cached_lhs
2586 && (TREE_CODE (cached_lhs) != SSA_NAME
2587 || may_propagate_copy (*expr_p, cached_lhs)))
2589 if (dump_file && (dump_flags & TDF_DETAILS))
2591 fprintf (dump_file, " Replaced redundant expr '");
2592 print_generic_expr (dump_file, *expr_p, dump_flags);
2593 fprintf (dump_file, "' with '");
2594 print_generic_expr (dump_file, cached_lhs, dump_flags);
2595 fprintf (dump_file, "'\n");
2598 opt_stats.num_re++;
2600 #if defined ENABLE_CHECKING
2601 gcc_assert (TREE_CODE (cached_lhs) == SSA_NAME
2602 || is_gimple_min_invariant (cached_lhs));
2603 #endif
2605 if (TREE_CODE (cached_lhs) == ADDR_EXPR
2606 || (POINTER_TYPE_P (TREE_TYPE (*expr_p))
2607 && is_gimple_min_invariant (cached_lhs)))
2608 retval = true;
2610 propagate_tree_value (expr_p, cached_lhs);
2611 modify_stmt (stmt);
2613 return retval;
2616 /* STMT, a MODIFY_EXPR, may create certain equivalences, in either
2617 the available expressions table or the const_and_copies table.
2618 Detect and record those equivalences. */
2620 static void
2621 record_equivalences_from_stmt (tree stmt,
2622 int may_optimize_p,
2623 stmt_ann_t ann)
2625 tree lhs = TREE_OPERAND (stmt, 0);
2626 enum tree_code lhs_code = TREE_CODE (lhs);
2627 int i;
2629 if (lhs_code == SSA_NAME)
2631 tree rhs = TREE_OPERAND (stmt, 1);
2633 /* Strip away any useless type conversions. */
2634 STRIP_USELESS_TYPE_CONVERSION (rhs);
2636 /* If the RHS of the assignment is a constant or another variable that
2637 may be propagated, register it in the CONST_AND_COPIES table. We
2638 do not need to record unwind data for this, since this is a true
2639 assignment and not an equivalence inferred from a comparison. All
2640 uses of this ssa name are dominated by this assignment, so unwinding
2641 just costs time and space. */
2642 if (may_optimize_p
2643 && (TREE_CODE (rhs) == SSA_NAME
2644 || is_gimple_min_invariant (rhs)))
2645 SSA_NAME_VALUE (lhs) = rhs;
2647 /* alloca never returns zero and the address of a non-weak symbol
2648 is never zero. NOP_EXPRs and CONVERT_EXPRs can be completely
2649 stripped as they do not affect this equivalence. */
2650 while (TREE_CODE (rhs) == NOP_EXPR
2651 || TREE_CODE (rhs) == CONVERT_EXPR)
2652 rhs = TREE_OPERAND (rhs, 0);
2654 if (alloca_call_p (rhs)
2655 || (TREE_CODE (rhs) == ADDR_EXPR
2656 && DECL_P (TREE_OPERAND (rhs, 0))
2657 && ! DECL_WEAK (TREE_OPERAND (rhs, 0))))
2658 record_var_is_nonzero (lhs);
2660 /* IOR of any value with a nonzero value will result in a nonzero
2661 value. Even if we do not know the exact result recording that
2662 the result is nonzero is worth the effort. */
2663 if (TREE_CODE (rhs) == BIT_IOR_EXPR
2664 && integer_nonzerop (TREE_OPERAND (rhs, 1)))
2665 record_var_is_nonzero (lhs);
2668 /* Look at both sides for pointer dereferences. If we find one, then
2669 the pointer must be nonnull and we can enter that equivalence into
2670 the hash tables. */
2671 if (flag_delete_null_pointer_checks)
2672 for (i = 0; i < 2; i++)
2674 tree t = TREE_OPERAND (stmt, i);
2676 /* Strip away any COMPONENT_REFs. */
2677 while (TREE_CODE (t) == COMPONENT_REF)
2678 t = TREE_OPERAND (t, 0);
2680 /* Now see if this is a pointer dereference. */
2681 if (INDIRECT_REF_P (t))
2683 tree op = TREE_OPERAND (t, 0);
2685 /* If the pointer is a SSA variable, then enter new
2686 equivalences into the hash table. */
2687 while (TREE_CODE (op) == SSA_NAME)
2689 tree def = SSA_NAME_DEF_STMT (op);
2691 record_var_is_nonzero (op);
2693 /* And walk up the USE-DEF chains noting other SSA_NAMEs
2694 which are known to have a nonzero value. */
2695 if (def
2696 && TREE_CODE (def) == MODIFY_EXPR
2697 && TREE_CODE (TREE_OPERAND (def, 1)) == NOP_EXPR)
2698 op = TREE_OPERAND (TREE_OPERAND (def, 1), 0);
2699 else
2700 break;
2705 /* A memory store, even an aliased store, creates a useful
2706 equivalence. By exchanging the LHS and RHS, creating suitable
2707 vops and recording the result in the available expression table,
2708 we may be able to expose more redundant loads. */
2709 if (!ann->has_volatile_ops
2710 && (TREE_CODE (TREE_OPERAND (stmt, 1)) == SSA_NAME
2711 || is_gimple_min_invariant (TREE_OPERAND (stmt, 1)))
2712 && !is_gimple_reg (lhs))
2714 tree rhs = TREE_OPERAND (stmt, 1);
2715 tree new;
2717 /* FIXME: If the LHS of the assignment is a bitfield and the RHS
2718 is a constant, we need to adjust the constant to fit into the
2719 type of the LHS. If the LHS is a bitfield and the RHS is not
2720 a constant, then we can not record any equivalences for this
2721 statement since we would need to represent the widening or
2722 narrowing of RHS. This fixes gcc.c-torture/execute/921016-1.c
2723 and should not be necessary if GCC represented bitfields
2724 properly. */
2725 if (lhs_code == COMPONENT_REF
2726 && DECL_BIT_FIELD (TREE_OPERAND (lhs, 1)))
2728 if (TREE_CONSTANT (rhs))
2729 rhs = widen_bitfield (rhs, TREE_OPERAND (lhs, 1), lhs);
2730 else
2731 rhs = NULL;
2733 /* If the value overflowed, then we can not use this equivalence. */
2734 if (rhs && ! is_gimple_min_invariant (rhs))
2735 rhs = NULL;
2738 if (rhs)
2740 /* Build a new statement with the RHS and LHS exchanged. */
2741 new = build (MODIFY_EXPR, TREE_TYPE (stmt), rhs, lhs);
2743 create_ssa_artficial_load_stmt (&(ann->operands), new);
2745 /* Finally enter the statement into the available expression
2746 table. */
2747 lookup_avail_expr (new, true);
2752 /* Replace *OP_P in STMT with any known equivalent value for *OP_P from
2753 CONST_AND_COPIES. */
2755 static bool
2756 cprop_operand (tree stmt, use_operand_p op_p)
2758 bool may_have_exposed_new_symbols = false;
2759 tree val;
2760 tree op = USE_FROM_PTR (op_p);
2762 /* If the operand has a known constant value or it is known to be a
2763 copy of some other variable, use the value or copy stored in
2764 CONST_AND_COPIES. */
2765 val = SSA_NAME_VALUE (op);
2766 if (val && TREE_CODE (val) != VALUE_HANDLE)
2768 tree op_type, val_type;
2770 /* Do not change the base variable in the virtual operand
2771 tables. That would make it impossible to reconstruct
2772 the renamed virtual operand if we later modify this
2773 statement. Also only allow the new value to be an SSA_NAME
2774 for propagation into virtual operands. */
2775 if (!is_gimple_reg (op)
2776 && (get_virtual_var (val) != get_virtual_var (op)
2777 || TREE_CODE (val) != SSA_NAME))
2778 return false;
2780 /* Do not replace hard register operands in asm statements. */
2781 if (TREE_CODE (stmt) == ASM_EXPR
2782 && !may_propagate_copy_into_asm (op))
2783 return false;
2785 /* Get the toplevel type of each operand. */
2786 op_type = TREE_TYPE (op);
2787 val_type = TREE_TYPE (val);
2789 /* While both types are pointers, get the type of the object
2790 pointed to. */
2791 while (POINTER_TYPE_P (op_type) && POINTER_TYPE_P (val_type))
2793 op_type = TREE_TYPE (op_type);
2794 val_type = TREE_TYPE (val_type);
2797 /* Make sure underlying types match before propagating a constant by
2798 converting the constant to the proper type. Note that convert may
2799 return a non-gimple expression, in which case we ignore this
2800 propagation opportunity. */
2801 if (TREE_CODE (val) != SSA_NAME)
2803 if (!lang_hooks.types_compatible_p (op_type, val_type))
2805 val = fold_convert (TREE_TYPE (op), val);
2806 if (!is_gimple_min_invariant (val))
2807 return false;
2811 /* Certain operands are not allowed to be copy propagated due
2812 to their interaction with exception handling and some GCC
2813 extensions. */
2814 else if (!may_propagate_copy (op, val))
2815 return false;
2817 /* Dump details. */
2818 if (dump_file && (dump_flags & TDF_DETAILS))
2820 fprintf (dump_file, " Replaced '");
2821 print_generic_expr (dump_file, op, dump_flags);
2822 fprintf (dump_file, "' with %s '",
2823 (TREE_CODE (val) != SSA_NAME ? "constant" : "variable"));
2824 print_generic_expr (dump_file, val, dump_flags);
2825 fprintf (dump_file, "'\n");
2828 /* If VAL is an ADDR_EXPR or a constant of pointer type, note
2829 that we may have exposed a new symbol for SSA renaming. */
2830 if (TREE_CODE (val) == ADDR_EXPR
2831 || (POINTER_TYPE_P (TREE_TYPE (op))
2832 && is_gimple_min_invariant (val)))
2833 may_have_exposed_new_symbols = true;
2835 propagate_value (op_p, val);
2837 /* And note that we modified this statement. This is now
2838 safe, even if we changed virtual operands since we will
2839 rescan the statement and rewrite its operands again. */
2840 modify_stmt (stmt);
2842 return may_have_exposed_new_symbols;
2845 /* CONST_AND_COPIES is a table which maps an SSA_NAME to the current
2846 known value for that SSA_NAME (or NULL if no value is known).
2848 Propagate values from CONST_AND_COPIES into the uses, vuses and
2849 v_may_def_ops of STMT. */
2851 static bool
2852 cprop_into_stmt (tree stmt)
2854 bool may_have_exposed_new_symbols = false;
2855 use_operand_p op_p;
2856 ssa_op_iter iter;
2857 tree rhs;
2859 FOR_EACH_SSA_USE_OPERAND (op_p, stmt, iter, SSA_OP_ALL_USES)
2861 if (TREE_CODE (USE_FROM_PTR (op_p)) == SSA_NAME)
2862 may_have_exposed_new_symbols |= cprop_operand (stmt, op_p);
2865 if (may_have_exposed_new_symbols)
2867 rhs = get_rhs (stmt);
2868 if (rhs && TREE_CODE (rhs) == ADDR_EXPR)
2869 recompute_tree_invarant_for_addr_expr (rhs);
2872 return may_have_exposed_new_symbols;
2876 /* Optimize the statement pointed by iterator SI.
2878 We try to perform some simplistic global redundancy elimination and
2879 constant propagation:
2881 1- To detect global redundancy, we keep track of expressions that have
2882 been computed in this block and its dominators. If we find that the
2883 same expression is computed more than once, we eliminate repeated
2884 computations by using the target of the first one.
2886 2- Constant values and copy assignments. This is used to do very
2887 simplistic constant and copy propagation. When a constant or copy
2888 assignment is found, we map the value on the RHS of the assignment to
2889 the variable in the LHS in the CONST_AND_COPIES table. */
2891 static void
2892 optimize_stmt (struct dom_walk_data *walk_data, basic_block bb,
2893 block_stmt_iterator si)
2895 stmt_ann_t ann;
2896 tree stmt;
2897 bool may_optimize_p;
2898 bool may_have_exposed_new_symbols = false;
2900 stmt = bsi_stmt (si);
2902 get_stmt_operands (stmt);
2903 ann = stmt_ann (stmt);
2904 opt_stats.num_stmts++;
2905 may_have_exposed_new_symbols = false;
2907 if (dump_file && (dump_flags & TDF_DETAILS))
2909 fprintf (dump_file, "Optimizing statement ");
2910 print_generic_stmt (dump_file, stmt, TDF_SLIM);
2913 /* Const/copy propagate into USES, VUSES and the RHS of V_MAY_DEFs. */
2914 may_have_exposed_new_symbols = cprop_into_stmt (stmt);
2916 /* If the statement has been modified with constant replacements,
2917 fold its RHS before checking for redundant computations. */
2918 if (ann->modified)
2920 /* Try to fold the statement making sure that STMT is kept
2921 up to date. */
2922 if (fold_stmt (bsi_stmt_ptr (si)))
2924 stmt = bsi_stmt (si);
2925 ann = stmt_ann (stmt);
2927 if (dump_file && (dump_flags & TDF_DETAILS))
2929 fprintf (dump_file, " Folded to: ");
2930 print_generic_stmt (dump_file, stmt, TDF_SLIM);
2934 /* Constant/copy propagation above may change the set of
2935 virtual operands associated with this statement. Folding
2936 may remove the need for some virtual operands.
2938 Indicate we will need to rescan and rewrite the statement. */
2939 may_have_exposed_new_symbols = true;
2942 /* Check for redundant computations. Do this optimization only
2943 for assignments that have no volatile ops and conditionals. */
2944 may_optimize_p = (!ann->has_volatile_ops
2945 && ((TREE_CODE (stmt) == RETURN_EXPR
2946 && TREE_OPERAND (stmt, 0)
2947 && TREE_CODE (TREE_OPERAND (stmt, 0)) == MODIFY_EXPR
2948 && ! (TREE_SIDE_EFFECTS
2949 (TREE_OPERAND (TREE_OPERAND (stmt, 0), 1))))
2950 || (TREE_CODE (stmt) == MODIFY_EXPR
2951 && ! TREE_SIDE_EFFECTS (TREE_OPERAND (stmt, 1)))
2952 || TREE_CODE (stmt) == COND_EXPR
2953 || TREE_CODE (stmt) == SWITCH_EXPR));
2955 if (may_optimize_p)
2956 may_have_exposed_new_symbols
2957 |= eliminate_redundant_computations (walk_data, stmt, ann);
2959 /* Record any additional equivalences created by this statement. */
2960 if (TREE_CODE (stmt) == MODIFY_EXPR)
2961 record_equivalences_from_stmt (stmt,
2962 may_optimize_p,
2963 ann);
2965 register_definitions_for_stmt (stmt);
2967 /* If STMT is a COND_EXPR and it was modified, then we may know
2968 where it goes. If that is the case, then mark the CFG as altered.
2970 This will cause us to later call remove_unreachable_blocks and
2971 cleanup_tree_cfg when it is safe to do so. It is not safe to
2972 clean things up here since removal of edges and such can trigger
2973 the removal of PHI nodes, which in turn can release SSA_NAMEs to
2974 the manager.
2976 That's all fine and good, except that once SSA_NAMEs are released
2977 to the manager, we must not call create_ssa_name until all references
2978 to released SSA_NAMEs have been eliminated.
2980 All references to the deleted SSA_NAMEs can not be eliminated until
2981 we remove unreachable blocks.
2983 We can not remove unreachable blocks until after we have completed
2984 any queued jump threading.
2986 We can not complete any queued jump threads until we have taken
2987 appropriate variables out of SSA form. Taking variables out of
2988 SSA form can call create_ssa_name and thus we lose.
2990 Ultimately I suspect we're going to need to change the interface
2991 into the SSA_NAME manager. */
2993 if (ann->modified)
2995 tree val = NULL;
2997 if (TREE_CODE (stmt) == COND_EXPR)
2998 val = COND_EXPR_COND (stmt);
2999 else if (TREE_CODE (stmt) == SWITCH_EXPR)
3000 val = SWITCH_COND (stmt);
3002 if (val && TREE_CODE (val) == INTEGER_CST && find_taken_edge (bb, val))
3003 cfg_altered = true;
3005 /* If we simplified a statement in such a way as to be shown that it
3006 cannot trap, update the eh information and the cfg to match. */
3007 if (maybe_clean_eh_stmt (stmt))
3009 bitmap_set_bit (need_eh_cleanup, bb->index);
3010 if (dump_file && (dump_flags & TDF_DETAILS))
3011 fprintf (dump_file, " Flagged to clear EH edges.\n");
3015 if (may_have_exposed_new_symbols)
3016 VEC_safe_push (tree_on_heap, stmts_to_rescan, bsi_stmt (si));
3019 /* Replace the RHS of STMT with NEW_RHS. If RHS can be found in the
3020 available expression hashtable, then return the LHS from the hash
3021 table.
3023 If INSERT is true, then we also update the available expression
3024 hash table to account for the changes made to STMT. */
3026 static tree
3027 update_rhs_and_lookup_avail_expr (tree stmt, tree new_rhs, bool insert)
3029 tree cached_lhs = NULL;
3031 /* Remove the old entry from the hash table. */
3032 if (insert)
3034 struct expr_hash_elt element;
3036 initialize_hash_element (stmt, NULL, &element);
3037 htab_remove_elt_with_hash (avail_exprs, &element, element.hash);
3040 /* Now update the RHS of the assignment. */
3041 TREE_OPERAND (stmt, 1) = new_rhs;
3043 /* Now lookup the updated statement in the hash table. */
3044 cached_lhs = lookup_avail_expr (stmt, insert);
3046 /* We have now called lookup_avail_expr twice with two different
3047 versions of this same statement, once in optimize_stmt, once here.
3049 We know the call in optimize_stmt did not find an existing entry
3050 in the hash table, so a new entry was created. At the same time
3051 this statement was pushed onto the AVAIL_EXPRS_STACK vector.
3053 If this call failed to find an existing entry on the hash table,
3054 then the new version of this statement was entered into the
3055 hash table. And this statement was pushed onto BLOCK_AVAIL_EXPR
3056 for the second time. So there are two copies on BLOCK_AVAIL_EXPRs
3058 If this call succeeded, we still have one copy of this statement
3059 on the BLOCK_AVAIL_EXPRs vector.
3061 For both cases, we need to pop the most recent entry off the
3062 BLOCK_AVAIL_EXPRs vector. For the case where we never found this
3063 statement in the hash tables, that will leave precisely one
3064 copy of this statement on BLOCK_AVAIL_EXPRs. For the case where
3065 we found a copy of this statement in the second hash table lookup
3066 we want _no_ copies of this statement in BLOCK_AVAIL_EXPRs. */
3067 if (insert)
3068 VEC_pop (tree_on_heap, avail_exprs_stack);
3070 /* And make sure we record the fact that we modified this
3071 statement. */
3072 modify_stmt (stmt);
3074 return cached_lhs;
3077 /* Search for an existing instance of STMT in the AVAIL_EXPRS table. If
3078 found, return its LHS. Otherwise insert STMT in the table and return
3079 NULL_TREE.
3081 Also, when an expression is first inserted in the AVAIL_EXPRS table, it
3082 is also added to the stack pointed by BLOCK_AVAIL_EXPRS_P, so that they
3083 can be removed when we finish processing this block and its children.
3085 NOTE: This function assumes that STMT is a MODIFY_EXPR node that
3086 contains no CALL_EXPR on its RHS and makes no volatile nor
3087 aliased references. */
3089 static tree
3090 lookup_avail_expr (tree stmt, bool insert)
3092 void **slot;
3093 tree lhs;
3094 tree temp;
3095 struct expr_hash_elt *element = xcalloc (sizeof (struct expr_hash_elt), 1);
3097 lhs = TREE_CODE (stmt) == MODIFY_EXPR ? TREE_OPERAND (stmt, 0) : NULL;
3099 initialize_hash_element (stmt, lhs, element);
3101 /* Don't bother remembering constant assignments and copy operations.
3102 Constants and copy operations are handled by the constant/copy propagator
3103 in optimize_stmt. */
3104 if (TREE_CODE (element->rhs) == SSA_NAME
3105 || is_gimple_min_invariant (element->rhs))
3107 free (element);
3108 return NULL_TREE;
3111 /* If this is an equality test against zero, see if we have recorded a
3112 nonzero value for the variable in question. */
3113 if ((TREE_CODE (element->rhs) == EQ_EXPR
3114 || TREE_CODE (element->rhs) == NE_EXPR)
3115 && TREE_CODE (TREE_OPERAND (element->rhs, 0)) == SSA_NAME
3116 && integer_zerop (TREE_OPERAND (element->rhs, 1)))
3118 int indx = SSA_NAME_VERSION (TREE_OPERAND (element->rhs, 0));
3120 if (bitmap_bit_p (nonzero_vars, indx))
3122 tree t = element->rhs;
3123 free (element);
3125 if (TREE_CODE (t) == EQ_EXPR)
3126 return boolean_false_node;
3127 else
3128 return boolean_true_node;
3132 /* Finally try to find the expression in the main expression hash table. */
3133 slot = htab_find_slot_with_hash (avail_exprs, element, element->hash,
3134 (insert ? INSERT : NO_INSERT));
3135 if (slot == NULL)
3137 free (element);
3138 return NULL_TREE;
3141 if (*slot == NULL)
3143 *slot = (void *) element;
3144 VEC_safe_push (tree_on_heap, avail_exprs_stack,
3145 stmt ? stmt : element->rhs);
3146 return NULL_TREE;
3149 /* Extract the LHS of the assignment so that it can be used as the current
3150 definition of another variable. */
3151 lhs = ((struct expr_hash_elt *)*slot)->lhs;
3153 /* See if the LHS appears in the CONST_AND_COPIES table. If it does, then
3154 use the value from the const_and_copies table. */
3155 if (TREE_CODE (lhs) == SSA_NAME)
3157 temp = SSA_NAME_VALUE (lhs);
3158 if (temp && TREE_CODE (temp) != VALUE_HANDLE)
3159 lhs = temp;
3162 free (element);
3163 return lhs;
3166 /* Given a condition COND, record into HI_P, LO_P and INVERTED_P the
3167 range of values that result in the conditional having a true value.
3169 Return true if we are successful in extracting a range from COND and
3170 false if we are unsuccessful. */
3172 static bool
3173 extract_range_from_cond (tree cond, tree *hi_p, tree *lo_p, int *inverted_p)
3175 tree op1 = TREE_OPERAND (cond, 1);
3176 tree high, low, type;
3177 int inverted;
3179 /* Experiments have shown that it's rarely, if ever useful to
3180 record ranges for enumerations. Presumably this is due to
3181 the fact that they're rarely used directly. They are typically
3182 cast into an integer type and used that way. */
3183 if (TREE_CODE (TREE_TYPE (op1)) != INTEGER_TYPE)
3184 return 0;
3186 type = TREE_TYPE (op1);
3188 switch (TREE_CODE (cond))
3190 case EQ_EXPR:
3191 high = low = op1;
3192 inverted = 0;
3193 break;
3195 case NE_EXPR:
3196 high = low = op1;
3197 inverted = 1;
3198 break;
3200 case GE_EXPR:
3201 low = op1;
3202 high = TYPE_MAX_VALUE (type);
3203 inverted = 0;
3204 break;
3206 case GT_EXPR:
3207 low = int_const_binop (PLUS_EXPR, op1, integer_one_node, 1);
3208 high = TYPE_MAX_VALUE (type);
3209 inverted = 0;
3210 break;
3212 case LE_EXPR:
3213 high = op1;
3214 low = TYPE_MIN_VALUE (type);
3215 inverted = 0;
3216 break;
3218 case LT_EXPR:
3219 high = int_const_binop (MINUS_EXPR, op1, integer_one_node, 1);
3220 low = TYPE_MIN_VALUE (type);
3221 inverted = 0;
3222 break;
3224 default:
3225 return 0;
3228 *hi_p = high;
3229 *lo_p = low;
3230 *inverted_p = inverted;
3231 return 1;
3234 /* Record a range created by COND for basic block BB. */
3236 static void
3237 record_range (tree cond, basic_block bb)
3239 enum tree_code code = TREE_CODE (cond);
3241 /* We explicitly ignore NE_EXPRs and all the unordered comparisons.
3242 They rarely allow for meaningful range optimizations and significantly
3243 complicate the implementation. */
3244 if ((code == LT_EXPR || code == LE_EXPR || code == GT_EXPR
3245 || code == GE_EXPR || code == EQ_EXPR)
3246 && TREE_CODE (TREE_TYPE (TREE_OPERAND (cond, 1))) == INTEGER_TYPE)
3248 struct vrp_hash_elt *vrp_hash_elt;
3249 struct vrp_element *element;
3250 varray_type *vrp_records_p;
3251 void **slot;
3254 vrp_hash_elt = xmalloc (sizeof (struct vrp_hash_elt));
3255 vrp_hash_elt->var = TREE_OPERAND (cond, 0);
3256 vrp_hash_elt->records = NULL;
3257 slot = htab_find_slot (vrp_data, vrp_hash_elt, INSERT);
3259 if (*slot == NULL)
3260 *slot = (void *) vrp_hash_elt;
3261 else
3262 free (vrp_hash_elt);
3264 vrp_hash_elt = (struct vrp_hash_elt *) *slot;
3265 vrp_records_p = &vrp_hash_elt->records;
3267 element = ggc_alloc (sizeof (struct vrp_element));
3268 element->low = NULL;
3269 element->high = NULL;
3270 element->cond = cond;
3271 element->bb = bb;
3273 if (*vrp_records_p == NULL)
3274 VARRAY_GENERIC_PTR_INIT (*vrp_records_p, 2, "vrp records");
3276 VARRAY_PUSH_GENERIC_PTR (*vrp_records_p, element);
3277 VEC_safe_push (tree_on_heap, vrp_variables_stack, TREE_OPERAND (cond, 0));
3281 /* Hashing and equality functions for VRP_DATA.
3283 Since this hash table is addressed by SSA_NAMEs, we can hash on
3284 their version number and equality can be determined with a
3285 pointer comparison. */
3287 static hashval_t
3288 vrp_hash (const void *p)
3290 tree var = ((struct vrp_hash_elt *)p)->var;
3292 return SSA_NAME_VERSION (var);
3295 static int
3296 vrp_eq (const void *p1, const void *p2)
3298 tree var1 = ((struct vrp_hash_elt *)p1)->var;
3299 tree var2 = ((struct vrp_hash_elt *)p2)->var;
3301 return var1 == var2;
3304 /* Hashing and equality functions for AVAIL_EXPRS. The table stores
3305 MODIFY_EXPR statements. We compute a value number for expressions using
3306 the code of the expression and the SSA numbers of its operands. */
3308 static hashval_t
3309 avail_expr_hash (const void *p)
3311 stmt_ann_t ann = ((struct expr_hash_elt *)p)->ann;
3312 tree rhs = ((struct expr_hash_elt *)p)->rhs;
3313 hashval_t val = 0;
3314 size_t i;
3315 vuse_optype vuses;
3317 /* iterative_hash_expr knows how to deal with any expression and
3318 deals with commutative operators as well, so just use it instead
3319 of duplicating such complexities here. */
3320 val = iterative_hash_expr (rhs, val);
3322 /* If the hash table entry is not associated with a statement, then we
3323 can just hash the expression and not worry about virtual operands
3324 and such. */
3325 if (!ann)
3326 return val;
3328 /* Add the SSA version numbers of every vuse operand. This is important
3329 because compound variables like arrays are not renamed in the
3330 operands. Rather, the rename is done on the virtual variable
3331 representing all the elements of the array. */
3332 vuses = VUSE_OPS (ann);
3333 for (i = 0; i < NUM_VUSES (vuses); i++)
3334 val = iterative_hash_expr (VUSE_OP (vuses, i), val);
3336 return val;
3339 static hashval_t
3340 real_avail_expr_hash (const void *p)
3342 return ((const struct expr_hash_elt *)p)->hash;
3345 static int
3346 avail_expr_eq (const void *p1, const void *p2)
3348 stmt_ann_t ann1 = ((struct expr_hash_elt *)p1)->ann;
3349 tree rhs1 = ((struct expr_hash_elt *)p1)->rhs;
3350 stmt_ann_t ann2 = ((struct expr_hash_elt *)p2)->ann;
3351 tree rhs2 = ((struct expr_hash_elt *)p2)->rhs;
3353 /* If they are the same physical expression, return true. */
3354 if (rhs1 == rhs2 && ann1 == ann2)
3355 return true;
3357 /* If their codes are not equal, then quit now. */
3358 if (TREE_CODE (rhs1) != TREE_CODE (rhs2))
3359 return false;
3361 /* In case of a collision, both RHS have to be identical and have the
3362 same VUSE operands. */
3363 if ((TREE_TYPE (rhs1) == TREE_TYPE (rhs2)
3364 || lang_hooks.types_compatible_p (TREE_TYPE (rhs1), TREE_TYPE (rhs2)))
3365 && operand_equal_p (rhs1, rhs2, OEP_PURE_SAME))
3367 vuse_optype ops1 = NULL;
3368 vuse_optype ops2 = NULL;
3369 size_t num_ops1 = 0;
3370 size_t num_ops2 = 0;
3371 size_t i;
3373 if (ann1)
3375 ops1 = VUSE_OPS (ann1);
3376 num_ops1 = NUM_VUSES (ops1);
3379 if (ann2)
3381 ops2 = VUSE_OPS (ann2);
3382 num_ops2 = NUM_VUSES (ops2);
3385 /* If the number of virtual uses is different, then we consider
3386 them not equal. */
3387 if (num_ops1 != num_ops2)
3388 return false;
3390 for (i = 0; i < num_ops1; i++)
3391 if (VUSE_OP (ops1, i) != VUSE_OP (ops2, i))
3392 return false;
3394 gcc_assert (((struct expr_hash_elt *)p1)->hash
3395 == ((struct expr_hash_elt *)p2)->hash);
3396 return true;
3399 return false;
3402 /* Given STMT and a pointer to the block local definitions BLOCK_DEFS_P,
3403 register register all objects set by this statement into BLOCK_DEFS_P
3404 and CURRDEFS. */
3406 static void
3407 register_definitions_for_stmt (tree stmt)
3409 tree def;
3410 ssa_op_iter iter;
3412 FOR_EACH_SSA_TREE_OPERAND (def, stmt, iter, SSA_OP_ALL_DEFS)
3415 /* FIXME: We shouldn't be registering new defs if the variable
3416 doesn't need to be renamed. */
3417 register_new_def (def, &block_defs_stack);